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"
37 #include "xfs_sysfs.h"
39 kmem_zone_t *xfs_log_ticket_zone;
41 /* Local miscellaneous function prototypes */
45 struct xlog_ticket *ticket,
46 struct xlog_in_core **iclog,
47 xfs_lsn_t *commitlsnp);
52 struct xfs_buftarg *log_target,
53 xfs_daddr_t blk_offset,
62 struct xlog_in_core *iclog);
67 /* local state machine functions */
68 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
70 xlog_state_do_callback(
73 struct xlog_in_core *iclog);
75 xlog_state_get_iclog_space(
78 struct xlog_in_core **iclog,
79 struct xlog_ticket *ticket,
83 xlog_state_release_iclog(
85 struct xlog_in_core *iclog);
87 xlog_state_switch_iclogs(
89 struct xlog_in_core *iclog,
94 struct xlog_in_core *iclog);
101 xlog_regrant_reserve_log_space(
103 struct xlog_ticket *ticket);
105 xlog_ungrant_log_space(
107 struct xlog_ticket *ticket);
111 xlog_verify_dest_ptr(
115 xlog_verify_grant_tail(
120 struct xlog_in_core *iclog,
124 xlog_verify_tail_lsn(
126 struct xlog_in_core *iclog,
129 #define xlog_verify_dest_ptr(a,b)
130 #define xlog_verify_grant_tail(a)
131 #define xlog_verify_iclog(a,b,c,d)
132 #define xlog_verify_tail_lsn(a,b,c)
140 xlog_grant_sub_space(
145 int64_t head_val = atomic64_read(head);
151 xlog_crack_grant_head_val(head_val, &cycle, &space);
155 space += log->l_logsize;
160 new = xlog_assign_grant_head_val(cycle, space);
161 head_val = atomic64_cmpxchg(head, old, new);
162 } while (head_val != old);
166 xlog_grant_add_space(
171 int64_t head_val = atomic64_read(head);
178 xlog_crack_grant_head_val(head_val, &cycle, &space);
180 tmp = log->l_logsize - space;
189 new = xlog_assign_grant_head_val(cycle, space);
190 head_val = atomic64_cmpxchg(head, old, new);
191 } while (head_val != old);
195 xlog_grant_head_init(
196 struct xlog_grant_head *head)
198 xlog_assign_grant_head(&head->grant, 1, 0);
199 INIT_LIST_HEAD(&head->waiters);
200 spin_lock_init(&head->lock);
204 xlog_grant_head_wake_all(
205 struct xlog_grant_head *head)
207 struct xlog_ticket *tic;
209 spin_lock(&head->lock);
210 list_for_each_entry(tic, &head->waiters, t_queue)
211 wake_up_process(tic->t_task);
212 spin_unlock(&head->lock);
216 xlog_ticket_reservation(
218 struct xlog_grant_head *head,
219 struct xlog_ticket *tic)
221 if (head == &log->l_write_head) {
222 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
223 return tic->t_unit_res;
225 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
226 return tic->t_unit_res * tic->t_cnt;
228 return tic->t_unit_res;
233 xlog_grant_head_wake(
235 struct xlog_grant_head *head,
238 struct xlog_ticket *tic;
241 list_for_each_entry(tic, &head->waiters, t_queue) {
242 need_bytes = xlog_ticket_reservation(log, head, tic);
243 if (*free_bytes < need_bytes)
246 *free_bytes -= need_bytes;
247 trace_xfs_log_grant_wake_up(log, tic);
248 wake_up_process(tic->t_task);
255 xlog_grant_head_wait(
257 struct xlog_grant_head *head,
258 struct xlog_ticket *tic,
259 int need_bytes) __releases(&head->lock)
260 __acquires(&head->lock)
262 list_add_tail(&tic->t_queue, &head->waiters);
265 if (XLOG_FORCED_SHUTDOWN(log))
267 xlog_grant_push_ail(log, need_bytes);
269 __set_current_state(TASK_UNINTERRUPTIBLE);
270 spin_unlock(&head->lock);
272 XFS_STATS_INC(xs_sleep_logspace);
274 trace_xfs_log_grant_sleep(log, tic);
276 trace_xfs_log_grant_wake(log, tic);
278 spin_lock(&head->lock);
279 if (XLOG_FORCED_SHUTDOWN(log))
281 } while (xlog_space_left(log, &head->grant) < need_bytes);
283 list_del_init(&tic->t_queue);
286 list_del_init(&tic->t_queue);
291 * Atomically get the log space required for a log ticket.
293 * Once a ticket gets put onto head->waiters, it will only return after the
294 * needed reservation is satisfied.
296 * This function is structured so that it has a lock free fast path. This is
297 * necessary because every new transaction reservation will come through this
298 * path. Hence any lock will be globally hot if we take it unconditionally on
301 * As tickets are only ever moved on and off head->waiters under head->lock, we
302 * only need to take that lock if we are going to add the ticket to the queue
303 * and sleep. We can avoid taking the lock if the ticket was never added to
304 * head->waiters because the t_queue list head will be empty and we hold the
305 * only reference to it so it can safely be checked unlocked.
308 xlog_grant_head_check(
310 struct xlog_grant_head *head,
311 struct xlog_ticket *tic,
317 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
320 * If there are other waiters on the queue then give them a chance at
321 * logspace before us. Wake up the first waiters, if we do not wake
322 * up all the waiters then go to sleep waiting for more free space,
323 * otherwise try to get some space for this transaction.
325 *need_bytes = xlog_ticket_reservation(log, head, tic);
326 free_bytes = xlog_space_left(log, &head->grant);
327 if (!list_empty_careful(&head->waiters)) {
328 spin_lock(&head->lock);
329 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
330 free_bytes < *need_bytes) {
331 error = xlog_grant_head_wait(log, head, tic,
334 spin_unlock(&head->lock);
335 } else if (free_bytes < *need_bytes) {
336 spin_lock(&head->lock);
337 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
338 spin_unlock(&head->lock);
345 xlog_tic_reset_res(xlog_ticket_t *tic)
348 tic->t_res_arr_sum = 0;
349 tic->t_res_num_ophdrs = 0;
353 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
355 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
356 /* add to overflow and start again */
357 tic->t_res_o_flow += tic->t_res_arr_sum;
359 tic->t_res_arr_sum = 0;
362 tic->t_res_arr[tic->t_res_num].r_len = len;
363 tic->t_res_arr[tic->t_res_num].r_type = type;
364 tic->t_res_arr_sum += len;
369 * Replenish the byte reservation required by moving the grant write head.
373 struct xfs_mount *mp,
374 struct xlog_ticket *tic)
376 struct xlog *log = mp->m_log;
380 if (XLOG_FORCED_SHUTDOWN(log))
383 XFS_STATS_INC(xs_try_logspace);
386 * This is a new transaction on the ticket, so we need to change the
387 * transaction ID so that the next transaction has a different TID in
388 * the log. Just add one to the existing tid so that we can see chains
389 * of rolling transactions in the log easily.
393 xlog_grant_push_ail(log, tic->t_unit_res);
395 tic->t_curr_res = tic->t_unit_res;
396 xlog_tic_reset_res(tic);
401 trace_xfs_log_regrant(log, tic);
403 error = xlog_grant_head_check(log, &log->l_write_head, tic,
408 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
409 trace_xfs_log_regrant_exit(log, tic);
410 xlog_verify_grant_tail(log);
415 * If we are failing, make sure the ticket doesn't have any current
416 * reservations. We don't want to add this back when the ticket/
417 * transaction gets cancelled.
420 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
425 * Reserve log space and return a ticket corresponding the reservation.
427 * Each reservation is going to reserve extra space for a log record header.
428 * When writes happen to the on-disk log, we don't subtract the length of the
429 * log record header from any reservation. By wasting space in each
430 * reservation, we prevent over allocation problems.
434 struct xfs_mount *mp,
437 struct xlog_ticket **ticp,
442 struct xlog *log = mp->m_log;
443 struct xlog_ticket *tic;
447 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
449 if (XLOG_FORCED_SHUTDOWN(log))
452 XFS_STATS_INC(xs_try_logspace);
454 ASSERT(*ticp == NULL);
455 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
456 KM_SLEEP | KM_MAYFAIL);
460 tic->t_trans_type = t_type;
463 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
466 trace_xfs_log_reserve(log, tic);
468 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
473 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
474 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
475 trace_xfs_log_reserve_exit(log, tic);
476 xlog_verify_grant_tail(log);
481 * If we are failing, make sure the ticket doesn't have any current
482 * reservations. We don't want to add this back when the ticket/
483 * transaction gets cancelled.
486 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
494 * 1. currblock field gets updated at startup and after in-core logs
495 * marked as with WANT_SYNC.
499 * This routine is called when a user of a log manager ticket is done with
500 * the reservation. If the ticket was ever used, then a commit record for
501 * the associated transaction is written out as a log operation header with
502 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
503 * a given ticket. If the ticket was one with a permanent reservation, then
504 * a few operations are done differently. Permanent reservation tickets by
505 * default don't release the reservation. They just commit the current
506 * transaction with the belief that the reservation is still needed. A flag
507 * must be passed in before permanent reservations are actually released.
508 * When these type of tickets are not released, they need to be set into
509 * the inited state again. By doing this, a start record will be written
510 * out when the next write occurs.
514 struct xfs_mount *mp,
515 struct xlog_ticket *ticket,
516 struct xlog_in_core **iclog,
519 struct xlog *log = mp->m_log;
522 if (XLOG_FORCED_SHUTDOWN(log) ||
524 * If nothing was ever written, don't write out commit record.
525 * If we get an error, just continue and give back the log ticket.
527 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
528 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
529 lsn = (xfs_lsn_t) -1;
530 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
531 flags |= XFS_LOG_REL_PERM_RESERV;
536 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
537 (flags & XFS_LOG_REL_PERM_RESERV)) {
538 trace_xfs_log_done_nonperm(log, ticket);
541 * Release ticket if not permanent reservation or a specific
542 * request has been made to release a permanent reservation.
544 xlog_ungrant_log_space(log, ticket);
545 xfs_log_ticket_put(ticket);
547 trace_xfs_log_done_perm(log, ticket);
549 xlog_regrant_reserve_log_space(log, ticket);
550 /* If this ticket was a permanent reservation and we aren't
551 * trying to release it, reset the inited flags; so next time
552 * we write, a start record will be written out.
554 ticket->t_flags |= XLOG_TIC_INITED;
561 * Attaches a new iclog I/O completion callback routine during
562 * transaction commit. If the log is in error state, a non-zero
563 * return code is handed back and the caller is responsible for
564 * executing the callback at an appropriate time.
568 struct xfs_mount *mp,
569 struct xlog_in_core *iclog,
570 xfs_log_callback_t *cb)
574 spin_lock(&iclog->ic_callback_lock);
575 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
577 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
578 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
580 *(iclog->ic_callback_tail) = cb;
581 iclog->ic_callback_tail = &(cb->cb_next);
583 spin_unlock(&iclog->ic_callback_lock);
588 xfs_log_release_iclog(
589 struct xfs_mount *mp,
590 struct xlog_in_core *iclog)
592 if (xlog_state_release_iclog(mp->m_log, iclog)) {
593 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
601 * Mount a log filesystem
603 * mp - ubiquitous xfs mount point structure
604 * log_target - buftarg of on-disk log device
605 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
606 * num_bblocks - Number of BBSIZE blocks in on-disk log
608 * Return error or zero.
