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 */
731 error = xlog_write(log, &vec, tic, &lsn,
732 NULL, XLOG_UNMOUNT_TRANS);
734 * At this point, we're umounting anyway,
735 * so there's no point in transitioning log state
736 * to IOERROR. Just continue...
741 xfs_alert(mp, "%s: unmount record failed", __func__);
744 spin_lock(&log->l_icloglock);
745 iclog = log->l_iclog;
746 atomic_inc(&iclog->ic_refcnt);
747 xlog_state_want_sync(log, iclog);
748 spin_unlock(&log->l_icloglock);
749 error = xlog_state_release_iclog(log, iclog);
751 spin_lock(&log->l_icloglock);
752 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
753 iclog->ic_state == XLOG_STATE_DIRTY)) {
754 if (!XLOG_FORCED_SHUTDOWN(log)) {
755 xlog_wait(&iclog->ic_force_wait,
758 spin_unlock(&log->l_icloglock);
761 spin_unlock(&log->l_icloglock);
764 trace_xfs_log_umount_write(log, tic);
765 xlog_ungrant_log_space(log, tic);
766 xfs_log_ticket_put(tic);
770 * We're already in forced_shutdown mode, couldn't
771 * even attempt to write out the unmount transaction.
773 * Go through the motions of sync'ing and releasing
774 * the iclog, even though no I/O will actually happen,
775 * we need to wait for other log I/Os that may already
776 * be in progress. Do this as a separate section of
777 * code so we'll know if we ever get stuck here that
778 * we're in this odd situation of trying to unmount
779 * a file system that went into forced_shutdown as
780 * the result of an unmount..
782 spin_lock(&log->l_icloglock);
783 iclog = log->l_iclog;
784 atomic_inc(&iclog->ic_refcnt);
786 xlog_state_want_sync(log, iclog);
787 spin_unlock(&log->l_icloglock);
788 error = xlog_state_release_iclog(log, iclog);
790 spin_lock(&log->l_icloglock);
792 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
793 || iclog->ic_state == XLOG_STATE_DIRTY
794 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
796 xlog_wait(&iclog->ic_force_wait,
799 spin_unlock(&log->l_icloglock);
804 } /* xfs_log_unmount_write */
807 * Deallocate log structures for unmount/relocation.
809 * We need to stop the aild from running before we destroy
810 * and deallocate the log as the aild references the log.
813 xfs_log_unmount(xfs_mount_t *mp)
815 xfs_trans_ail_destroy(mp);
816 xlog_dealloc_log(mp->m_log);
821 struct xfs_mount *mp,
822 struct xfs_log_item *item,
824 const struct xfs_item_ops *ops)
826 item->li_mountp = mp;
827 item->li_ailp = mp->m_ail;
828 item->li_type = type;
832 INIT_LIST_HEAD(&item->li_ail);
833 INIT_LIST_HEAD(&item->li_cil);
837 * Wake up processes waiting for log space after we have moved the log tail.
841 struct xfs_mount *mp)
843 struct log *log = mp->m_log;
846 if (XLOG_FORCED_SHUTDOWN(log))
849 if (!list_empty_careful(&log->l_write_head.waiters)) {
850 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
852 spin_lock(&log->l_write_head.lock);
853 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
854 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
855 spin_unlock(&log->l_write_head.lock);
858 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
859 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
861 spin_lock(&log->l_reserve_head.lock);
862 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
863 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
864 spin_unlock(&log->l_reserve_head.lock);
869 * Determine if we have a transaction that has gone to disk
870 * that needs to be covered. To begin the transition to the idle state
871 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
872 * If we are then in a state where covering is needed, the caller is informed
873 * that dummy transactions are required to move the log into the idle state.
875 * Because this is called as part of the sync process, we should also indicate
876 * that dummy transactions should be issued in anything but the covered or
877 * idle states. This ensures that the log tail is accurately reflected in
878 * the log at the end of the sync, hence if a crash occurrs avoids replay
879 * of transactions where the metadata is already on disk.
882 xfs_log_need_covered(xfs_mount_t *mp)
885 xlog_t *log = mp->m_log;
887 if (!xfs_fs_writable(mp))
890 spin_lock(&log->l_icloglock);
891 switch (log->l_covered_state) {
892 case XLOG_STATE_COVER_DONE:
893 case XLOG_STATE_COVER_DONE2:
894 case XLOG_STATE_COVER_IDLE:
896 case XLOG_STATE_COVER_NEED:
897 case XLOG_STATE_COVER_NEED2:
898 if (!xfs_ail_min_lsn(log->l_ailp) &&
899 xlog_iclogs_empty(log)) {
900 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
901 log->l_covered_state = XLOG_STATE_COVER_DONE;
903 log->l_covered_state = XLOG_STATE_COVER_DONE2;
910 spin_unlock(&log->l_icloglock);
915 * We may be holding the log iclog lock upon entering this routine.
918 xlog_assign_tail_lsn(
919 struct xfs_mount *mp)
922 struct log *log = mp->m_log;
925 * To make sure we always have a valid LSN for the log tail we keep
926 * track of the last LSN which was committed in log->l_last_sync_lsn,
927 * and use that when the AIL was empty and xfs_ail_min_lsn returns 0.
929 * If the AIL has been emptied we also need to wake any process
930 * waiting for this condition.
932 tail_lsn = xfs_ail_min_lsn(mp->m_ail);
934 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
935 atomic64_set(&log->l_tail_lsn, tail_lsn);
940 * Return the space in the log between the tail and the head. The head
941 * is passed in the cycle/bytes formal parms. In the special case where
942 * the reserve head has wrapped passed the tail, this calculation is no
943 * longer valid. In this case, just return 0 which means there is no space
944 * in the log. This works for all places where this function is called
945 * with the reserve head. Of course, if the write head were to ever
946 * wrap the tail, we should blow up. Rather than catch this case here,
947 * we depend on other ASSERTions in other parts of the code. XXXmiken
949 * This code also handles the case where the reservation head is behind
950 * the tail. The details of this case are described below, but the end
951 * result is that we return the size of the log as the amount of space left.
964 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
965 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
966 tail_bytes = BBTOB(tail_bytes);
967 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
968 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
969 else if (tail_cycle + 1 < head_cycle)
971 else if (tail_cycle < head_cycle) {
972 ASSERT(tail_cycle == (head_cycle - 1));
973 free_bytes = tail_bytes - head_bytes;
976 * The reservation head is behind the tail.
977 * In this case we just want to return the size of the
978 * log as the amount of space left.
981 "xlog_space_left: head behind tail\n"
982 " tail_cycle = %d, tail_bytes = %d\n"
983 " GH cycle = %d, GH bytes = %d",
984 tail_cycle, tail_bytes, head_cycle, head_bytes);
986 free_bytes = log->l_logsize;
993 * Log function which is called when an io completes.
995 * The log manager needs its own routine, in order to control what
996 * happens with the buffer after the write completes.
999 xlog_iodone(xfs_buf_t *bp)
1001 xlog_in_core_t *iclog = bp->b_fspriv;
1002 xlog_t *l = iclog->ic_log;
1006 * Race to shutdown the filesystem if we see an error.
1008 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1009 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1010 xfs_buf_ioerror_alert(bp, __func__);
1012 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1014 * This flag will be propagated to the trans-committed
1015 * callback routines to let them know that the log-commit
1018 aborted = XFS_LI_ABORTED;
1019 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1020 aborted = XFS_LI_ABORTED;
1023 /* log I/O is always issued ASYNC */
1024 ASSERT(XFS_BUF_ISASYNC(bp));
1025 xlog_state_done_syncing(iclog, aborted);
1027 * do not reference the buffer (bp) here as we could race
1028 * with it being freed after writing the unmount record to the
1035 * Return size of each in-core log record buffer.
1037 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1039 * If the filesystem blocksize is too large, we may need to choose a
1040 * larger size since the directory code currently logs entire blocks.
1044 xlog_get_iclog_buffer_size(xfs_mount_t *mp,
1050 if (mp->m_logbufs <= 0)
1051 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1053 log->l_iclog_bufs = mp->m_logbufs;
1056 * Buffer size passed in from mount system call.
