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 * Empty the log for unmount/freeze.
860 * To do this, we first need to shut down the background log work so it is not
861 * trying to cover the log as we clean up. We then need to unpin all objects in
862 * the log so we can then flush them out. Once they have completed their IO and
863 * run the callbacks removing themselves from the AIL, we can write the unmount
868 struct xfs_mount
*mp
)
870 cancel_delayed_work_sync(&mp
->m_log
->l_work
);
871 xfs_log_force(mp
, XFS_LOG_SYNC
);
874 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
875 * will push it, xfs_wait_buftarg() will not wait for it. Further,
876 * xfs_buf_iowait() cannot be used because it was pushed with the
877 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
878 * the IO to complete.
880 xfs_ail_push_all_sync(mp
->m_ail
);
881 xfs_wait_buftarg(mp
->m_ddev_targp
);
882 xfs_buf_lock(mp
->m_sb_bp
);
883 xfs_buf_unlock(mp
->m_sb_bp
);
885 xfs_log_unmount_write(mp
);
889 * Shut down and release the AIL and Log.
891 * During unmount, we need to ensure we flush all the dirty metadata objects
892 * from the AIL so that the log is empty before we write the unmount record to
893 * the log. Once this is done, we can tear down the AIL and the log.
897 struct xfs_mount
*mp
)
901 xfs_trans_ail_destroy(mp
);
902 xlog_dealloc_log(mp
->m_log
);
907 struct xfs_mount
*mp
,
908 struct xfs_log_item
*item
,
910 const struct xfs_item_ops
*ops
)
912 item
->li_mountp
= mp
;
913 item
->li_ailp
= mp
->m_ail
;
914 item
->li_type
= type
;
918 INIT_LIST_HEAD(&item
->li_ail
);
919 INIT_LIST_HEAD(&item
->li_cil
);
923 * Wake up processes waiting for log space after we have moved the log tail.
927 struct xfs_mount
*mp
)
929 struct xlog
*log
= mp
->m_log
;
932 if (XLOG_FORCED_SHUTDOWN(log
))
935 if (!list_empty_careful(&log
->l_write_head
.waiters
)) {
936 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
938 spin_lock(&log
->l_write_head
.lock
);
939 free_bytes
= xlog_space_left(log
, &log
->l_write_head
.grant
);
940 xlog_grant_head_wake(log
, &log
->l_write_head
, &free_bytes
);
941 spin_unlock(&log
->l_write_head
.lock
);
944 if (!list_empty_careful(&log
->l_reserve_head
.waiters
)) {
945 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
947 spin_lock(&log
->l_reserve_head
.lock
);
948 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
949 xlog_grant_head_wake(log
, &log
->l_reserve_head
, &free_bytes
);
950 spin_unlock(&log
->l_reserve_head
.lock
);
955 * Determine if we have a transaction that has gone to disk
956 * that needs to be covered. To begin the transition to the idle state
957 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
958 * If we are then in a state where covering is needed, the caller is informed
959 * that dummy transactions are required to move the log into the idle state.
961 * Because this is called as part of the sync process, we should also indicate
962 * that dummy transactions should be issued in anything but the covered or
963 * idle states. This ensures that the log tail is accurately reflected in
964 * the log at the end of the sync, hence if a crash occurrs avoids replay
965 * of transactions where the metadata is already on disk.
968 xfs_log_need_covered(xfs_mount_t
*mp
)
971 struct xlog
*log
= mp
->m_log
;
973 if (!xfs_fs_writable(mp
))
976 spin_lock(&log
->l_icloglock
);
977 switch (log
->l_covered_state
) {
978 case XLOG_STATE_COVER_DONE
:
979 case XLOG_STATE_COVER_DONE2
:
980 case XLOG_STATE_COVER_IDLE
:
982 case XLOG_STATE_COVER_NEED
:
983 case XLOG_STATE_COVER_NEED2
:
984 if (!xfs_ail_min_lsn(log
->l_ailp
) &&
985 xlog_iclogs_empty(log
)) {
986 if (log
->l_covered_state
== XLOG_STATE_COVER_NEED
)
987 log
->l_covered_state
= XLOG_STATE_COVER_DONE
;
989 log
->l_covered_state
= XLOG_STATE_COVER_DONE2
;
996 spin_unlock(&log
->l_icloglock
);
1001 * We may be holding the log iclog lock upon entering this routine.
1004 xlog_assign_tail_lsn_locked(
1005 struct xfs_mount
*mp
)
1007 struct xlog
*log
= mp
->m_log
;
1008 struct xfs_log_item
*lip
;
1011 assert_spin_locked(&mp
->m_ail
->xa_lock
);
1014 * To make sure we always have a valid LSN for the log tail we keep
1015 * track of the last LSN which was committed in log->l_last_sync_lsn,
1016 * and use that when the AIL was empty.
1018 lip
= xfs_ail_min(mp
->m_ail
);
1020 tail_lsn
= lip
->li_lsn
;
1022 tail_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1023 atomic64_set(&log
->l_tail_lsn
, tail_lsn
);
1028 xlog_assign_tail_lsn(
1029 struct xfs_mount
*mp
)
1033 spin_lock(&mp
->m_ail
->xa_lock
);
1034 tail_lsn
= xlog_assign_tail_lsn_locked(mp
);
1035 spin_unlock(&mp
->m_ail
->xa_lock
);
1041 * Return the space in the log between the tail and the head. The head
1042 * is passed in the cycle/bytes formal parms. In the special case where
1043 * the reserve head has wrapped passed the tail, this calculation is no
1044 * longer valid. In this case, just return 0 which means there is no space
1045 * in the log. This works for all places where this function is called
1046 * with the reserve head. Of course, if the write head were to ever
1047 * wrap the tail, we should blow up. Rather than catch this case here,
1048 * we depend on other ASSERTions in other parts of the code. XXXmiken
1050 * This code also handles the case where the reservation head is behind
1051 * the tail. The details of this case are described below, but the end
1052 * result is that we return the size of the log as the amount of space left.
1065 xlog_crack_grant_head(head
, &head_cycle
, &head_bytes
);
1066 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_bytes
);
1067 tail_bytes
= BBTOB(tail_bytes
);
1068 if (tail_cycle
== head_cycle
&& head_bytes
>= tail_bytes
)
1069 free_bytes
= log
->l_logsize
- (head_bytes
- tail_bytes
);
1070 else if (tail_cycle
+ 1 < head_cycle
)
1072 else if (tail_cycle
< head_cycle
) {
1073 ASSERT(tail_cycle
== (head_cycle
- 1));
1074 free_bytes
= tail_bytes
- head_bytes
;
1077 * The reservation head is behind the tail.
1078 * In this case we just want to return the size of the
1079 * log as the amount of space left.
1081 xfs_alert(log
->l_mp
,
1082 "xlog_space_left: head behind tail\n"
1083 " tail_cycle = %d, tail_bytes = %d\n"
1084 " GH cycle = %d, GH bytes = %d",
1085 tail_cycle
, tail_bytes
, head_cycle
, head_bytes
);
1087 free_bytes
= log
->l_logsize
;
1094 * Log function which is called when an io completes.
1096 * The log manager needs its own routine, in order to control what
1097 * happens with the buffer after the write completes.
1100 xlog_iodone(xfs_buf_t
*bp
)
1102 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1103 struct xlog
*l
= iclog
->ic_log
;
1107 * Race to shutdown the filesystem if we see an error.
1109 if (XFS_TEST_ERROR((xfs_buf_geterror(bp
)), l
->l_mp
,
1110 XFS_ERRTAG_IODONE_IOERR
, XFS_RANDOM_IODONE_IOERR
)) {
1111 xfs_buf_ioerror_alert(bp
, __func__
);
1113 xfs_force_shutdown(l
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1115 * This flag will be propagated to the trans-committed
1116 * callback routines to let them know that the log-commit
1119 aborted
= XFS_LI_ABORTED
;
1120 } else if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1121 aborted
= XFS_LI_ABORTED
;
1124 /* log I/O is always issued ASYNC */
1125 ASSERT(XFS_BUF_ISASYNC(bp
));
1126 xlog_state_done_syncing(iclog
, aborted
);
1128 * do not reference the buffer (bp) here as we could race
1129 * with it being freed after writing the unmount record to the
1136 * Return size of each in-core log record buffer.
1138 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1140 * If the filesystem blocksize is too large, we may need to choose a
1141 * larger size since the directory code currently logs entire blocks.
1145 xlog_get_iclog_buffer_size(
1146 struct xfs_mount
*mp
,
1152 if (mp
->m_logbufs
<= 0)
1153 log
->l_iclog_bufs
= XLOG_MAX_ICLOGS
;
1155 log
->l_iclog_bufs
= mp
->m_logbufs
;
1158 * Buffer size passed in from mount system call.
1160 if (mp
->m_logbsize
> 0) {
1161 size
= log
->l_iclog_size
= mp
->m_logbsize
;
1162 log
->l_iclog_size_log
= 0;
1164 log
->l_iclog_size_log
++;
1168 if (xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1169 /* # headers = size / 32k
1170 * one header holds cycles from 32k of data
1173 xhdrs
= mp
->m_logbsize
/ XLOG_HEADER_CYCLE_SIZE
;
1174 if (mp
->m_logbsize
% XLOG_HEADER_CYCLE_SIZE
)
1176 log
->l_iclog_hsize
= xhdrs
<< BBSHIFT
;
1177 log
->l_iclog_heads
= xhdrs
;
1179 ASSERT(mp
->m_logbsize
<= XLOG_BIG_RECORD_BSIZE
);
1180 log
->l_iclog_hsize
= BBSIZE
;
1181 log
->l_iclog_heads
= 1;
1186 /* All machines use 32kB buffers by default. */
1187 log
->l_iclog_size
= XLOG_BIG_RECORD_BSIZE
;
1188 log
->l_iclog_size_log
= XLOG_BIG_RECORD_BSHIFT
;
1190 /* the default log size is 16k or 32k which is one header sector */
1191 log
->l_iclog_hsize
= BBSIZE
;
1192 log
->l_iclog_heads
= 1;
1195 /* are we being asked to make the sizes selected above visible? */
1196 if (mp
->m_logbufs
== 0)
1197 mp
->m_logbufs
= log
->l_iclog_bufs
;
1198 if (mp
->m_logbsize
== 0)
1199 mp
->m_logbsize
= log
->l_iclog_size
;
1200 } /* xlog_get_iclog_buffer_size */
1205 struct xfs_mount
*mp
)
1207 queue_delayed_work(mp
->m_log_workqueue
, &mp
->m_log
->l_work
,
1208 msecs_to_jiffies(xfs_syncd_centisecs
* 10));
1212 * Every sync period we need to unpin all items in the AIL and push them to
1213 * disk. If there is nothing dirty, then we might need to cover the log to
1214 * indicate that the filesystem is idle.
