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 (mp->m_flags & XFS_MOUNT_NORECOVERY)
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.
492 * We avoid processing actively referenced iclogs so that we don't run callbacks
493 * while the iclog owner might still be preparing the iclog for IO submssion.
494 * These will be caught by xlog_state_iclog_release() and call this function
495 * again to process any callbacks that may have been added to that iclog.
498 xlog_state_shutdown_callbacks(
501 struct xlog_in_core *iclog;
504 spin_lock(&log->l_icloglock);
505 iclog = log->l_iclog;
507 if (atomic_read(&iclog->ic_refcnt)) {
508 /* Reference holder will re-run iclog callbacks. */
511 list_splice_init(&iclog->ic_callbacks, &cb_list);
512 wake_up_all(&iclog->ic_write_wait);
513 wake_up_all(&iclog->ic_force_wait);
514 } while ((iclog = iclog->ic_next) != log->l_iclog);
516 wake_up_all(&log->l_flush_wait);
517 spin_unlock(&log->l_icloglock);
519 xlog_cil_process_committed(&cb_list);
523 * Flush iclog to disk if this is the last reference to the given iclog and the
524 * it is in the WANT_SYNC state.
526 * If the caller passes in a non-zero @old_tail_lsn and the current log tail
527 * does not match, there may be metadata on disk that must be persisted before
528 * this iclog is written. To satisfy that requirement, set the
529 * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
532 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
533 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
534 * written to stable storage, and implies that a commit record is contained
535 * within the iclog. We need to ensure that the log tail does not move beyond
536 * the tail that the first commit record in the iclog ordered against, otherwise
537 * correct recovery of that checkpoint becomes dependent on future operations
538 * performed on this iclog.
540 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
541 * current tail into iclog. Once the iclog tail is set, future operations must
542 * not modify it, otherwise they potentially violate ordering constraints for
543 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
544 * the iclog will get zeroed on activation of the iclog after sync, so we
545 * always capture the tail lsn on the iclog on the first NEED_FUA release
546 * regardless of the number of active reference counts on this iclog.
550 xlog_state_release_iclog(
552 struct xlog_in_core *iclog,
553 xfs_lsn_t old_tail_lsn)
558 lockdep_assert_held(&log->l_icloglock);
560 trace_xlog_iclog_release(iclog, _RET_IP_);
562 * Grabbing the current log tail needs to be atomic w.r.t. the writing
563 * of the tail LSN into the iclog so we guarantee that the log tail does
564 * not move between deciding if a cache flush is required and writing
565 * the LSN into the iclog below.
567 if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
568 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
570 if (old_tail_lsn && tail_lsn != old_tail_lsn)
571 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
573 if ((iclog->ic_flags & XLOG_ICL_NEED_FUA) &&
574 !iclog->ic_header.h_tail_lsn)
575 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
578 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
580 if (xlog_is_shutdown(log)) {
582 * If there are no more references to this iclog, process the
583 * pending iclog callbacks that were waiting on the release of
587 spin_unlock(&log->l_icloglock);
588 xlog_state_shutdown_callbacks(log);
589 spin_lock(&log->l_icloglock);
597 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
598 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
602 iclog->ic_state = XLOG_STATE_SYNCING;
603 if (!iclog->ic_header.h_tail_lsn)
604 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
605 xlog_verify_tail_lsn(log, iclog);
606 trace_xlog_iclog_syncing(iclog, _RET_IP_);
608 spin_unlock(&log->l_icloglock);
609 xlog_sync(log, iclog);
610 spin_lock(&log->l_icloglock);
615 * Mount a log filesystem
617 * mp - ubiquitous xfs mount point structure
618 * log_target - buftarg of on-disk log device
619 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
620 * num_bblocks - Number of BBSIZE blocks in on-disk log
622 * Return error or zero.
627 xfs_buftarg_t *log_target,
628 xfs_daddr_t blk_offset,
632 bool fatal = xfs_has_crc(mp);
636 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
637 xfs_notice(mp, "Mounting V%d Filesystem",
638 XFS_SB_VERSION_NUM(&mp->m_sb));
641 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
642 XFS_SB_VERSION_NUM(&mp->m_sb));
643 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
646 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
648 error = PTR_ERR(log);
654 * Validate the given log space and drop a critical message via syslog
655 * if the log size is too small that would lead to some unexpected
656 * situations in transaction log space reservation stage.
658 * Note: we can't just reject the mount if the validation fails. This
659 * would mean that people would have to downgrade their kernel just to
660 * remedy the situation as there is no way to grow the log (short of
661 * black magic surgery with xfs_db).
663 * We can, however, reject mounts for CRC format filesystems, as the
664 * mkfs binary being used to make the filesystem should never create a
665 * filesystem with a log that is too small.
667 min_logfsbs = xfs_log_calc_minimum_size(mp);
669 if (mp->m_sb.sb_logblocks < min_logfsbs) {
671 "Log size %d blocks too small, minimum size is %d blocks",
672 mp->m_sb.sb_logblocks, min_logfsbs);
674 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
676 "Log size %d blocks too large, maximum size is %lld blocks",
677 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
679 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
681 "log size %lld bytes too large, maximum size is %lld bytes",
682 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
685 } else if (mp->m_sb.sb_logsunit > 1 &&
686 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
688 "log stripe unit %u bytes must be a multiple of block size",
689 mp->m_sb.sb_logsunit);
695 * Log check errors are always fatal on v5; or whenever bad
696 * metadata leads to a crash.
699 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
703 xfs_crit(mp, "Log size out of supported range.");
705 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
709 * Initialize the AIL now we have a log.
711 error = xfs_trans_ail_init(mp);
713 xfs_warn(mp, "AIL initialisation failed: error %d", error);
716 log->l_ailp = mp->m_ail;
719 * skip log recovery on a norecovery mount. pretend it all
722 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
723 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
726 mp->m_flags &= ~XFS_MOUNT_RDONLY;
728 error = xlog_recover(log);
731 mp->m_flags |= XFS_MOUNT_RDONLY;
733 xfs_warn(mp, "log mount/recovery failed: error %d",
735 xlog_recover_cancel(log);
736 goto out_destroy_ail;
740 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
743 goto out_destroy_ail;
745 /* Normal transactions can now occur */
746 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
749 * Now the log has been fully initialised and we know were our
750 * space grant counters are, we can initialise the permanent ticket
751 * needed for delayed logging to work.
753 xlog_cil_init_post_recovery(log);
758 xfs_trans_ail_destroy(mp);
760 xlog_dealloc_log(log);
766 * Finish the recovery of the file system. This is separate from the
767 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
768 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
771 * If we finish recovery successfully, start the background log work. If we are
772 * not doing recovery, then we have a RO filesystem and we don't need to start
776 xfs_log_mount_finish(
777 struct xfs_mount *mp)
779 struct xlog *log = mp->m_log;
780 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
783 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
784 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
786 } else if (readonly) {
787 /* Allow unlinked processing to proceed */
788 mp->m_flags &= ~XFS_MOUNT_RDONLY;
792 * During the second phase of log recovery, we need iget and
793 * iput to behave like they do for an active filesystem.
794 * xfs_fs_drop_inode needs to be able to prevent the deletion
795 * of inodes before we're done replaying log items on those
796 * inodes. Turn it off immediately after recovery finishes
797 * so that we don't leak the quota inodes if subsequent mount
800 * We let all inodes involved in redo item processing end up on
801 * the LRU instead of being evicted immediately so that if we do
802 * something to an unlinked inode, the irele won't cause
803 * premature truncation and freeing of the inode, which results
804 * in log recovery failure. We have to evict the unreferenced
805 * lru inodes after clearing SB_ACTIVE because we don't
806 * otherwise clean up the lru if there's a subsequent failure in
807 * xfs_mountfs, which leads to us leaking the inodes if nothing
808 * else (e.g. quotacheck) references the inodes before the
809 * mount failure occurs.
811 mp->m_super->s_flags |= SB_ACTIVE;
812 if (xlog_recovery_needed(log))
813 error = xlog_recover_finish(log);
815 xfs_log_work_queue(mp);
816 mp->m_super->s_flags &= ~SB_ACTIVE;
817 evict_inodes(mp->m_super);
820 * Drain the buffer LRU after log recovery. This is required for v4
821 * filesystems to avoid leaving around buffers with NULL verifier ops,
822 * but we do it unconditionally to make sure we're always in a clean
823 * cache state after mount.
825 * Don't push in the error case because the AIL may have pending intents
826 * that aren't removed until recovery is cancelled.
828 if (xlog_recovery_needed(log)) {
830 xfs_log_force(mp, XFS_LOG_SYNC);
831 xfs_ail_push_all_sync(mp->m_ail);
833 xfs_notice(mp, "Ending recovery (logdev: %s)",
834 mp->m_logname ? mp->m_logname : "internal");
836 xfs_info(mp, "Ending clean mount");
838 xfs_buftarg_drain(mp->m_ddev_targp);
840 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
842 mp->m_flags |= XFS_MOUNT_RDONLY;
844 /* Make sure the log is dead if we're returning failure. */
845 ASSERT(!error || xlog_is_shutdown(log));
851 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
855 xfs_log_mount_cancel(
856 struct xfs_mount *mp)
858 xlog_recover_cancel(mp->m_log);
863 * Flush out the iclog to disk ensuring that device caches are flushed and
864 * the iclog hits stable storage before any completion waiters are woken.
868 struct xlog_in_core *iclog)
870 atomic_inc(&iclog->ic_refcnt);
871 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
872 if (iclog->ic_state == XLOG_STATE_ACTIVE)
873 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
874 return xlog_state_release_iclog(iclog->ic_log, iclog, 0);
878 * Wait for the iclog and all prior iclogs to be written disk as required by the
879 * log force state machine. Waiting on ic_force_wait ensures iclog completions
880 * have been ordered and callbacks run before we are woken here, hence
881 * guaranteeing that all the iclogs up to this one are on stable storage.
885 struct xlog_in_core *iclog)
886 __releases(iclog->ic_log->l_icloglock)
888 struct xlog *log = iclog->ic_log;
890 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
891 if (!xlog_is_shutdown(log) &&
892 iclog->ic_state != XLOG_STATE_ACTIVE &&
893 iclog->ic_state != XLOG_STATE_DIRTY) {
894 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
895 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
897 spin_unlock(&log->l_icloglock);
900 if (xlog_is_shutdown(log))
906 * Write out an unmount record using the ticket provided. We have to account for
907 * the data space used in the unmount ticket as this write is not done from a
908 * transaction context that has already done the accounting for us.
