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 struct kmem_cache *xfs_log_ticket_cache;
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,
60 struct xlog_in_core *iclog,
61 struct xlog_ticket *ticket);
64 xlog_verify_grant_tail(
69 struct xlog_in_core *iclog,
74 struct xlog_in_core *iclog);
76 #define xlog_verify_grant_tail(a)
77 #define xlog_verify_iclog(a,b,c)
78 #define xlog_verify_tail_lsn(a,b)
86 xfs_log_cover(struct xfs_mount *);
89 * We need to make sure the buffer pointer returned is naturally aligned for the
90 * biggest basic data type we put into it. We have already accounted for this
91 * padding when sizing the buffer.
93 * However, this padding does not get written into the log, and hence we have to
94 * track the space used by the log vectors separately to prevent log space hangs
95 * due to inaccurate accounting (i.e. a leak) of the used log space through the
98 * We also add space for the xlog_op_header that describes this region in the
99 * log. This prepends the data region we return to the caller to copy their data
100 * into, so do all the static initialisation of the ophdr now. Because the ophdr
101 * is not 8 byte aligned, we have to be careful to ensure that we align the
102 * start of the buffer such that the region we return to the call is 8 byte
103 * aligned and packed against the tail of the ophdr.
107 struct xfs_log_vec *lv,
108 struct xfs_log_iovec **vecp,
111 struct xfs_log_iovec *vec = *vecp;
112 struct xlog_op_header *oph;
117 ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
120 vec = &lv->lv_iovecp[0];
123 len = lv->lv_buf_len + sizeof(struct xlog_op_header);
124 if (!IS_ALIGNED(len, sizeof(uint64_t))) {
125 lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
126 sizeof(struct xlog_op_header);
130 vec->i_addr = lv->lv_buf + lv->lv_buf_len;
133 oph->oh_clientid = XFS_TRANSACTION;
137 buf = vec->i_addr + sizeof(struct xlog_op_header);
138 ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
145 xlog_grant_sub_space(
150 int64_t head_val = atomic64_read(head);
156 xlog_crack_grant_head_val(head_val, &cycle, &space);
160 space += log->l_logsize;
165 new = xlog_assign_grant_head_val(cycle, space);
166 head_val = atomic64_cmpxchg(head, old, new);
167 } while (head_val != old);
171 xlog_grant_add_space(
176 int64_t head_val = atomic64_read(head);
183 xlog_crack_grant_head_val(head_val, &cycle, &space);
185 tmp = log->l_logsize - space;
194 new = xlog_assign_grant_head_val(cycle, space);
195 head_val = atomic64_cmpxchg(head, old, new);
196 } while (head_val != old);
200 xlog_grant_head_init(
201 struct xlog_grant_head *head)
203 xlog_assign_grant_head(&head->grant, 1, 0);
204 INIT_LIST_HEAD(&head->waiters);
205 spin_lock_init(&head->lock);
209 xlog_grant_head_wake_all(
210 struct xlog_grant_head *head)
212 struct xlog_ticket *tic;
214 spin_lock(&head->lock);
215 list_for_each_entry(tic, &head->waiters, t_queue)
216 wake_up_process(tic->t_task);
217 spin_unlock(&head->lock);
221 xlog_ticket_reservation(
223 struct xlog_grant_head *head,
224 struct xlog_ticket *tic)
226 if (head == &log->l_write_head) {
227 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
228 return tic->t_unit_res;
231 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
232 return tic->t_unit_res * tic->t_cnt;
234 return tic->t_unit_res;
238 xlog_grant_head_wake(
240 struct xlog_grant_head *head,
243 struct xlog_ticket *tic;
245 bool woken_task = false;
247 list_for_each_entry(tic, &head->waiters, t_queue) {
250 * There is a chance that the size of the CIL checkpoints in
251 * progress at the last AIL push target calculation resulted in
252 * limiting the target to the log head (l_last_sync_lsn) at the
253 * time. This may not reflect where the log head is now as the
254 * CIL checkpoints may have completed.
256 * Hence when we are woken here, it may be that the head of the
257 * log that has moved rather than the tail. As the tail didn't
258 * move, there still won't be space available for the
259 * reservation we require. However, if the AIL has already
260 * pushed to the target defined by the old log head location, we
261 * will hang here waiting for something else to update the AIL
264 * Therefore, if there isn't space to wake the first waiter on
265 * the grant head, we need to push the AIL again to ensure the
266 * target reflects both the current log tail and log head
267 * position before we wait for the tail to move again.
270 need_bytes = xlog_ticket_reservation(log, head, tic);
271 if (*free_bytes < need_bytes) {
273 xlog_grant_push_ail(log, need_bytes);
277 *free_bytes -= need_bytes;
278 trace_xfs_log_grant_wake_up(log, tic);
279 wake_up_process(tic->t_task);
287 xlog_grant_head_wait(
289 struct xlog_grant_head *head,
290 struct xlog_ticket *tic,
291 int need_bytes) __releases(&head->lock)
292 __acquires(&head->lock)
294 list_add_tail(&tic->t_queue, &head->waiters);
297 if (xlog_is_shutdown(log))
299 xlog_grant_push_ail(log, need_bytes);
301 __set_current_state(TASK_UNINTERRUPTIBLE);
302 spin_unlock(&head->lock);
304 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
306 trace_xfs_log_grant_sleep(log, tic);
308 trace_xfs_log_grant_wake(log, tic);
310 spin_lock(&head->lock);
311 if (xlog_is_shutdown(log))
313 } while (xlog_space_left(log, &head->grant) < need_bytes);
315 list_del_init(&tic->t_queue);
318 list_del_init(&tic->t_queue);
323 * Atomically get the log space required for a log ticket.
325 * Once a ticket gets put onto head->waiters, it will only return after the
326 * needed reservation is satisfied.
328 * This function is structured so that it has a lock free fast path. This is
329 * necessary because every new transaction reservation will come through this
330 * path. Hence any lock will be globally hot if we take it unconditionally on
333 * As tickets are only ever moved on and off head->waiters under head->lock, we
334 * only need to take that lock if we are going to add the ticket to the queue
335 * and sleep. We can avoid taking the lock if the ticket was never added to
336 * head->waiters because the t_queue list head will be empty and we hold the
337 * only reference to it so it can safely be checked unlocked.
340 xlog_grant_head_check(
342 struct xlog_grant_head *head,
343 struct xlog_ticket *tic,
349 ASSERT(!xlog_in_recovery(log));
352 * If there are other waiters on the queue then give them a chance at
353 * logspace before us. Wake up the first waiters, if we do not wake
354 * up all the waiters then go to sleep waiting for more free space,
355 * otherwise try to get some space for this transaction.
357 *need_bytes = xlog_ticket_reservation(log, head, tic);
358 free_bytes = xlog_space_left(log, &head->grant);
359 if (!list_empty_careful(&head->waiters)) {
360 spin_lock(&head->lock);
361 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
362 free_bytes < *need_bytes) {
363 error = xlog_grant_head_wait(log, head, tic,
366 spin_unlock(&head->lock);
367 } else if (free_bytes < *need_bytes) {
368 spin_lock(&head->lock);
369 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
370 spin_unlock(&head->lock);
378 struct xfs_mount *mp)
381 * Do not write to the log on norecovery mounts, if the data or log
382 * devices are read-only, or if the filesystem is shutdown. Read-only
383 * mounts allow internal writes for log recovery and unmount purposes,
384 * so don't restrict that case.
386 if (xfs_has_norecovery(mp))
388 if (xfs_readonly_buftarg(mp->m_ddev_targp))
390 if (xfs_readonly_buftarg(mp->m_log->l_targ))
392 if (xlog_is_shutdown(mp->m_log))
398 * Replenish the byte reservation required by moving the grant write head.
402 struct xfs_mount *mp,
403 struct xlog_ticket *tic)
405 struct xlog *log = mp->m_log;
409 if (xlog_is_shutdown(log))
412 XFS_STATS_INC(mp, xs_try_logspace);
415 * This is a new transaction on the ticket, so we need to change the
416 * transaction ID so that the next transaction has a different TID in
417 * the log. Just add one to the existing tid so that we can see chains
418 * of rolling transactions in the log easily.
422 xlog_grant_push_ail(log, tic->t_unit_res);
424 tic->t_curr_res = tic->t_unit_res;
428 trace_xfs_log_regrant(log, tic);
430 error = xlog_grant_head_check(log, &log->l_write_head, tic,
435 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
436 trace_xfs_log_regrant_exit(log, tic);
437 xlog_verify_grant_tail(log);
442 * If we are failing, make sure the ticket doesn't have any current
443 * reservations. We don't want to add this back when the ticket/
444 * transaction gets cancelled.
447 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
452 * Reserve log space and return a ticket corresponding to the reservation.
454 * Each reservation is going to reserve extra space for a log record header.
455 * When writes happen to the on-disk log, we don't subtract the length of the
456 * log record header from any reservation. By wasting space in each
457 * reservation, we prevent over allocation problems.
461 struct xfs_mount *mp,
464 struct xlog_ticket **ticp,
467 struct xlog *log = mp->m_log;
468 struct xlog_ticket *tic;
472 if (xlog_is_shutdown(log))
475 XFS_STATS_INC(mp, xs_try_logspace);
477 ASSERT(*ticp == NULL);
478 tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
481 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
484 trace_xfs_log_reserve(log, tic);
486 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
491 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
492 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
493 trace_xfs_log_reserve_exit(log, tic);
494 xlog_verify_grant_tail(log);
499 * If we are failing, make sure the ticket doesn't have any current
500 * reservations. We don't want to add this back when the ticket/
501 * transaction gets cancelled.
504 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
509 * Run all the pending iclog callbacks and wake log force waiters and iclog
510 * space waiters so they can process the newly set shutdown state. We really
511 * don't care what order we process callbacks here because the log is shut down
512 * and so state cannot change on disk anymore. However, we cannot wake waiters
513 * until the callbacks have been processed because we may be in unmount and
514 * we must ensure that all AIL operations the callbacks perform have completed
515 * before we tear down the AIL.
517 * We avoid processing actively referenced iclogs so that we don't run callbacks
518 * while the iclog owner might still be preparing the iclog for IO submssion.
519 * These will be caught by xlog_state_iclog_release() and call this function
520 * again to process any callbacks that may have been added to that iclog.
523 xlog_state_shutdown_callbacks(
526 struct xlog_in_core *iclog;
529 iclog = log->l_iclog;
531 if (atomic_read(&iclog->ic_refcnt)) {
532 /* Reference holder will re-run iclog callbacks. */
535 list_splice_init(&iclog->ic_callbacks, &cb_list);
536 spin_unlock(&log->l_icloglock);
538 xlog_cil_process_committed(&cb_list);
540 spin_lock(&log->l_icloglock);
541 wake_up_all(&iclog->ic_write_wait);
542 wake_up_all(&iclog->ic_force_wait);
543 } while ((iclog = iclog->ic_next) != log->l_iclog);
545 wake_up_all(&log->l_flush_wait);
549 * Flush iclog to disk if this is the last reference to the given iclog and the
550 * it is in the WANT_SYNC state.
552 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
553 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
554 * written to stable storage, and implies that a commit record is contained
555 * within the iclog. We need to ensure that the log tail does not move beyond
556 * the tail that the first commit record in the iclog ordered against, otherwise
557 * correct recovery of that checkpoint becomes dependent on future operations
558 * performed on this iclog.
560 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
561 * current tail into iclog. Once the iclog tail is set, future operations must
562 * not modify it, otherwise they potentially violate ordering constraints for
563 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
564 * the iclog will get zeroed on activation of the iclog after sync, so we
565 * always capture the tail lsn on the iclog on the first NEED_FUA release
566 * regardless of the number of active reference counts on this iclog.