1218 struct work_struct
*work
)
1220 struct xlog
*log
= container_of(to_delayed_work(work
),
1221 struct xlog
, l_work
);
1222 struct xfs_mount
*mp
= log
->l_mp
;
1224 /* dgc: errors ignored - not fatal and nowhere to report them */
1225 if (xfs_log_need_covered(mp
))
1226 xfs_fs_log_dummy(mp
);
1228 xfs_log_force(mp
, 0);
1230 /* start pushing all the metadata that is currently dirty */
1231 xfs_ail_push_all(mp
->m_ail
);
1233 /* queue us up again */
1234 xfs_log_work_queue(mp
);
1238 * This routine initializes some of the log structure for a given mount point.
1239 * Its primary purpose is to fill in enough, so recovery can occur. However,
1240 * some other stuff may be filled in too.
1242 STATIC
struct xlog
*
1244 struct xfs_mount
*mp
,
1245 struct xfs_buftarg
*log_target
,
1246 xfs_daddr_t blk_offset
,
1250 xlog_rec_header_t
*head
;
1251 xlog_in_core_t
**iclogp
;
1252 xlog_in_core_t
*iclog
, *prev_iclog
=NULL
;
1258 log
= kmem_zalloc(sizeof(struct xlog
), KM_MAYFAIL
);
1260 xfs_warn(mp
, "Log allocation failed: No memory!");
1265 log
->l_targ
= log_target
;
1266 log
->l_logsize
= BBTOB(num_bblks
);
1267 log
->l_logBBstart
= blk_offset
;
1268 log
->l_logBBsize
= num_bblks
;
1269 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
1270 log
->l_flags
|= XLOG_ACTIVE_RECOVERY
;
1271 INIT_DELAYED_WORK(&log
->l_work
, xfs_log_worker
);
1273 log
->l_prev_block
= -1;
1274 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1275 xlog_assign_atomic_lsn(&log
->l_tail_lsn
, 1, 0);
1276 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
, 1, 0);
1277 log
->l_curr_cycle
= 1; /* 0 is bad since this is initial value */
1279 xlog_grant_head_init(&log
->l_reserve_head
);
1280 xlog_grant_head_init(&log
->l_write_head
);
1282 error
= EFSCORRUPTED
;
1283 if (xfs_sb_version_hassector(&mp
->m_sb
)) {
1284 log2_size
= mp
->m_sb
.sb_logsectlog
;
1285 if (log2_size
< BBSHIFT
) {
1286 xfs_warn(mp
, "Log sector size too small (0x%x < 0x%x)",
1287 log2_size
, BBSHIFT
);
1291 log2_size
-= BBSHIFT
;
1292 if (log2_size
> mp
->m_sectbb_log
) {
1293 xfs_warn(mp
, "Log sector size too large (0x%x > 0x%x)",
1294 log2_size
, mp
->m_sectbb_log
);
1298 /* for larger sector sizes, must have v2 or external log */
1299 if (log2_size
&& log
->l_logBBstart
> 0 &&
1300 !xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1302 "log sector size (0x%x) invalid for configuration.",
1307 log
->l_sectBBsize
= 1 << log2_size
;
1309 xlog_get_iclog_buffer_size(mp
, log
);
1312 bp
= xfs_buf_alloc(mp
->m_logdev_targp
, 0, BTOBB(log
->l_iclog_size
), 0);
1315 bp
->b_iodone
= xlog_iodone
;
1316 ASSERT(xfs_buf_islocked(bp
));
1319 spin_lock_init(&log
->l_icloglock
);
1320 init_waitqueue_head(&log
->l_flush_wait
);
1322 iclogp
= &log
->l_iclog
;
1324 * The amount of memory to allocate for the iclog structure is
1325 * rather funky due to the way the structure is defined. It is
1326 * done this way so that we can use different sizes for machines
1327 * with different amounts of memory. See the definition of
1328 * xlog_in_core_t in xfs_log_priv.h for details.
1330 ASSERT(log
->l_iclog_size
>= 4096);
1331 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
1332 *iclogp
= kmem_zalloc(sizeof(xlog_in_core_t
), KM_MAYFAIL
);
1334 goto out_free_iclog
;
1337 iclog
->ic_prev
= prev_iclog
;
1340 bp
= xfs_buf_get_uncached(mp
->m_logdev_targp
,
1341 BTOBB(log
->l_iclog_size
), 0);
1343 goto out_free_iclog
;
1345 bp
->b_iodone
= xlog_iodone
;
1347 iclog
->ic_data
= bp
->b_addr
;
1349 log
->l_iclog_bak
[i
] = (xfs_caddr_t
)&(iclog
->ic_header
);
1351 head
= &iclog
->ic_header
;
1352 memset(head
, 0, sizeof(xlog_rec_header_t
));
1353 head
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1354 head
->h_version
= cpu_to_be32(
1355 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1356 head
->h_size
= cpu_to_be32(log
->l_iclog_size
);
1358 head
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1359 memcpy(&head
->h_fs_uuid
, &mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1361 iclog
->ic_size
= BBTOB(bp
->b_length
) - log
->l_iclog_hsize
;
1362 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1363 iclog
->ic_log
= log
;
1364 atomic_set(&iclog
->ic_refcnt
, 0);
1365 spin_lock_init(&iclog
->ic_callback_lock
);
1366 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
1367 iclog
->ic_datap
= (char *)iclog
->ic_data
+ log
->l_iclog_hsize
;
1369 ASSERT(xfs_buf_islocked(iclog
->ic_bp
));
1370 init_waitqueue_head(&iclog
->ic_force_wait
);
1371 init_waitqueue_head(&iclog
->ic_write_wait
);
1373 iclogp
= &iclog
->ic_next
;
1375 *iclogp
= log
->l_iclog
; /* complete ring */
1376 log
->l_iclog
->ic_prev
= prev_iclog
; /* re-write 1st prev ptr */
1378 error
= xlog_cil_init(log
);
1380 goto out_free_iclog
;
1384 for (iclog
= log
->l_iclog
; iclog
; iclog
= prev_iclog
) {
1385 prev_iclog
= iclog
->ic_next
;
1387 xfs_buf_free(iclog
->ic_bp
);
1390 spinlock_destroy(&log
->l_icloglock
);
1391 xfs_buf_free(log
->l_xbuf
);
1395 return ERR_PTR(-error
);
1396 } /* xlog_alloc_log */
1400 * Write out the commit record of a transaction associated with the given
1401 * ticket. Return the lsn of the commit record.
1406 struct xlog_ticket
*ticket
,
1407 struct xlog_in_core
**iclog
,
1408 xfs_lsn_t
*commitlsnp
)
1410 struct xfs_mount
*mp
= log
->l_mp
;
1412 struct xfs_log_iovec reg
= {
1415 .i_type
= XLOG_REG_TYPE_COMMIT
,
1417 struct xfs_log_vec vec
= {
1422 ASSERT_ALWAYS(iclog
);
1423 error
= xlog_write(log
, &vec
, ticket
, commitlsnp
, iclog
,
1426 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
1431 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1432 * log space. This code pushes on the lsn which would supposedly free up
1433 * the 25% which we want to leave free. We may need to adopt a policy which
1434 * pushes on an lsn which is further along in the log once we reach the high
1435 * water mark. In this manner, we would be creating a low water mark.
1438 xlog_grant_push_ail(
1442 xfs_lsn_t threshold_lsn
= 0;
1443 xfs_lsn_t last_sync_lsn
;
1446 int threshold_block
;
1447 int threshold_cycle
;
1450 ASSERT(BTOBB(need_bytes
) < log
->l_logBBsize
);
1452 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1453 free_blocks
= BTOBBT(free_bytes
);
1456 * Set the threshold for the minimum number of free blocks in the
1457 * log to the maximum of what the caller needs, one quarter of the
1458 * log, and 256 blocks.
1460 free_threshold
= BTOBB(need_bytes
);
1461 free_threshold
= MAX(free_threshold
, (log
->l_logBBsize
>> 2));
1462 free_threshold
= MAX(free_threshold
, 256);
1463 if (free_blocks
>= free_threshold
)
1466 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &threshold_cycle
,
1468 threshold_block
+= free_threshold
;
1469 if (threshold_block
>= log
->l_logBBsize
) {
1470 threshold_block
-= log
->l_logBBsize
;
1471 threshold_cycle
+= 1;
1473 threshold_lsn
= xlog_assign_lsn(threshold_cycle
,
1476 * Don't pass in an lsn greater than the lsn of the last
1477 * log record known to be on disk. Use a snapshot of the last sync lsn
1478 * so that it doesn't change between the compare and the set.
1480 last_sync_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1481 if (XFS_LSN_CMP(threshold_lsn
, last_sync_lsn
) > 0)
1482 threshold_lsn
= last_sync_lsn
;
1485 * Get the transaction layer to kick the dirty buffers out to
1486 * disk asynchronously. No point in trying to do this if
1487 * the filesystem is shutting down.
1489 if (!XLOG_FORCED_SHUTDOWN(log
))
1490 xfs_ail_push(log
->l_ailp
, threshold_lsn
);
1494 * The bdstrat callback function for log bufs. This gives us a central
1495 * place to trap bufs in case we get hit by a log I/O error and need to
1496 * shutdown. Actually, in practice, even when we didn't get a log error,
1497 * we transition the iclogs to IOERROR state *after* flushing all existing
1498 * iclogs to disk. This is because we don't want anymore new transactions to be
1499 * started or completed afterwards.
1505 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1507 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1508 xfs_buf_ioerror(bp
, EIO
);
1510 xfs_buf_ioend(bp
, 0);
1512 * It would seem logical to return EIO here, but we rely on
1513 * the log state machine to propagate I/O errors instead of
1519 xfs_buf_iorequest(bp
);
1524 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1525 * fashion. Previously, we should have moved the current iclog
1526 * ptr in the log to point to the next available iclog. This allows further
1527 * write to continue while this code syncs out an iclog ready to go.
1528 * Before an in-core log can be written out, the data section must be scanned
1529 * to save away the 1st word of each BBSIZE block into the header. We replace
1530 * it with the current cycle count. Each BBSIZE block is tagged with the
1531 * cycle count because there in an implicit assumption that drives will
1532 * guarantee that entire 512 byte blocks get written at once. In other words,
1533 * we can't have part of a 512 byte block written and part not written. By
1534 * tagging each block, we will know which blocks are valid when recovering
1535 * after an unclean shutdown.
