2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_types.h"
24 #include "xfs_trans.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
44 kmem_zone_t
*xfs_log_ticket_zone
;
46 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
51 /* Local miscellaneous function prototypes */
52 STATIC
int xlog_bdstrat_cb(struct xfs_buf
*);
53 STATIC
int xlog_commit_record(xfs_mount_t
*mp
, xlog_ticket_t
*ticket
,
54 xlog_in_core_t
**, xfs_lsn_t
*);
55 STATIC xlog_t
* xlog_alloc_log(xfs_mount_t
*mp
,
56 xfs_buftarg_t
*log_target
,
57 xfs_daddr_t blk_offset
,
59 STATIC
int xlog_space_left(xlog_t
*log
, int cycle
, int bytes
);
60 STATIC
int xlog_sync(xlog_t
*log
, xlog_in_core_t
*iclog
);
61 STATIC
void xlog_dealloc_log(xlog_t
*log
);
62 STATIC
int xlog_write(xfs_mount_t
*mp
, xfs_log_iovec_t region
[],
63 int nentries
, xfs_log_ticket_t tic
,
65 xlog_in_core_t
**commit_iclog
,
68 /* local state machine functions */
69 STATIC
void xlog_state_done_syncing(xlog_in_core_t
*iclog
, int);
70 STATIC
void xlog_state_do_callback(xlog_t
*log
,int aborted
, xlog_in_core_t
*iclog
);
71 STATIC
int xlog_state_get_iclog_space(xlog_t
*log
,
73 xlog_in_core_t
**iclog
,
74 xlog_ticket_t
*ticket
,
77 STATIC
int xlog_state_release_iclog(xlog_t
*log
,
78 xlog_in_core_t
*iclog
);
79 STATIC
void xlog_state_switch_iclogs(xlog_t
*log
,
80 xlog_in_core_t
*iclog
,
82 STATIC
int xlog_state_sync(xlog_t
*log
,
86 STATIC
int xlog_state_sync_all(xlog_t
*log
, uint flags
, int *log_flushed
);
87 STATIC
void xlog_state_want_sync(xlog_t
*log
, xlog_in_core_t
*iclog
);
89 /* local functions to manipulate grant head */
90 STATIC
int xlog_grant_log_space(xlog_t
*log
,
92 STATIC
void xlog_grant_push_ail(xfs_mount_t
*mp
,
94 STATIC
void xlog_regrant_reserve_log_space(xlog_t
*log
,
95 xlog_ticket_t
*ticket
);
96 STATIC
int xlog_regrant_write_log_space(xlog_t
*log
,
97 xlog_ticket_t
*ticket
);
98 STATIC
void xlog_ungrant_log_space(xlog_t
*log
,
99 xlog_ticket_t
*ticket
);
102 /* local ticket functions */
103 STATIC xlog_ticket_t
*xlog_ticket_get(xlog_t
*log
,
108 STATIC
void xlog_ticket_put(xlog_t
*log
, xlog_ticket_t
*ticket
);
111 STATIC
void xlog_verify_dest_ptr(xlog_t
*log
, __psint_t ptr
);
112 STATIC
void xlog_verify_grant_head(xlog_t
*log
, int equals
);
113 STATIC
void xlog_verify_iclog(xlog_t
*log
, xlog_in_core_t
*iclog
,
114 int count
, boolean_t syncing
);
115 STATIC
void xlog_verify_tail_lsn(xlog_t
*log
, xlog_in_core_t
*iclog
,
118 #define xlog_verify_dest_ptr(a,b)
119 #define xlog_verify_grant_head(a,b)
120 #define xlog_verify_iclog(a,b,c,d)
121 #define xlog_verify_tail_lsn(a,b,c)
124 STATIC
int xlog_iclogs_empty(xlog_t
*log
);
126 #if defined(XFS_LOG_TRACE)
128 xlog_trace_loggrant(xlog_t
*log
, xlog_ticket_t
*tic
, xfs_caddr_t string
)
132 if (!log
->l_grant_trace
) {
133 log
->l_grant_trace
= ktrace_alloc(2048, KM_NOSLEEP
);
134 if (!log
->l_grant_trace
)
137 /* ticket counts are 1 byte each */
138 cnts
= ((unsigned long)tic
->t_ocnt
) | ((unsigned long)tic
->t_cnt
) << 8;
140 ktrace_enter(log
->l_grant_trace
,
142 (void *)log
->l_reserve_headq
,
143 (void *)log
->l_write_headq
,
144 (void *)((unsigned long)log
->l_grant_reserve_cycle
),
145 (void *)((unsigned long)log
->l_grant_reserve_bytes
),
146 (void *)((unsigned long)log
->l_grant_write_cycle
),
147 (void *)((unsigned long)log
->l_grant_write_bytes
),
148 (void *)((unsigned long)log
->l_curr_cycle
),
149 (void *)((unsigned long)log
->l_curr_block
),
150 (void *)((unsigned long)CYCLE_LSN(log
->l_tail_lsn
)),
151 (void *)((unsigned long)BLOCK_LSN(log
->l_tail_lsn
)),
153 (void *)((unsigned long)tic
->t_trans_type
),
155 (void *)((unsigned long)tic
->t_curr_res
),
156 (void *)((unsigned long)tic
->t_unit_res
));
160 xlog_trace_iclog(xlog_in_core_t
*iclog
, uint state
)
162 if (!iclog
->ic_trace
)
163 iclog
->ic_trace
= ktrace_alloc(256, KM_SLEEP
);
164 ktrace_enter(iclog
->ic_trace
,
165 (void *)((unsigned long)state
),
166 (void *)((unsigned long)current_pid()),
167 (void *)NULL
, (void *)NULL
, (void *)NULL
, (void *)NULL
,
168 (void *)NULL
, (void *)NULL
, (void *)NULL
, (void *)NULL
,
169 (void *)NULL
, (void *)NULL
, (void *)NULL
, (void *)NULL
,
170 (void *)NULL
, (void *)NULL
);
173 #define xlog_trace_loggrant(log,tic,string)
174 #define xlog_trace_iclog(iclog,state)
175 #endif /* XFS_LOG_TRACE */
179 xlog_ins_ticketq(struct xlog_ticket
**qp
, struct xlog_ticket
*tic
)
183 tic
->t_prev
= (*qp
)->t_prev
;
184 (*qp
)->t_prev
->t_next
= tic
;
187 tic
->t_prev
= tic
->t_next
= tic
;
191 tic
->t_flags
|= XLOG_TIC_IN_Q
;
195 xlog_del_ticketq(struct xlog_ticket
**qp
, struct xlog_ticket
*tic
)
197 if (tic
== tic
->t_next
) {
201 tic
->t_next
->t_prev
= tic
->t_prev
;
202 tic
->t_prev
->t_next
= tic
->t_next
;
205 tic
->t_next
= tic
->t_prev
= NULL
;
206 tic
->t_flags
&= ~XLOG_TIC_IN_Q
;
210 xlog_grant_sub_space(struct log
*log
, int bytes
)
212 log
->l_grant_write_bytes
-= bytes
;
213 if (log
->l_grant_write_bytes
< 0) {
214 log
->l_grant_write_bytes
+= log
->l_logsize
;
215 log
->l_grant_write_cycle
--;
218 log
->l_grant_reserve_bytes
-= bytes
;
219 if ((log
)->l_grant_reserve_bytes
< 0) {
220 log
->l_grant_reserve_bytes
+= log
->l_logsize
;
221 log
->l_grant_reserve_cycle
--;
227 xlog_grant_add_space_write(struct log
*log
, int bytes
)
229 int tmp
= log
->l_logsize
- log
->l_grant_write_bytes
;
231 log
->l_grant_write_bytes
+= bytes
;
233 log
->l_grant_write_cycle
++;
234 log
->l_grant_write_bytes
= bytes
- tmp
;
239 xlog_grant_add_space_reserve(struct log
*log
, int bytes
)
241 int tmp
= log
->l_logsize
- log
->l_grant_reserve_bytes
;
243 log
->l_grant_reserve_bytes
+= bytes
;
245 log
->l_grant_reserve_cycle
++;
246 log
->l_grant_reserve_bytes
= bytes
- tmp
;
251 xlog_grant_add_space(struct log
*log
, int bytes
)
253 xlog_grant_add_space_write(log
, bytes
);
254 xlog_grant_add_space_reserve(log
, bytes
);
258 xlog_tic_reset_res(xlog_ticket_t
*tic
)
261 tic
->t_res_arr_sum
= 0;
262 tic
->t_res_num_ophdrs
= 0;
266 xlog_tic_add_region(xlog_ticket_t
*tic
, uint len
, uint type
)
268 if (tic
->t_res_num
== XLOG_TIC_LEN_MAX
) {
269 /* add to overflow and start again */
270 tic
->t_res_o_flow
+= tic
->t_res_arr_sum
;
272 tic
->t_res_arr_sum
= 0;
275 tic
->t_res_arr
[tic
->t_res_num
].r_len
= len
;
276 tic
->t_res_arr
[tic
->t_res_num
].r_type
= type
;
277 tic
->t_res_arr_sum
+= len
;
284 * 1. currblock field gets updated at startup and after in-core logs
285 * marked as with WANT_SYNC.
289 * This routine is called when a user of a log manager ticket is done with
290 * the reservation. If the ticket was ever used, then a commit record for
291 * the associated transaction is written out as a log operation header with
292 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
293 * a given ticket. If the ticket was one with a permanent reservation, then
294 * a few operations are done differently. Permanent reservation tickets by
295 * default don't release the reservation. They just commit the current
296 * transaction with the belief that the reservation is still needed. A flag
297 * must be passed in before permanent reservations are actually released.
298 * When these type of tickets are not released, they need to be set into
299 * the inited state again. By doing this, a start record will be written
300 * out when the next write occurs.
303 xfs_log_done(xfs_mount_t
*mp
,
304 xfs_log_ticket_t xtic
,
308 xlog_t
*log
= mp
->m_log
;
309 xlog_ticket_t
*ticket
= (xfs_log_ticket_t
) xtic
;
312 if (XLOG_FORCED_SHUTDOWN(log
) ||
314 * If nothing was ever written, don't write out commit record.
315 * If we get an error, just continue and give back the log ticket.
317 (((ticket
->t_flags
& XLOG_TIC_INITED
) == 0) &&
318 (xlog_commit_record(mp
, ticket
,
319 (xlog_in_core_t
**)iclog
, &lsn
)))) {
320 lsn
= (xfs_lsn_t
) -1;
321 if (ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) {
322 flags
|= XFS_LOG_REL_PERM_RESERV
;
327 if ((ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) == 0 ||
328 (flags
& XFS_LOG_REL_PERM_RESERV
)) {
330 * Release ticket if not permanent reservation or a specific
331 * request has been made to release a permanent reservation.
333 xlog_trace_loggrant(log
, ticket
, "xfs_log_done: (non-permanent)");
334 xlog_ungrant_log_space(log
, ticket
);
335 xlog_ticket_put(log
, ticket
);
337 xlog_trace_loggrant(log
, ticket
, "xfs_log_done: (permanent)");
338 xlog_regrant_reserve_log_space(log
, ticket
);
341 /* If this ticket was a permanent reservation and we aren't
342 * trying to release it, reset the inited flags; so next time
343 * we write, a start record will be written out.
345 if ((ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) &&
346 (flags
& XFS_LOG_REL_PERM_RESERV
) == 0)
347 ticket
->t_flags
|= XLOG_TIC_INITED
;
354 * Force the in-core log to disk. If flags == XFS_LOG_SYNC,
355 * the force is done synchronously.
357 * Asynchronous forces are implemented by setting the WANT_SYNC
358 * bit in the appropriate in-core log and then returning.
360 * Synchronous forces are implemented with a semaphore. All callers
361 * to force a given lsn to disk will wait on a semaphore attached to the
362 * specific in-core log. When given in-core log finally completes its
363 * write to disk, that thread will wake up all threads waiting on the
373 xlog_t
*log
= mp
->m_log
;
377 log_flushed
= &dummy
;
379 ASSERT(flags
& XFS_LOG_FORCE
);
381 XFS_STATS_INC(xs_log_force
);
383 if (log
->l_flags
& XLOG_IO_ERROR
)
384 return XFS_ERROR(EIO
);
386 return xlog_state_sync_all(log
, flags
, log_flushed
);
388 return xlog_state_sync(log
, lsn
, flags
, log_flushed
);
389 } /* _xfs_log_force */
392 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
393 * about errors or whether the log was flushed or not. This is the normal
394 * interface to use when trying to unpin items or move the log forward.
403 error
= _xfs_log_force(mp
, lsn
, flags
, NULL
);
405 xfs_fs_cmn_err(CE_WARN
, mp
, "xfs_log_force: "
406 "error %d returned.", error
);
412 * Attaches a new iclog I/O completion callback routine during
413 * transaction commit. If the log is in error state, a non-zero
414 * return code is handed back and the caller is responsible for
415 * executing the callback at an appropriate time.
418 xfs_log_notify(xfs_mount_t
*mp
, /* mount of partition */
419 void *iclog_hndl
, /* iclog to hang callback off */
420 xfs_log_callback_t
*cb
)
422 xlog_in_core_t
*iclog
= (xlog_in_core_t
*)iclog_hndl
;
425 spin_lock(&iclog
->ic_callback_lock
);
426 abortflg
= (iclog
->ic_state
& XLOG_STATE_IOERROR
);
428 ASSERT_ALWAYS((iclog
->ic_state
== XLOG_STATE_ACTIVE
) ||
429 (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
));
431 *(iclog
->ic_callback_tail
) = cb
;
432 iclog
->ic_callback_tail
= &(cb
->cb_next
);
434 spin_unlock(&iclog
->ic_callback_lock
);
436 } /* xfs_log_notify */
439 xfs_log_release_iclog(xfs_mount_t
*mp
,
442 xlog_t
*log
= mp
->m_log
;
443 xlog_in_core_t
*iclog
= (xlog_in_core_t
*)iclog_hndl
;
445 if (xlog_state_release_iclog(log
, iclog
)) {
446 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
454 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
455 * to the reservation.
456 * 2. Potentially, push buffers at tail of log to disk.
458 * Each reservation is going to reserve extra space for a log record header.
459 * When writes happen to the on-disk log, we don't subtract the length of the
460 * log record header from any reservation. By wasting space in each
461 * reservation, we prevent over allocation problems.
