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 int
34xlog_space_left(
35 struct xlog *log,
36 atomic64_t *head);
37STATIC void
38xlog_dealloc_log(
39 struct xlog *log);
40
41/* local state machine functions */
42STATIC void xlog_state_done_syncing(
43 struct xlog_in_core *iclog);
44STATIC void xlog_state_do_callback(
45 struct xlog *log);
46STATIC int
47xlog_state_get_iclog_space(
48 struct xlog *log,
49 int len,
50 struct xlog_in_core **iclog,
51 struct xlog_ticket *ticket,
52 int *logoffsetp);
53STATIC void
54xlog_grant_push_ail(
55 struct xlog *log,
56 int need_bytes);
57STATIC void
58xlog_sync(
59 struct xlog *log,
60 struct xlog_in_core *iclog,
61 struct xlog_ticket *ticket);
62#if defined(DEBUG)
63STATIC void
64xlog_verify_grant_tail(
65 struct xlog *log);
66STATIC void
67xlog_verify_iclog(
68 struct xlog *log,
69 struct xlog_in_core *iclog,
70 int count);
71STATIC void
72xlog_verify_tail_lsn(
73 struct xlog *log,
74 struct xlog_in_core *iclog);
75#else
76#define xlog_verify_grant_tail(a)
77#define xlog_verify_iclog(a,b,c)
78#define xlog_verify_tail_lsn(a,b)
79#endif
80
81STATIC int
82xlog_iclogs_empty(
83 struct xlog *log);
84
85static int
86xfs_log_cover(struct xfs_mount *);
87
88/*
89 * We need to make sure the buffer pointer returned is naturally aligned for the
90 * biggest basic data type we put into it. We have already accounted for this
91 * padding when sizing the buffer.
92 *
93 * However, this padding does not get written into the log, and hence we have to
94 * track the space used by the log vectors separately to prevent log space hangs
95 * due to inaccurate accounting (i.e. a leak) of the used log space through the
96 * CIL context ticket.
97 *
98 * We also add space for the xlog_op_header that describes this region in the
99 * log. This prepends the data region we return to the caller to copy their data
100 * into, so do all the static initialisation of the ophdr now. Because the ophdr
101 * is not 8 byte aligned, we have to be careful to ensure that we align the
102 * start of the buffer such that the region we return to the call is 8 byte
103 * aligned and packed against the tail of the ophdr.
104 */
105void *
106xlog_prepare_iovec(
107 struct xfs_log_vec *lv,
108 struct xfs_log_iovec **vecp,
109 uint type)
110{
111 struct xfs_log_iovec *vec = *vecp;
112 struct xlog_op_header *oph;
113 uint32_t len;
114 void *buf;
115
116 if (vec) {
117 ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
118 vec++;
119 } else {
120 vec = &lv->lv_iovecp[0];
121 }
122
123 len = lv->lv_buf_len + sizeof(struct xlog_op_header);
124 if (!IS_ALIGNED(len, sizeof(uint64_t))) {
125 lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
126 sizeof(struct xlog_op_header);
127 }
128
129 vec->i_type = type;
130 vec->i_addr = lv->lv_buf + lv->lv_buf_len;
131
132 oph = vec->i_addr;
133 oph->oh_clientid = XFS_TRANSACTION;
134 oph->oh_res2 = 0;
135 oph->oh_flags = 0;
136
137 buf = vec->i_addr + sizeof(struct xlog_op_header);
138 ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
139
140 *vecp = vec;
141 return buf;
142}
143
144static void
145xlog_grant_sub_space(
146 struct xlog *log,
147 atomic64_t *head,
148 int bytes)
149{
150 int64_t head_val = atomic64_read(head);
151 int64_t new, old;
152
153 do {
154 int cycle, space;
155
156 xlog_crack_grant_head_val(head_val, &cycle, &space);
157
158 space -= bytes;
159 if (space < 0) {
160 space += log->l_logsize;
161 cycle--;
162 }
163
164 old = head_val;
165 new = xlog_assign_grant_head_val(cycle, space);
166 head_val = atomic64_cmpxchg(head, old, new);
167 } while (head_val != old);
168}
169
170static void
171xlog_grant_add_space(
172 struct xlog *log,
173 atomic64_t *head,
174 int bytes)
175{
176 int64_t head_val = atomic64_read(head);
177 int64_t new, old;
178
179 do {
180 int tmp;
181 int cycle, space;
182
183 xlog_crack_grant_head_val(head_val, &cycle, &space);
184
185 tmp = log->l_logsize - space;
186 if (tmp > bytes)
187 space += bytes;
188 else {
189 space = bytes - tmp;
190 cycle++;
191 }
192
193 old = head_val;
194 new = xlog_assign_grant_head_val(cycle, space);
195 head_val = atomic64_cmpxchg(head, old, new);
196 } while (head_val != old);
197}
198
199STATIC void
200xlog_grant_head_init(
201 struct xlog_grant_head *head)
202{
203 xlog_assign_grant_head(&head->grant, 1, 0);
204 INIT_LIST_HEAD(&head->waiters);
205 spin_lock_init(&head->lock);
206}
207
208STATIC void
209xlog_grant_head_wake_all(
210 struct xlog_grant_head *head)
211{
212 struct xlog_ticket *tic;
213
214 spin_lock(&head->lock);
215 list_for_each_entry(tic, &head->waiters, t_queue)
216 wake_up_process(tic->t_task);
217 spin_unlock(&head->lock);
218}
219
220static inline int
221xlog_ticket_reservation(
222 struct xlog *log,
223 struct xlog_grant_head *head,
224 struct xlog_ticket *tic)
225{
226 if (head == &log->l_write_head) {
227 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
228 return tic->t_unit_res;
229 }
230
231 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
232 return tic->t_unit_res * tic->t_cnt;
233
234 return tic->t_unit_res;
235}
236
237STATIC bool
238xlog_grant_head_wake(
239 struct xlog *log,
240 struct xlog_grant_head *head,
241 int *free_bytes)
242{
243 struct xlog_ticket *tic;
244 int need_bytes;
245 bool woken_task = false;
246
247 list_for_each_entry(tic, &head->waiters, t_queue) {
248
249 /*
250 * There is a chance that the size of the CIL checkpoints in
251 * progress at the last AIL push target calculation resulted in
252 * limiting the target to the log head (l_last_sync_lsn) at the
253 * time. This may not reflect where the log head is now as the
254 * CIL checkpoints may have completed.
255 *
256 * Hence when we are woken here, it may be that the head of the
257 * log that has moved rather than the tail. As the tail didn't
258 * move, there still won't be space available for the
259 * reservation we require. However, if the AIL has already
260 * pushed to the target defined by the old log head location, we
261 * will hang here waiting for something else to update the AIL
262 * push target.
263 *
264 * Therefore, if there isn't space to wake the first waiter on
265 * the grant head, we need to push the AIL again to ensure the
266 * target reflects both the current log tail and log head
267 * position before we wait for the tail to move again.
268 */
269
270 need_bytes = xlog_ticket_reservation(log, head, tic);
271 if (*free_bytes < need_bytes) {
272 if (!woken_task)
273 xlog_grant_push_ail(log, need_bytes);
274 return false;
275 }
276
277 *free_bytes -= need_bytes;
278 trace_xfs_log_grant_wake_up(log, tic);
279 wake_up_process(tic->t_task);
280 woken_task = true;
281 }
282
283 return true;
284}
285
286STATIC int
287xlog_grant_head_wait(
288 struct xlog *log,
289 struct xlog_grant_head *head,
290 struct xlog_ticket *tic,
291 int need_bytes) __releases(&head->lock)
292 __acquires(&head->lock)
293{
294 list_add_tail(&tic->t_queue, &head->waiters);
295
296 do {
297 if (xlog_is_shutdown(log))
298 goto shutdown;
299 xlog_grant_push_ail(log, need_bytes);
300
301 __set_current_state(TASK_UNINTERRUPTIBLE);
302 spin_unlock(&head->lock);
303
304 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
305
306 trace_xfs_log_grant_sleep(log, tic);
307 schedule();
308 trace_xfs_log_grant_wake(log, tic);
309
310 spin_lock(&head->lock);
311 if (xlog_is_shutdown(log))
312 goto shutdown;
313 } while (xlog_space_left(log, &head->grant) < need_bytes);
314
315 list_del_init(&tic->t_queue);
316 return 0;
317shutdown:
318 list_del_init(&tic->t_queue);
319 return -EIO;
320}
321
322/*
323 * Atomically get the log space required for a log ticket.
324 *
325 * Once a ticket gets put onto head->waiters, it will only return after the
326 * needed reservation is satisfied.
327 *
328 * This function is structured so that it has a lock free fast path. This is
329 * necessary because every new transaction reservation will come through this
330 * path. Hence any lock will be globally hot if we take it unconditionally on
331 * every pass.
332 *
333 * As tickets are only ever moved on and off head->waiters under head->lock, we
334 * only need to take that lock if we are going to add the ticket to the queue
335 * and sleep. We can avoid taking the lock if the ticket was never added to
336 * head->waiters because the t_queue list head will be empty and we hold the
337 * only reference to it so it can safely be checked unlocked.
338 */
339STATIC int
340xlog_grant_head_check(
341 struct xlog *log,
342 struct xlog_grant_head *head,
343 struct xlog_ticket *tic,
344 int *need_bytes)
345{
346 int free_bytes;
347 int error = 0;
348
349 ASSERT(!xlog_in_recovery(log));
350
351 /*
352 * If there are other waiters on the queue then give them a chance at
353 * logspace before us. Wake up the first waiters, if we do not wake
354 * up all the waiters then go to sleep waiting for more free space,
355 * otherwise try to get some space for this transaction.
356 */
357 *need_bytes = xlog_ticket_reservation(log, head, tic);
358 free_bytes = xlog_space_left(log, &head->grant);
359 if (!list_empty_careful(&head->waiters)) {
360 spin_lock(&head->lock);
361 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
362 free_bytes < *need_bytes) {
363 error = xlog_grant_head_wait(log, head, tic,
364 *need_bytes);
365 }
366 spin_unlock(&head->lock);
367 } else if (free_bytes < *need_bytes) {
368 spin_lock(&head->lock);
369 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
370 spin_unlock(&head->lock);
371 }
372
373 return error;
374}
375
376bool
377xfs_log_writable(
378 struct xfs_mount *mp)
379{
380 /*
381 * Do not write to the log on norecovery mounts, if the data or log
382 * devices are read-only, or if the filesystem is shutdown. Read-only
383 * mounts allow internal writes for log recovery and unmount purposes,
384 * so don't restrict that case.
385 */
386 if (xfs_has_norecovery(mp))
387 return false;
388 if (xfs_readonly_buftarg(mp->m_ddev_targp))
389 return false;
390 if (xfs_readonly_buftarg(mp->m_log->l_targ))
391 return false;
392 if (xlog_is_shutdown(mp->m_log))
393 return false;
394 return true;
395}
396
397/*
398 * Replenish the byte reservation required by moving the grant write head.
399 */
400int
401xfs_log_regrant(
402 struct xfs_mount *mp,
403 struct xlog_ticket *tic)
404{
405 struct xlog *log = mp->m_log;
406 int need_bytes;
407 int error = 0;
408
409 if (xlog_is_shutdown(log))
410 return -EIO;
411
412 XFS_STATS_INC(mp, xs_try_logspace);
413
414 /*
415 * This is a new transaction on the ticket, so we need to change the
416 * transaction ID so that the next transaction has a different TID in
417 * the log. Just add one to the existing tid so that we can see chains
418 * of rolling transactions in the log easily.
419 */
420 tic->t_tid++;
421
422 xlog_grant_push_ail(log, tic->t_unit_res);
423
424 tic->t_curr_res = tic->t_unit_res;
425 if (tic->t_cnt > 0)
426 return 0;
427
428 trace_xfs_log_regrant(log, tic);
429
430 error = xlog_grant_head_check(log, &log->l_write_head, tic,
431 &need_bytes);
432 if (error)
433 goto out_error;
434
435 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
436 trace_xfs_log_regrant_exit(log, tic);
437 xlog_verify_grant_tail(log);
438 return 0;
439
440out_error:
441 /*
442 * If we are failing, make sure the ticket doesn't have any current
443 * reservations. We don't want to add this back when the ticket/
444 * transaction gets cancelled.
445 */
446 tic->t_curr_res = 0;
447 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
448 return error;
449}
450
451/*
452 * Reserve log space and return a ticket corresponding to the reservation.
453 *
454 * Each reservation is going to reserve extra space for a log record header.
455 * When writes happen to the on-disk log, we don't subtract the length of the
456 * log record header from any reservation. By wasting space in each
457 * reservation, we prevent over allocation problems.
458 */
459int
460xfs_log_reserve(
461 struct xfs_mount *mp,
462 int unit_bytes,
463 int cnt,
464 struct xlog_ticket **ticp,
465 bool permanent)
466{
467 struct xlog *log = mp->m_log;
468 struct xlog_ticket *tic;
469 int need_bytes;
470 int error = 0;
471
472 if (xlog_is_shutdown(log))
473 return -EIO;
474
475 XFS_STATS_INC(mp, xs_try_logspace);
476
477 ASSERT(*ticp == NULL);
478 tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
479 *ticp = tic;
480
481 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
482 : tic->t_unit_res);
483
484 trace_xfs_log_reserve(log, tic);
485
486 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
487 &need_bytes);
488 if (error)
489 goto out_error;
490
491 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
492 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
493 trace_xfs_log_reserve_exit(log, tic);
494 xlog_verify_grant_tail(log);
495 return 0;
496
497out_error:
498 /*
499 * If we are failing, make sure the ticket doesn't have any current
500 * reservations. We don't want to add this back when the ticket/
501 * transaction gets cancelled.
502 */
503 tic->t_curr_res = 0;
504 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
505 return error;
506}
507
508/*
509 * Run all the pending iclog callbacks and wake log force waiters and iclog
510 * space waiters so they can process the newly set shutdown state. We really
511 * don't care what order we process callbacks here because the log is shut down
512 * and so state cannot change on disk anymore. However, we cannot wake waiters
513 * until the callbacks have been processed because we may be in unmount and
514 * we must ensure that all AIL operations the callbacks perform have completed
515 * before we tear down the AIL.
516 *
517 * We avoid processing actively referenced iclogs so that we don't run callbacks
518 * while the iclog owner might still be preparing the iclog for IO submssion.
519 * These will be caught by xlog_state_iclog_release() and call this function
520 * again to process any callbacks that may have been added to that iclog.
521 */
522static void
523xlog_state_shutdown_callbacks(
524 struct xlog *log)
525{
526 struct xlog_in_core *iclog;
527 LIST_HEAD(cb_list);
528
529 iclog = log->l_iclog;
530 do {
531 if (atomic_read(&iclog->ic_refcnt)) {
532 /* Reference holder will re-run iclog callbacks. */
533 continue;
534 }
535 list_splice_init(&iclog->ic_callbacks, &cb_list);
536 spin_unlock(&log->l_icloglock);
537
538 xlog_cil_process_committed(&cb_list);
539
540 spin_lock(&log->l_icloglock);
541 wake_up_all(&iclog->ic_write_wait);
542 wake_up_all(&iclog->ic_force_wait);
543 } while ((iclog = iclog->ic_next) != log->l_iclog);
544
545 wake_up_all(&log->l_flush_wait);
546}
547
548/*
549 * Flush iclog to disk if this is the last reference to the given iclog and the
550 * it is in the WANT_SYNC state.
551 *
552 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
553 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
554 * written to stable storage, and implies that a commit record is contained
555 * within the iclog. We need to ensure that the log tail does not move beyond
556 * the tail that the first commit record in the iclog ordered against, otherwise
557 * correct recovery of that checkpoint becomes dependent on future operations
558 * performed on this iclog.
559 *
560 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
561 * current tail into iclog. Once the iclog tail is set, future operations must
562 * not modify it, otherwise they potentially violate ordering constraints for
563 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
564 * the iclog will get zeroed on activation of the iclog after sync, so we
565 * always capture the tail lsn on the iclog on the first NEED_FUA release
566 * regardless of the number of active reference counts on this iclog.
567 */
568int
569xlog_state_release_iclog(
570 struct xlog *log,
571 struct xlog_in_core *iclog,
572 struct xlog_ticket *ticket)
573{
574 xfs_lsn_t tail_lsn;
575 bool last_ref;
576
577 lockdep_assert_held(&log->l_icloglock);
578
579 trace_xlog_iclog_release(iclog, _RET_IP_);
580 /*
581 * Grabbing the current log tail needs to be atomic w.r.t. the writing
582 * of the tail LSN into the iclog so we guarantee that the log tail does
583 * not move between the first time we know that the iclog needs to be
584 * made stable and when we eventually submit it.
585 */
586 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
587 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
588 !iclog->ic_header.h_tail_lsn) {
589 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
590 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
591 }
592
593 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
594
595 if (xlog_is_shutdown(log)) {
596 /*
597 * If there are no more references to this iclog, process the
598 * pending iclog callbacks that were waiting on the release of
599 * this iclog.
600 */
601 if (last_ref)
602 xlog_state_shutdown_callbacks(log);
603 return -EIO;
604 }
605
606 if (!last_ref)
607 return 0;
608
609 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
610 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
611 return 0;
612 }
613
614 iclog->ic_state = XLOG_STATE_SYNCING;
615 xlog_verify_tail_lsn(log, iclog);
616 trace_xlog_iclog_syncing(iclog, _RET_IP_);
617
618 spin_unlock(&log->l_icloglock);
619 xlog_sync(log, iclog, ticket);
620 spin_lock(&log->l_icloglock);
621 return 0;
622}
623
624/*
625 * Mount a log filesystem
626 *
627 * mp - ubiquitous xfs mount point structure
628 * log_target - buftarg of on-disk log device
629 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
630 * num_bblocks - Number of BBSIZE blocks in on-disk log
631 *
632 * Return error or zero.
633 */
634int
635xfs_log_mount(
636 xfs_mount_t *mp,
637 xfs_buftarg_t *log_target,
638 xfs_daddr_t blk_offset,
639 int num_bblks)
640{
641 struct xlog *log;
642 bool fatal = xfs_has_crc(mp);
643 int error = 0;
644 int min_logfsbs;
645
646 if (!xfs_has_norecovery(mp)) {
647 xfs_notice(mp, "Mounting V%d Filesystem %pU",
648 XFS_SB_VERSION_NUM(&mp->m_sb),
649 &mp->m_sb.sb_uuid);
650 } else {
651 xfs_notice(mp,
652"Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
653 XFS_SB_VERSION_NUM(&mp->m_sb),
654 &mp->m_sb.sb_uuid);
655 ASSERT(xfs_is_readonly(mp));
656 }
657
658 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
659 if (IS_ERR(log)) {
660 error = PTR_ERR(log);
661 goto out;
662 }
663 mp->m_log = log;
664
665 /*
666 * Validate the given log space and drop a critical message via syslog
667 * if the log size is too small that would lead to some unexpected
668 * situations in transaction log space reservation stage.
669 *
670 * Note: we can't just reject the mount if the validation fails. This
671 * would mean that people would have to downgrade their kernel just to
672 * remedy the situation as there is no way to grow the log (short of
673 * black magic surgery with xfs_db).
674 *
675 * We can, however, reject mounts for CRC format filesystems, as the
676 * mkfs binary being used to make the filesystem should never create a
677 * filesystem with a log that is too small.
678 */
679 min_logfsbs = xfs_log_calc_minimum_size(mp);
680
681 if (mp->m_sb.sb_logblocks < min_logfsbs) {
682 xfs_warn(mp,
683 "Log size %d blocks too small, minimum size is %d blocks",
684 mp->m_sb.sb_logblocks, min_logfsbs);
685 error = -EINVAL;
686 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
687 xfs_warn(mp,
688 "Log size %d blocks too large, maximum size is %lld blocks",
689 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
690 error = -EINVAL;
691 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
692 xfs_warn(mp,
693 "log size %lld bytes too large, maximum size is %lld bytes",
694 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
695 XFS_MAX_LOG_BYTES);
696 error = -EINVAL;
697 } else if (mp->m_sb.sb_logsunit > 1 &&
698 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
699 xfs_warn(mp,
700 "log stripe unit %u bytes must be a multiple of block size",
701 mp->m_sb.sb_logsunit);
702 error = -EINVAL;
703 fatal = true;
704 }
705 if (error) {
706 /*
707 * Log check errors are always fatal on v5; or whenever bad
708 * metadata leads to a crash.
709 */
710 if (fatal) {
711 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
712 ASSERT(0);
713 goto out_free_log;
714 }
715 xfs_crit(mp, "Log size out of supported range.");
716 xfs_crit(mp,
717"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
718 }
719
720 /*
721 * Initialize the AIL now we have a log.
722 */
723 error = xfs_trans_ail_init(mp);
724 if (error) {
725 xfs_warn(mp, "AIL initialisation failed: error %d", error);
726 goto out_free_log;
727 }
728 log->l_ailp = mp->m_ail;
729
730 /*
731 * skip log recovery on a norecovery mount. pretend it all
732 * just worked.
733 */
734 if (!xfs_has_norecovery(mp)) {
735 /*
736 * log recovery ignores readonly state and so we need to clear
737 * mount-based read only state so it can write to disk.
738 */
739 bool readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
740 &mp->m_opstate);
741 error = xlog_recover(log);
742 if (readonly)
743 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
744 if (error) {
745 xfs_warn(mp, "log mount/recovery failed: error %d",
746 error);
747 xlog_recover_cancel(log);
748 goto out_destroy_ail;
749 }
750 }
751
752 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
753 "log");
754 if (error)
755 goto out_destroy_ail;
756
757 /* Normal transactions can now occur */
758 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
759
760 /*
761 * Now the log has been fully initialised and we know were our
762 * space grant counters are, we can initialise the permanent ticket
763 * needed for delayed logging to work.
764 */
765 xlog_cil_init_post_recovery(log);
766
767 return 0;
768
769out_destroy_ail:
770 xfs_trans_ail_destroy(mp);
771out_free_log:
772 xlog_dealloc_log(log);
773out:
774 return error;
775}
776
777/*
778 * Finish the recovery of the file system. This is separate from the
779 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
780 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
781 * here.
782 *
783 * If we finish recovery successfully, start the background log work. If we are
784 * not doing recovery, then we have a RO filesystem and we don't need to start
785 * it.
786 */
787int
788xfs_log_mount_finish(
789 struct xfs_mount *mp)
790{
791 struct xlog *log = mp->m_log;
792 bool readonly;
793 int error = 0;
794
795 if (xfs_has_norecovery(mp)) {
796 ASSERT(xfs_is_readonly(mp));
797 return 0;
798 }
799
800 /*
801 * log recovery ignores readonly state and so we need to clear
802 * mount-based read only state so it can write to disk.
803 */
804 readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
805
806 /*
807 * During the second phase of log recovery, we need iget and
808 * iput to behave like they do for an active filesystem.
809 * xfs_fs_drop_inode needs to be able to prevent the deletion
810 * of inodes before we're done replaying log items on those
811 * inodes. Turn it off immediately after recovery finishes
812 * so that we don't leak the quota inodes if subsequent mount
813 * activities fail.
814 *
815 * We let all inodes involved in redo item processing end up on
816 * the LRU instead of being evicted immediately so that if we do
817 * something to an unlinked inode, the irele won't cause
818 * premature truncation and freeing of the inode, which results
819 * in log recovery failure. We have to evict the unreferenced
820 * lru inodes after clearing SB_ACTIVE because we don't
821 * otherwise clean up the lru if there's a subsequent failure in
822 * xfs_mountfs, which leads to us leaking the inodes if nothing
823 * else (e.g. quotacheck) references the inodes before the
824 * mount failure occurs.
825 */
826 mp->m_super->s_flags |= SB_ACTIVE;
827 xfs_log_work_queue(mp);
828 if (xlog_recovery_needed(log))
829 error = xlog_recover_finish(log);
830 mp->m_super->s_flags &= ~SB_ACTIVE;
831 evict_inodes(mp->m_super);
832
833 /*
834 * Drain the buffer LRU after log recovery. This is required for v4
835 * filesystems to avoid leaving around buffers with NULL verifier ops,
836 * but we do it unconditionally to make sure we're always in a clean
837 * cache state after mount.
838 *
839 * Don't push in the error case because the AIL may have pending intents
840 * that aren't removed until recovery is cancelled.
841 */
842 if (xlog_recovery_needed(log)) {
843 if (!error) {
844 xfs_log_force(mp, XFS_LOG_SYNC);
845 xfs_ail_push_all_sync(mp->m_ail);
846 }
847 xfs_notice(mp, "Ending recovery (logdev: %s)",
848 mp->m_logname ? mp->m_logname : "internal");
849 } else {
850 xfs_info(mp, "Ending clean mount");
851 }
852 xfs_buftarg_drain(mp->m_ddev_targp);
853
854 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
855 if (readonly)
856 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
857
858 /* Make sure the log is dead if we're returning failure. */
859 ASSERT(!error || xlog_is_shutdown(log));
860
861 return error;
862}
863
864/*
865 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
866 * the log.
867 */
868void
869xfs_log_mount_cancel(
870 struct xfs_mount *mp)
871{
872 xlog_recover_cancel(mp->m_log);
873 xfs_log_unmount(mp);
874}
875
876/*
877 * Flush out the iclog to disk ensuring that device caches are flushed and
878 * the iclog hits stable storage before any completion waiters are woken.
879 */
880static inline int
881xlog_force_iclog(
882 struct xlog_in_core *iclog)
883{
884 atomic_inc(&iclog->ic_refcnt);
885 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
886 if (iclog->ic_state == XLOG_STATE_ACTIVE)
887 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
888 return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
889}
890
891/*
892 * Cycle all the iclogbuf locks to make sure all log IO completion
893 * is done before we tear down these buffers.
894 */
895static void
896xlog_wait_iclog_completion(struct xlog *log)
897{
898 int i;
899 struct xlog_in_core *iclog = log->l_iclog;
900
901 for (i = 0; i < log->l_iclog_bufs; i++) {
902 down(&iclog->ic_sema);
903 up(&iclog->ic_sema);
904 iclog = iclog->ic_next;
905 }
906}
907
908/*
909 * Wait for the iclog and all prior iclogs to be written disk as required by the
910 * log force state machine. Waiting on ic_force_wait ensures iclog completions
911 * have been ordered and callbacks run before we are woken here, hence
912 * guaranteeing that all the iclogs up to this one are on stable storage.
913 */
914int
915xlog_wait_on_iclog(
916 struct xlog_in_core *iclog)
917 __releases(iclog->ic_log->l_icloglock)
918{
919 struct xlog *log = iclog->ic_log;
920
921 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
922 if (!xlog_is_shutdown(log) &&
923 iclog->ic_state != XLOG_STATE_ACTIVE &&
924 iclog->ic_state != XLOG_STATE_DIRTY) {
925 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
926 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
927 } else {
928 spin_unlock(&log->l_icloglock);
929 }
930
931 if (xlog_is_shutdown(log))
932 return -EIO;
933 return 0;
934}
935
936/*
937 * Write out an unmount record using the ticket provided. We have to account for
938 * the data space used in the unmount ticket as this write is not done from a
939 * transaction context that has already done the accounting for us.
940 */
941static int
942xlog_write_unmount_record(
943 struct xlog *log,
944 struct xlog_ticket *ticket)
945{
946 struct {
947 struct xlog_op_header ophdr;
948 struct xfs_unmount_log_format ulf;
949 } unmount_rec = {
950 .ophdr = {
951 .oh_clientid = XFS_LOG,
952 .oh_tid = cpu_to_be32(ticket->t_tid),
953 .oh_flags = XLOG_UNMOUNT_TRANS,
954 },
955 .ulf = {
956 .magic = XLOG_UNMOUNT_TYPE,
957 },
958 };
959 struct xfs_log_iovec reg = {
960 .i_addr = &unmount_rec,
961 .i_len = sizeof(unmount_rec),
962 .i_type = XLOG_REG_TYPE_UNMOUNT,
963 };
964 struct xfs_log_vec vec = {
965 .lv_niovecs = 1,
966 .lv_iovecp = ®,
967 };
968 LIST_HEAD(lv_chain);
969 list_add(&vec.lv_list, &lv_chain);
970
971 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
972 sizeof(struct xfs_unmount_log_format)) !=
973 sizeof(unmount_rec));
974
975 /* account for space used by record data */
976 ticket->t_curr_res -= sizeof(unmount_rec);
977
978 return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
979}
980
981/*
982 * Mark the filesystem clean by writing an unmount record to the head of the
983 * log.
984 */
985static void
986xlog_unmount_write(
987 struct xlog *log)
988{
989 struct xfs_mount *mp = log->l_mp;
990 struct xlog_in_core *iclog;
991 struct xlog_ticket *tic = NULL;
992 int error;
993
994 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
995 if (error)
996 goto out_err;
997
998 error = xlog_write_unmount_record(log, tic);
999 /*
1000 * At this point, we're umounting anyway, so there's no point in
1001 * transitioning log state to shutdown. Just continue...
1002 */
1003out_err:
1004 if (error)
1005 xfs_alert(mp, "%s: unmount record failed", __func__);
1006
1007 spin_lock(&log->l_icloglock);
1008 iclog = log->l_iclog;
1009 error = xlog_force_iclog(iclog);
1010 xlog_wait_on_iclog(iclog);
1011
1012 if (tic) {
1013 trace_xfs_log_umount_write(log, tic);
1014 xfs_log_ticket_ungrant(log, tic);
1015 }
1016}
1017
1018static void
1019xfs_log_unmount_verify_iclog(
1020 struct xlog *log)
1021{
1022 struct xlog_in_core *iclog = log->l_iclog;
1023
1024 do {
1025 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
1026 ASSERT(iclog->ic_offset == 0);
1027 } while ((iclog = iclog->ic_next) != log->l_iclog);
1028}
1029
1030/*
1031 * Unmount record used to have a string "Unmount filesystem--" in the
1032 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
1033 * We just write the magic number now since that particular field isn't
1034 * currently architecture converted and "Unmount" is a bit foo.
1035 * As far as I know, there weren't any dependencies on the old behaviour.
1036 */
1037static void
1038xfs_log_unmount_write(
1039 struct xfs_mount *mp)
1040{
1041 struct xlog *log = mp->m_log;
1042
1043 if (!xfs_log_writable(mp))
1044 return;
1045
1046 xfs_log_force(mp, XFS_LOG_SYNC);
1047
1048 if (xlog_is_shutdown(log))
1049 return;
1050
1051 /*
1052 * If we think the summary counters are bad, avoid writing the unmount
1053 * record to force log recovery at next mount, after which the summary
1054 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1055 * more details.
