2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_error.h"
26 #include "xfs_trans.h"
27 #include "xfs_trans_priv.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_log_recover.h"
31 #include "xfs_inode.h"
32 #include "xfs_trace.h"
33 #include "xfs_fsops.h"
34 #include "xfs_cksum.h"
35 #include "xfs_sysfs.h"
38 kmem_zone_t *xfs_log_ticket_zone;
40 /* Local miscellaneous function prototypes */
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
61 struct xlog_in_core *iclog);
66 /* local state machine functions */
67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
69 xlog_state_do_callback(
72 struct xlog_in_core *iclog);
74 xlog_state_get_iclog_space(
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
82 xlog_state_release_iclog(
84 struct xlog_in_core *iclog);
86 xlog_state_switch_iclogs(
88 struct xlog_in_core *iclog,
93 struct xlog_in_core *iclog);
100 xlog_regrant_reserve_log_space(
102 struct xlog_ticket *ticket);
104 xlog_ungrant_log_space(
106 struct xlog_ticket *ticket);
110 xlog_verify_dest_ptr(
114 xlog_verify_grant_tail(
119 struct xlog_in_core *iclog,
123 xlog_verify_tail_lsn(
125 struct xlog_in_core *iclog,
128 #define xlog_verify_dest_ptr(a,b)
129 #define xlog_verify_grant_tail(a)
130 #define xlog_verify_iclog(a,b,c,d)
131 #define xlog_verify_tail_lsn(a,b,c)
139 xlog_grant_sub_space(
144 int64_t head_val = atomic64_read(head);
150 xlog_crack_grant_head_val(head_val, &cycle, &space);
154 space += log->l_logsize;
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_add_space(
170 int64_t head_val = atomic64_read(head);
177 xlog_crack_grant_head_val(head_val, &cycle, &space);
179 tmp = log->l_logsize - space;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
194 xlog_grant_head_init(
195 struct xlog_grant_head *head)
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
203 xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
206 struct xlog_ticket *tic;
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
215 xlog_ticket_reservation(
217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
225 return tic->t_unit_res * tic->t_cnt;
227 return tic->t_unit_res;
232 xlog_grant_head_wake(
234 struct xlog_grant_head *head,
237 struct xlog_ticket *tic;
240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
242 if (*free_bytes < need_bytes)
245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
247 wake_up_process(tic->t_task);
254 xlog_grant_head_wait(
256 struct xlog_grant_head *head,
257 struct xlog_ticket *tic,
258 int need_bytes) __releases(&head->lock)
259 __acquires(&head->lock)
261 list_add_tail(&tic->t_queue, &head->waiters);
264 if (XLOG_FORCED_SHUTDOWN(log))
266 xlog_grant_push_ail(log, need_bytes);
268 __set_current_state(TASK_UNINTERRUPTIBLE);
269 spin_unlock(&head->lock);
271 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
273 trace_xfs_log_grant_sleep(log, tic);
275 trace_xfs_log_grant_wake(log, tic);
277 spin_lock(&head->lock);
278 if (XLOG_FORCED_SHUTDOWN(log))
280 } while (xlog_space_left(log, &head->grant) < need_bytes);
282 list_del_init(&tic->t_queue);
285 list_del_init(&tic->t_queue);
290 * Atomically get the log space required for a log ticket.
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
307 xlog_grant_head_check(
309 struct xlog_grant_head *head,
310 struct xlog_ticket *tic,
316 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
324 *need_bytes = xlog_ticket_reservation(log, head, tic);
325 free_bytes = xlog_space_left(log, &head->grant);
326 if (!list_empty_careful(&head->waiters)) {
327 spin_lock(&head->lock);
328 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
329 free_bytes < *need_bytes) {
330 error = xlog_grant_head_wait(log, head, tic,
333 spin_unlock(&head->lock);
334 } else if (free_bytes < *need_bytes) {
335 spin_lock(&head->lock);
336 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
337 spin_unlock(&head->lock);
344 xlog_tic_reset_res(xlog_ticket_t *tic)
347 tic->t_res_arr_sum = 0;
348 tic->t_res_num_ophdrs = 0;
352 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
354 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
355 /* add to overflow and start again */
356 tic->t_res_o_flow += tic->t_res_arr_sum;
358 tic->t_res_arr_sum = 0;
361 tic->t_res_arr[tic->t_res_num].r_len = len;
362 tic->t_res_arr[tic->t_res_num].r_type = type;
363 tic->t_res_arr_sum += len;
368 * Replenish the byte reservation required by moving the grant write head.
372 struct xfs_mount *mp,
373 struct xlog_ticket *tic)
375 struct xlog *log = mp->m_log;
379 if (XLOG_FORCED_SHUTDOWN(log))
382 XFS_STATS_INC(mp, xs_try_logspace);
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
392 xlog_grant_push_ail(log, tic->t_unit_res);
394 tic->t_curr_res = tic->t_unit_res;
395 xlog_tic_reset_res(tic);
400 trace_xfs_log_regrant(log, tic);
402 error = xlog_grant_head_check(log, &log->l_write_head, tic,
407 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
408 trace_xfs_log_regrant_exit(log, tic);
409 xlog_verify_grant_tail(log);
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
419 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
424 * Reserve log space and return a ticket corresponding the reservation.
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
433 struct xfs_mount *mp,
436 struct xlog_ticket **ticp,
440 struct xlog *log = mp->m_log;
441 struct xlog_ticket *tic;
445 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
447 if (XLOG_FORCED_SHUTDOWN(log))
450 XFS_STATS_INC(mp, xs_try_logspace);
452 ASSERT(*ticp == NULL);
453 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
454 KM_SLEEP | KM_MAYFAIL);
460 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
463 trace_xfs_log_reserve(log, tic);
465 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
470 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
471 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
472 trace_xfs_log_reserve_exit(log, tic);
473 xlog_verify_grant_tail(log);
478 * If we are failing, make sure the ticket doesn't have any current
479 * reservations. We don't want to add this back when the ticket/
480 * transaction gets cancelled.
483 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
491 * 1. currblock field gets updated at startup and after in-core logs
492 * marked as with WANT_SYNC.
496 * This routine is called when a user of a log manager ticket is done with
497 * the reservation. If the ticket was ever used, then a commit record for
498 * the associated transaction is written out as a log operation header with
499 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
500 * a given ticket. If the ticket was one with a permanent reservation, then
501 * a few operations are done differently. Permanent reservation tickets by
502 * default don't release the reservation. They just commit the current
503 * transaction with the belief that the reservation is still needed. A flag
504 * must be passed in before permanent reservations are actually released.
505 * When these type of tickets are not released, they need to be set into
506 * the inited state again. By doing this, a start record will be written
507 * out when the next write occurs.
511 struct xfs_mount *mp,
512 struct xlog_ticket *ticket,
513 struct xlog_in_core **iclog,
516 struct xlog *log = mp->m_log;
519 if (XLOG_FORCED_SHUTDOWN(log) ||
521 * If nothing was ever written, don't write out commit record.
522 * If we get an error, just continue and give back the log ticket.
524 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
525 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
526 lsn = (xfs_lsn_t) -1;
532 trace_xfs_log_done_nonperm(log, ticket);
535 * Release ticket if not permanent reservation or a specific
536 * request has been made to release a permanent reservation.
538 xlog_ungrant_log_space(log, ticket);
540 trace_xfs_log_done_perm(log, ticket);
542 xlog_regrant_reserve_log_space(log, ticket);
543 /* If this ticket was a permanent reservation and we aren't
544 * trying to release it, reset the inited flags; so next time
545 * we write, a start record will be written out.
547 ticket->t_flags |= XLOG_TIC_INITED;
550 xfs_log_ticket_put(ticket);
555 * Attaches a new iclog I/O completion callback routine during
556 * transaction commit. If the log is in error state, a non-zero
557 * return code is handed back and the caller is responsible for
558 * executing the callback at an appropriate time.
562 struct xfs_mount *mp,
563 struct xlog_in_core *iclog,
564 xfs_log_callback_t *cb)
568 spin_lock(&iclog->ic_callback_lock);
569 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
571 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
572 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
574 *(iclog->ic_callback_tail) = cb;
575 iclog->ic_callback_tail = &(cb->cb_next);
577 spin_unlock(&iclog->ic_callback_lock);
582 xfs_log_release_iclog(
583 struct xfs_mount *mp,
584 struct xlog_in_core *iclog)
586 if (xlog_state_release_iclog(mp->m_log, iclog)) {
587 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
595 * Mount a log filesystem
597 * mp - ubiquitous xfs mount point structure
598 * log_target - buftarg of on-disk log device
599 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
600 * num_bblocks - Number of BBSIZE blocks in on-disk log
602 * Return error or zero.
607 xfs_buftarg_t *log_target,
608 xfs_daddr_t blk_offset,
614 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
615 xfs_notice(mp, "Mounting V%d Filesystem",
616 XFS_SB_VERSION_NUM(&mp->m_sb));
619 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
620 XFS_SB_VERSION_NUM(&mp->m_sb));
621 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
624 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
625 if (IS_ERR(mp->m_log)) {
626 error = PTR_ERR(mp->m_log);
631 * Validate the given log space and drop a critical message via syslog
632 * if the log size is too small that would lead to some unexpected
633 * situations in transaction log space reservation stage.
635 * Note: we can't just reject the mount if the validation fails. This
636 * would mean that people would have to downgrade their kernel just to
637 * remedy the situation as there is no way to grow the log (short of
638 * black magic surgery with xfs_db).
640 * We can, however, reject mounts for CRC format filesystems, as the
641 * mkfs binary being used to make the filesystem should never create a
642 * filesystem with a log that is too small.
644 min_logfsbs = xfs_log_calc_minimum_size(mp);
646 if (mp->m_sb.sb_logblocks < min_logfsbs) {
648 "Log size %d blocks too small, minimum size is %d blocks",
649 mp->m_sb.sb_logblocks, min_logfsbs);
651 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
653 "Log size %d blocks too large, maximum size is %lld blocks",
654 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
656 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
658 "log size %lld bytes too large, maximum size is %lld bytes",
659 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
664 if (xfs_sb_version_hascrc(&mp->m_sb)) {
665 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
669 xfs_crit(mp, "Log size out of supported range.");
671 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
675 * Initialize the AIL now we have a log.
677 error = xfs_trans_ail_init(mp);
679 xfs_warn(mp, "AIL initialisation failed: error %d", error);
682 mp->m_log->l_ailp = mp->m_ail;
685 * skip log recovery on a norecovery mount. pretend it all
688 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
689 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
692 mp->m_flags &= ~XFS_MOUNT_RDONLY;
694 error = xlog_recover(mp->m_log);
697 mp->m_flags |= XFS_MOUNT_RDONLY;
699 xfs_warn(mp, "log mount/recovery failed: error %d",
701 xlog_recover_cancel(mp->m_log);
702 goto out_destroy_ail;
706 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
709 goto out_destroy_ail;
711 /* Normal transactions can now occur */
712 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
715 * Now the log has been fully initialised and we know were our
716 * space grant counters are, we can initialise the permanent ticket
717 * needed for delayed logging to work.
