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"
39 kmem_zone_t
*xfs_log_ticket_zone
;
41 /* Local miscellaneous function prototypes */
45 struct xlog_ticket
*ticket
,
46 struct xlog_in_core
**iclog
,
47 xfs_lsn_t
*commitlsnp
);
52 struct xfs_buftarg
*log_target
,
53 xfs_daddr_t blk_offset
,
62 struct xlog_in_core
*iclog
);
67 /* local state machine functions */
68 STATIC
void xlog_state_done_syncing(xlog_in_core_t
*iclog
, int);
70 xlog_state_do_callback(
73 struct xlog_in_core
*iclog
);
75 xlog_state_get_iclog_space(
78 struct xlog_in_core
**iclog
,
79 struct xlog_ticket
*ticket
,
83 xlog_state_release_iclog(
85 struct xlog_in_core
*iclog
);
87 xlog_state_switch_iclogs(
89 struct xlog_in_core
*iclog
,
94 struct xlog_in_core
*iclog
);
101 xlog_regrant_reserve_log_space(
103 struct xlog_ticket
*ticket
);
105 xlog_ungrant_log_space(
107 struct xlog_ticket
*ticket
);
111 xlog_verify_dest_ptr(
115 xlog_verify_grant_tail(
120 struct xlog_in_core
*iclog
,
124 xlog_verify_tail_lsn(
126 struct xlog_in_core
*iclog
,
129 #define xlog_verify_dest_ptr(a,b)
130 #define xlog_verify_grant_tail(a)
131 #define xlog_verify_iclog(a,b,c,d)
132 #define xlog_verify_tail_lsn(a,b,c)
140 xlog_grant_sub_space(
145 int64_t head_val
= atomic64_read(head
);
151 xlog_crack_grant_head_val(head_val
, &cycle
, &space
);
155 space
+= log
->l_logsize
;
160 new = xlog_assign_grant_head_val(cycle
, space
);
161 head_val
= atomic64_cmpxchg(head
, old
, new);
162 } while (head_val
!= old
);
166 xlog_grant_add_space(
171 int64_t head_val
= atomic64_read(head
);
178 xlog_crack_grant_head_val(head_val
, &cycle
, &space
);
180 tmp
= log
->l_logsize
- space
;
189 new = xlog_assign_grant_head_val(cycle
, space
);
190 head_val
= atomic64_cmpxchg(head
, old
, new);
191 } while (head_val
!= old
);
195 xlog_grant_head_init(
196 struct xlog_grant_head
*head
)
198 xlog_assign_grant_head(&head
->grant
, 1, 0);
199 INIT_LIST_HEAD(&head
->waiters
);
200 spin_lock_init(&head
->lock
);
204 xlog_grant_head_wake_all(
205 struct xlog_grant_head
*head
)
207 struct xlog_ticket
*tic
;
209 spin_lock(&head
->lock
);
210 list_for_each_entry(tic
, &head
->waiters
, t_queue
)
211 wake_up_process(tic
->t_task
);
212 spin_unlock(&head
->lock
);
216 xlog_ticket_reservation(
218 struct xlog_grant_head
*head
,
219 struct xlog_ticket
*tic
)
221 if (head
== &log
->l_write_head
) {
222 ASSERT(tic
->t_flags
& XLOG_TIC_PERM_RESERV
);
223 return tic
->t_unit_res
;
225 if (tic
->t_flags
& XLOG_TIC_PERM_RESERV
)
226 return tic
->t_unit_res
* tic
->t_cnt
;
228 return tic
->t_unit_res
;
233 xlog_grant_head_wake(
235 struct xlog_grant_head
*head
,
238 struct xlog_ticket
*tic
;
241 list_for_each_entry(tic
, &head
->waiters
, t_queue
) {
242 need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
243 if (*free_bytes
< need_bytes
)
246 *free_bytes
-= need_bytes
;
247 trace_xfs_log_grant_wake_up(log
, tic
);
248 wake_up_process(tic
->t_task
);
255 xlog_grant_head_wait(
257 struct xlog_grant_head
*head
,
258 struct xlog_ticket
*tic
,
261 list_add_tail(&tic
->t_queue
, &head
->waiters
);
264 if (XLOG_FORCED_SHUTDOWN(log
))
266 xlog_grant_push_ail(log
, need_bytes
);
268 __set_current_state(TASK_UNINTERRUPTIBLE
);
269 spin_unlock(&head
->lock
);
271 XFS_STATS_INC(xs_sleep_logspace
);
273 trace_xfs_log_grant_sleep(log
, tic
);
275 trace_xfs_log_grant_wake(log
, tic
);
277 spin_lock(&head
->lock
);
278 if (XLOG_FORCED_SHUTDOWN(log
))
280 } while (xlog_space_left(log
, &head
->grant
) < need_bytes
);
282 list_del_init(&tic
->t_queue
);
285 list_del_init(&tic
->t_queue
);
286 return XFS_ERROR(EIO
);
290 * Atomically get the log space required for a log ticket.
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
307 xlog_grant_head_check(
309 struct xlog_grant_head
*head
,
310 struct xlog_ticket
*tic
,
316 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
324 *need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
325 free_bytes
= xlog_space_left(log
, &head
->grant
);
326 if (!list_empty_careful(&head
->waiters
)) {
327 spin_lock(&head
->lock
);
328 if (!xlog_grant_head_wake(log
, head
, &free_bytes
) ||
329 free_bytes
< *need_bytes
) {
330 error
= xlog_grant_head_wait(log
, head
, tic
,
333 spin_unlock(&head
->lock
);
334 } else if (free_bytes
< *need_bytes
) {
335 spin_lock(&head
->lock
);
336 error
= xlog_grant_head_wait(log
, head
, tic
, *need_bytes
);
337 spin_unlock(&head
->lock
);
344 xlog_tic_reset_res(xlog_ticket_t
*tic
)
347 tic
->t_res_arr_sum
= 0;
348 tic
->t_res_num_ophdrs
= 0;
352 xlog_tic_add_region(xlog_ticket_t
*tic
, uint len
, uint type
)
354 if (tic
->t_res_num
== XLOG_TIC_LEN_MAX
) {
355 /* add to overflow and start again */
356 tic
->t_res_o_flow
+= tic
->t_res_arr_sum
;
358 tic
->t_res_arr_sum
= 0;
361 tic
->t_res_arr
[tic
->t_res_num
].r_len
= len
;
362 tic
->t_res_arr
[tic
->t_res_num
].r_type
= type
;
363 tic
->t_res_arr_sum
+= len
;
368 * Replenish the byte reservation required by moving the grant write head.
372 struct xfs_mount
*mp
,
373 struct xlog_ticket
*tic
)
375 struct xlog
*log
= mp
->m_log
;
379 if (XLOG_FORCED_SHUTDOWN(log
))
380 return XFS_ERROR(EIO
);
382 XFS_STATS_INC(xs_try_logspace
);
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
392 xlog_grant_push_ail(log
, tic
->t_unit_res
);
394 tic
->t_curr_res
= tic
->t_unit_res
;
395 xlog_tic_reset_res(tic
);
400 trace_xfs_log_regrant(log
, tic
);
402 error
= xlog_grant_head_check(log
, &log
->l_write_head
, tic
,
407 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
408 trace_xfs_log_regrant_exit(log
, tic
);
409 xlog_verify_grant_tail(log
);
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
419 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
424 * Reserve log space and return a ticket corresponding the reservation.
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
433 struct xfs_mount
*mp
,
436 struct xlog_ticket
**ticp
,
441 struct xlog
*log
= mp
->m_log
;
442 struct xlog_ticket
*tic
;
446 ASSERT(client
== XFS_TRANSACTION
|| client
== XFS_LOG
);
448 if (XLOG_FORCED_SHUTDOWN(log
))
449 return XFS_ERROR(EIO
);
451 XFS_STATS_INC(xs_try_logspace
);
453 ASSERT(*ticp
== NULL
);
454 tic
= xlog_ticket_alloc(log
, unit_bytes
, cnt
, client
, permanent
,
455 KM_SLEEP
| KM_MAYFAIL
);
457 return XFS_ERROR(ENOMEM
);
459 tic
->t_trans_type
= t_type
;
462 xlog_grant_push_ail(log
, tic
->t_unit_res
* tic
->t_cnt
);
464 trace_xfs_log_reserve(log
, tic
);
466 error
= xlog_grant_head_check(log
, &log
->l_reserve_head
, tic
,
471 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, need_bytes
);
472 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
473 trace_xfs_log_reserve_exit(log
, tic
);
474 xlog_verify_grant_tail(log
);
479 * If we are failing, make sure the ticket doesn't have any current
480 * reservations. We don't want to add this back when the ticket/
481 * transaction gets cancelled.
484 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
492 * 1. currblock field gets updated at startup and after in-core logs
493 * marked as with WANT_SYNC.
497 * This routine is called when a user of a log manager ticket is done with
498 * the reservation. If the ticket was ever used, then a commit record for
499 * the associated transaction is written out as a log operation header with
500 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
501 * a given ticket. If the ticket was one with a permanent reservation, then
502 * a few operations are done differently. Permanent reservation tickets by
503 * default don't release the reservation. They just commit the current
504 * transaction with the belief that the reservation is still needed. A flag
505 * must be passed in before permanent reservations are actually released.
506 * When these type of tickets are not released, they need to be set into
507 * the inited state again. By doing this, a start record will be written
508 * out when the next write occurs.
512 struct xfs_mount
*mp
,
513 struct xlog_ticket
*ticket
,
514 struct xlog_in_core
**iclog
,
517 struct xlog
*log
= mp
->m_log
;
520 if (XLOG_FORCED_SHUTDOWN(log
) ||
522 * If nothing was ever written, don't write out commit record.
523 * If we get an error, just continue and give back the log ticket.
525 (((ticket
->t_flags
& XLOG_TIC_INITED
) == 0) &&
526 (xlog_commit_record(log
, ticket
, iclog
, &lsn
)))) {
527 lsn
= (xfs_lsn_t
) -1;
528 if (ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) {
529 flags
|= XFS_LOG_REL_PERM_RESERV
;
534 if ((ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) == 0 ||
535 (flags
& XFS_LOG_REL_PERM_RESERV
)) {
536 trace_xfs_log_done_nonperm(log
, ticket
);
539 * Release ticket if not permanent reservation or a specific
540 * request has been made to release a permanent reservation.
542 xlog_ungrant_log_space(log
, ticket
);
543 xfs_log_ticket_put(ticket
);
545 trace_xfs_log_done_perm(log
, ticket
);
547 xlog_regrant_reserve_log_space(log
, ticket
);
548 /* If this ticket was a permanent reservation and we aren't
549 * trying to release it, reset the inited flags; so next time
550 * we write, a start record will be written out.
552 ticket
->t_flags
|= XLOG_TIC_INITED
;
559 * Attaches a new iclog I/O completion callback routine during
560 * transaction commit. If the log is in error state, a non-zero
561 * return code is handed back and the caller is responsible for
562 * executing the callback at an appropriate time.
566 struct xfs_mount
*mp
,
567 struct xlog_in_core
*iclog
,
568 xfs_log_callback_t
*cb
)
572 spin_lock(&iclog
->ic_callback_lock
);
573 abortflg
= (iclog
->ic_state
& XLOG_STATE_IOERROR
);
575 ASSERT_ALWAYS((iclog
->ic_state
== XLOG_STATE_ACTIVE
) ||
576 (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
));
578 *(iclog
->ic_callback_tail
) = cb
;
579 iclog
->ic_callback_tail
= &(cb
->cb_next
);
581 spin_unlock(&iclog
->ic_callback_lock
);
586 xfs_log_release_iclog(
587 struct xfs_mount
*mp
,
588 struct xlog_in_core
*iclog
)
590 if (xlog_state_release_iclog(mp
->m_log
, iclog
)) {
591 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
599 * Mount a log filesystem
601 * mp - ubiquitous xfs mount point structure
602 * log_target - buftarg of on-disk log device
603 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
604 * num_bblocks - Number of BBSIZE blocks in on-disk log
606 * Return error or zero.
611 xfs_buftarg_t
*log_target
,
612 xfs_daddr_t blk_offset
,
617 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
))
618 xfs_notice(mp
, "Mounting Filesystem");
621 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
622 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
625 mp
->m_log
= xlog_alloc_log(mp
, log_target
, blk_offset
, num_bblks
);
626 if (IS_ERR(mp
->m_log
)) {
627 error
= -PTR_ERR(mp
->m_log
);
632 * Initialize the AIL now we have a log.
