2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_error.h"
28 #include "xfs_trans.h"
29 #include "xfs_trans_priv.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_log_recover.h"
33 #include "xfs_inode.h"
34 #include "xfs_trace.h"
35 #include "xfs_fsops.h"
36 #include "xfs_cksum.h"
38 kmem_zone_t *xfs_log_ticket_zone;
40 /* Local miscellaneous function prototypes */
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
61 struct xlog_in_core *iclog);
66 /* local state machine functions */
67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
69 xlog_state_do_callback(
72 struct xlog_in_core *iclog);
74 xlog_state_get_iclog_space(
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
82 xlog_state_release_iclog(
84 struct xlog_in_core *iclog);
86 xlog_state_switch_iclogs(
88 struct xlog_in_core *iclog,
93 struct xlog_in_core *iclog);
100 xlog_regrant_reserve_log_space(
102 struct xlog_ticket *ticket);
104 xlog_ungrant_log_space(
106 struct xlog_ticket *ticket);
110 xlog_verify_dest_ptr(
114 xlog_verify_grant_tail(
119 struct xlog_in_core *iclog,
123 xlog_verify_tail_lsn(
125 struct xlog_in_core *iclog,
128 #define xlog_verify_dest_ptr(a,b)
129 #define xlog_verify_grant_tail(a)
130 #define xlog_verify_iclog(a,b,c,d)
131 #define xlog_verify_tail_lsn(a,b,c)
139 xlog_grant_sub_space(
144 int64_t head_val = atomic64_read(head);
150 xlog_crack_grant_head_val(head_val, &cycle, &space);
154 space += log->l_logsize;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
165 xlog_grant_add_space(
170 int64_t head_val = atomic64_read(head);
177 xlog_crack_grant_head_val(head_val, &cycle, &space);
179 tmp = log->l_logsize - space;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
194 xlog_grant_head_init(
195 struct xlog_grant_head *head)
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
203 xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
206 struct xlog_ticket *tic;
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
215 xlog_ticket_reservation(
217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
225 return tic->t_unit_res * tic->t_cnt;
227 return tic->t_unit_res;
232 xlog_grant_head_wake(
234 struct xlog_grant_head *head,
237 struct xlog_ticket *tic;
240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
242 if (*free_bytes < need_bytes)
245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
247 wake_up_process(tic->t_task);
254 xlog_grant_head_wait(
256 struct xlog_grant_head *head,
257 struct xlog_ticket *tic,
258 int need_bytes) __releases(&head->lock)
259 __acquires(&head->lock)
261 list_add_tail(&tic->t_queue, &head->waiters);
264 if (XLOG_FORCED_SHUTDOWN(log))
266 xlog_grant_push_ail(log, need_bytes);
268 __set_current_state(TASK_UNINTERRUPTIBLE);
269 spin_unlock(&head->lock);
271 XFS_STATS_INC(xs_sleep_logspace);
273 trace_xfs_log_grant_sleep(log, tic);
275 trace_xfs_log_grant_wake(log, tic);
277 spin_lock(&head->lock);
278 if (XLOG_FORCED_SHUTDOWN(log))
280 } while (xlog_space_left(log, &head->grant) < need_bytes);
282 list_del_init(&tic->t_queue);
285 list_del_init(&tic->t_queue);
286 return XFS_ERROR(EIO);
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))
380 return XFS_ERROR(EIO);
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))
449 return XFS_ERROR(EIO);
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);
457 return XFS_ERROR(ENOMEM);
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 Filesystem");
623 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
624 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
627 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
628 if (IS_ERR(mp->m_log)) {
629 error = -PTR_ERR(mp->m_log);
634 * Validate the given log space and drop a critical message via syslog
635 * if the log size is too small that would lead to some unexpected
636 * situations in transaction log space reservation stage.
638 * Note: we can't just reject the mount if the validation fails. This
639 * would mean that people would have to downgrade their kernel just to
640 * remedy the situation as there is no way to grow the log (short of
641 * black magic surgery with xfs_db).
643 * We can, however, reject mounts for CRC format filesystems, as the
644 * mkfs binary being used to make the filesystem should never create a
645 * filesystem with a log that is too small.
647 min_logfsbs = xfs_log_calc_minimum_size(mp);
649 if (mp->m_sb.sb_logblocks < min_logfsbs) {
651 "Log size %d blocks too small, minimum size is %d blocks",
652 mp->m_sb.sb_logblocks, min_logfsbs);
654 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
656 "Log size %d blocks too large, maximum size is %lld blocks",
657 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
659 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
661 "log size %lld bytes too large, maximum size is %lld bytes",
662 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
667 if (xfs_sb_version_hascrc(&mp->m_sb)) {
668 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
673 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
674 "experienced then please report this message in the bug report.");
678 * Initialize the AIL now we have a log.
680 error = xfs_trans_ail_init(mp);
682 xfs_warn(mp, "AIL initialisation failed: error %d", error);
685 mp->m_log->l_ailp = mp->m_ail;
688 * skip log recovery on a norecovery mount. pretend it all
691 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
692 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
695 mp->m_flags &= ~XFS_MOUNT_RDONLY;
697 error = xlog_recover(mp->m_log);
700 mp->m_flags |= XFS_MOUNT_RDONLY;
702 xfs_warn(mp, "log mount/recovery failed: error %d",
704 goto out_destroy_ail;
708 /* Normal transactions can now occur */
709 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
712 * Now the log has been fully initialised and we know were our
713 * space grant counters are, we can initialise the permanent ticket
714 * needed for delayed logging to work.
716 xlog_cil_init_post_recovery(mp->m_log);
721 xfs_trans_ail_destroy(mp);
723 xlog_dealloc_log(mp->m_log);
729 * Finish the recovery of the file system. This is separate from the
730 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
731 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
734 * If we finish recovery successfully, start the background log work. If we are
735 * not doing recovery, then we have a RO filesystem and we don't need to start
739 xfs_log_mount_finish(xfs_mount_t *mp)
743 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
744 error = xlog_recover_finish(mp->m_log);
746 xfs_log_work_queue(mp);
748 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
756 * Final log writes as part of unmount.
758 * Mark the filesystem clean as unmount happens. Note that during relocation
759 * this routine needs to be executed as part of source-bag while the
760 * deallocation must not be done until source-end.
764 * Unmount record used to have a string "Unmount filesystem--" in the
765 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
766 * We just write the magic number now since that particular field isn't
767 * currently architecture converted and "Unmount" is a bit foo.
768 * As far as I know, there weren't any dependencies on the old behaviour.
772 xfs_log_unmount_write(xfs_mount_t *mp)
774 struct xlog *log = mp->m_log;
775 xlog_in_core_t *iclog;
777 xlog_in_core_t *first_iclog;
779 xlog_ticket_t *tic = NULL;
784 * Don't write out unmount record on read-only mounts.
785 * Or, if we are doing a forced umount (typically because of IO errors).
787 if (mp->m_flags & XFS_MOUNT_RDONLY)
790 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
791 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
794 first_iclog = iclog = log->l_iclog;
796 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
797 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
798 ASSERT(iclog->ic_offset == 0);
800 iclog = iclog->ic_next;
801 } while (iclog != first_iclog);
803 if (! (XLOG_FORCED_SHUTDOWN(log))) {
804 error = xfs_log_reserve(mp, 600, 1, &tic,
805 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
807 /* the data section must be 32 bit size aligned */
811 __uint32_t pad2; /* may as well make it 64 bits */
813 .magic = XLOG_UNMOUNT_TYPE,
815 struct xfs_log_iovec reg = {
817 .i_len = sizeof(magic),
818 .i_type = XLOG_REG_TYPE_UNMOUNT,
820 struct xfs_log_vec vec = {
825 /* remove inited flag, and account for space used */
827 tic->t_curr_res -= sizeof(magic);
828 error = xlog_write(log, &vec, tic, &lsn,
829 NULL, XLOG_UNMOUNT_TRANS);
831 * At this point, we're umounting anyway,
832 * so there's no point in transitioning log state
833 * to IOERROR. Just continue...
838 xfs_alert(mp, "%s: unmount record failed", __func__);
841 spin_lock(&log->l_icloglock);
842 iclog = log->l_iclog;
843 atomic_inc(&iclog->ic_refcnt);
844 xlog_state_want_sync(log, iclog);
845 spin_unlock(&log->l_icloglock);
846 error = xlog_state_release_iclog(log, iclog);
848 spin_lock(&log->l_icloglock);
849 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
850 iclog->ic_state == XLOG_STATE_DIRTY)) {
851 if (!XLOG_FORCED_SHUTDOWN(log)) {
852 xlog_wait(&iclog->ic_force_wait,
855 spin_unlock(&log->l_icloglock);
858 spin_unlock(&log->l_icloglock);
861 trace_xfs_log_umount_write(log, tic);
862 xlog_ungrant_log_space(log, tic);
863 xfs_log_ticket_put(tic);
867 * We're already in forced_shutdown mode, couldn't
868 * even attempt to write out the unmount transaction.
870 * Go through the motions of sync'ing and releasing
871 * the iclog, even though no I/O will actually happen,
872 * we need to wait for other log I/Os that may already
873 * be in progress. Do this as a separate section of
874 * code so we'll know if we ever get stuck here that
875 * we're in this odd situation of trying to unmount
876 * a file system that went into forced_shutdown as
877 * the result of an unmount..
879 spin_lock(&log->l_icloglock);
880 iclog = log->l_iclog;
881 atomic_inc(&iclog->ic_refcnt);
883 xlog_state_want_sync(log, iclog);
884 spin_unlock(&log->l_icloglock);
885 error = xlog_state_release_iclog(log, iclog);
887 spin_lock(&log->l_icloglock);
889 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
890 || iclog->ic_state == XLOG_STATE_DIRTY
891 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
893 xlog_wait(&iclog->ic_force_wait,
896 spin_unlock(&log->l_icloglock);
901 } /* xfs_log_unmount_write */
904 * Empty the log for unmount/freeze.
