1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_errortag.h"
14 #include "xfs_error.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
18 #include "xfs_log_priv.h"
19 #include "xfs_trace.h"
20 #include "xfs_sysfs.h"
22 #include "xfs_health.h"
24 kmem_zone_t *xfs_log_ticket_zone;
26 /* Local miscellaneous function prototypes */
30 struct xlog_ticket *ticket,
31 struct xlog_in_core **iclog,
32 xfs_lsn_t *commitlsnp);
37 struct xfs_buftarg *log_target,
38 xfs_daddr_t blk_offset,
48 /* local state machine functions */
49 STATIC void xlog_state_done_syncing(
50 struct xlog_in_core *iclog);
52 xlog_state_get_iclog_space(
55 struct xlog_in_core **iclog,
56 struct xlog_ticket *ticket,
60 xlog_state_switch_iclogs(
62 struct xlog_in_core *iclog,
69 xlog_regrant_reserve_log_space(
71 struct xlog_ticket *ticket);
73 xlog_ungrant_log_space(
75 struct xlog_ticket *ticket);
79 struct xlog_in_core *iclog);
86 xlog_verify_grant_tail(
91 struct xlog_in_core *iclog,
96 struct xlog_in_core *iclog,
99 #define xlog_verify_dest_ptr(a,b)
100 #define xlog_verify_grant_tail(a)
101 #define xlog_verify_iclog(a,b,c)
102 #define xlog_verify_tail_lsn(a,b,c)
110 xlog_grant_sub_space(
115 int64_t head_val = atomic64_read(head);
121 xlog_crack_grant_head_val(head_val, &cycle, &space);
125 space += log->l_logsize;
130 new = xlog_assign_grant_head_val(cycle, space);
131 head_val = atomic64_cmpxchg(head, old, new);
132 } while (head_val != old);
136 xlog_grant_add_space(
141 int64_t head_val = atomic64_read(head);
148 xlog_crack_grant_head_val(head_val, &cycle, &space);
150 tmp = log->l_logsize - space;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
165 xlog_grant_head_init(
166 struct xlog_grant_head *head)
168 xlog_assign_grant_head(&head->grant, 1, 0);
169 INIT_LIST_HEAD(&head->waiters);
170 spin_lock_init(&head->lock);
174 xlog_grant_head_wake_all(
175 struct xlog_grant_head *head)
177 struct xlog_ticket *tic;
179 spin_lock(&head->lock);
180 list_for_each_entry(tic, &head->waiters, t_queue)
181 wake_up_process(tic->t_task);
182 spin_unlock(&head->lock);
186 xlog_ticket_reservation(
188 struct xlog_grant_head *head,
189 struct xlog_ticket *tic)
191 if (head == &log->l_write_head) {
192 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
193 return tic->t_unit_res;
195 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
196 return tic->t_unit_res * tic->t_cnt;
198 return tic->t_unit_res;
203 xlog_grant_head_wake(
205 struct xlog_grant_head *head,
208 struct xlog_ticket *tic;
210 bool woken_task = false;
212 list_for_each_entry(tic, &head->waiters, t_queue) {
215 * There is a chance that the size of the CIL checkpoints in
216 * progress at the last AIL push target calculation resulted in
217 * limiting the target to the log head (l_last_sync_lsn) at the
218 * time. This may not reflect where the log head is now as the
219 * CIL checkpoints may have completed.
221 * Hence when we are woken here, it may be that the head of the
222 * log that has moved rather than the tail. As the tail didn't
223 * move, there still won't be space available for the
224 * reservation we require. However, if the AIL has already
225 * pushed to the target defined by the old log head location, we
226 * will hang here waiting for something else to update the AIL
229 * Therefore, if there isn't space to wake the first waiter on
230 * the grant head, we need to push the AIL again to ensure the
231 * target reflects both the current log tail and log head
232 * position before we wait for the tail to move again.
235 need_bytes = xlog_ticket_reservation(log, head, tic);
236 if (*free_bytes < need_bytes) {
238 xlog_grant_push_ail(log, need_bytes);
242 *free_bytes -= need_bytes;
243 trace_xfs_log_grant_wake_up(log, tic);
244 wake_up_process(tic->t_task);
252 xlog_grant_head_wait(
254 struct xlog_grant_head *head,
255 struct xlog_ticket *tic,
256 int need_bytes) __releases(&head->lock)
257 __acquires(&head->lock)
259 list_add_tail(&tic->t_queue, &head->waiters);
262 if (XLOG_FORCED_SHUTDOWN(log))
264 xlog_grant_push_ail(log, need_bytes);
266 __set_current_state(TASK_UNINTERRUPTIBLE);
267 spin_unlock(&head->lock);
269 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
271 trace_xfs_log_grant_sleep(log, tic);
273 trace_xfs_log_grant_wake(log, tic);
275 spin_lock(&head->lock);
276 if (XLOG_FORCED_SHUTDOWN(log))
278 } while (xlog_space_left(log, &head->grant) < need_bytes);
280 list_del_init(&tic->t_queue);
283 list_del_init(&tic->t_queue);
288 * Atomically get the log space required for a log ticket.
290 * Once a ticket gets put onto head->waiters, it will only return after the
291 * needed reservation is satisfied.
293 * This function is structured so that it has a lock free fast path. This is
294 * necessary because every new transaction reservation will come through this
295 * path. Hence any lock will be globally hot if we take it unconditionally on
298 * As tickets are only ever moved on and off head->waiters under head->lock, we
299 * only need to take that lock if we are going to add the ticket to the queue
300 * and sleep. We can avoid taking the lock if the ticket was never added to
301 * head->waiters because the t_queue list head will be empty and we hold the
302 * only reference to it so it can safely be checked unlocked.
305 xlog_grant_head_check(
307 struct xlog_grant_head *head,
308 struct xlog_ticket *tic,
314 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
317 * If there are other waiters on the queue then give them a chance at
318 * logspace before us. Wake up the first waiters, if we do not wake
319 * up all the waiters then go to sleep waiting for more free space,
320 * otherwise try to get some space for this transaction.
322 *need_bytes = xlog_ticket_reservation(log, head, tic);
323 free_bytes = xlog_space_left(log, &head->grant);
324 if (!list_empty_careful(&head->waiters)) {
325 spin_lock(&head->lock);
326 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
327 free_bytes < *need_bytes) {
328 error = xlog_grant_head_wait(log, head, tic,
331 spin_unlock(&head->lock);
332 } else if (free_bytes < *need_bytes) {
333 spin_lock(&head->lock);
334 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
335 spin_unlock(&head->lock);
342 xlog_tic_reset_res(xlog_ticket_t *tic)
345 tic->t_res_arr_sum = 0;
346 tic->t_res_num_ophdrs = 0;
350 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
352 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
353 /* add to overflow and start again */
354 tic->t_res_o_flow += tic->t_res_arr_sum;
356 tic->t_res_arr_sum = 0;
359 tic->t_res_arr[tic->t_res_num].r_len = len;
360 tic->t_res_arr[tic->t_res_num].r_type = type;
361 tic->t_res_arr_sum += len;
366 * Replenish the byte reservation required by moving the grant write head.
370 struct xfs_mount *mp,
371 struct xlog_ticket *tic)
373 struct xlog *log = mp->m_log;
377 if (XLOG_FORCED_SHUTDOWN(log))
380 XFS_STATS_INC(mp, xs_try_logspace);
383 * This is a new transaction on the ticket, so we need to change the
384 * transaction ID so that the next transaction has a different TID in
385 * the log. Just add one to the existing tid so that we can see chains
386 * of rolling transactions in the log easily.
390 xlog_grant_push_ail(log, tic->t_unit_res);
392 tic->t_curr_res = tic->t_unit_res;
393 xlog_tic_reset_res(tic);
398 trace_xfs_log_regrant(log, tic);
400 error = xlog_grant_head_check(log, &log->l_write_head, tic,
405 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
406 trace_xfs_log_regrant_exit(log, tic);
407 xlog_verify_grant_tail(log);
412 * If we are failing, make sure the ticket doesn't have any current
413 * reservations. We don't want to add this back when the ticket/
414 * transaction gets cancelled.
417 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
422 * Reserve log space and return a ticket corresponding to the reservation.
424 * Each reservation is going to reserve extra space for a log record header.
425 * When writes happen to the on-disk log, we don't subtract the length of the
426 * log record header from any reservation. By wasting space in each
427 * reservation, we prevent over allocation problems.
431 struct xfs_mount *mp,
434 struct xlog_ticket **ticp,
438 struct xlog *log = mp->m_log;
439 struct xlog_ticket *tic;
443 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
445 if (XLOG_FORCED_SHUTDOWN(log))
448 XFS_STATS_INC(mp, xs_try_logspace);
450 ASSERT(*ticp == NULL);
451 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent, 0);
454 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
457 trace_xfs_log_reserve(log, tic);
459 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
464 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
465 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
466 trace_xfs_log_reserve_exit(log, tic);
467 xlog_verify_grant_tail(log);
472 * If we are failing, make sure the ticket doesn't have any current
473 * reservations. We don't want to add this back when the ticket/
474 * transaction gets cancelled.
477 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
485 * 1. currblock field gets updated at startup and after in-core logs
486 * marked as with WANT_SYNC.
490 * This routine is called when a user of a log manager ticket is done with
491 * the reservation. If the ticket was ever used, then a commit record for
492 * the associated transaction is written out as a log operation header with
493 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
494 * a given ticket. If the ticket was one with a permanent reservation, then
495 * a few operations are done differently. Permanent reservation tickets by
496 * default don't release the reservation. They just commit the current
497 * transaction with the belief that the reservation is still needed. A flag
498 * must be passed in before permanent reservations are actually released.
499 * When these type of tickets are not released, they need to be set into
500 * the inited state again. By doing this, a start record will be written
501 * out when the next write occurs.
505 struct xfs_mount *mp,
506 struct xlog_ticket *ticket,
507 struct xlog_in_core **iclog,
510 struct xlog *log = mp->m_log;
513 if (XLOG_FORCED_SHUTDOWN(log) ||
515 * If nothing was ever written, don't write out commit record.
516 * If we get an error, just continue and give back the log ticket.
518 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
519 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
520 lsn = (xfs_lsn_t) -1;
526 trace_xfs_log_done_nonperm(log, ticket);
529 * Release ticket if not permanent reservation or a specific
530 * request has been made to release a permanent reservation.
532 xlog_ungrant_log_space(log, ticket);
534 trace_xfs_log_done_perm(log, ticket);
536 xlog_regrant_reserve_log_space(log, ticket);
537 /* If this ticket was a permanent reservation and we aren't
538 * trying to release it, reset the inited flags; so next time
539 * we write, a start record will be written out.
541 ticket->t_flags |= XLOG_TIC_INITED;
544 xfs_log_ticket_put(ticket);
549 __xlog_state_release_iclog(
551 struct xlog_in_core *iclog)
553 lockdep_assert_held(&log->l_icloglock);
555 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
556 /* update tail before writing to iclog */
557 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
559 iclog->ic_state = XLOG_STATE_SYNCING;
560 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
561 xlog_verify_tail_lsn(log, iclog, tail_lsn);
562 /* cycle incremented when incrementing curr_block */
566 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
571 * Flush iclog to disk if this is the last reference to the given iclog and the
572 * it is in the WANT_SYNC state.
