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