Loading...
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_shared.h"
21#include "xfs_format.h"
22#include "xfs_log_format.h"
23#include "xfs_trans_resv.h"
24#include "xfs_sb.h"
25#include "xfs_ag.h"
26#include "xfs_mount.h"
27#include "xfs_error.h"
28#include "xfs_trans.h"
29#include "xfs_trans_priv.h"
30#include "xfs_log.h"
31#include "xfs_log_priv.h"
32#include "xfs_log_recover.h"
33#include "xfs_inode.h"
34#include "xfs_trace.h"
35#include "xfs_fsops.h"
36#include "xfs_cksum.h"
37
38kmem_zone_t *xfs_log_ticket_zone;
39
40/* Local miscellaneous function prototypes */
41STATIC int
42xlog_commit_record(
43 struct xlog *log,
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
47
48STATIC struct xlog *
49xlog_alloc_log(
50 struct xfs_mount *mp,
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
53 int num_bblks);
54STATIC int
55xlog_space_left(
56 struct xlog *log,
57 atomic64_t *head);
58STATIC int
59xlog_sync(
60 struct xlog *log,
61 struct xlog_in_core *iclog);
62STATIC void
63xlog_dealloc_log(
64 struct xlog *log);
65
66/* local state machine functions */
67STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
68STATIC void
69xlog_state_do_callback(
70 struct xlog *log,
71 int aborted,
72 struct xlog_in_core *iclog);
73STATIC int
74xlog_state_get_iclog_space(
75 struct xlog *log,
76 int len,
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
79 int *continued_write,
80 int *logoffsetp);
81STATIC int
82xlog_state_release_iclog(
83 struct xlog *log,
84 struct xlog_in_core *iclog);
85STATIC void
86xlog_state_switch_iclogs(
87 struct xlog *log,
88 struct xlog_in_core *iclog,
89 int eventual_size);
90STATIC void
91xlog_state_want_sync(
92 struct xlog *log,
93 struct xlog_in_core *iclog);
94
95STATIC void
96xlog_grant_push_ail(
97 struct xlog *log,
98 int need_bytes);
99STATIC void
100xlog_regrant_reserve_log_space(
101 struct xlog *log,
102 struct xlog_ticket *ticket);
103STATIC void
104xlog_ungrant_log_space(
105 struct xlog *log,
106 struct xlog_ticket *ticket);
107
108#if defined(DEBUG)
109STATIC void
110xlog_verify_dest_ptr(
111 struct xlog *log,
112 char *ptr);
113STATIC void
114xlog_verify_grant_tail(
115 struct xlog *log);
116STATIC void
117xlog_verify_iclog(
118 struct xlog *log,
119 struct xlog_in_core *iclog,
120 int count,
121 bool syncing);
122STATIC void
123xlog_verify_tail_lsn(
124 struct xlog *log,
125 struct xlog_in_core *iclog,
126 xfs_lsn_t tail_lsn);
127#else
128#define xlog_verify_dest_ptr(a,b)
129#define xlog_verify_grant_tail(a)
130#define xlog_verify_iclog(a,b,c,d)
131#define xlog_verify_tail_lsn(a,b,c)
132#endif
133
134STATIC int
135xlog_iclogs_empty(
136 struct xlog *log);
137
138static void
139xlog_grant_sub_space(
140 struct xlog *log,
141 atomic64_t *head,
142 int bytes)
143{
144 int64_t head_val = atomic64_read(head);
145 int64_t new, old;
146
147 do {
148 int cycle, space;
149
150 xlog_crack_grant_head_val(head_val, &cycle, &space);
151
152 space -= bytes;
153 if (space < 0) {
154 space += log->l_logsize;
155 cycle--;
156 }
157
158 old = head_val;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
162}
163
164static void
165xlog_grant_add_space(
166 struct xlog *log,
167 atomic64_t *head,
168 int bytes)
169{
170 int64_t head_val = atomic64_read(head);
171 int64_t new, old;
172
173 do {
174 int tmp;
175 int cycle, space;
176
177 xlog_crack_grant_head_val(head_val, &cycle, &space);
178
179 tmp = log->l_logsize - space;
180 if (tmp > bytes)
181 space += bytes;
182 else {
183 space = bytes - tmp;
184 cycle++;
185 }
186
187 old = head_val;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
191}
192
193STATIC void
194xlog_grant_head_init(
195 struct xlog_grant_head *head)
196{
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
200}
201
202STATIC void
203xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
205{
206 struct xlog_ticket *tic;
207
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
212}
213
214static inline int
215xlog_ticket_reservation(
216 struct xlog *log,
217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
219{
220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
223 } else {
224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
225 return tic->t_unit_res * tic->t_cnt;
226 else
227 return tic->t_unit_res;
228 }
229}
230
231STATIC bool
232xlog_grant_head_wake(
233 struct xlog *log,
234 struct xlog_grant_head *head,
235 int *free_bytes)
236{
237 struct xlog_ticket *tic;
238 int need_bytes;
239
240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
242 if (*free_bytes < need_bytes)
243 return false;
244
245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
247 wake_up_process(tic->t_task);
248 }
249
250 return true;
251}
252
253STATIC int
254xlog_grant_head_wait(
255 struct xlog *log,
256 struct xlog_grant_head *head,
257 struct xlog_ticket *tic,
258 int need_bytes) __releases(&head->lock)
259 __acquires(&head->lock)
260{
261 list_add_tail(&tic->t_queue, &head->waiters);
262
263 do {
264 if (XLOG_FORCED_SHUTDOWN(log))
265 goto shutdown;
266 xlog_grant_push_ail(log, need_bytes);
267
268 __set_current_state(TASK_UNINTERRUPTIBLE);
269 spin_unlock(&head->lock);
270
271 XFS_STATS_INC(xs_sleep_logspace);
272
273 trace_xfs_log_grant_sleep(log, tic);
274 schedule();
275 trace_xfs_log_grant_wake(log, tic);
276
277 spin_lock(&head->lock);
278 if (XLOG_FORCED_SHUTDOWN(log))
279 goto shutdown;
280 } while (xlog_space_left(log, &head->grant) < need_bytes);
281
282 list_del_init(&tic->t_queue);
283 return 0;
284shutdown:
285 list_del_init(&tic->t_queue);
286 return XFS_ERROR(EIO);
287}
288
289/*
290 * Atomically get the log space required for a log ticket.
291 *
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
294 *
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
298 * every pass.
299 *
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
305 */
306STATIC int
307xlog_grant_head_check(
308 struct xlog *log,
309 struct xlog_grant_head *head,
310 struct xlog_ticket *tic,
311 int *need_bytes)
312{
313 int free_bytes;
314 int error = 0;
315
316 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
317
318 /*
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
323 */
324 *need_bytes = xlog_ticket_reservation(log, head, tic);
325 free_bytes = xlog_space_left(log, &head->grant);
326 if (!list_empty_careful(&head->waiters)) {
327 spin_lock(&head->lock);
328 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
329 free_bytes < *need_bytes) {
330 error = xlog_grant_head_wait(log, head, tic,
331 *need_bytes);
332 }
333 spin_unlock(&head->lock);
334 } else if (free_bytes < *need_bytes) {
335 spin_lock(&head->lock);
336 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
337 spin_unlock(&head->lock);
338 }
339
340 return error;
341}
342
343static void
344xlog_tic_reset_res(xlog_ticket_t *tic)
345{
346 tic->t_res_num = 0;
347 tic->t_res_arr_sum = 0;
348 tic->t_res_num_ophdrs = 0;
349}
350
351static void
352xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
353{
354 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
355 /* add to overflow and start again */
356 tic->t_res_o_flow += tic->t_res_arr_sum;
357 tic->t_res_num = 0;
358 tic->t_res_arr_sum = 0;
359 }
360
361 tic->t_res_arr[tic->t_res_num].r_len = len;
362 tic->t_res_arr[tic->t_res_num].r_type = type;
363 tic->t_res_arr_sum += len;
364 tic->t_res_num++;
365}
366
367/*
368 * Replenish the byte reservation required by moving the grant write head.
369 */
370int
371xfs_log_regrant(
372 struct xfs_mount *mp,
373 struct xlog_ticket *tic)
374{
375 struct xlog *log = mp->m_log;
376 int need_bytes;
377 int error = 0;
378
379 if (XLOG_FORCED_SHUTDOWN(log))
380 return XFS_ERROR(EIO);
381
382 XFS_STATS_INC(xs_try_logspace);
383
384 /*
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
389 */
390 tic->t_tid++;
391
392 xlog_grant_push_ail(log, tic->t_unit_res);
393
394 tic->t_curr_res = tic->t_unit_res;
395 xlog_tic_reset_res(tic);
396
397 if (tic->t_cnt > 0)
398 return 0;
399
400 trace_xfs_log_regrant(log, tic);
401
402 error = xlog_grant_head_check(log, &log->l_write_head, tic,
403 &need_bytes);
404 if (error)
405 goto out_error;
406
407 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
408 trace_xfs_log_regrant_exit(log, tic);
409 xlog_verify_grant_tail(log);
410 return 0;
411
412out_error:
413 /*
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
417 */
418 tic->t_curr_res = 0;
419 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
420 return error;
421}
422
423/*
424 * Reserve log space and return a ticket corresponding the reservation.
425 *
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
430 */
431int
432xfs_log_reserve(
433 struct xfs_mount *mp,
434 int unit_bytes,
435 int cnt,
436 struct xlog_ticket **ticp,
437 __uint8_t client,
438 bool permanent,
439 uint t_type)
440{
441 struct xlog *log = mp->m_log;
442 struct xlog_ticket *tic;
443 int need_bytes;
444 int error = 0;
445
446 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
447
448 if (XLOG_FORCED_SHUTDOWN(log))
449 return XFS_ERROR(EIO);
450
451 XFS_STATS_INC(xs_try_logspace);
452
453 ASSERT(*ticp == NULL);
454 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
455 KM_SLEEP | KM_MAYFAIL);
456 if (!tic)
457 return XFS_ERROR(ENOMEM);
458
459 tic->t_trans_type = t_type;
460 *ticp = tic;
461
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
463 : tic->t_unit_res);
464
465 trace_xfs_log_reserve(log, tic);
466
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
468 &need_bytes);
469 if (error)
470 goto out_error;
471
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
476 return 0;
477
478out_error:
479 /*
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
483 */
484 tic->t_curr_res = 0;
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
486 return error;
487}
488
489
490/*
491 * NOTES:
492 *
493 * 1. currblock field gets updated at startup and after in-core logs
494 * marked as with WANT_SYNC.
495 */
496
497/*
498 * This routine is called when a user of a log manager ticket is done with
499 * the reservation. If the ticket was ever used, then a commit record for
500 * the associated transaction is written out as a log operation header with
501 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
502 * a given ticket. If the ticket was one with a permanent reservation, then
503 * a few operations are done differently. Permanent reservation tickets by
504 * default don't release the reservation. They just commit the current
505 * transaction with the belief that the reservation is still needed. A flag
506 * must be passed in before permanent reservations are actually released.
507 * When these type of tickets are not released, they need to be set into
508 * the inited state again. By doing this, a start record will be written
509 * out when the next write occurs.
510 */
511xfs_lsn_t
512xfs_log_done(
513 struct xfs_mount *mp,
514 struct xlog_ticket *ticket,
515 struct xlog_in_core **iclog,
516 uint flags)
517{
518 struct xlog *log = mp->m_log;
519 xfs_lsn_t lsn = 0;
520
521 if (XLOG_FORCED_SHUTDOWN(log) ||
522 /*
523 * If nothing was ever written, don't write out commit record.
524 * If we get an error, just continue and give back the log ticket.
525 */
526 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
527 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
528 lsn = (xfs_lsn_t) -1;
529 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
530 flags |= XFS_LOG_REL_PERM_RESERV;
531 }
532 }
533
534
535 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
536 (flags & XFS_LOG_REL_PERM_RESERV)) {
537 trace_xfs_log_done_nonperm(log, ticket);
538
539 /*
540 * Release ticket if not permanent reservation or a specific
541 * request has been made to release a permanent reservation.
542 */
543 xlog_ungrant_log_space(log, ticket);
544 xfs_log_ticket_put(ticket);
545 } else {
546 trace_xfs_log_done_perm(log, ticket);
547
548 xlog_regrant_reserve_log_space(log, ticket);
549 /* If this ticket was a permanent reservation and we aren't
550 * trying to release it, reset the inited flags; so next time
551 * we write, a start record will be written out.
552 */
553 ticket->t_flags |= XLOG_TIC_INITED;
554 }
555
556 return lsn;
557}
558
559/*
560 * Attaches a new iclog I/O completion callback routine during
561 * transaction commit. If the log is in error state, a non-zero
562 * return code is handed back and the caller is responsible for
563 * executing the callback at an appropriate time.
564 */
565int
566xfs_log_notify(
567 struct xfs_mount *mp,
568 struct xlog_in_core *iclog,
569 xfs_log_callback_t *cb)
570{
571 int abortflg;
572
573 spin_lock(&iclog->ic_callback_lock);
574 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
575 if (!abortflg) {
576 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
577 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
578 cb->cb_next = NULL;
579 *(iclog->ic_callback_tail) = cb;
580 iclog->ic_callback_tail = &(cb->cb_next);
581 }
582 spin_unlock(&iclog->ic_callback_lock);
583 return abortflg;
584}
585
586int
587xfs_log_release_iclog(
588 struct xfs_mount *mp,
589 struct xlog_in_core *iclog)
590{
591 if (xlog_state_release_iclog(mp->m_log, iclog)) {
592 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
593 return EIO;
594 }
595
596 return 0;
597}
598
599/*
600 * Mount a log filesystem
601 *
602 * mp - ubiquitous xfs mount point structure
603 * log_target - buftarg of on-disk log device
604 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
605 * num_bblocks - Number of BBSIZE blocks in on-disk log
606 *
607 * Return error or zero.
608 */
609int
610xfs_log_mount(
611 xfs_mount_t *mp,
612 xfs_buftarg_t *log_target,
613 xfs_daddr_t blk_offset,
614 int num_bblks)
615{
616 int error = 0;
617 int min_logfsbs;
618
619 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
620 xfs_notice(mp, "Mounting V%d Filesystem",
621 XFS_SB_VERSION_NUM(&mp->m_sb));
622 } else {
623 xfs_notice(mp,
624"Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
625 XFS_SB_VERSION_NUM(&mp->m_sb));
626 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
627 }
628
629 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
630 if (IS_ERR(mp->m_log)) {
631 error = -PTR_ERR(mp->m_log);
632 goto out;
633 }
634
635 /*
636 * Validate the given log space and drop a critical message via syslog
637 * if the log size is too small that would lead to some unexpected
638 * situations in transaction log space reservation stage.
639 *
640 * Note: we can't just reject the mount if the validation fails. This
641 * would mean that people would have to downgrade their kernel just to
642 * remedy the situation as there is no way to grow the log (short of
643 * black magic surgery with xfs_db).
644 *
645 * We can, however, reject mounts for CRC format filesystems, as the
646 * mkfs binary being used to make the filesystem should never create a
647 * filesystem with a log that is too small.
648 */
649 min_logfsbs = xfs_log_calc_minimum_size(mp);
650
651 if (mp->m_sb.sb_logblocks < min_logfsbs) {
652 xfs_warn(mp,
653 "Log size %d blocks too small, minimum size is %d blocks",
654 mp->m_sb.sb_logblocks, min_logfsbs);
655 error = EINVAL;
656 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
657 xfs_warn(mp,
658 "Log size %d blocks too large, maximum size is %lld blocks",
659 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
660 error = EINVAL;
661 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
662 xfs_warn(mp,
663 "log size %lld bytes too large, maximum size is %lld bytes",
664 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
665 XFS_MAX_LOG_BYTES);
666 error = EINVAL;
667 }
668 if (error) {
669 if (xfs_sb_version_hascrc(&mp->m_sb)) {
670 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
671 ASSERT(0);
672 goto out_free_log;
673 }
674 xfs_crit(mp,
675"Log size out of supported range. Continuing onwards, but if log hangs are\n"
676"experienced then please report this message in the bug report.");
677 }
678
679 /*
680 * Initialize the AIL now we have a log.
681 */
682 error = xfs_trans_ail_init(mp);
683 if (error) {
684 xfs_warn(mp, "AIL initialisation failed: error %d", error);
685 goto out_free_log;
686 }
687 mp->m_log->l_ailp = mp->m_ail;
688
689 /*
690 * skip log recovery on a norecovery mount. pretend it all
691 * just worked.
692 */
693 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
694 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
695
696 if (readonly)
697 mp->m_flags &= ~XFS_MOUNT_RDONLY;
698
699 error = xlog_recover(mp->m_log);
700
701 if (readonly)
702 mp->m_flags |= XFS_MOUNT_RDONLY;
703 if (error) {
704 xfs_warn(mp, "log mount/recovery failed: error %d",
705 error);
706 goto out_destroy_ail;
707 }
708 }
709
710 /* Normal transactions can now occur */
711 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
712
713 /*
714 * Now the log has been fully initialised and we know were our
715 * space grant counters are, we can initialise the permanent ticket
716 * needed for delayed logging to work.
717 */
718 xlog_cil_init_post_recovery(mp->m_log);
719
720 return 0;
721
722out_destroy_ail:
723 xfs_trans_ail_destroy(mp);
724out_free_log:
725 xlog_dealloc_log(mp->m_log);
726out:
727 return error;
728}
729
730/*
731 * Finish the recovery of the file system. This is separate from the
732 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
733 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
734 * here.
735 *
736 * If we finish recovery successfully, start the background log work. If we are
737 * not doing recovery, then we have a RO filesystem and we don't need to start
738 * it.
739 */
740int
741xfs_log_mount_finish(xfs_mount_t *mp)
742{
743 int error = 0;
744
745 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
746 error = xlog_recover_finish(mp->m_log);
747 if (!error)
748 xfs_log_work_queue(mp);
749 } else {
750 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
751 }
752
753
754 return error;
755}
756
757/*
758 * Final log writes as part of unmount.
759 *
760 * Mark the filesystem clean as unmount happens. Note that during relocation
761 * this routine needs to be executed as part of source-bag while the
762 * deallocation must not be done until source-end.
763 */
764
765/*
766 * Unmount record used to have a string "Unmount filesystem--" in the
767 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
768 * We just write the magic number now since that particular field isn't
769 * currently architecture converted and "Unmount" is a bit foo.
770 * As far as I know, there weren't any dependencies on the old behaviour.
771 */
772
773int
774xfs_log_unmount_write(xfs_mount_t *mp)
775{
776 struct xlog *log = mp->m_log;
777 xlog_in_core_t *iclog;
778#ifdef DEBUG
779 xlog_in_core_t *first_iclog;
780#endif
781 xlog_ticket_t *tic = NULL;
782 xfs_lsn_t lsn;
783 int error;
784
785 /*
786 * Don't write out unmount record on read-only mounts.
787 * Or, if we are doing a forced umount (typically because of IO errors).
788 */
789 if (mp->m_flags & XFS_MOUNT_RDONLY)
790 return 0;
791
792 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
793 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
794
795#ifdef DEBUG
796 first_iclog = iclog = log->l_iclog;
797 do {
798 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
799 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
800 ASSERT(iclog->ic_offset == 0);
801 }
802 iclog = iclog->ic_next;
803 } while (iclog != first_iclog);
804#endif
805 if (! (XLOG_FORCED_SHUTDOWN(log))) {
806 error = xfs_log_reserve(mp, 600, 1, &tic,
807 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
808 if (!error) {
809 /* the data section must be 32 bit size aligned */
810 struct {
811 __uint16_t magic;
812 __uint16_t pad1;
813 __uint32_t pad2; /* may as well make it 64 bits */
814 } magic = {
815 .magic = XLOG_UNMOUNT_TYPE,
816 };
817 struct xfs_log_iovec reg = {
818 .i_addr = &magic,
819 .i_len = sizeof(magic),
820 .i_type = XLOG_REG_TYPE_UNMOUNT,
821 };
822 struct xfs_log_vec vec = {
823 .lv_niovecs = 1,
824 .lv_iovecp = ®,
825 };
826
827 /* remove inited flag, and account for space used */
828 tic->t_flags = 0;
829 tic->t_curr_res -= sizeof(magic);
830 error = xlog_write(log, &vec, tic, &lsn,
831 NULL, XLOG_UNMOUNT_TRANS);
832 /*
833 * At this point, we're umounting anyway,
834 * so there's no point in transitioning log state
835 * to IOERROR. Just continue...
836 */
837 }
838
839 if (error)
840 xfs_alert(mp, "%s: unmount record failed", __func__);
841
842
843 spin_lock(&log->l_icloglock);
844 iclog = log->l_iclog;
845 atomic_inc(&iclog->ic_refcnt);
846 xlog_state_want_sync(log, iclog);
847 spin_unlock(&log->l_icloglock);
848 error = xlog_state_release_iclog(log, iclog);
849
850 spin_lock(&log->l_icloglock);
851 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
852 iclog->ic_state == XLOG_STATE_DIRTY)) {
853 if (!XLOG_FORCED_SHUTDOWN(log)) {
854 xlog_wait(&iclog->ic_force_wait,
855 &log->l_icloglock);
856 } else {
857 spin_unlock(&log->l_icloglock);
858 }
859 } else {
860 spin_unlock(&log->l_icloglock);
861 }
862 if (tic) {
863 trace_xfs_log_umount_write(log, tic);
864 xlog_ungrant_log_space(log, tic);
865 xfs_log_ticket_put(tic);
866 }
867 } else {
868 /*
869 * We're already in forced_shutdown mode, couldn't
870 * even attempt to write out the unmount transaction.
871 *
872 * Go through the motions of sync'ing and releasing
873 * the iclog, even though no I/O will actually happen,
874 * we need to wait for other log I/Os that may already
875 * be in progress. Do this as a separate section of
876 * code so we'll know if we ever get stuck here that
877 * we're in this odd situation of trying to unmount
878 * a file system that went into forced_shutdown as
879 * the result of an unmount..
880 */
881 spin_lock(&log->l_icloglock);
882 iclog = log->l_iclog;
883 atomic_inc(&iclog->ic_refcnt);
884
885 xlog_state_want_sync(log, iclog);
886 spin_unlock(&log->l_icloglock);
887 error = xlog_state_release_iclog(log, iclog);
888
889 spin_lock(&log->l_icloglock);
890
891 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
892 || iclog->ic_state == XLOG_STATE_DIRTY
893 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
894
895 xlog_wait(&iclog->ic_force_wait,
896 &log->l_icloglock);
897 } else {
898 spin_unlock(&log->l_icloglock);
899 }
900 }
901
902 return error;
903} /* xfs_log_unmount_write */
904
905/*
906 * Empty the log for unmount/freeze.
907 *
908 * To do this, we first need to shut down the background log work so it is not
909 * trying to cover the log as we clean up. We then need to unpin all objects in
910 * the log so we can then flush them out. Once they have completed their IO and
911 * run the callbacks removing themselves from the AIL, we can write the unmount
912 * record.
913 */
914void
915xfs_log_quiesce(
916 struct xfs_mount *mp)
917{
918 cancel_delayed_work_sync(&mp->m_log->l_work);
919 xfs_log_force(mp, XFS_LOG_SYNC);
920
921 /*
922 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
923 * will push it, xfs_wait_buftarg() will not wait for it. Further,
924 * xfs_buf_iowait() cannot be used because it was pushed with the
925 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
926 * the IO to complete.
927 */
928 xfs_ail_push_all_sync(mp->m_ail);
929 xfs_wait_buftarg(mp->m_ddev_targp);
930 xfs_buf_lock(mp->m_sb_bp);
931 xfs_buf_unlock(mp->m_sb_bp);
932
933 xfs_log_unmount_write(mp);
934}
935
936/*
937 * Shut down and release the AIL and Log.
938 *
939 * During unmount, we need to ensure we flush all the dirty metadata objects
940 * from the AIL so that the log is empty before we write the unmount record to
941 * the log. Once this is done, we can tear down the AIL and the log.
942 */
943void
944xfs_log_unmount(
945 struct xfs_mount *mp)
946{
947 xfs_log_quiesce(mp);
948
949 xfs_trans_ail_destroy(mp);
950 xlog_dealloc_log(mp->m_log);
951}
952
953void
954xfs_log_item_init(
955 struct xfs_mount *mp,
956 struct xfs_log_item *item,
957 int type,
958 const struct xfs_item_ops *ops)
959{
960 item->li_mountp = mp;
961 item->li_ailp = mp->m_ail;
962 item->li_type = type;
963 item->li_ops = ops;
964 item->li_lv = NULL;
965
966 INIT_LIST_HEAD(&item->li_ail);
967 INIT_LIST_HEAD(&item->li_cil);
968}
969
970/*
971 * Wake up processes waiting for log space after we have moved the log tail.
972 */
973void
974xfs_log_space_wake(
975 struct xfs_mount *mp)
976{
977 struct xlog *log = mp->m_log;
978 int free_bytes;
979
980 if (XLOG_FORCED_SHUTDOWN(log))
981 return;
982
983 if (!list_empty_careful(&log->l_write_head.waiters)) {
984 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
985
986 spin_lock(&log->l_write_head.lock);
987 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
988 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
989 spin_unlock(&log->l_write_head.lock);
990 }
991
992 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
993 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
994
995 spin_lock(&log->l_reserve_head.lock);
996 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
997 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
998 spin_unlock(&log->l_reserve_head.lock);
999 }
1000}
1001
1002/*
1003 * Determine if we have a transaction that has gone to disk that needs to be
1004 * covered. To begin the transition to the idle state firstly the log needs to
1005 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1006 * we start attempting to cover the log.
1007 *
1008 * Only if we are then in a state where covering is needed, the caller is
1009 * informed that dummy transactions are required to move the log into the idle
1010 * state.
1011 *
1012 * If there are any items in the AIl or CIL, then we do not want to attempt to
1013 * cover the log as we may be in a situation where there isn't log space
1014 * available to run a dummy transaction and this can lead to deadlocks when the
1015 * tail of the log is pinned by an item that is modified in the CIL. Hence
1016 * there's no point in running a dummy transaction at this point because we
1017 * can't start trying to idle the log until both the CIL and AIL are empty.
1018 */
1019int
1020xfs_log_need_covered(xfs_mount_t *mp)
1021{
1022 struct xlog *log = mp->m_log;
1023 int needed = 0;
1024
1025 if (!xfs_fs_writable(mp))
1026 return 0;
1027
1028 if (!xlog_cil_empty(log))
1029 return 0;
1030
1031 spin_lock(&log->l_icloglock);
1032 switch (log->l_covered_state) {
1033 case XLOG_STATE_COVER_DONE:
1034 case XLOG_STATE_COVER_DONE2:
1035 case XLOG_STATE_COVER_IDLE:
1036 break;
1037 case XLOG_STATE_COVER_NEED:
1038 case XLOG_STATE_COVER_NEED2:
1039 if (xfs_ail_min_lsn(log->l_ailp))
1040 break;
1041 if (!xlog_iclogs_empty(log))
1042 break;
1043
1044 needed = 1;
1045 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1046 log->l_covered_state = XLOG_STATE_COVER_DONE;
1047 else
1048 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1049 break;
1050 default:
1051 needed = 1;
1052 break;
1053 }
1054 spin_unlock(&log->l_icloglock);
1055 return needed;
1056}
1057
1058/*
1059 * We may be holding the log iclog lock upon entering this routine.
1060 */
1061xfs_lsn_t
1062xlog_assign_tail_lsn_locked(
1063 struct xfs_mount *mp)
1064{
1065 struct xlog *log = mp->m_log;
1066 struct xfs_log_item *lip;
1067 xfs_lsn_t tail_lsn;
1068
1069 assert_spin_locked(&mp->m_ail->xa_lock);
1070
1071 /*
1072 * To make sure we always have a valid LSN for the log tail we keep
1073 * track of the last LSN which was committed in log->l_last_sync_lsn,
1074 * and use that when the AIL was empty.
1075 */
1076 lip = xfs_ail_min(mp->m_ail);
1077 if (lip)
1078 tail_lsn = lip->li_lsn;
1079 else
1080 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1081 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1082 atomic64_set(&log->l_tail_lsn, tail_lsn);
1083 return tail_lsn;
1084}
1085
1086xfs_lsn_t
1087xlog_assign_tail_lsn(
1088 struct xfs_mount *mp)
1089{
1090 xfs_lsn_t tail_lsn;
1091
1092 spin_lock(&mp->m_ail->xa_lock);
1093 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1094 spin_unlock(&mp->m_ail->xa_lock);
1095
1096 return tail_lsn;
1097}
1098
1099/*
1100 * Return the space in the log between the tail and the head. The head
1101 * is passed in the cycle/bytes formal parms. In the special case where
1102 * the reserve head has wrapped passed the tail, this calculation is no
1103 * longer valid. In this case, just return 0 which means there is no space
1104 * in the log. This works for all places where this function is called
1105 * with the reserve head. Of course, if the write head were to ever
1106 * wrap the tail, we should blow up. Rather than catch this case here,
1107 * we depend on other ASSERTions in other parts of the code. XXXmiken
1108 *
1109 * This code also handles the case where the reservation head is behind
1110 * the tail. The details of this case are described below, but the end
1111 * result is that we return the size of the log as the amount of space left.
