xref: /openbmc/linux/fs/xfs/xfs_log.c (revision 8938c48f)
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 
24 kmem_zone_t	*xfs_log_ticket_zone;
25 
26 /* Local miscellaneous function prototypes */
27 STATIC struct xlog *
28 xlog_alloc_log(
29 	struct xfs_mount	*mp,
30 	struct xfs_buftarg	*log_target,
31 	xfs_daddr_t		blk_offset,
32 	int			num_bblks);
33 STATIC int
34 xlog_space_left(
35 	struct xlog		*log,
36 	atomic64_t		*head);
37 STATIC void
38 xlog_dealloc_log(
39 	struct xlog		*log);
40 
41 /* local state machine functions */
42 STATIC void xlog_state_done_syncing(
43 	struct xlog_in_core	*iclog);
44 STATIC int
45 xlog_state_get_iclog_space(
46 	struct xlog		*log,
47 	int			len,
48 	struct xlog_in_core	**iclog,
49 	struct xlog_ticket	*ticket,
50 	int			*continued_write,
51 	int			*logoffsetp);
52 STATIC void
53 xlog_state_switch_iclogs(
54 	struct xlog		*log,
55 	struct xlog_in_core	*iclog,
56 	int			eventual_size);
57 STATIC void
58 xlog_grant_push_ail(
59 	struct xlog		*log,
60 	int			need_bytes);
61 STATIC void
62 xlog_sync(
63 	struct xlog		*log,
64 	struct xlog_in_core	*iclog);
65 #if defined(DEBUG)
66 STATIC void
67 xlog_verify_dest_ptr(
68 	struct xlog		*log,
69 	void			*ptr);
70 STATIC void
71 xlog_verify_grant_tail(
72 	struct xlog *log);
73 STATIC void
74 xlog_verify_iclog(
75 	struct xlog		*log,
76 	struct xlog_in_core	*iclog,
77 	int			count);
78 STATIC void
79 xlog_verify_tail_lsn(
80 	struct xlog		*log,
81 	struct xlog_in_core	*iclog,
82 	xfs_lsn_t		tail_lsn);
83 #else
84 #define xlog_verify_dest_ptr(a,b)
85 #define xlog_verify_grant_tail(a)
86 #define xlog_verify_iclog(a,b,c)
87 #define xlog_verify_tail_lsn(a,b,c)
88 #endif
89 
90 STATIC int
91 xlog_iclogs_empty(
92 	struct xlog		*log);
93 
94 static int
95 xfs_log_cover(struct xfs_mount *);
96 
97 static void
98 xlog_grant_sub_space(
99 	struct xlog		*log,
100 	atomic64_t		*head,
101 	int			bytes)
102 {
103 	int64_t	head_val = atomic64_read(head);
104 	int64_t new, old;
105 
106 	do {
107 		int	cycle, space;
108 
109 		xlog_crack_grant_head_val(head_val, &cycle, &space);
110 
111 		space -= bytes;
112 		if (space < 0) {
113 			space += log->l_logsize;
114 			cycle--;
115 		}
116 
117 		old = head_val;
118 		new = xlog_assign_grant_head_val(cycle, space);
119 		head_val = atomic64_cmpxchg(head, old, new);
120 	} while (head_val != old);
121 }
122 
123 static void
124 xlog_grant_add_space(
125 	struct xlog		*log,
126 	atomic64_t		*head,
127 	int			bytes)
128 {
129 	int64_t	head_val = atomic64_read(head);
130 	int64_t new, old;
131 
132 	do {
133 		int		tmp;
134 		int		cycle, space;
135 
136 		xlog_crack_grant_head_val(head_val, &cycle, &space);
137 
138 		tmp = log->l_logsize - space;
139 		if (tmp > bytes)
140 			space += bytes;
141 		else {
142 			space = bytes - tmp;
143 			cycle++;
144 		}
145 
146 		old = head_val;
147 		new = xlog_assign_grant_head_val(cycle, space);
148 		head_val = atomic64_cmpxchg(head, old, new);
149 	} while (head_val != old);
150 }
151 
152 STATIC void
153 xlog_grant_head_init(
154 	struct xlog_grant_head	*head)
155 {
156 	xlog_assign_grant_head(&head->grant, 1, 0);
157 	INIT_LIST_HEAD(&head->waiters);
158 	spin_lock_init(&head->lock);
159 }
160 
161 STATIC void
162 xlog_grant_head_wake_all(
163 	struct xlog_grant_head	*head)
164 {
165 	struct xlog_ticket	*tic;
166 
167 	spin_lock(&head->lock);
168 	list_for_each_entry(tic, &head->waiters, t_queue)
169 		wake_up_process(tic->t_task);
170 	spin_unlock(&head->lock);
171 }
172 
173 static inline int
174 xlog_ticket_reservation(
175 	struct xlog		*log,
176 	struct xlog_grant_head	*head,
177 	struct xlog_ticket	*tic)
178 {
179 	if (head == &log->l_write_head) {
180 		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
181 		return tic->t_unit_res;
182 	} else {
183 		if (tic->t_flags & XLOG_TIC_PERM_RESERV)
184 			return tic->t_unit_res * tic->t_cnt;
185 		else
186 			return tic->t_unit_res;
187 	}
188 }
189 
190 STATIC bool
191 xlog_grant_head_wake(
192 	struct xlog		*log,
193 	struct xlog_grant_head	*head,
194 	int			*free_bytes)
195 {
196 	struct xlog_ticket	*tic;
197 	int			need_bytes;
198 	bool			woken_task = false;
199 
200 	list_for_each_entry(tic, &head->waiters, t_queue) {
201 
202 		/*
203 		 * There is a chance that the size of the CIL checkpoints in
204 		 * progress at the last AIL push target calculation resulted in
205 		 * limiting the target to the log head (l_last_sync_lsn) at the
206 		 * time. This may not reflect where the log head is now as the
207 		 * CIL checkpoints may have completed.
208 		 *
209 		 * Hence when we are woken here, it may be that the head of the
210 		 * log that has moved rather than the tail. As the tail didn't
211 		 * move, there still won't be space available for the
212 		 * reservation we require.  However, if the AIL has already
213 		 * pushed to the target defined by the old log head location, we
214 		 * will hang here waiting for something else to update the AIL
215 		 * push target.
216 		 *
217 		 * Therefore, if there isn't space to wake the first waiter on
218 		 * the grant head, we need to push the AIL again to ensure the
219 		 * target reflects both the current log tail and log head
220 		 * position before we wait for the tail to move again.
221 		 */
222 
223 		need_bytes = xlog_ticket_reservation(log, head, tic);
224 		if (*free_bytes < need_bytes) {
225 			if (!woken_task)
226 				xlog_grant_push_ail(log, need_bytes);
227 			return false;
228 		}
229 
230 		*free_bytes -= need_bytes;
231 		trace_xfs_log_grant_wake_up(log, tic);
232 		wake_up_process(tic->t_task);
233 		woken_task = true;
234 	}
235 
236 	return true;
237 }
238 
239 STATIC int
240 xlog_grant_head_wait(
241 	struct xlog		*log,
242 	struct xlog_grant_head	*head,
243 	struct xlog_ticket	*tic,
244 	int			need_bytes) __releases(&head->lock)
245 					    __acquires(&head->lock)
246 {
247 	list_add_tail(&tic->t_queue, &head->waiters);
248 
249 	do {
250 		if (XLOG_FORCED_SHUTDOWN(log))
251 			goto shutdown;
252 		xlog_grant_push_ail(log, need_bytes);
253 
254 		__set_current_state(TASK_UNINTERRUPTIBLE);
255 		spin_unlock(&head->lock);
256 
257 		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
258 
259 		trace_xfs_log_grant_sleep(log, tic);
260 		schedule();
261 		trace_xfs_log_grant_wake(log, tic);
262 
263 		spin_lock(&head->lock);
264 		if (XLOG_FORCED_SHUTDOWN(log))
265 			goto shutdown;
266 	} while (xlog_space_left(log, &head->grant) < need_bytes);
267 
268 	list_del_init(&tic->t_queue);
269 	return 0;
270 shutdown:
271 	list_del_init(&tic->t_queue);
272 	return -EIO;
273 }
274 
275 /*
276  * Atomically get the log space required for a log ticket.
277  *
278  * Once a ticket gets put onto head->waiters, it will only return after the
279  * needed reservation is satisfied.
280  *
281  * This function is structured so that it has a lock free fast path. This is
282  * necessary because every new transaction reservation will come through this
283  * path. Hence any lock will be globally hot if we take it unconditionally on
284  * every pass.
285  *
286  * As tickets are only ever moved on and off head->waiters under head->lock, we
287  * only need to take that lock if we are going to add the ticket to the queue
288  * and sleep. We can avoid taking the lock if the ticket was never added to
289  * head->waiters because the t_queue list head will be empty and we hold the
290  * only reference to it so it can safely be checked unlocked.
291  */
292 STATIC int
293 xlog_grant_head_check(
294 	struct xlog		*log,
295 	struct xlog_grant_head	*head,
296 	struct xlog_ticket	*tic,
297 	int			*need_bytes)
298 {
299 	int			free_bytes;
300 	int			error = 0;
301 
302 	ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
303 
304 	/*
305 	 * If there are other waiters on the queue then give them a chance at
306 	 * logspace before us.  Wake up the first waiters, if we do not wake
307 	 * up all the waiters then go to sleep waiting for more free space,
308 	 * otherwise try to get some space for this transaction.
309 	 */
310 	*need_bytes = xlog_ticket_reservation(log, head, tic);
311 	free_bytes = xlog_space_left(log, &head->grant);
312 	if (!list_empty_careful(&head->waiters)) {
313 		spin_lock(&head->lock);
314 		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
315 		    free_bytes < *need_bytes) {
316 			error = xlog_grant_head_wait(log, head, tic,
317 						     *need_bytes);
318 		}
319 		spin_unlock(&head->lock);
320 	} else if (free_bytes < *need_bytes) {
321 		spin_lock(&head->lock);
322 		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
323 		spin_unlock(&head->lock);
324 	}
325 
326 	return error;
327 }
328 
329 static void
330 xlog_tic_reset_res(xlog_ticket_t *tic)
331 {
332 	tic->t_res_num = 0;
333 	tic->t_res_arr_sum = 0;
334 	tic->t_res_num_ophdrs = 0;
335 }
336 
337 static void
338 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
339 {
340 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
341 		/* add to overflow and start again */
342 		tic->t_res_o_flow += tic->t_res_arr_sum;
343 		tic->t_res_num = 0;
344 		tic->t_res_arr_sum = 0;
345 	}
346 
347 	tic->t_res_arr[tic->t_res_num].r_len = len;
348 	tic->t_res_arr[tic->t_res_num].r_type = type;
349 	tic->t_res_arr_sum += len;
350 	tic->t_res_num++;
351 }
352 
353 bool
354 xfs_log_writable(
355 	struct xfs_mount	*mp)
356 {
357 	/*
358 	 * Never write to the log on norecovery mounts, if the block device is
359 	 * read-only, or if the filesystem is shutdown. Read-only mounts still
360 	 * allow internal writes for log recovery and unmount purposes, so don't
361 	 * restrict that case here.
362 	 */
363 	if (mp->m_flags & XFS_MOUNT_NORECOVERY)
364 		return false;
365 	if (xfs_readonly_buftarg(mp->m_log->l_targ))
366 		return false;
367 	if (XFS_FORCED_SHUTDOWN(mp))
368 		return false;
369 	return true;
370 }
371 
372 /*
373  * Replenish the byte reservation required by moving the grant write head.
374  */
375 int
376 xfs_log_regrant(
377 	struct xfs_mount	*mp,
378 	struct xlog_ticket	*tic)
379 {
380 	struct xlog		*log = mp->m_log;
381 	int			need_bytes;
382 	int			error = 0;
383 
384 	if (XLOG_FORCED_SHUTDOWN(log))
385 		return -EIO;
386 
387 	XFS_STATS_INC(mp, xs_try_logspace);
388 
389 	/*
390 	 * This is a new transaction on the ticket, so we need to change the
391 	 * transaction ID so that the next transaction has a different TID in
392 	 * the log. Just add one to the existing tid so that we can see chains
393 	 * of rolling transactions in the log easily.
394 	 */
395 	tic->t_tid++;
396 
397 	xlog_grant_push_ail(log, tic->t_unit_res);
398 
399 	tic->t_curr_res = tic->t_unit_res;
400 	xlog_tic_reset_res(tic);
401 
402 	if (tic->t_cnt > 0)
403 		return 0;
404 
405 	trace_xfs_log_regrant(log, tic);
406 
407 	error = xlog_grant_head_check(log, &log->l_write_head, tic,
408 				      &need_bytes);
409 	if (error)
410 		goto out_error;
411 
412 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
413 	trace_xfs_log_regrant_exit(log, tic);
414 	xlog_verify_grant_tail(log);
415 	return 0;
416 
417 out_error:
418 	/*
419 	 * If we are failing, make sure the ticket doesn't have any current
420 	 * reservations.  We don't want to add this back when the ticket/
421 	 * transaction gets cancelled.
422 	 */
423 	tic->t_curr_res = 0;
424 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
425 	return error;
426 }
427 
428 /*
429  * Reserve log space and return a ticket corresponding to the reservation.
430  *
431  * Each reservation is going to reserve extra space for a log record header.
432  * When writes happen to the on-disk log, we don't subtract the length of the
433  * log record header from any reservation.  By wasting space in each
434  * reservation, we prevent over allocation problems.
435  */
436 int
437 xfs_log_reserve(
438 	struct xfs_mount	*mp,
439 	int		 	unit_bytes,
440 	int		 	cnt,
441 	struct xlog_ticket	**ticp,
442 	uint8_t		 	client,
443 	bool			permanent)
444 {
445 	struct xlog		*log = mp->m_log;
446 	struct xlog_ticket	*tic;
447 	int			need_bytes;
448 	int			error = 0;
449 
450 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
451 
452 	if (XLOG_FORCED_SHUTDOWN(log))
453 		return -EIO;
454 
455 	XFS_STATS_INC(mp, xs_try_logspace);
456 
457 	ASSERT(*ticp == NULL);
458 	tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
459 	*ticp = tic;
460 
461 	xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
462 					    : tic->t_unit_res);
463 
464 	trace_xfs_log_reserve(log, tic);
465 
466 	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
467 				      &need_bytes);
468 	if (error)
469 		goto out_error;
470 
471 	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
472 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
473 	trace_xfs_log_reserve_exit(log, tic);
474 	xlog_verify_grant_tail(log);
475 	return 0;
476 
477 out_error:
478 	/*
479 	 * If we are failing, make sure the ticket doesn't have any current
480 	 * reservations.  We don't want to add this back when the ticket/
481 	 * transaction gets cancelled.
482 	 */
483 	tic->t_curr_res = 0;
484 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
485 	return error;
486 }
487 
488 static bool
489 __xlog_state_release_iclog(
490 	struct xlog		*log,
491 	struct xlog_in_core	*iclog)
492 {
493 	lockdep_assert_held(&log->l_icloglock);
494 
495 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
496 		/* update tail before writing to iclog */
497 		xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
498 
499 		iclog->ic_state = XLOG_STATE_SYNCING;
500 		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
501 		xlog_verify_tail_lsn(log, iclog, tail_lsn);
502 		/* cycle incremented when incrementing curr_block */
503 		return true;
504 	}
505 
506 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
507 	return false;
508 }
509 
510 /*
511  * Flush iclog to disk if this is the last reference to the given iclog and the
512  * it is in the WANT_SYNC state.
513  */
514 static int
515 xlog_state_release_iclog(
516 	struct xlog		*log,
517 	struct xlog_in_core	*iclog)
518 {
519 	lockdep_assert_held(&log->l_icloglock);
520 
521 	if (iclog->ic_state == XLOG_STATE_IOERROR)
522 		return -EIO;
523 
524 	if (atomic_dec_and_test(&iclog->ic_refcnt) &&
525 	    __xlog_state_release_iclog(log, iclog)) {
526 		spin_unlock(&log->l_icloglock);
527 		xlog_sync(log, iclog);
528 		spin_lock(&log->l_icloglock);
529 	}
530 
531 	return 0;
532 }
533 
534 void
535 xfs_log_release_iclog(
536 	struct xlog_in_core	*iclog)
537 {
538 	struct xlog		*log = iclog->ic_log;
539 	bool			sync = false;
540 
541 	if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) {
542 		if (iclog->ic_state != XLOG_STATE_IOERROR)
543 			sync = __xlog_state_release_iclog(log, iclog);
544 		spin_unlock(&log->l_icloglock);
545 	}
546 
547 	if (sync)
548 		xlog_sync(log, iclog);
549 }
550 
551 /*
552  * Mount a log filesystem
553  *
554  * mp		- ubiquitous xfs mount point structure
555  * log_target	- buftarg of on-disk log device
556  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
557  * num_bblocks	- Number of BBSIZE blocks in on-disk log
558  *
559  * Return error or zero.
