xref: /linux/fs/xfs/xfs_log.c (revision b8321ed4a40c02054f930ca59d3570caa27bc86c)
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 struct kmem_cache	*xfs_log_ticket_cache;
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 void xlog_state_do_callback(
45 	struct xlog		*log);
46 STATIC int
47 xlog_state_get_iclog_space(
48 	struct xlog		*log,
49 	int			len,
50 	struct xlog_in_core	**iclog,
51 	struct xlog_ticket	*ticket,
52 	int			*continued_write,
53 	int			*logoffsetp);
54 STATIC void
55 xlog_grant_push_ail(
56 	struct xlog		*log,
57 	int			need_bytes);
58 STATIC void
59 xlog_sync(
60 	struct xlog		*log,
61 	struct xlog_in_core	*iclog);
62 #if defined(DEBUG)
63 STATIC void
64 xlog_verify_dest_ptr(
65 	struct xlog		*log,
66 	void			*ptr);
67 STATIC void
68 xlog_verify_grant_tail(
69 	struct xlog *log);
70 STATIC void
71 xlog_verify_iclog(
72 	struct xlog		*log,
73 	struct xlog_in_core	*iclog,
74 	int			count);
75 STATIC void
76 xlog_verify_tail_lsn(
77 	struct xlog		*log,
78 	struct xlog_in_core	*iclog);
79 #else
80 #define xlog_verify_dest_ptr(a,b)
81 #define xlog_verify_grant_tail(a)
82 #define xlog_verify_iclog(a,b,c)
83 #define xlog_verify_tail_lsn(a,b)
84 #endif
85 
86 STATIC int
87 xlog_iclogs_empty(
88 	struct xlog		*log);
89 
90 static int
91 xfs_log_cover(struct xfs_mount *);
92 
93 static void
94 xlog_grant_sub_space(
95 	struct xlog		*log,
96 	atomic64_t		*head,
97 	int			bytes)
98 {
99 	int64_t	head_val = atomic64_read(head);
100 	int64_t new, old;
101 
102 	do {
103 		int	cycle, space;
104 
105 		xlog_crack_grant_head_val(head_val, &cycle, &space);
106 
107 		space -= bytes;
108 		if (space < 0) {
109 			space += log->l_logsize;
110 			cycle--;
111 		}
112 
113 		old = head_val;
114 		new = xlog_assign_grant_head_val(cycle, space);
115 		head_val = atomic64_cmpxchg(head, old, new);
116 	} while (head_val != old);
117 }
118 
119 static void
120 xlog_grant_add_space(
121 	struct xlog		*log,
122 	atomic64_t		*head,
123 	int			bytes)
124 {
125 	int64_t	head_val = atomic64_read(head);
126 	int64_t new, old;
127 
128 	do {
129 		int		tmp;
130 		int		cycle, space;
131 
132 		xlog_crack_grant_head_val(head_val, &cycle, &space);
133 
134 		tmp = log->l_logsize - space;
135 		if (tmp > bytes)
136 			space += bytes;
137 		else {
138 			space = bytes - tmp;
139 			cycle++;
140 		}
141 
142 		old = head_val;
143 		new = xlog_assign_grant_head_val(cycle, space);
144 		head_val = atomic64_cmpxchg(head, old, new);
145 	} while (head_val != old);
146 }
147 
148 STATIC void
149 xlog_grant_head_init(
150 	struct xlog_grant_head	*head)
151 {
152 	xlog_assign_grant_head(&head->grant, 1, 0);
153 	INIT_LIST_HEAD(&head->waiters);
154 	spin_lock_init(&head->lock);
155 }
156 
157 STATIC void
158 xlog_grant_head_wake_all(
159 	struct xlog_grant_head	*head)
160 {
161 	struct xlog_ticket	*tic;
162 
163 	spin_lock(&head->lock);
164 	list_for_each_entry(tic, &head->waiters, t_queue)
165 		wake_up_process(tic->t_task);
166 	spin_unlock(&head->lock);
167 }
168 
169 static inline int
170 xlog_ticket_reservation(
171 	struct xlog		*log,
172 	struct xlog_grant_head	*head,
173 	struct xlog_ticket	*tic)
174 {
175 	if (head == &log->l_write_head) {
176 		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
177 		return tic->t_unit_res;
178 	} else {
179 		if (tic->t_flags & XLOG_TIC_PERM_RESERV)
180 			return tic->t_unit_res * tic->t_cnt;
181 		else
182 			return tic->t_unit_res;
183 	}
184 }
185 
186 STATIC bool
187 xlog_grant_head_wake(
188 	struct xlog		*log,
189 	struct xlog_grant_head	*head,
190 	int			*free_bytes)
191 {
192 	struct xlog_ticket	*tic;
193 	int			need_bytes;
194 	bool			woken_task = false;
195 
196 	list_for_each_entry(tic, &head->waiters, t_queue) {
197 
198 		/*
199 		 * There is a chance that the size of the CIL checkpoints in
200 		 * progress at the last AIL push target calculation resulted in
201 		 * limiting the target to the log head (l_last_sync_lsn) at the
202 		 * time. This may not reflect where the log head is now as the
203 		 * CIL checkpoints may have completed.
