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"
26 #include "xfs_mount.h"
27 #include "xfs_error.h"
28 #include "xfs_trans.h"
29 #include "xfs_trans_priv.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_log_recover.h"
33 #include "xfs_inode.h"
34 #include "xfs_trace.h"
35 #include "xfs_fsops.h"
36 #include "xfs_cksum.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(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(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(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;
529 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
530 flags |= XFS_LOG_REL_PERM_RESERV;
535 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
536 (flags & XFS_LOG_REL_PERM_RESERV)) {
537 trace_xfs_log_done_nonperm(log, ticket);
540 * Release ticket if not permanent reservation or a specific
541 * request has been made to release a permanent reservation.
543 xlog_ungrant_log_space(log, ticket);
544 xfs_log_ticket_put(ticket);
546 trace_xfs_log_done_perm(log, ticket);
548 xlog_regrant_reserve_log_space(log, ticket);
549 /* If this ticket was a permanent reservation and we aren't
550 * trying to release it, reset the inited flags; so next time
551 * we write, a start record will be written out.
553 ticket->t_flags |= XLOG_TIC_INITED;
560 * Attaches a new iclog I/O completion callback routine during
561 * transaction commit. If the log is in error state, a non-zero
562 * return code is handed back and the caller is responsible for
563 * executing the callback at an appropriate time.
567 struct xfs_mount *mp,
568 struct xlog_in_core *iclog,
569 xfs_log_callback_t *cb)
573 spin_lock(&iclog->ic_callback_lock);
574 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
576 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
577 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
579 *(iclog->ic_callback_tail) = cb;
580 iclog->ic_callback_tail = &(cb->cb_next);
582 spin_unlock(&iclog->ic_callback_lock);
587 xfs_log_release_iclog(
588 struct xfs_mount *mp,
589 struct xlog_in_core *iclog)
591 if (xlog_state_release_iclog(mp->m_log, iclog)) {
592 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
600 * Mount a log filesystem
602 * mp - ubiquitous xfs mount point structure
603 * log_target - buftarg of on-disk log device
604 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
605 * num_bblocks - Number of BBSIZE blocks in on-disk log
607 * Return error or zero.
612 xfs_buftarg_t *log_target,
613 xfs_daddr_t blk_offset,
619 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
620 xfs_notice(mp, "Mounting V%d Filesystem",
621 XFS_SB_VERSION_NUM(&mp->m_sb));
624 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
625 XFS_SB_VERSION_NUM(&mp->m_sb));
626 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
629 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
630 if (IS_ERR(mp->m_log)) {
631 error = PTR_ERR(mp->m_log);
636 * Validate the given log space and drop a critical message via syslog
637 * if the log size is too small that would lead to some unexpected
638 * situations in transaction log space reservation stage.
640 * Note: we can't just reject the mount if the validation fails. This
641 * would mean that people would have to downgrade their kernel just to
642 * remedy the situation as there is no way to grow the log (short of
643 * black magic surgery with xfs_db).
645 * We can, however, reject mounts for CRC format filesystems, as the
646 * mkfs binary being used to make the filesystem should never create a
647 * filesystem with a log that is too small.
649 min_logfsbs = xfs_log_calc_minimum_size(mp);
651 if (mp->m_sb.sb_logblocks < min_logfsbs) {
653 "Log size %d blocks too small, minimum size is %d blocks",
654 mp->m_sb.sb_logblocks, min_logfsbs);
656 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
658 "Log size %d blocks too large, maximum size is %lld blocks",
659 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
661 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
663 "log size %lld bytes too large, maximum size is %lld bytes",
664 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
669 if (xfs_sb_version_hascrc(&mp->m_sb)) {
670 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
675 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
676 "experienced then please report this message in the bug report.");
680 * Initialize the AIL now we have a log.
682 error = xfs_trans_ail_init(mp);
684 xfs_warn(mp, "AIL initialisation failed: error %d", error);
687 mp->m_log->l_ailp = mp->m_ail;
690 * skip log recovery on a norecovery mount. pretend it all
693 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
694 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
697 mp->m_flags &= ~XFS_MOUNT_RDONLY;
699 error = xlog_recover(mp->m_log);
702 mp->m_flags |= XFS_MOUNT_RDONLY;
704 xfs_warn(mp, "log mount/recovery failed: error %d",
706 goto out_destroy_ail;
710 /* Normal transactions can now occur */
711 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
714 * Now the log has been fully initialised and we know were our
715 * space grant counters are, we can initialise the permanent ticket
716 * needed for delayed logging to work.
718 xlog_cil_init_post_recovery(mp->m_log);
723 xfs_trans_ail_destroy(mp);
725 xlog_dealloc_log(mp->m_log);
731 * Finish the recovery of the file system. This is separate from the
732 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
733 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
736 * If we finish recovery successfully, start the background log work. If we are
737 * not doing recovery, then we have a RO filesystem and we don't need to start
741 xfs_log_mount_finish(xfs_mount_t *mp)
745 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
746 error = xlog_recover_finish(mp->m_log);
748 xfs_log_work_queue(mp);
750 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
758 * Final log writes as part of unmount.
760 * Mark the filesystem clean as unmount happens. Note that during relocation
761 * this routine needs to be executed as part of source-bag while the
762 * deallocation must not be done until source-end.
766 * Unmount record used to have a string "Unmount filesystem--" in the
767 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
768 * We just write the magic number now since that particular field isn't
769 * currently architecture converted and "Unmount" is a bit foo.
770 * As far as I know, there weren't any dependencies on the old behaviour.
774 xfs_log_unmount_write(xfs_mount_t *mp)
776 struct xlog *log = mp->m_log;
777 xlog_in_core_t *iclog;
779 xlog_in_core_t *first_iclog;
781 xlog_ticket_t *tic = NULL;
786 * Don't write out unmount record on read-only mounts.
787 * Or, if we are doing a forced umount (typically because of IO errors).
789 if (mp->m_flags & XFS_MOUNT_RDONLY)
792 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
793 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
796 first_iclog = iclog = log->l_iclog;
798 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
799 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
800 ASSERT(iclog->ic_offset == 0);
802 iclog = iclog->ic_next;
803 } while (iclog != first_iclog);
805 if (! (XLOG_FORCED_SHUTDOWN(log))) {
806 error = xfs_log_reserve(mp, 600, 1, &tic,
807 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
809 /* the data section must be 32 bit size aligned */
813 __uint32_t pad2; /* may as well make it 64 bits */
815 .magic = XLOG_UNMOUNT_TYPE,
817 struct xfs_log_iovec reg = {
819 .i_len = sizeof(magic),
820 .i_type = XLOG_REG_TYPE_UNMOUNT,
822 struct xfs_log_vec vec = {
827 /* remove inited flag, and account for space used */
829 tic->t_curr_res -= sizeof(magic);
830 error = xlog_write(log, &vec, tic, &lsn,
831 NULL, XLOG_UNMOUNT_TRANS);
833 * At this point, we're umounting anyway,
834 * so there's no point in transitioning log state
835 * to IOERROR. Just continue...
840 xfs_alert(mp, "%s: unmount record failed", __func__);
843 spin_lock(&log->l_icloglock);
844 iclog = log->l_iclog;
845 atomic_inc(&iclog->ic_refcnt);
846 xlog_state_want_sync(log, iclog);
847 spin_unlock(&log->l_icloglock);
848 error = xlog_state_release_iclog(log, iclog);
850 spin_lock(&log->l_icloglock);
851 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
852 iclog->ic_state == XLOG_STATE_DIRTY)) {
853 if (!XLOG_FORCED_SHUTDOWN(log)) {
854 xlog_wait(&iclog->ic_force_wait,
857 spin_unlock(&log->l_icloglock);
860 spin_unlock(&log->l_icloglock);
863 trace_xfs_log_umount_write(log, tic);
864 xlog_ungrant_log_space(log, tic);
865 xfs_log_ticket_put(tic);
869 * We're already in forced_shutdown mode, couldn't
870 * even attempt to write out the unmount transaction.
872 * Go through the motions of sync'ing and releasing
873 * the iclog, even though no I/O will actually happen,
874 * we need to wait for other log I/Os that may already
875 * be in progress. Do this as a separate section of
876 * code so we'll know if we ever get stuck here that
877 * we're in this odd situation of trying to unmount
878 * a file system that went into forced_shutdown as
879 * the result of an unmount..
881 spin_lock(&log->l_icloglock);
882 iclog = log->l_iclog;
883 atomic_inc(&iclog->ic_refcnt);
885 xlog_state_want_sync(log, iclog);
886 spin_unlock(&log->l_icloglock);
887 error = xlog_state_release_iclog(log, iclog);
889 spin_lock(&log->l_icloglock);
891 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
892 || iclog->ic_state == XLOG_STATE_DIRTY
893 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
895 xlog_wait(&iclog->ic_force_wait,
898 spin_unlock(&log->l_icloglock);
903 } /* xfs_log_unmount_write */
906 * Empty the log for unmount/freeze.
908 * To do this, we first need to shut down the background log work so it is not
909 * trying to cover the log as we clean up. We then need to unpin all objects in
910 * the log so we can then flush them out. Once they have completed their IO and
911 * run the callbacks removing themselves from the AIL, we can write the unmount
916 struct xfs_mount *mp)
918 cancel_delayed_work_sync(&mp->m_log->l_work);
919 xfs_log_force(mp, XFS_LOG_SYNC);
922 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
923 * will push it, xfs_wait_buftarg() will not wait for it. Further,
924 * xfs_buf_iowait() cannot be used because it was pushed with the
925 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
926 * the IO to complete.
928 xfs_ail_push_all_sync(mp->m_ail);
929 xfs_wait_buftarg(mp->m_ddev_targp);
930 xfs_buf_lock(mp->m_sb_bp);
931 xfs_buf_unlock(mp->m_sb_bp);
933 xfs_log_unmount_write(mp);
937 * Shut down and release the AIL and Log.
939 * During unmount, we need to ensure we flush all the dirty metadata objects
940 * from the AIL so that the log is empty before we write the unmount record to
941 * the log. Once this is done, we can tear down the AIL and the log.
