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,
441 struct xlog *log = mp->m_log;
442 struct xlog_ticket *tic;
446 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
448 if (XLOG_FORCED_SHUTDOWN(log))
451 XFS_STATS_INC(mp, xs_try_logspace);
453 ASSERT(*ticp == NULL);
454 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
455 KM_SLEEP | KM_MAYFAIL);
459 tic->t_trans_type = t_type;
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
465 trace_xfs_log_reserve(log, tic);
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
493 * 1. currblock field gets updated at startup and after in-core logs
494 * marked as with WANT_SYNC.
498 * This routine is called when a user of a log manager ticket is done with
499 * the reservation. If the ticket was ever used, then a commit record for
500 * the associated transaction is written out as a log operation header with
501 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
502 * a given ticket. If the ticket was one with a permanent reservation, then
503 * a few operations are done differently. Permanent reservation tickets by
504 * default don't release the reservation. They just commit the current
505 * transaction with the belief that the reservation is still needed. A flag
506 * must be passed in before permanent reservations are actually released.
507 * When these type of tickets are not released, they need to be set into
508 * the inited state again. By doing this, a start record will be written
509 * out when the next write occurs.
513 struct xfs_mount *mp,
514 struct xlog_ticket *ticket,
515 struct xlog_in_core **iclog,
518 struct xlog *log = mp->m_log;
521 if (XLOG_FORCED_SHUTDOWN(log) ||
523 * If nothing was ever written, don't write out commit record.
524 * If we get an error, just continue and give back the log ticket.
526 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
527 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
528 lsn = (xfs_lsn_t) -1;
534 trace_xfs_log_done_nonperm(log, ticket);
537 * Release ticket if not permanent reservation or a specific
538 * request has been made to release a permanent reservation.
540 xlog_ungrant_log_space(log, ticket);
542 trace_xfs_log_done_perm(log, ticket);
544 xlog_regrant_reserve_log_space(log, ticket);
545 /* If this ticket was a permanent reservation and we aren't
546 * trying to release it, reset the inited flags; so next time
547 * we write, a start record will be written out.
549 ticket->t_flags |= XLOG_TIC_INITED;
552 xfs_log_ticket_put(ticket);
557 * Attaches a new iclog I/O completion callback routine during
558 * transaction commit. If the log is in error state, a non-zero
559 * return code is handed back and the caller is responsible for
560 * executing the callback at an appropriate time.
564 struct xfs_mount *mp,
565 struct xlog_in_core *iclog,
566 xfs_log_callback_t *cb)
570 spin_lock(&iclog->ic_callback_lock);
571 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
573 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
574 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
576 *(iclog->ic_callback_tail) = cb;
577 iclog->ic_callback_tail = &(cb->cb_next);
579 spin_unlock(&iclog->ic_callback_lock);
584 xfs_log_release_iclog(
585 struct xfs_mount *mp,
586 struct xlog_in_core *iclog)
588 if (xlog_state_release_iclog(mp->m_log, iclog)) {
589 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
597 * Mount a log filesystem
599 * mp - ubiquitous xfs mount point structure
600 * log_target - buftarg of on-disk log device
601 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
602 * num_bblocks - Number of BBSIZE blocks in on-disk log
604 * Return error or zero.
609 xfs_buftarg_t *log_target,
610 xfs_daddr_t blk_offset,
616 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
617 xfs_notice(mp, "Mounting V%d Filesystem",
618 XFS_SB_VERSION_NUM(&mp->m_sb));
621 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
622 XFS_SB_VERSION_NUM(&mp->m_sb));
623 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
626 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
627 if (IS_ERR(mp->m_log)) {
628 error = PTR_ERR(mp->m_log);
633 * Validate the given log space and drop a critical message via syslog
634 * if the log size is too small that would lead to some unexpected
635 * situations in transaction log space reservation stage.
637 * Note: we can't just reject the mount if the validation fails. This
638 * would mean that people would have to downgrade their kernel just to
639 * remedy the situation as there is no way to grow the log (short of
640 * black magic surgery with xfs_db).
642 * We can, however, reject mounts for CRC format filesystems, as the
643 * mkfs binary being used to make the filesystem should never create a
644 * filesystem with a log that is too small.
646 min_logfsbs = xfs_log_calc_minimum_size(mp);
648 if (mp->m_sb.sb_logblocks < min_logfsbs) {
650 "Log size %d blocks too small, minimum size is %d blocks",
651 mp->m_sb.sb_logblocks, min_logfsbs);
653 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
655 "Log size %d blocks too large, maximum size is %lld blocks",
656 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
658 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
660 "log size %lld bytes too large, maximum size is %lld bytes",
661 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
666 if (xfs_sb_version_hascrc(&mp->m_sb)) {
667 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
671 xfs_crit(mp, "Log size out of supported range.");
673 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
677 * Initialize the AIL now we have a log.
679 error = xfs_trans_ail_init(mp);
681 xfs_warn(mp, "AIL initialisation failed: error %d", error);
684 mp->m_log->l_ailp = mp->m_ail;
687 * skip log recovery on a norecovery mount. pretend it all
690 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
691 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
694 mp->m_flags &= ~XFS_MOUNT_RDONLY;
696 error = xlog_recover(mp->m_log);
699 mp->m_flags |= XFS_MOUNT_RDONLY;
701 xfs_warn(mp, "log mount/recovery failed: error %d",
703 xlog_recover_cancel(mp->m_log);
704 goto out_destroy_ail;
708 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
711 goto out_destroy_ail;
713 /* Normal transactions can now occur */
714 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
717 * Now the log has been fully initialised and we know were our
718 * space grant counters are, we can initialise the permanent ticket
719 * needed for delayed logging to work.
721 xlog_cil_init_post_recovery(mp->m_log);
726 xfs_trans_ail_destroy(mp);
728 xlog_dealloc_log(mp->m_log);
734 * Finish the recovery of the file system. This is separate from the
735 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
736 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
739 * If we finish recovery successfully, start the background log work. If we are
740 * not doing recovery, then we have a RO filesystem and we don't need to start
744 xfs_log_mount_finish(
745 struct xfs_mount *mp)
749 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
750 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
754 error = xlog_recover_finish(mp->m_log);
756 xfs_log_work_queue(mp);
762 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
766 xfs_log_mount_cancel(
767 struct xfs_mount *mp)
771 error = xlog_recover_cancel(mp->m_log);
778 * Final log writes as part of unmount.
780 * Mark the filesystem clean as unmount happens. Note that during relocation
781 * this routine needs to be executed as part of source-bag while the
782 * deallocation must not be done until source-end.
786 * Unmount record used to have a string "Unmount filesystem--" in the
787 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
788 * We just write the magic number now since that particular field isn't
789 * currently architecture converted and "Unmount" is a bit foo.
790 * As far as I know, there weren't any dependencies on the old behaviour.
794 xfs_log_unmount_write(xfs_mount_t *mp)
796 struct xlog *log = mp->m_log;
797 xlog_in_core_t *iclog;
799 xlog_in_core_t *first_iclog;
801 xlog_ticket_t *tic = NULL;
806 * Don't write out unmount record on read-only mounts.
807 * Or, if we are doing a forced umount (typically because of IO errors).
809 if (mp->m_flags & XFS_MOUNT_RDONLY)
812 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
813 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
816 first_iclog = iclog = log->l_iclog;
818 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
819 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
820 ASSERT(iclog->ic_offset == 0);
822 iclog = iclog->ic_next;
823 } while (iclog != first_iclog);
825 if (! (XLOG_FORCED_SHUTDOWN(log))) {
826 error = xfs_log_reserve(mp, 600, 1, &tic,
827 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
829 /* the data section must be 32 bit size aligned */
833 __uint32_t pad2; /* may as well make it 64 bits */
835 .magic = XLOG_UNMOUNT_TYPE,
837 struct xfs_log_iovec reg = {
839 .i_len = sizeof(magic),
840 .i_type = XLOG_REG_TYPE_UNMOUNT,
842 struct xfs_log_vec vec = {
847 /* remove inited flag, and account for space used */
849 tic->t_curr_res -= sizeof(magic);
850 error = xlog_write(log, &vec, tic, &lsn,
851 NULL, XLOG_UNMOUNT_TRANS);
853 * At this point, we're umounting anyway,
854 * so there's no point in transitioning log state
855 * to IOERROR. Just continue...
860 xfs_alert(mp, "%s: unmount record failed", __func__);
863 spin_lock(&log->l_icloglock);
864 iclog = log->l_iclog;
865 atomic_inc(&iclog->ic_refcnt);
866 xlog_state_want_sync(log, iclog);
867 spin_unlock(&log->l_icloglock);
868 error = xlog_state_release_iclog(log, iclog);
870 spin_lock(&log->l_icloglock);
871 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
872 iclog->ic_state == XLOG_STATE_DIRTY)) {
873 if (!XLOG_FORCED_SHUTDOWN(log)) {
874 xlog_wait(&iclog->ic_force_wait,
877 spin_unlock(&log->l_icloglock);
880 spin_unlock(&log->l_icloglock);
883 trace_xfs_log_umount_write(log, tic);
884 xlog_ungrant_log_space(log, tic);
885 xfs_log_ticket_put(tic);
889 * We're already in forced_shutdown mode, couldn't
890 * even attempt to write out the unmount transaction.
892 * Go through the motions of sync'ing and releasing
893 * the iclog, even though no I/O will actually happen,
894 * we need to wait for other log I/Os that may already
895 * be in progress. Do this as a separate section of
896 * code so we'll know if we ever get stuck here that
897 * we're in this odd situation of trying to unmount
898 * a file system that went into forced_shutdown as
899 * the result of an unmount..
901 spin_lock(&log->l_icloglock);
902 iclog = log->l_iclog;
903 atomic_inc(&iclog->ic_refcnt);
905 xlog_state_want_sync(log, iclog);
906 spin_unlock(&log->l_icloglock);
907 error = xlog_state_release_iclog(log, iclog);
909 spin_lock(&log->l_icloglock);
911 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
912 || iclog->ic_state == XLOG_STATE_DIRTY
913 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
915 xlog_wait(&iclog->ic_force_wait,
918 spin_unlock(&log->l_icloglock);
923 } /* xfs_log_unmount_write */
926 * Empty the log for unmount/freeze.
928 * To do this, we first need to shut down the background log work so it is not
929 * trying to cover the log as we clean up. We then need to unpin all objects in
930 * the log so we can then flush them out. Once they have completed their IO and
931 * run the callbacks removing themselves from the AIL, we can write the unmount
936 struct xfs_mount *mp)
938 cancel_delayed_work_sync(&mp->m_log->l_work);
939 xfs_log_force(mp, XFS_LOG_SYNC);
942 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
943 * will push it, xfs_wait_buftarg() will not wait for it. Further,
944 * xfs_buf_iowait() cannot be used because it was pushed with the
945 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
946 * the IO to complete.