464 xfs_log_reserve(xfs_mount_t
*mp
,
467 xfs_log_ticket_t
*ticket
,
472 xlog_t
*log
= mp
->m_log
;
473 xlog_ticket_t
*internal_ticket
;
476 ASSERT(client
== XFS_TRANSACTION
|| client
== XFS_LOG
);
477 ASSERT((flags
& XFS_LOG_NOSLEEP
) == 0);
479 if (XLOG_FORCED_SHUTDOWN(log
))
480 return XFS_ERROR(EIO
);
482 XFS_STATS_INC(xs_try_logspace
);
484 if (*ticket
!= NULL
) {
485 ASSERT(flags
& XFS_LOG_PERM_RESERV
);
486 internal_ticket
= (xlog_ticket_t
*)*ticket
;
487 xlog_trace_loggrant(log
, internal_ticket
, "xfs_log_reserve: existing ticket (permanent trans)");
488 xlog_grant_push_ail(mp
, internal_ticket
->t_unit_res
);
489 retval
= xlog_regrant_write_log_space(log
, internal_ticket
);
491 /* may sleep if need to allocate more tickets */
492 internal_ticket
= xlog_ticket_get(log
, unit_bytes
, cnt
,
494 if (!internal_ticket
)
495 return XFS_ERROR(ENOMEM
);
496 internal_ticket
->t_trans_type
= t_type
;
497 *ticket
= internal_ticket
;
498 xlog_trace_loggrant(log
, internal_ticket
,
499 (internal_ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) ?
500 "xfs_log_reserve: create new ticket (permanent trans)" :
501 "xfs_log_reserve: create new ticket");
502 xlog_grant_push_ail(mp
,
503 (internal_ticket
->t_unit_res
*
504 internal_ticket
->t_cnt
));
505 retval
= xlog_grant_log_space(log
, internal_ticket
);
509 } /* xfs_log_reserve */
513 * Mount a log filesystem
515 * mp - ubiquitous xfs mount point structure
516 * log_target - buftarg of on-disk log device
517 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
518 * num_bblocks - Number of BBSIZE blocks in on-disk log
520 * Return error or zero.
525 xfs_buftarg_t
*log_target
,
526 xfs_daddr_t blk_offset
,
531 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
))
532 cmn_err(CE_NOTE
, "XFS mounting filesystem %s", mp
->m_fsname
);
535 "!Mounting filesystem \"%s\" in no-recovery mode. Filesystem will be inconsistent.",
537 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
540 mp
->m_log
= xlog_alloc_log(mp
, log_target
, blk_offset
, num_bblks
);
543 * Initialize the AIL now we have a log.
545 spin_lock_init(&mp
->m_ail_lock
);
546 error
= xfs_trans_ail_init(mp
);
548 cmn_err(CE_WARN
, "XFS: AIL initialisation failed: error %d", error
);
553 * skip log recovery on a norecovery mount. pretend it all
556 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
)) {
557 int readonly
= (mp
->m_flags
& XFS_MOUNT_RDONLY
);
560 mp
->m_flags
&= ~XFS_MOUNT_RDONLY
;
562 error
= xlog_recover(mp
->m_log
);
565 mp
->m_flags
|= XFS_MOUNT_RDONLY
;
567 cmn_err(CE_WARN
, "XFS: log mount/recovery failed: error %d", error
);
572 /* Normal transactions can now occur */
573 mp
->m_log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
575 /* End mounting message in xfs_log_mount_finish */
578 xfs_log_unmount_dealloc(mp
);
580 } /* xfs_log_mount */
583 * Finish the recovery of the file system. This is separate from
584 * the xfs_log_mount() call, because it depends on the code in
585 * xfs_mountfs() to read in the root and real-time bitmap inodes
586 * between calling xfs_log_mount() and here.
588 * mp - ubiquitous xfs mount point structure
591 xfs_log_mount_finish(xfs_mount_t
*mp
, int mfsi_flags
)
595 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
))
596 error
= xlog_recover_finish(mp
->m_log
, mfsi_flags
);
599 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
606 * Unmount processing for the log.
609 xfs_log_unmount(xfs_mount_t
*mp
)
613 error
= xfs_log_unmount_write(mp
);
614 xfs_log_unmount_dealloc(mp
);
619 * Final log writes as part of unmount.
621 * Mark the filesystem clean as unmount happens. Note that during relocation
622 * this routine needs to be executed as part of source-bag while the
623 * deallocation must not be done until source-end.
627 * Unmount record used to have a string "Unmount filesystem--" in the
628 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
629 * We just write the magic number now since that particular field isn't
630 * currently architecture converted and "nUmount" is a bit foo.
631 * As far as I know, there weren't any dependencies on the old behaviour.
635 xfs_log_unmount_write(xfs_mount_t
*mp
)
637 xlog_t
*log
= mp
->m_log
;
638 xlog_in_core_t
*iclog
;
640 xlog_in_core_t
*first_iclog
;
642 xfs_log_iovec_t reg
[1];
643 xfs_log_ticket_t tic
= NULL
;
647 /* the data section must be 32 bit size aligned */
651 __uint32_t pad2
; /* may as well make it 64 bits */
652 } magic
= { XLOG_UNMOUNT_TYPE
, 0, 0 };
655 * Don't write out unmount record on read-only mounts.
656 * Or, if we are doing a forced umount (typically because of IO errors).
658 if (mp
->m_flags
& XFS_MOUNT_RDONLY
)
661 error
= _xfs_log_force(mp
, 0, XFS_LOG_FORCE
|XFS_LOG_SYNC
, NULL
);
662 ASSERT(error
|| !(XLOG_FORCED_SHUTDOWN(log
)));
665 first_iclog
= iclog
= log
->l_iclog
;
667 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
668 ASSERT(iclog
->ic_state
& XLOG_STATE_ACTIVE
);
669 ASSERT(iclog
->ic_offset
== 0);
671 iclog
= iclog
->ic_next
;
672 } while (iclog
!= first_iclog
);
674 if (! (XLOG_FORCED_SHUTDOWN(log
))) {
675 reg
[0].i_addr
= (void*)&magic
;
676 reg
[0].i_len
= sizeof(magic
);
677 XLOG_VEC_SET_TYPE(®
[0], XLOG_REG_TYPE_UNMOUNT
);
679 error
= xfs_log_reserve(mp
, 600, 1, &tic
,
680 XFS_LOG
, 0, XLOG_UNMOUNT_REC_TYPE
);
682 /* remove inited flag */
683 ((xlog_ticket_t
*)tic
)->t_flags
= 0;
684 error
= xlog_write(mp
, reg
, 1, tic
, &lsn
,
685 NULL
, XLOG_UNMOUNT_TRANS
);
687 * At this point, we're umounting anyway,
688 * so there's no point in transitioning log state
689 * to IOERROR. Just continue...
694 xfs_fs_cmn_err(CE_ALERT
, mp
,
695 "xfs_log_unmount: unmount record failed");
699 spin_lock(&log
->l_icloglock
);
700 iclog
= log
->l_iclog
;
701 atomic_inc(&iclog
->ic_refcnt
);
702 spin_unlock(&log
->l_icloglock
);
703 xlog_state_want_sync(log
, iclog
);
704 error
= xlog_state_release_iclog(log
, iclog
);
706 spin_lock(&log
->l_icloglock
);
707 if (!(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
708 iclog
->ic_state
== XLOG_STATE_DIRTY
)) {
709 if (!XLOG_FORCED_SHUTDOWN(log
)) {
710 sv_wait(&iclog
->ic_forcesema
, PMEM
,
711 &log
->l_icloglock
, s
);
713 spin_unlock(&log
->l_icloglock
);
716 spin_unlock(&log
->l_icloglock
);
719 xlog_trace_loggrant(log
, tic
, "unmount rec");
720 xlog_ungrant_log_space(log
, tic
);
721 xlog_ticket_put(log
, tic
);
725 * We're already in forced_shutdown mode, couldn't
726 * even attempt to write out the unmount transaction.
728 * Go through the motions of sync'ing and releasing
729 * the iclog, even though no I/O will actually happen,
730 * we need to wait for other log I/Os that may already
731 * be in progress. Do this as a separate section of
732 * code so we'll know if we ever get stuck here that
733 * we're in this odd situation of trying to unmount
734 * a file system that went into forced_shutdown as
735 * the result of an unmount..
737 spin_lock(&log
->l_icloglock
);
738 iclog
= log
->l_iclog
;
739 atomic_inc(&iclog
->ic_refcnt
);
740 spin_unlock(&log
->l_icloglock
);
742 xlog_state_want_sync(log
, iclog
);
743 error
= xlog_state_release_iclog(log
, iclog
);
745 spin_lock(&log
->l_icloglock
);
747 if ( ! ( iclog
->ic_state
== XLOG_STATE_ACTIVE
748 || iclog
->ic_state
== XLOG_STATE_DIRTY
749 || iclog
->ic_state
== XLOG_STATE_IOERROR
) ) {
751 sv_wait(&iclog
->ic_forcesema
, PMEM
,
752 &log
->l_icloglock
, s
);
754 spin_unlock(&log
->l_icloglock
);
759 } /* xfs_log_unmount_write */
762 * Deallocate log structures for unmount/relocation.
764 * We need to stop the aild from running before we destroy
765 * and deallocate the log as the aild references the log.
768 xfs_log_unmount_dealloc(xfs_mount_t
*mp
)
770 xfs_trans_ail_destroy(mp
);
771 xlog_dealloc_log(mp
->m_log
);
775 * Write region vectors to log. The write happens using the space reservation
776 * of the ticket (tic). It is not a requirement that all writes for a given
777 * transaction occur with one call to xfs_log_write().
780 xfs_log_write(xfs_mount_t
* mp
,
781 xfs_log_iovec_t reg
[],
783 xfs_log_ticket_t tic
,
784 xfs_lsn_t
*start_lsn
)
787 xlog_t
*log
= mp
->m_log
;
789 if (XLOG_FORCED_SHUTDOWN(log
))
790 return XFS_ERROR(EIO
);
792 if ((error
= xlog_write(mp
, reg
, nentries
, tic
, start_lsn
, NULL
, 0))) {
793 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
796 } /* xfs_log_write */
800 xfs_log_move_tail(xfs_mount_t
*mp
,
804 xlog_t
*log
= mp
->m_log
;
805 int need_bytes
, free_bytes
, cycle
, bytes
;
807 if (XLOG_FORCED_SHUTDOWN(log
))
811 /* needed since sync_lsn is 64 bits */
812 spin_lock(&log
->l_icloglock
);
813 tail_lsn
= log
->l_last_sync_lsn
;
814 spin_unlock(&log
->l_icloglock
);
817 spin_lock(&log
->l_grant_lock
);
819 /* Also an invalid lsn. 1 implies that we aren't passing in a valid
823 log
->l_tail_lsn
= tail_lsn
;
826 if ((tic
= log
->l_write_headq
)) {
828 if (log
->l_flags
& XLOG_ACTIVE_RECOVERY
)
829 panic("Recovery problem");
831 cycle
= log
->l_grant_write_cycle
;
832 bytes
= log
->l_grant_write_bytes
;
833 free_bytes
= xlog_space_left(log
, cycle
, bytes
);
835 ASSERT(tic
->t_flags
& XLOG_TIC_PERM_RESERV
);
837 if (free_bytes
< tic
->t_unit_res
&& tail_lsn
!= 1)
840 free_bytes
-= tic
->t_unit_res
;
841 sv_signal(&tic
->t_sema
);
843 } while (tic
!= log
->l_write_headq
);
845 if ((tic
= log
->l_reserve_headq
)) {
847 if (log
->l_flags
& XLOG_ACTIVE_RECOVERY
)
848 panic("Recovery problem");
850 cycle
= log
->l_grant_reserve_cycle
;
851 bytes
= log
->l_grant_reserve_bytes
;
852 free_bytes
= xlog_space_left(log
, cycle
, bytes
);
854 if (tic
->t_flags
& XLOG_TIC_PERM_RESERV
)
855 need_bytes
= tic
->t_unit_res
*tic
->t_cnt
;
857 need_bytes
= tic
->t_unit_res
;
858 if (free_bytes
< need_bytes
&& tail_lsn
!= 1)
861 free_bytes
-= need_bytes
;
862 sv_signal(&tic
->t_sema
);
864 } while (tic
!= log
->l_reserve_headq
);
866 spin_unlock(&log
->l_grant_lock
);
867 } /* xfs_log_move_tail */
870 * Determine if we have a transaction that has gone to disk
871 * that needs to be covered. Log activity needs to be idle (no AIL and
872 * nothing in the iclogs). And, we need to be in the right state indicating
873 * something has gone out.
876 xfs_log_need_covered(xfs_mount_t
*mp
)
879 xlog_t
*log
= mp
->m_log
;
881 if (!xfs_fs_writable(mp
))
884 spin_lock(&log
->l_icloglock
);
885 if (((log
->l_covered_state
== XLOG_STATE_COVER_NEED
) ||
886 (log
->l_covered_state
== XLOG_STATE_COVER_NEED2
))
887 && !xfs_trans_first_ail(mp
, &gen
)
888 && xlog_iclogs_empty(log
)) {
889 if (log
->l_covered_state
== XLOG_STATE_COVER_NEED
)
890 log
->l_covered_state
= XLOG_STATE_COVER_DONE
;
892 ASSERT(log
->l_covered_state
== XLOG_STATE_COVER_NEED2
);
893 log
->l_covered_state
= XLOG_STATE_COVER_DONE2
;
897 spin_unlock(&log
->l_icloglock
);
901 /******************************************************************************
905 ******************************************************************************
908 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
909 * The log manager must keep track of the last LR which was committed
910 * to disk. The lsn of this LR will become the new tail_lsn whenever
911 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
912 * the situation where stuff could be written into the log but nothing
913 * was ever in the AIL when asked. Eventually, we panic since the
914 * tail hits the head.
916 * We may be holding the log iclog lock upon entering this routine.
919 xlog_assign_tail_lsn(xfs_mount_t
*mp
)
922 xlog_t
*log
= mp
->m_log
;
924 tail_lsn
= xfs_trans_tail_ail(mp
);
925 spin_lock(&log
->l_grant_lock
);
927 log
->l_tail_lsn
= tail_lsn
;
929 tail_lsn
= log
->l_tail_lsn
= log
->l_last_sync_lsn
;
931 spin_unlock(&log
->l_grant_lock
);
934 } /* xlog_assign_tail_lsn */
938 * Return the space in the log between the tail and the head. The head
939 * is passed in the cycle/bytes formal parms. In the special case where
940 * the reserve head has wrapped passed the tail, this calculation is no
941 * longer valid. In this case, just return 0 which means there is no space
942 * in the log. This works for all places where this function is called
943 * with the reserve head. Of course, if the write head were to ever
944 * wrap the tail, we should blow up. Rather than catch this case here,
945 * we depend on other ASSERTions in other parts of the code. XXXmiken
947 * This code also handles the case where the reservation head is behind
948 * the tail. The details of this case are described below, but the end
949 * result is that we return the size of the log as the amount of space left.
952 xlog_space_left(xlog_t
*log
, int cycle
, int bytes
)
958 tail_bytes
= BBTOB(BLOCK_LSN(log
->l_tail_lsn
));
959 tail_cycle
= CYCLE_LSN(log
->l_tail_lsn
);
960 if ((tail_cycle
== cycle
) && (bytes
>= tail_bytes
)) {
961 free_bytes
= log
->l_logsize
- (bytes
- tail_bytes
);
962 } else if ((tail_cycle
+ 1) < cycle
) {
964 } else if (tail_cycle
< cycle
) {
965 ASSERT(tail_cycle
== (cycle
- 1));
966 free_bytes
= tail_bytes
- bytes
;
969 * The reservation head is behind the tail.
