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
24 kmem_zone_t *xfs_log_ticket_zone;
25
26 /* Local miscellaneous function prototypes */
27 STATIC int
28 xlog_commit_record(
29 struct xlog *log,
30 struct xlog_ticket *ticket,
31 struct xlog_in_core **iclog,
32 xfs_lsn_t *commitlsnp);
33
34 STATIC struct xlog *
35 xlog_alloc_log(
36 struct xfs_mount *mp,
37 struct xfs_buftarg *log_target,
38 xfs_daddr_t blk_offset,
39 int num_bblks);
40 STATIC int
41 xlog_space_left(
42 struct xlog *log,
43 atomic64_t *head);
44 STATIC void
45 xlog_dealloc_log(
46 struct xlog *log);
47
48 /* local state machine functions */
49 STATIC void xlog_state_done_syncing(
50 struct xlog_in_core *iclog,
51 bool aborted);
52 STATIC int
53 xlog_state_get_iclog_space(
54 struct xlog *log,
55 int len,
56 struct xlog_in_core **iclog,
57 struct xlog_ticket *ticket,
58 int *continued_write,
59 int *logoffsetp);
60 STATIC int
61 xlog_state_release_iclog(
62 struct xlog *log,
63 struct xlog_in_core *iclog);
64 STATIC void
65 xlog_state_switch_iclogs(
66 struct xlog *log,
67 struct xlog_in_core *iclog,
68 int eventual_size);
69 STATIC void
70 xlog_state_want_sync(
71 struct xlog *log,
72 struct xlog_in_core *iclog);
73
74 STATIC void
75 xlog_grant_push_ail(
76 struct xlog *log,
77 int need_bytes);
78 STATIC void
79 xlog_regrant_reserve_log_space(
80 struct xlog *log,
81 struct xlog_ticket *ticket);
82 STATIC void
83 xlog_ungrant_log_space(
84 struct xlog *log,
85 struct xlog_ticket *ticket);
86
87 #if defined(DEBUG)
88 STATIC void
89 xlog_verify_dest_ptr(
90 struct xlog *log,
91 void *ptr);
92 STATIC void
93 xlog_verify_grant_tail(
94 struct xlog *log);
95 STATIC void
96 xlog_verify_iclog(
97 struct xlog *log,
98 struct xlog_in_core *iclog,
99 int count);
100 STATIC void
101 xlog_verify_tail_lsn(
102 struct xlog *log,
103 struct xlog_in_core *iclog,
104 xfs_lsn_t tail_lsn);
105 #else
106 #define xlog_verify_dest_ptr(a,b)
107 #define xlog_verify_grant_tail(a)
108 #define xlog_verify_iclog(a,b,c)
109 #define xlog_verify_tail_lsn(a,b,c)
110 #endif
111
112 STATIC int
113 xlog_iclogs_empty(
114 struct xlog *log);
115
116 static void
117 xlog_grant_sub_space(
118 struct xlog *log,
119 atomic64_t *head,
120 int bytes)
121 {
122 int64_t head_val = atomic64_read(head);
123 int64_t new, old;
124
125 do {
126 int cycle, space;
127
128 xlog_crack_grant_head_val(head_val, &cycle, &space);
129
130 space -= bytes;
131 if (space < 0) {
132 space += log->l_logsize;
133 cycle--;
134 }
135
136 old = head_val;
137 new = xlog_assign_grant_head_val(cycle, space);
138 head_val = atomic64_cmpxchg(head, old, new);
139 } while (head_val != old);
140 }
141
142 static void
143 xlog_grant_add_space(
144 struct xlog *log,
145 atomic64_t *head,
146 int bytes)
147 {
148 int64_t head_val = atomic64_read(head);
149 int64_t new, old;
150
151 do {
152 int tmp;
153 int cycle, space;
154
155 xlog_crack_grant_head_val(head_val, &cycle, &space);
156
157 tmp = log->l_logsize - space;
158 if (tmp > bytes)
159 space += bytes;
160 else {
161 space = bytes - tmp;
162 cycle++;
163 }
164
165 old = head_val;
166 new = xlog_assign_grant_head_val(cycle, space);
167 head_val = atomic64_cmpxchg(head, old, new);
168 } while (head_val != old);
169 }
170
171 STATIC void
172 xlog_grant_head_init(
173 struct xlog_grant_head *head)
174 {
175 xlog_assign_grant_head(&head->grant, 1, 0);
176 INIT_LIST_HEAD(&head->waiters);
177 spin_lock_init(&head->lock);
178 }
179
180 STATIC void
181 xlog_grant_head_wake_all(
182 struct xlog_grant_head *head)
183 {
184 struct xlog_ticket *tic;
185
186 spin_lock(&head->lock);
187 list_for_each_entry(tic, &head->waiters, t_queue)
188 wake_up_process(tic->t_task);
189 spin_unlock(&head->lock);
190 }
191
192 static inline int
193 xlog_ticket_reservation(
194 struct xlog *log,
195 struct xlog_grant_head *head,
196 struct xlog_ticket *tic)
197 {
198 if (head == &log->l_write_head) {
199 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
200 return tic->t_unit_res;
201 } else {
202 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
203 return tic->t_unit_res * tic->t_cnt;
204 else
205 return tic->t_unit_res;
206 }
207 }
208
209 STATIC bool
210 xlog_grant_head_wake(
211 struct xlog *log,
212 struct xlog_grant_head *head,
213 int *free_bytes)
214 {
215 struct xlog_ticket *tic;
216 int need_bytes;
217 bool woken_task = false;
218
219 list_for_each_entry(tic, &head->waiters, t_queue) {
220
221 /*
222 * There is a chance that the size of the CIL checkpoints in
223 * progress at the last AIL push target calculation resulted in
224 * limiting the target to the log head (l_last_sync_lsn) at the
225 * time. This may not reflect where the log head is now as the
226 * CIL checkpoints may have completed.
227 *
228 * Hence when we are woken here, it may be that the head of the
229 * log that has moved rather than the tail. As the tail didn't
230 * move, there still won't be space available for the
231 * reservation we require. However, if the AIL has already
232 * pushed to the target defined by the old log head location, we
233 * will hang here waiting for something else to update the AIL
234 * push target.
235 *
236 * Therefore, if there isn't space to wake the first waiter on
237 * the grant head, we need to push the AIL again to ensure the
238 * target reflects both the current log tail and log head
239 * position before we wait for the tail to move again.
240 */
241
242 need_bytes = xlog_ticket_reservation(log, head, tic);
243 if (*free_bytes < need_bytes) {
244 if (!woken_task)
245 xlog_grant_push_ail(log, need_bytes);
246 return false;
247 }
248
249 *free_bytes -= need_bytes;
250 trace_xfs_log_grant_wake_up(log, tic);
251 wake_up_process(tic->t_task);
252 woken_task = true;
253 }
254
255 return true;
256 }
257
258 STATIC int
259 xlog_grant_head_wait(
260 struct xlog *log,
261 struct xlog_grant_head *head,
262 struct xlog_ticket *tic,
263 int need_bytes) __releases(&head->lock)
264 __acquires(&head->lock)
265 {
266 list_add_tail(&tic->t_queue, &head->waiters);
267
268 do {
269 if (XLOG_FORCED_SHUTDOWN(log))
270 goto shutdown;
271 xlog_grant_push_ail(log, need_bytes);
272
273 __set_current_state(TASK_UNINTERRUPTIBLE);
274 spin_unlock(&head->lock);
275
276 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
277
278 trace_xfs_log_grant_sleep(log, tic);
279 schedule();
280 trace_xfs_log_grant_wake(log, tic);
281
282 spin_lock(&head->lock);
283 if (XLOG_FORCED_SHUTDOWN(log))
284 goto shutdown;
285 } while (xlog_space_left(log, &head->grant) < need_bytes);
286
287 list_del_init(&tic->t_queue);
288 return 0;
289 shutdown:
290 list_del_init(&tic->t_queue);
291 return -EIO;
292 }
293
294 /*
295 * Atomically get the log space required for a log ticket.
296 *
297 * Once a ticket gets put onto head->waiters, it will only return after the
298 * needed reservation is satisfied.
299 *
300 * This function is structured so that it has a lock free fast path. This is
301 * necessary because every new transaction reservation will come through this
302 * path. Hence any lock will be globally hot if we take it unconditionally on
303 * every pass.
304 *
305 * As tickets are only ever moved on and off head->waiters under head->lock, we
306 * only need to take that lock if we are going to add the ticket to the queue
307 * and sleep. We can avoid taking the lock if the ticket was never added to
308 * head->waiters because the t_queue list head will be empty and we hold the
309 * only reference to it so it can safely be checked unlocked.
310 */
311 STATIC int
312 xlog_grant_head_check(
313 struct xlog *log,
314 struct xlog_grant_head *head,
315 struct xlog_ticket *tic,
316 int *need_bytes)
317 {
318 int free_bytes;
319 int error = 0;
320
321 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
322
323 /*
324 * If there are other waiters on the queue then give them a chance at
325 * logspace before us. Wake up the first waiters, if we do not wake
326 * up all the waiters then go to sleep waiting for more free space,
327 * otherwise try to get some space for this transaction.
328 */
329 *need_bytes = xlog_ticket_reservation(log, head, tic);
330 free_bytes = xlog_space_left(log, &head->grant);
331 if (!list_empty_careful(&head->waiters)) {
332 spin_lock(&head->lock);
333 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
334 free_bytes < *need_bytes) {
335 error = xlog_grant_head_wait(log, head, tic,
336 *need_bytes);
337 }
338 spin_unlock(&head->lock);
339 } else if (free_bytes < *need_bytes) {
340 spin_lock(&head->lock);
341 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
342 spin_unlock(&head->lock);
343 }
344
345 return error;
346 }
347
348 static void
349 xlog_tic_reset_res(xlog_ticket_t *tic)
350 {
351 tic->t_res_num = 0;
352 tic->t_res_arr_sum = 0;
353 tic->t_res_num_ophdrs = 0;
354 }
355
356 static void
357 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
358 {
359 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
360 /* add to overflow and start again */
361 tic->t_res_o_flow += tic->t_res_arr_sum;
362 tic->t_res_num = 0;
363 tic->t_res_arr_sum = 0;
364 }
365
366 tic->t_res_arr[tic->t_res_num].r_len = len;
367 tic->t_res_arr[tic->t_res_num].r_type = type;
368 tic->t_res_arr_sum += len;
369 tic->t_res_num++;
370 }
371
372 /*
373 * Replenish the byte reservation required by moving the grant write head.
374 */
375 int
376 xfs_log_regrant(
377 struct xfs_mount *mp,
378 struct xlog_ticket *tic)
379 {
380 struct xlog *log = mp->m_log;
381 int need_bytes;
382 int error = 0;
383
384 if (XLOG_FORCED_SHUTDOWN(log))
385 return -EIO;
386
387 XFS_STATS_INC(mp, xs_try_logspace);
388
389 /*
390 * This is a new transaction on the ticket, so we need to change the
391 * transaction ID so that the next transaction has a different TID in
392 * the log. Just add one to the existing tid so that we can see chains
393 * of rolling transactions in the log easily.
394 */
395 tic->t_tid++;
396
397 xlog_grant_push_ail(log, tic->t_unit_res);
398
399 tic->t_curr_res = tic->t_unit_res;
400 xlog_tic_reset_res(tic);
401
402 if (tic->t_cnt > 0)
403 return 0;
404
405 trace_xfs_log_regrant(log, tic);
406
407 error = xlog_grant_head_check(log, &log->l_write_head, tic,
408 &need_bytes);
409 if (error)
410 goto out_error;
411
412 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
413 trace_xfs_log_regrant_exit(log, tic);
414 xlog_verify_grant_tail(log);
415 return 0;
416
417 out_error:
418 /*
419 * If we are failing, make sure the ticket doesn't have any current
420 * reservations. We don't want to add this back when the ticket/
421 * transaction gets cancelled.
422 */
423 tic->t_curr_res = 0;
424 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
425 return error;
426 }
427
428 /*
429 * Reserve log space and return a ticket corresponding to the reservation.
430 *
431 * Each reservation is going to reserve extra space for a log record header.
432 * When writes happen to the on-disk log, we don't subtract the length of the
433 * log record header from any reservation. By wasting space in each
434 * reservation, we prevent over allocation problems.
435 */
436 int
437 xfs_log_reserve(
438 struct xfs_mount *mp,
439 int unit_bytes,
440 int cnt,
441 struct xlog_ticket **ticp,
442 uint8_t client,
443 bool permanent)
444 {
445 struct xlog *log = mp->m_log;
446 struct xlog_ticket *tic;
447 int need_bytes;
448 int error = 0;
449
450 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
451
452 if (XLOG_FORCED_SHUTDOWN(log))
453 return -EIO;
454
455 XFS_STATS_INC(mp, xs_try_logspace);
456
457 ASSERT(*ticp == NULL);
458 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent, 0);
459 *ticp = tic;
460
461 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
462 : tic->t_unit_res);
463
464 trace_xfs_log_reserve(log, tic);
465
466 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
467 &need_bytes);
468 if (error)
469 goto out_error;
470
471 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
472 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
473 trace_xfs_log_reserve_exit(log, tic);
474 xlog_verify_grant_tail(log);
475 return 0;
476
477 out_error:
478 /*
479 * If we are failing, make sure the ticket doesn't have any current
480 * reservations. We don't want to add this back when the ticket/
481 * transaction gets cancelled.
482 */
483 tic->t_curr_res = 0;
484 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
485 return error;
486 }
487
488
489 /*
490 * NOTES:
491 *
492 * 1. currblock field gets updated at startup and after in-core logs
493 * marked as with WANT_SYNC.
494 */
495
496 /*
497 * This routine is called when a user of a log manager ticket is done with
498 * the reservation. If the ticket was ever used, then a commit record for
499 * the associated transaction is written out as a log operation header with
500 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
501 * a given ticket. If the ticket was one with a permanent reservation, then
502 * a few operations are done differently. Permanent reservation tickets by
503 * default don't release the reservation. They just commit the current
504 * transaction with the belief that the reservation is still needed. A flag
505 * must be passed in before permanent reservations are actually released.
506 * When these type of tickets are not released, they need to be set into
507 * the inited state again. By doing this, a start record will be written
508 * out when the next write occurs.
509 */
510 xfs_lsn_t
511 xfs_log_done(
512 struct xfs_mount *mp,
513 struct xlog_ticket *ticket,
514 struct xlog_in_core **iclog,
515 bool regrant)
516 {
517 struct xlog *log = mp->m_log;
518 xfs_lsn_t lsn = 0;
519
520 if (XLOG_FORCED_SHUTDOWN(log) ||
521 /*
522 * If nothing was ever written, don't write out commit record.
523 * If we get an error, just continue and give back the log ticket.
524 */
525 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
526 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
527 lsn = (xfs_lsn_t) -1;
528 regrant = false;
529 }
530
531
532 if (!regrant) {
533 trace_xfs_log_done_nonperm(log, ticket);
534
535 /*
536 * Release ticket if not permanent reservation or a specific
537 * request has been made to release a permanent reservation.
538 */
539 xlog_ungrant_log_space(log, ticket);
540 } else {
541 trace_xfs_log_done_perm(log, ticket);
542
543 xlog_regrant_reserve_log_space(log, ticket);
544 /* If this ticket was a permanent reservation and we aren't
545 * trying to release it, reset the inited flags; so next time
546 * we write, a start record will be written out.
547 */
548 ticket->t_flags |= XLOG_TIC_INITED;
549 }
550
551 xfs_log_ticket_put(ticket);
552 return lsn;
553 }
554
555 int
556 xfs_log_release_iclog(
557 struct xfs_mount *mp,
558 struct xlog_in_core *iclog)
559 {
560 if (xlog_state_release_iclog(mp->m_log, iclog)) {
561 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
562 return -EIO;
563 }
564
565 return 0;
566 }
567
568 /*
569 * Mount a log filesystem
570 *
571 * mp - ubiquitous xfs mount point structure
572 * log_target - buftarg of on-disk log device
573 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
574 * num_bblocks - Number of BBSIZE blocks in on-disk log
575 *
576 * Return error or zero.