948 xfs_ail_push_all_sync(mp->m_ail);
949 xfs_wait_buftarg(mp->m_ddev_targp);
950 xfs_buf_lock(mp->m_sb_bp);
951 xfs_buf_unlock(mp->m_sb_bp);
953 xfs_log_unmount_write(mp);
957 * Shut down and release the AIL and Log.
959 * During unmount, we need to ensure we flush all the dirty metadata objects
960 * from the AIL so that the log is empty before we write the unmount record to
961 * the log. Once this is done, we can tear down the AIL and the log.
965 struct xfs_mount *mp)
969 xfs_trans_ail_destroy(mp);
971 xfs_sysfs_del(&mp->m_log->l_kobj);
973 xlog_dealloc_log(mp->m_log);
978 struct xfs_mount *mp,
979 struct xfs_log_item *item,
981 const struct xfs_item_ops *ops)
983 item->li_mountp = mp;
984 item->li_ailp = mp->m_ail;
985 item->li_type = type;
989 INIT_LIST_HEAD(&item->li_ail);
990 INIT_LIST_HEAD(&item->li_cil);
994 * Wake up processes waiting for log space after we have moved the log tail.
998 struct xfs_mount *mp)
1000 struct xlog *log = mp->m_log;
1003 if (XLOG_FORCED_SHUTDOWN(log))
1006 if (!list_empty_careful(&log->l_write_head.waiters)) {
1007 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1009 spin_lock(&log->l_write_head.lock);
1010 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1011 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1012 spin_unlock(&log->l_write_head.lock);
1015 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1016 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1018 spin_lock(&log->l_reserve_head.lock);
1019 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1020 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1021 spin_unlock(&log->l_reserve_head.lock);
1026 * Determine if we have a transaction that has gone to disk that needs to be
1027 * covered. To begin the transition to the idle state firstly the log needs to
1028 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1029 * we start attempting to cover the log.
1031 * Only if we are then in a state where covering is needed, the caller is
1032 * informed that dummy transactions are required to move the log into the idle
1035 * If there are any items in the AIl or CIL, then we do not want to attempt to
1036 * cover the log as we may be in a situation where there isn't log space
1037 * available to run a dummy transaction and this can lead to deadlocks when the
1038 * tail of the log is pinned by an item that is modified in the CIL. Hence
1039 * there's no point in running a dummy transaction at this point because we
1040 * can't start trying to idle the log until both the CIL and AIL are empty.
1043 xfs_log_need_covered(xfs_mount_t *mp)
1045 struct xlog *log = mp->m_log;
1048 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1051 if (!xlog_cil_empty(log))
1054 spin_lock(&log->l_icloglock);
1055 switch (log->l_covered_state) {
1056 case XLOG_STATE_COVER_DONE:
1057 case XLOG_STATE_COVER_DONE2:
1058 case XLOG_STATE_COVER_IDLE:
1060 case XLOG_STATE_COVER_NEED:
1061 case XLOG_STATE_COVER_NEED2:
1062 if (xfs_ail_min_lsn(log->l_ailp))
1064 if (!xlog_iclogs_empty(log))
1068 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1069 log->l_covered_state = XLOG_STATE_COVER_DONE;
1071 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1077 spin_unlock(&log->l_icloglock);
1082 * We may be holding the log iclog lock upon entering this routine.
1085 xlog_assign_tail_lsn_locked(
1086 struct xfs_mount *mp)
1088 struct xlog *log = mp->m_log;
1089 struct xfs_log_item *lip;
1092 assert_spin_locked(&mp->m_ail->xa_lock);
1095 * To make sure we always have a valid LSN for the log tail we keep
1096 * track of the last LSN which was committed in log->l_last_sync_lsn,
1097 * and use that when the AIL was empty.
1099 lip = xfs_ail_min(mp->m_ail);
1101 tail_lsn = lip->li_lsn;
1103 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1104 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1105 atomic64_set(&log->l_tail_lsn, tail_lsn);
1110 xlog_assign_tail_lsn(
1111 struct xfs_mount *mp)
1115 spin_lock(&mp->m_ail->xa_lock);
1116 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1117 spin_unlock(&mp->m_ail->xa_lock);
1123 * Return the space in the log between the tail and the head. The head
1124 * is passed in the cycle/bytes formal parms. In the special case where
1125 * the reserve head has wrapped passed the tail, this calculation is no
1126 * longer valid. In this case, just return 0 which means there is no space
1127 * in the log. This works for all places where this function is called
1128 * with the reserve head. Of course, if the write head were to ever
1129 * wrap the tail, we should blow up. Rather than catch this case here,
1130 * we depend on other ASSERTions in other parts of the code. XXXmiken
1132 * This code also handles the case where the reservation head is behind
1133 * the tail. The details of this case are described below, but the end
1134 * result is that we return the size of the log as the amount of space left.
1147 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1148 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1149 tail_bytes = BBTOB(tail_bytes);
1150 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1151 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1152 else if (tail_cycle + 1 < head_cycle)
1154 else if (tail_cycle < head_cycle) {
1155 ASSERT(tail_cycle == (head_cycle - 1));
1156 free_bytes = tail_bytes - head_bytes;
1159 * The reservation head is behind the tail.
1160 * In this case we just want to return the size of the
1161 * log as the amount of space left.
1163 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1164 xfs_alert(log->l_mp,
1165 " tail_cycle = %d, tail_bytes = %d",
1166 tail_cycle, tail_bytes);
1167 xfs_alert(log->l_mp,
1168 " GH cycle = %d, GH bytes = %d",
1169 head_cycle, head_bytes);
1171 free_bytes = log->l_logsize;
1178 * Log function which is called when an io completes.
1180 * The log manager needs its own routine, in order to control what
1181 * happens with the buffer after the write completes.
1184 xlog_iodone(xfs_buf_t *bp)
1186 struct xlog_in_core *iclog = bp->b_fspriv;
1187 struct xlog *l = iclog->ic_log;
1191 * Race to shutdown the filesystem if we see an error or the iclog is in
1192 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1193 * CRC errors into log recovery.
1195 if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR,
1196 XFS_RANDOM_IODONE_IOERR) ||
1197 iclog->ic_state & XLOG_STATE_IOABORT) {
1198 if (iclog->ic_state & XLOG_STATE_IOABORT)
1199 iclog->ic_state &= ~XLOG_STATE_IOABORT;
1201 xfs_buf_ioerror_alert(bp, __func__);
1203 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1205 * This flag will be propagated to the trans-committed
1206 * callback routines to let them know that the log-commit
1209 aborted = XFS_LI_ABORTED;
1210 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1211 aborted = XFS_LI_ABORTED;
1214 /* log I/O is always issued ASYNC */
1215 ASSERT(bp->b_flags & XBF_ASYNC);
1216 xlog_state_done_syncing(iclog, aborted);
1219 * drop the buffer lock now that we are done. Nothing references
1220 * the buffer after this, so an unmount waiting on this lock can now
1221 * tear it down safely. As such, it is unsafe to reference the buffer
1222 * (bp) after the unlock as we could race with it being freed.
1228 * Return size of each in-core log record buffer.
1230 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1232 * If the filesystem blocksize is too large, we may need to choose a
1233 * larger size since the directory code currently logs entire blocks.
1237 xlog_get_iclog_buffer_size(
1238 struct xfs_mount *mp,
1244 if (mp->m_logbufs <= 0)
1245 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1247 log->l_iclog_bufs = mp->m_logbufs;
1250 * Buffer size passed in from mount system call.
1252 if (mp->m_logbsize > 0) {
1253 size = log->l_iclog_size = mp->m_logbsize;
1254 log->l_iclog_size_log = 0;
1256 log->l_iclog_size_log++;
1260 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1261 /* # headers = size / 32k
1262 * one header holds cycles from 32k of data
1265 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1266 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1268 log->l_iclog_hsize = xhdrs << BBSHIFT;
1269 log->l_iclog_heads = xhdrs;
1271 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1272 log->l_iclog_hsize = BBSIZE;
1273 log->l_iclog_heads = 1;
1278 /* All machines use 32kB buffers by default. */
1279 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1280 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1282 /* the default log size is 16k or 32k which is one header sector */
1283 log->l_iclog_hsize = BBSIZE;
1284 log->l_iclog_heads = 1;
1287 /* are we being asked to make the sizes selected above visible? */
1288 if (mp->m_logbufs == 0)
1289 mp->m_logbufs = log->l_iclog_bufs;
1290 if (mp->m_logbsize == 0)
1291 mp->m_logbsize = log->l_iclog_size;
1292 } /* xlog_get_iclog_buffer_size */
1297 struct xfs_mount *mp)
1299 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1300 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1304 * Every sync period we need to unpin all items in the AIL and push them to
1305 * disk. If there is nothing dirty, then we might need to cover the log to
1306 * indicate that the filesystem is idle.
1310 struct work_struct *work)
1312 struct xlog *log = container_of(to_delayed_work(work),
1313 struct xlog, l_work);
1314 struct xfs_mount *mp = log->l_mp;
1316 /* dgc: errors ignored - not fatal and nowhere to report them */
1317 if (xfs_log_need_covered(mp)) {
1319 * Dump a transaction into the log that contains no real change.
1320 * This is needed to stamp the current tail LSN into the log
1321 * during the covering operation.
1323 * We cannot use an inode here for this - that will push dirty
1324 * state back up into the VFS and then periodic inode flushing
1325 * will prevent log covering from making progress. Hence we
1326 * synchronously log the superblock instead to ensure the
1327 * superblock is immediately unpinned and can be written back.
1329 xfs_sync_sb(mp, true);
1331 xfs_log_force(mp, 0);
1333 /* start pushing all the metadata that is currently dirty */
1334 xfs_ail_push_all(mp->m_ail);
1336 /* queue us up again */
1337 xfs_log_work_queue(mp);
1341 * This routine initializes some of the log structure for a given mount point.
1342 * Its primary purpose is to fill in enough, so recovery can occur. However,
1343 * some other stuff may be filled in too.