613 xfs_buftarg_t *log_target,
614 xfs_daddr_t blk_offset,
620 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
621 xfs_notice(mp, "Mounting V%d Filesystem",
622 XFS_SB_VERSION_NUM(&mp->m_sb));
625 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
626 XFS_SB_VERSION_NUM(&mp->m_sb));
627 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
630 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
631 if (IS_ERR(mp->m_log)) {
632 error = PTR_ERR(mp->m_log);
637 * Validate the given log space and drop a critical message via syslog
638 * if the log size is too small that would lead to some unexpected
639 * situations in transaction log space reservation stage.
641 * Note: we can't just reject the mount if the validation fails. This
642 * would mean that people would have to downgrade their kernel just to
643 * remedy the situation as there is no way to grow the log (short of
644 * black magic surgery with xfs_db).
646 * We can, however, reject mounts for CRC format filesystems, as the
647 * mkfs binary being used to make the filesystem should never create a
648 * filesystem with a log that is too small.
650 min_logfsbs = xfs_log_calc_minimum_size(mp);
652 if (mp->m_sb.sb_logblocks < min_logfsbs) {
654 "Log size %d blocks too small, minimum size is %d blocks",
655 mp->m_sb.sb_logblocks, min_logfsbs);
657 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
659 "Log size %d blocks too large, maximum size is %lld blocks",
660 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
662 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
664 "log size %lld bytes too large, maximum size is %lld bytes",
665 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
670 if (xfs_sb_version_hascrc(&mp->m_sb)) {
671 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
676 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
677 "experienced then please report this message in the bug report.");
681 * Initialize the AIL now we have a log.
683 error = xfs_trans_ail_init(mp);
685 xfs_warn(mp, "AIL initialisation failed: error %d", error);
688 mp->m_log->l_ailp = mp->m_ail;
691 * skip log recovery on a norecovery mount. pretend it all
694 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
695 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
698 mp->m_flags &= ~XFS_MOUNT_RDONLY;
700 error = xlog_recover(mp->m_log);
703 mp->m_flags |= XFS_MOUNT_RDONLY;
705 xfs_warn(mp, "log mount/recovery failed: error %d",
707 goto out_destroy_ail;
711 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
714 goto out_destroy_ail;
716 /* Normal transactions can now occur */
717 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
720 * Now the log has been fully initialised and we know were our
721 * space grant counters are, we can initialise the permanent ticket
722 * needed for delayed logging to work.
724 xlog_cil_init_post_recovery(mp->m_log);
729 xfs_trans_ail_destroy(mp);
731 xlog_dealloc_log(mp->m_log);
737 * Finish the recovery of the file system. This is separate from the
738 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
739 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
742 * If we finish recovery successfully, start the background log work. If we are
743 * not doing recovery, then we have a RO filesystem and we don't need to start
747 xfs_log_mount_finish(xfs_mount_t *mp)
751 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
752 error = xlog_recover_finish(mp->m_log);
754 xfs_log_work_queue(mp);
756 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
764 * Final log writes as part of unmount.
766 * Mark the filesystem clean as unmount happens. Note that during relocation
767 * this routine needs to be executed as part of source-bag while the
768 * deallocation must not be done until source-end.
772 * Unmount record used to have a string "Unmount filesystem--" in the
773 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
774 * We just write the magic number now since that particular field isn't
775 * currently architecture converted and "Unmount" is a bit foo.
776 * As far as I know, there weren't any dependencies on the old behaviour.
780 xfs_log_unmount_write(xfs_mount_t *mp)
782 struct xlog *log = mp->m_log;
783 xlog_in_core_t *iclog;
785 xlog_in_core_t *first_iclog;
787 xlog_ticket_t *tic = NULL;
792 * Don't write out unmount record on read-only mounts.
793 * Or, if we are doing a forced umount (typically because of IO errors).
795 if (mp->m_flags & XFS_MOUNT_RDONLY)
798 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
799 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
802 first_iclog = iclog = log->l_iclog;
804 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
805 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
806 ASSERT(iclog->ic_offset == 0);
808 iclog = iclog->ic_next;
809 } while (iclog != first_iclog);
811 if (! (XLOG_FORCED_SHUTDOWN(log))) {
812 error = xfs_log_reserve(mp, 600, 1, &tic,
813 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
815 /* the data section must be 32 bit size aligned */
819 __uint32_t pad2; /* may as well make it 64 bits */
821 .magic = XLOG_UNMOUNT_TYPE,
823 struct xfs_log_iovec reg = {
825 .i_len = sizeof(magic),
826 .i_type = XLOG_REG_TYPE_UNMOUNT,
828 struct xfs_log_vec vec = {
833 /* remove inited flag, and account for space used */
835 tic->t_curr_res -= sizeof(magic);
836 error = xlog_write(log, &vec, tic, &lsn,
837 NULL, XLOG_UNMOUNT_TRANS);
839 * At this point, we're umounting anyway,
840 * so there's no point in transitioning log state
841 * to IOERROR. Just continue...
846 xfs_alert(mp, "%s: unmount record failed", __func__);
849 spin_lock(&log->l_icloglock);
850 iclog = log->l_iclog;
851 atomic_inc(&iclog->ic_refcnt);
852 xlog_state_want_sync(log, iclog);
853 spin_unlock(&log->l_icloglock);
854 error = xlog_state_release_iclog(log, iclog);
856 spin_lock(&log->l_icloglock);
857 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
858 iclog->ic_state == XLOG_STATE_DIRTY)) {
859 if (!XLOG_FORCED_SHUTDOWN(log)) {
860 xlog_wait(&iclog->ic_force_wait,
863 spin_unlock(&log->l_icloglock);
866 spin_unlock(&log->l_icloglock);
869 trace_xfs_log_umount_write(log, tic);
870 xlog_ungrant_log_space(log, tic);
871 xfs_log_ticket_put(tic);
875 * We're already in forced_shutdown mode, couldn't
876 * even attempt to write out the unmount transaction.
878 * Go through the motions of sync'ing and releasing
879 * the iclog, even though no I/O will actually happen,
880 * we need to wait for other log I/Os that may already
881 * be in progress. Do this as a separate section of
882 * code so we'll know if we ever get stuck here that
883 * we're in this odd situation of trying to unmount
884 * a file system that went into forced_shutdown as
885 * the result of an unmount..
887 spin_lock(&log->l_icloglock);
888 iclog = log->l_iclog;
889 atomic_inc(&iclog->ic_refcnt);
891 xlog_state_want_sync(log, iclog);
892 spin_unlock(&log->l_icloglock);
893 error = xlog_state_release_iclog(log, iclog);
895 spin_lock(&log->l_icloglock);
897 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
898 || iclog->ic_state == XLOG_STATE_DIRTY
899 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
901 xlog_wait(&iclog->ic_force_wait,
904 spin_unlock(&log->l_icloglock);
909 } /* xfs_log_unmount_write */
912 * Empty the log for unmount/freeze.
914 * To do this, we first need to shut down the background log work so it is not
915 * trying to cover the log as we clean up. We then need to unpin all objects in
916 * the log so we can then flush them out. Once they have completed their IO and
917 * run the callbacks removing themselves from the AIL, we can write the unmount
922 struct xfs_mount *mp)
924 cancel_delayed_work_sync(&mp->m_log->l_work);
925 xfs_log_force(mp, XFS_LOG_SYNC);
928 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
929 * will push it, xfs_wait_buftarg() will not wait for it. Further,
930 * xfs_buf_iowait() cannot be used because it was pushed with the
931 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
932 * the IO to complete.
934 xfs_ail_push_all_sync(mp->m_ail);
935 xfs_wait_buftarg(mp->m_ddev_targp);
936 xfs_buf_lock(mp->m_sb_bp);
937 xfs_buf_unlock(mp->m_sb_bp);
939 xfs_log_unmount_write(mp);
943 * Shut down and release the AIL and Log.
945 * During unmount, we need to ensure we flush all the dirty metadata objects
946 * from the AIL so that the log is empty before we write the unmount record to
947 * the log. Once this is done, we can tear down the AIL and the log.
951 struct xfs_mount *mp)
955 xfs_trans_ail_destroy(mp);
957 xfs_sysfs_del(&mp->m_log->l_kobj);
959 xlog_dealloc_log(mp->m_log);
964 struct xfs_mount *mp,
965 struct xfs_log_item *item,
967 const struct xfs_item_ops *ops)
969 item->li_mountp = mp;
970 item->li_ailp = mp->m_ail;
971 item->li_type = type;
975 INIT_LIST_HEAD(&item->li_ail);
976 INIT_LIST_HEAD(&item->li_cil);
980 * Wake up processes waiting for log space after we have moved the log tail.
984 struct xfs_mount *mp)
986 struct xlog *log = mp->m_log;
989 if (XLOG_FORCED_SHUTDOWN(log))
992 if (!list_empty_careful(&log->l_write_head.waiters)) {
993 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
995 spin_lock(&log->l_write_head.lock);
996 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
997 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
998 spin_unlock(&log->l_write_head.lock);
1001 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1002 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1004 spin_lock(&log->l_reserve_head.lock);
1005 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1006 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1007 spin_unlock(&log->l_reserve_head.lock);
1012 * Determine if we have a transaction that has gone to disk that needs to be
1013 * covered. To begin the transition to the idle state firstly the log needs to
1014 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1015 * we start attempting to cover the log.
1017 * Only if we are then in a state where covering is needed, the caller is
1018 * informed that dummy transactions are required to move the log into the idle
1021 * If there are any items in the AIl or CIL, then we do not want to attempt to
1022 * cover the log as we may be in a situation where there isn't log space
1023 * available to run a dummy transaction and this can lead to deadlocks when the
1024 * tail of the log is pinned by an item that is modified in the CIL. Hence
1025 * there's no point in running a dummy transaction at this point because we
1026 * can't start trying to idle the log until both the CIL and AIL are empty.
1029 xfs_log_need_covered(xfs_mount_t *mp)
1031 struct xlog *log = mp->m_log;
1034 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1037 if (!xlog_cil_empty(log))
1040 spin_lock(&log->l_icloglock);
1041 switch (log->l_covered_state) {
1042 case XLOG_STATE_COVER_DONE:
1043 case XLOG_STATE_COVER_DONE2:
1044 case XLOG_STATE_COVER_IDLE:
1046 case XLOG_STATE_COVER_NEED:
1047 case XLOG_STATE_COVER_NEED2:
1048 if (xfs_ail_min_lsn(log->l_ailp))
1050 if (!xlog_iclogs_empty(log))
1054 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1055 log->l_covered_state = XLOG_STATE_COVER_DONE;
1057 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1063 spin_unlock(&log->l_icloglock);
1068 * We may be holding the log iclog lock upon entering this routine.
1071 xlog_assign_tail_lsn_locked(
1072 struct xfs_mount *mp)
1074 struct xlog *log = mp->m_log;
1075 struct xfs_log_item *lip;
1078 assert_spin_locked(&mp->m_ail->xa_lock);
1081 * To make sure we always have a valid LSN for the log tail we keep
1082 * track of the last LSN which was committed in log->l_last_sync_lsn,
1083 * and use that when the AIL was empty.