945 struct xfs_mount *mp)
949 xfs_trans_ail_destroy(mp);
950 xlog_dealloc_log(mp->m_log);
955 struct xfs_mount *mp,
956 struct xfs_log_item *item,
958 const struct xfs_item_ops *ops)
960 item->li_mountp = mp;
961 item->li_ailp = mp->m_ail;
962 item->li_type = type;
966 INIT_LIST_HEAD(&item->li_ail);
967 INIT_LIST_HEAD(&item->li_cil);
971 * Wake up processes waiting for log space after we have moved the log tail.
975 struct xfs_mount *mp)
977 struct xlog *log = mp->m_log;
980 if (XLOG_FORCED_SHUTDOWN(log))
983 if (!list_empty_careful(&log->l_write_head.waiters)) {
984 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
986 spin_lock(&log->l_write_head.lock);
987 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
988 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
989 spin_unlock(&log->l_write_head.lock);
992 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
993 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
995 spin_lock(&log->l_reserve_head.lock);
996 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
997 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
998 spin_unlock(&log->l_reserve_head.lock);
1003 * Determine if we have a transaction that has gone to disk that needs to be
1004 * covered. To begin the transition to the idle state firstly the log needs to
1005 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1006 * we start attempting to cover the log.
1008 * Only if we are then in a state where covering is needed, the caller is
1009 * informed that dummy transactions are required to move the log into the idle
1012 * If there are any items in the AIl or CIL, then we do not want to attempt to
1013 * cover the log as we may be in a situation where there isn't log space
1014 * available to run a dummy transaction and this can lead to deadlocks when the
1015 * tail of the log is pinned by an item that is modified in the CIL. Hence
1016 * there's no point in running a dummy transaction at this point because we
1017 * can't start trying to idle the log until both the CIL and AIL are empty.
1020 xfs_log_need_covered(xfs_mount_t *mp)
1022 struct xlog *log = mp->m_log;
1025 if (!xfs_fs_writable(mp))
1028 if (!xlog_cil_empty(log))
1031 spin_lock(&log->l_icloglock);
1032 switch (log->l_covered_state) {
1033 case XLOG_STATE_COVER_DONE:
1034 case XLOG_STATE_COVER_DONE2:
1035 case XLOG_STATE_COVER_IDLE:
1037 case XLOG_STATE_COVER_NEED:
1038 case XLOG_STATE_COVER_NEED2:
1039 if (xfs_ail_min_lsn(log->l_ailp))
1041 if (!xlog_iclogs_empty(log))
1045 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1046 log->l_covered_state = XLOG_STATE_COVER_DONE;
1048 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1054 spin_unlock(&log->l_icloglock);
1059 * We may be holding the log iclog lock upon entering this routine.
1062 xlog_assign_tail_lsn_locked(
1063 struct xfs_mount *mp)
1065 struct xlog *log = mp->m_log;
1066 struct xfs_log_item *lip;
1069 assert_spin_locked(&mp->m_ail->xa_lock);
1072 * To make sure we always have a valid LSN for the log tail we keep
1073 * track of the last LSN which was committed in log->l_last_sync_lsn,
1074 * and use that when the AIL was empty.
1076 lip = xfs_ail_min(mp->m_ail);
1078 tail_lsn = lip->li_lsn;
1080 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1081 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1082 atomic64_set(&log->l_tail_lsn, tail_lsn);
1087 xlog_assign_tail_lsn(
1088 struct xfs_mount *mp)
1092 spin_lock(&mp->m_ail->xa_lock);
1093 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1094 spin_unlock(&mp->m_ail->xa_lock);
1100 * Return the space in the log between the tail and the head. The head
1101 * is passed in the cycle/bytes formal parms. In the special case where
1102 * the reserve head has wrapped passed the tail, this calculation is no
1103 * longer valid. In this case, just return 0 which means there is no space
1104 * in the log. This works for all places where this function is called
1105 * with the reserve head. Of course, if the write head were to ever
1106 * wrap the tail, we should blow up. Rather than catch this case here,
1107 * we depend on other ASSERTions in other parts of the code. XXXmiken
1109 * This code also handles the case where the reservation head is behind
1110 * the tail. The details of this case are described below, but the end
1111 * result is that we return the size of the log as the amount of space left.
1124 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1125 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1126 tail_bytes = BBTOB(tail_bytes);
1127 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1128 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1129 else if (tail_cycle + 1 < head_cycle)
1131 else if (tail_cycle < head_cycle) {
1132 ASSERT(tail_cycle == (head_cycle - 1));
1133 free_bytes = tail_bytes - head_bytes;
1136 * The reservation head is behind the tail.
1137 * In this case we just want to return the size of the
1138 * log as the amount of space left.
1140 xfs_alert(log->l_mp,
1141 "xlog_space_left: head behind tail\n"
1142 " tail_cycle = %d, tail_bytes = %d\n"
1143 " GH cycle = %d, GH bytes = %d",
1144 tail_cycle, tail_bytes, head_cycle, head_bytes);
1146 free_bytes = log->l_logsize;
1153 * Log function which is called when an io completes.
1155 * The log manager needs its own routine, in order to control what
1156 * happens with the buffer after the write completes.
1159 xlog_iodone(xfs_buf_t *bp)
1161 struct xlog_in_core *iclog = bp->b_fspriv;
1162 struct xlog *l = iclog->ic_log;
1166 * Race to shutdown the filesystem if we see an error.
1168 if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1169 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1170 xfs_buf_ioerror_alert(bp, __func__);
1172 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1174 * This flag will be propagated to the trans-committed
1175 * callback routines to let them know that the log-commit
1178 aborted = XFS_LI_ABORTED;
1179 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1180 aborted = XFS_LI_ABORTED;
1183 /* log I/O is always issued ASYNC */
1184 ASSERT(XFS_BUF_ISASYNC(bp));
1185 xlog_state_done_syncing(iclog, aborted);
1188 * drop the buffer lock now that we are done. Nothing references
1189 * the buffer after this, so an unmount waiting on this lock can now
1190 * tear it down safely. As such, it is unsafe to reference the buffer
1191 * (bp) after the unlock as we could race with it being freed.
1197 * Return size of each in-core log record buffer.
1199 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1201 * If the filesystem blocksize is too large, we may need to choose a
1202 * larger size since the directory code currently logs entire blocks.
1206 xlog_get_iclog_buffer_size(
1207 struct xfs_mount *mp,
1213 if (mp->m_logbufs <= 0)
1214 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1216 log->l_iclog_bufs = mp->m_logbufs;
1219 * Buffer size passed in from mount system call.
1221 if (mp->m_logbsize > 0) {
1222 size = log->l_iclog_size = mp->m_logbsize;
1223 log->l_iclog_size_log = 0;
1225 log->l_iclog_size_log++;
1229 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1230 /* # headers = size / 32k
1231 * one header holds cycles from 32k of data
1234 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1235 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1237 log->l_iclog_hsize = xhdrs << BBSHIFT;
1238 log->l_iclog_heads = xhdrs;
1240 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1241 log->l_iclog_hsize = BBSIZE;
1242 log->l_iclog_heads = 1;
1247 /* All machines use 32kB buffers by default. */
1248 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1249 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1251 /* the default log size is 16k or 32k which is one header sector */
1252 log->l_iclog_hsize = BBSIZE;
1253 log->l_iclog_heads = 1;
1256 /* are we being asked to make the sizes selected above visible? */
1257 if (mp->m_logbufs == 0)
1258 mp->m_logbufs = log->l_iclog_bufs;
1259 if (mp->m_logbsize == 0)
1260 mp->m_logbsize = log->l_iclog_size;
1261 } /* xlog_get_iclog_buffer_size */
1266 struct xfs_mount *mp)
1268 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1269 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1273 * Every sync period we need to unpin all items in the AIL and push them to
1274 * disk. If there is nothing dirty, then we might need to cover the log to
1275 * indicate that the filesystem is idle.
1279 struct work_struct *work)
1281 struct xlog *log = container_of(to_delayed_work(work),
1282 struct xlog, l_work);
1283 struct xfs_mount *mp = log->l_mp;
1285 /* dgc: errors ignored - not fatal and nowhere to report them */
1286 if (xfs_log_need_covered(mp))
1287 xfs_fs_log_dummy(mp);
1289 xfs_log_force(mp, 0);
1291 /* start pushing all the metadata that is currently dirty */
1292 xfs_ail_push_all(mp->m_ail);
1294 /* queue us up again */
1295 xfs_log_work_queue(mp);
1299 * This routine initializes some of the log structure for a given mount point.
1300 * Its primary purpose is to fill in enough, so recovery can occur. However,
1301 * some other stuff may be filled in too.
1303 STATIC struct xlog *
1305 struct xfs_mount *mp,
1306 struct xfs_buftarg *log_target,
1307 xfs_daddr_t blk_offset,
1311 xlog_rec_header_t *head;
1312 xlog_in_core_t **iclogp;
1313 xlog_in_core_t *iclog, *prev_iclog=NULL;
1316 int error = -ENOMEM;
1319 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1321 xfs_warn(mp, "Log allocation failed: No memory!");
1326 log->l_targ = log_target;
1327 log->l_logsize = BBTOB(num_bblks);
1328 log->l_logBBstart = blk_offset;
1329 log->l_logBBsize = num_bblks;
1330 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1331 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1332 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1334 log->l_prev_block = -1;
1335 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1336 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1337 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1338 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1340 xlog_grant_head_init(&log->l_reserve_head);
1341 xlog_grant_head_init(&log->l_write_head);
1343 error = -EFSCORRUPTED;
1344 if (xfs_sb_version_hassector(&mp->m_sb)) {
1345 log2_size = mp->m_sb.sb_logsectlog;
1346 if (log2_size < BBSHIFT) {
1347 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1348 log2_size, BBSHIFT);
1352 log2_size -= BBSHIFT;
1353 if (log2_size > mp->m_sectbb_log) {
1354 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1355 log2_size, mp->m_sectbb_log);
1359 /* for larger sector sizes, must have v2 or external log */
1360 if (log2_size && log->l_logBBstart > 0 &&
1361 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1363 "log sector size (0x%x) invalid for configuration.",
1368 log->l_sectBBsize = 1 << log2_size;
1370 xlog_get_iclog_buffer_size(mp, log);
1373 * Use a NULL block for the extra log buffer used during splits so that
1374 * it will trigger errors if we ever try to do IO on it without first
1375 * having set it up properly.