1056 */
1057 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1058 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1059 xfs_alert(mp, "%s: will fix summary counters at next mount",
1060 __func__);
1061 return;
1062 }
1063
1064 xfs_log_unmount_verify_iclog(log);
1065 xlog_unmount_write(log);
1066}
1067
1068/*
1069 * Empty the log for unmount/freeze.
1070 *
1071 * To do this, we first need to shut down the background log work so it is not
1072 * trying to cover the log as we clean up. We then need to unpin all objects in
1073 * the log so we can then flush them out. Once they have completed their IO and
1074 * run the callbacks removing themselves from the AIL, we can cover the log.
1075 */
1076int
1077xfs_log_quiesce(
1078 struct xfs_mount *mp)
1079{
1080 /*
1081 * Clear log incompat features since we're quiescing the log. Report
1082 * failures, though it's not fatal to have a higher log feature
1083 * protection level than the log contents actually require.
1084 */
1085 if (xfs_clear_incompat_log_features(mp)) {
1086 int error;
1087
1088 error = xfs_sync_sb(mp, false);
1089 if (error)
1090 xfs_warn(mp,
1091 "Failed to clear log incompat features on quiesce");
1092 }
1093
1094 cancel_delayed_work_sync(&mp->m_log->l_work);
1095 xfs_log_force(mp, XFS_LOG_SYNC);
1096
1097 /*
1098 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1099 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1100 * xfs_buf_iowait() cannot be used because it was pushed with the
1101 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1102 * the IO to complete.
1103 */
1104 xfs_ail_push_all_sync(mp->m_ail);
1105 xfs_buftarg_wait(mp->m_ddev_targp);
1106 xfs_buf_lock(mp->m_sb_bp);
1107 xfs_buf_unlock(mp->m_sb_bp);
1108
1109 return xfs_log_cover(mp);
1110}
1111
1112void
1113xfs_log_clean(
1114 struct xfs_mount *mp)
1115{
1116 xfs_log_quiesce(mp);
1117 xfs_log_unmount_write(mp);
1118}
1119
1120/*
1121 * Shut down and release the AIL and Log.
1122 *
1123 * During unmount, we need to ensure we flush all the dirty metadata objects
1124 * from the AIL so that the log is empty before we write the unmount record to
1125 * the log. Once this is done, we can tear down the AIL and the log.
1126 */
1127void
1128xfs_log_unmount(
1129 struct xfs_mount *mp)
1130{
1131 xfs_log_clean(mp);
1132
1133 /*
1134 * If shutdown has come from iclog IO context, the log
1135 * cleaning will have been skipped and so we need to wait
1136 * for the iclog to complete shutdown processing before we
1137 * tear anything down.
1138 */
1139 xlog_wait_iclog_completion(mp->m_log);
1140
1141 xfs_buftarg_drain(mp->m_ddev_targp);
1142
1143 xfs_trans_ail_destroy(mp);
1144
1145 xfs_sysfs_del(&mp->m_log->l_kobj);
1146
1147 xlog_dealloc_log(mp->m_log);
1148}
1149
1150void
1151xfs_log_item_init(
1152 struct xfs_mount *mp,
1153 struct xfs_log_item *item,
1154 int type,
1155 const struct xfs_item_ops *ops)
1156{
1157 item->li_log = mp->m_log;
1158 item->li_ailp = mp->m_ail;
1159 item->li_type = type;
1160 item->li_ops = ops;
1161 item->li_lv = NULL;
1162
1163 INIT_LIST_HEAD(&item->li_ail);
1164 INIT_LIST_HEAD(&item->li_cil);
1165 INIT_LIST_HEAD(&item->li_bio_list);
1166 INIT_LIST_HEAD(&item->li_trans);
1167}
1168
1169/*
1170 * Wake up processes waiting for log space after we have moved the log tail.
1171 */
1172void
1173xfs_log_space_wake(
1174 struct xfs_mount *mp)
1175{
1176 struct xlog *log = mp->m_log;
1177 int free_bytes;
1178
1179 if (xlog_is_shutdown(log))
1180 return;
1181
1182 if (!list_empty_careful(&log->l_write_head.waiters)) {
1183 ASSERT(!xlog_in_recovery(log));
1184
1185 spin_lock(&log->l_write_head.lock);
1186 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1187 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1188 spin_unlock(&log->l_write_head.lock);
1189 }
1190
1191 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1192 ASSERT(!xlog_in_recovery(log));
1193
1194 spin_lock(&log->l_reserve_head.lock);
1195 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1196 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1197 spin_unlock(&log->l_reserve_head.lock);
1198 }
1199}
1200
1201/*
1202 * Determine if we have a transaction that has gone to disk that needs to be
1203 * covered. To begin the transition to the idle state firstly the log needs to
1204 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1205 * we start attempting to cover the log.
1206 *
1207 * Only if we are then in a state where covering is needed, the caller is
1208 * informed that dummy transactions are required to move the log into the idle
1209 * state.
1210 *
1211 * If there are any items in the AIl or CIL, then we do not want to attempt to
1212 * cover the log as we may be in a situation where there isn't log space
1213 * available to run a dummy transaction and this can lead to deadlocks when the
1214 * tail of the log is pinned by an item that is modified in the CIL. Hence
1215 * there's no point in running a dummy transaction at this point because we
1216 * can't start trying to idle the log until both the CIL and AIL are empty.
1217 */
1218static bool
1219xfs_log_need_covered(
1220 struct xfs_mount *mp)
1221{
1222 struct xlog *log = mp->m_log;
1223 bool needed = false;
1224
1225 if (!xlog_cil_empty(log))
1226 return false;
1227
1228 spin_lock(&log->l_icloglock);
1229 switch (log->l_covered_state) {
1230 case XLOG_STATE_COVER_DONE:
1231 case XLOG_STATE_COVER_DONE2:
1232 case XLOG_STATE_COVER_IDLE:
1233 break;
1234 case XLOG_STATE_COVER_NEED:
1235 case XLOG_STATE_COVER_NEED2:
1236 if (xfs_ail_min_lsn(log->l_ailp))
1237 break;
1238 if (!xlog_iclogs_empty(log))
1239 break;
1240
1241 needed = true;
1242 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1243 log->l_covered_state = XLOG_STATE_COVER_DONE;
1244 else
1245 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1246 break;
1247 default:
1248 needed = true;
1249 break;
1250 }
1251 spin_unlock(&log->l_icloglock);
1252 return needed;
1253}
1254
1255/*
1256 * Explicitly cover the log. This is similar to background log covering but
1257 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1258 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1259 * must all be empty.
1260 */
1261static int
1262xfs_log_cover(
1263 struct xfs_mount *mp)
1264{
1265 int error = 0;
1266 bool need_covered;
1267
1268 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1269 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1270 xlog_is_shutdown(mp->m_log));
1271
1272 if (!xfs_log_writable(mp))
1273 return 0;
1274
1275 /*
1276 * xfs_log_need_covered() is not idempotent because it progresses the
1277 * state machine if the log requires covering. Therefore, we must call
1278 * this function once and use the result until we've issued an sb sync.
1279 * Do so first to make that abundantly clear.
1280 *
1281 * Fall into the covering sequence if the log needs covering or the
1282 * mount has lazy superblock accounting to sync to disk. The sb sync
1283 * used for covering accumulates the in-core counters, so covering
1284 * handles this for us.
1285 */
1286 need_covered = xfs_log_need_covered(mp);
1287 if (!need_covered && !xfs_has_lazysbcount(mp))
1288 return 0;
1289
1290 /*
1291 * To cover the log, commit the superblock twice (at most) in
1292 * independent checkpoints. The first serves as a reference for the
1293 * tail pointer. The sync transaction and AIL push empties the AIL and
1294 * updates the in-core tail to the LSN of the first checkpoint. The
1295 * second commit updates the on-disk tail with the in-core LSN,
1296 * covering the log. Push the AIL one more time to leave it empty, as
1297 * we found it.
1298 */
1299 do {
1300 error = xfs_sync_sb(mp, true);
1301 if (error)
1302 break;
1303 xfs_ail_push_all_sync(mp->m_ail);
1304 } while (xfs_log_need_covered(mp));
1305
1306 return error;
1307}
1308
1309/*
1310 * We may be holding the log iclog lock upon entering this routine.
1311 */
1312xfs_lsn_t
1313xlog_assign_tail_lsn_locked(
1314 struct xfs_mount *mp)
1315{
1316 struct xlog *log = mp->m_log;
1317 struct xfs_log_item *lip;
1318 xfs_lsn_t tail_lsn;
1319
1320 assert_spin_locked(&mp->m_ail->ail_lock);
1321
1322 /*
1323 * To make sure we always have a valid LSN for the log tail we keep
1324 * track of the last LSN which was committed in log->l_last_sync_lsn,
1325 * and use that when the AIL was empty.
1326 */
1327 lip = xfs_ail_min(mp->m_ail);
1328 if (lip)
1329 tail_lsn = lip->li_lsn;
1330 else
1331 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1332 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1333 atomic64_set(&log->l_tail_lsn, tail_lsn);
1334 return tail_lsn;
1335}
1336
1337xfs_lsn_t
1338xlog_assign_tail_lsn(
1339 struct xfs_mount *mp)
1340{
1341 xfs_lsn_t tail_lsn;
1342
1343 spin_lock(&mp->m_ail->ail_lock);
1344 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1345 spin_unlock(&mp->m_ail->ail_lock);
1346
1347 return tail_lsn;
1348}
1349
1350/*
1351 * Return the space in the log between the tail and the head. The head
1352 * is passed in the cycle/bytes formal parms. In the special case where
1353 * the reserve head has wrapped passed the tail, this calculation is no
1354 * longer valid. In this case, just return 0 which means there is no space
1355 * in the log. This works for all places where this function is called
1356 * with the reserve head. Of course, if the write head were to ever
1357 * wrap the tail, we should blow up. Rather than catch this case here,
1358 * we depend on other ASSERTions in other parts of the code. XXXmiken
1359 *
1360 * If reservation head is behind the tail, we have a problem. Warn about it,
1361 * but then treat it as if the log is empty.
1362 *
1363 * If the log is shut down, the head and tail may be invalid or out of whack, so
1364 * shortcut invalidity asserts in this case so that we don't trigger them
1365 * falsely.
1366 */
1367STATIC int
1368xlog_space_left(
1369 struct xlog *log,
1370 atomic64_t *head)
1371{
1372 int tail_bytes;
1373 int tail_cycle;
1374 int head_cycle;
1375 int head_bytes;
1376
1377 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1378 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1379 tail_bytes = BBTOB(tail_bytes);
1380 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1381 return log->l_logsize - (head_bytes - tail_bytes);
1382 if (tail_cycle + 1 < head_cycle)
1383 return 0;
1384
1385 /* Ignore potential inconsistency when shutdown. */
1386 if (xlog_is_shutdown(log))
1387 return log->l_logsize;
1388
1389 if (tail_cycle < head_cycle) {
1390 ASSERT(tail_cycle == (head_cycle - 1));
1391 return tail_bytes - head_bytes;
1392 }
1393
1394 /*
1395 * The reservation head is behind the tail. In this case we just want to
1396 * return the size of the log as the amount of space left.
1397 */
1398 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1399 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1400 tail_cycle, tail_bytes);
1401 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1402 head_cycle, head_bytes);
1403 ASSERT(0);
1404 return log->l_logsize;
1405}
1406
1407
1408static void
1409xlog_ioend_work(
1410 struct work_struct *work)
1411{
1412 struct xlog_in_core *iclog =
1413 container_of(work, struct xlog_in_core, ic_end_io_work);
1414 struct xlog *log = iclog->ic_log;
1415 int error;
1416
1417 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1418#ifdef DEBUG
1419 /* treat writes with injected CRC errors as failed */
1420 if (iclog->ic_fail_crc)
1421 error = -EIO;
1422#endif
1423
1424 /*
1425 * Race to shutdown the filesystem if we see an error.
1426 */
1427 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1428 xfs_alert(log->l_mp, "log I/O error %d", error);
1429 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1430 }
1431
1432 xlog_state_done_syncing(iclog);
1433 bio_uninit(&iclog->ic_bio);
1434
1435 /*
1436 * Drop the lock to signal that we are done. Nothing references the
1437 * iclog after this, so an unmount waiting on this lock can now tear it
1438 * down safely. As such, it is unsafe to reference the iclog after the
1439 * unlock as we could race with it being freed.
1440 */
1441 up(&iclog->ic_sema);
1442}
1443
1444/*
1445 * Return size of each in-core log record buffer.
1446 *
1447 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1448 *
1449 * If the filesystem blocksize is too large, we may need to choose a
1450 * larger size since the directory code currently logs entire blocks.
1451 */
1452STATIC void
1453xlog_get_iclog_buffer_size(
1454 struct xfs_mount *mp,
1455 struct xlog *log)
1456{
1457 if (mp->m_logbufs <= 0)
1458 mp->m_logbufs = XLOG_MAX_ICLOGS;
1459 if (mp->m_logbsize <= 0)
1460 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1461
1462 log->l_iclog_bufs = mp->m_logbufs;
1463 log->l_iclog_size = mp->m_logbsize;
1464
1465 /*
1466 * # headers = size / 32k - one header holds cycles from 32k of data.
1467 */
1468 log->l_iclog_heads =
1469 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1470 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1471}
1472
1473void
1474xfs_log_work_queue(
1475 struct xfs_mount *mp)
1476{
1477 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1478 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1479}
1480
1481/*
1482 * Clear the log incompat flags if we have the opportunity.
1483 *
1484 * This only happens if we're about to log the second dummy transaction as part
1485 * of covering the log and we can get the log incompat feature usage lock.
1486 */
1487static inline void
1488xlog_clear_incompat(
1489 struct xlog *log)
1490{
1491 struct xfs_mount *mp = log->l_mp;
1492
1493 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1494 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1495 return;
1496
1497 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1498 return;
1499
1500 if (!down_write_trylock(&log->l_incompat_users))
1501 return;
1502
1503 xfs_clear_incompat_log_features(mp);
1504 up_write(&log->l_incompat_users);
1505}
1506
1507/*
1508 * Every sync period we need to unpin all items in the AIL and push them to
1509 * disk. If there is nothing dirty, then we might need to cover the log to
1510 * indicate that the filesystem is idle.
1511 */
1512static void
1513xfs_log_worker(
1514 struct work_struct *work)
1515{
1516 struct xlog *log = container_of(to_delayed_work(work),
1517 struct xlog, l_work);
1518 struct xfs_mount *mp = log->l_mp;
1519
1520 /* dgc: errors ignored - not fatal and nowhere to report them */
1521 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1522 /*
1523 * Dump a transaction into the log that contains no real change.
1524 * This is needed to stamp the current tail LSN into the log
1525 * during the covering operation.
1526 *
1527 * We cannot use an inode here for this - that will push dirty
1528 * state back up into the VFS and then periodic inode flushing
1529 * will prevent log covering from making progress. Hence we
1530 * synchronously log the superblock instead to ensure the
1531 * superblock is immediately unpinned and can be written back.
1532 */
1533 xlog_clear_incompat(log);
1534 xfs_sync_sb(mp, true);
1535 } else
1536 xfs_log_force(mp, 0);
1537
1538 /* start pushing all the metadata that is currently dirty */
1539 xfs_ail_push_all(mp->m_ail);
1540
1541 /* queue us up again */
1542 xfs_log_work_queue(mp);
1543}
1544
1545/*
1546 * This routine initializes some of the log structure for a given mount point.
1547 * Its primary purpose is to fill in enough, so recovery can occur. However,
1548 * some other stuff may be filled in too.
1549 */
1550STATIC struct xlog *
1551xlog_alloc_log(
1552 struct xfs_mount *mp,
1553 struct xfs_buftarg *log_target,
1554 xfs_daddr_t blk_offset,
1555 int num_bblks)
1556{
1557 struct xlog *log;
1558 xlog_rec_header_t *head;
1559 xlog_in_core_t **iclogp;
1560 xlog_in_core_t *iclog, *prev_iclog=NULL;
1561 int i;
1562 int error = -ENOMEM;
1563 uint log2_size = 0;
1564
1565 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1566 if (!log) {
1567 xfs_warn(mp, "Log allocation failed: No memory!");
1568 goto out;
1569 }
1570
1571 log->l_mp = mp;
1572 log->l_targ = log_target;
1573 log->l_logsize = BBTOB(num_bblks);
1574 log->l_logBBstart = blk_offset;
1575 log->l_logBBsize = num_bblks;
1576 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1577 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1578 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1579
1580 log->l_prev_block = -1;
1581 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1582 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1583 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1584 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1585
1586 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1587 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1588 else
1589 log->l_iclog_roundoff = BBSIZE;
1590
1591 xlog_grant_head_init(&log->l_reserve_head);
1592 xlog_grant_head_init(&log->l_write_head);
1593
1594 error = -EFSCORRUPTED;
1595 if (xfs_has_sector(mp)) {
1596 log2_size = mp->m_sb.sb_logsectlog;
1597 if (log2_size < BBSHIFT) {
1598 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1599 log2_size, BBSHIFT);
1600 goto out_free_log;
1601 }
1602
1603 log2_size -= BBSHIFT;
1604 if (log2_size > mp->m_sectbb_log) {
1605 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1606 log2_size, mp->m_sectbb_log);
1607 goto out_free_log;
1608 }
1609
1610 /* for larger sector sizes, must have v2 or external log */
1611 if (log2_size && log->l_logBBstart > 0 &&
1612 !xfs_has_logv2(mp)) {
1613 xfs_warn(mp,
1614 "log sector size (0x%x) invalid for configuration.",
1615 log2_size);
1616 goto out_free_log;
1617 }
1618 }
1619 log->l_sectBBsize = 1 << log2_size;
1620
1621 init_rwsem(&log->l_incompat_users);
1622
1623 xlog_get_iclog_buffer_size(mp, log);
1624
1625 spin_lock_init(&log->l_icloglock);
1626 init_waitqueue_head(&log->l_flush_wait);
1627
1628 iclogp = &log->l_iclog;
1629 /*
1630 * The amount of memory to allocate for the iclog structure is
1631 * rather funky due to the way the structure is defined. It is
1632 * done this way so that we can use different sizes for machines
1633 * with different amounts of memory. See the definition of
1634 * xlog_in_core_t in xfs_log_priv.h for details.
1635 */
1636 ASSERT(log->l_iclog_size >= 4096);
1637 for (i = 0; i < log->l_iclog_bufs; i++) {
1638 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1639 sizeof(struct bio_vec);
1640
1641 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1642 if (!iclog)
1643 goto out_free_iclog;
1644
1645 *iclogp = iclog;
1646 iclog->ic_prev = prev_iclog;
1647 prev_iclog = iclog;
1648
1649 iclog->ic_data = kvzalloc(log->l_iclog_size,
1650 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1651 if (!iclog->ic_data)
1652 goto out_free_iclog;
1653 head = &iclog->ic_header;
1654 memset(head, 0, sizeof(xlog_rec_header_t));
1655 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1656 head->h_version = cpu_to_be32(
1657 xfs_has_logv2(log->l_mp) ? 2 : 1);
1658 head->h_size = cpu_to_be32(log->l_iclog_size);
1659 /* new fields */
1660 head->h_fmt = cpu_to_be32(XLOG_FMT);
1661 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1662
1663 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1664 iclog->ic_state = XLOG_STATE_ACTIVE;
1665 iclog->ic_log = log;
1666 atomic_set(&iclog->ic_refcnt, 0);
1667 INIT_LIST_HEAD(&iclog->ic_callbacks);
1668 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1669
1670 init_waitqueue_head(&iclog->ic_force_wait);
1671 init_waitqueue_head(&iclog->ic_write_wait);
1672 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1673 sema_init(&iclog->ic_sema, 1);
1674
1675 iclogp = &iclog->ic_next;
1676 }
1677 *iclogp = log->l_iclog; /* complete ring */
1678 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1679
1680 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1681 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1682 WQ_HIGHPRI),
1683 0, mp->m_super->s_id);
1684 if (!log->l_ioend_workqueue)
1685 goto out_free_iclog;
1686
1687 error = xlog_cil_init(log);
1688 if (error)
1689 goto out_destroy_workqueue;
1690 return log;
1691
1692out_destroy_workqueue:
1693 destroy_workqueue(log->l_ioend_workqueue);
1694out_free_iclog:
1695 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1696 prev_iclog = iclog->ic_next;
1697 kmem_free(iclog->ic_data);
1698 kmem_free(iclog);
1699 if (prev_iclog == log->l_iclog)
1700 break;
1701 }
1702out_free_log:
1703 kmem_free(log);
1704out:
1705 return ERR_PTR(error);
1706} /* xlog_alloc_log */
1707
1708/*
1709 * Compute the LSN that we'd need to push the log tail towards in order to have
1710 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1711 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1712 * log free space already meets all three thresholds, this function returns
1713 * NULLCOMMITLSN.
1714 */
1715xfs_lsn_t
1716xlog_grant_push_threshold(
1717 struct xlog *log,
1718 int need_bytes)
1719{
1720 xfs_lsn_t threshold_lsn = 0;
1721 xfs_lsn_t last_sync_lsn;
1722 int free_blocks;
1723 int free_bytes;
1724 int threshold_block;
1725 int threshold_cycle;
1726 int free_threshold;
1727
1728 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1729
1730 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1731 free_blocks = BTOBBT(free_bytes);
1732
1733 /*
1734 * Set the threshold for the minimum number of free blocks in the
1735 * log to the maximum of what the caller needs, one quarter of the
1736 * log, and 256 blocks.
1737 */
1738 free_threshold = BTOBB(need_bytes);
1739 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1740 free_threshold = max(free_threshold, 256);
1741 if (free_blocks >= free_threshold)
1742 return NULLCOMMITLSN;
1743
1744 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1745 &threshold_block);
1746 threshold_block += free_threshold;
1747 if (threshold_block >= log->l_logBBsize) {
1748 threshold_block -= log->l_logBBsize;
1749 threshold_cycle += 1;
1750 }
1751 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1752 threshold_block);
1753 /*
1754 * Don't pass in an lsn greater than the lsn of the last
1755 * log record known to be on disk. Use a snapshot of the last sync lsn
1756 * so that it doesn't change between the compare and the set.
1757 */
1758 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1759 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1760 threshold_lsn = last_sync_lsn;
1761
1762 return threshold_lsn;
1763}
1764
1765/*
1766 * Push the tail of the log if we need to do so to maintain the free log space
1767 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1768 * policy which pushes on an lsn which is further along in the log once we
1769 * reach the high water mark. In this manner, we would be creating a low water
1770 * mark.
1771 */
1772STATIC void
1773xlog_grant_push_ail(
1774 struct xlog *log,
1775 int need_bytes)
1776{
1777 xfs_lsn_t threshold_lsn;
1778
1779 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1780 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1781 return;
1782
1783 /*
1784 * Get the transaction layer to kick the dirty buffers out to
1785 * disk asynchronously. No point in trying to do this if
1786 * the filesystem is shutting down.
1787 */
1788 xfs_ail_push(log->l_ailp, threshold_lsn);
1789}
1790
1791/*
1792 * Stamp cycle number in every block
1793 */
1794STATIC void
1795xlog_pack_data(
1796 struct xlog *log,
1797 struct xlog_in_core *iclog,
1798 int roundoff)
1799{
1800 int i, j, k;
1801 int size = iclog->ic_offset + roundoff;
1802 __be32 cycle_lsn;
1803 char *dp;
1804
1805 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1806
1807 dp = iclog->ic_datap;
1808 for (i = 0; i < BTOBB(size); i++) {
1809 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1810 break;
1811 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1812 *(__be32 *)dp = cycle_lsn;
1813 dp += BBSIZE;
1814 }
1815
1816 if (xfs_has_logv2(log->l_mp)) {
1817 xlog_in_core_2_t *xhdr = iclog->ic_data;
1818
1819 for ( ; i < BTOBB(size); i++) {
1820 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1821 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1822 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1823 *(__be32 *)dp = cycle_lsn;
1824 dp += BBSIZE;
1825 }
1826
1827 for (i = 1; i < log->l_iclog_heads; i++)
1828 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1829 }
1830}
1831
1832/*
1833 * Calculate the checksum for a log buffer.
1834 *
1835 * This is a little more complicated than it should be because the various
1836 * headers and the actual data are non-contiguous.
1837 */
1838__le32
1839xlog_cksum(
1840 struct xlog *log,
1841 struct xlog_rec_header *rhead,
1842 char *dp,
1843 int size)
1844{
1845 uint32_t crc;
1846
1847 /* first generate the crc for the record header ... */
1848 crc = xfs_start_cksum_update((char *)rhead,
1849 sizeof(struct xlog_rec_header),
1850 offsetof(struct xlog_rec_header, h_crc));
1851
1852 /* ... then for additional cycle data for v2 logs ... */
1853 if (xfs_has_logv2(log->l_mp)) {
1854 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1855 int i;
1856 int xheads;
1857
1858 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1859
1860 for (i = 1; i < xheads; i++) {
1861 crc = crc32c(crc, &xhdr[i].hic_xheader,
1862 sizeof(struct xlog_rec_ext_header));
1863 }
1864 }
1865
1866 /* ... and finally for the payload */
1867 crc = crc32c(crc, dp, size);
1868
1869 return xfs_end_cksum(crc);
1870}
1871
1872static void
1873xlog_bio_end_io(
1874 struct bio *bio)
1875{
1876 struct xlog_in_core *iclog = bio->bi_private;
1877
1878 queue_work(iclog->ic_log->l_ioend_workqueue,
1879 &iclog->ic_end_io_work);
1880}
1881
1882static int
1883xlog_map_iclog_data(
1884 struct bio *bio,
1885 void *data,
1886 size_t count)
1887{
1888 do {
1889 struct page *page = kmem_to_page(data);
1890 unsigned int off = offset_in_page(data);
1891 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1892
1893 if (bio_add_page(bio, page, len, off) != len)
1894 return -EIO;
1895
1896 data += len;
1897 count -= len;
1898 } while (count);
1899
1900 return 0;
1901}
1902
1903STATIC void
1904xlog_write_iclog(
1905 struct xlog *log,
1906 struct xlog_in_core *iclog,
1907 uint64_t bno,
1908 unsigned int count)
1909{
1910 ASSERT(bno < log->l_logBBsize);
1911 trace_xlog_iclog_write(iclog, _RET_IP_);
1912
1913 /*
1914 * We lock the iclogbufs here so that we can serialise against I/O
1915 * completion during unmount. We might be processing a shutdown
1916 * triggered during unmount, and that can occur asynchronously to the
1917 * unmount thread, and hence we need to ensure that completes before
1918 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1919 * across the log IO to archieve that.
1920 */
1921 down(&iclog->ic_sema);
1922 if (xlog_is_shutdown(log)) {
1923 /*
1924 * It would seem logical to return EIO here, but we rely on
1925 * the log state machine to propagate I/O errors instead of
1926 * doing it here. We kick of the state machine and unlock
1927 * the buffer manually, the code needs to be kept in sync
1928 * with the I/O completion path.
1929 */
1930 xlog_state_done_syncing(iclog);
1931 up(&iclog->ic_sema);
1932 return;
1933 }
1934
1935 /*
1936 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1937 * IOs coming immediately after this one. This prevents the block layer
1938 * writeback throttle from throttling log writes behind background
1939 * metadata writeback and causing priority inversions.
1940 */
1941 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1942 howmany(count, PAGE_SIZE),
1943 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1944 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1945 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1946 iclog->ic_bio.bi_private = iclog;
1947
1948 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1949 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1950 /*
1951 * For external log devices, we also need to flush the data
1952 * device cache first to ensure all metadata writeback covered
1953 * by the LSN in this iclog is on stable storage. This is slow,
1954 * but it *must* complete before we issue the external log IO.
1955 *
1956 * If the flush fails, we cannot conclude that past metadata
1957 * writeback from the log succeeded. Repeating the flush is
1958 * not possible, hence we must shut down with log IO error to
1959 * avoid shutdown re-entering this path and erroring out again.
1960 */
1961 if (log->l_targ != log->l_mp->m_ddev_targp &&
1962 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev)) {
1963 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1964 return;
1965 }
1966 }
1967 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1968 iclog->ic_bio.bi_opf |= REQ_FUA;
1969
1970 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1971
1972 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1973 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1974 return;
1975 }
1976 if (is_vmalloc_addr(iclog->ic_data))
1977 flush_kernel_vmap_range(iclog->ic_data, count);
1978
1979 /*
1980 * If this log buffer would straddle the end of the log we will have
1981 * to split it up into two bios, so that we can continue at the start.
1982 */
1983 if (bno + BTOBB(count) > log->l_logBBsize) {
1984 struct bio *split;
1985
1986 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1987 GFP_NOIO, &fs_bio_set);
1988 bio_chain(split, &iclog->ic_bio);
1989 submit_bio(split);
1990
1991 /* restart at logical offset zero for the remainder */
1992 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1993 }
1994
1995 submit_bio(&iclog->ic_bio);
1996}
1997
1998/*
1999 * We need to bump cycle number for the part of the iclog that is
2000 * written to the start of the log. Watch out for the header magic
2001 * number case, though.
2002 */
2003static void
2004xlog_split_iclog(
2005 struct xlog *log,
2006 void *data,
2007 uint64_t bno,
2008 unsigned int count)
2009{
2010 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
2011 unsigned int i;
2012
2013 for (i = split_offset; i < count; i += BBSIZE) {
2014 uint32_t cycle = get_unaligned_be32(data + i);
2015
2016 if (++cycle == XLOG_HEADER_MAGIC_NUM)
2017 cycle++;
2018 put_unaligned_be32(cycle, data + i);
2019 }
2020}
2021
2022static int
2023xlog_calc_iclog_size(
2024 struct xlog *log,
2025 struct xlog_in_core *iclog,
2026 uint32_t *roundoff)
2027{
2028 uint32_t count_init, count;
2029
2030 /* Add for LR header */
2031 count_init = log->l_iclog_hsize + iclog->ic_offset;
2032 count = roundup(count_init, log->l_iclog_roundoff);
2033
2034 *roundoff = count - count_init;
2035
2036 ASSERT(count >= count_init);
2037 ASSERT(*roundoff < log->l_iclog_roundoff);
2038 return count;
2039}
2040
2041/*
2042 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2043 * fashion. Previously, we should have moved the current iclog
2044 * ptr in the log to point to the next available iclog. This allows further
2045 * write to continue while this code syncs out an iclog ready to go.