970 * In this case we just want to return the size of the
971 * log as the amount of space left.
973 xfs_fs_cmn_err(CE_ALERT
, log
->l_mp
,
974 "xlog_space_left: head behind tail\n"
975 " tail_cycle = %d, tail_bytes = %d\n"
976 " GH cycle = %d, GH bytes = %d",
977 tail_cycle
, tail_bytes
, cycle
, bytes
);
979 free_bytes
= log
->l_logsize
;
982 } /* xlog_space_left */
986 * Log function which is called when an io completes.
988 * The log manager needs its own routine, in order to control what
989 * happens with the buffer after the write completes.
992 xlog_iodone(xfs_buf_t
*bp
)
994 xlog_in_core_t
*iclog
;
998 iclog
= XFS_BUF_FSPRIVATE(bp
, xlog_in_core_t
*);
999 ASSERT(XFS_BUF_FSPRIVATE2(bp
, unsigned long) == (unsigned long) 2);
1000 XFS_BUF_SET_FSPRIVATE2(bp
, (unsigned long)1);
1004 * Some versions of cpp barf on the recursive definition of
1005 * ic_log -> hic_fields.ic_log and expand ic_log twice when
1006 * it is passed through two macros. Workaround broken cpp.
1011 * If the ordered flag has been removed by a lower
1012 * layer, it means the underlyin device no longer supports
1013 * barrier I/O. Warn loudly and turn off barriers.
1015 if ((l
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
) && !XFS_BUF_ORDERED(bp
)) {
1016 l
->l_mp
->m_flags
&= ~XFS_MOUNT_BARRIER
;
1017 xfs_fs_cmn_err(CE_WARN
, l
->l_mp
,
1018 "xlog_iodone: Barriers are no longer supported"
1019 " by device. Disabling barriers\n");
1020 xfs_buftrace("XLOG_IODONE BARRIERS OFF", bp
);
1024 * Race to shutdown the filesystem if we see an error.
1026 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp
)), l
->l_mp
,
1027 XFS_ERRTAG_IODONE_IOERR
, XFS_RANDOM_IODONE_IOERR
)) {
1028 xfs_ioerror_alert("xlog_iodone", l
->l_mp
, bp
, XFS_BUF_ADDR(bp
));
1030 xfs_force_shutdown(l
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1032 * This flag will be propagated to the trans-committed
1033 * callback routines to let them know that the log-commit
1036 aborted
= XFS_LI_ABORTED
;
1037 } else if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1038 aborted
= XFS_LI_ABORTED
;
1041 /* log I/O is always issued ASYNC */
1042 ASSERT(XFS_BUF_ISASYNC(bp
));
1043 xlog_state_done_syncing(iclog
, aborted
);
1045 * do not reference the buffer (bp) here as we could race
1046 * with it being freed after writing the unmount record to the
1053 * The bdstrat callback function for log bufs. This gives us a central
1054 * place to trap bufs in case we get hit by a log I/O error and need to
1055 * shutdown. Actually, in practice, even when we didn't get a log error,
1056 * we transition the iclogs to IOERROR state *after* flushing all existing
1057 * iclogs to disk. This is because we don't want anymore new transactions to be
1058 * started or completed afterwards.
1061 xlog_bdstrat_cb(struct xfs_buf
*bp
)
1063 xlog_in_core_t
*iclog
;
1065 iclog
= XFS_BUF_FSPRIVATE(bp
, xlog_in_core_t
*);
1067 if ((iclog
->ic_state
& XLOG_STATE_IOERROR
) == 0) {
1068 /* note for irix bstrat will need struct bdevsw passed
1069 * Fix the following macro if the code ever is merged
1075 xfs_buftrace("XLOG__BDSTRAT IOERROR", bp
);
1076 XFS_BUF_ERROR(bp
, EIO
);
1079 return XFS_ERROR(EIO
);
1085 * Return size of each in-core log record buffer.
1087 * All machines get 8 x 32KB buffers by default, unless tuned otherwise.
1089 * If the filesystem blocksize is too large, we may need to choose a
1090 * larger size since the directory code currently logs entire blocks.
1094 xlog_get_iclog_buffer_size(xfs_mount_t
*mp
,
1100 if (mp
->m_logbufs
<= 0)
1101 log
->l_iclog_bufs
= XLOG_MAX_ICLOGS
;
1103 log
->l_iclog_bufs
= mp
->m_logbufs
;
1106 * Buffer size passed in from mount system call.
1108 if (mp
->m_logbsize
> 0) {
1109 size
= log
->l_iclog_size
= mp
->m_logbsize
;
1110 log
->l_iclog_size_log
= 0;
1112 log
->l_iclog_size_log
++;
1116 if (xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1117 /* # headers = size / 32K
1118 * one header holds cycles from 32K of data
1121 xhdrs
= mp
->m_logbsize
/ XLOG_HEADER_CYCLE_SIZE
;
1122 if (mp
->m_logbsize
% XLOG_HEADER_CYCLE_SIZE
)
1124 log
->l_iclog_hsize
= xhdrs
<< BBSHIFT
;
1125 log
->l_iclog_heads
= xhdrs
;
1127 ASSERT(mp
->m_logbsize
<= XLOG_BIG_RECORD_BSIZE
);
1128 log
->l_iclog_hsize
= BBSIZE
;
1129 log
->l_iclog_heads
= 1;
1134 /* All machines use 32KB buffers by default. */
1135 log
->l_iclog_size
= XLOG_BIG_RECORD_BSIZE
;
1136 log
->l_iclog_size_log
= XLOG_BIG_RECORD_BSHIFT
;
1138 /* the default log size is 16k or 32k which is one header sector */
1139 log
->l_iclog_hsize
= BBSIZE
;
1140 log
->l_iclog_heads
= 1;
1143 * For 16KB, we use 3 32KB buffers. For 32KB block sizes, we use
1144 * 4 32KB buffers. For 64KB block sizes, we use 8 32KB buffers.
1146 if (mp
->m_sb
.sb_blocksize
>= 16*1024) {
1147 log
->l_iclog_size
= XLOG_BIG_RECORD_BSIZE
;
1148 log
->l_iclog_size_log
= XLOG_BIG_RECORD_BSHIFT
;
1149 if (mp
->m_logbufs
<= 0) {
1150 switch (mp
->m_sb
.sb_blocksize
) {
1151 case 16*1024: /* 16 KB */
1152 log
->l_iclog_bufs
= 3;
1154 case 32*1024: /* 32 KB */
1155 log
->l_iclog_bufs
= 4;
1157 case 64*1024: /* 64 KB */
1158 log
->l_iclog_bufs
= 8;
1161 xlog_panic("XFS: Invalid blocksize");
1167 done
: /* are we being asked to make the sizes selected above visible? */
1168 if (mp
->m_logbufs
== 0)
1169 mp
->m_logbufs
= log
->l_iclog_bufs
;
1170 if (mp
->m_logbsize
== 0)
1171 mp
->m_logbsize
= log
->l_iclog_size
;
1172 } /* xlog_get_iclog_buffer_size */
1176 * This routine initializes some of the log structure for a given mount point.
1177 * Its primary purpose is to fill in enough, so recovery can occur. However,
1178 * some other stuff may be filled in too.
1181 xlog_alloc_log(xfs_mount_t
*mp
,
1182 xfs_buftarg_t
*log_target
,
1183 xfs_daddr_t blk_offset
,
1187 xlog_rec_header_t
*head
;
1188 xlog_in_core_t
**iclogp
;
1189 xlog_in_core_t
*iclog
, *prev_iclog
=NULL
;
1194 log
= (xlog_t
*)kmem_zalloc(sizeof(xlog_t
), KM_SLEEP
);
1197 log
->l_targ
= log_target
;
1198 log
->l_logsize
= BBTOB(num_bblks
);
1199 log
->l_logBBstart
= blk_offset
;
1200 log
->l_logBBsize
= num_bblks
;
1201 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
1202 log
->l_flags
|= XLOG_ACTIVE_RECOVERY
;
1204 log
->l_prev_block
= -1;
1205 log
->l_tail_lsn
= xlog_assign_lsn(1, 0);
1206 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1207 log
->l_last_sync_lsn
= log
->l_tail_lsn
;
1208 log
->l_curr_cycle
= 1; /* 0 is bad since this is initial value */
1209 log
->l_grant_reserve_cycle
= 1;
1210 log
->l_grant_write_cycle
= 1;
1212 if (xfs_sb_version_hassector(&mp
->m_sb
)) {
1213 log
->l_sectbb_log
= mp
->m_sb
.sb_logsectlog
- BBSHIFT
;
1214 ASSERT(log
->l_sectbb_log
<= mp
->m_sectbb_log
);
1215 /* for larger sector sizes, must have v2 or external log */
1216 ASSERT(log
->l_sectbb_log
== 0 ||
1217 log
->l_logBBstart
== 0 ||
1218 xfs_sb_version_haslogv2(&mp
->m_sb
));
1219 ASSERT(mp
->m_sb
.sb_logsectlog
>= BBSHIFT
);
1221 log
->l_sectbb_mask
= (1 << log
->l_sectbb_log
) - 1;
1223 xlog_get_iclog_buffer_size(mp
, log
);
1225 bp
= xfs_buf_get_empty(log
->l_iclog_size
, mp
->m_logdev_targp
);
1226 XFS_BUF_SET_IODONE_FUNC(bp
, xlog_iodone
);
1227 XFS_BUF_SET_BDSTRAT_FUNC(bp
, xlog_bdstrat_cb
);
1228 XFS_BUF_SET_FSPRIVATE2(bp
, (unsigned long)1);
1229 ASSERT(XFS_BUF_ISBUSY(bp
));
1230 ASSERT(XFS_BUF_VALUSEMA(bp
) <= 0);
1233 spin_lock_init(&log
->l_icloglock
);
1234 spin_lock_init(&log
->l_grant_lock
);
1235 sv_init(&log
->l_flush_wait
, 0, "flush_wait");
1237 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1238 ASSERT((XFS_BUF_SIZE(bp
) & BBMASK
) == 0);
1240 iclogp
= &log
->l_iclog
;
1242 * The amount of memory to allocate for the iclog structure is
1243 * rather funky due to the way the structure is defined. It is
1244 * done this way so that we can use different sizes for machines
1245 * with different amounts of memory. See the definition of
1246 * xlog_in_core_t in xfs_log_priv.h for details.
1248 iclogsize
= log
->l_iclog_size
;
1249 ASSERT(log
->l_iclog_size
>= 4096);
1250 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
1251 *iclogp
= (xlog_in_core_t
*)
1252 kmem_zalloc(sizeof(xlog_in_core_t
), KM_SLEEP
);
1254 iclog
->ic_prev
= prev_iclog
;
1257 bp
= xfs_buf_get_noaddr(log
->l_iclog_size
, mp
->m_logdev_targp
);
1258 if (!XFS_BUF_CPSEMA(bp
))
1260 XFS_BUF_SET_IODONE_FUNC(bp
, xlog_iodone
);
1261 XFS_BUF_SET_BDSTRAT_FUNC(bp
, xlog_bdstrat_cb
);
1262 XFS_BUF_SET_FSPRIVATE2(bp
, (unsigned long)1);
1264 iclog
->hic_data
= bp
->b_addr
;
1266 log
->l_iclog_bak
[i
] = (xfs_caddr_t
)&(iclog
->ic_header
);
1268 head
= &iclog
->ic_header
;
1269 memset(head
, 0, sizeof(xlog_rec_header_t
));
1270 head
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1271 head
->h_version
= cpu_to_be32(
1272 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1273 head
->h_size
= cpu_to_be32(log
->l_iclog_size
);
1275 head
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1276 memcpy(&head
->h_fs_uuid
, &mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1278 iclog
->ic_size
= XFS_BUF_SIZE(bp
) - log
->l_iclog_hsize
;
1279 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1280 iclog
->ic_log
= log
;
1281 atomic_set(&iclog
->ic_refcnt
, 0);
1282 spin_lock_init(&iclog
->ic_callback_lock
);
1283 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
1284 iclog
->ic_datap
= (char *)iclog
->hic_data
+ log
->l_iclog_hsize
;
1286 ASSERT(XFS_BUF_ISBUSY(iclog
->ic_bp
));
1287 ASSERT(XFS_BUF_VALUSEMA(iclog
->ic_bp
) <= 0);
1288 sv_init(&iclog
->ic_forcesema
, SV_DEFAULT
, "iclog-force");
1289 sv_init(&iclog
->ic_writesema
, SV_DEFAULT
, "iclog-write");
1291 iclogp
= &iclog
->ic_next
;
1293 *iclogp
= log
->l_iclog
; /* complete ring */
1294 log
->l_iclog
->ic_prev
= prev_iclog
; /* re-write 1st prev ptr */
1297 } /* xlog_alloc_log */
1301 * Write out the commit record of a transaction associated with the given
1302 * ticket. Return the lsn of the commit record.
1305 xlog_commit_record(xfs_mount_t
*mp
,
1306 xlog_ticket_t
*ticket
,
1307 xlog_in_core_t
**iclog
,
1308 xfs_lsn_t
*commitlsnp
)
1311 xfs_log_iovec_t reg
[1];
1313 reg
[0].i_addr
= NULL
;
1315 XLOG_VEC_SET_TYPE(®
[0], XLOG_REG_TYPE_COMMIT
);
1317 ASSERT_ALWAYS(iclog
);
1318 if ((error
= xlog_write(mp
, reg
, 1, ticket
, commitlsnp
,
1319 iclog
, XLOG_COMMIT_TRANS
))) {
1320 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
1323 } /* xlog_commit_record */
1327 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1328 * log space. This code pushes on the lsn which would supposedly free up
1329 * the 25% which we want to leave free. We may need to adopt a policy which
1330 * pushes on an lsn which is further along in the log once we reach the high
1331 * water mark. In this manner, we would be creating a low water mark.
1334 xlog_grant_push_ail(xfs_mount_t
*mp
,
1337 xlog_t
*log
= mp
->m_log
; /* pointer to the log */
1338 xfs_lsn_t tail_lsn
; /* lsn of the log tail */
1339 xfs_lsn_t threshold_lsn
= 0; /* lsn we'd like to be at */
1340 int free_blocks
; /* free blocks left to write to */
1341 int free_bytes
; /* free bytes left to write to */
1342 int threshold_block
; /* block in lsn we'd like to be at */
1343 int threshold_cycle
; /* lsn cycle we'd like to be at */
1346 ASSERT(BTOBB(need_bytes
) < log
->l_logBBsize
);
1348 spin_lock(&log
->l_grant_lock
);
1349 free_bytes
= xlog_space_left(log
,
1350 log
->l_grant_reserve_cycle
,
1351 log
->l_grant_reserve_bytes
);
1352 tail_lsn
= log
->l_tail_lsn
;
1353 free_blocks
= BTOBBT(free_bytes
);
1356 * Set the threshold for the minimum number of free blocks in the
1357 * log to the maximum of what the caller needs, one quarter of the
1358 * log, and 256 blocks.
