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"
25 #include "xfs_mount.h"
26 #include "xfs_error.h"
27 #include "xfs_trans.h"
28 #include "xfs_trans_priv.h"
30 #include "xfs_log_priv.h"
31 #include "xfs_log_recover.h"
32 #include "xfs_inode.h"
33 #include "xfs_trace.h"
34 #include "xfs_fsops.h"
35 #include "xfs_cksum.h"
36 #include "xfs_sysfs.h"
38 kmem_zone_t *xfs_log_ticket_zone;
40 /* Local miscellaneous function prototypes */
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
61 struct xlog_in_core *iclog);
66 /* local state machine functions */
67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
69 xlog_state_do_callback(
72 struct xlog_in_core *iclog);
74 xlog_state_get_iclog_space(
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
82 xlog_state_release_iclog(
84 struct xlog_in_core *iclog);
86 xlog_state_switch_iclogs(
88 struct xlog_in_core *iclog,
93 struct xlog_in_core *iclog);
100 xlog_regrant_reserve_log_space(
102 struct xlog_ticket *ticket);
104 xlog_ungrant_log_space(
106 struct xlog_ticket *ticket);
110 xlog_verify_dest_ptr(
114 xlog_verify_grant_tail(
119 struct xlog_in_core *iclog,
123 xlog_verify_tail_lsn(
125 struct xlog_in_core *iclog,
128 #define xlog_verify_dest_ptr(a,b)
129 #define xlog_verify_grant_tail(a)
130 #define xlog_verify_iclog(a,b,c,d)
131 #define xlog_verify_tail_lsn(a,b,c)
139 xlog_grant_sub_space(
144 int64_t head_val = atomic64_read(head);
150 xlog_crack_grant_head_val(head_val, &cycle, &space);
154 space += log->l_logsize;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
165 xlog_grant_add_space(
170 int64_t head_val = atomic64_read(head);
177 xlog_crack_grant_head_val(head_val, &cycle, &space);
179 tmp = log->l_logsize - space;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
194 xlog_grant_head_init(
195 struct xlog_grant_head *head)
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
203 xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
206 struct xlog_ticket *tic;
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
215 xlog_ticket_reservation(
217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
225 return tic->t_unit_res * tic->t_cnt;
227 return tic->t_unit_res;
232 xlog_grant_head_wake(
234 struct xlog_grant_head *head,
237 struct xlog_ticket *tic;
240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
242 if (*free_bytes < need_bytes)
245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
247 wake_up_process(tic->t_task);
254 xlog_grant_head_wait(
256 struct xlog_grant_head *head,
257 struct xlog_ticket *tic,
258 int need_bytes) __releases(&head->lock)
259 __acquires(&head->lock)
261 list_add_tail(&tic->t_queue, &head->waiters);
264 if (XLOG_FORCED_SHUTDOWN(log))
266 xlog_grant_push_ail(log, need_bytes);
268 __set_current_state(TASK_UNINTERRUPTIBLE);
269 spin_unlock(&head->lock);
271 XFS_STATS_INC(xs_sleep_logspace);
273 trace_xfs_log_grant_sleep(log, tic);
275 trace_xfs_log_grant_wake(log, tic);
277 spin_lock(&head->lock);
278 if (XLOG_FORCED_SHUTDOWN(log))
280 } while (xlog_space_left(log, &head->grant) < need_bytes);
282 list_del_init(&tic->t_queue);
285 list_del_init(&tic->t_queue);
290 * Atomically get the log space required for a log ticket.
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
307 xlog_grant_head_check(
309 struct xlog_grant_head *head,
310 struct xlog_ticket *tic,
316 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
324 *need_bytes = xlog_ticket_reservation(log, head, tic);
325 free_bytes = xlog_space_left(log, &head->grant);
326 if (!list_empty_careful(&head->waiters)) {
327 spin_lock(&head->lock);
328 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
329 free_bytes < *need_bytes) {
330 error = xlog_grant_head_wait(log, head, tic,
333 spin_unlock(&head->lock);
334 } else if (free_bytes < *need_bytes) {
335 spin_lock(&head->lock);
336 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
337 spin_unlock(&head->lock);
344 xlog_tic_reset_res(xlog_ticket_t *tic)
347 tic->t_res_arr_sum = 0;
348 tic->t_res_num_ophdrs = 0;
352 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
354 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
355 /* add to overflow and start again */
356 tic->t_res_o_flow += tic->t_res_arr_sum;
358 tic->t_res_arr_sum = 0;
361 tic->t_res_arr[tic->t_res_num].r_len = len;
362 tic->t_res_arr[tic->t_res_num].r_type = type;
363 tic->t_res_arr_sum += len;
368 * Replenish the byte reservation required by moving the grant write head.
372 struct xfs_mount *mp,
373 struct xlog_ticket *tic)
375 struct xlog *log = mp->m_log;
379 if (XLOG_FORCED_SHUTDOWN(log))
382 XFS_STATS_INC(xs_try_logspace);
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
392 xlog_grant_push_ail(log, tic->t_unit_res);
394 tic->t_curr_res = tic->t_unit_res;
395 xlog_tic_reset_res(tic);
400 trace_xfs_log_regrant(log, tic);
402 error = xlog_grant_head_check(log, &log->l_write_head, tic,
407 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
408 trace_xfs_log_regrant_exit(log, tic);
409 xlog_verify_grant_tail(log);
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
419 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
424 * Reserve log space and return a ticket corresponding the reservation.
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
433 struct xfs_mount *mp,
436 struct xlog_ticket **ticp,
441 struct xlog *log = mp->m_log;
442 struct xlog_ticket *tic;
446 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
448 if (XLOG_FORCED_SHUTDOWN(log))
451 XFS_STATS_INC(xs_try_logspace);
453 ASSERT(*ticp == NULL);
454 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
455 KM_SLEEP | KM_MAYFAIL);
459 tic->t_trans_type = t_type;
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
465 trace_xfs_log_reserve(log, tic);
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
493 * 1. currblock field gets updated at startup and after in-core logs
494 * marked as with WANT_SYNC.
498 * This routine is called when a user of a log manager ticket is done with
499 * the reservation. If the ticket was ever used, then a commit record for
500 * the associated transaction is written out as a log operation header with
501 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
502 * a given ticket. If the ticket was one with a permanent reservation, then
503 * a few operations are done differently. Permanent reservation tickets by
504 * default don't release the reservation. They just commit the current
505 * transaction with the belief that the reservation is still needed. A flag
506 * must be passed in before permanent reservations are actually released.
507 * When these type of tickets are not released, they need to be set into
508 * the inited state again. By doing this, a start record will be written
509 * out when the next write occurs.
513 struct xfs_mount *mp,
514 struct xlog_ticket *ticket,
515 struct xlog_in_core **iclog,
518 struct xlog *log = mp->m_log;
521 if (XLOG_FORCED_SHUTDOWN(log) ||
523 * If nothing was ever written, don't write out commit record.
524 * If we get an error, just continue and give back the log ticket.
526 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
527 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
528 lsn = (xfs_lsn_t) -1;
529 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
530 flags |= XFS_LOG_REL_PERM_RESERV;
535 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
536 (flags & XFS_LOG_REL_PERM_RESERV)) {
537 trace_xfs_log_done_nonperm(log, ticket);
540 * Release ticket if not permanent reservation or a specific
541 * request has been made to release a permanent reservation.
543 xlog_ungrant_log_space(log, ticket);
544 xfs_log_ticket_put(ticket);
546 trace_xfs_log_done_perm(log, ticket);
548 xlog_regrant_reserve_log_space(log, ticket);
549 /* If this ticket was a permanent reservation and we aren't
550 * trying to release it, reset the inited flags; so next time
551 * we write, a start record will be written out.
553 ticket->t_flags |= XLOG_TIC_INITED;
560 * Attaches a new iclog I/O completion callback routine during
561 * transaction commit. If the log is in error state, a non-zero
562 * return code is handed back and the caller is responsible for
563 * executing the callback at an appropriate time.
567 struct xfs_mount *mp,
568 struct xlog_in_core *iclog,
569 xfs_log_callback_t *cb)
573 spin_lock(&iclog->ic_callback_lock);
574 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
576 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
577 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
579 *(iclog->ic_callback_tail) = cb;
580 iclog->ic_callback_tail = &(cb->cb_next);
582 spin_unlock(&iclog->ic_callback_lock);
587 xfs_log_release_iclog(
588 struct xfs_mount *mp,
589 struct xlog_in_core *iclog)
591 if (xlog_state_release_iclog(mp->m_log, iclog)) {
592 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
600 * Mount a log filesystem
602 * mp - ubiquitous xfs mount point structure
603 * log_target - buftarg of on-disk log device
604 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
605 * num_bblocks - Number of BBSIZE blocks in on-disk log
607 * Return error or zero.
612 xfs_buftarg_t *log_target,
613 xfs_daddr_t blk_offset,
619 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
620 xfs_notice(mp, "Mounting V%d Filesystem",
621 XFS_SB_VERSION_NUM(&mp->m_sb));
624 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
625 XFS_SB_VERSION_NUM(&mp->m_sb));
626 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
629 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
630 if (IS_ERR(mp->m_log)) {
631 error = PTR_ERR(mp->m_log);
636 * Validate the given log space and drop a critical message via syslog
637 * if the log size is too small that would lead to some unexpected
638 * situations in transaction log space reservation stage.
640 * Note: we can't just reject the mount if the validation fails. This
641 * would mean that people would have to downgrade their kernel just to
642 * remedy the situation as there is no way to grow the log (short of
643 * black magic surgery with xfs_db).
645 * We can, however, reject mounts for CRC format filesystems, as the
646 * mkfs binary being used to make the filesystem should never create a
647 * filesystem with a log that is too small.
649 min_logfsbs = xfs_log_calc_minimum_size(mp);
651 if (mp->m_sb.sb_logblocks < min_logfsbs) {
653 "Log size %d blocks too small, minimum size is %d blocks",
654 mp->m_sb.sb_logblocks, min_logfsbs);
656 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
658 "Log size %d blocks too large, maximum size is %lld blocks",
659 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
661 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
663 "log size %lld bytes too large, maximum size is %lld bytes",
664 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
669 if (xfs_sb_version_hascrc(&mp->m_sb)) {
670 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
675 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
676 "experienced then please report this message in the bug report.");
680 * Initialize the AIL now we have a log.
682 error = xfs_trans_ail_init(mp);
684 xfs_warn(mp, "AIL initialisation failed: error %d", error);
687 mp->m_log->l_ailp = mp->m_ail;
690 * skip log recovery on a norecovery mount. pretend it all
693 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
694 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
697 mp->m_flags &= ~XFS_MOUNT_RDONLY;
699 error = xlog_recover(mp->m_log);
702 mp->m_flags |= XFS_MOUNT_RDONLY;
704 xfs_warn(mp, "log mount/recovery failed: error %d",
706 goto out_destroy_ail;
710 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
713 goto out_destroy_ail;
715 /* Normal transactions can now occur */
716 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
719 * Now the log has been fully initialised and we know were our
720 * space grant counters are, we can initialise the permanent ticket
721 * needed for delayed logging to work.