577 */
578 int
579 xfs_log_mount(
580 xfs_mount_t *mp,
581 xfs_buftarg_t *log_target,
582 xfs_daddr_t blk_offset,
583 int num_bblks)
584 {
585 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
586 int error = 0;
587 int min_logfsbs;
588
589 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
590 xfs_notice(mp, "Mounting V%d Filesystem",
591 XFS_SB_VERSION_NUM(&mp->m_sb));
592 } else {
593 xfs_notice(mp,
594 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
595 XFS_SB_VERSION_NUM(&mp->m_sb));
596 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
597 }
598
599 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
600 if (IS_ERR(mp->m_log)) {
601 error = PTR_ERR(mp->m_log);
602 goto out;
603 }
604
605 /*
606 * Validate the given log space and drop a critical message via syslog
607 * if the log size is too small that would lead to some unexpected
608 * situations in transaction log space reservation stage.
609 *
610 * Note: we can't just reject the mount if the validation fails. This
611 * would mean that people would have to downgrade their kernel just to
612 * remedy the situation as there is no way to grow the log (short of
613 * black magic surgery with xfs_db).
614 *
615 * We can, however, reject mounts for CRC format filesystems, as the
616 * mkfs binary being used to make the filesystem should never create a
617 * filesystem with a log that is too small.
618 */
619 min_logfsbs = xfs_log_calc_minimum_size(mp);
620
621 if (mp->m_sb.sb_logblocks < min_logfsbs) {
622 xfs_warn(mp,
623 "Log size %d blocks too small, minimum size is %d blocks",
624 mp->m_sb.sb_logblocks, min_logfsbs);
625 error = -EINVAL;
626 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
627 xfs_warn(mp,
628 "Log size %d blocks too large, maximum size is %lld blocks",
629 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
630 error = -EINVAL;
631 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
632 xfs_warn(mp,
633 "log size %lld bytes too large, maximum size is %lld bytes",
634 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
635 XFS_MAX_LOG_BYTES);
636 error = -EINVAL;
637 } else if (mp->m_sb.sb_logsunit > 1 &&
638 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
639 xfs_warn(mp,
640 "log stripe unit %u bytes must be a multiple of block size",
641 mp->m_sb.sb_logsunit);
642 error = -EINVAL;
643 fatal = true;
644 }
645 if (error) {
646 /*
647 * Log check errors are always fatal on v5; or whenever bad
648 * metadata leads to a crash.
649 */
650 if (fatal) {
651 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
652 ASSERT(0);
653 goto out_free_log;
654 }
655 xfs_crit(mp, "Log size out of supported range.");
656 xfs_crit(mp,
657 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
658 }
659
660 /*
661 * Initialize the AIL now we have a log.
662 */
663 error = xfs_trans_ail_init(mp);
664 if (error) {
665 xfs_warn(mp, "AIL initialisation failed: error %d", error);
666 goto out_free_log;
667 }
668 mp->m_log->l_ailp = mp->m_ail;
669
670 /*
671 * skip log recovery on a norecovery mount. pretend it all
672 * just worked.
673 */
674 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
675 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
676
677 if (readonly)
678 mp->m_flags &= ~XFS_MOUNT_RDONLY;
679
680 error = xlog_recover(mp->m_log);
681
682 if (readonly)
683 mp->m_flags |= XFS_MOUNT_RDONLY;
684 if (error) {
685 xfs_warn(mp, "log mount/recovery failed: error %d",
686 error);
687 xlog_recover_cancel(mp->m_log);
688 goto out_destroy_ail;
689 }
690 }
691
692 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
693 "log");
694 if (error)
695 goto out_destroy_ail;
696
697 /* Normal transactions can now occur */
698 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
699
700 /*
701 * Now the log has been fully initialised and we know were our
702 * space grant counters are, we can initialise the permanent ticket
703 * needed for delayed logging to work.
704 */
705 xlog_cil_init_post_recovery(mp->m_log);
706
707 return 0;
708
709 out_destroy_ail:
710 xfs_trans_ail_destroy(mp);
711 out_free_log:
712 xlog_dealloc_log(mp->m_log);
713 out:
714 return error;
715 }
716
717 /*
718 * Finish the recovery of the file system. This is separate from the
719 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
720 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
721 * here.
722 *
723 * If we finish recovery successfully, start the background log work. If we are
724 * not doing recovery, then we have a RO filesystem and we don't need to start
725 * it.
726 */
727 int
728 xfs_log_mount_finish(
729 struct xfs_mount *mp)
730 {
731 int error = 0;
732 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
733 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
734
735 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
736 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
737 return 0;
738 } else if (readonly) {
739 /* Allow unlinked processing to proceed */
740 mp->m_flags &= ~XFS_MOUNT_RDONLY;
741 }
742
743 /*
744 * During the second phase of log recovery, we need iget and
745 * iput to behave like they do for an active filesystem.
746 * xfs_fs_drop_inode needs to be able to prevent the deletion
747 * of inodes before we're done replaying log items on those
748 * inodes. Turn it off immediately after recovery finishes
749 * so that we don't leak the quota inodes if subsequent mount
750 * activities fail.
751 *
752 * We let all inodes involved in redo item processing end up on
753 * the LRU instead of being evicted immediately so that if we do
754 * something to an unlinked inode, the irele won't cause
755 * premature truncation and freeing of the inode, which results
756 * in log recovery failure. We have to evict the unreferenced
757 * lru inodes after clearing SB_ACTIVE because we don't
758 * otherwise clean up the lru if there's a subsequent failure in
759 * xfs_mountfs, which leads to us leaking the inodes if nothing
760 * else (e.g. quotacheck) references the inodes before the
761 * mount failure occurs.
762 */
763 mp->m_super->s_flags |= SB_ACTIVE;
764 error = xlog_recover_finish(mp->m_log);
765 if (!error)
766 xfs_log_work_queue(mp);
767 mp->m_super->s_flags &= ~SB_ACTIVE;
768 evict_inodes(mp->m_super);
769
770 /*
771 * Drain the buffer LRU after log recovery. This is required for v4
772 * filesystems to avoid leaving around buffers with NULL verifier ops,
773 * but we do it unconditionally to make sure we're always in a clean
774 * cache state after mount.
775 *
776 * Don't push in the error case because the AIL may have pending intents
777 * that aren't removed until recovery is cancelled.
778 */
779 if (!error && recovered) {
780 xfs_log_force(mp, XFS_LOG_SYNC);
781 xfs_ail_push_all_sync(mp->m_ail);
782 }
783 xfs_wait_buftarg(mp->m_ddev_targp);
784
785 if (readonly)
786 mp->m_flags |= XFS_MOUNT_RDONLY;
787
788 return error;
789 }
790
791 /*
792 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
793 * the log.
794 */
795 void
796 xfs_log_mount_cancel(
797 struct xfs_mount *mp)
798 {
799 xlog_recover_cancel(mp->m_log);
800 xfs_log_unmount(mp);
801 }
802
803 /*
804 * Final log writes as part of unmount.
805 *
806 * Mark the filesystem clean as unmount happens. Note that during relocation
807 * this routine needs to be executed as part of source-bag while the
808 * deallocation must not be done until source-end.
809 */
810
811 /* Actually write the unmount record to disk. */
812 static void
813 xfs_log_write_unmount_record(
814 struct xfs_mount *mp)
815 {
816 /* the data section must be 32 bit size aligned */
817 struct xfs_unmount_log_format magic = {
818 .magic = XLOG_UNMOUNT_TYPE,
819 };
820 struct xfs_log_iovec reg = {
821 .i_addr = &magic,
822 .i_len = sizeof(magic),
823 .i_type = XLOG_REG_TYPE_UNMOUNT,
824 };
825 struct xfs_log_vec vec = {
826 .lv_niovecs = 1,
827 .lv_iovecp = ®,
828 };
829 struct xlog *log = mp->m_log;
830 struct xlog_in_core *iclog;
831 struct xlog_ticket *tic = NULL;
832 xfs_lsn_t lsn;
833 uint flags = XLOG_UNMOUNT_TRANS;
834 int error;
835
836 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
837 if (error)
838 goto out_err;
839
840 /*
841 * If we think the summary counters are bad, clear the unmount header
842 * flag in the unmount record so that the summary counters will be
843 * recalculated during log recovery at next mount. Refer to
844 * xlog_check_unmount_rec for more details.
845 */
846 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
847 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
848 xfs_alert(mp, "%s: will fix summary counters at next mount",
849 __func__);
850 flags &= ~XLOG_UNMOUNT_TRANS;
851 }
852
853 /* remove inited flag, and account for space used */
854 tic->t_flags = 0;
855 tic->t_curr_res -= sizeof(magic);
856 error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
857 /*
858 * At this point, we're umounting anyway, so there's no point in
859 * transitioning log state to IOERROR. Just continue...
860 */
861 out_err:
862 if (error)
863 xfs_alert(mp, "%s: unmount record failed", __func__);
864
865 spin_lock(&log->l_icloglock);
866 iclog = log->l_iclog;
867 atomic_inc(&iclog->ic_refcnt);
868 xlog_state_want_sync(log, iclog);
869 spin_unlock(&log->l_icloglock);
870 error = xlog_state_release_iclog(log, iclog);
871
872 spin_lock(&log->l_icloglock);
873 switch (iclog->ic_state) {
874 default:
875 if (!XLOG_FORCED_SHUTDOWN(log)) {
876 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
877 break;
878 }
879 /* fall through */
880 case XLOG_STATE_ACTIVE:
881 case XLOG_STATE_DIRTY:
882 spin_unlock(&log->l_icloglock);
883 break;
884 }
885
886 if (tic) {
887 trace_xfs_log_umount_write(log, tic);
888 xlog_ungrant_log_space(log, tic);
889 xfs_log_ticket_put(tic);
890 }
891 }
892
893 /*
894 * Unmount record used to have a string "Unmount filesystem--" in the
895 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
896 * We just write the magic number now since that particular field isn't
897 * currently architecture converted and "Unmount" is a bit foo.
898 * As far as I know, there weren't any dependencies on the old behaviour.
899 */
900
901 static int
902 xfs_log_unmount_write(xfs_mount_t *mp)
903 {
904 struct xlog *log = mp->m_log;
905 xlog_in_core_t *iclog;
906 #ifdef DEBUG
907 xlog_in_core_t *first_iclog;
908 #endif
909 int error;
910
911 /*
912 * Don't write out unmount record on norecovery mounts or ro devices.
913 * Or, if we are doing a forced umount (typically because of IO errors).
914 */
915 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
916 xfs_readonly_buftarg(log->l_targ)) {
917 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
918 return 0;
919 }
920
921 error = xfs_log_force(mp, XFS_LOG_SYNC);
922 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
923
924 #ifdef DEBUG
925 first_iclog = iclog = log->l_iclog;
926 do {
927 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
928 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
929 ASSERT(iclog->ic_offset == 0);
930 }
931 iclog = iclog->ic_next;
932 } while (iclog != first_iclog);
933 #endif
934 if (! (XLOG_FORCED_SHUTDOWN(log))) {
935 xfs_log_write_unmount_record(mp);
936 } else {
937 /*
938 * We're already in forced_shutdown mode, couldn't
939 * even attempt to write out the unmount transaction.
940 *
941 * Go through the motions of sync'ing and releasing
942 * the iclog, even though no I/O will actually happen,
943 * we need to wait for other log I/Os that may already
944 * be in progress. Do this as a separate section of
945 * code so we'll know if we ever get stuck here that
946 * we're in this odd situation of trying to unmount
947 * a file system that went into forced_shutdown as
948 * the result of an unmount..
949 */
950 spin_lock(&log->l_icloglock);
951 iclog = log->l_iclog;
952 atomic_inc(&iclog->ic_refcnt);
953
954 xlog_state_want_sync(log, iclog);
955 spin_unlock(&log->l_icloglock);
956 error = xlog_state_release_iclog(log, iclog);
957
958 spin_lock(&log->l_icloglock);
959
960 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
961 || iclog->ic_state == XLOG_STATE_DIRTY
962 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
963
964 xlog_wait(&iclog->ic_force_wait,
965 &log->l_icloglock);
966 } else {
967 spin_unlock(&log->l_icloglock);
968 }
969 }
970
971 return error;
972 } /* xfs_log_unmount_write */
973
974 /*
975 * Empty the log for unmount/freeze.
976 *
977 * To do this, we first need to shut down the background log work so it is not
978 * trying to cover the log as we clean up. We then need to unpin all objects in
979 * the log so we can then flush them out. Once they have completed their IO and
980 * run the callbacks removing themselves from the AIL, we can write the unmount
981 * record.
982 */
983 void
984 xfs_log_quiesce(
985 struct xfs_mount *mp)
986 {
987 cancel_delayed_work_sync(&mp->m_log->l_work);
988 xfs_log_force(mp, XFS_LOG_SYNC);
989
990 /*
991 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
992 * will push it, xfs_wait_buftarg() will not wait for it. Further,
993 * xfs_buf_iowait() cannot be used because it was pushed with the
994 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
995 * the IO to complete.
996 */
997 xfs_ail_push_all_sync(mp->m_ail);
998 xfs_wait_buftarg(mp->m_ddev_targp);
999 xfs_buf_lock(mp->m_sb_bp);
1000 xfs_buf_unlock(mp->m_sb_bp);
1001
1002 xfs_log_unmount_write(mp);
1003 }
1004
1005 /*
1006 * Shut down and release the AIL and Log.
1007 *
1008 * During unmount, we need to ensure we flush all the dirty metadata objects
1009 * from the AIL so that the log is empty before we write the unmount record to
1010 * the log. Once this is done, we can tear down the AIL and the log.
1011 */
1012 void
1013 xfs_log_unmount(
1014 struct xfs_mount *mp)
1015 {
1016 xfs_log_quiesce(mp);
1017
1018 xfs_trans_ail_destroy(mp);
1019
1020 xfs_sysfs_del(&mp->m_log->l_kobj);
1021
1022 xlog_dealloc_log(mp->m_log);
1023 }
1024
1025 void
1026 xfs_log_item_init(
1027 struct xfs_mount *mp,
1028 struct xfs_log_item *item,
1029 int type,
1030 const struct xfs_item_ops *ops)
1031 {
1032 item->li_mountp = mp;
1033 item->li_ailp = mp->m_ail;
1034 item->li_type = type;
1035 item->li_ops = ops;
1036 item->li_lv = NULL;
1037
1038 INIT_LIST_HEAD(&item->li_ail);
1039 INIT_LIST_HEAD(&item->li_cil);
1040 INIT_LIST_HEAD(&item->li_bio_list);
1041 INIT_LIST_HEAD(&item->li_trans);
1042 }
1043
1044 /*
1045 * Wake up processes waiting for log space after we have moved the log tail.
1046 */
1047 void
1048 xfs_log_space_wake(
1049 struct xfs_mount *mp)
1050 {
1051 struct xlog *log = mp->m_log;
1052 int free_bytes;
1053
1054 if (XLOG_FORCED_SHUTDOWN(log))
1055 return;
1056
1057 if (!list_empty_careful(&log->l_write_head.waiters)) {
1058 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1059
1060 spin_lock(&log->l_write_head.lock);
1061 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1062 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1063 spin_unlock(&log->l_write_head.lock);
1064 }
1065
1066 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1067 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1068
1069 spin_lock(&log->l_reserve_head.lock);
1070 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1071 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1072 spin_unlock(&log->l_reserve_head.lock);
1073 }
1074 }
1075
1076 /*
1077 * Determine if we have a transaction that has gone to disk that needs to be
1078 * covered. To begin the transition to the idle state firstly the log needs to
1079 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1080 * we start attempting to cover the log.
1081 *
1082 * Only if we are then in a state where covering is needed, the caller is
1083 * informed that dummy transactions are required to move the log into the idle
1084 * state.