1360 free_threshold
= BTOBB(need_bytes
);
1361 free_threshold
= MAX(free_threshold
, (log
->l_logBBsize
>> 2));
1362 free_threshold
= MAX(free_threshold
, 256);
1363 if (free_blocks
< free_threshold
) {
1364 threshold_block
= BLOCK_LSN(tail_lsn
) + free_threshold
;
1365 threshold_cycle
= CYCLE_LSN(tail_lsn
);
1366 if (threshold_block
>= log
->l_logBBsize
) {
1367 threshold_block
-= log
->l_logBBsize
;
1368 threshold_cycle
+= 1;
1370 threshold_lsn
= xlog_assign_lsn(threshold_cycle
, threshold_block
);
1372 /* Don't pass in an lsn greater than the lsn of the last
1373 * log record known to be on disk.
1375 if (XFS_LSN_CMP(threshold_lsn
, log
->l_last_sync_lsn
) > 0)
1376 threshold_lsn
= log
->l_last_sync_lsn
;
1378 spin_unlock(&log
->l_grant_lock
);
1381 * Get the transaction layer to kick the dirty buffers out to
1382 * disk asynchronously. No point in trying to do this if
1383 * the filesystem is shutting down.
1385 if (threshold_lsn
&&
1386 !XLOG_FORCED_SHUTDOWN(log
))
1387 xfs_trans_push_ail(mp
, threshold_lsn
);
1388 } /* xlog_grant_push_ail */
1392 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1393 * fashion. Previously, we should have moved the current iclog
1394 * ptr in the log to point to the next available iclog. This allows further
1395 * write to continue while this code syncs out an iclog ready to go.
1396 * Before an in-core log can be written out, the data section must be scanned
1397 * to save away the 1st word of each BBSIZE block into the header. We replace
1398 * it with the current cycle count. Each BBSIZE block is tagged with the
1399 * cycle count because there in an implicit assumption that drives will
1400 * guarantee that entire 512 byte blocks get written at once. In other words,
1401 * we can't have part of a 512 byte block written and part not written. By
1402 * tagging each block, we will know which blocks are valid when recovering
1403 * after an unclean shutdown.
1405 * This routine is single threaded on the iclog. No other thread can be in
1406 * this routine with the same iclog. Changing contents of iclog can there-
1407 * fore be done without grabbing the state machine lock. Updating the global
1408 * log will require grabbing the lock though.
1410 * The entire log manager uses a logical block numbering scheme. Only
1411 * log_sync (and then only bwrite()) know about the fact that the log may
1412 * not start with block zero on a given device. The log block start offset
1413 * is added immediately before calling bwrite().
1417 xlog_sync(xlog_t
*log
,
1418 xlog_in_core_t
*iclog
)
1420 xfs_caddr_t dptr
; /* pointer to byte sized element */
1423 uint count
; /* byte count of bwrite */
1424 uint count_init
; /* initial count before roundup */
1425 int roundoff
; /* roundoff to BB or stripe */
1426 int split
= 0; /* split write into two regions */
1428 int v2
= xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
);
1430 XFS_STATS_INC(xs_log_writes
);
1431 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
1433 /* Add for LR header */
1434 count_init
= log
->l_iclog_hsize
+ iclog
->ic_offset
;
1436 /* Round out the log write size */
1437 if (v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1) {
1438 /* we have a v2 stripe unit to use */
1439 count
= XLOG_LSUNITTOB(log
, XLOG_BTOLSUNIT(log
, count_init
));
1441 count
= BBTOB(BTOBB(count_init
));
1443 roundoff
= count
- count_init
;
1444 ASSERT(roundoff
>= 0);
1445 ASSERT((v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1 &&
1446 roundoff
< log
->l_mp
->m_sb
.sb_logsunit
)
1448 (log
->l_mp
->m_sb
.sb_logsunit
<= 1 &&
1449 roundoff
< BBTOB(1)));
1451 /* move grant heads by roundoff in sync */
1452 spin_lock(&log
->l_grant_lock
);
1453 xlog_grant_add_space(log
, roundoff
);
1454 spin_unlock(&log
->l_grant_lock
);
1456 /* put cycle number in every block */
1457 xlog_pack_data(log
, iclog
, roundoff
);
1459 /* real byte length */
1461 iclog
->ic_header
.h_len
=
1462 cpu_to_be32(iclog
->ic_offset
+ roundoff
);
1464 iclog
->ic_header
.h_len
=
1465 cpu_to_be32(iclog
->ic_offset
);
1469 ASSERT(XFS_BUF_FSPRIVATE2(bp
, unsigned long) == (unsigned long)1);
1470 XFS_BUF_SET_FSPRIVATE2(bp
, (unsigned long)2);
1471 XFS_BUF_SET_ADDR(bp
, BLOCK_LSN(be64_to_cpu(iclog
->ic_header
.h_lsn
)));
1473 XFS_STATS_ADD(xs_log_blocks
, BTOBB(count
));
1475 /* Do we need to split this write into 2 parts? */
1476 if (XFS_BUF_ADDR(bp
) + BTOBB(count
) > log
->l_logBBsize
) {
1477 split
= count
- (BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
)));
1478 count
= BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
));
1479 iclog
->ic_bwritecnt
= 2; /* split into 2 writes */
1481 iclog
->ic_bwritecnt
= 1;
1483 XFS_BUF_SET_COUNT(bp
, count
);
1484 XFS_BUF_SET_FSPRIVATE(bp
, iclog
); /* save for later */
1485 XFS_BUF_ZEROFLAGS(bp
);
1489 * Do an ordered write for the log block.
1490 * Its unnecessary to flush the first split block in the log wrap case.
1492 if (!split
&& (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
))
1493 XFS_BUF_ORDERED(bp
);
1495 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1496 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1498 xlog_verify_iclog(log
, iclog
, count
, B_TRUE
);
1500 /* account for log which doesn't start at block #0 */
1501 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1503 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1508 if ((error
= XFS_bwrite(bp
))) {
1509 xfs_ioerror_alert("xlog_sync", log
->l_mp
, bp
,
1514 bp
= iclog
->ic_log
->l_xbuf
;
1515 ASSERT(XFS_BUF_FSPRIVATE2(bp
, unsigned long) ==
1517 XFS_BUF_SET_FSPRIVATE2(bp
, (unsigned long)2);
1518 XFS_BUF_SET_ADDR(bp
, 0); /* logical 0 */
1519 XFS_BUF_SET_PTR(bp
, (xfs_caddr_t
)((__psint_t
)&(iclog
->ic_header
)+
1520 (__psint_t
)count
), split
);
1521 XFS_BUF_SET_FSPRIVATE(bp
, iclog
);
1522 XFS_BUF_ZEROFLAGS(bp
);
1525 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
)
1526 XFS_BUF_ORDERED(bp
);
1527 dptr
= XFS_BUF_PTR(bp
);
1529 * Bump the cycle numbers at the start of each block
1530 * since this part of the buffer is at the start of
1531 * a new cycle. Watch out for the header magic number
1534 for (i
= 0; i
< split
; i
+= BBSIZE
) {
1535 be32_add_cpu((__be32
*)dptr
, 1);
1536 if (be32_to_cpu(*(__be32
*)dptr
) == XLOG_HEADER_MAGIC_NUM
)
1537 be32_add_cpu((__be32
*)dptr
, 1);
1541 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1542 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1544 /* account for internal log which doesn't start at block #0 */
1545 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1547 if ((error
= XFS_bwrite(bp
))) {
1548 xfs_ioerror_alert("xlog_sync (split)", log
->l_mp
,
1549 bp
, XFS_BUF_ADDR(bp
));
1558 * Deallocate a log structure
1561 xlog_dealloc_log(xlog_t
*log
)
1563 xlog_in_core_t
*iclog
, *next_iclog
;
1566 iclog
= log
->l_iclog
;
1567 for (i
=0; i
<log
->l_iclog_bufs
; i
++) {
1568 sv_destroy(&iclog
->ic_forcesema
);
1569 sv_destroy(&iclog
->ic_writesema
);
1570 xfs_buf_free(iclog
->ic_bp
);
1571 #ifdef XFS_LOG_TRACE
1572 if (iclog
->ic_trace
!= NULL
) {
1573 ktrace_free(iclog
->ic_trace
);
1576 next_iclog
= iclog
->ic_next
;
1580 spinlock_destroy(&log
->l_icloglock
);
1581 spinlock_destroy(&log
->l_grant_lock
);
1583 xfs_buf_free(log
->l_xbuf
);
1584 #ifdef XFS_LOG_TRACE
1585 if (log
->l_trace
!= NULL
) {
1586 ktrace_free(log
->l_trace
);
1588 if (log
->l_grant_trace
!= NULL
) {
1589 ktrace_free(log
->l_grant_trace
);
1592 log
->l_mp
->m_log
= NULL
;
1594 } /* xlog_dealloc_log */
1597 * Update counters atomically now that memcpy is done.
1601 xlog_state_finish_copy(xlog_t
*log
,
1602 xlog_in_core_t
*iclog
,
1606 spin_lock(&log
->l_icloglock
);
1608 be32_add_cpu(&iclog
->ic_header
.h_num_logops
, record_cnt
);
1609 iclog
->ic_offset
+= copy_bytes
;
1611 spin_unlock(&log
->l_icloglock
);
1612 } /* xlog_state_finish_copy */
1618 * print out info relating to regions written which consume
1622 xlog_print_tic_res(xfs_mount_t
*mp
, xlog_ticket_t
*ticket
)
1625 uint ophdr_spc
= ticket
->t_res_num_ophdrs
* (uint
)sizeof(xlog_op_header_t
);
1627 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1628 static char *res_type_str
[XLOG_REG_TYPE_MAX
] = {
1649 static char *trans_type_str
[XFS_TRANS_TYPE_MAX
] = {
1692 xfs_fs_cmn_err(CE_WARN
, mp
,
1693 "xfs_log_write: reservation summary:\n"
1694 " trans type = %s (%u)\n"
1695 " unit res = %d bytes\n"
1696 " current res = %d bytes\n"
1697 " total reg = %u bytes (o/flow = %u bytes)\n"
1698 " ophdrs = %u (ophdr space = %u bytes)\n"
1699 " ophdr + reg = %u bytes\n"
1700 " num regions = %u\n",
1701 ((ticket
->t_trans_type
<= 0 ||
1702 ticket
->t_trans_type
> XFS_TRANS_TYPE_MAX
) ?
1703 "bad-trans-type" : trans_type_str
[ticket
->t_trans_type
-1]),
1704 ticket
->t_trans_type
,
1707 ticket
->t_res_arr_sum
, ticket
->t_res_o_flow
,
1708 ticket
->t_res_num_ophdrs
, ophdr_spc
,
1709 ticket
->t_res_arr_sum
+
1710 ticket
->t_res_o_flow
+ ophdr_spc
,
1713 for (i
= 0; i
< ticket
->t_res_num
; i
++) {
1714 uint r_type
= ticket
->t_res_arr
[i
].r_type
;
1716 "region[%u]: %s - %u bytes\n",
1718 ((r_type
<= 0 || r_type
> XLOG_REG_TYPE_MAX
) ?
1719 "bad-rtype" : res_type_str
[r_type
-1]),
1720 ticket
->t_res_arr
[i
].r_len
);
1725 * Write some region out to in-core log
1727 * This will be called when writing externally provided regions or when
1728 * writing out a commit record for a given transaction.
1730 * General algorithm:
1731 * 1. Find total length of this write. This may include adding to the
1732 * lengths passed in.
1733 * 2. Check whether we violate the tickets reservation.
1734 * 3. While writing to this iclog
1735 * A. Reserve as much space in this iclog as can get
1736 * B. If this is first write, save away start lsn
1737 * C. While writing this region:
1738 * 1. If first write of transaction, write start record
1739 * 2. Write log operation header (header per region)
1740 * 3. Find out if we can fit entire region into this iclog
1741 * 4. Potentially, verify destination memcpy ptr
1742 * 5. Memcpy (partial) region
1743 * 6. If partial copy, release iclog; otherwise, continue
1744 * copying more regions into current iclog
1745 * 4. Mark want sync bit (in simulation mode)
1746 * 5. Release iclog for potential flush to on-disk log.
1749 * 1. Panic if reservation is overrun. This should never happen since
1750 * reservation amounts are generated internal to the filesystem.
1752 * 1. Tickets are single threaded data structures.
1753 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1754 * syncing routine. When a single log_write region needs to span
1755 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1756 * on all log operation writes which don't contain the end of the
1757 * region. The XLOG_END_TRANS bit is used for the in-core log
1758 * operation which contains the end of the continued log_write region.
1759 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1760 * we don't really know exactly how much space will be used. As a result,
1761 * we don't update ic_offset until the end when we know exactly how many
1762 * bytes have been written out.
1765 xlog_write(xfs_mount_t
* mp
,
1766 xfs_log_iovec_t reg
[],
1768 xfs_log_ticket_t tic
,
1769 xfs_lsn_t
*start_lsn
,
1770 xlog_in_core_t
**commit_iclog
,
1773 xlog_t
*log
= mp
->m_log
;
1774 xlog_ticket_t
*ticket
= (xlog_ticket_t
*)tic
;
1775 xlog_in_core_t
*iclog
= NULL
; /* ptr to current in-core log */
1776 xlog_op_header_t
*logop_head
; /* ptr to log operation header */
1777 __psint_t ptr
; /* copy address into data region */
1778 int len
; /* # xlog_write() bytes 2 still copy */
1779 int index
; /* region index currently copying */
1780 int log_offset
; /* offset (from 0) into data region */
1781 int start_rec_copy
; /* # bytes to copy for start record */
1782 int partial_copy
; /* did we split a region? */
1783 int partial_copy_len
;/* # bytes copied if split region */
1784 int need_copy
; /* # bytes need to memcpy this region */
1785 int copy_len
; /* # bytes actually memcpy'ing */
1786 int copy_off
; /* # bytes from entry start */
1787 int contwr
; /* continued write of in-core log? */
1789 int record_cnt
= 0, data_cnt
= 0;
1791 partial_copy_len
= partial_copy
= 0;
1793 /* Calculate potential maximum space. Each region gets its own
1794 * xlog_op_header_t and may need to be double word aligned.