1345 STATIC struct xlog *
1347 struct xfs_mount *mp,
1348 struct xfs_buftarg *log_target,
1349 xfs_daddr_t blk_offset,
1353 xlog_rec_header_t *head;
1354 xlog_in_core_t **iclogp;
1355 xlog_in_core_t *iclog, *prev_iclog=NULL;
1358 int error = -ENOMEM;
1361 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1363 xfs_warn(mp, "Log allocation failed: No memory!");
1368 log->l_targ = log_target;
1369 log->l_logsize = BBTOB(num_bblks);
1370 log->l_logBBstart = blk_offset;
1371 log->l_logBBsize = num_bblks;
1372 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1373 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1374 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1376 log->l_prev_block = -1;
1377 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1378 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1379 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1380 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1382 xlog_grant_head_init(&log->l_reserve_head);
1383 xlog_grant_head_init(&log->l_write_head);
1385 error = -EFSCORRUPTED;
1386 if (xfs_sb_version_hassector(&mp->m_sb)) {
1387 log2_size = mp->m_sb.sb_logsectlog;
1388 if (log2_size < BBSHIFT) {
1389 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1390 log2_size, BBSHIFT);
1394 log2_size -= BBSHIFT;
1395 if (log2_size > mp->m_sectbb_log) {
1396 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1397 log2_size, mp->m_sectbb_log);
1401 /* for larger sector sizes, must have v2 or external log */
1402 if (log2_size && log->l_logBBstart > 0 &&
1403 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1405 "log sector size (0x%x) invalid for configuration.",
1410 log->l_sectBBsize = 1 << log2_size;
1412 xlog_get_iclog_buffer_size(mp, log);
1415 * Use a NULL block for the extra log buffer used during splits so that
1416 * it will trigger errors if we ever try to do IO on it without first
1417 * having set it up properly.
1420 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1421 BTOBB(log->l_iclog_size), 0);
1426 * The iclogbuf buffer locks are held over IO but we are not going to do
1427 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1428 * when appropriately.
1430 ASSERT(xfs_buf_islocked(bp));
1433 /* use high priority wq for log I/O completion */
1434 bp->b_ioend_wq = mp->m_log_workqueue;
1435 bp->b_iodone = xlog_iodone;
1438 spin_lock_init(&log->l_icloglock);
1439 init_waitqueue_head(&log->l_flush_wait);
1441 iclogp = &log->l_iclog;
1443 * The amount of memory to allocate for the iclog structure is
1444 * rather funky due to the way the structure is defined. It is
1445 * done this way so that we can use different sizes for machines
1446 * with different amounts of memory. See the definition of
1447 * xlog_in_core_t in xfs_log_priv.h for details.
1449 ASSERT(log->l_iclog_size >= 4096);
1450 for (i=0; i < log->l_iclog_bufs; i++) {
1451 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1453 goto out_free_iclog;
1456 iclog->ic_prev = prev_iclog;
1459 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1460 BTOBB(log->l_iclog_size), 0);
1462 goto out_free_iclog;
1464 ASSERT(xfs_buf_islocked(bp));
1467 /* use high priority wq for log I/O completion */
1468 bp->b_ioend_wq = mp->m_log_workqueue;
1469 bp->b_iodone = xlog_iodone;
1471 iclog->ic_data = bp->b_addr;
1473 log->l_iclog_bak[i] = &iclog->ic_header;
1475 head = &iclog->ic_header;
1476 memset(head, 0, sizeof(xlog_rec_header_t));
1477 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1478 head->h_version = cpu_to_be32(
1479 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1480 head->h_size = cpu_to_be32(log->l_iclog_size);
1482 head->h_fmt = cpu_to_be32(XLOG_FMT);
1483 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1485 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1486 iclog->ic_state = XLOG_STATE_ACTIVE;
1487 iclog->ic_log = log;
1488 atomic_set(&iclog->ic_refcnt, 0);
1489 spin_lock_init(&iclog->ic_callback_lock);
1490 iclog->ic_callback_tail = &(iclog->ic_callback);
1491 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1493 init_waitqueue_head(&iclog->ic_force_wait);
1494 init_waitqueue_head(&iclog->ic_write_wait);
1496 iclogp = &iclog->ic_next;
1498 *iclogp = log->l_iclog; /* complete ring */
1499 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1501 error = xlog_cil_init(log);
1503 goto out_free_iclog;
1507 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1508 prev_iclog = iclog->ic_next;
1510 xfs_buf_free(iclog->ic_bp);
1513 spinlock_destroy(&log->l_icloglock);
1514 xfs_buf_free(log->l_xbuf);
1518 return ERR_PTR(error);
1519 } /* xlog_alloc_log */
1523 * Write out the commit record of a transaction associated with the given
1524 * ticket. Return the lsn of the commit record.
1529 struct xlog_ticket *ticket,
1530 struct xlog_in_core **iclog,
1531 xfs_lsn_t *commitlsnp)
1533 struct xfs_mount *mp = log->l_mp;
1535 struct xfs_log_iovec reg = {
1538 .i_type = XLOG_REG_TYPE_COMMIT,
1540 struct xfs_log_vec vec = {
1545 ASSERT_ALWAYS(iclog);
1546 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1549 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1554 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1555 * log space. This code pushes on the lsn which would supposedly free up
1556 * the 25% which we want to leave free. We may need to adopt a policy which
1557 * pushes on an lsn which is further along in the log once we reach the high
1558 * water mark. In this manner, we would be creating a low water mark.
1561 xlog_grant_push_ail(
1565 xfs_lsn_t threshold_lsn = 0;
1566 xfs_lsn_t last_sync_lsn;
1569 int threshold_block;
1570 int threshold_cycle;
1573 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1575 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1576 free_blocks = BTOBBT(free_bytes);
1579 * Set the threshold for the minimum number of free blocks in the
1580 * log to the maximum of what the caller needs, one quarter of the
1581 * log, and 256 blocks.
1583 free_threshold = BTOBB(need_bytes);
1584 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1585 free_threshold = MAX(free_threshold, 256);
1586 if (free_blocks >= free_threshold)
1589 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1591 threshold_block += free_threshold;
1592 if (threshold_block >= log->l_logBBsize) {
1593 threshold_block -= log->l_logBBsize;
1594 threshold_cycle += 1;
1596 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1599 * Don't pass in an lsn greater than the lsn of the last
1600 * log record known to be on disk. Use a snapshot of the last sync lsn
1601 * so that it doesn't change between the compare and the set.
1603 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1604 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1605 threshold_lsn = last_sync_lsn;
1608 * Get the transaction layer to kick the dirty buffers out to
1609 * disk asynchronously. No point in trying to do this if
1610 * the filesystem is shutting down.
1612 if (!XLOG_FORCED_SHUTDOWN(log))
1613 xfs_ail_push(log->l_ailp, threshold_lsn);
1617 * Stamp cycle number in every block
1622 struct xlog_in_core *iclog,
1626 int size = iclog->ic_offset + roundoff;
1630 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1632 dp = iclog->ic_datap;
1633 for (i = 0; i < BTOBB(size); i++) {
1634 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1636 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1637 *(__be32 *)dp = cycle_lsn;
1641 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1642 xlog_in_core_2_t *xhdr = iclog->ic_data;
1644 for ( ; i < BTOBB(size); i++) {
1645 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1646 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1647 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1648 *(__be32 *)dp = cycle_lsn;
1652 for (i = 1; i < log->l_iclog_heads; i++)
1653 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1658 * Calculate the checksum for a log buffer.
1660 * This is a little more complicated than it should be because the various
1661 * headers and the actual data are non-contiguous.
1666 struct xlog_rec_header *rhead,
1672 /* first generate the crc for the record header ... */
1673 crc = xfs_start_cksum((char *)rhead,
1674 sizeof(struct xlog_rec_header),
1675 offsetof(struct xlog_rec_header, h_crc));
1677 /* ... then for additional cycle data for v2 logs ... */
1678 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1679 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1683 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1684 if (size % XLOG_HEADER_CYCLE_SIZE)
1687 for (i = 1; i < xheads; i++) {
1688 crc = crc32c(crc, &xhdr[i].hic_xheader,
1689 sizeof(struct xlog_rec_ext_header));
1693 /* ... and finally for the payload */
1694 crc = crc32c(crc, dp, size);
1696 return xfs_end_cksum(crc);
1700 * The bdstrat callback function for log bufs. This gives us a central
1701 * place to trap bufs in case we get hit by a log I/O error and need to
1702 * shutdown. Actually, in practice, even when we didn't get a log error,
1703 * we transition the iclogs to IOERROR state *after* flushing all existing
1704 * iclogs to disk. This is because we don't want anymore new transactions to be
1705 * started or completed afterwards.
1707 * We lock the iclogbufs here so that we can serialise against IO completion
1708 * during unmount. We might be processing a shutdown triggered during unmount,
1709 * and that can occur asynchronously to the unmount thread, and hence we need to
1710 * ensure that completes before tearing down the iclogbufs. Hence we need to
1711 * hold the buffer lock across the log IO to acheive that.
1717 struct xlog_in_core *iclog = bp->b_fspriv;
1720 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1721 xfs_buf_ioerror(bp, -EIO);
1725 * It would seem logical to return EIO here, but we rely on
1726 * the log state machine to propagate I/O errors instead of
1727 * doing it here. Similarly, IO completion will unlock the
1728 * buffer, so we don't do it here.
1738 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1739 * fashion. Previously, we should have moved the current iclog
1740 * ptr in the log to point to the next available iclog. This allows further
1741 * write to continue while this code syncs out an iclog ready to go.
1742 * Before an in-core log can be written out, the data section must be scanned
1743 * to save away the 1st word of each BBSIZE block into the header. We replace
1744 * it with the current cycle count. Each BBSIZE block is tagged with the
1745 * cycle count because there in an implicit assumption that drives will
1746 * guarantee that entire 512 byte blocks get written at once. In other words,
1747 * we can't have part of a 512 byte block written and part not written. By
1748 * tagging each block, we will know which blocks are valid when recovering
1749 * after an unclean shutdown.
1751 * This routine is single threaded on the iclog. No other thread can be in
1752 * this routine with the same iclog. Changing contents of iclog can there-
1753 * fore be done without grabbing the state machine lock. Updating the global
1754 * log will require grabbing the lock though.
1756 * The entire log manager uses a logical block numbering scheme. Only
1757 * log_sync (and then only bwrite()) know about the fact that the log may
1758 * not start with block zero on a given device. The log block start offset
1759 * is added immediately before calling bwrite().
1765 struct xlog_in_core *iclog)
1769 uint count; /* byte count of bwrite */
1770 uint count_init; /* initial count before roundup */
1771 int roundoff; /* roundoff to BB or stripe */
1772 int split = 0; /* split write into two regions */
1774 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1777 XFS_STATS_INC(log->l_mp, xs_log_writes);
1778 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1780 /* Add for LR header */
1781 count_init = log->l_iclog_hsize + iclog->ic_offset;
1783 /* Round out the log write size */
1784 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1785 /* we have a v2 stripe unit to use */
1786 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1788 count = BBTOB(BTOBB(count_init));
1790 roundoff = count - count_init;
1791 ASSERT(roundoff >= 0);
1792 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1793 roundoff < log->l_mp->m_sb.sb_logsunit)
1795 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1796 roundoff < BBTOB(1)));
1798 /* move grant heads by roundoff in sync */
1799 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1800 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1802 /* put cycle number in every block */
1803 xlog_pack_data(log, iclog, roundoff);
1805 /* real byte length */
1806 size = iclog->ic_offset;
1809 iclog->ic_header.h_len = cpu_to_be32(size);
1812 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1814 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1816 /* Do we need to split this write into 2 parts? */
1817 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1820 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1821 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1822 iclog->ic_bwritecnt = 2;
1825 * Bump the cycle numbers at the start of each block in the
1826 * part of the iclog that ends up in the buffer that gets
1827 * written to the start of the log.