911 xlog_write_unmount_record(
913 struct xlog_ticket *ticket)
915 struct xfs_unmount_log_format ulf = {
916 .magic = XLOG_UNMOUNT_TYPE,
918 struct xfs_log_iovec reg = {
920 .i_len = sizeof(ulf),
921 .i_type = XLOG_REG_TYPE_UNMOUNT,
923 struct xfs_log_vec vec = {
928 /* account for space used by record data */
929 ticket->t_curr_res -= sizeof(ulf);
931 return xlog_write(log, NULL, &vec, ticket, XLOG_UNMOUNT_TRANS);
935 * Mark the filesystem clean by writing an unmount record to the head of the
942 struct xfs_mount *mp = log->l_mp;
943 struct xlog_in_core *iclog;
944 struct xlog_ticket *tic = NULL;
947 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
951 error = xlog_write_unmount_record(log, tic);
953 * At this point, we're umounting anyway, so there's no point in
954 * transitioning log state to shutdown. Just continue...
958 xfs_alert(mp, "%s: unmount record failed", __func__);
960 spin_lock(&log->l_icloglock);
961 iclog = log->l_iclog;
962 error = xlog_force_iclog(iclog);
963 xlog_wait_on_iclog(iclog);
966 trace_xfs_log_umount_write(log, tic);
967 xfs_log_ticket_ungrant(log, tic);
972 xfs_log_unmount_verify_iclog(
975 struct xlog_in_core *iclog = log->l_iclog;
978 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
979 ASSERT(iclog->ic_offset == 0);
980 } while ((iclog = iclog->ic_next) != log->l_iclog);
984 * Unmount record used to have a string "Unmount filesystem--" in the
985 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
986 * We just write the magic number now since that particular field isn't
987 * currently architecture converted and "Unmount" is a bit foo.
988 * As far as I know, there weren't any dependencies on the old behaviour.
991 xfs_log_unmount_write(
992 struct xfs_mount *mp)
994 struct xlog *log = mp->m_log;
996 if (!xfs_log_writable(mp))
999 xfs_log_force(mp, XFS_LOG_SYNC);
1001 if (xlog_is_shutdown(log))
1005 * If we think the summary counters are bad, avoid writing the unmount
1006 * record to force log recovery at next mount, after which the summary
1007 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1010 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1011 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1012 xfs_alert(mp, "%s: will fix summary counters at next mount",
1017 xfs_log_unmount_verify_iclog(log);
1018 xlog_unmount_write(log);
1022 * Empty the log for unmount/freeze.
1024 * To do this, we first need to shut down the background log work so it is not
1025 * trying to cover the log as we clean up. We then need to unpin all objects in
1026 * the log so we can then flush them out. Once they have completed their IO and
1027 * run the callbacks removing themselves from the AIL, we can cover the log.
1031 struct xfs_mount *mp)
1034 * Clear log incompat features since we're quiescing the log. Report
1035 * failures, though it's not fatal to have a higher log feature
1036 * protection level than the log contents actually require.
1038 if (xfs_clear_incompat_log_features(mp)) {
1041 error = xfs_sync_sb(mp, false);
1044 "Failed to clear log incompat features on quiesce");
1047 cancel_delayed_work_sync(&mp->m_log->l_work);
1048 xfs_log_force(mp, XFS_LOG_SYNC);
1051 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1052 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1053 * xfs_buf_iowait() cannot be used because it was pushed with the
1054 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1055 * the IO to complete.
1057 xfs_ail_push_all_sync(mp->m_ail);
1058 xfs_buftarg_wait(mp->m_ddev_targp);
1059 xfs_buf_lock(mp->m_sb_bp);
1060 xfs_buf_unlock(mp->m_sb_bp);
1062 return xfs_log_cover(mp);
1067 struct xfs_mount *mp)
1069 xfs_log_quiesce(mp);
1070 xfs_log_unmount_write(mp);
1074 * Shut down and release the AIL and Log.
1076 * During unmount, we need to ensure we flush all the dirty metadata objects
1077 * from the AIL so that the log is empty before we write the unmount record to
1078 * the log. Once this is done, we can tear down the AIL and the log.
1082 struct xfs_mount *mp)
1086 xfs_buftarg_drain(mp->m_ddev_targp);
1088 xfs_trans_ail_destroy(mp);
1090 xfs_sysfs_del(&mp->m_log->l_kobj);
1092 xlog_dealloc_log(mp->m_log);
1097 struct xfs_mount *mp,
1098 struct xfs_log_item *item,
1100 const struct xfs_item_ops *ops)
1102 item->li_mountp = mp;
1103 item->li_ailp = mp->m_ail;
1104 item->li_type = type;
1108 INIT_LIST_HEAD(&item->li_ail);
1109 INIT_LIST_HEAD(&item->li_cil);
1110 INIT_LIST_HEAD(&item->li_bio_list);
1111 INIT_LIST_HEAD(&item->li_trans);
1115 * Wake up processes waiting for log space after we have moved the log tail.
1119 struct xfs_mount *mp)
1121 struct xlog *log = mp->m_log;
1124 if (xlog_is_shutdown(log))
1127 if (!list_empty_careful(&log->l_write_head.waiters)) {
1128 ASSERT(!xlog_in_recovery(log));
1130 spin_lock(&log->l_write_head.lock);
1131 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1132 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1133 spin_unlock(&log->l_write_head.lock);
1136 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1137 ASSERT(!xlog_in_recovery(log));
1139 spin_lock(&log->l_reserve_head.lock);
1140 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1141 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1142 spin_unlock(&log->l_reserve_head.lock);
1147 * Determine if we have a transaction that has gone to disk that needs to be
1148 * covered. To begin the transition to the idle state firstly the log needs to
1149 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1150 * we start attempting to cover the log.
1152 * Only if we are then in a state where covering is needed, the caller is
1153 * informed that dummy transactions are required to move the log into the idle
1156 * If there are any items in the AIl or CIL, then we do not want to attempt to
1157 * cover the log as we may be in a situation where there isn't log space
1158 * available to run a dummy transaction and this can lead to deadlocks when the
1159 * tail of the log is pinned by an item that is modified in the CIL. Hence
1160 * there's no point in running a dummy transaction at this point because we
1161 * can't start trying to idle the log until both the CIL and AIL are empty.
1164 xfs_log_need_covered(
1165 struct xfs_mount *mp)
1167 struct xlog *log = mp->m_log;
1168 bool needed = false;
1170 if (!xlog_cil_empty(log))
1173 spin_lock(&log->l_icloglock);
1174 switch (log->l_covered_state) {
1175 case XLOG_STATE_COVER_DONE:
1176 case XLOG_STATE_COVER_DONE2:
1177 case XLOG_STATE_COVER_IDLE:
1179 case XLOG_STATE_COVER_NEED:
1180 case XLOG_STATE_COVER_NEED2:
1181 if (xfs_ail_min_lsn(log->l_ailp))
1183 if (!xlog_iclogs_empty(log))
1187 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1188 log->l_covered_state = XLOG_STATE_COVER_DONE;
1190 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1196 spin_unlock(&log->l_icloglock);
1201 * Explicitly cover the log. This is similar to background log covering but
1202 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1203 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1204 * must all be empty.
1208 struct xfs_mount *mp)
1213 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1214 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1215 xlog_is_shutdown(mp->m_log));
1217 if (!xfs_log_writable(mp))
1221 * xfs_log_need_covered() is not idempotent because it progresses the
1222 * state machine if the log requires covering. Therefore, we must call
1223 * this function once and use the result until we've issued an sb sync.
1224 * Do so first to make that abundantly clear.
1226 * Fall into the covering sequence if the log needs covering or the
1227 * mount has lazy superblock accounting to sync to disk. The sb sync
1228 * used for covering accumulates the in-core counters, so covering
1229 * handles this for us.
1231 need_covered = xfs_log_need_covered(mp);
1232 if (!need_covered && !xfs_has_lazysbcount(mp))
1236 * To cover the log, commit the superblock twice (at most) in
1237 * independent checkpoints. The first serves as a reference for the
1238 * tail pointer. The sync transaction and AIL push empties the AIL and
1239 * updates the in-core tail to the LSN of the first checkpoint. The
1240 * second commit updates the on-disk tail with the in-core LSN,
1241 * covering the log. Push the AIL one more time to leave it empty, as
1245 error = xfs_sync_sb(mp, true);
1248 xfs_ail_push_all_sync(mp->m_ail);
1249 } while (xfs_log_need_covered(mp));
1255 * We may be holding the log iclog lock upon entering this routine.
1258 xlog_assign_tail_lsn_locked(
1259 struct xfs_mount *mp)
1261 struct xlog *log = mp->m_log;
1262 struct xfs_log_item *lip;
1265 assert_spin_locked(&mp->m_ail->ail_lock);
1268 * To make sure we always have a valid LSN for the log tail we keep
1269 * track of the last LSN which was committed in log->l_last_sync_lsn,
1270 * and use that when the AIL was empty.
1272 lip = xfs_ail_min(mp->m_ail);
1274 tail_lsn = lip->li_lsn;
1276 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1277 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1278 atomic64_set(&log->l_tail_lsn, tail_lsn);
1283 xlog_assign_tail_lsn(
1284 struct xfs_mount *mp)
1288 spin_lock(&mp->m_ail->ail_lock);
1289 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1290 spin_unlock(&mp->m_ail->ail_lock);
1296 * Return the space in the log between the tail and the head. The head
1297 * is passed in the cycle/bytes formal parms. In the special case where
1298 * the reserve head has wrapped passed the tail, this calculation is no
1299 * longer valid. In this case, just return 0 which means there is no space
1300 * in the log. This works for all places where this function is called
1301 * with the reserve head. Of course, if the write head were to ever
1302 * wrap the tail, we should blow up. Rather than catch this case here,
1303 * we depend on other ASSERTions in other parts of the code. XXXmiken
1305 * If reservation head is behind the tail, we have a problem. Warn about it,
1306 * but then treat it as if the log is empty.
1308 * If the log is shut down, the head and tail may be invalid or out of whack, so
1309 * shortcut invalidity asserts in this case so that we don't trigger them
1322 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1323 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1324 tail_bytes = BBTOB(tail_bytes);
1325 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1326 return log->l_logsize - (head_bytes - tail_bytes);
1327 if (tail_cycle + 1 < head_cycle)
1330 /* Ignore potential inconsistency when shutdown. */
1331 if (xlog_is_shutdown(log))
1332 return log->l_logsize;
1334 if (tail_cycle < head_cycle) {
1335 ASSERT(tail_cycle == (head_cycle - 1));
1336 return tail_bytes - head_bytes;
1340 * The reservation head is behind the tail. In this case we just want to
1341 * return the size of the log as the amount of space left.
