Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_errortag.h"
14#include "xfs_error.h"
15#include "xfs_trans.h"
16#include "xfs_trans_priv.h"
17#include "xfs_log.h"
18#include "xfs_log_priv.h"
19#include "xfs_trace.h"
20#include "xfs_sysfs.h"
21#include "xfs_sb.h"
22#include "xfs_health.h"
23
24struct kmem_cache *xfs_log_ticket_cache;
25
26/* Local miscellaneous function prototypes */
27STATIC struct xlog *
28xlog_alloc_log(
29 struct xfs_mount *mp,
30 struct xfs_buftarg *log_target,
31 xfs_daddr_t blk_offset,
32 int num_bblks);
33STATIC void
34xlog_dealloc_log(
35 struct xlog *log);
36
37/* local state machine functions */
38STATIC void xlog_state_done_syncing(
39 struct xlog_in_core *iclog);
40STATIC void xlog_state_do_callback(
41 struct xlog *log);
42STATIC int
43xlog_state_get_iclog_space(
44 struct xlog *log,
45 int len,
46 struct xlog_in_core **iclog,
47 struct xlog_ticket *ticket,
48 int *logoffsetp);
49STATIC void
50xlog_sync(
51 struct xlog *log,
52 struct xlog_in_core *iclog,
53 struct xlog_ticket *ticket);
54#if defined(DEBUG)
55STATIC void
56xlog_verify_iclog(
57 struct xlog *log,
58 struct xlog_in_core *iclog,
59 int count);
60STATIC void
61xlog_verify_tail_lsn(
62 struct xlog *log,
63 struct xlog_in_core *iclog);
64#else
65#define xlog_verify_iclog(a,b,c)
66#define xlog_verify_tail_lsn(a,b)
67#endif
68
69STATIC int
70xlog_iclogs_empty(
71 struct xlog *log);
72
73static int
74xfs_log_cover(struct xfs_mount *);
75
76/*
77 * We need to make sure the buffer pointer returned is naturally aligned for the
78 * biggest basic data type we put into it. We have already accounted for this
79 * padding when sizing the buffer.
80 *
81 * However, this padding does not get written into the log, and hence we have to
82 * track the space used by the log vectors separately to prevent log space hangs
83 * due to inaccurate accounting (i.e. a leak) of the used log space through the
84 * CIL context ticket.
85 *
86 * We also add space for the xlog_op_header that describes this region in the
87 * log. This prepends the data region we return to the caller to copy their data
88 * into, so do all the static initialisation of the ophdr now. Because the ophdr
89 * is not 8 byte aligned, we have to be careful to ensure that we align the
90 * start of the buffer such that the region we return to the call is 8 byte
91 * aligned and packed against the tail of the ophdr.
92 */
93void *
94xlog_prepare_iovec(
95 struct xfs_log_vec *lv,
96 struct xfs_log_iovec **vecp,
97 uint type)
98{
99 struct xfs_log_iovec *vec = *vecp;
100 struct xlog_op_header *oph;
101 uint32_t len;
102 void *buf;
103
104 if (vec) {
105 ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
106 vec++;
107 } else {
108 vec = &lv->lv_iovecp[0];
109 }
110
111 len = lv->lv_buf_len + sizeof(struct xlog_op_header);
112 if (!IS_ALIGNED(len, sizeof(uint64_t))) {
113 lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
114 sizeof(struct xlog_op_header);
115 }
116
117 vec->i_type = type;
118 vec->i_addr = lv->lv_buf + lv->lv_buf_len;
119
120 oph = vec->i_addr;
121 oph->oh_clientid = XFS_TRANSACTION;
122 oph->oh_res2 = 0;
123 oph->oh_flags = 0;
124
125 buf = vec->i_addr + sizeof(struct xlog_op_header);
126 ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
127
128 *vecp = vec;
129 return buf;
130}
131
132static inline void
133xlog_grant_sub_space(
134 struct xlog_grant_head *head,
135 int64_t bytes)
136{
137 atomic64_sub(bytes, &head->grant);
138}
139
140static inline void
141xlog_grant_add_space(
142 struct xlog_grant_head *head,
143 int64_t bytes)
144{
145 atomic64_add(bytes, &head->grant);
146}
147
148static void
149xlog_grant_head_init(
150 struct xlog_grant_head *head)
151{
152 atomic64_set(&head->grant, 0);
153 INIT_LIST_HEAD(&head->waiters);
154 spin_lock_init(&head->lock);
155}
156
157void
158xlog_grant_return_space(
159 struct xlog *log,
160 xfs_lsn_t old_head,
161 xfs_lsn_t new_head)
162{
163 int64_t diff = xlog_lsn_sub(log, new_head, old_head);
164
165 xlog_grant_sub_space(&log->l_reserve_head, diff);
166 xlog_grant_sub_space(&log->l_write_head, diff);
167}
168
169/*
170 * Return the space in the log between the tail and the head. In the case where
171 * we have overrun available reservation space, return 0. The memory barrier
172 * pairs with the smp_wmb() in xlog_cil_ail_insert() to ensure that grant head
173 * vs tail space updates are seen in the correct order and hence avoid
174 * transients as space is transferred from the grant heads to the AIL on commit
175 * completion.
176 */
177static uint64_t
178xlog_grant_space_left(
179 struct xlog *log,
180 struct xlog_grant_head *head)
181{
182 int64_t free_bytes;
183
184 smp_rmb(); /* paired with smp_wmb in xlog_cil_ail_insert() */
185 free_bytes = log->l_logsize - READ_ONCE(log->l_tail_space) -
186 atomic64_read(&head->grant);
187 if (free_bytes > 0)
188 return free_bytes;
189 return 0;
190}
191
192STATIC void
193xlog_grant_head_wake_all(
194 struct xlog_grant_head *head)
195{
196 struct xlog_ticket *tic;
197
198 spin_lock(&head->lock);
199 list_for_each_entry(tic, &head->waiters, t_queue)
200 wake_up_process(tic->t_task);
201 spin_unlock(&head->lock);
202}
203
204static inline int
205xlog_ticket_reservation(
206 struct xlog *log,
207 struct xlog_grant_head *head,
208 struct xlog_ticket *tic)
209{
210 if (head == &log->l_write_head) {
211 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
212 return tic->t_unit_res;
213 }
214
215 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
216 return tic->t_unit_res * tic->t_cnt;
217
218 return tic->t_unit_res;
219}
220
221STATIC bool
222xlog_grant_head_wake(
223 struct xlog *log,
224 struct xlog_grant_head *head,
225 int *free_bytes)
226{
227 struct xlog_ticket *tic;
228 int need_bytes;
229
230 list_for_each_entry(tic, &head->waiters, t_queue) {
231 need_bytes = xlog_ticket_reservation(log, head, tic);
232 if (*free_bytes < need_bytes)
233 return false;
234
235 *free_bytes -= need_bytes;
236 trace_xfs_log_grant_wake_up(log, tic);
237 wake_up_process(tic->t_task);
238 }
239
240 return true;
241}
242
243STATIC int
244xlog_grant_head_wait(
245 struct xlog *log,
246 struct xlog_grant_head *head,
247 struct xlog_ticket *tic,
248 int need_bytes) __releases(&head->lock)
249 __acquires(&head->lock)
250{
251 list_add_tail(&tic->t_queue, &head->waiters);
252
253 do {
254 if (xlog_is_shutdown(log))
255 goto shutdown;
256
257 __set_current_state(TASK_UNINTERRUPTIBLE);
258 spin_unlock(&head->lock);
259
260 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
261
262 /* Push on the AIL to free up all the log space. */
263 xfs_ail_push_all(log->l_ailp);
264
265 trace_xfs_log_grant_sleep(log, tic);
266 schedule();
267 trace_xfs_log_grant_wake(log, tic);
268
269 spin_lock(&head->lock);
270 if (xlog_is_shutdown(log))
271 goto shutdown;
272 } while (xlog_grant_space_left(log, head) < need_bytes);
273
274 list_del_init(&tic->t_queue);
275 return 0;
276shutdown:
277 list_del_init(&tic->t_queue);
278 return -EIO;
279}
280
281/*
282 * Atomically get the log space required for a log ticket.
283 *
284 * Once a ticket gets put onto head->waiters, it will only return after the
285 * needed reservation is satisfied.
286 *
287 * This function is structured so that it has a lock free fast path. This is
288 * necessary because every new transaction reservation will come through this
289 * path. Hence any lock will be globally hot if we take it unconditionally on
290 * every pass.
291 *
292 * As tickets are only ever moved on and off head->waiters under head->lock, we
293 * only need to take that lock if we are going to add the ticket to the queue
294 * and sleep. We can avoid taking the lock if the ticket was never added to
295 * head->waiters because the t_queue list head will be empty and we hold the
296 * only reference to it so it can safely be checked unlocked.
297 */
298STATIC int
299xlog_grant_head_check(
300 struct xlog *log,
301 struct xlog_grant_head *head,
302 struct xlog_ticket *tic,
303 int *need_bytes)
304{
305 int free_bytes;
306 int error = 0;
307
308 ASSERT(!xlog_in_recovery(log));
309
310 /*
311 * If there are other waiters on the queue then give them a chance at
312 * logspace before us. Wake up the first waiters, if we do not wake
313 * up all the waiters then go to sleep waiting for more free space,
314 * otherwise try to get some space for this transaction.
315 */
316 *need_bytes = xlog_ticket_reservation(log, head, tic);
317 free_bytes = xlog_grant_space_left(log, head);
318 if (!list_empty_careful(&head->waiters)) {
319 spin_lock(&head->lock);
320 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
321 free_bytes < *need_bytes) {
322 error = xlog_grant_head_wait(log, head, tic,
323 *need_bytes);
324 }
325 spin_unlock(&head->lock);
326 } else if (free_bytes < *need_bytes) {
327 spin_lock(&head->lock);
328 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
329 spin_unlock(&head->lock);
330 }
331
332 return error;
333}
334
335bool
336xfs_log_writable(
337 struct xfs_mount *mp)
338{
339 /*
340 * Do not write to the log on norecovery mounts, if the data or log
341 * devices are read-only, or if the filesystem is shutdown. Read-only
342 * mounts allow internal writes for log recovery and unmount purposes,
343 * so don't restrict that case.
344 */
345 if (xfs_has_norecovery(mp))
346 return false;
347 if (xfs_readonly_buftarg(mp->m_ddev_targp))
348 return false;
349 if (xfs_readonly_buftarg(mp->m_log->l_targ))
350 return false;
351 if (xlog_is_shutdown(mp->m_log))
352 return false;
353 return true;
354}
355
356/*
357 * Replenish the byte reservation required by moving the grant write head.
358 */
359int
360xfs_log_regrant(
361 struct xfs_mount *mp,
362 struct xlog_ticket *tic)
363{
364 struct xlog *log = mp->m_log;
365 int need_bytes;
366 int error = 0;
367
368 if (xlog_is_shutdown(log))
369 return -EIO;
370
371 XFS_STATS_INC(mp, xs_try_logspace);
372
373 /*
374 * This is a new transaction on the ticket, so we need to change the
375 * transaction ID so that the next transaction has a different TID in
376 * the log. Just add one to the existing tid so that we can see chains
377 * of rolling transactions in the log easily.
378 */
379 tic->t_tid++;
380 tic->t_curr_res = tic->t_unit_res;
381 if (tic->t_cnt > 0)
382 return 0;
383
384 trace_xfs_log_regrant(log, tic);
385
386 error = xlog_grant_head_check(log, &log->l_write_head, tic,
387 &need_bytes);
388 if (error)
389 goto out_error;
390
391 xlog_grant_add_space(&log->l_write_head, need_bytes);
392 trace_xfs_log_regrant_exit(log, tic);
393 return 0;
394
395out_error:
396 /*
397 * If we are failing, make sure the ticket doesn't have any current
398 * reservations. We don't want to add this back when the ticket/
399 * transaction gets cancelled.
400 */
401 tic->t_curr_res = 0;
402 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
403 return error;
404}
405
406/*
407 * Reserve log space and return a ticket corresponding to the reservation.
408 *
409 * Each reservation is going to reserve extra space for a log record header.
410 * When writes happen to the on-disk log, we don't subtract the length of the
411 * log record header from any reservation. By wasting space in each
412 * reservation, we prevent over allocation problems.
413 */
414int
415xfs_log_reserve(
416 struct xfs_mount *mp,
417 int unit_bytes,
418 int cnt,
419 struct xlog_ticket **ticp,
420 bool permanent)
421{
422 struct xlog *log = mp->m_log;
423 struct xlog_ticket *tic;
424 int need_bytes;
425 int error = 0;
426
427 if (xlog_is_shutdown(log))
428 return -EIO;
429
430 XFS_STATS_INC(mp, xs_try_logspace);
431
432 ASSERT(*ticp == NULL);
433 tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
434 *ticp = tic;
435 trace_xfs_log_reserve(log, tic);
436 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
437 &need_bytes);
438 if (error)
439 goto out_error;
440
441 xlog_grant_add_space(&log->l_reserve_head, need_bytes);
442 xlog_grant_add_space(&log->l_write_head, need_bytes);
443 trace_xfs_log_reserve_exit(log, tic);
444 return 0;
445
446out_error:
447 /*
448 * If we are failing, make sure the ticket doesn't have any current
449 * reservations. We don't want to add this back when the ticket/
450 * transaction gets cancelled.
451 */
452 tic->t_curr_res = 0;
453 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
454 return error;
455}
456
457/*
458 * Run all the pending iclog callbacks and wake log force waiters and iclog
459 * space waiters so they can process the newly set shutdown state. We really
460 * don't care what order we process callbacks here because the log is shut down
461 * and so state cannot change on disk anymore. However, we cannot wake waiters
462 * until the callbacks have been processed because we may be in unmount and
463 * we must ensure that all AIL operations the callbacks perform have completed
464 * before we tear down the AIL.
465 *
466 * We avoid processing actively referenced iclogs so that we don't run callbacks
467 * while the iclog owner might still be preparing the iclog for IO submssion.
468 * These will be caught by xlog_state_iclog_release() and call this function
469 * again to process any callbacks that may have been added to that iclog.
470 */
471static void
472xlog_state_shutdown_callbacks(
473 struct xlog *log)
474{
475 struct xlog_in_core *iclog;
476 LIST_HEAD(cb_list);
477
478 iclog = log->l_iclog;
479 do {
480 if (atomic_read(&iclog->ic_refcnt)) {
481 /* Reference holder will re-run iclog callbacks. */
482 continue;
483 }
484 list_splice_init(&iclog->ic_callbacks, &cb_list);
485 spin_unlock(&log->l_icloglock);
486
487 xlog_cil_process_committed(&cb_list);
488
489 spin_lock(&log->l_icloglock);
490 wake_up_all(&iclog->ic_write_wait);
491 wake_up_all(&iclog->ic_force_wait);
492 } while ((iclog = iclog->ic_next) != log->l_iclog);
493
494 wake_up_all(&log->l_flush_wait);
495}
496
497/*
498 * Flush iclog to disk if this is the last reference to the given iclog and the
499 * it is in the WANT_SYNC state.
500 *
501 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
502 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
503 * written to stable storage, and implies that a commit record is contained
504 * within the iclog. We need to ensure that the log tail does not move beyond
505 * the tail that the first commit record in the iclog ordered against, otherwise
506 * correct recovery of that checkpoint becomes dependent on future operations
507 * performed on this iclog.
508 *
509 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
510 * current tail into iclog. Once the iclog tail is set, future operations must
511 * not modify it, otherwise they potentially violate ordering constraints for
512 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
513 * the iclog will get zeroed on activation of the iclog after sync, so we
514 * always capture the tail lsn on the iclog on the first NEED_FUA release
515 * regardless of the number of active reference counts on this iclog.
516 */
517int
518xlog_state_release_iclog(
519 struct xlog *log,
520 struct xlog_in_core *iclog,
521 struct xlog_ticket *ticket)
522{
523 bool last_ref;
524
525 lockdep_assert_held(&log->l_icloglock);
526
527 trace_xlog_iclog_release(iclog, _RET_IP_);
528 /*
529 * Grabbing the current log tail needs to be atomic w.r.t. the writing
530 * of the tail LSN into the iclog so we guarantee that the log tail does
531 * not move between the first time we know that the iclog needs to be
532 * made stable and when we eventually submit it.
533 */
534 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
535 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
536 !iclog->ic_header.h_tail_lsn) {
537 iclog->ic_header.h_tail_lsn =
538 cpu_to_be64(atomic64_read(&log->l_tail_lsn));
539 }
540
541 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
542
543 if (xlog_is_shutdown(log)) {
544 /*
545 * If there are no more references to this iclog, process the
546 * pending iclog callbacks that were waiting on the release of
547 * this iclog.
548 */
549 if (last_ref)
550 xlog_state_shutdown_callbacks(log);
551 return -EIO;
552 }
553
554 if (!last_ref)
555 return 0;
556
557 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
558 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
559 return 0;
560 }
561
562 iclog->ic_state = XLOG_STATE_SYNCING;
563 xlog_verify_tail_lsn(log, iclog);
564 trace_xlog_iclog_syncing(iclog, _RET_IP_);
565
566 spin_unlock(&log->l_icloglock);
567 xlog_sync(log, iclog, ticket);
568 spin_lock(&log->l_icloglock);
569 return 0;
570}
571
572/*
573 * Mount a log filesystem
574 *
575 * mp - ubiquitous xfs mount point structure
576 * log_target - buftarg of on-disk log device
577 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
578 * num_bblocks - Number of BBSIZE blocks in on-disk log
579 *
580 * Return error or zero.
581 */
582int
583xfs_log_mount(
584 xfs_mount_t *mp,
585 struct xfs_buftarg *log_target,
586 xfs_daddr_t blk_offset,
587 int num_bblks)
588{
589 struct xlog *log;
590 int error = 0;
591 int min_logfsbs;
592
593 if (!xfs_has_norecovery(mp)) {
594 xfs_notice(mp, "Mounting V%d Filesystem %pU",
595 XFS_SB_VERSION_NUM(&mp->m_sb),
596 &mp->m_sb.sb_uuid);
597 } else {
598 xfs_notice(mp,
599"Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
600 XFS_SB_VERSION_NUM(&mp->m_sb),
601 &mp->m_sb.sb_uuid);
602 ASSERT(xfs_is_readonly(mp));
603 }
604
605 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
606 if (IS_ERR(log)) {
607 error = PTR_ERR(log);
608 goto out;
609 }
610 mp->m_log = log;
611
612 /*
613 * Now that we have set up the log and it's internal geometry
614 * parameters, we can validate the given log space and drop a critical
615 * message via syslog if the log size is too small. A log that is too
616 * small can lead to unexpected situations in transaction log space
617 * reservation stage. The superblock verifier has already validated all
618 * the other log geometry constraints, so we don't have to check those
619 * here.
620 *
621 * Note: For v4 filesystems, we can't just reject the mount if the
622 * validation fails. This would mean that people would have to
623 * downgrade their kernel just to remedy the situation as there is no
624 * way to grow the log (short of black magic surgery with xfs_db).
625 *
626 * We can, however, reject mounts for V5 format filesystems, as the
627 * mkfs binary being used to make the filesystem should never create a
628 * filesystem with a log that is too small.
629 */
630 min_logfsbs = xfs_log_calc_minimum_size(mp);
631 if (mp->m_sb.sb_logblocks < min_logfsbs) {
632 xfs_warn(mp,
633 "Log size %d blocks too small, minimum size is %d blocks",
634 mp->m_sb.sb_logblocks, min_logfsbs);
635
636 /*
637 * Log check errors are always fatal on v5; or whenever bad
638 * metadata leads to a crash.
639 */
640 if (xfs_has_crc(mp)) {
641 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
642 ASSERT(0);
643 error = -EINVAL;
644 goto out_free_log;
645 }
646 xfs_crit(mp, "Log size out of supported range.");
647 xfs_crit(mp,
648"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
649 }
650
651 /*
652 * Initialize the AIL now we have a log.
653 */
654 error = xfs_trans_ail_init(mp);
655 if (error) {
656 xfs_warn(mp, "AIL initialisation failed: error %d", error);
657 goto out_free_log;
658 }
659 log->l_ailp = mp->m_ail;
660
661 /*
662 * skip log recovery on a norecovery mount. pretend it all
663 * just worked.
664 */
665 if (!xfs_has_norecovery(mp)) {
666 error = xlog_recover(log);
667 if (error) {
668 xfs_warn(mp, "log mount/recovery failed: error %d",
669 error);
670 xlog_recover_cancel(log);
671 goto out_destroy_ail;
672 }
673 }
674
675 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
676 "log");
677 if (error)
678 goto out_destroy_ail;
679
680 /* Normal transactions can now occur */
681 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
682
683 /*
684 * Now the log has been fully initialised and we know were our
685 * space grant counters are, we can initialise the permanent ticket
686 * needed for delayed logging to work.
687 */
688 xlog_cil_init_post_recovery(log);
689
690 return 0;
691
692out_destroy_ail:
693 xfs_trans_ail_destroy(mp);
694out_free_log:
695 xlog_dealloc_log(log);
696out:
697 return error;
698}
699
700/*
701 * Finish the recovery of the file system. This is separate from the
702 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
703 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
704 * here.
705 *
706 * If we finish recovery successfully, start the background log work. If we are
707 * not doing recovery, then we have a RO filesystem and we don't need to start
708 * it.
709 */
710int
711xfs_log_mount_finish(
712 struct xfs_mount *mp)
713{
714 struct xlog *log = mp->m_log;
715 int error = 0;
716
717 if (xfs_has_norecovery(mp)) {
718 ASSERT(xfs_is_readonly(mp));
719 return 0;
720 }
721
722 /*
723 * During the second phase of log recovery, we need iget and
724 * iput to behave like they do for an active filesystem.
725 * xfs_fs_drop_inode needs to be able to prevent the deletion
726 * of inodes before we're done replaying log items on those
727 * inodes. Turn it off immediately after recovery finishes
728 * so that we don't leak the quota inodes if subsequent mount
729 * activities fail.
730 *
731 * We let all inodes involved in redo item processing end up on
732 * the LRU instead of being evicted immediately so that if we do
733 * something to an unlinked inode, the irele won't cause
734 * premature truncation and freeing of the inode, which results
735 * in log recovery failure. We have to evict the unreferenced
736 * lru inodes after clearing SB_ACTIVE because we don't
737 * otherwise clean up the lru if there's a subsequent failure in
738 * xfs_mountfs, which leads to us leaking the inodes if nothing
739 * else (e.g. quotacheck) references the inodes before the
740 * mount failure occurs.
741 */
742 mp->m_super->s_flags |= SB_ACTIVE;
743 xfs_log_work_queue(mp);
744 if (xlog_recovery_needed(log))
745 error = xlog_recover_finish(log);
746 mp->m_super->s_flags &= ~SB_ACTIVE;
747 evict_inodes(mp->m_super);
748
749 /*
750 * Drain the buffer LRU after log recovery. This is required for v4
751 * filesystems to avoid leaving around buffers with NULL verifier ops,
752 * but we do it unconditionally to make sure we're always in a clean
753 * cache state after mount.
754 *
755 * Don't push in the error case because the AIL may have pending intents
756 * that aren't removed until recovery is cancelled.
757 */
758 if (xlog_recovery_needed(log)) {
759 if (!error) {
760 xfs_log_force(mp, XFS_LOG_SYNC);
761 xfs_ail_push_all_sync(mp->m_ail);
762 }
763 xfs_notice(mp, "Ending recovery (logdev: %s)",
764 mp->m_logname ? mp->m_logname : "internal");
765 } else {
766 xfs_info(mp, "Ending clean mount");
767 }
768 xfs_buftarg_drain(mp->m_ddev_targp);
769
770 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
771
772 /* Make sure the log is dead if we're returning failure. */
773 ASSERT(!error || xlog_is_shutdown(log));
774
775 return error;
776}
777
778/*
779 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
780 * the log.
781 */
782void
783xfs_log_mount_cancel(
784 struct xfs_mount *mp)
785{
786 xlog_recover_cancel(mp->m_log);
787 xfs_log_unmount(mp);
788}
789
790/*
791 * Flush out the iclog to disk ensuring that device caches are flushed and
792 * the iclog hits stable storage before any completion waiters are woken.
793 */
794static inline int
795xlog_force_iclog(
796 struct xlog_in_core *iclog)
797{
798 atomic_inc(&iclog->ic_refcnt);
799 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
800 if (iclog->ic_state == XLOG_STATE_ACTIVE)
801 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
802 return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
803}
804
805/*
806 * Cycle all the iclogbuf locks to make sure all log IO completion
807 * is done before we tear down these buffers.
808 */
809static void
810xlog_wait_iclog_completion(struct xlog *log)
811{
812 int i;
813 struct xlog_in_core *iclog = log->l_iclog;
814
815 for (i = 0; i < log->l_iclog_bufs; i++) {
816 down(&iclog->ic_sema);
817 up(&iclog->ic_sema);
818 iclog = iclog->ic_next;
819 }
820}
821
822/*
823 * Wait for the iclog and all prior iclogs to be written disk as required by the
824 * log force state machine. Waiting on ic_force_wait ensures iclog completions
825 * have been ordered and callbacks run before we are woken here, hence
826 * guaranteeing that all the iclogs up to this one are on stable storage.
827 */
828int
829xlog_wait_on_iclog(
830 struct xlog_in_core *iclog)
831 __releases(iclog->ic_log->l_icloglock)
832{
833 struct xlog *log = iclog->ic_log;
834
835 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
836 if (!xlog_is_shutdown(log) &&
837 iclog->ic_state != XLOG_STATE_ACTIVE &&
838 iclog->ic_state != XLOG_STATE_DIRTY) {
839 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
840 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
841 } else {
842 spin_unlock(&log->l_icloglock);
843 }
844
845 if (xlog_is_shutdown(log))
846 return -EIO;
847 return 0;
848}
849
850/*
851 * Write out an unmount record using the ticket provided. We have to account for
852 * the data space used in the unmount ticket as this write is not done from a
853 * transaction context that has already done the accounting for us.
854 */
855static int
856xlog_write_unmount_record(
857 struct xlog *log,
858 struct xlog_ticket *ticket)
859{
860 struct {
861 struct xlog_op_header ophdr;
862 struct xfs_unmount_log_format ulf;
863 } unmount_rec = {
864 .ophdr = {
865 .oh_clientid = XFS_LOG,
866 .oh_tid = cpu_to_be32(ticket->t_tid),
867 .oh_flags = XLOG_UNMOUNT_TRANS,
868 },
869 .ulf = {
870 .magic = XLOG_UNMOUNT_TYPE,
871 },
872 };
873 struct xfs_log_iovec reg = {
874 .i_addr = &unmount_rec,
875 .i_len = sizeof(unmount_rec),
876 .i_type = XLOG_REG_TYPE_UNMOUNT,
877 };
878 struct xfs_log_vec vec = {
879 .lv_niovecs = 1,
880 .lv_iovecp = ®,
881 };
882 LIST_HEAD(lv_chain);
883 list_add(&vec.lv_list, &lv_chain);
884
885 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
886 sizeof(struct xfs_unmount_log_format)) !=
887 sizeof(unmount_rec));
888
889 /* account for space used by record data */
890 ticket->t_curr_res -= sizeof(unmount_rec);
891
892 return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
893}
894
895/*
896 * Mark the filesystem clean by writing an unmount record to the head of the
897 * log.
898 */
899static void
900xlog_unmount_write(
901 struct xlog *log)
902{
903 struct xfs_mount *mp = log->l_mp;
904 struct xlog_in_core *iclog;
905 struct xlog_ticket *tic = NULL;
906 int error;
907
908 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
909 if (error)
910 goto out_err;
911
912 error = xlog_write_unmount_record(log, tic);
913 /*
914 * At this point, we're umounting anyway, so there's no point in
915 * transitioning log state to shutdown. Just continue...
916 */
917out_err:
918 if (error)
919 xfs_alert(mp, "%s: unmount record failed", __func__);
920
921 spin_lock(&log->l_icloglock);
922 iclog = log->l_iclog;
923 error = xlog_force_iclog(iclog);
924 xlog_wait_on_iclog(iclog);
925
926 if (tic) {
927 trace_xfs_log_umount_write(log, tic);
928 xfs_log_ticket_ungrant(log, tic);
929 }
930}
931
932static void
933xfs_log_unmount_verify_iclog(
934 struct xlog *log)
935{
936 struct xlog_in_core *iclog = log->l_iclog;
937
938 do {
939 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
940 ASSERT(iclog->ic_offset == 0);
941 } while ((iclog = iclog->ic_next) != log->l_iclog);
942}
943
944/*
945 * Unmount record used to have a string "Unmount filesystem--" in the
946 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
947 * We just write the magic number now since that particular field isn't
948 * currently architecture converted and "Unmount" is a bit foo.
949 * As far as I know, there weren't any dependencies on the old behaviour.