1058 if (mp->m_logbsize > 0) {
1059 size = log->l_iclog_size = mp->m_logbsize;
1060 log->l_iclog_size_log = 0;
1062 log->l_iclog_size_log++;
1066 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1067 /* # headers = size / 32k
1068 * one header holds cycles from 32k of data
1071 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1072 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1074 log->l_iclog_hsize = xhdrs << BBSHIFT;
1075 log->l_iclog_heads = xhdrs;
1077 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1078 log->l_iclog_hsize = BBSIZE;
1079 log->l_iclog_heads = 1;
1084 /* All machines use 32kB buffers by default. */
1085 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1086 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1088 /* the default log size is 16k or 32k which is one header sector */
1089 log->l_iclog_hsize = BBSIZE;
1090 log->l_iclog_heads = 1;
1093 /* are we being asked to make the sizes selected above visible? */
1094 if (mp->m_logbufs == 0)
1095 mp->m_logbufs = log->l_iclog_bufs;
1096 if (mp->m_logbsize == 0)
1097 mp->m_logbsize = log->l_iclog_size;
1098 } /* xlog_get_iclog_buffer_size */
1102 * This routine initializes some of the log structure for a given mount point.
1103 * Its primary purpose is to fill in enough, so recovery can occur. However,
1104 * some other stuff may be filled in too.
1107 xlog_alloc_log(xfs_mount_t *mp,
1108 xfs_buftarg_t *log_target,
1109 xfs_daddr_t blk_offset,
1113 xlog_rec_header_t *head;
1114 xlog_in_core_t **iclogp;
1115 xlog_in_core_t *iclog, *prev_iclog=NULL;
1121 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1123 xfs_warn(mp, "Log allocation failed: No memory!");
1128 log->l_targ = log_target;
1129 log->l_logsize = BBTOB(num_bblks);
1130 log->l_logBBstart = blk_offset;
1131 log->l_logBBsize = num_bblks;
1132 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1133 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1135 log->l_prev_block = -1;
1136 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1137 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1138 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1139 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1141 xlog_grant_head_init(&log->l_reserve_head);
1142 xlog_grant_head_init(&log->l_write_head);
1144 error = EFSCORRUPTED;
1145 if (xfs_sb_version_hassector(&mp->m_sb)) {
1146 log2_size = mp->m_sb.sb_logsectlog;
1147 if (log2_size < BBSHIFT) {
1148 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1149 log2_size, BBSHIFT);
1153 log2_size -= BBSHIFT;
1154 if (log2_size > mp->m_sectbb_log) {
1155 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1156 log2_size, mp->m_sectbb_log);
1160 /* for larger sector sizes, must have v2 or external log */
1161 if (log2_size && log->l_logBBstart > 0 &&
1162 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1164 "log sector size (0x%x) invalid for configuration.",
1169 log->l_sectBBsize = 1 << log2_size;
1171 xlog_get_iclog_buffer_size(mp, log);
1174 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, log->l_iclog_size, 0);
1177 bp->b_iodone = xlog_iodone;
1178 ASSERT(xfs_buf_islocked(bp));
1181 spin_lock_init(&log->l_icloglock);
1182 init_waitqueue_head(&log->l_flush_wait);
1184 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1185 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1187 iclogp = &log->l_iclog;
1189 * The amount of memory to allocate for the iclog structure is
1190 * rather funky due to the way the structure is defined. It is
1191 * done this way so that we can use different sizes for machines
1192 * with different amounts of memory. See the definition of
1193 * xlog_in_core_t in xfs_log_priv.h for details.
1195 ASSERT(log->l_iclog_size >= 4096);
1196 for (i=0; i < log->l_iclog_bufs; i++) {
1197 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1199 goto out_free_iclog;
1202 iclog->ic_prev = prev_iclog;
1205 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1206 log->l_iclog_size, 0);
1208 goto out_free_iclog;
1210 bp->b_iodone = xlog_iodone;
1212 iclog->ic_data = bp->b_addr;
1214 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1216 head = &iclog->ic_header;
1217 memset(head, 0, sizeof(xlog_rec_header_t));
1218 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1219 head->h_version = cpu_to_be32(
1220 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1221 head->h_size = cpu_to_be32(log->l_iclog_size);
1223 head->h_fmt = cpu_to_be32(XLOG_FMT);
1224 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1226 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1227 iclog->ic_state = XLOG_STATE_ACTIVE;
1228 iclog->ic_log = log;
1229 atomic_set(&iclog->ic_refcnt, 0);
1230 spin_lock_init(&iclog->ic_callback_lock);
1231 iclog->ic_callback_tail = &(iclog->ic_callback);
1232 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1234 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1235 init_waitqueue_head(&iclog->ic_force_wait);
1236 init_waitqueue_head(&iclog->ic_write_wait);
1238 iclogp = &iclog->ic_next;
1240 *iclogp = log->l_iclog; /* complete ring */
1241 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1243 error = xlog_cil_init(log);
1245 goto out_free_iclog;
1249 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1250 prev_iclog = iclog->ic_next;
1252 xfs_buf_free(iclog->ic_bp);
1255 spinlock_destroy(&log->l_icloglock);
1256 xfs_buf_free(log->l_xbuf);
1260 return ERR_PTR(-error);
1261 } /* xlog_alloc_log */
1265 * Write out the commit record of a transaction associated with the given
1266 * ticket. Return the lsn of the commit record.
1271 struct xlog_ticket *ticket,
1272 struct xlog_in_core **iclog,
1273 xfs_lsn_t *commitlsnp)
1275 struct xfs_mount *mp = log->l_mp;
1277 struct xfs_log_iovec reg = {
1280 .i_type = XLOG_REG_TYPE_COMMIT,
1282 struct xfs_log_vec vec = {
1287 ASSERT_ALWAYS(iclog);
1288 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1291 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1296 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1297 * log space. This code pushes on the lsn which would supposedly free up
1298 * the 25% which we want to leave free. We may need to adopt a policy which
1299 * pushes on an lsn which is further along in the log once we reach the high
1300 * water mark. In this manner, we would be creating a low water mark.
1303 xlog_grant_push_ail(
1307 xfs_lsn_t threshold_lsn = 0;
1308 xfs_lsn_t last_sync_lsn;
1311 int threshold_block;
1312 int threshold_cycle;
1315 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1317 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1318 free_blocks = BTOBBT(free_bytes);
1321 * Set the threshold for the minimum number of free blocks in the
1322 * log to the maximum of what the caller needs, one quarter of the
1323 * log, and 256 blocks.
1325 free_threshold = BTOBB(need_bytes);
1326 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1327 free_threshold = MAX(free_threshold, 256);
1328 if (free_blocks >= free_threshold)
1331 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1333 threshold_block += free_threshold;
1334 if (threshold_block >= log->l_logBBsize) {
1335 threshold_block -= log->l_logBBsize;
1336 threshold_cycle += 1;
1338 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1341 * Don't pass in an lsn greater than the lsn of the last
1342 * log record known to be on disk. Use a snapshot of the last sync lsn
1343 * so that it doesn't change between the compare and the set.
1345 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1346 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1347 threshold_lsn = last_sync_lsn;
1350 * Get the transaction layer to kick the dirty buffers out to
1351 * disk asynchronously. No point in trying to do this if
1352 * the filesystem is shutting down.
1354 if (!XLOG_FORCED_SHUTDOWN(log))
1355 xfs_ail_push(log->l_ailp, threshold_lsn);
1359 * The bdstrat callback function for log bufs. This gives us a central
1360 * place to trap bufs in case we get hit by a log I/O error and need to
1361 * shutdown. Actually, in practice, even when we didn't get a log error,
1362 * we transition the iclogs to IOERROR state *after* flushing all existing
1363 * iclogs to disk. This is because we don't want anymore new transactions to be
1364 * started or completed afterwards.
1370 struct xlog_in_core *iclog = bp->b_fspriv;
1372 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1373 xfs_buf_ioerror(bp, EIO);
1375 xfs_buf_ioend(bp, 0);
1377 * It would seem logical to return EIO here, but we rely on
1378 * the log state machine to propagate I/O errors instead of
1384 xfs_buf_iorequest(bp);
1389 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1390 * fashion. Previously, we should have moved the current iclog
1391 * ptr in the log to point to the next available iclog. This allows further
1392 * write to continue while this code syncs out an iclog ready to go.
1393 * Before an in-core log can be written out, the data section must be scanned
1394 * to save away the 1st word of each BBSIZE block into the header. We replace
1395 * it with the current cycle count. Each BBSIZE block is tagged with the
1396 * cycle count because there in an implicit assumption that drives will
1397 * guarantee that entire 512 byte blocks get written at once. In other words,
1398 * we can't have part of a 512 byte block written and part not written. By
1399 * tagging each block, we will know which blocks are valid when recovering
1400 * after an unclean shutdown.
1402 * This routine is single threaded on the iclog. No other thread can be in
1403 * this routine with the same iclog. Changing contents of iclog can there-
1404 * fore be done without grabbing the state machine lock. Updating the global
1405 * log will require grabbing the lock though.
1407 * The entire log manager uses a logical block numbering scheme. Only
1408 * log_sync (and then only bwrite()) know about the fact that the log may
1409 * not start with block zero on a given device. The log block start offset
1410 * is added immediately before calling bwrite().