723 xlog_cil_init_post_recovery(mp->m_log);
728 xfs_trans_ail_destroy(mp);
730 xlog_dealloc_log(mp->m_log);
736 * Finish the recovery of the file system. This is separate from the
737 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
738 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
741 * If we finish recovery successfully, start the background log work. If we are
742 * not doing recovery, then we have a RO filesystem and we don't need to start
746 xfs_log_mount_finish(xfs_mount_t *mp)
750 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
751 error = xlog_recover_finish(mp->m_log);
753 xfs_log_work_queue(mp);
755 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
763 * Final log writes as part of unmount.
765 * Mark the filesystem clean as unmount happens. Note that during relocation
766 * this routine needs to be executed as part of source-bag while the
767 * deallocation must not be done until source-end.
771 * Unmount record used to have a string "Unmount filesystem--" in the
772 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
773 * We just write the magic number now since that particular field isn't
774 * currently architecture converted and "Unmount" is a bit foo.
775 * As far as I know, there weren't any dependencies on the old behaviour.
779 xfs_log_unmount_write(xfs_mount_t *mp)
781 struct xlog *log = mp->m_log;
782 xlog_in_core_t *iclog;
784 xlog_in_core_t *first_iclog;
786 xlog_ticket_t *tic = NULL;
791 * Don't write out unmount record on read-only mounts.
792 * Or, if we are doing a forced umount (typically because of IO errors).
794 if (mp->m_flags & XFS_MOUNT_RDONLY)
797 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
798 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
801 first_iclog = iclog = log->l_iclog;
803 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
804 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
805 ASSERT(iclog->ic_offset == 0);
807 iclog = iclog->ic_next;
808 } while (iclog != first_iclog);
810 if (! (XLOG_FORCED_SHUTDOWN(log))) {
811 error = xfs_log_reserve(mp, 600, 1, &tic,
812 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
814 /* the data section must be 32 bit size aligned */
818 __uint32_t pad2; /* may as well make it 64 bits */
820 .magic = XLOG_UNMOUNT_TYPE,
822 struct xfs_log_iovec reg = {
824 .i_len = sizeof(magic),
825 .i_type = XLOG_REG_TYPE_UNMOUNT,
827 struct xfs_log_vec vec = {
832 /* remove inited flag, and account for space used */
834 tic->t_curr_res -= sizeof(magic);
835 error = xlog_write(log, &vec, tic, &lsn,
836 NULL, XLOG_UNMOUNT_TRANS);
838 * At this point, we're umounting anyway,
839 * so there's no point in transitioning log state
840 * to IOERROR. Just continue...
845 xfs_alert(mp, "%s: unmount record failed", __func__);
848 spin_lock(&log->l_icloglock);
849 iclog = log->l_iclog;
850 atomic_inc(&iclog->ic_refcnt);
851 xlog_state_want_sync(log, iclog);
852 spin_unlock(&log->l_icloglock);
853 error = xlog_state_release_iclog(log, iclog);
855 spin_lock(&log->l_icloglock);
856 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
857 iclog->ic_state == XLOG_STATE_DIRTY)) {
858 if (!XLOG_FORCED_SHUTDOWN(log)) {
859 xlog_wait(&iclog->ic_force_wait,
862 spin_unlock(&log->l_icloglock);
865 spin_unlock(&log->l_icloglock);
868 trace_xfs_log_umount_write(log, tic);
869 xlog_ungrant_log_space(log, tic);
870 xfs_log_ticket_put(tic);
874 * We're already in forced_shutdown mode, couldn't
875 * even attempt to write out the unmount transaction.
877 * Go through the motions of sync'ing and releasing
878 * the iclog, even though no I/O will actually happen,
879 * we need to wait for other log I/Os that may already
880 * be in progress. Do this as a separate section of
881 * code so we'll know if we ever get stuck here that
882 * we're in this odd situation of trying to unmount
883 * a file system that went into forced_shutdown as
884 * the result of an unmount..
886 spin_lock(&log->l_icloglock);
887 iclog = log->l_iclog;
888 atomic_inc(&iclog->ic_refcnt);
890 xlog_state_want_sync(log, iclog);
891 spin_unlock(&log->l_icloglock);
892 error = xlog_state_release_iclog(log, iclog);
894 spin_lock(&log->l_icloglock);
896 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
897 || iclog->ic_state == XLOG_STATE_DIRTY
898 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
900 xlog_wait(&iclog->ic_force_wait,
903 spin_unlock(&log->l_icloglock);
908 } /* xfs_log_unmount_write */
911 * Empty the log for unmount/freeze.
913 * To do this, we first need to shut down the background log work so it is not
914 * trying to cover the log as we clean up. We then need to unpin all objects in
915 * the log so we can then flush them out. Once they have completed their IO and
916 * run the callbacks removing themselves from the AIL, we can write the unmount
921 struct xfs_mount *mp)
923 cancel_delayed_work_sync(&mp->m_log->l_work);
924 xfs_log_force(mp, XFS_LOG_SYNC);
927 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
928 * will push it, xfs_wait_buftarg() will not wait for it. Further,
929 * xfs_buf_iowait() cannot be used because it was pushed with the
930 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
931 * the IO to complete.
933 xfs_ail_push_all_sync(mp->m_ail);
934 xfs_wait_buftarg(mp->m_ddev_targp);
935 xfs_buf_lock(mp->m_sb_bp);
936 xfs_buf_unlock(mp->m_sb_bp);
938 xfs_log_unmount_write(mp);
942 * Shut down and release the AIL and Log.
944 * During unmount, we need to ensure we flush all the dirty metadata objects
945 * from the AIL so that the log is empty before we write the unmount record to
946 * the log. Once this is done, we can tear down the AIL and the log.
950 struct xfs_mount *mp)
954 xfs_trans_ail_destroy(mp);
956 xfs_sysfs_del(&mp->m_log->l_kobj);
958 xlog_dealloc_log(mp->m_log);
963 struct xfs_mount *mp,
964 struct xfs_log_item *item,
966 const struct xfs_item_ops *ops)
968 item->li_mountp = mp;
969 item->li_ailp = mp->m_ail;
970 item->li_type = type;
974 INIT_LIST_HEAD(&item->li_ail);
975 INIT_LIST_HEAD(&item->li_cil);
979 * Wake up processes waiting for log space after we have moved the log tail.
983 struct xfs_mount *mp)
985 struct xlog *log = mp->m_log;
988 if (XLOG_FORCED_SHUTDOWN(log))
991 if (!list_empty_careful(&log->l_write_head.waiters)) {
992 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
994 spin_lock(&log->l_write_head.lock);
995 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
996 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
997 spin_unlock(&log->l_write_head.lock);
1000 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1001 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1003 spin_lock(&log->l_reserve_head.lock);
1004 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1005 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1006 spin_unlock(&log->l_reserve_head.lock);
1011 * Determine if we have a transaction that has gone to disk that needs to be
1012 * covered. To begin the transition to the idle state firstly the log needs to
1013 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1014 * we start attempting to cover the log.
1016 * Only if we are then in a state where covering is needed, the caller is
1017 * informed that dummy transactions are required to move the log into the idle
1020 * If there are any items in the AIl or CIL, then we do not want to attempt to
1021 * cover the log as we may be in a situation where there isn't log space
1022 * available to run a dummy transaction and this can lead to deadlocks when the
1023 * tail of the log is pinned by an item that is modified in the CIL. Hence
1024 * there's no point in running a dummy transaction at this point because we
1025 * can't start trying to idle the log until both the CIL and AIL are empty.
1028 xfs_log_need_covered(xfs_mount_t *mp)
1030 struct xlog *log = mp->m_log;
1033 if (!xfs_fs_writable(mp))
1036 if (!xlog_cil_empty(log))
1039 spin_lock(&log->l_icloglock);
1040 switch (log->l_covered_state) {
1041 case XLOG_STATE_COVER_DONE:
1042 case XLOG_STATE_COVER_DONE2:
1043 case XLOG_STATE_COVER_IDLE:
1045 case XLOG_STATE_COVER_NEED:
1046 case XLOG_STATE_COVER_NEED2:
1047 if (xfs_ail_min_lsn(log->l_ailp))
1049 if (!xlog_iclogs_empty(log))
1053 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1054 log->l_covered_state = XLOG_STATE_COVER_DONE;
1056 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1062 spin_unlock(&log->l_icloglock);
1067 * We may be holding the log iclog lock upon entering this routine.
1070 xlog_assign_tail_lsn_locked(
1071 struct xfs_mount *mp)
1073 struct xlog *log = mp->m_log;
1074 struct xfs_log_item *lip;
1077 assert_spin_locked(&mp->m_ail->xa_lock);
1080 * To make sure we always have a valid LSN for the log tail we keep
1081 * track of the last LSN which was committed in log->l_last_sync_lsn,
1082 * and use that when the AIL was empty.
1084 lip = xfs_ail_min(mp->m_ail);
1086 tail_lsn = lip->li_lsn;
1088 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1089 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1090 atomic64_set(&log->l_tail_lsn, tail_lsn);
1095 xlog_assign_tail_lsn(
1096 struct xfs_mount *mp)
1100 spin_lock(&mp->m_ail->xa_lock);
1101 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1102 spin_unlock(&mp->m_ail->xa_lock);
1108 * Return the space in the log between the tail and the head. The head
1109 * is passed in the cycle/bytes formal parms. In the special case where
1110 * the reserve head has wrapped passed the tail, this calculation is no
1111 * longer valid. In this case, just return 0 which means there is no space
1112 * in the log. This works for all places where this function is called
1113 * with the reserve head. Of course, if the write head were to ever
1114 * wrap the tail, we should blow up. Rather than catch this case here,
1115 * we depend on other ASSERTions in other parts of the code. XXXmiken
1117 * This code also handles the case where the reservation head is behind
1118 * the tail. The details of this case are described below, but the end
1119 * result is that we return the size of the log as the amount of space left.
1132 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1133 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1134 tail_bytes = BBTOB(tail_bytes);
1135 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1136 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1137 else if (tail_cycle + 1 < head_cycle)
1139 else if (tail_cycle < head_cycle) {
1140 ASSERT(tail_cycle == (head_cycle - 1));
1141 free_bytes = tail_bytes - head_bytes;
1144 * The reservation head is behind the tail.
1145 * In this case we just want to return the size of the
1146 * log as the amount of space left.
1148 xfs_alert(log->l_mp,
1149 "xlog_space_left: head behind tail\n"
1150 " tail_cycle = %d, tail_bytes = %d\n"
1151 " GH cycle = %d, GH bytes = %d",
1152 tail_cycle, tail_bytes, head_cycle, head_bytes);
1154 free_bytes = log->l_logsize;
1161 * Log function which is called when an io completes.
1163 * The log manager needs its own routine, in order to control what
1164 * happens with the buffer after the write completes.