1112 */
1113STATIC int
1114xlog_space_left(
1115 struct xlog *log,
1116 atomic64_t *head)
1117{
1118 int free_bytes;
1119 int tail_bytes;
1120 int tail_cycle;
1121 int head_cycle;
1122 int head_bytes;
1123
1124 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1125 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1126 tail_bytes = BBTOB(tail_bytes);
1127 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1128 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1129 else if (tail_cycle + 1 < head_cycle)
1130 return 0;
1131 else if (tail_cycle < head_cycle) {
1132 ASSERT(tail_cycle == (head_cycle - 1));
1133 free_bytes = tail_bytes - head_bytes;
1134 } else {
1135 /*
1136 * The reservation head is behind the tail.
1137 * In this case we just want to return the size of the
1138 * log as the amount of space left.
1139 */
1140 xfs_alert(log->l_mp,
1141 "xlog_space_left: head behind tail\n"
1142 " tail_cycle = %d, tail_bytes = %d\n"
1143 " GH cycle = %d, GH bytes = %d",
1144 tail_cycle, tail_bytes, head_cycle, head_bytes);
1145 ASSERT(0);
1146 free_bytes = log->l_logsize;
1147 }
1148 return free_bytes;
1149}
1150
1151
1152/*
1153 * Log function which is called when an io completes.
1154 *
1155 * The log manager needs its own routine, in order to control what
1156 * happens with the buffer after the write completes.
1157 */
1158void
1159xlog_iodone(xfs_buf_t *bp)
1160{
1161 struct xlog_in_core *iclog = bp->b_fspriv;
1162 struct xlog *l = iclog->ic_log;
1163 int aborted = 0;
1164
1165 /*
1166 * Race to shutdown the filesystem if we see an error.
1167 */
1168 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1169 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1170 xfs_buf_ioerror_alert(bp, __func__);
1171 xfs_buf_stale(bp);
1172 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1173 /*
1174 * This flag will be propagated to the trans-committed
1175 * callback routines to let them know that the log-commit
1176 * didn't succeed.
1177 */
1178 aborted = XFS_LI_ABORTED;
1179 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1180 aborted = XFS_LI_ABORTED;
1181 }
1182
1183 /* log I/O is always issued ASYNC */
1184 ASSERT(XFS_BUF_ISASYNC(bp));
1185 xlog_state_done_syncing(iclog, aborted);
1186
1187 /*
1188 * drop the buffer lock now that we are done. Nothing references
1189 * the buffer after this, so an unmount waiting on this lock can now
1190 * tear it down safely. As such, it is unsafe to reference the buffer
1191 * (bp) after the unlock as we could race with it being freed.
1192 */
1193 xfs_buf_unlock(bp);
1194}
1195
1196/*
1197 * Return size of each in-core log record buffer.
1198 *
1199 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1200 *
1201 * If the filesystem blocksize is too large, we may need to choose a
1202 * larger size since the directory code currently logs entire blocks.
1203 */
1204
1205STATIC void
1206xlog_get_iclog_buffer_size(
1207 struct xfs_mount *mp,
1208 struct xlog *log)
1209{
1210 int size;
1211 int xhdrs;
1212
1213 if (mp->m_logbufs <= 0)
1214 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1215 else
1216 log->l_iclog_bufs = mp->m_logbufs;
1217
1218 /*
1219 * Buffer size passed in from mount system call.
1220 */
1221 if (mp->m_logbsize > 0) {
1222 size = log->l_iclog_size = mp->m_logbsize;
1223 log->l_iclog_size_log = 0;
1224 while (size != 1) {
1225 log->l_iclog_size_log++;
1226 size >>= 1;
1227 }
1228
1229 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1230 /* # headers = size / 32k
1231 * one header holds cycles from 32k of data
1232 */
1233
1234 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1235 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1236 xhdrs++;
1237 log->l_iclog_hsize = xhdrs << BBSHIFT;
1238 log->l_iclog_heads = xhdrs;
1239 } else {
1240 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1241 log->l_iclog_hsize = BBSIZE;
1242 log->l_iclog_heads = 1;
1243 }
1244 goto done;
1245 }
1246
1247 /* All machines use 32kB buffers by default. */
1248 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1249 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1250
1251 /* the default log size is 16k or 32k which is one header sector */
1252 log->l_iclog_hsize = BBSIZE;
1253 log->l_iclog_heads = 1;
1254
1255done:
1256 /* are we being asked to make the sizes selected above visible? */
1257 if (mp->m_logbufs == 0)
1258 mp->m_logbufs = log->l_iclog_bufs;
1259 if (mp->m_logbsize == 0)
1260 mp->m_logbsize = log->l_iclog_size;
1261} /* xlog_get_iclog_buffer_size */
1262
1263
1264void
1265xfs_log_work_queue(
1266 struct xfs_mount *mp)
1267{
1268 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1269 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1270}
1271
1272/*
1273 * Every sync period we need to unpin all items in the AIL and push them to
1274 * disk. If there is nothing dirty, then we might need to cover the log to
1275 * indicate that the filesystem is idle.
1276 */
1277void
1278xfs_log_worker(
1279 struct work_struct *work)
1280{
1281 struct xlog *log = container_of(to_delayed_work(work),
1282 struct xlog, l_work);
1283 struct xfs_mount *mp = log->l_mp;
1284
1285 /* dgc: errors ignored - not fatal and nowhere to report them */
1286 if (xfs_log_need_covered(mp))
1287 xfs_fs_log_dummy(mp);
1288 else
1289 xfs_log_force(mp, 0);
1290
1291 /* start pushing all the metadata that is currently dirty */
1292 xfs_ail_push_all(mp->m_ail);
1293
1294 /* queue us up again */
1295 xfs_log_work_queue(mp);
1296}
1297
1298/*
1299 * This routine initializes some of the log structure for a given mount point.
1300 * Its primary purpose is to fill in enough, so recovery can occur. However,
1301 * some other stuff may be filled in too.
1302 */
1303STATIC struct xlog *
1304xlog_alloc_log(
1305 struct xfs_mount *mp,
1306 struct xfs_buftarg *log_target,
1307 xfs_daddr_t blk_offset,
1308 int num_bblks)
1309{
1310 struct xlog *log;
1311 xlog_rec_header_t *head;
1312 xlog_in_core_t **iclogp;
1313 xlog_in_core_t *iclog, *prev_iclog=NULL;
1314 xfs_buf_t *bp;
1315 int i;
1316 int error = ENOMEM;
1317 uint log2_size = 0;
1318
1319 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1320 if (!log) {
1321 xfs_warn(mp, "Log allocation failed: No memory!");
1322 goto out;
1323 }
1324
1325 log->l_mp = mp;
1326 log->l_targ = log_target;
1327 log->l_logsize = BBTOB(num_bblks);
1328 log->l_logBBstart = blk_offset;
1329 log->l_logBBsize = num_bblks;
1330 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1331 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1332 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1333
1334 log->l_prev_block = -1;
1335 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1336 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1337 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1338 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1339
1340 xlog_grant_head_init(&log->l_reserve_head);
1341 xlog_grant_head_init(&log->l_write_head);
1342
1343 error = EFSCORRUPTED;
1344 if (xfs_sb_version_hassector(&mp->m_sb)) {
1345 log2_size = mp->m_sb.sb_logsectlog;
1346 if (log2_size < BBSHIFT) {
1347 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1348 log2_size, BBSHIFT);
1349 goto out_free_log;
1350 }
1351
1352 log2_size -= BBSHIFT;
1353 if (log2_size > mp->m_sectbb_log) {
1354 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1355 log2_size, mp->m_sectbb_log);
1356 goto out_free_log;
1357 }
1358
1359 /* for larger sector sizes, must have v2 or external log */
1360 if (log2_size && log->l_logBBstart > 0 &&
1361 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1362 xfs_warn(mp,
1363 "log sector size (0x%x) invalid for configuration.",
1364 log2_size);
1365 goto out_free_log;
1366 }
1367 }
1368 log->l_sectBBsize = 1 << log2_size;
1369
1370 xlog_get_iclog_buffer_size(mp, log);
1371
1372 error = ENOMEM;
1373 bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1374 if (!bp)
1375 goto out_free_log;
1376
1377 /*
1378 * The iclogbuf buffer locks are held over IO but we are not going to do
1379 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1380 * when appropriately.
1381 */
1382 ASSERT(xfs_buf_islocked(bp));
1383 xfs_buf_unlock(bp);
1384
1385 bp->b_iodone = xlog_iodone;
1386 log->l_xbuf = bp;
1387
1388 spin_lock_init(&log->l_icloglock);
1389 init_waitqueue_head(&log->l_flush_wait);
1390
1391 iclogp = &log->l_iclog;
1392 /*
1393 * The amount of memory to allocate for the iclog structure is
1394 * rather funky due to the way the structure is defined. It is
1395 * done this way so that we can use different sizes for machines
1396 * with different amounts of memory. See the definition of
1397 * xlog_in_core_t in xfs_log_priv.h for details.
1398 */
1399 ASSERT(log->l_iclog_size >= 4096);
1400 for (i=0; i < log->l_iclog_bufs; i++) {
1401 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1402 if (!*iclogp)
1403 goto out_free_iclog;
1404
1405 iclog = *iclogp;
1406 iclog->ic_prev = prev_iclog;
1407 prev_iclog = iclog;
1408
1409 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1410 BTOBB(log->l_iclog_size), 0);
1411 if (!bp)
1412 goto out_free_iclog;
1413
1414 ASSERT(xfs_buf_islocked(bp));
1415 xfs_buf_unlock(bp);
1416
1417 bp->b_iodone = xlog_iodone;
1418 iclog->ic_bp = bp;
1419 iclog->ic_data = bp->b_addr;
1420#ifdef DEBUG
1421 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1422#endif
1423 head = &iclog->ic_header;
1424 memset(head, 0, sizeof(xlog_rec_header_t));
1425 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1426 head->h_version = cpu_to_be32(
1427 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1428 head->h_size = cpu_to_be32(log->l_iclog_size);
1429 /* new fields */
1430 head->h_fmt = cpu_to_be32(XLOG_FMT);
1431 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1432
1433 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1434 iclog->ic_state = XLOG_STATE_ACTIVE;
1435 iclog->ic_log = log;
1436 atomic_set(&iclog->ic_refcnt, 0);
1437 spin_lock_init(&iclog->ic_callback_lock);
1438 iclog->ic_callback_tail = &(iclog->ic_callback);
1439 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1440
1441 init_waitqueue_head(&iclog->ic_force_wait);
1442 init_waitqueue_head(&iclog->ic_write_wait);
1443
1444 iclogp = &iclog->ic_next;
1445 }
1446 *iclogp = log->l_iclog; /* complete ring */
1447 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1448
1449 error = xlog_cil_init(log);
1450 if (error)
1451 goto out_free_iclog;
1452 return log;
1453
1454out_free_iclog:
1455 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1456 prev_iclog = iclog->ic_next;
1457 if (iclog->ic_bp)
1458 xfs_buf_free(iclog->ic_bp);
1459 kmem_free(iclog);
1460 }
1461 spinlock_destroy(&log->l_icloglock);
1462 xfs_buf_free(log->l_xbuf);
1463out_free_log:
1464 kmem_free(log);
1465out:
1466 return ERR_PTR(-error);
1467} /* xlog_alloc_log */
1468
1469
1470/*
1471 * Write out the commit record of a transaction associated with the given
1472 * ticket. Return the lsn of the commit record.
1473 */
1474STATIC int
1475xlog_commit_record(
1476 struct xlog *log,
1477 struct xlog_ticket *ticket,
1478 struct xlog_in_core **iclog,
1479 xfs_lsn_t *commitlsnp)
1480{
1481 struct xfs_mount *mp = log->l_mp;
1482 int error;
1483 struct xfs_log_iovec reg = {
1484 .i_addr = NULL,
1485 .i_len = 0,
1486 .i_type = XLOG_REG_TYPE_COMMIT,
1487 };
1488 struct xfs_log_vec vec = {
1489 .lv_niovecs = 1,
1490 .lv_iovecp = ®,
1491 };
1492
1493 ASSERT_ALWAYS(iclog);
1494 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1495 XLOG_COMMIT_TRANS);
1496 if (error)
1497 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1498 return error;
1499}
1500
1501/*
1502 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1503 * log space. This code pushes on the lsn which would supposedly free up
1504 * the 25% which we want to leave free. We may need to adopt a policy which
1505 * pushes on an lsn which is further along in the log once we reach the high
1506 * water mark. In this manner, we would be creating a low water mark.
1507 */
1508STATIC void
1509xlog_grant_push_ail(
1510 struct xlog *log,
1511 int need_bytes)
1512{
1513 xfs_lsn_t threshold_lsn = 0;
1514 xfs_lsn_t last_sync_lsn;
1515 int free_blocks;
1516 int free_bytes;
1517 int threshold_block;
1518 int threshold_cycle;
1519 int free_threshold;
1520
1521 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1522
1523 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1524 free_blocks = BTOBBT(free_bytes);
1525
1526 /*
1527 * Set the threshold for the minimum number of free blocks in the
1528 * log to the maximum of what the caller needs, one quarter of the
1529 * log, and 256 blocks.
1530 */
1531 free_threshold = BTOBB(need_bytes);
1532 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1533 free_threshold = MAX(free_threshold, 256);
1534 if (free_blocks >= free_threshold)
1535 return;
1536
1537 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1538 &threshold_block);
1539 threshold_block += free_threshold;
1540 if (threshold_block >= log->l_logBBsize) {
1541 threshold_block -= log->l_logBBsize;
1542 threshold_cycle += 1;
1543 }
1544 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1545 threshold_block);
1546 /*
1547 * Don't pass in an lsn greater than the lsn of the last
1548 * log record known to be on disk. Use a snapshot of the last sync lsn
1549 * so that it doesn't change between the compare and the set.
1550 */
1551 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1552 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1553 threshold_lsn = last_sync_lsn;
1554
1555 /*
1556 * Get the transaction layer to kick the dirty buffers out to
1557 * disk asynchronously. No point in trying to do this if
1558 * the filesystem is shutting down.
1559 */
1560 if (!XLOG_FORCED_SHUTDOWN(log))
1561 xfs_ail_push(log->l_ailp, threshold_lsn);
1562}
1563
1564/*
1565 * Stamp cycle number in every block
1566 */
1567STATIC void
1568xlog_pack_data(
1569 struct xlog *log,
1570 struct xlog_in_core *iclog,
1571 int roundoff)
1572{
1573 int i, j, k;
1574 int size = iclog->ic_offset + roundoff;
1575 __be32 cycle_lsn;
1576 xfs_caddr_t dp;
1577
1578 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1579
1580 dp = iclog->ic_datap;
1581 for (i = 0; i < BTOBB(size); i++) {
1582 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1583 break;
1584 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1585 *(__be32 *)dp = cycle_lsn;
1586 dp += BBSIZE;
1587 }
1588
1589 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1590 xlog_in_core_2_t *xhdr = iclog->ic_data;
1591
1592 for ( ; i < BTOBB(size); i++) {
1593 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1594 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1595 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1596 *(__be32 *)dp = cycle_lsn;
1597 dp += BBSIZE;
1598 }
1599
1600 for (i = 1; i < log->l_iclog_heads; i++)
1601 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1602 }
1603}
1604
1605/*
1606 * Calculate the checksum for a log buffer.
1607 *
1608 * This is a little more complicated than it should be because the various
1609 * headers and the actual data are non-contiguous.
1610 */
1611__le32
1612xlog_cksum(
1613 struct xlog *log,
1614 struct xlog_rec_header *rhead,
1615 char *dp,
1616 int size)
1617{
1618 __uint32_t crc;
1619
1620 /* first generate the crc for the record header ... */
1621 crc = xfs_start_cksum((char *)rhead,
1622 sizeof(struct xlog_rec_header),
1623 offsetof(struct xlog_rec_header, h_crc));
1624
1625 /* ... then for additional cycle data for v2 logs ... */
1626 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1627 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1628 int i;
1629
1630 for (i = 1; i < log->l_iclog_heads; i++) {
1631 crc = crc32c(crc, &xhdr[i].hic_xheader,
1632 sizeof(struct xlog_rec_ext_header));
1633 }
1634 }
1635
1636 /* ... and finally for the payload */
1637 crc = crc32c(crc, dp, size);
1638
1639 return xfs_end_cksum(crc);
1640}
1641
1642/*
1643 * The bdstrat callback function for log bufs. This gives us a central
1644 * place to trap bufs in case we get hit by a log I/O error and need to
1645 * shutdown. Actually, in practice, even when we didn't get a log error,
1646 * we transition the iclogs to IOERROR state *after* flushing all existing
1647 * iclogs to disk. This is because we don't want anymore new transactions to be
1648 * started or completed afterwards.
1649 *
1650 * We lock the iclogbufs here so that we can serialise against IO completion
1651 * during unmount. We might be processing a shutdown triggered during unmount,
1652 * and that can occur asynchronously to the unmount thread, and hence we need to
1653 * ensure that completes before tearing down the iclogbufs. Hence we need to
1654 * hold the buffer lock across the log IO to acheive that.
1655 */
1656STATIC int
1657xlog_bdstrat(
1658 struct xfs_buf *bp)
1659{
1660 struct xlog_in_core *iclog = bp->b_fspriv;
1661
1662 xfs_buf_lock(bp);
1663 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1664 xfs_buf_ioerror(bp, EIO);
1665 xfs_buf_stale(bp);
1666 xfs_buf_ioend(bp, 0);
1667 /*
1668 * It would seem logical to return EIO here, but we rely on
1669 * the log state machine to propagate I/O errors instead of
1670 * doing it here. Similarly, IO completion will unlock the
1671 * buffer, so we don't do it here.
1672 */
1673 return 0;
1674 }
1675
1676 xfs_buf_iorequest(bp);
1677 return 0;
1678}
1679
1680/*
1681 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1682 * fashion. Previously, we should have moved the current iclog
1683 * ptr in the log to point to the next available iclog. This allows further
1684 * write to continue while this code syncs out an iclog ready to go.
1685 * Before an in-core log can be written out, the data section must be scanned
1686 * to save away the 1st word of each BBSIZE block into the header. We replace
1687 * it with the current cycle count. Each BBSIZE block is tagged with the
1688 * cycle count because there in an implicit assumption that drives will
1689 * guarantee that entire 512 byte blocks get written at once. In other words,
1690 * we can't have part of a 512 byte block written and part not written. By
1691 * tagging each block, we will know which blocks are valid when recovering
1692 * after an unclean shutdown.
1693 *
1694 * This routine is single threaded on the iclog. No other thread can be in
1695 * this routine with the same iclog. Changing contents of iclog can there-
1696 * fore be done without grabbing the state machine lock. Updating the global
1697 * log will require grabbing the lock though.
1698 *
1699 * The entire log manager uses a logical block numbering scheme. Only
1700 * log_sync (and then only bwrite()) know about the fact that the log may
1701 * not start with block zero on a given device. The log block start offset
1702 * is added immediately before calling bwrite().
1703 */
1704
1705STATIC int
1706xlog_sync(
1707 struct xlog *log,
1708 struct xlog_in_core *iclog)
1709{
1710 xfs_buf_t *bp;
1711 int i;
1712 uint count; /* byte count of bwrite */
1713 uint count_init; /* initial count before roundup */
1714 int roundoff; /* roundoff to BB or stripe */
1715 int split = 0; /* split write into two regions */
1716 int error;
1717 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1718 int size;
1719
1720 XFS_STATS_INC(xs_log_writes);
1721 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1722
1723 /* Add for LR header */
1724 count_init = log->l_iclog_hsize + iclog->ic_offset;
1725
1726 /* Round out the log write size */
1727 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1728 /* we have a v2 stripe unit to use */
1729 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1730 } else {
1731 count = BBTOB(BTOBB(count_init));
1732 }
1733 roundoff = count - count_init;
1734 ASSERT(roundoff >= 0);
1735 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1736 roundoff < log->l_mp->m_sb.sb_logsunit)
1737 ||
1738 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1739 roundoff < BBTOB(1)));
1740
1741 /* move grant heads by roundoff in sync */
1742 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1743 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1744
1745 /* put cycle number in every block */
1746 xlog_pack_data(log, iclog, roundoff);
1747
1748 /* real byte length */
1749 size = iclog->ic_offset;
1750 if (v2)
1751 size += roundoff;
1752 iclog->ic_header.h_len = cpu_to_be32(size);
1753
1754 bp = iclog->ic_bp;
1755 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1756
1757 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1758
1759 /* Do we need to split this write into 2 parts? */
1760 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1761 char *dptr;
1762
1763 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1764 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1765 iclog->ic_bwritecnt = 2;
1766
1767 /*
1768 * Bump the cycle numbers at the start of each block in the
1769 * part of the iclog that ends up in the buffer that gets
1770 * written to the start of the log.
1771 *
1772 * Watch out for the header magic number case, though.
1773 */
1774 dptr = (char *)&iclog->ic_header + count;
1775 for (i = 0; i < split; i += BBSIZE) {
1776 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1777 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1778 cycle++;
1779 *(__be32 *)dptr = cpu_to_be32(cycle);
1780
1781 dptr += BBSIZE;
1782 }
1783 } else {
1784 iclog->ic_bwritecnt = 1;
1785 }
1786
1787 /* calculcate the checksum */
1788 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1789 iclog->ic_datap, size);
1790
1791 bp->b_io_length = BTOBB(count);
1792 bp->b_fspriv = iclog;
1793 XFS_BUF_ZEROFLAGS(bp);
1794 XFS_BUF_ASYNC(bp);
1795 bp->b_flags |= XBF_SYNCIO;
1796
1797 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1798 bp->b_flags |= XBF_FUA;
1799
1800 /*
1801 * Flush the data device before flushing the log to make
1802 * sure all meta data written back from the AIL actually made
1803 * it to disk before stamping the new log tail LSN into the
1804 * log buffer. For an external log we need to issue the
1805 * flush explicitly, and unfortunately synchronously here;
1806 * for an internal log we can simply use the block layer
1807 * state machine for preflushes.
1808 */
1809 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1810 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1811 else
1812 bp->b_flags |= XBF_FLUSH;
1813 }
1814
1815 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1816 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1817
1818 xlog_verify_iclog(log, iclog, count, true);
1819
1820 /* account for log which doesn't start at block #0 */
1821 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1822 /*
1823 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1824 * is shutting down.
1825 */
1826 XFS_BUF_WRITE(bp);
1827
1828 error = xlog_bdstrat(bp);
1829 if (error) {
1830 xfs_buf_ioerror_alert(bp, "xlog_sync");
1831 return error;
1832 }
1833 if (split) {
1834 bp = iclog->ic_log->l_xbuf;
1835 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1836 xfs_buf_associate_memory(bp,
1837 (char *)&iclog->ic_header + count, split);
1838 bp->b_fspriv = iclog;
1839 XFS_BUF_ZEROFLAGS(bp);
1840 XFS_BUF_ASYNC(bp);
1841 bp->b_flags |= XBF_SYNCIO;
1842 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1843 bp->b_flags |= XBF_FUA;
1844
1845 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1846 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1847
1848 /* account for internal log which doesn't start at block #0 */
1849 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1850 XFS_BUF_WRITE(bp);
1851 error = xlog_bdstrat(bp);
1852 if (error) {
1853 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1854 return error;
1855 }
1856 }
1857 return 0;
1858} /* xlog_sync */
1859
1860/*
1861 * Deallocate a log structure
1862 */
1863STATIC void
1864xlog_dealloc_log(
1865 struct xlog *log)
1866{
1867 xlog_in_core_t *iclog, *next_iclog;
1868 int i;
1869
1870 xlog_cil_destroy(log);
1871
1872 /*
1873 * Cycle all the iclogbuf locks to make sure all log IO completion
1874 * is done before we tear down these buffers.
1875 */
1876 iclog = log->l_iclog;
1877 for (i = 0; i < log->l_iclog_bufs; i++) {
1878 xfs_buf_lock(iclog->ic_bp);
1879 xfs_buf_unlock(iclog->ic_bp);
1880 iclog = iclog->ic_next;
1881 }
1882
1883 /*
1884 * Always need to ensure that the extra buffer does not point to memory
1885 * owned by another log buffer before we free it. Also, cycle the lock
1886 * first to ensure we've completed IO on it.
1887 */
1888 xfs_buf_lock(log->l_xbuf);
1889 xfs_buf_unlock(log->l_xbuf);
1890 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1891 xfs_buf_free(log->l_xbuf);
1892
1893 iclog = log->l_iclog;
1894 for (i = 0; i < log->l_iclog_bufs; i++) {
1895 xfs_buf_free(iclog->ic_bp);
1896 next_iclog = iclog->ic_next;
1897 kmem_free(iclog);
1898 iclog = next_iclog;
1899 }
1900 spinlock_destroy(&log->l_icloglock);
1901
1902 log->l_mp->m_log = NULL;
1903 kmem_free(log);
1904} /* xlog_dealloc_log */
1905
1906/*
1907 * Update counters atomically now that memcpy is done.
1908 */
1909/* ARGSUSED */
1910static inline void
1911xlog_state_finish_copy(
1912 struct xlog *log,
1913 struct xlog_in_core *iclog,
1914 int record_cnt,
1915 int copy_bytes)
1916{
1917 spin_lock(&log->l_icloglock);
1918
1919 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1920 iclog->ic_offset += copy_bytes;
1921
1922 spin_unlock(&log->l_icloglock);
1923} /* xlog_state_finish_copy */
1924
1925
1926
1927
1928/*
1929 * print out info relating to regions written which consume
1930 * the reservation
1931 */
1932void
1933xlog_print_tic_res(
1934 struct xfs_mount *mp,
1935 struct xlog_ticket *ticket)
1936{
1937 uint i;
1938 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1939
1940 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1941 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1942 "bformat",
1943 "bchunk",
1944 "efi_format",
1945 "efd_format",
1946 "iformat",
1947 "icore",
1948 "iext",
1949 "ibroot",
1950 "ilocal",
1951 "iattr_ext",
1952 "iattr_broot",
1953 "iattr_local",
1954 "qformat",
1955 "dquot",
1956 "quotaoff",
1957 "LR header",
1958 "unmount",
1959 "commit",
1960 "trans header"
1961 };
1962 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1963 "SETATTR_NOT_SIZE",
1964 "SETATTR_SIZE",
1965 "INACTIVE",
1966 "CREATE",
1967 "CREATE_TRUNC",
1968 "TRUNCATE_FILE",
1969 "REMOVE",
1970 "LINK",
1971 "RENAME",
1972 "MKDIR",
1973 "RMDIR",
1974 "SYMLINK",
1975 "SET_DMATTRS",
1976 "GROWFS",
1977 "STRAT_WRITE",
1978 "DIOSTRAT",
1979 "WRITE_SYNC",
1980 "WRITEID",
1981 "ADDAFORK",
1982 "ATTRINVAL",
1983 "ATRUNCATE",
1984 "ATTR_SET",
1985 "ATTR_RM",
1986 "ATTR_FLAG",
1987 "CLEAR_AGI_BUCKET",
1988 "QM_SBCHANGE",
1989 "DUMMY1",
1990 "DUMMY2",
1991 "QM_QUOTAOFF",
1992 "QM_DQALLOC",
1993 "QM_SETQLIM",
1994 "QM_DQCLUSTER",
1995 "QM_QINOCREATE",
1996 "QM_QUOTAOFF_END",
1997 "SB_UNIT",
1998 "FSYNC_TS",
1999 "GROWFSRT_ALLOC",
2000 "GROWFSRT_ZERO",
2001 "GROWFSRT_FREE",
2002 "SWAPEXT"
2003 };
2004
2005 xfs_warn(mp,
2006 "xlog_write: reservation summary:\n"
2007 " trans type = %s (%u)\n"
2008 " unit res = %d bytes\n"
2009 " current res = %d bytes\n"
2010 " total reg = %u bytes (o/flow = %u bytes)\n"
2011 " ophdrs = %u (ophdr space = %u bytes)\n"
2012 " ophdr + reg = %u bytes\n"
2013 " num regions = %u\n",
2014 ((ticket->t_trans_type <= 0 ||
2015 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2016 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2017 ticket->t_trans_type,
2018 ticket->t_unit_res,
2019 ticket->t_curr_res,
2020 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2021 ticket->t_res_num_ophdrs, ophdr_spc,
2022 ticket->t_res_arr_sum +
2023 ticket->t_res_o_flow + ophdr_spc,
2024 ticket->t_res_num);
2025
2026 for (i = 0; i < ticket->t_res_num; i++) {
2027 uint r_type = ticket->t_res_arr[i].r_type;
2028 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2029 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2030 "bad-rtype" : res_type_str[r_type-1]),
2031 ticket->t_res_arr[i].r_len);
2032 }
2033
2034 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2035 "xlog_write: reservation ran out. Need to up reservation");
2036 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2037}
2038
2039/*
2040 * Calculate the potential space needed by the log vector. Each region gets
2041 * its own xlog_op_header_t and may need to be double word aligned.