1414 xlog_sync(xlog_t *log,
1415 xlog_in_core_t *iclog)
1417 xfs_caddr_t dptr; /* pointer to byte sized element */
1420 uint count; /* byte count of bwrite */
1421 uint count_init; /* initial count before roundup */
1422 int roundoff; /* roundoff to BB or stripe */
1423 int split = 0; /* split write into two regions */
1425 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1427 XFS_STATS_INC(xs_log_writes);
1428 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1430 /* Add for LR header */
1431 count_init = log->l_iclog_hsize + iclog->ic_offset;
1433 /* Round out the log write size */
1434 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1435 /* we have a v2 stripe unit to use */
1436 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1438 count = BBTOB(BTOBB(count_init));
1440 roundoff = count - count_init;
1441 ASSERT(roundoff >= 0);
1442 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1443 roundoff < log->l_mp->m_sb.sb_logsunit)
1445 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1446 roundoff < BBTOB(1)));
1448 /* move grant heads by roundoff in sync */
1449 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1450 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1452 /* put cycle number in every block */
1453 xlog_pack_data(log, iclog, roundoff);
1455 /* real byte length */
1457 iclog->ic_header.h_len =
1458 cpu_to_be32(iclog->ic_offset + roundoff);
1460 iclog->ic_header.h_len =
1461 cpu_to_be32(iclog->ic_offset);
1465 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1467 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1469 /* Do we need to split this write into 2 parts? */
1470 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1471 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1472 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1473 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1475 iclog->ic_bwritecnt = 1;
1477 XFS_BUF_SET_COUNT(bp, count);
1478 bp->b_fspriv = iclog;
1479 XFS_BUF_ZEROFLAGS(bp);
1481 bp->b_flags |= XBF_SYNCIO;
1483 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1484 bp->b_flags |= XBF_FUA;
1487 * Flush the data device before flushing the log to make
1488 * sure all meta data written back from the AIL actually made
1489 * it to disk before stamping the new log tail LSN into the
1490 * log buffer. For an external log we need to issue the
1491 * flush explicitly, and unfortunately synchronously here;
1492 * for an internal log we can simply use the block layer
1493 * state machine for preflushes.
1495 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1496 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1498 bp->b_flags |= XBF_FLUSH;
1501 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1502 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1504 xlog_verify_iclog(log, iclog, count, B_TRUE);
1506 /* account for log which doesn't start at block #0 */
1507 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1509 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1514 error = xlog_bdstrat(bp);
1516 xfs_buf_ioerror_alert(bp, "xlog_sync");
1520 bp = iclog->ic_log->l_xbuf;
1521 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1522 xfs_buf_associate_memory(bp,
1523 (char *)&iclog->ic_header + count, split);
1524 bp->b_fspriv = iclog;
1525 XFS_BUF_ZEROFLAGS(bp);
1527 bp->b_flags |= XBF_SYNCIO;
1528 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1529 bp->b_flags |= XBF_FUA;
1532 * Bump the cycle numbers at the start of each block
1533 * since this part of the buffer is at the start of
1534 * a new cycle. Watch out for the header magic number
1537 for (i = 0; i < split; i += BBSIZE) {
1538 be32_add_cpu((__be32 *)dptr, 1);
1539 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1540 be32_add_cpu((__be32 *)dptr, 1);
1544 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1545 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1547 /* account for internal log which doesn't start at block #0 */
1548 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1550 error = xlog_bdstrat(bp);
1552 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1561 * Deallocate a log structure
1564 xlog_dealloc_log(xlog_t *log)
1566 xlog_in_core_t *iclog, *next_iclog;
1569 xlog_cil_destroy(log);
1572 * always need to ensure that the extra buffer does not point to memory
1573 * owned by another log buffer before we free it.
1575 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size);
1576 xfs_buf_free(log->l_xbuf);
1578 iclog = log->l_iclog;
1579 for (i=0; i<log->l_iclog_bufs; i++) {
1580 xfs_buf_free(iclog->ic_bp);
1581 next_iclog = iclog->ic_next;
1585 spinlock_destroy(&log->l_icloglock);
1587 log->l_mp->m_log = NULL;
1589 } /* xlog_dealloc_log */
1592 * Update counters atomically now that memcpy is done.
1596 xlog_state_finish_copy(xlog_t *log,
1597 xlog_in_core_t *iclog,
1601 spin_lock(&log->l_icloglock);
1603 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1604 iclog->ic_offset += copy_bytes;
1606 spin_unlock(&log->l_icloglock);
1607 } /* xlog_state_finish_copy */
1613 * print out info relating to regions written which consume
1618 struct xfs_mount *mp,
1619 struct xlog_ticket *ticket)
1622 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1624 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1625 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1646 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1690 "xlog_write: reservation summary:\n"
1691 " trans type = %s (%u)\n"
1692 " unit res = %d bytes\n"
1693 " current res = %d bytes\n"
1694 " total reg = %u bytes (o/flow = %u bytes)\n"
1695 " ophdrs = %u (ophdr space = %u bytes)\n"
1696 " ophdr + reg = %u bytes\n"
1697 " num regions = %u\n",
1698 ((ticket->t_trans_type <= 0 ||
1699 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1700 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1701 ticket->t_trans_type,
1704 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1705 ticket->t_res_num_ophdrs, ophdr_spc,
1706 ticket->t_res_arr_sum +
1707 ticket->t_res_o_flow + ophdr_spc,
1710 for (i = 0; i < ticket->t_res_num; i++) {
1711 uint r_type = ticket->t_res_arr[i].r_type;
1712 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1713 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1714 "bad-rtype" : res_type_str[r_type-1]),
1715 ticket->t_res_arr[i].r_len);
1718 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1719 "xlog_write: reservation ran out. Need to up reservation");
1720 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1724 * Calculate the potential space needed by the log vector. Each region gets
1725 * its own xlog_op_header_t and may need to be double word aligned.
1728 xlog_write_calc_vec_length(
1729 struct xlog_ticket *ticket,
1730 struct xfs_log_vec *log_vector)
1732 struct xfs_log_vec *lv;
1737 /* acct for start rec of xact */
1738 if (ticket->t_flags & XLOG_TIC_INITED)
1741 for (lv = log_vector; lv; lv = lv->lv_next) {
1742 headers += lv->lv_niovecs;
1744 for (i = 0; i < lv->lv_niovecs; i++) {
1745 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1748 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1752 ticket->t_res_num_ophdrs += headers;
1753 len += headers * sizeof(struct xlog_op_header);
1759 * If first write for transaction, insert start record We can't be trying to
1760 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1763 xlog_write_start_rec(
1764 struct xlog_op_header *ophdr,
1765 struct xlog_ticket *ticket)
1767 if (!(ticket->t_flags & XLOG_TIC_INITED))
1770 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1771 ophdr->oh_clientid = ticket->t_clientid;
1773 ophdr->oh_flags = XLOG_START_TRANS;
1776 ticket->t_flags &= ~XLOG_TIC_INITED;
1778 return sizeof(struct xlog_op_header);
1781 static xlog_op_header_t *
1782 xlog_write_setup_ophdr(
1784 struct xlog_op_header *ophdr,
1785 struct xlog_ticket *ticket,
1788 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1789 ophdr->oh_clientid = ticket->t_clientid;
1792 /* are we copying a commit or unmount record? */
1793 ophdr->oh_flags = flags;
1796 * We've seen logs corrupted with bad transaction client ids. This
1797 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1798 * and shut down the filesystem.
1800 switch (ophdr->oh_clientid) {
1801 case XFS_TRANSACTION:
1807 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1808 ophdr->oh_clientid, ticket);
1816 * Set up the parameters of the region copy into the log. This has
1817 * to handle region write split across multiple log buffers - this
1818 * state is kept external to this function so that this code can
1819 * can be written in an obvious, self documenting manner.
1822 xlog_write_setup_copy(
1823 struct xlog_ticket *ticket,
1824 struct xlog_op_header *ophdr,
1825 int space_available,
1829 int *last_was_partial_copy,
1830 int *bytes_consumed)
1834 still_to_copy = space_required - *bytes_consumed;
1835 *copy_off = *bytes_consumed;
1837 if (still_to_copy <= space_available) {
1838 /* write of region completes here */
1839 *copy_len = still_to_copy;
1840 ophdr->oh_len = cpu_to_be32(*copy_len);
1841 if (*last_was_partial_copy)
1842 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1843 *last_was_partial_copy = 0;
1844 *bytes_consumed = 0;
1848 /* partial write of region, needs extra log op header reservation */
1849 *copy_len = space_available;
1850 ophdr->oh_len = cpu_to_be32(*copy_len);
1851 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1852 if (*last_was_partial_copy)
1853 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1854 *bytes_consumed += *copy_len;
1855 (*last_was_partial_copy)++;
1857 /* account for new log op header */
1858 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1859 ticket->t_res_num_ophdrs++;
1861 return sizeof(struct xlog_op_header);
1865 xlog_write_copy_finish(
1867 struct xlog_in_core *iclog,
1872 int *partial_copy_len,
1874 struct xlog_in_core **commit_iclog)
1876 if (*partial_copy) {
1878 * This iclog has already been marked WANT_SYNC by
1879 * xlog_state_get_iclog_space.