204 		 *
205 		 * Hence when we are woken here, it may be that the head of the
206 		 * log that has moved rather than the tail. As the tail didn't
207 		 * move, there still won't be space available for the
208 		 * reservation we require.  However, if the AIL has already
209 		 * pushed to the target defined by the old log head location, we
210 		 * will hang here waiting for something else to update the AIL
211 		 * push target.
212 		 *
213 		 * Therefore, if there isn't space to wake the first waiter on
214 		 * the grant head, we need to push the AIL again to ensure the
215 		 * target reflects both the current log tail and log head
216 		 * position before we wait for the tail to move again.
217 		 */
218 
219 		need_bytes = xlog_ticket_reservation(log, head, tic);
220 		if (*free_bytes < need_bytes) {
221 			if (!woken_task)
222 				xlog_grant_push_ail(log, need_bytes);
223 			return false;
224 		}
225 
226 		*free_bytes -= need_bytes;
227 		trace_xfs_log_grant_wake_up(log, tic);
228 		wake_up_process(tic->t_task);
229 		woken_task = true;
230 	}
231 
232 	return true;
233 }
234 
235 STATIC int
236 xlog_grant_head_wait(
237 	struct xlog		*log,
238 	struct xlog_grant_head	*head,
239 	struct xlog_ticket	*tic,
240 	int			need_bytes) __releases(&head->lock)
241 					    __acquires(&head->lock)
242 {
243 	list_add_tail(&tic->t_queue, &head->waiters);
244 
245 	do {
246 		if (xlog_is_shutdown(log))
247 			goto shutdown;
248 		xlog_grant_push_ail(log, need_bytes);
249 
250 		__set_current_state(TASK_UNINTERRUPTIBLE);
251 		spin_unlock(&head->lock);
252 
253 		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
254 
255 		trace_xfs_log_grant_sleep(log, tic);
256 		schedule();
257 		trace_xfs_log_grant_wake(log, tic);
258 
259 		spin_lock(&head->lock);
260 		if (xlog_is_shutdown(log))
261 			goto shutdown;
262 	} while (xlog_space_left(log, &head->grant) < need_bytes);
263 
264 	list_del_init(&tic->t_queue);
265 	return 0;
266 shutdown:
267 	list_del_init(&tic->t_queue);
268 	return -EIO;
269 }
270 
271 /*
272  * Atomically get the log space required for a log ticket.
273  *
274  * Once a ticket gets put onto head->waiters, it will only return after the
275  * needed reservation is satisfied.
276  *
277  * This function is structured so that it has a lock free fast path. This is
278  * necessary because every new transaction reservation will come through this
279  * path. Hence any lock will be globally hot if we take it unconditionally on
280  * every pass.
281  *
282  * As tickets are only ever moved on and off head->waiters under head->lock, we
283  * only need to take that lock if we are going to add the ticket to the queue
284  * and sleep. We can avoid taking the lock if the ticket was never added to
285  * head->waiters because the t_queue list head will be empty and we hold the
286  * only reference to it so it can safely be checked unlocked.
287  */
288 STATIC int
289 xlog_grant_head_check(
290 	struct xlog		*log,
291 	struct xlog_grant_head	*head,
292 	struct xlog_ticket	*tic,
293 	int			*need_bytes)
294 {
295 	int			free_bytes;
296 	int			error = 0;
297 
298 	ASSERT(!xlog_in_recovery(log));
299 
300 	/*
301 	 * If there are other waiters on the queue then give them a chance at
302 	 * logspace before us.  Wake up the first waiters, if we do not wake
303 	 * up all the waiters then go to sleep waiting for more free space,
304 	 * otherwise try to get some space for this transaction.