569 xlog_state_release_iclog(
571 struct xlog_in_core *iclog,
572 struct xlog_ticket *ticket)
577 lockdep_assert_held(&log->l_icloglock);
579 trace_xlog_iclog_release(iclog, _RET_IP_);
581 * Grabbing the current log tail needs to be atomic w.r.t. the writing
582 * of the tail LSN into the iclog so we guarantee that the log tail does
583 * not move between the first time we know that the iclog needs to be
584 * made stable and when we eventually submit it.
586 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
587 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
588 !iclog->ic_header.h_tail_lsn) {
589 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
590 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
593 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
595 if (xlog_is_shutdown(log)) {
597 * If there are no more references to this iclog, process the
598 * pending iclog callbacks that were waiting on the release of
602 xlog_state_shutdown_callbacks(log);
609 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
610 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
614 iclog->ic_state = XLOG_STATE_SYNCING;
615 xlog_verify_tail_lsn(log, iclog);
616 trace_xlog_iclog_syncing(iclog, _RET_IP_);
618 spin_unlock(&log->l_icloglock);
619 xlog_sync(log, iclog, ticket);
620 spin_lock(&log->l_icloglock);
625 * Mount a log filesystem
627 * mp - ubiquitous xfs mount point structure
628 * log_target - buftarg of on-disk log device
629 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
630 * num_bblocks - Number of BBSIZE blocks in on-disk log
632 * Return error or zero.
637 struct xfs_buftarg *log_target,
638 xfs_daddr_t blk_offset,
645 if (!xfs_has_norecovery(mp)) {
646 xfs_notice(mp, "Mounting V%d Filesystem %pU",
647 XFS_SB_VERSION_NUM(&mp->m_sb),
651 "Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
652 XFS_SB_VERSION_NUM(&mp->m_sb),
654 ASSERT(xfs_is_readonly(mp));
657 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
659 error = PTR_ERR(log);
665 * Now that we have set up the log and it's internal geometry
666 * parameters, we can validate the given log space and drop a critical
667 * message via syslog if the log size is too small. A log that is too
668 * small can lead to unexpected situations in transaction log space
669 * reservation stage. The superblock verifier has already validated all
670 * the other log geometry constraints, so we don't have to check those
673 * Note: For v4 filesystems, we can't just reject the mount if the
674 * validation fails. This would mean that people would have to
675 * downgrade their kernel just to remedy the situation as there is no
676 * way to grow the log (short of black magic surgery with xfs_db).
678 * We can, however, reject mounts for V5 format filesystems, as the
679 * mkfs binary being used to make the filesystem should never create a
680 * filesystem with a log that is too small.
682 min_logfsbs = xfs_log_calc_minimum_size(mp);
683 if (mp->m_sb.sb_logblocks < min_logfsbs) {
685 "Log size %d blocks too small, minimum size is %d blocks",
686 mp->m_sb.sb_logblocks, min_logfsbs);
689 * Log check errors are always fatal on v5; or whenever bad
690 * metadata leads to a crash.
692 if (xfs_has_crc(mp)) {
693 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
698 xfs_crit(mp, "Log size out of supported range.");
700 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
704 * Initialize the AIL now we have a log.
706 error = xfs_trans_ail_init(mp);
708 xfs_warn(mp, "AIL initialisation failed: error %d", error);
711 log->l_ailp = mp->m_ail;
714 * skip log recovery on a norecovery mount. pretend it all
717 if (!xfs_has_norecovery(mp)) {
718 error = xlog_recover(log);
720 xfs_warn(mp, "log mount/recovery failed: error %d",
722 xlog_recover_cancel(log);
723 goto out_destroy_ail;
727 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
730 goto out_destroy_ail;
732 /* Normal transactions can now occur */
733 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
736 * Now the log has been fully initialised and we know were our
737 * space grant counters are, we can initialise the permanent ticket
738 * needed for delayed logging to work.
740 xlog_cil_init_post_recovery(log);
745 xfs_trans_ail_destroy(mp);
747 xlog_dealloc_log(log);
753 * Finish the recovery of the file system. This is separate from the
754 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
755 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
758 * If we finish recovery successfully, start the background log work. If we are
759 * not doing recovery, then we have a RO filesystem and we don't need to start
763 xfs_log_mount_finish(
764 struct xfs_mount *mp)
766 struct xlog *log = mp->m_log;
769 if (xfs_has_norecovery(mp)) {
770 ASSERT(xfs_is_readonly(mp));
775 * During the second phase of log recovery, we need iget and
776 * iput to behave like they do for an active filesystem.
777 * xfs_fs_drop_inode needs to be able to prevent the deletion
778 * of inodes before we're done replaying log items on those
779 * inodes. Turn it off immediately after recovery finishes
780 * so that we don't leak the quota inodes if subsequent mount
783 * We let all inodes involved in redo item processing end up on
784 * the LRU instead of being evicted immediately so that if we do
785 * something to an unlinked inode, the irele won't cause
786 * premature truncation and freeing of the inode, which results
787 * in log recovery failure. We have to evict the unreferenced
788 * lru inodes after clearing SB_ACTIVE because we don't
789 * otherwise clean up the lru if there's a subsequent failure in
790 * xfs_mountfs, which leads to us leaking the inodes if nothing
791 * else (e.g. quotacheck) references the inodes before the
792 * mount failure occurs.
794 mp->m_super->s_flags |= SB_ACTIVE;
795 xfs_log_work_queue(mp);
796 if (xlog_recovery_needed(log))
797 error = xlog_recover_finish(log);
798 mp->m_super->s_flags &= ~SB_ACTIVE;
799 evict_inodes(mp->m_super);
802 * Drain the buffer LRU after log recovery. This is required for v4
803 * filesystems to avoid leaving around buffers with NULL verifier ops,
804 * but we do it unconditionally to make sure we're always in a clean
805 * cache state after mount.
807 * Don't push in the error case because the AIL may have pending intents
808 * that aren't removed until recovery is cancelled.
810 if (xlog_recovery_needed(log)) {
812 xfs_log_force(mp, XFS_LOG_SYNC);
813 xfs_ail_push_all_sync(mp->m_ail);
815 xfs_notice(mp, "Ending recovery (logdev: %s)",
816 mp->m_logname ? mp->m_logname : "internal");
818 xfs_info(mp, "Ending clean mount");
820 xfs_buftarg_drain(mp->m_ddev_targp);
822 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
824 /* Make sure the log is dead if we're returning failure. */
825 ASSERT(!error || xlog_is_shutdown(log));
831 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
835 xfs_log_mount_cancel(
836 struct xfs_mount *mp)
838 xlog_recover_cancel(mp->m_log);
843 * Flush out the iclog to disk ensuring that device caches are flushed and
844 * the iclog hits stable storage before any completion waiters are woken.
848 struct xlog_in_core *iclog)
850 atomic_inc(&iclog->ic_refcnt);
851 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
852 if (iclog->ic_state == XLOG_STATE_ACTIVE)
853 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
854 return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
858 * Cycle all the iclogbuf locks to make sure all log IO completion
859 * is done before we tear down these buffers.
862 xlog_wait_iclog_completion(struct xlog *log)
865 struct xlog_in_core *iclog = log->l_iclog;
867 for (i = 0; i < log->l_iclog_bufs; i++) {
868 down(&iclog->ic_sema);
870 iclog = iclog->ic_next;
875 * Wait for the iclog and all prior iclogs to be written disk as required by the
876 * log force state machine. Waiting on ic_force_wait ensures iclog completions
877 * have been ordered and callbacks run before we are woken here, hence
878 * guaranteeing that all the iclogs up to this one are on stable storage.
882 struct xlog_in_core *iclog)
883 __releases(iclog->ic_log->l_icloglock)
885 struct xlog *log = iclog->ic_log;
887 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
888 if (!xlog_is_shutdown(log) &&
889 iclog->ic_state != XLOG_STATE_ACTIVE &&
890 iclog->ic_state != XLOG_STATE_DIRTY) {
891 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
892 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
894 spin_unlock(&log->l_icloglock);
897 if (xlog_is_shutdown(log))
903 * Write out an unmount record using the ticket provided. We have to account for
904 * the data space used in the unmount ticket as this write is not done from a
905 * transaction context that has already done the accounting for us.
908 xlog_write_unmount_record(
910 struct xlog_ticket *ticket)
913 struct xlog_op_header ophdr;
914 struct xfs_unmount_log_format ulf;
917 .oh_clientid = XFS_LOG,
918 .oh_tid = cpu_to_be32(ticket->t_tid),
919 .oh_flags = XLOG_UNMOUNT_TRANS,
922 .magic = XLOG_UNMOUNT_TYPE,
925 struct xfs_log_iovec reg = {
926 .i_addr = &unmount_rec,
927 .i_len = sizeof(unmount_rec),
928 .i_type = XLOG_REG_TYPE_UNMOUNT,
930 struct xfs_log_vec vec = {
935 list_add(&vec.lv_list, &lv_chain);
937 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
938 sizeof(struct xfs_unmount_log_format)) !=
939 sizeof(unmount_rec));
941 /* account for space used by record data */
942 ticket->t_curr_res -= sizeof(unmount_rec);
944 return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
948 * Mark the filesystem clean by writing an unmount record to the head of the
955 struct xfs_mount *mp = log->l_mp;
956 struct xlog_in_core *iclog;
957 struct xlog_ticket *tic = NULL;
960 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
964 error = xlog_write_unmount_record(log, tic);
966 * At this point, we're umounting anyway, so there's no point in
967 * transitioning log state to shutdown. Just continue...
971 xfs_alert(mp, "%s: unmount record failed", __func__);
973 spin_lock(&log->l_icloglock);
974 iclog = log->l_iclog;
975 error = xlog_force_iclog(iclog);
976 xlog_wait_on_iclog(iclog);
979 trace_xfs_log_umount_write(log, tic);
980 xfs_log_ticket_ungrant(log, tic);
985 xfs_log_unmount_verify_iclog(
988 struct xlog_in_core *iclog = log->l_iclog;
991 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
992 ASSERT(iclog->ic_offset == 0);
993 } while ((iclog = iclog->ic_next) != log->l_iclog);
997 * Unmount record used to have a string "Unmount filesystem--" in the
998 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
999 * We just write the magic number now since that particular field isn't
1000 * currently architecture converted and "Unmount" is a bit foo.
1001 * As far as I know, there weren't any dependencies on the old behaviour.
1004 xfs_log_unmount_write(
1005 struct xfs_mount *mp)
1007 struct xlog *log = mp->m_log;
1009 if (!xfs_log_writable(mp))
1012 xfs_log_force(mp, XFS_LOG_SYNC);
1014 if (xlog_is_shutdown(log))
1018 * If we think the summary counters are bad, avoid writing the unmount
1019 * record to force log recovery at next mount, after which the summary
1020 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1023 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1024 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1025 xfs_alert(mp, "%s: will fix summary counters at next mount",
1030 xfs_log_unmount_verify_iclog(log);
1031 xlog_unmount_write(log);
1035 * Empty the log for unmount/freeze.
1037 * To do this, we first need to shut down the background log work so it is not
1038 * trying to cover the log as we clean up. We then need to unpin all objects in
1039 * the log so we can then flush them out. Once they have completed their IO and
1040 * run the callbacks removing themselves from the AIL, we can cover the log.
1044 struct xfs_mount *mp)
1047 * Clear log incompat features since we're quiescing the log. Report
1048 * failures, though it's not fatal to have a higher log feature
1049 * protection level than the log contents actually require.
