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 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1080 atomic64_set(&log->l_tail_lsn, tail_lsn);
1085 xlog_assign_tail_lsn(
1086 struct xfs_mount *mp)
1090 spin_lock(&mp->m_ail->xa_lock);
1091 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1092 spin_unlock(&mp->m_ail->xa_lock);
1098 * Return the space in the log between the tail and the head. The head
1099 * is passed in the cycle/bytes formal parms. In the special case where
1100 * the reserve head has wrapped passed the tail, this calculation is no
1101 * longer valid. In this case, just return 0 which means there is no space
1102 * in the log. This works for all places where this function is called
1103 * with the reserve head. Of course, if the write head were to ever
1104 * wrap the tail, we should blow up. Rather than catch this case here,
1105 * we depend on other ASSERTions in other parts of the code. XXXmiken
1107 * This code also handles the case where the reservation head is behind
1108 * the tail. The details of this case are described below, but the end
1109 * result is that we return the size of the log as the amount of space left.
1122 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1123 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1124 tail_bytes = BBTOB(tail_bytes);
1125 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1126 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1127 else if (tail_cycle + 1 < head_cycle)
1129 else if (tail_cycle < head_cycle) {
1130 ASSERT(tail_cycle == (head_cycle - 1));
1131 free_bytes = tail_bytes - head_bytes;
1134 * The reservation head is behind the tail.
1135 * In this case we just want to return the size of the
1136 * log as the amount of space left.
1138 xfs_alert(log->l_mp,
1139 "xlog_space_left: head behind tail\n"
1140 " tail_cycle = %d, tail_bytes = %d\n"
1141 " GH cycle = %d, GH bytes = %d",
1142 tail_cycle, tail_bytes, head_cycle, head_bytes);
1144 free_bytes = log->l_logsize;
1151 * Log function which is called when an io completes.
1153 * The log manager needs its own routine, in order to control what
1154 * happens with the buffer after the write completes.
1157 xlog_iodone(xfs_buf_t *bp)
1159 struct xlog_in_core *iclog = bp->b_fspriv;
1160 struct xlog *l = iclog->ic_log;
1164 * Race to shutdown the filesystem if we see an error.
1166 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1167 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1168 xfs_buf_ioerror_alert(bp, __func__);
1170 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1172 * This flag will be propagated to the trans-committed
1173 * callback routines to let them know that the log-commit
1176 aborted = XFS_LI_ABORTED;
1177 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1178 aborted = XFS_LI_ABORTED;
1181 /* log I/O is always issued ASYNC */
1182 ASSERT(XFS_BUF_ISASYNC(bp));
1183 xlog_state_done_syncing(iclog, aborted);
1186 * drop the buffer lock now that we are done. Nothing references
1187 * the buffer after this, so an unmount waiting on this lock can now
1188 * tear it down safely. As such, it is unsafe to reference the buffer
1189 * (bp) after the unlock as we could race with it being freed.
1195 * Return size of each in-core log record buffer.
1197 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1199 * If the filesystem blocksize is too large, we may need to choose a
1200 * larger size since the directory code currently logs entire blocks.
1204 xlog_get_iclog_buffer_size(
1205 struct xfs_mount *mp,
1211 if (mp->m_logbufs <= 0)
1212 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1214 log->l_iclog_bufs = mp->m_logbufs;
1217 * Buffer size passed in from mount system call.
1219 if (mp->m_logbsize > 0) {
1220 size = log->l_iclog_size = mp->m_logbsize;
1221 log->l_iclog_size_log = 0;
1223 log->l_iclog_size_log++;
1227 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1228 /* # headers = size / 32k
1229 * one header holds cycles from 32k of data
1232 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1233 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1235 log->l_iclog_hsize = xhdrs << BBSHIFT;
1236 log->l_iclog_heads = xhdrs;
1238 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1239 log->l_iclog_hsize = BBSIZE;
1240 log->l_iclog_heads = 1;
1245 /* All machines use 32kB buffers by default. */
1246 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1247 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1249 /* the default log size is 16k or 32k which is one header sector */
1250 log->l_iclog_hsize = BBSIZE;
1251 log->l_iclog_heads = 1;
1254 /* are we being asked to make the sizes selected above visible? */
1255 if (mp->m_logbufs == 0)
1256 mp->m_logbufs = log->l_iclog_bufs;
1257 if (mp->m_logbsize == 0)
1258 mp->m_logbsize = log->l_iclog_size;
1259 } /* xlog_get_iclog_buffer_size */
1264 struct xfs_mount *mp)
1266 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1267 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1271 * Every sync period we need to unpin all items in the AIL and push them to
1272 * disk. If there is nothing dirty, then we might need to cover the log to
1273 * indicate that the filesystem is idle.
1277 struct work_struct *work)
1279 struct xlog *log = container_of(to_delayed_work(work),
1280 struct xlog, l_work);
1281 struct xfs_mount *mp = log->l_mp;
1283 /* dgc: errors ignored - not fatal and nowhere to report them */
1284 if (xfs_log_need_covered(mp))
1285 xfs_fs_log_dummy(mp);
1287 xfs_log_force(mp, 0);
1289 /* start pushing all the metadata that is currently dirty */
1290 xfs_ail_push_all(mp->m_ail);
1292 /* queue us up again */
1293 xfs_log_work_queue(mp);
1297 * This routine initializes some of the log structure for a given mount point.
1298 * Its primary purpose is to fill in enough, so recovery can occur. However,
1299 * some other stuff may be filled in too.
1301 STATIC struct xlog *
1303 struct xfs_mount *mp,
1304 struct xfs_buftarg *log_target,
1305 xfs_daddr_t blk_offset,
1309 xlog_rec_header_t *head;
1310 xlog_in_core_t **iclogp;
1311 xlog_in_core_t *iclog, *prev_iclog=NULL;
1317 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1319 xfs_warn(mp, "Log allocation failed: No memory!");
1324 log->l_targ = log_target;
1325 log->l_logsize = BBTOB(num_bblks);
1326 log->l_logBBstart = blk_offset;
1327 log->l_logBBsize = num_bblks;
1328 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1329 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1330 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1332 log->l_prev_block = -1;
1333 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1334 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1335 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1336 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1338 xlog_grant_head_init(&log->l_reserve_head);
1339 xlog_grant_head_init(&log->l_write_head);
1341 error = EFSCORRUPTED;
1342 if (xfs_sb_version_hassector(&mp->m_sb)) {
1343 log2_size = mp->m_sb.sb_logsectlog;
1344 if (log2_size < BBSHIFT) {
1345 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1346 log2_size, BBSHIFT);
1350 log2_size -= BBSHIFT;
1351 if (log2_size > mp->m_sectbb_log) {
1352 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1353 log2_size, mp->m_sectbb_log);
1357 /* for larger sector sizes, must have v2 or external log */
1358 if (log2_size && log->l_logBBstart > 0 &&
1359 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1361 "log sector size (0x%x) invalid for configuration.",
1366 log->l_sectBBsize = 1 << log2_size;
1368 xlog_get_iclog_buffer_size(mp, log);
1371 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1376 * The iclogbuf buffer locks are held over IO but we are not going to do
1377 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1378 * when appropriately.
1380 ASSERT(xfs_buf_islocked(bp));
1383 bp->b_iodone = xlog_iodone;
1386 spin_lock_init(&log->l_icloglock);
1387 init_waitqueue_head(&log->l_flush_wait);
1389 iclogp = &log->l_iclog;
1391 * The amount of memory to allocate for the iclog structure is
1392 * rather funky due to the way the structure is defined. It is
1393 * done this way so that we can use different sizes for machines
1394 * with different amounts of memory. See the definition of
1395 * xlog_in_core_t in xfs_log_priv.h for details.