1085 lip = xfs_ail_min(mp->m_ail);
1087 tail_lsn = lip->li_lsn;
1089 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1090 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1091 atomic64_set(&log->l_tail_lsn, tail_lsn);
1096 xlog_assign_tail_lsn(
1097 struct xfs_mount *mp)
1101 spin_lock(&mp->m_ail->xa_lock);
1102 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1103 spin_unlock(&mp->m_ail->xa_lock);
1109 * Return the space in the log between the tail and the head. The head
1110 * is passed in the cycle/bytes formal parms. In the special case where
1111 * the reserve head has wrapped passed the tail, this calculation is no
1112 * longer valid. In this case, just return 0 which means there is no space
1113 * in the log. This works for all places where this function is called
1114 * with the reserve head. Of course, if the write head were to ever
1115 * wrap the tail, we should blow up. Rather than catch this case here,
1116 * we depend on other ASSERTions in other parts of the code. XXXmiken
1118 * This code also handles the case where the reservation head is behind
1119 * the tail. The details of this case are described below, but the end
1120 * result is that we return the size of the log as the amount of space left.
1133 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1134 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1135 tail_bytes = BBTOB(tail_bytes);
1136 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1137 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1138 else if (tail_cycle + 1 < head_cycle)
1140 else if (tail_cycle < head_cycle) {
1141 ASSERT(tail_cycle == (head_cycle - 1));
1142 free_bytes = tail_bytes - head_bytes;
1145 * The reservation head is behind the tail.
1146 * In this case we just want to return the size of the
1147 * log as the amount of space left.
1149 xfs_alert(log->l_mp,
1150 "xlog_space_left: head behind tail\n"
1151 " tail_cycle = %d, tail_bytes = %d\n"
1152 " GH cycle = %d, GH bytes = %d",
1153 tail_cycle, tail_bytes, head_cycle, head_bytes);
1155 free_bytes = log->l_logsize;
1162 * Log function which is called when an io completes.
1164 * The log manager needs its own routine, in order to control what
1165 * happens with the buffer after the write completes.
1168 xlog_iodone(xfs_buf_t *bp)
1170 struct xlog_in_core *iclog = bp->b_fspriv;
1171 struct xlog *l = iclog->ic_log;
1175 * Race to shutdown the filesystem if we see an error.
1177 if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1178 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1179 xfs_buf_ioerror_alert(bp, __func__);
1181 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1183 * This flag will be propagated to the trans-committed
1184 * callback routines to let them know that the log-commit
1187 aborted = XFS_LI_ABORTED;
1188 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1189 aborted = XFS_LI_ABORTED;
1192 /* log I/O is always issued ASYNC */
1193 ASSERT(XFS_BUF_ISASYNC(bp));
1194 xlog_state_done_syncing(iclog, aborted);
1197 * drop the buffer lock now that we are done. Nothing references
1198 * the buffer after this, so an unmount waiting on this lock can now
1199 * tear it down safely. As such, it is unsafe to reference the buffer
1200 * (bp) after the unlock as we could race with it being freed.
1206 * Return size of each in-core log record buffer.
1208 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1210 * If the filesystem blocksize is too large, we may need to choose a
1211 * larger size since the directory code currently logs entire blocks.
1215 xlog_get_iclog_buffer_size(
1216 struct xfs_mount *mp,
1222 if (mp->m_logbufs <= 0)
1223 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1225 log->l_iclog_bufs = mp->m_logbufs;
1228 * Buffer size passed in from mount system call.
1230 if (mp->m_logbsize > 0) {
1231 size = log->l_iclog_size = mp->m_logbsize;
1232 log->l_iclog_size_log = 0;
1234 log->l_iclog_size_log++;
1238 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1239 /* # headers = size / 32k
1240 * one header holds cycles from 32k of data
1243 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1244 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1246 log->l_iclog_hsize = xhdrs << BBSHIFT;
1247 log->l_iclog_heads = xhdrs;
1249 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1250 log->l_iclog_hsize = BBSIZE;
1251 log->l_iclog_heads = 1;
1256 /* All machines use 32kB buffers by default. */
1257 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1258 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1260 /* the default log size is 16k or 32k which is one header sector */
1261 log->l_iclog_hsize = BBSIZE;
1262 log->l_iclog_heads = 1;
1265 /* are we being asked to make the sizes selected above visible? */
1266 if (mp->m_logbufs == 0)
1267 mp->m_logbufs = log->l_iclog_bufs;
1268 if (mp->m_logbsize == 0)
1269 mp->m_logbsize = log->l_iclog_size;
1270 } /* xlog_get_iclog_buffer_size */
1275 struct xfs_mount *mp)
1277 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1278 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1282 * Every sync period we need to unpin all items in the AIL and push them to
1283 * disk. If there is nothing dirty, then we might need to cover the log to
1284 * indicate that the filesystem is idle.
1288 struct work_struct *work)
1290 struct xlog *log = container_of(to_delayed_work(work),
1291 struct xlog, l_work);
1292 struct xfs_mount *mp = log->l_mp;
1294 /* dgc: errors ignored - not fatal and nowhere to report them */
1295 if (xfs_log_need_covered(mp))
1296 xfs_fs_log_dummy(mp);
1298 xfs_log_force(mp, 0);
1300 /* start pushing all the metadata that is currently dirty */
1301 xfs_ail_push_all(mp->m_ail);
1303 /* queue us up again */
1304 xfs_log_work_queue(mp);
1308 * This routine initializes some of the log structure for a given mount point.
1309 * Its primary purpose is to fill in enough, so recovery can occur. However,
1310 * some other stuff may be filled in too.
1312 STATIC struct xlog *
1314 struct xfs_mount *mp,
1315 struct xfs_buftarg *log_target,
1316 xfs_daddr_t blk_offset,
1320 xlog_rec_header_t *head;
1321 xlog_in_core_t **iclogp;
1322 xlog_in_core_t *iclog, *prev_iclog=NULL;
1325 int error = -ENOMEM;
1328 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1330 xfs_warn(mp, "Log allocation failed: No memory!");
1335 log->l_targ = log_target;
1336 log->l_logsize = BBTOB(num_bblks);
1337 log->l_logBBstart = blk_offset;
1338 log->l_logBBsize = num_bblks;
1339 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1340 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1341 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1343 log->l_prev_block = -1;
1344 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1345 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1346 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1347 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1349 xlog_grant_head_init(&log->l_reserve_head);
1350 xlog_grant_head_init(&log->l_write_head);
1352 error = -EFSCORRUPTED;
1353 if (xfs_sb_version_hassector(&mp->m_sb)) {
1354 log2_size = mp->m_sb.sb_logsectlog;
1355 if (log2_size < BBSHIFT) {
1356 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1357 log2_size, BBSHIFT);
1361 log2_size -= BBSHIFT;
1362 if (log2_size > mp->m_sectbb_log) {
1363 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1364 log2_size, mp->m_sectbb_log);
1368 /* for larger sector sizes, must have v2 or external log */
1369 if (log2_size && log->l_logBBstart > 0 &&
1370 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1372 "log sector size (0x%x) invalid for configuration.",
1377 log->l_sectBBsize = 1 << log2_size;
1379 xlog_get_iclog_buffer_size(mp, log);
1382 * Use a NULL block for the extra log buffer used during splits so that
1383 * it will trigger errors if we ever try to do IO on it without first
1384 * having set it up properly.
1387 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1388 BTOBB(log->l_iclog_size), 0);
1393 * The iclogbuf buffer locks are held over IO but we are not going to do
1394 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1395 * when appropriately.
1397 ASSERT(xfs_buf_islocked(bp));
1400 bp->b_iodone = xlog_iodone;
1403 spin_lock_init(&log->l_icloglock);
1404 init_waitqueue_head(&log->l_flush_wait);
1406 iclogp = &log->l_iclog;
1408 * The amount of memory to allocate for the iclog structure is
1409 * rather funky due to the way the structure is defined. It is
1410 * done this way so that we can use different sizes for machines
1411 * with different amounts of memory. See the definition of
1412 * xlog_in_core_t in xfs_log_priv.h for details.
1414 ASSERT(log->l_iclog_size >= 4096);
1415 for (i=0; i < log->l_iclog_bufs; i++) {
1416 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1418 goto out_free_iclog;
1421 iclog->ic_prev = prev_iclog;
1424 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1425 BTOBB(log->l_iclog_size), 0);
1427 goto out_free_iclog;
1429 ASSERT(xfs_buf_islocked(bp));
1432 bp->b_iodone = xlog_iodone;
1434 iclog->ic_data = bp->b_addr;
1436 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1438 head = &iclog->ic_header;
1439 memset(head, 0, sizeof(xlog_rec_header_t));
1440 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1441 head->h_version = cpu_to_be32(
1442 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1443 head->h_size = cpu_to_be32(log->l_iclog_size);
1445 head->h_fmt = cpu_to_be32(XLOG_FMT);
1446 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1448 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1449 iclog->ic_state = XLOG_STATE_ACTIVE;
1450 iclog->ic_log = log;
1451 atomic_set(&iclog->ic_refcnt, 0);
1452 spin_lock_init(&iclog->ic_callback_lock);
1453 iclog->ic_callback_tail = &(iclog->ic_callback);
1454 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1456 init_waitqueue_head(&iclog->ic_force_wait);
1457 init_waitqueue_head(&iclog->ic_write_wait);
1459 iclogp = &iclog->ic_next;
1461 *iclogp = log->l_iclog; /* complete ring */
1462 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1464 error = xlog_cil_init(log);
1466 goto out_free_iclog;
1470 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1471 prev_iclog = iclog->ic_next;
1473 xfs_buf_free(iclog->ic_bp);
1476 spinlock_destroy(&log->l_icloglock);
1477 xfs_buf_free(log->l_xbuf);
1481 return ERR_PTR(error);
1482 } /* xlog_alloc_log */
1486 * Write out the commit record of a transaction associated with the given
1487 * ticket. Return the lsn of the commit record.
1492 struct xlog_ticket *ticket,
1493 struct xlog_in_core **iclog,
1494 xfs_lsn_t *commitlsnp)
1496 struct xfs_mount *mp = log->l_mp;
1498 struct xfs_log_iovec reg = {
1501 .i_type = XLOG_REG_TYPE_COMMIT,
1503 struct xfs_log_vec vec = {
1508 ASSERT_ALWAYS(iclog);
1509 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1512 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1517 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1518 * log space. This code pushes on the lsn which would supposedly free up
1519 * the 25% which we want to leave free. We may need to adopt a policy which
1520 * pushes on an lsn which is further along in the log once we reach the high
1521 * water mark. In this manner, we would be creating a low water mark.
1524 xlog_grant_push_ail(
1528 xfs_lsn_t threshold_lsn = 0;
1529 xfs_lsn_t last_sync_lsn;
1532 int threshold_block;
1533 int threshold_cycle;
1536 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1538 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1539 free_blocks = BTOBBT(free_bytes);
1542 * Set the threshold for the minimum number of free blocks in the
1543 * log to the maximum of what the caller needs, one quarter of the
1544 * log, and 256 blocks.
1546 free_threshold = BTOBB(need_bytes);
1547 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1548 free_threshold = MAX(free_threshold, 256);
1549 if (free_blocks >= free_threshold)
1552 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1554 threshold_block += free_threshold;
1555 if (threshold_block >= log->l_logBBsize) {
1556 threshold_block -= log->l_logBBsize;
1557 threshold_cycle += 1;
1559 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1562 * Don't pass in an lsn greater than the lsn of the last
1563 * log record known to be on disk. Use a snapshot of the last sync lsn
1564 * so that it doesn't change between the compare and the set.