1797 if (ticket
->t_flags
& XLOG_TIC_INITED
) { /* acct for start rec of xact */
1798 len
+= sizeof(xlog_op_header_t
);
1799 ticket
->t_res_num_ophdrs
++;
1802 for (index
= 0; index
< nentries
; index
++) {
1803 len
+= sizeof(xlog_op_header_t
); /* each region gets >= 1 */
1804 ticket
->t_res_num_ophdrs
++;
1805 len
+= reg
[index
].i_len
;
1806 xlog_tic_add_region(ticket
, reg
[index
].i_len
, reg
[index
].i_type
);
1808 contwr
= *start_lsn
= 0;
1810 if (ticket
->t_curr_res
< len
) {
1811 xlog_print_tic_res(mp
, ticket
);
1814 "xfs_log_write: reservation ran out. Need to up reservation");
1816 /* Customer configurable panic */
1817 xfs_cmn_err(XFS_PTAG_LOGRES
, CE_ALERT
, mp
,
1818 "xfs_log_write: reservation ran out. Need to up reservation");
1819 /* If we did not panic, shutdown the filesystem */
1820 xfs_force_shutdown(mp
, SHUTDOWN_CORRUPT_INCORE
);
1823 ticket
->t_curr_res
-= len
;
1825 for (index
= 0; index
< nentries
; ) {
1826 if ((error
= xlog_state_get_iclog_space(log
, len
, &iclog
, ticket
,
1827 &contwr
, &log_offset
)))
1830 ASSERT(log_offset
<= iclog
->ic_size
- 1);
1831 ptr
= (__psint_t
) ((char *)iclog
->ic_datap
+log_offset
);
1833 /* start_lsn is the first lsn written to. That's all we need. */
1835 *start_lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
1837 /* This loop writes out as many regions as can fit in the amount
1838 * of space which was allocated by xlog_state_get_iclog_space().
1840 while (index
< nentries
) {
1841 ASSERT(reg
[index
].i_len
% sizeof(__int32_t
) == 0);
1842 ASSERT((__psint_t
)ptr
% sizeof(__int32_t
) == 0);
1845 /* If first write for transaction, insert start record.
1846 * We can't be trying to commit if we are inited. We can't
1847 * have any "partial_copy" if we are inited.
1849 if (ticket
->t_flags
& XLOG_TIC_INITED
) {
1850 logop_head
= (xlog_op_header_t
*)ptr
;
1851 logop_head
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1852 logop_head
->oh_clientid
= ticket
->t_clientid
;
1853 logop_head
->oh_len
= 0;
1854 logop_head
->oh_flags
= XLOG_START_TRANS
;
1855 logop_head
->oh_res2
= 0;
1856 ticket
->t_flags
&= ~XLOG_TIC_INITED
; /* clear bit */
1859 start_rec_copy
= sizeof(xlog_op_header_t
);
1860 xlog_write_adv_cnt(ptr
, len
, log_offset
, start_rec_copy
);
1863 /* Copy log operation header directly into data section */
1864 logop_head
= (xlog_op_header_t
*)ptr
;
1865 logop_head
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1866 logop_head
->oh_clientid
= ticket
->t_clientid
;
1867 logop_head
->oh_res2
= 0;
1869 /* header copied directly */
1870 xlog_write_adv_cnt(ptr
, len
, log_offset
, sizeof(xlog_op_header_t
));
1872 /* are we copying a commit or unmount record? */
1873 logop_head
->oh_flags
= flags
;
1876 * We've seen logs corrupted with bad transaction client
1877 * ids. This makes sure that XFS doesn't generate them on.
1878 * Turn this into an EIO and shut down the filesystem.
1880 switch (logop_head
->oh_clientid
) {
1881 case XFS_TRANSACTION
:
1886 xfs_fs_cmn_err(CE_WARN
, mp
,
1887 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1888 logop_head
->oh_clientid
, tic
);
1889 return XFS_ERROR(EIO
);
1892 /* Partial write last time? => (partial_copy != 0)
1893 * need_copy is the amount we'd like to copy if everything could
1894 * fit in the current memcpy.
1896 need_copy
= reg
[index
].i_len
- partial_copy_len
;
1898 copy_off
= partial_copy_len
;
1899 if (need_copy
<= iclog
->ic_size
- log_offset
) { /*complete write */
1900 copy_len
= need_copy
;
1901 logop_head
->oh_len
= cpu_to_be32(copy_len
);
1903 logop_head
->oh_flags
|= (XLOG_END_TRANS
|XLOG_WAS_CONT_TRANS
);
1904 partial_copy_len
= partial_copy
= 0;
1905 } else { /* partial write */
1906 copy_len
= iclog
->ic_size
- log_offset
;
1907 logop_head
->oh_len
= cpu_to_be32(copy_len
);
1908 logop_head
->oh_flags
|= XLOG_CONTINUE_TRANS
;
1910 logop_head
->oh_flags
|= XLOG_WAS_CONT_TRANS
;
1911 partial_copy_len
+= copy_len
;
1913 len
+= sizeof(xlog_op_header_t
); /* from splitting of region */
1914 /* account for new log op header */
1915 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
1916 ticket
->t_res_num_ophdrs
++;
1918 xlog_verify_dest_ptr(log
, ptr
);
1921 ASSERT(copy_len
>= 0);
1922 memcpy((xfs_caddr_t
)ptr
, reg
[index
].i_addr
+ copy_off
, copy_len
);
1923 xlog_write_adv_cnt(ptr
, len
, log_offset
, copy_len
);
1925 /* make copy_len total bytes copied, including headers */
1926 copy_len
+= start_rec_copy
+ sizeof(xlog_op_header_t
);
1928 data_cnt
+= contwr
? copy_len
: 0;
1929 if (partial_copy
) { /* copied partial region */
1930 /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1931 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
1932 record_cnt
= data_cnt
= 0;
1933 if ((error
= xlog_state_release_iclog(log
, iclog
)))
1935 break; /* don't increment index */
1936 } else { /* copied entire region */
1938 partial_copy_len
= partial_copy
= 0;
1940 if (iclog
->ic_size
- log_offset
<= sizeof(xlog_op_header_t
)) {
1941 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
1942 record_cnt
= data_cnt
= 0;
1943 xlog_state_want_sync(log
, iclog
);
1945 ASSERT(flags
& XLOG_COMMIT_TRANS
);
1946 *commit_iclog
= iclog
;
1947 } else if ((error
= xlog_state_release_iclog(log
, iclog
)))
1949 if (index
== nentries
)
1950 return 0; /* we are done */
1954 } /* if (partial_copy) */
1955 } /* while (index < nentries) */
1956 } /* for (index = 0; index < nentries; ) */
1959 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
1961 ASSERT(flags
& XLOG_COMMIT_TRANS
);
1962 *commit_iclog
= iclog
;
1965 return xlog_state_release_iclog(log
, iclog
);
1969 /*****************************************************************************
1971 * State Machine functions
1973 *****************************************************************************
1976 /* Clean iclogs starting from the head. This ordering must be
1977 * maintained, so an iclog doesn't become ACTIVE beyond one that
1978 * is SYNCING. This is also required to maintain the notion that we use
1979 * a counting semaphore to hold off would be writers to the log when every
1980 * iclog is trying to sync to disk.
1982 * State Change: DIRTY -> ACTIVE
1985 xlog_state_clean_log(xlog_t
*log
)
1987 xlog_in_core_t
*iclog
;
1990 iclog
= log
->l_iclog
;
1992 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
1993 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1994 iclog
->ic_offset
= 0;
1995 ASSERT(iclog
->ic_callback
== NULL
);
1997 * If the number of ops in this iclog indicate it just
1998 * contains the dummy transaction, we can
1999 * change state into IDLE (the second time around).
2000 * Otherwise we should change the state into
2002 * We don't need to cover the dummy.
2005 (be32_to_cpu(iclog
->ic_header
.h_num_logops
) ==
2010 * We have two dirty iclogs so start over
2011 * This could also be num of ops indicates
2012 * this is not the dummy going out.
2016 iclog
->ic_header
.h_num_logops
= 0;
2017 memset(iclog
->ic_header
.h_cycle_data
, 0,
2018 sizeof(iclog
->ic_header
.h_cycle_data
));
2019 iclog
->ic_header
.h_lsn
= 0;
2020 } else if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2023 break; /* stop cleaning */
2024 iclog
= iclog
->ic_next
;
2025 } while (iclog
!= log
->l_iclog
);
2027 /* log is locked when we are called */
2029 * Change state for the dummy log recording.
2030 * We usually go to NEED. But we go to NEED2 if the changed indicates
2031 * we are done writing the dummy record.
2032 * If we are done with the second dummy recored (DONE2), then
2036 switch (log
->l_covered_state
) {
2037 case XLOG_STATE_COVER_IDLE
:
2038 case XLOG_STATE_COVER_NEED
:
2039 case XLOG_STATE_COVER_NEED2
:
2040 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2043 case XLOG_STATE_COVER_DONE
:
2045 log
->l_covered_state
= XLOG_STATE_COVER_NEED2
;
2047 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2050 case XLOG_STATE_COVER_DONE2
:
2052 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
2054 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2061 } /* xlog_state_clean_log */
2064 xlog_get_lowest_lsn(
2067 xlog_in_core_t
*lsn_log
;
2068 xfs_lsn_t lowest_lsn
, lsn
;
2070 lsn_log
= log
->l_iclog
;
2073 if (!(lsn_log
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
))) {
2074 lsn
= be64_to_cpu(lsn_log
->ic_header
.h_lsn
);
2075 if ((lsn
&& !lowest_lsn
) ||
2076 (XFS_LSN_CMP(lsn
, lowest_lsn
) < 0)) {
2080 lsn_log
= lsn_log
->ic_next
;
2081 } while (lsn_log
!= log
->l_iclog
);
2087 xlog_state_do_callback(
2090 xlog_in_core_t
*ciclog
)
2092 xlog_in_core_t
*iclog
;
2093 xlog_in_core_t
*first_iclog
; /* used to know when we've
2094 * processed all iclogs once */
2095 xfs_log_callback_t
*cb
, *cb_next
;
2097 xfs_lsn_t lowest_lsn
;
2098 int ioerrors
; /* counter: iclogs with errors */
2099 int loopdidcallbacks
; /* flag: inner loop did callbacks*/
2100 int funcdidcallbacks
; /* flag: function did callbacks */
2101 int repeats
; /* for issuing console warnings if
2102 * looping too many times */
2105 spin_lock(&log
->l_icloglock
);
2106 first_iclog
= iclog
= log
->l_iclog
;
2108 funcdidcallbacks
= 0;
2113 * Scan all iclogs starting with the one pointed to by the
2114 * log. Reset this starting point each time the log is
2115 * unlocked (during callbacks).
2117 * Keep looping through iclogs until one full pass is made
2118 * without running any callbacks.
2120 first_iclog
= log
->l_iclog
;
2121 iclog
= log
->l_iclog
;
2122 loopdidcallbacks
= 0;
2127 /* skip all iclogs in the ACTIVE & DIRTY states */
2128 if (iclog
->ic_state
&
2129 (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
)) {
2130 iclog
= iclog
->ic_next
;
2135 * Between marking a filesystem SHUTDOWN and stopping
2136 * the log, we do flush all iclogs to disk (if there
2137 * wasn't a log I/O error). So, we do want things to
2138 * go smoothly in case of just a SHUTDOWN w/o a
2141 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
2143 * Can only perform callbacks in order. Since
2144 * this iclog is not in the DONE_SYNC/
2145 * DO_CALLBACK state, we skip the rest and
2146 * just try to clean up. If we set our iclog
2147 * to DO_CALLBACK, we will not process it when
2148 * we retry since a previous iclog is in the
2149 * CALLBACK and the state cannot change since
2150 * we are holding the l_icloglock.
2152 if (!(iclog
->ic_state
&
2153 (XLOG_STATE_DONE_SYNC
|
2154 XLOG_STATE_DO_CALLBACK
))) {
2155 if (ciclog
&& (ciclog
->ic_state
==
2156 XLOG_STATE_DONE_SYNC
)) {
2157 ciclog
->ic_state
= XLOG_STATE_DO_CALLBACK
;
2162 * We now have an iclog that is in either the
2163 * DO_CALLBACK or DONE_SYNC states. The other
2164 * states (WANT_SYNC, SYNCING, or CALLBACK were
2165 * caught by the above if and are going to
2166 * clean (i.e. we aren't doing their callbacks)
2171 * We will do one more check here to see if we
2172 * have chased our tail around.
2175 lowest_lsn
= xlog_get_lowest_lsn(log
);
2177 XFS_LSN_CMP(lowest_lsn
,
2178 be64_to_cpu(iclog
->ic_header
.h_lsn
)) < 0) {
2179 iclog
= iclog
->ic_next
;
2180 continue; /* Leave this iclog for
2184 iclog
->ic_state
= XLOG_STATE_CALLBACK
;
2186 spin_unlock(&log
->l_icloglock
);
2188 /* l_last_sync_lsn field protected by
2189 * l_grant_lock. Don't worry about iclog's lsn.
2190 * No one else can be here except us.
2192 spin_lock(&log
->l_grant_lock
);
2193 ASSERT(XFS_LSN_CMP(log
->l_last_sync_lsn
,
2194 be64_to_cpu(iclog
->ic_header
.h_lsn
)) <= 0);
2195 log
->l_last_sync_lsn
=
2196 be64_to_cpu(iclog
->ic_header
.h_lsn
);
2197 spin_unlock(&log
->l_grant_lock
);
2200 spin_unlock(&log
->l_icloglock
);
2205 * Keep processing entries in the callback list until
2206 * we come around and it is empty. We need to
2207 * atomically see that the list is empty and change the
2208 * state to DIRTY so that we don't miss any more
2209 * callbacks being added.
2211 spin_lock(&iclog
->ic_callback_lock
);
2212 cb
= iclog
->ic_callback
;
2214 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
2215 iclog
->ic_callback
= NULL
;
2216 spin_unlock(&iclog
->ic_callback_lock
);
2218 /* perform callbacks in the order given */
2219 for (; cb
; cb
= cb_next
) {
2220 cb_next
= cb
->cb_next
;
2221 cb
->cb_func(cb
->cb_arg
, aborted
);
2223 spin_lock(&iclog
->ic_callback_lock
);
2224 cb
= iclog
->ic_callback
;
2230 spin_lock(&log
->l_icloglock
);
2231 ASSERT(iclog
->ic_callback
== NULL
);
2232 spin_unlock(&iclog
->ic_callback_lock
);
2233 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
))
2234 iclog
->ic_state
= XLOG_STATE_DIRTY
;
2237 * Transition from DIRTY to ACTIVE if applicable.
2238 * NOP if STATE_IOERROR.
2240 xlog_state_clean_log(log
);
2242 /* wake up threads waiting in xfs_log_force() */
2243 sv_broadcast(&iclog
->ic_forcesema
);
2245 iclog
= iclog
->ic_next
;
2246 } while (first_iclog
!= iclog
);
2248 if (repeats
> 5000) {
2249 flushcnt
+= repeats
;
2251 xfs_fs_cmn_err(CE_WARN
, log
->l_mp
,
2252 "%s: possible infinite loop (%d iterations)",
2253 __func__
, flushcnt
);
2255 } while (!ioerrors
&& loopdidcallbacks
);
2258 * make one last gasp attempt to see if iclogs are being left in
2262 if (funcdidcallbacks
) {
2263 first_iclog
= iclog
= log
->l_iclog
;
2265 ASSERT(iclog
->ic_state
!= XLOG_STATE_DO_CALLBACK
);
2267 * Terminate the loop if iclogs are found in states
2268 * which will cause other threads to clean up iclogs.