1167 xlog_iodone(xfs_buf_t *bp)
1169 struct xlog_in_core *iclog = bp->b_fspriv;
1170 struct xlog *l = iclog->ic_log;
1174 * Race to shutdown the filesystem if we see an error.
1176 if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1177 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1178 xfs_buf_ioerror_alert(bp, __func__);
1180 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1182 * This flag will be propagated to the trans-committed
1183 * callback routines to let them know that the log-commit
1186 aborted = XFS_LI_ABORTED;
1187 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1188 aborted = XFS_LI_ABORTED;
1191 /* log I/O is always issued ASYNC */
1192 ASSERT(XFS_BUF_ISASYNC(bp));
1193 xlog_state_done_syncing(iclog, aborted);
1196 * drop the buffer lock now that we are done. Nothing references
1197 * the buffer after this, so an unmount waiting on this lock can now
1198 * tear it down safely. As such, it is unsafe to reference the buffer
1199 * (bp) after the unlock as we could race with it being freed.
1205 * Return size of each in-core log record buffer.
1207 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1209 * If the filesystem blocksize is too large, we may need to choose a
1210 * larger size since the directory code currently logs entire blocks.
1214 xlog_get_iclog_buffer_size(
1215 struct xfs_mount *mp,
1221 if (mp->m_logbufs <= 0)
1222 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1224 log->l_iclog_bufs = mp->m_logbufs;
1227 * Buffer size passed in from mount system call.
1229 if (mp->m_logbsize > 0) {
1230 size = log->l_iclog_size = mp->m_logbsize;
1231 log->l_iclog_size_log = 0;
1233 log->l_iclog_size_log++;
1237 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1238 /* # headers = size / 32k
1239 * one header holds cycles from 32k of data
1242 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1243 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1245 log->l_iclog_hsize = xhdrs << BBSHIFT;
1246 log->l_iclog_heads = xhdrs;
1248 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1249 log->l_iclog_hsize = BBSIZE;
1250 log->l_iclog_heads = 1;
1255 /* All machines use 32kB buffers by default. */
1256 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1257 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1259 /* the default log size is 16k or 32k which is one header sector */
1260 log->l_iclog_hsize = BBSIZE;
1261 log->l_iclog_heads = 1;
1264 /* are we being asked to make the sizes selected above visible? */
1265 if (mp->m_logbufs == 0)
1266 mp->m_logbufs = log->l_iclog_bufs;
1267 if (mp->m_logbsize == 0)
1268 mp->m_logbsize = log->l_iclog_size;
1269 } /* xlog_get_iclog_buffer_size */
1274 struct xfs_mount *mp)
1276 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1277 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1281 * Every sync period we need to unpin all items in the AIL and push them to
1282 * disk. If there is nothing dirty, then we might need to cover the log to
1283 * indicate that the filesystem is idle.
1287 struct work_struct *work)
1289 struct xlog *log = container_of(to_delayed_work(work),
1290 struct xlog, l_work);
1291 struct xfs_mount *mp = log->l_mp;
1293 /* dgc: errors ignored - not fatal and nowhere to report them */
1294 if (xfs_log_need_covered(mp))
1295 xfs_fs_log_dummy(mp);
1297 xfs_log_force(mp, 0);
1299 /* start pushing all the metadata that is currently dirty */
1300 xfs_ail_push_all(mp->m_ail);
1302 /* queue us up again */
1303 xfs_log_work_queue(mp);
1307 * This routine initializes some of the log structure for a given mount point.
1308 * Its primary purpose is to fill in enough, so recovery can occur. However,
1309 * some other stuff may be filled in too.
1311 STATIC struct xlog *
1313 struct xfs_mount *mp,
1314 struct xfs_buftarg *log_target,
1315 xfs_daddr_t blk_offset,
1319 xlog_rec_header_t *head;
1320 xlog_in_core_t **iclogp;
1321 xlog_in_core_t *iclog, *prev_iclog=NULL;
1324 int error = -ENOMEM;
1327 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1329 xfs_warn(mp, "Log allocation failed: No memory!");
1334 log->l_targ = log_target;
1335 log->l_logsize = BBTOB(num_bblks);
1336 log->l_logBBstart = blk_offset;
1337 log->l_logBBsize = num_bblks;
1338 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1339 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1340 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1342 log->l_prev_block = -1;
1343 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1344 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1345 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1346 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1348 xlog_grant_head_init(&log->l_reserve_head);
1349 xlog_grant_head_init(&log->l_write_head);
1351 error = -EFSCORRUPTED;
1352 if (xfs_sb_version_hassector(&mp->m_sb)) {
1353 log2_size = mp->m_sb.sb_logsectlog;
1354 if (log2_size < BBSHIFT) {
1355 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1356 log2_size, BBSHIFT);
1360 log2_size -= BBSHIFT;
1361 if (log2_size > mp->m_sectbb_log) {
1362 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1363 log2_size, mp->m_sectbb_log);
1367 /* for larger sector sizes, must have v2 or external log */
1368 if (log2_size && log->l_logBBstart > 0 &&
1369 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1371 "log sector size (0x%x) invalid for configuration.",
1376 log->l_sectBBsize = 1 << log2_size;
1378 xlog_get_iclog_buffer_size(mp, log);
1381 * Use a NULL block for the extra log buffer used during splits so that
1382 * it will trigger errors if we ever try to do IO on it without first
1383 * having set it up properly.
1386 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1387 BTOBB(log->l_iclog_size), 0);
1392 * The iclogbuf buffer locks are held over IO but we are not going to do
1393 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1394 * when appropriately.
1396 ASSERT(xfs_buf_islocked(bp));
1399 bp->b_iodone = xlog_iodone;
1402 spin_lock_init(&log->l_icloglock);
1403 init_waitqueue_head(&log->l_flush_wait);
1405 iclogp = &log->l_iclog;
1407 * The amount of memory to allocate for the iclog structure is
1408 * rather funky due to the way the structure is defined. It is
1409 * done this way so that we can use different sizes for machines
1410 * with different amounts of memory. See the definition of
1411 * xlog_in_core_t in xfs_log_priv.h for details.
1413 ASSERT(log->l_iclog_size >= 4096);
1414 for (i=0; i < log->l_iclog_bufs; i++) {
1415 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1417 goto out_free_iclog;
1420 iclog->ic_prev = prev_iclog;
1423 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1424 BTOBB(log->l_iclog_size), 0);
1426 goto out_free_iclog;
1428 ASSERT(xfs_buf_islocked(bp));
1431 bp->b_iodone = xlog_iodone;
1433 iclog->ic_data = bp->b_addr;
1435 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1437 head = &iclog->ic_header;
1438 memset(head, 0, sizeof(xlog_rec_header_t));
1439 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1440 head->h_version = cpu_to_be32(
1441 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1442 head->h_size = cpu_to_be32(log->l_iclog_size);
1444 head->h_fmt = cpu_to_be32(XLOG_FMT);
1445 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1447 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1448 iclog->ic_state = XLOG_STATE_ACTIVE;
1449 iclog->ic_log = log;
1450 atomic_set(&iclog->ic_refcnt, 0);
1451 spin_lock_init(&iclog->ic_callback_lock);
1452 iclog->ic_callback_tail = &(iclog->ic_callback);
1453 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1455 init_waitqueue_head(&iclog->ic_force_wait);
1456 init_waitqueue_head(&iclog->ic_write_wait);
1458 iclogp = &iclog->ic_next;
1460 *iclogp = log->l_iclog; /* complete ring */
1461 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1463 error = xlog_cil_init(log);
1465 goto out_free_iclog;
1469 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1470 prev_iclog = iclog->ic_next;
1472 xfs_buf_free(iclog->ic_bp);
1475 spinlock_destroy(&log->l_icloglock);
1476 xfs_buf_free(log->l_xbuf);
1480 return ERR_PTR(error);
1481 } /* xlog_alloc_log */
1485 * Write out the commit record of a transaction associated with the given
1486 * ticket. Return the lsn of the commit record.
1491 struct xlog_ticket *ticket,
1492 struct xlog_in_core **iclog,
1493 xfs_lsn_t *commitlsnp)
1495 struct xfs_mount *mp = log->l_mp;
1497 struct xfs_log_iovec reg = {
1500 .i_type = XLOG_REG_TYPE_COMMIT,
1502 struct xfs_log_vec vec = {
1507 ASSERT_ALWAYS(iclog);
1508 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1511 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1516 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1517 * log space. This code pushes on the lsn which would supposedly free up
1518 * the 25% which we want to leave free. We may need to adopt a policy which
1519 * pushes on an lsn which is further along in the log once we reach the high
1520 * water mark. In this manner, we would be creating a low water mark.
1523 xlog_grant_push_ail(
1527 xfs_lsn_t threshold_lsn = 0;
1528 xfs_lsn_t last_sync_lsn;
1531 int threshold_block;
1532 int threshold_cycle;
1535 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1537 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1538 free_blocks = BTOBBT(free_bytes);
1541 * Set the threshold for the minimum number of free blocks in the
1542 * log to the maximum of what the caller needs, one quarter of the
1543 * log, and 256 blocks.
1545 free_threshold = BTOBB(need_bytes);
1546 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1547 free_threshold = MAX(free_threshold, 256);
1548 if (free_blocks >= free_threshold)
1551 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1553 threshold_block += free_threshold;
1554 if (threshold_block >= log->l_logBBsize) {
1555 threshold_block -= log->l_logBBsize;
1556 threshold_cycle += 1;
1558 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1561 * Don't pass in an lsn greater than the lsn of the last
1562 * log record known to be on disk. Use a snapshot of the last sync lsn
1563 * so that it doesn't change between the compare and the set.
1565 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1566 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1567 threshold_lsn = last_sync_lsn;
1570 * Get the transaction layer to kick the dirty buffers out to
1571 * disk asynchronously. No point in trying to do this if
1572 * the filesystem is shutting down.
1574 if (!XLOG_FORCED_SHUTDOWN(log))
1575 xfs_ail_push(log->l_ailp, threshold_lsn);
1579 * Stamp cycle number in every block
1584 struct xlog_in_core *iclog,
1588 int size = iclog->ic_offset + roundoff;
1592 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1594 dp = iclog->ic_datap;
1595 for (i = 0; i < BTOBB(size); i++) {
1596 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1598 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1599 *(__be32 *)dp = cycle_lsn;
1603 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1604 xlog_in_core_2_t *xhdr = iclog->ic_data;
1606 for ( ; i < BTOBB(size); i++) {
1607 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1608 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1609 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1610 *(__be32 *)dp = cycle_lsn;
1614 for (i = 1; i < log->l_iclog_heads; i++)
1615 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1620 * Calculate the checksum for a log buffer.
1622 * This is a little more complicated than it should be because the various
1623 * headers and the actual data are non-contiguous.