719 xlog_cil_init_post_recovery(mp->m_log);
724 xfs_trans_ail_destroy(mp);
726 xlog_dealloc_log(mp->m_log);
732 * Finish the recovery of the file system. This is separate from the
733 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
734 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
737 * If we finish recovery successfully, start the background log work. If we are
738 * not doing recovery, then we have a RO filesystem and we don't need to start
742 xfs_log_mount_finish(
743 struct xfs_mount *mp)
747 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
748 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
752 error = xlog_recover_finish(mp->m_log);
754 xfs_log_work_queue(mp);
760 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
764 xfs_log_mount_cancel(
765 struct xfs_mount *mp)
769 error = xlog_recover_cancel(mp->m_log);
776 * Final log writes as part of unmount.
778 * Mark the filesystem clean as unmount happens. Note that during relocation
779 * this routine needs to be executed as part of source-bag while the
780 * deallocation must not be done until source-end.
784 * Unmount record used to have a string "Unmount filesystem--" in the
785 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
786 * We just write the magic number now since that particular field isn't
787 * currently architecture converted and "Unmount" is a bit foo.
788 * As far as I know, there weren't any dependencies on the old behaviour.
792 xfs_log_unmount_write(xfs_mount_t *mp)
794 struct xlog *log = mp->m_log;
795 xlog_in_core_t *iclog;
797 xlog_in_core_t *first_iclog;
799 xlog_ticket_t *tic = NULL;
804 * Don't write out unmount record on read-only mounts.
805 * Or, if we are doing a forced umount (typically because of IO errors).
807 if (mp->m_flags & XFS_MOUNT_RDONLY)
810 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
811 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
814 first_iclog = iclog = log->l_iclog;
816 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
817 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
818 ASSERT(iclog->ic_offset == 0);
820 iclog = iclog->ic_next;
821 } while (iclog != first_iclog);
823 if (! (XLOG_FORCED_SHUTDOWN(log))) {
824 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
826 /* the data section must be 32 bit size aligned */
830 __uint32_t pad2; /* may as well make it 64 bits */
832 .magic = XLOG_UNMOUNT_TYPE,
834 struct xfs_log_iovec reg = {
836 .i_len = sizeof(magic),
837 .i_type = XLOG_REG_TYPE_UNMOUNT,
839 struct xfs_log_vec vec = {
844 /* remove inited flag, and account for space used */
846 tic->t_curr_res -= sizeof(magic);
847 error = xlog_write(log, &vec, tic, &lsn,
848 NULL, XLOG_UNMOUNT_TRANS);
850 * At this point, we're umounting anyway,
851 * so there's no point in transitioning log state
852 * to IOERROR. Just continue...
857 xfs_alert(mp, "%s: unmount record failed", __func__);
860 spin_lock(&log->l_icloglock);
861 iclog = log->l_iclog;
862 atomic_inc(&iclog->ic_refcnt);
863 xlog_state_want_sync(log, iclog);
864 spin_unlock(&log->l_icloglock);
865 error = xlog_state_release_iclog(log, iclog);
867 spin_lock(&log->l_icloglock);
868 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
869 iclog->ic_state == XLOG_STATE_DIRTY)) {
870 if (!XLOG_FORCED_SHUTDOWN(log)) {
871 xlog_wait(&iclog->ic_force_wait,
874 spin_unlock(&log->l_icloglock);
877 spin_unlock(&log->l_icloglock);
880 trace_xfs_log_umount_write(log, tic);
881 xlog_ungrant_log_space(log, tic);
882 xfs_log_ticket_put(tic);
886 * We're already in forced_shutdown mode, couldn't
887 * even attempt to write out the unmount transaction.
889 * Go through the motions of sync'ing and releasing
890 * the iclog, even though no I/O will actually happen,
891 * we need to wait for other log I/Os that may already
892 * be in progress. Do this as a separate section of
893 * code so we'll know if we ever get stuck here that
894 * we're in this odd situation of trying to unmount
895 * a file system that went into forced_shutdown as
896 * the result of an unmount..
898 spin_lock(&log->l_icloglock);
899 iclog = log->l_iclog;
900 atomic_inc(&iclog->ic_refcnt);
902 xlog_state_want_sync(log, iclog);
903 spin_unlock(&log->l_icloglock);
904 error = xlog_state_release_iclog(log, iclog);
906 spin_lock(&log->l_icloglock);
908 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
909 || iclog->ic_state == XLOG_STATE_DIRTY
910 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
912 xlog_wait(&iclog->ic_force_wait,
915 spin_unlock(&log->l_icloglock);
920 } /* xfs_log_unmount_write */
923 * Empty the log for unmount/freeze.
925 * To do this, we first need to shut down the background log work so it is not
926 * trying to cover the log as we clean up. We then need to unpin all objects in
927 * the log so we can then flush them out. Once they have completed their IO and
928 * run the callbacks removing themselves from the AIL, we can write the unmount
933 struct xfs_mount *mp)
935 cancel_delayed_work_sync(&mp->m_log->l_work);
936 xfs_log_force(mp, XFS_LOG_SYNC);
939 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
940 * will push it, xfs_wait_buftarg() will not wait for it. Further,
941 * xfs_buf_iowait() cannot be used because it was pushed with the
942 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
943 * the IO to complete.
945 xfs_ail_push_all_sync(mp->m_ail);
946 xfs_wait_buftarg(mp->m_ddev_targp);
947 xfs_buf_lock(mp->m_sb_bp);
948 xfs_buf_unlock(mp->m_sb_bp);
950 xfs_log_unmount_write(mp);
954 * Shut down and release the AIL and Log.
956 * During unmount, we need to ensure we flush all the dirty metadata objects
957 * from the AIL so that the log is empty before we write the unmount record to
958 * the log. Once this is done, we can tear down the AIL and the log.
962 struct xfs_mount *mp)
966 xfs_trans_ail_destroy(mp);
968 xfs_sysfs_del(&mp->m_log->l_kobj);
970 xlog_dealloc_log(mp->m_log);
975 struct xfs_mount *mp,
976 struct xfs_log_item *item,
978 const struct xfs_item_ops *ops)
980 item->li_mountp = mp;
981 item->li_ailp = mp->m_ail;
982 item->li_type = type;
986 INIT_LIST_HEAD(&item->li_ail);
987 INIT_LIST_HEAD(&item->li_cil);
991 * Wake up processes waiting for log space after we have moved the log tail.
995 struct xfs_mount *mp)
997 struct xlog *log = mp->m_log;
1000 if (XLOG_FORCED_SHUTDOWN(log))
1003 if (!list_empty_careful(&log->l_write_head.waiters)) {
1004 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1006 spin_lock(&log->l_write_head.lock);
1007 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1008 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1009 spin_unlock(&log->l_write_head.lock);
1012 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1013 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1015 spin_lock(&log->l_reserve_head.lock);
1016 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1017 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1018 spin_unlock(&log->l_reserve_head.lock);
1023 * Determine if we have a transaction that has gone to disk that needs to be
1024 * covered. To begin the transition to the idle state firstly the log needs to
1025 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1026 * we start attempting to cover the log.
1028 * Only if we are then in a state where covering is needed, the caller is
1029 * informed that dummy transactions are required to move the log into the idle
1032 * If there are any items in the AIl or CIL, then we do not want to attempt to
1033 * cover the log as we may be in a situation where there isn't log space
1034 * available to run a dummy transaction and this can lead to deadlocks when the
1035 * tail of the log is pinned by an item that is modified in the CIL. Hence
1036 * there's no point in running a dummy transaction at this point because we
1037 * can't start trying to idle the log until both the CIL and AIL are empty.
1040 xfs_log_need_covered(xfs_mount_t *mp)
1042 struct xlog *log = mp->m_log;
1045 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1048 if (!xlog_cil_empty(log))
1051 spin_lock(&log->l_icloglock);
1052 switch (log->l_covered_state) {
1053 case XLOG_STATE_COVER_DONE:
1054 case XLOG_STATE_COVER_DONE2:
1055 case XLOG_STATE_COVER_IDLE:
1057 case XLOG_STATE_COVER_NEED:
1058 case XLOG_STATE_COVER_NEED2:
1059 if (xfs_ail_min_lsn(log->l_ailp))
1061 if (!xlog_iclogs_empty(log))
1065 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1066 log->l_covered_state = XLOG_STATE_COVER_DONE;
1068 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1074 spin_unlock(&log->l_icloglock);
1079 * We may be holding the log iclog lock upon entering this routine.
1082 xlog_assign_tail_lsn_locked(
1083 struct xfs_mount *mp)
1085 struct xlog *log = mp->m_log;
1086 struct xfs_log_item *lip;
1089 assert_spin_locked(&mp->m_ail->xa_lock);
1092 * To make sure we always have a valid LSN for the log tail we keep
1093 * track of the last LSN which was committed in log->l_last_sync_lsn,
1094 * and use that when the AIL was empty.
1096 lip = xfs_ail_min(mp->m_ail);
1098 tail_lsn = lip->li_lsn;
1100 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1101 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1102 atomic64_set(&log->l_tail_lsn, tail_lsn);
1107 xlog_assign_tail_lsn(
1108 struct xfs_mount *mp)
1112 spin_lock(&mp->m_ail->xa_lock);
1113 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1114 spin_unlock(&mp->m_ail->xa_lock);
1120 * Return the space in the log between the tail and the head. The head
1121 * is passed in the cycle/bytes formal parms. In the special case where
1122 * the reserve head has wrapped passed the tail, this calculation is no
1123 * longer valid. In this case, just return 0 which means there is no space
1124 * in the log. This works for all places where this function is called
1125 * with the reserve head. Of course, if the write head were to ever
1126 * wrap the tail, we should blow up. Rather than catch this case here,
1127 * we depend on other ASSERTions in other parts of the code. XXXmiken
1129 * This code also handles the case where the reservation head is behind
1130 * the tail. The details of this case are described below, but the end
1131 * result is that we return the size of the log as the amount of space left.
1144 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1145 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1146 tail_bytes = BBTOB(tail_bytes);
1147 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1148 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1149 else if (tail_cycle + 1 < head_cycle)
1151 else if (tail_cycle < head_cycle) {
1152 ASSERT(tail_cycle == (head_cycle - 1));
1153 free_bytes = tail_bytes - head_bytes;
1156 * The reservation head is behind the tail.
1157 * In this case we just want to return the size of the
1158 * log as the amount of space left.
1160 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1161 xfs_alert(log->l_mp,
1162 " tail_cycle = %d, tail_bytes = %d",
1163 tail_cycle, tail_bytes);
1164 xfs_alert(log->l_mp,
1165 " GH cycle = %d, GH bytes = %d",
1166 head_cycle, head_bytes);
1168 free_bytes = log->l_logsize;
1175 * Log function which is called when an io completes.
1177 * The log manager needs its own routine, in order to control what
1178 * happens with the buffer after the write completes.
1181 xlog_iodone(xfs_buf_t *bp)
1183 struct xlog_in_core *iclog = bp->b_fspriv;
1184 struct xlog *l = iclog->ic_log;
1188 * Race to shutdown the filesystem if we see an error or the iclog is in
1189 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1190 * CRC errors into log recovery.
1192 if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR,
1193 XFS_RANDOM_IODONE_IOERR) ||
1194 iclog->ic_state & XLOG_STATE_IOABORT) {
1195 if (iclog->ic_state & XLOG_STATE_IOABORT)
1196 iclog->ic_state &= ~XLOG_STATE_IOABORT;
1198 xfs_buf_ioerror_alert(bp, __func__);
1200 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1202 * This flag will be propagated to the trans-committed
1203 * callback routines to let them know that the log-commit
1206 aborted = XFS_LI_ABORTED;
1207 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1208 aborted = XFS_LI_ABORTED;
1211 /* log I/O is always issued ASYNC */
1212 ASSERT(bp->b_flags & XBF_ASYNC);
1213 xlog_state_done_syncing(iclog, aborted);
1216 * drop the buffer lock now that we are done. Nothing references
1217 * the buffer after this, so an unmount waiting on this lock can now
1218 * tear it down safely. As such, it is unsafe to reference the buffer
1219 * (bp) after the unlock as we could race with it being freed.
