xref: /openbmc/linux/fs/xfs/xfs_log.c (revision b4e18b29)
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 void
95 xlog_grant_sub_space(
96 	struct xlog		*log,
97 	atomic64_t		*head,
98 	int			bytes)
99 {
100 	int64_t	head_val = atomic64_read(head);
101 	int64_t new, old;
102 
103 	do {
104 		int	cycle, space;
105 
106 		xlog_crack_grant_head_val(head_val, &cycle, &space);
107 
108 		space -= bytes;
109 		if (space < 0) {
110 			space += log->l_logsize;
111 			cycle--;
112 		}
113 
114 		old = head_val;
115 		new = xlog_assign_grant_head_val(cycle, space);
116 		head_val = atomic64_cmpxchg(head, old, new);
117 	} while (head_val != old);
118 }
119 
120 static void
121 xlog_grant_add_space(
122 	struct xlog		*log,
123 	atomic64_t		*head,
124 	int			bytes)
125 {
126 	int64_t	head_val = atomic64_read(head);
127 	int64_t new, old;
128 
129 	do {
130 		int		tmp;
131 		int		cycle, space;
132 
133 		xlog_crack_grant_head_val(head_val, &cycle, &space);
134 
135 		tmp = log->l_logsize - space;
136 		if (tmp > bytes)
137 			space += bytes;
138 		else {
139 			space = bytes - tmp;
140 			cycle++;
141 		}
142 
143 		old = head_val;
144 		new = xlog_assign_grant_head_val(cycle, space);
145 		head_val = atomic64_cmpxchg(head, old, new);
146 	} while (head_val != old);
147 }
148 
149 STATIC void
150 xlog_grant_head_init(
151 	struct xlog_grant_head	*head)
152 {
153 	xlog_assign_grant_head(&head->grant, 1, 0);
154 	INIT_LIST_HEAD(&head->waiters);
155 	spin_lock_init(&head->lock);
156 }
157 
158 STATIC void
159 xlog_grant_head_wake_all(
160 	struct xlog_grant_head	*head)
161 {
162 	struct xlog_ticket	*tic;
163 
164 	spin_lock(&head->lock);
165 	list_for_each_entry(tic, &head->waiters, t_queue)
166 		wake_up_process(tic->t_task);
167 	spin_unlock(&head->lock);
168 }
169 
170 static inline int
171 xlog_ticket_reservation(
172 	struct xlog		*log,
173 	struct xlog_grant_head	*head,
174 	struct xlog_ticket	*tic)
175 {
176 	if (head == &log->l_write_head) {
177 		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
178 		return tic->t_unit_res;
179 	} else {
180 		if (tic->t_flags & XLOG_TIC_PERM_RESERV)
181 			return tic->t_unit_res * tic->t_cnt;
182 		else
183 			return tic->t_unit_res;
184 	}
185 }
186 
187 STATIC bool
188 xlog_grant_head_wake(
189 	struct xlog		*log,
190 	struct xlog_grant_head	*head,
191 	int			*free_bytes)
192 {
193 	struct xlog_ticket	*tic;
194 	int			need_bytes;
195 	bool			woken_task = false;
196 
197 	list_for_each_entry(tic, &head->waiters, t_queue) {
198 
199 		/*
200 		 * There is a chance that the size of the CIL checkpoints in
201 		 * progress at the last AIL push target calculation resulted in
202 		 * limiting the target to the log head (l_last_sync_lsn) at the
203 		 * time. This may not reflect where the log head is now as the
204 		 * CIL checkpoints may have completed.
205 		 *
206 		 * Hence when we are woken here, it may be that the head of the
207 		 * log that has moved rather than the tail. As the tail didn't
208 		 * move, there still won't be space available for the
209 		 * reservation we require.  However, if the AIL has already
210 		 * pushed to the target defined by the old log head location, we
211 		 * will hang here waiting for something else to update the AIL
212 		 * push target.
213 		 *
214 		 * Therefore, if there isn't space to wake the first waiter on
215 		 * the grant head, we need to push the AIL again to ensure the
216 		 * target reflects both the current log tail and log head
217 		 * position before we wait for the tail to move again.
218 		 */
219 
220 		need_bytes = xlog_ticket_reservation(log, head, tic);
221 		if (*free_bytes < need_bytes) {
222 			if (!woken_task)
223 				xlog_grant_push_ail(log, need_bytes);
224 			return false;
225 		}
226 
227 		*free_bytes -= need_bytes;
228 		trace_xfs_log_grant_wake_up(log, tic);
229 		wake_up_process(tic->t_task);
230 		woken_task = true;
231 	}
232 
233 	return true;
234 }
235 
236 STATIC int
237 xlog_grant_head_wait(
238 	struct xlog		*log,
239 	struct xlog_grant_head	*head,
240 	struct xlog_ticket	*tic,
241 	int			need_bytes) __releases(&head->lock)
242 					    __acquires(&head->lock)
243 {
244 	list_add_tail(&tic->t_queue, &head->waiters);
245 
246 	do {
247 		if (XLOG_FORCED_SHUTDOWN(log))
248 			goto shutdown;
249 		xlog_grant_push_ail(log, need_bytes);
250 
251 		__set_current_state(TASK_UNINTERRUPTIBLE);
252 		spin_unlock(&head->lock);
253 
254 		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
255 
256 		trace_xfs_log_grant_sleep(log, tic);
257 		schedule();
258 		trace_xfs_log_grant_wake(log, tic);
259 
260 		spin_lock(&head->lock);
261 		if (XLOG_FORCED_SHUTDOWN(log))
262 			goto shutdown;
263 	} while (xlog_space_left(log, &head->grant) < need_bytes);
264 
265 	list_del_init(&tic->t_queue);
266 	return 0;
267 shutdown:
268 	list_del_init(&tic->t_queue);
269 	return -EIO;
270 }
271 
272 /*
273  * Atomically get the log space required for a log ticket.
274  *
275  * Once a ticket gets put onto head->waiters, it will only return after the
276  * needed reservation is satisfied.
277  *
278  * This function is structured so that it has a lock free fast path. This is
279  * necessary because every new transaction reservation will come through this
280  * path. Hence any lock will be globally hot if we take it unconditionally on
281  * every pass.
282  *
283  * As tickets are only ever moved on and off head->waiters under head->lock, we
284  * only need to take that lock if we are going to add the ticket to the queue
285  * and sleep. We can avoid taking the lock if the ticket was never added to
286  * head->waiters because the t_queue list head will be empty and we hold the
287  * only reference to it so it can safely be checked unlocked.
288  */
289 STATIC int
290 xlog_grant_head_check(
291 	struct xlog		*log,
292 	struct xlog_grant_head	*head,
293 	struct xlog_ticket	*tic,
294 	int			*need_bytes)
295 {
296 	int			free_bytes;
297 	int			error = 0;
298 
299 	ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
300 
301 	/*
302 	 * If there are other waiters on the queue then give them a chance at
303 	 * logspace before us.  Wake up the first waiters, if we do not wake
304 	 * up all the waiters then go to sleep waiting for more free space,
305 	 * otherwise try to get some space for this transaction.
306 	 */
307 	*need_bytes = xlog_ticket_reservation(log, head, tic);
308 	free_bytes = xlog_space_left(log, &head->grant);
309 	if (!list_empty_careful(&head->waiters)) {
310 		spin_lock(&head->lock);
311 		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
312 		    free_bytes < *need_bytes) {
313 			error = xlog_grant_head_wait(log, head, tic,
314 						     *need_bytes);
315 		}
316 		spin_unlock(&head->lock);
317 	} else if (free_bytes < *need_bytes) {
318 		spin_lock(&head->lock);
319 		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
320 		spin_unlock(&head->lock);
321 	}
322 
323 	return error;
324 }
325 
326 static void
327 xlog_tic_reset_res(xlog_ticket_t *tic)
328 {
329 	tic->t_res_num = 0;
330 	tic->t_res_arr_sum = 0;
331 	tic->t_res_num_ophdrs = 0;
332 }
333 
334 static void
335 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
336 {
337 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
338 		/* add to overflow and start again */
339 		tic->t_res_o_flow += tic->t_res_arr_sum;
340 		tic->t_res_num = 0;
341 		tic->t_res_arr_sum = 0;
342 	}
343 
344 	tic->t_res_arr[tic->t_res_num].r_len = len;
345 	tic->t_res_arr[tic->t_res_num].r_type = type;
346 	tic->t_res_arr_sum += len;
347 	tic->t_res_num++;
348 }
349 
350 /*
351  * Replenish the byte reservation required by moving the grant write head.
352  */
353 int
354 xfs_log_regrant(
355 	struct xfs_mount	*mp,
356 	struct xlog_ticket	*tic)
357 {
358 	struct xlog		*log = mp->m_log;
359 	int			need_bytes;
360 	int			error = 0;
361 
362 	if (XLOG_FORCED_SHUTDOWN(log))
363 		return -EIO;
364 
365 	XFS_STATS_INC(mp, xs_try_logspace);
366 
367 	/*
368 	 * This is a new transaction on the ticket, so we need to change the
369 	 * transaction ID so that the next transaction has a different TID in
370 	 * the log. Just add one to the existing tid so that we can see chains
371 	 * of rolling transactions in the log easily.
372 	 */
373 	tic->t_tid++;
374 
375 	xlog_grant_push_ail(log, tic->t_unit_res);
376 
377 	tic->t_curr_res = tic->t_unit_res;
378 	xlog_tic_reset_res(tic);
379 
380 	if (tic->t_cnt > 0)
381 		return 0;
382 
383 	trace_xfs_log_regrant(log, tic);
384 
385 	error = xlog_grant_head_check(log, &log->l_write_head, tic,
386 				      &need_bytes);
387 	if (error)
388 		goto out_error;
389 
390 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
391 	trace_xfs_log_regrant_exit(log, tic);
392 	xlog_verify_grant_tail(log);
393 	return 0;
394 
395 out_error:
396 	/*
397 	 * If we are failing, make sure the ticket doesn't have any current
398 	 * reservations.  We don't want to add this back when the ticket/
399 	 * transaction gets cancelled.
400 	 */
401 	tic->t_curr_res = 0;
402 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
403 	return error;
404 }
405 
406 /*
407  * Reserve log space and return a ticket corresponding to the reservation.
408  *
409  * Each reservation is going to reserve extra space for a log record header.
410  * When writes happen to the on-disk log, we don't subtract the length of the
411  * log record header from any reservation.  By wasting space in each
412  * reservation, we prevent over allocation problems.
413  */
414 int
415 xfs_log_reserve(
416 	struct xfs_mount	*mp,
417 	int		 	unit_bytes,
418 	int		 	cnt,
419 	struct xlog_ticket	**ticp,
420 	uint8_t		 	client,
421 	bool			permanent)
422 {
423 	struct xlog		*log = mp->m_log;
424 	struct xlog_ticket	*tic;
425 	int			need_bytes;
426 	int			error = 0;
427 
428 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
429 
430 	if (XLOG_FORCED_SHUTDOWN(log))
431 		return -EIO;
432 
433 	XFS_STATS_INC(mp, xs_try_logspace);
434 
435 	ASSERT(*ticp == NULL);
436 	tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
437 	*ticp = tic;
438 
439 	xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
440 					    : tic->t_unit_res);
441 
442 	trace_xfs_log_reserve(log, tic);
443 
444 	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
445 				      &need_bytes);
446 	if (error)
447 		goto out_error;
448 
449 	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
450 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
451 	trace_xfs_log_reserve_exit(log, tic);
452 	xlog_verify_grant_tail(log);
453 	return 0;
454 
455 out_error:
456 	/*
457 	 * If we are failing, make sure the ticket doesn't have any current
458 	 * reservations.  We don't want to add this back when the ticket/
459 	 * transaction gets cancelled.
460 	 */
461 	tic->t_curr_res = 0;
462 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
463 	return error;
464 }
465 
466 static bool
467 __xlog_state_release_iclog(
468 	struct xlog		*log,
469 	struct xlog_in_core	*iclog)
470 {
471 	lockdep_assert_held(&log->l_icloglock);
472 
473 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
474 		/* update tail before writing to iclog */
475 		xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
476 
477 		iclog->ic_state = XLOG_STATE_SYNCING;
478 		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
479 		xlog_verify_tail_lsn(log, iclog, tail_lsn);
480 		/* cycle incremented when incrementing curr_block */
481 		return true;
482 	}
483 
484 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
485 	return false;
486 }
487 
488 /*
489  * Flush iclog to disk if this is the last reference to the given iclog and the
490  * it is in the WANT_SYNC state.
491  */
492 static int
493 xlog_state_release_iclog(
494 	struct xlog		*log,
495 	struct xlog_in_core	*iclog)
496 {
497 	lockdep_assert_held(&log->l_icloglock);
498 
499 	if (iclog->ic_state == XLOG_STATE_IOERROR)
500 		return -EIO;
501 
502 	if (atomic_dec_and_test(&iclog->ic_refcnt) &&
503 	    __xlog_state_release_iclog(log, iclog)) {
504 		spin_unlock(&log->l_icloglock);
505 		xlog_sync(log, iclog);
506 		spin_lock(&log->l_icloglock);
507 	}
508 
509 	return 0;
510 }
511 
512 void
513 xfs_log_release_iclog(
514 	struct xlog_in_core	*iclog)
515 {
516 	struct xlog		*log = iclog->ic_log;
517 	bool			sync = false;
518 
519 	if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) {
520 		if (iclog->ic_state != XLOG_STATE_IOERROR)
521 			sync = __xlog_state_release_iclog(log, iclog);
522 		spin_unlock(&log->l_icloglock);
523 	}
524 
525 	if (sync)
526 		xlog_sync(log, iclog);
527 }
528 
529 /*
530  * Mount a log filesystem
531  *
532  * mp		- ubiquitous xfs mount point structure
533  * log_target	- buftarg of on-disk log device
534  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
535  * num_bblocks	- Number of BBSIZE blocks in on-disk log
536  *
537  * Return error or zero.
538  */
539 int
540 xfs_log_mount(
541 	xfs_mount_t	*mp,
542 	xfs_buftarg_t	*log_target,
543 	xfs_daddr_t	blk_offset,
544 	int		num_bblks)
545 {
546 	bool		fatal = xfs_sb_version_hascrc(&mp->m_sb);
547 	int		error = 0;
548 	int		min_logfsbs;
549 
550 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
551 		xfs_notice(mp, "Mounting V%d Filesystem",
552 			   XFS_SB_VERSION_NUM(&mp->m_sb));
553 	} else {
554 		xfs_notice(mp,
555 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
556 			   XFS_SB_VERSION_NUM(&mp->m_sb));
557 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
558 	}
559 
560 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
561 	if (IS_ERR(mp->m_log)) {
562 		error = PTR_ERR(mp->m_log);
563 		goto out;
564 	}
565 
566 	/*
567 	 * Validate the given log space and drop a critical message via syslog
568 	 * if the log size is too small that would lead to some unexpected
569 	 * situations in transaction log space reservation stage.
570 	 *
571 	 * Note: we can't just reject the mount if the validation fails.  This
572 	 * would mean that people would have to downgrade their kernel just to
573 	 * remedy the situation as there is no way to grow the log (short of
574 	 * black magic surgery with xfs_db).
575 	 *
576 	 * We can, however, reject mounts for CRC format filesystems, as the
577 	 * mkfs binary being used to make the filesystem should never create a
578 	 * filesystem with a log that is too small.