1343 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1344 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1345 tail_cycle, tail_bytes);
1346 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1347 head_cycle, head_bytes);
1349 return log->l_logsize;
1355 struct work_struct *work)
1357 struct xlog_in_core *iclog =
1358 container_of(work, struct xlog_in_core, ic_end_io_work);
1359 struct xlog *log = iclog->ic_log;
1362 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1364 /* treat writes with injected CRC errors as failed */
1365 if (iclog->ic_fail_crc)
1370 * Race to shutdown the filesystem if we see an error.
1372 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1373 xfs_alert(log->l_mp, "log I/O error %d", error);
1374 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1377 xlog_state_done_syncing(iclog);
1378 bio_uninit(&iclog->ic_bio);
1381 * Drop the lock to signal that we are done. Nothing references the
1382 * iclog after this, so an unmount waiting on this lock can now tear it
1383 * down safely. As such, it is unsafe to reference the iclog after the
1384 * unlock as we could race with it being freed.
1386 up(&iclog->ic_sema);
1390 * Return size of each in-core log record buffer.
1392 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1394 * If the filesystem blocksize is too large, we may need to choose a
1395 * larger size since the directory code currently logs entire blocks.
1398 xlog_get_iclog_buffer_size(
1399 struct xfs_mount *mp,
1402 if (mp->m_logbufs <= 0)
1403 mp->m_logbufs = XLOG_MAX_ICLOGS;
1404 if (mp->m_logbsize <= 0)
1405 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1407 log->l_iclog_bufs = mp->m_logbufs;
1408 log->l_iclog_size = mp->m_logbsize;
1411 * # headers = size / 32k - one header holds cycles from 32k of data.
1413 log->l_iclog_heads =
1414 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1415 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1420 struct xfs_mount *mp)
1422 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1423 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1427 * Clear the log incompat flags if we have the opportunity.
1429 * This only happens if we're about to log the second dummy transaction as part
1430 * of covering the log and we can get the log incompat feature usage lock.
1433 xlog_clear_incompat(
1436 struct xfs_mount *mp = log->l_mp;
1438 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1439 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1442 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1445 if (!down_write_trylock(&log->l_incompat_users))
1448 xfs_clear_incompat_log_features(mp);
1449 up_write(&log->l_incompat_users);
1453 * Every sync period we need to unpin all items in the AIL and push them to
1454 * disk. If there is nothing dirty, then we might need to cover the log to
1455 * indicate that the filesystem is idle.
1459 struct work_struct *work)
1461 struct xlog *log = container_of(to_delayed_work(work),
1462 struct xlog, l_work);
1463 struct xfs_mount *mp = log->l_mp;
1465 /* dgc: errors ignored - not fatal and nowhere to report them */
1466 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1468 * Dump a transaction into the log that contains no real change.
1469 * This is needed to stamp the current tail LSN into the log
1470 * during the covering operation.
1472 * We cannot use an inode here for this - that will push dirty
1473 * state back up into the VFS and then periodic inode flushing
1474 * will prevent log covering from making progress. Hence we
1475 * synchronously log the superblock instead to ensure the
1476 * superblock is immediately unpinned and can be written back.
1478 xlog_clear_incompat(log);
1479 xfs_sync_sb(mp, true);
1481 xfs_log_force(mp, 0);
1483 /* start pushing all the metadata that is currently dirty */
1484 xfs_ail_push_all(mp->m_ail);
1486 /* queue us up again */
1487 xfs_log_work_queue(mp);
1491 * This routine initializes some of the log structure for a given mount point.
1492 * Its primary purpose is to fill in enough, so recovery can occur. However,
1493 * some other stuff may be filled in too.
1495 STATIC struct xlog *
1497 struct xfs_mount *mp,
1498 struct xfs_buftarg *log_target,
1499 xfs_daddr_t blk_offset,
1503 xlog_rec_header_t *head;
1504 xlog_in_core_t **iclogp;
1505 xlog_in_core_t *iclog, *prev_iclog=NULL;
1507 int error = -ENOMEM;
1510 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1512 xfs_warn(mp, "Log allocation failed: No memory!");
1517 log->l_targ = log_target;
1518 log->l_logsize = BBTOB(num_bblks);
1519 log->l_logBBstart = blk_offset;
1520 log->l_logBBsize = num_bblks;
1521 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1522 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1523 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1525 log->l_prev_block = -1;
1526 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1527 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1528 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1529 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1531 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1532 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1534 log->l_iclog_roundoff = BBSIZE;
1536 xlog_grant_head_init(&log->l_reserve_head);
1537 xlog_grant_head_init(&log->l_write_head);
1539 error = -EFSCORRUPTED;
1540 if (xfs_has_sector(mp)) {
1541 log2_size = mp->m_sb.sb_logsectlog;
1542 if (log2_size < BBSHIFT) {
1543 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1544 log2_size, BBSHIFT);
1548 log2_size -= BBSHIFT;
1549 if (log2_size > mp->m_sectbb_log) {
1550 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1551 log2_size, mp->m_sectbb_log);
1555 /* for larger sector sizes, must have v2 or external log */
1556 if (log2_size && log->l_logBBstart > 0 &&
1557 !xfs_has_logv2(mp)) {
1559 "log sector size (0x%x) invalid for configuration.",
1564 log->l_sectBBsize = 1 << log2_size;
1566 init_rwsem(&log->l_incompat_users);
1568 xlog_get_iclog_buffer_size(mp, log);
1570 spin_lock_init(&log->l_icloglock);
1571 init_waitqueue_head(&log->l_flush_wait);
1573 iclogp = &log->l_iclog;
1575 * The amount of memory to allocate for the iclog structure is
1576 * rather funky due to the way the structure is defined. It is
1577 * done this way so that we can use different sizes for machines
1578 * with different amounts of memory. See the definition of
1579 * xlog_in_core_t in xfs_log_priv.h for details.
1581 ASSERT(log->l_iclog_size >= 4096);
1582 for (i = 0; i < log->l_iclog_bufs; i++) {
1583 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1584 sizeof(struct bio_vec);
1586 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1588 goto out_free_iclog;
1591 iclog->ic_prev = prev_iclog;
1594 iclog->ic_data = kvzalloc(log->l_iclog_size,
1595 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1596 if (!iclog->ic_data)
1597 goto out_free_iclog;
1599 log->l_iclog_bak[i] = &iclog->ic_header;
1601 head = &iclog->ic_header;
1602 memset(head, 0, sizeof(xlog_rec_header_t));
1603 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1604 head->h_version = cpu_to_be32(
1605 xfs_has_logv2(log->l_mp) ? 2 : 1);
1606 head->h_size = cpu_to_be32(log->l_iclog_size);
1608 head->h_fmt = cpu_to_be32(XLOG_FMT);
1609 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1611 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1612 iclog->ic_state = XLOG_STATE_ACTIVE;
1613 iclog->ic_log = log;
1614 atomic_set(&iclog->ic_refcnt, 0);
1615 INIT_LIST_HEAD(&iclog->ic_callbacks);
1616 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1618 init_waitqueue_head(&iclog->ic_force_wait);
1619 init_waitqueue_head(&iclog->ic_write_wait);
1620 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1621 sema_init(&iclog->ic_sema, 1);
1623 iclogp = &iclog->ic_next;
1625 *iclogp = log->l_iclog; /* complete ring */
1626 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1628 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1629 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1631 0, mp->m_super->s_id);
1632 if (!log->l_ioend_workqueue)
1633 goto out_free_iclog;
1635 error = xlog_cil_init(log);
1637 goto out_destroy_workqueue;
1640 out_destroy_workqueue:
1641 destroy_workqueue(log->l_ioend_workqueue);
1643 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1644 prev_iclog = iclog->ic_next;
1645 kmem_free(iclog->ic_data);
1647 if (prev_iclog == log->l_iclog)
1653 return ERR_PTR(error);
1654 } /* xlog_alloc_log */
1657 * Compute the LSN that we'd need to push the log tail towards in order to have
1658 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1659 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1660 * log free space already meets all three thresholds, this function returns
1664 xlog_grant_push_threshold(
1668 xfs_lsn_t threshold_lsn = 0;
1669 xfs_lsn_t last_sync_lsn;
1672 int threshold_block;
1673 int threshold_cycle;
1676 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1678 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1679 free_blocks = BTOBBT(free_bytes);
1682 * Set the threshold for the minimum number of free blocks in the
1683 * log to the maximum of what the caller needs, one quarter of the
1684 * log, and 256 blocks.
1686 free_threshold = BTOBB(need_bytes);
1687 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1688 free_threshold = max(free_threshold, 256);
1689 if (free_blocks >= free_threshold)
1690 return NULLCOMMITLSN;
1692 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1694 threshold_block += free_threshold;
1695 if (threshold_block >= log->l_logBBsize) {
1696 threshold_block -= log->l_logBBsize;
1697 threshold_cycle += 1;
1699 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1702 * Don't pass in an lsn greater than the lsn of the last
1703 * log record known to be on disk. Use a snapshot of the last sync lsn
1704 * so that it doesn't change between the compare and the set.
1706 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1707 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1708 threshold_lsn = last_sync_lsn;
1710 return threshold_lsn;
1714 * Push the tail of the log if we need to do so to maintain the free log space
1715 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1716 * policy which pushes on an lsn which is further along in the log once we
1717 * reach the high water mark. In this manner, we would be creating a low water
1721 xlog_grant_push_ail(
1725 xfs_lsn_t threshold_lsn;
1727 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1728 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1732 * Get the transaction layer to kick the dirty buffers out to
1733 * disk asynchronously. No point in trying to do this if
1734 * the filesystem is shutting down.
1736 xfs_ail_push(log->l_ailp, threshold_lsn);
1740 * Stamp cycle number in every block
1745 struct xlog_in_core *iclog,
1749 int size = iclog->ic_offset + roundoff;
1753 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1755 dp = iclog->ic_datap;
1756 for (i = 0; i < BTOBB(size); i++) {
1757 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1759 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1760 *(__be32 *)dp = cycle_lsn;
1764 if (xfs_has_logv2(log->l_mp)) {
1765 xlog_in_core_2_t *xhdr = iclog->ic_data;
1767 for ( ; i < BTOBB(size); i++) {
1768 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1769 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1770 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1771 *(__be32 *)dp = cycle_lsn;
1775 for (i = 1; i < log->l_iclog_heads; i++)
1776 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1781 * Calculate the checksum for a log buffer.
1783 * This is a little more complicated than it should be because the various
1784 * headers and the actual data are non-contiguous.