2042 */
2043static int
2044xlog_write_calc_vec_length(
2045 struct xlog_ticket *ticket,
2046 struct xfs_log_vec *log_vector)
2047{
2048 struct xfs_log_vec *lv;
2049 int headers = 0;
2050 int len = 0;
2051 int i;
2052
2053 /* acct for start rec of xact */
2054 if (ticket->t_flags & XLOG_TIC_INITED)
2055 headers++;
2056
2057 for (lv = log_vector; lv; lv = lv->lv_next) {
2058 /* we don't write ordered log vectors */
2059 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2060 continue;
2061
2062 headers += lv->lv_niovecs;
2063
2064 for (i = 0; i < lv->lv_niovecs; i++) {
2065 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2066
2067 len += vecp->i_len;
2068 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2069 }
2070 }
2071
2072 ticket->t_res_num_ophdrs += headers;
2073 len += headers * sizeof(struct xlog_op_header);
2074
2075 return len;
2076}
2077
2078/*
2079 * If first write for transaction, insert start record We can't be trying to
2080 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2081 */
2082static int
2083xlog_write_start_rec(
2084 struct xlog_op_header *ophdr,
2085 struct xlog_ticket *ticket)
2086{
2087 if (!(ticket->t_flags & XLOG_TIC_INITED))
2088 return 0;
2089
2090 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2091 ophdr->oh_clientid = ticket->t_clientid;
2092 ophdr->oh_len = 0;
2093 ophdr->oh_flags = XLOG_START_TRANS;
2094 ophdr->oh_res2 = 0;
2095
2096 ticket->t_flags &= ~XLOG_TIC_INITED;
2097
2098 return sizeof(struct xlog_op_header);
2099}
2100
2101static xlog_op_header_t *
2102xlog_write_setup_ophdr(
2103 struct xlog *log,
2104 struct xlog_op_header *ophdr,
2105 struct xlog_ticket *ticket,
2106 uint flags)
2107{
2108 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2109 ophdr->oh_clientid = ticket->t_clientid;
2110 ophdr->oh_res2 = 0;
2111
2112 /* are we copying a commit or unmount record? */
2113 ophdr->oh_flags = flags;
2114
2115 /*
2116 * We've seen logs corrupted with bad transaction client ids. This
2117 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2118 * and shut down the filesystem.
2119 */
2120 switch (ophdr->oh_clientid) {
2121 case XFS_TRANSACTION:
2122 case XFS_VOLUME:
2123 case XFS_LOG:
2124 break;
2125 default:
2126 xfs_warn(log->l_mp,
2127 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2128 ophdr->oh_clientid, ticket);
2129 return NULL;
2130 }
2131
2132 return ophdr;
2133}
2134
2135/*
2136 * Set up the parameters of the region copy into the log. This has
2137 * to handle region write split across multiple log buffers - this
2138 * state is kept external to this function so that this code can
2139 * be written in an obvious, self documenting manner.
2140 */
2141static int
2142xlog_write_setup_copy(
2143 struct xlog_ticket *ticket,
2144 struct xlog_op_header *ophdr,
2145 int space_available,
2146 int space_required,
2147 int *copy_off,
2148 int *copy_len,
2149 int *last_was_partial_copy,
2150 int *bytes_consumed)
2151{
2152 int still_to_copy;
2153
2154 still_to_copy = space_required - *bytes_consumed;
2155 *copy_off = *bytes_consumed;
2156
2157 if (still_to_copy <= space_available) {
2158 /* write of region completes here */
2159 *copy_len = still_to_copy;
2160 ophdr->oh_len = cpu_to_be32(*copy_len);
2161 if (*last_was_partial_copy)
2162 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2163 *last_was_partial_copy = 0;
2164 *bytes_consumed = 0;
2165 return 0;
2166 }
2167
2168 /* partial write of region, needs extra log op header reservation */
2169 *copy_len = space_available;
2170 ophdr->oh_len = cpu_to_be32(*copy_len);
2171 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2172 if (*last_was_partial_copy)
2173 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2174 *bytes_consumed += *copy_len;
2175 (*last_was_partial_copy)++;
2176
2177 /* account for new log op header */
2178 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2179 ticket->t_res_num_ophdrs++;
2180
2181 return sizeof(struct xlog_op_header);
2182}
2183
2184static int
2185xlog_write_copy_finish(
2186 struct xlog *log,
2187 struct xlog_in_core *iclog,
2188 uint flags,
2189 int *record_cnt,
2190 int *data_cnt,
2191 int *partial_copy,
2192 int *partial_copy_len,
2193 int log_offset,
2194 struct xlog_in_core **commit_iclog)
2195{
2196 if (*partial_copy) {
2197 /*
2198 * This iclog has already been marked WANT_SYNC by
2199 * xlog_state_get_iclog_space.
2200 */
2201 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2202 *record_cnt = 0;
2203 *data_cnt = 0;
2204 return xlog_state_release_iclog(log, iclog);
2205 }
2206
2207 *partial_copy = 0;
2208 *partial_copy_len = 0;
2209
2210 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2211 /* no more space in this iclog - push it. */
2212 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2213 *record_cnt = 0;
2214 *data_cnt = 0;
2215
2216 spin_lock(&log->l_icloglock);
2217 xlog_state_want_sync(log, iclog);
2218 spin_unlock(&log->l_icloglock);
2219
2220 if (!commit_iclog)
2221 return xlog_state_release_iclog(log, iclog);
2222 ASSERT(flags & XLOG_COMMIT_TRANS);
2223 *commit_iclog = iclog;
2224 }
2225
2226 return 0;
2227}
2228
2229/*
2230 * Write some region out to in-core log
2231 *
2232 * This will be called when writing externally provided regions or when
2233 * writing out a commit record for a given transaction.
2234 *
2235 * General algorithm:
2236 * 1. Find total length of this write. This may include adding to the
2237 * lengths passed in.
2238 * 2. Check whether we violate the tickets reservation.
2239 * 3. While writing to this iclog
2240 * A. Reserve as much space in this iclog as can get
2241 * B. If this is first write, save away start lsn
2242 * C. While writing this region:
2243 * 1. If first write of transaction, write start record
2244 * 2. Write log operation header (header per region)
2245 * 3. Find out if we can fit entire region into this iclog
2246 * 4. Potentially, verify destination memcpy ptr
2247 * 5. Memcpy (partial) region
2248 * 6. If partial copy, release iclog; otherwise, continue
2249 * copying more regions into current iclog
2250 * 4. Mark want sync bit (in simulation mode)
2251 * 5. Release iclog for potential flush to on-disk log.
2252 *
2253 * ERRORS:
2254 * 1. Panic if reservation is overrun. This should never happen since
2255 * reservation amounts are generated internal to the filesystem.
2256 * NOTES:
2257 * 1. Tickets are single threaded data structures.
2258 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2259 * syncing routine. When a single log_write region needs to span
2260 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2261 * on all log operation writes which don't contain the end of the
2262 * region. The XLOG_END_TRANS bit is used for the in-core log
2263 * operation which contains the end of the continued log_write region.
2264 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2265 * we don't really know exactly how much space will be used. As a result,
2266 * we don't update ic_offset until the end when we know exactly how many
2267 * bytes have been written out.
2268 */
2269int
2270xlog_write(
2271 struct xlog *log,
2272 struct xfs_log_vec *log_vector,
2273 struct xlog_ticket *ticket,
2274 xfs_lsn_t *start_lsn,
2275 struct xlog_in_core **commit_iclog,
2276 uint flags)
2277{
2278 struct xlog_in_core *iclog = NULL;
2279 struct xfs_log_iovec *vecp;
2280 struct xfs_log_vec *lv;
2281 int len;
2282 int index;
2283 int partial_copy = 0;
2284 int partial_copy_len = 0;
2285 int contwr = 0;
2286 int record_cnt = 0;
2287 int data_cnt = 0;
2288 int error;
2289
2290 *start_lsn = 0;
2291
2292 len = xlog_write_calc_vec_length(ticket, log_vector);
2293
2294 /*
2295 * Region headers and bytes are already accounted for.
2296 * We only need to take into account start records and
2297 * split regions in this function.
2298 */
2299 if (ticket->t_flags & XLOG_TIC_INITED)
2300 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2301
2302 /*
2303 * Commit record headers need to be accounted for. These
2304 * come in as separate writes so are easy to detect.
2305 */
2306 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2307 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2308
2309 if (ticket->t_curr_res < 0)
2310 xlog_print_tic_res(log->l_mp, ticket);
2311
2312 index = 0;
2313 lv = log_vector;
2314 vecp = lv->lv_iovecp;
2315 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2316 void *ptr;
2317 int log_offset;
2318
2319 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2320 &contwr, &log_offset);
2321 if (error)
2322 return error;
2323
2324 ASSERT(log_offset <= iclog->ic_size - 1);
2325 ptr = iclog->ic_datap + log_offset;
2326
2327 /* start_lsn is the first lsn written to. That's all we need. */
2328 if (!*start_lsn)
2329 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2330
2331 /*
2332 * This loop writes out as many regions as can fit in the amount
2333 * of space which was allocated by xlog_state_get_iclog_space().
2334 */
2335 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2336 struct xfs_log_iovec *reg;
2337 struct xlog_op_header *ophdr;
2338 int start_rec_copy;
2339 int copy_len;
2340 int copy_off;
2341 bool ordered = false;
2342
2343 /* ordered log vectors have no regions to write */
2344 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2345 ASSERT(lv->lv_niovecs == 0);
2346 ordered = true;
2347 goto next_lv;
2348 }
2349
2350 reg = &vecp[index];
2351 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2352 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2353
2354 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2355 if (start_rec_copy) {
2356 record_cnt++;
2357 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2358 start_rec_copy);
2359 }
2360
2361 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2362 if (!ophdr)
2363 return XFS_ERROR(EIO);
2364
2365 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2366 sizeof(struct xlog_op_header));
2367
2368 len += xlog_write_setup_copy(ticket, ophdr,
2369 iclog->ic_size-log_offset,
2370 reg->i_len,
2371 ©_off, ©_len,
2372 &partial_copy,
2373 &partial_copy_len);
2374 xlog_verify_dest_ptr(log, ptr);
2375
2376 /* copy region */
2377 ASSERT(copy_len >= 0);
2378 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2379 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2380
2381 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2382 record_cnt++;
2383 data_cnt += contwr ? copy_len : 0;
2384
2385 error = xlog_write_copy_finish(log, iclog, flags,
2386 &record_cnt, &data_cnt,
2387 &partial_copy,
2388 &partial_copy_len,
2389 log_offset,
2390 commit_iclog);
2391 if (error)
2392 return error;
2393
2394 /*
2395 * if we had a partial copy, we need to get more iclog
2396 * space but we don't want to increment the region
2397 * index because there is still more is this region to
2398 * write.
2399 *
2400 * If we completed writing this region, and we flushed
2401 * the iclog (indicated by resetting of the record
2402 * count), then we also need to get more log space. If
2403 * this was the last record, though, we are done and
2404 * can just return.
2405 */
2406 if (partial_copy)
2407 break;
2408
2409 if (++index == lv->lv_niovecs) {
2410next_lv:
2411 lv = lv->lv_next;
2412 index = 0;
2413 if (lv)
2414 vecp = lv->lv_iovecp;
2415 }
2416 if (record_cnt == 0 && ordered == false) {
2417 if (!lv)
2418 return 0;
2419 break;
2420 }
2421 }
2422 }
2423
2424 ASSERT(len == 0);
2425
2426 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2427 if (!commit_iclog)
2428 return xlog_state_release_iclog(log, iclog);
2429
2430 ASSERT(flags & XLOG_COMMIT_TRANS);
2431 *commit_iclog = iclog;
2432 return 0;
2433}
2434
2435
2436/*****************************************************************************
2437 *
2438 * State Machine functions
2439 *
2440 *****************************************************************************
2441 */
2442
2443/* Clean iclogs starting from the head. This ordering must be
2444 * maintained, so an iclog doesn't become ACTIVE beyond one that
2445 * is SYNCING. This is also required to maintain the notion that we use
2446 * a ordered wait queue to hold off would be writers to the log when every
2447 * iclog is trying to sync to disk.
2448 *
2449 * State Change: DIRTY -> ACTIVE
2450 */
2451STATIC void
2452xlog_state_clean_log(
2453 struct xlog *log)
2454{
2455 xlog_in_core_t *iclog;
2456 int changed = 0;
2457
2458 iclog = log->l_iclog;
2459 do {
2460 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2461 iclog->ic_state = XLOG_STATE_ACTIVE;
2462 iclog->ic_offset = 0;
2463 ASSERT(iclog->ic_callback == NULL);
2464 /*
2465 * If the number of ops in this iclog indicate it just
2466 * contains the dummy transaction, we can
2467 * change state into IDLE (the second time around).
2468 * Otherwise we should change the state into
2469 * NEED a dummy.
2470 * We don't need to cover the dummy.
2471 */
2472 if (!changed &&
2473 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2474 XLOG_COVER_OPS)) {
2475 changed = 1;
2476 } else {
2477 /*
2478 * We have two dirty iclogs so start over
2479 * This could also be num of ops indicates
2480 * this is not the dummy going out.
2481 */
2482 changed = 2;
2483 }
2484 iclog->ic_header.h_num_logops = 0;
2485 memset(iclog->ic_header.h_cycle_data, 0,
2486 sizeof(iclog->ic_header.h_cycle_data));
2487 iclog->ic_header.h_lsn = 0;
2488 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2489 /* do nothing */;
2490 else
2491 break; /* stop cleaning */
2492 iclog = iclog->ic_next;
2493 } while (iclog != log->l_iclog);
2494
2495 /* log is locked when we are called */
2496 /*
2497 * Change state for the dummy log recording.
2498 * We usually go to NEED. But we go to NEED2 if the changed indicates
2499 * we are done writing the dummy record.
2500 * If we are done with the second dummy recored (DONE2), then
2501 * we go to IDLE.
2502 */
2503 if (changed) {
2504 switch (log->l_covered_state) {
2505 case XLOG_STATE_COVER_IDLE:
2506 case XLOG_STATE_COVER_NEED:
2507 case XLOG_STATE_COVER_NEED2:
2508 log->l_covered_state = XLOG_STATE_COVER_NEED;
2509 break;
2510
2511 case XLOG_STATE_COVER_DONE:
2512 if (changed == 1)
2513 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2514 else
2515 log->l_covered_state = XLOG_STATE_COVER_NEED;
2516 break;
2517
2518 case XLOG_STATE_COVER_DONE2:
2519 if (changed == 1)
2520 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2521 else
2522 log->l_covered_state = XLOG_STATE_COVER_NEED;
2523 break;
2524
2525 default:
2526 ASSERT(0);
2527 }
2528 }
2529} /* xlog_state_clean_log */
2530
2531STATIC xfs_lsn_t
2532xlog_get_lowest_lsn(
2533 struct xlog *log)
2534{
2535 xlog_in_core_t *lsn_log;
2536 xfs_lsn_t lowest_lsn, lsn;
2537
2538 lsn_log = log->l_iclog;
2539 lowest_lsn = 0;
2540 do {
2541 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2542 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2543 if ((lsn && !lowest_lsn) ||
2544 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2545 lowest_lsn = lsn;
2546 }
2547 }
2548 lsn_log = lsn_log->ic_next;
2549 } while (lsn_log != log->l_iclog);
2550 return lowest_lsn;
2551}
2552
2553
2554STATIC void
2555xlog_state_do_callback(
2556 struct xlog *log,
2557 int aborted,
2558 struct xlog_in_core *ciclog)
2559{
2560 xlog_in_core_t *iclog;
2561 xlog_in_core_t *first_iclog; /* used to know when we've
2562 * processed all iclogs once */
2563 xfs_log_callback_t *cb, *cb_next;
2564 int flushcnt = 0;
2565 xfs_lsn_t lowest_lsn;
2566 int ioerrors; /* counter: iclogs with errors */
2567 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2568 int funcdidcallbacks; /* flag: function did callbacks */
2569 int repeats; /* for issuing console warnings if
2570 * looping too many times */
2571 int wake = 0;
2572
2573 spin_lock(&log->l_icloglock);
2574 first_iclog = iclog = log->l_iclog;
2575 ioerrors = 0;
2576 funcdidcallbacks = 0;
2577 repeats = 0;
2578
2579 do {
2580 /*
2581 * Scan all iclogs starting with the one pointed to by the
2582 * log. Reset this starting point each time the log is
2583 * unlocked (during callbacks).
2584 *
2585 * Keep looping through iclogs until one full pass is made
2586 * without running any callbacks.
2587 */
2588 first_iclog = log->l_iclog;
2589 iclog = log->l_iclog;
2590 loopdidcallbacks = 0;
2591 repeats++;
2592
2593 do {
2594
2595 /* skip all iclogs in the ACTIVE & DIRTY states */
2596 if (iclog->ic_state &
2597 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2598 iclog = iclog->ic_next;
2599 continue;
2600 }
2601
2602 /*
2603 * Between marking a filesystem SHUTDOWN and stopping
2604 * the log, we do flush all iclogs to disk (if there
2605 * wasn't a log I/O error). So, we do want things to
2606 * go smoothly in case of just a SHUTDOWN w/o a
2607 * LOG_IO_ERROR.
2608 */
2609 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2610 /*
2611 * Can only perform callbacks in order. Since
2612 * this iclog is not in the DONE_SYNC/
2613 * DO_CALLBACK state, we skip the rest and
2614 * just try to clean up. If we set our iclog
2615 * to DO_CALLBACK, we will not process it when
2616 * we retry since a previous iclog is in the
2617 * CALLBACK and the state cannot change since
2618 * we are holding the l_icloglock.
2619 */
2620 if (!(iclog->ic_state &
2621 (XLOG_STATE_DONE_SYNC |
2622 XLOG_STATE_DO_CALLBACK))) {
2623 if (ciclog && (ciclog->ic_state ==
2624 XLOG_STATE_DONE_SYNC)) {
2625 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2626 }
2627 break;
2628 }
2629 /*
2630 * We now have an iclog that is in either the
2631 * DO_CALLBACK or DONE_SYNC states. The other
2632 * states (WANT_SYNC, SYNCING, or CALLBACK were
2633 * caught by the above if and are going to
2634 * clean (i.e. we aren't doing their callbacks)
2635 * see the above if.
2636 */
2637
2638 /*
2639 * We will do one more check here to see if we
2640 * have chased our tail around.
2641 */
2642
2643 lowest_lsn = xlog_get_lowest_lsn(log);
2644 if (lowest_lsn &&
2645 XFS_LSN_CMP(lowest_lsn,
2646 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2647 iclog = iclog->ic_next;
2648 continue; /* Leave this iclog for
2649 * another thread */
2650 }
2651
2652 iclog->ic_state = XLOG_STATE_CALLBACK;
2653
2654
2655 /*
2656 * Completion of a iclog IO does not imply that
2657 * a transaction has completed, as transactions
2658 * can be large enough to span many iclogs. We
2659 * cannot change the tail of the log half way
2660 * through a transaction as this may be the only
2661 * transaction in the log and moving th etail to
2662 * point to the middle of it will prevent
2663 * recovery from finding the start of the
2664 * transaction. Hence we should only update the
2665 * last_sync_lsn if this iclog contains
2666 * transaction completion callbacks on it.
2667 *
2668 * We have to do this before we drop the
2669 * icloglock to ensure we are the only one that
2670 * can update it.
2671 */
2672 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2673 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2674 if (iclog->ic_callback)
2675 atomic64_set(&log->l_last_sync_lsn,
2676 be64_to_cpu(iclog->ic_header.h_lsn));
2677
2678 } else
2679 ioerrors++;
2680
2681 spin_unlock(&log->l_icloglock);
2682
2683 /*
2684 * Keep processing entries in the callback list until
2685 * we come around and it is empty. We need to
2686 * atomically see that the list is empty and change the
2687 * state to DIRTY so that we don't miss any more
2688 * callbacks being added.
2689 */
2690 spin_lock(&iclog->ic_callback_lock);
2691 cb = iclog->ic_callback;
2692 while (cb) {
2693 iclog->ic_callback_tail = &(iclog->ic_callback);
2694 iclog->ic_callback = NULL;
2695 spin_unlock(&iclog->ic_callback_lock);
2696
2697 /* perform callbacks in the order given */
2698 for (; cb; cb = cb_next) {
2699 cb_next = cb->cb_next;
2700 cb->cb_func(cb->cb_arg, aborted);
2701 }
2702 spin_lock(&iclog->ic_callback_lock);
2703 cb = iclog->ic_callback;
2704 }
2705
2706 loopdidcallbacks++;
2707 funcdidcallbacks++;
2708
2709 spin_lock(&log->l_icloglock);
2710 ASSERT(iclog->ic_callback == NULL);
2711 spin_unlock(&iclog->ic_callback_lock);
2712 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2713 iclog->ic_state = XLOG_STATE_DIRTY;
2714
2715 /*
2716 * Transition from DIRTY to ACTIVE if applicable.
2717 * NOP if STATE_IOERROR.
2718 */
2719 xlog_state_clean_log(log);
2720
2721 /* wake up threads waiting in xfs_log_force() */
2722 wake_up_all(&iclog->ic_force_wait);
2723
2724 iclog = iclog->ic_next;
2725 } while (first_iclog != iclog);
2726
2727 if (repeats > 5000) {
2728 flushcnt += repeats;
2729 repeats = 0;
2730 xfs_warn(log->l_mp,
2731 "%s: possible infinite loop (%d iterations)",
2732 __func__, flushcnt);
2733 }
2734 } while (!ioerrors && loopdidcallbacks);
2735
2736 /*
2737 * make one last gasp attempt to see if iclogs are being left in
2738 * limbo..
2739 */
2740#ifdef DEBUG
2741 if (funcdidcallbacks) {
2742 first_iclog = iclog = log->l_iclog;
2743 do {
2744 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2745 /*
2746 * Terminate the loop if iclogs are found in states
2747 * which will cause other threads to clean up iclogs.
2748 *
2749 * SYNCING - i/o completion will go through logs
2750 * DONE_SYNC - interrupt thread should be waiting for
2751 * l_icloglock
2752 * IOERROR - give up hope all ye who enter here
2753 */
2754 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2755 iclog->ic_state == XLOG_STATE_SYNCING ||
2756 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2757 iclog->ic_state == XLOG_STATE_IOERROR )
2758 break;
2759 iclog = iclog->ic_next;
2760 } while (first_iclog != iclog);
2761 }
2762#endif
2763
2764 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2765 wake = 1;
2766 spin_unlock(&log->l_icloglock);
2767
2768 if (wake)
2769 wake_up_all(&log->l_flush_wait);
2770}
2771
2772
2773/*
2774 * Finish transitioning this iclog to the dirty state.
2775 *
2776 * Make sure that we completely execute this routine only when this is
2777 * the last call to the iclog. There is a good chance that iclog flushes,
2778 * when we reach the end of the physical log, get turned into 2 separate
2779 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2780 * routine. By using the reference count bwritecnt, we guarantee that only
2781 * the second completion goes through.
2782 *
2783 * Callbacks could take time, so they are done outside the scope of the
2784 * global state machine log lock.
2785 */
2786STATIC void
2787xlog_state_done_syncing(
2788 xlog_in_core_t *iclog,
2789 int aborted)
2790{
2791 struct xlog *log = iclog->ic_log;
2792
2793 spin_lock(&log->l_icloglock);
2794
2795 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2796 iclog->ic_state == XLOG_STATE_IOERROR);
2797 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2798 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2799
2800
2801 /*
2802 * If we got an error, either on the first buffer, or in the case of
2803 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2804 * and none should ever be attempted to be written to disk
2805 * again.
2806 */
2807 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2808 if (--iclog->ic_bwritecnt == 1) {
2809 spin_unlock(&log->l_icloglock);
2810 return;
2811 }
2812 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2813 }
2814
2815 /*
2816 * Someone could be sleeping prior to writing out the next
2817 * iclog buffer, we wake them all, one will get to do the
2818 * I/O, the others get to wait for the result.
2819 */
2820 wake_up_all(&iclog->ic_write_wait);
2821 spin_unlock(&log->l_icloglock);
2822 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2823} /* xlog_state_done_syncing */
2824
2825
2826/*
2827 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2828 * sleep. We wait on the flush queue on the head iclog as that should be
2829 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2830 * we will wait here and all new writes will sleep until a sync completes.
2831 *
2832 * The in-core logs are used in a circular fashion. They are not used
2833 * out-of-order even when an iclog past the head is free.
2834 *
2835 * return:
2836 * * log_offset where xlog_write() can start writing into the in-core
2837 * log's data space.
2838 * * in-core log pointer to which xlog_write() should write.
2839 * * boolean indicating this is a continued write to an in-core log.
2840 * If this is the last write, then the in-core log's offset field
2841 * needs to be incremented, depending on the amount of data which
2842 * is copied.
2843 */
2844STATIC int
2845xlog_state_get_iclog_space(
2846 struct xlog *log,
2847 int len,
2848 struct xlog_in_core **iclogp,
2849 struct xlog_ticket *ticket,
2850 int *continued_write,
2851 int *logoffsetp)
2852{
2853 int log_offset;
2854 xlog_rec_header_t *head;
2855 xlog_in_core_t *iclog;
2856 int error;
2857
2858restart:
2859 spin_lock(&log->l_icloglock);
2860 if (XLOG_FORCED_SHUTDOWN(log)) {
2861 spin_unlock(&log->l_icloglock);
2862 return XFS_ERROR(EIO);
2863 }
2864
2865 iclog = log->l_iclog;
2866 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2867 XFS_STATS_INC(xs_log_noiclogs);
2868
2869 /* Wait for log writes to have flushed */
2870 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2871 goto restart;
2872 }
2873
2874 head = &iclog->ic_header;
2875
2876 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2877 log_offset = iclog->ic_offset;
2878
2879 /* On the 1st write to an iclog, figure out lsn. This works
2880 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2881 * committing to. If the offset is set, that's how many blocks
2882 * must be written.
2883 */
2884 if (log_offset == 0) {
2885 ticket->t_curr_res -= log->l_iclog_hsize;
2886 xlog_tic_add_region(ticket,
2887 log->l_iclog_hsize,
2888 XLOG_REG_TYPE_LRHEADER);
2889 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2890 head->h_lsn = cpu_to_be64(
2891 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2892 ASSERT(log->l_curr_block >= 0);
2893 }
2894
2895 /* If there is enough room to write everything, then do it. Otherwise,
2896 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2897 * bit is on, so this will get flushed out. Don't update ic_offset
2898 * until you know exactly how many bytes get copied. Therefore, wait
2899 * until later to update ic_offset.
2900 *
2901 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2902 * can fit into remaining data section.
2903 */
2904 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2905 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2906
2907 /*
2908 * If I'm the only one writing to this iclog, sync it to disk.
2909 * We need to do an atomic compare and decrement here to avoid
2910 * racing with concurrent atomic_dec_and_lock() calls in
2911 * xlog_state_release_iclog() when there is more than one
2912 * reference to the iclog.
2913 */
2914 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2915 /* we are the only one */
2916 spin_unlock(&log->l_icloglock);
2917 error = xlog_state_release_iclog(log, iclog);
2918 if (error)
2919 return error;
2920 } else {
2921 spin_unlock(&log->l_icloglock);
2922 }
2923 goto restart;
2924 }
2925
2926 /* Do we have enough room to write the full amount in the remainder
2927 * of this iclog? Or must we continue a write on the next iclog and
2928 * mark this iclog as completely taken? In the case where we switch
2929 * iclogs (to mark it taken), this particular iclog will release/sync
2930 * to disk in xlog_write().
2931 */
2932 if (len <= iclog->ic_size - iclog->ic_offset) {
2933 *continued_write = 0;
2934 iclog->ic_offset += len;
2935 } else {
2936 *continued_write = 1;
2937 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2938 }
2939 *iclogp = iclog;
2940
2941 ASSERT(iclog->ic_offset <= iclog->ic_size);
2942 spin_unlock(&log->l_icloglock);
2943
2944 *logoffsetp = log_offset;
2945 return 0;
2946} /* xlog_state_get_iclog_space */
2947
2948/* The first cnt-1 times through here we don't need to
2949 * move the grant write head because the permanent
2950 * reservation has reserved cnt times the unit amount.
2951 * Release part of current permanent unit reservation and
2952 * reset current reservation to be one units worth. Also
2953 * move grant reservation head forward.
2954 */
2955STATIC void
2956xlog_regrant_reserve_log_space(
2957 struct xlog *log,
2958 struct xlog_ticket *ticket)
2959{
2960 trace_xfs_log_regrant_reserve_enter(log, ticket);
2961
2962 if (ticket->t_cnt > 0)
2963 ticket->t_cnt--;
2964
2965 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2966 ticket->t_curr_res);
2967 xlog_grant_sub_space(log, &log->l_write_head.grant,
2968 ticket->t_curr_res);
2969 ticket->t_curr_res = ticket->t_unit_res;
2970 xlog_tic_reset_res(ticket);
2971
2972 trace_xfs_log_regrant_reserve_sub(log, ticket);
2973
2974 /* just return if we still have some of the pre-reserved space */
2975 if (ticket->t_cnt > 0)
2976 return;
2977
2978 xlog_grant_add_space(log, &log->l_reserve_head.grant,
2979 ticket->t_unit_res);
2980
2981 trace_xfs_log_regrant_reserve_exit(log, ticket);
2982
2983 ticket->t_curr_res = ticket->t_unit_res;
2984 xlog_tic_reset_res(ticket);
2985} /* xlog_regrant_reserve_log_space */
2986
2987
2988/*
2989 * Give back the space left from a reservation.