1397 ASSERT(log->l_iclog_size >= 4096);
1398 for (i=0; i < log->l_iclog_bufs; i++) {
1399 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1401 goto out_free_iclog;
1404 iclog->ic_prev = prev_iclog;
1407 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1408 BTOBB(log->l_iclog_size), 0);
1410 goto out_free_iclog;
1412 ASSERT(xfs_buf_islocked(bp));
1415 bp->b_iodone = xlog_iodone;
1417 iclog->ic_data = bp->b_addr;
1419 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1421 head = &iclog->ic_header;
1422 memset(head, 0, sizeof(xlog_rec_header_t));
1423 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1424 head->h_version = cpu_to_be32(
1425 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1426 head->h_size = cpu_to_be32(log->l_iclog_size);
1428 head->h_fmt = cpu_to_be32(XLOG_FMT);
1429 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1431 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1432 iclog->ic_state = XLOG_STATE_ACTIVE;
1433 iclog->ic_log = log;
1434 atomic_set(&iclog->ic_refcnt, 0);
1435 spin_lock_init(&iclog->ic_callback_lock);
1436 iclog->ic_callback_tail = &(iclog->ic_callback);
1437 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1439 init_waitqueue_head(&iclog->ic_force_wait);
1440 init_waitqueue_head(&iclog->ic_write_wait);
1442 iclogp = &iclog->ic_next;
1444 *iclogp = log->l_iclog; /* complete ring */
1445 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1447 error = xlog_cil_init(log);
1449 goto out_free_iclog;
1453 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1454 prev_iclog = iclog->ic_next;
1456 xfs_buf_free(iclog->ic_bp);
1459 spinlock_destroy(&log->l_icloglock);
1460 xfs_buf_free(log->l_xbuf);
1464 return ERR_PTR(-error);
1465 } /* xlog_alloc_log */
1469 * Write out the commit record of a transaction associated with the given
1470 * ticket. Return the lsn of the commit record.
1475 struct xlog_ticket *ticket,
1476 struct xlog_in_core **iclog,
1477 xfs_lsn_t *commitlsnp)
1479 struct xfs_mount *mp = log->l_mp;
1481 struct xfs_log_iovec reg = {
1484 .i_type = XLOG_REG_TYPE_COMMIT,
1486 struct xfs_log_vec vec = {
1491 ASSERT_ALWAYS(iclog);
1492 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1495 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1500 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1501 * log space. This code pushes on the lsn which would supposedly free up
1502 * the 25% which we want to leave free. We may need to adopt a policy which
1503 * pushes on an lsn which is further along in the log once we reach the high
1504 * water mark. In this manner, we would be creating a low water mark.
1507 xlog_grant_push_ail(
1511 xfs_lsn_t threshold_lsn = 0;
1512 xfs_lsn_t last_sync_lsn;
1515 int threshold_block;
1516 int threshold_cycle;
1519 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1521 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1522 free_blocks = BTOBBT(free_bytes);
1525 * Set the threshold for the minimum number of free blocks in the
1526 * log to the maximum of what the caller needs, one quarter of the
1527 * log, and 256 blocks.
1529 free_threshold = BTOBB(need_bytes);
1530 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1531 free_threshold = MAX(free_threshold, 256);
1532 if (free_blocks >= free_threshold)
1535 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1537 threshold_block += free_threshold;
1538 if (threshold_block >= log->l_logBBsize) {
1539 threshold_block -= log->l_logBBsize;
1540 threshold_cycle += 1;
1542 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1545 * Don't pass in an lsn greater than the lsn of the last
1546 * log record known to be on disk. Use a snapshot of the last sync lsn
1547 * so that it doesn't change between the compare and the set.
1549 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1550 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1551 threshold_lsn = last_sync_lsn;
1554 * Get the transaction layer to kick the dirty buffers out to
1555 * disk asynchronously. No point in trying to do this if
1556 * the filesystem is shutting down.
1558 if (!XLOG_FORCED_SHUTDOWN(log))
1559 xfs_ail_push(log->l_ailp, threshold_lsn);
1563 * Stamp cycle number in every block
1568 struct xlog_in_core *iclog,
1572 int size = iclog->ic_offset + roundoff;
1576 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1578 dp = iclog->ic_datap;
1579 for (i = 0; i < BTOBB(size); i++) {
1580 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1582 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1583 *(__be32 *)dp = cycle_lsn;
1587 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1588 xlog_in_core_2_t *xhdr = iclog->ic_data;
1590 for ( ; i < BTOBB(size); i++) {
1591 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1592 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1593 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1594 *(__be32 *)dp = cycle_lsn;
1598 for (i = 1; i < log->l_iclog_heads; i++)
1599 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1604 * Calculate the checksum for a log buffer.
1606 * This is a little more complicated than it should be because the various
1607 * headers and the actual data are non-contiguous.
1612 struct xlog_rec_header *rhead,
1618 /* first generate the crc for the record header ... */
1619 crc = xfs_start_cksum((char *)rhead,
1620 sizeof(struct xlog_rec_header),
1621 offsetof(struct xlog_rec_header, h_crc));
1623 /* ... then for additional cycle data for v2 logs ... */
1624 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1625 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1628 for (i = 1; i < log->l_iclog_heads; i++) {
1629 crc = crc32c(crc, &xhdr[i].hic_xheader,
1630 sizeof(struct xlog_rec_ext_header));
1634 /* ... and finally for the payload */
1635 crc = crc32c(crc, dp, size);
1637 return xfs_end_cksum(crc);
1641 * The bdstrat callback function for log bufs. This gives us a central
1642 * place to trap bufs in case we get hit by a log I/O error and need to
1643 * shutdown. Actually, in practice, even when we didn't get a log error,
1644 * we transition the iclogs to IOERROR state *after* flushing all existing
1645 * iclogs to disk. This is because we don't want anymore new transactions to be
1646 * started or completed afterwards.
1648 * We lock the iclogbufs here so that we can serialise against IO completion
1649 * during unmount. We might be processing a shutdown triggered during unmount,
1650 * and that can occur asynchronously to the unmount thread, and hence we need to
1651 * ensure that completes before tearing down the iclogbufs. Hence we need to
1652 * hold the buffer lock across the log IO to acheive that.
1658 struct xlog_in_core *iclog = bp->b_fspriv;
1661 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1662 xfs_buf_ioerror(bp, EIO);
1664 xfs_buf_ioend(bp, 0);
1666 * It would seem logical to return EIO here, but we rely on
1667 * the log state machine to propagate I/O errors instead of
1668 * doing it here. Similarly, IO completion will unlock the
1669 * buffer, so we don't do it here.
1674 xfs_buf_iorequest(bp);
1679 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1680 * fashion. Previously, we should have moved the current iclog
1681 * ptr in the log to point to the next available iclog. This allows further
1682 * write to continue while this code syncs out an iclog ready to go.
1683 * Before an in-core log can be written out, the data section must be scanned
1684 * to save away the 1st word of each BBSIZE block into the header. We replace
1685 * it with the current cycle count. Each BBSIZE block is tagged with the
1686 * cycle count because there in an implicit assumption that drives will
1687 * guarantee that entire 512 byte blocks get written at once. In other words,
1688 * we can't have part of a 512 byte block written and part not written. By
1689 * tagging each block, we will know which blocks are valid when recovering
1690 * after an unclean shutdown.
1692 * This routine is single threaded on the iclog. No other thread can be in
1693 * this routine with the same iclog. Changing contents of iclog can there-
1694 * fore be done without grabbing the state machine lock. Updating the global
1695 * log will require grabbing the lock though.
1697 * The entire log manager uses a logical block numbering scheme. Only
1698 * log_sync (and then only bwrite()) know about the fact that the log may
1699 * not start with block zero on a given device. The log block start offset
1700 * is added immediately before calling bwrite().
1706 struct xlog_in_core *iclog)
1710 uint count; /* byte count of bwrite */
1711 uint count_init; /* initial count before roundup */
1712 int roundoff; /* roundoff to BB or stripe */
1713 int split = 0; /* split write into two regions */
1715 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1718 XFS_STATS_INC(xs_log_writes);
1719 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1721 /* Add for LR header */
1722 count_init = log->l_iclog_hsize + iclog->ic_offset;
1724 /* Round out the log write size */
1725 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1726 /* we have a v2 stripe unit to use */
1727 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1729 count = BBTOB(BTOBB(count_init));
1731 roundoff = count - count_init;
1732 ASSERT(roundoff >= 0);
1733 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1734 roundoff < log->l_mp->m_sb.sb_logsunit)
1736 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1737 roundoff < BBTOB(1)));
1739 /* move grant heads by roundoff in sync */
1740 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1741 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1743 /* put cycle number in every block */
1744 xlog_pack_data(log, iclog, roundoff);
1746 /* real byte length */
1747 size = iclog->ic_offset;
1750 iclog->ic_header.h_len = cpu_to_be32(size);
1753 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1755 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1757 /* Do we need to split this write into 2 parts? */
1758 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1761 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1762 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1763 iclog->ic_bwritecnt = 2;
1766 * Bump the cycle numbers at the start of each block in the
1767 * part of the iclog that ends up in the buffer that gets
1768 * written to the start of the log.
1770 * Watch out for the header magic number case, though.
