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 xfs_buftarg_t *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)) {
719 * log recovery ignores readonly state and so we need to clear
720 * mount-based read only state so it can write to disk.
722 bool readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
724 error = xlog_recover(log);
726 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
728 xfs_warn(mp, "log mount/recovery failed: error %d",
730 xlog_recover_cancel(log);
731 goto out_destroy_ail;
735 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
738 goto out_destroy_ail;
740 /* Normal transactions can now occur */
741 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
744 * Now the log has been fully initialised and we know were our
745 * space grant counters are, we can initialise the permanent ticket
746 * needed for delayed logging to work.
748 xlog_cil_init_post_recovery(log);
753 xfs_trans_ail_destroy(mp);
755 xlog_dealloc_log(log);
761 * Finish the recovery of the file system. This is separate from the
762 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
763 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
766 * If we finish recovery successfully, start the background log work. If we are
767 * not doing recovery, then we have a RO filesystem and we don't need to start
771 xfs_log_mount_finish(
772 struct xfs_mount *mp)
774 struct xlog *log = mp->m_log;
778 if (xfs_has_norecovery(mp)) {
779 ASSERT(xfs_is_readonly(mp));
784 * log recovery ignores readonly state and so we need to clear
785 * mount-based read only state so it can write to disk.
787 readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
790 * During the second phase of log recovery, we need iget and
791 * iput to behave like they do for an active filesystem.
792 * xfs_fs_drop_inode needs to be able to prevent the deletion
793 * of inodes before we're done replaying log items on those
794 * inodes. Turn it off immediately after recovery finishes
795 * so that we don't leak the quota inodes if subsequent mount
798 * We let all inodes involved in redo item processing end up on
799 * the LRU instead of being evicted immediately so that if we do
800 * something to an unlinked inode, the irele won't cause
801 * premature truncation and freeing of the inode, which results
802 * in log recovery failure. We have to evict the unreferenced
803 * lru inodes after clearing SB_ACTIVE because we don't
804 * otherwise clean up the lru if there's a subsequent failure in
805 * xfs_mountfs, which leads to us leaking the inodes if nothing
806 * else (e.g. quotacheck) references the inodes before the
807 * mount failure occurs.
809 mp->m_super->s_flags |= SB_ACTIVE;
810 xfs_log_work_queue(mp);
811 if (xlog_recovery_needed(log))
812 error = xlog_recover_finish(log);
813 mp->m_super->s_flags &= ~SB_ACTIVE;
814 evict_inodes(mp->m_super);
817 * Drain the buffer LRU after log recovery. This is required for v4
818 * filesystems to avoid leaving around buffers with NULL verifier ops,
819 * but we do it unconditionally to make sure we're always in a clean
820 * cache state after mount.
822 * Don't push in the error case because the AIL may have pending intents
823 * that aren't removed until recovery is cancelled.
825 if (xlog_recovery_needed(log)) {
827 xfs_log_force(mp, XFS_LOG_SYNC);
828 xfs_ail_push_all_sync(mp->m_ail);
830 xfs_notice(mp, "Ending recovery (logdev: %s)",
831 mp->m_logname ? mp->m_logname : "internal");
833 xfs_info(mp, "Ending clean mount");
835 xfs_buftarg_drain(mp->m_ddev_targp);
837 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
839 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
841 /* Make sure the log is dead if we're returning failure. */
842 ASSERT(!error || xlog_is_shutdown(log));
848 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
852 xfs_log_mount_cancel(
853 struct xfs_mount *mp)
855 xlog_recover_cancel(mp->m_log);
860 * Flush out the iclog to disk ensuring that device caches are flushed and
861 * the iclog hits stable storage before any completion waiters are woken.
865 struct xlog_in_core *iclog)
867 atomic_inc(&iclog->ic_refcnt);
868 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
869 if (iclog->ic_state == XLOG_STATE_ACTIVE)
870 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
871 return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
875 * Cycle all the iclogbuf locks to make sure all log IO completion
876 * is done before we tear down these buffers.
879 xlog_wait_iclog_completion(struct xlog *log)
882 struct xlog_in_core *iclog = log->l_iclog;
884 for (i = 0; i < log->l_iclog_bufs; i++) {
885 down(&iclog->ic_sema);
887 iclog = iclog->ic_next;
892 * Wait for the iclog and all prior iclogs to be written disk as required by the
893 * log force state machine. Waiting on ic_force_wait ensures iclog completions
894 * have been ordered and callbacks run before we are woken here, hence
895 * guaranteeing that all the iclogs up to this one are on stable storage.
899 struct xlog_in_core *iclog)
900 __releases(iclog->ic_log->l_icloglock)
902 struct xlog *log = iclog->ic_log;
904 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
905 if (!xlog_is_shutdown(log) &&
906 iclog->ic_state != XLOG_STATE_ACTIVE &&
907 iclog->ic_state != XLOG_STATE_DIRTY) {
908 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
909 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
911 spin_unlock(&log->l_icloglock);
914 if (xlog_is_shutdown(log))
920 * Write out an unmount record using the ticket provided. We have to account for
921 * the data space used in the unmount ticket as this write is not done from a
922 * transaction context that has already done the accounting for us.
925 xlog_write_unmount_record(
927 struct xlog_ticket *ticket)
930 struct xlog_op_header ophdr;
931 struct xfs_unmount_log_format ulf;
934 .oh_clientid = XFS_LOG,
935 .oh_tid = cpu_to_be32(ticket->t_tid),
936 .oh_flags = XLOG_UNMOUNT_TRANS,
939 .magic = XLOG_UNMOUNT_TYPE,
942 struct xfs_log_iovec reg = {
943 .i_addr = &unmount_rec,
944 .i_len = sizeof(unmount_rec),
945 .i_type = XLOG_REG_TYPE_UNMOUNT,
947 struct xfs_log_vec vec = {
952 list_add(&vec.lv_list, &lv_chain);
954 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
955 sizeof(struct xfs_unmount_log_format)) !=
956 sizeof(unmount_rec));
958 /* account for space used by record data */
959 ticket->t_curr_res -= sizeof(unmount_rec);
961 return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
965 * Mark the filesystem clean by writing an unmount record to the head of the
972 struct xfs_mount *mp = log->l_mp;
973 struct xlog_in_core *iclog;
974 struct xlog_ticket *tic = NULL;
977 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
981 error = xlog_write_unmount_record(log, tic);
983 * At this point, we're umounting anyway, so there's no point in
984 * transitioning log state to shutdown. Just continue...
988 xfs_alert(mp, "%s: unmount record failed", __func__);
990 spin_lock(&log->l_icloglock);
991 iclog = log->l_iclog;
992 error = xlog_force_iclog(iclog);
993 xlog_wait_on_iclog(iclog);
996 trace_xfs_log_umount_write(log, tic);
997 xfs_log_ticket_ungrant(log, tic);
1002 xfs_log_unmount_verify_iclog(
1005 struct xlog_in_core *iclog = log->l_iclog;
1008 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
1009 ASSERT(iclog->ic_offset == 0);
1010 } while ((iclog = iclog->ic_next) != log->l_iclog);
1014 * Unmount record used to have a string "Unmount filesystem--" in the
1015 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
1016 * We just write the magic number now since that particular field isn't
1017 * currently architecture converted and "Unmount" is a bit foo.
1018 * As far as I know, there weren't any dependencies on the old behaviour.
1021 xfs_log_unmount_write(
1022 struct xfs_mount *mp)
1024 struct xlog *log = mp->m_log;
1026 if (!xfs_log_writable(mp))
1029 xfs_log_force(mp, XFS_LOG_SYNC);
1031 if (xlog_is_shutdown(log))
1035 * If we think the summary counters are bad, avoid writing the unmount
1036 * record to force log recovery at next mount, after which the summary
1037 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1040 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1041 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1042 xfs_alert(mp, "%s: will fix summary counters at next mount",
1047 xfs_log_unmount_verify_iclog(log);
1048 xlog_unmount_write(log);
1052 * Empty the log for unmount/freeze.
1054 * To do this, we first need to shut down the background log work so it is not
1055 * trying to cover the log as we clean up. We then need to unpin all objects in
1056 * the log so we can then flush them out. Once they have completed their IO and
1057 * run the callbacks removing themselves from the AIL, we can cover the log.
1061 struct xfs_mount *mp)
1064 * Clear log incompat features since we're quiescing the log. Report
1065 * failures, though it's not fatal to have a higher log feature
1066 * protection level than the log contents actually require.
1068 if (xfs_clear_incompat_log_features(mp)) {
1071 error = xfs_sync_sb(mp, false);
1074 "Failed to clear log incompat features on quiesce");
1077 cancel_delayed_work_sync(&mp->m_log->l_work);
1078 xfs_log_force(mp, XFS_LOG_SYNC);
1081 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1082 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1083 * xfs_buf_iowait() cannot be used because it was pushed with the
1084 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1085 * the IO to complete.
1087 xfs_ail_push_all_sync(mp->m_ail);
1088 xfs_buftarg_wait(mp->m_ddev_targp);
1089 xfs_buf_lock(mp->m_sb_bp);
1090 xfs_buf_unlock(mp->m_sb_bp);
1092 return xfs_log_cover(mp);
1097 struct xfs_mount *mp)
1099 xfs_log_quiesce(mp);
1100 xfs_log_unmount_write(mp);
1104 * Shut down and release the AIL and Log.
1106 * During unmount, we need to ensure we flush all the dirty metadata objects
1107 * from the AIL so that the log is empty before we write the unmount record to
1108 * the log. Once this is done, we can tear down the AIL and the log.
1112 struct xfs_mount *mp)
1117 * If shutdown has come from iclog IO context, the log
1118 * cleaning will have been skipped and so we need to wait
1119 * for the iclog to complete shutdown processing before we
1120 * tear anything down.