575 xlog_state_release_iclog(
577 struct xlog_in_core *iclog)
579 lockdep_assert_held(&log->l_icloglock);
581 if (iclog->ic_state == XLOG_STATE_IOERROR)
584 if (atomic_dec_and_test(&iclog->ic_refcnt) &&
585 __xlog_state_release_iclog(log, iclog)) {
586 spin_unlock(&log->l_icloglock);
587 xlog_sync(log, iclog);
588 spin_lock(&log->l_icloglock);
595 xfs_log_release_iclog(
596 struct xlog_in_core *iclog)
598 struct xlog *log = iclog->ic_log;
601 if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) {
602 if (iclog->ic_state != XLOG_STATE_IOERROR)
603 sync = __xlog_state_release_iclog(log, iclog);
604 spin_unlock(&log->l_icloglock);
608 xlog_sync(log, iclog);
612 * Mount a log filesystem
614 * mp - ubiquitous xfs mount point structure
615 * log_target - buftarg of on-disk log device
616 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
617 * num_bblocks - Number of BBSIZE blocks in on-disk log
619 * Return error or zero.
624 xfs_buftarg_t *log_target,
625 xfs_daddr_t blk_offset,
628 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
632 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
633 xfs_notice(mp, "Mounting V%d Filesystem",
634 XFS_SB_VERSION_NUM(&mp->m_sb));
637 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
638 XFS_SB_VERSION_NUM(&mp->m_sb));
639 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
642 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
643 if (IS_ERR(mp->m_log)) {
644 error = PTR_ERR(mp->m_log);
649 * Validate the given log space and drop a critical message via syslog
650 * if the log size is too small that would lead to some unexpected
651 * situations in transaction log space reservation stage.
653 * Note: we can't just reject the mount if the validation fails. This
654 * would mean that people would have to downgrade their kernel just to
655 * remedy the situation as there is no way to grow the log (short of
656 * black magic surgery with xfs_db).
658 * We can, however, reject mounts for CRC format filesystems, as the
659 * mkfs binary being used to make the filesystem should never create a
660 * filesystem with a log that is too small.
662 min_logfsbs = xfs_log_calc_minimum_size(mp);
664 if (mp->m_sb.sb_logblocks < min_logfsbs) {
666 "Log size %d blocks too small, minimum size is %d blocks",
667 mp->m_sb.sb_logblocks, min_logfsbs);
669 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
671 "Log size %d blocks too large, maximum size is %lld blocks",
672 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
674 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
676 "log size %lld bytes too large, maximum size is %lld bytes",
677 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
680 } else if (mp->m_sb.sb_logsunit > 1 &&
681 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
683 "log stripe unit %u bytes must be a multiple of block size",
684 mp->m_sb.sb_logsunit);
690 * Log check errors are always fatal on v5; or whenever bad
691 * metadata leads to a crash.
694 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
698 xfs_crit(mp, "Log size out of supported range.");
700 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
704 * Initialize the AIL now we have a log.
706 error = xfs_trans_ail_init(mp);
708 xfs_warn(mp, "AIL initialisation failed: error %d", error);
711 mp->m_log->l_ailp = mp->m_ail;
714 * skip log recovery on a norecovery mount. pretend it all
717 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
718 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
721 mp->m_flags &= ~XFS_MOUNT_RDONLY;
723 error = xlog_recover(mp->m_log);
726 mp->m_flags |= XFS_MOUNT_RDONLY;
728 xfs_warn(mp, "log mount/recovery failed: error %d",
730 xlog_recover_cancel(mp->m_log);
731 goto out_destroy_ail;
735 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
738 goto out_destroy_ail;
740 /* Normal transactions can now occur */
741 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
744 * Now the log has been fully initialised and we know were our
745 * space grant counters are, we can initialise the permanent ticket
746 * needed for delayed logging to work.
748 xlog_cil_init_post_recovery(mp->m_log);
753 xfs_trans_ail_destroy(mp);
755 xlog_dealloc_log(mp->m_log);
761 * Finish the recovery of the file system. This is separate from the
762 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
763 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
766 * If we finish recovery successfully, start the background log work. If we are
767 * not doing recovery, then we have a RO filesystem and we don't need to start
771 xfs_log_mount_finish(
772 struct xfs_mount *mp)
775 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
776 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
778 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
779 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
781 } else if (readonly) {
782 /* Allow unlinked processing to proceed */
783 mp->m_flags &= ~XFS_MOUNT_RDONLY;
787 * During the second phase of log recovery, we need iget and
788 * iput to behave like they do for an active filesystem.
789 * xfs_fs_drop_inode needs to be able to prevent the deletion
790 * of inodes before we're done replaying log items on those
791 * inodes. Turn it off immediately after recovery finishes
792 * so that we don't leak the quota inodes if subsequent mount
795 * We let all inodes involved in redo item processing end up on
796 * the LRU instead of being evicted immediately so that if we do
797 * something to an unlinked inode, the irele won't cause
798 * premature truncation and freeing of the inode, which results
799 * in log recovery failure. We have to evict the unreferenced
800 * lru inodes after clearing SB_ACTIVE because we don't
801 * otherwise clean up the lru if there's a subsequent failure in
802 * xfs_mountfs, which leads to us leaking the inodes if nothing
803 * else (e.g. quotacheck) references the inodes before the
804 * mount failure occurs.
806 mp->m_super->s_flags |= SB_ACTIVE;
807 error = xlog_recover_finish(mp->m_log);
809 xfs_log_work_queue(mp);
810 mp->m_super->s_flags &= ~SB_ACTIVE;
811 evict_inodes(mp->m_super);
814 * Drain the buffer LRU after log recovery. This is required for v4
815 * filesystems to avoid leaving around buffers with NULL verifier ops,
816 * but we do it unconditionally to make sure we're always in a clean
817 * cache state after mount.
819 * Don't push in the error case because the AIL may have pending intents
820 * that aren't removed until recovery is cancelled.
822 if (!error && recovered) {
823 xfs_log_force(mp, XFS_LOG_SYNC);
824 xfs_ail_push_all_sync(mp->m_ail);
826 xfs_wait_buftarg(mp->m_ddev_targp);
829 mp->m_flags |= XFS_MOUNT_RDONLY;
835 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
839 xfs_log_mount_cancel(
840 struct xfs_mount *mp)
842 xlog_recover_cancel(mp->m_log);
847 * Wait for the iclog to be written disk, or return an error if the log has been
852 struct xlog_in_core *iclog)
853 __releases(iclog->ic_log->l_icloglock)
855 struct xlog *log = iclog->ic_log;
857 if (!XLOG_FORCED_SHUTDOWN(log) &&
858 iclog->ic_state != XLOG_STATE_ACTIVE &&
859 iclog->ic_state != XLOG_STATE_DIRTY) {
860 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
861 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
863 spin_unlock(&log->l_icloglock);
866 if (XLOG_FORCED_SHUTDOWN(log))
872 * Final log writes as part of unmount.
874 * Mark the filesystem clean as unmount happens. Note that during relocation
875 * this routine needs to be executed as part of source-bag while the
876 * deallocation must not be done until source-end.
879 /* Actually write the unmount record to disk. */
881 xfs_log_write_unmount_record(
882 struct xfs_mount *mp)
884 /* the data section must be 32 bit size aligned */
885 struct xfs_unmount_log_format magic = {
886 .magic = XLOG_UNMOUNT_TYPE,
888 struct xfs_log_iovec reg = {
890 .i_len = sizeof(magic),
891 .i_type = XLOG_REG_TYPE_UNMOUNT,
893 struct xfs_log_vec vec = {
897 struct xlog *log = mp->m_log;
898 struct xlog_in_core *iclog;
899 struct xlog_ticket *tic = NULL;
901 uint flags = XLOG_UNMOUNT_TRANS;
904 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
909 * If we think the summary counters are bad, clear the unmount header
910 * flag in the unmount record so that the summary counters will be
911 * recalculated during log recovery at next mount. Refer to
912 * xlog_check_unmount_rec for more details.
914 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
915 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
916 xfs_alert(mp, "%s: will fix summary counters at next mount",
918 flags &= ~XLOG_UNMOUNT_TRANS;
921 /* remove inited flag, and account for space used */
923 tic->t_curr_res -= sizeof(magic);
924 error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
926 * At this point, we're umounting anyway, so there's no point in
927 * transitioning log state to IOERROR. Just continue...
931 xfs_alert(mp, "%s: unmount record failed", __func__);
933 spin_lock(&log->l_icloglock);
934 iclog = log->l_iclog;
935 atomic_inc(&iclog->ic_refcnt);
936 if (iclog->ic_state == XLOG_STATE_ACTIVE)
937 xlog_state_switch_iclogs(log, iclog, 0);
939 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
940 iclog->ic_state == XLOG_STATE_IOERROR);
941 error = xlog_state_release_iclog(log, iclog);
942 xlog_wait_on_iclog(iclog);
945 trace_xfs_log_umount_write(log, tic);
946 xlog_ungrant_log_space(log, tic);
947 xfs_log_ticket_put(tic);
952 xfs_log_unmount_verify_iclog(
955 struct xlog_in_core *iclog = log->l_iclog;
958 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
959 ASSERT(iclog->ic_offset == 0);
960 } while ((iclog = iclog->ic_next) != log->l_iclog);
964 * Unmount record used to have a string "Unmount filesystem--" in the
965 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
966 * We just write the magic number now since that particular field isn't
967 * currently architecture converted and "Unmount" is a bit foo.
968 * As far as I know, there weren't any dependencies on the old behaviour.
971 xfs_log_unmount_write(
972 struct xfs_mount *mp)
974 struct xlog *log = mp->m_log;
977 * Don't write out unmount record on norecovery mounts or ro devices.
978 * Or, if we are doing a forced umount (typically because of IO errors).
980 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
981 xfs_readonly_buftarg(log->l_targ)) {
982 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
986 xfs_log_force(mp, XFS_LOG_SYNC);
988 if (XLOG_FORCED_SHUTDOWN(log))
990 xfs_log_unmount_verify_iclog(log);
991 xfs_log_write_unmount_record(mp);
995 * Empty the log for unmount/freeze.
997 * To do this, we first need to shut down the background log work so it is not
998 * trying to cover the log as we clean up. We then need to unpin all objects in
999 * the log so we can then flush them out. Once they have completed their IO and
1000 * run the callbacks removing themselves from the AIL, we can write the unmount
1005 struct xfs_mount *mp)
1007 cancel_delayed_work_sync(&mp->m_log->l_work);
1008 xfs_log_force(mp, XFS_LOG_SYNC);
1011 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1012 * will push it, xfs_wait_buftarg() will not wait for it. Further,
1013 * xfs_buf_iowait() cannot be used because it was pushed with the
1014 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1015 * the IO to complete.
1017 xfs_ail_push_all_sync(mp->m_ail);
1018 xfs_wait_buftarg(mp->m_ddev_targp);
1019 xfs_buf_lock(mp->m_sb_bp);
1020 xfs_buf_unlock(mp->m_sb_bp);
1022 xfs_log_unmount_write(mp);
1026 * Shut down and release the AIL and Log.
1028 * During unmount, we need to ensure we flush all the dirty metadata objects
1029 * from the AIL so that the log is empty before we write the unmount record to
1030 * the log. Once this is done, we can tear down the AIL and the log.
1034 struct xfs_mount *mp)
1036 xfs_log_quiesce(mp);
1038 xfs_trans_ail_destroy(mp);
1040 xfs_sysfs_del(&mp->m_log->l_kobj);
1042 xlog_dealloc_log(mp->m_log);
1047 struct xfs_mount *mp,
1048 struct xfs_log_item *item,
1050 const struct xfs_item_ops *ops)
1052 item->li_mountp = mp;
1053 item->li_ailp = mp->m_ail;
1054 item->li_type = type;
1058 INIT_LIST_HEAD(&item->li_ail);
1059 INIT_LIST_HEAD(&item->li_cil);
1060 INIT_LIST_HEAD(&item->li_bio_list);
1061 INIT_LIST_HEAD(&item->li_trans);
1065 * Wake up processes waiting for log space after we have moved the log tail.