560  */
561 int
562 xfs_log_mount(
563 	xfs_mount_t	*mp,
564 	xfs_buftarg_t	*log_target,
565 	xfs_daddr_t	blk_offset,
566 	int		num_bblks)
567 {
568 	bool		fatal = xfs_sb_version_hascrc(&mp->m_sb);
569 	int		error = 0;
570 	int		min_logfsbs;
571 
572 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
573 		xfs_notice(mp, "Mounting V%d Filesystem",
574 			   XFS_SB_VERSION_NUM(&mp->m_sb));
575 	} else {
576 		xfs_notice(mp,
577 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
578 			   XFS_SB_VERSION_NUM(&mp->m_sb));
579 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
580 	}
581 
582 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
583 	if (IS_ERR(mp->m_log)) {
584 		error = PTR_ERR(mp->m_log);
585 		goto out;
586 	}
587 
588 	/*
589 	 * Validate the given log space and drop a critical message via syslog
590 	 * if the log size is too small that would lead to some unexpected
591 	 * situations in transaction log space reservation stage.
592 	 *
593 	 * Note: we can't just reject the mount if the validation fails.  This
594 	 * would mean that people would have to downgrade their kernel just to
595 	 * remedy the situation as there is no way to grow the log (short of
596 	 * black magic surgery with xfs_db).
597 	 *
598 	 * We can, however, reject mounts for CRC format filesystems, as the
599 	 * mkfs binary being used to make the filesystem should never create a
600 	 * filesystem with a log that is too small.
601 	 */
602 	min_logfsbs = xfs_log_calc_minimum_size(mp);
603 
604 	if (mp->m_sb.sb_logblocks < min_logfsbs) {
605 		xfs_warn(mp,
606 		"Log size %d blocks too small, minimum size is %d blocks",
607 			 mp->m_sb.sb_logblocks, min_logfsbs);
608 		error = -EINVAL;
609 	} else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
610 		xfs_warn(mp,
611 		"Log size %d blocks too large, maximum size is %lld blocks",
612 			 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
613 		error = -EINVAL;
614 	} else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
615 		xfs_warn(mp,
616 		"log size %lld bytes too large, maximum size is %lld bytes",
617 			 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
618 			 XFS_MAX_LOG_BYTES);
619 		error = -EINVAL;
620 	} else if (mp->m_sb.sb_logsunit > 1 &&
621 		   mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
622 		xfs_warn(mp,
623 		"log stripe unit %u bytes must be a multiple of block size",
624 			 mp->m_sb.sb_logsunit);
625 		error = -EINVAL;
626 		fatal = true;
627 	}
628 	if (error) {
629 		/*
630 		 * Log check errors are always fatal on v5; or whenever bad
631 		 * metadata leads to a crash.
632 		 */
633 		if (fatal) {
634 			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
635 			ASSERT(0);
636 			goto out_free_log;
637 		}
638 		xfs_crit(mp, "Log size out of supported range.");
639 		xfs_crit(mp,
640 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
641 	}
642 
643 	/*
644 	 * Initialize the AIL now we have a log.
645 	 */
646 	error = xfs_trans_ail_init(mp);
647 	if (error) {
648 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
649 		goto out_free_log;
650 	}
651 	mp->m_log->l_ailp = mp->m_ail;
652 
653 	/*
654 	 * skip log recovery on a norecovery mount.  pretend it all
655 	 * just worked.
656 	 */
657 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
658 		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
659 
660 		if (readonly)
661 			mp->m_flags &= ~XFS_MOUNT_RDONLY;
662 
663 		error = xlog_recover(mp->m_log);
664 
665 		if (readonly)
666 			mp->m_flags |= XFS_MOUNT_RDONLY;
667 		if (error) {
668 			xfs_warn(mp, "log mount/recovery failed: error %d",
669 				error);
670 			xlog_recover_cancel(mp->m_log);
671 			goto out_destroy_ail;
672 		}
673 	}
674 
675 	error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
676 			       "log");
677 	if (error)
678 		goto out_destroy_ail;
679 
680 	/* Normal transactions can now occur */
681 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
682 
683 	/*
684 	 * Now the log has been fully initialised and we know were our
685 	 * space grant counters are, we can initialise the permanent ticket
686 	 * needed for delayed logging to work.
687 	 */
688 	xlog_cil_init_post_recovery(mp->m_log);
689 
690 	return 0;
691 
692 out_destroy_ail:
693 	xfs_trans_ail_destroy(mp);
694 out_free_log:
695 	xlog_dealloc_log(mp->m_log);
696 out:
697 	return error;
698 }
699 
700 /*
701  * Finish the recovery of the file system.  This is separate from the
702  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
703  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
704  * here.
705  *
706  * If we finish recovery successfully, start the background log work. If we are
707  * not doing recovery, then we have a RO filesystem and we don't need to start
708  * it.
709  */
710 int
711 xfs_log_mount_finish(
712 	struct xfs_mount	*mp)
713 {
714 	int	error = 0;
715 	bool	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
716 	bool	recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
717 
718 	if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
719 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
720 		return 0;
721 	} else if (readonly) {
722 		/* Allow unlinked processing to proceed */
723 		mp->m_flags &= ~XFS_MOUNT_RDONLY;
724 	}
725 
726 	/*
727 	 * During the second phase of log recovery, we need iget and
728 	 * iput to behave like they do for an active filesystem.
729 	 * xfs_fs_drop_inode needs to be able to prevent the deletion
730 	 * of inodes before we're done replaying log items on those
731 	 * inodes.  Turn it off immediately after recovery finishes
732 	 * so that we don't leak the quota inodes if subsequent mount
733 	 * activities fail.
734 	 *
735 	 * We let all inodes involved in redo item processing end up on
736 	 * the LRU instead of being evicted immediately so that if we do
737 	 * something to an unlinked inode, the irele won't cause
738 	 * premature truncation and freeing of the inode, which results
739 	 * in log recovery failure.  We have to evict the unreferenced
740 	 * lru inodes after clearing SB_ACTIVE because we don't
741 	 * otherwise clean up the lru if there's a subsequent failure in
742 	 * xfs_mountfs, which leads to us leaking the inodes if nothing
743 	 * else (e.g. quotacheck) references the inodes before the
744 	 * mount failure occurs.
745 	 */
746 	mp->m_super->s_flags |= SB_ACTIVE;
747 	error = xlog_recover_finish(mp->m_log);
748 	if (!error)
749 		xfs_log_work_queue(mp);
750 	mp->m_super->s_flags &= ~SB_ACTIVE;
751 	evict_inodes(mp->m_super);
752 
753 	/*
754 	 * Drain the buffer LRU after log recovery. This is required for v4
755 	 * filesystems to avoid leaving around buffers with NULL verifier ops,
756 	 * but we do it unconditionally to make sure we're always in a clean
757 	 * cache state after mount.
758 	 *
759 	 * Don't push in the error case because the AIL may have pending intents
760 	 * that aren't removed until recovery is cancelled.
761 	 */
762 	if (!error && recovered) {
763 		xfs_log_force(mp, XFS_LOG_SYNC);
764 		xfs_ail_push_all_sync(mp->m_ail);
765 	}
766 	xfs_buftarg_drain(mp->m_ddev_targp);
767 
768 	if (readonly)
769 		mp->m_flags |= XFS_MOUNT_RDONLY;
770 
771 	return error;
772 }
773 
774 /*
775  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
776  * the log.
777  */
778 void
779 xfs_log_mount_cancel(
780 	struct xfs_mount	*mp)
781 {
782 	xlog_recover_cancel(mp->m_log);
783 	xfs_log_unmount(mp);
784 }
785 
786 /*
787  * Wait for the iclog to be written disk, or return an error if the log has been
788  * shut down.
789  */
790 static int
791 xlog_wait_on_iclog(
792 	struct xlog_in_core	*iclog)
793 		__releases(iclog->ic_log->l_icloglock)
794 {
795 	struct xlog		*log = iclog->ic_log;
796 
797 	if (!XLOG_FORCED_SHUTDOWN(log) &&
798 	    iclog->ic_state != XLOG_STATE_ACTIVE &&
799 	    iclog->ic_state != XLOG_STATE_DIRTY) {
800 		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
801 		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
802 	} else {
803 		spin_unlock(&log->l_icloglock);
804 	}
805 
806 	if (XLOG_FORCED_SHUTDOWN(log))
807 		return -EIO;
808 	return 0;
809 }
810 
811 /*
812  * Write out an unmount record using the ticket provided. We have to account for
813  * the data space used in the unmount ticket as this write is not done from a
814  * transaction context that has already done the accounting for us.
815  */
816 static int
817 xlog_write_unmount_record(
818 	struct xlog		*log,
819 	struct xlog_ticket	*ticket,
820 	xfs_lsn_t		*lsn,
821 	uint			flags)
822 {
823 	struct xfs_unmount_log_format ulf = {
824 		.magic = XLOG_UNMOUNT_TYPE,
825 	};
826 	struct xfs_log_iovec reg = {
827 		.i_addr = &ulf,
828 		.i_len = sizeof(ulf),
829 		.i_type = XLOG_REG_TYPE_UNMOUNT,
830 	};
831 	struct xfs_log_vec vec = {
832 		.lv_niovecs = 1,
833 		.lv_iovecp = &reg,
834 	};
835 
836 	/* account for space used by record data */
837 	ticket->t_curr_res -= sizeof(ulf);
838 	return xlog_write(log, &vec, ticket, lsn, NULL, flags, false);
839 }
840 
841 /*
842  * Mark the filesystem clean by writing an unmount record to the head of the
843  * log.
844  */
845 static void
846 xlog_unmount_write(
847 	struct xlog		*log)
848 {
849 	struct xfs_mount	*mp = log->l_mp;
850 	struct xlog_in_core	*iclog;
851 	struct xlog_ticket	*tic = NULL;
852 	xfs_lsn_t		lsn;
853 	uint			flags = XLOG_UNMOUNT_TRANS;
854 	int			error;
855 
856 	error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
857 	if (error)
858 		goto out_err;
859 
860 	error = xlog_write_unmount_record(log, tic, &lsn, flags);
861 	/*
862 	 * At this point, we're umounting anyway, so there's no point in
863 	 * transitioning log state to IOERROR. Just continue...
864 	 */
865 out_err:
866 	if (error)
867 		xfs_alert(mp, "%s: unmount record failed", __func__);
868 
869 	spin_lock(&log->l_icloglock);
870 	iclog = log->l_iclog;
871 	atomic_inc(&iclog->ic_refcnt);
872 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
873 		xlog_state_switch_iclogs(log, iclog, 0);
874 	else
875 		ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
876 		       iclog->ic_state == XLOG_STATE_IOERROR);
877 	error = xlog_state_release_iclog(log, iclog);
878 	xlog_wait_on_iclog(iclog);
879 
880 	if (tic) {
881 		trace_xfs_log_umount_write(log, tic);
882 		xfs_log_ticket_ungrant(log, tic);
883 	}
884 }
885 
886 static void
887 xfs_log_unmount_verify_iclog(
888 	struct xlog		*log)
889 {
890 	struct xlog_in_core	*iclog = log->l_iclog;
891 
892 	do {
893 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
894 		ASSERT(iclog->ic_offset == 0);
895 	} while ((iclog = iclog->ic_next) != log->l_iclog);
896 }
897 
898 /*
899  * Unmount record used to have a string "Unmount filesystem--" in the
900  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
901  * We just write the magic number now since that particular field isn't
902  * currently architecture converted and "Unmount" is a bit foo.
903  * As far as I know, there weren't any dependencies on the old behaviour.
904  */
905 static void
906 xfs_log_unmount_write(
907 	struct xfs_mount	*mp)
908 {
909 	struct xlog		*log = mp->m_log;
910 
911 	if (!xfs_log_writable(mp))
912 		return;
913 
914 	xfs_log_force(mp, XFS_LOG_SYNC);
915 
916 	if (XLOG_FORCED_SHUTDOWN(log))
917 		return;
918 
919 	/*
920 	 * If we think the summary counters are bad, avoid writing the unmount
921 	 * record to force log recovery at next mount, after which the summary
922 	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
923 	 * more details.
924 	 */
925 	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
926 			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
927 		xfs_alert(mp, "%s: will fix summary counters at next mount",
928 				__func__);
929 		return;
930 	}
931 
932 	xfs_log_unmount_verify_iclog(log);
933 	xlog_unmount_write(log);
934 }
935 
936 /*
937  * Empty the log for unmount/freeze.
938  *
939  * To do this, we first need to shut down the background log work so it is not
940  * trying to cover the log as we clean up. We then need to unpin all objects in
941  * the log so we can then flush them out. Once they have completed their IO and
942  * run the callbacks removing themselves from the AIL, we can cover the log.
943  */
944 int
945 xfs_log_quiesce(
946 	struct xfs_mount	*mp)
947 {
948 	cancel_delayed_work_sync(&mp->m_log->l_work);
949 	xfs_log_force(mp, XFS_LOG_SYNC);
950 
951 	/*
952 	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
953 	 * will push it, xfs_buftarg_wait() will not wait for it. Further,
954 	 * xfs_buf_iowait() cannot be used because it was pushed with the
955 	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
956 	 * the IO to complete.
957 	 */
958 	xfs_ail_push_all_sync(mp->m_ail);
959 	xfs_buftarg_wait(mp->m_ddev_targp);
960 	xfs_buf_lock(mp->m_sb_bp);
961 	xfs_buf_unlock(mp->m_sb_bp);
962 
963 	return xfs_log_cover(mp);
964 }
965 
966 void
967 xfs_log_clean(
968 	struct xfs_mount	*mp)
969 {
970 	xfs_log_quiesce(mp);
971 	xfs_log_unmount_write(mp);
972 }
973 
974 /*
975  * Shut down and release the AIL and Log.
976  *
977  * During unmount, we need to ensure we flush all the dirty metadata objects
978  * from the AIL so that the log is empty before we write the unmount record to
979  * the log. Once this is done, we can tear down the AIL and the log.
980  */
981 void
982 xfs_log_unmount(
983 	struct xfs_mount	*mp)
984 {
985 	xfs_log_clean(mp);
986 
987 	xfs_buftarg_drain(mp->m_ddev_targp);
988 
989 	xfs_trans_ail_destroy(mp);
990 
991 	xfs_sysfs_del(&mp->m_log->l_kobj);
992 
993 	xlog_dealloc_log(mp->m_log);
994 }
995 
996 void
997 xfs_log_item_init(
998 	struct xfs_mount	*mp,
999 	struct xfs_log_item	*item,
1000 	int			type,
1001 	const struct xfs_item_ops *ops)
1002 {
1003 	item->li_mountp = mp;
1004 	item->li_ailp = mp->m_ail;
1005 	item->li_type = type;
1006 	item->li_ops = ops;
1007 	item->li_lv = NULL;
1008 
1009 	INIT_LIST_HEAD(&item->li_ail);
1010 	INIT_LIST_HEAD(&item->li_cil);
1011 	INIT_LIST_HEAD(&item->li_bio_list);
1012 	INIT_LIST_HEAD(&item->li_trans);
1013 }
1014 
1015 /*
1016  * Wake up processes waiting for log space after we have moved the log tail.
1017  */
1018 void
1019 xfs_log_space_wake(
1020 	struct xfs_mount	*mp)
1021 {
1022 	struct xlog		*log = mp->m_log;
1023 	int			free_bytes;
1024 
1025 	if (XLOG_FORCED_SHUTDOWN(log))
1026 		return;
1027 
1028 	if (!list_empty_careful(&log->l_write_head.waiters)) {
1029 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1030 
1031 		spin_lock(&log->l_write_head.lock);
1032 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1033 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1034 		spin_unlock(&log->l_write_head.lock);
1035 	}
1036 
1037 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1038 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1039 
1040 		spin_lock(&log->l_reserve_head.lock);
1041 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1042 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1043 		spin_unlock(&log->l_reserve_head.lock);
1044 	}
1045 }
1046 
1047 /*
1048  * Determine if we have a transaction that has gone to disk that needs to be
1049  * covered. To begin the transition to the idle state firstly the log needs to
1050  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1051  * we start attempting to cover the log.