1225 * Return size of each in-core log record buffer.
1227 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1229 * If the filesystem blocksize is too large, we may need to choose a
1230 * larger size since the directory code currently logs entire blocks.
1234 xlog_get_iclog_buffer_size(
1235 struct xfs_mount *mp,
1241 if (mp->m_logbufs <= 0)
1242 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1244 log->l_iclog_bufs = mp->m_logbufs;
1247 * Buffer size passed in from mount system call.
1249 if (mp->m_logbsize > 0) {
1250 size = log->l_iclog_size = mp->m_logbsize;
1251 log->l_iclog_size_log = 0;
1253 log->l_iclog_size_log++;
1257 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1258 /* # headers = size / 32k
1259 * one header holds cycles from 32k of data
1262 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1263 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1265 log->l_iclog_hsize = xhdrs << BBSHIFT;
1266 log->l_iclog_heads = xhdrs;
1268 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1269 log->l_iclog_hsize = BBSIZE;
1270 log->l_iclog_heads = 1;
1275 /* All machines use 32kB buffers by default. */
1276 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1277 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1279 /* the default log size is 16k or 32k which is one header sector */
1280 log->l_iclog_hsize = BBSIZE;
1281 log->l_iclog_heads = 1;
1284 /* are we being asked to make the sizes selected above visible? */
1285 if (mp->m_logbufs == 0)
1286 mp->m_logbufs = log->l_iclog_bufs;
1287 if (mp->m_logbsize == 0)
1288 mp->m_logbsize = log->l_iclog_size;
1289 } /* xlog_get_iclog_buffer_size */
1294 struct xfs_mount *mp)
1296 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1297 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1301 * Every sync period we need to unpin all items in the AIL and push them to
1302 * disk. If there is nothing dirty, then we might need to cover the log to
1303 * indicate that the filesystem is idle.
1307 struct work_struct *work)
1309 struct xlog *log = container_of(to_delayed_work(work),
1310 struct xlog, l_work);
1311 struct xfs_mount *mp = log->l_mp;
1313 /* dgc: errors ignored - not fatal and nowhere to report them */
1314 if (xfs_log_need_covered(mp)) {
1316 * Dump a transaction into the log that contains no real change.
1317 * This is needed to stamp the current tail LSN into the log
1318 * during the covering operation.
1320 * We cannot use an inode here for this - that will push dirty
1321 * state back up into the VFS and then periodic inode flushing
1322 * will prevent log covering from making progress. Hence we
1323 * synchronously log the superblock instead to ensure the
1324 * superblock is immediately unpinned and can be written back.
1326 xfs_sync_sb(mp, true);
1328 xfs_log_force(mp, 0);
1330 /* start pushing all the metadata that is currently dirty */
1331 xfs_ail_push_all(mp->m_ail);
1333 /* queue us up again */
1334 xfs_log_work_queue(mp);
1338 * This routine initializes some of the log structure for a given mount point.
1339 * Its primary purpose is to fill in enough, so recovery can occur. However,
1340 * some other stuff may be filled in too.
1342 STATIC struct xlog *
1344 struct xfs_mount *mp,
1345 struct xfs_buftarg *log_target,
1346 xfs_daddr_t blk_offset,
1350 xlog_rec_header_t *head;
1351 xlog_in_core_t **iclogp;
1352 xlog_in_core_t *iclog, *prev_iclog=NULL;
1355 int error = -ENOMEM;
1358 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1360 xfs_warn(mp, "Log allocation failed: No memory!");
1365 log->l_targ = log_target;
1366 log->l_logsize = BBTOB(num_bblks);
1367 log->l_logBBstart = blk_offset;
1368 log->l_logBBsize = num_bblks;
1369 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1370 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1371 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1373 log->l_prev_block = -1;
1374 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1375 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1376 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1377 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1379 xlog_grant_head_init(&log->l_reserve_head);
1380 xlog_grant_head_init(&log->l_write_head);
1382 error = -EFSCORRUPTED;
1383 if (xfs_sb_version_hassector(&mp->m_sb)) {
1384 log2_size = mp->m_sb.sb_logsectlog;
1385 if (log2_size < BBSHIFT) {
1386 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1387 log2_size, BBSHIFT);
1391 log2_size -= BBSHIFT;
1392 if (log2_size > mp->m_sectbb_log) {
1393 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1394 log2_size, mp->m_sectbb_log);
1398 /* for larger sector sizes, must have v2 or external log */
1399 if (log2_size && log->l_logBBstart > 0 &&
1400 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1402 "log sector size (0x%x) invalid for configuration.",
1407 log->l_sectBBsize = 1 << log2_size;
1409 xlog_get_iclog_buffer_size(mp, log);
1412 * Use a NULL block for the extra log buffer used during splits so that
1413 * it will trigger errors if we ever try to do IO on it without first
1414 * having set it up properly.
1417 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1418 BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1423 * The iclogbuf buffer locks are held over IO but we are not going to do
1424 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1425 * when appropriately.
1427 ASSERT(xfs_buf_islocked(bp));
1430 /* use high priority wq for log I/O completion */
1431 bp->b_ioend_wq = mp->m_log_workqueue;
1432 bp->b_iodone = xlog_iodone;
1435 spin_lock_init(&log->l_icloglock);
1436 init_waitqueue_head(&log->l_flush_wait);
1438 iclogp = &log->l_iclog;
1440 * The amount of memory to allocate for the iclog structure is
1441 * rather funky due to the way the structure is defined. It is
1442 * done this way so that we can use different sizes for machines
1443 * with different amounts of memory. See the definition of
1444 * xlog_in_core_t in xfs_log_priv.h for details.
1446 ASSERT(log->l_iclog_size >= 4096);
1447 for (i=0; i < log->l_iclog_bufs; i++) {
1448 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1450 goto out_free_iclog;
1453 iclog->ic_prev = prev_iclog;
1456 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1457 BTOBB(log->l_iclog_size),
1460 goto out_free_iclog;
1462 ASSERT(xfs_buf_islocked(bp));
1465 /* use high priority wq for log I/O completion */
1466 bp->b_ioend_wq = mp->m_log_workqueue;
1467 bp->b_iodone = xlog_iodone;
1469 iclog->ic_data = bp->b_addr;
1471 log->l_iclog_bak[i] = &iclog->ic_header;
1473 head = &iclog->ic_header;
1474 memset(head, 0, sizeof(xlog_rec_header_t));
1475 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1476 head->h_version = cpu_to_be32(
1477 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1478 head->h_size = cpu_to_be32(log->l_iclog_size);
1480 head->h_fmt = cpu_to_be32(XLOG_FMT);
1481 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1483 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1484 iclog->ic_state = XLOG_STATE_ACTIVE;
1485 iclog->ic_log = log;
1486 atomic_set(&iclog->ic_refcnt, 0);
1487 spin_lock_init(&iclog->ic_callback_lock);
1488 iclog->ic_callback_tail = &(iclog->ic_callback);
1489 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1491 init_waitqueue_head(&iclog->ic_force_wait);
1492 init_waitqueue_head(&iclog->ic_write_wait);
1494 iclogp = &iclog->ic_next;
1496 *iclogp = log->l_iclog; /* complete ring */
1497 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1499 error = xlog_cil_init(log);
1501 goto out_free_iclog;
1505 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1506 prev_iclog = iclog->ic_next;
1508 xfs_buf_free(iclog->ic_bp);
1511 spinlock_destroy(&log->l_icloglock);
1512 xfs_buf_free(log->l_xbuf);
1516 return ERR_PTR(error);
1517 } /* xlog_alloc_log */
1521 * Write out the commit record of a transaction associated with the given
1522 * ticket. Return the lsn of the commit record.
1527 struct xlog_ticket *ticket,
1528 struct xlog_in_core **iclog,
1529 xfs_lsn_t *commitlsnp)
1531 struct xfs_mount *mp = log->l_mp;
1533 struct xfs_log_iovec reg = {
1536 .i_type = XLOG_REG_TYPE_COMMIT,
1538 struct xfs_log_vec vec = {
1543 ASSERT_ALWAYS(iclog);
1544 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1547 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1552 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1553 * log space. This code pushes on the lsn which would supposedly free up
1554 * the 25% which we want to leave free. We may need to adopt a policy which
1555 * pushes on an lsn which is further along in the log once we reach the high
1556 * water mark. In this manner, we would be creating a low water mark.
1559 xlog_grant_push_ail(
1563 xfs_lsn_t threshold_lsn = 0;
1564 xfs_lsn_t last_sync_lsn;
1567 int threshold_block;
1568 int threshold_cycle;
1571 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1573 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1574 free_blocks = BTOBBT(free_bytes);
1577 * Set the threshold for the minimum number of free blocks in the
1578 * log to the maximum of what the caller needs, one quarter of the
1579 * log, and 256 blocks.
1581 free_threshold = BTOBB(need_bytes);
1582 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1583 free_threshold = MAX(free_threshold, 256);
1584 if (free_blocks >= free_threshold)
1587 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1589 threshold_block += free_threshold;
1590 if (threshold_block >= log->l_logBBsize) {
1591 threshold_block -= log->l_logBBsize;
1592 threshold_cycle += 1;
1594 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1597 * Don't pass in an lsn greater than the lsn of the last
1598 * log record known to be on disk. Use a snapshot of the last sync lsn
1599 * so that it doesn't change between the compare and the set.
1601 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1602 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1603 threshold_lsn = last_sync_lsn;
1606 * Get the transaction layer to kick the dirty buffers out to
1607 * disk asynchronously. No point in trying to do this if
1608 * the filesystem is shutting down.
1610 if (!XLOG_FORCED_SHUTDOWN(log))
1611 xfs_ail_push(log->l_ailp, threshold_lsn);
1615 * Stamp cycle number in every block
1620 struct xlog_in_core *iclog,
1624 int size = iclog->ic_offset + roundoff;
1628 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1630 dp = iclog->ic_datap;
1631 for (i = 0; i < BTOBB(size); i++) {
1632 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1634 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1635 *(__be32 *)dp = cycle_lsn;
1639 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1640 xlog_in_core_2_t *xhdr = iclog->ic_data;
1642 for ( ; i < BTOBB(size); i++) {
1643 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1644 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1645 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1646 *(__be32 *)dp = cycle_lsn;
1650 for (i = 1; i < log->l_iclog_heads; i++)
1651 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1656 * Calculate the checksum for a log buffer.
1658 * This is a little more complicated than it should be because the various
1659 * headers and the actual data are non-contiguous.
1664 struct xlog_rec_header *rhead,
1670 /* first generate the crc for the record header ... */
1671 crc = xfs_start_cksum_update((char *)rhead,
1672 sizeof(struct xlog_rec_header),
1673 offsetof(struct xlog_rec_header, h_crc));
1675 /* ... then for additional cycle data for v2 logs ... */
1676 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1677 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1681 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1682 if (size % XLOG_HEADER_CYCLE_SIZE)
1685 for (i = 1; i < xheads; i++) {
1686 crc = crc32c(crc, &xhdr[i].hic_xheader,
1687 sizeof(struct xlog_rec_ext_header));
1691 /* ... and finally for the payload */
1692 crc = crc32c(crc, dp, size);
1694 return xfs_end_cksum(crc);
1698 * The bdstrat callback function for log bufs. This gives us a central
1699 * place to trap bufs in case we get hit by a log I/O error and need to
1700 * shutdown. Actually, in practice, even when we didn't get a log error,
1701 * we transition the iclogs to IOERROR state *after* flushing all existing
1702 * iclogs to disk. This is because we don't want anymore new transactions to be
1703 * started or completed afterwards.
