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"
22 #include "xfs_trans.h"
25 #include "xfs_mount.h"
26 #include "xfs_error.h"
27 #include "xfs_log_priv.h"
28 #include "xfs_buf_item.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_log_recover.h"
33 #include "xfs_trans_priv.h"
34 #include "xfs_dinode.h"
35 #include "xfs_inode.h"
36 #include "xfs_trace.h"
37 #include "xfs_fsops.h"
38 #include "xfs_cksum.h"
40 kmem_zone_t *xfs_log_ticket_zone;
42 /* Local miscellaneous function prototypes */
46 struct xlog_ticket *ticket,
47 struct xlog_in_core **iclog,
48 xfs_lsn_t *commitlsnp);
53 struct xfs_buftarg *log_target,
54 xfs_daddr_t blk_offset,
63 struct xlog_in_core *iclog);
68 /* local state machine functions */
69 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
71 xlog_state_do_callback(
74 struct xlog_in_core *iclog);
76 xlog_state_get_iclog_space(
79 struct xlog_in_core **iclog,
80 struct xlog_ticket *ticket,
84 xlog_state_release_iclog(
86 struct xlog_in_core *iclog);
88 xlog_state_switch_iclogs(
90 struct xlog_in_core *iclog,
95 struct xlog_in_core *iclog);
102 xlog_regrant_reserve_log_space(
104 struct xlog_ticket *ticket);
106 xlog_ungrant_log_space(
108 struct xlog_ticket *ticket);
112 xlog_verify_dest_ptr(
116 xlog_verify_grant_tail(
121 struct xlog_in_core *iclog,
125 xlog_verify_tail_lsn(
127 struct xlog_in_core *iclog,
130 #define xlog_verify_dest_ptr(a,b)
131 #define xlog_verify_grant_tail(a)
132 #define xlog_verify_iclog(a,b,c,d)
133 #define xlog_verify_tail_lsn(a,b,c)
141 xlog_grant_sub_space(
146 int64_t head_val = atomic64_read(head);
152 xlog_crack_grant_head_val(head_val, &cycle, &space);
156 space += log->l_logsize;
161 new = xlog_assign_grant_head_val(cycle, space);
162 head_val = atomic64_cmpxchg(head, old, new);
163 } while (head_val != old);
167 xlog_grant_add_space(
172 int64_t head_val = atomic64_read(head);
179 xlog_crack_grant_head_val(head_val, &cycle, &space);
181 tmp = log->l_logsize - space;
190 new = xlog_assign_grant_head_val(cycle, space);
191 head_val = atomic64_cmpxchg(head, old, new);
192 } while (head_val != old);
196 xlog_grant_head_init(
197 struct xlog_grant_head *head)
199 xlog_assign_grant_head(&head->grant, 1, 0);
200 INIT_LIST_HEAD(&head->waiters);
201 spin_lock_init(&head->lock);
205 xlog_grant_head_wake_all(
206 struct xlog_grant_head *head)
208 struct xlog_ticket *tic;
210 spin_lock(&head->lock);
211 list_for_each_entry(tic, &head->waiters, t_queue)
212 wake_up_process(tic->t_task);
213 spin_unlock(&head->lock);
217 xlog_ticket_reservation(
219 struct xlog_grant_head *head,
220 struct xlog_ticket *tic)
222 if (head == &log->l_write_head) {
223 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
224 return tic->t_unit_res;
226 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
227 return tic->t_unit_res * tic->t_cnt;
229 return tic->t_unit_res;
234 xlog_grant_head_wake(
236 struct xlog_grant_head *head,
239 struct xlog_ticket *tic;
242 list_for_each_entry(tic, &head->waiters, t_queue) {
243 need_bytes = xlog_ticket_reservation(log, head, tic);
244 if (*free_bytes < need_bytes)
247 *free_bytes -= need_bytes;
248 trace_xfs_log_grant_wake_up(log, tic);
249 wake_up_process(tic->t_task);
256 xlog_grant_head_wait(
258 struct xlog_grant_head *head,
259 struct xlog_ticket *tic,
262 list_add_tail(&tic->t_queue, &head->waiters);
265 if (XLOG_FORCED_SHUTDOWN(log))
267 xlog_grant_push_ail(log, need_bytes);
269 __set_current_state(TASK_UNINTERRUPTIBLE);
270 spin_unlock(&head->lock);
272 XFS_STATS_INC(xs_sleep_logspace);
274 trace_xfs_log_grant_sleep(log, tic);
276 trace_xfs_log_grant_wake(log, tic);
278 spin_lock(&head->lock);
279 if (XLOG_FORCED_SHUTDOWN(log))
281 } while (xlog_space_left(log, &head->grant) < need_bytes);
283 list_del_init(&tic->t_queue);
286 list_del_init(&tic->t_queue);
287 return XFS_ERROR(EIO);
291 * Atomically get the log space required for a log ticket.
293 * Once a ticket gets put onto head->waiters, it will only return after the
294 * needed reservation is satisfied.
296 * This function is structured so that it has a lock free fast path. This is
297 * necessary because every new transaction reservation will come through this
298 * path. Hence any lock will be globally hot if we take it unconditionally on
301 * As tickets are only ever moved on and off head->waiters under head->lock, we
302 * only need to take that lock if we are going to add the ticket to the queue
303 * and sleep. We can avoid taking the lock if the ticket was never added to
304 * head->waiters because the t_queue list head will be empty and we hold the
305 * only reference to it so it can safely be checked unlocked.
308 xlog_grant_head_check(
310 struct xlog_grant_head *head,
311 struct xlog_ticket *tic,
317 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
320 * If there are other waiters on the queue then give them a chance at
321 * logspace before us. Wake up the first waiters, if we do not wake
322 * up all the waiters then go to sleep waiting for more free space,
323 * otherwise try to get some space for this transaction.
325 *need_bytes = xlog_ticket_reservation(log, head, tic);
326 free_bytes = xlog_space_left(log, &head->grant);
327 if (!list_empty_careful(&head->waiters)) {
328 spin_lock(&head->lock);
329 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
330 free_bytes < *need_bytes) {
331 error = xlog_grant_head_wait(log, head, tic,
334 spin_unlock(&head->lock);
335 } else if (free_bytes < *need_bytes) {
336 spin_lock(&head->lock);
337 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
338 spin_unlock(&head->lock);
345 xlog_tic_reset_res(xlog_ticket_t *tic)
348 tic->t_res_arr_sum = 0;
349 tic->t_res_num_ophdrs = 0;
353 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
355 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
356 /* add to overflow and start again */
357 tic->t_res_o_flow += tic->t_res_arr_sum;
359 tic->t_res_arr_sum = 0;
362 tic->t_res_arr[tic->t_res_num].r_len = len;
363 tic->t_res_arr[tic->t_res_num].r_type = type;
364 tic->t_res_arr_sum += len;
369 * Replenish the byte reservation required by moving the grant write head.
373 struct xfs_mount *mp,
374 struct xlog_ticket *tic)
376 struct xlog *log = mp->m_log;
380 if (XLOG_FORCED_SHUTDOWN(log))
381 return XFS_ERROR(EIO);
383 XFS_STATS_INC(xs_try_logspace);
386 * This is a new transaction on the ticket, so we need to change the
387 * transaction ID so that the next transaction has a different TID in
388 * the log. Just add one to the existing tid so that we can see chains
389 * of rolling transactions in the log easily.
393 xlog_grant_push_ail(log, tic->t_unit_res);
395 tic->t_curr_res = tic->t_unit_res;
396 xlog_tic_reset_res(tic);
401 trace_xfs_log_regrant(log, tic);
403 error = xlog_grant_head_check(log, &log->l_write_head, tic,
408 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
409 trace_xfs_log_regrant_exit(log, tic);
410 xlog_verify_grant_tail(log);
415 * If we are failing, make sure the ticket doesn't have any current
416 * reservations. We don't want to add this back when the ticket/
417 * transaction gets cancelled.
420 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
425 * Reserve log space and return a ticket corresponding the reservation.
427 * Each reservation is going to reserve extra space for a log record header.
428 * When writes happen to the on-disk log, we don't subtract the length of the
429 * log record header from any reservation. By wasting space in each
430 * reservation, we prevent over allocation problems.
434 struct xfs_mount *mp,
437 struct xlog_ticket **ticp,
442 struct xlog *log = mp->m_log;
443 struct xlog_ticket *tic;
447 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
449 if (XLOG_FORCED_SHUTDOWN(log))
450 return XFS_ERROR(EIO);
452 XFS_STATS_INC(xs_try_logspace);
454 ASSERT(*ticp == NULL);
455 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
456 KM_SLEEP | KM_MAYFAIL);
458 return XFS_ERROR(ENOMEM);
460 tic->t_trans_type = t_type;
463 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
466 trace_xfs_log_reserve(log, tic);
468 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
473 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
474 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
475 trace_xfs_log_reserve_exit(log, tic);
476 xlog_verify_grant_tail(log);
481 * If we are failing, make sure the ticket doesn't have any current
482 * reservations. We don't want to add this back when the ticket/
483 * transaction gets cancelled.
486 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
494 * 1. currblock field gets updated at startup and after in-core logs
495 * marked as with WANT_SYNC.
499 * This routine is called when a user of a log manager ticket is done with
500 * the reservation. If the ticket was ever used, then a commit record for
501 * the associated transaction is written out as a log operation header with
502 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
503 * a given ticket. If the ticket was one with a permanent reservation, then
504 * a few operations are done differently. Permanent reservation tickets by
505 * default don't release the reservation. They just commit the current
506 * transaction with the belief that the reservation is still needed. A flag
507 * must be passed in before permanent reservations are actually released.
508 * When these type of tickets are not released, they need to be set into
509 * the inited state again. By doing this, a start record will be written
510 * out when the next write occurs.
514 struct xfs_mount *mp,
515 struct xlog_ticket *ticket,
516 struct xlog_in_core **iclog,
519 struct xlog *log = mp->m_log;
522 if (XLOG_FORCED_SHUTDOWN(log) ||
524 * If nothing was ever written, don't write out commit record.
525 * If we get an error, just continue and give back the log ticket.
527 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
528 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
529 lsn = (xfs_lsn_t) -1;
530 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
531 flags |= XFS_LOG_REL_PERM_RESERV;
536 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
537 (flags & XFS_LOG_REL_PERM_RESERV)) {
538 trace_xfs_log_done_nonperm(log, ticket);
541 * Release ticket if not permanent reservation or a specific
542 * request has been made to release a permanent reservation.
544 xlog_ungrant_log_space(log, ticket);
545 xfs_log_ticket_put(ticket);
547 trace_xfs_log_done_perm(log, ticket);
549 xlog_regrant_reserve_log_space(log, ticket);
550 /* If this ticket was a permanent reservation and we aren't
551 * trying to release it, reset the inited flags; so next time
552 * we write, a start record will be written out.
554 ticket->t_flags |= XLOG_TIC_INITED;
561 * Attaches a new iclog I/O completion callback routine during
562 * transaction commit. If the log is in error state, a non-zero
563 * return code is handed back and the caller is responsible for
564 * executing the callback at an appropriate time.