1069 struct xfs_mount *mp)
1071 struct xlog *log = mp->m_log;
1074 if (XLOG_FORCED_SHUTDOWN(log))
1077 if (!list_empty_careful(&log->l_write_head.waiters)) {
1078 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1080 spin_lock(&log->l_write_head.lock);
1081 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1082 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1083 spin_unlock(&log->l_write_head.lock);
1086 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1087 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1089 spin_lock(&log->l_reserve_head.lock);
1090 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1091 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1092 spin_unlock(&log->l_reserve_head.lock);
1097 * Determine if we have a transaction that has gone to disk that needs to be
1098 * covered. To begin the transition to the idle state firstly the log needs to
1099 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1100 * we start attempting to cover the log.
1102 * Only if we are then in a state where covering is needed, the caller is
1103 * informed that dummy transactions are required to move the log into the idle
1106 * If there are any items in the AIl or CIL, then we do not want to attempt to
1107 * cover the log as we may be in a situation where there isn't log space
1108 * available to run a dummy transaction and this can lead to deadlocks when the
1109 * tail of the log is pinned by an item that is modified in the CIL. Hence
1110 * there's no point in running a dummy transaction at this point because we
1111 * can't start trying to idle the log until both the CIL and AIL are empty.
1114 xfs_log_need_covered(xfs_mount_t *mp)
1116 struct xlog *log = mp->m_log;
1119 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1122 if (!xlog_cil_empty(log))
1125 spin_lock(&log->l_icloglock);
1126 switch (log->l_covered_state) {
1127 case XLOG_STATE_COVER_DONE:
1128 case XLOG_STATE_COVER_DONE2:
1129 case XLOG_STATE_COVER_IDLE:
1131 case XLOG_STATE_COVER_NEED:
1132 case XLOG_STATE_COVER_NEED2:
1133 if (xfs_ail_min_lsn(log->l_ailp))
1135 if (!xlog_iclogs_empty(log))
1139 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1140 log->l_covered_state = XLOG_STATE_COVER_DONE;
1142 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1148 spin_unlock(&log->l_icloglock);
1153 * We may be holding the log iclog lock upon entering this routine.
1156 xlog_assign_tail_lsn_locked(
1157 struct xfs_mount *mp)
1159 struct xlog *log = mp->m_log;
1160 struct xfs_log_item *lip;
1163 assert_spin_locked(&mp->m_ail->ail_lock);
1166 * To make sure we always have a valid LSN for the log tail we keep
1167 * track of the last LSN which was committed in log->l_last_sync_lsn,
1168 * and use that when the AIL was empty.
1170 lip = xfs_ail_min(mp->m_ail);
1172 tail_lsn = lip->li_lsn;
1174 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1175 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1176 atomic64_set(&log->l_tail_lsn, tail_lsn);
1181 xlog_assign_tail_lsn(
1182 struct xfs_mount *mp)
1186 spin_lock(&mp->m_ail->ail_lock);
1187 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1188 spin_unlock(&mp->m_ail->ail_lock);
1194 * Return the space in the log between the tail and the head. The head
1195 * is passed in the cycle/bytes formal parms. In the special case where
1196 * the reserve head has wrapped passed the tail, this calculation is no
1197 * longer valid. In this case, just return 0 which means there is no space
1198 * in the log. This works for all places where this function is called
1199 * with the reserve head. Of course, if the write head were to ever
1200 * wrap the tail, we should blow up. Rather than catch this case here,
1201 * we depend on other ASSERTions in other parts of the code. XXXmiken
1203 * This code also handles the case where the reservation head is behind
1204 * the tail. The details of this case are described below, but the end
1205 * result is that we return the size of the log as the amount of space left.
1218 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1219 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1220 tail_bytes = BBTOB(tail_bytes);
1221 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1222 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1223 else if (tail_cycle + 1 < head_cycle)
1225 else if (tail_cycle < head_cycle) {
1226 ASSERT(tail_cycle == (head_cycle - 1));
1227 free_bytes = tail_bytes - head_bytes;
1230 * The reservation head is behind the tail.
1231 * In this case we just want to return the size of the
1232 * log as the amount of space left.
1234 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1235 xfs_alert(log->l_mp,
1236 " tail_cycle = %d, tail_bytes = %d",
1237 tail_cycle, tail_bytes);
1238 xfs_alert(log->l_mp,
1239 " GH cycle = %d, GH bytes = %d",
1240 head_cycle, head_bytes);
1242 free_bytes = log->l_logsize;
1250 struct work_struct *work)
1252 struct xlog_in_core *iclog =
1253 container_of(work, struct xlog_in_core, ic_end_io_work);
1254 struct xlog *log = iclog->ic_log;
1257 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1259 /* treat writes with injected CRC errors as failed */
1260 if (iclog->ic_fail_crc)
1265 * Race to shutdown the filesystem if we see an error.
1267 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1268 xfs_alert(log->l_mp, "log I/O error %d", error);
1269 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1272 xlog_state_done_syncing(iclog);
1273 bio_uninit(&iclog->ic_bio);
1276 * Drop the lock to signal that we are done. Nothing references the
1277 * iclog after this, so an unmount waiting on this lock can now tear it
1278 * down safely. As such, it is unsafe to reference the iclog after the
1279 * unlock as we could race with it being freed.
1281 up(&iclog->ic_sema);
1285 * Return size of each in-core log record buffer.
1287 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1289 * If the filesystem blocksize is too large, we may need to choose a
1290 * larger size since the directory code currently logs entire blocks.
1293 xlog_get_iclog_buffer_size(
1294 struct xfs_mount *mp,
1297 if (mp->m_logbufs <= 0)
1298 mp->m_logbufs = XLOG_MAX_ICLOGS;
1299 if (mp->m_logbsize <= 0)
1300 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1302 log->l_iclog_bufs = mp->m_logbufs;
1303 log->l_iclog_size = mp->m_logbsize;
1306 * # headers = size / 32k - one header holds cycles from 32k of data.
1308 log->l_iclog_heads =
1309 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1310 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1315 struct xfs_mount *mp)
1317 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1318 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1322 * Every sync period we need to unpin all items in the AIL and push them to
1323 * disk. If there is nothing dirty, then we might need to cover the log to
1324 * indicate that the filesystem is idle.
1328 struct work_struct *work)
1330 struct xlog *log = container_of(to_delayed_work(work),
1331 struct xlog, l_work);
1332 struct xfs_mount *mp = log->l_mp;
1334 /* dgc: errors ignored - not fatal and nowhere to report them */
1335 if (xfs_log_need_covered(mp)) {
1337 * Dump a transaction into the log that contains no real change.
1338 * This is needed to stamp the current tail LSN into the log
1339 * during the covering operation.
1341 * We cannot use an inode here for this - that will push dirty
1342 * state back up into the VFS and then periodic inode flushing
1343 * will prevent log covering from making progress. Hence we
1344 * synchronously log the superblock instead to ensure the
1345 * superblock is immediately unpinned and can be written back.
1347 xfs_sync_sb(mp, true);
1349 xfs_log_force(mp, 0);
1351 /* start pushing all the metadata that is currently dirty */
1352 xfs_ail_push_all(mp->m_ail);
1354 /* queue us up again */
1355 xfs_log_work_queue(mp);
1359 * This routine initializes some of the log structure for a given mount point.
1360 * Its primary purpose is to fill in enough, so recovery can occur. However,
1361 * some other stuff may be filled in too.
1363 STATIC struct xlog *
1365 struct xfs_mount *mp,
1366 struct xfs_buftarg *log_target,
1367 xfs_daddr_t blk_offset,
1371 xlog_rec_header_t *head;
1372 xlog_in_core_t **iclogp;
1373 xlog_in_core_t *iclog, *prev_iclog=NULL;
1375 int error = -ENOMEM;
1378 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1380 xfs_warn(mp, "Log allocation failed: No memory!");
1385 log->l_targ = log_target;
1386 log->l_logsize = BBTOB(num_bblks);
1387 log->l_logBBstart = blk_offset;
1388 log->l_logBBsize = num_bblks;
1389 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1390 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1391 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1393 log->l_prev_block = -1;
1394 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1395 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1396 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1397 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1399 xlog_grant_head_init(&log->l_reserve_head);
1400 xlog_grant_head_init(&log->l_write_head);
1402 error = -EFSCORRUPTED;
1403 if (xfs_sb_version_hassector(&mp->m_sb)) {
1404 log2_size = mp->m_sb.sb_logsectlog;
1405 if (log2_size < BBSHIFT) {
1406 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1407 log2_size, BBSHIFT);
1411 log2_size -= BBSHIFT;
1412 if (log2_size > mp->m_sectbb_log) {
1413 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1414 log2_size, mp->m_sectbb_log);
1418 /* for larger sector sizes, must have v2 or external log */
1419 if (log2_size && log->l_logBBstart > 0 &&
1420 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1422 "log sector size (0x%x) invalid for configuration.",
1427 log->l_sectBBsize = 1 << log2_size;
1429 xlog_get_iclog_buffer_size(mp, log);
1431 spin_lock_init(&log->l_icloglock);
1432 init_waitqueue_head(&log->l_flush_wait);
1434 iclogp = &log->l_iclog;
1436 * The amount of memory to allocate for the iclog structure is
1437 * rather funky due to the way the structure is defined. It is
1438 * done this way so that we can use different sizes for machines
1439 * with different amounts of memory. See the definition of
1440 * xlog_in_core_t in xfs_log_priv.h for details.
1442 ASSERT(log->l_iclog_size >= 4096);
1443 for (i = 0; i < log->l_iclog_bufs; i++) {
1444 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1445 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1446 sizeof(struct bio_vec);
1448 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1450 goto out_free_iclog;
1453 iclog->ic_prev = prev_iclog;
1456 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1457 KM_MAYFAIL | KM_ZERO);
1458 if (!iclog->ic_data)
1459 goto out_free_iclog;
1461 log->l_iclog_bak[i] = &iclog->ic_header;
1463 head = &iclog->ic_header;
1464 memset(head, 0, sizeof(xlog_rec_header_t));
1465 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1466 head->h_version = cpu_to_be32(
1467 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1468 head->h_size = cpu_to_be32(log->l_iclog_size);
1470 head->h_fmt = cpu_to_be32(XLOG_FMT);
1471 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1473 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1474 iclog->ic_state = XLOG_STATE_ACTIVE;
1475 iclog->ic_log = log;
1476 atomic_set(&iclog->ic_refcnt, 0);
1477 spin_lock_init(&iclog->ic_callback_lock);
1478 INIT_LIST_HEAD(&iclog->ic_callbacks);
1479 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1481 init_waitqueue_head(&iclog->ic_force_wait);
1482 init_waitqueue_head(&iclog->ic_write_wait);
1483 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1484 sema_init(&iclog->ic_sema, 1);
1486 iclogp = &iclog->ic_next;
1488 *iclogp = log->l_iclog; /* complete ring */
1489 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1491 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1492 WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1494 if (!log->l_ioend_workqueue)
1495 goto out_free_iclog;
1497 error = xlog_cil_init(log);
1499 goto out_destroy_workqueue;
1502 out_destroy_workqueue:
1503 destroy_workqueue(log->l_ioend_workqueue);
1505 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1506 prev_iclog = iclog->ic_next;
1507 kmem_free(iclog->ic_data);
1509 if (prev_iclog == log->l_iclog)
1515 return ERR_PTR(error);
1516 } /* xlog_alloc_log */
1520 * Write out the commit record of a transaction associated with the given
1521 * ticket. Return the lsn of the commit record.