2046 * Before an in-core log can be written out, the data section must be scanned
2047 * to save away the 1st word of each BBSIZE block into the header. We replace
2048 * it with the current cycle count. Each BBSIZE block is tagged with the
2049 * cycle count because there in an implicit assumption that drives will
2050 * guarantee that entire 512 byte blocks get written at once. In other words,
2051 * we can't have part of a 512 byte block written and part not written. By
2052 * tagging each block, we will know which blocks are valid when recovering
2053 * after an unclean shutdown.
2054 *
2055 * This routine is single threaded on the iclog. No other thread can be in
2056 * this routine with the same iclog. Changing contents of iclog can there-
2057 * fore be done without grabbing the state machine lock. Updating the global
2058 * log will require grabbing the lock though.
2059 *
2060 * The entire log manager uses a logical block numbering scheme. Only
2061 * xlog_write_iclog knows about the fact that the log may not start with
2062 * block zero on a given device.
2063 */
2064STATIC void
2065xlog_sync(
2066 struct xlog *log,
2067 struct xlog_in_core *iclog,
2068 struct xlog_ticket *ticket)
2069{
2070 unsigned int count; /* byte count of bwrite */
2071 unsigned int roundoff; /* roundoff to BB or stripe */
2072 uint64_t bno;
2073 unsigned int size;
2074
2075 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2076 trace_xlog_iclog_sync(iclog, _RET_IP_);
2077
2078 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2079
2080 /*
2081 * If we have a ticket, account for the roundoff via the ticket
2082 * reservation to avoid touching the hot grant heads needlessly.
2083 * Otherwise, we have to move grant heads directly.
2084 */
2085 if (ticket) {
2086 ticket->t_curr_res -= roundoff;
2087 } else {
2088 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2089 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2090 }
2091
2092 /* put cycle number in every block */
2093 xlog_pack_data(log, iclog, roundoff);
2094
2095 /* real byte length */
2096 size = iclog->ic_offset;
2097 if (xfs_has_logv2(log->l_mp))
2098 size += roundoff;
2099 iclog->ic_header.h_len = cpu_to_be32(size);
2100
2101 XFS_STATS_INC(log->l_mp, xs_log_writes);
2102 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2103
2104 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2105
2106 /* Do we need to split this write into 2 parts? */
2107 if (bno + BTOBB(count) > log->l_logBBsize)
2108 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2109
2110 /* calculcate the checksum */
2111 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2112 iclog->ic_datap, size);
2113 /*
2114 * Intentionally corrupt the log record CRC based on the error injection
2115 * frequency, if defined. This facilitates testing log recovery in the
2116 * event of torn writes. Hence, set the IOABORT state to abort the log
2117 * write on I/O completion and shutdown the fs. The subsequent mount
2118 * detects the bad CRC and attempts to recover.
2119 */
2120#ifdef DEBUG
2121 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2122 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2123 iclog->ic_fail_crc = true;
2124 xfs_warn(log->l_mp,
2125 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2126 be64_to_cpu(iclog->ic_header.h_lsn));
2127 }
2128#endif
2129 xlog_verify_iclog(log, iclog, count);
2130 xlog_write_iclog(log, iclog, bno, count);
2131}
2132
2133/*
2134 * Deallocate a log structure
2135 */
2136STATIC void
2137xlog_dealloc_log(
2138 struct xlog *log)
2139{
2140 xlog_in_core_t *iclog, *next_iclog;
2141 int i;
2142
2143 /*
2144 * Destroy the CIL after waiting for iclog IO completion because an
2145 * iclog EIO error will try to shut down the log, which accesses the
2146 * CIL to wake up the waiters.
2147 */
2148 xlog_cil_destroy(log);
2149
2150 iclog = log->l_iclog;
2151 for (i = 0; i < log->l_iclog_bufs; i++) {
2152 next_iclog = iclog->ic_next;
2153 kmem_free(iclog->ic_data);
2154 kmem_free(iclog);
2155 iclog = next_iclog;
2156 }
2157
2158 log->l_mp->m_log = NULL;
2159 destroy_workqueue(log->l_ioend_workqueue);
2160 kmem_free(log);
2161}
2162
2163/*
2164 * Update counters atomically now that memcpy is done.
2165 */
2166static inline void
2167xlog_state_finish_copy(
2168 struct xlog *log,
2169 struct xlog_in_core *iclog,
2170 int record_cnt,
2171 int copy_bytes)
2172{
2173 lockdep_assert_held(&log->l_icloglock);
2174
2175 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2176 iclog->ic_offset += copy_bytes;
2177}
2178
2179/*
2180 * print out info relating to regions written which consume
2181 * the reservation
2182 */
2183void
2184xlog_print_tic_res(
2185 struct xfs_mount *mp,
2186 struct xlog_ticket *ticket)
2187{
2188 xfs_warn(mp, "ticket reservation summary:");
2189 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
2190 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
2191 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
2192 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
2193}
2194
2195/*
2196 * Print a summary of the transaction.
2197 */
2198void
2199xlog_print_trans(
2200 struct xfs_trans *tp)
2201{
2202 struct xfs_mount *mp = tp->t_mountp;
2203 struct xfs_log_item *lip;
2204
2205 /* dump core transaction and ticket info */
2206 xfs_warn(mp, "transaction summary:");
2207 xfs_warn(mp, " log res = %d", tp->t_log_res);
2208 xfs_warn(mp, " log count = %d", tp->t_log_count);
2209 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2210
2211 xlog_print_tic_res(mp, tp->t_ticket);
2212
2213 /* dump each log item */
2214 list_for_each_entry(lip, &tp->t_items, li_trans) {
2215 struct xfs_log_vec *lv = lip->li_lv;
2216 struct xfs_log_iovec *vec;
2217 int i;
2218
2219 xfs_warn(mp, "log item: ");
2220 xfs_warn(mp, " type = 0x%x", lip->li_type);
2221 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2222 if (!lv)
2223 continue;
2224 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2225 xfs_warn(mp, " size = %d", lv->lv_size);
2226 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2227 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2228
2229 /* dump each iovec for the log item */
2230 vec = lv->lv_iovecp;
2231 for (i = 0; i < lv->lv_niovecs; i++) {
2232 int dumplen = min(vec->i_len, 32);
2233
2234 xfs_warn(mp, " iovec[%d]", i);
2235 xfs_warn(mp, " type = 0x%x", vec->i_type);
2236 xfs_warn(mp, " len = %d", vec->i_len);
2237 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2238 xfs_hex_dump(vec->i_addr, dumplen);
2239
2240 vec++;
2241 }
2242 }
2243}
2244
2245static inline void
2246xlog_write_iovec(
2247 struct xlog_in_core *iclog,
2248 uint32_t *log_offset,
2249 void *data,
2250 uint32_t write_len,
2251 int *bytes_left,
2252 uint32_t *record_cnt,
2253 uint32_t *data_cnt)
2254{
2255 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2256 ASSERT(*log_offset % sizeof(int32_t) == 0);
2257 ASSERT(write_len % sizeof(int32_t) == 0);
2258
2259 memcpy(iclog->ic_datap + *log_offset, data, write_len);
2260 *log_offset += write_len;
2261 *bytes_left -= write_len;
2262 (*record_cnt)++;
2263 *data_cnt += write_len;
2264}
2265
2266/*
2267 * Write log vectors into a single iclog which is guaranteed by the caller
2268 * to have enough space to write the entire log vector into.
2269 */
2270static void
2271xlog_write_full(
2272 struct xfs_log_vec *lv,
2273 struct xlog_ticket *ticket,
2274 struct xlog_in_core *iclog,
2275 uint32_t *log_offset,
2276 uint32_t *len,
2277 uint32_t *record_cnt,
2278 uint32_t *data_cnt)
2279{
2280 int index;
2281
2282 ASSERT(*log_offset + *len <= iclog->ic_size ||
2283 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2284
2285 /*
2286 * Ordered log vectors have no regions to write so this
2287 * loop will naturally skip them.
2288 */
2289 for (index = 0; index < lv->lv_niovecs; index++) {
2290 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2291 struct xlog_op_header *ophdr = reg->i_addr;
2292
2293 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2294 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2295 reg->i_len, len, record_cnt, data_cnt);
2296 }
2297}
2298
2299static int
2300xlog_write_get_more_iclog_space(
2301 struct xlog_ticket *ticket,
2302 struct xlog_in_core **iclogp,
2303 uint32_t *log_offset,
2304 uint32_t len,
2305 uint32_t *record_cnt,
2306 uint32_t *data_cnt)
2307{
2308 struct xlog_in_core *iclog = *iclogp;
2309 struct xlog *log = iclog->ic_log;
2310 int error;
2311
2312 spin_lock(&log->l_icloglock);
2313 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2314 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2315 error = xlog_state_release_iclog(log, iclog, ticket);
2316 spin_unlock(&log->l_icloglock);
2317 if (error)
2318 return error;
2319
2320 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2321 log_offset);
2322 if (error)
2323 return error;
2324 *record_cnt = 0;
2325 *data_cnt = 0;
2326 *iclogp = iclog;
2327 return 0;
2328}
2329
2330/*
2331 * Write log vectors into a single iclog which is smaller than the current chain
2332 * length. We write until we cannot fit a full record into the remaining space
2333 * and then stop. We return the log vector that is to be written that cannot
2334 * wholly fit in the iclog.
2335 */
2336static int
2337xlog_write_partial(
2338 struct xfs_log_vec *lv,
2339 struct xlog_ticket *ticket,
2340 struct xlog_in_core **iclogp,
2341 uint32_t *log_offset,
2342 uint32_t *len,
2343 uint32_t *record_cnt,
2344 uint32_t *data_cnt)
2345{
2346 struct xlog_in_core *iclog = *iclogp;
2347 struct xlog_op_header *ophdr;
2348 int index = 0;
2349 uint32_t rlen;
2350 int error;
2351
2352 /* walk the logvec, copying until we run out of space in the iclog */
2353 for (index = 0; index < lv->lv_niovecs; index++) {
2354 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2355 uint32_t reg_offset = 0;
2356
2357 /*
2358 * The first region of a continuation must have a non-zero
2359 * length otherwise log recovery will just skip over it and
2360 * start recovering from the next opheader it finds. Because we
2361 * mark the next opheader as a continuation, recovery will then
2362 * incorrectly add the continuation to the previous region and
2363 * that breaks stuff.
2364 *
2365 * Hence if there isn't space for region data after the
2366 * opheader, then we need to start afresh with a new iclog.
2367 */
2368 if (iclog->ic_size - *log_offset <=
2369 sizeof(struct xlog_op_header)) {
2370 error = xlog_write_get_more_iclog_space(ticket,
2371 &iclog, log_offset, *len, record_cnt,
2372 data_cnt);
2373 if (error)
2374 return error;
2375 }
2376
2377 ophdr = reg->i_addr;
2378 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2379
2380 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2381 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2382 if (rlen != reg->i_len)
2383 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2384
2385 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2386 rlen, len, record_cnt, data_cnt);
2387
2388 /* If we wrote the whole region, move to the next. */
2389 if (rlen == reg->i_len)
2390 continue;
2391
2392 /*
2393 * We now have a partially written iovec, but it can span
2394 * multiple iclogs so we loop here. First we release the iclog
2395 * we currently have, then we get a new iclog and add a new
2396 * opheader. Then we continue copying from where we were until
2397 * we either complete the iovec or fill the iclog. If we
2398 * complete the iovec, then we increment the index and go right
2399 * back to the top of the outer loop. if we fill the iclog, we
2400 * run the inner loop again.
2401 *
2402 * This is complicated by the tail of a region using all the
2403 * space in an iclog and hence requiring us to release the iclog
2404 * and get a new one before returning to the outer loop. We must
2405 * always guarantee that we exit this inner loop with at least
2406 * space for log transaction opheaders left in the current
2407 * iclog, hence we cannot just terminate the loop at the end
2408 * of the of the continuation. So we loop while there is no
2409 * space left in the current iclog, and check for the end of the
2410 * continuation after getting a new iclog.
2411 */
2412 do {
2413 /*
2414 * Ensure we include the continuation opheader in the
2415 * space we need in the new iclog by adding that size
2416 * to the length we require. This continuation opheader
2417 * needs to be accounted to the ticket as the space it
2418 * consumes hasn't been accounted to the lv we are
2419 * writing.
2420 */
2421 error = xlog_write_get_more_iclog_space(ticket,
2422 &iclog, log_offset,
2423 *len + sizeof(struct xlog_op_header),
2424 record_cnt, data_cnt);
2425 if (error)
2426 return error;
2427
2428 ophdr = iclog->ic_datap + *log_offset;
2429 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2430 ophdr->oh_clientid = XFS_TRANSACTION;
2431 ophdr->oh_res2 = 0;
2432 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2433
2434 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2435 *log_offset += sizeof(struct xlog_op_header);
2436 *data_cnt += sizeof(struct xlog_op_header);
2437
2438 /*
2439 * If rlen fits in the iclog, then end the region
2440 * continuation. Otherwise we're going around again.
2441 */
2442 reg_offset += rlen;
2443 rlen = reg->i_len - reg_offset;
2444 if (rlen <= iclog->ic_size - *log_offset)
2445 ophdr->oh_flags |= XLOG_END_TRANS;
2446 else
2447 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2448
2449 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2450 ophdr->oh_len = cpu_to_be32(rlen);
2451
2452 xlog_write_iovec(iclog, log_offset,
2453 reg->i_addr + reg_offset,
2454 rlen, len, record_cnt, data_cnt);
2455
2456 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2457 }
2458
2459 /*
2460 * No more iovecs remain in this logvec so return the next log vec to
2461 * the caller so it can go back to fast path copying.
2462 */
2463 *iclogp = iclog;
2464 return 0;
2465}
2466
2467/*
2468 * Write some region out to in-core log
2469 *
2470 * This will be called when writing externally provided regions or when
2471 * writing out a commit record for a given transaction.
2472 *
2473 * General algorithm:
2474 * 1. Find total length of this write. This may include adding to the
2475 * lengths passed in.
2476 * 2. Check whether we violate the tickets reservation.
2477 * 3. While writing to this iclog
2478 * A. Reserve as much space in this iclog as can get
2479 * B. If this is first write, save away start lsn
2480 * C. While writing this region:
2481 * 1. If first write of transaction, write start record
2482 * 2. Write log operation header (header per region)
2483 * 3. Find out if we can fit entire region into this iclog
2484 * 4. Potentially, verify destination memcpy ptr
2485 * 5. Memcpy (partial) region
2486 * 6. If partial copy, release iclog; otherwise, continue
2487 * copying more regions into current iclog
2488 * 4. Mark want sync bit (in simulation mode)
2489 * 5. Release iclog for potential flush to on-disk log.
2490 *
2491 * ERRORS:
2492 * 1. Panic if reservation is overrun. This should never happen since
2493 * reservation amounts are generated internal to the filesystem.
2494 * NOTES:
2495 * 1. Tickets are single threaded data structures.
2496 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2497 * syncing routine. When a single log_write region needs to span
2498 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2499 * on all log operation writes which don't contain the end of the
2500 * region. The XLOG_END_TRANS bit is used for the in-core log
2501 * operation which contains the end of the continued log_write region.
2502 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2503 * we don't really know exactly how much space will be used. As a result,
2504 * we don't update ic_offset until the end when we know exactly how many
2505 * bytes have been written out.
2506 */
2507int
2508xlog_write(
2509 struct xlog *log,
2510 struct xfs_cil_ctx *ctx,
2511 struct list_head *lv_chain,
2512 struct xlog_ticket *ticket,
2513 uint32_t len)
2514
2515{
2516 struct xlog_in_core *iclog = NULL;
2517 struct xfs_log_vec *lv;
2518 uint32_t record_cnt = 0;
2519 uint32_t data_cnt = 0;
2520 int error = 0;
2521 int log_offset;
2522
2523 if (ticket->t_curr_res < 0) {
2524 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2525 "ctx ticket reservation ran out. Need to up reservation");
2526 xlog_print_tic_res(log->l_mp, ticket);
2527 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2528 }
2529
2530 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2531 &log_offset);
2532 if (error)
2533 return error;
2534
2535 ASSERT(log_offset <= iclog->ic_size - 1);
2536
2537 /*
2538 * If we have a context pointer, pass it the first iclog we are
2539 * writing to so it can record state needed for iclog write
2540 * ordering.
2541 */
2542 if (ctx)
2543 xlog_cil_set_ctx_write_state(ctx, iclog);
2544
2545 list_for_each_entry(lv, lv_chain, lv_list) {
2546 /*
2547 * If the entire log vec does not fit in the iclog, punt it to
2548 * the partial copy loop which can handle this case.
2549 */
2550 if (lv->lv_niovecs &&
2551 lv->lv_bytes > iclog->ic_size - log_offset) {
2552 error = xlog_write_partial(lv, ticket, &iclog,
2553 &log_offset, &len, &record_cnt,
2554 &data_cnt);
2555 if (error) {
2556 /*
2557 * We have no iclog to release, so just return
2558 * the error immediately.
2559 */
2560 return error;
2561 }
2562 } else {
2563 xlog_write_full(lv, ticket, iclog, &log_offset,
2564 &len, &record_cnt, &data_cnt);
2565 }
2566 }
2567 ASSERT(len == 0);
2568
2569 /*
2570 * We've already been guaranteed that the last writes will fit inside
2571 * the current iclog, and hence it will already have the space used by
2572 * those writes accounted to it. Hence we do not need to update the
2573 * iclog with the number of bytes written here.
2574 */
2575 spin_lock(&log->l_icloglock);
2576 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2577 error = xlog_state_release_iclog(log, iclog, ticket);
2578 spin_unlock(&log->l_icloglock);
2579
2580 return error;
2581}
2582
2583static void
2584xlog_state_activate_iclog(
2585 struct xlog_in_core *iclog,
2586 int *iclogs_changed)
2587{
2588 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2589 trace_xlog_iclog_activate(iclog, _RET_IP_);
2590
2591 /*
2592 * If the number of ops in this iclog indicate it just contains the
2593 * dummy transaction, we can change state into IDLE (the second time
2594 * around). Otherwise we should change the state into NEED a dummy.
2595 * We don't need to cover the dummy.
2596 */
2597 if (*iclogs_changed == 0 &&
2598 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2599 *iclogs_changed = 1;
2600 } else {
2601 /*
2602 * We have two dirty iclogs so start over. This could also be
2603 * num of ops indicating this is not the dummy going out.
2604 */
2605 *iclogs_changed = 2;
2606 }
2607
2608 iclog->ic_state = XLOG_STATE_ACTIVE;
2609 iclog->ic_offset = 0;
2610 iclog->ic_header.h_num_logops = 0;
2611 memset(iclog->ic_header.h_cycle_data, 0,
2612 sizeof(iclog->ic_header.h_cycle_data));
2613 iclog->ic_header.h_lsn = 0;
2614 iclog->ic_header.h_tail_lsn = 0;
2615}
2616
2617/*
2618 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2619 * ACTIVE after iclog I/O has completed.
2620 */
2621static void
2622xlog_state_activate_iclogs(
2623 struct xlog *log,
2624 int *iclogs_changed)
2625{
2626 struct xlog_in_core *iclog = log->l_iclog;
2627
2628 do {
2629 if (iclog->ic_state == XLOG_STATE_DIRTY)
2630 xlog_state_activate_iclog(iclog, iclogs_changed);
2631 /*
2632 * The ordering of marking iclogs ACTIVE must be maintained, so
2633 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2634 */
2635 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2636 break;
2637 } while ((iclog = iclog->ic_next) != log->l_iclog);
2638}
2639
2640static int
2641xlog_covered_state(
2642 int prev_state,
2643 int iclogs_changed)
2644{
2645 /*
2646 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2647 * wrote the first covering record (DONE). We go to IDLE if we just
2648 * wrote the second covering record (DONE2) and remain in IDLE until a
2649 * non-covering write occurs.
2650 */
2651 switch (prev_state) {
2652 case XLOG_STATE_COVER_IDLE:
2653 if (iclogs_changed == 1)
2654 return XLOG_STATE_COVER_IDLE;
2655 fallthrough;
2656 case XLOG_STATE_COVER_NEED:
2657 case XLOG_STATE_COVER_NEED2:
2658 break;
2659 case XLOG_STATE_COVER_DONE:
2660 if (iclogs_changed == 1)
2661 return XLOG_STATE_COVER_NEED2;
2662 break;
2663 case XLOG_STATE_COVER_DONE2:
2664 if (iclogs_changed == 1)
2665 return XLOG_STATE_COVER_IDLE;
2666 break;
2667 default:
2668 ASSERT(0);
2669 }
2670
2671 return XLOG_STATE_COVER_NEED;
2672}
2673
2674STATIC void
2675xlog_state_clean_iclog(
2676 struct xlog *log,
2677 struct xlog_in_core *dirty_iclog)
2678{
2679 int iclogs_changed = 0;
2680
2681 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2682
2683 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2684
2685 xlog_state_activate_iclogs(log, &iclogs_changed);
2686 wake_up_all(&dirty_iclog->ic_force_wait);
2687
2688 if (iclogs_changed) {
2689 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2690 iclogs_changed);
2691 }
2692}
2693
2694STATIC xfs_lsn_t
2695xlog_get_lowest_lsn(
2696 struct xlog *log)
2697{
2698 struct xlog_in_core *iclog = log->l_iclog;
2699 xfs_lsn_t lowest_lsn = 0, lsn;
2700
2701 do {
2702 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2703 iclog->ic_state == XLOG_STATE_DIRTY)
2704 continue;
2705
2706 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2707 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2708 lowest_lsn = lsn;
2709 } while ((iclog = iclog->ic_next) != log->l_iclog);
2710
2711 return lowest_lsn;
2712}
2713
2714/*
2715 * Completion of a iclog IO does not imply that a transaction has completed, as
2716 * transactions can be large enough to span many iclogs. We cannot change the
2717 * tail of the log half way through a transaction as this may be the only
2718 * transaction in the log and moving the tail to point to the middle of it
2719 * will prevent recovery from finding the start of the transaction. Hence we
2720 * should only update the last_sync_lsn if this iclog contains transaction
2721 * completion callbacks on it.
2722 *
2723 * We have to do this before we drop the icloglock to ensure we are the only one
2724 * that can update it.
2725 *
2726 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2727 * the reservation grant head pushing. This is due to the fact that the push
2728 * target is bound by the current last_sync_lsn value. Hence if we have a large
2729 * amount of log space bound up in this committing transaction then the
2730 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2731 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2732 * should push the AIL to ensure the push target (and hence the grant head) is
2733 * no longer bound by the old log head location and can move forwards and make
2734 * progress again.
2735 */
2736static void
2737xlog_state_set_callback(
2738 struct xlog *log,
2739 struct xlog_in_core *iclog,
2740 xfs_lsn_t header_lsn)
2741{
2742 trace_xlog_iclog_callback(iclog, _RET_IP_);
2743 iclog->ic_state = XLOG_STATE_CALLBACK;
2744
2745 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2746 header_lsn) <= 0);
2747
2748 if (list_empty_careful(&iclog->ic_callbacks))
2749 return;
2750
2751 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2752 xlog_grant_push_ail(log, 0);
2753}
2754
2755/*
2756 * Return true if we need to stop processing, false to continue to the next
2757 * iclog. The caller will need to run callbacks if the iclog is returned in the
2758 * XLOG_STATE_CALLBACK state.
2759 */
2760static bool
2761xlog_state_iodone_process_iclog(
2762 struct xlog *log,
2763 struct xlog_in_core *iclog)
2764{
2765 xfs_lsn_t lowest_lsn;
2766 xfs_lsn_t header_lsn;
2767
2768 switch (iclog->ic_state) {
2769 case XLOG_STATE_ACTIVE:
2770 case XLOG_STATE_DIRTY:
2771 /*
2772 * Skip all iclogs in the ACTIVE & DIRTY states:
2773 */
2774 return false;
2775 case XLOG_STATE_DONE_SYNC:
2776 /*
2777 * Now that we have an iclog that is in the DONE_SYNC state, do
2778 * one more check here to see if we have chased our tail around.
2779 * If this is not the lowest lsn iclog, then we will leave it
2780 * for another completion to process.
2781 */
2782 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2783 lowest_lsn = xlog_get_lowest_lsn(log);
2784 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2785 return false;
2786 xlog_state_set_callback(log, iclog, header_lsn);
2787 return false;
2788 default:
2789 /*
2790 * Can only perform callbacks in order. Since this iclog is not
2791 * in the DONE_SYNC state, we skip the rest and just try to
2792 * clean up.
2793 */
2794 return true;
2795 }
2796}
2797
2798/*
2799 * Loop over all the iclogs, running attached callbacks on them. Return true if
2800 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2801 * to handle transient shutdown state here at all because
2802 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2803 * cleanup of the callbacks.
2804 */
2805static bool
2806xlog_state_do_iclog_callbacks(
2807 struct xlog *log)
2808 __releases(&log->l_icloglock)
2809 __acquires(&log->l_icloglock)
2810{
2811 struct xlog_in_core *first_iclog = log->l_iclog;
2812 struct xlog_in_core *iclog = first_iclog;
2813 bool ran_callback = false;
2814
2815 do {
2816 LIST_HEAD(cb_list);
2817
2818 if (xlog_state_iodone_process_iclog(log, iclog))
2819 break;
2820 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2821 iclog = iclog->ic_next;
2822 continue;
2823 }
2824 list_splice_init(&iclog->ic_callbacks, &cb_list);
2825 spin_unlock(&log->l_icloglock);
2826
2827 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2828 xlog_cil_process_committed(&cb_list);
2829 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2830 ran_callback = true;
2831
2832 spin_lock(&log->l_icloglock);
2833 xlog_state_clean_iclog(log, iclog);
2834 iclog = iclog->ic_next;
2835 } while (iclog != first_iclog);
2836
2837 return ran_callback;
2838}
2839
2840
2841/*
2842 * Loop running iclog completion callbacks until there are no more iclogs in a
2843 * state that can run callbacks.
2844 */
2845STATIC void
2846xlog_state_do_callback(
2847 struct xlog *log)
2848{
2849 int flushcnt = 0;
2850 int repeats = 0;
2851
2852 spin_lock(&log->l_icloglock);
2853 while (xlog_state_do_iclog_callbacks(log)) {
2854 if (xlog_is_shutdown(log))
2855 break;
2856
2857 if (++repeats > 5000) {
2858 flushcnt += repeats;
2859 repeats = 0;
2860 xfs_warn(log->l_mp,
2861 "%s: possible infinite loop (%d iterations)",
2862 __func__, flushcnt);
2863 }
2864 }
2865
2866 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2867 wake_up_all(&log->l_flush_wait);
2868
2869 spin_unlock(&log->l_icloglock);
2870}
2871
2872
2873/*
2874 * Finish transitioning this iclog to the dirty state.
2875 *
2876 * Callbacks could take time, so they are done outside the scope of the
2877 * global state machine log lock.
2878 */
2879STATIC void
2880xlog_state_done_syncing(
2881 struct xlog_in_core *iclog)
2882{
2883 struct xlog *log = iclog->ic_log;
2884
2885 spin_lock(&log->l_icloglock);
2886 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2887 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2888
2889 /*
2890 * If we got an error, either on the first buffer, or in the case of
2891 * split log writes, on the second, we shut down the file system and
2892 * no iclogs should ever be attempted to be written to disk again.
2893 */
2894 if (!xlog_is_shutdown(log)) {
2895 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2896 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2897 }
2898
2899 /*
2900 * Someone could be sleeping prior to writing out the next
2901 * iclog buffer, we wake them all, one will get to do the
2902 * I/O, the others get to wait for the result.
2903 */
2904 wake_up_all(&iclog->ic_write_wait);
2905 spin_unlock(&log->l_icloglock);
2906 xlog_state_do_callback(log);
2907}
2908
2909/*
2910 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2911 * sleep. We wait on the flush queue on the head iclog as that should be
2912 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2913 * we will wait here and all new writes will sleep until a sync completes.
2914 *
2915 * The in-core logs are used in a circular fashion. They are not used
2916 * out-of-order even when an iclog past the head is free.
2917 *
2918 * return:
2919 * * log_offset where xlog_write() can start writing into the in-core
2920 * log's data space.
2921 * * in-core log pointer to which xlog_write() should write.
2922 * * boolean indicating this is a continued write to an in-core log.
2923 * If this is the last write, then the in-core log's offset field
2924 * needs to be incremented, depending on the amount of data which
2925 * is copied.
2926 */
2927STATIC int
2928xlog_state_get_iclog_space(
2929 struct xlog *log,
2930 int len,
2931 struct xlog_in_core **iclogp,
2932 struct xlog_ticket *ticket,
2933 int *logoffsetp)
2934{
2935 int log_offset;
2936 xlog_rec_header_t *head;
2937 xlog_in_core_t *iclog;
2938
2939restart:
2940 spin_lock(&log->l_icloglock);
2941 if (xlog_is_shutdown(log)) {
2942 spin_unlock(&log->l_icloglock);
2943 return -EIO;
2944 }
2945
2946 iclog = log->l_iclog;
2947 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2948 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2949
2950 /* Wait for log writes to have flushed */
2951 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2952 goto restart;
2953 }
2954
2955 head = &iclog->ic_header;
2956
2957 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2958 log_offset = iclog->ic_offset;
2959
2960 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2961
2962 /* On the 1st write to an iclog, figure out lsn. This works
2963 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2964 * committing to. If the offset is set, that's how many blocks
2965 * must be written.
2966 */
2967 if (log_offset == 0) {
2968 ticket->t_curr_res -= log->l_iclog_hsize;
2969 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2970 head->h_lsn = cpu_to_be64(
2971 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2972 ASSERT(log->l_curr_block >= 0);
2973 }
2974
2975 /* If there is enough room to write everything, then do it. Otherwise,
2976 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2977 * bit is on, so this will get flushed out. Don't update ic_offset
2978 * until you know exactly how many bytes get copied. Therefore, wait
2979 * until later to update ic_offset.
2980 *
2981 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2982 * can fit into remaining data section.