1829 * Watch out for the header magic number case, though.
1831 dptr = (char *)&iclog->ic_header + count;
1832 for (i = 0; i < split; i += BBSIZE) {
1833 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1834 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1836 *(__be32 *)dptr = cpu_to_be32(cycle);
1841 iclog->ic_bwritecnt = 1;
1844 /* calculcate the checksum */
1845 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1846 iclog->ic_datap, size);
1849 * Intentionally corrupt the log record CRC based on the error injection
1850 * frequency, if defined. This facilitates testing log recovery in the
1851 * event of torn writes. Hence, set the IOABORT state to abort the log
1852 * write on I/O completion and shutdown the fs. The subsequent mount
1853 * detects the bad CRC and attempts to recover.
1855 if (log->l_badcrc_factor &&
1856 (prandom_u32() % log->l_badcrc_factor == 0)) {
1857 iclog->ic_header.h_crc &= 0xAAAAAAAA;
1858 iclog->ic_state |= XLOG_STATE_IOABORT;
1860 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1861 be64_to_cpu(iclog->ic_header.h_lsn));
1865 bp->b_io_length = BTOBB(count);
1866 bp->b_fspriv = iclog;
1867 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1868 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1870 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1871 bp->b_flags |= XBF_FUA;
1874 * Flush the data device before flushing the log to make
1875 * sure all meta data written back from the AIL actually made
1876 * it to disk before stamping the new log tail LSN into the
1877 * log buffer. For an external log we need to issue the
1878 * flush explicitly, and unfortunately synchronously here;
1879 * for an internal log we can simply use the block layer
1880 * state machine for preflushes.
1882 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1883 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1885 bp->b_flags |= XBF_FLUSH;
1888 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1889 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1891 xlog_verify_iclog(log, iclog, count, true);
1893 /* account for log which doesn't start at block #0 */
1894 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1897 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1900 error = xlog_bdstrat(bp);
1902 xfs_buf_ioerror_alert(bp, "xlog_sync");
1906 bp = iclog->ic_log->l_xbuf;
1907 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1908 xfs_buf_associate_memory(bp,
1909 (char *)&iclog->ic_header + count, split);
1910 bp->b_fspriv = iclog;
1911 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1912 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1913 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1914 bp->b_flags |= XBF_FUA;
1916 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1917 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1919 /* account for internal log which doesn't start at block #0 */
1920 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1921 error = xlog_bdstrat(bp);
1923 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1931 * Deallocate a log structure
1937 xlog_in_core_t *iclog, *next_iclog;
1940 xlog_cil_destroy(log);
1943 * Cycle all the iclogbuf locks to make sure all log IO completion
1944 * is done before we tear down these buffers.
1946 iclog = log->l_iclog;
1947 for (i = 0; i < log->l_iclog_bufs; i++) {
1948 xfs_buf_lock(iclog->ic_bp);
1949 xfs_buf_unlock(iclog->ic_bp);
1950 iclog = iclog->ic_next;
1954 * Always need to ensure that the extra buffer does not point to memory
1955 * owned by another log buffer before we free it. Also, cycle the lock
1956 * first to ensure we've completed IO on it.
1958 xfs_buf_lock(log->l_xbuf);
1959 xfs_buf_unlock(log->l_xbuf);
1960 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1961 xfs_buf_free(log->l_xbuf);
1963 iclog = log->l_iclog;
1964 for (i = 0; i < log->l_iclog_bufs; i++) {
1965 xfs_buf_free(iclog->ic_bp);
1966 next_iclog = iclog->ic_next;
1970 spinlock_destroy(&log->l_icloglock);
1972 log->l_mp->m_log = NULL;
1974 } /* xlog_dealloc_log */
1977 * Update counters atomically now that memcpy is done.
1981 xlog_state_finish_copy(
1983 struct xlog_in_core *iclog,
1987 spin_lock(&log->l_icloglock);
1989 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1990 iclog->ic_offset += copy_bytes;
1992 spin_unlock(&log->l_icloglock);
1993 } /* xlog_state_finish_copy */
1999 * print out info relating to regions written which consume
2004 struct xfs_mount *mp,
2005 struct xlog_ticket *ticket)
2008 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2010 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2011 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
2032 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2077 xfs_warn(mp, "xlog_write: reservation summary:");
2078 xfs_warn(mp, " trans type = %s (%u)",
2079 ((ticket->t_trans_type <= 0 ||
2080 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2081 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2082 ticket->t_trans_type);
2083 xfs_warn(mp, " unit res = %d bytes",
2084 ticket->t_unit_res);
2085 xfs_warn(mp, " current res = %d bytes",
2086 ticket->t_curr_res);
2087 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2088 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2089 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2090 ticket->t_res_num_ophdrs, ophdr_spc);
2091 xfs_warn(mp, " ophdr + reg = %u bytes",
2092 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2093 xfs_warn(mp, " num regions = %u",
2096 for (i = 0; i < ticket->t_res_num; i++) {
2097 uint r_type = ticket->t_res_arr[i].r_type;
2098 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2099 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2100 "bad-rtype" : res_type_str[r_type-1]),
2101 ticket->t_res_arr[i].r_len);
2104 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2105 "xlog_write: reservation ran out. Need to up reservation");
2106 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2110 * Calculate the potential space needed by the log vector. Each region gets
2111 * its own xlog_op_header_t and may need to be double word aligned.
2114 xlog_write_calc_vec_length(
2115 struct xlog_ticket *ticket,
2116 struct xfs_log_vec *log_vector)
2118 struct xfs_log_vec *lv;
2123 /* acct for start rec of xact */
2124 if (ticket->t_flags & XLOG_TIC_INITED)
2127 for (lv = log_vector; lv; lv = lv->lv_next) {
2128 /* we don't write ordered log vectors */
2129 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2132 headers += lv->lv_niovecs;
2134 for (i = 0; i < lv->lv_niovecs; i++) {
2135 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2138 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2142 ticket->t_res_num_ophdrs += headers;
2143 len += headers * sizeof(struct xlog_op_header);
2149 * If first write for transaction, insert start record We can't be trying to
2150 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2153 xlog_write_start_rec(
2154 struct xlog_op_header *ophdr,
2155 struct xlog_ticket *ticket)
2157 if (!(ticket->t_flags & XLOG_TIC_INITED))
2160 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2161 ophdr->oh_clientid = ticket->t_clientid;
2163 ophdr->oh_flags = XLOG_START_TRANS;
2166 ticket->t_flags &= ~XLOG_TIC_INITED;
2168 return sizeof(struct xlog_op_header);
2171 static xlog_op_header_t *
2172 xlog_write_setup_ophdr(
2174 struct xlog_op_header *ophdr,
2175 struct xlog_ticket *ticket,
2178 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2179 ophdr->oh_clientid = ticket->t_clientid;
2182 /* are we copying a commit or unmount record? */
2183 ophdr->oh_flags = flags;
2186 * We've seen logs corrupted with bad transaction client ids. This
2187 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2188 * and shut down the filesystem.
2190 switch (ophdr->oh_clientid) {
2191 case XFS_TRANSACTION:
2197 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2198 ophdr->oh_clientid, ticket);
2206 * Set up the parameters of the region copy into the log. This has
2207 * to handle region write split across multiple log buffers - this
2208 * state is kept external to this function so that this code can
2209 * be written in an obvious, self documenting manner.
2212 xlog_write_setup_copy(
2213 struct xlog_ticket *ticket,
2214 struct xlog_op_header *ophdr,
2215 int space_available,
2219 int *last_was_partial_copy,
2220 int *bytes_consumed)
2224 still_to_copy = space_required - *bytes_consumed;
2225 *copy_off = *bytes_consumed;
2227 if (still_to_copy <= space_available) {
2228 /* write of region completes here */
2229 *copy_len = still_to_copy;
2230 ophdr->oh_len = cpu_to_be32(*copy_len);
2231 if (*last_was_partial_copy)
2232 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2233 *last_was_partial_copy = 0;
2234 *bytes_consumed = 0;
2238 /* partial write of region, needs extra log op header reservation */
2239 *copy_len = space_available;
2240 ophdr->oh_len = cpu_to_be32(*copy_len);
2241 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2242 if (*last_was_partial_copy)
2243 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2244 *bytes_consumed += *copy_len;
2245 (*last_was_partial_copy)++;
2247 /* account for new log op header */
2248 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2249 ticket->t_res_num_ophdrs++;
2251 return sizeof(struct xlog_op_header);
2255 xlog_write_copy_finish(
2257 struct xlog_in_core *iclog,
2262 int *partial_copy_len,
2264 struct xlog_in_core **commit_iclog)
2266 if (*partial_copy) {
2268 * This iclog has already been marked WANT_SYNC by
2269 * xlog_state_get_iclog_space.
2271 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2274 return xlog_state_release_iclog(log, iclog);
2278 *partial_copy_len = 0;
2280 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2281 /* no more space in this iclog - push it. */
2282 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2286 spin_lock(&log->l_icloglock);
2287 xlog_state_want_sync(log, iclog);
2288 spin_unlock(&log->l_icloglock);
2291 return xlog_state_release_iclog(log, iclog);
2292 ASSERT(flags & XLOG_COMMIT_TRANS);
2293 *commit_iclog = iclog;
2300 * Write some region out to in-core log
2302 * This will be called when writing externally provided regions or when
2303 * writing out a commit record for a given transaction.
2305 * General algorithm:
2306 * 1. Find total length of this write. This may include adding to the
2307 * lengths passed in.
2308 * 2. Check whether we violate the tickets reservation.
2309 * 3. While writing to this iclog
2310 * A. Reserve as much space in this iclog as can get
2311 * B. If this is first write, save away start lsn
2312 * C. While writing this region:
2313 * 1. If first write of transaction, write start record
2314 * 2. Write log operation header (header per region)
2315 * 3. Find out if we can fit entire region into this iclog
2316 * 4. Potentially, verify destination memcpy ptr
2317 * 5. Memcpy (partial) region
2318 * 6. If partial copy, release iclog; otherwise, continue
2319 * copying more regions into current iclog
2320 * 4. Mark want sync bit (in simulation mode)
2321 * 5. Release iclog for potential flush to on-disk log.