1526 struct xlog_ticket *ticket,
1527 struct xlog_in_core **iclog,
1528 xfs_lsn_t *commitlsnp)
1530 struct xfs_mount *mp = log->l_mp;
1532 struct xfs_log_iovec reg = {
1535 .i_type = XLOG_REG_TYPE_COMMIT,
1537 struct xfs_log_vec vec = {
1542 ASSERT_ALWAYS(iclog);
1543 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1546 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1551 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1552 * log space. This code pushes on the lsn which would supposedly free up
1553 * the 25% which we want to leave free. We may need to adopt a policy which
1554 * pushes on an lsn which is further along in the log once we reach the high
1555 * water mark. In this manner, we would be creating a low water mark.
1558 xlog_grant_push_ail(
1562 xfs_lsn_t threshold_lsn = 0;
1563 xfs_lsn_t last_sync_lsn;
1566 int threshold_block;
1567 int threshold_cycle;
1570 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1572 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1573 free_blocks = BTOBBT(free_bytes);
1576 * Set the threshold for the minimum number of free blocks in the
1577 * log to the maximum of what the caller needs, one quarter of the
1578 * log, and 256 blocks.
1580 free_threshold = BTOBB(need_bytes);
1581 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1582 free_threshold = max(free_threshold, 256);
1583 if (free_blocks >= free_threshold)
1586 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1588 threshold_block += free_threshold;
1589 if (threshold_block >= log->l_logBBsize) {
1590 threshold_block -= log->l_logBBsize;
1591 threshold_cycle += 1;
1593 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1596 * Don't pass in an lsn greater than the lsn of the last
1597 * log record known to be on disk. Use a snapshot of the last sync lsn
1598 * so that it doesn't change between the compare and the set.
1600 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1601 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1602 threshold_lsn = last_sync_lsn;
1605 * Get the transaction layer to kick the dirty buffers out to
1606 * disk asynchronously. No point in trying to do this if
1607 * the filesystem is shutting down.
1609 if (!XLOG_FORCED_SHUTDOWN(log))
1610 xfs_ail_push(log->l_ailp, threshold_lsn);
1614 * Stamp cycle number in every block
1619 struct xlog_in_core *iclog,
1623 int size = iclog->ic_offset + roundoff;
1627 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1629 dp = iclog->ic_datap;
1630 for (i = 0; i < BTOBB(size); i++) {
1631 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1633 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1634 *(__be32 *)dp = cycle_lsn;
1638 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1639 xlog_in_core_2_t *xhdr = iclog->ic_data;
1641 for ( ; i < BTOBB(size); i++) {
1642 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1643 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1644 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1645 *(__be32 *)dp = cycle_lsn;
1649 for (i = 1; i < log->l_iclog_heads; i++)
1650 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1655 * Calculate the checksum for a log buffer.
1657 * This is a little more complicated than it should be because the various
1658 * headers and the actual data are non-contiguous.
1663 struct xlog_rec_header *rhead,
1669 /* first generate the crc for the record header ... */
1670 crc = xfs_start_cksum_update((char *)rhead,
1671 sizeof(struct xlog_rec_header),
1672 offsetof(struct xlog_rec_header, h_crc));
1674 /* ... then for additional cycle data for v2 logs ... */
1675 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1676 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1680 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1681 if (size % XLOG_HEADER_CYCLE_SIZE)
1684 for (i = 1; i < xheads; i++) {
1685 crc = crc32c(crc, &xhdr[i].hic_xheader,
1686 sizeof(struct xlog_rec_ext_header));
1690 /* ... and finally for the payload */
1691 crc = crc32c(crc, dp, size);
1693 return xfs_end_cksum(crc);
1700 struct xlog_in_core *iclog = bio->bi_private;
1702 queue_work(iclog->ic_log->l_ioend_workqueue,
1703 &iclog->ic_end_io_work);
1707 xlog_map_iclog_data(
1713 struct page *page = kmem_to_page(data);
1714 unsigned int off = offset_in_page(data);
1715 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1717 if (bio_add_page(bio, page, len, off) != len)
1730 struct xlog_in_core *iclog,
1735 ASSERT(bno < log->l_logBBsize);
1738 * We lock the iclogbufs here so that we can serialise against I/O
1739 * completion during unmount. We might be processing a shutdown
1740 * triggered during unmount, and that can occur asynchronously to the
1741 * unmount thread, and hence we need to ensure that completes before
1742 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1743 * across the log IO to archieve that.
1745 down(&iclog->ic_sema);
1746 if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1748 * It would seem logical to return EIO here, but we rely on
1749 * the log state machine to propagate I/O errors instead of
1750 * doing it here. We kick of the state machine and unlock
1751 * the buffer manually, the code needs to be kept in sync
1752 * with the I/O completion path.
1754 xlog_state_done_syncing(iclog);
1755 up(&iclog->ic_sema);
1759 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1760 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1761 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1762 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1763 iclog->ic_bio.bi_private = iclog;
1764 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1766 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1768 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1769 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1772 if (is_vmalloc_addr(iclog->ic_data))
1773 flush_kernel_vmap_range(iclog->ic_data, count);
1776 * If this log buffer would straddle the end of the log we will have
1777 * to split it up into two bios, so that we can continue at the start.
1779 if (bno + BTOBB(count) > log->l_logBBsize) {
1782 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1783 GFP_NOIO, &fs_bio_set);
1784 bio_chain(split, &iclog->ic_bio);
1787 /* restart at logical offset zero for the remainder */
1788 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1791 submit_bio(&iclog->ic_bio);
1795 * We need to bump cycle number for the part of the iclog that is
1796 * written to the start of the log. Watch out for the header magic
1797 * number case, though.
1806 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1809 for (i = split_offset; i < count; i += BBSIZE) {
1810 uint32_t cycle = get_unaligned_be32(data + i);
1812 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1814 put_unaligned_be32(cycle, data + i);
1819 xlog_calc_iclog_size(
1821 struct xlog_in_core *iclog,
1824 uint32_t count_init, count;
1827 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1828 log->l_mp->m_sb.sb_logsunit > 1;
1830 /* Add for LR header */
1831 count_init = log->l_iclog_hsize + iclog->ic_offset;
1833 /* Round out the log write size */
1835 /* we have a v2 stripe unit to use */
1836 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1838 count = BBTOB(BTOBB(count_init));
1841 ASSERT(count >= count_init);
1842 *roundoff = count - count_init;
1845 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1847 ASSERT(*roundoff < BBTOB(1));
1852 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1853 * fashion. Previously, we should have moved the current iclog
1854 * ptr in the log to point to the next available iclog. This allows further
1855 * write to continue while this code syncs out an iclog ready to go.
1856 * Before an in-core log can be written out, the data section must be scanned
1857 * to save away the 1st word of each BBSIZE block into the header. We replace
1858 * it with the current cycle count. Each BBSIZE block is tagged with the
1859 * cycle count because there in an implicit assumption that drives will
1860 * guarantee that entire 512 byte blocks get written at once. In other words,
1861 * we can't have part of a 512 byte block written and part not written. By
1862 * tagging each block, we will know which blocks are valid when recovering
1863 * after an unclean shutdown.
1865 * This routine is single threaded on the iclog. No other thread can be in
1866 * this routine with the same iclog. Changing contents of iclog can there-
1867 * fore be done without grabbing the state machine lock. Updating the global
1868 * log will require grabbing the lock though.
1870 * The entire log manager uses a logical block numbering scheme. Only
1871 * xlog_write_iclog knows about the fact that the log may not start with
1872 * block zero on a given device.
1877 struct xlog_in_core *iclog)
1879 unsigned int count; /* byte count of bwrite */
1880 unsigned int roundoff; /* roundoff to BB or stripe */
1883 bool need_flush = true, split = false;
1885 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1887 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1889 /* move grant heads by roundoff in sync */
1890 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1891 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1893 /* put cycle number in every block */
1894 xlog_pack_data(log, iclog, roundoff);
1896 /* real byte length */
1897 size = iclog->ic_offset;
1898 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1900 iclog->ic_header.h_len = cpu_to_be32(size);
1902 XFS_STATS_INC(log->l_mp, xs_log_writes);
1903 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1905 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1907 /* Do we need to split this write into 2 parts? */
1908 if (bno + BTOBB(count) > log->l_logBBsize) {
1909 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1913 /* calculcate the checksum */
1914 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1915 iclog->ic_datap, size);
1917 * Intentionally corrupt the log record CRC based on the error injection
1918 * frequency, if defined. This facilitates testing log recovery in the
1919 * event of torn writes. Hence, set the IOABORT state to abort the log
1920 * write on I/O completion and shutdown the fs. The subsequent mount
1921 * detects the bad CRC and attempts to recover.
1924 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1925 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1926 iclog->ic_fail_crc = true;
1928 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1929 be64_to_cpu(iclog->ic_header.h_lsn));
1934 * Flush the data device before flushing the log to make sure all meta
1935 * data written back from the AIL actually made it to disk before
1936 * stamping the new log tail LSN into the log buffer. For an external
1937 * log we need to issue the flush explicitly, and unfortunately
1938 * synchronously here; for an internal log we can simply use the block
1939 * layer state machine for preflushes.
1941 if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1942 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1946 xlog_verify_iclog(log, iclog, count);
1947 xlog_write_iclog(log, iclog, bno, count, need_flush);
1951 * Deallocate a log structure
1957 xlog_in_core_t *iclog, *next_iclog;
1960 xlog_cil_destroy(log);
1963 * Cycle all the iclogbuf locks to make sure all log IO completion
1964 * is done before we tear down these buffers.
1966 iclog = log->l_iclog;
1967 for (i = 0; i < log->l_iclog_bufs; i++) {
1968 down(&iclog->ic_sema);
1969 up(&iclog->ic_sema);
1970 iclog = iclog->ic_next;
1973 iclog = log->l_iclog;
1974 for (i = 0; i < log->l_iclog_bufs; i++) {
1975 next_iclog = iclog->ic_next;
1976 kmem_free(iclog->ic_data);
1981 log->l_mp->m_log = NULL;
1982 destroy_workqueue(log->l_ioend_workqueue);
1984 } /* xlog_dealloc_log */
1987 * Update counters atomically now that memcpy is done.
1990 xlog_state_finish_copy(
1992 struct xlog_in_core *iclog,
1996 lockdep_assert_held(&log->l_icloglock);
1998 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1999 iclog->ic_offset += copy_bytes;
2003 * print out info relating to regions written which consume
2008 struct xfs_mount *mp,
2009 struct xlog_ticket *ticket)
2012 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2014 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2015 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2016 static char *res_type_str[] = {
2017 REG_TYPE_STR(BFORMAT, "bformat"),
2018 REG_TYPE_STR(BCHUNK, "bchunk"),
2019 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2020 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2021 REG_TYPE_STR(IFORMAT, "iformat"),
2022 REG_TYPE_STR(ICORE, "icore"),
2023 REG_TYPE_STR(IEXT, "iext"),
2024 REG_TYPE_STR(IBROOT, "ibroot"),
2025 REG_TYPE_STR(ILOCAL, "ilocal"),
2026 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2027 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2028 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2029 REG_TYPE_STR(QFORMAT, "qformat"),
2030 REG_TYPE_STR(DQUOT, "dquot"),
2031 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2032 REG_TYPE_STR(LRHEADER, "LR header"),
2033 REG_TYPE_STR(UNMOUNT, "unmount"),
2034 REG_TYPE_STR(COMMIT, "commit"),
2035 REG_TYPE_STR(TRANSHDR, "trans header"),
2036 REG_TYPE_STR(ICREATE, "inode create"),
2037 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2038 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2039 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2040 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2041 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2042 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2044 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2047 xfs_warn(mp, "ticket reservation summary:");
2048 xfs_warn(mp, " unit res = %d bytes",
2049 ticket->t_unit_res);
2050 xfs_warn(mp, " current res = %d bytes",
2051 ticket->t_curr_res);
2052 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2053 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2054 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2055 ticket->t_res_num_ophdrs, ophdr_spc);
2056 xfs_warn(mp, " ophdr + reg = %u bytes",
2057 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2058 xfs_warn(mp, " num regions = %u",
2061 for (i = 0; i < ticket->t_res_num; i++) {
2062 uint r_type = ticket->t_res_arr[i].r_type;
2063 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2064 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2065 "bad-rtype" : res_type_str[r_type]),
2066 ticket->t_res_arr[i].r_len);
2071 * Print a summary of the transaction.