1085 *
1086 * If there are any items in the AIl or CIL, then we do not want to attempt to
1087 * cover the log as we may be in a situation where there isn't log space
1088 * available to run a dummy transaction and this can lead to deadlocks when the
1089 * tail of the log is pinned by an item that is modified in the CIL. Hence
1090 * there's no point in running a dummy transaction at this point because we
1091 * can't start trying to idle the log until both the CIL and AIL are empty.
1092 */
1093 static int
1094 xfs_log_need_covered(xfs_mount_t *mp)
1095 {
1096 struct xlog *log = mp->m_log;
1097 int needed = 0;
1098
1099 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1100 return 0;
1101
1102 if (!xlog_cil_empty(log))
1103 return 0;
1104
1105 spin_lock(&log->l_icloglock);
1106 switch (log->l_covered_state) {
1107 case XLOG_STATE_COVER_DONE:
1108 case XLOG_STATE_COVER_DONE2:
1109 case XLOG_STATE_COVER_IDLE:
1110 break;
1111 case XLOG_STATE_COVER_NEED:
1112 case XLOG_STATE_COVER_NEED2:
1113 if (xfs_ail_min_lsn(log->l_ailp))
1114 break;
1115 if (!xlog_iclogs_empty(log))
1116 break;
1117
1118 needed = 1;
1119 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1120 log->l_covered_state = XLOG_STATE_COVER_DONE;
1121 else
1122 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1123 break;
1124 default:
1125 needed = 1;
1126 break;
1127 }
1128 spin_unlock(&log->l_icloglock);
1129 return needed;
1130 }
1131
1132 /*
1133 * We may be holding the log iclog lock upon entering this routine.
1134 */
1135 xfs_lsn_t
1136 xlog_assign_tail_lsn_locked(
1137 struct xfs_mount *mp)
1138 {
1139 struct xlog *log = mp->m_log;
1140 struct xfs_log_item *lip;
1141 xfs_lsn_t tail_lsn;
1142
1143 assert_spin_locked(&mp->m_ail->ail_lock);
1144
1145 /*
1146 * To make sure we always have a valid LSN for the log tail we keep
1147 * track of the last LSN which was committed in log->l_last_sync_lsn,
1148 * and use that when the AIL was empty.
1149 */
1150 lip = xfs_ail_min(mp->m_ail);
1151 if (lip)
1152 tail_lsn = lip->li_lsn;
1153 else
1154 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1155 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1156 atomic64_set(&log->l_tail_lsn, tail_lsn);
1157 return tail_lsn;
1158 }
1159
1160 xfs_lsn_t
1161 xlog_assign_tail_lsn(
1162 struct xfs_mount *mp)
1163 {
1164 xfs_lsn_t tail_lsn;
1165
1166 spin_lock(&mp->m_ail->ail_lock);
1167 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1168 spin_unlock(&mp->m_ail->ail_lock);
1169
1170 return tail_lsn;
1171 }
1172
1173 /*
1174 * Return the space in the log between the tail and the head. The head
1175 * is passed in the cycle/bytes formal parms. In the special case where
1176 * the reserve head has wrapped passed the tail, this calculation is no
1177 * longer valid. In this case, just return 0 which means there is no space
1178 * in the log. This works for all places where this function is called
1179 * with the reserve head. Of course, if the write head were to ever
1180 * wrap the tail, we should blow up. Rather than catch this case here,
1181 * we depend on other ASSERTions in other parts of the code. XXXmiken
1182 *
1183 * This code also handles the case where the reservation head is behind
1184 * the tail. The details of this case are described below, but the end
1185 * result is that we return the size of the log as the amount of space left.
1186 */
1187 STATIC int
1188 xlog_space_left(
1189 struct xlog *log,
1190 atomic64_t *head)
1191 {
1192 int free_bytes;
1193 int tail_bytes;
1194 int tail_cycle;
1195 int head_cycle;
1196 int head_bytes;
1197
1198 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1199 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1200 tail_bytes = BBTOB(tail_bytes);
1201 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1202 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1203 else if (tail_cycle + 1 < head_cycle)
1204 return 0;
1205 else if (tail_cycle < head_cycle) {
1206 ASSERT(tail_cycle == (head_cycle - 1));
1207 free_bytes = tail_bytes - head_bytes;
1208 } else {
1209 /*
1210 * The reservation head is behind the tail.
1211 * In this case we just want to return the size of the
1212 * log as the amount of space left.
1213 */
1214 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1215 xfs_alert(log->l_mp,
1216 " tail_cycle = %d, tail_bytes = %d",
1217 tail_cycle, tail_bytes);
1218 xfs_alert(log->l_mp,
1219 " GH cycle = %d, GH bytes = %d",
1220 head_cycle, head_bytes);
1221 ASSERT(0);
1222 free_bytes = log->l_logsize;
1223 }
1224 return free_bytes;
1225 }
1226
1227
1228 static void
1229 xlog_ioend_work(
1230 struct work_struct *work)
1231 {
1232 struct xlog_in_core *iclog =
1233 container_of(work, struct xlog_in_core, ic_end_io_work);
1234 struct xlog *log = iclog->ic_log;
1235 bool aborted = false;
1236 int error;
1237
1238 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1239 #ifdef DEBUG
1240 /* treat writes with injected CRC errors as failed */
1241 if (iclog->ic_fail_crc)
1242 error = -EIO;
1243 #endif
1244
1245 /*
1246 * Race to shutdown the filesystem if we see an error.
1247 */
1248 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1249 xfs_alert(log->l_mp, "log I/O error %d", error);
1250 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1251 /*
1252 * This flag will be propagated to the trans-committed
1253 * callback routines to let them know that the log-commit
1254 * didn't succeed.
1255 */
1256 aborted = true;
1257 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1258 aborted = true;
1259 }
1260
1261 xlog_state_done_syncing(iclog, aborted);
1262 bio_uninit(&iclog->ic_bio);
1263
1264 /*
1265 * Drop the lock to signal that we are done. Nothing references the
1266 * iclog after this, so an unmount waiting on this lock can now tear it
1267 * down safely. As such, it is unsafe to reference the iclog after the
1268 * unlock as we could race with it being freed.
1269 */
1270 up(&iclog->ic_sema);
1271 }
1272
1273 /*
1274 * Return size of each in-core log record buffer.
1275 *
1276 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1277 *
1278 * If the filesystem blocksize is too large, we may need to choose a
1279 * larger size since the directory code currently logs entire blocks.
1280 */
1281 STATIC void
1282 xlog_get_iclog_buffer_size(
1283 struct xfs_mount *mp,
1284 struct xlog *log)
1285 {
1286 if (mp->m_logbufs <= 0)
1287 mp->m_logbufs = XLOG_MAX_ICLOGS;
1288 if (mp->m_logbsize <= 0)
1289 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1290
1291 log->l_iclog_bufs = mp->m_logbufs;
1292 log->l_iclog_size = mp->m_logbsize;
1293
1294 /*
1295 * # headers = size / 32k - one header holds cycles from 32k of data.
1296 */
1297 log->l_iclog_heads =
1298 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1299 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1300 }
1301
1302 void
1303 xfs_log_work_queue(
1304 struct xfs_mount *mp)
1305 {
1306 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1307 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1308 }
1309
1310 /*
1311 * Every sync period we need to unpin all items in the AIL and push them to
1312 * disk. If there is nothing dirty, then we might need to cover the log to
1313 * indicate that the filesystem is idle.
1314 */
1315 static void
1316 xfs_log_worker(
1317 struct work_struct *work)
1318 {
1319 struct xlog *log = container_of(to_delayed_work(work),
1320 struct xlog, l_work);
1321 struct xfs_mount *mp = log->l_mp;
1322
1323 /* dgc: errors ignored - not fatal and nowhere to report them */
1324 if (xfs_log_need_covered(mp)) {
1325 /*
1326 * Dump a transaction into the log that contains no real change.
1327 * This is needed to stamp the current tail LSN into the log
1328 * during the covering operation.
1329 *
1330 * We cannot use an inode here for this - that will push dirty
1331 * state back up into the VFS and then periodic inode flushing
1332 * will prevent log covering from making progress. Hence we
1333 * synchronously log the superblock instead to ensure the
1334 * superblock is immediately unpinned and can be written back.
1335 */
1336 xfs_sync_sb(mp, true);
1337 } else
1338 xfs_log_force(mp, 0);
1339
1340 /* start pushing all the metadata that is currently dirty */
1341 xfs_ail_push_all(mp->m_ail);
1342
1343 /* queue us up again */
1344 xfs_log_work_queue(mp);
1345 }
1346
1347 /*
1348 * This routine initializes some of the log structure for a given mount point.
1349 * Its primary purpose is to fill in enough, so recovery can occur. However,
1350 * some other stuff may be filled in too.
1351 */
1352 STATIC struct xlog *
1353 xlog_alloc_log(
1354 struct xfs_mount *mp,
1355 struct xfs_buftarg *log_target,
1356 xfs_daddr_t blk_offset,
1357 int num_bblks)
1358 {
1359 struct xlog *log;
1360 xlog_rec_header_t *head;
1361 xlog_in_core_t **iclogp;
1362 xlog_in_core_t *iclog, *prev_iclog=NULL;
1363 int i;
1364 int error = -ENOMEM;
1365 uint log2_size = 0;
1366
1367 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1368 if (!log) {
1369 xfs_warn(mp, "Log allocation failed: No memory!");
1370 goto out;
1371 }
1372
1373 log->l_mp = mp;
1374 log->l_targ = log_target;
1375 log->l_logsize = BBTOB(num_bblks);
1376 log->l_logBBstart = blk_offset;
1377 log->l_logBBsize = num_bblks;
1378 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1379 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1380 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1381
1382 log->l_prev_block = -1;
1383 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1384 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1385 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1386 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1387
1388 xlog_grant_head_init(&log->l_reserve_head);
1389 xlog_grant_head_init(&log->l_write_head);
1390
1391 error = -EFSCORRUPTED;
1392 if (xfs_sb_version_hassector(&mp->m_sb)) {
1393 log2_size = mp->m_sb.sb_logsectlog;
1394 if (log2_size < BBSHIFT) {
1395 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1396 log2_size, BBSHIFT);
1397 goto out_free_log;
1398 }
1399
1400 log2_size -= BBSHIFT;
1401 if (log2_size > mp->m_sectbb_log) {
1402 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1403 log2_size, mp->m_sectbb_log);
1404 goto out_free_log;
1405 }
1406
1407 /* for larger sector sizes, must have v2 or external log */
1408 if (log2_size && log->l_logBBstart > 0 &&
1409 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1410 xfs_warn(mp,
1411 "log sector size (0x%x) invalid for configuration.",
1412 log2_size);
1413 goto out_free_log;
1414 }
1415 }
1416 log->l_sectBBsize = 1 << log2_size;
1417
1418 xlog_get_iclog_buffer_size(mp, log);
1419
1420 spin_lock_init(&log->l_icloglock);
1421 init_waitqueue_head(&log->l_flush_wait);
1422
1423 iclogp = &log->l_iclog;
1424 /*
1425 * The amount of memory to allocate for the iclog structure is
1426 * rather funky due to the way the structure is defined. It is
1427 * done this way so that we can use different sizes for machines
1428 * with different amounts of memory. See the definition of
1429 * xlog_in_core_t in xfs_log_priv.h for details.
1430 */
1431 ASSERT(log->l_iclog_size >= 4096);
1432 for (i = 0; i < log->l_iclog_bufs; i++) {
1433 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1434 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1435 sizeof(struct bio_vec);
1436
1437 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1438 if (!iclog)
1439 goto out_free_iclog;
1440
1441 *iclogp = iclog;
1442 iclog->ic_prev = prev_iclog;
1443 prev_iclog = iclog;
1444
1445 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1446 KM_MAYFAIL | KM_ZERO);
1447 if (!iclog->ic_data)
1448 goto out_free_iclog;
1449 #ifdef DEBUG
1450 log->l_iclog_bak[i] = &iclog->ic_header;
1451 #endif
1452 head = &iclog->ic_header;
1453 memset(head, 0, sizeof(xlog_rec_header_t));
1454 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1455 head->h_version = cpu_to_be32(
1456 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1457 head->h_size = cpu_to_be32(log->l_iclog_size);
1458 /* new fields */
1459 head->h_fmt = cpu_to_be32(XLOG_FMT);
1460 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1461
1462 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1463 iclog->ic_state = XLOG_STATE_ACTIVE;
1464 iclog->ic_log = log;
1465 atomic_set(&iclog->ic_refcnt, 0);
1466 spin_lock_init(&iclog->ic_callback_lock);
1467 INIT_LIST_HEAD(&iclog->ic_callbacks);
1468 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1469
1470 init_waitqueue_head(&iclog->ic_force_wait);
1471 init_waitqueue_head(&iclog->ic_write_wait);
1472 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1473 sema_init(&iclog->ic_sema, 1);
1474
1475 iclogp = &iclog->ic_next;
1476 }
1477 *iclogp = log->l_iclog; /* complete ring */
1478 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1479
1480 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1481 WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1482 mp->m_fsname);
1483 if (!log->l_ioend_workqueue)
1484 goto out_free_iclog;
1485
1486 error = xlog_cil_init(log);
1487 if (error)
1488 goto out_destroy_workqueue;
1489 return log;
1490
1491 out_destroy_workqueue:
1492 destroy_workqueue(log->l_ioend_workqueue);
1493 out_free_iclog:
1494 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1495 prev_iclog = iclog->ic_next;
1496 kmem_free(iclog->ic_data);
1497 kmem_free(iclog);
1498 if (prev_iclog == log->l_iclog)
1499 break;
1500 }
1501 out_free_log:
1502 kmem_free(log);
1503 out:
1504 return ERR_PTR(error);
1505 } /* xlog_alloc_log */
1506
1507
1508 /*
1509 * Write out the commit record of a transaction associated with the given
1510 * ticket. Return the lsn of the commit record.
1511 */
1512 STATIC int
1513 xlog_commit_record(
1514 struct xlog *log,
1515 struct xlog_ticket *ticket,
1516 struct xlog_in_core **iclog,
1517 xfs_lsn_t *commitlsnp)
1518 {
1519 struct xfs_mount *mp = log->l_mp;
1520 int error;
1521 struct xfs_log_iovec reg = {
1522 .i_addr = NULL,
1523 .i_len = 0,
1524 .i_type = XLOG_REG_TYPE_COMMIT,
1525 };
1526 struct xfs_log_vec vec = {
1527 .lv_niovecs = 1,
1528 .lv_iovecp = ®,
1529 };
1530
1531 ASSERT_ALWAYS(iclog);
1532 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1533 XLOG_COMMIT_TRANS);
1534 if (error)
1535 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1536 return error;
1537 }
1538
1539 /*
1540 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1541 * log space. This code pushes on the lsn which would supposedly free up
1542 * the 25% which we want to leave free. We may need to adopt a policy which
1543 * pushes on an lsn which is further along in the log once we reach the high
1544 * water mark. In this manner, we would be creating a low water mark.
1545 */
1546 STATIC void
1547 xlog_grant_push_ail(
1548 struct xlog *log,
1549 int need_bytes)
1550 {
1551 xfs_lsn_t threshold_lsn = 0;
1552 xfs_lsn_t last_sync_lsn;
1553 int free_blocks;
1554 int free_bytes;
1555 int threshold_block;
1556 int threshold_cycle;
1557 int free_threshold;
1558
1559 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1560
1561 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1562 free_blocks = BTOBBT(free_bytes);
1563
1564 /*
1565 * Set the threshold for the minimum number of free blocks in the
1566 * log to the maximum of what the caller needs, one quarter of the
1567 * log, and 256 blocks.
1568 */
1569 free_threshold = BTOBB(need_bytes);
1570 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1571 free_threshold = max(free_threshold, 256);
1572 if (free_blocks >= free_threshold)
1573 return;
1574
1575 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1576 &threshold_block);
1577 threshold_block += free_threshold;
1578 if (threshold_block >= log->l_logBBsize) {
1579 threshold_block -= log->l_logBBsize;
1580 threshold_cycle += 1;
1581 }
1582 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1583 threshold_block);
1584 /*
1585 * Don't pass in an lsn greater than the lsn of the last
1586 * log record known to be on disk. Use a snapshot of the last sync lsn
1587 * so that it doesn't change between the compare and the set.