906 * To do this, we first need to shut down the background log work so it is not
907 * trying to cover the log as we clean up. We then need to unpin all objects in
908 * the log so we can then flush them out. Once they have completed their IO and
909 * run the callbacks removing themselves from the AIL, we can write the unmount
914 struct xfs_mount *mp)
916 cancel_delayed_work_sync(&mp->m_log->l_work);
917 xfs_log_force(mp, XFS_LOG_SYNC);
920 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
921 * will push it, xfs_wait_buftarg() will not wait for it. Further,
922 * xfs_buf_iowait() cannot be used because it was pushed with the
923 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
924 * the IO to complete.
926 xfs_ail_push_all_sync(mp->m_ail);
927 xfs_wait_buftarg(mp->m_ddev_targp);
928 xfs_buf_lock(mp->m_sb_bp);
929 xfs_buf_unlock(mp->m_sb_bp);
931 xfs_log_unmount_write(mp);
935 * Shut down and release the AIL and Log.
937 * During unmount, we need to ensure we flush all the dirty metadata objects
938 * from the AIL so that the log is empty before we write the unmount record to
939 * the log. Once this is done, we can tear down the AIL and the log.
943 struct xfs_mount *mp)
947 xfs_trans_ail_destroy(mp);
948 xlog_dealloc_log(mp->m_log);
953 struct xfs_mount *mp,
954 struct xfs_log_item *item,
956 const struct xfs_item_ops *ops)
958 item->li_mountp = mp;
959 item->li_ailp = mp->m_ail;
960 item->li_type = type;
964 INIT_LIST_HEAD(&item->li_ail);
965 INIT_LIST_HEAD(&item->li_cil);
969 * Wake up processes waiting for log space after we have moved the log tail.
973 struct xfs_mount *mp)
975 struct xlog *log = mp->m_log;
978 if (XLOG_FORCED_SHUTDOWN(log))
981 if (!list_empty_careful(&log->l_write_head.waiters)) {
982 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
984 spin_lock(&log->l_write_head.lock);
985 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
986 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
987 spin_unlock(&log->l_write_head.lock);
990 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
991 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
993 spin_lock(&log->l_reserve_head.lock);
994 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
995 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
996 spin_unlock(&log->l_reserve_head.lock);
1001 * Determine if we have a transaction that has gone to disk that needs to be
1002 * covered. To begin the transition to the idle state firstly the log needs to
1003 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1004 * we start attempting to cover the log.
1006 * Only if we are then in a state where covering is needed, the caller is
1007 * informed that dummy transactions are required to move the log into the idle
1010 * If there are any items in the AIl or CIL, then we do not want to attempt to
1011 * cover the log as we may be in a situation where there isn't log space
1012 * available to run a dummy transaction and this can lead to deadlocks when the
1013 * tail of the log is pinned by an item that is modified in the CIL. Hence
1014 * there's no point in running a dummy transaction at this point because we
1015 * can't start trying to idle the log until both the CIL and AIL are empty.
1018 xfs_log_need_covered(xfs_mount_t *mp)
1020 struct xlog *log = mp->m_log;
1023 if (!xfs_fs_writable(mp))
1026 if (!xlog_cil_empty(log))
1029 spin_lock(&log->l_icloglock);
1030 switch (log->l_covered_state) {
1031 case XLOG_STATE_COVER_DONE:
1032 case XLOG_STATE_COVER_DONE2:
1033 case XLOG_STATE_COVER_IDLE:
1035 case XLOG_STATE_COVER_NEED:
1036 case XLOG_STATE_COVER_NEED2:
1037 if (xfs_ail_min_lsn(log->l_ailp))
1039 if (!xlog_iclogs_empty(log))
1043 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1044 log->l_covered_state = XLOG_STATE_COVER_DONE;
1046 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1052 spin_unlock(&log->l_icloglock);
1057 * We may be holding the log iclog lock upon entering this routine.
1060 xlog_assign_tail_lsn_locked(
1061 struct xfs_mount *mp)
1063 struct xlog *log = mp->m_log;
1064 struct xfs_log_item *lip;
1067 assert_spin_locked(&mp->m_ail->xa_lock);
1070 * To make sure we always have a valid LSN for the log tail we keep
1071 * track of the last LSN which was committed in log->l_last_sync_lsn,
1072 * and use that when the AIL was empty.
1074 lip = xfs_ail_min(mp->m_ail);
1076 tail_lsn = lip->li_lsn;
1078 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1079 atomic64_set(&log->l_tail_lsn, tail_lsn);
1084 xlog_assign_tail_lsn(
1085 struct xfs_mount *mp)
1089 spin_lock(&mp->m_ail->xa_lock);
1090 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1091 spin_unlock(&mp->m_ail->xa_lock);
1097 * Return the space in the log between the tail and the head. The head
1098 * is passed in the cycle/bytes formal parms. In the special case where
1099 * the reserve head has wrapped passed the tail, this calculation is no
1100 * longer valid. In this case, just return 0 which means there is no space
1101 * in the log. This works for all places where this function is called
1102 * with the reserve head. Of course, if the write head were to ever
1103 * wrap the tail, we should blow up. Rather than catch this case here,
1104 * we depend on other ASSERTions in other parts of the code. XXXmiken
1106 * This code also handles the case where the reservation head is behind
1107 * the tail. The details of this case are described below, but the end
1108 * result is that we return the size of the log as the amount of space left.
1121 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1122 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1123 tail_bytes = BBTOB(tail_bytes);
1124 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1125 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1126 else if (tail_cycle + 1 < head_cycle)
1128 else if (tail_cycle < head_cycle) {
1129 ASSERT(tail_cycle == (head_cycle - 1));
1130 free_bytes = tail_bytes - head_bytes;
1133 * The reservation head is behind the tail.
1134 * In this case we just want to return the size of the
1135 * log as the amount of space left.
1137 xfs_alert(log->l_mp,
1138 "xlog_space_left: head behind tail\n"
1139 " tail_cycle = %d, tail_bytes = %d\n"
1140 " GH cycle = %d, GH bytes = %d",
1141 tail_cycle, tail_bytes, head_cycle, head_bytes);
1143 free_bytes = log->l_logsize;
1150 * Log function which is called when an io completes.
1152 * The log manager needs its own routine, in order to control what
1153 * happens with the buffer after the write completes.
1156 xlog_iodone(xfs_buf_t *bp)
1158 struct xlog_in_core *iclog = bp->b_fspriv;
1159 struct xlog *l = iclog->ic_log;
1163 * Race to shutdown the filesystem if we see an error.
1165 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1166 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1167 xfs_buf_ioerror_alert(bp, __func__);
1169 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1171 * This flag will be propagated to the trans-committed
1172 * callback routines to let them know that the log-commit
1175 aborted = XFS_LI_ABORTED;
1176 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1177 aborted = XFS_LI_ABORTED;
1180 /* log I/O is always issued ASYNC */
1181 ASSERT(XFS_BUF_ISASYNC(bp));
1182 xlog_state_done_syncing(iclog, aborted);
1184 * do not reference the buffer (bp) here as we could race
1185 * with it being freed after writing the unmount record to the
1191 * Return size of each in-core log record buffer.
1193 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1195 * If the filesystem blocksize is too large, we may need to choose a
1196 * larger size since the directory code currently logs entire blocks.
1200 xlog_get_iclog_buffer_size(
1201 struct xfs_mount *mp,
1207 if (mp->m_logbufs <= 0)
1208 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1210 log->l_iclog_bufs = mp->m_logbufs;
1213 * Buffer size passed in from mount system call.
1215 if (mp->m_logbsize > 0) {
1216 size = log->l_iclog_size = mp->m_logbsize;
1217 log->l_iclog_size_log = 0;
1219 log->l_iclog_size_log++;
1223 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1224 /* # headers = size / 32k
1225 * one header holds cycles from 32k of data
1228 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1229 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1231 log->l_iclog_hsize = xhdrs << BBSHIFT;
1232 log->l_iclog_heads = xhdrs;
1234 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1235 log->l_iclog_hsize = BBSIZE;
1236 log->l_iclog_heads = 1;
1241 /* All machines use 32kB buffers by default. */
1242 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1243 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1245 /* the default log size is 16k or 32k which is one header sector */
1246 log->l_iclog_hsize = BBSIZE;
1247 log->l_iclog_heads = 1;
1250 /* are we being asked to make the sizes selected above visible? */
1251 if (mp->m_logbufs == 0)
1252 mp->m_logbufs = log->l_iclog_bufs;
1253 if (mp->m_logbsize == 0)
1254 mp->m_logbsize = log->l_iclog_size;
1255 } /* xlog_get_iclog_buffer_size */
1260 struct xfs_mount *mp)
1262 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1263 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1267 * Every sync period we need to unpin all items in the AIL and push them to
1268 * disk. If there is nothing dirty, then we might need to cover the log to
1269 * indicate that the filesystem is idle.
1273 struct work_struct *work)
1275 struct xlog *log = container_of(to_delayed_work(work),
1276 struct xlog, l_work);
1277 struct xfs_mount *mp = log->l_mp;
1279 /* dgc: errors ignored - not fatal and nowhere to report them */
1280 if (xfs_log_need_covered(mp))
1281 xfs_fs_log_dummy(mp);
1283 xfs_log_force(mp, 0);
1285 /* start pushing all the metadata that is currently dirty */
1286 xfs_ail_push_all(mp->m_ail);
1288 /* queue us up again */
1289 xfs_log_work_queue(mp);
1293 * This routine initializes some of the log structure for a given mount point.
1294 * Its primary purpose is to fill in enough, so recovery can occur. However,
1295 * some other stuff may be filled in too.