1789 struct xlog_rec_header *rhead,
1795 /* first generate the crc for the record header ... */
1796 crc = xfs_start_cksum_update((char *)rhead,
1797 sizeof(struct xlog_rec_header),
1798 offsetof(struct xlog_rec_header, h_crc));
1800 /* ... then for additional cycle data for v2 logs ... */
1801 if (xfs_has_logv2(log->l_mp)) {
1802 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1806 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1808 for (i = 1; i < xheads; i++) {
1809 crc = crc32c(crc, &xhdr[i].hic_xheader,
1810 sizeof(struct xlog_rec_ext_header));
1814 /* ... and finally for the payload */
1815 crc = crc32c(crc, dp, size);
1817 return xfs_end_cksum(crc);
1824 struct xlog_in_core *iclog = bio->bi_private;
1826 queue_work(iclog->ic_log->l_ioend_workqueue,
1827 &iclog->ic_end_io_work);
1831 xlog_map_iclog_data(
1837 struct page *page = kmem_to_page(data);
1838 unsigned int off = offset_in_page(data);
1839 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1841 if (bio_add_page(bio, page, len, off) != len)
1854 struct xlog_in_core *iclog,
1858 ASSERT(bno < log->l_logBBsize);
1859 trace_xlog_iclog_write(iclog, _RET_IP_);
1862 * We lock the iclogbufs here so that we can serialise against I/O
1863 * completion during unmount. We might be processing a shutdown
1864 * triggered during unmount, and that can occur asynchronously to the
1865 * unmount thread, and hence we need to ensure that completes before
1866 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1867 * across the log IO to archieve that.
1869 down(&iclog->ic_sema);
1870 if (xlog_is_shutdown(log)) {
1872 * It would seem logical to return EIO here, but we rely on
1873 * the log state machine to propagate I/O errors instead of
1874 * doing it here. We kick of the state machine and unlock
1875 * the buffer manually, the code needs to be kept in sync
1876 * with the I/O completion path.
1878 xlog_state_done_syncing(iclog);
1879 up(&iclog->ic_sema);
1883 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1884 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1885 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1886 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1887 iclog->ic_bio.bi_private = iclog;
1890 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1891 * IOs coming immediately after this one. This prevents the block layer
1892 * writeback throttle from throttling log writes behind background
1893 * metadata writeback and causing priority inversions.
1895 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE;
1896 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1897 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1899 * For external log devices, we also need to flush the data
1900 * device cache first to ensure all metadata writeback covered
1901 * by the LSN in this iclog is on stable storage. This is slow,
1902 * but it *must* complete before we issue the external log IO.
1904 if (log->l_targ != log->l_mp->m_ddev_targp)
1905 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
1907 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1908 iclog->ic_bio.bi_opf |= REQ_FUA;
1910 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1912 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1913 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1916 if (is_vmalloc_addr(iclog->ic_data))
1917 flush_kernel_vmap_range(iclog->ic_data, count);
1920 * If this log buffer would straddle the end of the log we will have
1921 * to split it up into two bios, so that we can continue at the start.
1923 if (bno + BTOBB(count) > log->l_logBBsize) {
1926 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1927 GFP_NOIO, &fs_bio_set);
1928 bio_chain(split, &iclog->ic_bio);
1931 /* restart at logical offset zero for the remainder */
1932 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1935 submit_bio(&iclog->ic_bio);
1939 * We need to bump cycle number for the part of the iclog that is
1940 * written to the start of the log. Watch out for the header magic
1941 * number case, though.
1950 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1953 for (i = split_offset; i < count; i += BBSIZE) {
1954 uint32_t cycle = get_unaligned_be32(data + i);
1956 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1958 put_unaligned_be32(cycle, data + i);
1963 xlog_calc_iclog_size(
1965 struct xlog_in_core *iclog,
1968 uint32_t count_init, count;
1970 /* Add for LR header */
1971 count_init = log->l_iclog_hsize + iclog->ic_offset;
1972 count = roundup(count_init, log->l_iclog_roundoff);
1974 *roundoff = count - count_init;
1976 ASSERT(count >= count_init);
1977 ASSERT(*roundoff < log->l_iclog_roundoff);
1982 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1983 * fashion. Previously, we should have moved the current iclog
1984 * ptr in the log to point to the next available iclog. This allows further
1985 * write to continue while this code syncs out an iclog ready to go.
1986 * Before an in-core log can be written out, the data section must be scanned
1987 * to save away the 1st word of each BBSIZE block into the header. We replace
1988 * it with the current cycle count. Each BBSIZE block is tagged with the
1989 * cycle count because there in an implicit assumption that drives will
1990 * guarantee that entire 512 byte blocks get written at once. In other words,
1991 * we can't have part of a 512 byte block written and part not written. By
1992 * tagging each block, we will know which blocks are valid when recovering
1993 * after an unclean shutdown.
1995 * This routine is single threaded on the iclog. No other thread can be in
1996 * this routine with the same iclog. Changing contents of iclog can there-
1997 * fore be done without grabbing the state machine lock. Updating the global
1998 * log will require grabbing the lock though.
2000 * The entire log manager uses a logical block numbering scheme. Only
2001 * xlog_write_iclog knows about the fact that the log may not start with
2002 * block zero on a given device.
2007 struct xlog_in_core *iclog)
2009 unsigned int count; /* byte count of bwrite */
2010 unsigned int roundoff; /* roundoff to BB or stripe */
2014 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2015 trace_xlog_iclog_sync(iclog, _RET_IP_);
2017 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2019 /* move grant heads by roundoff in sync */
2020 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2021 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2023 /* put cycle number in every block */
2024 xlog_pack_data(log, iclog, roundoff);
2026 /* real byte length */
2027 size = iclog->ic_offset;
2028 if (xfs_has_logv2(log->l_mp))
2030 iclog->ic_header.h_len = cpu_to_be32(size);
2032 XFS_STATS_INC(log->l_mp, xs_log_writes);
2033 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2035 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2037 /* Do we need to split this write into 2 parts? */
2038 if (bno + BTOBB(count) > log->l_logBBsize)
2039 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2041 /* calculcate the checksum */
2042 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2043 iclog->ic_datap, size);
2045 * Intentionally corrupt the log record CRC based on the error injection
2046 * frequency, if defined. This facilitates testing log recovery in the
2047 * event of torn writes. Hence, set the IOABORT state to abort the log
2048 * write on I/O completion and shutdown the fs. The subsequent mount
2049 * detects the bad CRC and attempts to recover.
2052 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2053 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2054 iclog->ic_fail_crc = true;
2056 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2057 be64_to_cpu(iclog->ic_header.h_lsn));
2060 xlog_verify_iclog(log, iclog, count);
2061 xlog_write_iclog(log, iclog, bno, count);
2065 * Deallocate a log structure
2071 xlog_in_core_t *iclog, *next_iclog;
2074 xlog_cil_destroy(log);
2077 * Cycle all the iclogbuf locks to make sure all log IO completion
2078 * is done before we tear down these buffers.
2080 iclog = log->l_iclog;
2081 for (i = 0; i < log->l_iclog_bufs; i++) {
2082 down(&iclog->ic_sema);
2083 up(&iclog->ic_sema);
2084 iclog = iclog->ic_next;
2087 iclog = log->l_iclog;
2088 for (i = 0; i < log->l_iclog_bufs; i++) {
2089 next_iclog = iclog->ic_next;
2090 kmem_free(iclog->ic_data);
2095 log->l_mp->m_log = NULL;
2096 destroy_workqueue(log->l_ioend_workqueue);
2101 * Update counters atomically now that memcpy is done.
2104 xlog_state_finish_copy(
2106 struct xlog_in_core *iclog,
2110 lockdep_assert_held(&log->l_icloglock);
2112 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2113 iclog->ic_offset += copy_bytes;
2117 * print out info relating to regions written which consume
2122 struct xfs_mount *mp,
2123 struct xlog_ticket *ticket)
2126 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2128 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2129 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2130 static char *res_type_str[] = {
2131 REG_TYPE_STR(BFORMAT, "bformat"),
2132 REG_TYPE_STR(BCHUNK, "bchunk"),
2133 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2134 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2135 REG_TYPE_STR(IFORMAT, "iformat"),
2136 REG_TYPE_STR(ICORE, "icore"),
2137 REG_TYPE_STR(IEXT, "iext"),
2138 REG_TYPE_STR(IBROOT, "ibroot"),
2139 REG_TYPE_STR(ILOCAL, "ilocal"),
2140 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2141 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2142 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2143 REG_TYPE_STR(QFORMAT, "qformat"),
2144 REG_TYPE_STR(DQUOT, "dquot"),
2145 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2146 REG_TYPE_STR(LRHEADER, "LR header"),
2147 REG_TYPE_STR(UNMOUNT, "unmount"),
2148 REG_TYPE_STR(COMMIT, "commit"),
2149 REG_TYPE_STR(TRANSHDR, "trans header"),
2150 REG_TYPE_STR(ICREATE, "inode create"),
2151 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2152 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2153 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2154 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2155 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2156 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2158 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2161 xfs_warn(mp, "ticket reservation summary:");
2162 xfs_warn(mp, " unit res = %d bytes",
2163 ticket->t_unit_res);
2164 xfs_warn(mp, " current res = %d bytes",
2165 ticket->t_curr_res);
2166 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2167 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2168 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2169 ticket->t_res_num_ophdrs, ophdr_spc);
2170 xfs_warn(mp, " ophdr + reg = %u bytes",
2171 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2172 xfs_warn(mp, " num regions = %u",
2175 for (i = 0; i < ticket->t_res_num; i++) {
2176 uint r_type = ticket->t_res_arr[i].r_type;
2177 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2178 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2179 "bad-rtype" : res_type_str[r_type]),
2180 ticket->t_res_arr[i].r_len);
2185 * Print a summary of the transaction.
2189 struct xfs_trans *tp)
2191 struct xfs_mount *mp = tp->t_mountp;
2192 struct xfs_log_item *lip;
2194 /* dump core transaction and ticket info */
2195 xfs_warn(mp, "transaction summary:");
2196 xfs_warn(mp, " log res = %d", tp->t_log_res);
2197 xfs_warn(mp, " log count = %d", tp->t_log_count);
2198 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2200 xlog_print_tic_res(mp, tp->t_ticket);
2202 /* dump each log item */
2203 list_for_each_entry(lip, &tp->t_items, li_trans) {
2204 struct xfs_log_vec *lv = lip->li_lv;
2205 struct xfs_log_iovec *vec;
2208 xfs_warn(mp, "log item: ");
2209 xfs_warn(mp, " type = 0x%x", lip->li_type);
2210 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2213 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2214 xfs_warn(mp, " size = %d", lv->lv_size);
2215 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2216 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2218 /* dump each iovec for the log item */
2219 vec = lv->lv_iovecp;
2220 for (i = 0; i < lv->lv_niovecs; i++) {
2221 int dumplen = min(vec->i_len, 32);
2223 xfs_warn(mp, " iovec[%d]", i);
2224 xfs_warn(mp, " type = 0x%x", vec->i_type);
2225 xfs_warn(mp, " len = %d", vec->i_len);
2226 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2227 xfs_hex_dump(vec->i_addr, dumplen);
2235 * Calculate the potential space needed by the log vector. We may need a start
2236 * record, and each region gets its own struct xlog_op_header and may need to be
2237 * double word aligned.