1588 */
1589 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1590 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1591 threshold_lsn = last_sync_lsn;
1592
1593 /*
1594 * Get the transaction layer to kick the dirty buffers out to
1595 * disk asynchronously. No point in trying to do this if
1596 * the filesystem is shutting down.
1597 */
1598 if (!XLOG_FORCED_SHUTDOWN(log))
1599 xfs_ail_push(log->l_ailp, threshold_lsn);
1600 }
1601
1602 /*
1603 * Stamp cycle number in every block
1604 */
1605 STATIC void
1606 xlog_pack_data(
1607 struct xlog *log,
1608 struct xlog_in_core *iclog,
1609 int roundoff)
1610 {
1611 int i, j, k;
1612 int size = iclog->ic_offset + roundoff;
1613 __be32 cycle_lsn;
1614 char *dp;
1615
1616 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1617
1618 dp = iclog->ic_datap;
1619 for (i = 0; i < BTOBB(size); i++) {
1620 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1621 break;
1622 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1623 *(__be32 *)dp = cycle_lsn;
1624 dp += BBSIZE;
1625 }
1626
1627 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1628 xlog_in_core_2_t *xhdr = iclog->ic_data;
1629
1630 for ( ; i < BTOBB(size); i++) {
1631 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1632 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1633 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1634 *(__be32 *)dp = cycle_lsn;
1635 dp += BBSIZE;
1636 }
1637
1638 for (i = 1; i < log->l_iclog_heads; i++)
1639 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1640 }
1641 }
1642
1643 /*
1644 * Calculate the checksum for a log buffer.
1645 *
1646 * This is a little more complicated than it should be because the various
1647 * headers and the actual data are non-contiguous.
1648 */
1649 __le32
1650 xlog_cksum(
1651 struct xlog *log,
1652 struct xlog_rec_header *rhead,
1653 char *dp,
1654 int size)
1655 {
1656 uint32_t crc;
1657
1658 /* first generate the crc for the record header ... */
1659 crc = xfs_start_cksum_update((char *)rhead,
1660 sizeof(struct xlog_rec_header),
1661 offsetof(struct xlog_rec_header, h_crc));
1662
1663 /* ... then for additional cycle data for v2 logs ... */
1664 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1665 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1666 int i;
1667 int xheads;
1668
1669 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1670 if (size % XLOG_HEADER_CYCLE_SIZE)
1671 xheads++;
1672
1673 for (i = 1; i < xheads; i++) {
1674 crc = crc32c(crc, &xhdr[i].hic_xheader,
1675 sizeof(struct xlog_rec_ext_header));
1676 }
1677 }
1678
1679 /* ... and finally for the payload */
1680 crc = crc32c(crc, dp, size);
1681
1682 return xfs_end_cksum(crc);
1683 }
1684
1685 static void
1686 xlog_bio_end_io(
1687 struct bio *bio)
1688 {
1689 struct xlog_in_core *iclog = bio->bi_private;
1690
1691 queue_work(iclog->ic_log->l_ioend_workqueue,
1692 &iclog->ic_end_io_work);
1693 }
1694
1695 static void
1696 xlog_map_iclog_data(
1697 struct bio *bio,
1698 void *data,
1699 size_t count)
1700 {
1701 do {
1702 struct page *page = kmem_to_page(data);
1703 unsigned int off = offset_in_page(data);
1704 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1705
1706 WARN_ON_ONCE(bio_add_page(bio, page, len, off) != len);
1707
1708 data += len;
1709 count -= len;
1710 } while (count);
1711 }
1712
1713 STATIC void
1714 xlog_write_iclog(
1715 struct xlog *log,
1716 struct xlog_in_core *iclog,
1717 uint64_t bno,
1718 unsigned int count,
1719 bool need_flush)
1720 {
1721 ASSERT(bno < log->l_logBBsize);
1722
1723 /*
1724 * We lock the iclogbufs here so that we can serialise against I/O
1725 * completion during unmount. We might be processing a shutdown
1726 * triggered during unmount, and that can occur asynchronously to the
1727 * unmount thread, and hence we need to ensure that completes before
1728 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1729 * across the log IO to archieve that.
1730 */
1731 down(&iclog->ic_sema);
1732 if (unlikely(iclog->ic_state & XLOG_STATE_IOERROR)) {
1733 /*
1734 * It would seem logical to return EIO here, but we rely on
1735 * the log state machine to propagate I/O errors instead of
1736 * doing it here. We kick of the state machine and unlock
1737 * the buffer manually, the code needs to be kept in sync
1738 * with the I/O completion path.
1739 */
1740 xlog_state_done_syncing(iclog, XFS_LI_ABORTED);
1741 up(&iclog->ic_sema);
1742 return;
1743 }
1744
1745 iclog->ic_io_size = count;
1746
1747 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1748 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1749 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1750 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1751 iclog->ic_bio.bi_private = iclog;
1752 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1753 if (need_flush)
1754 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1755
1756 xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, iclog->ic_io_size);
1757 if (is_vmalloc_addr(iclog->ic_data))
1758 flush_kernel_vmap_range(iclog->ic_data, iclog->ic_io_size);
1759
1760 /*
1761 * If this log buffer would straddle the end of the log we will have
1762 * to split it up into two bios, so that we can continue at the start.
1763 */
1764 if (bno + BTOBB(count) > log->l_logBBsize) {
1765 struct bio *split;
1766
1767 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1768 GFP_NOIO, &fs_bio_set);
1769 bio_chain(split, &iclog->ic_bio);
1770 submit_bio(split);
1771
1772 /* restart at logical offset zero for the remainder */
1773 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1774 }
1775
1776 submit_bio(&iclog->ic_bio);
1777 }
1778
1779 /*
1780 * We need to bump cycle number for the part of the iclog that is
1781 * written to the start of the log. Watch out for the header magic
1782 * number case, though.
1783 */
1784 static void
1785 xlog_split_iclog(
1786 struct xlog *log,
1787 void *data,
1788 uint64_t bno,
1789 unsigned int count)
1790 {
1791 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1792 unsigned int i;
1793
1794 for (i = split_offset; i < count; i += BBSIZE) {
1795 uint32_t cycle = get_unaligned_be32(data + i);
1796
1797 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1798 cycle++;
1799 put_unaligned_be32(cycle, data + i);
1800 }
1801 }
1802
1803 static int
1804 xlog_calc_iclog_size(
1805 struct xlog *log,
1806 struct xlog_in_core *iclog,
1807 uint32_t *roundoff)
1808 {
1809 uint32_t count_init, count;
1810 bool use_lsunit;
1811
1812 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1813 log->l_mp->m_sb.sb_logsunit > 1;
1814
1815 /* Add for LR header */
1816 count_init = log->l_iclog_hsize + iclog->ic_offset;
1817
1818 /* Round out the log write size */
1819 if (use_lsunit) {
1820 /* we have a v2 stripe unit to use */
1821 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1822 } else {
1823 count = BBTOB(BTOBB(count_init));
1824 }
1825
1826 ASSERT(count >= count_init);
1827 *roundoff = count - count_init;
1828
1829 if (use_lsunit)
1830 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1831 else
1832 ASSERT(*roundoff < BBTOB(1));
1833 return count;
1834 }
1835
1836 /*
1837 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1838 * fashion. Previously, we should have moved the current iclog
1839 * ptr in the log to point to the next available iclog. This allows further
1840 * write to continue while this code syncs out an iclog ready to go.
1841 * Before an in-core log can be written out, the data section must be scanned
1842 * to save away the 1st word of each BBSIZE block into the header. We replace
1843 * it with the current cycle count. Each BBSIZE block is tagged with the
1844 * cycle count because there in an implicit assumption that drives will
1845 * guarantee that entire 512 byte blocks get written at once. In other words,
1846 * we can't have part of a 512 byte block written and part not written. By
1847 * tagging each block, we will know which blocks are valid when recovering
1848 * after an unclean shutdown.
1849 *
1850 * This routine is single threaded on the iclog. No other thread can be in
1851 * this routine with the same iclog. Changing contents of iclog can there-
1852 * fore be done without grabbing the state machine lock. Updating the global
1853 * log will require grabbing the lock though.
1854 *
1855 * The entire log manager uses a logical block numbering scheme. Only
1856 * xlog_write_iclog knows about the fact that the log may not start with
1857 * block zero on a given device.
1858 */
1859 STATIC void
1860 xlog_sync(
1861 struct xlog *log,
1862 struct xlog_in_core *iclog)
1863 {
1864 unsigned int count; /* byte count of bwrite */
1865 unsigned int roundoff; /* roundoff to BB or stripe */
1866 uint64_t bno;
1867 unsigned int size;
1868 bool need_flush = true, split = false;
1869
1870 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1871
1872 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1873
1874 /* move grant heads by roundoff in sync */
1875 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1876 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1877
1878 /* put cycle number in every block */
1879 xlog_pack_data(log, iclog, roundoff);
1880
1881 /* real byte length */
1882 size = iclog->ic_offset;
1883 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1884 size += roundoff;
1885 iclog->ic_header.h_len = cpu_to_be32(size);
1886
1887 XFS_STATS_INC(log->l_mp, xs_log_writes);
1888 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1889
1890 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1891
1892 /* Do we need to split this write into 2 parts? */
1893 if (bno + BTOBB(count) > log->l_logBBsize) {
1894 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1895 split = true;
1896 }
1897
1898 /* calculcate the checksum */
1899 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1900 iclog->ic_datap, size);
1901 /*
1902 * Intentionally corrupt the log record CRC based on the error injection
1903 * frequency, if defined. This facilitates testing log recovery in the
1904 * event of torn writes. Hence, set the IOABORT state to abort the log
1905 * write on I/O completion and shutdown the fs. The subsequent mount
1906 * detects the bad CRC and attempts to recover.
1907 */
1908 #ifdef DEBUG
1909 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1910 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1911 iclog->ic_fail_crc = true;
1912 xfs_warn(log->l_mp,
1913 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1914 be64_to_cpu(iclog->ic_header.h_lsn));
1915 }
1916 #endif
1917
1918 /*
1919 * Flush the data device before flushing the log to make sure all meta
1920 * data written back from the AIL actually made it to disk before
1921 * stamping the new log tail LSN into the log buffer. For an external
1922 * log we need to issue the flush explicitly, and unfortunately
1923 * synchronously here; for an internal log we can simply use the block
1924 * layer state machine for preflushes.
1925 */
1926 if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1927 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1928 need_flush = false;
1929 }
1930
1931 xlog_verify_iclog(log, iclog, count);
1932 xlog_write_iclog(log, iclog, bno, count, need_flush);
1933 }
1934
1935 /*
1936 * Deallocate a log structure
1937 */
1938 STATIC void
1939 xlog_dealloc_log(
1940 struct xlog *log)
1941 {
1942 xlog_in_core_t *iclog, *next_iclog;
1943 int i;
1944
1945 xlog_cil_destroy(log);
1946
1947 /*
1948 * Cycle all the iclogbuf locks to make sure all log IO completion
1949 * is done before we tear down these buffers.
1950 */
1951 iclog = log->l_iclog;
1952 for (i = 0; i < log->l_iclog_bufs; i++) {
1953 down(&iclog->ic_sema);
1954 up(&iclog->ic_sema);
1955 iclog = iclog->ic_next;
1956 }
1957
1958 iclog = log->l_iclog;
1959 for (i = 0; i < log->l_iclog_bufs; i++) {
1960 next_iclog = iclog->ic_next;
1961 kmem_free(iclog->ic_data);
1962 kmem_free(iclog);
1963 iclog = next_iclog;
1964 }
1965
1966 log->l_mp->m_log = NULL;
1967 destroy_workqueue(log->l_ioend_workqueue);
1968 kmem_free(log);
1969 } /* xlog_dealloc_log */
1970
1971 /*
1972 * Update counters atomically now that memcpy is done.
1973 */
1974 /* ARGSUSED */
1975 static inline void
1976 xlog_state_finish_copy(
1977 struct xlog *log,
1978 struct xlog_in_core *iclog,
1979 int record_cnt,
1980 int copy_bytes)
1981 {
1982 spin_lock(&log->l_icloglock);
1983
1984 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1985 iclog->ic_offset += copy_bytes;
1986
1987 spin_unlock(&log->l_icloglock);
1988 } /* xlog_state_finish_copy */
1989
1990
1991
1992
1993 /*
1994 * print out info relating to regions written which consume
1995 * the reservation
1996 */
1997 void
1998 xlog_print_tic_res(
1999 struct xfs_mount *mp,
2000 struct xlog_ticket *ticket)
2001 {
2002 uint i;
2003 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2004
2005 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2006 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2007 static char *res_type_str[] = {
2008 REG_TYPE_STR(BFORMAT, "bformat"),
2009 REG_TYPE_STR(BCHUNK, "bchunk"),
2010 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2011 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2012 REG_TYPE_STR(IFORMAT, "iformat"),
2013 REG_TYPE_STR(ICORE, "icore"),
2014 REG_TYPE_STR(IEXT, "iext"),
2015 REG_TYPE_STR(IBROOT, "ibroot"),
2016 REG_TYPE_STR(ILOCAL, "ilocal"),
2017 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2018 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2019 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2020 REG_TYPE_STR(QFORMAT, "qformat"),
2021 REG_TYPE_STR(DQUOT, "dquot"),
2022 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2023 REG_TYPE_STR(LRHEADER, "LR header"),
2024 REG_TYPE_STR(UNMOUNT, "unmount"),
2025 REG_TYPE_STR(COMMIT, "commit"),
2026 REG_TYPE_STR(TRANSHDR, "trans header"),
2027 REG_TYPE_STR(ICREATE, "inode create"),
2028 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2029 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2030 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2031 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2032 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2033 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2034 };
2035 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2036 #undef REG_TYPE_STR
2037
2038 xfs_warn(mp, "ticket reservation summary:");
2039 xfs_warn(mp, " unit res = %d bytes",
2040 ticket->t_unit_res);
2041 xfs_warn(mp, " current res = %d bytes",
2042 ticket->t_curr_res);
2043 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2044 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2045 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2046 ticket->t_res_num_ophdrs, ophdr_spc);
2047 xfs_warn(mp, " ophdr + reg = %u bytes",
2048 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2049 xfs_warn(mp, " num regions = %u",
2050 ticket->t_res_num);
2051
2052 for (i = 0; i < ticket->t_res_num; i++) {
2053 uint r_type = ticket->t_res_arr[i].r_type;
2054 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2055 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2056 "bad-rtype" : res_type_str[r_type]),
2057 ticket->t_res_arr[i].r_len);
2058 }
2059 }
2060
2061 /*
2062 * Print a summary of the transaction.
2063 */
2064 void
2065 xlog_print_trans(
2066 struct xfs_trans *tp)
2067 {
2068 struct xfs_mount *mp = tp->t_mountp;
2069 struct xfs_log_item *lip;
2070
2071 /* dump core transaction and ticket info */
2072 xfs_warn(mp, "transaction summary:");
2073 xfs_warn(mp, " log res = %d", tp->t_log_res);
2074 xfs_warn(mp, " log count = %d", tp->t_log_count);
2075 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2076
2077 xlog_print_tic_res(mp, tp->t_ticket);
2078
2079 /* dump each log item */
2080 list_for_each_entry(lip, &tp->t_items, li_trans) {
2081 struct xfs_log_vec *lv = lip->li_lv;
2082 struct xfs_log_iovec *vec;
2083 int i;
2084
2085 xfs_warn(mp, "log item: ");
2086 xfs_warn(mp, " type = 0x%x", lip->li_type);
2087 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2088 if (!lv)
2089 continue;
2090 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2091 xfs_warn(mp, " size = %d", lv->lv_size);
2092 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2093 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2094
2095 /* dump each iovec for the log item */
2096 vec = lv->lv_iovecp;
2097 for (i = 0; i < lv->lv_niovecs; i++) {
2098 int dumplen = min(vec->i_len, 32);
2099
2100 xfs_warn(mp, " iovec[%d]", i);
2101 xfs_warn(mp, " type = 0x%x", vec->i_type);
2102 xfs_warn(mp, " len = %d", vec->i_len);
2103 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2104 xfs_hex_dump(vec->i_addr, dumplen);
2105
2106 vec++;
2107 }
2108 }
2109 }
2110
2111 /*
2112 * Calculate the potential space needed by the log vector. Each region gets
2113 * its own xlog_op_header_t and may need to be double word aligned.