2990 *
2991 * All the information we need to make a correct determination of space left
2992 * is present. For non-permanent reservations, things are quite easy. The
2993 * count should have been decremented to zero. We only need to deal with the
2994 * space remaining in the current reservation part of the ticket. If the
2995 * ticket contains a permanent reservation, there may be left over space which
2996 * needs to be released. A count of N means that N-1 refills of the current
2997 * reservation can be done before we need to ask for more space. The first
2998 * one goes to fill up the first current reservation. Once we run out of
2999 * space, the count will stay at zero and the only space remaining will be
3000 * in the current reservation field.
3001 */
3002STATIC void
3003xlog_ungrant_log_space(
3004 struct xlog *log,
3005 struct xlog_ticket *ticket)
3006{
3007 int bytes;
3008
3009 if (ticket->t_cnt > 0)
3010 ticket->t_cnt--;
3011
3012 trace_xfs_log_ungrant_enter(log, ticket);
3013 trace_xfs_log_ungrant_sub(log, ticket);
3014
3015 /*
3016 * If this is a permanent reservation ticket, we may be able to free
3017 * up more space based on the remaining count.
3018 */
3019 bytes = ticket->t_curr_res;
3020 if (ticket->t_cnt > 0) {
3021 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3022 bytes += ticket->t_unit_res*ticket->t_cnt;
3023 }
3024
3025 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3026 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3027
3028 trace_xfs_log_ungrant_exit(log, ticket);
3029
3030 xfs_log_space_wake(log->l_mp);
3031}
3032
3033/*
3034 * Flush iclog to disk if this is the last reference to the given iclog and
3035 * the WANT_SYNC bit is set.
3036 *
3037 * When this function is entered, the iclog is not necessarily in the
3038 * WANT_SYNC state. It may be sitting around waiting to get filled.
3039 *
3040 *
3041 */
3042STATIC int
3043xlog_state_release_iclog(
3044 struct xlog *log,
3045 struct xlog_in_core *iclog)
3046{
3047 int sync = 0; /* do we sync? */
3048
3049 if (iclog->ic_state & XLOG_STATE_IOERROR)
3050 return XFS_ERROR(EIO);
3051
3052 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3053 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3054 return 0;
3055
3056 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3057 spin_unlock(&log->l_icloglock);
3058 return XFS_ERROR(EIO);
3059 }
3060 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3061 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3062
3063 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3064 /* update tail before writing to iclog */
3065 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3066 sync++;
3067 iclog->ic_state = XLOG_STATE_SYNCING;
3068 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3069 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3070 /* cycle incremented when incrementing curr_block */
3071 }
3072 spin_unlock(&log->l_icloglock);
3073
3074 /*
3075 * We let the log lock go, so it's possible that we hit a log I/O
3076 * error or some other SHUTDOWN condition that marks the iclog
3077 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3078 * this iclog has consistent data, so we ignore IOERROR
3079 * flags after this point.
3080 */
3081 if (sync)
3082 return xlog_sync(log, iclog);
3083 return 0;
3084} /* xlog_state_release_iclog */
3085
3086
3087/*
3088 * This routine will mark the current iclog in the ring as WANT_SYNC
3089 * and move the current iclog pointer to the next iclog in the ring.
3090 * When this routine is called from xlog_state_get_iclog_space(), the
3091 * exact size of the iclog has not yet been determined. All we know is
3092 * that every data block. We have run out of space in this log record.
3093 */
3094STATIC void
3095xlog_state_switch_iclogs(
3096 struct xlog *log,
3097 struct xlog_in_core *iclog,
3098 int eventual_size)
3099{
3100 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3101 if (!eventual_size)
3102 eventual_size = iclog->ic_offset;
3103 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3104 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3105 log->l_prev_block = log->l_curr_block;
3106 log->l_prev_cycle = log->l_curr_cycle;
3107
3108 /* roll log?: ic_offset changed later */
3109 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3110
3111 /* Round up to next log-sunit */
3112 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3113 log->l_mp->m_sb.sb_logsunit > 1) {
3114 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3115 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3116 }
3117
3118 if (log->l_curr_block >= log->l_logBBsize) {
3119 log->l_curr_cycle++;
3120 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3121 log->l_curr_cycle++;
3122 log->l_curr_block -= log->l_logBBsize;
3123 ASSERT(log->l_curr_block >= 0);
3124 }
3125 ASSERT(iclog == log->l_iclog);
3126 log->l_iclog = iclog->ic_next;
3127} /* xlog_state_switch_iclogs */
3128
3129/*
3130 * Write out all data in the in-core log as of this exact moment in time.
3131 *
3132 * Data may be written to the in-core log during this call. However,
3133 * we don't guarantee this data will be written out. A change from past
3134 * implementation means this routine will *not* write out zero length LRs.
3135 *
3136 * Basically, we try and perform an intelligent scan of the in-core logs.
3137 * If we determine there is no flushable data, we just return. There is no
3138 * flushable data if:
3139 *
3140 * 1. the current iclog is active and has no data; the previous iclog
3141 * is in the active or dirty state.
3142 * 2. the current iclog is drity, and the previous iclog is in the
3143 * active or dirty state.
3144 *
3145 * We may sleep if:
3146 *
3147 * 1. the current iclog is not in the active nor dirty state.
3148 * 2. the current iclog dirty, and the previous iclog is not in the
3149 * active nor dirty state.
3150 * 3. the current iclog is active, and there is another thread writing
3151 * to this particular iclog.
3152 * 4. a) the current iclog is active and has no other writers
3153 * b) when we return from flushing out this iclog, it is still
3154 * not in the active nor dirty state.
3155 */
3156int
3157_xfs_log_force(
3158 struct xfs_mount *mp,
3159 uint flags,
3160 int *log_flushed)
3161{
3162 struct xlog *log = mp->m_log;
3163 struct xlog_in_core *iclog;
3164 xfs_lsn_t lsn;
3165
3166 XFS_STATS_INC(xs_log_force);
3167
3168 xlog_cil_force(log);
3169
3170 spin_lock(&log->l_icloglock);
3171
3172 iclog = log->l_iclog;
3173 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3174 spin_unlock(&log->l_icloglock);
3175 return XFS_ERROR(EIO);
3176 }
3177
3178 /* If the head iclog is not active nor dirty, we just attach
3179 * ourselves to the head and go to sleep.
3180 */
3181 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3182 iclog->ic_state == XLOG_STATE_DIRTY) {
3183 /*
3184 * If the head is dirty or (active and empty), then
3185 * we need to look at the previous iclog. If the previous
3186 * iclog is active or dirty we are done. There is nothing
3187 * to sync out. Otherwise, we attach ourselves to the
3188 * previous iclog and go to sleep.
3189 */
3190 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3191 (atomic_read(&iclog->ic_refcnt) == 0
3192 && iclog->ic_offset == 0)) {
3193 iclog = iclog->ic_prev;
3194 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3195 iclog->ic_state == XLOG_STATE_DIRTY)
3196 goto no_sleep;
3197 else
3198 goto maybe_sleep;
3199 } else {
3200 if (atomic_read(&iclog->ic_refcnt) == 0) {
3201 /* We are the only one with access to this
3202 * iclog. Flush it out now. There should
3203 * be a roundoff of zero to show that someone
3204 * has already taken care of the roundoff from
3205 * the previous sync.
3206 */
3207 atomic_inc(&iclog->ic_refcnt);
3208 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3209 xlog_state_switch_iclogs(log, iclog, 0);
3210 spin_unlock(&log->l_icloglock);
3211
3212 if (xlog_state_release_iclog(log, iclog))
3213 return XFS_ERROR(EIO);
3214
3215 if (log_flushed)
3216 *log_flushed = 1;
3217 spin_lock(&log->l_icloglock);
3218 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3219 iclog->ic_state != XLOG_STATE_DIRTY)
3220 goto maybe_sleep;
3221 else
3222 goto no_sleep;
3223 } else {
3224 /* Someone else is writing to this iclog.
3225 * Use its call to flush out the data. However,
3226 * the other thread may not force out this LR,
3227 * so we mark it WANT_SYNC.
3228 */
3229 xlog_state_switch_iclogs(log, iclog, 0);
3230 goto maybe_sleep;
3231 }
3232 }
3233 }
3234
3235 /* By the time we come around again, the iclog could've been filled
3236 * which would give it another lsn. If we have a new lsn, just
3237 * return because the relevant data has been flushed.
3238 */
3239maybe_sleep:
3240 if (flags & XFS_LOG_SYNC) {
3241 /*
3242 * We must check if we're shutting down here, before
3243 * we wait, while we're holding the l_icloglock.
3244 * Then we check again after waking up, in case our
3245 * sleep was disturbed by a bad news.
3246 */
3247 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3248 spin_unlock(&log->l_icloglock);
3249 return XFS_ERROR(EIO);
3250 }
3251 XFS_STATS_INC(xs_log_force_sleep);
3252 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3253 /*
3254 * No need to grab the log lock here since we're
3255 * only deciding whether or not to return EIO
3256 * and the memory read should be atomic.
3257 */
3258 if (iclog->ic_state & XLOG_STATE_IOERROR)
3259 return XFS_ERROR(EIO);
3260 if (log_flushed)
3261 *log_flushed = 1;
3262 } else {
3263
3264no_sleep:
3265 spin_unlock(&log->l_icloglock);
3266 }
3267 return 0;
3268}
3269
3270/*
3271 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3272 * about errors or whether the log was flushed or not. This is the normal
3273 * interface to use when trying to unpin items or move the log forward.
3274 */
3275void
3276xfs_log_force(
3277 xfs_mount_t *mp,
3278 uint flags)
3279{
3280 int error;
3281
3282 trace_xfs_log_force(mp, 0);
3283 error = _xfs_log_force(mp, flags, NULL);
3284 if (error)
3285 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3286}
3287
3288/*
3289 * Force the in-core log to disk for a specific LSN.
3290 *
3291 * Find in-core log with lsn.
3292 * If it is in the DIRTY state, just return.
3293 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3294 * state and go to sleep or return.
3295 * If it is in any other state, go to sleep or return.
3296 *
3297 * Synchronous forces are implemented with a signal variable. All callers
3298 * to force a given lsn to disk will wait on a the sv attached to the
3299 * specific in-core log. When given in-core log finally completes its
3300 * write to disk, that thread will wake up all threads waiting on the
3301 * sv.
3302 */
3303int
3304_xfs_log_force_lsn(
3305 struct xfs_mount *mp,
3306 xfs_lsn_t lsn,
3307 uint flags,
3308 int *log_flushed)
3309{
3310 struct xlog *log = mp->m_log;
3311 struct xlog_in_core *iclog;
3312 int already_slept = 0;
3313
3314 ASSERT(lsn != 0);
3315
3316 XFS_STATS_INC(xs_log_force);
3317
3318 lsn = xlog_cil_force_lsn(log, lsn);
3319 if (lsn == NULLCOMMITLSN)
3320 return 0;
3321
3322try_again:
3323 spin_lock(&log->l_icloglock);
3324 iclog = log->l_iclog;
3325 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3326 spin_unlock(&log->l_icloglock);
3327 return XFS_ERROR(EIO);
3328 }
3329
3330 do {
3331 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3332 iclog = iclog->ic_next;
3333 continue;
3334 }
3335
3336 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3337 spin_unlock(&log->l_icloglock);
3338 return 0;
3339 }
3340
3341 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3342 /*
3343 * We sleep here if we haven't already slept (e.g.
3344 * this is the first time we've looked at the correct
3345 * iclog buf) and the buffer before us is going to
3346 * be sync'ed. The reason for this is that if we
3347 * are doing sync transactions here, by waiting for
3348 * the previous I/O to complete, we can allow a few
3349 * more transactions into this iclog before we close
3350 * it down.
3351 *
3352 * Otherwise, we mark the buffer WANT_SYNC, and bump
3353 * up the refcnt so we can release the log (which
3354 * drops the ref count). The state switch keeps new
3355 * transaction commits from using this buffer. When
3356 * the current commits finish writing into the buffer,
3357 * the refcount will drop to zero and the buffer will
3358 * go out then.
3359 */
3360 if (!already_slept &&
3361 (iclog->ic_prev->ic_state &
3362 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3363 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3364
3365 XFS_STATS_INC(xs_log_force_sleep);
3366
3367 xlog_wait(&iclog->ic_prev->ic_write_wait,
3368 &log->l_icloglock);
3369 if (log_flushed)
3370 *log_flushed = 1;
3371 already_slept = 1;
3372 goto try_again;
3373 }
3374 atomic_inc(&iclog->ic_refcnt);
3375 xlog_state_switch_iclogs(log, iclog, 0);
3376 spin_unlock(&log->l_icloglock);
3377 if (xlog_state_release_iclog(log, iclog))
3378 return XFS_ERROR(EIO);
3379 if (log_flushed)
3380 *log_flushed = 1;
3381 spin_lock(&log->l_icloglock);
3382 }
3383
3384 if ((flags & XFS_LOG_SYNC) && /* sleep */
3385 !(iclog->ic_state &
3386 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3387 /*
3388 * Don't wait on completion if we know that we've
3389 * gotten a log write error.
3390 */
3391 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3392 spin_unlock(&log->l_icloglock);
3393 return XFS_ERROR(EIO);
3394 }
3395 XFS_STATS_INC(xs_log_force_sleep);
3396 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3397 /*
3398 * No need to grab the log lock here since we're
3399 * only deciding whether or not to return EIO
3400 * and the memory read should be atomic.
3401 */
3402 if (iclog->ic_state & XLOG_STATE_IOERROR)
3403 return XFS_ERROR(EIO);
3404
3405 if (log_flushed)
3406 *log_flushed = 1;
3407 } else { /* just return */
3408 spin_unlock(&log->l_icloglock);
3409 }
3410
3411 return 0;
3412 } while (iclog != log->l_iclog);
3413
3414 spin_unlock(&log->l_icloglock);
3415 return 0;
3416}
3417
3418/*
3419 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3420 * about errors or whether the log was flushed or not. This is the normal
3421 * interface to use when trying to unpin items or move the log forward.
3422 */
3423void
3424xfs_log_force_lsn(
3425 xfs_mount_t *mp,
3426 xfs_lsn_t lsn,
3427 uint flags)
3428{
3429 int error;
3430
3431 trace_xfs_log_force(mp, lsn);
3432 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3433 if (error)
3434 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3435}
3436
3437/*
3438 * Called when we want to mark the current iclog as being ready to sync to
3439 * disk.
3440 */
3441STATIC void
3442xlog_state_want_sync(
3443 struct xlog *log,
3444 struct xlog_in_core *iclog)
3445{
3446 assert_spin_locked(&log->l_icloglock);
3447
3448 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3449 xlog_state_switch_iclogs(log, iclog, 0);
3450 } else {
3451 ASSERT(iclog->ic_state &
3452 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3453 }
3454}
3455
3456
3457/*****************************************************************************
3458 *
3459 * TICKET functions
3460 *
3461 *****************************************************************************
3462 */
3463
3464/*
3465 * Free a used ticket when its refcount falls to zero.
3466 */
3467void
3468xfs_log_ticket_put(
3469 xlog_ticket_t *ticket)
3470{
3471 ASSERT(atomic_read(&ticket->t_ref) > 0);
3472 if (atomic_dec_and_test(&ticket->t_ref))
3473 kmem_zone_free(xfs_log_ticket_zone, ticket);
3474}
3475
3476xlog_ticket_t *
3477xfs_log_ticket_get(
3478 xlog_ticket_t *ticket)
3479{
3480 ASSERT(atomic_read(&ticket->t_ref) > 0);
3481 atomic_inc(&ticket->t_ref);
3482 return ticket;
3483}
3484
3485/*
3486 * Figure out the total log space unit (in bytes) that would be
3487 * required for a log ticket.
3488 */
3489int
3490xfs_log_calc_unit_res(
3491 struct xfs_mount *mp,
3492 int unit_bytes)
3493{
3494 struct xlog *log = mp->m_log;
3495 int iclog_space;
3496 uint num_headers;
3497
3498 /*
3499 * Permanent reservations have up to 'cnt'-1 active log operations
3500 * in the log. A unit in this case is the amount of space for one
3501 * of these log operations. Normal reservations have a cnt of 1
3502 * and their unit amount is the total amount of space required.
3503 *
3504 * The following lines of code account for non-transaction data
3505 * which occupy space in the on-disk log.
3506 *
3507 * Normal form of a transaction is:
3508 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3509 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3510 *
3511 * We need to account for all the leadup data and trailer data
3512 * around the transaction data.
3513 * And then we need to account for the worst case in terms of using
3514 * more space.
3515 * The worst case will happen if:
3516 * - the placement of the transaction happens to be such that the
3517 * roundoff is at its maximum
3518 * - the transaction data is synced before the commit record is synced
3519 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3520 * Therefore the commit record is in its own Log Record.
3521 * This can happen as the commit record is called with its
3522 * own region to xlog_write().
3523 * This then means that in the worst case, roundoff can happen for
3524 * the commit-rec as well.
3525 * The commit-rec is smaller than padding in this scenario and so it is
3526 * not added separately.
3527 */
3528
3529 /* for trans header */
3530 unit_bytes += sizeof(xlog_op_header_t);
3531 unit_bytes += sizeof(xfs_trans_header_t);
3532
3533 /* for start-rec */
3534 unit_bytes += sizeof(xlog_op_header_t);
3535
3536 /*
3537 * for LR headers - the space for data in an iclog is the size minus
3538 * the space used for the headers. If we use the iclog size, then we
3539 * undercalculate the number of headers required.
3540 *
3541 * Furthermore - the addition of op headers for split-recs might
3542 * increase the space required enough to require more log and op
3543 * headers, so take that into account too.
3544 *
3545 * IMPORTANT: This reservation makes the assumption that if this
3546 * transaction is the first in an iclog and hence has the LR headers
3547 * accounted to it, then the remaining space in the iclog is
3548 * exclusively for this transaction. i.e. if the transaction is larger
3549 * than the iclog, it will be the only thing in that iclog.
3550 * Fundamentally, this means we must pass the entire log vector to
3551 * xlog_write to guarantee this.
3552 */
3553 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3554 num_headers = howmany(unit_bytes, iclog_space);
3555
3556 /* for split-recs - ophdrs added when data split over LRs */
3557 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3558
3559 /* add extra header reservations if we overrun */
3560 while (!num_headers ||
3561 howmany(unit_bytes, iclog_space) > num_headers) {
3562 unit_bytes += sizeof(xlog_op_header_t);
3563 num_headers++;
3564 }
3565 unit_bytes += log->l_iclog_hsize * num_headers;
3566
3567 /* for commit-rec LR header - note: padding will subsume the ophdr */
3568 unit_bytes += log->l_iclog_hsize;
3569
3570 /* for roundoff padding for transaction data and one for commit record */
3571 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3572 /* log su roundoff */
3573 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3574 } else {
3575 /* BB roundoff */
3576 unit_bytes += 2 * BBSIZE;
3577 }
3578
3579 return unit_bytes;
3580}
3581
3582/*
3583 * Allocate and initialise a new log ticket.
3584 */
3585struct xlog_ticket *
3586xlog_ticket_alloc(
3587 struct xlog *log,
3588 int unit_bytes,
3589 int cnt,
3590 char client,
3591 bool permanent,
3592 xfs_km_flags_t alloc_flags)
3593{
3594 struct xlog_ticket *tic;
3595 int unit_res;
3596
3597 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3598 if (!tic)
3599 return NULL;
3600
3601 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3602
3603 atomic_set(&tic->t_ref, 1);
3604 tic->t_task = current;
3605 INIT_LIST_HEAD(&tic->t_queue);
3606 tic->t_unit_res = unit_res;
3607 tic->t_curr_res = unit_res;
3608 tic->t_cnt = cnt;
3609 tic->t_ocnt = cnt;
3610 tic->t_tid = prandom_u32();
3611 tic->t_clientid = client;
3612 tic->t_flags = XLOG_TIC_INITED;
3613 tic->t_trans_type = 0;
3614 if (permanent)
3615 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3616
3617 xlog_tic_reset_res(tic);
3618
3619 return tic;
3620}
3621
3622
3623/******************************************************************************
3624 *
3625 * Log debug routines
3626 *
3627 ******************************************************************************
3628 */
3629#if defined(DEBUG)
3630/*
3631 * Make sure that the destination ptr is within the valid data region of
3632 * one of the iclogs. This uses backup pointers stored in a different
3633 * part of the log in case we trash the log structure.
3634 */
3635void
3636xlog_verify_dest_ptr(
3637 struct xlog *log,
3638 char *ptr)
3639{
3640 int i;
3641 int good_ptr = 0;
3642
3643 for (i = 0; i < log->l_iclog_bufs; i++) {
3644 if (ptr >= log->l_iclog_bak[i] &&
3645 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3646 good_ptr++;
3647 }
3648
3649 if (!good_ptr)
3650 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3651}
3652
3653/*
3654 * Check to make sure the grant write head didn't just over lap the tail. If
3655 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3656 * the cycles differ by exactly one and check the byte count.
3657 *
3658 * This check is run unlocked, so can give false positives. Rather than assert
3659 * on failures, use a warn-once flag and a panic tag to allow the admin to
3660 * determine if they want to panic the machine when such an error occurs. For
3661 * debug kernels this will have the same effect as using an assert but, unlinke
3662 * an assert, it can be turned off at runtime.
3663 */
3664STATIC void
3665xlog_verify_grant_tail(
3666 struct xlog *log)
3667{
3668 int tail_cycle, tail_blocks;
3669 int cycle, space;
3670
3671 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3672 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3673 if (tail_cycle != cycle) {
3674 if (cycle - 1 != tail_cycle &&
3675 !(log->l_flags & XLOG_TAIL_WARN)) {
3676 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3677 "%s: cycle - 1 != tail_cycle", __func__);
3678 log->l_flags |= XLOG_TAIL_WARN;
3679 }
3680
3681 if (space > BBTOB(tail_blocks) &&
3682 !(log->l_flags & XLOG_TAIL_WARN)) {
3683 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3684 "%s: space > BBTOB(tail_blocks)", __func__);
3685 log->l_flags |= XLOG_TAIL_WARN;
3686 }
3687 }
3688}
3689
3690/* check if it will fit */
3691STATIC void
3692xlog_verify_tail_lsn(
3693 struct xlog *log,
3694 struct xlog_in_core *iclog,
3695 xfs_lsn_t tail_lsn)
3696{
3697 int blocks;
3698
3699 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3700 blocks =
3701 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3702 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3703 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3704 } else {
3705 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3706
3707 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3708 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3709
3710 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3711 if (blocks < BTOBB(iclog->ic_offset) + 1)
3712 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3713 }
3714} /* xlog_verify_tail_lsn */
3715
3716/*
3717 * Perform a number of checks on the iclog before writing to disk.
3718 *
3719 * 1. Make sure the iclogs are still circular
3720 * 2. Make sure we have a good magic number
3721 * 3. Make sure we don't have magic numbers in the data
3722 * 4. Check fields of each log operation header for:
3723 * A. Valid client identifier
3724 * B. tid ptr value falls in valid ptr space (user space code)
3725 * C. Length in log record header is correct according to the
3726 * individual operation headers within record.
3727 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3728 * log, check the preceding blocks of the physical log to make sure all
3729 * the cycle numbers agree with the current cycle number.
3730 */
3731STATIC void
3732xlog_verify_iclog(
3733 struct xlog *log,
3734 struct xlog_in_core *iclog,
3735 int count,
3736 bool syncing)
3737{
3738 xlog_op_header_t *ophead;
3739 xlog_in_core_t *icptr;
3740 xlog_in_core_2_t *xhdr;
3741 xfs_caddr_t ptr;
3742 xfs_caddr_t base_ptr;
3743 __psint_t field_offset;
3744 __uint8_t clientid;
3745 int len, i, j, k, op_len;
3746 int idx;
3747
3748 /* check validity of iclog pointers */
3749 spin_lock(&log->l_icloglock);
3750 icptr = log->l_iclog;
3751 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3752 ASSERT(icptr);
3753
3754 if (icptr != log->l_iclog)
3755 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3756 spin_unlock(&log->l_icloglock);
3757
3758 /* check log magic numbers */
3759 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3760 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3761
3762 ptr = (xfs_caddr_t) &iclog->ic_header;
3763 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3764 ptr += BBSIZE) {
3765 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3766 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3767 __func__);
3768 }
3769
3770 /* check fields */
3771 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3772 ptr = iclog->ic_datap;
3773 base_ptr = ptr;
3774 ophead = (xlog_op_header_t *)ptr;
3775 xhdr = iclog->ic_data;
3776 for (i = 0; i < len; i++) {
3777 ophead = (xlog_op_header_t *)ptr;
3778
3779 /* clientid is only 1 byte */
3780 field_offset = (__psint_t)
3781 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3782 if (!syncing || (field_offset & 0x1ff)) {
3783 clientid = ophead->oh_clientid;
3784 } else {
3785 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3786 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3787 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3788 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3789 clientid = xlog_get_client_id(
3790 xhdr[j].hic_xheader.xh_cycle_data[k]);
3791 } else {
3792 clientid = xlog_get_client_id(
3793 iclog->ic_header.h_cycle_data[idx]);
3794 }
3795 }
3796 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3797 xfs_warn(log->l_mp,
3798 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3799 __func__, clientid, ophead,
3800 (unsigned long)field_offset);
3801
3802 /* check length */
3803 field_offset = (__psint_t)
3804 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3805 if (!syncing || (field_offset & 0x1ff)) {
3806 op_len = be32_to_cpu(ophead->oh_len);
3807 } else {
3808 idx = BTOBBT((__psint_t)&ophead->oh_len -
3809 (__psint_t)iclog->ic_datap);
3810 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3811 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3812 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3813 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3814 } else {
3815 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3816 }
3817 }
3818 ptr += sizeof(xlog_op_header_t) + op_len;
3819 }
3820} /* xlog_verify_iclog */
3821#endif
3822
3823/*
3824 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3825 */
3826STATIC int
3827xlog_state_ioerror(
3828 struct xlog *log)
3829{
3830 xlog_in_core_t *iclog, *ic;
3831
3832 iclog = log->l_iclog;
3833 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3834 /*
3835 * Mark all the incore logs IOERROR.
3836 * From now on, no log flushes will result.
3837 */
3838 ic = iclog;
3839 do {
3840 ic->ic_state = XLOG_STATE_IOERROR;
3841 ic = ic->ic_next;
3842 } while (ic != iclog);
3843 return 0;
3844 }
3845 /*
3846 * Return non-zero, if state transition has already happened.
3847 */
3848 return 1;
3849}
3850
3851/*
3852 * This is called from xfs_force_shutdown, when we're forcibly
3853 * shutting down the filesystem, typically because of an IO error.
3854 * Our main objectives here are to make sure that:
3855 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3856 * parties to find out, 'atomically'.
3857 * b. those who're sleeping on log reservations, pinned objects and
3858 * other resources get woken up, and be told the bad news.
3859 * c. nothing new gets queued up after (a) and (b) are done.
3860 * d. if !logerror, flush the iclogs to disk, then seal them off
3861 * for business.
3862 *
3863 * Note: for delayed logging the !logerror case needs to flush the regions
3864 * held in memory out to the iclogs before flushing them to disk. This needs
3865 * to be done before the log is marked as shutdown, otherwise the flush to the
3866 * iclogs will fail.
3867 */
3868int
3869xfs_log_force_umount(
3870 struct xfs_mount *mp,
3871 int logerror)
3872{
3873 struct xlog *log;
3874 int retval;
3875
3876 log = mp->m_log;
3877
3878 /*
3879 * If this happens during log recovery, don't worry about
3880 * locking; the log isn't open for business yet.
3881 */
3882 if (!log ||
3883 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3884 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3885 if (mp->m_sb_bp)
3886 XFS_BUF_DONE(mp->m_sb_bp);
3887 return 0;
3888 }
3889
3890 /*
3891 * Somebody could've already done the hard work for us.
3892 * No need to get locks for this.
3893 */
3894 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3895 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3896 return 1;
3897 }
3898 retval = 0;
3899
3900 /*
3901 * Flush the in memory commit item list before marking the log as
3902 * being shut down. We need to do it in this order to ensure all the
3903 * completed transactions are flushed to disk with the xfs_log_force()
3904 * call below.
3905 */
3906 if (!logerror)
3907 xlog_cil_force(log);
3908
3909 /*
3910 * mark the filesystem and the as in a shutdown state and wake
3911 * everybody up to tell them the bad news.
3912 */
3913 spin_lock(&log->l_icloglock);
3914 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3915 if (mp->m_sb_bp)
3916 XFS_BUF_DONE(mp->m_sb_bp);
3917
3918 /*
3919 * This flag is sort of redundant because of the mount flag, but
3920 * it's good to maintain the separation between the log and the rest
3921 * of XFS.