1297 STATIC struct xlog *
1299 struct xfs_mount *mp,
1300 struct xfs_buftarg *log_target,
1301 xfs_daddr_t blk_offset,
1305 xlog_rec_header_t *head;
1306 xlog_in_core_t **iclogp;
1307 xlog_in_core_t *iclog, *prev_iclog=NULL;
1313 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1315 xfs_warn(mp, "Log allocation failed: No memory!");
1320 log->l_targ = log_target;
1321 log->l_logsize = BBTOB(num_bblks);
1322 log->l_logBBstart = blk_offset;
1323 log->l_logBBsize = num_bblks;
1324 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1325 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1326 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1328 log->l_prev_block = -1;
1329 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1330 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1331 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1332 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1334 xlog_grant_head_init(&log->l_reserve_head);
1335 xlog_grant_head_init(&log->l_write_head);
1337 error = EFSCORRUPTED;
1338 if (xfs_sb_version_hassector(&mp->m_sb)) {
1339 log2_size = mp->m_sb.sb_logsectlog;
1340 if (log2_size < BBSHIFT) {
1341 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1342 log2_size, BBSHIFT);
1346 log2_size -= BBSHIFT;
1347 if (log2_size > mp->m_sectbb_log) {
1348 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1349 log2_size, mp->m_sectbb_log);
1353 /* for larger sector sizes, must have v2 or external log */
1354 if (log2_size && log->l_logBBstart > 0 &&
1355 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1357 "log sector size (0x%x) invalid for configuration.",
1362 log->l_sectBBsize = 1 << log2_size;
1364 xlog_get_iclog_buffer_size(mp, log);
1367 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1370 bp->b_iodone = xlog_iodone;
1371 ASSERT(xfs_buf_islocked(bp));
1374 spin_lock_init(&log->l_icloglock);
1375 init_waitqueue_head(&log->l_flush_wait);
1377 iclogp = &log->l_iclog;
1379 * The amount of memory to allocate for the iclog structure is
1380 * rather funky due to the way the structure is defined. It is
1381 * done this way so that we can use different sizes for machines
1382 * with different amounts of memory. See the definition of
1383 * xlog_in_core_t in xfs_log_priv.h for details.
1385 ASSERT(log->l_iclog_size >= 4096);
1386 for (i=0; i < log->l_iclog_bufs; i++) {
1387 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1389 goto out_free_iclog;
1392 iclog->ic_prev = prev_iclog;
1395 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1396 BTOBB(log->l_iclog_size), 0);
1398 goto out_free_iclog;
1400 bp->b_iodone = xlog_iodone;
1402 iclog->ic_data = bp->b_addr;
1404 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1406 head = &iclog->ic_header;
1407 memset(head, 0, sizeof(xlog_rec_header_t));
1408 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1409 head->h_version = cpu_to_be32(
1410 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1411 head->h_size = cpu_to_be32(log->l_iclog_size);
1413 head->h_fmt = cpu_to_be32(XLOG_FMT);
1414 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1416 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1417 iclog->ic_state = XLOG_STATE_ACTIVE;
1418 iclog->ic_log = log;
1419 atomic_set(&iclog->ic_refcnt, 0);
1420 spin_lock_init(&iclog->ic_callback_lock);
1421 iclog->ic_callback_tail = &(iclog->ic_callback);
1422 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1424 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1425 init_waitqueue_head(&iclog->ic_force_wait);
1426 init_waitqueue_head(&iclog->ic_write_wait);
1428 iclogp = &iclog->ic_next;
1430 *iclogp = log->l_iclog; /* complete ring */
1431 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1433 error = xlog_cil_init(log);
1435 goto out_free_iclog;
1439 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1440 prev_iclog = iclog->ic_next;
1442 xfs_buf_free(iclog->ic_bp);
1445 spinlock_destroy(&log->l_icloglock);
1446 xfs_buf_free(log->l_xbuf);
1450 return ERR_PTR(-error);
1451 } /* xlog_alloc_log */
1455 * Write out the commit record of a transaction associated with the given
1456 * ticket. Return the lsn of the commit record.
1461 struct xlog_ticket *ticket,
1462 struct xlog_in_core **iclog,
1463 xfs_lsn_t *commitlsnp)
1465 struct xfs_mount *mp = log->l_mp;
1467 struct xfs_log_iovec reg = {
1470 .i_type = XLOG_REG_TYPE_COMMIT,
1472 struct xfs_log_vec vec = {
1477 ASSERT_ALWAYS(iclog);
1478 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1481 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1486 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1487 * log space. This code pushes on the lsn which would supposedly free up
1488 * the 25% which we want to leave free. We may need to adopt a policy which
1489 * pushes on an lsn which is further along in the log once we reach the high
1490 * water mark. In this manner, we would be creating a low water mark.
1493 xlog_grant_push_ail(
1497 xfs_lsn_t threshold_lsn = 0;
1498 xfs_lsn_t last_sync_lsn;
1501 int threshold_block;
1502 int threshold_cycle;
1505 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1507 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1508 free_blocks = BTOBBT(free_bytes);
1511 * Set the threshold for the minimum number of free blocks in the
1512 * log to the maximum of what the caller needs, one quarter of the
1513 * log, and 256 blocks.
1515 free_threshold = BTOBB(need_bytes);
1516 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1517 free_threshold = MAX(free_threshold, 256);
1518 if (free_blocks >= free_threshold)
1521 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1523 threshold_block += free_threshold;
1524 if (threshold_block >= log->l_logBBsize) {
1525 threshold_block -= log->l_logBBsize;
1526 threshold_cycle += 1;
1528 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1531 * Don't pass in an lsn greater than the lsn of the last
1532 * log record known to be on disk. Use a snapshot of the last sync lsn
1533 * so that it doesn't change between the compare and the set.
1535 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1536 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1537 threshold_lsn = last_sync_lsn;
1540 * Get the transaction layer to kick the dirty buffers out to
1541 * disk asynchronously. No point in trying to do this if
1542 * the filesystem is shutting down.
1544 if (!XLOG_FORCED_SHUTDOWN(log))
1545 xfs_ail_push(log->l_ailp, threshold_lsn);
1549 * Stamp cycle number in every block
1554 struct xlog_in_core *iclog,
1558 int size = iclog->ic_offset + roundoff;
1562 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1564 dp = iclog->ic_datap;
1565 for (i = 0; i < BTOBB(size); i++) {
1566 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1568 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1569 *(__be32 *)dp = cycle_lsn;
1573 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1574 xlog_in_core_2_t *xhdr = iclog->ic_data;
1576 for ( ; i < BTOBB(size); i++) {
1577 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1578 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1579 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1580 *(__be32 *)dp = cycle_lsn;
1584 for (i = 1; i < log->l_iclog_heads; i++)
1585 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1590 * Calculate the checksum for a log buffer.
1592 * This is a little more complicated than it should be because the various
1593 * headers and the actual data are non-contiguous.
1598 struct xlog_rec_header *rhead,
1604 /* first generate the crc for the record header ... */
1605 crc = xfs_start_cksum((char *)rhead,
1606 sizeof(struct xlog_rec_header),
1607 offsetof(struct xlog_rec_header, h_crc));
1609 /* ... then for additional cycle data for v2 logs ... */
1610 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1611 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1614 for (i = 1; i < log->l_iclog_heads; i++) {
1615 crc = crc32c(crc, &xhdr[i].hic_xheader,
1616 sizeof(struct xlog_rec_ext_header));
1620 /* ... and finally for the payload */
1621 crc = crc32c(crc, dp, size);
1623 return xfs_end_cksum(crc);
1627 * The bdstrat callback function for log bufs. This gives us a central
1628 * place to trap bufs in case we get hit by a log I/O error and need to
1629 * shutdown. Actually, in practice, even when we didn't get a log error,
1630 * we transition the iclogs to IOERROR state *after* flushing all existing
1631 * iclogs to disk. This is because we don't want anymore new transactions to be
1632 * started or completed afterwards.
1638 struct xlog_in_core *iclog = bp->b_fspriv;
1640 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1641 xfs_buf_ioerror(bp, EIO);
1643 xfs_buf_ioend(bp, 0);
1645 * It would seem logical to return EIO here, but we rely on
1646 * the log state machine to propagate I/O errors instead of
1652 xfs_buf_iorequest(bp);
1657 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1658 * fashion. Previously, we should have moved the current iclog
1659 * ptr in the log to point to the next available iclog. This allows further
1660 * write to continue while this code syncs out an iclog ready to go.
1661 * Before an in-core log can be written out, the data section must be scanned
1662 * to save away the 1st word of each BBSIZE block into the header. We replace
1663 * it with the current cycle count. Each BBSIZE block is tagged with the
1664 * cycle count because there in an implicit assumption that drives will
1665 * guarantee that entire 512 byte blocks get written at once. In other words,
1666 * we can't have part of a 512 byte block written and part not written. By
1667 * tagging each block, we will know which blocks are valid when recovering
1668 * after an unclean shutdown.
1670 * This routine is single threaded on the iclog. No other thread can be in
1671 * this routine with the same iclog. Changing contents of iclog can there-
1672 * fore be done without grabbing the state machine lock. Updating the global
1673 * log will require grabbing the lock though.
1675 * The entire log manager uses a logical block numbering scheme. Only
1676 * log_sync (and then only bwrite()) know about the fact that the log may
1677 * not start with block zero on a given device. The log block start offset
1678 * is added immediately before calling bwrite().
1684 struct xlog_in_core *iclog)
1688 uint count; /* byte count of bwrite */
1689 uint count_init; /* initial count before roundup */
1690 int roundoff; /* roundoff to BB or stripe */
1691 int split = 0; /* split write into two regions */
1693 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1696 XFS_STATS_INC(xs_log_writes);
1697 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1699 /* Add for LR header */
1700 count_init = log->l_iclog_hsize + iclog->ic_offset;
1702 /* Round out the log write size */
1703 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1704 /* we have a v2 stripe unit to use */
1705 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1707 count = BBTOB(BTOBB(count_init));
1709 roundoff = count - count_init;
1710 ASSERT(roundoff >= 0);
1711 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1712 roundoff < log->l_mp->m_sb.sb_logsunit)
1714 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1715 roundoff < BBTOB(1)));
1717 /* move grant heads by roundoff in sync */
1718 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1719 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1721 /* put cycle number in every block */
1722 xlog_pack_data(log, iclog, roundoff);
1724 /* real byte length */
1725 size = iclog->ic_offset;
1728 iclog->ic_header.h_len = cpu_to_be32(size);
1731 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1733 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1735 /* Do we need to split this write into 2 parts? */
1736 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1739 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1740 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1741 iclog->ic_bwritecnt = 2;
1744 * Bump the cycle numbers at the start of each block in the
1745 * part of the iclog that ends up in the buffer that gets
1746 * written to the start of the log.
1748 * Watch out for the header magic number case, though.
