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);
63 xlog_verify_grant_tail(
68 struct xlog_in_core *iclog,
73 struct xlog_in_core *iclog);
75 #define xlog_verify_grant_tail(a)
76 #define xlog_verify_iclog(a,b,c)
77 #define xlog_verify_tail_lsn(a,b)
85 xfs_log_cover(struct xfs_mount *);
88 * We need to make sure the buffer pointer returned is naturally aligned for the
89 * biggest basic data type we put into it. We have already accounted for this
90 * padding when sizing the buffer.
92 * However, this padding does not get written into the log, and hence we have to
93 * track the space used by the log vectors separately to prevent log space hangs
94 * due to inaccurate accounting (i.e. a leak) of the used log space through the
97 * We also add space for the xlog_op_header that describes this region in the
98 * log. This prepends the data region we return to the caller to copy their data
99 * into, so do all the static initialisation of the ophdr now. Because the ophdr
100 * is not 8 byte aligned, we have to be careful to ensure that we align the
101 * start of the buffer such that the region we return to the call is 8 byte
102 * aligned and packed against the tail of the ophdr.
106 struct xfs_log_vec *lv,
107 struct xfs_log_iovec **vecp,
110 struct xfs_log_iovec *vec = *vecp;
111 struct xlog_op_header *oph;
116 ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
119 vec = &lv->lv_iovecp[0];
122 len = lv->lv_buf_len + sizeof(struct xlog_op_header);
123 if (!IS_ALIGNED(len, sizeof(uint64_t))) {
124 lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
125 sizeof(struct xlog_op_header);
129 vec->i_addr = lv->lv_buf + lv->lv_buf_len;
132 oph->oh_clientid = XFS_TRANSACTION;
136 buf = vec->i_addr + sizeof(struct xlog_op_header);
137 ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
144 xlog_grant_sub_space(
149 int64_t head_val = atomic64_read(head);
155 xlog_crack_grant_head_val(head_val, &cycle, &space);
159 space += log->l_logsize;
164 new = xlog_assign_grant_head_val(cycle, space);
165 head_val = atomic64_cmpxchg(head, old, new);
166 } while (head_val != old);
170 xlog_grant_add_space(
175 int64_t head_val = atomic64_read(head);
182 xlog_crack_grant_head_val(head_val, &cycle, &space);
184 tmp = log->l_logsize - space;
193 new = xlog_assign_grant_head_val(cycle, space);
194 head_val = atomic64_cmpxchg(head, old, new);
195 } while (head_val != old);
199 xlog_grant_head_init(
200 struct xlog_grant_head *head)
202 xlog_assign_grant_head(&head->grant, 1, 0);
203 INIT_LIST_HEAD(&head->waiters);
204 spin_lock_init(&head->lock);
208 xlog_grant_head_wake_all(
209 struct xlog_grant_head *head)
211 struct xlog_ticket *tic;
213 spin_lock(&head->lock);
214 list_for_each_entry(tic, &head->waiters, t_queue)
215 wake_up_process(tic->t_task);
216 spin_unlock(&head->lock);
220 xlog_ticket_reservation(
222 struct xlog_grant_head *head,
223 struct xlog_ticket *tic)
225 if (head == &log->l_write_head) {
226 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
227 return tic->t_unit_res;
229 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
230 return tic->t_unit_res * tic->t_cnt;
232 return tic->t_unit_res;
237 xlog_grant_head_wake(
239 struct xlog_grant_head *head,
242 struct xlog_ticket *tic;
244 bool woken_task = false;
246 list_for_each_entry(tic, &head->waiters, t_queue) {
249 * There is a chance that the size of the CIL checkpoints in
250 * progress at the last AIL push target calculation resulted in
251 * limiting the target to the log head (l_last_sync_lsn) at the
252 * time. This may not reflect where the log head is now as the
253 * CIL checkpoints may have completed.
255 * Hence when we are woken here, it may be that the head of the
256 * log that has moved rather than the tail. As the tail didn't
257 * move, there still won't be space available for the
258 * reservation we require. However, if the AIL has already
259 * pushed to the target defined by the old log head location, we
260 * will hang here waiting for something else to update the AIL
263 * Therefore, if there isn't space to wake the first waiter on
264 * the grant head, we need to push the AIL again to ensure the
265 * target reflects both the current log tail and log head
266 * position before we wait for the tail to move again.
269 need_bytes = xlog_ticket_reservation(log, head, tic);
270 if (*free_bytes < need_bytes) {
272 xlog_grant_push_ail(log, need_bytes);
276 *free_bytes -= need_bytes;
277 trace_xfs_log_grant_wake_up(log, tic);
278 wake_up_process(tic->t_task);
286 xlog_grant_head_wait(
288 struct xlog_grant_head *head,
289 struct xlog_ticket *tic,
290 int need_bytes) __releases(&head->lock)
291 __acquires(&head->lock)
293 list_add_tail(&tic->t_queue, &head->waiters);
296 if (xlog_is_shutdown(log))
298 xlog_grant_push_ail(log, need_bytes);
300 __set_current_state(TASK_UNINTERRUPTIBLE);
301 spin_unlock(&head->lock);
303 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
305 trace_xfs_log_grant_sleep(log, tic);
307 trace_xfs_log_grant_wake(log, tic);
309 spin_lock(&head->lock);
310 if (xlog_is_shutdown(log))
312 } while (xlog_space_left(log, &head->grant) < need_bytes);
314 list_del_init(&tic->t_queue);
317 list_del_init(&tic->t_queue);
322 * Atomically get the log space required for a log ticket.
324 * Once a ticket gets put onto head->waiters, it will only return after the
325 * needed reservation is satisfied.
327 * This function is structured so that it has a lock free fast path. This is
328 * necessary because every new transaction reservation will come through this
329 * path. Hence any lock will be globally hot if we take it unconditionally on
332 * As tickets are only ever moved on and off head->waiters under head->lock, we
333 * only need to take that lock if we are going to add the ticket to the queue
334 * and sleep. We can avoid taking the lock if the ticket was never added to
335 * head->waiters because the t_queue list head will be empty and we hold the
336 * only reference to it so it can safely be checked unlocked.
339 xlog_grant_head_check(
341 struct xlog_grant_head *head,
342 struct xlog_ticket *tic,
348 ASSERT(!xlog_in_recovery(log));
351 * If there are other waiters on the queue then give them a chance at
352 * logspace before us. Wake up the first waiters, if we do not wake
353 * up all the waiters then go to sleep waiting for more free space,
354 * otherwise try to get some space for this transaction.
356 *need_bytes = xlog_ticket_reservation(log, head, tic);
357 free_bytes = xlog_space_left(log, &head->grant);
358 if (!list_empty_careful(&head->waiters)) {
359 spin_lock(&head->lock);
360 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
361 free_bytes < *need_bytes) {
362 error = xlog_grant_head_wait(log, head, tic,
365 spin_unlock(&head->lock);
366 } else if (free_bytes < *need_bytes) {
367 spin_lock(&head->lock);
368 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
369 spin_unlock(&head->lock);
377 struct xfs_mount *mp)
380 * Do not write to the log on norecovery mounts, if the data or log
381 * devices are read-only, or if the filesystem is shutdown. Read-only
382 * mounts allow internal writes for log recovery and unmount purposes,
383 * so don't restrict that case.
385 if (xfs_has_norecovery(mp))
387 if (xfs_readonly_buftarg(mp->m_ddev_targp))
389 if (xfs_readonly_buftarg(mp->m_log->l_targ))
391 if (xlog_is_shutdown(mp->m_log))
397 * Replenish the byte reservation required by moving the grant write head.
401 struct xfs_mount *mp,
402 struct xlog_ticket *tic)
404 struct xlog *log = mp->m_log;
408 if (xlog_is_shutdown(log))
411 XFS_STATS_INC(mp, xs_try_logspace);
414 * This is a new transaction on the ticket, so we need to change the
415 * transaction ID so that the next transaction has a different TID in
416 * the log. Just add one to the existing tid so that we can see chains
417 * of rolling transactions in the log easily.
421 xlog_grant_push_ail(log, tic->t_unit_res);
423 tic->t_curr_res = tic->t_unit_res;
427 trace_xfs_log_regrant(log, tic);
429 error = xlog_grant_head_check(log, &log->l_write_head, tic,
434 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
435 trace_xfs_log_regrant_exit(log, tic);
436 xlog_verify_grant_tail(log);
441 * If we are failing, make sure the ticket doesn't have any current
442 * reservations. We don't want to add this back when the ticket/
443 * transaction gets cancelled.
446 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
451 * Reserve log space and return a ticket corresponding to the reservation.
453 * Each reservation is going to reserve extra space for a log record header.
454 * When writes happen to the on-disk log, we don't subtract the length of the
455 * log record header from any reservation. By wasting space in each
456 * reservation, we prevent over allocation problems.
460 struct xfs_mount *mp,
463 struct xlog_ticket **ticp,
466 struct xlog *log = mp->m_log;
467 struct xlog_ticket *tic;
471 if (xlog_is_shutdown(log))
474 XFS_STATS_INC(mp, xs_try_logspace);
476 ASSERT(*ticp == NULL);
477 tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
480 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
483 trace_xfs_log_reserve(log, tic);
485 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
490 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
491 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
492 trace_xfs_log_reserve_exit(log, tic);
493 xlog_verify_grant_tail(log);
498 * If we are failing, make sure the ticket doesn't have any current
499 * reservations. We don't want to add this back when the ticket/
500 * transaction gets cancelled.
503 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
508 * Run all the pending iclog callbacks and wake log force waiters and iclog
509 * space waiters so they can process the newly set shutdown state. We really
510 * don't care what order we process callbacks here because the log is shut down
511 * and so state cannot change on disk anymore. However, we cannot wake waiters
512 * until the callbacks have been processed because we may be in unmount and
513 * we must ensure that all AIL operations the callbacks perform have completed
514 * before we tear down the AIL.
516 * We avoid processing actively referenced iclogs so that we don't run callbacks
517 * while the iclog owner might still be preparing the iclog for IO submssion.
518 * These will be caught by xlog_state_iclog_release() and call this function
519 * again to process any callbacks that may have been added to that iclog.
522 xlog_state_shutdown_callbacks(
525 struct xlog_in_core *iclog;
528 iclog = log->l_iclog;
530 if (atomic_read(&iclog->ic_refcnt)) {
531 /* Reference holder will re-run iclog callbacks. */
534 list_splice_init(&iclog->ic_callbacks, &cb_list);
535 spin_unlock(&log->l_icloglock);
537 xlog_cil_process_committed(&cb_list);
539 spin_lock(&log->l_icloglock);
540 wake_up_all(&iclog->ic_write_wait);
541 wake_up_all(&iclog->ic_force_wait);
542 } while ((iclog = iclog->ic_next) != log->l_iclog);
544 wake_up_all(&log->l_flush_wait);
548 * Flush iclog to disk if this is the last reference to the given iclog and the
549 * it is in the WANT_SYNC state.
551 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
552 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
553 * written to stable storage, and implies that a commit record is contained
554 * within the iclog. We need to ensure that the log tail does not move beyond
555 * the tail that the first commit record in the iclog ordered against, otherwise
556 * correct recovery of that checkpoint becomes dependent on future operations
557 * performed on this iclog.
559 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
560 * current tail into iclog. Once the iclog tail is set, future operations must
561 * not modify it, otherwise they potentially violate ordering constraints for
562 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
563 * the iclog will get zeroed on activation of the iclog after sync, so we
564 * always capture the tail lsn on the iclog on the first NEED_FUA release
565 * regardless of the number of active reference counts on this iclog.