568 struct xfs_mount *mp,
569 struct xlog_in_core *iclog,
570 xfs_log_callback_t *cb)
574 spin_lock(&iclog->ic_callback_lock);
575 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
577 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
578 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
580 *(iclog->ic_callback_tail) = cb;
581 iclog->ic_callback_tail = &(cb->cb_next);
583 spin_unlock(&iclog->ic_callback_lock);
588 xfs_log_release_iclog(
589 struct xfs_mount *mp,
590 struct xlog_in_core *iclog)
592 if (xlog_state_release_iclog(mp->m_log, iclog)) {
593 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
601 * Mount a log filesystem
603 * mp - ubiquitous xfs mount point structure
604 * log_target - buftarg of on-disk log device
605 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
606 * num_bblocks - Number of BBSIZE blocks in on-disk log
608 * Return error or zero.
613 xfs_buftarg_t *log_target,
614 xfs_daddr_t blk_offset,
620 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
621 xfs_notice(mp, "Mounting Filesystem");
624 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
625 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
628 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
629 if (IS_ERR(mp->m_log)) {
630 error = -PTR_ERR(mp->m_log);
635 * Validate the given log space and drop a critical message via syslog
636 * if the log size is too small that would lead to some unexpected
637 * situations in transaction log space reservation stage.
639 * Note: we can't just reject the mount if the validation fails. This
640 * would mean that people would have to downgrade their kernel just to
641 * remedy the situation as there is no way to grow the log (short of
642 * black magic surgery with xfs_db).
644 * We can, however, reject mounts for CRC format filesystems, as the
645 * mkfs binary being used to make the filesystem should never create a
646 * filesystem with a log that is too small.
648 min_logfsbs = xfs_log_calc_minimum_size(mp);
650 if (mp->m_sb.sb_logblocks < min_logfsbs) {
652 "Log size %d blocks too small, minimum size is %d blocks",
653 mp->m_sb.sb_logblocks, min_logfsbs);
655 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
657 "Log size %d blocks too large, maximum size is %lld blocks",
658 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
660 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
662 "log size %lld bytes too large, maximum size is %lld bytes",
663 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
668 if (xfs_sb_version_hascrc(&mp->m_sb)) {
669 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
674 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
675 "experienced then please report this message in the bug report.");
679 * Initialize the AIL now we have a log.
681 error = xfs_trans_ail_init(mp);
683 xfs_warn(mp, "AIL initialisation failed: error %d", error);
686 mp->m_log->l_ailp = mp->m_ail;
689 * skip log recovery on a norecovery mount. pretend it all
692 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
693 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
696 mp->m_flags &= ~XFS_MOUNT_RDONLY;
698 error = xlog_recover(mp->m_log);
701 mp->m_flags |= XFS_MOUNT_RDONLY;
703 xfs_warn(mp, "log mount/recovery failed: error %d",
705 goto out_destroy_ail;
709 /* Normal transactions can now occur */
710 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
713 * Now the log has been fully initialised and we know were our
714 * space grant counters are, we can initialise the permanent ticket
715 * needed for delayed logging to work.
717 xlog_cil_init_post_recovery(mp->m_log);
722 xfs_trans_ail_destroy(mp);
724 xlog_dealloc_log(mp->m_log);
730 * Finish the recovery of the file system. This is separate from the
731 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
732 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
735 * If we finish recovery successfully, start the background log work. If we are
736 * not doing recovery, then we have a RO filesystem and we don't need to start
740 xfs_log_mount_finish(xfs_mount_t *mp)
744 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
745 error = xlog_recover_finish(mp->m_log);
747 xfs_log_work_queue(mp);
749 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
757 * Final log writes as part of unmount.
759 * Mark the filesystem clean as unmount happens. Note that during relocation
760 * this routine needs to be executed as part of source-bag while the
761 * deallocation must not be done until source-end.
765 * Unmount record used to have a string "Unmount filesystem--" in the
766 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
767 * We just write the magic number now since that particular field isn't
768 * currently architecture converted and "Unmount" is a bit foo.
769 * As far as I know, there weren't any dependencies on the old behaviour.
773 xfs_log_unmount_write(xfs_mount_t *mp)
775 struct xlog *log = mp->m_log;
776 xlog_in_core_t *iclog;
778 xlog_in_core_t *first_iclog;
780 xlog_ticket_t *tic = NULL;
785 * Don't write out unmount record on read-only mounts.
786 * Or, if we are doing a forced umount (typically because of IO errors).
788 if (mp->m_flags & XFS_MOUNT_RDONLY)
791 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
792 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
795 first_iclog = iclog = log->l_iclog;
797 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
798 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
799 ASSERT(iclog->ic_offset == 0);
801 iclog = iclog->ic_next;
802 } while (iclog != first_iclog);
804 if (! (XLOG_FORCED_SHUTDOWN(log))) {
805 error = xfs_log_reserve(mp, 600, 1, &tic,
806 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
808 /* the data section must be 32 bit size aligned */
812 __uint32_t pad2; /* may as well make it 64 bits */
814 .magic = XLOG_UNMOUNT_TYPE,
816 struct xfs_log_iovec reg = {
818 .i_len = sizeof(magic),
819 .i_type = XLOG_REG_TYPE_UNMOUNT,
821 struct xfs_log_vec vec = {
826 /* remove inited flag, and account for space used */
828 tic->t_curr_res -= sizeof(magic);
829 error = xlog_write(log, &vec, tic, &lsn,
830 NULL, XLOG_UNMOUNT_TRANS);
832 * At this point, we're umounting anyway,
833 * so there's no point in transitioning log state
834 * to IOERROR. Just continue...
839 xfs_alert(mp, "%s: unmount record failed", __func__);
842 spin_lock(&log->l_icloglock);
843 iclog = log->l_iclog;
844 atomic_inc(&iclog->ic_refcnt);
845 xlog_state_want_sync(log, iclog);
846 spin_unlock(&log->l_icloglock);
847 error = xlog_state_release_iclog(log, iclog);
849 spin_lock(&log->l_icloglock);
850 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
851 iclog->ic_state == XLOG_STATE_DIRTY)) {
852 if (!XLOG_FORCED_SHUTDOWN(log)) {
853 xlog_wait(&iclog->ic_force_wait,
856 spin_unlock(&log->l_icloglock);
859 spin_unlock(&log->l_icloglock);
862 trace_xfs_log_umount_write(log, tic);
863 xlog_ungrant_log_space(log, tic);
864 xfs_log_ticket_put(tic);
868 * We're already in forced_shutdown mode, couldn't
869 * even attempt to write out the unmount transaction.
871 * Go through the motions of sync'ing and releasing
872 * the iclog, even though no I/O will actually happen,
873 * we need to wait for other log I/Os that may already
874 * be in progress. Do this as a separate section of
875 * code so we'll know if we ever get stuck here that
876 * we're in this odd situation of trying to unmount
877 * a file system that went into forced_shutdown as
878 * the result of an unmount..
880 spin_lock(&log->l_icloglock);
881 iclog = log->l_iclog;
882 atomic_inc(&iclog->ic_refcnt);
884 xlog_state_want_sync(log, iclog);
885 spin_unlock(&log->l_icloglock);
886 error = xlog_state_release_iclog(log, iclog);
888 spin_lock(&log->l_icloglock);
890 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
891 || iclog->ic_state == XLOG_STATE_DIRTY
892 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
894 xlog_wait(&iclog->ic_force_wait,
897 spin_unlock(&log->l_icloglock);
902 } /* xfs_log_unmount_write */
905 * Empty the log for unmount/freeze.
907 * To do this, we first need to shut down the background log work so it is not
908 * trying to cover the log as we clean up. We then need to unpin all objects in
909 * the log so we can then flush them out. Once they have completed their IO and
910 * run the callbacks removing themselves from the AIL, we can write the unmount
915 struct xfs_mount *mp)
917 cancel_delayed_work_sync(&mp->m_log->l_work);
918 xfs_log_force(mp, XFS_LOG_SYNC);
921 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
922 * will push it, xfs_wait_buftarg() will not wait for it. Further,
923 * xfs_buf_iowait() cannot be used because it was pushed with the
924 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
925 * the IO to complete.
927 xfs_ail_push_all_sync(mp->m_ail);
928 xfs_wait_buftarg(mp->m_ddev_targp);
929 xfs_buf_lock(mp->m_sb_bp);
930 xfs_buf_unlock(mp->m_sb_bp);
932 xfs_log_unmount_write(mp);
936 * Shut down and release the AIL and Log.
938 * During unmount, we need to ensure we flush all the dirty metadata objects
939 * from the AIL so that the log is empty before we write the unmount record to
940 * the log. Once this is done, we can tear down the AIL and the log.
944 struct xfs_mount *mp)
948 xfs_trans_ail_destroy(mp);
949 xlog_dealloc_log(mp->m_log);
954 struct xfs_mount *mp,
955 struct xfs_log_item *item,
957 const struct xfs_item_ops *ops)
959 item->li_mountp = mp;
960 item->li_ailp = mp->m_ail;
961 item->li_type = type;
965 INIT_LIST_HEAD(&item->li_ail);
966 INIT_LIST_HEAD(&item->li_cil);
970 * Wake up processes waiting for log space after we have moved the log tail.
974 struct xfs_mount *mp)
976 struct xlog *log = mp->m_log;
979 if (XLOG_FORCED_SHUTDOWN(log))
982 if (!list_empty_careful(&log->l_write_head.waiters)) {
983 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
985 spin_lock(&log->l_write_head.lock);
986 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
987 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
988 spin_unlock(&log->l_write_head.lock);
991 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
992 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
994 spin_lock(&log->l_reserve_head.lock);
995 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
996 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
997 spin_unlock(&log->l_reserve_head.lock);
1002 * Determine if we have a transaction that has gone to disk
1003 * that needs to be covered. To begin the transition to the idle state
1004 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
1005 * If we are then in a state where covering is needed, the caller is informed
1006 * that dummy transactions are required to move the log into the idle state.
1008 * Because this is called as part of the sync process, we should also indicate
1009 * that dummy transactions should be issued in anything but the covered or
1010 * idle states. This ensures that the log tail is accurately reflected in
1011 * the log at the end of the sync, hence if a crash occurrs avoids replay
1012 * of transactions where the metadata is already on disk.
1015 xfs_log_need_covered(xfs_mount_t *mp)
1018 struct xlog *log = mp->m_log;
1020 if (!xfs_fs_writable(mp))
1023 spin_lock(&log->l_icloglock);
1024 switch (log->l_covered_state) {
1025 case XLOG_STATE_COVER_DONE:
1026 case XLOG_STATE_COVER_DONE2:
1027 case XLOG_STATE_COVER_IDLE:
1029 case XLOG_STATE_COVER_NEED:
1030 case XLOG_STATE_COVER_NEED2:
1031 if (!xfs_ail_min_lsn(log->l_ailp) &&
1032 xlog_iclogs_empty(log)) {
1033 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1034 log->l_covered_state = XLOG_STATE_COVER_DONE;
1036 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1043 spin_unlock(&log->l_icloglock);
1048 * We may be holding the log iclog lock upon entering this routine.
1051 xlog_assign_tail_lsn_locked(
1052 struct xfs_mount *mp)
1054 struct xlog *log = mp->m_log;
1055 struct xfs_log_item *lip;
1058 assert_spin_locked(&mp->m_ail->xa_lock);
1061 * To make sure we always have a valid LSN for the log tail we keep
1062 * track of the last LSN which was committed in log->l_last_sync_lsn,
1063 * and use that when the AIL was empty.