1537 * This routine is single threaded on the iclog. No other thread can be in
1538 * this routine with the same iclog. Changing contents of iclog can there-
1539 * fore be done without grabbing the state machine lock. Updating the global
1540 * log will require grabbing the lock though.
1542 * The entire log manager uses a logical block numbering scheme. Only
1543 * log_sync (and then only bwrite()) know about the fact that the log may
1544 * not start with block zero on a given device. The log block start offset
1545 * is added immediately before calling bwrite().
1551 struct xlog_in_core
*iclog
)
1553 xfs_caddr_t dptr
; /* pointer to byte sized element */
1556 uint count
; /* byte count of bwrite */
1557 uint count_init
; /* initial count before roundup */
1558 int roundoff
; /* roundoff to BB or stripe */
1559 int split
= 0; /* split write into two regions */
1561 int v2
= xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
);
1563 XFS_STATS_INC(xs_log_writes
);
1564 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
1566 /* Add for LR header */
1567 count_init
= log
->l_iclog_hsize
+ iclog
->ic_offset
;
1569 /* Round out the log write size */
1570 if (v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1) {
1571 /* we have a v2 stripe unit to use */
1572 count
= XLOG_LSUNITTOB(log
, XLOG_BTOLSUNIT(log
, count_init
));
1574 count
= BBTOB(BTOBB(count_init
));
1576 roundoff
= count
- count_init
;
1577 ASSERT(roundoff
>= 0);
1578 ASSERT((v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1 &&
1579 roundoff
< log
->l_mp
->m_sb
.sb_logsunit
)
1581 (log
->l_mp
->m_sb
.sb_logsunit
<= 1 &&
1582 roundoff
< BBTOB(1)));
1584 /* move grant heads by roundoff in sync */
1585 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, roundoff
);
1586 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, roundoff
);
1588 /* put cycle number in every block */
1589 xlog_pack_data(log
, iclog
, roundoff
);
1591 /* real byte length */
1593 iclog
->ic_header
.h_len
=
1594 cpu_to_be32(iclog
->ic_offset
+ roundoff
);
1596 iclog
->ic_header
.h_len
=
1597 cpu_to_be32(iclog
->ic_offset
);
1601 XFS_BUF_SET_ADDR(bp
, BLOCK_LSN(be64_to_cpu(iclog
->ic_header
.h_lsn
)));
1603 XFS_STATS_ADD(xs_log_blocks
, BTOBB(count
));
1605 /* Do we need to split this write into 2 parts? */
1606 if (XFS_BUF_ADDR(bp
) + BTOBB(count
) > log
->l_logBBsize
) {
1607 split
= count
- (BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
)));
1608 count
= BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
));
1609 iclog
->ic_bwritecnt
= 2; /* split into 2 writes */
1611 iclog
->ic_bwritecnt
= 1;
1613 bp
->b_io_length
= BTOBB(count
);
1614 bp
->b_fspriv
= iclog
;
1615 XFS_BUF_ZEROFLAGS(bp
);
1617 bp
->b_flags
|= XBF_SYNCIO
;
1619 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
) {
1620 bp
->b_flags
|= XBF_FUA
;
1623 * Flush the data device before flushing the log to make
1624 * sure all meta data written back from the AIL actually made
1625 * it to disk before stamping the new log tail LSN into the
1626 * log buffer. For an external log we need to issue the
1627 * flush explicitly, and unfortunately synchronously here;
1628 * for an internal log we can simply use the block layer
1629 * state machine for preflushes.
1631 if (log
->l_mp
->m_logdev_targp
!= log
->l_mp
->m_ddev_targp
)
1632 xfs_blkdev_issue_flush(log
->l_mp
->m_ddev_targp
);
1634 bp
->b_flags
|= XBF_FLUSH
;
1637 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1638 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1640 xlog_verify_iclog(log
, iclog
, count
, B_TRUE
);
1642 /* account for log which doesn't start at block #0 */
1643 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1645 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1650 error
= xlog_bdstrat(bp
);
1652 xfs_buf_ioerror_alert(bp
, "xlog_sync");
1656 bp
= iclog
->ic_log
->l_xbuf
;
1657 XFS_BUF_SET_ADDR(bp
, 0); /* logical 0 */
1658 xfs_buf_associate_memory(bp
,
1659 (char *)&iclog
->ic_header
+ count
, split
);
1660 bp
->b_fspriv
= iclog
;
1661 XFS_BUF_ZEROFLAGS(bp
);
1663 bp
->b_flags
|= XBF_SYNCIO
;
1664 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
)
1665 bp
->b_flags
|= XBF_FUA
;
1668 * Bump the cycle numbers at the start of each block
1669 * since this part of the buffer is at the start of
1670 * a new cycle. Watch out for the header magic number
1673 for (i
= 0; i
< split
; i
+= BBSIZE
) {
1674 be32_add_cpu((__be32
*)dptr
, 1);
1675 if (be32_to_cpu(*(__be32
*)dptr
) == XLOG_HEADER_MAGIC_NUM
)
1676 be32_add_cpu((__be32
*)dptr
, 1);
1680 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1681 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1683 /* account for internal log which doesn't start at block #0 */
1684 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1686 error
= xlog_bdstrat(bp
);
1688 xfs_buf_ioerror_alert(bp
, "xlog_sync (split)");
1697 * Deallocate a log structure
1703 xlog_in_core_t
*iclog
, *next_iclog
;
1706 xlog_cil_destroy(log
);
1709 * always need to ensure that the extra buffer does not point to memory
1710 * owned by another log buffer before we free it.
1712 xfs_buf_set_empty(log
->l_xbuf
, BTOBB(log
->l_iclog_size
));
1713 xfs_buf_free(log
->l_xbuf
);
1715 iclog
= log
->l_iclog
;
1716 for (i
=0; i
<log
->l_iclog_bufs
; i
++) {
1717 xfs_buf_free(iclog
->ic_bp
);
1718 next_iclog
= iclog
->ic_next
;
1722 spinlock_destroy(&log
->l_icloglock
);
1724 log
->l_mp
->m_log
= NULL
;
1726 } /* xlog_dealloc_log */
1729 * Update counters atomically now that memcpy is done.
1733 xlog_state_finish_copy(
1735 struct xlog_in_core
*iclog
,
1739 spin_lock(&log
->l_icloglock
);
1741 be32_add_cpu(&iclog
->ic_header
.h_num_logops
, record_cnt
);
1742 iclog
->ic_offset
+= copy_bytes
;
1744 spin_unlock(&log
->l_icloglock
);
1745 } /* xlog_state_finish_copy */
1751 * print out info relating to regions written which consume
1756 struct xfs_mount
*mp
,
1757 struct xlog_ticket
*ticket
)
1760 uint ophdr_spc
= ticket
->t_res_num_ophdrs
* (uint
)sizeof(xlog_op_header_t
);
1762 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1763 static char *res_type_str
[XLOG_REG_TYPE_MAX
] = {
1784 static char *trans_type_str
[XFS_TRANS_TYPE_MAX
] = {
1828 "xlog_write: reservation summary:\n"
1829 " trans type = %s (%u)\n"
1830 " unit res = %d bytes\n"
1831 " current res = %d bytes\n"
1832 " total reg = %u bytes (o/flow = %u bytes)\n"
1833 " ophdrs = %u (ophdr space = %u bytes)\n"
1834 " ophdr + reg = %u bytes\n"
1835 " num regions = %u\n",
1836 ((ticket
->t_trans_type
<= 0 ||
1837 ticket
->t_trans_type
> XFS_TRANS_TYPE_MAX
) ?
1838 "bad-trans-type" : trans_type_str
[ticket
->t_trans_type
-1]),
1839 ticket
->t_trans_type
,
1842 ticket
->t_res_arr_sum
, ticket
->t_res_o_flow
,
1843 ticket
->t_res_num_ophdrs
, ophdr_spc
,
1844 ticket
->t_res_arr_sum
+
1845 ticket
->t_res_o_flow
+ ophdr_spc
,
1848 for (i
= 0; i
< ticket
->t_res_num
; i
++) {
1849 uint r_type
= ticket
->t_res_arr
[i
].r_type
;
1850 xfs_warn(mp
, "region[%u]: %s - %u bytes\n", i
,
1851 ((r_type
<= 0 || r_type
> XLOG_REG_TYPE_MAX
) ?
1852 "bad-rtype" : res_type_str
[r_type
-1]),
1853 ticket
->t_res_arr
[i
].r_len
);
1856 xfs_alert_tag(mp
, XFS_PTAG_LOGRES
,
1857 "xlog_write: reservation ran out. Need to up reservation");
1858 xfs_force_shutdown(mp
, SHUTDOWN_CORRUPT_INCORE
);
1862 * Calculate the potential space needed by the log vector. Each region gets
1863 * its own xlog_op_header_t and may need to be double word aligned.
1866 xlog_write_calc_vec_length(
1867 struct xlog_ticket
*ticket
,
1868 struct xfs_log_vec
*log_vector
)
1870 struct xfs_log_vec
*lv
;
1875 /* acct for start rec of xact */
1876 if (ticket
->t_flags
& XLOG_TIC_INITED
)
1879 for (lv
= log_vector
; lv
; lv
= lv
->lv_next
) {
1880 headers
+= lv
->lv_niovecs
;
1882 for (i
= 0; i
< lv
->lv_niovecs
; i
++) {
1883 struct xfs_log_iovec
*vecp
= &lv
->lv_iovecp
[i
];
1886 xlog_tic_add_region(ticket
, vecp
->i_len
, vecp
->i_type
);
1890 ticket
->t_res_num_ophdrs
+= headers
;
1891 len
+= headers
* sizeof(struct xlog_op_header
);
1897 * If first write for transaction, insert start record We can't be trying to
1898 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1901 xlog_write_start_rec(
1902 struct xlog_op_header
*ophdr
,
1903 struct xlog_ticket
*ticket
)
1905 if (!(ticket
->t_flags
& XLOG_TIC_INITED
))
1908 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1909 ophdr
->oh_clientid
= ticket
->t_clientid
;
1911 ophdr
->oh_flags
= XLOG_START_TRANS
;
1914 ticket
->t_flags
&= ~XLOG_TIC_INITED
;
1916 return sizeof(struct xlog_op_header
);
1919 static xlog_op_header_t
*
1920 xlog_write_setup_ophdr(
1922 struct xlog_op_header
*ophdr
,
1923 struct xlog_ticket
*ticket
,
1926 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1927 ophdr
->oh_clientid
= ticket
->t_clientid
;
1930 /* are we copying a commit or unmount record? */
1931 ophdr
->oh_flags
= flags
;
1934 * We've seen logs corrupted with bad transaction client ids. This
1935 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1936 * and shut down the filesystem.