1051 if (xfs_clear_incompat_log_features(mp)) {
1054 error = xfs_sync_sb(mp, false);
1057 "Failed to clear log incompat features on quiesce");
1060 cancel_delayed_work_sync(&mp->m_log->l_work);
1061 xfs_log_force(mp, XFS_LOG_SYNC);
1064 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1065 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1066 * xfs_buf_iowait() cannot be used because it was pushed with the
1067 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1068 * the IO to complete.
1070 xfs_ail_push_all_sync(mp->m_ail);
1071 xfs_buftarg_wait(mp->m_ddev_targp);
1072 xfs_buf_lock(mp->m_sb_bp);
1073 xfs_buf_unlock(mp->m_sb_bp);
1075 return xfs_log_cover(mp);
1080 struct xfs_mount *mp)
1082 xfs_log_quiesce(mp);
1083 xfs_log_unmount_write(mp);
1087 * Shut down and release the AIL and Log.
1089 * During unmount, we need to ensure we flush all the dirty metadata objects
1090 * from the AIL so that the log is empty before we write the unmount record to
1091 * the log. Once this is done, we can tear down the AIL and the log.
1095 struct xfs_mount *mp)
1100 * If shutdown has come from iclog IO context, the log
1101 * cleaning will have been skipped and so we need to wait
1102 * for the iclog to complete shutdown processing before we
1103 * tear anything down.
1105 xlog_wait_iclog_completion(mp->m_log);
1107 xfs_buftarg_drain(mp->m_ddev_targp);
1109 xfs_trans_ail_destroy(mp);
1111 xfs_sysfs_del(&mp->m_log->l_kobj);
1113 xlog_dealloc_log(mp->m_log);
1118 struct xfs_mount *mp,
1119 struct xfs_log_item *item,
1121 const struct xfs_item_ops *ops)
1123 item->li_log = mp->m_log;
1124 item->li_ailp = mp->m_ail;
1125 item->li_type = type;
1129 INIT_LIST_HEAD(&item->li_ail);
1130 INIT_LIST_HEAD(&item->li_cil);
1131 INIT_LIST_HEAD(&item->li_bio_list);
1132 INIT_LIST_HEAD(&item->li_trans);
1136 * Wake up processes waiting for log space after we have moved the log tail.
1140 struct xfs_mount *mp)
1142 struct xlog *log = mp->m_log;
1145 if (xlog_is_shutdown(log))
1148 if (!list_empty_careful(&log->l_write_head.waiters)) {
1149 ASSERT(!xlog_in_recovery(log));
1151 spin_lock(&log->l_write_head.lock);
1152 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1153 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1154 spin_unlock(&log->l_write_head.lock);
1157 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1158 ASSERT(!xlog_in_recovery(log));
1160 spin_lock(&log->l_reserve_head.lock);
1161 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1162 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1163 spin_unlock(&log->l_reserve_head.lock);
1168 * Determine if we have a transaction that has gone to disk that needs to be
1169 * covered. To begin the transition to the idle state firstly the log needs to
1170 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1171 * we start attempting to cover the log.
1173 * Only if we are then in a state where covering is needed, the caller is
1174 * informed that dummy transactions are required to move the log into the idle
1177 * If there are any items in the AIl or CIL, then we do not want to attempt to
1178 * cover the log as we may be in a situation where there isn't log space
1179 * available to run a dummy transaction and this can lead to deadlocks when the
1180 * tail of the log is pinned by an item that is modified in the CIL. Hence
1181 * there's no point in running a dummy transaction at this point because we
1182 * can't start trying to idle the log until both the CIL and AIL are empty.
1185 xfs_log_need_covered(
1186 struct xfs_mount *mp)
1188 struct xlog *log = mp->m_log;
1189 bool needed = false;
1191 if (!xlog_cil_empty(log))
1194 spin_lock(&log->l_icloglock);
1195 switch (log->l_covered_state) {
1196 case XLOG_STATE_COVER_DONE:
1197 case XLOG_STATE_COVER_DONE2:
1198 case XLOG_STATE_COVER_IDLE:
1200 case XLOG_STATE_COVER_NEED:
1201 case XLOG_STATE_COVER_NEED2:
1202 if (xfs_ail_min_lsn(log->l_ailp))
1204 if (!xlog_iclogs_empty(log))
1208 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1209 log->l_covered_state = XLOG_STATE_COVER_DONE;
1211 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1217 spin_unlock(&log->l_icloglock);
1222 * Explicitly cover the log. This is similar to background log covering but
1223 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1224 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1225 * must all be empty.
1229 struct xfs_mount *mp)
1234 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1235 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1236 xlog_is_shutdown(mp->m_log));
1238 if (!xfs_log_writable(mp))
1242 * xfs_log_need_covered() is not idempotent because it progresses the
1243 * state machine if the log requires covering. Therefore, we must call
1244 * this function once and use the result until we've issued an sb sync.
1245 * Do so first to make that abundantly clear.
1247 * Fall into the covering sequence if the log needs covering or the
1248 * mount has lazy superblock accounting to sync to disk. The sb sync
1249 * used for covering accumulates the in-core counters, so covering
1250 * handles this for us.
1252 need_covered = xfs_log_need_covered(mp);
1253 if (!need_covered && !xfs_has_lazysbcount(mp))
1257 * To cover the log, commit the superblock twice (at most) in
1258 * independent checkpoints. The first serves as a reference for the
1259 * tail pointer. The sync transaction and AIL push empties the AIL and
1260 * updates the in-core tail to the LSN of the first checkpoint. The
1261 * second commit updates the on-disk tail with the in-core LSN,
1262 * covering the log. Push the AIL one more time to leave it empty, as
1266 error = xfs_sync_sb(mp, true);
1269 xfs_ail_push_all_sync(mp->m_ail);
1270 } while (xfs_log_need_covered(mp));
1276 * We may be holding the log iclog lock upon entering this routine.
1279 xlog_assign_tail_lsn_locked(
1280 struct xfs_mount *mp)
1282 struct xlog *log = mp->m_log;
1283 struct xfs_log_item *lip;
1286 assert_spin_locked(&mp->m_ail->ail_lock);
1289 * To make sure we always have a valid LSN for the log tail we keep
1290 * track of the last LSN which was committed in log->l_last_sync_lsn,
1291 * and use that when the AIL was empty.
1293 lip = xfs_ail_min(mp->m_ail);
1295 tail_lsn = lip->li_lsn;
1297 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1298 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1299 atomic64_set(&log->l_tail_lsn, tail_lsn);
1304 xlog_assign_tail_lsn(
1305 struct xfs_mount *mp)
1309 spin_lock(&mp->m_ail->ail_lock);
1310 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1311 spin_unlock(&mp->m_ail->ail_lock);
1317 * Return the space in the log between the tail and the head. The head
1318 * is passed in the cycle/bytes formal parms. In the special case where
1319 * the reserve head has wrapped passed the tail, this calculation is no
1320 * longer valid. In this case, just return 0 which means there is no space
1321 * in the log. This works for all places where this function is called
1322 * with the reserve head. Of course, if the write head were to ever
1323 * wrap the tail, we should blow up. Rather than catch this case here,
1324 * we depend on other ASSERTions in other parts of the code. XXXmiken
1326 * If reservation head is behind the tail, we have a problem. Warn about it,
1327 * but then treat it as if the log is empty.
1329 * If the log is shut down, the head and tail may be invalid or out of whack, so
1330 * shortcut invalidity asserts in this case so that we don't trigger them
1343 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1344 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1345 tail_bytes = BBTOB(tail_bytes);
1346 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1347 return log->l_logsize - (head_bytes - tail_bytes);
1348 if (tail_cycle + 1 < head_cycle)
1351 /* Ignore potential inconsistency when shutdown. */
1352 if (xlog_is_shutdown(log))
1353 return log->l_logsize;
1355 if (tail_cycle < head_cycle) {
1356 ASSERT(tail_cycle == (head_cycle - 1));
1357 return tail_bytes - head_bytes;
1361 * The reservation head is behind the tail. In this case we just want to
1362 * return the size of the log as the amount of space left.
1364 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1365 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1366 tail_cycle, tail_bytes);
1367 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1368 head_cycle, head_bytes);
1370 return log->l_logsize;
1376 struct work_struct *work)
1378 struct xlog_in_core *iclog =
1379 container_of(work, struct xlog_in_core, ic_end_io_work);
1380 struct xlog *log = iclog->ic_log;
1383 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1385 /* treat writes with injected CRC errors as failed */
1386 if (iclog->ic_fail_crc)
1391 * Race to shutdown the filesystem if we see an error.
1393 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1394 xfs_alert(log->l_mp, "log I/O error %d", error);
1395 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1398 xlog_state_done_syncing(iclog);
1399 bio_uninit(&iclog->ic_bio);
1402 * Drop the lock to signal that we are done. Nothing references the
1403 * iclog after this, so an unmount waiting on this lock can now tear it
1404 * down safely. As such, it is unsafe to reference the iclog after the
1405 * unlock as we could race with it being freed.
1407 up(&iclog->ic_sema);
1411 * Return size of each in-core log record buffer.
1413 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1415 * If the filesystem blocksize is too large, we may need to choose a
1416 * larger size since the directory code currently logs entire blocks.
1419 xlog_get_iclog_buffer_size(
1420 struct xfs_mount *mp,
1423 if (mp->m_logbufs <= 0)
1424 mp->m_logbufs = XLOG_MAX_ICLOGS;
1425 if (mp->m_logbsize <= 0)
1426 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1428 log->l_iclog_bufs = mp->m_logbufs;
1429 log->l_iclog_size = mp->m_logbsize;
1432 * # headers = size / 32k - one header holds cycles from 32k of data.
1434 log->l_iclog_heads =
1435 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1436 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1441 struct xfs_mount *mp)
1443 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1444 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1448 * Clear the log incompat flags if we have the opportunity.
1450 * This only happens if we're about to log the second dummy transaction as part
1451 * of covering the log and we can get the log incompat feature usage lock.
1454 xlog_clear_incompat(
1457 struct xfs_mount *mp = log->l_mp;
1459 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1460 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1463 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1466 if (!down_write_trylock(&log->l_incompat_users))
1469 xfs_clear_incompat_log_features(mp);
1470 up_write(&log->l_incompat_users);
1474 * Every sync period we need to unpin all items in the AIL and push them to
1475 * disk. If there is nothing dirty, then we might need to cover the log to
1476 * indicate that the filesystem is idle.
1480 struct work_struct *work)
1482 struct xlog *log = container_of(to_delayed_work(work),
1483 struct xlog, l_work);
1484 struct xfs_mount *mp = log->l_mp;
1486 /* dgc: errors ignored - not fatal and nowhere to report them */
1487 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1489 * Dump a transaction into the log that contains no real change.
1490 * This is needed to stamp the current tail LSN into the log
1491 * during the covering operation.
1493 * We cannot use an inode here for this - that will push dirty
1494 * state back up into the VFS and then periodic inode flushing
1495 * will prevent log covering from making progress. Hence we
1496 * synchronously log the superblock instead to ensure the
1497 * superblock is immediately unpinned and can be written back.
1499 xlog_clear_incompat(log);
1500 xfs_sync_sb(mp, true);
1502 xfs_log_force(mp, 0);
1504 /* start pushing all the metadata that is currently dirty */
1505 xfs_ail_push_all(mp->m_ail);
1507 /* queue us up again */
1508 xfs_log_work_queue(mp);
1512 * This routine initializes some of the log structure for a given mount point.
1513 * Its primary purpose is to fill in enough, so recovery can occur. However,
1514 * some other stuff may be filled in too.