1065 lip = xfs_ail_min(mp->m_ail);
1067 tail_lsn = lip->li_lsn;
1069 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1070 atomic64_set(&log->l_tail_lsn, tail_lsn);
1075 xlog_assign_tail_lsn(
1076 struct xfs_mount *mp)
1080 spin_lock(&mp->m_ail->xa_lock);
1081 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1082 spin_unlock(&mp->m_ail->xa_lock);
1088 * Return the space in the log between the tail and the head. The head
1089 * is passed in the cycle/bytes formal parms. In the special case where
1090 * the reserve head has wrapped passed the tail, this calculation is no
1091 * longer valid. In this case, just return 0 which means there is no space
1092 * in the log. This works for all places where this function is called
1093 * with the reserve head. Of course, if the write head were to ever
1094 * wrap the tail, we should blow up. Rather than catch this case here,
1095 * we depend on other ASSERTions in other parts of the code. XXXmiken
1097 * This code also handles the case where the reservation head is behind
1098 * the tail. The details of this case are described below, but the end
1099 * result is that we return the size of the log as the amount of space left.
1112 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1113 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1114 tail_bytes = BBTOB(tail_bytes);
1115 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1116 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1117 else if (tail_cycle + 1 < head_cycle)
1119 else if (tail_cycle < head_cycle) {
1120 ASSERT(tail_cycle == (head_cycle - 1));
1121 free_bytes = tail_bytes - head_bytes;
1124 * The reservation head is behind the tail.
1125 * In this case we just want to return the size of the
1126 * log as the amount of space left.
1128 xfs_alert(log->l_mp,
1129 "xlog_space_left: head behind tail\n"
1130 " tail_cycle = %d, tail_bytes = %d\n"
1131 " GH cycle = %d, GH bytes = %d",
1132 tail_cycle, tail_bytes, head_cycle, head_bytes);
1134 free_bytes = log->l_logsize;
1141 * Log function which is called when an io completes.
1143 * The log manager needs its own routine, in order to control what
1144 * happens with the buffer after the write completes.
1147 xlog_iodone(xfs_buf_t *bp)
1149 struct xlog_in_core *iclog = bp->b_fspriv;
1150 struct xlog *l = iclog->ic_log;
1154 * Race to shutdown the filesystem if we see an error.
1156 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1157 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1158 xfs_buf_ioerror_alert(bp, __func__);
1160 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1162 * This flag will be propagated to the trans-committed
1163 * callback routines to let them know that the log-commit
1166 aborted = XFS_LI_ABORTED;
1167 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1168 aborted = XFS_LI_ABORTED;
1171 /* log I/O is always issued ASYNC */
1172 ASSERT(XFS_BUF_ISASYNC(bp));
1173 xlog_state_done_syncing(iclog, aborted);
1175 * do not reference the buffer (bp) here as we could race
1176 * with it being freed after writing the unmount record to the
1182 * Return size of each in-core log record buffer.
1184 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1186 * If the filesystem blocksize is too large, we may need to choose a
1187 * larger size since the directory code currently logs entire blocks.
1191 xlog_get_iclog_buffer_size(
1192 struct xfs_mount *mp,
1198 if (mp->m_logbufs <= 0)
1199 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1201 log->l_iclog_bufs = mp->m_logbufs;
1204 * Buffer size passed in from mount system call.
1206 if (mp->m_logbsize > 0) {
1207 size = log->l_iclog_size = mp->m_logbsize;
1208 log->l_iclog_size_log = 0;
1210 log->l_iclog_size_log++;
1214 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1215 /* # headers = size / 32k
1216 * one header holds cycles from 32k of data
1219 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1220 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1222 log->l_iclog_hsize = xhdrs << BBSHIFT;
1223 log->l_iclog_heads = xhdrs;
1225 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1226 log->l_iclog_hsize = BBSIZE;
1227 log->l_iclog_heads = 1;
1232 /* All machines use 32kB buffers by default. */
1233 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1234 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1236 /* the default log size is 16k or 32k which is one header sector */
1237 log->l_iclog_hsize = BBSIZE;
1238 log->l_iclog_heads = 1;
1241 /* are we being asked to make the sizes selected above visible? */
1242 if (mp->m_logbufs == 0)
1243 mp->m_logbufs = log->l_iclog_bufs;
1244 if (mp->m_logbsize == 0)
1245 mp->m_logbsize = log->l_iclog_size;
1246 } /* xlog_get_iclog_buffer_size */
1251 struct xfs_mount *mp)
1253 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1254 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1258 * Every sync period we need to unpin all items in the AIL and push them to
1259 * disk. If there is nothing dirty, then we might need to cover the log to
1260 * indicate that the filesystem is idle.
1264 struct work_struct *work)
1266 struct xlog *log = container_of(to_delayed_work(work),
1267 struct xlog, l_work);
1268 struct xfs_mount *mp = log->l_mp;
1270 /* dgc: errors ignored - not fatal and nowhere to report them */
1271 if (xfs_log_need_covered(mp))
1272 xfs_fs_log_dummy(mp);
1274 xfs_log_force(mp, 0);
1276 /* start pushing all the metadata that is currently dirty */
1277 xfs_ail_push_all(mp->m_ail);
1279 /* queue us up again */
1280 xfs_log_work_queue(mp);
1284 * This routine initializes some of the log structure for a given mount point.
1285 * Its primary purpose is to fill in enough, so recovery can occur. However,
1286 * some other stuff may be filled in too.
1288 STATIC struct xlog *
1290 struct xfs_mount *mp,
1291 struct xfs_buftarg *log_target,
1292 xfs_daddr_t blk_offset,
1296 xlog_rec_header_t *head;
1297 xlog_in_core_t **iclogp;
1298 xlog_in_core_t *iclog, *prev_iclog=NULL;
1304 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1306 xfs_warn(mp, "Log allocation failed: No memory!");
1311 log->l_targ = log_target;
1312 log->l_logsize = BBTOB(num_bblks);
1313 log->l_logBBstart = blk_offset;
1314 log->l_logBBsize = num_bblks;
1315 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1316 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1317 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1319 log->l_prev_block = -1;
1320 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1321 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1322 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1323 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1325 xlog_grant_head_init(&log->l_reserve_head);
1326 xlog_grant_head_init(&log->l_write_head);
1328 error = EFSCORRUPTED;
1329 if (xfs_sb_version_hassector(&mp->m_sb)) {
1330 log2_size = mp->m_sb.sb_logsectlog;
1331 if (log2_size < BBSHIFT) {
1332 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1333 log2_size, BBSHIFT);
1337 log2_size -= BBSHIFT;
1338 if (log2_size > mp->m_sectbb_log) {
1339 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1340 log2_size, mp->m_sectbb_log);
1344 /* for larger sector sizes, must have v2 or external log */
1345 if (log2_size && log->l_logBBstart > 0 &&
1346 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1348 "log sector size (0x%x) invalid for configuration.",
1353 log->l_sectBBsize = 1 << log2_size;
1355 xlog_get_iclog_buffer_size(mp, log);
1358 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1361 bp->b_iodone = xlog_iodone;
1362 ASSERT(xfs_buf_islocked(bp));
1365 spin_lock_init(&log->l_icloglock);
1366 init_waitqueue_head(&log->l_flush_wait);
1368 iclogp = &log->l_iclog;
1370 * The amount of memory to allocate for the iclog structure is
1371 * rather funky due to the way the structure is defined. It is
1372 * done this way so that we can use different sizes for machines
1373 * with different amounts of memory. See the definition of
1374 * xlog_in_core_t in xfs_log_priv.h for details.
1376 ASSERT(log->l_iclog_size >= 4096);
1377 for (i=0; i < log->l_iclog_bufs; i++) {
1378 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1380 goto out_free_iclog;
1383 iclog->ic_prev = prev_iclog;
1386 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1387 BTOBB(log->l_iclog_size), 0);
1389 goto out_free_iclog;
1391 bp->b_iodone = xlog_iodone;
1393 iclog->ic_data = bp->b_addr;
1395 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1397 head = &iclog->ic_header;
1398 memset(head, 0, sizeof(xlog_rec_header_t));
1399 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1400 head->h_version = cpu_to_be32(
1401 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1402 head->h_size = cpu_to_be32(log->l_iclog_size);
1404 head->h_fmt = cpu_to_be32(XLOG_FMT);
1405 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1407 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1408 iclog->ic_state = XLOG_STATE_ACTIVE;
1409 iclog->ic_log = log;
1410 atomic_set(&iclog->ic_refcnt, 0);
1411 spin_lock_init(&iclog->ic_callback_lock);
1412 iclog->ic_callback_tail = &(iclog->ic_callback);
1413 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1415 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1416 init_waitqueue_head(&iclog->ic_force_wait);
1417 init_waitqueue_head(&iclog->ic_write_wait);
1419 iclogp = &iclog->ic_next;
1421 *iclogp = log->l_iclog; /* complete ring */
1422 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1424 error = xlog_cil_init(log);
1426 goto out_free_iclog;
1430 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1431 prev_iclog = iclog->ic_next;
1433 xfs_buf_free(iclog->ic_bp);
1436 spinlock_destroy(&log->l_icloglock);
1437 xfs_buf_free(log->l_xbuf);
1441 return ERR_PTR(-error);
1442 } /* xlog_alloc_log */
1446 * Write out the commit record of a transaction associated with the given
1447 * ticket. Return the lsn of the commit record.
1452 struct xlog_ticket *ticket,
1453 struct xlog_in_core **iclog,
1454 xfs_lsn_t *commitlsnp)
1456 struct xfs_mount *mp = log->l_mp;
1458 struct xfs_log_iovec reg = {
1461 .i_type = XLOG_REG_TYPE_COMMIT,
1463 struct xfs_log_vec vec = {
1468 ASSERT_ALWAYS(iclog);
1469 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1472 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1477 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1478 * log space. This code pushes on the lsn which would supposedly free up
1479 * the 25% which we want to leave free. We may need to adopt a policy which
1480 * pushes on an lsn which is further along in the log once we reach the high
1481 * water mark. In this manner, we would be creating a low water mark.
1484 xlog_grant_push_ail(
1488 xfs_lsn_t threshold_lsn = 0;
1489 xfs_lsn_t last_sync_lsn;
1492 int threshold_block;
1493 int threshold_cycle;
1496 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1498 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1499 free_blocks = BTOBBT(free_bytes);
1502 * Set the threshold for the minimum number of free blocks in the
1503 * log to the maximum of what the caller needs, one quarter of the
1504 * log, and 256 blocks.
1506 free_threshold = BTOBB(need_bytes);
1507 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1508 free_threshold = MAX(free_threshold, 256);
1509 if (free_blocks >= free_threshold)
1512 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1514 threshold_block += free_threshold;
1515 if (threshold_block >= log->l_logBBsize) {
1516 threshold_block -= log->l_logBBsize;
1517 threshold_cycle += 1;
1519 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1522 * Don't pass in an lsn greater than the lsn of the last
1523 * log record known to be on disk. Use a snapshot of the last sync lsn
1524 * so that it doesn't change between the compare and the set.