634 error
= xfs_trans_ail_init(mp
);
636 xfs_warn(mp
, "AIL initialisation failed: error %d", error
);
639 mp
->m_log
->l_ailp
= mp
->m_ail
;
642 * skip log recovery on a norecovery mount. pretend it all
645 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
)) {
646 int readonly
= (mp
->m_flags
& XFS_MOUNT_RDONLY
);
649 mp
->m_flags
&= ~XFS_MOUNT_RDONLY
;
651 error
= xlog_recover(mp
->m_log
);
654 mp
->m_flags
|= XFS_MOUNT_RDONLY
;
656 xfs_warn(mp
, "log mount/recovery failed: error %d",
658 goto out_destroy_ail
;
662 /* Normal transactions can now occur */
663 mp
->m_log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
666 * Now the log has been fully initialised and we know were our
667 * space grant counters are, we can initialise the permanent ticket
668 * needed for delayed logging to work.
670 xlog_cil_init_post_recovery(mp
->m_log
);
675 xfs_trans_ail_destroy(mp
);
677 xlog_dealloc_log(mp
->m_log
);
683 * Finish the recovery of the file system. This is separate from the
684 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
685 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
688 * If we finish recovery successfully, start the background log work. If we are
689 * not doing recovery, then we have a RO filesystem and we don't need to start
693 xfs_log_mount_finish(xfs_mount_t
*mp
)
697 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
)) {
698 error
= xlog_recover_finish(mp
->m_log
);
700 xfs_log_work_queue(mp
);
702 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
710 * Final log writes as part of unmount.
712 * Mark the filesystem clean as unmount happens. Note that during relocation
713 * this routine needs to be executed as part of source-bag while the
714 * deallocation must not be done until source-end.
718 * Unmount record used to have a string "Unmount filesystem--" in the
719 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
720 * We just write the magic number now since that particular field isn't
721 * currently architecture converted and "nUmount" is a bit foo.
722 * As far as I know, there weren't any dependencies on the old behaviour.
726 xfs_log_unmount_write(xfs_mount_t
*mp
)
728 struct xlog
*log
= mp
->m_log
;
729 xlog_in_core_t
*iclog
;
731 xlog_in_core_t
*first_iclog
;
733 xlog_ticket_t
*tic
= NULL
;
738 * Don't write out unmount record on read-only mounts.
739 * Or, if we are doing a forced umount (typically because of IO errors).
741 if (mp
->m_flags
& XFS_MOUNT_RDONLY
)
744 error
= _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
745 ASSERT(error
|| !(XLOG_FORCED_SHUTDOWN(log
)));
748 first_iclog
= iclog
= log
->l_iclog
;
750 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
751 ASSERT(iclog
->ic_state
& XLOG_STATE_ACTIVE
);
752 ASSERT(iclog
->ic_offset
== 0);
754 iclog
= iclog
->ic_next
;
755 } while (iclog
!= first_iclog
);
757 if (! (XLOG_FORCED_SHUTDOWN(log
))) {
758 error
= xfs_log_reserve(mp
, 600, 1, &tic
,
759 XFS_LOG
, 0, XLOG_UNMOUNT_REC_TYPE
);
761 /* the data section must be 32 bit size aligned */
765 __uint32_t pad2
; /* may as well make it 64 bits */
767 .magic
= XLOG_UNMOUNT_TYPE
,
769 struct xfs_log_iovec reg
= {
771 .i_len
= sizeof(magic
),
772 .i_type
= XLOG_REG_TYPE_UNMOUNT
,
774 struct xfs_log_vec vec
= {
779 /* remove inited flag, and account for space used */
781 tic
->t_curr_res
-= sizeof(magic
);
782 error
= xlog_write(log
, &vec
, tic
, &lsn
,
783 NULL
, XLOG_UNMOUNT_TRANS
);
785 * At this point, we're umounting anyway,
786 * so there's no point in transitioning log state
787 * to IOERROR. Just continue...
792 xfs_alert(mp
, "%s: unmount record failed", __func__
);
795 spin_lock(&log
->l_icloglock
);
796 iclog
= log
->l_iclog
;
797 atomic_inc(&iclog
->ic_refcnt
);
798 xlog_state_want_sync(log
, iclog
);
799 spin_unlock(&log
->l_icloglock
);
800 error
= xlog_state_release_iclog(log
, iclog
);
802 spin_lock(&log
->l_icloglock
);
803 if (!(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
804 iclog
->ic_state
== XLOG_STATE_DIRTY
)) {
805 if (!XLOG_FORCED_SHUTDOWN(log
)) {
806 xlog_wait(&iclog
->ic_force_wait
,
809 spin_unlock(&log
->l_icloglock
);
812 spin_unlock(&log
->l_icloglock
);
815 trace_xfs_log_umount_write(log
, tic
);
816 xlog_ungrant_log_space(log
, tic
);
817 xfs_log_ticket_put(tic
);
821 * We're already in forced_shutdown mode, couldn't
822 * even attempt to write out the unmount transaction.
824 * Go through the motions of sync'ing and releasing
825 * the iclog, even though no I/O will actually happen,
826 * we need to wait for other log I/Os that may already
827 * be in progress. Do this as a separate section of
828 * code so we'll know if we ever get stuck here that
829 * we're in this odd situation of trying to unmount
830 * a file system that went into forced_shutdown as
831 * the result of an unmount..
833 spin_lock(&log
->l_icloglock
);
834 iclog
= log
->l_iclog
;
835 atomic_inc(&iclog
->ic_refcnt
);
837 xlog_state_want_sync(log
, iclog
);
838 spin_unlock(&log
->l_icloglock
);
839 error
= xlog_state_release_iclog(log
, iclog
);
841 spin_lock(&log
->l_icloglock
);
843 if ( ! ( iclog
->ic_state
== XLOG_STATE_ACTIVE
844 || iclog
->ic_state
== XLOG_STATE_DIRTY
845 || iclog
->ic_state
== XLOG_STATE_IOERROR
) ) {
847 xlog_wait(&iclog
->ic_force_wait
,
850 spin_unlock(&log
->l_icloglock
);
855 } /* xfs_log_unmount_write */
858 * Shut down and release the AIL and Log.
860 * During unmount, we need to ensure we flush all the dirty metadata objects
861 * from the AIL so that the log is empty before we write the unmount record to
864 * To do this, we first need to shut down the background log work so it is not
865 * trying to cover the log as we clean up. We then need to unpin all objects in
866 * the log so we can then flush them out. Once they have completed their IO and
867 * run the callbacks removing themselves from the AIL, we can write the unmount
868 * record, tear down the AIL and finally free the log.
871 xfs_log_unmount(xfs_mount_t
*mp
)
873 cancel_delayed_work_sync(&mp
->m_log
->l_work
);
874 xfs_log_force(mp
, XFS_LOG_SYNC
);
877 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
878 * will push it, xfs_wait_buftarg() will not wait for it. Further,
879 * xfs_buf_iowait() cannot be used because it was pushed with the
880 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
881 * the IO to complete.
883 xfs_ail_push_all_sync(mp
->m_ail
);
884 xfs_wait_buftarg(mp
->m_ddev_targp
);
885 xfs_buf_lock(mp
->m_sb_bp
);
886 xfs_buf_unlock(mp
->m_sb_bp
);
888 xfs_log_unmount_write(mp
);
890 xfs_trans_ail_destroy(mp
);
891 xlog_dealloc_log(mp
->m_log
);
896 struct xfs_mount
*mp
,
897 struct xfs_log_item
*item
,
899 const struct xfs_item_ops
*ops
)
901 item
->li_mountp
= mp
;
902 item
->li_ailp
= mp
->m_ail
;
903 item
->li_type
= type
;
907 INIT_LIST_HEAD(&item
->li_ail
);
908 INIT_LIST_HEAD(&item
->li_cil
);
912 * Wake up processes waiting for log space after we have moved the log tail.
916 struct xfs_mount
*mp
)
918 struct xlog
*log
= mp
->m_log
;
921 if (XLOG_FORCED_SHUTDOWN(log
))
924 if (!list_empty_careful(&log
->l_write_head
.waiters
)) {
925 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
927 spin_lock(&log
->l_write_head
.lock
);
928 free_bytes
= xlog_space_left(log
, &log
->l_write_head
.grant
);
929 xlog_grant_head_wake(log
, &log
->l_write_head
, &free_bytes
);
930 spin_unlock(&log
->l_write_head
.lock
);
933 if (!list_empty_careful(&log
->l_reserve_head
.waiters
)) {
934 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
936 spin_lock(&log
->l_reserve_head
.lock
);
937 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
938 xlog_grant_head_wake(log
, &log
->l_reserve_head
, &free_bytes
);
939 spin_unlock(&log
->l_reserve_head
.lock
);
944 * Determine if we have a transaction that has gone to disk
945 * that needs to be covered. To begin the transition to the idle state
946 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
947 * If we are then in a state where covering is needed, the caller is informed
948 * that dummy transactions are required to move the log into the idle state.
950 * Because this is called as part of the sync process, we should also indicate
951 * that dummy transactions should be issued in anything but the covered or
952 * idle states. This ensures that the log tail is accurately reflected in
953 * the log at the end of the sync, hence if a crash occurrs avoids replay
954 * of transactions where the metadata is already on disk.
957 xfs_log_need_covered(xfs_mount_t
*mp
)
960 struct xlog
*log
= mp
->m_log
;
962 if (!xfs_fs_writable(mp
))
965 spin_lock(&log
->l_icloglock
);
966 switch (log
->l_covered_state
) {
967 case XLOG_STATE_COVER_DONE
:
968 case XLOG_STATE_COVER_DONE2
:
969 case XLOG_STATE_COVER_IDLE
:
971 case XLOG_STATE_COVER_NEED
:
972 case XLOG_STATE_COVER_NEED2
:
973 if (!xfs_ail_min_lsn(log
->l_ailp
) &&
974 xlog_iclogs_empty(log
)) {
975 if (log
->l_covered_state
== XLOG_STATE_COVER_NEED
)
976 log
->l_covered_state
= XLOG_STATE_COVER_DONE
;
978 log
->l_covered_state
= XLOG_STATE_COVER_DONE2
;
985 spin_unlock(&log
->l_icloglock
);
990 * We may be holding the log iclog lock upon entering this routine.
993 xlog_assign_tail_lsn_locked(
994 struct xfs_mount
*mp
)
996 struct xlog
*log
= mp
->m_log
;
997 struct xfs_log_item
*lip
;
1000 assert_spin_locked(&mp
->m_ail
->xa_lock
);
1003 * To make sure we always have a valid LSN for the log tail we keep
1004 * track of the last LSN which was committed in log->l_last_sync_lsn,
1005 * and use that when the AIL was empty.
1007 lip
= xfs_ail_min(mp
->m_ail
);
1009 tail_lsn
= lip
->li_lsn
;
1011 tail_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1012 atomic64_set(&log
->l_tail_lsn
, tail_lsn
);
1017 xlog_assign_tail_lsn(
1018 struct xfs_mount
*mp
)
1022 spin_lock(&mp
->m_ail
->xa_lock
);
1023 tail_lsn
= xlog_assign_tail_lsn_locked(mp
);
1024 spin_unlock(&mp
->m_ail
->xa_lock
);
1030 * Return the space in the log between the tail and the head. The head
1031 * is passed in the cycle/bytes formal parms. In the special case where
1032 * the reserve head has wrapped passed the tail, this calculation is no
1033 * longer valid. In this case, just return 0 which means there is no space
1034 * in the log. This works for all places where this function is called
1035 * with the reserve head. Of course, if the write head were to ever
1036 * wrap the tail, we should blow up. Rather than catch this case here,
1037 * we depend on other ASSERTions in other parts of the code. XXXmiken
1039 * This code also handles the case where the reservation head is behind
1040 * the tail. The details of this case are described below, but the end
1041 * result is that we return the size of the log as the amount of space left.
1054 xlog_crack_grant_head(head
, &head_cycle
, &head_bytes
);
1055 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_bytes
);
1056 tail_bytes
= BBTOB(tail_bytes
);
1057 if (tail_cycle
== head_cycle
&& head_bytes
>= tail_bytes
)
1058 free_bytes
= log
->l_logsize
- (head_bytes
- tail_bytes
);
1059 else if (tail_cycle
+ 1 < head_cycle
)
1061 else if (tail_cycle
< head_cycle
) {
1062 ASSERT(tail_cycle
== (head_cycle
- 1));
1063 free_bytes
= tail_bytes
- head_bytes
;
1066 * The reservation head is behind the tail.
1067 * In this case we just want to return the size of the
1068 * log as the amount of space left.
1070 xfs_alert(log
->l_mp
,
1071 "xlog_space_left: head behind tail\n"
1072 " tail_cycle = %d, tail_bytes = %d\n"
1073 " GH cycle = %d, GH bytes = %d",
1074 tail_cycle
, tail_bytes
, head_cycle
, head_bytes
);
1076 free_bytes
= log
->l_logsize
;
1083 * Log function which is called when an io completes.
1085 * The log manager needs its own routine, in order to control what
1086 * happens with the buffer after the write completes.