1628 struct xlog_rec_header *rhead,
1634 /* first generate the crc for the record header ... */
1635 crc = xfs_start_cksum((char *)rhead,
1636 sizeof(struct xlog_rec_header),
1637 offsetof(struct xlog_rec_header, h_crc));
1639 /* ... then for additional cycle data for v2 logs ... */
1640 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1641 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1644 for (i = 1; i < log->l_iclog_heads; i++) {
1645 crc = crc32c(crc, &xhdr[i].hic_xheader,
1646 sizeof(struct xlog_rec_ext_header));
1650 /* ... and finally for the payload */
1651 crc = crc32c(crc, dp, size);
1653 return xfs_end_cksum(crc);
1657 * The bdstrat callback function for log bufs. This gives us a central
1658 * place to trap bufs in case we get hit by a log I/O error and need to
1659 * shutdown. Actually, in practice, even when we didn't get a log error,
1660 * we transition the iclogs to IOERROR state *after* flushing all existing
1661 * iclogs to disk. This is because we don't want anymore new transactions to be
1662 * started or completed afterwards.
1664 * We lock the iclogbufs here so that we can serialise against IO completion
1665 * during unmount. We might be processing a shutdown triggered during unmount,
1666 * and that can occur asynchronously to the unmount thread, and hence we need to
1667 * ensure that completes before tearing down the iclogbufs. Hence we need to
1668 * hold the buffer lock across the log IO to acheive that.
1674 struct xlog_in_core *iclog = bp->b_fspriv;
1677 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1678 xfs_buf_ioerror(bp, -EIO);
1682 * It would seem logical to return EIO here, but we rely on
1683 * the log state machine to propagate I/O errors instead of
1684 * doing it here. Similarly, IO completion will unlock the
1685 * buffer, so we don't do it here.
1695 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1696 * fashion. Previously, we should have moved the current iclog
1697 * ptr in the log to point to the next available iclog. This allows further
1698 * write to continue while this code syncs out an iclog ready to go.
1699 * Before an in-core log can be written out, the data section must be scanned
1700 * to save away the 1st word of each BBSIZE block into the header. We replace
1701 * it with the current cycle count. Each BBSIZE block is tagged with the
1702 * cycle count because there in an implicit assumption that drives will
1703 * guarantee that entire 512 byte blocks get written at once. In other words,
1704 * we can't have part of a 512 byte block written and part not written. By
1705 * tagging each block, we will know which blocks are valid when recovering
1706 * after an unclean shutdown.
1708 * This routine is single threaded on the iclog. No other thread can be in
1709 * this routine with the same iclog. Changing contents of iclog can there-
1710 * fore be done without grabbing the state machine lock. Updating the global
1711 * log will require grabbing the lock though.
1713 * The entire log manager uses a logical block numbering scheme. Only
1714 * log_sync (and then only bwrite()) know about the fact that the log may
1715 * not start with block zero on a given device. The log block start offset
1716 * is added immediately before calling bwrite().
1722 struct xlog_in_core *iclog)
1726 uint count; /* byte count of bwrite */
1727 uint count_init; /* initial count before roundup */
1728 int roundoff; /* roundoff to BB or stripe */
1729 int split = 0; /* split write into two regions */
1731 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1734 XFS_STATS_INC(xs_log_writes);
1735 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1737 /* Add for LR header */
1738 count_init = log->l_iclog_hsize + iclog->ic_offset;
1740 /* Round out the log write size */
1741 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1742 /* we have a v2 stripe unit to use */
1743 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1745 count = BBTOB(BTOBB(count_init));
1747 roundoff = count - count_init;
1748 ASSERT(roundoff >= 0);
1749 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1750 roundoff < log->l_mp->m_sb.sb_logsunit)
1752 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1753 roundoff < BBTOB(1)));
1755 /* move grant heads by roundoff in sync */
1756 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1757 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1759 /* put cycle number in every block */
1760 xlog_pack_data(log, iclog, roundoff);
1762 /* real byte length */
1763 size = iclog->ic_offset;
1766 iclog->ic_header.h_len = cpu_to_be32(size);
1769 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1771 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1773 /* Do we need to split this write into 2 parts? */
1774 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1777 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1778 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1779 iclog->ic_bwritecnt = 2;
1782 * Bump the cycle numbers at the start of each block in the
1783 * part of the iclog that ends up in the buffer that gets
1784 * written to the start of the log.
1786 * Watch out for the header magic number case, though.
1788 dptr = (char *)&iclog->ic_header + count;
1789 for (i = 0; i < split; i += BBSIZE) {
1790 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1791 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1793 *(__be32 *)dptr = cpu_to_be32(cycle);
1798 iclog->ic_bwritecnt = 1;
1801 /* calculcate the checksum */
1802 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1803 iclog->ic_datap, size);
1805 bp->b_io_length = BTOBB(count);
1806 bp->b_fspriv = iclog;
1807 XFS_BUF_ZEROFLAGS(bp);
1809 bp->b_flags |= XBF_SYNCIO;
1811 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1812 bp->b_flags |= XBF_FUA;
1815 * Flush the data device before flushing the log to make
1816 * sure all meta data written back from the AIL actually made
1817 * it to disk before stamping the new log tail LSN into the
1818 * log buffer. For an external log we need to issue the
1819 * flush explicitly, and unfortunately synchronously here;
1820 * for an internal log we can simply use the block layer
1821 * state machine for preflushes.
1823 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1824 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1826 bp->b_flags |= XBF_FLUSH;
1829 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1830 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1832 xlog_verify_iclog(log, iclog, count, true);
1834 /* account for log which doesn't start at block #0 */
1835 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1837 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1842 error = xlog_bdstrat(bp);
1844 xfs_buf_ioerror_alert(bp, "xlog_sync");
1848 bp = iclog->ic_log->l_xbuf;
1849 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1850 xfs_buf_associate_memory(bp,
1851 (char *)&iclog->ic_header + count, split);
1852 bp->b_fspriv = iclog;
1853 XFS_BUF_ZEROFLAGS(bp);
1855 bp->b_flags |= XBF_SYNCIO;
1856 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1857 bp->b_flags |= XBF_FUA;
1859 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1860 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1862 /* account for internal log which doesn't start at block #0 */
1863 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1865 error = xlog_bdstrat(bp);
1867 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1875 * Deallocate a log structure
1881 xlog_in_core_t *iclog, *next_iclog;
1884 xlog_cil_destroy(log);
1887 * Cycle all the iclogbuf locks to make sure all log IO completion
1888 * is done before we tear down these buffers.
1890 iclog = log->l_iclog;
1891 for (i = 0; i < log->l_iclog_bufs; i++) {
1892 xfs_buf_lock(iclog->ic_bp);
1893 xfs_buf_unlock(iclog->ic_bp);
1894 iclog = iclog->ic_next;
1898 * Always need to ensure that the extra buffer does not point to memory
1899 * owned by another log buffer before we free it. Also, cycle the lock
1900 * first to ensure we've completed IO on it.
1902 xfs_buf_lock(log->l_xbuf);
1903 xfs_buf_unlock(log->l_xbuf);
1904 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1905 xfs_buf_free(log->l_xbuf);
1907 iclog = log->l_iclog;
1908 for (i = 0; i < log->l_iclog_bufs; i++) {
1909 xfs_buf_free(iclog->ic_bp);
1910 next_iclog = iclog->ic_next;
1914 spinlock_destroy(&log->l_icloglock);
1916 log->l_mp->m_log = NULL;
1918 } /* xlog_dealloc_log */
1921 * Update counters atomically now that memcpy is done.
1925 xlog_state_finish_copy(
1927 struct xlog_in_core *iclog,
1931 spin_lock(&log->l_icloglock);
1933 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1934 iclog->ic_offset += copy_bytes;
1936 spin_unlock(&log->l_icloglock);
1937 } /* xlog_state_finish_copy */
1943 * print out info relating to regions written which consume
1948 struct xfs_mount *mp,
1949 struct xlog_ticket *ticket)
1952 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1954 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1955 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1976 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2020 "xlog_write: reservation summary:\n"
2021 " trans type = %s (%u)\n"
2022 " unit res = %d bytes\n"
2023 " current res = %d bytes\n"
2024 " total reg = %u bytes (o/flow = %u bytes)\n"
2025 " ophdrs = %u (ophdr space = %u bytes)\n"
2026 " ophdr + reg = %u bytes\n"
2027 " num regions = %u\n",
2028 ((ticket->t_trans_type <= 0 ||
2029 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2030 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2031 ticket->t_trans_type,
2034 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2035 ticket->t_res_num_ophdrs, ophdr_spc,
2036 ticket->t_res_arr_sum +
2037 ticket->t_res_o_flow + ophdr_spc,
2040 for (i = 0; i < ticket->t_res_num; i++) {
2041 uint r_type = ticket->t_res_arr[i].r_type;
2042 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2043 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2044 "bad-rtype" : res_type_str[r_type-1]),
2045 ticket->t_res_arr[i].r_len);
2048 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2049 "xlog_write: reservation ran out. Need to up reservation");
2050 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2054 * Calculate the potential space needed by the log vector. Each region gets
2055 * its own xlog_op_header_t and may need to be double word aligned.
2058 xlog_write_calc_vec_length(
2059 struct xlog_ticket *ticket,
2060 struct xfs_log_vec *log_vector)
2062 struct xfs_log_vec *lv;
2067 /* acct for start rec of xact */
2068 if (ticket->t_flags & XLOG_TIC_INITED)
2071 for (lv = log_vector; lv; lv = lv->lv_next) {
2072 /* we don't write ordered log vectors */
2073 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2076 headers += lv->lv_niovecs;
2078 for (i = 0; i < lv->lv_niovecs; i++) {
2079 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2082 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2086 ticket->t_res_num_ophdrs += headers;
2087 len += headers * sizeof(struct xlog_op_header);
2093 * If first write for transaction, insert start record We can't be trying to
2094 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2097 xlog_write_start_rec(
2098 struct xlog_op_header *ophdr,
2099 struct xlog_ticket *ticket)
2101 if (!(ticket->t_flags & XLOG_TIC_INITED))
2104 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2105 ophdr->oh_clientid = ticket->t_clientid;
2107 ophdr->oh_flags = XLOG_START_TRANS;
2110 ticket->t_flags &= ~XLOG_TIC_INITED;
2112 return sizeof(struct xlog_op_header);
2115 static xlog_op_header_t *
2116 xlog_write_setup_ophdr(
2118 struct xlog_op_header *ophdr,
2119 struct xlog_ticket *ticket,
2122 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2123 ophdr->oh_clientid = ticket->t_clientid;
2126 /* are we copying a commit or unmount record? */
2127 ophdr->oh_flags = flags;
2130 * We've seen logs corrupted with bad transaction client ids. This
2131 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2132 * and shut down the filesystem.
2134 switch (ophdr->oh_clientid) {
2135 case XFS_TRANSACTION:
2141 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2142 ophdr->oh_clientid, ticket);
2150 * Set up the parameters of the region copy into the log. This has
2151 * to handle region write split across multiple log buffers - this
2152 * state is kept external to this function so that this code can
2153 * be written in an obvious, self documenting manner.