1566 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1567 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1568 threshold_lsn = last_sync_lsn;
1571 * Get the transaction layer to kick the dirty buffers out to
1572 * disk asynchronously. No point in trying to do this if
1573 * the filesystem is shutting down.
1575 if (!XLOG_FORCED_SHUTDOWN(log))
1576 xfs_ail_push(log->l_ailp, threshold_lsn);
1580 * Stamp cycle number in every block
1585 struct xlog_in_core *iclog,
1589 int size = iclog->ic_offset + roundoff;
1593 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1595 dp = iclog->ic_datap;
1596 for (i = 0; i < BTOBB(size); i++) {
1597 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1599 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1600 *(__be32 *)dp = cycle_lsn;
1604 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1605 xlog_in_core_2_t *xhdr = iclog->ic_data;
1607 for ( ; i < BTOBB(size); i++) {
1608 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1609 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1610 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1611 *(__be32 *)dp = cycle_lsn;
1615 for (i = 1; i < log->l_iclog_heads; i++)
1616 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1621 * Calculate the checksum for a log buffer.
1623 * This is a little more complicated than it should be because the various
1624 * headers and the actual data are non-contiguous.
1629 struct xlog_rec_header *rhead,
1635 /* first generate the crc for the record header ... */
1636 crc = xfs_start_cksum((char *)rhead,
1637 sizeof(struct xlog_rec_header),
1638 offsetof(struct xlog_rec_header, h_crc));
1640 /* ... then for additional cycle data for v2 logs ... */
1641 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1642 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1645 for (i = 1; i < log->l_iclog_heads; i++) {
1646 crc = crc32c(crc, &xhdr[i].hic_xheader,
1647 sizeof(struct xlog_rec_ext_header));
1651 /* ... and finally for the payload */
1652 crc = crc32c(crc, dp, size);
1654 return xfs_end_cksum(crc);
1658 * The bdstrat callback function for log bufs. This gives us a central
1659 * place to trap bufs in case we get hit by a log I/O error and need to
1660 * shutdown. Actually, in practice, even when we didn't get a log error,
1661 * we transition the iclogs to IOERROR state *after* flushing all existing
1662 * iclogs to disk. This is because we don't want anymore new transactions to be
1663 * started or completed afterwards.
1665 * We lock the iclogbufs here so that we can serialise against IO completion
1666 * during unmount. We might be processing a shutdown triggered during unmount,
1667 * and that can occur asynchronously to the unmount thread, and hence we need to
1668 * ensure that completes before tearing down the iclogbufs. Hence we need to
1669 * hold the buffer lock across the log IO to acheive that.
1675 struct xlog_in_core *iclog = bp->b_fspriv;
1678 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1679 xfs_buf_ioerror(bp, -EIO);
1683 * It would seem logical to return EIO here, but we rely on
1684 * the log state machine to propagate I/O errors instead of
1685 * doing it here. Similarly, IO completion will unlock the
1686 * buffer, so we don't do it here.
1696 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1697 * fashion. Previously, we should have moved the current iclog
1698 * ptr in the log to point to the next available iclog. This allows further
1699 * write to continue while this code syncs out an iclog ready to go.
1700 * Before an in-core log can be written out, the data section must be scanned
1701 * to save away the 1st word of each BBSIZE block into the header. We replace
1702 * it with the current cycle count. Each BBSIZE block is tagged with the
1703 * cycle count because there in an implicit assumption that drives will
1704 * guarantee that entire 512 byte blocks get written at once. In other words,
1705 * we can't have part of a 512 byte block written and part not written. By
1706 * tagging each block, we will know which blocks are valid when recovering
1707 * after an unclean shutdown.
1709 * This routine is single threaded on the iclog. No other thread can be in
1710 * this routine with the same iclog. Changing contents of iclog can there-
1711 * fore be done without grabbing the state machine lock. Updating the global
1712 * log will require grabbing the lock though.
1714 * The entire log manager uses a logical block numbering scheme. Only
1715 * log_sync (and then only bwrite()) know about the fact that the log may
1716 * not start with block zero on a given device. The log block start offset
1717 * is added immediately before calling bwrite().
1723 struct xlog_in_core *iclog)
1727 uint count; /* byte count of bwrite */
1728 uint count_init; /* initial count before roundup */
1729 int roundoff; /* roundoff to BB or stripe */
1730 int split = 0; /* split write into two regions */
1732 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1735 XFS_STATS_INC(xs_log_writes);
1736 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1738 /* Add for LR header */
1739 count_init = log->l_iclog_hsize + iclog->ic_offset;
1741 /* Round out the log write size */
1742 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1743 /* we have a v2 stripe unit to use */
1744 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1746 count = BBTOB(BTOBB(count_init));
1748 roundoff = count - count_init;
1749 ASSERT(roundoff >= 0);
1750 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1751 roundoff < log->l_mp->m_sb.sb_logsunit)
1753 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1754 roundoff < BBTOB(1)));
1756 /* move grant heads by roundoff in sync */
1757 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1758 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1760 /* put cycle number in every block */
1761 xlog_pack_data(log, iclog, roundoff);
1763 /* real byte length */
1764 size = iclog->ic_offset;
1767 iclog->ic_header.h_len = cpu_to_be32(size);
1770 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1772 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1774 /* Do we need to split this write into 2 parts? */
1775 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1778 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1779 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1780 iclog->ic_bwritecnt = 2;
1783 * Bump the cycle numbers at the start of each block in the
1784 * part of the iclog that ends up in the buffer that gets
1785 * written to the start of the log.
1787 * Watch out for the header magic number case, though.
1789 dptr = (char *)&iclog->ic_header + count;
1790 for (i = 0; i < split; i += BBSIZE) {
1791 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1792 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1794 *(__be32 *)dptr = cpu_to_be32(cycle);
1799 iclog->ic_bwritecnt = 1;
1802 /* calculcate the checksum */
1803 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1804 iclog->ic_datap, size);
1806 bp->b_io_length = BTOBB(count);
1807 bp->b_fspriv = iclog;
1808 XFS_BUF_ZEROFLAGS(bp);
1810 bp->b_flags |= XBF_SYNCIO;
1811 /* use high priority completion wq */
1812 bp->b_ioend_wq = log->l_mp->m_log_workqueue;
1814 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1815 bp->b_flags |= XBF_FUA;
1818 * Flush the data device before flushing the log to make
1819 * sure all meta data written back from the AIL actually made
1820 * it to disk before stamping the new log tail LSN into the
1821 * log buffer. For an external log we need to issue the
1822 * flush explicitly, and unfortunately synchronously here;
1823 * for an internal log we can simply use the block layer
1824 * state machine for preflushes.
1826 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1827 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1829 bp->b_flags |= XBF_FLUSH;
1832 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1833 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1835 xlog_verify_iclog(log, iclog, count, true);
1837 /* account for log which doesn't start at block #0 */
1838 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1840 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1845 error = xlog_bdstrat(bp);
1847 xfs_buf_ioerror_alert(bp, "xlog_sync");
1851 bp = iclog->ic_log->l_xbuf;
1852 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1853 xfs_buf_associate_memory(bp,
1854 (char *)&iclog->ic_header + count, split);
1855 bp->b_fspriv = iclog;
1856 XFS_BUF_ZEROFLAGS(bp);
1858 bp->b_flags |= XBF_SYNCIO;
1859 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1860 bp->b_flags |= XBF_FUA;
1861 /* use high priority completion wq */
1862 bp->b_ioend_wq = log->l_mp->m_log_workqueue;
1864 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1865 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1867 /* account for internal log which doesn't start at block #0 */
1868 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1870 error = xlog_bdstrat(bp);
1872 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1880 * Deallocate a log structure
1886 xlog_in_core_t *iclog, *next_iclog;
1889 xlog_cil_destroy(log);
1892 * Cycle all the iclogbuf locks to make sure all log IO completion
1893 * is done before we tear down these buffers.
1895 iclog = log->l_iclog;
1896 for (i = 0; i < log->l_iclog_bufs; i++) {
1897 xfs_buf_lock(iclog->ic_bp);
1898 xfs_buf_unlock(iclog->ic_bp);
1899 iclog = iclog->ic_next;
1903 * Always need to ensure that the extra buffer does not point to memory
1904 * owned by another log buffer before we free it. Also, cycle the lock
1905 * first to ensure we've completed IO on it.
1907 xfs_buf_lock(log->l_xbuf);
1908 xfs_buf_unlock(log->l_xbuf);
1909 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1910 xfs_buf_free(log->l_xbuf);
1912 iclog = log->l_iclog;
1913 for (i = 0; i < log->l_iclog_bufs; i++) {
1914 xfs_buf_free(iclog->ic_bp);
1915 next_iclog = iclog->ic_next;
1919 spinlock_destroy(&log->l_icloglock);
1921 log->l_mp->m_log = NULL;
1923 } /* xlog_dealloc_log */
1926 * Update counters atomically now that memcpy is done.
1930 xlog_state_finish_copy(
1932 struct xlog_in_core *iclog,
1936 spin_lock(&log->l_icloglock);
1938 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1939 iclog->ic_offset += copy_bytes;
1941 spin_unlock(&log->l_icloglock);
1942 } /* xlog_state_finish_copy */
1948 * print out info relating to regions written which consume
1953 struct xfs_mount *mp,
1954 struct xlog_ticket *ticket)
1957 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1959 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1960 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1981 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2025 "xlog_write: reservation summary:\n"
2026 " trans type = %s (%u)\n"
2027 " unit res = %d bytes\n"
2028 " current res = %d bytes\n"
2029 " total reg = %u bytes (o/flow = %u bytes)\n"
2030 " ophdrs = %u (ophdr space = %u bytes)\n"
2031 " ophdr + reg = %u bytes\n"
2032 " num regions = %u\n",
2033 ((ticket->t_trans_type <= 0 ||
2034 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2035 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2036 ticket->t_trans_type,
2039 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2040 ticket->t_res_num_ophdrs, ophdr_spc,
2041 ticket->t_res_arr_sum +
2042 ticket->t_res_o_flow + ophdr_spc,
2045 for (i = 0; i < ticket->t_res_num; i++) {
2046 uint r_type = ticket->t_res_arr[i].r_type;
2047 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2048 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2049 "bad-rtype" : res_type_str[r_type-1]),
2050 ticket->t_res_arr[i].r_len);
2053 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2054 "xlog_write: reservation ran out. Need to up reservation");
2055 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2059 * Calculate the potential space needed by the log vector. Each region gets
2060 * its own xlog_op_header_t and may need to be double word aligned.