2114 */
2115 static int
2116 xlog_write_calc_vec_length(
2117 struct xlog_ticket *ticket,
2118 struct xfs_log_vec *log_vector)
2119 {
2120 struct xfs_log_vec *lv;
2121 int headers = 0;
2122 int len = 0;
2123 int i;
2124
2125 /* acct for start rec of xact */
2126 if (ticket->t_flags & XLOG_TIC_INITED)
2127 headers++;
2128
2129 for (lv = log_vector; lv; lv = lv->lv_next) {
2130 /* we don't write ordered log vectors */
2131 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2132 continue;
2133
2134 headers += lv->lv_niovecs;
2135
2136 for (i = 0; i < lv->lv_niovecs; i++) {
2137 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2138
2139 len += vecp->i_len;
2140 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2141 }
2142 }
2143
2144 ticket->t_res_num_ophdrs += headers;
2145 len += headers * sizeof(struct xlog_op_header);
2146
2147 return len;
2148 }
2149
2150 /*
2151 * If first write for transaction, insert start record We can't be trying to
2152 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2153 */
2154 static int
2155 xlog_write_start_rec(
2156 struct xlog_op_header *ophdr,
2157 struct xlog_ticket *ticket)
2158 {
2159 if (!(ticket->t_flags & XLOG_TIC_INITED))
2160 return 0;
2161
2162 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2163 ophdr->oh_clientid = ticket->t_clientid;
2164 ophdr->oh_len = 0;
2165 ophdr->oh_flags = XLOG_START_TRANS;
2166 ophdr->oh_res2 = 0;
2167
2168 ticket->t_flags &= ~XLOG_TIC_INITED;
2169
2170 return sizeof(struct xlog_op_header);
2171 }
2172
2173 static xlog_op_header_t *
2174 xlog_write_setup_ophdr(
2175 struct xlog *log,
2176 struct xlog_op_header *ophdr,
2177 struct xlog_ticket *ticket,
2178 uint flags)
2179 {
2180 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2181 ophdr->oh_clientid = ticket->t_clientid;
2182 ophdr->oh_res2 = 0;
2183
2184 /* are we copying a commit or unmount record? */
2185 ophdr->oh_flags = flags;
2186
2187 /*
2188 * We've seen logs corrupted with bad transaction client ids. This
2189 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2190 * and shut down the filesystem.
2191 */
2192 switch (ophdr->oh_clientid) {
2193 case XFS_TRANSACTION:
2194 case XFS_VOLUME:
2195 case XFS_LOG:
2196 break;
2197 default:
2198 xfs_warn(log->l_mp,
2199 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2200 ophdr->oh_clientid, ticket);
2201 return NULL;
2202 }
2203
2204 return ophdr;
2205 }
2206
2207 /*
2208 * Set up the parameters of the region copy into the log. This has
2209 * to handle region write split across multiple log buffers - this
2210 * state is kept external to this function so that this code can
2211 * be written in an obvious, self documenting manner.
2212 */
2213 static int
2214 xlog_write_setup_copy(
2215 struct xlog_ticket *ticket,
2216 struct xlog_op_header *ophdr,
2217 int space_available,
2218 int space_required,
2219 int *copy_off,
2220 int *copy_len,
2221 int *last_was_partial_copy,
2222 int *bytes_consumed)
2223 {
2224 int still_to_copy;
2225
2226 still_to_copy = space_required - *bytes_consumed;
2227 *copy_off = *bytes_consumed;
2228
2229 if (still_to_copy <= space_available) {
2230 /* write of region completes here */
2231 *copy_len = still_to_copy;
2232 ophdr->oh_len = cpu_to_be32(*copy_len);
2233 if (*last_was_partial_copy)
2234 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2235 *last_was_partial_copy = 0;
2236 *bytes_consumed = 0;
2237 return 0;
2238 }
2239
2240 /* partial write of region, needs extra log op header reservation */
2241 *copy_len = space_available;
2242 ophdr->oh_len = cpu_to_be32(*copy_len);
2243 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2244 if (*last_was_partial_copy)
2245 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2246 *bytes_consumed += *copy_len;
2247 (*last_was_partial_copy)++;
2248
2249 /* account for new log op header */
2250 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2251 ticket->t_res_num_ophdrs++;
2252
2253 return sizeof(struct xlog_op_header);
2254 }
2255
2256 static int
2257 xlog_write_copy_finish(
2258 struct xlog *log,
2259 struct xlog_in_core *iclog,
2260 uint flags,
2261 int *record_cnt,
2262 int *data_cnt,
2263 int *partial_copy,
2264 int *partial_copy_len,
2265 int log_offset,
2266 struct xlog_in_core **commit_iclog)
2267 {
2268 if (*partial_copy) {
2269 /*
2270 * This iclog has already been marked WANT_SYNC by
2271 * xlog_state_get_iclog_space.
2272 */
2273 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2274 *record_cnt = 0;
2275 *data_cnt = 0;
2276 return xlog_state_release_iclog(log, iclog);
2277 }
2278
2279 *partial_copy = 0;
2280 *partial_copy_len = 0;
2281
2282 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2283 /* no more space in this iclog - push it. */
2284 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2285 *record_cnt = 0;
2286 *data_cnt = 0;
2287
2288 spin_lock(&log->l_icloglock);
2289 xlog_state_want_sync(log, iclog);
2290 spin_unlock(&log->l_icloglock);
2291
2292 if (!commit_iclog)
2293 return xlog_state_release_iclog(log, iclog);
2294 ASSERT(flags & XLOG_COMMIT_TRANS);
2295 *commit_iclog = iclog;
2296 }
2297
2298 return 0;
2299 }
2300
2301 /*
2302 * Write some region out to in-core log
2303 *
2304 * This will be called when writing externally provided regions or when
2305 * writing out a commit record for a given transaction.
2306 *
2307 * General algorithm:
2308 * 1. Find total length of this write. This may include adding to the
2309 * lengths passed in.
2310 * 2. Check whether we violate the tickets reservation.
2311 * 3. While writing to this iclog
2312 * A. Reserve as much space in this iclog as can get
2313 * B. If this is first write, save away start lsn
2314 * C. While writing this region:
2315 * 1. If first write of transaction, write start record
2316 * 2. Write log operation header (header per region)
2317 * 3. Find out if we can fit entire region into this iclog
2318 * 4. Potentially, verify destination memcpy ptr
2319 * 5. Memcpy (partial) region
2320 * 6. If partial copy, release iclog; otherwise, continue
2321 * copying more regions into current iclog
2322 * 4. Mark want sync bit (in simulation mode)
2323 * 5. Release iclog for potential flush to on-disk log.
2324 *
2325 * ERRORS:
2326 * 1. Panic if reservation is overrun. This should never happen since
2327 * reservation amounts are generated internal to the filesystem.
2328 * NOTES:
2329 * 1. Tickets are single threaded data structures.
2330 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2331 * syncing routine. When a single log_write region needs to span
2332 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2333 * on all log operation writes which don't contain the end of the
2334 * region. The XLOG_END_TRANS bit is used for the in-core log
2335 * operation which contains the end of the continued log_write region.
2336 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2337 * we don't really know exactly how much space will be used. As a result,
2338 * we don't update ic_offset until the end when we know exactly how many
2339 * bytes have been written out.
2340 */
2341 int
2342 xlog_write(
2343 struct xlog *log,
2344 struct xfs_log_vec *log_vector,
2345 struct xlog_ticket *ticket,
2346 xfs_lsn_t *start_lsn,
2347 struct xlog_in_core **commit_iclog,
2348 uint flags)
2349 {
2350 struct xlog_in_core *iclog = NULL;
2351 struct xfs_log_iovec *vecp;
2352 struct xfs_log_vec *lv;
2353 int len;
2354 int index;
2355 int partial_copy = 0;
2356 int partial_copy_len = 0;
2357 int contwr = 0;
2358 int record_cnt = 0;
2359 int data_cnt = 0;
2360 int error;
2361
2362 *start_lsn = 0;
2363
2364 len = xlog_write_calc_vec_length(ticket, log_vector);
2365
2366 /*
2367 * Region headers and bytes are already accounted for.
2368 * We only need to take into account start records and
2369 * split regions in this function.
2370 */
2371 if (ticket->t_flags & XLOG_TIC_INITED)
2372 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2373
2374 /*
2375 * Commit record headers need to be accounted for. These
2376 * come in as separate writes so are easy to detect.
2377 */
2378 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2379 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2380
2381 if (ticket->t_curr_res < 0) {
2382 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2383 "ctx ticket reservation ran out. Need to up reservation");
2384 xlog_print_tic_res(log->l_mp, ticket);
2385 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2386 }
2387
2388 index = 0;
2389 lv = log_vector;
2390 vecp = lv->lv_iovecp;
2391 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2392 void *ptr;
2393 int log_offset;
2394
2395 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2396 &contwr, &log_offset);
2397 if (error)
2398 return error;
2399
2400 ASSERT(log_offset <= iclog->ic_size - 1);
2401 ptr = iclog->ic_datap + log_offset;
2402
2403 /* start_lsn is the first lsn written to. That's all we need. */
2404 if (!*start_lsn)
2405 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2406
2407 /*
2408 * This loop writes out as many regions as can fit in the amount
2409 * of space which was allocated by xlog_state_get_iclog_space().
2410 */
2411 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2412 struct xfs_log_iovec *reg;
2413 struct xlog_op_header *ophdr;
2414 int start_rec_copy;
2415 int copy_len;
2416 int copy_off;
2417 bool ordered = false;
2418
2419 /* ordered log vectors have no regions to write */
2420 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2421 ASSERT(lv->lv_niovecs == 0);
2422 ordered = true;
2423 goto next_lv;
2424 }
2425
2426 reg = &vecp[index];
2427 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2428 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2429
2430 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2431 if (start_rec_copy) {
2432 record_cnt++;
2433 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2434 start_rec_copy);
2435 }
2436
2437 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2438 if (!ophdr)
2439 return -EIO;
2440
2441 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2442 sizeof(struct xlog_op_header));
2443
2444 len += xlog_write_setup_copy(ticket, ophdr,
2445 iclog->ic_size-log_offset,
2446 reg->i_len,
2447 ©_off, ©_len,
2448 &partial_copy,
2449 &partial_copy_len);
2450 xlog_verify_dest_ptr(log, ptr);
2451
2452 /*
2453 * Copy region.
2454 *
2455 * Unmount records just log an opheader, so can have
2456 * empty payloads with no data region to copy. Hence we
2457 * only copy the payload if the vector says it has data
2458 * to copy.
2459 */
2460 ASSERT(copy_len >= 0);
2461 if (copy_len > 0) {
2462 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2463 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2464 copy_len);
2465 }
2466 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2467 record_cnt++;
2468 data_cnt += contwr ? copy_len : 0;
2469
2470 error = xlog_write_copy_finish(log, iclog, flags,
2471 &record_cnt, &data_cnt,
2472 &partial_copy,
2473 &partial_copy_len,
2474 log_offset,
2475 commit_iclog);
2476 if (error)
2477 return error;
2478
2479 /*
2480 * if we had a partial copy, we need to get more iclog
2481 * space but we don't want to increment the region
2482 * index because there is still more is this region to
2483 * write.
2484 *
2485 * If we completed writing this region, and we flushed
2486 * the iclog (indicated by resetting of the record
2487 * count), then we also need to get more log space. If
2488 * this was the last record, though, we are done and
2489 * can just return.
2490 */
2491 if (partial_copy)
2492 break;
2493
2494 if (++index == lv->lv_niovecs) {
2495 next_lv:
2496 lv = lv->lv_next;
2497 index = 0;
2498 if (lv)
2499 vecp = lv->lv_iovecp;
2500 }
2501 if (record_cnt == 0 && !ordered) {
2502 if (!lv)
2503 return 0;
2504 break;
2505 }
2506 }
2507 }
2508
2509 ASSERT(len == 0);
2510
2511 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2512 if (!commit_iclog)
2513 return xlog_state_release_iclog(log, iclog);
2514
2515 ASSERT(flags & XLOG_COMMIT_TRANS);
2516 *commit_iclog = iclog;
2517 return 0;
2518 }
2519
2520
2521 /*****************************************************************************
2522 *
2523 * State Machine functions
2524 *
2525 *****************************************************************************
2526 */
2527
2528 /*
2529 * An iclog has just finished IO completion processing, so we need to update
2530 * the iclog state and propagate that up into the overall log state. Hence we
2531 * prepare the iclog for cleaning, and then clean all the pending dirty iclogs
2532 * starting from the head, and then wake up any threads that are waiting for the
2533 * iclog to be marked clean.
2534 *
2535 * The ordering of marking iclogs ACTIVE must be maintained, so an iclog
2536 * doesn't become ACTIVE beyond one that is SYNCING. This is also required to
2537 * maintain the notion that we use a ordered wait queue to hold off would be
2538 * writers to the log when every iclog is trying to sync to disk.
2539 *
2540 * Caller must hold the icloglock before calling us.
2541 *
2542 * State Change: !IOERROR -> DIRTY -> ACTIVE
2543 */
2544 STATIC void
2545 xlog_state_clean_iclog(
2546 struct xlog *log,
2547 struct xlog_in_core *dirty_iclog)
2548 {
2549 struct xlog_in_core *iclog;
2550 int changed = 0;
2551
2552 /* Prepare the completed iclog. */
2553 if (!(dirty_iclog->ic_state & XLOG_STATE_IOERROR))
2554 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2555
2556 /* Walk all the iclogs to update the ordered active state. */
2557 iclog = log->l_iclog;
2558 do {
2559 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2560 iclog->ic_state = XLOG_STATE_ACTIVE;
2561 iclog->ic_offset = 0;
2562 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2563 /*
2564 * If the number of ops in this iclog indicate it just
2565 * contains the dummy transaction, we can
2566 * change state into IDLE (the second time around).
2567 * Otherwise we should change the state into
2568 * NEED a dummy.
2569 * We don't need to cover the dummy.
2570 */
2571 if (!changed &&
2572 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2573 XLOG_COVER_OPS)) {
2574 changed = 1;
2575 } else {
2576 /*
2577 * We have two dirty iclogs so start over
2578 * This could also be num of ops indicates
2579 * this is not the dummy going out.
2580 */
2581 changed = 2;
2582 }
2583 iclog->ic_header.h_num_logops = 0;
2584 memset(iclog->ic_header.h_cycle_data, 0,
2585 sizeof(iclog->ic_header.h_cycle_data));
2586 iclog->ic_header.h_lsn = 0;
2587 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2588 /* do nothing */;
2589 else
2590 break; /* stop cleaning */
2591 iclog = iclog->ic_next;
2592 } while (iclog != log->l_iclog);
2593
2594
2595 /*
2596 * Wake up threads waiting in xfs_log_force() for the dirty iclog
2597 * to be cleaned.
2598 */
2599 wake_up_all(&dirty_iclog->ic_force_wait);
2600
2601 /*
2602 * Change state for the dummy log recording.
2603 * We usually go to NEED. But we go to NEED2 if the changed indicates
2604 * we are done writing the dummy record.
2605 * If we are done with the second dummy recored (DONE2), then
2606 * we go to IDLE.