2075 struct xfs_trans *tp)
2077 struct xfs_mount *mp = tp->t_mountp;
2078 struct xfs_log_item *lip;
2080 /* dump core transaction and ticket info */
2081 xfs_warn(mp, "transaction summary:");
2082 xfs_warn(mp, " log res = %d", tp->t_log_res);
2083 xfs_warn(mp, " log count = %d", tp->t_log_count);
2084 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2086 xlog_print_tic_res(mp, tp->t_ticket);
2088 /* dump each log item */
2089 list_for_each_entry(lip, &tp->t_items, li_trans) {
2090 struct xfs_log_vec *lv = lip->li_lv;
2091 struct xfs_log_iovec *vec;
2094 xfs_warn(mp, "log item: ");
2095 xfs_warn(mp, " type = 0x%x", lip->li_type);
2096 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2099 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2100 xfs_warn(mp, " size = %d", lv->lv_size);
2101 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2102 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2104 /* dump each iovec for the log item */
2105 vec = lv->lv_iovecp;
2106 for (i = 0; i < lv->lv_niovecs; i++) {
2107 int dumplen = min(vec->i_len, 32);
2109 xfs_warn(mp, " iovec[%d]", i);
2110 xfs_warn(mp, " type = 0x%x", vec->i_type);
2111 xfs_warn(mp, " len = %d", vec->i_len);
2112 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2113 xfs_hex_dump(vec->i_addr, dumplen);
2121 * Calculate the potential space needed by the log vector. Each region gets
2122 * its own xlog_op_header_t and may need to be double word aligned.
2125 xlog_write_calc_vec_length(
2126 struct xlog_ticket *ticket,
2127 struct xfs_log_vec *log_vector)
2129 struct xfs_log_vec *lv;
2134 /* acct for start rec of xact */
2135 if (ticket->t_flags & XLOG_TIC_INITED)
2138 for (lv = log_vector; lv; lv = lv->lv_next) {
2139 /* we don't write ordered log vectors */
2140 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2143 headers += lv->lv_niovecs;
2145 for (i = 0; i < lv->lv_niovecs; i++) {
2146 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2149 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2153 ticket->t_res_num_ophdrs += headers;
2154 len += headers * sizeof(struct xlog_op_header);
2160 * If first write for transaction, insert start record We can't be trying to
2161 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2164 xlog_write_start_rec(
2165 struct xlog_op_header *ophdr,
2166 struct xlog_ticket *ticket)
2168 if (!(ticket->t_flags & XLOG_TIC_INITED))
2171 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2172 ophdr->oh_clientid = ticket->t_clientid;
2174 ophdr->oh_flags = XLOG_START_TRANS;
2177 ticket->t_flags &= ~XLOG_TIC_INITED;
2179 return sizeof(struct xlog_op_header);
2182 static xlog_op_header_t *
2183 xlog_write_setup_ophdr(
2185 struct xlog_op_header *ophdr,
2186 struct xlog_ticket *ticket,
2189 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2190 ophdr->oh_clientid = ticket->t_clientid;
2193 /* are we copying a commit or unmount record? */
2194 ophdr->oh_flags = flags;
2197 * We've seen logs corrupted with bad transaction client ids. This
2198 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2199 * and shut down the filesystem.
2201 switch (ophdr->oh_clientid) {
2202 case XFS_TRANSACTION:
2208 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2209 ophdr->oh_clientid, ticket);
2217 * Set up the parameters of the region copy into the log. This has
2218 * to handle region write split across multiple log buffers - this
2219 * state is kept external to this function so that this code can
2220 * be written in an obvious, self documenting manner.
2223 xlog_write_setup_copy(
2224 struct xlog_ticket *ticket,
2225 struct xlog_op_header *ophdr,
2226 int space_available,
2230 int *last_was_partial_copy,
2231 int *bytes_consumed)
2235 still_to_copy = space_required - *bytes_consumed;
2236 *copy_off = *bytes_consumed;
2238 if (still_to_copy <= space_available) {
2239 /* write of region completes here */
2240 *copy_len = still_to_copy;
2241 ophdr->oh_len = cpu_to_be32(*copy_len);
2242 if (*last_was_partial_copy)
2243 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2244 *last_was_partial_copy = 0;
2245 *bytes_consumed = 0;
2249 /* partial write of region, needs extra log op header reservation */
2250 *copy_len = space_available;
2251 ophdr->oh_len = cpu_to_be32(*copy_len);
2252 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2253 if (*last_was_partial_copy)
2254 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2255 *bytes_consumed += *copy_len;
2256 (*last_was_partial_copy)++;
2258 /* account for new log op header */
2259 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2260 ticket->t_res_num_ophdrs++;
2262 return sizeof(struct xlog_op_header);
2266 xlog_write_copy_finish(
2268 struct xlog_in_core *iclog,
2273 int *partial_copy_len,
2275 struct xlog_in_core **commit_iclog)
2279 if (*partial_copy) {
2281 * This iclog has already been marked WANT_SYNC by
2282 * xlog_state_get_iclog_space.
2284 spin_lock(&log->l_icloglock);
2285 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2292 *partial_copy_len = 0;
2294 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2295 /* no more space in this iclog - push it. */
2296 spin_lock(&log->l_icloglock);
2297 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2301 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2302 xlog_state_switch_iclogs(log, iclog, 0);
2304 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2305 iclog->ic_state == XLOG_STATE_IOERROR);
2308 spin_unlock(&log->l_icloglock);
2309 ASSERT(flags & XLOG_COMMIT_TRANS);
2310 *commit_iclog = iclog;
2316 error = xlog_state_release_iclog(log, iclog);
2317 spin_unlock(&log->l_icloglock);
2322 * Write some region out to in-core log
2324 * This will be called when writing externally provided regions or when
2325 * writing out a commit record for a given transaction.
2327 * General algorithm:
2328 * 1. Find total length of this write. This may include adding to the
2329 * lengths passed in.
2330 * 2. Check whether we violate the tickets reservation.
2331 * 3. While writing to this iclog
2332 * A. Reserve as much space in this iclog as can get
2333 * B. If this is first write, save away start lsn
2334 * C. While writing this region:
2335 * 1. If first write of transaction, write start record
2336 * 2. Write log operation header (header per region)
2337 * 3. Find out if we can fit entire region into this iclog
2338 * 4. Potentially, verify destination memcpy ptr
2339 * 5. Memcpy (partial) region
2340 * 6. If partial copy, release iclog; otherwise, continue
2341 * copying more regions into current iclog
2342 * 4. Mark want sync bit (in simulation mode)
2343 * 5. Release iclog for potential flush to on-disk log.
2346 * 1. Panic if reservation is overrun. This should never happen since
2347 * reservation amounts are generated internal to the filesystem.
2349 * 1. Tickets are single threaded data structures.
2350 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2351 * syncing routine. When a single log_write region needs to span
2352 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2353 * on all log operation writes which don't contain the end of the
2354 * region. The XLOG_END_TRANS bit is used for the in-core log
2355 * operation which contains the end of the continued log_write region.
2356 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2357 * we don't really know exactly how much space will be used. As a result,
2358 * we don't update ic_offset until the end when we know exactly how many
2359 * bytes have been written out.
2364 struct xfs_log_vec *log_vector,
2365 struct xlog_ticket *ticket,
2366 xfs_lsn_t *start_lsn,
2367 struct xlog_in_core **commit_iclog,
2370 struct xlog_in_core *iclog = NULL;
2371 struct xfs_log_iovec *vecp;
2372 struct xfs_log_vec *lv;
2375 int partial_copy = 0;
2376 int partial_copy_len = 0;
2384 len = xlog_write_calc_vec_length(ticket, log_vector);
2387 * Region headers and bytes are already accounted for.
2388 * We only need to take into account start records and
2389 * split regions in this function.
2391 if (ticket->t_flags & XLOG_TIC_INITED)
2392 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2395 * Commit record headers need to be accounted for. These
2396 * come in as separate writes so are easy to detect.
2398 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2399 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2401 if (ticket->t_curr_res < 0) {
2402 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2403 "ctx ticket reservation ran out. Need to up reservation");
2404 xlog_print_tic_res(log->l_mp, ticket);
2405 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2410 vecp = lv->lv_iovecp;
2411 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2415 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2416 &contwr, &log_offset);
2420 ASSERT(log_offset <= iclog->ic_size - 1);
2421 ptr = iclog->ic_datap + log_offset;
2423 /* start_lsn is the first lsn written to. That's all we need. */
2425 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2428 * This loop writes out as many regions as can fit in the amount
2429 * of space which was allocated by xlog_state_get_iclog_space().
2431 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2432 struct xfs_log_iovec *reg;
2433 struct xlog_op_header *ophdr;
2437 bool ordered = false;
2439 /* ordered log vectors have no regions to write */
2440 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2441 ASSERT(lv->lv_niovecs == 0);
2447 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2448 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2450 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2451 if (start_rec_copy) {
2453 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2457 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2461 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2462 sizeof(struct xlog_op_header));
2464 len += xlog_write_setup_copy(ticket, ophdr,
2465 iclog->ic_size-log_offset,
2467 ©_off, ©_len,
2470 xlog_verify_dest_ptr(log, ptr);
2475 * Unmount records just log an opheader, so can have
2476 * empty payloads with no data region to copy. Hence we
2477 * only copy the payload if the vector says it has data
2480 ASSERT(copy_len >= 0);
2482 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2483 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2486 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2488 data_cnt += contwr ? copy_len : 0;
2490 error = xlog_write_copy_finish(log, iclog, flags,
2491 &record_cnt, &data_cnt,
2500 * if we had a partial copy, we need to get more iclog
2501 * space but we don't want to increment the region
2502 * index because there is still more is this region to
2505 * If we completed writing this region, and we flushed
2506 * the iclog (indicated by resetting of the record
2507 * count), then we also need to get more log space. If
2508 * this was the last record, though, we are done and
2514 if (++index == lv->lv_niovecs) {
2519 vecp = lv->lv_iovecp;
2521 if (record_cnt == 0 && !ordered) {
2531 spin_lock(&log->l_icloglock);
2532 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2534 ASSERT(flags & XLOG_COMMIT_TRANS);
2535 *commit_iclog = iclog;
2537 error = xlog_state_release_iclog(log, iclog);
2539 spin_unlock(&log->l_icloglock);
2545 /*****************************************************************************
2547 * State Machine functions
2549 *****************************************************************************
2553 xlog_state_activate_iclog(
2554 struct xlog_in_core *iclog,
2555 int *iclogs_changed)
2557 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2560 * If the number of ops in this iclog indicate it just contains the
2561 * dummy transaction, we can change state into IDLE (the second time
2562 * around). Otherwise we should change the state into NEED a dummy.