1052  *
1053  * Only if we are then in a state where covering is needed, the caller is
1054  * informed that dummy transactions are required to move the log into the idle
1055  * state.
1056  *
1057  * If there are any items in the AIl or CIL, then we do not want to attempt to
1058  * cover the log as we may be in a situation where there isn't log space
1059  * available to run a dummy transaction and this can lead to deadlocks when the
1060  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1061  * there's no point in running a dummy transaction at this point because we
1062  * can't start trying to idle the log until both the CIL and AIL are empty.
1063  */
1064 static bool
1065 xfs_log_need_covered(
1066 	struct xfs_mount	*mp)
1067 {
1068 	struct xlog		*log = mp->m_log;
1069 	bool			needed = false;
1070 
1071 	if (!xlog_cil_empty(log))
1072 		return false;
1073 
1074 	spin_lock(&log->l_icloglock);
1075 	switch (log->l_covered_state) {
1076 	case XLOG_STATE_COVER_DONE:
1077 	case XLOG_STATE_COVER_DONE2:
1078 	case XLOG_STATE_COVER_IDLE:
1079 		break;
1080 	case XLOG_STATE_COVER_NEED:
1081 	case XLOG_STATE_COVER_NEED2:
1082 		if (xfs_ail_min_lsn(log->l_ailp))
1083 			break;
1084 		if (!xlog_iclogs_empty(log))
1085 			break;
1086 
1087 		needed = true;
1088 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1089 			log->l_covered_state = XLOG_STATE_COVER_DONE;
1090 		else
1091 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1092 		break;
1093 	default:
1094 		needed = true;
1095 		break;
1096 	}
1097 	spin_unlock(&log->l_icloglock);
1098 	return needed;
1099 }
1100 
1101 /*
1102  * Explicitly cover the log. This is similar to background log covering but
1103  * intended for usage in quiesce codepaths. The caller is responsible to ensure
1104  * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1105  * must all be empty.
1106  */
1107 static int
1108 xfs_log_cover(
1109 	struct xfs_mount	*mp)
1110 {
1111 	int			error = 0;
1112 	bool			need_covered;
1113 
1114 	ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1115 	        !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1116 	       XFS_FORCED_SHUTDOWN(mp));
1117 
1118 	if (!xfs_log_writable(mp))
1119 		return 0;
1120 
1121 	/*
1122 	 * xfs_log_need_covered() is not idempotent because it progresses the
1123 	 * state machine if the log requires covering. Therefore, we must call
1124 	 * this function once and use the result until we've issued an sb sync.
1125 	 * Do so first to make that abundantly clear.
1126 	 *
1127 	 * Fall into the covering sequence if the log needs covering or the
1128 	 * mount has lazy superblock accounting to sync to disk. The sb sync
1129 	 * used for covering accumulates the in-core counters, so covering
1130 	 * handles this for us.
1131 	 */
1132 	need_covered = xfs_log_need_covered(mp);
1133 	if (!need_covered && !xfs_sb_version_haslazysbcount(&mp->m_sb))
1134 		return 0;
1135 
1136 	/*
1137 	 * To cover the log, commit the superblock twice (at most) in
1138 	 * independent checkpoints. The first serves as a reference for the
1139 	 * tail pointer. The sync transaction and AIL push empties the AIL and
1140 	 * updates the in-core tail to the LSN of the first checkpoint. The
1141 	 * second commit updates the on-disk tail with the in-core LSN,
1142 	 * covering the log. Push the AIL one more time to leave it empty, as
1143 	 * we found it.
1144 	 */
1145 	do {
1146 		error = xfs_sync_sb(mp, true);
1147 		if (error)
1148 			break;
1149 		xfs_ail_push_all_sync(mp->m_ail);
1150 	} while (xfs_log_need_covered(mp));
1151 
1152 	return error;
1153 }
1154 
1155 /*
1156  * We may be holding the log iclog lock upon entering this routine.
1157  */
1158 xfs_lsn_t
1159 xlog_assign_tail_lsn_locked(
1160 	struct xfs_mount	*mp)
1161 {
1162 	struct xlog		*log = mp->m_log;
1163 	struct xfs_log_item	*lip;
1164 	xfs_lsn_t		tail_lsn;
1165 
1166 	assert_spin_locked(&mp->m_ail->ail_lock);
1167 
1168 	/*
1169 	 * To make sure we always have a valid LSN for the log tail we keep
1170 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1171 	 * and use that when the AIL was empty.
1172 	 */
1173 	lip = xfs_ail_min(mp->m_ail);
1174 	if (lip)
1175 		tail_lsn = lip->li_lsn;
1176 	else
1177 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1178 	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1179 	atomic64_set(&log->l_tail_lsn, tail_lsn);
1180 	return tail_lsn;
1181 }
1182 
1183 xfs_lsn_t
1184 xlog_assign_tail_lsn(
1185 	struct xfs_mount	*mp)
1186 {
1187 	xfs_lsn_t		tail_lsn;
1188 
1189 	spin_lock(&mp->m_ail->ail_lock);
1190 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1191 	spin_unlock(&mp->m_ail->ail_lock);
1192 
1193 	return tail_lsn;
1194 }
1195 
1196 /*
1197  * Return the space in the log between the tail and the head.  The head
1198  * is passed in the cycle/bytes formal parms.  In the special case where
1199  * the reserve head has wrapped passed the tail, this calculation is no
1200  * longer valid.  In this case, just return 0 which means there is no space
1201  * in the log.  This works for all places where this function is called
1202  * with the reserve head.  Of course, if the write head were to ever
1203  * wrap the tail, we should blow up.  Rather than catch this case here,
1204  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1205  *
1206  * This code also handles the case where the reservation head is behind
1207  * the tail.  The details of this case are described below, but the end
1208  * result is that we return the size of the log as the amount of space left.
1209  */
1210 STATIC int
1211 xlog_space_left(
1212 	struct xlog	*log,
1213 	atomic64_t	*head)
1214 {
1215 	int		free_bytes;
1216 	int		tail_bytes;
1217 	int		tail_cycle;
1218 	int		head_cycle;
1219 	int		head_bytes;
1220 
1221 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1222 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1223 	tail_bytes = BBTOB(tail_bytes);
1224 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1225 		free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1226 	else if (tail_cycle + 1 < head_cycle)
1227 		return 0;
1228 	else if (tail_cycle < head_cycle) {
1229 		ASSERT(tail_cycle == (head_cycle - 1));
1230 		free_bytes = tail_bytes - head_bytes;
1231 	} else {
1232 		/*
1233 		 * The reservation head is behind the tail.
1234 		 * In this case we just want to return the size of the
1235 		 * log as the amount of space left.
1236 		 */
1237 		xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1238 		xfs_alert(log->l_mp,
1239 			  "  tail_cycle = %d, tail_bytes = %d",
1240 			  tail_cycle, tail_bytes);
1241 		xfs_alert(log->l_mp,
1242 			  "  GH   cycle = %d, GH   bytes = %d",
1243 			  head_cycle, head_bytes);
1244 		ASSERT(0);
1245 		free_bytes = log->l_logsize;
1246 	}
1247 	return free_bytes;
1248 }
1249 
1250 
1251 static void
1252 xlog_ioend_work(
1253 	struct work_struct	*work)
1254 {
1255 	struct xlog_in_core     *iclog =
1256 		container_of(work, struct xlog_in_core, ic_end_io_work);
1257 	struct xlog		*log = iclog->ic_log;
1258 	int			error;
1259 
1260 	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1261 #ifdef DEBUG
1262 	/* treat writes with injected CRC errors as failed */
1263 	if (iclog->ic_fail_crc)
1264 		error = -EIO;
1265 #endif
1266 
1267 	/*
1268 	 * Race to shutdown the filesystem if we see an error.
1269 	 */
1270 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1271 		xfs_alert(log->l_mp, "log I/O error %d", error);
1272 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1273 	}
1274 
1275 	xlog_state_done_syncing(iclog);
1276 	bio_uninit(&iclog->ic_bio);
1277 
1278 	/*
1279 	 * Drop the lock to signal that we are done. Nothing references the
1280 	 * iclog after this, so an unmount waiting on this lock can now tear it
1281 	 * down safely. As such, it is unsafe to reference the iclog after the
1282 	 * unlock as we could race with it being freed.
1283 	 */
1284 	up(&iclog->ic_sema);
1285 }
1286 
1287 /*
1288  * Return size of each in-core log record buffer.
1289  *
1290  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1291  *
1292  * If the filesystem blocksize is too large, we may need to choose a
1293  * larger size since the directory code currently logs entire blocks.
1294  */
1295 STATIC void
1296 xlog_get_iclog_buffer_size(
1297 	struct xfs_mount	*mp,
1298 	struct xlog		*log)
1299 {
1300 	if (mp->m_logbufs <= 0)
1301 		mp->m_logbufs = XLOG_MAX_ICLOGS;
1302 	if (mp->m_logbsize <= 0)
1303 		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1304 
1305 	log->l_iclog_bufs = mp->m_logbufs;
1306 	log->l_iclog_size = mp->m_logbsize;
1307 
1308 	/*
1309 	 * # headers = size / 32k - one header holds cycles from 32k of data.
1310 	 */
1311 	log->l_iclog_heads =
1312 		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1313 	log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1314 }
1315 
1316 void
1317 xfs_log_work_queue(
1318 	struct xfs_mount        *mp)
1319 {
1320 	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1321 				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1322 }
1323 
1324 /*
1325  * Every sync period we need to unpin all items in the AIL and push them to
1326  * disk. If there is nothing dirty, then we might need to cover the log to
1327  * indicate that the filesystem is idle.
1328  */
1329 static void
1330 xfs_log_worker(
1331 	struct work_struct	*work)
1332 {
1333 	struct xlog		*log = container_of(to_delayed_work(work),
1334 						struct xlog, l_work);
1335 	struct xfs_mount	*mp = log->l_mp;
1336 
1337 	/* dgc: errors ignored - not fatal and nowhere to report them */
1338 	if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1339 		/*
1340 		 * Dump a transaction into the log that contains no real change.
1341 		 * This is needed to stamp the current tail LSN into the log
1342 		 * during the covering operation.
1343 		 *
1344 		 * We cannot use an inode here for this - that will push dirty
1345 		 * state back up into the VFS and then periodic inode flushing
1346 		 * will prevent log covering from making progress. Hence we
1347 		 * synchronously log the superblock instead to ensure the
1348 		 * superblock is immediately unpinned and can be written back.
1349 		 */
1350 		xfs_sync_sb(mp, true);
1351 	} else
1352 		xfs_log_force(mp, 0);
1353 
1354 	/* start pushing all the metadata that is currently dirty */
1355 	xfs_ail_push_all(mp->m_ail);
1356 
1357 	/* queue us up again */
1358 	xfs_log_work_queue(mp);
1359 }
1360 
1361 /*
1362  * This routine initializes some of the log structure for a given mount point.
1363  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1364  * some other stuff may be filled in too.
1365  */
1366 STATIC struct xlog *
1367 xlog_alloc_log(
1368 	struct xfs_mount	*mp,
1369 	struct xfs_buftarg	*log_target,
1370 	xfs_daddr_t		blk_offset,
1371 	int			num_bblks)
1372 {
1373 	struct xlog		*log;
1374 	xlog_rec_header_t	*head;
1375 	xlog_in_core_t		**iclogp;
1376 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1377 	int			i;
1378 	int			error = -ENOMEM;
1379 	uint			log2_size = 0;
1380 
1381 	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1382 	if (!log) {
1383 		xfs_warn(mp, "Log allocation failed: No memory!");
1384 		goto out;
1385 	}
1386 
1387 	log->l_mp	   = mp;
1388 	log->l_targ	   = log_target;
1389 	log->l_logsize     = BBTOB(num_bblks);
1390 	log->l_logBBstart  = blk_offset;
1391 	log->l_logBBsize   = num_bblks;
1392 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1393 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1394 	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1395 
1396 	log->l_prev_block  = -1;
1397 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1398 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1399 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1400 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1401 
1402 	xlog_grant_head_init(&log->l_reserve_head);
1403 	xlog_grant_head_init(&log->l_write_head);
1404 
1405 	error = -EFSCORRUPTED;
1406 	if (xfs_sb_version_hassector(&mp->m_sb)) {
1407 	        log2_size = mp->m_sb.sb_logsectlog;
1408 		if (log2_size < BBSHIFT) {
1409 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1410 				log2_size, BBSHIFT);
1411 			goto out_free_log;
1412 		}
1413 
1414 	        log2_size -= BBSHIFT;
1415 		if (log2_size > mp->m_sectbb_log) {
1416 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1417 				log2_size, mp->m_sectbb_log);
1418 			goto out_free_log;
1419 		}
1420 
1421 		/* for larger sector sizes, must have v2 or external log */
1422 		if (log2_size && log->l_logBBstart > 0 &&
1423 			    !xfs_sb_version_haslogv2(&mp->m_sb)) {
1424 			xfs_warn(mp,
1425 		"log sector size (0x%x) invalid for configuration.",
1426 				log2_size);
1427 			goto out_free_log;
1428 		}
1429 	}
1430 	log->l_sectBBsize = 1 << log2_size;
1431 
1432 	xlog_get_iclog_buffer_size(mp, log);
1433 
1434 	spin_lock_init(&log->l_icloglock);
1435 	init_waitqueue_head(&log->l_flush_wait);
1436 
1437 	iclogp = &log->l_iclog;
1438 	/*
1439 	 * The amount of memory to allocate for the iclog structure is
1440 	 * rather funky due to the way the structure is defined.  It is
1441 	 * done this way so that we can use different sizes for machines
1442 	 * with different amounts of memory.  See the definition of
1443 	 * xlog_in_core_t in xfs_log_priv.h for details.
1444 	 */
1445 	ASSERT(log->l_iclog_size >= 4096);
1446 	for (i = 0; i < log->l_iclog_bufs; i++) {
1447 		int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1448 		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1449 				sizeof(struct bio_vec);
1450 
1451 		iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1452 		if (!iclog)
1453 			goto out_free_iclog;
1454 
1455 		*iclogp = iclog;
1456 		iclog->ic_prev = prev_iclog;
1457 		prev_iclog = iclog;
1458 
1459 		iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1460 						KM_MAYFAIL | KM_ZERO);
1461 		if (!iclog->ic_data)
1462 			goto out_free_iclog;
1463 #ifdef DEBUG
1464 		log->l_iclog_bak[i] = &iclog->ic_header;
1465 #endif
1466 		head = &iclog->ic_header;
1467 		memset(head, 0, sizeof(xlog_rec_header_t));
1468 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1469 		head->h_version = cpu_to_be32(
1470 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1471 		head->h_size = cpu_to_be32(log->l_iclog_size);
1472 		/* new fields */
1473 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1474 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1475 
1476 		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1477 		iclog->ic_state = XLOG_STATE_ACTIVE;
1478 		iclog->ic_log = log;
1479 		atomic_set(&iclog->ic_refcnt, 0);
1480 		spin_lock_init(&iclog->ic_callback_lock);
1481 		INIT_LIST_HEAD(&iclog->ic_callbacks);
1482 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1483 
1484 		init_waitqueue_head(&iclog->ic_force_wait);
1485 		init_waitqueue_head(&iclog->ic_write_wait);
1486 		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1487 		sema_init(&iclog->ic_sema, 1);
1488 
1489 		iclogp = &iclog->ic_next;
1490 	}
1491 	*iclogp = log->l_iclog;			/* complete ring */
1492 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1493 
1494 	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1495 			XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1496 				    WQ_HIGHPRI),
1497 			0, mp->m_super->s_id);
1498 	if (!log->l_ioend_workqueue)
1499 		goto out_free_iclog;
1500 
1501 	error = xlog_cil_init(log);
1502 	if (error)
1503 		goto out_destroy_workqueue;
1504 	return log;
1505 
1506 out_destroy_workqueue:
1507 	destroy_workqueue(log->l_ioend_workqueue);
1508 out_free_iclog:
1509 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1510 		prev_iclog = iclog->ic_next;
1511 		kmem_free(iclog->ic_data);
1512 		kmem_free(iclog);
1513 		if (prev_iclog == log->l_iclog)
1514 			break;
1515 	}
1516 out_free_log:
1517 	kmem_free(log);
1518 out:
1519 	return ERR_PTR(error);
1520 }	/* xlog_alloc_log */
1521 
1522 /*
1523  * Write out the commit record of a transaction associated with the given
1524  * ticket to close off a running log write. Return the lsn of the commit record.