1705 * We lock the iclogbufs here so that we can serialise against IO completion
1706 * during unmount. We might be processing a shutdown triggered during unmount,
1707 * and that can occur asynchronously to the unmount thread, and hence we need to
1708 * ensure that completes before tearing down the iclogbufs. Hence we need to
1709 * hold the buffer lock across the log IO to acheive that.
1715 struct xlog_in_core *iclog = bp->b_fspriv;
1718 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1719 xfs_buf_ioerror(bp, -EIO);
1723 * It would seem logical to return EIO here, but we rely on
1724 * the log state machine to propagate I/O errors instead of
1725 * doing it here. Similarly, IO completion will unlock the
1726 * buffer, so we don't do it here.
1736 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1737 * fashion. Previously, we should have moved the current iclog
1738 * ptr in the log to point to the next available iclog. This allows further
1739 * write to continue while this code syncs out an iclog ready to go.
1740 * Before an in-core log can be written out, the data section must be scanned
1741 * to save away the 1st word of each BBSIZE block into the header. We replace
1742 * it with the current cycle count. Each BBSIZE block is tagged with the
1743 * cycle count because there in an implicit assumption that drives will
1744 * guarantee that entire 512 byte blocks get written at once. In other words,
1745 * we can't have part of a 512 byte block written and part not written. By
1746 * tagging each block, we will know which blocks are valid when recovering
1747 * after an unclean shutdown.
1749 * This routine is single threaded on the iclog. No other thread can be in
1750 * this routine with the same iclog. Changing contents of iclog can there-
1751 * fore be done without grabbing the state machine lock. Updating the global
1752 * log will require grabbing the lock though.
1754 * The entire log manager uses a logical block numbering scheme. Only
1755 * log_sync (and then only bwrite()) know about the fact that the log may
1756 * not start with block zero on a given device. The log block start offset
1757 * is added immediately before calling bwrite().
1763 struct xlog_in_core *iclog)
1767 uint count; /* byte count of bwrite */
1768 uint count_init; /* initial count before roundup */
1769 int roundoff; /* roundoff to BB or stripe */
1770 int split = 0; /* split write into two regions */
1772 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1775 XFS_STATS_INC(log->l_mp, xs_log_writes);
1776 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1778 /* Add for LR header */
1779 count_init = log->l_iclog_hsize + iclog->ic_offset;
1781 /* Round out the log write size */
1782 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1783 /* we have a v2 stripe unit to use */
1784 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1786 count = BBTOB(BTOBB(count_init));
1788 roundoff = count - count_init;
1789 ASSERT(roundoff >= 0);
1790 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1791 roundoff < log->l_mp->m_sb.sb_logsunit)
1793 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1794 roundoff < BBTOB(1)));
1796 /* move grant heads by roundoff in sync */
1797 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1798 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1800 /* put cycle number in every block */
1801 xlog_pack_data(log, iclog, roundoff);
1803 /* real byte length */
1804 size = iclog->ic_offset;
1807 iclog->ic_header.h_len = cpu_to_be32(size);
1810 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1812 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1814 /* Do we need to split this write into 2 parts? */
1815 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1818 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1819 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1820 iclog->ic_bwritecnt = 2;
1823 * Bump the cycle numbers at the start of each block in the
1824 * part of the iclog that ends up in the buffer that gets
1825 * written to the start of the log.
1827 * Watch out for the header magic number case, though.
1829 dptr = (char *)&iclog->ic_header + count;
1830 for (i = 0; i < split; i += BBSIZE) {
1831 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1832 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1834 *(__be32 *)dptr = cpu_to_be32(cycle);
1839 iclog->ic_bwritecnt = 1;
1842 /* calculcate the checksum */
1843 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1844 iclog->ic_datap, size);
1847 * Intentionally corrupt the log record CRC based on the error injection
1848 * frequency, if defined. This facilitates testing log recovery in the
1849 * event of torn writes. Hence, set the IOABORT state to abort the log
1850 * write on I/O completion and shutdown the fs. The subsequent mount
1851 * detects the bad CRC and attempts to recover.
1853 if (log->l_badcrc_factor &&
1854 (prandom_u32() % log->l_badcrc_factor == 0)) {
1855 iclog->ic_header.h_crc &= 0xAAAAAAAA;
1856 iclog->ic_state |= XLOG_STATE_IOABORT;
1858 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1859 be64_to_cpu(iclog->ic_header.h_lsn));
1863 bp->b_io_length = BTOBB(count);
1864 bp->b_fspriv = iclog;
1865 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1866 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1868 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1869 bp->b_flags |= XBF_FUA;
1872 * Flush the data device before flushing the log to make
1873 * sure all meta data written back from the AIL actually made
1874 * it to disk before stamping the new log tail LSN into the
1875 * log buffer. For an external log we need to issue the
1876 * flush explicitly, and unfortunately synchronously here;
1877 * for an internal log we can simply use the block layer
1878 * state machine for preflushes.
1880 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1881 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1883 bp->b_flags |= XBF_FLUSH;
1886 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1887 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1889 xlog_verify_iclog(log, iclog, count, true);
1891 /* account for log which doesn't start at block #0 */
1892 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1895 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1898 error = xlog_bdstrat(bp);
1900 xfs_buf_ioerror_alert(bp, "xlog_sync");
1904 bp = iclog->ic_log->l_xbuf;
1905 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1906 xfs_buf_associate_memory(bp,
1907 (char *)&iclog->ic_header + count, split);
1908 bp->b_fspriv = iclog;
1909 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1910 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1911 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1912 bp->b_flags |= XBF_FUA;
1914 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1915 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1917 /* account for internal log which doesn't start at block #0 */
1918 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1919 error = xlog_bdstrat(bp);
1921 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1929 * Deallocate a log structure
1935 xlog_in_core_t *iclog, *next_iclog;
1938 xlog_cil_destroy(log);
1941 * Cycle all the iclogbuf locks to make sure all log IO completion
1942 * is done before we tear down these buffers.
1944 iclog = log->l_iclog;
1945 for (i = 0; i < log->l_iclog_bufs; i++) {
1946 xfs_buf_lock(iclog->ic_bp);
1947 xfs_buf_unlock(iclog->ic_bp);
1948 iclog = iclog->ic_next;
1952 * Always need to ensure that the extra buffer does not point to memory
1953 * owned by another log buffer before we free it. Also, cycle the lock
1954 * first to ensure we've completed IO on it.
1956 xfs_buf_lock(log->l_xbuf);
1957 xfs_buf_unlock(log->l_xbuf);
1958 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1959 xfs_buf_free(log->l_xbuf);
1961 iclog = log->l_iclog;
1962 for (i = 0; i < log->l_iclog_bufs; i++) {
1963 xfs_buf_free(iclog->ic_bp);
1964 next_iclog = iclog->ic_next;
1968 spinlock_destroy(&log->l_icloglock);
1970 log->l_mp->m_log = NULL;
1972 } /* xlog_dealloc_log */
1975 * Update counters atomically now that memcpy is done.
1979 xlog_state_finish_copy(
1981 struct xlog_in_core *iclog,
1985 spin_lock(&log->l_icloglock);
1987 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1988 iclog->ic_offset += copy_bytes;
1990 spin_unlock(&log->l_icloglock);
1991 } /* xlog_state_finish_copy */
1997 * print out info relating to regions written which consume
2002 struct xfs_mount *mp,
2003 struct xlog_ticket *ticket)
2006 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2008 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2009 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2010 static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2011 REG_TYPE_STR(BFORMAT, "bformat"),
2012 REG_TYPE_STR(BCHUNK, "bchunk"),
2013 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2014 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2015 REG_TYPE_STR(IFORMAT, "iformat"),
2016 REG_TYPE_STR(ICORE, "icore"),
2017 REG_TYPE_STR(IEXT, "iext"),
2018 REG_TYPE_STR(IBROOT, "ibroot"),
2019 REG_TYPE_STR(ILOCAL, "ilocal"),
2020 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2021 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2022 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2023 REG_TYPE_STR(QFORMAT, "qformat"),
2024 REG_TYPE_STR(DQUOT, "dquot"),
2025 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2026 REG_TYPE_STR(LRHEADER, "LR header"),
2027 REG_TYPE_STR(UNMOUNT, "unmount"),
2028 REG_TYPE_STR(COMMIT, "commit"),
2029 REG_TYPE_STR(TRANSHDR, "trans header"),
2030 REG_TYPE_STR(ICREATE, "inode create")
2034 xfs_warn(mp, "xlog_write: reservation summary:");
2035 xfs_warn(mp, " unit res = %d bytes",
2036 ticket->t_unit_res);
2037 xfs_warn(mp, " current res = %d bytes",
2038 ticket->t_curr_res);
2039 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2040 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2041 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2042 ticket->t_res_num_ophdrs, ophdr_spc);
2043 xfs_warn(mp, " ophdr + reg = %u bytes",
2044 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2045 xfs_warn(mp, " num regions = %u",
2048 for (i = 0; i < ticket->t_res_num; i++) {
2049 uint r_type = ticket->t_res_arr[i].r_type;
2050 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2051 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2052 "bad-rtype" : res_type_str[r_type]),
2053 ticket->t_res_arr[i].r_len);
2056 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2057 "xlog_write: reservation ran out. Need to up reservation");
2058 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2062 * Calculate the potential space needed by the log vector. Each region gets
2063 * its own xlog_op_header_t and may need to be double word aligned.
2066 xlog_write_calc_vec_length(
2067 struct xlog_ticket *ticket,
2068 struct xfs_log_vec *log_vector)
2070 struct xfs_log_vec *lv;
2075 /* acct for start rec of xact */
2076 if (ticket->t_flags & XLOG_TIC_INITED)
2079 for (lv = log_vector; lv; lv = lv->lv_next) {
2080 /* we don't write ordered log vectors */
2081 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2084 headers += lv->lv_niovecs;
2086 for (i = 0; i < lv->lv_niovecs; i++) {
2087 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2090 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2094 ticket->t_res_num_ophdrs += headers;
2095 len += headers * sizeof(struct xlog_op_header);
2101 * If first write for transaction, insert start record We can't be trying to
2102 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2105 xlog_write_start_rec(
2106 struct xlog_op_header *ophdr,
2107 struct xlog_ticket *ticket)
2109 if (!(ticket->t_flags & XLOG_TIC_INITED))
2112 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2113 ophdr->oh_clientid = ticket->t_clientid;
2115 ophdr->oh_flags = XLOG_START_TRANS;
2118 ticket->t_flags &= ~XLOG_TIC_INITED;
2120 return sizeof(struct xlog_op_header);
2123 static xlog_op_header_t *
2124 xlog_write_setup_ophdr(
2126 struct xlog_op_header *ophdr,
2127 struct xlog_ticket *ticket,
2130 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2131 ophdr->oh_clientid = ticket->t_clientid;
2134 /* are we copying a commit or unmount record? */
2135 ophdr->oh_flags = flags;
2138 * We've seen logs corrupted with bad transaction client ids. This
2139 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2140 * and shut down the filesystem.