568 xlog_state_release_iclog(
570 struct xlog_in_core *iclog)
575 lockdep_assert_held(&log->l_icloglock);
577 trace_xlog_iclog_release(iclog, _RET_IP_);
579 * Grabbing the current log tail needs to be atomic w.r.t. the writing
580 * of the tail LSN into the iclog so we guarantee that the log tail does
581 * not move between the first time we know that the iclog needs to be
582 * made stable and when we eventually submit it.
584 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
585 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
586 !iclog->ic_header.h_tail_lsn) {
587 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
588 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
591 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
593 if (xlog_is_shutdown(log)) {
595 * If there are no more references to this iclog, process the
596 * pending iclog callbacks that were waiting on the release of
600 xlog_state_shutdown_callbacks(log);
607 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
608 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
612 iclog->ic_state = XLOG_STATE_SYNCING;
613 xlog_verify_tail_lsn(log, iclog);
614 trace_xlog_iclog_syncing(iclog, _RET_IP_);
616 spin_unlock(&log->l_icloglock);
617 xlog_sync(log, iclog);
618 spin_lock(&log->l_icloglock);
623 * Mount a log filesystem
625 * mp - ubiquitous xfs mount point structure
626 * log_target - buftarg of on-disk log device
627 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
628 * num_bblocks - Number of BBSIZE blocks in on-disk log
630 * Return error or zero.
635 xfs_buftarg_t *log_target,
636 xfs_daddr_t blk_offset,
640 bool fatal = xfs_has_crc(mp);
644 if (!xfs_has_norecovery(mp)) {
645 xfs_notice(mp, "Mounting V%d Filesystem",
646 XFS_SB_VERSION_NUM(&mp->m_sb));
649 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
650 XFS_SB_VERSION_NUM(&mp->m_sb));
651 ASSERT(xfs_is_readonly(mp));
654 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
656 error = PTR_ERR(log);
662 * Validate the given log space and drop a critical message via syslog
663 * if the log size is too small that would lead to some unexpected
664 * situations in transaction log space reservation stage.
666 * Note: we can't just reject the mount if the validation fails. This
667 * would mean that people would have to downgrade their kernel just to
668 * remedy the situation as there is no way to grow the log (short of
669 * black magic surgery with xfs_db).
671 * We can, however, reject mounts for CRC format filesystems, as the
672 * mkfs binary being used to make the filesystem should never create a
673 * filesystem with a log that is too small.
675 min_logfsbs = xfs_log_calc_minimum_size(mp);
677 if (mp->m_sb.sb_logblocks < min_logfsbs) {
679 "Log size %d blocks too small, minimum size is %d blocks",
680 mp->m_sb.sb_logblocks, min_logfsbs);
682 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
684 "Log size %d blocks too large, maximum size is %lld blocks",
685 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
687 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
689 "log size %lld bytes too large, maximum size is %lld bytes",
690 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
693 } else if (mp->m_sb.sb_logsunit > 1 &&
694 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
696 "log stripe unit %u bytes must be a multiple of block size",
697 mp->m_sb.sb_logsunit);
703 * Log check errors are always fatal on v5; or whenever bad
704 * metadata leads to a crash.
707 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
711 xfs_crit(mp, "Log size out of supported range.");
713 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
717 * Initialize the AIL now we have a log.
719 error = xfs_trans_ail_init(mp);
721 xfs_warn(mp, "AIL initialisation failed: error %d", error);
724 log->l_ailp = mp->m_ail;
727 * skip log recovery on a norecovery mount. pretend it all
730 if (!xfs_has_norecovery(mp)) {
732 * log recovery ignores readonly state and so we need to clear
733 * mount-based read only state so it can write to disk.
735 bool readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
737 error = xlog_recover(log);
739 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
741 xfs_warn(mp, "log mount/recovery failed: error %d",
743 xlog_recover_cancel(log);
744 goto out_destroy_ail;
748 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
751 goto out_destroy_ail;
753 /* Normal transactions can now occur */
754 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
757 * Now the log has been fully initialised and we know were our
758 * space grant counters are, we can initialise the permanent ticket
759 * needed for delayed logging to work.
761 xlog_cil_init_post_recovery(log);
766 xfs_trans_ail_destroy(mp);
768 xlog_dealloc_log(log);
774 * Finish the recovery of the file system. This is separate from the
775 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
776 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
779 * If we finish recovery successfully, start the background log work. If we are
780 * not doing recovery, then we have a RO filesystem and we don't need to start
784 xfs_log_mount_finish(
785 struct xfs_mount *mp)
787 struct xlog *log = mp->m_log;
791 if (xfs_has_norecovery(mp)) {
792 ASSERT(xfs_is_readonly(mp));
797 * log recovery ignores readonly state and so we need to clear
798 * mount-based read only state so it can write to disk.
800 readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
803 * During the second phase of log recovery, we need iget and
804 * iput to behave like they do for an active filesystem.
805 * xfs_fs_drop_inode needs to be able to prevent the deletion
806 * of inodes before we're done replaying log items on those
807 * inodes. Turn it off immediately after recovery finishes
808 * so that we don't leak the quota inodes if subsequent mount
811 * We let all inodes involved in redo item processing end up on
812 * the LRU instead of being evicted immediately so that if we do
813 * something to an unlinked inode, the irele won't cause
814 * premature truncation and freeing of the inode, which results
815 * in log recovery failure. We have to evict the unreferenced
816 * lru inodes after clearing SB_ACTIVE because we don't
817 * otherwise clean up the lru if there's a subsequent failure in
818 * xfs_mountfs, which leads to us leaking the inodes if nothing
819 * else (e.g. quotacheck) references the inodes before the
820 * mount failure occurs.
822 mp->m_super->s_flags |= SB_ACTIVE;
823 xfs_log_work_queue(mp);
824 if (xlog_recovery_needed(log))
825 error = xlog_recover_finish(log);
826 mp->m_super->s_flags &= ~SB_ACTIVE;
827 evict_inodes(mp->m_super);
830 * Drain the buffer LRU after log recovery. This is required for v4
831 * filesystems to avoid leaving around buffers with NULL verifier ops,
832 * but we do it unconditionally to make sure we're always in a clean
833 * cache state after mount.
835 * Don't push in the error case because the AIL may have pending intents
836 * that aren't removed until recovery is cancelled.
838 if (xlog_recovery_needed(log)) {
840 xfs_log_force(mp, XFS_LOG_SYNC);
841 xfs_ail_push_all_sync(mp->m_ail);
843 xfs_notice(mp, "Ending recovery (logdev: %s)",
844 mp->m_logname ? mp->m_logname : "internal");
846 xfs_info(mp, "Ending clean mount");
848 xfs_buftarg_drain(mp->m_ddev_targp);
850 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
852 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
854 /* Make sure the log is dead if we're returning failure. */
855 ASSERT(!error || xlog_is_shutdown(log));
861 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
865 xfs_log_mount_cancel(
866 struct xfs_mount *mp)
868 xlog_recover_cancel(mp->m_log);
873 * Flush out the iclog to disk ensuring that device caches are flushed and
874 * the iclog hits stable storage before any completion waiters are woken.
878 struct xlog_in_core *iclog)
880 atomic_inc(&iclog->ic_refcnt);
881 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
882 if (iclog->ic_state == XLOG_STATE_ACTIVE)
883 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
884 return xlog_state_release_iclog(iclog->ic_log, iclog);
888 * Wait for the iclog and all prior iclogs to be written disk as required by the
889 * log force state machine. Waiting on ic_force_wait ensures iclog completions
890 * have been ordered and callbacks run before we are woken here, hence
891 * guaranteeing that all the iclogs up to this one are on stable storage.
895 struct xlog_in_core *iclog)
896 __releases(iclog->ic_log->l_icloglock)
898 struct xlog *log = iclog->ic_log;
900 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
901 if (!xlog_is_shutdown(log) &&
902 iclog->ic_state != XLOG_STATE_ACTIVE &&
903 iclog->ic_state != XLOG_STATE_DIRTY) {
904 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
905 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
907 spin_unlock(&log->l_icloglock);
910 if (xlog_is_shutdown(log))
916 * Write out an unmount record using the ticket provided. We have to account for
917 * the data space used in the unmount ticket as this write is not done from a
918 * transaction context that has already done the accounting for us.
921 xlog_write_unmount_record(
923 struct xlog_ticket *ticket)
926 struct xlog_op_header ophdr;
927 struct xfs_unmount_log_format ulf;
930 .oh_clientid = XFS_LOG,
931 .oh_tid = cpu_to_be32(ticket->t_tid),
932 .oh_flags = XLOG_UNMOUNT_TRANS,
935 .magic = XLOG_UNMOUNT_TYPE,
938 struct xfs_log_iovec reg = {
939 .i_addr = &unmount_rec,
940 .i_len = sizeof(unmount_rec),
941 .i_type = XLOG_REG_TYPE_UNMOUNT,
943 struct xfs_log_vec vec = {
948 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
949 sizeof(struct xfs_unmount_log_format)) !=
950 sizeof(unmount_rec));
952 /* account for space used by record data */
953 ticket->t_curr_res -= sizeof(unmount_rec);
955 return xlog_write(log, NULL, &vec, ticket, reg.i_len);
959 * Mark the filesystem clean by writing an unmount record to the head of the
966 struct xfs_mount *mp = log->l_mp;
967 struct xlog_in_core *iclog;
968 struct xlog_ticket *tic = NULL;
971 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
975 error = xlog_write_unmount_record(log, tic);
977 * At this point, we're umounting anyway, so there's no point in
978 * transitioning log state to shutdown. Just continue...
982 xfs_alert(mp, "%s: unmount record failed", __func__);
984 spin_lock(&log->l_icloglock);
985 iclog = log->l_iclog;
986 error = xlog_force_iclog(iclog);
987 xlog_wait_on_iclog(iclog);
990 trace_xfs_log_umount_write(log, tic);
991 xfs_log_ticket_ungrant(log, tic);
996 xfs_log_unmount_verify_iclog(
999 struct xlog_in_core *iclog = log->l_iclog;
1002 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
1003 ASSERT(iclog->ic_offset == 0);
1004 } while ((iclog = iclog->ic_next) != log->l_iclog);
1008 * Unmount record used to have a string "Unmount filesystem--" in the
1009 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
1010 * We just write the magic number now since that particular field isn't
1011 * currently architecture converted and "Unmount" is a bit foo.
1012 * As far as I know, there weren't any dependencies on the old behaviour.
1015 xfs_log_unmount_write(
1016 struct xfs_mount *mp)
1018 struct xlog *log = mp->m_log;
1020 if (!xfs_log_writable(mp))
1023 xfs_log_force(mp, XFS_LOG_SYNC);
1025 if (xlog_is_shutdown(log))
1029 * If we think the summary counters are bad, avoid writing the unmount
1030 * record to force log recovery at next mount, after which the summary
1031 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1034 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1035 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1036 xfs_alert(mp, "%s: will fix summary counters at next mount",
1041 xfs_log_unmount_verify_iclog(log);
1042 xlog_unmount_write(log);
1046 * Empty the log for unmount/freeze.
1048 * To do this, we first need to shut down the background log work so it is not
1049 * trying to cover the log as we clean up. We then need to unpin all objects in
1050 * the log so we can then flush them out. Once they have completed their IO and
1051 * run the callbacks removing themselves from the AIL, we can cover the log.
1055 struct xfs_mount *mp)
1058 * Clear log incompat features since we're quiescing the log. Report
1059 * failures, though it's not fatal to have a higher log feature
1060 * protection level than the log contents actually require.