1525  */
1526 int
1527 xlog_commit_record(
1528 	struct xlog		*log,
1529 	struct xlog_ticket	*ticket,
1530 	struct xlog_in_core	**iclog,
1531 	xfs_lsn_t		*lsn)
1532 {
1533 	struct xfs_log_iovec reg = {
1534 		.i_addr = NULL,
1535 		.i_len = 0,
1536 		.i_type = XLOG_REG_TYPE_COMMIT,
1537 	};
1538 	struct xfs_log_vec vec = {
1539 		.lv_niovecs = 1,
1540 		.lv_iovecp = &reg,
1541 	};
1542 	int	error;
1543 
1544 	if (XLOG_FORCED_SHUTDOWN(log))
1545 		return -EIO;
1546 
1547 	error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS,
1548 			   false);
1549 	if (error)
1550 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1551 	return error;
1552 }
1553 
1554 /*
1555  * Compute the LSN that we'd need to push the log tail towards in order to have
1556  * (a) enough on-disk log space to log the number of bytes specified, (b) at
1557  * least 25% of the log space free, and (c) at least 256 blocks free.  If the
1558  * log free space already meets all three thresholds, this function returns
1559  * NULLCOMMITLSN.
1560  */
1561 xfs_lsn_t
1562 xlog_grant_push_threshold(
1563 	struct xlog	*log,
1564 	int		need_bytes)
1565 {
1566 	xfs_lsn_t	threshold_lsn = 0;
1567 	xfs_lsn_t	last_sync_lsn;
1568 	int		free_blocks;
1569 	int		free_bytes;
1570 	int		threshold_block;
1571 	int		threshold_cycle;
1572 	int		free_threshold;
1573 
1574 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1575 
1576 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1577 	free_blocks = BTOBBT(free_bytes);
1578 
1579 	/*
1580 	 * Set the threshold for the minimum number of free blocks in the
1581 	 * log to the maximum of what the caller needs, one quarter of the
1582 	 * log, and 256 blocks.
1583 	 */
1584 	free_threshold = BTOBB(need_bytes);
1585 	free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1586 	free_threshold = max(free_threshold, 256);
1587 	if (free_blocks >= free_threshold)
1588 		return NULLCOMMITLSN;
1589 
1590 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1591 						&threshold_block);
1592 	threshold_block += free_threshold;
1593 	if (threshold_block >= log->l_logBBsize) {
1594 		threshold_block -= log->l_logBBsize;
1595 		threshold_cycle += 1;
1596 	}
1597 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1598 					threshold_block);
1599 	/*
1600 	 * Don't pass in an lsn greater than the lsn of the last
1601 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1602 	 * so that it doesn't change between the compare and the set.
1603 	 */
1604 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1605 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1606 		threshold_lsn = last_sync_lsn;
1607 
1608 	return threshold_lsn;
1609 }
1610 
1611 /*
1612  * Push the tail of the log if we need to do so to maintain the free log space
1613  * thresholds set out by xlog_grant_push_threshold.  We may need to adopt a
1614  * policy which pushes on an lsn which is further along in the log once we
1615  * reach the high water mark.  In this manner, we would be creating a low water
1616  * mark.
1617  */
1618 STATIC void
1619 xlog_grant_push_ail(
1620 	struct xlog	*log,
1621 	int		need_bytes)
1622 {
1623 	xfs_lsn_t	threshold_lsn;
1624 
1625 	threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1626 	if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log))
1627 		return;
1628 
1629 	/*
1630 	 * Get the transaction layer to kick the dirty buffers out to
1631 	 * disk asynchronously. No point in trying to do this if
1632 	 * the filesystem is shutting down.
1633 	 */
1634 	xfs_ail_push(log->l_ailp, threshold_lsn);
1635 }
1636 
1637 /*
1638  * Stamp cycle number in every block
1639  */
1640 STATIC void
1641 xlog_pack_data(
1642 	struct xlog		*log,
1643 	struct xlog_in_core	*iclog,
1644 	int			roundoff)
1645 {
1646 	int			i, j, k;
1647 	int			size = iclog->ic_offset + roundoff;
1648 	__be32			cycle_lsn;
1649 	char			*dp;
1650 
1651 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1652 
1653 	dp = iclog->ic_datap;
1654 	for (i = 0; i < BTOBB(size); i++) {
1655 		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1656 			break;
1657 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1658 		*(__be32 *)dp = cycle_lsn;
1659 		dp += BBSIZE;
1660 	}
1661 
1662 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1663 		xlog_in_core_2_t *xhdr = iclog->ic_data;
1664 
1665 		for ( ; i < BTOBB(size); i++) {
1666 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1667 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1668 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1669 			*(__be32 *)dp = cycle_lsn;
1670 			dp += BBSIZE;
1671 		}
1672 
1673 		for (i = 1; i < log->l_iclog_heads; i++)
1674 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1675 	}
1676 }
1677 
1678 /*
1679  * Calculate the checksum for a log buffer.
1680  *
1681  * This is a little more complicated than it should be because the various
1682  * headers and the actual data are non-contiguous.
1683  */
1684 __le32
1685 xlog_cksum(
1686 	struct xlog		*log,
1687 	struct xlog_rec_header	*rhead,
1688 	char			*dp,
1689 	int			size)
1690 {
1691 	uint32_t		crc;
1692 
1693 	/* first generate the crc for the record header ... */
1694 	crc = xfs_start_cksum_update((char *)rhead,
1695 			      sizeof(struct xlog_rec_header),
1696 			      offsetof(struct xlog_rec_header, h_crc));
1697 
1698 	/* ... then for additional cycle data for v2 logs ... */
1699 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1700 		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1701 		int		i;
1702 		int		xheads;
1703 
1704 		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1705 
1706 		for (i = 1; i < xheads; i++) {
1707 			crc = crc32c(crc, &xhdr[i].hic_xheader,
1708 				     sizeof(struct xlog_rec_ext_header));
1709 		}
1710 	}
1711 
1712 	/* ... and finally for the payload */
1713 	crc = crc32c(crc, dp, size);
1714 
1715 	return xfs_end_cksum(crc);
1716 }
1717 
1718 static void
1719 xlog_bio_end_io(
1720 	struct bio		*bio)
1721 {
1722 	struct xlog_in_core	*iclog = bio->bi_private;
1723 
1724 	queue_work(iclog->ic_log->l_ioend_workqueue,
1725 		   &iclog->ic_end_io_work);
1726 }
1727 
1728 static int
1729 xlog_map_iclog_data(
1730 	struct bio		*bio,
1731 	void			*data,
1732 	size_t			count)
1733 {
1734 	do {
1735 		struct page	*page = kmem_to_page(data);
1736 		unsigned int	off = offset_in_page(data);
1737 		size_t		len = min_t(size_t, count, PAGE_SIZE - off);
1738 
1739 		if (bio_add_page(bio, page, len, off) != len)
1740 			return -EIO;
1741 
1742 		data += len;
1743 		count -= len;
1744 	} while (count);
1745 
1746 	return 0;
1747 }
1748 
1749 STATIC void
1750 xlog_write_iclog(
1751 	struct xlog		*log,
1752 	struct xlog_in_core	*iclog,
1753 	uint64_t		bno,
1754 	unsigned int		count,
1755 	bool			need_flush)
1756 {
1757 	ASSERT(bno < log->l_logBBsize);
1758 
1759 	/*
1760 	 * We lock the iclogbufs here so that we can serialise against I/O
1761 	 * completion during unmount.  We might be processing a shutdown
1762 	 * triggered during unmount, and that can occur asynchronously to the
1763 	 * unmount thread, and hence we need to ensure that completes before
1764 	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1765 	 * across the log IO to archieve that.
1766 	 */
1767 	down(&iclog->ic_sema);
1768 	if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1769 		/*
1770 		 * It would seem logical to return EIO here, but we rely on
1771 		 * the log state machine to propagate I/O errors instead of
1772 		 * doing it here.  We kick of the state machine and unlock
1773 		 * the buffer manually, the code needs to be kept in sync
1774 		 * with the I/O completion path.
1775 		 */
1776 		xlog_state_done_syncing(iclog);
1777 		up(&iclog->ic_sema);
1778 		return;
1779 	}
1780 
1781 	bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1782 	bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1783 	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1784 	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1785 	iclog->ic_bio.bi_private = iclog;
1786 
1787 	/*
1788 	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1789 	 * IOs coming immediately after this one. This prevents the block layer
1790 	 * writeback throttle from throttling log writes behind background
1791 	 * metadata writeback and causing priority inversions.
1792 	 */
1793 	iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC |
1794 				REQ_IDLE | REQ_FUA;
1795 	if (need_flush)
1796 		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1797 
1798 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1799 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1800 		return;
1801 	}
1802 	if (is_vmalloc_addr(iclog->ic_data))
1803 		flush_kernel_vmap_range(iclog->ic_data, count);
1804 
1805 	/*
1806 	 * If this log buffer would straddle the end of the log we will have
1807 	 * to split it up into two bios, so that we can continue at the start.
1808 	 */
1809 	if (bno + BTOBB(count) > log->l_logBBsize) {
1810 		struct bio *split;
1811 
1812 		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1813 				  GFP_NOIO, &fs_bio_set);
1814 		bio_chain(split, &iclog->ic_bio);
1815 		submit_bio(split);
1816 
1817 		/* restart at logical offset zero for the remainder */
1818 		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1819 	}
1820 
1821 	submit_bio(&iclog->ic_bio);
1822 }
1823 
1824 /*
1825  * We need to bump cycle number for the part of the iclog that is
1826  * written to the start of the log. Watch out for the header magic
1827  * number case, though.
1828  */
1829 static void
1830 xlog_split_iclog(
1831 	struct xlog		*log,
1832 	void			*data,
1833 	uint64_t		bno,
1834 	unsigned int		count)
1835 {
1836 	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1837 	unsigned int		i;
1838 
1839 	for (i = split_offset; i < count; i += BBSIZE) {
1840 		uint32_t cycle = get_unaligned_be32(data + i);
1841 
1842 		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1843 			cycle++;
1844 		put_unaligned_be32(cycle, data + i);
1845 	}
1846 }
1847 
1848 static int
1849 xlog_calc_iclog_size(
1850 	struct xlog		*log,
1851 	struct xlog_in_core	*iclog,
1852 	uint32_t		*roundoff)
1853 {
1854 	uint32_t		count_init, count;
1855 	bool			use_lsunit;
1856 
1857 	use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1858 			log->l_mp->m_sb.sb_logsunit > 1;
1859 
1860 	/* Add for LR header */
1861 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1862 
1863 	/* Round out the log write size */
1864 	if (use_lsunit) {
1865 		/* we have a v2 stripe unit to use */
1866 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1867 	} else {
1868 		count = BBTOB(BTOBB(count_init));
1869 	}
1870 
1871 	ASSERT(count >= count_init);
1872 	*roundoff = count - count_init;
1873 
1874 	if (use_lsunit)
1875 		ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1876 	else
1877 		ASSERT(*roundoff < BBTOB(1));
1878 	return count;
1879 }
1880 
1881 /*
1882  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1883  * fashion.  Previously, we should have moved the current iclog
1884  * ptr in the log to point to the next available iclog.  This allows further
1885  * write to continue while this code syncs out an iclog ready to go.
1886  * Before an in-core log can be written out, the data section must be scanned
1887  * to save away the 1st word of each BBSIZE block into the header.  We replace
1888  * it with the current cycle count.  Each BBSIZE block is tagged with the
1889  * cycle count because there in an implicit assumption that drives will
1890  * guarantee that entire 512 byte blocks get written at once.  In other words,
1891  * we can't have part of a 512 byte block written and part not written.  By
1892  * tagging each block, we will know which blocks are valid when recovering
1893  * after an unclean shutdown.
1894  *
1895  * This routine is single threaded on the iclog.  No other thread can be in
1896  * this routine with the same iclog.  Changing contents of iclog can there-
1897  * fore be done without grabbing the state machine lock.  Updating the global
1898  * log will require grabbing the lock though.
1899  *
1900  * The entire log manager uses a logical block numbering scheme.  Only
1901  * xlog_write_iclog knows about the fact that the log may not start with
1902  * block zero on a given device.
1903  */
1904 STATIC void
1905 xlog_sync(
1906 	struct xlog		*log,
1907 	struct xlog_in_core	*iclog)
1908 {
1909 	unsigned int		count;		/* byte count of bwrite */
1910 	unsigned int		roundoff;       /* roundoff to BB or stripe */
1911 	uint64_t		bno;
1912 	unsigned int		size;
1913 	bool			need_flush = true, split = false;
1914 
1915 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1916 
1917 	count = xlog_calc_iclog_size(log, iclog, &roundoff);
1918 
1919 	/* move grant heads by roundoff in sync */
1920 	xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1921 	xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1922 
1923 	/* put cycle number in every block */
1924 	xlog_pack_data(log, iclog, roundoff);
1925 
1926 	/* real byte length */
1927 	size = iclog->ic_offset;
1928 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1929 		size += roundoff;
1930 	iclog->ic_header.h_len = cpu_to_be32(size);
1931 
1932 	XFS_STATS_INC(log->l_mp, xs_log_writes);
1933 	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1934 
1935 	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1936 
1937 	/* Do we need to split this write into 2 parts? */
1938 	if (bno + BTOBB(count) > log->l_logBBsize) {
1939 		xlog_split_iclog(log, &iclog->ic_header, bno, count);
1940 		split = true;
1941 	}
1942 
1943 	/* calculcate the checksum */
1944 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1945 					    iclog->ic_datap, size);
1946 	/*
1947 	 * Intentionally corrupt the log record CRC based on the error injection
1948 	 * frequency, if defined. This facilitates testing log recovery in the
1949 	 * event of torn writes. Hence, set the IOABORT state to abort the log
1950 	 * write on I/O completion and shutdown the fs. The subsequent mount
1951 	 * detects the bad CRC and attempts to recover.
1952 	 */
1953 #ifdef DEBUG
1954 	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1955 		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1956 		iclog->ic_fail_crc = true;
1957 		xfs_warn(log->l_mp,
1958 	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1959 			 be64_to_cpu(iclog->ic_header.h_lsn));
1960 	}
1961 #endif
1962 
1963 	/*
1964 	 * Flush the data device before flushing the log to make sure all meta
1965 	 * data written back from the AIL actually made it to disk before
1966 	 * stamping the new log tail LSN into the log buffer.  For an external
1967 	 * log we need to issue the flush explicitly, and unfortunately
1968 	 * synchronously here; for an internal log we can simply use the block
1969 	 * layer state machine for preflushes.
1970 	 */
1971 	if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1972 		xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1973 		need_flush = false;
1974 	}
1975 
1976 	xlog_verify_iclog(log, iclog, count);
1977 	xlog_write_iclog(log, iclog, bno, count, need_flush);
1978 }
1979 
1980 /*
1981  * Deallocate a log structure
1982  */
1983 STATIC void
1984 xlog_dealloc_log(
1985 	struct xlog	*log)
1986 {
1987 	xlog_in_core_t	*iclog, *next_iclog;
1988 	int		i;
1989 
1990 	xlog_cil_destroy(log);
1991 
1992 	/*
1993 	 * Cycle all the iclogbuf locks to make sure all log IO completion
1994 	 * is done before we tear down these buffers.
1995 	 */
1996 	iclog = log->l_iclog;
1997 	for (i = 0; i < log->l_iclog_bufs; i++) {
1998 		down(&iclog->ic_sema);
1999 		up(&iclog->ic_sema);
2000 		iclog = iclog->ic_next;
2001 	}
2002 
2003 	iclog = log->l_iclog;
2004 	for (i = 0; i < log->l_iclog_bufs; i++) {
2005 		next_iclog = iclog->ic_next;
2006 		kmem_free(iclog->ic_data);
2007 		kmem_free(iclog);
2008 		iclog = next_iclog;
2009 	}
2010 
2011 	log->l_mp->m_log = NULL;
2012 	destroy_workqueue(log->l_ioend_workqueue);
2013 	kmem_free(log);
2014 }
2015 
2016 /*
2017  * Update counters atomically now that memcpy is done.
