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_types.h"
24 #include "xfs_trans.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_buf_item.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_log_recover.h"
35 #include "xfs_trans_priv.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
39 #include "xfs_trace.h"
41 kmem_zone_t *xfs_log_ticket_zone;
43 /* Local miscellaneous function prototypes */
44 STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45 xlog_in_core_t **, xfs_lsn_t *);
46 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
47 xfs_buftarg_t *log_target,
48 xfs_daddr_t blk_offset,
50 STATIC int xlog_space_left(struct log *log, atomic64_t *head);
51 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52 STATIC void xlog_dealloc_log(xlog_t *log);
54 /* local state machine functions */
55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57 STATIC int xlog_state_get_iclog_space(xlog_t *log,
59 xlog_in_core_t **iclog,
60 xlog_ticket_t *ticket,
63 STATIC int xlog_state_release_iclog(xlog_t *log,
64 xlog_in_core_t *iclog);
65 STATIC void xlog_state_switch_iclogs(xlog_t *log,
66 xlog_in_core_t *iclog,
68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
70 STATIC void xlog_grant_push_ail(struct log *log,
72 STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
73 xlog_ticket_t *ticket);
74 STATIC void xlog_ungrant_log_space(xlog_t *log,
75 xlog_ticket_t *ticket);
78 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
79 STATIC void xlog_verify_grant_tail(struct log *log);
80 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
81 int count, boolean_t syncing);
82 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
85 #define xlog_verify_dest_ptr(a,b)
86 #define xlog_verify_grant_tail(a)
87 #define xlog_verify_iclog(a,b,c,d)
88 #define xlog_verify_tail_lsn(a,b,c)
91 STATIC int xlog_iclogs_empty(xlog_t *log);
99 int64_t head_val = atomic64_read(head);
105 xlog_crack_grant_head_val(head_val, &cycle, &space);
109 space += log->l_logsize;
114 new = xlog_assign_grant_head_val(cycle, space);
115 head_val = atomic64_cmpxchg(head, old, new);
116 } while (head_val != old);
120 xlog_grant_add_space(
125 int64_t head_val = atomic64_read(head);
132 xlog_crack_grant_head_val(head_val, &cycle, &space);
134 tmp = log->l_logsize - space;
143 new = xlog_assign_grant_head_val(cycle, space);
144 head_val = atomic64_cmpxchg(head, old, new);
145 } while (head_val != old);
149 xlog_grant_head_init(
150 struct xlog_grant_head *head)
152 xlog_assign_grant_head(&head->grant, 1, 0);
153 INIT_LIST_HEAD(&head->waiters);
154 spin_lock_init(&head->lock);
158 xlog_grant_head_wake_all(
159 struct xlog_grant_head *head)
161 struct xlog_ticket *tic;
163 spin_lock(&head->lock);
164 list_for_each_entry(tic, &head->waiters, t_queue)
165 wake_up_process(tic->t_task);
166 spin_unlock(&head->lock);
170 xlog_ticket_reservation(
172 struct xlog_grant_head *head,
173 struct xlog_ticket *tic)
175 if (head == &log->l_write_head) {
176 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
177 return tic->t_unit_res;
179 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
180 return tic->t_unit_res * tic->t_cnt;
182 return tic->t_unit_res;
187 xlog_grant_head_wake(
189 struct xlog_grant_head *head,
192 struct xlog_ticket *tic;
195 list_for_each_entry(tic, &head->waiters, t_queue) {
196 need_bytes = xlog_ticket_reservation(log, head, tic);
197 if (*free_bytes < need_bytes)
200 *free_bytes -= need_bytes;
201 trace_xfs_log_grant_wake_up(log, tic);
202 wake_up_process(tic->t_task);
209 xlog_grant_head_wait(
211 struct xlog_grant_head *head,
212 struct xlog_ticket *tic,
215 list_add_tail(&tic->t_queue, &head->waiters);
218 if (XLOG_FORCED_SHUTDOWN(log))
220 xlog_grant_push_ail(log, need_bytes);
222 __set_current_state(TASK_UNINTERRUPTIBLE);
223 spin_unlock(&head->lock);
225 XFS_STATS_INC(xs_sleep_logspace);
227 trace_xfs_log_grant_sleep(log, tic);
229 trace_xfs_log_grant_wake(log, tic);
231 spin_lock(&head->lock);
232 if (XLOG_FORCED_SHUTDOWN(log))
234 } while (xlog_space_left(log, &head->grant) < need_bytes);
236 list_del_init(&tic->t_queue);
239 list_del_init(&tic->t_queue);
240 return XFS_ERROR(EIO);
244 * Atomically get the log space required for a log ticket.
246 * Once a ticket gets put onto head->waiters, it will only return after the
247 * needed reservation is satisfied.
249 * This function is structured so that it has a lock free fast path. This is
250 * necessary because every new transaction reservation will come through this
251 * path. Hence any lock will be globally hot if we take it unconditionally on
254 * As tickets are only ever moved on and off head->waiters under head->lock, we
255 * only need to take that lock if we are going to add the ticket to the queue
256 * and sleep. We can avoid taking the lock if the ticket was never added to
257 * head->waiters because the t_queue list head will be empty and we hold the
258 * only reference to it so it can safely be checked unlocked.
261 xlog_grant_head_check(
263 struct xlog_grant_head *head,
264 struct xlog_ticket *tic,
270 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
273 * If there are other waiters on the queue then give them a chance at
274 * logspace before us. Wake up the first waiters, if we do not wake
275 * up all the waiters then go to sleep waiting for more free space,
276 * otherwise try to get some space for this transaction.
278 *need_bytes = xlog_ticket_reservation(log, head, tic);
279 free_bytes = xlog_space_left(log, &head->grant);
280 if (!list_empty_careful(&head->waiters)) {
281 spin_lock(&head->lock);
282 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
283 free_bytes < *need_bytes) {
284 error = xlog_grant_head_wait(log, head, tic,
287 spin_unlock(&head->lock);
288 } else if (free_bytes < *need_bytes) {
289 spin_lock(&head->lock);
290 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
291 spin_unlock(&head->lock);
298 xlog_tic_reset_res(xlog_ticket_t *tic)
301 tic->t_res_arr_sum = 0;
302 tic->t_res_num_ophdrs = 0;
306 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
308 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
309 /* add to overflow and start again */
310 tic->t_res_o_flow += tic->t_res_arr_sum;
312 tic->t_res_arr_sum = 0;
315 tic->t_res_arr[tic->t_res_num].r_len = len;
316 tic->t_res_arr[tic->t_res_num].r_type = type;
317 tic->t_res_arr_sum += len;
322 * Replenish the byte reservation required by moving the grant write head.
326 struct xfs_mount *mp,
327 struct xlog_ticket *tic)
329 struct log *log = mp->m_log;
333 if (XLOG_FORCED_SHUTDOWN(log))
334 return XFS_ERROR(EIO);
336 XFS_STATS_INC(xs_try_logspace);
339 * This is a new transaction on the ticket, so we need to change the
340 * transaction ID so that the next transaction has a different TID in
341 * the log. Just add one to the existing tid so that we can see chains
342 * of rolling transactions in the log easily.
346 xlog_grant_push_ail(log, tic->t_unit_res);
348 tic->t_curr_res = tic->t_unit_res;
349 xlog_tic_reset_res(tic);
354 trace_xfs_log_regrant(log, tic);
356 error = xlog_grant_head_check(log, &log->l_write_head, tic,
361 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
362 trace_xfs_log_regrant_exit(log, tic);
363 xlog_verify_grant_tail(log);
368 * If we are failing, make sure the ticket doesn't have any current
369 * reservations. We don't want to add this back when the ticket/
370 * transaction gets cancelled.
373 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
378 * Reserve log space and return a ticket corresponding the reservation.
380 * Each reservation is going to reserve extra space for a log record header.
381 * When writes happen to the on-disk log, we don't subtract the length of the
382 * log record header from any reservation. By wasting space in each
383 * reservation, we prevent over allocation problems.
387 struct xfs_mount *mp,
390 struct xlog_ticket **ticp,
395 struct log *log = mp->m_log;
396 struct xlog_ticket *tic;
400 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
402 if (XLOG_FORCED_SHUTDOWN(log))
403 return XFS_ERROR(EIO);
405 XFS_STATS_INC(xs_try_logspace);
407 ASSERT(*ticp == NULL);
408 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
409 KM_SLEEP | KM_MAYFAIL);
411 return XFS_ERROR(ENOMEM);
413 tic->t_trans_type = t_type;
416 xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
418 trace_xfs_log_reserve(log, tic);
420 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
425 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
426 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
427 trace_xfs_log_reserve_exit(log, tic);
428 xlog_verify_grant_tail(log);
433 * If we are failing, make sure the ticket doesn't have any current
434 * reservations. We don't want to add this back when the ticket/
435 * transaction gets cancelled.
438 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
446 * 1. currblock field gets updated at startup and after in-core logs
447 * marked as with WANT_SYNC.
451 * This routine is called when a user of a log manager ticket is done with
452 * the reservation. If the ticket was ever used, then a commit record for
453 * the associated transaction is written out as a log operation header with
454 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
455 * a given ticket. If the ticket was one with a permanent reservation, then
456 * a few operations are done differently. Permanent reservation tickets by
457 * default don't release the reservation. They just commit the current
458 * transaction with the belief that the reservation is still needed. A flag
459 * must be passed in before permanent reservations are actually released.
460 * When these type of tickets are not released, they need to be set into
461 * the inited state again. By doing this, a start record will be written
462 * out when the next write occurs.
466 struct xfs_mount *mp,
467 struct xlog_ticket *ticket,
468 struct xlog_in_core **iclog,
471 struct log *log = mp->m_log;
474 if (XLOG_FORCED_SHUTDOWN(log) ||
476 * If nothing was ever written, don't write out commit record.
477 * If we get an error, just continue and give back the log ticket.
479 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
480 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
481 lsn = (xfs_lsn_t) -1;
482 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
483 flags |= XFS_LOG_REL_PERM_RESERV;
488 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
489 (flags & XFS_LOG_REL_PERM_RESERV)) {
490 trace_xfs_log_done_nonperm(log, ticket);
493 * Release ticket if not permanent reservation or a specific
494 * request has been made to release a permanent reservation.
496 xlog_ungrant_log_space(log, ticket);
497 xfs_log_ticket_put(ticket);
499 trace_xfs_log_done_perm(log, ticket);
501 xlog_regrant_reserve_log_space(log, ticket);
502 /* If this ticket was a permanent reservation and we aren't
503 * trying to release it, reset the inited flags; so next time
504 * we write, a start record will be written out.