1881 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1884 return xlog_state_release_iclog(log, iclog);
1888 *partial_copy_len = 0;
1890 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1891 /* no more space in this iclog - push it. */
1892 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1896 spin_lock(&log->l_icloglock);
1897 xlog_state_want_sync(log, iclog);
1898 spin_unlock(&log->l_icloglock);
1901 return xlog_state_release_iclog(log, iclog);
1902 ASSERT(flags & XLOG_COMMIT_TRANS);
1903 *commit_iclog = iclog;
1910 * Write some region out to in-core log
1912 * This will be called when writing externally provided regions or when
1913 * writing out a commit record for a given transaction.
1915 * General algorithm:
1916 * 1. Find total length of this write. This may include adding to the
1917 * lengths passed in.
1918 * 2. Check whether we violate the tickets reservation.
1919 * 3. While writing to this iclog
1920 * A. Reserve as much space in this iclog as can get
1921 * B. If this is first write, save away start lsn
1922 * C. While writing this region:
1923 * 1. If first write of transaction, write start record
1924 * 2. Write log operation header (header per region)
1925 * 3. Find out if we can fit entire region into this iclog
1926 * 4. Potentially, verify destination memcpy ptr
1927 * 5. Memcpy (partial) region
1928 * 6. If partial copy, release iclog; otherwise, continue
1929 * copying more regions into current iclog
1930 * 4. Mark want sync bit (in simulation mode)
1931 * 5. Release iclog for potential flush to on-disk log.
1934 * 1. Panic if reservation is overrun. This should never happen since
1935 * reservation amounts are generated internal to the filesystem.
1937 * 1. Tickets are single threaded data structures.
1938 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1939 * syncing routine. When a single log_write region needs to span
1940 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1941 * on all log operation writes which don't contain the end of the
1942 * region. The XLOG_END_TRANS bit is used for the in-core log
1943 * operation which contains the end of the continued log_write region.
1944 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1945 * we don't really know exactly how much space will be used. As a result,
1946 * we don't update ic_offset until the end when we know exactly how many
1947 * bytes have been written out.
1952 struct xfs_log_vec *log_vector,
1953 struct xlog_ticket *ticket,
1954 xfs_lsn_t *start_lsn,
1955 struct xlog_in_core **commit_iclog,
1958 struct xlog_in_core *iclog = NULL;
1959 struct xfs_log_iovec *vecp;
1960 struct xfs_log_vec *lv;
1963 int partial_copy = 0;
1964 int partial_copy_len = 0;
1972 len = xlog_write_calc_vec_length(ticket, log_vector);
1975 * Region headers and bytes are already accounted for.
1976 * We only need to take into account start records and
1977 * split regions in this function.
1979 if (ticket->t_flags & XLOG_TIC_INITED)
1980 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1983 * Commit record headers need to be accounted for. These
1984 * come in as separate writes so are easy to detect.
1986 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1987 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1989 if (ticket->t_curr_res < 0)
1990 xlog_print_tic_res(log->l_mp, ticket);
1994 vecp = lv->lv_iovecp;
1995 while (lv && index < lv->lv_niovecs) {
1999 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2000 &contwr, &log_offset);
2004 ASSERT(log_offset <= iclog->ic_size - 1);
2005 ptr = iclog->ic_datap + log_offset;
2007 /* start_lsn is the first lsn written to. That's all we need. */
2009 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2012 * This loop writes out as many regions as can fit in the amount
2013 * of space which was allocated by xlog_state_get_iclog_space().
2015 while (lv && index < lv->lv_niovecs) {
2016 struct xfs_log_iovec *reg = &vecp[index];
2017 struct xlog_op_header *ophdr;
2022 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2023 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2025 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2026 if (start_rec_copy) {
2028 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2032 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2034 return XFS_ERROR(EIO);
2036 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2037 sizeof(struct xlog_op_header));
2039 len += xlog_write_setup_copy(ticket, ophdr,
2040 iclog->ic_size-log_offset,
2042 ©_off, ©_len,
2045 xlog_verify_dest_ptr(log, ptr);
2048 ASSERT(copy_len >= 0);
2049 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2050 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2052 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2054 data_cnt += contwr ? copy_len : 0;
2056 error = xlog_write_copy_finish(log, iclog, flags,
2057 &record_cnt, &data_cnt,
2066 * if we had a partial copy, we need to get more iclog
2067 * space but we don't want to increment the region
2068 * index because there is still more is this region to
2071 * If we completed writing this region, and we flushed
2072 * the iclog (indicated by resetting of the record
2073 * count), then we also need to get more log space. If
2074 * this was the last record, though, we are done and
2080 if (++index == lv->lv_niovecs) {
2084 vecp = lv->lv_iovecp;
2086 if (record_cnt == 0) {
2096 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2098 return xlog_state_release_iclog(log, iclog);
2100 ASSERT(flags & XLOG_COMMIT_TRANS);
2101 *commit_iclog = iclog;
2106 /*****************************************************************************
2108 * State Machine functions
2110 *****************************************************************************
2113 /* Clean iclogs starting from the head. This ordering must be
2114 * maintained, so an iclog doesn't become ACTIVE beyond one that
2115 * is SYNCING. This is also required to maintain the notion that we use
2116 * a ordered wait queue to hold off would be writers to the log when every
2117 * iclog is trying to sync to disk.
2119 * State Change: DIRTY -> ACTIVE
2122 xlog_state_clean_log(xlog_t *log)
2124 xlog_in_core_t *iclog;
2127 iclog = log->l_iclog;
2129 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2130 iclog->ic_state = XLOG_STATE_ACTIVE;
2131 iclog->ic_offset = 0;
2132 ASSERT(iclog->ic_callback == NULL);
2134 * If the number of ops in this iclog indicate it just
2135 * contains the dummy transaction, we can
2136 * change state into IDLE (the second time around).
2137 * Otherwise we should change the state into
2139 * We don't need to cover the dummy.
2142 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2147 * We have two dirty iclogs so start over
2148 * This could also be num of ops indicates
2149 * this is not the dummy going out.
2153 iclog->ic_header.h_num_logops = 0;
2154 memset(iclog->ic_header.h_cycle_data, 0,
2155 sizeof(iclog->ic_header.h_cycle_data));
2156 iclog->ic_header.h_lsn = 0;
2157 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2160 break; /* stop cleaning */
2161 iclog = iclog->ic_next;
2162 } while (iclog != log->l_iclog);
2164 /* log is locked when we are called */
2166 * Change state for the dummy log recording.
2167 * We usually go to NEED. But we go to NEED2 if the changed indicates
2168 * we are done writing the dummy record.
2169 * If we are done with the second dummy recored (DONE2), then
2173 switch (log->l_covered_state) {
2174 case XLOG_STATE_COVER_IDLE:
2175 case XLOG_STATE_COVER_NEED:
2176 case XLOG_STATE_COVER_NEED2:
2177 log->l_covered_state = XLOG_STATE_COVER_NEED;
2180 case XLOG_STATE_COVER_DONE:
2182 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2184 log->l_covered_state = XLOG_STATE_COVER_NEED;
2187 case XLOG_STATE_COVER_DONE2:
2189 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2191 log->l_covered_state = XLOG_STATE_COVER_NEED;
2198 } /* xlog_state_clean_log */
2201 xlog_get_lowest_lsn(
2204 xlog_in_core_t *lsn_log;
2205 xfs_lsn_t lowest_lsn, lsn;
2207 lsn_log = log->l_iclog;
2210 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2211 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2212 if ((lsn && !lowest_lsn) ||
2213 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2217 lsn_log = lsn_log->ic_next;
2218 } while (lsn_log != log->l_iclog);
2224 xlog_state_do_callback(
2227 xlog_in_core_t *ciclog)
2229 xlog_in_core_t *iclog;
2230 xlog_in_core_t *first_iclog; /* used to know when we've
2231 * processed all iclogs once */
2232 xfs_log_callback_t *cb, *cb_next;
2234 xfs_lsn_t lowest_lsn;
2235 int ioerrors; /* counter: iclogs with errors */
2236 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2237 int funcdidcallbacks; /* flag: function did callbacks */
2238 int repeats; /* for issuing console warnings if
2239 * looping too many times */
2242 spin_lock(&log->l_icloglock);
2243 first_iclog = iclog = log->l_iclog;
2245 funcdidcallbacks = 0;
2250 * Scan all iclogs starting with the one pointed to by the
2251 * log. Reset this starting point each time the log is
2252 * unlocked (during callbacks).