1122 xlog_wait_iclog_completion(mp->m_log);
1124 xfs_buftarg_drain(mp->m_ddev_targp);
1126 xfs_trans_ail_destroy(mp);
1128 xfs_sysfs_del(&mp->m_log->l_kobj);
1130 xlog_dealloc_log(mp->m_log);
1135 struct xfs_mount *mp,
1136 struct xfs_log_item *item,
1138 const struct xfs_item_ops *ops)
1140 item->li_log = mp->m_log;
1141 item->li_ailp = mp->m_ail;
1142 item->li_type = type;
1146 INIT_LIST_HEAD(&item->li_ail);
1147 INIT_LIST_HEAD(&item->li_cil);
1148 INIT_LIST_HEAD(&item->li_bio_list);
1149 INIT_LIST_HEAD(&item->li_trans);
1153 * Wake up processes waiting for log space after we have moved the log tail.
1157 struct xfs_mount *mp)
1159 struct xlog *log = mp->m_log;
1162 if (xlog_is_shutdown(log))
1165 if (!list_empty_careful(&log->l_write_head.waiters)) {
1166 ASSERT(!xlog_in_recovery(log));
1168 spin_lock(&log->l_write_head.lock);
1169 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1170 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1171 spin_unlock(&log->l_write_head.lock);
1174 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1175 ASSERT(!xlog_in_recovery(log));
1177 spin_lock(&log->l_reserve_head.lock);
1178 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1179 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1180 spin_unlock(&log->l_reserve_head.lock);
1185 * Determine if we have a transaction that has gone to disk that needs to be
1186 * covered. To begin the transition to the idle state firstly the log needs to
1187 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1188 * we start attempting to cover the log.
1190 * Only if we are then in a state where covering is needed, the caller is
1191 * informed that dummy transactions are required to move the log into the idle
1194 * If there are any items in the AIl or CIL, then we do not want to attempt to
1195 * cover the log as we may be in a situation where there isn't log space
1196 * available to run a dummy transaction and this can lead to deadlocks when the
1197 * tail of the log is pinned by an item that is modified in the CIL. Hence
1198 * there's no point in running a dummy transaction at this point because we
1199 * can't start trying to idle the log until both the CIL and AIL are empty.
1202 xfs_log_need_covered(
1203 struct xfs_mount *mp)
1205 struct xlog *log = mp->m_log;
1206 bool needed = false;
1208 if (!xlog_cil_empty(log))
1211 spin_lock(&log->l_icloglock);
1212 switch (log->l_covered_state) {
1213 case XLOG_STATE_COVER_DONE:
1214 case XLOG_STATE_COVER_DONE2:
1215 case XLOG_STATE_COVER_IDLE:
1217 case XLOG_STATE_COVER_NEED:
1218 case XLOG_STATE_COVER_NEED2:
1219 if (xfs_ail_min_lsn(log->l_ailp))
1221 if (!xlog_iclogs_empty(log))
1225 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1226 log->l_covered_state = XLOG_STATE_COVER_DONE;
1228 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1234 spin_unlock(&log->l_icloglock);
1239 * Explicitly cover the log. This is similar to background log covering but
1240 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1241 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1242 * must all be empty.
1246 struct xfs_mount *mp)
1251 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1252 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1253 xlog_is_shutdown(mp->m_log));
1255 if (!xfs_log_writable(mp))
1259 * xfs_log_need_covered() is not idempotent because it progresses the
1260 * state machine if the log requires covering. Therefore, we must call
1261 * this function once and use the result until we've issued an sb sync.
1262 * Do so first to make that abundantly clear.
1264 * Fall into the covering sequence if the log needs covering or the
1265 * mount has lazy superblock accounting to sync to disk. The sb sync
1266 * used for covering accumulates the in-core counters, so covering
1267 * handles this for us.
1269 need_covered = xfs_log_need_covered(mp);
1270 if (!need_covered && !xfs_has_lazysbcount(mp))
1274 * To cover the log, commit the superblock twice (at most) in
1275 * independent checkpoints. The first serves as a reference for the
1276 * tail pointer. The sync transaction and AIL push empties the AIL and
1277 * updates the in-core tail to the LSN of the first checkpoint. The
1278 * second commit updates the on-disk tail with the in-core LSN,
1279 * covering the log. Push the AIL one more time to leave it empty, as
1283 error = xfs_sync_sb(mp, true);
1286 xfs_ail_push_all_sync(mp->m_ail);
1287 } while (xfs_log_need_covered(mp));
1293 * We may be holding the log iclog lock upon entering this routine.
1296 xlog_assign_tail_lsn_locked(
1297 struct xfs_mount *mp)
1299 struct xlog *log = mp->m_log;
1300 struct xfs_log_item *lip;
1303 assert_spin_locked(&mp->m_ail->ail_lock);
1306 * To make sure we always have a valid LSN for the log tail we keep
1307 * track of the last LSN which was committed in log->l_last_sync_lsn,
1308 * and use that when the AIL was empty.
1310 lip = xfs_ail_min(mp->m_ail);
1312 tail_lsn = lip->li_lsn;
1314 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1315 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1316 atomic64_set(&log->l_tail_lsn, tail_lsn);
1321 xlog_assign_tail_lsn(
1322 struct xfs_mount *mp)
1326 spin_lock(&mp->m_ail->ail_lock);
1327 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1328 spin_unlock(&mp->m_ail->ail_lock);
1334 * Return the space in the log between the tail and the head. The head
1335 * is passed in the cycle/bytes formal parms. In the special case where
1336 * the reserve head has wrapped passed the tail, this calculation is no
1337 * longer valid. In this case, just return 0 which means there is no space
1338 * in the log. This works for all places where this function is called
1339 * with the reserve head. Of course, if the write head were to ever
1340 * wrap the tail, we should blow up. Rather than catch this case here,
1341 * we depend on other ASSERTions in other parts of the code. XXXmiken
1343 * If reservation head is behind the tail, we have a problem. Warn about it,
1344 * but then treat it as if the log is empty.
1346 * If the log is shut down, the head and tail may be invalid or out of whack, so
1347 * shortcut invalidity asserts in this case so that we don't trigger them
1360 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1361 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1362 tail_bytes = BBTOB(tail_bytes);
1363 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1364 return log->l_logsize - (head_bytes - tail_bytes);
1365 if (tail_cycle + 1 < head_cycle)
1368 /* Ignore potential inconsistency when shutdown. */
1369 if (xlog_is_shutdown(log))
1370 return log->l_logsize;
1372 if (tail_cycle < head_cycle) {
1373 ASSERT(tail_cycle == (head_cycle - 1));
1374 return tail_bytes - head_bytes;
1378 * The reservation head is behind the tail. In this case we just want to
1379 * return the size of the log as the amount of space left.
1381 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1382 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1383 tail_cycle, tail_bytes);
1384 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1385 head_cycle, head_bytes);
1387 return log->l_logsize;
1393 struct work_struct *work)
1395 struct xlog_in_core *iclog =
1396 container_of(work, struct xlog_in_core, ic_end_io_work);
1397 struct xlog *log = iclog->ic_log;
1400 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1402 /* treat writes with injected CRC errors as failed */
1403 if (iclog->ic_fail_crc)
1408 * Race to shutdown the filesystem if we see an error.
1410 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1411 xfs_alert(log->l_mp, "log I/O error %d", error);
1412 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1415 xlog_state_done_syncing(iclog);
1416 bio_uninit(&iclog->ic_bio);
1419 * Drop the lock to signal that we are done. Nothing references the
1420 * iclog after this, so an unmount waiting on this lock can now tear it
1421 * down safely. As such, it is unsafe to reference the iclog after the
1422 * unlock as we could race with it being freed.
1424 up(&iclog->ic_sema);
1428 * Return size of each in-core log record buffer.
1430 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1432 * If the filesystem blocksize is too large, we may need to choose a
1433 * larger size since the directory code currently logs entire blocks.
1436 xlog_get_iclog_buffer_size(
1437 struct xfs_mount *mp,
1440 if (mp->m_logbufs <= 0)
1441 mp->m_logbufs = XLOG_MAX_ICLOGS;
1442 if (mp->m_logbsize <= 0)
1443 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1445 log->l_iclog_bufs = mp->m_logbufs;
1446 log->l_iclog_size = mp->m_logbsize;
1449 * # headers = size / 32k - one header holds cycles from 32k of data.
1451 log->l_iclog_heads =
1452 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1453 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1458 struct xfs_mount *mp)
1460 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1461 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1465 * Clear the log incompat flags if we have the opportunity.
1467 * This only happens if we're about to log the second dummy transaction as part
1468 * of covering the log and we can get the log incompat feature usage lock.
1471 xlog_clear_incompat(
1474 struct xfs_mount *mp = log->l_mp;
1476 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1477 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1480 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1483 if (!down_write_trylock(&log->l_incompat_users))
1486 xfs_clear_incompat_log_features(mp);
1487 up_write(&log->l_incompat_users);
1491 * Every sync period we need to unpin all items in the AIL and push them to
1492 * disk. If there is nothing dirty, then we might need to cover the log to
1493 * indicate that the filesystem is idle.
1497 struct work_struct *work)
1499 struct xlog *log = container_of(to_delayed_work(work),
1500 struct xlog, l_work);
1501 struct xfs_mount *mp = log->l_mp;
1503 /* dgc: errors ignored - not fatal and nowhere to report them */
1504 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1506 * Dump a transaction into the log that contains no real change.
1507 * This is needed to stamp the current tail LSN into the log
1508 * during the covering operation.
1510 * We cannot use an inode here for this - that will push dirty
1511 * state back up into the VFS and then periodic inode flushing
1512 * will prevent log covering from making progress. Hence we
1513 * synchronously log the superblock instead to ensure the
1514 * superblock is immediately unpinned and can be written back.