2063 xlog_write_calc_vec_length(
2064 struct xlog_ticket *ticket,
2065 struct xfs_log_vec *log_vector)
2067 struct xfs_log_vec *lv;
2072 /* acct for start rec of xact */
2073 if (ticket->t_flags & XLOG_TIC_INITED)
2076 for (lv = log_vector; lv; lv = lv->lv_next) {
2077 /* we don't write ordered log vectors */
2078 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2081 headers += lv->lv_niovecs;
2083 for (i = 0; i < lv->lv_niovecs; i++) {
2084 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2087 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2091 ticket->t_res_num_ophdrs += headers;
2092 len += headers * sizeof(struct xlog_op_header);
2098 * If first write for transaction, insert start record We can't be trying to
2099 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2102 xlog_write_start_rec(
2103 struct xlog_op_header *ophdr,
2104 struct xlog_ticket *ticket)
2106 if (!(ticket->t_flags & XLOG_TIC_INITED))
2109 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2110 ophdr->oh_clientid = ticket->t_clientid;
2112 ophdr->oh_flags = XLOG_START_TRANS;
2115 ticket->t_flags &= ~XLOG_TIC_INITED;
2117 return sizeof(struct xlog_op_header);
2120 static xlog_op_header_t *
2121 xlog_write_setup_ophdr(
2123 struct xlog_op_header *ophdr,
2124 struct xlog_ticket *ticket,
2127 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2128 ophdr->oh_clientid = ticket->t_clientid;
2131 /* are we copying a commit or unmount record? */
2132 ophdr->oh_flags = flags;
2135 * We've seen logs corrupted with bad transaction client ids. This
2136 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2137 * and shut down the filesystem.
2139 switch (ophdr->oh_clientid) {
2140 case XFS_TRANSACTION:
2146 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2147 ophdr->oh_clientid, ticket);
2155 * Set up the parameters of the region copy into the log. This has
2156 * to handle region write split across multiple log buffers - this
2157 * state is kept external to this function so that this code can
2158 * be written in an obvious, self documenting manner.
2161 xlog_write_setup_copy(
2162 struct xlog_ticket *ticket,
2163 struct xlog_op_header *ophdr,
2164 int space_available,
2168 int *last_was_partial_copy,
2169 int *bytes_consumed)
2173 still_to_copy = space_required - *bytes_consumed;
2174 *copy_off = *bytes_consumed;
2176 if (still_to_copy <= space_available) {
2177 /* write of region completes here */
2178 *copy_len = still_to_copy;
2179 ophdr->oh_len = cpu_to_be32(*copy_len);
2180 if (*last_was_partial_copy)
2181 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2182 *last_was_partial_copy = 0;
2183 *bytes_consumed = 0;
2187 /* partial write of region, needs extra log op header reservation */
2188 *copy_len = space_available;
2189 ophdr->oh_len = cpu_to_be32(*copy_len);
2190 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2191 if (*last_was_partial_copy)
2192 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2193 *bytes_consumed += *copy_len;
2194 (*last_was_partial_copy)++;
2196 /* account for new log op header */
2197 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2198 ticket->t_res_num_ophdrs++;
2200 return sizeof(struct xlog_op_header);
2204 xlog_write_copy_finish(
2206 struct xlog_in_core *iclog,
2211 int *partial_copy_len,
2213 struct xlog_in_core **commit_iclog)
2215 if (*partial_copy) {
2217 * This iclog has already been marked WANT_SYNC by
2218 * xlog_state_get_iclog_space.
2220 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2223 return xlog_state_release_iclog(log, iclog);
2227 *partial_copy_len = 0;
2229 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2230 /* no more space in this iclog - push it. */
2231 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2235 spin_lock(&log->l_icloglock);
2236 xlog_state_want_sync(log, iclog);
2237 spin_unlock(&log->l_icloglock);
2240 return xlog_state_release_iclog(log, iclog);
2241 ASSERT(flags & XLOG_COMMIT_TRANS);
2242 *commit_iclog = iclog;
2249 * Write some region out to in-core log
2251 * This will be called when writing externally provided regions or when
2252 * writing out a commit record for a given transaction.
2254 * General algorithm:
2255 * 1. Find total length of this write. This may include adding to the
2256 * lengths passed in.
2257 * 2. Check whether we violate the tickets reservation.
2258 * 3. While writing to this iclog
2259 * A. Reserve as much space in this iclog as can get
2260 * B. If this is first write, save away start lsn
2261 * C. While writing this region:
2262 * 1. If first write of transaction, write start record
2263 * 2. Write log operation header (header per region)
2264 * 3. Find out if we can fit entire region into this iclog
2265 * 4. Potentially, verify destination memcpy ptr
2266 * 5. Memcpy (partial) region
2267 * 6. If partial copy, release iclog; otherwise, continue
2268 * copying more regions into current iclog
2269 * 4. Mark want sync bit (in simulation mode)
2270 * 5. Release iclog for potential flush to on-disk log.
2273 * 1. Panic if reservation is overrun. This should never happen since
2274 * reservation amounts are generated internal to the filesystem.
2276 * 1. Tickets are single threaded data structures.
2277 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2278 * syncing routine. When a single log_write region needs to span
2279 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2280 * on all log operation writes which don't contain the end of the
2281 * region. The XLOG_END_TRANS bit is used for the in-core log
2282 * operation which contains the end of the continued log_write region.
2283 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2284 * we don't really know exactly how much space will be used. As a result,
2285 * we don't update ic_offset until the end when we know exactly how many
2286 * bytes have been written out.
2291 struct xfs_log_vec *log_vector,
2292 struct xlog_ticket *ticket,
2293 xfs_lsn_t *start_lsn,
2294 struct xlog_in_core **commit_iclog,
2297 struct xlog_in_core *iclog = NULL;
2298 struct xfs_log_iovec *vecp;
2299 struct xfs_log_vec *lv;
2302 int partial_copy = 0;
2303 int partial_copy_len = 0;
2311 len = xlog_write_calc_vec_length(ticket, log_vector);
2314 * Region headers and bytes are already accounted for.
2315 * We only need to take into account start records and
2316 * split regions in this function.
2318 if (ticket->t_flags & XLOG_TIC_INITED)
2319 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2322 * Commit record headers need to be accounted for. These
2323 * come in as separate writes so are easy to detect.
2325 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2326 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2328 if (ticket->t_curr_res < 0)
2329 xlog_print_tic_res(log->l_mp, ticket);
2333 vecp = lv->lv_iovecp;
2334 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2338 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2339 &contwr, &log_offset);
2343 ASSERT(log_offset <= iclog->ic_size - 1);
2344 ptr = iclog->ic_datap + log_offset;
2346 /* start_lsn is the first lsn written to. That's all we need. */
2348 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2351 * This loop writes out as many regions as can fit in the amount
2352 * of space which was allocated by xlog_state_get_iclog_space().
2354 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2355 struct xfs_log_iovec *reg;
2356 struct xlog_op_header *ophdr;
2360 bool ordered = false;
2362 /* ordered log vectors have no regions to write */
2363 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2364 ASSERT(lv->lv_niovecs == 0);
2370 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2371 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2373 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2374 if (start_rec_copy) {
2376 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2380 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2384 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2385 sizeof(struct xlog_op_header));
2387 len += xlog_write_setup_copy(ticket, ophdr,
2388 iclog->ic_size-log_offset,
2390 ©_off, ©_len,
2393 xlog_verify_dest_ptr(log, ptr);
2396 ASSERT(copy_len >= 0);
2397 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2398 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2400 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2402 data_cnt += contwr ? copy_len : 0;
2404 error = xlog_write_copy_finish(log, iclog, flags,
2405 &record_cnt, &data_cnt,
2414 * if we had a partial copy, we need to get more iclog
2415 * space but we don't want to increment the region
2416 * index because there is still more is this region to
2419 * If we completed writing this region, and we flushed
2420 * the iclog (indicated by resetting of the record
2421 * count), then we also need to get more log space. If
2422 * this was the last record, though, we are done and
2428 if (++index == lv->lv_niovecs) {
2433 vecp = lv->lv_iovecp;
2435 if (record_cnt == 0 && ordered == false) {
2445 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2447 return xlog_state_release_iclog(log, iclog);
2449 ASSERT(flags & XLOG_COMMIT_TRANS);
2450 *commit_iclog = iclog;
2455 /*****************************************************************************
2457 * State Machine functions
2459 *****************************************************************************
2462 /* Clean iclogs starting from the head. This ordering must be
2463 * maintained, so an iclog doesn't become ACTIVE beyond one that
2464 * is SYNCING. This is also required to maintain the notion that we use
2465 * a ordered wait queue to hold off would be writers to the log when every
2466 * iclog is trying to sync to disk.
2468 * State Change: DIRTY -> ACTIVE
2471 xlog_state_clean_log(
2474 xlog_in_core_t *iclog;
2477 iclog = log->l_iclog;
2479 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2480 iclog->ic_state = XLOG_STATE_ACTIVE;
2481 iclog->ic_offset = 0;
2482 ASSERT(iclog->ic_callback == NULL);
2484 * If the number of ops in this iclog indicate it just
2485 * contains the dummy transaction, we can
2486 * change state into IDLE (the second time around).
2487 * Otherwise we should change the state into
2489 * We don't need to cover the dummy.
2492 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2497 * We have two dirty iclogs so start over
2498 * This could also be num of ops indicates
2499 * this is not the dummy going out.
2503 iclog->ic_header.h_num_logops = 0;
2504 memset(iclog->ic_header.h_cycle_data, 0,
2505 sizeof(iclog->ic_header.h_cycle_data));
2506 iclog->ic_header.h_lsn = 0;
2507 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2510 break; /* stop cleaning */
2511 iclog = iclog->ic_next;
2512 } while (iclog != log->l_iclog);
2514 /* log is locked when we are called */
2516 * Change state for the dummy log recording.
2517 * We usually go to NEED. But we go to NEED2 if the changed indicates
2518 * we are done writing the dummy record.
2519 * If we are done with the second dummy recored (DONE2), then
2523 switch (log->l_covered_state) {
2524 case XLOG_STATE_COVER_IDLE:
2525 case XLOG_STATE_COVER_NEED:
2526 case XLOG_STATE_COVER_NEED2:
2527 log->l_covered_state = XLOG_STATE_COVER_NEED;
2530 case XLOG_STATE_COVER_DONE:
2532 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2534 log->l_covered_state = XLOG_STATE_COVER_NEED;
2537 case XLOG_STATE_COVER_DONE2:
2539 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2541 log->l_covered_state = XLOG_STATE_COVER_NEED;
2548 } /* xlog_state_clean_log */
2551 xlog_get_lowest_lsn(
2554 xlog_in_core_t *lsn_log;
2555 xfs_lsn_t lowest_lsn, lsn;
2557 lsn_log = log->l_iclog;
2560 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2561 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2562 if ((lsn && !lowest_lsn) ||
2563 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2567 lsn_log = lsn_log->ic_next;
2568 } while (lsn_log != log->l_iclog);
2574 xlog_state_do_callback(
2577 struct xlog_in_core *ciclog)
2579 xlog_in_core_t *iclog;
2580 xlog_in_core_t *first_iclog; /* used to know when we've
2581 * processed all iclogs once */
2582 xfs_log_callback_t *cb, *cb_next;
2584 xfs_lsn_t lowest_lsn;
2585 int ioerrors; /* counter: iclogs with errors */
2586 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2587 int funcdidcallbacks; /* flag: function did callbacks */
2588 int repeats; /* for issuing console warnings if
2589 * looping too many times */
2592 spin_lock(&log->l_icloglock);
2593 first_iclog = iclog = log->l_iclog;
2595 funcdidcallbacks = 0;
2600 * Scan all iclogs starting with the one pointed to by the
2601 * log. Reset this starting point each time the log is
2602 * unlocked (during callbacks).
2604 * Keep looping through iclogs until one full pass is made
2605 * without running any callbacks.