506 ticket->t_flags |= XLOG_TIC_INITED;
513 * Attaches a new iclog I/O completion callback routine during
514 * transaction commit. If the log is in error state, a non-zero
515 * return code is handed back and the caller is responsible for
516 * executing the callback at an appropriate time.
520 struct xfs_mount *mp,
521 struct xlog_in_core *iclog,
522 xfs_log_callback_t *cb)
526 spin_lock(&iclog->ic_callback_lock);
527 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
529 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
530 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
532 *(iclog->ic_callback_tail) = cb;
533 iclog->ic_callback_tail = &(cb->cb_next);
535 spin_unlock(&iclog->ic_callback_lock);
540 xfs_log_release_iclog(
541 struct xfs_mount *mp,
542 struct xlog_in_core *iclog)
544 if (xlog_state_release_iclog(mp->m_log, iclog)) {
545 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
553 * Mount a log filesystem
555 * mp - ubiquitous xfs mount point structure
556 * log_target - buftarg of on-disk log device
557 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
558 * num_bblocks - Number of BBSIZE blocks in on-disk log
560 * Return error or zero.
565 xfs_buftarg_t *log_target,
566 xfs_daddr_t blk_offset,
571 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
572 xfs_notice(mp, "Mounting Filesystem");
575 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
576 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
579 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
580 if (IS_ERR(mp->m_log)) {
581 error = -PTR_ERR(mp->m_log);
586 * Initialize the AIL now we have a log.
588 error = xfs_trans_ail_init(mp);
590 xfs_warn(mp, "AIL initialisation failed: error %d", error);
593 mp->m_log->l_ailp = mp->m_ail;
596 * skip log recovery on a norecovery mount. pretend it all
599 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
600 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
603 mp->m_flags &= ~XFS_MOUNT_RDONLY;
605 error = xlog_recover(mp->m_log);
608 mp->m_flags |= XFS_MOUNT_RDONLY;
610 xfs_warn(mp, "log mount/recovery failed: error %d",
612 goto out_destroy_ail;
616 /* Normal transactions can now occur */
617 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
620 * Now the log has been fully initialised and we know were our
621 * space grant counters are, we can initialise the permanent ticket
622 * needed for delayed logging to work.
624 xlog_cil_init_post_recovery(mp->m_log);
629 xfs_trans_ail_destroy(mp);
631 xlog_dealloc_log(mp->m_log);
637 * Finish the recovery of the file system. This is separate from
638 * the xfs_log_mount() call, because it depends on the code in
639 * xfs_mountfs() to read in the root and real-time bitmap inodes
640 * between calling xfs_log_mount() and here.
642 * mp - ubiquitous xfs mount point structure
645 xfs_log_mount_finish(xfs_mount_t *mp)
649 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
650 error = xlog_recover_finish(mp->m_log);
653 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
660 * Final log writes as part of unmount.
662 * Mark the filesystem clean as unmount happens. Note that during relocation
663 * this routine needs to be executed as part of source-bag while the
664 * deallocation must not be done until source-end.
668 * Unmount record used to have a string "Unmount filesystem--" in the
669 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
670 * We just write the magic number now since that particular field isn't
671 * currently architecture converted and "nUmount" is a bit foo.
672 * As far as I know, there weren't any dependencies on the old behaviour.
676 xfs_log_unmount_write(xfs_mount_t *mp)
678 xlog_t *log = mp->m_log;
679 xlog_in_core_t *iclog;
681 xlog_in_core_t *first_iclog;
683 xlog_ticket_t *tic = NULL;
688 * Don't write out unmount record on read-only mounts.
689 * Or, if we are doing a forced umount (typically because of IO errors).
691 if (mp->m_flags & XFS_MOUNT_RDONLY)
694 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
695 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
698 first_iclog = iclog = log->l_iclog;
700 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
701 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
702 ASSERT(iclog->ic_offset == 0);
704 iclog = iclog->ic_next;
705 } while (iclog != first_iclog);
707 if (! (XLOG_FORCED_SHUTDOWN(log))) {
708 error = xfs_log_reserve(mp, 600, 1, &tic,
709 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
711 /* the data section must be 32 bit size aligned */
715 __uint32_t pad2; /* may as well make it 64 bits */
717 .magic = XLOG_UNMOUNT_TYPE,
719 struct xfs_log_iovec reg = {
721 .i_len = sizeof(magic),
722 .i_type = XLOG_REG_TYPE_UNMOUNT,
724 struct xfs_log_vec vec = {
729 /* remove inited flag, and account for space used */
731 tic->t_curr_res -= sizeof(magic);
732 error = xlog_write(log, &vec, tic, &lsn,
733 NULL, XLOG_UNMOUNT_TRANS);
735 * At this point, we're umounting anyway,
736 * so there's no point in transitioning log state
737 * to IOERROR. Just continue...
742 xfs_alert(mp, "%s: unmount record failed", __func__);
745 spin_lock(&log->l_icloglock);
746 iclog = log->l_iclog;
747 atomic_inc(&iclog->ic_refcnt);
748 xlog_state_want_sync(log, iclog);
749 spin_unlock(&log->l_icloglock);
750 error = xlog_state_release_iclog(log, iclog);
752 spin_lock(&log->l_icloglock);
753 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
754 iclog->ic_state == XLOG_STATE_DIRTY)) {
755 if (!XLOG_FORCED_SHUTDOWN(log)) {
756 xlog_wait(&iclog->ic_force_wait,
759 spin_unlock(&log->l_icloglock);
762 spin_unlock(&log->l_icloglock);
765 trace_xfs_log_umount_write(log, tic);
766 xlog_ungrant_log_space(log, tic);
767 xfs_log_ticket_put(tic);
771 * We're already in forced_shutdown mode, couldn't
772 * even attempt to write out the unmount transaction.
774 * Go through the motions of sync'ing and releasing
775 * the iclog, even though no I/O will actually happen,
776 * we need to wait for other log I/Os that may already
777 * be in progress. Do this as a separate section of
778 * code so we'll know if we ever get stuck here that
779 * we're in this odd situation of trying to unmount
780 * a file system that went into forced_shutdown as
781 * the result of an unmount..
783 spin_lock(&log->l_icloglock);
784 iclog = log->l_iclog;
785 atomic_inc(&iclog->ic_refcnt);
787 xlog_state_want_sync(log, iclog);
788 spin_unlock(&log->l_icloglock);
789 error = xlog_state_release_iclog(log, iclog);
791 spin_lock(&log->l_icloglock);
793 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
794 || iclog->ic_state == XLOG_STATE_DIRTY
795 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
797 xlog_wait(&iclog->ic_force_wait,
800 spin_unlock(&log->l_icloglock);
805 } /* xfs_log_unmount_write */
808 * Deallocate log structures for unmount/relocation.
810 * We need to stop the aild from running before we destroy
811 * and deallocate the log as the aild references the log.
814 xfs_log_unmount(xfs_mount_t *mp)
816 xfs_trans_ail_destroy(mp);
817 xlog_dealloc_log(mp->m_log);
822 struct xfs_mount *mp,
823 struct xfs_log_item *item,
825 const struct xfs_item_ops *ops)
827 item->li_mountp = mp;
828 item->li_ailp = mp->m_ail;
829 item->li_type = type;
833 INIT_LIST_HEAD(&item->li_ail);
834 INIT_LIST_HEAD(&item->li_cil);
838 * Wake up processes waiting for log space after we have moved the log tail.
842 struct xfs_mount *mp)
844 struct log *log = mp->m_log;
847 if (XLOG_FORCED_SHUTDOWN(log))
850 if (!list_empty_careful(&log->l_write_head.waiters)) {
851 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
853 spin_lock(&log->l_write_head.lock);
854 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
855 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
856 spin_unlock(&log->l_write_head.lock);
859 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
860 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
862 spin_lock(&log->l_reserve_head.lock);
863 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
864 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
865 spin_unlock(&log->l_reserve_head.lock);
870 * Determine if we have a transaction that has gone to disk
871 * that needs to be covered. To begin the transition to the idle state
872 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
873 * If we are then in a state where covering is needed, the caller is informed
874 * that dummy transactions are required to move the log into the idle state.
876 * Because this is called as part of the sync process, we should also indicate
877 * that dummy transactions should be issued in anything but the covered or
878 * idle states. This ensures that the log tail is accurately reflected in
879 * the log at the end of the sync, hence if a crash occurrs avoids replay
880 * of transactions where the metadata is already on disk.
883 xfs_log_need_covered(xfs_mount_t *mp)
886 xlog_t *log = mp->m_log;
888 if (!xfs_fs_writable(mp))
891 spin_lock(&log->l_icloglock);
892 switch (log->l_covered_state) {
893 case XLOG_STATE_COVER_DONE:
894 case XLOG_STATE_COVER_DONE2:
895 case XLOG_STATE_COVER_IDLE:
897 case XLOG_STATE_COVER_NEED:
898 case XLOG_STATE_COVER_NEED2:
899 if (!xfs_ail_min_lsn(log->l_ailp) &&
900 xlog_iclogs_empty(log)) {
901 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
902 log->l_covered_state = XLOG_STATE_COVER_DONE;
904 log->l_covered_state = XLOG_STATE_COVER_DONE2;
911 spin_unlock(&log->l_icloglock);
916 * We may be holding the log iclog lock upon entering this routine.
919 xlog_assign_tail_lsn_locked(
920 struct xfs_mount *mp)
922 struct log *log = mp->m_log;
923 struct xfs_log_item *lip;
926 assert_spin_locked(&mp->m_ail->xa_lock);
929 * To make sure we always have a valid LSN for the log tail we keep
930 * track of the last LSN which was committed in log->l_last_sync_lsn,
931 * and use that when the AIL was empty.
933 lip = xfs_ail_min(mp->m_ail);
935 tail_lsn = lip->li_lsn;
937 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
938 atomic64_set(&log->l_tail_lsn, tail_lsn);
943 xlog_assign_tail_lsn(
944 struct xfs_mount *mp)
948 spin_lock(&mp->m_ail->xa_lock);
949 tail_lsn = xlog_assign_tail_lsn_locked(mp);
950 spin_unlock(&mp->m_ail->xa_lock);
956 * Return the space in the log between the tail and the head. The head
957 * is passed in the cycle/bytes formal parms. In the special case where
958 * the reserve head has wrapped passed the tail, this calculation is no
959 * longer valid. In this case, just return 0 which means there is no space
960 * in the log. This works for all places where this function is called
961 * with the reserve head. Of course, if the write head were to ever
962 * wrap the tail, we should blow up. Rather than catch this case here,
963 * we depend on other ASSERTions in other parts of the code. XXXmiken
965 * This code also handles the case where the reservation head is behind
966 * the tail. The details of this case are described below, but the end
967 * result is that we return the size of the log as the amount of space left.