579 	 */
580 	min_logfsbs = xfs_log_calc_minimum_size(mp);
581 
582 	if (mp->m_sb.sb_logblocks < min_logfsbs) {
583 		xfs_warn(mp,
584 		"Log size %d blocks too small, minimum size is %d blocks",
585 			 mp->m_sb.sb_logblocks, min_logfsbs);
586 		error = -EINVAL;
587 	} else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
588 		xfs_warn(mp,
589 		"Log size %d blocks too large, maximum size is %lld blocks",
590 			 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
591 		error = -EINVAL;
592 	} else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
593 		xfs_warn(mp,
594 		"log size %lld bytes too large, maximum size is %lld bytes",
595 			 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
596 			 XFS_MAX_LOG_BYTES);
597 		error = -EINVAL;
598 	} else if (mp->m_sb.sb_logsunit > 1 &&
599 		   mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
600 		xfs_warn(mp,
601 		"log stripe unit %u bytes must be a multiple of block size",
602 			 mp->m_sb.sb_logsunit);
603 		error = -EINVAL;
604 		fatal = true;
605 	}
606 	if (error) {
607 		/*
608 		 * Log check errors are always fatal on v5; or whenever bad
609 		 * metadata leads to a crash.
610 		 */
611 		if (fatal) {
612 			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
613 			ASSERT(0);
614 			goto out_free_log;
615 		}
616 		xfs_crit(mp, "Log size out of supported range.");
617 		xfs_crit(mp,
618 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
619 	}
620 
621 	/*
622 	 * Initialize the AIL now we have a log.
623 	 */
624 	error = xfs_trans_ail_init(mp);
625 	if (error) {
626 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
627 		goto out_free_log;
628 	}
629 	mp->m_log->l_ailp = mp->m_ail;
630 
631 	/*
632 	 * skip log recovery on a norecovery mount.  pretend it all
633 	 * just worked.
634 	 */
635 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
636 		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
637 
638 		if (readonly)
639 			mp->m_flags &= ~XFS_MOUNT_RDONLY;
640 
641 		error = xlog_recover(mp->m_log);
642 
643 		if (readonly)
644 			mp->m_flags |= XFS_MOUNT_RDONLY;
645 		if (error) {
646 			xfs_warn(mp, "log mount/recovery failed: error %d",
647 				error);
648 			xlog_recover_cancel(mp->m_log);
649 			goto out_destroy_ail;
650 		}
651 	}
652 
653 	error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
654 			       "log");
655 	if (error)
656 		goto out_destroy_ail;
657 
658 	/* Normal transactions can now occur */
659 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
660 
661 	/*
662 	 * Now the log has been fully initialised and we know were our
663 	 * space grant counters are, we can initialise the permanent ticket
664 	 * needed for delayed logging to work.
665 	 */
666 	xlog_cil_init_post_recovery(mp->m_log);
667 
668 	return 0;
669 
670 out_destroy_ail:
671 	xfs_trans_ail_destroy(mp);
672 out_free_log:
673 	xlog_dealloc_log(mp->m_log);
674 out:
675 	return error;
676 }
677 
678 /*
679  * Finish the recovery of the file system.  This is separate from the
680  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
681  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
682  * here.
683  *
684  * If we finish recovery successfully, start the background log work. If we are
685  * not doing recovery, then we have a RO filesystem and we don't need to start
686  * it.
687  */
688 int
689 xfs_log_mount_finish(
690 	struct xfs_mount	*mp)
691 {
692 	int	error = 0;
693 	bool	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
694 	bool	recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
695 
696 	if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
697 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
698 		return 0;
699 	} else if (readonly) {
700 		/* Allow unlinked processing to proceed */
701 		mp->m_flags &= ~XFS_MOUNT_RDONLY;
702 	}
703 
704 	/*
705 	 * During the second phase of log recovery, we need iget and
706 	 * iput to behave like they do for an active filesystem.
707 	 * xfs_fs_drop_inode needs to be able to prevent the deletion
708 	 * of inodes before we're done replaying log items on those
709 	 * inodes.  Turn it off immediately after recovery finishes
710 	 * so that we don't leak the quota inodes if subsequent mount
711 	 * activities fail.
712 	 *
713 	 * We let all inodes involved in redo item processing end up on
714 	 * the LRU instead of being evicted immediately so that if we do
715 	 * something to an unlinked inode, the irele won't cause
716 	 * premature truncation and freeing of the inode, which results
717 	 * in log recovery failure.  We have to evict the unreferenced
718 	 * lru inodes after clearing SB_ACTIVE because we don't
719 	 * otherwise clean up the lru if there's a subsequent failure in
720 	 * xfs_mountfs, which leads to us leaking the inodes if nothing
721 	 * else (e.g. quotacheck) references the inodes before the
722 	 * mount failure occurs.
723 	 */
724 	mp->m_super->s_flags |= SB_ACTIVE;
725 	error = xlog_recover_finish(mp->m_log);
726 	if (!error)
727 		xfs_log_work_queue(mp);
728 	mp->m_super->s_flags &= ~SB_ACTIVE;
729 	evict_inodes(mp->m_super);
730 
731 	/*
732 	 * Drain the buffer LRU after log recovery. This is required for v4
733 	 * filesystems to avoid leaving around buffers with NULL verifier ops,
734 	 * but we do it unconditionally to make sure we're always in a clean
735 	 * cache state after mount.
736 	 *
737 	 * Don't push in the error case because the AIL may have pending intents
738 	 * that aren't removed until recovery is cancelled.
739 	 */
740 	if (!error && recovered) {
741 		xfs_log_force(mp, XFS_LOG_SYNC);
742 		xfs_ail_push_all_sync(mp->m_ail);
743 	}
744 	xfs_wait_buftarg(mp->m_ddev_targp);
745 
746 	if (readonly)
747 		mp->m_flags |= XFS_MOUNT_RDONLY;
748 
749 	return error;
750 }
751 
752 /*
753  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
754  * the log.
755  */
756 void
757 xfs_log_mount_cancel(
758 	struct xfs_mount	*mp)
759 {
760 	xlog_recover_cancel(mp->m_log);
761 	xfs_log_unmount(mp);
762 }
763 
764 /*
765  * Wait for the iclog to be written disk, or return an error if the log has been
766  * shut down.
767  */
768 static int
769 xlog_wait_on_iclog(
770 	struct xlog_in_core	*iclog)
771 		__releases(iclog->ic_log->l_icloglock)
772 {
773 	struct xlog		*log = iclog->ic_log;
774 
775 	if (!XLOG_FORCED_SHUTDOWN(log) &&
776 	    iclog->ic_state != XLOG_STATE_ACTIVE &&
777 	    iclog->ic_state != XLOG_STATE_DIRTY) {
778 		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
779 		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
780 	} else {
781 		spin_unlock(&log->l_icloglock);
782 	}
783 
784 	if (XLOG_FORCED_SHUTDOWN(log))
785 		return -EIO;
786 	return 0;
787 }
788 
789 /*
790  * Write out an unmount record using the ticket provided. We have to account for
791  * the data space used in the unmount ticket as this write is not done from a
792  * transaction context that has already done the accounting for us.
793  */
794 static int
795 xlog_write_unmount_record(
796 	struct xlog		*log,
797 	struct xlog_ticket	*ticket,
798 	xfs_lsn_t		*lsn,
799 	uint			flags)
800 {
801 	struct xfs_unmount_log_format ulf = {
802 		.magic = XLOG_UNMOUNT_TYPE,
803 	};
804 	struct xfs_log_iovec reg = {
805 		.i_addr = &ulf,
806 		.i_len = sizeof(ulf),
807 		.i_type = XLOG_REG_TYPE_UNMOUNT,
808 	};
809 	struct xfs_log_vec vec = {
810 		.lv_niovecs = 1,
811 		.lv_iovecp = &reg,
812 	};
813 
814 	/* account for space used by record data */
815 	ticket->t_curr_res -= sizeof(ulf);
816 	return xlog_write(log, &vec, ticket, lsn, NULL, flags, false);
817 }
818 
819 /*
820  * Mark the filesystem clean by writing an unmount record to the head of the
821  * log.
822  */
823 static void
824 xlog_unmount_write(
825 	struct xlog		*log)
826 {
827 	struct xfs_mount	*mp = log->l_mp;
828 	struct xlog_in_core	*iclog;
829 	struct xlog_ticket	*tic = NULL;
830 	xfs_lsn_t		lsn;
831 	uint			flags = XLOG_UNMOUNT_TRANS;
832 	int			error;
833 
834 	error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
835 	if (error)
836 		goto out_err;
837 
838 	error = xlog_write_unmount_record(log, tic, &lsn, flags);
839 	/*
840 	 * At this point, we're umounting anyway, so there's no point in
841 	 * transitioning log state to IOERROR. Just continue...
842 	 */
843 out_err:
844 	if (error)
845 		xfs_alert(mp, "%s: unmount record failed", __func__);
846 
847 	spin_lock(&log->l_icloglock);
848 	iclog = log->l_iclog;
849 	atomic_inc(&iclog->ic_refcnt);
850 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
851 		xlog_state_switch_iclogs(log, iclog, 0);
852 	else
853 		ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
854 		       iclog->ic_state == XLOG_STATE_IOERROR);
855 	error = xlog_state_release_iclog(log, iclog);
856 	xlog_wait_on_iclog(iclog);
857 
858 	if (tic) {
859 		trace_xfs_log_umount_write(log, tic);
860 		xfs_log_ticket_ungrant(log, tic);
861 	}
862 }
863 
864 static void
865 xfs_log_unmount_verify_iclog(
866 	struct xlog		*log)
867 {
868 	struct xlog_in_core	*iclog = log->l_iclog;
869 
870 	do {
871 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
872 		ASSERT(iclog->ic_offset == 0);
873 	} while ((iclog = iclog->ic_next) != log->l_iclog);
874 }
875 
876 /*
877  * Unmount record used to have a string "Unmount filesystem--" in the
878  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
879  * We just write the magic number now since that particular field isn't
880  * currently architecture converted and "Unmount" is a bit foo.
881  * As far as I know, there weren't any dependencies on the old behaviour.
882  */
883 static void
884 xfs_log_unmount_write(
885 	struct xfs_mount	*mp)
886 {
887 	struct xlog		*log = mp->m_log;
888 
889 	/*
890 	 * Don't write out unmount record on norecovery mounts or ro devices.
891 	 * Or, if we are doing a forced umount (typically because of IO errors).
892 	 */
893 	if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
894 	    xfs_readonly_buftarg(log->l_targ)) {
895 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
896 		return;
897 	}
898 
899 	xfs_log_force(mp, XFS_LOG_SYNC);
900 
901 	if (XLOG_FORCED_SHUTDOWN(log))
902 		return;
903 
904 	/*
905 	 * If we think the summary counters are bad, avoid writing the unmount
906 	 * record to force log recovery at next mount, after which the summary
907 	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
908 	 * more details.
909 	 */
910 	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
911 			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
912 		xfs_alert(mp, "%s: will fix summary counters at next mount",
913 				__func__);
914 		return;
915 	}
916 
917 	xfs_log_unmount_verify_iclog(log);
918 	xlog_unmount_write(log);
919 }
920 
921 /*
922  * Empty the log for unmount/freeze.
923  *
924  * To do this, we first need to shut down the background log work so it is not
925  * trying to cover the log as we clean up. We then need to unpin all objects in
926  * the log so we can then flush them out. Once they have completed their IO and
927  * run the callbacks removing themselves from the AIL, we can write the unmount
928  * record.
929  */
930 void
931 xfs_log_quiesce(
932 	struct xfs_mount	*mp)
933 {
934 	cancel_delayed_work_sync(&mp->m_log->l_work);
935 	xfs_log_force(mp, XFS_LOG_SYNC);
936 
937 	/*
938 	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
939 	 * will push it, xfs_wait_buftarg() will not wait for it. Further,
940 	 * xfs_buf_iowait() cannot be used because it was pushed with the
941 	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
942 	 * the IO to complete.
943 	 */
944 	xfs_ail_push_all_sync(mp->m_ail);
945 	xfs_wait_buftarg(mp->m_ddev_targp);
946 	xfs_buf_lock(mp->m_sb_bp);
947 	xfs_buf_unlock(mp->m_sb_bp);
948 
949 	xfs_log_unmount_write(mp);
950 }
951 
952 /*
953  * Shut down and release the AIL and Log.
954  *
955  * During unmount, we need to ensure we flush all the dirty metadata objects
956  * from the AIL so that the log is empty before we write the unmount record to
957  * the log. Once this is done, we can tear down the AIL and the log.
958  */
959 void
960 xfs_log_unmount(
961 	struct xfs_mount	*mp)
962 {
963 	xfs_log_quiesce(mp);
964 
965 	xfs_trans_ail_destroy(mp);
966 
967 	xfs_sysfs_del(&mp->m_log->l_kobj);
968 
969 	xlog_dealloc_log(mp->m_log);
970 }
971 
972 void
973 xfs_log_item_init(
974 	struct xfs_mount	*mp,
975 	struct xfs_log_item	*item,
976 	int			type,
977 	const struct xfs_item_ops *ops)
978 {
979 	item->li_mountp = mp;
980 	item->li_ailp = mp->m_ail;
981 	item->li_type = type;
982 	item->li_ops = ops;
983 	item->li_lv = NULL;
984 
985 	INIT_LIST_HEAD(&item->li_ail);
986 	INIT_LIST_HEAD(&item->li_cil);
987 	INIT_LIST_HEAD(&item->li_bio_list);
988 	INIT_LIST_HEAD(&item->li_trans);
989 }
990 
991 /*
992  * Wake up processes waiting for log space after we have moved the log tail.
993  */
994 void
995 xfs_log_space_wake(
996 	struct xfs_mount	*mp)
997 {
998 	struct xlog		*log = mp->m_log;
999 	int			free_bytes;
1000 
1001 	if (XLOG_FORCED_SHUTDOWN(log))
1002 		return;
1003 
1004 	if (!list_empty_careful(&log->l_write_head.waiters)) {
1005 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1006 
1007 		spin_lock(&log->l_write_head.lock);
1008 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1009 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1010 		spin_unlock(&log->l_write_head.lock);
1011 	}
1012 
1013 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1014 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1015 
1016 		spin_lock(&log->l_reserve_head.lock);
1017 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1018 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1019 		spin_unlock(&log->l_reserve_head.lock);
1020 	}
1021 }
1022 
1023 /*
1024  * Determine if we have a transaction that has gone to disk that needs to be
1025  * covered. To begin the transition to the idle state firstly the log needs to
1026  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1027  * we start attempting to cover the log.
1028  *
1029  * Only if we are then in a state where covering is needed, the caller is
1030  * informed that dummy transactions are required to move the log into the idle
1031  * state.
1032  *
1033  * If there are any items in the AIl or CIL, then we do not want to attempt to
1034  * cover the log as we may be in a situation where there isn't log space
1035  * available to run a dummy transaction and this can lead to deadlocks when the
1036  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1037  * there's no point in running a dummy transaction at this point because we
1038  * can't start trying to idle the log until both the CIL and AIL are empty.
1039  */
1040 static int
1041 xfs_log_need_covered(xfs_mount_t *mp)
1042 {
1043 	struct xlog	*log = mp->m_log;
1044 	int		needed = 0;
1045 
1046 	if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1047 		return 0;
1048 
1049 	if (!xlog_cil_empty(log))
1050 		return 0;
1051 
1052 	spin_lock(&log->l_icloglock);
1053 	switch (log->l_covered_state) {
1054 	case XLOG_STATE_COVER_DONE:
1055 	case XLOG_STATE_COVER_DONE2:
1056 	case XLOG_STATE_COVER_IDLE:
1057 		break;
1058 	case XLOG_STATE_COVER_NEED:
1059 	case XLOG_STATE_COVER_NEED2:
1060 		if (xfs_ail_min_lsn(log->l_ailp))
1061 			break;
1062 		if (!xlog_iclogs_empty(log))
1063 			break;
1064 
1065 		needed = 1;
1066 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1067 			log->l_covered_state = XLOG_STATE_COVER_DONE;
1068 		else
1069 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1070 		break;
1071 	default:
1072 		needed = 1;
1073 		break;
1074 	}
1075 	spin_unlock(&log->l_icloglock);
1076 	return needed;
1077 }
1078 
1079 /*
1080  * We may be holding the log iclog lock upon entering this routine.