1062 if (xfs_clear_incompat_log_features(mp)) {
1065 error = xfs_sync_sb(mp, false);
1068 "Failed to clear log incompat features on quiesce");
1071 cancel_delayed_work_sync(&mp->m_log->l_work);
1072 xfs_log_force(mp, XFS_LOG_SYNC);
1075 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1076 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1077 * xfs_buf_iowait() cannot be used because it was pushed with the
1078 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1079 * the IO to complete.
1081 xfs_ail_push_all_sync(mp->m_ail);
1082 xfs_buftarg_wait(mp->m_ddev_targp);
1083 xfs_buf_lock(mp->m_sb_bp);
1084 xfs_buf_unlock(mp->m_sb_bp);
1086 return xfs_log_cover(mp);
1091 struct xfs_mount *mp)
1093 xfs_log_quiesce(mp);
1094 xfs_log_unmount_write(mp);
1098 * Shut down and release the AIL and Log.
1100 * During unmount, we need to ensure we flush all the dirty metadata objects
1101 * from the AIL so that the log is empty before we write the unmount record to
1102 * the log. Once this is done, we can tear down the AIL and the log.
1106 struct xfs_mount *mp)
1110 xfs_buftarg_drain(mp->m_ddev_targp);
1112 xfs_trans_ail_destroy(mp);
1114 xfs_sysfs_del(&mp->m_log->l_kobj);
1116 xlog_dealloc_log(mp->m_log);
1121 struct xfs_mount *mp,
1122 struct xfs_log_item *item,
1124 const struct xfs_item_ops *ops)
1126 item->li_log = mp->m_log;
1127 item->li_ailp = mp->m_ail;
1128 item->li_type = type;
1132 INIT_LIST_HEAD(&item->li_ail);
1133 INIT_LIST_HEAD(&item->li_cil);
1134 INIT_LIST_HEAD(&item->li_bio_list);
1135 INIT_LIST_HEAD(&item->li_trans);
1139 * Wake up processes waiting for log space after we have moved the log tail.
1143 struct xfs_mount *mp)
1145 struct xlog *log = mp->m_log;
1148 if (xlog_is_shutdown(log))
1151 if (!list_empty_careful(&log->l_write_head.waiters)) {
1152 ASSERT(!xlog_in_recovery(log));
1154 spin_lock(&log->l_write_head.lock);
1155 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1156 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1157 spin_unlock(&log->l_write_head.lock);
1160 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1161 ASSERT(!xlog_in_recovery(log));
1163 spin_lock(&log->l_reserve_head.lock);
1164 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1165 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1166 spin_unlock(&log->l_reserve_head.lock);
1171 * Determine if we have a transaction that has gone to disk that needs to be
1172 * covered. To begin the transition to the idle state firstly the log needs to
1173 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1174 * we start attempting to cover the log.
1176 * Only if we are then in a state where covering is needed, the caller is
1177 * informed that dummy transactions are required to move the log into the idle
1180 * If there are any items in the AIl or CIL, then we do not want to attempt to
1181 * cover the log as we may be in a situation where there isn't log space
1182 * available to run a dummy transaction and this can lead to deadlocks when the
1183 * tail of the log is pinned by an item that is modified in the CIL. Hence
1184 * there's no point in running a dummy transaction at this point because we
1185 * can't start trying to idle the log until both the CIL and AIL are empty.
1188 xfs_log_need_covered(
1189 struct xfs_mount *mp)
1191 struct xlog *log = mp->m_log;
1192 bool needed = false;
1194 if (!xlog_cil_empty(log))
1197 spin_lock(&log->l_icloglock);
1198 switch (log->l_covered_state) {
1199 case XLOG_STATE_COVER_DONE:
1200 case XLOG_STATE_COVER_DONE2:
1201 case XLOG_STATE_COVER_IDLE:
1203 case XLOG_STATE_COVER_NEED:
1204 case XLOG_STATE_COVER_NEED2:
1205 if (xfs_ail_min_lsn(log->l_ailp))
1207 if (!xlog_iclogs_empty(log))
1211 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1212 log->l_covered_state = XLOG_STATE_COVER_DONE;
1214 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1220 spin_unlock(&log->l_icloglock);
1225 * Explicitly cover the log. This is similar to background log covering but
1226 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1227 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1228 * must all be empty.
1232 struct xfs_mount *mp)
1237 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1238 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1239 xlog_is_shutdown(mp->m_log));
1241 if (!xfs_log_writable(mp))
1245 * xfs_log_need_covered() is not idempotent because it progresses the
1246 * state machine if the log requires covering. Therefore, we must call
1247 * this function once and use the result until we've issued an sb sync.
1248 * Do so first to make that abundantly clear.
1250 * Fall into the covering sequence if the log needs covering or the
1251 * mount has lazy superblock accounting to sync to disk. The sb sync
1252 * used for covering accumulates the in-core counters, so covering
1253 * handles this for us.
1255 need_covered = xfs_log_need_covered(mp);
1256 if (!need_covered && !xfs_has_lazysbcount(mp))
1260 * To cover the log, commit the superblock twice (at most) in
1261 * independent checkpoints. The first serves as a reference for the
1262 * tail pointer. The sync transaction and AIL push empties the AIL and
1263 * updates the in-core tail to the LSN of the first checkpoint. The
1264 * second commit updates the on-disk tail with the in-core LSN,
1265 * covering the log. Push the AIL one more time to leave it empty, as
1269 error = xfs_sync_sb(mp, true);
1272 xfs_ail_push_all_sync(mp->m_ail);
1273 } while (xfs_log_need_covered(mp));
1279 * We may be holding the log iclog lock upon entering this routine.
1282 xlog_assign_tail_lsn_locked(
1283 struct xfs_mount *mp)
1285 struct xlog *log = mp->m_log;
1286 struct xfs_log_item *lip;
1289 assert_spin_locked(&mp->m_ail->ail_lock);
1292 * To make sure we always have a valid LSN for the log tail we keep
1293 * track of the last LSN which was committed in log->l_last_sync_lsn,
1294 * and use that when the AIL was empty.
1296 lip = xfs_ail_min(mp->m_ail);
1298 tail_lsn = lip->li_lsn;
1300 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1301 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1302 atomic64_set(&log->l_tail_lsn, tail_lsn);
1307 xlog_assign_tail_lsn(
1308 struct xfs_mount *mp)
1312 spin_lock(&mp->m_ail->ail_lock);
1313 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1314 spin_unlock(&mp->m_ail->ail_lock);
1320 * Return the space in the log between the tail and the head. The head
1321 * is passed in the cycle/bytes formal parms. In the special case where
1322 * the reserve head has wrapped passed the tail, this calculation is no
1323 * longer valid. In this case, just return 0 which means there is no space
1324 * in the log. This works for all places where this function is called
1325 * with the reserve head. Of course, if the write head were to ever
1326 * wrap the tail, we should blow up. Rather than catch this case here,
1327 * we depend on other ASSERTions in other parts of the code. XXXmiken
1329 * If reservation head is behind the tail, we have a problem. Warn about it,
1330 * but then treat it as if the log is empty.
1332 * If the log is shut down, the head and tail may be invalid or out of whack, so
1333 * shortcut invalidity asserts in this case so that we don't trigger them
1346 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1347 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1348 tail_bytes = BBTOB(tail_bytes);
1349 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1350 return log->l_logsize - (head_bytes - tail_bytes);
1351 if (tail_cycle + 1 < head_cycle)
1354 /* Ignore potential inconsistency when shutdown. */
1355 if (xlog_is_shutdown(log))
1356 return log->l_logsize;
1358 if (tail_cycle < head_cycle) {
1359 ASSERT(tail_cycle == (head_cycle - 1));
1360 return tail_bytes - head_bytes;
1364 * The reservation head is behind the tail. In this case we just want to
1365 * return the size of the log as the amount of space left.
1367 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1368 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1369 tail_cycle, tail_bytes);
1370 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1371 head_cycle, head_bytes);
1373 return log->l_logsize;
1379 struct work_struct *work)
1381 struct xlog_in_core *iclog =
1382 container_of(work, struct xlog_in_core, ic_end_io_work);
1383 struct xlog *log = iclog->ic_log;
1386 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1388 /* treat writes with injected CRC errors as failed */
1389 if (iclog->ic_fail_crc)
1394 * Race to shutdown the filesystem if we see an error.
1396 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1397 xfs_alert(log->l_mp, "log I/O error %d", error);
1398 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1401 xlog_state_done_syncing(iclog);
1402 bio_uninit(&iclog->ic_bio);
1405 * Drop the lock to signal that we are done. Nothing references the
1406 * iclog after this, so an unmount waiting on this lock can now tear it
1407 * down safely. As such, it is unsafe to reference the iclog after the
1408 * unlock as we could race with it being freed.
1410 up(&iclog->ic_sema);
1414 * Return size of each in-core log record buffer.
1416 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1418 * If the filesystem blocksize is too large, we may need to choose a
1419 * larger size since the directory code currently logs entire blocks.
1422 xlog_get_iclog_buffer_size(
1423 struct xfs_mount *mp,
1426 if (mp->m_logbufs <= 0)
1427 mp->m_logbufs = XLOG_MAX_ICLOGS;
1428 if (mp->m_logbsize <= 0)
1429 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1431 log->l_iclog_bufs = mp->m_logbufs;
1432 log->l_iclog_size = mp->m_logbsize;
1435 * # headers = size / 32k - one header holds cycles from 32k of data.
1437 log->l_iclog_heads =
1438 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1439 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1444 struct xfs_mount *mp)
1446 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1447 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1451 * Clear the log incompat flags if we have the opportunity.
1453 * This only happens if we're about to log the second dummy transaction as part
1454 * of covering the log and we can get the log incompat feature usage lock.
1457 xlog_clear_incompat(
1460 struct xfs_mount *mp = log->l_mp;
1462 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1463 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1466 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1469 if (!down_write_trylock(&log->l_incompat_users))
1472 xfs_clear_incompat_log_features(mp);
1473 up_write(&log->l_incompat_users);
1477 * Every sync period we need to unpin all items in the AIL and push them to
1478 * disk. If there is nothing dirty, then we might need to cover the log to
1479 * indicate that the filesystem is idle.
1483 struct work_struct *work)
1485 struct xlog *log = container_of(to_delayed_work(work),
1486 struct xlog, l_work);
1487 struct xfs_mount *mp = log->l_mp;
1489 /* dgc: errors ignored - not fatal and nowhere to report them */
1490 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1492 * Dump a transaction into the log that contains no real change.
1493 * This is needed to stamp the current tail LSN into the log
1494 * during the covering operation.
1496 * We cannot use an inode here for this - that will push dirty
1497 * state back up into the VFS and then periodic inode flushing
1498 * will prevent log covering from making progress. Hence we
1499 * synchronously log the superblock instead to ensure the
1500 * superblock is immediately unpinned and can be written back.
1502 xlog_clear_incompat(log);
1503 xfs_sync_sb(mp, true);
1505 xfs_log_force(mp, 0);
1507 /* start pushing all the metadata that is currently dirty */
1508 xfs_ail_push_all(mp->m_ail);
1510 /* queue us up again */
1511 xfs_log_work_queue(mp);
1515 * This routine initializes some of the log structure for a given mount point.
1516 * Its primary purpose is to fill in enough, so recovery can occur. However,
1517 * some other stuff may be filled in too.