1772 dptr = (char *)&iclog->ic_header + count;
1773 for (i = 0; i < split; i += BBSIZE) {
1774 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1775 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1777 *(__be32 *)dptr = cpu_to_be32(cycle);
1782 iclog->ic_bwritecnt = 1;
1785 /* calculcate the checksum */
1786 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1787 iclog->ic_datap, size);
1789 bp->b_io_length = BTOBB(count);
1790 bp->b_fspriv = iclog;
1791 XFS_BUF_ZEROFLAGS(bp);
1793 bp->b_flags |= XBF_SYNCIO;
1795 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1796 bp->b_flags |= XBF_FUA;
1799 * Flush the data device before flushing the log to make
1800 * sure all meta data written back from the AIL actually made
1801 * it to disk before stamping the new log tail LSN into the
1802 * log buffer. For an external log we need to issue the
1803 * flush explicitly, and unfortunately synchronously here;
1804 * for an internal log we can simply use the block layer
1805 * state machine for preflushes.
1807 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1808 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1810 bp->b_flags |= XBF_FLUSH;
1813 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1814 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1816 xlog_verify_iclog(log, iclog, count, true);
1818 /* account for log which doesn't start at block #0 */
1819 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1821 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1826 error = xlog_bdstrat(bp);
1828 xfs_buf_ioerror_alert(bp, "xlog_sync");
1832 bp = iclog->ic_log->l_xbuf;
1833 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1834 xfs_buf_associate_memory(bp,
1835 (char *)&iclog->ic_header + count, split);
1836 bp->b_fspriv = iclog;
1837 XFS_BUF_ZEROFLAGS(bp);
1839 bp->b_flags |= XBF_SYNCIO;
1840 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1841 bp->b_flags |= XBF_FUA;
1843 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1844 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1846 /* account for internal log which doesn't start at block #0 */
1847 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1849 error = xlog_bdstrat(bp);
1851 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1859 * Deallocate a log structure
1865 xlog_in_core_t *iclog, *next_iclog;
1868 xlog_cil_destroy(log);
1871 * Cycle all the iclogbuf locks to make sure all log IO completion
1872 * is done before we tear down these buffers.
1874 iclog = log->l_iclog;
1875 for (i = 0; i < log->l_iclog_bufs; i++) {
1876 xfs_buf_lock(iclog->ic_bp);
1877 xfs_buf_unlock(iclog->ic_bp);
1878 iclog = iclog->ic_next;
1882 * Always need to ensure that the extra buffer does not point to memory
1883 * owned by another log buffer before we free it. Also, cycle the lock
1884 * first to ensure we've completed IO on it.
1886 xfs_buf_lock(log->l_xbuf);
1887 xfs_buf_unlock(log->l_xbuf);
1888 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1889 xfs_buf_free(log->l_xbuf);
1891 iclog = log->l_iclog;
1892 for (i = 0; i < log->l_iclog_bufs; i++) {
1893 xfs_buf_free(iclog->ic_bp);
1894 next_iclog = iclog->ic_next;
1898 spinlock_destroy(&log->l_icloglock);
1900 log->l_mp->m_log = NULL;
1902 } /* xlog_dealloc_log */
1905 * Update counters atomically now that memcpy is done.
1909 xlog_state_finish_copy(
1911 struct xlog_in_core *iclog,
1915 spin_lock(&log->l_icloglock);
1917 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1918 iclog->ic_offset += copy_bytes;
1920 spin_unlock(&log->l_icloglock);
1921 } /* xlog_state_finish_copy */
1927 * print out info relating to regions written which consume
1932 struct xfs_mount *mp,
1933 struct xlog_ticket *ticket)
1936 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1938 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1939 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1960 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2004 "xlog_write: reservation summary:\n"
2005 " trans type = %s (%u)\n"
2006 " unit res = %d bytes\n"
2007 " current res = %d bytes\n"
2008 " total reg = %u bytes (o/flow = %u bytes)\n"
2009 " ophdrs = %u (ophdr space = %u bytes)\n"
2010 " ophdr + reg = %u bytes\n"
2011 " num regions = %u\n",
2012 ((ticket->t_trans_type <= 0 ||
2013 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2014 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2015 ticket->t_trans_type,
2018 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2019 ticket->t_res_num_ophdrs, ophdr_spc,
2020 ticket->t_res_arr_sum +
2021 ticket->t_res_o_flow + ophdr_spc,
2024 for (i = 0; i < ticket->t_res_num; i++) {
2025 uint r_type = ticket->t_res_arr[i].r_type;
2026 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2027 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2028 "bad-rtype" : res_type_str[r_type-1]),
2029 ticket->t_res_arr[i].r_len);
2032 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2033 "xlog_write: reservation ran out. Need to up reservation");
2034 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2038 * Calculate the potential space needed by the log vector. Each region gets
2039 * its own xlog_op_header_t and may need to be double word aligned.
2042 xlog_write_calc_vec_length(
2043 struct xlog_ticket *ticket,
2044 struct xfs_log_vec *log_vector)
2046 struct xfs_log_vec *lv;
2051 /* acct for start rec of xact */
2052 if (ticket->t_flags & XLOG_TIC_INITED)
2055 for (lv = log_vector; lv; lv = lv->lv_next) {
2056 /* we don't write ordered log vectors */
2057 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2060 headers += lv->lv_niovecs;
2062 for (i = 0; i < lv->lv_niovecs; i++) {
2063 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2066 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2070 ticket->t_res_num_ophdrs += headers;
2071 len += headers * sizeof(struct xlog_op_header);
2077 * If first write for transaction, insert start record We can't be trying to
2078 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2081 xlog_write_start_rec(
2082 struct xlog_op_header *ophdr,
2083 struct xlog_ticket *ticket)
2085 if (!(ticket->t_flags & XLOG_TIC_INITED))
2088 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2089 ophdr->oh_clientid = ticket->t_clientid;
2091 ophdr->oh_flags = XLOG_START_TRANS;
2094 ticket->t_flags &= ~XLOG_TIC_INITED;
2096 return sizeof(struct xlog_op_header);
2099 static xlog_op_header_t *
2100 xlog_write_setup_ophdr(
2102 struct xlog_op_header *ophdr,
2103 struct xlog_ticket *ticket,
2106 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2107 ophdr->oh_clientid = ticket->t_clientid;
2110 /* are we copying a commit or unmount record? */
2111 ophdr->oh_flags = flags;
2114 * We've seen logs corrupted with bad transaction client ids. This
2115 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2116 * and shut down the filesystem.
2118 switch (ophdr->oh_clientid) {
2119 case XFS_TRANSACTION:
2125 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2126 ophdr->oh_clientid, ticket);
2134 * Set up the parameters of the region copy into the log. This has
2135 * to handle region write split across multiple log buffers - this
2136 * state is kept external to this function so that this code can
2137 * be written in an obvious, self documenting manner.
2140 xlog_write_setup_copy(
2141 struct xlog_ticket *ticket,
2142 struct xlog_op_header *ophdr,
2143 int space_available,
2147 int *last_was_partial_copy,
2148 int *bytes_consumed)
2152 still_to_copy = space_required - *bytes_consumed;
2153 *copy_off = *bytes_consumed;
2155 if (still_to_copy <= space_available) {
2156 /* write of region completes here */
2157 *copy_len = still_to_copy;
2158 ophdr->oh_len = cpu_to_be32(*copy_len);
2159 if (*last_was_partial_copy)
2160 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2161 *last_was_partial_copy = 0;
2162 *bytes_consumed = 0;
2166 /* partial write of region, needs extra log op header reservation */
2167 *copy_len = space_available;
2168 ophdr->oh_len = cpu_to_be32(*copy_len);
2169 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2170 if (*last_was_partial_copy)
2171 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2172 *bytes_consumed += *copy_len;
2173 (*last_was_partial_copy)++;
2175 /* account for new log op header */
2176 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2177 ticket->t_res_num_ophdrs++;
2179 return sizeof(struct xlog_op_header);
2183 xlog_write_copy_finish(
2185 struct xlog_in_core *iclog,
2190 int *partial_copy_len,
2192 struct xlog_in_core **commit_iclog)
2194 if (*partial_copy) {
2196 * This iclog has already been marked WANT_SYNC by
2197 * xlog_state_get_iclog_space.
2199 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2202 return xlog_state_release_iclog(log, iclog);
2206 *partial_copy_len = 0;
2208 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2209 /* no more space in this iclog - push it. */
2210 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2214 spin_lock(&log->l_icloglock);
2215 xlog_state_want_sync(log, iclog);
2216 spin_unlock(&log->l_icloglock);
2219 return xlog_state_release_iclog(log, iclog);
2220 ASSERT(flags & XLOG_COMMIT_TRANS);
2221 *commit_iclog = iclog;
2228 * Write some region out to in-core log
2230 * This will be called when writing externally provided regions or when
2231 * writing out a commit record for a given transaction.