1081  */
1082 xfs_lsn_t
1083 xlog_assign_tail_lsn_locked(
1084 	struct xfs_mount	*mp)
1085 {
1086 	struct xlog		*log = mp->m_log;
1087 	struct xfs_log_item	*lip;
1088 	xfs_lsn_t		tail_lsn;
1089 
1090 	assert_spin_locked(&mp->m_ail->ail_lock);
1091 
1092 	/*
1093 	 * To make sure we always have a valid LSN for the log tail we keep
1094 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1095 	 * and use that when the AIL was empty.
1096 	 */
1097 	lip = xfs_ail_min(mp->m_ail);
1098 	if (lip)
1099 		tail_lsn = lip->li_lsn;
1100 	else
1101 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1102 	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1103 	atomic64_set(&log->l_tail_lsn, tail_lsn);
1104 	return tail_lsn;
1105 }
1106 
1107 xfs_lsn_t
1108 xlog_assign_tail_lsn(
1109 	struct xfs_mount	*mp)
1110 {
1111 	xfs_lsn_t		tail_lsn;
1112 
1113 	spin_lock(&mp->m_ail->ail_lock);
1114 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1115 	spin_unlock(&mp->m_ail->ail_lock);
1116 
1117 	return tail_lsn;
1118 }
1119 
1120 /*
1121  * Return the space in the log between the tail and the head.  The head
1122  * is passed in the cycle/bytes formal parms.  In the special case where
1123  * the reserve head has wrapped passed the tail, this calculation is no
1124  * longer valid.  In this case, just return 0 which means there is no space
1125  * in the log.  This works for all places where this function is called
1126  * with the reserve head.  Of course, if the write head were to ever
1127  * wrap the tail, we should blow up.  Rather than catch this case here,
1128  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1129  *
1130  * This code also handles the case where the reservation head is behind
1131  * the tail.  The details of this case are described below, but the end
1132  * result is that we return the size of the log as the amount of space left.
1133  */
1134 STATIC int
1135 xlog_space_left(
1136 	struct xlog	*log,
1137 	atomic64_t	*head)
1138 {
1139 	int		free_bytes;
1140 	int		tail_bytes;
1141 	int		tail_cycle;
1142 	int		head_cycle;
1143 	int		head_bytes;
1144 
1145 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1146 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1147 	tail_bytes = BBTOB(tail_bytes);
1148 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1149 		free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1150 	else if (tail_cycle + 1 < head_cycle)
1151 		return 0;
1152 	else if (tail_cycle < head_cycle) {
1153 		ASSERT(tail_cycle == (head_cycle - 1));
1154 		free_bytes = tail_bytes - head_bytes;
1155 	} else {
1156 		/*
1157 		 * The reservation head is behind the tail.
1158 		 * In this case we just want to return the size of the
1159 		 * log as the amount of space left.
1160 		 */
1161 		xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1162 		xfs_alert(log->l_mp,
1163 			  "  tail_cycle = %d, tail_bytes = %d",
1164 			  tail_cycle, tail_bytes);
1165 		xfs_alert(log->l_mp,
1166 			  "  GH   cycle = %d, GH   bytes = %d",
1167 			  head_cycle, head_bytes);
1168 		ASSERT(0);
1169 		free_bytes = log->l_logsize;
1170 	}
1171 	return free_bytes;
1172 }
1173 
1174 
1175 static void
1176 xlog_ioend_work(
1177 	struct work_struct	*work)
1178 {
1179 	struct xlog_in_core     *iclog =
1180 		container_of(work, struct xlog_in_core, ic_end_io_work);
1181 	struct xlog		*log = iclog->ic_log;
1182 	int			error;
1183 
1184 	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1185 #ifdef DEBUG
1186 	/* treat writes with injected CRC errors as failed */
1187 	if (iclog->ic_fail_crc)
1188 		error = -EIO;
1189 #endif
1190 
1191 	/*
1192 	 * Race to shutdown the filesystem if we see an error.
1193 	 */
1194 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1195 		xfs_alert(log->l_mp, "log I/O error %d", error);
1196 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1197 	}
1198 
1199 	xlog_state_done_syncing(iclog);
1200 	bio_uninit(&iclog->ic_bio);
1201 
1202 	/*
1203 	 * Drop the lock to signal that we are done. Nothing references the
1204 	 * iclog after this, so an unmount waiting on this lock can now tear it
1205 	 * down safely. As such, it is unsafe to reference the iclog after the
1206 	 * unlock as we could race with it being freed.
1207 	 */
1208 	up(&iclog->ic_sema);
1209 }
1210 
1211 /*
1212  * Return size of each in-core log record buffer.
1213  *
1214  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1215  *
1216  * If the filesystem blocksize is too large, we may need to choose a
1217  * larger size since the directory code currently logs entire blocks.
1218  */
1219 STATIC void
1220 xlog_get_iclog_buffer_size(
1221 	struct xfs_mount	*mp,
1222 	struct xlog		*log)
1223 {
1224 	if (mp->m_logbufs <= 0)
1225 		mp->m_logbufs = XLOG_MAX_ICLOGS;
1226 	if (mp->m_logbsize <= 0)
1227 		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1228 
1229 	log->l_iclog_bufs = mp->m_logbufs;
1230 	log->l_iclog_size = mp->m_logbsize;
1231 
1232 	/*
1233 	 * # headers = size / 32k - one header holds cycles from 32k of data.
1234 	 */
1235 	log->l_iclog_heads =
1236 		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1237 	log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1238 }
1239 
1240 void
1241 xfs_log_work_queue(
1242 	struct xfs_mount        *mp)
1243 {
1244 	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1245 				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1246 }
1247 
1248 /*
1249  * Every sync period we need to unpin all items in the AIL and push them to
1250  * disk. If there is nothing dirty, then we might need to cover the log to
1251  * indicate that the filesystem is idle.
1252  */
1253 static void
1254 xfs_log_worker(
1255 	struct work_struct	*work)
1256 {
1257 	struct xlog		*log = container_of(to_delayed_work(work),
1258 						struct xlog, l_work);
1259 	struct xfs_mount	*mp = log->l_mp;
1260 
1261 	/* dgc: errors ignored - not fatal and nowhere to report them */
1262 	if (xfs_log_need_covered(mp)) {
1263 		/*
1264 		 * Dump a transaction into the log that contains no real change.
1265 		 * This is needed to stamp the current tail LSN into the log
1266 		 * during the covering operation.
1267 		 *
1268 		 * We cannot use an inode here for this - that will push dirty
1269 		 * state back up into the VFS and then periodic inode flushing
1270 		 * will prevent log covering from making progress. Hence we
1271 		 * synchronously log the superblock instead to ensure the
1272 		 * superblock is immediately unpinned and can be written back.
1273 		 */
1274 		xfs_sync_sb(mp, true);
1275 	} else
1276 		xfs_log_force(mp, 0);
1277 
1278 	/* start pushing all the metadata that is currently dirty */
1279 	xfs_ail_push_all(mp->m_ail);
1280 
1281 	/* queue us up again */
1282 	xfs_log_work_queue(mp);
1283 }
1284 
1285 /*
1286  * This routine initializes some of the log structure for a given mount point.
1287  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1288  * some other stuff may be filled in too.
1289  */
1290 STATIC struct xlog *
1291 xlog_alloc_log(
1292 	struct xfs_mount	*mp,
1293 	struct xfs_buftarg	*log_target,
1294 	xfs_daddr_t		blk_offset,
1295 	int			num_bblks)
1296 {
1297 	struct xlog		*log;
1298 	xlog_rec_header_t	*head;
1299 	xlog_in_core_t		**iclogp;
1300 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1301 	int			i;
1302 	int			error = -ENOMEM;
1303 	uint			log2_size = 0;
1304 
1305 	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1306 	if (!log) {
1307 		xfs_warn(mp, "Log allocation failed: No memory!");
1308 		goto out;
1309 	}
1310 
1311 	log->l_mp	   = mp;
1312 	log->l_targ	   = log_target;
1313 	log->l_logsize     = BBTOB(num_bblks);
1314 	log->l_logBBstart  = blk_offset;
1315 	log->l_logBBsize   = num_bblks;
1316 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1317 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1318 	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1319 
1320 	log->l_prev_block  = -1;
1321 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1322 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1323 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1324 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1325 
1326 	xlog_grant_head_init(&log->l_reserve_head);
1327 	xlog_grant_head_init(&log->l_write_head);
1328 
1329 	error = -EFSCORRUPTED;
1330 	if (xfs_sb_version_hassector(&mp->m_sb)) {
1331 	        log2_size = mp->m_sb.sb_logsectlog;
1332 		if (log2_size < BBSHIFT) {
1333 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1334 				log2_size, BBSHIFT);
1335 			goto out_free_log;
1336 		}
1337 
1338 	        log2_size -= BBSHIFT;
1339 		if (log2_size > mp->m_sectbb_log) {
1340 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1341 				log2_size, mp->m_sectbb_log);
1342 			goto out_free_log;
1343 		}
1344 
1345 		/* for larger sector sizes, must have v2 or external log */
1346 		if (log2_size && log->l_logBBstart > 0 &&
1347 			    !xfs_sb_version_haslogv2(&mp->m_sb)) {
1348 			xfs_warn(mp,
1349 		"log sector size (0x%x) invalid for configuration.",
1350 				log2_size);
1351 			goto out_free_log;
1352 		}
1353 	}
1354 	log->l_sectBBsize = 1 << log2_size;
1355 
1356 	xlog_get_iclog_buffer_size(mp, log);
1357 
1358 	spin_lock_init(&log->l_icloglock);
1359 	init_waitqueue_head(&log->l_flush_wait);
1360 
1361 	iclogp = &log->l_iclog;
1362 	/*
1363 	 * The amount of memory to allocate for the iclog structure is
1364 	 * rather funky due to the way the structure is defined.  It is
1365 	 * done this way so that we can use different sizes for machines
1366 	 * with different amounts of memory.  See the definition of
1367 	 * xlog_in_core_t in xfs_log_priv.h for details.
1368 	 */
1369 	ASSERT(log->l_iclog_size >= 4096);
1370 	for (i = 0; i < log->l_iclog_bufs; i++) {
1371 		int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1372 		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1373 				sizeof(struct bio_vec);
1374 
1375 		iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1376 		if (!iclog)
1377 			goto out_free_iclog;
1378 
1379 		*iclogp = iclog;
1380 		iclog->ic_prev = prev_iclog;
1381 		prev_iclog = iclog;
1382 
1383 		iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1384 						KM_MAYFAIL | KM_ZERO);
1385 		if (!iclog->ic_data)
1386 			goto out_free_iclog;
1387 #ifdef DEBUG
1388 		log->l_iclog_bak[i] = &iclog->ic_header;
1389 #endif
1390 		head = &iclog->ic_header;
1391 		memset(head, 0, sizeof(xlog_rec_header_t));
1392 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1393 		head->h_version = cpu_to_be32(
1394 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1395 		head->h_size = cpu_to_be32(log->l_iclog_size);
1396 		/* new fields */
1397 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1398 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1399 
1400 		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1401 		iclog->ic_state = XLOG_STATE_ACTIVE;
1402 		iclog->ic_log = log;
1403 		atomic_set(&iclog->ic_refcnt, 0);
1404 		spin_lock_init(&iclog->ic_callback_lock);
1405 		INIT_LIST_HEAD(&iclog->ic_callbacks);
1406 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1407 
1408 		init_waitqueue_head(&iclog->ic_force_wait);
1409 		init_waitqueue_head(&iclog->ic_write_wait);
1410 		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1411 		sema_init(&iclog->ic_sema, 1);
1412 
1413 		iclogp = &iclog->ic_next;
1414 	}
1415 	*iclogp = log->l_iclog;			/* complete ring */
1416 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1417 
1418 	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1419 			WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1420 			mp->m_super->s_id);
1421 	if (!log->l_ioend_workqueue)
1422 		goto out_free_iclog;
1423 
1424 	error = xlog_cil_init(log);
1425 	if (error)
1426 		goto out_destroy_workqueue;
1427 	return log;
1428 
1429 out_destroy_workqueue:
1430 	destroy_workqueue(log->l_ioend_workqueue);
1431 out_free_iclog:
1432 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1433 		prev_iclog = iclog->ic_next;
1434 		kmem_free(iclog->ic_data);
1435 		kmem_free(iclog);
1436 		if (prev_iclog == log->l_iclog)
1437 			break;
1438 	}
1439 out_free_log:
1440 	kmem_free(log);
1441 out:
1442 	return ERR_PTR(error);
1443 }	/* xlog_alloc_log */
1444 
1445 /*
1446  * Write out the commit record of a transaction associated with the given
1447  * ticket to close off a running log write. Return the lsn of the commit record.
1448  */
1449 int
1450 xlog_commit_record(
1451 	struct xlog		*log,
1452 	struct xlog_ticket	*ticket,
1453 	struct xlog_in_core	**iclog,
1454 	xfs_lsn_t		*lsn)
1455 {
1456 	struct xfs_log_iovec reg = {
1457 		.i_addr = NULL,
1458 		.i_len = 0,
1459 		.i_type = XLOG_REG_TYPE_COMMIT,
1460 	};
1461 	struct xfs_log_vec vec = {
1462 		.lv_niovecs = 1,
1463 		.lv_iovecp = &reg,
1464 	};
1465 	int	error;
1466 
1467 	if (XLOG_FORCED_SHUTDOWN(log))
1468 		return -EIO;
1469 
1470 	error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS,
1471 			   false);
1472 	if (error)
1473 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1474 	return error;
1475 }
1476 
1477 /*
1478  * Compute the LSN that we'd need to push the log tail towards in order to have
1479  * (a) enough on-disk log space to log the number of bytes specified, (b) at
1480  * least 25% of the log space free, and (c) at least 256 blocks free.  If the
1481  * log free space already meets all three thresholds, this function returns
1482  * NULLCOMMITLSN.
1483  */
1484 xfs_lsn_t
1485 xlog_grant_push_threshold(
1486 	struct xlog	*log,
1487 	int		need_bytes)
1488 {
1489 	xfs_lsn_t	threshold_lsn = 0;
1490 	xfs_lsn_t	last_sync_lsn;
1491 	int		free_blocks;
1492 	int		free_bytes;
1493 	int		threshold_block;
1494 	int		threshold_cycle;
1495 	int		free_threshold;
1496 
1497 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1498 
1499 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1500 	free_blocks = BTOBBT(free_bytes);
1501 
1502 	/*
1503 	 * Set the threshold for the minimum number of free blocks in the
1504 	 * log to the maximum of what the caller needs, one quarter of the
1505 	 * log, and 256 blocks.
1506 	 */
1507 	free_threshold = BTOBB(need_bytes);
1508 	free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1509 	free_threshold = max(free_threshold, 256);
1510 	if (free_blocks >= free_threshold)
1511 		return NULLCOMMITLSN;
1512 
1513 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1514 						&threshold_block);
1515 	threshold_block += free_threshold;
1516 	if (threshold_block >= log->l_logBBsize) {
1517 		threshold_block -= log->l_logBBsize;
1518 		threshold_cycle += 1;
1519 	}
1520 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1521 					threshold_block);
1522 	/*
1523 	 * Don't pass in an lsn greater than the lsn of the last
1524 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1525 	 * so that it doesn't change between the compare and the set.
1526 	 */
1527 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1528 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1529 		threshold_lsn = last_sync_lsn;
1530 
1531 	return threshold_lsn;
1532 }
1533 
1534 /*
1535  * Push the tail of the log if we need to do so to maintain the free log space
1536  * thresholds set out by xlog_grant_push_threshold.  We may need to adopt a
1537  * policy which pushes on an lsn which is further along in the log once we
1538  * reach the high water mark.  In this manner, we would be creating a low water
1539  * mark.