2254 * Keep looping through iclogs until one full pass is made
2255 * without running any callbacks.
2257 first_iclog = log->l_iclog;
2258 iclog = log->l_iclog;
2259 loopdidcallbacks = 0;
2264 /* skip all iclogs in the ACTIVE & DIRTY states */
2265 if (iclog->ic_state &
2266 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2267 iclog = iclog->ic_next;
2272 * Between marking a filesystem SHUTDOWN and stopping
2273 * the log, we do flush all iclogs to disk (if there
2274 * wasn't a log I/O error). So, we do want things to
2275 * go smoothly in case of just a SHUTDOWN w/o a
2278 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2280 * Can only perform callbacks in order. Since
2281 * this iclog is not in the DONE_SYNC/
2282 * DO_CALLBACK state, we skip the rest and
2283 * just try to clean up. If we set our iclog
2284 * to DO_CALLBACK, we will not process it when
2285 * we retry since a previous iclog is in the
2286 * CALLBACK and the state cannot change since
2287 * we are holding the l_icloglock.
2289 if (!(iclog->ic_state &
2290 (XLOG_STATE_DONE_SYNC |
2291 XLOG_STATE_DO_CALLBACK))) {
2292 if (ciclog && (ciclog->ic_state ==
2293 XLOG_STATE_DONE_SYNC)) {
2294 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2299 * We now have an iclog that is in either the
2300 * DO_CALLBACK or DONE_SYNC states. The other
2301 * states (WANT_SYNC, SYNCING, or CALLBACK were
2302 * caught by the above if and are going to
2303 * clean (i.e. we aren't doing their callbacks)
2308 * We will do one more check here to see if we
2309 * have chased our tail around.
2312 lowest_lsn = xlog_get_lowest_lsn(log);
2314 XFS_LSN_CMP(lowest_lsn,
2315 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2316 iclog = iclog->ic_next;
2317 continue; /* Leave this iclog for
2321 iclog->ic_state = XLOG_STATE_CALLBACK;
2325 * update the last_sync_lsn before we drop the
2326 * icloglock to ensure we are the only one that
2329 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2330 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2331 atomic64_set(&log->l_last_sync_lsn,
2332 be64_to_cpu(iclog->ic_header.h_lsn));
2337 spin_unlock(&log->l_icloglock);
2340 * Keep processing entries in the callback list until
2341 * we come around and it is empty. We need to
2342 * atomically see that the list is empty and change the
2343 * state to DIRTY so that we don't miss any more
2344 * callbacks being added.
2346 spin_lock(&iclog->ic_callback_lock);
2347 cb = iclog->ic_callback;
2349 iclog->ic_callback_tail = &(iclog->ic_callback);
2350 iclog->ic_callback = NULL;
2351 spin_unlock(&iclog->ic_callback_lock);
2353 /* perform callbacks in the order given */
2354 for (; cb; cb = cb_next) {
2355 cb_next = cb->cb_next;
2356 cb->cb_func(cb->cb_arg, aborted);
2358 spin_lock(&iclog->ic_callback_lock);
2359 cb = iclog->ic_callback;
2365 spin_lock(&log->l_icloglock);
2366 ASSERT(iclog->ic_callback == NULL);
2367 spin_unlock(&iclog->ic_callback_lock);
2368 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2369 iclog->ic_state = XLOG_STATE_DIRTY;
2372 * Transition from DIRTY to ACTIVE if applicable.
2373 * NOP if STATE_IOERROR.
2375 xlog_state_clean_log(log);
2377 /* wake up threads waiting in xfs_log_force() */
2378 wake_up_all(&iclog->ic_force_wait);
2380 iclog = iclog->ic_next;
2381 } while (first_iclog != iclog);
2383 if (repeats > 5000) {
2384 flushcnt += repeats;
2387 "%s: possible infinite loop (%d iterations)",
2388 __func__, flushcnt);
2390 } while (!ioerrors && loopdidcallbacks);
2393 * make one last gasp attempt to see if iclogs are being left in
2397 if (funcdidcallbacks) {
2398 first_iclog = iclog = log->l_iclog;
2400 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2402 * Terminate the loop if iclogs are found in states
2403 * which will cause other threads to clean up iclogs.
2405 * SYNCING - i/o completion will go through logs
2406 * DONE_SYNC - interrupt thread should be waiting for
2408 * IOERROR - give up hope all ye who enter here
2410 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2411 iclog->ic_state == XLOG_STATE_SYNCING ||
2412 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2413 iclog->ic_state == XLOG_STATE_IOERROR )
2415 iclog = iclog->ic_next;
2416 } while (first_iclog != iclog);
2420 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2422 spin_unlock(&log->l_icloglock);
2425 wake_up_all(&log->l_flush_wait);
2430 * Finish transitioning this iclog to the dirty state.
2432 * Make sure that we completely execute this routine only when this is
2433 * the last call to the iclog. There is a good chance that iclog flushes,
2434 * when we reach the end of the physical log, get turned into 2 separate
2435 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2436 * routine. By using the reference count bwritecnt, we guarantee that only
2437 * the second completion goes through.
2439 * Callbacks could take time, so they are done outside the scope of the
2440 * global state machine log lock.
2443 xlog_state_done_syncing(
2444 xlog_in_core_t *iclog,
2447 xlog_t *log = iclog->ic_log;
2449 spin_lock(&log->l_icloglock);
2451 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2452 iclog->ic_state == XLOG_STATE_IOERROR);
2453 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2454 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2458 * If we got an error, either on the first buffer, or in the case of
2459 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2460 * and none should ever be attempted to be written to disk
2463 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2464 if (--iclog->ic_bwritecnt == 1) {
2465 spin_unlock(&log->l_icloglock);
2468 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2472 * Someone could be sleeping prior to writing out the next
2473 * iclog buffer, we wake them all, one will get to do the
2474 * I/O, the others get to wait for the result.
2476 wake_up_all(&iclog->ic_write_wait);
2477 spin_unlock(&log->l_icloglock);
2478 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2479 } /* xlog_state_done_syncing */
2483 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2484 * sleep. We wait on the flush queue on the head iclog as that should be
2485 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2486 * we will wait here and all new writes will sleep until a sync completes.
2488 * The in-core logs are used in a circular fashion. They are not used
2489 * out-of-order even when an iclog past the head is free.
2492 * * log_offset where xlog_write() can start writing into the in-core
2494 * * in-core log pointer to which xlog_write() should write.
2495 * * boolean indicating this is a continued write to an in-core log.
2496 * If this is the last write, then the in-core log's offset field
2497 * needs to be incremented, depending on the amount of data which
2501 xlog_state_get_iclog_space(xlog_t *log,
2503 xlog_in_core_t **iclogp,
2504 xlog_ticket_t *ticket,
2505 int *continued_write,
2509 xlog_rec_header_t *head;
2510 xlog_in_core_t *iclog;
2514 spin_lock(&log->l_icloglock);
2515 if (XLOG_FORCED_SHUTDOWN(log)) {
2516 spin_unlock(&log->l_icloglock);
2517 return XFS_ERROR(EIO);
2520 iclog = log->l_iclog;
2521 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2522 XFS_STATS_INC(xs_log_noiclogs);
2524 /* Wait for log writes to have flushed */
2525 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2529 head = &iclog->ic_header;
2531 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2532 log_offset = iclog->ic_offset;
2534 /* On the 1st write to an iclog, figure out lsn. This works
2535 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2536 * committing to. If the offset is set, that's how many blocks
2539 if (log_offset == 0) {
2540 ticket->t_curr_res -= log->l_iclog_hsize;
2541 xlog_tic_add_region(ticket,
2543 XLOG_REG_TYPE_LRHEADER);
2544 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2545 head->h_lsn = cpu_to_be64(
2546 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2547 ASSERT(log->l_curr_block >= 0);
2550 /* If there is enough room to write everything, then do it. Otherwise,
2551 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2552 * bit is on, so this will get flushed out. Don't update ic_offset
2553 * until you know exactly how many bytes get copied. Therefore, wait
2554 * until later to update ic_offset.
2556 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2557 * can fit into remaining data section.
2559 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2560 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2563 * If I'm the only one writing to this iclog, sync it to disk.
2564 * We need to do an atomic compare and decrement here to avoid
2565 * racing with concurrent atomic_dec_and_lock() calls in
2566 * xlog_state_release_iclog() when there is more than one
2567 * reference to the iclog.