1089 xlog_iodone(xfs_buf_t
*bp
)
1091 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1092 struct xlog
*l
= iclog
->ic_log
;
1096 * Race to shutdown the filesystem if we see an error.
1098 if (XFS_TEST_ERROR((xfs_buf_geterror(bp
)), l
->l_mp
,
1099 XFS_ERRTAG_IODONE_IOERR
, XFS_RANDOM_IODONE_IOERR
)) {
1100 xfs_buf_ioerror_alert(bp
, __func__
);
1102 xfs_force_shutdown(l
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1104 * This flag will be propagated to the trans-committed
1105 * callback routines to let them know that the log-commit
1108 aborted
= XFS_LI_ABORTED
;
1109 } else if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1110 aborted
= XFS_LI_ABORTED
;
1113 /* log I/O is always issued ASYNC */
1114 ASSERT(XFS_BUF_ISASYNC(bp
));
1115 xlog_state_done_syncing(iclog
, aborted
);
1117 * do not reference the buffer (bp) here as we could race
1118 * with it being freed after writing the unmount record to the
1125 * Return size of each in-core log record buffer.
1127 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1129 * If the filesystem blocksize is too large, we may need to choose a
1130 * larger size since the directory code currently logs entire blocks.
1134 xlog_get_iclog_buffer_size(
1135 struct xfs_mount
*mp
,
1141 if (mp
->m_logbufs
<= 0)
1142 log
->l_iclog_bufs
= XLOG_MAX_ICLOGS
;
1144 log
->l_iclog_bufs
= mp
->m_logbufs
;
1147 * Buffer size passed in from mount system call.
1149 if (mp
->m_logbsize
> 0) {
1150 size
= log
->l_iclog_size
= mp
->m_logbsize
;
1151 log
->l_iclog_size_log
= 0;
1153 log
->l_iclog_size_log
++;
1157 if (xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1158 /* # headers = size / 32k
1159 * one header holds cycles from 32k of data
1162 xhdrs
= mp
->m_logbsize
/ XLOG_HEADER_CYCLE_SIZE
;
1163 if (mp
->m_logbsize
% XLOG_HEADER_CYCLE_SIZE
)
1165 log
->l_iclog_hsize
= xhdrs
<< BBSHIFT
;
1166 log
->l_iclog_heads
= xhdrs
;
1168 ASSERT(mp
->m_logbsize
<= XLOG_BIG_RECORD_BSIZE
);
1169 log
->l_iclog_hsize
= BBSIZE
;
1170 log
->l_iclog_heads
= 1;
1175 /* All machines use 32kB buffers by default. */
1176 log
->l_iclog_size
= XLOG_BIG_RECORD_BSIZE
;
1177 log
->l_iclog_size_log
= XLOG_BIG_RECORD_BSHIFT
;
1179 /* the default log size is 16k or 32k which is one header sector */
1180 log
->l_iclog_hsize
= BBSIZE
;
1181 log
->l_iclog_heads
= 1;
1184 /* are we being asked to make the sizes selected above visible? */
1185 if (mp
->m_logbufs
== 0)
1186 mp
->m_logbufs
= log
->l_iclog_bufs
;
1187 if (mp
->m_logbsize
== 0)
1188 mp
->m_logbsize
= log
->l_iclog_size
;
1189 } /* xlog_get_iclog_buffer_size */
1194 struct xfs_mount
*mp
)
1196 queue_delayed_work(mp
->m_log_workqueue
, &mp
->m_log
->l_work
,
1197 msecs_to_jiffies(xfs_syncd_centisecs
* 10));
1201 * Every sync period we need to unpin all items in the AIL and push them to
1202 * disk. If there is nothing dirty, then we might need to cover the log to
1203 * indicate that the filesystem is idle.
1207 struct work_struct
*work
)
1209 struct xlog
*log
= container_of(to_delayed_work(work
),
1210 struct xlog
, l_work
);
1211 struct xfs_mount
*mp
= log
->l_mp
;
1213 /* dgc: errors ignored - not fatal and nowhere to report them */
1214 if (xfs_log_need_covered(mp
))
1215 xfs_fs_log_dummy(mp
);
1217 xfs_log_force(mp
, 0);
1219 /* start pushing all the metadata that is currently dirty */
1220 xfs_ail_push_all(mp
->m_ail
);
1222 /* queue us up again */
1223 xfs_log_work_queue(mp
);
1227 * This routine initializes some of the log structure for a given mount point.
1228 * Its primary purpose is to fill in enough, so recovery can occur. However,
1229 * some other stuff may be filled in too.
1231 STATIC
struct xlog
*
1233 struct xfs_mount
*mp
,
1234 struct xfs_buftarg
*log_target
,
1235 xfs_daddr_t blk_offset
,
1239 xlog_rec_header_t
*head
;
1240 xlog_in_core_t
**iclogp
;
1241 xlog_in_core_t
*iclog
, *prev_iclog
=NULL
;
1247 log
= kmem_zalloc(sizeof(struct xlog
), KM_MAYFAIL
);
1249 xfs_warn(mp
, "Log allocation failed: No memory!");
1254 log
->l_targ
= log_target
;
1255 log
->l_logsize
= BBTOB(num_bblks
);
1256 log
->l_logBBstart
= blk_offset
;
1257 log
->l_logBBsize
= num_bblks
;
1258 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
1259 log
->l_flags
|= XLOG_ACTIVE_RECOVERY
;
1260 INIT_DELAYED_WORK(&log
->l_work
, xfs_log_worker
);
1262 log
->l_prev_block
= -1;
1263 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1264 xlog_assign_atomic_lsn(&log
->l_tail_lsn
, 1, 0);
1265 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
, 1, 0);
1266 log
->l_curr_cycle
= 1; /* 0 is bad since this is initial value */
1268 xlog_grant_head_init(&log
->l_reserve_head
);
1269 xlog_grant_head_init(&log
->l_write_head
);
1271 error
= EFSCORRUPTED
;
1272 if (xfs_sb_version_hassector(&mp
->m_sb
)) {
1273 log2_size
= mp
->m_sb
.sb_logsectlog
;
1274 if (log2_size
< BBSHIFT
) {
1275 xfs_warn(mp
, "Log sector size too small (0x%x < 0x%x)",
1276 log2_size
, BBSHIFT
);
1280 log2_size
-= BBSHIFT
;
1281 if (log2_size
> mp
->m_sectbb_log
) {
1282 xfs_warn(mp
, "Log sector size too large (0x%x > 0x%x)",
1283 log2_size
, mp
->m_sectbb_log
);
1287 /* for larger sector sizes, must have v2 or external log */
1288 if (log2_size
&& log
->l_logBBstart
> 0 &&
1289 !xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1291 "log sector size (0x%x) invalid for configuration.",
1296 log
->l_sectBBsize
= 1 << log2_size
;
1298 xlog_get_iclog_buffer_size(mp
, log
);
1301 bp
= xfs_buf_alloc(mp
->m_logdev_targp
, 0, BTOBB(log
->l_iclog_size
), 0);
1304 bp
->b_iodone
= xlog_iodone
;
1305 ASSERT(xfs_buf_islocked(bp
));
1308 spin_lock_init(&log
->l_icloglock
);
1309 init_waitqueue_head(&log
->l_flush_wait
);
1311 iclogp
= &log
->l_iclog
;
1313 * The amount of memory to allocate for the iclog structure is
1314 * rather funky due to the way the structure is defined. It is
1315 * done this way so that we can use different sizes for machines
1316 * with different amounts of memory. See the definition of
1317 * xlog_in_core_t in xfs_log_priv.h for details.
1319 ASSERT(log
->l_iclog_size
>= 4096);
1320 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
1321 *iclogp
= kmem_zalloc(sizeof(xlog_in_core_t
), KM_MAYFAIL
);
1323 goto out_free_iclog
;
1326 iclog
->ic_prev
= prev_iclog
;
1329 bp
= xfs_buf_get_uncached(mp
->m_logdev_targp
,
1330 BTOBB(log
->l_iclog_size
), 0);
1332 goto out_free_iclog
;
1334 bp
->b_iodone
= xlog_iodone
;
1336 iclog
->ic_data
= bp
->b_addr
;
1338 log
->l_iclog_bak
[i
] = (xfs_caddr_t
)&(iclog
->ic_header
);
1340 head
= &iclog
->ic_header
;
1341 memset(head
, 0, sizeof(xlog_rec_header_t
));
1342 head
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1343 head
->h_version
= cpu_to_be32(
1344 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1345 head
->h_size
= cpu_to_be32(log
->l_iclog_size
);
1347 head
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1348 memcpy(&head
->h_fs_uuid
, &mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1350 iclog
->ic_size
= BBTOB(bp
->b_length
) - log
->l_iclog_hsize
;
1351 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1352 iclog
->ic_log
= log
;
1353 atomic_set(&iclog
->ic_refcnt
, 0);
1354 spin_lock_init(&iclog
->ic_callback_lock
);
1355 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
1356 iclog
->ic_datap
= (char *)iclog
->ic_data
+ log
->l_iclog_hsize
;
1358 ASSERT(xfs_buf_islocked(iclog
->ic_bp
));
1359 init_waitqueue_head(&iclog
->ic_force_wait
);
1360 init_waitqueue_head(&iclog
->ic_write_wait
);
1362 iclogp
= &iclog
->ic_next
;
1364 *iclogp
= log
->l_iclog
; /* complete ring */
1365 log
->l_iclog
->ic_prev
= prev_iclog
; /* re-write 1st prev ptr */
1367 error
= xlog_cil_init(log
);
1369 goto out_free_iclog
;
1373 for (iclog
= log
->l_iclog
; iclog
; iclog
= prev_iclog
) {
1374 prev_iclog
= iclog
->ic_next
;
1376 xfs_buf_free(iclog
->ic_bp
);
1379 spinlock_destroy(&log
->l_icloglock
);
1380 xfs_buf_free(log
->l_xbuf
);
1384 return ERR_PTR(-error
);
1385 } /* xlog_alloc_log */
1389 * Write out the commit record of a transaction associated with the given
1390 * ticket. Return the lsn of the commit record.
1395 struct xlog_ticket
*ticket
,
1396 struct xlog_in_core
**iclog
,
1397 xfs_lsn_t
*commitlsnp
)
1399 struct xfs_mount
*mp
= log
->l_mp
;
1401 struct xfs_log_iovec reg
= {
1404 .i_type
= XLOG_REG_TYPE_COMMIT
,
1406 struct xfs_log_vec vec
= {
1411 ASSERT_ALWAYS(iclog
);
1412 error
= xlog_write(log
, &vec
, ticket
, commitlsnp
, iclog
,
1415 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
1420 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1421 * log space. This code pushes on the lsn which would supposedly free up
1422 * the 25% which we want to leave free. We may need to adopt a policy which
1423 * pushes on an lsn which is further along in the log once we reach the high
1424 * water mark. In this manner, we would be creating a low water mark.
1427 xlog_grant_push_ail(
1431 xfs_lsn_t threshold_lsn
= 0;
1432 xfs_lsn_t last_sync_lsn
;
1435 int threshold_block
;
1436 int threshold_cycle
;
1439 ASSERT(BTOBB(need_bytes
) < log
->l_logBBsize
);
1441 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1442 free_blocks
= BTOBBT(free_bytes
);
1445 * Set the threshold for the minimum number of free blocks in the
1446 * log to the maximum of what the caller needs, one quarter of the
1447 * log, and 256 blocks.
1449 free_threshold
= BTOBB(need_bytes
);
1450 free_threshold
= MAX(free_threshold
, (log
->l_logBBsize
>> 2));
1451 free_threshold
= MAX(free_threshold
, 256);
1452 if (free_blocks
>= free_threshold
)
1455 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &threshold_cycle
,
1457 threshold_block
+= free_threshold
;
1458 if (threshold_block
>= log
->l_logBBsize
) {
1459 threshold_block
-= log
->l_logBBsize
;
1460 threshold_cycle
+= 1;
1462 threshold_lsn
= xlog_assign_lsn(threshold_cycle
,
1465 * Don't pass in an lsn greater than the lsn of the last
1466 * log record known to be on disk. Use a snapshot of the last sync lsn
1467 * so that it doesn't change between the compare and the set.
1469 last_sync_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1470 if (XFS_LSN_CMP(threshold_lsn
, last_sync_lsn
) > 0)
1471 threshold_lsn
= last_sync_lsn
;
1474 * Get the transaction layer to kick the dirty buffers out to
1475 * disk asynchronously. No point in trying to do this if
1476 * the filesystem is shutting down.
1478 if (!XLOG_FORCED_SHUTDOWN(log
))
1479 xfs_ail_push(log
->l_ailp
, threshold_lsn
);
1483 * The bdstrat callback function for log bufs. This gives us a central
1484 * place to trap bufs in case we get hit by a log I/O error and need to
1485 * shutdown. Actually, in practice, even when we didn't get a log error,
1486 * we transition the iclogs to IOERROR state *after* flushing all existing
1487 * iclogs to disk. This is because we don't want anymore new transactions to be
1488 * started or completed afterwards.