305 	 */
306 	*need_bytes = xlog_ticket_reservation(log, head, tic);
307 	free_bytes = xlog_space_left(log, &head->grant);
308 	if (!list_empty_careful(&head->waiters)) {
309 		spin_lock(&head->lock);
310 		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
311 		    free_bytes < *need_bytes) {
312 			error = xlog_grant_head_wait(log, head, tic,
313 						     *need_bytes);
314 		}
315 		spin_unlock(&head->lock);
316 	} else if (free_bytes < *need_bytes) {
317 		spin_lock(&head->lock);
318 		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
319 		spin_unlock(&head->lock);
320 	}
321 
322 	return error;
323 }
324 
325 static void
326 xlog_tic_reset_res(xlog_ticket_t *tic)
327 {
328 	tic->t_res_num = 0;
329 	tic->t_res_arr_sum = 0;
330 	tic->t_res_num_ophdrs = 0;
331 }
332 
333 static void
334 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
335 {
336 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
337 		/* add to overflow and start again */
338 		tic->t_res_o_flow += tic->t_res_arr_sum;
339 		tic->t_res_num = 0;
340 		tic->t_res_arr_sum = 0;
341 	}
342 
343 	tic->t_res_arr[tic->t_res_num].r_len = len;
344 	tic->t_res_arr[tic->t_res_num].r_type = type;
345 	tic->t_res_arr_sum += len;
346 	tic->t_res_num++;
347 }
348 
349 bool
350 xfs_log_writable(
351 	struct xfs_mount	*mp)
352 {
353 	/*
354 	 * Do not write to the log on norecovery mounts, if the data or log
355 	 * devices are read-only, or if the filesystem is shutdown. Read-only
356 	 * mounts allow internal writes for log recovery and unmount purposes,
357 	 * so don't restrict that case.
358 	 */
359 	if (xfs_has_norecovery(mp))
360 		return false;
361 	if (xfs_readonly_buftarg(mp->m_ddev_targp))
362 		return false;
363 	if (xfs_readonly_buftarg(mp->m_log->l_targ))
364 		return false;
365 	if (xlog_is_shutdown(mp->m_log))
366 		return false;
367 	return true;
368 }
369 
370 /*
371  * Replenish the byte reservation required by moving the grant write head.
372  */
373 int
374 xfs_log_regrant(
375 	struct xfs_mount	*mp,
376 	struct xlog_ticket	*tic)
377 {
378 	struct xlog		*log = mp->m_log;
379 	int			need_bytes;
380 	int			error = 0;
381 
382 	if (xlog_is_shutdown(log))
383 		return -EIO;
384 
385 	XFS_STATS_INC(mp, xs_try_logspace);
386 
387 	/*
388 	 * This is a new transaction on the ticket, so we need to change the
389 	 * transaction ID so that the next transaction has a different TID in
390 	 * the log. Just add one to the existing tid so that we can see chains
391 	 * of rolling transactions in the log easily.
392 	 */
393 	tic->t_tid++;
394 
395 	xlog_grant_push_ail(log, tic->t_unit_res);
396 
397 	tic->t_curr_res = tic->t_unit_res;
398 	xlog_tic_reset_res(tic);
399 
400 	if (tic->t_cnt > 0)
401 		return 0;
402 
403 	trace_xfs_log_regrant(log, tic);
404 
405 	error = xlog_grant_head_check(log, &log->l_write_head, tic,
406 				      &need_bytes);
407 	if (error)
408 		goto out_error;
409 
410 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
411 	trace_xfs_log_regrant_exit(log, tic);
412 	xlog_verify_grant_tail(log);
413 	return 0;
414 
415 out_error:
416 	/*
417 	 * If we are failing, make sure the ticket doesn't have any current
418 	 * reservations.  We don't want to add this back when the ticket/
419 	 * transaction gets cancelled.
420 	 */
421 	tic->t_curr_res = 0;
422 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
423 	return error;
424 }
425 
426 /*
427  * Reserve log space and return a ticket corresponding to the reservation.
428  *
429  * Each reservation is going to reserve extra space for a log record header.
430  * When writes happen to the on-disk log, we don't subtract the length of the
431  * log record header from any reservation.  By wasting space in each
432  * reservation, we prevent over allocation problems.