2233 * General algorithm:
2234 * 1. Find total length of this write. This may include adding to the
2235 * lengths passed in.
2236 * 2. Check whether we violate the tickets reservation.
2237 * 3. While writing to this iclog
2238 * A. Reserve as much space in this iclog as can get
2239 * B. If this is first write, save away start lsn
2240 * C. While writing this region:
2241 * 1. If first write of transaction, write start record
2242 * 2. Write log operation header (header per region)
2243 * 3. Find out if we can fit entire region into this iclog
2244 * 4. Potentially, verify destination memcpy ptr
2245 * 5. Memcpy (partial) region
2246 * 6. If partial copy, release iclog; otherwise, continue
2247 * copying more regions into current iclog
2248 * 4. Mark want sync bit (in simulation mode)
2249 * 5. Release iclog for potential flush to on-disk log.
2252 * 1. Panic if reservation is overrun. This should never happen since
2253 * reservation amounts are generated internal to the filesystem.
2255 * 1. Tickets are single threaded data structures.
2256 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2257 * syncing routine. When a single log_write region needs to span
2258 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2259 * on all log operation writes which don't contain the end of the
2260 * region. The XLOG_END_TRANS bit is used for the in-core log
2261 * operation which contains the end of the continued log_write region.
2262 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2263 * we don't really know exactly how much space will be used. As a result,
2264 * we don't update ic_offset until the end when we know exactly how many
2265 * bytes have been written out.
2270 struct xfs_log_vec *log_vector,
2271 struct xlog_ticket *ticket,
2272 xfs_lsn_t *start_lsn,
2273 struct xlog_in_core **commit_iclog,
2276 struct xlog_in_core *iclog = NULL;
2277 struct xfs_log_iovec *vecp;
2278 struct xfs_log_vec *lv;
2281 int partial_copy = 0;
2282 int partial_copy_len = 0;
2290 len = xlog_write_calc_vec_length(ticket, log_vector);
2293 * Region headers and bytes are already accounted for.
2294 * We only need to take into account start records and
2295 * split regions in this function.
2297 if (ticket->t_flags & XLOG_TIC_INITED)
2298 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2301 * Commit record headers need to be accounted for. These
2302 * come in as separate writes so are easy to detect.
2304 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2305 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2307 if (ticket->t_curr_res < 0)
2308 xlog_print_tic_res(log->l_mp, ticket);
2312 vecp = lv->lv_iovecp;
2313 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2317 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2318 &contwr, &log_offset);
2322 ASSERT(log_offset <= iclog->ic_size - 1);
2323 ptr = iclog->ic_datap + log_offset;
2325 /* start_lsn is the first lsn written to. That's all we need. */
2327 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2330 * This loop writes out as many regions as can fit in the amount
2331 * of space which was allocated by xlog_state_get_iclog_space().
2333 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2334 struct xfs_log_iovec *reg;
2335 struct xlog_op_header *ophdr;
2339 bool ordered = false;
2341 /* ordered log vectors have no regions to write */
2342 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2343 ASSERT(lv->lv_niovecs == 0);
2349 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2350 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2352 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2353 if (start_rec_copy) {
2355 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2359 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2361 return XFS_ERROR(EIO);
2363 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2364 sizeof(struct xlog_op_header));
2366 len += xlog_write_setup_copy(ticket, ophdr,
2367 iclog->ic_size-log_offset,
2369 ©_off, ©_len,
2372 xlog_verify_dest_ptr(log, ptr);
2375 ASSERT(copy_len >= 0);
2376 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2377 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2379 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2381 data_cnt += contwr ? copy_len : 0;
2383 error = xlog_write_copy_finish(log, iclog, flags,
2384 &record_cnt, &data_cnt,
2393 * if we had a partial copy, we need to get more iclog
2394 * space but we don't want to increment the region
2395 * index because there is still more is this region to
2398 * If we completed writing this region, and we flushed
2399 * the iclog (indicated by resetting of the record
2400 * count), then we also need to get more log space. If
2401 * this was the last record, though, we are done and
2407 if (++index == lv->lv_niovecs) {
2412 vecp = lv->lv_iovecp;
2414 if (record_cnt == 0 && ordered == false) {
2424 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2426 return xlog_state_release_iclog(log, iclog);
2428 ASSERT(flags & XLOG_COMMIT_TRANS);
2429 *commit_iclog = iclog;
2434 /*****************************************************************************
2436 * State Machine functions
2438 *****************************************************************************
2441 /* Clean iclogs starting from the head. This ordering must be
2442 * maintained, so an iclog doesn't become ACTIVE beyond one that
2443 * is SYNCING. This is also required to maintain the notion that we use
2444 * a ordered wait queue to hold off would be writers to the log when every
2445 * iclog is trying to sync to disk.
2447 * State Change: DIRTY -> ACTIVE
2450 xlog_state_clean_log(
2453 xlog_in_core_t *iclog;
2456 iclog = log->l_iclog;
2458 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2459 iclog->ic_state = XLOG_STATE_ACTIVE;
2460 iclog->ic_offset = 0;
2461 ASSERT(iclog->ic_callback == NULL);
2463 * If the number of ops in this iclog indicate it just
2464 * contains the dummy transaction, we can
2465 * change state into IDLE (the second time around).
2466 * Otherwise we should change the state into
2468 * We don't need to cover the dummy.
2471 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2476 * We have two dirty iclogs so start over
2477 * This could also be num of ops indicates
2478 * this is not the dummy going out.
2482 iclog->ic_header.h_num_logops = 0;
2483 memset(iclog->ic_header.h_cycle_data, 0,
2484 sizeof(iclog->ic_header.h_cycle_data));
2485 iclog->ic_header.h_lsn = 0;
2486 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2489 break; /* stop cleaning */
2490 iclog = iclog->ic_next;
2491 } while (iclog != log->l_iclog);
2493 /* log is locked when we are called */
2495 * Change state for the dummy log recording.
2496 * We usually go to NEED. But we go to NEED2 if the changed indicates
2497 * we are done writing the dummy record.
2498 * If we are done with the second dummy recored (DONE2), then
2502 switch (log->l_covered_state) {
2503 case XLOG_STATE_COVER_IDLE:
2504 case XLOG_STATE_COVER_NEED:
2505 case XLOG_STATE_COVER_NEED2:
2506 log->l_covered_state = XLOG_STATE_COVER_NEED;
2509 case XLOG_STATE_COVER_DONE:
2511 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2513 log->l_covered_state = XLOG_STATE_COVER_NEED;
2516 case XLOG_STATE_COVER_DONE2:
2518 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2520 log->l_covered_state = XLOG_STATE_COVER_NEED;
2527 } /* xlog_state_clean_log */
2530 xlog_get_lowest_lsn(
2533 xlog_in_core_t *lsn_log;
2534 xfs_lsn_t lowest_lsn, lsn;
2536 lsn_log = log->l_iclog;
2539 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2540 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2541 if ((lsn && !lowest_lsn) ||
2542 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2546 lsn_log = lsn_log->ic_next;
2547 } while (lsn_log != log->l_iclog);
2553 xlog_state_do_callback(
2556 struct xlog_in_core *ciclog)
2558 xlog_in_core_t *iclog;
2559 xlog_in_core_t *first_iclog; /* used to know when we've
2560 * processed all iclogs once */
2561 xfs_log_callback_t *cb, *cb_next;
2563 xfs_lsn_t lowest_lsn;
2564 int ioerrors; /* counter: iclogs with errors */
2565 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2566 int funcdidcallbacks; /* flag: function did callbacks */
2567 int repeats; /* for issuing console warnings if
2568 * looping too many times */
2571 spin_lock(&log->l_icloglock);
2572 first_iclog = iclog = log->l_iclog;
2574 funcdidcallbacks = 0;
2579 * Scan all iclogs starting with the one pointed to by the
2580 * log. Reset this starting point each time the log is
2581 * unlocked (during callbacks).
2583 * Keep looping through iclogs until one full pass is made
2584 * without running any callbacks.
2586 first_iclog = log->l_iclog;
2587 iclog = log->l_iclog;
2588 loopdidcallbacks = 0;
2593 /* skip all iclogs in the ACTIVE & DIRTY states */
2594 if (iclog->ic_state &
2595 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2596 iclog = iclog->ic_next;
2601 * Between marking a filesystem SHUTDOWN and stopping
2602 * the log, we do flush all iclogs to disk (if there
2603 * wasn't a log I/O error). So, we do want things to
2604 * go smoothly in case of just a SHUTDOWN w/o a
2607 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2609 * Can only perform callbacks in order. Since
2610 * this iclog is not in the DONE_SYNC/
2611 * DO_CALLBACK state, we skip the rest and
2612 * just try to clean up. If we set our iclog
2613 * to DO_CALLBACK, we will not process it when
2614 * we retry since a previous iclog is in the
2615 * CALLBACK and the state cannot change since
2616 * we are holding the l_icloglock.