1938 switch (ophdr
->oh_clientid
) {
1939 case XFS_TRANSACTION
:
1945 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1946 ophdr
->oh_clientid
, ticket
);
1954 * Set up the parameters of the region copy into the log. This has
1955 * to handle region write split across multiple log buffers - this
1956 * state is kept external to this function so that this code can
1957 * can be written in an obvious, self documenting manner.
1960 xlog_write_setup_copy(
1961 struct xlog_ticket
*ticket
,
1962 struct xlog_op_header
*ophdr
,
1963 int space_available
,
1967 int *last_was_partial_copy
,
1968 int *bytes_consumed
)
1972 still_to_copy
= space_required
- *bytes_consumed
;
1973 *copy_off
= *bytes_consumed
;
1975 if (still_to_copy
<= space_available
) {
1976 /* write of region completes here */
1977 *copy_len
= still_to_copy
;
1978 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
1979 if (*last_was_partial_copy
)
1980 ophdr
->oh_flags
|= (XLOG_END_TRANS
|XLOG_WAS_CONT_TRANS
);
1981 *last_was_partial_copy
= 0;
1982 *bytes_consumed
= 0;
1986 /* partial write of region, needs extra log op header reservation */
1987 *copy_len
= space_available
;
1988 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
1989 ophdr
->oh_flags
|= XLOG_CONTINUE_TRANS
;
1990 if (*last_was_partial_copy
)
1991 ophdr
->oh_flags
|= XLOG_WAS_CONT_TRANS
;
1992 *bytes_consumed
+= *copy_len
;
1993 (*last_was_partial_copy
)++;
1995 /* account for new log op header */
1996 ticket
->t_curr_res
-= sizeof(struct xlog_op_header
);
1997 ticket
->t_res_num_ophdrs
++;
1999 return sizeof(struct xlog_op_header
);
2003 xlog_write_copy_finish(
2005 struct xlog_in_core
*iclog
,
2010 int *partial_copy_len
,
2012 struct xlog_in_core
**commit_iclog
)
2014 if (*partial_copy
) {
2016 * This iclog has already been marked WANT_SYNC by
2017 * xlog_state_get_iclog_space.
2019 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2022 return xlog_state_release_iclog(log
, iclog
);
2026 *partial_copy_len
= 0;
2028 if (iclog
->ic_size
- log_offset
<= sizeof(xlog_op_header_t
)) {
2029 /* no more space in this iclog - push it. */
2030 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2034 spin_lock(&log
->l_icloglock
);
2035 xlog_state_want_sync(log
, iclog
);
2036 spin_unlock(&log
->l_icloglock
);
2039 return xlog_state_release_iclog(log
, iclog
);
2040 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2041 *commit_iclog
= iclog
;
2048 * Write some region out to in-core log
2050 * This will be called when writing externally provided regions or when
2051 * writing out a commit record for a given transaction.
2053 * General algorithm:
2054 * 1. Find total length of this write. This may include adding to the
2055 * lengths passed in.
2056 * 2. Check whether we violate the tickets reservation.
2057 * 3. While writing to this iclog
2058 * A. Reserve as much space in this iclog as can get
2059 * B. If this is first write, save away start lsn
2060 * C. While writing this region:
2061 * 1. If first write of transaction, write start record
2062 * 2. Write log operation header (header per region)
2063 * 3. Find out if we can fit entire region into this iclog
2064 * 4. Potentially, verify destination memcpy ptr
2065 * 5. Memcpy (partial) region
2066 * 6. If partial copy, release iclog; otherwise, continue
2067 * copying more regions into current iclog
2068 * 4. Mark want sync bit (in simulation mode)
2069 * 5. Release iclog for potential flush to on-disk log.
2072 * 1. Panic if reservation is overrun. This should never happen since
2073 * reservation amounts are generated internal to the filesystem.
2075 * 1. Tickets are single threaded data structures.
2076 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2077 * syncing routine. When a single log_write region needs to span
2078 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2079 * on all log operation writes which don't contain the end of the
2080 * region. The XLOG_END_TRANS bit is used for the in-core log
2081 * operation which contains the end of the continued log_write region.
2082 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2083 * we don't really know exactly how much space will be used. As a result,
2084 * we don't update ic_offset until the end when we know exactly how many
2085 * bytes have been written out.
2090 struct xfs_log_vec
*log_vector
,
2091 struct xlog_ticket
*ticket
,
2092 xfs_lsn_t
*start_lsn
,
2093 struct xlog_in_core
**commit_iclog
,
2096 struct xlog_in_core
*iclog
= NULL
;
2097 struct xfs_log_iovec
*vecp
;
2098 struct xfs_log_vec
*lv
;
2101 int partial_copy
= 0;
2102 int partial_copy_len
= 0;
2110 len
= xlog_write_calc_vec_length(ticket
, log_vector
);
2113 * Region headers and bytes are already accounted for.
2114 * We only need to take into account start records and
2115 * split regions in this function.
2117 if (ticket
->t_flags
& XLOG_TIC_INITED
)
2118 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2121 * Commit record headers need to be accounted for. These
2122 * come in as separate writes so are easy to detect.
2124 if (flags
& (XLOG_COMMIT_TRANS
| XLOG_UNMOUNT_TRANS
))
2125 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2127 if (ticket
->t_curr_res
< 0)
2128 xlog_print_tic_res(log
->l_mp
, ticket
);
2132 vecp
= lv
->lv_iovecp
;
2133 while (lv
&& index
< lv
->lv_niovecs
) {
2137 error
= xlog_state_get_iclog_space(log
, len
, &iclog
, ticket
,
2138 &contwr
, &log_offset
);
2142 ASSERT(log_offset
<= iclog
->ic_size
- 1);
2143 ptr
= iclog
->ic_datap
+ log_offset
;
2145 /* start_lsn is the first lsn written to. That's all we need. */
2147 *start_lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2150 * This loop writes out as many regions as can fit in the amount
2151 * of space which was allocated by xlog_state_get_iclog_space().
2153 while (lv
&& index
< lv
->lv_niovecs
) {
2154 struct xfs_log_iovec
*reg
= &vecp
[index
];
2155 struct xlog_op_header
*ophdr
;
2160 ASSERT(reg
->i_len
% sizeof(__int32_t
) == 0);
2161 ASSERT((unsigned long)ptr
% sizeof(__int32_t
) == 0);
2163 start_rec_copy
= xlog_write_start_rec(ptr
, ticket
);
2164 if (start_rec_copy
) {
2166 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2170 ophdr
= xlog_write_setup_ophdr(log
, ptr
, ticket
, flags
);
2172 return XFS_ERROR(EIO
);
2174 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2175 sizeof(struct xlog_op_header
));
2177 len
+= xlog_write_setup_copy(ticket
, ophdr
,
2178 iclog
->ic_size
-log_offset
,
2180 ©_off
, ©_len
,
2183 xlog_verify_dest_ptr(log
, ptr
);
2186 ASSERT(copy_len
>= 0);
2187 memcpy(ptr
, reg
->i_addr
+ copy_off
, copy_len
);
2188 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
, copy_len
);
2190 copy_len
+= start_rec_copy
+ sizeof(xlog_op_header_t
);
2192 data_cnt
+= contwr
? copy_len
: 0;
2194 error
= xlog_write_copy_finish(log
, iclog
, flags
,
2195 &record_cnt
, &data_cnt
,
2204 * if we had a partial copy, we need to get more iclog
2205 * space but we don't want to increment the region
2206 * index because there is still more is this region to
2209 * If we completed writing this region, and we flushed
2210 * the iclog (indicated by resetting of the record
2211 * count), then we also need to get more log space. If
2212 * this was the last record, though, we are done and
2218 if (++index
== lv
->lv_niovecs
) {
2222 vecp
= lv
->lv_iovecp
;
2224 if (record_cnt
== 0) {
2234 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
2236 return xlog_state_release_iclog(log
, iclog
);
2238 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2239 *commit_iclog
= iclog
;
2244 /*****************************************************************************
2246 * State Machine functions
2248 *****************************************************************************
2251 /* Clean iclogs starting from the head. This ordering must be
2252 * maintained, so an iclog doesn't become ACTIVE beyond one that
2253 * is SYNCING. This is also required to maintain the notion that we use
2254 * a ordered wait queue to hold off would be writers to the log when every
2255 * iclog is trying to sync to disk.
2257 * State Change: DIRTY -> ACTIVE
2260 xlog_state_clean_log(
2263 xlog_in_core_t
*iclog
;
2266 iclog
= log
->l_iclog
;
2268 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2269 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
2270 iclog
->ic_offset
= 0;
2271 ASSERT(iclog
->ic_callback
== NULL
);
2273 * If the number of ops in this iclog indicate it just
2274 * contains the dummy transaction, we can
2275 * change state into IDLE (the second time around).
2276 * Otherwise we should change the state into
2278 * We don't need to cover the dummy.
2281 (be32_to_cpu(iclog
->ic_header
.h_num_logops
) ==
2286 * We have two dirty iclogs so start over
2287 * This could also be num of ops indicates
2288 * this is not the dummy going out.
2292 iclog
->ic_header
.h_num_logops
= 0;
2293 memset(iclog
->ic_header
.h_cycle_data
, 0,
2294 sizeof(iclog
->ic_header
.h_cycle_data
));
2295 iclog
->ic_header
.h_lsn
= 0;
2296 } else if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2299 break; /* stop cleaning */
2300 iclog
= iclog
->ic_next
;
2301 } while (iclog
!= log
->l_iclog
);
2303 /* log is locked when we are called */
2305 * Change state for the dummy log recording.
2306 * We usually go to NEED. But we go to NEED2 if the changed indicates
2307 * we are done writing the dummy record.