2563 * We don't need to cover the dummy.
2565 if (*iclogs_changed == 0 &&
2566 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2567 *iclogs_changed = 1;
2570 * We have two dirty iclogs so start over. This could also be
2571 * num of ops indicating this is not the dummy going out.
2573 *iclogs_changed = 2;
2576 iclog->ic_state = XLOG_STATE_ACTIVE;
2577 iclog->ic_offset = 0;
2578 iclog->ic_header.h_num_logops = 0;
2579 memset(iclog->ic_header.h_cycle_data, 0,
2580 sizeof(iclog->ic_header.h_cycle_data));
2581 iclog->ic_header.h_lsn = 0;
2585 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2586 * ACTIVE after iclog I/O has completed.
2589 xlog_state_activate_iclogs(
2591 int *iclogs_changed)
2593 struct xlog_in_core *iclog = log->l_iclog;
2596 if (iclog->ic_state == XLOG_STATE_DIRTY)
2597 xlog_state_activate_iclog(iclog, iclogs_changed);
2599 * The ordering of marking iclogs ACTIVE must be maintained, so
2600 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2602 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2604 } while ((iclog = iclog->ic_next) != log->l_iclog);
2613 * We usually go to NEED. But we go to NEED2 if the changed indicates we
2614 * are done writing the dummy record. If we are done with the second
2615 * dummy recored (DONE2), then we go to IDLE.
2617 switch (prev_state) {
2618 case XLOG_STATE_COVER_IDLE:
2619 case XLOG_STATE_COVER_NEED:
2620 case XLOG_STATE_COVER_NEED2:
2622 case XLOG_STATE_COVER_DONE:
2623 if (iclogs_changed == 1)
2624 return XLOG_STATE_COVER_NEED2;
2626 case XLOG_STATE_COVER_DONE2:
2627 if (iclogs_changed == 1)
2628 return XLOG_STATE_COVER_IDLE;
2634 return XLOG_STATE_COVER_NEED;
2638 xlog_state_clean_iclog(
2640 struct xlog_in_core *dirty_iclog)
2642 int iclogs_changed = 0;
2644 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2646 xlog_state_activate_iclogs(log, &iclogs_changed);
2647 wake_up_all(&dirty_iclog->ic_force_wait);
2649 if (iclogs_changed) {
2650 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2656 xlog_get_lowest_lsn(
2659 struct xlog_in_core *iclog = log->l_iclog;
2660 xfs_lsn_t lowest_lsn = 0, lsn;
2663 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2664 iclog->ic_state == XLOG_STATE_DIRTY)
2667 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2668 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2670 } while ((iclog = iclog->ic_next) != log->l_iclog);
2676 * Completion of a iclog IO does not imply that a transaction has completed, as
2677 * transactions can be large enough to span many iclogs. We cannot change the
2678 * tail of the log half way through a transaction as this may be the only
2679 * transaction in the log and moving the tail to point to the middle of it
2680 * will prevent recovery from finding the start of the transaction. Hence we
2681 * should only update the last_sync_lsn if this iclog contains transaction
2682 * completion callbacks on it.
2684 * We have to do this before we drop the icloglock to ensure we are the only one
2685 * that can update it.
2687 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2688 * the reservation grant head pushing. This is due to the fact that the push
2689 * target is bound by the current last_sync_lsn value. Hence if we have a large
2690 * amount of log space bound up in this committing transaction then the
2691 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2692 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2693 * should push the AIL to ensure the push target (and hence the grant head) is
2694 * no longer bound by the old log head location and can move forwards and make
2698 xlog_state_set_callback(
2700 struct xlog_in_core *iclog,
2701 xfs_lsn_t header_lsn)
2703 iclog->ic_state = XLOG_STATE_CALLBACK;
2705 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2708 if (list_empty_careful(&iclog->ic_callbacks))
2711 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2712 xlog_grant_push_ail(log, 0);
2716 * Return true if we need to stop processing, false to continue to the next
2717 * iclog. The caller will need to run callbacks if the iclog is returned in the
2718 * XLOG_STATE_CALLBACK state.
2721 xlog_state_iodone_process_iclog(
2723 struct xlog_in_core *iclog,
2726 xfs_lsn_t lowest_lsn;
2727 xfs_lsn_t header_lsn;
2729 switch (iclog->ic_state) {
2730 case XLOG_STATE_ACTIVE:
2731 case XLOG_STATE_DIRTY:
2733 * Skip all iclogs in the ACTIVE & DIRTY states:
2736 case XLOG_STATE_IOERROR:
2738 * Between marking a filesystem SHUTDOWN and stopping the log,
2739 * we do flush all iclogs to disk (if there wasn't a log I/O
2740 * error). So, we do want things to go smoothly in case of just
2741 * a SHUTDOWN w/o a LOG_IO_ERROR.
2745 case XLOG_STATE_DONE_SYNC:
2747 * Now that we have an iclog that is in the DONE_SYNC state, do
2748 * one more check here to see if we have chased our tail around.
2749 * If this is not the lowest lsn iclog, then we will leave it
2750 * for another completion to process.
2752 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2753 lowest_lsn = xlog_get_lowest_lsn(log);
2754 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2756 xlog_state_set_callback(log, iclog, header_lsn);
2760 * Can only perform callbacks in order. Since this iclog is not
2761 * in the DONE_SYNC state, we skip the rest and just try to
2769 * Keep processing entries in the iclog callback list until we come around and
2770 * it is empty. We need to atomically see that the list is empty and change the
2771 * state to DIRTY so that we don't miss any more callbacks being added.
2773 * This function is called with the icloglock held and returns with it held. We
2774 * drop it while running callbacks, however, as holding it over thousands of
2775 * callbacks is unnecessary and causes excessive contention if we do.
2778 xlog_state_do_iclog_callbacks(
2780 struct xlog_in_core *iclog)
2781 __releases(&log->l_icloglock)
2782 __acquires(&log->l_icloglock)
2784 spin_unlock(&log->l_icloglock);
2785 spin_lock(&iclog->ic_callback_lock);
2786 while (!list_empty(&iclog->ic_callbacks)) {
2789 list_splice_init(&iclog->ic_callbacks, &tmp);
2791 spin_unlock(&iclog->ic_callback_lock);
2792 xlog_cil_process_committed(&tmp);
2793 spin_lock(&iclog->ic_callback_lock);
2797 * Pick up the icloglock while still holding the callback lock so we
2798 * serialise against anyone trying to add more callbacks to this iclog
2799 * now we've finished processing.
2801 spin_lock(&log->l_icloglock);
2802 spin_unlock(&iclog->ic_callback_lock);
2806 xlog_state_do_callback(
2809 struct xlog_in_core *iclog;
2810 struct xlog_in_core *first_iclog;
2811 bool cycled_icloglock;
2816 spin_lock(&log->l_icloglock);
2819 * Scan all iclogs starting with the one pointed to by the
2820 * log. Reset this starting point each time the log is
2821 * unlocked (during callbacks).
2823 * Keep looping through iclogs until one full pass is made
2824 * without running any callbacks.
2826 first_iclog = log->l_iclog;
2827 iclog = log->l_iclog;
2828 cycled_icloglock = false;
2833 if (xlog_state_iodone_process_iclog(log, iclog,
2837 if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2838 iclog->ic_state != XLOG_STATE_IOERROR) {
2839 iclog = iclog->ic_next;
2844 * Running callbacks will drop the icloglock which means
2845 * we'll have to run at least one more complete loop.
2847 cycled_icloglock = true;
2848 xlog_state_do_iclog_callbacks(log, iclog);
2849 if (XLOG_FORCED_SHUTDOWN(log))
2850 wake_up_all(&iclog->ic_force_wait);
2852 xlog_state_clean_iclog(log, iclog);
2853 iclog = iclog->ic_next;
2854 } while (first_iclog != iclog);
2856 if (repeats > 5000) {
2857 flushcnt += repeats;
2860 "%s: possible infinite loop (%d iterations)",
2861 __func__, flushcnt);
2863 } while (!ioerror && cycled_icloglock);
2865 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2866 log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2867 wake_up_all(&log->l_flush_wait);
2869 spin_unlock(&log->l_icloglock);
2874 * Finish transitioning this iclog to the dirty state.
2876 * Make sure that we completely execute this routine only when this is
2877 * the last call to the iclog. There is a good chance that iclog flushes,
2878 * when we reach the end of the physical log, get turned into 2 separate
2879 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2880 * routine. By using the reference count bwritecnt, we guarantee that only
2881 * the second completion goes through.
2883 * Callbacks could take time, so they are done outside the scope of the
2884 * global state machine log lock.
2887 xlog_state_done_syncing(
2888 struct xlog_in_core *iclog)
2890 struct xlog *log = iclog->ic_log;
2892 spin_lock(&log->l_icloglock);
2893 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2896 * If we got an error, either on the first buffer, or in the case of
2897 * split log writes, on the second, we shut down the file system and
2898 * no iclogs should ever be attempted to be written to disk again.
2900 if (!XLOG_FORCED_SHUTDOWN(log)) {
2901 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2902 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2906 * Someone could be sleeping prior to writing out the next
2907 * iclog buffer, we wake them all, one will get to do the
2908 * I/O, the others get to wait for the result.
2910 wake_up_all(&iclog->ic_write_wait);
2911 spin_unlock(&log->l_icloglock);
2912 xlog_state_do_callback(log); /* also cleans log */
2916 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2917 * sleep. We wait on the flush queue on the head iclog as that should be
2918 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2919 * we will wait here and all new writes will sleep until a sync completes.
2921 * The in-core logs are used in a circular fashion. They are not used
2922 * out-of-order even when an iclog past the head is free.
2925 * * log_offset where xlog_write() can start writing into the in-core
2927 * * in-core log pointer to which xlog_write() should write.
2928 * * boolean indicating this is a continued write to an in-core log.
2929 * If this is the last write, then the in-core log's offset field
2930 * needs to be incremented, depending on the amount of data which
2934 xlog_state_get_iclog_space(
2937 struct xlog_in_core **iclogp,
2938 struct xlog_ticket *ticket,
2939 int *continued_write,
2943 xlog_rec_header_t *head;
2944 xlog_in_core_t *iclog;
2947 spin_lock(&log->l_icloglock);
2948 if (XLOG_FORCED_SHUTDOWN(log)) {
2949 spin_unlock(&log->l_icloglock);
2953 iclog = log->l_iclog;
2954 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2955 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2957 /* Wait for log writes to have flushed */
2958 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2962 head = &iclog->ic_header;
2964 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2965 log_offset = iclog->ic_offset;
2967 /* On the 1st write to an iclog, figure out lsn. This works
2968 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2969 * committing to. If the offset is set, that's how many blocks
2972 if (log_offset == 0) {
2973 ticket->t_curr_res -= log->l_iclog_hsize;
2974 xlog_tic_add_region(ticket,
2976 XLOG_REG_TYPE_LRHEADER);
2977 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2978 head->h_lsn = cpu_to_be64(
2979 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2980 ASSERT(log->l_curr_block >= 0);
2983 /* If there is enough room to write everything, then do it. Otherwise,
2984 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2985 * bit is on, so this will get flushed out. Don't update ic_offset
2986 * until you know exactly how many bytes get copied. Therefore, wait
2987 * until later to update ic_offset.
2989 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2990 * can fit into remaining data section.