3922 */
3923 log->l_flags |= XLOG_IO_ERROR;
3924
3925 /*
3926 * If we hit a log error, we want to mark all the iclogs IOERROR
3927 * while we're still holding the loglock.
3928 */
3929 if (logerror)
3930 retval = xlog_state_ioerror(log);
3931 spin_unlock(&log->l_icloglock);
3932
3933 /*
3934 * We don't want anybody waiting for log reservations after this. That
3935 * means we have to wake up everybody queued up on reserveq as well as
3936 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3937 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3938 * action is protected by the grant locks.
3939 */
3940 xlog_grant_head_wake_all(&log->l_reserve_head);
3941 xlog_grant_head_wake_all(&log->l_write_head);
3942
3943 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3944 ASSERT(!logerror);
3945 /*
3946 * Force the incore logs to disk before shutting the
3947 * log down completely.
3948 */
3949 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3950
3951 spin_lock(&log->l_icloglock);
3952 retval = xlog_state_ioerror(log);
3953 spin_unlock(&log->l_icloglock);
3954 }
3955 /*
3956 * Wake up everybody waiting on xfs_log_force.
3957 * Callback all log item committed functions as if the
3958 * log writes were completed.
3959 */
3960 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3961
3962#ifdef XFSERRORDEBUG
3963 {
3964 xlog_in_core_t *iclog;
3965
3966 spin_lock(&log->l_icloglock);
3967 iclog = log->l_iclog;
3968 do {
3969 ASSERT(iclog->ic_callback == 0);
3970 iclog = iclog->ic_next;
3971 } while (iclog != log->l_iclog);
3972 spin_unlock(&log->l_icloglock);
3973 }
3974#endif
3975 /* return non-zero if log IOERROR transition had already happened */
3976 return retval;
3977}
3978
3979STATIC int
3980xlog_iclogs_empty(
3981 struct xlog *log)
3982{
3983 xlog_in_core_t *iclog;
3984
3985 iclog = log->l_iclog;
3986 do {
3987 /* endianness does not matter here, zero is zero in
3988 * any language.
3989 */
3990 if (iclog->ic_header.h_num_logops)
3991 return 0;
3992 iclog = iclog->ic_next;
3993 } while (iclog != log->l_iclog);
3994 return 1;
3995}
3996
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_errortag.h"
14#include "xfs_error.h"
15#include "xfs_trans.h"
16#include "xfs_trans_priv.h"
17#include "xfs_log.h"
18#include "xfs_log_priv.h"
19#include "xfs_trace.h"
20#include "xfs_sysfs.h"
21#include "xfs_sb.h"
22#include "xfs_health.h"
23
24struct kmem_cache *xfs_log_ticket_cache;
25
26/* Local miscellaneous function prototypes */
27STATIC struct xlog *
28xlog_alloc_log(
29 struct xfs_mount *mp,
30 struct xfs_buftarg *log_target,
31 xfs_daddr_t blk_offset,
32 int num_bblks);
33STATIC void
34xlog_dealloc_log(
35 struct xlog *log);
36
37/* local state machine functions */
38STATIC void xlog_state_done_syncing(
39 struct xlog_in_core *iclog);
40STATIC void xlog_state_do_callback(
41 struct xlog *log);
42STATIC int
43xlog_state_get_iclog_space(
44 struct xlog *log,
45 int len,
46 struct xlog_in_core **iclog,
47 struct xlog_ticket *ticket,
48 int *logoffsetp);
49STATIC void
50xlog_sync(
51 struct xlog *log,
52 struct xlog_in_core *iclog,
53 struct xlog_ticket *ticket);
54#if defined(DEBUG)
55STATIC void
56xlog_verify_iclog(
57 struct xlog *log,
58 struct xlog_in_core *iclog,
59 int count);
60STATIC void
61xlog_verify_tail_lsn(
62 struct xlog *log,
63 struct xlog_in_core *iclog);
64#else
65#define xlog_verify_iclog(a,b,c)
66#define xlog_verify_tail_lsn(a,b)
67#endif
68
69STATIC int
70xlog_iclogs_empty(
71 struct xlog *log);
72
73static int
74xfs_log_cover(struct xfs_mount *);
75
76/*
77 * We need to make sure the buffer pointer returned is naturally aligned for the
78 * biggest basic data type we put into it. We have already accounted for this
79 * padding when sizing the buffer.
80 *
81 * However, this padding does not get written into the log, and hence we have to
82 * track the space used by the log vectors separately to prevent log space hangs
83 * due to inaccurate accounting (i.e. a leak) of the used log space through the
84 * CIL context ticket.
85 *
86 * We also add space for the xlog_op_header that describes this region in the
87 * log. This prepends the data region we return to the caller to copy their data
88 * into, so do all the static initialisation of the ophdr now. Because the ophdr
89 * is not 8 byte aligned, we have to be careful to ensure that we align the
90 * start of the buffer such that the region we return to the call is 8 byte
91 * aligned and packed against the tail of the ophdr.
92 */
93void *
94xlog_prepare_iovec(
95 struct xfs_log_vec *lv,
96 struct xfs_log_iovec **vecp,
97 uint type)
98{
99 struct xfs_log_iovec *vec = *vecp;
100 struct xlog_op_header *oph;
101 uint32_t len;
102 void *buf;
103
104 if (vec) {
105 ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
106 vec++;
107 } else {
108 vec = &lv->lv_iovecp[0];
109 }
110
111 len = lv->lv_buf_len + sizeof(struct xlog_op_header);
112 if (!IS_ALIGNED(len, sizeof(uint64_t))) {
113 lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
114 sizeof(struct xlog_op_header);
115 }
116
117 vec->i_type = type;
118 vec->i_addr = lv->lv_buf + lv->lv_buf_len;
119
120 oph = vec->i_addr;
121 oph->oh_clientid = XFS_TRANSACTION;
122 oph->oh_res2 = 0;
123 oph->oh_flags = 0;
124
125 buf = vec->i_addr + sizeof(struct xlog_op_header);
126 ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
127
128 *vecp = vec;
129 return buf;
130}
131
132static inline void
133xlog_grant_sub_space(
134 struct xlog_grant_head *head,
135 int64_t bytes)
136{
137 atomic64_sub(bytes, &head->grant);
138}
139
140static inline void
141xlog_grant_add_space(
142 struct xlog_grant_head *head,
143 int64_t bytes)
144{
145 atomic64_add(bytes, &head->grant);
146}
147
148static void
149xlog_grant_head_init(
150 struct xlog_grant_head *head)
151{
152 atomic64_set(&head->grant, 0);
153 INIT_LIST_HEAD(&head->waiters);
154 spin_lock_init(&head->lock);
155}
156
157void
158xlog_grant_return_space(
159 struct xlog *log,
160 xfs_lsn_t old_head,
161 xfs_lsn_t new_head)
162{
163 int64_t diff = xlog_lsn_sub(log, new_head, old_head);
164
165 xlog_grant_sub_space(&log->l_reserve_head, diff);
166 xlog_grant_sub_space(&log->l_write_head, diff);
167}
168
169/*
170 * Return the space in the log between the tail and the head. In the case where
171 * we have overrun available reservation space, return 0. The memory barrier
172 * pairs with the smp_wmb() in xlog_cil_ail_insert() to ensure that grant head
173 * vs tail space updates are seen in the correct order and hence avoid
174 * transients as space is transferred from the grant heads to the AIL on commit
175 * completion.
176 */
177static uint64_t
178xlog_grant_space_left(
179 struct xlog *log,
180 struct xlog_grant_head *head)
181{
182 int64_t free_bytes;
183
184 smp_rmb(); /* paired with smp_wmb in xlog_cil_ail_insert() */
185 free_bytes = log->l_logsize - READ_ONCE(log->l_tail_space) -
186 atomic64_read(&head->grant);
187 if (free_bytes > 0)
188 return free_bytes;
189 return 0;
190}
191
192STATIC void
193xlog_grant_head_wake_all(
194 struct xlog_grant_head *head)
195{
196 struct xlog_ticket *tic;
197
198 spin_lock(&head->lock);
199 list_for_each_entry(tic, &head->waiters, t_queue)
200 wake_up_process(tic->t_task);
201 spin_unlock(&head->lock);
202}
203
204static inline int
205xlog_ticket_reservation(
206 struct xlog *log,
207 struct xlog_grant_head *head,
208 struct xlog_ticket *tic)
209{
210 if (head == &log->l_write_head) {
211 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
212 return tic->t_unit_res;
213 }
214
215 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
216 return tic->t_unit_res * tic->t_cnt;
217
218 return tic->t_unit_res;
219}
220
221STATIC bool
222xlog_grant_head_wake(
223 struct xlog *log,
224 struct xlog_grant_head *head,
225 int *free_bytes)
226{
227 struct xlog_ticket *tic;
228 int need_bytes;
229
230 list_for_each_entry(tic, &head->waiters, t_queue) {
231 need_bytes = xlog_ticket_reservation(log, head, tic);
232 if (*free_bytes < need_bytes)
233 return false;
234
235 *free_bytes -= need_bytes;
236 trace_xfs_log_grant_wake_up(log, tic);
237 wake_up_process(tic->t_task);
238 }
239
240 return true;
241}
242
243STATIC int
244xlog_grant_head_wait(
245 struct xlog *log,
246 struct xlog_grant_head *head,
247 struct xlog_ticket *tic,
248 int need_bytes) __releases(&head->lock)
249 __acquires(&head->lock)
250{
251 list_add_tail(&tic->t_queue, &head->waiters);
252
253 do {
254 if (xlog_is_shutdown(log))
255 goto shutdown;
256
257 __set_current_state(TASK_UNINTERRUPTIBLE);
258 spin_unlock(&head->lock);
259
260 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
261
262 /* Push on the AIL to free up all the log space. */
263 xfs_ail_push_all(log->l_ailp);
264
265 trace_xfs_log_grant_sleep(log, tic);
266 schedule();
267 trace_xfs_log_grant_wake(log, tic);
268
269 spin_lock(&head->lock);
270 if (xlog_is_shutdown(log))
271 goto shutdown;
272 } while (xlog_grant_space_left(log, head) < need_bytes);
273
274 list_del_init(&tic->t_queue);
275 return 0;
276shutdown:
277 list_del_init(&tic->t_queue);
278 return -EIO;
279}
280
281/*
282 * Atomically get the log space required for a log ticket.
283 *
284 * Once a ticket gets put onto head->waiters, it will only return after the
285 * needed reservation is satisfied.
286 *
287 * This function is structured so that it has a lock free fast path. This is
288 * necessary because every new transaction reservation will come through this
289 * path. Hence any lock will be globally hot if we take it unconditionally on
290 * every pass.
291 *
292 * As tickets are only ever moved on and off head->waiters under head->lock, we
293 * only need to take that lock if we are going to add the ticket to the queue
294 * and sleep. We can avoid taking the lock if the ticket was never added to
295 * head->waiters because the t_queue list head will be empty and we hold the
296 * only reference to it so it can safely be checked unlocked.
297 */
298STATIC int
299xlog_grant_head_check(
300 struct xlog *log,
301 struct xlog_grant_head *head,
302 struct xlog_ticket *tic,
303 int *need_bytes)
304{
305 int free_bytes;
306 int error = 0;
307
308 ASSERT(!xlog_in_recovery(log));
309
310 /*
311 * If there are other waiters on the queue then give them a chance at
312 * logspace before us. Wake up the first waiters, if we do not wake
313 * up all the waiters then go to sleep waiting for more free space,
314 * otherwise try to get some space for this transaction.
315 */
316 *need_bytes = xlog_ticket_reservation(log, head, tic);
317 free_bytes = xlog_grant_space_left(log, head);
318 if (!list_empty_careful(&head->waiters)) {
319 spin_lock(&head->lock);
320 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
321 free_bytes < *need_bytes) {
322 error = xlog_grant_head_wait(log, head, tic,
323 *need_bytes);
324 }
325 spin_unlock(&head->lock);
326 } else if (free_bytes < *need_bytes) {
327 spin_lock(&head->lock);
328 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
329 spin_unlock(&head->lock);
330 }
331
332 return error;
333}
334
335bool
336xfs_log_writable(
337 struct xfs_mount *mp)
338{
339 /*
340 * Do not write to the log on norecovery mounts, if the data or log
341 * devices are read-only, or if the filesystem is shutdown. Read-only
342 * mounts allow internal writes for log recovery and unmount purposes,
343 * so don't restrict that case.
344 */
345 if (xfs_has_norecovery(mp))
346 return false;
347 if (xfs_readonly_buftarg(mp->m_ddev_targp))
348 return false;
349 if (xfs_readonly_buftarg(mp->m_log->l_targ))
350 return false;
351 if (xlog_is_shutdown(mp->m_log))
352 return false;
353 return true;
354}
355
356/*
357 * Replenish the byte reservation required by moving the grant write head.
358 */
359int
360xfs_log_regrant(
361 struct xfs_mount *mp,
362 struct xlog_ticket *tic)
363{
364 struct xlog *log = mp->m_log;
365 int need_bytes;
366 int error = 0;
367
368 if (xlog_is_shutdown(log))
369 return -EIO;
370
371 XFS_STATS_INC(mp, xs_try_logspace);
372
373 /*
374 * This is a new transaction on the ticket, so we need to change the
375 * transaction ID so that the next transaction has a different TID in
376 * the log. Just add one to the existing tid so that we can see chains
377 * of rolling transactions in the log easily.
378 */
379 tic->t_tid++;
380 tic->t_curr_res = tic->t_unit_res;
381 if (tic->t_cnt > 0)
382 return 0;
383
384 trace_xfs_log_regrant(log, tic);
385
386 error = xlog_grant_head_check(log, &log->l_write_head, tic,
387 &need_bytes);
388 if (error)
389 goto out_error;
390
391 xlog_grant_add_space(&log->l_write_head, need_bytes);
392 trace_xfs_log_regrant_exit(log, tic);
393 return 0;
394
395out_error:
396 /*
397 * If we are failing, make sure the ticket doesn't have any current
398 * reservations. We don't want to add this back when the ticket/
399 * transaction gets cancelled.
400 */
401 tic->t_curr_res = 0;
402 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
403 return error;
404}
405
406/*
407 * Reserve log space and return a ticket corresponding to the reservation.
408 *
409 * Each reservation is going to reserve extra space for a log record header.
410 * When writes happen to the on-disk log, we don't subtract the length of the
411 * log record header from any reservation. By wasting space in each
412 * reservation, we prevent over allocation problems.
413 */
414int
415xfs_log_reserve(
416 struct xfs_mount *mp,
417 int unit_bytes,
418 int cnt,
419 struct xlog_ticket **ticp,
420 bool permanent)
421{
422 struct xlog *log = mp->m_log;
423 struct xlog_ticket *tic;
424 int need_bytes;
425 int error = 0;
426
427 if (xlog_is_shutdown(log))
428 return -EIO;
429
430 XFS_STATS_INC(mp, xs_try_logspace);
431
432 ASSERT(*ticp == NULL);
433 tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
434 *ticp = tic;
435 trace_xfs_log_reserve(log, tic);
436 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
437 &need_bytes);
438 if (error)
439 goto out_error;
440
441 xlog_grant_add_space(&log->l_reserve_head, need_bytes);
442 xlog_grant_add_space(&log->l_write_head, need_bytes);
443 trace_xfs_log_reserve_exit(log, tic);
444 return 0;
445
446out_error:
447 /*
448 * If we are failing, make sure the ticket doesn't have any current
449 * reservations. We don't want to add this back when the ticket/
450 * transaction gets cancelled.
451 */
452 tic->t_curr_res = 0;
453 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
454 return error;
455}
456
457/*
458 * Run all the pending iclog callbacks and wake log force waiters and iclog
459 * space waiters so they can process the newly set shutdown state. We really
460 * don't care what order we process callbacks here because the log is shut down
461 * and so state cannot change on disk anymore. However, we cannot wake waiters
462 * until the callbacks have been processed because we may be in unmount and
463 * we must ensure that all AIL operations the callbacks perform have completed
464 * before we tear down the AIL.
465 *
466 * We avoid processing actively referenced iclogs so that we don't run callbacks
467 * while the iclog owner might still be preparing the iclog for IO submssion.
468 * These will be caught by xlog_state_iclog_release() and call this function
469 * again to process any callbacks that may have been added to that iclog.
470 */
471static void
472xlog_state_shutdown_callbacks(
473 struct xlog *log)
474{
475 struct xlog_in_core *iclog;
476 LIST_HEAD(cb_list);
477
478 iclog = log->l_iclog;
479 do {
480 if (atomic_read(&iclog->ic_refcnt)) {
481 /* Reference holder will re-run iclog callbacks. */
482 continue;
483 }
484 list_splice_init(&iclog->ic_callbacks, &cb_list);
485 spin_unlock(&log->l_icloglock);
486
487 xlog_cil_process_committed(&cb_list);
488
489 spin_lock(&log->l_icloglock);
490 wake_up_all(&iclog->ic_write_wait);
491 wake_up_all(&iclog->ic_force_wait);
492 } while ((iclog = iclog->ic_next) != log->l_iclog);
493
494 wake_up_all(&log->l_flush_wait);
495}
496
497/*
498 * Flush iclog to disk if this is the last reference to the given iclog and the
499 * it is in the WANT_SYNC state.
500 *
501 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
502 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
503 * written to stable storage, and implies that a commit record is contained
504 * within the iclog. We need to ensure that the log tail does not move beyond
505 * the tail that the first commit record in the iclog ordered against, otherwise
506 * correct recovery of that checkpoint becomes dependent on future operations
507 * performed on this iclog.
508 *
509 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
510 * current tail into iclog. Once the iclog tail is set, future operations must
511 * not modify it, otherwise they potentially violate ordering constraints for
512 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
513 * the iclog will get zeroed on activation of the iclog after sync, so we
514 * always capture the tail lsn on the iclog on the first NEED_FUA release
515 * regardless of the number of active reference counts on this iclog.
516 */
517int
518xlog_state_release_iclog(
519 struct xlog *log,
520 struct xlog_in_core *iclog,
521 struct xlog_ticket *ticket)
522{
523 bool last_ref;
524
525 lockdep_assert_held(&log->l_icloglock);
526
527 trace_xlog_iclog_release(iclog, _RET_IP_);
528 /*
529 * Grabbing the current log tail needs to be atomic w.r.t. the writing
530 * of the tail LSN into the iclog so we guarantee that the log tail does
531 * not move between the first time we know that the iclog needs to be
532 * made stable and when we eventually submit it.
533 */
534 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
535 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
536 !iclog->ic_header.h_tail_lsn) {
537 iclog->ic_header.h_tail_lsn =
538 cpu_to_be64(atomic64_read(&log->l_tail_lsn));
539 }
540
541 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
542
543 if (xlog_is_shutdown(log)) {
544 /*
545 * If there are no more references to this iclog, process the
546 * pending iclog callbacks that were waiting on the release of
547 * this iclog.
548 */
549 if (last_ref)
550 xlog_state_shutdown_callbacks(log);
551 return -EIO;
552 }
553
554 if (!last_ref)
555 return 0;
556
557 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
558 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
559 return 0;
560 }
561
562 iclog->ic_state = XLOG_STATE_SYNCING;
563 xlog_verify_tail_lsn(log, iclog);
564 trace_xlog_iclog_syncing(iclog, _RET_IP_);
565
566 spin_unlock(&log->l_icloglock);
567 xlog_sync(log, iclog, ticket);
568 spin_lock(&log->l_icloglock);
569 return 0;
570}
571
572/*
573 * Mount a log filesystem
574 *
575 * mp - ubiquitous xfs mount point structure
576 * log_target - buftarg of on-disk log device
577 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
578 * num_bblocks - Number of BBSIZE blocks in on-disk log
579 *
580 * Return error or zero.
581 */
582int
583xfs_log_mount(
584 xfs_mount_t *mp,
585 struct xfs_buftarg *log_target,
586 xfs_daddr_t blk_offset,
587 int num_bblks)
588{
589 struct xlog *log;
590 int error = 0;
591 int min_logfsbs;
592
593 if (!xfs_has_norecovery(mp)) {
594 xfs_notice(mp, "Mounting V%d Filesystem %pU",
595 XFS_SB_VERSION_NUM(&mp->m_sb),
596 &mp->m_sb.sb_uuid);
597 } else {
598 xfs_notice(mp,
599"Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
600 XFS_SB_VERSION_NUM(&mp->m_sb),
601 &mp->m_sb.sb_uuid);
602 ASSERT(xfs_is_readonly(mp));
603 }
604
605 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
606 if (IS_ERR(log)) {
607 error = PTR_ERR(log);
608 goto out;
609 }
610 mp->m_log = log;
611
612 /*
613 * Now that we have set up the log and it's internal geometry
614 * parameters, we can validate the given log space and drop a critical
615 * message via syslog if the log size is too small. A log that is too
616 * small can lead to unexpected situations in transaction log space
617 * reservation stage. The superblock verifier has already validated all
618 * the other log geometry constraints, so we don't have to check those
619 * here.
620 *
621 * Note: For v4 filesystems, we can't just reject the mount if the
622 * validation fails. This would mean that people would have to
623 * downgrade their kernel just to remedy the situation as there is no
624 * way to grow the log (short of black magic surgery with xfs_db).
625 *
626 * We can, however, reject mounts for V5 format filesystems, as the
627 * mkfs binary being used to make the filesystem should never create a
628 * filesystem with a log that is too small.
629 */
630 min_logfsbs = xfs_log_calc_minimum_size(mp);
631 if (mp->m_sb.sb_logblocks < min_logfsbs) {
632 xfs_warn(mp,
633 "Log size %d blocks too small, minimum size is %d blocks",
634 mp->m_sb.sb_logblocks, min_logfsbs);
635
636 /*
637 * Log check errors are always fatal on v5; or whenever bad
638 * metadata leads to a crash.
639 */
640 if (xfs_has_crc(mp)) {
641 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
642 ASSERT(0);
643 error = -EINVAL;
644 goto out_free_log;
645 }
646 xfs_crit(mp, "Log size out of supported range.");
647 xfs_crit(mp,
648"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
649 }
650
651 /*
652 * Initialize the AIL now we have a log.
653 */
654 error = xfs_trans_ail_init(mp);
655 if (error) {
656 xfs_warn(mp, "AIL initialisation failed: error %d", error);
657 goto out_free_log;
658 }
659 log->l_ailp = mp->m_ail;
660
661 /*
662 * skip log recovery on a norecovery mount. pretend it all
663 * just worked.
664 */
665 if (!xfs_has_norecovery(mp)) {
666 error = xlog_recover(log);
667 if (error) {
668 xfs_warn(mp, "log mount/recovery failed: error %d",
669 error);
670 xlog_recover_cancel(log);
671 goto out_destroy_ail;
672 }
673 }
674
675 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
676 "log");
677 if (error)
678 goto out_destroy_ail;
679
680 /* Normal transactions can now occur */
681 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
682
683 /*
684 * Now the log has been fully initialised and we know were our
685 * space grant counters are, we can initialise the permanent ticket
686 * needed for delayed logging to work.
687 */
688 xlog_cil_init_post_recovery(log);
689
690 return 0;
691
692out_destroy_ail:
693 xfs_trans_ail_destroy(mp);
694out_free_log:
695 xlog_dealloc_log(log);
696out:
697 return error;
698}
699
700/*
701 * Finish the recovery of the file system. This is separate from the
702 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
703 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
704 * here.
705 *
706 * If we finish recovery successfully, start the background log work. If we are
707 * not doing recovery, then we have a RO filesystem and we don't need to start
708 * it.
709 */
710int
711xfs_log_mount_finish(
712 struct xfs_mount *mp)
713{
714 struct xlog *log = mp->m_log;
715 int error = 0;
716
717 if (xfs_has_norecovery(mp)) {
718 ASSERT(xfs_is_readonly(mp));
719 return 0;
720 }
721
722 /*
723 * During the second phase of log recovery, we need iget and
724 * iput to behave like they do for an active filesystem.
725 * xfs_fs_drop_inode needs to be able to prevent the deletion
726 * of inodes before we're done replaying log items on those
727 * inodes. Turn it off immediately after recovery finishes
728 * so that we don't leak the quota inodes if subsequent mount
729 * activities fail.
730 *
731 * We let all inodes involved in redo item processing end up on
732 * the LRU instead of being evicted immediately so that if we do
733 * something to an unlinked inode, the irele won't cause
734 * premature truncation and freeing of the inode, which results
735 * in log recovery failure. We have to evict the unreferenced
736 * lru inodes after clearing SB_ACTIVE because we don't
737 * otherwise clean up the lru if there's a subsequent failure in
738 * xfs_mountfs, which leads to us leaking the inodes if nothing
739 * else (e.g. quotacheck) references the inodes before the
740 * mount failure occurs.
741 */
742 mp->m_super->s_flags |= SB_ACTIVE;
743 xfs_log_work_queue(mp);
744 if (xlog_recovery_needed(log))
745 error = xlog_recover_finish(log);
746 mp->m_super->s_flags &= ~SB_ACTIVE;
747 evict_inodes(mp->m_super);
748
749 /*
750 * Drain the buffer LRU after log recovery. This is required for v4
751 * filesystems to avoid leaving around buffers with NULL verifier ops,
752 * but we do it unconditionally to make sure we're always in a clean
753 * cache state after mount.
754 *
755 * Don't push in the error case because the AIL may have pending intents
756 * that aren't removed until recovery is cancelled.
757 */
758 if (xlog_recovery_needed(log)) {
759 if (!error) {
760 xfs_log_force(mp, XFS_LOG_SYNC);
761 xfs_ail_push_all_sync(mp->m_ail);
762 }
763 xfs_notice(mp, "Ending recovery (logdev: %s)",
764 mp->m_logname ? mp->m_logname : "internal");
765 } else {
766 xfs_info(mp, "Ending clean mount");
767 }
768 xfs_buftarg_drain(mp->m_ddev_targp);
769
770 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
771
772 /* Make sure the log is dead if we're returning failure. */
773 ASSERT(!error || xlog_is_shutdown(log));
774
775 return error;
776}
777
778/*
779 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
780 * the log.
781 */
782void
783xfs_log_mount_cancel(
784 struct xfs_mount *mp)
785{
786 xlog_recover_cancel(mp->m_log);
787 xfs_log_unmount(mp);
788}
789
790/*
791 * Flush out the iclog to disk ensuring that device caches are flushed and
792 * the iclog hits stable storage before any completion waiters are woken.
793 */
794static inline int
795xlog_force_iclog(
796 struct xlog_in_core *iclog)
797{
798 atomic_inc(&iclog->ic_refcnt);
799 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
800 if (iclog->ic_state == XLOG_STATE_ACTIVE)
801 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
802 return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
803}
804
805/*
806 * Cycle all the iclogbuf locks to make sure all log IO completion
807 * is done before we tear down these buffers.
808 */
809static void
810xlog_wait_iclog_completion(struct xlog *log)
811{
812 int i;
813 struct xlog_in_core *iclog = log->l_iclog;
814
815 for (i = 0; i < log->l_iclog_bufs; i++) {
816 down(&iclog->ic_sema);
817 up(&iclog->ic_sema);
818 iclog = iclog->ic_next;
819 }
820}
821
822/*
823 * Wait for the iclog and all prior iclogs to be written disk as required by the
824 * log force state machine. Waiting on ic_force_wait ensures iclog completions
825 * have been ordered and callbacks run before we are woken here, hence
826 * guaranteeing that all the iclogs up to this one are on stable storage.
827 */
828int
829xlog_wait_on_iclog(
830 struct xlog_in_core *iclog)
831 __releases(iclog->ic_log->l_icloglock)
832{
833 struct xlog *log = iclog->ic_log;
834
835 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
836 if (!xlog_is_shutdown(log) &&
837 iclog->ic_state != XLOG_STATE_ACTIVE &&
838 iclog->ic_state != XLOG_STATE_DIRTY) {
839 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
840 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
841 } else {
842 spin_unlock(&log->l_icloglock);
843 }
844
845 if (xlog_is_shutdown(log))
846 return -EIO;
847 return 0;
848}
849
850/*
851 * Write out an unmount record using the ticket provided. We have to account for
852 * the data space used in the unmount ticket as this write is not done from a
853 * transaction context that has already done the accounting for us.
854 */
855static int
856xlog_write_unmount_record(
857 struct xlog *log,
858 struct xlog_ticket *ticket)
859{
860 struct {
861 struct xlog_op_header ophdr;
862 struct xfs_unmount_log_format ulf;
863 } unmount_rec = {
864 .ophdr = {
865 .oh_clientid = XFS_LOG,
866 .oh_tid = cpu_to_be32(ticket->t_tid),
867 .oh_flags = XLOG_UNMOUNT_TRANS,
868 },
869 .ulf = {
870 .magic = XLOG_UNMOUNT_TYPE,
871 },
872 };
873 struct xfs_log_iovec reg = {
874 .i_addr = &unmount_rec,
875 .i_len = sizeof(unmount_rec),
876 .i_type = XLOG_REG_TYPE_UNMOUNT,
877 };
878 struct xfs_log_vec vec = {
879 .lv_niovecs = 1,
880 .lv_iovecp = ®,
881 };
882 LIST_HEAD(lv_chain);
883 list_add(&vec.lv_list, &lv_chain);
884
885 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
886 sizeof(struct xfs_unmount_log_format)) !=
887 sizeof(unmount_rec));
888
889 /* account for space used by record data */
890 ticket->t_curr_res -= sizeof(unmount_rec);
891
892 return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
893}
894
895/*
896 * Mark the filesystem clean by writing an unmount record to the head of the
897 * log.