1526 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1527 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1528 threshold_lsn = last_sync_lsn;
1531 * Get the transaction layer to kick the dirty buffers out to
1532 * disk asynchronously. No point in trying to do this if
1533 * the filesystem is shutting down.
1535 if (!XLOG_FORCED_SHUTDOWN(log))
1536 xfs_ail_push(log->l_ailp, threshold_lsn);
1540 * Stamp cycle number in every block
1545 struct xlog_in_core *iclog,
1549 int size = iclog->ic_offset + roundoff;
1553 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1555 dp = iclog->ic_datap;
1556 for (i = 0; i < BTOBB(size); i++) {
1557 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1559 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1560 *(__be32 *)dp = cycle_lsn;
1564 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1565 xlog_in_core_2_t *xhdr = iclog->ic_data;
1567 for ( ; i < BTOBB(size); i++) {
1568 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1569 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1570 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1571 *(__be32 *)dp = cycle_lsn;
1575 for (i = 1; i < log->l_iclog_heads; i++)
1576 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1581 * Calculate the checksum for a log buffer.
1583 * This is a little more complicated than it should be because the various
1584 * headers and the actual data are non-contiguous.
1589 struct xlog_rec_header *rhead,
1595 /* first generate the crc for the record header ... */
1596 crc = xfs_start_cksum((char *)rhead,
1597 sizeof(struct xlog_rec_header),
1598 offsetof(struct xlog_rec_header, h_crc));
1600 /* ... then for additional cycle data for v2 logs ... */
1601 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1602 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1605 for (i = 1; i < log->l_iclog_heads; i++) {
1606 crc = crc32c(crc, &xhdr[i].hic_xheader,
1607 sizeof(struct xlog_rec_ext_header));
1611 /* ... and finally for the payload */
1612 crc = crc32c(crc, dp, size);
1614 return xfs_end_cksum(crc);
1618 * The bdstrat callback function for log bufs. This gives us a central
1619 * place to trap bufs in case we get hit by a log I/O error and need to
1620 * shutdown. Actually, in practice, even when we didn't get a log error,
1621 * we transition the iclogs to IOERROR state *after* flushing all existing
1622 * iclogs to disk. This is because we don't want anymore new transactions to be
1623 * started or completed afterwards.
1629 struct xlog_in_core *iclog = bp->b_fspriv;
1631 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1632 xfs_buf_ioerror(bp, EIO);
1634 xfs_buf_ioend(bp, 0);
1636 * It would seem logical to return EIO here, but we rely on
1637 * the log state machine to propagate I/O errors instead of
1643 xfs_buf_iorequest(bp);
1648 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1649 * fashion. Previously, we should have moved the current iclog
1650 * ptr in the log to point to the next available iclog. This allows further
1651 * write to continue while this code syncs out an iclog ready to go.
1652 * Before an in-core log can be written out, the data section must be scanned
1653 * to save away the 1st word of each BBSIZE block into the header. We replace
1654 * it with the current cycle count. Each BBSIZE block is tagged with the
1655 * cycle count because there in an implicit assumption that drives will
1656 * guarantee that entire 512 byte blocks get written at once. In other words,
1657 * we can't have part of a 512 byte block written and part not written. By
1658 * tagging each block, we will know which blocks are valid when recovering
1659 * after an unclean shutdown.
1661 * This routine is single threaded on the iclog. No other thread can be in
1662 * this routine with the same iclog. Changing contents of iclog can there-
1663 * fore be done without grabbing the state machine lock. Updating the global
1664 * log will require grabbing the lock though.
1666 * The entire log manager uses a logical block numbering scheme. Only
1667 * log_sync (and then only bwrite()) know about the fact that the log may
1668 * not start with block zero on a given device. The log block start offset
1669 * is added immediately before calling bwrite().
1675 struct xlog_in_core *iclog)
1679 uint count; /* byte count of bwrite */
1680 uint count_init; /* initial count before roundup */
1681 int roundoff; /* roundoff to BB or stripe */
1682 int split = 0; /* split write into two regions */
1684 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1687 XFS_STATS_INC(xs_log_writes);
1688 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1690 /* Add for LR header */
1691 count_init = log->l_iclog_hsize + iclog->ic_offset;
1693 /* Round out the log write size */
1694 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1695 /* we have a v2 stripe unit to use */
1696 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1698 count = BBTOB(BTOBB(count_init));
1700 roundoff = count - count_init;
1701 ASSERT(roundoff >= 0);
1702 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1703 roundoff < log->l_mp->m_sb.sb_logsunit)
1705 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1706 roundoff < BBTOB(1)));
1708 /* move grant heads by roundoff in sync */
1709 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1710 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1712 /* put cycle number in every block */
1713 xlog_pack_data(log, iclog, roundoff);
1715 /* real byte length */
1716 size = iclog->ic_offset;
1719 iclog->ic_header.h_len = cpu_to_be32(size);
1722 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1724 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1726 /* Do we need to split this write into 2 parts? */
1727 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1730 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1731 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1732 iclog->ic_bwritecnt = 2;
1735 * Bump the cycle numbers at the start of each block in the
1736 * part of the iclog that ends up in the buffer that gets
1737 * written to the start of the log.
1739 * Watch out for the header magic number case, though.
1741 dptr = (char *)&iclog->ic_header + count;
1742 for (i = 0; i < split; i += BBSIZE) {
1743 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1744 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1746 *(__be32 *)dptr = cpu_to_be32(cycle);
1751 iclog->ic_bwritecnt = 1;
1754 /* calculcate the checksum */
1755 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1756 iclog->ic_datap, size);
1758 bp->b_io_length = BTOBB(count);
1759 bp->b_fspriv = iclog;
1760 XFS_BUF_ZEROFLAGS(bp);
1762 bp->b_flags |= XBF_SYNCIO;
1764 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1765 bp->b_flags |= XBF_FUA;
1768 * Flush the data device before flushing the log to make
1769 * sure all meta data written back from the AIL actually made
1770 * it to disk before stamping the new log tail LSN into the
1771 * log buffer. For an external log we need to issue the
1772 * flush explicitly, and unfortunately synchronously here;
1773 * for an internal log we can simply use the block layer
1774 * state machine for preflushes.
1776 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1777 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1779 bp->b_flags |= XBF_FLUSH;
1782 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1783 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1785 xlog_verify_iclog(log, iclog, count, true);
1787 /* account for log which doesn't start at block #0 */
1788 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1790 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1795 error = xlog_bdstrat(bp);
1797 xfs_buf_ioerror_alert(bp, "xlog_sync");
1801 bp = iclog->ic_log->l_xbuf;
1802 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1803 xfs_buf_associate_memory(bp,
1804 (char *)&iclog->ic_header + count, split);
1805 bp->b_fspriv = iclog;
1806 XFS_BUF_ZEROFLAGS(bp);
1808 bp->b_flags |= XBF_SYNCIO;
1809 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1810 bp->b_flags |= XBF_FUA;
1812 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1813 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1815 /* account for internal log which doesn't start at block #0 */
1816 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1818 error = xlog_bdstrat(bp);
1820 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1828 * Deallocate a log structure
1834 xlog_in_core_t *iclog, *next_iclog;
1837 xlog_cil_destroy(log);
1840 * always need to ensure that the extra buffer does not point to memory
1841 * owned by another log buffer before we free it.
1843 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1844 xfs_buf_free(log->l_xbuf);
1846 iclog = log->l_iclog;
1847 for (i=0; i<log->l_iclog_bufs; i++) {
1848 xfs_buf_free(iclog->ic_bp);
1849 next_iclog = iclog->ic_next;
1853 spinlock_destroy(&log->l_icloglock);
1855 log->l_mp->m_log = NULL;
1857 } /* xlog_dealloc_log */
1860 * Update counters atomically now that memcpy is done.
1864 xlog_state_finish_copy(
1866 struct xlog_in_core *iclog,
1870 spin_lock(&log->l_icloglock);
1872 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1873 iclog->ic_offset += copy_bytes;
1875 spin_unlock(&log->l_icloglock);
1876 } /* xlog_state_finish_copy */
1882 * print out info relating to regions written which consume
1887 struct xfs_mount *mp,
1888 struct xlog_ticket *ticket)
1891 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1893 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1894 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1915 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1959 "xlog_write: reservation summary:\n"
1960 " trans type = %s (%u)\n"
1961 " unit res = %d bytes\n"
1962 " current res = %d bytes\n"
1963 " total reg = %u bytes (o/flow = %u bytes)\n"
1964 " ophdrs = %u (ophdr space = %u bytes)\n"
1965 " ophdr + reg = %u bytes\n"
1966 " num regions = %u\n",
1967 ((ticket->t_trans_type <= 0 ||
1968 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1969 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1970 ticket->t_trans_type,
1973 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1974 ticket->t_res_num_ophdrs, ophdr_spc,
1975 ticket->t_res_arr_sum +
1976 ticket->t_res_o_flow + ophdr_spc,
1979 for (i = 0; i < ticket->t_res_num; i++) {
1980 uint r_type = ticket->t_res_arr[i].r_type;
1981 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1982 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1983 "bad-rtype" : res_type_str[r_type-1]),
1984 ticket->t_res_arr[i].r_len);
1987 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1988 "xlog_write: reservation ran out. Need to up reservation");
1989 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1993 * Calculate the potential space needed by the log vector. Each region gets
1994 * its own xlog_op_header_t and may need to be double word aligned.
1997 xlog_write_calc_vec_length(
1998 struct xlog_ticket *ticket,
1999 struct xfs_log_vec *log_vector)
2001 struct xfs_log_vec *lv;
2006 /* acct for start rec of xact */
2007 if (ticket->t_flags & XLOG_TIC_INITED)
2010 for (lv = log_vector; lv; lv = lv->lv_next) {
2011 /* we don't write ordered log vectors */
2012 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2015 headers += lv->lv_niovecs;
2017 for (i = 0; i < lv->lv_niovecs; i++) {
2018 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2021 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2025 ticket->t_res_num_ophdrs += headers;
2026 len += headers * sizeof(struct xlog_op_header);
2032 * If first write for transaction, insert start record We can't be trying to
2033 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2036 xlog_write_start_rec(
2037 struct xlog_op_header *ophdr,
2038 struct xlog_ticket *ticket)
2040 if (!(ticket->t_flags & XLOG_TIC_INITED))
2043 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2044 ophdr->oh_clientid = ticket->t_clientid;
2046 ophdr->oh_flags = XLOG_START_TRANS;
2049 ticket->t_flags &= ~XLOG_TIC_INITED;
2051 return sizeof(struct xlog_op_header);
2054 static xlog_op_header_t *
2055 xlog_write_setup_ophdr(
2057 struct xlog_op_header *ophdr,
2058 struct xlog_ticket *ticket,
2061 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2062 ophdr->oh_clientid = ticket->t_clientid;
2065 /* are we copying a commit or unmount record? */
2066 ophdr->oh_flags = flags;
2069 * We've seen logs corrupted with bad transaction client ids. This
2070 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2071 * and shut down the filesystem.
2073 switch (ophdr->oh_clientid) {
2074 case XFS_TRANSACTION:
2080 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2081 ophdr->oh_clientid, ticket);
2089 * Set up the parameters of the region copy into the log. This has
2090 * to handle region write split across multiple log buffers - this
2091 * state is kept external to this function so that this code can
2092 * be written in an obvious, self documenting manner.