1519 STATIC struct xlog *
1521 struct xfs_mount *mp,
1522 struct xfs_buftarg *log_target,
1523 xfs_daddr_t blk_offset,
1527 xlog_rec_header_t *head;
1528 xlog_in_core_t **iclogp;
1529 xlog_in_core_t *iclog, *prev_iclog=NULL;
1531 int error = -ENOMEM;
1534 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1536 xfs_warn(mp, "Log allocation failed: No memory!");
1541 log->l_targ = log_target;
1542 log->l_logsize = BBTOB(num_bblks);
1543 log->l_logBBstart = blk_offset;
1544 log->l_logBBsize = num_bblks;
1545 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1546 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1547 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1549 log->l_prev_block = -1;
1550 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1551 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1552 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1553 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1555 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1556 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1558 log->l_iclog_roundoff = BBSIZE;
1560 xlog_grant_head_init(&log->l_reserve_head);
1561 xlog_grant_head_init(&log->l_write_head);
1563 error = -EFSCORRUPTED;
1564 if (xfs_has_sector(mp)) {
1565 log2_size = mp->m_sb.sb_logsectlog;
1566 if (log2_size < BBSHIFT) {
1567 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1568 log2_size, BBSHIFT);
1572 log2_size -= BBSHIFT;
1573 if (log2_size > mp->m_sectbb_log) {
1574 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1575 log2_size, mp->m_sectbb_log);
1579 /* for larger sector sizes, must have v2 or external log */
1580 if (log2_size && log->l_logBBstart > 0 &&
1581 !xfs_has_logv2(mp)) {
1583 "log sector size (0x%x) invalid for configuration.",
1588 log->l_sectBBsize = 1 << log2_size;
1590 init_rwsem(&log->l_incompat_users);
1592 xlog_get_iclog_buffer_size(mp, log);
1594 spin_lock_init(&log->l_icloglock);
1595 init_waitqueue_head(&log->l_flush_wait);
1597 iclogp = &log->l_iclog;
1599 * The amount of memory to allocate for the iclog structure is
1600 * rather funky due to the way the structure is defined. It is
1601 * done this way so that we can use different sizes for machines
1602 * with different amounts of memory. See the definition of
1603 * xlog_in_core_t in xfs_log_priv.h for details.
1605 ASSERT(log->l_iclog_size >= 4096);
1606 for (i = 0; i < log->l_iclog_bufs; i++) {
1607 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1608 sizeof(struct bio_vec);
1610 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1612 goto out_free_iclog;
1615 iclog->ic_prev = prev_iclog;
1618 iclog->ic_data = kvzalloc(log->l_iclog_size,
1619 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1620 if (!iclog->ic_data)
1621 goto out_free_iclog;
1622 head = &iclog->ic_header;
1623 memset(head, 0, sizeof(xlog_rec_header_t));
1624 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1625 head->h_version = cpu_to_be32(
1626 xfs_has_logv2(log->l_mp) ? 2 : 1);
1627 head->h_size = cpu_to_be32(log->l_iclog_size);
1629 head->h_fmt = cpu_to_be32(XLOG_FMT);
1630 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1632 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1633 iclog->ic_state = XLOG_STATE_ACTIVE;
1634 iclog->ic_log = log;
1635 atomic_set(&iclog->ic_refcnt, 0);
1636 INIT_LIST_HEAD(&iclog->ic_callbacks);
1637 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1639 init_waitqueue_head(&iclog->ic_force_wait);
1640 init_waitqueue_head(&iclog->ic_write_wait);
1641 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1642 sema_init(&iclog->ic_sema, 1);
1644 iclogp = &iclog->ic_next;
1646 *iclogp = log->l_iclog; /* complete ring */
1647 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1649 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1650 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1652 0, mp->m_super->s_id);
1653 if (!log->l_ioend_workqueue)
1654 goto out_free_iclog;
1656 error = xlog_cil_init(log);
1658 goto out_destroy_workqueue;
1661 out_destroy_workqueue:
1662 destroy_workqueue(log->l_ioend_workqueue);
1664 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1665 prev_iclog = iclog->ic_next;
1666 kmem_free(iclog->ic_data);
1668 if (prev_iclog == log->l_iclog)
1674 return ERR_PTR(error);
1675 } /* xlog_alloc_log */
1678 * Compute the LSN that we'd need to push the log tail towards in order to have
1679 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1680 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1681 * log free space already meets all three thresholds, this function returns
1685 xlog_grant_push_threshold(
1689 xfs_lsn_t threshold_lsn = 0;
1690 xfs_lsn_t last_sync_lsn;
1693 int threshold_block;
1694 int threshold_cycle;
1697 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1699 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1700 free_blocks = BTOBBT(free_bytes);
1703 * Set the threshold for the minimum number of free blocks in the
1704 * log to the maximum of what the caller needs, one quarter of the
1705 * log, and 256 blocks.
1707 free_threshold = BTOBB(need_bytes);
1708 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1709 free_threshold = max(free_threshold, 256);
1710 if (free_blocks >= free_threshold)
1711 return NULLCOMMITLSN;
1713 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1715 threshold_block += free_threshold;
1716 if (threshold_block >= log->l_logBBsize) {
1717 threshold_block -= log->l_logBBsize;
1718 threshold_cycle += 1;
1720 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1723 * Don't pass in an lsn greater than the lsn of the last
1724 * log record known to be on disk. Use a snapshot of the last sync lsn
1725 * so that it doesn't change between the compare and the set.
1727 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1728 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1729 threshold_lsn = last_sync_lsn;
1731 return threshold_lsn;
1735 * Push the tail of the log if we need to do so to maintain the free log space
1736 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1737 * policy which pushes on an lsn which is further along in the log once we
1738 * reach the high water mark. In this manner, we would be creating a low water
1742 xlog_grant_push_ail(
1746 xfs_lsn_t threshold_lsn;
1748 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1749 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1753 * Get the transaction layer to kick the dirty buffers out to
1754 * disk asynchronously. No point in trying to do this if
1755 * the filesystem is shutting down.
1757 xfs_ail_push(log->l_ailp, threshold_lsn);
1761 * Stamp cycle number in every block
1766 struct xlog_in_core *iclog,
1770 int size = iclog->ic_offset + roundoff;
1774 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1776 dp = iclog->ic_datap;
1777 for (i = 0; i < BTOBB(size); i++) {
1778 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1780 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1781 *(__be32 *)dp = cycle_lsn;
1785 if (xfs_has_logv2(log->l_mp)) {
1786 xlog_in_core_2_t *xhdr = iclog->ic_data;
1788 for ( ; i < BTOBB(size); i++) {
1789 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1790 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1791 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1792 *(__be32 *)dp = cycle_lsn;
1796 for (i = 1; i < log->l_iclog_heads; i++)
1797 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1802 * Calculate the checksum for a log buffer.
1804 * This is a little more complicated than it should be because the various
1805 * headers and the actual data are non-contiguous.
1810 struct xlog_rec_header *rhead,
1816 /* first generate the crc for the record header ... */
1817 crc = xfs_start_cksum_update((char *)rhead,
1818 sizeof(struct xlog_rec_header),
1819 offsetof(struct xlog_rec_header, h_crc));
1821 /* ... then for additional cycle data for v2 logs ... */
1822 if (xfs_has_logv2(log->l_mp)) {
1823 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1827 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1829 for (i = 1; i < xheads; i++) {
1830 crc = crc32c(crc, &xhdr[i].hic_xheader,
1831 sizeof(struct xlog_rec_ext_header));
1835 /* ... and finally for the payload */
1836 crc = crc32c(crc, dp, size);
1838 return xfs_end_cksum(crc);
1845 struct xlog_in_core *iclog = bio->bi_private;
1847 queue_work(iclog->ic_log->l_ioend_workqueue,
1848 &iclog->ic_end_io_work);
1852 xlog_map_iclog_data(
1858 struct page *page = kmem_to_page(data);
1859 unsigned int off = offset_in_page(data);
1860 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1862 if (bio_add_page(bio, page, len, off) != len)
1875 struct xlog_in_core *iclog,
1879 ASSERT(bno < log->l_logBBsize);
1880 trace_xlog_iclog_write(iclog, _RET_IP_);
1883 * We lock the iclogbufs here so that we can serialise against I/O
1884 * completion during unmount. We might be processing a shutdown
1885 * triggered during unmount, and that can occur asynchronously to the
1886 * unmount thread, and hence we need to ensure that completes before
1887 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1888 * across the log IO to archieve that.
1890 down(&iclog->ic_sema);
1891 if (xlog_is_shutdown(log)) {
1893 * It would seem logical to return EIO here, but we rely on
1894 * the log state machine to propagate I/O errors instead of
1895 * doing it here. We kick of the state machine and unlock
1896 * the buffer manually, the code needs to be kept in sync
1897 * with the I/O completion path.
1899 xlog_state_done_syncing(iclog);
1900 up(&iclog->ic_sema);
1905 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1906 * IOs coming immediately after this one. This prevents the block layer
1907 * writeback throttle from throttling log writes behind background
1908 * metadata writeback and causing priority inversions.
1910 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1911 howmany(count, PAGE_SIZE),
1912 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1913 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1914 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1915 iclog->ic_bio.bi_private = iclog;
1917 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1918 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1920 * For external log devices, we also need to flush the data
1921 * device cache first to ensure all metadata writeback covered
1922 * by the LSN in this iclog is on stable storage. This is slow,
1923 * but it *must* complete before we issue the external log IO.
1925 if (log->l_targ != log->l_mp->m_ddev_targp)
1926 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
1928 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1929 iclog->ic_bio.bi_opf |= REQ_FUA;
1931 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1933 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1934 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1937 if (is_vmalloc_addr(iclog->ic_data))
1938 flush_kernel_vmap_range(iclog->ic_data, count);
1941 * If this log buffer would straddle the end of the log we will have
1942 * to split it up into two bios, so that we can continue at the start.
1944 if (bno + BTOBB(count) > log->l_logBBsize) {
1947 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1948 GFP_NOIO, &fs_bio_set);
1949 bio_chain(split, &iclog->ic_bio);
1952 /* restart at logical offset zero for the remainder */
1953 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1956 submit_bio(&iclog->ic_bio);
1960 * We need to bump cycle number for the part of the iclog that is
1961 * written to the start of the log. Watch out for the header magic
1962 * number case, though.
1971 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1974 for (i = split_offset; i < count; i += BBSIZE) {
1975 uint32_t cycle = get_unaligned_be32(data + i);
1977 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1979 put_unaligned_be32(cycle, data + i);
1984 xlog_calc_iclog_size(
1986 struct xlog_in_core *iclog,
1989 uint32_t count_init, count;
1991 /* Add for LR header */
1992 count_init = log->l_iclog_hsize + iclog->ic_offset;
1993 count = roundup(count_init, log->l_iclog_roundoff);
1995 *roundoff = count - count_init;
1997 ASSERT(count >= count_init);
1998 ASSERT(*roundoff < log->l_iclog_roundoff);
2003 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2004 * fashion. Previously, we should have moved the current iclog
2005 * ptr in the log to point to the next available iclog. This allows further
2006 * write to continue while this code syncs out an iclog ready to go.
2007 * Before an in-core log can be written out, the data section must be scanned
2008 * to save away the 1st word of each BBSIZE block into the header. We replace
2009 * it with the current cycle count. Each BBSIZE block is tagged with the
2010 * cycle count because there in an implicit assumption that drives will
2011 * guarantee that entire 512 byte blocks get written at once. In other words,
2012 * we can't have part of a 512 byte block written and part not written. By
2013 * tagging each block, we will know which blocks are valid when recovering
2014 * after an unclean shutdown.
2016 * This routine is single threaded on the iclog. No other thread can be in
2017 * this routine with the same iclog. Changing contents of iclog can there-
2018 * fore be done without grabbing the state machine lock. Updating the global
2019 * log will require grabbing the lock though.