2607 */
2608 if (changed) {
2609 switch (log->l_covered_state) {
2610 case XLOG_STATE_COVER_IDLE:
2611 case XLOG_STATE_COVER_NEED:
2612 case XLOG_STATE_COVER_NEED2:
2613 log->l_covered_state = XLOG_STATE_COVER_NEED;
2614 break;
2615
2616 case XLOG_STATE_COVER_DONE:
2617 if (changed == 1)
2618 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2619 else
2620 log->l_covered_state = XLOG_STATE_COVER_NEED;
2621 break;
2622
2623 case XLOG_STATE_COVER_DONE2:
2624 if (changed == 1)
2625 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2626 else
2627 log->l_covered_state = XLOG_STATE_COVER_NEED;
2628 break;
2629
2630 default:
2631 ASSERT(0);
2632 }
2633 }
2634 }
2635
2636 STATIC xfs_lsn_t
2637 xlog_get_lowest_lsn(
2638 struct xlog *log)
2639 {
2640 struct xlog_in_core *iclog = log->l_iclog;
2641 xfs_lsn_t lowest_lsn = 0, lsn;
2642
2643 do {
2644 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2645 continue;
2646
2647 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2648 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2649 lowest_lsn = lsn;
2650 } while ((iclog = iclog->ic_next) != log->l_iclog);
2651
2652 return lowest_lsn;
2653 }
2654
2655 /*
2656 * Completion of a iclog IO does not imply that a transaction has completed, as
2657 * transactions can be large enough to span many iclogs. We cannot change the
2658 * tail of the log half way through a transaction as this may be the only
2659 * transaction in the log and moving the tail to point to the middle of it
2660 * will prevent recovery from finding the start of the transaction. Hence we
2661 * should only update the last_sync_lsn if this iclog contains transaction
2662 * completion callbacks on it.
2663 *
2664 * We have to do this before we drop the icloglock to ensure we are the only one
2665 * that can update it.
2666 *
2667 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2668 * the reservation grant head pushing. This is due to the fact that the push
2669 * target is bound by the current last_sync_lsn value. Hence if we have a large
2670 * amount of log space bound up in this committing transaction then the
2671 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2672 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2673 * should push the AIL to ensure the push target (and hence the grant head) is
2674 * no longer bound by the old log head location and can move forwards and make
2675 * progress again.
2676 */
2677 static void
2678 xlog_state_set_callback(
2679 struct xlog *log,
2680 struct xlog_in_core *iclog,
2681 xfs_lsn_t header_lsn)
2682 {
2683 iclog->ic_state = XLOG_STATE_CALLBACK;
2684
2685 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2686 header_lsn) <= 0);
2687
2688 if (list_empty_careful(&iclog->ic_callbacks))
2689 return;
2690
2691 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2692 xlog_grant_push_ail(log, 0);
2693 }
2694
2695 /*
2696 * Return true if we need to stop processing, false to continue to the next
2697 * iclog. The caller will need to run callbacks if the iclog is returned in the
2698 * XLOG_STATE_CALLBACK state.
2699 */
2700 static bool
2701 xlog_state_iodone_process_iclog(
2702 struct xlog *log,
2703 struct xlog_in_core *iclog,
2704 struct xlog_in_core *completed_iclog,
2705 bool *ioerror)
2706 {
2707 xfs_lsn_t lowest_lsn;
2708 xfs_lsn_t header_lsn;
2709
2710 /* Skip all iclogs in the ACTIVE & DIRTY states */
2711 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2712 return false;
2713
2714 /*
2715 * Between marking a filesystem SHUTDOWN and stopping the log, we do
2716 * flush all iclogs to disk (if there wasn't a log I/O error). So, we do
2717 * want things to go smoothly in case of just a SHUTDOWN w/o a
2718 * LOG_IO_ERROR.
2719 */
2720 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2721 *ioerror = true;
2722 return false;
2723 }
2724
2725 /*
2726 * Can only perform callbacks in order. Since this iclog is not in the
2727 * DONE_SYNC/ DO_CALLBACK state, we skip the rest and just try to clean
2728 * up. If we set our iclog to DO_CALLBACK, we will not process it when
2729 * we retry since a previous iclog is in the CALLBACK and the state
2730 * cannot change since we are holding the l_icloglock.
2731 */
2732 if (!(iclog->ic_state &
2733 (XLOG_STATE_DONE_SYNC | XLOG_STATE_DO_CALLBACK))) {
2734 if (completed_iclog &&
2735 (completed_iclog->ic_state == XLOG_STATE_DONE_SYNC)) {
2736 completed_iclog->ic_state = XLOG_STATE_DO_CALLBACK;
2737 }
2738 return true;
2739 }
2740
2741 /*
2742 * We now have an iclog that is in either the DO_CALLBACK or DONE_SYNC
2743 * states. The other states (WANT_SYNC, SYNCING, or CALLBACK were caught
2744 * by the above if and are going to clean (i.e. we aren't doing their
2745 * callbacks) see the above if.
2746 *
2747 * We will do one more check here to see if we have chased our tail
2748 * around. If this is not the lowest lsn iclog, then we will leave it
2749 * for another completion to process.
2750 */
2751 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2752 lowest_lsn = xlog_get_lowest_lsn(log);
2753 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2754 return false;
2755
2756 xlog_state_set_callback(log, iclog, header_lsn);
2757 return false;
2758
2759 }
2760
2761 /*
2762 * Keep processing entries in the iclog callback list until we come around and
2763 * it is empty. We need to atomically see that the list is empty and change the
2764 * state to DIRTY so that we don't miss any more callbacks being added.
2765 *
2766 * This function is called with the icloglock held and returns with it held. We
2767 * drop it while running callbacks, however, as holding it over thousands of
2768 * callbacks is unnecessary and causes excessive contention if we do.
2769 */
2770 static void
2771 xlog_state_do_iclog_callbacks(
2772 struct xlog *log,
2773 struct xlog_in_core *iclog,
2774 bool aborted)
2775 {
2776 spin_unlock(&log->l_icloglock);
2777 spin_lock(&iclog->ic_callback_lock);
2778 while (!list_empty(&iclog->ic_callbacks)) {
2779 LIST_HEAD(tmp);
2780
2781 list_splice_init(&iclog->ic_callbacks, &tmp);
2782
2783 spin_unlock(&iclog->ic_callback_lock);
2784 xlog_cil_process_committed(&tmp, aborted);
2785 spin_lock(&iclog->ic_callback_lock);
2786 }
2787
2788 /*
2789 * Pick up the icloglock while still holding the callback lock so we
2790 * serialise against anyone trying to add more callbacks to this iclog
2791 * now we've finished processing.
2792 */
2793 spin_lock(&log->l_icloglock);
2794 spin_unlock(&iclog->ic_callback_lock);
2795 }
2796
2797 #ifdef DEBUG
2798 /*
2799 * Make one last gasp attempt to see if iclogs are being left in limbo. If the
2800 * above loop finds an iclog earlier than the current iclog and in one of the
2801 * syncing states, the current iclog is put into DO_CALLBACK and the callbacks
2802 * are deferred to the completion of the earlier iclog. Walk the iclogs in order
2803 * and make sure that no iclog is in DO_CALLBACK unless an earlier iclog is in
2804 * one of the syncing states.
2805 *
2806 * Note that SYNCING|IOERROR is a valid state so we cannot just check for
2807 * ic_state == SYNCING.
2808 */
2809 static void
2810 xlog_state_callback_check_state(
2811 struct xlog *log)
2812 {
2813 struct xlog_in_core *first_iclog = log->l_iclog;
2814 struct xlog_in_core *iclog = first_iclog;
2815
2816 do {
2817 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2818 /*
2819 * Terminate the loop if iclogs are found in states
2820 * which will cause other threads to clean up iclogs.
2821 *
2822 * SYNCING - i/o completion will go through logs
2823 * DONE_SYNC - interrupt thread should be waiting for
2824 * l_icloglock
2825 * IOERROR - give up hope all ye who enter here
2826 */
2827 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2828 iclog->ic_state & XLOG_STATE_SYNCING ||
2829 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2830 iclog->ic_state == XLOG_STATE_IOERROR )
2831 break;
2832 iclog = iclog->ic_next;
2833 } while (first_iclog != iclog);
2834 }
2835 #else
2836 #define xlog_state_callback_check_state(l) ((void)0)
2837 #endif
2838
2839 STATIC void
2840 xlog_state_do_callback(
2841 struct xlog *log,
2842 bool aborted,
2843 struct xlog_in_core *ciclog)
2844 {
2845 struct xlog_in_core *iclog;
2846 struct xlog_in_core *first_iclog;
2847 bool did_callbacks = false;
2848 bool cycled_icloglock;
2849 bool ioerror;
2850 int flushcnt = 0;
2851 int repeats = 0;
2852
2853 spin_lock(&log->l_icloglock);
2854 do {
2855 /*
2856 * Scan all iclogs starting with the one pointed to by the
2857 * log. Reset this starting point each time the log is
2858 * unlocked (during callbacks).
2859 *
2860 * Keep looping through iclogs until one full pass is made
2861 * without running any callbacks.
2862 */
2863 first_iclog = log->l_iclog;
2864 iclog = log->l_iclog;
2865 cycled_icloglock = false;
2866 ioerror = false;
2867 repeats++;
2868
2869 do {
2870 if (xlog_state_iodone_process_iclog(log, iclog,
2871 ciclog, &ioerror))
2872 break;
2873
2874 if (!(iclog->ic_state &
2875 (XLOG_STATE_CALLBACK | XLOG_STATE_IOERROR))) {
2876 iclog = iclog->ic_next;
2877 continue;
2878 }
2879
2880 /*
2881 * Running callbacks will drop the icloglock which means
2882 * we'll have to run at least one more complete loop.
2883 */
2884 cycled_icloglock = true;
2885 xlog_state_do_iclog_callbacks(log, iclog, aborted);
2886
2887 xlog_state_clean_iclog(log, iclog);
2888 iclog = iclog->ic_next;
2889 } while (first_iclog != iclog);
2890
2891 did_callbacks |= cycled_icloglock;
2892
2893 if (repeats > 5000) {
2894 flushcnt += repeats;
2895 repeats = 0;
2896 xfs_warn(log->l_mp,
2897 "%s: possible infinite loop (%d iterations)",
2898 __func__, flushcnt);
2899 }
2900 } while (!ioerror && cycled_icloglock);
2901
2902 if (did_callbacks)
2903 xlog_state_callback_check_state(log);
2904
2905 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2906 wake_up_all(&log->l_flush_wait);
2907
2908 spin_unlock(&log->l_icloglock);
2909 }
2910
2911
2912 /*
2913 * Finish transitioning this iclog to the dirty state.
2914 *
2915 * Make sure that we completely execute this routine only when this is
2916 * the last call to the iclog. There is a good chance that iclog flushes,
2917 * when we reach the end of the physical log, get turned into 2 separate
2918 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2919 * routine. By using the reference count bwritecnt, we guarantee that only
2920 * the second completion goes through.
2921 *
2922 * Callbacks could take time, so they are done outside the scope of the
2923 * global state machine log lock.
2924 */
2925 STATIC void
2926 xlog_state_done_syncing(
2927 struct xlog_in_core *iclog,
2928 bool aborted)
2929 {
2930 struct xlog *log = iclog->ic_log;
2931
2932 spin_lock(&log->l_icloglock);
2933
2934 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2935 iclog->ic_state == XLOG_STATE_IOERROR);
2936 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2937
2938 /*
2939 * If we got an error, either on the first buffer, or in the case of
2940 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2941 * and none should ever be attempted to be written to disk
2942 * again.
2943 */
2944 if (iclog->ic_state != XLOG_STATE_IOERROR)
2945 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2946
2947 /*
2948 * Someone could be sleeping prior to writing out the next
2949 * iclog buffer, we wake them all, one will get to do the
2950 * I/O, the others get to wait for the result.
2951 */
2952 wake_up_all(&iclog->ic_write_wait);
2953 spin_unlock(&log->l_icloglock);
2954 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2955 } /* xlog_state_done_syncing */
2956
2957
2958 /*
2959 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2960 * sleep. We wait on the flush queue on the head iclog as that should be
2961 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2962 * we will wait here and all new writes will sleep until a sync completes.
2963 *
2964 * The in-core logs are used in a circular fashion. They are not used
2965 * out-of-order even when an iclog past the head is free.
2966 *
2967 * return:
2968 * * log_offset where xlog_write() can start writing into the in-core
2969 * log's data space.
2970 * * in-core log pointer to which xlog_write() should write.
2971 * * boolean indicating this is a continued write to an in-core log.
2972 * If this is the last write, then the in-core log's offset field
2973 * needs to be incremented, depending on the amount of data which
2974 * is copied.
2975 */
2976 STATIC int
2977 xlog_state_get_iclog_space(
2978 struct xlog *log,
2979 int len,
2980 struct xlog_in_core **iclogp,
2981 struct xlog_ticket *ticket,
2982 int *continued_write,
2983 int *logoffsetp)
2984 {
2985 int log_offset;
2986 xlog_rec_header_t *head;
2987 xlog_in_core_t *iclog;
2988 int error;
2989
2990 restart:
2991 spin_lock(&log->l_icloglock);
2992 if (XLOG_FORCED_SHUTDOWN(log)) {
2993 spin_unlock(&log->l_icloglock);
2994 return -EIO;
2995 }
2996
2997 iclog = log->l_iclog;
2998 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2999 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3000
3001 /* Wait for log writes to have flushed */
3002 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3003 goto restart;
3004 }
3005
3006 head = &iclog->ic_header;
3007
3008 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
3009 log_offset = iclog->ic_offset;
3010
3011 /* On the 1st write to an iclog, figure out lsn. This works
3012 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3013 * committing to. If the offset is set, that's how many blocks
3014 * must be written.
3015 */
3016 if (log_offset == 0) {
3017 ticket->t_curr_res -= log->l_iclog_hsize;
3018 xlog_tic_add_region(ticket,
3019 log->l_iclog_hsize,
3020 XLOG_REG_TYPE_LRHEADER);
3021 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3022 head->h_lsn = cpu_to_be64(
3023 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3024 ASSERT(log->l_curr_block >= 0);
3025 }
3026
3027 /* If there is enough room to write everything, then do it. Otherwise,
3028 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3029 * bit is on, so this will get flushed out. Don't update ic_offset
3030 * until you know exactly how many bytes get copied. Therefore, wait
3031 * until later to update ic_offset.
3032 *
3033 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3034 * can fit into remaining data section.
3035 */
3036 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3037 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3038
3039 /*
3040 * If I'm the only one writing to this iclog, sync it to disk.
3041 * We need to do an atomic compare and decrement here to avoid
3042 * racing with concurrent atomic_dec_and_lock() calls in
3043 * xlog_state_release_iclog() when there is more than one
3044 * reference to the iclog.
3045 */
3046 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3047 /* we are the only one */
3048 spin_unlock(&log->l_icloglock);
3049 error = xlog_state_release_iclog(log, iclog);
3050 if (error)
3051 return error;
3052 } else {
3053 spin_unlock(&log->l_icloglock);
3054 }
3055 goto restart;
3056 }
3057
3058 /* Do we have enough room to write the full amount in the remainder
3059 * of this iclog? Or must we continue a write on the next iclog and
3060 * mark this iclog as completely taken? In the case where we switch
3061 * iclogs (to mark it taken), this particular iclog will release/sync
3062 * to disk in xlog_write().
3063 */
3064 if (len <= iclog->ic_size - iclog->ic_offset) {
3065 *continued_write = 0;
3066 iclog->ic_offset += len;
3067 } else {
3068 *continued_write = 1;
3069 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3070 }
3071 *iclogp = iclog;
3072
3073 ASSERT(iclog->ic_offset <= iclog->ic_size);
3074 spin_unlock(&log->l_icloglock);
3075
3076 *logoffsetp = log_offset;
3077 return 0;
3078 } /* xlog_state_get_iclog_space */
3079
3080 /* The first cnt-1 times through here we don't need to
3081 * move the grant write head because the permanent
3082 * reservation has reserved cnt times the unit amount.
3083 * Release part of current permanent unit reservation and
3084 * reset current reservation to be one units worth. Also
3085 * move grant reservation head forward.