2156 xlog_write_setup_copy(
2157 struct xlog_ticket *ticket,
2158 struct xlog_op_header *ophdr,
2159 int space_available,
2163 int *last_was_partial_copy,
2164 int *bytes_consumed)
2168 still_to_copy = space_required - *bytes_consumed;
2169 *copy_off = *bytes_consumed;
2171 if (still_to_copy <= space_available) {
2172 /* write of region completes here */
2173 *copy_len = still_to_copy;
2174 ophdr->oh_len = cpu_to_be32(*copy_len);
2175 if (*last_was_partial_copy)
2176 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2177 *last_was_partial_copy = 0;
2178 *bytes_consumed = 0;
2182 /* partial write of region, needs extra log op header reservation */
2183 *copy_len = space_available;
2184 ophdr->oh_len = cpu_to_be32(*copy_len);
2185 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2186 if (*last_was_partial_copy)
2187 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2188 *bytes_consumed += *copy_len;
2189 (*last_was_partial_copy)++;
2191 /* account for new log op header */
2192 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2193 ticket->t_res_num_ophdrs++;
2195 return sizeof(struct xlog_op_header);
2199 xlog_write_copy_finish(
2201 struct xlog_in_core *iclog,
2206 int *partial_copy_len,
2208 struct xlog_in_core **commit_iclog)
2210 if (*partial_copy) {
2212 * This iclog has already been marked WANT_SYNC by
2213 * xlog_state_get_iclog_space.
2215 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2218 return xlog_state_release_iclog(log, iclog);
2222 *partial_copy_len = 0;
2224 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2225 /* no more space in this iclog - push it. */
2226 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2230 spin_lock(&log->l_icloglock);
2231 xlog_state_want_sync(log, iclog);
2232 spin_unlock(&log->l_icloglock);
2235 return xlog_state_release_iclog(log, iclog);
2236 ASSERT(flags & XLOG_COMMIT_TRANS);
2237 *commit_iclog = iclog;
2244 * Write some region out to in-core log
2246 * This will be called when writing externally provided regions or when
2247 * writing out a commit record for a given transaction.
2249 * General algorithm:
2250 * 1. Find total length of this write. This may include adding to the
2251 * lengths passed in.
2252 * 2. Check whether we violate the tickets reservation.
2253 * 3. While writing to this iclog
2254 * A. Reserve as much space in this iclog as can get
2255 * B. If this is first write, save away start lsn
2256 * C. While writing this region:
2257 * 1. If first write of transaction, write start record
2258 * 2. Write log operation header (header per region)
2259 * 3. Find out if we can fit entire region into this iclog
2260 * 4. Potentially, verify destination memcpy ptr
2261 * 5. Memcpy (partial) region
2262 * 6. If partial copy, release iclog; otherwise, continue
2263 * copying more regions into current iclog
2264 * 4. Mark want sync bit (in simulation mode)
2265 * 5. Release iclog for potential flush to on-disk log.
2268 * 1. Panic if reservation is overrun. This should never happen since
2269 * reservation amounts are generated internal to the filesystem.
2271 * 1. Tickets are single threaded data structures.
2272 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2273 * syncing routine. When a single log_write region needs to span
2274 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2275 * on all log operation writes which don't contain the end of the
2276 * region. The XLOG_END_TRANS bit is used for the in-core log
2277 * operation which contains the end of the continued log_write region.
2278 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2279 * we don't really know exactly how much space will be used. As a result,
2280 * we don't update ic_offset until the end when we know exactly how many
2281 * bytes have been written out.
2286 struct xfs_log_vec *log_vector,
2287 struct xlog_ticket *ticket,
2288 xfs_lsn_t *start_lsn,
2289 struct xlog_in_core **commit_iclog,
2292 struct xlog_in_core *iclog = NULL;
2293 struct xfs_log_iovec *vecp;
2294 struct xfs_log_vec *lv;
2297 int partial_copy = 0;
2298 int partial_copy_len = 0;
2306 len = xlog_write_calc_vec_length(ticket, log_vector);
2309 * Region headers and bytes are already accounted for.
2310 * We only need to take into account start records and
2311 * split regions in this function.
2313 if (ticket->t_flags & XLOG_TIC_INITED)
2314 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2317 * Commit record headers need to be accounted for. These
2318 * come in as separate writes so are easy to detect.
2320 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2321 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2323 if (ticket->t_curr_res < 0)
2324 xlog_print_tic_res(log->l_mp, ticket);
2328 vecp = lv->lv_iovecp;
2329 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2333 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2334 &contwr, &log_offset);
2338 ASSERT(log_offset <= iclog->ic_size - 1);
2339 ptr = iclog->ic_datap + log_offset;
2341 /* start_lsn is the first lsn written to. That's all we need. */
2343 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2346 * This loop writes out as many regions as can fit in the amount
2347 * of space which was allocated by xlog_state_get_iclog_space().
2349 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2350 struct xfs_log_iovec *reg;
2351 struct xlog_op_header *ophdr;
2355 bool ordered = false;
2357 /* ordered log vectors have no regions to write */
2358 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2359 ASSERT(lv->lv_niovecs == 0);
2365 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2366 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2368 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2369 if (start_rec_copy) {
2371 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2375 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2379 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2380 sizeof(struct xlog_op_header));
2382 len += xlog_write_setup_copy(ticket, ophdr,
2383 iclog->ic_size-log_offset,
2385 ©_off, ©_len,
2388 xlog_verify_dest_ptr(log, ptr);
2391 ASSERT(copy_len >= 0);
2392 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2393 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2395 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2397 data_cnt += contwr ? copy_len : 0;
2399 error = xlog_write_copy_finish(log, iclog, flags,
2400 &record_cnt, &data_cnt,
2409 * if we had a partial copy, we need to get more iclog
2410 * space but we don't want to increment the region
2411 * index because there is still more is this region to
2414 * If we completed writing this region, and we flushed
2415 * the iclog (indicated by resetting of the record
2416 * count), then we also need to get more log space. If
2417 * this was the last record, though, we are done and
2423 if (++index == lv->lv_niovecs) {
2428 vecp = lv->lv_iovecp;
2430 if (record_cnt == 0 && ordered == false) {
2440 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2442 return xlog_state_release_iclog(log, iclog);
2444 ASSERT(flags & XLOG_COMMIT_TRANS);
2445 *commit_iclog = iclog;
2450 /*****************************************************************************
2452 * State Machine functions
2454 *****************************************************************************
2457 /* Clean iclogs starting from the head. This ordering must be
2458 * maintained, so an iclog doesn't become ACTIVE beyond one that
2459 * is SYNCING. This is also required to maintain the notion that we use
2460 * a ordered wait queue to hold off would be writers to the log when every
2461 * iclog is trying to sync to disk.
2463 * State Change: DIRTY -> ACTIVE
2466 xlog_state_clean_log(
2469 xlog_in_core_t *iclog;
2472 iclog = log->l_iclog;
2474 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2475 iclog->ic_state = XLOG_STATE_ACTIVE;
2476 iclog->ic_offset = 0;
2477 ASSERT(iclog->ic_callback == NULL);
2479 * If the number of ops in this iclog indicate it just
2480 * contains the dummy transaction, we can
2481 * change state into IDLE (the second time around).
2482 * Otherwise we should change the state into
2484 * We don't need to cover the dummy.
2487 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2492 * We have two dirty iclogs so start over
2493 * This could also be num of ops indicates
2494 * this is not the dummy going out.
2498 iclog->ic_header.h_num_logops = 0;
2499 memset(iclog->ic_header.h_cycle_data, 0,
2500 sizeof(iclog->ic_header.h_cycle_data));
2501 iclog->ic_header.h_lsn = 0;
2502 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2505 break; /* stop cleaning */
2506 iclog = iclog->ic_next;
2507 } while (iclog != log->l_iclog);
2509 /* log is locked when we are called */
2511 * Change state for the dummy log recording.
2512 * We usually go to NEED. But we go to NEED2 if the changed indicates
2513 * we are done writing the dummy record.
2514 * If we are done with the second dummy recored (DONE2), then
2518 switch (log->l_covered_state) {
2519 case XLOG_STATE_COVER_IDLE:
2520 case XLOG_STATE_COVER_NEED:
2521 case XLOG_STATE_COVER_NEED2:
2522 log->l_covered_state = XLOG_STATE_COVER_NEED;
2525 case XLOG_STATE_COVER_DONE:
2527 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2529 log->l_covered_state = XLOG_STATE_COVER_NEED;
2532 case XLOG_STATE_COVER_DONE2:
2534 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2536 log->l_covered_state = XLOG_STATE_COVER_NEED;
2543 } /* xlog_state_clean_log */
2546 xlog_get_lowest_lsn(
2549 xlog_in_core_t *lsn_log;
2550 xfs_lsn_t lowest_lsn, lsn;
2552 lsn_log = log->l_iclog;
2555 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2556 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2557 if ((lsn && !lowest_lsn) ||
2558 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2562 lsn_log = lsn_log->ic_next;
2563 } while (lsn_log != log->l_iclog);
2569 xlog_state_do_callback(
2572 struct xlog_in_core *ciclog)
2574 xlog_in_core_t *iclog;
2575 xlog_in_core_t *first_iclog; /* used to know when we've
2576 * processed all iclogs once */
2577 xfs_log_callback_t *cb, *cb_next;
2579 xfs_lsn_t lowest_lsn;
2580 int ioerrors; /* counter: iclogs with errors */
2581 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2582 int funcdidcallbacks; /* flag: function did callbacks */
2583 int repeats; /* for issuing console warnings if
2584 * looping too many times */
2587 spin_lock(&log->l_icloglock);
2588 first_iclog = iclog = log->l_iclog;
2590 funcdidcallbacks = 0;
2595 * Scan all iclogs starting with the one pointed to by the
2596 * log. Reset this starting point each time the log is
2597 * unlocked (during callbacks).
2599 * Keep looping through iclogs until one full pass is made
2600 * without running any callbacks.
2602 first_iclog = log->l_iclog;
2603 iclog = log->l_iclog;
2604 loopdidcallbacks = 0;
2609 /* skip all iclogs in the ACTIVE & DIRTY states */
2610 if (iclog->ic_state &
2611 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2612 iclog = iclog->ic_next;
2617 * Between marking a filesystem SHUTDOWN and stopping
2618 * the log, we do flush all iclogs to disk (if there
2619 * wasn't a log I/O error). So, we do want things to
2620 * go smoothly in case of just a SHUTDOWN w/o a
2623 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2625 * Can only perform callbacks in order. Since
2626 * this iclog is not in the DONE_SYNC/
2627 * DO_CALLBACK state, we skip the rest and
2628 * just try to clean up. If we set our iclog
2629 * to DO_CALLBACK, we will not process it when
2630 * we retry since a previous iclog is in the
2631 * CALLBACK and the state cannot change since
2632 * we are holding the l_icloglock.