980 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
981 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
982 tail_bytes = BBTOB(tail_bytes);
983 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
984 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
985 else if (tail_cycle + 1 < head_cycle)
987 else if (tail_cycle < head_cycle) {
988 ASSERT(tail_cycle == (head_cycle - 1));
989 free_bytes = tail_bytes - head_bytes;
992 * The reservation head is behind the tail.
993 * In this case we just want to return the size of the
994 * log as the amount of space left.
997 "xlog_space_left: head behind tail\n"
998 " tail_cycle = %d, tail_bytes = %d\n"
999 " GH cycle = %d, GH bytes = %d",
1000 tail_cycle, tail_bytes, head_cycle, head_bytes);
1002 free_bytes = log->l_logsize;
1009 * Log function which is called when an io completes.
1011 * The log manager needs its own routine, in order to control what
1012 * happens with the buffer after the write completes.
1015 xlog_iodone(xfs_buf_t *bp)
1017 xlog_in_core_t *iclog = bp->b_fspriv;
1018 xlog_t *l = iclog->ic_log;
1022 * Race to shutdown the filesystem if we see an error.
1024 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1025 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1026 xfs_buf_ioerror_alert(bp, __func__);
1028 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1030 * This flag will be propagated to the trans-committed
1031 * callback routines to let them know that the log-commit
1034 aborted = XFS_LI_ABORTED;
1035 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1036 aborted = XFS_LI_ABORTED;
1039 /* log I/O is always issued ASYNC */
1040 ASSERT(XFS_BUF_ISASYNC(bp));
1041 xlog_state_done_syncing(iclog, aborted);
1043 * do not reference the buffer (bp) here as we could race
1044 * with it being freed after writing the unmount record to the
1051 * Return size of each in-core log record buffer.
1053 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1055 * If the filesystem blocksize is too large, we may need to choose a
1056 * larger size since the directory code currently logs entire blocks.
1060 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1066 if (mp->m_logbufs <= 0)
1067 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1069 log->l_iclog_bufs = mp->m_logbufs;
1072 * Buffer size passed in from mount system call.
1074 if (mp->m_logbsize > 0) {
1075 size = log->l_iclog_size = mp->m_logbsize;
1076 log->l_iclog_size_log = 0;
1078 log->l_iclog_size_log++;
1082 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1083 /* # headers = size / 32k
1084 * one header holds cycles from 32k of data
1087 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1088 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1090 log->l_iclog_hsize = xhdrs << BBSHIFT;
1091 log->l_iclog_heads = xhdrs;
1093 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1094 log->l_iclog_hsize = BBSIZE;
1095 log->l_iclog_heads = 1;
1100 /* All machines use 32kB buffers by default. */
1101 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1102 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1104 /* the default log size is 16k or 32k which is one header sector */
1105 log->l_iclog_hsize = BBSIZE;
1106 log->l_iclog_heads = 1;
1109 /* are we being asked to make the sizes selected above visible? */
1110 if (mp->m_logbufs == 0)
1111 mp->m_logbufs = log->l_iclog_bufs;
1112 if (mp->m_logbsize == 0)
1113 mp->m_logbsize = log->l_iclog_size;
1114 } /* xlog_get_iclog_buffer_size */
1118 * This routine initializes some of the log structure for a given mount point.
1119 * Its primary purpose is to fill in enough, so recovery can occur. However,
1120 * some other stuff may be filled in too.
1123 xlog_alloc_log(xfs_mount_t *mp,
1124 xfs_buftarg_t *log_target,
1125 xfs_daddr_t blk_offset,
1129 xlog_rec_header_t *head;
1130 xlog_in_core_t **iclogp;
1131 xlog_in_core_t *iclog, *prev_iclog=NULL;
1137 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1139 xfs_warn(mp, "Log allocation failed: No memory!");
1144 log->l_targ = log_target;
1145 log->l_logsize = BBTOB(num_bblks);
1146 log->l_logBBstart = blk_offset;
1147 log->l_logBBsize = num_bblks;
1148 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1149 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1151 log->l_prev_block = -1;
1152 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1153 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1154 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1155 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1157 xlog_grant_head_init(&log->l_reserve_head);
1158 xlog_grant_head_init(&log->l_write_head);
1160 error = EFSCORRUPTED;
1161 if (xfs_sb_version_hassector(&mp->m_sb)) {
1162 log2_size = mp->m_sb.sb_logsectlog;
1163 if (log2_size < BBSHIFT) {
1164 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1165 log2_size, BBSHIFT);
1169 log2_size -= BBSHIFT;
1170 if (log2_size > mp->m_sectbb_log) {
1171 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1172 log2_size, mp->m_sectbb_log);
1176 /* for larger sector sizes, must have v2 or external log */
1177 if (log2_size && log->l_logBBstart > 0 &&
1178 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1180 "log sector size (0x%x) invalid for configuration.",
1185 log->l_sectBBsize = 1 << log2_size;
1187 xlog_get_iclog_buffer_size(mp, log);
1190 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, log->l_iclog_size, 0);
1193 bp->b_iodone = xlog_iodone;
1194 ASSERT(xfs_buf_islocked(bp));
1197 spin_lock_init(&log->l_icloglock);
1198 init_waitqueue_head(&log->l_flush_wait);
1200 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1201 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1203 iclogp = &log->l_iclog;
1205 * The amount of memory to allocate for the iclog structure is
1206 * rather funky due to the way the structure is defined. It is
1207 * done this way so that we can use different sizes for machines
1208 * with different amounts of memory. See the definition of
1209 * xlog_in_core_t in xfs_log_priv.h for details.
1211 ASSERT(log->l_iclog_size >= 4096);
1212 for (i=0; i < log->l_iclog_bufs; i++) {
1213 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1215 goto out_free_iclog;
1218 iclog->ic_prev = prev_iclog;
1221 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1222 log->l_iclog_size, 0);
1224 goto out_free_iclog;
1226 bp->b_iodone = xlog_iodone;
1228 iclog->ic_data = bp->b_addr;
1230 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1232 head = &iclog->ic_header;
1233 memset(head, 0, sizeof(xlog_rec_header_t));
1234 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1235 head->h_version = cpu_to_be32(
1236 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1237 head->h_size = cpu_to_be32(log->l_iclog_size);
1239 head->h_fmt = cpu_to_be32(XLOG_FMT);
1240 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1242 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1243 iclog->ic_state = XLOG_STATE_ACTIVE;
1244 iclog->ic_log = log;
1245 atomic_set(&iclog->ic_refcnt, 0);
1246 spin_lock_init(&iclog->ic_callback_lock);
1247 iclog->ic_callback_tail = &(iclog->ic_callback);
1248 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1250 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1251 init_waitqueue_head(&iclog->ic_force_wait);
1252 init_waitqueue_head(&iclog->ic_write_wait);
1254 iclogp = &iclog->ic_next;
1256 *iclogp = log->l_iclog; /* complete ring */
1257 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1259 error = xlog_cil_init(log);
1261 goto out_free_iclog;
1265 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1266 prev_iclog = iclog->ic_next;
1268 xfs_buf_free(iclog->ic_bp);
1271 spinlock_destroy(&log->l_icloglock);
1272 xfs_buf_free(log->l_xbuf);
1276 return ERR_PTR(-error);
1277 } /* xlog_alloc_log */
1281 * Write out the commit record of a transaction associated with the given
1282 * ticket. Return the lsn of the commit record.
1287 struct xlog_ticket *ticket,
1288 struct xlog_in_core **iclog,
1289 xfs_lsn_t *commitlsnp)
1291 struct xfs_mount *mp = log->l_mp;
1293 struct xfs_log_iovec reg = {
1296 .i_type = XLOG_REG_TYPE_COMMIT,
1298 struct xfs_log_vec vec = {
1303 ASSERT_ALWAYS(iclog);
1304 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1307 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1312 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1313 * log space. This code pushes on the lsn which would supposedly free up
1314 * the 25% which we want to leave free. We may need to adopt a policy which
1315 * pushes on an lsn which is further along in the log once we reach the high
1316 * water mark. In this manner, we would be creating a low water mark.
1319 xlog_grant_push_ail(
1323 xfs_lsn_t threshold_lsn = 0;
1324 xfs_lsn_t last_sync_lsn;
1327 int threshold_block;
1328 int threshold_cycle;
1331 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1333 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1334 free_blocks = BTOBBT(free_bytes);
1337 * Set the threshold for the minimum number of free blocks in the
1338 * log to the maximum of what the caller needs, one quarter of the
1339 * log, and 256 blocks.
1341 free_threshold = BTOBB(need_bytes);
1342 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1343 free_threshold = MAX(free_threshold, 256);
1344 if (free_blocks >= free_threshold)
1347 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1349 threshold_block += free_threshold;
1350 if (threshold_block >= log->l_logBBsize) {
1351 threshold_block -= log->l_logBBsize;
1352 threshold_cycle += 1;
1354 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1357 * Don't pass in an lsn greater than the lsn of the last
1358 * log record known to be on disk. Use a snapshot of the last sync lsn
1359 * so that it doesn't change between the compare and the set.
1361 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1362 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1363 threshold_lsn = last_sync_lsn;
1366 * Get the transaction layer to kick the dirty buffers out to
1367 * disk asynchronously. No point in trying to do this if
1368 * the filesystem is shutting down.
1370 if (!XLOG_FORCED_SHUTDOWN(log))
1371 xfs_ail_push(log->l_ailp, threshold_lsn);
1375 * The bdstrat callback function for log bufs. This gives us a central
1376 * place to trap bufs in case we get hit by a log I/O error and need to
1377 * shutdown. Actually, in practice, even when we didn't get a log error,
1378 * we transition the iclogs to IOERROR state *after* flushing all existing
1379 * iclogs to disk. This is because we don't want anymore new transactions to be
1380 * started or completed afterwards.
1386 struct xlog_in_core *iclog = bp->b_fspriv;
1388 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1389 xfs_buf_ioerror(bp, EIO);
1391 xfs_buf_ioend(bp, 0);
1393 * It would seem logical to return EIO here, but we rely on
1394 * the log state machine to propagate I/O errors instead of
1400 xfs_buf_iorequest(bp);
1405 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1406 * fashion. Previously, we should have moved the current iclog
1407 * ptr in the log to point to the next available iclog. This allows further
1408 * write to continue while this code syncs out an iclog ready to go.