950 */
951static void
952xfs_log_unmount_write(
953 struct xfs_mount *mp)
954{
955 struct xlog *log = mp->m_log;
956
957 if (!xfs_log_writable(mp))
958 return;
959
960 xfs_log_force(mp, XFS_LOG_SYNC);
961
962 if (xlog_is_shutdown(log))
963 return;
964
965 /*
966 * If we think the summary counters are bad, avoid writing the unmount
967 * record to force log recovery at next mount, after which the summary
968 * counters will be recalculated. Refer to xlog_check_unmount_rec for
969 * more details.
970 */
971 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
972 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
973 xfs_alert(mp, "%s: will fix summary counters at next mount",
974 __func__);
975 return;
976 }
977
978 xfs_log_unmount_verify_iclog(log);
979 xlog_unmount_write(log);
980}
981
982/*
983 * Empty the log for unmount/freeze.
984 *
985 * To do this, we first need to shut down the background log work so it is not
986 * trying to cover the log as we clean up. We then need to unpin all objects in
987 * the log so we can then flush them out. Once they have completed their IO and
988 * run the callbacks removing themselves from the AIL, we can cover the log.
989 */
990int
991xfs_log_quiesce(
992 struct xfs_mount *mp)
993{
994 /*
995 * Clear log incompat features since we're quiescing the log. Report
996 * failures, though it's not fatal to have a higher log feature
997 * protection level than the log contents actually require.
998 */
999 if (xfs_clear_incompat_log_features(mp)) {
1000 int error;
1001
1002 error = xfs_sync_sb(mp, false);
1003 if (error)
1004 xfs_warn(mp,
1005 "Failed to clear log incompat features on quiesce");
1006 }
1007
1008 cancel_delayed_work_sync(&mp->m_log->l_work);
1009 xfs_log_force(mp, XFS_LOG_SYNC);
1010
1011 /*
1012 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1013 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1014 * xfs_buf_iowait() cannot be used because it was pushed with the
1015 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1016 * the IO to complete.
1017 */
1018 xfs_ail_push_all_sync(mp->m_ail);
1019 xfs_buftarg_wait(mp->m_ddev_targp);
1020 xfs_buf_lock(mp->m_sb_bp);
1021 xfs_buf_unlock(mp->m_sb_bp);
1022
1023 return xfs_log_cover(mp);
1024}
1025
1026void
1027xfs_log_clean(
1028 struct xfs_mount *mp)
1029{
1030 xfs_log_quiesce(mp);
1031 xfs_log_unmount_write(mp);
1032}
1033
1034/*
1035 * Shut down and release the AIL and Log.
1036 *
1037 * During unmount, we need to ensure we flush all the dirty metadata objects
1038 * from the AIL so that the log is empty before we write the unmount record to
1039 * the log. Once this is done, we can tear down the AIL and the log.
1040 */
1041void
1042xfs_log_unmount(
1043 struct xfs_mount *mp)
1044{
1045 xfs_log_clean(mp);
1046
1047 /*
1048 * If shutdown has come from iclog IO context, the log
1049 * cleaning will have been skipped and so we need to wait
1050 * for the iclog to complete shutdown processing before we
1051 * tear anything down.
1052 */
1053 xlog_wait_iclog_completion(mp->m_log);
1054
1055 xfs_buftarg_drain(mp->m_ddev_targp);
1056
1057 xfs_trans_ail_destroy(mp);
1058
1059 xfs_sysfs_del(&mp->m_log->l_kobj);
1060
1061 xlog_dealloc_log(mp->m_log);
1062}
1063
1064void
1065xfs_log_item_init(
1066 struct xfs_mount *mp,
1067 struct xfs_log_item *item,
1068 int type,
1069 const struct xfs_item_ops *ops)
1070{
1071 item->li_log = mp->m_log;
1072 item->li_ailp = mp->m_ail;
1073 item->li_type = type;
1074 item->li_ops = ops;
1075 item->li_lv = NULL;
1076
1077 INIT_LIST_HEAD(&item->li_ail);
1078 INIT_LIST_HEAD(&item->li_cil);
1079 INIT_LIST_HEAD(&item->li_bio_list);
1080 INIT_LIST_HEAD(&item->li_trans);
1081}
1082
1083/*
1084 * Wake up processes waiting for log space after we have moved the log tail.
1085 */
1086void
1087xfs_log_space_wake(
1088 struct xfs_mount *mp)
1089{
1090 struct xlog *log = mp->m_log;
1091 int free_bytes;
1092
1093 if (xlog_is_shutdown(log))
1094 return;
1095
1096 if (!list_empty_careful(&log->l_write_head.waiters)) {
1097 ASSERT(!xlog_in_recovery(log));
1098
1099 spin_lock(&log->l_write_head.lock);
1100 free_bytes = xlog_grant_space_left(log, &log->l_write_head);
1101 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1102 spin_unlock(&log->l_write_head.lock);
1103 }
1104
1105 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1106 ASSERT(!xlog_in_recovery(log));
1107
1108 spin_lock(&log->l_reserve_head.lock);
1109 free_bytes = xlog_grant_space_left(log, &log->l_reserve_head);
1110 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1111 spin_unlock(&log->l_reserve_head.lock);
1112 }
1113}
1114
1115/*
1116 * Determine if we have a transaction that has gone to disk that needs to be
1117 * covered. To begin the transition to the idle state firstly the log needs to
1118 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1119 * we start attempting to cover the log.
1120 *
1121 * Only if we are then in a state where covering is needed, the caller is
1122 * informed that dummy transactions are required to move the log into the idle
1123 * state.
1124 *
1125 * If there are any items in the AIl or CIL, then we do not want to attempt to
1126 * cover the log as we may be in a situation where there isn't log space
1127 * available to run a dummy transaction and this can lead to deadlocks when the
1128 * tail of the log is pinned by an item that is modified in the CIL. Hence
1129 * there's no point in running a dummy transaction at this point because we
1130 * can't start trying to idle the log until both the CIL and AIL are empty.
1131 */
1132static bool
1133xfs_log_need_covered(
1134 struct xfs_mount *mp)
1135{
1136 struct xlog *log = mp->m_log;
1137 bool needed = false;
1138
1139 if (!xlog_cil_empty(log))
1140 return false;
1141
1142 spin_lock(&log->l_icloglock);
1143 switch (log->l_covered_state) {
1144 case XLOG_STATE_COVER_DONE:
1145 case XLOG_STATE_COVER_DONE2:
1146 case XLOG_STATE_COVER_IDLE:
1147 break;
1148 case XLOG_STATE_COVER_NEED:
1149 case XLOG_STATE_COVER_NEED2:
1150 if (xfs_ail_min_lsn(log->l_ailp))
1151 break;
1152 if (!xlog_iclogs_empty(log))
1153 break;
1154
1155 needed = true;
1156 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1157 log->l_covered_state = XLOG_STATE_COVER_DONE;
1158 else
1159 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1160 break;
1161 default:
1162 needed = true;
1163 break;
1164 }
1165 spin_unlock(&log->l_icloglock);
1166 return needed;
1167}
1168
1169/*
1170 * Explicitly cover the log. This is similar to background log covering but
1171 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1172 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1173 * must all be empty.
1174 */
1175static int
1176xfs_log_cover(
1177 struct xfs_mount *mp)
1178{
1179 int error = 0;
1180 bool need_covered;
1181
1182 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1183 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1184 xlog_is_shutdown(mp->m_log));
1185
1186 if (!xfs_log_writable(mp))
1187 return 0;
1188
1189 /*
1190 * xfs_log_need_covered() is not idempotent because it progresses the
1191 * state machine if the log requires covering. Therefore, we must call
1192 * this function once and use the result until we've issued an sb sync.
1193 * Do so first to make that abundantly clear.
1194 *
1195 * Fall into the covering sequence if the log needs covering or the
1196 * mount has lazy superblock accounting to sync to disk. The sb sync
1197 * used for covering accumulates the in-core counters, so covering
1198 * handles this for us.
1199 */
1200 need_covered = xfs_log_need_covered(mp);
1201 if (!need_covered && !xfs_has_lazysbcount(mp))
1202 return 0;
1203
1204 /*
1205 * To cover the log, commit the superblock twice (at most) in
1206 * independent checkpoints. The first serves as a reference for the
1207 * tail pointer. The sync transaction and AIL push empties the AIL and
1208 * updates the in-core tail to the LSN of the first checkpoint. The
1209 * second commit updates the on-disk tail with the in-core LSN,
1210 * covering the log. Push the AIL one more time to leave it empty, as
1211 * we found it.
1212 */
1213 do {
1214 error = xfs_sync_sb(mp, true);
1215 if (error)
1216 break;
1217 xfs_ail_push_all_sync(mp->m_ail);
1218 } while (xfs_log_need_covered(mp));
1219
1220 return error;
1221}
1222
1223static void
1224xlog_ioend_work(
1225 struct work_struct *work)
1226{
1227 struct xlog_in_core *iclog =
1228 container_of(work, struct xlog_in_core, ic_end_io_work);
1229 struct xlog *log = iclog->ic_log;
1230 int error;
1231
1232 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1233#ifdef DEBUG
1234 /* treat writes with injected CRC errors as failed */
1235 if (iclog->ic_fail_crc)
1236 error = -EIO;
1237#endif
1238
1239 /*
1240 * Race to shutdown the filesystem if we see an error.
1241 */
1242 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1243 xfs_alert(log->l_mp, "log I/O error %d", error);
1244 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1245 }
1246
1247 xlog_state_done_syncing(iclog);
1248 bio_uninit(&iclog->ic_bio);
1249
1250 /*
1251 * Drop the lock to signal that we are done. Nothing references the
1252 * iclog after this, so an unmount waiting on this lock can now tear it
1253 * down safely. As such, it is unsafe to reference the iclog after the
1254 * unlock as we could race with it being freed.
1255 */
1256 up(&iclog->ic_sema);
1257}
1258
1259/*
1260 * Return size of each in-core log record buffer.
1261 *
1262 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1263 *
1264 * If the filesystem blocksize is too large, we may need to choose a
1265 * larger size since the directory code currently logs entire blocks.
1266 */
1267STATIC void
1268xlog_get_iclog_buffer_size(
1269 struct xfs_mount *mp,
1270 struct xlog *log)
1271{
1272 if (mp->m_logbufs <= 0)
1273 mp->m_logbufs = XLOG_MAX_ICLOGS;
1274 if (mp->m_logbsize <= 0)
1275 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1276
1277 log->l_iclog_bufs = mp->m_logbufs;
1278 log->l_iclog_size = mp->m_logbsize;
1279
1280 /*
1281 * # headers = size / 32k - one header holds cycles from 32k of data.
1282 */
1283 log->l_iclog_heads =
1284 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1285 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1286}
1287
1288void
1289xfs_log_work_queue(
1290 struct xfs_mount *mp)
1291{
1292 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1293 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1294}
1295
1296/*
1297 * Clear the log incompat flags if we have the opportunity.
1298 *
1299 * This only happens if we're about to log the second dummy transaction as part
1300 * of covering the log.
1301 */
1302static inline void
1303xlog_clear_incompat(
1304 struct xlog *log)
1305{
1306 struct xfs_mount *mp = log->l_mp;
1307
1308 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1309 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1310 return;
1311
1312 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1313 return;
1314
1315 xfs_clear_incompat_log_features(mp);
1316}
1317
1318/*
1319 * Every sync period we need to unpin all items in the AIL and push them to
1320 * disk. If there is nothing dirty, then we might need to cover the log to
1321 * indicate that the filesystem is idle.
1322 */
1323static void
1324xfs_log_worker(
1325 struct work_struct *work)
1326{
1327 struct xlog *log = container_of(to_delayed_work(work),
1328 struct xlog, l_work);
1329 struct xfs_mount *mp = log->l_mp;
1330
1331 /* dgc: errors ignored - not fatal and nowhere to report them */
1332 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1333 /*
1334 * Dump a transaction into the log that contains no real change.
1335 * This is needed to stamp the current tail LSN into the log
1336 * during the covering operation.
1337 *
1338 * We cannot use an inode here for this - that will push dirty
1339 * state back up into the VFS and then periodic inode flushing
1340 * will prevent log covering from making progress. Hence we
1341 * synchronously log the superblock instead to ensure the
1342 * superblock is immediately unpinned and can be written back.
1343 */
1344 xlog_clear_incompat(log);
1345 xfs_sync_sb(mp, true);
1346 } else
1347 xfs_log_force(mp, 0);
1348
1349 /* start pushing all the metadata that is currently dirty */
1350 xfs_ail_push_all(mp->m_ail);
1351
1352 /* queue us up again */
1353 xfs_log_work_queue(mp);
1354}
1355
1356/*
1357 * This routine initializes some of the log structure for a given mount point.
1358 * Its primary purpose is to fill in enough, so recovery can occur. However,
1359 * some other stuff may be filled in too.
1360 */
1361STATIC struct xlog *
1362xlog_alloc_log(
1363 struct xfs_mount *mp,
1364 struct xfs_buftarg *log_target,
1365 xfs_daddr_t blk_offset,
1366 int num_bblks)
1367{
1368 struct xlog *log;
1369 xlog_rec_header_t *head;
1370 xlog_in_core_t **iclogp;
1371 xlog_in_core_t *iclog, *prev_iclog=NULL;
1372 int i;
1373 int error = -ENOMEM;
1374 uint log2_size = 0;
1375
1376 log = kzalloc(sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1377 if (!log) {
1378 xfs_warn(mp, "Log allocation failed: No memory!");
1379 goto out;
1380 }
1381
1382 log->l_mp = mp;
1383 log->l_targ = log_target;
1384 log->l_logsize = BBTOB(num_bblks);
1385 log->l_logBBstart = blk_offset;
1386 log->l_logBBsize = num_bblks;
1387 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1388 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1389 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1390 INIT_LIST_HEAD(&log->r_dfops);
1391
1392 log->l_prev_block = -1;
1393 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1394 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1395 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1396
1397 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1398 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1399 else
1400 log->l_iclog_roundoff = BBSIZE;
1401
1402 xlog_grant_head_init(&log->l_reserve_head);
1403 xlog_grant_head_init(&log->l_write_head);
1404
1405 error = -EFSCORRUPTED;
1406 if (xfs_has_sector(mp)) {
1407 log2_size = mp->m_sb.sb_logsectlog;
1408 if (log2_size < BBSHIFT) {
1409 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1410 log2_size, BBSHIFT);
1411 goto out_free_log;
1412 }
1413
1414 log2_size -= BBSHIFT;
1415 if (log2_size > mp->m_sectbb_log) {
1416 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1417 log2_size, mp->m_sectbb_log);
1418 goto out_free_log;
1419 }
1420
1421 /* for larger sector sizes, must have v2 or external log */
1422 if (log2_size && log->l_logBBstart > 0 &&
1423 !xfs_has_logv2(mp)) {
1424 xfs_warn(mp,
1425 "log sector size (0x%x) invalid for configuration.",
1426 log2_size);
1427 goto out_free_log;
1428 }
1429 }
1430 log->l_sectBBsize = 1 << log2_size;
1431
1432 xlog_get_iclog_buffer_size(mp, log);
1433
1434 spin_lock_init(&log->l_icloglock);
1435 init_waitqueue_head(&log->l_flush_wait);
1436
1437 iclogp = &log->l_iclog;
1438 /*
1439 * The amount of memory to allocate for the iclog structure is
1440 * rather funky due to the way the structure is defined. It is
1441 * done this way so that we can use different sizes for machines
1442 * with different amounts of memory. See the definition of
1443 * xlog_in_core_t in xfs_log_priv.h for details.
1444 */
1445 ASSERT(log->l_iclog_size >= 4096);
1446 for (i = 0; i < log->l_iclog_bufs; i++) {
1447 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1448 sizeof(struct bio_vec);
1449
1450 iclog = kzalloc(sizeof(*iclog) + bvec_size,
1451 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1452 if (!iclog)
1453 goto out_free_iclog;
1454
1455 *iclogp = iclog;
1456 iclog->ic_prev = prev_iclog;
1457 prev_iclog = iclog;
1458
1459 iclog->ic_data = kvzalloc(log->l_iclog_size,
1460 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1461 if (!iclog->ic_data)
1462 goto out_free_iclog;
1463 head = &iclog->ic_header;
1464 memset(head, 0, sizeof(xlog_rec_header_t));
1465 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1466 head->h_version = cpu_to_be32(
1467 xfs_has_logv2(log->l_mp) ? 2 : 1);
1468 head->h_size = cpu_to_be32(log->l_iclog_size);
1469 /* new fields */
1470 head->h_fmt = cpu_to_be32(XLOG_FMT);
1471 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1472
1473 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1474 iclog->ic_state = XLOG_STATE_ACTIVE;
1475 iclog->ic_log = log;
1476 atomic_set(&iclog->ic_refcnt, 0);
1477 INIT_LIST_HEAD(&iclog->ic_callbacks);
1478 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1479
1480 init_waitqueue_head(&iclog->ic_force_wait);
1481 init_waitqueue_head(&iclog->ic_write_wait);
1482 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1483 sema_init(&iclog->ic_sema, 1);
1484
1485 iclogp = &iclog->ic_next;
1486 }
1487 *iclogp = log->l_iclog; /* complete ring */
1488 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1489
1490 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1491 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1492 WQ_HIGHPRI),
1493 0, mp->m_super->s_id);
1494 if (!log->l_ioend_workqueue)
1495 goto out_free_iclog;
1496
1497 error = xlog_cil_init(log);
1498 if (error)
1499 goto out_destroy_workqueue;
1500 return log;
1501
1502out_destroy_workqueue:
1503 destroy_workqueue(log->l_ioend_workqueue);
1504out_free_iclog:
1505 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1506 prev_iclog = iclog->ic_next;
1507 kvfree(iclog->ic_data);
1508 kfree(iclog);
1509 if (prev_iclog == log->l_iclog)
1510 break;
1511 }
1512out_free_log:
1513 kfree(log);
1514out:
1515 return ERR_PTR(error);
1516} /* xlog_alloc_log */
1517
1518/*
1519 * Stamp cycle number in every block
1520 */
1521STATIC void
1522xlog_pack_data(
1523 struct xlog *log,
1524 struct xlog_in_core *iclog,
1525 int roundoff)
1526{
1527 int i, j, k;
1528 int size = iclog->ic_offset + roundoff;
1529 __be32 cycle_lsn;
1530 char *dp;
1531
1532 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1533
1534 dp = iclog->ic_datap;
1535 for (i = 0; i < BTOBB(size); i++) {
1536 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1537 break;
1538 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1539 *(__be32 *)dp = cycle_lsn;
1540 dp += BBSIZE;
1541 }
1542
1543 if (xfs_has_logv2(log->l_mp)) {
1544 xlog_in_core_2_t *xhdr = iclog->ic_data;
1545
1546 for ( ; i < BTOBB(size); i++) {
1547 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1548 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1549 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1550 *(__be32 *)dp = cycle_lsn;
1551 dp += BBSIZE;
1552 }
1553
1554 for (i = 1; i < log->l_iclog_heads; i++)
1555 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1556 }
1557}
1558
1559/*
1560 * Calculate the checksum for a log buffer.
1561 *
1562 * This is a little more complicated than it should be because the various
1563 * headers and the actual data are non-contiguous.
1564 */
1565__le32
1566xlog_cksum(
1567 struct xlog *log,
1568 struct xlog_rec_header *rhead,
1569 char *dp,
1570 int size)
1571{
1572 uint32_t crc;
1573
1574 /* first generate the crc for the record header ... */
1575 crc = xfs_start_cksum_update((char *)rhead,
1576 sizeof(struct xlog_rec_header),
1577 offsetof(struct xlog_rec_header, h_crc));
1578
1579 /* ... then for additional cycle data for v2 logs ... */
1580 if (xfs_has_logv2(log->l_mp)) {
1581 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1582 int i;
1583 int xheads;
1584
1585 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1586
1587 for (i = 1; i < xheads; i++) {
1588 crc = crc32c(crc, &xhdr[i].hic_xheader,
1589 sizeof(struct xlog_rec_ext_header));
1590 }
1591 }
1592
1593 /* ... and finally for the payload */
1594 crc = crc32c(crc, dp, size);
1595
1596 return xfs_end_cksum(crc);
1597}
1598
1599static void
1600xlog_bio_end_io(
1601 struct bio *bio)
1602{
1603 struct xlog_in_core *iclog = bio->bi_private;
1604
1605 queue_work(iclog->ic_log->l_ioend_workqueue,
1606 &iclog->ic_end_io_work);
1607}
1608
1609static int
1610xlog_map_iclog_data(
1611 struct bio *bio,
1612 void *data,
1613 size_t count)
1614{
1615 do {
1616 struct page *page = kmem_to_page(data);
1617 unsigned int off = offset_in_page(data);
1618 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1619
1620 if (bio_add_page(bio, page, len, off) != len)
1621 return -EIO;
1622
1623 data += len;
1624 count -= len;
1625 } while (count);
1626
1627 return 0;
1628}
1629
1630STATIC void
1631xlog_write_iclog(
1632 struct xlog *log,
1633 struct xlog_in_core *iclog,
1634 uint64_t bno,
1635 unsigned int count)
1636{
1637 ASSERT(bno < log->l_logBBsize);
1638 trace_xlog_iclog_write(iclog, _RET_IP_);
1639
1640 /*
1641 * We lock the iclogbufs here so that we can serialise against I/O
1642 * completion during unmount. We might be processing a shutdown
1643 * triggered during unmount, and that can occur asynchronously to the
1644 * unmount thread, and hence we need to ensure that completes before
1645 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1646 * across the log IO to archieve that.
1647 */
1648 down(&iclog->ic_sema);
1649 if (xlog_is_shutdown(log)) {
1650 /*
1651 * It would seem logical to return EIO here, but we rely on
1652 * the log state machine to propagate I/O errors instead of
1653 * doing it here. We kick of the state machine and unlock
1654 * the buffer manually, the code needs to be kept in sync
1655 * with the I/O completion path.
1656 */
1657 goto sync;
1658 }
1659
1660 /*
1661 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1662 * IOs coming immediately after this one. This prevents the block layer
1663 * writeback throttle from throttling log writes behind background
1664 * metadata writeback and causing priority inversions.
1665 */
1666 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1667 howmany(count, PAGE_SIZE),
1668 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1669 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1670 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1671 iclog->ic_bio.bi_private = iclog;
1672
1673 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1674 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1675 /*
1676 * For external log devices, we also need to flush the data
1677 * device cache first to ensure all metadata writeback covered
1678 * by the LSN in this iclog is on stable storage. This is slow,
1679 * but it *must* complete before we issue the external log IO.
1680 *
1681 * If the flush fails, we cannot conclude that past metadata
1682 * writeback from the log succeeded. Repeating the flush is
1683 * not possible, hence we must shut down with log IO error to
1684 * avoid shutdown re-entering this path and erroring out again.
1685 */
1686 if (log->l_targ != log->l_mp->m_ddev_targp &&
1687 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev))
1688 goto shutdown;
1689 }
1690 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1691 iclog->ic_bio.bi_opf |= REQ_FUA;
1692
1693 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1694
1695 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count))
1696 goto shutdown;
1697
1698 if (is_vmalloc_addr(iclog->ic_data))
1699 flush_kernel_vmap_range(iclog->ic_data, count);
1700
1701 /*
1702 * If this log buffer would straddle the end of the log we will have
1703 * to split it up into two bios, so that we can continue at the start.
1704 */
1705 if (bno + BTOBB(count) > log->l_logBBsize) {
1706 struct bio *split;
1707
1708 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1709 GFP_NOIO, &fs_bio_set);
1710 bio_chain(split, &iclog->ic_bio);
1711 submit_bio(split);
1712
1713 /* restart at logical offset zero for the remainder */
1714 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1715 }
1716
1717 submit_bio(&iclog->ic_bio);
1718 return;
1719shutdown:
1720 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1721sync:
1722 xlog_state_done_syncing(iclog);
1723 up(&iclog->ic_sema);
1724}
1725
1726/*
1727 * We need to bump cycle number for the part of the iclog that is
1728 * written to the start of the log. Watch out for the header magic
1729 * number case, though.
1730 */
1731static void
1732xlog_split_iclog(
1733 struct xlog *log,
1734 void *data,
1735 uint64_t bno,
1736 unsigned int count)
1737{
1738 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1739 unsigned int i;
1740
1741 for (i = split_offset; i < count; i += BBSIZE) {
1742 uint32_t cycle = get_unaligned_be32(data + i);
1743
1744 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1745 cycle++;
1746 put_unaligned_be32(cycle, data + i);
1747 }
1748}
1749
1750static int
1751xlog_calc_iclog_size(
1752 struct xlog *log,
1753 struct xlog_in_core *iclog,
1754 uint32_t *roundoff)
1755{
1756 uint32_t count_init, count;
1757
1758 /* Add for LR header */
1759 count_init = log->l_iclog_hsize + iclog->ic_offset;
1760 count = roundup(count_init, log->l_iclog_roundoff);
1761
1762 *roundoff = count - count_init;
1763
1764 ASSERT(count >= count_init);
1765 ASSERT(*roundoff < log->l_iclog_roundoff);
1766 return count;
1767}
1768
1769/*
1770 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1771 * fashion. Previously, we should have moved the current iclog
1772 * ptr in the log to point to the next available iclog. This allows further
1773 * write to continue while this code syncs out an iclog ready to go.
1774 * Before an in-core log can be written out, the data section must be scanned
1775 * to save away the 1st word of each BBSIZE block into the header. We replace
1776 * it with the current cycle count. Each BBSIZE block is tagged with the
1777 * cycle count because there in an implicit assumption that drives will
1778 * guarantee that entire 512 byte blocks get written at once. In other words,
1779 * we can't have part of a 512 byte block written and part not written. By
1780 * tagging each block, we will know which blocks are valid when recovering
1781 * after an unclean shutdown.
1782 *
1783 * This routine is single threaded on the iclog. No other thread can be in
1784 * this routine with the same iclog. Changing contents of iclog can there-
1785 * fore be done without grabbing the state machine lock. Updating the global
1786 * log will require grabbing the lock though.
1787 *
1788 * The entire log manager uses a logical block numbering scheme. Only
1789 * xlog_write_iclog knows about the fact that the log may not start with
1790 * block zero on a given device.
1791 */
1792STATIC void
1793xlog_sync(
1794 struct xlog *log,
1795 struct xlog_in_core *iclog,
1796 struct xlog_ticket *ticket)
1797{
1798 unsigned int count; /* byte count of bwrite */
1799 unsigned int roundoff; /* roundoff to BB or stripe */
1800 uint64_t bno;
1801 unsigned int size;
1802
1803 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1804 trace_xlog_iclog_sync(iclog, _RET_IP_);
1805
1806 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1807
1808 /*
1809 * If we have a ticket, account for the roundoff via the ticket
1810 * reservation to avoid touching the hot grant heads needlessly.
1811 * Otherwise, we have to move grant heads directly.
1812 */
1813 if (ticket) {
1814 ticket->t_curr_res -= roundoff;
1815 } else {
1816 xlog_grant_add_space(&log->l_reserve_head, roundoff);
1817 xlog_grant_add_space(&log->l_write_head, roundoff);
1818 }
1819
1820 /* put cycle number in every block */
1821 xlog_pack_data(log, iclog, roundoff);
1822
1823 /* real byte length */
1824 size = iclog->ic_offset;
1825 if (xfs_has_logv2(log->l_mp))
1826 size += roundoff;
1827 iclog->ic_header.h_len = cpu_to_be32(size);
1828
1829 XFS_STATS_INC(log->l_mp, xs_log_writes);
1830 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1831
1832 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1833
1834 /* Do we need to split this write into 2 parts? */
1835 if (bno + BTOBB(count) > log->l_logBBsize)
1836 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1837
1838 /* calculcate the checksum */
1839 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1840 iclog->ic_datap, size);
1841 /*
1842 * Intentionally corrupt the log record CRC based on the error injection
1843 * frequency, if defined. This facilitates testing log recovery in the
1844 * event of torn writes. Hence, set the IOABORT state to abort the log
1845 * write on I/O completion and shutdown the fs. The subsequent mount
1846 * detects the bad CRC and attempts to recover.
1847 */
1848#ifdef DEBUG
1849 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1850 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1851 iclog->ic_fail_crc = true;
1852 xfs_warn(log->l_mp,
1853 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1854 be64_to_cpu(iclog->ic_header.h_lsn));
1855 }
1856#endif
1857 xlog_verify_iclog(log, iclog, count);
1858 xlog_write_iclog(log, iclog, bno, count);
1859}
1860
1861/*
1862 * Deallocate a log structure
1863 */
1864STATIC void
1865xlog_dealloc_log(
1866 struct xlog *log)
1867{
1868 xlog_in_core_t *iclog, *next_iclog;
1869 int i;
1870
1871 /*
1872 * Destroy the CIL after waiting for iclog IO completion because an
1873 * iclog EIO error will try to shut down the log, which accesses the
1874 * CIL to wake up the waiters.