1494 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1496 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1497 xfs_buf_ioerror(bp
, EIO
);
1499 xfs_buf_ioend(bp
, 0);
1501 * It would seem logical to return EIO here, but we rely on
1502 * the log state machine to propagate I/O errors instead of
1508 xfs_buf_iorequest(bp
);
1513 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1514 * fashion. Previously, we should have moved the current iclog
1515 * ptr in the log to point to the next available iclog. This allows further
1516 * write to continue while this code syncs out an iclog ready to go.
1517 * Before an in-core log can be written out, the data section must be scanned
1518 * to save away the 1st word of each BBSIZE block into the header. We replace
1519 * it with the current cycle count. Each BBSIZE block is tagged with the
1520 * cycle count because there in an implicit assumption that drives will
1521 * guarantee that entire 512 byte blocks get written at once. In other words,
1522 * we can't have part of a 512 byte block written and part not written. By
1523 * tagging each block, we will know which blocks are valid when recovering
1524 * after an unclean shutdown.
1526 * This routine is single threaded on the iclog. No other thread can be in
1527 * this routine with the same iclog. Changing contents of iclog can there-
1528 * fore be done without grabbing the state machine lock. Updating the global
1529 * log will require grabbing the lock though.
1531 * The entire log manager uses a logical block numbering scheme. Only
1532 * log_sync (and then only bwrite()) know about the fact that the log may
1533 * not start with block zero on a given device. The log block start offset
1534 * is added immediately before calling bwrite().
1540 struct xlog_in_core
*iclog
)
1542 xfs_caddr_t dptr
; /* pointer to byte sized element */
1545 uint count
; /* byte count of bwrite */
1546 uint count_init
; /* initial count before roundup */
1547 int roundoff
; /* roundoff to BB or stripe */
1548 int split
= 0; /* split write into two regions */
1550 int v2
= xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
);
1552 XFS_STATS_INC(xs_log_writes
);
1553 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
1555 /* Add for LR header */
1556 count_init
= log
->l_iclog_hsize
+ iclog
->ic_offset
;
1558 /* Round out the log write size */
1559 if (v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1) {
1560 /* we have a v2 stripe unit to use */
1561 count
= XLOG_LSUNITTOB(log
, XLOG_BTOLSUNIT(log
, count_init
));
1563 count
= BBTOB(BTOBB(count_init
));
1565 roundoff
= count
- count_init
;
1566 ASSERT(roundoff
>= 0);
1567 ASSERT((v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1 &&
1568 roundoff
< log
->l_mp
->m_sb
.sb_logsunit
)
1570 (log
->l_mp
->m_sb
.sb_logsunit
<= 1 &&
1571 roundoff
< BBTOB(1)));
1573 /* move grant heads by roundoff in sync */
1574 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, roundoff
);
1575 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, roundoff
);
1577 /* put cycle number in every block */
1578 xlog_pack_data(log
, iclog
, roundoff
);
1580 /* real byte length */
1582 iclog
->ic_header
.h_len
=
1583 cpu_to_be32(iclog
->ic_offset
+ roundoff
);
1585 iclog
->ic_header
.h_len
=
1586 cpu_to_be32(iclog
->ic_offset
);
1590 XFS_BUF_SET_ADDR(bp
, BLOCK_LSN(be64_to_cpu(iclog
->ic_header
.h_lsn
)));
1592 XFS_STATS_ADD(xs_log_blocks
, BTOBB(count
));
1594 /* Do we need to split this write into 2 parts? */
1595 if (XFS_BUF_ADDR(bp
) + BTOBB(count
) > log
->l_logBBsize
) {
1596 split
= count
- (BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
)));
1597 count
= BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
));
1598 iclog
->ic_bwritecnt
= 2; /* split into 2 writes */
1600 iclog
->ic_bwritecnt
= 1;
1602 bp
->b_io_length
= BTOBB(count
);
1603 bp
->b_fspriv
= iclog
;
1604 XFS_BUF_ZEROFLAGS(bp
);
1606 bp
->b_flags
|= XBF_SYNCIO
;
1608 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
) {
1609 bp
->b_flags
|= XBF_FUA
;
1612 * Flush the data device before flushing the log to make
1613 * sure all meta data written back from the AIL actually made
1614 * it to disk before stamping the new log tail LSN into the
1615 * log buffer. For an external log we need to issue the
1616 * flush explicitly, and unfortunately synchronously here;
1617 * for an internal log we can simply use the block layer
1618 * state machine for preflushes.
1620 if (log
->l_mp
->m_logdev_targp
!= log
->l_mp
->m_ddev_targp
)
1621 xfs_blkdev_issue_flush(log
->l_mp
->m_ddev_targp
);
1623 bp
->b_flags
|= XBF_FLUSH
;
1626 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1627 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1629 xlog_verify_iclog(log
, iclog
, count
, B_TRUE
);
1631 /* account for log which doesn't start at block #0 */
1632 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1634 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1639 error
= xlog_bdstrat(bp
);
1641 xfs_buf_ioerror_alert(bp
, "xlog_sync");
1645 bp
= iclog
->ic_log
->l_xbuf
;
1646 XFS_BUF_SET_ADDR(bp
, 0); /* logical 0 */
1647 xfs_buf_associate_memory(bp
,
1648 (char *)&iclog
->ic_header
+ count
, split
);
1649 bp
->b_fspriv
= iclog
;
1650 XFS_BUF_ZEROFLAGS(bp
);
1652 bp
->b_flags
|= XBF_SYNCIO
;
1653 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
)
1654 bp
->b_flags
|= XBF_FUA
;
1657 * Bump the cycle numbers at the start of each block
1658 * since this part of the buffer is at the start of
1659 * a new cycle. Watch out for the header magic number
1662 for (i
= 0; i
< split
; i
+= BBSIZE
) {
1663 be32_add_cpu((__be32
*)dptr
, 1);
1664 if (be32_to_cpu(*(__be32
*)dptr
) == XLOG_HEADER_MAGIC_NUM
)
1665 be32_add_cpu((__be32
*)dptr
, 1);
1669 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1670 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1672 /* account for internal log which doesn't start at block #0 */
1673 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1675 error
= xlog_bdstrat(bp
);
1677 xfs_buf_ioerror_alert(bp
, "xlog_sync (split)");
1686 * Deallocate a log structure
1692 xlog_in_core_t
*iclog
, *next_iclog
;
1695 xlog_cil_destroy(log
);
1698 * always need to ensure that the extra buffer does not point to memory
1699 * owned by another log buffer before we free it.
1701 xfs_buf_set_empty(log
->l_xbuf
, BTOBB(log
->l_iclog_size
));
1702 xfs_buf_free(log
->l_xbuf
);
1704 iclog
= log
->l_iclog
;
1705 for (i
=0; i
<log
->l_iclog_bufs
; i
++) {
1706 xfs_buf_free(iclog
->ic_bp
);
1707 next_iclog
= iclog
->ic_next
;
1711 spinlock_destroy(&log
->l_icloglock
);
1713 log
->l_mp
->m_log
= NULL
;
1715 } /* xlog_dealloc_log */
1718 * Update counters atomically now that memcpy is done.
1722 xlog_state_finish_copy(
1724 struct xlog_in_core
*iclog
,
1728 spin_lock(&log
->l_icloglock
);
1730 be32_add_cpu(&iclog
->ic_header
.h_num_logops
, record_cnt
);
1731 iclog
->ic_offset
+= copy_bytes
;
1733 spin_unlock(&log
->l_icloglock
);
1734 } /* xlog_state_finish_copy */
1740 * print out info relating to regions written which consume
1745 struct xfs_mount
*mp
,
1746 struct xlog_ticket
*ticket
)
1749 uint ophdr_spc
= ticket
->t_res_num_ophdrs
* (uint
)sizeof(xlog_op_header_t
);
1751 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1752 static char *res_type_str
[XLOG_REG_TYPE_MAX
] = {
1773 static char *trans_type_str
[XFS_TRANS_TYPE_MAX
] = {
1817 "xlog_write: reservation summary:\n"
1818 " trans type = %s (%u)\n"
1819 " unit res = %d bytes\n"
1820 " current res = %d bytes\n"
1821 " total reg = %u bytes (o/flow = %u bytes)\n"
1822 " ophdrs = %u (ophdr space = %u bytes)\n"
1823 " ophdr + reg = %u bytes\n"
1824 " num regions = %u\n",
1825 ((ticket
->t_trans_type
<= 0 ||
1826 ticket
->t_trans_type
> XFS_TRANS_TYPE_MAX
) ?
1827 "bad-trans-type" : trans_type_str
[ticket
->t_trans_type
-1]),
1828 ticket
->t_trans_type
,
1831 ticket
->t_res_arr_sum
, ticket
->t_res_o_flow
,
1832 ticket
->t_res_num_ophdrs
, ophdr_spc
,
1833 ticket
->t_res_arr_sum
+
1834 ticket
->t_res_o_flow
+ ophdr_spc
,
1837 for (i
= 0; i
< ticket
->t_res_num
; i
++) {
1838 uint r_type
= ticket
->t_res_arr
[i
].r_type
;
1839 xfs_warn(mp
, "region[%u]: %s - %u bytes\n", i
,
1840 ((r_type
<= 0 || r_type
> XLOG_REG_TYPE_MAX
) ?
1841 "bad-rtype" : res_type_str
[r_type
-1]),
1842 ticket
->t_res_arr
[i
].r_len
);
1845 xfs_alert_tag(mp
, XFS_PTAG_LOGRES
,
1846 "xlog_write: reservation ran out. Need to up reservation");
1847 xfs_force_shutdown(mp
, SHUTDOWN_CORRUPT_INCORE
);
1851 * Calculate the potential space needed by the log vector. Each region gets
1852 * its own xlog_op_header_t and may need to be double word aligned.
1855 xlog_write_calc_vec_length(
1856 struct xlog_ticket
*ticket
,
1857 struct xfs_log_vec
*log_vector
)
1859 struct xfs_log_vec
*lv
;
1864 /* acct for start rec of xact */
1865 if (ticket
->t_flags
& XLOG_TIC_INITED
)
1868 for (lv
= log_vector
; lv
; lv
= lv
->lv_next
) {
1869 headers
+= lv
->lv_niovecs
;
1871 for (i
= 0; i
< lv
->lv_niovecs
; i
++) {
1872 struct xfs_log_iovec
*vecp
= &lv
->lv_iovecp
[i
];
1875 xlog_tic_add_region(ticket
, vecp
->i_len
, vecp
->i_type
);
1879 ticket
->t_res_num_ophdrs
+= headers
;
1880 len
+= headers
* sizeof(struct xlog_op_header
);
1886 * If first write for transaction, insert start record We can't be trying to
1887 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1890 xlog_write_start_rec(
1891 struct xlog_op_header
*ophdr
,
1892 struct xlog_ticket
*ticket
)
1894 if (!(ticket
->t_flags
& XLOG_TIC_INITED
))
1897 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1898 ophdr
->oh_clientid
= ticket
->t_clientid
;
1900 ophdr
->oh_flags
= XLOG_START_TRANS
;
1903 ticket
->t_flags
&= ~XLOG_TIC_INITED
;
1905 return sizeof(struct xlog_op_header
);
1908 static xlog_op_header_t
*
1909 xlog_write_setup_ophdr(
1911 struct xlog_op_header
*ophdr
,
1912 struct xlog_ticket
*ticket
,
1915 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1916 ophdr
->oh_clientid
= ticket
->t_clientid
;
1919 /* are we copying a commit or unmount record? */
1920 ophdr
->oh_flags
= flags
;
1923 * We've seen logs corrupted with bad transaction client ids. This
1924 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1925 * and shut down the filesystem.
1927 switch (ophdr
->oh_clientid
) {
1928 case XFS_TRANSACTION
:
1934 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1935 ophdr
->oh_clientid
, ticket
);
1943 * Set up the parameters of the region copy into the log. This has
1944 * to handle region write split across multiple log buffers - this
1945 * state is kept external to this function so that this code can
1946 * can be written in an obvious, self documenting manner.