2018  */
2019 static inline void
2020 xlog_state_finish_copy(
2021 	struct xlog		*log,
2022 	struct xlog_in_core	*iclog,
2023 	int			record_cnt,
2024 	int			copy_bytes)
2025 {
2026 	lockdep_assert_held(&log->l_icloglock);
2027 
2028 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2029 	iclog->ic_offset += copy_bytes;
2030 }
2031 
2032 /*
2033  * print out info relating to regions written which consume
2034  * the reservation
2035  */
2036 void
2037 xlog_print_tic_res(
2038 	struct xfs_mount	*mp,
2039 	struct xlog_ticket	*ticket)
2040 {
2041 	uint i;
2042 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2043 
2044 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
2045 #define REG_TYPE_STR(type, str)	[XLOG_REG_TYPE_##type] = str
2046 	static char *res_type_str[] = {
2047 	    REG_TYPE_STR(BFORMAT, "bformat"),
2048 	    REG_TYPE_STR(BCHUNK, "bchunk"),
2049 	    REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2050 	    REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2051 	    REG_TYPE_STR(IFORMAT, "iformat"),
2052 	    REG_TYPE_STR(ICORE, "icore"),
2053 	    REG_TYPE_STR(IEXT, "iext"),
2054 	    REG_TYPE_STR(IBROOT, "ibroot"),
2055 	    REG_TYPE_STR(ILOCAL, "ilocal"),
2056 	    REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2057 	    REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2058 	    REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2059 	    REG_TYPE_STR(QFORMAT, "qformat"),
2060 	    REG_TYPE_STR(DQUOT, "dquot"),
2061 	    REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2062 	    REG_TYPE_STR(LRHEADER, "LR header"),
2063 	    REG_TYPE_STR(UNMOUNT, "unmount"),
2064 	    REG_TYPE_STR(COMMIT, "commit"),
2065 	    REG_TYPE_STR(TRANSHDR, "trans header"),
2066 	    REG_TYPE_STR(ICREATE, "inode create"),
2067 	    REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2068 	    REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2069 	    REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2070 	    REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2071 	    REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2072 	    REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2073 	};
2074 	BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2075 #undef REG_TYPE_STR
2076 
2077 	xfs_warn(mp, "ticket reservation summary:");
2078 	xfs_warn(mp, "  unit res    = %d bytes",
2079 		 ticket->t_unit_res);
2080 	xfs_warn(mp, "  current res = %d bytes",
2081 		 ticket->t_curr_res);
2082 	xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2083 		 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2084 	xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2085 		 ticket->t_res_num_ophdrs, ophdr_spc);
2086 	xfs_warn(mp, "  ophdr + reg = %u bytes",
2087 		 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2088 	xfs_warn(mp, "  num regions = %u",
2089 		 ticket->t_res_num);
2090 
2091 	for (i = 0; i < ticket->t_res_num; i++) {
2092 		uint r_type = ticket->t_res_arr[i].r_type;
2093 		xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2094 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2095 			    "bad-rtype" : res_type_str[r_type]),
2096 			    ticket->t_res_arr[i].r_len);
2097 	}
2098 }
2099 
2100 /*
2101  * Print a summary of the transaction.
2102  */
2103 void
2104 xlog_print_trans(
2105 	struct xfs_trans	*tp)
2106 {
2107 	struct xfs_mount	*mp = tp->t_mountp;
2108 	struct xfs_log_item	*lip;
2109 
2110 	/* dump core transaction and ticket info */
2111 	xfs_warn(mp, "transaction summary:");
2112 	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2113 	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2114 	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2115 
2116 	xlog_print_tic_res(mp, tp->t_ticket);
2117 
2118 	/* dump each log item */
2119 	list_for_each_entry(lip, &tp->t_items, li_trans) {
2120 		struct xfs_log_vec	*lv = lip->li_lv;
2121 		struct xfs_log_iovec	*vec;
2122 		int			i;
2123 
2124 		xfs_warn(mp, "log item: ");
2125 		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2126 		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
2127 		if (!lv)
2128 			continue;
2129 		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2130 		xfs_warn(mp, "  size	= %d", lv->lv_size);
2131 		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2132 		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2133 
2134 		/* dump each iovec for the log item */
2135 		vec = lv->lv_iovecp;
2136 		for (i = 0; i < lv->lv_niovecs; i++) {
2137 			int dumplen = min(vec->i_len, 32);
2138 
2139 			xfs_warn(mp, "  iovec[%d]", i);
2140 			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2141 			xfs_warn(mp, "    len	= %d", vec->i_len);
2142 			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2143 			xfs_hex_dump(vec->i_addr, dumplen);
2144 
2145 			vec++;
2146 		}
2147 	}
2148 }
2149 
2150 /*
2151  * Calculate the potential space needed by the log vector.  We may need a start
2152  * record, and each region gets its own struct xlog_op_header and may need to be
2153  * double word aligned.
2154  */
2155 static int
2156 xlog_write_calc_vec_length(
2157 	struct xlog_ticket	*ticket,
2158 	struct xfs_log_vec	*log_vector,
2159 	bool			need_start_rec)
2160 {
2161 	struct xfs_log_vec	*lv;
2162 	int			headers = need_start_rec ? 1 : 0;
2163 	int			len = 0;
2164 	int			i;
2165 
2166 	for (lv = log_vector; lv; lv = lv->lv_next) {
2167 		/* we don't write ordered log vectors */
2168 		if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2169 			continue;
2170 
2171 		headers += lv->lv_niovecs;
2172 
2173 		for (i = 0; i < lv->lv_niovecs; i++) {
2174 			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
2175 
2176 			len += vecp->i_len;
2177 			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2178 		}
2179 	}
2180 
2181 	ticket->t_res_num_ophdrs += headers;
2182 	len += headers * sizeof(struct xlog_op_header);
2183 
2184 	return len;
2185 }
2186 
2187 static void
2188 xlog_write_start_rec(
2189 	struct xlog_op_header	*ophdr,
2190 	struct xlog_ticket	*ticket)
2191 {
2192 	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
2193 	ophdr->oh_clientid = ticket->t_clientid;
2194 	ophdr->oh_len = 0;
2195 	ophdr->oh_flags = XLOG_START_TRANS;
2196 	ophdr->oh_res2 = 0;
2197 }
2198 
2199 static xlog_op_header_t *
2200 xlog_write_setup_ophdr(
2201 	struct xlog		*log,
2202 	struct xlog_op_header	*ophdr,
2203 	struct xlog_ticket	*ticket,
2204 	uint			flags)
2205 {
2206 	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2207 	ophdr->oh_clientid = ticket->t_clientid;
2208 	ophdr->oh_res2 = 0;
2209 
2210 	/* are we copying a commit or unmount record? */
2211 	ophdr->oh_flags = flags;
2212 
2213 	/*
2214 	 * We've seen logs corrupted with bad transaction client ids.  This
2215 	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2216 	 * and shut down the filesystem.
2217 	 */
2218 	switch (ophdr->oh_clientid)  {
2219 	case XFS_TRANSACTION:
2220 	case XFS_VOLUME:
2221 	case XFS_LOG:
2222 		break;
2223 	default:
2224 		xfs_warn(log->l_mp,
2225 			"Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2226 			ophdr->oh_clientid, ticket);
2227 		return NULL;
2228 	}
2229 
2230 	return ophdr;
2231 }
2232 
2233 /*
2234  * Set up the parameters of the region copy into the log. This has
2235  * to handle region write split across multiple log buffers - this
2236  * state is kept external to this function so that this code can
2237  * be written in an obvious, self documenting manner.
2238  */
2239 static int
2240 xlog_write_setup_copy(
2241 	struct xlog_ticket	*ticket,
2242 	struct xlog_op_header	*ophdr,
2243 	int			space_available,
2244 	int			space_required,
2245 	int			*copy_off,
2246 	int			*copy_len,
2247 	int			*last_was_partial_copy,
2248 	int			*bytes_consumed)
2249 {
2250 	int			still_to_copy;
2251 
2252 	still_to_copy = space_required - *bytes_consumed;
2253 	*copy_off = *bytes_consumed;
2254 
2255 	if (still_to_copy <= space_available) {
2256 		/* write of region completes here */
2257 		*copy_len = still_to_copy;
2258 		ophdr->oh_len = cpu_to_be32(*copy_len);
2259 		if (*last_was_partial_copy)
2260 			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2261 		*last_was_partial_copy = 0;
2262 		*bytes_consumed = 0;
2263 		return 0;
2264 	}
2265 
2266 	/* partial write of region, needs extra log op header reservation */
2267 	*copy_len = space_available;
2268 	ophdr->oh_len = cpu_to_be32(*copy_len);
2269 	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2270 	if (*last_was_partial_copy)
2271 		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2272 	*bytes_consumed += *copy_len;
2273 	(*last_was_partial_copy)++;
2274 
2275 	/* account for new log op header */
2276 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2277 	ticket->t_res_num_ophdrs++;
2278 
2279 	return sizeof(struct xlog_op_header);
2280 }
2281 
2282 static int
2283 xlog_write_copy_finish(
2284 	struct xlog		*log,
2285 	struct xlog_in_core	*iclog,
2286 	uint			flags,
2287 	int			*record_cnt,
2288 	int			*data_cnt,
2289 	int			*partial_copy,
2290 	int			*partial_copy_len,
2291 	int			log_offset,
2292 	struct xlog_in_core	**commit_iclog)
2293 {
2294 	int			error;
2295 
2296 	if (*partial_copy) {
2297 		/*
2298 		 * This iclog has already been marked WANT_SYNC by
2299 		 * xlog_state_get_iclog_space.
2300 		 */
2301 		spin_lock(&log->l_icloglock);
2302 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2303 		*record_cnt = 0;
2304 		*data_cnt = 0;
2305 		goto release_iclog;
2306 	}
2307 
2308 	*partial_copy = 0;
2309 	*partial_copy_len = 0;
2310 
2311 	if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2312 		/* no more space in this iclog - push it. */
2313 		spin_lock(&log->l_icloglock);
2314 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2315 		*record_cnt = 0;
2316 		*data_cnt = 0;
2317 
2318 		if (iclog->ic_state == XLOG_STATE_ACTIVE)
2319 			xlog_state_switch_iclogs(log, iclog, 0);
2320 		else
2321 			ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2322 			       iclog->ic_state == XLOG_STATE_IOERROR);
2323 		if (!commit_iclog)
2324 			goto release_iclog;
2325 		spin_unlock(&log->l_icloglock);
2326 		ASSERT(flags & XLOG_COMMIT_TRANS);
2327 		*commit_iclog = iclog;
2328 	}
2329 
2330 	return 0;
2331 
2332 release_iclog:
2333 	error = xlog_state_release_iclog(log, iclog);
2334 	spin_unlock(&log->l_icloglock);
2335 	return error;
2336 }
2337 
2338 /*
2339  * Write some region out to in-core log
2340  *
2341  * This will be called when writing externally provided regions or when
2342  * writing out a commit record for a given transaction.
2343  *
2344  * General algorithm:
2345  *	1. Find total length of this write.  This may include adding to the
2346  *		lengths passed in.
2347  *	2. Check whether we violate the tickets reservation.
2348  *	3. While writing to this iclog
2349  *	    A. Reserve as much space in this iclog as can get
2350  *	    B. If this is first write, save away start lsn
2351  *	    C. While writing this region:
2352  *		1. If first write of transaction, write start record
2353  *		2. Write log operation header (header per region)
2354  *		3. Find out if we can fit entire region into this iclog
2355  *		4. Potentially, verify destination memcpy ptr
2356  *		5. Memcpy (partial) region
2357  *		6. If partial copy, release iclog; otherwise, continue
2358  *			copying more regions into current iclog
2359  *	4. Mark want sync bit (in simulation mode)
2360  *	5. Release iclog for potential flush to on-disk log.
2361  *
2362  * ERRORS:
2363  * 1.	Panic if reservation is overrun.  This should never happen since
2364  *	reservation amounts are generated internal to the filesystem.
2365  * NOTES:
2366  * 1. Tickets are single threaded data structures.
2367  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2368  *	syncing routine.  When a single log_write region needs to span
2369  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2370  *	on all log operation writes which don't contain the end of the
2371  *	region.  The XLOG_END_TRANS bit is used for the in-core log
2372  *	operation which contains the end of the continued log_write region.
2373  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2374  *	we don't really know exactly how much space will be used.  As a result,
2375  *	we don't update ic_offset until the end when we know exactly how many
2376  *	bytes have been written out.
2377  */
2378 int
2379 xlog_write(
2380 	struct xlog		*log,
2381 	struct xfs_log_vec	*log_vector,
2382 	struct xlog_ticket	*ticket,
2383 	xfs_lsn_t		*start_lsn,
2384 	struct xlog_in_core	**commit_iclog,
2385 	uint			flags,
2386 	bool			need_start_rec)
2387 {
2388 	struct xlog_in_core	*iclog = NULL;
2389 	struct xfs_log_vec	*lv = log_vector;
2390 	struct xfs_log_iovec	*vecp = lv->lv_iovecp;
2391 	int			index = 0;
2392 	int			len;
2393 	int			partial_copy = 0;
2394 	int			partial_copy_len = 0;
2395 	int			contwr = 0;
2396 	int			record_cnt = 0;
2397 	int			data_cnt = 0;
2398 	int			error = 0;
2399 
2400 	/*
2401 	 * If this is a commit or unmount transaction, we don't need a start
2402 	 * record to be written.  We do, however, have to account for the
2403 	 * commit or unmount header that gets written. Hence we always have
2404 	 * to account for an extra xlog_op_header here.
2405 	 */
2406 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2407 	if (ticket->t_curr_res < 0) {
2408 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2409 		     "ctx ticket reservation ran out. Need to up reservation");
2410 		xlog_print_tic_res(log->l_mp, ticket);
2411 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2412 	}
2413 
2414 	len = xlog_write_calc_vec_length(ticket, log_vector, need_start_rec);
2415 	*start_lsn = 0;
2416 	while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2417 		void		*ptr;
2418 		int		log_offset;
2419 
2420 		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2421 						   &contwr, &log_offset);
2422 		if (error)
2423 			return error;
2424 
2425 		ASSERT(log_offset <= iclog->ic_size - 1);
2426 		ptr = iclog->ic_datap + log_offset;
2427 
2428 		/* start_lsn is the first lsn written to. That's all we need. */
2429 		if (!*start_lsn)
2430 			*start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2431 
2432 		/*
2433 		 * This loop writes out as many regions as can fit in the amount
2434 		 * of space which was allocated by xlog_state_get_iclog_space().
2435 		 */
2436 		while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2437 			struct xfs_log_iovec	*reg;
2438 			struct xlog_op_header	*ophdr;
2439 			int			copy_len;
2440 			int			copy_off;
2441 			bool			ordered = false;
2442 
2443 			/* ordered log vectors have no regions to write */
2444 			if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2445 				ASSERT(lv->lv_niovecs == 0);
2446 				ordered = true;
2447 				goto next_lv;
2448 			}
2449 
2450 			reg = &vecp[index];
2451 			ASSERT(reg->i_len % sizeof(int32_t) == 0);
2452 			ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2453 
2454 			/*
2455 			 * Before we start formatting log vectors, we need to
2456 			 * write a start record. Only do this for the first
2457 			 * iclog we write to.
2458 			 */
2459 			if (need_start_rec) {
2460 				xlog_write_start_rec(ptr, ticket);
2461 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2462 						sizeof(struct xlog_op_header));
2463 			}
2464 
2465 			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2466 			if (!ophdr)
2467 				return -EIO;
2468 
2469 			xlog_write_adv_cnt(&ptr, &len, &log_offset,
2470 					   sizeof(struct xlog_op_header));
2471 
2472 			len += xlog_write_setup_copy(ticket, ophdr,
2473 						     iclog->ic_size-log_offset,
2474 						     reg->i_len,
2475 						     &copy_off, &copy_len,
2476 						     &partial_copy,
2477 						     &partial_copy_len);
2478 			xlog_verify_dest_ptr(log, ptr);
2479 
2480 			/*
2481 			 * Copy region.
2482 			 *
2483 			 * Unmount records just log an opheader, so can have
2484 			 * empty payloads with no data region to copy. Hence we
2485 			 * only copy the payload if the vector says it has data
2486 			 * to copy.