1875 */
1876 xlog_cil_destroy(log);
1877
1878 iclog = log->l_iclog;
1879 for (i = 0; i < log->l_iclog_bufs; i++) {
1880 next_iclog = iclog->ic_next;
1881 kvfree(iclog->ic_data);
1882 kfree(iclog);
1883 iclog = next_iclog;
1884 }
1885
1886 log->l_mp->m_log = NULL;
1887 destroy_workqueue(log->l_ioend_workqueue);
1888 kfree(log);
1889}
1890
1891/*
1892 * Update counters atomically now that memcpy is done.
1893 */
1894static inline void
1895xlog_state_finish_copy(
1896 struct xlog *log,
1897 struct xlog_in_core *iclog,
1898 int record_cnt,
1899 int copy_bytes)
1900{
1901 lockdep_assert_held(&log->l_icloglock);
1902
1903 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1904 iclog->ic_offset += copy_bytes;
1905}
1906
1907/*
1908 * print out info relating to regions written which consume
1909 * the reservation
1910 */
1911void
1912xlog_print_tic_res(
1913 struct xfs_mount *mp,
1914 struct xlog_ticket *ticket)
1915{
1916 xfs_warn(mp, "ticket reservation summary:");
1917 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
1918 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
1919 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
1920 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
1921}
1922
1923/*
1924 * Print a summary of the transaction.
1925 */
1926void
1927xlog_print_trans(
1928 struct xfs_trans *tp)
1929{
1930 struct xfs_mount *mp = tp->t_mountp;
1931 struct xfs_log_item *lip;
1932
1933 /* dump core transaction and ticket info */
1934 xfs_warn(mp, "transaction summary:");
1935 xfs_warn(mp, " log res = %d", tp->t_log_res);
1936 xfs_warn(mp, " log count = %d", tp->t_log_count);
1937 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
1938
1939 xlog_print_tic_res(mp, tp->t_ticket);
1940
1941 /* dump each log item */
1942 list_for_each_entry(lip, &tp->t_items, li_trans) {
1943 struct xfs_log_vec *lv = lip->li_lv;
1944 struct xfs_log_iovec *vec;
1945 int i;
1946
1947 xfs_warn(mp, "log item: ");
1948 xfs_warn(mp, " type = 0x%x", lip->li_type);
1949 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
1950 if (!lv)
1951 continue;
1952 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
1953 xfs_warn(mp, " size = %d", lv->lv_size);
1954 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
1955 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
1956
1957 /* dump each iovec for the log item */
1958 vec = lv->lv_iovecp;
1959 for (i = 0; i < lv->lv_niovecs; i++) {
1960 int dumplen = min(vec->i_len, 32);
1961
1962 xfs_warn(mp, " iovec[%d]", i);
1963 xfs_warn(mp, " type = 0x%x", vec->i_type);
1964 xfs_warn(mp, " len = %d", vec->i_len);
1965 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
1966 xfs_hex_dump(vec->i_addr, dumplen);
1967
1968 vec++;
1969 }
1970 }
1971}
1972
1973static inline void
1974xlog_write_iovec(
1975 struct xlog_in_core *iclog,
1976 uint32_t *log_offset,
1977 void *data,
1978 uint32_t write_len,
1979 int *bytes_left,
1980 uint32_t *record_cnt,
1981 uint32_t *data_cnt)
1982{
1983 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
1984 ASSERT(*log_offset % sizeof(int32_t) == 0);
1985 ASSERT(write_len % sizeof(int32_t) == 0);
1986
1987 memcpy(iclog->ic_datap + *log_offset, data, write_len);
1988 *log_offset += write_len;
1989 *bytes_left -= write_len;
1990 (*record_cnt)++;
1991 *data_cnt += write_len;
1992}
1993
1994/*
1995 * Write log vectors into a single iclog which is guaranteed by the caller
1996 * to have enough space to write the entire log vector into.
1997 */
1998static void
1999xlog_write_full(
2000 struct xfs_log_vec *lv,
2001 struct xlog_ticket *ticket,
2002 struct xlog_in_core *iclog,
2003 uint32_t *log_offset,
2004 uint32_t *len,
2005 uint32_t *record_cnt,
2006 uint32_t *data_cnt)
2007{
2008 int index;
2009
2010 ASSERT(*log_offset + *len <= iclog->ic_size ||
2011 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2012
2013 /*
2014 * Ordered log vectors have no regions to write so this
2015 * loop will naturally skip them.
2016 */
2017 for (index = 0; index < lv->lv_niovecs; index++) {
2018 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2019 struct xlog_op_header *ophdr = reg->i_addr;
2020
2021 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2022 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2023 reg->i_len, len, record_cnt, data_cnt);
2024 }
2025}
2026
2027static int
2028xlog_write_get_more_iclog_space(
2029 struct xlog_ticket *ticket,
2030 struct xlog_in_core **iclogp,
2031 uint32_t *log_offset,
2032 uint32_t len,
2033 uint32_t *record_cnt,
2034 uint32_t *data_cnt)
2035{
2036 struct xlog_in_core *iclog = *iclogp;
2037 struct xlog *log = iclog->ic_log;
2038 int error;
2039
2040 spin_lock(&log->l_icloglock);
2041 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2042 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2043 error = xlog_state_release_iclog(log, iclog, ticket);
2044 spin_unlock(&log->l_icloglock);
2045 if (error)
2046 return error;
2047
2048 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2049 log_offset);
2050 if (error)
2051 return error;
2052 *record_cnt = 0;
2053 *data_cnt = 0;
2054 *iclogp = iclog;
2055 return 0;
2056}
2057
2058/*
2059 * Write log vectors into a single iclog which is smaller than the current chain
2060 * length. We write until we cannot fit a full record into the remaining space
2061 * and then stop. We return the log vector that is to be written that cannot
2062 * wholly fit in the iclog.
2063 */
2064static int
2065xlog_write_partial(
2066 struct xfs_log_vec *lv,
2067 struct xlog_ticket *ticket,
2068 struct xlog_in_core **iclogp,
2069 uint32_t *log_offset,
2070 uint32_t *len,
2071 uint32_t *record_cnt,
2072 uint32_t *data_cnt)
2073{
2074 struct xlog_in_core *iclog = *iclogp;
2075 struct xlog_op_header *ophdr;
2076 int index = 0;
2077 uint32_t rlen;
2078 int error;
2079
2080 /* walk the logvec, copying until we run out of space in the iclog */
2081 for (index = 0; index < lv->lv_niovecs; index++) {
2082 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2083 uint32_t reg_offset = 0;
2084
2085 /*
2086 * The first region of a continuation must have a non-zero
2087 * length otherwise log recovery will just skip over it and
2088 * start recovering from the next opheader it finds. Because we
2089 * mark the next opheader as a continuation, recovery will then
2090 * incorrectly add the continuation to the previous region and
2091 * that breaks stuff.
2092 *
2093 * Hence if there isn't space for region data after the
2094 * opheader, then we need to start afresh with a new iclog.
2095 */
2096 if (iclog->ic_size - *log_offset <=
2097 sizeof(struct xlog_op_header)) {
2098 error = xlog_write_get_more_iclog_space(ticket,
2099 &iclog, log_offset, *len, record_cnt,
2100 data_cnt);
2101 if (error)
2102 return error;
2103 }
2104
2105 ophdr = reg->i_addr;
2106 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2107
2108 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2109 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2110 if (rlen != reg->i_len)
2111 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2112
2113 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2114 rlen, len, record_cnt, data_cnt);
2115
2116 /* If we wrote the whole region, move to the next. */
2117 if (rlen == reg->i_len)
2118 continue;
2119
2120 /*
2121 * We now have a partially written iovec, but it can span
2122 * multiple iclogs so we loop here. First we release the iclog
2123 * we currently have, then we get a new iclog and add a new
2124 * opheader. Then we continue copying from where we were until
2125 * we either complete the iovec or fill the iclog. If we
2126 * complete the iovec, then we increment the index and go right
2127 * back to the top of the outer loop. if we fill the iclog, we
2128 * run the inner loop again.
2129 *
2130 * This is complicated by the tail of a region using all the
2131 * space in an iclog and hence requiring us to release the iclog
2132 * and get a new one before returning to the outer loop. We must
2133 * always guarantee that we exit this inner loop with at least
2134 * space for log transaction opheaders left in the current
2135 * iclog, hence we cannot just terminate the loop at the end
2136 * of the of the continuation. So we loop while there is no
2137 * space left in the current iclog, and check for the end of the
2138 * continuation after getting a new iclog.
2139 */
2140 do {
2141 /*
2142 * Ensure we include the continuation opheader in the
2143 * space we need in the new iclog by adding that size
2144 * to the length we require. This continuation opheader
2145 * needs to be accounted to the ticket as the space it
2146 * consumes hasn't been accounted to the lv we are
2147 * writing.
2148 */
2149 error = xlog_write_get_more_iclog_space(ticket,
2150 &iclog, log_offset,
2151 *len + sizeof(struct xlog_op_header),
2152 record_cnt, data_cnt);
2153 if (error)
2154 return error;
2155
2156 ophdr = iclog->ic_datap + *log_offset;
2157 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2158 ophdr->oh_clientid = XFS_TRANSACTION;
2159 ophdr->oh_res2 = 0;
2160 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2161
2162 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2163 *log_offset += sizeof(struct xlog_op_header);
2164 *data_cnt += sizeof(struct xlog_op_header);
2165
2166 /*
2167 * If rlen fits in the iclog, then end the region
2168 * continuation. Otherwise we're going around again.
2169 */
2170 reg_offset += rlen;
2171 rlen = reg->i_len - reg_offset;
2172 if (rlen <= iclog->ic_size - *log_offset)
2173 ophdr->oh_flags |= XLOG_END_TRANS;
2174 else
2175 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2176
2177 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2178 ophdr->oh_len = cpu_to_be32(rlen);
2179
2180 xlog_write_iovec(iclog, log_offset,
2181 reg->i_addr + reg_offset,
2182 rlen, len, record_cnt, data_cnt);
2183
2184 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2185 }
2186
2187 /*
2188 * No more iovecs remain in this logvec so return the next log vec to
2189 * the caller so it can go back to fast path copying.
2190 */
2191 *iclogp = iclog;
2192 return 0;
2193}
2194
2195/*
2196 * Write some region out to in-core log
2197 *
2198 * This will be called when writing externally provided regions or when
2199 * writing out a commit record for a given transaction.
2200 *
2201 * General algorithm:
2202 * 1. Find total length of this write. This may include adding to the
2203 * lengths passed in.
2204 * 2. Check whether we violate the tickets reservation.
2205 * 3. While writing to this iclog
2206 * A. Reserve as much space in this iclog as can get
2207 * B. If this is first write, save away start lsn
2208 * C. While writing this region:
2209 * 1. If first write of transaction, write start record
2210 * 2. Write log operation header (header per region)
2211 * 3. Find out if we can fit entire region into this iclog
2212 * 4. Potentially, verify destination memcpy ptr
2213 * 5. Memcpy (partial) region
2214 * 6. If partial copy, release iclog; otherwise, continue
2215 * copying more regions into current iclog
2216 * 4. Mark want sync bit (in simulation mode)
2217 * 5. Release iclog for potential flush to on-disk log.
2218 *
2219 * ERRORS:
2220 * 1. Panic if reservation is overrun. This should never happen since
2221 * reservation amounts are generated internal to the filesystem.
2222 * NOTES:
2223 * 1. Tickets are single threaded data structures.
2224 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2225 * syncing routine. When a single log_write region needs to span
2226 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2227 * on all log operation writes which don't contain the end of the
2228 * region. The XLOG_END_TRANS bit is used for the in-core log
2229 * operation which contains the end of the continued log_write region.
2230 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2231 * we don't really know exactly how much space will be used. As a result,
2232 * we don't update ic_offset until the end when we know exactly how many
2233 * bytes have been written out.
2234 */
2235int
2236xlog_write(
2237 struct xlog *log,
2238 struct xfs_cil_ctx *ctx,
2239 struct list_head *lv_chain,
2240 struct xlog_ticket *ticket,
2241 uint32_t len)
2242
2243{
2244 struct xlog_in_core *iclog = NULL;
2245 struct xfs_log_vec *lv;
2246 uint32_t record_cnt = 0;
2247 uint32_t data_cnt = 0;
2248 int error = 0;
2249 int log_offset;
2250
2251 if (ticket->t_curr_res < 0) {
2252 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2253 "ctx ticket reservation ran out. Need to up reservation");
2254 xlog_print_tic_res(log->l_mp, ticket);
2255 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2256 }
2257
2258 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2259 &log_offset);
2260 if (error)
2261 return error;
2262
2263 ASSERT(log_offset <= iclog->ic_size - 1);
2264
2265 /*
2266 * If we have a context pointer, pass it the first iclog we are
2267 * writing to so it can record state needed for iclog write
2268 * ordering.
2269 */
2270 if (ctx)
2271 xlog_cil_set_ctx_write_state(ctx, iclog);
2272
2273 list_for_each_entry(lv, lv_chain, lv_list) {
2274 /*
2275 * If the entire log vec does not fit in the iclog, punt it to
2276 * the partial copy loop which can handle this case.
2277 */
2278 if (lv->lv_niovecs &&
2279 lv->lv_bytes > iclog->ic_size - log_offset) {
2280 error = xlog_write_partial(lv, ticket, &iclog,
2281 &log_offset, &len, &record_cnt,
2282 &data_cnt);
2283 if (error) {
2284 /*
2285 * We have no iclog to release, so just return
2286 * the error immediately.
2287 */
2288 return error;
2289 }
2290 } else {
2291 xlog_write_full(lv, ticket, iclog, &log_offset,
2292 &len, &record_cnt, &data_cnt);
2293 }
2294 }
2295 ASSERT(len == 0);
2296
2297 /*
2298 * We've already been guaranteed that the last writes will fit inside
2299 * the current iclog, and hence it will already have the space used by
2300 * those writes accounted to it. Hence we do not need to update the
2301 * iclog with the number of bytes written here.
2302 */
2303 spin_lock(&log->l_icloglock);
2304 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2305 error = xlog_state_release_iclog(log, iclog, ticket);
2306 spin_unlock(&log->l_icloglock);
2307
2308 return error;
2309}
2310
2311static void
2312xlog_state_activate_iclog(
2313 struct xlog_in_core *iclog,
2314 int *iclogs_changed)
2315{
2316 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2317 trace_xlog_iclog_activate(iclog, _RET_IP_);
2318
2319 /*
2320 * If the number of ops in this iclog indicate it just contains the
2321 * dummy transaction, we can change state into IDLE (the second time
2322 * around). Otherwise we should change the state into NEED a dummy.
2323 * We don't need to cover the dummy.
2324 */
2325 if (*iclogs_changed == 0 &&
2326 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2327 *iclogs_changed = 1;
2328 } else {
2329 /*
2330 * We have two dirty iclogs so start over. This could also be
2331 * num of ops indicating this is not the dummy going out.
2332 */
2333 *iclogs_changed = 2;
2334 }
2335
2336 iclog->ic_state = XLOG_STATE_ACTIVE;
2337 iclog->ic_offset = 0;
2338 iclog->ic_header.h_num_logops = 0;
2339 memset(iclog->ic_header.h_cycle_data, 0,
2340 sizeof(iclog->ic_header.h_cycle_data));
2341 iclog->ic_header.h_lsn = 0;
2342 iclog->ic_header.h_tail_lsn = 0;
2343}
2344
2345/*
2346 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2347 * ACTIVE after iclog I/O has completed.
2348 */
2349static void
2350xlog_state_activate_iclogs(
2351 struct xlog *log,
2352 int *iclogs_changed)
2353{
2354 struct xlog_in_core *iclog = log->l_iclog;
2355
2356 do {
2357 if (iclog->ic_state == XLOG_STATE_DIRTY)
2358 xlog_state_activate_iclog(iclog, iclogs_changed);
2359 /*
2360 * The ordering of marking iclogs ACTIVE must be maintained, so
2361 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2362 */
2363 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2364 break;
2365 } while ((iclog = iclog->ic_next) != log->l_iclog);
2366}
2367
2368static int
2369xlog_covered_state(
2370 int prev_state,
2371 int iclogs_changed)
2372{
2373 /*
2374 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2375 * wrote the first covering record (DONE). We go to IDLE if we just
2376 * wrote the second covering record (DONE2) and remain in IDLE until a
2377 * non-covering write occurs.
2378 */
2379 switch (prev_state) {
2380 case XLOG_STATE_COVER_IDLE:
2381 if (iclogs_changed == 1)
2382 return XLOG_STATE_COVER_IDLE;
2383 fallthrough;
2384 case XLOG_STATE_COVER_NEED:
2385 case XLOG_STATE_COVER_NEED2:
2386 break;
2387 case XLOG_STATE_COVER_DONE:
2388 if (iclogs_changed == 1)
2389 return XLOG_STATE_COVER_NEED2;
2390 break;
2391 case XLOG_STATE_COVER_DONE2:
2392 if (iclogs_changed == 1)
2393 return XLOG_STATE_COVER_IDLE;
2394 break;
2395 default:
2396 ASSERT(0);
2397 }
2398
2399 return XLOG_STATE_COVER_NEED;
2400}
2401
2402STATIC void
2403xlog_state_clean_iclog(
2404 struct xlog *log,
2405 struct xlog_in_core *dirty_iclog)
2406{
2407 int iclogs_changed = 0;
2408
2409 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2410
2411 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2412
2413 xlog_state_activate_iclogs(log, &iclogs_changed);
2414 wake_up_all(&dirty_iclog->ic_force_wait);
2415
2416 if (iclogs_changed) {
2417 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2418 iclogs_changed);
2419 }
2420}
2421
2422STATIC xfs_lsn_t
2423xlog_get_lowest_lsn(
2424 struct xlog *log)
2425{
2426 struct xlog_in_core *iclog = log->l_iclog;
2427 xfs_lsn_t lowest_lsn = 0, lsn;
2428
2429 do {
2430 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2431 iclog->ic_state == XLOG_STATE_DIRTY)
2432 continue;
2433
2434 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2435 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2436 lowest_lsn = lsn;
2437 } while ((iclog = iclog->ic_next) != log->l_iclog);
2438
2439 return lowest_lsn;
2440}
2441
2442/*
2443 * Return true if we need to stop processing, false to continue to the next
2444 * iclog. The caller will need to run callbacks if the iclog is returned in the
2445 * XLOG_STATE_CALLBACK state.
2446 */
2447static bool
2448xlog_state_iodone_process_iclog(
2449 struct xlog *log,
2450 struct xlog_in_core *iclog)
2451{
2452 xfs_lsn_t lowest_lsn;
2453 xfs_lsn_t header_lsn;
2454
2455 switch (iclog->ic_state) {
2456 case XLOG_STATE_ACTIVE:
2457 case XLOG_STATE_DIRTY:
2458 /*
2459 * Skip all iclogs in the ACTIVE & DIRTY states:
2460 */
2461 return false;
2462 case XLOG_STATE_DONE_SYNC:
2463 /*
2464 * Now that we have an iclog that is in the DONE_SYNC state, do
2465 * one more check here to see if we have chased our tail around.
2466 * If this is not the lowest lsn iclog, then we will leave it
2467 * for another completion to process.
2468 */
2469 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2470 lowest_lsn = xlog_get_lowest_lsn(log);
2471 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2472 return false;
2473 /*
2474 * If there are no callbacks on this iclog, we can mark it clean
2475 * immediately and return. Otherwise we need to run the
2476 * callbacks.
2477 */
2478 if (list_empty(&iclog->ic_callbacks)) {
2479 xlog_state_clean_iclog(log, iclog);
2480 return false;
2481 }
2482 trace_xlog_iclog_callback(iclog, _RET_IP_);
2483 iclog->ic_state = XLOG_STATE_CALLBACK;
2484 return false;
2485 default:
2486 /*
2487 * Can only perform callbacks in order. Since this iclog is not
2488 * in the DONE_SYNC state, we skip the rest and just try to
2489 * clean up.
2490 */
2491 return true;
2492 }
2493}
2494
2495/*
2496 * Loop over all the iclogs, running attached callbacks on them. Return true if
2497 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2498 * to handle transient shutdown state here at all because
2499 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2500 * cleanup of the callbacks.
2501 */
2502static bool
2503xlog_state_do_iclog_callbacks(
2504 struct xlog *log)
2505 __releases(&log->l_icloglock)
2506 __acquires(&log->l_icloglock)
2507{
2508 struct xlog_in_core *first_iclog = log->l_iclog;
2509 struct xlog_in_core *iclog = first_iclog;
2510 bool ran_callback = false;
2511
2512 do {
2513 LIST_HEAD(cb_list);
2514
2515 if (xlog_state_iodone_process_iclog(log, iclog))
2516 break;
2517 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2518 iclog = iclog->ic_next;
2519 continue;
2520 }
2521 list_splice_init(&iclog->ic_callbacks, &cb_list);
2522 spin_unlock(&log->l_icloglock);
2523
2524 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2525 xlog_cil_process_committed(&cb_list);
2526 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2527 ran_callback = true;
2528
2529 spin_lock(&log->l_icloglock);
2530 xlog_state_clean_iclog(log, iclog);
2531 iclog = iclog->ic_next;
2532 } while (iclog != first_iclog);
2533
2534 return ran_callback;
2535}
2536
2537
2538/*
2539 * Loop running iclog completion callbacks until there are no more iclogs in a
2540 * state that can run callbacks.
2541 */
2542STATIC void
2543xlog_state_do_callback(
2544 struct xlog *log)
2545{
2546 int flushcnt = 0;
2547 int repeats = 0;
2548
2549 spin_lock(&log->l_icloglock);
2550 while (xlog_state_do_iclog_callbacks(log)) {
2551 if (xlog_is_shutdown(log))
2552 break;
2553
2554 if (++repeats > 5000) {
2555 flushcnt += repeats;
2556 repeats = 0;
2557 xfs_warn(log->l_mp,
2558 "%s: possible infinite loop (%d iterations)",
2559 __func__, flushcnt);
2560 }
2561 }
2562
2563 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2564 wake_up_all(&log->l_flush_wait);
2565
2566 spin_unlock(&log->l_icloglock);
2567}
2568
2569
2570/*
2571 * Finish transitioning this iclog to the dirty state.
2572 *
2573 * Callbacks could take time, so they are done outside the scope of the
2574 * global state machine log lock.
2575 */
2576STATIC void
2577xlog_state_done_syncing(
2578 struct xlog_in_core *iclog)
2579{
2580 struct xlog *log = iclog->ic_log;
2581
2582 spin_lock(&log->l_icloglock);
2583 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2584 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2585
2586 /*
2587 * If we got an error, either on the first buffer, or in the case of
2588 * split log writes, on the second, we shut down the file system and
2589 * no iclogs should ever be attempted to be written to disk again.
2590 */
2591 if (!xlog_is_shutdown(log)) {
2592 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2593 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2594 }
2595
2596 /*
2597 * Someone could be sleeping prior to writing out the next
2598 * iclog buffer, we wake them all, one will get to do the
2599 * I/O, the others get to wait for the result.
2600 */
2601 wake_up_all(&iclog->ic_write_wait);
2602 spin_unlock(&log->l_icloglock);
2603 xlog_state_do_callback(log);
2604}
2605
2606/*
2607 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2608 * sleep. We wait on the flush queue on the head iclog as that should be
2609 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2610 * we will wait here and all new writes will sleep until a sync completes.
2611 *
2612 * The in-core logs are used in a circular fashion. They are not used
2613 * out-of-order even when an iclog past the head is free.
2614 *
2615 * return:
2616 * * log_offset where xlog_write() can start writing into the in-core
2617 * log's data space.
2618 * * in-core log pointer to which xlog_write() should write.
2619 * * boolean indicating this is a continued write to an in-core log.
2620 * If this is the last write, then the in-core log's offset field
2621 * needs to be incremented, depending on the amount of data which
2622 * is copied.
2623 */
2624STATIC int
2625xlog_state_get_iclog_space(
2626 struct xlog *log,
2627 int len,
2628 struct xlog_in_core **iclogp,
2629 struct xlog_ticket *ticket,
2630 int *logoffsetp)
2631{
2632 int log_offset;
2633 xlog_rec_header_t *head;
2634 xlog_in_core_t *iclog;
2635
2636restart:
2637 spin_lock(&log->l_icloglock);
2638 if (xlog_is_shutdown(log)) {
2639 spin_unlock(&log->l_icloglock);
2640 return -EIO;
2641 }
2642
2643 iclog = log->l_iclog;
2644 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2645 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2646
2647 /* Wait for log writes to have flushed */
2648 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2649 goto restart;
2650 }
2651
2652 head = &iclog->ic_header;
2653
2654 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2655 log_offset = iclog->ic_offset;
2656
2657 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2658
2659 /* On the 1st write to an iclog, figure out lsn. This works
2660 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2661 * committing to. If the offset is set, that's how many blocks
2662 * must be written.
2663 */
2664 if (log_offset == 0) {
2665 ticket->t_curr_res -= log->l_iclog_hsize;
2666 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2667 head->h_lsn = cpu_to_be64(
2668 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2669 ASSERT(log->l_curr_block >= 0);
2670 }
2671
2672 /* If there is enough room to write everything, then do it. Otherwise,
2673 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2674 * bit is on, so this will get flushed out. Don't update ic_offset
2675 * until you know exactly how many bytes get copied. Therefore, wait
2676 * until later to update ic_offset.
2677 *
2678 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2679 * can fit into remaining data section.
2680 */
2681 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2682 int error = 0;
2683
2684 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2685
2686 /*
2687 * If we are the only one writing to this iclog, sync it to
2688 * disk. We need to do an atomic compare and decrement here to
2689 * avoid racing with concurrent atomic_dec_and_lock() calls in
2690 * xlog_state_release_iclog() when there is more than one
2691 * reference to the iclog.
2692 */
2693 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2694 error = xlog_state_release_iclog(log, iclog, ticket);
2695 spin_unlock(&log->l_icloglock);
2696 if (error)
2697 return error;
2698 goto restart;
2699 }
2700
2701 /* Do we have enough room to write the full amount in the remainder
2702 * of this iclog? Or must we continue a write on the next iclog and
2703 * mark this iclog as completely taken? In the case where we switch
2704 * iclogs (to mark it taken), this particular iclog will release/sync
2705 * to disk in xlog_write().
2706 */
2707 if (len <= iclog->ic_size - iclog->ic_offset)
2708 iclog->ic_offset += len;
2709 else
2710 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2711 *iclogp = iclog;
2712
2713 ASSERT(iclog->ic_offset <= iclog->ic_size);
2714 spin_unlock(&log->l_icloglock);
2715
2716 *logoffsetp = log_offset;
2717 return 0;
2718}
2719
2720/*
2721 * The first cnt-1 times a ticket goes through here we don't need to move the
2722 * grant write head because the permanent reservation has reserved cnt times the
2723 * unit amount. Release part of current permanent unit reservation and reset
2724 * current reservation to be one units worth. Also move grant reservation head
2725 * forward.
2726 */
2727void
2728xfs_log_ticket_regrant(
2729 struct xlog *log,
2730 struct xlog_ticket *ticket)
2731{
2732 trace_xfs_log_ticket_regrant(log, ticket);
2733
2734 if (ticket->t_cnt > 0)
2735 ticket->t_cnt--;
2736
2737 xlog_grant_sub_space(&log->l_reserve_head, ticket->t_curr_res);
2738 xlog_grant_sub_space(&log->l_write_head, ticket->t_curr_res);
2739 ticket->t_curr_res = ticket->t_unit_res;
2740
2741 trace_xfs_log_ticket_regrant_sub(log, ticket);
2742
2743 /* just return if we still have some of the pre-reserved space */
2744 if (!ticket->t_cnt) {
2745 xlog_grant_add_space(&log->l_reserve_head, ticket->t_unit_res);
2746 trace_xfs_log_ticket_regrant_exit(log, ticket);
2747
2748 ticket->t_curr_res = ticket->t_unit_res;
2749 }
2750
2751 xfs_log_ticket_put(ticket);
2752}
2753
2754/*
2755 * Give back the space left from a reservation.
2756 *
2757 * All the information we need to make a correct determination of space left
2758 * is present. For non-permanent reservations, things are quite easy. The
2759 * count should have been decremented to zero. We only need to deal with the
2760 * space remaining in the current reservation part of the ticket. If the
2761 * ticket contains a permanent reservation, there may be left over space which
2762 * needs to be released. A count of N means that N-1 refills of the current
2763 * reservation can be done before we need to ask for more space. The first
2764 * one goes to fill up the first current reservation. Once we run out of
2765 * space, the count will stay at zero and the only space remaining will be
2766 * in the current reservation field.
2767 */
2768void
2769xfs_log_ticket_ungrant(
2770 struct xlog *log,
2771 struct xlog_ticket *ticket)
2772{
2773 int bytes;
2774
2775 trace_xfs_log_ticket_ungrant(log, ticket);
2776
2777 if (ticket->t_cnt > 0)
2778 ticket->t_cnt--;
2779
2780 trace_xfs_log_ticket_ungrant_sub(log, ticket);
2781
2782 /*
2783 * If this is a permanent reservation ticket, we may be able to free
2784 * up more space based on the remaining count.