2142 switch (ophdr->oh_clientid) {
2143 case XFS_TRANSACTION:
2149 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2150 ophdr->oh_clientid, ticket);
2158 * Set up the parameters of the region copy into the log. This has
2159 * to handle region write split across multiple log buffers - this
2160 * state is kept external to this function so that this code can
2161 * be written in an obvious, self documenting manner.
2164 xlog_write_setup_copy(
2165 struct xlog_ticket *ticket,
2166 struct xlog_op_header *ophdr,
2167 int space_available,
2171 int *last_was_partial_copy,
2172 int *bytes_consumed)
2176 still_to_copy = space_required - *bytes_consumed;
2177 *copy_off = *bytes_consumed;
2179 if (still_to_copy <= space_available) {
2180 /* write of region completes here */
2181 *copy_len = still_to_copy;
2182 ophdr->oh_len = cpu_to_be32(*copy_len);
2183 if (*last_was_partial_copy)
2184 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2185 *last_was_partial_copy = 0;
2186 *bytes_consumed = 0;
2190 /* partial write of region, needs extra log op header reservation */
2191 *copy_len = space_available;
2192 ophdr->oh_len = cpu_to_be32(*copy_len);
2193 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2194 if (*last_was_partial_copy)
2195 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2196 *bytes_consumed += *copy_len;
2197 (*last_was_partial_copy)++;
2199 /* account for new log op header */
2200 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2201 ticket->t_res_num_ophdrs++;
2203 return sizeof(struct xlog_op_header);
2207 xlog_write_copy_finish(
2209 struct xlog_in_core *iclog,
2214 int *partial_copy_len,
2216 struct xlog_in_core **commit_iclog)
2218 if (*partial_copy) {
2220 * This iclog has already been marked WANT_SYNC by
2221 * xlog_state_get_iclog_space.
2223 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2226 return xlog_state_release_iclog(log, iclog);
2230 *partial_copy_len = 0;
2232 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2233 /* no more space in this iclog - push it. */
2234 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2238 spin_lock(&log->l_icloglock);
2239 xlog_state_want_sync(log, iclog);
2240 spin_unlock(&log->l_icloglock);
2243 return xlog_state_release_iclog(log, iclog);
2244 ASSERT(flags & XLOG_COMMIT_TRANS);
2245 *commit_iclog = iclog;
2252 * Write some region out to in-core log
2254 * This will be called when writing externally provided regions or when
2255 * writing out a commit record for a given transaction.
2257 * General algorithm:
2258 * 1. Find total length of this write. This may include adding to the
2259 * lengths passed in.
2260 * 2. Check whether we violate the tickets reservation.
2261 * 3. While writing to this iclog
2262 * A. Reserve as much space in this iclog as can get
2263 * B. If this is first write, save away start lsn
2264 * C. While writing this region:
2265 * 1. If first write of transaction, write start record
2266 * 2. Write log operation header (header per region)
2267 * 3. Find out if we can fit entire region into this iclog
2268 * 4. Potentially, verify destination memcpy ptr
2269 * 5. Memcpy (partial) region
2270 * 6. If partial copy, release iclog; otherwise, continue
2271 * copying more regions into current iclog
2272 * 4. Mark want sync bit (in simulation mode)
2273 * 5. Release iclog for potential flush to on-disk log.
2276 * 1. Panic if reservation is overrun. This should never happen since
2277 * reservation amounts are generated internal to the filesystem.
2279 * 1. Tickets are single threaded data structures.
2280 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2281 * syncing routine. When a single log_write region needs to span
2282 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2283 * on all log operation writes which don't contain the end of the
2284 * region. The XLOG_END_TRANS bit is used for the in-core log
2285 * operation which contains the end of the continued log_write region.
2286 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2287 * we don't really know exactly how much space will be used. As a result,
2288 * we don't update ic_offset until the end when we know exactly how many
2289 * bytes have been written out.
2294 struct xfs_log_vec *log_vector,
2295 struct xlog_ticket *ticket,
2296 xfs_lsn_t *start_lsn,
2297 struct xlog_in_core **commit_iclog,
2300 struct xlog_in_core *iclog = NULL;
2301 struct xfs_log_iovec *vecp;
2302 struct xfs_log_vec *lv;
2305 int partial_copy = 0;
2306 int partial_copy_len = 0;
2314 len = xlog_write_calc_vec_length(ticket, log_vector);
2317 * Region headers and bytes are already accounted for.
2318 * We only need to take into account start records and
2319 * split regions in this function.
2321 if (ticket->t_flags & XLOG_TIC_INITED)
2322 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2325 * Commit record headers need to be accounted for. These
2326 * come in as separate writes so are easy to detect.
2328 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2329 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2331 if (ticket->t_curr_res < 0)
2332 xlog_print_tic_res(log->l_mp, ticket);
2336 vecp = lv->lv_iovecp;
2337 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2341 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2342 &contwr, &log_offset);
2346 ASSERT(log_offset <= iclog->ic_size - 1);
2347 ptr = iclog->ic_datap + log_offset;
2349 /* start_lsn is the first lsn written to. That's all we need. */
2351 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2354 * This loop writes out as many regions as can fit in the amount
2355 * of space which was allocated by xlog_state_get_iclog_space().
2357 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2358 struct xfs_log_iovec *reg;
2359 struct xlog_op_header *ophdr;
2363 bool ordered = false;
2365 /* ordered log vectors have no regions to write */
2366 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2367 ASSERT(lv->lv_niovecs == 0);
2373 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2374 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2376 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2377 if (start_rec_copy) {
2379 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2383 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2387 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2388 sizeof(struct xlog_op_header));
2390 len += xlog_write_setup_copy(ticket, ophdr,
2391 iclog->ic_size-log_offset,
2393 ©_off, ©_len,
2396 xlog_verify_dest_ptr(log, ptr);
2401 * Unmount records just log an opheader, so can have
2402 * empty payloads with no data region to copy. Hence we
2403 * only copy the payload if the vector says it has data
2406 ASSERT(copy_len >= 0);
2408 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2409 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2412 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2414 data_cnt += contwr ? copy_len : 0;
2416 error = xlog_write_copy_finish(log, iclog, flags,
2417 &record_cnt, &data_cnt,
2426 * if we had a partial copy, we need to get more iclog
2427 * space but we don't want to increment the region
2428 * index because there is still more is this region to
2431 * If we completed writing this region, and we flushed
2432 * the iclog (indicated by resetting of the record
2433 * count), then we also need to get more log space. If
2434 * this was the last record, though, we are done and
2440 if (++index == lv->lv_niovecs) {
2445 vecp = lv->lv_iovecp;
2447 if (record_cnt == 0 && ordered == false) {
2457 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2459 return xlog_state_release_iclog(log, iclog);
2461 ASSERT(flags & XLOG_COMMIT_TRANS);
2462 *commit_iclog = iclog;
2467 /*****************************************************************************
2469 * State Machine functions
2471 *****************************************************************************
2474 /* Clean iclogs starting from the head. This ordering must be
2475 * maintained, so an iclog doesn't become ACTIVE beyond one that
2476 * is SYNCING. This is also required to maintain the notion that we use
2477 * a ordered wait queue to hold off would be writers to the log when every
2478 * iclog is trying to sync to disk.
2480 * State Change: DIRTY -> ACTIVE
2483 xlog_state_clean_log(
2486 xlog_in_core_t *iclog;
2489 iclog = log->l_iclog;
2491 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2492 iclog->ic_state = XLOG_STATE_ACTIVE;
2493 iclog->ic_offset = 0;
2494 ASSERT(iclog->ic_callback == NULL);
2496 * If the number of ops in this iclog indicate it just
2497 * contains the dummy transaction, we can
2498 * change state into IDLE (the second time around).
2499 * Otherwise we should change the state into
2501 * We don't need to cover the dummy.
2504 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2509 * We have two dirty iclogs so start over
2510 * This could also be num of ops indicates
2511 * this is not the dummy going out.
2515 iclog->ic_header.h_num_logops = 0;
2516 memset(iclog->ic_header.h_cycle_data, 0,
2517 sizeof(iclog->ic_header.h_cycle_data));
2518 iclog->ic_header.h_lsn = 0;
2519 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2522 break; /* stop cleaning */
2523 iclog = iclog->ic_next;
2524 } while (iclog != log->l_iclog);
2526 /* log is locked when we are called */
2528 * Change state for the dummy log recording.
2529 * We usually go to NEED. But we go to NEED2 if the changed indicates
2530 * we are done writing the dummy record.
2531 * If we are done with the second dummy recored (DONE2), then
2535 switch (log->l_covered_state) {
2536 case XLOG_STATE_COVER_IDLE:
2537 case XLOG_STATE_COVER_NEED:
2538 case XLOG_STATE_COVER_NEED2:
2539 log->l_covered_state = XLOG_STATE_COVER_NEED;
2542 case XLOG_STATE_COVER_DONE:
2544 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2546 log->l_covered_state = XLOG_STATE_COVER_NEED;
2549 case XLOG_STATE_COVER_DONE2:
2551 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2553 log->l_covered_state = XLOG_STATE_COVER_NEED;
2560 } /* xlog_state_clean_log */
2563 xlog_get_lowest_lsn(
2566 xlog_in_core_t *lsn_log;
2567 xfs_lsn_t lowest_lsn, lsn;
2569 lsn_log = log->l_iclog;
2572 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2573 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2574 if ((lsn && !lowest_lsn) ||
2575 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2579 lsn_log = lsn_log->ic_next;
2580 } while (lsn_log != log->l_iclog);
2586 xlog_state_do_callback(
2589 struct xlog_in_core *ciclog)
2591 xlog_in_core_t *iclog;
2592 xlog_in_core_t *first_iclog; /* used to know when we've
2593 * processed all iclogs once */
2594 xfs_log_callback_t *cb, *cb_next;
2596 xfs_lsn_t lowest_lsn;
2597 int ioerrors; /* counter: iclogs with errors */
2598 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2599 int funcdidcallbacks; /* flag: function did callbacks */
2600 int repeats; /* for issuing console warnings if
2601 * looping too many times */
2604 spin_lock(&log->l_icloglock);
2605 first_iclog = iclog = log->l_iclog;
2607 funcdidcallbacks = 0;
2612 * Scan all iclogs starting with the one pointed to by the
2613 * log. Reset this starting point each time the log is
2614 * unlocked (during callbacks).
2616 * Keep looping through iclogs until one full pass is made
2617 * without running any callbacks.
2619 first_iclog = log->l_iclog;
2620 iclog = log->l_iclog;
2621 loopdidcallbacks = 0;
2626 /* skip all iclogs in the ACTIVE & DIRTY states */
2627 if (iclog->ic_state &
2628 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2629 iclog = iclog->ic_next;
2634 * Between marking a filesystem SHUTDOWN and stopping
2635 * the log, we do flush all iclogs to disk (if there
2636 * wasn't a log I/O error). So, we do want things to
2637 * go smoothly in case of just a SHUTDOWN w/o a
2640 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2642 * Can only perform callbacks in order. Since
2643 * this iclog is not in the DONE_SYNC/
2644 * DO_CALLBACK state, we skip the rest and
2645 * just try to clean up. If we set our iclog
2646 * to DO_CALLBACK, we will not process it when
2647 * we retry since a previous iclog is in the
2648 * CALLBACK and the state cannot change since
2649 * we are holding the l_icloglock.