1378 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1379 BTOBB(log->l_iclog_size), 0);
1384 * The iclogbuf buffer locks are held over IO but we are not going to do
1385 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1386 * when appropriately.
1388 ASSERT(xfs_buf_islocked(bp));
1391 bp->b_iodone = xlog_iodone;
1394 spin_lock_init(&log->l_icloglock);
1395 init_waitqueue_head(&log->l_flush_wait);
1397 iclogp = &log->l_iclog;
1399 * The amount of memory to allocate for the iclog structure is
1400 * rather funky due to the way the structure is defined. It is
1401 * done this way so that we can use different sizes for machines
1402 * with different amounts of memory. See the definition of
1403 * xlog_in_core_t in xfs_log_priv.h for details.
1405 ASSERT(log->l_iclog_size >= 4096);
1406 for (i=0; i < log->l_iclog_bufs; i++) {
1407 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1409 goto out_free_iclog;
1412 iclog->ic_prev = prev_iclog;
1415 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1416 BTOBB(log->l_iclog_size), 0);
1418 goto out_free_iclog;
1420 ASSERT(xfs_buf_islocked(bp));
1423 bp->b_iodone = xlog_iodone;
1425 iclog->ic_data = bp->b_addr;
1427 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1429 head = &iclog->ic_header;
1430 memset(head, 0, sizeof(xlog_rec_header_t));
1431 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1432 head->h_version = cpu_to_be32(
1433 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1434 head->h_size = cpu_to_be32(log->l_iclog_size);
1436 head->h_fmt = cpu_to_be32(XLOG_FMT);
1437 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1439 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1440 iclog->ic_state = XLOG_STATE_ACTIVE;
1441 iclog->ic_log = log;
1442 atomic_set(&iclog->ic_refcnt, 0);
1443 spin_lock_init(&iclog->ic_callback_lock);
1444 iclog->ic_callback_tail = &(iclog->ic_callback);
1445 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1447 init_waitqueue_head(&iclog->ic_force_wait);
1448 init_waitqueue_head(&iclog->ic_write_wait);
1450 iclogp = &iclog->ic_next;
1452 *iclogp = log->l_iclog; /* complete ring */
1453 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1455 error = xlog_cil_init(log);
1457 goto out_free_iclog;
1461 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1462 prev_iclog = iclog->ic_next;
1464 xfs_buf_free(iclog->ic_bp);
1467 spinlock_destroy(&log->l_icloglock);
1468 xfs_buf_free(log->l_xbuf);
1472 return ERR_PTR(error);
1473 } /* xlog_alloc_log */
1477 * Write out the commit record of a transaction associated with the given
1478 * ticket. Return the lsn of the commit record.
1483 struct xlog_ticket *ticket,
1484 struct xlog_in_core **iclog,
1485 xfs_lsn_t *commitlsnp)
1487 struct xfs_mount *mp = log->l_mp;
1489 struct xfs_log_iovec reg = {
1492 .i_type = XLOG_REG_TYPE_COMMIT,
1494 struct xfs_log_vec vec = {
1499 ASSERT_ALWAYS(iclog);
1500 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1503 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1508 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1509 * log space. This code pushes on the lsn which would supposedly free up
1510 * the 25% which we want to leave free. We may need to adopt a policy which
1511 * pushes on an lsn which is further along in the log once we reach the high
1512 * water mark. In this manner, we would be creating a low water mark.
1515 xlog_grant_push_ail(
1519 xfs_lsn_t threshold_lsn = 0;
1520 xfs_lsn_t last_sync_lsn;
1523 int threshold_block;
1524 int threshold_cycle;
1527 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1529 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1530 free_blocks = BTOBBT(free_bytes);
1533 * Set the threshold for the minimum number of free blocks in the
1534 * log to the maximum of what the caller needs, one quarter of the
1535 * log, and 256 blocks.
1537 free_threshold = BTOBB(need_bytes);
1538 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1539 free_threshold = MAX(free_threshold, 256);
1540 if (free_blocks >= free_threshold)
1543 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1545 threshold_block += free_threshold;
1546 if (threshold_block >= log->l_logBBsize) {
1547 threshold_block -= log->l_logBBsize;
1548 threshold_cycle += 1;
1550 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1553 * Don't pass in an lsn greater than the lsn of the last
1554 * log record known to be on disk. Use a snapshot of the last sync lsn
1555 * so that it doesn't change between the compare and the set.
1557 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1558 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1559 threshold_lsn = last_sync_lsn;
1562 * Get the transaction layer to kick the dirty buffers out to
1563 * disk asynchronously. No point in trying to do this if
1564 * the filesystem is shutting down.
1566 if (!XLOG_FORCED_SHUTDOWN(log))
1567 xfs_ail_push(log->l_ailp, threshold_lsn);
1571 * Stamp cycle number in every block
1576 struct xlog_in_core *iclog,
1580 int size = iclog->ic_offset + roundoff;
1584 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1586 dp = iclog->ic_datap;
1587 for (i = 0; i < BTOBB(size); i++) {
1588 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1590 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1591 *(__be32 *)dp = cycle_lsn;
1595 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1596 xlog_in_core_2_t *xhdr = iclog->ic_data;
1598 for ( ; i < BTOBB(size); i++) {
1599 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1600 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1601 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1602 *(__be32 *)dp = cycle_lsn;
1606 for (i = 1; i < log->l_iclog_heads; i++)
1607 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1612 * Calculate the checksum for a log buffer.
1614 * This is a little more complicated than it should be because the various
1615 * headers and the actual data are non-contiguous.
1620 struct xlog_rec_header *rhead,
1626 /* first generate the crc for the record header ... */
1627 crc = xfs_start_cksum((char *)rhead,
1628 sizeof(struct xlog_rec_header),
1629 offsetof(struct xlog_rec_header, h_crc));
1631 /* ... then for additional cycle data for v2 logs ... */
1632 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1633 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1636 for (i = 1; i < log->l_iclog_heads; i++) {
1637 crc = crc32c(crc, &xhdr[i].hic_xheader,
1638 sizeof(struct xlog_rec_ext_header));
1642 /* ... and finally for the payload */
1643 crc = crc32c(crc, dp, size);
1645 return xfs_end_cksum(crc);
1649 * The bdstrat callback function for log bufs. This gives us a central
1650 * place to trap bufs in case we get hit by a log I/O error and need to
1651 * shutdown. Actually, in practice, even when we didn't get a log error,
1652 * we transition the iclogs to IOERROR state *after* flushing all existing
1653 * iclogs to disk. This is because we don't want anymore new transactions to be
1654 * started or completed afterwards.
1656 * We lock the iclogbufs here so that we can serialise against IO completion
1657 * during unmount. We might be processing a shutdown triggered during unmount,
1658 * and that can occur asynchronously to the unmount thread, and hence we need to
1659 * ensure that completes before tearing down the iclogbufs. Hence we need to
1660 * hold the buffer lock across the log IO to acheive that.
1666 struct xlog_in_core *iclog = bp->b_fspriv;
1669 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1670 xfs_buf_ioerror(bp, -EIO);
1672 xfs_buf_ioend(bp, 0);
1674 * It would seem logical to return EIO here, but we rely on
1675 * the log state machine to propagate I/O errors instead of
1676 * doing it here. Similarly, IO completion will unlock the
1677 * buffer, so we don't do it here.
1682 xfs_buf_iorequest(bp);
1687 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1688 * fashion. Previously, we should have moved the current iclog
1689 * ptr in the log to point to the next available iclog. This allows further
1690 * write to continue while this code syncs out an iclog ready to go.
1691 * Before an in-core log can be written out, the data section must be scanned
1692 * to save away the 1st word of each BBSIZE block into the header. We replace
1693 * it with the current cycle count. Each BBSIZE block is tagged with the
1694 * cycle count because there in an implicit assumption that drives will
1695 * guarantee that entire 512 byte blocks get written at once. In other words,
1696 * we can't have part of a 512 byte block written and part not written. By
1697 * tagging each block, we will know which blocks are valid when recovering
1698 * after an unclean shutdown.
1700 * This routine is single threaded on the iclog. No other thread can be in
1701 * this routine with the same iclog. Changing contents of iclog can there-
1702 * fore be done without grabbing the state machine lock. Updating the global
1703 * log will require grabbing the lock though.
1705 * The entire log manager uses a logical block numbering scheme. Only
1706 * log_sync (and then only bwrite()) know about the fact that the log may
1707 * not start with block zero on a given device. The log block start offset
1708 * is added immediately before calling bwrite().
1714 struct xlog_in_core *iclog)
1718 uint count; /* byte count of bwrite */
1719 uint count_init; /* initial count before roundup */
1720 int roundoff; /* roundoff to BB or stripe */
1721 int split = 0; /* split write into two regions */
1723 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1726 XFS_STATS_INC(xs_log_writes);
1727 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1729 /* Add for LR header */
1730 count_init = log->l_iclog_hsize + iclog->ic_offset;
1732 /* Round out the log write size */
1733 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1734 /* we have a v2 stripe unit to use */
1735 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1737 count = BBTOB(BTOBB(count_init));
1739 roundoff = count - count_init;
1740 ASSERT(roundoff >= 0);
1741 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1742 roundoff < log->l_mp->m_sb.sb_logsunit)
1744 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1745 roundoff < BBTOB(1)));
1747 /* move grant heads by roundoff in sync */
1748 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1749 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1751 /* put cycle number in every block */
1752 xlog_pack_data(log, iclog, roundoff);
1754 /* real byte length */
1755 size = iclog->ic_offset;
1758 iclog->ic_header.h_len = cpu_to_be32(size);
1761 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1763 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1765 /* Do we need to split this write into 2 parts? */
1766 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1769 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1770 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1771 iclog->ic_bwritecnt = 2;
1774 * Bump the cycle numbers at the start of each block in the
1775 * part of the iclog that ends up in the buffer that gets
1776 * written to the start of the log.
1778 * Watch out for the header magic number case, though.