898 */
899static void
900xlog_unmount_write(
901 struct xlog *log)
902{
903 struct xfs_mount *mp = log->l_mp;
904 struct xlog_in_core *iclog;
905 struct xlog_ticket *tic = NULL;
906 int error;
907
908 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
909 if (error)
910 goto out_err;
911
912 error = xlog_write_unmount_record(log, tic);
913 /*
914 * At this point, we're umounting anyway, so there's no point in
915 * transitioning log state to shutdown. Just continue...
916 */
917out_err:
918 if (error)
919 xfs_alert(mp, "%s: unmount record failed", __func__);
920
921 spin_lock(&log->l_icloglock);
922 iclog = log->l_iclog;
923 error = xlog_force_iclog(iclog);
924 xlog_wait_on_iclog(iclog);
925
926 if (tic) {
927 trace_xfs_log_umount_write(log, tic);
928 xfs_log_ticket_ungrant(log, tic);
929 }
930}
931
932static void
933xfs_log_unmount_verify_iclog(
934 struct xlog *log)
935{
936 struct xlog_in_core *iclog = log->l_iclog;
937
938 do {
939 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
940 ASSERT(iclog->ic_offset == 0);
941 } while ((iclog = iclog->ic_next) != log->l_iclog);
942}
943
944/*
945 * Unmount record used to have a string "Unmount filesystem--" in the
946 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
947 * We just write the magic number now since that particular field isn't
948 * currently architecture converted and "Unmount" is a bit foo.
949 * As far as I know, there weren't any dependencies on the old behaviour.
950 */
951static void
952xfs_log_unmount_write(
953 struct xfs_mount *mp)
954{
955 struct xlog *log = mp->m_log;
956
957 if (!xfs_log_writable(mp))
958 return;
959
960 xfs_log_force(mp, XFS_LOG_SYNC);
961
962 if (xlog_is_shutdown(log))
963 return;
964
965 /*
966 * If we think the summary counters are bad, avoid writing the unmount
967 * record to force log recovery at next mount, after which the summary
968 * counters will be recalculated. Refer to xlog_check_unmount_rec for
969 * more details.
970 */
971 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
972 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
973 xfs_alert(mp, "%s: will fix summary counters at next mount",
974 __func__);
975 return;
976 }
977
978 xfs_log_unmount_verify_iclog(log);
979 xlog_unmount_write(log);
980}
981
982/*
983 * Empty the log for unmount/freeze.
984 *
985 * To do this, we first need to shut down the background log work so it is not
986 * trying to cover the log as we clean up. We then need to unpin all objects in
987 * the log so we can then flush them out. Once they have completed their IO and
988 * run the callbacks removing themselves from the AIL, we can cover the log.
989 */
990int
991xfs_log_quiesce(
992 struct xfs_mount *mp)
993{
994 /*
995 * Clear log incompat features since we're quiescing the log. Report
996 * failures, though it's not fatal to have a higher log feature
997 * protection level than the log contents actually require.
998 */
999 if (xfs_clear_incompat_log_features(mp)) {
1000 int error;
1001
1002 error = xfs_sync_sb(mp, false);
1003 if (error)
1004 xfs_warn(mp,
1005 "Failed to clear log incompat features on quiesce");
1006 }
1007
1008 cancel_delayed_work_sync(&mp->m_log->l_work);
1009 xfs_log_force(mp, XFS_LOG_SYNC);
1010
1011 /*
1012 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1013 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1014 * xfs_buf_iowait() cannot be used because it was pushed with the
1015 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1016 * the IO to complete.
1017 */
1018 xfs_ail_push_all_sync(mp->m_ail);
1019 xfs_buftarg_wait(mp->m_ddev_targp);
1020 xfs_buf_lock(mp->m_sb_bp);
1021 xfs_buf_unlock(mp->m_sb_bp);
1022
1023 return xfs_log_cover(mp);
1024}
1025
1026void
1027xfs_log_clean(
1028 struct xfs_mount *mp)
1029{
1030 xfs_log_quiesce(mp);
1031 xfs_log_unmount_write(mp);
1032}
1033
1034/*
1035 * Shut down and release the AIL and Log.
1036 *
1037 * During unmount, we need to ensure we flush all the dirty metadata objects
1038 * from the AIL so that the log is empty before we write the unmount record to
1039 * the log. Once this is done, we can tear down the AIL and the log.
1040 */
1041void
1042xfs_log_unmount(
1043 struct xfs_mount *mp)
1044{
1045 xfs_log_clean(mp);
1046
1047 /*
1048 * If shutdown has come from iclog IO context, the log
1049 * cleaning will have been skipped and so we need to wait
1050 * for the iclog to complete shutdown processing before we
1051 * tear anything down.
1052 */
1053 xlog_wait_iclog_completion(mp->m_log);
1054
1055 xfs_buftarg_drain(mp->m_ddev_targp);
1056
1057 xfs_trans_ail_destroy(mp);
1058
1059 xfs_sysfs_del(&mp->m_log->l_kobj);
1060
1061 xlog_dealloc_log(mp->m_log);
1062}
1063
1064void
1065xfs_log_item_init(
1066 struct xfs_mount *mp,
1067 struct xfs_log_item *item,
1068 int type,
1069 const struct xfs_item_ops *ops)
1070{
1071 item->li_log = mp->m_log;
1072 item->li_ailp = mp->m_ail;
1073 item->li_type = type;
1074 item->li_ops = ops;
1075 item->li_lv = NULL;
1076
1077 INIT_LIST_HEAD(&item->li_ail);
1078 INIT_LIST_HEAD(&item->li_cil);
1079 INIT_LIST_HEAD(&item->li_bio_list);
1080 INIT_LIST_HEAD(&item->li_trans);
1081}
1082
1083/*
1084 * Wake up processes waiting for log space after we have moved the log tail.
1085 */
1086void
1087xfs_log_space_wake(
1088 struct xfs_mount *mp)
1089{
1090 struct xlog *log = mp->m_log;
1091 int free_bytes;
1092
1093 if (xlog_is_shutdown(log))
1094 return;
1095
1096 if (!list_empty_careful(&log->l_write_head.waiters)) {
1097 ASSERT(!xlog_in_recovery(log));
1098
1099 spin_lock(&log->l_write_head.lock);
1100 free_bytes = xlog_grant_space_left(log, &log->l_write_head);
1101 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1102 spin_unlock(&log->l_write_head.lock);
1103 }
1104
1105 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1106 ASSERT(!xlog_in_recovery(log));
1107
1108 spin_lock(&log->l_reserve_head.lock);
1109 free_bytes = xlog_grant_space_left(log, &log->l_reserve_head);
1110 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1111 spin_unlock(&log->l_reserve_head.lock);
1112 }
1113}
1114
1115/*
1116 * Determine if we have a transaction that has gone to disk that needs to be
1117 * covered. To begin the transition to the idle state firstly the log needs to
1118 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1119 * we start attempting to cover the log.
1120 *
1121 * Only if we are then in a state where covering is needed, the caller is
1122 * informed that dummy transactions are required to move the log into the idle
1123 * state.
1124 *
1125 * If there are any items in the AIl or CIL, then we do not want to attempt to
1126 * cover the log as we may be in a situation where there isn't log space
1127 * available to run a dummy transaction and this can lead to deadlocks when the
1128 * tail of the log is pinned by an item that is modified in the CIL. Hence
1129 * there's no point in running a dummy transaction at this point because we
1130 * can't start trying to idle the log until both the CIL and AIL are empty.
1131 */
1132static bool
1133xfs_log_need_covered(
1134 struct xfs_mount *mp)
1135{
1136 struct xlog *log = mp->m_log;
1137 bool needed = false;
1138
1139 if (!xlog_cil_empty(log))
1140 return false;
1141
1142 spin_lock(&log->l_icloglock);
1143 switch (log->l_covered_state) {
1144 case XLOG_STATE_COVER_DONE:
1145 case XLOG_STATE_COVER_DONE2:
1146 case XLOG_STATE_COVER_IDLE:
1147 break;
1148 case XLOG_STATE_COVER_NEED:
1149 case XLOG_STATE_COVER_NEED2:
1150 if (xfs_ail_min_lsn(log->l_ailp))
1151 break;
1152 if (!xlog_iclogs_empty(log))
1153 break;
1154
1155 needed = true;
1156 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1157 log->l_covered_state = XLOG_STATE_COVER_DONE;
1158 else
1159 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1160 break;
1161 default:
1162 needed = true;
1163 break;
1164 }
1165 spin_unlock(&log->l_icloglock);
1166 return needed;
1167}
1168
1169/*
1170 * Explicitly cover the log. This is similar to background log covering but
1171 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1172 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1173 * must all be empty.
1174 */
1175static int
1176xfs_log_cover(
1177 struct xfs_mount *mp)
1178{
1179 int error = 0;
1180 bool need_covered;
1181
1182 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1183 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1184 xlog_is_shutdown(mp->m_log));
1185
1186 if (!xfs_log_writable(mp))
1187 return 0;
1188
1189 /*
1190 * xfs_log_need_covered() is not idempotent because it progresses the
1191 * state machine if the log requires covering. Therefore, we must call
1192 * this function once and use the result until we've issued an sb sync.
1193 * Do so first to make that abundantly clear.
1194 *
1195 * Fall into the covering sequence if the log needs covering or the
1196 * mount has lazy superblock accounting to sync to disk. The sb sync
1197 * used for covering accumulates the in-core counters, so covering
1198 * handles this for us.
1199 */
1200 need_covered = xfs_log_need_covered(mp);
1201 if (!need_covered && !xfs_has_lazysbcount(mp))
1202 return 0;
1203
1204 /*
1205 * To cover the log, commit the superblock twice (at most) in
1206 * independent checkpoints. The first serves as a reference for the
1207 * tail pointer. The sync transaction and AIL push empties the AIL and
1208 * updates the in-core tail to the LSN of the first checkpoint. The
1209 * second commit updates the on-disk tail with the in-core LSN,
1210 * covering the log. Push the AIL one more time to leave it empty, as
1211 * we found it.
1212 */
1213 do {
1214 error = xfs_sync_sb(mp, true);
1215 if (error)
1216 break;
1217 xfs_ail_push_all_sync(mp->m_ail);
1218 } while (xfs_log_need_covered(mp));
1219
1220 return error;
1221}
1222
1223static void
1224xlog_ioend_work(
1225 struct work_struct *work)
1226{
1227 struct xlog_in_core *iclog =
1228 container_of(work, struct xlog_in_core, ic_end_io_work);
1229 struct xlog *log = iclog->ic_log;
1230 int error;
1231
1232 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1233#ifdef DEBUG
1234 /* treat writes with injected CRC errors as failed */
1235 if (iclog->ic_fail_crc)
1236 error = -EIO;
1237#endif
1238
1239 /*
1240 * Race to shutdown the filesystem if we see an error.
1241 */
1242 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1243 xfs_alert(log->l_mp, "log I/O error %d", error);
1244 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1245 }
1246
1247 xlog_state_done_syncing(iclog);
1248 bio_uninit(&iclog->ic_bio);
1249
1250 /*
1251 * Drop the lock to signal that we are done. Nothing references the
1252 * iclog after this, so an unmount waiting on this lock can now tear it
1253 * down safely. As such, it is unsafe to reference the iclog after the
1254 * unlock as we could race with it being freed.
1255 */
1256 up(&iclog->ic_sema);
1257}
1258
1259/*
1260 * Return size of each in-core log record buffer.
1261 *
1262 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1263 *
1264 * If the filesystem blocksize is too large, we may need to choose a
1265 * larger size since the directory code currently logs entire blocks.
1266 */
1267STATIC void
1268xlog_get_iclog_buffer_size(
1269 struct xfs_mount *mp,
1270 struct xlog *log)
1271{
1272 if (mp->m_logbufs <= 0)
1273 mp->m_logbufs = XLOG_MAX_ICLOGS;
1274 if (mp->m_logbsize <= 0)
1275 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1276
1277 log->l_iclog_bufs = mp->m_logbufs;
1278 log->l_iclog_size = mp->m_logbsize;
1279
1280 /*
1281 * # headers = size / 32k - one header holds cycles from 32k of data.
1282 */
1283 log->l_iclog_heads =
1284 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1285 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1286}
1287
1288void
1289xfs_log_work_queue(
1290 struct xfs_mount *mp)
1291{
1292 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1293 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1294}
1295
1296/*
1297 * Clear the log incompat flags if we have the opportunity.
1298 *
1299 * This only happens if we're about to log the second dummy transaction as part
1300 * of covering the log.
1301 */
1302static inline void
1303xlog_clear_incompat(
1304 struct xlog *log)
1305{
1306 struct xfs_mount *mp = log->l_mp;
1307
1308 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1309 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1310 return;
1311
1312 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1313 return;
1314
1315 xfs_clear_incompat_log_features(mp);
1316}
1317
1318/*
1319 * Every sync period we need to unpin all items in the AIL and push them to
1320 * disk. If there is nothing dirty, then we might need to cover the log to
1321 * indicate that the filesystem is idle.
1322 */
1323static void
1324xfs_log_worker(
1325 struct work_struct *work)
1326{
1327 struct xlog *log = container_of(to_delayed_work(work),
1328 struct xlog, l_work);
1329 struct xfs_mount *mp = log->l_mp;
1330
1331 /* dgc: errors ignored - not fatal and nowhere to report them */
1332 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1333 /*
1334 * Dump a transaction into the log that contains no real change.
1335 * This is needed to stamp the current tail LSN into the log
1336 * during the covering operation.
1337 *
1338 * We cannot use an inode here for this - that will push dirty
1339 * state back up into the VFS and then periodic inode flushing
1340 * will prevent log covering from making progress. Hence we
1341 * synchronously log the superblock instead to ensure the
1342 * superblock is immediately unpinned and can be written back.
1343 */
1344 xlog_clear_incompat(log);
1345 xfs_sync_sb(mp, true);
1346 } else
1347 xfs_log_force(mp, 0);
1348
1349 /* start pushing all the metadata that is currently dirty */
1350 xfs_ail_push_all(mp->m_ail);
1351
1352 /* queue us up again */
1353 xfs_log_work_queue(mp);
1354}
1355
1356/*
1357 * This routine initializes some of the log structure for a given mount point.
1358 * Its primary purpose is to fill in enough, so recovery can occur. However,
1359 * some other stuff may be filled in too.
1360 */
1361STATIC struct xlog *
1362xlog_alloc_log(
1363 struct xfs_mount *mp,
1364 struct xfs_buftarg *log_target,
1365 xfs_daddr_t blk_offset,
1366 int num_bblks)
1367{
1368 struct xlog *log;
1369 xlog_rec_header_t *head;
1370 xlog_in_core_t **iclogp;
1371 xlog_in_core_t *iclog, *prev_iclog=NULL;
1372 int i;
1373 int error = -ENOMEM;
1374 uint log2_size = 0;
1375
1376 log = kzalloc(sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1377 if (!log) {
1378 xfs_warn(mp, "Log allocation failed: No memory!");
1379 goto out;
1380 }
1381
1382 log->l_mp = mp;
1383 log->l_targ = log_target;
1384 log->l_logsize = BBTOB(num_bblks);
1385 log->l_logBBstart = blk_offset;
1386 log->l_logBBsize = num_bblks;
1387 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1388 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1389 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1390 INIT_LIST_HEAD(&log->r_dfops);
1391
1392 log->l_prev_block = -1;
1393 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1394 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1395 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1396
1397 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1398 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1399 else
1400 log->l_iclog_roundoff = BBSIZE;
1401
1402 xlog_grant_head_init(&log->l_reserve_head);
1403 xlog_grant_head_init(&log->l_write_head);
1404
1405 error = -EFSCORRUPTED;
1406 if (xfs_has_sector(mp)) {
1407 log2_size = mp->m_sb.sb_logsectlog;
1408 if (log2_size < BBSHIFT) {
1409 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1410 log2_size, BBSHIFT);
1411 goto out_free_log;
1412 }
1413
1414 log2_size -= BBSHIFT;
1415 if (log2_size > mp->m_sectbb_log) {
1416 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1417 log2_size, mp->m_sectbb_log);
1418 goto out_free_log;
1419 }
1420
1421 /* for larger sector sizes, must have v2 or external log */
1422 if (log2_size && log->l_logBBstart > 0 &&
1423 !xfs_has_logv2(mp)) {
1424 xfs_warn(mp,
1425 "log sector size (0x%x) invalid for configuration.",
1426 log2_size);
1427 goto out_free_log;
1428 }
1429 }
1430 log->l_sectBBsize = 1 << log2_size;
1431
1432 xlog_get_iclog_buffer_size(mp, log);
1433
1434 spin_lock_init(&log->l_icloglock);
1435 init_waitqueue_head(&log->l_flush_wait);
1436
1437 iclogp = &log->l_iclog;
1438 /*
1439 * The amount of memory to allocate for the iclog structure is
1440 * rather funky due to the way the structure is defined. It is
1441 * done this way so that we can use different sizes for machines
1442 * with different amounts of memory. See the definition of
1443 * xlog_in_core_t in xfs_log_priv.h for details.
1444 */
1445 ASSERT(log->l_iclog_size >= 4096);
1446 for (i = 0; i < log->l_iclog_bufs; i++) {
1447 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1448 sizeof(struct bio_vec);
1449
1450 iclog = kzalloc(sizeof(*iclog) + bvec_size,
1451 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1452 if (!iclog)
1453 goto out_free_iclog;
1454
1455 *iclogp = iclog;
1456 iclog->ic_prev = prev_iclog;
1457 prev_iclog = iclog;
1458
1459 iclog->ic_data = kvzalloc(log->l_iclog_size,
1460 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1461 if (!iclog->ic_data)
1462 goto out_free_iclog;
1463 head = &iclog->ic_header;
1464 memset(head, 0, sizeof(xlog_rec_header_t));
1465 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1466 head->h_version = cpu_to_be32(
1467 xfs_has_logv2(log->l_mp) ? 2 : 1);
1468 head->h_size = cpu_to_be32(log->l_iclog_size);
1469 /* new fields */
1470 head->h_fmt = cpu_to_be32(XLOG_FMT);
1471 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1472
1473 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1474 iclog->ic_state = XLOG_STATE_ACTIVE;
1475 iclog->ic_log = log;
1476 atomic_set(&iclog->ic_refcnt, 0);
1477 INIT_LIST_HEAD(&iclog->ic_callbacks);
1478 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1479
1480 init_waitqueue_head(&iclog->ic_force_wait);
1481 init_waitqueue_head(&iclog->ic_write_wait);
1482 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1483 sema_init(&iclog->ic_sema, 1);
1484
1485 iclogp = &iclog->ic_next;
1486 }
1487 *iclogp = log->l_iclog; /* complete ring */
1488 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1489
1490 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1491 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1492 WQ_HIGHPRI),
1493 0, mp->m_super->s_id);
1494 if (!log->l_ioend_workqueue)
1495 goto out_free_iclog;
1496
1497 error = xlog_cil_init(log);
1498 if (error)
1499 goto out_destroy_workqueue;
1500 return log;
1501
1502out_destroy_workqueue:
1503 destroy_workqueue(log->l_ioend_workqueue);
1504out_free_iclog:
1505 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1506 prev_iclog = iclog->ic_next;
1507 kvfree(iclog->ic_data);
1508 kfree(iclog);
1509 if (prev_iclog == log->l_iclog)
1510 break;
1511 }
1512out_free_log:
1513 kfree(log);
1514out:
1515 return ERR_PTR(error);
1516} /* xlog_alloc_log */
1517
1518/*
1519 * Stamp cycle number in every block
1520 */
1521STATIC void
1522xlog_pack_data(
1523 struct xlog *log,
1524 struct xlog_in_core *iclog,
1525 int roundoff)
1526{
1527 int i, j, k;
1528 int size = iclog->ic_offset + roundoff;
1529 __be32 cycle_lsn;
1530 char *dp;
1531
1532 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1533
1534 dp = iclog->ic_datap;
1535 for (i = 0; i < BTOBB(size); i++) {
1536 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1537 break;
1538 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1539 *(__be32 *)dp = cycle_lsn;
1540 dp += BBSIZE;
1541 }
1542
1543 if (xfs_has_logv2(log->l_mp)) {
1544 xlog_in_core_2_t *xhdr = iclog->ic_data;
1545
1546 for ( ; i < BTOBB(size); i++) {
1547 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1548 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1549 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1550 *(__be32 *)dp = cycle_lsn;
1551 dp += BBSIZE;
1552 }
1553
1554 for (i = 1; i < log->l_iclog_heads; i++)
1555 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1556 }
1557}
1558
1559/*
1560 * Calculate the checksum for a log buffer.
1561 *
1562 * This is a little more complicated than it should be because the various
1563 * headers and the actual data are non-contiguous.
1564 */
1565__le32
1566xlog_cksum(
1567 struct xlog *log,
1568 struct xlog_rec_header *rhead,
1569 char *dp,
1570 int size)
1571{
1572 uint32_t crc;
1573
1574 /* first generate the crc for the record header ... */
1575 crc = xfs_start_cksum_update((char *)rhead,
1576 sizeof(struct xlog_rec_header),
1577 offsetof(struct xlog_rec_header, h_crc));
1578
1579 /* ... then for additional cycle data for v2 logs ... */
1580 if (xfs_has_logv2(log->l_mp)) {
1581 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1582 int i;
1583 int xheads;
1584
1585 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1586
1587 for (i = 1; i < xheads; i++) {
1588 crc = crc32c(crc, &xhdr[i].hic_xheader,
1589 sizeof(struct xlog_rec_ext_header));
1590 }
1591 }
1592
1593 /* ... and finally for the payload */
1594 crc = crc32c(crc, dp, size);
1595
1596 return xfs_end_cksum(crc);
1597}
1598
1599static void
1600xlog_bio_end_io(
1601 struct bio *bio)
1602{
1603 struct xlog_in_core *iclog = bio->bi_private;
1604
1605 queue_work(iclog->ic_log->l_ioend_workqueue,
1606 &iclog->ic_end_io_work);
1607}
1608
1609static int
1610xlog_map_iclog_data(
1611 struct bio *bio,
1612 void *data,
1613 size_t count)
1614{
1615 do {
1616 struct page *page = kmem_to_page(data);
1617 unsigned int off = offset_in_page(data);
1618 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1619
1620 if (bio_add_page(bio, page, len, off) != len)
1621 return -EIO;
1622
1623 data += len;
1624 count -= len;
1625 } while (count);
1626
1627 return 0;
1628}
1629
1630STATIC void
1631xlog_write_iclog(
1632 struct xlog *log,
1633 struct xlog_in_core *iclog,
1634 uint64_t bno,
1635 unsigned int count)
1636{
1637 ASSERT(bno < log->l_logBBsize);
1638 trace_xlog_iclog_write(iclog, _RET_IP_);
1639
1640 /*
1641 * We lock the iclogbufs here so that we can serialise against I/O
1642 * completion during unmount. We might be processing a shutdown
1643 * triggered during unmount, and that can occur asynchronously to the
1644 * unmount thread, and hence we need to ensure that completes before
1645 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1646 * across the log IO to archieve that.
1647 */
1648 down(&iclog->ic_sema);
1649 if (xlog_is_shutdown(log)) {
1650 /*
1651 * It would seem logical to return EIO here, but we rely on
1652 * the log state machine to propagate I/O errors instead of
1653 * doing it here. We kick of the state machine and unlock
1654 * the buffer manually, the code needs to be kept in sync
1655 * with the I/O completion path.
1656 */
1657 goto sync;
1658 }
1659
1660 /*
1661 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1662 * IOs coming immediately after this one. This prevents the block layer
1663 * writeback throttle from throttling log writes behind background
1664 * metadata writeback and causing priority inversions.
1665 */
1666 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1667 howmany(count, PAGE_SIZE),
1668 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1669 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1670 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1671 iclog->ic_bio.bi_private = iclog;
1672
1673 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1674 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1675 /*
1676 * For external log devices, we also need to flush the data
1677 * device cache first to ensure all metadata writeback covered
1678 * by the LSN in this iclog is on stable storage. This is slow,
1679 * but it *must* complete before we issue the external log IO.
1680 *
1681 * If the flush fails, we cannot conclude that past metadata
1682 * writeback from the log succeeded. Repeating the flush is
1683 * not possible, hence we must shut down with log IO error to
1684 * avoid shutdown re-entering this path and erroring out again.
1685 */
1686 if (log->l_targ != log->l_mp->m_ddev_targp &&
1687 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev))
1688 goto shutdown;
1689 }
1690 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1691 iclog->ic_bio.bi_opf |= REQ_FUA;
1692
1693 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1694
1695 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count))
1696 goto shutdown;
1697
1698 if (is_vmalloc_addr(iclog->ic_data))
1699 flush_kernel_vmap_range(iclog->ic_data, count);
1700
1701 /*
1702 * If this log buffer would straddle the end of the log we will have
1703 * to split it up into two bios, so that we can continue at the start.
1704 */
1705 if (bno + BTOBB(count) > log->l_logBBsize) {
1706 struct bio *split;
1707
1708 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1709 GFP_NOIO, &fs_bio_set);
1710 bio_chain(split, &iclog->ic_bio);
1711 submit_bio(split);
1712
1713 /* restart at logical offset zero for the remainder */
1714 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1715 }
1716
1717 submit_bio(&iclog->ic_bio);
1718 return;
1719shutdown:
1720 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1721sync:
1722 xlog_state_done_syncing(iclog);
1723 up(&iclog->ic_sema);
1724}
1725
1726/*
1727 * We need to bump cycle number for the part of the iclog that is
1728 * written to the start of the log. Watch out for the header magic
1729 * number case, though.
1730 */
1731static void
1732xlog_split_iclog(
1733 struct xlog *log,
1734 void *data,
1735 uint64_t bno,
1736 unsigned int count)
1737{
1738 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1739 unsigned int i;
1740
1741 for (i = split_offset; i < count; i += BBSIZE) {
1742 uint32_t cycle = get_unaligned_be32(data + i);
1743
1744 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1745 cycle++;
1746 put_unaligned_be32(cycle, data + i);
1747 }
1748}
1749
1750static int
1751xlog_calc_iclog_size(
1752 struct xlog *log,
1753 struct xlog_in_core *iclog,
1754 uint32_t *roundoff)
1755{
1756 uint32_t count_init, count;
1757
1758 /* Add for LR header */
1759 count_init = log->l_iclog_hsize + iclog->ic_offset;
1760 count = roundup(count_init, log->l_iclog_roundoff);
1761
1762 *roundoff = count - count_init;
1763
1764 ASSERT(count >= count_init);
1765 ASSERT(*roundoff < log->l_iclog_roundoff);
1766 return count;
1767}
1768
1769/*
1770 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1771 * fashion. Previously, we should have moved the current iclog
1772 * ptr in the log to point to the next available iclog. This allows further
1773 * write to continue while this code syncs out an iclog ready to go.
1774 * Before an in-core log can be written out, the data section must be scanned
1775 * to save away the 1st word of each BBSIZE block into the header. We replace
1776 * it with the current cycle count. Each BBSIZE block is tagged with the
1777 * cycle count because there in an implicit assumption that drives will
1778 * guarantee that entire 512 byte blocks get written at once. In other words,
1779 * we can't have part of a 512 byte block written and part not written. By
1780 * tagging each block, we will know which blocks are valid when recovering
1781 * after an unclean shutdown.
1782 *
1783 * This routine is single threaded on the iclog. No other thread can be in
1784 * this routine with the same iclog. Changing contents of iclog can there-
1785 * fore be done without grabbing the state machine lock. Updating the global
1786 * log will require grabbing the lock though.
1787 *
1788 * The entire log manager uses a logical block numbering scheme. Only
1789 * xlog_write_iclog knows about the fact that the log may not start with
1790 * block zero on a given device.
1791 */
1792STATIC void
1793xlog_sync(
1794 struct xlog *log,
1795 struct xlog_in_core *iclog,
1796 struct xlog_ticket *ticket)
1797{
1798 unsigned int count; /* byte count of bwrite */
1799 unsigned int roundoff; /* roundoff to BB or stripe */
1800 uint64_t bno;
1801 unsigned int size;
1802
1803 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1804 trace_xlog_iclog_sync(iclog, _RET_IP_);
1805
1806 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1807
1808 /*
1809 * If we have a ticket, account for the roundoff via the ticket
1810 * reservation to avoid touching the hot grant heads needlessly.
1811 * Otherwise, we have to move grant heads directly.