2634 if (!(iclog->ic_state &
2635 (XLOG_STATE_DONE_SYNC |
2636 XLOG_STATE_DO_CALLBACK))) {
2637 if (ciclog && (ciclog->ic_state ==
2638 XLOG_STATE_DONE_SYNC)) {
2639 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2644 * We now have an iclog that is in either the
2645 * DO_CALLBACK or DONE_SYNC states. The other
2646 * states (WANT_SYNC, SYNCING, or CALLBACK were
2647 * caught by the above if and are going to
2648 * clean (i.e. we aren't doing their callbacks)
2653 * We will do one more check here to see if we
2654 * have chased our tail around.
2657 lowest_lsn = xlog_get_lowest_lsn(log);
2659 XFS_LSN_CMP(lowest_lsn,
2660 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2661 iclog = iclog->ic_next;
2662 continue; /* Leave this iclog for
2666 iclog->ic_state = XLOG_STATE_CALLBACK;
2670 * Completion of a iclog IO does not imply that
2671 * a transaction has completed, as transactions
2672 * can be large enough to span many iclogs. We
2673 * cannot change the tail of the log half way
2674 * through a transaction as this may be the only
2675 * transaction in the log and moving th etail to
2676 * point to the middle of it will prevent
2677 * recovery from finding the start of the
2678 * transaction. Hence we should only update the
2679 * last_sync_lsn if this iclog contains
2680 * transaction completion callbacks on it.
2682 * We have to do this before we drop the
2683 * icloglock to ensure we are the only one that
2686 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2687 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2688 if (iclog->ic_callback)
2689 atomic64_set(&log->l_last_sync_lsn,
2690 be64_to_cpu(iclog->ic_header.h_lsn));
2695 spin_unlock(&log->l_icloglock);
2698 * Keep processing entries in the callback list until
2699 * we come around and it is empty. We need to
2700 * atomically see that the list is empty and change the
2701 * state to DIRTY so that we don't miss any more
2702 * callbacks being added.
2704 spin_lock(&iclog->ic_callback_lock);
2705 cb = iclog->ic_callback;
2707 iclog->ic_callback_tail = &(iclog->ic_callback);
2708 iclog->ic_callback = NULL;
2709 spin_unlock(&iclog->ic_callback_lock);
2711 /* perform callbacks in the order given */
2712 for (; cb; cb = cb_next) {
2713 cb_next = cb->cb_next;
2714 cb->cb_func(cb->cb_arg, aborted);
2716 spin_lock(&iclog->ic_callback_lock);
2717 cb = iclog->ic_callback;
2723 spin_lock(&log->l_icloglock);
2724 ASSERT(iclog->ic_callback == NULL);
2725 spin_unlock(&iclog->ic_callback_lock);
2726 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2727 iclog->ic_state = XLOG_STATE_DIRTY;
2730 * Transition from DIRTY to ACTIVE if applicable.
2731 * NOP if STATE_IOERROR.
2733 xlog_state_clean_log(log);
2735 /* wake up threads waiting in xfs_log_force() */
2736 wake_up_all(&iclog->ic_force_wait);
2738 iclog = iclog->ic_next;
2739 } while (first_iclog != iclog);
2741 if (repeats > 5000) {
2742 flushcnt += repeats;
2745 "%s: possible infinite loop (%d iterations)",
2746 __func__, flushcnt);
2748 } while (!ioerrors && loopdidcallbacks);
2751 * make one last gasp attempt to see if iclogs are being left in
2755 if (funcdidcallbacks) {
2756 first_iclog = iclog = log->l_iclog;
2758 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2760 * Terminate the loop if iclogs are found in states
2761 * which will cause other threads to clean up iclogs.
2763 * SYNCING - i/o completion will go through logs
2764 * DONE_SYNC - interrupt thread should be waiting for
2766 * IOERROR - give up hope all ye who enter here
2768 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2769 iclog->ic_state == XLOG_STATE_SYNCING ||
2770 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2771 iclog->ic_state == XLOG_STATE_IOERROR )
2773 iclog = iclog->ic_next;
2774 } while (first_iclog != iclog);
2778 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2780 spin_unlock(&log->l_icloglock);
2783 wake_up_all(&log->l_flush_wait);
2788 * Finish transitioning this iclog to the dirty state.
2790 * Make sure that we completely execute this routine only when this is
2791 * the last call to the iclog. There is a good chance that iclog flushes,
2792 * when we reach the end of the physical log, get turned into 2 separate
2793 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2794 * routine. By using the reference count bwritecnt, we guarantee that only
2795 * the second completion goes through.
2797 * Callbacks could take time, so they are done outside the scope of the
2798 * global state machine log lock.
2801 xlog_state_done_syncing(
2802 xlog_in_core_t *iclog,
2805 struct xlog *log = iclog->ic_log;
2807 spin_lock(&log->l_icloglock);
2809 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2810 iclog->ic_state == XLOG_STATE_IOERROR);
2811 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2812 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2816 * If we got an error, either on the first buffer, or in the case of
2817 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2818 * and none should ever be attempted to be written to disk
2821 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2822 if (--iclog->ic_bwritecnt == 1) {
2823 spin_unlock(&log->l_icloglock);
2826 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2830 * Someone could be sleeping prior to writing out the next
2831 * iclog buffer, we wake them all, one will get to do the
2832 * I/O, the others get to wait for the result.
2834 wake_up_all(&iclog->ic_write_wait);
2835 spin_unlock(&log->l_icloglock);
2836 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2837 } /* xlog_state_done_syncing */
2841 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2842 * sleep. We wait on the flush queue on the head iclog as that should be
2843 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2844 * we will wait here and all new writes will sleep until a sync completes.
2846 * The in-core logs are used in a circular fashion. They are not used
2847 * out-of-order even when an iclog past the head is free.
2850 * * log_offset where xlog_write() can start writing into the in-core
2852 * * in-core log pointer to which xlog_write() should write.
2853 * * boolean indicating this is a continued write to an in-core log.
2854 * If this is the last write, then the in-core log's offset field
2855 * needs to be incremented, depending on the amount of data which
2859 xlog_state_get_iclog_space(
2862 struct xlog_in_core **iclogp,
2863 struct xlog_ticket *ticket,
2864 int *continued_write,
2868 xlog_rec_header_t *head;
2869 xlog_in_core_t *iclog;
2873 spin_lock(&log->l_icloglock);
2874 if (XLOG_FORCED_SHUTDOWN(log)) {
2875 spin_unlock(&log->l_icloglock);
2879 iclog = log->l_iclog;
2880 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2881 XFS_STATS_INC(xs_log_noiclogs);
2883 /* Wait for log writes to have flushed */
2884 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2888 head = &iclog->ic_header;
2890 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2891 log_offset = iclog->ic_offset;
2893 /* On the 1st write to an iclog, figure out lsn. This works
2894 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2895 * committing to. If the offset is set, that's how many blocks
2898 if (log_offset == 0) {
2899 ticket->t_curr_res -= log->l_iclog_hsize;
2900 xlog_tic_add_region(ticket,
2902 XLOG_REG_TYPE_LRHEADER);
2903 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2904 head->h_lsn = cpu_to_be64(
2905 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2906 ASSERT(log->l_curr_block >= 0);
2909 /* If there is enough room to write everything, then do it. Otherwise,
2910 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2911 * bit is on, so this will get flushed out. Don't update ic_offset
2912 * until you know exactly how many bytes get copied. Therefore, wait
2913 * until later to update ic_offset.
2915 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2916 * can fit into remaining data section.
2918 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2919 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2922 * If I'm the only one writing to this iclog, sync it to disk.
2923 * We need to do an atomic compare and decrement here to avoid
2924 * racing with concurrent atomic_dec_and_lock() calls in
2925 * xlog_state_release_iclog() when there is more than one
2926 * reference to the iclog.
2928 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2929 /* we are the only one */
2930 spin_unlock(&log->l_icloglock);
2931 error = xlog_state_release_iclog(log, iclog);
2935 spin_unlock(&log->l_icloglock);
2940 /* Do we have enough room to write the full amount in the remainder
2941 * of this iclog? Or must we continue a write on the next iclog and
2942 * mark this iclog as completely taken? In the case where we switch
2943 * iclogs (to mark it taken), this particular iclog will release/sync
2944 * to disk in xlog_write().
2946 if (len <= iclog->ic_size - iclog->ic_offset) {
2947 *continued_write = 0;
2948 iclog->ic_offset += len;
2950 *continued_write = 1;
2951 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2955 ASSERT(iclog->ic_offset <= iclog->ic_size);
2956 spin_unlock(&log->l_icloglock);
2958 *logoffsetp = log_offset;
2960 } /* xlog_state_get_iclog_space */
2962 /* The first cnt-1 times through here we don't need to
2963 * move the grant write head because the permanent
2964 * reservation has reserved cnt times the unit amount.
2965 * Release part of current permanent unit reservation and
2966 * reset current reservation to be one units worth. Also
2967 * move grant reservation head forward.
2970 xlog_regrant_reserve_log_space(
2972 struct xlog_ticket *ticket)
2974 trace_xfs_log_regrant_reserve_enter(log, ticket);
2976 if (ticket->t_cnt > 0)
2979 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2980 ticket->t_curr_res);
2981 xlog_grant_sub_space(log, &log->l_write_head.grant,
2982 ticket->t_curr_res);
2983 ticket->t_curr_res = ticket->t_unit_res;
2984 xlog_tic_reset_res(ticket);
2986 trace_xfs_log_regrant_reserve_sub(log, ticket);
2988 /* just return if we still have some of the pre-reserved space */
2989 if (ticket->t_cnt > 0)
2992 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2993 ticket->t_unit_res);
2995 trace_xfs_log_regrant_reserve_exit(log, ticket);
2997 ticket->t_curr_res = ticket->t_unit_res;
2998 xlog_tic_reset_res(ticket);
2999 } /* xlog_regrant_reserve_log_space */
3003 * Give back the space left from a reservation.
3005 * All the information we need to make a correct determination of space left
3006 * is present. For non-permanent reservations, things are quite easy. The
3007 * count should have been decremented to zero. We only need to deal with the
3008 * space remaining in the current reservation part of the ticket. If the
3009 * ticket contains a permanent reservation, there may be left over space which
3010 * needs to be released. A count of N means that N-1 refills of the current
3011 * reservation can be done before we need to ask for more space. The first
3012 * one goes to fill up the first current reservation. Once we run out of
3013 * space, the count will stay at zero and the only space remaining will be
3014 * in the current reservation field.
3017 xlog_ungrant_log_space(
3019 struct xlog_ticket *ticket)
3023 if (ticket->t_cnt > 0)
3026 trace_xfs_log_ungrant_enter(log, ticket);
3027 trace_xfs_log_ungrant_sub(log, ticket);
3030 * If this is a permanent reservation ticket, we may be able to free
3031 * up more space based on the remaining count.
3033 bytes = ticket->t_curr_res;
3034 if (ticket->t_cnt > 0) {
3035 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3036 bytes += ticket->t_unit_res*ticket->t_cnt;
3039 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3040 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3042 trace_xfs_log_ungrant_exit(log, ticket);
3044 xfs_log_space_wake(log->l_mp);
3048 * Flush iclog to disk if this is the last reference to the given iclog and
3049 * the WANT_SYNC bit is set.
