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