2308 * If we are done with the second dummy recored (DONE2), then
2312 switch (log
->l_covered_state
) {
2313 case XLOG_STATE_COVER_IDLE
:
2314 case XLOG_STATE_COVER_NEED
:
2315 case XLOG_STATE_COVER_NEED2
:
2316 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2319 case XLOG_STATE_COVER_DONE
:
2321 log
->l_covered_state
= XLOG_STATE_COVER_NEED2
;
2323 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2326 case XLOG_STATE_COVER_DONE2
:
2328 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
2330 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2337 } /* xlog_state_clean_log */
2340 xlog_get_lowest_lsn(
2343 xlog_in_core_t
*lsn_log
;
2344 xfs_lsn_t lowest_lsn
, lsn
;
2346 lsn_log
= log
->l_iclog
;
2349 if (!(lsn_log
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
))) {
2350 lsn
= be64_to_cpu(lsn_log
->ic_header
.h_lsn
);
2351 if ((lsn
&& !lowest_lsn
) ||
2352 (XFS_LSN_CMP(lsn
, lowest_lsn
) < 0)) {
2356 lsn_log
= lsn_log
->ic_next
;
2357 } while (lsn_log
!= log
->l_iclog
);
2363 xlog_state_do_callback(
2366 struct xlog_in_core
*ciclog
)
2368 xlog_in_core_t
*iclog
;
2369 xlog_in_core_t
*first_iclog
; /* used to know when we've
2370 * processed all iclogs once */
2371 xfs_log_callback_t
*cb
, *cb_next
;
2373 xfs_lsn_t lowest_lsn
;
2374 int ioerrors
; /* counter: iclogs with errors */
2375 int loopdidcallbacks
; /* flag: inner loop did callbacks*/
2376 int funcdidcallbacks
; /* flag: function did callbacks */
2377 int repeats
; /* for issuing console warnings if
2378 * looping too many times */
2381 spin_lock(&log
->l_icloglock
);
2382 first_iclog
= iclog
= log
->l_iclog
;
2384 funcdidcallbacks
= 0;
2389 * Scan all iclogs starting with the one pointed to by the
2390 * log. Reset this starting point each time the log is
2391 * unlocked (during callbacks).
2393 * Keep looping through iclogs until one full pass is made
2394 * without running any callbacks.
2396 first_iclog
= log
->l_iclog
;
2397 iclog
= log
->l_iclog
;
2398 loopdidcallbacks
= 0;
2403 /* skip all iclogs in the ACTIVE & DIRTY states */
2404 if (iclog
->ic_state
&
2405 (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
)) {
2406 iclog
= iclog
->ic_next
;
2411 * Between marking a filesystem SHUTDOWN and stopping
2412 * the log, we do flush all iclogs to disk (if there
2413 * wasn't a log I/O error). So, we do want things to
2414 * go smoothly in case of just a SHUTDOWN w/o a
2417 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
2419 * Can only perform callbacks in order. Since
2420 * this iclog is not in the DONE_SYNC/
2421 * DO_CALLBACK state, we skip the rest and
2422 * just try to clean up. If we set our iclog
2423 * to DO_CALLBACK, we will not process it when
2424 * we retry since a previous iclog is in the
2425 * CALLBACK and the state cannot change since
2426 * we are holding the l_icloglock.
2428 if (!(iclog
->ic_state
&
2429 (XLOG_STATE_DONE_SYNC
|
2430 XLOG_STATE_DO_CALLBACK
))) {
2431 if (ciclog
&& (ciclog
->ic_state
==
2432 XLOG_STATE_DONE_SYNC
)) {
2433 ciclog
->ic_state
= XLOG_STATE_DO_CALLBACK
;
2438 * We now have an iclog that is in either the
2439 * DO_CALLBACK or DONE_SYNC states. The other
2440 * states (WANT_SYNC, SYNCING, or CALLBACK were
2441 * caught by the above if and are going to
2442 * clean (i.e. we aren't doing their callbacks)
2447 * We will do one more check here to see if we
2448 * have chased our tail around.
2451 lowest_lsn
= xlog_get_lowest_lsn(log
);
2453 XFS_LSN_CMP(lowest_lsn
,
2454 be64_to_cpu(iclog
->ic_header
.h_lsn
)) < 0) {
2455 iclog
= iclog
->ic_next
;
2456 continue; /* Leave this iclog for
2460 iclog
->ic_state
= XLOG_STATE_CALLBACK
;
2464 * update the last_sync_lsn before we drop the
2465 * icloglock to ensure we are the only one that
2468 ASSERT(XFS_LSN_CMP(atomic64_read(&log
->l_last_sync_lsn
),
2469 be64_to_cpu(iclog
->ic_header
.h_lsn
)) <= 0);
2470 atomic64_set(&log
->l_last_sync_lsn
,
2471 be64_to_cpu(iclog
->ic_header
.h_lsn
));
2476 spin_unlock(&log
->l_icloglock
);
2479 * Keep processing entries in the callback list until
2480 * we come around and it is empty. We need to
2481 * atomically see that the list is empty and change the
2482 * state to DIRTY so that we don't miss any more
2483 * callbacks being added.
2485 spin_lock(&iclog
->ic_callback_lock
);
2486 cb
= iclog
->ic_callback
;
2488 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
2489 iclog
->ic_callback
= NULL
;
2490 spin_unlock(&iclog
->ic_callback_lock
);
2492 /* perform callbacks in the order given */
2493 for (; cb
; cb
= cb_next
) {
2494 cb_next
= cb
->cb_next
;
2495 cb
->cb_func(cb
->cb_arg
, aborted
);
2497 spin_lock(&iclog
->ic_callback_lock
);
2498 cb
= iclog
->ic_callback
;
2504 spin_lock(&log
->l_icloglock
);
2505 ASSERT(iclog
->ic_callback
== NULL
);
2506 spin_unlock(&iclog
->ic_callback_lock
);
2507 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
))
2508 iclog
->ic_state
= XLOG_STATE_DIRTY
;
2511 * Transition from DIRTY to ACTIVE if applicable.
2512 * NOP if STATE_IOERROR.
2514 xlog_state_clean_log(log
);
2516 /* wake up threads waiting in xfs_log_force() */
2517 wake_up_all(&iclog
->ic_force_wait
);
2519 iclog
= iclog
->ic_next
;
2520 } while (first_iclog
!= iclog
);
2522 if (repeats
> 5000) {
2523 flushcnt
+= repeats
;
2526 "%s: possible infinite loop (%d iterations)",
2527 __func__
, flushcnt
);
2529 } while (!ioerrors
&& loopdidcallbacks
);
2532 * make one last gasp attempt to see if iclogs are being left in
2536 if (funcdidcallbacks
) {
2537 first_iclog
= iclog
= log
->l_iclog
;
2539 ASSERT(iclog
->ic_state
!= XLOG_STATE_DO_CALLBACK
);
2541 * Terminate the loop if iclogs are found in states
2542 * which will cause other threads to clean up iclogs.
2544 * SYNCING - i/o completion will go through logs
2545 * DONE_SYNC - interrupt thread should be waiting for
2547 * IOERROR - give up hope all ye who enter here
2549 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
||
2550 iclog
->ic_state
== XLOG_STATE_SYNCING
||
2551 iclog
->ic_state
== XLOG_STATE_DONE_SYNC
||
2552 iclog
->ic_state
== XLOG_STATE_IOERROR
)
2554 iclog
= iclog
->ic_next
;
2555 } while (first_iclog
!= iclog
);
2559 if (log
->l_iclog
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_IOERROR
))
2561 spin_unlock(&log
->l_icloglock
);
2564 wake_up_all(&log
->l_flush_wait
);
2569 * Finish transitioning this iclog to the dirty state.
2571 * Make sure that we completely execute this routine only when this is
2572 * the last call to the iclog. There is a good chance that iclog flushes,
2573 * when we reach the end of the physical log, get turned into 2 separate
2574 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2575 * routine. By using the reference count bwritecnt, we guarantee that only
2576 * the second completion goes through.
2578 * Callbacks could take time, so they are done outside the scope of the
2579 * global state machine log lock.
2582 xlog_state_done_syncing(
2583 xlog_in_core_t
*iclog
,
2586 struct xlog
*log
= iclog
->ic_log
;
2588 spin_lock(&log
->l_icloglock
);
2590 ASSERT(iclog
->ic_state
== XLOG_STATE_SYNCING
||
2591 iclog
->ic_state
== XLOG_STATE_IOERROR
);
2592 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2593 ASSERT(iclog
->ic_bwritecnt
== 1 || iclog
->ic_bwritecnt
== 2);
2597 * If we got an error, either on the first buffer, or in the case of
2598 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2599 * and none should ever be attempted to be written to disk
2602 if (iclog
->ic_state
!= XLOG_STATE_IOERROR
) {
2603 if (--iclog
->ic_bwritecnt
== 1) {
2604 spin_unlock(&log
->l_icloglock
);
2607 iclog
->ic_state
= XLOG_STATE_DONE_SYNC
;
2611 * Someone could be sleeping prior to writing out the next
2612 * iclog buffer, we wake them all, one will get to do the
2613 * I/O, the others get to wait for the result.
2615 wake_up_all(&iclog
->ic_write_wait
);
2616 spin_unlock(&log
->l_icloglock
);
2617 xlog_state_do_callback(log
, aborted
, iclog
); /* also cleans log */
2618 } /* xlog_state_done_syncing */
2622 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2623 * sleep. We wait on the flush queue on the head iclog as that should be
2624 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2625 * we will wait here and all new writes will sleep until a sync completes.
2627 * The in-core logs are used in a circular fashion. They are not used
2628 * out-of-order even when an iclog past the head is free.
2631 * * log_offset where xlog_write() can start writing into the in-core
2633 * * in-core log pointer to which xlog_write() should write.
2634 * * boolean indicating this is a continued write to an in-core log.
2635 * If this is the last write, then the in-core log's offset field
2636 * needs to be incremented, depending on the amount of data which
2640 xlog_state_get_iclog_space(
2643 struct xlog_in_core
**iclogp
,
2644 struct xlog_ticket
*ticket
,
2645 int *continued_write
,
2649 xlog_rec_header_t
*head
;
2650 xlog_in_core_t
*iclog
;
2654 spin_lock(&log
->l_icloglock
);
2655 if (XLOG_FORCED_SHUTDOWN(log
)) {
2656 spin_unlock(&log
->l_icloglock
);
2657 return XFS_ERROR(EIO
);
2660 iclog
= log
->l_iclog
;
2661 if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
) {
2662 XFS_STATS_INC(xs_log_noiclogs
);
2664 /* Wait for log writes to have flushed */
2665 xlog_wait(&log
->l_flush_wait
, &log
->l_icloglock
);
2669 head
= &iclog
->ic_header
;
2671 atomic_inc(&iclog
->ic_refcnt
); /* prevents sync */
2672 log_offset
= iclog
->ic_offset
;
2674 /* On the 1st write to an iclog, figure out lsn. This works
2675 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2676 * committing to. If the offset is set, that's how many blocks
2679 if (log_offset
== 0) {
2680 ticket
->t_curr_res
-= log
->l_iclog_hsize
;
2681 xlog_tic_add_region(ticket
,
2683 XLOG_REG_TYPE_LRHEADER
);
2684 head
->h_cycle
= cpu_to_be32(log
->l_curr_cycle
);
2685 head
->h_lsn
= cpu_to_be64(
2686 xlog_assign_lsn(log
->l_curr_cycle
, log
->l_curr_block
));
2687 ASSERT(log
->l_curr_block
>= 0);
2690 /* If there is enough room to write everything, then do it. Otherwise,
2691 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2692 * bit is on, so this will get flushed out. Don't update ic_offset
2693 * until you know exactly how many bytes get copied. Therefore, wait
2694 * until later to update ic_offset.