1516 xlog_clear_incompat(log);
1517 xfs_sync_sb(mp, true);
1519 xfs_log_force(mp, 0);
1521 /* start pushing all the metadata that is currently dirty */
1522 xfs_ail_push_all(mp->m_ail);
1524 /* queue us up again */
1525 xfs_log_work_queue(mp);
1529 * This routine initializes some of the log structure for a given mount point.
1530 * Its primary purpose is to fill in enough, so recovery can occur. However,
1531 * some other stuff may be filled in too.
1533 STATIC struct xlog *
1535 struct xfs_mount *mp,
1536 struct xfs_buftarg *log_target,
1537 xfs_daddr_t blk_offset,
1541 xlog_rec_header_t *head;
1542 xlog_in_core_t **iclogp;
1543 xlog_in_core_t *iclog, *prev_iclog=NULL;
1545 int error = -ENOMEM;
1548 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1550 xfs_warn(mp, "Log allocation failed: No memory!");
1555 log->l_targ = log_target;
1556 log->l_logsize = BBTOB(num_bblks);
1557 log->l_logBBstart = blk_offset;
1558 log->l_logBBsize = num_bblks;
1559 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1560 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1561 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1563 log->l_prev_block = -1;
1564 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1565 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1566 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1567 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1569 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1570 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1572 log->l_iclog_roundoff = BBSIZE;
1574 xlog_grant_head_init(&log->l_reserve_head);
1575 xlog_grant_head_init(&log->l_write_head);
1577 error = -EFSCORRUPTED;
1578 if (xfs_has_sector(mp)) {
1579 log2_size = mp->m_sb.sb_logsectlog;
1580 if (log2_size < BBSHIFT) {
1581 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1582 log2_size, BBSHIFT);
1586 log2_size -= BBSHIFT;
1587 if (log2_size > mp->m_sectbb_log) {
1588 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1589 log2_size, mp->m_sectbb_log);
1593 /* for larger sector sizes, must have v2 or external log */
1594 if (log2_size && log->l_logBBstart > 0 &&
1595 !xfs_has_logv2(mp)) {
1597 "log sector size (0x%x) invalid for configuration.",
1602 log->l_sectBBsize = 1 << log2_size;
1604 init_rwsem(&log->l_incompat_users);
1606 xlog_get_iclog_buffer_size(mp, log);
1608 spin_lock_init(&log->l_icloglock);
1609 init_waitqueue_head(&log->l_flush_wait);
1611 iclogp = &log->l_iclog;
1613 * The amount of memory to allocate for the iclog structure is
1614 * rather funky due to the way the structure is defined. It is
1615 * done this way so that we can use different sizes for machines
1616 * with different amounts of memory. See the definition of
1617 * xlog_in_core_t in xfs_log_priv.h for details.
1619 ASSERT(log->l_iclog_size >= 4096);
1620 for (i = 0; i < log->l_iclog_bufs; i++) {
1621 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1622 sizeof(struct bio_vec);
1624 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1626 goto out_free_iclog;
1629 iclog->ic_prev = prev_iclog;
1632 iclog->ic_data = kvzalloc(log->l_iclog_size,
1633 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1634 if (!iclog->ic_data)
1635 goto out_free_iclog;
1636 head = &iclog->ic_header;
1637 memset(head, 0, sizeof(xlog_rec_header_t));
1638 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1639 head->h_version = cpu_to_be32(
1640 xfs_has_logv2(log->l_mp) ? 2 : 1);
1641 head->h_size = cpu_to_be32(log->l_iclog_size);
1643 head->h_fmt = cpu_to_be32(XLOG_FMT);
1644 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1646 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1647 iclog->ic_state = XLOG_STATE_ACTIVE;
1648 iclog->ic_log = log;
1649 atomic_set(&iclog->ic_refcnt, 0);
1650 INIT_LIST_HEAD(&iclog->ic_callbacks);
1651 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1653 init_waitqueue_head(&iclog->ic_force_wait);
1654 init_waitqueue_head(&iclog->ic_write_wait);
1655 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1656 sema_init(&iclog->ic_sema, 1);
1658 iclogp = &iclog->ic_next;
1660 *iclogp = log->l_iclog; /* complete ring */
1661 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1663 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1664 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1666 0, mp->m_super->s_id);
1667 if (!log->l_ioend_workqueue)
1668 goto out_free_iclog;
1670 error = xlog_cil_init(log);
1672 goto out_destroy_workqueue;
1675 out_destroy_workqueue:
1676 destroy_workqueue(log->l_ioend_workqueue);
1678 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1679 prev_iclog = iclog->ic_next;
1680 kmem_free(iclog->ic_data);
1682 if (prev_iclog == log->l_iclog)
1688 return ERR_PTR(error);
1689 } /* xlog_alloc_log */
1692 * Compute the LSN that we'd need to push the log tail towards in order to have
1693 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1694 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1695 * log free space already meets all three thresholds, this function returns
1699 xlog_grant_push_threshold(
1703 xfs_lsn_t threshold_lsn = 0;
1704 xfs_lsn_t last_sync_lsn;
1707 int threshold_block;
1708 int threshold_cycle;
1711 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1713 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1714 free_blocks = BTOBBT(free_bytes);
1717 * Set the threshold for the minimum number of free blocks in the
1718 * log to the maximum of what the caller needs, one quarter of the
1719 * log, and 256 blocks.
1721 free_threshold = BTOBB(need_bytes);
1722 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1723 free_threshold = max(free_threshold, 256);
1724 if (free_blocks >= free_threshold)
1725 return NULLCOMMITLSN;
1727 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1729 threshold_block += free_threshold;
1730 if (threshold_block >= log->l_logBBsize) {
1731 threshold_block -= log->l_logBBsize;
1732 threshold_cycle += 1;
1734 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1737 * Don't pass in an lsn greater than the lsn of the last
1738 * log record known to be on disk. Use a snapshot of the last sync lsn
1739 * so that it doesn't change between the compare and the set.
1741 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1742 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1743 threshold_lsn = last_sync_lsn;
1745 return threshold_lsn;
1749 * Push the tail of the log if we need to do so to maintain the free log space
1750 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1751 * policy which pushes on an lsn which is further along in the log once we
1752 * reach the high water mark. In this manner, we would be creating a low water
1756 xlog_grant_push_ail(
1760 xfs_lsn_t threshold_lsn;
1762 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1763 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1767 * Get the transaction layer to kick the dirty buffers out to
1768 * disk asynchronously. No point in trying to do this if
1769 * the filesystem is shutting down.
1771 xfs_ail_push(log->l_ailp, threshold_lsn);
1775 * Stamp cycle number in every block
1780 struct xlog_in_core *iclog,
1784 int size = iclog->ic_offset + roundoff;
1788 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1790 dp = iclog->ic_datap;
1791 for (i = 0; i < BTOBB(size); i++) {
1792 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1794 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1795 *(__be32 *)dp = cycle_lsn;
1799 if (xfs_has_logv2(log->l_mp)) {
1800 xlog_in_core_2_t *xhdr = iclog->ic_data;
1802 for ( ; i < BTOBB(size); i++) {
1803 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1804 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1805 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1806 *(__be32 *)dp = cycle_lsn;
1810 for (i = 1; i < log->l_iclog_heads; i++)
1811 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1816 * Calculate the checksum for a log buffer.
1818 * This is a little more complicated than it should be because the various
1819 * headers and the actual data are non-contiguous.
1824 struct xlog_rec_header *rhead,
1830 /* first generate the crc for the record header ... */
1831 crc = xfs_start_cksum_update((char *)rhead,
1832 sizeof(struct xlog_rec_header),
1833 offsetof(struct xlog_rec_header, h_crc));
1835 /* ... then for additional cycle data for v2 logs ... */
1836 if (xfs_has_logv2(log->l_mp)) {
1837 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1841 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1843 for (i = 1; i < xheads; i++) {
1844 crc = crc32c(crc, &xhdr[i].hic_xheader,
1845 sizeof(struct xlog_rec_ext_header));
1849 /* ... and finally for the payload */
1850 crc = crc32c(crc, dp, size);
1852 return xfs_end_cksum(crc);
1859 struct xlog_in_core *iclog = bio->bi_private;
1861 queue_work(iclog->ic_log->l_ioend_workqueue,
1862 &iclog->ic_end_io_work);
1866 xlog_map_iclog_data(
1872 struct page *page = kmem_to_page(data);
1873 unsigned int off = offset_in_page(data);
1874 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1876 if (bio_add_page(bio, page, len, off) != len)
1889 struct xlog_in_core *iclog,
1893 ASSERT(bno < log->l_logBBsize);
1894 trace_xlog_iclog_write(iclog, _RET_IP_);
1897 * We lock the iclogbufs here so that we can serialise against I/O
1898 * completion during unmount. We might be processing a shutdown
1899 * triggered during unmount, and that can occur asynchronously to the
1900 * unmount thread, and hence we need to ensure that completes before
1901 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1902 * across the log IO to archieve that.
1904 down(&iclog->ic_sema);
1905 if (xlog_is_shutdown(log)) {
1907 * It would seem logical to return EIO here, but we rely on
1908 * the log state machine to propagate I/O errors instead of
1909 * doing it here. We kick of the state machine and unlock
1910 * the buffer manually, the code needs to be kept in sync
1911 * with the I/O completion path.
1913 xlog_state_done_syncing(iclog);
1914 up(&iclog->ic_sema);
1919 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1920 * IOs coming immediately after this one. This prevents the block layer
1921 * writeback throttle from throttling log writes behind background
1922 * metadata writeback and causing priority inversions.
1924 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1925 howmany(count, PAGE_SIZE),
1926 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1927 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1928 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1929 iclog->ic_bio.bi_private = iclog;
1931 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1932 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1934 * For external log devices, we also need to flush the data
1935 * device cache first to ensure all metadata writeback covered
1936 * by the LSN in this iclog is on stable storage. This is slow,
1937 * but it *must* complete before we issue the external log IO.