2785 */
2786 bytes = ticket->t_curr_res;
2787 if (ticket->t_cnt > 0) {
2788 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2789 bytes += ticket->t_unit_res*ticket->t_cnt;
2790 }
2791
2792 xlog_grant_sub_space(&log->l_reserve_head, bytes);
2793 xlog_grant_sub_space(&log->l_write_head, bytes);
2794
2795 trace_xfs_log_ticket_ungrant_exit(log, ticket);
2796
2797 xfs_log_space_wake(log->l_mp);
2798 xfs_log_ticket_put(ticket);
2799}
2800
2801/*
2802 * This routine will mark the current iclog in the ring as WANT_SYNC and move
2803 * the current iclog pointer to the next iclog in the ring.
2804 */
2805void
2806xlog_state_switch_iclogs(
2807 struct xlog *log,
2808 struct xlog_in_core *iclog,
2809 int eventual_size)
2810{
2811 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2812 assert_spin_locked(&log->l_icloglock);
2813 trace_xlog_iclog_switch(iclog, _RET_IP_);
2814
2815 if (!eventual_size)
2816 eventual_size = iclog->ic_offset;
2817 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2818 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2819 log->l_prev_block = log->l_curr_block;
2820 log->l_prev_cycle = log->l_curr_cycle;
2821
2822 /* roll log?: ic_offset changed later */
2823 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2824
2825 /* Round up to next log-sunit */
2826 if (log->l_iclog_roundoff > BBSIZE) {
2827 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
2828 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2829 }
2830
2831 if (log->l_curr_block >= log->l_logBBsize) {
2832 /*
2833 * Rewind the current block before the cycle is bumped to make
2834 * sure that the combined LSN never transiently moves forward
2835 * when the log wraps to the next cycle. This is to support the
2836 * unlocked sample of these fields from xlog_valid_lsn(). Most
2837 * other cases should acquire l_icloglock.
2838 */
2839 log->l_curr_block -= log->l_logBBsize;
2840 ASSERT(log->l_curr_block >= 0);
2841 smp_wmb();
2842 log->l_curr_cycle++;
2843 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2844 log->l_curr_cycle++;
2845 }
2846 ASSERT(iclog == log->l_iclog);
2847 log->l_iclog = iclog->ic_next;
2848}
2849
2850/*
2851 * Force the iclog to disk and check if the iclog has been completed before
2852 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
2853 * pmem) or fast async storage because we drop the icloglock to issue the IO.
2854 * If completion has already occurred, tell the caller so that it can avoid an
2855 * unnecessary wait on the iclog.
2856 */
2857static int
2858xlog_force_and_check_iclog(
2859 struct xlog_in_core *iclog,
2860 bool *completed)
2861{
2862 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2863 int error;
2864
2865 *completed = false;
2866 error = xlog_force_iclog(iclog);
2867 if (error)
2868 return error;
2869
2870 /*
2871 * If the iclog has already been completed and reused the header LSN
2872 * will have been rewritten by completion
2873 */
2874 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
2875 *completed = true;
2876 return 0;
2877}
2878
2879/*
2880 * Write out all data in the in-core log as of this exact moment in time.
2881 *
2882 * Data may be written to the in-core log during this call. However,
2883 * we don't guarantee this data will be written out. A change from past
2884 * implementation means this routine will *not* write out zero length LRs.
2885 *
2886 * Basically, we try and perform an intelligent scan of the in-core logs.
2887 * If we determine there is no flushable data, we just return. There is no
2888 * flushable data if:
2889 *
2890 * 1. the current iclog is active and has no data; the previous iclog
2891 * is in the active or dirty state.
2892 * 2. the current iclog is drity, and the previous iclog is in the
2893 * active or dirty state.
2894 *
2895 * We may sleep if:
2896 *
2897 * 1. the current iclog is not in the active nor dirty state.
2898 * 2. the current iclog dirty, and the previous iclog is not in the
2899 * active nor dirty state.
2900 * 3. the current iclog is active, and there is another thread writing
2901 * to this particular iclog.
2902 * 4. a) the current iclog is active and has no other writers
2903 * b) when we return from flushing out this iclog, it is still
2904 * not in the active nor dirty state.
2905 */
2906int
2907xfs_log_force(
2908 struct xfs_mount *mp,
2909 uint flags)
2910{
2911 struct xlog *log = mp->m_log;
2912 struct xlog_in_core *iclog;
2913
2914 XFS_STATS_INC(mp, xs_log_force);
2915 trace_xfs_log_force(mp, 0, _RET_IP_);
2916
2917 xlog_cil_force(log);
2918
2919 spin_lock(&log->l_icloglock);
2920 if (xlog_is_shutdown(log))
2921 goto out_error;
2922
2923 iclog = log->l_iclog;
2924 trace_xlog_iclog_force(iclog, _RET_IP_);
2925
2926 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2927 (iclog->ic_state == XLOG_STATE_ACTIVE &&
2928 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
2929 /*
2930 * If the head is dirty or (active and empty), then we need to
2931 * look at the previous iclog.
2932 *
2933 * If the previous iclog is active or dirty we are done. There
2934 * is nothing to sync out. Otherwise, we attach ourselves to the
2935 * previous iclog and go to sleep.
2936 */
2937 iclog = iclog->ic_prev;
2938 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2939 if (atomic_read(&iclog->ic_refcnt) == 0) {
2940 /* We have exclusive access to this iclog. */
2941 bool completed;
2942
2943 if (xlog_force_and_check_iclog(iclog, &completed))
2944 goto out_error;
2945
2946 if (completed)
2947 goto out_unlock;
2948 } else {
2949 /*
2950 * Someone else is still writing to this iclog, so we
2951 * need to ensure that when they release the iclog it
2952 * gets synced immediately as we may be waiting on it.
2953 */
2954 xlog_state_switch_iclogs(log, iclog, 0);
2955 }
2956 }
2957
2958 /*
2959 * The iclog we are about to wait on may contain the checkpoint pushed
2960 * by the above xlog_cil_force() call, but it may not have been pushed
2961 * to disk yet. Like the ACTIVE case above, we need to make sure caches
2962 * are flushed when this iclog is written.
2963 */
2964 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
2965 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
2966
2967 if (flags & XFS_LOG_SYNC)
2968 return xlog_wait_on_iclog(iclog);
2969out_unlock:
2970 spin_unlock(&log->l_icloglock);
2971 return 0;
2972out_error:
2973 spin_unlock(&log->l_icloglock);
2974 return -EIO;
2975}
2976
2977/*
2978 * Force the log to a specific LSN.
2979 *
2980 * If an iclog with that lsn can be found:
2981 * If it is in the DIRTY state, just return.
2982 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2983 * state and go to sleep or return.
2984 * If it is in any other state, go to sleep or return.
2985 *
2986 * Synchronous forces are implemented with a wait queue. All callers trying
2987 * to force a given lsn to disk must wait on the queue attached to the
2988 * specific in-core log. When given in-core log finally completes its write
2989 * to disk, that thread will wake up all threads waiting on the queue.
2990 */
2991static int
2992xlog_force_lsn(
2993 struct xlog *log,
2994 xfs_lsn_t lsn,
2995 uint flags,
2996 int *log_flushed,
2997 bool already_slept)
2998{
2999 struct xlog_in_core *iclog;
3000 bool completed;
3001
3002 spin_lock(&log->l_icloglock);
3003 if (xlog_is_shutdown(log))
3004 goto out_error;
3005
3006 iclog = log->l_iclog;
3007 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3008 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3009 iclog = iclog->ic_next;
3010 if (iclog == log->l_iclog)
3011 goto out_unlock;
3012 }
3013
3014 switch (iclog->ic_state) {
3015 case XLOG_STATE_ACTIVE:
3016 /*
3017 * We sleep here if we haven't already slept (e.g. this is the
3018 * first time we've looked at the correct iclog buf) and the
3019 * buffer before us is going to be sync'ed. The reason for this
3020 * is that if we are doing sync transactions here, by waiting
3021 * for the previous I/O to complete, we can allow a few more
3022 * transactions into this iclog before we close it down.
3023 *
3024 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3025 * refcnt so we can release the log (which drops the ref count).
3026 * The state switch keeps new transaction commits from using
3027 * this buffer. When the current commits finish writing into
3028 * the buffer, the refcount will drop to zero and the buffer
3029 * will go out then.
3030 */
3031 if (!already_slept &&
3032 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3033 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3034 xlog_wait(&iclog->ic_prev->ic_write_wait,
3035 &log->l_icloglock);
3036 return -EAGAIN;
3037 }
3038 if (xlog_force_and_check_iclog(iclog, &completed))
3039 goto out_error;
3040 if (log_flushed)
3041 *log_flushed = 1;
3042 if (completed)
3043 goto out_unlock;
3044 break;
3045 case XLOG_STATE_WANT_SYNC:
3046 /*
3047 * This iclog may contain the checkpoint pushed by the
3048 * xlog_cil_force_seq() call, but there are other writers still
3049 * accessing it so it hasn't been pushed to disk yet. Like the
3050 * ACTIVE case above, we need to make sure caches are flushed
3051 * when this iclog is written.
3052 */
3053 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3054 break;
3055 default:
3056 /*
3057 * The entire checkpoint was written by the CIL force and is on
3058 * its way to disk already. It will be stable when it
3059 * completes, so we don't need to manipulate caches here at all.
3060 * We just need to wait for completion if necessary.
3061 */
3062 break;
3063 }
3064
3065 if (flags & XFS_LOG_SYNC)
3066 return xlog_wait_on_iclog(iclog);
3067out_unlock:
3068 spin_unlock(&log->l_icloglock);
3069 return 0;
3070out_error:
3071 spin_unlock(&log->l_icloglock);
3072 return -EIO;
3073}
3074
3075/*
3076 * Force the log to a specific checkpoint sequence.
3077 *
3078 * First force the CIL so that all the required changes have been flushed to the
3079 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3080 * the iclog that needs to be flushed to stable storage. If the caller needs
3081 * a synchronous log force, we will wait on the iclog with the LSN returned by
3082 * xlog_cil_force_seq() to be completed.
3083 */
3084int
3085xfs_log_force_seq(
3086 struct xfs_mount *mp,
3087 xfs_csn_t seq,
3088 uint flags,
3089 int *log_flushed)
3090{
3091 struct xlog *log = mp->m_log;
3092 xfs_lsn_t lsn;
3093 int ret;
3094 ASSERT(seq != 0);
3095
3096 XFS_STATS_INC(mp, xs_log_force);
3097 trace_xfs_log_force(mp, seq, _RET_IP_);
3098
3099 lsn = xlog_cil_force_seq(log, seq);
3100 if (lsn == NULLCOMMITLSN)
3101 return 0;
3102
3103 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3104 if (ret == -EAGAIN) {
3105 XFS_STATS_INC(mp, xs_log_force_sleep);
3106 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3107 }
3108 return ret;
3109}
3110
3111/*
3112 * Free a used ticket when its refcount falls to zero.
3113 */
3114void
3115xfs_log_ticket_put(
3116 xlog_ticket_t *ticket)
3117{
3118 ASSERT(atomic_read(&ticket->t_ref) > 0);
3119 if (atomic_dec_and_test(&ticket->t_ref))
3120 kmem_cache_free(xfs_log_ticket_cache, ticket);
3121}
3122
3123xlog_ticket_t *
3124xfs_log_ticket_get(
3125 xlog_ticket_t *ticket)
3126{
3127 ASSERT(atomic_read(&ticket->t_ref) > 0);
3128 atomic_inc(&ticket->t_ref);
3129 return ticket;
3130}
3131
3132/*
3133 * Figure out the total log space unit (in bytes) that would be
3134 * required for a log ticket.
3135 */
3136static int
3137xlog_calc_unit_res(
3138 struct xlog *log,
3139 int unit_bytes,
3140 int *niclogs)
3141{
3142 int iclog_space;
3143 uint num_headers;
3144
3145 /*
3146 * Permanent reservations have up to 'cnt'-1 active log operations
3147 * in the log. A unit in this case is the amount of space for one
3148 * of these log operations. Normal reservations have a cnt of 1
3149 * and their unit amount is the total amount of space required.
3150 *
3151 * The following lines of code account for non-transaction data
3152 * which occupy space in the on-disk log.
3153 *
3154 * Normal form of a transaction is:
3155 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3156 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3157 *
3158 * We need to account for all the leadup data and trailer data
3159 * around the transaction data.
3160 * And then we need to account for the worst case in terms of using
3161 * more space.
3162 * The worst case will happen if:
3163 * - the placement of the transaction happens to be such that the
3164 * roundoff is at its maximum
3165 * - the transaction data is synced before the commit record is synced
3166 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3167 * Therefore the commit record is in its own Log Record.
3168 * This can happen as the commit record is called with its
3169 * own region to xlog_write().
3170 * This then means that in the worst case, roundoff can happen for
3171 * the commit-rec as well.
3172 * The commit-rec is smaller than padding in this scenario and so it is
3173 * not added separately.
3174 */
3175
3176 /* for trans header */
3177 unit_bytes += sizeof(xlog_op_header_t);
3178 unit_bytes += sizeof(xfs_trans_header_t);
3179
3180 /* for start-rec */
3181 unit_bytes += sizeof(xlog_op_header_t);
3182
3183 /*
3184 * for LR headers - the space for data in an iclog is the size minus
3185 * the space used for the headers. If we use the iclog size, then we
3186 * undercalculate the number of headers required.
3187 *
3188 * Furthermore - the addition of op headers for split-recs might
3189 * increase the space required enough to require more log and op
3190 * headers, so take that into account too.
3191 *
3192 * IMPORTANT: This reservation makes the assumption that if this
3193 * transaction is the first in an iclog and hence has the LR headers
3194 * accounted to it, then the remaining space in the iclog is
3195 * exclusively for this transaction. i.e. if the transaction is larger
3196 * than the iclog, it will be the only thing in that iclog.
3197 * Fundamentally, this means we must pass the entire log vector to
3198 * xlog_write to guarantee this.
3199 */
3200 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3201 num_headers = howmany(unit_bytes, iclog_space);
3202
3203 /* for split-recs - ophdrs added when data split over LRs */
3204 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3205
3206 /* add extra header reservations if we overrun */
3207 while (!num_headers ||
3208 howmany(unit_bytes, iclog_space) > num_headers) {
3209 unit_bytes += sizeof(xlog_op_header_t);
3210 num_headers++;
3211 }
3212 unit_bytes += log->l_iclog_hsize * num_headers;
3213
3214 /* for commit-rec LR header - note: padding will subsume the ophdr */
3215 unit_bytes += log->l_iclog_hsize;
3216
3217 /* roundoff padding for transaction data and one for commit record */
3218 unit_bytes += 2 * log->l_iclog_roundoff;
3219
3220 if (niclogs)
3221 *niclogs = num_headers;
3222 return unit_bytes;
3223}
3224
3225int
3226xfs_log_calc_unit_res(
3227 struct xfs_mount *mp,
3228 int unit_bytes)
3229{
3230 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3231}
3232
3233/*
3234 * Allocate and initialise a new log ticket.
3235 */
3236struct xlog_ticket *
3237xlog_ticket_alloc(
3238 struct xlog *log,
3239 int unit_bytes,
3240 int cnt,
3241 bool permanent)
3242{
3243 struct xlog_ticket *tic;
3244 int unit_res;
3245
3246 tic = kmem_cache_zalloc(xfs_log_ticket_cache,
3247 GFP_KERNEL | __GFP_NOFAIL);
3248
3249 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3250
3251 atomic_set(&tic->t_ref, 1);
3252 tic->t_task = current;
3253 INIT_LIST_HEAD(&tic->t_queue);
3254 tic->t_unit_res = unit_res;
3255 tic->t_curr_res = unit_res;
3256 tic->t_cnt = cnt;
3257 tic->t_ocnt = cnt;
3258 tic->t_tid = get_random_u32();
3259 if (permanent)
3260 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3261
3262 return tic;
3263}
3264
3265#if defined(DEBUG)
3266static void
3267xlog_verify_dump_tail(
3268 struct xlog *log,
3269 struct xlog_in_core *iclog)
3270{
3271 xfs_alert(log->l_mp,
3272"ran out of log space tail 0x%llx/0x%llx, head lsn 0x%llx, head 0x%x/0x%x, prev head 0x%x/0x%x",
3273 iclog ? be64_to_cpu(iclog->ic_header.h_tail_lsn) : -1,
3274 atomic64_read(&log->l_tail_lsn),
3275 log->l_ailp->ail_head_lsn,
3276 log->l_curr_cycle, log->l_curr_block,
3277 log->l_prev_cycle, log->l_prev_block);
3278 xfs_alert(log->l_mp,
3279"write grant 0x%llx, reserve grant 0x%llx, tail_space 0x%llx, size 0x%x, iclog flags 0x%x",
3280 atomic64_read(&log->l_write_head.grant),
3281 atomic64_read(&log->l_reserve_head.grant),
3282 log->l_tail_space, log->l_logsize,
3283 iclog ? iclog->ic_flags : -1);
3284}
3285
3286/* Check if the new iclog will fit in the log. */
3287STATIC void
3288xlog_verify_tail_lsn(
3289 struct xlog *log,
3290 struct xlog_in_core *iclog)
3291{
3292 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3293 int blocks;
3294
3295 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3296 blocks = log->l_logBBsize -
3297 (log->l_prev_block - BLOCK_LSN(tail_lsn));
3298 if (blocks < BTOBB(iclog->ic_offset) +
3299 BTOBB(log->l_iclog_hsize)) {
3300 xfs_emerg(log->l_mp,
3301 "%s: ran out of log space", __func__);
3302 xlog_verify_dump_tail(log, iclog);
3303 }
3304 return;
3305 }
3306
3307 if (CYCLE_LSN(tail_lsn) + 1 != log->l_prev_cycle) {
3308 xfs_emerg(log->l_mp, "%s: head has wrapped tail.", __func__);
3309 xlog_verify_dump_tail(log, iclog);
3310 return;
3311 }
3312 if (BLOCK_LSN(tail_lsn) == log->l_prev_block) {
3313 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3314 xlog_verify_dump_tail(log, iclog);
3315 return;
3316 }
3317
3318 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3319 if (blocks < BTOBB(iclog->ic_offset) + 1) {
3320 xfs_emerg(log->l_mp, "%s: ran out of iclog space", __func__);
3321 xlog_verify_dump_tail(log, iclog);
3322 }
3323}
3324
3325/*
3326 * Perform a number of checks on the iclog before writing to disk.
3327 *
3328 * 1. Make sure the iclogs are still circular
3329 * 2. Make sure we have a good magic number
3330 * 3. Make sure we don't have magic numbers in the data
3331 * 4. Check fields of each log operation header for:
3332 * A. Valid client identifier
3333 * B. tid ptr value falls in valid ptr space (user space code)
3334 * C. Length in log record header is correct according to the
3335 * individual operation headers within record.
3336 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3337 * log, check the preceding blocks of the physical log to make sure all
3338 * the cycle numbers agree with the current cycle number.
3339 */
3340STATIC void
3341xlog_verify_iclog(
3342 struct xlog *log,
3343 struct xlog_in_core *iclog,
3344 int count)
3345{
3346 xlog_op_header_t *ophead;
3347 xlog_in_core_t *icptr;
3348 xlog_in_core_2_t *xhdr;
3349 void *base_ptr, *ptr, *p;
3350 ptrdiff_t field_offset;
3351 uint8_t clientid;
3352 int len, i, j, k, op_len;
3353 int idx;
3354
3355 /* check validity of iclog pointers */
3356 spin_lock(&log->l_icloglock);
3357 icptr = log->l_iclog;
3358 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3359 ASSERT(icptr);
3360
3361 if (icptr != log->l_iclog)
3362 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3363 spin_unlock(&log->l_icloglock);
3364
3365 /* check log magic numbers */
3366 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3367 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3368
3369 base_ptr = ptr = &iclog->ic_header;
3370 p = &iclog->ic_header;
3371 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3372 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3373 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3374 __func__);
3375 }
3376
3377 /* check fields */
3378 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3379 base_ptr = ptr = iclog->ic_datap;
3380 ophead = ptr;
3381 xhdr = iclog->ic_data;
3382 for (i = 0; i < len; i++) {
3383 ophead = ptr;
3384
3385 /* clientid is only 1 byte */
3386 p = &ophead->oh_clientid;
3387 field_offset = p - base_ptr;
3388 if (field_offset & 0x1ff) {
3389 clientid = ophead->oh_clientid;
3390 } else {
3391 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3392 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3393 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3394 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3395 clientid = xlog_get_client_id(
3396 xhdr[j].hic_xheader.xh_cycle_data[k]);
3397 } else {
3398 clientid = xlog_get_client_id(
3399 iclog->ic_header.h_cycle_data[idx]);
3400 }
3401 }
3402 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3403 xfs_warn(log->l_mp,
3404 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3405 __func__, i, clientid, ophead,
3406 (unsigned long)field_offset);
3407 }
3408
3409 /* check length */
3410 p = &ophead->oh_len;
3411 field_offset = p - base_ptr;
3412 if (field_offset & 0x1ff) {
3413 op_len = be32_to_cpu(ophead->oh_len);
3414 } else {
3415 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3416 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3417 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3418 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3419 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3420 } else {
3421 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3422 }
3423 }
3424 ptr += sizeof(xlog_op_header_t) + op_len;
3425 }
3426}
3427#endif
3428
3429/*
3430 * Perform a forced shutdown on the log.
3431 *
3432 * This can be called from low level log code to trigger a shutdown, or from the
3433 * high level mount shutdown code when the mount shuts down.
3434 *
3435 * Our main objectives here are to make sure that:
3436 * a. if the shutdown was not due to a log IO error, flush the logs to
3437 * disk. Anything modified after this is ignored.
3438 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3439 * parties to find out. Nothing new gets queued after this is done.
3440 * c. Tasks sleeping on log reservations, pinned objects and
3441 * other resources get woken up.
3442 * d. The mount is also marked as shut down so that log triggered shutdowns
3443 * still behave the same as if they called xfs_forced_shutdown().
3444 *
3445 * Return true if the shutdown cause was a log IO error and we actually shut the
3446 * log down.
3447 */
3448bool
3449xlog_force_shutdown(
3450 struct xlog *log,
3451 uint32_t shutdown_flags)
3452{
3453 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3454
3455 if (!log)
3456 return false;
3457
3458 /*
3459 * Ensure that there is only ever one log shutdown being processed.
3460 * If we allow the log force below on a second pass after shutting
3461 * down the log, we risk deadlocking the CIL push as it may require
3462 * locks on objects the current shutdown context holds (e.g. taking
3463 * buffer locks to abort buffers on last unpin of buf log items).
3464 */
3465 if (test_and_set_bit(XLOG_SHUTDOWN_STARTED, &log->l_opstate))
3466 return false;
3467
3468 /*
3469 * Flush all the completed transactions to disk before marking the log
3470 * being shut down. We need to do this first as shutting down the log
3471 * before the force will prevent the log force from flushing the iclogs
3472 * to disk.
3473 *
3474 * When we are in recovery, there are no transactions to flush, and
3475 * we don't want to touch the log because we don't want to perturb the
3476 * current head/tail for future recovery attempts. Hence we need to
3477 * avoid a log force in this case.
3478 *
3479 * If we are shutting down due to a log IO error, then we must avoid
3480 * trying to write the log as that may just result in more IO errors and
3481 * an endless shutdown/force loop.
3482 */
3483 if (!log_error && !xlog_in_recovery(log))
3484 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3485
3486 /*
3487 * Atomically set the shutdown state. If the shutdown state is already
3488 * set, there someone else is performing the shutdown and so we are done
3489 * here. This should never happen because we should only ever get called
3490 * once by the first shutdown caller.
3491 *
3492 * Much of the log state machine transitions assume that shutdown state
3493 * cannot change once they hold the log->l_icloglock. Hence we need to
3494 * hold that lock here, even though we use the atomic test_and_set_bit()
3495 * operation to set the shutdown state.
3496 */
3497 spin_lock(&log->l_icloglock);
3498 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3499 spin_unlock(&log->l_icloglock);
3500 ASSERT(0);
3501 return false;
3502 }
3503 spin_unlock(&log->l_icloglock);
3504
3505 /*
3506 * If this log shutdown also sets the mount shutdown state, issue a
3507 * shutdown warning message.
3508 */
3509 if (!xfs_set_shutdown(log->l_mp)) {
3510 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3511"Filesystem has been shut down due to log error (0x%x).",
3512 shutdown_flags);
3513 xfs_alert(log->l_mp,
3514"Please unmount the filesystem and rectify the problem(s).");
3515 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3516 xfs_stack_trace();
3517 }
3518
3519 /*
3520 * We don't want anybody waiting for log reservations after this. That
3521 * means we have to wake up everybody queued up on reserveq as well as
3522 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3523 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3524 * action is protected by the grant locks.
3525 */
3526 xlog_grant_head_wake_all(&log->l_reserve_head);
3527 xlog_grant_head_wake_all(&log->l_write_head);
3528
3529 /*
3530 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3531 * as if the log writes were completed. The abort handling in the log
3532 * item committed callback functions will do this again under lock to
3533 * avoid races.
3534 */
3535 spin_lock(&log->l_cilp->xc_push_lock);
3536 wake_up_all(&log->l_cilp->xc_start_wait);
3537 wake_up_all(&log->l_cilp->xc_commit_wait);
3538 spin_unlock(&log->l_cilp->xc_push_lock);
3539
3540 spin_lock(&log->l_icloglock);
3541 xlog_state_shutdown_callbacks(log);
3542 spin_unlock(&log->l_icloglock);
3543
3544 wake_up_var(&log->l_opstate);
3545 return log_error;
3546}
3547
3548STATIC int
3549xlog_iclogs_empty(
3550 struct xlog *log)
3551{
3552 xlog_in_core_t *iclog;
3553
3554 iclog = log->l_iclog;
3555 do {
3556 /* endianness does not matter here, zero is zero in
3557 * any language.
3558 */
3559 if (iclog->ic_header.h_num_logops)
3560 return 0;
3561 iclog = iclog->ic_next;
3562 } while (iclog != log->l_iclog);
3563 return 1;
3564}
3565
3566/*
3567 * Verify that an LSN stamped into a piece of metadata is valid. This is
3568 * intended for use in read verifiers on v5 superblocks.
3569 */
3570bool
3571xfs_log_check_lsn(
3572 struct xfs_mount *mp,
3573 xfs_lsn_t lsn)
3574{
3575 struct xlog *log = mp->m_log;
3576 bool valid;
3577
3578 /*
3579 * norecovery mode skips mount-time log processing and unconditionally
3580 * resets the in-core LSN. We can't validate in this mode, but
3581 * modifications are not allowed anyways so just return true.
3582 */
3583 if (xfs_has_norecovery(mp))
3584 return true;
3585
3586 /*
3587 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3588 * handled by recovery and thus safe to ignore here.