2983 */
2984 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2985 int error = 0;
2986
2987 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2988
2989 /*
2990 * If we are the only one writing to this iclog, sync it to
2991 * disk. We need to do an atomic compare and decrement here to
2992 * avoid racing with concurrent atomic_dec_and_lock() calls in
2993 * xlog_state_release_iclog() when there is more than one
2994 * reference to the iclog.
2995 */
2996 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2997 error = xlog_state_release_iclog(log, iclog, ticket);
2998 spin_unlock(&log->l_icloglock);
2999 if (error)
3000 return error;
3001 goto restart;
3002 }
3003
3004 /* Do we have enough room to write the full amount in the remainder
3005 * of this iclog? Or must we continue a write on the next iclog and
3006 * mark this iclog as completely taken? In the case where we switch
3007 * iclogs (to mark it taken), this particular iclog will release/sync
3008 * to disk in xlog_write().
3009 */
3010 if (len <= iclog->ic_size - iclog->ic_offset)
3011 iclog->ic_offset += len;
3012 else
3013 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3014 *iclogp = iclog;
3015
3016 ASSERT(iclog->ic_offset <= iclog->ic_size);
3017 spin_unlock(&log->l_icloglock);
3018
3019 *logoffsetp = log_offset;
3020 return 0;
3021}
3022
3023/*
3024 * The first cnt-1 times a ticket goes through here we don't need to move the
3025 * grant write head because the permanent reservation has reserved cnt times the
3026 * unit amount. Release part of current permanent unit reservation and reset
3027 * current reservation to be one units worth. Also move grant reservation head
3028 * forward.
3029 */
3030void
3031xfs_log_ticket_regrant(
3032 struct xlog *log,
3033 struct xlog_ticket *ticket)
3034{
3035 trace_xfs_log_ticket_regrant(log, ticket);
3036
3037 if (ticket->t_cnt > 0)
3038 ticket->t_cnt--;
3039
3040 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3041 ticket->t_curr_res);
3042 xlog_grant_sub_space(log, &log->l_write_head.grant,
3043 ticket->t_curr_res);
3044 ticket->t_curr_res = ticket->t_unit_res;
3045
3046 trace_xfs_log_ticket_regrant_sub(log, ticket);
3047
3048 /* just return if we still have some of the pre-reserved space */
3049 if (!ticket->t_cnt) {
3050 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3051 ticket->t_unit_res);
3052 trace_xfs_log_ticket_regrant_exit(log, ticket);
3053
3054 ticket->t_curr_res = ticket->t_unit_res;
3055 }
3056
3057 xfs_log_ticket_put(ticket);
3058}
3059
3060/*
3061 * Give back the space left from a reservation.
3062 *
3063 * All the information we need to make a correct determination of space left
3064 * is present. For non-permanent reservations, things are quite easy. The
3065 * count should have been decremented to zero. We only need to deal with the
3066 * space remaining in the current reservation part of the ticket. If the
3067 * ticket contains a permanent reservation, there may be left over space which
3068 * needs to be released. A count of N means that N-1 refills of the current
3069 * reservation can be done before we need to ask for more space. The first
3070 * one goes to fill up the first current reservation. Once we run out of
3071 * space, the count will stay at zero and the only space remaining will be
3072 * in the current reservation field.
3073 */
3074void
3075xfs_log_ticket_ungrant(
3076 struct xlog *log,
3077 struct xlog_ticket *ticket)
3078{
3079 int bytes;
3080
3081 trace_xfs_log_ticket_ungrant(log, ticket);
3082
3083 if (ticket->t_cnt > 0)
3084 ticket->t_cnt--;
3085
3086 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3087
3088 /*
3089 * If this is a permanent reservation ticket, we may be able to free
3090 * up more space based on the remaining count.
3091 */
3092 bytes = ticket->t_curr_res;
3093 if (ticket->t_cnt > 0) {
3094 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3095 bytes += ticket->t_unit_res*ticket->t_cnt;
3096 }
3097
3098 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3099 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3100
3101 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3102
3103 xfs_log_space_wake(log->l_mp);
3104 xfs_log_ticket_put(ticket);
3105}
3106
3107/*
3108 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3109 * the current iclog pointer to the next iclog in the ring.
3110 */
3111void
3112xlog_state_switch_iclogs(
3113 struct xlog *log,
3114 struct xlog_in_core *iclog,
3115 int eventual_size)
3116{
3117 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3118 assert_spin_locked(&log->l_icloglock);
3119 trace_xlog_iclog_switch(iclog, _RET_IP_);
3120
3121 if (!eventual_size)
3122 eventual_size = iclog->ic_offset;
3123 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3124 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3125 log->l_prev_block = log->l_curr_block;
3126 log->l_prev_cycle = log->l_curr_cycle;
3127
3128 /* roll log?: ic_offset changed later */
3129 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3130
3131 /* Round up to next log-sunit */
3132 if (log->l_iclog_roundoff > BBSIZE) {
3133 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3134 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3135 }
3136
3137 if (log->l_curr_block >= log->l_logBBsize) {
3138 /*
3139 * Rewind the current block before the cycle is bumped to make
3140 * sure that the combined LSN never transiently moves forward
3141 * when the log wraps to the next cycle. This is to support the
3142 * unlocked sample of these fields from xlog_valid_lsn(). Most
3143 * other cases should acquire l_icloglock.
3144 */
3145 log->l_curr_block -= log->l_logBBsize;
3146 ASSERT(log->l_curr_block >= 0);
3147 smp_wmb();
3148 log->l_curr_cycle++;
3149 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3150 log->l_curr_cycle++;
3151 }
3152 ASSERT(iclog == log->l_iclog);
3153 log->l_iclog = iclog->ic_next;
3154}
3155
3156/*
3157 * Force the iclog to disk and check if the iclog has been completed before
3158 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3159 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3160 * If completion has already occurred, tell the caller so that it can avoid an
3161 * unnecessary wait on the iclog.
3162 */
3163static int
3164xlog_force_and_check_iclog(
3165 struct xlog_in_core *iclog,
3166 bool *completed)
3167{
3168 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3169 int error;
3170
3171 *completed = false;
3172 error = xlog_force_iclog(iclog);
3173 if (error)
3174 return error;
3175
3176 /*
3177 * If the iclog has already been completed and reused the header LSN
3178 * will have been rewritten by completion
3179 */
3180 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3181 *completed = true;
3182 return 0;
3183}
3184
3185/*
3186 * Write out all data in the in-core log as of this exact moment in time.
3187 *
3188 * Data may be written to the in-core log during this call. However,
3189 * we don't guarantee this data will be written out. A change from past
3190 * implementation means this routine will *not* write out zero length LRs.
3191 *
3192 * Basically, we try and perform an intelligent scan of the in-core logs.
3193 * If we determine there is no flushable data, we just return. There is no
3194 * flushable data if:
3195 *
3196 * 1. the current iclog is active and has no data; the previous iclog
3197 * is in the active or dirty state.
3198 * 2. the current iclog is drity, and the previous iclog is in the
3199 * active or dirty state.
3200 *
3201 * We may sleep if:
3202 *
3203 * 1. the current iclog is not in the active nor dirty state.
3204 * 2. the current iclog dirty, and the previous iclog is not in the
3205 * active nor dirty state.
3206 * 3. the current iclog is active, and there is another thread writing
3207 * to this particular iclog.
3208 * 4. a) the current iclog is active and has no other writers
3209 * b) when we return from flushing out this iclog, it is still
3210 * not in the active nor dirty state.
3211 */
3212int
3213xfs_log_force(
3214 struct xfs_mount *mp,
3215 uint flags)
3216{
3217 struct xlog *log = mp->m_log;
3218 struct xlog_in_core *iclog;
3219
3220 XFS_STATS_INC(mp, xs_log_force);
3221 trace_xfs_log_force(mp, 0, _RET_IP_);
3222
3223 xlog_cil_force(log);
3224
3225 spin_lock(&log->l_icloglock);
3226 if (xlog_is_shutdown(log))
3227 goto out_error;
3228
3229 iclog = log->l_iclog;
3230 trace_xlog_iclog_force(iclog, _RET_IP_);
3231
3232 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3233 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3234 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3235 /*
3236 * If the head is dirty or (active and empty), then we need to
3237 * look at the previous iclog.
3238 *
3239 * If the previous iclog is active or dirty we are done. There
3240 * is nothing to sync out. Otherwise, we attach ourselves to the
3241 * previous iclog and go to sleep.
3242 */
3243 iclog = iclog->ic_prev;
3244 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3245 if (atomic_read(&iclog->ic_refcnt) == 0) {
3246 /* We have exclusive access to this iclog. */
3247 bool completed;
3248
3249 if (xlog_force_and_check_iclog(iclog, &completed))
3250 goto out_error;
3251
3252 if (completed)
3253 goto out_unlock;
3254 } else {
3255 /*
3256 * Someone else is still writing to this iclog, so we
3257 * need to ensure that when they release the iclog it
3258 * gets synced immediately as we may be waiting on it.
3259 */
3260 xlog_state_switch_iclogs(log, iclog, 0);
3261 }
3262 }
3263
3264 /*
3265 * The iclog we are about to wait on may contain the checkpoint pushed
3266 * by the above xlog_cil_force() call, but it may not have been pushed
3267 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3268 * are flushed when this iclog is written.
3269 */
3270 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3271 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3272
3273 if (flags & XFS_LOG_SYNC)
3274 return xlog_wait_on_iclog(iclog);
3275out_unlock:
3276 spin_unlock(&log->l_icloglock);
3277 return 0;
3278out_error:
3279 spin_unlock(&log->l_icloglock);
3280 return -EIO;
3281}
3282
3283/*
3284 * Force the log to a specific LSN.
3285 *
3286 * If an iclog with that lsn can be found:
3287 * If it is in the DIRTY state, just return.
3288 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3289 * state and go to sleep or return.
3290 * If it is in any other state, go to sleep or return.
3291 *
3292 * Synchronous forces are implemented with a wait queue. All callers trying
3293 * to force a given lsn to disk must wait on the queue attached to the
3294 * specific in-core log. When given in-core log finally completes its write
3295 * to disk, that thread will wake up all threads waiting on the queue.
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 if (xlog_is_shutdown(log))
3310 goto out_error;
3311
3312 iclog = log->l_iclog;
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 log to a specific checkpoint sequence.
3383 *
3384 * First force the CIL so that all the required changes have been flushed to the
3385 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3386 * the iclog that needs to be flushed to stable storage. If the caller needs
3387 * a synchronous log force, we will wait on the iclog with the LSN returned by
3388 * xlog_cil_force_seq() to be completed.
3389 */
3390int
3391xfs_log_force_seq(
3392 struct xfs_mount *mp,
3393 xfs_csn_t seq,
3394 uint flags,
3395 int *log_flushed)
3396{
3397 struct xlog *log = mp->m_log;
3398 xfs_lsn_t lsn;
3399 int ret;
3400 ASSERT(seq != 0);
3401
3402 XFS_STATS_INC(mp, xs_log_force);
3403 trace_xfs_log_force(mp, seq, _RET_IP_);
3404
3405 lsn = xlog_cil_force_seq(log, seq);
3406 if (lsn == NULLCOMMITLSN)
3407 return 0;
3408
3409 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3410 if (ret == -EAGAIN) {
3411 XFS_STATS_INC(mp, xs_log_force_sleep);
3412 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3413 }
3414 return ret;
3415}
3416
3417/*
3418 * Free a used ticket when its refcount falls to zero.
3419 */
3420void
3421xfs_log_ticket_put(
3422 xlog_ticket_t *ticket)
3423{
3424 ASSERT(atomic_read(&ticket->t_ref) > 0);
3425 if (atomic_dec_and_test(&ticket->t_ref))
3426 kmem_cache_free(xfs_log_ticket_cache, ticket);
3427}
3428
3429xlog_ticket_t *
3430xfs_log_ticket_get(
3431 xlog_ticket_t *ticket)
3432{
3433 ASSERT(atomic_read(&ticket->t_ref) > 0);
3434 atomic_inc(&ticket->t_ref);
3435 return ticket;
3436}
3437
3438/*
3439 * Figure out the total log space unit (in bytes) that would be
3440 * required for a log ticket.
3441 */
3442static int
3443xlog_calc_unit_res(
3444 struct xlog *log,
3445 int unit_bytes,
3446 int *niclogs)
3447{
3448 int iclog_space;
3449 uint num_headers;
3450
3451 /*
3452 * Permanent reservations have up to 'cnt'-1 active log operations
3453 * in the log. A unit in this case is the amount of space for one
3454 * of these log operations. Normal reservations have a cnt of 1
3455 * and their unit amount is the total amount of space required.
3456 *
3457 * The following lines of code account for non-transaction data
3458 * which occupy space in the on-disk log.
3459 *
3460 * Normal form of a transaction is:
3461 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3462 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3463 *
3464 * We need to account for all the leadup data and trailer data
3465 * around the transaction data.
3466 * And then we need to account for the worst case in terms of using
3467 * more space.
3468 * The worst case will happen if:
3469 * - the placement of the transaction happens to be such that the
3470 * roundoff is at its maximum
3471 * - the transaction data is synced before the commit record is synced
3472 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3473 * Therefore the commit record is in its own Log Record.
3474 * This can happen as the commit record is called with its
3475 * own region to xlog_write().
3476 * This then means that in the worst case, roundoff can happen for
3477 * the commit-rec as well.
3478 * The commit-rec is smaller than padding in this scenario and so it is
3479 * not added separately.
3480 */
3481
3482 /* for trans header */
3483 unit_bytes += sizeof(xlog_op_header_t);
3484 unit_bytes += sizeof(xfs_trans_header_t);
3485
3486 /* for start-rec */
3487 unit_bytes += sizeof(xlog_op_header_t);
3488
3489 /*
3490 * for LR headers - the space for data in an iclog is the size minus
3491 * the space used for the headers. If we use the iclog size, then we
3492 * undercalculate the number of headers required.
3493 *
3494 * Furthermore - the addition of op headers for split-recs might
3495 * increase the space required enough to require more log and op
3496 * headers, so take that into account too.
3497 *
3498 * IMPORTANT: This reservation makes the assumption that if this
3499 * transaction is the first in an iclog and hence has the LR headers
3500 * accounted to it, then the remaining space in the iclog is
3501 * exclusively for this transaction. i.e. if the transaction is larger
3502 * than the iclog, it will be the only thing in that iclog.
3503 * Fundamentally, this means we must pass the entire log vector to
3504 * xlog_write to guarantee this.
3505 */
3506 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3507 num_headers = howmany(unit_bytes, iclog_space);
3508
3509 /* for split-recs - ophdrs added when data split over LRs */
3510 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3511
3512 /* add extra header reservations if we overrun */
3513 while (!num_headers ||
3514 howmany(unit_bytes, iclog_space) > num_headers) {
3515 unit_bytes += sizeof(xlog_op_header_t);
3516 num_headers++;
3517 }
3518 unit_bytes += log->l_iclog_hsize * num_headers;
3519
3520 /* for commit-rec LR header - note: padding will subsume the ophdr */
3521 unit_bytes += log->l_iclog_hsize;
3522
3523 /* roundoff padding for transaction data and one for commit record */
3524 unit_bytes += 2 * log->l_iclog_roundoff;
3525
3526 if (niclogs)
3527 *niclogs = num_headers;
3528 return unit_bytes;
3529}
3530
3531int
3532xfs_log_calc_unit_res(
3533 struct xfs_mount *mp,
3534 int unit_bytes)
3535{
3536 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3537}
3538
3539/*
3540 * Allocate and initialise a new log ticket.
3541 */
3542struct xlog_ticket *
3543xlog_ticket_alloc(
3544 struct xlog *log,
3545 int unit_bytes,
3546 int cnt,
3547 bool permanent)
3548{
3549 struct xlog_ticket *tic;
3550 int unit_res;
3551
3552 tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3553
3554 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3555
3556 atomic_set(&tic->t_ref, 1);
3557 tic->t_task = current;
3558 INIT_LIST_HEAD(&tic->t_queue);
3559 tic->t_unit_res = unit_res;
3560 tic->t_curr_res = unit_res;
3561 tic->t_cnt = cnt;
3562 tic->t_ocnt = cnt;
3563 tic->t_tid = get_random_u32();
3564 if (permanent)
3565 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3566
3567 return tic;
3568}
3569
3570#if defined(DEBUG)
3571/*
3572 * Check to make sure the grant write head didn't just over lap the tail. If
3573 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3574 * the cycles differ by exactly one and check the byte count.
3575 *
3576 * This check is run unlocked, so can give false positives. Rather than assert
3577 * on failures, use a warn-once flag and a panic tag to allow the admin to
3578 * determine if they want to panic the machine when such an error occurs. For
3579 * debug kernels this will have the same effect as using an assert but, unlinke
3580 * an assert, it can be turned off at runtime.
3581 */
3582STATIC void
3583xlog_verify_grant_tail(
3584 struct xlog *log)
3585{
3586 int tail_cycle, tail_blocks;
3587 int cycle, space;
3588
3589 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3590 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3591 if (tail_cycle != cycle) {
3592 if (cycle - 1 != tail_cycle &&
3593 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3594 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3595 "%s: cycle - 1 != tail_cycle", __func__);
3596 }
3597
3598 if (space > BBTOB(tail_blocks) &&
3599 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3600 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3601 "%s: space > BBTOB(tail_blocks)", __func__);
3602 }
3603 }
3604}
3605
3606/* check if it will fit */
3607STATIC void
3608xlog_verify_tail_lsn(
3609 struct xlog *log,
3610 struct xlog_in_core *iclog)
3611{
3612 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3613 int blocks;
3614
3615 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3616 blocks =
3617 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3618 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3619 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3620 } else {
3621 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3622
3623 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3624 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3625
3626 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3627 if (blocks < BTOBB(iclog->ic_offset) + 1)
3628 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3629 }
3630}
3631
3632/*
3633 * Perform a number of checks on the iclog before writing to disk.
3634 *
3635 * 1. Make sure the iclogs are still circular
3636 * 2. Make sure we have a good magic number
3637 * 3. Make sure we don't have magic numbers in the data
3638 * 4. Check fields of each log operation header for:
3639 * A. Valid client identifier
3640 * B. tid ptr value falls in valid ptr space (user space code)
3641 * C. Length in log record header is correct according to the
3642 * individual operation headers within record.
3643 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3644 * log, check the preceding blocks of the physical log to make sure all
3645 * the cycle numbers agree with the current cycle number.
3646 */
3647STATIC void
3648xlog_verify_iclog(
3649 struct xlog *log,
3650 struct xlog_in_core *iclog,
3651 int count)
3652{
3653 xlog_op_header_t *ophead;
3654 xlog_in_core_t *icptr;
3655 xlog_in_core_2_t *xhdr;
3656 void *base_ptr, *ptr, *p;
3657 ptrdiff_t field_offset;
3658 uint8_t clientid;
3659 int len, i, j, k, op_len;
3660 int idx;
3661
3662 /* check validity of iclog pointers */
3663 spin_lock(&log->l_icloglock);
3664 icptr = log->l_iclog;
3665 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3666 ASSERT(icptr);
3667
3668 if (icptr != log->l_iclog)
3669 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3670 spin_unlock(&log->l_icloglock);
3671
3672 /* check log magic numbers */
3673 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3674 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3675
3676 base_ptr = ptr = &iclog->ic_header;
3677 p = &iclog->ic_header;
3678 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3679 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3680 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3681 __func__);
3682 }
3683
3684 /* check fields */
3685 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3686 base_ptr = ptr = iclog->ic_datap;
3687 ophead = ptr;
3688 xhdr = iclog->ic_data;
3689 for (i = 0; i < len; i++) {
3690 ophead = ptr;
3691
3692 /* clientid is only 1 byte */
3693 p = &ophead->oh_clientid;
3694 field_offset = p - base_ptr;
3695 if (field_offset & 0x1ff) {
3696 clientid = ophead->oh_clientid;
3697 } else {
3698 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3699 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3700 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3701 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3702 clientid = xlog_get_client_id(
3703 xhdr[j].hic_xheader.xh_cycle_data[k]);
3704 } else {
3705 clientid = xlog_get_client_id(
3706 iclog->ic_header.h_cycle_data[idx]);
3707 }
3708 }
3709 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3710 xfs_warn(log->l_mp,
3711 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3712 __func__, i, clientid, ophead,
3713 (unsigned long)field_offset);
3714 }
3715
3716 /* check length */
3717 p = &ophead->oh_len;
3718 field_offset = p - base_ptr;
3719 if (field_offset & 0x1ff) {
3720 op_len = be32_to_cpu(ophead->oh_len);
3721 } else {
3722 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3723 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3724 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3725 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3726 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3727 } else {
3728 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3729 }
3730 }
3731 ptr += sizeof(xlog_op_header_t) + op_len;
3732 }
3733}
3734#endif
3735
3736/*
3737 * Perform a forced shutdown on the log.
3738 *
3739 * This can be called from low level log code to trigger a shutdown, or from the
3740 * high level mount shutdown code when the mount shuts down.
3741 *
3742 * Our main objectives here are to make sure that:
3743 * a. if the shutdown was not due to a log IO error, flush the logs to
3744 * disk. Anything modified after this is ignored.
3745 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3746 * parties to find out. Nothing new gets queued after this is done.
3747 * c. Tasks sleeping on log reservations, pinned objects and
3748 * other resources get woken up.
3749 * d. The mount is also marked as shut down so that log triggered shutdowns
3750 * still behave the same as if they called xfs_forced_shutdown().
3751 *
3752 * Return true if the shutdown cause was a log IO error and we actually shut the
3753 * log down.
3754 */
3755bool
3756xlog_force_shutdown(
3757 struct xlog *log,
3758 uint32_t shutdown_flags)
3759{
3760 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3761
3762 if (!log)
3763 return false;
3764
3765 /*
3766 * Flush all the completed transactions to disk before marking the log
3767 * being shut down. We need to do this first as shutting down the log
3768 * before the force will prevent the log force from flushing the iclogs
3769 * to disk.
3770 *
3771 * When we are in recovery, there are no transactions to flush, and
3772 * we don't want to touch the log because we don't want to perturb the
3773 * current head/tail for future recovery attempts. Hence we need to
3774 * avoid a log force in this case.
3775 *
3776 * If we are shutting down due to a log IO error, then we must avoid
3777 * trying to write the log as that may just result in more IO errors and
3778 * an endless shutdown/force loop.
3779 */
3780 if (!log_error && !xlog_in_recovery(log))
3781 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3782
3783 /*
3784 * Atomically set the shutdown state. If the shutdown state is already
3785 * set, there someone else is performing the shutdown and so we are done
3786 * here. This should never happen because we should only ever get called
3787 * once by the first shutdown caller.
3788 *
3789 * Much of the log state machine transitions assume that shutdown state
3790 * cannot change once they hold the log->l_icloglock. Hence we need to
3791 * hold that lock here, even though we use the atomic test_and_set_bit()
3792 * operation to set the shutdown state.
3793 */
3794 spin_lock(&log->l_icloglock);
3795 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3796 spin_unlock(&log->l_icloglock);
3797 return false;
3798 }
3799 spin_unlock(&log->l_icloglock);
3800
3801 /*
3802 * If this log shutdown also sets the mount shutdown state, issue a
3803 * shutdown warning message.
3804 */
3805 if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3806 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3807"Filesystem has been shut down due to log error (0x%x).",
3808 shutdown_flags);
3809 xfs_alert(log->l_mp,
3810"Please unmount the filesystem and rectify the problem(s).");
3811 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3812 xfs_stack_trace();
3813 }
3814
3815 /*
3816 * We don't want anybody waiting for log reservations after this. That
3817 * means we have to wake up everybody queued up on reserveq as well as
3818 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3819 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3820 * action is protected by the grant locks.
3821 */
3822 xlog_grant_head_wake_all(&log->l_reserve_head);
3823 xlog_grant_head_wake_all(&log->l_write_head);
3824
3825 /*
3826 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3827 * as if the log writes were completed. The abort handling in the log
3828 * item committed callback functions will do this again under lock to
3829 * avoid races.
3830 */
3831 spin_lock(&log->l_cilp->xc_push_lock);
3832 wake_up_all(&log->l_cilp->xc_start_wait);
3833 wake_up_all(&log->l_cilp->xc_commit_wait);
3834 spin_unlock(&log->l_cilp->xc_push_lock);
3835
3836 spin_lock(&log->l_icloglock);
3837 xlog_state_shutdown_callbacks(log);
3838 spin_unlock(&log->l_icloglock);
3839
3840 wake_up_var(&log->l_opstate);
3841 return log_error;
3842}
3843
3844STATIC int
3845xlog_iclogs_empty(
3846 struct xlog *log)
3847{
3848 xlog_in_core_t *iclog;
3849
3850 iclog = log->l_iclog;
3851 do {
3852 /* endianness does not matter here, zero is zero in
3853 * any language.
3854 */
3855 if (iclog->ic_header.h_num_logops)
3856 return 0;
3857 iclog = iclog->ic_next;
3858 } while (iclog != log->l_iclog);
3859 return 1;
3860}
3861
3862/*
3863 * Verify that an LSN stamped into a piece of metadata is valid. This is
3864 * intended for use in read verifiers on v5 superblocks.
3865 */
3866bool
3867xfs_log_check_lsn(
3868 struct xfs_mount *mp,
3869 xfs_lsn_t lsn)
3870{
3871 struct xlog *log = mp->m_log;
3872 bool valid;
3873
3874 /*
3875 * norecovery mode skips mount-time log processing and unconditionally
3876 * resets the in-core LSN. We can't validate in this mode, but
3877 * modifications are not allowed anyways so just return true.
3878 */
3879 if (xfs_has_norecovery(mp))
3880 return true;
3881
3882 /*
3883 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3884 * handled by recovery and thus safe to ignore here.
3885 */
3886 if (lsn == NULLCOMMITLSN)
3887 return true;
3888
3889 valid = xlog_valid_lsn(mp->m_log, lsn);
3890
3891 /* warn the user about what's gone wrong before verifier failure */
3892 if (!valid) {
3893 spin_lock(&log->l_icloglock);
3894 xfs_warn(mp,
3895"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3896"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3897 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3898 log->l_curr_cycle, log->l_curr_block);
3899 spin_unlock(&log->l_icloglock);
3900 }
3901
3902 return valid;
3903}
3904
3905/*
3906 * Notify the log that we're about to start using a feature that is protected
3907 * by a log incompat feature flag. This will prevent log covering from
3908 * clearing those flags.
3909 */
3910void
3911xlog_use_incompat_feat(
3912 struct xlog *log)
3913{
3914 down_read(&log->l_incompat_users);
3915}
3916
3917/* Notify the log that we've finished using log incompat features. */
3918void
3919xlog_drop_incompat_feat(
3920 struct xlog *log)
3921{
3922 up_read(&log->l_incompat_users);
3923}
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_shared.h"
21#include "xfs_format.h"
22#include "xfs_log_format.h"
23#include "xfs_trans_resv.h"
24#include "xfs_mount.h"
25#include "xfs_error.h"
26#include "xfs_trans.h"
27#include "xfs_trans_priv.h"
28#include "xfs_log.h"
29#include "xfs_log_priv.h"
30#include "xfs_log_recover.h"
31#include "xfs_inode.h"
32#include "xfs_trace.h"
33#include "xfs_fsops.h"
34#include "xfs_cksum.h"
35#include "xfs_sysfs.h"
36#include "xfs_sb.h"
37
38kmem_zone_t *xfs_log_ticket_zone;
39
40/* Local miscellaneous function prototypes */
41STATIC int
42xlog_commit_record(
43 struct xlog *log,
44 struct xlog_ticket *ticket,
45 struct xlog_in_core **iclog,
46 xfs_lsn_t *commitlsnp);
47
48STATIC struct xlog *
49xlog_alloc_log(
50 struct xfs_mount *mp,
51 struct xfs_buftarg *log_target,
52 xfs_daddr_t blk_offset,
53 int num_bblks);
54STATIC int
55xlog_space_left(
56 struct xlog *log,
57 atomic64_t *head);
58STATIC int
59xlog_sync(
60 struct xlog *log,
61 struct xlog_in_core *iclog);
62STATIC void
63xlog_dealloc_log(
64 struct xlog *log);
65
66/* local state machine functions */
67STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
68STATIC void
69xlog_state_do_callback(
70 struct xlog *log,
71 int aborted,
72 struct xlog_in_core *iclog);
73STATIC int
74xlog_state_get_iclog_space(
75 struct xlog *log,
76 int len,
77 struct xlog_in_core **iclog,
78 struct xlog_ticket *ticket,
79 int *continued_write,
80 int *logoffsetp);
81STATIC int
82xlog_state_release_iclog(
83 struct xlog *log,
84 struct xlog_in_core *iclog);
85STATIC void
86xlog_state_switch_iclogs(
87 struct xlog *log,
88 struct xlog_in_core *iclog,
89 int eventual_size);
90STATIC void
91xlog_state_want_sync(
92 struct xlog *log,
93 struct xlog_in_core *iclog);
94
95STATIC void
96xlog_grant_push_ail(
97 struct xlog *log,
98 int need_bytes);
99STATIC void
100xlog_regrant_reserve_log_space(
101 struct xlog *log,
102 struct xlog_ticket *ticket);
103STATIC void
104xlog_ungrant_log_space(
105 struct xlog *log,
106 struct xlog_ticket *ticket);
107
108#if defined(DEBUG)
109STATIC void
110xlog_verify_dest_ptr(
111 struct xlog *log,
112 void *ptr);
113STATIC void
114xlog_verify_grant_tail(
115 struct xlog *log);
116STATIC void
117xlog_verify_iclog(
118 struct xlog *log,
119 struct xlog_in_core *iclog,
120 int count,
121 bool syncing);
122STATIC void
123xlog_verify_tail_lsn(
124 struct xlog *log,
125 struct xlog_in_core *iclog,
126 xfs_lsn_t tail_lsn);
127#else
128#define xlog_verify_dest_ptr(a,b)
129#define xlog_verify_grant_tail(a)
130#define xlog_verify_iclog(a,b,c,d)
131#define xlog_verify_tail_lsn(a,b,c)
132#endif
133
134STATIC int
135xlog_iclogs_empty(
136 struct xlog *log);
137
138static void
139xlog_grant_sub_space(
140 struct xlog *log,
141 atomic64_t *head,
142 int bytes)
143{
144 int64_t head_val = atomic64_read(head);
145 int64_t new, old;
146
147 do {
148 int cycle, space;
149
150 xlog_crack_grant_head_val(head_val, &cycle, &space);
151
152 space -= bytes;
153 if (space < 0) {
154 space += log->l_logsize;
155 cycle--;
156 }
157
158 old = head_val;
159 new = xlog_assign_grant_head_val(cycle, space);
160 head_val = atomic64_cmpxchg(head, old, new);
161 } while (head_val != old);
162}
163
164static void
165xlog_grant_add_space(
166 struct xlog *log,
167 atomic64_t *head,
168 int bytes)
169{
170 int64_t head_val = atomic64_read(head);
171 int64_t new, old;
172
173 do {
174 int tmp;
175 int cycle, space;
176
177 xlog_crack_grant_head_val(head_val, &cycle, &space);
178
179 tmp = log->l_logsize - space;
180 if (tmp > bytes)
181 space += bytes;
182 else {
183 space = bytes - tmp;
184 cycle++;
185 }
186
187 old = head_val;
188 new = xlog_assign_grant_head_val(cycle, space);
189 head_val = atomic64_cmpxchg(head, old, new);
190 } while (head_val != old);
191}
192
193STATIC void
194xlog_grant_head_init(
195 struct xlog_grant_head *head)
196{
197 xlog_assign_grant_head(&head->grant, 1, 0);
198 INIT_LIST_HEAD(&head->waiters);
199 spin_lock_init(&head->lock);
200}
201
202STATIC void
203xlog_grant_head_wake_all(
204 struct xlog_grant_head *head)
205{
206 struct xlog_ticket *tic;
207
208 spin_lock(&head->lock);
209 list_for_each_entry(tic, &head->waiters, t_queue)
210 wake_up_process(tic->t_task);
211 spin_unlock(&head->lock);
212}
213
214static inline int
215xlog_ticket_reservation(
216 struct xlog *log,
217 struct xlog_grant_head *head,
218 struct xlog_ticket *tic)
219{
220 if (head == &log->l_write_head) {
221 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
222 return tic->t_unit_res;
223 } else {
224 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
225 return tic->t_unit_res * tic->t_cnt;
226 else
227 return tic->t_unit_res;
228 }
229}
230
231STATIC bool
232xlog_grant_head_wake(
233 struct xlog *log,
234 struct xlog_grant_head *head,
235 int *free_bytes)
236{
237 struct xlog_ticket *tic;
238 int need_bytes;
239
240 list_for_each_entry(tic, &head->waiters, t_queue) {
241 need_bytes = xlog_ticket_reservation(log, head, tic);
242 if (*free_bytes < need_bytes)
243 return false;
244
245 *free_bytes -= need_bytes;
246 trace_xfs_log_grant_wake_up(log, tic);
247 wake_up_process(tic->t_task);
248 }
249
250 return true;
251}
252
253STATIC int
254xlog_grant_head_wait(
255 struct xlog *log,
256 struct xlog_grant_head *head,
257 struct xlog_ticket *tic,
258 int need_bytes) __releases(&head->lock)
259 __acquires(&head->lock)
260{
261 list_add_tail(&tic->t_queue, &head->waiters);
262
263 do {
264 if (XLOG_FORCED_SHUTDOWN(log))
265 goto shutdown;
266 xlog_grant_push_ail(log, need_bytes);
267
268 __set_current_state(TASK_UNINTERRUPTIBLE);
269 spin_unlock(&head->lock);
270
271 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
272
273 trace_xfs_log_grant_sleep(log, tic);
274 schedule();
275 trace_xfs_log_grant_wake(log, tic);
276
277 spin_lock(&head->lock);
278 if (XLOG_FORCED_SHUTDOWN(log))
279 goto shutdown;
280 } while (xlog_space_left(log, &head->grant) < need_bytes);
281
282 list_del_init(&tic->t_queue);
283 return 0;
284shutdown:
285 list_del_init(&tic->t_queue);
286 return -EIO;
287}
288
289/*
290 * Atomically get the log space required for a log ticket.