1939 * If the flush fails, we cannot conclude that past metadata
1940 * writeback from the log succeeded. Repeating the flush is
1941 * not possible, hence we must shut down with log IO error to
1942 * avoid shutdown re-entering this path and erroring out again.
1944 if (log->l_targ != log->l_mp->m_ddev_targp &&
1945 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev)) {
1946 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1950 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1951 iclog->ic_bio.bi_opf |= REQ_FUA;
1953 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1955 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1956 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1959 if (is_vmalloc_addr(iclog->ic_data))
1960 flush_kernel_vmap_range(iclog->ic_data, count);
1963 * If this log buffer would straddle the end of the log we will have
1964 * to split it up into two bios, so that we can continue at the start.
1966 if (bno + BTOBB(count) > log->l_logBBsize) {
1969 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1970 GFP_NOIO, &fs_bio_set);
1971 bio_chain(split, &iclog->ic_bio);
1974 /* restart at logical offset zero for the remainder */
1975 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1978 submit_bio(&iclog->ic_bio);
1982 * We need to bump cycle number for the part of the iclog that is
1983 * written to the start of the log. Watch out for the header magic
1984 * number case, though.
1993 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1996 for (i = split_offset; i < count; i += BBSIZE) {
1997 uint32_t cycle = get_unaligned_be32(data + i);
1999 if (++cycle == XLOG_HEADER_MAGIC_NUM)
2001 put_unaligned_be32(cycle, data + i);
2006 xlog_calc_iclog_size(
2008 struct xlog_in_core *iclog,
2011 uint32_t count_init, count;
2013 /* Add for LR header */
2014 count_init = log->l_iclog_hsize + iclog->ic_offset;
2015 count = roundup(count_init, log->l_iclog_roundoff);
2017 *roundoff = count - count_init;
2019 ASSERT(count >= count_init);
2020 ASSERT(*roundoff < log->l_iclog_roundoff);
2025 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2026 * fashion. Previously, we should have moved the current iclog
2027 * ptr in the log to point to the next available iclog. This allows further
2028 * write to continue while this code syncs out an iclog ready to go.
2029 * Before an in-core log can be written out, the data section must be scanned
2030 * to save away the 1st word of each BBSIZE block into the header. We replace
2031 * it with the current cycle count. Each BBSIZE block is tagged with the
2032 * cycle count because there in an implicit assumption that drives will
2033 * guarantee that entire 512 byte blocks get written at once. In other words,
2034 * we can't have part of a 512 byte block written and part not written. By
2035 * tagging each block, we will know which blocks are valid when recovering
2036 * after an unclean shutdown.
2038 * This routine is single threaded on the iclog. No other thread can be in
2039 * this routine with the same iclog. Changing contents of iclog can there-
2040 * fore be done without grabbing the state machine lock. Updating the global
2041 * log will require grabbing the lock though.
2043 * The entire log manager uses a logical block numbering scheme. Only
2044 * xlog_write_iclog knows about the fact that the log may not start with
2045 * block zero on a given device.
2050 struct xlog_in_core *iclog,
2051 struct xlog_ticket *ticket)
2053 unsigned int count; /* byte count of bwrite */
2054 unsigned int roundoff; /* roundoff to BB or stripe */
2058 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2059 trace_xlog_iclog_sync(iclog, _RET_IP_);
2061 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2064 * If we have a ticket, account for the roundoff via the ticket
2065 * reservation to avoid touching the hot grant heads needlessly.
2066 * Otherwise, we have to move grant heads directly.
2069 ticket->t_curr_res -= roundoff;
2071 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2072 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2075 /* put cycle number in every block */
2076 xlog_pack_data(log, iclog, roundoff);
2078 /* real byte length */
2079 size = iclog->ic_offset;
2080 if (xfs_has_logv2(log->l_mp))
2082 iclog->ic_header.h_len = cpu_to_be32(size);
2084 XFS_STATS_INC(log->l_mp, xs_log_writes);
2085 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2087 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2089 /* Do we need to split this write into 2 parts? */
2090 if (bno + BTOBB(count) > log->l_logBBsize)
2091 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2093 /* calculcate the checksum */
2094 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2095 iclog->ic_datap, size);
2097 * Intentionally corrupt the log record CRC based on the error injection
2098 * frequency, if defined. This facilitates testing log recovery in the
2099 * event of torn writes. Hence, set the IOABORT state to abort the log
2100 * write on I/O completion and shutdown the fs. The subsequent mount
2101 * detects the bad CRC and attempts to recover.
2104 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2105 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2106 iclog->ic_fail_crc = true;
2108 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2109 be64_to_cpu(iclog->ic_header.h_lsn));
2112 xlog_verify_iclog(log, iclog, count);
2113 xlog_write_iclog(log, iclog, bno, count);
2117 * Deallocate a log structure
2123 xlog_in_core_t *iclog, *next_iclog;
2127 * Destroy the CIL after waiting for iclog IO completion because an
2128 * iclog EIO error will try to shut down the log, which accesses the
2129 * CIL to wake up the waiters.
2131 xlog_cil_destroy(log);
2133 iclog = log->l_iclog;
2134 for (i = 0; i < log->l_iclog_bufs; i++) {
2135 next_iclog = iclog->ic_next;
2136 kmem_free(iclog->ic_data);
2141 log->l_mp->m_log = NULL;
2142 destroy_workqueue(log->l_ioend_workqueue);
2147 * Update counters atomically now that memcpy is done.
2150 xlog_state_finish_copy(
2152 struct xlog_in_core *iclog,
2156 lockdep_assert_held(&log->l_icloglock);
2158 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2159 iclog->ic_offset += copy_bytes;
2163 * print out info relating to regions written which consume
2168 struct xfs_mount *mp,
2169 struct xlog_ticket *ticket)
2171 xfs_warn(mp, "ticket reservation summary:");
2172 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
2173 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
2174 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
2175 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
2179 * Print a summary of the transaction.
2183 struct xfs_trans *tp)
2185 struct xfs_mount *mp = tp->t_mountp;
2186 struct xfs_log_item *lip;
2188 /* dump core transaction and ticket info */
2189 xfs_warn(mp, "transaction summary:");
2190 xfs_warn(mp, " log res = %d", tp->t_log_res);
2191 xfs_warn(mp, " log count = %d", tp->t_log_count);
2192 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2194 xlog_print_tic_res(mp, tp->t_ticket);
2196 /* dump each log item */
2197 list_for_each_entry(lip, &tp->t_items, li_trans) {
2198 struct xfs_log_vec *lv = lip->li_lv;
2199 struct xfs_log_iovec *vec;
2202 xfs_warn(mp, "log item: ");
2203 xfs_warn(mp, " type = 0x%x", lip->li_type);
2204 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2207 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2208 xfs_warn(mp, " size = %d", lv->lv_size);
2209 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2210 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2212 /* dump each iovec for the log item */
2213 vec = lv->lv_iovecp;
2214 for (i = 0; i < lv->lv_niovecs; i++) {
2215 int dumplen = min(vec->i_len, 32);
2217 xfs_warn(mp, " iovec[%d]", i);
2218 xfs_warn(mp, " type = 0x%x", vec->i_type);
2219 xfs_warn(mp, " len = %d", vec->i_len);
2220 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2221 xfs_hex_dump(vec->i_addr, dumplen);
2230 struct xlog_in_core *iclog,
2231 uint32_t *log_offset,
2235 uint32_t *record_cnt,
2238 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2239 ASSERT(*log_offset % sizeof(int32_t) == 0);
2240 ASSERT(write_len % sizeof(int32_t) == 0);
2242 memcpy(iclog->ic_datap + *log_offset, data, write_len);
2243 *log_offset += write_len;
2244 *bytes_left -= write_len;
2246 *data_cnt += write_len;
2250 * Write log vectors into a single iclog which is guaranteed by the caller
2251 * to have enough space to write the entire log vector into.
2255 struct xfs_log_vec *lv,
2256 struct xlog_ticket *ticket,
2257 struct xlog_in_core *iclog,
2258 uint32_t *log_offset,
2260 uint32_t *record_cnt,
2265 ASSERT(*log_offset + *len <= iclog->ic_size ||
2266 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2269 * Ordered log vectors have no regions to write so this
2270 * loop will naturally skip them.
2272 for (index = 0; index < lv->lv_niovecs; index++) {
2273 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2274 struct xlog_op_header *ophdr = reg->i_addr;
2276 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2277 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2278 reg->i_len, len, record_cnt, data_cnt);
2283 xlog_write_get_more_iclog_space(
2284 struct xlog_ticket *ticket,
2285 struct xlog_in_core **iclogp,
2286 uint32_t *log_offset,
2288 uint32_t *record_cnt,
2291 struct xlog_in_core *iclog = *iclogp;
2292 struct xlog *log = iclog->ic_log;
2295 spin_lock(&log->l_icloglock);
2296 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2297 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2298 error = xlog_state_release_iclog(log, iclog, ticket);
2299 spin_unlock(&log->l_icloglock);
2303 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2314 * Write log vectors into a single iclog which is smaller than the current chain
2315 * length. We write until we cannot fit a full record into the remaining space
2316 * and then stop. We return the log vector that is to be written that cannot
2317 * wholly fit in the iclog.
2321 struct xfs_log_vec *lv,
2322 struct xlog_ticket *ticket,
2323 struct xlog_in_core **iclogp,
2324 uint32_t *log_offset,
2326 uint32_t *record_cnt,
2329 struct xlog_in_core *iclog = *iclogp;
2330 struct xlog_op_header *ophdr;
2335 /* walk the logvec, copying until we run out of space in the iclog */
2336 for (index = 0; index < lv->lv_niovecs; index++) {
2337 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2338 uint32_t reg_offset = 0;
2341 * The first region of a continuation must have a non-zero
2342 * length otherwise log recovery will just skip over it and
2343 * start recovering from the next opheader it finds. Because we
2344 * mark the next opheader as a continuation, recovery will then
2345 * incorrectly add the continuation to the previous region and
2346 * that breaks stuff.