2270 * SYNCING - i/o completion will go through logs
2271 * DONE_SYNC - interrupt thread should be waiting for
2273 * IOERROR - give up hope all ye who enter here
2275 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
||
2276 iclog
->ic_state
== XLOG_STATE_SYNCING
||
2277 iclog
->ic_state
== XLOG_STATE_DONE_SYNC
||
2278 iclog
->ic_state
== XLOG_STATE_IOERROR
)
2280 iclog
= iclog
->ic_next
;
2281 } while (first_iclog
!= iclog
);
2285 if (log
->l_iclog
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_IOERROR
))
2287 spin_unlock(&log
->l_icloglock
);
2290 sv_broadcast(&log
->l_flush_wait
);
2295 * Finish transitioning this iclog to the dirty state.
2297 * Make sure that we completely execute this routine only when this is
2298 * the last call to the iclog. There is a good chance that iclog flushes,
2299 * when we reach the end of the physical log, get turned into 2 separate
2300 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2301 * routine. By using the reference count bwritecnt, we guarantee that only
2302 * the second completion goes through.
2304 * Callbacks could take time, so they are done outside the scope of the
2305 * global state machine log lock. Assume that the calls to cvsema won't
2306 * take a long time. At least we know it won't sleep.
2309 xlog_state_done_syncing(
2310 xlog_in_core_t
*iclog
,
2313 xlog_t
*log
= iclog
->ic_log
;
2315 spin_lock(&log
->l_icloglock
);
2317 ASSERT(iclog
->ic_state
== XLOG_STATE_SYNCING
||
2318 iclog
->ic_state
== XLOG_STATE_IOERROR
);
2319 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2320 ASSERT(iclog
->ic_bwritecnt
== 1 || iclog
->ic_bwritecnt
== 2);
2324 * If we got an error, either on the first buffer, or in the case of
2325 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2326 * and none should ever be attempted to be written to disk
2329 if (iclog
->ic_state
!= XLOG_STATE_IOERROR
) {
2330 if (--iclog
->ic_bwritecnt
== 1) {
2331 spin_unlock(&log
->l_icloglock
);
2334 iclog
->ic_state
= XLOG_STATE_DONE_SYNC
;
2338 * Someone could be sleeping prior to writing out the next
2339 * iclog buffer, we wake them all, one will get to do the
2340 * I/O, the others get to wait for the result.
2342 sv_broadcast(&iclog
->ic_writesema
);
2343 spin_unlock(&log
->l_icloglock
);
2344 xlog_state_do_callback(log
, aborted
, iclog
); /* also cleans log */
2345 } /* xlog_state_done_syncing */
2349 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2350 * sleep. The flush semaphore is set to the number of in-core buffers and
2351 * decremented around disk syncing. Therefore, if all buffers are syncing,
2352 * this semaphore will cause new writes to sleep until a sync completes.
2353 * Otherwise, this code just does p() followed by v(). This approximates
2354 * a sleep/wakeup except we can't race.
2356 * The in-core logs are used in a circular fashion. They are not used
2357 * out-of-order even when an iclog past the head is free.
2360 * * log_offset where xlog_write() can start writing into the in-core
2362 * * in-core log pointer to which xlog_write() should write.
2363 * * boolean indicating this is a continued write to an in-core log.
2364 * If this is the last write, then the in-core log's offset field
2365 * needs to be incremented, depending on the amount of data which
2369 xlog_state_get_iclog_space(xlog_t
*log
,
2371 xlog_in_core_t
**iclogp
,
2372 xlog_ticket_t
*ticket
,
2373 int *continued_write
,
2377 xlog_rec_header_t
*head
;
2378 xlog_in_core_t
*iclog
;
2382 spin_lock(&log
->l_icloglock
);
2383 if (XLOG_FORCED_SHUTDOWN(log
)) {
2384 spin_unlock(&log
->l_icloglock
);
2385 return XFS_ERROR(EIO
);
2388 iclog
= log
->l_iclog
;
2389 if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
) {
2390 xlog_trace_iclog(iclog
, XLOG_TRACE_SLEEP_FLUSH
);
2391 XFS_STATS_INC(xs_log_noiclogs
);
2393 /* Wait for log writes to have flushed */
2394 sv_wait(&log
->l_flush_wait
, 0, &log
->l_icloglock
, 0);
2398 head
= &iclog
->ic_header
;
2400 atomic_inc(&iclog
->ic_refcnt
); /* prevents sync */
2401 log_offset
= iclog
->ic_offset
;
2403 /* On the 1st write to an iclog, figure out lsn. This works
2404 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2405 * committing to. If the offset is set, that's how many blocks
2408 if (log_offset
== 0) {
2409 ticket
->t_curr_res
-= log
->l_iclog_hsize
;
2410 xlog_tic_add_region(ticket
,
2412 XLOG_REG_TYPE_LRHEADER
);
2413 head
->h_cycle
= cpu_to_be32(log
->l_curr_cycle
);
2414 head
->h_lsn
= cpu_to_be64(
2415 xlog_assign_lsn(log
->l_curr_cycle
, log
->l_curr_block
));
2416 ASSERT(log
->l_curr_block
>= 0);
2419 /* If there is enough room to write everything, then do it. Otherwise,
2420 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2421 * bit is on, so this will get flushed out. Don't update ic_offset
2422 * until you know exactly how many bytes get copied. Therefore, wait
2423 * until later to update ic_offset.
2425 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2426 * can fit into remaining data section.
2428 if (iclog
->ic_size
- iclog
->ic_offset
< 2*sizeof(xlog_op_header_t
)) {
2429 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2432 * If I'm the only one writing to this iclog, sync it to disk.
2433 * We need to do an atomic compare and decrement here to avoid
2434 * racing with concurrent atomic_dec_and_lock() calls in
2435 * xlog_state_release_iclog() when there is more than one
2436 * reference to the iclog.
2438 if (!atomic_add_unless(&iclog
->ic_refcnt
, -1, 1)) {
2439 /* we are the only one */
2440 spin_unlock(&log
->l_icloglock
);
2441 error
= xlog_state_release_iclog(log
, iclog
);
2445 spin_unlock(&log
->l_icloglock
);
2450 /* Do we have enough room to write the full amount in the remainder
2451 * of this iclog? Or must we continue a write on the next iclog and
2452 * mark this iclog as completely taken? In the case where we switch
2453 * iclogs (to mark it taken), this particular iclog will release/sync
2454 * to disk in xlog_write().
2456 if (len
<= iclog
->ic_size
- iclog
->ic_offset
) {
2457 *continued_write
= 0;
2458 iclog
->ic_offset
+= len
;
2460 *continued_write
= 1;
2461 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2465 ASSERT(iclog
->ic_offset
<= iclog
->ic_size
);
2466 spin_unlock(&log
->l_icloglock
);
2468 *logoffsetp
= log_offset
;
2470 } /* xlog_state_get_iclog_space */
2473 * Atomically get the log space required for a log ticket.
2475 * Once a ticket gets put onto the reserveq, it will only return after
2476 * the needed reservation is satisfied.
2479 xlog_grant_log_space(xlog_t
*log
,
2490 if (log
->l_flags
& XLOG_ACTIVE_RECOVERY
)
2491 panic("grant Recovery problem");
2494 /* Is there space or do we need to sleep? */
2495 spin_lock(&log
->l_grant_lock
);
2496 xlog_trace_loggrant(log
, tic
, "xlog_grant_log_space: enter");
2498 /* something is already sleeping; insert new transaction at end */
2499 if (log
->l_reserve_headq
) {
2500 xlog_ins_ticketq(&log
->l_reserve_headq
, tic
);
2501 xlog_trace_loggrant(log
, tic
,
2502 "xlog_grant_log_space: sleep 1");
2504 * Gotta check this before going to sleep, while we're
2505 * holding the grant lock.
2507 if (XLOG_FORCED_SHUTDOWN(log
))
2510 XFS_STATS_INC(xs_sleep_logspace
);
2511 sv_wait(&tic
->t_sema
, PINOD
|PLTWAIT
, &log
->l_grant_lock
, s
);
2513 * If we got an error, and the filesystem is shutting down,
2514 * we'll catch it down below. So just continue...
2516 xlog_trace_loggrant(log
, tic
,
2517 "xlog_grant_log_space: wake 1");
2518 spin_lock(&log
->l_grant_lock
);
2520 if (tic
->t_flags
& XFS_LOG_PERM_RESERV
)
2521 need_bytes
= tic
->t_unit_res
*tic
->t_ocnt
;
2523 need_bytes
= tic
->t_unit_res
;
2526 if (XLOG_FORCED_SHUTDOWN(log
))
2529 free_bytes
= xlog_space_left(log
, log
->l_grant_reserve_cycle
,
2530 log
->l_grant_reserve_bytes
);
2531 if (free_bytes
< need_bytes
) {
2532 if ((tic
->t_flags
& XLOG_TIC_IN_Q
) == 0)
2533 xlog_ins_ticketq(&log
->l_reserve_headq
, tic
);
2534 xlog_trace_loggrant(log
, tic
,
2535 "xlog_grant_log_space: sleep 2");
2536 XFS_STATS_INC(xs_sleep_logspace
);
2537 sv_wait(&tic
->t_sema
, PINOD
|PLTWAIT
, &log
->l_grant_lock
, s
);
2539 if (XLOG_FORCED_SHUTDOWN(log
)) {
2540 spin_lock(&log
->l_grant_lock
);
2544 xlog_trace_loggrant(log
, tic
,
2545 "xlog_grant_log_space: wake 2");
2546 xlog_grant_push_ail(log
->l_mp
, need_bytes
);
2547 spin_lock(&log
->l_grant_lock
);
2549 } else if (tic
->t_flags
& XLOG_TIC_IN_Q
)
2550 xlog_del_ticketq(&log
->l_reserve_headq
, tic
);
2552 /* we've got enough space */
2553 xlog_grant_add_space(log
, need_bytes
);
2555 tail_lsn
= log
->l_tail_lsn
;
2557 * Check to make sure the grant write head didn't just over lap the
2558 * tail. If the cycles are the same, we can't be overlapping.
2559 * Otherwise, make sure that the cycles differ by exactly one and
2560 * check the byte count.
2562 if (CYCLE_LSN(tail_lsn
) != log
->l_grant_write_cycle
) {
2563 ASSERT(log
->l_grant_write_cycle
-1 == CYCLE_LSN(tail_lsn
));
2564 ASSERT(log
->l_grant_write_bytes
<= BBTOB(BLOCK_LSN(tail_lsn
)));
2567 xlog_trace_loggrant(log
, tic
, "xlog_grant_log_space: exit");
2568 xlog_verify_grant_head(log
, 1);
2569 spin_unlock(&log
->l_grant_lock
);
2573 if (tic
->t_flags
& XLOG_TIC_IN_Q
)
2574 xlog_del_ticketq(&log
->l_reserve_headq
, tic
);
2575 xlog_trace_loggrant(log
, tic
, "xlog_grant_log_space: err_ret");
2577 * If we are failing, make sure the ticket doesn't have any
2578 * current reservations. We don't want to add this back when
2579 * the ticket/transaction gets cancelled.
2581 tic
->t_curr_res
= 0;
2582 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
2583 spin_unlock(&log
->l_grant_lock
);
2584 return XFS_ERROR(EIO
);
2585 } /* xlog_grant_log_space */
2589 * Replenish the byte reservation required by moving the grant write head.
2594 xlog_regrant_write_log_space(xlog_t
*log
,
2597 int free_bytes
, need_bytes
;
2598 xlog_ticket_t
*ntic
;
2603 tic
->t_curr_res
= tic
->t_unit_res
;
2604 xlog_tic_reset_res(tic
);
2610 if (log
->l_flags
& XLOG_ACTIVE_RECOVERY
)
2611 panic("regrant Recovery problem");
2614 spin_lock(&log
->l_grant_lock
);
2615 xlog_trace_loggrant(log
, tic
, "xlog_regrant_write_log_space: enter");
2617 if (XLOG_FORCED_SHUTDOWN(log
))
2620 /* If there are other waiters on the queue then give them a
2621 * chance at logspace before us. Wake up the first waiters,
2622 * if we do not wake up all the waiters then go to sleep waiting
2623 * for more free space, otherwise try to get some space for
2627 if ((ntic
= log
->l_write_headq
)) {
2628 free_bytes
= xlog_space_left(log
, log
->l_grant_write_cycle
,
2629 log
->l_grant_write_bytes
);
2631 ASSERT(ntic
->t_flags
& XLOG_TIC_PERM_RESERV
);
2633 if (free_bytes
< ntic
->t_unit_res
)
2635 free_bytes
-= ntic
->t_unit_res
;
2636 sv_signal(&ntic
->t_sema
);
2637 ntic
= ntic
->t_next
;
2638 } while (ntic
!= log
->l_write_headq
);
2640 if (ntic
!= log
->l_write_headq
) {
2641 if ((tic
->t_flags
& XLOG_TIC_IN_Q
) == 0)
2642 xlog_ins_ticketq(&log
->l_write_headq
, tic
);
2644 xlog_trace_loggrant(log
, tic
,
2645 "xlog_regrant_write_log_space: sleep 1");
2646 XFS_STATS_INC(xs_sleep_logspace
);
2647 sv_wait(&tic
->t_sema
, PINOD
|PLTWAIT
,
2648 &log
->l_grant_lock
, s
);
2650 /* If we're shutting down, this tic is already
2652 if (XLOG_FORCED_SHUTDOWN(log
)) {
2653 spin_lock(&log
->l_grant_lock
);
2657 xlog_trace_loggrant(log
, tic
,
2658 "xlog_regrant_write_log_space: wake 1");
2659 xlog_grant_push_ail(log
->l_mp
, tic
->t_unit_res
);
2660 spin_lock(&log
->l_grant_lock
);
2664 need_bytes
= tic
->t_unit_res
;
2667 if (XLOG_FORCED_SHUTDOWN(log
))
2670 free_bytes
= xlog_space_left(log
, log
->l_grant_write_cycle
,
2671 log
->l_grant_write_bytes
);
2672 if (free_bytes
< need_bytes
) {
2673 if ((tic
->t_flags
& XLOG_TIC_IN_Q
) == 0)
2674 xlog_ins_ticketq(&log
->l_write_headq
, tic
);
2675 XFS_STATS_INC(xs_sleep_logspace
);
2676 sv_wait(&tic
->t_sema
, PINOD
|PLTWAIT
, &log
->l_grant_lock
, s
);
2678 /* If we're shutting down, this tic is already off the queue */
2679 if (XLOG_FORCED_SHUTDOWN(log
)) {
2680 spin_lock(&log
->l_grant_lock
);
2684 xlog_trace_loggrant(log
, tic
,
2685 "xlog_regrant_write_log_space: wake 2");
2686 xlog_grant_push_ail(log
->l_mp
, need_bytes
);
2687 spin_lock(&log
->l_grant_lock
);
2689 } else if (tic
->t_flags
& XLOG_TIC_IN_Q
)
2690 xlog_del_ticketq(&log
->l_write_headq
, tic
);
2692 /* we've got enough space */
2693 xlog_grant_add_space_write(log
, need_bytes
);
2695 tail_lsn
= log
->l_tail_lsn
;
2696 if (CYCLE_LSN(tail_lsn
) != log
->l_grant_write_cycle
) {
2697 ASSERT(log
->l_grant_write_cycle
-1 == CYCLE_LSN(tail_lsn
));
2698 ASSERT(log
->l_grant_write_bytes
<= BBTOB(BLOCK_LSN(tail_lsn
)));
2702 xlog_trace_loggrant(log
, tic
, "xlog_regrant_write_log_space: exit");
2703 xlog_verify_grant_head(log
, 1);
2704 spin_unlock(&log
->l_grant_lock
);
2709 if (tic
->t_flags
& XLOG_TIC_IN_Q
)
2710 xlog_del_ticketq(&log
->l_reserve_headq
, tic
);
2711 xlog_trace_loggrant(log
, tic
, "xlog_regrant_write_log_space: err_ret");
2713 * If we are failing, make sure the ticket doesn't have any
2714 * current reservations. We don't want to add this back when
2715 * the ticket/transaction gets cancelled.