2992 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2995 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2998 * If we are the only one writing to this iclog, sync it to
2999 * disk. We need to do an atomic compare and decrement here to
3000 * avoid racing with concurrent atomic_dec_and_lock() calls in
3001 * xlog_state_release_iclog() when there is more than one
3002 * reference to the iclog.
3004 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3005 error = xlog_state_release_iclog(log, iclog);
3006 spin_unlock(&log->l_icloglock);
3012 /* Do we have enough room to write the full amount in the remainder
3013 * of this iclog? Or must we continue a write on the next iclog and
3014 * mark this iclog as completely taken? In the case where we switch
3015 * iclogs (to mark it taken), this particular iclog will release/sync
3016 * to disk in xlog_write().
3018 if (len <= iclog->ic_size - iclog->ic_offset) {
3019 *continued_write = 0;
3020 iclog->ic_offset += len;
3022 *continued_write = 1;
3023 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3027 ASSERT(iclog->ic_offset <= iclog->ic_size);
3028 spin_unlock(&log->l_icloglock);
3030 *logoffsetp = log_offset;
3032 } /* xlog_state_get_iclog_space */
3034 /* The first cnt-1 times through here we don't need to
3035 * move the grant write head because the permanent
3036 * reservation has reserved cnt times the unit amount.
3037 * Release part of current permanent unit reservation and
3038 * reset current reservation to be one units worth. Also
3039 * move grant reservation head forward.
3042 xlog_regrant_reserve_log_space(
3044 struct xlog_ticket *ticket)
3046 trace_xfs_log_regrant_reserve_enter(log, ticket);
3048 if (ticket->t_cnt > 0)
3051 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3052 ticket->t_curr_res);
3053 xlog_grant_sub_space(log, &log->l_write_head.grant,
3054 ticket->t_curr_res);
3055 ticket->t_curr_res = ticket->t_unit_res;
3056 xlog_tic_reset_res(ticket);
3058 trace_xfs_log_regrant_reserve_sub(log, ticket);
3060 /* just return if we still have some of the pre-reserved space */
3061 if (ticket->t_cnt > 0)
3064 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3065 ticket->t_unit_res);
3067 trace_xfs_log_regrant_reserve_exit(log, ticket);
3069 ticket->t_curr_res = ticket->t_unit_res;
3070 xlog_tic_reset_res(ticket);
3071 } /* xlog_regrant_reserve_log_space */
3075 * Give back the space left from a reservation.
3077 * All the information we need to make a correct determination of space left
3078 * is present. For non-permanent reservations, things are quite easy. The
3079 * count should have been decremented to zero. We only need to deal with the
3080 * space remaining in the current reservation part of the ticket. If the
3081 * ticket contains a permanent reservation, there may be left over space which
3082 * needs to be released. A count of N means that N-1 refills of the current
3083 * reservation can be done before we need to ask for more space. The first
3084 * one goes to fill up the first current reservation. Once we run out of
3085 * space, the count will stay at zero and the only space remaining will be
3086 * in the current reservation field.
3089 xlog_ungrant_log_space(
3091 struct xlog_ticket *ticket)
3095 if (ticket->t_cnt > 0)
3098 trace_xfs_log_ungrant_enter(log, ticket);
3099 trace_xfs_log_ungrant_sub(log, ticket);
3102 * If this is a permanent reservation ticket, we may be able to free
3103 * up more space based on the remaining count.
3105 bytes = ticket->t_curr_res;
3106 if (ticket->t_cnt > 0) {
3107 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3108 bytes += ticket->t_unit_res*ticket->t_cnt;
3111 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3112 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3114 trace_xfs_log_ungrant_exit(log, ticket);
3116 xfs_log_space_wake(log->l_mp);
3120 * Mark the current iclog in the ring as WANT_SYNC and move the current iclog
3121 * pointer to the next iclog in the ring.
3123 * When called from xlog_state_get_iclog_space(), the exact size of the iclog
3124 * has not yet been determined, all we know is that we have run out of space in
3125 * the current iclog.
3128 xlog_state_switch_iclogs(
3130 struct xlog_in_core *iclog,
3133 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3134 assert_spin_locked(&log->l_icloglock);
3137 eventual_size = iclog->ic_offset;
3138 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3139 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3140 log->l_prev_block = log->l_curr_block;
3141 log->l_prev_cycle = log->l_curr_cycle;
3143 /* roll log?: ic_offset changed later */
3144 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3146 /* Round up to next log-sunit */
3147 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3148 log->l_mp->m_sb.sb_logsunit > 1) {
3149 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3150 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3153 if (log->l_curr_block >= log->l_logBBsize) {
3155 * Rewind the current block before the cycle is bumped to make
3156 * sure that the combined LSN never transiently moves forward
3157 * when the log wraps to the next cycle. This is to support the
3158 * unlocked sample of these fields from xlog_valid_lsn(). Most
3159 * other cases should acquire l_icloglock.
3161 log->l_curr_block -= log->l_logBBsize;
3162 ASSERT(log->l_curr_block >= 0);
3164 log->l_curr_cycle++;
3165 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3166 log->l_curr_cycle++;
3168 ASSERT(iclog == log->l_iclog);
3169 log->l_iclog = iclog->ic_next;
3170 } /* xlog_state_switch_iclogs */
3173 * Write out all data in the in-core log as of this exact moment in time.
3175 * Data may be written to the in-core log during this call. However,
3176 * we don't guarantee this data will be written out. A change from past
3177 * implementation means this routine will *not* write out zero length LRs.
3179 * Basically, we try and perform an intelligent scan of the in-core logs.
3180 * If we determine there is no flushable data, we just return. There is no
3181 * flushable data if:
3183 * 1. the current iclog is active and has no data; the previous iclog
3184 * is in the active or dirty state.
3185 * 2. the current iclog is drity, and the previous iclog is in the
3186 * active or dirty state.
3190 * 1. the current iclog is not in the active nor dirty state.
3191 * 2. the current iclog dirty, and the previous iclog is not in the
3192 * active nor dirty state.
3193 * 3. the current iclog is active, and there is another thread writing
3194 * to this particular iclog.
3195 * 4. a) the current iclog is active and has no other writers
3196 * b) when we return from flushing out this iclog, it is still
3197 * not in the active nor dirty state.
3201 struct xfs_mount *mp,
3204 struct xlog *log = mp->m_log;
3205 struct xlog_in_core *iclog;
3208 XFS_STATS_INC(mp, xs_log_force);
3209 trace_xfs_log_force(mp, 0, _RET_IP_);
3211 xlog_cil_force(log);
3213 spin_lock(&log->l_icloglock);
3214 iclog = log->l_iclog;
3215 if (iclog->ic_state == XLOG_STATE_IOERROR)
3218 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3219 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3220 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3222 * If the head is dirty or (active and empty), then we need to
3223 * look at the previous iclog.
3225 * If the previous iclog is active or dirty we are done. There
3226 * is nothing to sync out. Otherwise, we attach ourselves to the
3227 * previous iclog and go to sleep.
3229 iclog = iclog->ic_prev;
3230 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3231 if (atomic_read(&iclog->ic_refcnt) == 0) {
3233 * We are the only one with access to this iclog.
3235 * Flush it out now. There should be a roundoff of zero
3236 * to show that someone has already taken care of the
3237 * roundoff from the previous sync.
3239 atomic_inc(&iclog->ic_refcnt);
3240 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3241 xlog_state_switch_iclogs(log, iclog, 0);
3242 if (xlog_state_release_iclog(log, iclog))
3245 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3249 * Someone else is writing to this iclog.
3251 * Use its call to flush out the data. However, the
3252 * other thread may not force out this LR, so we mark
3255 xlog_state_switch_iclogs(log, iclog, 0);
3259 * If the head iclog is not active nor dirty, we just attach
3260 * ourselves to the head and go to sleep if necessary.
3265 if (flags & XFS_LOG_SYNC)
3266 return xlog_wait_on_iclog(iclog);
3268 spin_unlock(&log->l_icloglock);
3271 spin_unlock(&log->l_icloglock);
3276 __xfs_log_force_lsn(
3277 struct xfs_mount *mp,
3283 struct xlog *log = mp->m_log;
3284 struct xlog_in_core *iclog;
3286 spin_lock(&log->l_icloglock);
3287 iclog = log->l_iclog;
3288 if (iclog->ic_state == XLOG_STATE_IOERROR)
3291 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3292 iclog = iclog->ic_next;
3293 if (iclog == log->l_iclog)
3297 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3299 * We sleep here if we haven't already slept (e.g. this is the
3300 * first time we've looked at the correct iclog buf) and the
3301 * buffer before us is going to be sync'ed. The reason for this
3302 * is that if we are doing sync transactions here, by waiting
3303 * for the previous I/O to complete, we can allow a few more
3304 * transactions into this iclog before we close it down.
3306 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3307 * refcnt so we can release the log (which drops the ref count).
3308 * The state switch keeps new transaction commits from using
3309 * this buffer. When the current commits finish writing into
3310 * the buffer, the refcount will drop to zero and the buffer
3313 if (!already_slept &&
3314 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3315 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3316 XFS_STATS_INC(mp, xs_log_force_sleep);
3318 xlog_wait(&iclog->ic_prev->ic_write_wait,
3322 atomic_inc(&iclog->ic_refcnt);
3323 xlog_state_switch_iclogs(log, iclog, 0);
3324 if (xlog_state_release_iclog(log, iclog))
3330 if (flags & XFS_LOG_SYNC)
3331 return xlog_wait_on_iclog(iclog);
3333 spin_unlock(&log->l_icloglock);
3336 spin_unlock(&log->l_icloglock);
3341 * Force the in-core log to disk for a specific LSN.
3343 * Find in-core log with lsn.
3344 * If it is in the DIRTY state, just return.
3345 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3346 * state and go to sleep or return.
3347 * If it is in any other state, go to sleep or return.
3349 * Synchronous forces are implemented with a wait queue. All callers trying
3350 * to force a given lsn to disk must wait on the queue attached to the
3351 * specific in-core log. When given in-core log finally completes its write
3352 * to disk, that thread will wake up all threads waiting on the queue.
3356 struct xfs_mount *mp,
3364 XFS_STATS_INC(mp, xs_log_force);
3365 trace_xfs_log_force(mp, lsn, _RET_IP_);
3367 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3368 if (lsn == NULLCOMMITLSN)
3371 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3373 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3377 /*****************************************************************************
3381 *****************************************************************************
3385 * Free a used ticket when its refcount falls to zero.
3389 xlog_ticket_t *ticket)
3391 ASSERT(atomic_read(&ticket->t_ref) > 0);
3392 if (atomic_dec_and_test(&ticket->t_ref))
3393 kmem_cache_free(xfs_log_ticket_zone, ticket);
3398 xlog_ticket_t *ticket)
3400 ASSERT(atomic_read(&ticket->t_ref) > 0);
3401 atomic_inc(&ticket->t_ref);
3406 * Figure out the total log space unit (in bytes) that would be
3407 * required for a log ticket.
3410 xfs_log_calc_unit_res(
3411 struct xfs_mount *mp,
3414 struct xlog *log = mp->m_log;
3419 * Permanent reservations have up to 'cnt'-1 active log operations
3420 * in the log. A unit in this case is the amount of space for one
3421 * of these log operations. Normal reservations have a cnt of 1
3422 * and their unit amount is the total amount of space required.
3424 * The following lines of code account for non-transaction data
3425 * which occupy space in the on-disk log.
3427 * Normal form of a transaction is:
3428 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3429 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3431 * We need to account for all the leadup data and trailer data
3432 * around the transaction data.