2487 			 */
2488 			ASSERT(copy_len >= 0);
2489 			if (copy_len > 0) {
2490 				memcpy(ptr, reg->i_addr + copy_off, copy_len);
2491 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2492 						   copy_len);
2493 			}
2494 			copy_len += sizeof(struct xlog_op_header);
2495 			record_cnt++;
2496 			if (need_start_rec) {
2497 				copy_len += sizeof(struct xlog_op_header);
2498 				record_cnt++;
2499 				need_start_rec = false;
2500 			}
2501 			data_cnt += contwr ? copy_len : 0;
2502 
2503 			error = xlog_write_copy_finish(log, iclog, flags,
2504 						       &record_cnt, &data_cnt,
2505 						       &partial_copy,
2506 						       &partial_copy_len,
2507 						       log_offset,
2508 						       commit_iclog);
2509 			if (error)
2510 				return error;
2511 
2512 			/*
2513 			 * if we had a partial copy, we need to get more iclog
2514 			 * space but we don't want to increment the region
2515 			 * index because there is still more is this region to
2516 			 * write.
2517 			 *
2518 			 * If we completed writing this region, and we flushed
2519 			 * the iclog (indicated by resetting of the record
2520 			 * count), then we also need to get more log space. If
2521 			 * this was the last record, though, we are done and
2522 			 * can just return.
2523 			 */
2524 			if (partial_copy)
2525 				break;
2526 
2527 			if (++index == lv->lv_niovecs) {
2528 next_lv:
2529 				lv = lv->lv_next;
2530 				index = 0;
2531 				if (lv)
2532 					vecp = lv->lv_iovecp;
2533 			}
2534 			if (record_cnt == 0 && !ordered) {
2535 				if (!lv)
2536 					return 0;
2537 				break;
2538 			}
2539 		}
2540 	}
2541 
2542 	ASSERT(len == 0);
2543 
2544 	spin_lock(&log->l_icloglock);
2545 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2546 	if (commit_iclog) {
2547 		ASSERT(flags & XLOG_COMMIT_TRANS);
2548 		*commit_iclog = iclog;
2549 	} else {
2550 		error = xlog_state_release_iclog(log, iclog);
2551 	}
2552 	spin_unlock(&log->l_icloglock);
2553 
2554 	return error;
2555 }
2556 
2557 static void
2558 xlog_state_activate_iclog(
2559 	struct xlog_in_core	*iclog,
2560 	int			*iclogs_changed)
2561 {
2562 	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2563 
2564 	/*
2565 	 * If the number of ops in this iclog indicate it just contains the
2566 	 * dummy transaction, we can change state into IDLE (the second time
2567 	 * around). Otherwise we should change the state into NEED a dummy.
2568 	 * We don't need to cover the dummy.
2569 	 */
2570 	if (*iclogs_changed == 0 &&
2571 	    iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2572 		*iclogs_changed = 1;
2573 	} else {
2574 		/*
2575 		 * We have two dirty iclogs so start over.  This could also be
2576 		 * num of ops indicating this is not the dummy going out.
2577 		 */
2578 		*iclogs_changed = 2;
2579 	}
2580 
2581 	iclog->ic_state	= XLOG_STATE_ACTIVE;
2582 	iclog->ic_offset = 0;
2583 	iclog->ic_header.h_num_logops = 0;
2584 	memset(iclog->ic_header.h_cycle_data, 0,
2585 		sizeof(iclog->ic_header.h_cycle_data));
2586 	iclog->ic_header.h_lsn = 0;
2587 }
2588 
2589 /*
2590  * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2591  * ACTIVE after iclog I/O has completed.
2592  */
2593 static void
2594 xlog_state_activate_iclogs(
2595 	struct xlog		*log,
2596 	int			*iclogs_changed)
2597 {
2598 	struct xlog_in_core	*iclog = log->l_iclog;
2599 
2600 	do {
2601 		if (iclog->ic_state == XLOG_STATE_DIRTY)
2602 			xlog_state_activate_iclog(iclog, iclogs_changed);
2603 		/*
2604 		 * The ordering of marking iclogs ACTIVE must be maintained, so
2605 		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2606 		 */
2607 		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2608 			break;
2609 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2610 }
2611 
2612 static int
2613 xlog_covered_state(
2614 	int			prev_state,
2615 	int			iclogs_changed)
2616 {
2617 	/*
2618 	 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2619 	 * wrote the first covering record (DONE). We go to IDLE if we just
2620 	 * wrote the second covering record (DONE2) and remain in IDLE until a
2621 	 * non-covering write occurs.
2622 	 */
2623 	switch (prev_state) {
2624 	case XLOG_STATE_COVER_IDLE:
2625 		if (iclogs_changed == 1)
2626 			return XLOG_STATE_COVER_IDLE;
2627 	case XLOG_STATE_COVER_NEED:
2628 	case XLOG_STATE_COVER_NEED2:
2629 		break;
2630 	case XLOG_STATE_COVER_DONE:
2631 		if (iclogs_changed == 1)
2632 			return XLOG_STATE_COVER_NEED2;
2633 		break;
2634 	case XLOG_STATE_COVER_DONE2:
2635 		if (iclogs_changed == 1)
2636 			return XLOG_STATE_COVER_IDLE;
2637 		break;
2638 	default:
2639 		ASSERT(0);
2640 	}
2641 
2642 	return XLOG_STATE_COVER_NEED;
2643 }
2644 
2645 STATIC void
2646 xlog_state_clean_iclog(
2647 	struct xlog		*log,
2648 	struct xlog_in_core	*dirty_iclog)
2649 {
2650 	int			iclogs_changed = 0;
2651 
2652 	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2653 
2654 	xlog_state_activate_iclogs(log, &iclogs_changed);
2655 	wake_up_all(&dirty_iclog->ic_force_wait);
2656 
2657 	if (iclogs_changed) {
2658 		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2659 				iclogs_changed);
2660 	}
2661 }
2662 
2663 STATIC xfs_lsn_t
2664 xlog_get_lowest_lsn(
2665 	struct xlog		*log)
2666 {
2667 	struct xlog_in_core	*iclog = log->l_iclog;
2668 	xfs_lsn_t		lowest_lsn = 0, lsn;
2669 
2670 	do {
2671 		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2672 		    iclog->ic_state == XLOG_STATE_DIRTY)
2673 			continue;
2674 
2675 		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2676 		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2677 			lowest_lsn = lsn;
2678 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2679 
2680 	return lowest_lsn;
2681 }
2682 
2683 /*
2684  * Completion of a iclog IO does not imply that a transaction has completed, as
2685  * transactions can be large enough to span many iclogs. We cannot change the
2686  * tail of the log half way through a transaction as this may be the only
2687  * transaction in the log and moving the tail to point to the middle of it
2688  * will prevent recovery from finding the start of the transaction. Hence we
2689  * should only update the last_sync_lsn if this iclog contains transaction
2690  * completion callbacks on it.
2691  *
2692  * We have to do this before we drop the icloglock to ensure we are the only one
2693  * that can update it.
2694  *
2695  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2696  * the reservation grant head pushing. This is due to the fact that the push
2697  * target is bound by the current last_sync_lsn value. Hence if we have a large
2698  * amount of log space bound up in this committing transaction then the
2699  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2700  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2701  * should push the AIL to ensure the push target (and hence the grant head) is
2702  * no longer bound by the old log head location and can move forwards and make
2703  * progress again.
2704  */
2705 static void
2706 xlog_state_set_callback(
2707 	struct xlog		*log,
2708 	struct xlog_in_core	*iclog,
2709 	xfs_lsn_t		header_lsn)
2710 {
2711 	iclog->ic_state = XLOG_STATE_CALLBACK;
2712 
2713 	ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2714 			   header_lsn) <= 0);
2715 
2716 	if (list_empty_careful(&iclog->ic_callbacks))
2717 		return;
2718 
2719 	atomic64_set(&log->l_last_sync_lsn, header_lsn);
2720 	xlog_grant_push_ail(log, 0);
2721 }
2722 
2723 /*
2724  * Return true if we need to stop processing, false to continue to the next
2725  * iclog. The caller will need to run callbacks if the iclog is returned in the
2726  * XLOG_STATE_CALLBACK state.
2727  */
2728 static bool
2729 xlog_state_iodone_process_iclog(
2730 	struct xlog		*log,
2731 	struct xlog_in_core	*iclog,
2732 	bool			*ioerror)
2733 {
2734 	xfs_lsn_t		lowest_lsn;
2735 	xfs_lsn_t		header_lsn;
2736 
2737 	switch (iclog->ic_state) {
2738 	case XLOG_STATE_ACTIVE:
2739 	case XLOG_STATE_DIRTY:
2740 		/*
2741 		 * Skip all iclogs in the ACTIVE & DIRTY states:
2742 		 */
2743 		return false;
2744 	case XLOG_STATE_IOERROR:
2745 		/*
2746 		 * Between marking a filesystem SHUTDOWN and stopping the log,
2747 		 * we do flush all iclogs to disk (if there wasn't a log I/O
2748 		 * error). So, we do want things to go smoothly in case of just
2749 		 * a SHUTDOWN w/o a LOG_IO_ERROR.
2750 		 */
2751 		*ioerror = true;
2752 		return false;
2753 	case XLOG_STATE_DONE_SYNC:
2754 		/*
2755 		 * Now that we have an iclog that is in the DONE_SYNC state, do
2756 		 * one more check here to see if we have chased our tail around.
2757 		 * If this is not the lowest lsn iclog, then we will leave it
2758 		 * for another completion to process.
2759 		 */
2760 		header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2761 		lowest_lsn = xlog_get_lowest_lsn(log);
2762 		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2763 			return false;
2764 		xlog_state_set_callback(log, iclog, header_lsn);
2765 		return false;
2766 	default:
2767 		/*
2768 		 * Can only perform callbacks in order.  Since this iclog is not
2769 		 * in the DONE_SYNC state, we skip the rest and just try to
2770 		 * clean up.
2771 		 */
2772 		return true;
2773 	}
2774 }
2775 
2776 /*
2777  * Keep processing entries in the iclog callback list until we come around and
2778  * it is empty.  We need to atomically see that the list is empty and change the
2779  * state to DIRTY so that we don't miss any more callbacks being added.
2780  *
2781  * This function is called with the icloglock held and returns with it held. We
2782  * drop it while running callbacks, however, as holding it over thousands of
2783  * callbacks is unnecessary and causes excessive contention if we do.
2784  */
2785 static void
2786 xlog_state_do_iclog_callbacks(
2787 	struct xlog		*log,
2788 	struct xlog_in_core	*iclog)
2789 		__releases(&log->l_icloglock)
2790 		__acquires(&log->l_icloglock)
2791 {
2792 	spin_unlock(&log->l_icloglock);
2793 	spin_lock(&iclog->ic_callback_lock);
2794 	while (!list_empty(&iclog->ic_callbacks)) {
2795 		LIST_HEAD(tmp);
2796 
2797 		list_splice_init(&iclog->ic_callbacks, &tmp);
2798 
2799 		spin_unlock(&iclog->ic_callback_lock);
2800 		xlog_cil_process_committed(&tmp);
2801 		spin_lock(&iclog->ic_callback_lock);
2802 	}
2803 
2804 	/*
2805 	 * Pick up the icloglock while still holding the callback lock so we
2806 	 * serialise against anyone trying to add more callbacks to this iclog
2807 	 * now we've finished processing.
2808 	 */
2809 	spin_lock(&log->l_icloglock);
2810 	spin_unlock(&iclog->ic_callback_lock);
2811 }
2812 
2813 STATIC void
2814 xlog_state_do_callback(
2815 	struct xlog		*log)
2816 {
2817 	struct xlog_in_core	*iclog;
2818 	struct xlog_in_core	*first_iclog;
2819 	bool			cycled_icloglock;
2820 	bool			ioerror;
2821 	int			flushcnt = 0;
2822 	int			repeats = 0;
2823 
2824 	spin_lock(&log->l_icloglock);
2825 	do {
2826 		/*
2827 		 * Scan all iclogs starting with the one pointed to by the
2828 		 * log.  Reset this starting point each time the log is
2829 		 * unlocked (during callbacks).
2830 		 *
2831 		 * Keep looping through iclogs until one full pass is made
2832 		 * without running any callbacks.
2833 		 */
2834 		first_iclog = log->l_iclog;
2835 		iclog = log->l_iclog;
2836 		cycled_icloglock = false;
2837 		ioerror = false;
2838 		repeats++;
2839 
2840 		do {
2841 			if (xlog_state_iodone_process_iclog(log, iclog,
2842 							&ioerror))
2843 				break;
2844 
2845 			if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2846 			    iclog->ic_state != XLOG_STATE_IOERROR) {
2847 				iclog = iclog->ic_next;
2848 				continue;
2849 			}
2850 
2851 			/*
2852 			 * Running callbacks will drop the icloglock which means
2853 			 * we'll have to run at least one more complete loop.
2854 			 */
2855 			cycled_icloglock = true;
2856 			xlog_state_do_iclog_callbacks(log, iclog);
2857 			if (XLOG_FORCED_SHUTDOWN(log))
2858 				wake_up_all(&iclog->ic_force_wait);
2859 			else
2860 				xlog_state_clean_iclog(log, iclog);
2861 			iclog = iclog->ic_next;
2862 		} while (first_iclog != iclog);
2863 
2864 		if (repeats > 5000) {
2865 			flushcnt += repeats;
2866 			repeats = 0;
2867 			xfs_warn(log->l_mp,
2868 				"%s: possible infinite loop (%d iterations)",
2869 				__func__, flushcnt);
2870 		}
2871 	} while (!ioerror && cycled_icloglock);
2872 
2873 	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2874 	    log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2875 		wake_up_all(&log->l_flush_wait);
2876 
2877 	spin_unlock(&log->l_icloglock);
2878 }
2879 
2880 
2881 /*
2882  * Finish transitioning this iclog to the dirty state.
2883  *
2884  * Make sure that we completely execute this routine only when this is
2885  * the last call to the iclog.  There is a good chance that iclog flushes,
2886  * when we reach the end of the physical log, get turned into 2 separate
2887  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2888  * routine.  By using the reference count bwritecnt, we guarantee that only
2889  * the second completion goes through.
2890  *
2891  * Callbacks could take time, so they are done outside the scope of the
2892  * global state machine log lock.
2893  */
2894 STATIC void
2895 xlog_state_done_syncing(
2896 	struct xlog_in_core	*iclog)
2897 {
2898 	struct xlog		*log = iclog->ic_log;
2899 
2900 	spin_lock(&log->l_icloglock);
2901 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2902 
2903 	/*
2904 	 * If we got an error, either on the first buffer, or in the case of
2905 	 * split log writes, on the second, we shut down the file system and
2906 	 * no iclogs should ever be attempted to be written to disk again.
2907 	 */
2908 	if (!XLOG_FORCED_SHUTDOWN(log)) {
2909 		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2910 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2911 	}
2912 
2913 	/*
2914 	 * Someone could be sleeping prior to writing out the next
2915 	 * iclog buffer, we wake them all, one will get to do the
2916 	 * I/O, the others get to wait for the result.
2917 	 */
2918 	wake_up_all(&iclog->ic_write_wait);
2919 	spin_unlock(&log->l_icloglock);
2920 	xlog_state_do_callback(log);
2921 }
2922 
2923 /*
2924  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2925  * sleep.  We wait on the flush queue on the head iclog as that should be
2926  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2927  * we will wait here and all new writes will sleep until a sync completes.
2928  *
2929  * The in-core logs are used in a circular fashion. They are not used
2930  * out-of-order even when an iclog past the head is free.
2931  *
2932  * return:
2933  *	* log_offset where xlog_write() can start writing into the in-core
2934  *		log's data space.
2935  *	* in-core log pointer to which xlog_write() should write.
2936  *	* boolean indicating this is a continued write to an in-core log.
2937  *		If this is the last write, then the in-core log's offset field
2938  *		needs to be incremented, depending on the amount of data which
2939  *		is copied.