3589 */
3590 if (lsn == NULLCOMMITLSN)
3591 return true;
3592
3593 valid = xlog_valid_lsn(mp->m_log, lsn);
3594
3595 /* warn the user about what's gone wrong before verifier failure */
3596 if (!valid) {
3597 spin_lock(&log->l_icloglock);
3598 xfs_warn(mp,
3599"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3600"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3601 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3602 log->l_curr_cycle, log->l_curr_block);
3603 spin_unlock(&log->l_icloglock);
3604 }
3605
3606 return valid;
3607}
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_bit.h"
22#include "xfs_log.h"
23#include "xfs_inum.h"
24#include "xfs_trans.h"
25#include "xfs_sb.h"
26#include "xfs_ag.h"
27#include "xfs_mount.h"
28#include "xfs_error.h"
29#include "xfs_log_priv.h"
30#include "xfs_buf_item.h"
31#include "xfs_bmap_btree.h"
32#include "xfs_alloc_btree.h"
33#include "xfs_ialloc_btree.h"
34#include "xfs_log_recover.h"
35#include "xfs_trans_priv.h"
36#include "xfs_dinode.h"
37#include "xfs_inode.h"
38#include "xfs_rw.h"
39#include "xfs_trace.h"
40
41kmem_zone_t *xfs_log_ticket_zone;
42
43/* Local miscellaneous function prototypes */
44STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45 xlog_in_core_t **, xfs_lsn_t *);
46STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp,
47 xfs_buftarg_t *log_target,
48 xfs_daddr_t blk_offset,
49 int num_bblks);
50STATIC int xlog_space_left(struct log *log, atomic64_t *head);
51STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52STATIC void xlog_dealloc_log(xlog_t *log);
53
54/* local state machine functions */
55STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57STATIC int xlog_state_get_iclog_space(xlog_t *log,
58 int len,
59 xlog_in_core_t **iclog,
60 xlog_ticket_t *ticket,
61 int *continued_write,
62 int *logoffsetp);
63STATIC int xlog_state_release_iclog(xlog_t *log,
64 xlog_in_core_t *iclog);
65STATIC void xlog_state_switch_iclogs(xlog_t *log,
66 xlog_in_core_t *iclog,
67 int eventual_size);
68STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
69
70/* local functions to manipulate grant head */
71STATIC int xlog_grant_log_space(xlog_t *log,
72 xlog_ticket_t *xtic);
73STATIC void xlog_grant_push_ail(struct log *log,
74 int need_bytes);
75STATIC void xlog_regrant_reserve_log_space(xlog_t *log,
76 xlog_ticket_t *ticket);
77STATIC int xlog_regrant_write_log_space(xlog_t *log,
78 xlog_ticket_t *ticket);
79STATIC void xlog_ungrant_log_space(xlog_t *log,
80 xlog_ticket_t *ticket);
81
82#if defined(DEBUG)
83STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr);
84STATIC void xlog_verify_grant_tail(struct log *log);
85STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
86 int count, boolean_t syncing);
87STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
88 xfs_lsn_t tail_lsn);
89#else
90#define xlog_verify_dest_ptr(a,b)
91#define xlog_verify_grant_tail(a)
92#define xlog_verify_iclog(a,b,c,d)
93#define xlog_verify_tail_lsn(a,b,c)
94#endif
95
96STATIC int xlog_iclogs_empty(xlog_t *log);
97
98static void
99xlog_grant_sub_space(
100 struct log *log,
101 atomic64_t *head,
102 int bytes)
103{
104 int64_t head_val = atomic64_read(head);
105 int64_t new, old;
106
107 do {
108 int cycle, space;
109
110 xlog_crack_grant_head_val(head_val, &cycle, &space);
111
112 space -= bytes;
113 if (space < 0) {
114 space += log->l_logsize;
115 cycle--;
116 }
117
118 old = head_val;
119 new = xlog_assign_grant_head_val(cycle, space);
120 head_val = atomic64_cmpxchg(head, old, new);
121 } while (head_val != old);
122}
123
124static void
125xlog_grant_add_space(
126 struct log *log,
127 atomic64_t *head,
128 int bytes)
129{
130 int64_t head_val = atomic64_read(head);
131 int64_t new, old;
132
133 do {
134 int tmp;
135 int cycle, space;
136
137 xlog_crack_grant_head_val(head_val, &cycle, &space);
138
139 tmp = log->l_logsize - space;
140 if (tmp > bytes)
141 space += bytes;
142 else {
143 space = bytes - tmp;
144 cycle++;
145 }
146
147 old = head_val;
148 new = xlog_assign_grant_head_val(cycle, space);
149 head_val = atomic64_cmpxchg(head, old, new);
150 } while (head_val != old);
151}
152
153static void
154xlog_tic_reset_res(xlog_ticket_t *tic)
155{
156 tic->t_res_num = 0;
157 tic->t_res_arr_sum = 0;
158 tic->t_res_num_ophdrs = 0;
159}
160
161static void
162xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
163{
164 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
165 /* add to overflow and start again */
166 tic->t_res_o_flow += tic->t_res_arr_sum;
167 tic->t_res_num = 0;
168 tic->t_res_arr_sum = 0;
169 }
170
171 tic->t_res_arr[tic->t_res_num].r_len = len;
172 tic->t_res_arr[tic->t_res_num].r_type = type;
173 tic->t_res_arr_sum += len;
174 tic->t_res_num++;
175}
176
177/*
178 * NOTES:
179 *
180 * 1. currblock field gets updated at startup and after in-core logs
181 * marked as with WANT_SYNC.
182 */
183
184/*
185 * This routine is called when a user of a log manager ticket is done with
186 * the reservation. If the ticket was ever used, then a commit record for
187 * the associated transaction is written out as a log operation header with
188 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
189 * a given ticket. If the ticket was one with a permanent reservation, then
190 * a few operations are done differently. Permanent reservation tickets by
191 * default don't release the reservation. They just commit the current
192 * transaction with the belief that the reservation is still needed. A flag
193 * must be passed in before permanent reservations are actually released.
194 * When these type of tickets are not released, they need to be set into
195 * the inited state again. By doing this, a start record will be written
196 * out when the next write occurs.
197 */
198xfs_lsn_t
199xfs_log_done(
200 struct xfs_mount *mp,
201 struct xlog_ticket *ticket,
202 struct xlog_in_core **iclog,
203 uint flags)
204{
205 struct log *log = mp->m_log;
206 xfs_lsn_t lsn = 0;
207
208 if (XLOG_FORCED_SHUTDOWN(log) ||
209 /*
210 * If nothing was ever written, don't write out commit record.
211 * If we get an error, just continue and give back the log ticket.
212 */
213 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
214 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
215 lsn = (xfs_lsn_t) -1;
216 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
217 flags |= XFS_LOG_REL_PERM_RESERV;
218 }
219 }
220
221
222 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
223 (flags & XFS_LOG_REL_PERM_RESERV)) {
224 trace_xfs_log_done_nonperm(log, ticket);
225
226 /*
227 * Release ticket if not permanent reservation or a specific
228 * request has been made to release a permanent reservation.
229 */
230 xlog_ungrant_log_space(log, ticket);
231 xfs_log_ticket_put(ticket);
232 } else {
233 trace_xfs_log_done_perm(log, ticket);
234
235 xlog_regrant_reserve_log_space(log, ticket);
236 /* If this ticket was a permanent reservation and we aren't
237 * trying to release it, reset the inited flags; so next time
238 * we write, a start record will be written out.
239 */
240 ticket->t_flags |= XLOG_TIC_INITED;
241 }
242
243 return lsn;
244}
245
246/*
247 * Attaches a new iclog I/O completion callback routine during
248 * transaction commit. If the log is in error state, a non-zero
249 * return code is handed back and the caller is responsible for
250 * executing the callback at an appropriate time.
251 */
252int
253xfs_log_notify(
254 struct xfs_mount *mp,
255 struct xlog_in_core *iclog,
256 xfs_log_callback_t *cb)
257{
258 int abortflg;
259
260 spin_lock(&iclog->ic_callback_lock);
261 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
262 if (!abortflg) {
263 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
264 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
265 cb->cb_next = NULL;
266 *(iclog->ic_callback_tail) = cb;
267 iclog->ic_callback_tail = &(cb->cb_next);
268 }
269 spin_unlock(&iclog->ic_callback_lock);
270 return abortflg;
271}
272
273int
274xfs_log_release_iclog(
275 struct xfs_mount *mp,
276 struct xlog_in_core *iclog)
277{
278 if (xlog_state_release_iclog(mp->m_log, iclog)) {
279 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
280 return EIO;
281 }
282
283 return 0;
284}
285
286/*
287 * 1. Reserve an amount of on-disk log space and return a ticket corresponding
288 * to the reservation.
289 * 2. Potentially, push buffers at tail of log to disk.
290 *
291 * Each reservation is going to reserve extra space for a log record header.
292 * When writes happen to the on-disk log, we don't subtract the length of the
293 * log record header from any reservation. By wasting space in each
294 * reservation, we prevent over allocation problems.
295 */
296int
297xfs_log_reserve(
298 struct xfs_mount *mp,
299 int unit_bytes,
300 int cnt,
301 struct xlog_ticket **ticket,
302 __uint8_t client,
303 uint flags,
304 uint t_type)
305{
306 struct log *log = mp->m_log;
307 struct xlog_ticket *internal_ticket;
308 int retval = 0;
309
310 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
311
312 if (XLOG_FORCED_SHUTDOWN(log))
313 return XFS_ERROR(EIO);
314
315 XFS_STATS_INC(xs_try_logspace);
316
317
318 if (*ticket != NULL) {
319 ASSERT(flags & XFS_LOG_PERM_RESERV);
320 internal_ticket = *ticket;
321
322 /*
323 * this is a new transaction on the ticket, so we need to
324 * change the transaction ID so that the next transaction has a
325 * different TID in the log. Just add one to the existing tid
326 * so that we can see chains of rolling transactions in the log
327 * easily.
328 */
329 internal_ticket->t_tid++;
330
331 trace_xfs_log_reserve(log, internal_ticket);
332
333 xlog_grant_push_ail(log, internal_ticket->t_unit_res);
334 retval = xlog_regrant_write_log_space(log, internal_ticket);
335 } else {
336 /* may sleep if need to allocate more tickets */
337 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
338 client, flags,
339 KM_SLEEP|KM_MAYFAIL);
340 if (!internal_ticket)
341 return XFS_ERROR(ENOMEM);
342 internal_ticket->t_trans_type = t_type;
343 *ticket = internal_ticket;
344
345 trace_xfs_log_reserve(log, internal_ticket);
346
347 xlog_grant_push_ail(log,
348 (internal_ticket->t_unit_res *
349 internal_ticket->t_cnt));
350 retval = xlog_grant_log_space(log, internal_ticket);
351 }
352
353 return retval;
354} /* xfs_log_reserve */
355
356
357/*
358 * Mount a log filesystem
359 *
360 * mp - ubiquitous xfs mount point structure
361 * log_target - buftarg of on-disk log device
362 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
363 * num_bblocks - Number of BBSIZE blocks in on-disk log
364 *
365 * Return error or zero.
366 */
367int
368xfs_log_mount(
369 xfs_mount_t *mp,
370 xfs_buftarg_t *log_target,
371 xfs_daddr_t blk_offset,
372 int num_bblks)
373{
374 int error;
375
376 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
377 xfs_notice(mp, "Mounting Filesystem");
378 else {
379 xfs_notice(mp,
380"Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
381 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
382 }
383
384 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
385 if (IS_ERR(mp->m_log)) {
386 error = -PTR_ERR(mp->m_log);
387 goto out;
388 }
389
390 /*
391 * Initialize the AIL now we have a log.
392 */
393 error = xfs_trans_ail_init(mp);
394 if (error) {
395 xfs_warn(mp, "AIL initialisation failed: error %d", error);
396 goto out_free_log;
397 }
398 mp->m_log->l_ailp = mp->m_ail;
399
400 /*
401 * skip log recovery on a norecovery mount. pretend it all
402 * just worked.
403 */
404 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
405 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
406
407 if (readonly)
408 mp->m_flags &= ~XFS_MOUNT_RDONLY;
409
410 error = xlog_recover(mp->m_log);
411
412 if (readonly)
413 mp->m_flags |= XFS_MOUNT_RDONLY;
414 if (error) {
415 xfs_warn(mp, "log mount/recovery failed: error %d",
416 error);
417 goto out_destroy_ail;
418 }
419 }
420
421 /* Normal transactions can now occur */
422 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
423
424 /*
425 * Now the log has been fully initialised and we know were our
426 * space grant counters are, we can initialise the permanent ticket
427 * needed for delayed logging to work.
428 */
429 xlog_cil_init_post_recovery(mp->m_log);
430
431 return 0;
432
433out_destroy_ail:
434 xfs_trans_ail_destroy(mp);
435out_free_log:
436 xlog_dealloc_log(mp->m_log);
437out:
438 return error;
439}
440
441/*
442 * Finish the recovery of the file system. This is separate from
443 * the xfs_log_mount() call, because it depends on the code in
444 * xfs_mountfs() to read in the root and real-time bitmap inodes
445 * between calling xfs_log_mount() and here.
446 *
447 * mp - ubiquitous xfs mount point structure
448 */
449int
450xfs_log_mount_finish(xfs_mount_t *mp)
451{
452 int error;
453
454 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
455 error = xlog_recover_finish(mp->m_log);
456 else {
457 error = 0;
458 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
459 }
460
461 return error;
462}
463
464/*
465 * Final log writes as part of unmount.
466 *
467 * Mark the filesystem clean as unmount happens. Note that during relocation
468 * this routine needs to be executed as part of source-bag while the
469 * deallocation must not be done until source-end.
470 */
471
472/*
473 * Unmount record used to have a string "Unmount filesystem--" in the
474 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
475 * We just write the magic number now since that particular field isn't
476 * currently architecture converted and "nUmount" is a bit foo.
477 * As far as I know, there weren't any dependencies on the old behaviour.
478 */
479
480int
481xfs_log_unmount_write(xfs_mount_t *mp)
482{
483 xlog_t *log = mp->m_log;
484 xlog_in_core_t *iclog;
485#ifdef DEBUG
486 xlog_in_core_t *first_iclog;
487#endif
488 xlog_ticket_t *tic = NULL;
489 xfs_lsn_t lsn;
490 int error;
491
492 /*
493 * Don't write out unmount record on read-only mounts.
494 * Or, if we are doing a forced umount (typically because of IO errors).
495 */
496 if (mp->m_flags & XFS_MOUNT_RDONLY)
497 return 0;
498
499 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
500 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
501
502#ifdef DEBUG
503 first_iclog = iclog = log->l_iclog;
504 do {
505 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
506 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
507 ASSERT(iclog->ic_offset == 0);
508 }
509 iclog = iclog->ic_next;
510 } while (iclog != first_iclog);
511#endif
512 if (! (XLOG_FORCED_SHUTDOWN(log))) {
513 error = xfs_log_reserve(mp, 600, 1, &tic,
514 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
515 if (!error) {
516 /* the data section must be 32 bit size aligned */
517 struct {
518 __uint16_t magic;
519 __uint16_t pad1;
520 __uint32_t pad2; /* may as well make it 64 bits */
521 } magic = {
522 .magic = XLOG_UNMOUNT_TYPE,
523 };
524 struct xfs_log_iovec reg = {
525 .i_addr = &magic,
526 .i_len = sizeof(magic),
527 .i_type = XLOG_REG_TYPE_UNMOUNT,
528 };
529 struct xfs_log_vec vec = {
530 .lv_niovecs = 1,
531 .lv_iovecp = ®,
532 };
533
534 /* remove inited flag */
535 tic->t_flags = 0;
536 error = xlog_write(log, &vec, tic, &lsn,
537 NULL, XLOG_UNMOUNT_TRANS);
538 /*
539 * At this point, we're umounting anyway,
540 * so there's no point in transitioning log state
541 * to IOERROR. Just continue...
542 */
543 }
544
545 if (error)
546 xfs_alert(mp, "%s: unmount record failed", __func__);
547
548
549 spin_lock(&log->l_icloglock);
550 iclog = log->l_iclog;
551 atomic_inc(&iclog->ic_refcnt);
552 xlog_state_want_sync(log, iclog);
553 spin_unlock(&log->l_icloglock);
554 error = xlog_state_release_iclog(log, iclog);
555
556 spin_lock(&log->l_icloglock);
557 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
558 iclog->ic_state == XLOG_STATE_DIRTY)) {
559 if (!XLOG_FORCED_SHUTDOWN(log)) {
560 xlog_wait(&iclog->ic_force_wait,
561 &log->l_icloglock);
562 } else {
563 spin_unlock(&log->l_icloglock);
564 }
565 } else {
566 spin_unlock(&log->l_icloglock);
567 }
568 if (tic) {
569 trace_xfs_log_umount_write(log, tic);
570 xlog_ungrant_log_space(log, tic);
571 xfs_log_ticket_put(tic);
572 }
573 } else {
574 /*
575 * We're already in forced_shutdown mode, couldn't
576 * even attempt to write out the unmount transaction.
577 *
578 * Go through the motions of sync'ing and releasing
579 * the iclog, even though no I/O will actually happen,
580 * we need to wait for other log I/Os that may already
581 * be in progress. Do this as a separate section of
582 * code so we'll know if we ever get stuck here that
583 * we're in this odd situation of trying to unmount
584 * a file system that went into forced_shutdown as
585 * the result of an unmount..
586 */
587 spin_lock(&log->l_icloglock);
588 iclog = log->l_iclog;
589 atomic_inc(&iclog->ic_refcnt);
590
591 xlog_state_want_sync(log, iclog);
592 spin_unlock(&log->l_icloglock);
593 error = xlog_state_release_iclog(log, iclog);
594
595 spin_lock(&log->l_icloglock);
596
597 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
598 || iclog->ic_state == XLOG_STATE_DIRTY
599 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
600
601 xlog_wait(&iclog->ic_force_wait,
602 &log->l_icloglock);
603 } else {
604 spin_unlock(&log->l_icloglock);
605 }
606 }
607
608 return error;
609} /* xfs_log_unmount_write */
610
611/*
612 * Deallocate log structures for unmount/relocation.
613 *
614 * We need to stop the aild from running before we destroy
615 * and deallocate the log as the aild references the log.
616 */
617void
618xfs_log_unmount(xfs_mount_t *mp)
619{
620 xfs_trans_ail_destroy(mp);
621 xlog_dealloc_log(mp->m_log);
622}
623
624void
625xfs_log_item_init(
626 struct xfs_mount *mp,
627 struct xfs_log_item *item,
628 int type,
629 struct xfs_item_ops *ops)
630{
631 item->li_mountp = mp;
632 item->li_ailp = mp->m_ail;
633 item->li_type = type;
634 item->li_ops = ops;
635 item->li_lv = NULL;
636
637 INIT_LIST_HEAD(&item->li_ail);
638 INIT_LIST_HEAD(&item->li_cil);
639}
640
641/*
642 * Write region vectors to log. The write happens using the space reservation
643 * of the ticket (tic). It is not a requirement that all writes for a given
644 * transaction occur with one call to xfs_log_write(). However, it is important
645 * to note that the transaction reservation code makes an assumption about the
646 * number of log headers a transaction requires that may be violated if you
647 * don't pass all the transaction vectors in one call....
648 */
649int
650xfs_log_write(
651 struct xfs_mount *mp,
652 struct xfs_log_iovec reg[],
653 int nentries,
654 struct xlog_ticket *tic,
655 xfs_lsn_t *start_lsn)
656{
657 struct log *log = mp->m_log;
658 int error;
659 struct xfs_log_vec vec = {
660 .lv_niovecs = nentries,
661 .lv_iovecp = reg,
662 };
663
664 if (XLOG_FORCED_SHUTDOWN(log))
665 return XFS_ERROR(EIO);
666
667 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
668 if (error)
669 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
670 return error;
671}
672
673void
674xfs_log_move_tail(xfs_mount_t *mp,
675 xfs_lsn_t tail_lsn)
676{
677 xlog_ticket_t *tic;
678 xlog_t *log = mp->m_log;
679 int need_bytes, free_bytes;
680
681 if (XLOG_FORCED_SHUTDOWN(log))
682 return;
683
684 if (tail_lsn == 0)
685 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
686
687 /* tail_lsn == 1 implies that we weren't passed a valid value. */
688 if (tail_lsn != 1)
689 atomic64_set(&log->l_tail_lsn, tail_lsn);
690
691 if (!list_empty_careful(&log->l_writeq)) {
692#ifdef DEBUG
693 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
694 panic("Recovery problem");
695#endif
696 spin_lock(&log->l_grant_write_lock);
697 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
698 list_for_each_entry(tic, &log->l_writeq, t_queue) {
699 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
700
701 if (free_bytes < tic->t_unit_res && tail_lsn != 1)
702 break;
703 tail_lsn = 0;
704 free_bytes -= tic->t_unit_res;
705 trace_xfs_log_regrant_write_wake_up(log, tic);
706 wake_up(&tic->t_wait);
707 }
708 spin_unlock(&log->l_grant_write_lock);
709 }
710
711 if (!list_empty_careful(&log->l_reserveq)) {
712#ifdef DEBUG
713 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
714 panic("Recovery problem");
715#endif
716 spin_lock(&log->l_grant_reserve_lock);
717 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
718 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
719 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
720 need_bytes = tic->t_unit_res*tic->t_cnt;
721 else
722 need_bytes = tic->t_unit_res;
723 if (free_bytes < need_bytes && tail_lsn != 1)
724 break;
725 tail_lsn = 0;
726 free_bytes -= need_bytes;
727 trace_xfs_log_grant_wake_up(log, tic);
728 wake_up(&tic->t_wait);
729 }
730 spin_unlock(&log->l_grant_reserve_lock);
731 }
732}
733
734/*
735 * Determine if we have a transaction that has gone to disk
736 * that needs to be covered. To begin the transition to the idle state
737 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
738 * If we are then in a state where covering is needed, the caller is informed
739 * that dummy transactions are required to move the log into the idle state.
740 *
741 * Because this is called as part of the sync process, we should also indicate
742 * that dummy transactions should be issued in anything but the covered or
743 * idle states. This ensures that the log tail is accurately reflected in
744 * the log at the end of the sync, hence if a crash occurrs avoids replay
745 * of transactions where the metadata is already on disk.
746 */
747int
748xfs_log_need_covered(xfs_mount_t *mp)
749{
750 int needed = 0;
751 xlog_t *log = mp->m_log;
752
753 if (!xfs_fs_writable(mp))
754 return 0;
755
756 spin_lock(&log->l_icloglock);
757 switch (log->l_covered_state) {
758 case XLOG_STATE_COVER_DONE:
759 case XLOG_STATE_COVER_DONE2:
760 case XLOG_STATE_COVER_IDLE:
761 break;
762 case XLOG_STATE_COVER_NEED:
763 case XLOG_STATE_COVER_NEED2:
764 if (!xfs_ail_min_lsn(log->l_ailp) &&
765 xlog_iclogs_empty(log)) {
766 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
767 log->l_covered_state = XLOG_STATE_COVER_DONE;
768 else
769 log->l_covered_state = XLOG_STATE_COVER_DONE2;
770 }
771 /* FALLTHRU */
772 default:
773 needed = 1;
774 break;
775 }
776 spin_unlock(&log->l_icloglock);
777 return needed;
778}
779
780/******************************************************************************
781 *
782 * local routines
783 *
784 ******************************************************************************
785 */
786
787/* xfs_trans_tail_ail returns 0 when there is nothing in the list.
788 * The log manager must keep track of the last LR which was committed
789 * to disk. The lsn of this LR will become the new tail_lsn whenever
790 * xfs_trans_tail_ail returns 0. If we don't do this, we run into
791 * the situation where stuff could be written into the log but nothing
792 * was ever in the AIL when asked. Eventually, we panic since the
793 * tail hits the head.
794 *
795 * We may be holding the log iclog lock upon entering this routine.
796 */
797xfs_lsn_t
798xlog_assign_tail_lsn(
799 struct xfs_mount *mp)
800{
801 xfs_lsn_t tail_lsn;
802 struct log *log = mp->m_log;
803
804 tail_lsn = xfs_ail_min_lsn(mp->m_ail);
805 if (!tail_lsn)
806 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
807
808 atomic64_set(&log->l_tail_lsn, tail_lsn);
809 return tail_lsn;
810}
811
812/*
813 * Return the space in the log between the tail and the head. The head
814 * is passed in the cycle/bytes formal parms. In the special case where
815 * the reserve head has wrapped passed the tail, this calculation is no
816 * longer valid. In this case, just return 0 which means there is no space
817 * in the log. This works for all places where this function is called
818 * with the reserve head. Of course, if the write head were to ever
819 * wrap the tail, we should blow up. Rather than catch this case here,
820 * we depend on other ASSERTions in other parts of the code. XXXmiken
821 *
822 * This code also handles the case where the reservation head is behind
823 * the tail. The details of this case are described below, but the end
824 * result is that we return the size of the log as the amount of space left.
825 */
826STATIC int
827xlog_space_left(
828 struct log *log,
829 atomic64_t *head)
830{
831 int free_bytes;
832 int tail_bytes;
833 int tail_cycle;
834 int head_cycle;
835 int head_bytes;
836
837 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
838 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
839 tail_bytes = BBTOB(tail_bytes);
840 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
841 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
842 else if (tail_cycle + 1 < head_cycle)
843 return 0;
844 else if (tail_cycle < head_cycle) {
845 ASSERT(tail_cycle == (head_cycle - 1));
846 free_bytes = tail_bytes - head_bytes;
847 } else {
848 /*
849 * The reservation head is behind the tail.
850 * In this case we just want to return the size of the
851 * log as the amount of space left.
852 */
853 xfs_alert(log->l_mp,
854 "xlog_space_left: head behind tail\n"
855 " tail_cycle = %d, tail_bytes = %d\n"
856 " GH cycle = %d, GH bytes = %d",
857 tail_cycle, tail_bytes, head_cycle, head_bytes);
858 ASSERT(0);
859 free_bytes = log->l_logsize;
860 }
861 return free_bytes;
862}
863
864
865/*
866 * Log function which is called when an io completes.
867 *
868 * The log manager needs its own routine, in order to control what
869 * happens with the buffer after the write completes.
870 */
871void
872xlog_iodone(xfs_buf_t *bp)
873{
874 xlog_in_core_t *iclog = bp->b_fspriv;
875 xlog_t *l = iclog->ic_log;
876 int aborted = 0;
877
878 /*
879 * Race to shutdown the filesystem if we see an error.
880 */
881 if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
882 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
883 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
884 XFS_BUF_STALE(bp);
885 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
886 /*
887 * This flag will be propagated to the trans-committed
888 * callback routines to let them know that the log-commit
889 * didn't succeed.
890 */
891 aborted = XFS_LI_ABORTED;
892 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
893 aborted = XFS_LI_ABORTED;
894 }
895
896 /* log I/O is always issued ASYNC */
897 ASSERT(XFS_BUF_ISASYNC(bp));
898 xlog_state_done_syncing(iclog, aborted);
899 /*
900 * do not reference the buffer (bp) here as we could race
901 * with it being freed after writing the unmount record to the
902 * log.
903 */
904
905} /* xlog_iodone */
906
907/*
908 * Return size of each in-core log record buffer.
909 *
910 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
911 *
912 * If the filesystem blocksize is too large, we may need to choose a
913 * larger size since the directory code currently logs entire blocks.
914 */
915
916STATIC void
917xlog_get_iclog_buffer_size(xfs_mount_t *mp,
918 xlog_t *log)
919{
920 int size;
921 int xhdrs;
922
923 if (mp->m_logbufs <= 0)
924 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
925 else
926 log->l_iclog_bufs = mp->m_logbufs;
927
928 /*
929 * Buffer size passed in from mount system call.
930 */
931 if (mp->m_logbsize > 0) {
932 size = log->l_iclog_size = mp->m_logbsize;
933 log->l_iclog_size_log = 0;
934 while (size != 1) {
935 log->l_iclog_size_log++;
936 size >>= 1;
937 }
938
939 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
940 /* # headers = size / 32k
941 * one header holds cycles from 32k of data
942 */
943
944 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
945 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
946 xhdrs++;
947 log->l_iclog_hsize = xhdrs << BBSHIFT;
948 log->l_iclog_heads = xhdrs;
949 } else {
950 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
951 log->l_iclog_hsize = BBSIZE;
952 log->l_iclog_heads = 1;
953 }
954 goto done;
955 }
956
957 /* All machines use 32kB buffers by default. */
958 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
959 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
960
961 /* the default log size is 16k or 32k which is one header sector */
962 log->l_iclog_hsize = BBSIZE;
963 log->l_iclog_heads = 1;
964
965done:
966 /* are we being asked to make the sizes selected above visible? */
967 if (mp->m_logbufs == 0)
968 mp->m_logbufs = log->l_iclog_bufs;
969 if (mp->m_logbsize == 0)
970 mp->m_logbsize = log->l_iclog_size;
971} /* xlog_get_iclog_buffer_size */
972
973
974/*
975 * This routine initializes some of the log structure for a given mount point.
976 * Its primary purpose is to fill in enough, so recovery can occur. However,
977 * some other stuff may be filled in too.
978 */
979STATIC xlog_t *
980xlog_alloc_log(xfs_mount_t *mp,
981 xfs_buftarg_t *log_target,
982 xfs_daddr_t blk_offset,
983 int num_bblks)
984{
985 xlog_t *log;
986 xlog_rec_header_t *head;
987 xlog_in_core_t **iclogp;
988 xlog_in_core_t *iclog, *prev_iclog=NULL;
989 xfs_buf_t *bp;
990 int i;
991 int error = ENOMEM;
992 uint log2_size = 0;
993
994 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
995 if (!log) {
996 xfs_warn(mp, "Log allocation failed: No memory!");
997 goto out;
998 }
999
1000 log->l_mp = mp;
1001 log->l_targ = log_target;
1002 log->l_logsize = BBTOB(num_bblks);
1003 log->l_logBBstart = blk_offset;
1004 log->l_logBBsize = num_bblks;
1005 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1006 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1007
1008 log->l_prev_block = -1;
1009 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1010 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1011 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1012 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1013 xlog_assign_grant_head(&log->l_grant_reserve_head, 1, 0);
1014 xlog_assign_grant_head(&log->l_grant_write_head, 1, 0);
1015 INIT_LIST_HEAD(&log->l_reserveq);
1016 INIT_LIST_HEAD(&log->l_writeq);
1017 spin_lock_init(&log->l_grant_reserve_lock);
1018 spin_lock_init(&log->l_grant_write_lock);
1019
1020 error = EFSCORRUPTED;
1021 if (xfs_sb_version_hassector(&mp->m_sb)) {
1022 log2_size = mp->m_sb.sb_logsectlog;
1023 if (log2_size < BBSHIFT) {
1024 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1025 log2_size, BBSHIFT);
1026 goto out_free_log;
1027 }
1028
1029 log2_size -= BBSHIFT;
1030 if (log2_size > mp->m_sectbb_log) {
1031 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1032 log2_size, mp->m_sectbb_log);
1033 goto out_free_log;
1034 }
1035
1036 /* for larger sector sizes, must have v2 or external log */
1037 if (log2_size && log->l_logBBstart > 0 &&
1038 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1039 xfs_warn(mp,
1040 "log sector size (0x%x) invalid for configuration.",
1041 log2_size);
1042 goto out_free_log;
1043 }
1044 }
1045 log->l_sectBBsize = 1 << log2_size;
1046
1047 xlog_get_iclog_buffer_size(mp, log);
1048
1049 error = ENOMEM;
1050 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1051 if (!bp)
1052 goto out_free_log;
1053 bp->b_iodone = xlog_iodone;
1054 ASSERT(xfs_buf_islocked(bp));
1055 log->l_xbuf = bp;
1056
1057 spin_lock_init(&log->l_icloglock);
1058 init_waitqueue_head(&log->l_flush_wait);
1059
1060 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1061 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1062
1063 iclogp = &log->l_iclog;
1064 /*
1065 * The amount of memory to allocate for the iclog structure is
1066 * rather funky due to the way the structure is defined. It is
1067 * done this way so that we can use different sizes for machines
1068 * with different amounts of memory. See the definition of
1069 * xlog_in_core_t in xfs_log_priv.h for details.