2717 tic
->t_curr_res
= 0;
2718 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
2719 spin_unlock(&log
->l_grant_lock
);
2720 return XFS_ERROR(EIO
);
2721 } /* xlog_regrant_write_log_space */
2724 /* The first cnt-1 times through here we don't need to
2725 * move the grant write head because the permanent
2726 * reservation has reserved cnt times the unit amount.
2727 * Release part of current permanent unit reservation and
2728 * reset current reservation to be one units worth. Also
2729 * move grant reservation head forward.
2732 xlog_regrant_reserve_log_space(xlog_t
*log
,
2733 xlog_ticket_t
*ticket
)
2735 xlog_trace_loggrant(log
, ticket
,
2736 "xlog_regrant_reserve_log_space: enter");
2737 if (ticket
->t_cnt
> 0)
2740 spin_lock(&log
->l_grant_lock
);
2741 xlog_grant_sub_space(log
, ticket
->t_curr_res
);
2742 ticket
->t_curr_res
= ticket
->t_unit_res
;
2743 xlog_tic_reset_res(ticket
);
2744 xlog_trace_loggrant(log
, ticket
,
2745 "xlog_regrant_reserve_log_space: sub current res");
2746 xlog_verify_grant_head(log
, 1);
2748 /* just return if we still have some of the pre-reserved space */
2749 if (ticket
->t_cnt
> 0) {
2750 spin_unlock(&log
->l_grant_lock
);
2754 xlog_grant_add_space_reserve(log
, ticket
->t_unit_res
);
2755 xlog_trace_loggrant(log
, ticket
,
2756 "xlog_regrant_reserve_log_space: exit");
2757 xlog_verify_grant_head(log
, 0);
2758 spin_unlock(&log
->l_grant_lock
);
2759 ticket
->t_curr_res
= ticket
->t_unit_res
;
2760 xlog_tic_reset_res(ticket
);
2761 } /* xlog_regrant_reserve_log_space */
2765 * Give back the space left from a reservation.
2767 * All the information we need to make a correct determination of space left
2768 * is present. For non-permanent reservations, things are quite easy. The
2769 * count should have been decremented to zero. We only need to deal with the
2770 * space remaining in the current reservation part of the ticket. If the
2771 * ticket contains a permanent reservation, there may be left over space which
2772 * needs to be released. A count of N means that N-1 refills of the current
2773 * reservation can be done before we need to ask for more space. The first
2774 * one goes to fill up the first current reservation. Once we run out of
2775 * space, the count will stay at zero and the only space remaining will be
2776 * in the current reservation field.
2779 xlog_ungrant_log_space(xlog_t
*log
,
2780 xlog_ticket_t
*ticket
)
2782 if (ticket
->t_cnt
> 0)
2785 spin_lock(&log
->l_grant_lock
);
2786 xlog_trace_loggrant(log
, ticket
, "xlog_ungrant_log_space: enter");
2788 xlog_grant_sub_space(log
, ticket
->t_curr_res
);
2790 xlog_trace_loggrant(log
, ticket
, "xlog_ungrant_log_space: sub current");
2792 /* If this is a permanent reservation ticket, we may be able to free
2793 * up more space based on the remaining count.
2795 if (ticket
->t_cnt
> 0) {
2796 ASSERT(ticket
->t_flags
& XLOG_TIC_PERM_RESERV
);
2797 xlog_grant_sub_space(log
, ticket
->t_unit_res
*ticket
->t_cnt
);
2800 xlog_trace_loggrant(log
, ticket
, "xlog_ungrant_log_space: exit");
2801 xlog_verify_grant_head(log
, 1);
2802 spin_unlock(&log
->l_grant_lock
);
2803 xfs_log_move_tail(log
->l_mp
, 1);
2804 } /* xlog_ungrant_log_space */
2808 * Flush iclog to disk if this is the last reference to the given iclog and
2809 * the WANT_SYNC bit is set.
2811 * When this function is entered, the iclog is not necessarily in the
2812 * WANT_SYNC state. It may be sitting around waiting to get filled.
2817 xlog_state_release_iclog(
2819 xlog_in_core_t
*iclog
)
2821 int sync
= 0; /* do we sync? */
2823 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
2824 return XFS_ERROR(EIO
);
2826 ASSERT(atomic_read(&iclog
->ic_refcnt
) > 0);
2827 if (!atomic_dec_and_lock(&iclog
->ic_refcnt
, &log
->l_icloglock
))
2830 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2831 spin_unlock(&log
->l_icloglock
);
2832 return XFS_ERROR(EIO
);
2834 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2835 iclog
->ic_state
== XLOG_STATE_WANT_SYNC
);
2837 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
) {
2838 /* update tail before writing to iclog */
2839 xlog_assign_tail_lsn(log
->l_mp
);
2841 iclog
->ic_state
= XLOG_STATE_SYNCING
;
2842 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(log
->l_tail_lsn
);
2843 xlog_verify_tail_lsn(log
, iclog
, log
->l_tail_lsn
);
2844 /* cycle incremented when incrementing curr_block */
2846 spin_unlock(&log
->l_icloglock
);
2849 * We let the log lock go, so it's possible that we hit a log I/O
2850 * error or some other SHUTDOWN condition that marks the iclog
2851 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2852 * this iclog has consistent data, so we ignore IOERROR
2853 * flags after this point.
2856 return xlog_sync(log
, iclog
);
2858 } /* xlog_state_release_iclog */
2862 * This routine will mark the current iclog in the ring as WANT_SYNC
2863 * and move the current iclog pointer to the next iclog in the ring.
2864 * When this routine is called from xlog_state_get_iclog_space(), the
2865 * exact size of the iclog has not yet been determined. All we know is
2866 * that every data block. We have run out of space in this log record.
2869 xlog_state_switch_iclogs(xlog_t
*log
,
2870 xlog_in_core_t
*iclog
,
2873 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
2875 eventual_size
= iclog
->ic_offset
;
2876 iclog
->ic_state
= XLOG_STATE_WANT_SYNC
;
2877 iclog
->ic_header
.h_prev_block
= cpu_to_be32(log
->l_prev_block
);
2878 log
->l_prev_block
= log
->l_curr_block
;
2879 log
->l_prev_cycle
= log
->l_curr_cycle
;
2881 /* roll log?: ic_offset changed later */
2882 log
->l_curr_block
+= BTOBB(eventual_size
)+BTOBB(log
->l_iclog_hsize
);
2884 /* Round up to next log-sunit */
2885 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
2886 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
2887 __uint32_t sunit_bb
= BTOBB(log
->l_mp
->m_sb
.sb_logsunit
);
2888 log
->l_curr_block
= roundup(log
->l_curr_block
, sunit_bb
);
2891 if (log
->l_curr_block
>= log
->l_logBBsize
) {
2892 log
->l_curr_cycle
++;
2893 if (log
->l_curr_cycle
== XLOG_HEADER_MAGIC_NUM
)
2894 log
->l_curr_cycle
++;
2895 log
->l_curr_block
-= log
->l_logBBsize
;
2896 ASSERT(log
->l_curr_block
>= 0);
2898 ASSERT(iclog
== log
->l_iclog
);
2899 log
->l_iclog
= iclog
->ic_next
;
2900 } /* xlog_state_switch_iclogs */
2904 * Write out all data in the in-core log as of this exact moment in time.
2906 * Data may be written to the in-core log during this call. However,
2907 * we don't guarantee this data will be written out. A change from past
2908 * implementation means this routine will *not* write out zero length LRs.
2910 * Basically, we try and perform an intelligent scan of the in-core logs.
2911 * If we determine there is no flushable data, we just return. There is no
2912 * flushable data if:
2914 * 1. the current iclog is active and has no data; the previous iclog
2915 * is in the active or dirty state.
2916 * 2. the current iclog is drity, and the previous iclog is in the
2917 * active or dirty state.
2919 * We may sleep (call psema) if:
2921 * 1. the current iclog is not in the active nor dirty state.
2922 * 2. the current iclog dirty, and the previous iclog is not in the
2923 * active nor dirty state.
2924 * 3. the current iclog is active, and there is another thread writing
2925 * to this particular iclog.
2926 * 4. a) the current iclog is active and has no other writers
2927 * b) when we return from flushing out this iclog, it is still
2928 * not in the active nor dirty state.
2931 xlog_state_sync_all(xlog_t
*log
, uint flags
, int *log_flushed
)
2933 xlog_in_core_t
*iclog
;
2936 spin_lock(&log
->l_icloglock
);
2938 iclog
= log
->l_iclog
;
2939 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2940 spin_unlock(&log
->l_icloglock
);
2941 return XFS_ERROR(EIO
);
2944 /* If the head iclog is not active nor dirty, we just attach
2945 * ourselves to the head and go to sleep.
2947 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2948 iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2950 * If the head is dirty or (active and empty), then
2951 * we need to look at the previous iclog. If the previous
2952 * iclog is active or dirty we are done. There is nothing
2953 * to sync out. Otherwise, we attach ourselves to the
2954 * previous iclog and go to sleep.
2956 if (iclog
->ic_state
== XLOG_STATE_DIRTY
||
2957 (atomic_read(&iclog
->ic_refcnt
) == 0
2958 && iclog
->ic_offset
== 0)) {
2959 iclog
= iclog
->ic_prev
;
2960 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2961 iclog
->ic_state
== XLOG_STATE_DIRTY
)
2966 if (atomic_read(&iclog
->ic_refcnt
) == 0) {
2967 /* We are the only one with access to this
2968 * iclog. Flush it out now. There should
2969 * be a roundoff of zero to show that someone
2970 * has already taken care of the roundoff from
2971 * the previous sync.
2973 atomic_inc(&iclog
->ic_refcnt
);
2974 lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2975 xlog_state_switch_iclogs(log
, iclog
, 0);
2976 spin_unlock(&log
->l_icloglock
);
2978 if (xlog_state_release_iclog(log
, iclog
))
2979 return XFS_ERROR(EIO
);
2981 spin_lock(&log
->l_icloglock
);
2982 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) == lsn
&&
2983 iclog
->ic_state
!= XLOG_STATE_DIRTY
)
2988 /* Someone else is writing to this iclog.
2989 * Use its call to flush out the data. However,
2990 * the other thread may not force out this LR,
2991 * so we mark it WANT_SYNC.
2993 xlog_state_switch_iclogs(log
, iclog
, 0);
2999 /* By the time we come around again, the iclog could've been filled
3000 * which would give it another lsn. If we have a new lsn, just
3001 * return because the relevant data has been flushed.
3004 if (flags
& XFS_LOG_SYNC
) {
3006 * We must check if we're shutting down here, before
3007 * we wait, while we're holding the l_icloglock.
3008 * Then we check again after waking up, in case our
3009 * sleep was disturbed by a bad news.
3011 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3012 spin_unlock(&log
->l_icloglock
);
3013 return XFS_ERROR(EIO
);
3015 XFS_STATS_INC(xs_log_force_sleep
);
3016 sv_wait(&iclog
->ic_forcesema
, PINOD
, &log
->l_icloglock
, s
);
3018 * No need to grab the log lock here since we're
3019 * only deciding whether or not to return EIO
3020 * and the memory read should be atomic.
3022 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3023 return XFS_ERROR(EIO
);
3029 spin_unlock(&log
->l_icloglock
);
3032 } /* xlog_state_sync_all */
3036 * Used by code which implements synchronous log forces.
3038 * Find in-core log with lsn.
3039 * If it is in the DIRTY state, just return.
3040 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3041 * state and go to sleep or return.
3042 * If it is in any other state, go to sleep or return.
3044 * If filesystem activity goes to zero, the iclog will get flushed only by
3048 xlog_state_sync(xlog_t
*log
,
3053 xlog_in_core_t
*iclog
;
3054 int already_slept
= 0;
3057 spin_lock(&log
->l_icloglock
);
3058 iclog
= log
->l_iclog
;
3060 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3061 spin_unlock(&log
->l_icloglock
);
3062 return XFS_ERROR(EIO
);
3066 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
) {
3067 iclog
= iclog
->ic_next
;
3071 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
3072 spin_unlock(&log
->l_icloglock
);
3076 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3078 * We sleep here if we haven't already slept (e.g.
3079 * this is the first time we've looked at the correct
3080 * iclog buf) and the buffer before us is going to
3081 * be sync'ed. The reason for this is that if we
3082 * are doing sync transactions here, by waiting for
3083 * the previous I/O to complete, we can allow a few
3084 * more transactions into this iclog before we close
3087 * Otherwise, we mark the buffer WANT_SYNC, and bump
3088 * up the refcnt so we can release the log (which drops
3089 * the ref count). The state switch keeps new transaction
3090 * commits from using this buffer. When the current commits
3091 * finish writing into the buffer, the refcount will drop to
3092 * zero and the buffer will go out then.
3094 if (!already_slept
&&
3095 (iclog
->ic_prev
->ic_state
& (XLOG_STATE_WANT_SYNC
|
3096 XLOG_STATE_SYNCING
))) {
3097 ASSERT(!(iclog
->ic_state
& XLOG_STATE_IOERROR
));
3098 XFS_STATS_INC(xs_log_force_sleep
);
3099 sv_wait(&iclog
->ic_prev
->ic_writesema
, PSWP
,
3100 &log
->l_icloglock
, s
);
3105 atomic_inc(&iclog
->ic_refcnt
);
3106 xlog_state_switch_iclogs(log
, iclog
, 0);
3107 spin_unlock(&log
->l_icloglock
);
3108 if (xlog_state_release_iclog(log
, iclog
))
3109 return XFS_ERROR(EIO
);
3111 spin_lock(&log
->l_icloglock
);
3115 if ((flags
& XFS_LOG_SYNC
) && /* sleep */
3116 !(iclog
->ic_state
& (XLOG_STATE_ACTIVE
| XLOG_STATE_DIRTY
))) {
3119 * Don't wait on the forcesema if we know that we've
3120 * gotten a log write error.
3122 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3123 spin_unlock(&log
->l_icloglock
);
3124 return XFS_ERROR(EIO
);
3126 XFS_STATS_INC(xs_log_force_sleep
);
3127 sv_wait(&iclog
->ic_forcesema
, PSWP
, &log
->l_icloglock
, s
);
3129 * No need to grab the log lock here since we're
3130 * only deciding whether or not to return EIO
3131 * and the memory read should be atomic.