2348 * Hence if there isn't space for region data after the
2349 * opheader, then we need to start afresh with a new iclog.
2351 if (iclog->ic_size - *log_offset <=
2352 sizeof(struct xlog_op_header)) {
2353 error = xlog_write_get_more_iclog_space(ticket,
2354 &iclog, log_offset, *len, record_cnt,
2360 ophdr = reg->i_addr;
2361 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2363 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2364 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2365 if (rlen != reg->i_len)
2366 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2368 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2369 rlen, len, record_cnt, data_cnt);
2371 /* If we wrote the whole region, move to the next. */
2372 if (rlen == reg->i_len)
2376 * We now have a partially written iovec, but it can span
2377 * multiple iclogs so we loop here. First we release the iclog
2378 * we currently have, then we get a new iclog and add a new
2379 * opheader. Then we continue copying from where we were until
2380 * we either complete the iovec or fill the iclog. If we
2381 * complete the iovec, then we increment the index and go right
2382 * back to the top of the outer loop. if we fill the iclog, we
2383 * run the inner loop again.
2385 * This is complicated by the tail of a region using all the
2386 * space in an iclog and hence requiring us to release the iclog
2387 * and get a new one before returning to the outer loop. We must
2388 * always guarantee that we exit this inner loop with at least
2389 * space for log transaction opheaders left in the current
2390 * iclog, hence we cannot just terminate the loop at the end
2391 * of the of the continuation. So we loop while there is no
2392 * space left in the current iclog, and check for the end of the
2393 * continuation after getting a new iclog.
2397 * Ensure we include the continuation opheader in the
2398 * space we need in the new iclog by adding that size
2399 * to the length we require. This continuation opheader
2400 * needs to be accounted to the ticket as the space it
2401 * consumes hasn't been accounted to the lv we are
2404 error = xlog_write_get_more_iclog_space(ticket,
2406 *len + sizeof(struct xlog_op_header),
2407 record_cnt, data_cnt);
2411 ophdr = iclog->ic_datap + *log_offset;
2412 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2413 ophdr->oh_clientid = XFS_TRANSACTION;
2415 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2417 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2418 *log_offset += sizeof(struct xlog_op_header);
2419 *data_cnt += sizeof(struct xlog_op_header);
2422 * If rlen fits in the iclog, then end the region
2423 * continuation. Otherwise we're going around again.
2426 rlen = reg->i_len - reg_offset;
2427 if (rlen <= iclog->ic_size - *log_offset)
2428 ophdr->oh_flags |= XLOG_END_TRANS;
2430 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2432 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2433 ophdr->oh_len = cpu_to_be32(rlen);
2435 xlog_write_iovec(iclog, log_offset,
2436 reg->i_addr + reg_offset,
2437 rlen, len, record_cnt, data_cnt);
2439 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2443 * No more iovecs remain in this logvec so return the next log vec to
2444 * the caller so it can go back to fast path copying.
2451 * Write some region out to in-core log
2453 * This will be called when writing externally provided regions or when
2454 * writing out a commit record for a given transaction.
2456 * General algorithm:
2457 * 1. Find total length of this write. This may include adding to the
2458 * lengths passed in.
2459 * 2. Check whether we violate the tickets reservation.
2460 * 3. While writing to this iclog
2461 * A. Reserve as much space in this iclog as can get
2462 * B. If this is first write, save away start lsn
2463 * C. While writing this region:
2464 * 1. If first write of transaction, write start record
2465 * 2. Write log operation header (header per region)
2466 * 3. Find out if we can fit entire region into this iclog
2467 * 4. Potentially, verify destination memcpy ptr
2468 * 5. Memcpy (partial) region
2469 * 6. If partial copy, release iclog; otherwise, continue
2470 * copying more regions into current iclog
2471 * 4. Mark want sync bit (in simulation mode)
2472 * 5. Release iclog for potential flush to on-disk log.
2475 * 1. Panic if reservation is overrun. This should never happen since
2476 * reservation amounts are generated internal to the filesystem.
2478 * 1. Tickets are single threaded data structures.
2479 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2480 * syncing routine. When a single log_write region needs to span
2481 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2482 * on all log operation writes which don't contain the end of the
2483 * region. The XLOG_END_TRANS bit is used for the in-core log
2484 * operation which contains the end of the continued log_write region.
2485 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2486 * we don't really know exactly how much space will be used. As a result,
2487 * we don't update ic_offset until the end when we know exactly how many
2488 * bytes have been written out.
2493 struct xfs_cil_ctx *ctx,
2494 struct list_head *lv_chain,
2495 struct xlog_ticket *ticket,
2499 struct xlog_in_core *iclog = NULL;
2500 struct xfs_log_vec *lv;
2501 uint32_t record_cnt = 0;
2502 uint32_t data_cnt = 0;
2506 if (ticket->t_curr_res < 0) {
2507 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2508 "ctx ticket reservation ran out. Need to up reservation");
2509 xlog_print_tic_res(log->l_mp, ticket);
2510 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2513 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2518 ASSERT(log_offset <= iclog->ic_size - 1);
2521 * If we have a context pointer, pass it the first iclog we are
2522 * writing to so it can record state needed for iclog write
2526 xlog_cil_set_ctx_write_state(ctx, iclog);
2528 list_for_each_entry(lv, lv_chain, lv_list) {
2530 * If the entire log vec does not fit in the iclog, punt it to
2531 * the partial copy loop which can handle this case.
2533 if (lv->lv_niovecs &&
2534 lv->lv_bytes > iclog->ic_size - log_offset) {
2535 error = xlog_write_partial(lv, ticket, &iclog,
2536 &log_offset, &len, &record_cnt,
2540 * We have no iclog to release, so just return
2541 * the error immediately.
2546 xlog_write_full(lv, ticket, iclog, &log_offset,
2547 &len, &record_cnt, &data_cnt);
2553 * We've already been guaranteed that the last writes will fit inside
2554 * the current iclog, and hence it will already have the space used by
2555 * those writes accounted to it. Hence we do not need to update the
2556 * iclog with the number of bytes written here.
2558 spin_lock(&log->l_icloglock);
2559 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2560 error = xlog_state_release_iclog(log, iclog, ticket);
2561 spin_unlock(&log->l_icloglock);
2567 xlog_state_activate_iclog(
2568 struct xlog_in_core *iclog,
2569 int *iclogs_changed)
2571 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2572 trace_xlog_iclog_activate(iclog, _RET_IP_);
2575 * If the number of ops in this iclog indicate it just contains the
2576 * dummy transaction, we can change state into IDLE (the second time
2577 * around). Otherwise we should change the state into NEED a dummy.
2578 * We don't need to cover the dummy.
2580 if (*iclogs_changed == 0 &&
2581 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2582 *iclogs_changed = 1;
2585 * We have two dirty iclogs so start over. This could also be
2586 * num of ops indicating this is not the dummy going out.
2588 *iclogs_changed = 2;
2591 iclog->ic_state = XLOG_STATE_ACTIVE;
2592 iclog->ic_offset = 0;
2593 iclog->ic_header.h_num_logops = 0;
2594 memset(iclog->ic_header.h_cycle_data, 0,
2595 sizeof(iclog->ic_header.h_cycle_data));
2596 iclog->ic_header.h_lsn = 0;
2597 iclog->ic_header.h_tail_lsn = 0;
2601 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2602 * ACTIVE after iclog I/O has completed.
2605 xlog_state_activate_iclogs(
2607 int *iclogs_changed)
2609 struct xlog_in_core *iclog = log->l_iclog;
2612 if (iclog->ic_state == XLOG_STATE_DIRTY)
2613 xlog_state_activate_iclog(iclog, iclogs_changed);
2615 * The ordering of marking iclogs ACTIVE must be maintained, so
2616 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2618 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2620 } while ((iclog = iclog->ic_next) != log->l_iclog);
2629 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2630 * wrote the first covering record (DONE). We go to IDLE if we just
2631 * wrote the second covering record (DONE2) and remain in IDLE until a
2632 * non-covering write occurs.
2634 switch (prev_state) {
2635 case XLOG_STATE_COVER_IDLE:
2636 if (iclogs_changed == 1)
2637 return XLOG_STATE_COVER_IDLE;
2639 case XLOG_STATE_COVER_NEED:
2640 case XLOG_STATE_COVER_NEED2:
2642 case XLOG_STATE_COVER_DONE:
2643 if (iclogs_changed == 1)
2644 return XLOG_STATE_COVER_NEED2;
2646 case XLOG_STATE_COVER_DONE2:
2647 if (iclogs_changed == 1)
2648 return XLOG_STATE_COVER_IDLE;
2654 return XLOG_STATE_COVER_NEED;
2658 xlog_state_clean_iclog(
2660 struct xlog_in_core *dirty_iclog)
2662 int iclogs_changed = 0;
2664 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2666 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2668 xlog_state_activate_iclogs(log, &iclogs_changed);
2669 wake_up_all(&dirty_iclog->ic_force_wait);
2671 if (iclogs_changed) {
2672 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2678 xlog_get_lowest_lsn(
2681 struct xlog_in_core *iclog = log->l_iclog;
2682 xfs_lsn_t lowest_lsn = 0, lsn;
2685 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2686 iclog->ic_state == XLOG_STATE_DIRTY)
2689 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2690 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2692 } while ((iclog = iclog->ic_next) != log->l_iclog);
2698 * Completion of a iclog IO does not imply that a transaction has completed, as
2699 * transactions can be large enough to span many iclogs. We cannot change the
2700 * tail of the log half way through a transaction as this may be the only
2701 * transaction in the log and moving the tail to point to the middle of it
2702 * will prevent recovery from finding the start of the transaction. Hence we
2703 * should only update the last_sync_lsn if this iclog contains transaction
2704 * completion callbacks on it.