1516 STATIC struct xlog *
1518 struct xfs_mount *mp,
1519 struct xfs_buftarg *log_target,
1520 xfs_daddr_t blk_offset,
1524 xlog_rec_header_t *head;
1525 xlog_in_core_t **iclogp;
1526 xlog_in_core_t *iclog, *prev_iclog=NULL;
1528 int error = -ENOMEM;
1531 log = kzalloc(sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1533 xfs_warn(mp, "Log allocation failed: No memory!");
1538 log->l_targ = log_target;
1539 log->l_logsize = BBTOB(num_bblks);
1540 log->l_logBBstart = blk_offset;
1541 log->l_logBBsize = num_bblks;
1542 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1543 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1544 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1545 INIT_LIST_HEAD(&log->r_dfops);
1547 log->l_prev_block = -1;
1548 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1549 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1550 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1551 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1553 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1554 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1556 log->l_iclog_roundoff = BBSIZE;
1558 xlog_grant_head_init(&log->l_reserve_head);
1559 xlog_grant_head_init(&log->l_write_head);
1561 error = -EFSCORRUPTED;
1562 if (xfs_has_sector(mp)) {
1563 log2_size = mp->m_sb.sb_logsectlog;
1564 if (log2_size < BBSHIFT) {
1565 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1566 log2_size, BBSHIFT);
1570 log2_size -= BBSHIFT;
1571 if (log2_size > mp->m_sectbb_log) {
1572 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1573 log2_size, mp->m_sectbb_log);
1577 /* for larger sector sizes, must have v2 or external log */
1578 if (log2_size && log->l_logBBstart > 0 &&
1579 !xfs_has_logv2(mp)) {
1581 "log sector size (0x%x) invalid for configuration.",
1586 log->l_sectBBsize = 1 << log2_size;
1588 init_rwsem(&log->l_incompat_users);
1590 xlog_get_iclog_buffer_size(mp, log);
1592 spin_lock_init(&log->l_icloglock);
1593 init_waitqueue_head(&log->l_flush_wait);
1595 iclogp = &log->l_iclog;
1597 * The amount of memory to allocate for the iclog structure is
1598 * rather funky due to the way the structure is defined. It is
1599 * done this way so that we can use different sizes for machines
1600 * with different amounts of memory. See the definition of
1601 * xlog_in_core_t in xfs_log_priv.h for details.
1603 ASSERT(log->l_iclog_size >= 4096);
1604 for (i = 0; i < log->l_iclog_bufs; i++) {
1605 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1606 sizeof(struct bio_vec);
1608 iclog = kzalloc(sizeof(*iclog) + bvec_size,
1609 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1611 goto out_free_iclog;
1614 iclog->ic_prev = prev_iclog;
1617 iclog->ic_data = kvzalloc(log->l_iclog_size,
1618 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1619 if (!iclog->ic_data)
1620 goto out_free_iclog;
1621 head = &iclog->ic_header;
1622 memset(head, 0, sizeof(xlog_rec_header_t));
1623 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1624 head->h_version = cpu_to_be32(
1625 xfs_has_logv2(log->l_mp) ? 2 : 1);
1626 head->h_size = cpu_to_be32(log->l_iclog_size);
1628 head->h_fmt = cpu_to_be32(XLOG_FMT);
1629 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1631 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1632 iclog->ic_state = XLOG_STATE_ACTIVE;
1633 iclog->ic_log = log;
1634 atomic_set(&iclog->ic_refcnt, 0);
1635 INIT_LIST_HEAD(&iclog->ic_callbacks);
1636 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1638 init_waitqueue_head(&iclog->ic_force_wait);
1639 init_waitqueue_head(&iclog->ic_write_wait);
1640 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1641 sema_init(&iclog->ic_sema, 1);
1643 iclogp = &iclog->ic_next;
1645 *iclogp = log->l_iclog; /* complete ring */
1646 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1648 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1649 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1651 0, mp->m_super->s_id);
1652 if (!log->l_ioend_workqueue)
1653 goto out_free_iclog;
1655 error = xlog_cil_init(log);
1657 goto out_destroy_workqueue;
1660 out_destroy_workqueue:
1661 destroy_workqueue(log->l_ioend_workqueue);
1663 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1664 prev_iclog = iclog->ic_next;
1665 kvfree(iclog->ic_data);
1667 if (prev_iclog == log->l_iclog)
1673 return ERR_PTR(error);
1674 } /* xlog_alloc_log */
1677 * Compute the LSN that we'd need to push the log tail towards in order to have
1678 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1679 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1680 * log free space already meets all three thresholds, this function returns
1684 xlog_grant_push_threshold(
1688 xfs_lsn_t threshold_lsn = 0;
1689 xfs_lsn_t last_sync_lsn;
1692 int threshold_block;
1693 int threshold_cycle;
1696 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1698 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1699 free_blocks = BTOBBT(free_bytes);
1702 * Set the threshold for the minimum number of free blocks in the
1703 * log to the maximum of what the caller needs, one quarter of the
1704 * log, and 256 blocks.
1706 free_threshold = BTOBB(need_bytes);
1707 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1708 free_threshold = max(free_threshold, 256);
1709 if (free_blocks >= free_threshold)
1710 return NULLCOMMITLSN;
1712 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1714 threshold_block += free_threshold;
1715 if (threshold_block >= log->l_logBBsize) {
1716 threshold_block -= log->l_logBBsize;
1717 threshold_cycle += 1;
1719 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1722 * Don't pass in an lsn greater than the lsn of the last
1723 * log record known to be on disk. Use a snapshot of the last sync lsn
1724 * so that it doesn't change between the compare and the set.
1726 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1727 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1728 threshold_lsn = last_sync_lsn;
1730 return threshold_lsn;
1734 * Push the tail of the log if we need to do so to maintain the free log space
1735 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1736 * policy which pushes on an lsn which is further along in the log once we
1737 * reach the high water mark. In this manner, we would be creating a low water
1741 xlog_grant_push_ail(
1745 xfs_lsn_t threshold_lsn;
1747 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1748 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1752 * Get the transaction layer to kick the dirty buffers out to
1753 * disk asynchronously. No point in trying to do this if
1754 * the filesystem is shutting down.
1756 xfs_ail_push(log->l_ailp, threshold_lsn);
1760 * Stamp cycle number in every block
1765 struct xlog_in_core *iclog,
1769 int size = iclog->ic_offset + roundoff;
1773 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1775 dp = iclog->ic_datap;
1776 for (i = 0; i < BTOBB(size); i++) {
1777 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1779 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1780 *(__be32 *)dp = cycle_lsn;
1784 if (xfs_has_logv2(log->l_mp)) {
1785 xlog_in_core_2_t *xhdr = iclog->ic_data;
1787 for ( ; i < BTOBB(size); i++) {
1788 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1789 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1790 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1791 *(__be32 *)dp = cycle_lsn;
1795 for (i = 1; i < log->l_iclog_heads; i++)
1796 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1801 * Calculate the checksum for a log buffer.
1803 * This is a little more complicated than it should be because the various
1804 * headers and the actual data are non-contiguous.
1809 struct xlog_rec_header *rhead,
1815 /* first generate the crc for the record header ... */
1816 crc = xfs_start_cksum_update((char *)rhead,
1817 sizeof(struct xlog_rec_header),
1818 offsetof(struct xlog_rec_header, h_crc));
1820 /* ... then for additional cycle data for v2 logs ... */
1821 if (xfs_has_logv2(log->l_mp)) {
1822 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1826 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1828 for (i = 1; i < xheads; i++) {
1829 crc = crc32c(crc, &xhdr[i].hic_xheader,
1830 sizeof(struct xlog_rec_ext_header));
1834 /* ... and finally for the payload */
1835 crc = crc32c(crc, dp, size);
1837 return xfs_end_cksum(crc);
1844 struct xlog_in_core *iclog = bio->bi_private;
1846 queue_work(iclog->ic_log->l_ioend_workqueue,
1847 &iclog->ic_end_io_work);
1851 xlog_map_iclog_data(
1857 struct page *page = kmem_to_page(data);
1858 unsigned int off = offset_in_page(data);
1859 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1861 if (bio_add_page(bio, page, len, off) != len)
1874 struct xlog_in_core *iclog,
1878 ASSERT(bno < log->l_logBBsize);
1879 trace_xlog_iclog_write(iclog, _RET_IP_);
1882 * We lock the iclogbufs here so that we can serialise against I/O
1883 * completion during unmount. We might be processing a shutdown
1884 * triggered during unmount, and that can occur asynchronously to the
1885 * unmount thread, and hence we need to ensure that completes before
1886 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1887 * across the log IO to archieve that.
1889 down(&iclog->ic_sema);
1890 if (xlog_is_shutdown(log)) {
1892 * It would seem logical to return EIO here, but we rely on
1893 * the log state machine to propagate I/O errors instead of
1894 * doing it here. We kick of the state machine and unlock
1895 * the buffer manually, the code needs to be kept in sync
1896 * with the I/O completion path.
1902 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1903 * IOs coming immediately after this one. This prevents the block layer
1904 * writeback throttle from throttling log writes behind background
1905 * metadata writeback and causing priority inversions.
1907 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1908 howmany(count, PAGE_SIZE),
1909 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1910 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1911 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1912 iclog->ic_bio.bi_private = iclog;
1914 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1915 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1917 * For external log devices, we also need to flush the data
1918 * device cache first to ensure all metadata writeback covered
1919 * by the LSN in this iclog is on stable storage. This is slow,
1920 * but it *must* complete before we issue the external log IO.
1922 * If the flush fails, we cannot conclude that past metadata
1923 * writeback from the log succeeded. Repeating the flush is
1924 * not possible, hence we must shut down with log IO error to
1925 * avoid shutdown re-entering this path and erroring out again.
1927 if (log->l_targ != log->l_mp->m_ddev_targp &&
1928 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev))
1931 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1932 iclog->ic_bio.bi_opf |= REQ_FUA;
1934 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1936 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count))
1939 if (is_vmalloc_addr(iclog->ic_data))
1940 flush_kernel_vmap_range(iclog->ic_data, count);
1943 * If this log buffer would straddle the end of the log we will have
1944 * to split it up into two bios, so that we can continue at the start.
1946 if (bno + BTOBB(count) > log->l_logBBsize) {
1949 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1950 GFP_NOIO, &fs_bio_set);
1951 bio_chain(split, &iclog->ic_bio);
1954 /* restart at logical offset zero for the remainder */
1955 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1958 submit_bio(&iclog->ic_bio);
1961 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1963 xlog_state_done_syncing(iclog);
1964 up(&iclog->ic_sema);
1968 * We need to bump cycle number for the part of the iclog that is
1969 * written to the start of the log. Watch out for the header magic
1970 * number case, though.
1979 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1982 for (i = split_offset; i < count; i += BBSIZE) {
1983 uint32_t cycle = get_unaligned_be32(data + i);
1985 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1987 put_unaligned_be32(cycle, data + i);
1992 xlog_calc_iclog_size(
1994 struct xlog_in_core *iclog,
1997 uint32_t count_init, count;
1999 /* Add for LR header */
2000 count_init = log->l_iclog_hsize + iclog->ic_offset;
2001 count = roundup(count_init, log->l_iclog_roundoff);
2003 *roundoff = count - count_init;
2005 ASSERT(count >= count_init);
2006 ASSERT(*roundoff < log->l_iclog_roundoff);
2011 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2012 * fashion. Previously, we should have moved the current iclog
2013 * ptr in the log to point to the next available iclog. This allows further
2014 * write to continue while this code syncs out an iclog ready to go.
2015 * Before an in-core log can be written out, the data section must be scanned
2016 * to save away the 1st word of each BBSIZE block into the header. We replace
2017 * it with the current cycle count. Each BBSIZE block is tagged with the
2018 * cycle count because there in an implicit assumption that drives will
2019 * guarantee that entire 512 byte blocks get written at once. In other words,
2020 * we can't have part of a 512 byte block written and part not written. By
2021 * tagging each block, we will know which blocks are valid when recovering
2022 * after an unclean shutdown.