1409 * Before an in-core log can be written out, the data section must be scanned
1410 * to save away the 1st word of each BBSIZE block into the header. We replace
1411 * it with the current cycle count. Each BBSIZE block is tagged with the
1412 * cycle count because there in an implicit assumption that drives will
1413 * guarantee that entire 512 byte blocks get written at once. In other words,
1414 * we can't have part of a 512 byte block written and part not written. By
1415 * tagging each block, we will know which blocks are valid when recovering
1416 * after an unclean shutdown.
1418 * This routine is single threaded on the iclog. No other thread can be in
1419 * this routine with the same iclog. Changing contents of iclog can there-
1420 * fore be done without grabbing the state machine lock. Updating the global
1421 * log will require grabbing the lock though.
1423 * The entire log manager uses a logical block numbering scheme. Only
1424 * log_sync (and then only bwrite()) know about the fact that the log may
1425 * not start with block zero on a given device. The log block start offset
1426 * is added immediately before calling bwrite().
1430 xlog_sync(xlog_t *log,
1431 xlog_in_core_t *iclog)
1433 xfs_caddr_t dptr; /* pointer to byte sized element */
1436 uint count; /* byte count of bwrite */
1437 uint count_init; /* initial count before roundup */
1438 int roundoff; /* roundoff to BB or stripe */
1439 int split = 0; /* split write into two regions */
1441 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1443 XFS_STATS_INC(xs_log_writes);
1444 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1446 /* Add for LR header */
1447 count_init = log->l_iclog_hsize + iclog->ic_offset;
1449 /* Round out the log write size */
1450 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1451 /* we have a v2 stripe unit to use */
1452 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1454 count = BBTOB(BTOBB(count_init));
1456 roundoff = count - count_init;
1457 ASSERT(roundoff >= 0);
1458 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1459 roundoff < log->l_mp->m_sb.sb_logsunit)
1461 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1462 roundoff < BBTOB(1)));
1464 /* move grant heads by roundoff in sync */
1465 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1466 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1468 /* put cycle number in every block */
1469 xlog_pack_data(log, iclog, roundoff);
1471 /* real byte length */
1473 iclog->ic_header.h_len =
1474 cpu_to_be32(iclog->ic_offset + roundoff);
1476 iclog->ic_header.h_len =
1477 cpu_to_be32(iclog->ic_offset);
1481 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1483 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1485 /* Do we need to split this write into 2 parts? */
1486 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1487 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1488 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1489 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1491 iclog->ic_bwritecnt = 1;
1493 XFS_BUF_SET_COUNT(bp, count);
1494 bp->b_fspriv = iclog;
1495 XFS_BUF_ZEROFLAGS(bp);
1497 bp->b_flags |= XBF_SYNCIO;
1499 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1500 bp->b_flags |= XBF_FUA;
1503 * Flush the data device before flushing the log to make
1504 * sure all meta data written back from the AIL actually made
1505 * it to disk before stamping the new log tail LSN into the
1506 * log buffer. For an external log we need to issue the
1507 * flush explicitly, and unfortunately synchronously here;
1508 * for an internal log we can simply use the block layer
1509 * state machine for preflushes.
1511 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1512 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1514 bp->b_flags |= XBF_FLUSH;
1517 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1518 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1520 xlog_verify_iclog(log, iclog, count, B_TRUE);
1522 /* account for log which doesn't start at block #0 */
1523 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1525 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1530 error = xlog_bdstrat(bp);
1532 xfs_buf_ioerror_alert(bp, "xlog_sync");
1536 bp = iclog->ic_log->l_xbuf;
1537 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1538 xfs_buf_associate_memory(bp,
1539 (char *)&iclog->ic_header + count, split);
1540 bp->b_fspriv = iclog;
1541 XFS_BUF_ZEROFLAGS(bp);
1543 bp->b_flags |= XBF_SYNCIO;
1544 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1545 bp->b_flags |= XBF_FUA;
1548 * Bump the cycle numbers at the start of each block
1549 * since this part of the buffer is at the start of
1550 * a new cycle. Watch out for the header magic number
1553 for (i = 0; i < split; i += BBSIZE) {
1554 be32_add_cpu((__be32 *)dptr, 1);
1555 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1556 be32_add_cpu((__be32 *)dptr, 1);
1560 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1561 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1563 /* account for internal log which doesn't start at block #0 */
1564 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1566 error = xlog_bdstrat(bp);
1568 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1577 * Deallocate a log structure
1580 xlog_dealloc_log(xlog_t *log)
1582 xlog_in_core_t *iclog, *next_iclog;
1585 xlog_cil_destroy(log);
1588 * always need to ensure that the extra buffer does not point to memory
1589 * owned by another log buffer before we free it.
1591 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size);
1592 xfs_buf_free(log->l_xbuf);
1594 iclog = log->l_iclog;
1595 for (i=0; i<log->l_iclog_bufs; i++) {
1596 xfs_buf_free(iclog->ic_bp);
1597 next_iclog = iclog->ic_next;
1601 spinlock_destroy(&log->l_icloglock);
1603 log->l_mp->m_log = NULL;
1605 } /* xlog_dealloc_log */
1608 * Update counters atomically now that memcpy is done.
1612 xlog_state_finish_copy(xlog_t *log,
1613 xlog_in_core_t *iclog,
1617 spin_lock(&log->l_icloglock);
1619 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1620 iclog->ic_offset += copy_bytes;
1622 spin_unlock(&log->l_icloglock);
1623 } /* xlog_state_finish_copy */
1629 * print out info relating to regions written which consume
1634 struct xfs_mount *mp,
1635 struct xlog_ticket *ticket)
1638 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1640 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1641 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1662 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1706 "xlog_write: reservation summary:\n"
1707 " trans type = %s (%u)\n"
1708 " unit res = %d bytes\n"
1709 " current res = %d bytes\n"
1710 " total reg = %u bytes (o/flow = %u bytes)\n"
1711 " ophdrs = %u (ophdr space = %u bytes)\n"
1712 " ophdr + reg = %u bytes\n"
1713 " num regions = %u\n",
1714 ((ticket->t_trans_type <= 0 ||
1715 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1716 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1717 ticket->t_trans_type,
1720 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1721 ticket->t_res_num_ophdrs, ophdr_spc,
1722 ticket->t_res_arr_sum +
1723 ticket->t_res_o_flow + ophdr_spc,
1726 for (i = 0; i < ticket->t_res_num; i++) {
1727 uint r_type = ticket->t_res_arr[i].r_type;
1728 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1729 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1730 "bad-rtype" : res_type_str[r_type-1]),
1731 ticket->t_res_arr[i].r_len);
1734 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1735 "xlog_write: reservation ran out. Need to up reservation");
1736 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1740 * Calculate the potential space needed by the log vector. Each region gets
1741 * its own xlog_op_header_t and may need to be double word aligned.
1744 xlog_write_calc_vec_length(
1745 struct xlog_ticket *ticket,
1746 struct xfs_log_vec *log_vector)
1748 struct xfs_log_vec *lv;
1753 /* acct for start rec of xact */
1754 if (ticket->t_flags & XLOG_TIC_INITED)
1757 for (lv = log_vector; lv; lv = lv->lv_next) {
1758 headers += lv->lv_niovecs;
1760 for (i = 0; i < lv->lv_niovecs; i++) {
1761 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1764 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1768 ticket->t_res_num_ophdrs += headers;
1769 len += headers * sizeof(struct xlog_op_header);
1775 * If first write for transaction, insert start record We can't be trying to
1776 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1779 xlog_write_start_rec(
1780 struct xlog_op_header *ophdr,
1781 struct xlog_ticket *ticket)
1783 if (!(ticket->t_flags & XLOG_TIC_INITED))
1786 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1787 ophdr->oh_clientid = ticket->t_clientid;
1789 ophdr->oh_flags = XLOG_START_TRANS;
1792 ticket->t_flags &= ~XLOG_TIC_INITED;
1794 return sizeof(struct xlog_op_header);
1797 static xlog_op_header_t *
1798 xlog_write_setup_ophdr(
1800 struct xlog_op_header *ophdr,
1801 struct xlog_ticket *ticket,
1804 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1805 ophdr->oh_clientid = ticket->t_clientid;
1808 /* are we copying a commit or unmount record? */
1809 ophdr->oh_flags = flags;
1812 * We've seen logs corrupted with bad transaction client ids. This
1813 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1814 * and shut down the filesystem.
1816 switch (ophdr->oh_clientid) {
1817 case XFS_TRANSACTION:
1823 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1824 ophdr->oh_clientid, ticket);
1832 * Set up the parameters of the region copy into the log. This has
1833 * to handle region write split across multiple log buffers - this
1834 * state is kept external to this function so that this code can
1835 * can be written in an obvious, self documenting manner.
1838 xlog_write_setup_copy(
1839 struct xlog_ticket *ticket,
1840 struct xlog_op_header *ophdr,
1841 int space_available,
1845 int *last_was_partial_copy,
1846 int *bytes_consumed)
1850 still_to_copy = space_required - *bytes_consumed;
1851 *copy_off = *bytes_consumed;
1853 if (still_to_copy <= space_available) {
1854 /* write of region completes here */
1855 *copy_len = still_to_copy;
1856 ophdr->oh_len = cpu_to_be32(*copy_len);
1857 if (*last_was_partial_copy)
1858 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1859 *last_was_partial_copy = 0;
1860 *bytes_consumed = 0;
1864 /* partial write of region, needs extra log op header reservation */
1865 *copy_len = space_available;
1866 ophdr->oh_len = cpu_to_be32(*copy_len);
1867 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1868 if (*last_was_partial_copy)
1869 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1870 *bytes_consumed += *copy_len;
1871 (*last_was_partial_copy)++;
1873 /* account for new log op header */
1874 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1875 ticket->t_res_num_ophdrs++;
1877 return sizeof(struct xlog_op_header);
1881 xlog_write_copy_finish(
1883 struct xlog_in_core *iclog,
1888 int *partial_copy_len,
1890 struct xlog_in_core **commit_iclog)
1892 if (*partial_copy) {
1894 * This iclog has already been marked WANT_SYNC by
1895 * xlog_state_get_iclog_space.
1897 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1900 return xlog_state_release_iclog(log, iclog);
1904 *partial_copy_len = 0;
1906 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1907 /* no more space in this iclog - push it. */
1908 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1912 spin_lock(&log->l_icloglock);
1913 xlog_state_want_sync(log, iclog);
1914 spin_unlock(&log->l_icloglock);
1917 return xlog_state_release_iclog(log, iclog);
1918 ASSERT(flags & XLOG_COMMIT_TRANS);
1919 *commit_iclog = iclog;
1926 * Write some region out to in-core log
1928 * This will be called when writing externally provided regions or when
1929 * writing out a commit record for a given transaction.