2324 * 1. Panic if reservation is overrun. This should never happen since
2325 * reservation amounts are generated internal to the filesystem.
2327 * 1. Tickets are single threaded data structures.
2328 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2329 * syncing routine. When a single log_write region needs to span
2330 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2331 * on all log operation writes which don't contain the end of the
2332 * region. The XLOG_END_TRANS bit is used for the in-core log
2333 * operation which contains the end of the continued log_write region.
2334 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2335 * we don't really know exactly how much space will be used. As a result,
2336 * we don't update ic_offset until the end when we know exactly how many
2337 * bytes have been written out.
2342 struct xfs_log_vec *log_vector,
2343 struct xlog_ticket *ticket,
2344 xfs_lsn_t *start_lsn,
2345 struct xlog_in_core **commit_iclog,
2348 struct xlog_in_core *iclog = NULL;
2349 struct xfs_log_iovec *vecp;
2350 struct xfs_log_vec *lv;
2353 int partial_copy = 0;
2354 int partial_copy_len = 0;
2362 len = xlog_write_calc_vec_length(ticket, log_vector);
2365 * Region headers and bytes are already accounted for.
2366 * We only need to take into account start records and
2367 * split regions in this function.
2369 if (ticket->t_flags & XLOG_TIC_INITED)
2370 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2373 * Commit record headers need to be accounted for. These
2374 * come in as separate writes so are easy to detect.
2376 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2377 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2379 if (ticket->t_curr_res < 0)
2380 xlog_print_tic_res(log->l_mp, ticket);
2384 vecp = lv->lv_iovecp;
2385 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2389 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2390 &contwr, &log_offset);
2394 ASSERT(log_offset <= iclog->ic_size - 1);
2395 ptr = iclog->ic_datap + log_offset;
2397 /* start_lsn is the first lsn written to. That's all we need. */
2399 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2402 * This loop writes out as many regions as can fit in the amount
2403 * of space which was allocated by xlog_state_get_iclog_space().
2405 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2406 struct xfs_log_iovec *reg;
2407 struct xlog_op_header *ophdr;
2411 bool ordered = false;
2413 /* ordered log vectors have no regions to write */
2414 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2415 ASSERT(lv->lv_niovecs == 0);
2421 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2422 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2424 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2425 if (start_rec_copy) {
2427 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2431 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2435 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2436 sizeof(struct xlog_op_header));
2438 len += xlog_write_setup_copy(ticket, ophdr,
2439 iclog->ic_size-log_offset,
2441 ©_off, ©_len,
2444 xlog_verify_dest_ptr(log, ptr);
2449 * Unmount records just log an opheader, so can have
2450 * empty payloads with no data region to copy. Hence we
2451 * only copy the payload if the vector says it has data
2454 ASSERT(copy_len >= 0);
2456 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2457 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2460 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2462 data_cnt += contwr ? copy_len : 0;
2464 error = xlog_write_copy_finish(log, iclog, flags,
2465 &record_cnt, &data_cnt,
2474 * if we had a partial copy, we need to get more iclog
2475 * space but we don't want to increment the region
2476 * index because there is still more is this region to
2479 * If we completed writing this region, and we flushed
2480 * the iclog (indicated by resetting of the record
2481 * count), then we also need to get more log space. If
2482 * this was the last record, though, we are done and
2488 if (++index == lv->lv_niovecs) {
2493 vecp = lv->lv_iovecp;
2495 if (record_cnt == 0 && ordered == false) {
2505 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2507 return xlog_state_release_iclog(log, iclog);
2509 ASSERT(flags & XLOG_COMMIT_TRANS);
2510 *commit_iclog = iclog;
2515 /*****************************************************************************
2517 * State Machine functions
2519 *****************************************************************************
2522 /* Clean iclogs starting from the head. This ordering must be
2523 * maintained, so an iclog doesn't become ACTIVE beyond one that
2524 * is SYNCING. This is also required to maintain the notion that we use
2525 * a ordered wait queue to hold off would be writers to the log when every
2526 * iclog is trying to sync to disk.
2528 * State Change: DIRTY -> ACTIVE
2531 xlog_state_clean_log(
2534 xlog_in_core_t *iclog;
2537 iclog = log->l_iclog;
2539 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2540 iclog->ic_state = XLOG_STATE_ACTIVE;
2541 iclog->ic_offset = 0;
2542 ASSERT(iclog->ic_callback == NULL);
2544 * If the number of ops in this iclog indicate it just
2545 * contains the dummy transaction, we can
2546 * change state into IDLE (the second time around).
2547 * Otherwise we should change the state into
2549 * We don't need to cover the dummy.
2552 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2557 * We have two dirty iclogs so start over
2558 * This could also be num of ops indicates
2559 * this is not the dummy going out.
2563 iclog->ic_header.h_num_logops = 0;
2564 memset(iclog->ic_header.h_cycle_data, 0,
2565 sizeof(iclog->ic_header.h_cycle_data));
2566 iclog->ic_header.h_lsn = 0;
2567 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2570 break; /* stop cleaning */
2571 iclog = iclog->ic_next;
2572 } while (iclog != log->l_iclog);
2574 /* log is locked when we are called */
2576 * Change state for the dummy log recording.
2577 * We usually go to NEED. But we go to NEED2 if the changed indicates
2578 * we are done writing the dummy record.
2579 * If we are done with the second dummy recored (DONE2), then
2583 switch (log->l_covered_state) {
2584 case XLOG_STATE_COVER_IDLE:
2585 case XLOG_STATE_COVER_NEED:
2586 case XLOG_STATE_COVER_NEED2:
2587 log->l_covered_state = XLOG_STATE_COVER_NEED;
2590 case XLOG_STATE_COVER_DONE:
2592 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2594 log->l_covered_state = XLOG_STATE_COVER_NEED;
2597 case XLOG_STATE_COVER_DONE2:
2599 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2601 log->l_covered_state = XLOG_STATE_COVER_NEED;
2608 } /* xlog_state_clean_log */
2611 xlog_get_lowest_lsn(
2614 xlog_in_core_t *lsn_log;
2615 xfs_lsn_t lowest_lsn, lsn;
2617 lsn_log = log->l_iclog;
2620 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2621 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2622 if ((lsn && !lowest_lsn) ||
2623 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2627 lsn_log = lsn_log->ic_next;
2628 } while (lsn_log != log->l_iclog);
2634 xlog_state_do_callback(
2637 struct xlog_in_core *ciclog)
2639 xlog_in_core_t *iclog;
2640 xlog_in_core_t *first_iclog; /* used to know when we've
2641 * processed all iclogs once */
2642 xfs_log_callback_t *cb, *cb_next;
2644 xfs_lsn_t lowest_lsn;
2645 int ioerrors; /* counter: iclogs with errors */
2646 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2647 int funcdidcallbacks; /* flag: function did callbacks */
2648 int repeats; /* for issuing console warnings if
2649 * looping too many times */
2652 spin_lock(&log->l_icloglock);
2653 first_iclog = iclog = log->l_iclog;
2655 funcdidcallbacks = 0;
2660 * Scan all iclogs starting with the one pointed to by the
2661 * log. Reset this starting point each time the log is
2662 * unlocked (during callbacks).
2664 * Keep looping through iclogs until one full pass is made
2665 * without running any callbacks.
2667 first_iclog = log->l_iclog;
2668 iclog = log->l_iclog;
2669 loopdidcallbacks = 0;
2674 /* skip all iclogs in the ACTIVE & DIRTY states */
2675 if (iclog->ic_state &
2676 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2677 iclog = iclog->ic_next;
2682 * Between marking a filesystem SHUTDOWN and stopping
2683 * the log, we do flush all iclogs to disk (if there
2684 * wasn't a log I/O error). So, we do want things to
2685 * go smoothly in case of just a SHUTDOWN w/o a
2688 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2690 * Can only perform callbacks in order. Since
2691 * this iclog is not in the DONE_SYNC/
2692 * DO_CALLBACK state, we skip the rest and
2693 * just try to clean up. If we set our iclog
2694 * to DO_CALLBACK, we will not process it when
2695 * we retry since a previous iclog is in the
2696 * CALLBACK and the state cannot change since
2697 * we are holding the l_icloglock.
2699 if (!(iclog->ic_state &
2700 (XLOG_STATE_DONE_SYNC |
2701 XLOG_STATE_DO_CALLBACK))) {
2702 if (ciclog && (ciclog->ic_state ==
2703 XLOG_STATE_DONE_SYNC)) {
2704 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2709 * We now have an iclog that is in either the
2710 * DO_CALLBACK or DONE_SYNC states. The other
2711 * states (WANT_SYNC, SYNCING, or CALLBACK were
2712 * caught by the above if and are going to
2713 * clean (i.e. we aren't doing their callbacks)
2718 * We will do one more check here to see if we
2719 * have chased our tail around.
2722 lowest_lsn = xlog_get_lowest_lsn(log);
2724 XFS_LSN_CMP(lowest_lsn,
2725 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2726 iclog = iclog->ic_next;
2727 continue; /* Leave this iclog for
2731 iclog->ic_state = XLOG_STATE_CALLBACK;
2735 * Completion of a iclog IO does not imply that
2736 * a transaction has completed, as transactions
2737 * can be large enough to span many iclogs. We
2738 * cannot change the tail of the log half way
2739 * through a transaction as this may be the only
2740 * transaction in the log and moving th etail to
2741 * point to the middle of it will prevent
2742 * recovery from finding the start of the
2743 * transaction. Hence we should only update the
2744 * last_sync_lsn if this iclog contains
2745 * transaction completion callbacks on it.
2747 * We have to do this before we drop the
2748 * icloglock to ensure we are the only one that
2751 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2752 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2753 if (iclog->ic_callback)
2754 atomic64_set(&log->l_last_sync_lsn,
2755 be64_to_cpu(iclog->ic_header.h_lsn));
2760 spin_unlock(&log->l_icloglock);
2763 * Keep processing entries in the callback list until
2764 * we come around and it is empty. We need to
2765 * atomically see that the list is empty and change the
2766 * state to DIRTY so that we don't miss any more
2767 * callbacks being added.
2769 spin_lock(&iclog->ic_callback_lock);
2770 cb = iclog->ic_callback;
2772 iclog->ic_callback_tail = &(iclog->ic_callback);
2773 iclog->ic_callback = NULL;
2774 spin_unlock(&iclog->ic_callback_lock);
2776 /* perform callbacks in the order given */
2777 for (; cb; cb = cb_next) {
2778 cb_next = cb->cb_next;
2779 cb->cb_func(cb->cb_arg, aborted);
2781 spin_lock(&iclog->ic_callback_lock);
2782 cb = iclog->ic_callback;
2788 spin_lock(&log->l_icloglock);
2789 ASSERT(iclog->ic_callback == NULL);
2790 spin_unlock(&iclog->ic_callback_lock);
2791 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2792 iclog->ic_state = XLOG_STATE_DIRTY;
2795 * Transition from DIRTY to ACTIVE if applicable.