2706 * We have to do this before we drop the icloglock to ensure we are the only one
2707 * that can update it.
2709 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2710 * the reservation grant head pushing. This is due to the fact that the push
2711 * target is bound by the current last_sync_lsn value. Hence if we have a large
2712 * amount of log space bound up in this committing transaction then the
2713 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2714 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2715 * should push the AIL to ensure the push target (and hence the grant head) is
2716 * no longer bound by the old log head location and can move forwards and make
2720 xlog_state_set_callback(
2722 struct xlog_in_core *iclog,
2723 xfs_lsn_t header_lsn)
2725 trace_xlog_iclog_callback(iclog, _RET_IP_);
2726 iclog->ic_state = XLOG_STATE_CALLBACK;
2728 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2731 if (list_empty_careful(&iclog->ic_callbacks))
2734 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2735 xlog_grant_push_ail(log, 0);
2739 * Return true if we need to stop processing, false to continue to the next
2740 * iclog. The caller will need to run callbacks if the iclog is returned in the
2741 * XLOG_STATE_CALLBACK state.
2744 xlog_state_iodone_process_iclog(
2746 struct xlog_in_core *iclog)
2748 xfs_lsn_t lowest_lsn;
2749 xfs_lsn_t header_lsn;
2751 switch (iclog->ic_state) {
2752 case XLOG_STATE_ACTIVE:
2753 case XLOG_STATE_DIRTY:
2755 * Skip all iclogs in the ACTIVE & DIRTY states:
2758 case XLOG_STATE_DONE_SYNC:
2760 * Now that we have an iclog that is in the DONE_SYNC state, do
2761 * one more check here to see if we have chased our tail around.
2762 * If this is not the lowest lsn iclog, then we will leave it
2763 * for another completion to process.
2765 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2766 lowest_lsn = xlog_get_lowest_lsn(log);
2767 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2769 xlog_state_set_callback(log, iclog, header_lsn);
2773 * Can only perform callbacks in order. Since this iclog is not
2774 * in the DONE_SYNC state, we skip the rest and just try to
2782 * Loop over all the iclogs, running attached callbacks on them. Return true if
2783 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2784 * to handle transient shutdown state here at all because
2785 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2786 * cleanup of the callbacks.
2789 xlog_state_do_iclog_callbacks(
2791 __releases(&log->l_icloglock)
2792 __acquires(&log->l_icloglock)
2794 struct xlog_in_core *first_iclog = log->l_iclog;
2795 struct xlog_in_core *iclog = first_iclog;
2796 bool ran_callback = false;
2801 if (xlog_state_iodone_process_iclog(log, iclog))
2803 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2804 iclog = iclog->ic_next;
2807 list_splice_init(&iclog->ic_callbacks, &cb_list);
2808 spin_unlock(&log->l_icloglock);
2810 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2811 xlog_cil_process_committed(&cb_list);
2812 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2813 ran_callback = true;
2815 spin_lock(&log->l_icloglock);
2816 xlog_state_clean_iclog(log, iclog);
2817 iclog = iclog->ic_next;
2818 } while (iclog != first_iclog);
2820 return ran_callback;
2825 * Loop running iclog completion callbacks until there are no more iclogs in a
2826 * state that can run callbacks.
2829 xlog_state_do_callback(
2835 spin_lock(&log->l_icloglock);
2836 while (xlog_state_do_iclog_callbacks(log)) {
2837 if (xlog_is_shutdown(log))
2840 if (++repeats > 5000) {
2841 flushcnt += repeats;
2844 "%s: possible infinite loop (%d iterations)",
2845 __func__, flushcnt);
2849 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2850 wake_up_all(&log->l_flush_wait);
2852 spin_unlock(&log->l_icloglock);
2857 * Finish transitioning this iclog to the dirty state.
2859 * Callbacks could take time, so they are done outside the scope of the
2860 * global state machine log lock.
2863 xlog_state_done_syncing(
2864 struct xlog_in_core *iclog)
2866 struct xlog *log = iclog->ic_log;
2868 spin_lock(&log->l_icloglock);
2869 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2870 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2873 * If we got an error, either on the first buffer, or in the case of
2874 * split log writes, on the second, we shut down the file system and
2875 * no iclogs should ever be attempted to be written to disk again.
2877 if (!xlog_is_shutdown(log)) {
2878 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2879 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2883 * Someone could be sleeping prior to writing out the next
2884 * iclog buffer, we wake them all, one will get to do the
2885 * I/O, the others get to wait for the result.
2887 wake_up_all(&iclog->ic_write_wait);
2888 spin_unlock(&log->l_icloglock);
2889 xlog_state_do_callback(log);
2893 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2894 * sleep. We wait on the flush queue on the head iclog as that should be
2895 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2896 * we will wait here and all new writes will sleep until a sync completes.
2898 * The in-core logs are used in a circular fashion. They are not used
2899 * out-of-order even when an iclog past the head is free.
2902 * * log_offset where xlog_write() can start writing into the in-core
2904 * * in-core log pointer to which xlog_write() should write.
2905 * * boolean indicating this is a continued write to an in-core log.
2906 * If this is the last write, then the in-core log's offset field
2907 * needs to be incremented, depending on the amount of data which
2911 xlog_state_get_iclog_space(
2914 struct xlog_in_core **iclogp,
2915 struct xlog_ticket *ticket,
2919 xlog_rec_header_t *head;
2920 xlog_in_core_t *iclog;
2923 spin_lock(&log->l_icloglock);
2924 if (xlog_is_shutdown(log)) {
2925 spin_unlock(&log->l_icloglock);
2929 iclog = log->l_iclog;
2930 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2931 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2933 /* Wait for log writes to have flushed */
2934 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2938 head = &iclog->ic_header;
2940 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2941 log_offset = iclog->ic_offset;
2943 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2945 /* On the 1st write to an iclog, figure out lsn. This works
2946 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2947 * committing to. If the offset is set, that's how many blocks
2950 if (log_offset == 0) {
2951 ticket->t_curr_res -= log->l_iclog_hsize;
2952 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2953 head->h_lsn = cpu_to_be64(
2954 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2955 ASSERT(log->l_curr_block >= 0);
2958 /* If there is enough room to write everything, then do it. Otherwise,
2959 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2960 * bit is on, so this will get flushed out. Don't update ic_offset
2961 * until you know exactly how many bytes get copied. Therefore, wait
2962 * until later to update ic_offset.
2964 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2965 * can fit into remaining data section.
2967 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2970 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2973 * If we are the only one writing to this iclog, sync it to
2974 * disk. We need to do an atomic compare and decrement here to
2975 * avoid racing with concurrent atomic_dec_and_lock() calls in
2976 * xlog_state_release_iclog() when there is more than one
2977 * reference to the iclog.
2979 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2980 error = xlog_state_release_iclog(log, iclog, ticket);
2981 spin_unlock(&log->l_icloglock);
2987 /* Do we have enough room to write the full amount in the remainder
2988 * of this iclog? Or must we continue a write on the next iclog and
2989 * mark this iclog as completely taken? In the case where we switch
2990 * iclogs (to mark it taken), this particular iclog will release/sync
2991 * to disk in xlog_write().
2993 if (len <= iclog->ic_size - iclog->ic_offset)
2994 iclog->ic_offset += len;
2996 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2999 ASSERT(iclog->ic_offset <= iclog->ic_size);
3000 spin_unlock(&log->l_icloglock);
3002 *logoffsetp = log_offset;
3007 * The first cnt-1 times a ticket goes through here we don't need to move the
3008 * grant write head because the permanent reservation has reserved cnt times the
3009 * unit amount. Release part of current permanent unit reservation and reset
3010 * current reservation to be one units worth. Also move grant reservation head
3014 xfs_log_ticket_regrant(
3016 struct xlog_ticket *ticket)
3018 trace_xfs_log_ticket_regrant(log, ticket);
3020 if (ticket->t_cnt > 0)
3023 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3024 ticket->t_curr_res);
3025 xlog_grant_sub_space(log, &log->l_write_head.grant,
3026 ticket->t_curr_res);
3027 ticket->t_curr_res = ticket->t_unit_res;
3029 trace_xfs_log_ticket_regrant_sub(log, ticket);
3031 /* just return if we still have some of the pre-reserved space */
3032 if (!ticket->t_cnt) {
3033 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3034 ticket->t_unit_res);
3035 trace_xfs_log_ticket_regrant_exit(log, ticket);
3037 ticket->t_curr_res = ticket->t_unit_res;
3040 xfs_log_ticket_put(ticket);
3044 * Give back the space left from a reservation.
3046 * All the information we need to make a correct determination of space left
3047 * is present. For non-permanent reservations, things are quite easy. The
3048 * count should have been decremented to zero. We only need to deal with the
3049 * space remaining in the current reservation part of the ticket. If the
3050 * ticket contains a permanent reservation, there may be left over space which
3051 * needs to be released. A count of N means that N-1 refills of the current
3052 * reservation can be done before we need to ask for more space. The first
3053 * one goes to fill up the first current reservation. Once we run out of
3054 * space, the count will stay at zero and the only space remaining will be
3055 * in the current reservation field.
3058 xfs_log_ticket_ungrant(
3060 struct xlog_ticket *ticket)
3064 trace_xfs_log_ticket_ungrant(log, ticket);
3066 if (ticket->t_cnt > 0)
3069 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3072 * If this is a permanent reservation ticket, we may be able to free
3073 * up more space based on the remaining count.
3075 bytes = ticket->t_curr_res;
3076 if (ticket->t_cnt > 0) {
3077 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3078 bytes += ticket->t_unit_res*ticket->t_cnt;
3081 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3082 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3084 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3086 xfs_log_space_wake(log->l_mp);
3087 xfs_log_ticket_put(ticket);
3091 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3092 * the current iclog pointer to the next iclog in the ring.