291 *
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
294 *
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
298 * every pass.
299 *
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
305 */
306STATIC int
307xlog_grant_head_check(
308 struct xlog *log,
309 struct xlog_grant_head *head,
310 struct xlog_ticket *tic,
311 int *need_bytes)
312{
313 int free_bytes;
314 int error = 0;
315
316 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
317
318 /*
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
323 */
324 *need_bytes = xlog_ticket_reservation(log, head, tic);
325 free_bytes = xlog_space_left(log, &head->grant);
326 if (!list_empty_careful(&head->waiters)) {
327 spin_lock(&head->lock);
328 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
329 free_bytes < *need_bytes) {
330 error = xlog_grant_head_wait(log, head, tic,
331 *need_bytes);
332 }
333 spin_unlock(&head->lock);
334 } else if (free_bytes < *need_bytes) {
335 spin_lock(&head->lock);
336 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
337 spin_unlock(&head->lock);
338 }
339
340 return error;
341}
342
343static void
344xlog_tic_reset_res(xlog_ticket_t *tic)
345{
346 tic->t_res_num = 0;
347 tic->t_res_arr_sum = 0;
348 tic->t_res_num_ophdrs = 0;
349}
350
351static void
352xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
353{
354 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
355 /* add to overflow and start again */
356 tic->t_res_o_flow += tic->t_res_arr_sum;
357 tic->t_res_num = 0;
358 tic->t_res_arr_sum = 0;
359 }
360
361 tic->t_res_arr[tic->t_res_num].r_len = len;
362 tic->t_res_arr[tic->t_res_num].r_type = type;
363 tic->t_res_arr_sum += len;
364 tic->t_res_num++;
365}
366
367/*
368 * Replenish the byte reservation required by moving the grant write head.
369 */
370int
371xfs_log_regrant(
372 struct xfs_mount *mp,
373 struct xlog_ticket *tic)
374{
375 struct xlog *log = mp->m_log;
376 int need_bytes;
377 int error = 0;
378
379 if (XLOG_FORCED_SHUTDOWN(log))
380 return -EIO;
381
382 XFS_STATS_INC(mp, xs_try_logspace);
383
384 /*
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
389 */
390 tic->t_tid++;
391
392 xlog_grant_push_ail(log, tic->t_unit_res);
393
394 tic->t_curr_res = tic->t_unit_res;
395 xlog_tic_reset_res(tic);
396
397 if (tic->t_cnt > 0)
398 return 0;
399
400 trace_xfs_log_regrant(log, tic);
401
402 error = xlog_grant_head_check(log, &log->l_write_head, tic,
403 &need_bytes);
404 if (error)
405 goto out_error;
406
407 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
408 trace_xfs_log_regrant_exit(log, tic);
409 xlog_verify_grant_tail(log);
410 return 0;
411
412out_error:
413 /*
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
417 */
418 tic->t_curr_res = 0;
419 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
420 return error;
421}
422
423/*
424 * Reserve log space and return a ticket corresponding the reservation.
425 *
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
430 */
431int
432xfs_log_reserve(
433 struct xfs_mount *mp,
434 int unit_bytes,
435 int cnt,
436 struct xlog_ticket **ticp,
437 __uint8_t client,
438 bool permanent,
439 uint t_type)
440{
441 struct xlog *log = mp->m_log;
442 struct xlog_ticket *tic;
443 int need_bytes;
444 int error = 0;
445
446 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
447
448 if (XLOG_FORCED_SHUTDOWN(log))
449 return -EIO;
450
451 XFS_STATS_INC(mp, xs_try_logspace);
452
453 ASSERT(*ticp == NULL);
454 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
455 KM_SLEEP | KM_MAYFAIL);
456 if (!tic)
457 return -ENOMEM;
458
459 tic->t_trans_type = t_type;
460 *ticp = tic;
461
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
463 : tic->t_unit_res);
464
465 trace_xfs_log_reserve(log, tic);
466
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
468 &need_bytes);
469 if (error)
470 goto out_error;
471
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
476 return 0;
477
478out_error:
479 /*
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
483 */
484 tic->t_curr_res = 0;
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
486 return error;
487}
488
489
490/*
491 * NOTES:
492 *
493 * 1. currblock field gets updated at startup and after in-core logs
494 * marked as with WANT_SYNC.
495 */
496
497/*
498 * This routine is called when a user of a log manager ticket is done with
499 * the reservation. If the ticket was ever used, then a commit record for
500 * the associated transaction is written out as a log operation header with
501 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
502 * a given ticket. If the ticket was one with a permanent reservation, then
503 * a few operations are done differently. Permanent reservation tickets by
504 * default don't release the reservation. They just commit the current
505 * transaction with the belief that the reservation is still needed. A flag
506 * must be passed in before permanent reservations are actually released.
507 * When these type of tickets are not released, they need to be set into
508 * the inited state again. By doing this, a start record will be written
509 * out when the next write occurs.
510 */
511xfs_lsn_t
512xfs_log_done(
513 struct xfs_mount *mp,
514 struct xlog_ticket *ticket,
515 struct xlog_in_core **iclog,
516 bool regrant)
517{
518 struct xlog *log = mp->m_log;
519 xfs_lsn_t lsn = 0;
520
521 if (XLOG_FORCED_SHUTDOWN(log) ||
522 /*
523 * If nothing was ever written, don't write out commit record.
524 * If we get an error, just continue and give back the log ticket.
525 */
526 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
527 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
528 lsn = (xfs_lsn_t) -1;
529 regrant = false;
530 }
531
532
533 if (!regrant) {
534 trace_xfs_log_done_nonperm(log, ticket);
535
536 /*
537 * Release ticket if not permanent reservation or a specific
538 * request has been made to release a permanent reservation.
539 */
540 xlog_ungrant_log_space(log, ticket);
541 } else {
542 trace_xfs_log_done_perm(log, ticket);
543
544 xlog_regrant_reserve_log_space(log, ticket);
545 /* If this ticket was a permanent reservation and we aren't
546 * trying to release it, reset the inited flags; so next time
547 * we write, a start record will be written out.
548 */
549 ticket->t_flags |= XLOG_TIC_INITED;
550 }
551
552 xfs_log_ticket_put(ticket);
553 return lsn;
554}
555
556/*
557 * Attaches a new iclog I/O completion callback routine during
558 * transaction commit. If the log is in error state, a non-zero
559 * return code is handed back and the caller is responsible for
560 * executing the callback at an appropriate time.
561 */
562int
563xfs_log_notify(
564 struct xfs_mount *mp,
565 struct xlog_in_core *iclog,
566 xfs_log_callback_t *cb)
567{
568 int abortflg;
569
570 spin_lock(&iclog->ic_callback_lock);
571 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
572 if (!abortflg) {
573 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
574 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
575 cb->cb_next = NULL;
576 *(iclog->ic_callback_tail) = cb;
577 iclog->ic_callback_tail = &(cb->cb_next);
578 }
579 spin_unlock(&iclog->ic_callback_lock);
580 return abortflg;
581}
582
583int
584xfs_log_release_iclog(
585 struct xfs_mount *mp,
586 struct xlog_in_core *iclog)
587{
588 if (xlog_state_release_iclog(mp->m_log, iclog)) {
589 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
590 return -EIO;
591 }
592
593 return 0;
594}
595
596/*
597 * Mount a log filesystem
598 *
599 * mp - ubiquitous xfs mount point structure
600 * log_target - buftarg of on-disk log device
601 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
602 * num_bblocks - Number of BBSIZE blocks in on-disk log
603 *
604 * Return error or zero.
605 */
606int
607xfs_log_mount(
608 xfs_mount_t *mp,
609 xfs_buftarg_t *log_target,
610 xfs_daddr_t blk_offset,
611 int num_bblks)
612{
613 int error = 0;
614 int min_logfsbs;
615
616 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
617 xfs_notice(mp, "Mounting V%d Filesystem",
618 XFS_SB_VERSION_NUM(&mp->m_sb));
619 } else {
620 xfs_notice(mp,
621"Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
622 XFS_SB_VERSION_NUM(&mp->m_sb));
623 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
624 }
625
626 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
627 if (IS_ERR(mp->m_log)) {
628 error = PTR_ERR(mp->m_log);
629 goto out;
630 }
631
632 /*
633 * Validate the given log space and drop a critical message via syslog
634 * if the log size is too small that would lead to some unexpected
635 * situations in transaction log space reservation stage.
636 *
637 * Note: we can't just reject the mount if the validation fails. This
638 * would mean that people would have to downgrade their kernel just to
639 * remedy the situation as there is no way to grow the log (short of
640 * black magic surgery with xfs_db).
641 *
642 * We can, however, reject mounts for CRC format filesystems, as the
643 * mkfs binary being used to make the filesystem should never create a
644 * filesystem with a log that is too small.
645 */
646 min_logfsbs = xfs_log_calc_minimum_size(mp);
647
648 if (mp->m_sb.sb_logblocks < min_logfsbs) {
649 xfs_warn(mp,
650 "Log size %d blocks too small, minimum size is %d blocks",
651 mp->m_sb.sb_logblocks, min_logfsbs);
652 error = -EINVAL;
653 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
654 xfs_warn(mp,
655 "Log size %d blocks too large, maximum size is %lld blocks",
656 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
657 error = -EINVAL;
658 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
659 xfs_warn(mp,
660 "log size %lld bytes too large, maximum size is %lld bytes",
661 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
662 XFS_MAX_LOG_BYTES);
663 error = -EINVAL;
664 }
665 if (error) {
666 if (xfs_sb_version_hascrc(&mp->m_sb)) {
667 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
668 ASSERT(0);
669 goto out_free_log;
670 }
671 xfs_crit(mp, "Log size out of supported range.");
672 xfs_crit(mp,
673"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
674 }
675
676 /*
677 * Initialize the AIL now we have a log.
678 */
679 error = xfs_trans_ail_init(mp);
680 if (error) {
681 xfs_warn(mp, "AIL initialisation failed: error %d", error);
682 goto out_free_log;
683 }
684 mp->m_log->l_ailp = mp->m_ail;
685
686 /*
687 * skip log recovery on a norecovery mount. pretend it all
688 * just worked.
689 */
690 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
691 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
692
693 if (readonly)
694 mp->m_flags &= ~XFS_MOUNT_RDONLY;
695
696 error = xlog_recover(mp->m_log);
697
698 if (readonly)
699 mp->m_flags |= XFS_MOUNT_RDONLY;
700 if (error) {
701 xfs_warn(mp, "log mount/recovery failed: error %d",
702 error);
703 xlog_recover_cancel(mp->m_log);
704 goto out_destroy_ail;
705 }
706 }
707
708 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
709 "log");
710 if (error)
711 goto out_destroy_ail;
712
713 /* Normal transactions can now occur */
714 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
715
716 /*
717 * Now the log has been fully initialised and we know were our
718 * space grant counters are, we can initialise the permanent ticket
719 * needed for delayed logging to work.
720 */
721 xlog_cil_init_post_recovery(mp->m_log);
722
723 return 0;
724
725out_destroy_ail:
726 xfs_trans_ail_destroy(mp);
727out_free_log:
728 xlog_dealloc_log(mp->m_log);
729out:
730 return error;
731}
732
733/*
734 * Finish the recovery of the file system. This is separate from the
735 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
736 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
737 * here.
738 *
739 * If we finish recovery successfully, start the background log work. If we are
740 * not doing recovery, then we have a RO filesystem and we don't need to start
741 * it.
742 */
743int
744xfs_log_mount_finish(
745 struct xfs_mount *mp)
746{
747 int error = 0;
748
749 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
750 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
751 return 0;
752 }
753
754 error = xlog_recover_finish(mp->m_log);
755 if (!error)
756 xfs_log_work_queue(mp);
757
758 return error;
759}
760
761/*
762 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
763 * the log.
764 */
765int
766xfs_log_mount_cancel(
767 struct xfs_mount *mp)
768{
769 int error;
770
771 error = xlog_recover_cancel(mp->m_log);
772 xfs_log_unmount(mp);
773
774 return error;
775}
776
777/*
778 * Final log writes as part of unmount.
779 *
780 * Mark the filesystem clean as unmount happens. Note that during relocation
781 * this routine needs to be executed as part of source-bag while the
782 * deallocation must not be done until source-end.
783 */
784
785/*
786 * Unmount record used to have a string "Unmount filesystem--" in the
787 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
788 * We just write the magic number now since that particular field isn't
789 * currently architecture converted and "Unmount" is a bit foo.
790 * As far as I know, there weren't any dependencies on the old behaviour.
791 */
792
793int
794xfs_log_unmount_write(xfs_mount_t *mp)
795{
796 struct xlog *log = mp->m_log;
797 xlog_in_core_t *iclog;
798#ifdef DEBUG
799 xlog_in_core_t *first_iclog;
800#endif
801 xlog_ticket_t *tic = NULL;
802 xfs_lsn_t lsn;
803 int error;
804
805 /*
806 * Don't write out unmount record on read-only mounts.
807 * Or, if we are doing a forced umount (typically because of IO errors).
808 */
809 if (mp->m_flags & XFS_MOUNT_RDONLY)
810 return 0;
811
812 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
813 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
814
815#ifdef DEBUG
816 first_iclog = iclog = log->l_iclog;
817 do {
818 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
819 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
820 ASSERT(iclog->ic_offset == 0);
821 }
822 iclog = iclog->ic_next;
823 } while (iclog != first_iclog);
824#endif
825 if (! (XLOG_FORCED_SHUTDOWN(log))) {
826 error = xfs_log_reserve(mp, 600, 1, &tic,
827 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
828 if (!error) {
829 /* the data section must be 32 bit size aligned */
830 struct {
831 __uint16_t magic;
832 __uint16_t pad1;
833 __uint32_t pad2; /* may as well make it 64 bits */
834 } magic = {
835 .magic = XLOG_UNMOUNT_TYPE,
836 };
837 struct xfs_log_iovec reg = {
838 .i_addr = &magic,
839 .i_len = sizeof(magic),
840 .i_type = XLOG_REG_TYPE_UNMOUNT,
841 };
842 struct xfs_log_vec vec = {
843 .lv_niovecs = 1,
844 .lv_iovecp = ®,
845 };
846
847 /* remove inited flag, and account for space used */
848 tic->t_flags = 0;
849 tic->t_curr_res -= sizeof(magic);
850 error = xlog_write(log, &vec, tic, &lsn,
851 NULL, XLOG_UNMOUNT_TRANS);
852 /*
853 * At this point, we're umounting anyway,
854 * so there's no point in transitioning log state
855 * to IOERROR. Just continue...
856 */
857 }
858
859 if (error)
860 xfs_alert(mp, "%s: unmount record failed", __func__);
861
862
863 spin_lock(&log->l_icloglock);
864 iclog = log->l_iclog;
865 atomic_inc(&iclog->ic_refcnt);
866 xlog_state_want_sync(log, iclog);
867 spin_unlock(&log->l_icloglock);
868 error = xlog_state_release_iclog(log, iclog);
869
870 spin_lock(&log->l_icloglock);
871 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
872 iclog->ic_state == XLOG_STATE_DIRTY)) {
873 if (!XLOG_FORCED_SHUTDOWN(log)) {
874 xlog_wait(&iclog->ic_force_wait,
875 &log->l_icloglock);
876 } else {
877 spin_unlock(&log->l_icloglock);
878 }
879 } else {
880 spin_unlock(&log->l_icloglock);
881 }
882 if (tic) {
883 trace_xfs_log_umount_write(log, tic);
884 xlog_ungrant_log_space(log, tic);
885 xfs_log_ticket_put(tic);
886 }
887 } else {
888 /*
889 * We're already in forced_shutdown mode, couldn't
890 * even attempt to write out the unmount transaction.
891 *
892 * Go through the motions of sync'ing and releasing
893 * the iclog, even though no I/O will actually happen,
894 * we need to wait for other log I/Os that may already
895 * be in progress. Do this as a separate section of
896 * code so we'll know if we ever get stuck here that
897 * we're in this odd situation of trying to unmount
898 * a file system that went into forced_shutdown as
899 * the result of an unmount..
900 */
901 spin_lock(&log->l_icloglock);
902 iclog = log->l_iclog;
903 atomic_inc(&iclog->ic_refcnt);
904
905 xlog_state_want_sync(log, iclog);
906 spin_unlock(&log->l_icloglock);
907 error = xlog_state_release_iclog(log, iclog);
908
909 spin_lock(&log->l_icloglock);
910
911 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
912 || iclog->ic_state == XLOG_STATE_DIRTY
913 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
914
915 xlog_wait(&iclog->ic_force_wait,
916 &log->l_icloglock);
917 } else {
918 spin_unlock(&log->l_icloglock);
919 }
920 }
921
922 return error;
923} /* xfs_log_unmount_write */
924
925/*
926 * Empty the log for unmount/freeze.
927 *
928 * To do this, we first need to shut down the background log work so it is not
929 * trying to cover the log as we clean up. We then need to unpin all objects in
930 * the log so we can then flush them out. Once they have completed their IO and
931 * run the callbacks removing themselves from the AIL, we can write the unmount
932 * record.
933 */
934void
935xfs_log_quiesce(
936 struct xfs_mount *mp)
937{
938 cancel_delayed_work_sync(&mp->m_log->l_work);
939 xfs_log_force(mp, XFS_LOG_SYNC);
940
941 /*
942 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
943 * will push it, xfs_wait_buftarg() will not wait for it. Further,
944 * xfs_buf_iowait() cannot be used because it was pushed with the
945 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
946 * the IO to complete.
947 */
948 xfs_ail_push_all_sync(mp->m_ail);
949 xfs_wait_buftarg(mp->m_ddev_targp);
950 xfs_buf_lock(mp->m_sb_bp);
951 xfs_buf_unlock(mp->m_sb_bp);
952
953 xfs_log_unmount_write(mp);
954}
955
956/*
957 * Shut down and release the AIL and Log.
958 *
959 * During unmount, we need to ensure we flush all the dirty metadata objects
960 * from the AIL so that the log is empty before we write the unmount record to
961 * the log. Once this is done, we can tear down the AIL and the log.
962 */
963void
964xfs_log_unmount(
965 struct xfs_mount *mp)
966{
967 xfs_log_quiesce(mp);
968
969 xfs_trans_ail_destroy(mp);
970
971 xfs_sysfs_del(&mp->m_log->l_kobj);
972
973 xlog_dealloc_log(mp->m_log);
974}
975
976void
977xfs_log_item_init(
978 struct xfs_mount *mp,
979 struct xfs_log_item *item,
980 int type,
981 const struct xfs_item_ops *ops)
982{
983 item->li_mountp = mp;
984 item->li_ailp = mp->m_ail;
985 item->li_type = type;
986 item->li_ops = ops;
987 item->li_lv = NULL;
988
989 INIT_LIST_HEAD(&item->li_ail);
990 INIT_LIST_HEAD(&item->li_cil);
991}
992
993/*
994 * Wake up processes waiting for log space after we have moved the log tail.
995 */
996void
997xfs_log_space_wake(
998 struct xfs_mount *mp)
999{
1000 struct xlog *log = mp->m_log;
1001 int free_bytes;
1002
1003 if (XLOG_FORCED_SHUTDOWN(log))
1004 return;
1005
1006 if (!list_empty_careful(&log->l_write_head.waiters)) {
1007 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1008
1009 spin_lock(&log->l_write_head.lock);
1010 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1011 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1012 spin_unlock(&log->l_write_head.lock);
1013 }
1014
1015 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1016 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1017
1018 spin_lock(&log->l_reserve_head.lock);
1019 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1020 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1021 spin_unlock(&log->l_reserve_head.lock);
1022 }
1023}
1024
1025/*
1026 * Determine if we have a transaction that has gone to disk that needs to be
1027 * covered. To begin the transition to the idle state firstly the log needs to
1028 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1029 * we start attempting to cover the log.
1030 *
1031 * Only if we are then in a state where covering is needed, the caller is
1032 * informed that dummy transactions are required to move the log into the idle
1033 * state.
1034 *
1035 * If there are any items in the AIl or CIL, then we do not want to attempt to
1036 * cover the log as we may be in a situation where there isn't log space
1037 * available to run a dummy transaction and this can lead to deadlocks when the
1038 * tail of the log is pinned by an item that is modified in the CIL. Hence
1039 * there's no point in running a dummy transaction at this point because we
1040 * can't start trying to idle the log until both the CIL and AIL are empty.
1041 */
1042int
1043xfs_log_need_covered(xfs_mount_t *mp)
1044{
1045 struct xlog *log = mp->m_log;
1046 int needed = 0;
1047
1048 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1049 return 0;
1050
1051 if (!xlog_cil_empty(log))
1052 return 0;
1053
1054 spin_lock(&log->l_icloglock);
1055 switch (log->l_covered_state) {
1056 case XLOG_STATE_COVER_DONE:
1057 case XLOG_STATE_COVER_DONE2:
1058 case XLOG_STATE_COVER_IDLE:
1059 break;
1060 case XLOG_STATE_COVER_NEED:
1061 case XLOG_STATE_COVER_NEED2:
1062 if (xfs_ail_min_lsn(log->l_ailp))
1063 break;
1064 if (!xlog_iclogs_empty(log))
1065 break;
1066
1067 needed = 1;
1068 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1069 log->l_covered_state = XLOG_STATE_COVER_DONE;
1070 else
1071 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1072 break;
1073 default:
1074 needed = 1;
1075 break;
1076 }
1077 spin_unlock(&log->l_icloglock);
1078 return needed;
1079}
1080
1081/*
1082 * We may be holding the log iclog lock upon entering this routine.
1083 */
1084xfs_lsn_t
1085xlog_assign_tail_lsn_locked(
1086 struct xfs_mount *mp)
1087{
1088 struct xlog *log = mp->m_log;
1089 struct xfs_log_item *lip;
1090 xfs_lsn_t tail_lsn;
1091
1092 assert_spin_locked(&mp->m_ail->xa_lock);
1093
1094 /*
1095 * To make sure we always have a valid LSN for the log tail we keep
1096 * track of the last LSN which was committed in log->l_last_sync_lsn,
1097 * and use that when the AIL was empty.
1098 */
1099 lip = xfs_ail_min(mp->m_ail);
1100 if (lip)
1101 tail_lsn = lip->li_lsn;
1102 else
1103 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1104 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1105 atomic64_set(&log->l_tail_lsn, tail_lsn);
1106 return tail_lsn;
1107}
1108
1109xfs_lsn_t
1110xlog_assign_tail_lsn(
1111 struct xfs_mount *mp)
1112{
1113 xfs_lsn_t tail_lsn;
1114
1115 spin_lock(&mp->m_ail->xa_lock);
1116 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1117 spin_unlock(&mp->m_ail->xa_lock);
1118
1119 return tail_lsn;
1120}
1121
1122/*
1123 * Return the space in the log between the tail and the head. The head
1124 * is passed in the cycle/bytes formal parms. In the special case where
1125 * the reserve head has wrapped passed the tail, this calculation is no
1126 * longer valid. In this case, just return 0 which means there is no space
1127 * in the log. This works for all places where this function is called
1128 * with the reserve head. Of course, if the write head were to ever
1129 * wrap the tail, we should blow up. Rather than catch this case here,
1130 * we depend on other ASSERTions in other parts of the code. XXXmiken
1131 *
1132 * This code also handles the case where the reservation head is behind
1133 * the tail. The details of this case are described below, but the end
1134 * result is that we return the size of the log as the amount of space left.
1135 */
1136STATIC int
1137xlog_space_left(
1138 struct xlog *log,
1139 atomic64_t *head)
1140{
1141 int free_bytes;
1142 int tail_bytes;
1143 int tail_cycle;
1144 int head_cycle;
1145 int head_bytes;
1146
1147 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1148 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1149 tail_bytes = BBTOB(tail_bytes);
1150 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1151 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1152 else if (tail_cycle + 1 < head_cycle)
1153 return 0;
1154 else if (tail_cycle < head_cycle) {
1155 ASSERT(tail_cycle == (head_cycle - 1));
1156 free_bytes = tail_bytes - head_bytes;
1157 } else {
1158 /*
1159 * The reservation head is behind the tail.
1160 * In this case we just want to return the size of the
1161 * log as the amount of space left.
1162 */
1163 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1164 xfs_alert(log->l_mp,
1165 " tail_cycle = %d, tail_bytes = %d",
1166 tail_cycle, tail_bytes);
1167 xfs_alert(log->l_mp,
1168 " GH cycle = %d, GH bytes = %d",
1169 head_cycle, head_bytes);
1170 ASSERT(0);
1171 free_bytes = log->l_logsize;
1172 }
1173 return free_bytes;
1174}
1175
1176
1177/*
1178 * Log function which is called when an io completes.
1179 *
1180 * The log manager needs its own routine, in order to control what
1181 * happens with the buffer after the write completes.
1182 */
1183void
1184xlog_iodone(xfs_buf_t *bp)
1185{
1186 struct xlog_in_core *iclog = bp->b_fspriv;
1187 struct xlog *l = iclog->ic_log;
1188 int aborted = 0;
1189
1190 /*
1191 * Race to shutdown the filesystem if we see an error or the iclog is in
1192 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1193 * CRC errors into log recovery.
1194 */
1195 if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR,
1196 XFS_RANDOM_IODONE_IOERR) ||
1197 iclog->ic_state & XLOG_STATE_IOABORT) {
1198 if (iclog->ic_state & XLOG_STATE_IOABORT)
1199 iclog->ic_state &= ~XLOG_STATE_IOABORT;
1200
1201 xfs_buf_ioerror_alert(bp, __func__);
1202 xfs_buf_stale(bp);
1203 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1204 /*
1205 * This flag will be propagated to the trans-committed
1206 * callback routines to let them know that the log-commit
1207 * didn't succeed.
1208 */
1209 aborted = XFS_LI_ABORTED;
1210 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1211 aborted = XFS_LI_ABORTED;
1212 }
1213
1214 /* log I/O is always issued ASYNC */
1215 ASSERT(bp->b_flags & XBF_ASYNC);
1216 xlog_state_done_syncing(iclog, aborted);
1217
1218 /*
1219 * drop the buffer lock now that we are done. Nothing references
1220 * the buffer after this, so an unmount waiting on this lock can now
1221 * tear it down safely. As such, it is unsafe to reference the buffer
1222 * (bp) after the unlock as we could race with it being freed.
1223 */
1224 xfs_buf_unlock(bp);
1225}
1226
1227/*
1228 * Return size of each in-core log record buffer.
1229 *
1230 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1231 *
1232 * If the filesystem blocksize is too large, we may need to choose a
1233 * larger size since the directory code currently logs entire blocks.