2569 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2570 /* we are the only one */
2571 spin_unlock(&log->l_icloglock);
2572 error = xlog_state_release_iclog(log, iclog);
2576 spin_unlock(&log->l_icloglock);
2581 /* Do we have enough room to write the full amount in the remainder
2582 * of this iclog? Or must we continue a write on the next iclog and
2583 * mark this iclog as completely taken? In the case where we switch
2584 * iclogs (to mark it taken), this particular iclog will release/sync
2585 * to disk in xlog_write().
2587 if (len <= iclog->ic_size - iclog->ic_offset) {
2588 *continued_write = 0;
2589 iclog->ic_offset += len;
2591 *continued_write = 1;
2592 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2596 ASSERT(iclog->ic_offset <= iclog->ic_size);
2597 spin_unlock(&log->l_icloglock);
2599 *logoffsetp = log_offset;
2601 } /* xlog_state_get_iclog_space */
2603 /* The first cnt-1 times through here we don't need to
2604 * move the grant write head because the permanent
2605 * reservation has reserved cnt times the unit amount.
2606 * Release part of current permanent unit reservation and
2607 * reset current reservation to be one units worth. Also
2608 * move grant reservation head forward.
2611 xlog_regrant_reserve_log_space(xlog_t *log,
2612 xlog_ticket_t *ticket)
2614 trace_xfs_log_regrant_reserve_enter(log, ticket);
2616 if (ticket->t_cnt > 0)
2619 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2620 ticket->t_curr_res);
2621 xlog_grant_sub_space(log, &log->l_write_head.grant,
2622 ticket->t_curr_res);
2623 ticket->t_curr_res = ticket->t_unit_res;
2624 xlog_tic_reset_res(ticket);
2626 trace_xfs_log_regrant_reserve_sub(log, ticket);
2628 /* just return if we still have some of the pre-reserved space */
2629 if (ticket->t_cnt > 0)
2632 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2633 ticket->t_unit_res);
2635 trace_xfs_log_regrant_reserve_exit(log, ticket);
2637 ticket->t_curr_res = ticket->t_unit_res;
2638 xlog_tic_reset_res(ticket);
2639 } /* xlog_regrant_reserve_log_space */
2643 * Give back the space left from a reservation.
2645 * All the information we need to make a correct determination of space left
2646 * is present. For non-permanent reservations, things are quite easy. The
2647 * count should have been decremented to zero. We only need to deal with the
2648 * space remaining in the current reservation part of the ticket. If the
2649 * ticket contains a permanent reservation, there may be left over space which
2650 * needs to be released. A count of N means that N-1 refills of the current
2651 * reservation can be done before we need to ask for more space. The first
2652 * one goes to fill up the first current reservation. Once we run out of
2653 * space, the count will stay at zero and the only space remaining will be
2654 * in the current reservation field.
2657 xlog_ungrant_log_space(xlog_t *log,
2658 xlog_ticket_t *ticket)
2662 if (ticket->t_cnt > 0)
2665 trace_xfs_log_ungrant_enter(log, ticket);
2666 trace_xfs_log_ungrant_sub(log, ticket);
2669 * If this is a permanent reservation ticket, we may be able to free
2670 * up more space based on the remaining count.
2672 bytes = ticket->t_curr_res;
2673 if (ticket->t_cnt > 0) {
2674 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2675 bytes += ticket->t_unit_res*ticket->t_cnt;
2678 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2679 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2681 trace_xfs_log_ungrant_exit(log, ticket);
2683 xfs_log_space_wake(log->l_mp);
2687 * Flush iclog to disk if this is the last reference to the given iclog and
2688 * the WANT_SYNC bit is set.
2690 * When this function is entered, the iclog is not necessarily in the
2691 * WANT_SYNC state. It may be sitting around waiting to get filled.
2696 xlog_state_release_iclog(
2698 xlog_in_core_t *iclog)
2700 int sync = 0; /* do we sync? */
2702 if (iclog->ic_state & XLOG_STATE_IOERROR)
2703 return XFS_ERROR(EIO);
2705 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2706 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2709 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2710 spin_unlock(&log->l_icloglock);
2711 return XFS_ERROR(EIO);
2713 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2714 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2716 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2717 /* update tail before writing to iclog */
2718 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2720 iclog->ic_state = XLOG_STATE_SYNCING;
2721 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2722 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2723 /* cycle incremented when incrementing curr_block */
2725 spin_unlock(&log->l_icloglock);
2728 * We let the log lock go, so it's possible that we hit a log I/O
2729 * error or some other SHUTDOWN condition that marks the iclog
2730 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2731 * this iclog has consistent data, so we ignore IOERROR
2732 * flags after this point.
2735 return xlog_sync(log, iclog);
2737 } /* xlog_state_release_iclog */
2741 * This routine will mark the current iclog in the ring as WANT_SYNC
2742 * and move the current iclog pointer to the next iclog in the ring.
2743 * When this routine is called from xlog_state_get_iclog_space(), the
2744 * exact size of the iclog has not yet been determined. All we know is
2745 * that every data block. We have run out of space in this log record.
2748 xlog_state_switch_iclogs(xlog_t *log,
2749 xlog_in_core_t *iclog,
2752 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2754 eventual_size = iclog->ic_offset;
2755 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2756 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2757 log->l_prev_block = log->l_curr_block;
2758 log->l_prev_cycle = log->l_curr_cycle;
2760 /* roll log?: ic_offset changed later */
2761 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2763 /* Round up to next log-sunit */
2764 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2765 log->l_mp->m_sb.sb_logsunit > 1) {
2766 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2767 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2770 if (log->l_curr_block >= log->l_logBBsize) {
2771 log->l_curr_cycle++;
2772 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2773 log->l_curr_cycle++;
2774 log->l_curr_block -= log->l_logBBsize;
2775 ASSERT(log->l_curr_block >= 0);
2777 ASSERT(iclog == log->l_iclog);
2778 log->l_iclog = iclog->ic_next;
2779 } /* xlog_state_switch_iclogs */
2782 * Write out all data in the in-core log as of this exact moment in time.
2784 * Data may be written to the in-core log during this call. However,
2785 * we don't guarantee this data will be written out. A change from past
2786 * implementation means this routine will *not* write out zero length LRs.
2788 * Basically, we try and perform an intelligent scan of the in-core logs.
2789 * If we determine there is no flushable data, we just return. There is no
2790 * flushable data if:
2792 * 1. the current iclog is active and has no data; the previous iclog
2793 * is in the active or dirty state.
2794 * 2. the current iclog is drity, and the previous iclog is in the
2795 * active or dirty state.
2799 * 1. the current iclog is not in the active nor dirty state.
2800 * 2. the current iclog dirty, and the previous iclog is not in the
2801 * active nor dirty state.
2802 * 3. the current iclog is active, and there is another thread writing
2803 * to this particular iclog.
2804 * 4. a) the current iclog is active and has no other writers
2805 * b) when we return from flushing out this iclog, it is still
2806 * not in the active nor dirty state.
2810 struct xfs_mount *mp,
2814 struct log *log = mp->m_log;
2815 struct xlog_in_core *iclog;
2818 XFS_STATS_INC(xs_log_force);
2820 xlog_cil_force(log);
2822 spin_lock(&log->l_icloglock);
2824 iclog = log->l_iclog;
2825 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2826 spin_unlock(&log->l_icloglock);
2827 return XFS_ERROR(EIO);
2830 /* If the head iclog is not active nor dirty, we just attach
2831 * ourselves to the head and go to sleep.
2833 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2834 iclog->ic_state == XLOG_STATE_DIRTY) {
2836 * If the head is dirty or (active and empty), then
2837 * we need to look at the previous iclog. If the previous
2838 * iclog is active or dirty we are done. There is nothing
2839 * to sync out. Otherwise, we attach ourselves to the
2840 * previous iclog and go to sleep.
2842 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2843 (atomic_read(&iclog->ic_refcnt) == 0
2844 && iclog->ic_offset == 0)) {
2845 iclog = iclog->ic_prev;
2846 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2847 iclog->ic_state == XLOG_STATE_DIRTY)
2852 if (atomic_read(&iclog->ic_refcnt) == 0) {
2853 /* We are the only one with access to this
2854 * iclog. Flush it out now. There should
2855 * be a roundoff of zero to show that someone
2856 * has already taken care of the roundoff from
2857 * the previous sync.
2859 atomic_inc(&iclog->ic_refcnt);
2860 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2861 xlog_state_switch_iclogs(log, iclog, 0);
2862 spin_unlock(&log->l_icloglock);
2864 if (xlog_state_release_iclog(log, iclog))
2865 return XFS_ERROR(EIO);
2869 spin_lock(&log->l_icloglock);
2870 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2871 iclog->ic_state != XLOG_STATE_DIRTY)
2876 /* Someone else is writing to this iclog.