1540  */
1541 STATIC void
1542 xlog_grant_push_ail(
1543 	struct xlog	*log,
1544 	int		need_bytes)
1545 {
1546 	xfs_lsn_t	threshold_lsn;
1547 
1548 	threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1549 	if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log))
1550 		return;
1551 
1552 	/*
1553 	 * Get the transaction layer to kick the dirty buffers out to
1554 	 * disk asynchronously. No point in trying to do this if
1555 	 * the filesystem is shutting down.
1556 	 */
1557 	xfs_ail_push(log->l_ailp, threshold_lsn);
1558 }
1559 
1560 /*
1561  * Stamp cycle number in every block
1562  */
1563 STATIC void
1564 xlog_pack_data(
1565 	struct xlog		*log,
1566 	struct xlog_in_core	*iclog,
1567 	int			roundoff)
1568 {
1569 	int			i, j, k;
1570 	int			size = iclog->ic_offset + roundoff;
1571 	__be32			cycle_lsn;
1572 	char			*dp;
1573 
1574 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1575 
1576 	dp = iclog->ic_datap;
1577 	for (i = 0; i < BTOBB(size); i++) {
1578 		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1579 			break;
1580 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1581 		*(__be32 *)dp = cycle_lsn;
1582 		dp += BBSIZE;
1583 	}
1584 
1585 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1586 		xlog_in_core_2_t *xhdr = iclog->ic_data;
1587 
1588 		for ( ; i < BTOBB(size); i++) {
1589 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1590 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1591 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1592 			*(__be32 *)dp = cycle_lsn;
1593 			dp += BBSIZE;
1594 		}
1595 
1596 		for (i = 1; i < log->l_iclog_heads; i++)
1597 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1598 	}
1599 }
1600 
1601 /*
1602  * Calculate the checksum for a log buffer.
1603  *
1604  * This is a little more complicated than it should be because the various
1605  * headers and the actual data are non-contiguous.
1606  */
1607 __le32
1608 xlog_cksum(
1609 	struct xlog		*log,
1610 	struct xlog_rec_header	*rhead,
1611 	char			*dp,
1612 	int			size)
1613 {
1614 	uint32_t		crc;
1615 
1616 	/* first generate the crc for the record header ... */
1617 	crc = xfs_start_cksum_update((char *)rhead,
1618 			      sizeof(struct xlog_rec_header),
1619 			      offsetof(struct xlog_rec_header, h_crc));
1620 
1621 	/* ... then for additional cycle data for v2 logs ... */
1622 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1623 		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1624 		int		i;
1625 		int		xheads;
1626 
1627 		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1628 
1629 		for (i = 1; i < xheads; i++) {
1630 			crc = crc32c(crc, &xhdr[i].hic_xheader,
1631 				     sizeof(struct xlog_rec_ext_header));
1632 		}
1633 	}
1634 
1635 	/* ... and finally for the payload */
1636 	crc = crc32c(crc, dp, size);
1637 
1638 	return xfs_end_cksum(crc);
1639 }
1640 
1641 static void
1642 xlog_bio_end_io(
1643 	struct bio		*bio)
1644 {
1645 	struct xlog_in_core	*iclog = bio->bi_private;
1646 
1647 	queue_work(iclog->ic_log->l_ioend_workqueue,
1648 		   &iclog->ic_end_io_work);
1649 }
1650 
1651 static int
1652 xlog_map_iclog_data(
1653 	struct bio		*bio,
1654 	void			*data,
1655 	size_t			count)
1656 {
1657 	do {
1658 		struct page	*page = kmem_to_page(data);
1659 		unsigned int	off = offset_in_page(data);
1660 		size_t		len = min_t(size_t, count, PAGE_SIZE - off);
1661 
1662 		if (bio_add_page(bio, page, len, off) != len)
1663 			return -EIO;
1664 
1665 		data += len;
1666 		count -= len;
1667 	} while (count);
1668 
1669 	return 0;
1670 }
1671 
1672 STATIC void
1673 xlog_write_iclog(
1674 	struct xlog		*log,
1675 	struct xlog_in_core	*iclog,
1676 	uint64_t		bno,
1677 	unsigned int		count,
1678 	bool			need_flush)
1679 {
1680 	ASSERT(bno < log->l_logBBsize);
1681 
1682 	/*
1683 	 * We lock the iclogbufs here so that we can serialise against I/O
1684 	 * completion during unmount.  We might be processing a shutdown
1685 	 * triggered during unmount, and that can occur asynchronously to the
1686 	 * unmount thread, and hence we need to ensure that completes before
1687 	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1688 	 * across the log IO to archieve that.
1689 	 */
1690 	down(&iclog->ic_sema);
1691 	if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1692 		/*
1693 		 * It would seem logical to return EIO here, but we rely on
1694 		 * the log state machine to propagate I/O errors instead of
1695 		 * doing it here.  We kick of the state machine and unlock
1696 		 * the buffer manually, the code needs to be kept in sync
1697 		 * with the I/O completion path.
1698 		 */
1699 		xlog_state_done_syncing(iclog);
1700 		up(&iclog->ic_sema);
1701 		return;
1702 	}
1703 
1704 	bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1705 	bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1706 	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1707 	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1708 	iclog->ic_bio.bi_private = iclog;
1709 
1710 	/*
1711 	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1712 	 * IOs coming immediately after this one. This prevents the block layer
1713 	 * writeback throttle from throttling log writes behind background
1714 	 * metadata writeback and causing priority inversions.
1715 	 */
1716 	iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC |
1717 				REQ_IDLE | REQ_FUA;
1718 	if (need_flush)
1719 		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1720 
1721 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1722 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1723 		return;
1724 	}
1725 	if (is_vmalloc_addr(iclog->ic_data))
1726 		flush_kernel_vmap_range(iclog->ic_data, count);
1727 
1728 	/*
1729 	 * If this log buffer would straddle the end of the log we will have
1730 	 * to split it up into two bios, so that we can continue at the start.
1731 	 */
1732 	if (bno + BTOBB(count) > log->l_logBBsize) {
1733 		struct bio *split;
1734 
1735 		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1736 				  GFP_NOIO, &fs_bio_set);
1737 		bio_chain(split, &iclog->ic_bio);
1738 		submit_bio(split);
1739 
1740 		/* restart at logical offset zero for the remainder */
1741 		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1742 	}
1743 
1744 	submit_bio(&iclog->ic_bio);
1745 }
1746 
1747 /*
1748  * We need to bump cycle number for the part of the iclog that is
1749  * written to the start of the log. Watch out for the header magic
1750  * number case, though.
1751  */
1752 static void
1753 xlog_split_iclog(
1754 	struct xlog		*log,
1755 	void			*data,
1756 	uint64_t		bno,
1757 	unsigned int		count)
1758 {
1759 	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1760 	unsigned int		i;
1761 
1762 	for (i = split_offset; i < count; i += BBSIZE) {
1763 		uint32_t cycle = get_unaligned_be32(data + i);
1764 
1765 		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1766 			cycle++;
1767 		put_unaligned_be32(cycle, data + i);
1768 	}
1769 }
1770 
1771 static int
1772 xlog_calc_iclog_size(
1773 	struct xlog		*log,
1774 	struct xlog_in_core	*iclog,
1775 	uint32_t		*roundoff)
1776 {
1777 	uint32_t		count_init, count;
1778 	bool			use_lsunit;
1779 
1780 	use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1781 			log->l_mp->m_sb.sb_logsunit > 1;
1782 
1783 	/* Add for LR header */
1784 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1785 
1786 	/* Round out the log write size */
1787 	if (use_lsunit) {
1788 		/* we have a v2 stripe unit to use */
1789 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1790 	} else {
1791 		count = BBTOB(BTOBB(count_init));
1792 	}
1793 
1794 	ASSERT(count >= count_init);
1795 	*roundoff = count - count_init;
1796 
1797 	if (use_lsunit)
1798 		ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1799 	else
1800 		ASSERT(*roundoff < BBTOB(1));
1801 	return count;
1802 }
1803 
1804 /*
1805  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1806  * fashion.  Previously, we should have moved the current iclog
1807  * ptr in the log to point to the next available iclog.  This allows further
1808  * write to continue while this code syncs out an iclog ready to go.
1809  * Before an in-core log can be written out, the data section must be scanned
1810  * to save away the 1st word of each BBSIZE block into the header.  We replace
1811  * it with the current cycle count.  Each BBSIZE block is tagged with the
1812  * cycle count because there in an implicit assumption that drives will
1813  * guarantee that entire 512 byte blocks get written at once.  In other words,
1814  * we can't have part of a 512 byte block written and part not written.  By
1815  * tagging each block, we will know which blocks are valid when recovering
1816  * after an unclean shutdown.
1817  *
1818  * This routine is single threaded on the iclog.  No other thread can be in
1819  * this routine with the same iclog.  Changing contents of iclog can there-
1820  * fore be done without grabbing the state machine lock.  Updating the global
1821  * log will require grabbing the lock though.
1822  *
1823  * The entire log manager uses a logical block numbering scheme.  Only
1824  * xlog_write_iclog knows about the fact that the log may not start with
1825  * block zero on a given device.
1826  */
1827 STATIC void
1828 xlog_sync(
1829 	struct xlog		*log,
1830 	struct xlog_in_core	*iclog)
1831 {
1832 	unsigned int		count;		/* byte count of bwrite */
1833 	unsigned int		roundoff;       /* roundoff to BB or stripe */
1834 	uint64_t		bno;
1835 	unsigned int		size;
1836 	bool			need_flush = true, split = false;
1837 
1838 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1839 
1840 	count = xlog_calc_iclog_size(log, iclog, &roundoff);
1841 
1842 	/* move grant heads by roundoff in sync */
1843 	xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1844 	xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1845 
1846 	/* put cycle number in every block */
1847 	xlog_pack_data(log, iclog, roundoff);
1848 
1849 	/* real byte length */
1850 	size = iclog->ic_offset;
1851 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1852 		size += roundoff;
1853 	iclog->ic_header.h_len = cpu_to_be32(size);
1854 
1855 	XFS_STATS_INC(log->l_mp, xs_log_writes);
1856 	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1857 
1858 	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1859 
1860 	/* Do we need to split this write into 2 parts? */
1861 	if (bno + BTOBB(count) > log->l_logBBsize) {
1862 		xlog_split_iclog(log, &iclog->ic_header, bno, count);
1863 		split = true;
1864 	}
1865 
1866 	/* calculcate the checksum */
1867 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1868 					    iclog->ic_datap, size);
1869 	/*
1870 	 * Intentionally corrupt the log record CRC based on the error injection
1871 	 * frequency, if defined. This facilitates testing log recovery in the
1872 	 * event of torn writes. Hence, set the IOABORT state to abort the log
1873 	 * write on I/O completion and shutdown the fs. The subsequent mount
1874 	 * detects the bad CRC and attempts to recover.
1875 	 */
1876 #ifdef DEBUG
1877 	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1878 		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1879 		iclog->ic_fail_crc = true;
1880 		xfs_warn(log->l_mp,
1881 	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1882 			 be64_to_cpu(iclog->ic_header.h_lsn));
1883 	}
1884 #endif
1885 
1886 	/*
1887 	 * Flush the data device before flushing the log to make sure all meta
1888 	 * data written back from the AIL actually made it to disk before
1889 	 * stamping the new log tail LSN into the log buffer.  For an external
1890 	 * log we need to issue the flush explicitly, and unfortunately
1891 	 * synchronously here; for an internal log we can simply use the block
1892 	 * layer state machine for preflushes.
1893 	 */
1894 	if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1895 		xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1896 		need_flush = false;
1897 	}
1898 
1899 	xlog_verify_iclog(log, iclog, count);
1900 	xlog_write_iclog(log, iclog, bno, count, need_flush);
1901 }
1902 
1903 /*
1904  * Deallocate a log structure
1905  */
1906 STATIC void
1907 xlog_dealloc_log(
1908 	struct xlog	*log)
1909 {
1910 	xlog_in_core_t	*iclog, *next_iclog;
1911 	int		i;
1912 
1913 	xlog_cil_destroy(log);
1914 
1915 	/*
1916 	 * Cycle all the iclogbuf locks to make sure all log IO completion
1917 	 * is done before we tear down these buffers.
1918 	 */
1919 	iclog = log->l_iclog;
1920 	for (i = 0; i < log->l_iclog_bufs; i++) {
1921 		down(&iclog->ic_sema);
1922 		up(&iclog->ic_sema);
1923 		iclog = iclog->ic_next;
1924 	}
1925 
1926 	iclog = log->l_iclog;
1927 	for (i = 0; i < log->l_iclog_bufs; i++) {
1928 		next_iclog = iclog->ic_next;
1929 		kmem_free(iclog->ic_data);
1930 		kmem_free(iclog);
1931 		iclog = next_iclog;
1932 	}
1933 
1934 	log->l_mp->m_log = NULL;
1935 	destroy_workqueue(log->l_ioend_workqueue);
1936 	kmem_free(log);
1937 }
1938 
1939 /*
1940  * Update counters atomically now that memcpy is done.
1941  */
1942 static inline void
1943 xlog_state_finish_copy(
1944 	struct xlog		*log,
1945 	struct xlog_in_core	*iclog,
1946 	int			record_cnt,
1947 	int			copy_bytes)
1948 {
1949 	lockdep_assert_held(&log->l_icloglock);
1950 
1951 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1952 	iclog->ic_offset += copy_bytes;
1953 }
1954 
1955 /*
1956  * print out info relating to regions written which consume
1957  * the reservation
1958  */
1959 void
1960 xlog_print_tic_res(
1961 	struct xfs_mount	*mp,
1962 	struct xlog_ticket	*ticket)
1963 {
1964 	uint i;
1965 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1966 
1967 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1968 #define REG_TYPE_STR(type, str)	[XLOG_REG_TYPE_##type] = str
1969 	static char *res_type_str[] = {
1970 	    REG_TYPE_STR(BFORMAT, "bformat"),
1971 	    REG_TYPE_STR(BCHUNK, "bchunk"),
1972 	    REG_TYPE_STR(EFI_FORMAT, "efi_format"),
1973 	    REG_TYPE_STR(EFD_FORMAT, "efd_format"),
1974 	    REG_TYPE_STR(IFORMAT, "iformat"),
1975 	    REG_TYPE_STR(ICORE, "icore"),
1976 	    REG_TYPE_STR(IEXT, "iext"),
1977 	    REG_TYPE_STR(IBROOT, "ibroot"),
1978 	    REG_TYPE_STR(ILOCAL, "ilocal"),
1979 	    REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
1980 	    REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
1981 	    REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
1982 	    REG_TYPE_STR(QFORMAT, "qformat"),
1983 	    REG_TYPE_STR(DQUOT, "dquot"),
1984 	    REG_TYPE_STR(QUOTAOFF, "quotaoff"),
1985 	    REG_TYPE_STR(LRHEADER, "LR header"),
1986 	    REG_TYPE_STR(UNMOUNT, "unmount"),
1987 	    REG_TYPE_STR(COMMIT, "commit"),
1988 	    REG_TYPE_STR(TRANSHDR, "trans header"),
1989 	    REG_TYPE_STR(ICREATE, "inode create"),
1990 	    REG_TYPE_STR(RUI_FORMAT, "rui_format"),
1991 	    REG_TYPE_STR(RUD_FORMAT, "rud_format"),
1992 	    REG_TYPE_STR(CUI_FORMAT, "cui_format"),
1993 	    REG_TYPE_STR(CUD_FORMAT, "cud_format"),
1994 	    REG_TYPE_STR(BUI_FORMAT, "bui_format"),
1995 	    REG_TYPE_STR(BUD_FORMAT, "bud_format"),
1996 	};
1997 	BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
1998 #undef REG_TYPE_STR
1999 
2000 	xfs_warn(mp, "ticket reservation summary:");
2001 	xfs_warn(mp, "  unit res    = %d bytes",
2002 		 ticket->t_unit_res);
2003 	xfs_warn(mp, "  current res = %d bytes",
2004 		 ticket->t_curr_res);
2005 	xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2006 		 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2007 	xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2008 		 ticket->t_res_num_ophdrs, ophdr_spc);
2009 	xfs_warn(mp, "  ophdr + reg = %u bytes",
2010 		 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2011 	xfs_warn(mp, "  num regions = %u",
2012 		 ticket->t_res_num);
2013 
2014 	for (i = 0; i < ticket->t_res_num; i++) {
2015 		uint r_type = ticket->t_res_arr[i].r_type;
2016 		xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2017 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2018 			    "bad-rtype" : res_type_str[r_type]),
2019 			    ticket->t_res_arr[i].r_len);
2020 	}
2021 }
2022 
2023 /*
2024  * Print a summary of the transaction.