1949 xlog_write_setup_copy(
1950 struct xlog_ticket
*ticket
,
1951 struct xlog_op_header
*ophdr
,
1952 int space_available
,
1956 int *last_was_partial_copy
,
1957 int *bytes_consumed
)
1961 still_to_copy
= space_required
- *bytes_consumed
;
1962 *copy_off
= *bytes_consumed
;
1964 if (still_to_copy
<= space_available
) {
1965 /* write of region completes here */
1966 *copy_len
= still_to_copy
;
1967 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
1968 if (*last_was_partial_copy
)
1969 ophdr
->oh_flags
|= (XLOG_END_TRANS
|XLOG_WAS_CONT_TRANS
);
1970 *last_was_partial_copy
= 0;
1971 *bytes_consumed
= 0;
1975 /* partial write of region, needs extra log op header reservation */
1976 *copy_len
= space_available
;
1977 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
1978 ophdr
->oh_flags
|= XLOG_CONTINUE_TRANS
;
1979 if (*last_was_partial_copy
)
1980 ophdr
->oh_flags
|= XLOG_WAS_CONT_TRANS
;
1981 *bytes_consumed
+= *copy_len
;
1982 (*last_was_partial_copy
)++;
1984 /* account for new log op header */
1985 ticket
->t_curr_res
-= sizeof(struct xlog_op_header
);
1986 ticket
->t_res_num_ophdrs
++;
1988 return sizeof(struct xlog_op_header
);
1992 xlog_write_copy_finish(
1994 struct xlog_in_core
*iclog
,
1999 int *partial_copy_len
,
2001 struct xlog_in_core
**commit_iclog
)
2003 if (*partial_copy
) {
2005 * This iclog has already been marked WANT_SYNC by
2006 * xlog_state_get_iclog_space.
2008 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2011 return xlog_state_release_iclog(log
, iclog
);
2015 *partial_copy_len
= 0;
2017 if (iclog
->ic_size
- log_offset
<= sizeof(xlog_op_header_t
)) {
2018 /* no more space in this iclog - push it. */
2019 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2023 spin_lock(&log
->l_icloglock
);
2024 xlog_state_want_sync(log
, iclog
);
2025 spin_unlock(&log
->l_icloglock
);
2028 return xlog_state_release_iclog(log
, iclog
);
2029 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2030 *commit_iclog
= iclog
;
2037 * Write some region out to in-core log
2039 * This will be called when writing externally provided regions or when
2040 * writing out a commit record for a given transaction.
2042 * General algorithm:
2043 * 1. Find total length of this write. This may include adding to the
2044 * lengths passed in.
2045 * 2. Check whether we violate the tickets reservation.
2046 * 3. While writing to this iclog
2047 * A. Reserve as much space in this iclog as can get
2048 * B. If this is first write, save away start lsn
2049 * C. While writing this region:
2050 * 1. If first write of transaction, write start record
2051 * 2. Write log operation header (header per region)
2052 * 3. Find out if we can fit entire region into this iclog
2053 * 4. Potentially, verify destination memcpy ptr
2054 * 5. Memcpy (partial) region
2055 * 6. If partial copy, release iclog; otherwise, continue
2056 * copying more regions into current iclog
2057 * 4. Mark want sync bit (in simulation mode)
2058 * 5. Release iclog for potential flush to on-disk log.
2061 * 1. Panic if reservation is overrun. This should never happen since
2062 * reservation amounts are generated internal to the filesystem.
2064 * 1. Tickets are single threaded data structures.
2065 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2066 * syncing routine. When a single log_write region needs to span
2067 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2068 * on all log operation writes which don't contain the end of the
2069 * region. The XLOG_END_TRANS bit is used for the in-core log
2070 * operation which contains the end of the continued log_write region.
2071 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2072 * we don't really know exactly how much space will be used. As a result,
2073 * we don't update ic_offset until the end when we know exactly how many
2074 * bytes have been written out.
2079 struct xfs_log_vec
*log_vector
,
2080 struct xlog_ticket
*ticket
,
2081 xfs_lsn_t
*start_lsn
,
2082 struct xlog_in_core
**commit_iclog
,
2085 struct xlog_in_core
*iclog
= NULL
;
2086 struct xfs_log_iovec
*vecp
;
2087 struct xfs_log_vec
*lv
;
2090 int partial_copy
= 0;
2091 int partial_copy_len
= 0;
2099 len
= xlog_write_calc_vec_length(ticket
, log_vector
);
2102 * Region headers and bytes are already accounted for.
2103 * We only need to take into account start records and
2104 * split regions in this function.
2106 if (ticket
->t_flags
& XLOG_TIC_INITED
)
2107 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2110 * Commit record headers need to be accounted for. These
2111 * come in as separate writes so are easy to detect.
2113 if (flags
& (XLOG_COMMIT_TRANS
| XLOG_UNMOUNT_TRANS
))
2114 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2116 if (ticket
->t_curr_res
< 0)
2117 xlog_print_tic_res(log
->l_mp
, ticket
);
2121 vecp
= lv
->lv_iovecp
;
2122 while (lv
&& index
< lv
->lv_niovecs
) {
2126 error
= xlog_state_get_iclog_space(log
, len
, &iclog
, ticket
,
2127 &contwr
, &log_offset
);
2131 ASSERT(log_offset
<= iclog
->ic_size
- 1);
2132 ptr
= iclog
->ic_datap
+ log_offset
;
2134 /* start_lsn is the first lsn written to. That's all we need. */
2136 *start_lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2139 * This loop writes out as many regions as can fit in the amount
2140 * of space which was allocated by xlog_state_get_iclog_space().
2142 while (lv
&& index
< lv
->lv_niovecs
) {
2143 struct xfs_log_iovec
*reg
= &vecp
[index
];
2144 struct xlog_op_header
*ophdr
;
2149 ASSERT(reg
->i_len
% sizeof(__int32_t
) == 0);
2150 ASSERT((unsigned long)ptr
% sizeof(__int32_t
) == 0);
2152 start_rec_copy
= xlog_write_start_rec(ptr
, ticket
);
2153 if (start_rec_copy
) {
2155 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2159 ophdr
= xlog_write_setup_ophdr(log
, ptr
, ticket
, flags
);
2161 return XFS_ERROR(EIO
);
2163 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2164 sizeof(struct xlog_op_header
));
2166 len
+= xlog_write_setup_copy(ticket
, ophdr
,
2167 iclog
->ic_size
-log_offset
,
2169 ©_off
, ©_len
,
2172 xlog_verify_dest_ptr(log
, ptr
);
2175 ASSERT(copy_len
>= 0);
2176 memcpy(ptr
, reg
->i_addr
+ copy_off
, copy_len
);
2177 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
, copy_len
);
2179 copy_len
+= start_rec_copy
+ sizeof(xlog_op_header_t
);
2181 data_cnt
+= contwr
? copy_len
: 0;
2183 error
= xlog_write_copy_finish(log
, iclog
, flags
,
2184 &record_cnt
, &data_cnt
,
2193 * if we had a partial copy, we need to get more iclog
2194 * space but we don't want to increment the region
2195 * index because there is still more is this region to
2198 * If we completed writing this region, and we flushed
2199 * the iclog (indicated by resetting of the record
2200 * count), then we also need to get more log space. If
2201 * this was the last record, though, we are done and
2207 if (++index
== lv
->lv_niovecs
) {
2211 vecp
= lv
->lv_iovecp
;
2213 if (record_cnt
== 0) {
2223 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
2225 return xlog_state_release_iclog(log
, iclog
);
2227 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2228 *commit_iclog
= iclog
;
2233 /*****************************************************************************
2235 * State Machine functions
2237 *****************************************************************************
2240 /* Clean iclogs starting from the head. This ordering must be
2241 * maintained, so an iclog doesn't become ACTIVE beyond one that
2242 * is SYNCING. This is also required to maintain the notion that we use
2243 * a ordered wait queue to hold off would be writers to the log when every
2244 * iclog is trying to sync to disk.
2246 * State Change: DIRTY -> ACTIVE
2249 xlog_state_clean_log(
2252 xlog_in_core_t
*iclog
;
2255 iclog
= log
->l_iclog
;
2257 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2258 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
2259 iclog
->ic_offset
= 0;
2260 ASSERT(iclog
->ic_callback
== NULL
);
2262 * If the number of ops in this iclog indicate it just
2263 * contains the dummy transaction, we can
2264 * change state into IDLE (the second time around).
2265 * Otherwise we should change the state into
2267 * We don't need to cover the dummy.
2270 (be32_to_cpu(iclog
->ic_header
.h_num_logops
) ==
2275 * We have two dirty iclogs so start over
2276 * This could also be num of ops indicates
2277 * this is not the dummy going out.
2281 iclog
->ic_header
.h_num_logops
= 0;
2282 memset(iclog
->ic_header
.h_cycle_data
, 0,
2283 sizeof(iclog
->ic_header
.h_cycle_data
));
2284 iclog
->ic_header
.h_lsn
= 0;
2285 } else if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2288 break; /* stop cleaning */
2289 iclog
= iclog
->ic_next
;
2290 } while (iclog
!= log
->l_iclog
);
2292 /* log is locked when we are called */
2294 * Change state for the dummy log recording.
2295 * We usually go to NEED. But we go to NEED2 if the changed indicates
2296 * we are done writing the dummy record.
2297 * If we are done with the second dummy recored (DONE2), then
2301 switch (log
->l_covered_state
) {
2302 case XLOG_STATE_COVER_IDLE
:
2303 case XLOG_STATE_COVER_NEED
:
2304 case XLOG_STATE_COVER_NEED2
:
2305 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2308 case XLOG_STATE_COVER_DONE
:
2310 log
->l_covered_state
= XLOG_STATE_COVER_NEED2
;
2312 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2315 case XLOG_STATE_COVER_DONE2
:
2317 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
2319 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2326 } /* xlog_state_clean_log */
2329 xlog_get_lowest_lsn(
2332 xlog_in_core_t
*lsn_log
;
2333 xfs_lsn_t lowest_lsn
, lsn
;
2335 lsn_log
= log
->l_iclog
;
2338 if (!(lsn_log
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
))) {
2339 lsn
= be64_to_cpu(lsn_log
->ic_header
.h_lsn
);
2340 if ((lsn
&& !lowest_lsn
) ||
2341 (XFS_LSN_CMP(lsn
, lowest_lsn
) < 0)) {
2345 lsn_log
= lsn_log
->ic_next
;
2346 } while (lsn_log
!= log
->l_iclog
);
2352 xlog_state_do_callback(
2355 struct xlog_in_core
*ciclog
)
2357 xlog_in_core_t
*iclog
;
2358 xlog_in_core_t
*first_iclog
; /* used to know when we've
2359 * processed all iclogs once */
2360 xfs_log_callback_t
*cb
, *cb_next
;
2362 xfs_lsn_t lowest_lsn
;
2363 int ioerrors
; /* counter: iclogs with errors */
2364 int loopdidcallbacks
; /* flag: inner loop did callbacks*/
2365 int funcdidcallbacks
; /* flag: function did callbacks */
2366 int repeats
; /* for issuing console warnings if
2367 * looping too many times */
2370 spin_lock(&log
->l_icloglock
);
2371 first_iclog
= iclog
= log
->l_iclog
;
2373 funcdidcallbacks
= 0;
2378 * Scan all iclogs starting with the one pointed to by the
2379 * log. Reset this starting point each time the log is
2380 * unlocked (during callbacks).
2382 * Keep looping through iclogs until one full pass is made
2383 * without running any callbacks.
2385 first_iclog
= log
->l_iclog
;
2386 iclog
= log
->l_iclog
;
2387 loopdidcallbacks
= 0;
2392 /* skip all iclogs in the ACTIVE & DIRTY states */
2393 if (iclog
->ic_state
&
2394 (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
)) {
2395 iclog
= iclog
->ic_next
;
2400 * Between marking a filesystem SHUTDOWN and stopping
2401 * the log, we do flush all iclogs to disk (if there
2402 * wasn't a log I/O error). So, we do want things to
2403 * go smoothly in case of just a SHUTDOWN w/o a
2406 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
2408 * Can only perform callbacks in order. Since
2409 * this iclog is not in the DONE_SYNC/
2410 * DO_CALLBACK state, we skip the rest and
2411 * just try to clean up. If we set our iclog
2412 * to DO_CALLBACK, we will not process it when
2413 * we retry since a previous iclog is in the
2414 * CALLBACK and the state cannot change since
2415 * we are holding the l_icloglock.
2417 if (!(iclog
->ic_state
&
2418 (XLOG_STATE_DONE_SYNC
|
2419 XLOG_STATE_DO_CALLBACK
))) {
2420 if (ciclog
&& (ciclog
->ic_state
==
2421 XLOG_STATE_DONE_SYNC
)) {
2422 ciclog
->ic_state
= XLOG_STATE_DO_CALLBACK
;
2427 * We now have an iclog that is in either the
2428 * DO_CALLBACK or DONE_SYNC states. The other
2429 * states (WANT_SYNC, SYNCING, or CALLBACK were
2430 * caught by the above if and are going to
2431 * clean (i.e. we aren't doing their callbacks)
2436 * We will do one more check here to see if we
2437 * have chased our tail around.