2651 if (!(iclog->ic_state &
2652 (XLOG_STATE_DONE_SYNC |
2653 XLOG_STATE_DO_CALLBACK))) {
2654 if (ciclog && (ciclog->ic_state ==
2655 XLOG_STATE_DONE_SYNC)) {
2656 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2661 * We now have an iclog that is in either the
2662 * DO_CALLBACK or DONE_SYNC states. The other
2663 * states (WANT_SYNC, SYNCING, or CALLBACK were
2664 * caught by the above if and are going to
2665 * clean (i.e. we aren't doing their callbacks)
2670 * We will do one more check here to see if we
2671 * have chased our tail around.
2674 lowest_lsn = xlog_get_lowest_lsn(log);
2676 XFS_LSN_CMP(lowest_lsn,
2677 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2678 iclog = iclog->ic_next;
2679 continue; /* Leave this iclog for
2683 iclog->ic_state = XLOG_STATE_CALLBACK;
2687 * Completion of a iclog IO does not imply that
2688 * a transaction has completed, as transactions
2689 * can be large enough to span many iclogs. We
2690 * cannot change the tail of the log half way
2691 * through a transaction as this may be the only
2692 * transaction in the log and moving th etail to
2693 * point to the middle of it will prevent
2694 * recovery from finding the start of the
2695 * transaction. Hence we should only update the
2696 * last_sync_lsn if this iclog contains
2697 * transaction completion callbacks on it.
2699 * We have to do this before we drop the
2700 * icloglock to ensure we are the only one that
2703 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2704 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2705 if (iclog->ic_callback)
2706 atomic64_set(&log->l_last_sync_lsn,
2707 be64_to_cpu(iclog->ic_header.h_lsn));
2712 spin_unlock(&log->l_icloglock);
2715 * Keep processing entries in the callback list until
2716 * we come around and it is empty. We need to
2717 * atomically see that the list is empty and change the
2718 * state to DIRTY so that we don't miss any more
2719 * callbacks being added.
2721 spin_lock(&iclog->ic_callback_lock);
2722 cb = iclog->ic_callback;
2724 iclog->ic_callback_tail = &(iclog->ic_callback);
2725 iclog->ic_callback = NULL;
2726 spin_unlock(&iclog->ic_callback_lock);
2728 /* perform callbacks in the order given */
2729 for (; cb; cb = cb_next) {
2730 cb_next = cb->cb_next;
2731 cb->cb_func(cb->cb_arg, aborted);
2733 spin_lock(&iclog->ic_callback_lock);
2734 cb = iclog->ic_callback;
2740 spin_lock(&log->l_icloglock);
2741 ASSERT(iclog->ic_callback == NULL);
2742 spin_unlock(&iclog->ic_callback_lock);
2743 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2744 iclog->ic_state = XLOG_STATE_DIRTY;
2747 * Transition from DIRTY to ACTIVE if applicable.
2748 * NOP if STATE_IOERROR.
2750 xlog_state_clean_log(log);
2752 /* wake up threads waiting in xfs_log_force() */
2753 wake_up_all(&iclog->ic_force_wait);
2755 iclog = iclog->ic_next;
2756 } while (first_iclog != iclog);
2758 if (repeats > 5000) {
2759 flushcnt += repeats;
2762 "%s: possible infinite loop (%d iterations)",
2763 __func__, flushcnt);
2765 } while (!ioerrors && loopdidcallbacks);
2769 * Make one last gasp attempt to see if iclogs are being left in limbo.
2770 * If the above loop finds an iclog earlier than the current iclog and
2771 * in one of the syncing states, the current iclog is put into
2772 * DO_CALLBACK and the callbacks are deferred to the completion of the
2773 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2774 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2777 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2778 * for ic_state == SYNCING.
2780 if (funcdidcallbacks) {
2781 first_iclog = iclog = log->l_iclog;
2783 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2785 * Terminate the loop if iclogs are found in states
2786 * which will cause other threads to clean up iclogs.
2788 * SYNCING - i/o completion will go through logs
2789 * DONE_SYNC - interrupt thread should be waiting for
2791 * IOERROR - give up hope all ye who enter here
2793 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2794 iclog->ic_state & XLOG_STATE_SYNCING ||
2795 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2796 iclog->ic_state == XLOG_STATE_IOERROR )
2798 iclog = iclog->ic_next;
2799 } while (first_iclog != iclog);
2803 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2805 spin_unlock(&log->l_icloglock);
2808 wake_up_all(&log->l_flush_wait);
2813 * Finish transitioning this iclog to the dirty state.
2815 * Make sure that we completely execute this routine only when this is
2816 * the last call to the iclog. There is a good chance that iclog flushes,
2817 * when we reach the end of the physical log, get turned into 2 separate
2818 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2819 * routine. By using the reference count bwritecnt, we guarantee that only
2820 * the second completion goes through.
2822 * Callbacks could take time, so they are done outside the scope of the
2823 * global state machine log lock.
2826 xlog_state_done_syncing(
2827 xlog_in_core_t *iclog,
2830 struct xlog *log = iclog->ic_log;
2832 spin_lock(&log->l_icloglock);
2834 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2835 iclog->ic_state == XLOG_STATE_IOERROR);
2836 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2837 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2841 * If we got an error, either on the first buffer, or in the case of
2842 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2843 * and none should ever be attempted to be written to disk
2846 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2847 if (--iclog->ic_bwritecnt == 1) {
2848 spin_unlock(&log->l_icloglock);
2851 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2855 * Someone could be sleeping prior to writing out the next
2856 * iclog buffer, we wake them all, one will get to do the
2857 * I/O, the others get to wait for the result.
2859 wake_up_all(&iclog->ic_write_wait);
2860 spin_unlock(&log->l_icloglock);
2861 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2862 } /* xlog_state_done_syncing */
2866 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2867 * sleep. We wait on the flush queue on the head iclog as that should be
2868 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2869 * we will wait here and all new writes will sleep until a sync completes.
2871 * The in-core logs are used in a circular fashion. They are not used
2872 * out-of-order even when an iclog past the head is free.
2875 * * log_offset where xlog_write() can start writing into the in-core
2877 * * in-core log pointer to which xlog_write() should write.
2878 * * boolean indicating this is a continued write to an in-core log.
2879 * If this is the last write, then the in-core log's offset field
2880 * needs to be incremented, depending on the amount of data which
2884 xlog_state_get_iclog_space(
2887 struct xlog_in_core **iclogp,
2888 struct xlog_ticket *ticket,
2889 int *continued_write,
2893 xlog_rec_header_t *head;
2894 xlog_in_core_t *iclog;
2898 spin_lock(&log->l_icloglock);
2899 if (XLOG_FORCED_SHUTDOWN(log)) {
2900 spin_unlock(&log->l_icloglock);
2904 iclog = log->l_iclog;
2905 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2906 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2908 /* Wait for log writes to have flushed */
2909 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2913 head = &iclog->ic_header;
2915 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2916 log_offset = iclog->ic_offset;
2918 /* On the 1st write to an iclog, figure out lsn. This works
2919 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2920 * committing to. If the offset is set, that's how many blocks
2923 if (log_offset == 0) {
2924 ticket->t_curr_res -= log->l_iclog_hsize;
2925 xlog_tic_add_region(ticket,
2927 XLOG_REG_TYPE_LRHEADER);
2928 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2929 head->h_lsn = cpu_to_be64(
2930 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2931 ASSERT(log->l_curr_block >= 0);
2934 /* If there is enough room to write everything, then do it. Otherwise,
2935 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2936 * bit is on, so this will get flushed out. Don't update ic_offset
2937 * until you know exactly how many bytes get copied. Therefore, wait
2938 * until later to update ic_offset.
2940 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2941 * can fit into remaining data section.
2943 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2944 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2947 * If I'm the only one writing to this iclog, sync it to disk.
2948 * We need to do an atomic compare and decrement here to avoid
2949 * racing with concurrent atomic_dec_and_lock() calls in
2950 * xlog_state_release_iclog() when there is more than one
2951 * reference to the iclog.
2953 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2954 /* we are the only one */
2955 spin_unlock(&log->l_icloglock);
2956 error = xlog_state_release_iclog(log, iclog);
2960 spin_unlock(&log->l_icloglock);
2965 /* Do we have enough room to write the full amount in the remainder
2966 * of this iclog? Or must we continue a write on the next iclog and
2967 * mark this iclog as completely taken? In the case where we switch
2968 * iclogs (to mark it taken), this particular iclog will release/sync
2969 * to disk in xlog_write().
2971 if (len <= iclog->ic_size - iclog->ic_offset) {
2972 *continued_write = 0;
2973 iclog->ic_offset += len;
2975 *continued_write = 1;
2976 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2980 ASSERT(iclog->ic_offset <= iclog->ic_size);
2981 spin_unlock(&log->l_icloglock);
2983 *logoffsetp = log_offset;
2985 } /* xlog_state_get_iclog_space */
2987 /* The first cnt-1 times through here we don't need to
2988 * move the grant write head because the permanent
2989 * reservation has reserved cnt times the unit amount.
2990 * Release part of current permanent unit reservation and
2991 * reset current reservation to be one units worth. Also
2992 * move grant reservation head forward.
2995 xlog_regrant_reserve_log_space(
2997 struct xlog_ticket *ticket)
2999 trace_xfs_log_regrant_reserve_enter(log, ticket);
3001 if (ticket->t_cnt > 0)
3004 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3005 ticket->t_curr_res);
3006 xlog_grant_sub_space(log, &log->l_write_head.grant,
3007 ticket->t_curr_res);
3008 ticket->t_curr_res = ticket->t_unit_res;
3009 xlog_tic_reset_res(ticket);
3011 trace_xfs_log_regrant_reserve_sub(log, ticket);
3013 /* just return if we still have some of the pre-reserved space */
3014 if (ticket->t_cnt > 0)
3017 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3018 ticket->t_unit_res);
3020 trace_xfs_log_regrant_reserve_exit(log, ticket);
3022 ticket->t_curr_res = ticket->t_unit_res;
3023 xlog_tic_reset_res(ticket);
3024 } /* xlog_regrant_reserve_log_space */
3028 * Give back the space left from a reservation.
3030 * All the information we need to make a correct determination of space left
3031 * is present. For non-permanent reservations, things are quite easy. The
3032 * count should have been decremented to zero. We only need to deal with the
3033 * space remaining in the current reservation part of the ticket. If the
3034 * ticket contains a permanent reservation, there may be left over space which
3035 * needs to be released. A count of N means that N-1 refills of the current
3036 * reservation can be done before we need to ask for more space. The first
3037 * one goes to fill up the first current reservation. Once we run out of
3038 * space, the count will stay at zero and the only space remaining will be
3039 * in the current reservation field.
3042 xlog_ungrant_log_space(
3044 struct xlog_ticket *ticket)
3048 if (ticket->t_cnt > 0)
3051 trace_xfs_log_ungrant_enter(log, ticket);
3052 trace_xfs_log_ungrant_sub(log, ticket);
3055 * If this is a permanent reservation ticket, we may be able to free
3056 * up more space based on the remaining count.
3058 bytes = ticket->t_curr_res;
3059 if (ticket->t_cnt > 0) {
3060 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3061 bytes += ticket->t_unit_res*ticket->t_cnt;
3064 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3065 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3067 trace_xfs_log_ungrant_exit(log, ticket);
3069 xfs_log_space_wake(log->l_mp);
3073 * Flush iclog to disk if this is the last reference to the given iclog and
3074 * the WANT_SYNC bit is set.