1070 */
1071 ASSERT(log->l_iclog_size >= 4096);
1072 for (i=0; i < log->l_iclog_bufs; i++) {
1073 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1074 if (!*iclogp)
1075 goto out_free_iclog;
1076
1077 iclog = *iclogp;
1078 iclog->ic_prev = prev_iclog;
1079 prev_iclog = iclog;
1080
1081 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1082 log->l_iclog_size, 0);
1083 if (!bp)
1084 goto out_free_iclog;
1085
1086 bp->b_iodone = xlog_iodone;
1087 iclog->ic_bp = bp;
1088 iclog->ic_data = bp->b_addr;
1089#ifdef DEBUG
1090 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1091#endif
1092 head = &iclog->ic_header;
1093 memset(head, 0, sizeof(xlog_rec_header_t));
1094 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1095 head->h_version = cpu_to_be32(
1096 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1097 head->h_size = cpu_to_be32(log->l_iclog_size);
1098 /* new fields */
1099 head->h_fmt = cpu_to_be32(XLOG_FMT);
1100 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1101
1102 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1103 iclog->ic_state = XLOG_STATE_ACTIVE;
1104 iclog->ic_log = log;
1105 atomic_set(&iclog->ic_refcnt, 0);
1106 spin_lock_init(&iclog->ic_callback_lock);
1107 iclog->ic_callback_tail = &(iclog->ic_callback);
1108 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1109
1110 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1111 init_waitqueue_head(&iclog->ic_force_wait);
1112 init_waitqueue_head(&iclog->ic_write_wait);
1113
1114 iclogp = &iclog->ic_next;
1115 }
1116 *iclogp = log->l_iclog; /* complete ring */
1117 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1118
1119 error = xlog_cil_init(log);
1120 if (error)
1121 goto out_free_iclog;
1122 return log;
1123
1124out_free_iclog:
1125 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1126 prev_iclog = iclog->ic_next;
1127 if (iclog->ic_bp)
1128 xfs_buf_free(iclog->ic_bp);
1129 kmem_free(iclog);
1130 }
1131 spinlock_destroy(&log->l_icloglock);
1132 xfs_buf_free(log->l_xbuf);
1133out_free_log:
1134 kmem_free(log);
1135out:
1136 return ERR_PTR(-error);
1137} /* xlog_alloc_log */
1138
1139
1140/*
1141 * Write out the commit record of a transaction associated with the given
1142 * ticket. Return the lsn of the commit record.
1143 */
1144STATIC int
1145xlog_commit_record(
1146 struct log *log,
1147 struct xlog_ticket *ticket,
1148 struct xlog_in_core **iclog,
1149 xfs_lsn_t *commitlsnp)
1150{
1151 struct xfs_mount *mp = log->l_mp;
1152 int error;
1153 struct xfs_log_iovec reg = {
1154 .i_addr = NULL,
1155 .i_len = 0,
1156 .i_type = XLOG_REG_TYPE_COMMIT,
1157 };
1158 struct xfs_log_vec vec = {
1159 .lv_niovecs = 1,
1160 .lv_iovecp = ®,
1161 };
1162
1163 ASSERT_ALWAYS(iclog);
1164 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1165 XLOG_COMMIT_TRANS);
1166 if (error)
1167 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1168 return error;
1169}
1170
1171/*
1172 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1173 * log space. This code pushes on the lsn which would supposedly free up
1174 * the 25% which we want to leave free. We may need to adopt a policy which
1175 * pushes on an lsn which is further along in the log once we reach the high
1176 * water mark. In this manner, we would be creating a low water mark.
1177 */
1178STATIC void
1179xlog_grant_push_ail(
1180 struct log *log,
1181 int need_bytes)
1182{
1183 xfs_lsn_t threshold_lsn = 0;
1184 xfs_lsn_t last_sync_lsn;
1185 int free_blocks;
1186 int free_bytes;
1187 int threshold_block;
1188 int threshold_cycle;
1189 int free_threshold;
1190
1191 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1192
1193 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
1194 free_blocks = BTOBBT(free_bytes);
1195
1196 /*
1197 * Set the threshold for the minimum number of free blocks in the
1198 * log to the maximum of what the caller needs, one quarter of the
1199 * log, and 256 blocks.
1200 */
1201 free_threshold = BTOBB(need_bytes);
1202 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1203 free_threshold = MAX(free_threshold, 256);
1204 if (free_blocks >= free_threshold)
1205 return;
1206
1207 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1208 &threshold_block);
1209 threshold_block += free_threshold;
1210 if (threshold_block >= log->l_logBBsize) {
1211 threshold_block -= log->l_logBBsize;
1212 threshold_cycle += 1;
1213 }
1214 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1215 threshold_block);
1216 /*
1217 * Don't pass in an lsn greater than the lsn of the last
1218 * log record known to be on disk. Use a snapshot of the last sync lsn
1219 * so that it doesn't change between the compare and the set.
1220 */
1221 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1222 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1223 threshold_lsn = last_sync_lsn;
1224
1225 /*
1226 * Get the transaction layer to kick the dirty buffers out to
1227 * disk asynchronously. No point in trying to do this if
1228 * the filesystem is shutting down.
1229 */
1230 if (!XLOG_FORCED_SHUTDOWN(log))
1231 xfs_ail_push(log->l_ailp, threshold_lsn);
1232}
1233
1234/*
1235 * The bdstrat callback function for log bufs. This gives us a central
1236 * place to trap bufs in case we get hit by a log I/O error and need to
1237 * shutdown. Actually, in practice, even when we didn't get a log error,
1238 * we transition the iclogs to IOERROR state *after* flushing all existing
1239 * iclogs to disk. This is because we don't want anymore new transactions to be
1240 * started or completed afterwards.
1241 */
1242STATIC int
1243xlog_bdstrat(
1244 struct xfs_buf *bp)
1245{
1246 struct xlog_in_core *iclog = bp->b_fspriv;
1247
1248 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1249 xfs_buf_ioerror(bp, EIO);
1250 XFS_BUF_STALE(bp);
1251 xfs_buf_ioend(bp, 0);
1252 /*
1253 * It would seem logical to return EIO here, but we rely on
1254 * the log state machine to propagate I/O errors instead of
1255 * doing it here.
1256 */
1257 return 0;
1258 }
1259
1260 xfs_buf_iorequest(bp);
1261 return 0;
1262}
1263
1264/*
1265 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1266 * fashion. Previously, we should have moved the current iclog
1267 * ptr in the log to point to the next available iclog. This allows further
1268 * write to continue while this code syncs out an iclog ready to go.
1269 * Before an in-core log can be written out, the data section must be scanned
1270 * to save away the 1st word of each BBSIZE block into the header. We replace
1271 * it with the current cycle count. Each BBSIZE block is tagged with the
1272 * cycle count because there in an implicit assumption that drives will
1273 * guarantee that entire 512 byte blocks get written at once. In other words,
1274 * we can't have part of a 512 byte block written and part not written. By
1275 * tagging each block, we will know which blocks are valid when recovering
1276 * after an unclean shutdown.
1277 *
1278 * This routine is single threaded on the iclog. No other thread can be in
1279 * this routine with the same iclog. Changing contents of iclog can there-
1280 * fore be done without grabbing the state machine lock. Updating the global
1281 * log will require grabbing the lock though.
1282 *
1283 * The entire log manager uses a logical block numbering scheme. Only
1284 * log_sync (and then only bwrite()) know about the fact that the log may
1285 * not start with block zero on a given device. The log block start offset
1286 * is added immediately before calling bwrite().
1287 */
1288
1289STATIC int
1290xlog_sync(xlog_t *log,
1291 xlog_in_core_t *iclog)
1292{
1293 xfs_caddr_t dptr; /* pointer to byte sized element */
1294 xfs_buf_t *bp;
1295 int i;
1296 uint count; /* byte count of bwrite */
1297 uint count_init; /* initial count before roundup */
1298 int roundoff; /* roundoff to BB or stripe */
1299 int split = 0; /* split write into two regions */
1300 int error;
1301 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1302
1303 XFS_STATS_INC(xs_log_writes);
1304 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1305
1306 /* Add for LR header */
1307 count_init = log->l_iclog_hsize + iclog->ic_offset;
1308
1309 /* Round out the log write size */
1310 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1311 /* we have a v2 stripe unit to use */
1312 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1313 } else {
1314 count = BBTOB(BTOBB(count_init));
1315 }
1316 roundoff = count - count_init;
1317 ASSERT(roundoff >= 0);
1318 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1319 roundoff < log->l_mp->m_sb.sb_logsunit)
1320 ||
1321 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1322 roundoff < BBTOB(1)));
1323
1324 /* move grant heads by roundoff in sync */
1325 xlog_grant_add_space(log, &log->l_grant_reserve_head, roundoff);
1326 xlog_grant_add_space(log, &log->l_grant_write_head, roundoff);
1327
1328 /* put cycle number in every block */
1329 xlog_pack_data(log, iclog, roundoff);
1330
1331 /* real byte length */
1332 if (v2) {
1333 iclog->ic_header.h_len =
1334 cpu_to_be32(iclog->ic_offset + roundoff);
1335 } else {
1336 iclog->ic_header.h_len =
1337 cpu_to_be32(iclog->ic_offset);
1338 }
1339
1340 bp = iclog->ic_bp;
1341 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1342
1343 XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1344
1345 /* Do we need to split this write into 2 parts? */
1346 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1347 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1348 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1349 iclog->ic_bwritecnt = 2; /* split into 2 writes */
1350 } else {
1351 iclog->ic_bwritecnt = 1;
1352 }
1353 XFS_BUF_SET_COUNT(bp, count);
1354 bp->b_fspriv = iclog;
1355 XFS_BUF_ZEROFLAGS(bp);
1356 XFS_BUF_ASYNC(bp);
1357 bp->b_flags |= XBF_SYNCIO;
1358
1359 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1360 bp->b_flags |= XBF_FUA;
1361
1362 /*
1363 * Flush the data device before flushing the log to make
1364 * sure all meta data written back from the AIL actually made
1365 * it to disk before stamping the new log tail LSN into the
1366 * log buffer. For an external log we need to issue the
1367 * flush explicitly, and unfortunately synchronously here;
1368 * for an internal log we can simply use the block layer
1369 * state machine for preflushes.
1370 */
1371 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1372 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1373 else
1374 bp->b_flags |= XBF_FLUSH;
1375 }
1376
1377 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1378 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1379
1380 xlog_verify_iclog(log, iclog, count, B_TRUE);
1381
1382 /* account for log which doesn't start at block #0 */
1383 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1384 /*
1385 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1386 * is shutting down.
1387 */
1388 XFS_BUF_WRITE(bp);
1389
1390 if ((error = xlog_bdstrat(bp))) {
1391 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1392 XFS_BUF_ADDR(bp));
1393 return error;
1394 }
1395 if (split) {
1396 bp = iclog->ic_log->l_xbuf;
1397 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1398 xfs_buf_associate_memory(bp,
1399 (char *)&iclog->ic_header + count, split);
1400 bp->b_fspriv = iclog;
1401 XFS_BUF_ZEROFLAGS(bp);
1402 XFS_BUF_ASYNC(bp);
1403 bp->b_flags |= XBF_SYNCIO;
1404 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1405 bp->b_flags |= XBF_FUA;
1406 dptr = bp->b_addr;
1407 /*
1408 * Bump the cycle numbers at the start of each block
1409 * since this part of the buffer is at the start of
1410 * a new cycle. Watch out for the header magic number
1411 * case, though.
1412 */
1413 for (i = 0; i < split; i += BBSIZE) {
1414 be32_add_cpu((__be32 *)dptr, 1);
1415 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1416 be32_add_cpu((__be32 *)dptr, 1);
1417 dptr += BBSIZE;
1418 }
1419
1420 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1421 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1422
1423 /* account for internal log which doesn't start at block #0 */
1424 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1425 XFS_BUF_WRITE(bp);
1426 if ((error = xlog_bdstrat(bp))) {
1427 xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1428 bp, XFS_BUF_ADDR(bp));
1429 return error;
1430 }
1431 }
1432 return 0;
1433} /* xlog_sync */
1434
1435
1436/*
1437 * Deallocate a log structure
1438 */
1439STATIC void
1440xlog_dealloc_log(xlog_t *log)
1441{
1442 xlog_in_core_t *iclog, *next_iclog;
1443 int i;
1444
1445 xlog_cil_destroy(log);
1446
1447 /*
1448 * always need to ensure that the extra buffer does not point to memory
1449 * owned by another log buffer before we free it.
1450 */
1451 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size);
1452 xfs_buf_free(log->l_xbuf);
1453
1454 iclog = log->l_iclog;
1455 for (i=0; i<log->l_iclog_bufs; i++) {
1456 xfs_buf_free(iclog->ic_bp);
1457 next_iclog = iclog->ic_next;
1458 kmem_free(iclog);
1459 iclog = next_iclog;
1460 }
1461 spinlock_destroy(&log->l_icloglock);
1462
1463 log->l_mp->m_log = NULL;
1464 kmem_free(log);
1465} /* xlog_dealloc_log */
1466
1467/*
1468 * Update counters atomically now that memcpy is done.
1469 */
1470/* ARGSUSED */
1471static inline void
1472xlog_state_finish_copy(xlog_t *log,
1473 xlog_in_core_t *iclog,
1474 int record_cnt,
1475 int copy_bytes)
1476{
1477 spin_lock(&log->l_icloglock);
1478
1479 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1480 iclog->ic_offset += copy_bytes;
1481
1482 spin_unlock(&log->l_icloglock);
1483} /* xlog_state_finish_copy */
1484
1485
1486
1487
1488/*
1489 * print out info relating to regions written which consume
1490 * the reservation
1491 */
1492void
1493xlog_print_tic_res(
1494 struct xfs_mount *mp,
1495 struct xlog_ticket *ticket)
1496{
1497 uint i;
1498 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1499
1500 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1501 static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1502 "bformat",
1503 "bchunk",
1504 "efi_format",
1505 "efd_format",
1506 "iformat",
1507 "icore",
1508 "iext",
1509 "ibroot",
1510 "ilocal",
1511 "iattr_ext",
1512 "iattr_broot",
1513 "iattr_local",
1514 "qformat",
1515 "dquot",
1516 "quotaoff",
1517 "LR header",
1518 "unmount",
1519 "commit",
1520 "trans header"
1521 };
1522 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1523 "SETATTR_NOT_SIZE",
1524 "SETATTR_SIZE",
1525 "INACTIVE",
1526 "CREATE",
1527 "CREATE_TRUNC",
1528 "TRUNCATE_FILE",
1529 "REMOVE",
1530 "LINK",
1531 "RENAME",
1532 "MKDIR",
1533 "RMDIR",
1534 "SYMLINK",
1535 "SET_DMATTRS",
1536 "GROWFS",
1537 "STRAT_WRITE",
1538 "DIOSTRAT",
1539 "WRITE_SYNC",
1540 "WRITEID",
1541 "ADDAFORK",
1542 "ATTRINVAL",
1543 "ATRUNCATE",
1544 "ATTR_SET",
1545 "ATTR_RM",
1546 "ATTR_FLAG",
1547 "CLEAR_AGI_BUCKET",
1548 "QM_SBCHANGE",
1549 "DUMMY1",
1550 "DUMMY2",
1551 "QM_QUOTAOFF",
1552 "QM_DQALLOC",
1553 "QM_SETQLIM",
1554 "QM_DQCLUSTER",
1555 "QM_QINOCREATE",
1556 "QM_QUOTAOFF_END",
1557 "SB_UNIT",
1558 "FSYNC_TS",
1559 "GROWFSRT_ALLOC",
1560 "GROWFSRT_ZERO",
1561 "GROWFSRT_FREE",
1562 "SWAPEXT"
1563 };
1564
1565 xfs_warn(mp,
1566 "xfs_log_write: reservation summary:\n"
1567 " trans type = %s (%u)\n"
1568 " unit res = %d bytes\n"
1569 " current res = %d bytes\n"
1570 " total reg = %u bytes (o/flow = %u bytes)\n"
1571 " ophdrs = %u (ophdr space = %u bytes)\n"
1572 " ophdr + reg = %u bytes\n"
1573 " num regions = %u\n",
1574 ((ticket->t_trans_type <= 0 ||
1575 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1576 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1577 ticket->t_trans_type,
1578 ticket->t_unit_res,
1579 ticket->t_curr_res,
1580 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1581 ticket->t_res_num_ophdrs, ophdr_spc,
1582 ticket->t_res_arr_sum +
1583 ticket->t_res_o_flow + ophdr_spc,
1584 ticket->t_res_num);
1585
1586 for (i = 0; i < ticket->t_res_num; i++) {
1587 uint r_type = ticket->t_res_arr[i].r_type;
1588 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1589 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1590 "bad-rtype" : res_type_str[r_type-1]),
1591 ticket->t_res_arr[i].r_len);
1592 }
1593
1594 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1595 "xfs_log_write: reservation ran out. Need to up reservation");
1596 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1597}
1598
1599/*
1600 * Calculate the potential space needed by the log vector. Each region gets
1601 * its own xlog_op_header_t and may need to be double word aligned.
1602 */
1603static int
1604xlog_write_calc_vec_length(
1605 struct xlog_ticket *ticket,
1606 struct xfs_log_vec *log_vector)
1607{
1608 struct xfs_log_vec *lv;
1609 int headers = 0;
1610 int len = 0;
1611 int i;
1612
1613 /* acct for start rec of xact */
1614 if (ticket->t_flags & XLOG_TIC_INITED)
1615 headers++;
1616
1617 for (lv = log_vector; lv; lv = lv->lv_next) {
1618 headers += lv->lv_niovecs;
1619
1620 for (i = 0; i < lv->lv_niovecs; i++) {
1621 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
1622
1623 len += vecp->i_len;
1624 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1625 }
1626 }
1627
1628 ticket->t_res_num_ophdrs += headers;
1629 len += headers * sizeof(struct xlog_op_header);
1630
1631 return len;
1632}
1633
1634/*
1635 * If first write for transaction, insert start record We can't be trying to
1636 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1637 */
1638static int
1639xlog_write_start_rec(
1640 struct xlog_op_header *ophdr,
1641 struct xlog_ticket *ticket)
1642{
1643 if (!(ticket->t_flags & XLOG_TIC_INITED))
1644 return 0;
1645
1646 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1647 ophdr->oh_clientid = ticket->t_clientid;
1648 ophdr->oh_len = 0;
1649 ophdr->oh_flags = XLOG_START_TRANS;
1650 ophdr->oh_res2 = 0;
1651
1652 ticket->t_flags &= ~XLOG_TIC_INITED;
1653
1654 return sizeof(struct xlog_op_header);
1655}
1656
1657static xlog_op_header_t *
1658xlog_write_setup_ophdr(
1659 struct log *log,
1660 struct xlog_op_header *ophdr,
1661 struct xlog_ticket *ticket,
1662 uint flags)
1663{
1664 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1665 ophdr->oh_clientid = ticket->t_clientid;
1666 ophdr->oh_res2 = 0;
1667
1668 /* are we copying a commit or unmount record? */
1669 ophdr->oh_flags = flags;
1670
1671 /*
1672 * We've seen logs corrupted with bad transaction client ids. This
1673 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1674 * and shut down the filesystem.
1675 */
1676 switch (ophdr->oh_clientid) {
1677 case XFS_TRANSACTION:
1678 case XFS_VOLUME:
1679 case XFS_LOG:
1680 break;
1681 default:
1682 xfs_warn(log->l_mp,
1683 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1684 ophdr->oh_clientid, ticket);
1685 return NULL;
1686 }
1687
1688 return ophdr;
1689}
1690
1691/*
1692 * Set up the parameters of the region copy into the log. This has
1693 * to handle region write split across multiple log buffers - this
1694 * state is kept external to this function so that this code can
1695 * can be written in an obvious, self documenting manner.
1696 */
1697static int
1698xlog_write_setup_copy(
1699 struct xlog_ticket *ticket,
1700 struct xlog_op_header *ophdr,
1701 int space_available,
1702 int space_required,
1703 int *copy_off,
1704 int *copy_len,
1705 int *last_was_partial_copy,
1706 int *bytes_consumed)
1707{
1708 int still_to_copy;
1709
1710 still_to_copy = space_required - *bytes_consumed;
1711 *copy_off = *bytes_consumed;
1712
1713 if (still_to_copy <= space_available) {
1714 /* write of region completes here */
1715 *copy_len = still_to_copy;
1716 ophdr->oh_len = cpu_to_be32(*copy_len);
1717 if (*last_was_partial_copy)
1718 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1719 *last_was_partial_copy = 0;
1720 *bytes_consumed = 0;
1721 return 0;
1722 }
1723
1724 /* partial write of region, needs extra log op header reservation */
1725 *copy_len = space_available;
1726 ophdr->oh_len = cpu_to_be32(*copy_len);
1727 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1728 if (*last_was_partial_copy)
1729 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1730 *bytes_consumed += *copy_len;
1731 (*last_was_partial_copy)++;
1732
1733 /* account for new log op header */
1734 ticket->t_curr_res -= sizeof(struct xlog_op_header);
1735 ticket->t_res_num_ophdrs++;
1736
1737 return sizeof(struct xlog_op_header);
1738}
1739
1740static int
1741xlog_write_copy_finish(
1742 struct log *log,
1743 struct xlog_in_core *iclog,
1744 uint flags,
1745 int *record_cnt,
1746 int *data_cnt,
1747 int *partial_copy,
1748 int *partial_copy_len,
1749 int log_offset,
1750 struct xlog_in_core **commit_iclog)
1751{
1752 if (*partial_copy) {
1753 /*
1754 * This iclog has already been marked WANT_SYNC by
1755 * xlog_state_get_iclog_space.
1756 */
1757 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1758 *record_cnt = 0;
1759 *data_cnt = 0;
1760 return xlog_state_release_iclog(log, iclog);
1761 }
1762
1763 *partial_copy = 0;
1764 *partial_copy_len = 0;
1765
1766 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1767 /* no more space in this iclog - push it. */
1768 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1769 *record_cnt = 0;
1770 *data_cnt = 0;
1771
1772 spin_lock(&log->l_icloglock);
1773 xlog_state_want_sync(log, iclog);
1774 spin_unlock(&log->l_icloglock);
1775
1776 if (!commit_iclog)
1777 return xlog_state_release_iclog(log, iclog);
1778 ASSERT(flags & XLOG_COMMIT_TRANS);
1779 *commit_iclog = iclog;
1780 }
1781
1782 return 0;
1783}
1784
1785/*
1786 * Write some region out to in-core log
1787 *
1788 * This will be called when writing externally provided regions or when
1789 * writing out a commit record for a given transaction.
1790 *
1791 * General algorithm:
1792 * 1. Find total length of this write. This may include adding to the
1793 * lengths passed in.
1794 * 2. Check whether we violate the tickets reservation.
1795 * 3. While writing to this iclog
1796 * A. Reserve as much space in this iclog as can get
1797 * B. If this is first write, save away start lsn
1798 * C. While writing this region:
1799 * 1. If first write of transaction, write start record
1800 * 2. Write log operation header (header per region)
1801 * 3. Find out if we can fit entire region into this iclog
1802 * 4. Potentially, verify destination memcpy ptr
1803 * 5. Memcpy (partial) region
1804 * 6. If partial copy, release iclog; otherwise, continue
1805 * copying more regions into current iclog
1806 * 4. Mark want sync bit (in simulation mode)
1807 * 5. Release iclog for potential flush to on-disk log.
1808 *
1809 * ERRORS:
1810 * 1. Panic if reservation is overrun. This should never happen since
1811 * reservation amounts are generated internal to the filesystem.
1812 * NOTES:
1813 * 1. Tickets are single threaded data structures.
1814 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1815 * syncing routine. When a single log_write region needs to span
1816 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1817 * on all log operation writes which don't contain the end of the
1818 * region. The XLOG_END_TRANS bit is used for the in-core log
1819 * operation which contains the end of the continued log_write region.
1820 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1821 * we don't really know exactly how much space will be used. As a result,
1822 * we don't update ic_offset until the end when we know exactly how many
1823 * bytes have been written out.
1824 */
1825int
1826xlog_write(
1827 struct log *log,
1828 struct xfs_log_vec *log_vector,
1829 struct xlog_ticket *ticket,
1830 xfs_lsn_t *start_lsn,
1831 struct xlog_in_core **commit_iclog,
1832 uint flags)
1833{
1834 struct xlog_in_core *iclog = NULL;
1835 struct xfs_log_iovec *vecp;
1836 struct xfs_log_vec *lv;
1837 int len;
1838 int index;
1839 int partial_copy = 0;
1840 int partial_copy_len = 0;
1841 int contwr = 0;
1842 int record_cnt = 0;
1843 int data_cnt = 0;
1844 int error;
1845
1846 *start_lsn = 0;
1847
1848 len = xlog_write_calc_vec_length(ticket, log_vector);
1849 if (log->l_cilp) {
1850 /*
1851 * Region headers and bytes are already accounted for.
1852 * We only need to take into account start records and
1853 * split regions in this function.
1854 */
1855 if (ticket->t_flags & XLOG_TIC_INITED)
1856 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1857
1858 /*
1859 * Commit record headers need to be accounted for. These
1860 * come in as separate writes so are easy to detect.
1861 */
1862 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1863 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1864 } else
1865 ticket->t_curr_res -= len;
1866
1867 if (ticket->t_curr_res < 0)
1868 xlog_print_tic_res(log->l_mp, ticket);
1869
1870 index = 0;
1871 lv = log_vector;
1872 vecp = lv->lv_iovecp;
1873 while (lv && index < lv->lv_niovecs) {
1874 void *ptr;
1875 int log_offset;
1876
1877 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1878 &contwr, &log_offset);
1879 if (error)
1880 return error;
1881
1882 ASSERT(log_offset <= iclog->ic_size - 1);
1883 ptr = iclog->ic_datap + log_offset;
1884
1885 /* start_lsn is the first lsn written to. That's all we need. */
1886 if (!*start_lsn)
1887 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1888
1889 /*
1890 * This loop writes out as many regions as can fit in the amount
1891 * of space which was allocated by xlog_state_get_iclog_space().
1892 */
1893 while (lv && index < lv->lv_niovecs) {
1894 struct xfs_log_iovec *reg = &vecp[index];
1895 struct xlog_op_header *ophdr;
1896 int start_rec_copy;
1897 int copy_len;
1898 int copy_off;
1899
1900 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1901 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1902
1903 start_rec_copy = xlog_write_start_rec(ptr, ticket);
1904 if (start_rec_copy) {
1905 record_cnt++;
1906 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1907 start_rec_copy);
1908 }
1909
1910 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1911 if (!ophdr)
1912 return XFS_ERROR(EIO);
1913
1914 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1915 sizeof(struct xlog_op_header));
1916
1917 len += xlog_write_setup_copy(ticket, ophdr,
1918 iclog->ic_size-log_offset,
1919 reg->i_len,
1920 ©_off, ©_len,
1921 &partial_copy,
1922 &partial_copy_len);
1923 xlog_verify_dest_ptr(log, ptr);
1924
1925 /* copy region */
1926 ASSERT(copy_len >= 0);
1927 memcpy(ptr, reg->i_addr + copy_off, copy_len);
1928 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1929
1930 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1931 record_cnt++;
1932 data_cnt += contwr ? copy_len : 0;
1933
1934 error = xlog_write_copy_finish(log, iclog, flags,
1935 &record_cnt, &data_cnt,
1936 &partial_copy,
1937 &partial_copy_len,
1938 log_offset,
1939 commit_iclog);
1940 if (error)
1941 return error;
1942
1943 /*
1944 * if we had a partial copy, we need to get more iclog
1945 * space but we don't want to increment the region
1946 * index because there is still more is this region to
1947 * write.