2440 lowest_lsn
= xlog_get_lowest_lsn(log
);
2442 XFS_LSN_CMP(lowest_lsn
,
2443 be64_to_cpu(iclog
->ic_header
.h_lsn
)) < 0) {
2444 iclog
= iclog
->ic_next
;
2445 continue; /* Leave this iclog for
2449 iclog
->ic_state
= XLOG_STATE_CALLBACK
;
2453 * update the last_sync_lsn before we drop the
2454 * icloglock to ensure we are the only one that
2457 ASSERT(XFS_LSN_CMP(atomic64_read(&log
->l_last_sync_lsn
),
2458 be64_to_cpu(iclog
->ic_header
.h_lsn
)) <= 0);
2459 atomic64_set(&log
->l_last_sync_lsn
,
2460 be64_to_cpu(iclog
->ic_header
.h_lsn
));
2465 spin_unlock(&log
->l_icloglock
);
2468 * Keep processing entries in the callback list until
2469 * we come around and it is empty. We need to
2470 * atomically see that the list is empty and change the
2471 * state to DIRTY so that we don't miss any more
2472 * callbacks being added.
2474 spin_lock(&iclog
->ic_callback_lock
);
2475 cb
= iclog
->ic_callback
;
2477 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
2478 iclog
->ic_callback
= NULL
;
2479 spin_unlock(&iclog
->ic_callback_lock
);
2481 /* perform callbacks in the order given */
2482 for (; cb
; cb
= cb_next
) {
2483 cb_next
= cb
->cb_next
;
2484 cb
->cb_func(cb
->cb_arg
, aborted
);
2486 spin_lock(&iclog
->ic_callback_lock
);
2487 cb
= iclog
->ic_callback
;
2493 spin_lock(&log
->l_icloglock
);
2494 ASSERT(iclog
->ic_callback
== NULL
);
2495 spin_unlock(&iclog
->ic_callback_lock
);
2496 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
))
2497 iclog
->ic_state
= XLOG_STATE_DIRTY
;
2500 * Transition from DIRTY to ACTIVE if applicable.
2501 * NOP if STATE_IOERROR.
2503 xlog_state_clean_log(log
);
2505 /* wake up threads waiting in xfs_log_force() */
2506 wake_up_all(&iclog
->ic_force_wait
);
2508 iclog
= iclog
->ic_next
;
2509 } while (first_iclog
!= iclog
);
2511 if (repeats
> 5000) {
2512 flushcnt
+= repeats
;
2515 "%s: possible infinite loop (%d iterations)",
2516 __func__
, flushcnt
);
2518 } while (!ioerrors
&& loopdidcallbacks
);
2521 * make one last gasp attempt to see if iclogs are being left in
2525 if (funcdidcallbacks
) {
2526 first_iclog
= iclog
= log
->l_iclog
;
2528 ASSERT(iclog
->ic_state
!= XLOG_STATE_DO_CALLBACK
);
2530 * Terminate the loop if iclogs are found in states
2531 * which will cause other threads to clean up iclogs.
2533 * SYNCING - i/o completion will go through logs
2534 * DONE_SYNC - interrupt thread should be waiting for
2536 * IOERROR - give up hope all ye who enter here
2538 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
||
2539 iclog
->ic_state
== XLOG_STATE_SYNCING
||
2540 iclog
->ic_state
== XLOG_STATE_DONE_SYNC
||
2541 iclog
->ic_state
== XLOG_STATE_IOERROR
)
2543 iclog
= iclog
->ic_next
;
2544 } while (first_iclog
!= iclog
);
2548 if (log
->l_iclog
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_IOERROR
))
2550 spin_unlock(&log
->l_icloglock
);
2553 wake_up_all(&log
->l_flush_wait
);
2558 * Finish transitioning this iclog to the dirty state.
2560 * Make sure that we completely execute this routine only when this is
2561 * the last call to the iclog. There is a good chance that iclog flushes,
2562 * when we reach the end of the physical log, get turned into 2 separate
2563 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2564 * routine. By using the reference count bwritecnt, we guarantee that only
2565 * the second completion goes through.
2567 * Callbacks could take time, so they are done outside the scope of the
2568 * global state machine log lock.
2571 xlog_state_done_syncing(
2572 xlog_in_core_t
*iclog
,
2575 struct xlog
*log
= iclog
->ic_log
;
2577 spin_lock(&log
->l_icloglock
);
2579 ASSERT(iclog
->ic_state
== XLOG_STATE_SYNCING
||
2580 iclog
->ic_state
== XLOG_STATE_IOERROR
);
2581 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2582 ASSERT(iclog
->ic_bwritecnt
== 1 || iclog
->ic_bwritecnt
== 2);
2586 * If we got an error, either on the first buffer, or in the case of
2587 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2588 * and none should ever be attempted to be written to disk
2591 if (iclog
->ic_state
!= XLOG_STATE_IOERROR
) {
2592 if (--iclog
->ic_bwritecnt
== 1) {
2593 spin_unlock(&log
->l_icloglock
);
2596 iclog
->ic_state
= XLOG_STATE_DONE_SYNC
;
2600 * Someone could be sleeping prior to writing out the next
2601 * iclog buffer, we wake them all, one will get to do the
2602 * I/O, the others get to wait for the result.
2604 wake_up_all(&iclog
->ic_write_wait
);
2605 spin_unlock(&log
->l_icloglock
);
2606 xlog_state_do_callback(log
, aborted
, iclog
); /* also cleans log */
2607 } /* xlog_state_done_syncing */
2611 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2612 * sleep. We wait on the flush queue on the head iclog as that should be
2613 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2614 * we will wait here and all new writes will sleep until a sync completes.
2616 * The in-core logs are used in a circular fashion. They are not used
2617 * out-of-order even when an iclog past the head is free.
2620 * * log_offset where xlog_write() can start writing into the in-core
2622 * * in-core log pointer to which xlog_write() should write.
2623 * * boolean indicating this is a continued write to an in-core log.
2624 * If this is the last write, then the in-core log's offset field
2625 * needs to be incremented, depending on the amount of data which
2629 xlog_state_get_iclog_space(
2632 struct xlog_in_core
**iclogp
,
2633 struct xlog_ticket
*ticket
,
2634 int *continued_write
,
2638 xlog_rec_header_t
*head
;
2639 xlog_in_core_t
*iclog
;
2643 spin_lock(&log
->l_icloglock
);
2644 if (XLOG_FORCED_SHUTDOWN(log
)) {
2645 spin_unlock(&log
->l_icloglock
);
2646 return XFS_ERROR(EIO
);
2649 iclog
= log
->l_iclog
;
2650 if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
) {
2651 XFS_STATS_INC(xs_log_noiclogs
);
2653 /* Wait for log writes to have flushed */
2654 xlog_wait(&log
->l_flush_wait
, &log
->l_icloglock
);
2658 head
= &iclog
->ic_header
;
2660 atomic_inc(&iclog
->ic_refcnt
); /* prevents sync */
2661 log_offset
= iclog
->ic_offset
;
2663 /* On the 1st write to an iclog, figure out lsn. This works
2664 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2665 * committing to. If the offset is set, that's how many blocks
2668 if (log_offset
== 0) {
2669 ticket
->t_curr_res
-= log
->l_iclog_hsize
;
2670 xlog_tic_add_region(ticket
,
2672 XLOG_REG_TYPE_LRHEADER
);
2673 head
->h_cycle
= cpu_to_be32(log
->l_curr_cycle
);
2674 head
->h_lsn
= cpu_to_be64(
2675 xlog_assign_lsn(log
->l_curr_cycle
, log
->l_curr_block
));
2676 ASSERT(log
->l_curr_block
>= 0);
2679 /* If there is enough room to write everything, then do it. Otherwise,
2680 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2681 * bit is on, so this will get flushed out. Don't update ic_offset
2682 * until you know exactly how many bytes get copied. Therefore, wait
2683 * until later to update ic_offset.
2685 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2686 * can fit into remaining data section.
2688 if (iclog
->ic_size
- iclog
->ic_offset
< 2*sizeof(xlog_op_header_t
)) {
2689 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2692 * If I'm the only one writing to this iclog, sync it to disk.
2693 * We need to do an atomic compare and decrement here to avoid
2694 * racing with concurrent atomic_dec_and_lock() calls in
2695 * xlog_state_release_iclog() when there is more than one
2696 * reference to the iclog.
2698 if (!atomic_add_unless(&iclog
->ic_refcnt
, -1, 1)) {
2699 /* we are the only one */
2700 spin_unlock(&log
->l_icloglock
);
2701 error
= xlog_state_release_iclog(log
, iclog
);
2705 spin_unlock(&log
->l_icloglock
);
2710 /* Do we have enough room to write the full amount in the remainder
2711 * of this iclog? Or must we continue a write on the next iclog and
2712 * mark this iclog as completely taken? In the case where we switch
2713 * iclogs (to mark it taken), this particular iclog will release/sync
2714 * to disk in xlog_write().
2716 if (len
<= iclog
->ic_size
- iclog
->ic_offset
) {
2717 *continued_write
= 0;
2718 iclog
->ic_offset
+= len
;
2720 *continued_write
= 1;
2721 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2725 ASSERT(iclog
->ic_offset
<= iclog
->ic_size
);
2726 spin_unlock(&log
->l_icloglock
);
2728 *logoffsetp
= log_offset
;
2730 } /* xlog_state_get_iclog_space */
2732 /* The first cnt-1 times through here we don't need to
2733 * move the grant write head because the permanent
2734 * reservation has reserved cnt times the unit amount.
2735 * Release part of current permanent unit reservation and
2736 * reset current reservation to be one units worth. Also
2737 * move grant reservation head forward.
2740 xlog_regrant_reserve_log_space(
2742 struct xlog_ticket
*ticket
)
2744 trace_xfs_log_regrant_reserve_enter(log
, ticket
);
2746 if (ticket
->t_cnt
> 0)
2749 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
,
2750 ticket
->t_curr_res
);
2751 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
,
2752 ticket
->t_curr_res
);
2753 ticket
->t_curr_res
= ticket
->t_unit_res
;
2754 xlog_tic_reset_res(ticket
);
2756 trace_xfs_log_regrant_reserve_sub(log
, ticket
);
2758 /* just return if we still have some of the pre-reserved space */
2759 if (ticket
->t_cnt
> 0)
2762 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
,
2763 ticket
->t_unit_res
);
2765 trace_xfs_log_regrant_reserve_exit(log
, ticket
);
2767 ticket
->t_curr_res
= ticket
->t_unit_res
;
2768 xlog_tic_reset_res(ticket
);
2769 } /* xlog_regrant_reserve_log_space */
2773 * Give back the space left from a reservation.
2775 * All the information we need to make a correct determination of space left
2776 * is present. For non-permanent reservations, things are quite easy. The
2777 * count should have been decremented to zero. We only need to deal with the
2778 * space remaining in the current reservation part of the ticket. If the
2779 * ticket contains a permanent reservation, there may be left over space which
2780 * needs to be released. A count of N means that N-1 refills of the current
2781 * reservation can be done before we need to ask for more space. The first
2782 * one goes to fill up the first current reservation. Once we run out of
2783 * space, the count will stay at zero and the only space remaining will be
2784 * in the current reservation field.
2787 xlog_ungrant_log_space(
2789 struct xlog_ticket
*ticket
)
2793 if (ticket
->t_cnt
> 0)
2796 trace_xfs_log_ungrant_enter(log
, ticket
);
2797 trace_xfs_log_ungrant_sub(log
, ticket
);
2800 * If this is a permanent reservation ticket, we may be able to free
2801 * up more space based on the remaining count.
2803 bytes
= ticket
->t_curr_res
;
2804 if (ticket
->t_cnt
> 0) {
2805 ASSERT(ticket
->t_flags
& XLOG_TIC_PERM_RESERV
);
2806 bytes
+= ticket
->t_unit_res
*ticket
->t_cnt
;
2809 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
, bytes
);
2810 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
, bytes
);
2812 trace_xfs_log_ungrant_exit(log
, ticket
);
2814 xfs_log_space_wake(log
->l_mp
);
2818 * Flush iclog to disk if this is the last reference to the given iclog and
2819 * the WANT_SYNC bit is set.
2821 * When this function is entered, the iclog is not necessarily in the
2822 * WANT_SYNC state. It may be sitting around waiting to get filled.
2827 xlog_state_release_iclog(
2829 struct xlog_in_core
*iclog
)
2831 int sync
= 0; /* do we sync? */
2833 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
2834 return XFS_ERROR(EIO
);
2836 ASSERT(atomic_read(&iclog
->ic_refcnt
) > 0);
2837 if (!atomic_dec_and_lock(&iclog
->ic_refcnt
, &log
->l_icloglock
))
2840 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2841 spin_unlock(&log
->l_icloglock
);
2842 return XFS_ERROR(EIO
);
2844 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2845 iclog
->ic_state
== XLOG_STATE_WANT_SYNC
);
2847 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
) {
2848 /* update tail before writing to iclog */
2849 xfs_lsn_t tail_lsn
= xlog_assign_tail_lsn(log
->l_mp
);
2851 iclog
->ic_state
= XLOG_STATE_SYNCING
;
2852 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(tail_lsn
);
2853 xlog_verify_tail_lsn(log
, iclog
, tail_lsn
);
2854 /* cycle incremented when incrementing curr_block */
2856 spin_unlock(&log
->l_icloglock
);
2859 * We let the log lock go, so it's possible that we hit a log I/O
2860 * error or some other SHUTDOWN condition that marks the iclog
2861 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2862 * this iclog has consistent data, so we ignore IOERROR
2863 * flags after this point.