2095 xlog_write_setup_copy(
2096 struct xlog_ticket *ticket,
2097 struct xlog_op_header *ophdr,
2098 int space_available,
2102 int *last_was_partial_copy,
2103 int *bytes_consumed)
2107 still_to_copy = space_required - *bytes_consumed;
2108 *copy_off = *bytes_consumed;
2110 if (still_to_copy <= space_available) {
2111 /* write of region completes here */
2112 *copy_len = still_to_copy;
2113 ophdr->oh_len = cpu_to_be32(*copy_len);
2114 if (*last_was_partial_copy)
2115 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2116 *last_was_partial_copy = 0;
2117 *bytes_consumed = 0;
2121 /* partial write of region, needs extra log op header reservation */
2122 *copy_len = space_available;
2123 ophdr->oh_len = cpu_to_be32(*copy_len);
2124 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2125 if (*last_was_partial_copy)
2126 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2127 *bytes_consumed += *copy_len;
2128 (*last_was_partial_copy)++;
2130 /* account for new log op header */
2131 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2132 ticket->t_res_num_ophdrs++;
2134 return sizeof(struct xlog_op_header);
2138 xlog_write_copy_finish(
2140 struct xlog_in_core *iclog,
2145 int *partial_copy_len,
2147 struct xlog_in_core **commit_iclog)
2149 if (*partial_copy) {
2151 * This iclog has already been marked WANT_SYNC by
2152 * xlog_state_get_iclog_space.
2154 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2157 return xlog_state_release_iclog(log, iclog);
2161 *partial_copy_len = 0;
2163 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2164 /* no more space in this iclog - push it. */
2165 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2169 spin_lock(&log->l_icloglock);
2170 xlog_state_want_sync(log, iclog);
2171 spin_unlock(&log->l_icloglock);
2174 return xlog_state_release_iclog(log, iclog);
2175 ASSERT(flags & XLOG_COMMIT_TRANS);
2176 *commit_iclog = iclog;
2183 * Write some region out to in-core log
2185 * This will be called when writing externally provided regions or when
2186 * writing out a commit record for a given transaction.
2188 * General algorithm:
2189 * 1. Find total length of this write. This may include adding to the
2190 * lengths passed in.
2191 * 2. Check whether we violate the tickets reservation.
2192 * 3. While writing to this iclog
2193 * A. Reserve as much space in this iclog as can get
2194 * B. If this is first write, save away start lsn
2195 * C. While writing this region:
2196 * 1. If first write of transaction, write start record
2197 * 2. Write log operation header (header per region)
2198 * 3. Find out if we can fit entire region into this iclog
2199 * 4. Potentially, verify destination memcpy ptr
2200 * 5. Memcpy (partial) region
2201 * 6. If partial copy, release iclog; otherwise, continue
2202 * copying more regions into current iclog
2203 * 4. Mark want sync bit (in simulation mode)
2204 * 5. Release iclog for potential flush to on-disk log.
2207 * 1. Panic if reservation is overrun. This should never happen since
2208 * reservation amounts are generated internal to the filesystem.
2210 * 1. Tickets are single threaded data structures.
2211 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2212 * syncing routine. When a single log_write region needs to span
2213 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2214 * on all log operation writes which don't contain the end of the
2215 * region. The XLOG_END_TRANS bit is used for the in-core log
2216 * operation which contains the end of the continued log_write region.
2217 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2218 * we don't really know exactly how much space will be used. As a result,
2219 * we don't update ic_offset until the end when we know exactly how many
2220 * bytes have been written out.
2225 struct xfs_log_vec *log_vector,
2226 struct xlog_ticket *ticket,
2227 xfs_lsn_t *start_lsn,
2228 struct xlog_in_core **commit_iclog,
2231 struct xlog_in_core *iclog = NULL;
2232 struct xfs_log_iovec *vecp;
2233 struct xfs_log_vec *lv;
2236 int partial_copy = 0;
2237 int partial_copy_len = 0;
2245 len = xlog_write_calc_vec_length(ticket, log_vector);
2248 * Region headers and bytes are already accounted for.
2249 * We only need to take into account start records and
2250 * split regions in this function.
2252 if (ticket->t_flags & XLOG_TIC_INITED)
2253 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2256 * Commit record headers need to be accounted for. These
2257 * come in as separate writes so are easy to detect.
2259 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2260 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2262 if (ticket->t_curr_res < 0)
2263 xlog_print_tic_res(log->l_mp, ticket);
2267 vecp = lv->lv_iovecp;
2268 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2272 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2273 &contwr, &log_offset);
2277 ASSERT(log_offset <= iclog->ic_size - 1);
2278 ptr = iclog->ic_datap + log_offset;
2280 /* start_lsn is the first lsn written to. That's all we need. */
2282 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2285 * This loop writes out as many regions as can fit in the amount
2286 * of space which was allocated by xlog_state_get_iclog_space().
2288 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2289 struct xfs_log_iovec *reg;
2290 struct xlog_op_header *ophdr;
2294 bool ordered = false;
2296 /* ordered log vectors have no regions to write */
2297 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2298 ASSERT(lv->lv_niovecs == 0);
2304 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2305 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2307 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2308 if (start_rec_copy) {
2310 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2314 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2316 return XFS_ERROR(EIO);
2318 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2319 sizeof(struct xlog_op_header));
2321 len += xlog_write_setup_copy(ticket, ophdr,
2322 iclog->ic_size-log_offset,
2324 ©_off, ©_len,
2327 xlog_verify_dest_ptr(log, ptr);
2330 ASSERT(copy_len >= 0);
2331 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2332 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2334 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2336 data_cnt += contwr ? copy_len : 0;
2338 error = xlog_write_copy_finish(log, iclog, flags,
2339 &record_cnt, &data_cnt,
2348 * if we had a partial copy, we need to get more iclog
2349 * space but we don't want to increment the region
2350 * index because there is still more is this region to
2353 * If we completed writing this region, and we flushed
2354 * the iclog (indicated by resetting of the record
2355 * count), then we also need to get more log space. If
2356 * this was the last record, though, we are done and
2362 if (++index == lv->lv_niovecs) {
2367 vecp = lv->lv_iovecp;
2369 if (record_cnt == 0 && ordered == false) {
2379 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2381 return xlog_state_release_iclog(log, iclog);
2383 ASSERT(flags & XLOG_COMMIT_TRANS);
2384 *commit_iclog = iclog;
2389 /*****************************************************************************
2391 * State Machine functions
2393 *****************************************************************************
2396 /* Clean iclogs starting from the head. This ordering must be
2397 * maintained, so an iclog doesn't become ACTIVE beyond one that
2398 * is SYNCING. This is also required to maintain the notion that we use
2399 * a ordered wait queue to hold off would be writers to the log when every
2400 * iclog is trying to sync to disk.
2402 * State Change: DIRTY -> ACTIVE
2405 xlog_state_clean_log(
2408 xlog_in_core_t *iclog;
2411 iclog = log->l_iclog;
2413 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2414 iclog->ic_state = XLOG_STATE_ACTIVE;
2415 iclog->ic_offset = 0;
2416 ASSERT(iclog->ic_callback == NULL);
2418 * If the number of ops in this iclog indicate it just
2419 * contains the dummy transaction, we can
2420 * change state into IDLE (the second time around).
2421 * Otherwise we should change the state into
2423 * We don't need to cover the dummy.
2426 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2431 * We have two dirty iclogs so start over
2432 * This could also be num of ops indicates
2433 * this is not the dummy going out.
2437 iclog->ic_header.h_num_logops = 0;
2438 memset(iclog->ic_header.h_cycle_data, 0,
2439 sizeof(iclog->ic_header.h_cycle_data));
2440 iclog->ic_header.h_lsn = 0;
2441 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2444 break; /* stop cleaning */
2445 iclog = iclog->ic_next;
2446 } while (iclog != log->l_iclog);
2448 /* log is locked when we are called */
2450 * Change state for the dummy log recording.
2451 * We usually go to NEED. But we go to NEED2 if the changed indicates
2452 * we are done writing the dummy record.
2453 * If we are done with the second dummy recored (DONE2), then
2457 switch (log->l_covered_state) {
2458 case XLOG_STATE_COVER_IDLE:
2459 case XLOG_STATE_COVER_NEED:
2460 case XLOG_STATE_COVER_NEED2:
2461 log->l_covered_state = XLOG_STATE_COVER_NEED;
2464 case XLOG_STATE_COVER_DONE:
2466 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2468 log->l_covered_state = XLOG_STATE_COVER_NEED;
2471 case XLOG_STATE_COVER_DONE2:
2473 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2475 log->l_covered_state = XLOG_STATE_COVER_NEED;
2482 } /* xlog_state_clean_log */
2485 xlog_get_lowest_lsn(
2488 xlog_in_core_t *lsn_log;
2489 xfs_lsn_t lowest_lsn, lsn;
2491 lsn_log = log->l_iclog;
2494 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2495 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2496 if ((lsn && !lowest_lsn) ||
2497 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2501 lsn_log = lsn_log->ic_next;
2502 } while (lsn_log != log->l_iclog);
2508 xlog_state_do_callback(
2511 struct xlog_in_core *ciclog)
2513 xlog_in_core_t *iclog;
2514 xlog_in_core_t *first_iclog; /* used to know when we've
2515 * processed all iclogs once */
2516 xfs_log_callback_t *cb, *cb_next;
2518 xfs_lsn_t lowest_lsn;
2519 int ioerrors; /* counter: iclogs with errors */
2520 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2521 int funcdidcallbacks; /* flag: function did callbacks */
2522 int repeats; /* for issuing console warnings if
2523 * looping too many times */
2526 spin_lock(&log->l_icloglock);
2527 first_iclog = iclog = log->l_iclog;
2529 funcdidcallbacks = 0;
2534 * Scan all iclogs starting with the one pointed to by the
2535 * log. Reset this starting point each time the log is
2536 * unlocked (during callbacks).
2538 * Keep looping through iclogs until one full pass is made
2539 * without running any callbacks.
2541 first_iclog = log->l_iclog;
2542 iclog = log->l_iclog;
2543 loopdidcallbacks = 0;
2548 /* skip all iclogs in the ACTIVE & DIRTY states */
2549 if (iclog->ic_state &
2550 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2551 iclog = iclog->ic_next;
2556 * Between marking a filesystem SHUTDOWN and stopping
2557 * the log, we do flush all iclogs to disk (if there
2558 * wasn't a log I/O error). So, we do want things to
2559 * go smoothly in case of just a SHUTDOWN w/o a
2562 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2564 * Can only perform callbacks in order. Since
2565 * this iclog is not in the DONE_SYNC/
2566 * DO_CALLBACK state, we skip the rest and
2567 * just try to clean up. If we set our iclog
2568 * to DO_CALLBACK, we will not process it when
2569 * we retry since a previous iclog is in the
2570 * CALLBACK and the state cannot change since
2571 * we are holding the l_icloglock.