2877 * Use its call to flush out the data. However,
2878 * the other thread may not force out this LR,
2879 * so we mark it WANT_SYNC.
2881 xlog_state_switch_iclogs(log, iclog, 0);
2887 /* By the time we come around again, the iclog could've been filled
2888 * which would give it another lsn. If we have a new lsn, just
2889 * return because the relevant data has been flushed.
2892 if (flags & XFS_LOG_SYNC) {
2894 * We must check if we're shutting down here, before
2895 * we wait, while we're holding the l_icloglock.
2896 * Then we check again after waking up, in case our
2897 * sleep was disturbed by a bad news.
2899 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2900 spin_unlock(&log->l_icloglock);
2901 return XFS_ERROR(EIO);
2903 XFS_STATS_INC(xs_log_force_sleep);
2904 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2906 * No need to grab the log lock here since we're
2907 * only deciding whether or not to return EIO
2908 * and the memory read should be atomic.
2910 if (iclog->ic_state & XLOG_STATE_IOERROR)
2911 return XFS_ERROR(EIO);
2917 spin_unlock(&log->l_icloglock);
2923 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2924 * about errors or whether the log was flushed or not. This is the normal
2925 * interface to use when trying to unpin items or move the log forward.
2934 error = _xfs_log_force(mp, flags, NULL);
2936 xfs_warn(mp, "%s: error %d returned.", __func__, error);
2940 * Force the in-core log to disk for a specific LSN.
2942 * Find in-core log with lsn.
2943 * If it is in the DIRTY state, just return.
2944 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2945 * state and go to sleep or return.
2946 * If it is in any other state, go to sleep or return.
2948 * Synchronous forces are implemented with a signal variable. All callers
2949 * to force a given lsn to disk will wait on a the sv attached to the
2950 * specific in-core log. When given in-core log finally completes its
2951 * write to disk, that thread will wake up all threads waiting on the
2956 struct xfs_mount *mp,
2961 struct log *log = mp->m_log;
2962 struct xlog_in_core *iclog;
2963 int already_slept = 0;
2967 XFS_STATS_INC(xs_log_force);
2969 lsn = xlog_cil_force_lsn(log, lsn);
2970 if (lsn == NULLCOMMITLSN)
2974 spin_lock(&log->l_icloglock);
2975 iclog = log->l_iclog;
2976 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2977 spin_unlock(&log->l_icloglock);
2978 return XFS_ERROR(EIO);
2982 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
2983 iclog = iclog->ic_next;
2987 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2988 spin_unlock(&log->l_icloglock);
2992 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2994 * We sleep here if we haven't already slept (e.g.
2995 * this is the first time we've looked at the correct
2996 * iclog buf) and the buffer before us is going to
2997 * be sync'ed. The reason for this is that if we
2998 * are doing sync transactions here, by waiting for
2999 * the previous I/O to complete, we can allow a few
3000 * more transactions into this iclog before we close
3003 * Otherwise, we mark the buffer WANT_SYNC, and bump
3004 * up the refcnt so we can release the log (which
3005 * drops the ref count). The state switch keeps new
3006 * transaction commits from using this buffer. When
3007 * the current commits finish writing into the buffer,
3008 * the refcount will drop to zero and the buffer will
3011 if (!already_slept &&
3012 (iclog->ic_prev->ic_state &
3013 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3014 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3016 XFS_STATS_INC(xs_log_force_sleep);
3018 xlog_wait(&iclog->ic_prev->ic_write_wait,
3025 atomic_inc(&iclog->ic_refcnt);
3026 xlog_state_switch_iclogs(log, iclog, 0);
3027 spin_unlock(&log->l_icloglock);
3028 if (xlog_state_release_iclog(log, iclog))
3029 return XFS_ERROR(EIO);
3032 spin_lock(&log->l_icloglock);
3035 if ((flags & XFS_LOG_SYNC) && /* sleep */
3037 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3039 * Don't wait on completion if we know that we've
3040 * gotten a log write error.
3042 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3043 spin_unlock(&log->l_icloglock);
3044 return XFS_ERROR(EIO);
3046 XFS_STATS_INC(xs_log_force_sleep);
3047 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3049 * No need to grab the log lock here since we're
3050 * only deciding whether or not to return EIO
3051 * and the memory read should be atomic.
3053 if (iclog->ic_state & XLOG_STATE_IOERROR)
3054 return XFS_ERROR(EIO);
3058 } else { /* just return */
3059 spin_unlock(&log->l_icloglock);
3063 } while (iclog != log->l_iclog);
3065 spin_unlock(&log->l_icloglock);
3070 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3071 * about errors or whether the log was flushed or not. This is the normal
3072 * interface to use when trying to unpin items or move the log forward.
3082 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3084 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3088 * Called when we want to mark the current iclog as being ready to sync to
3092 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3094 assert_spin_locked(&log->l_icloglock);
3096 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3097 xlog_state_switch_iclogs(log, iclog, 0);
3099 ASSERT(iclog->ic_state &
3100 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3105 /*****************************************************************************
3109 *****************************************************************************
3113 * Free a used ticket when its refcount falls to zero.
3117 xlog_ticket_t *ticket)
3119 ASSERT(atomic_read(&ticket->t_ref) > 0);
3120 if (atomic_dec_and_test(&ticket->t_ref))
3121 kmem_zone_free(xfs_log_ticket_zone, ticket);
3126 xlog_ticket_t *ticket)
3128 ASSERT(atomic_read(&ticket->t_ref) > 0);
3129 atomic_inc(&ticket->t_ref);
3134 * Allocate and initialise a new log ticket.
3145 struct xlog_ticket *tic;
3149 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3154 * Permanent reservations have up to 'cnt'-1 active log operations
3155 * in the log. A unit in this case is the amount of space for one
3156 * of these log operations. Normal reservations have a cnt of 1
3157 * and their unit amount is the total amount of space required.
3159 * The following lines of code account for non-transaction data
3160 * which occupy space in the on-disk log.
3162 * Normal form of a transaction is:
3163 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3164 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3166 * We need to account for all the leadup data and trailer data
3167 * around the transaction data.
3168 * And then we need to account for the worst case in terms of using
3170 * The worst case will happen if:
3171 * - the placement of the transaction happens to be such that the
3172 * roundoff is at its maximum
3173 * - the transaction data is synced before the commit record is synced
3174 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3175 * Therefore the commit record is in its own Log Record.
3176 * This can happen as the commit record is called with its
3177 * own region to xlog_write().
3178 * This then means that in the worst case, roundoff can happen for
3179 * the commit-rec as well.
3180 * The commit-rec is smaller than padding in this scenario and so it is
3181 * not added separately.
3184 /* for trans header */
3185 unit_bytes += sizeof(xlog_op_header_t);
3186 unit_bytes += sizeof(xfs_trans_header_t);
3189 unit_bytes += sizeof(xlog_op_header_t);
3192 * for LR headers - the space for data in an iclog is the size minus
3193 * the space used for the headers. If we use the iclog size, then we
3194 * undercalculate the number of headers required.
3196 * Furthermore - the addition of op headers for split-recs might
3197 * increase the space required enough to require more log and op
3198 * headers, so take that into account too.
3200 * IMPORTANT: This reservation makes the assumption that if this
3201 * transaction is the first in an iclog and hence has the LR headers
3202 * accounted to it, then the remaining space in the iclog is
3203 * exclusively for this transaction. i.e. if the transaction is larger
3204 * than the iclog, it will be the only thing in that iclog.
3205 * Fundamentally, this means we must pass the entire log vector to
3206 * xlog_write to guarantee this.
3208 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3209 num_headers = howmany(unit_bytes, iclog_space);
3211 /* for split-recs - ophdrs added when data split over LRs */
3212 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3214 /* add extra header reservations if we overrun */
3215 while (!num_headers ||
3216 howmany(unit_bytes, iclog_space) > num_headers) {
3217 unit_bytes += sizeof(xlog_op_header_t);
3220 unit_bytes += log->l_iclog_hsize * num_headers;
3222 /* for commit-rec LR header - note: padding will subsume the ophdr */
3223 unit_bytes += log->l_iclog_hsize;
3225 /* for roundoff padding for transaction data and one for commit record */
3226 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3227 log->l_mp->m_sb.sb_logsunit > 1) {
3228 /* log su roundoff */
3229 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3232 unit_bytes += 2*BBSIZE;
3235 atomic_set(&tic->t_ref, 1);
3236 tic->t_task = current;
3237 INIT_LIST_HEAD(&tic->t_queue);
3238 tic->t_unit_res = unit_bytes;
3239 tic->t_curr_res = unit_bytes;
3242 tic->t_tid = random32();
3243 tic->t_clientid = client;
3244 tic->t_flags = XLOG_TIC_INITED;
3245 tic->t_trans_type = 0;
3247 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3249 xlog_tic_reset_res(tic);
3255 /******************************************************************************
3257 * Log debug routines
3259 ******************************************************************************
3263 * Make sure that the destination ptr is within the valid data region of
3264 * one of the iclogs. This uses backup pointers stored in a different
3265 * part of the log in case we trash the log structure.