2696 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2697 * can fit into remaining data section.
2699 if (iclog
->ic_size
- iclog
->ic_offset
< 2*sizeof(xlog_op_header_t
)) {
2700 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2703 * If I'm the only one writing to this iclog, sync it to disk.
2704 * We need to do an atomic compare and decrement here to avoid
2705 * racing with concurrent atomic_dec_and_lock() calls in
2706 * xlog_state_release_iclog() when there is more than one
2707 * reference to the iclog.
2709 if (!atomic_add_unless(&iclog
->ic_refcnt
, -1, 1)) {
2710 /* we are the only one */
2711 spin_unlock(&log
->l_icloglock
);
2712 error
= xlog_state_release_iclog(log
, iclog
);
2716 spin_unlock(&log
->l_icloglock
);
2721 /* Do we have enough room to write the full amount in the remainder
2722 * of this iclog? Or must we continue a write on the next iclog and
2723 * mark this iclog as completely taken? In the case where we switch
2724 * iclogs (to mark it taken), this particular iclog will release/sync
2725 * to disk in xlog_write().
2727 if (len
<= iclog
->ic_size
- iclog
->ic_offset
) {
2728 *continued_write
= 0;
2729 iclog
->ic_offset
+= len
;
2731 *continued_write
= 1;
2732 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2736 ASSERT(iclog
->ic_offset
<= iclog
->ic_size
);
2737 spin_unlock(&log
->l_icloglock
);
2739 *logoffsetp
= log_offset
;
2741 } /* xlog_state_get_iclog_space */
2743 /* The first cnt-1 times through here we don't need to
2744 * move the grant write head because the permanent
2745 * reservation has reserved cnt times the unit amount.
2746 * Release part of current permanent unit reservation and
2747 * reset current reservation to be one units worth. Also
2748 * move grant reservation head forward.
2751 xlog_regrant_reserve_log_space(
2753 struct xlog_ticket
*ticket
)
2755 trace_xfs_log_regrant_reserve_enter(log
, ticket
);
2757 if (ticket
->t_cnt
> 0)
2760 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
,
2761 ticket
->t_curr_res
);
2762 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
,
2763 ticket
->t_curr_res
);
2764 ticket
->t_curr_res
= ticket
->t_unit_res
;
2765 xlog_tic_reset_res(ticket
);
2767 trace_xfs_log_regrant_reserve_sub(log
, ticket
);
2769 /* just return if we still have some of the pre-reserved space */
2770 if (ticket
->t_cnt
> 0)
2773 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
,
2774 ticket
->t_unit_res
);
2776 trace_xfs_log_regrant_reserve_exit(log
, ticket
);
2778 ticket
->t_curr_res
= ticket
->t_unit_res
;
2779 xlog_tic_reset_res(ticket
);
2780 } /* xlog_regrant_reserve_log_space */
2784 * Give back the space left from a reservation.
2786 * All the information we need to make a correct determination of space left
2787 * is present. For non-permanent reservations, things are quite easy. The
2788 * count should have been decremented to zero. We only need to deal with the
2789 * space remaining in the current reservation part of the ticket. If the
2790 * ticket contains a permanent reservation, there may be left over space which
2791 * needs to be released. A count of N means that N-1 refills of the current
2792 * reservation can be done before we need to ask for more space. The first
2793 * one goes to fill up the first current reservation. Once we run out of
2794 * space, the count will stay at zero and the only space remaining will be
2795 * in the current reservation field.
2798 xlog_ungrant_log_space(
2800 struct xlog_ticket
*ticket
)
2804 if (ticket
->t_cnt
> 0)
2807 trace_xfs_log_ungrant_enter(log
, ticket
);
2808 trace_xfs_log_ungrant_sub(log
, ticket
);
2811 * If this is a permanent reservation ticket, we may be able to free
2812 * up more space based on the remaining count.
2814 bytes
= ticket
->t_curr_res
;
2815 if (ticket
->t_cnt
> 0) {
2816 ASSERT(ticket
->t_flags
& XLOG_TIC_PERM_RESERV
);
2817 bytes
+= ticket
->t_unit_res
*ticket
->t_cnt
;
2820 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
, bytes
);
2821 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
, bytes
);
2823 trace_xfs_log_ungrant_exit(log
, ticket
);
2825 xfs_log_space_wake(log
->l_mp
);
2829 * Flush iclog to disk if this is the last reference to the given iclog and
2830 * the WANT_SYNC bit is set.
2832 * When this function is entered, the iclog is not necessarily in the
2833 * WANT_SYNC state. It may be sitting around waiting to get filled.
2838 xlog_state_release_iclog(
2840 struct xlog_in_core
*iclog
)
2842 int sync
= 0; /* do we sync? */
2844 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
2845 return XFS_ERROR(EIO
);
2847 ASSERT(atomic_read(&iclog
->ic_refcnt
) > 0);
2848 if (!atomic_dec_and_lock(&iclog
->ic_refcnt
, &log
->l_icloglock
))
2851 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2852 spin_unlock(&log
->l_icloglock
);
2853 return XFS_ERROR(EIO
);
2855 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2856 iclog
->ic_state
== XLOG_STATE_WANT_SYNC
);
2858 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
) {
2859 /* update tail before writing to iclog */
2860 xfs_lsn_t tail_lsn
= xlog_assign_tail_lsn(log
->l_mp
);
2862 iclog
->ic_state
= XLOG_STATE_SYNCING
;
2863 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(tail_lsn
);
2864 xlog_verify_tail_lsn(log
, iclog
, tail_lsn
);
2865 /* cycle incremented when incrementing curr_block */
2867 spin_unlock(&log
->l_icloglock
);
2870 * We let the log lock go, so it's possible that we hit a log I/O
2871 * error or some other SHUTDOWN condition that marks the iclog
2872 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2873 * this iclog has consistent data, so we ignore IOERROR
2874 * flags after this point.
2877 return xlog_sync(log
, iclog
);
2879 } /* xlog_state_release_iclog */
2883 * This routine will mark the current iclog in the ring as WANT_SYNC
2884 * and move the current iclog pointer to the next iclog in the ring.
2885 * When this routine is called from xlog_state_get_iclog_space(), the
2886 * exact size of the iclog has not yet been determined. All we know is
2887 * that every data block. We have run out of space in this log record.
2890 xlog_state_switch_iclogs(
2892 struct xlog_in_core
*iclog
,
2895 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
2897 eventual_size
= iclog
->ic_offset
;
2898 iclog
->ic_state
= XLOG_STATE_WANT_SYNC
;
2899 iclog
->ic_header
.h_prev_block
= cpu_to_be32(log
->l_prev_block
);
2900 log
->l_prev_block
= log
->l_curr_block
;
2901 log
->l_prev_cycle
= log
->l_curr_cycle
;
2903 /* roll log?: ic_offset changed later */
2904 log
->l_curr_block
+= BTOBB(eventual_size
)+BTOBB(log
->l_iclog_hsize
);
2906 /* Round up to next log-sunit */
2907 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
2908 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
2909 __uint32_t sunit_bb
= BTOBB(log
->l_mp
->m_sb
.sb_logsunit
);
2910 log
->l_curr_block
= roundup(log
->l_curr_block
, sunit_bb
);
2913 if (log
->l_curr_block
>= log
->l_logBBsize
) {
2914 log
->l_curr_cycle
++;
2915 if (log
->l_curr_cycle
== XLOG_HEADER_MAGIC_NUM
)
2916 log
->l_curr_cycle
++;
2917 log
->l_curr_block
-= log
->l_logBBsize
;
2918 ASSERT(log
->l_curr_block
>= 0);
2920 ASSERT(iclog
== log
->l_iclog
);
2921 log
->l_iclog
= iclog
->ic_next
;
2922 } /* xlog_state_switch_iclogs */
2925 * Write out all data in the in-core log as of this exact moment in time.
2927 * Data may be written to the in-core log during this call. However,
2928 * we don't guarantee this data will be written out. A change from past
2929 * implementation means this routine will *not* write out zero length LRs.
2931 * Basically, we try and perform an intelligent scan of the in-core logs.
2932 * If we determine there is no flushable data, we just return. There is no
2933 * flushable data if:
2935 * 1. the current iclog is active and has no data; the previous iclog
2936 * is in the active or dirty state.
2937 * 2. the current iclog is drity, and the previous iclog is in the
2938 * active or dirty state.
2942 * 1. the current iclog is not in the active nor dirty state.
2943 * 2. the current iclog dirty, and the previous iclog is not in the
2944 * active nor dirty state.
2945 * 3. the current iclog is active, and there is another thread writing
2946 * to this particular iclog.
2947 * 4. a) the current iclog is active and has no other writers
2948 * b) when we return from flushing out this iclog, it is still
2949 * not in the active nor dirty state.
2953 struct xfs_mount
*mp
,
2957 struct xlog
*log
= mp
->m_log
;
2958 struct xlog_in_core
*iclog
;
2961 XFS_STATS_INC(xs_log_force
);
2963 xlog_cil_force(log
);
2965 spin_lock(&log
->l_icloglock
);
2967 iclog
= log
->l_iclog
;
2968 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2969 spin_unlock(&log
->l_icloglock
);
2970 return XFS_ERROR(EIO
);
2973 /* If the head iclog is not active nor dirty, we just attach
2974 * ourselves to the head and go to sleep.
2976 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2977 iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2979 * If the head is dirty or (active and empty), then
2980 * we need to look at the previous iclog. If the previous
2981 * iclog is active or dirty we are done. There is nothing
2982 * to sync out. Otherwise, we attach ourselves to the
2983 * previous iclog and go to sleep.
2985 if (iclog
->ic_state
== XLOG_STATE_DIRTY
||
2986 (atomic_read(&iclog
->ic_refcnt
) == 0
2987 && iclog
->ic_offset
== 0)) {
2988 iclog
= iclog
->ic_prev
;
2989 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2990 iclog
->ic_state
== XLOG_STATE_DIRTY
)
2995 if (atomic_read(&iclog
->ic_refcnt
) == 0) {
2996 /* We are the only one with access to this
2997 * iclog. Flush it out now. There should
2998 * be a roundoff of zero to show that someone
2999 * has already taken care of the roundoff from
3000 * the previous sync.