1931 * General algorithm:
1932 * 1. Find total length of this write. This may include adding to the
1933 * lengths passed in.
1934 * 2. Check whether we violate the tickets reservation.
1935 * 3. While writing to this iclog
1936 * A. Reserve as much space in this iclog as can get
1937 * B. If this is first write, save away start lsn
1938 * C. While writing this region:
1939 * 1. If first write of transaction, write start record
1940 * 2. Write log operation header (header per region)
1941 * 3. Find out if we can fit entire region into this iclog
1942 * 4. Potentially, verify destination memcpy ptr
1943 * 5. Memcpy (partial) region
1944 * 6. If partial copy, release iclog; otherwise, continue
1945 * copying more regions into current iclog
1946 * 4. Mark want sync bit (in simulation mode)
1947 * 5. Release iclog for potential flush to on-disk log.
1950 * 1. Panic if reservation is overrun. This should never happen since
1951 * reservation amounts are generated internal to the filesystem.
1953 * 1. Tickets are single threaded data structures.
1954 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1955 * syncing routine. When a single log_write region needs to span
1956 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1957 * on all log operation writes which don't contain the end of the
1958 * region. The XLOG_END_TRANS bit is used for the in-core log
1959 * operation which contains the end of the continued log_write region.
1960 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1961 * we don't really know exactly how much space will be used. As a result,
1962 * we don't update ic_offset until the end when we know exactly how many
1963 * bytes have been written out.
1968 struct xfs_log_vec *log_vector,
1969 struct xlog_ticket *ticket,
1970 xfs_lsn_t *start_lsn,
1971 struct xlog_in_core **commit_iclog,
1974 struct xlog_in_core *iclog = NULL;
1975 struct xfs_log_iovec *vecp;
1976 struct xfs_log_vec *lv;
1979 int partial_copy = 0;
1980 int partial_copy_len = 0;
1988 len = xlog_write_calc_vec_length(ticket, log_vector);
1991 * Region headers and bytes are already accounted for.
1992 * We only need to take into account start records and
1993 * split regions in this function.
1995 if (ticket->t_flags & XLOG_TIC_INITED)
1996 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1999 * Commit record headers need to be accounted for. These
2000 * come in as separate writes so are easy to detect.
2002 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2003 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2005 if (ticket->t_curr_res < 0)
2006 xlog_print_tic_res(log->l_mp, ticket);
2010 vecp = lv->lv_iovecp;
2011 while (lv && index < lv->lv_niovecs) {
2015 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2016 &contwr, &log_offset);
2020 ASSERT(log_offset <= iclog->ic_size - 1);
2021 ptr = iclog->ic_datap + log_offset;
2023 /* start_lsn is the first lsn written to. That's all we need. */
2025 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2028 * This loop writes out as many regions as can fit in the amount
2029 * of space which was allocated by xlog_state_get_iclog_space().
2031 while (lv && index < lv->lv_niovecs) {
2032 struct xfs_log_iovec *reg = &vecp[index];
2033 struct xlog_op_header *ophdr;
2038 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2039 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2041 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2042 if (start_rec_copy) {
2044 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2048 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2050 return XFS_ERROR(EIO);
2052 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2053 sizeof(struct xlog_op_header));
2055 len += xlog_write_setup_copy(ticket, ophdr,
2056 iclog->ic_size-log_offset,
2058 ©_off, ©_len,
2061 xlog_verify_dest_ptr(log, ptr);
2064 ASSERT(copy_len >= 0);
2065 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2066 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2068 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2070 data_cnt += contwr ? copy_len : 0;
2072 error = xlog_write_copy_finish(log, iclog, flags,
2073 &record_cnt, &data_cnt,
2082 * if we had a partial copy, we need to get more iclog
2083 * space but we don't want to increment the region
2084 * index because there is still more is this region to
2087 * If we completed writing this region, and we flushed
2088 * the iclog (indicated by resetting of the record
2089 * count), then we also need to get more log space. If
2090 * this was the last record, though, we are done and
2096 if (++index == lv->lv_niovecs) {
2100 vecp = lv->lv_iovecp;
2102 if (record_cnt == 0) {
2112 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2114 return xlog_state_release_iclog(log, iclog);
2116 ASSERT(flags & XLOG_COMMIT_TRANS);
2117 *commit_iclog = iclog;
2122 /*****************************************************************************
2124 * State Machine functions
2126 *****************************************************************************
2129 /* Clean iclogs starting from the head. This ordering must be
2130 * maintained, so an iclog doesn't become ACTIVE beyond one that
2131 * is SYNCING. This is also required to maintain the notion that we use
2132 * a ordered wait queue to hold off would be writers to the log when every
2133 * iclog is trying to sync to disk.
2135 * State Change: DIRTY -> ACTIVE
2138 xlog_state_clean_log(xlog_t *log)
2140 xlog_in_core_t *iclog;
2143 iclog = log->l_iclog;
2145 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2146 iclog->ic_state = XLOG_STATE_ACTIVE;
2147 iclog->ic_offset = 0;
2148 ASSERT(iclog->ic_callback == NULL);
2150 * If the number of ops in this iclog indicate it just
2151 * contains the dummy transaction, we can
2152 * change state into IDLE (the second time around).
2153 * Otherwise we should change the state into
2155 * We don't need to cover the dummy.
2158 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2163 * We have two dirty iclogs so start over
2164 * This could also be num of ops indicates
2165 * this is not the dummy going out.
2169 iclog->ic_header.h_num_logops = 0;
2170 memset(iclog->ic_header.h_cycle_data, 0,
2171 sizeof(iclog->ic_header.h_cycle_data));
2172 iclog->ic_header.h_lsn = 0;
2173 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2176 break; /* stop cleaning */
2177 iclog = iclog->ic_next;
2178 } while (iclog != log->l_iclog);
2180 /* log is locked when we are called */
2182 * Change state for the dummy log recording.
2183 * We usually go to NEED. But we go to NEED2 if the changed indicates
2184 * we are done writing the dummy record.
2185 * If we are done with the second dummy recored (DONE2), then
2189 switch (log->l_covered_state) {
2190 case XLOG_STATE_COVER_IDLE:
2191 case XLOG_STATE_COVER_NEED:
2192 case XLOG_STATE_COVER_NEED2:
2193 log->l_covered_state = XLOG_STATE_COVER_NEED;
2196 case XLOG_STATE_COVER_DONE:
2198 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2200 log->l_covered_state = XLOG_STATE_COVER_NEED;
2203 case XLOG_STATE_COVER_DONE2:
2205 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2207 log->l_covered_state = XLOG_STATE_COVER_NEED;
2214 } /* xlog_state_clean_log */
2217 xlog_get_lowest_lsn(
2220 xlog_in_core_t *lsn_log;
2221 xfs_lsn_t lowest_lsn, lsn;
2223 lsn_log = log->l_iclog;
2226 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2227 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2228 if ((lsn && !lowest_lsn) ||
2229 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2233 lsn_log = lsn_log->ic_next;
2234 } while (lsn_log != log->l_iclog);
2240 xlog_state_do_callback(
2243 xlog_in_core_t *ciclog)
2245 xlog_in_core_t *iclog;
2246 xlog_in_core_t *first_iclog; /* used to know when we've
2247 * processed all iclogs once */
2248 xfs_log_callback_t *cb, *cb_next;
2250 xfs_lsn_t lowest_lsn;
2251 int ioerrors; /* counter: iclogs with errors */
2252 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2253 int funcdidcallbacks; /* flag: function did callbacks */
2254 int repeats; /* for issuing console warnings if
2255 * looping too many times */
2258 spin_lock(&log->l_icloglock);
2259 first_iclog = iclog = log->l_iclog;
2261 funcdidcallbacks = 0;
2266 * Scan all iclogs starting with the one pointed to by the
2267 * log. Reset this starting point each time the log is
2268 * unlocked (during callbacks).
2270 * Keep looping through iclogs until one full pass is made
2271 * without running any callbacks.
2273 first_iclog = log->l_iclog;
2274 iclog = log->l_iclog;
2275 loopdidcallbacks = 0;
2280 /* skip all iclogs in the ACTIVE & DIRTY states */
2281 if (iclog->ic_state &
2282 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2283 iclog = iclog->ic_next;
2288 * Between marking a filesystem SHUTDOWN and stopping
2289 * the log, we do flush all iclogs to disk (if there
2290 * wasn't a log I/O error). So, we do want things to
2291 * go smoothly in case of just a SHUTDOWN w/o a
2294 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2296 * Can only perform callbacks in order. Since
2297 * this iclog is not in the DONE_SYNC/
2298 * DO_CALLBACK state, we skip the rest and
2299 * just try to clean up. If we set our iclog
2300 * to DO_CALLBACK, we will not process it when
2301 * we retry since a previous iclog is in the
2302 * CALLBACK and the state cannot change since
2303 * we are holding the l_icloglock.
2305 if (!(iclog->ic_state &
2306 (XLOG_STATE_DONE_SYNC |
2307 XLOG_STATE_DO_CALLBACK))) {
2308 if (ciclog && (ciclog->ic_state ==
2309 XLOG_STATE_DONE_SYNC)) {
2310 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2315 * We now have an iclog that is in either the
2316 * DO_CALLBACK or DONE_SYNC states. The other
2317 * states (WANT_SYNC, SYNCING, or CALLBACK were
2318 * caught by the above if and are going to
2319 * clean (i.e. we aren't doing their callbacks)
2324 * We will do one more check here to see if we
2325 * have chased our tail around.
2328 lowest_lsn = xlog_get_lowest_lsn(log);
2330 XFS_LSN_CMP(lowest_lsn,
2331 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2332 iclog = iclog->ic_next;
2333 continue; /* Leave this iclog for
2337 iclog->ic_state = XLOG_STATE_CALLBACK;
2341 * update the last_sync_lsn before we drop the
2342 * icloglock to ensure we are the only one that
2345 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2346 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2347 atomic64_set(&log->l_last_sync_lsn,
2348 be64_to_cpu(iclog->ic_header.h_lsn));
2353 spin_unlock(&log->l_icloglock);
2356 * Keep processing entries in the callback list until
2357 * we come around and it is empty. We need to
2358 * atomically see that the list is empty and change the
2359 * state to DIRTY so that we don't miss any more
2360 * callbacks being added.