2573 if (!(iclog->ic_state &
2574 (XLOG_STATE_DONE_SYNC |
2575 XLOG_STATE_DO_CALLBACK))) {
2576 if (ciclog && (ciclog->ic_state ==
2577 XLOG_STATE_DONE_SYNC)) {
2578 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2583 * We now have an iclog that is in either the
2584 * DO_CALLBACK or DONE_SYNC states. The other
2585 * states (WANT_SYNC, SYNCING, or CALLBACK were
2586 * caught by the above if and are going to
2587 * clean (i.e. we aren't doing their callbacks)
2592 * We will do one more check here to see if we
2593 * have chased our tail around.
2596 lowest_lsn = xlog_get_lowest_lsn(log);
2598 XFS_LSN_CMP(lowest_lsn,
2599 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2600 iclog = iclog->ic_next;
2601 continue; /* Leave this iclog for
2605 iclog->ic_state = XLOG_STATE_CALLBACK;
2609 * Completion of a iclog IO does not imply that
2610 * a transaction has completed, as transactions
2611 * can be large enough to span many iclogs. We
2612 * cannot change the tail of the log half way
2613 * through a transaction as this may be the only
2614 * transaction in the log and moving th etail to
2615 * point to the middle of it will prevent
2616 * recovery from finding the start of the
2617 * transaction. Hence we should only update the
2618 * last_sync_lsn if this iclog contains
2619 * transaction completion callbacks on it.
2621 * We have to do this before we drop the
2622 * icloglock to ensure we are the only one that
2625 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2626 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2627 if (iclog->ic_callback)
2628 atomic64_set(&log->l_last_sync_lsn,
2629 be64_to_cpu(iclog->ic_header.h_lsn));
2634 spin_unlock(&log->l_icloglock);
2637 * Keep processing entries in the callback list until
2638 * we come around and it is empty. We need to
2639 * atomically see that the list is empty and change the
2640 * state to DIRTY so that we don't miss any more
2641 * callbacks being added.
2643 spin_lock(&iclog->ic_callback_lock);
2644 cb = iclog->ic_callback;
2646 iclog->ic_callback_tail = &(iclog->ic_callback);
2647 iclog->ic_callback = NULL;
2648 spin_unlock(&iclog->ic_callback_lock);
2650 /* perform callbacks in the order given */
2651 for (; cb; cb = cb_next) {
2652 cb_next = cb->cb_next;
2653 cb->cb_func(cb->cb_arg, aborted);
2655 spin_lock(&iclog->ic_callback_lock);
2656 cb = iclog->ic_callback;
2662 spin_lock(&log->l_icloglock);
2663 ASSERT(iclog->ic_callback == NULL);
2664 spin_unlock(&iclog->ic_callback_lock);
2665 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2666 iclog->ic_state = XLOG_STATE_DIRTY;
2669 * Transition from DIRTY to ACTIVE if applicable.
2670 * NOP if STATE_IOERROR.
2672 xlog_state_clean_log(log);
2674 /* wake up threads waiting in xfs_log_force() */
2675 wake_up_all(&iclog->ic_force_wait);
2677 iclog = iclog->ic_next;
2678 } while (first_iclog != iclog);
2680 if (repeats > 5000) {
2681 flushcnt += repeats;
2684 "%s: possible infinite loop (%d iterations)",
2685 __func__, flushcnt);
2687 } while (!ioerrors && loopdidcallbacks);
2690 * make one last gasp attempt to see if iclogs are being left in
2694 if (funcdidcallbacks) {
2695 first_iclog = iclog = log->l_iclog;
2697 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2699 * Terminate the loop if iclogs are found in states
2700 * which will cause other threads to clean up iclogs.
2702 * SYNCING - i/o completion will go through logs
2703 * DONE_SYNC - interrupt thread should be waiting for
2705 * IOERROR - give up hope all ye who enter here
2707 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2708 iclog->ic_state == XLOG_STATE_SYNCING ||
2709 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2710 iclog->ic_state == XLOG_STATE_IOERROR )
2712 iclog = iclog->ic_next;
2713 } while (first_iclog != iclog);
2717 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2719 spin_unlock(&log->l_icloglock);
2722 wake_up_all(&log->l_flush_wait);
2727 * Finish transitioning this iclog to the dirty state.
2729 * Make sure that we completely execute this routine only when this is
2730 * the last call to the iclog. There is a good chance that iclog flushes,
2731 * when we reach the end of the physical log, get turned into 2 separate
2732 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2733 * routine. By using the reference count bwritecnt, we guarantee that only
2734 * the second completion goes through.
2736 * Callbacks could take time, so they are done outside the scope of the
2737 * global state machine log lock.
2740 xlog_state_done_syncing(
2741 xlog_in_core_t *iclog,
2744 struct xlog *log = iclog->ic_log;
2746 spin_lock(&log->l_icloglock);
2748 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2749 iclog->ic_state == XLOG_STATE_IOERROR);
2750 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2751 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2755 * If we got an error, either on the first buffer, or in the case of
2756 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2757 * and none should ever be attempted to be written to disk
2760 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2761 if (--iclog->ic_bwritecnt == 1) {
2762 spin_unlock(&log->l_icloglock);
2765 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2769 * Someone could be sleeping prior to writing out the next
2770 * iclog buffer, we wake them all, one will get to do the
2771 * I/O, the others get to wait for the result.
2773 wake_up_all(&iclog->ic_write_wait);
2774 spin_unlock(&log->l_icloglock);
2775 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2776 } /* xlog_state_done_syncing */
2780 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2781 * sleep. We wait on the flush queue on the head iclog as that should be
2782 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2783 * we will wait here and all new writes will sleep until a sync completes.
2785 * The in-core logs are used in a circular fashion. They are not used
2786 * out-of-order even when an iclog past the head is free.
2789 * * log_offset where xlog_write() can start writing into the in-core
2791 * * in-core log pointer to which xlog_write() should write.
2792 * * boolean indicating this is a continued write to an in-core log.
2793 * If this is the last write, then the in-core log's offset field
2794 * needs to be incremented, depending on the amount of data which
2798 xlog_state_get_iclog_space(
2801 struct xlog_in_core **iclogp,
2802 struct xlog_ticket *ticket,
2803 int *continued_write,
2807 xlog_rec_header_t *head;
2808 xlog_in_core_t *iclog;
2812 spin_lock(&log->l_icloglock);
2813 if (XLOG_FORCED_SHUTDOWN(log)) {
2814 spin_unlock(&log->l_icloglock);
2815 return XFS_ERROR(EIO);
2818 iclog = log->l_iclog;
2819 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2820 XFS_STATS_INC(xs_log_noiclogs);
2822 /* Wait for log writes to have flushed */
2823 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2827 head = &iclog->ic_header;
2829 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2830 log_offset = iclog->ic_offset;
2832 /* On the 1st write to an iclog, figure out lsn. This works
2833 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2834 * committing to. If the offset is set, that's how many blocks
2837 if (log_offset == 0) {
2838 ticket->t_curr_res -= log->l_iclog_hsize;
2839 xlog_tic_add_region(ticket,
2841 XLOG_REG_TYPE_LRHEADER);
2842 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2843 head->h_lsn = cpu_to_be64(
2844 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2845 ASSERT(log->l_curr_block >= 0);
2848 /* If there is enough room to write everything, then do it. Otherwise,
2849 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2850 * bit is on, so this will get flushed out. Don't update ic_offset
2851 * until you know exactly how many bytes get copied. Therefore, wait
2852 * until later to update ic_offset.
2854 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2855 * can fit into remaining data section.
2857 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2858 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2861 * If I'm the only one writing to this iclog, sync it to disk.
2862 * We need to do an atomic compare and decrement here to avoid
2863 * racing with concurrent atomic_dec_and_lock() calls in
2864 * xlog_state_release_iclog() when there is more than one
2865 * reference to the iclog.
2867 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2868 /* we are the only one */
2869 spin_unlock(&log->l_icloglock);
2870 error = xlog_state_release_iclog(log, iclog);
2874 spin_unlock(&log->l_icloglock);
2879 /* Do we have enough room to write the full amount in the remainder
2880 * of this iclog? Or must we continue a write on the next iclog and
2881 * mark this iclog as completely taken? In the case where we switch
2882 * iclogs (to mark it taken), this particular iclog will release/sync
2883 * to disk in xlog_write().
2885 if (len <= iclog->ic_size - iclog->ic_offset) {
2886 *continued_write = 0;
2887 iclog->ic_offset += len;
2889 *continued_write = 1;
2890 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2894 ASSERT(iclog->ic_offset <= iclog->ic_size);
2895 spin_unlock(&log->l_icloglock);
2897 *logoffsetp = log_offset;
2899 } /* xlog_state_get_iclog_space */
2901 /* The first cnt-1 times through here we don't need to
2902 * move the grant write head because the permanent
2903 * reservation has reserved cnt times the unit amount.
2904 * Release part of current permanent unit reservation and
2905 * reset current reservation to be one units worth. Also
2906 * move grant reservation head forward.
2909 xlog_regrant_reserve_log_space(
2911 struct xlog_ticket *ticket)
2913 trace_xfs_log_regrant_reserve_enter(log, ticket);
2915 if (ticket->t_cnt > 0)
2918 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2919 ticket->t_curr_res);
2920 xlog_grant_sub_space(log, &log->l_write_head.grant,
2921 ticket->t_curr_res);
2922 ticket->t_curr_res = ticket->t_unit_res;
2923 xlog_tic_reset_res(ticket);
2925 trace_xfs_log_regrant_reserve_sub(log, ticket);
2927 /* just return if we still have some of the pre-reserved space */
2928 if (ticket->t_cnt > 0)
2931 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2932 ticket->t_unit_res);
2934 trace_xfs_log_regrant_reserve_exit(log, ticket);
2936 ticket->t_curr_res = ticket->t_unit_res;
2937 xlog_tic_reset_res(ticket);
2938 } /* xlog_regrant_reserve_log_space */
2942 * Give back the space left from a reservation.
2944 * All the information we need to make a correct determination of space left
2945 * is present. For non-permanent reservations, things are quite easy. The
2946 * count should have been decremented to zero. We only need to deal with the
2947 * space remaining in the current reservation part of the ticket. If the
2948 * ticket contains a permanent reservation, there may be left over space which
2949 * needs to be released. A count of N means that N-1 refills of the current
2950 * reservation can be done before we need to ask for more space. The first
2951 * one goes to fill up the first current reservation. Once we run out of
2952 * space, the count will stay at zero and the only space remaining will be
2953 * in the current reservation field.
2956 xlog_ungrant_log_space(
2958 struct xlog_ticket *ticket)
2962 if (ticket->t_cnt > 0)
2965 trace_xfs_log_ungrant_enter(log, ticket);
2966 trace_xfs_log_ungrant_sub(log, ticket);
2969 * If this is a permanent reservation ticket, we may be able to free
2970 * up more space based on the remaining count.
2972 bytes = ticket->t_curr_res;
2973 if (ticket->t_cnt > 0) {
2974 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2975 bytes += ticket->t_unit_res*ticket->t_cnt;
2978 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2979 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2981 trace_xfs_log_ungrant_exit(log, ticket);
2983 xfs_log_space_wake(log->l_mp);
2987 * Flush iclog to disk if this is the last reference to the given iclog and
2988 * the WANT_SYNC bit is set.
2990 * When this function is entered, the iclog is not necessarily in the
2991 * WANT_SYNC state. It may be sitting around waiting to get filled.