2607 first_iclog = log->l_iclog;
2608 iclog = log->l_iclog;
2609 loopdidcallbacks = 0;
2614 /* skip all iclogs in the ACTIVE & DIRTY states */
2615 if (iclog->ic_state &
2616 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2617 iclog = iclog->ic_next;
2622 * Between marking a filesystem SHUTDOWN and stopping
2623 * the log, we do flush all iclogs to disk (if there
2624 * wasn't a log I/O error). So, we do want things to
2625 * go smoothly in case of just a SHUTDOWN w/o a
2628 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2630 * Can only perform callbacks in order. Since
2631 * this iclog is not in the DONE_SYNC/
2632 * DO_CALLBACK state, we skip the rest and
2633 * just try to clean up. If we set our iclog
2634 * to DO_CALLBACK, we will not process it when
2635 * we retry since a previous iclog is in the
2636 * CALLBACK and the state cannot change since
2637 * we are holding the l_icloglock.
2639 if (!(iclog->ic_state &
2640 (XLOG_STATE_DONE_SYNC |
2641 XLOG_STATE_DO_CALLBACK))) {
2642 if (ciclog && (ciclog->ic_state ==
2643 XLOG_STATE_DONE_SYNC)) {
2644 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2649 * We now have an iclog that is in either the
2650 * DO_CALLBACK or DONE_SYNC states. The other
2651 * states (WANT_SYNC, SYNCING, or CALLBACK were
2652 * caught by the above if and are going to
2653 * clean (i.e. we aren't doing their callbacks)
2658 * We will do one more check here to see if we
2659 * have chased our tail around.
2662 lowest_lsn = xlog_get_lowest_lsn(log);
2664 XFS_LSN_CMP(lowest_lsn,
2665 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2666 iclog = iclog->ic_next;
2667 continue; /* Leave this iclog for
2671 iclog->ic_state = XLOG_STATE_CALLBACK;
2675 * Completion of a iclog IO does not imply that
2676 * a transaction has completed, as transactions
2677 * can be large enough to span many iclogs. We
2678 * cannot change the tail of the log half way
2679 * through a transaction as this may be the only
2680 * transaction in the log and moving th etail to
2681 * point to the middle of it will prevent
2682 * recovery from finding the start of the
2683 * transaction. Hence we should only update the
2684 * last_sync_lsn if this iclog contains
2685 * transaction completion callbacks on it.
2687 * We have to do this before we drop the
2688 * icloglock to ensure we are the only one that
2691 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2692 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2693 if (iclog->ic_callback)
2694 atomic64_set(&log->l_last_sync_lsn,
2695 be64_to_cpu(iclog->ic_header.h_lsn));
2700 spin_unlock(&log->l_icloglock);
2703 * Keep processing entries in the callback list until
2704 * we come around and it is empty. We need to
2705 * atomically see that the list is empty and change the
2706 * state to DIRTY so that we don't miss any more
2707 * callbacks being added.
2709 spin_lock(&iclog->ic_callback_lock);
2710 cb = iclog->ic_callback;
2712 iclog->ic_callback_tail = &(iclog->ic_callback);
2713 iclog->ic_callback = NULL;
2714 spin_unlock(&iclog->ic_callback_lock);
2716 /* perform callbacks in the order given */
2717 for (; cb; cb = cb_next) {
2718 cb_next = cb->cb_next;
2719 cb->cb_func(cb->cb_arg, aborted);
2721 spin_lock(&iclog->ic_callback_lock);
2722 cb = iclog->ic_callback;
2728 spin_lock(&log->l_icloglock);
2729 ASSERT(iclog->ic_callback == NULL);
2730 spin_unlock(&iclog->ic_callback_lock);
2731 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2732 iclog->ic_state = XLOG_STATE_DIRTY;
2735 * Transition from DIRTY to ACTIVE if applicable.
2736 * NOP if STATE_IOERROR.
2738 xlog_state_clean_log(log);
2740 /* wake up threads waiting in xfs_log_force() */
2741 wake_up_all(&iclog->ic_force_wait);
2743 iclog = iclog->ic_next;
2744 } while (first_iclog != iclog);
2746 if (repeats > 5000) {
2747 flushcnt += repeats;
2750 "%s: possible infinite loop (%d iterations)",
2751 __func__, flushcnt);
2753 } while (!ioerrors && loopdidcallbacks);
2756 * make one last gasp attempt to see if iclogs are being left in
2760 if (funcdidcallbacks) {
2761 first_iclog = iclog = log->l_iclog;
2763 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2765 * Terminate the loop if iclogs are found in states
2766 * which will cause other threads to clean up iclogs.
2768 * SYNCING - i/o completion will go through logs
2769 * DONE_SYNC - interrupt thread should be waiting for
2771 * IOERROR - give up hope all ye who enter here
2773 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2774 iclog->ic_state == XLOG_STATE_SYNCING ||
2775 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2776 iclog->ic_state == XLOG_STATE_IOERROR )
2778 iclog = iclog->ic_next;
2779 } while (first_iclog != iclog);
2783 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2785 spin_unlock(&log->l_icloglock);
2788 wake_up_all(&log->l_flush_wait);
2793 * Finish transitioning this iclog to the dirty state.
2795 * Make sure that we completely execute this routine only when this is
2796 * the last call to the iclog. There is a good chance that iclog flushes,
2797 * when we reach the end of the physical log, get turned into 2 separate
2798 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2799 * routine. By using the reference count bwritecnt, we guarantee that only
2800 * the second completion goes through.
2802 * Callbacks could take time, so they are done outside the scope of the
2803 * global state machine log lock.
2806 xlog_state_done_syncing(
2807 xlog_in_core_t *iclog,
2810 struct xlog *log = iclog->ic_log;
2812 spin_lock(&log->l_icloglock);
2814 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2815 iclog->ic_state == XLOG_STATE_IOERROR);
2816 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2817 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2821 * If we got an error, either on the first buffer, or in the case of
2822 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2823 * and none should ever be attempted to be written to disk
2826 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2827 if (--iclog->ic_bwritecnt == 1) {
2828 spin_unlock(&log->l_icloglock);
2831 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2835 * Someone could be sleeping prior to writing out the next
2836 * iclog buffer, we wake them all, one will get to do the
2837 * I/O, the others get to wait for the result.
2839 wake_up_all(&iclog->ic_write_wait);
2840 spin_unlock(&log->l_icloglock);
2841 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2842 } /* xlog_state_done_syncing */
2846 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2847 * sleep. We wait on the flush queue on the head iclog as that should be
2848 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2849 * we will wait here and all new writes will sleep until a sync completes.
2851 * The in-core logs are used in a circular fashion. They are not used
2852 * out-of-order even when an iclog past the head is free.
2855 * * log_offset where xlog_write() can start writing into the in-core
2857 * * in-core log pointer to which xlog_write() should write.
2858 * * boolean indicating this is a continued write to an in-core log.
2859 * If this is the last write, then the in-core log's offset field
2860 * needs to be incremented, depending on the amount of data which
2864 xlog_state_get_iclog_space(
2867 struct xlog_in_core **iclogp,
2868 struct xlog_ticket *ticket,
2869 int *continued_write,
2873 xlog_rec_header_t *head;
2874 xlog_in_core_t *iclog;
2878 spin_lock(&log->l_icloglock);
2879 if (XLOG_FORCED_SHUTDOWN(log)) {
2880 spin_unlock(&log->l_icloglock);
2884 iclog = log->l_iclog;
2885 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2886 XFS_STATS_INC(xs_log_noiclogs);
2888 /* Wait for log writes to have flushed */
2889 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2893 head = &iclog->ic_header;
2895 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2896 log_offset = iclog->ic_offset;
2898 /* On the 1st write to an iclog, figure out lsn. This works
2899 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2900 * committing to. If the offset is set, that's how many blocks
2903 if (log_offset == 0) {
2904 ticket->t_curr_res -= log->l_iclog_hsize;
2905 xlog_tic_add_region(ticket,
2907 XLOG_REG_TYPE_LRHEADER);
2908 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2909 head->h_lsn = cpu_to_be64(
2910 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2911 ASSERT(log->l_curr_block >= 0);
2914 /* If there is enough room to write everything, then do it. Otherwise,
2915 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2916 * bit is on, so this will get flushed out. Don't update ic_offset
2917 * until you know exactly how many bytes get copied. Therefore, wait
2918 * until later to update ic_offset.
2920 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2921 * can fit into remaining data section.
2923 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2924 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2927 * If I'm the only one writing to this iclog, sync it to disk.
2928 * We need to do an atomic compare and decrement here to avoid
2929 * racing with concurrent atomic_dec_and_lock() calls in
2930 * xlog_state_release_iclog() when there is more than one
2931 * reference to the iclog.
2933 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2934 /* we are the only one */
2935 spin_unlock(&log->l_icloglock);
2936 error = xlog_state_release_iclog(log, iclog);
2940 spin_unlock(&log->l_icloglock);
2945 /* Do we have enough room to write the full amount in the remainder
2946 * of this iclog? Or must we continue a write on the next iclog and
2947 * mark this iclog as completely taken? In the case where we switch
2948 * iclogs (to mark it taken), this particular iclog will release/sync
2949 * to disk in xlog_write().
2951 if (len <= iclog->ic_size - iclog->ic_offset) {
2952 *continued_write = 0;
2953 iclog->ic_offset += len;
2955 *continued_write = 1;
2956 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2960 ASSERT(iclog->ic_offset <= iclog->ic_size);
2961 spin_unlock(&log->l_icloglock);
2963 *logoffsetp = log_offset;
2965 } /* xlog_state_get_iclog_space */
2967 /* The first cnt-1 times through here we don't need to
2968 * move the grant write head because the permanent
2969 * reservation has reserved cnt times the unit amount.
2970 * Release part of current permanent unit reservation and
2971 * reset current reservation to be one units worth. Also
2972 * move grant reservation head forward.
2975 xlog_regrant_reserve_log_space(
2977 struct xlog_ticket *ticket)
2979 trace_xfs_log_regrant_reserve_enter(log, ticket);
2981 if (ticket->t_cnt > 0)
2984 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2985 ticket->t_curr_res);
2986 xlog_grant_sub_space(log, &log->l_write_head.grant,
2987 ticket->t_curr_res);
2988 ticket->t_curr_res = ticket->t_unit_res;
2989 xlog_tic_reset_res(ticket);
2991 trace_xfs_log_regrant_reserve_sub(log, ticket);
2993 /* just return if we still have some of the pre-reserved space */
2994 if (ticket->t_cnt > 0)
2997 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2998 ticket->t_unit_res);
3000 trace_xfs_log_regrant_reserve_exit(log, ticket);
3002 ticket->t_curr_res = ticket->t_unit_res;
3003 xlog_tic_reset_res(ticket);
3004 } /* xlog_regrant_reserve_log_space */
3008 * Give back the space left from a reservation.
3010 * All the information we need to make a correct determination of space left
3011 * is present. For non-permanent reservations, things are quite easy. The
3012 * count should have been decremented to zero. We only need to deal with the
3013 * space remaining in the current reservation part of the ticket. If the
3014 * ticket contains a permanent reservation, there may be left over space which
3015 * needs to be released. A count of N means that N-1 refills of the current
3016 * reservation can be done before we need to ask for more space. The first
3017 * one goes to fill up the first current reservation. Once we run out of
3018 * space, the count will stay at zero and the only space remaining will be
3019 * in the current reservation field.