2796 * NOP if STATE_IOERROR.
2798 xlog_state_clean_log(log);
2800 /* wake up threads waiting in xfs_log_force() */
2801 wake_up_all(&iclog->ic_force_wait);
2803 iclog = iclog->ic_next;
2804 } while (first_iclog != iclog);
2806 if (repeats > 5000) {
2807 flushcnt += repeats;
2810 "%s: possible infinite loop (%d iterations)",
2811 __func__, flushcnt);
2813 } while (!ioerrors && loopdidcallbacks);
2817 * Make one last gasp attempt to see if iclogs are being left in limbo.
2818 * If the above loop finds an iclog earlier than the current iclog and
2819 * in one of the syncing states, the current iclog is put into
2820 * DO_CALLBACK and the callbacks are deferred to the completion of the
2821 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2822 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2825 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2826 * for ic_state == SYNCING.
2828 if (funcdidcallbacks) {
2829 first_iclog = iclog = log->l_iclog;
2831 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2833 * Terminate the loop if iclogs are found in states
2834 * which will cause other threads to clean up iclogs.
2836 * SYNCING - i/o completion will go through logs
2837 * DONE_SYNC - interrupt thread should be waiting for
2839 * IOERROR - give up hope all ye who enter here
2841 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2842 iclog->ic_state & XLOG_STATE_SYNCING ||
2843 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2844 iclog->ic_state == XLOG_STATE_IOERROR )
2846 iclog = iclog->ic_next;
2847 } while (first_iclog != iclog);
2851 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2853 spin_unlock(&log->l_icloglock);
2856 wake_up_all(&log->l_flush_wait);
2861 * Finish transitioning this iclog to the dirty state.
2863 * Make sure that we completely execute this routine only when this is
2864 * the last call to the iclog. There is a good chance that iclog flushes,
2865 * when we reach the end of the physical log, get turned into 2 separate
2866 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2867 * routine. By using the reference count bwritecnt, we guarantee that only
2868 * the second completion goes through.
2870 * Callbacks could take time, so they are done outside the scope of the
2871 * global state machine log lock.
2874 xlog_state_done_syncing(
2875 xlog_in_core_t *iclog,
2878 struct xlog *log = iclog->ic_log;
2880 spin_lock(&log->l_icloglock);
2882 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2883 iclog->ic_state == XLOG_STATE_IOERROR);
2884 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2885 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2889 * If we got an error, either on the first buffer, or in the case of
2890 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2891 * and none should ever be attempted to be written to disk
2894 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2895 if (--iclog->ic_bwritecnt == 1) {
2896 spin_unlock(&log->l_icloglock);
2899 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2903 * Someone could be sleeping prior to writing out the next
2904 * iclog buffer, we wake them all, one will get to do the
2905 * I/O, the others get to wait for the result.
2907 wake_up_all(&iclog->ic_write_wait);
2908 spin_unlock(&log->l_icloglock);
2909 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2910 } /* xlog_state_done_syncing */
2914 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2915 * sleep. We wait on the flush queue on the head iclog as that should be
2916 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2917 * we will wait here and all new writes will sleep until a sync completes.
2919 * The in-core logs are used in a circular fashion. They are not used
2920 * out-of-order even when an iclog past the head is free.
2923 * * log_offset where xlog_write() can start writing into the in-core
2925 * * in-core log pointer to which xlog_write() should write.
2926 * * boolean indicating this is a continued write to an in-core log.
2927 * If this is the last write, then the in-core log's offset field
2928 * needs to be incremented, depending on the amount of data which
2932 xlog_state_get_iclog_space(
2935 struct xlog_in_core **iclogp,
2936 struct xlog_ticket *ticket,
2937 int *continued_write,
2941 xlog_rec_header_t *head;
2942 xlog_in_core_t *iclog;
2946 spin_lock(&log->l_icloglock);
2947 if (XLOG_FORCED_SHUTDOWN(log)) {
2948 spin_unlock(&log->l_icloglock);
2952 iclog = log->l_iclog;
2953 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2954 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2956 /* Wait for log writes to have flushed */
2957 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2961 head = &iclog->ic_header;
2963 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2964 log_offset = iclog->ic_offset;
2966 /* On the 1st write to an iclog, figure out lsn. This works
2967 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2968 * committing to. If the offset is set, that's how many blocks
2971 if (log_offset == 0) {
2972 ticket->t_curr_res -= log->l_iclog_hsize;
2973 xlog_tic_add_region(ticket,
2975 XLOG_REG_TYPE_LRHEADER);
2976 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2977 head->h_lsn = cpu_to_be64(
2978 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2979 ASSERT(log->l_curr_block >= 0);
2982 /* If there is enough room to write everything, then do it. Otherwise,
2983 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2984 * bit is on, so this will get flushed out. Don't update ic_offset
2985 * until you know exactly how many bytes get copied. Therefore, wait
2986 * until later to update ic_offset.
2988 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2989 * can fit into remaining data section.
2991 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2992 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2995 * If I'm the only one writing to this iclog, sync it to disk.
2996 * We need to do an atomic compare and decrement here to avoid
2997 * racing with concurrent atomic_dec_and_lock() calls in
2998 * xlog_state_release_iclog() when there is more than one
2999 * reference to the iclog.
3001 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3002 /* we are the only one */
3003 spin_unlock(&log->l_icloglock);
3004 error = xlog_state_release_iclog(log, iclog);
3008 spin_unlock(&log->l_icloglock);
3013 /* Do we have enough room to write the full amount in the remainder
3014 * of this iclog? Or must we continue a write on the next iclog and
3015 * mark this iclog as completely taken? In the case where we switch
3016 * iclogs (to mark it taken), this particular iclog will release/sync
3017 * to disk in xlog_write().
3019 if (len <= iclog->ic_size - iclog->ic_offset) {
3020 *continued_write = 0;
3021 iclog->ic_offset += len;
3023 *continued_write = 1;
3024 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3028 ASSERT(iclog->ic_offset <= iclog->ic_size);
3029 spin_unlock(&log->l_icloglock);
3031 *logoffsetp = log_offset;
3033 } /* xlog_state_get_iclog_space */
3035 /* The first cnt-1 times through here we don't need to
3036 * move the grant write head because the permanent
3037 * reservation has reserved cnt times the unit amount.
3038 * Release part of current permanent unit reservation and
3039 * reset current reservation to be one units worth. Also
3040 * move grant reservation head forward.
3043 xlog_regrant_reserve_log_space(
3045 struct xlog_ticket *ticket)
3047 trace_xfs_log_regrant_reserve_enter(log, ticket);
3049 if (ticket->t_cnt > 0)
3052 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3053 ticket->t_curr_res);
3054 xlog_grant_sub_space(log, &log->l_write_head.grant,
3055 ticket->t_curr_res);
3056 ticket->t_curr_res = ticket->t_unit_res;
3057 xlog_tic_reset_res(ticket);
3059 trace_xfs_log_regrant_reserve_sub(log, ticket);
3061 /* just return if we still have some of the pre-reserved space */
3062 if (ticket->t_cnt > 0)
3065 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3066 ticket->t_unit_res);
3068 trace_xfs_log_regrant_reserve_exit(log, ticket);
3070 ticket->t_curr_res = ticket->t_unit_res;
3071 xlog_tic_reset_res(ticket);
3072 } /* xlog_regrant_reserve_log_space */
3076 * Give back the space left from a reservation.
3078 * All the information we need to make a correct determination of space left
3079 * is present. For non-permanent reservations, things are quite easy. The
3080 * count should have been decremented to zero. We only need to deal with the
3081 * space remaining in the current reservation part of the ticket. If the
3082 * ticket contains a permanent reservation, there may be left over space which
3083 * needs to be released. A count of N means that N-1 refills of the current
3084 * reservation can be done before we need to ask for more space. The first
3085 * one goes to fill up the first current reservation. Once we run out of
3086 * space, the count will stay at zero and the only space remaining will be
3087 * in the current reservation field.
3090 xlog_ungrant_log_space(
3092 struct xlog_ticket *ticket)
3096 if (ticket->t_cnt > 0)
3099 trace_xfs_log_ungrant_enter(log, ticket);
3100 trace_xfs_log_ungrant_sub(log, ticket);
3103 * If this is a permanent reservation ticket, we may be able to free
3104 * up more space based on the remaining count.
3106 bytes = ticket->t_curr_res;
3107 if (ticket->t_cnt > 0) {
3108 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3109 bytes += ticket->t_unit_res*ticket->t_cnt;
3112 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3113 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3115 trace_xfs_log_ungrant_exit(log, ticket);
3117 xfs_log_space_wake(log->l_mp);
3121 * Flush iclog to disk if this is the last reference to the given iclog and
3122 * the WANT_SYNC bit is set.
3124 * When this function is entered, the iclog is not necessarily in the
3125 * WANT_SYNC state. It may be sitting around waiting to get filled.
3130 xlog_state_release_iclog(
3132 struct xlog_in_core *iclog)
3134 int sync = 0; /* do we sync? */
3136 if (iclog->ic_state & XLOG_STATE_IOERROR)
3139 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3140 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3143 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3144 spin_unlock(&log->l_icloglock);
3147 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3148 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3150 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3151 /* update tail before writing to iclog */
3152 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3154 iclog->ic_state = XLOG_STATE_SYNCING;
3155 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3156 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3157 /* cycle incremented when incrementing curr_block */
3159 spin_unlock(&log->l_icloglock);
3162 * We let the log lock go, so it's possible that we hit a log I/O
3163 * error or some other SHUTDOWN condition that marks the iclog
3164 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3165 * this iclog has consistent data, so we ignore IOERROR
3166 * flags after this point.
3169 return xlog_sync(log, iclog);
3171 } /* xlog_state_release_iclog */
3175 * This routine will mark the current iclog in the ring as WANT_SYNC
3176 * and move the current iclog pointer to the next iclog in the ring.
3177 * When this routine is called from xlog_state_get_iclog_space(), the
3178 * exact size of the iclog has not yet been determined. All we know is
3179 * that every data block. We have run out of space in this log record.