1234 */
1235
1236STATIC void
1237xlog_get_iclog_buffer_size(
1238 struct xfs_mount *mp,
1239 struct xlog *log)
1240{
1241 int size;
1242 int xhdrs;
1243
1244 if (mp->m_logbufs <= 0)
1245 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1246 else
1247 log->l_iclog_bufs = mp->m_logbufs;
1248
1249 /*
1250 * Buffer size passed in from mount system call.
1251 */
1252 if (mp->m_logbsize > 0) {
1253 size = log->l_iclog_size = mp->m_logbsize;
1254 log->l_iclog_size_log = 0;
1255 while (size != 1) {
1256 log->l_iclog_size_log++;
1257 size >>= 1;
1258 }
1259
1260 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1261 /* # headers = size / 32k
1262 * one header holds cycles from 32k of data
1263 */
1264
1265 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1266 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1267 xhdrs++;
1268 log->l_iclog_hsize = xhdrs << BBSHIFT;
1269 log->l_iclog_heads = xhdrs;
1270 } else {
1271 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1272 log->l_iclog_hsize = BBSIZE;
1273 log->l_iclog_heads = 1;
1274 }
1275 goto done;
1276 }
1277
1278 /* All machines use 32kB buffers by default. */
1279 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1280 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1281
1282 /* the default log size is 16k or 32k which is one header sector */
1283 log->l_iclog_hsize = BBSIZE;
1284 log->l_iclog_heads = 1;
1285
1286done:
1287 /* are we being asked to make the sizes selected above visible? */
1288 if (mp->m_logbufs == 0)
1289 mp->m_logbufs = log->l_iclog_bufs;
1290 if (mp->m_logbsize == 0)
1291 mp->m_logbsize = log->l_iclog_size;
1292} /* xlog_get_iclog_buffer_size */
1293
1294
1295void
1296xfs_log_work_queue(
1297 struct xfs_mount *mp)
1298{
1299 queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1300 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1301}
1302
1303/*
1304 * Every sync period we need to unpin all items in the AIL and push them to
1305 * disk. If there is nothing dirty, then we might need to cover the log to
1306 * indicate that the filesystem is idle.
1307 */
1308void
1309xfs_log_worker(
1310 struct work_struct *work)
1311{
1312 struct xlog *log = container_of(to_delayed_work(work),
1313 struct xlog, l_work);
1314 struct xfs_mount *mp = log->l_mp;
1315
1316 /* dgc: errors ignored - not fatal and nowhere to report them */
1317 if (xfs_log_need_covered(mp)) {
1318 /*
1319 * Dump a transaction into the log that contains no real change.
1320 * This is needed to stamp the current tail LSN into the log
1321 * during the covering operation.
1322 *
1323 * We cannot use an inode here for this - that will push dirty
1324 * state back up into the VFS and then periodic inode flushing
1325 * will prevent log covering from making progress. Hence we
1326 * synchronously log the superblock instead to ensure the
1327 * superblock is immediately unpinned and can be written back.
1328 */
1329 xfs_sync_sb(mp, true);
1330 } else
1331 xfs_log_force(mp, 0);
1332
1333 /* start pushing all the metadata that is currently dirty */
1334 xfs_ail_push_all(mp->m_ail);
1335
1336 /* queue us up again */
1337 xfs_log_work_queue(mp);
1338}
1339
1340/*
1341 * This routine initializes some of the log structure for a given mount point.
1342 * Its primary purpose is to fill in enough, so recovery can occur. However,
1343 * some other stuff may be filled in too.
1344 */
1345STATIC struct xlog *
1346xlog_alloc_log(
1347 struct xfs_mount *mp,
1348 struct xfs_buftarg *log_target,
1349 xfs_daddr_t blk_offset,
1350 int num_bblks)
1351{
1352 struct xlog *log;
1353 xlog_rec_header_t *head;
1354 xlog_in_core_t **iclogp;
1355 xlog_in_core_t *iclog, *prev_iclog=NULL;
1356 xfs_buf_t *bp;
1357 int i;
1358 int error = -ENOMEM;
1359 uint log2_size = 0;
1360
1361 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1362 if (!log) {
1363 xfs_warn(mp, "Log allocation failed: No memory!");
1364 goto out;
1365 }
1366
1367 log->l_mp = mp;
1368 log->l_targ = log_target;
1369 log->l_logsize = BBTOB(num_bblks);
1370 log->l_logBBstart = blk_offset;
1371 log->l_logBBsize = num_bblks;
1372 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1373 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1374 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1375
1376 log->l_prev_block = -1;
1377 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1378 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1379 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1380 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1381
1382 xlog_grant_head_init(&log->l_reserve_head);
1383 xlog_grant_head_init(&log->l_write_head);
1384
1385 error = -EFSCORRUPTED;
1386 if (xfs_sb_version_hassector(&mp->m_sb)) {
1387 log2_size = mp->m_sb.sb_logsectlog;
1388 if (log2_size < BBSHIFT) {
1389 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1390 log2_size, BBSHIFT);
1391 goto out_free_log;
1392 }
1393
1394 log2_size -= BBSHIFT;
1395 if (log2_size > mp->m_sectbb_log) {
1396 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1397 log2_size, mp->m_sectbb_log);
1398 goto out_free_log;
1399 }
1400
1401 /* for larger sector sizes, must have v2 or external log */
1402 if (log2_size && log->l_logBBstart > 0 &&
1403 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1404 xfs_warn(mp,
1405 "log sector size (0x%x) invalid for configuration.",
1406 log2_size);
1407 goto out_free_log;
1408 }
1409 }
1410 log->l_sectBBsize = 1 << log2_size;
1411
1412 xlog_get_iclog_buffer_size(mp, log);
1413
1414 /*
1415 * Use a NULL block for the extra log buffer used during splits so that
1416 * it will trigger errors if we ever try to do IO on it without first
1417 * having set it up properly.
1418 */
1419 error = -ENOMEM;
1420 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1421 BTOBB(log->l_iclog_size), 0);
1422 if (!bp)
1423 goto out_free_log;
1424
1425 /*
1426 * The iclogbuf buffer locks are held over IO but we are not going to do
1427 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1428 * when appropriately.
1429 */
1430 ASSERT(xfs_buf_islocked(bp));
1431 xfs_buf_unlock(bp);
1432
1433 /* use high priority wq for log I/O completion */
1434 bp->b_ioend_wq = mp->m_log_workqueue;
1435 bp->b_iodone = xlog_iodone;
1436 log->l_xbuf = bp;
1437
1438 spin_lock_init(&log->l_icloglock);
1439 init_waitqueue_head(&log->l_flush_wait);
1440
1441 iclogp = &log->l_iclog;
1442 /*
1443 * The amount of memory to allocate for the iclog structure is
1444 * rather funky due to the way the structure is defined. It is
1445 * done this way so that we can use different sizes for machines
1446 * with different amounts of memory. See the definition of
1447 * xlog_in_core_t in xfs_log_priv.h for details.
1448 */
1449 ASSERT(log->l_iclog_size >= 4096);
1450 for (i=0; i < log->l_iclog_bufs; i++) {
1451 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1452 if (!*iclogp)
1453 goto out_free_iclog;
1454
1455 iclog = *iclogp;
1456 iclog->ic_prev = prev_iclog;
1457 prev_iclog = iclog;
1458
1459 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1460 BTOBB(log->l_iclog_size), 0);
1461 if (!bp)
1462 goto out_free_iclog;
1463
1464 ASSERT(xfs_buf_islocked(bp));
1465 xfs_buf_unlock(bp);
1466
1467 /* use high priority wq for log I/O completion */
1468 bp->b_ioend_wq = mp->m_log_workqueue;
1469 bp->b_iodone = xlog_iodone;
1470 iclog->ic_bp = bp;
1471 iclog->ic_data = bp->b_addr;
1472#ifdef DEBUG
1473 log->l_iclog_bak[i] = &iclog->ic_header;
1474#endif
1475 head = &iclog->ic_header;
1476 memset(head, 0, sizeof(xlog_rec_header_t));
1477 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1478 head->h_version = cpu_to_be32(
1479 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1480 head->h_size = cpu_to_be32(log->l_iclog_size);
1481 /* new fields */
1482 head->h_fmt = cpu_to_be32(XLOG_FMT);
1483 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1484
1485 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1486 iclog->ic_state = XLOG_STATE_ACTIVE;
1487 iclog->ic_log = log;
1488 atomic_set(&iclog->ic_refcnt, 0);
1489 spin_lock_init(&iclog->ic_callback_lock);
1490 iclog->ic_callback_tail = &(iclog->ic_callback);
1491 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1492
1493 init_waitqueue_head(&iclog->ic_force_wait);
1494 init_waitqueue_head(&iclog->ic_write_wait);
1495
1496 iclogp = &iclog->ic_next;
1497 }
1498 *iclogp = log->l_iclog; /* complete ring */
1499 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1500
1501 error = xlog_cil_init(log);
1502 if (error)
1503 goto out_free_iclog;
1504 return log;
1505
1506out_free_iclog:
1507 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1508 prev_iclog = iclog->ic_next;
1509 if (iclog->ic_bp)
1510 xfs_buf_free(iclog->ic_bp);
1511 kmem_free(iclog);
1512 }
1513 spinlock_destroy(&log->l_icloglock);
1514 xfs_buf_free(log->l_xbuf);
1515out_free_log:
1516 kmem_free(log);
1517out:
1518 return ERR_PTR(error);
1519} /* xlog_alloc_log */
1520
1521
1522/*
1523 * Write out the commit record of a transaction associated with the given
1524 * ticket. Return the lsn of the commit record.
1525 */
1526STATIC int
1527xlog_commit_record(
1528 struct xlog *log,
1529 struct xlog_ticket *ticket,
1530 struct xlog_in_core **iclog,
1531 xfs_lsn_t *commitlsnp)
1532{
1533 struct xfs_mount *mp = log->l_mp;
1534 int error;
1535 struct xfs_log_iovec reg = {
1536 .i_addr = NULL,
1537 .i_len = 0,
1538 .i_type = XLOG_REG_TYPE_COMMIT,
1539 };
1540 struct xfs_log_vec vec = {
1541 .lv_niovecs = 1,
1542 .lv_iovecp = ®,
1543 };
1544
1545 ASSERT_ALWAYS(iclog);
1546 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1547 XLOG_COMMIT_TRANS);
1548 if (error)
1549 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1550 return error;
1551}
1552
1553/*
1554 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1555 * log space. This code pushes on the lsn which would supposedly free up
1556 * the 25% which we want to leave free. We may need to adopt a policy which
1557 * pushes on an lsn which is further along in the log once we reach the high
1558 * water mark. In this manner, we would be creating a low water mark.
1559 */
1560STATIC void
1561xlog_grant_push_ail(
1562 struct xlog *log,
1563 int need_bytes)
1564{
1565 xfs_lsn_t threshold_lsn = 0;
1566 xfs_lsn_t last_sync_lsn;
1567 int free_blocks;
1568 int free_bytes;
1569 int threshold_block;
1570 int threshold_cycle;
1571 int free_threshold;
1572
1573 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1574
1575 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1576 free_blocks = BTOBBT(free_bytes);
1577
1578 /*
1579 * Set the threshold for the minimum number of free blocks in the
1580 * log to the maximum of what the caller needs, one quarter of the
1581 * log, and 256 blocks.
1582 */
1583 free_threshold = BTOBB(need_bytes);
1584 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1585 free_threshold = MAX(free_threshold, 256);
1586 if (free_blocks >= free_threshold)
1587 return;
1588
1589 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1590 &threshold_block);
1591 threshold_block += free_threshold;
1592 if (threshold_block >= log->l_logBBsize) {
1593 threshold_block -= log->l_logBBsize;
1594 threshold_cycle += 1;
1595 }
1596 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1597 threshold_block);
1598 /*
1599 * Don't pass in an lsn greater than the lsn of the last
1600 * log record known to be on disk. Use a snapshot of the last sync lsn
1601 * so that it doesn't change between the compare and the set.
1602 */
1603 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1604 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1605 threshold_lsn = last_sync_lsn;
1606
1607 /*
1608 * Get the transaction layer to kick the dirty buffers out to
1609 * disk asynchronously. No point in trying to do this if
1610 * the filesystem is shutting down.
1611 */
1612 if (!XLOG_FORCED_SHUTDOWN(log))
1613 xfs_ail_push(log->l_ailp, threshold_lsn);
1614}
1615
1616/*
1617 * Stamp cycle number in every block
1618 */
1619STATIC void
1620xlog_pack_data(
1621 struct xlog *log,
1622 struct xlog_in_core *iclog,
1623 int roundoff)
1624{
1625 int i, j, k;
1626 int size = iclog->ic_offset + roundoff;
1627 __be32 cycle_lsn;
1628 char *dp;
1629
1630 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1631
1632 dp = iclog->ic_datap;
1633 for (i = 0; i < BTOBB(size); i++) {
1634 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1635 break;
1636 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1637 *(__be32 *)dp = cycle_lsn;
1638 dp += BBSIZE;
1639 }
1640
1641 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1642 xlog_in_core_2_t *xhdr = iclog->ic_data;
1643
1644 for ( ; i < BTOBB(size); i++) {
1645 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1646 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1647 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1648 *(__be32 *)dp = cycle_lsn;
1649 dp += BBSIZE;
1650 }
1651
1652 for (i = 1; i < log->l_iclog_heads; i++)
1653 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1654 }
1655}
1656
1657/*
1658 * Calculate the checksum for a log buffer.
1659 *
1660 * This is a little more complicated than it should be because the various
1661 * headers and the actual data are non-contiguous.
1662 */
1663__le32
1664xlog_cksum(
1665 struct xlog *log,
1666 struct xlog_rec_header *rhead,
1667 char *dp,
1668 int size)
1669{
1670 __uint32_t crc;
1671
1672 /* first generate the crc for the record header ... */
1673 crc = xfs_start_cksum((char *)rhead,
1674 sizeof(struct xlog_rec_header),
1675 offsetof(struct xlog_rec_header, h_crc));
1676
1677 /* ... then for additional cycle data for v2 logs ... */
1678 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1679 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1680 int i;
1681 int xheads;
1682
1683 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1684 if (size % XLOG_HEADER_CYCLE_SIZE)
1685 xheads++;
1686
1687 for (i = 1; i < xheads; i++) {
1688 crc = crc32c(crc, &xhdr[i].hic_xheader,
1689 sizeof(struct xlog_rec_ext_header));
1690 }
1691 }
1692
1693 /* ... and finally for the payload */
1694 crc = crc32c(crc, dp, size);
1695
1696 return xfs_end_cksum(crc);
1697}
1698
1699/*
1700 * The bdstrat callback function for log bufs. This gives us a central
1701 * place to trap bufs in case we get hit by a log I/O error and need to
1702 * shutdown. Actually, in practice, even when we didn't get a log error,
1703 * we transition the iclogs to IOERROR state *after* flushing all existing
1704 * iclogs to disk. This is because we don't want anymore new transactions to be
1705 * started or completed afterwards.
1706 *
1707 * We lock the iclogbufs here so that we can serialise against IO completion
1708 * during unmount. We might be processing a shutdown triggered during unmount,
1709 * and that can occur asynchronously to the unmount thread, and hence we need to
1710 * ensure that completes before tearing down the iclogbufs. Hence we need to
1711 * hold the buffer lock across the log IO to acheive that.
1712 */
1713STATIC int
1714xlog_bdstrat(
1715 struct xfs_buf *bp)
1716{
1717 struct xlog_in_core *iclog = bp->b_fspriv;
1718
1719 xfs_buf_lock(bp);
1720 if (iclog->ic_state & XLOG_STATE_IOERROR) {
1721 xfs_buf_ioerror(bp, -EIO);
1722 xfs_buf_stale(bp);
1723 xfs_buf_ioend(bp);
1724 /*
1725 * It would seem logical to return EIO here, but we rely on
1726 * the log state machine to propagate I/O errors instead of
1727 * doing it here. Similarly, IO completion will unlock the
1728 * buffer, so we don't do it here.
1729 */
1730 return 0;
1731 }
1732
1733 xfs_buf_submit(bp);
1734 return 0;
1735}
1736
1737/*
1738 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1739 * fashion. Previously, we should have moved the current iclog
1740 * ptr in the log to point to the next available iclog. This allows further
1741 * write to continue while this code syncs out an iclog ready to go.
1742 * Before an in-core log can be written out, the data section must be scanned
1743 * to save away the 1st word of each BBSIZE block into the header. We replace
1744 * it with the current cycle count. Each BBSIZE block is tagged with the
1745 * cycle count because there in an implicit assumption that drives will
1746 * guarantee that entire 512 byte blocks get written at once. In other words,
1747 * we can't have part of a 512 byte block written and part not written. By
1748 * tagging each block, we will know which blocks are valid when recovering
1749 * after an unclean shutdown.
1750 *
1751 * This routine is single threaded on the iclog. No other thread can be in
1752 * this routine with the same iclog. Changing contents of iclog can there-
1753 * fore be done without grabbing the state machine lock. Updating the global
1754 * log will require grabbing the lock though.
1755 *
1756 * The entire log manager uses a logical block numbering scheme. Only
1757 * log_sync (and then only bwrite()) know about the fact that the log may
1758 * not start with block zero on a given device. The log block start offset
1759 * is added immediately before calling bwrite().
1760 */
1761
1762STATIC int
1763xlog_sync(
1764 struct xlog *log,
1765 struct xlog_in_core *iclog)
1766{
1767 xfs_buf_t *bp;
1768 int i;
1769 uint count; /* byte count of bwrite */
1770 uint count_init; /* initial count before roundup */
1771 int roundoff; /* roundoff to BB or stripe */
1772 int split = 0; /* split write into two regions */
1773 int error;
1774 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1775 int size;
1776
1777 XFS_STATS_INC(log->l_mp, xs_log_writes);
1778 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1779
1780 /* Add for LR header */
1781 count_init = log->l_iclog_hsize + iclog->ic_offset;
1782
1783 /* Round out the log write size */
1784 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1785 /* we have a v2 stripe unit to use */
1786 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1787 } else {
1788 count = BBTOB(BTOBB(count_init));
1789 }
1790 roundoff = count - count_init;
1791 ASSERT(roundoff >= 0);
1792 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1793 roundoff < log->l_mp->m_sb.sb_logsunit)
1794 ||
1795 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1796 roundoff < BBTOB(1)));
1797
1798 /* move grant heads by roundoff in sync */
1799 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1800 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1801
1802 /* put cycle number in every block */
1803 xlog_pack_data(log, iclog, roundoff);
1804
1805 /* real byte length */
1806 size = iclog->ic_offset;
1807 if (v2)
1808 size += roundoff;
1809 iclog->ic_header.h_len = cpu_to_be32(size);
1810
1811 bp = iclog->ic_bp;
1812 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1813
1814 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1815
1816 /* Do we need to split this write into 2 parts? */
1817 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1818 char *dptr;
1819
1820 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1821 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1822 iclog->ic_bwritecnt = 2;
1823
1824 /*
1825 * Bump the cycle numbers at the start of each block in the
1826 * part of the iclog that ends up in the buffer that gets
1827 * written to the start of the log.
1828 *
1829 * Watch out for the header magic number case, though.
1830 */
1831 dptr = (char *)&iclog->ic_header + count;
1832 for (i = 0; i < split; i += BBSIZE) {
1833 __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1834 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1835 cycle++;
1836 *(__be32 *)dptr = cpu_to_be32(cycle);
1837
1838 dptr += BBSIZE;
1839 }
1840 } else {
1841 iclog->ic_bwritecnt = 1;
1842 }
1843
1844 /* calculcate the checksum */
1845 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1846 iclog->ic_datap, size);
1847#ifdef DEBUG
1848 /*
1849 * Intentionally corrupt the log record CRC based on the error injection
1850 * frequency, if defined. This facilitates testing log recovery in the
1851 * event of torn writes. Hence, set the IOABORT state to abort the log
1852 * write on I/O completion and shutdown the fs. The subsequent mount
1853 * detects the bad CRC and attempts to recover.
1854 */
1855 if (log->l_badcrc_factor &&
1856 (prandom_u32() % log->l_badcrc_factor == 0)) {
1857 iclog->ic_header.h_crc &= 0xAAAAAAAA;
1858 iclog->ic_state |= XLOG_STATE_IOABORT;
1859 xfs_warn(log->l_mp,
1860 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1861 be64_to_cpu(iclog->ic_header.h_lsn));
1862 }
1863#endif
1864
1865 bp->b_io_length = BTOBB(count);
1866 bp->b_fspriv = iclog;
1867 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1868 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1869
1870 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1871 bp->b_flags |= XBF_FUA;
1872
1873 /*
1874 * Flush the data device before flushing the log to make
1875 * sure all meta data written back from the AIL actually made
1876 * it to disk before stamping the new log tail LSN into the
1877 * log buffer. For an external log we need to issue the
1878 * flush explicitly, and unfortunately synchronously here;
1879 * for an internal log we can simply use the block layer
1880 * state machine for preflushes.
1881 */
1882 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1883 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1884 else
1885 bp->b_flags |= XBF_FLUSH;
1886 }
1887
1888 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1889 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1890
1891 xlog_verify_iclog(log, iclog, count, true);
1892
1893 /* account for log which doesn't start at block #0 */
1894 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1895
1896 /*
1897 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1898 * is shutting down.
1899 */
1900 error = xlog_bdstrat(bp);
1901 if (error) {
1902 xfs_buf_ioerror_alert(bp, "xlog_sync");
1903 return error;
1904 }
1905 if (split) {
1906 bp = iclog->ic_log->l_xbuf;
1907 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */
1908 xfs_buf_associate_memory(bp,
1909 (char *)&iclog->ic_header + count, split);
1910 bp->b_fspriv = iclog;
1911 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1912 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1913 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1914 bp->b_flags |= XBF_FUA;
1915
1916 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1917 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1918
1919 /* account for internal log which doesn't start at block #0 */
1920 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1921 error = xlog_bdstrat(bp);
1922 if (error) {
1923 xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1924 return error;
1925 }
1926 }
1927 return 0;
1928} /* xlog_sync */
1929
1930/*
1931 * Deallocate a log structure
1932 */
1933STATIC void
1934xlog_dealloc_log(
1935 struct xlog *log)
1936{
1937 xlog_in_core_t *iclog, *next_iclog;
1938 int i;
1939
1940 xlog_cil_destroy(log);
1941
1942 /*
1943 * Cycle all the iclogbuf locks to make sure all log IO completion
1944 * is done before we tear down these buffers.
1945 */
1946 iclog = log->l_iclog;
1947 for (i = 0; i < log->l_iclog_bufs; i++) {
1948 xfs_buf_lock(iclog->ic_bp);
1949 xfs_buf_unlock(iclog->ic_bp);
1950 iclog = iclog->ic_next;
1951 }
1952
1953 /*
1954 * Always need to ensure that the extra buffer does not point to memory
1955 * owned by another log buffer before we free it. Also, cycle the lock
1956 * first to ensure we've completed IO on it.
1957 */
1958 xfs_buf_lock(log->l_xbuf);
1959 xfs_buf_unlock(log->l_xbuf);
1960 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1961 xfs_buf_free(log->l_xbuf);
1962
1963 iclog = log->l_iclog;
1964 for (i = 0; i < log->l_iclog_bufs; i++) {
1965 xfs_buf_free(iclog->ic_bp);
1966 next_iclog = iclog->ic_next;
1967 kmem_free(iclog);
1968 iclog = next_iclog;
1969 }
1970 spinlock_destroy(&log->l_icloglock);
1971
1972 log->l_mp->m_log = NULL;
1973 kmem_free(log);
1974} /* xlog_dealloc_log */
1975
1976/*
1977 * Update counters atomically now that memcpy is done.
1978 */
1979/* ARGSUSED */
1980static inline void
1981xlog_state_finish_copy(
1982 struct xlog *log,
1983 struct xlog_in_core *iclog,
1984 int record_cnt,
1985 int copy_bytes)
1986{
1987 spin_lock(&log->l_icloglock);
1988
1989 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1990 iclog->ic_offset += copy_bytes;
1991
1992 spin_unlock(&log->l_icloglock);
1993} /* xlog_state_finish_copy */
1994
1995
1996
1997
1998/*
1999 * print out info relating to regions written which consume
2000 * the reservation
2001 */
2002void
2003xlog_print_tic_res(
2004 struct xfs_mount *mp,
2005 struct xlog_ticket *ticket)
2006{
2007 uint i;
2008 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2009
2010 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2011#define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2012 static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2013 REG_TYPE_STR(BFORMAT, "bformat"),
2014 REG_TYPE_STR(BCHUNK, "bchunk"),
2015 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2016 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2017 REG_TYPE_STR(IFORMAT, "iformat"),
2018 REG_TYPE_STR(ICORE, "icore"),
2019 REG_TYPE_STR(IEXT, "iext"),
2020 REG_TYPE_STR(IBROOT, "ibroot"),
2021 REG_TYPE_STR(ILOCAL, "ilocal"),
2022 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2023 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2024 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2025 REG_TYPE_STR(QFORMAT, "qformat"),
2026 REG_TYPE_STR(DQUOT, "dquot"),
2027 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2028 REG_TYPE_STR(LRHEADER, "LR header"),
2029 REG_TYPE_STR(UNMOUNT, "unmount"),
2030 REG_TYPE_STR(COMMIT, "commit"),
2031 REG_TYPE_STR(TRANSHDR, "trans header"),
2032 REG_TYPE_STR(ICREATE, "inode create")
2033 };
2034#undef REG_TYPE_STR
2035#define TRANS_TYPE_STR(type) [XFS_TRANS_##type] = #type
2036 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2037 TRANS_TYPE_STR(SETATTR_NOT_SIZE),
2038 TRANS_TYPE_STR(SETATTR_SIZE),
2039 TRANS_TYPE_STR(INACTIVE),
2040 TRANS_TYPE_STR(CREATE),
2041 TRANS_TYPE_STR(CREATE_TRUNC),
2042 TRANS_TYPE_STR(TRUNCATE_FILE),
2043 TRANS_TYPE_STR(REMOVE),
2044 TRANS_TYPE_STR(LINK),
2045 TRANS_TYPE_STR(RENAME),
2046 TRANS_TYPE_STR(MKDIR),
2047 TRANS_TYPE_STR(RMDIR),
2048 TRANS_TYPE_STR(SYMLINK),
2049 TRANS_TYPE_STR(SET_DMATTRS),
2050 TRANS_TYPE_STR(GROWFS),
2051 TRANS_TYPE_STR(STRAT_WRITE),
2052 TRANS_TYPE_STR(DIOSTRAT),
2053 TRANS_TYPE_STR(WRITEID),
2054 TRANS_TYPE_STR(ADDAFORK),
2055 TRANS_TYPE_STR(ATTRINVAL),
2056 TRANS_TYPE_STR(ATRUNCATE),
2057 TRANS_TYPE_STR(ATTR_SET),
2058 TRANS_TYPE_STR(ATTR_RM),
2059 TRANS_TYPE_STR(ATTR_FLAG),
2060 TRANS_TYPE_STR(CLEAR_AGI_BUCKET),
2061 TRANS_TYPE_STR(SB_CHANGE),
2062 TRANS_TYPE_STR(DUMMY1),
2063 TRANS_TYPE_STR(DUMMY2),
2064 TRANS_TYPE_STR(QM_QUOTAOFF),
2065 TRANS_TYPE_STR(QM_DQALLOC),
2066 TRANS_TYPE_STR(QM_SETQLIM),
2067 TRANS_TYPE_STR(QM_DQCLUSTER),
2068 TRANS_TYPE_STR(QM_QINOCREATE),
2069 TRANS_TYPE_STR(QM_QUOTAOFF_END),
2070 TRANS_TYPE_STR(FSYNC_TS),
2071 TRANS_TYPE_STR(GROWFSRT_ALLOC),
2072 TRANS_TYPE_STR(GROWFSRT_ZERO),
2073 TRANS_TYPE_STR(GROWFSRT_FREE),
2074 TRANS_TYPE_STR(SWAPEXT),
2075 TRANS_TYPE_STR(CHECKPOINT),
2076 TRANS_TYPE_STR(ICREATE),
2077 TRANS_TYPE_STR(CREATE_TMPFILE)
2078 };
2079#undef TRANS_TYPE_STR
2080
2081 xfs_warn(mp, "xlog_write: reservation summary:");
2082 xfs_warn(mp, " trans type = %s (%u)",
2083 ((ticket->t_trans_type <= 0 ||
2084 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2085 "bad-trans-type" : trans_type_str[ticket->t_trans_type]),
2086 ticket->t_trans_type);
2087 xfs_warn(mp, " unit res = %d bytes",
2088 ticket->t_unit_res);
2089 xfs_warn(mp, " current res = %d bytes",
2090 ticket->t_curr_res);
2091 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2092 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2093 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2094 ticket->t_res_num_ophdrs, ophdr_spc);
2095 xfs_warn(mp, " ophdr + reg = %u bytes",
2096 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2097 xfs_warn(mp, " num regions = %u",
2098 ticket->t_res_num);
2099
2100 for (i = 0; i < ticket->t_res_num; i++) {
2101 uint r_type = ticket->t_res_arr[i].r_type;
2102 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2103 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2104 "bad-rtype" : res_type_str[r_type]),
2105 ticket->t_res_arr[i].r_len);
2106 }
2107
2108 xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2109 "xlog_write: reservation ran out. Need to up reservation");
2110 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2111}
2112
2113/*
2114 * Calculate the potential space needed by the log vector. Each region gets
2115 * its own xlog_op_header_t and may need to be double word aligned.