2025  */
2026 void
2027 xlog_print_trans(
2028 	struct xfs_trans	*tp)
2029 {
2030 	struct xfs_mount	*mp = tp->t_mountp;
2031 	struct xfs_log_item	*lip;
2032 
2033 	/* dump core transaction and ticket info */
2034 	xfs_warn(mp, "transaction summary:");
2035 	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2036 	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2037 	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2038 
2039 	xlog_print_tic_res(mp, tp->t_ticket);
2040 
2041 	/* dump each log item */
2042 	list_for_each_entry(lip, &tp->t_items, li_trans) {
2043 		struct xfs_log_vec	*lv = lip->li_lv;
2044 		struct xfs_log_iovec	*vec;
2045 		int			i;
2046 
2047 		xfs_warn(mp, "log item: ");
2048 		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2049 		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
2050 		if (!lv)
2051 			continue;
2052 		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2053 		xfs_warn(mp, "  size	= %d", lv->lv_size);
2054 		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2055 		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2056 
2057 		/* dump each iovec for the log item */
2058 		vec = lv->lv_iovecp;
2059 		for (i = 0; i < lv->lv_niovecs; i++) {
2060 			int dumplen = min(vec->i_len, 32);
2061 
2062 			xfs_warn(mp, "  iovec[%d]", i);
2063 			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2064 			xfs_warn(mp, "    len	= %d", vec->i_len);
2065 			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2066 			xfs_hex_dump(vec->i_addr, dumplen);
2067 
2068 			vec++;
2069 		}
2070 	}
2071 }
2072 
2073 /*
2074  * Calculate the potential space needed by the log vector.  We may need a start
2075  * record, and each region gets its own struct xlog_op_header and may need to be
2076  * double word aligned.
2077  */
2078 static int
2079 xlog_write_calc_vec_length(
2080 	struct xlog_ticket	*ticket,
2081 	struct xfs_log_vec	*log_vector,
2082 	bool			need_start_rec)
2083 {
2084 	struct xfs_log_vec	*lv;
2085 	int			headers = need_start_rec ? 1 : 0;
2086 	int			len = 0;
2087 	int			i;
2088 
2089 	for (lv = log_vector; lv; lv = lv->lv_next) {
2090 		/* we don't write ordered log vectors */
2091 		if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2092 			continue;
2093 
2094 		headers += lv->lv_niovecs;
2095 
2096 		for (i = 0; i < lv->lv_niovecs; i++) {
2097 			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
2098 
2099 			len += vecp->i_len;
2100 			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2101 		}
2102 	}
2103 
2104 	ticket->t_res_num_ophdrs += headers;
2105 	len += headers * sizeof(struct xlog_op_header);
2106 
2107 	return len;
2108 }
2109 
2110 static void
2111 xlog_write_start_rec(
2112 	struct xlog_op_header	*ophdr,
2113 	struct xlog_ticket	*ticket)
2114 {
2115 	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
2116 	ophdr->oh_clientid = ticket->t_clientid;
2117 	ophdr->oh_len = 0;
2118 	ophdr->oh_flags = XLOG_START_TRANS;
2119 	ophdr->oh_res2 = 0;
2120 }
2121 
2122 static xlog_op_header_t *
2123 xlog_write_setup_ophdr(
2124 	struct xlog		*log,
2125 	struct xlog_op_header	*ophdr,
2126 	struct xlog_ticket	*ticket,
2127 	uint			flags)
2128 {
2129 	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2130 	ophdr->oh_clientid = ticket->t_clientid;
2131 	ophdr->oh_res2 = 0;
2132 
2133 	/* are we copying a commit or unmount record? */
2134 	ophdr->oh_flags = flags;
2135 
2136 	/*
2137 	 * We've seen logs corrupted with bad transaction client ids.  This
2138 	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2139 	 * and shut down the filesystem.
2140 	 */
2141 	switch (ophdr->oh_clientid)  {
2142 	case XFS_TRANSACTION:
2143 	case XFS_VOLUME:
2144 	case XFS_LOG:
2145 		break;
2146 	default:
2147 		xfs_warn(log->l_mp,
2148 			"Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2149 			ophdr->oh_clientid, ticket);
2150 		return NULL;
2151 	}
2152 
2153 	return ophdr;
2154 }
2155 
2156 /*
2157  * Set up the parameters of the region copy into the log. This has
2158  * to handle region write split across multiple log buffers - this
2159  * state is kept external to this function so that this code can
2160  * be written in an obvious, self documenting manner.
2161  */
2162 static int
2163 xlog_write_setup_copy(
2164 	struct xlog_ticket	*ticket,
2165 	struct xlog_op_header	*ophdr,
2166 	int			space_available,
2167 	int			space_required,
2168 	int			*copy_off,
2169 	int			*copy_len,
2170 	int			*last_was_partial_copy,
2171 	int			*bytes_consumed)
2172 {
2173 	int			still_to_copy;
2174 
2175 	still_to_copy = space_required - *bytes_consumed;
2176 	*copy_off = *bytes_consumed;
2177 
2178 	if (still_to_copy <= space_available) {
2179 		/* write of region completes here */
2180 		*copy_len = still_to_copy;
2181 		ophdr->oh_len = cpu_to_be32(*copy_len);
2182 		if (*last_was_partial_copy)
2183 			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2184 		*last_was_partial_copy = 0;
2185 		*bytes_consumed = 0;
2186 		return 0;
2187 	}
2188 
2189 	/* partial write of region, needs extra log op header reservation */
2190 	*copy_len = space_available;
2191 	ophdr->oh_len = cpu_to_be32(*copy_len);
2192 	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2193 	if (*last_was_partial_copy)
2194 		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2195 	*bytes_consumed += *copy_len;
2196 	(*last_was_partial_copy)++;
2197 
2198 	/* account for new log op header */
2199 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2200 	ticket->t_res_num_ophdrs++;
2201 
2202 	return sizeof(struct xlog_op_header);
2203 }
2204 
2205 static int
2206 xlog_write_copy_finish(
2207 	struct xlog		*log,
2208 	struct xlog_in_core	*iclog,
2209 	uint			flags,
2210 	int			*record_cnt,
2211 	int			*data_cnt,
2212 	int			*partial_copy,
2213 	int			*partial_copy_len,
2214 	int			log_offset,
2215 	struct xlog_in_core	**commit_iclog)
2216 {
2217 	int			error;
2218 
2219 	if (*partial_copy) {
2220 		/*
2221 		 * This iclog has already been marked WANT_SYNC by
2222 		 * xlog_state_get_iclog_space.
2223 		 */
2224 		spin_lock(&log->l_icloglock);
2225 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2226 		*record_cnt = 0;
2227 		*data_cnt = 0;
2228 		goto release_iclog;
2229 	}
2230 
2231 	*partial_copy = 0;
2232 	*partial_copy_len = 0;
2233 
2234 	if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2235 		/* no more space in this iclog - push it. */
2236 		spin_lock(&log->l_icloglock);
2237 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2238 		*record_cnt = 0;
2239 		*data_cnt = 0;
2240 
2241 		if (iclog->ic_state == XLOG_STATE_ACTIVE)
2242 			xlog_state_switch_iclogs(log, iclog, 0);
2243 		else
2244 			ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2245 			       iclog->ic_state == XLOG_STATE_IOERROR);
2246 		if (!commit_iclog)
2247 			goto release_iclog;
2248 		spin_unlock(&log->l_icloglock);
2249 		ASSERT(flags & XLOG_COMMIT_TRANS);
2250 		*commit_iclog = iclog;
2251 	}
2252 
2253 	return 0;
2254 
2255 release_iclog:
2256 	error = xlog_state_release_iclog(log, iclog);
2257 	spin_unlock(&log->l_icloglock);
2258 	return error;
2259 }
2260 
2261 /*
2262  * Write some region out to in-core log
2263  *
2264  * This will be called when writing externally provided regions or when
2265  * writing out a commit record for a given transaction.
2266  *
2267  * General algorithm:
2268  *	1. Find total length of this write.  This may include adding to the
2269  *		lengths passed in.
2270  *	2. Check whether we violate the tickets reservation.
2271  *	3. While writing to this iclog
2272  *	    A. Reserve as much space in this iclog as can get
2273  *	    B. If this is first write, save away start lsn
2274  *	    C. While writing this region:
2275  *		1. If first write of transaction, write start record
2276  *		2. Write log operation header (header per region)
2277  *		3. Find out if we can fit entire region into this iclog
2278  *		4. Potentially, verify destination memcpy ptr
2279  *		5. Memcpy (partial) region
2280  *		6. If partial copy, release iclog; otherwise, continue
2281  *			copying more regions into current iclog
2282  *	4. Mark want sync bit (in simulation mode)
2283  *	5. Release iclog for potential flush to on-disk log.
2284  *
2285  * ERRORS:
2286  * 1.	Panic if reservation is overrun.  This should never happen since
2287  *	reservation amounts are generated internal to the filesystem.
2288  * NOTES:
2289  * 1. Tickets are single threaded data structures.
2290  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2291  *	syncing routine.  When a single log_write region needs to span
2292  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2293  *	on all log operation writes which don't contain the end of the
2294  *	region.  The XLOG_END_TRANS bit is used for the in-core log
2295  *	operation which contains the end of the continued log_write region.
2296  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2297  *	we don't really know exactly how much space will be used.  As a result,
2298  *	we don't update ic_offset until the end when we know exactly how many
2299  *	bytes have been written out.
2300  */
2301 int
2302 xlog_write(
2303 	struct xlog		*log,
2304 	struct xfs_log_vec	*log_vector,
2305 	struct xlog_ticket	*ticket,
2306 	xfs_lsn_t		*start_lsn,
2307 	struct xlog_in_core	**commit_iclog,
2308 	uint			flags,
2309 	bool			need_start_rec)
2310 {
2311 	struct xlog_in_core	*iclog = NULL;
2312 	struct xfs_log_vec	*lv = log_vector;
2313 	struct xfs_log_iovec	*vecp = lv->lv_iovecp;
2314 	int			index = 0;
2315 	int			len;
2316 	int			partial_copy = 0;
2317 	int			partial_copy_len = 0;
2318 	int			contwr = 0;
2319 	int			record_cnt = 0;
2320 	int			data_cnt = 0;
2321 	int			error = 0;
2322 
2323 	/*
2324 	 * If this is a commit or unmount transaction, we don't need a start
2325 	 * record to be written.  We do, however, have to account for the
2326 	 * commit or unmount header that gets written. Hence we always have
2327 	 * to account for an extra xlog_op_header here.
2328 	 */
2329 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2330 	if (ticket->t_curr_res < 0) {
2331 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2332 		     "ctx ticket reservation ran out. Need to up reservation");
2333 		xlog_print_tic_res(log->l_mp, ticket);
2334 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2335 	}
2336 
2337 	len = xlog_write_calc_vec_length(ticket, log_vector, need_start_rec);
2338 	*start_lsn = 0;
2339 	while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2340 		void		*ptr;
2341 		int		log_offset;
2342 
2343 		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2344 						   &contwr, &log_offset);
2345 		if (error)
2346 			return error;
2347 
2348 		ASSERT(log_offset <= iclog->ic_size - 1);
2349 		ptr = iclog->ic_datap + log_offset;
2350 
2351 		/* start_lsn is the first lsn written to. That's all we need. */
2352 		if (!*start_lsn)
2353 			*start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2354 
2355 		/*
2356 		 * This loop writes out as many regions as can fit in the amount
2357 		 * of space which was allocated by xlog_state_get_iclog_space().
2358 		 */
2359 		while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2360 			struct xfs_log_iovec	*reg;
2361 			struct xlog_op_header	*ophdr;
2362 			int			copy_len;
2363 			int			copy_off;
2364 			bool			ordered = false;
2365 
2366 			/* ordered log vectors have no regions to write */
2367 			if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2368 				ASSERT(lv->lv_niovecs == 0);
2369 				ordered = true;
2370 				goto next_lv;
2371 			}
2372 
2373 			reg = &vecp[index];
2374 			ASSERT(reg->i_len % sizeof(int32_t) == 0);
2375 			ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2376 
2377 			/*
2378 			 * Before we start formatting log vectors, we need to
2379 			 * write a start record. Only do this for the first
2380 			 * iclog we write to.
2381 			 */
2382 			if (need_start_rec) {
2383 				xlog_write_start_rec(ptr, ticket);
2384 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2385 						sizeof(struct xlog_op_header));
2386 			}
2387 
2388 			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2389 			if (!ophdr)
2390 				return -EIO;
2391 
2392 			xlog_write_adv_cnt(&ptr, &len, &log_offset,
2393 					   sizeof(struct xlog_op_header));
2394 
2395 			len += xlog_write_setup_copy(ticket, ophdr,
2396 						     iclog->ic_size-log_offset,
2397 						     reg->i_len,
2398 						     &copy_off, &copy_len,
2399 						     &partial_copy,
2400 						     &partial_copy_len);
2401 			xlog_verify_dest_ptr(log, ptr);
2402 
2403 			/*
2404 			 * Copy region.
2405 			 *
2406 			 * Unmount records just log an opheader, so can have
2407 			 * empty payloads with no data region to copy. Hence we
2408 			 * only copy the payload if the vector says it has data
2409 			 * to copy.
2410 			 */
2411 			ASSERT(copy_len >= 0);
2412 			if (copy_len > 0) {
2413 				memcpy(ptr, reg->i_addr + copy_off, copy_len);
2414 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2415 						   copy_len);
2416 			}
2417 			copy_len += sizeof(struct xlog_op_header);
2418 			record_cnt++;
2419 			if (need_start_rec) {
2420 				copy_len += sizeof(struct xlog_op_header);
2421 				record_cnt++;
2422 				need_start_rec = false;
2423 			}
2424 			data_cnt += contwr ? copy_len : 0;
2425 
2426 			error = xlog_write_copy_finish(log, iclog, flags,
2427 						       &record_cnt, &data_cnt,
2428 						       &partial_copy,
2429 						       &partial_copy_len,
2430 						       log_offset,
2431 						       commit_iclog);
2432 			if (error)
2433 				return error;
2434 
2435 			/*
2436 			 * if we had a partial copy, we need to get more iclog
2437 			 * space but we don't want to increment the region
2438 			 * index because there is still more is this region to
2439 			 * write.
2440 			 *
2441 			 * If we completed writing this region, and we flushed
2442 			 * the iclog (indicated by resetting of the record
2443 			 * count), then we also need to get more log space. If
2444 			 * this was the last record, though, we are done and
2445 			 * can just return.