3051 * When this function is entered, the iclog is not necessarily in the
3052 * WANT_SYNC state. It may be sitting around waiting to get filled.
3057 xlog_state_release_iclog(
3059 struct xlog_in_core *iclog)
3061 int sync = 0; /* do we sync? */
3063 if (iclog->ic_state & XLOG_STATE_IOERROR)
3066 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3067 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3070 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3071 spin_unlock(&log->l_icloglock);
3074 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3075 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3077 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3078 /* update tail before writing to iclog */
3079 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3081 iclog->ic_state = XLOG_STATE_SYNCING;
3082 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3083 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3084 /* cycle incremented when incrementing curr_block */
3086 spin_unlock(&log->l_icloglock);
3089 * We let the log lock go, so it's possible that we hit a log I/O
3090 * error or some other SHUTDOWN condition that marks the iclog
3091 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3092 * this iclog has consistent data, so we ignore IOERROR
3093 * flags after this point.
3096 return xlog_sync(log, iclog);
3098 } /* xlog_state_release_iclog */
3102 * This routine will mark the current iclog in the ring as WANT_SYNC
3103 * and move the current iclog pointer to the next iclog in the ring.
3104 * When this routine is called from xlog_state_get_iclog_space(), the
3105 * exact size of the iclog has not yet been determined. All we know is
3106 * that every data block. We have run out of space in this log record.
3109 xlog_state_switch_iclogs(
3111 struct xlog_in_core *iclog,
3114 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3116 eventual_size = iclog->ic_offset;
3117 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3118 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3119 log->l_prev_block = log->l_curr_block;
3120 log->l_prev_cycle = log->l_curr_cycle;
3122 /* roll log?: ic_offset changed later */
3123 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3125 /* Round up to next log-sunit */
3126 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3127 log->l_mp->m_sb.sb_logsunit > 1) {
3128 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3129 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3132 if (log->l_curr_block >= log->l_logBBsize) {
3133 log->l_curr_cycle++;
3134 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3135 log->l_curr_cycle++;
3136 log->l_curr_block -= log->l_logBBsize;
3137 ASSERT(log->l_curr_block >= 0);
3139 ASSERT(iclog == log->l_iclog);
3140 log->l_iclog = iclog->ic_next;
3141 } /* xlog_state_switch_iclogs */
3144 * Write out all data in the in-core log as of this exact moment in time.
3146 * Data may be written to the in-core log during this call. However,
3147 * we don't guarantee this data will be written out. A change from past
3148 * implementation means this routine will *not* write out zero length LRs.
3150 * Basically, we try and perform an intelligent scan of the in-core logs.
3151 * If we determine there is no flushable data, we just return. There is no
3152 * flushable data if:
3154 * 1. the current iclog is active and has no data; the previous iclog
3155 * is in the active or dirty state.
3156 * 2. the current iclog is drity, and the previous iclog is in the
3157 * active or dirty state.
3161 * 1. the current iclog is not in the active nor dirty state.
3162 * 2. the current iclog dirty, and the previous iclog is not in the
3163 * active nor dirty state.
3164 * 3. the current iclog is active, and there is another thread writing
3165 * to this particular iclog.
3166 * 4. a) the current iclog is active and has no other writers
3167 * b) when we return from flushing out this iclog, it is still
3168 * not in the active nor dirty state.
3172 struct xfs_mount *mp,
3176 struct xlog *log = mp->m_log;
3177 struct xlog_in_core *iclog;
3180 XFS_STATS_INC(xs_log_force);
3182 xlog_cil_force(log);
3184 spin_lock(&log->l_icloglock);
3186 iclog = log->l_iclog;
3187 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3188 spin_unlock(&log->l_icloglock);
3192 /* If the head iclog is not active nor dirty, we just attach
3193 * ourselves to the head and go to sleep.
3195 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3196 iclog->ic_state == XLOG_STATE_DIRTY) {
3198 * If the head is dirty or (active and empty), then
3199 * we need to look at the previous iclog. If the previous
3200 * iclog is active or dirty we are done. There is nothing
3201 * to sync out. Otherwise, we attach ourselves to the
3202 * previous iclog and go to sleep.
3204 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3205 (atomic_read(&iclog->ic_refcnt) == 0
3206 && iclog->ic_offset == 0)) {
3207 iclog = iclog->ic_prev;
3208 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3209 iclog->ic_state == XLOG_STATE_DIRTY)
3214 if (atomic_read(&iclog->ic_refcnt) == 0) {
3215 /* We are the only one with access to this
3216 * iclog. Flush it out now. There should
3217 * be a roundoff of zero to show that someone
3218 * has already taken care of the roundoff from
3219 * the previous sync.
3221 atomic_inc(&iclog->ic_refcnt);
3222 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3223 xlog_state_switch_iclogs(log, iclog, 0);
3224 spin_unlock(&log->l_icloglock);
3226 if (xlog_state_release_iclog(log, iclog))
3231 spin_lock(&log->l_icloglock);
3232 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3233 iclog->ic_state != XLOG_STATE_DIRTY)
3238 /* Someone else is writing to this iclog.
3239 * Use its call to flush out the data. However,
3240 * the other thread may not force out this LR,
3241 * so we mark it WANT_SYNC.
3243 xlog_state_switch_iclogs(log, iclog, 0);
3249 /* By the time we come around again, the iclog could've been filled
3250 * which would give it another lsn. If we have a new lsn, just
3251 * return because the relevant data has been flushed.
3254 if (flags & XFS_LOG_SYNC) {
3256 * We must check if we're shutting down here, before
3257 * we wait, while we're holding the l_icloglock.
3258 * Then we check again after waking up, in case our
3259 * sleep was disturbed by a bad news.
3261 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3262 spin_unlock(&log->l_icloglock);
3265 XFS_STATS_INC(xs_log_force_sleep);
3266 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3268 * No need to grab the log lock here since we're
3269 * only deciding whether or not to return EIO
3270 * and the memory read should be atomic.
3272 if (iclog->ic_state & XLOG_STATE_IOERROR)
3279 spin_unlock(&log->l_icloglock);
3285 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3286 * about errors or whether the log was flushed or not. This is the normal
3287 * interface to use when trying to unpin items or move the log forward.
3296 trace_xfs_log_force(mp, 0);
3297 error = _xfs_log_force(mp, flags, NULL);
3299 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3303 * Force the in-core log to disk for a specific LSN.
3305 * Find in-core log with lsn.
3306 * If it is in the DIRTY state, just return.
3307 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3308 * state and go to sleep or return.
3309 * If it is in any other state, go to sleep or return.
3311 * Synchronous forces are implemented with a signal variable. All callers
3312 * to force a given lsn to disk will wait on a the sv attached to the
3313 * specific in-core log. When given in-core log finally completes its
3314 * write to disk, that thread will wake up all threads waiting on the
3319 struct xfs_mount *mp,
3324 struct xlog *log = mp->m_log;
3325 struct xlog_in_core *iclog;
3326 int already_slept = 0;
3330 XFS_STATS_INC(xs_log_force);
3332 lsn = xlog_cil_force_lsn(log, lsn);
3333 if (lsn == NULLCOMMITLSN)
3337 spin_lock(&log->l_icloglock);
3338 iclog = log->l_iclog;
3339 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3340 spin_unlock(&log->l_icloglock);
3345 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3346 iclog = iclog->ic_next;
3350 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3351 spin_unlock(&log->l_icloglock);
3355 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3357 * We sleep here if we haven't already slept (e.g.
3358 * this is the first time we've looked at the correct
3359 * iclog buf) and the buffer before us is going to
3360 * be sync'ed. The reason for this is that if we
3361 * are doing sync transactions here, by waiting for
3362 * the previous I/O to complete, we can allow a few
3363 * more transactions into this iclog before we close
3366 * Otherwise, we mark the buffer WANT_SYNC, and bump
3367 * up the refcnt so we can release the log (which
3368 * drops the ref count). The state switch keeps new
3369 * transaction commits from using this buffer. When
3370 * the current commits finish writing into the buffer,
3371 * the refcount will drop to zero and the buffer will
3374 if (!already_slept &&
3375 (iclog->ic_prev->ic_state &
3376 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3377 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3379 XFS_STATS_INC(xs_log_force_sleep);
3381 xlog_wait(&iclog->ic_prev->ic_write_wait,
3388 atomic_inc(&iclog->ic_refcnt);
3389 xlog_state_switch_iclogs(log, iclog, 0);
3390 spin_unlock(&log->l_icloglock);
3391 if (xlog_state_release_iclog(log, iclog))
3395 spin_lock(&log->l_icloglock);
3398 if ((flags & XFS_LOG_SYNC) && /* sleep */
3400 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3402 * Don't wait on completion if we know that we've
3403 * gotten a log write error.
3405 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3406 spin_unlock(&log->l_icloglock);
3409 XFS_STATS_INC(xs_log_force_sleep);
3410 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3412 * No need to grab the log lock here since we're
3413 * only deciding whether or not to return EIO
3414 * and the memory read should be atomic.
3416 if (iclog->ic_state & XLOG_STATE_IOERROR)
3421 } else { /* just return */
3422 spin_unlock(&log->l_icloglock);
3426 } while (iclog != log->l_iclog);
3428 spin_unlock(&log->l_icloglock);
3433 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3434 * about errors or whether the log was flushed or not. This is the normal
3435 * interface to use when trying to unpin items or move the log forward.
3445 trace_xfs_log_force(mp, lsn);
3446 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3448 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3452 * Called when we want to mark the current iclog as being ready to sync to
3456 xlog_state_want_sync(
3458 struct xlog_in_core *iclog)
3460 assert_spin_locked(&log->l_icloglock);
3462 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3463 xlog_state_switch_iclogs(log, iclog, 0);
3465 ASSERT(iclog->ic_state &
3466 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3471 /*****************************************************************************
3475 *****************************************************************************
3479 * Free a used ticket when its refcount falls to zero.
3483 xlog_ticket_t *ticket)
3485 ASSERT(atomic_read(&ticket->t_ref) > 0);
3486 if (atomic_dec_and_test(&ticket->t_ref))
3487 kmem_zone_free(xfs_log_ticket_zone, ticket);
3492 xlog_ticket_t *ticket)
3494 ASSERT(atomic_read(&ticket->t_ref) > 0);
3495 atomic_inc(&ticket->t_ref);
3500 * Figure out the total log space unit (in bytes) that would be
3501 * required for a log ticket.
3504 xfs_log_calc_unit_res(
3505 struct xfs_mount *mp,
3508 struct xlog *log = mp->m_log;
3513 * Permanent reservations have up to 'cnt'-1 active log operations
3514 * in the log. A unit in this case is the amount of space for one
3515 * of these log operations. Normal reservations have a cnt of 1
3516 * and their unit amount is the total amount of space required.
3518 * The following lines of code account for non-transaction data
3519 * which occupy space in the on-disk log.