3182 xlog_state_switch_iclogs(
3184 struct xlog_in_core *iclog,
3187 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3189 eventual_size = iclog->ic_offset;
3190 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3191 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3192 log->l_prev_block = log->l_curr_block;
3193 log->l_prev_cycle = log->l_curr_cycle;
3195 /* roll log?: ic_offset changed later */
3196 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3198 /* Round up to next log-sunit */
3199 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3200 log->l_mp->m_sb.sb_logsunit > 1) {
3201 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3202 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3205 if (log->l_curr_block >= log->l_logBBsize) {
3207 * Rewind the current block before the cycle is bumped to make
3208 * sure that the combined LSN never transiently moves forward
3209 * when the log wraps to the next cycle. This is to support the
3210 * unlocked sample of these fields from xlog_valid_lsn(). Most
3211 * other cases should acquire l_icloglock.
3213 log->l_curr_block -= log->l_logBBsize;
3214 ASSERT(log->l_curr_block >= 0);
3216 log->l_curr_cycle++;
3217 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3218 log->l_curr_cycle++;
3220 ASSERT(iclog == log->l_iclog);
3221 log->l_iclog = iclog->ic_next;
3222 } /* xlog_state_switch_iclogs */
3225 * Write out all data in the in-core log as of this exact moment in time.
3227 * Data may be written to the in-core log during this call. However,
3228 * we don't guarantee this data will be written out. A change from past
3229 * implementation means this routine will *not* write out zero length LRs.
3231 * Basically, we try and perform an intelligent scan of the in-core logs.
3232 * If we determine there is no flushable data, we just return. There is no
3233 * flushable data if:
3235 * 1. the current iclog is active and has no data; the previous iclog
3236 * is in the active or dirty state.
3237 * 2. the current iclog is drity, and the previous iclog is in the
3238 * active or dirty state.
3242 * 1. the current iclog is not in the active nor dirty state.
3243 * 2. the current iclog dirty, and the previous iclog is not in the
3244 * active nor dirty state.
3245 * 3. the current iclog is active, and there is another thread writing
3246 * to this particular iclog.
3247 * 4. a) the current iclog is active and has no other writers
3248 * b) when we return from flushing out this iclog, it is still
3249 * not in the active nor dirty state.
3253 struct xfs_mount *mp,
3257 struct xlog *log = mp->m_log;
3258 struct xlog_in_core *iclog;
3261 XFS_STATS_INC(mp, xs_log_force);
3263 xlog_cil_force(log);
3265 spin_lock(&log->l_icloglock);
3267 iclog = log->l_iclog;
3268 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3269 spin_unlock(&log->l_icloglock);
3273 /* If the head iclog is not active nor dirty, we just attach
3274 * ourselves to the head and go to sleep.
3276 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3277 iclog->ic_state == XLOG_STATE_DIRTY) {
3279 * If the head is dirty or (active and empty), then
3280 * we need to look at the previous iclog. If the previous
3281 * iclog is active or dirty we are done. There is nothing
3282 * to sync out. Otherwise, we attach ourselves to the
3283 * previous iclog and go to sleep.
3285 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3286 (atomic_read(&iclog->ic_refcnt) == 0
3287 && iclog->ic_offset == 0)) {
3288 iclog = iclog->ic_prev;
3289 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3290 iclog->ic_state == XLOG_STATE_DIRTY)
3295 if (atomic_read(&iclog->ic_refcnt) == 0) {
3296 /* We are the only one with access to this
3297 * iclog. Flush it out now. There should
3298 * be a roundoff of zero to show that someone
3299 * has already taken care of the roundoff from
3300 * the previous sync.
3302 atomic_inc(&iclog->ic_refcnt);
3303 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3304 xlog_state_switch_iclogs(log, iclog, 0);
3305 spin_unlock(&log->l_icloglock);
3307 if (xlog_state_release_iclog(log, iclog))
3312 spin_lock(&log->l_icloglock);
3313 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3314 iclog->ic_state != XLOG_STATE_DIRTY)
3319 /* Someone else is writing to this iclog.
3320 * Use its call to flush out the data. However,
3321 * the other thread may not force out this LR,
3322 * so we mark it WANT_SYNC.
3324 xlog_state_switch_iclogs(log, iclog, 0);
3330 /* By the time we come around again, the iclog could've been filled
3331 * which would give it another lsn. If we have a new lsn, just
3332 * return because the relevant data has been flushed.
3335 if (flags & XFS_LOG_SYNC) {
3337 * We must check if we're shutting down here, before
3338 * we wait, while we're holding the l_icloglock.
3339 * Then we check again after waking up, in case our
3340 * sleep was disturbed by a bad news.
3342 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3343 spin_unlock(&log->l_icloglock);
3346 XFS_STATS_INC(mp, xs_log_force_sleep);
3347 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3349 * No need to grab the log lock here since we're
3350 * only deciding whether or not to return EIO
3351 * and the memory read should be atomic.
3353 if (iclog->ic_state & XLOG_STATE_IOERROR)
3360 spin_unlock(&log->l_icloglock);
3366 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3367 * about errors or whether the log was flushed or not. This is the normal
3368 * interface to use when trying to unpin items or move the log forward.
3377 trace_xfs_log_force(mp, 0);
3378 error = _xfs_log_force(mp, flags, NULL);
3380 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3384 * Force the in-core log to disk for a specific LSN.
3386 * Find in-core log with lsn.
3387 * If it is in the DIRTY state, just return.
3388 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3389 * state and go to sleep or return.
3390 * If it is in any other state, go to sleep or return.
3392 * Synchronous forces are implemented with a signal variable. All callers
3393 * to force a given lsn to disk will wait on a the sv attached to the
3394 * specific in-core log. When given in-core log finally completes its
3395 * write to disk, that thread will wake up all threads waiting on the
3400 struct xfs_mount *mp,
3405 struct xlog *log = mp->m_log;
3406 struct xlog_in_core *iclog;
3407 int already_slept = 0;
3411 XFS_STATS_INC(mp, xs_log_force);
3413 lsn = xlog_cil_force_lsn(log, lsn);
3414 if (lsn == NULLCOMMITLSN)
3418 spin_lock(&log->l_icloglock);
3419 iclog = log->l_iclog;
3420 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3421 spin_unlock(&log->l_icloglock);
3426 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3427 iclog = iclog->ic_next;
3431 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3432 spin_unlock(&log->l_icloglock);
3436 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3438 * We sleep here if we haven't already slept (e.g.
3439 * this is the first time we've looked at the correct
3440 * iclog buf) and the buffer before us is going to
3441 * be sync'ed. The reason for this is that if we
3442 * are doing sync transactions here, by waiting for
3443 * the previous I/O to complete, we can allow a few
3444 * more transactions into this iclog before we close
3447 * Otherwise, we mark the buffer WANT_SYNC, and bump
3448 * up the refcnt so we can release the log (which
3449 * drops the ref count). The state switch keeps new
3450 * transaction commits from using this buffer. When
3451 * the current commits finish writing into the buffer,
3452 * the refcount will drop to zero and the buffer will
3455 if (!already_slept &&
3456 (iclog->ic_prev->ic_state &
3457 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3458 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3460 XFS_STATS_INC(mp, xs_log_force_sleep);
3462 xlog_wait(&iclog->ic_prev->ic_write_wait,
3469 atomic_inc(&iclog->ic_refcnt);
3470 xlog_state_switch_iclogs(log, iclog, 0);
3471 spin_unlock(&log->l_icloglock);
3472 if (xlog_state_release_iclog(log, iclog))
3476 spin_lock(&log->l_icloglock);
3479 if ((flags & XFS_LOG_SYNC) && /* sleep */
3481 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3483 * Don't wait on completion if we know that we've
3484 * gotten a log write error.
3486 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3487 spin_unlock(&log->l_icloglock);
3490 XFS_STATS_INC(mp, xs_log_force_sleep);
3491 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3493 * No need to grab the log lock here since we're
3494 * only deciding whether or not to return EIO
3495 * and the memory read should be atomic.
3497 if (iclog->ic_state & XLOG_STATE_IOERROR)
3502 } else { /* just return */
3503 spin_unlock(&log->l_icloglock);
3507 } while (iclog != log->l_iclog);
3509 spin_unlock(&log->l_icloglock);
3514 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3515 * about errors or whether the log was flushed or not. This is the normal
3516 * interface to use when trying to unpin items or move the log forward.
3526 trace_xfs_log_force(mp, lsn);
3527 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3529 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3533 * Called when we want to mark the current iclog as being ready to sync to
3537 xlog_state_want_sync(
3539 struct xlog_in_core *iclog)
3541 assert_spin_locked(&log->l_icloglock);
3543 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3544 xlog_state_switch_iclogs(log, iclog, 0);
3546 ASSERT(iclog->ic_state &
3547 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3552 /*****************************************************************************
3556 *****************************************************************************
3560 * Free a used ticket when its refcount falls to zero.
3564 xlog_ticket_t *ticket)
3566 ASSERT(atomic_read(&ticket->t_ref) > 0);
3567 if (atomic_dec_and_test(&ticket->t_ref))
3568 kmem_zone_free(xfs_log_ticket_zone, ticket);
3573 xlog_ticket_t *ticket)
3575 ASSERT(atomic_read(&ticket->t_ref) > 0);
3576 atomic_inc(&ticket->t_ref);
3581 * Figure out the total log space unit (in bytes) that would be
3582 * required for a log ticket.
3585 xfs_log_calc_unit_res(
3586 struct xfs_mount *mp,
3589 struct xlog *log = mp->m_log;
3594 * Permanent reservations have up to 'cnt'-1 active log operations
3595 * in the log. A unit in this case is the amount of space for one
3596 * of these log operations. Normal reservations have a cnt of 1
3597 * and their unit amount is the total amount of space required.
3599 * The following lines of code account for non-transaction data
3600 * which occupy space in the on-disk log.
3602 * Normal form of a transaction is:
3603 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3604 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3606 * We need to account for all the leadup data and trailer data
3607 * around the transaction data.
3608 * And then we need to account for the worst case in terms of using
3610 * The worst case will happen if:
3611 * - the placement of the transaction happens to be such that the
3612 * roundoff is at its maximum
3613 * - the transaction data is synced before the commit record is synced
3614 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3615 * Therefore the commit record is in its own Log Record.
3616 * This can happen as the commit record is called with its
3617 * own region to xlog_write().
3618 * This then means that in the worst case, roundoff can happen for
3619 * the commit-rec as well.
3620 * The commit-rec is smaller than padding in this scenario and so it is
3621 * not added separately.
3624 /* for trans header */
3625 unit_bytes += sizeof(xlog_op_header_t);
3626 unit_bytes += sizeof(xfs_trans_header_t);
3629 unit_bytes += sizeof(xlog_op_header_t);
3632 * for LR headers - the space for data in an iclog is the size minus
3633 * the space used for the headers. If we use the iclog size, then we
3634 * undercalculate the number of headers required.