2021 * The entire log manager uses a logical block numbering scheme. Only
2022 * xlog_write_iclog knows about the fact that the log may not start with
2023 * block zero on a given device.
2028 struct xlog_in_core *iclog)
2030 unsigned int count; /* byte count of bwrite */
2031 unsigned int roundoff; /* roundoff to BB or stripe */
2035 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2036 trace_xlog_iclog_sync(iclog, _RET_IP_);
2038 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2040 /* move grant heads by roundoff in sync */
2041 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2042 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2044 /* put cycle number in every block */
2045 xlog_pack_data(log, iclog, roundoff);
2047 /* real byte length */
2048 size = iclog->ic_offset;
2049 if (xfs_has_logv2(log->l_mp))
2051 iclog->ic_header.h_len = cpu_to_be32(size);
2053 XFS_STATS_INC(log->l_mp, xs_log_writes);
2054 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2056 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2058 /* Do we need to split this write into 2 parts? */
2059 if (bno + BTOBB(count) > log->l_logBBsize)
2060 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2062 /* calculcate the checksum */
2063 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2064 iclog->ic_datap, size);
2066 * Intentionally corrupt the log record CRC based on the error injection
2067 * frequency, if defined. This facilitates testing log recovery in the
2068 * event of torn writes. Hence, set the IOABORT state to abort the log
2069 * write on I/O completion and shutdown the fs. The subsequent mount
2070 * detects the bad CRC and attempts to recover.
2073 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2074 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2075 iclog->ic_fail_crc = true;
2077 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2078 be64_to_cpu(iclog->ic_header.h_lsn));
2081 xlog_verify_iclog(log, iclog, count);
2082 xlog_write_iclog(log, iclog, bno, count);
2086 * Deallocate a log structure
2092 xlog_in_core_t *iclog, *next_iclog;
2096 * Cycle all the iclogbuf locks to make sure all log IO completion
2097 * is done before we tear down these buffers.
2099 iclog = log->l_iclog;
2100 for (i = 0; i < log->l_iclog_bufs; i++) {
2101 down(&iclog->ic_sema);
2102 up(&iclog->ic_sema);
2103 iclog = iclog->ic_next;
2107 * Destroy the CIL after waiting for iclog IO completion because an
2108 * iclog EIO error will try to shut down the log, which accesses the
2109 * CIL to wake up the waiters.
2111 xlog_cil_destroy(log);
2113 iclog = log->l_iclog;
2114 for (i = 0; i < log->l_iclog_bufs; i++) {
2115 next_iclog = iclog->ic_next;
2116 kmem_free(iclog->ic_data);
2121 log->l_mp->m_log = NULL;
2122 destroy_workqueue(log->l_ioend_workqueue);
2127 * Update counters atomically now that memcpy is done.
2130 xlog_state_finish_copy(
2132 struct xlog_in_core *iclog,
2136 lockdep_assert_held(&log->l_icloglock);
2138 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2139 iclog->ic_offset += copy_bytes;
2143 * print out info relating to regions written which consume
2148 struct xfs_mount *mp,
2149 struct xlog_ticket *ticket)
2151 xfs_warn(mp, "ticket reservation summary:");
2152 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
2153 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
2154 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
2155 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
2159 * Print a summary of the transaction.
2163 struct xfs_trans *tp)
2165 struct xfs_mount *mp = tp->t_mountp;
2166 struct xfs_log_item *lip;
2168 /* dump core transaction and ticket info */
2169 xfs_warn(mp, "transaction summary:");
2170 xfs_warn(mp, " log res = %d", tp->t_log_res);
2171 xfs_warn(mp, " log count = %d", tp->t_log_count);
2172 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2174 xlog_print_tic_res(mp, tp->t_ticket);
2176 /* dump each log item */
2177 list_for_each_entry(lip, &tp->t_items, li_trans) {
2178 struct xfs_log_vec *lv = lip->li_lv;
2179 struct xfs_log_iovec *vec;
2182 xfs_warn(mp, "log item: ");
2183 xfs_warn(mp, " type = 0x%x", lip->li_type);
2184 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2187 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2188 xfs_warn(mp, " size = %d", lv->lv_size);
2189 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2190 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2192 /* dump each iovec for the log item */
2193 vec = lv->lv_iovecp;
2194 for (i = 0; i < lv->lv_niovecs; i++) {
2195 int dumplen = min(vec->i_len, 32);
2197 xfs_warn(mp, " iovec[%d]", i);
2198 xfs_warn(mp, " type = 0x%x", vec->i_type);
2199 xfs_warn(mp, " len = %d", vec->i_len);
2200 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2201 xfs_hex_dump(vec->i_addr, dumplen);
2210 struct xlog_in_core *iclog,
2211 uint32_t *log_offset,
2215 uint32_t *record_cnt,
2218 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2219 ASSERT(*log_offset % sizeof(int32_t) == 0);
2220 ASSERT(write_len % sizeof(int32_t) == 0);
2222 memcpy(iclog->ic_datap + *log_offset, data, write_len);
2223 *log_offset += write_len;
2224 *bytes_left -= write_len;
2226 *data_cnt += write_len;
2230 * Write log vectors into a single iclog which is guaranteed by the caller
2231 * to have enough space to write the entire log vector into.
2235 struct xfs_log_vec *lv,
2236 struct xlog_ticket *ticket,
2237 struct xlog_in_core *iclog,
2238 uint32_t *log_offset,
2240 uint32_t *record_cnt,
2245 ASSERT(*log_offset + *len <= iclog->ic_size ||
2246 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2249 * Ordered log vectors have no regions to write so this
2250 * loop will naturally skip them.
2252 for (index = 0; index < lv->lv_niovecs; index++) {
2253 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2254 struct xlog_op_header *ophdr = reg->i_addr;
2256 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2257 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2258 reg->i_len, len, record_cnt, data_cnt);
2263 xlog_write_get_more_iclog_space(
2264 struct xlog_ticket *ticket,
2265 struct xlog_in_core **iclogp,
2266 uint32_t *log_offset,
2268 uint32_t *record_cnt,
2271 struct xlog_in_core *iclog = *iclogp;
2272 struct xlog *log = iclog->ic_log;
2275 spin_lock(&log->l_icloglock);
2276 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2277 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2278 error = xlog_state_release_iclog(log, iclog);
2279 spin_unlock(&log->l_icloglock);
2283 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2294 * Write log vectors into a single iclog which is smaller than the current chain
2295 * length. We write until we cannot fit a full record into the remaining space
2296 * and then stop. We return the log vector that is to be written that cannot
2297 * wholly fit in the iclog.
2301 struct xfs_log_vec *lv,
2302 struct xlog_ticket *ticket,
2303 struct xlog_in_core **iclogp,
2304 uint32_t *log_offset,
2306 uint32_t *record_cnt,
2309 struct xlog_in_core *iclog = *iclogp;
2310 struct xlog_op_header *ophdr;
2315 /* walk the logvec, copying until we run out of space in the iclog */
2316 for (index = 0; index < lv->lv_niovecs; index++) {
2317 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2318 uint32_t reg_offset = 0;
2321 * The first region of a continuation must have a non-zero
2322 * length otherwise log recovery will just skip over it and
2323 * start recovering from the next opheader it finds. Because we
2324 * mark the next opheader as a continuation, recovery will then
2325 * incorrectly add the continuation to the previous region and
2326 * that breaks stuff.
2328 * Hence if there isn't space for region data after the
2329 * opheader, then we need to start afresh with a new iclog.
2331 if (iclog->ic_size - *log_offset <=
2332 sizeof(struct xlog_op_header)) {
2333 error = xlog_write_get_more_iclog_space(ticket,
2334 &iclog, log_offset, *len, record_cnt,
2340 ophdr = reg->i_addr;
2341 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2343 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2344 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2345 if (rlen != reg->i_len)
2346 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2348 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2349 rlen, len, record_cnt, data_cnt);
2351 /* If we wrote the whole region, move to the next. */
2352 if (rlen == reg->i_len)
2356 * We now have a partially written iovec, but it can span
2357 * multiple iclogs so we loop here. First we release the iclog
2358 * we currently have, then we get a new iclog and add a new
2359 * opheader. Then we continue copying from where we were until
2360 * we either complete the iovec or fill the iclog. If we
2361 * complete the iovec, then we increment the index and go right
2362 * back to the top of the outer loop. if we fill the iclog, we
2363 * run the inner loop again.
2365 * This is complicated by the tail of a region using all the
2366 * space in an iclog and hence requiring us to release the iclog
2367 * and get a new one before returning to the outer loop. We must
2368 * always guarantee that we exit this inner loop with at least
2369 * space for log transaction opheaders left in the current
2370 * iclog, hence we cannot just terminate the loop at the end
2371 * of the of the continuation. So we loop while there is no
2372 * space left in the current iclog, and check for the end of the
2373 * continuation after getting a new iclog.
2377 * Ensure we include the continuation opheader in the
2378 * space we need in the new iclog by adding that size
2379 * to the length we require. This continuation opheader
2380 * needs to be accounted to the ticket as the space it
2381 * consumes hasn't been accounted to the lv we are
2384 error = xlog_write_get_more_iclog_space(ticket,
2386 *len + sizeof(struct xlog_op_header),
2387 record_cnt, data_cnt);
2391 ophdr = iclog->ic_datap + *log_offset;
2392 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2393 ophdr->oh_clientid = XFS_TRANSACTION;
2395 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2397 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2398 *log_offset += sizeof(struct xlog_op_header);
2399 *data_cnt += sizeof(struct xlog_op_header);
2402 * If rlen fits in the iclog, then end the region
2403 * continuation. Otherwise we're going around again.
2406 rlen = reg->i_len - reg_offset;
2407 if (rlen <= iclog->ic_size - *log_offset)
2408 ophdr->oh_flags |= XLOG_END_TRANS;
2410 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2412 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2413 ophdr->oh_len = cpu_to_be32(rlen);
2415 xlog_write_iovec(iclog, log_offset,
2416 reg->i_addr + reg_offset,
2417 rlen, len, record_cnt, data_cnt);
2419 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2423 * No more iovecs remain in this logvec so return the next log vec to
2424 * the caller so it can go back to fast path copying.
2431 * Write some region out to in-core log
2433 * This will be called when writing externally provided regions or when
2434 * writing out a commit record for a given transaction.
2436 * General algorithm:
2437 * 1. Find total length of this write. This may include adding to the
2438 * lengths passed in.
2439 * 2. Check whether we violate the tickets reservation.
2440 * 3. While writing to this iclog
2441 * A. Reserve as much space in this iclog as can get
2442 * B. If this is first write, save away start lsn
2443 * C. While writing this region:
2444 * 1. If first write of transaction, write start record
2445 * 2. Write log operation header (header per region)
2446 * 3. Find out if we can fit entire region into this iclog
2447 * 4. Potentially, verify destination memcpy ptr
2448 * 5. Memcpy (partial) region
2449 * 6. If partial copy, release iclog; otherwise, continue
2450 * copying more regions into current iclog
2451 * 4. Mark want sync bit (in simulation mode)
2452 * 5. Release iclog for potential flush to on-disk log.
2455 * 1. Panic if reservation is overrun. This should never happen since
2456 * reservation amounts are generated internal to the filesystem.
2458 * 1. Tickets are single threaded data structures.
2459 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2460 * syncing routine. When a single log_write region needs to span
2461 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2462 * on all log operation writes which don't contain the end of the
2463 * region. The XLOG_END_TRANS bit is used for the in-core log
2464 * operation which contains the end of the continued log_write region.
2465 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2466 * we don't really know exactly how much space will be used. As a result,
2467 * we don't update ic_offset until the end when we know exactly how many
2468 * bytes have been written out.