1812 */
1813 if (ticket) {
1814 ticket->t_curr_res -= roundoff;
1815 } else {
1816 xlog_grant_add_space(&log->l_reserve_head, roundoff);
1817 xlog_grant_add_space(&log->l_write_head, roundoff);
1818 }
1819
1820 /* put cycle number in every block */
1821 xlog_pack_data(log, iclog, roundoff);
1822
1823 /* real byte length */
1824 size = iclog->ic_offset;
1825 if (xfs_has_logv2(log->l_mp))
1826 size += roundoff;
1827 iclog->ic_header.h_len = cpu_to_be32(size);
1828
1829 XFS_STATS_INC(log->l_mp, xs_log_writes);
1830 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1831
1832 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1833
1834 /* Do we need to split this write into 2 parts? */
1835 if (bno + BTOBB(count) > log->l_logBBsize)
1836 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1837
1838 /* calculcate the checksum */
1839 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1840 iclog->ic_datap, size);
1841 /*
1842 * Intentionally corrupt the log record CRC based on the error injection
1843 * frequency, if defined. This facilitates testing log recovery in the
1844 * event of torn writes. Hence, set the IOABORT state to abort the log
1845 * write on I/O completion and shutdown the fs. The subsequent mount
1846 * detects the bad CRC and attempts to recover.
1847 */
1848#ifdef DEBUG
1849 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1850 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1851 iclog->ic_fail_crc = true;
1852 xfs_warn(log->l_mp,
1853 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1854 be64_to_cpu(iclog->ic_header.h_lsn));
1855 }
1856#endif
1857 xlog_verify_iclog(log, iclog, count);
1858 xlog_write_iclog(log, iclog, bno, count);
1859}
1860
1861/*
1862 * Deallocate a log structure
1863 */
1864STATIC void
1865xlog_dealloc_log(
1866 struct xlog *log)
1867{
1868 xlog_in_core_t *iclog, *next_iclog;
1869 int i;
1870
1871 /*
1872 * Destroy the CIL after waiting for iclog IO completion because an
1873 * iclog EIO error will try to shut down the log, which accesses the
1874 * CIL to wake up the waiters.
1875 */
1876 xlog_cil_destroy(log);
1877
1878 iclog = log->l_iclog;
1879 for (i = 0; i < log->l_iclog_bufs; i++) {
1880 next_iclog = iclog->ic_next;
1881 kvfree(iclog->ic_data);
1882 kfree(iclog);
1883 iclog = next_iclog;
1884 }
1885
1886 log->l_mp->m_log = NULL;
1887 destroy_workqueue(log->l_ioend_workqueue);
1888 kfree(log);
1889}
1890
1891/*
1892 * Update counters atomically now that memcpy is done.
1893 */
1894static inline void
1895xlog_state_finish_copy(
1896 struct xlog *log,
1897 struct xlog_in_core *iclog,
1898 int record_cnt,
1899 int copy_bytes)
1900{
1901 lockdep_assert_held(&log->l_icloglock);
1902
1903 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1904 iclog->ic_offset += copy_bytes;
1905}
1906
1907/*
1908 * print out info relating to regions written which consume
1909 * the reservation
1910 */
1911void
1912xlog_print_tic_res(
1913 struct xfs_mount *mp,
1914 struct xlog_ticket *ticket)
1915{
1916 xfs_warn(mp, "ticket reservation summary:");
1917 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
1918 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
1919 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
1920 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
1921}
1922
1923/*
1924 * Print a summary of the transaction.
1925 */
1926void
1927xlog_print_trans(
1928 struct xfs_trans *tp)
1929{
1930 struct xfs_mount *mp = tp->t_mountp;
1931 struct xfs_log_item *lip;
1932
1933 /* dump core transaction and ticket info */
1934 xfs_warn(mp, "transaction summary:");
1935 xfs_warn(mp, " log res = %d", tp->t_log_res);
1936 xfs_warn(mp, " log count = %d", tp->t_log_count);
1937 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
1938
1939 xlog_print_tic_res(mp, tp->t_ticket);
1940
1941 /* dump each log item */
1942 list_for_each_entry(lip, &tp->t_items, li_trans) {
1943 struct xfs_log_vec *lv = lip->li_lv;
1944 struct xfs_log_iovec *vec;
1945 int i;
1946
1947 xfs_warn(mp, "log item: ");
1948 xfs_warn(mp, " type = 0x%x", lip->li_type);
1949 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
1950 if (!lv)
1951 continue;
1952 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
1953 xfs_warn(mp, " size = %d", lv->lv_size);
1954 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
1955 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
1956
1957 /* dump each iovec for the log item */
1958 vec = lv->lv_iovecp;
1959 for (i = 0; i < lv->lv_niovecs; i++) {
1960 int dumplen = min(vec->i_len, 32);
1961
1962 xfs_warn(mp, " iovec[%d]", i);
1963 xfs_warn(mp, " type = 0x%x", vec->i_type);
1964 xfs_warn(mp, " len = %d", vec->i_len);
1965 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
1966 xfs_hex_dump(vec->i_addr, dumplen);
1967
1968 vec++;
1969 }
1970 }
1971}
1972
1973static inline void
1974xlog_write_iovec(
1975 struct xlog_in_core *iclog,
1976 uint32_t *log_offset,
1977 void *data,
1978 uint32_t write_len,
1979 int *bytes_left,
1980 uint32_t *record_cnt,
1981 uint32_t *data_cnt)
1982{
1983 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
1984 ASSERT(*log_offset % sizeof(int32_t) == 0);
1985 ASSERT(write_len % sizeof(int32_t) == 0);
1986
1987 memcpy(iclog->ic_datap + *log_offset, data, write_len);
1988 *log_offset += write_len;
1989 *bytes_left -= write_len;
1990 (*record_cnt)++;
1991 *data_cnt += write_len;
1992}
1993
1994/*
1995 * Write log vectors into a single iclog which is guaranteed by the caller
1996 * to have enough space to write the entire log vector into.
1997 */
1998static void
1999xlog_write_full(
2000 struct xfs_log_vec *lv,
2001 struct xlog_ticket *ticket,
2002 struct xlog_in_core *iclog,
2003 uint32_t *log_offset,
2004 uint32_t *len,
2005 uint32_t *record_cnt,
2006 uint32_t *data_cnt)
2007{
2008 int index;
2009
2010 ASSERT(*log_offset + *len <= iclog->ic_size ||
2011 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2012
2013 /*
2014 * Ordered log vectors have no regions to write so this
2015 * loop will naturally skip them.
2016 */
2017 for (index = 0; index < lv->lv_niovecs; index++) {
2018 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2019 struct xlog_op_header *ophdr = reg->i_addr;
2020
2021 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2022 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2023 reg->i_len, len, record_cnt, data_cnt);
2024 }
2025}
2026
2027static int
2028xlog_write_get_more_iclog_space(
2029 struct xlog_ticket *ticket,
2030 struct xlog_in_core **iclogp,
2031 uint32_t *log_offset,
2032 uint32_t len,
2033 uint32_t *record_cnt,
2034 uint32_t *data_cnt)
2035{
2036 struct xlog_in_core *iclog = *iclogp;
2037 struct xlog *log = iclog->ic_log;
2038 int error;
2039
2040 spin_lock(&log->l_icloglock);
2041 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2042 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2043 error = xlog_state_release_iclog(log, iclog, ticket);
2044 spin_unlock(&log->l_icloglock);
2045 if (error)
2046 return error;
2047
2048 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2049 log_offset);
2050 if (error)
2051 return error;
2052 *record_cnt = 0;
2053 *data_cnt = 0;
2054 *iclogp = iclog;
2055 return 0;
2056}
2057
2058/*
2059 * Write log vectors into a single iclog which is smaller than the current chain
2060 * length. We write until we cannot fit a full record into the remaining space
2061 * and then stop. We return the log vector that is to be written that cannot
2062 * wholly fit in the iclog.
2063 */
2064static int
2065xlog_write_partial(
2066 struct xfs_log_vec *lv,
2067 struct xlog_ticket *ticket,
2068 struct xlog_in_core **iclogp,
2069 uint32_t *log_offset,
2070 uint32_t *len,
2071 uint32_t *record_cnt,
2072 uint32_t *data_cnt)
2073{
2074 struct xlog_in_core *iclog = *iclogp;
2075 struct xlog_op_header *ophdr;
2076 int index = 0;
2077 uint32_t rlen;
2078 int error;
2079
2080 /* walk the logvec, copying until we run out of space in the iclog */
2081 for (index = 0; index < lv->lv_niovecs; index++) {
2082 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2083 uint32_t reg_offset = 0;
2084
2085 /*
2086 * The first region of a continuation must have a non-zero
2087 * length otherwise log recovery will just skip over it and
2088 * start recovering from the next opheader it finds. Because we
2089 * mark the next opheader as a continuation, recovery will then
2090 * incorrectly add the continuation to the previous region and
2091 * that breaks stuff.
2092 *
2093 * Hence if there isn't space for region data after the
2094 * opheader, then we need to start afresh with a new iclog.
2095 */
2096 if (iclog->ic_size - *log_offset <=
2097 sizeof(struct xlog_op_header)) {
2098 error = xlog_write_get_more_iclog_space(ticket,
2099 &iclog, log_offset, *len, record_cnt,
2100 data_cnt);
2101 if (error)
2102 return error;
2103 }
2104
2105 ophdr = reg->i_addr;
2106 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2107
2108 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2109 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2110 if (rlen != reg->i_len)
2111 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2112
2113 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2114 rlen, len, record_cnt, data_cnt);
2115
2116 /* If we wrote the whole region, move to the next. */
2117 if (rlen == reg->i_len)
2118 continue;
2119
2120 /*
2121 * We now have a partially written iovec, but it can span
2122 * multiple iclogs so we loop here. First we release the iclog
2123 * we currently have, then we get a new iclog and add a new
2124 * opheader. Then we continue copying from where we were until
2125 * we either complete the iovec or fill the iclog. If we
2126 * complete the iovec, then we increment the index and go right
2127 * back to the top of the outer loop. if we fill the iclog, we
2128 * run the inner loop again.
2129 *
2130 * This is complicated by the tail of a region using all the
2131 * space in an iclog and hence requiring us to release the iclog
2132 * and get a new one before returning to the outer loop. We must
2133 * always guarantee that we exit this inner loop with at least
2134 * space for log transaction opheaders left in the current
2135 * iclog, hence we cannot just terminate the loop at the end
2136 * of the of the continuation. So we loop while there is no
2137 * space left in the current iclog, and check for the end of the
2138 * continuation after getting a new iclog.
2139 */
2140 do {
2141 /*
2142 * Ensure we include the continuation opheader in the
2143 * space we need in the new iclog by adding that size
2144 * to the length we require. This continuation opheader
2145 * needs to be accounted to the ticket as the space it
2146 * consumes hasn't been accounted to the lv we are
2147 * writing.
2148 */
2149 error = xlog_write_get_more_iclog_space(ticket,
2150 &iclog, log_offset,
2151 *len + sizeof(struct xlog_op_header),
2152 record_cnt, data_cnt);
2153 if (error)
2154 return error;
2155
2156 ophdr = iclog->ic_datap + *log_offset;
2157 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2158 ophdr->oh_clientid = XFS_TRANSACTION;
2159 ophdr->oh_res2 = 0;
2160 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2161
2162 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2163 *log_offset += sizeof(struct xlog_op_header);
2164 *data_cnt += sizeof(struct xlog_op_header);
2165
2166 /*
2167 * If rlen fits in the iclog, then end the region
2168 * continuation. Otherwise we're going around again.
2169 */
2170 reg_offset += rlen;
2171 rlen = reg->i_len - reg_offset;
2172 if (rlen <= iclog->ic_size - *log_offset)
2173 ophdr->oh_flags |= XLOG_END_TRANS;
2174 else
2175 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2176
2177 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2178 ophdr->oh_len = cpu_to_be32(rlen);
2179
2180 xlog_write_iovec(iclog, log_offset,
2181 reg->i_addr + reg_offset,
2182 rlen, len, record_cnt, data_cnt);
2183
2184 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2185 }
2186
2187 /*
2188 * No more iovecs remain in this logvec so return the next log vec to
2189 * the caller so it can go back to fast path copying.
2190 */
2191 *iclogp = iclog;
2192 return 0;
2193}
2194
2195/*
2196 * Write some region out to in-core log
2197 *
2198 * This will be called when writing externally provided regions or when
2199 * writing out a commit record for a given transaction.
2200 *
2201 * General algorithm:
2202 * 1. Find total length of this write. This may include adding to the
2203 * lengths passed in.
2204 * 2. Check whether we violate the tickets reservation.
2205 * 3. While writing to this iclog
2206 * A. Reserve as much space in this iclog as can get
2207 * B. If this is first write, save away start lsn
2208 * C. While writing this region:
2209 * 1. If first write of transaction, write start record
2210 * 2. Write log operation header (header per region)
2211 * 3. Find out if we can fit entire region into this iclog
2212 * 4. Potentially, verify destination memcpy ptr
2213 * 5. Memcpy (partial) region
2214 * 6. If partial copy, release iclog; otherwise, continue
2215 * copying more regions into current iclog
2216 * 4. Mark want sync bit (in simulation mode)
2217 * 5. Release iclog for potential flush to on-disk log.
2218 *
2219 * ERRORS:
2220 * 1. Panic if reservation is overrun. This should never happen since
2221 * reservation amounts are generated internal to the filesystem.
2222 * NOTES:
2223 * 1. Tickets are single threaded data structures.
2224 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2225 * syncing routine. When a single log_write region needs to span
2226 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2227 * on all log operation writes which don't contain the end of the
2228 * region. The XLOG_END_TRANS bit is used for the in-core log
2229 * operation which contains the end of the continued log_write region.
2230 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2231 * we don't really know exactly how much space will be used. As a result,
2232 * we don't update ic_offset until the end when we know exactly how many
2233 * bytes have been written out.
2234 */
2235int
2236xlog_write(
2237 struct xlog *log,
2238 struct xfs_cil_ctx *ctx,
2239 struct list_head *lv_chain,
2240 struct xlog_ticket *ticket,
2241 uint32_t len)
2242
2243{
2244 struct xlog_in_core *iclog = NULL;
2245 struct xfs_log_vec *lv;
2246 uint32_t record_cnt = 0;
2247 uint32_t data_cnt = 0;
2248 int error = 0;
2249 int log_offset;
2250
2251 if (ticket->t_curr_res < 0) {
2252 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2253 "ctx ticket reservation ran out. Need to up reservation");
2254 xlog_print_tic_res(log->l_mp, ticket);
2255 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2256 }
2257
2258 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2259 &log_offset);
2260 if (error)
2261 return error;
2262
2263 ASSERT(log_offset <= iclog->ic_size - 1);
2264
2265 /*
2266 * If we have a context pointer, pass it the first iclog we are
2267 * writing to so it can record state needed for iclog write
2268 * ordering.
2269 */
2270 if (ctx)
2271 xlog_cil_set_ctx_write_state(ctx, iclog);
2272
2273 list_for_each_entry(lv, lv_chain, lv_list) {
2274 /*
2275 * If the entire log vec does not fit in the iclog, punt it to
2276 * the partial copy loop which can handle this case.
2277 */
2278 if (lv->lv_niovecs &&
2279 lv->lv_bytes > iclog->ic_size - log_offset) {
2280 error = xlog_write_partial(lv, ticket, &iclog,
2281 &log_offset, &len, &record_cnt,
2282 &data_cnt);
2283 if (error) {
2284 /*
2285 * We have no iclog to release, so just return
2286 * the error immediately.
2287 */
2288 return error;
2289 }
2290 } else {
2291 xlog_write_full(lv, ticket, iclog, &log_offset,
2292 &len, &record_cnt, &data_cnt);
2293 }
2294 }
2295 ASSERT(len == 0);
2296
2297 /*
2298 * We've already been guaranteed that the last writes will fit inside
2299 * the current iclog, and hence it will already have the space used by
2300 * those writes accounted to it. Hence we do not need to update the
2301 * iclog with the number of bytes written here.
2302 */
2303 spin_lock(&log->l_icloglock);
2304 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2305 error = xlog_state_release_iclog(log, iclog, ticket);
2306 spin_unlock(&log->l_icloglock);
2307
2308 return error;
2309}
2310
2311static void
2312xlog_state_activate_iclog(
2313 struct xlog_in_core *iclog,
2314 int *iclogs_changed)
2315{
2316 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2317 trace_xlog_iclog_activate(iclog, _RET_IP_);
2318
2319 /*
2320 * If the number of ops in this iclog indicate it just contains the
2321 * dummy transaction, we can change state into IDLE (the second time
2322 * around). Otherwise we should change the state into NEED a dummy.
2323 * We don't need to cover the dummy.
2324 */
2325 if (*iclogs_changed == 0 &&
2326 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2327 *iclogs_changed = 1;
2328 } else {
2329 /*
2330 * We have two dirty iclogs so start over. This could also be
2331 * num of ops indicating this is not the dummy going out.
2332 */
2333 *iclogs_changed = 2;
2334 }
2335
2336 iclog->ic_state = XLOG_STATE_ACTIVE;
2337 iclog->ic_offset = 0;
2338 iclog->ic_header.h_num_logops = 0;
2339 memset(iclog->ic_header.h_cycle_data, 0,
2340 sizeof(iclog->ic_header.h_cycle_data));
2341 iclog->ic_header.h_lsn = 0;
2342 iclog->ic_header.h_tail_lsn = 0;
2343}
2344
2345/*
2346 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2347 * ACTIVE after iclog I/O has completed.
2348 */
2349static void
2350xlog_state_activate_iclogs(
2351 struct xlog *log,
2352 int *iclogs_changed)
2353{
2354 struct xlog_in_core *iclog = log->l_iclog;
2355
2356 do {
2357 if (iclog->ic_state == XLOG_STATE_DIRTY)
2358 xlog_state_activate_iclog(iclog, iclogs_changed);
2359 /*
2360 * The ordering of marking iclogs ACTIVE must be maintained, so
2361 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2362 */
2363 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2364 break;
2365 } while ((iclog = iclog->ic_next) != log->l_iclog);
2366}
2367
2368static int
2369xlog_covered_state(
2370 int prev_state,
2371 int iclogs_changed)
2372{
2373 /*
2374 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2375 * wrote the first covering record (DONE). We go to IDLE if we just
2376 * wrote the second covering record (DONE2) and remain in IDLE until a
2377 * non-covering write occurs.
2378 */
2379 switch (prev_state) {
2380 case XLOG_STATE_COVER_IDLE:
2381 if (iclogs_changed == 1)
2382 return XLOG_STATE_COVER_IDLE;
2383 fallthrough;
2384 case XLOG_STATE_COVER_NEED:
2385 case XLOG_STATE_COVER_NEED2:
2386 break;
2387 case XLOG_STATE_COVER_DONE:
2388 if (iclogs_changed == 1)
2389 return XLOG_STATE_COVER_NEED2;
2390 break;
2391 case XLOG_STATE_COVER_DONE2:
2392 if (iclogs_changed == 1)
2393 return XLOG_STATE_COVER_IDLE;
2394 break;
2395 default:
2396 ASSERT(0);
2397 }
2398
2399 return XLOG_STATE_COVER_NEED;
2400}
2401
2402STATIC void
2403xlog_state_clean_iclog(
2404 struct xlog *log,
2405 struct xlog_in_core *dirty_iclog)
2406{
2407 int iclogs_changed = 0;
2408
2409 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2410
2411 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2412
2413 xlog_state_activate_iclogs(log, &iclogs_changed);
2414 wake_up_all(&dirty_iclog->ic_force_wait);
2415
2416 if (iclogs_changed) {
2417 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2418 iclogs_changed);
2419 }
2420}
2421
2422STATIC xfs_lsn_t
2423xlog_get_lowest_lsn(
2424 struct xlog *log)
2425{
2426 struct xlog_in_core *iclog = log->l_iclog;
2427 xfs_lsn_t lowest_lsn = 0, lsn;
2428
2429 do {
2430 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2431 iclog->ic_state == XLOG_STATE_DIRTY)
2432 continue;
2433
2434 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2435 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2436 lowest_lsn = lsn;
2437 } while ((iclog = iclog->ic_next) != log->l_iclog);
2438
2439 return lowest_lsn;
2440}
2441
2442/*
2443 * Return true if we need to stop processing, false to continue to the next
2444 * iclog. The caller will need to run callbacks if the iclog is returned in the
2445 * XLOG_STATE_CALLBACK state.
2446 */
2447static bool
2448xlog_state_iodone_process_iclog(
2449 struct xlog *log,
2450 struct xlog_in_core *iclog)
2451{
2452 xfs_lsn_t lowest_lsn;
2453 xfs_lsn_t header_lsn;
2454
2455 switch (iclog->ic_state) {
2456 case XLOG_STATE_ACTIVE:
2457 case XLOG_STATE_DIRTY:
2458 /*
2459 * Skip all iclogs in the ACTIVE & DIRTY states:
2460 */
2461 return false;
2462 case XLOG_STATE_DONE_SYNC:
2463 /*
2464 * Now that we have an iclog that is in the DONE_SYNC state, do
2465 * one more check here to see if we have chased our tail around.
2466 * If this is not the lowest lsn iclog, then we will leave it
2467 * for another completion to process.
2468 */
2469 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2470 lowest_lsn = xlog_get_lowest_lsn(log);
2471 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2472 return false;
2473 /*
2474 * If there are no callbacks on this iclog, we can mark it clean
2475 * immediately and return. Otherwise we need to run the
2476 * callbacks.
2477 */
2478 if (list_empty(&iclog->ic_callbacks)) {
2479 xlog_state_clean_iclog(log, iclog);
2480 return false;
2481 }
2482 trace_xlog_iclog_callback(iclog, _RET_IP_);
2483 iclog->ic_state = XLOG_STATE_CALLBACK;
2484 return false;
2485 default:
2486 /*
2487 * Can only perform callbacks in order. Since this iclog is not
2488 * in the DONE_SYNC state, we skip the rest and just try to
2489 * clean up.
2490 */
2491 return true;
2492 }
2493}
2494
2495/*
2496 * Loop over all the iclogs, running attached callbacks on them. Return true if
2497 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2498 * to handle transient shutdown state here at all because
2499 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2500 * cleanup of the callbacks.
2501 */
2502static bool
2503xlog_state_do_iclog_callbacks(
2504 struct xlog *log)
2505 __releases(&log->l_icloglock)
2506 __acquires(&log->l_icloglock)
2507{
2508 struct xlog_in_core *first_iclog = log->l_iclog;
2509 struct xlog_in_core *iclog = first_iclog;
2510 bool ran_callback = false;
2511
2512 do {
2513 LIST_HEAD(cb_list);
2514
2515 if (xlog_state_iodone_process_iclog(log, iclog))
2516 break;
2517 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2518 iclog = iclog->ic_next;
2519 continue;
2520 }
2521 list_splice_init(&iclog->ic_callbacks, &cb_list);
2522 spin_unlock(&log->l_icloglock);
2523
2524 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2525 xlog_cil_process_committed(&cb_list);
2526 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2527 ran_callback = true;
2528
2529 spin_lock(&log->l_icloglock);
2530 xlog_state_clean_iclog(log, iclog);
2531 iclog = iclog->ic_next;
2532 } while (iclog != first_iclog);
2533
2534 return ran_callback;
2535}
2536
2537
2538/*
2539 * Loop running iclog completion callbacks until there are no more iclogs in a
2540 * state that can run callbacks.
2541 */
2542STATIC void
2543xlog_state_do_callback(
2544 struct xlog *log)
2545{
2546 int flushcnt = 0;
2547 int repeats = 0;
2548
2549 spin_lock(&log->l_icloglock);
2550 while (xlog_state_do_iclog_callbacks(log)) {
2551 if (xlog_is_shutdown(log))
2552 break;
2553
2554 if (++repeats > 5000) {
2555 flushcnt += repeats;
2556 repeats = 0;
2557 xfs_warn(log->l_mp,
2558 "%s: possible infinite loop (%d iterations)",
2559 __func__, flushcnt);
2560 }
2561 }
2562
2563 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2564 wake_up_all(&log->l_flush_wait);
2565
2566 spin_unlock(&log->l_icloglock);
2567}
2568
2569
2570/*
2571 * Finish transitioning this iclog to the dirty state.
2572 *
2573 * Callbacks could take time, so they are done outside the scope of the
2574 * global state machine log lock.
2575 */
2576STATIC void
2577xlog_state_done_syncing(
2578 struct xlog_in_core *iclog)
2579{
2580 struct xlog *log = iclog->ic_log;
2581
2582 spin_lock(&log->l_icloglock);
2583 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2584 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2585
2586 /*
2587 * If we got an error, either on the first buffer, or in the case of
2588 * split log writes, on the second, we shut down the file system and
2589 * no iclogs should ever be attempted to be written to disk again.
2590 */
2591 if (!xlog_is_shutdown(log)) {
2592 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2593 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2594 }
2595
2596 /*
2597 * Someone could be sleeping prior to writing out the next
2598 * iclog buffer, we wake them all, one will get to do the
2599 * I/O, the others get to wait for the result.
2600 */
2601 wake_up_all(&iclog->ic_write_wait);
2602 spin_unlock(&log->l_icloglock);
2603 xlog_state_do_callback(log);
2604}
2605
2606/*
2607 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2608 * sleep. We wait on the flush queue on the head iclog as that should be
2609 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2610 * we will wait here and all new writes will sleep until a sync completes.
2611 *
2612 * The in-core logs are used in a circular fashion. They are not used
2613 * out-of-order even when an iclog past the head is free.
2614 *
2615 * return:
2616 * * log_offset where xlog_write() can start writing into the in-core
2617 * log's data space.
2618 * * in-core log pointer to which xlog_write() should write.
2619 * * boolean indicating this is a continued write to an in-core log.
2620 * If this is the last write, then the in-core log's offset field
2621 * needs to be incremented, depending on the amount of data which
2622 * is copied.
2623 */
2624STATIC int
2625xlog_state_get_iclog_space(
2626 struct xlog *log,
2627 int len,
2628 struct xlog_in_core **iclogp,
2629 struct xlog_ticket *ticket,
2630 int *logoffsetp)
2631{
2632 int log_offset;
2633 xlog_rec_header_t *head;
2634 xlog_in_core_t *iclog;
2635
2636restart:
2637 spin_lock(&log->l_icloglock);
2638 if (xlog_is_shutdown(log)) {
2639 spin_unlock(&log->l_icloglock);
2640 return -EIO;
2641 }
2642
2643 iclog = log->l_iclog;
2644 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2645 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2646
2647 /* Wait for log writes to have flushed */
2648 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2649 goto restart;
2650 }
2651
2652 head = &iclog->ic_header;
2653
2654 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2655 log_offset = iclog->ic_offset;
2656
2657 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2658
2659 /* On the 1st write to an iclog, figure out lsn. This works
2660 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2661 * committing to. If the offset is set, that's how many blocks
2662 * must be written.
2663 */
2664 if (log_offset == 0) {
2665 ticket->t_curr_res -= log->l_iclog_hsize;
2666 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2667 head->h_lsn = cpu_to_be64(
2668 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2669 ASSERT(log->l_curr_block >= 0);
2670 }
2671
2672 /* If there is enough room to write everything, then do it. Otherwise,
2673 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2674 * bit is on, so this will get flushed out. Don't update ic_offset
2675 * until you know exactly how many bytes get copied. Therefore, wait
2676 * until later to update ic_offset.
2677 *
2678 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2679 * can fit into remaining data section.
2680 */
2681 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2682 int error = 0;
2683
2684 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2685
2686 /*
2687 * If we are the only one writing to this iclog, sync it to
2688 * disk. We need to do an atomic compare and decrement here to
2689 * avoid racing with concurrent atomic_dec_and_lock() calls in
2690 * xlog_state_release_iclog() when there is more than one
2691 * reference to the iclog.
2692 */
2693 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2694 error = xlog_state_release_iclog(log, iclog, ticket);
2695 spin_unlock(&log->l_icloglock);
2696 if (error)
2697 return error;
2698 goto restart;
2699 }
2700
2701 /* Do we have enough room to write the full amount in the remainder
2702 * of this iclog? Or must we continue a write on the next iclog and
2703 * mark this iclog as completely taken? In the case where we switch
2704 * iclogs (to mark it taken), this particular iclog will release/sync
2705 * to disk in xlog_write().
2706 */
2707 if (len <= iclog->ic_size - iclog->ic_offset)
2708 iclog->ic_offset += len;
2709 else
2710 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2711 *iclogp = iclog;
2712
2713 ASSERT(iclog->ic_offset <= iclog->ic_size);
2714 spin_unlock(&log->l_icloglock);
2715
2716 *logoffsetp = log_offset;
2717 return 0;
2718}
2719
2720/*
2721 * The first cnt-1 times a ticket goes through here we don't need to move the
2722 * grant write head because the permanent reservation has reserved cnt times the
2723 * unit amount. Release part of current permanent unit reservation and reset
2724 * current reservation to be one units worth. Also move grant reservation head
2725 * forward.