3002 atomic_inc(&iclog
->ic_refcnt
);
3003 lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
3004 xlog_state_switch_iclogs(log
, iclog
, 0);
3005 spin_unlock(&log
->l_icloglock
);
3007 if (xlog_state_release_iclog(log
, iclog
))
3008 return XFS_ERROR(EIO
);
3012 spin_lock(&log
->l_icloglock
);
3013 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) == lsn
&&
3014 iclog
->ic_state
!= XLOG_STATE_DIRTY
)
3019 /* Someone else is writing to this iclog.
3020 * Use its call to flush out the data. However,
3021 * the other thread may not force out this LR,
3022 * so we mark it WANT_SYNC.
3024 xlog_state_switch_iclogs(log
, iclog
, 0);
3030 /* By the time we come around again, the iclog could've been filled
3031 * which would give it another lsn. If we have a new lsn, just
3032 * return because the relevant data has been flushed.
3035 if (flags
& XFS_LOG_SYNC
) {
3037 * We must check if we're shutting down here, before
3038 * we wait, while we're holding the l_icloglock.
3039 * Then we check again after waking up, in case our
3040 * sleep was disturbed by a bad news.
3042 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3043 spin_unlock(&log
->l_icloglock
);
3044 return XFS_ERROR(EIO
);
3046 XFS_STATS_INC(xs_log_force_sleep
);
3047 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3049 * No need to grab the log lock here since we're
3050 * only deciding whether or not to return EIO
3051 * and the memory read should be atomic.
3053 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3054 return XFS_ERROR(EIO
);
3060 spin_unlock(&log
->l_icloglock
);
3066 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3067 * about errors or whether the log was flushed or not. This is the normal
3068 * interface to use when trying to unpin items or move the log forward.
3077 trace_xfs_log_force(mp
, 0);
3078 error
= _xfs_log_force(mp
, flags
, NULL
);
3080 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3084 * Force the in-core log to disk for a specific LSN.
3086 * Find in-core log with lsn.
3087 * If it is in the DIRTY state, just return.
3088 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3089 * state and go to sleep or return.
3090 * If it is in any other state, go to sleep or return.
3092 * Synchronous forces are implemented with a signal variable. All callers
3093 * to force a given lsn to disk will wait on a the sv attached to the
3094 * specific in-core log. When given in-core log finally completes its
3095 * write to disk, that thread will wake up all threads waiting on the
3100 struct xfs_mount
*mp
,
3105 struct xlog
*log
= mp
->m_log
;
3106 struct xlog_in_core
*iclog
;
3107 int already_slept
= 0;
3111 XFS_STATS_INC(xs_log_force
);
3113 lsn
= xlog_cil_force_lsn(log
, lsn
);
3114 if (lsn
== NULLCOMMITLSN
)
3118 spin_lock(&log
->l_icloglock
);
3119 iclog
= log
->l_iclog
;
3120 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3121 spin_unlock(&log
->l_icloglock
);
3122 return XFS_ERROR(EIO
);
3126 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
) {
3127 iclog
= iclog
->ic_next
;
3131 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
3132 spin_unlock(&log
->l_icloglock
);
3136 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3138 * We sleep here if we haven't already slept (e.g.
3139 * this is the first time we've looked at the correct
3140 * iclog buf) and the buffer before us is going to
3141 * be sync'ed. The reason for this is that if we
3142 * are doing sync transactions here, by waiting for
3143 * the previous I/O to complete, we can allow a few
3144 * more transactions into this iclog before we close
3147 * Otherwise, we mark the buffer WANT_SYNC, and bump
3148 * up the refcnt so we can release the log (which
3149 * drops the ref count). The state switch keeps new
3150 * transaction commits from using this buffer. When
3151 * the current commits finish writing into the buffer,
3152 * the refcount will drop to zero and the buffer will
3155 if (!already_slept
&&
3156 (iclog
->ic_prev
->ic_state
&
3157 (XLOG_STATE_WANT_SYNC
| XLOG_STATE_SYNCING
))) {
3158 ASSERT(!(iclog
->ic_state
& XLOG_STATE_IOERROR
));
3160 XFS_STATS_INC(xs_log_force_sleep
);
3162 xlog_wait(&iclog
->ic_prev
->ic_write_wait
,
3169 atomic_inc(&iclog
->ic_refcnt
);
3170 xlog_state_switch_iclogs(log
, iclog
, 0);
3171 spin_unlock(&log
->l_icloglock
);
3172 if (xlog_state_release_iclog(log
, iclog
))
3173 return XFS_ERROR(EIO
);
3176 spin_lock(&log
->l_icloglock
);
3179 if ((flags
& XFS_LOG_SYNC
) && /* sleep */
3181 (XLOG_STATE_ACTIVE
| XLOG_STATE_DIRTY
))) {
3183 * Don't wait on completion if we know that we've
3184 * gotten a log write error.
3186 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3187 spin_unlock(&log
->l_icloglock
);
3188 return XFS_ERROR(EIO
);
3190 XFS_STATS_INC(xs_log_force_sleep
);
3191 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3193 * No need to grab the log lock here since we're
3194 * only deciding whether or not to return EIO
3195 * and the memory read should be atomic.
3197 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3198 return XFS_ERROR(EIO
);
3202 } else { /* just return */
3203 spin_unlock(&log
->l_icloglock
);
3207 } while (iclog
!= log
->l_iclog
);
3209 spin_unlock(&log
->l_icloglock
);
3214 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3215 * about errors or whether the log was flushed or not. This is the normal
3216 * interface to use when trying to unpin items or move the log forward.
3226 trace_xfs_log_force(mp
, lsn
);
3227 error
= _xfs_log_force_lsn(mp
, lsn
, flags
, NULL
);
3229 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3233 * Called when we want to mark the current iclog as being ready to sync to
3237 xlog_state_want_sync(
3239 struct xlog_in_core
*iclog
)
3241 assert_spin_locked(&log
->l_icloglock
);
3243 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3244 xlog_state_switch_iclogs(log
, iclog
, 0);
3246 ASSERT(iclog
->ic_state
&
3247 (XLOG_STATE_WANT_SYNC
|XLOG_STATE_IOERROR
));
3252 /*****************************************************************************
3256 *****************************************************************************
3260 * Free a used ticket when its refcount falls to zero.
3264 xlog_ticket_t
*ticket
)
3266 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3267 if (atomic_dec_and_test(&ticket
->t_ref
))
3268 kmem_zone_free(xfs_log_ticket_zone
, ticket
);
3273 xlog_ticket_t
*ticket
)
3275 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3276 atomic_inc(&ticket
->t_ref
);
3281 * Allocate and initialise a new log ticket.
3283 struct xlog_ticket
*
3290 xfs_km_flags_t alloc_flags
)
3292 struct xlog_ticket
*tic
;
3296 tic
= kmem_zone_zalloc(xfs_log_ticket_zone
, alloc_flags
);
3301 * Permanent reservations have up to 'cnt'-1 active log operations
3302 * in the log. A unit in this case is the amount of space for one
3303 * of these log operations. Normal reservations have a cnt of 1
3304 * and their unit amount is the total amount of space required.
3306 * The following lines of code account for non-transaction data
3307 * which occupy space in the on-disk log.
3309 * Normal form of a transaction is:
3310 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3311 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3313 * We need to account for all the leadup data and trailer data
3314 * around the transaction data.
3315 * And then we need to account for the worst case in terms of using
3317 * The worst case will happen if:
3318 * - the placement of the transaction happens to be such that the
3319 * roundoff is at its maximum
3320 * - the transaction data is synced before the commit record is synced
3321 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3322 * Therefore the commit record is in its own Log Record.
3323 * This can happen as the commit record is called with its
3324 * own region to xlog_write().
3325 * This then means that in the worst case, roundoff can happen for
3326 * the commit-rec as well.
3327 * The commit-rec is smaller than padding in this scenario and so it is
3328 * not added separately.
3331 /* for trans header */
3332 unit_bytes
+= sizeof(xlog_op_header_t
);
3333 unit_bytes
+= sizeof(xfs_trans_header_t
);
3336 unit_bytes
+= sizeof(xlog_op_header_t
);
3339 * for LR headers - the space for data in an iclog is the size minus
3340 * the space used for the headers. If we use the iclog size, then we
3341 * undercalculate the number of headers required.
3343 * Furthermore - the addition of op headers for split-recs might
3344 * increase the space required enough to require more log and op
3345 * headers, so take that into account too.
3347 * IMPORTANT: This reservation makes the assumption that if this
3348 * transaction is the first in an iclog and hence has the LR headers
3349 * accounted to it, then the remaining space in the iclog is
3350 * exclusively for this transaction. i.e. if the transaction is larger
3351 * than the iclog, it will be the only thing in that iclog.
3352 * Fundamentally, this means we must pass the entire log vector to
3353 * xlog_write to guarantee this.
3355 iclog_space
= log
->l_iclog_size
- log
->l_iclog_hsize
;
3356 num_headers
= howmany(unit_bytes
, iclog_space
);
3358 /* for split-recs - ophdrs added when data split over LRs */
3359 unit_bytes
+= sizeof(xlog_op_header_t
) * num_headers
;
3361 /* add extra header reservations if we overrun */
3362 while (!num_headers
||
3363 howmany(unit_bytes
, iclog_space
) > num_headers
) {
3364 unit_bytes
+= sizeof(xlog_op_header_t
);
3367 unit_bytes
+= log
->l_iclog_hsize
* num_headers
;
3369 /* for commit-rec LR header - note: padding will subsume the ophdr */
3370 unit_bytes
+= log
->l_iclog_hsize
;
3372 /* for roundoff padding for transaction data and one for commit record */
3373 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
3374 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
3375 /* log su roundoff */
3376 unit_bytes
+= 2*log
->l_mp
->m_sb
.sb_logsunit
;
3379 unit_bytes
+= 2*BBSIZE
;
3382 atomic_set(&tic
->t_ref
, 1);
3383 tic
->t_task
= current
;
3384 INIT_LIST_HEAD(&tic
->t_queue
);
3385 tic
->t_unit_res
= unit_bytes
;
3386 tic
->t_curr_res
= unit_bytes
;
3389 tic
->t_tid
= random32();
3390 tic
->t_clientid
= client
;
3391 tic
->t_flags
= XLOG_TIC_INITED
;
3392 tic
->t_trans_type
= 0;
3394 tic
->t_flags
|= XLOG_TIC_PERM_RESERV
;
3396 xlog_tic_reset_res(tic
);
3402 /******************************************************************************
3404 * Log debug routines
3406 ******************************************************************************
3410 * Make sure that the destination ptr is within the valid data region of
3411 * one of the iclogs. This uses backup pointers stored in a different
3412 * part of the log in case we trash the log structure.