2024 * This routine is single threaded on the iclog. No other thread can be in
2025 * this routine with the same iclog. Changing contents of iclog can there-
2026 * fore be done without grabbing the state machine lock. Updating the global
2027 * log will require grabbing the lock though.
2029 * The entire log manager uses a logical block numbering scheme. Only
2030 * xlog_write_iclog knows about the fact that the log may not start with
2031 * block zero on a given device.
2036 struct xlog_in_core *iclog,
2037 struct xlog_ticket *ticket)
2039 unsigned int count; /* byte count of bwrite */
2040 unsigned int roundoff; /* roundoff to BB or stripe */
2044 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2045 trace_xlog_iclog_sync(iclog, _RET_IP_);
2047 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2050 * If we have a ticket, account for the roundoff via the ticket
2051 * reservation to avoid touching the hot grant heads needlessly.
2052 * Otherwise, we have to move grant heads directly.
2055 ticket->t_curr_res -= roundoff;
2057 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2058 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2061 /* put cycle number in every block */
2062 xlog_pack_data(log, iclog, roundoff);
2064 /* real byte length */
2065 size = iclog->ic_offset;
2066 if (xfs_has_logv2(log->l_mp))
2068 iclog->ic_header.h_len = cpu_to_be32(size);
2070 XFS_STATS_INC(log->l_mp, xs_log_writes);
2071 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2073 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2075 /* Do we need to split this write into 2 parts? */
2076 if (bno + BTOBB(count) > log->l_logBBsize)
2077 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2079 /* calculcate the checksum */
2080 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2081 iclog->ic_datap, size);
2083 * Intentionally corrupt the log record CRC based on the error injection
2084 * frequency, if defined. This facilitates testing log recovery in the
2085 * event of torn writes. Hence, set the IOABORT state to abort the log
2086 * write on I/O completion and shutdown the fs. The subsequent mount
2087 * detects the bad CRC and attempts to recover.
2090 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2091 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2092 iclog->ic_fail_crc = true;
2094 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2095 be64_to_cpu(iclog->ic_header.h_lsn));
2098 xlog_verify_iclog(log, iclog, count);
2099 xlog_write_iclog(log, iclog, bno, count);
2103 * Deallocate a log structure
2109 xlog_in_core_t *iclog, *next_iclog;
2113 * Destroy the CIL after waiting for iclog IO completion because an
2114 * iclog EIO error will try to shut down the log, which accesses the
2115 * CIL to wake up the waiters.
2117 xlog_cil_destroy(log);
2119 iclog = log->l_iclog;
2120 for (i = 0; i < log->l_iclog_bufs; i++) {
2121 next_iclog = iclog->ic_next;
2122 kvfree(iclog->ic_data);
2127 log->l_mp->m_log = NULL;
2128 destroy_workqueue(log->l_ioend_workqueue);
2133 * Update counters atomically now that memcpy is done.
2136 xlog_state_finish_copy(
2138 struct xlog_in_core *iclog,
2142 lockdep_assert_held(&log->l_icloglock);
2144 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2145 iclog->ic_offset += copy_bytes;
2149 * print out info relating to regions written which consume
2154 struct xfs_mount *mp,
2155 struct xlog_ticket *ticket)
2157 xfs_warn(mp, "ticket reservation summary:");
2158 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
2159 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
2160 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
2161 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
2165 * Print a summary of the transaction.
2169 struct xfs_trans *tp)
2171 struct xfs_mount *mp = tp->t_mountp;
2172 struct xfs_log_item *lip;
2174 /* dump core transaction and ticket info */
2175 xfs_warn(mp, "transaction summary:");
2176 xfs_warn(mp, " log res = %d", tp->t_log_res);
2177 xfs_warn(mp, " log count = %d", tp->t_log_count);
2178 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2180 xlog_print_tic_res(mp, tp->t_ticket);
2182 /* dump each log item */
2183 list_for_each_entry(lip, &tp->t_items, li_trans) {
2184 struct xfs_log_vec *lv = lip->li_lv;
2185 struct xfs_log_iovec *vec;
2188 xfs_warn(mp, "log item: ");
2189 xfs_warn(mp, " type = 0x%x", lip->li_type);
2190 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2193 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2194 xfs_warn(mp, " size = %d", lv->lv_size);
2195 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2196 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2198 /* dump each iovec for the log item */
2199 vec = lv->lv_iovecp;
2200 for (i = 0; i < lv->lv_niovecs; i++) {
2201 int dumplen = min(vec->i_len, 32);
2203 xfs_warn(mp, " iovec[%d]", i);
2204 xfs_warn(mp, " type = 0x%x", vec->i_type);
2205 xfs_warn(mp, " len = %d", vec->i_len);
2206 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2207 xfs_hex_dump(vec->i_addr, dumplen);
2216 struct xlog_in_core *iclog,
2217 uint32_t *log_offset,
2221 uint32_t *record_cnt,
2224 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2225 ASSERT(*log_offset % sizeof(int32_t) == 0);
2226 ASSERT(write_len % sizeof(int32_t) == 0);
2228 memcpy(iclog->ic_datap + *log_offset, data, write_len);
2229 *log_offset += write_len;
2230 *bytes_left -= write_len;
2232 *data_cnt += write_len;
2236 * Write log vectors into a single iclog which is guaranteed by the caller
2237 * to have enough space to write the entire log vector into.
2241 struct xfs_log_vec *lv,
2242 struct xlog_ticket *ticket,
2243 struct xlog_in_core *iclog,
2244 uint32_t *log_offset,
2246 uint32_t *record_cnt,
2251 ASSERT(*log_offset + *len <= iclog->ic_size ||
2252 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2255 * Ordered log vectors have no regions to write so this
2256 * loop will naturally skip them.
2258 for (index = 0; index < lv->lv_niovecs; index++) {
2259 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2260 struct xlog_op_header *ophdr = reg->i_addr;
2262 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2263 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2264 reg->i_len, len, record_cnt, data_cnt);
2269 xlog_write_get_more_iclog_space(
2270 struct xlog_ticket *ticket,
2271 struct xlog_in_core **iclogp,
2272 uint32_t *log_offset,
2274 uint32_t *record_cnt,
2277 struct xlog_in_core *iclog = *iclogp;
2278 struct xlog *log = iclog->ic_log;
2281 spin_lock(&log->l_icloglock);
2282 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2283 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2284 error = xlog_state_release_iclog(log, iclog, ticket);
2285 spin_unlock(&log->l_icloglock);
2289 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2300 * Write log vectors into a single iclog which is smaller than the current chain
2301 * length. We write until we cannot fit a full record into the remaining space
2302 * and then stop. We return the log vector that is to be written that cannot
2303 * wholly fit in the iclog.
2307 struct xfs_log_vec *lv,
2308 struct xlog_ticket *ticket,
2309 struct xlog_in_core **iclogp,
2310 uint32_t *log_offset,
2312 uint32_t *record_cnt,
2315 struct xlog_in_core *iclog = *iclogp;
2316 struct xlog_op_header *ophdr;
2321 /* walk the logvec, copying until we run out of space in the iclog */
2322 for (index = 0; index < lv->lv_niovecs; index++) {
2323 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2324 uint32_t reg_offset = 0;
2327 * The first region of a continuation must have a non-zero
2328 * length otherwise log recovery will just skip over it and
2329 * start recovering from the next opheader it finds. Because we
2330 * mark the next opheader as a continuation, recovery will then
2331 * incorrectly add the continuation to the previous region and
2332 * that breaks stuff.
2334 * Hence if there isn't space for region data after the
2335 * opheader, then we need to start afresh with a new iclog.
2337 if (iclog->ic_size - *log_offset <=
2338 sizeof(struct xlog_op_header)) {
2339 error = xlog_write_get_more_iclog_space(ticket,
2340 &iclog, log_offset, *len, record_cnt,
2346 ophdr = reg->i_addr;
2347 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2349 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2350 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2351 if (rlen != reg->i_len)
2352 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2354 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2355 rlen, len, record_cnt, data_cnt);
2357 /* If we wrote the whole region, move to the next. */
2358 if (rlen == reg->i_len)
2362 * We now have a partially written iovec, but it can span
2363 * multiple iclogs so we loop here. First we release the iclog
2364 * we currently have, then we get a new iclog and add a new
2365 * opheader. Then we continue copying from where we were until
2366 * we either complete the iovec or fill the iclog. If we
2367 * complete the iovec, then we increment the index and go right
2368 * back to the top of the outer loop. if we fill the iclog, we
2369 * run the inner loop again.
2371 * This is complicated by the tail of a region using all the
2372 * space in an iclog and hence requiring us to release the iclog
2373 * and get a new one before returning to the outer loop. We must
2374 * always guarantee that we exit this inner loop with at least
2375 * space for log transaction opheaders left in the current
2376 * iclog, hence we cannot just terminate the loop at the end
2377 * of the of the continuation. So we loop while there is no
2378 * space left in the current iclog, and check for the end of the
2379 * continuation after getting a new iclog.
2383 * Ensure we include the continuation opheader in the
2384 * space we need in the new iclog by adding that size
2385 * to the length we require. This continuation opheader
2386 * needs to be accounted to the ticket as the space it
2387 * consumes hasn't been accounted to the lv we are
2390 error = xlog_write_get_more_iclog_space(ticket,
2392 *len + sizeof(struct xlog_op_header),
2393 record_cnt, data_cnt);
2397 ophdr = iclog->ic_datap + *log_offset;
2398 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2399 ophdr->oh_clientid = XFS_TRANSACTION;
2401 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2403 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2404 *log_offset += sizeof(struct xlog_op_header);
2405 *data_cnt += sizeof(struct xlog_op_header);
2408 * If rlen fits in the iclog, then end the region
2409 * continuation. Otherwise we're going around again.
2412 rlen = reg->i_len - reg_offset;
2413 if (rlen <= iclog->ic_size - *log_offset)
2414 ophdr->oh_flags |= XLOG_END_TRANS;
2416 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2418 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2419 ophdr->oh_len = cpu_to_be32(rlen);
2421 xlog_write_iovec(iclog, log_offset,
2422 reg->i_addr + reg_offset,
2423 rlen, len, record_cnt, data_cnt);
2425 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2429 * No more iovecs remain in this logvec so return the next log vec to
2430 * the caller so it can go back to fast path copying.
2437 * Write some region out to in-core log
2439 * This will be called when writing externally provided regions or when
2440 * writing out a commit record for a given transaction.
2442 * General algorithm:
2443 * 1. Find total length of this write. This may include adding to the
2444 * lengths passed in.
2445 * 2. Check whether we violate the tickets reservation.
2446 * 3. While writing to this iclog
2447 * A. Reserve as much space in this iclog as can get
2448 * B. If this is first write, save away start lsn
2449 * C. While writing this region:
2450 * 1. If first write of transaction, write start record
2451 * 2. Write log operation header (header per region)
2452 * 3. Find out if we can fit entire region into this iclog
2453 * 4. Potentially, verify destination memcpy ptr
2454 * 5. Memcpy (partial) region
2455 * 6. If partial copy, release iclog; otherwise, continue
2456 * copying more regions into current iclog
2457 * 4. Mark want sync bit (in simulation mode)
2458 * 5. Release iclog for potential flush to on-disk log.
2461 * 1. Panic if reservation is overrun. This should never happen since
2462 * reservation amounts are generated internal to the filesystem.
2464 * 1. Tickets are single threaded data structures.
2465 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2466 * syncing routine. When a single log_write region needs to span
2467 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2468 * on all log operation writes which don't contain the end of the
2469 * region. The XLOG_END_TRANS bit is used for the in-core log
2470 * operation which contains the end of the continued log_write region.