3268 xlog_verify_dest_ptr(
3275 for (i = 0; i < log->l_iclog_bufs; i++) {
3276 if (ptr >= log->l_iclog_bak[i] &&
3277 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3282 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3286 * Check to make sure the grant write head didn't just over lap the tail. If
3287 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3288 * the cycles differ by exactly one and check the byte count.
3290 * This check is run unlocked, so can give false positives. Rather than assert
3291 * on failures, use a warn-once flag and a panic tag to allow the admin to
3292 * determine if they want to panic the machine when such an error occurs. For
3293 * debug kernels this will have the same effect as using an assert but, unlinke
3294 * an assert, it can be turned off at runtime.
3297 xlog_verify_grant_tail(
3300 int tail_cycle, tail_blocks;
3303 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3304 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3305 if (tail_cycle != cycle) {
3306 if (cycle - 1 != tail_cycle &&
3307 !(log->l_flags & XLOG_TAIL_WARN)) {
3308 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3309 "%s: cycle - 1 != tail_cycle", __func__);
3310 log->l_flags |= XLOG_TAIL_WARN;
3313 if (space > BBTOB(tail_blocks) &&
3314 !(log->l_flags & XLOG_TAIL_WARN)) {
3315 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3316 "%s: space > BBTOB(tail_blocks)", __func__);
3317 log->l_flags |= XLOG_TAIL_WARN;
3322 /* check if it will fit */
3324 xlog_verify_tail_lsn(xlog_t *log,
3325 xlog_in_core_t *iclog,
3330 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3332 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3333 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3334 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3336 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3338 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3339 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3341 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3342 if (blocks < BTOBB(iclog->ic_offset) + 1)
3343 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3345 } /* xlog_verify_tail_lsn */
3348 * Perform a number of checks on the iclog before writing to disk.
3350 * 1. Make sure the iclogs are still circular
3351 * 2. Make sure we have a good magic number
3352 * 3. Make sure we don't have magic numbers in the data
3353 * 4. Check fields of each log operation header for:
3354 * A. Valid client identifier
3355 * B. tid ptr value falls in valid ptr space (user space code)
3356 * C. Length in log record header is correct according to the
3357 * individual operation headers within record.
3358 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3359 * log, check the preceding blocks of the physical log to make sure all
3360 * the cycle numbers agree with the current cycle number.
3363 xlog_verify_iclog(xlog_t *log,
3364 xlog_in_core_t *iclog,
3368 xlog_op_header_t *ophead;
3369 xlog_in_core_t *icptr;
3370 xlog_in_core_2_t *xhdr;
3372 xfs_caddr_t base_ptr;
3373 __psint_t field_offset;
3375 int len, i, j, k, op_len;
3378 /* check validity of iclog pointers */
3379 spin_lock(&log->l_icloglock);
3380 icptr = log->l_iclog;
3381 for (i=0; i < log->l_iclog_bufs; i++) {
3383 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3384 icptr = icptr->ic_next;
3386 if (icptr != log->l_iclog)
3387 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3388 spin_unlock(&log->l_icloglock);
3390 /* check log magic numbers */
3391 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3392 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3394 ptr = (xfs_caddr_t) &iclog->ic_header;
3395 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3397 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3398 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3403 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3404 ptr = iclog->ic_datap;
3406 ophead = (xlog_op_header_t *)ptr;
3407 xhdr = iclog->ic_data;
3408 for (i = 0; i < len; i++) {
3409 ophead = (xlog_op_header_t *)ptr;
3411 /* clientid is only 1 byte */
3412 field_offset = (__psint_t)
3413 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3414 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3415 clientid = ophead->oh_clientid;
3417 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3418 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3419 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3420 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3421 clientid = xlog_get_client_id(
3422 xhdr[j].hic_xheader.xh_cycle_data[k]);
3424 clientid = xlog_get_client_id(
3425 iclog->ic_header.h_cycle_data[idx]);
3428 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3430 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3431 __func__, clientid, ophead,
3432 (unsigned long)field_offset);
3435 field_offset = (__psint_t)
3436 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3437 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3438 op_len = be32_to_cpu(ophead->oh_len);
3440 idx = BTOBBT((__psint_t)&ophead->oh_len -
3441 (__psint_t)iclog->ic_datap);
3442 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3443 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3444 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3445 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3447 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3450 ptr += sizeof(xlog_op_header_t) + op_len;
3452 } /* xlog_verify_iclog */
3456 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3462 xlog_in_core_t *iclog, *ic;
3464 iclog = log->l_iclog;
3465 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3467 * Mark all the incore logs IOERROR.
3468 * From now on, no log flushes will result.
3472 ic->ic_state = XLOG_STATE_IOERROR;
3474 } while (ic != iclog);
3478 * Return non-zero, if state transition has already happened.
3484 * This is called from xfs_force_shutdown, when we're forcibly
3485 * shutting down the filesystem, typically because of an IO error.
3486 * Our main objectives here are to make sure that:
3487 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3488 * parties to find out, 'atomically'.
3489 * b. those who're sleeping on log reservations, pinned objects and
3490 * other resources get woken up, and be told the bad news.
3491 * c. nothing new gets queued up after (a) and (b) are done.
3492 * d. if !logerror, flush the iclogs to disk, then seal them off
3495 * Note: for delayed logging the !logerror case needs to flush the regions
3496 * held in memory out to the iclogs before flushing them to disk. This needs
3497 * to be done before the log is marked as shutdown, otherwise the flush to the
3501 xfs_log_force_umount(
3502 struct xfs_mount *mp,
3511 * If this happens during log recovery, don't worry about
3512 * locking; the log isn't open for business yet.
3515 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3516 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3518 XFS_BUF_DONE(mp->m_sb_bp);
3523 * Somebody could've already done the hard work for us.
3524 * No need to get locks for this.
3526 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3527 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3533 * Flush the in memory commit item list before marking the log as
3534 * being shut down. We need to do it in this order to ensure all the
3535 * completed transactions are flushed to disk with the xfs_log_force()
3539 xlog_cil_force(log);
3542 * mark the filesystem and the as in a shutdown state and wake
3543 * everybody up to tell them the bad news.
3545 spin_lock(&log->l_icloglock);
3546 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3548 XFS_BUF_DONE(mp->m_sb_bp);
3551 * This flag is sort of redundant because of the mount flag, but
3552 * it's good to maintain the separation between the log and the rest
3555 log->l_flags |= XLOG_IO_ERROR;
3558 * If we hit a log error, we want to mark all the iclogs IOERROR
3559 * while we're still holding the loglock.
3562 retval = xlog_state_ioerror(log);
3563 spin_unlock(&log->l_icloglock);
3566 * We don't want anybody waiting for log reservations after this. That
3567 * means we have to wake up everybody queued up on reserveq as well as
3568 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3569 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3570 * action is protected by the grant locks.
3572 xlog_grant_head_wake_all(&log->l_reserve_head);
3573 xlog_grant_head_wake_all(&log->l_write_head);
3575 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3578 * Force the incore logs to disk before shutting the
3579 * log down completely.
3581 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3583 spin_lock(&log->l_icloglock);
3584 retval = xlog_state_ioerror(log);
3585 spin_unlock(&log->l_icloglock);
3588 * Wake up everybody waiting on xfs_log_force.
3589 * Callback all log item committed functions as if the
3590 * log writes were completed.
3592 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3594 #ifdef XFSERRORDEBUG
3596 xlog_in_core_t *iclog;
3598 spin_lock(&log->l_icloglock);
3599 iclog = log->l_iclog;
3601 ASSERT(iclog->ic_callback == 0);
3602 iclog = iclog->ic_next;
3603 } while (iclog != log->l_iclog);
3604 spin_unlock(&log->l_icloglock);
3607 /* return non-zero if log IOERROR transition had already happened */
3612 xlog_iclogs_empty(xlog_t *log)
3614 xlog_in_core_t *iclog;
3616 iclog = log->l_iclog;
3618 /* endianness does not matter here, zero is zero in
3621 if (iclog->ic_header.h_num_logops)
3623 iclog = iclog->ic_next;
3624 } while (iclog != log->l_iclog);