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