1780 dptr = (char *)&iclog->ic_header + count;
1781 for (i = 0; i < split; i += BBSIZE) {
1782 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1783 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1785 *(__be32 *)dptr = cpu_to_be32(cycle);
1790 iclog->ic_bwritecnt = 1;
1793 /* calculcate the checksum */
1794 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1795 iclog->ic_datap, size);
1797 bp->b_io_length = BTOBB(count);
1798 bp->b_fspriv = iclog;
1799 XFS_BUF_ZEROFLAGS(bp);
1801 bp->b_flags |= XBF_SYNCIO;
1803 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1804 bp->b_flags |= XBF_FUA;
1807 * Flush the data device before flushing the log to make
1808 * sure all meta data written back from the AIL actually made
1809 * it to disk before stamping the new log tail LSN into the
1810 * log buffer. For an external log we need to issue the
1811 * flush explicitly, and unfortunately synchronously here;
1812 * for an internal log we can simply use the block layer
1813 * state machine for preflushes.
1815 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1816 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1818 bp->b_flags |= XBF_FLUSH;
1821 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1822 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1824 xlog_verify_iclog(log, iclog, count, true);
1826 /* account for log which doesn't start at block #0 */
1827 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1829 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1834 error = xlog_bdstrat(bp);
1836 xfs_buf_ioerror_alert(bp, "xlog_sync");
1840 bp = iclog->ic_log->l_xbuf;
1841 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1842 xfs_buf_associate_memory(bp,
1843 (char *)&iclog->ic_header + count, split);
1844 bp->b_fspriv = iclog;
1845 XFS_BUF_ZEROFLAGS(bp);
1847 bp->b_flags |= XBF_SYNCIO;
1848 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1849 bp->b_flags |= XBF_FUA;
1851 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1852 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1854 /* account for internal log which doesn't start at block #0 */
1855 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1857 error = xlog_bdstrat(bp);
1859 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1867 * Deallocate a log structure
1873 xlog_in_core_t *iclog, *next_iclog;
1876 xlog_cil_destroy(log);
1879 * Cycle all the iclogbuf locks to make sure all log IO completion
1880 * is done before we tear down these buffers.
1882 iclog = log->l_iclog;
1883 for (i = 0; i < log->l_iclog_bufs; i++) {
1884 xfs_buf_lock(iclog->ic_bp);
1885 xfs_buf_unlock(iclog->ic_bp);
1886 iclog = iclog->ic_next;
1890 * Always need to ensure that the extra buffer does not point to memory
1891 * owned by another log buffer before we free it. Also, cycle the lock
1892 * first to ensure we've completed IO on it.
1894 xfs_buf_lock(log->l_xbuf);
1895 xfs_buf_unlock(log->l_xbuf);
1896 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1897 xfs_buf_free(log->l_xbuf);
1899 iclog = log->l_iclog;
1900 for (i = 0; i < log->l_iclog_bufs; i++) {
1901 xfs_buf_free(iclog->ic_bp);
1902 next_iclog = iclog->ic_next;
1906 spinlock_destroy(&log->l_icloglock);
1908 log->l_mp->m_log = NULL;
1910 } /* xlog_dealloc_log */
1913 * Update counters atomically now that memcpy is done.
1917 xlog_state_finish_copy(
1919 struct xlog_in_core *iclog,
1923 spin_lock(&log->l_icloglock);
1925 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1926 iclog->ic_offset += copy_bytes;
1928 spin_unlock(&log->l_icloglock);
1929 } /* xlog_state_finish_copy */
1935 * print out info relating to regions written which consume
1940 struct xfs_mount *mp,
1941 struct xlog_ticket *ticket)
1944 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1946 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1947 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1968 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2012 "xlog_write: reservation summary:\n"
2013 " trans type = %s (%u)\n"
2014 " unit res = %d bytes\n"
2015 " current res = %d bytes\n"
2016 " total reg = %u bytes (o/flow = %u bytes)\n"
2017 " ophdrs = %u (ophdr space = %u bytes)\n"
2018 " ophdr + reg = %u bytes\n"
2019 " num regions = %u\n",
2020 ((ticket->t_trans_type <= 0 ||
2021 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2022 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2023 ticket->t_trans_type,
2026 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2027 ticket->t_res_num_ophdrs, ophdr_spc,
2028 ticket->t_res_arr_sum +
2029 ticket->t_res_o_flow + ophdr_spc,
2032 for (i = 0; i < ticket->t_res_num; i++) {
2033 uint r_type = ticket->t_res_arr[i].r_type;
2034 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2035 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2036 "bad-rtype" : res_type_str[r_type-1]),
2037 ticket->t_res_arr[i].r_len);
2040 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2041 "xlog_write: reservation ran out. Need to up reservation");
2042 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2046 * Calculate the potential space needed by the log vector. Each region gets
2047 * its own xlog_op_header_t and may need to be double word aligned.
2050 xlog_write_calc_vec_length(
2051 struct xlog_ticket *ticket,
2052 struct xfs_log_vec *log_vector)
2054 struct xfs_log_vec *lv;
2059 /* acct for start rec of xact */
2060 if (ticket->t_flags & XLOG_TIC_INITED)
2063 for (lv = log_vector; lv; lv = lv->lv_next) {
2064 /* we don't write ordered log vectors */
2065 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2068 headers += lv->lv_niovecs;
2070 for (i = 0; i < lv->lv_niovecs; i++) {
2071 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2074 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2078 ticket->t_res_num_ophdrs += headers;
2079 len += headers * sizeof(struct xlog_op_header);
2085 * If first write for transaction, insert start record We can't be trying to
2086 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2089 xlog_write_start_rec(
2090 struct xlog_op_header *ophdr,
2091 struct xlog_ticket *ticket)
2093 if (!(ticket->t_flags & XLOG_TIC_INITED))
2096 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2097 ophdr->oh_clientid = ticket->t_clientid;
2099 ophdr->oh_flags = XLOG_START_TRANS;
2102 ticket->t_flags &= ~XLOG_TIC_INITED;
2104 return sizeof(struct xlog_op_header);
2107 static xlog_op_header_t *
2108 xlog_write_setup_ophdr(
2110 struct xlog_op_header *ophdr,
2111 struct xlog_ticket *ticket,
2114 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2115 ophdr->oh_clientid = ticket->t_clientid;
2118 /* are we copying a commit or unmount record? */
2119 ophdr->oh_flags = flags;
2122 * We've seen logs corrupted with bad transaction client ids. This
2123 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2124 * and shut down the filesystem.
2126 switch (ophdr->oh_clientid) {
2127 case XFS_TRANSACTION:
2133 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2134 ophdr->oh_clientid, ticket);
2142 * Set up the parameters of the region copy into the log. This has
2143 * to handle region write split across multiple log buffers - this
2144 * state is kept external to this function so that this code can
2145 * be written in an obvious, self documenting manner.
2148 xlog_write_setup_copy(
2149 struct xlog_ticket *ticket,
2150 struct xlog_op_header *ophdr,
2151 int space_available,
2155 int *last_was_partial_copy,
2156 int *bytes_consumed)
2160 still_to_copy = space_required - *bytes_consumed;
2161 *copy_off = *bytes_consumed;
2163 if (still_to_copy <= space_available) {
2164 /* write of region completes here */
2165 *copy_len = still_to_copy;
2166 ophdr->oh_len = cpu_to_be32(*copy_len);
2167 if (*last_was_partial_copy)
2168 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2169 *last_was_partial_copy = 0;
2170 *bytes_consumed = 0;
2174 /* partial write of region, needs extra log op header reservation */
2175 *copy_len = space_available;
2176 ophdr->oh_len = cpu_to_be32(*copy_len);
2177 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2178 if (*last_was_partial_copy)
2179 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2180 *bytes_consumed += *copy_len;
2181 (*last_was_partial_copy)++;
2183 /* account for new log op header */
2184 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2185 ticket->t_res_num_ophdrs++;
2187 return sizeof(struct xlog_op_header);
2191 xlog_write_copy_finish(
2193 struct xlog_in_core *iclog,
2198 int *partial_copy_len,
2200 struct xlog_in_core **commit_iclog)
2202 if (*partial_copy) {
2204 * This iclog has already been marked WANT_SYNC by
2205 * xlog_state_get_iclog_space.
2207 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2210 return xlog_state_release_iclog(log, iclog);
2214 *partial_copy_len = 0;
2216 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2217 /* no more space in this iclog - push it. */
2218 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2222 spin_lock(&log->l_icloglock);
2223 xlog_state_want_sync(log, iclog);
2224 spin_unlock(&log->l_icloglock);
2227 return xlog_state_release_iclog(log, iclog);
2228 ASSERT(flags & XLOG_COMMIT_TRANS);
2229 *commit_iclog = iclog;
2236 * Write some region out to in-core log
2238 * This will be called when writing externally provided regions or when
2239 * writing out a commit record for a given transaction.
2241 * General algorithm:
2242 * 1. Find total length of this write. This may include adding to the
2243 * lengths passed in.
2244 * 2. Check whether we violate the tickets reservation.
2245 * 3. While writing to this iclog
2246 * A. Reserve as much space in this iclog as can get
2247 * B. If this is first write, save away start lsn
2248 * C. While writing this region:
2249 * 1. If first write of transaction, write start record
2250 * 2. Write log operation header (header per region)
2251 * 3. Find out if we can fit entire region into this iclog
2252 * 4. Potentially, verify destination memcpy ptr
2253 * 5. Memcpy (partial) region
2254 * 6. If partial copy, release iclog; otherwise, continue
2255 * copying more regions into current iclog
2256 * 4. Mark want sync bit (in simulation mode)
2257 * 5. Release iclog for potential flush to on-disk log.
2260 * 1. Panic if reservation is overrun. This should never happen since
2261 * reservation amounts are generated internal to the filesystem.
2263 * 1. Tickets are single threaded data structures.
2264 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2265 * syncing routine. When a single log_write region needs to span
2266 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2267 * on all log operation writes which don't contain the end of the
2268 * region. The XLOG_END_TRANS bit is used for the in-core log
2269 * operation which contains the end of the continued log_write region.
2270 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2271 * we don't really know exactly how much space will be used. As a result,
2272 * we don't update ic_offset until the end when we know exactly how many
2273 * bytes have been written out.