2240 xlog_write_calc_vec_length(
2241 struct xlog_ticket *ticket,
2242 struct xfs_log_vec *log_vector,
2245 struct xfs_log_vec *lv;
2250 if (optype & XLOG_START_TRANS)
2253 for (lv = log_vector; lv; lv = lv->lv_next) {
2254 /* we don't write ordered log vectors */
2255 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2258 headers += lv->lv_niovecs;
2260 for (i = 0; i < lv->lv_niovecs; i++) {
2261 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2264 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2268 ticket->t_res_num_ophdrs += headers;
2269 len += headers * sizeof(struct xlog_op_header);
2275 xlog_write_start_rec(
2276 struct xlog_op_header *ophdr,
2277 struct xlog_ticket *ticket)
2279 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2280 ophdr->oh_clientid = ticket->t_clientid;
2282 ophdr->oh_flags = XLOG_START_TRANS;
2286 static xlog_op_header_t *
2287 xlog_write_setup_ophdr(
2289 struct xlog_op_header *ophdr,
2290 struct xlog_ticket *ticket,
2293 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2294 ophdr->oh_clientid = ticket->t_clientid;
2297 /* are we copying a commit or unmount record? */
2298 ophdr->oh_flags = flags;
2301 * We've seen logs corrupted with bad transaction client ids. This
2302 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2303 * and shut down the filesystem.
2305 switch (ophdr->oh_clientid) {
2306 case XFS_TRANSACTION:
2312 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2313 ophdr->oh_clientid, ticket);
2321 * Set up the parameters of the region copy into the log. This has
2322 * to handle region write split across multiple log buffers - this
2323 * state is kept external to this function so that this code can
2324 * be written in an obvious, self documenting manner.
2327 xlog_write_setup_copy(
2328 struct xlog_ticket *ticket,
2329 struct xlog_op_header *ophdr,
2330 int space_available,
2334 int *last_was_partial_copy,
2335 int *bytes_consumed)
2339 still_to_copy = space_required - *bytes_consumed;
2340 *copy_off = *bytes_consumed;
2342 if (still_to_copy <= space_available) {
2343 /* write of region completes here */
2344 *copy_len = still_to_copy;
2345 ophdr->oh_len = cpu_to_be32(*copy_len);
2346 if (*last_was_partial_copy)
2347 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2348 *last_was_partial_copy = 0;
2349 *bytes_consumed = 0;
2353 /* partial write of region, needs extra log op header reservation */
2354 *copy_len = space_available;
2355 ophdr->oh_len = cpu_to_be32(*copy_len);
2356 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2357 if (*last_was_partial_copy)
2358 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2359 *bytes_consumed += *copy_len;
2360 (*last_was_partial_copy)++;
2362 /* account for new log op header */
2363 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2364 ticket->t_res_num_ophdrs++;
2366 return sizeof(struct xlog_op_header);
2370 xlog_write_copy_finish(
2372 struct xlog_in_core *iclog,
2377 int *partial_copy_len,
2382 if (*partial_copy) {
2384 * This iclog has already been marked WANT_SYNC by
2385 * xlog_state_get_iclog_space.
2387 spin_lock(&log->l_icloglock);
2388 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2395 *partial_copy_len = 0;
2397 if (iclog->ic_size - log_offset > sizeof(xlog_op_header_t))
2400 /* no more space in this iclog - push it. */
2401 spin_lock(&log->l_icloglock);
2402 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2406 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2407 xlog_state_switch_iclogs(log, iclog, 0);
2409 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2410 xlog_is_shutdown(log));
2412 error = xlog_state_release_iclog(log, iclog, 0);
2413 spin_unlock(&log->l_icloglock);
2418 * Write some region out to in-core log
2420 * This will be called when writing externally provided regions or when
2421 * writing out a commit record for a given transaction.
2423 * General algorithm:
2424 * 1. Find total length of this write. This may include adding to the
2425 * lengths passed in.
2426 * 2. Check whether we violate the tickets reservation.
2427 * 3. While writing to this iclog
2428 * A. Reserve as much space in this iclog as can get
2429 * B. If this is first write, save away start lsn
2430 * C. While writing this region:
2431 * 1. If first write of transaction, write start record
2432 * 2. Write log operation header (header per region)
2433 * 3. Find out if we can fit entire region into this iclog
2434 * 4. Potentially, verify destination memcpy ptr
2435 * 5. Memcpy (partial) region
2436 * 6. If partial copy, release iclog; otherwise, continue
2437 * copying more regions into current iclog
2438 * 4. Mark want sync bit (in simulation mode)
2439 * 5. Release iclog for potential flush to on-disk log.
2442 * 1. Panic if reservation is overrun. This should never happen since
2443 * reservation amounts are generated internal to the filesystem.
2445 * 1. Tickets are single threaded data structures.
2446 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2447 * syncing routine. When a single log_write region needs to span
2448 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2449 * on all log operation writes which don't contain the end of the
2450 * region. The XLOG_END_TRANS bit is used for the in-core log
2451 * operation which contains the end of the continued log_write region.
2452 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2453 * we don't really know exactly how much space will be used. As a result,
2454 * we don't update ic_offset until the end when we know exactly how many
2455 * bytes have been written out.
2460 struct xfs_cil_ctx *ctx,
2461 struct xfs_log_vec *log_vector,
2462 struct xlog_ticket *ticket,
2465 struct xlog_in_core *iclog = NULL;
2466 struct xfs_log_vec *lv = log_vector;
2467 struct xfs_log_iovec *vecp = lv->lv_iovecp;
2470 int partial_copy = 0;
2471 int partial_copy_len = 0;
2478 * If this is a commit or unmount transaction, we don't need a start
2479 * record to be written. We do, however, have to account for the
2480 * commit or unmount header that gets written. Hence we always have
2481 * to account for an extra xlog_op_header here.
2483 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2484 if (ticket->t_curr_res < 0) {
2485 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2486 "ctx ticket reservation ran out. Need to up reservation");
2487 xlog_print_tic_res(log->l_mp, ticket);
2488 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2491 len = xlog_write_calc_vec_length(ticket, log_vector, optype);
2492 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2496 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2497 &contwr, &log_offset);
2501 ASSERT(log_offset <= iclog->ic_size - 1);
2502 ptr = iclog->ic_datap + log_offset;
2505 * If we have a context pointer, pass it the first iclog we are
2506 * writing to so it can record state needed for iclog write
2510 xlog_cil_set_ctx_write_state(ctx, iclog);
2515 * This loop writes out as many regions as can fit in the amount
2516 * of space which was allocated by xlog_state_get_iclog_space().
2518 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2519 struct xfs_log_iovec *reg;
2520 struct xlog_op_header *ophdr;
2523 bool ordered = false;
2524 bool wrote_start_rec = false;
2526 /* ordered log vectors have no regions to write */
2527 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2528 ASSERT(lv->lv_niovecs == 0);
2534 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2535 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2538 * Before we start formatting log vectors, we need to
2539 * write a start record. Only do this for the first
2540 * iclog we write to.
2542 if (optype & XLOG_START_TRANS) {
2543 xlog_write_start_rec(ptr, ticket);
2544 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2545 sizeof(struct xlog_op_header));
2546 optype &= ~XLOG_START_TRANS;
2547 wrote_start_rec = true;
2550 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, optype);
2554 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2555 sizeof(struct xlog_op_header));
2557 len += xlog_write_setup_copy(ticket, ophdr,
2558 iclog->ic_size-log_offset,
2560 ©_off, ©_len,
2563 xlog_verify_dest_ptr(log, ptr);
2568 * Unmount records just log an opheader, so can have
2569 * empty payloads with no data region to copy. Hence we
2570 * only copy the payload if the vector says it has data
2573 ASSERT(copy_len >= 0);
2575 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2576 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2579 copy_len += sizeof(struct xlog_op_header);
2581 if (wrote_start_rec) {
2582 copy_len += sizeof(struct xlog_op_header);
2585 data_cnt += contwr ? copy_len : 0;
2587 error = xlog_write_copy_finish(log, iclog, optype,
2588 &record_cnt, &data_cnt,
2596 * if we had a partial copy, we need to get more iclog
2597 * space but we don't want to increment the region
2598 * index because there is still more is this region to
2601 * If we completed writing this region, and we flushed
2602 * the iclog (indicated by resetting of the record
2603 * count), then we also need to get more log space. If
2604 * this was the last record, though, we are done and
2610 if (++index == lv->lv_niovecs) {
2615 vecp = lv->lv_iovecp;
2617 if (record_cnt == 0 && !ordered) {
2627 spin_lock(&log->l_icloglock);
2628 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2629 error = xlog_state_release_iclog(log, iclog, 0);
2630 spin_unlock(&log->l_icloglock);
2636 xlog_state_activate_iclog(
2637 struct xlog_in_core *iclog,
2638 int *iclogs_changed)
2640 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2641 trace_xlog_iclog_activate(iclog, _RET_IP_);
2644 * If the number of ops in this iclog indicate it just contains the
2645 * dummy transaction, we can change state into IDLE (the second time
2646 * around). Otherwise we should change the state into NEED a dummy.
2647 * We don't need to cover the dummy.
2649 if (*iclogs_changed == 0 &&
2650 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2651 *iclogs_changed = 1;
2654 * We have two dirty iclogs so start over. This could also be
2655 * num of ops indicating this is not the dummy going out.
2657 *iclogs_changed = 2;
2660 iclog->ic_state = XLOG_STATE_ACTIVE;
2661 iclog->ic_offset = 0;
2662 iclog->ic_header.h_num_logops = 0;
2663 memset(iclog->ic_header.h_cycle_data, 0,
2664 sizeof(iclog->ic_header.h_cycle_data));
2665 iclog->ic_header.h_lsn = 0;
2666 iclog->ic_header.h_tail_lsn = 0;
2670 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2671 * ACTIVE after iclog I/O has completed.