2116 */
2117static int
2118xlog_write_calc_vec_length(
2119 struct xlog_ticket *ticket,
2120 struct xfs_log_vec *log_vector)
2121{
2122 struct xfs_log_vec *lv;
2123 int headers = 0;
2124 int len = 0;
2125 int i;
2126
2127 /* acct for start rec of xact */
2128 if (ticket->t_flags & XLOG_TIC_INITED)
2129 headers++;
2130
2131 for (lv = log_vector; lv; lv = lv->lv_next) {
2132 /* we don't write ordered log vectors */
2133 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2134 continue;
2135
2136 headers += lv->lv_niovecs;
2137
2138 for (i = 0; i < lv->lv_niovecs; i++) {
2139 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2140
2141 len += vecp->i_len;
2142 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2143 }
2144 }
2145
2146 ticket->t_res_num_ophdrs += headers;
2147 len += headers * sizeof(struct xlog_op_header);
2148
2149 return len;
2150}
2151
2152/*
2153 * If first write for transaction, insert start record We can't be trying to
2154 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2155 */
2156static int
2157xlog_write_start_rec(
2158 struct xlog_op_header *ophdr,
2159 struct xlog_ticket *ticket)
2160{
2161 if (!(ticket->t_flags & XLOG_TIC_INITED))
2162 return 0;
2163
2164 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2165 ophdr->oh_clientid = ticket->t_clientid;
2166 ophdr->oh_len = 0;
2167 ophdr->oh_flags = XLOG_START_TRANS;
2168 ophdr->oh_res2 = 0;
2169
2170 ticket->t_flags &= ~XLOG_TIC_INITED;
2171
2172 return sizeof(struct xlog_op_header);
2173}
2174
2175static xlog_op_header_t *
2176xlog_write_setup_ophdr(
2177 struct xlog *log,
2178 struct xlog_op_header *ophdr,
2179 struct xlog_ticket *ticket,
2180 uint flags)
2181{
2182 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2183 ophdr->oh_clientid = ticket->t_clientid;
2184 ophdr->oh_res2 = 0;
2185
2186 /* are we copying a commit or unmount record? */
2187 ophdr->oh_flags = flags;
2188
2189 /*
2190 * We've seen logs corrupted with bad transaction client ids. This
2191 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2192 * and shut down the filesystem.
2193 */
2194 switch (ophdr->oh_clientid) {
2195 case XFS_TRANSACTION:
2196 case XFS_VOLUME:
2197 case XFS_LOG:
2198 break;
2199 default:
2200 xfs_warn(log->l_mp,
2201 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2202 ophdr->oh_clientid, ticket);
2203 return NULL;
2204 }
2205
2206 return ophdr;
2207}
2208
2209/*
2210 * Set up the parameters of the region copy into the log. This has
2211 * to handle region write split across multiple log buffers - this
2212 * state is kept external to this function so that this code can
2213 * be written in an obvious, self documenting manner.
2214 */
2215static int
2216xlog_write_setup_copy(
2217 struct xlog_ticket *ticket,
2218 struct xlog_op_header *ophdr,
2219 int space_available,
2220 int space_required,
2221 int *copy_off,
2222 int *copy_len,
2223 int *last_was_partial_copy,
2224 int *bytes_consumed)
2225{
2226 int still_to_copy;
2227
2228 still_to_copy = space_required - *bytes_consumed;
2229 *copy_off = *bytes_consumed;
2230
2231 if (still_to_copy <= space_available) {
2232 /* write of region completes here */
2233 *copy_len = still_to_copy;
2234 ophdr->oh_len = cpu_to_be32(*copy_len);
2235 if (*last_was_partial_copy)
2236 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2237 *last_was_partial_copy = 0;
2238 *bytes_consumed = 0;
2239 return 0;
2240 }
2241
2242 /* partial write of region, needs extra log op header reservation */
2243 *copy_len = space_available;
2244 ophdr->oh_len = cpu_to_be32(*copy_len);
2245 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2246 if (*last_was_partial_copy)
2247 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2248 *bytes_consumed += *copy_len;
2249 (*last_was_partial_copy)++;
2250
2251 /* account for new log op header */
2252 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2253 ticket->t_res_num_ophdrs++;
2254
2255 return sizeof(struct xlog_op_header);
2256}
2257
2258static int
2259xlog_write_copy_finish(
2260 struct xlog *log,
2261 struct xlog_in_core *iclog,
2262 uint flags,
2263 int *record_cnt,
2264 int *data_cnt,
2265 int *partial_copy,
2266 int *partial_copy_len,
2267 int log_offset,
2268 struct xlog_in_core **commit_iclog)
2269{
2270 if (*partial_copy) {
2271 /*
2272 * This iclog has already been marked WANT_SYNC by
2273 * xlog_state_get_iclog_space.
2274 */
2275 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2276 *record_cnt = 0;
2277 *data_cnt = 0;
2278 return xlog_state_release_iclog(log, iclog);
2279 }
2280
2281 *partial_copy = 0;
2282 *partial_copy_len = 0;
2283
2284 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2285 /* no more space in this iclog - push it. */
2286 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2287 *record_cnt = 0;
2288 *data_cnt = 0;
2289
2290 spin_lock(&log->l_icloglock);
2291 xlog_state_want_sync(log, iclog);
2292 spin_unlock(&log->l_icloglock);
2293
2294 if (!commit_iclog)
2295 return xlog_state_release_iclog(log, iclog);
2296 ASSERT(flags & XLOG_COMMIT_TRANS);
2297 *commit_iclog = iclog;
2298 }
2299
2300 return 0;
2301}
2302
2303/*
2304 * Write some region out to in-core log
2305 *
2306 * This will be called when writing externally provided regions or when
2307 * writing out a commit record for a given transaction.
2308 *
2309 * General algorithm:
2310 * 1. Find total length of this write. This may include adding to the
2311 * lengths passed in.
2312 * 2. Check whether we violate the tickets reservation.
2313 * 3. While writing to this iclog
2314 * A. Reserve as much space in this iclog as can get
2315 * B. If this is first write, save away start lsn
2316 * C. While writing this region:
2317 * 1. If first write of transaction, write start record
2318 * 2. Write log operation header (header per region)
2319 * 3. Find out if we can fit entire region into this iclog
2320 * 4. Potentially, verify destination memcpy ptr
2321 * 5. Memcpy (partial) region
2322 * 6. If partial copy, release iclog; otherwise, continue
2323 * copying more regions into current iclog
2324 * 4. Mark want sync bit (in simulation mode)
2325 * 5. Release iclog for potential flush to on-disk log.
2326 *
2327 * ERRORS:
2328 * 1. Panic if reservation is overrun. This should never happen since
2329 * reservation amounts are generated internal to the filesystem.
2330 * NOTES:
2331 * 1. Tickets are single threaded data structures.
2332 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2333 * syncing routine. When a single log_write region needs to span
2334 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2335 * on all log operation writes which don't contain the end of the
2336 * region. The XLOG_END_TRANS bit is used for the in-core log
2337 * operation which contains the end of the continued log_write region.
2338 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2339 * we don't really know exactly how much space will be used. As a result,
2340 * we don't update ic_offset until the end when we know exactly how many
2341 * bytes have been written out.
2342 */
2343int
2344xlog_write(
2345 struct xlog *log,
2346 struct xfs_log_vec *log_vector,
2347 struct xlog_ticket *ticket,
2348 xfs_lsn_t *start_lsn,
2349 struct xlog_in_core **commit_iclog,
2350 uint flags)
2351{
2352 struct xlog_in_core *iclog = NULL;
2353 struct xfs_log_iovec *vecp;
2354 struct xfs_log_vec *lv;
2355 int len;
2356 int index;
2357 int partial_copy = 0;
2358 int partial_copy_len = 0;
2359 int contwr = 0;
2360 int record_cnt = 0;
2361 int data_cnt = 0;
2362 int error;
2363
2364 *start_lsn = 0;
2365
2366 len = xlog_write_calc_vec_length(ticket, log_vector);
2367
2368 /*
2369 * Region headers and bytes are already accounted for.
2370 * We only need to take into account start records and
2371 * split regions in this function.
2372 */
2373 if (ticket->t_flags & XLOG_TIC_INITED)
2374 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2375
2376 /*
2377 * Commit record headers need to be accounted for. These
2378 * come in as separate writes so are easy to detect.
2379 */
2380 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2381 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2382
2383 if (ticket->t_curr_res < 0)
2384 xlog_print_tic_res(log->l_mp, ticket);
2385
2386 index = 0;
2387 lv = log_vector;
2388 vecp = lv->lv_iovecp;
2389 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2390 void *ptr;
2391 int log_offset;
2392
2393 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2394 &contwr, &log_offset);
2395 if (error)
2396 return error;
2397
2398 ASSERT(log_offset <= iclog->ic_size - 1);
2399 ptr = iclog->ic_datap + log_offset;
2400
2401 /* start_lsn is the first lsn written to. That's all we need. */
2402 if (!*start_lsn)
2403 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2404
2405 /*
2406 * This loop writes out as many regions as can fit in the amount
2407 * of space which was allocated by xlog_state_get_iclog_space().
2408 */
2409 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2410 struct xfs_log_iovec *reg;
2411 struct xlog_op_header *ophdr;
2412 int start_rec_copy;
2413 int copy_len;
2414 int copy_off;
2415 bool ordered = false;
2416
2417 /* ordered log vectors have no regions to write */
2418 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2419 ASSERT(lv->lv_niovecs == 0);
2420 ordered = true;
2421 goto next_lv;
2422 }
2423
2424 reg = &vecp[index];
2425 ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2426 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2427
2428 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2429 if (start_rec_copy) {
2430 record_cnt++;
2431 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2432 start_rec_copy);
2433 }
2434
2435 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2436 if (!ophdr)
2437 return -EIO;
2438
2439 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2440 sizeof(struct xlog_op_header));
2441
2442 len += xlog_write_setup_copy(ticket, ophdr,
2443 iclog->ic_size-log_offset,
2444 reg->i_len,
2445 ©_off, ©_len,
2446 &partial_copy,
2447 &partial_copy_len);
2448 xlog_verify_dest_ptr(log, ptr);
2449
2450 /*
2451 * Copy region.
2452 *
2453 * Unmount records just log an opheader, so can have
2454 * empty payloads with no data region to copy. Hence we
2455 * only copy the payload if the vector says it has data
2456 * to copy.
2457 */
2458 ASSERT(copy_len >= 0);
2459 if (copy_len > 0) {
2460 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2461 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2462 copy_len);
2463 }
2464 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2465 record_cnt++;
2466 data_cnt += contwr ? copy_len : 0;
2467
2468 error = xlog_write_copy_finish(log, iclog, flags,
2469 &record_cnt, &data_cnt,
2470 &partial_copy,
2471 &partial_copy_len,
2472 log_offset,
2473 commit_iclog);
2474 if (error)
2475 return error;
2476
2477 /*
2478 * if we had a partial copy, we need to get more iclog
2479 * space but we don't want to increment the region
2480 * index because there is still more is this region to
2481 * write.
2482 *
2483 * If we completed writing this region, and we flushed
2484 * the iclog (indicated by resetting of the record
2485 * count), then we also need to get more log space. If
2486 * this was the last record, though, we are done and
2487 * can just return.
2488 */
2489 if (partial_copy)
2490 break;
2491
2492 if (++index == lv->lv_niovecs) {
2493next_lv:
2494 lv = lv->lv_next;
2495 index = 0;
2496 if (lv)
2497 vecp = lv->lv_iovecp;
2498 }
2499 if (record_cnt == 0 && ordered == false) {
2500 if (!lv)
2501 return 0;
2502 break;
2503 }
2504 }
2505 }
2506
2507 ASSERT(len == 0);
2508
2509 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2510 if (!commit_iclog)
2511 return xlog_state_release_iclog(log, iclog);
2512
2513 ASSERT(flags & XLOG_COMMIT_TRANS);
2514 *commit_iclog = iclog;
2515 return 0;
2516}
2517
2518
2519/*****************************************************************************
2520 *
2521 * State Machine functions
2522 *
2523 *****************************************************************************
2524 */
2525
2526/* Clean iclogs starting from the head. This ordering must be
2527 * maintained, so an iclog doesn't become ACTIVE beyond one that
2528 * is SYNCING. This is also required to maintain the notion that we use
2529 * a ordered wait queue to hold off would be writers to the log when every
2530 * iclog is trying to sync to disk.
2531 *
2532 * State Change: DIRTY -> ACTIVE
2533 */
2534STATIC void
2535xlog_state_clean_log(
2536 struct xlog *log)
2537{
2538 xlog_in_core_t *iclog;
2539 int changed = 0;
2540
2541 iclog = log->l_iclog;
2542 do {
2543 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2544 iclog->ic_state = XLOG_STATE_ACTIVE;
2545 iclog->ic_offset = 0;
2546 ASSERT(iclog->ic_callback == NULL);
2547 /*
2548 * If the number of ops in this iclog indicate it just
2549 * contains the dummy transaction, we can
2550 * change state into IDLE (the second time around).
2551 * Otherwise we should change the state into
2552 * NEED a dummy.
2553 * We don't need to cover the dummy.
2554 */
2555 if (!changed &&
2556 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2557 XLOG_COVER_OPS)) {
2558 changed = 1;
2559 } else {
2560 /*
2561 * We have two dirty iclogs so start over
2562 * This could also be num of ops indicates
2563 * this is not the dummy going out.
2564 */
2565 changed = 2;
2566 }
2567 iclog->ic_header.h_num_logops = 0;
2568 memset(iclog->ic_header.h_cycle_data, 0,
2569 sizeof(iclog->ic_header.h_cycle_data));
2570 iclog->ic_header.h_lsn = 0;
2571 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2572 /* do nothing */;
2573 else
2574 break; /* stop cleaning */
2575 iclog = iclog->ic_next;
2576 } while (iclog != log->l_iclog);
2577
2578 /* log is locked when we are called */
2579 /*
2580 * Change state for the dummy log recording.
2581 * We usually go to NEED. But we go to NEED2 if the changed indicates
2582 * we are done writing the dummy record.
2583 * If we are done with the second dummy recored (DONE2), then
2584 * we go to IDLE.
2585 */
2586 if (changed) {
2587 switch (log->l_covered_state) {
2588 case XLOG_STATE_COVER_IDLE:
2589 case XLOG_STATE_COVER_NEED:
2590 case XLOG_STATE_COVER_NEED2:
2591 log->l_covered_state = XLOG_STATE_COVER_NEED;
2592 break;
2593
2594 case XLOG_STATE_COVER_DONE:
2595 if (changed == 1)
2596 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2597 else
2598 log->l_covered_state = XLOG_STATE_COVER_NEED;
2599 break;
2600
2601 case XLOG_STATE_COVER_DONE2:
2602 if (changed == 1)
2603 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2604 else
2605 log->l_covered_state = XLOG_STATE_COVER_NEED;
2606 break;
2607
2608 default:
2609 ASSERT(0);
2610 }
2611 }
2612} /* xlog_state_clean_log */
2613
2614STATIC xfs_lsn_t
2615xlog_get_lowest_lsn(
2616 struct xlog *log)
2617{
2618 xlog_in_core_t *lsn_log;
2619 xfs_lsn_t lowest_lsn, lsn;
2620
2621 lsn_log = log->l_iclog;
2622 lowest_lsn = 0;
2623 do {
2624 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2625 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2626 if ((lsn && !lowest_lsn) ||
2627 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2628 lowest_lsn = lsn;
2629 }
2630 }
2631 lsn_log = lsn_log->ic_next;
2632 } while (lsn_log != log->l_iclog);
2633 return lowest_lsn;
2634}
2635
2636
2637STATIC void
2638xlog_state_do_callback(
2639 struct xlog *log,
2640 int aborted,
2641 struct xlog_in_core *ciclog)
2642{
2643 xlog_in_core_t *iclog;
2644 xlog_in_core_t *first_iclog; /* used to know when we've
2645 * processed all iclogs once */
2646 xfs_log_callback_t *cb, *cb_next;
2647 int flushcnt = 0;
2648 xfs_lsn_t lowest_lsn;
2649 int ioerrors; /* counter: iclogs with errors */
2650 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2651 int funcdidcallbacks; /* flag: function did callbacks */
2652 int repeats; /* for issuing console warnings if
2653 * looping too many times */
2654 int wake = 0;
2655
2656 spin_lock(&log->l_icloglock);
2657 first_iclog = iclog = log->l_iclog;
2658 ioerrors = 0;
2659 funcdidcallbacks = 0;
2660 repeats = 0;
2661
2662 do {
2663 /*
2664 * Scan all iclogs starting with the one pointed to by the
2665 * log. Reset this starting point each time the log is
2666 * unlocked (during callbacks).
2667 *
2668 * Keep looping through iclogs until one full pass is made
2669 * without running any callbacks.
2670 */
2671 first_iclog = log->l_iclog;
2672 iclog = log->l_iclog;
2673 loopdidcallbacks = 0;
2674 repeats++;
2675
2676 do {
2677
2678 /* skip all iclogs in the ACTIVE & DIRTY states */
2679 if (iclog->ic_state &
2680 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2681 iclog = iclog->ic_next;
2682 continue;
2683 }
2684
2685 /*
2686 * Between marking a filesystem SHUTDOWN and stopping
2687 * the log, we do flush all iclogs to disk (if there
2688 * wasn't a log I/O error). So, we do want things to
2689 * go smoothly in case of just a SHUTDOWN w/o a
2690 * LOG_IO_ERROR.
2691 */
2692 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2693 /*
2694 * Can only perform callbacks in order. Since
2695 * this iclog is not in the DONE_SYNC/
2696 * DO_CALLBACK state, we skip the rest and
2697 * just try to clean up. If we set our iclog
2698 * to DO_CALLBACK, we will not process it when
2699 * we retry since a previous iclog is in the
2700 * CALLBACK and the state cannot change since
2701 * we are holding the l_icloglock.
2702 */
2703 if (!(iclog->ic_state &
2704 (XLOG_STATE_DONE_SYNC |
2705 XLOG_STATE_DO_CALLBACK))) {
2706 if (ciclog && (ciclog->ic_state ==
2707 XLOG_STATE_DONE_SYNC)) {
2708 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2709 }
2710 break;
2711 }
2712 /*
2713 * We now have an iclog that is in either the
2714 * DO_CALLBACK or DONE_SYNC states. The other
2715 * states (WANT_SYNC, SYNCING, or CALLBACK were
2716 * caught by the above if and are going to
2717 * clean (i.e. we aren't doing their callbacks)
2718 * see the above if.
2719 */
2720
2721 /*
2722 * We will do one more check here to see if we
2723 * have chased our tail around.
2724 */
2725
2726 lowest_lsn = xlog_get_lowest_lsn(log);
2727 if (lowest_lsn &&
2728 XFS_LSN_CMP(lowest_lsn,
2729 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2730 iclog = iclog->ic_next;
2731 continue; /* Leave this iclog for
2732 * another thread */
2733 }
2734
2735 iclog->ic_state = XLOG_STATE_CALLBACK;
2736
2737
2738 /*
2739 * Completion of a iclog IO does not imply that
2740 * a transaction has completed, as transactions
2741 * can be large enough to span many iclogs. We
2742 * cannot change the tail of the log half way
2743 * through a transaction as this may be the only
2744 * transaction in the log and moving th etail to
2745 * point to the middle of it will prevent
2746 * recovery from finding the start of the
2747 * transaction. Hence we should only update the
2748 * last_sync_lsn if this iclog contains
2749 * transaction completion callbacks on it.
2750 *
2751 * We have to do this before we drop the
2752 * icloglock to ensure we are the only one that
2753 * can update it.
2754 */
2755 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2756 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2757 if (iclog->ic_callback)
2758 atomic64_set(&log->l_last_sync_lsn,
2759 be64_to_cpu(iclog->ic_header.h_lsn));
2760
2761 } else
2762 ioerrors++;
2763
2764 spin_unlock(&log->l_icloglock);
2765
2766 /*
2767 * Keep processing entries in the callback list until
2768 * we come around and it is empty. We need to
2769 * atomically see that the list is empty and change the
2770 * state to DIRTY so that we don't miss any more
2771 * callbacks being added.
2772 */
2773 spin_lock(&iclog->ic_callback_lock);
2774 cb = iclog->ic_callback;
2775 while (cb) {
2776 iclog->ic_callback_tail = &(iclog->ic_callback);
2777 iclog->ic_callback = NULL;
2778 spin_unlock(&iclog->ic_callback_lock);
2779
2780 /* perform callbacks in the order given */
2781 for (; cb; cb = cb_next) {
2782 cb_next = cb->cb_next;
2783 cb->cb_func(cb->cb_arg, aborted);
2784 }
2785 spin_lock(&iclog->ic_callback_lock);
2786 cb = iclog->ic_callback;
2787 }
2788
2789 loopdidcallbacks++;
2790 funcdidcallbacks++;
2791
2792 spin_lock(&log->l_icloglock);
2793 ASSERT(iclog->ic_callback == NULL);
2794 spin_unlock(&iclog->ic_callback_lock);
2795 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2796 iclog->ic_state = XLOG_STATE_DIRTY;
2797
2798 /*
2799 * Transition from DIRTY to ACTIVE if applicable.
2800 * NOP if STATE_IOERROR.
2801 */
2802 xlog_state_clean_log(log);
2803
2804 /* wake up threads waiting in xfs_log_force() */
2805 wake_up_all(&iclog->ic_force_wait);
2806
2807 iclog = iclog->ic_next;
2808 } while (first_iclog != iclog);
2809
2810 if (repeats > 5000) {
2811 flushcnt += repeats;
2812 repeats = 0;
2813 xfs_warn(log->l_mp,
2814 "%s: possible infinite loop (%d iterations)",
2815 __func__, flushcnt);
2816 }
2817 } while (!ioerrors && loopdidcallbacks);
2818
2819#ifdef DEBUG
2820 /*
2821 * Make one last gasp attempt to see if iclogs are being left in limbo.
2822 * If the above loop finds an iclog earlier than the current iclog and
2823 * in one of the syncing states, the current iclog is put into
2824 * DO_CALLBACK and the callbacks are deferred to the completion of the
2825 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2826 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2827 * states.
2828 *
2829 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2830 * for ic_state == SYNCING.
2831 */
2832 if (funcdidcallbacks) {
2833 first_iclog = iclog = log->l_iclog;
2834 do {
2835 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2836 /*
2837 * Terminate the loop if iclogs are found in states
2838 * which will cause other threads to clean up iclogs.
2839 *
2840 * SYNCING - i/o completion will go through logs
2841 * DONE_SYNC - interrupt thread should be waiting for
2842 * l_icloglock
2843 * IOERROR - give up hope all ye who enter here
2844 */
2845 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2846 iclog->ic_state & XLOG_STATE_SYNCING ||
2847 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2848 iclog->ic_state == XLOG_STATE_IOERROR )
2849 break;
2850 iclog = iclog->ic_next;
2851 } while (first_iclog != iclog);
2852 }
2853#endif
2854
2855 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2856 wake = 1;
2857 spin_unlock(&log->l_icloglock);
2858
2859 if (wake)
2860 wake_up_all(&log->l_flush_wait);
2861}
2862
2863
2864/*
2865 * Finish transitioning this iclog to the dirty state.
2866 *
2867 * Make sure that we completely execute this routine only when this is
2868 * the last call to the iclog. There is a good chance that iclog flushes,
2869 * when we reach the end of the physical log, get turned into 2 separate
2870 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2871 * routine. By using the reference count bwritecnt, we guarantee that only
2872 * the second completion goes through.
2873 *
2874 * Callbacks could take time, so they are done outside the scope of the
2875 * global state machine log lock.
2876 */
2877STATIC void
2878xlog_state_done_syncing(
2879 xlog_in_core_t *iclog,
2880 int aborted)
2881{
2882 struct xlog *log = iclog->ic_log;
2883
2884 spin_lock(&log->l_icloglock);
2885
2886 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2887 iclog->ic_state == XLOG_STATE_IOERROR);
2888 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2889 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2890
2891
2892 /*
2893 * If we got an error, either on the first buffer, or in the case of
2894 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2895 * and none should ever be attempted to be written to disk
2896 * again.
2897 */
2898 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2899 if (--iclog->ic_bwritecnt == 1) {
2900 spin_unlock(&log->l_icloglock);
2901 return;
2902 }
2903 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2904 }
2905
2906 /*
2907 * Someone could be sleeping prior to writing out the next
2908 * iclog buffer, we wake them all, one will get to do the
2909 * I/O, the others get to wait for the result.
2910 */
2911 wake_up_all(&iclog->ic_write_wait);
2912 spin_unlock(&log->l_icloglock);
2913 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2914} /* xlog_state_done_syncing */
2915
2916
2917/*
2918 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2919 * sleep. We wait on the flush queue on the head iclog as that should be
2920 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2921 * we will wait here and all new writes will sleep until a sync completes.
2922 *
2923 * The in-core logs are used in a circular fashion. They are not used
2924 * out-of-order even when an iclog past the head is free.
2925 *
2926 * return:
2927 * * log_offset where xlog_write() can start writing into the in-core
2928 * log's data space.
2929 * * in-core log pointer to which xlog_write() should write.
2930 * * boolean indicating this is a continued write to an in-core log.
2931 * If this is the last write, then the in-core log's offset field
2932 * needs to be incremented, depending on the amount of data which
2933 * is copied.
2934 */
2935STATIC int
2936xlog_state_get_iclog_space(
2937 struct xlog *log,
2938 int len,
2939 struct xlog_in_core **iclogp,
2940 struct xlog_ticket *ticket,
2941 int *continued_write,
2942 int *logoffsetp)
2943{
2944 int log_offset;
2945 xlog_rec_header_t *head;
2946 xlog_in_core_t *iclog;
2947 int error;
2948
2949restart:
2950 spin_lock(&log->l_icloglock);
2951 if (XLOG_FORCED_SHUTDOWN(log)) {
2952 spin_unlock(&log->l_icloglock);
2953 return -EIO;
2954 }
2955
2956 iclog = log->l_iclog;
2957 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2958 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2959
2960 /* Wait for log writes to have flushed */
2961 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2962 goto restart;
2963 }
2964
2965 head = &iclog->ic_header;
2966
2967 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2968 log_offset = iclog->ic_offset;
2969
2970 /* On the 1st write to an iclog, figure out lsn. This works
2971 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2972 * committing to. If the offset is set, that's how many blocks
2973 * must be written.
2974 */
2975 if (log_offset == 0) {
2976 ticket->t_curr_res -= log->l_iclog_hsize;
2977 xlog_tic_add_region(ticket,
2978 log->l_iclog_hsize,
2979 XLOG_REG_TYPE_LRHEADER);
2980 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2981 head->h_lsn = cpu_to_be64(
2982 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2983 ASSERT(log->l_curr_block >= 0);
2984 }
2985
2986 /* If there is enough room to write everything, then do it. Otherwise,
2987 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2988 * bit is on, so this will get flushed out. Don't update ic_offset
2989 * until you know exactly how many bytes get copied. Therefore, wait
2990 * until later to update ic_offset.
2991 *
2992 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2993 * can fit into remaining data section.
2994 */
2995 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2996 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2997
2998 /*
2999 * If I'm the only one writing to this iclog, sync it to disk.
3000 * We need to do an atomic compare and decrement here to avoid
3001 * 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 /* we are the only one */
3007 spin_unlock(&log->l_icloglock);
3008 error = xlog_state_release_iclog(log, iclog);
3009 if (error)
3010 return error;
3011 } else {
3012 spin_unlock(&log->l_icloglock);
3013 }
3014 goto restart;
3015 }
3016
3017 /* Do we have enough room to write the full amount in the remainder
3018 * of this iclog? Or must we continue a write on the next iclog and
3019 * mark this iclog as completely taken? In the case where we switch
3020 * iclogs (to mark it taken), this particular iclog will release/sync
3021 * to disk in xlog_write().
3022 */
3023 if (len <= iclog->ic_size - iclog->ic_offset) {
3024 *continued_write = 0;
3025 iclog->ic_offset += len;
3026 } else {
3027 *continued_write = 1;
3028 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3029 }
3030 *iclogp = iclog;
3031
3032 ASSERT(iclog->ic_offset <= iclog->ic_size);
3033 spin_unlock(&log->l_icloglock);
3034
3035 *logoffsetp = log_offset;
3036 return 0;
3037} /* xlog_state_get_iclog_space */
3038
3039/* The first cnt-1 times through here we don't need to
3040 * move the grant write head because the permanent
3041 * reservation has reserved cnt times the unit amount.
3042 * Release part of current permanent unit reservation and
3043 * reset current reservation to be one units worth. Also
3044 * move grant reservation head forward.
3045 */
3046STATIC void
3047xlog_regrant_reserve_log_space(
3048 struct xlog *log,
3049 struct xlog_ticket *ticket)
3050{
3051 trace_xfs_log_regrant_reserve_enter(log, ticket);
3052
3053 if (ticket->t_cnt > 0)
3054 ticket->t_cnt--;
3055
3056 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3057 ticket->t_curr_res);
3058 xlog_grant_sub_space(log, &log->l_write_head.grant,
3059 ticket->t_curr_res);
3060 ticket->t_curr_res = ticket->t_unit_res;
3061 xlog_tic_reset_res(ticket);
3062
3063 trace_xfs_log_regrant_reserve_sub(log, ticket);
3064
3065 /* just return if we still have some of the pre-reserved space */
3066 if (ticket->t_cnt > 0)
3067 return;
3068
3069 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3070 ticket->t_unit_res);
3071
3072 trace_xfs_log_regrant_reserve_exit(log, ticket);
3073
3074 ticket->t_curr_res = ticket->t_unit_res;
3075 xlog_tic_reset_res(ticket);
3076} /* xlog_regrant_reserve_log_space */
3077
3078
3079/*
3080 * Give back the space left from a reservation.
3081 *
3082 * All the information we need to make a correct determination of space left
3083 * is present. For non-permanent reservations, things are quite easy. The
3084 * count should have been decremented to zero. We only need to deal with the
3085 * space remaining in the current reservation part of the ticket. If the
3086 * ticket contains a permanent reservation, there may be left over space which
3087 * needs to be released. A count of N means that N-1 refills of the current
3088 * reservation can be done before we need to ask for more space. The first
3089 * one goes to fill up the first current reservation. Once we run out of
3090 * space, the count will stay at zero and the only space remaining will be
3091 * in the current reservation field.
3092 */
3093STATIC void
3094xlog_ungrant_log_space(
3095 struct xlog *log,
3096 struct xlog_ticket *ticket)
3097{
3098 int bytes;
3099
3100 if (ticket->t_cnt > 0)
3101 ticket->t_cnt--;
3102
3103 trace_xfs_log_ungrant_enter(log, ticket);
3104 trace_xfs_log_ungrant_sub(log, ticket);
3105
3106 /*
3107 * If this is a permanent reservation ticket, we may be able to free
3108 * up more space based on the remaining count.
3109 */
3110 bytes = ticket->t_curr_res;
3111 if (ticket->t_cnt > 0) {
3112 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3113 bytes += ticket->t_unit_res*ticket->t_cnt;
3114 }
3115
3116 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3117 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3118
3119 trace_xfs_log_ungrant_exit(log, ticket);
3120
3121 xfs_log_space_wake(log->l_mp);
3122}
3123
3124/*
3125 * Flush iclog to disk if this is the last reference to the given iclog and
3126 * the WANT_SYNC bit is set.