2278 struct xfs_log_vec *log_vector,
2279 struct xlog_ticket *ticket,
2280 xfs_lsn_t *start_lsn,
2281 struct xlog_in_core **commit_iclog,
2284 struct xlog_in_core *iclog = NULL;
2285 struct xfs_log_iovec *vecp;
2286 struct xfs_log_vec *lv;
2289 int partial_copy = 0;
2290 int partial_copy_len = 0;
2298 len = xlog_write_calc_vec_length(ticket, log_vector);
2301 * Region headers and bytes are already accounted for.
2302 * We only need to take into account start records and
2303 * split regions in this function.
2305 if (ticket->t_flags & XLOG_TIC_INITED)
2306 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2309 * Commit record headers need to be accounted for. These
2310 * come in as separate writes so are easy to detect.
2312 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2313 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2315 if (ticket->t_curr_res < 0)
2316 xlog_print_tic_res(log->l_mp, ticket);
2320 vecp = lv->lv_iovecp;
2321 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2325 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2326 &contwr, &log_offset);
2330 ASSERT(log_offset <= iclog->ic_size - 1);
2331 ptr = iclog->ic_datap + log_offset;
2333 /* start_lsn is the first lsn written to. That's all we need. */
2335 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2338 * This loop writes out as many regions as can fit in the amount
2339 * of space which was allocated by xlog_state_get_iclog_space().
2341 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2342 struct xfs_log_iovec *reg;
2343 struct xlog_op_header *ophdr;
2347 bool ordered = false;
2349 /* ordered log vectors have no regions to write */
2350 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2351 ASSERT(lv->lv_niovecs == 0);
2357 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2358 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2360 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2361 if (start_rec_copy) {
2363 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2367 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2371 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2372 sizeof(struct xlog_op_header));
2374 len += xlog_write_setup_copy(ticket, ophdr,
2375 iclog->ic_size-log_offset,
2377 ©_off, ©_len,
2380 xlog_verify_dest_ptr(log, ptr);
2383 ASSERT(copy_len >= 0);
2384 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2385 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2387 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2389 data_cnt += contwr ? copy_len : 0;
2391 error = xlog_write_copy_finish(log, iclog, flags,
2392 &record_cnt, &data_cnt,
2401 * if we had a partial copy, we need to get more iclog
2402 * space but we don't want to increment the region
2403 * index because there is still more is this region to
2406 * If we completed writing this region, and we flushed
2407 * the iclog (indicated by resetting of the record
2408 * count), then we also need to get more log space. If
2409 * this was the last record, though, we are done and
2415 if (++index == lv->lv_niovecs) {
2420 vecp = lv->lv_iovecp;
2422 if (record_cnt == 0 && ordered == false) {
2432 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2434 return xlog_state_release_iclog(log, iclog);
2436 ASSERT(flags & XLOG_COMMIT_TRANS);
2437 *commit_iclog = iclog;
2442 /*****************************************************************************
2444 * State Machine functions
2446 *****************************************************************************
2449 /* Clean iclogs starting from the head. This ordering must be
2450 * maintained, so an iclog doesn't become ACTIVE beyond one that
2451 * is SYNCING. This is also required to maintain the notion that we use
2452 * a ordered wait queue to hold off would be writers to the log when every
2453 * iclog is trying to sync to disk.
2455 * State Change: DIRTY -> ACTIVE
2458 xlog_state_clean_log(
2461 xlog_in_core_t *iclog;
2464 iclog = log->l_iclog;
2466 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2467 iclog->ic_state = XLOG_STATE_ACTIVE;
2468 iclog->ic_offset = 0;
2469 ASSERT(iclog->ic_callback == NULL);
2471 * If the number of ops in this iclog indicate it just
2472 * contains the dummy transaction, we can
2473 * change state into IDLE (the second time around).
2474 * Otherwise we should change the state into
2476 * We don't need to cover the dummy.
2479 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2484 * We have two dirty iclogs so start over
2485 * This could also be num of ops indicates
2486 * this is not the dummy going out.
2490 iclog->ic_header.h_num_logops = 0;
2491 memset(iclog->ic_header.h_cycle_data, 0,
2492 sizeof(iclog->ic_header.h_cycle_data));
2493 iclog->ic_header.h_lsn = 0;
2494 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2497 break; /* stop cleaning */
2498 iclog = iclog->ic_next;
2499 } while (iclog != log->l_iclog);
2501 /* log is locked when we are called */
2503 * Change state for the dummy log recording.
2504 * We usually go to NEED. But we go to NEED2 if the changed indicates
2505 * we are done writing the dummy record.
2506 * If we are done with the second dummy recored (DONE2), then
2510 switch (log->l_covered_state) {
2511 case XLOG_STATE_COVER_IDLE:
2512 case XLOG_STATE_COVER_NEED:
2513 case XLOG_STATE_COVER_NEED2:
2514 log->l_covered_state = XLOG_STATE_COVER_NEED;
2517 case XLOG_STATE_COVER_DONE:
2519 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2521 log->l_covered_state = XLOG_STATE_COVER_NEED;
2524 case XLOG_STATE_COVER_DONE2:
2526 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2528 log->l_covered_state = XLOG_STATE_COVER_NEED;
2535 } /* xlog_state_clean_log */
2538 xlog_get_lowest_lsn(
2541 xlog_in_core_t *lsn_log;
2542 xfs_lsn_t lowest_lsn, lsn;
2544 lsn_log = log->l_iclog;
2547 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2548 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2549 if ((lsn && !lowest_lsn) ||
2550 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2554 lsn_log = lsn_log->ic_next;
2555 } while (lsn_log != log->l_iclog);
2561 xlog_state_do_callback(
2564 struct xlog_in_core *ciclog)
2566 xlog_in_core_t *iclog;
2567 xlog_in_core_t *first_iclog; /* used to know when we've
2568 * processed all iclogs once */
2569 xfs_log_callback_t *cb, *cb_next;
2571 xfs_lsn_t lowest_lsn;
2572 int ioerrors; /* counter: iclogs with errors */
2573 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2574 int funcdidcallbacks; /* flag: function did callbacks */
2575 int repeats; /* for issuing console warnings if
2576 * looping too many times */
2579 spin_lock(&log->l_icloglock);
2580 first_iclog = iclog = log->l_iclog;
2582 funcdidcallbacks = 0;
2587 * Scan all iclogs starting with the one pointed to by the
2588 * log. Reset this starting point each time the log is
2589 * unlocked (during callbacks).
2591 * Keep looping through iclogs until one full pass is made
2592 * without running any callbacks.
2594 first_iclog = log->l_iclog;
2595 iclog = log->l_iclog;
2596 loopdidcallbacks = 0;
2601 /* skip all iclogs in the ACTIVE & DIRTY states */
2602 if (iclog->ic_state &
2603 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2604 iclog = iclog->ic_next;
2609 * Between marking a filesystem SHUTDOWN and stopping
2610 * the log, we do flush all iclogs to disk (if there
2611 * wasn't a log I/O error). So, we do want things to
2612 * go smoothly in case of just a SHUTDOWN w/o a
2615 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2617 * Can only perform callbacks in order. Since
2618 * this iclog is not in the DONE_SYNC/
2619 * DO_CALLBACK state, we skip the rest and
2620 * just try to clean up. If we set our iclog
2621 * to DO_CALLBACK, we will not process it when
2622 * we retry since a previous iclog is in the
2623 * CALLBACK and the state cannot change since
2624 * we are holding the l_icloglock.
2626 if (!(iclog->ic_state &
2627 (XLOG_STATE_DONE_SYNC |
2628 XLOG_STATE_DO_CALLBACK))) {
2629 if (ciclog && (ciclog->ic_state ==
2630 XLOG_STATE_DONE_SYNC)) {
2631 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2636 * We now have an iclog that is in either the
2637 * DO_CALLBACK or DONE_SYNC states. The other
2638 * states (WANT_SYNC, SYNCING, or CALLBACK were
2639 * caught by the above if and are going to
2640 * clean (i.e. we aren't doing their callbacks)
2645 * We will do one more check here to see if we
2646 * have chased our tail around.
2649 lowest_lsn = xlog_get_lowest_lsn(log);
2651 XFS_LSN_CMP(lowest_lsn,
2652 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2653 iclog = iclog->ic_next;
2654 continue; /* Leave this iclog for
2658 iclog->ic_state = XLOG_STATE_CALLBACK;
2662 * Completion of a iclog IO does not imply that
2663 * a transaction has completed, as transactions
2664 * can be large enough to span many iclogs. We
2665 * cannot change the tail of the log half way
2666 * through a transaction as this may be the only
2667 * transaction in the log and moving th etail to
2668 * point to the middle of it will prevent
2669 * recovery from finding the start of the
2670 * transaction. Hence we should only update the
2671 * last_sync_lsn if this iclog contains
2672 * transaction completion callbacks on it.
2674 * We have to do this before we drop the
2675 * icloglock to ensure we are the only one that
2678 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2679 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2680 if (iclog->ic_callback)
2681 atomic64_set(&log->l_last_sync_lsn,
2682 be64_to_cpu(iclog->ic_header.h_lsn));
2687 spin_unlock(&log->l_icloglock);
2690 * Keep processing entries in the callback list until
2691 * we come around and it is empty. We need to
2692 * atomically see that the list is empty and change the
2693 * state to DIRTY so that we don't miss any more
2694 * callbacks being added.
2696 spin_lock(&iclog->ic_callback_lock);
2697 cb = iclog->ic_callback;
2699 iclog->ic_callback_tail = &(iclog->ic_callback);
2700 iclog->ic_callback = NULL;
2701 spin_unlock(&iclog->ic_callback_lock);
2703 /* perform callbacks in the order given */
2704 for (; cb; cb = cb_next) {
2705 cb_next = cb->cb_next;
2706 cb->cb_func(cb->cb_arg, aborted);
2708 spin_lock(&iclog->ic_callback_lock);
2709 cb = iclog->ic_callback;
2715 spin_lock(&log->l_icloglock);
2716 ASSERT(iclog->ic_callback == NULL);
2717 spin_unlock(&iclog->ic_callback_lock);
2718 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2719 iclog->ic_state = XLOG_STATE_DIRTY;
2722 * Transition from DIRTY to ACTIVE if applicable.
2723 * NOP if STATE_IOERROR.