2471 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2472 * we don't really know exactly how much space will be used. As a result,
2473 * we don't update ic_offset until the end when we know exactly how many
2474 * bytes have been written out.
2479 struct xfs_cil_ctx *ctx,
2480 struct list_head *lv_chain,
2481 struct xlog_ticket *ticket,
2485 struct xlog_in_core *iclog = NULL;
2486 struct xfs_log_vec *lv;
2487 uint32_t record_cnt = 0;
2488 uint32_t data_cnt = 0;
2492 if (ticket->t_curr_res < 0) {
2493 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2494 "ctx ticket reservation ran out. Need to up reservation");
2495 xlog_print_tic_res(log->l_mp, ticket);
2496 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2499 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2504 ASSERT(log_offset <= iclog->ic_size - 1);
2507 * If we have a context pointer, pass it the first iclog we are
2508 * writing to so it can record state needed for iclog write
2512 xlog_cil_set_ctx_write_state(ctx, iclog);
2514 list_for_each_entry(lv, lv_chain, lv_list) {
2516 * If the entire log vec does not fit in the iclog, punt it to
2517 * the partial copy loop which can handle this case.
2519 if (lv->lv_niovecs &&
2520 lv->lv_bytes > iclog->ic_size - log_offset) {
2521 error = xlog_write_partial(lv, ticket, &iclog,
2522 &log_offset, &len, &record_cnt,
2526 * We have no iclog to release, so just return
2527 * the error immediately.
2532 xlog_write_full(lv, ticket, iclog, &log_offset,
2533 &len, &record_cnt, &data_cnt);
2539 * We've already been guaranteed that the last writes will fit inside
2540 * the current iclog, and hence it will already have the space used by
2541 * those writes accounted to it. Hence we do not need to update the
2542 * iclog with the number of bytes written here.
2544 spin_lock(&log->l_icloglock);
2545 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2546 error = xlog_state_release_iclog(log, iclog, ticket);
2547 spin_unlock(&log->l_icloglock);
2553 xlog_state_activate_iclog(
2554 struct xlog_in_core *iclog,
2555 int *iclogs_changed)
2557 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2558 trace_xlog_iclog_activate(iclog, _RET_IP_);
2561 * If the number of ops in this iclog indicate it just contains the
2562 * dummy transaction, we can change state into IDLE (the second time
2563 * around). Otherwise we should change the state into NEED a dummy.
2564 * We don't need to cover the dummy.
2566 if (*iclogs_changed == 0 &&
2567 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2568 *iclogs_changed = 1;
2571 * We have two dirty iclogs so start over. This could also be
2572 * num of ops indicating this is not the dummy going out.
2574 *iclogs_changed = 2;
2577 iclog->ic_state = XLOG_STATE_ACTIVE;
2578 iclog->ic_offset = 0;
2579 iclog->ic_header.h_num_logops = 0;
2580 memset(iclog->ic_header.h_cycle_data, 0,
2581 sizeof(iclog->ic_header.h_cycle_data));
2582 iclog->ic_header.h_lsn = 0;
2583 iclog->ic_header.h_tail_lsn = 0;
2587 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2588 * ACTIVE after iclog I/O has completed.
2591 xlog_state_activate_iclogs(
2593 int *iclogs_changed)
2595 struct xlog_in_core *iclog = log->l_iclog;
2598 if (iclog->ic_state == XLOG_STATE_DIRTY)
2599 xlog_state_activate_iclog(iclog, iclogs_changed);
2601 * The ordering of marking iclogs ACTIVE must be maintained, so
2602 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2604 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2606 } while ((iclog = iclog->ic_next) != log->l_iclog);
2615 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2616 * wrote the first covering record (DONE). We go to IDLE if we just
2617 * wrote the second covering record (DONE2) and remain in IDLE until a
2618 * non-covering write occurs.
2620 switch (prev_state) {
2621 case XLOG_STATE_COVER_IDLE:
2622 if (iclogs_changed == 1)
2623 return XLOG_STATE_COVER_IDLE;
2625 case XLOG_STATE_COVER_NEED:
2626 case XLOG_STATE_COVER_NEED2:
2628 case XLOG_STATE_COVER_DONE:
2629 if (iclogs_changed == 1)
2630 return XLOG_STATE_COVER_NEED2;
2632 case XLOG_STATE_COVER_DONE2:
2633 if (iclogs_changed == 1)
2634 return XLOG_STATE_COVER_IDLE;
2640 return XLOG_STATE_COVER_NEED;
2644 xlog_state_clean_iclog(
2646 struct xlog_in_core *dirty_iclog)
2648 int iclogs_changed = 0;
2650 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2652 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2654 xlog_state_activate_iclogs(log, &iclogs_changed);
2655 wake_up_all(&dirty_iclog->ic_force_wait);
2657 if (iclogs_changed) {
2658 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2664 xlog_get_lowest_lsn(
2667 struct xlog_in_core *iclog = log->l_iclog;
2668 xfs_lsn_t lowest_lsn = 0, lsn;
2671 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2672 iclog->ic_state == XLOG_STATE_DIRTY)
2675 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2676 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2678 } while ((iclog = iclog->ic_next) != log->l_iclog);
2684 * Completion of a iclog IO does not imply that a transaction has completed, as
2685 * transactions can be large enough to span many iclogs. We cannot change the
2686 * tail of the log half way through a transaction as this may be the only
2687 * transaction in the log and moving the tail to point to the middle of it
2688 * will prevent recovery from finding the start of the transaction. Hence we
2689 * should only update the last_sync_lsn if this iclog contains transaction
2690 * completion callbacks on it.
2692 * We have to do this before we drop the icloglock to ensure we are the only one
2693 * that can update it.
2695 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2696 * the reservation grant head pushing. This is due to the fact that the push
2697 * target is bound by the current last_sync_lsn value. Hence if we have a large
2698 * amount of log space bound up in this committing transaction then the
2699 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2700 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2701 * should push the AIL to ensure the push target (and hence the grant head) is
2702 * no longer bound by the old log head location and can move forwards and make
2706 xlog_state_set_callback(
2708 struct xlog_in_core *iclog,
2709 xfs_lsn_t header_lsn)
2711 trace_xlog_iclog_callback(iclog, _RET_IP_);
2712 iclog->ic_state = XLOG_STATE_CALLBACK;
2714 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2717 if (list_empty_careful(&iclog->ic_callbacks))
2720 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2721 xlog_grant_push_ail(log, 0);
2725 * Return true if we need to stop processing, false to continue to the next
2726 * iclog. The caller will need to run callbacks if the iclog is returned in the
2727 * XLOG_STATE_CALLBACK state.
2730 xlog_state_iodone_process_iclog(
2732 struct xlog_in_core *iclog)
2734 xfs_lsn_t lowest_lsn;
2735 xfs_lsn_t header_lsn;
2737 switch (iclog->ic_state) {
2738 case XLOG_STATE_ACTIVE:
2739 case XLOG_STATE_DIRTY:
2741 * Skip all iclogs in the ACTIVE & DIRTY states:
2744 case XLOG_STATE_DONE_SYNC:
2746 * Now that we have an iclog that is in the DONE_SYNC state, do
2747 * one more check here to see if we have chased our tail around.
2748 * If this is not the lowest lsn iclog, then we will leave it
2749 * for another completion to process.
2751 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2752 lowest_lsn = xlog_get_lowest_lsn(log);
2753 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2755 xlog_state_set_callback(log, iclog, header_lsn);
2759 * Can only perform callbacks in order. Since this iclog is not
2760 * in the DONE_SYNC state, we skip the rest and just try to
2768 * Loop over all the iclogs, running attached callbacks on them. Return true if
2769 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2770 * to handle transient shutdown state here at all because
2771 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2772 * cleanup of the callbacks.
2775 xlog_state_do_iclog_callbacks(
2777 __releases(&log->l_icloglock)
2778 __acquires(&log->l_icloglock)
2780 struct xlog_in_core *first_iclog = log->l_iclog;
2781 struct xlog_in_core *iclog = first_iclog;
2782 bool ran_callback = false;
2787 if (xlog_state_iodone_process_iclog(log, iclog))
2789 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2790 iclog = iclog->ic_next;
2793 list_splice_init(&iclog->ic_callbacks, &cb_list);
2794 spin_unlock(&log->l_icloglock);
2796 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2797 xlog_cil_process_committed(&cb_list);
2798 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2799 ran_callback = true;
2801 spin_lock(&log->l_icloglock);
2802 xlog_state_clean_iclog(log, iclog);
2803 iclog = iclog->ic_next;
2804 } while (iclog != first_iclog);
2806 return ran_callback;
2811 * Loop running iclog completion callbacks until there are no more iclogs in a
2812 * state that can run callbacks.
2815 xlog_state_do_callback(
2821 spin_lock(&log->l_icloglock);
2822 while (xlog_state_do_iclog_callbacks(log)) {
2823 if (xlog_is_shutdown(log))
2826 if (++repeats > 5000) {
2827 flushcnt += repeats;
2830 "%s: possible infinite loop (%d iterations)",
2831 __func__, flushcnt);
2835 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2836 wake_up_all(&log->l_flush_wait);
2838 spin_unlock(&log->l_icloglock);
2843 * Finish transitioning this iclog to the dirty state.
2845 * Callbacks could take time, so they are done outside the scope of the
2846 * global state machine log lock.
2849 xlog_state_done_syncing(
2850 struct xlog_in_core *iclog)
2852 struct xlog *log = iclog->ic_log;
2854 spin_lock(&log->l_icloglock);
2855 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2856 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2859 * If we got an error, either on the first buffer, or in the case of
2860 * split log writes, on the second, we shut down the file system and
2861 * no iclogs should ever be attempted to be written to disk again.
2863 if (!xlog_is_shutdown(log)) {
2864 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2865 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2869 * Someone could be sleeping prior to writing out the next
2870 * iclog buffer, we wake them all, one will get to do the
2871 * I/O, the others get to wait for the result.
2873 wake_up_all(&iclog->ic_write_wait);
2874 spin_unlock(&log->l_icloglock);
2875 xlog_state_do_callback(log);
2879 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2880 * sleep. We wait on the flush queue on the head iclog as that should be
2881 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2882 * we will wait here and all new writes will sleep until a sync completes.
2884 * The in-core logs are used in a circular fashion. They are not used
2885 * out-of-order even when an iclog past the head is free.
2888 * * log_offset where xlog_write() can start writing into the in-core
2890 * * in-core log pointer to which xlog_write() should write.
2891 * * boolean indicating this is a continued write to an in-core log.
2892 * If this is the last write, then the in-core log's offset field
2893 * needs to be incremented, depending on the amount of data which
2897 xlog_state_get_iclog_space(
2900 struct xlog_in_core **iclogp,
2901 struct xlog_ticket *ticket,
2905 xlog_rec_header_t *head;
2906 xlog_in_core_t *iclog;
2909 spin_lock(&log->l_icloglock);
2910 if (xlog_is_shutdown(log)) {
2911 spin_unlock(&log->l_icloglock);
2915 iclog = log->l_iclog;
2916 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2917 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2919 /* Wait for log writes to have flushed */
2920 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2924 head = &iclog->ic_header;
2926 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2927 log_offset = iclog->ic_offset;
2929 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2931 /* On the 1st write to an iclog, figure out lsn. This works
2932 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2933 * committing to. If the offset is set, that's how many blocks
2936 if (log_offset == 0) {
2937 ticket->t_curr_res -= log->l_iclog_hsize;
2938 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2939 head->h_lsn = cpu_to_be64(
2940 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2941 ASSERT(log->l_curr_block >= 0);
2944 /* If there is enough room to write everything, then do it. Otherwise,
2945 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2946 * bit is on, so this will get flushed out. Don't update ic_offset
2947 * until you know exactly how many bytes get copied. Therefore, wait
2948 * until later to update ic_offset.