2866 return xlog_sync(log
, iclog
);
2868 } /* xlog_state_release_iclog */
2872 * This routine will mark the current iclog in the ring as WANT_SYNC
2873 * and move the current iclog pointer to the next iclog in the ring.
2874 * When this routine is called from xlog_state_get_iclog_space(), the
2875 * exact size of the iclog has not yet been determined. All we know is
2876 * that every data block. We have run out of space in this log record.
2879 xlog_state_switch_iclogs(
2881 struct xlog_in_core
*iclog
,
2884 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
2886 eventual_size
= iclog
->ic_offset
;
2887 iclog
->ic_state
= XLOG_STATE_WANT_SYNC
;
2888 iclog
->ic_header
.h_prev_block
= cpu_to_be32(log
->l_prev_block
);
2889 log
->l_prev_block
= log
->l_curr_block
;
2890 log
->l_prev_cycle
= log
->l_curr_cycle
;
2892 /* roll log?: ic_offset changed later */
2893 log
->l_curr_block
+= BTOBB(eventual_size
)+BTOBB(log
->l_iclog_hsize
);
2895 /* Round up to next log-sunit */
2896 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
2897 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
2898 __uint32_t sunit_bb
= BTOBB(log
->l_mp
->m_sb
.sb_logsunit
);
2899 log
->l_curr_block
= roundup(log
->l_curr_block
, sunit_bb
);
2902 if (log
->l_curr_block
>= log
->l_logBBsize
) {
2903 log
->l_curr_cycle
++;
2904 if (log
->l_curr_cycle
== XLOG_HEADER_MAGIC_NUM
)
2905 log
->l_curr_cycle
++;
2906 log
->l_curr_block
-= log
->l_logBBsize
;
2907 ASSERT(log
->l_curr_block
>= 0);
2909 ASSERT(iclog
== log
->l_iclog
);
2910 log
->l_iclog
= iclog
->ic_next
;
2911 } /* xlog_state_switch_iclogs */
2914 * Write out all data in the in-core log as of this exact moment in time.
2916 * Data may be written to the in-core log during this call. However,
2917 * we don't guarantee this data will be written out. A change from past
2918 * implementation means this routine will *not* write out zero length LRs.
2920 * Basically, we try and perform an intelligent scan of the in-core logs.
2921 * If we determine there is no flushable data, we just return. There is no
2922 * flushable data if:
2924 * 1. the current iclog is active and has no data; the previous iclog
2925 * is in the active or dirty state.
2926 * 2. the current iclog is drity, and the previous iclog is in the
2927 * active or dirty state.
2931 * 1. the current iclog is not in the active nor dirty state.
2932 * 2. the current iclog dirty, and the previous iclog is not in the
2933 * active nor dirty state.
2934 * 3. the current iclog is active, and there is another thread writing
2935 * to this particular iclog.
2936 * 4. a) the current iclog is active and has no other writers
2937 * b) when we return from flushing out this iclog, it is still
2938 * not in the active nor dirty state.
2942 struct xfs_mount
*mp
,
2946 struct xlog
*log
= mp
->m_log
;
2947 struct xlog_in_core
*iclog
;
2950 XFS_STATS_INC(xs_log_force
);
2952 xlog_cil_force(log
);
2954 spin_lock(&log
->l_icloglock
);
2956 iclog
= log
->l_iclog
;
2957 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2958 spin_unlock(&log
->l_icloglock
);
2959 return XFS_ERROR(EIO
);
2962 /* If the head iclog is not active nor dirty, we just attach
2963 * ourselves to the head and go to sleep.
2965 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2966 iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2968 * If the head is dirty or (active and empty), then
2969 * we need to look at the previous iclog. If the previous
2970 * iclog is active or dirty we are done. There is nothing
2971 * to sync out. Otherwise, we attach ourselves to the
2972 * previous iclog and go to sleep.
2974 if (iclog
->ic_state
== XLOG_STATE_DIRTY
||
2975 (atomic_read(&iclog
->ic_refcnt
) == 0
2976 && iclog
->ic_offset
== 0)) {
2977 iclog
= iclog
->ic_prev
;
2978 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2979 iclog
->ic_state
== XLOG_STATE_DIRTY
)
2984 if (atomic_read(&iclog
->ic_refcnt
) == 0) {
2985 /* We are the only one with access to this
2986 * iclog. Flush it out now. There should
2987 * be a roundoff of zero to show that someone
2988 * has already taken care of the roundoff from
2989 * the previous sync.
2991 atomic_inc(&iclog
->ic_refcnt
);
2992 lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2993 xlog_state_switch_iclogs(log
, iclog
, 0);
2994 spin_unlock(&log
->l_icloglock
);
2996 if (xlog_state_release_iclog(log
, iclog
))
2997 return XFS_ERROR(EIO
);
3001 spin_lock(&log
->l_icloglock
);
3002 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) == lsn
&&
3003 iclog
->ic_state
!= XLOG_STATE_DIRTY
)
3008 /* Someone else is writing to this iclog.
3009 * Use its call to flush out the data. However,
3010 * the other thread may not force out this LR,
3011 * so we mark it WANT_SYNC.
3013 xlog_state_switch_iclogs(log
, iclog
, 0);
3019 /* By the time we come around again, the iclog could've been filled
3020 * which would give it another lsn. If we have a new lsn, just
3021 * return because the relevant data has been flushed.
3024 if (flags
& XFS_LOG_SYNC
) {
3026 * We must check if we're shutting down here, before
3027 * we wait, while we're holding the l_icloglock.
3028 * Then we check again after waking up, in case our
3029 * sleep was disturbed by a bad news.
3031 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3032 spin_unlock(&log
->l_icloglock
);
3033 return XFS_ERROR(EIO
);
3035 XFS_STATS_INC(xs_log_force_sleep
);
3036 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3038 * No need to grab the log lock here since we're
3039 * only deciding whether or not to return EIO
3040 * and the memory read should be atomic.
3042 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3043 return XFS_ERROR(EIO
);
3049 spin_unlock(&log
->l_icloglock
);
3055 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3056 * about errors or whether the log was flushed or not. This is the normal
3057 * interface to use when trying to unpin items or move the log forward.
3066 trace_xfs_log_force(mp
, 0);
3067 error
= _xfs_log_force(mp
, flags
, NULL
);
3069 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3073 * Force the in-core log to disk for a specific LSN.
3075 * Find in-core log with lsn.
3076 * If it is in the DIRTY state, just return.
3077 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3078 * state and go to sleep or return.
3079 * If it is in any other state, go to sleep or return.
3081 * Synchronous forces are implemented with a signal variable. All callers
3082 * to force a given lsn to disk will wait on a the sv attached to the
3083 * specific in-core log. When given in-core log finally completes its
3084 * write to disk, that thread will wake up all threads waiting on the
3089 struct xfs_mount
*mp
,
3094 struct xlog
*log
= mp
->m_log
;
3095 struct xlog_in_core
*iclog
;
3096 int already_slept
= 0;
3100 XFS_STATS_INC(xs_log_force
);
3102 lsn
= xlog_cil_force_lsn(log
, lsn
);
3103 if (lsn
== NULLCOMMITLSN
)
3107 spin_lock(&log
->l_icloglock
);
3108 iclog
= log
->l_iclog
;
3109 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3110 spin_unlock(&log
->l_icloglock
);
3111 return XFS_ERROR(EIO
);
3115 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
) {
3116 iclog
= iclog
->ic_next
;
3120 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
3121 spin_unlock(&log
->l_icloglock
);
3125 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3127 * We sleep here if we haven't already slept (e.g.
3128 * this is the first time we've looked at the correct
3129 * iclog buf) and the buffer before us is going to
3130 * be sync'ed. The reason for this is that if we
3131 * are doing sync transactions here, by waiting for
3132 * the previous I/O to complete, we can allow a few
3133 * more transactions into this iclog before we close
3136 * Otherwise, we mark the buffer WANT_SYNC, and bump
3137 * up the refcnt so we can release the log (which
3138 * drops the ref count). The state switch keeps new
3139 * transaction commits from using this buffer. When
3140 * the current commits finish writing into the buffer,
3141 * the refcount will drop to zero and the buffer will
3144 if (!already_slept
&&
3145 (iclog
->ic_prev
->ic_state
&
3146 (XLOG_STATE_WANT_SYNC
| XLOG_STATE_SYNCING
))) {
3147 ASSERT(!(iclog
->ic_state
& XLOG_STATE_IOERROR
));
3149 XFS_STATS_INC(xs_log_force_sleep
);
3151 xlog_wait(&iclog
->ic_prev
->ic_write_wait
,
3158 atomic_inc(&iclog
->ic_refcnt
);
3159 xlog_state_switch_iclogs(log
, iclog
, 0);
3160 spin_unlock(&log
->l_icloglock
);
3161 if (xlog_state_release_iclog(log
, iclog
))
3162 return XFS_ERROR(EIO
);
3165 spin_lock(&log
->l_icloglock
);
3168 if ((flags
& XFS_LOG_SYNC
) && /* sleep */
3170 (XLOG_STATE_ACTIVE
| XLOG_STATE_DIRTY
))) {
3172 * Don't wait on completion if we know that we've
3173 * gotten a log write error.
3175 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3176 spin_unlock(&log
->l_icloglock
);
3177 return XFS_ERROR(EIO
);
3179 XFS_STATS_INC(xs_log_force_sleep
);
3180 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3182 * No need to grab the log lock here since we're
3183 * only deciding whether or not to return EIO
3184 * and the memory read should be atomic.
3186 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3187 return XFS_ERROR(EIO
);
3191 } else { /* just return */
3192 spin_unlock(&log
->l_icloglock
);
3196 } while (iclog
!= log
->l_iclog
);
3198 spin_unlock(&log
->l_icloglock
);
3203 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3204 * about errors or whether the log was flushed or not. This is the normal
3205 * interface to use when trying to unpin items or move the log forward.
3215 trace_xfs_log_force(mp
, lsn
);
3216 error
= _xfs_log_force_lsn(mp
, lsn
, flags
, NULL
);
3218 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3222 * Called when we want to mark the current iclog as being ready to sync to
3226 xlog_state_want_sync(
3228 struct xlog_in_core
*iclog
)
3230 assert_spin_locked(&log
->l_icloglock
);
3232 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3233 xlog_state_switch_iclogs(log
, iclog
, 0);
3235 ASSERT(iclog
->ic_state
&
3236 (XLOG_STATE_WANT_SYNC
|XLOG_STATE_IOERROR
));
3241 /*****************************************************************************
3245 *****************************************************************************
3249 * Free a used ticket when its refcount falls to zero.
3253 xlog_ticket_t
*ticket
)
3255 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3256 if (atomic_dec_and_test(&ticket
->t_ref
))
3257 kmem_zone_free(xfs_log_ticket_zone
, ticket
);
3262 xlog_ticket_t
*ticket
)
3264 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3265 atomic_inc(&ticket
->t_ref
);
3270 * Allocate and initialise a new log ticket.
3272 struct xlog_ticket
*
3279 xfs_km_flags_t alloc_flags
)
3281 struct xlog_ticket
*tic
;
3285 tic
= kmem_zone_zalloc(xfs_log_ticket_zone
, alloc_flags
);
3290 * Permanent reservations have up to 'cnt'-1 active log operations
3291 * in the log. A unit in this case is the amount of space for one
3292 * of these log operations. Normal reservations have a cnt of 1
3293 * and their unit amount is the total amount of space required.
3295 * The following lines of code account for non-transaction data
3296 * which occupy space in the on-disk log.
3298 * Normal form of a transaction is:
3299 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3300 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3302 * We need to account for all the leadup data and trailer data
3303 * around the transaction data.
3304 * And then we need to account for the worst case in terms of using
3306 * The worst case will happen if:
3307 * - the placement of the transaction happens to be such that the
3308 * roundoff is at its maximum
3309 * - the transaction data is synced before the commit record is synced
3310 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3311 * Therefore the commit record is in its own Log Record.
3312 * This can happen as the commit record is called with its
3313 * own region to xlog_write().
3314 * This then means that in the worst case, roundoff can happen for
3315 * the commit-rec as well.
3316 * The commit-rec is smaller than padding in this scenario and so it is
3317 * not added separately.