1750 dptr = (char *)&iclog->ic_header + count;
1751 for (i = 0; i < split; i += BBSIZE) {
1752 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1753 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1755 *(__be32 *)dptr = cpu_to_be32(cycle);
1760 iclog->ic_bwritecnt = 1;
1763 /* calculcate the checksum */
1764 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1765 iclog->ic_datap, size);
1767 bp->b_io_length = BTOBB(count);
1768 bp->b_fspriv = iclog;
1769 XFS_BUF_ZEROFLAGS(bp);
1771 bp->b_flags |= XBF_SYNCIO;
1773 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1774 bp->b_flags |= XBF_FUA;
1777 * Flush the data device before flushing the log to make
1778 * sure all meta data written back from the AIL actually made
1779 * it to disk before stamping the new log tail LSN into the
1780 * log buffer. For an external log we need to issue the
1781 * flush explicitly, and unfortunately synchronously here;
1782 * for an internal log we can simply use the block layer
1783 * state machine for preflushes.
1785 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1786 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1788 bp->b_flags |= XBF_FLUSH;
1791 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1792 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1794 xlog_verify_iclog(log, iclog, count, true);
1796 /* account for log which doesn't start at block #0 */
1797 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1799 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1804 error = xlog_bdstrat(bp);
1806 xfs_buf_ioerror_alert(bp, "xlog_sync");
1810 bp = iclog->ic_log->l_xbuf;
1811 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1812 xfs_buf_associate_memory(bp,
1813 (char *)&iclog->ic_header + count, split);
1814 bp->b_fspriv = iclog;
1815 XFS_BUF_ZEROFLAGS(bp);
1817 bp->b_flags |= XBF_SYNCIO;
1818 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1819 bp->b_flags |= XBF_FUA;
1821 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1822 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1824 /* account for internal log which doesn't start at block #0 */
1825 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1827 error = xlog_bdstrat(bp);
1829 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1837 * Deallocate a log structure
1843 xlog_in_core_t *iclog, *next_iclog;
1846 xlog_cil_destroy(log);
1849 * always need to ensure that the extra buffer does not point to memory
1850 * owned by another log buffer before we free it.
1852 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1853 xfs_buf_free(log->l_xbuf);
1855 iclog = log->l_iclog;
1856 for (i=0; i<log->l_iclog_bufs; i++) {
1857 xfs_buf_free(iclog->ic_bp);
1858 next_iclog = iclog->ic_next;
1862 spinlock_destroy(&log->l_icloglock);
1864 log->l_mp->m_log = NULL;
1866 } /* xlog_dealloc_log */
1869 * Update counters atomically now that memcpy is done.
1873 xlog_state_finish_copy(
1875 struct xlog_in_core *iclog,
1879 spin_lock(&log->l_icloglock);
1881 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1882 iclog->ic_offset += copy_bytes;
1884 spin_unlock(&log->l_icloglock);
1885 } /* xlog_state_finish_copy */
1891 * print out info relating to regions written which consume
1896 struct xfs_mount *mp,
1897 struct xlog_ticket *ticket)
1900 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1902 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1903 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1924 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1968 "xlog_write: reservation summary:\n"
1969 " trans type = %s (%u)\n"
1970 " unit res = %d bytes\n"
1971 " current res = %d bytes\n"
1972 " total reg = %u bytes (o/flow = %u bytes)\n"
1973 " ophdrs = %u (ophdr space = %u bytes)\n"
1974 " ophdr + reg = %u bytes\n"
1975 " num regions = %u\n",
1976 ((ticket->t_trans_type <= 0 ||
1977 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1978 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1979 ticket->t_trans_type,
1982 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1983 ticket->t_res_num_ophdrs, ophdr_spc,
1984 ticket->t_res_arr_sum +
1985 ticket->t_res_o_flow + ophdr_spc,
1988 for (i = 0; i < ticket->t_res_num; i++) {
1989 uint r_type = ticket->t_res_arr[i].r_type;
1990 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
1991 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1992 "bad-rtype" : res_type_str[r_type-1]),
1993 ticket->t_res_arr[i].r_len);
1996 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1997 "xlog_write: reservation ran out. Need to up reservation");
1998 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2002 * Calculate the potential space needed by the log vector. Each region gets
2003 * its own xlog_op_header_t and may need to be double word aligned.
2006 xlog_write_calc_vec_length(
2007 struct xlog_ticket *ticket,
2008 struct xfs_log_vec *log_vector)
2010 struct xfs_log_vec *lv;
2015 /* acct for start rec of xact */
2016 if (ticket->t_flags & XLOG_TIC_INITED)
2019 for (lv = log_vector; lv; lv = lv->lv_next) {
2020 /* we don't write ordered log vectors */
2021 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2024 headers += lv->lv_niovecs;
2026 for (i = 0; i < lv->lv_niovecs; i++) {
2027 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2030 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2034 ticket->t_res_num_ophdrs += headers;
2035 len += headers * sizeof(struct xlog_op_header);
2041 * If first write for transaction, insert start record We can't be trying to
2042 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2045 xlog_write_start_rec(
2046 struct xlog_op_header *ophdr,
2047 struct xlog_ticket *ticket)
2049 if (!(ticket->t_flags & XLOG_TIC_INITED))
2052 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2053 ophdr->oh_clientid = ticket->t_clientid;
2055 ophdr->oh_flags = XLOG_START_TRANS;
2058 ticket->t_flags &= ~XLOG_TIC_INITED;
2060 return sizeof(struct xlog_op_header);
2063 static xlog_op_header_t *
2064 xlog_write_setup_ophdr(
2066 struct xlog_op_header *ophdr,
2067 struct xlog_ticket *ticket,
2070 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2071 ophdr->oh_clientid = ticket->t_clientid;
2074 /* are we copying a commit or unmount record? */
2075 ophdr->oh_flags = flags;
2078 * We've seen logs corrupted with bad transaction client ids. This
2079 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2080 * and shut down the filesystem.
2082 switch (ophdr->oh_clientid) {
2083 case XFS_TRANSACTION:
2089 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2090 ophdr->oh_clientid, ticket);
2098 * Set up the parameters of the region copy into the log. This has
2099 * to handle region write split across multiple log buffers - this
2100 * state is kept external to this function so that this code can
2101 * be written in an obvious, self documenting manner.
2104 xlog_write_setup_copy(
2105 struct xlog_ticket *ticket,
2106 struct xlog_op_header *ophdr,
2107 int space_available,
2111 int *last_was_partial_copy,
2112 int *bytes_consumed)
2116 still_to_copy = space_required - *bytes_consumed;
2117 *copy_off = *bytes_consumed;
2119 if (still_to_copy <= space_available) {
2120 /* write of region completes here */
2121 *copy_len = still_to_copy;
2122 ophdr->oh_len = cpu_to_be32(*copy_len);
2123 if (*last_was_partial_copy)
2124 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2125 *last_was_partial_copy = 0;
2126 *bytes_consumed = 0;
2130 /* partial write of region, needs extra log op header reservation */
2131 *copy_len = space_available;
2132 ophdr->oh_len = cpu_to_be32(*copy_len);
2133 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2134 if (*last_was_partial_copy)
2135 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2136 *bytes_consumed += *copy_len;
2137 (*last_was_partial_copy)++;
2139 /* account for new log op header */
2140 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2141 ticket->t_res_num_ophdrs++;
2143 return sizeof(struct xlog_op_header);
2147 xlog_write_copy_finish(
2149 struct xlog_in_core *iclog,
2154 int *partial_copy_len,
2156 struct xlog_in_core **commit_iclog)
2158 if (*partial_copy) {
2160 * This iclog has already been marked WANT_SYNC by
2161 * xlog_state_get_iclog_space.
2163 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2166 return xlog_state_release_iclog(log, iclog);
2170 *partial_copy_len = 0;
2172 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2173 /* no more space in this iclog - push it. */
2174 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2178 spin_lock(&log->l_icloglock);
2179 xlog_state_want_sync(log, iclog);
2180 spin_unlock(&log->l_icloglock);
2183 return xlog_state_release_iclog(log, iclog);
2184 ASSERT(flags & XLOG_COMMIT_TRANS);
2185 *commit_iclog = iclog;
2192 * Write some region out to in-core log
2194 * This will be called when writing externally provided regions or when
2195 * writing out a commit record for a given transaction.
2197 * General algorithm:
2198 * 1. Find total length of this write. This may include adding to the
2199 * lengths passed in.
2200 * 2. Check whether we violate the tickets reservation.
2201 * 3. While writing to this iclog
2202 * A. Reserve as much space in this iclog as can get
2203 * B. If this is first write, save away start lsn
2204 * C. While writing this region:
2205 * 1. If first write of transaction, write start record
2206 * 2. Write log operation header (header per region)
2207 * 3. Find out if we can fit entire region into this iclog
2208 * 4. Potentially, verify destination memcpy ptr
2209 * 5. Memcpy (partial) region
2210 * 6. If partial copy, release iclog; otherwise, continue
2211 * copying more regions into current iclog
2212 * 4. Mark want sync bit (in simulation mode)
2213 * 5. Release iclog for potential flush to on-disk log.
2216 * 1. Panic if reservation is overrun. This should never happen since
2217 * reservation amounts are generated internal to the filesystem.
2219 * 1. Tickets are single threaded data structures.
2220 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2221 * syncing routine. When a single log_write region needs to span
2222 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2223 * on all log operation writes which don't contain the end of the
2224 * region. The XLOG_END_TRANS bit is used for the in-core log
2225 * operation which contains the end of the continued log_write region.
2226 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2227 * we don't really know exactly how much space will be used. As a result,
2228 * we don't update ic_offset until the end when we know exactly how many
2229 * bytes have been written out.
2234 struct xfs_log_vec *log_vector,
2235 struct xlog_ticket *ticket,
2236 xfs_lsn_t *start_lsn,
2237 struct xlog_in_core **commit_iclog,
2240 struct xlog_in_core *iclog = NULL;
2241 struct xfs_log_iovec *vecp;
2242 struct xfs_log_vec *lv;
2245 int partial_copy = 0;
2246 int partial_copy_len = 0;
2254 len = xlog_write_calc_vec_length(ticket, log_vector);
2257 * Region headers and bytes are already accounted for.
2258 * We only need to take into account start records and
2259 * split regions in this function.