2674 xlog_state_activate_iclogs(
2676 int *iclogs_changed)
2678 struct xlog_in_core *iclog = log->l_iclog;
2681 if (iclog->ic_state == XLOG_STATE_DIRTY)
2682 xlog_state_activate_iclog(iclog, iclogs_changed);
2684 * The ordering of marking iclogs ACTIVE must be maintained, so
2685 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2687 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2689 } while ((iclog = iclog->ic_next) != log->l_iclog);
2698 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2699 * wrote the first covering record (DONE). We go to IDLE if we just
2700 * wrote the second covering record (DONE2) and remain in IDLE until a
2701 * non-covering write occurs.
2703 switch (prev_state) {
2704 case XLOG_STATE_COVER_IDLE:
2705 if (iclogs_changed == 1)
2706 return XLOG_STATE_COVER_IDLE;
2708 case XLOG_STATE_COVER_NEED:
2709 case XLOG_STATE_COVER_NEED2:
2711 case XLOG_STATE_COVER_DONE:
2712 if (iclogs_changed == 1)
2713 return XLOG_STATE_COVER_NEED2;
2715 case XLOG_STATE_COVER_DONE2:
2716 if (iclogs_changed == 1)
2717 return XLOG_STATE_COVER_IDLE;
2723 return XLOG_STATE_COVER_NEED;
2727 xlog_state_clean_iclog(
2729 struct xlog_in_core *dirty_iclog)
2731 int iclogs_changed = 0;
2733 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2735 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2737 xlog_state_activate_iclogs(log, &iclogs_changed);
2738 wake_up_all(&dirty_iclog->ic_force_wait);
2740 if (iclogs_changed) {
2741 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2747 xlog_get_lowest_lsn(
2750 struct xlog_in_core *iclog = log->l_iclog;
2751 xfs_lsn_t lowest_lsn = 0, lsn;
2754 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2755 iclog->ic_state == XLOG_STATE_DIRTY)
2758 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2759 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2761 } while ((iclog = iclog->ic_next) != log->l_iclog);
2767 * Completion of a iclog IO does not imply that a transaction has completed, as
2768 * transactions can be large enough to span many iclogs. We cannot change the
2769 * tail of the log half way through a transaction as this may be the only
2770 * transaction in the log and moving the tail to point to the middle of it
2771 * will prevent recovery from finding the start of the transaction. Hence we
2772 * should only update the last_sync_lsn if this iclog contains transaction
2773 * completion callbacks on it.
2775 * We have to do this before we drop the icloglock to ensure we are the only one
2776 * that can update it.
2778 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2779 * the reservation grant head pushing. This is due to the fact that the push
2780 * target is bound by the current last_sync_lsn value. Hence if we have a large
2781 * amount of log space bound up in this committing transaction then the
2782 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2783 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2784 * should push the AIL to ensure the push target (and hence the grant head) is
2785 * no longer bound by the old log head location and can move forwards and make
2789 xlog_state_set_callback(
2791 struct xlog_in_core *iclog,
2792 xfs_lsn_t header_lsn)
2794 trace_xlog_iclog_callback(iclog, _RET_IP_);
2795 iclog->ic_state = XLOG_STATE_CALLBACK;
2797 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2800 if (list_empty_careful(&iclog->ic_callbacks))
2803 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2804 xlog_grant_push_ail(log, 0);
2808 * Return true if we need to stop processing, false to continue to the next
2809 * iclog. The caller will need to run callbacks if the iclog is returned in the
2810 * XLOG_STATE_CALLBACK state.
2813 xlog_state_iodone_process_iclog(
2815 struct xlog_in_core *iclog)
2817 xfs_lsn_t lowest_lsn;
2818 xfs_lsn_t header_lsn;
2820 switch (iclog->ic_state) {
2821 case XLOG_STATE_ACTIVE:
2822 case XLOG_STATE_DIRTY:
2824 * Skip all iclogs in the ACTIVE & DIRTY states:
2827 case XLOG_STATE_DONE_SYNC:
2829 * Now that we have an iclog that is in the DONE_SYNC state, do
2830 * one more check here to see if we have chased our tail around.
2831 * If this is not the lowest lsn iclog, then we will leave it
2832 * for another completion to process.
2834 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2835 lowest_lsn = xlog_get_lowest_lsn(log);
2836 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2838 xlog_state_set_callback(log, iclog, header_lsn);
2842 * Can only perform callbacks in order. Since this iclog is not
2843 * in the DONE_SYNC state, we skip the rest and just try to
2851 * Loop over all the iclogs, running attached callbacks on them. Return true if
2852 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2853 * to handle transient shutdown state here at all because
2854 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2855 * cleanup of the callbacks.
2858 xlog_state_do_iclog_callbacks(
2860 __releases(&log->l_icloglock)
2861 __acquires(&log->l_icloglock)
2863 struct xlog_in_core *first_iclog = log->l_iclog;
2864 struct xlog_in_core *iclog = first_iclog;
2865 bool ran_callback = false;
2870 if (xlog_state_iodone_process_iclog(log, iclog))
2872 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2873 iclog = iclog->ic_next;
2876 list_splice_init(&iclog->ic_callbacks, &cb_list);
2877 spin_unlock(&log->l_icloglock);
2879 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2880 xlog_cil_process_committed(&cb_list);
2881 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2882 ran_callback = true;
2884 spin_lock(&log->l_icloglock);
2885 xlog_state_clean_iclog(log, iclog);
2886 iclog = iclog->ic_next;
2887 } while (iclog != first_iclog);
2889 return ran_callback;
2894 * Loop running iclog completion callbacks until there are no more iclogs in a
2895 * state that can run callbacks.
2898 xlog_state_do_callback(
2904 spin_lock(&log->l_icloglock);
2905 while (xlog_state_do_iclog_callbacks(log)) {
2906 if (xlog_is_shutdown(log))
2909 if (++repeats > 5000) {
2910 flushcnt += repeats;
2913 "%s: possible infinite loop (%d iterations)",
2914 __func__, flushcnt);
2918 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2919 wake_up_all(&log->l_flush_wait);
2921 spin_unlock(&log->l_icloglock);
2926 * Finish transitioning this iclog to the dirty state.
2928 * Callbacks could take time, so they are done outside the scope of the
2929 * global state machine log lock.
2932 xlog_state_done_syncing(
2933 struct xlog_in_core *iclog)
2935 struct xlog *log = iclog->ic_log;
2937 spin_lock(&log->l_icloglock);
2938 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2939 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2942 * If we got an error, either on the first buffer, or in the case of
2943 * split log writes, on the second, we shut down the file system and
2944 * no iclogs should ever be attempted to be written to disk again.
2946 if (!xlog_is_shutdown(log)) {
2947 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2948 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2952 * Someone could be sleeping prior to writing out the next
2953 * iclog buffer, we wake them all, one will get to do the
2954 * I/O, the others get to wait for the result.
2956 wake_up_all(&iclog->ic_write_wait);
2957 spin_unlock(&log->l_icloglock);
2958 xlog_state_do_callback(log);
2962 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2963 * sleep. We wait on the flush queue on the head iclog as that should be
2964 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2965 * we will wait here and all new writes will sleep until a sync completes.
2967 * The in-core logs are used in a circular fashion. They are not used
2968 * out-of-order even when an iclog past the head is free.
2971 * * log_offset where xlog_write() can start writing into the in-core
2973 * * in-core log pointer to which xlog_write() should write.
2974 * * boolean indicating this is a continued write to an in-core log.
2975 * If this is the last write, then the in-core log's offset field
2976 * needs to be incremented, depending on the amount of data which
2980 xlog_state_get_iclog_space(
2983 struct xlog_in_core **iclogp,
2984 struct xlog_ticket *ticket,
2985 int *continued_write,
2989 xlog_rec_header_t *head;
2990 xlog_in_core_t *iclog;
2993 spin_lock(&log->l_icloglock);
2994 if (xlog_is_shutdown(log)) {
2995 spin_unlock(&log->l_icloglock);
2999 iclog = log->l_iclog;
3000 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3001 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3003 /* Wait for log writes to have flushed */
3004 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3008 head = &iclog->ic_header;
3010 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
3011 log_offset = iclog->ic_offset;
3013 trace_xlog_iclog_get_space(iclog, _RET_IP_);
3015 /* On the 1st write to an iclog, figure out lsn. This works
3016 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3017 * committing to. If the offset is set, that's how many blocks
3020 if (log_offset == 0) {
3021 ticket->t_curr_res -= log->l_iclog_hsize;
3022 xlog_tic_add_region(ticket,
3024 XLOG_REG_TYPE_LRHEADER);
3025 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3026 head->h_lsn = cpu_to_be64(
3027 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3028 ASSERT(log->l_curr_block >= 0);
3031 /* If there is enough room to write everything, then do it. Otherwise,
3032 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3033 * bit is on, so this will get flushed out. Don't update ic_offset
3034 * until you know exactly how many bytes get copied. Therefore, wait
3035 * until later to update ic_offset.
3037 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3038 * can fit into remaining data section.
3040 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3043 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3046 * If we are the only one writing to this iclog, sync it to
3047 * disk. We need to do an atomic compare and decrement here to
3048 * avoid racing with concurrent atomic_dec_and_lock() calls in
3049 * xlog_state_release_iclog() when there is more than one
3050 * reference to the iclog.
3052 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3053 error = xlog_state_release_iclog(log, iclog, 0);
3054 spin_unlock(&log->l_icloglock);
3060 /* Do we have enough room to write the full amount in the remainder
3061 * of this iclog? Or must we continue a write on the next iclog and
3062 * mark this iclog as completely taken? In the case where we switch
3063 * iclogs (to mark it taken), this particular iclog will release/sync
3064 * to disk in xlog_write().
3066 if (len <= iclog->ic_size - iclog->ic_offset) {
3067 *continued_write = 0;
3068 iclog->ic_offset += len;
3070 *continued_write = 1;
3071 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3075 ASSERT(iclog->ic_offset <= iclog->ic_size);
3076 spin_unlock(&log->l_icloglock);
3078 *logoffsetp = log_offset;
3083 * The first cnt-1 times a ticket goes through here we don't need to move the
3084 * grant write head because the permanent reservation has reserved cnt times the
3085 * unit amount. Release part of current permanent unit reservation and reset
3086 * current reservation to be one units worth. Also move grant reservation head
3090 xfs_log_ticket_regrant(
3092 struct xlog_ticket *ticket)
3094 trace_xfs_log_ticket_regrant(log, ticket);
3096 if (ticket->t_cnt > 0)
3099 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3100 ticket->t_curr_res);
3101 xlog_grant_sub_space(log, &log->l_write_head.grant,
3102 ticket->t_curr_res);
3103 ticket->t_curr_res = ticket->t_unit_res;
3104 xlog_tic_reset_res(ticket);
3106 trace_xfs_log_ticket_regrant_sub(log, ticket);
3108 /* just return if we still have some of the pre-reserved space */
3109 if (!ticket->t_cnt) {
3110 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3111 ticket->t_unit_res);
3112 trace_xfs_log_ticket_regrant_exit(log, ticket);
3114 ticket->t_curr_res = ticket->t_unit_res;
3115 xlog_tic_reset_res(ticket);
3118 xfs_log_ticket_put(ticket);
3122 * Give back the space left from a reservation.