1948 *
1949 * If we completed writing this region, and we flushed
1950 * the iclog (indicated by resetting of the record
1951 * count), then we also need to get more log space. If
1952 * this was the last record, though, we are done and
1953 * can just return.
1954 */
1955 if (partial_copy)
1956 break;
1957
1958 if (++index == lv->lv_niovecs) {
1959 lv = lv->lv_next;
1960 index = 0;
1961 if (lv)
1962 vecp = lv->lv_iovecp;
1963 }
1964 if (record_cnt == 0) {
1965 if (!lv)
1966 return 0;
1967 break;
1968 }
1969 }
1970 }
1971
1972 ASSERT(len == 0);
1973
1974 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1975 if (!commit_iclog)
1976 return xlog_state_release_iclog(log, iclog);
1977
1978 ASSERT(flags & XLOG_COMMIT_TRANS);
1979 *commit_iclog = iclog;
1980 return 0;
1981}
1982
1983
1984/*****************************************************************************
1985 *
1986 * State Machine functions
1987 *
1988 *****************************************************************************
1989 */
1990
1991/* Clean iclogs starting from the head. This ordering must be
1992 * maintained, so an iclog doesn't become ACTIVE beyond one that
1993 * is SYNCING. This is also required to maintain the notion that we use
1994 * a ordered wait queue to hold off would be writers to the log when every
1995 * iclog is trying to sync to disk.
1996 *
1997 * State Change: DIRTY -> ACTIVE
1998 */
1999STATIC void
2000xlog_state_clean_log(xlog_t *log)
2001{
2002 xlog_in_core_t *iclog;
2003 int changed = 0;
2004
2005 iclog = log->l_iclog;
2006 do {
2007 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2008 iclog->ic_state = XLOG_STATE_ACTIVE;
2009 iclog->ic_offset = 0;
2010 ASSERT(iclog->ic_callback == NULL);
2011 /*
2012 * If the number of ops in this iclog indicate it just
2013 * contains the dummy transaction, we can
2014 * change state into IDLE (the second time around).
2015 * Otherwise we should change the state into
2016 * NEED a dummy.
2017 * We don't need to cover the dummy.
2018 */
2019 if (!changed &&
2020 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2021 XLOG_COVER_OPS)) {
2022 changed = 1;
2023 } else {
2024 /*
2025 * We have two dirty iclogs so start over
2026 * This could also be num of ops indicates
2027 * this is not the dummy going out.
2028 */
2029 changed = 2;
2030 }
2031 iclog->ic_header.h_num_logops = 0;
2032 memset(iclog->ic_header.h_cycle_data, 0,
2033 sizeof(iclog->ic_header.h_cycle_data));
2034 iclog->ic_header.h_lsn = 0;
2035 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2036 /* do nothing */;
2037 else
2038 break; /* stop cleaning */
2039 iclog = iclog->ic_next;
2040 } while (iclog != log->l_iclog);
2041
2042 /* log is locked when we are called */
2043 /*
2044 * Change state for the dummy log recording.
2045 * We usually go to NEED. But we go to NEED2 if the changed indicates
2046 * we are done writing the dummy record.
2047 * If we are done with the second dummy recored (DONE2), then
2048 * we go to IDLE.
2049 */
2050 if (changed) {
2051 switch (log->l_covered_state) {
2052 case XLOG_STATE_COVER_IDLE:
2053 case XLOG_STATE_COVER_NEED:
2054 case XLOG_STATE_COVER_NEED2:
2055 log->l_covered_state = XLOG_STATE_COVER_NEED;
2056 break;
2057
2058 case XLOG_STATE_COVER_DONE:
2059 if (changed == 1)
2060 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2061 else
2062 log->l_covered_state = XLOG_STATE_COVER_NEED;
2063 break;
2064
2065 case XLOG_STATE_COVER_DONE2:
2066 if (changed == 1)
2067 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2068 else
2069 log->l_covered_state = XLOG_STATE_COVER_NEED;
2070 break;
2071
2072 default:
2073 ASSERT(0);
2074 }
2075 }
2076} /* xlog_state_clean_log */
2077
2078STATIC xfs_lsn_t
2079xlog_get_lowest_lsn(
2080 xlog_t *log)
2081{
2082 xlog_in_core_t *lsn_log;
2083 xfs_lsn_t lowest_lsn, lsn;
2084
2085 lsn_log = log->l_iclog;
2086 lowest_lsn = 0;
2087 do {
2088 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2089 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2090 if ((lsn && !lowest_lsn) ||
2091 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2092 lowest_lsn = lsn;
2093 }
2094 }
2095 lsn_log = lsn_log->ic_next;
2096 } while (lsn_log != log->l_iclog);
2097 return lowest_lsn;
2098}
2099
2100
2101STATIC void
2102xlog_state_do_callback(
2103 xlog_t *log,
2104 int aborted,
2105 xlog_in_core_t *ciclog)
2106{
2107 xlog_in_core_t *iclog;
2108 xlog_in_core_t *first_iclog; /* used to know when we've
2109 * processed all iclogs once */
2110 xfs_log_callback_t *cb, *cb_next;
2111 int flushcnt = 0;
2112 xfs_lsn_t lowest_lsn;
2113 int ioerrors; /* counter: iclogs with errors */
2114 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2115 int funcdidcallbacks; /* flag: function did callbacks */
2116 int repeats; /* for issuing console warnings if
2117 * looping too many times */
2118 int wake = 0;
2119
2120 spin_lock(&log->l_icloglock);
2121 first_iclog = iclog = log->l_iclog;
2122 ioerrors = 0;
2123 funcdidcallbacks = 0;
2124 repeats = 0;
2125
2126 do {
2127 /*
2128 * Scan all iclogs starting with the one pointed to by the
2129 * log. Reset this starting point each time the log is
2130 * unlocked (during callbacks).
2131 *
2132 * Keep looping through iclogs until one full pass is made
2133 * without running any callbacks.
2134 */
2135 first_iclog = log->l_iclog;
2136 iclog = log->l_iclog;
2137 loopdidcallbacks = 0;
2138 repeats++;
2139
2140 do {
2141
2142 /* skip all iclogs in the ACTIVE & DIRTY states */
2143 if (iclog->ic_state &
2144 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2145 iclog = iclog->ic_next;
2146 continue;
2147 }
2148
2149 /*
2150 * Between marking a filesystem SHUTDOWN and stopping
2151 * the log, we do flush all iclogs to disk (if there
2152 * wasn't a log I/O error). So, we do want things to
2153 * go smoothly in case of just a SHUTDOWN w/o a
2154 * LOG_IO_ERROR.
2155 */
2156 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2157 /*
2158 * Can only perform callbacks in order. Since
2159 * this iclog is not in the DONE_SYNC/
2160 * DO_CALLBACK state, we skip the rest and
2161 * just try to clean up. If we set our iclog
2162 * to DO_CALLBACK, we will not process it when
2163 * we retry since a previous iclog is in the
2164 * CALLBACK and the state cannot change since
2165 * we are holding the l_icloglock.
2166 */
2167 if (!(iclog->ic_state &
2168 (XLOG_STATE_DONE_SYNC |
2169 XLOG_STATE_DO_CALLBACK))) {
2170 if (ciclog && (ciclog->ic_state ==
2171 XLOG_STATE_DONE_SYNC)) {
2172 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2173 }
2174 break;
2175 }
2176 /*
2177 * We now have an iclog that is in either the
2178 * DO_CALLBACK or DONE_SYNC states. The other
2179 * states (WANT_SYNC, SYNCING, or CALLBACK were
2180 * caught by the above if and are going to
2181 * clean (i.e. we aren't doing their callbacks)
2182 * see the above if.
2183 */
2184
2185 /*
2186 * We will do one more check here to see if we
2187 * have chased our tail around.
2188 */
2189
2190 lowest_lsn = xlog_get_lowest_lsn(log);
2191 if (lowest_lsn &&
2192 XFS_LSN_CMP(lowest_lsn,
2193 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2194 iclog = iclog->ic_next;
2195 continue; /* Leave this iclog for
2196 * another thread */
2197 }
2198
2199 iclog->ic_state = XLOG_STATE_CALLBACK;
2200
2201
2202 /*
2203 * update the last_sync_lsn before we drop the
2204 * icloglock to ensure we are the only one that
2205 * can update it.
2206 */
2207 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2208 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2209 atomic64_set(&log->l_last_sync_lsn,
2210 be64_to_cpu(iclog->ic_header.h_lsn));
2211
2212 } else
2213 ioerrors++;
2214
2215 spin_unlock(&log->l_icloglock);
2216
2217 /*
2218 * Keep processing entries in the callback list until
2219 * we come around and it is empty. We need to
2220 * atomically see that the list is empty and change the
2221 * state to DIRTY so that we don't miss any more
2222 * callbacks being added.
2223 */
2224 spin_lock(&iclog->ic_callback_lock);
2225 cb = iclog->ic_callback;
2226 while (cb) {
2227 iclog->ic_callback_tail = &(iclog->ic_callback);
2228 iclog->ic_callback = NULL;
2229 spin_unlock(&iclog->ic_callback_lock);
2230
2231 /* perform callbacks in the order given */
2232 for (; cb; cb = cb_next) {
2233 cb_next = cb->cb_next;
2234 cb->cb_func(cb->cb_arg, aborted);
2235 }
2236 spin_lock(&iclog->ic_callback_lock);
2237 cb = iclog->ic_callback;
2238 }
2239
2240 loopdidcallbacks++;
2241 funcdidcallbacks++;
2242
2243 spin_lock(&log->l_icloglock);
2244 ASSERT(iclog->ic_callback == NULL);
2245 spin_unlock(&iclog->ic_callback_lock);
2246 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2247 iclog->ic_state = XLOG_STATE_DIRTY;
2248
2249 /*
2250 * Transition from DIRTY to ACTIVE if applicable.
2251 * NOP if STATE_IOERROR.
2252 */
2253 xlog_state_clean_log(log);
2254
2255 /* wake up threads waiting in xfs_log_force() */
2256 wake_up_all(&iclog->ic_force_wait);
2257
2258 iclog = iclog->ic_next;
2259 } while (first_iclog != iclog);
2260
2261 if (repeats > 5000) {
2262 flushcnt += repeats;
2263 repeats = 0;
2264 xfs_warn(log->l_mp,
2265 "%s: possible infinite loop (%d iterations)",
2266 __func__, flushcnt);
2267 }
2268 } while (!ioerrors && loopdidcallbacks);
2269
2270 /*
2271 * make one last gasp attempt to see if iclogs are being left in
2272 * limbo..
2273 */
2274#ifdef DEBUG
2275 if (funcdidcallbacks) {
2276 first_iclog = iclog = log->l_iclog;
2277 do {
2278 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2279 /*
2280 * Terminate the loop if iclogs are found in states
2281 * which will cause other threads to clean up iclogs.
2282 *
2283 * SYNCING - i/o completion will go through logs
2284 * DONE_SYNC - interrupt thread should be waiting for
2285 * l_icloglock
2286 * IOERROR - give up hope all ye who enter here
2287 */
2288 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2289 iclog->ic_state == XLOG_STATE_SYNCING ||
2290 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2291 iclog->ic_state == XLOG_STATE_IOERROR )
2292 break;
2293 iclog = iclog->ic_next;
2294 } while (first_iclog != iclog);
2295 }
2296#endif
2297
2298 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2299 wake = 1;
2300 spin_unlock(&log->l_icloglock);
2301
2302 if (wake)
2303 wake_up_all(&log->l_flush_wait);
2304}
2305
2306
2307/*
2308 * Finish transitioning this iclog to the dirty state.
2309 *
2310 * Make sure that we completely execute this routine only when this is
2311 * the last call to the iclog. There is a good chance that iclog flushes,
2312 * when we reach the end of the physical log, get turned into 2 separate
2313 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2314 * routine. By using the reference count bwritecnt, we guarantee that only
2315 * the second completion goes through.
2316 *
2317 * Callbacks could take time, so they are done outside the scope of the
2318 * global state machine log lock.
2319 */
2320STATIC void
2321xlog_state_done_syncing(
2322 xlog_in_core_t *iclog,
2323 int aborted)
2324{
2325 xlog_t *log = iclog->ic_log;
2326
2327 spin_lock(&log->l_icloglock);
2328
2329 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2330 iclog->ic_state == XLOG_STATE_IOERROR);
2331 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2332 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2333
2334
2335 /*
2336 * If we got an error, either on the first buffer, or in the case of
2337 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2338 * and none should ever be attempted to be written to disk
2339 * again.
2340 */
2341 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2342 if (--iclog->ic_bwritecnt == 1) {
2343 spin_unlock(&log->l_icloglock);
2344 return;
2345 }
2346 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2347 }
2348
2349 /*
2350 * Someone could be sleeping prior to writing out the next
2351 * iclog buffer, we wake them all, one will get to do the
2352 * I/O, the others get to wait for the result.
2353 */
2354 wake_up_all(&iclog->ic_write_wait);
2355 spin_unlock(&log->l_icloglock);
2356 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2357} /* xlog_state_done_syncing */
2358
2359
2360/*
2361 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2362 * sleep. We wait on the flush queue on the head iclog as that should be
2363 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2364 * we will wait here and all new writes will sleep until a sync completes.
2365 *
2366 * The in-core logs are used in a circular fashion. They are not used
2367 * out-of-order even when an iclog past the head is free.
2368 *
2369 * return:
2370 * * log_offset where xlog_write() can start writing into the in-core
2371 * log's data space.
2372 * * in-core log pointer to which xlog_write() should write.
2373 * * boolean indicating this is a continued write to an in-core log.
2374 * If this is the last write, then the in-core log's offset field
2375 * needs to be incremented, depending on the amount of data which
2376 * is copied.
2377 */
2378STATIC int
2379xlog_state_get_iclog_space(xlog_t *log,
2380 int len,
2381 xlog_in_core_t **iclogp,
2382 xlog_ticket_t *ticket,
2383 int *continued_write,
2384 int *logoffsetp)
2385{
2386 int log_offset;
2387 xlog_rec_header_t *head;
2388 xlog_in_core_t *iclog;
2389 int error;
2390
2391restart:
2392 spin_lock(&log->l_icloglock);
2393 if (XLOG_FORCED_SHUTDOWN(log)) {
2394 spin_unlock(&log->l_icloglock);
2395 return XFS_ERROR(EIO);
2396 }
2397
2398 iclog = log->l_iclog;
2399 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2400 XFS_STATS_INC(xs_log_noiclogs);
2401
2402 /* Wait for log writes to have flushed */
2403 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2404 goto restart;
2405 }
2406
2407 head = &iclog->ic_header;
2408
2409 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2410 log_offset = iclog->ic_offset;
2411
2412 /* On the 1st write to an iclog, figure out lsn. This works
2413 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2414 * committing to. If the offset is set, that's how many blocks
2415 * must be written.
2416 */
2417 if (log_offset == 0) {
2418 ticket->t_curr_res -= log->l_iclog_hsize;
2419 xlog_tic_add_region(ticket,
2420 log->l_iclog_hsize,
2421 XLOG_REG_TYPE_LRHEADER);
2422 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2423 head->h_lsn = cpu_to_be64(
2424 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2425 ASSERT(log->l_curr_block >= 0);
2426 }
2427
2428 /* If there is enough room to write everything, then do it. Otherwise,
2429 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2430 * bit is on, so this will get flushed out. Don't update ic_offset
2431 * until you know exactly how many bytes get copied. Therefore, wait
2432 * until later to update ic_offset.
2433 *
2434 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2435 * can fit into remaining data section.
2436 */
2437 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2438 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2439
2440 /*
2441 * If I'm the only one writing to this iclog, sync it to disk.
2442 * We need to do an atomic compare and decrement here to avoid
2443 * racing with concurrent atomic_dec_and_lock() calls in
2444 * xlog_state_release_iclog() when there is more than one
2445 * reference to the iclog.
2446 */
2447 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2448 /* we are the only one */
2449 spin_unlock(&log->l_icloglock);
2450 error = xlog_state_release_iclog(log, iclog);
2451 if (error)
2452 return error;
2453 } else {
2454 spin_unlock(&log->l_icloglock);
2455 }
2456 goto restart;
2457 }
2458
2459 /* Do we have enough room to write the full amount in the remainder
2460 * of this iclog? Or must we continue a write on the next iclog and
2461 * mark this iclog as completely taken? In the case where we switch
2462 * iclogs (to mark it taken), this particular iclog will release/sync
2463 * to disk in xlog_write().
2464 */
2465 if (len <= iclog->ic_size - iclog->ic_offset) {
2466 *continued_write = 0;
2467 iclog->ic_offset += len;
2468 } else {
2469 *continued_write = 1;
2470 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2471 }
2472 *iclogp = iclog;
2473
2474 ASSERT(iclog->ic_offset <= iclog->ic_size);
2475 spin_unlock(&log->l_icloglock);
2476
2477 *logoffsetp = log_offset;
2478 return 0;
2479} /* xlog_state_get_iclog_space */
2480
2481/*
2482 * Atomically get the log space required for a log ticket.
2483 *
2484 * Once a ticket gets put onto the reserveq, it will only return after
2485 * the needed reservation is satisfied.
2486 *
2487 * This function is structured so that it has a lock free fast path. This is
2488 * necessary because every new transaction reservation will come through this
2489 * path. Hence any lock will be globally hot if we take it unconditionally on
2490 * every pass.
2491 *
2492 * As tickets are only ever moved on and off the reserveq under the
2493 * l_grant_reserve_lock, we only need to take that lock if we are going
2494 * to add the ticket to the queue and sleep. We can avoid taking the lock if the
2495 * ticket was never added to the reserveq because the t_queue list head will be
2496 * empty and we hold the only reference to it so it can safely be checked
2497 * unlocked.
2498 */
2499STATIC int
2500xlog_grant_log_space(xlog_t *log,
2501 xlog_ticket_t *tic)
2502{
2503 int free_bytes;
2504 int need_bytes;
2505
2506#ifdef DEBUG
2507 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2508 panic("grant Recovery problem");
2509#endif
2510
2511 trace_xfs_log_grant_enter(log, tic);
2512
2513 need_bytes = tic->t_unit_res;
2514 if (tic->t_flags & XFS_LOG_PERM_RESERV)
2515 need_bytes *= tic->t_ocnt;
2516
2517 /* something is already sleeping; insert new transaction at end */
2518 if (!list_empty_careful(&log->l_reserveq)) {
2519 spin_lock(&log->l_grant_reserve_lock);
2520 /* recheck the queue now we are locked */
2521 if (list_empty(&log->l_reserveq)) {
2522 spin_unlock(&log->l_grant_reserve_lock);
2523 goto redo;
2524 }
2525 list_add_tail(&tic->t_queue, &log->l_reserveq);
2526
2527 trace_xfs_log_grant_sleep1(log, tic);
2528
2529 /*
2530 * Gotta check this before going to sleep, while we're
2531 * holding the grant lock.
2532 */
2533 if (XLOG_FORCED_SHUTDOWN(log))
2534 goto error_return;
2535
2536 XFS_STATS_INC(xs_sleep_logspace);
2537 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2538
2539 /*
2540 * If we got an error, and the filesystem is shutting down,
2541 * we'll catch it down below. So just continue...
2542 */
2543 trace_xfs_log_grant_wake1(log, tic);
2544 }
2545
2546redo:
2547 if (XLOG_FORCED_SHUTDOWN(log))
2548 goto error_return_unlocked;
2549
2550 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
2551 if (free_bytes < need_bytes) {
2552 spin_lock(&log->l_grant_reserve_lock);
2553 if (list_empty(&tic->t_queue))
2554 list_add_tail(&tic->t_queue, &log->l_reserveq);
2555
2556 trace_xfs_log_grant_sleep2(log, tic);
2557
2558 if (XLOG_FORCED_SHUTDOWN(log))
2559 goto error_return;
2560
2561 xlog_grant_push_ail(log, need_bytes);
2562
2563 XFS_STATS_INC(xs_sleep_logspace);
2564 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2565
2566 trace_xfs_log_grant_wake2(log, tic);
2567 goto redo;
2568 }
2569
2570 if (!list_empty(&tic->t_queue)) {
2571 spin_lock(&log->l_grant_reserve_lock);
2572 list_del_init(&tic->t_queue);
2573 spin_unlock(&log->l_grant_reserve_lock);
2574 }
2575
2576 /* we've got enough space */
2577 xlog_grant_add_space(log, &log->l_grant_reserve_head, need_bytes);
2578 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2579 trace_xfs_log_grant_exit(log, tic);
2580 xlog_verify_grant_tail(log);
2581 return 0;
2582
2583error_return_unlocked:
2584 spin_lock(&log->l_grant_reserve_lock);
2585error_return:
2586 list_del_init(&tic->t_queue);
2587 spin_unlock(&log->l_grant_reserve_lock);
2588 trace_xfs_log_grant_error(log, tic);
2589
2590 /*
2591 * If we are failing, make sure the ticket doesn't have any
2592 * current reservations. We don't want to add this back when
2593 * the ticket/transaction gets cancelled.
2594 */
2595 tic->t_curr_res = 0;
2596 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2597 return XFS_ERROR(EIO);
2598} /* xlog_grant_log_space */
2599
2600
2601/*
2602 * Replenish the byte reservation required by moving the grant write head.
2603 *
2604 * Similar to xlog_grant_log_space, the function is structured to have a lock
2605 * free fast path.
2606 */
2607STATIC int
2608xlog_regrant_write_log_space(xlog_t *log,
2609 xlog_ticket_t *tic)
2610{
2611 int free_bytes, need_bytes;
2612
2613 tic->t_curr_res = tic->t_unit_res;
2614 xlog_tic_reset_res(tic);
2615
2616 if (tic->t_cnt > 0)
2617 return 0;
2618
2619#ifdef DEBUG
2620 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2621 panic("regrant Recovery problem");
2622#endif
2623
2624 trace_xfs_log_regrant_write_enter(log, tic);
2625 if (XLOG_FORCED_SHUTDOWN(log))
2626 goto error_return_unlocked;
2627
2628 /* If there are other waiters on the queue then give them a
2629 * chance at logspace before us. Wake up the first waiters,
2630 * if we do not wake up all the waiters then go to sleep waiting
2631 * for more free space, otherwise try to get some space for
2632 * this transaction.
2633 */
2634 need_bytes = tic->t_unit_res;
2635 if (!list_empty_careful(&log->l_writeq)) {
2636 struct xlog_ticket *ntic;
2637
2638 spin_lock(&log->l_grant_write_lock);
2639 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2640 list_for_each_entry(ntic, &log->l_writeq, t_queue) {
2641 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2642
2643 if (free_bytes < ntic->t_unit_res)
2644 break;
2645 free_bytes -= ntic->t_unit_res;
2646 wake_up(&ntic->t_wait);
2647 }
2648
2649 if (ntic != list_first_entry(&log->l_writeq,
2650 struct xlog_ticket, t_queue)) {
2651 if (list_empty(&tic->t_queue))
2652 list_add_tail(&tic->t_queue, &log->l_writeq);
2653 trace_xfs_log_regrant_write_sleep1(log, tic);
2654
2655 xlog_grant_push_ail(log, need_bytes);
2656
2657 XFS_STATS_INC(xs_sleep_logspace);
2658 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2659 trace_xfs_log_regrant_write_wake1(log, tic);
2660 } else
2661 spin_unlock(&log->l_grant_write_lock);
2662 }
2663
2664redo:
2665 if (XLOG_FORCED_SHUTDOWN(log))
2666 goto error_return_unlocked;
2667
2668 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2669 if (free_bytes < need_bytes) {
2670 spin_lock(&log->l_grant_write_lock);
2671 if (list_empty(&tic->t_queue))
2672 list_add_tail(&tic->t_queue, &log->l_writeq);
2673
2674 if (XLOG_FORCED_SHUTDOWN(log))
2675 goto error_return;
2676
2677 xlog_grant_push_ail(log, need_bytes);
2678
2679 XFS_STATS_INC(xs_sleep_logspace);
2680 trace_xfs_log_regrant_write_sleep2(log, tic);
2681 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2682
2683 trace_xfs_log_regrant_write_wake2(log, tic);
2684 goto redo;
2685 }
2686
2687 if (!list_empty(&tic->t_queue)) {
2688 spin_lock(&log->l_grant_write_lock);
2689 list_del_init(&tic->t_queue);
2690 spin_unlock(&log->l_grant_write_lock);
2691 }
2692
2693 /* we've got enough space */
2694 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2695 trace_xfs_log_regrant_write_exit(log, tic);
2696 xlog_verify_grant_tail(log);
2697 return 0;
2698
2699
2700 error_return_unlocked:
2701 spin_lock(&log->l_grant_write_lock);
2702 error_return:
2703 list_del_init(&tic->t_queue);
2704 spin_unlock(&log->l_grant_write_lock);
2705 trace_xfs_log_regrant_write_error(log, tic);
2706
2707 /*
2708 * If we are failing, make sure the ticket doesn't have any
2709 * current reservations. We don't want to add this back when
2710 * the ticket/transaction gets cancelled.
2711 */
2712 tic->t_curr_res = 0;
2713 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2714 return XFS_ERROR(EIO);
2715} /* xlog_regrant_write_log_space */
2716
2717
2718/* The first cnt-1 times through here we don't need to
2719 * move the grant write head because the permanent
2720 * reservation has reserved cnt times the unit amount.
2721 * Release part of current permanent unit reservation and
2722 * reset current reservation to be one units worth. Also
2723 * move grant reservation head forward.
2724 */
2725STATIC void
2726xlog_regrant_reserve_log_space(xlog_t *log,
2727 xlog_ticket_t *ticket)
2728{
2729 trace_xfs_log_regrant_reserve_enter(log, ticket);
2730
2731 if (ticket->t_cnt > 0)
2732 ticket->t_cnt--;
2733
2734 xlog_grant_sub_space(log, &log->l_grant_reserve_head,
2735 ticket->t_curr_res);
2736 xlog_grant_sub_space(log, &log->l_grant_write_head,
2737 ticket->t_curr_res);
2738 ticket->t_curr_res = ticket->t_unit_res;
2739 xlog_tic_reset_res(ticket);
2740
2741 trace_xfs_log_regrant_reserve_sub(log, ticket);
2742
2743 /* just return if we still have some of the pre-reserved space */
2744 if (ticket->t_cnt > 0)
2745 return;
2746
2747 xlog_grant_add_space(log, &log->l_grant_reserve_head,
2748 ticket->t_unit_res);
2749
2750 trace_xfs_log_regrant_reserve_exit(log, ticket);
2751
2752 ticket->t_curr_res = ticket->t_unit_res;
2753 xlog_tic_reset_res(ticket);
2754} /* xlog_regrant_reserve_log_space */
2755
2756
2757/*
2758 * Give back the space left from a reservation.
2759 *
2760 * All the information we need to make a correct determination of space left
2761 * is present. For non-permanent reservations, things are quite easy. The
2762 * count should have been decremented to zero. We only need to deal with the
2763 * space remaining in the current reservation part of the ticket. If the
2764 * ticket contains a permanent reservation, there may be left over space which
2765 * needs to be released. A count of N means that N-1 refills of the current
2766 * reservation can be done before we need to ask for more space. The first
2767 * one goes to fill up the first current reservation. Once we run out of
2768 * space, the count will stay at zero and the only space remaining will be
2769 * in the current reservation field.
2770 */
2771STATIC void
2772xlog_ungrant_log_space(xlog_t *log,
2773 xlog_ticket_t *ticket)
2774{
2775 int bytes;
2776
2777 if (ticket->t_cnt > 0)
2778 ticket->t_cnt--;
2779
2780 trace_xfs_log_ungrant_enter(log, ticket);
2781 trace_xfs_log_ungrant_sub(log, ticket);
2782
2783 /*
2784 * If this is a permanent reservation ticket, we may be able to free
2785 * up more space based on the remaining count.
2786 */
2787 bytes = ticket->t_curr_res;
2788 if (ticket->t_cnt > 0) {
2789 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2790 bytes += ticket->t_unit_res*ticket->t_cnt;
2791 }
2792
2793 xlog_grant_sub_space(log, &log->l_grant_reserve_head, bytes);
2794 xlog_grant_sub_space(log, &log->l_grant_write_head, bytes);
2795
2796 trace_xfs_log_ungrant_exit(log, ticket);
2797
2798 xfs_log_move_tail(log->l_mp, 1);
2799} /* xlog_ungrant_log_space */
2800
2801
2802/*
2803 * Flush iclog to disk if this is the last reference to the given iclog and
2804 * the WANT_SYNC bit is set.
2805 *
2806 * When this function is entered, the iclog is not necessarily in the
2807 * WANT_SYNC state. It may be sitting around waiting to get filled.