2473 struct xfs_cil_ctx *ctx,
2474 struct xfs_log_vec *log_vector,
2475 struct xlog_ticket *ticket,
2479 struct xlog_in_core *iclog = NULL;
2480 struct xfs_log_vec *lv = log_vector;
2481 uint32_t record_cnt = 0;
2482 uint32_t data_cnt = 0;
2486 if (ticket->t_curr_res < 0) {
2487 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2488 "ctx ticket reservation ran out. Need to up reservation");
2489 xlog_print_tic_res(log->l_mp, ticket);
2490 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2493 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2498 ASSERT(log_offset <= iclog->ic_size - 1);
2501 * If we have a context pointer, pass it the first iclog we are
2502 * writing to so it can record state needed for iclog write
2506 xlog_cil_set_ctx_write_state(ctx, iclog);
2510 * If the entire log vec does not fit in the iclog, punt it to
2511 * the partial copy loop which can handle this case.
2513 if (lv->lv_niovecs &&
2514 lv->lv_bytes > iclog->ic_size - log_offset) {
2515 error = xlog_write_partial(lv, ticket, &iclog,
2516 &log_offset, &len, &record_cnt,
2520 * We have no iclog to release, so just return
2521 * the error immediately.
2526 xlog_write_full(lv, ticket, iclog, &log_offset,
2527 &len, &record_cnt, &data_cnt);
2534 * We've already been guaranteed that the last writes will fit inside
2535 * the current iclog, and hence it will already have the space used by
2536 * those writes accounted to it. Hence we do not need to update the
2537 * iclog with the number of bytes written here.
2539 spin_lock(&log->l_icloglock);
2540 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2541 error = xlog_state_release_iclog(log, iclog);
2542 spin_unlock(&log->l_icloglock);
2548 xlog_state_activate_iclog(
2549 struct xlog_in_core *iclog,
2550 int *iclogs_changed)
2552 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2553 trace_xlog_iclog_activate(iclog, _RET_IP_);
2556 * If the number of ops in this iclog indicate it just contains the
2557 * dummy transaction, we can change state into IDLE (the second time
2558 * around). Otherwise we should change the state into NEED a dummy.
2559 * We don't need to cover the dummy.
2561 if (*iclogs_changed == 0 &&
2562 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2563 *iclogs_changed = 1;
2566 * We have two dirty iclogs so start over. This could also be
2567 * num of ops indicating this is not the dummy going out.
2569 *iclogs_changed = 2;
2572 iclog->ic_state = XLOG_STATE_ACTIVE;
2573 iclog->ic_offset = 0;
2574 iclog->ic_header.h_num_logops = 0;
2575 memset(iclog->ic_header.h_cycle_data, 0,
2576 sizeof(iclog->ic_header.h_cycle_data));
2577 iclog->ic_header.h_lsn = 0;
2578 iclog->ic_header.h_tail_lsn = 0;
2582 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2583 * ACTIVE after iclog I/O has completed.
2586 xlog_state_activate_iclogs(
2588 int *iclogs_changed)
2590 struct xlog_in_core *iclog = log->l_iclog;
2593 if (iclog->ic_state == XLOG_STATE_DIRTY)
2594 xlog_state_activate_iclog(iclog, iclogs_changed);
2596 * The ordering of marking iclogs ACTIVE must be maintained, so
2597 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2599 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2601 } while ((iclog = iclog->ic_next) != log->l_iclog);
2610 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2611 * wrote the first covering record (DONE). We go to IDLE if we just
2612 * wrote the second covering record (DONE2) and remain in IDLE until a
2613 * non-covering write occurs.
2615 switch (prev_state) {
2616 case XLOG_STATE_COVER_IDLE:
2617 if (iclogs_changed == 1)
2618 return XLOG_STATE_COVER_IDLE;
2620 case XLOG_STATE_COVER_NEED:
2621 case XLOG_STATE_COVER_NEED2:
2623 case XLOG_STATE_COVER_DONE:
2624 if (iclogs_changed == 1)
2625 return XLOG_STATE_COVER_NEED2;
2627 case XLOG_STATE_COVER_DONE2:
2628 if (iclogs_changed == 1)
2629 return XLOG_STATE_COVER_IDLE;
2635 return XLOG_STATE_COVER_NEED;
2639 xlog_state_clean_iclog(
2641 struct xlog_in_core *dirty_iclog)
2643 int iclogs_changed = 0;
2645 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2647 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2649 xlog_state_activate_iclogs(log, &iclogs_changed);
2650 wake_up_all(&dirty_iclog->ic_force_wait);
2652 if (iclogs_changed) {
2653 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2659 xlog_get_lowest_lsn(
2662 struct xlog_in_core *iclog = log->l_iclog;
2663 xfs_lsn_t lowest_lsn = 0, lsn;
2666 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2667 iclog->ic_state == XLOG_STATE_DIRTY)
2670 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2671 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2673 } while ((iclog = iclog->ic_next) != log->l_iclog);
2679 * Completion of a iclog IO does not imply that a transaction has completed, as
2680 * transactions can be large enough to span many iclogs. We cannot change the
2681 * tail of the log half way through a transaction as this may be the only
2682 * transaction in the log and moving the tail to point to the middle of it
2683 * will prevent recovery from finding the start of the transaction. Hence we
2684 * should only update the last_sync_lsn if this iclog contains transaction
2685 * completion callbacks on it.
2687 * We have to do this before we drop the icloglock to ensure we are the only one
2688 * that can update it.
2690 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2691 * the reservation grant head pushing. This is due to the fact that the push
2692 * target is bound by the current last_sync_lsn value. Hence if we have a large
2693 * amount of log space bound up in this committing transaction then the
2694 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2695 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2696 * should push the AIL to ensure the push target (and hence the grant head) is
2697 * no longer bound by the old log head location and can move forwards and make
2701 xlog_state_set_callback(
2703 struct xlog_in_core *iclog,
2704 xfs_lsn_t header_lsn)
2706 trace_xlog_iclog_callback(iclog, _RET_IP_);
2707 iclog->ic_state = XLOG_STATE_CALLBACK;
2709 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2712 if (list_empty_careful(&iclog->ic_callbacks))
2715 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2716 xlog_grant_push_ail(log, 0);
2720 * Return true if we need to stop processing, false to continue to the next
2721 * iclog. The caller will need to run callbacks if the iclog is returned in the
2722 * XLOG_STATE_CALLBACK state.
2725 xlog_state_iodone_process_iclog(
2727 struct xlog_in_core *iclog)
2729 xfs_lsn_t lowest_lsn;
2730 xfs_lsn_t header_lsn;
2732 switch (iclog->ic_state) {
2733 case XLOG_STATE_ACTIVE:
2734 case XLOG_STATE_DIRTY:
2736 * Skip all iclogs in the ACTIVE & DIRTY states:
2739 case XLOG_STATE_DONE_SYNC:
2741 * Now that we have an iclog that is in the DONE_SYNC state, do
2742 * one more check here to see if we have chased our tail around.
2743 * If this is not the lowest lsn iclog, then we will leave it
2744 * for another completion to process.
2746 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2747 lowest_lsn = xlog_get_lowest_lsn(log);
2748 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2750 xlog_state_set_callback(log, iclog, header_lsn);
2754 * Can only perform callbacks in order. Since this iclog is not
2755 * in the DONE_SYNC state, we skip the rest and just try to
2763 * Loop over all the iclogs, running attached callbacks on them. Return true if
2764 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2765 * to handle transient shutdown state here at all because
2766 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2767 * cleanup of the callbacks.
2770 xlog_state_do_iclog_callbacks(
2772 __releases(&log->l_icloglock)
2773 __acquires(&log->l_icloglock)
2775 struct xlog_in_core *first_iclog = log->l_iclog;
2776 struct xlog_in_core *iclog = first_iclog;
2777 bool ran_callback = false;
2782 if (xlog_state_iodone_process_iclog(log, iclog))
2784 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2785 iclog = iclog->ic_next;
2788 list_splice_init(&iclog->ic_callbacks, &cb_list);
2789 spin_unlock(&log->l_icloglock);
2791 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2792 xlog_cil_process_committed(&cb_list);
2793 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2794 ran_callback = true;
2796 spin_lock(&log->l_icloglock);
2797 xlog_state_clean_iclog(log, iclog);
2798 iclog = iclog->ic_next;
2799 } while (iclog != first_iclog);
2801 return ran_callback;
2806 * Loop running iclog completion callbacks until there are no more iclogs in a
2807 * state that can run callbacks.
2810 xlog_state_do_callback(
2816 spin_lock(&log->l_icloglock);
2817 while (xlog_state_do_iclog_callbacks(log)) {
2818 if (xlog_is_shutdown(log))
2821 if (++repeats > 5000) {
2822 flushcnt += repeats;
2825 "%s: possible infinite loop (%d iterations)",
2826 __func__, flushcnt);
2830 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2831 wake_up_all(&log->l_flush_wait);
2833 spin_unlock(&log->l_icloglock);
2838 * Finish transitioning this iclog to the dirty state.
2840 * Callbacks could take time, so they are done outside the scope of the
2841 * global state machine log lock.
2844 xlog_state_done_syncing(
2845 struct xlog_in_core *iclog)
2847 struct xlog *log = iclog->ic_log;
2849 spin_lock(&log->l_icloglock);
2850 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2851 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2854 * If we got an error, either on the first buffer, or in the case of
2855 * split log writes, on the second, we shut down the file system and
2856 * no iclogs should ever be attempted to be written to disk again.
2858 if (!xlog_is_shutdown(log)) {
2859 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2860 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2864 * Someone could be sleeping prior to writing out the next
2865 * iclog buffer, we wake them all, one will get to do the
2866 * I/O, the others get to wait for the result.
2868 wake_up_all(&iclog->ic_write_wait);
2869 spin_unlock(&log->l_icloglock);
2870 xlog_state_do_callback(log);
2874 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2875 * sleep. We wait on the flush queue on the head iclog as that should be
2876 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2877 * we will wait here and all new writes will sleep until a sync completes.
2879 * The in-core logs are used in a circular fashion. They are not used
2880 * out-of-order even when an iclog past the head is free.
2883 * * log_offset where xlog_write() can start writing into the in-core
2885 * * in-core log pointer to which xlog_write() should write.
2886 * * boolean indicating this is a continued write to an in-core log.
2887 * If this is the last write, then the in-core log's offset field
2888 * needs to be incremented, depending on the amount of data which
2892 xlog_state_get_iclog_space(
2895 struct xlog_in_core **iclogp,
2896 struct xlog_ticket *ticket,
2900 xlog_rec_header_t *head;
2901 xlog_in_core_t *iclog;
2904 spin_lock(&log->l_icloglock);
2905 if (xlog_is_shutdown(log)) {
2906 spin_unlock(&log->l_icloglock);
2910 iclog = log->l_iclog;
2911 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2912 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2914 /* Wait for log writes to have flushed */
2915 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2919 head = &iclog->ic_header;
2921 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2922 log_offset = iclog->ic_offset;
2924 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2926 /* On the 1st write to an iclog, figure out lsn. This works
2927 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2928 * committing to. If the offset is set, that's how many blocks
2931 if (log_offset == 0) {
2932 ticket->t_curr_res -= log->l_iclog_hsize;
2933 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2934 head->h_lsn = cpu_to_be64(
2935 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2936 ASSERT(log->l_curr_block >= 0);
2939 /* If there is enough room to write everything, then do it. Otherwise,
2940 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2941 * bit is on, so this will get flushed out. Don't update ic_offset
2942 * until you know exactly how many bytes get copied. Therefore, wait
2943 * until later to update ic_offset.