3076 * When this function is entered, the iclog is not necessarily in the
3077 * WANT_SYNC state. It may be sitting around waiting to get filled.
3082 xlog_state_release_iclog(
3084 struct xlog_in_core *iclog)
3086 int sync = 0; /* do we sync? */
3088 if (iclog->ic_state & XLOG_STATE_IOERROR)
3091 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3092 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3095 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3096 spin_unlock(&log->l_icloglock);
3099 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3100 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3102 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3103 /* update tail before writing to iclog */
3104 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3106 iclog->ic_state = XLOG_STATE_SYNCING;
3107 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3108 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3109 /* cycle incremented when incrementing curr_block */
3111 spin_unlock(&log->l_icloglock);
3114 * We let the log lock go, so it's possible that we hit a log I/O
3115 * error or some other SHUTDOWN condition that marks the iclog
3116 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3117 * this iclog has consistent data, so we ignore IOERROR
3118 * flags after this point.
3121 return xlog_sync(log, iclog);
3123 } /* xlog_state_release_iclog */
3127 * This routine will mark the current iclog in the ring as WANT_SYNC
3128 * and move the current iclog pointer to the next iclog in the ring.
3129 * When this routine is called from xlog_state_get_iclog_space(), the
3130 * exact size of the iclog has not yet been determined. All we know is
3131 * that every data block. We have run out of space in this log record.
3134 xlog_state_switch_iclogs(
3136 struct xlog_in_core *iclog,
3139 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3141 eventual_size = iclog->ic_offset;
3142 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3143 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3144 log->l_prev_block = log->l_curr_block;
3145 log->l_prev_cycle = log->l_curr_cycle;
3147 /* roll log?: ic_offset changed later */
3148 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3150 /* Round up to next log-sunit */
3151 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3152 log->l_mp->m_sb.sb_logsunit > 1) {
3153 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3154 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3157 if (log->l_curr_block >= log->l_logBBsize) {
3159 * Rewind the current block before the cycle is bumped to make
3160 * sure that the combined LSN never transiently moves forward
3161 * when the log wraps to the next cycle. This is to support the
3162 * unlocked sample of these fields from xlog_valid_lsn(). Most
3163 * other cases should acquire l_icloglock.
3165 log->l_curr_block -= log->l_logBBsize;
3166 ASSERT(log->l_curr_block >= 0);
3168 log->l_curr_cycle++;
3169 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3170 log->l_curr_cycle++;
3172 ASSERT(iclog == log->l_iclog);
3173 log->l_iclog = iclog->ic_next;
3174 } /* xlog_state_switch_iclogs */
3177 * Write out all data in the in-core log as of this exact moment in time.
3179 * Data may be written to the in-core log during this call. However,
3180 * we don't guarantee this data will be written out. A change from past
3181 * implementation means this routine will *not* write out zero length LRs.
3183 * Basically, we try and perform an intelligent scan of the in-core logs.
3184 * If we determine there is no flushable data, we just return. There is no
3185 * flushable data if:
3187 * 1. the current iclog is active and has no data; the previous iclog
3188 * is in the active or dirty state.
3189 * 2. the current iclog is drity, and the previous iclog is in the
3190 * active or dirty state.
3194 * 1. the current iclog is not in the active nor dirty state.
3195 * 2. the current iclog dirty, and the previous iclog is not in the
3196 * active nor dirty state.
3197 * 3. the current iclog is active, and there is another thread writing
3198 * to this particular iclog.
3199 * 4. a) the current iclog is active and has no other writers
3200 * b) when we return from flushing out this iclog, it is still
3201 * not in the active nor dirty state.
3205 struct xfs_mount *mp,
3209 struct xlog *log = mp->m_log;
3210 struct xlog_in_core *iclog;
3213 XFS_STATS_INC(mp, xs_log_force);
3215 xlog_cil_force(log);
3217 spin_lock(&log->l_icloglock);
3219 iclog = log->l_iclog;
3220 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3221 spin_unlock(&log->l_icloglock);
3225 /* If the head iclog is not active nor dirty, we just attach
3226 * ourselves to the head and go to sleep.
3228 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3229 iclog->ic_state == XLOG_STATE_DIRTY) {
3231 * If the head is dirty or (active and empty), then
3232 * we need to look at the previous iclog. If the previous
3233 * iclog is active or dirty we are done. There is nothing
3234 * to sync out. Otherwise, we attach ourselves to the
3235 * previous iclog and go to sleep.
3237 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3238 (atomic_read(&iclog->ic_refcnt) == 0
3239 && iclog->ic_offset == 0)) {
3240 iclog = iclog->ic_prev;
3241 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3242 iclog->ic_state == XLOG_STATE_DIRTY)
3247 if (atomic_read(&iclog->ic_refcnt) == 0) {
3248 /* We are the only one with access to this
3249 * iclog. Flush it out now. There should
3250 * be a roundoff of zero to show that someone
3251 * has already taken care of the roundoff from
3252 * the previous sync.
3254 atomic_inc(&iclog->ic_refcnt);
3255 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3256 xlog_state_switch_iclogs(log, iclog, 0);
3257 spin_unlock(&log->l_icloglock);
3259 if (xlog_state_release_iclog(log, iclog))
3264 spin_lock(&log->l_icloglock);
3265 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3266 iclog->ic_state != XLOG_STATE_DIRTY)
3271 /* Someone else is writing to this iclog.
3272 * Use its call to flush out the data. However,
3273 * the other thread may not force out this LR,
3274 * so we mark it WANT_SYNC.
3276 xlog_state_switch_iclogs(log, iclog, 0);
3282 /* By the time we come around again, the iclog could've been filled
3283 * which would give it another lsn. If we have a new lsn, just
3284 * return because the relevant data has been flushed.
3287 if (flags & XFS_LOG_SYNC) {
3289 * We must check if we're shutting down here, before
3290 * we wait, while we're holding the l_icloglock.
3291 * Then we check again after waking up, in case our
3292 * sleep was disturbed by a bad news.
3294 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3295 spin_unlock(&log->l_icloglock);
3298 XFS_STATS_INC(mp, xs_log_force_sleep);
3299 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3301 * No need to grab the log lock here since we're
3302 * only deciding whether or not to return EIO
3303 * and the memory read should be atomic.
3305 if (iclog->ic_state & XLOG_STATE_IOERROR)
3312 spin_unlock(&log->l_icloglock);
3318 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3319 * about errors or whether the log was flushed or not. This is the normal
3320 * interface to use when trying to unpin items or move the log forward.
3329 trace_xfs_log_force(mp, 0, _RET_IP_);
3330 error = _xfs_log_force(mp, flags, NULL);
3332 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3336 * Force the in-core log to disk for a specific LSN.
3338 * Find in-core log with lsn.
3339 * If it is in the DIRTY state, just return.
3340 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3341 * state and go to sleep or return.
3342 * If it is in any other state, go to sleep or return.
3344 * Synchronous forces are implemented with a signal variable. All callers
3345 * to force a given lsn to disk will wait on a the sv attached to the
3346 * specific in-core log. When given in-core log finally completes its
3347 * write to disk, that thread will wake up all threads waiting on the
3352 struct xfs_mount *mp,
3357 struct xlog *log = mp->m_log;
3358 struct xlog_in_core *iclog;
3359 int already_slept = 0;
3363 XFS_STATS_INC(mp, xs_log_force);
3365 lsn = xlog_cil_force_lsn(log, lsn);
3366 if (lsn == NULLCOMMITLSN)
3370 spin_lock(&log->l_icloglock);
3371 iclog = log->l_iclog;
3372 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3373 spin_unlock(&log->l_icloglock);
3378 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3379 iclog = iclog->ic_next;
3383 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3384 spin_unlock(&log->l_icloglock);
3388 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3390 * We sleep here if we haven't already slept (e.g.
3391 * this is the first time we've looked at the correct
3392 * iclog buf) and the buffer before us is going to
3393 * be sync'ed. The reason for this is that if we
3394 * are doing sync transactions here, by waiting for
3395 * the previous I/O to complete, we can allow a few
3396 * more transactions into this iclog before we close
3399 * Otherwise, we mark the buffer WANT_SYNC, and bump
3400 * up the refcnt so we can release the log (which
3401 * drops the ref count). The state switch keeps new
3402 * transaction commits from using this buffer. When
3403 * the current commits finish writing into the buffer,
3404 * the refcount will drop to zero and the buffer will
3407 if (!already_slept &&
3408 (iclog->ic_prev->ic_state &
3409 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3410 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3412 XFS_STATS_INC(mp, xs_log_force_sleep);
3414 xlog_wait(&iclog->ic_prev->ic_write_wait,
3421 atomic_inc(&iclog->ic_refcnt);
3422 xlog_state_switch_iclogs(log, iclog, 0);
3423 spin_unlock(&log->l_icloglock);
3424 if (xlog_state_release_iclog(log, iclog))
3428 spin_lock(&log->l_icloglock);
3431 if ((flags & XFS_LOG_SYNC) && /* sleep */
3433 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3435 * Don't wait on completion if we know that we've
3436 * gotten a log write error.
3438 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3439 spin_unlock(&log->l_icloglock);
3442 XFS_STATS_INC(mp, xs_log_force_sleep);
3443 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3445 * No need to grab the log lock here since we're
3446 * only deciding whether or not to return EIO
3447 * and the memory read should be atomic.
3449 if (iclog->ic_state & XLOG_STATE_IOERROR)
3454 } else { /* just return */
3455 spin_unlock(&log->l_icloglock);
3459 } while (iclog != log->l_iclog);
3461 spin_unlock(&log->l_icloglock);
3466 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3467 * about errors or whether the log was flushed or not. This is the normal
3468 * interface to use when trying to unpin items or move the log forward.
3478 trace_xfs_log_force(mp, lsn, _RET_IP_);
3479 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3481 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3485 * Called when we want to mark the current iclog as being ready to sync to
3489 xlog_state_want_sync(
3491 struct xlog_in_core *iclog)
3493 assert_spin_locked(&log->l_icloglock);
3495 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3496 xlog_state_switch_iclogs(log, iclog, 0);
3498 ASSERT(iclog->ic_state &
3499 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3504 /*****************************************************************************
3508 *****************************************************************************
3512 * Free a used ticket when its refcount falls to zero.
3516 xlog_ticket_t *ticket)
3518 ASSERT(atomic_read(&ticket->t_ref) > 0);
3519 if (atomic_dec_and_test(&ticket->t_ref))
3520 kmem_zone_free(xfs_log_ticket_zone, ticket);
3525 xlog_ticket_t *ticket)
3527 ASSERT(atomic_read(&ticket->t_ref) > 0);
3528 atomic_inc(&ticket->t_ref);
3533 * Figure out the total log space unit (in bytes) that would be
3534 * required for a log ticket.
3537 xfs_log_calc_unit_res(
3538 struct xfs_mount *mp,
3541 struct xlog *log = mp->m_log;
3546 * Permanent reservations have up to 'cnt'-1 active log operations
3547 * in the log. A unit in this case is the amount of space for one
3548 * of these log operations. Normal reservations have a cnt of 1
3549 * and their unit amount is the total amount of space required.
3551 * The following lines of code account for non-transaction data
3552 * which occupy space in the on-disk log.
3554 * Normal form of a transaction is:
3555 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3556 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3558 * We need to account for all the leadup data and trailer data
3559 * around the transaction data.