3022 xlog_ungrant_log_space(
3024 struct xlog_ticket *ticket)
3028 if (ticket->t_cnt > 0)
3031 trace_xfs_log_ungrant_enter(log, ticket);
3032 trace_xfs_log_ungrant_sub(log, ticket);
3035 * If this is a permanent reservation ticket, we may be able to free
3036 * up more space based on the remaining count.
3038 bytes = ticket->t_curr_res;
3039 if (ticket->t_cnt > 0) {
3040 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3041 bytes += ticket->t_unit_res*ticket->t_cnt;
3044 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3045 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3047 trace_xfs_log_ungrant_exit(log, ticket);
3049 xfs_log_space_wake(log->l_mp);
3053 * Flush iclog to disk if this is the last reference to the given iclog and
3054 * the WANT_SYNC bit is set.
3056 * When this function is entered, the iclog is not necessarily in the
3057 * WANT_SYNC state. It may be sitting around waiting to get filled.
3062 xlog_state_release_iclog(
3064 struct xlog_in_core *iclog)
3066 int sync = 0; /* do we sync? */
3068 if (iclog->ic_state & XLOG_STATE_IOERROR)
3071 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3072 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3075 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3076 spin_unlock(&log->l_icloglock);
3079 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3080 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3082 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3083 /* update tail before writing to iclog */
3084 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3086 iclog->ic_state = XLOG_STATE_SYNCING;
3087 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3088 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3089 /* cycle incremented when incrementing curr_block */
3091 spin_unlock(&log->l_icloglock);
3094 * We let the log lock go, so it's possible that we hit a log I/O
3095 * error or some other SHUTDOWN condition that marks the iclog
3096 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3097 * this iclog has consistent data, so we ignore IOERROR
3098 * flags after this point.
3101 return xlog_sync(log, iclog);
3103 } /* xlog_state_release_iclog */
3107 * This routine will mark the current iclog in the ring as WANT_SYNC
3108 * and move the current iclog pointer to the next iclog in the ring.
3109 * When this routine is called from xlog_state_get_iclog_space(), the
3110 * exact size of the iclog has not yet been determined. All we know is
3111 * that every data block. We have run out of space in this log record.
3114 xlog_state_switch_iclogs(
3116 struct xlog_in_core *iclog,
3119 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3121 eventual_size = iclog->ic_offset;
3122 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3123 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3124 log->l_prev_block = log->l_curr_block;
3125 log->l_prev_cycle = log->l_curr_cycle;
3127 /* roll log?: ic_offset changed later */
3128 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3130 /* Round up to next log-sunit */
3131 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3132 log->l_mp->m_sb.sb_logsunit > 1) {
3133 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3134 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3137 if (log->l_curr_block >= log->l_logBBsize) {
3138 log->l_curr_cycle++;
3139 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3140 log->l_curr_cycle++;
3141 log->l_curr_block -= log->l_logBBsize;
3142 ASSERT(log->l_curr_block >= 0);
3144 ASSERT(iclog == log->l_iclog);
3145 log->l_iclog = iclog->ic_next;
3146 } /* xlog_state_switch_iclogs */
3149 * Write out all data in the in-core log as of this exact moment in time.
3151 * Data may be written to the in-core log during this call. However,
3152 * we don't guarantee this data will be written out. A change from past
3153 * implementation means this routine will *not* write out zero length LRs.
3155 * Basically, we try and perform an intelligent scan of the in-core logs.
3156 * If we determine there is no flushable data, we just return. There is no
3157 * flushable data if:
3159 * 1. the current iclog is active and has no data; the previous iclog
3160 * is in the active or dirty state.
3161 * 2. the current iclog is drity, and the previous iclog is in the
3162 * active or dirty state.
3166 * 1. the current iclog is not in the active nor dirty state.
3167 * 2. the current iclog dirty, and the previous iclog is not in the
3168 * active nor dirty state.
3169 * 3. the current iclog is active, and there is another thread writing
3170 * to this particular iclog.
3171 * 4. a) the current iclog is active and has no other writers
3172 * b) when we return from flushing out this iclog, it is still
3173 * not in the active nor dirty state.
3177 struct xfs_mount *mp,
3181 struct xlog *log = mp->m_log;
3182 struct xlog_in_core *iclog;
3185 XFS_STATS_INC(xs_log_force);
3187 xlog_cil_force(log);
3189 spin_lock(&log->l_icloglock);
3191 iclog = log->l_iclog;
3192 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3193 spin_unlock(&log->l_icloglock);
3197 /* If the head iclog is not active nor dirty, we just attach
3198 * ourselves to the head and go to sleep.
3200 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3201 iclog->ic_state == XLOG_STATE_DIRTY) {
3203 * If the head is dirty or (active and empty), then
3204 * we need to look at the previous iclog. If the previous
3205 * iclog is active or dirty we are done. There is nothing
3206 * to sync out. Otherwise, we attach ourselves to the
3207 * previous iclog and go to sleep.
3209 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3210 (atomic_read(&iclog->ic_refcnt) == 0
3211 && iclog->ic_offset == 0)) {
3212 iclog = iclog->ic_prev;
3213 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3214 iclog->ic_state == XLOG_STATE_DIRTY)
3219 if (atomic_read(&iclog->ic_refcnt) == 0) {
3220 /* We are the only one with access to this
3221 * iclog. Flush it out now. There should
3222 * be a roundoff of zero to show that someone
3223 * has already taken care of the roundoff from
3224 * the previous sync.
3226 atomic_inc(&iclog->ic_refcnt);
3227 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3228 xlog_state_switch_iclogs(log, iclog, 0);
3229 spin_unlock(&log->l_icloglock);
3231 if (xlog_state_release_iclog(log, iclog))
3236 spin_lock(&log->l_icloglock);
3237 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3238 iclog->ic_state != XLOG_STATE_DIRTY)
3243 /* Someone else is writing to this iclog.
3244 * Use its call to flush out the data. However,
3245 * the other thread may not force out this LR,
3246 * so we mark it WANT_SYNC.
3248 xlog_state_switch_iclogs(log, iclog, 0);
3254 /* By the time we come around again, the iclog could've been filled
3255 * which would give it another lsn. If we have a new lsn, just
3256 * return because the relevant data has been flushed.
3259 if (flags & XFS_LOG_SYNC) {
3261 * We must check if we're shutting down here, before
3262 * we wait, while we're holding the l_icloglock.
3263 * Then we check again after waking up, in case our
3264 * sleep was disturbed by a bad news.
3266 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3267 spin_unlock(&log->l_icloglock);
3270 XFS_STATS_INC(xs_log_force_sleep);
3271 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3273 * No need to grab the log lock here since we're
3274 * only deciding whether or not to return EIO
3275 * and the memory read should be atomic.
3277 if (iclog->ic_state & XLOG_STATE_IOERROR)
3284 spin_unlock(&log->l_icloglock);
3290 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3291 * about errors or whether the log was flushed or not. This is the normal
3292 * interface to use when trying to unpin items or move the log forward.
3301 trace_xfs_log_force(mp, 0);
3302 error = _xfs_log_force(mp, flags, NULL);
3304 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3308 * Force the in-core log to disk for a specific LSN.
3310 * Find in-core log with lsn.
3311 * If it is in the DIRTY state, just return.
3312 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3313 * state and go to sleep or return.
3314 * If it is in any other state, go to sleep or return.
3316 * Synchronous forces are implemented with a signal variable. All callers
3317 * to force a given lsn to disk will wait on a the sv attached to the
3318 * specific in-core log. When given in-core log finally completes its
3319 * write to disk, that thread will wake up all threads waiting on the
3324 struct xfs_mount *mp,
3329 struct xlog *log = mp->m_log;
3330 struct xlog_in_core *iclog;
3331 int already_slept = 0;
3335 XFS_STATS_INC(xs_log_force);
3337 lsn = xlog_cil_force_lsn(log, lsn);
3338 if (lsn == NULLCOMMITLSN)
3342 spin_lock(&log->l_icloglock);
3343 iclog = log->l_iclog;
3344 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3345 spin_unlock(&log->l_icloglock);
3350 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3351 iclog = iclog->ic_next;
3355 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3356 spin_unlock(&log->l_icloglock);
3360 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3362 * We sleep here if we haven't already slept (e.g.
3363 * this is the first time we've looked at the correct
3364 * iclog buf) and the buffer before us is going to
3365 * be sync'ed. The reason for this is that if we
3366 * are doing sync transactions here, by waiting for
3367 * the previous I/O to complete, we can allow a few
3368 * more transactions into this iclog before we close
3371 * Otherwise, we mark the buffer WANT_SYNC, and bump
3372 * up the refcnt so we can release the log (which
3373 * drops the ref count). The state switch keeps new
3374 * transaction commits from using this buffer. When
3375 * the current commits finish writing into the buffer,
3376 * the refcount will drop to zero and the buffer will
3379 if (!already_slept &&
3380 (iclog->ic_prev->ic_state &
3381 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3382 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3384 XFS_STATS_INC(xs_log_force_sleep);
3386 xlog_wait(&iclog->ic_prev->ic_write_wait,
3393 atomic_inc(&iclog->ic_refcnt);
3394 xlog_state_switch_iclogs(log, iclog, 0);
3395 spin_unlock(&log->l_icloglock);
3396 if (xlog_state_release_iclog(log, iclog))
3400 spin_lock(&log->l_icloglock);
3403 if ((flags & XFS_LOG_SYNC) && /* sleep */
3405 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3407 * Don't wait on completion if we know that we've
3408 * gotten a log write error.
3410 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3411 spin_unlock(&log->l_icloglock);
3414 XFS_STATS_INC(xs_log_force_sleep);
3415 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3417 * No need to grab the log lock here since we're
3418 * only deciding whether or not to return EIO
3419 * and the memory read should be atomic.
3421 if (iclog->ic_state & XLOG_STATE_IOERROR)
3426 } else { /* just return */
3427 spin_unlock(&log->l_icloglock);
3431 } while (iclog != log->l_iclog);
3433 spin_unlock(&log->l_icloglock);
3438 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3439 * about errors or whether the log was flushed or not. This is the normal
3440 * interface to use when trying to unpin items or move the log forward.
3450 trace_xfs_log_force(mp, lsn);
3451 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3453 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3457 * Called when we want to mark the current iclog as being ready to sync to
3461 xlog_state_want_sync(
3463 struct xlog_in_core *iclog)
3465 assert_spin_locked(&log->l_icloglock);
3467 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3468 xlog_state_switch_iclogs(log, iclog, 0);
3470 ASSERT(iclog->ic_state &
3471 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3476 /*****************************************************************************
3480 *****************************************************************************
3484 * Free a used ticket when its refcount falls to zero.
3488 xlog_ticket_t *ticket)
3490 ASSERT(atomic_read(&ticket->t_ref) > 0);
3491 if (atomic_dec_and_test(&ticket->t_ref))
3492 kmem_zone_free(xfs_log_ticket_zone, ticket);
3497 xlog_ticket_t *ticket)
3499 ASSERT(atomic_read(&ticket->t_ref) > 0);
3500 atomic_inc(&ticket->t_ref);
3505 * Figure out the total log space unit (in bytes) that would be
3506 * required for a log ticket.