2945 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2946 * can fit into remaining data section.
2948 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2951 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2954 * If we are the only one writing to this iclog, sync it to
2955 * disk. We need to do an atomic compare and decrement here to
2956 * avoid racing with concurrent atomic_dec_and_lock() calls in
2957 * xlog_state_release_iclog() when there is more than one
2958 * reference to the iclog.
2960 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2961 error = xlog_state_release_iclog(log, iclog);
2962 spin_unlock(&log->l_icloglock);
2968 /* Do we have enough room to write the full amount in the remainder
2969 * of this iclog? Or must we continue a write on the next iclog and
2970 * mark this iclog as completely taken? In the case where we switch
2971 * iclogs (to mark it taken), this particular iclog will release/sync
2972 * to disk in xlog_write().
2974 if (len <= iclog->ic_size - iclog->ic_offset)
2975 iclog->ic_offset += len;
2977 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2980 ASSERT(iclog->ic_offset <= iclog->ic_size);
2981 spin_unlock(&log->l_icloglock);
2983 *logoffsetp = log_offset;
2988 * The first cnt-1 times a ticket goes through here we don't need to move the
2989 * grant write head because the permanent reservation has reserved cnt times the
2990 * unit amount. Release part of current permanent unit reservation and reset
2991 * current reservation to be one units worth. Also move grant reservation head
2995 xfs_log_ticket_regrant(
2997 struct xlog_ticket *ticket)
2999 trace_xfs_log_ticket_regrant(log, ticket);
3001 if (ticket->t_cnt > 0)
3004 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3005 ticket->t_curr_res);
3006 xlog_grant_sub_space(log, &log->l_write_head.grant,
3007 ticket->t_curr_res);
3008 ticket->t_curr_res = ticket->t_unit_res;
3010 trace_xfs_log_ticket_regrant_sub(log, ticket);
3012 /* just return if we still have some of the pre-reserved space */
3013 if (!ticket->t_cnt) {
3014 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3015 ticket->t_unit_res);
3016 trace_xfs_log_ticket_regrant_exit(log, ticket);
3018 ticket->t_curr_res = ticket->t_unit_res;
3021 xfs_log_ticket_put(ticket);
3025 * Give back the space left from a reservation.
3027 * All the information we need to make a correct determination of space left
3028 * is present. For non-permanent reservations, things are quite easy. The
3029 * count should have been decremented to zero. We only need to deal with the
3030 * space remaining in the current reservation part of the ticket. If the
3031 * ticket contains a permanent reservation, there may be left over space which
3032 * needs to be released. A count of N means that N-1 refills of the current
3033 * reservation can be done before we need to ask for more space. The first
3034 * one goes to fill up the first current reservation. Once we run out of
3035 * space, the count will stay at zero and the only space remaining will be
3036 * in the current reservation field.
3039 xfs_log_ticket_ungrant(
3041 struct xlog_ticket *ticket)
3045 trace_xfs_log_ticket_ungrant(log, ticket);
3047 if (ticket->t_cnt > 0)
3050 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3053 * If this is a permanent reservation ticket, we may be able to free
3054 * up more space based on the remaining count.
3056 bytes = ticket->t_curr_res;
3057 if (ticket->t_cnt > 0) {
3058 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3059 bytes += ticket->t_unit_res*ticket->t_cnt;
3062 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3063 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3065 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3067 xfs_log_space_wake(log->l_mp);
3068 xfs_log_ticket_put(ticket);
3072 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3073 * the current iclog pointer to the next iclog in the ring.
3076 xlog_state_switch_iclogs(
3078 struct xlog_in_core *iclog,
3081 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3082 assert_spin_locked(&log->l_icloglock);
3083 trace_xlog_iclog_switch(iclog, _RET_IP_);
3086 eventual_size = iclog->ic_offset;
3087 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3088 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3089 log->l_prev_block = log->l_curr_block;
3090 log->l_prev_cycle = log->l_curr_cycle;
3092 /* roll log?: ic_offset changed later */
3093 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3095 /* Round up to next log-sunit */
3096 if (log->l_iclog_roundoff > BBSIZE) {
3097 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3098 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3101 if (log->l_curr_block >= log->l_logBBsize) {
3103 * Rewind the current block before the cycle is bumped to make
3104 * sure that the combined LSN never transiently moves forward
3105 * when the log wraps to the next cycle. This is to support the
3106 * unlocked sample of these fields from xlog_valid_lsn(). Most
3107 * other cases should acquire l_icloglock.
3109 log->l_curr_block -= log->l_logBBsize;
3110 ASSERT(log->l_curr_block >= 0);
3112 log->l_curr_cycle++;
3113 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3114 log->l_curr_cycle++;
3116 ASSERT(iclog == log->l_iclog);
3117 log->l_iclog = iclog->ic_next;
3121 * Force the iclog to disk and check if the iclog has been completed before
3122 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3123 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3124 * If completion has already occurred, tell the caller so that it can avoid an
3125 * unnecessary wait on the iclog.
3128 xlog_force_and_check_iclog(
3129 struct xlog_in_core *iclog,
3132 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3136 error = xlog_force_iclog(iclog);
3141 * If the iclog has already been completed and reused the header LSN
3142 * will have been rewritten by completion
3144 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3150 * Write out all data in the in-core log as of this exact moment in time.
3152 * Data may be written to the in-core log during this call. However,
3153 * we don't guarantee this data will be written out. A change from past
3154 * implementation means this routine will *not* write out zero length LRs.
3156 * Basically, we try and perform an intelligent scan of the in-core logs.
3157 * If we determine there is no flushable data, we just return. There is no
3158 * flushable data if:
3160 * 1. the current iclog is active and has no data; the previous iclog
3161 * is in the active or dirty state.
3162 * 2. the current iclog is drity, and the previous iclog is in the
3163 * active or dirty state.
3167 * 1. the current iclog is not in the active nor dirty state.
3168 * 2. the current iclog dirty, and the previous iclog is not in the
3169 * active nor dirty state.
3170 * 3. the current iclog is active, and there is another thread writing
3171 * to this particular iclog.
3172 * 4. a) the current iclog is active and has no other writers
3173 * b) when we return from flushing out this iclog, it is still
3174 * not in the active nor dirty state.
3178 struct xfs_mount *mp,
3181 struct xlog *log = mp->m_log;
3182 struct xlog_in_core *iclog;
3184 XFS_STATS_INC(mp, xs_log_force);
3185 trace_xfs_log_force(mp, 0, _RET_IP_);
3187 xlog_cil_force(log);
3189 spin_lock(&log->l_icloglock);
3190 if (xlog_is_shutdown(log))
3193 iclog = log->l_iclog;
3194 trace_xlog_iclog_force(iclog, _RET_IP_);
3196 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3197 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3198 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3200 * If the head is dirty or (active and empty), then we need to
3201 * look at the previous iclog.
3203 * If the previous iclog is active or dirty we are done. There
3204 * is nothing to sync out. Otherwise, we attach ourselves to the
3205 * previous iclog and go to sleep.
3207 iclog = iclog->ic_prev;
3208 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3209 if (atomic_read(&iclog->ic_refcnt) == 0) {
3210 /* We have exclusive access to this iclog. */
3213 if (xlog_force_and_check_iclog(iclog, &completed))
3220 * Someone else is still writing to this iclog, so we
3221 * need to ensure that when they release the iclog it
3222 * gets synced immediately as we may be waiting on it.
3224 xlog_state_switch_iclogs(log, iclog, 0);
3229 * The iclog we are about to wait on may contain the checkpoint pushed
3230 * by the above xlog_cil_force() call, but it may not have been pushed
3231 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3232 * are flushed when this iclog is written.
3234 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3235 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3237 if (flags & XFS_LOG_SYNC)
3238 return xlog_wait_on_iclog(iclog);
3240 spin_unlock(&log->l_icloglock);
3243 spin_unlock(&log->l_icloglock);
3248 * Force the log to a specific LSN.
3250 * If an iclog with that lsn can be found:
3251 * If it is in the DIRTY state, just return.
3252 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3253 * state and go to sleep or return.
3254 * If it is in any other state, go to sleep or return.
3256 * Synchronous forces are implemented with a wait queue. All callers trying
3257 * to force a given lsn to disk must wait on the queue attached to the
3258 * specific in-core log. When given in-core log finally completes its write
3259 * to disk, that thread will wake up all threads waiting on the queue.
3269 struct xlog_in_core *iclog;
3272 spin_lock(&log->l_icloglock);
3273 if (xlog_is_shutdown(log))
3276 iclog = log->l_iclog;
3277 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3278 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3279 iclog = iclog->ic_next;
3280 if (iclog == log->l_iclog)
3284 switch (iclog->ic_state) {
3285 case XLOG_STATE_ACTIVE:
3287 * We sleep here if we haven't already slept (e.g. this is the
3288 * first time we've looked at the correct iclog buf) and the
3289 * buffer before us is going to be sync'ed. The reason for this
3290 * is that if we are doing sync transactions here, by waiting
3291 * for the previous I/O to complete, we can allow a few more
3292 * transactions into this iclog before we close it down.
3294 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3295 * refcnt so we can release the log (which drops the ref count).
3296 * The state switch keeps new transaction commits from using
3297 * this buffer. When the current commits finish writing into
3298 * the buffer, the refcount will drop to zero and the buffer
3301 if (!already_slept &&
3302 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3303 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3304 xlog_wait(&iclog->ic_prev->ic_write_wait,
3308 if (xlog_force_and_check_iclog(iclog, &completed))
3315 case XLOG_STATE_WANT_SYNC:
3317 * This iclog may contain the checkpoint pushed by the
3318 * xlog_cil_force_seq() call, but there are other writers still
3319 * accessing it so it hasn't been pushed to disk yet. Like the
3320 * ACTIVE case above, we need to make sure caches are flushed
3321 * when this iclog is written.
3323 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3327 * The entire checkpoint was written by the CIL force and is on
3328 * its way to disk already. It will be stable when it
3329 * completes, so we don't need to manipulate caches here at all.
3330 * We just need to wait for completion if necessary.
3335 if (flags & XFS_LOG_SYNC)
3336 return xlog_wait_on_iclog(iclog);
3338 spin_unlock(&log->l_icloglock);
3341 spin_unlock(&log->l_icloglock);
3346 * Force the log to a specific checkpoint sequence.
3348 * First force the CIL so that all the required changes have been flushed to the
3349 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3350 * the iclog that needs to be flushed to stable storage. If the caller needs
3351 * a synchronous log force, we will wait on the iclog with the LSN returned by
3352 * xlog_cil_force_seq() to be completed.
3356 struct xfs_mount *mp,
3361 struct xlog *log = mp->m_log;
3366 XFS_STATS_INC(mp, xs_log_force);
3367 trace_xfs_log_force(mp, seq, _RET_IP_);
3369 lsn = xlog_cil_force_seq(log, seq);
3370 if (lsn == NULLCOMMITLSN)
3373 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3374 if (ret == -EAGAIN) {
3375 XFS_STATS_INC(mp, xs_log_force_sleep);
3376 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3382 * Free a used ticket when its refcount falls to zero.
3386 xlog_ticket_t *ticket)
3388 ASSERT(atomic_read(&ticket->t_ref) > 0);
3389 if (atomic_dec_and_test(&ticket->t_ref))
3390 kmem_cache_free(xfs_log_ticket_cache, ticket);
3395 xlog_ticket_t *ticket)
3397 ASSERT(atomic_read(&ticket->t_ref) > 0);
3398 atomic_inc(&ticket->t_ref);
3403 * Figure out the total log space unit (in bytes) that would be
3404 * required for a log ticket.