3433 * And then we need to account for the worst case in terms of using
3435 * The worst case will happen if:
3436 * - the placement of the transaction happens to be such that the
3437 * roundoff is at its maximum
3438 * - the transaction data is synced before the commit record is synced
3439 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3440 * Therefore the commit record is in its own Log Record.
3441 * This can happen as the commit record is called with its
3442 * own region to xlog_write().
3443 * This then means that in the worst case, roundoff can happen for
3444 * the commit-rec as well.
3445 * The commit-rec is smaller than padding in this scenario and so it is
3446 * not added separately.
3449 /* for trans header */
3450 unit_bytes += sizeof(xlog_op_header_t);
3451 unit_bytes += sizeof(xfs_trans_header_t);
3454 unit_bytes += sizeof(xlog_op_header_t);
3457 * for LR headers - the space for data in an iclog is the size minus
3458 * the space used for the headers. If we use the iclog size, then we
3459 * undercalculate the number of headers required.
3461 * Furthermore - the addition of op headers for split-recs might
3462 * increase the space required enough to require more log and op
3463 * headers, so take that into account too.
3465 * IMPORTANT: This reservation makes the assumption that if this
3466 * transaction is the first in an iclog and hence has the LR headers
3467 * accounted to it, then the remaining space in the iclog is
3468 * exclusively for this transaction. i.e. if the transaction is larger
3469 * than the iclog, it will be the only thing in that iclog.
3470 * Fundamentally, this means we must pass the entire log vector to
3471 * xlog_write to guarantee this.
3473 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3474 num_headers = howmany(unit_bytes, iclog_space);
3476 /* for split-recs - ophdrs added when data split over LRs */
3477 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3479 /* add extra header reservations if we overrun */
3480 while (!num_headers ||
3481 howmany(unit_bytes, iclog_space) > num_headers) {
3482 unit_bytes += sizeof(xlog_op_header_t);
3485 unit_bytes += log->l_iclog_hsize * num_headers;
3487 /* for commit-rec LR header - note: padding will subsume the ophdr */
3488 unit_bytes += log->l_iclog_hsize;
3490 /* for roundoff padding for transaction data and one for commit record */
3491 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3492 /* log su roundoff */
3493 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3496 unit_bytes += 2 * BBSIZE;
3503 * Allocate and initialise a new log ticket.
3505 struct xlog_ticket *
3512 xfs_km_flags_t alloc_flags)
3514 struct xlog_ticket *tic;
3517 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3521 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3523 atomic_set(&tic->t_ref, 1);
3524 tic->t_task = current;
3525 INIT_LIST_HEAD(&tic->t_queue);
3526 tic->t_unit_res = unit_res;
3527 tic->t_curr_res = unit_res;
3530 tic->t_tid = prandom_u32();
3531 tic->t_clientid = client;
3532 tic->t_flags = XLOG_TIC_INITED;
3534 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3536 xlog_tic_reset_res(tic);
3542 /******************************************************************************
3544 * Log debug routines
3546 ******************************************************************************
3550 * Make sure that the destination ptr is within the valid data region of
3551 * one of the iclogs. This uses backup pointers stored in a different
3552 * part of the log in case we trash the log structure.
3555 xlog_verify_dest_ptr(
3562 for (i = 0; i < log->l_iclog_bufs; i++) {
3563 if (ptr >= log->l_iclog_bak[i] &&
3564 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3569 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3573 * Check to make sure the grant write head didn't just over lap the tail. If
3574 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3575 * the cycles differ by exactly one and check the byte count.
3577 * This check is run unlocked, so can give false positives. Rather than assert
3578 * on failures, use a warn-once flag and a panic tag to allow the admin to
3579 * determine if they want to panic the machine when such an error occurs. For
3580 * debug kernels this will have the same effect as using an assert but, unlinke
3581 * an assert, it can be turned off at runtime.
3584 xlog_verify_grant_tail(
3587 int tail_cycle, tail_blocks;
3590 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3591 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3592 if (tail_cycle != cycle) {
3593 if (cycle - 1 != tail_cycle &&
3594 !(log->l_flags & XLOG_TAIL_WARN)) {
3595 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3596 "%s: cycle - 1 != tail_cycle", __func__);
3597 log->l_flags |= XLOG_TAIL_WARN;
3600 if (space > BBTOB(tail_blocks) &&
3601 !(log->l_flags & XLOG_TAIL_WARN)) {
3602 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3603 "%s: space > BBTOB(tail_blocks)", __func__);
3604 log->l_flags |= XLOG_TAIL_WARN;
3609 /* check if it will fit */
3611 xlog_verify_tail_lsn(
3613 struct xlog_in_core *iclog,
3618 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3620 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3621 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3622 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3624 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3626 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3627 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3629 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3630 if (blocks < BTOBB(iclog->ic_offset) + 1)
3631 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3633 } /* xlog_verify_tail_lsn */
3636 * Perform a number of checks on the iclog before writing to disk.
3638 * 1. Make sure the iclogs are still circular
3639 * 2. Make sure we have a good magic number
3640 * 3. Make sure we don't have magic numbers in the data
3641 * 4. Check fields of each log operation header for:
3642 * A. Valid client identifier
3643 * B. tid ptr value falls in valid ptr space (user space code)
3644 * C. Length in log record header is correct according to the
3645 * individual operation headers within record.
3646 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3647 * log, check the preceding blocks of the physical log to make sure all
3648 * the cycle numbers agree with the current cycle number.
3653 struct xlog_in_core *iclog,
3656 xlog_op_header_t *ophead;
3657 xlog_in_core_t *icptr;
3658 xlog_in_core_2_t *xhdr;
3659 void *base_ptr, *ptr, *p;
3660 ptrdiff_t field_offset;
3662 int len, i, j, k, op_len;
3665 /* check validity of iclog pointers */
3666 spin_lock(&log->l_icloglock);
3667 icptr = log->l_iclog;
3668 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3671 if (icptr != log->l_iclog)
3672 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3673 spin_unlock(&log->l_icloglock);
3675 /* check log magic numbers */
3676 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3677 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3679 base_ptr = ptr = &iclog->ic_header;
3680 p = &iclog->ic_header;
3681 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3682 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3683 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3688 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3689 base_ptr = ptr = iclog->ic_datap;
3691 xhdr = iclog->ic_data;
3692 for (i = 0; i < len; i++) {
3695 /* clientid is only 1 byte */
3696 p = &ophead->oh_clientid;
3697 field_offset = p - base_ptr;
3698 if (field_offset & 0x1ff) {
3699 clientid = ophead->oh_clientid;
3701 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3702 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3703 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3704 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3705 clientid = xlog_get_client_id(
3706 xhdr[j].hic_xheader.xh_cycle_data[k]);
3708 clientid = xlog_get_client_id(
3709 iclog->ic_header.h_cycle_data[idx]);
3712 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3714 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3715 __func__, clientid, ophead,
3716 (unsigned long)field_offset);
3719 p = &ophead->oh_len;
3720 field_offset = p - base_ptr;
3721 if (field_offset & 0x1ff) {
3722 op_len = be32_to_cpu(ophead->oh_len);
3724 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3725 (uintptr_t)iclog->ic_datap);
3726 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3727 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3728 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3729 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3731 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3734 ptr += sizeof(xlog_op_header_t) + op_len;
3736 } /* xlog_verify_iclog */
3740 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3746 xlog_in_core_t *iclog, *ic;
3748 iclog = log->l_iclog;
3749 if (iclog->ic_state != XLOG_STATE_IOERROR) {
3751 * Mark all the incore logs IOERROR.
3752 * From now on, no log flushes will result.
3756 ic->ic_state = XLOG_STATE_IOERROR;
3758 } while (ic != iclog);
3762 * Return non-zero, if state transition has already happened.
3768 * This is called from xfs_force_shutdown, when we're forcibly
3769 * shutting down the filesystem, typically because of an IO error.
3770 * Our main objectives here are to make sure that:
3771 * a. if !logerror, flush the logs to disk. Anything modified
3772 * after this is ignored.
3773 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3774 * parties to find out, 'atomically'.
3775 * c. those who're sleeping on log reservations, pinned objects and
3776 * other resources get woken up, and be told the bad news.
3777 * d. nothing new gets queued up after (b) and (c) are done.
3779 * Note: for the !logerror case we need to flush the regions held in memory out
3780 * to disk first. This needs to be done before the log is marked as shutdown,
3781 * otherwise the iclog writes will fail.
3784 xfs_log_force_umount(
3785 struct xfs_mount *mp,
3794 * If this happens during log recovery, don't worry about
3795 * locking; the log isn't open for business yet.
3798 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3799 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3801 mp->m_sb_bp->b_flags |= XBF_DONE;
3806 * Somebody could've already done the hard work for us.
3807 * No need to get locks for this.
3809 if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3810 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3815 * Flush all the completed transactions to disk before marking the log
3816 * being shut down. We need to do it in this order to ensure that
3817 * completed operations are safely on disk before we shut down, and that
3818 * we don't have to issue any buffer IO after the shutdown flags are set
3819 * to guarantee this.
3822 xfs_log_force(mp, XFS_LOG_SYNC);
3825 * mark the filesystem and the as in a shutdown state and wake
3826 * everybody up to tell them the bad news.
3828 spin_lock(&log->l_icloglock);
3829 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3831 mp->m_sb_bp->b_flags |= XBF_DONE;
3834 * Mark the log and the iclogs with IO error flags to prevent any
3835 * further log IO from being issued or completed.
3837 log->l_flags |= XLOG_IO_ERROR;
3838 retval = xlog_state_ioerror(log);
3839 spin_unlock(&log->l_icloglock);
3842 * We don't want anybody waiting for log reservations after this. That
3843 * means we have to wake up everybody queued up on reserveq as well as
3844 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3845 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3846 * action is protected by the grant locks.
3848 xlog_grant_head_wake_all(&log->l_reserve_head);
3849 xlog_grant_head_wake_all(&log->l_write_head);
3852 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3853 * as if the log writes were completed. The abort handling in the log
3854 * item committed callback functions will do this again under lock to
3857 spin_lock(&log->l_cilp->xc_push_lock);
3858 wake_up_all(&log->l_cilp->xc_commit_wait);
3859 spin_unlock(&log->l_cilp->xc_push_lock);
3860 xlog_state_do_callback(log);
3862 /* return non-zero if log IOERROR transition had already happened */
3870 xlog_in_core_t *iclog;
3872 iclog = log->l_iclog;
3874 /* endianness does not matter here, zero is zero in
3877 if (iclog->ic_header.h_num_logops)
3879 iclog = iclog->ic_next;
3880 } while (iclog != log->l_iclog);
3885 * Verify that an LSN stamped into a piece of metadata is valid. This is
3886 * intended for use in read verifiers on v5 superblocks.
3890 struct xfs_mount *mp,
3893 struct xlog *log = mp->m_log;
3897 * norecovery mode skips mount-time log processing and unconditionally
3898 * resets the in-core LSN. We can't validate in this mode, but
3899 * modifications are not allowed anyways so just return true.
3901 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3905 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3906 * handled by recovery and thus safe to ignore here.
3908 if (lsn == NULLCOMMITLSN)
3911 valid = xlog_valid_lsn(mp->m_log, lsn);
3913 /* warn the user about what's gone wrong before verifier failure */
3915 spin_lock(&log->l_icloglock);
3917 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3918 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3919 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3920 log->l_curr_cycle, log->l_curr_block);
3921 spin_unlock(&log->l_icloglock);
3928 xfs_log_in_recovery(
3929 struct xfs_mount *mp)
3931 struct xlog *log = mp->m_log;
3933 return log->l_flags & XLOG_ACTIVE_RECOVERY;