3086 */
3087 STATIC void
3088 xlog_regrant_reserve_log_space(
3089 struct xlog *log,
3090 struct xlog_ticket *ticket)
3091 {
3092 trace_xfs_log_regrant_reserve_enter(log, ticket);
3093
3094 if (ticket->t_cnt > 0)
3095 ticket->t_cnt--;
3096
3097 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3098 ticket->t_curr_res);
3099 xlog_grant_sub_space(log, &log->l_write_head.grant,
3100 ticket->t_curr_res);
3101 ticket->t_curr_res = ticket->t_unit_res;
3102 xlog_tic_reset_res(ticket);
3103
3104 trace_xfs_log_regrant_reserve_sub(log, ticket);
3105
3106 /* just return if we still have some of the pre-reserved space */
3107 if (ticket->t_cnt > 0)
3108 return;
3109
3110 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3111 ticket->t_unit_res);
3112
3113 trace_xfs_log_regrant_reserve_exit(log, ticket);
3114
3115 ticket->t_curr_res = ticket->t_unit_res;
3116 xlog_tic_reset_res(ticket);
3117 } /* xlog_regrant_reserve_log_space */
3118
3119
3120 /*
3121 * Give back the space left from a reservation.
3122 *
3123 * All the information we need to make a correct determination of space left
3124 * is present. For non-permanent reservations, things are quite easy. The
3125 * count should have been decremented to zero. We only need to deal with the
3126 * space remaining in the current reservation part of the ticket. If the
3127 * ticket contains a permanent reservation, there may be left over space which
3128 * needs to be released. A count of N means that N-1 refills of the current
3129 * reservation can be done before we need to ask for more space. The first
3130 * one goes to fill up the first current reservation. Once we run out of
3131 * space, the count will stay at zero and the only space remaining will be
3132 * in the current reservation field.
3133 */
3134 STATIC void
3135 xlog_ungrant_log_space(
3136 struct xlog *log,
3137 struct xlog_ticket *ticket)
3138 {
3139 int bytes;
3140
3141 if (ticket->t_cnt > 0)
3142 ticket->t_cnt--;
3143
3144 trace_xfs_log_ungrant_enter(log, ticket);
3145 trace_xfs_log_ungrant_sub(log, ticket);
3146
3147 /*
3148 * If this is a permanent reservation ticket, we may be able to free
3149 * up more space based on the remaining count.
3150 */
3151 bytes = ticket->t_curr_res;
3152 if (ticket->t_cnt > 0) {
3153 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3154 bytes += ticket->t_unit_res*ticket->t_cnt;
3155 }
3156
3157 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3158 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3159
3160 trace_xfs_log_ungrant_exit(log, ticket);
3161
3162 xfs_log_space_wake(log->l_mp);
3163 }
3164
3165 /*
3166 * Flush iclog to disk if this is the last reference to the given iclog and
3167 * the WANT_SYNC bit is set.
3168 *
3169 * When this function is entered, the iclog is not necessarily in the
3170 * WANT_SYNC state. It may be sitting around waiting to get filled.
3171 *
3172 *
3173 */
3174 STATIC int
3175 xlog_state_release_iclog(
3176 struct xlog *log,
3177 struct xlog_in_core *iclog)
3178 {
3179 int sync = 0; /* do we sync? */
3180
3181 if (iclog->ic_state & XLOG_STATE_IOERROR)
3182 return -EIO;
3183
3184 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3185 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3186 return 0;
3187
3188 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3189 spin_unlock(&log->l_icloglock);
3190 return -EIO;
3191 }
3192 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3193 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3194
3195 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3196 /* update tail before writing to iclog */
3197 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3198 sync++;
3199 iclog->ic_state = XLOG_STATE_SYNCING;
3200 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3201 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3202 /* cycle incremented when incrementing curr_block */
3203 }
3204 spin_unlock(&log->l_icloglock);
3205
3206 /*
3207 * We let the log lock go, so it's possible that we hit a log I/O
3208 * error or some other SHUTDOWN condition that marks the iclog
3209 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3210 * this iclog has consistent data, so we ignore IOERROR
3211 * flags after this point.
3212 */
3213 if (sync)
3214 xlog_sync(log, iclog);
3215 return 0;
3216 } /* xlog_state_release_iclog */
3217
3218
3219 /*
3220 * This routine will mark the current iclog in the ring as WANT_SYNC
3221 * and move the current iclog pointer to the next iclog in the ring.
3222 * When this routine is called from xlog_state_get_iclog_space(), the
3223 * exact size of the iclog has not yet been determined. All we know is
3224 * that every data block. We have run out of space in this log record.
3225 */
3226 STATIC void
3227 xlog_state_switch_iclogs(
3228 struct xlog *log,
3229 struct xlog_in_core *iclog,
3230 int eventual_size)
3231 {
3232 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3233 if (!eventual_size)
3234 eventual_size = iclog->ic_offset;
3235 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3236 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3237 log->l_prev_block = log->l_curr_block;
3238 log->l_prev_cycle = log->l_curr_cycle;
3239
3240 /* roll log?: ic_offset changed later */
3241 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3242
3243 /* Round up to next log-sunit */
3244 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3245 log->l_mp->m_sb.sb_logsunit > 1) {
3246 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3247 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3248 }
3249
3250 if (log->l_curr_block >= log->l_logBBsize) {
3251 /*
3252 * Rewind the current block before the cycle is bumped to make
3253 * sure that the combined LSN never transiently moves forward
3254 * when the log wraps to the next cycle. This is to support the
3255 * unlocked sample of these fields from xlog_valid_lsn(). Most
3256 * other cases should acquire l_icloglock.
3257 */
3258 log->l_curr_block -= log->l_logBBsize;
3259 ASSERT(log->l_curr_block >= 0);
3260 smp_wmb();
3261 log->l_curr_cycle++;
3262 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3263 log->l_curr_cycle++;
3264 }
3265 ASSERT(iclog == log->l_iclog);
3266 log->l_iclog = iclog->ic_next;
3267 } /* xlog_state_switch_iclogs */
3268
3269 /*
3270 * Write out all data in the in-core log as of this exact moment in time.
3271 *
3272 * Data may be written to the in-core log during this call. However,
3273 * we don't guarantee this data will be written out. A change from past
3274 * implementation means this routine will *not* write out zero length LRs.
3275 *
3276 * Basically, we try and perform an intelligent scan of the in-core logs.
3277 * If we determine there is no flushable data, we just return. There is no
3278 * flushable data if:
3279 *
3280 * 1. the current iclog is active and has no data; the previous iclog
3281 * is in the active or dirty state.
3282 * 2. the current iclog is drity, and the previous iclog is in the
3283 * active or dirty state.
3284 *
3285 * We may sleep if:
3286 *
3287 * 1. the current iclog is not in the active nor dirty state.
3288 * 2. the current iclog dirty, and the previous iclog is not in the
3289 * active nor dirty state.
3290 * 3. the current iclog is active, and there is another thread writing
3291 * to this particular iclog.
3292 * 4. a) the current iclog is active and has no other writers
3293 * b) when we return from flushing out this iclog, it is still
3294 * not in the active nor dirty state.
3295 */
3296 int
3297 xfs_log_force(
3298 struct xfs_mount *mp,
3299 uint flags)
3300 {
3301 struct xlog *log = mp->m_log;
3302 struct xlog_in_core *iclog;
3303 xfs_lsn_t lsn;
3304
3305 XFS_STATS_INC(mp, xs_log_force);
3306 trace_xfs_log_force(mp, 0, _RET_IP_);
3307
3308 xlog_cil_force(log);
3309
3310 spin_lock(&log->l_icloglock);
3311 iclog = log->l_iclog;
3312 if (iclog->ic_state & XLOG_STATE_IOERROR)
3313 goto out_error;
3314
3315 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3316 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3317 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3318 /*
3319 * If the head is dirty or (active and empty), then we need to
3320 * look at the previous iclog.
3321 *
3322 * If the previous iclog is active or dirty we are done. There
3323 * is nothing to sync out. Otherwise, we attach ourselves to the
3324 * previous iclog and go to sleep.
3325 */
3326 iclog = iclog->ic_prev;
3327 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3328 iclog->ic_state == XLOG_STATE_DIRTY)
3329 goto out_unlock;
3330 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3331 if (atomic_read(&iclog->ic_refcnt) == 0) {
3332 /*
3333 * We are the only one with access to this iclog.
3334 *
3335 * Flush it out now. There should be a roundoff of zero
3336 * to show that someone has already taken care of the
3337 * roundoff from the previous sync.
3338 */
3339 atomic_inc(&iclog->ic_refcnt);
3340 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3341 xlog_state_switch_iclogs(log, iclog, 0);
3342 spin_unlock(&log->l_icloglock);
3343
3344 if (xlog_state_release_iclog(log, iclog))
3345 return -EIO;
3346
3347 spin_lock(&log->l_icloglock);
3348 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3349 iclog->ic_state == XLOG_STATE_DIRTY)
3350 goto out_unlock;
3351 } else {
3352 /*
3353 * Someone else is writing to this iclog.
3354 *
3355 * Use its call to flush out the data. However, the
3356 * other thread may not force out this LR, so we mark
3357 * it WANT_SYNC.
3358 */
3359 xlog_state_switch_iclogs(log, iclog, 0);
3360 }
3361 } else {
3362 /*
3363 * If the head iclog is not active nor dirty, we just attach
3364 * ourselves to the head and go to sleep if necessary.
3365 */
3366 ;
3367 }
3368
3369 if (!(flags & XFS_LOG_SYNC))
3370 goto out_unlock;
3371
3372 if (iclog->ic_state & XLOG_STATE_IOERROR)
3373 goto out_error;
3374 XFS_STATS_INC(mp, xs_log_force_sleep);
3375 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3376 if (iclog->ic_state & XLOG_STATE_IOERROR)
3377 return -EIO;
3378 return 0;
3379
3380 out_unlock:
3381 spin_unlock(&log->l_icloglock);
3382 return 0;
3383 out_error:
3384 spin_unlock(&log->l_icloglock);
3385 return -EIO;
3386 }
3387
3388 static int
3389 __xfs_log_force_lsn(
3390 struct xfs_mount *mp,
3391 xfs_lsn_t lsn,
3392 uint flags,
3393 int *log_flushed,
3394 bool already_slept)
3395 {
3396 struct xlog *log = mp->m_log;
3397 struct xlog_in_core *iclog;
3398
3399 spin_lock(&log->l_icloglock);
3400 iclog = log->l_iclog;
3401 if (iclog->ic_state & XLOG_STATE_IOERROR)
3402 goto out_error;
3403
3404 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3405 iclog = iclog->ic_next;
3406 if (iclog == log->l_iclog)
3407 goto out_unlock;
3408 }
3409
3410 if (iclog->ic_state == XLOG_STATE_DIRTY)
3411 goto out_unlock;
3412
3413 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3414 /*
3415 * We sleep here if we haven't already slept (e.g. this is the
3416 * first time we've looked at the correct iclog buf) and the
3417 * buffer before us is going to be sync'ed. The reason for this
3418 * is that if we are doing sync transactions here, by waiting
3419 * for the previous I/O to complete, we can allow a few more
3420 * transactions into this iclog before we close it down.
3421 *
3422 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3423 * refcnt so we can release the log (which drops the ref count).
3424 * The state switch keeps new transaction commits from using
3425 * this buffer. When the current commits finish writing into
3426 * the buffer, the refcount will drop to zero and the buffer
3427 * will go out then.
3428 */
3429 if (!already_slept &&
3430 (iclog->ic_prev->ic_state &
3431 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3432 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3433
3434 XFS_STATS_INC(mp, xs_log_force_sleep);
3435
3436 xlog_wait(&iclog->ic_prev->ic_write_wait,
3437 &log->l_icloglock);
3438 return -EAGAIN;
3439 }
3440 atomic_inc(&iclog->ic_refcnt);
3441 xlog_state_switch_iclogs(log, iclog, 0);
3442 spin_unlock(&log->l_icloglock);
3443 if (xlog_state_release_iclog(log, iclog))
3444 return -EIO;
3445 if (log_flushed)
3446 *log_flushed = 1;
3447 spin_lock(&log->l_icloglock);
3448 }
3449
3450 if (!(flags & XFS_LOG_SYNC) ||
3451 (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3452 goto out_unlock;
3453
3454 if (iclog->ic_state & XLOG_STATE_IOERROR)
3455 goto out_error;
3456
3457 XFS_STATS_INC(mp, xs_log_force_sleep);
3458 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3459 if (iclog->ic_state & XLOG_STATE_IOERROR)
3460 return -EIO;
3461 return 0;
3462
3463 out_unlock:
3464 spin_unlock(&log->l_icloglock);
3465 return 0;
3466 out_error:
3467 spin_unlock(&log->l_icloglock);
3468 return -EIO;
3469 }
3470
3471 /*
3472 * Force the in-core log to disk for a specific LSN.
3473 *
3474 * Find in-core log with lsn.
3475 * If it is in the DIRTY state, just return.
3476 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3477 * state and go to sleep or return.
3478 * If it is in any other state, go to sleep or return.
3479 *
3480 * Synchronous forces are implemented with a wait queue. All callers trying
3481 * to force a given lsn to disk must wait on the queue attached to the
3482 * specific in-core log. When given in-core log finally completes its write
3483 * to disk, that thread will wake up all threads waiting on the queue.
3484 */
3485 int
3486 xfs_log_force_lsn(
3487 struct xfs_mount *mp,
3488 xfs_lsn_t lsn,
3489 uint flags,
3490 int *log_flushed)
3491 {
3492 int ret;
3493 ASSERT(lsn != 0);
3494
3495 XFS_STATS_INC(mp, xs_log_force);
3496 trace_xfs_log_force(mp, lsn, _RET_IP_);
3497
3498 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3499 if (lsn == NULLCOMMITLSN)
3500 return 0;
3501
3502 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3503 if (ret == -EAGAIN)
3504 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3505 return ret;
3506 }
3507
3508 /*
3509 * Called when we want to mark the current iclog as being ready to sync to
3510 * disk.
3511 */
3512 STATIC void
3513 xlog_state_want_sync(
3514 struct xlog *log,
3515 struct xlog_in_core *iclog)
3516 {
3517 assert_spin_locked(&log->l_icloglock);
3518
3519 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3520 xlog_state_switch_iclogs(log, iclog, 0);
3521 } else {
3522 ASSERT(iclog->ic_state &
3523 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3524 }
3525 }
3526
3527
3528 /*****************************************************************************
3529 *
3530 * TICKET functions
3531 *
3532 *****************************************************************************
3533 */
3534
3535 /*
3536 * Free a used ticket when its refcount falls to zero.
3537 */
3538 void
3539 xfs_log_ticket_put(
3540 xlog_ticket_t *ticket)
3541 {
3542 ASSERT(atomic_read(&ticket->t_ref) > 0);
3543 if (atomic_dec_and_test(&ticket->t_ref))
3544 kmem_zone_free(xfs_log_ticket_zone, ticket);
3545 }
3546
3547 xlog_ticket_t *
3548 xfs_log_ticket_get(
3549 xlog_ticket_t *ticket)
3550 {
3551 ASSERT(atomic_read(&ticket->t_ref) > 0);
3552 atomic_inc(&ticket->t_ref);
3553 return ticket;
3554 }
3555
3556 /*
3557 * Figure out the total log space unit (in bytes) that would be
3558 * required for a log ticket.
3559 */
3560 int
3561 xfs_log_calc_unit_res(
3562 struct xfs_mount *mp,
3563 int unit_bytes)
3564 {
3565 struct xlog *log = mp->m_log;
3566 int iclog_space;
3567 uint num_headers;
3568
3569 /*
3570 * Permanent reservations have up to 'cnt'-1 active log operations
3571 * in the log. A unit in this case is the amount of space for one
3572 * of these log operations. Normal reservations have a cnt of 1
3573 * and their unit amount is the total amount of space required.
3574 *
3575 * The following lines of code account for non-transaction data
3576 * which occupy space in the on-disk log.