2950 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2951 * can fit into remaining data section.
2953 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2956 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2959 * If we are the only one writing to this iclog, sync it to
2960 * disk. We need to do an atomic compare and decrement here to
2961 * avoid racing with concurrent atomic_dec_and_lock() calls in
2962 * xlog_state_release_iclog() when there is more than one
2963 * reference to the iclog.
2965 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2966 error = xlog_state_release_iclog(log, iclog, ticket);
2967 spin_unlock(&log->l_icloglock);
2973 /* Do we have enough room to write the full amount in the remainder
2974 * of this iclog? Or must we continue a write on the next iclog and
2975 * mark this iclog as completely taken? In the case where we switch
2976 * iclogs (to mark it taken), this particular iclog will release/sync
2977 * to disk in xlog_write().
2979 if (len <= iclog->ic_size - iclog->ic_offset)
2980 iclog->ic_offset += len;
2982 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2985 ASSERT(iclog->ic_offset <= iclog->ic_size);
2986 spin_unlock(&log->l_icloglock);
2988 *logoffsetp = log_offset;
2993 * The first cnt-1 times a ticket goes through here we don't need to move the
2994 * grant write head because the permanent reservation has reserved cnt times the
2995 * unit amount. Release part of current permanent unit reservation and reset
2996 * current reservation to be one units worth. Also move grant reservation head
3000 xfs_log_ticket_regrant(
3002 struct xlog_ticket *ticket)
3004 trace_xfs_log_ticket_regrant(log, ticket);
3006 if (ticket->t_cnt > 0)
3009 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3010 ticket->t_curr_res);
3011 xlog_grant_sub_space(log, &log->l_write_head.grant,
3012 ticket->t_curr_res);
3013 ticket->t_curr_res = ticket->t_unit_res;
3015 trace_xfs_log_ticket_regrant_sub(log, ticket);
3017 /* just return if we still have some of the pre-reserved space */
3018 if (!ticket->t_cnt) {
3019 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3020 ticket->t_unit_res);
3021 trace_xfs_log_ticket_regrant_exit(log, ticket);
3023 ticket->t_curr_res = ticket->t_unit_res;
3026 xfs_log_ticket_put(ticket);
3030 * Give back the space left from a reservation.
3032 * All the information we need to make a correct determination of space left
3033 * is present. For non-permanent reservations, things are quite easy. The
3034 * count should have been decremented to zero. We only need to deal with the
3035 * space remaining in the current reservation part of the ticket. If the
3036 * ticket contains a permanent reservation, there may be left over space which
3037 * needs to be released. A count of N means that N-1 refills of the current
3038 * reservation can be done before we need to ask for more space. The first
3039 * one goes to fill up the first current reservation. Once we run out of
3040 * space, the count will stay at zero and the only space remaining will be
3041 * in the current reservation field.
3044 xfs_log_ticket_ungrant(
3046 struct xlog_ticket *ticket)
3050 trace_xfs_log_ticket_ungrant(log, ticket);
3052 if (ticket->t_cnt > 0)
3055 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3058 * If this is a permanent reservation ticket, we may be able to free
3059 * up more space based on the remaining count.
3061 bytes = ticket->t_curr_res;
3062 if (ticket->t_cnt > 0) {
3063 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3064 bytes += ticket->t_unit_res*ticket->t_cnt;
3067 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3068 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3070 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3072 xfs_log_space_wake(log->l_mp);
3073 xfs_log_ticket_put(ticket);
3077 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3078 * the current iclog pointer to the next iclog in the ring.
3081 xlog_state_switch_iclogs(
3083 struct xlog_in_core *iclog,
3086 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3087 assert_spin_locked(&log->l_icloglock);
3088 trace_xlog_iclog_switch(iclog, _RET_IP_);
3091 eventual_size = iclog->ic_offset;
3092 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3093 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3094 log->l_prev_block = log->l_curr_block;
3095 log->l_prev_cycle = log->l_curr_cycle;
3097 /* roll log?: ic_offset changed later */
3098 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3100 /* Round up to next log-sunit */
3101 if (log->l_iclog_roundoff > BBSIZE) {
3102 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3103 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3106 if (log->l_curr_block >= log->l_logBBsize) {
3108 * Rewind the current block before the cycle is bumped to make
3109 * sure that the combined LSN never transiently moves forward
3110 * when the log wraps to the next cycle. This is to support the
3111 * unlocked sample of these fields from xlog_valid_lsn(). Most
3112 * other cases should acquire l_icloglock.
3114 log->l_curr_block -= log->l_logBBsize;
3115 ASSERT(log->l_curr_block >= 0);
3117 log->l_curr_cycle++;
3118 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3119 log->l_curr_cycle++;
3121 ASSERT(iclog == log->l_iclog);
3122 log->l_iclog = iclog->ic_next;
3126 * Force the iclog to disk and check if the iclog has been completed before
3127 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3128 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3129 * If completion has already occurred, tell the caller so that it can avoid an
3130 * unnecessary wait on the iclog.
3133 xlog_force_and_check_iclog(
3134 struct xlog_in_core *iclog,
3137 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3141 error = xlog_force_iclog(iclog);
3146 * If the iclog has already been completed and reused the header LSN
3147 * will have been rewritten by completion
3149 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3155 * Write out all data in the in-core log as of this exact moment in time.
3157 * Data may be written to the in-core log during this call. However,
3158 * we don't guarantee this data will be written out. A change from past
3159 * implementation means this routine will *not* write out zero length LRs.
3161 * Basically, we try and perform an intelligent scan of the in-core logs.
3162 * If we determine there is no flushable data, we just return. There is no
3163 * flushable data if:
3165 * 1. the current iclog is active and has no data; the previous iclog
3166 * is in the active or dirty state.
3167 * 2. the current iclog is drity, and the previous iclog is in the
3168 * active or dirty state.
3172 * 1. the current iclog is not in the active nor dirty state.
3173 * 2. the current iclog dirty, and the previous iclog is not in the
3174 * active nor dirty state.
3175 * 3. the current iclog is active, and there is another thread writing
3176 * to this particular iclog.
3177 * 4. a) the current iclog is active and has no other writers
3178 * b) when we return from flushing out this iclog, it is still
3179 * not in the active nor dirty state.
3183 struct xfs_mount *mp,
3186 struct xlog *log = mp->m_log;
3187 struct xlog_in_core *iclog;
3189 XFS_STATS_INC(mp, xs_log_force);
3190 trace_xfs_log_force(mp, 0, _RET_IP_);
3192 xlog_cil_force(log);
3194 spin_lock(&log->l_icloglock);
3195 if (xlog_is_shutdown(log))
3198 iclog = log->l_iclog;
3199 trace_xlog_iclog_force(iclog, _RET_IP_);
3201 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3202 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3203 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3205 * If the head is dirty or (active and empty), then we need to
3206 * look at the previous iclog.
3208 * If the previous iclog is active or dirty we are done. There
3209 * is nothing to sync out. Otherwise, we attach ourselves to the
3210 * previous iclog and go to sleep.
3212 iclog = iclog->ic_prev;
3213 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3214 if (atomic_read(&iclog->ic_refcnt) == 0) {
3215 /* We have exclusive access to this iclog. */
3218 if (xlog_force_and_check_iclog(iclog, &completed))
3225 * Someone else is still writing to this iclog, so we
3226 * need to ensure that when they release the iclog it
3227 * gets synced immediately as we may be waiting on it.
3229 xlog_state_switch_iclogs(log, iclog, 0);
3234 * The iclog we are about to wait on may contain the checkpoint pushed
3235 * by the above xlog_cil_force() call, but it may not have been pushed
3236 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3237 * are flushed when this iclog is written.
3239 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3240 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3242 if (flags & XFS_LOG_SYNC)
3243 return xlog_wait_on_iclog(iclog);
3245 spin_unlock(&log->l_icloglock);
3248 spin_unlock(&log->l_icloglock);
3253 * Force the log to a specific LSN.
3255 * If an iclog with that lsn can be found:
3256 * If it is in the DIRTY state, just return.
3257 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3258 * state and go to sleep or return.
3259 * If it is in any other state, go to sleep or return.
3261 * Synchronous forces are implemented with a wait queue. All callers trying
3262 * to force a given lsn to disk must wait on the queue attached to the
3263 * specific in-core log. When given in-core log finally completes its write
3264 * to disk, that thread will wake up all threads waiting on the queue.
3274 struct xlog_in_core *iclog;
3277 spin_lock(&log->l_icloglock);
3278 if (xlog_is_shutdown(log))
3281 iclog = log->l_iclog;
3282 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3283 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3284 iclog = iclog->ic_next;
3285 if (iclog == log->l_iclog)
3289 switch (iclog->ic_state) {
3290 case XLOG_STATE_ACTIVE:
3292 * We sleep here if we haven't already slept (e.g. this is the
3293 * first time we've looked at the correct iclog buf) and the
3294 * buffer before us is going to be sync'ed. The reason for this
3295 * is that if we are doing sync transactions here, by waiting
3296 * for the previous I/O to complete, we can allow a few more
3297 * transactions into this iclog before we close it down.
3299 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3300 * refcnt so we can release the log (which drops the ref count).
3301 * The state switch keeps new transaction commits from using
3302 * this buffer. When the current commits finish writing into
3303 * the buffer, the refcount will drop to zero and the buffer
3306 if (!already_slept &&
3307 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3308 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3309 xlog_wait(&iclog->ic_prev->ic_write_wait,
3313 if (xlog_force_and_check_iclog(iclog, &completed))
3320 case XLOG_STATE_WANT_SYNC:
3322 * This iclog may contain the checkpoint pushed by the
3323 * xlog_cil_force_seq() call, but there are other writers still
3324 * accessing it so it hasn't been pushed to disk yet. Like the
3325 * ACTIVE case above, we need to make sure caches are flushed
3326 * when this iclog is written.
3328 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3332 * The entire checkpoint was written by the CIL force and is on
3333 * its way to disk already. It will be stable when it
3334 * completes, so we don't need to manipulate caches here at all.
3335 * We just need to wait for completion if necessary.
3340 if (flags & XFS_LOG_SYNC)
3341 return xlog_wait_on_iclog(iclog);
3343 spin_unlock(&log->l_icloglock);
3346 spin_unlock(&log->l_icloglock);
3351 * Force the log to a specific checkpoint sequence.
3353 * First force the CIL so that all the required changes have been flushed to the
3354 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3355 * the iclog that needs to be flushed to stable storage. If the caller needs
3356 * a synchronous log force, we will wait on the iclog with the LSN returned by
3357 * xlog_cil_force_seq() to be completed.
3361 struct xfs_mount *mp,
3366 struct xlog *log = mp->m_log;
3371 XFS_STATS_INC(mp, xs_log_force);
3372 trace_xfs_log_force(mp, seq, _RET_IP_);
3374 lsn = xlog_cil_force_seq(log, seq);
3375 if (lsn == NULLCOMMITLSN)
3378 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3379 if (ret == -EAGAIN) {
3380 XFS_STATS_INC(mp, xs_log_force_sleep);
3381 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3387 * Free a used ticket when its refcount falls to zero.
3391 xlog_ticket_t *ticket)
3393 ASSERT(atomic_read(&ticket->t_ref) > 0);
3394 if (atomic_dec_and_test(&ticket->t_ref))
3395 kmem_cache_free(xfs_log_ticket_cache, ticket);
3400 xlog_ticket_t *ticket)
3402 ASSERT(atomic_read(&ticket->t_ref) > 0);
3403 atomic_inc(&ticket->t_ref);
3408 * Figure out the total log space unit (in bytes) that would be
3409 * required for a log ticket.