2618 if (!(iclog->ic_state &
2619 (XLOG_STATE_DONE_SYNC |
2620 XLOG_STATE_DO_CALLBACK))) {
2621 if (ciclog && (ciclog->ic_state ==
2622 XLOG_STATE_DONE_SYNC)) {
2623 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2628 * We now have an iclog that is in either the
2629 * DO_CALLBACK or DONE_SYNC states. The other
2630 * states (WANT_SYNC, SYNCING, or CALLBACK were
2631 * caught by the above if and are going to
2632 * clean (i.e. we aren't doing their callbacks)
2637 * We will do one more check here to see if we
2638 * have chased our tail around.
2641 lowest_lsn = xlog_get_lowest_lsn(log);
2643 XFS_LSN_CMP(lowest_lsn,
2644 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2645 iclog = iclog->ic_next;
2646 continue; /* Leave this iclog for
2650 iclog->ic_state = XLOG_STATE_CALLBACK;
2654 * Completion of a iclog IO does not imply that
2655 * a transaction has completed, as transactions
2656 * can be large enough to span many iclogs. We
2657 * cannot change the tail of the log half way
2658 * through a transaction as this may be the only
2659 * transaction in the log and moving th etail to
2660 * point to the middle of it will prevent
2661 * recovery from finding the start of the
2662 * transaction. Hence we should only update the
2663 * last_sync_lsn if this iclog contains
2664 * transaction completion callbacks on it.
2666 * We have to do this before we drop the
2667 * icloglock to ensure we are the only one that
2670 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2671 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2672 if (iclog->ic_callback)
2673 atomic64_set(&log->l_last_sync_lsn,
2674 be64_to_cpu(iclog->ic_header.h_lsn));
2679 spin_unlock(&log->l_icloglock);
2682 * Keep processing entries in the callback list until
2683 * we come around and it is empty. We need to
2684 * atomically see that the list is empty and change the
2685 * state to DIRTY so that we don't miss any more
2686 * callbacks being added.
2688 spin_lock(&iclog->ic_callback_lock);
2689 cb = iclog->ic_callback;
2691 iclog->ic_callback_tail = &(iclog->ic_callback);
2692 iclog->ic_callback = NULL;
2693 spin_unlock(&iclog->ic_callback_lock);
2695 /* perform callbacks in the order given */
2696 for (; cb; cb = cb_next) {
2697 cb_next = cb->cb_next;
2698 cb->cb_func(cb->cb_arg, aborted);
2700 spin_lock(&iclog->ic_callback_lock);
2701 cb = iclog->ic_callback;
2707 spin_lock(&log->l_icloglock);
2708 ASSERT(iclog->ic_callback == NULL);
2709 spin_unlock(&iclog->ic_callback_lock);
2710 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2711 iclog->ic_state = XLOG_STATE_DIRTY;
2714 * Transition from DIRTY to ACTIVE if applicable.
2715 * NOP if STATE_IOERROR.
2717 xlog_state_clean_log(log);
2719 /* wake up threads waiting in xfs_log_force() */
2720 wake_up_all(&iclog->ic_force_wait);
2722 iclog = iclog->ic_next;
2723 } while (first_iclog != iclog);
2725 if (repeats > 5000) {
2726 flushcnt += repeats;
2729 "%s: possible infinite loop (%d iterations)",
2730 __func__, flushcnt);
2732 } while (!ioerrors && loopdidcallbacks);
2735 * make one last gasp attempt to see if iclogs are being left in
2739 if (funcdidcallbacks) {
2740 first_iclog = iclog = log->l_iclog;
2742 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2744 * Terminate the loop if iclogs are found in states
2745 * which will cause other threads to clean up iclogs.
2747 * SYNCING - i/o completion will go through logs
2748 * DONE_SYNC - interrupt thread should be waiting for
2750 * IOERROR - give up hope all ye who enter here
2752 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2753 iclog->ic_state == XLOG_STATE_SYNCING ||
2754 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2755 iclog->ic_state == XLOG_STATE_IOERROR )
2757 iclog = iclog->ic_next;
2758 } while (first_iclog != iclog);
2762 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2764 spin_unlock(&log->l_icloglock);
2767 wake_up_all(&log->l_flush_wait);
2772 * Finish transitioning this iclog to the dirty state.
2774 * Make sure that we completely execute this routine only when this is
2775 * the last call to the iclog. There is a good chance that iclog flushes,
2776 * when we reach the end of the physical log, get turned into 2 separate
2777 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2778 * routine. By using the reference count bwritecnt, we guarantee that only
2779 * the second completion goes through.
2781 * Callbacks could take time, so they are done outside the scope of the
2782 * global state machine log lock.
2785 xlog_state_done_syncing(
2786 xlog_in_core_t *iclog,
2789 struct xlog *log = iclog->ic_log;
2791 spin_lock(&log->l_icloglock);
2793 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2794 iclog->ic_state == XLOG_STATE_IOERROR);
2795 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2796 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2800 * If we got an error, either on the first buffer, or in the case of
2801 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2802 * and none should ever be attempted to be written to disk
2805 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2806 if (--iclog->ic_bwritecnt == 1) {
2807 spin_unlock(&log->l_icloglock);
2810 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2814 * Someone could be sleeping prior to writing out the next
2815 * iclog buffer, we wake them all, one will get to do the
2816 * I/O, the others get to wait for the result.
2818 wake_up_all(&iclog->ic_write_wait);
2819 spin_unlock(&log->l_icloglock);
2820 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2821 } /* xlog_state_done_syncing */
2825 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2826 * sleep. We wait on the flush queue on the head iclog as that should be
2827 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2828 * we will wait here and all new writes will sleep until a sync completes.
2830 * The in-core logs are used in a circular fashion. They are not used
2831 * out-of-order even when an iclog past the head is free.
2834 * * log_offset where xlog_write() can start writing into the in-core
2836 * * in-core log pointer to which xlog_write() should write.
2837 * * boolean indicating this is a continued write to an in-core log.
2838 * If this is the last write, then the in-core log's offset field
2839 * needs to be incremented, depending on the amount of data which
2843 xlog_state_get_iclog_space(
2846 struct xlog_in_core **iclogp,
2847 struct xlog_ticket *ticket,
2848 int *continued_write,
2852 xlog_rec_header_t *head;
2853 xlog_in_core_t *iclog;
2857 spin_lock(&log->l_icloglock);
2858 if (XLOG_FORCED_SHUTDOWN(log)) {
2859 spin_unlock(&log->l_icloglock);
2860 return XFS_ERROR(EIO);
2863 iclog = log->l_iclog;
2864 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2865 XFS_STATS_INC(xs_log_noiclogs);
2867 /* Wait for log writes to have flushed */
2868 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2872 head = &iclog->ic_header;
2874 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2875 log_offset = iclog->ic_offset;
2877 /* On the 1st write to an iclog, figure out lsn. This works
2878 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2879 * committing to. If the offset is set, that's how many blocks
2882 if (log_offset == 0) {
2883 ticket->t_curr_res -= log->l_iclog_hsize;
2884 xlog_tic_add_region(ticket,
2886 XLOG_REG_TYPE_LRHEADER);
2887 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2888 head->h_lsn = cpu_to_be64(
2889 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2890 ASSERT(log->l_curr_block >= 0);
2893 /* If there is enough room to write everything, then do it. Otherwise,
2894 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2895 * bit is on, so this will get flushed out. Don't update ic_offset
2896 * until you know exactly how many bytes get copied. Therefore, wait
2897 * until later to update ic_offset.
2899 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2900 * can fit into remaining data section.
2902 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2903 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2906 * If I'm the only one writing to this iclog, sync it to disk.
2907 * We need to do an atomic compare and decrement here to avoid
2908 * racing with concurrent atomic_dec_and_lock() calls in
2909 * xlog_state_release_iclog() when there is more than one
2910 * reference to the iclog.
2912 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2913 /* we are the only one */
2914 spin_unlock(&log->l_icloglock);
2915 error = xlog_state_release_iclog(log, iclog);
2919 spin_unlock(&log->l_icloglock);
2924 /* Do we have enough room to write the full amount in the remainder
2925 * of this iclog? Or must we continue a write on the next iclog and
2926 * mark this iclog as completely taken? In the case where we switch
2927 * iclogs (to mark it taken), this particular iclog will release/sync
2928 * to disk in xlog_write().