2940  */
2941 STATIC int
2942 xlog_state_get_iclog_space(
2943 	struct xlog		*log,
2944 	int			len,
2945 	struct xlog_in_core	**iclogp,
2946 	struct xlog_ticket	*ticket,
2947 	int			*continued_write,
2948 	int			*logoffsetp)
2949 {
2950 	int		  log_offset;
2951 	xlog_rec_header_t *head;
2952 	xlog_in_core_t	  *iclog;
2953 
2954 restart:
2955 	spin_lock(&log->l_icloglock);
2956 	if (XLOG_FORCED_SHUTDOWN(log)) {
2957 		spin_unlock(&log->l_icloglock);
2958 		return -EIO;
2959 	}
2960 
2961 	iclog = log->l_iclog;
2962 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2963 		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2964 
2965 		/* Wait for log writes to have flushed */
2966 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2967 		goto restart;
2968 	}
2969 
2970 	head = &iclog->ic_header;
2971 
2972 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2973 	log_offset = iclog->ic_offset;
2974 
2975 	/* On the 1st write to an iclog, figure out lsn.  This works
2976 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2977 	 * committing to.  If the offset is set, that's how many blocks
2978 	 * must be written.
2979 	 */
2980 	if (log_offset == 0) {
2981 		ticket->t_curr_res -= log->l_iclog_hsize;
2982 		xlog_tic_add_region(ticket,
2983 				    log->l_iclog_hsize,
2984 				    XLOG_REG_TYPE_LRHEADER);
2985 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2986 		head->h_lsn = cpu_to_be64(
2987 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2988 		ASSERT(log->l_curr_block >= 0);
2989 	}
2990 
2991 	/* If there is enough room to write everything, then do it.  Otherwise,
2992 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2993 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2994 	 * until you know exactly how many bytes get copied.  Therefore, wait
2995 	 * until later to update ic_offset.
2996 	 *
2997 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2998 	 * can fit into remaining data section.
2999 	 */
3000 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3001 		int		error = 0;
3002 
3003 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3004 
3005 		/*
3006 		 * If we are the only one writing to this iclog, sync it to
3007 		 * disk.  We need to do an atomic compare and decrement here to
3008 		 * avoid racing with concurrent atomic_dec_and_lock() calls in
3009 		 * xlog_state_release_iclog() when there is more than one
3010 		 * reference to the iclog.
3011 		 */
3012 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3013 			error = xlog_state_release_iclog(log, iclog);
3014 		spin_unlock(&log->l_icloglock);
3015 		if (error)
3016 			return error;
3017 		goto restart;
3018 	}
3019 
3020 	/* Do we have enough room to write the full amount in the remainder
3021 	 * of this iclog?  Or must we continue a write on the next iclog and
3022 	 * mark this iclog as completely taken?  In the case where we switch
3023 	 * iclogs (to mark it taken), this particular iclog will release/sync
3024 	 * to disk in xlog_write().
3025 	 */
3026 	if (len <= iclog->ic_size - iclog->ic_offset) {
3027 		*continued_write = 0;
3028 		iclog->ic_offset += len;
3029 	} else {
3030 		*continued_write = 1;
3031 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3032 	}
3033 	*iclogp = iclog;
3034 
3035 	ASSERT(iclog->ic_offset <= iclog->ic_size);
3036 	spin_unlock(&log->l_icloglock);
3037 
3038 	*logoffsetp = log_offset;
3039 	return 0;
3040 }
3041 
3042 /*
3043  * The first cnt-1 times a ticket goes through here we don't need to move the
3044  * grant write head because the permanent reservation has reserved cnt times the
3045  * unit amount.  Release part of current permanent unit reservation and reset
3046  * current reservation to be one units worth.  Also move grant reservation head
3047  * forward.
3048  */
3049 void
3050 xfs_log_ticket_regrant(
3051 	struct xlog		*log,
3052 	struct xlog_ticket	*ticket)
3053 {
3054 	trace_xfs_log_ticket_regrant(log, ticket);
3055 
3056 	if (ticket->t_cnt > 0)
3057 		ticket->t_cnt--;
3058 
3059 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3060 					ticket->t_curr_res);
3061 	xlog_grant_sub_space(log, &log->l_write_head.grant,
3062 					ticket->t_curr_res);
3063 	ticket->t_curr_res = ticket->t_unit_res;
3064 	xlog_tic_reset_res(ticket);
3065 
3066 	trace_xfs_log_ticket_regrant_sub(log, ticket);
3067 
3068 	/* just return if we still have some of the pre-reserved space */
3069 	if (!ticket->t_cnt) {
3070 		xlog_grant_add_space(log, &log->l_reserve_head.grant,
3071 				     ticket->t_unit_res);
3072 		trace_xfs_log_ticket_regrant_exit(log, ticket);
3073 
3074 		ticket->t_curr_res = ticket->t_unit_res;
3075 		xlog_tic_reset_res(ticket);
3076 	}
3077 
3078 	xfs_log_ticket_put(ticket);
3079 }
3080 
3081 /*
3082  * Give back the space left from a reservation.
3083  *
3084  * All the information we need to make a correct determination of space left
3085  * is present.  For non-permanent reservations, things are quite easy.  The
3086  * count should have been decremented to zero.  We only need to deal with the
3087  * space remaining in the current reservation part of the ticket.  If the
3088  * ticket contains a permanent reservation, there may be left over space which
3089  * needs to be released.  A count of N means that N-1 refills of the current
3090  * reservation can be done before we need to ask for more space.  The first
3091  * one goes to fill up the first current reservation.  Once we run out of
3092  * space, the count will stay at zero and the only space remaining will be
3093  * in the current reservation field.
3094  */
3095 void
3096 xfs_log_ticket_ungrant(
3097 	struct xlog		*log,
3098 	struct xlog_ticket	*ticket)
3099 {
3100 	int			bytes;
3101 
3102 	trace_xfs_log_ticket_ungrant(log, ticket);
3103 
3104 	if (ticket->t_cnt > 0)
3105 		ticket->t_cnt--;
3106 
3107 	trace_xfs_log_ticket_ungrant_sub(log, ticket);
3108 
3109 	/*
3110 	 * If this is a permanent reservation ticket, we may be able to free
3111 	 * up more space based on the remaining count.
3112 	 */
3113 	bytes = ticket->t_curr_res;
3114 	if (ticket->t_cnt > 0) {
3115 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3116 		bytes += ticket->t_unit_res*ticket->t_cnt;
3117 	}
3118 
3119 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3120 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3121 
3122 	trace_xfs_log_ticket_ungrant_exit(log, ticket);
3123 
3124 	xfs_log_space_wake(log->l_mp);
3125 	xfs_log_ticket_put(ticket);
3126 }
3127 
3128 /*
3129  * This routine will mark the current iclog in the ring as WANT_SYNC and move
3130  * the current iclog pointer to the next iclog in the ring.
3131  */
3132 STATIC void
3133 xlog_state_switch_iclogs(
3134 	struct xlog		*log,
3135 	struct xlog_in_core	*iclog,
3136 	int			eventual_size)
3137 {
3138 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3139 	assert_spin_locked(&log->l_icloglock);
3140 
3141 	if (!eventual_size)
3142 		eventual_size = iclog->ic_offset;
3143 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3144 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3145 	log->l_prev_block = log->l_curr_block;
3146 	log->l_prev_cycle = log->l_curr_cycle;
3147 
3148 	/* roll log?: ic_offset changed later */
3149 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3150 
3151 	/* Round up to next log-sunit */
3152 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3153 	    log->l_mp->m_sb.sb_logsunit > 1) {
3154 		uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3155 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3156 	}
3157 
3158 	if (log->l_curr_block >= log->l_logBBsize) {
3159 		/*
3160 		 * Rewind the current block before the cycle is bumped to make
3161 		 * sure that the combined LSN never transiently moves forward
3162 		 * when the log wraps to the next cycle. This is to support the
3163 		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3164 		 * other cases should acquire l_icloglock.
3165 		 */
3166 		log->l_curr_block -= log->l_logBBsize;
3167 		ASSERT(log->l_curr_block >= 0);
3168 		smp_wmb();
3169 		log->l_curr_cycle++;
3170 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3171 			log->l_curr_cycle++;
3172 	}
3173 	ASSERT(iclog == log->l_iclog);
3174 	log->l_iclog = iclog->ic_next;
3175 }
3176 
3177 /*
3178  * Write out all data in the in-core log as of this exact moment in time.
3179  *
3180  * Data may be written to the in-core log during this call.  However,
3181  * we don't guarantee this data will be written out.  A change from past
3182  * implementation means this routine will *not* write out zero length LRs.
3183  *
3184  * Basically, we try and perform an intelligent scan of the in-core logs.
3185  * If we determine there is no flushable data, we just return.  There is no
3186  * flushable data if:
3187  *
3188  *	1. the current iclog is active and has no data; the previous iclog
3189  *		is in the active or dirty state.
3190  *	2. the current iclog is drity, and the previous iclog is in the
3191  *		active or dirty state.
3192  *
3193  * We may sleep if:
3194  *
3195  *	1. the current iclog is not in the active nor dirty state.
3196  *	2. the current iclog dirty, and the previous iclog is not in the
3197  *		active nor dirty state.
3198  *	3. the current iclog is active, and there is another thread writing
3199  *		to this particular iclog.
3200  *	4. a) the current iclog is active and has no other writers
3201  *	   b) when we return from flushing out this iclog, it is still
3202  *		not in the active nor dirty state.
3203  */
3204 int
3205 xfs_log_force(
3206 	struct xfs_mount	*mp,
3207 	uint			flags)
3208 {
3209 	struct xlog		*log = mp->m_log;
3210 	struct xlog_in_core	*iclog;
3211 	xfs_lsn_t		lsn;
3212 
3213 	XFS_STATS_INC(mp, xs_log_force);
3214 	trace_xfs_log_force(mp, 0, _RET_IP_);
3215 
3216 	xlog_cil_force(log);
3217 
3218 	spin_lock(&log->l_icloglock);
3219 	iclog = log->l_iclog;
3220 	if (iclog->ic_state == XLOG_STATE_IOERROR)
3221 		goto out_error;
3222 
3223 	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3224 	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3225 	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3226 		/*
3227 		 * If the head is dirty or (active and empty), then we need to
3228 		 * look at the previous iclog.
3229 		 *
3230 		 * If the previous iclog is active or dirty we are done.  There
3231 		 * is nothing to sync out. Otherwise, we attach ourselves to the
3232 		 * previous iclog and go to sleep.
3233 		 */
3234 		iclog = iclog->ic_prev;
3235 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3236 		if (atomic_read(&iclog->ic_refcnt) == 0) {
3237 			/*
3238 			 * We are the only one with access to this iclog.
3239 			 *
3240 			 * Flush it out now.  There should be a roundoff of zero
3241 			 * to show that someone has already taken care of the
3242 			 * roundoff from the previous sync.
3243 			 */
3244 			atomic_inc(&iclog->ic_refcnt);
3245 			lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3246 			xlog_state_switch_iclogs(log, iclog, 0);
3247 			if (xlog_state_release_iclog(log, iclog))
3248 				goto out_error;
3249 
3250 			if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3251 				goto out_unlock;
3252 		} else {
3253 			/*
3254 			 * Someone else is writing to this iclog.
3255 			 *
3256 			 * Use its call to flush out the data.  However, the
3257 			 * other thread may not force out this LR, so we mark
3258 			 * it WANT_SYNC.
3259 			 */
3260 			xlog_state_switch_iclogs(log, iclog, 0);
3261 		}
3262 	} else {
3263 		/*
3264 		 * If the head iclog is not active nor dirty, we just attach
3265 		 * ourselves to the head and go to sleep if necessary.
3266 		 */
3267 		;
3268 	}
3269 
3270 	if (flags & XFS_LOG_SYNC)
3271 		return xlog_wait_on_iclog(iclog);
3272 out_unlock:
3273 	spin_unlock(&log->l_icloglock);
3274 	return 0;
3275 out_error:
3276 	spin_unlock(&log->l_icloglock);
3277 	return -EIO;
3278 }
3279 
3280 static int
3281 __xfs_log_force_lsn(
3282 	struct xfs_mount	*mp,
3283 	xfs_lsn_t		lsn,
3284 	uint			flags,
3285 	int			*log_flushed,
3286 	bool			already_slept)
3287 {
3288 	struct xlog		*log = mp->m_log;
3289 	struct xlog_in_core	*iclog;
3290 
3291 	spin_lock(&log->l_icloglock);
3292 	iclog = log->l_iclog;
3293 	if (iclog->ic_state == XLOG_STATE_IOERROR)
3294 		goto out_error;
3295 
3296 	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3297 		iclog = iclog->ic_next;
3298 		if (iclog == log->l_iclog)
3299 			goto out_unlock;
3300 	}
3301 
3302 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3303 		/*
3304 		 * We sleep here if we haven't already slept (e.g. this is the
3305 		 * first time we've looked at the correct iclog buf) and the
3306 		 * buffer before us is going to be sync'ed.  The reason for this
3307 		 * is that if we are doing sync transactions here, by waiting
3308 		 * for the previous I/O to complete, we can allow a few more
3309 		 * transactions into this iclog before we close it down.
3310 		 *
3311 		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3312 		 * refcnt so we can release the log (which drops the ref count).
3313 		 * The state switch keeps new transaction commits from using
3314 		 * this buffer.  When the current commits finish writing into
3315 		 * the buffer, the refcount will drop to zero and the buffer
3316 		 * will go out then.
3317 		 */
3318 		if (!already_slept &&
3319 		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3320 		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3321 			XFS_STATS_INC(mp, xs_log_force_sleep);
3322 
3323 			xlog_wait(&iclog->ic_prev->ic_write_wait,
3324 					&log->l_icloglock);
3325 			return -EAGAIN;
3326 		}
3327 		atomic_inc(&iclog->ic_refcnt);
3328 		xlog_state_switch_iclogs(log, iclog, 0);
3329 		if (xlog_state_release_iclog(log, iclog))
3330 			goto out_error;
3331 		if (log_flushed)
3332 			*log_flushed = 1;
3333 	}
3334 
3335 	if (flags & XFS_LOG_SYNC)
3336 		return xlog_wait_on_iclog(iclog);
3337 out_unlock:
3338 	spin_unlock(&log->l_icloglock);
3339 	return 0;
3340 out_error:
3341 	spin_unlock(&log->l_icloglock);
3342 	return -EIO;
3343 }
3344 
3345 /*
3346  * Force the in-core log to disk for a specific LSN.
3347  *
3348  * Find in-core log with lsn.
3349  *	If it is in the DIRTY state, just return.
3350  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3351  *		state and go to sleep or return.
3352  *	If it is in any other state, go to sleep or return.
3353  *
3354  * Synchronous forces are implemented with a wait queue.  All callers trying
3355  * to force a given lsn to disk must wait on the queue attached to the
3356  * specific in-core log.  When given in-core log finally completes its write
3357  * to disk, that thread will wake up all threads waiting on the queue.
3358  */
3359 int
3360 xfs_log_force_lsn(
3361 	struct xfs_mount	*mp,
3362 	xfs_lsn_t		lsn,
3363 	uint			flags,
3364 	int			*log_flushed)
3365 {
3366 	int			ret;
3367 	ASSERT(lsn != 0);
3368 
3369 	XFS_STATS_INC(mp, xs_log_force);
3370 	trace_xfs_log_force(mp, lsn, _RET_IP_);
3371 
3372 	lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3373 	if (lsn == NULLCOMMITLSN)
3374 		return 0;
3375 
3376 	ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3377 	if (ret == -EAGAIN)
3378 		ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3379 	return ret;
3380 }
3381 
3382 /*
3383  * Free a used ticket when its refcount falls to zero.
3384  */
3385 void
3386 xfs_log_ticket_put(
3387 	xlog_ticket_t	*ticket)
3388 {
3389 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3390 	if (atomic_dec_and_test(&ticket->t_ref))
3391 		kmem_cache_free(xfs_log_ticket_zone, ticket);
3392 }
3393 
3394 xlog_ticket_t *
3395 xfs_log_ticket_get(
3396 	xlog_ticket_t	*ticket)
3397 {
3398 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3399 	atomic_inc(&ticket->t_ref);
3400 	return ticket;
3401 }
3402 
3403 /*
3404  * Figure out the total log space unit (in bytes) that would be
3405  * required for a log ticket.
3406  */
3407 int
3408 xfs_log_calc_unit_res(
3409 	struct xfs_mount	*mp,
3410 	int			unit_bytes)
3411 {
3412 	struct xlog		*log = mp->m_log;
3413 	int			iclog_space;
3414 	uint			num_headers;
3415 
3416 	/*
3417 	 * Permanent reservations have up to 'cnt'-1 active log operations
3418 	 * in the log.  A unit in this case is the amount of space for one
3419 	 * of these log operations.  Normal reservations have a cnt of 1
3420 	 * and their unit amount is the total amount of space required.