2362 spin_lock(&iclog->ic_callback_lock);
2363 cb = iclog->ic_callback;
2365 iclog->ic_callback_tail = &(iclog->ic_callback);
2366 iclog->ic_callback = NULL;
2367 spin_unlock(&iclog->ic_callback_lock);
2369 /* perform callbacks in the order given */
2370 for (; cb; cb = cb_next) {
2371 cb_next = cb->cb_next;
2372 cb->cb_func(cb->cb_arg, aborted);
2374 spin_lock(&iclog->ic_callback_lock);
2375 cb = iclog->ic_callback;
2381 spin_lock(&log->l_icloglock);
2382 ASSERT(iclog->ic_callback == NULL);
2383 spin_unlock(&iclog->ic_callback_lock);
2384 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2385 iclog->ic_state = XLOG_STATE_DIRTY;
2388 * Transition from DIRTY to ACTIVE if applicable.
2389 * NOP if STATE_IOERROR.
2391 xlog_state_clean_log(log);
2393 /* wake up threads waiting in xfs_log_force() */
2394 wake_up_all(&iclog->ic_force_wait);
2396 iclog = iclog->ic_next;
2397 } while (first_iclog != iclog);
2399 if (repeats > 5000) {
2400 flushcnt += repeats;
2403 "%s: possible infinite loop (%d iterations)",
2404 __func__, flushcnt);
2406 } while (!ioerrors && loopdidcallbacks);
2409 * make one last gasp attempt to see if iclogs are being left in
2413 if (funcdidcallbacks) {
2414 first_iclog = iclog = log->l_iclog;
2416 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2418 * Terminate the loop if iclogs are found in states
2419 * which will cause other threads to clean up iclogs.
2421 * SYNCING - i/o completion will go through logs
2422 * DONE_SYNC - interrupt thread should be waiting for
2424 * IOERROR - give up hope all ye who enter here
2426 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2427 iclog->ic_state == XLOG_STATE_SYNCING ||
2428 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2429 iclog->ic_state == XLOG_STATE_IOERROR )
2431 iclog = iclog->ic_next;
2432 } while (first_iclog != iclog);
2436 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2438 spin_unlock(&log->l_icloglock);
2441 wake_up_all(&log->l_flush_wait);
2446 * Finish transitioning this iclog to the dirty state.
2448 * Make sure that we completely execute this routine only when this is
2449 * the last call to the iclog. There is a good chance that iclog flushes,
2450 * when we reach the end of the physical log, get turned into 2 separate
2451 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2452 * routine. By using the reference count bwritecnt, we guarantee that only
2453 * the second completion goes through.
2455 * Callbacks could take time, so they are done outside the scope of the
2456 * global state machine log lock.
2459 xlog_state_done_syncing(
2460 xlog_in_core_t *iclog,
2463 xlog_t *log = iclog->ic_log;
2465 spin_lock(&log->l_icloglock);
2467 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2468 iclog->ic_state == XLOG_STATE_IOERROR);
2469 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2470 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2474 * If we got an error, either on the first buffer, or in the case of
2475 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2476 * and none should ever be attempted to be written to disk
2479 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2480 if (--iclog->ic_bwritecnt == 1) {
2481 spin_unlock(&log->l_icloglock);
2484 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2488 * Someone could be sleeping prior to writing out the next
2489 * iclog buffer, we wake them all, one will get to do the
2490 * I/O, the others get to wait for the result.
2492 wake_up_all(&iclog->ic_write_wait);
2493 spin_unlock(&log->l_icloglock);
2494 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2495 } /* xlog_state_done_syncing */
2499 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2500 * sleep. We wait on the flush queue on the head iclog as that should be
2501 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2502 * we will wait here and all new writes will sleep until a sync completes.
2504 * The in-core logs are used in a circular fashion. They are not used
2505 * out-of-order even when an iclog past the head is free.
2508 * * log_offset where xlog_write() can start writing into the in-core
2510 * * in-core log pointer to which xlog_write() should write.
2511 * * boolean indicating this is a continued write to an in-core log.
2512 * If this is the last write, then the in-core log's offset field
2513 * needs to be incremented, depending on the amount of data which
2517 xlog_state_get_iclog_space(xlog_t *log,
2519 xlog_in_core_t **iclogp,
2520 xlog_ticket_t *ticket,
2521 int *continued_write,
2525 xlog_rec_header_t *head;
2526 xlog_in_core_t *iclog;
2530 spin_lock(&log->l_icloglock);
2531 if (XLOG_FORCED_SHUTDOWN(log)) {
2532 spin_unlock(&log->l_icloglock);
2533 return XFS_ERROR(EIO);
2536 iclog = log->l_iclog;
2537 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2538 XFS_STATS_INC(xs_log_noiclogs);
2540 /* Wait for log writes to have flushed */
2541 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2545 head = &iclog->ic_header;
2547 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2548 log_offset = iclog->ic_offset;
2550 /* On the 1st write to an iclog, figure out lsn. This works
2551 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2552 * committing to. If the offset is set, that's how many blocks
2555 if (log_offset == 0) {
2556 ticket->t_curr_res -= log->l_iclog_hsize;
2557 xlog_tic_add_region(ticket,
2559 XLOG_REG_TYPE_LRHEADER);
2560 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2561 head->h_lsn = cpu_to_be64(
2562 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2563 ASSERT(log->l_curr_block >= 0);
2566 /* If there is enough room to write everything, then do it. Otherwise,
2567 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2568 * bit is on, so this will get flushed out. Don't update ic_offset
2569 * until you know exactly how many bytes get copied. Therefore, wait
2570 * until later to update ic_offset.
2572 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2573 * can fit into remaining data section.
2575 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2576 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2579 * If I'm the only one writing to this iclog, sync it to disk.
2580 * We need to do an atomic compare and decrement here to avoid
2581 * racing with concurrent atomic_dec_and_lock() calls in
2582 * xlog_state_release_iclog() when there is more than one
2583 * reference to the iclog.
2585 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2586 /* we are the only one */
2587 spin_unlock(&log->l_icloglock);
2588 error = xlog_state_release_iclog(log, iclog);
2592 spin_unlock(&log->l_icloglock);
2597 /* Do we have enough room to write the full amount in the remainder
2598 * of this iclog? Or must we continue a write on the next iclog and
2599 * mark this iclog as completely taken? In the case where we switch
2600 * iclogs (to mark it taken), this particular iclog will release/sync
2601 * to disk in xlog_write().
2603 if (len <= iclog->ic_size - iclog->ic_offset) {
2604 *continued_write = 0;
2605 iclog->ic_offset += len;
2607 *continued_write = 1;
2608 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2612 ASSERT(iclog->ic_offset <= iclog->ic_size);
2613 spin_unlock(&log->l_icloglock);
2615 *logoffsetp = log_offset;
2617 } /* xlog_state_get_iclog_space */
2619 /* The first cnt-1 times through here we don't need to
2620 * move the grant write head because the permanent
2621 * reservation has reserved cnt times the unit amount.
2622 * Release part of current permanent unit reservation and
2623 * reset current reservation to be one units worth. Also
2624 * move grant reservation head forward.
2627 xlog_regrant_reserve_log_space(xlog_t *log,
2628 xlog_ticket_t *ticket)
2630 trace_xfs_log_regrant_reserve_enter(log, ticket);
2632 if (ticket->t_cnt > 0)
2635 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2636 ticket->t_curr_res);
2637 xlog_grant_sub_space(log, &log->l_write_head.grant,
2638 ticket->t_curr_res);
2639 ticket->t_curr_res = ticket->t_unit_res;
2640 xlog_tic_reset_res(ticket);
2642 trace_xfs_log_regrant_reserve_sub(log, ticket);
2644 /* just return if we still have some of the pre-reserved space */
2645 if (ticket->t_cnt > 0)
2648 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2649 ticket->t_unit_res);
2651 trace_xfs_log_regrant_reserve_exit(log, ticket);
2653 ticket->t_curr_res = ticket->t_unit_res;
2654 xlog_tic_reset_res(ticket);
2655 } /* xlog_regrant_reserve_log_space */
2659 * Give back the space left from a reservation.
2661 * All the information we need to make a correct determination of space left
2662 * is present. For non-permanent reservations, things are quite easy. The
2663 * count should have been decremented to zero. We only need to deal with the
2664 * space remaining in the current reservation part of the ticket. If the
2665 * ticket contains a permanent reservation, there may be left over space which
2666 * needs to be released. A count of N means that N-1 refills of the current
2667 * reservation can be done before we need to ask for more space. The first
2668 * one goes to fill up the first current reservation. Once we run out of
2669 * space, the count will stay at zero and the only space remaining will be
2670 * in the current reservation field.
2673 xlog_ungrant_log_space(xlog_t *log,
2674 xlog_ticket_t *ticket)
2678 if (ticket->t_cnt > 0)
2681 trace_xfs_log_ungrant_enter(log, ticket);
2682 trace_xfs_log_ungrant_sub(log, ticket);
2685 * If this is a permanent reservation ticket, we may be able to free
2686 * up more space based on the remaining count.
2688 bytes = ticket->t_curr_res;
2689 if (ticket->t_cnt > 0) {
2690 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2691 bytes += ticket->t_unit_res*ticket->t_cnt;
2694 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2695 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2697 trace_xfs_log_ungrant_exit(log, ticket);
2699 xfs_log_space_wake(log->l_mp);
2703 * Flush iclog to disk if this is the last reference to the given iclog and
2704 * the WANT_SYNC bit is set.
2706 * When this function is entered, the iclog is not necessarily in the
2707 * WANT_SYNC state. It may be sitting around waiting to get filled.
2712 xlog_state_release_iclog(
2714 xlog_in_core_t *iclog)
2716 int sync = 0; /* do we sync? */
2718 if (iclog->ic_state & XLOG_STATE_IOERROR)
2719 return XFS_ERROR(EIO);
2721 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2722 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2725 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2726 spin_unlock(&log->l_icloglock);
2727 return XFS_ERROR(EIO);
2729 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2730 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2732 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2733 /* update tail before writing to iclog */
2734 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2736 iclog->ic_state = XLOG_STATE_SYNCING;
2737 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2738 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2739 /* cycle incremented when incrementing curr_block */
2741 spin_unlock(&log->l_icloglock);
2744 * We let the log lock go, so it's possible that we hit a log I/O
2745 * error or some other SHUTDOWN condition that marks the iclog
2746 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2747 * this iclog has consistent data, so we ignore IOERROR
2748 * flags after this point.
2751 return xlog_sync(log, iclog);
2753 } /* xlog_state_release_iclog */
2757 * This routine will mark the current iclog in the ring as WANT_SYNC
2758 * and move the current iclog pointer to the next iclog in the ring.