2446 			 */
2447 			if (partial_copy)
2448 				break;
2449 
2450 			if (++index == lv->lv_niovecs) {
2451 next_lv:
2452 				lv = lv->lv_next;
2453 				index = 0;
2454 				if (lv)
2455 					vecp = lv->lv_iovecp;
2456 			}
2457 			if (record_cnt == 0 && !ordered) {
2458 				if (!lv)
2459 					return 0;
2460 				break;
2461 			}
2462 		}
2463 	}
2464 
2465 	ASSERT(len == 0);
2466 
2467 	spin_lock(&log->l_icloglock);
2468 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2469 	if (commit_iclog) {
2470 		ASSERT(flags & XLOG_COMMIT_TRANS);
2471 		*commit_iclog = iclog;
2472 	} else {
2473 		error = xlog_state_release_iclog(log, iclog);
2474 	}
2475 	spin_unlock(&log->l_icloglock);
2476 
2477 	return error;
2478 }
2479 
2480 static void
2481 xlog_state_activate_iclog(
2482 	struct xlog_in_core	*iclog,
2483 	int			*iclogs_changed)
2484 {
2485 	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2486 
2487 	/*
2488 	 * If the number of ops in this iclog indicate it just contains the
2489 	 * dummy transaction, we can change state into IDLE (the second time
2490 	 * around). Otherwise we should change the state into NEED a dummy.
2491 	 * We don't need to cover the dummy.
2492 	 */
2493 	if (*iclogs_changed == 0 &&
2494 	    iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2495 		*iclogs_changed = 1;
2496 	} else {
2497 		/*
2498 		 * We have two dirty iclogs so start over.  This could also be
2499 		 * num of ops indicating this is not the dummy going out.
2500 		 */
2501 		*iclogs_changed = 2;
2502 	}
2503 
2504 	iclog->ic_state	= XLOG_STATE_ACTIVE;
2505 	iclog->ic_offset = 0;
2506 	iclog->ic_header.h_num_logops = 0;
2507 	memset(iclog->ic_header.h_cycle_data, 0,
2508 		sizeof(iclog->ic_header.h_cycle_data));
2509 	iclog->ic_header.h_lsn = 0;
2510 }
2511 
2512 /*
2513  * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2514  * ACTIVE after iclog I/O has completed.
2515  */
2516 static void
2517 xlog_state_activate_iclogs(
2518 	struct xlog		*log,
2519 	int			*iclogs_changed)
2520 {
2521 	struct xlog_in_core	*iclog = log->l_iclog;
2522 
2523 	do {
2524 		if (iclog->ic_state == XLOG_STATE_DIRTY)
2525 			xlog_state_activate_iclog(iclog, iclogs_changed);
2526 		/*
2527 		 * The ordering of marking iclogs ACTIVE must be maintained, so
2528 		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2529 		 */
2530 		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2531 			break;
2532 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2533 }
2534 
2535 static int
2536 xlog_covered_state(
2537 	int			prev_state,
2538 	int			iclogs_changed)
2539 {
2540 	/*
2541 	 * We usually go to NEED. But we go to NEED2 if the changed indicates we
2542 	 * are done writing the dummy record.  If we are done with the second
2543 	 * dummy recored (DONE2), then we go to IDLE.
2544 	 */
2545 	switch (prev_state) {
2546 	case XLOG_STATE_COVER_IDLE:
2547 	case XLOG_STATE_COVER_NEED:
2548 	case XLOG_STATE_COVER_NEED2:
2549 		break;
2550 	case XLOG_STATE_COVER_DONE:
2551 		if (iclogs_changed == 1)
2552 			return XLOG_STATE_COVER_NEED2;
2553 		break;
2554 	case XLOG_STATE_COVER_DONE2:
2555 		if (iclogs_changed == 1)
2556 			return XLOG_STATE_COVER_IDLE;
2557 		break;
2558 	default:
2559 		ASSERT(0);
2560 	}
2561 
2562 	return XLOG_STATE_COVER_NEED;
2563 }
2564 
2565 STATIC void
2566 xlog_state_clean_iclog(
2567 	struct xlog		*log,
2568 	struct xlog_in_core	*dirty_iclog)
2569 {
2570 	int			iclogs_changed = 0;
2571 
2572 	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2573 
2574 	xlog_state_activate_iclogs(log, &iclogs_changed);
2575 	wake_up_all(&dirty_iclog->ic_force_wait);
2576 
2577 	if (iclogs_changed) {
2578 		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2579 				iclogs_changed);
2580 	}
2581 }
2582 
2583 STATIC xfs_lsn_t
2584 xlog_get_lowest_lsn(
2585 	struct xlog		*log)
2586 {
2587 	struct xlog_in_core	*iclog = log->l_iclog;
2588 	xfs_lsn_t		lowest_lsn = 0, lsn;
2589 
2590 	do {
2591 		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2592 		    iclog->ic_state == XLOG_STATE_DIRTY)
2593 			continue;
2594 
2595 		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2596 		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2597 			lowest_lsn = lsn;
2598 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2599 
2600 	return lowest_lsn;
2601 }
2602 
2603 /*
2604  * Completion of a iclog IO does not imply that a transaction has completed, as
2605  * transactions can be large enough to span many iclogs. We cannot change the
2606  * tail of the log half way through a transaction as this may be the only
2607  * transaction in the log and moving the tail to point to the middle of it
2608  * will prevent recovery from finding the start of the transaction. Hence we
2609  * should only update the last_sync_lsn if this iclog contains transaction
2610  * completion callbacks on it.
2611  *
2612  * We have to do this before we drop the icloglock to ensure we are the only one
2613  * that can update it.
2614  *
2615  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2616  * the reservation grant head pushing. This is due to the fact that the push
2617  * target is bound by the current last_sync_lsn value. Hence if we have a large
2618  * amount of log space bound up in this committing transaction then the
2619  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2620  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2621  * should push the AIL to ensure the push target (and hence the grant head) is
2622  * no longer bound by the old log head location and can move forwards and make
2623  * progress again.
2624  */
2625 static void
2626 xlog_state_set_callback(
2627 	struct xlog		*log,
2628 	struct xlog_in_core	*iclog,
2629 	xfs_lsn_t		header_lsn)
2630 {
2631 	iclog->ic_state = XLOG_STATE_CALLBACK;
2632 
2633 	ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2634 			   header_lsn) <= 0);
2635 
2636 	if (list_empty_careful(&iclog->ic_callbacks))
2637 		return;
2638 
2639 	atomic64_set(&log->l_last_sync_lsn, header_lsn);
2640 	xlog_grant_push_ail(log, 0);
2641 }
2642 
2643 /*
2644  * Return true if we need to stop processing, false to continue to the next
2645  * iclog. The caller will need to run callbacks if the iclog is returned in the
2646  * XLOG_STATE_CALLBACK state.
2647  */
2648 static bool
2649 xlog_state_iodone_process_iclog(
2650 	struct xlog		*log,
2651 	struct xlog_in_core	*iclog,
2652 	bool			*ioerror)
2653 {
2654 	xfs_lsn_t		lowest_lsn;
2655 	xfs_lsn_t		header_lsn;
2656 
2657 	switch (iclog->ic_state) {
2658 	case XLOG_STATE_ACTIVE:
2659 	case XLOG_STATE_DIRTY:
2660 		/*
2661 		 * Skip all iclogs in the ACTIVE & DIRTY states:
2662 		 */
2663 		return false;
2664 	case XLOG_STATE_IOERROR:
2665 		/*
2666 		 * Between marking a filesystem SHUTDOWN and stopping the log,
2667 		 * we do flush all iclogs to disk (if there wasn't a log I/O
2668 		 * error). So, we do want things to go smoothly in case of just
2669 		 * a SHUTDOWN w/o a LOG_IO_ERROR.
2670 		 */
2671 		*ioerror = true;
2672 		return false;
2673 	case XLOG_STATE_DONE_SYNC:
2674 		/*
2675 		 * Now that we have an iclog that is in the DONE_SYNC state, do
2676 		 * one more check here to see if we have chased our tail around.
2677 		 * If this is not the lowest lsn iclog, then we will leave it
2678 		 * for another completion to process.
2679 		 */
2680 		header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2681 		lowest_lsn = xlog_get_lowest_lsn(log);
2682 		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2683 			return false;
2684 		xlog_state_set_callback(log, iclog, header_lsn);
2685 		return false;
2686 	default:
2687 		/*
2688 		 * Can only perform callbacks in order.  Since this iclog is not
2689 		 * in the DONE_SYNC state, we skip the rest and just try to
2690 		 * clean up.
2691 		 */
2692 		return true;
2693 	}
2694 }
2695 
2696 /*
2697  * Keep processing entries in the iclog callback list until we come around and
2698  * it is empty.  We need to atomically see that the list is empty and change the
2699  * state to DIRTY so that we don't miss any more callbacks being added.
2700  *
2701  * This function is called with the icloglock held and returns with it held. We
2702  * drop it while running callbacks, however, as holding it over thousands of
2703  * callbacks is unnecessary and causes excessive contention if we do.
2704  */
2705 static void
2706 xlog_state_do_iclog_callbacks(
2707 	struct xlog		*log,
2708 	struct xlog_in_core	*iclog)
2709 		__releases(&log->l_icloglock)
2710 		__acquires(&log->l_icloglock)
2711 {
2712 	spin_unlock(&log->l_icloglock);
2713 	spin_lock(&iclog->ic_callback_lock);
2714 	while (!list_empty(&iclog->ic_callbacks)) {
2715 		LIST_HEAD(tmp);
2716 
2717 		list_splice_init(&iclog->ic_callbacks, &tmp);
2718 
2719 		spin_unlock(&iclog->ic_callback_lock);
2720 		xlog_cil_process_committed(&tmp);
2721 		spin_lock(&iclog->ic_callback_lock);
2722 	}
2723 
2724 	/*
2725 	 * Pick up the icloglock while still holding the callback lock so we
2726 	 * serialise against anyone trying to add more callbacks to this iclog
2727 	 * now we've finished processing.
2728 	 */
2729 	spin_lock(&log->l_icloglock);
2730 	spin_unlock(&iclog->ic_callback_lock);
2731 }
2732 
2733 STATIC void
2734 xlog_state_do_callback(
2735 	struct xlog		*log)
2736 {
2737 	struct xlog_in_core	*iclog;
2738 	struct xlog_in_core	*first_iclog;
2739 	bool			cycled_icloglock;
2740 	bool			ioerror;
2741 	int			flushcnt = 0;
2742 	int			repeats = 0;
2743 
2744 	spin_lock(&log->l_icloglock);
2745 	do {
2746 		/*
2747 		 * Scan all iclogs starting with the one pointed to by the
2748 		 * log.  Reset this starting point each time the log is
2749 		 * unlocked (during callbacks).
2750 		 *
2751 		 * Keep looping through iclogs until one full pass is made
2752 		 * without running any callbacks.
2753 		 */
2754 		first_iclog = log->l_iclog;
2755 		iclog = log->l_iclog;
2756 		cycled_icloglock = false;
2757 		ioerror = false;
2758 		repeats++;
2759 
2760 		do {
2761 			if (xlog_state_iodone_process_iclog(log, iclog,
2762 							&ioerror))
2763 				break;
2764 
2765 			if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2766 			    iclog->ic_state != XLOG_STATE_IOERROR) {
2767 				iclog = iclog->ic_next;
2768 				continue;
2769 			}
2770 
2771 			/*
2772 			 * Running callbacks will drop the icloglock which means
2773 			 * we'll have to run at least one more complete loop.
2774 			 */
2775 			cycled_icloglock = true;
2776 			xlog_state_do_iclog_callbacks(log, iclog);
2777 			if (XLOG_FORCED_SHUTDOWN(log))
2778 				wake_up_all(&iclog->ic_force_wait);
2779 			else
2780 				xlog_state_clean_iclog(log, iclog);
2781 			iclog = iclog->ic_next;
2782 		} while (first_iclog != iclog);
2783 
2784 		if (repeats > 5000) {
2785 			flushcnt += repeats;
2786 			repeats = 0;
2787 			xfs_warn(log->l_mp,
2788 				"%s: possible infinite loop (%d iterations)",
2789 				__func__, flushcnt);
2790 		}
2791 	} while (!ioerror && cycled_icloglock);
2792 
2793 	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2794 	    log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2795 		wake_up_all(&log->l_flush_wait);
2796 
2797 	spin_unlock(&log->l_icloglock);
2798 }
2799 
2800 
2801 /*
2802  * Finish transitioning this iclog to the dirty state.
2803  *
2804  * Make sure that we completely execute this routine only when this is
2805  * the last call to the iclog.  There is a good chance that iclog flushes,
2806  * when we reach the end of the physical log, get turned into 2 separate
2807  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2808  * routine.  By using the reference count bwritecnt, we guarantee that only
2809  * the second completion goes through.
2810  *
2811  * Callbacks could take time, so they are done outside the scope of the
2812  * global state machine log lock.
2813  */
2814 STATIC void
2815 xlog_state_done_syncing(
2816 	struct xlog_in_core	*iclog)
2817 {
2818 	struct xlog		*log = iclog->ic_log;
2819 
2820 	spin_lock(&log->l_icloglock);
2821 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2822 
2823 	/*
2824 	 * If we got an error, either on the first buffer, or in the case of
2825 	 * split log writes, on the second, we shut down the file system and
2826 	 * no iclogs should ever be attempted to be written to disk again.
2827 	 */
2828 	if (!XLOG_FORCED_SHUTDOWN(log)) {
2829 		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2830 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2831 	}
2832 
2833 	/*
2834 	 * Someone could be sleeping prior to writing out the next
2835 	 * iclog buffer, we wake them all, one will get to do the
2836 	 * I/O, the others get to wait for the result.
2837 	 */
2838 	wake_up_all(&iclog->ic_write_wait);
2839 	spin_unlock(&log->l_icloglock);
2840 	xlog_state_do_callback(log);
2841 }
2842 
2843 /*
2844  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2845  * sleep.  We wait on the flush queue on the head iclog as that should be
2846  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2847  * we will wait here and all new writes will sleep until a sync completes.
2848  *
2849  * The in-core logs are used in a circular fashion. They are not used
2850  * out-of-order even when an iclog past the head is free.
2851  *
2852  * return:
2853  *	* log_offset where xlog_write() can start writing into the in-core
2854  *		log's data space.
2855  *	* in-core log pointer to which xlog_write() should write.
2856  *	* boolean indicating this is a continued write to an in-core log.
2857  *		If this is the last write, then the in-core log's offset field
2858  *		needs to be incremented, depending on the amount of data which
2859  *		is copied.
2860  */
2861 STATIC int
2862 xlog_state_get_iclog_space(
2863 	struct xlog		*log,
2864 	int			len,
2865 	struct xlog_in_core	**iclogp,
2866 	struct xlog_ticket	*ticket,
2867 	int			*continued_write,
2868 	int			*logoffsetp)
2869 {
2870 	int		  log_offset;
2871 	xlog_rec_header_t *head;
2872 	xlog_in_core_t	  *iclog;
2873 
2874 restart:
2875 	spin_lock(&log->l_icloglock);
2876 	if (XLOG_FORCED_SHUTDOWN(log)) {
2877 		spin_unlock(&log->l_icloglock);
2878 		return -EIO;
2879 	}
2880 
2881 	iclog = log->l_iclog;
2882 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2883 		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2884 
2885 		/* Wait for log writes to have flushed */
2886 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2887 		goto restart;
2888 	}
2889 
2890 	head = &iclog->ic_header;
2891 
2892 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2893 	log_offset = iclog->ic_offset;
2894 
2895 	/* On the 1st write to an iclog, figure out lsn.  This works
2896 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2897 	 * committing to.  If the offset is set, that's how many blocks
2898 	 * must be written.
2899 	 */
2900 	if (log_offset == 0) {
2901 		ticket->t_curr_res -= log->l_iclog_hsize;
2902 		xlog_tic_add_region(ticket,
2903 				    log->l_iclog_hsize,
2904 				    XLOG_REG_TYPE_LRHEADER);
2905 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2906 		head->h_lsn = cpu_to_be64(
2907 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2908 		ASSERT(log->l_curr_block >= 0);
2909 	}
2910 
2911 	/* If there is enough room to write everything, then do it.  Otherwise,
2912 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2913 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2914 	 * until you know exactly how many bytes get copied.  Therefore, wait
2915 	 * until later to update ic_offset.