3415 xlog_verify_dest_ptr(
3422 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
3423 if (ptr
>= log
->l_iclog_bak
[i
] &&
3424 ptr
<= log
->l_iclog_bak
[i
] + log
->l_iclog_size
)
3429 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3433 * Check to make sure the grant write head didn't just over lap the tail. If
3434 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3435 * the cycles differ by exactly one and check the byte count.
3437 * This check is run unlocked, so can give false positives. Rather than assert
3438 * on failures, use a warn-once flag and a panic tag to allow the admin to
3439 * determine if they want to panic the machine when such an error occurs. For
3440 * debug kernels this will have the same effect as using an assert but, unlinke
3441 * an assert, it can be turned off at runtime.
3444 xlog_verify_grant_tail(
3447 int tail_cycle
, tail_blocks
;
3450 xlog_crack_grant_head(&log
->l_write_head
.grant
, &cycle
, &space
);
3451 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_blocks
);
3452 if (tail_cycle
!= cycle
) {
3453 if (cycle
- 1 != tail_cycle
&&
3454 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3455 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3456 "%s: cycle - 1 != tail_cycle", __func__
);
3457 log
->l_flags
|= XLOG_TAIL_WARN
;
3460 if (space
> BBTOB(tail_blocks
) &&
3461 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3462 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3463 "%s: space > BBTOB(tail_blocks)", __func__
);
3464 log
->l_flags
|= XLOG_TAIL_WARN
;
3469 /* check if it will fit */
3471 xlog_verify_tail_lsn(
3473 struct xlog_in_core
*iclog
,
3478 if (CYCLE_LSN(tail_lsn
) == log
->l_prev_cycle
) {
3480 log
->l_logBBsize
- (log
->l_prev_block
- BLOCK_LSN(tail_lsn
));
3481 if (blocks
< BTOBB(iclog
->ic_offset
)+BTOBB(log
->l_iclog_hsize
))
3482 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3484 ASSERT(CYCLE_LSN(tail_lsn
)+1 == log
->l_prev_cycle
);
3486 if (BLOCK_LSN(tail_lsn
) == log
->l_prev_block
)
3487 xfs_emerg(log
->l_mp
, "%s: tail wrapped", __func__
);
3489 blocks
= BLOCK_LSN(tail_lsn
) - log
->l_prev_block
;
3490 if (blocks
< BTOBB(iclog
->ic_offset
) + 1)
3491 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3493 } /* xlog_verify_tail_lsn */
3496 * Perform a number of checks on the iclog before writing to disk.
3498 * 1. Make sure the iclogs are still circular
3499 * 2. Make sure we have a good magic number
3500 * 3. Make sure we don't have magic numbers in the data
3501 * 4. Check fields of each log operation header for:
3502 * A. Valid client identifier
3503 * B. tid ptr value falls in valid ptr space (user space code)
3504 * C. Length in log record header is correct according to the
3505 * individual operation headers within record.
3506 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3507 * log, check the preceding blocks of the physical log to make sure all
3508 * the cycle numbers agree with the current cycle number.
3513 struct xlog_in_core
*iclog
,
3517 xlog_op_header_t
*ophead
;
3518 xlog_in_core_t
*icptr
;
3519 xlog_in_core_2_t
*xhdr
;
3521 xfs_caddr_t base_ptr
;
3522 __psint_t field_offset
;
3524 int len
, i
, j
, k
, op_len
;
3527 /* check validity of iclog pointers */
3528 spin_lock(&log
->l_icloglock
);
3529 icptr
= log
->l_iclog
;
3530 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
3532 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3533 icptr
= icptr
->ic_next
;
3535 if (icptr
!= log
->l_iclog
)
3536 xfs_emerg(log
->l_mp
, "%s: corrupt iclog ring", __func__
);
3537 spin_unlock(&log
->l_icloglock
);
3539 /* check log magic numbers */
3540 if (iclog
->ic_header
.h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3541 xfs_emerg(log
->l_mp
, "%s: invalid magic num", __func__
);
3543 ptr
= (xfs_caddr_t
) &iclog
->ic_header
;
3544 for (ptr
+= BBSIZE
; ptr
< ((xfs_caddr_t
)&iclog
->ic_header
) + count
;
3546 if (*(__be32
*)ptr
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3547 xfs_emerg(log
->l_mp
, "%s: unexpected magic num",
3552 len
= be32_to_cpu(iclog
->ic_header
.h_num_logops
);
3553 ptr
= iclog
->ic_datap
;
3555 ophead
= (xlog_op_header_t
*)ptr
;
3556 xhdr
= iclog
->ic_data
;
3557 for (i
= 0; i
< len
; i
++) {
3558 ophead
= (xlog_op_header_t
*)ptr
;
3560 /* clientid is only 1 byte */
3561 field_offset
= (__psint_t
)
3562 ((xfs_caddr_t
)&(ophead
->oh_clientid
) - base_ptr
);
3563 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3564 clientid
= ophead
->oh_clientid
;
3566 idx
= BTOBBT((xfs_caddr_t
)&(ophead
->oh_clientid
) - iclog
->ic_datap
);
3567 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3568 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3569 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3570 clientid
= xlog_get_client_id(
3571 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3573 clientid
= xlog_get_client_id(
3574 iclog
->ic_header
.h_cycle_data
[idx
]);
3577 if (clientid
!= XFS_TRANSACTION
&& clientid
!= XFS_LOG
)
3579 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3580 __func__
, clientid
, ophead
,
3581 (unsigned long)field_offset
);
3584 field_offset
= (__psint_t
)
3585 ((xfs_caddr_t
)&(ophead
->oh_len
) - base_ptr
);
3586 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3587 op_len
= be32_to_cpu(ophead
->oh_len
);
3589 idx
= BTOBBT((__psint_t
)&ophead
->oh_len
-
3590 (__psint_t
)iclog
->ic_datap
);
3591 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3592 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3593 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3594 op_len
= be32_to_cpu(xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3596 op_len
= be32_to_cpu(iclog
->ic_header
.h_cycle_data
[idx
]);
3599 ptr
+= sizeof(xlog_op_header_t
) + op_len
;
3601 } /* xlog_verify_iclog */
3605 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3611 xlog_in_core_t
*iclog
, *ic
;
3613 iclog
= log
->l_iclog
;
3614 if (! (iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3616 * Mark all the incore logs IOERROR.
3617 * From now on, no log flushes will result.
3621 ic
->ic_state
= XLOG_STATE_IOERROR
;
3623 } while (ic
!= iclog
);
3627 * Return non-zero, if state transition has already happened.
3633 * This is called from xfs_force_shutdown, when we're forcibly
3634 * shutting down the filesystem, typically because of an IO error.
3635 * Our main objectives here are to make sure that:
3636 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3637 * parties to find out, 'atomically'.
3638 * b. those who're sleeping on log reservations, pinned objects and
3639 * other resources get woken up, and be told the bad news.
3640 * c. nothing new gets queued up after (a) and (b) are done.
3641 * d. if !logerror, flush the iclogs to disk, then seal them off
3644 * Note: for delayed logging the !logerror case needs to flush the regions
3645 * held in memory out to the iclogs before flushing them to disk. This needs
3646 * to be done before the log is marked as shutdown, otherwise the flush to the
3650 xfs_log_force_umount(
3651 struct xfs_mount
*mp
,
3660 * If this happens during log recovery, don't worry about
3661 * locking; the log isn't open for business yet.
3664 log
->l_flags
& XLOG_ACTIVE_RECOVERY
) {
3665 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3667 XFS_BUF_DONE(mp
->m_sb_bp
);
3672 * Somebody could've already done the hard work for us.
3673 * No need to get locks for this.
3675 if (logerror
&& log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3676 ASSERT(XLOG_FORCED_SHUTDOWN(log
));
3682 * Flush the in memory commit item list before marking the log as
3683 * being shut down. We need to do it in this order to ensure all the
3684 * completed transactions are flushed to disk with the xfs_log_force()
3688 xlog_cil_force(log
);
3691 * mark the filesystem and the as in a shutdown state and wake
3692 * everybody up to tell them the bad news.
3694 spin_lock(&log
->l_icloglock
);
3695 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3697 XFS_BUF_DONE(mp
->m_sb_bp
);
3700 * This flag is sort of redundant because of the mount flag, but
3701 * it's good to maintain the separation between the log and the rest
3704 log
->l_flags
|= XLOG_IO_ERROR
;
3707 * If we hit a log error, we want to mark all the iclogs IOERROR
3708 * while we're still holding the loglock.
3711 retval
= xlog_state_ioerror(log
);
3712 spin_unlock(&log
->l_icloglock
);
3715 * We don't want anybody waiting for log reservations after this. That
3716 * means we have to wake up everybody queued up on reserveq as well as
3717 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3718 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3719 * action is protected by the grant locks.
3721 xlog_grant_head_wake_all(&log
->l_reserve_head
);
3722 xlog_grant_head_wake_all(&log
->l_write_head
);
3724 if (!(log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3727 * Force the incore logs to disk before shutting the
3728 * log down completely.
3730 _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
3732 spin_lock(&log
->l_icloglock
);
3733 retval
= xlog_state_ioerror(log
);
3734 spin_unlock(&log
->l_icloglock
);
3737 * Wake up everybody waiting on xfs_log_force.
3738 * Callback all log item committed functions as if the
3739 * log writes were completed.
3741 xlog_state_do_callback(log
, XFS_LI_ABORTED
, NULL
);
3743 #ifdef XFSERRORDEBUG
3745 xlog_in_core_t
*iclog
;
3747 spin_lock(&log
->l_icloglock
);
3748 iclog
= log
->l_iclog
;
3750 ASSERT(iclog
->ic_callback
== 0);
3751 iclog
= iclog
->ic_next
;
3752 } while (iclog
!= log
->l_iclog
);
3753 spin_unlock(&log
->l_icloglock
);
3756 /* return non-zero if log IOERROR transition had already happened */
3764 xlog_in_core_t
*iclog
;
3766 iclog
= log
->l_iclog
;
3768 /* endianness does not matter here, zero is zero in
3771 if (iclog
->ic_header
.h_num_logops
)
3773 iclog
= iclog
->ic_next
;
3774 } while (iclog
!= log
->l_iclog
);