2759 * When this routine is called from xlog_state_get_iclog_space(), the
2760 * exact size of the iclog has not yet been determined. All we know is
2761 * that every data block. We have run out of space in this log record.
2764 xlog_state_switch_iclogs(xlog_t *log,
2765 xlog_in_core_t *iclog,
2768 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2770 eventual_size = iclog->ic_offset;
2771 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2772 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2773 log->l_prev_block = log->l_curr_block;
2774 log->l_prev_cycle = log->l_curr_cycle;
2776 /* roll log?: ic_offset changed later */
2777 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2779 /* Round up to next log-sunit */
2780 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2781 log->l_mp->m_sb.sb_logsunit > 1) {
2782 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2783 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2786 if (log->l_curr_block >= log->l_logBBsize) {
2787 log->l_curr_cycle++;
2788 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2789 log->l_curr_cycle++;
2790 log->l_curr_block -= log->l_logBBsize;
2791 ASSERT(log->l_curr_block >= 0);
2793 ASSERT(iclog == log->l_iclog);
2794 log->l_iclog = iclog->ic_next;
2795 } /* xlog_state_switch_iclogs */
2798 * Write out all data in the in-core log as of this exact moment in time.
2800 * Data may be written to the in-core log during this call. However,
2801 * we don't guarantee this data will be written out. A change from past
2802 * implementation means this routine will *not* write out zero length LRs.
2804 * Basically, we try and perform an intelligent scan of the in-core logs.
2805 * If we determine there is no flushable data, we just return. There is no
2806 * flushable data if:
2808 * 1. the current iclog is active and has no data; the previous iclog
2809 * is in the active or dirty state.
2810 * 2. the current iclog is drity, and the previous iclog is in the
2811 * active or dirty state.
2815 * 1. the current iclog is not in the active nor dirty state.
2816 * 2. the current iclog dirty, and the previous iclog is not in the
2817 * active nor dirty state.
2818 * 3. the current iclog is active, and there is another thread writing
2819 * to this particular iclog.
2820 * 4. a) the current iclog is active and has no other writers
2821 * b) when we return from flushing out this iclog, it is still
2822 * not in the active nor dirty state.
2826 struct xfs_mount *mp,
2830 struct log *log = mp->m_log;
2831 struct xlog_in_core *iclog;
2834 XFS_STATS_INC(xs_log_force);
2836 xlog_cil_force(log);
2838 spin_lock(&log->l_icloglock);
2840 iclog = log->l_iclog;
2841 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2842 spin_unlock(&log->l_icloglock);
2843 return XFS_ERROR(EIO);
2846 /* If the head iclog is not active nor dirty, we just attach
2847 * ourselves to the head and go to sleep.
2849 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2850 iclog->ic_state == XLOG_STATE_DIRTY) {
2852 * If the head is dirty or (active and empty), then
2853 * we need to look at the previous iclog. If the previous
2854 * iclog is active or dirty we are done. There is nothing
2855 * to sync out. Otherwise, we attach ourselves to the
2856 * previous iclog and go to sleep.
2858 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2859 (atomic_read(&iclog->ic_refcnt) == 0
2860 && iclog->ic_offset == 0)) {
2861 iclog = iclog->ic_prev;
2862 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2863 iclog->ic_state == XLOG_STATE_DIRTY)
2868 if (atomic_read(&iclog->ic_refcnt) == 0) {
2869 /* We are the only one with access to this
2870 * iclog. Flush it out now. There should
2871 * be a roundoff of zero to show that someone
2872 * has already taken care of the roundoff from
2873 * the previous sync.
2875 atomic_inc(&iclog->ic_refcnt);
2876 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2877 xlog_state_switch_iclogs(log, iclog, 0);
2878 spin_unlock(&log->l_icloglock);
2880 if (xlog_state_release_iclog(log, iclog))
2881 return XFS_ERROR(EIO);
2885 spin_lock(&log->l_icloglock);
2886 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2887 iclog->ic_state != XLOG_STATE_DIRTY)
2892 /* Someone else is writing to this iclog.
2893 * Use its call to flush out the data. However,
2894 * the other thread may not force out this LR,
2895 * so we mark it WANT_SYNC.
2897 xlog_state_switch_iclogs(log, iclog, 0);
2903 /* By the time we come around again, the iclog could've been filled
2904 * which would give it another lsn. If we have a new lsn, just
2905 * return because the relevant data has been flushed.
2908 if (flags & XFS_LOG_SYNC) {
2910 * We must check if we're shutting down here, before
2911 * we wait, while we're holding the l_icloglock.
2912 * Then we check again after waking up, in case our
2913 * sleep was disturbed by a bad news.
2915 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2916 spin_unlock(&log->l_icloglock);
2917 return XFS_ERROR(EIO);
2919 XFS_STATS_INC(xs_log_force_sleep);
2920 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2922 * No need to grab the log lock here since we're
2923 * only deciding whether or not to return EIO
2924 * and the memory read should be atomic.
2926 if (iclog->ic_state & XLOG_STATE_IOERROR)
2927 return XFS_ERROR(EIO);
2933 spin_unlock(&log->l_icloglock);
2939 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2940 * about errors or whether the log was flushed or not. This is the normal
2941 * interface to use when trying to unpin items or move the log forward.
2950 error = _xfs_log_force(mp, flags, NULL);
2952 xfs_warn(mp, "%s: error %d returned.", __func__, error);
2956 * Force the in-core log to disk for a specific LSN.
2958 * Find in-core log with lsn.
2959 * If it is in the DIRTY state, just return.
2960 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2961 * state and go to sleep or return.
2962 * If it is in any other state, go to sleep or return.
2964 * Synchronous forces are implemented with a signal variable. All callers
2965 * to force a given lsn to disk will wait on a the sv attached to the
2966 * specific in-core log. When given in-core log finally completes its
2967 * write to disk, that thread will wake up all threads waiting on the
2972 struct xfs_mount *mp,
2977 struct log *log = mp->m_log;
2978 struct xlog_in_core *iclog;
2979 int already_slept = 0;
2983 XFS_STATS_INC(xs_log_force);
2985 lsn = xlog_cil_force_lsn(log, lsn);
2986 if (lsn == NULLCOMMITLSN)
2990 spin_lock(&log->l_icloglock);
2991 iclog = log->l_iclog;
2992 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2993 spin_unlock(&log->l_icloglock);
2994 return XFS_ERROR(EIO);
2998 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
2999 iclog = iclog->ic_next;
3003 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3004 spin_unlock(&log->l_icloglock);
3008 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3010 * We sleep here if we haven't already slept (e.g.
3011 * this is the first time we've looked at the correct
3012 * iclog buf) and the buffer before us is going to
3013 * be sync'ed. The reason for this is that if we
3014 * are doing sync transactions here, by waiting for
3015 * the previous I/O to complete, we can allow a few
3016 * more transactions into this iclog before we close
3019 * Otherwise, we mark the buffer WANT_SYNC, and bump
3020 * up the refcnt so we can release the log (which
3021 * drops the ref count). The state switch keeps new
3022 * transaction commits from using this buffer. When
3023 * the current commits finish writing into the buffer,
3024 * the refcount will drop to zero and the buffer will
3027 if (!already_slept &&
3028 (iclog->ic_prev->ic_state &
3029 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3030 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3032 XFS_STATS_INC(xs_log_force_sleep);
3034 xlog_wait(&iclog->ic_prev->ic_write_wait,
3041 atomic_inc(&iclog->ic_refcnt);
3042 xlog_state_switch_iclogs(log, iclog, 0);
3043 spin_unlock(&log->l_icloglock);
3044 if (xlog_state_release_iclog(log, iclog))
3045 return XFS_ERROR(EIO);
3048 spin_lock(&log->l_icloglock);
3051 if ((flags & XFS_LOG_SYNC) && /* sleep */
3053 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3055 * Don't wait on completion if we know that we've
3056 * gotten a log write error.
3058 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3059 spin_unlock(&log->l_icloglock);
3060 return XFS_ERROR(EIO);
3062 XFS_STATS_INC(xs_log_force_sleep);
3063 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3065 * No need to grab the log lock here since we're
3066 * only deciding whether or not to return EIO
3067 * and the memory read should be atomic.
3069 if (iclog->ic_state & XLOG_STATE_IOERROR)
3070 return XFS_ERROR(EIO);
3074 } else { /* just return */
3075 spin_unlock(&log->l_icloglock);
3079 } while (iclog != log->l_iclog);
3081 spin_unlock(&log->l_icloglock);
3086 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3087 * about errors or whether the log was flushed or not. This is the normal
3088 * interface to use when trying to unpin items or move the log forward.
3098 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3100 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3104 * Called when we want to mark the current iclog as being ready to sync to
3108 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3110 assert_spin_locked(&log->l_icloglock);
3112 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3113 xlog_state_switch_iclogs(log, iclog, 0);
3115 ASSERT(iclog->ic_state &
3116 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3121 /*****************************************************************************
3125 *****************************************************************************
3129 * Free a used ticket when its refcount falls to zero.
3133 xlog_ticket_t *ticket)
3135 ASSERT(atomic_read(&ticket->t_ref) > 0);
3136 if (atomic_dec_and_test(&ticket->t_ref))
3137 kmem_zone_free(xfs_log_ticket_zone, ticket);
3142 xlog_ticket_t *ticket)
3144 ASSERT(atomic_read(&ticket->t_ref) > 0);
3145 atomic_inc(&ticket->t_ref);
3150 * Allocate and initialise a new log ticket.
3161 struct xlog_ticket *tic;
3165 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3170 * Permanent reservations have up to 'cnt'-1 active log operations
3171 * in the log. A unit in this case is the amount of space for one
3172 * of these log operations. Normal reservations have a cnt of 1
3173 * and their unit amount is the total amount of space required.
3175 * The following lines of code account for non-transaction data
3176 * which occupy space in the on-disk log.
3178 * Normal form of a transaction is:
3179 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3180 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3182 * We need to account for all the leadup data and trailer data
3183 * around the transaction data.
3184 * And then we need to account for the worst case in terms of using
3186 * The worst case will happen if:
3187 * - the placement of the transaction happens to be such that the
3188 * roundoff is at its maximum
3189 * - the transaction data is synced before the commit record is synced
3190 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3191 * Therefore the commit record is in its own Log Record.
3192 * This can happen as the commit record is called with its
3193 * own region to xlog_write().
3194 * This then means that in the worst case, roundoff can happen for
3195 * the commit-rec as well.
3196 * The commit-rec is smaller than padding in this scenario and so it is
3197 * not added separately.