3133 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3134 return XFS_ERROR(EIO
);
3136 } else { /* just return */
3137 spin_unlock(&log
->l_icloglock
);
3141 } while (iclog
!= log
->l_iclog
);
3143 spin_unlock(&log
->l_icloglock
);
3145 } /* xlog_state_sync */
3149 * Called when we want to mark the current iclog as being ready to sync to
3153 xlog_state_want_sync(xlog_t
*log
, xlog_in_core_t
*iclog
)
3155 spin_lock(&log
->l_icloglock
);
3157 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3158 xlog_state_switch_iclogs(log
, iclog
, 0);
3160 ASSERT(iclog
->ic_state
&
3161 (XLOG_STATE_WANT_SYNC
|XLOG_STATE_IOERROR
));
3164 spin_unlock(&log
->l_icloglock
);
3165 } /* xlog_state_want_sync */
3169 /*****************************************************************************
3173 *****************************************************************************
3177 * Free a used ticket.
3180 xlog_ticket_put(xlog_t
*log
,
3181 xlog_ticket_t
*ticket
)
3183 sv_destroy(&ticket
->t_sema
);
3184 kmem_zone_free(xfs_log_ticket_zone
, ticket
);
3185 } /* xlog_ticket_put */
3189 * Allocate and initialise a new log ticket.
3191 STATIC xlog_ticket_t
*
3192 xlog_ticket_get(xlog_t
*log
,
3201 tic
= kmem_zone_zalloc(xfs_log_ticket_zone
, KM_SLEEP
|KM_MAYFAIL
);
3206 * Permanent reservations have up to 'cnt'-1 active log operations
3207 * in the log. A unit in this case is the amount of space for one
3208 * of these log operations. Normal reservations have a cnt of 1
3209 * and their unit amount is the total amount of space required.
3211 * The following lines of code account for non-transaction data
3212 * which occupy space in the on-disk log.
3214 * Normal form of a transaction is:
3215 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3216 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3218 * We need to account for all the leadup data and trailer data
3219 * around the transaction data.
3220 * And then we need to account for the worst case in terms of using
3222 * The worst case will happen if:
3223 * - the placement of the transaction happens to be such that the
3224 * roundoff is at its maximum
3225 * - the transaction data is synced before the commit record is synced
3226 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3227 * Therefore the commit record is in its own Log Record.
3228 * This can happen as the commit record is called with its
3229 * own region to xlog_write().
3230 * This then means that in the worst case, roundoff can happen for
3231 * the commit-rec as well.
3232 * The commit-rec is smaller than padding in this scenario and so it is
3233 * not added separately.
3236 /* for trans header */
3237 unit_bytes
+= sizeof(xlog_op_header_t
);
3238 unit_bytes
+= sizeof(xfs_trans_header_t
);
3241 unit_bytes
+= sizeof(xlog_op_header_t
);
3243 /* for LR headers */
3244 num_headers
= ((unit_bytes
+ log
->l_iclog_size
-1) >> log
->l_iclog_size_log
);
3245 unit_bytes
+= log
->l_iclog_hsize
* num_headers
;
3247 /* for commit-rec LR header - note: padding will subsume the ophdr */
3248 unit_bytes
+= log
->l_iclog_hsize
;
3250 /* for split-recs - ophdrs added when data split over LRs */
3251 unit_bytes
+= sizeof(xlog_op_header_t
) * num_headers
;
3253 /* for roundoff padding for transaction data and one for commit record */
3254 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
3255 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
3256 /* log su roundoff */
3257 unit_bytes
+= 2*log
->l_mp
->m_sb
.sb_logsunit
;
3260 unit_bytes
+= 2*BBSIZE
;
3263 tic
->t_unit_res
= unit_bytes
;
3264 tic
->t_curr_res
= unit_bytes
;
3267 tic
->t_tid
= (xlog_tid_t
)((__psint_t
)tic
& 0xffffffff);
3268 tic
->t_clientid
= client
;
3269 tic
->t_flags
= XLOG_TIC_INITED
;
3270 tic
->t_trans_type
= 0;
3271 if (xflags
& XFS_LOG_PERM_RESERV
)
3272 tic
->t_flags
|= XLOG_TIC_PERM_RESERV
;
3273 sv_init(&(tic
->t_sema
), SV_DEFAULT
, "logtick");
3275 xlog_tic_reset_res(tic
);
3278 } /* xlog_ticket_get */
3281 /******************************************************************************
3283 * Log debug routines
3285 ******************************************************************************
3289 * Make sure that the destination ptr is within the valid data region of
3290 * one of the iclogs. This uses backup pointers stored in a different
3291 * part of the log in case we trash the log structure.
3294 xlog_verify_dest_ptr(xlog_t
*log
,
3300 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
3301 if (ptr
>= (__psint_t
)log
->l_iclog_bak
[i
] &&
3302 ptr
<= (__psint_t
)log
->l_iclog_bak
[i
]+log
->l_iclog_size
)
3306 xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3307 } /* xlog_verify_dest_ptr */
3310 xlog_verify_grant_head(xlog_t
*log
, int equals
)
3312 if (log
->l_grant_reserve_cycle
== log
->l_grant_write_cycle
) {
3314 ASSERT(log
->l_grant_reserve_bytes
>= log
->l_grant_write_bytes
);
3316 ASSERT(log
->l_grant_reserve_bytes
> log
->l_grant_write_bytes
);
3318 ASSERT(log
->l_grant_reserve_cycle
-1 == log
->l_grant_write_cycle
);
3319 ASSERT(log
->l_grant_write_bytes
>= log
->l_grant_reserve_bytes
);
3321 } /* xlog_verify_grant_head */
3323 /* check if it will fit */
3325 xlog_verify_tail_lsn(xlog_t
*log
,
3326 xlog_in_core_t
*iclog
,
3331 if (CYCLE_LSN(tail_lsn
) == log
->l_prev_cycle
) {
3333 log
->l_logBBsize
- (log
->l_prev_block
- BLOCK_LSN(tail_lsn
));
3334 if (blocks
< BTOBB(iclog
->ic_offset
)+BTOBB(log
->l_iclog_hsize
))
3335 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3337 ASSERT(CYCLE_LSN(tail_lsn
)+1 == log
->l_prev_cycle
);
3339 if (BLOCK_LSN(tail_lsn
) == log
->l_prev_block
)
3340 xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3342 blocks
= BLOCK_LSN(tail_lsn
) - log
->l_prev_block
;
3343 if (blocks
< BTOBB(iclog
->ic_offset
) + 1)
3344 xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3346 } /* xlog_verify_tail_lsn */
3349 * Perform a number of checks on the iclog before writing to disk.
3351 * 1. Make sure the iclogs are still circular
3352 * 2. Make sure we have a good magic number
3353 * 3. Make sure we don't have magic numbers in the data
3354 * 4. Check fields of each log operation header for:
3355 * A. Valid client identifier
3356 * B. tid ptr value falls in valid ptr space (user space code)
3357 * C. Length in log record header is correct according to the
3358 * individual operation headers within record.
3359 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3360 * log, check the preceding blocks of the physical log to make sure all
3361 * the cycle numbers agree with the current cycle number.
3364 xlog_verify_iclog(xlog_t
*log
,
3365 xlog_in_core_t
*iclog
,
3369 xlog_op_header_t
*ophead
;
3370 xlog_in_core_t
*icptr
;
3371 xlog_in_core_2_t
*xhdr
;
3373 xfs_caddr_t base_ptr
;
3374 __psint_t field_offset
;
3376 int len
, i
, j
, k
, op_len
;
3379 /* check validity of iclog pointers */
3380 spin_lock(&log
->l_icloglock
);
3381 icptr
= log
->l_iclog
;
3382 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
3384 xlog_panic("xlog_verify_iclog: invalid ptr");
3385 icptr
= icptr
->ic_next
;
3387 if (icptr
!= log
->l_iclog
)
3388 xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3389 spin_unlock(&log
->l_icloglock
);
3391 /* check log magic numbers */
3392 if (be32_to_cpu(iclog
->ic_header
.h_magicno
) != XLOG_HEADER_MAGIC_NUM
)
3393 xlog_panic("xlog_verify_iclog: invalid magic num");
3395 ptr
= (xfs_caddr_t
) &iclog
->ic_header
;
3396 for (ptr
+= BBSIZE
; ptr
< ((xfs_caddr_t
)&iclog
->ic_header
) + count
;
3398 if (be32_to_cpu(*(__be32
*)ptr
) == XLOG_HEADER_MAGIC_NUM
)
3399 xlog_panic("xlog_verify_iclog: unexpected magic num");
3403 len
= be32_to_cpu(iclog
->ic_header
.h_num_logops
);
3404 ptr
= iclog
->ic_datap
;
3406 ophead
= (xlog_op_header_t
*)ptr
;
3407 xhdr
= (xlog_in_core_2_t
*)&iclog
->ic_header
;
3408 for (i
= 0; i
< len
; i
++) {
3409 ophead
= (xlog_op_header_t
*)ptr
;
3411 /* clientid is only 1 byte */
3412 field_offset
= (__psint_t
)
3413 ((xfs_caddr_t
)&(ophead
->oh_clientid
) - base_ptr
);
3414 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3415 clientid
= ophead
->oh_clientid
;
3417 idx
= BTOBBT((xfs_caddr_t
)&(ophead
->oh_clientid
) - iclog
->ic_datap
);
3418 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3419 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3420 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3421 clientid
= xlog_get_client_id(
3422 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3424 clientid
= xlog_get_client_id(
3425 iclog
->ic_header
.h_cycle_data
[idx
]);
3428 if (clientid
!= XFS_TRANSACTION
&& clientid
!= XFS_LOG
)
3429 cmn_err(CE_WARN
, "xlog_verify_iclog: "
3430 "invalid clientid %d op 0x%p offset 0x%lx",
3431 clientid
, ophead
, (unsigned long)field_offset
);
3434 field_offset
= (__psint_t
)
3435 ((xfs_caddr_t
)&(ophead
->oh_len
) - base_ptr
);
3436 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3437 op_len
= be32_to_cpu(ophead
->oh_len
);
3439 idx
= BTOBBT((__psint_t
)&ophead
->oh_len
-
3440 (__psint_t
)iclog
->ic_datap
);
3441 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3442 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3443 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3444 op_len
= be32_to_cpu(xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3446 op_len
= be32_to_cpu(iclog
->ic_header
.h_cycle_data
[idx
]);
3449 ptr
+= sizeof(xlog_op_header_t
) + op_len
;
3451 } /* xlog_verify_iclog */
3455 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3461 xlog_in_core_t
*iclog
, *ic
;
3463 iclog
= log
->l_iclog
;
3464 if (! (iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3466 * Mark all the incore logs IOERROR.
3467 * From now on, no log flushes will result.
3471 ic
->ic_state
= XLOG_STATE_IOERROR
;
3473 } while (ic
!= iclog
);
3477 * Return non-zero, if state transition has already happened.
3483 * This is called from xfs_force_shutdown, when we're forcibly
3484 * shutting down the filesystem, typically because of an IO error.
3485 * Our main objectives here are to make sure that:
3486 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3487 * parties to find out, 'atomically'.
3488 * b. those who're sleeping on log reservations, pinned objects and
3489 * other resources get woken up, and be told the bad news.
3490 * c. nothing new gets queued up after (a) and (b) are done.
3491 * d. if !logerror, flush the iclogs to disk, then seal them off
3495 xfs_log_force_umount(
3496 struct xfs_mount
*mp
,
3507 * If this happens during log recovery, don't worry about
3508 * locking; the log isn't open for business yet.
3511 log
->l_flags
& XLOG_ACTIVE_RECOVERY
) {
3512 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3513 XFS_BUF_DONE(mp
->m_sb_bp
);
3518 * Somebody could've already done the hard work for us.
3519 * No need to get locks for this.
3521 if (logerror
&& log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3522 ASSERT(XLOG_FORCED_SHUTDOWN(log
));
3527 * We must hold both the GRANT lock and the LOG lock,
3528 * before we mark the filesystem SHUTDOWN and wake
3529 * everybody up to tell the bad news.
3531 spin_lock(&log
->l_icloglock
);
3532 spin_lock(&log
->l_grant_lock
);
3533 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3534 XFS_BUF_DONE(mp
->m_sb_bp
);
3536 * This flag is sort of redundant because of the mount flag, but
3537 * it's good to maintain the separation between the log and the rest
3540 log
->l_flags
|= XLOG_IO_ERROR
;
3543 * If we hit a log error, we want to mark all the iclogs IOERROR
3544 * while we're still holding the loglock.
3547 retval
= xlog_state_ioerror(log
);
3548 spin_unlock(&log
->l_icloglock
);
3551 * We don't want anybody waiting for log reservations
3552 * after this. That means we have to wake up everybody
3553 * queued up on reserve_headq as well as write_headq.
3554 * In addition, we make sure in xlog_{re}grant_log_space
3555 * that we don't enqueue anything once the SHUTDOWN flag
3556 * is set, and this action is protected by the GRANTLOCK.
3558 if ((tic
= log
->l_reserve_headq
)) {
3560 sv_signal(&tic
->t_sema
);
3562 } while (tic
!= log
->l_reserve_headq
);
3565 if ((tic
= log
->l_write_headq
)) {
3567 sv_signal(&tic
->t_sema
);
3569 } while (tic
!= log
->l_write_headq
);
3571 spin_unlock(&log
->l_grant_lock
);
3573 if (! (log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3576 * Force the incore logs to disk before shutting the
3577 * log down completely.
3579 xlog_state_sync_all(log
, XFS_LOG_FORCE
|XFS_LOG_SYNC
, &dummy
);
3580 spin_lock(&log
->l_icloglock
);
3581 retval
= xlog_state_ioerror(log
);
3582 spin_unlock(&log
->l_icloglock
);
3585 * Wake up everybody waiting on xfs_log_force.
3586 * Callback all log item committed functions as if the
3587 * log writes were completed.
3589 xlog_state_do_callback(log
, XFS_LI_ABORTED
, NULL
);
3591 #ifdef XFSERRORDEBUG
3593 xlog_in_core_t
*iclog
;
3595 spin_lock(&log
->l_icloglock
);
3596 iclog
= log
->l_iclog
;
3598 ASSERT(iclog
->ic_callback
== 0);
3599 iclog
= iclog
->ic_next
;
3600 } while (iclog
!= log
->l_iclog
);
3601 spin_unlock(&log
->l_icloglock
);
3604 /* return non-zero if log IOERROR transition had already happened */
3609 xlog_iclogs_empty(xlog_t
*log
)
3611 xlog_in_core_t
*iclog
;
3613 iclog
= log
->l_iclog
;
3615 /* endianness does not matter here, zero is zero in
3618 if (iclog
->ic_header
.h_num_logops
)
3620 iclog
= iclog
->ic_next
;
3621 } while (iclog
!= log
->l_iclog
);