3095 xlog_state_switch_iclogs(
3097 struct xlog_in_core *iclog,
3100 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3101 assert_spin_locked(&log->l_icloglock);
3102 trace_xlog_iclog_switch(iclog, _RET_IP_);
3105 eventual_size = iclog->ic_offset;
3106 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3107 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3108 log->l_prev_block = log->l_curr_block;
3109 log->l_prev_cycle = log->l_curr_cycle;
3111 /* roll log?: ic_offset changed later */
3112 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3114 /* Round up to next log-sunit */
3115 if (log->l_iclog_roundoff > BBSIZE) {
3116 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3117 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3120 if (log->l_curr_block >= log->l_logBBsize) {
3122 * Rewind the current block before the cycle is bumped to make
3123 * sure that the combined LSN never transiently moves forward
3124 * when the log wraps to the next cycle. This is to support the
3125 * unlocked sample of these fields from xlog_valid_lsn(). Most
3126 * other cases should acquire l_icloglock.
3128 log->l_curr_block -= log->l_logBBsize;
3129 ASSERT(log->l_curr_block >= 0);
3131 log->l_curr_cycle++;
3132 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3133 log->l_curr_cycle++;
3135 ASSERT(iclog == log->l_iclog);
3136 log->l_iclog = iclog->ic_next;
3140 * Force the iclog to disk and check if the iclog has been completed before
3141 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3142 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3143 * If completion has already occurred, tell the caller so that it can avoid an
3144 * unnecessary wait on the iclog.
3147 xlog_force_and_check_iclog(
3148 struct xlog_in_core *iclog,
3151 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3155 error = xlog_force_iclog(iclog);
3160 * If the iclog has already been completed and reused the header LSN
3161 * will have been rewritten by completion
3163 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3169 * Write out all data in the in-core log as of this exact moment in time.
3171 * Data may be written to the in-core log during this call. However,
3172 * we don't guarantee this data will be written out. A change from past
3173 * implementation means this routine will *not* write out zero length LRs.
3175 * Basically, we try and perform an intelligent scan of the in-core logs.
3176 * If we determine there is no flushable data, we just return. There is no
3177 * flushable data if:
3179 * 1. the current iclog is active and has no data; the previous iclog
3180 * is in the active or dirty state.
3181 * 2. the current iclog is drity, and the previous iclog is in the
3182 * active or dirty state.
3186 * 1. the current iclog is not in the active nor dirty state.
3187 * 2. the current iclog dirty, and the previous iclog is not in the
3188 * active nor dirty state.
3189 * 3. the current iclog is active, and there is another thread writing
3190 * to this particular iclog.
3191 * 4. a) the current iclog is active and has no other writers
3192 * b) when we return from flushing out this iclog, it is still
3193 * not in the active nor dirty state.
3197 struct xfs_mount *mp,
3200 struct xlog *log = mp->m_log;
3201 struct xlog_in_core *iclog;
3203 XFS_STATS_INC(mp, xs_log_force);
3204 trace_xfs_log_force(mp, 0, _RET_IP_);
3206 xlog_cil_force(log);
3208 spin_lock(&log->l_icloglock);
3209 if (xlog_is_shutdown(log))
3212 iclog = log->l_iclog;
3213 trace_xlog_iclog_force(iclog, _RET_IP_);
3215 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3216 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3217 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3219 * If the head is dirty or (active and empty), then we need to
3220 * look at the previous iclog.
3222 * If the previous iclog is active or dirty we are done. There
3223 * is nothing to sync out. Otherwise, we attach ourselves to the
3224 * previous iclog and go to sleep.
3226 iclog = iclog->ic_prev;
3227 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3228 if (atomic_read(&iclog->ic_refcnt) == 0) {
3229 /* We have exclusive access to this iclog. */
3232 if (xlog_force_and_check_iclog(iclog, &completed))
3239 * Someone else is still writing to this iclog, so we
3240 * need to ensure that when they release the iclog it
3241 * gets synced immediately as we may be waiting on it.
3243 xlog_state_switch_iclogs(log, iclog, 0);
3248 * The iclog we are about to wait on may contain the checkpoint pushed
3249 * by the above xlog_cil_force() call, but it may not have been pushed
3250 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3251 * are flushed when this iclog is written.
3253 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3254 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3256 if (flags & XFS_LOG_SYNC)
3257 return xlog_wait_on_iclog(iclog);
3259 spin_unlock(&log->l_icloglock);
3262 spin_unlock(&log->l_icloglock);
3267 * Force the log to a specific LSN.
3269 * If an iclog with that lsn can be found:
3270 * If it is in the DIRTY state, just return.
3271 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3272 * state and go to sleep or return.
3273 * If it is in any other state, go to sleep or return.
3275 * Synchronous forces are implemented with a wait queue. All callers trying
3276 * to force a given lsn to disk must wait on the queue attached to the
3277 * specific in-core log. When given in-core log finally completes its write
3278 * to disk, that thread will wake up all threads waiting on the queue.
3288 struct xlog_in_core *iclog;
3291 spin_lock(&log->l_icloglock);
3292 if (xlog_is_shutdown(log))
3295 iclog = log->l_iclog;
3296 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3297 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3298 iclog = iclog->ic_next;
3299 if (iclog == log->l_iclog)
3303 switch (iclog->ic_state) {
3304 case XLOG_STATE_ACTIVE:
3306 * We sleep here if we haven't already slept (e.g. this is the
3307 * first time we've looked at the correct iclog buf) and the
3308 * buffer before us is going to be sync'ed. The reason for this
3309 * is that if we are doing sync transactions here, by waiting
3310 * for the previous I/O to complete, we can allow a few more
3311 * transactions into this iclog before we close it down.
3313 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3314 * refcnt so we can release the log (which drops the ref count).
3315 * The state switch keeps new transaction commits from using
3316 * this buffer. When the current commits finish writing into
3317 * the buffer, the refcount will drop to zero and the buffer
3320 if (!already_slept &&
3321 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3322 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3323 xlog_wait(&iclog->ic_prev->ic_write_wait,
3327 if (xlog_force_and_check_iclog(iclog, &completed))
3334 case XLOG_STATE_WANT_SYNC:
3336 * This iclog may contain the checkpoint pushed by the
3337 * xlog_cil_force_seq() call, but there are other writers still
3338 * accessing it so it hasn't been pushed to disk yet. Like the
3339 * ACTIVE case above, we need to make sure caches are flushed
3340 * when this iclog is written.
3342 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3346 * The entire checkpoint was written by the CIL force and is on
3347 * its way to disk already. It will be stable when it
3348 * completes, so we don't need to manipulate caches here at all.
3349 * We just need to wait for completion if necessary.
3354 if (flags & XFS_LOG_SYNC)
3355 return xlog_wait_on_iclog(iclog);
3357 spin_unlock(&log->l_icloglock);
3360 spin_unlock(&log->l_icloglock);
3365 * Force the log to a specific checkpoint sequence.
3367 * First force the CIL so that all the required changes have been flushed to the
3368 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3369 * the iclog that needs to be flushed to stable storage. If the caller needs
3370 * a synchronous log force, we will wait on the iclog with the LSN returned by
3371 * xlog_cil_force_seq() to be completed.
3375 struct xfs_mount *mp,
3380 struct xlog *log = mp->m_log;
3385 XFS_STATS_INC(mp, xs_log_force);
3386 trace_xfs_log_force(mp, seq, _RET_IP_);
3388 lsn = xlog_cil_force_seq(log, seq);
3389 if (lsn == NULLCOMMITLSN)
3392 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3393 if (ret == -EAGAIN) {
3394 XFS_STATS_INC(mp, xs_log_force_sleep);
3395 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3401 * Free a used ticket when its refcount falls to zero.
3405 xlog_ticket_t *ticket)
3407 ASSERT(atomic_read(&ticket->t_ref) > 0);
3408 if (atomic_dec_and_test(&ticket->t_ref))
3409 kmem_cache_free(xfs_log_ticket_cache, ticket);
3414 xlog_ticket_t *ticket)
3416 ASSERT(atomic_read(&ticket->t_ref) > 0);
3417 atomic_inc(&ticket->t_ref);
3422 * Figure out the total log space unit (in bytes) that would be
3423 * required for a log ticket.
3435 * Permanent reservations have up to 'cnt'-1 active log operations
3436 * in the log. A unit in this case is the amount of space for one
3437 * of these log operations. Normal reservations have a cnt of 1
3438 * and their unit amount is the total amount of space required.
3440 * The following lines of code account for non-transaction data
3441 * which occupy space in the on-disk log.
3443 * Normal form of a transaction is:
3444 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3445 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3447 * We need to account for all the leadup data and trailer data
3448 * around the transaction data.
3449 * And then we need to account for the worst case in terms of using
3451 * The worst case will happen if:
3452 * - the placement of the transaction happens to be such that the
3453 * roundoff is at its maximum
3454 * - the transaction data is synced before the commit record is synced
3455 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3456 * Therefore the commit record is in its own Log Record.
3457 * This can happen as the commit record is called with its
3458 * own region to xlog_write().
3459 * This then means that in the worst case, roundoff can happen for
3460 * the commit-rec as well.
3461 * The commit-rec is smaller than padding in this scenario and so it is
3462 * not added separately.
3465 /* for trans header */
3466 unit_bytes += sizeof(xlog_op_header_t);
3467 unit_bytes += sizeof(xfs_trans_header_t);
3470 unit_bytes += sizeof(xlog_op_header_t);
3473 * for LR headers - the space for data in an iclog is the size minus
3474 * the space used for the headers. If we use the iclog size, then we
3475 * undercalculate the number of headers required.
3477 * Furthermore - the addition of op headers for split-recs might
3478 * increase the space required enough to require more log and op
3479 * headers, so take that into account too.
3481 * IMPORTANT: This reservation makes the assumption that if this
3482 * transaction is the first in an iclog and hence has the LR headers
3483 * accounted to it, then the remaining space in the iclog is
3484 * exclusively for this transaction. i.e. if the transaction is larger
3485 * than the iclog, it will be the only thing in that iclog.