2996 xlog_state_release_iclog(
2998 struct xlog_in_core *iclog)
3000 int sync = 0; /* do we sync? */
3002 if (iclog->ic_state & XLOG_STATE_IOERROR)
3003 return XFS_ERROR(EIO);
3005 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3006 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3009 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3010 spin_unlock(&log->l_icloglock);
3011 return XFS_ERROR(EIO);
3013 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3014 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3016 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3017 /* update tail before writing to iclog */
3018 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3020 iclog->ic_state = XLOG_STATE_SYNCING;
3021 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3022 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3023 /* cycle incremented when incrementing curr_block */
3025 spin_unlock(&log->l_icloglock);
3028 * We let the log lock go, so it's possible that we hit a log I/O
3029 * error or some other SHUTDOWN condition that marks the iclog
3030 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3031 * this iclog has consistent data, so we ignore IOERROR
3032 * flags after this point.
3035 return xlog_sync(log, iclog);
3037 } /* xlog_state_release_iclog */
3041 * This routine will mark the current iclog in the ring as WANT_SYNC
3042 * and move the current iclog pointer to the next iclog in the ring.
3043 * When this routine is called from xlog_state_get_iclog_space(), the
3044 * exact size of the iclog has not yet been determined. All we know is
3045 * that every data block. We have run out of space in this log record.
3048 xlog_state_switch_iclogs(
3050 struct xlog_in_core *iclog,
3053 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3055 eventual_size = iclog->ic_offset;
3056 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3057 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3058 log->l_prev_block = log->l_curr_block;
3059 log->l_prev_cycle = log->l_curr_cycle;
3061 /* roll log?: ic_offset changed later */
3062 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3064 /* Round up to next log-sunit */
3065 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3066 log->l_mp->m_sb.sb_logsunit > 1) {
3067 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3068 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3071 if (log->l_curr_block >= log->l_logBBsize) {
3072 log->l_curr_cycle++;
3073 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3074 log->l_curr_cycle++;
3075 log->l_curr_block -= log->l_logBBsize;
3076 ASSERT(log->l_curr_block >= 0);
3078 ASSERT(iclog == log->l_iclog);
3079 log->l_iclog = iclog->ic_next;
3080 } /* xlog_state_switch_iclogs */
3083 * Write out all data in the in-core log as of this exact moment in time.
3085 * Data may be written to the in-core log during this call. However,
3086 * we don't guarantee this data will be written out. A change from past
3087 * implementation means this routine will *not* write out zero length LRs.
3089 * Basically, we try and perform an intelligent scan of the in-core logs.
3090 * If we determine there is no flushable data, we just return. There is no
3091 * flushable data if:
3093 * 1. the current iclog is active and has no data; the previous iclog
3094 * is in the active or dirty state.
3095 * 2. the current iclog is drity, and the previous iclog is in the
3096 * active or dirty state.
3100 * 1. the current iclog is not in the active nor dirty state.
3101 * 2. the current iclog dirty, and the previous iclog is not in the
3102 * active nor dirty state.
3103 * 3. the current iclog is active, and there is another thread writing
3104 * to this particular iclog.
3105 * 4. a) the current iclog is active and has no other writers
3106 * b) when we return from flushing out this iclog, it is still
3107 * not in the active nor dirty state.
3111 struct xfs_mount *mp,
3115 struct xlog *log = mp->m_log;
3116 struct xlog_in_core *iclog;
3119 XFS_STATS_INC(xs_log_force);
3121 xlog_cil_force(log);
3123 spin_lock(&log->l_icloglock);
3125 iclog = log->l_iclog;
3126 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3127 spin_unlock(&log->l_icloglock);
3128 return XFS_ERROR(EIO);
3131 /* If the head iclog is not active nor dirty, we just attach
3132 * ourselves to the head and go to sleep.
3134 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3135 iclog->ic_state == XLOG_STATE_DIRTY) {
3137 * If the head is dirty or (active and empty), then
3138 * we need to look at the previous iclog. If the previous
3139 * iclog is active or dirty we are done. There is nothing
3140 * to sync out. Otherwise, we attach ourselves to the
3141 * previous iclog and go to sleep.
3143 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3144 (atomic_read(&iclog->ic_refcnt) == 0
3145 && iclog->ic_offset == 0)) {
3146 iclog = iclog->ic_prev;
3147 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3148 iclog->ic_state == XLOG_STATE_DIRTY)
3153 if (atomic_read(&iclog->ic_refcnt) == 0) {
3154 /* We are the only one with access to this
3155 * iclog. Flush it out now. There should
3156 * be a roundoff of zero to show that someone
3157 * has already taken care of the roundoff from
3158 * the previous sync.
3160 atomic_inc(&iclog->ic_refcnt);
3161 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3162 xlog_state_switch_iclogs(log, iclog, 0);
3163 spin_unlock(&log->l_icloglock);
3165 if (xlog_state_release_iclog(log, iclog))
3166 return XFS_ERROR(EIO);
3170 spin_lock(&log->l_icloglock);
3171 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3172 iclog->ic_state != XLOG_STATE_DIRTY)
3177 /* Someone else is writing to this iclog.
3178 * Use its call to flush out the data. However,
3179 * the other thread may not force out this LR,
3180 * so we mark it WANT_SYNC.
3182 xlog_state_switch_iclogs(log, iclog, 0);
3188 /* By the time we come around again, the iclog could've been filled
3189 * which would give it another lsn. If we have a new lsn, just
3190 * return because the relevant data has been flushed.
3193 if (flags & XFS_LOG_SYNC) {
3195 * We must check if we're shutting down here, before
3196 * we wait, while we're holding the l_icloglock.
3197 * Then we check again after waking up, in case our
3198 * sleep was disturbed by a bad news.
3200 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3201 spin_unlock(&log->l_icloglock);
3202 return XFS_ERROR(EIO);
3204 XFS_STATS_INC(xs_log_force_sleep);
3205 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3207 * No need to grab the log lock here since we're
3208 * only deciding whether or not to return EIO
3209 * and the memory read should be atomic.
3211 if (iclog->ic_state & XLOG_STATE_IOERROR)
3212 return XFS_ERROR(EIO);
3218 spin_unlock(&log->l_icloglock);
3224 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3225 * about errors or whether the log was flushed or not. This is the normal
3226 * interface to use when trying to unpin items or move the log forward.
3235 trace_xfs_log_force(mp, 0);
3236 error = _xfs_log_force(mp, flags, NULL);
3238 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3242 * Force the in-core log to disk for a specific LSN.
3244 * Find in-core log with lsn.
3245 * If it is in the DIRTY state, just return.
3246 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3247 * state and go to sleep or return.
3248 * If it is in any other state, go to sleep or return.
3250 * Synchronous forces are implemented with a signal variable. All callers
3251 * to force a given lsn to disk will wait on a the sv attached to the
3252 * specific in-core log. When given in-core log finally completes its
3253 * write to disk, that thread will wake up all threads waiting on the
3258 struct xfs_mount *mp,
3263 struct xlog *log = mp->m_log;
3264 struct xlog_in_core *iclog;
3265 int already_slept = 0;
3269 XFS_STATS_INC(xs_log_force);
3271 lsn = xlog_cil_force_lsn(log, lsn);
3272 if (lsn == NULLCOMMITLSN)
3276 spin_lock(&log->l_icloglock);
3277 iclog = log->l_iclog;
3278 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3279 spin_unlock(&log->l_icloglock);
3280 return XFS_ERROR(EIO);
3284 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3285 iclog = iclog->ic_next;
3289 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3290 spin_unlock(&log->l_icloglock);
3294 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3296 * We sleep here if we haven't already slept (e.g.
3297 * this is the first time we've looked at the correct
3298 * iclog buf) and the buffer before us is going to
3299 * be sync'ed. The reason for this is that if we
3300 * are doing sync transactions here, by waiting for
3301 * the previous I/O to complete, we can allow a few
3302 * more transactions into this iclog before we close
3305 * Otherwise, we mark the buffer WANT_SYNC, and bump
3306 * up the refcnt so we can release the log (which
3307 * drops the ref count). The state switch keeps new
3308 * transaction commits from using this buffer. When
3309 * the current commits finish writing into the buffer,
3310 * the refcount will drop to zero and the buffer will
3313 if (!already_slept &&
3314 (iclog->ic_prev->ic_state &
3315 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3316 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3318 XFS_STATS_INC(xs_log_force_sleep);
3320 xlog_wait(&iclog->ic_prev->ic_write_wait,
3327 atomic_inc(&iclog->ic_refcnt);
3328 xlog_state_switch_iclogs(log, iclog, 0);
3329 spin_unlock(&log->l_icloglock);
3330 if (xlog_state_release_iclog(log, iclog))
3331 return XFS_ERROR(EIO);
3334 spin_lock(&log->l_icloglock);
3337 if ((flags & XFS_LOG_SYNC) && /* sleep */
3339 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3341 * Don't wait on completion if we know that we've
3342 * gotten a log write error.
3344 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3345 spin_unlock(&log->l_icloglock);
3346 return XFS_ERROR(EIO);
3348 XFS_STATS_INC(xs_log_force_sleep);
3349 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3351 * No need to grab the log lock here since we're
3352 * only deciding whether or not to return EIO
3353 * and the memory read should be atomic.
3355 if (iclog->ic_state & XLOG_STATE_IOERROR)
3356 return XFS_ERROR(EIO);
3360 } else { /* just return */
3361 spin_unlock(&log->l_icloglock);
3365 } while (iclog != log->l_iclog);
3367 spin_unlock(&log->l_icloglock);
3372 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3373 * about errors or whether the log was flushed or not. This is the normal
3374 * interface to use when trying to unpin items or move the log forward.
3384 trace_xfs_log_force(mp, lsn);
3385 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3387 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3391 * Called when we want to mark the current iclog as being ready to sync to
3395 xlog_state_want_sync(
3397 struct xlog_in_core *iclog)
3399 assert_spin_locked(&log->l_icloglock);
3401 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3402 xlog_state_switch_iclogs(log, iclog, 0);
3404 ASSERT(iclog->ic_state &
3405 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3410 /*****************************************************************************
3414 *****************************************************************************
3418 * Free a used ticket when its refcount falls to zero.
3422 xlog_ticket_t *ticket)
3424 ASSERT(atomic_read(&ticket->t_ref) > 0);
3425 if (atomic_dec_and_test(&ticket->t_ref))
3426 kmem_zone_free(xfs_log_ticket_zone, ticket);
3431 xlog_ticket_t *ticket)
3433 ASSERT(atomic_read(&ticket->t_ref) > 0);
3434 atomic_inc(&ticket->t_ref);
3439 * Figure out the total log space unit (in bytes) that would be
3440 * required for a log ticket.
3443 xfs_log_calc_unit_res(
3444 struct xfs_mount *mp,
3447 struct xlog *log = mp->m_log;
3452 * Permanent reservations have up to 'cnt'-1 active log operations
3453 * in the log. A unit in this case is the amount of space for one
3454 * of these log operations. Normal reservations have a cnt of 1
3455 * and their unit amount is the total amount of space required.
3457 * The following lines of code account for non-transaction data
3458 * which occupy space in the on-disk log.
3460 * Normal form of a transaction is:
3461 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3462 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3464 * We need to account for all the leadup data and trailer data
3465 * around the transaction data.