3416 * Permanent reservations have up to 'cnt'-1 active log operations
3417 * in the log. A unit in this case is the amount of space for one
3418 * of these log operations. Normal reservations have a cnt of 1
3419 * and their unit amount is the total amount of space required.
3421 * The following lines of code account for non-transaction data
3422 * which occupy space in the on-disk log.
3424 * Normal form of a transaction is:
3425 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3426 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3428 * We need to account for all the leadup data and trailer data
3429 * around the transaction data.
3430 * And then we need to account for the worst case in terms of using
3432 * The worst case will happen if:
3433 * - the placement of the transaction happens to be such that the
3434 * roundoff is at its maximum
3435 * - the transaction data is synced before the commit record is synced
3436 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3437 * Therefore the commit record is in its own Log Record.
3438 * This can happen as the commit record is called with its
3439 * own region to xlog_write().
3440 * This then means that in the worst case, roundoff can happen for
3441 * the commit-rec as well.
3442 * The commit-rec is smaller than padding in this scenario and so it is
3443 * not added separately.
3446 /* for trans header */
3447 unit_bytes += sizeof(xlog_op_header_t);
3448 unit_bytes += sizeof(xfs_trans_header_t);
3451 unit_bytes += sizeof(xlog_op_header_t);
3454 * for LR headers - the space for data in an iclog is the size minus
3455 * the space used for the headers. If we use the iclog size, then we
3456 * undercalculate the number of headers required.
3458 * Furthermore - the addition of op headers for split-recs might
3459 * increase the space required enough to require more log and op
3460 * headers, so take that into account too.
3462 * IMPORTANT: This reservation makes the assumption that if this
3463 * transaction is the first in an iclog and hence has the LR headers
3464 * accounted to it, then the remaining space in the iclog is
3465 * exclusively for this transaction. i.e. if the transaction is larger
3466 * than the iclog, it will be the only thing in that iclog.
3467 * Fundamentally, this means we must pass the entire log vector to
3468 * xlog_write to guarantee this.
3470 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3471 num_headers = howmany(unit_bytes, iclog_space);
3473 /* for split-recs - ophdrs added when data split over LRs */
3474 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3476 /* add extra header reservations if we overrun */
3477 while (!num_headers ||
3478 howmany(unit_bytes, iclog_space) > num_headers) {
3479 unit_bytes += sizeof(xlog_op_header_t);
3482 unit_bytes += log->l_iclog_hsize * num_headers;
3484 /* for commit-rec LR header - note: padding will subsume the ophdr */
3485 unit_bytes += log->l_iclog_hsize;
3487 /* roundoff padding for transaction data and one for commit record */
3488 unit_bytes += 2 * log->l_iclog_roundoff;
3491 *niclogs = num_headers;
3496 xfs_log_calc_unit_res(
3497 struct xfs_mount *mp,
3500 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3504 * Allocate and initialise a new log ticket.
3506 struct xlog_ticket *
3513 struct xlog_ticket *tic;
3516 tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3518 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3520 atomic_set(&tic->t_ref, 1);
3521 tic->t_task = current;
3522 INIT_LIST_HEAD(&tic->t_queue);
3523 tic->t_unit_res = unit_res;
3524 tic->t_curr_res = unit_res;
3527 tic->t_tid = prandom_u32();
3529 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3536 * Check to make sure the grant write head didn't just over lap the tail. If
3537 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3538 * the cycles differ by exactly one and check the byte count.
3540 * This check is run unlocked, so can give false positives. Rather than assert
3541 * on failures, use a warn-once flag and a panic tag to allow the admin to
3542 * determine if they want to panic the machine when such an error occurs. For
3543 * debug kernels this will have the same effect as using an assert but, unlinke
3544 * an assert, it can be turned off at runtime.
3547 xlog_verify_grant_tail(
3550 int tail_cycle, tail_blocks;
3553 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3554 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3555 if (tail_cycle != cycle) {
3556 if (cycle - 1 != tail_cycle &&
3557 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3558 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3559 "%s: cycle - 1 != tail_cycle", __func__);
3562 if (space > BBTOB(tail_blocks) &&
3563 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3564 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3565 "%s: space > BBTOB(tail_blocks)", __func__);
3570 /* check if it will fit */
3572 xlog_verify_tail_lsn(
3574 struct xlog_in_core *iclog)
3576 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3579 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3581 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3582 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3583 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3585 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3587 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3588 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3590 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3591 if (blocks < BTOBB(iclog->ic_offset) + 1)
3592 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3597 * Perform a number of checks on the iclog before writing to disk.
3599 * 1. Make sure the iclogs are still circular
3600 * 2. Make sure we have a good magic number
3601 * 3. Make sure we don't have magic numbers in the data
3602 * 4. Check fields of each log operation header for:
3603 * A. Valid client identifier
3604 * B. tid ptr value falls in valid ptr space (user space code)
3605 * C. Length in log record header is correct according to the
3606 * individual operation headers within record.
3607 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3608 * log, check the preceding blocks of the physical log to make sure all
3609 * the cycle numbers agree with the current cycle number.
3614 struct xlog_in_core *iclog,
3617 xlog_op_header_t *ophead;
3618 xlog_in_core_t *icptr;
3619 xlog_in_core_2_t *xhdr;
3620 void *base_ptr, *ptr, *p;
3621 ptrdiff_t field_offset;
3623 int len, i, j, k, op_len;
3626 /* check validity of iclog pointers */
3627 spin_lock(&log->l_icloglock);
3628 icptr = log->l_iclog;
3629 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3632 if (icptr != log->l_iclog)
3633 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3634 spin_unlock(&log->l_icloglock);
3636 /* check log magic numbers */
3637 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3638 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3640 base_ptr = ptr = &iclog->ic_header;
3641 p = &iclog->ic_header;
3642 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3643 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3644 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3649 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3650 base_ptr = ptr = iclog->ic_datap;
3652 xhdr = iclog->ic_data;
3653 for (i = 0; i < len; i++) {
3656 /* clientid is only 1 byte */
3657 p = &ophead->oh_clientid;
3658 field_offset = p - base_ptr;
3659 if (field_offset & 0x1ff) {
3660 clientid = ophead->oh_clientid;
3662 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3663 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3664 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3665 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3666 clientid = xlog_get_client_id(
3667 xhdr[j].hic_xheader.xh_cycle_data[k]);
3669 clientid = xlog_get_client_id(
3670 iclog->ic_header.h_cycle_data[idx]);
3673 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3675 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3676 __func__, i, clientid, ophead,
3677 (unsigned long)field_offset);
3681 p = &ophead->oh_len;
3682 field_offset = p - base_ptr;
3683 if (field_offset & 0x1ff) {
3684 op_len = be32_to_cpu(ophead->oh_len);
3686 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3687 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3688 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3689 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3690 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3692 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3695 ptr += sizeof(xlog_op_header_t) + op_len;
3701 * Perform a forced shutdown on the log.
3703 * This can be called from low level log code to trigger a shutdown, or from the
3704 * high level mount shutdown code when the mount shuts down.
3706 * Our main objectives here are to make sure that:
3707 * a. if the shutdown was not due to a log IO error, flush the logs to
3708 * disk. Anything modified after this is ignored.
3709 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3710 * parties to find out. Nothing new gets queued after this is done.
3711 * c. Tasks sleeping on log reservations, pinned objects and
3712 * other resources get woken up.
3713 * d. The mount is also marked as shut down so that log triggered shutdowns
3714 * still behave the same as if they called xfs_forced_shutdown().
3716 * Return true if the shutdown cause was a log IO error and we actually shut the
3720 xlog_force_shutdown(
3722 uint32_t shutdown_flags)
3724 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3730 * Flush all the completed transactions to disk before marking the log
3731 * being shut down. We need to do this first as shutting down the log
3732 * before the force will prevent the log force from flushing the iclogs
3735 * When we are in recovery, there are no transactions to flush, and
3736 * we don't want to touch the log because we don't want to perturb the
3737 * current head/tail for future recovery attempts. Hence we need to
3738 * avoid a log force in this case.
3740 * If we are shutting down due to a log IO error, then we must avoid
3741 * trying to write the log as that may just result in more IO errors and
3742 * an endless shutdown/force loop.
3744 if (!log_error && !xlog_in_recovery(log))
3745 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3748 * Atomically set the shutdown state. If the shutdown state is already
3749 * set, there someone else is performing the shutdown and so we are done
3750 * here. This should never happen because we should only ever get called
3751 * once by the first shutdown caller.
3753 * Much of the log state machine transitions assume that shutdown state
3754 * cannot change once they hold the log->l_icloglock. Hence we need to
3755 * hold that lock here, even though we use the atomic test_and_set_bit()
3756 * operation to set the shutdown state.
3758 spin_lock(&log->l_icloglock);
3759 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3760 spin_unlock(&log->l_icloglock);
3763 spin_unlock(&log->l_icloglock);
3766 * If this log shutdown also sets the mount shutdown state, issue a
3767 * shutdown warning message.
3769 if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3770 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3771 "Filesystem has been shut down due to log error (0x%x).",
3773 xfs_alert(log->l_mp,
3774 "Please unmount the filesystem and rectify the problem(s).");
3775 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3780 * We don't want anybody waiting for log reservations after this. That
3781 * means we have to wake up everybody queued up on reserveq as well as
3782 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3783 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3784 * action is protected by the grant locks.
3786 xlog_grant_head_wake_all(&log->l_reserve_head);
3787 xlog_grant_head_wake_all(&log->l_write_head);
3790 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3791 * as if the log writes were completed. The abort handling in the log
3792 * item committed callback functions will do this again under lock to
3795 spin_lock(&log->l_cilp->xc_push_lock);
3796 wake_up_all(&log->l_cilp->xc_start_wait);
3797 wake_up_all(&log->l_cilp->xc_commit_wait);
3798 spin_unlock(&log->l_cilp->xc_push_lock);
3800 spin_lock(&log->l_icloglock);
3801 xlog_state_shutdown_callbacks(log);
3802 spin_unlock(&log->l_icloglock);
3804 wake_up_var(&log->l_opstate);
3812 xlog_in_core_t *iclog;
3814 iclog = log->l_iclog;
3816 /* endianness does not matter here, zero is zero in
3819 if (iclog->ic_header.h_num_logops)
3821 iclog = iclog->ic_next;
3822 } while (iclog != log->l_iclog);
3827 * Verify that an LSN stamped into a piece of metadata is valid. This is
3828 * intended for use in read verifiers on v5 superblocks.
3832 struct xfs_mount *mp,
3835 struct xlog *log = mp->m_log;
3839 * norecovery mode skips mount-time log processing and unconditionally
3840 * resets the in-core LSN. We can't validate in this mode, but
3841 * modifications are not allowed anyways so just return true.
3843 if (xfs_has_norecovery(mp))
3847 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3848 * handled by recovery and thus safe to ignore here.
3850 if (lsn == NULLCOMMITLSN)
3853 valid = xlog_valid_lsn(mp->m_log, lsn);
3855 /* warn the user about what's gone wrong before verifier failure */
3857 spin_lock(&log->l_icloglock);
3859 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3860 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3861 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3862 log->l_curr_cycle, log->l_curr_block);
3863 spin_unlock(&log->l_icloglock);
3870 * Notify the log that we're about to start using a feature that is protected
3871 * by a log incompat feature flag. This will prevent log covering from
3872 * clearing those flags.
3875 xlog_use_incompat_feat(
3878 down_read(&log->l_incompat_users);
3881 /* Notify the log that we've finished using log incompat features. */
3883 xlog_drop_incompat_feat(
3886 up_read(&log->l_incompat_users);