3124 * All the information we need to make a correct determination of space left
3125 * is present. For non-permanent reservations, things are quite easy. The
3126 * count should have been decremented to zero. We only need to deal with the
3127 * space remaining in the current reservation part of the ticket. If the
3128 * ticket contains a permanent reservation, there may be left over space which
3129 * needs to be released. A count of N means that N-1 refills of the current
3130 * reservation can be done before we need to ask for more space. The first
3131 * one goes to fill up the first current reservation. Once we run out of
3132 * space, the count will stay at zero and the only space remaining will be
3133 * in the current reservation field.
3136 xfs_log_ticket_ungrant(
3138 struct xlog_ticket *ticket)
3142 trace_xfs_log_ticket_ungrant(log, ticket);
3144 if (ticket->t_cnt > 0)
3147 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3150 * If this is a permanent reservation ticket, we may be able to free
3151 * up more space based on the remaining count.
3153 bytes = ticket->t_curr_res;
3154 if (ticket->t_cnt > 0) {
3155 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3156 bytes += ticket->t_unit_res*ticket->t_cnt;
3159 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3160 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3162 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3164 xfs_log_space_wake(log->l_mp);
3165 xfs_log_ticket_put(ticket);
3169 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3170 * the current iclog pointer to the next iclog in the ring.
3173 xlog_state_switch_iclogs(
3175 struct xlog_in_core *iclog,
3178 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3179 assert_spin_locked(&log->l_icloglock);
3180 trace_xlog_iclog_switch(iclog, _RET_IP_);
3183 eventual_size = iclog->ic_offset;
3184 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3185 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3186 log->l_prev_block = log->l_curr_block;
3187 log->l_prev_cycle = log->l_curr_cycle;
3189 /* roll log?: ic_offset changed later */
3190 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3192 /* Round up to next log-sunit */
3193 if (log->l_iclog_roundoff > BBSIZE) {
3194 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3195 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3198 if (log->l_curr_block >= log->l_logBBsize) {
3200 * Rewind the current block before the cycle is bumped to make
3201 * sure that the combined LSN never transiently moves forward
3202 * when the log wraps to the next cycle. This is to support the
3203 * unlocked sample of these fields from xlog_valid_lsn(). Most
3204 * other cases should acquire l_icloglock.
3206 log->l_curr_block -= log->l_logBBsize;
3207 ASSERT(log->l_curr_block >= 0);
3209 log->l_curr_cycle++;
3210 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3211 log->l_curr_cycle++;
3213 ASSERT(iclog == log->l_iclog);
3214 log->l_iclog = iclog->ic_next;
3218 * Force the iclog to disk and check if the iclog has been completed before
3219 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3220 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3221 * If completion has already occurred, tell the caller so that it can avoid an
3222 * unnecessary wait on the iclog.
3225 xlog_force_and_check_iclog(
3226 struct xlog_in_core *iclog,
3229 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3233 error = xlog_force_iclog(iclog);
3238 * If the iclog has already been completed and reused the header LSN
3239 * will have been rewritten by completion
3241 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3247 * Write out all data in the in-core log as of this exact moment in time.
3249 * Data may be written to the in-core log during this call. However,
3250 * we don't guarantee this data will be written out. A change from past
3251 * implementation means this routine will *not* write out zero length LRs.
3253 * Basically, we try and perform an intelligent scan of the in-core logs.
3254 * If we determine there is no flushable data, we just return. There is no
3255 * flushable data if:
3257 * 1. the current iclog is active and has no data; the previous iclog
3258 * is in the active or dirty state.
3259 * 2. the current iclog is drity, and the previous iclog is in the
3260 * active or dirty state.
3264 * 1. the current iclog is not in the active nor dirty state.
3265 * 2. the current iclog dirty, and the previous iclog is not in the
3266 * active nor dirty state.
3267 * 3. the current iclog is active, and there is another thread writing
3268 * to this particular iclog.
3269 * 4. a) the current iclog is active and has no other writers
3270 * b) when we return from flushing out this iclog, it is still
3271 * not in the active nor dirty state.
3275 struct xfs_mount *mp,
3278 struct xlog *log = mp->m_log;
3279 struct xlog_in_core *iclog;
3281 XFS_STATS_INC(mp, xs_log_force);
3282 trace_xfs_log_force(mp, 0, _RET_IP_);
3284 xlog_cil_force(log);
3286 spin_lock(&log->l_icloglock);
3287 if (xlog_is_shutdown(log))
3290 iclog = log->l_iclog;
3291 trace_xlog_iclog_force(iclog, _RET_IP_);
3293 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3294 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3295 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3297 * If the head is dirty or (active and empty), then we need to
3298 * look at the previous iclog.
3300 * If the previous iclog is active or dirty we are done. There
3301 * is nothing to sync out. Otherwise, we attach ourselves to the
3302 * previous iclog and go to sleep.
3304 iclog = iclog->ic_prev;
3305 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3306 if (atomic_read(&iclog->ic_refcnt) == 0) {
3307 /* We have exclusive access to this iclog. */
3310 if (xlog_force_and_check_iclog(iclog, &completed))
3317 * Someone else is still writing to this iclog, so we
3318 * need to ensure that when they release the iclog it
3319 * gets synced immediately as we may be waiting on it.
3321 xlog_state_switch_iclogs(log, iclog, 0);
3326 * The iclog we are about to wait on may contain the checkpoint pushed
3327 * by the above xlog_cil_force() call, but it may not have been pushed
3328 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3329 * are flushed when this iclog is written.
3331 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3332 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3334 if (flags & XFS_LOG_SYNC)
3335 return xlog_wait_on_iclog(iclog);
3337 spin_unlock(&log->l_icloglock);
3340 spin_unlock(&log->l_icloglock);
3345 * Force the log to a specific LSN.
3347 * If an iclog with that lsn can be found:
3348 * If it is in the DIRTY state, just return.
3349 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3350 * state and go to sleep or return.
3351 * If it is in any other state, go to sleep or return.
3353 * Synchronous forces are implemented with a wait queue. All callers trying
3354 * to force a given lsn to disk must wait on the queue attached to the
3355 * specific in-core log. When given in-core log finally completes its write
3356 * to disk, that thread will wake up all threads waiting on the queue.
3366 struct xlog_in_core *iclog;
3369 spin_lock(&log->l_icloglock);
3370 if (xlog_is_shutdown(log))
3373 iclog = log->l_iclog;
3374 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3375 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3376 iclog = iclog->ic_next;
3377 if (iclog == log->l_iclog)
3381 switch (iclog->ic_state) {
3382 case XLOG_STATE_ACTIVE:
3384 * We sleep here if we haven't already slept (e.g. this is the
3385 * first time we've looked at the correct iclog buf) and the
3386 * buffer before us is going to be sync'ed. The reason for this
3387 * is that if we are doing sync transactions here, by waiting
3388 * for the previous I/O to complete, we can allow a few more
3389 * transactions into this iclog before we close it down.
3391 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3392 * refcnt so we can release the log (which drops the ref count).
3393 * The state switch keeps new transaction commits from using
3394 * this buffer. When the current commits finish writing into
3395 * the buffer, the refcount will drop to zero and the buffer
3398 if (!already_slept &&
3399 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3400 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3401 xlog_wait(&iclog->ic_prev->ic_write_wait,
3405 if (xlog_force_and_check_iclog(iclog, &completed))
3412 case XLOG_STATE_WANT_SYNC:
3414 * This iclog may contain the checkpoint pushed by the
3415 * xlog_cil_force_seq() call, but there are other writers still
3416 * accessing it so it hasn't been pushed to disk yet. Like the
3417 * ACTIVE case above, we need to make sure caches are flushed
3418 * when this iclog is written.
3420 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3424 * The entire checkpoint was written by the CIL force and is on
3425 * its way to disk already. It will be stable when it
3426 * completes, so we don't need to manipulate caches here at all.
3427 * We just need to wait for completion if necessary.
3432 if (flags & XFS_LOG_SYNC)
3433 return xlog_wait_on_iclog(iclog);
3435 spin_unlock(&log->l_icloglock);
3438 spin_unlock(&log->l_icloglock);
3443 * Force the log to a specific checkpoint sequence.
3445 * First force the CIL so that all the required changes have been flushed to the
3446 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3447 * the iclog that needs to be flushed to stable storage. If the caller needs
3448 * a synchronous log force, we will wait on the iclog with the LSN returned by
3449 * xlog_cil_force_seq() to be completed.
3453 struct xfs_mount *mp,
3458 struct xlog *log = mp->m_log;
3463 XFS_STATS_INC(mp, xs_log_force);
3464 trace_xfs_log_force(mp, seq, _RET_IP_);
3466 lsn = xlog_cil_force_seq(log, seq);
3467 if (lsn == NULLCOMMITLSN)
3470 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3471 if (ret == -EAGAIN) {
3472 XFS_STATS_INC(mp, xs_log_force_sleep);
3473 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3479 * Free a used ticket when its refcount falls to zero.
3483 xlog_ticket_t *ticket)
3485 ASSERT(atomic_read(&ticket->t_ref) > 0);
3486 if (atomic_dec_and_test(&ticket->t_ref))
3487 kmem_cache_free(xfs_log_ticket_zone, ticket);
3492 xlog_ticket_t *ticket)
3494 ASSERT(atomic_read(&ticket->t_ref) > 0);
3495 atomic_inc(&ticket->t_ref);
3500 * Figure out the total log space unit (in bytes) that would be
3501 * required for a log ticket.
3512 * Permanent reservations have up to 'cnt'-1 active log operations
3513 * in the log. A unit in this case is the amount of space for one
3514 * of these log operations. Normal reservations have a cnt of 1
3515 * and their unit amount is the total amount of space required.
3517 * The following lines of code account for non-transaction data
3518 * which occupy space in the on-disk log.
3520 * Normal form of a transaction is:
3521 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3522 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3524 * We need to account for all the leadup data and trailer data
3525 * around the transaction data.
3526 * And then we need to account for the worst case in terms of using
3528 * The worst case will happen if:
3529 * - the placement of the transaction happens to be such that the
3530 * roundoff is at its maximum
3531 * - the transaction data is synced before the commit record is synced
3532 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3533 * Therefore the commit record is in its own Log Record.