2725 xlog_state_clean_log(log);
2727 /* wake up threads waiting in xfs_log_force() */
2728 wake_up_all(&iclog->ic_force_wait);
2730 iclog = iclog->ic_next;
2731 } while (first_iclog != iclog);
2733 if (repeats > 5000) {
2734 flushcnt += repeats;
2737 "%s: possible infinite loop (%d iterations)",
2738 __func__, flushcnt);
2740 } while (!ioerrors && loopdidcallbacks);
2743 * make one last gasp attempt to see if iclogs are being left in
2747 if (funcdidcallbacks) {
2748 first_iclog = iclog = log->l_iclog;
2750 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2752 * Terminate the loop if iclogs are found in states
2753 * which will cause other threads to clean up iclogs.
2755 * SYNCING - i/o completion will go through logs
2756 * DONE_SYNC - interrupt thread should be waiting for
2758 * IOERROR - give up hope all ye who enter here
2760 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2761 iclog->ic_state == XLOG_STATE_SYNCING ||
2762 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2763 iclog->ic_state == XLOG_STATE_IOERROR )
2765 iclog = iclog->ic_next;
2766 } while (first_iclog != iclog);
2770 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2772 spin_unlock(&log->l_icloglock);
2775 wake_up_all(&log->l_flush_wait);
2780 * Finish transitioning this iclog to the dirty state.
2782 * Make sure that we completely execute this routine only when this is
2783 * the last call to the iclog. There is a good chance that iclog flushes,
2784 * when we reach the end of the physical log, get turned into 2 separate
2785 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2786 * routine. By using the reference count bwritecnt, we guarantee that only
2787 * the second completion goes through.
2789 * Callbacks could take time, so they are done outside the scope of the
2790 * global state machine log lock.
2793 xlog_state_done_syncing(
2794 xlog_in_core_t *iclog,
2797 struct xlog *log = iclog->ic_log;
2799 spin_lock(&log->l_icloglock);
2801 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2802 iclog->ic_state == XLOG_STATE_IOERROR);
2803 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2804 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2808 * If we got an error, either on the first buffer, or in the case of
2809 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2810 * and none should ever be attempted to be written to disk
2813 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2814 if (--iclog->ic_bwritecnt == 1) {
2815 spin_unlock(&log->l_icloglock);
2818 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2822 * Someone could be sleeping prior to writing out the next
2823 * iclog buffer, we wake them all, one will get to do the
2824 * I/O, the others get to wait for the result.
2826 wake_up_all(&iclog->ic_write_wait);
2827 spin_unlock(&log->l_icloglock);
2828 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2829 } /* xlog_state_done_syncing */
2833 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2834 * sleep. We wait on the flush queue on the head iclog as that should be
2835 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2836 * we will wait here and all new writes will sleep until a sync completes.
2838 * The in-core logs are used in a circular fashion. They are not used
2839 * out-of-order even when an iclog past the head is free.
2842 * * log_offset where xlog_write() can start writing into the in-core
2844 * * in-core log pointer to which xlog_write() should write.
2845 * * boolean indicating this is a continued write to an in-core log.
2846 * If this is the last write, then the in-core log's offset field
2847 * needs to be incremented, depending on the amount of data which
2851 xlog_state_get_iclog_space(
2854 struct xlog_in_core **iclogp,
2855 struct xlog_ticket *ticket,
2856 int *continued_write,
2860 xlog_rec_header_t *head;
2861 xlog_in_core_t *iclog;
2865 spin_lock(&log->l_icloglock);
2866 if (XLOG_FORCED_SHUTDOWN(log)) {
2867 spin_unlock(&log->l_icloglock);
2871 iclog = log->l_iclog;
2872 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2873 XFS_STATS_INC(xs_log_noiclogs);
2875 /* Wait for log writes to have flushed */
2876 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2880 head = &iclog->ic_header;
2882 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2883 log_offset = iclog->ic_offset;
2885 /* On the 1st write to an iclog, figure out lsn. This works
2886 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2887 * committing to. If the offset is set, that's how many blocks
2890 if (log_offset == 0) {
2891 ticket->t_curr_res -= log->l_iclog_hsize;
2892 xlog_tic_add_region(ticket,
2894 XLOG_REG_TYPE_LRHEADER);
2895 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2896 head->h_lsn = cpu_to_be64(
2897 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2898 ASSERT(log->l_curr_block >= 0);
2901 /* If there is enough room to write everything, then do it. Otherwise,
2902 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2903 * bit is on, so this will get flushed out. Don't update ic_offset
2904 * until you know exactly how many bytes get copied. Therefore, wait
2905 * until later to update ic_offset.
2907 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2908 * can fit into remaining data section.
2910 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2911 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2914 * If I'm the only one writing to this iclog, sync it to disk.
2915 * We need to do an atomic compare and decrement here to avoid
2916 * racing with concurrent atomic_dec_and_lock() calls in
2917 * xlog_state_release_iclog() when there is more than one
2918 * reference to the iclog.
2920 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2921 /* we are the only one */
2922 spin_unlock(&log->l_icloglock);
2923 error = xlog_state_release_iclog(log, iclog);
2927 spin_unlock(&log->l_icloglock);
2932 /* Do we have enough room to write the full amount in the remainder
2933 * of this iclog? Or must we continue a write on the next iclog and
2934 * mark this iclog as completely taken? In the case where we switch
2935 * iclogs (to mark it taken), this particular iclog will release/sync
2936 * to disk in xlog_write().
2938 if (len <= iclog->ic_size - iclog->ic_offset) {
2939 *continued_write = 0;
2940 iclog->ic_offset += len;
2942 *continued_write = 1;
2943 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2947 ASSERT(iclog->ic_offset <= iclog->ic_size);
2948 spin_unlock(&log->l_icloglock);
2950 *logoffsetp = log_offset;
2952 } /* xlog_state_get_iclog_space */
2954 /* The first cnt-1 times through here we don't need to
2955 * move the grant write head because the permanent
2956 * reservation has reserved cnt times the unit amount.
2957 * Release part of current permanent unit reservation and
2958 * reset current reservation to be one units worth. Also
2959 * move grant reservation head forward.
2962 xlog_regrant_reserve_log_space(
2964 struct xlog_ticket *ticket)
2966 trace_xfs_log_regrant_reserve_enter(log, ticket);
2968 if (ticket->t_cnt > 0)
2971 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2972 ticket->t_curr_res);
2973 xlog_grant_sub_space(log, &log->l_write_head.grant,
2974 ticket->t_curr_res);
2975 ticket->t_curr_res = ticket->t_unit_res;
2976 xlog_tic_reset_res(ticket);
2978 trace_xfs_log_regrant_reserve_sub(log, ticket);
2980 /* just return if we still have some of the pre-reserved space */
2981 if (ticket->t_cnt > 0)
2984 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2985 ticket->t_unit_res);
2987 trace_xfs_log_regrant_reserve_exit(log, ticket);
2989 ticket->t_curr_res = ticket->t_unit_res;
2990 xlog_tic_reset_res(ticket);
2991 } /* xlog_regrant_reserve_log_space */
2995 * Give back the space left from a reservation.
2997 * All the information we need to make a correct determination of space left
2998 * is present. For non-permanent reservations, things are quite easy. The
2999 * count should have been decremented to zero. We only need to deal with the
3000 * space remaining in the current reservation part of the ticket. If the
3001 * ticket contains a permanent reservation, there may be left over space which
3002 * needs to be released. A count of N means that N-1 refills of the current
3003 * reservation can be done before we need to ask for more space. The first
3004 * one goes to fill up the first current reservation. Once we run out of
3005 * space, the count will stay at zero and the only space remaining will be
3006 * in the current reservation field.
3009 xlog_ungrant_log_space(
3011 struct xlog_ticket *ticket)
3015 if (ticket->t_cnt > 0)
3018 trace_xfs_log_ungrant_enter(log, ticket);
3019 trace_xfs_log_ungrant_sub(log, ticket);
3022 * If this is a permanent reservation ticket, we may be able to free
3023 * up more space based on the remaining count.
3025 bytes = ticket->t_curr_res;
3026 if (ticket->t_cnt > 0) {
3027 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3028 bytes += ticket->t_unit_res*ticket->t_cnt;
3031 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3032 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3034 trace_xfs_log_ungrant_exit(log, ticket);
3036 xfs_log_space_wake(log->l_mp);
3040 * Flush iclog to disk if this is the last reference to the given iclog and
3041 * the WANT_SYNC bit is set.
3043 * When this function is entered, the iclog is not necessarily in the
3044 * WANT_SYNC state. It may be sitting around waiting to get filled.
3049 xlog_state_release_iclog(
3051 struct xlog_in_core *iclog)
3053 int sync = 0; /* do we sync? */
3055 if (iclog->ic_state & XLOG_STATE_IOERROR)
3058 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3059 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3062 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3063 spin_unlock(&log->l_icloglock);
3066 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3067 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3069 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3070 /* update tail before writing to iclog */
3071 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3073 iclog->ic_state = XLOG_STATE_SYNCING;
3074 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3075 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3076 /* cycle incremented when incrementing curr_block */
3078 spin_unlock(&log->l_icloglock);
3081 * We let the log lock go, so it's possible that we hit a log I/O
3082 * error or some other SHUTDOWN condition that marks the iclog
3083 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3084 * this iclog has consistent data, so we ignore IOERROR
3085 * flags after this point.
3088 return xlog_sync(log, iclog);
3090 } /* xlog_state_release_iclog */
3094 * This routine will mark the current iclog in the ring as WANT_SYNC
3095 * and move the current iclog pointer to the next iclog in the ring.
3096 * When this routine is called from xlog_state_get_iclog_space(), the
3097 * exact size of the iclog has not yet been determined. All we know is
3098 * that every data block. We have run out of space in this log record.