3636 * Furthermore - the addition of op headers for split-recs might
3637 * increase the space required enough to require more log and op
3638 * headers, so take that into account too.
3640 * IMPORTANT: This reservation makes the assumption that if this
3641 * transaction is the first in an iclog and hence has the LR headers
3642 * accounted to it, then the remaining space in the iclog is
3643 * exclusively for this transaction. i.e. if the transaction is larger
3644 * than the iclog, it will be the only thing in that iclog.
3645 * Fundamentally, this means we must pass the entire log vector to
3646 * xlog_write to guarantee this.
3648 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3649 num_headers = howmany(unit_bytes, iclog_space);
3651 /* for split-recs - ophdrs added when data split over LRs */
3652 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3654 /* add extra header reservations if we overrun */
3655 while (!num_headers ||
3656 howmany(unit_bytes, iclog_space) > num_headers) {
3657 unit_bytes += sizeof(xlog_op_header_t);
3660 unit_bytes += log->l_iclog_hsize * num_headers;
3662 /* for commit-rec LR header - note: padding will subsume the ophdr */
3663 unit_bytes += log->l_iclog_hsize;
3665 /* for roundoff padding for transaction data and one for commit record */
3666 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3667 /* log su roundoff */
3668 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3671 unit_bytes += 2 * BBSIZE;
3678 * Allocate and initialise a new log ticket.
3680 struct xlog_ticket *
3687 xfs_km_flags_t alloc_flags)
3689 struct xlog_ticket *tic;
3692 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3696 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3698 atomic_set(&tic->t_ref, 1);
3699 tic->t_task = current;
3700 INIT_LIST_HEAD(&tic->t_queue);
3701 tic->t_unit_res = unit_res;
3702 tic->t_curr_res = unit_res;
3705 tic->t_tid = prandom_u32();
3706 tic->t_clientid = client;
3707 tic->t_flags = XLOG_TIC_INITED;
3708 tic->t_trans_type = 0;
3710 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3712 xlog_tic_reset_res(tic);
3718 /******************************************************************************
3720 * Log debug routines
3722 ******************************************************************************
3726 * Make sure that the destination ptr is within the valid data region of
3727 * one of the iclogs. This uses backup pointers stored in a different
3728 * part of the log in case we trash the log structure.
3731 xlog_verify_dest_ptr(
3738 for (i = 0; i < log->l_iclog_bufs; i++) {
3739 if (ptr >= log->l_iclog_bak[i] &&
3740 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3745 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3749 * Check to make sure the grant write head didn't just over lap the tail. If
3750 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3751 * the cycles differ by exactly one and check the byte count.
3753 * This check is run unlocked, so can give false positives. Rather than assert
3754 * on failures, use a warn-once flag and a panic tag to allow the admin to
3755 * determine if they want to panic the machine when such an error occurs. For
3756 * debug kernels this will have the same effect as using an assert but, unlinke
3757 * an assert, it can be turned off at runtime.
3760 xlog_verify_grant_tail(
3763 int tail_cycle, tail_blocks;
3766 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3767 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3768 if (tail_cycle != cycle) {
3769 if (cycle - 1 != tail_cycle &&
3770 !(log->l_flags & XLOG_TAIL_WARN)) {
3771 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3772 "%s: cycle - 1 != tail_cycle", __func__);
3773 log->l_flags |= XLOG_TAIL_WARN;
3776 if (space > BBTOB(tail_blocks) &&
3777 !(log->l_flags & XLOG_TAIL_WARN)) {
3778 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3779 "%s: space > BBTOB(tail_blocks)", __func__);
3780 log->l_flags |= XLOG_TAIL_WARN;
3785 /* check if it will fit */
3787 xlog_verify_tail_lsn(
3789 struct xlog_in_core *iclog,
3794 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3796 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3797 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3798 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3800 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3802 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3803 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3805 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3806 if (blocks < BTOBB(iclog->ic_offset) + 1)
3807 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3809 } /* xlog_verify_tail_lsn */
3812 * Perform a number of checks on the iclog before writing to disk.
3814 * 1. Make sure the iclogs are still circular
3815 * 2. Make sure we have a good magic number
3816 * 3. Make sure we don't have magic numbers in the data
3817 * 4. Check fields of each log operation header for:
3818 * A. Valid client identifier
3819 * B. tid ptr value falls in valid ptr space (user space code)
3820 * C. Length in log record header is correct according to the
3821 * individual operation headers within record.
3822 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3823 * log, check the preceding blocks of the physical log to make sure all
3824 * the cycle numbers agree with the current cycle number.
3829 struct xlog_in_core *iclog,
3833 xlog_op_header_t *ophead;
3834 xlog_in_core_t *icptr;
3835 xlog_in_core_2_t *xhdr;
3836 void *base_ptr, *ptr, *p;
3837 ptrdiff_t field_offset;
3839 int len, i, j, k, op_len;
3842 /* check validity of iclog pointers */
3843 spin_lock(&log->l_icloglock);
3844 icptr = log->l_iclog;
3845 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3848 if (icptr != log->l_iclog)
3849 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3850 spin_unlock(&log->l_icloglock);
3852 /* check log magic numbers */
3853 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3854 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3856 base_ptr = ptr = &iclog->ic_header;
3857 p = &iclog->ic_header;
3858 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3859 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3860 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3865 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3866 base_ptr = ptr = iclog->ic_datap;
3868 xhdr = iclog->ic_data;
3869 for (i = 0; i < len; i++) {
3872 /* clientid is only 1 byte */
3873 p = &ophead->oh_clientid;
3874 field_offset = p - base_ptr;
3875 if (!syncing || (field_offset & 0x1ff)) {
3876 clientid = ophead->oh_clientid;
3878 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3879 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3880 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3881 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3882 clientid = xlog_get_client_id(
3883 xhdr[j].hic_xheader.xh_cycle_data[k]);
3885 clientid = xlog_get_client_id(
3886 iclog->ic_header.h_cycle_data[idx]);
3889 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3891 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3892 __func__, clientid, ophead,
3893 (unsigned long)field_offset);
3896 p = &ophead->oh_len;
3897 field_offset = p - base_ptr;
3898 if (!syncing || (field_offset & 0x1ff)) {
3899 op_len = be32_to_cpu(ophead->oh_len);
3901 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3902 (uintptr_t)iclog->ic_datap);
3903 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3904 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3905 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3906 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3908 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3911 ptr += sizeof(xlog_op_header_t) + op_len;
3913 } /* xlog_verify_iclog */
3917 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3923 xlog_in_core_t *iclog, *ic;
3925 iclog = log->l_iclog;
3926 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3928 * Mark all the incore logs IOERROR.
3929 * From now on, no log flushes will result.
3933 ic->ic_state = XLOG_STATE_IOERROR;
3935 } while (ic != iclog);
3939 * Return non-zero, if state transition has already happened.
3945 * This is called from xfs_force_shutdown, when we're forcibly
3946 * shutting down the filesystem, typically because of an IO error.
3947 * Our main objectives here are to make sure that:
3948 * a. if !logerror, flush the logs to disk. Anything modified
3949 * after this is ignored.
3950 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3951 * parties to find out, 'atomically'.
3952 * c. those who're sleeping on log reservations, pinned objects and
3953 * other resources get woken up, and be told the bad news.
3954 * d. nothing new gets queued up after (b) and (c) are done.
3956 * Note: for the !logerror case we need to flush the regions held in memory out
3957 * to disk first. This needs to be done before the log is marked as shutdown,
3958 * otherwise the iclog writes will fail.
3961 xfs_log_force_umount(
3962 struct xfs_mount *mp,
3971 * If this happens during log recovery, don't worry about
3972 * locking; the log isn't open for business yet.
3975 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3976 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3978 mp->m_sb_bp->b_flags |= XBF_DONE;
3983 * Somebody could've already done the hard work for us.
3984 * No need to get locks for this.
3986 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3987 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3992 * Flush all the completed transactions to disk before marking the log
3993 * being shut down. We need to do it in this order to ensure that
3994 * completed operations are safely on disk before we shut down, and that
3995 * we don't have to issue any buffer IO after the shutdown flags are set
3996 * to guarantee this.
3999 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
4002 * mark the filesystem and the as in a shutdown state and wake
4003 * everybody up to tell them the bad news.
4005 spin_lock(&log->l_icloglock);
4006 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4008 mp->m_sb_bp->b_flags |= XBF_DONE;
4011 * Mark the log and the iclogs with IO error flags to prevent any
4012 * further log IO from being issued or completed.
4014 log->l_flags |= XLOG_IO_ERROR;
4015 retval = xlog_state_ioerror(log);
4016 spin_unlock(&log->l_icloglock);
4019 * We don't want anybody waiting for log reservations after this. That
4020 * means we have to wake up everybody queued up on reserveq as well as
4021 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
4022 * we don't enqueue anything once the SHUTDOWN flag is set, and this
4023 * action is protected by the grant locks.
4025 xlog_grant_head_wake_all(&log->l_reserve_head);
4026 xlog_grant_head_wake_all(&log->l_write_head);
4029 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4030 * as if the log writes were completed. The abort handling in the log
4031 * item committed callback functions will do this again under lock to
4034 wake_up_all(&log->l_cilp->xc_commit_wait);
4035 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4037 #ifdef XFSERRORDEBUG
4039 xlog_in_core_t *iclog;
4041 spin_lock(&log->l_icloglock);
4042 iclog = log->l_iclog;
4044 ASSERT(iclog->ic_callback == 0);
4045 iclog = iclog->ic_next;
4046 } while (iclog != log->l_iclog);
4047 spin_unlock(&log->l_icloglock);
4050 /* return non-zero if log IOERROR transition had already happened */
4058 xlog_in_core_t *iclog;
4060 iclog = log->l_iclog;
4062 /* endianness does not matter here, zero is zero in
4065 if (iclog->ic_header.h_num_logops)
4067 iclog = iclog->ic_next;
4068 } while (iclog != log->l_iclog);
4073 * Verify that an LSN stamped into a piece of metadata is valid. This is
4074 * intended for use in read verifiers on v5 superblocks.
4078 struct xfs_mount *mp,
4081 struct xlog *log = mp->m_log;
4085 * norecovery mode skips mount-time log processing and unconditionally
4086 * resets the in-core LSN. We can't validate in this mode, but
4087 * modifications are not allowed anyways so just return true.
4089 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4093 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4094 * handled by recovery and thus safe to ignore here.
4096 if (lsn == NULLCOMMITLSN)
4099 valid = xlog_valid_lsn(mp->m_log, lsn);
4101 /* warn the user about what's gone wrong before verifier failure */
4103 spin_lock(&log->l_icloglock);
4105 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4106 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4107 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4108 log->l_curr_cycle, log->l_curr_block);
4109 spin_unlock(&log->l_icloglock);