2930 if (len <= iclog->ic_size - iclog->ic_offset) {
2931 *continued_write = 0;
2932 iclog->ic_offset += len;
2934 *continued_write = 1;
2935 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2939 ASSERT(iclog->ic_offset <= iclog->ic_size);
2940 spin_unlock(&log->l_icloglock);
2942 *logoffsetp = log_offset;
2944 } /* xlog_state_get_iclog_space */
2946 /* The first cnt-1 times through here we don't need to
2947 * move the grant write head because the permanent
2948 * reservation has reserved cnt times the unit amount.
2949 * Release part of current permanent unit reservation and
2950 * reset current reservation to be one units worth. Also
2951 * move grant reservation head forward.
2954 xlog_regrant_reserve_log_space(
2956 struct xlog_ticket *ticket)
2958 trace_xfs_log_regrant_reserve_enter(log, ticket);
2960 if (ticket->t_cnt > 0)
2963 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2964 ticket->t_curr_res);
2965 xlog_grant_sub_space(log, &log->l_write_head.grant,
2966 ticket->t_curr_res);
2967 ticket->t_curr_res = ticket->t_unit_res;
2968 xlog_tic_reset_res(ticket);
2970 trace_xfs_log_regrant_reserve_sub(log, ticket);
2972 /* just return if we still have some of the pre-reserved space */
2973 if (ticket->t_cnt > 0)
2976 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2977 ticket->t_unit_res);
2979 trace_xfs_log_regrant_reserve_exit(log, ticket);
2981 ticket->t_curr_res = ticket->t_unit_res;
2982 xlog_tic_reset_res(ticket);
2983 } /* xlog_regrant_reserve_log_space */
2987 * Give back the space left from a reservation.
2989 * All the information we need to make a correct determination of space left
2990 * is present. For non-permanent reservations, things are quite easy. The
2991 * count should have been decremented to zero. We only need to deal with the
2992 * space remaining in the current reservation part of the ticket. If the
2993 * ticket contains a permanent reservation, there may be left over space which
2994 * needs to be released. A count of N means that N-1 refills of the current
2995 * reservation can be done before we need to ask for more space. The first
2996 * one goes to fill up the first current reservation. Once we run out of
2997 * space, the count will stay at zero and the only space remaining will be
2998 * in the current reservation field.
3001 xlog_ungrant_log_space(
3003 struct xlog_ticket *ticket)
3007 if (ticket->t_cnt > 0)
3010 trace_xfs_log_ungrant_enter(log, ticket);
3011 trace_xfs_log_ungrant_sub(log, ticket);
3014 * If this is a permanent reservation ticket, we may be able to free
3015 * up more space based on the remaining count.
3017 bytes = ticket->t_curr_res;
3018 if (ticket->t_cnt > 0) {
3019 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3020 bytes += ticket->t_unit_res*ticket->t_cnt;
3023 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3024 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3026 trace_xfs_log_ungrant_exit(log, ticket);
3028 xfs_log_space_wake(log->l_mp);
3032 * Flush iclog to disk if this is the last reference to the given iclog and
3033 * the WANT_SYNC bit is set.
3035 * When this function is entered, the iclog is not necessarily in the
3036 * WANT_SYNC state. It may be sitting around waiting to get filled.
3041 xlog_state_release_iclog(
3043 struct xlog_in_core *iclog)
3045 int sync = 0; /* do we sync? */
3047 if (iclog->ic_state & XLOG_STATE_IOERROR)
3048 return XFS_ERROR(EIO);
3050 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3051 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3054 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3055 spin_unlock(&log->l_icloglock);
3056 return XFS_ERROR(EIO);
3058 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3059 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3061 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3062 /* update tail before writing to iclog */
3063 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3065 iclog->ic_state = XLOG_STATE_SYNCING;
3066 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3067 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3068 /* cycle incremented when incrementing curr_block */
3070 spin_unlock(&log->l_icloglock);
3073 * We let the log lock go, so it's possible that we hit a log I/O
3074 * error or some other SHUTDOWN condition that marks the iclog
3075 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3076 * this iclog has consistent data, so we ignore IOERROR
3077 * flags after this point.
3080 return xlog_sync(log, iclog);
3082 } /* xlog_state_release_iclog */
3086 * This routine will mark the current iclog in the ring as WANT_SYNC
3087 * and move the current iclog pointer to the next iclog in the ring.
3088 * When this routine is called from xlog_state_get_iclog_space(), the
3089 * exact size of the iclog has not yet been determined. All we know is
3090 * that every data block. We have run out of space in this log record.
3093 xlog_state_switch_iclogs(
3095 struct xlog_in_core *iclog,
3098 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3100 eventual_size = iclog->ic_offset;
3101 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3102 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3103 log->l_prev_block = log->l_curr_block;
3104 log->l_prev_cycle = log->l_curr_cycle;
3106 /* roll log?: ic_offset changed later */
3107 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3109 /* Round up to next log-sunit */
3110 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3111 log->l_mp->m_sb.sb_logsunit > 1) {
3112 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3113 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3116 if (log->l_curr_block >= log->l_logBBsize) {
3117 log->l_curr_cycle++;
3118 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3119 log->l_curr_cycle++;
3120 log->l_curr_block -= log->l_logBBsize;
3121 ASSERT(log->l_curr_block >= 0);
3123 ASSERT(iclog == log->l_iclog);
3124 log->l_iclog = iclog->ic_next;
3125 } /* xlog_state_switch_iclogs */
3128 * Write out all data in the in-core log as of this exact moment in time.
3130 * Data may be written to the in-core log during this call. However,
3131 * we don't guarantee this data will be written out. A change from past
3132 * implementation means this routine will *not* write out zero length LRs.
3134 * Basically, we try and perform an intelligent scan of the in-core logs.
3135 * If we determine there is no flushable data, we just return. There is no
3136 * flushable data if:
3138 * 1. the current iclog is active and has no data; the previous iclog
3139 * is in the active or dirty state.
3140 * 2. the current iclog is drity, and the previous iclog is in the
3141 * active or dirty state.
3145 * 1. the current iclog is not in the active nor dirty state.
3146 * 2. the current iclog dirty, and the previous iclog is not in the
3147 * active nor dirty state.
3148 * 3. the current iclog is active, and there is another thread writing
3149 * to this particular iclog.
3150 * 4. a) the current iclog is active and has no other writers
3151 * b) when we return from flushing out this iclog, it is still
3152 * not in the active nor dirty state.
3156 struct xfs_mount *mp,
3160 struct xlog *log = mp->m_log;
3161 struct xlog_in_core *iclog;
3164 XFS_STATS_INC(xs_log_force);
3166 xlog_cil_force(log);
3168 spin_lock(&log->l_icloglock);
3170 iclog = log->l_iclog;
3171 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3172 spin_unlock(&log->l_icloglock);
3173 return XFS_ERROR(EIO);
3176 /* If the head iclog is not active nor dirty, we just attach
3177 * ourselves to the head and go to sleep.
3179 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3180 iclog->ic_state == XLOG_STATE_DIRTY) {
3182 * If the head is dirty or (active and empty), then
3183 * we need to look at the previous iclog. If the previous
3184 * iclog is active or dirty we are done. There is nothing
3185 * to sync out. Otherwise, we attach ourselves to the
3186 * previous iclog and go to sleep.
3188 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3189 (atomic_read(&iclog->ic_refcnt) == 0
3190 && iclog->ic_offset == 0)) {
3191 iclog = iclog->ic_prev;
3192 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3193 iclog->ic_state == XLOG_STATE_DIRTY)
3198 if (atomic_read(&iclog->ic_refcnt) == 0) {
3199 /* We are the only one with access to this
3200 * iclog. Flush it out now. There should
3201 * be a roundoff of zero to show that someone
3202 * has already taken care of the roundoff from
3203 * the previous sync.
3205 atomic_inc(&iclog->ic_refcnt);
3206 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3207 xlog_state_switch_iclogs(log, iclog, 0);
3208 spin_unlock(&log->l_icloglock);
3210 if (xlog_state_release_iclog(log, iclog))
3211 return XFS_ERROR(EIO);
3215 spin_lock(&log->l_icloglock);
3216 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3217 iclog->ic_state != XLOG_STATE_DIRTY)
3222 /* Someone else is writing to this iclog.
3223 * Use its call to flush out the data. However,
3224 * the other thread may not force out this LR,
3225 * so we mark it WANT_SYNC.