2261 if (ticket->t_flags & XLOG_TIC_INITED)
2262 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2265 * Commit record headers need to be accounted for. These
2266 * come in as separate writes so are easy to detect.
2268 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2269 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2271 if (ticket->t_curr_res < 0)
2272 xlog_print_tic_res(log->l_mp, ticket);
2276 vecp = lv->lv_iovecp;
2277 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2281 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2282 &contwr, &log_offset);
2286 ASSERT(log_offset <= iclog->ic_size - 1);
2287 ptr = iclog->ic_datap + log_offset;
2289 /* start_lsn is the first lsn written to. That's all we need. */
2291 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2294 * This loop writes out as many regions as can fit in the amount
2295 * of space which was allocated by xlog_state_get_iclog_space().
2297 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2298 struct xfs_log_iovec *reg;
2299 struct xlog_op_header *ophdr;
2303 bool ordered = false;
2305 /* ordered log vectors have no regions to write */
2306 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2307 ASSERT(lv->lv_niovecs == 0);
2313 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2314 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2316 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2317 if (start_rec_copy) {
2319 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2323 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2325 return XFS_ERROR(EIO);
2327 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2328 sizeof(struct xlog_op_header));
2330 len += xlog_write_setup_copy(ticket, ophdr,
2331 iclog->ic_size-log_offset,
2333 ©_off, ©_len,
2336 xlog_verify_dest_ptr(log, ptr);
2339 ASSERT(copy_len >= 0);
2340 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2341 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2343 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2345 data_cnt += contwr ? copy_len : 0;
2347 error = xlog_write_copy_finish(log, iclog, flags,
2348 &record_cnt, &data_cnt,
2357 * if we had a partial copy, we need to get more iclog
2358 * space but we don't want to increment the region
2359 * index because there is still more is this region to
2362 * If we completed writing this region, and we flushed
2363 * the iclog (indicated by resetting of the record
2364 * count), then we also need to get more log space. If
2365 * this was the last record, though, we are done and
2371 if (++index == lv->lv_niovecs) {
2376 vecp = lv->lv_iovecp;
2378 if (record_cnt == 0 && ordered == false) {
2388 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2390 return xlog_state_release_iclog(log, iclog);
2392 ASSERT(flags & XLOG_COMMIT_TRANS);
2393 *commit_iclog = iclog;
2398 /*****************************************************************************
2400 * State Machine functions
2402 *****************************************************************************
2405 /* Clean iclogs starting from the head. This ordering must be
2406 * maintained, so an iclog doesn't become ACTIVE beyond one that
2407 * is SYNCING. This is also required to maintain the notion that we use
2408 * a ordered wait queue to hold off would be writers to the log when every
2409 * iclog is trying to sync to disk.
2411 * State Change: DIRTY -> ACTIVE
2414 xlog_state_clean_log(
2417 xlog_in_core_t *iclog;
2420 iclog = log->l_iclog;
2422 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2423 iclog->ic_state = XLOG_STATE_ACTIVE;
2424 iclog->ic_offset = 0;
2425 ASSERT(iclog->ic_callback == NULL);
2427 * If the number of ops in this iclog indicate it just
2428 * contains the dummy transaction, we can
2429 * change state into IDLE (the second time around).
2430 * Otherwise we should change the state into
2432 * We don't need to cover the dummy.
2435 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2440 * We have two dirty iclogs so start over
2441 * This could also be num of ops indicates
2442 * this is not the dummy going out.
2446 iclog->ic_header.h_num_logops = 0;
2447 memset(iclog->ic_header.h_cycle_data, 0,
2448 sizeof(iclog->ic_header.h_cycle_data));
2449 iclog->ic_header.h_lsn = 0;
2450 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2453 break; /* stop cleaning */
2454 iclog = iclog->ic_next;
2455 } while (iclog != log->l_iclog);
2457 /* log is locked when we are called */
2459 * Change state for the dummy log recording.
2460 * We usually go to NEED. But we go to NEED2 if the changed indicates
2461 * we are done writing the dummy record.
2462 * If we are done with the second dummy recored (DONE2), then
2466 switch (log->l_covered_state) {
2467 case XLOG_STATE_COVER_IDLE:
2468 case XLOG_STATE_COVER_NEED:
2469 case XLOG_STATE_COVER_NEED2:
2470 log->l_covered_state = XLOG_STATE_COVER_NEED;
2473 case XLOG_STATE_COVER_DONE:
2475 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2477 log->l_covered_state = XLOG_STATE_COVER_NEED;
2480 case XLOG_STATE_COVER_DONE2:
2482 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2484 log->l_covered_state = XLOG_STATE_COVER_NEED;
2491 } /* xlog_state_clean_log */
2494 xlog_get_lowest_lsn(
2497 xlog_in_core_t *lsn_log;
2498 xfs_lsn_t lowest_lsn, lsn;
2500 lsn_log = log->l_iclog;
2503 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2504 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2505 if ((lsn && !lowest_lsn) ||
2506 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2510 lsn_log = lsn_log->ic_next;
2511 } while (lsn_log != log->l_iclog);
2517 xlog_state_do_callback(
2520 struct xlog_in_core *ciclog)
2522 xlog_in_core_t *iclog;
2523 xlog_in_core_t *first_iclog; /* used to know when we've
2524 * processed all iclogs once */
2525 xfs_log_callback_t *cb, *cb_next;
2527 xfs_lsn_t lowest_lsn;
2528 int ioerrors; /* counter: iclogs with errors */
2529 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2530 int funcdidcallbacks; /* flag: function did callbacks */
2531 int repeats; /* for issuing console warnings if
2532 * looping too many times */
2535 spin_lock(&log->l_icloglock);
2536 first_iclog = iclog = log->l_iclog;
2538 funcdidcallbacks = 0;
2543 * Scan all iclogs starting with the one pointed to by the
2544 * log. Reset this starting point each time the log is
2545 * unlocked (during callbacks).
2547 * Keep looping through iclogs until one full pass is made
2548 * without running any callbacks.
2550 first_iclog = log->l_iclog;
2551 iclog = log->l_iclog;
2552 loopdidcallbacks = 0;
2557 /* skip all iclogs in the ACTIVE & DIRTY states */
2558 if (iclog->ic_state &
2559 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2560 iclog = iclog->ic_next;
2565 * Between marking a filesystem SHUTDOWN and stopping
2566 * the log, we do flush all iclogs to disk (if there
2567 * wasn't a log I/O error). So, we do want things to
2568 * go smoothly in case of just a SHUTDOWN w/o a
2571 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2573 * Can only perform callbacks in order. Since
2574 * this iclog is not in the DONE_SYNC/
2575 * DO_CALLBACK state, we skip the rest and
2576 * just try to clean up. If we set our iclog
2577 * to DO_CALLBACK, we will not process it when
2578 * we retry since a previous iclog is in the
2579 * CALLBACK and the state cannot change since
2580 * we are holding the l_icloglock.
2582 if (!(iclog->ic_state &
2583 (XLOG_STATE_DONE_SYNC |
2584 XLOG_STATE_DO_CALLBACK))) {
2585 if (ciclog && (ciclog->ic_state ==
2586 XLOG_STATE_DONE_SYNC)) {
2587 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2592 * We now have an iclog that is in either the
2593 * DO_CALLBACK or DONE_SYNC states. The other
2594 * states (WANT_SYNC, SYNCING, or CALLBACK were
2595 * caught by the above if and are going to
2596 * clean (i.e. we aren't doing their callbacks)
2601 * We will do one more check here to see if we
2602 * have chased our tail around.
2605 lowest_lsn = xlog_get_lowest_lsn(log);
2607 XFS_LSN_CMP(lowest_lsn,
2608 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2609 iclog = iclog->ic_next;
2610 continue; /* Leave this iclog for
2614 iclog->ic_state = XLOG_STATE_CALLBACK;
2618 * Completion of a iclog IO does not imply that
2619 * a transaction has completed, as transactions
2620 * can be large enough to span many iclogs. We
2621 * cannot change the tail of the log half way
2622 * through a transaction as this may be the only
2623 * transaction in the log and moving th etail to
2624 * point to the middle of it will prevent
2625 * recovery from finding the start of the
2626 * transaction. Hence we should only update the
2627 * last_sync_lsn if this iclog contains
2628 * transaction completion callbacks on it.
2630 * We have to do this before we drop the
2631 * icloglock to ensure we are the only one that
2634 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2635 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2636 if (iclog->ic_callback)
2637 atomic64_set(&log->l_last_sync_lsn,
2638 be64_to_cpu(iclog->ic_header.h_lsn));
2643 spin_unlock(&log->l_icloglock);
2646 * Keep processing entries in the callback list until
2647 * we come around and it is empty. We need to
2648 * atomically see that the list is empty and change the
2649 * state to DIRTY so that we don't miss any more
2650 * callbacks being added.
2652 spin_lock(&iclog->ic_callback_lock);
2653 cb = iclog->ic_callback;
2655 iclog->ic_callback_tail = &(iclog->ic_callback);
2656 iclog->ic_callback = NULL;
2657 spin_unlock(&iclog->ic_callback_lock);
2659 /* perform callbacks in the order given */
2660 for (; cb; cb = cb_next) {
2661 cb_next = cb->cb_next;
2662 cb->cb_func(cb->cb_arg, aborted);
2664 spin_lock(&iclog->ic_callback_lock);
2665 cb = iclog->ic_callback;
2671 spin_lock(&log->l_icloglock);
2672 ASSERT(iclog->ic_callback == NULL);
2673 spin_unlock(&iclog->ic_callback_lock);
2674 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2675 iclog->ic_state = XLOG_STATE_DIRTY;
2678 * Transition from DIRTY to ACTIVE if applicable.
2679 * NOP if STATE_IOERROR.
2681 xlog_state_clean_log(log);
2683 /* wake up threads waiting in xfs_log_force() */
2684 wake_up_all(&iclog->ic_force_wait);
2686 iclog = iclog->ic_next;
2687 } while (first_iclog != iclog);
2689 if (repeats > 5000) {
2690 flushcnt += repeats;
2693 "%s: possible infinite loop (%d iterations)",
2694 __func__, flushcnt);
2696 } while (!ioerrors && loopdidcallbacks);
2699 * make one last gasp attempt to see if iclogs are being left in
2703 if (funcdidcallbacks) {
2704 first_iclog = iclog = log->l_iclog;
2706 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2708 * Terminate the loop if iclogs are found in states
2709 * which will cause other threads to clean up iclogs.