3127 *
3128 * When this function is entered, the iclog is not necessarily in the
3129 * WANT_SYNC state. It may be sitting around waiting to get filled.
3130 *
3131 *
3132 */
3133STATIC int
3134xlog_state_release_iclog(
3135 struct xlog *log,
3136 struct xlog_in_core *iclog)
3137{
3138 int sync = 0; /* do we sync? */
3139
3140 if (iclog->ic_state & XLOG_STATE_IOERROR)
3141 return -EIO;
3142
3143 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3144 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3145 return 0;
3146
3147 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3148 spin_unlock(&log->l_icloglock);
3149 return -EIO;
3150 }
3151 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3152 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3153
3154 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3155 /* update tail before writing to iclog */
3156 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3157 sync++;
3158 iclog->ic_state = XLOG_STATE_SYNCING;
3159 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3160 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3161 /* cycle incremented when incrementing curr_block */
3162 }
3163 spin_unlock(&log->l_icloglock);
3164
3165 /*
3166 * We let the log lock go, so it's possible that we hit a log I/O
3167 * error or some other SHUTDOWN condition that marks the iclog
3168 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3169 * this iclog has consistent data, so we ignore IOERROR
3170 * flags after this point.
3171 */
3172 if (sync)
3173 return xlog_sync(log, iclog);
3174 return 0;
3175} /* xlog_state_release_iclog */
3176
3177
3178/*
3179 * This routine will mark the current iclog in the ring as WANT_SYNC
3180 * and move the current iclog pointer to the next iclog in the ring.
3181 * When this routine is called from xlog_state_get_iclog_space(), the
3182 * exact size of the iclog has not yet been determined. All we know is
3183 * that every data block. We have run out of space in this log record.
3184 */
3185STATIC void
3186xlog_state_switch_iclogs(
3187 struct xlog *log,
3188 struct xlog_in_core *iclog,
3189 int eventual_size)
3190{
3191 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3192 if (!eventual_size)
3193 eventual_size = iclog->ic_offset;
3194 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3195 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3196 log->l_prev_block = log->l_curr_block;
3197 log->l_prev_cycle = log->l_curr_cycle;
3198
3199 /* roll log?: ic_offset changed later */
3200 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3201
3202 /* Round up to next log-sunit */
3203 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3204 log->l_mp->m_sb.sb_logsunit > 1) {
3205 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3206 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3207 }
3208
3209 if (log->l_curr_block >= log->l_logBBsize) {
3210 /*
3211 * Rewind the current block before the cycle is bumped to make
3212 * sure that the combined LSN never transiently moves forward
3213 * when the log wraps to the next cycle. This is to support the
3214 * unlocked sample of these fields from xlog_valid_lsn(). Most
3215 * other cases should acquire l_icloglock.
3216 */
3217 log->l_curr_block -= log->l_logBBsize;
3218 ASSERT(log->l_curr_block >= 0);
3219 smp_wmb();
3220 log->l_curr_cycle++;
3221 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3222 log->l_curr_cycle++;
3223 }
3224 ASSERT(iclog == log->l_iclog);
3225 log->l_iclog = iclog->ic_next;
3226} /* xlog_state_switch_iclogs */
3227
3228/*
3229 * Write out all data in the in-core log as of this exact moment in time.
3230 *
3231 * Data may be written to the in-core log during this call. However,
3232 * we don't guarantee this data will be written out. A change from past
3233 * implementation means this routine will *not* write out zero length LRs.
3234 *
3235 * Basically, we try and perform an intelligent scan of the in-core logs.
3236 * If we determine there is no flushable data, we just return. There is no
3237 * flushable data if:
3238 *
3239 * 1. the current iclog is active and has no data; the previous iclog
3240 * is in the active or dirty state.
3241 * 2. the current iclog is drity, and the previous iclog is in the
3242 * active or dirty state.
3243 *
3244 * We may sleep if:
3245 *
3246 * 1. the current iclog is not in the active nor dirty state.
3247 * 2. the current iclog dirty, and the previous iclog is not in the
3248 * active nor dirty state.
3249 * 3. the current iclog is active, and there is another thread writing
3250 * to this particular iclog.
3251 * 4. a) the current iclog is active and has no other writers
3252 * b) when we return from flushing out this iclog, it is still
3253 * not in the active nor dirty state.
3254 */
3255int
3256_xfs_log_force(
3257 struct xfs_mount *mp,
3258 uint flags,
3259 int *log_flushed)
3260{
3261 struct xlog *log = mp->m_log;
3262 struct xlog_in_core *iclog;
3263 xfs_lsn_t lsn;
3264
3265 XFS_STATS_INC(mp, xs_log_force);
3266
3267 xlog_cil_force(log);
3268
3269 spin_lock(&log->l_icloglock);
3270
3271 iclog = log->l_iclog;
3272 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3273 spin_unlock(&log->l_icloglock);
3274 return -EIO;
3275 }
3276
3277 /* If the head iclog is not active nor dirty, we just attach
3278 * ourselves to the head and go to sleep.
3279 */
3280 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3281 iclog->ic_state == XLOG_STATE_DIRTY) {
3282 /*
3283 * If the head is dirty or (active and empty), then
3284 * we need to look at the previous iclog. If the previous
3285 * iclog is active or dirty we are done. There is nothing
3286 * to sync out. Otherwise, we attach ourselves to the
3287 * previous iclog and go to sleep.
3288 */
3289 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3290 (atomic_read(&iclog->ic_refcnt) == 0
3291 && iclog->ic_offset == 0)) {
3292 iclog = iclog->ic_prev;
3293 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3294 iclog->ic_state == XLOG_STATE_DIRTY)
3295 goto no_sleep;
3296 else
3297 goto maybe_sleep;
3298 } else {
3299 if (atomic_read(&iclog->ic_refcnt) == 0) {
3300 /* We are the only one with access to this
3301 * iclog. Flush it out now. There should
3302 * be a roundoff of zero to show that someone
3303 * has already taken care of the roundoff from
3304 * the previous sync.
3305 */
3306 atomic_inc(&iclog->ic_refcnt);
3307 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3308 xlog_state_switch_iclogs(log, iclog, 0);
3309 spin_unlock(&log->l_icloglock);
3310
3311 if (xlog_state_release_iclog(log, iclog))
3312 return -EIO;
3313
3314 if (log_flushed)
3315 *log_flushed = 1;
3316 spin_lock(&log->l_icloglock);
3317 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3318 iclog->ic_state != XLOG_STATE_DIRTY)
3319 goto maybe_sleep;
3320 else
3321 goto no_sleep;
3322 } else {
3323 /* Someone else is writing to this iclog.
3324 * Use its call to flush out the data. However,
3325 * the other thread may not force out this LR,
3326 * so we mark it WANT_SYNC.
3327 */
3328 xlog_state_switch_iclogs(log, iclog, 0);
3329 goto maybe_sleep;
3330 }
3331 }
3332 }
3333
3334 /* By the time we come around again, the iclog could've been filled
3335 * which would give it another lsn. If we have a new lsn, just
3336 * return because the relevant data has been flushed.
3337 */
3338maybe_sleep:
3339 if (flags & XFS_LOG_SYNC) {
3340 /*
3341 * We must check if we're shutting down here, before
3342 * we wait, while we're holding the l_icloglock.
3343 * Then we check again after waking up, in case our
3344 * sleep was disturbed by a bad news.
3345 */
3346 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3347 spin_unlock(&log->l_icloglock);
3348 return -EIO;
3349 }
3350 XFS_STATS_INC(mp, xs_log_force_sleep);
3351 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3352 /*
3353 * No need to grab the log lock here since we're
3354 * only deciding whether or not to return EIO
3355 * and the memory read should be atomic.
3356 */
3357 if (iclog->ic_state & XLOG_STATE_IOERROR)
3358 return -EIO;
3359 if (log_flushed)
3360 *log_flushed = 1;
3361 } else {
3362
3363no_sleep:
3364 spin_unlock(&log->l_icloglock);
3365 }
3366 return 0;
3367}
3368
3369/*
3370 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3371 * about errors or whether the log was flushed or not. This is the normal
3372 * interface to use when trying to unpin items or move the log forward.
3373 */
3374void
3375xfs_log_force(
3376 xfs_mount_t *mp,
3377 uint flags)
3378{
3379 int error;
3380
3381 trace_xfs_log_force(mp, 0);
3382 error = _xfs_log_force(mp, flags, NULL);
3383 if (error)
3384 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3385}
3386
3387/*
3388 * Force the in-core log to disk for a specific LSN.
3389 *
3390 * Find in-core log with lsn.
3391 * If it is in the DIRTY state, just return.
3392 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3393 * state and go to sleep or return.
3394 * If it is in any other state, go to sleep or return.
3395 *
3396 * Synchronous forces are implemented with a signal variable. All callers
3397 * to force a given lsn to disk will wait on a the sv attached to the
3398 * specific in-core log. When given in-core log finally completes its
3399 * write to disk, that thread will wake up all threads waiting on the
3400 * sv.
3401 */
3402int
3403_xfs_log_force_lsn(
3404 struct xfs_mount *mp,
3405 xfs_lsn_t lsn,
3406 uint flags,
3407 int *log_flushed)
3408{
3409 struct xlog *log = mp->m_log;
3410 struct xlog_in_core *iclog;
3411 int already_slept = 0;
3412
3413 ASSERT(lsn != 0);
3414
3415 XFS_STATS_INC(mp, xs_log_force);
3416
3417 lsn = xlog_cil_force_lsn(log, lsn);
3418 if (lsn == NULLCOMMITLSN)
3419 return 0;
3420
3421try_again:
3422 spin_lock(&log->l_icloglock);
3423 iclog = log->l_iclog;
3424 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3425 spin_unlock(&log->l_icloglock);
3426 return -EIO;
3427 }
3428
3429 do {
3430 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3431 iclog = iclog->ic_next;
3432 continue;
3433 }
3434
3435 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3436 spin_unlock(&log->l_icloglock);
3437 return 0;
3438 }
3439
3440 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3441 /*
3442 * We sleep here if we haven't already slept (e.g.
3443 * this is the first time we've looked at the correct
3444 * iclog buf) and the buffer before us is going to
3445 * be sync'ed. The reason for this is that if we
3446 * are doing sync transactions here, by waiting for
3447 * the previous I/O to complete, we can allow a few
3448 * more transactions into this iclog before we close
3449 * it down.
3450 *
3451 * Otherwise, we mark the buffer WANT_SYNC, and bump
3452 * up the refcnt so we can release the log (which
3453 * drops the ref count). The state switch keeps new
3454 * transaction commits from using this buffer. When
3455 * the current commits finish writing into the buffer,
3456 * the refcount will drop to zero and the buffer will
3457 * go out then.
3458 */
3459 if (!already_slept &&
3460 (iclog->ic_prev->ic_state &
3461 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3462 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3463
3464 XFS_STATS_INC(mp, xs_log_force_sleep);
3465
3466 xlog_wait(&iclog->ic_prev->ic_write_wait,
3467 &log->l_icloglock);
3468 if (log_flushed)
3469 *log_flushed = 1;
3470 already_slept = 1;
3471 goto try_again;
3472 }
3473 atomic_inc(&iclog->ic_refcnt);
3474 xlog_state_switch_iclogs(log, iclog, 0);
3475 spin_unlock(&log->l_icloglock);
3476 if (xlog_state_release_iclog(log, iclog))
3477 return -EIO;
3478 if (log_flushed)
3479 *log_flushed = 1;
3480 spin_lock(&log->l_icloglock);
3481 }
3482
3483 if ((flags & XFS_LOG_SYNC) && /* sleep */
3484 !(iclog->ic_state &
3485 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3486 /*
3487 * Don't wait on completion if we know that we've
3488 * gotten a log write error.
3489 */
3490 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3491 spin_unlock(&log->l_icloglock);
3492 return -EIO;
3493 }
3494 XFS_STATS_INC(mp, xs_log_force_sleep);
3495 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3496 /*
3497 * No need to grab the log lock here since we're
3498 * only deciding whether or not to return EIO
3499 * and the memory read should be atomic.
3500 */
3501 if (iclog->ic_state & XLOG_STATE_IOERROR)
3502 return -EIO;
3503
3504 if (log_flushed)
3505 *log_flushed = 1;
3506 } else { /* just return */
3507 spin_unlock(&log->l_icloglock);
3508 }
3509
3510 return 0;
3511 } while (iclog != log->l_iclog);
3512
3513 spin_unlock(&log->l_icloglock);
3514 return 0;
3515}
3516
3517/*
3518 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3519 * about errors or whether the log was flushed or not. This is the normal
3520 * interface to use when trying to unpin items or move the log forward.
3521 */
3522void
3523xfs_log_force_lsn(
3524 xfs_mount_t *mp,
3525 xfs_lsn_t lsn,
3526 uint flags)
3527{
3528 int error;
3529
3530 trace_xfs_log_force(mp, lsn);
3531 error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3532 if (error)
3533 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3534}
3535
3536/*
3537 * Called when we want to mark the current iclog as being ready to sync to
3538 * disk.
3539 */
3540STATIC void
3541xlog_state_want_sync(
3542 struct xlog *log,
3543 struct xlog_in_core *iclog)
3544{
3545 assert_spin_locked(&log->l_icloglock);
3546
3547 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3548 xlog_state_switch_iclogs(log, iclog, 0);
3549 } else {
3550 ASSERT(iclog->ic_state &
3551 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3552 }
3553}
3554
3555
3556/*****************************************************************************
3557 *
3558 * TICKET functions
3559 *
3560 *****************************************************************************
3561 */
3562
3563/*
3564 * Free a used ticket when its refcount falls to zero.
3565 */
3566void
3567xfs_log_ticket_put(
3568 xlog_ticket_t *ticket)
3569{
3570 ASSERT(atomic_read(&ticket->t_ref) > 0);
3571 if (atomic_dec_and_test(&ticket->t_ref))
3572 kmem_zone_free(xfs_log_ticket_zone, ticket);
3573}
3574
3575xlog_ticket_t *
3576xfs_log_ticket_get(
3577 xlog_ticket_t *ticket)
3578{
3579 ASSERT(atomic_read(&ticket->t_ref) > 0);
3580 atomic_inc(&ticket->t_ref);
3581 return ticket;
3582}
3583
3584/*
3585 * Figure out the total log space unit (in bytes) that would be
3586 * required for a log ticket.
3587 */
3588int
3589xfs_log_calc_unit_res(
3590 struct xfs_mount *mp,
3591 int unit_bytes)
3592{
3593 struct xlog *log = mp->m_log;
3594 int iclog_space;
3595 uint num_headers;
3596
3597 /*
3598 * Permanent reservations have up to 'cnt'-1 active log operations
3599 * in the log. A unit in this case is the amount of space for one
3600 * of these log operations. Normal reservations have a cnt of 1
3601 * and their unit amount is the total amount of space required.
3602 *
3603 * The following lines of code account for non-transaction data
3604 * which occupy space in the on-disk log.
3605 *
3606 * Normal form of a transaction is:
3607 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3608 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3609 *
3610 * We need to account for all the leadup data and trailer data
3611 * around the transaction data.
3612 * And then we need to account for the worst case in terms of using
3613 * more space.
3614 * The worst case will happen if:
3615 * - the placement of the transaction happens to be such that the
3616 * roundoff is at its maximum
3617 * - the transaction data is synced before the commit record is synced
3618 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3619 * Therefore the commit record is in its own Log Record.
3620 * This can happen as the commit record is called with its
3621 * own region to xlog_write().
3622 * This then means that in the worst case, roundoff can happen for
3623 * the commit-rec as well.
3624 * The commit-rec is smaller than padding in this scenario and so it is
3625 * not added separately.
3626 */
3627
3628 /* for trans header */
3629 unit_bytes += sizeof(xlog_op_header_t);
3630 unit_bytes += sizeof(xfs_trans_header_t);
3631
3632 /* for start-rec */
3633 unit_bytes += sizeof(xlog_op_header_t);
3634
3635 /*
3636 * for LR headers - the space for data in an iclog is the size minus
3637 * the space used for the headers. If we use the iclog size, then we
3638 * undercalculate the number of headers required.
3639 *
3640 * Furthermore - the addition of op headers for split-recs might
3641 * increase the space required enough to require more log and op
3642 * headers, so take that into account too.
3643 *
3644 * IMPORTANT: This reservation makes the assumption that if this
3645 * transaction is the first in an iclog and hence has the LR headers
3646 * accounted to it, then the remaining space in the iclog is
3647 * exclusively for this transaction. i.e. if the transaction is larger
3648 * than the iclog, it will be the only thing in that iclog.
3649 * Fundamentally, this means we must pass the entire log vector to
3650 * xlog_write to guarantee this.
3651 */
3652 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3653 num_headers = howmany(unit_bytes, iclog_space);
3654
3655 /* for split-recs - ophdrs added when data split over LRs */
3656 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3657
3658 /* add extra header reservations if we overrun */
3659 while (!num_headers ||
3660 howmany(unit_bytes, iclog_space) > num_headers) {
3661 unit_bytes += sizeof(xlog_op_header_t);
3662 num_headers++;
3663 }
3664 unit_bytes += log->l_iclog_hsize * num_headers;
3665
3666 /* for commit-rec LR header - note: padding will subsume the ophdr */
3667 unit_bytes += log->l_iclog_hsize;
3668
3669 /* for roundoff padding for transaction data and one for commit record */
3670 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3671 /* log su roundoff */
3672 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3673 } else {
3674 /* BB roundoff */
3675 unit_bytes += 2 * BBSIZE;
3676 }
3677
3678 return unit_bytes;
3679}
3680
3681/*
3682 * Allocate and initialise a new log ticket.
3683 */
3684struct xlog_ticket *
3685xlog_ticket_alloc(
3686 struct xlog *log,
3687 int unit_bytes,
3688 int cnt,
3689 char client,
3690 bool permanent,
3691 xfs_km_flags_t alloc_flags)
3692{
3693 struct xlog_ticket *tic;
3694 int unit_res;
3695
3696 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3697 if (!tic)
3698 return NULL;
3699
3700 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3701
3702 atomic_set(&tic->t_ref, 1);
3703 tic->t_task = current;
3704 INIT_LIST_HEAD(&tic->t_queue);
3705 tic->t_unit_res = unit_res;
3706 tic->t_curr_res = unit_res;
3707 tic->t_cnt = cnt;
3708 tic->t_ocnt = cnt;
3709 tic->t_tid = prandom_u32();
3710 tic->t_clientid = client;
3711 tic->t_flags = XLOG_TIC_INITED;
3712 tic->t_trans_type = 0;
3713 if (permanent)
3714 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3715
3716 xlog_tic_reset_res(tic);
3717
3718 return tic;
3719}
3720
3721
3722/******************************************************************************
3723 *
3724 * Log debug routines
3725 *
3726 ******************************************************************************
3727 */
3728#if defined(DEBUG)
3729/*
3730 * Make sure that the destination ptr is within the valid data region of
3731 * one of the iclogs. This uses backup pointers stored in a different
3732 * part of the log in case we trash the log structure.
3733 */
3734void
3735xlog_verify_dest_ptr(
3736 struct xlog *log,
3737 void *ptr)
3738{
3739 int i;
3740 int good_ptr = 0;
3741
3742 for (i = 0; i < log->l_iclog_bufs; i++) {
3743 if (ptr >= log->l_iclog_bak[i] &&
3744 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3745 good_ptr++;
3746 }
3747
3748 if (!good_ptr)
3749 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3750}
3751
3752/*
3753 * Check to make sure the grant write head didn't just over lap the tail. If
3754 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3755 * the cycles differ by exactly one and check the byte count.
3756 *
3757 * This check is run unlocked, so can give false positives. Rather than assert
3758 * on failures, use a warn-once flag and a panic tag to allow the admin to
3759 * determine if they want to panic the machine when such an error occurs. For
3760 * debug kernels this will have the same effect as using an assert but, unlinke
3761 * an assert, it can be turned off at runtime.
3762 */
3763STATIC void
3764xlog_verify_grant_tail(
3765 struct xlog *log)
3766{
3767 int tail_cycle, tail_blocks;
3768 int cycle, space;
3769
3770 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3771 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3772 if (tail_cycle != cycle) {
3773 if (cycle - 1 != tail_cycle &&
3774 !(log->l_flags & XLOG_TAIL_WARN)) {
3775 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3776 "%s: cycle - 1 != tail_cycle", __func__);
3777 log->l_flags |= XLOG_TAIL_WARN;
3778 }
3779
3780 if (space > BBTOB(tail_blocks) &&
3781 !(log->l_flags & XLOG_TAIL_WARN)) {
3782 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3783 "%s: space > BBTOB(tail_blocks)", __func__);
3784 log->l_flags |= XLOG_TAIL_WARN;
3785 }
3786 }
3787}
3788
3789/* check if it will fit */
3790STATIC void
3791xlog_verify_tail_lsn(
3792 struct xlog *log,
3793 struct xlog_in_core *iclog,
3794 xfs_lsn_t tail_lsn)
3795{
3796 int blocks;
3797
3798 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3799 blocks =
3800 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3801 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3802 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3803 } else {
3804 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3805
3806 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3807 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3808
3809 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3810 if (blocks < BTOBB(iclog->ic_offset) + 1)
3811 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3812 }
3813} /* xlog_verify_tail_lsn */
3814
3815/*
3816 * Perform a number of checks on the iclog before writing to disk.
3817 *
3818 * 1. Make sure the iclogs are still circular
3819 * 2. Make sure we have a good magic number
3820 * 3. Make sure we don't have magic numbers in the data
3821 * 4. Check fields of each log operation header for:
3822 * A. Valid client identifier
3823 * B. tid ptr value falls in valid ptr space (user space code)
3824 * C. Length in log record header is correct according to the
3825 * individual operation headers within record.
3826 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3827 * log, check the preceding blocks of the physical log to make sure all
3828 * the cycle numbers agree with the current cycle number.
3829 */
3830STATIC void
3831xlog_verify_iclog(
3832 struct xlog *log,
3833 struct xlog_in_core *iclog,
3834 int count,
3835 bool syncing)
3836{
3837 xlog_op_header_t *ophead;
3838 xlog_in_core_t *icptr;
3839 xlog_in_core_2_t *xhdr;
3840 void *base_ptr, *ptr, *p;
3841 ptrdiff_t field_offset;
3842 __uint8_t clientid;
3843 int len, i, j, k, op_len;
3844 int idx;
3845
3846 /* check validity of iclog pointers */
3847 spin_lock(&log->l_icloglock);
3848 icptr = log->l_iclog;
3849 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3850 ASSERT(icptr);
3851
3852 if (icptr != log->l_iclog)
3853 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3854 spin_unlock(&log->l_icloglock);
3855
3856 /* check log magic numbers */
3857 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3858 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3859
3860 base_ptr = ptr = &iclog->ic_header;
3861 p = &iclog->ic_header;
3862 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3863 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3864 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3865 __func__);
3866 }
3867
3868 /* check fields */
3869 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3870 base_ptr = ptr = iclog->ic_datap;
3871 ophead = ptr;
3872 xhdr = iclog->ic_data;
3873 for (i = 0; i < len; i++) {
3874 ophead = ptr;
3875
3876 /* clientid is only 1 byte */
3877 p = &ophead->oh_clientid;
3878 field_offset = p - base_ptr;
3879 if (!syncing || (field_offset & 0x1ff)) {
3880 clientid = ophead->oh_clientid;
3881 } else {
3882 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3883 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3884 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3885 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3886 clientid = xlog_get_client_id(
3887 xhdr[j].hic_xheader.xh_cycle_data[k]);
3888 } else {
3889 clientid = xlog_get_client_id(
3890 iclog->ic_header.h_cycle_data[idx]);
3891 }
3892 }
3893 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3894 xfs_warn(log->l_mp,
3895 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3896 __func__, clientid, ophead,
3897 (unsigned long)field_offset);
3898
3899 /* check length */
3900 p = &ophead->oh_len;
3901 field_offset = p - base_ptr;
3902 if (!syncing || (field_offset & 0x1ff)) {
3903 op_len = be32_to_cpu(ophead->oh_len);
3904 } else {
3905 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3906 (uintptr_t)iclog->ic_datap);
3907 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3908 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3909 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3910 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3911 } else {
3912 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3913 }
3914 }
3915 ptr += sizeof(xlog_op_header_t) + op_len;
3916 }
3917} /* xlog_verify_iclog */
3918#endif
3919
3920/*
3921 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3922 */
3923STATIC int
3924xlog_state_ioerror(
3925 struct xlog *log)
3926{
3927 xlog_in_core_t *iclog, *ic;
3928
3929 iclog = log->l_iclog;
3930 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3931 /*
3932 * Mark all the incore logs IOERROR.
3933 * From now on, no log flushes will result.
3934 */
3935 ic = iclog;
3936 do {
3937 ic->ic_state = XLOG_STATE_IOERROR;
3938 ic = ic->ic_next;
3939 } while (ic != iclog);
3940 return 0;
3941 }
3942 /*
3943 * Return non-zero, if state transition has already happened.
3944 */
3945 return 1;
3946}
3947
3948/*
3949 * This is called from xfs_force_shutdown, when we're forcibly
3950 * shutting down the filesystem, typically because of an IO error.
3951 * Our main objectives here are to make sure that:
3952 * a. if !logerror, flush the logs to disk. Anything modified
3953 * after this is ignored.
3954 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3955 * parties to find out, 'atomically'.
3956 * c. those who're sleeping on log reservations, pinned objects and
3957 * other resources get woken up, and be told the bad news.
3958 * d. nothing new gets queued up after (b) and (c) are done.
3959 *
3960 * Note: for the !logerror case we need to flush the regions held in memory out
3961 * to disk first. This needs to be done before the log is marked as shutdown,
3962 * otherwise the iclog writes will fail.
3963 */
3964int
3965xfs_log_force_umount(
3966 struct xfs_mount *mp,
3967 int logerror)
3968{
3969 struct xlog *log;
3970 int retval;
3971
3972 log = mp->m_log;
3973
3974 /*
3975 * If this happens during log recovery, don't worry about
3976 * locking; the log isn't open for business yet.
3977 */
3978 if (!log ||
3979 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3980 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3981 if (mp->m_sb_bp)
3982 mp->m_sb_bp->b_flags |= XBF_DONE;
3983 return 0;
3984 }
3985
3986 /*
3987 * Somebody could've already done the hard work for us.
3988 * No need to get locks for this.
3989 */
3990 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3991 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3992 return 1;
3993 }
3994
3995 /*
3996 * Flush all the completed transactions to disk before marking the log
3997 * being shut down. We need to do it in this order to ensure that
3998 * completed operations are safely on disk before we shut down, and that
3999 * we don't have to issue any buffer IO after the shutdown flags are set
4000 * to guarantee this.
4001 */
4002 if (!logerror)
4003 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
4004
4005 /*
4006 * mark the filesystem and the as in a shutdown state and wake
4007 * everybody up to tell them the bad news.
4008 */
4009 spin_lock(&log->l_icloglock);
4010 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4011 if (mp->m_sb_bp)
4012 mp->m_sb_bp->b_flags |= XBF_DONE;
4013
4014 /*
4015 * Mark the log and the iclogs with IO error flags to prevent any
4016 * further log IO from being issued or completed.
4017 */
4018 log->l_flags |= XLOG_IO_ERROR;
4019 retval = xlog_state_ioerror(log);
4020 spin_unlock(&log->l_icloglock);
4021
4022 /*
4023 * We don't want anybody waiting for log reservations after this. That
4024 * means we have to wake up everybody queued up on reserveq as well as
4025 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
4026 * we don't enqueue anything once the SHUTDOWN flag is set, and this
4027 * action is protected by the grant locks.
4028 */
4029 xlog_grant_head_wake_all(&log->l_reserve_head);
4030 xlog_grant_head_wake_all(&log->l_write_head);
4031
4032 /*
4033 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4034 * as if the log writes were completed. The abort handling in the log
4035 * item committed callback functions will do this again under lock to
4036 * avoid races.
4037 */
4038 wake_up_all(&log->l_cilp->xc_commit_wait);
4039 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4040
4041#ifdef XFSERRORDEBUG
4042 {
4043 xlog_in_core_t *iclog;
4044
4045 spin_lock(&log->l_icloglock);
4046 iclog = log->l_iclog;
4047 do {
4048 ASSERT(iclog->ic_callback == 0);
4049 iclog = iclog->ic_next;
4050 } while (iclog != log->l_iclog);
4051 spin_unlock(&log->l_icloglock);
4052 }
4053#endif
4054 /* return non-zero if log IOERROR transition had already happened */
4055 return retval;
4056}
4057
4058STATIC int
4059xlog_iclogs_empty(
4060 struct xlog *log)
4061{
4062 xlog_in_core_t *iclog;
4063
4064 iclog = log->l_iclog;
4065 do {
4066 /* endianness does not matter here, zero is zero in
4067 * any language.
4068 */
4069 if (iclog->ic_header.h_num_logops)
4070 return 0;
4071 iclog = iclog->ic_next;
4072 } while (iclog != log->l_iclog);
4073 return 1;
4074}
4075
4076/*
4077 * Verify that an LSN stamped into a piece of metadata is valid. This is
4078 * intended for use in read verifiers on v5 superblocks.
4079 */
4080bool
4081xfs_log_check_lsn(
4082 struct xfs_mount *mp,
4083 xfs_lsn_t lsn)
4084{
4085 struct xlog *log = mp->m_log;
4086 bool valid;
4087
4088 /*
4089 * norecovery mode skips mount-time log processing and unconditionally
4090 * resets the in-core LSN. We can't validate in this mode, but
4091 * modifications are not allowed anyways so just return true.
4092 */
4093 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4094 return true;
4095
4096 /*
4097 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4098 * handled by recovery and thus safe to ignore here.
4099 */
4100 if (lsn == NULLCOMMITLSN)
4101 return true;
4102
4103 valid = xlog_valid_lsn(mp->m_log, lsn);
4104
4105 /* warn the user about what's gone wrong before verifier failure */
4106 if (!valid) {
4107 spin_lock(&log->l_icloglock);
4108 xfs_warn(mp,
4109"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4110"Please unmount and run xfs_repair (>= v4.3) to resolve.",
4111 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4112 log->l_curr_cycle, log->l_curr_block);
4113 spin_unlock(&log->l_icloglock);
4114 }
4115
4116 return valid;
4117}