3577 *
3578 * Normal form of a transaction is:
3579 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3580 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3581 *
3582 * We need to account for all the leadup data and trailer data
3583 * around the transaction data.
3584 * And then we need to account for the worst case in terms of using
3585 * more space.
3586 * The worst case will happen if:
3587 * - the placement of the transaction happens to be such that the
3588 * roundoff is at its maximum
3589 * - the transaction data is synced before the commit record is synced
3590 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3591 * Therefore the commit record is in its own Log Record.
3592 * This can happen as the commit record is called with its
3593 * own region to xlog_write().
3594 * This then means that in the worst case, roundoff can happen for
3595 * the commit-rec as well.
3596 * The commit-rec is smaller than padding in this scenario and so it is
3597 * not added separately.
3598 */
3599
3600 /* for trans header */
3601 unit_bytes += sizeof(xlog_op_header_t);
3602 unit_bytes += sizeof(xfs_trans_header_t);
3603
3604 /* for start-rec */
3605 unit_bytes += sizeof(xlog_op_header_t);
3606
3607 /*
3608 * for LR headers - the space for data in an iclog is the size minus
3609 * the space used for the headers. If we use the iclog size, then we
3610 * undercalculate the number of headers required.
3611 *
3612 * Furthermore - the addition of op headers for split-recs might
3613 * increase the space required enough to require more log and op
3614 * headers, so take that into account too.
3615 *
3616 * IMPORTANT: This reservation makes the assumption that if this
3617 * transaction is the first in an iclog and hence has the LR headers
3618 * accounted to it, then the remaining space in the iclog is
3619 * exclusively for this transaction. i.e. if the transaction is larger
3620 * than the iclog, it will be the only thing in that iclog.
3621 * Fundamentally, this means we must pass the entire log vector to
3622 * xlog_write to guarantee this.
3623 */
3624 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3625 num_headers = howmany(unit_bytes, iclog_space);
3626
3627 /* for split-recs - ophdrs added when data split over LRs */
3628 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3629
3630 /* add extra header reservations if we overrun */
3631 while (!num_headers ||
3632 howmany(unit_bytes, iclog_space) > num_headers) {
3633 unit_bytes += sizeof(xlog_op_header_t);
3634 num_headers++;
3635 }
3636 unit_bytes += log->l_iclog_hsize * num_headers;
3637
3638 /* for commit-rec LR header - note: padding will subsume the ophdr */
3639 unit_bytes += log->l_iclog_hsize;
3640
3641 /* for roundoff padding for transaction data and one for commit record */
3642 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3643 /* log su roundoff */
3644 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3645 } else {
3646 /* BB roundoff */
3647 unit_bytes += 2 * BBSIZE;
3648 }
3649
3650 return unit_bytes;
3651 }
3652
3653 /*
3654 * Allocate and initialise a new log ticket.
3655 */
3656 struct xlog_ticket *
3657 xlog_ticket_alloc(
3658 struct xlog *log,
3659 int unit_bytes,
3660 int cnt,
3661 char client,
3662 bool permanent,
3663 xfs_km_flags_t alloc_flags)
3664 {
3665 struct xlog_ticket *tic;
3666 int unit_res;
3667
3668 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3669 if (!tic)
3670 return NULL;
3671
3672 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3673
3674 atomic_set(&tic->t_ref, 1);
3675 tic->t_task = current;
3676 INIT_LIST_HEAD(&tic->t_queue);
3677 tic->t_unit_res = unit_res;
3678 tic->t_curr_res = unit_res;
3679 tic->t_cnt = cnt;
3680 tic->t_ocnt = cnt;
3681 tic->t_tid = prandom_u32();
3682 tic->t_clientid = client;
3683 tic->t_flags = XLOG_TIC_INITED;
3684 if (permanent)
3685 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3686
3687 xlog_tic_reset_res(tic);
3688
3689 return tic;
3690 }
3691
3692
3693 /******************************************************************************
3694 *
3695 * Log debug routines
3696 *
3697 ******************************************************************************
3698 */
3699 #if defined(DEBUG)
3700 /*
3701 * Make sure that the destination ptr is within the valid data region of
3702 * one of the iclogs. This uses backup pointers stored in a different
3703 * part of the log in case we trash the log structure.
3704 */
3705 STATIC void
3706 xlog_verify_dest_ptr(
3707 struct xlog *log,
3708 void *ptr)
3709 {
3710 int i;
3711 int good_ptr = 0;
3712
3713 for (i = 0; i < log->l_iclog_bufs; i++) {
3714 if (ptr >= log->l_iclog_bak[i] &&
3715 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3716 good_ptr++;
3717 }
3718
3719 if (!good_ptr)
3720 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3721 }
3722
3723 /*
3724 * Check to make sure the grant write head didn't just over lap the tail. If
3725 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3726 * the cycles differ by exactly one and check the byte count.
3727 *
3728 * This check is run unlocked, so can give false positives. Rather than assert
3729 * on failures, use a warn-once flag and a panic tag to allow the admin to
3730 * determine if they want to panic the machine when such an error occurs. For
3731 * debug kernels this will have the same effect as using an assert but, unlinke
3732 * an assert, it can be turned off at runtime.
3733 */
3734 STATIC void
3735 xlog_verify_grant_tail(
3736 struct xlog *log)
3737 {
3738 int tail_cycle, tail_blocks;
3739 int cycle, space;
3740
3741 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3742 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3743 if (tail_cycle != cycle) {
3744 if (cycle - 1 != tail_cycle &&
3745 !(log->l_flags & XLOG_TAIL_WARN)) {
3746 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3747 "%s: cycle - 1 != tail_cycle", __func__);
3748 log->l_flags |= XLOG_TAIL_WARN;
3749 }
3750
3751 if (space > BBTOB(tail_blocks) &&
3752 !(log->l_flags & XLOG_TAIL_WARN)) {
3753 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3754 "%s: space > BBTOB(tail_blocks)", __func__);
3755 log->l_flags |= XLOG_TAIL_WARN;
3756 }
3757 }
3758 }
3759
3760 /* check if it will fit */
3761 STATIC void
3762 xlog_verify_tail_lsn(
3763 struct xlog *log,
3764 struct xlog_in_core *iclog,
3765 xfs_lsn_t tail_lsn)
3766 {
3767 int blocks;
3768
3769 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3770 blocks =
3771 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3772 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3773 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3774 } else {
3775 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3776
3777 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3778 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3779
3780 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3781 if (blocks < BTOBB(iclog->ic_offset) + 1)
3782 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3783 }
3784 } /* xlog_verify_tail_lsn */
3785
3786 /*
3787 * Perform a number of checks on the iclog before writing to disk.
3788 *
3789 * 1. Make sure the iclogs are still circular
3790 * 2. Make sure we have a good magic number
3791 * 3. Make sure we don't have magic numbers in the data
3792 * 4. Check fields of each log operation header for:
3793 * A. Valid client identifier
3794 * B. tid ptr value falls in valid ptr space (user space code)
3795 * C. Length in log record header is correct according to the
3796 * individual operation headers within record.
3797 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3798 * log, check the preceding blocks of the physical log to make sure all
3799 * the cycle numbers agree with the current cycle number.
3800 */
3801 STATIC void
3802 xlog_verify_iclog(
3803 struct xlog *log,
3804 struct xlog_in_core *iclog,
3805 int count)
3806 {
3807 xlog_op_header_t *ophead;
3808 xlog_in_core_t *icptr;
3809 xlog_in_core_2_t *xhdr;
3810 void *base_ptr, *ptr, *p;
3811 ptrdiff_t field_offset;
3812 uint8_t clientid;
3813 int len, i, j, k, op_len;
3814 int idx;
3815
3816 /* check validity of iclog pointers */
3817 spin_lock(&log->l_icloglock);
3818 icptr = log->l_iclog;
3819 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3820 ASSERT(icptr);
3821
3822 if (icptr != log->l_iclog)
3823 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3824 spin_unlock(&log->l_icloglock);
3825
3826 /* check log magic numbers */
3827 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3828 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3829
3830 base_ptr = ptr = &iclog->ic_header;
3831 p = &iclog->ic_header;
3832 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3833 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3834 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3835 __func__);
3836 }
3837
3838 /* check fields */
3839 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3840 base_ptr = ptr = iclog->ic_datap;
3841 ophead = ptr;
3842 xhdr = iclog->ic_data;
3843 for (i = 0; i < len; i++) {
3844 ophead = ptr;
3845
3846 /* clientid is only 1 byte */
3847 p = &ophead->oh_clientid;
3848 field_offset = p - base_ptr;
3849 if (field_offset & 0x1ff) {
3850 clientid = ophead->oh_clientid;
3851 } else {
3852 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3853 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3854 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3855 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3856 clientid = xlog_get_client_id(
3857 xhdr[j].hic_xheader.xh_cycle_data[k]);
3858 } else {
3859 clientid = xlog_get_client_id(
3860 iclog->ic_header.h_cycle_data[idx]);
3861 }
3862 }
3863 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3864 xfs_warn(log->l_mp,
3865 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3866 __func__, clientid, ophead,
3867 (unsigned long)field_offset);
3868
3869 /* check length */
3870 p = &ophead->oh_len;
3871 field_offset = p - base_ptr;
3872 if (field_offset & 0x1ff) {
3873 op_len = be32_to_cpu(ophead->oh_len);
3874 } else {
3875 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3876 (uintptr_t)iclog->ic_datap);
3877 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3878 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3879 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3880 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3881 } else {
3882 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3883 }
3884 }
3885 ptr += sizeof(xlog_op_header_t) + op_len;
3886 }
3887 } /* xlog_verify_iclog */
3888 #endif
3889
3890 /*
3891 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3892 */
3893 STATIC int
3894 xlog_state_ioerror(
3895 struct xlog *log)
3896 {
3897 xlog_in_core_t *iclog, *ic;
3898
3899 iclog = log->l_iclog;
3900 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3901 /*
3902 * Mark all the incore logs IOERROR.
3903 * From now on, no log flushes will result.
3904 */
3905 ic = iclog;
3906 do {
3907 ic->ic_state = XLOG_STATE_IOERROR;
3908 ic = ic->ic_next;
3909 } while (ic != iclog);
3910 return 0;
3911 }
3912 /*
3913 * Return non-zero, if state transition has already happened.
3914 */
3915 return 1;
3916 }
3917
3918 /*
3919 * This is called from xfs_force_shutdown, when we're forcibly
3920 * shutting down the filesystem, typically because of an IO error.
3921 * Our main objectives here are to make sure that:
3922 * a. if !logerror, flush the logs to disk. Anything modified
3923 * after this is ignored.
3924 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3925 * parties to find out, 'atomically'.
3926 * c. those who're sleeping on log reservations, pinned objects and
3927 * other resources get woken up, and be told the bad news.
3928 * d. nothing new gets queued up after (b) and (c) are done.
3929 *
3930 * Note: for the !logerror case we need to flush the regions held in memory out
3931 * to disk first. This needs to be done before the log is marked as shutdown,
3932 * otherwise the iclog writes will fail.
3933 */
3934 int
3935 xfs_log_force_umount(
3936 struct xfs_mount *mp,
3937 int logerror)
3938 {
3939 struct xlog *log;
3940 int retval;
3941
3942 log = mp->m_log;
3943
3944 /*
3945 * If this happens during log recovery, don't worry about
3946 * locking; the log isn't open for business yet.
3947 */
3948 if (!log ||
3949 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3950 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3951 if (mp->m_sb_bp)
3952 mp->m_sb_bp->b_flags |= XBF_DONE;
3953 return 0;
3954 }
3955
3956 /*
3957 * Somebody could've already done the hard work for us.
3958 * No need to get locks for this.
3959 */
3960 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3961 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3962 return 1;
3963 }
3964
3965 /*
3966 * Flush all the completed transactions to disk before marking the log
3967 * being shut down. We need to do it in this order to ensure that
3968 * completed operations are safely on disk before we shut down, and that
3969 * we don't have to issue any buffer IO after the shutdown flags are set
3970 * to guarantee this.
3971 */
3972 if (!logerror)
3973 xfs_log_force(mp, XFS_LOG_SYNC);
3974
3975 /*
3976 * mark the filesystem and the as in a shutdown state and wake
3977 * everybody up to tell them the bad news.
3978 */
3979 spin_lock(&log->l_icloglock);
3980 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3981 if (mp->m_sb_bp)
3982 mp->m_sb_bp->b_flags |= XBF_DONE;
3983
3984 /*
3985 * Mark the log and the iclogs with IO error flags to prevent any
3986 * further log IO from being issued or completed.
3987 */
3988 log->l_flags |= XLOG_IO_ERROR;
3989 retval = xlog_state_ioerror(log);
3990 spin_unlock(&log->l_icloglock);
3991
3992 /*
3993 * We don't want anybody waiting for log reservations after this. That
3994 * means we have to wake up everybody queued up on reserveq as well as
3995 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3996 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3997 * action is protected by the grant locks.
3998 */
3999 xlog_grant_head_wake_all(&log->l_reserve_head);
4000 xlog_grant_head_wake_all(&log->l_write_head);
4001
4002 /*
4003 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4004 * as if the log writes were completed. The abort handling in the log
4005 * item committed callback functions will do this again under lock to
4006 * avoid races.
4007 */
4008 spin_lock(&log->l_cilp->xc_push_lock);
4009 wake_up_all(&log->l_cilp->xc_commit_wait);
4010 spin_unlock(&log->l_cilp->xc_push_lock);
4011 xlog_state_do_callback(log, true, NULL);
4012
4013 #ifdef XFSERRORDEBUG
4014 {
4015 xlog_in_core_t *iclog;
4016
4017 spin_lock(&log->l_icloglock);
4018 iclog = log->l_iclog;
4019 do {
4020 ASSERT(iclog->ic_callback == 0);
4021 iclog = iclog->ic_next;
4022 } while (iclog != log->l_iclog);
4023 spin_unlock(&log->l_icloglock);
4024 }
4025 #endif
4026 /* return non-zero if log IOERROR transition had already happened */
4027 return retval;
4028 }
4029
4030 STATIC int
4031 xlog_iclogs_empty(
4032 struct xlog *log)
4033 {
4034 xlog_in_core_t *iclog;
4035
4036 iclog = log->l_iclog;
4037 do {
4038 /* endianness does not matter here, zero is zero in
4039 * any language.
4040 */
4041 if (iclog->ic_header.h_num_logops)
4042 return 0;
4043 iclog = iclog->ic_next;
4044 } while (iclog != log->l_iclog);
4045 return 1;
4046 }
4047
4048 /*
4049 * Verify that an LSN stamped into a piece of metadata is valid. This is
4050 * intended for use in read verifiers on v5 superblocks.
4051 */
4052 bool
4053 xfs_log_check_lsn(
4054 struct xfs_mount *mp,
4055 xfs_lsn_t lsn)
4056 {
4057 struct xlog *log = mp->m_log;
4058 bool valid;
4059
4060 /*
4061 * norecovery mode skips mount-time log processing and unconditionally
4062 * resets the in-core LSN. We can't validate in this mode, but
4063 * modifications are not allowed anyways so just return true.
4064 */
4065 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4066 return true;
4067
4068 /*
4069 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4070 * handled by recovery and thus safe to ignore here.
4071 */
4072 if (lsn == NULLCOMMITLSN)
4073 return true;
4074
4075 valid = xlog_valid_lsn(mp->m_log, lsn);
4076
4077 /* warn the user about what's gone wrong before verifier failure */
4078 if (!valid) {
4079 spin_lock(&log->l_icloglock);
4080 xfs_warn(mp,
4081 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4082 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4083 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4084 log->l_curr_cycle, log->l_curr_block);
4085 spin_unlock(&log->l_icloglock);
4086 }
4087
4088 return valid;
4089 }
4090
4091 bool
4092 xfs_log_in_recovery(
4093 struct xfs_mount *mp)
4094 {
4095 struct xlog *log = mp->m_log;
4096
4097 return log->l_flags & XLOG_ACTIVE_RECOVERY;
4098 }