3200 /* for trans header */
3201 unit_bytes += sizeof(xlog_op_header_t);
3202 unit_bytes += sizeof(xfs_trans_header_t);
3205 unit_bytes += sizeof(xlog_op_header_t);
3208 * for LR headers - the space for data in an iclog is the size minus
3209 * the space used for the headers. If we use the iclog size, then we
3210 * undercalculate the number of headers required.
3212 * Furthermore - the addition of op headers for split-recs might
3213 * increase the space required enough to require more log and op
3214 * headers, so take that into account too.
3216 * IMPORTANT: This reservation makes the assumption that if this
3217 * transaction is the first in an iclog and hence has the LR headers
3218 * accounted to it, then the remaining space in the iclog is
3219 * exclusively for this transaction. i.e. if the transaction is larger
3220 * than the iclog, it will be the only thing in that iclog.
3221 * Fundamentally, this means we must pass the entire log vector to
3222 * xlog_write to guarantee this.
3224 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3225 num_headers = howmany(unit_bytes, iclog_space);
3227 /* for split-recs - ophdrs added when data split over LRs */
3228 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3230 /* add extra header reservations if we overrun */
3231 while (!num_headers ||
3232 howmany(unit_bytes, iclog_space) > num_headers) {
3233 unit_bytes += sizeof(xlog_op_header_t);
3236 unit_bytes += log->l_iclog_hsize * num_headers;
3238 /* for commit-rec LR header - note: padding will subsume the ophdr */
3239 unit_bytes += log->l_iclog_hsize;
3241 /* for roundoff padding for transaction data and one for commit record */
3242 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3243 log->l_mp->m_sb.sb_logsunit > 1) {
3244 /* log su roundoff */
3245 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3248 unit_bytes += 2*BBSIZE;
3251 atomic_set(&tic->t_ref, 1);
3252 tic->t_task = current;
3253 INIT_LIST_HEAD(&tic->t_queue);
3254 tic->t_unit_res = unit_bytes;
3255 tic->t_curr_res = unit_bytes;
3258 tic->t_tid = random32();
3259 tic->t_clientid = client;
3260 tic->t_flags = XLOG_TIC_INITED;
3261 tic->t_trans_type = 0;
3263 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3265 xlog_tic_reset_res(tic);
3271 /******************************************************************************
3273 * Log debug routines
3275 ******************************************************************************
3279 * Make sure that the destination ptr is within the valid data region of
3280 * one of the iclogs. This uses backup pointers stored in a different
3281 * part of the log in case we trash the log structure.
3284 xlog_verify_dest_ptr(
3291 for (i = 0; i < log->l_iclog_bufs; i++) {
3292 if (ptr >= log->l_iclog_bak[i] &&
3293 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3298 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3302 * Check to make sure the grant write head didn't just over lap the tail. If
3303 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3304 * the cycles differ by exactly one and check the byte count.
3306 * This check is run unlocked, so can give false positives. Rather than assert
3307 * on failures, use a warn-once flag and a panic tag to allow the admin to
3308 * determine if they want to panic the machine when such an error occurs. For
3309 * debug kernels this will have the same effect as using an assert but, unlinke
3310 * an assert, it can be turned off at runtime.
3313 xlog_verify_grant_tail(
3316 int tail_cycle, tail_blocks;
3319 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3320 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3321 if (tail_cycle != cycle) {
3322 if (cycle - 1 != tail_cycle &&
3323 !(log->l_flags & XLOG_TAIL_WARN)) {
3324 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3325 "%s: cycle - 1 != tail_cycle", __func__);
3326 log->l_flags |= XLOG_TAIL_WARN;
3329 if (space > BBTOB(tail_blocks) &&
3330 !(log->l_flags & XLOG_TAIL_WARN)) {
3331 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3332 "%s: space > BBTOB(tail_blocks)", __func__);
3333 log->l_flags |= XLOG_TAIL_WARN;
3338 /* check if it will fit */
3340 xlog_verify_tail_lsn(xlog_t *log,
3341 xlog_in_core_t *iclog,
3346 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3348 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3349 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3350 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3352 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3354 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3355 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3357 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3358 if (blocks < BTOBB(iclog->ic_offset) + 1)
3359 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3361 } /* xlog_verify_tail_lsn */
3364 * Perform a number of checks on the iclog before writing to disk.
3366 * 1. Make sure the iclogs are still circular
3367 * 2. Make sure we have a good magic number
3368 * 3. Make sure we don't have magic numbers in the data
3369 * 4. Check fields of each log operation header for:
3370 * A. Valid client identifier
3371 * B. tid ptr value falls in valid ptr space (user space code)
3372 * C. Length in log record header is correct according to the
3373 * individual operation headers within record.
3374 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3375 * log, check the preceding blocks of the physical log to make sure all
3376 * the cycle numbers agree with the current cycle number.
3379 xlog_verify_iclog(xlog_t *log,
3380 xlog_in_core_t *iclog,
3384 xlog_op_header_t *ophead;
3385 xlog_in_core_t *icptr;
3386 xlog_in_core_2_t *xhdr;
3388 xfs_caddr_t base_ptr;
3389 __psint_t field_offset;
3391 int len, i, j, k, op_len;
3394 /* check validity of iclog pointers */
3395 spin_lock(&log->l_icloglock);
3396 icptr = log->l_iclog;
3397 for (i=0; i < log->l_iclog_bufs; i++) {
3399 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3400 icptr = icptr->ic_next;
3402 if (icptr != log->l_iclog)
3403 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3404 spin_unlock(&log->l_icloglock);
3406 /* check log magic numbers */
3407 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3408 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3410 ptr = (xfs_caddr_t) &iclog->ic_header;
3411 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3413 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3414 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3419 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3420 ptr = iclog->ic_datap;
3422 ophead = (xlog_op_header_t *)ptr;
3423 xhdr = iclog->ic_data;
3424 for (i = 0; i < len; i++) {
3425 ophead = (xlog_op_header_t *)ptr;
3427 /* clientid is only 1 byte */
3428 field_offset = (__psint_t)
3429 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3430 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3431 clientid = ophead->oh_clientid;
3433 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3434 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3435 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3436 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3437 clientid = xlog_get_client_id(
3438 xhdr[j].hic_xheader.xh_cycle_data[k]);
3440 clientid = xlog_get_client_id(
3441 iclog->ic_header.h_cycle_data[idx]);
3444 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3446 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3447 __func__, clientid, ophead,
3448 (unsigned long)field_offset);
3451 field_offset = (__psint_t)
3452 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3453 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3454 op_len = be32_to_cpu(ophead->oh_len);
3456 idx = BTOBBT((__psint_t)&ophead->oh_len -
3457 (__psint_t)iclog->ic_datap);
3458 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3459 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3460 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3461 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3463 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3466 ptr += sizeof(xlog_op_header_t) + op_len;
3468 } /* xlog_verify_iclog */
3472 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3478 xlog_in_core_t *iclog, *ic;
3480 iclog = log->l_iclog;
3481 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3483 * Mark all the incore logs IOERROR.
3484 * From now on, no log flushes will result.
3488 ic->ic_state = XLOG_STATE_IOERROR;
3490 } while (ic != iclog);
3494 * Return non-zero, if state transition has already happened.
3500 * This is called from xfs_force_shutdown, when we're forcibly
3501 * shutting down the filesystem, typically because of an IO error.
3502 * Our main objectives here are to make sure that:
3503 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3504 * parties to find out, 'atomically'.
3505 * b. those who're sleeping on log reservations, pinned objects and
3506 * other resources get woken up, and be told the bad news.
3507 * c. nothing new gets queued up after (a) and (b) are done.
3508 * d. if !logerror, flush the iclogs to disk, then seal them off
3511 * Note: for delayed logging the !logerror case needs to flush the regions
3512 * held in memory out to the iclogs before flushing them to disk. This needs
3513 * to be done before the log is marked as shutdown, otherwise the flush to the
3517 xfs_log_force_umount(
3518 struct xfs_mount *mp,
3527 * If this happens during log recovery, don't worry about
3528 * locking; the log isn't open for business yet.
3531 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3532 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3534 XFS_BUF_DONE(mp->m_sb_bp);
3539 * Somebody could've already done the hard work for us.
3540 * No need to get locks for this.
3542 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3543 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3549 * Flush the in memory commit item list before marking the log as
3550 * being shut down. We need to do it in this order to ensure all the
3551 * completed transactions are flushed to disk with the xfs_log_force()
3555 xlog_cil_force(log);
3558 * mark the filesystem and the as in a shutdown state and wake
3559 * everybody up to tell them the bad news.
3561 spin_lock(&log->l_icloglock);
3562 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3564 XFS_BUF_DONE(mp->m_sb_bp);
3567 * This flag is sort of redundant because of the mount flag, but
3568 * it's good to maintain the separation between the log and the rest
3571 log->l_flags |= XLOG_IO_ERROR;
3574 * If we hit a log error, we want to mark all the iclogs IOERROR
3575 * while we're still holding the loglock.
3578 retval = xlog_state_ioerror(log);
3579 spin_unlock(&log->l_icloglock);
3582 * We don't want anybody waiting for log reservations after this. That
3583 * means we have to wake up everybody queued up on reserveq as well as
3584 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3585 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3586 * action is protected by the grant locks.
3588 xlog_grant_head_wake_all(&log->l_reserve_head);
3589 xlog_grant_head_wake_all(&log->l_write_head);
3591 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3594 * Force the incore logs to disk before shutting the
3595 * log down completely.
3597 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3599 spin_lock(&log->l_icloglock);
3600 retval = xlog_state_ioerror(log);
3601 spin_unlock(&log->l_icloglock);
3604 * Wake up everybody waiting on xfs_log_force.
3605 * Callback all log item committed functions as if the
3606 * log writes were completed.
3608 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3610 #ifdef XFSERRORDEBUG
3612 xlog_in_core_t *iclog;
3614 spin_lock(&log->l_icloglock);
3615 iclog = log->l_iclog;
3617 ASSERT(iclog->ic_callback == 0);
3618 iclog = iclog->ic_next;
3619 } while (iclog != log->l_iclog);
3620 spin_unlock(&log->l_icloglock);
3623 /* return non-zero if log IOERROR transition had already happened */
3628 xlog_iclogs_empty(xlog_t *log)
3630 xlog_in_core_t *iclog;
3632 iclog = log->l_iclog;
3634 /* endianness does not matter here, zero is zero in
3637 if (iclog->ic_header.h_num_logops)
3639 iclog = iclog->ic_next;
3640 } while (iclog != log->l_iclog);