3521 * Normal form of a transaction is:
3522 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3523 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3525 * We need to account for all the leadup data and trailer data
3526 * around the transaction data.
3527 * And then we need to account for the worst case in terms of using
3529 * The worst case will happen if:
3530 * - the placement of the transaction happens to be such that the
3531 * roundoff is at its maximum
3532 * - the transaction data is synced before the commit record is synced
3533 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3534 * Therefore the commit record is in its own Log Record.
3535 * This can happen as the commit record is called with its
3536 * own region to xlog_write().
3537 * This then means that in the worst case, roundoff can happen for
3538 * the commit-rec as well.
3539 * The commit-rec is smaller than padding in this scenario and so it is
3540 * not added separately.
3543 /* for trans header */
3544 unit_bytes += sizeof(xlog_op_header_t);
3545 unit_bytes += sizeof(xfs_trans_header_t);
3548 unit_bytes += sizeof(xlog_op_header_t);
3551 * for LR headers - the space for data in an iclog is the size minus
3552 * the space used for the headers. If we use the iclog size, then we
3553 * undercalculate the number of headers required.
3555 * Furthermore - the addition of op headers for split-recs might
3556 * increase the space required enough to require more log and op
3557 * headers, so take that into account too.
3559 * IMPORTANT: This reservation makes the assumption that if this
3560 * transaction is the first in an iclog and hence has the LR headers
3561 * accounted to it, then the remaining space in the iclog is
3562 * exclusively for this transaction. i.e. if the transaction is larger
3563 * than the iclog, it will be the only thing in that iclog.
3564 * Fundamentally, this means we must pass the entire log vector to
3565 * xlog_write to guarantee this.
3567 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3568 num_headers = howmany(unit_bytes, iclog_space);
3570 /* for split-recs - ophdrs added when data split over LRs */
3571 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3573 /* add extra header reservations if we overrun */
3574 while (!num_headers ||
3575 howmany(unit_bytes, iclog_space) > num_headers) {
3576 unit_bytes += sizeof(xlog_op_header_t);
3579 unit_bytes += log->l_iclog_hsize * num_headers;
3581 /* for commit-rec LR header - note: padding will subsume the ophdr */
3582 unit_bytes += log->l_iclog_hsize;
3584 /* for roundoff padding for transaction data and one for commit record */
3585 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3586 /* log su roundoff */
3587 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3590 unit_bytes += 2 * BBSIZE;
3597 * Allocate and initialise a new log ticket.
3599 struct xlog_ticket *
3606 xfs_km_flags_t alloc_flags)
3608 struct xlog_ticket *tic;
3611 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3615 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3617 atomic_set(&tic->t_ref, 1);
3618 tic->t_task = current;
3619 INIT_LIST_HEAD(&tic->t_queue);
3620 tic->t_unit_res = unit_res;
3621 tic->t_curr_res = unit_res;
3624 tic->t_tid = prandom_u32();
3625 tic->t_clientid = client;
3626 tic->t_flags = XLOG_TIC_INITED;
3627 tic->t_trans_type = 0;
3629 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3631 xlog_tic_reset_res(tic);
3637 /******************************************************************************
3639 * Log debug routines
3641 ******************************************************************************
3645 * Make sure that the destination ptr is within the valid data region of
3646 * one of the iclogs. This uses backup pointers stored in a different
3647 * part of the log in case we trash the log structure.
3650 xlog_verify_dest_ptr(
3657 for (i = 0; i < log->l_iclog_bufs; i++) {
3658 if (ptr >= log->l_iclog_bak[i] &&
3659 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3664 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3668 * Check to make sure the grant write head didn't just over lap the tail. If
3669 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3670 * the cycles differ by exactly one and check the byte count.
3672 * This check is run unlocked, so can give false positives. Rather than assert
3673 * on failures, use a warn-once flag and a panic tag to allow the admin to
3674 * determine if they want to panic the machine when such an error occurs. For
3675 * debug kernels this will have the same effect as using an assert but, unlinke
3676 * an assert, it can be turned off at runtime.
3679 xlog_verify_grant_tail(
3682 int tail_cycle, tail_blocks;
3685 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3686 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3687 if (tail_cycle != cycle) {
3688 if (cycle - 1 != tail_cycle &&
3689 !(log->l_flags & XLOG_TAIL_WARN)) {
3690 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3691 "%s: cycle - 1 != tail_cycle", __func__);
3692 log->l_flags |= XLOG_TAIL_WARN;
3695 if (space > BBTOB(tail_blocks) &&
3696 !(log->l_flags & XLOG_TAIL_WARN)) {
3697 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3698 "%s: space > BBTOB(tail_blocks)", __func__);
3699 log->l_flags |= XLOG_TAIL_WARN;
3704 /* check if it will fit */
3706 xlog_verify_tail_lsn(
3708 struct xlog_in_core *iclog,
3713 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3715 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3716 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3717 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3719 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3721 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3722 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3724 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3725 if (blocks < BTOBB(iclog->ic_offset) + 1)
3726 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3728 } /* xlog_verify_tail_lsn */
3731 * Perform a number of checks on the iclog before writing to disk.
3733 * 1. Make sure the iclogs are still circular
3734 * 2. Make sure we have a good magic number
3735 * 3. Make sure we don't have magic numbers in the data
3736 * 4. Check fields of each log operation header for:
3737 * A. Valid client identifier
3738 * B. tid ptr value falls in valid ptr space (user space code)
3739 * C. Length in log record header is correct according to the
3740 * individual operation headers within record.
3741 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3742 * log, check the preceding blocks of the physical log to make sure all
3743 * the cycle numbers agree with the current cycle number.
3748 struct xlog_in_core *iclog,
3752 xlog_op_header_t *ophead;
3753 xlog_in_core_t *icptr;
3754 xlog_in_core_2_t *xhdr;
3756 xfs_caddr_t base_ptr;
3757 __psint_t field_offset;
3759 int len, i, j, k, op_len;
3762 /* check validity of iclog pointers */
3763 spin_lock(&log->l_icloglock);
3764 icptr = log->l_iclog;
3765 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3768 if (icptr != log->l_iclog)
3769 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3770 spin_unlock(&log->l_icloglock);
3772 /* check log magic numbers */
3773 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3774 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3776 ptr = (xfs_caddr_t) &iclog->ic_header;
3777 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3779 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3780 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3785 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3786 ptr = iclog->ic_datap;
3788 ophead = (xlog_op_header_t *)ptr;
3789 xhdr = iclog->ic_data;
3790 for (i = 0; i < len; i++) {
3791 ophead = (xlog_op_header_t *)ptr;
3793 /* clientid is only 1 byte */
3794 field_offset = (__psint_t)
3795 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3796 if (!syncing || (field_offset & 0x1ff)) {
3797 clientid = ophead->oh_clientid;
3799 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3800 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3801 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3802 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3803 clientid = xlog_get_client_id(
3804 xhdr[j].hic_xheader.xh_cycle_data[k]);
3806 clientid = xlog_get_client_id(
3807 iclog->ic_header.h_cycle_data[idx]);
3810 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3812 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3813 __func__, clientid, ophead,
3814 (unsigned long)field_offset);
3817 field_offset = (__psint_t)
3818 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3819 if (!syncing || (field_offset & 0x1ff)) {
3820 op_len = be32_to_cpu(ophead->oh_len);
3822 idx = BTOBBT((__psint_t)&ophead->oh_len -
3823 (__psint_t)iclog->ic_datap);
3824 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3825 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3826 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3827 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3829 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3832 ptr += sizeof(xlog_op_header_t) + op_len;
3834 } /* xlog_verify_iclog */
3838 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3844 xlog_in_core_t *iclog, *ic;
3846 iclog = log->l_iclog;
3847 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3849 * Mark all the incore logs IOERROR.
3850 * From now on, no log flushes will result.
3854 ic->ic_state = XLOG_STATE_IOERROR;
3856 } while (ic != iclog);
3860 * Return non-zero, if state transition has already happened.
3866 * This is called from xfs_force_shutdown, when we're forcibly
3867 * shutting down the filesystem, typically because of an IO error.
3868 * Our main objectives here are to make sure that:
3869 * a. if !logerror, flush the logs to disk. Anything modified
3870 * after this is ignored.
3871 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3872 * parties to find out, 'atomically'.
3873 * c. those who're sleeping on log reservations, pinned objects and
3874 * other resources get woken up, and be told the bad news.
3875 * d. nothing new gets queued up after (b) and (c) are done.
3877 * Note: for the !logerror case we need to flush the regions held in memory out
3878 * to disk first. This needs to be done before the log is marked as shutdown,
3879 * otherwise the iclog writes will fail.
3882 xfs_log_force_umount(
3883 struct xfs_mount *mp,
3892 * If this happens during log recovery, don't worry about
3893 * locking; the log isn't open for business yet.
3896 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3897 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3899 XFS_BUF_DONE(mp->m_sb_bp);
3904 * Somebody could've already done the hard work for us.
3905 * No need to get locks for this.
3907 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3908 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3913 * Flush all the completed transactions to disk before marking the log
3914 * being shut down. We need to do it in this order to ensure that
3915 * completed operations are safely on disk before we shut down, and that
3916 * we don't have to issue any buffer IO after the shutdown flags are set
3917 * to guarantee this.
3920 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3923 * mark the filesystem and the as in a shutdown state and wake
3924 * everybody up to tell them the bad news.
3926 spin_lock(&log->l_icloglock);
3927 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3929 XFS_BUF_DONE(mp->m_sb_bp);
3932 * Mark the log and the iclogs with IO error flags to prevent any
3933 * further log IO from being issued or completed.
3935 log->l_flags |= XLOG_IO_ERROR;
3936 retval = xlog_state_ioerror(log);
3937 spin_unlock(&log->l_icloglock);
3940 * We don't want anybody waiting for log reservations after this. That
3941 * means we have to wake up everybody queued up on reserveq as well as
3942 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3943 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3944 * action is protected by the grant locks.
3946 xlog_grant_head_wake_all(&log->l_reserve_head);
3947 xlog_grant_head_wake_all(&log->l_write_head);
3950 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3951 * as if the log writes were completed. The abort handling in the log
3952 * item committed callback functions will do this again under lock to
3955 wake_up_all(&log->l_cilp->xc_commit_wait);
3956 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3958 #ifdef XFSERRORDEBUG
3960 xlog_in_core_t *iclog;
3962 spin_lock(&log->l_icloglock);
3963 iclog = log->l_iclog;
3965 ASSERT(iclog->ic_callback == 0);
3966 iclog = iclog->ic_next;
3967 } while (iclog != log->l_iclog);
3968 spin_unlock(&log->l_icloglock);
3971 /* return non-zero if log IOERROR transition had already happened */
3979 xlog_in_core_t *iclog;
3981 iclog = log->l_iclog;
3983 /* endianness does not matter here, zero is zero in
3986 if (iclog->ic_header.h_num_logops)
3988 iclog = iclog->ic_next;
3989 } while (iclog != log->l_iclog);