3101 xlog_state_switch_iclogs(
3103 struct xlog_in_core *iclog,
3106 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3108 eventual_size = iclog->ic_offset;
3109 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3110 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3111 log->l_prev_block = log->l_curr_block;
3112 log->l_prev_cycle = log->l_curr_cycle;
3114 /* roll log?: ic_offset changed later */
3115 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3117 /* Round up to next log-sunit */
3118 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3119 log->l_mp->m_sb.sb_logsunit > 1) {
3120 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3121 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3124 if (log->l_curr_block >= log->l_logBBsize) {
3125 log->l_curr_cycle++;
3126 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3127 log->l_curr_cycle++;
3128 log->l_curr_block -= log->l_logBBsize;
3129 ASSERT(log->l_curr_block >= 0);
3131 ASSERT(iclog == log->l_iclog);
3132 log->l_iclog = iclog->ic_next;
3133 } /* xlog_state_switch_iclogs */
3136 * Write out all data in the in-core log as of this exact moment in time.
3138 * Data may be written to the in-core log during this call. However,
3139 * we don't guarantee this data will be written out. A change from past
3140 * implementation means this routine will *not* write out zero length LRs.
3142 * Basically, we try and perform an intelligent scan of the in-core logs.
3143 * If we determine there is no flushable data, we just return. There is no
3144 * flushable data if:
3146 * 1. the current iclog is active and has no data; the previous iclog
3147 * is in the active or dirty state.
3148 * 2. the current iclog is drity, and the previous iclog is in the
3149 * active or dirty state.
3153 * 1. the current iclog is not in the active nor dirty state.
3154 * 2. the current iclog dirty, and the previous iclog is not in the
3155 * active nor dirty state.
3156 * 3. the current iclog is active, and there is another thread writing
3157 * to this particular iclog.
3158 * 4. a) the current iclog is active and has no other writers
3159 * b) when we return from flushing out this iclog, it is still
3160 * not in the active nor dirty state.
3164 struct xfs_mount *mp,
3168 struct xlog *log = mp->m_log;
3169 struct xlog_in_core *iclog;
3172 XFS_STATS_INC(xs_log_force);
3174 xlog_cil_force(log);
3176 spin_lock(&log->l_icloglock);
3178 iclog = log->l_iclog;
3179 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3180 spin_unlock(&log->l_icloglock);
3184 /* If the head iclog is not active nor dirty, we just attach
3185 * ourselves to the head and go to sleep.
3187 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3188 iclog->ic_state == XLOG_STATE_DIRTY) {
3190 * If the head is dirty or (active and empty), then
3191 * we need to look at the previous iclog. If the previous
3192 * iclog is active or dirty we are done. There is nothing
3193 * to sync out. Otherwise, we attach ourselves to the
3194 * previous iclog and go to sleep.
3196 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3197 (atomic_read(&iclog->ic_refcnt) == 0
3198 && iclog->ic_offset == 0)) {
3199 iclog = iclog->ic_prev;
3200 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3201 iclog->ic_state == XLOG_STATE_DIRTY)
3206 if (atomic_read(&iclog->ic_refcnt) == 0) {
3207 /* We are the only one with access to this
3208 * iclog. Flush it out now. There should
3209 * be a roundoff of zero to show that someone
3210 * has already taken care of the roundoff from
3211 * the previous sync.
3213 atomic_inc(&iclog->ic_refcnt);
3214 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3215 xlog_state_switch_iclogs(log, iclog, 0);
3216 spin_unlock(&log->l_icloglock);
3218 if (xlog_state_release_iclog(log, iclog))
3223 spin_lock(&log->l_icloglock);
3224 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3225 iclog->ic_state != XLOG_STATE_DIRTY)
3230 /* Someone else is writing to this iclog.
3231 * Use its call to flush out the data. However,
3232 * the other thread may not force out this LR,
3233 * so we mark it WANT_SYNC.
3235 xlog_state_switch_iclogs(log, iclog, 0);
3241 /* By the time we come around again, the iclog could've been filled
3242 * which would give it another lsn. If we have a new lsn, just
3243 * return because the relevant data has been flushed.
3246 if (flags & XFS_LOG_SYNC) {
3248 * We must check if we're shutting down here, before
3249 * we wait, while we're holding the l_icloglock.
3250 * Then we check again after waking up, in case our
3251 * sleep was disturbed by a bad news.
3253 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3254 spin_unlock(&log->l_icloglock);
3257 XFS_STATS_INC(xs_log_force_sleep);
3258 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3260 * No need to grab the log lock here since we're
3261 * only deciding whether or not to return EIO
3262 * and the memory read should be atomic.
3264 if (iclog->ic_state & XLOG_STATE_IOERROR)
3271 spin_unlock(&log->l_icloglock);
3277 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3278 * about errors or whether the log was flushed or not. This is the normal
3279 * interface to use when trying to unpin items or move the log forward.
3288 trace_xfs_log_force(mp, 0);
3289 error = _xfs_log_force(mp, flags, NULL);
3291 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3295 * Force the in-core log to disk for a specific LSN.
3297 * Find in-core log with lsn.
3298 * If it is in the DIRTY state, just return.
3299 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3300 * state and go to sleep or return.
3301 * If it is in any other state, go to sleep or return.
3303 * Synchronous forces are implemented with a signal variable. All callers
3304 * to force a given lsn to disk will wait on a the sv attached to the
3305 * specific in-core log. When given in-core log finally completes its
3306 * write to disk, that thread will wake up all threads waiting on the
3311 struct xfs_mount *mp,
3316 struct xlog *log = mp->m_log;
3317 struct xlog_in_core *iclog;
3318 int already_slept = 0;
3322 XFS_STATS_INC(xs_log_force);
3324 lsn = xlog_cil_force_lsn(log, lsn);
3325 if (lsn == NULLCOMMITLSN)
3329 spin_lock(&log->l_icloglock);
3330 iclog = log->l_iclog;
3331 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3332 spin_unlock(&log->l_icloglock);
3337 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3338 iclog = iclog->ic_next;
3342 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3343 spin_unlock(&log->l_icloglock);
3347 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3349 * We sleep here if we haven't already slept (e.g.
3350 * this is the first time we've looked at the correct
3351 * iclog buf) and the buffer before us is going to
3352 * be sync'ed. The reason for this is that if we
3353 * are doing sync transactions here, by waiting for
3354 * the previous I/O to complete, we can allow a few
3355 * more transactions into this iclog before we close
3358 * Otherwise, we mark the buffer WANT_SYNC, and bump
3359 * up the refcnt so we can release the log (which
3360 * drops the ref count). The state switch keeps new
3361 * transaction commits from using this buffer. When
3362 * the current commits finish writing into the buffer,
3363 * the refcount will drop to zero and the buffer will
3366 if (!already_slept &&
3367 (iclog->ic_prev->ic_state &
3368 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3369 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3371 XFS_STATS_INC(xs_log_force_sleep);
3373 xlog_wait(&iclog->ic_prev->ic_write_wait,
3380 atomic_inc(&iclog->ic_refcnt);
3381 xlog_state_switch_iclogs(log, iclog, 0);
3382 spin_unlock(&log->l_icloglock);
3383 if (xlog_state_release_iclog(log, iclog))
3387 spin_lock(&log->l_icloglock);
3390 if ((flags & XFS_LOG_SYNC) && /* sleep */
3392 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3394 * Don't wait on completion if we know that we've
3395 * gotten a log write error.
3397 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3398 spin_unlock(&log->l_icloglock);
3401 XFS_STATS_INC(xs_log_force_sleep);
3402 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3404 * No need to grab the log lock here since we're
3405 * only deciding whether or not to return EIO
3406 * and the memory read should be atomic.
3408 if (iclog->ic_state & XLOG_STATE_IOERROR)
3413 } else { /* just return */
3414 spin_unlock(&log->l_icloglock);
3418 } while (iclog != log->l_iclog);
3420 spin_unlock(&log->l_icloglock);
3425 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3426 * about errors or whether the log was flushed or not. This is the normal
3427 * interface to use when trying to unpin items or move the log forward.
3437 trace_xfs_log_force(mp, lsn);
3438 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3440 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3444 * Called when we want to mark the current iclog as being ready to sync to
3448 xlog_state_want_sync(
3450 struct xlog_in_core *iclog)
3452 assert_spin_locked(&log->l_icloglock);
3454 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3455 xlog_state_switch_iclogs(log, iclog, 0);
3457 ASSERT(iclog->ic_state &
3458 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3463 /*****************************************************************************
3467 *****************************************************************************
3471 * Free a used ticket when its refcount falls to zero.
3475 xlog_ticket_t *ticket)
3477 ASSERT(atomic_read(&ticket->t_ref) > 0);
3478 if (atomic_dec_and_test(&ticket->t_ref))
3479 kmem_zone_free(xfs_log_ticket_zone, ticket);
3484 xlog_ticket_t *ticket)
3486 ASSERT(atomic_read(&ticket->t_ref) > 0);
3487 atomic_inc(&ticket->t_ref);
3492 * Figure out the total log space unit (in bytes) that would be
3493 * required for a log ticket.
3496 xfs_log_calc_unit_res(
3497 struct xfs_mount *mp,
3500 struct xlog *log = mp->m_log;
3505 * Permanent reservations have up to 'cnt'-1 active log operations
3506 * in the log. A unit in this case is the amount of space for one
3507 * of these log operations. Normal reservations have a cnt of 1
3508 * and their unit amount is the total amount of space required.
3510 * The following lines of code account for non-transaction data
3511 * which occupy space in the on-disk log.
3513 * Normal form of a transaction is:
3514 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3515 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3517 * We need to account for all the leadup data and trailer data
3518 * around the transaction data.
3519 * And then we need to account for the worst case in terms of using
3521 * The worst case will happen if:
3522 * - the placement of the transaction happens to be such that the
3523 * roundoff is at its maximum
3524 * - the transaction data is synced before the commit record is synced
3525 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3526 * Therefore the commit record is in its own Log Record.
3527 * This can happen as the commit record is called with its
3528 * own region to xlog_write().
3529 * This then means that in the worst case, roundoff can happen for
3530 * the commit-rec as well.
3531 * The commit-rec is smaller than padding in this scenario and so it is
3532 * not added separately.
3535 /* for trans header */
3536 unit_bytes += sizeof(xlog_op_header_t);
3537 unit_bytes += sizeof(xfs_trans_header_t);
3540 unit_bytes += sizeof(xlog_op_header_t);
3543 * for LR headers - the space for data in an iclog is the size minus
3544 * the space used for the headers. If we use the iclog size, then we
3545 * undercalculate the number of headers required.