2808 *
2809 *
2810 */
2811STATIC int
2812xlog_state_release_iclog(
2813 xlog_t *log,
2814 xlog_in_core_t *iclog)
2815{
2816 int sync = 0; /* do we sync? */
2817
2818 if (iclog->ic_state & XLOG_STATE_IOERROR)
2819 return XFS_ERROR(EIO);
2820
2821 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2822 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2823 return 0;
2824
2825 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2826 spin_unlock(&log->l_icloglock);
2827 return XFS_ERROR(EIO);
2828 }
2829 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2830 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2831
2832 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2833 /* update tail before writing to iclog */
2834 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2835 sync++;
2836 iclog->ic_state = XLOG_STATE_SYNCING;
2837 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2838 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2839 /* cycle incremented when incrementing curr_block */
2840 }
2841 spin_unlock(&log->l_icloglock);
2842
2843 /*
2844 * We let the log lock go, so it's possible that we hit a log I/O
2845 * error or some other SHUTDOWN condition that marks the iclog
2846 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2847 * this iclog has consistent data, so we ignore IOERROR
2848 * flags after this point.
2849 */
2850 if (sync)
2851 return xlog_sync(log, iclog);
2852 return 0;
2853} /* xlog_state_release_iclog */
2854
2855
2856/*
2857 * This routine will mark the current iclog in the ring as WANT_SYNC
2858 * and move the current iclog pointer to the next iclog in the ring.
2859 * When this routine is called from xlog_state_get_iclog_space(), the
2860 * exact size of the iclog has not yet been determined. All we know is
2861 * that every data block. We have run out of space in this log record.
2862 */
2863STATIC void
2864xlog_state_switch_iclogs(xlog_t *log,
2865 xlog_in_core_t *iclog,
2866 int eventual_size)
2867{
2868 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2869 if (!eventual_size)
2870 eventual_size = iclog->ic_offset;
2871 iclog->ic_state = XLOG_STATE_WANT_SYNC;
2872 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2873 log->l_prev_block = log->l_curr_block;
2874 log->l_prev_cycle = log->l_curr_cycle;
2875
2876 /* roll log?: ic_offset changed later */
2877 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2878
2879 /* Round up to next log-sunit */
2880 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2881 log->l_mp->m_sb.sb_logsunit > 1) {
2882 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2883 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2884 }
2885
2886 if (log->l_curr_block >= log->l_logBBsize) {
2887 log->l_curr_cycle++;
2888 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2889 log->l_curr_cycle++;
2890 log->l_curr_block -= log->l_logBBsize;
2891 ASSERT(log->l_curr_block >= 0);
2892 }
2893 ASSERT(iclog == log->l_iclog);
2894 log->l_iclog = iclog->ic_next;
2895} /* xlog_state_switch_iclogs */
2896
2897/*
2898 * Write out all data in the in-core log as of this exact moment in time.
2899 *
2900 * Data may be written to the in-core log during this call. However,
2901 * we don't guarantee this data will be written out. A change from past
2902 * implementation means this routine will *not* write out zero length LRs.
2903 *
2904 * Basically, we try and perform an intelligent scan of the in-core logs.
2905 * If we determine there is no flushable data, we just return. There is no
2906 * flushable data if:
2907 *
2908 * 1. the current iclog is active and has no data; the previous iclog
2909 * is in the active or dirty state.
2910 * 2. the current iclog is drity, and the previous iclog is in the
2911 * active or dirty state.
2912 *
2913 * We may sleep if:
2914 *
2915 * 1. the current iclog is not in the active nor dirty state.
2916 * 2. the current iclog dirty, and the previous iclog is not in the
2917 * active nor dirty state.
2918 * 3. the current iclog is active, and there is another thread writing
2919 * to this particular iclog.
2920 * 4. a) the current iclog is active and has no other writers
2921 * b) when we return from flushing out this iclog, it is still
2922 * not in the active nor dirty state.
2923 */
2924int
2925_xfs_log_force(
2926 struct xfs_mount *mp,
2927 uint flags,
2928 int *log_flushed)
2929{
2930 struct log *log = mp->m_log;
2931 struct xlog_in_core *iclog;
2932 xfs_lsn_t lsn;
2933
2934 XFS_STATS_INC(xs_log_force);
2935
2936 if (log->l_cilp)
2937 xlog_cil_force(log);
2938
2939 spin_lock(&log->l_icloglock);
2940
2941 iclog = log->l_iclog;
2942 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2943 spin_unlock(&log->l_icloglock);
2944 return XFS_ERROR(EIO);
2945 }
2946
2947 /* If the head iclog is not active nor dirty, we just attach
2948 * ourselves to the head and go to sleep.
2949 */
2950 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2951 iclog->ic_state == XLOG_STATE_DIRTY) {
2952 /*
2953 * If the head is dirty or (active and empty), then
2954 * we need to look at the previous iclog. If the previous
2955 * iclog is active or dirty we are done. There is nothing
2956 * to sync out. Otherwise, we attach ourselves to the
2957 * previous iclog and go to sleep.
2958 */
2959 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2960 (atomic_read(&iclog->ic_refcnt) == 0
2961 && iclog->ic_offset == 0)) {
2962 iclog = iclog->ic_prev;
2963 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2964 iclog->ic_state == XLOG_STATE_DIRTY)
2965 goto no_sleep;
2966 else
2967 goto maybe_sleep;
2968 } else {
2969 if (atomic_read(&iclog->ic_refcnt) == 0) {
2970 /* We are the only one with access to this
2971 * iclog. Flush it out now. There should
2972 * be a roundoff of zero to show that someone
2973 * has already taken care of the roundoff from
2974 * the previous sync.
2975 */
2976 atomic_inc(&iclog->ic_refcnt);
2977 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2978 xlog_state_switch_iclogs(log, iclog, 0);
2979 spin_unlock(&log->l_icloglock);
2980
2981 if (xlog_state_release_iclog(log, iclog))
2982 return XFS_ERROR(EIO);
2983
2984 if (log_flushed)
2985 *log_flushed = 1;
2986 spin_lock(&log->l_icloglock);
2987 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2988 iclog->ic_state != XLOG_STATE_DIRTY)
2989 goto maybe_sleep;
2990 else
2991 goto no_sleep;
2992 } else {
2993 /* Someone else is writing to this iclog.
2994 * Use its call to flush out the data. However,
2995 * the other thread may not force out this LR,
2996 * so we mark it WANT_SYNC.
2997 */
2998 xlog_state_switch_iclogs(log, iclog, 0);
2999 goto maybe_sleep;
3000 }
3001 }
3002 }
3003
3004 /* By the time we come around again, the iclog could've been filled
3005 * which would give it another lsn. If we have a new lsn, just
3006 * return because the relevant data has been flushed.
3007 */
3008maybe_sleep:
3009 if (flags & XFS_LOG_SYNC) {
3010 /*
3011 * We must check if we're shutting down here, before
3012 * we wait, while we're holding the l_icloglock.
3013 * Then we check again after waking up, in case our
3014 * sleep was disturbed by a bad news.
3015 */
3016 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3017 spin_unlock(&log->l_icloglock);
3018 return XFS_ERROR(EIO);
3019 }
3020 XFS_STATS_INC(xs_log_force_sleep);
3021 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3022 /*
3023 * No need to grab the log lock here since we're
3024 * only deciding whether or not to return EIO
3025 * and the memory read should be atomic.
3026 */
3027 if (iclog->ic_state & XLOG_STATE_IOERROR)
3028 return XFS_ERROR(EIO);
3029 if (log_flushed)
3030 *log_flushed = 1;
3031 } else {
3032
3033no_sleep:
3034 spin_unlock(&log->l_icloglock);
3035 }
3036 return 0;
3037}
3038
3039/*
3040 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3041 * about errors or whether the log was flushed or not. This is the normal
3042 * interface to use when trying to unpin items or move the log forward.
3043 */
3044void
3045xfs_log_force(
3046 xfs_mount_t *mp,
3047 uint flags)
3048{
3049 int error;
3050
3051 error = _xfs_log_force(mp, flags, NULL);
3052 if (error)
3053 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3054}
3055
3056/*
3057 * Force the in-core log to disk for a specific LSN.
3058 *
3059 * Find in-core log with lsn.
3060 * If it is in the DIRTY state, just return.
3061 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3062 * state and go to sleep or return.
3063 * If it is in any other state, go to sleep or return.
3064 *
3065 * Synchronous forces are implemented with a signal variable. All callers
3066 * to force a given lsn to disk will wait on a the sv attached to the
3067 * specific in-core log. When given in-core log finally completes its
3068 * write to disk, that thread will wake up all threads waiting on the
3069 * sv.
3070 */
3071int
3072_xfs_log_force_lsn(
3073 struct xfs_mount *mp,
3074 xfs_lsn_t lsn,
3075 uint flags,
3076 int *log_flushed)
3077{
3078 struct log *log = mp->m_log;
3079 struct xlog_in_core *iclog;
3080 int already_slept = 0;
3081
3082 ASSERT(lsn != 0);
3083
3084 XFS_STATS_INC(xs_log_force);
3085
3086 if (log->l_cilp) {
3087 lsn = xlog_cil_force_lsn(log, lsn);
3088 if (lsn == NULLCOMMITLSN)
3089 return 0;
3090 }
3091
3092try_again:
3093 spin_lock(&log->l_icloglock);
3094 iclog = log->l_iclog;
3095 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3096 spin_unlock(&log->l_icloglock);
3097 return XFS_ERROR(EIO);
3098 }
3099
3100 do {
3101 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3102 iclog = iclog->ic_next;
3103 continue;
3104 }
3105
3106 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3107 spin_unlock(&log->l_icloglock);
3108 return 0;
3109 }
3110
3111 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3112 /*
3113 * We sleep here if we haven't already slept (e.g.
3114 * this is the first time we've looked at the correct
3115 * iclog buf) and the buffer before us is going to
3116 * be sync'ed. The reason for this is that if we
3117 * are doing sync transactions here, by waiting for
3118 * the previous I/O to complete, we can allow a few
3119 * more transactions into this iclog before we close
3120 * it down.
3121 *
3122 * Otherwise, we mark the buffer WANT_SYNC, and bump
3123 * up the refcnt so we can release the log (which
3124 * drops the ref count). The state switch keeps new
3125 * transaction commits from using this buffer. When
3126 * the current commits finish writing into the buffer,
3127 * the refcount will drop to zero and the buffer will
3128 * go out then.
3129 */
3130 if (!already_slept &&
3131 (iclog->ic_prev->ic_state &
3132 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3133 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3134
3135 XFS_STATS_INC(xs_log_force_sleep);
3136
3137 xlog_wait(&iclog->ic_prev->ic_write_wait,
3138 &log->l_icloglock);
3139 if (log_flushed)
3140 *log_flushed = 1;
3141 already_slept = 1;
3142 goto try_again;
3143 }
3144 atomic_inc(&iclog->ic_refcnt);
3145 xlog_state_switch_iclogs(log, iclog, 0);
3146 spin_unlock(&log->l_icloglock);
3147 if (xlog_state_release_iclog(log, iclog))
3148 return XFS_ERROR(EIO);
3149 if (log_flushed)
3150 *log_flushed = 1;
3151 spin_lock(&log->l_icloglock);
3152 }
3153
3154 if ((flags & XFS_LOG_SYNC) && /* sleep */
3155 !(iclog->ic_state &
3156 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3157 /*
3158 * Don't wait on completion if we know that we've
3159 * gotten a log write error.
3160 */
3161 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3162 spin_unlock(&log->l_icloglock);
3163 return XFS_ERROR(EIO);
3164 }
3165 XFS_STATS_INC(xs_log_force_sleep);
3166 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3167 /*
3168 * No need to grab the log lock here since we're
3169 * only deciding whether or not to return EIO
3170 * and the memory read should be atomic.
3171 */
3172 if (iclog->ic_state & XLOG_STATE_IOERROR)
3173 return XFS_ERROR(EIO);
3174
3175 if (log_flushed)
3176 *log_flushed = 1;
3177 } else { /* just return */
3178 spin_unlock(&log->l_icloglock);
3179 }
3180
3181 return 0;
3182 } while (iclog != log->l_iclog);
3183
3184 spin_unlock(&log->l_icloglock);
3185 return 0;
3186}
3187
3188/*
3189 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3190 * about errors or whether the log was flushed or not. This is the normal
3191 * interface to use when trying to unpin items or move the log forward.
3192 */
3193void
3194xfs_log_force_lsn(
3195 xfs_mount_t *mp,
3196 xfs_lsn_t lsn,
3197 uint flags)
3198{
3199 int error;
3200
3201 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3202 if (error)
3203 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3204}
3205
3206/*
3207 * Called when we want to mark the current iclog as being ready to sync to
3208 * disk.
3209 */
3210STATIC void
3211xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3212{
3213 assert_spin_locked(&log->l_icloglock);
3214
3215 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3216 xlog_state_switch_iclogs(log, iclog, 0);
3217 } else {
3218 ASSERT(iclog->ic_state &
3219 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3220 }
3221}
3222
3223
3224/*****************************************************************************
3225 *
3226 * TICKET functions
3227 *
3228 *****************************************************************************
3229 */
3230
3231/*
3232 * Free a used ticket when its refcount falls to zero.
3233 */
3234void
3235xfs_log_ticket_put(
3236 xlog_ticket_t *ticket)
3237{
3238 ASSERT(atomic_read(&ticket->t_ref) > 0);
3239 if (atomic_dec_and_test(&ticket->t_ref))
3240 kmem_zone_free(xfs_log_ticket_zone, ticket);
3241}
3242
3243xlog_ticket_t *
3244xfs_log_ticket_get(
3245 xlog_ticket_t *ticket)
3246{
3247 ASSERT(atomic_read(&ticket->t_ref) > 0);
3248 atomic_inc(&ticket->t_ref);
3249 return ticket;
3250}
3251
3252/*
3253 * Allocate and initialise a new log ticket.
3254 */
3255xlog_ticket_t *
3256xlog_ticket_alloc(
3257 struct log *log,
3258 int unit_bytes,
3259 int cnt,
3260 char client,
3261 uint xflags,
3262 int alloc_flags)
3263{
3264 struct xlog_ticket *tic;
3265 uint num_headers;
3266 int iclog_space;
3267
3268 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3269 if (!tic)
3270 return NULL;
3271
3272 /*
3273 * Permanent reservations have up to 'cnt'-1 active log operations
3274 * in the log. A unit in this case is the amount of space for one
3275 * of these log operations. Normal reservations have a cnt of 1
3276 * and their unit amount is the total amount of space required.
3277 *
3278 * The following lines of code account for non-transaction data
3279 * which occupy space in the on-disk log.
3280 *
3281 * Normal form of a transaction is:
3282 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3283 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3284 *
3285 * We need to account for all the leadup data and trailer data
3286 * around the transaction data.
3287 * And then we need to account for the worst case in terms of using
3288 * more space.
3289 * The worst case will happen if:
3290 * - the placement of the transaction happens to be such that the
3291 * roundoff is at its maximum
3292 * - the transaction data is synced before the commit record is synced
3293 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3294 * Therefore the commit record is in its own Log Record.
3295 * This can happen as the commit record is called with its
3296 * own region to xlog_write().
3297 * This then means that in the worst case, roundoff can happen for
3298 * the commit-rec as well.
3299 * The commit-rec is smaller than padding in this scenario and so it is
3300 * not added separately.
3301 */
3302
3303 /* for trans header */
3304 unit_bytes += sizeof(xlog_op_header_t);
3305 unit_bytes += sizeof(xfs_trans_header_t);
3306
3307 /* for start-rec */
3308 unit_bytes += sizeof(xlog_op_header_t);
3309
3310 /*
3311 * for LR headers - the space for data in an iclog is the size minus
3312 * the space used for the headers. If we use the iclog size, then we
3313 * undercalculate the number of headers required.
3314 *
3315 * Furthermore - the addition of op headers for split-recs might
3316 * increase the space required enough to require more log and op
3317 * headers, so take that into account too.
3318 *
3319 * IMPORTANT: This reservation makes the assumption that if this
3320 * transaction is the first in an iclog and hence has the LR headers
3321 * accounted to it, then the remaining space in the iclog is
3322 * exclusively for this transaction. i.e. if the transaction is larger
3323 * than the iclog, it will be the only thing in that iclog.
3324 * Fundamentally, this means we must pass the entire log vector to
3325 * xlog_write to guarantee this.
3326 */
3327 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3328 num_headers = howmany(unit_bytes, iclog_space);
3329
3330 /* for split-recs - ophdrs added when data split over LRs */
3331 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3332
3333 /* add extra header reservations if we overrun */
3334 while (!num_headers ||
3335 howmany(unit_bytes, iclog_space) > num_headers) {
3336 unit_bytes += sizeof(xlog_op_header_t);
3337 num_headers++;
3338 }
3339 unit_bytes += log->l_iclog_hsize * num_headers;
3340
3341 /* for commit-rec LR header - note: padding will subsume the ophdr */
3342 unit_bytes += log->l_iclog_hsize;
3343
3344 /* for roundoff padding for transaction data and one for commit record */
3345 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3346 log->l_mp->m_sb.sb_logsunit > 1) {
3347 /* log su roundoff */
3348 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3349 } else {
3350 /* BB roundoff */
3351 unit_bytes += 2*BBSIZE;
3352 }
3353
3354 atomic_set(&tic->t_ref, 1);
3355 INIT_LIST_HEAD(&tic->t_queue);
3356 tic->t_unit_res = unit_bytes;
3357 tic->t_curr_res = unit_bytes;
3358 tic->t_cnt = cnt;
3359 tic->t_ocnt = cnt;
3360 tic->t_tid = random32();
3361 tic->t_clientid = client;
3362 tic->t_flags = XLOG_TIC_INITED;
3363 tic->t_trans_type = 0;
3364 if (xflags & XFS_LOG_PERM_RESERV)
3365 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3366 init_waitqueue_head(&tic->t_wait);
3367
3368 xlog_tic_reset_res(tic);
3369
3370 return tic;
3371}
3372
3373
3374/******************************************************************************
3375 *
3376 * Log debug routines
3377 *
3378 ******************************************************************************
3379 */
3380#if defined(DEBUG)
3381/*
3382 * Make sure that the destination ptr is within the valid data region of
3383 * one of the iclogs. This uses backup pointers stored in a different
3384 * part of the log in case we trash the log structure.
3385 */
3386void
3387xlog_verify_dest_ptr(
3388 struct log *log,
3389 char *ptr)
3390{
3391 int i;
3392 int good_ptr = 0;
3393
3394 for (i = 0; i < log->l_iclog_bufs; i++) {
3395 if (ptr >= log->l_iclog_bak[i] &&
3396 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3397 good_ptr++;
3398 }
3399
3400 if (!good_ptr)
3401 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3402}
3403
3404/*
3405 * Check to make sure the grant write head didn't just over lap the tail. If
3406 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3407 * the cycles differ by exactly one and check the byte count.
3408 *
3409 * This check is run unlocked, so can give false positives. Rather than assert
3410 * on failures, use a warn-once flag and a panic tag to allow the admin to
3411 * determine if they want to panic the machine when such an error occurs. For
3412 * debug kernels this will have the same effect as using an assert but, unlinke
3413 * an assert, it can be turned off at runtime.
3414 */
3415STATIC void
3416xlog_verify_grant_tail(
3417 struct log *log)
3418{
3419 int tail_cycle, tail_blocks;
3420 int cycle, space;
3421
3422 xlog_crack_grant_head(&log->l_grant_write_head, &cycle, &space);
3423 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3424 if (tail_cycle != cycle) {
3425 if (cycle - 1 != tail_cycle &&
3426 !(log->l_flags & XLOG_TAIL_WARN)) {
3427 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3428 "%s: cycle - 1 != tail_cycle", __func__);
3429 log->l_flags |= XLOG_TAIL_WARN;
3430 }
3431
3432 if (space > BBTOB(tail_blocks) &&
3433 !(log->l_flags & XLOG_TAIL_WARN)) {
3434 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3435 "%s: space > BBTOB(tail_blocks)", __func__);
3436 log->l_flags |= XLOG_TAIL_WARN;
3437 }
3438 }
3439}
3440
3441/* check if it will fit */
3442STATIC void
3443xlog_verify_tail_lsn(xlog_t *log,
3444 xlog_in_core_t *iclog,
3445 xfs_lsn_t tail_lsn)
3446{
3447 int blocks;
3448
3449 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3450 blocks =
3451 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3452 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3453 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3454 } else {
3455 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3456
3457 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3458 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3459
3460 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3461 if (blocks < BTOBB(iclog->ic_offset) + 1)
3462 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3463 }
3464} /* xlog_verify_tail_lsn */
3465
3466/*
3467 * Perform a number of checks on the iclog before writing to disk.
3468 *
3469 * 1. Make sure the iclogs are still circular
3470 * 2. Make sure we have a good magic number
3471 * 3. Make sure we don't have magic numbers in the data
3472 * 4. Check fields of each log operation header for:
3473 * A. Valid client identifier
3474 * B. tid ptr value falls in valid ptr space (user space code)
3475 * C. Length in log record header is correct according to the
3476 * individual operation headers within record.
3477 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3478 * log, check the preceding blocks of the physical log to make sure all
3479 * the cycle numbers agree with the current cycle number.
3480 */
3481STATIC void
3482xlog_verify_iclog(xlog_t *log,
3483 xlog_in_core_t *iclog,
3484 int count,
3485 boolean_t syncing)
3486{
3487 xlog_op_header_t *ophead;
3488 xlog_in_core_t *icptr;
3489 xlog_in_core_2_t *xhdr;
3490 xfs_caddr_t ptr;
3491 xfs_caddr_t base_ptr;
3492 __psint_t field_offset;
3493 __uint8_t clientid;
3494 int len, i, j, k, op_len;
3495 int idx;
3496
3497 /* check validity of iclog pointers */
3498 spin_lock(&log->l_icloglock);
3499 icptr = log->l_iclog;
3500 for (i=0; i < log->l_iclog_bufs; i++) {
3501 if (icptr == NULL)
3502 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3503 icptr = icptr->ic_next;
3504 }
3505 if (icptr != log->l_iclog)
3506 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3507 spin_unlock(&log->l_icloglock);
3508
3509 /* check log magic numbers */
3510 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3511 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3512
3513 ptr = (xfs_caddr_t) &iclog->ic_header;
3514 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3515 ptr += BBSIZE) {
3516 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3517 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3518 __func__);
3519 }
3520
3521 /* check fields */
3522 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3523 ptr = iclog->ic_datap;
3524 base_ptr = ptr;
3525 ophead = (xlog_op_header_t *)ptr;
3526 xhdr = iclog->ic_data;
3527 for (i = 0; i < len; i++) {
3528 ophead = (xlog_op_header_t *)ptr;
3529
3530 /* clientid is only 1 byte */
3531 field_offset = (__psint_t)
3532 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3533 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3534 clientid = ophead->oh_clientid;
3535 } else {
3536 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3537 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3538 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3539 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3540 clientid = xlog_get_client_id(
3541 xhdr[j].hic_xheader.xh_cycle_data[k]);
3542 } else {
3543 clientid = xlog_get_client_id(
3544 iclog->ic_header.h_cycle_data[idx]);
3545 }
3546 }
3547 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3548 xfs_warn(log->l_mp,
3549 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3550 __func__, clientid, ophead,
3551 (unsigned long)field_offset);
3552
3553 /* check length */
3554 field_offset = (__psint_t)
3555 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3556 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3557 op_len = be32_to_cpu(ophead->oh_len);
3558 } else {
3559 idx = BTOBBT((__psint_t)&ophead->oh_len -
3560 (__psint_t)iclog->ic_datap);
3561 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3562 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3563 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3564 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3565 } else {
3566 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3567 }
3568 }
3569 ptr += sizeof(xlog_op_header_t) + op_len;
3570 }
3571} /* xlog_verify_iclog */
3572#endif
3573
3574/*
3575 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3576 */
3577STATIC int
3578xlog_state_ioerror(
3579 xlog_t *log)
3580{
3581 xlog_in_core_t *iclog, *ic;
3582
3583 iclog = log->l_iclog;
3584 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3585 /*
3586 * Mark all the incore logs IOERROR.
3587 * From now on, no log flushes will result.
3588 */
3589 ic = iclog;
3590 do {
3591 ic->ic_state = XLOG_STATE_IOERROR;
3592 ic = ic->ic_next;
3593 } while (ic != iclog);
3594 return 0;
3595 }
3596 /*
3597 * Return non-zero, if state transition has already happened.
3598 */
3599 return 1;
3600}
3601
3602/*
3603 * This is called from xfs_force_shutdown, when we're forcibly
3604 * shutting down the filesystem, typically because of an IO error.
3605 * Our main objectives here are to make sure that:
3606 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3607 * parties to find out, 'atomically'.
3608 * b. those who're sleeping on log reservations, pinned objects and
3609 * other resources get woken up, and be told the bad news.
3610 * c. nothing new gets queued up after (a) and (b) are done.
3611 * d. if !logerror, flush the iclogs to disk, then seal them off
3612 * for business.
3613 *
3614 * Note: for delayed logging the !logerror case needs to flush the regions
3615 * held in memory out to the iclogs before flushing them to disk. This needs
3616 * to be done before the log is marked as shutdown, otherwise the flush to the
3617 * iclogs will fail.
3618 */
3619int
3620xfs_log_force_umount(
3621 struct xfs_mount *mp,
3622 int logerror)
3623{
3624 xlog_ticket_t *tic;
3625 xlog_t *log;
3626 int retval;
3627
3628 log = mp->m_log;
3629
3630 /*
3631 * If this happens during log recovery, don't worry about
3632 * locking; the log isn't open for business yet.
3633 */
3634 if (!log ||
3635 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3636 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3637 if (mp->m_sb_bp)
3638 XFS_BUF_DONE(mp->m_sb_bp);
3639 return 0;
3640 }
3641
3642 /*
3643 * Somebody could've already done the hard work for us.
3644 * No need to get locks for this.
3645 */
3646 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3647 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3648 return 1;
3649 }
3650 retval = 0;
3651
3652 /*
3653 * Flush the in memory commit item list before marking the log as
3654 * being shut down. We need to do it in this order to ensure all the
3655 * completed transactions are flushed to disk with the xfs_log_force()
3656 * call below.
3657 */
3658 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG))
3659 xlog_cil_force(log);
3660
3661 /*
3662 * mark the filesystem and the as in a shutdown state and wake
3663 * everybody up to tell them the bad news.
3664 */
3665 spin_lock(&log->l_icloglock);
3666 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3667 if (mp->m_sb_bp)
3668 XFS_BUF_DONE(mp->m_sb_bp);
3669
3670 /*
3671 * This flag is sort of redundant because of the mount flag, but
3672 * it's good to maintain the separation between the log and the rest
3673 * of XFS.
3674 */
3675 log->l_flags |= XLOG_IO_ERROR;
3676
3677 /*
3678 * If we hit a log error, we want to mark all the iclogs IOERROR
3679 * while we're still holding the loglock.
3680 */
3681 if (logerror)
3682 retval = xlog_state_ioerror(log);
3683 spin_unlock(&log->l_icloglock);
3684
3685 /*
3686 * We don't want anybody waiting for log reservations after this. That
3687 * means we have to wake up everybody queued up on reserveq as well as
3688 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3689 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3690 * action is protected by the grant locks.
3691 */
3692 spin_lock(&log->l_grant_reserve_lock);
3693 list_for_each_entry(tic, &log->l_reserveq, t_queue)
3694 wake_up(&tic->t_wait);
3695 spin_unlock(&log->l_grant_reserve_lock);
3696
3697 spin_lock(&log->l_grant_write_lock);
3698 list_for_each_entry(tic, &log->l_writeq, t_queue)
3699 wake_up(&tic->t_wait);
3700 spin_unlock(&log->l_grant_write_lock);
3701
3702 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3703 ASSERT(!logerror);
3704 /*
3705 * Force the incore logs to disk before shutting the
3706 * log down completely.
3707 */
3708 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3709
3710 spin_lock(&log->l_icloglock);
3711 retval = xlog_state_ioerror(log);
3712 spin_unlock(&log->l_icloglock);
3713 }
3714 /*
3715 * Wake up everybody waiting on xfs_log_force.
3716 * Callback all log item committed functions as if the
3717 * log writes were completed.
3718 */
3719 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3720
3721#ifdef XFSERRORDEBUG
3722 {
3723 xlog_in_core_t *iclog;
3724
3725 spin_lock(&log->l_icloglock);
3726 iclog = log->l_iclog;
3727 do {
3728 ASSERT(iclog->ic_callback == 0);
3729 iclog = iclog->ic_next;
3730 } while (iclog != log->l_iclog);
3731 spin_unlock(&log->l_icloglock);
3732 }
3733#endif
3734 /* return non-zero if log IOERROR transition had already happened */
3735 return retval;
3736}
3737
3738STATIC int
3739xlog_iclogs_empty(xlog_t *log)
3740{
3741 xlog_in_core_t *iclog;
3742
3743 iclog = log->l_iclog;
3744 do {
3745 /* endianness does not matter here, zero is zero in
3746 * any language.
3747 */
3748 if (iclog->ic_header.h_num_logops)
3749 return 0;
3750 iclog = iclog->ic_next;
3751 } while (iclog != log->l_iclog);
3752 return 1;
3753}