3560 * And then we need to account for the worst case in terms of using
3562 * The worst case will happen if:
3563 * - the placement of the transaction happens to be such that the
3564 * roundoff is at its maximum
3565 * - the transaction data is synced before the commit record is synced
3566 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3567 * Therefore the commit record is in its own Log Record.
3568 * This can happen as the commit record is called with its
3569 * own region to xlog_write().
3570 * This then means that in the worst case, roundoff can happen for
3571 * the commit-rec as well.
3572 * The commit-rec is smaller than padding in this scenario and so it is
3573 * not added separately.
3576 /* for trans header */
3577 unit_bytes += sizeof(xlog_op_header_t);
3578 unit_bytes += sizeof(xfs_trans_header_t);
3581 unit_bytes += sizeof(xlog_op_header_t);
3584 * for LR headers - the space for data in an iclog is the size minus
3585 * the space used for the headers. If we use the iclog size, then we
3586 * undercalculate the number of headers required.
3588 * Furthermore - the addition of op headers for split-recs might
3589 * increase the space required enough to require more log and op
3590 * headers, so take that into account too.
3592 * IMPORTANT: This reservation makes the assumption that if this
3593 * transaction is the first in an iclog and hence has the LR headers
3594 * accounted to it, then the remaining space in the iclog is
3595 * exclusively for this transaction. i.e. if the transaction is larger
3596 * than the iclog, it will be the only thing in that iclog.
3597 * Fundamentally, this means we must pass the entire log vector to
3598 * xlog_write to guarantee this.
3600 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3601 num_headers = howmany(unit_bytes, iclog_space);
3603 /* for split-recs - ophdrs added when data split over LRs */
3604 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3606 /* add extra header reservations if we overrun */
3607 while (!num_headers ||
3608 howmany(unit_bytes, iclog_space) > num_headers) {
3609 unit_bytes += sizeof(xlog_op_header_t);
3612 unit_bytes += log->l_iclog_hsize * num_headers;
3614 /* for commit-rec LR header - note: padding will subsume the ophdr */
3615 unit_bytes += log->l_iclog_hsize;
3617 /* for roundoff padding for transaction data and one for commit record */
3618 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3619 /* log su roundoff */
3620 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3623 unit_bytes += 2 * BBSIZE;
3630 * Allocate and initialise a new log ticket.
3632 struct xlog_ticket *
3639 xfs_km_flags_t alloc_flags)
3641 struct xlog_ticket *tic;
3644 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3648 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3650 atomic_set(&tic->t_ref, 1);
3651 tic->t_task = current;
3652 INIT_LIST_HEAD(&tic->t_queue);
3653 tic->t_unit_res = unit_res;
3654 tic->t_curr_res = unit_res;
3657 tic->t_tid = prandom_u32();
3658 tic->t_clientid = client;
3659 tic->t_flags = XLOG_TIC_INITED;
3661 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3663 xlog_tic_reset_res(tic);
3669 /******************************************************************************
3671 * Log debug routines
3673 ******************************************************************************
3677 * Make sure that the destination ptr is within the valid data region of
3678 * one of the iclogs. This uses backup pointers stored in a different
3679 * part of the log in case we trash the log structure.
3682 xlog_verify_dest_ptr(
3689 for (i = 0; i < log->l_iclog_bufs; i++) {
3690 if (ptr >= log->l_iclog_bak[i] &&
3691 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3696 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3700 * Check to make sure the grant write head didn't just over lap the tail. If
3701 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3702 * the cycles differ by exactly one and check the byte count.
3704 * This check is run unlocked, so can give false positives. Rather than assert
3705 * on failures, use a warn-once flag and a panic tag to allow the admin to
3706 * determine if they want to panic the machine when such an error occurs. For
3707 * debug kernels this will have the same effect as using an assert but, unlinke
3708 * an assert, it can be turned off at runtime.
3711 xlog_verify_grant_tail(
3714 int tail_cycle, tail_blocks;
3717 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3718 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3719 if (tail_cycle != cycle) {
3720 if (cycle - 1 != tail_cycle &&
3721 !(log->l_flags & XLOG_TAIL_WARN)) {
3722 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3723 "%s: cycle - 1 != tail_cycle", __func__);
3724 log->l_flags |= XLOG_TAIL_WARN;
3727 if (space > BBTOB(tail_blocks) &&
3728 !(log->l_flags & XLOG_TAIL_WARN)) {
3729 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3730 "%s: space > BBTOB(tail_blocks)", __func__);
3731 log->l_flags |= XLOG_TAIL_WARN;
3736 /* check if it will fit */
3738 xlog_verify_tail_lsn(
3740 struct xlog_in_core *iclog,
3745 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3747 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3748 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3749 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3751 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3753 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3754 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3756 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3757 if (blocks < BTOBB(iclog->ic_offset) + 1)
3758 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3760 } /* xlog_verify_tail_lsn */
3763 * Perform a number of checks on the iclog before writing to disk.
3765 * 1. Make sure the iclogs are still circular
3766 * 2. Make sure we have a good magic number
3767 * 3. Make sure we don't have magic numbers in the data
3768 * 4. Check fields of each log operation header for:
3769 * A. Valid client identifier
3770 * B. tid ptr value falls in valid ptr space (user space code)
3771 * C. Length in log record header is correct according to the
3772 * individual operation headers within record.
3773 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3774 * log, check the preceding blocks of the physical log to make sure all
3775 * the cycle numbers agree with the current cycle number.
3780 struct xlog_in_core *iclog,
3784 xlog_op_header_t *ophead;
3785 xlog_in_core_t *icptr;
3786 xlog_in_core_2_t *xhdr;
3787 void *base_ptr, *ptr, *p;
3788 ptrdiff_t field_offset;
3790 int len, i, j, k, op_len;
3793 /* check validity of iclog pointers */
3794 spin_lock(&log->l_icloglock);
3795 icptr = log->l_iclog;
3796 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3799 if (icptr != log->l_iclog)
3800 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3801 spin_unlock(&log->l_icloglock);
3803 /* check log magic numbers */
3804 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3805 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3807 base_ptr = ptr = &iclog->ic_header;
3808 p = &iclog->ic_header;
3809 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3810 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3811 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3816 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3817 base_ptr = ptr = iclog->ic_datap;
3819 xhdr = iclog->ic_data;
3820 for (i = 0; i < len; i++) {
3823 /* clientid is only 1 byte */
3824 p = &ophead->oh_clientid;
3825 field_offset = p - base_ptr;
3826 if (!syncing || (field_offset & 0x1ff)) {
3827 clientid = ophead->oh_clientid;
3829 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3830 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3831 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3832 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3833 clientid = xlog_get_client_id(
3834 xhdr[j].hic_xheader.xh_cycle_data[k]);
3836 clientid = xlog_get_client_id(
3837 iclog->ic_header.h_cycle_data[idx]);
3840 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3842 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3843 __func__, clientid, ophead,
3844 (unsigned long)field_offset);
3847 p = &ophead->oh_len;
3848 field_offset = p - base_ptr;
3849 if (!syncing || (field_offset & 0x1ff)) {
3850 op_len = be32_to_cpu(ophead->oh_len);
3852 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3853 (uintptr_t)iclog->ic_datap);
3854 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3855 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3856 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3857 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3859 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3862 ptr += sizeof(xlog_op_header_t) + op_len;
3864 } /* xlog_verify_iclog */
3868 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3874 xlog_in_core_t *iclog, *ic;
3876 iclog = log->l_iclog;
3877 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3879 * Mark all the incore logs IOERROR.
3880 * From now on, no log flushes will result.
3884 ic->ic_state = XLOG_STATE_IOERROR;
3886 } while (ic != iclog);
3890 * Return non-zero, if state transition has already happened.
3896 * This is called from xfs_force_shutdown, when we're forcibly
3897 * shutting down the filesystem, typically because of an IO error.
3898 * Our main objectives here are to make sure that:
3899 * a. if !logerror, flush the logs to disk. Anything modified
3900 * after this is ignored.
3901 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3902 * parties to find out, 'atomically'.
3903 * c. those who're sleeping on log reservations, pinned objects and
3904 * other resources get woken up, and be told the bad news.
3905 * d. nothing new gets queued up after (b) and (c) are done.
3907 * Note: for the !logerror case we need to flush the regions held in memory out
3908 * to disk first. This needs to be done before the log is marked as shutdown,
3909 * otherwise the iclog writes will fail.
3912 xfs_log_force_umount(
3913 struct xfs_mount *mp,
3922 * If this happens during log recovery, don't worry about
3923 * locking; the log isn't open for business yet.
3926 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3927 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3929 mp->m_sb_bp->b_flags |= XBF_DONE;
3934 * Somebody could've already done the hard work for us.
3935 * No need to get locks for this.
3937 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3938 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3943 * Flush all the completed transactions to disk before marking the log
3944 * being shut down. We need to do it in this order to ensure that
3945 * completed operations are safely on disk before we shut down, and that
3946 * we don't have to issue any buffer IO after the shutdown flags are set
3947 * to guarantee this.
3950 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3953 * mark the filesystem and the as in a shutdown state and wake
3954 * everybody up to tell them the bad news.
3956 spin_lock(&log->l_icloglock);
3957 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3959 mp->m_sb_bp->b_flags |= XBF_DONE;
3962 * Mark the log and the iclogs with IO error flags to prevent any
3963 * further log IO from being issued or completed.
3965 log->l_flags |= XLOG_IO_ERROR;
3966 retval = xlog_state_ioerror(log);
3967 spin_unlock(&log->l_icloglock);
3970 * We don't want anybody waiting for log reservations after this. That
3971 * means we have to wake up everybody queued up on reserveq as well as
3972 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3973 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3974 * action is protected by the grant locks.
3976 xlog_grant_head_wake_all(&log->l_reserve_head);
3977 xlog_grant_head_wake_all(&log->l_write_head);
3980 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3981 * as if the log writes were completed. The abort handling in the log
3982 * item committed callback functions will do this again under lock to
3985 wake_up_all(&log->l_cilp->xc_commit_wait);
3986 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3988 #ifdef XFSERRORDEBUG
3990 xlog_in_core_t *iclog;
3992 spin_lock(&log->l_icloglock);
3993 iclog = log->l_iclog;
3995 ASSERT(iclog->ic_callback == 0);
3996 iclog = iclog->ic_next;
3997 } while (iclog != log->l_iclog);
3998 spin_unlock(&log->l_icloglock);
4001 /* return non-zero if log IOERROR transition had already happened */
4009 xlog_in_core_t *iclog;
4011 iclog = log->l_iclog;
4013 /* endianness does not matter here, zero is zero in
4016 if (iclog->ic_header.h_num_logops)
4018 iclog = iclog->ic_next;
4019 } while (iclog != log->l_iclog);
4024 * Verify that an LSN stamped into a piece of metadata is valid. This is
4025 * intended for use in read verifiers on v5 superblocks.
4029 struct xfs_mount *mp,
4032 struct xlog *log = mp->m_log;
4036 * norecovery mode skips mount-time log processing and unconditionally
4037 * resets the in-core LSN. We can't validate in this mode, but
4038 * modifications are not allowed anyways so just return true.
4040 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4044 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4045 * handled by recovery and thus safe to ignore here.
4047 if (lsn == NULLCOMMITLSN)
4050 valid = xlog_valid_lsn(mp->m_log, lsn);
4052 /* warn the user about what's gone wrong before verifier failure */
4054 spin_lock(&log->l_icloglock);
4056 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4057 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4058 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4059 log->l_curr_cycle, log->l_curr_block);
4060 spin_unlock(&log->l_icloglock);