3466 * And then we need to account for the worst case in terms of using
3468 * The worst case will happen if:
3469 * - the placement of the transaction happens to be such that the
3470 * roundoff is at its maximum
3471 * - the transaction data is synced before the commit record is synced
3472 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3473 * Therefore the commit record is in its own Log Record.
3474 * This can happen as the commit record is called with its
3475 * own region to xlog_write().
3476 * This then means that in the worst case, roundoff can happen for
3477 * the commit-rec as well.
3478 * The commit-rec is smaller than padding in this scenario and so it is
3479 * not added separately.
3482 /* for trans header */
3483 unit_bytes += sizeof(xlog_op_header_t);
3484 unit_bytes += sizeof(xfs_trans_header_t);
3487 unit_bytes += sizeof(xlog_op_header_t);
3490 * for LR headers - the space for data in an iclog is the size minus
3491 * the space used for the headers. If we use the iclog size, then we
3492 * undercalculate the number of headers required.
3494 * Furthermore - the addition of op headers for split-recs might
3495 * increase the space required enough to require more log and op
3496 * headers, so take that into account too.
3498 * IMPORTANT: This reservation makes the assumption that if this
3499 * transaction is the first in an iclog and hence has the LR headers
3500 * accounted to it, then the remaining space in the iclog is
3501 * exclusively for this transaction. i.e. if the transaction is larger
3502 * than the iclog, it will be the only thing in that iclog.
3503 * Fundamentally, this means we must pass the entire log vector to
3504 * xlog_write to guarantee this.
3506 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3507 num_headers = howmany(unit_bytes, iclog_space);
3509 /* for split-recs - ophdrs added when data split over LRs */
3510 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3512 /* add extra header reservations if we overrun */
3513 while (!num_headers ||
3514 howmany(unit_bytes, iclog_space) > num_headers) {
3515 unit_bytes += sizeof(xlog_op_header_t);
3518 unit_bytes += log->l_iclog_hsize * num_headers;
3520 /* for commit-rec LR header - note: padding will subsume the ophdr */
3521 unit_bytes += log->l_iclog_hsize;
3523 /* for roundoff padding for transaction data and one for commit record */
3524 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3525 /* log su roundoff */
3526 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3529 unit_bytes += 2 * BBSIZE;
3536 * Allocate and initialise a new log ticket.
3538 struct xlog_ticket *
3545 xfs_km_flags_t alloc_flags)
3547 struct xlog_ticket *tic;
3550 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3554 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3556 atomic_set(&tic->t_ref, 1);
3557 tic->t_task = current;
3558 INIT_LIST_HEAD(&tic->t_queue);
3559 tic->t_unit_res = unit_res;
3560 tic->t_curr_res = unit_res;
3563 tic->t_tid = prandom_u32();
3564 tic->t_clientid = client;
3565 tic->t_flags = XLOG_TIC_INITED;
3566 tic->t_trans_type = 0;
3568 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3570 xlog_tic_reset_res(tic);
3576 /******************************************************************************
3578 * Log debug routines
3580 ******************************************************************************
3584 * Make sure that the destination ptr is within the valid data region of
3585 * one of the iclogs. This uses backup pointers stored in a different
3586 * part of the log in case we trash the log structure.
3589 xlog_verify_dest_ptr(
3596 for (i = 0; i < log->l_iclog_bufs; i++) {
3597 if (ptr >= log->l_iclog_bak[i] &&
3598 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3603 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3607 * Check to make sure the grant write head didn't just over lap the tail. If
3608 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3609 * the cycles differ by exactly one and check the byte count.
3611 * This check is run unlocked, so can give false positives. Rather than assert
3612 * on failures, use a warn-once flag and a panic tag to allow the admin to
3613 * determine if they want to panic the machine when such an error occurs. For
3614 * debug kernels this will have the same effect as using an assert but, unlinke
3615 * an assert, it can be turned off at runtime.
3618 xlog_verify_grant_tail(
3621 int tail_cycle, tail_blocks;
3624 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3625 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3626 if (tail_cycle != cycle) {
3627 if (cycle - 1 != tail_cycle &&
3628 !(log->l_flags & XLOG_TAIL_WARN)) {
3629 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3630 "%s: cycle - 1 != tail_cycle", __func__);
3631 log->l_flags |= XLOG_TAIL_WARN;
3634 if (space > BBTOB(tail_blocks) &&
3635 !(log->l_flags & XLOG_TAIL_WARN)) {
3636 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3637 "%s: space > BBTOB(tail_blocks)", __func__);
3638 log->l_flags |= XLOG_TAIL_WARN;
3643 /* check if it will fit */
3645 xlog_verify_tail_lsn(
3647 struct xlog_in_core *iclog,
3652 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3654 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3655 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3656 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3658 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3660 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3661 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3663 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3664 if (blocks < BTOBB(iclog->ic_offset) + 1)
3665 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3667 } /* xlog_verify_tail_lsn */
3670 * Perform a number of checks on the iclog before writing to disk.
3672 * 1. Make sure the iclogs are still circular
3673 * 2. Make sure we have a good magic number
3674 * 3. Make sure we don't have magic numbers in the data
3675 * 4. Check fields of each log operation header for:
3676 * A. Valid client identifier
3677 * B. tid ptr value falls in valid ptr space (user space code)
3678 * C. Length in log record header is correct according to the
3679 * individual operation headers within record.
3680 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3681 * log, check the preceding blocks of the physical log to make sure all
3682 * the cycle numbers agree with the current cycle number.
3687 struct xlog_in_core *iclog,
3691 xlog_op_header_t *ophead;
3692 xlog_in_core_t *icptr;
3693 xlog_in_core_2_t *xhdr;
3695 xfs_caddr_t base_ptr;
3696 __psint_t field_offset;
3698 int len, i, j, k, op_len;
3701 /* check validity of iclog pointers */
3702 spin_lock(&log->l_icloglock);
3703 icptr = log->l_iclog;
3704 for (i=0; i < log->l_iclog_bufs; i++) {
3706 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3707 icptr = icptr->ic_next;
3709 if (icptr != log->l_iclog)
3710 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3711 spin_unlock(&log->l_icloglock);
3713 /* check log magic numbers */
3714 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3715 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3717 ptr = (xfs_caddr_t) &iclog->ic_header;
3718 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3720 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3721 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3726 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3727 ptr = iclog->ic_datap;
3729 ophead = (xlog_op_header_t *)ptr;
3730 xhdr = iclog->ic_data;
3731 for (i = 0; i < len; i++) {
3732 ophead = (xlog_op_header_t *)ptr;
3734 /* clientid is only 1 byte */
3735 field_offset = (__psint_t)
3736 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3737 if (!syncing || (field_offset & 0x1ff)) {
3738 clientid = ophead->oh_clientid;
3740 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3741 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3742 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3743 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3744 clientid = xlog_get_client_id(
3745 xhdr[j].hic_xheader.xh_cycle_data[k]);
3747 clientid = xlog_get_client_id(
3748 iclog->ic_header.h_cycle_data[idx]);
3751 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3753 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3754 __func__, clientid, ophead,
3755 (unsigned long)field_offset);
3758 field_offset = (__psint_t)
3759 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3760 if (!syncing || (field_offset & 0x1ff)) {
3761 op_len = be32_to_cpu(ophead->oh_len);
3763 idx = BTOBBT((__psint_t)&ophead->oh_len -
3764 (__psint_t)iclog->ic_datap);
3765 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3766 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3767 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3768 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3770 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3773 ptr += sizeof(xlog_op_header_t) + op_len;
3775 } /* xlog_verify_iclog */
3779 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3785 xlog_in_core_t *iclog, *ic;
3787 iclog = log->l_iclog;
3788 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3790 * Mark all the incore logs IOERROR.
3791 * From now on, no log flushes will result.
3795 ic->ic_state = XLOG_STATE_IOERROR;
3797 } while (ic != iclog);
3801 * Return non-zero, if state transition has already happened.
3807 * This is called from xfs_force_shutdown, when we're forcibly
3808 * shutting down the filesystem, typically because of an IO error.
3809 * Our main objectives here are to make sure that:
3810 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3811 * parties to find out, 'atomically'.
3812 * b. those who're sleeping on log reservations, pinned objects and
3813 * other resources get woken up, and be told the bad news.
3814 * c. nothing new gets queued up after (a) and (b) are done.
3815 * d. if !logerror, flush the iclogs to disk, then seal them off
3818 * Note: for delayed logging the !logerror case needs to flush the regions
3819 * held in memory out to the iclogs before flushing them to disk. This needs
3820 * to be done before the log is marked as shutdown, otherwise the flush to the
3824 xfs_log_force_umount(
3825 struct xfs_mount *mp,
3834 * If this happens during log recovery, don't worry about
3835 * locking; the log isn't open for business yet.
3838 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3839 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3841 XFS_BUF_DONE(mp->m_sb_bp);
3846 * Somebody could've already done the hard work for us.
3847 * No need to get locks for this.
3849 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3850 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3856 * Flush the in memory commit item list before marking the log as
3857 * being shut down. We need to do it in this order to ensure all the
3858 * completed transactions are flushed to disk with the xfs_log_force()
3862 xlog_cil_force(log);
3865 * mark the filesystem and the as in a shutdown state and wake
3866 * everybody up to tell them the bad news.
3868 spin_lock(&log->l_icloglock);
3869 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3871 XFS_BUF_DONE(mp->m_sb_bp);
3874 * This flag is sort of redundant because of the mount flag, but
3875 * it's good to maintain the separation between the log and the rest
3878 log->l_flags |= XLOG_IO_ERROR;
3881 * If we hit a log error, we want to mark all the iclogs IOERROR
3882 * while we're still holding the loglock.
3885 retval = xlog_state_ioerror(log);
3886 spin_unlock(&log->l_icloglock);
3889 * We don't want anybody waiting for log reservations after this. That
3890 * means we have to wake up everybody queued up on reserveq as well as
3891 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3892 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3893 * action is protected by the grant locks.
3895 xlog_grant_head_wake_all(&log->l_reserve_head);
3896 xlog_grant_head_wake_all(&log->l_write_head);
3898 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3901 * Force the incore logs to disk before shutting the
3902 * log down completely.
3904 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3906 spin_lock(&log->l_icloglock);
3907 retval = xlog_state_ioerror(log);
3908 spin_unlock(&log->l_icloglock);
3911 * Wake up everybody waiting on xfs_log_force.
3912 * Callback all log item committed functions as if the
3913 * log writes were completed.
3915 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3917 #ifdef XFSERRORDEBUG
3919 xlog_in_core_t *iclog;
3921 spin_lock(&log->l_icloglock);
3922 iclog = log->l_iclog;
3924 ASSERT(iclog->ic_callback == 0);
3925 iclog = iclog->ic_next;
3926 } while (iclog != log->l_iclog);
3927 spin_unlock(&log->l_icloglock);
3930 /* return non-zero if log IOERROR transition had already happened */
3938 xlog_in_core_t *iclog;
3940 iclog = log->l_iclog;
3942 /* endianness does not matter here, zero is zero in
3945 if (iclog->ic_header.h_num_logops)
3947 iclog = iclog->ic_next;
3948 } while (iclog != log->l_iclog);