2711 * SYNCING - i/o completion will go through logs
2712 * DONE_SYNC - interrupt thread should be waiting for
2714 * IOERROR - give up hope all ye who enter here
2716 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2717 iclog->ic_state == XLOG_STATE_SYNCING ||
2718 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2719 iclog->ic_state == XLOG_STATE_IOERROR )
2721 iclog = iclog->ic_next;
2722 } while (first_iclog != iclog);
2726 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2728 spin_unlock(&log->l_icloglock);
2731 wake_up_all(&log->l_flush_wait);
2736 * Finish transitioning this iclog to the dirty state.
2738 * Make sure that we completely execute this routine only when this is
2739 * the last call to the iclog. There is a good chance that iclog flushes,
2740 * when we reach the end of the physical log, get turned into 2 separate
2741 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2742 * routine. By using the reference count bwritecnt, we guarantee that only
2743 * the second completion goes through.
2745 * Callbacks could take time, so they are done outside the scope of the
2746 * global state machine log lock.
2749 xlog_state_done_syncing(
2750 xlog_in_core_t *iclog,
2753 struct xlog *log = iclog->ic_log;
2755 spin_lock(&log->l_icloglock);
2757 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2758 iclog->ic_state == XLOG_STATE_IOERROR);
2759 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2760 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2764 * If we got an error, either on the first buffer, or in the case of
2765 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2766 * and none should ever be attempted to be written to disk
2769 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2770 if (--iclog->ic_bwritecnt == 1) {
2771 spin_unlock(&log->l_icloglock);
2774 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2778 * Someone could be sleeping prior to writing out the next
2779 * iclog buffer, we wake them all, one will get to do the
2780 * I/O, the others get to wait for the result.
2782 wake_up_all(&iclog->ic_write_wait);
2783 spin_unlock(&log->l_icloglock);
2784 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2785 } /* xlog_state_done_syncing */
2789 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2790 * sleep. We wait on the flush queue on the head iclog as that should be
2791 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2792 * we will wait here and all new writes will sleep until a sync completes.
2794 * The in-core logs are used in a circular fashion. They are not used
2795 * out-of-order even when an iclog past the head is free.
2798 * * log_offset where xlog_write() can start writing into the in-core
2800 * * in-core log pointer to which xlog_write() should write.
2801 * * boolean indicating this is a continued write to an in-core log.
2802 * If this is the last write, then the in-core log's offset field
2803 * needs to be incremented, depending on the amount of data which
2807 xlog_state_get_iclog_space(
2810 struct xlog_in_core **iclogp,
2811 struct xlog_ticket *ticket,
2812 int *continued_write,
2816 xlog_rec_header_t *head;
2817 xlog_in_core_t *iclog;
2821 spin_lock(&log->l_icloglock);
2822 if (XLOG_FORCED_SHUTDOWN(log)) {
2823 spin_unlock(&log->l_icloglock);
2824 return XFS_ERROR(EIO);
2827 iclog = log->l_iclog;
2828 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2829 XFS_STATS_INC(xs_log_noiclogs);
2831 /* Wait for log writes to have flushed */
2832 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2836 head = &iclog->ic_header;
2838 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2839 log_offset = iclog->ic_offset;
2841 /* On the 1st write to an iclog, figure out lsn. This works
2842 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2843 * committing to. If the offset is set, that's how many blocks
2846 if (log_offset == 0) {
2847 ticket->t_curr_res -= log->l_iclog_hsize;
2848 xlog_tic_add_region(ticket,
2850 XLOG_REG_TYPE_LRHEADER);
2851 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2852 head->h_lsn = cpu_to_be64(
2853 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2854 ASSERT(log->l_curr_block >= 0);
2857 /* If there is enough room to write everything, then do it. Otherwise,
2858 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2859 * bit is on, so this will get flushed out. Don't update ic_offset
2860 * until you know exactly how many bytes get copied. Therefore, wait
2861 * until later to update ic_offset.
2863 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2864 * can fit into remaining data section.
2866 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2867 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2870 * If I'm the only one writing to this iclog, sync it to disk.
2871 * We need to do an atomic compare and decrement here to avoid
2872 * racing with concurrent atomic_dec_and_lock() calls in
2873 * xlog_state_release_iclog() when there is more than one
2874 * reference to the iclog.
2876 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2877 /* we are the only one */
2878 spin_unlock(&log->l_icloglock);
2879 error = xlog_state_release_iclog(log, iclog);
2883 spin_unlock(&log->l_icloglock);
2888 /* Do we have enough room to write the full amount in the remainder
2889 * of this iclog? Or must we continue a write on the next iclog and
2890 * mark this iclog as completely taken? In the case where we switch
2891 * iclogs (to mark it taken), this particular iclog will release/sync
2892 * to disk in xlog_write().
2894 if (len <= iclog->ic_size - iclog->ic_offset) {
2895 *continued_write = 0;
2896 iclog->ic_offset += len;
2898 *continued_write = 1;
2899 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2903 ASSERT(iclog->ic_offset <= iclog->ic_size);
2904 spin_unlock(&log->l_icloglock);
2906 *logoffsetp = log_offset;
2908 } /* xlog_state_get_iclog_space */
2910 /* The first cnt-1 times through here we don't need to
2911 * move the grant write head because the permanent
2912 * reservation has reserved cnt times the unit amount.
2913 * Release part of current permanent unit reservation and
2914 * reset current reservation to be one units worth. Also
2915 * move grant reservation head forward.
2918 xlog_regrant_reserve_log_space(
2920 struct xlog_ticket *ticket)
2922 trace_xfs_log_regrant_reserve_enter(log, ticket);
2924 if (ticket->t_cnt > 0)
2927 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2928 ticket->t_curr_res);
2929 xlog_grant_sub_space(log, &log->l_write_head.grant,
2930 ticket->t_curr_res);
2931 ticket->t_curr_res = ticket->t_unit_res;
2932 xlog_tic_reset_res(ticket);
2934 trace_xfs_log_regrant_reserve_sub(log, ticket);
2936 /* just return if we still have some of the pre-reserved space */
2937 if (ticket->t_cnt > 0)
2940 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2941 ticket->t_unit_res);
2943 trace_xfs_log_regrant_reserve_exit(log, ticket);
2945 ticket->t_curr_res = ticket->t_unit_res;
2946 xlog_tic_reset_res(ticket);
2947 } /* xlog_regrant_reserve_log_space */
2951 * Give back the space left from a reservation.
2953 * All the information we need to make a correct determination of space left
2954 * is present. For non-permanent reservations, things are quite easy. The
2955 * count should have been decremented to zero. We only need to deal with the
2956 * space remaining in the current reservation part of the ticket. If the
2957 * ticket contains a permanent reservation, there may be left over space which
2958 * needs to be released. A count of N means that N-1 refills of the current
2959 * reservation can be done before we need to ask for more space. The first
2960 * one goes to fill up the first current reservation. Once we run out of
2961 * space, the count will stay at zero and the only space remaining will be
2962 * in the current reservation field.
2965 xlog_ungrant_log_space(
2967 struct xlog_ticket *ticket)
2971 if (ticket->t_cnt > 0)
2974 trace_xfs_log_ungrant_enter(log, ticket);
2975 trace_xfs_log_ungrant_sub(log, ticket);
2978 * If this is a permanent reservation ticket, we may be able to free
2979 * up more space based on the remaining count.
2981 bytes = ticket->t_curr_res;
2982 if (ticket->t_cnt > 0) {
2983 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2984 bytes += ticket->t_unit_res*ticket->t_cnt;
2987 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2988 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2990 trace_xfs_log_ungrant_exit(log, ticket);
2992 xfs_log_space_wake(log->l_mp);
2996 * Flush iclog to disk if this is the last reference to the given iclog and
2997 * the WANT_SYNC bit is set.
2999 * When this function is entered, the iclog is not necessarily in the
3000 * WANT_SYNC state. It may be sitting around waiting to get filled.
3005 xlog_state_release_iclog(
3007 struct xlog_in_core *iclog)
3009 int sync = 0; /* do we sync? */
3011 if (iclog->ic_state & XLOG_STATE_IOERROR)
3012 return XFS_ERROR(EIO);
3014 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3015 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3018 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3019 spin_unlock(&log->l_icloglock);
3020 return XFS_ERROR(EIO);
3022 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3023 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3025 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3026 /* update tail before writing to iclog */
3027 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3029 iclog->ic_state = XLOG_STATE_SYNCING;
3030 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3031 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3032 /* cycle incremented when incrementing curr_block */
3034 spin_unlock(&log->l_icloglock);
3037 * We let the log lock go, so it's possible that we hit a log I/O
3038 * error or some other SHUTDOWN condition that marks the iclog
3039 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3040 * this iclog has consistent data, so we ignore IOERROR
3041 * flags after this point.
3044 return xlog_sync(log, iclog);
3046 } /* xlog_state_release_iclog */
3050 * This routine will mark the current iclog in the ring as WANT_SYNC
3051 * and move the current iclog pointer to the next iclog in the ring.
3052 * When this routine is called from xlog_state_get_iclog_space(), the
3053 * exact size of the iclog has not yet been determined. All we know is
3054 * that every data block. We have run out of space in this log record.
3057 xlog_state_switch_iclogs(
3059 struct xlog_in_core *iclog,
3062 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3064 eventual_size = iclog->ic_offset;
3065 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3066 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3067 log->l_prev_block = log->l_curr_block;
3068 log->l_prev_cycle = log->l_curr_cycle;
3070 /* roll log?: ic_offset changed later */
3071 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3073 /* Round up to next log-sunit */
3074 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3075 log->l_mp->m_sb.sb_logsunit > 1) {
3076 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3077 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3080 if (log->l_curr_block >= log->l_logBBsize) {
3081 log->l_curr_cycle++;
3082 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3083 log->l_curr_cycle++;
3084 log->l_curr_block -= log->l_logBBsize;
3085 ASSERT(log->l_curr_block >= 0);
3087 ASSERT(iclog == log->l_iclog);
3088 log->l_iclog = iclog->ic_next;
3089 } /* xlog_state_switch_iclogs */
3092 * Write out all data in the in-core log as of this exact moment in time.