3227 xlog_state_switch_iclogs(log, iclog, 0);
3233 /* By the time we come around again, the iclog could've been filled
3234 * which would give it another lsn. If we have a new lsn, just
3235 * return because the relevant data has been flushed.
3238 if (flags & XFS_LOG_SYNC) {
3240 * We must check if we're shutting down here, before
3241 * we wait, while we're holding the l_icloglock.
3242 * Then we check again after waking up, in case our
3243 * sleep was disturbed by a bad news.
3245 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3246 spin_unlock(&log->l_icloglock);
3247 return XFS_ERROR(EIO);
3249 XFS_STATS_INC(xs_log_force_sleep);
3250 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3252 * No need to grab the log lock here since we're
3253 * only deciding whether or not to return EIO
3254 * and the memory read should be atomic.
3256 if (iclog->ic_state & XLOG_STATE_IOERROR)
3257 return XFS_ERROR(EIO);
3263 spin_unlock(&log->l_icloglock);
3269 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3270 * about errors or whether the log was flushed or not. This is the normal
3271 * interface to use when trying to unpin items or move the log forward.
3280 trace_xfs_log_force(mp, 0);
3281 error = _xfs_log_force(mp, flags, NULL);
3283 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3287 * Force the in-core log to disk for a specific LSN.
3289 * Find in-core log with lsn.
3290 * If it is in the DIRTY state, just return.
3291 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3292 * state and go to sleep or return.
3293 * If it is in any other state, go to sleep or return.
3295 * Synchronous forces are implemented with a signal variable. All callers
3296 * to force a given lsn to disk will wait on a the sv attached to the
3297 * specific in-core log. When given in-core log finally completes its
3298 * write to disk, that thread will wake up all threads waiting on the
3303 struct xfs_mount *mp,
3308 struct xlog *log = mp->m_log;
3309 struct xlog_in_core *iclog;
3310 int already_slept = 0;
3314 XFS_STATS_INC(xs_log_force);
3316 lsn = xlog_cil_force_lsn(log, lsn);
3317 if (lsn == NULLCOMMITLSN)
3321 spin_lock(&log->l_icloglock);
3322 iclog = log->l_iclog;
3323 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3324 spin_unlock(&log->l_icloglock);
3325 return XFS_ERROR(EIO);
3329 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3330 iclog = iclog->ic_next;
3334 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3335 spin_unlock(&log->l_icloglock);
3339 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3341 * We sleep here if we haven't already slept (e.g.
3342 * this is the first time we've looked at the correct
3343 * iclog buf) and the buffer before us is going to
3344 * be sync'ed. The reason for this is that if we
3345 * are doing sync transactions here, by waiting for
3346 * the previous I/O to complete, we can allow a few
3347 * more transactions into this iclog before we close
3350 * Otherwise, we mark the buffer WANT_SYNC, and bump
3351 * up the refcnt so we can release the log (which
3352 * drops the ref count). The state switch keeps new
3353 * transaction commits from using this buffer. When
3354 * the current commits finish writing into the buffer,
3355 * the refcount will drop to zero and the buffer will
3358 if (!already_slept &&
3359 (iclog->ic_prev->ic_state &
3360 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3361 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3363 XFS_STATS_INC(xs_log_force_sleep);
3365 xlog_wait(&iclog->ic_prev->ic_write_wait,
3372 atomic_inc(&iclog->ic_refcnt);
3373 xlog_state_switch_iclogs(log, iclog, 0);
3374 spin_unlock(&log->l_icloglock);
3375 if (xlog_state_release_iclog(log, iclog))
3376 return XFS_ERROR(EIO);
3379 spin_lock(&log->l_icloglock);
3382 if ((flags & XFS_LOG_SYNC) && /* sleep */
3384 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3386 * Don't wait on completion if we know that we've
3387 * gotten a log write error.
3389 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3390 spin_unlock(&log->l_icloglock);
3391 return XFS_ERROR(EIO);
3393 XFS_STATS_INC(xs_log_force_sleep);
3394 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3396 * No need to grab the log lock here since we're
3397 * only deciding whether or not to return EIO
3398 * and the memory read should be atomic.
3400 if (iclog->ic_state & XLOG_STATE_IOERROR)
3401 return XFS_ERROR(EIO);
3405 } else { /* just return */
3406 spin_unlock(&log->l_icloglock);
3410 } while (iclog != log->l_iclog);
3412 spin_unlock(&log->l_icloglock);
3417 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3418 * about errors or whether the log was flushed or not. This is the normal
3419 * interface to use when trying to unpin items or move the log forward.
3429 trace_xfs_log_force(mp, lsn);
3430 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3432 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3436 * Called when we want to mark the current iclog as being ready to sync to
3440 xlog_state_want_sync(
3442 struct xlog_in_core *iclog)
3444 assert_spin_locked(&log->l_icloglock);
3446 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3447 xlog_state_switch_iclogs(log, iclog, 0);
3449 ASSERT(iclog->ic_state &
3450 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3455 /*****************************************************************************
3459 *****************************************************************************
3463 * Free a used ticket when its refcount falls to zero.
3467 xlog_ticket_t *ticket)
3469 ASSERT(atomic_read(&ticket->t_ref) > 0);
3470 if (atomic_dec_and_test(&ticket->t_ref))
3471 kmem_zone_free(xfs_log_ticket_zone, ticket);
3476 xlog_ticket_t *ticket)
3478 ASSERT(atomic_read(&ticket->t_ref) > 0);
3479 atomic_inc(&ticket->t_ref);
3484 * Figure out the total log space unit (in bytes) that would be
3485 * required for a log ticket.
3488 xfs_log_calc_unit_res(
3489 struct xfs_mount *mp,
3492 struct xlog *log = mp->m_log;
3497 * Permanent reservations have up to 'cnt'-1 active log operations
3498 * in the log. A unit in this case is the amount of space for one
3499 * of these log operations. Normal reservations have a cnt of 1
3500 * and their unit amount is the total amount of space required.
3502 * The following lines of code account for non-transaction data
3503 * which occupy space in the on-disk log.
3505 * Normal form of a transaction is:
3506 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3507 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3509 * We need to account for all the leadup data and trailer data
3510 * around the transaction data.
3511 * And then we need to account for the worst case in terms of using
3513 * The worst case will happen if:
3514 * - the placement of the transaction happens to be such that the
3515 * roundoff is at its maximum
3516 * - the transaction data is synced before the commit record is synced
3517 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3518 * Therefore the commit record is in its own Log Record.
3519 * This can happen as the commit record is called with its
3520 * own region to xlog_write().
3521 * This then means that in the worst case, roundoff can happen for
3522 * the commit-rec as well.
3523 * The commit-rec is smaller than padding in this scenario and so it is
3524 * not added separately.
3527 /* for trans header */
3528 unit_bytes += sizeof(xlog_op_header_t);
3529 unit_bytes += sizeof(xfs_trans_header_t);
3532 unit_bytes += sizeof(xlog_op_header_t);
3535 * for LR headers - the space for data in an iclog is the size minus
3536 * the space used for the headers. If we use the iclog size, then we
3537 * undercalculate the number of headers required.
3539 * Furthermore - the addition of op headers for split-recs might
3540 * increase the space required enough to require more log and op
3541 * headers, so take that into account too.
3543 * IMPORTANT: This reservation makes the assumption that if this
3544 * transaction is the first in an iclog and hence has the LR headers
3545 * accounted to it, then the remaining space in the iclog is
3546 * exclusively for this transaction. i.e. if the transaction is larger
3547 * than the iclog, it will be the only thing in that iclog.
3548 * Fundamentally, this means we must pass the entire log vector to
3549 * xlog_write to guarantee this.
3551 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3552 num_headers = howmany(unit_bytes, iclog_space);
3554 /* for split-recs - ophdrs added when data split over LRs */
3555 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3557 /* add extra header reservations if we overrun */
3558 while (!num_headers ||
3559 howmany(unit_bytes, iclog_space) > num_headers) {
3560 unit_bytes += sizeof(xlog_op_header_t);
3563 unit_bytes += log->l_iclog_hsize * num_headers;
3565 /* for commit-rec LR header - note: padding will subsume the ophdr */
3566 unit_bytes += log->l_iclog_hsize;
3568 /* for roundoff padding for transaction data and one for commit record */
3569 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3570 /* log su roundoff */
3571 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3574 unit_bytes += 2 * BBSIZE;
3581 * Allocate and initialise a new log ticket.