2916 	 *
2917 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2918 	 * can fit into remaining data section.
2919 	 */
2920 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2921 		int		error = 0;
2922 
2923 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2924 
2925 		/*
2926 		 * If we are the only one writing to this iclog, sync it to
2927 		 * disk.  We need to do an atomic compare and decrement here to
2928 		 * avoid racing with concurrent atomic_dec_and_lock() calls in
2929 		 * xlog_state_release_iclog() when there is more than one
2930 		 * reference to the iclog.
2931 		 */
2932 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2933 			error = xlog_state_release_iclog(log, iclog);
2934 		spin_unlock(&log->l_icloglock);
2935 		if (error)
2936 			return error;
2937 		goto restart;
2938 	}
2939 
2940 	/* Do we have enough room to write the full amount in the remainder
2941 	 * of this iclog?  Or must we continue a write on the next iclog and
2942 	 * mark this iclog as completely taken?  In the case where we switch
2943 	 * iclogs (to mark it taken), this particular iclog will release/sync
2944 	 * to disk in xlog_write().
2945 	 */
2946 	if (len <= iclog->ic_size - iclog->ic_offset) {
2947 		*continued_write = 0;
2948 		iclog->ic_offset += len;
2949 	} else {
2950 		*continued_write = 1;
2951 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2952 	}
2953 	*iclogp = iclog;
2954 
2955 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2956 	spin_unlock(&log->l_icloglock);
2957 
2958 	*logoffsetp = log_offset;
2959 	return 0;
2960 }
2961 
2962 /*
2963  * The first cnt-1 times a ticket goes through here we don't need to move the
2964  * grant write head because the permanent reservation has reserved cnt times the
2965  * unit amount.  Release part of current permanent unit reservation and reset
2966  * current reservation to be one units worth.  Also move grant reservation head
2967  * forward.
2968  */
2969 void
2970 xfs_log_ticket_regrant(
2971 	struct xlog		*log,
2972 	struct xlog_ticket	*ticket)
2973 {
2974 	trace_xfs_log_ticket_regrant(log, ticket);
2975 
2976 	if (ticket->t_cnt > 0)
2977 		ticket->t_cnt--;
2978 
2979 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2980 					ticket->t_curr_res);
2981 	xlog_grant_sub_space(log, &log->l_write_head.grant,
2982 					ticket->t_curr_res);
2983 	ticket->t_curr_res = ticket->t_unit_res;
2984 	xlog_tic_reset_res(ticket);
2985 
2986 	trace_xfs_log_ticket_regrant_sub(log, ticket);
2987 
2988 	/* just return if we still have some of the pre-reserved space */
2989 	if (!ticket->t_cnt) {
2990 		xlog_grant_add_space(log, &log->l_reserve_head.grant,
2991 				     ticket->t_unit_res);
2992 		trace_xfs_log_ticket_regrant_exit(log, ticket);
2993 
2994 		ticket->t_curr_res = ticket->t_unit_res;
2995 		xlog_tic_reset_res(ticket);
2996 	}
2997 
2998 	xfs_log_ticket_put(ticket);
2999 }
3000 
3001 /*
3002  * Give back the space left from a reservation.
3003  *
3004  * All the information we need to make a correct determination of space left
3005  * is present.  For non-permanent reservations, things are quite easy.  The
3006  * count should have been decremented to zero.  We only need to deal with the
3007  * space remaining in the current reservation part of the ticket.  If the
3008  * ticket contains a permanent reservation, there may be left over space which
3009  * needs to be released.  A count of N means that N-1 refills of the current
3010  * reservation can be done before we need to ask for more space.  The first
3011  * one goes to fill up the first current reservation.  Once we run out of
3012  * space, the count will stay at zero and the only space remaining will be
3013  * in the current reservation field.
3014  */
3015 void
3016 xfs_log_ticket_ungrant(
3017 	struct xlog		*log,
3018 	struct xlog_ticket	*ticket)
3019 {
3020 	int			bytes;
3021 
3022 	trace_xfs_log_ticket_ungrant(log, ticket);
3023 
3024 	if (ticket->t_cnt > 0)
3025 		ticket->t_cnt--;
3026 
3027 	trace_xfs_log_ticket_ungrant_sub(log, ticket);
3028 
3029 	/*
3030 	 * If this is a permanent reservation ticket, we may be able to free
3031 	 * up more space based on the remaining count.
3032 	 */
3033 	bytes = ticket->t_curr_res;
3034 	if (ticket->t_cnt > 0) {
3035 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3036 		bytes += ticket->t_unit_res*ticket->t_cnt;
3037 	}
3038 
3039 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3040 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3041 
3042 	trace_xfs_log_ticket_ungrant_exit(log, ticket);
3043 
3044 	xfs_log_space_wake(log->l_mp);
3045 	xfs_log_ticket_put(ticket);
3046 }
3047 
3048 /*
3049  * This routine will mark the current iclog in the ring as WANT_SYNC and move
3050  * the current iclog pointer to the next iclog in the ring.
3051  */
3052 STATIC void
3053 xlog_state_switch_iclogs(
3054 	struct xlog		*log,
3055 	struct xlog_in_core	*iclog,
3056 	int			eventual_size)
3057 {
3058 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3059 	assert_spin_locked(&log->l_icloglock);
3060 
3061 	if (!eventual_size)
3062 		eventual_size = iclog->ic_offset;
3063 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3064 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3065 	log->l_prev_block = log->l_curr_block;
3066 	log->l_prev_cycle = log->l_curr_cycle;
3067 
3068 	/* roll log?: ic_offset changed later */
3069 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3070 
3071 	/* Round up to next log-sunit */
3072 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3073 	    log->l_mp->m_sb.sb_logsunit > 1) {
3074 		uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3075 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3076 	}
3077 
3078 	if (log->l_curr_block >= log->l_logBBsize) {
3079 		/*
3080 		 * Rewind the current block before the cycle is bumped to make
3081 		 * sure that the combined LSN never transiently moves forward
3082 		 * when the log wraps to the next cycle. This is to support the
3083 		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3084 		 * other cases should acquire l_icloglock.
3085 		 */
3086 		log->l_curr_block -= log->l_logBBsize;
3087 		ASSERT(log->l_curr_block >= 0);
3088 		smp_wmb();
3089 		log->l_curr_cycle++;
3090 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3091 			log->l_curr_cycle++;
3092 	}
3093 	ASSERT(iclog == log->l_iclog);
3094 	log->l_iclog = iclog->ic_next;
3095 }
3096 
3097 /*
3098  * Write out all data in the in-core log as of this exact moment in time.
3099  *
3100  * Data may be written to the in-core log during this call.  However,
3101  * we don't guarantee this data will be written out.  A change from past
3102  * implementation means this routine will *not* write out zero length LRs.
3103  *
3104  * Basically, we try and perform an intelligent scan of the in-core logs.
3105  * If we determine there is no flushable data, we just return.  There is no
3106  * flushable data if:
3107  *
3108  *	1. the current iclog is active and has no data; the previous iclog
3109  *		is in the active or dirty state.
3110  *	2. the current iclog is drity, and the previous iclog is in the
3111  *		active or dirty state.
3112  *
3113  * We may sleep if:
3114  *
3115  *	1. the current iclog is not in the active nor dirty state.
3116  *	2. the current iclog dirty, and the previous iclog is not in the
3117  *		active nor dirty state.
3118  *	3. the current iclog is active, and there is another thread writing
3119  *		to this particular iclog.
3120  *	4. a) the current iclog is active and has no other writers
3121  *	   b) when we return from flushing out this iclog, it is still
3122  *		not in the active nor dirty state.
3123  */
3124 int
3125 xfs_log_force(
3126 	struct xfs_mount	*mp,
3127 	uint			flags)
3128 {
3129 	struct xlog		*log = mp->m_log;
3130 	struct xlog_in_core	*iclog;
3131 	xfs_lsn_t		lsn;
3132 
3133 	XFS_STATS_INC(mp, xs_log_force);
3134 	trace_xfs_log_force(mp, 0, _RET_IP_);
3135 
3136 	xlog_cil_force(log);
3137 
3138 	spin_lock(&log->l_icloglock);
3139 	iclog = log->l_iclog;
3140 	if (iclog->ic_state == XLOG_STATE_IOERROR)
3141 		goto out_error;
3142 
3143 	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3144 	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3145 	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3146 		/*
3147 		 * If the head is dirty or (active and empty), then we need to
3148 		 * look at the previous iclog.
3149 		 *
3150 		 * If the previous iclog is active or dirty we are done.  There
3151 		 * is nothing to sync out. Otherwise, we attach ourselves to the
3152 		 * previous iclog and go to sleep.
3153 		 */
3154 		iclog = iclog->ic_prev;
3155 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3156 		if (atomic_read(&iclog->ic_refcnt) == 0) {
3157 			/*
3158 			 * We are the only one with access to this iclog.
3159 			 *
3160 			 * Flush it out now.  There should be a roundoff of zero
3161 			 * to show that someone has already taken care of the
3162 			 * roundoff from the previous sync.
3163 			 */
3164 			atomic_inc(&iclog->ic_refcnt);
3165 			lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3166 			xlog_state_switch_iclogs(log, iclog, 0);
3167 			if (xlog_state_release_iclog(log, iclog))
3168 				goto out_error;
3169 
3170 			if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3171 				goto out_unlock;
3172 		} else {
3173 			/*
3174 			 * Someone else is writing to this iclog.
3175 			 *
3176 			 * Use its call to flush out the data.  However, the
3177 			 * other thread may not force out this LR, so we mark
3178 			 * it WANT_SYNC.
3179 			 */
3180 			xlog_state_switch_iclogs(log, iclog, 0);
3181 		}
3182 	} else {
3183 		/*
3184 		 * If the head iclog is not active nor dirty, we just attach
3185 		 * ourselves to the head and go to sleep if necessary.
3186 		 */
3187 		;
3188 	}
3189 
3190 	if (flags & XFS_LOG_SYNC)
3191 		return xlog_wait_on_iclog(iclog);
3192 out_unlock:
3193 	spin_unlock(&log->l_icloglock);
3194 	return 0;
3195 out_error:
3196 	spin_unlock(&log->l_icloglock);
3197 	return -EIO;
3198 }
3199 
3200 static int
3201 __xfs_log_force_lsn(
3202 	struct xfs_mount	*mp,
3203 	xfs_lsn_t		lsn,
3204 	uint			flags,
3205 	int			*log_flushed,
3206 	bool			already_slept)
3207 {
3208 	struct xlog		*log = mp->m_log;
3209 	struct xlog_in_core	*iclog;
3210 
3211 	spin_lock(&log->l_icloglock);
3212 	iclog = log->l_iclog;
3213 	if (iclog->ic_state == XLOG_STATE_IOERROR)
3214 		goto out_error;
3215 
3216 	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3217 		iclog = iclog->ic_next;
3218 		if (iclog == log->l_iclog)
3219 			goto out_unlock;
3220 	}
3221 
3222 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3223 		/*
3224 		 * We sleep here if we haven't already slept (e.g. this is the
3225 		 * first time we've looked at the correct iclog buf) and the
3226 		 * buffer before us is going to be sync'ed.  The reason for this
3227 		 * is that if we are doing sync transactions here, by waiting
3228 		 * for the previous I/O to complete, we can allow a few more
3229 		 * transactions into this iclog before we close it down.
3230 		 *
3231 		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3232 		 * refcnt so we can release the log (which drops the ref count).
3233 		 * The state switch keeps new transaction commits from using
3234 		 * this buffer.  When the current commits finish writing into
3235 		 * the buffer, the refcount will drop to zero and the buffer
3236 		 * will go out then.
3237 		 */
3238 		if (!already_slept &&
3239 		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3240 		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3241 			XFS_STATS_INC(mp, xs_log_force_sleep);
3242 
3243 			xlog_wait(&iclog->ic_prev->ic_write_wait,
3244 					&log->l_icloglock);
3245 			return -EAGAIN;
3246 		}
3247 		atomic_inc(&iclog->ic_refcnt);
3248 		xlog_state_switch_iclogs(log, iclog, 0);
3249 		if (xlog_state_release_iclog(log, iclog))
3250 			goto out_error;
3251 		if (log_flushed)
3252 			*log_flushed = 1;
3253 	}
3254 
3255 	if (flags & XFS_LOG_SYNC)
3256 		return xlog_wait_on_iclog(iclog);
3257 out_unlock:
3258 	spin_unlock(&log->l_icloglock);
3259 	return 0;
3260 out_error:
3261 	spin_unlock(&log->l_icloglock);
3262 	return -EIO;
3263 }
3264 
3265 /*
3266  * Force the in-core log to disk for a specific LSN.
3267  *
3268  * Find in-core log with lsn.
3269  *	If it is in the DIRTY state, just return.
3270  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3271  *		state and go to sleep or return.
3272  *	If it is in any other state, go to sleep or return.
3273  *
3274  * Synchronous forces are implemented with a wait queue.  All callers trying
3275  * to force a given lsn to disk must wait on the queue attached to the
3276  * specific in-core log.  When given in-core log finally completes its write
3277  * to disk, that thread will wake up all threads waiting on the queue.
3278  */
3279 int
3280 xfs_log_force_lsn(
3281 	struct xfs_mount	*mp,
3282 	xfs_lsn_t		lsn,
3283 	uint			flags,
3284 	int			*log_flushed)
3285 {
3286 	int			ret;
3287 	ASSERT(lsn != 0);
3288 
3289 	XFS_STATS_INC(mp, xs_log_force);
3290 	trace_xfs_log_force(mp, lsn, _RET_IP_);
3291 
3292 	lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3293 	if (lsn == NULLCOMMITLSN)
3294 		return 0;
3295 
3296 	ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3297 	if (ret == -EAGAIN)
3298 		ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3299 	return ret;
3300 }
3301 
3302 /*
3303  * Free a used ticket when its refcount falls to zero.
3304  */
3305 void
3306 xfs_log_ticket_put(
3307 	xlog_ticket_t	*ticket)
3308 {
3309 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3310 	if (atomic_dec_and_test(&ticket->t_ref))
3311 		kmem_cache_free(xfs_log_ticket_zone, ticket);
3312 }
3313 
3314 xlog_ticket_t *
3315 xfs_log_ticket_get(
3316 	xlog_ticket_t	*ticket)
3317 {
3318 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3319 	atomic_inc(&ticket->t_ref);
3320 	return ticket;
3321 }
3322 
3323 /*
3324  * Figure out the total log space unit (in bytes) that would be
3325  * required for a log ticket.
3326  */
3327 int
3328 xfs_log_calc_unit_res(
3329 	struct xfs_mount	*mp,
3330 	int			unit_bytes)
3331 {
3332 	struct xlog		*log = mp->m_log;
3333 	int			iclog_space;
3334 	uint			num_headers;
3335 
3336 	/*
3337 	 * Permanent reservations have up to 'cnt'-1 active log operations
3338 	 * in the log.  A unit in this case is the amount of space for one
3339 	 * of these log operations.  Normal reservations have a cnt of 1
3340 	 * and their unit amount is the total amount of space required.
3341 	 *
3342 	 * The following lines of code account for non-transaction data
3343 	 * which occupy space in the on-disk log.
3344 	 *
3345 	 * Normal form of a transaction is:
3346 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3347 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3348 	 *
3349 	 * We need to account for all the leadup data and trailer data
3350 	 * around the transaction data.
3351 	 * And then we need to account for the worst case in terms of using
3352 	 * more space.
3353 	 * The worst case will happen if:
3354 	 * - the placement of the transaction happens to be such that the
3355 	 *   roundoff is at its maximum
3356 	 * - the transaction data is synced before the commit record is synced
3357 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3358 	 *   Therefore the commit record is in its own Log Record.