3094 * Data may be written to the in-core log during this call. However,
3095 * we don't guarantee this data will be written out. A change from past
3096 * implementation means this routine will *not* write out zero length LRs.
3098 * Basically, we try and perform an intelligent scan of the in-core logs.
3099 * If we determine there is no flushable data, we just return. There is no
3100 * flushable data if:
3102 * 1. the current iclog is active and has no data; the previous iclog
3103 * is in the active or dirty state.
3104 * 2. the current iclog is drity, and the previous iclog is in the
3105 * active or dirty state.
3109 * 1. the current iclog is not in the active nor dirty state.
3110 * 2. the current iclog dirty, and the previous iclog is not in the
3111 * active nor dirty state.
3112 * 3. the current iclog is active, and there is another thread writing
3113 * to this particular iclog.
3114 * 4. a) the current iclog is active and has no other writers
3115 * b) when we return from flushing out this iclog, it is still
3116 * not in the active nor dirty state.
3120 struct xfs_mount *mp,
3124 struct xlog *log = mp->m_log;
3125 struct xlog_in_core *iclog;
3128 XFS_STATS_INC(xs_log_force);
3130 xlog_cil_force(log);
3132 spin_lock(&log->l_icloglock);
3134 iclog = log->l_iclog;
3135 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3136 spin_unlock(&log->l_icloglock);
3137 return XFS_ERROR(EIO);
3140 /* If the head iclog is not active nor dirty, we just attach
3141 * ourselves to the head and go to sleep.
3143 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3144 iclog->ic_state == XLOG_STATE_DIRTY) {
3146 * If the head is dirty or (active and empty), then
3147 * we need to look at the previous iclog. If the previous
3148 * iclog is active or dirty we are done. There is nothing
3149 * to sync out. Otherwise, we attach ourselves to the
3150 * previous iclog and go to sleep.
3152 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3153 (atomic_read(&iclog->ic_refcnt) == 0
3154 && iclog->ic_offset == 0)) {
3155 iclog = iclog->ic_prev;
3156 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3157 iclog->ic_state == XLOG_STATE_DIRTY)
3162 if (atomic_read(&iclog->ic_refcnt) == 0) {
3163 /* We are the only one with access to this
3164 * iclog. Flush it out now. There should
3165 * be a roundoff of zero to show that someone
3166 * has already taken care of the roundoff from
3167 * the previous sync.
3169 atomic_inc(&iclog->ic_refcnt);
3170 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3171 xlog_state_switch_iclogs(log, iclog, 0);
3172 spin_unlock(&log->l_icloglock);
3174 if (xlog_state_release_iclog(log, iclog))
3175 return XFS_ERROR(EIO);
3179 spin_lock(&log->l_icloglock);
3180 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3181 iclog->ic_state != XLOG_STATE_DIRTY)
3186 /* Someone else is writing to this iclog.
3187 * Use its call to flush out the data. However,
3188 * the other thread may not force out this LR,
3189 * so we mark it WANT_SYNC.
3191 xlog_state_switch_iclogs(log, iclog, 0);
3197 /* By the time we come around again, the iclog could've been filled
3198 * which would give it another lsn. If we have a new lsn, just
3199 * return because the relevant data has been flushed.
3202 if (flags & XFS_LOG_SYNC) {
3204 * We must check if we're shutting down here, before
3205 * we wait, while we're holding the l_icloglock.
3206 * Then we check again after waking up, in case our
3207 * sleep was disturbed by a bad news.
3209 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3210 spin_unlock(&log->l_icloglock);
3211 return XFS_ERROR(EIO);
3213 XFS_STATS_INC(xs_log_force_sleep);
3214 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3216 * No need to grab the log lock here since we're
3217 * only deciding whether or not to return EIO
3218 * and the memory read should be atomic.
3220 if (iclog->ic_state & XLOG_STATE_IOERROR)
3221 return XFS_ERROR(EIO);
3227 spin_unlock(&log->l_icloglock);
3233 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3234 * about errors or whether the log was flushed or not. This is the normal
3235 * interface to use when trying to unpin items or move the log forward.
3244 trace_xfs_log_force(mp, 0);
3245 error = _xfs_log_force(mp, flags, NULL);
3247 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3251 * Force the in-core log to disk for a specific LSN.
3253 * Find in-core log with lsn.
3254 * If it is in the DIRTY state, just return.
3255 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3256 * state and go to sleep or return.
3257 * If it is in any other state, go to sleep or return.
3259 * Synchronous forces are implemented with a signal variable. All callers
3260 * to force a given lsn to disk will wait on a the sv attached to the
3261 * specific in-core log. When given in-core log finally completes its
3262 * write to disk, that thread will wake up all threads waiting on the
3267 struct xfs_mount *mp,
3272 struct xlog *log = mp->m_log;
3273 struct xlog_in_core *iclog;
3274 int already_slept = 0;
3278 XFS_STATS_INC(xs_log_force);
3280 lsn = xlog_cil_force_lsn(log, lsn);
3281 if (lsn == NULLCOMMITLSN)
3285 spin_lock(&log->l_icloglock);
3286 iclog = log->l_iclog;
3287 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3288 spin_unlock(&log->l_icloglock);
3289 return XFS_ERROR(EIO);
3293 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3294 iclog = iclog->ic_next;
3298 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3299 spin_unlock(&log->l_icloglock);
3303 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3305 * We sleep here if we haven't already slept (e.g.
3306 * this is the first time we've looked at the correct
3307 * iclog buf) and the buffer before us is going to
3308 * be sync'ed. The reason for this is that if we
3309 * are doing sync transactions here, by waiting for
3310 * the previous I/O to complete, we can allow a few
3311 * more transactions into this iclog before we close
3314 * Otherwise, we mark the buffer WANT_SYNC, and bump
3315 * up the refcnt so we can release the log (which
3316 * drops the ref count). The state switch keeps new
3317 * transaction commits from using this buffer. When
3318 * the current commits finish writing into the buffer,
3319 * the refcount will drop to zero and the buffer will
3322 if (!already_slept &&
3323 (iclog->ic_prev->ic_state &
3324 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3325 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3327 XFS_STATS_INC(xs_log_force_sleep);
3329 xlog_wait(&iclog->ic_prev->ic_write_wait,
3336 atomic_inc(&iclog->ic_refcnt);
3337 xlog_state_switch_iclogs(log, iclog, 0);
3338 spin_unlock(&log->l_icloglock);
3339 if (xlog_state_release_iclog(log, iclog))
3340 return XFS_ERROR(EIO);
3343 spin_lock(&log->l_icloglock);
3346 if ((flags & XFS_LOG_SYNC) && /* sleep */
3348 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3350 * Don't wait on completion if we know that we've
3351 * gotten a log write error.
3353 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3354 spin_unlock(&log->l_icloglock);
3355 return XFS_ERROR(EIO);
3357 XFS_STATS_INC(xs_log_force_sleep);
3358 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3360 * No need to grab the log lock here since we're
3361 * only deciding whether or not to return EIO
3362 * and the memory read should be atomic.
3364 if (iclog->ic_state & XLOG_STATE_IOERROR)
3365 return XFS_ERROR(EIO);
3369 } else { /* just return */
3370 spin_unlock(&log->l_icloglock);
3374 } while (iclog != log->l_iclog);
3376 spin_unlock(&log->l_icloglock);
3381 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3382 * about errors or whether the log was flushed or not. This is the normal
3383 * interface to use when trying to unpin items or move the log forward.
3393 trace_xfs_log_force(mp, lsn);
3394 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3396 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3400 * Called when we want to mark the current iclog as being ready to sync to
3404 xlog_state_want_sync(
3406 struct xlog_in_core *iclog)
3408 assert_spin_locked(&log->l_icloglock);
3410 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3411 xlog_state_switch_iclogs(log, iclog, 0);
3413 ASSERT(iclog->ic_state &
3414 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3419 /*****************************************************************************
3423 *****************************************************************************
3427 * Free a used ticket when its refcount falls to zero.
3431 xlog_ticket_t *ticket)
3433 ASSERT(atomic_read(&ticket->t_ref) > 0);
3434 if (atomic_dec_and_test(&ticket->t_ref))
3435 kmem_zone_free(xfs_log_ticket_zone, ticket);
3440 xlog_ticket_t *ticket)
3442 ASSERT(atomic_read(&ticket->t_ref) > 0);
3443 atomic_inc(&ticket->t_ref);
3448 * Figure out the total log space unit (in bytes) that would be
3449 * required for a log ticket.
3452 xfs_log_calc_unit_res(
3453 struct xfs_mount *mp,
3456 struct xlog *log = mp->m_log;
3461 * Permanent reservations have up to 'cnt'-1 active log operations
3462 * in the log. A unit in this case is the amount of space for one
3463 * of these log operations. Normal reservations have a cnt of 1
3464 * and their unit amount is the total amount of space required.
3466 * The following lines of code account for non-transaction data
3467 * which occupy space in the on-disk log.
3469 * Normal form of a transaction is:
3470 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3471 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3473 * We need to account for all the leadup data and trailer data
3474 * around the transaction data.
3475 * And then we need to account for the worst case in terms of using
3477 * The worst case will happen if:
3478 * - the placement of the transaction happens to be such that the
3479 * roundoff is at its maximum
3480 * - the transaction data is synced before the commit record is synced
3481 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3482 * Therefore the commit record is in its own Log Record.
3483 * This can happen as the commit record is called with its
3484 * own region to xlog_write().
3485 * This then means that in the worst case, roundoff can happen for
3486 * the commit-rec as well.
3487 * The commit-rec is smaller than padding in this scenario and so it is
3488 * not added separately.
3491 /* for trans header */
3492 unit_bytes += sizeof(xlog_op_header_t);
3493 unit_bytes += sizeof(xfs_trans_header_t);
3496 unit_bytes += sizeof(xlog_op_header_t);
3499 * for LR headers - the space for data in an iclog is the size minus
3500 * the space used for the headers. If we use the iclog size, then we
3501 * undercalculate the number of headers required.