3534 * This can happen as the commit record is called with its
3535 * own region to xlog_write().
3536 * This then means that in the worst case, roundoff can happen for
3537 * the commit-rec as well.
3538 * The commit-rec is smaller than padding in this scenario and so it is
3539 * not added separately.
3542 /* for trans header */
3543 unit_bytes += sizeof(xlog_op_header_t);
3544 unit_bytes += sizeof(xfs_trans_header_t);
3547 unit_bytes += sizeof(xlog_op_header_t);
3550 * for LR headers - the space for data in an iclog is the size minus
3551 * the space used for the headers. If we use the iclog size, then we
3552 * undercalculate the number of headers required.
3554 * Furthermore - the addition of op headers for split-recs might
3555 * increase the space required enough to require more log and op
3556 * headers, so take that into account too.
3558 * IMPORTANT: This reservation makes the assumption that if this
3559 * transaction is the first in an iclog and hence has the LR headers
3560 * accounted to it, then the remaining space in the iclog is
3561 * exclusively for this transaction. i.e. if the transaction is larger
3562 * than the iclog, it will be the only thing in that iclog.
3563 * Fundamentally, this means we must pass the entire log vector to
3564 * xlog_write to guarantee this.
3566 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3567 num_headers = howmany(unit_bytes, iclog_space);
3569 /* for split-recs - ophdrs added when data split over LRs */
3570 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3572 /* add extra header reservations if we overrun */
3573 while (!num_headers ||
3574 howmany(unit_bytes, iclog_space) > num_headers) {
3575 unit_bytes += sizeof(xlog_op_header_t);
3578 unit_bytes += log->l_iclog_hsize * num_headers;
3580 /* for commit-rec LR header - note: padding will subsume the ophdr */
3581 unit_bytes += log->l_iclog_hsize;
3583 /* roundoff padding for transaction data and one for commit record */
3584 unit_bytes += 2 * log->l_iclog_roundoff;
3590 xfs_log_calc_unit_res(
3591 struct xfs_mount *mp,
3594 return xlog_calc_unit_res(mp->m_log, unit_bytes);
3598 * Allocate and initialise a new log ticket.
3600 struct xlog_ticket *
3608 struct xlog_ticket *tic;
3611 tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3613 unit_res = xlog_calc_unit_res(log, unit_bytes);
3615 atomic_set(&tic->t_ref, 1);
3616 tic->t_task = current;
3617 INIT_LIST_HEAD(&tic->t_queue);
3618 tic->t_unit_res = unit_res;
3619 tic->t_curr_res = unit_res;
3622 tic->t_tid = prandom_u32();
3623 tic->t_clientid = client;
3625 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3627 xlog_tic_reset_res(tic);
3634 * Make sure that the destination ptr is within the valid data region of
3635 * one of the iclogs. This uses backup pointers stored in a different
3636 * part of the log in case we trash the log structure.
3639 xlog_verify_dest_ptr(
3646 for (i = 0; i < log->l_iclog_bufs; i++) {
3647 if (ptr >= log->l_iclog_bak[i] &&
3648 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3653 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3657 * Check to make sure the grant write head didn't just over lap the tail. If
3658 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3659 * the cycles differ by exactly one and check the byte count.
3661 * This check is run unlocked, so can give false positives. Rather than assert
3662 * on failures, use a warn-once flag and a panic tag to allow the admin to
3663 * determine if they want to panic the machine when such an error occurs. For
3664 * debug kernels this will have the same effect as using an assert but, unlinke
3665 * an assert, it can be turned off at runtime.
3668 xlog_verify_grant_tail(
3671 int tail_cycle, tail_blocks;
3674 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3675 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3676 if (tail_cycle != cycle) {
3677 if (cycle - 1 != tail_cycle &&
3678 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3679 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3680 "%s: cycle - 1 != tail_cycle", __func__);
3683 if (space > BBTOB(tail_blocks) &&
3684 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3685 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3686 "%s: space > BBTOB(tail_blocks)", __func__);
3691 /* check if it will fit */
3693 xlog_verify_tail_lsn(
3695 struct xlog_in_core *iclog)
3697 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3700 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3702 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3703 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3704 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3706 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3708 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3709 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3711 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3712 if (blocks < BTOBB(iclog->ic_offset) + 1)
3713 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3718 * Perform a number of checks on the iclog before writing to disk.
3720 * 1. Make sure the iclogs are still circular
3721 * 2. Make sure we have a good magic number
3722 * 3. Make sure we don't have magic numbers in the data
3723 * 4. Check fields of each log operation header for:
3724 * A. Valid client identifier
3725 * B. tid ptr value falls in valid ptr space (user space code)
3726 * C. Length in log record header is correct according to the
3727 * individual operation headers within record.
3728 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3729 * log, check the preceding blocks of the physical log to make sure all
3730 * the cycle numbers agree with the current cycle number.
3735 struct xlog_in_core *iclog,
3738 xlog_op_header_t *ophead;
3739 xlog_in_core_t *icptr;
3740 xlog_in_core_2_t *xhdr;
3741 void *base_ptr, *ptr, *p;
3742 ptrdiff_t field_offset;
3744 int len, i, j, k, op_len;
3747 /* check validity of iclog pointers */
3748 spin_lock(&log->l_icloglock);
3749 icptr = log->l_iclog;
3750 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3753 if (icptr != log->l_iclog)
3754 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3755 spin_unlock(&log->l_icloglock);
3757 /* check log magic numbers */
3758 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3759 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3761 base_ptr = ptr = &iclog->ic_header;
3762 p = &iclog->ic_header;
3763 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3764 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3765 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3770 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3771 base_ptr = ptr = iclog->ic_datap;
3773 xhdr = iclog->ic_data;
3774 for (i = 0; i < len; i++) {
3777 /* clientid is only 1 byte */
3778 p = &ophead->oh_clientid;
3779 field_offset = p - base_ptr;
3780 if (field_offset & 0x1ff) {
3781 clientid = ophead->oh_clientid;
3783 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3784 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3785 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3786 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3787 clientid = xlog_get_client_id(
3788 xhdr[j].hic_xheader.xh_cycle_data[k]);
3790 clientid = xlog_get_client_id(
3791 iclog->ic_header.h_cycle_data[idx]);
3794 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3796 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3797 __func__, clientid, ophead,
3798 (unsigned long)field_offset);
3801 p = &ophead->oh_len;
3802 field_offset = p - base_ptr;
3803 if (field_offset & 0x1ff) {
3804 op_len = be32_to_cpu(ophead->oh_len);
3806 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3807 (uintptr_t)iclog->ic_datap);
3808 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3809 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3810 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3811 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3813 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3816 ptr += sizeof(xlog_op_header_t) + op_len;
3822 * Perform a forced shutdown on the log. This should be called once and once
3823 * only by the high level filesystem shutdown code to shut the log subsystem
3826 * Our main objectives here are to make sure that:
3827 * a. if the shutdown was not due to a log IO error, flush the logs to
3828 * disk. Anything modified after this is ignored.
3829 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3830 * parties to find out. Nothing new gets queued after this is done.
3831 * c. Tasks sleeping on log reservations, pinned objects and
3832 * other resources get woken up.
3834 * Return true if the shutdown cause was a log IO error and we actually shut the
3838 xlog_force_shutdown(
3842 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3845 * If this happens during log recovery then we aren't using the runtime
3846 * log mechanisms yet so there's nothing to shut down.
3848 if (!log || xlog_in_recovery(log))
3851 ASSERT(!xlog_is_shutdown(log));
3854 * Flush all the completed transactions to disk before marking the log
3855 * being shut down. We need to do this first as shutting down the log
3856 * before the force will prevent the log force from flushing the iclogs
3859 * Re-entry due to a log IO error shutdown during the log force is
3860 * prevented by the atomicity of higher level shutdown code.
3863 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3866 * Atomically set the shutdown state. If the shutdown state is already
3867 * set, there someone else is performing the shutdown and so we are done
3868 * here. This should never happen because we should only ever get called
3869 * once by the first shutdown caller.
3871 * Much of the log state machine transitions assume that shutdown state
3872 * cannot change once they hold the log->l_icloglock. Hence we need to
3873 * hold that lock here, even though we use the atomic test_and_set_bit()
3874 * operation to set the shutdown state.
3876 spin_lock(&log->l_icloglock);
3877 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3878 spin_unlock(&log->l_icloglock);
3882 spin_unlock(&log->l_icloglock);
3885 * We don't want anybody waiting for log reservations after this. That
3886 * means we have to wake up everybody queued up on reserveq as well as
3887 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3888 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3889 * action is protected by the grant locks.
3891 xlog_grant_head_wake_all(&log->l_reserve_head);
3892 xlog_grant_head_wake_all(&log->l_write_head);
3895 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3896 * as if the log writes were completed. The abort handling in the log
3897 * item committed callback functions will do this again under lock to
3900 spin_lock(&log->l_cilp->xc_push_lock);
3901 wake_up_all(&log->l_cilp->xc_start_wait);
3902 wake_up_all(&log->l_cilp->xc_commit_wait);
3903 spin_unlock(&log->l_cilp->xc_push_lock);
3904 xlog_state_shutdown_callbacks(log);
3913 xlog_in_core_t *iclog;
3915 iclog = log->l_iclog;
3917 /* endianness does not matter here, zero is zero in
3920 if (iclog->ic_header.h_num_logops)
3922 iclog = iclog->ic_next;
3923 } while (iclog != log->l_iclog);
3928 * Verify that an LSN stamped into a piece of metadata is valid. This is
3929 * intended for use in read verifiers on v5 superblocks.
3933 struct xfs_mount *mp,
3936 struct xlog *log = mp->m_log;
3940 * norecovery mode skips mount-time log processing and unconditionally
3941 * resets the in-core LSN. We can't validate in this mode, but
3942 * modifications are not allowed anyways so just return true.
3944 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3948 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3949 * handled by recovery and thus safe to ignore here.
3951 if (lsn == NULLCOMMITLSN)
3954 valid = xlog_valid_lsn(mp->m_log, lsn);
3956 /* warn the user about what's gone wrong before verifier failure */
3958 spin_lock(&log->l_icloglock);
3960 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3961 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3962 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3963 log->l_curr_cycle, log->l_curr_block);
3964 spin_unlock(&log->l_icloglock);
3971 * Notify the log that we're about to start using a feature that is protected
3972 * by a log incompat feature flag. This will prevent log covering from
3973 * clearing those flags.
3976 xlog_use_incompat_feat(
3979 down_read(&log->l_incompat_users);
3982 /* Notify the log that we've finished using log incompat features. */
3984 xlog_drop_incompat_feat(
3987 up_read(&log->l_incompat_users);