3421 * Permanent reservations have up to 'cnt'-1 active log operations
3422 * in the log. A unit in this case is the amount of space for one
3423 * of these log operations. Normal reservations have a cnt of 1
3424 * and their unit amount is the total amount of space required.
3426 * The following lines of code account for non-transaction data
3427 * which occupy space in the on-disk log.
3429 * Normal form of a transaction is:
3430 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3431 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3433 * We need to account for all the leadup data and trailer data
3434 * around the transaction data.
3435 * And then we need to account for the worst case in terms of using
3437 * The worst case will happen if:
3438 * - the placement of the transaction happens to be such that the
3439 * roundoff is at its maximum
3440 * - the transaction data is synced before the commit record is synced
3441 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3442 * Therefore the commit record is in its own Log Record.
3443 * This can happen as the commit record is called with its
3444 * own region to xlog_write().
3445 * This then means that in the worst case, roundoff can happen for
3446 * the commit-rec as well.
3447 * The commit-rec is smaller than padding in this scenario and so it is
3448 * not added separately.
3451 /* for trans header */
3452 unit_bytes += sizeof(xlog_op_header_t);
3453 unit_bytes += sizeof(xfs_trans_header_t);
3456 unit_bytes += sizeof(xlog_op_header_t);
3459 * for LR headers - the space for data in an iclog is the size minus
3460 * the space used for the headers. If we use the iclog size, then we
3461 * undercalculate the number of headers required.
3463 * Furthermore - the addition of op headers for split-recs might
3464 * increase the space required enough to require more log and op
3465 * headers, so take that into account too.
3467 * IMPORTANT: This reservation makes the assumption that if this
3468 * transaction is the first in an iclog and hence has the LR headers
3469 * accounted to it, then the remaining space in the iclog is
3470 * exclusively for this transaction. i.e. if the transaction is larger
3471 * than the iclog, it will be the only thing in that iclog.
3472 * Fundamentally, this means we must pass the entire log vector to
3473 * xlog_write to guarantee this.
3475 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3476 num_headers = howmany(unit_bytes, iclog_space);
3478 /* for split-recs - ophdrs added when data split over LRs */
3479 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3481 /* add extra header reservations if we overrun */
3482 while (!num_headers ||
3483 howmany(unit_bytes, iclog_space) > num_headers) {
3484 unit_bytes += sizeof(xlog_op_header_t);
3487 unit_bytes += log->l_iclog_hsize * num_headers;
3489 /* for commit-rec LR header - note: padding will subsume the ophdr */
3490 unit_bytes += log->l_iclog_hsize;
3492 /* roundoff padding for transaction data and one for commit record */
3493 unit_bytes += 2 * log->l_iclog_roundoff;
3496 *niclogs = num_headers;
3501 xfs_log_calc_unit_res(
3502 struct xfs_mount *mp,
3505 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3509 * Allocate and initialise a new log ticket.
3511 struct xlog_ticket *
3518 struct xlog_ticket *tic;
3521 tic = kmem_cache_zalloc(xfs_log_ticket_cache,
3522 GFP_KERNEL | __GFP_NOFAIL);
3524 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3526 atomic_set(&tic->t_ref, 1);
3527 tic->t_task = current;
3528 INIT_LIST_HEAD(&tic->t_queue);
3529 tic->t_unit_res = unit_res;
3530 tic->t_curr_res = unit_res;
3533 tic->t_tid = get_random_u32();
3535 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3542 * Check to make sure the grant write head didn't just over lap the tail. If
3543 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3544 * the cycles differ by exactly one and check the byte count.
3546 * This check is run unlocked, so can give false positives. Rather than assert
3547 * on failures, use a warn-once flag and a panic tag to allow the admin to
3548 * determine if they want to panic the machine when such an error occurs. For
3549 * debug kernels this will have the same effect as using an assert but, unlinke
3550 * an assert, it can be turned off at runtime.
3553 xlog_verify_grant_tail(
3556 int tail_cycle, tail_blocks;
3559 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3560 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3561 if (tail_cycle != cycle) {
3562 if (cycle - 1 != tail_cycle &&
3563 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3564 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3565 "%s: cycle - 1 != tail_cycle", __func__);
3568 if (space > BBTOB(tail_blocks) &&
3569 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3570 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3571 "%s: space > BBTOB(tail_blocks)", __func__);
3576 /* check if it will fit */
3578 xlog_verify_tail_lsn(
3580 struct xlog_in_core *iclog)
3582 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3585 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3587 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3588 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3589 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3591 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3593 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3594 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3596 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3597 if (blocks < BTOBB(iclog->ic_offset) + 1)
3598 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3603 * Perform a number of checks on the iclog before writing to disk.
3605 * 1. Make sure the iclogs are still circular
3606 * 2. Make sure we have a good magic number
3607 * 3. Make sure we don't have magic numbers in the data
3608 * 4. Check fields of each log operation header for:
3609 * A. Valid client identifier
3610 * B. tid ptr value falls in valid ptr space (user space code)
3611 * C. Length in log record header is correct according to the
3612 * individual operation headers within record.
3613 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3614 * log, check the preceding blocks of the physical log to make sure all
3615 * the cycle numbers agree with the current cycle number.
3620 struct xlog_in_core *iclog,
3623 xlog_op_header_t *ophead;
3624 xlog_in_core_t *icptr;
3625 xlog_in_core_2_t *xhdr;
3626 void *base_ptr, *ptr, *p;
3627 ptrdiff_t field_offset;
3629 int len, i, j, k, op_len;
3632 /* check validity of iclog pointers */
3633 spin_lock(&log->l_icloglock);
3634 icptr = log->l_iclog;
3635 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3638 if (icptr != log->l_iclog)
3639 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3640 spin_unlock(&log->l_icloglock);
3642 /* check log magic numbers */
3643 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3644 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3646 base_ptr = ptr = &iclog->ic_header;
3647 p = &iclog->ic_header;
3648 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3649 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3650 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3655 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3656 base_ptr = ptr = iclog->ic_datap;
3658 xhdr = iclog->ic_data;
3659 for (i = 0; i < len; i++) {
3662 /* clientid is only 1 byte */
3663 p = &ophead->oh_clientid;
3664 field_offset = p - base_ptr;
3665 if (field_offset & 0x1ff) {
3666 clientid = ophead->oh_clientid;
3668 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3669 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3670 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3671 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3672 clientid = xlog_get_client_id(
3673 xhdr[j].hic_xheader.xh_cycle_data[k]);
3675 clientid = xlog_get_client_id(
3676 iclog->ic_header.h_cycle_data[idx]);
3679 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3681 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3682 __func__, i, clientid, ophead,
3683 (unsigned long)field_offset);
3687 p = &ophead->oh_len;
3688 field_offset = p - base_ptr;
3689 if (field_offset & 0x1ff) {
3690 op_len = be32_to_cpu(ophead->oh_len);
3692 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3693 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3694 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3695 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3696 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3698 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3701 ptr += sizeof(xlog_op_header_t) + op_len;
3707 * Perform a forced shutdown on the log.
3709 * This can be called from low level log code to trigger a shutdown, or from the
3710 * high level mount shutdown code when the mount shuts down.
3712 * Our main objectives here are to make sure that:
3713 * a. if the shutdown was not due to a log IO error, flush the logs to
3714 * disk. Anything modified after this is ignored.
3715 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3716 * parties to find out. Nothing new gets queued after this is done.
3717 * c. Tasks sleeping on log reservations, pinned objects and
3718 * other resources get woken up.
3719 * d. The mount is also marked as shut down so that log triggered shutdowns
3720 * still behave the same as if they called xfs_forced_shutdown().
3722 * Return true if the shutdown cause was a log IO error and we actually shut the
3726 xlog_force_shutdown(
3728 uint32_t shutdown_flags)
3730 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3736 * Flush all the completed transactions to disk before marking the log
3737 * being shut down. We need to do this first as shutting down the log
3738 * before the force will prevent the log force from flushing the iclogs
3741 * When we are in recovery, there are no transactions to flush, and
3742 * we don't want to touch the log because we don't want to perturb the
3743 * current head/tail for future recovery attempts. Hence we need to
3744 * avoid a log force in this case.
3746 * If we are shutting down due to a log IO error, then we must avoid
3747 * trying to write the log as that may just result in more IO errors and
3748 * an endless shutdown/force loop.
3750 if (!log_error && !xlog_in_recovery(log))
3751 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3754 * Atomically set the shutdown state. If the shutdown state is already
3755 * set, there someone else is performing the shutdown and so we are done
3756 * here. This should never happen because we should only ever get called
3757 * once by the first shutdown caller.
3759 * Much of the log state machine transitions assume that shutdown state
3760 * cannot change once they hold the log->l_icloglock. Hence we need to
3761 * hold that lock here, even though we use the atomic test_and_set_bit()
3762 * operation to set the shutdown state.
3764 spin_lock(&log->l_icloglock);
3765 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3766 spin_unlock(&log->l_icloglock);
3769 spin_unlock(&log->l_icloglock);
3772 * If this log shutdown also sets the mount shutdown state, issue a
3773 * shutdown warning message.
3775 if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3776 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3777 "Filesystem has been shut down due to log error (0x%x).",
3779 xfs_alert(log->l_mp,
3780 "Please unmount the filesystem and rectify the problem(s).");
3781 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3786 * We don't want anybody waiting for log reservations after this. That
3787 * means we have to wake up everybody queued up on reserveq as well as
3788 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3789 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3790 * action is protected by the grant locks.
3792 xlog_grant_head_wake_all(&log->l_reserve_head);
3793 xlog_grant_head_wake_all(&log->l_write_head);
3796 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3797 * as if the log writes were completed. The abort handling in the log
3798 * item committed callback functions will do this again under lock to
3801 spin_lock(&log->l_cilp->xc_push_lock);
3802 wake_up_all(&log->l_cilp->xc_start_wait);
3803 wake_up_all(&log->l_cilp->xc_commit_wait);
3804 spin_unlock(&log->l_cilp->xc_push_lock);
3806 spin_lock(&log->l_icloglock);
3807 xlog_state_shutdown_callbacks(log);
3808 spin_unlock(&log->l_icloglock);
3810 wake_up_var(&log->l_opstate);
3818 xlog_in_core_t *iclog;
3820 iclog = log->l_iclog;
3822 /* endianness does not matter here, zero is zero in
3825 if (iclog->ic_header.h_num_logops)
3827 iclog = iclog->ic_next;
3828 } while (iclog != log->l_iclog);
3833 * Verify that an LSN stamped into a piece of metadata is valid. This is
3834 * intended for use in read verifiers on v5 superblocks.
3838 struct xfs_mount *mp,
3841 struct xlog *log = mp->m_log;
3845 * norecovery mode skips mount-time log processing and unconditionally
3846 * resets the in-core LSN. We can't validate in this mode, but
3847 * modifications are not allowed anyways so just return true.
3849 if (xfs_has_norecovery(mp))
3853 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3854 * handled by recovery and thus safe to ignore here.
3856 if (lsn == NULLCOMMITLSN)
3859 valid = xlog_valid_lsn(mp->m_log, lsn);
3861 /* warn the user about what's gone wrong before verifier failure */
3863 spin_lock(&log->l_icloglock);
3865 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3866 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3867 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3868 log->l_curr_cycle, log->l_curr_block);
3869 spin_unlock(&log->l_icloglock);
3876 * Notify the log that we're about to start using a feature that is protected
3877 * by a log incompat feature flag. This will prevent log covering from
3878 * clearing those flags.
3881 xlog_use_incompat_feat(
3884 down_read(&log->l_incompat_users);
3887 /* Notify the log that we've finished using log incompat features. */
3889 xlog_drop_incompat_feat(
3892 up_read(&log->l_incompat_users);