3359 	 *   This can happen as the commit record is called with its
3360 	 *   own region to xlog_write().
3361 	 *   This then means that in the worst case, roundoff can happen for
3362 	 *   the commit-rec as well.
3363 	 *   The commit-rec is smaller than padding in this scenario and so it is
3364 	 *   not added separately.
3365 	 */
3366 
3367 	/* for trans header */
3368 	unit_bytes += sizeof(xlog_op_header_t);
3369 	unit_bytes += sizeof(xfs_trans_header_t);
3370 
3371 	/* for start-rec */
3372 	unit_bytes += sizeof(xlog_op_header_t);
3373 
3374 	/*
3375 	 * for LR headers - the space for data in an iclog is the size minus
3376 	 * the space used for the headers. If we use the iclog size, then we
3377 	 * undercalculate the number of headers required.
3378 	 *
3379 	 * Furthermore - the addition of op headers for split-recs might
3380 	 * increase the space required enough to require more log and op
3381 	 * headers, so take that into account too.
3382 	 *
3383 	 * IMPORTANT: This reservation makes the assumption that if this
3384 	 * transaction is the first in an iclog and hence has the LR headers
3385 	 * accounted to it, then the remaining space in the iclog is
3386 	 * exclusively for this transaction.  i.e. if the transaction is larger
3387 	 * than the iclog, it will be the only thing in that iclog.
3388 	 * Fundamentally, this means we must pass the entire log vector to
3389 	 * xlog_write to guarantee this.
3390 	 */
3391 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3392 	num_headers = howmany(unit_bytes, iclog_space);
3393 
3394 	/* for split-recs - ophdrs added when data split over LRs */
3395 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3396 
3397 	/* add extra header reservations if we overrun */
3398 	while (!num_headers ||
3399 	       howmany(unit_bytes, iclog_space) > num_headers) {
3400 		unit_bytes += sizeof(xlog_op_header_t);
3401 		num_headers++;
3402 	}
3403 	unit_bytes += log->l_iclog_hsize * num_headers;
3404 
3405 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3406 	unit_bytes += log->l_iclog_hsize;
3407 
3408 	/* for roundoff padding for transaction data and one for commit record */
3409 	if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3410 		/* log su roundoff */
3411 		unit_bytes += 2 * mp->m_sb.sb_logsunit;
3412 	} else {
3413 		/* BB roundoff */
3414 		unit_bytes += 2 * BBSIZE;
3415         }
3416 
3417 	return unit_bytes;
3418 }
3419 
3420 /*
3421  * Allocate and initialise a new log ticket.
3422  */
3423 struct xlog_ticket *
3424 xlog_ticket_alloc(
3425 	struct xlog		*log,
3426 	int			unit_bytes,
3427 	int			cnt,
3428 	char			client,
3429 	bool			permanent)
3430 {
3431 	struct xlog_ticket	*tic;
3432 	int			unit_res;
3433 
3434 	tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3435 
3436 	unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3437 
3438 	atomic_set(&tic->t_ref, 1);
3439 	tic->t_task		= current;
3440 	INIT_LIST_HEAD(&tic->t_queue);
3441 	tic->t_unit_res		= unit_res;
3442 	tic->t_curr_res		= unit_res;
3443 	tic->t_cnt		= cnt;
3444 	tic->t_ocnt		= cnt;
3445 	tic->t_tid		= prandom_u32();
3446 	tic->t_clientid		= client;
3447 	if (permanent)
3448 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3449 
3450 	xlog_tic_reset_res(tic);
3451 
3452 	return tic;
3453 }
3454 
3455 #if defined(DEBUG)
3456 /*
3457  * Make sure that the destination ptr is within the valid data region of
3458  * one of the iclogs.  This uses backup pointers stored in a different
3459  * part of the log in case we trash the log structure.
3460  */
3461 STATIC void
3462 xlog_verify_dest_ptr(
3463 	struct xlog	*log,
3464 	void		*ptr)
3465 {
3466 	int i;
3467 	int good_ptr = 0;
3468 
3469 	for (i = 0; i < log->l_iclog_bufs; i++) {
3470 		if (ptr >= log->l_iclog_bak[i] &&
3471 		    ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3472 			good_ptr++;
3473 	}
3474 
3475 	if (!good_ptr)
3476 		xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3477 }
3478 
3479 /*
3480  * Check to make sure the grant write head didn't just over lap the tail.  If
3481  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3482  * the cycles differ by exactly one and check the byte count.
3483  *
3484  * This check is run unlocked, so can give false positives. Rather than assert
3485  * on failures, use a warn-once flag and a panic tag to allow the admin to
3486  * determine if they want to panic the machine when such an error occurs. For
3487  * debug kernels this will have the same effect as using an assert but, unlinke
3488  * an assert, it can be turned off at runtime.
3489  */
3490 STATIC void
3491 xlog_verify_grant_tail(
3492 	struct xlog	*log)
3493 {
3494 	int		tail_cycle, tail_blocks;
3495 	int		cycle, space;
3496 
3497 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3498 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3499 	if (tail_cycle != cycle) {
3500 		if (cycle - 1 != tail_cycle &&
3501 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3502 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3503 				"%s: cycle - 1 != tail_cycle", __func__);
3504 			log->l_flags |= XLOG_TAIL_WARN;
3505 		}
3506 
3507 		if (space > BBTOB(tail_blocks) &&
3508 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3509 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3510 				"%s: space > BBTOB(tail_blocks)", __func__);
3511 			log->l_flags |= XLOG_TAIL_WARN;
3512 		}
3513 	}
3514 }
3515 
3516 /* check if it will fit */
3517 STATIC void
3518 xlog_verify_tail_lsn(
3519 	struct xlog		*log,
3520 	struct xlog_in_core	*iclog,
3521 	xfs_lsn_t		tail_lsn)
3522 {
3523     int blocks;
3524 
3525     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3526 	blocks =
3527 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3528 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3529 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3530     } else {
3531 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3532 
3533 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3534 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3535 
3536 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3537 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3538 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3539     }
3540 }
3541 
3542 /*
3543  * Perform a number of checks on the iclog before writing to disk.
3544  *
3545  * 1. Make sure the iclogs are still circular
3546  * 2. Make sure we have a good magic number
3547  * 3. Make sure we don't have magic numbers in the data
3548  * 4. Check fields of each log operation header for:
3549  *	A. Valid client identifier
3550  *	B. tid ptr value falls in valid ptr space (user space code)
3551  *	C. Length in log record header is correct according to the
3552  *		individual operation headers within record.
3553  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3554  *	log, check the preceding blocks of the physical log to make sure all
3555  *	the cycle numbers agree with the current cycle number.
3556  */
3557 STATIC void
3558 xlog_verify_iclog(
3559 	struct xlog		*log,
3560 	struct xlog_in_core	*iclog,
3561 	int			count)
3562 {
3563 	xlog_op_header_t	*ophead;
3564 	xlog_in_core_t		*icptr;
3565 	xlog_in_core_2_t	*xhdr;
3566 	void			*base_ptr, *ptr, *p;
3567 	ptrdiff_t		field_offset;
3568 	uint8_t			clientid;
3569 	int			len, i, j, k, op_len;
3570 	int			idx;
3571 
3572 	/* check validity of iclog pointers */
3573 	spin_lock(&log->l_icloglock);
3574 	icptr = log->l_iclog;
3575 	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3576 		ASSERT(icptr);
3577 
3578 	if (icptr != log->l_iclog)
3579 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3580 	spin_unlock(&log->l_icloglock);
3581 
3582 	/* check log magic numbers */
3583 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3584 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3585 
3586 	base_ptr = ptr = &iclog->ic_header;
3587 	p = &iclog->ic_header;
3588 	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3589 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3590 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3591 				__func__);
3592 	}
3593 
3594 	/* check fields */
3595 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3596 	base_ptr = ptr = iclog->ic_datap;
3597 	ophead = ptr;
3598 	xhdr = iclog->ic_data;
3599 	for (i = 0; i < len; i++) {
3600 		ophead = ptr;
3601 
3602 		/* clientid is only 1 byte */
3603 		p = &ophead->oh_clientid;
3604 		field_offset = p - base_ptr;
3605 		if (field_offset & 0x1ff) {
3606 			clientid = ophead->oh_clientid;
3607 		} else {
3608 			idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3609 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3610 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3611 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3612 				clientid = xlog_get_client_id(
3613 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3614 			} else {
3615 				clientid = xlog_get_client_id(
3616 					iclog->ic_header.h_cycle_data[idx]);
3617 			}
3618 		}
3619 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3620 			xfs_warn(log->l_mp,
3621 				"%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3622 				__func__, clientid, ophead,
3623 				(unsigned long)field_offset);
3624 
3625 		/* check length */
3626 		p = &ophead->oh_len;
3627 		field_offset = p - base_ptr;
3628 		if (field_offset & 0x1ff) {
3629 			op_len = be32_to_cpu(ophead->oh_len);
3630 		} else {
3631 			idx = BTOBBT((uintptr_t)&ophead->oh_len -
3632 				    (uintptr_t)iclog->ic_datap);
3633 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3634 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3635 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3636 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3637 			} else {
3638 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3639 			}
3640 		}
3641 		ptr += sizeof(xlog_op_header_t) + op_len;
3642 	}
3643 }
3644 #endif
3645 
3646 /*
3647  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3648  */
3649 STATIC int
3650 xlog_state_ioerror(
3651 	struct xlog	*log)
3652 {
3653 	xlog_in_core_t	*iclog, *ic;
3654 
3655 	iclog = log->l_iclog;
3656 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
3657 		/*
3658 		 * Mark all the incore logs IOERROR.
3659 		 * From now on, no log flushes will result.
3660 		 */
3661 		ic = iclog;
3662 		do {
3663 			ic->ic_state = XLOG_STATE_IOERROR;
3664 			ic = ic->ic_next;
3665 		} while (ic != iclog);
3666 		return 0;
3667 	}
3668 	/*
3669 	 * Return non-zero, if state transition has already happened.
3670 	 */
3671 	return 1;
3672 }
3673 
3674 /*
3675  * This is called from xfs_force_shutdown, when we're forcibly
3676  * shutting down the filesystem, typically because of an IO error.
3677  * Our main objectives here are to make sure that:
3678  *	a. if !logerror, flush the logs to disk. Anything modified
3679  *	   after this is ignored.
3680  *	b. the filesystem gets marked 'SHUTDOWN' for all interested
3681  *	   parties to find out, 'atomically'.
3682  *	c. those who're sleeping on log reservations, pinned objects and
3683  *	    other resources get woken up, and be told the bad news.
3684  *	d. nothing new gets queued up after (b) and (c) are done.
3685  *
3686  * Note: for the !logerror case we need to flush the regions held in memory out
3687  * to disk first. This needs to be done before the log is marked as shutdown,
3688  * otherwise the iclog writes will fail.
3689  */
3690 int
3691 xfs_log_force_umount(
3692 	struct xfs_mount	*mp,
3693 	int			logerror)
3694 {
3695 	struct xlog	*log;
3696 	int		retval;
3697 
3698 	log = mp->m_log;
3699 
3700 	/*
3701 	 * If this happens during log recovery, don't worry about
3702 	 * locking; the log isn't open for business yet.
3703 	 */
3704 	if (!log ||
3705 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3706 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3707 		if (mp->m_sb_bp)
3708 			mp->m_sb_bp->b_flags |= XBF_DONE;
3709 		return 0;
3710 	}
3711 
3712 	/*
3713 	 * Somebody could've already done the hard work for us.
3714 	 * No need to get locks for this.
3715 	 */
3716 	if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3717 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3718 		return 1;
3719 	}
3720 
3721 	/*
3722 	 * Flush all the completed transactions to disk before marking the log
3723 	 * being shut down. We need to do it in this order to ensure that
3724 	 * completed operations are safely on disk before we shut down, and that
3725 	 * we don't have to issue any buffer IO after the shutdown flags are set
3726 	 * to guarantee this.
3727 	 */
3728 	if (!logerror)
3729 		xfs_log_force(mp, XFS_LOG_SYNC);
3730 
3731 	/*
3732 	 * mark the filesystem and the as in a shutdown state and wake
3733 	 * everybody up to tell them the bad news.
3734 	 */
3735 	spin_lock(&log->l_icloglock);
3736 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3737 	if (mp->m_sb_bp)
3738 		mp->m_sb_bp->b_flags |= XBF_DONE;
3739 
3740 	/*
3741 	 * Mark the log and the iclogs with IO error flags to prevent any
3742 	 * further log IO from being issued or completed.
3743 	 */
3744 	log->l_flags |= XLOG_IO_ERROR;
3745 	retval = xlog_state_ioerror(log);
3746 	spin_unlock(&log->l_icloglock);
3747 
3748 	/*
3749 	 * We don't want anybody waiting for log reservations after this. That
3750 	 * means we have to wake up everybody queued up on reserveq as well as
3751 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3752 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3753 	 * action is protected by the grant locks.
3754 	 */
3755 	xlog_grant_head_wake_all(&log->l_reserve_head);
3756 	xlog_grant_head_wake_all(&log->l_write_head);
3757 
3758 	/*
3759 	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3760 	 * as if the log writes were completed. The abort handling in the log
3761 	 * item committed callback functions will do this again under lock to
3762 	 * avoid races.
3763 	 */
3764 	spin_lock(&log->l_cilp->xc_push_lock);
3765 	wake_up_all(&log->l_cilp->xc_commit_wait);
3766 	spin_unlock(&log->l_cilp->xc_push_lock);
3767 	xlog_state_do_callback(log);
3768 
3769 	/* return non-zero if log IOERROR transition had already happened */
3770 	return retval;
3771 }
3772 
3773 STATIC int
3774 xlog_iclogs_empty(
3775 	struct xlog	*log)
3776 {
3777 	xlog_in_core_t	*iclog;
3778 
3779 	iclog = log->l_iclog;
3780 	do {
3781 		/* endianness does not matter here, zero is zero in
3782 		 * any language.
3783 		 */
3784 		if (iclog->ic_header.h_num_logops)
3785 			return 0;
3786 		iclog = iclog->ic_next;
3787 	} while (iclog != log->l_iclog);
3788 	return 1;
3789 }
3790 
3791 /*
3792  * Verify that an LSN stamped into a piece of metadata is valid. This is
3793  * intended for use in read verifiers on v5 superblocks.
3794  */
3795 bool
3796 xfs_log_check_lsn(
3797 	struct xfs_mount	*mp,
3798 	xfs_lsn_t		lsn)
3799 {
3800 	struct xlog		*log = mp->m_log;
3801 	bool			valid;
3802 
3803 	/*
3804 	 * norecovery mode skips mount-time log processing and unconditionally
3805 	 * resets the in-core LSN. We can't validate in this mode, but
3806 	 * modifications are not allowed anyways so just return true.
3807 	 */
3808 	if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3809 		return true;
3810 
3811 	/*
3812 	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3813 	 * handled by recovery and thus safe to ignore here.
3814 	 */
3815 	if (lsn == NULLCOMMITLSN)
3816 		return true;
3817 
3818 	valid = xlog_valid_lsn(mp->m_log, lsn);
3819 
3820 	/* warn the user about what's gone wrong before verifier failure */
3821 	if (!valid) {
3822 		spin_lock(&log->l_icloglock);
3823 		xfs_warn(mp,
3824 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3825 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3826 			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3827 			 log->l_curr_cycle, log->l_curr_block);
3828 		spin_unlock(&log->l_icloglock);
3829 	}
3830 
3831 	return valid;
3832 }
3833 
3834 bool
3835 xfs_log_in_recovery(
3836 	struct xfs_mount	*mp)
3837 {
3838 	struct xlog		*log = mp->m_log;
3839 
3840 	return log->l_flags & XLOG_ACTIVE_RECOVERY;
3841 }
3842