3547 * Furthermore - the addition of op headers for split-recs might
3548 * increase the space required enough to require more log and op
3549 * headers, so take that into account too.
3551 * IMPORTANT: This reservation makes the assumption that if this
3552 * transaction is the first in an iclog and hence has the LR headers
3553 * accounted to it, then the remaining space in the iclog is
3554 * exclusively for this transaction. i.e. if the transaction is larger
3555 * than the iclog, it will be the only thing in that iclog.
3556 * Fundamentally, this means we must pass the entire log vector to
3557 * xlog_write to guarantee this.
3559 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3560 num_headers = howmany(unit_bytes, iclog_space);
3562 /* for split-recs - ophdrs added when data split over LRs */
3563 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3565 /* add extra header reservations if we overrun */
3566 while (!num_headers ||
3567 howmany(unit_bytes, iclog_space) > num_headers) {
3568 unit_bytes += sizeof(xlog_op_header_t);
3571 unit_bytes += log->l_iclog_hsize * num_headers;
3573 /* for commit-rec LR header - note: padding will subsume the ophdr */
3574 unit_bytes += log->l_iclog_hsize;
3576 /* for roundoff padding for transaction data and one for commit record */
3577 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3578 /* log su roundoff */
3579 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3582 unit_bytes += 2 * BBSIZE;
3589 * Allocate and initialise a new log ticket.
3591 struct xlog_ticket *
3598 xfs_km_flags_t alloc_flags)
3600 struct xlog_ticket *tic;
3603 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3607 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3609 atomic_set(&tic->t_ref, 1);
3610 tic->t_task = current;
3611 INIT_LIST_HEAD(&tic->t_queue);
3612 tic->t_unit_res = unit_res;
3613 tic->t_curr_res = unit_res;
3616 tic->t_tid = prandom_u32();
3617 tic->t_clientid = client;
3618 tic->t_flags = XLOG_TIC_INITED;
3619 tic->t_trans_type = 0;
3621 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3623 xlog_tic_reset_res(tic);
3629 /******************************************************************************
3631 * Log debug routines
3633 ******************************************************************************
3637 * Make sure that the destination ptr is within the valid data region of
3638 * one of the iclogs. This uses backup pointers stored in a different
3639 * part of the log in case we trash the log structure.
3642 xlog_verify_dest_ptr(
3649 for (i = 0; i < log->l_iclog_bufs; i++) {
3650 if (ptr >= log->l_iclog_bak[i] &&
3651 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3656 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3660 * Check to make sure the grant write head didn't just over lap the tail. If
3661 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3662 * the cycles differ by exactly one and check the byte count.
3664 * This check is run unlocked, so can give false positives. Rather than assert
3665 * on failures, use a warn-once flag and a panic tag to allow the admin to
3666 * determine if they want to panic the machine when such an error occurs. For
3667 * debug kernels this will have the same effect as using an assert but, unlinke
3668 * an assert, it can be turned off at runtime.
3671 xlog_verify_grant_tail(
3674 int tail_cycle, tail_blocks;
3677 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3678 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3679 if (tail_cycle != cycle) {
3680 if (cycle - 1 != tail_cycle &&
3681 !(log->l_flags & XLOG_TAIL_WARN)) {
3682 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3683 "%s: cycle - 1 != tail_cycle", __func__);
3684 log->l_flags |= XLOG_TAIL_WARN;
3687 if (space > BBTOB(tail_blocks) &&
3688 !(log->l_flags & XLOG_TAIL_WARN)) {
3689 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3690 "%s: space > BBTOB(tail_blocks)", __func__);
3691 log->l_flags |= XLOG_TAIL_WARN;
3696 /* check if it will fit */
3698 xlog_verify_tail_lsn(
3700 struct xlog_in_core *iclog,
3705 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3707 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3708 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3709 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3711 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3713 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3714 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3716 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3717 if (blocks < BTOBB(iclog->ic_offset) + 1)
3718 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3720 } /* xlog_verify_tail_lsn */
3723 * Perform a number of checks on the iclog before writing to disk.
3725 * 1. Make sure the iclogs are still circular
3726 * 2. Make sure we have a good magic number
3727 * 3. Make sure we don't have magic numbers in the data
3728 * 4. Check fields of each log operation header for:
3729 * A. Valid client identifier
3730 * B. tid ptr value falls in valid ptr space (user space code)
3731 * C. Length in log record header is correct according to the
3732 * individual operation headers within record.
3733 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3734 * log, check the preceding blocks of the physical log to make sure all
3735 * the cycle numbers agree with the current cycle number.
3740 struct xlog_in_core *iclog,
3744 xlog_op_header_t *ophead;
3745 xlog_in_core_t *icptr;
3746 xlog_in_core_2_t *xhdr;
3748 xfs_caddr_t base_ptr;
3749 __psint_t field_offset;
3751 int len, i, j, k, op_len;
3754 /* check validity of iclog pointers */
3755 spin_lock(&log->l_icloglock);
3756 icptr = log->l_iclog;
3757 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3760 if (icptr != log->l_iclog)
3761 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3762 spin_unlock(&log->l_icloglock);
3764 /* check log magic numbers */
3765 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3766 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3768 ptr = (xfs_caddr_t) &iclog->ic_header;
3769 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3771 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3772 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3777 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3778 ptr = iclog->ic_datap;
3780 ophead = (xlog_op_header_t *)ptr;
3781 xhdr = iclog->ic_data;
3782 for (i = 0; i < len; i++) {
3783 ophead = (xlog_op_header_t *)ptr;
3785 /* clientid is only 1 byte */
3786 field_offset = (__psint_t)
3787 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3788 if (!syncing || (field_offset & 0x1ff)) {
3789 clientid = ophead->oh_clientid;
3791 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3792 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3793 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3794 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3795 clientid = xlog_get_client_id(
3796 xhdr[j].hic_xheader.xh_cycle_data[k]);
3798 clientid = xlog_get_client_id(
3799 iclog->ic_header.h_cycle_data[idx]);
3802 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3804 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3805 __func__, clientid, ophead,
3806 (unsigned long)field_offset);
3809 field_offset = (__psint_t)
3810 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3811 if (!syncing || (field_offset & 0x1ff)) {
3812 op_len = be32_to_cpu(ophead->oh_len);
3814 idx = BTOBBT((__psint_t)&ophead->oh_len -
3815 (__psint_t)iclog->ic_datap);
3816 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3817 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3818 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3819 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3821 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3824 ptr += sizeof(xlog_op_header_t) + op_len;
3826 } /* xlog_verify_iclog */
3830 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3836 xlog_in_core_t *iclog, *ic;
3838 iclog = log->l_iclog;
3839 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3841 * Mark all the incore logs IOERROR.
3842 * From now on, no log flushes will result.
3846 ic->ic_state = XLOG_STATE_IOERROR;
3848 } while (ic != iclog);
3852 * Return non-zero, if state transition has already happened.
3858 * This is called from xfs_force_shutdown, when we're forcibly
3859 * shutting down the filesystem, typically because of an IO error.
3860 * Our main objectives here are to make sure that:
3861 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3862 * parties to find out, 'atomically'.
3863 * b. those who're sleeping on log reservations, pinned objects and
3864 * other resources get woken up, and be told the bad news.
3865 * c. nothing new gets queued up after (a) and (b) are done.
3866 * d. if !logerror, flush the iclogs to disk, then seal them off
3869 * Note: for delayed logging the !logerror case needs to flush the regions
3870 * held in memory out to the iclogs before flushing them to disk. This needs
3871 * to be done before the log is marked as shutdown, otherwise the flush to the
3875 xfs_log_force_umount(
3876 struct xfs_mount *mp,
3885 * If this happens during log recovery, don't worry about
3886 * locking; the log isn't open for business yet.
3889 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3890 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3892 XFS_BUF_DONE(mp->m_sb_bp);
3897 * Somebody could've already done the hard work for us.
3898 * No need to get locks for this.
3900 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3901 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3907 * Flush the in memory commit item list before marking the log as
3908 * being shut down. We need to do it in this order to ensure all the
3909 * completed transactions are flushed to disk with the xfs_log_force()
3913 xlog_cil_force(log);
3916 * mark the filesystem and the as in a shutdown state and wake
3917 * everybody up to tell them the bad news.
3919 spin_lock(&log->l_icloglock);
3920 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3922 XFS_BUF_DONE(mp->m_sb_bp);
3925 * This flag is sort of redundant because of the mount flag, but
3926 * it's good to maintain the separation between the log and the rest
3929 log->l_flags |= XLOG_IO_ERROR;
3932 * If we hit a log error, we want to mark all the iclogs IOERROR
3933 * while we're still holding the loglock.
3936 retval = xlog_state_ioerror(log);
3937 spin_unlock(&log->l_icloglock);
3940 * We don't want anybody waiting for log reservations after this. That
3941 * means we have to wake up everybody queued up on reserveq as well as
3942 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3943 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3944 * action is protected by the grant locks.
3946 xlog_grant_head_wake_all(&log->l_reserve_head);
3947 xlog_grant_head_wake_all(&log->l_write_head);
3949 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3952 * Force the incore logs to disk before shutting the
3953 * log down completely.
3955 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3957 spin_lock(&log->l_icloglock);
3958 retval = xlog_state_ioerror(log);
3959 spin_unlock(&log->l_icloglock);
3963 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3964 * as if the log writes were completed. The abort handling in the log
3965 * item committed callback functions will do this again under lock to
3968 wake_up_all(&log->l_cilp->xc_commit_wait);
3969 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3971 #ifdef XFSERRORDEBUG
3973 xlog_in_core_t *iclog;
3975 spin_lock(&log->l_icloglock);
3976 iclog = log->l_iclog;
3978 ASSERT(iclog->ic_callback == 0);
3979 iclog = iclog->ic_next;
3980 } while (iclog != log->l_iclog);
3981 spin_unlock(&log->l_icloglock);
3984 /* return non-zero if log IOERROR transition had already happened */
3992 xlog_in_core_t *iclog;
3994 iclog = log->l_iclog;
3996 /* endianness does not matter here, zero is zero in
3999 if (iclog->ic_header.h_num_logops)
4001 iclog = iclog->ic_next;
4002 } while (iclog != log->l_iclog);