3486 * Fundamentally, this means we must pass the entire log vector to
3487 * xlog_write to guarantee this.
3489 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3490 num_headers = howmany(unit_bytes, iclog_space);
3492 /* for split-recs - ophdrs added when data split over LRs */
3493 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3495 /* add extra header reservations if we overrun */
3496 while (!num_headers ||
3497 howmany(unit_bytes, iclog_space) > num_headers) {
3498 unit_bytes += sizeof(xlog_op_header_t);
3501 unit_bytes += log->l_iclog_hsize * num_headers;
3503 /* for commit-rec LR header - note: padding will subsume the ophdr */
3504 unit_bytes += log->l_iclog_hsize;
3506 /* roundoff padding for transaction data and one for commit record */
3507 unit_bytes += 2 * log->l_iclog_roundoff;
3510 *niclogs = num_headers;
3515 xfs_log_calc_unit_res(
3516 struct xfs_mount *mp,
3519 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3523 * Allocate and initialise a new log ticket.
3525 struct xlog_ticket *
3532 struct xlog_ticket *tic;
3535 tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3537 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3539 atomic_set(&tic->t_ref, 1);
3540 tic->t_task = current;
3541 INIT_LIST_HEAD(&tic->t_queue);
3542 tic->t_unit_res = unit_res;
3543 tic->t_curr_res = unit_res;
3546 tic->t_tid = get_random_u32();
3548 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3555 * Check to make sure the grant write head didn't just over lap the tail. If
3556 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3557 * the cycles differ by exactly one and check the byte count.
3559 * This check is run unlocked, so can give false positives. Rather than assert
3560 * on failures, use a warn-once flag and a panic tag to allow the admin to
3561 * determine if they want to panic the machine when such an error occurs. For
3562 * debug kernels this will have the same effect as using an assert but, unlinke
3563 * an assert, it can be turned off at runtime.
3566 xlog_verify_grant_tail(
3569 int tail_cycle, tail_blocks;
3572 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3573 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3574 if (tail_cycle != cycle) {
3575 if (cycle - 1 != tail_cycle &&
3576 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3577 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3578 "%s: cycle - 1 != tail_cycle", __func__);
3581 if (space > BBTOB(tail_blocks) &&
3582 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3583 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3584 "%s: space > BBTOB(tail_blocks)", __func__);
3589 /* check if it will fit */
3591 xlog_verify_tail_lsn(
3593 struct xlog_in_core *iclog)
3595 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3598 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3600 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3601 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3602 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3604 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3606 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3607 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3609 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3610 if (blocks < BTOBB(iclog->ic_offset) + 1)
3611 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3616 * Perform a number of checks on the iclog before writing to disk.
3618 * 1. Make sure the iclogs are still circular
3619 * 2. Make sure we have a good magic number
3620 * 3. Make sure we don't have magic numbers in the data
3621 * 4. Check fields of each log operation header for:
3622 * A. Valid client identifier
3623 * B. tid ptr value falls in valid ptr space (user space code)
3624 * C. Length in log record header is correct according to the
3625 * individual operation headers within record.
3626 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3627 * log, check the preceding blocks of the physical log to make sure all
3628 * the cycle numbers agree with the current cycle number.
3633 struct xlog_in_core *iclog,
3636 xlog_op_header_t *ophead;
3637 xlog_in_core_t *icptr;
3638 xlog_in_core_2_t *xhdr;
3639 void *base_ptr, *ptr, *p;
3640 ptrdiff_t field_offset;
3642 int len, i, j, k, op_len;
3645 /* check validity of iclog pointers */
3646 spin_lock(&log->l_icloglock);
3647 icptr = log->l_iclog;
3648 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3651 if (icptr != log->l_iclog)
3652 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3653 spin_unlock(&log->l_icloglock);
3655 /* check log magic numbers */
3656 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3657 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3659 base_ptr = ptr = &iclog->ic_header;
3660 p = &iclog->ic_header;
3661 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3662 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3663 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3668 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3669 base_ptr = ptr = iclog->ic_datap;
3671 xhdr = iclog->ic_data;
3672 for (i = 0; i < len; i++) {
3675 /* clientid is only 1 byte */
3676 p = &ophead->oh_clientid;
3677 field_offset = p - base_ptr;
3678 if (field_offset & 0x1ff) {
3679 clientid = ophead->oh_clientid;
3681 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3682 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3683 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3684 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3685 clientid = xlog_get_client_id(
3686 xhdr[j].hic_xheader.xh_cycle_data[k]);
3688 clientid = xlog_get_client_id(
3689 iclog->ic_header.h_cycle_data[idx]);
3692 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3694 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3695 __func__, i, clientid, ophead,
3696 (unsigned long)field_offset);
3700 p = &ophead->oh_len;
3701 field_offset = p - base_ptr;
3702 if (field_offset & 0x1ff) {
3703 op_len = be32_to_cpu(ophead->oh_len);
3705 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3706 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3707 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3708 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3709 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3711 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3714 ptr += sizeof(xlog_op_header_t) + op_len;
3720 * Perform a forced shutdown on the log.
3722 * This can be called from low level log code to trigger a shutdown, or from the
3723 * high level mount shutdown code when the mount shuts down.
3725 * Our main objectives here are to make sure that:
3726 * a. if the shutdown was not due to a log IO error, flush the logs to
3727 * disk. Anything modified after this is ignored.
3728 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3729 * parties to find out. Nothing new gets queued after this is done.
3730 * c. Tasks sleeping on log reservations, pinned objects and
3731 * other resources get woken up.
3732 * d. The mount is also marked as shut down so that log triggered shutdowns
3733 * still behave the same as if they called xfs_forced_shutdown().
3735 * Return true if the shutdown cause was a log IO error and we actually shut the
3739 xlog_force_shutdown(
3741 uint32_t shutdown_flags)
3743 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3749 * Flush all the completed transactions to disk before marking the log
3750 * being shut down. We need to do this first as shutting down the log
3751 * before the force will prevent the log force from flushing the iclogs
3754 * When we are in recovery, there are no transactions to flush, and
3755 * we don't want to touch the log because we don't want to perturb the
3756 * current head/tail for future recovery attempts. Hence we need to
3757 * avoid a log force in this case.
3759 * If we are shutting down due to a log IO error, then we must avoid
3760 * trying to write the log as that may just result in more IO errors and
3761 * an endless shutdown/force loop.
3763 if (!log_error && !xlog_in_recovery(log))
3764 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3767 * Atomically set the shutdown state. If the shutdown state is already
3768 * set, there someone else is performing the shutdown and so we are done
3769 * here. This should never happen because we should only ever get called
3770 * once by the first shutdown caller.
3772 * Much of the log state machine transitions assume that shutdown state
3773 * cannot change once they hold the log->l_icloglock. Hence we need to
3774 * hold that lock here, even though we use the atomic test_and_set_bit()
3775 * operation to set the shutdown state.
3777 spin_lock(&log->l_icloglock);
3778 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3779 spin_unlock(&log->l_icloglock);
3782 spin_unlock(&log->l_icloglock);
3785 * If this log shutdown also sets the mount shutdown state, issue a
3786 * shutdown warning message.
3788 if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3789 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3790 "Filesystem has been shut down due to log error (0x%x).",
3792 xfs_alert(log->l_mp,
3793 "Please unmount the filesystem and rectify the problem(s).");
3794 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3799 * We don't want anybody waiting for log reservations after this. That
3800 * means we have to wake up everybody queued up on reserveq as well as
3801 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3802 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3803 * action is protected by the grant locks.
3805 xlog_grant_head_wake_all(&log->l_reserve_head);
3806 xlog_grant_head_wake_all(&log->l_write_head);
3809 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3810 * as if the log writes were completed. The abort handling in the log
3811 * item committed callback functions will do this again under lock to
3814 spin_lock(&log->l_cilp->xc_push_lock);
3815 wake_up_all(&log->l_cilp->xc_start_wait);
3816 wake_up_all(&log->l_cilp->xc_commit_wait);
3817 spin_unlock(&log->l_cilp->xc_push_lock);
3819 spin_lock(&log->l_icloglock);
3820 xlog_state_shutdown_callbacks(log);
3821 spin_unlock(&log->l_icloglock);
3823 wake_up_var(&log->l_opstate);
3831 xlog_in_core_t *iclog;
3833 iclog = log->l_iclog;
3835 /* endianness does not matter here, zero is zero in
3838 if (iclog->ic_header.h_num_logops)
3840 iclog = iclog->ic_next;
3841 } while (iclog != log->l_iclog);
3846 * Verify that an LSN stamped into a piece of metadata is valid. This is
3847 * intended for use in read verifiers on v5 superblocks.
3851 struct xfs_mount *mp,
3854 struct xlog *log = mp->m_log;
3858 * norecovery mode skips mount-time log processing and unconditionally
3859 * resets the in-core LSN. We can't validate in this mode, but
3860 * modifications are not allowed anyways so just return true.
3862 if (xfs_has_norecovery(mp))
3866 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3867 * handled by recovery and thus safe to ignore here.
3869 if (lsn == NULLCOMMITLSN)
3872 valid = xlog_valid_lsn(mp->m_log, lsn);
3874 /* warn the user about what's gone wrong before verifier failure */
3876 spin_lock(&log->l_icloglock);
3878 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3879 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3880 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3881 log->l_curr_cycle, log->l_curr_block);
3882 spin_unlock(&log->l_icloglock);
3889 * Notify the log that we're about to start using a feature that is protected
3890 * by a log incompat feature flag. This will prevent log covering from
3891 * clearing those flags.
3894 xlog_use_incompat_feat(
3897 down_read(&log->l_incompat_users);
3900 /* Notify the log that we've finished using log incompat features. */
3902 xlog_drop_incompat_feat(
3905 up_read(&log->l_incompat_users);