3503 * Furthermore - the addition of op headers for split-recs might
3504 * increase the space required enough to require more log and op
3505 * headers, so take that into account too.
3507 * IMPORTANT: This reservation makes the assumption that if this
3508 * transaction is the first in an iclog and hence has the LR headers
3509 * accounted to it, then the remaining space in the iclog is
3510 * exclusively for this transaction. i.e. if the transaction is larger
3511 * than the iclog, it will be the only thing in that iclog.
3512 * Fundamentally, this means we must pass the entire log vector to
3513 * xlog_write to guarantee this.
3515 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3516 num_headers = howmany(unit_bytes, iclog_space);
3518 /* for split-recs - ophdrs added when data split over LRs */
3519 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3521 /* add extra header reservations if we overrun */
3522 while (!num_headers ||
3523 howmany(unit_bytes, iclog_space) > num_headers) {
3524 unit_bytes += sizeof(xlog_op_header_t);
3527 unit_bytes += log->l_iclog_hsize * num_headers;
3529 /* for commit-rec LR header - note: padding will subsume the ophdr */
3530 unit_bytes += log->l_iclog_hsize;
3532 /* for roundoff padding for transaction data and one for commit record */
3533 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3534 /* log su roundoff */
3535 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3538 unit_bytes += 2 * BBSIZE;
3545 * Allocate and initialise a new log ticket.
3547 struct xlog_ticket *
3554 xfs_km_flags_t alloc_flags)
3556 struct xlog_ticket *tic;
3559 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3563 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3565 atomic_set(&tic->t_ref, 1);
3566 tic->t_task = current;
3567 INIT_LIST_HEAD(&tic->t_queue);
3568 tic->t_unit_res = unit_res;
3569 tic->t_curr_res = unit_res;
3572 tic->t_tid = prandom_u32();
3573 tic->t_clientid = client;
3574 tic->t_flags = XLOG_TIC_INITED;
3575 tic->t_trans_type = 0;
3577 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3579 xlog_tic_reset_res(tic);
3585 /******************************************************************************
3587 * Log debug routines
3589 ******************************************************************************
3593 * Make sure that the destination ptr is within the valid data region of
3594 * one of the iclogs. This uses backup pointers stored in a different
3595 * part of the log in case we trash the log structure.
3598 xlog_verify_dest_ptr(
3605 for (i = 0; i < log->l_iclog_bufs; i++) {
3606 if (ptr >= log->l_iclog_bak[i] &&
3607 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3612 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3616 * Check to make sure the grant write head didn't just over lap the tail. If
3617 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3618 * the cycles differ by exactly one and check the byte count.
3620 * This check is run unlocked, so can give false positives. Rather than assert
3621 * on failures, use a warn-once flag and a panic tag to allow the admin to
3622 * determine if they want to panic the machine when such an error occurs. For
3623 * debug kernels this will have the same effect as using an assert but, unlinke
3624 * an assert, it can be turned off at runtime.
3627 xlog_verify_grant_tail(
3630 int tail_cycle, tail_blocks;
3633 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3634 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3635 if (tail_cycle != cycle) {
3636 if (cycle - 1 != tail_cycle &&
3637 !(log->l_flags & XLOG_TAIL_WARN)) {
3638 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3639 "%s: cycle - 1 != tail_cycle", __func__);
3640 log->l_flags |= XLOG_TAIL_WARN;
3643 if (space > BBTOB(tail_blocks) &&
3644 !(log->l_flags & XLOG_TAIL_WARN)) {
3645 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3646 "%s: space > BBTOB(tail_blocks)", __func__);
3647 log->l_flags |= XLOG_TAIL_WARN;
3652 /* check if it will fit */
3654 xlog_verify_tail_lsn(
3656 struct xlog_in_core *iclog,
3661 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3663 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3664 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3665 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3667 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3669 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3670 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3672 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3673 if (blocks < BTOBB(iclog->ic_offset) + 1)
3674 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3676 } /* xlog_verify_tail_lsn */
3679 * Perform a number of checks on the iclog before writing to disk.
3681 * 1. Make sure the iclogs are still circular
3682 * 2. Make sure we have a good magic number
3683 * 3. Make sure we don't have magic numbers in the data
3684 * 4. Check fields of each log operation header for:
3685 * A. Valid client identifier
3686 * B. tid ptr value falls in valid ptr space (user space code)
3687 * C. Length in log record header is correct according to the
3688 * individual operation headers within record.
3689 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3690 * log, check the preceding blocks of the physical log to make sure all
3691 * the cycle numbers agree with the current cycle number.
3696 struct xlog_in_core *iclog,
3700 xlog_op_header_t *ophead;
3701 xlog_in_core_t *icptr;
3702 xlog_in_core_2_t *xhdr;
3704 xfs_caddr_t base_ptr;
3705 __psint_t field_offset;
3707 int len, i, j, k, op_len;
3710 /* check validity of iclog pointers */
3711 spin_lock(&log->l_icloglock);
3712 icptr = log->l_iclog;
3713 for (i=0; i < log->l_iclog_bufs; i++) {
3715 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3716 icptr = icptr->ic_next;
3718 if (icptr != log->l_iclog)
3719 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3720 spin_unlock(&log->l_icloglock);
3722 /* check log magic numbers */
3723 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3724 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3726 ptr = (xfs_caddr_t) &iclog->ic_header;
3727 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3729 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3730 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3735 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3736 ptr = iclog->ic_datap;
3738 ophead = (xlog_op_header_t *)ptr;
3739 xhdr = iclog->ic_data;
3740 for (i = 0; i < len; i++) {
3741 ophead = (xlog_op_header_t *)ptr;
3743 /* clientid is only 1 byte */
3744 field_offset = (__psint_t)
3745 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3746 if (!syncing || (field_offset & 0x1ff)) {
3747 clientid = ophead->oh_clientid;
3749 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3750 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3751 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3752 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3753 clientid = xlog_get_client_id(
3754 xhdr[j].hic_xheader.xh_cycle_data[k]);
3756 clientid = xlog_get_client_id(
3757 iclog->ic_header.h_cycle_data[idx]);
3760 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3762 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3763 __func__, clientid, ophead,
3764 (unsigned long)field_offset);
3767 field_offset = (__psint_t)
3768 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3769 if (!syncing || (field_offset & 0x1ff)) {
3770 op_len = be32_to_cpu(ophead->oh_len);
3772 idx = BTOBBT((__psint_t)&ophead->oh_len -
3773 (__psint_t)iclog->ic_datap);
3774 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3775 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3776 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3777 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3779 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3782 ptr += sizeof(xlog_op_header_t) + op_len;
3784 } /* xlog_verify_iclog */
3788 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3794 xlog_in_core_t *iclog, *ic;
3796 iclog = log->l_iclog;
3797 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3799 * Mark all the incore logs IOERROR.
3800 * From now on, no log flushes will result.
3804 ic->ic_state = XLOG_STATE_IOERROR;
3806 } while (ic != iclog);
3810 * Return non-zero, if state transition has already happened.
3816 * This is called from xfs_force_shutdown, when we're forcibly
3817 * shutting down the filesystem, typically because of an IO error.
3818 * Our main objectives here are to make sure that:
3819 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3820 * parties to find out, 'atomically'.
3821 * b. those who're sleeping on log reservations, pinned objects and
3822 * other resources get woken up, and be told the bad news.
3823 * c. nothing new gets queued up after (a) and (b) are done.
3824 * d. if !logerror, flush the iclogs to disk, then seal them off
3827 * Note: for delayed logging the !logerror case needs to flush the regions
3828 * held in memory out to the iclogs before flushing them to disk. This needs
3829 * to be done before the log is marked as shutdown, otherwise the flush to the
3833 xfs_log_force_umount(
3834 struct xfs_mount *mp,
3843 * If this happens during log recovery, don't worry about
3844 * locking; the log isn't open for business yet.
3847 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3848 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3850 XFS_BUF_DONE(mp->m_sb_bp);
3855 * Somebody could've already done the hard work for us.
3856 * No need to get locks for this.
3858 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3859 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3865 * Flush the in memory commit item list before marking the log as
3866 * being shut down. We need to do it in this order to ensure all the
3867 * completed transactions are flushed to disk with the xfs_log_force()
3871 xlog_cil_force(log);
3874 * mark the filesystem and the as in a shutdown state and wake
3875 * everybody up to tell them the bad news.
3877 spin_lock(&log->l_icloglock);
3878 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3880 XFS_BUF_DONE(mp->m_sb_bp);
3883 * This flag is sort of redundant because of the mount flag, but
3884 * it's good to maintain the separation between the log and the rest
3887 log->l_flags |= XLOG_IO_ERROR;
3890 * If we hit a log error, we want to mark all the iclogs IOERROR
3891 * while we're still holding the loglock.
3894 retval = xlog_state_ioerror(log);
3895 spin_unlock(&log->l_icloglock);
3898 * We don't want anybody waiting for log reservations after this. That
3899 * means we have to wake up everybody queued up on reserveq as well as
3900 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3901 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3902 * action is protected by the grant locks.
3904 xlog_grant_head_wake_all(&log->l_reserve_head);
3905 xlog_grant_head_wake_all(&log->l_write_head);
3907 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3910 * Force the incore logs to disk before shutting the
3911 * log down completely.
3913 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3915 spin_lock(&log->l_icloglock);
3916 retval = xlog_state_ioerror(log);
3917 spin_unlock(&log->l_icloglock);
3920 * Wake up everybody waiting on xfs_log_force.
3921 * Callback all log item committed functions as if the
3922 * log writes were completed.
3924 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3926 #ifdef XFSERRORDEBUG
3928 xlog_in_core_t *iclog;
3930 spin_lock(&log->l_icloglock);
3931 iclog = log->l_iclog;
3933 ASSERT(iclog->ic_callback == 0);
3934 iclog = iclog->ic_next;
3935 } while (iclog != log->l_iclog);
3936 spin_unlock(&log->l_icloglock);
3939 /* return non-zero if log IOERROR transition had already happened */
3947 xlog_in_core_t *iclog;
3949 iclog = log->l_iclog;
3951 /* endianness does not matter here, zero is zero in
3954 if (iclog->ic_header.h_num_logops)
3956 iclog = iclog->ic_next;
3957 } while (iclog != log->l_iclog);