2726 */
2727void
2728xfs_log_ticket_regrant(
2729 struct xlog *log,
2730 struct xlog_ticket *ticket)
2731{
2732 trace_xfs_log_ticket_regrant(log, ticket);
2733
2734 if (ticket->t_cnt > 0)
2735 ticket->t_cnt--;
2736
2737 xlog_grant_sub_space(&log->l_reserve_head, ticket->t_curr_res);
2738 xlog_grant_sub_space(&log->l_write_head, ticket->t_curr_res);
2739 ticket->t_curr_res = ticket->t_unit_res;
2740
2741 trace_xfs_log_ticket_regrant_sub(log, ticket);
2742
2743 /* just return if we still have some of the pre-reserved space */
2744 if (!ticket->t_cnt) {
2745 xlog_grant_add_space(&log->l_reserve_head, ticket->t_unit_res);
2746 trace_xfs_log_ticket_regrant_exit(log, ticket);
2747
2748 ticket->t_curr_res = ticket->t_unit_res;
2749 }
2750
2751 xfs_log_ticket_put(ticket);
2752}
2753
2754/*
2755 * Give back the space left from a reservation.
2756 *
2757 * All the information we need to make a correct determination of space left
2758 * is present. For non-permanent reservations, things are quite easy. The
2759 * count should have been decremented to zero. We only need to deal with the
2760 * space remaining in the current reservation part of the ticket. If the
2761 * ticket contains a permanent reservation, there may be left over space which
2762 * needs to be released. A count of N means that N-1 refills of the current
2763 * reservation can be done before we need to ask for more space. The first
2764 * one goes to fill up the first current reservation. Once we run out of
2765 * space, the count will stay at zero and the only space remaining will be
2766 * in the current reservation field.
2767 */
2768void
2769xfs_log_ticket_ungrant(
2770 struct xlog *log,
2771 struct xlog_ticket *ticket)
2772{
2773 int bytes;
2774
2775 trace_xfs_log_ticket_ungrant(log, ticket);
2776
2777 if (ticket->t_cnt > 0)
2778 ticket->t_cnt--;
2779
2780 trace_xfs_log_ticket_ungrant_sub(log, ticket);
2781
2782 /*
2783 * If this is a permanent reservation ticket, we may be able to free
2784 * up more space based on the remaining count.
2785 */
2786 bytes = ticket->t_curr_res;
2787 if (ticket->t_cnt > 0) {
2788 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2789 bytes += ticket->t_unit_res*ticket->t_cnt;
2790 }
2791
2792 xlog_grant_sub_space(&log->l_reserve_head, bytes);
2793 xlog_grant_sub_space(&log->l_write_head, bytes);
2794
2795 trace_xfs_log_ticket_ungrant_exit(log, ticket);
2796
2797 xfs_log_space_wake(log->l_mp);
2798 xfs_log_ticket_put(ticket);
2799}
2800
2801/*
2802 * This routine will mark the current iclog in the ring as WANT_SYNC and move
2803 * the current iclog pointer to the next iclog in the ring.
2804 */
2805void
2806xlog_state_switch_iclogs(
2807 struct xlog *log,
2808 struct xlog_in_core *iclog,
2809 int eventual_size)
2810{
2811 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2812 assert_spin_locked(&log->l_icloglock);
2813 trace_xlog_iclog_switch(iclog, _RET_IP_);
2814
2815 if (!eventual_size)
2816 eventual_size = iclog->ic_offset;
2817 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2818 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2819 log->l_prev_block = log->l_curr_block;
2820 log->l_prev_cycle = log->l_curr_cycle;
2821
2822 /* roll log?: ic_offset changed later */
2823 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2824
2825 /* Round up to next log-sunit */
2826 if (log->l_iclog_roundoff > BBSIZE) {
2827 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
2828 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2829 }
2830
2831 if (log->l_curr_block >= log->l_logBBsize) {
2832 /*
2833 * Rewind the current block before the cycle is bumped to make
2834 * sure that the combined LSN never transiently moves forward
2835 * when the log wraps to the next cycle. This is to support the
2836 * unlocked sample of these fields from xlog_valid_lsn(). Most
2837 * other cases should acquire l_icloglock.
2838 */
2839 log->l_curr_block -= log->l_logBBsize;
2840 ASSERT(log->l_curr_block >= 0);
2841 smp_wmb();
2842 log->l_curr_cycle++;
2843 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2844 log->l_curr_cycle++;
2845 }
2846 ASSERT(iclog == log->l_iclog);
2847 log->l_iclog = iclog->ic_next;
2848}
2849
2850/*
2851 * Force the iclog to disk and check if the iclog has been completed before
2852 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
2853 * pmem) or fast async storage because we drop the icloglock to issue the IO.
2854 * If completion has already occurred, tell the caller so that it can avoid an
2855 * unnecessary wait on the iclog.
2856 */
2857static int
2858xlog_force_and_check_iclog(
2859 struct xlog_in_core *iclog,
2860 bool *completed)
2861{
2862 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2863 int error;
2864
2865 *completed = false;
2866 error = xlog_force_iclog(iclog);
2867 if (error)
2868 return error;
2869
2870 /*
2871 * If the iclog has already been completed and reused the header LSN
2872 * will have been rewritten by completion
2873 */
2874 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
2875 *completed = true;
2876 return 0;
2877}
2878
2879/*
2880 * Write out all data in the in-core log as of this exact moment in time.
2881 *
2882 * Data may be written to the in-core log during this call. However,
2883 * we don't guarantee this data will be written out. A change from past
2884 * implementation means this routine will *not* write out zero length LRs.
2885 *
2886 * Basically, we try and perform an intelligent scan of the in-core logs.
2887 * If we determine there is no flushable data, we just return. There is no
2888 * flushable data if:
2889 *
2890 * 1. the current iclog is active and has no data; the previous iclog
2891 * is in the active or dirty state.
2892 * 2. the current iclog is drity, and the previous iclog is in the
2893 * active or dirty state.
2894 *
2895 * We may sleep if:
2896 *
2897 * 1. the current iclog is not in the active nor dirty state.
2898 * 2. the current iclog dirty, and the previous iclog is not in the
2899 * active nor dirty state.
2900 * 3. the current iclog is active, and there is another thread writing
2901 * to this particular iclog.
2902 * 4. a) the current iclog is active and has no other writers
2903 * b) when we return from flushing out this iclog, it is still
2904 * not in the active nor dirty state.
2905 */
2906int
2907xfs_log_force(
2908 struct xfs_mount *mp,
2909 uint flags)
2910{
2911 struct xlog *log = mp->m_log;
2912 struct xlog_in_core *iclog;
2913
2914 XFS_STATS_INC(mp, xs_log_force);
2915 trace_xfs_log_force(mp, 0, _RET_IP_);
2916
2917 xlog_cil_force(log);
2918
2919 spin_lock(&log->l_icloglock);
2920 if (xlog_is_shutdown(log))
2921 goto out_error;
2922
2923 iclog = log->l_iclog;
2924 trace_xlog_iclog_force(iclog, _RET_IP_);
2925
2926 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2927 (iclog->ic_state == XLOG_STATE_ACTIVE &&
2928 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
2929 /*
2930 * If the head is dirty or (active and empty), then we need to
2931 * look at the previous iclog.
2932 *
2933 * If the previous iclog is active or dirty we are done. There
2934 * is nothing to sync out. Otherwise, we attach ourselves to the
2935 * previous iclog and go to sleep.
2936 */
2937 iclog = iclog->ic_prev;
2938 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2939 if (atomic_read(&iclog->ic_refcnt) == 0) {
2940 /* We have exclusive access to this iclog. */
2941 bool completed;
2942
2943 if (xlog_force_and_check_iclog(iclog, &completed))
2944 goto out_error;
2945
2946 if (completed)
2947 goto out_unlock;
2948 } else {
2949 /*
2950 * Someone else is still writing to this iclog, so we
2951 * need to ensure that when they release the iclog it
2952 * gets synced immediately as we may be waiting on it.
2953 */
2954 xlog_state_switch_iclogs(log, iclog, 0);
2955 }
2956 }
2957
2958 /*
2959 * The iclog we are about to wait on may contain the checkpoint pushed
2960 * by the above xlog_cil_force() call, but it may not have been pushed
2961 * to disk yet. Like the ACTIVE case above, we need to make sure caches
2962 * are flushed when this iclog is written.
2963 */
2964 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
2965 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
2966
2967 if (flags & XFS_LOG_SYNC)
2968 return xlog_wait_on_iclog(iclog);
2969out_unlock:
2970 spin_unlock(&log->l_icloglock);
2971 return 0;
2972out_error:
2973 spin_unlock(&log->l_icloglock);
2974 return -EIO;
2975}
2976
2977/*
2978 * Force the log to a specific LSN.
2979 *
2980 * If an iclog with that lsn can be found:
2981 * If it is in the DIRTY state, just return.
2982 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2983 * state and go to sleep or return.
2984 * If it is in any other state, go to sleep or return.
2985 *
2986 * Synchronous forces are implemented with a wait queue. All callers trying
2987 * to force a given lsn to disk must wait on the queue attached to the
2988 * specific in-core log. When given in-core log finally completes its write
2989 * to disk, that thread will wake up all threads waiting on the queue.
2990 */
2991static int
2992xlog_force_lsn(
2993 struct xlog *log,
2994 xfs_lsn_t lsn,
2995 uint flags,
2996 int *log_flushed,
2997 bool already_slept)
2998{
2999 struct xlog_in_core *iclog;
3000 bool completed;
3001
3002 spin_lock(&log->l_icloglock);
3003 if (xlog_is_shutdown(log))
3004 goto out_error;
3005
3006 iclog = log->l_iclog;
3007 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3008 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3009 iclog = iclog->ic_next;
3010 if (iclog == log->l_iclog)
3011 goto out_unlock;
3012 }
3013
3014 switch (iclog->ic_state) {
3015 case XLOG_STATE_ACTIVE:
3016 /*
3017 * We sleep here if we haven't already slept (e.g. this is the
3018 * first time we've looked at the correct iclog buf) and the
3019 * buffer before us is going to be sync'ed. The reason for this
3020 * is that if we are doing sync transactions here, by waiting
3021 * for the previous I/O to complete, we can allow a few more
3022 * transactions into this iclog before we close it down.
3023 *
3024 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3025 * refcnt so we can release the log (which drops the ref count).
3026 * The state switch keeps new transaction commits from using
3027 * this buffer. When the current commits finish writing into
3028 * the buffer, the refcount will drop to zero and the buffer
3029 * will go out then.
3030 */
3031 if (!already_slept &&
3032 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3033 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3034 xlog_wait(&iclog->ic_prev->ic_write_wait,
3035 &log->l_icloglock);
3036 return -EAGAIN;
3037 }
3038 if (xlog_force_and_check_iclog(iclog, &completed))
3039 goto out_error;
3040 if (log_flushed)
3041 *log_flushed = 1;
3042 if (completed)
3043 goto out_unlock;
3044 break;
3045 case XLOG_STATE_WANT_SYNC:
3046 /*
3047 * This iclog may contain the checkpoint pushed by the
3048 * xlog_cil_force_seq() call, but there are other writers still
3049 * accessing it so it hasn't been pushed to disk yet. Like the
3050 * ACTIVE case above, we need to make sure caches are flushed
3051 * when this iclog is written.
3052 */
3053 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3054 break;
3055 default:
3056 /*
3057 * The entire checkpoint was written by the CIL force and is on
3058 * its way to disk already. It will be stable when it
3059 * completes, so we don't need to manipulate caches here at all.
3060 * We just need to wait for completion if necessary.
3061 */
3062 break;
3063 }
3064
3065 if (flags & XFS_LOG_SYNC)
3066 return xlog_wait_on_iclog(iclog);
3067out_unlock:
3068 spin_unlock(&log->l_icloglock);
3069 return 0;
3070out_error:
3071 spin_unlock(&log->l_icloglock);
3072 return -EIO;
3073}
3074
3075/*
3076 * Force the log to a specific checkpoint sequence.
3077 *
3078 * First force the CIL so that all the required changes have been flushed to the
3079 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3080 * the iclog that needs to be flushed to stable storage. If the caller needs
3081 * a synchronous log force, we will wait on the iclog with the LSN returned by
3082 * xlog_cil_force_seq() to be completed.
3083 */
3084int
3085xfs_log_force_seq(
3086 struct xfs_mount *mp,
3087 xfs_csn_t seq,
3088 uint flags,
3089 int *log_flushed)
3090{
3091 struct xlog *log = mp->m_log;
3092 xfs_lsn_t lsn;
3093 int ret;
3094 ASSERT(seq != 0);
3095
3096 XFS_STATS_INC(mp, xs_log_force);
3097 trace_xfs_log_force(mp, seq, _RET_IP_);
3098
3099 lsn = xlog_cil_force_seq(log, seq);
3100 if (lsn == NULLCOMMITLSN)
3101 return 0;
3102
3103 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3104 if (ret == -EAGAIN) {
3105 XFS_STATS_INC(mp, xs_log_force_sleep);
3106 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3107 }
3108 return ret;
3109}
3110
3111/*
3112 * Free a used ticket when its refcount falls to zero.
3113 */
3114void
3115xfs_log_ticket_put(
3116 xlog_ticket_t *ticket)
3117{
3118 ASSERT(atomic_read(&ticket->t_ref) > 0);
3119 if (atomic_dec_and_test(&ticket->t_ref))
3120 kmem_cache_free(xfs_log_ticket_cache, ticket);
3121}
3122
3123xlog_ticket_t *
3124xfs_log_ticket_get(
3125 xlog_ticket_t *ticket)
3126{
3127 ASSERT(atomic_read(&ticket->t_ref) > 0);
3128 atomic_inc(&ticket->t_ref);
3129 return ticket;
3130}
3131
3132/*
3133 * Figure out the total log space unit (in bytes) that would be
3134 * required for a log ticket.
3135 */
3136static int
3137xlog_calc_unit_res(
3138 struct xlog *log,
3139 int unit_bytes,
3140 int *niclogs)
3141{
3142 int iclog_space;
3143 uint num_headers;
3144
3145 /*
3146 * Permanent reservations have up to 'cnt'-1 active log operations
3147 * in the log. A unit in this case is the amount of space for one
3148 * of these log operations. Normal reservations have a cnt of 1
3149 * and their unit amount is the total amount of space required.
3150 *
3151 * The following lines of code account for non-transaction data
3152 * which occupy space in the on-disk log.
3153 *
3154 * Normal form of a transaction is:
3155 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3156 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3157 *
3158 * We need to account for all the leadup data and trailer data
3159 * around the transaction data.
3160 * And then we need to account for the worst case in terms of using
3161 * more space.
3162 * The worst case will happen if:
3163 * - the placement of the transaction happens to be such that the
3164 * roundoff is at its maximum
3165 * - the transaction data is synced before the commit record is synced
3166 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3167 * Therefore the commit record is in its own Log Record.
3168 * This can happen as the commit record is called with its
3169 * own region to xlog_write().
3170 * This then means that in the worst case, roundoff can happen for
3171 * the commit-rec as well.
3172 * The commit-rec is smaller than padding in this scenario and so it is
3173 * not added separately.
3174 */
3175
3176 /* for trans header */
3177 unit_bytes += sizeof(xlog_op_header_t);
3178 unit_bytes += sizeof(xfs_trans_header_t);
3179
3180 /* for start-rec */
3181 unit_bytes += sizeof(xlog_op_header_t);
3182
3183 /*
3184 * for LR headers - the space for data in an iclog is the size minus
3185 * the space used for the headers. If we use the iclog size, then we
3186 * undercalculate the number of headers required.
3187 *
3188 * Furthermore - the addition of op headers for split-recs might
3189 * increase the space required enough to require more log and op
3190 * headers, so take that into account too.
3191 *
3192 * IMPORTANT: This reservation makes the assumption that if this
3193 * transaction is the first in an iclog and hence has the LR headers
3194 * accounted to it, then the remaining space in the iclog is
3195 * exclusively for this transaction. i.e. if the transaction is larger
3196 * than the iclog, it will be the only thing in that iclog.
3197 * Fundamentally, this means we must pass the entire log vector to
3198 * xlog_write to guarantee this.
3199 */
3200 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3201 num_headers = howmany(unit_bytes, iclog_space);
3202
3203 /* for split-recs - ophdrs added when data split over LRs */
3204 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3205
3206 /* add extra header reservations if we overrun */
3207 while (!num_headers ||
3208 howmany(unit_bytes, iclog_space) > num_headers) {
3209 unit_bytes += sizeof(xlog_op_header_t);
3210 num_headers++;
3211 }
3212 unit_bytes += log->l_iclog_hsize * num_headers;
3213
3214 /* for commit-rec LR header - note: padding will subsume the ophdr */
3215 unit_bytes += log->l_iclog_hsize;
3216
3217 /* roundoff padding for transaction data and one for commit record */
3218 unit_bytes += 2 * log->l_iclog_roundoff;
3219
3220 if (niclogs)
3221 *niclogs = num_headers;
3222 return unit_bytes;
3223}
3224
3225int
3226xfs_log_calc_unit_res(
3227 struct xfs_mount *mp,
3228 int unit_bytes)
3229{
3230 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3231}
3232
3233/*
3234 * Allocate and initialise a new log ticket.
3235 */
3236struct xlog_ticket *
3237xlog_ticket_alloc(
3238 struct xlog *log,
3239 int unit_bytes,
3240 int cnt,
3241 bool permanent)
3242{
3243 struct xlog_ticket *tic;
3244 int unit_res;
3245
3246 tic = kmem_cache_zalloc(xfs_log_ticket_cache,
3247 GFP_KERNEL | __GFP_NOFAIL);
3248
3249 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3250
3251 atomic_set(&tic->t_ref, 1);
3252 tic->t_task = current;
3253 INIT_LIST_HEAD(&tic->t_queue);
3254 tic->t_unit_res = unit_res;
3255 tic->t_curr_res = unit_res;
3256 tic->t_cnt = cnt;
3257 tic->t_ocnt = cnt;
3258 tic->t_tid = get_random_u32();
3259 if (permanent)
3260 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3261
3262 return tic;
3263}
3264
3265#if defined(DEBUG)
3266static void
3267xlog_verify_dump_tail(
3268 struct xlog *log,
3269 struct xlog_in_core *iclog)
3270{
3271 xfs_alert(log->l_mp,
3272"ran out of log space tail 0x%llx/0x%llx, head lsn 0x%llx, head 0x%x/0x%x, prev head 0x%x/0x%x",
3273 iclog ? be64_to_cpu(iclog->ic_header.h_tail_lsn) : -1,
3274 atomic64_read(&log->l_tail_lsn),
3275 log->l_ailp->ail_head_lsn,
3276 log->l_curr_cycle, log->l_curr_block,
3277 log->l_prev_cycle, log->l_prev_block);
3278 xfs_alert(log->l_mp,
3279"write grant 0x%llx, reserve grant 0x%llx, tail_space 0x%llx, size 0x%x, iclog flags 0x%x",
3280 atomic64_read(&log->l_write_head.grant),
3281 atomic64_read(&log->l_reserve_head.grant),
3282 log->l_tail_space, log->l_logsize,
3283 iclog ? iclog->ic_flags : -1);
3284}
3285
3286/* Check if the new iclog will fit in the log. */
3287STATIC void
3288xlog_verify_tail_lsn(
3289 struct xlog *log,
3290 struct xlog_in_core *iclog)
3291{
3292 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3293 int blocks;
3294
3295 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3296 blocks = log->l_logBBsize -
3297 (log->l_prev_block - BLOCK_LSN(tail_lsn));
3298 if (blocks < BTOBB(iclog->ic_offset) +
3299 BTOBB(log->l_iclog_hsize)) {
3300 xfs_emerg(log->l_mp,
3301 "%s: ran out of log space", __func__);
3302 xlog_verify_dump_tail(log, iclog);
3303 }
3304 return;
3305 }
3306
3307 if (CYCLE_LSN(tail_lsn) + 1 != log->l_prev_cycle) {
3308 xfs_emerg(log->l_mp, "%s: head has wrapped tail.", __func__);
3309 xlog_verify_dump_tail(log, iclog);
3310 return;
3311 }
3312 if (BLOCK_LSN(tail_lsn) == log->l_prev_block) {
3313 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3314 xlog_verify_dump_tail(log, iclog);
3315 return;
3316 }
3317
3318 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3319 if (blocks < BTOBB(iclog->ic_offset) + 1) {
3320 xfs_emerg(log->l_mp, "%s: ran out of iclog space", __func__);
3321 xlog_verify_dump_tail(log, iclog);
3322 }
3323}
3324
3325/*
3326 * Perform a number of checks on the iclog before writing to disk.
3327 *
3328 * 1. Make sure the iclogs are still circular
3329 * 2. Make sure we have a good magic number
3330 * 3. Make sure we don't have magic numbers in the data
3331 * 4. Check fields of each log operation header for:
3332 * A. Valid client identifier
3333 * B. tid ptr value falls in valid ptr space (user space code)
3334 * C. Length in log record header is correct according to the
3335 * individual operation headers within record.
3336 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3337 * log, check the preceding blocks of the physical log to make sure all
3338 * the cycle numbers agree with the current cycle number.
3339 */
3340STATIC void
3341xlog_verify_iclog(
3342 struct xlog *log,
3343 struct xlog_in_core *iclog,
3344 int count)
3345{
3346 xlog_op_header_t *ophead;
3347 xlog_in_core_t *icptr;
3348 xlog_in_core_2_t *xhdr;
3349 void *base_ptr, *ptr, *p;
3350 ptrdiff_t field_offset;
3351 uint8_t clientid;
3352 int len, i, j, k, op_len;
3353 int idx;
3354
3355 /* check validity of iclog pointers */
3356 spin_lock(&log->l_icloglock);
3357 icptr = log->l_iclog;
3358 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3359 ASSERT(icptr);
3360
3361 if (icptr != log->l_iclog)
3362 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3363 spin_unlock(&log->l_icloglock);
3364
3365 /* check log magic numbers */
3366 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3367 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3368
3369 base_ptr = ptr = &iclog->ic_header;
3370 p = &iclog->ic_header;
3371 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3372 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3373 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3374 __func__);
3375 }
3376
3377 /* check fields */
3378 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3379 base_ptr = ptr = iclog->ic_datap;
3380 ophead = ptr;
3381 xhdr = iclog->ic_data;
3382 for (i = 0; i < len; i++) {
3383 ophead = ptr;
3384
3385 /* clientid is only 1 byte */
3386 p = &ophead->oh_clientid;
3387 field_offset = p - base_ptr;
3388 if (field_offset & 0x1ff) {
3389 clientid = ophead->oh_clientid;
3390 } else {
3391 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3392 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3393 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3394 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3395 clientid = xlog_get_client_id(
3396 xhdr[j].hic_xheader.xh_cycle_data[k]);
3397 } else {
3398 clientid = xlog_get_client_id(
3399 iclog->ic_header.h_cycle_data[idx]);
3400 }
3401 }
3402 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3403 xfs_warn(log->l_mp,
3404 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3405 __func__, i, clientid, ophead,
3406 (unsigned long)field_offset);
3407 }
3408
3409 /* check length */
3410 p = &ophead->oh_len;
3411 field_offset = p - base_ptr;
3412 if (field_offset & 0x1ff) {
3413 op_len = be32_to_cpu(ophead->oh_len);
3414 } else {
3415 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3416 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3417 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3418 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3419 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3420 } else {
3421 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3422 }
3423 }
3424 ptr += sizeof(xlog_op_header_t) + op_len;
3425 }
3426}
3427#endif
3428
3429/*
3430 * Perform a forced shutdown on the log.
3431 *
3432 * This can be called from low level log code to trigger a shutdown, or from the
3433 * high level mount shutdown code when the mount shuts down.
3434 *
3435 * Our main objectives here are to make sure that:
3436 * a. if the shutdown was not due to a log IO error, flush the logs to
3437 * disk. Anything modified after this is ignored.
3438 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3439 * parties to find out. Nothing new gets queued after this is done.
3440 * c. Tasks sleeping on log reservations, pinned objects and
3441 * other resources get woken up.
3442 * d. The mount is also marked as shut down so that log triggered shutdowns
3443 * still behave the same as if they called xfs_forced_shutdown().
3444 *
3445 * Return true if the shutdown cause was a log IO error and we actually shut the
3446 * log down.
3447 */
3448bool
3449xlog_force_shutdown(
3450 struct xlog *log,
3451 uint32_t shutdown_flags)
3452{
3453 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3454
3455 if (!log)
3456 return false;
3457
3458 /*
3459 * Ensure that there is only ever one log shutdown being processed.
3460 * If we allow the log force below on a second pass after shutting
3461 * down the log, we risk deadlocking the CIL push as it may require
3462 * locks on objects the current shutdown context holds (e.g. taking
3463 * buffer locks to abort buffers on last unpin of buf log items).
3464 */
3465 if (test_and_set_bit(XLOG_SHUTDOWN_STARTED, &log->l_opstate))
3466 return false;
3467
3468 /*
3469 * Flush all the completed transactions to disk before marking the log
3470 * being shut down. We need to do this first as shutting down the log
3471 * before the force will prevent the log force from flushing the iclogs
3472 * to disk.
3473 *
3474 * When we are in recovery, there are no transactions to flush, and
3475 * we don't want to touch the log because we don't want to perturb the
3476 * current head/tail for future recovery attempts. Hence we need to
3477 * avoid a log force in this case.
3478 *
3479 * If we are shutting down due to a log IO error, then we must avoid
3480 * trying to write the log as that may just result in more IO errors and
3481 * an endless shutdown/force loop.
3482 */
3483 if (!log_error && !xlog_in_recovery(log))
3484 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3485
3486 /*
3487 * Atomically set the shutdown state. If the shutdown state is already
3488 * set, there someone else is performing the shutdown and so we are done
3489 * here. This should never happen because we should only ever get called
3490 * once by the first shutdown caller.
3491 *
3492 * Much of the log state machine transitions assume that shutdown state
3493 * cannot change once they hold the log->l_icloglock. Hence we need to
3494 * hold that lock here, even though we use the atomic test_and_set_bit()
3495 * operation to set the shutdown state.
3496 */
3497 spin_lock(&log->l_icloglock);
3498 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3499 spin_unlock(&log->l_icloglock);
3500 ASSERT(0);
3501 return false;
3502 }
3503 spin_unlock(&log->l_icloglock);
3504
3505 /*
3506 * If this log shutdown also sets the mount shutdown state, issue a
3507 * shutdown warning message.
3508 */
3509 if (!xfs_set_shutdown(log->l_mp)) {
3510 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3511"Filesystem has been shut down due to log error (0x%x).",
3512 shutdown_flags);
3513 xfs_alert(log->l_mp,
3514"Please unmount the filesystem and rectify the problem(s).");
3515 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3516 xfs_stack_trace();
3517 }
3518
3519 /*
3520 * We don't want anybody waiting for log reservations after this. That
3521 * means we have to wake up everybody queued up on reserveq as well as
3522 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3523 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3524 * action is protected by the grant locks.
3525 */
3526 xlog_grant_head_wake_all(&log->l_reserve_head);
3527 xlog_grant_head_wake_all(&log->l_write_head);
3528
3529 /*
3530 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3531 * as if the log writes were completed. The abort handling in the log
3532 * item committed callback functions will do this again under lock to
3533 * avoid races.
3534 */
3535 spin_lock(&log->l_cilp->xc_push_lock);
3536 wake_up_all(&log->l_cilp->xc_start_wait);
3537 wake_up_all(&log->l_cilp->xc_commit_wait);
3538 spin_unlock(&log->l_cilp->xc_push_lock);
3539
3540 spin_lock(&log->l_icloglock);
3541 xlog_state_shutdown_callbacks(log);
3542 spin_unlock(&log->l_icloglock);
3543
3544 wake_up_var(&log->l_opstate);
3545 return log_error;
3546}
3547
3548STATIC int
3549xlog_iclogs_empty(
3550 struct xlog *log)
3551{
3552 xlog_in_core_t *iclog;
3553
3554 iclog = log->l_iclog;
3555 do {
3556 /* endianness does not matter here, zero is zero in
3557 * any language.
3558 */
3559 if (iclog->ic_header.h_num_logops)
3560 return 0;
3561 iclog = iclog->ic_next;
3562 } while (iclog != log->l_iclog);
3563 return 1;
3564}
3565
3566/*
3567 * Verify that an LSN stamped into a piece of metadata is valid. This is
3568 * intended for use in read verifiers on v5 superblocks.
3569 */
3570bool
3571xfs_log_check_lsn(
3572 struct xfs_mount *mp,
3573 xfs_lsn_t lsn)
3574{
3575 struct xlog *log = mp->m_log;
3576 bool valid;
3577
3578 /*
3579 * norecovery mode skips mount-time log processing and unconditionally
3580 * resets the in-core LSN. We can't validate in this mode, but
3581 * modifications are not allowed anyways so just return true.
3582 */
3583 if (xfs_has_norecovery(mp))
3584 return true;
3585
3586 /*
3587 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3588 * handled by recovery and thus safe to ignore here.
3589 */
3590 if (lsn == NULLCOMMITLSN)
3591 return true;
3592
3593 valid = xlog_valid_lsn(mp->m_log, lsn);
3594
3595 /* warn the user about what's gone wrong before verifier failure */
3596 if (!valid) {
3597 spin_lock(&log->l_icloglock);
3598 xfs_warn(mp,
3599"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3600"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3601 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3602 log->l_curr_cycle, log->l_curr_block);
3603 spin_unlock(&log->l_icloglock);
3604 }
3605
3606 return valid;
3607}