3583 struct xlog_ticket *
3590 xfs_km_flags_t alloc_flags)
3592 struct xlog_ticket *tic;
3595 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3599 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3601 atomic_set(&tic->t_ref, 1);
3602 tic->t_task = current;
3603 INIT_LIST_HEAD(&tic->t_queue);
3604 tic->t_unit_res = unit_res;
3605 tic->t_curr_res = unit_res;
3608 tic->t_tid = prandom_u32();
3609 tic->t_clientid = client;
3610 tic->t_flags = XLOG_TIC_INITED;
3611 tic->t_trans_type = 0;
3613 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3615 xlog_tic_reset_res(tic);
3621 /******************************************************************************
3623 * Log debug routines
3625 ******************************************************************************
3629 * Make sure that the destination ptr is within the valid data region of
3630 * one of the iclogs. This uses backup pointers stored in a different
3631 * part of the log in case we trash the log structure.
3634 xlog_verify_dest_ptr(
3641 for (i = 0; i < log->l_iclog_bufs; i++) {
3642 if (ptr >= log->l_iclog_bak[i] &&
3643 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3648 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3652 * Check to make sure the grant write head didn't just over lap the tail. If
3653 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3654 * the cycles differ by exactly one and check the byte count.
3656 * This check is run unlocked, so can give false positives. Rather than assert
3657 * on failures, use a warn-once flag and a panic tag to allow the admin to
3658 * determine if they want to panic the machine when such an error occurs. For
3659 * debug kernels this will have the same effect as using an assert but, unlinke
3660 * an assert, it can be turned off at runtime.
3663 xlog_verify_grant_tail(
3666 int tail_cycle, tail_blocks;
3669 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3670 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3671 if (tail_cycle != cycle) {
3672 if (cycle - 1 != tail_cycle &&
3673 !(log->l_flags & XLOG_TAIL_WARN)) {
3674 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3675 "%s: cycle - 1 != tail_cycle", __func__);
3676 log->l_flags |= XLOG_TAIL_WARN;
3679 if (space > BBTOB(tail_blocks) &&
3680 !(log->l_flags & XLOG_TAIL_WARN)) {
3681 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3682 "%s: space > BBTOB(tail_blocks)", __func__);
3683 log->l_flags |= XLOG_TAIL_WARN;
3688 /* check if it will fit */
3690 xlog_verify_tail_lsn(
3692 struct xlog_in_core *iclog,
3697 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3699 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3700 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3701 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3703 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3705 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3706 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3708 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3709 if (blocks < BTOBB(iclog->ic_offset) + 1)
3710 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3712 } /* xlog_verify_tail_lsn */
3715 * Perform a number of checks on the iclog before writing to disk.
3717 * 1. Make sure the iclogs are still circular
3718 * 2. Make sure we have a good magic number
3719 * 3. Make sure we don't have magic numbers in the data
3720 * 4. Check fields of each log operation header for:
3721 * A. Valid client identifier
3722 * B. tid ptr value falls in valid ptr space (user space code)
3723 * C. Length in log record header is correct according to the
3724 * individual operation headers within record.
3725 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3726 * log, check the preceding blocks of the physical log to make sure all
3727 * the cycle numbers agree with the current cycle number.
3732 struct xlog_in_core *iclog,
3736 xlog_op_header_t *ophead;
3737 xlog_in_core_t *icptr;
3738 xlog_in_core_2_t *xhdr;
3740 xfs_caddr_t base_ptr;
3741 __psint_t field_offset;
3743 int len, i, j, k, op_len;
3746 /* check validity of iclog pointers */
3747 spin_lock(&log->l_icloglock);
3748 icptr = log->l_iclog;
3749 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3752 if (icptr != log->l_iclog)
3753 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3754 spin_unlock(&log->l_icloglock);
3756 /* check log magic numbers */
3757 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3758 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3760 ptr = (xfs_caddr_t) &iclog->ic_header;
3761 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3763 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3764 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3769 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3770 ptr = iclog->ic_datap;
3772 ophead = (xlog_op_header_t *)ptr;
3773 xhdr = iclog->ic_data;
3774 for (i = 0; i < len; i++) {
3775 ophead = (xlog_op_header_t *)ptr;
3777 /* clientid is only 1 byte */
3778 field_offset = (__psint_t)
3779 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3780 if (!syncing || (field_offset & 0x1ff)) {
3781 clientid = ophead->oh_clientid;
3783 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3784 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3785 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3786 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3787 clientid = xlog_get_client_id(
3788 xhdr[j].hic_xheader.xh_cycle_data[k]);
3790 clientid = xlog_get_client_id(
3791 iclog->ic_header.h_cycle_data[idx]);
3794 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3796 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3797 __func__, clientid, ophead,
3798 (unsigned long)field_offset);
3801 field_offset = (__psint_t)
3802 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3803 if (!syncing || (field_offset & 0x1ff)) {
3804 op_len = be32_to_cpu(ophead->oh_len);
3806 idx = BTOBBT((__psint_t)&ophead->oh_len -
3807 (__psint_t)iclog->ic_datap);
3808 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3809 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3810 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3811 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3813 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3816 ptr += sizeof(xlog_op_header_t) + op_len;
3818 } /* xlog_verify_iclog */
3822 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3828 xlog_in_core_t *iclog, *ic;
3830 iclog = log->l_iclog;
3831 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3833 * Mark all the incore logs IOERROR.
3834 * From now on, no log flushes will result.
3838 ic->ic_state = XLOG_STATE_IOERROR;
3840 } while (ic != iclog);
3844 * Return non-zero, if state transition has already happened.
3850 * This is called from xfs_force_shutdown, when we're forcibly
3851 * shutting down the filesystem, typically because of an IO error.
3852 * Our main objectives here are to make sure that:
3853 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3854 * parties to find out, 'atomically'.
3855 * b. those who're sleeping on log reservations, pinned objects and
3856 * other resources get woken up, and be told the bad news.
3857 * c. nothing new gets queued up after (a) and (b) are done.
3858 * d. if !logerror, flush the iclogs to disk, then seal them off
3861 * Note: for delayed logging the !logerror case needs to flush the regions
3862 * held in memory out to the iclogs before flushing them to disk. This needs
3863 * to be done before the log is marked as shutdown, otherwise the flush to the
3867 xfs_log_force_umount(
3868 struct xfs_mount *mp,
3877 * If this happens during log recovery, don't worry about
3878 * locking; the log isn't open for business yet.
3881 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3882 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3884 XFS_BUF_DONE(mp->m_sb_bp);
3889 * Somebody could've already done the hard work for us.
3890 * No need to get locks for this.
3892 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3893 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3899 * Flush the in memory commit item list before marking the log as
3900 * being shut down. We need to do it in this order to ensure all the
3901 * completed transactions are flushed to disk with the xfs_log_force()
3905 xlog_cil_force(log);
3908 * mark the filesystem and the as in a shutdown state and wake
3909 * everybody up to tell them the bad news.
3911 spin_lock(&log->l_icloglock);
3912 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3914 XFS_BUF_DONE(mp->m_sb_bp);
3917 * This flag is sort of redundant because of the mount flag, but
3918 * it's good to maintain the separation between the log and the rest
3921 log->l_flags |= XLOG_IO_ERROR;
3924 * If we hit a log error, we want to mark all the iclogs IOERROR
3925 * while we're still holding the loglock.
3928 retval = xlog_state_ioerror(log);
3929 spin_unlock(&log->l_icloglock);
3932 * We don't want anybody waiting for log reservations after this. That
3933 * means we have to wake up everybody queued up on reserveq as well as
3934 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3935 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3936 * action is protected by the grant locks.
3938 xlog_grant_head_wake_all(&log->l_reserve_head);
3939 xlog_grant_head_wake_all(&log->l_write_head);
3941 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3944 * Force the incore logs to disk before shutting the
3945 * log down completely.
3947 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3949 spin_lock(&log->l_icloglock);
3950 retval = xlog_state_ioerror(log);
3951 spin_unlock(&log->l_icloglock);
3954 * Wake up everybody waiting on xfs_log_force.
3955 * Callback all log item committed functions as if the
3956 * log writes were completed.
3958 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3960 #ifdef XFSERRORDEBUG
3962 xlog_in_core_t *iclog;
3964 spin_lock(&log->l_icloglock);
3965 iclog = log->l_iclog;
3967 ASSERT(iclog->ic_callback == 0);
3968 iclog = iclog->ic_next;
3969 } while (iclog != log->l_iclog);
3970 spin_unlock(&log->l_icloglock);
3973 /* return non-zero if log IOERROR transition had already happened */
3981 xlog_in_core_t *iclog;
3983 iclog = log->l_iclog;
3985 /* endianness does not matter here, zero is zero in
3988 if (iclog->ic_header.h_num_logops)
3990 iclog = iclog->ic_next;
3991 } while (iclog != log->l_iclog);