3509 xfs_log_calc_unit_res(
3510 struct xfs_mount *mp,
3513 struct xlog *log = mp->m_log;
3518 * Permanent reservations have up to 'cnt'-1 active log operations
3519 * in the log. A unit in this case is the amount of space for one
3520 * of these log operations. Normal reservations have a cnt of 1
3521 * and their unit amount is the total amount of space required.
3523 * The following lines of code account for non-transaction data
3524 * which occupy space in the on-disk log.
3526 * Normal form of a transaction is:
3527 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3528 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3530 * We need to account for all the leadup data and trailer data
3531 * around the transaction data.
3532 * And then we need to account for the worst case in terms of using
3534 * The worst case will happen if:
3535 * - the placement of the transaction happens to be such that the
3536 * roundoff is at its maximum
3537 * - the transaction data is synced before the commit record is synced
3538 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3539 * Therefore the commit record is in its own Log Record.
3540 * This can happen as the commit record is called with its
3541 * own region to xlog_write().
3542 * This then means that in the worst case, roundoff can happen for
3543 * the commit-rec as well.
3544 * The commit-rec is smaller than padding in this scenario and so it is
3545 * not added separately.
3548 /* for trans header */
3549 unit_bytes += sizeof(xlog_op_header_t);
3550 unit_bytes += sizeof(xfs_trans_header_t);
3553 unit_bytes += sizeof(xlog_op_header_t);
3556 * for LR headers - the space for data in an iclog is the size minus
3557 * the space used for the headers. If we use the iclog size, then we
3558 * undercalculate the number of headers required.
3560 * Furthermore - the addition of op headers for split-recs might
3561 * increase the space required enough to require more log and op
3562 * headers, so take that into account too.
3564 * IMPORTANT: This reservation makes the assumption that if this
3565 * transaction is the first in an iclog and hence has the LR headers
3566 * accounted to it, then the remaining space in the iclog is
3567 * exclusively for this transaction. i.e. if the transaction is larger
3568 * than the iclog, it will be the only thing in that iclog.
3569 * Fundamentally, this means we must pass the entire log vector to
3570 * xlog_write to guarantee this.
3572 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3573 num_headers = howmany(unit_bytes, iclog_space);
3575 /* for split-recs - ophdrs added when data split over LRs */
3576 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3578 /* add extra header reservations if we overrun */
3579 while (!num_headers ||
3580 howmany(unit_bytes, iclog_space) > num_headers) {
3581 unit_bytes += sizeof(xlog_op_header_t);
3584 unit_bytes += log->l_iclog_hsize * num_headers;
3586 /* for commit-rec LR header - note: padding will subsume the ophdr */
3587 unit_bytes += log->l_iclog_hsize;
3589 /* for roundoff padding for transaction data and one for commit record */
3590 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3591 /* log su roundoff */
3592 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3595 unit_bytes += 2 * BBSIZE;
3602 * Allocate and initialise a new log ticket.
3604 struct xlog_ticket *
3611 xfs_km_flags_t alloc_flags)
3613 struct xlog_ticket *tic;
3616 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3620 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3622 atomic_set(&tic->t_ref, 1);
3623 tic->t_task = current;
3624 INIT_LIST_HEAD(&tic->t_queue);
3625 tic->t_unit_res = unit_res;
3626 tic->t_curr_res = unit_res;
3629 tic->t_tid = prandom_u32();
3630 tic->t_clientid = client;
3631 tic->t_flags = XLOG_TIC_INITED;
3632 tic->t_trans_type = 0;
3634 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3636 xlog_tic_reset_res(tic);
3642 /******************************************************************************
3644 * Log debug routines
3646 ******************************************************************************
3650 * Make sure that the destination ptr is within the valid data region of
3651 * one of the iclogs. This uses backup pointers stored in a different
3652 * part of the log in case we trash the log structure.
3655 xlog_verify_dest_ptr(
3662 for (i = 0; i < log->l_iclog_bufs; i++) {
3663 if (ptr >= log->l_iclog_bak[i] &&
3664 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3669 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3673 * Check to make sure the grant write head didn't just over lap the tail. If
3674 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3675 * the cycles differ by exactly one and check the byte count.
3677 * This check is run unlocked, so can give false positives. Rather than assert
3678 * on failures, use a warn-once flag and a panic tag to allow the admin to
3679 * determine if they want to panic the machine when such an error occurs. For
3680 * debug kernels this will have the same effect as using an assert but, unlinke
3681 * an assert, it can be turned off at runtime.
3684 xlog_verify_grant_tail(
3687 int tail_cycle, tail_blocks;
3690 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3691 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3692 if (tail_cycle != cycle) {
3693 if (cycle - 1 != tail_cycle &&
3694 !(log->l_flags & XLOG_TAIL_WARN)) {
3695 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3696 "%s: cycle - 1 != tail_cycle", __func__);
3697 log->l_flags |= XLOG_TAIL_WARN;
3700 if (space > BBTOB(tail_blocks) &&
3701 !(log->l_flags & XLOG_TAIL_WARN)) {
3702 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3703 "%s: space > BBTOB(tail_blocks)", __func__);
3704 log->l_flags |= XLOG_TAIL_WARN;
3709 /* check if it will fit */
3711 xlog_verify_tail_lsn(
3713 struct xlog_in_core *iclog,
3718 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3720 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3721 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3722 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3724 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3726 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3727 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3729 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3730 if (blocks < BTOBB(iclog->ic_offset) + 1)
3731 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3733 } /* xlog_verify_tail_lsn */
3736 * Perform a number of checks on the iclog before writing to disk.
3738 * 1. Make sure the iclogs are still circular
3739 * 2. Make sure we have a good magic number
3740 * 3. Make sure we don't have magic numbers in the data
3741 * 4. Check fields of each log operation header for:
3742 * A. Valid client identifier
3743 * B. tid ptr value falls in valid ptr space (user space code)
3744 * C. Length in log record header is correct according to the
3745 * individual operation headers within record.
3746 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3747 * log, check the preceding blocks of the physical log to make sure all
3748 * the cycle numbers agree with the current cycle number.
3753 struct xlog_in_core *iclog,
3757 xlog_op_header_t *ophead;
3758 xlog_in_core_t *icptr;
3759 xlog_in_core_2_t *xhdr;
3761 xfs_caddr_t base_ptr;
3762 __psint_t field_offset;
3764 int len, i, j, k, op_len;
3767 /* check validity of iclog pointers */
3768 spin_lock(&log->l_icloglock);
3769 icptr = log->l_iclog;
3770 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3773 if (icptr != log->l_iclog)
3774 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3775 spin_unlock(&log->l_icloglock);
3777 /* check log magic numbers */
3778 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3779 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3781 ptr = (xfs_caddr_t) &iclog->ic_header;
3782 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3784 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3785 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3790 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3791 ptr = iclog->ic_datap;
3793 ophead = (xlog_op_header_t *)ptr;
3794 xhdr = iclog->ic_data;
3795 for (i = 0; i < len; i++) {
3796 ophead = (xlog_op_header_t *)ptr;
3798 /* clientid is only 1 byte */
3799 field_offset = (__psint_t)
3800 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3801 if (!syncing || (field_offset & 0x1ff)) {
3802 clientid = ophead->oh_clientid;
3804 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3805 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3806 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3807 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3808 clientid = xlog_get_client_id(
3809 xhdr[j].hic_xheader.xh_cycle_data[k]);
3811 clientid = xlog_get_client_id(
3812 iclog->ic_header.h_cycle_data[idx]);
3815 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3817 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3818 __func__, clientid, ophead,
3819 (unsigned long)field_offset);
3822 field_offset = (__psint_t)
3823 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3824 if (!syncing || (field_offset & 0x1ff)) {
3825 op_len = be32_to_cpu(ophead->oh_len);
3827 idx = BTOBBT((__psint_t)&ophead->oh_len -
3828 (__psint_t)iclog->ic_datap);
3829 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3830 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3831 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3832 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3834 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3837 ptr += sizeof(xlog_op_header_t) + op_len;
3839 } /* xlog_verify_iclog */
3843 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3849 xlog_in_core_t *iclog, *ic;
3851 iclog = log->l_iclog;
3852 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3854 * Mark all the incore logs IOERROR.
3855 * From now on, no log flushes will result.
3859 ic->ic_state = XLOG_STATE_IOERROR;
3861 } while (ic != iclog);
3865 * Return non-zero, if state transition has already happened.
3871 * This is called from xfs_force_shutdown, when we're forcibly
3872 * shutting down the filesystem, typically because of an IO error.
3873 * Our main objectives here are to make sure that:
3874 * a. if !logerror, flush the logs to disk. Anything modified
3875 * after this is ignored.
3876 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3877 * parties to find out, 'atomically'.
3878 * c. those who're sleeping on log reservations, pinned objects and
3879 * other resources get woken up, and be told the bad news.
3880 * d. nothing new gets queued up after (b) and (c) are done.
3882 * Note: for the !logerror case we need to flush the regions held in memory out
3883 * to disk first. This needs to be done before the log is marked as shutdown,
3884 * otherwise the iclog writes will fail.
3887 xfs_log_force_umount(
3888 struct xfs_mount *mp,
3897 * If this happens during log recovery, don't worry about
3898 * locking; the log isn't open for business yet.
3901 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3902 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3904 XFS_BUF_DONE(mp->m_sb_bp);
3909 * Somebody could've already done the hard work for us.
3910 * No need to get locks for this.
3912 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3913 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3918 * Flush all the completed transactions to disk before marking the log
3919 * being shut down. We need to do it in this order to ensure that
3920 * completed operations are safely on disk before we shut down, and that
3921 * we don't have to issue any buffer IO after the shutdown flags are set
3922 * to guarantee this.
3925 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3928 * mark the filesystem and the as in a shutdown state and wake
3929 * everybody up to tell them the bad news.
3931 spin_lock(&log->l_icloglock);
3932 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3934 XFS_BUF_DONE(mp->m_sb_bp);
3937 * Mark the log and the iclogs with IO error flags to prevent any
3938 * further log IO from being issued or completed.
3940 log->l_flags |= XLOG_IO_ERROR;
3941 retval = xlog_state_ioerror(log);
3942 spin_unlock(&log->l_icloglock);
3945 * We don't want anybody waiting for log reservations after this. That
3946 * means we have to wake up everybody queued up on reserveq as well as
3947 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3948 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3949 * action is protected by the grant locks.
3951 xlog_grant_head_wake_all(&log->l_reserve_head);
3952 xlog_grant_head_wake_all(&log->l_write_head);
3955 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3956 * as if the log writes were completed. The abort handling in the log
3957 * item committed callback functions will do this again under lock to
3960 wake_up_all(&log->l_cilp->xc_commit_wait);
3961 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3963 #ifdef XFSERRORDEBUG
3965 xlog_in_core_t *iclog;
3967 spin_lock(&log->l_icloglock);
3968 iclog = log->l_iclog;
3970 ASSERT(iclog->ic_callback == 0);
3971 iclog = iclog->ic_next;
3972 } while (iclog != log->l_iclog);
3973 spin_unlock(&log->l_icloglock);
3976 /* return non-zero if log IOERROR transition had already happened */
3984 xlog_in_core_t *iclog;
3986 iclog = log->l_iclog;
3988 /* endianness does not matter here, zero is zero in
3991 if (iclog->ic_header.h_num_logops)
3993 iclog = iclog->ic_next;
3994 } while (iclog != log->l_iclog);