3421 	 *
3422 	 * The following lines of code account for non-transaction data
3423 	 * which occupy space in the on-disk log.
3424 	 *
3425 	 * Normal form of a transaction is:
3426 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3427 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3428 	 *
3429 	 * We need to account for all the leadup data and trailer data
3430 	 * around the transaction data.
3431 	 * And then we need to account for the worst case in terms of using
3432 	 * more space.
3433 	 * The worst case will happen if:
3434 	 * - the placement of the transaction happens to be such that the
3435 	 *   roundoff is at its maximum
3436 	 * - the transaction data is synced before the commit record is synced
3437 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3438 	 *   Therefore the commit record is in its own Log Record.
3439 	 *   This can happen as the commit record is called with its
3440 	 *   own region to xlog_write().
3441 	 *   This then means that in the worst case, roundoff can happen for
3442 	 *   the commit-rec as well.
3443 	 *   The commit-rec is smaller than padding in this scenario and so it is
3444 	 *   not added separately.
3445 	 */
3446 
3447 	/* for trans header */
3448 	unit_bytes += sizeof(xlog_op_header_t);
3449 	unit_bytes += sizeof(xfs_trans_header_t);
3450 
3451 	/* for start-rec */
3452 	unit_bytes += sizeof(xlog_op_header_t);
3453 
3454 	/*
3455 	 * for LR headers - the space for data in an iclog is the size minus
3456 	 * the space used for the headers. If we use the iclog size, then we
3457 	 * undercalculate the number of headers required.
3458 	 *
3459 	 * Furthermore - the addition of op headers for split-recs might
3460 	 * increase the space required enough to require more log and op
3461 	 * headers, so take that into account too.
3462 	 *
3463 	 * IMPORTANT: This reservation makes the assumption that if this
3464 	 * transaction is the first in an iclog and hence has the LR headers
3465 	 * accounted to it, then the remaining space in the iclog is
3466 	 * exclusively for this transaction.  i.e. if the transaction is larger
3467 	 * than the iclog, it will be the only thing in that iclog.
3468 	 * Fundamentally, this means we must pass the entire log vector to
3469 	 * xlog_write to guarantee this.
3470 	 */
3471 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3472 	num_headers = howmany(unit_bytes, iclog_space);
3473 
3474 	/* for split-recs - ophdrs added when data split over LRs */
3475 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3476 
3477 	/* add extra header reservations if we overrun */
3478 	while (!num_headers ||
3479 	       howmany(unit_bytes, iclog_space) > num_headers) {
3480 		unit_bytes += sizeof(xlog_op_header_t);
3481 		num_headers++;
3482 	}
3483 	unit_bytes += log->l_iclog_hsize * num_headers;
3484 
3485 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3486 	unit_bytes += log->l_iclog_hsize;
3487 
3488 	/* for roundoff padding for transaction data and one for commit record */
3489 	if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3490 		/* log su roundoff */
3491 		unit_bytes += 2 * mp->m_sb.sb_logsunit;
3492 	} else {
3493 		/* BB roundoff */
3494 		unit_bytes += 2 * BBSIZE;
3495         }
3496 
3497 	return unit_bytes;
3498 }
3499 
3500 /*
3501  * Allocate and initialise a new log ticket.
3502  */
3503 struct xlog_ticket *
3504 xlog_ticket_alloc(
3505 	struct xlog		*log,
3506 	int			unit_bytes,
3507 	int			cnt,
3508 	char			client,
3509 	bool			permanent)
3510 {
3511 	struct xlog_ticket	*tic;
3512 	int			unit_res;
3513 
3514 	tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3515 
3516 	unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3517 
3518 	atomic_set(&tic->t_ref, 1);
3519 	tic->t_task		= current;
3520 	INIT_LIST_HEAD(&tic->t_queue);
3521 	tic->t_unit_res		= unit_res;
3522 	tic->t_curr_res		= unit_res;
3523 	tic->t_cnt		= cnt;
3524 	tic->t_ocnt		= cnt;
3525 	tic->t_tid		= prandom_u32();
3526 	tic->t_clientid		= client;
3527 	if (permanent)
3528 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3529 
3530 	xlog_tic_reset_res(tic);
3531 
3532 	return tic;
3533 }
3534 
3535 #if defined(DEBUG)
3536 /*
3537  * Make sure that the destination ptr is within the valid data region of
3538  * one of the iclogs.  This uses backup pointers stored in a different
3539  * part of the log in case we trash the log structure.
3540  */
3541 STATIC void
3542 xlog_verify_dest_ptr(
3543 	struct xlog	*log,
3544 	void		*ptr)
3545 {
3546 	int i;
3547 	int good_ptr = 0;
3548 
3549 	for (i = 0; i < log->l_iclog_bufs; i++) {
3550 		if (ptr >= log->l_iclog_bak[i] &&
3551 		    ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3552 			good_ptr++;
3553 	}
3554 
3555 	if (!good_ptr)
3556 		xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3557 }
3558 
3559 /*
3560  * Check to make sure the grant write head didn't just over lap the tail.  If
3561  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3562  * the cycles differ by exactly one and check the byte count.
3563  *
3564  * This check is run unlocked, so can give false positives. Rather than assert
3565  * on failures, use a warn-once flag and a panic tag to allow the admin to
3566  * determine if they want to panic the machine when such an error occurs. For
3567  * debug kernels this will have the same effect as using an assert but, unlinke
3568  * an assert, it can be turned off at runtime.
3569  */
3570 STATIC void
3571 xlog_verify_grant_tail(
3572 	struct xlog	*log)
3573 {
3574 	int		tail_cycle, tail_blocks;
3575 	int		cycle, space;
3576 
3577 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3578 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3579 	if (tail_cycle != cycle) {
3580 		if (cycle - 1 != tail_cycle &&
3581 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3582 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3583 				"%s: cycle - 1 != tail_cycle", __func__);
3584 			log->l_flags |= XLOG_TAIL_WARN;
3585 		}
3586 
3587 		if (space > BBTOB(tail_blocks) &&
3588 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3589 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3590 				"%s: space > BBTOB(tail_blocks)", __func__);
3591 			log->l_flags |= XLOG_TAIL_WARN;
3592 		}
3593 	}
3594 }
3595 
3596 /* check if it will fit */
3597 STATIC void
3598 xlog_verify_tail_lsn(
3599 	struct xlog		*log,
3600 	struct xlog_in_core	*iclog,
3601 	xfs_lsn_t		tail_lsn)
3602 {
3603     int blocks;
3604 
3605     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3606 	blocks =
3607 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3608 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3609 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3610     } else {
3611 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3612 
3613 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3614 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3615 
3616 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3617 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3618 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3619     }
3620 }
3621 
3622 /*
3623  * Perform a number of checks on the iclog before writing to disk.
3624  *
3625  * 1. Make sure the iclogs are still circular
3626  * 2. Make sure we have a good magic number
3627  * 3. Make sure we don't have magic numbers in the data
3628  * 4. Check fields of each log operation header for:
3629  *	A. Valid client identifier
3630  *	B. tid ptr value falls in valid ptr space (user space code)
3631  *	C. Length in log record header is correct according to the
3632  *		individual operation headers within record.
3633  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3634  *	log, check the preceding blocks of the physical log to make sure all
3635  *	the cycle numbers agree with the current cycle number.
3636  */
3637 STATIC void
3638 xlog_verify_iclog(
3639 	struct xlog		*log,
3640 	struct xlog_in_core	*iclog,
3641 	int			count)
3642 {
3643 	xlog_op_header_t	*ophead;
3644 	xlog_in_core_t		*icptr;
3645 	xlog_in_core_2_t	*xhdr;
3646 	void			*base_ptr, *ptr, *p;
3647 	ptrdiff_t		field_offset;
3648 	uint8_t			clientid;
3649 	int			len, i, j, k, op_len;
3650 	int			idx;
3651 
3652 	/* check validity of iclog pointers */
3653 	spin_lock(&log->l_icloglock);
3654 	icptr = log->l_iclog;
3655 	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3656 		ASSERT(icptr);
3657 
3658 	if (icptr != log->l_iclog)
3659 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3660 	spin_unlock(&log->l_icloglock);
3661 
3662 	/* check log magic numbers */
3663 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3664 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3665 
3666 	base_ptr = ptr = &iclog->ic_header;
3667 	p = &iclog->ic_header;
3668 	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3669 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3670 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3671 				__func__);
3672 	}
3673 
3674 	/* check fields */
3675 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3676 	base_ptr = ptr = iclog->ic_datap;
3677 	ophead = ptr;
3678 	xhdr = iclog->ic_data;
3679 	for (i = 0; i < len; i++) {
3680 		ophead = ptr;
3681 
3682 		/* clientid is only 1 byte */
3683 		p = &ophead->oh_clientid;
3684 		field_offset = p - base_ptr;
3685 		if (field_offset & 0x1ff) {
3686 			clientid = ophead->oh_clientid;
3687 		} else {
3688 			idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3689 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3690 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3691 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3692 				clientid = xlog_get_client_id(
3693 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3694 			} else {
3695 				clientid = xlog_get_client_id(
3696 					iclog->ic_header.h_cycle_data[idx]);
3697 			}
3698 		}
3699 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3700 			xfs_warn(log->l_mp,
3701 				"%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3702 				__func__, clientid, ophead,
3703 				(unsigned long)field_offset);
3704 
3705 		/* check length */
3706 		p = &ophead->oh_len;
3707 		field_offset = p - base_ptr;
3708 		if (field_offset & 0x1ff) {
3709 			op_len = be32_to_cpu(ophead->oh_len);
3710 		} else {
3711 			idx = BTOBBT((uintptr_t)&ophead->oh_len -
3712 				    (uintptr_t)iclog->ic_datap);
3713 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3714 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3715 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3716 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3717 			} else {
3718 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3719 			}
3720 		}
3721 		ptr += sizeof(xlog_op_header_t) + op_len;
3722 	}
3723 }
3724 #endif
3725 
3726 /*
3727  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3728  */
3729 STATIC int
3730 xlog_state_ioerror(
3731 	struct xlog	*log)
3732 {
3733 	xlog_in_core_t	*iclog, *ic;
3734 
3735 	iclog = log->l_iclog;
3736 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
3737 		/*
3738 		 * Mark all the incore logs IOERROR.
3739 		 * From now on, no log flushes will result.
3740 		 */
3741 		ic = iclog;
3742 		do {
3743 			ic->ic_state = XLOG_STATE_IOERROR;
3744 			ic = ic->ic_next;
3745 		} while (ic != iclog);
3746 		return 0;
3747 	}
3748 	/*
3749 	 * Return non-zero, if state transition has already happened.
3750 	 */
3751 	return 1;
3752 }
3753 
3754 /*
3755  * This is called from xfs_force_shutdown, when we're forcibly
3756  * shutting down the filesystem, typically because of an IO error.
3757  * Our main objectives here are to make sure that:
3758  *	a. if !logerror, flush the logs to disk. Anything modified
3759  *	   after this is ignored.
3760  *	b. the filesystem gets marked 'SHUTDOWN' for all interested
3761  *	   parties to find out, 'atomically'.
3762  *	c. those who're sleeping on log reservations, pinned objects and
3763  *	    other resources get woken up, and be told the bad news.
3764  *	d. nothing new gets queued up after (b) and (c) are done.
3765  *
3766  * Note: for the !logerror case we need to flush the regions held in memory out
3767  * to disk first. This needs to be done before the log is marked as shutdown,
3768  * otherwise the iclog writes will fail.
3769  */
3770 int
3771 xfs_log_force_umount(
3772 	struct xfs_mount	*mp,
3773 	int			logerror)
3774 {
3775 	struct xlog	*log;
3776 	int		retval;
3777 
3778 	log = mp->m_log;
3779 
3780 	/*
3781 	 * If this happens during log recovery, don't worry about
3782 	 * locking; the log isn't open for business yet.
3783 	 */
3784 	if (!log ||
3785 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3786 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3787 		if (mp->m_sb_bp)
3788 			mp->m_sb_bp->b_flags |= XBF_DONE;
3789 		return 0;
3790 	}
3791 
3792 	/*
3793 	 * Somebody could've already done the hard work for us.
3794 	 * No need to get locks for this.
3795 	 */
3796 	if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3797 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3798 		return 1;
3799 	}
3800 
3801 	/*
3802 	 * Flush all the completed transactions to disk before marking the log
3803 	 * being shut down. We need to do it in this order to ensure that
3804 	 * completed operations are safely on disk before we shut down, and that
3805 	 * we don't have to issue any buffer IO after the shutdown flags are set
3806 	 * to guarantee this.
3807 	 */
3808 	if (!logerror)
3809 		xfs_log_force(mp, XFS_LOG_SYNC);
3810 
3811 	/*
3812 	 * mark the filesystem and the as in a shutdown state and wake
3813 	 * everybody up to tell them the bad news.
3814 	 */
3815 	spin_lock(&log->l_icloglock);
3816 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3817 	if (mp->m_sb_bp)
3818 		mp->m_sb_bp->b_flags |= XBF_DONE;
3819 
3820 	/*
3821 	 * Mark the log and the iclogs with IO error flags to prevent any
3822 	 * further log IO from being issued or completed.
3823 	 */
3824 	log->l_flags |= XLOG_IO_ERROR;
3825 	retval = xlog_state_ioerror(log);
3826 	spin_unlock(&log->l_icloglock);
3827 
3828 	/*
3829 	 * We don't want anybody waiting for log reservations after this. That
3830 	 * means we have to wake up everybody queued up on reserveq as well as
3831 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3832 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3833 	 * action is protected by the grant locks.
3834 	 */
3835 	xlog_grant_head_wake_all(&log->l_reserve_head);
3836 	xlog_grant_head_wake_all(&log->l_write_head);
3837 
3838 	/*
3839 	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3840 	 * as if the log writes were completed. The abort handling in the log
3841 	 * item committed callback functions will do this again under lock to
3842 	 * avoid races.
3843 	 */
3844 	spin_lock(&log->l_cilp->xc_push_lock);
3845 	wake_up_all(&log->l_cilp->xc_commit_wait);
3846 	spin_unlock(&log->l_cilp->xc_push_lock);
3847 	xlog_state_do_callback(log);
3848 
3849 	/* return non-zero if log IOERROR transition had already happened */
3850 	return retval;
3851 }
3852 
3853 STATIC int
3854 xlog_iclogs_empty(
3855 	struct xlog	*log)
3856 {
3857 	xlog_in_core_t	*iclog;
3858 
3859 	iclog = log->l_iclog;
3860 	do {
3861 		/* endianness does not matter here, zero is zero in
3862 		 * any language.
3863 		 */
3864 		if (iclog->ic_header.h_num_logops)
3865 			return 0;
3866 		iclog = iclog->ic_next;
3867 	} while (iclog != log->l_iclog);
3868 	return 1;
3869 }
3870 
3871 /*
3872  * Verify that an LSN stamped into a piece of metadata is valid. This is
3873  * intended for use in read verifiers on v5 superblocks.
3874  */
3875 bool
3876 xfs_log_check_lsn(
3877 	struct xfs_mount	*mp,
3878 	xfs_lsn_t		lsn)
3879 {
3880 	struct xlog		*log = mp->m_log;
3881 	bool			valid;
3882 
3883 	/*
3884 	 * norecovery mode skips mount-time log processing and unconditionally
3885 	 * resets the in-core LSN. We can't validate in this mode, but
3886 	 * modifications are not allowed anyways so just return true.
3887 	 */
3888 	if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3889 		return true;
3890 
3891 	/*
3892 	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3893 	 * handled by recovery and thus safe to ignore here.
3894 	 */
3895 	if (lsn == NULLCOMMITLSN)
3896 		return true;
3897 
3898 	valid = xlog_valid_lsn(mp->m_log, lsn);
3899 
3900 	/* warn the user about what's gone wrong before verifier failure */
3901 	if (!valid) {
3902 		spin_lock(&log->l_icloglock);
3903 		xfs_warn(mp,
3904 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3905 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3906 			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3907 			 log->l_curr_cycle, log->l_curr_block);
3908 		spin_unlock(&log->l_icloglock);
3909 	}
3910 
3911 	return valid;
3912 }
3913 
3914 bool
3915 xfs_log_in_recovery(
3916 	struct xfs_mount	*mp)
3917 {
3918 	struct xlog		*log = mp->m_log;
3919 
3920 	return log->l_flags & XLOG_ACTIVE_RECOVERY;
3921 }
3922