3320 /* for trans header */
3321 unit_bytes
+= sizeof(xlog_op_header_t
);
3322 unit_bytes
+= sizeof(xfs_trans_header_t
);
3325 unit_bytes
+= sizeof(xlog_op_header_t
);
3328 * for LR headers - the space for data in an iclog is the size minus
3329 * the space used for the headers. If we use the iclog size, then we
3330 * undercalculate the number of headers required.
3332 * Furthermore - the addition of op headers for split-recs might
3333 * increase the space required enough to require more log and op
3334 * headers, so take that into account too.
3336 * IMPORTANT: This reservation makes the assumption that if this
3337 * transaction is the first in an iclog and hence has the LR headers
3338 * accounted to it, then the remaining space in the iclog is
3339 * exclusively for this transaction. i.e. if the transaction is larger
3340 * than the iclog, it will be the only thing in that iclog.
3341 * Fundamentally, this means we must pass the entire log vector to
3342 * xlog_write to guarantee this.
3344 iclog_space
= log
->l_iclog_size
- log
->l_iclog_hsize
;
3345 num_headers
= howmany(unit_bytes
, iclog_space
);
3347 /* for split-recs - ophdrs added when data split over LRs */
3348 unit_bytes
+= sizeof(xlog_op_header_t
) * num_headers
;
3350 /* add extra header reservations if we overrun */
3351 while (!num_headers
||
3352 howmany(unit_bytes
, iclog_space
) > num_headers
) {
3353 unit_bytes
+= sizeof(xlog_op_header_t
);
3356 unit_bytes
+= log
->l_iclog_hsize
* num_headers
;
3358 /* for commit-rec LR header - note: padding will subsume the ophdr */
3359 unit_bytes
+= log
->l_iclog_hsize
;
3361 /* for roundoff padding for transaction data and one for commit record */
3362 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
3363 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
3364 /* log su roundoff */
3365 unit_bytes
+= 2*log
->l_mp
->m_sb
.sb_logsunit
;
3368 unit_bytes
+= 2*BBSIZE
;
3371 atomic_set(&tic
->t_ref
, 1);
3372 tic
->t_task
= current
;
3373 INIT_LIST_HEAD(&tic
->t_queue
);
3374 tic
->t_unit_res
= unit_bytes
;
3375 tic
->t_curr_res
= unit_bytes
;
3378 tic
->t_tid
= random32();
3379 tic
->t_clientid
= client
;
3380 tic
->t_flags
= XLOG_TIC_INITED
;
3381 tic
->t_trans_type
= 0;
3383 tic
->t_flags
|= XLOG_TIC_PERM_RESERV
;
3385 xlog_tic_reset_res(tic
);
3391 /******************************************************************************
3393 * Log debug routines
3395 ******************************************************************************
3399 * Make sure that the destination ptr is within the valid data region of
3400 * one of the iclogs. This uses backup pointers stored in a different
3401 * part of the log in case we trash the log structure.
3404 xlog_verify_dest_ptr(
3411 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
3412 if (ptr
>= log
->l_iclog_bak
[i
] &&
3413 ptr
<= log
->l_iclog_bak
[i
] + log
->l_iclog_size
)
3418 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3422 * Check to make sure the grant write head didn't just over lap the tail. If
3423 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3424 * the cycles differ by exactly one and check the byte count.
3426 * This check is run unlocked, so can give false positives. Rather than assert
3427 * on failures, use a warn-once flag and a panic tag to allow the admin to
3428 * determine if they want to panic the machine when such an error occurs. For
3429 * debug kernels this will have the same effect as using an assert but, unlinke
3430 * an assert, it can be turned off at runtime.
3433 xlog_verify_grant_tail(
3436 int tail_cycle
, tail_blocks
;
3439 xlog_crack_grant_head(&log
->l_write_head
.grant
, &cycle
, &space
);
3440 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_blocks
);
3441 if (tail_cycle
!= cycle
) {
3442 if (cycle
- 1 != tail_cycle
&&
3443 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3444 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3445 "%s: cycle - 1 != tail_cycle", __func__
);
3446 log
->l_flags
|= XLOG_TAIL_WARN
;
3449 if (space
> BBTOB(tail_blocks
) &&
3450 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3451 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3452 "%s: space > BBTOB(tail_blocks)", __func__
);
3453 log
->l_flags
|= XLOG_TAIL_WARN
;
3458 /* check if it will fit */
3460 xlog_verify_tail_lsn(
3462 struct xlog_in_core
*iclog
,
3467 if (CYCLE_LSN(tail_lsn
) == log
->l_prev_cycle
) {
3469 log
->l_logBBsize
- (log
->l_prev_block
- BLOCK_LSN(tail_lsn
));
3470 if (blocks
< BTOBB(iclog
->ic_offset
)+BTOBB(log
->l_iclog_hsize
))
3471 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3473 ASSERT(CYCLE_LSN(tail_lsn
)+1 == log
->l_prev_cycle
);
3475 if (BLOCK_LSN(tail_lsn
) == log
->l_prev_block
)
3476 xfs_emerg(log
->l_mp
, "%s: tail wrapped", __func__
);
3478 blocks
= BLOCK_LSN(tail_lsn
) - log
->l_prev_block
;
3479 if (blocks
< BTOBB(iclog
->ic_offset
) + 1)
3480 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3482 } /* xlog_verify_tail_lsn */
3485 * Perform a number of checks on the iclog before writing to disk.
3487 * 1. Make sure the iclogs are still circular
3488 * 2. Make sure we have a good magic number
3489 * 3. Make sure we don't have magic numbers in the data
3490 * 4. Check fields of each log operation header for:
3491 * A. Valid client identifier
3492 * B. tid ptr value falls in valid ptr space (user space code)
3493 * C. Length in log record header is correct according to the
3494 * individual operation headers within record.
3495 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3496 * log, check the preceding blocks of the physical log to make sure all
3497 * the cycle numbers agree with the current cycle number.
3502 struct xlog_in_core
*iclog
,
3506 xlog_op_header_t
*ophead
;
3507 xlog_in_core_t
*icptr
;
3508 xlog_in_core_2_t
*xhdr
;
3510 xfs_caddr_t base_ptr
;
3511 __psint_t field_offset
;
3513 int len
, i
, j
, k
, op_len
;
3516 /* check validity of iclog pointers */
3517 spin_lock(&log
->l_icloglock
);
3518 icptr
= log
->l_iclog
;
3519 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
3521 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3522 icptr
= icptr
->ic_next
;
3524 if (icptr
!= log
->l_iclog
)
3525 xfs_emerg(log
->l_mp
, "%s: corrupt iclog ring", __func__
);
3526 spin_unlock(&log
->l_icloglock
);
3528 /* check log magic numbers */
3529 if (iclog
->ic_header
.h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3530 xfs_emerg(log
->l_mp
, "%s: invalid magic num", __func__
);
3532 ptr
= (xfs_caddr_t
) &iclog
->ic_header
;
3533 for (ptr
+= BBSIZE
; ptr
< ((xfs_caddr_t
)&iclog
->ic_header
) + count
;
3535 if (*(__be32
*)ptr
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3536 xfs_emerg(log
->l_mp
, "%s: unexpected magic num",
3541 len
= be32_to_cpu(iclog
->ic_header
.h_num_logops
);
3542 ptr
= iclog
->ic_datap
;
3544 ophead
= (xlog_op_header_t
*)ptr
;
3545 xhdr
= iclog
->ic_data
;
3546 for (i
= 0; i
< len
; i
++) {
3547 ophead
= (xlog_op_header_t
*)ptr
;
3549 /* clientid is only 1 byte */
3550 field_offset
= (__psint_t
)
3551 ((xfs_caddr_t
)&(ophead
->oh_clientid
) - base_ptr
);
3552 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3553 clientid
= ophead
->oh_clientid
;
3555 idx
= BTOBBT((xfs_caddr_t
)&(ophead
->oh_clientid
) - iclog
->ic_datap
);
3556 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3557 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3558 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3559 clientid
= xlog_get_client_id(
3560 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3562 clientid
= xlog_get_client_id(
3563 iclog
->ic_header
.h_cycle_data
[idx
]);
3566 if (clientid
!= XFS_TRANSACTION
&& clientid
!= XFS_LOG
)
3568 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3569 __func__
, clientid
, ophead
,
3570 (unsigned long)field_offset
);
3573 field_offset
= (__psint_t
)
3574 ((xfs_caddr_t
)&(ophead
->oh_len
) - base_ptr
);
3575 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3576 op_len
= be32_to_cpu(ophead
->oh_len
);
3578 idx
= BTOBBT((__psint_t
)&ophead
->oh_len
-
3579 (__psint_t
)iclog
->ic_datap
);
3580 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3581 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3582 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3583 op_len
= be32_to_cpu(xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3585 op_len
= be32_to_cpu(iclog
->ic_header
.h_cycle_data
[idx
]);
3588 ptr
+= sizeof(xlog_op_header_t
) + op_len
;
3590 } /* xlog_verify_iclog */
3594 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3600 xlog_in_core_t
*iclog
, *ic
;
3602 iclog
= log
->l_iclog
;
3603 if (! (iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3605 * Mark all the incore logs IOERROR.
3606 * From now on, no log flushes will result.
3610 ic
->ic_state
= XLOG_STATE_IOERROR
;
3612 } while (ic
!= iclog
);
3616 * Return non-zero, if state transition has already happened.
3622 * This is called from xfs_force_shutdown, when we're forcibly
3623 * shutting down the filesystem, typically because of an IO error.
3624 * Our main objectives here are to make sure that:
3625 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3626 * parties to find out, 'atomically'.
3627 * b. those who're sleeping on log reservations, pinned objects and
3628 * other resources get woken up, and be told the bad news.
3629 * c. nothing new gets queued up after (a) and (b) are done.
3630 * d. if !logerror, flush the iclogs to disk, then seal them off
3633 * Note: for delayed logging the !logerror case needs to flush the regions
3634 * held in memory out to the iclogs before flushing them to disk. This needs
3635 * to be done before the log is marked as shutdown, otherwise the flush to the
3639 xfs_log_force_umount(
3640 struct xfs_mount
*mp
,
3649 * If this happens during log recovery, don't worry about
3650 * locking; the log isn't open for business yet.
3653 log
->l_flags
& XLOG_ACTIVE_RECOVERY
) {
3654 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3656 XFS_BUF_DONE(mp
->m_sb_bp
);
3661 * Somebody could've already done the hard work for us.
3662 * No need to get locks for this.
3664 if (logerror
&& log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3665 ASSERT(XLOG_FORCED_SHUTDOWN(log
));
3671 * Flush the in memory commit item list before marking the log as
3672 * being shut down. We need to do it in this order to ensure all the
3673 * completed transactions are flushed to disk with the xfs_log_force()
3677 xlog_cil_force(log
);
3680 * mark the filesystem and the as in a shutdown state and wake
3681 * everybody up to tell them the bad news.
3683 spin_lock(&log
->l_icloglock
);
3684 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3686 XFS_BUF_DONE(mp
->m_sb_bp
);
3689 * This flag is sort of redundant because of the mount flag, but
3690 * it's good to maintain the separation between the log and the rest
3693 log
->l_flags
|= XLOG_IO_ERROR
;
3696 * If we hit a log error, we want to mark all the iclogs IOERROR
3697 * while we're still holding the loglock.
3700 retval
= xlog_state_ioerror(log
);
3701 spin_unlock(&log
->l_icloglock
);
3704 * We don't want anybody waiting for log reservations after this. That
3705 * means we have to wake up everybody queued up on reserveq as well as
3706 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3707 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3708 * action is protected by the grant locks.
3710 xlog_grant_head_wake_all(&log
->l_reserve_head
);
3711 xlog_grant_head_wake_all(&log
->l_write_head
);
3713 if (!(log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3716 * Force the incore logs to disk before shutting the
3717 * log down completely.
3719 _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
3721 spin_lock(&log
->l_icloglock
);
3722 retval
= xlog_state_ioerror(log
);
3723 spin_unlock(&log
->l_icloglock
);
3726 * Wake up everybody waiting on xfs_log_force.
3727 * Callback all log item committed functions as if the
3728 * log writes were completed.
3730 xlog_state_do_callback(log
, XFS_LI_ABORTED
, NULL
);
3732 #ifdef XFSERRORDEBUG
3734 xlog_in_core_t
*iclog
;
3736 spin_lock(&log
->l_icloglock
);
3737 iclog
= log
->l_iclog
;
3739 ASSERT(iclog
->ic_callback
== 0);
3740 iclog
= iclog
->ic_next
;
3741 } while (iclog
!= log
->l_iclog
);
3742 spin_unlock(&log
->l_icloglock
);
3745 /* return non-zero if log IOERROR transition had already happened */
3753 xlog_in_core_t
*iclog
;
3755 iclog
= log
->l_iclog
;
3757 /* endianness does not matter here, zero is zero in
3760 if (iclog
->ic_header
.h_num_logops
)
3762 iclog
= iclog
->ic_next
;
3763 } while (iclog
!= log
->l_iclog
);