433  */
434 int
435 xfs_log_reserve(
436 	struct xfs_mount	*mp,
437 	int		 	unit_bytes,
438 	int		 	cnt,
439 	struct xlog_ticket	**ticp,
440 	uint8_t		 	client,
441 	bool			permanent)
442 {
443 	struct xlog		*log = mp->m_log;
444 	struct xlog_ticket	*tic;
445 	int			need_bytes;
446 	int			error = 0;
447 
448 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
449 
450 	if (xlog_is_shutdown(log))
451 		return -EIO;
452 
453 	XFS_STATS_INC(mp, xs_try_logspace);
454 
455 	ASSERT(*ticp == NULL);
456 	tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
457 	*ticp = tic;
458 
459 	xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
460 					    : tic->t_unit_res);
461 
462 	trace_xfs_log_reserve(log, tic);
463 
464 	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
465 				      &need_bytes);
466 	if (error)
467 		goto out_error;
468 
469 	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
470 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
471 	trace_xfs_log_reserve_exit(log, tic);
472 	xlog_verify_grant_tail(log);
473 	return 0;
474 
475 out_error:
476 	/*
477 	 * If we are failing, make sure the ticket doesn't have any current
478 	 * reservations.  We don't want to add this back when the ticket/
479 	 * transaction gets cancelled.
480 	 */
481 	tic->t_curr_res = 0;
482 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
483 	return error;
484 }
485 
486 /*
487  * Run all the pending iclog callbacks and wake log force waiters and iclog
488  * space waiters so they can process the newly set shutdown state. We really
489  * don't care what order we process callbacks here because the log is shut down
490  * and so state cannot change on disk anymore.
491  *
492  * We avoid processing actively referenced iclogs so that we don't run callbacks
493  * while the iclog owner might still be preparing the iclog for IO submssion.
494  * These will be caught by xlog_state_iclog_release() and call this function
495  * again to process any callbacks that may have been added to that iclog.
496  */
497 static void
498 xlog_state_shutdown_callbacks(
499 	struct xlog		*log)
500 {
501 	struct xlog_in_core	*iclog;
502 	LIST_HEAD(cb_list);
503 
504 	spin_lock(&log->l_icloglock);
505 	iclog = log->l_iclog;
506 	do {
507 		if (atomic_read(&iclog->ic_refcnt)) {
508 			/* Reference holder will re-run iclog callbacks. */
509 			continue;
510 		}
511 		list_splice_init(&iclog->ic_callbacks, &cb_list);
512 		wake_up_all(&iclog->ic_write_wait);
513 		wake_up_all(&iclog->ic_force_wait);
514 	} while ((iclog = iclog->ic_next) != log->l_iclog);
515 
516 	wake_up_all(&log->l_flush_wait);
517 	spin_unlock(&log->l_icloglock);
518 
519 	xlog_cil_process_committed(&cb_list);
520 }
521 
522 /*
523  * Flush iclog to disk if this is the last reference to the given iclog and the
524  * it is in the WANT_SYNC state.
525  *
526  * If the caller passes in a non-zero @old_tail_lsn and the current log tail
527  * does not match, there may be metadata on disk that must be persisted before
528  * this iclog is written.  To satisfy that requirement, set the
529  * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
530  * log tail value.
531  *
532  * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
533  * log tail is updated correctly. NEED_FUA indicates that the iclog will be
534  * written to stable storage, and implies that a commit record is contained
535  * within the iclog. We need to ensure that the log tail does not move beyond
536  * the tail that the first commit record in the iclog ordered against, otherwise
537  * correct recovery of that checkpoint becomes dependent on future operations
538  * performed on this iclog.
539  *
540  * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
541  * current tail into iclog. Once the iclog tail is set, future operations must
542  * not modify it, otherwise they potentially violate ordering constraints for
543  * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
544  * the iclog will get zeroed on activation of the iclog after sync, so we
545  * always capture the tail lsn on the iclog on the first NEED_FUA release
546  * regardless of the number of active reference counts on this iclog.
547  */
548 
549 int
550 xlog_state_release_iclog(
551 	struct xlog		*log,
552 	struct xlog_in_core	*iclog,
553 	xfs_lsn_t		old_tail_lsn)
554 {
555 	xfs_lsn_t		tail_lsn;
556 	bool			last_ref;
557 
558 	lockdep_assert_held(&log->l_icloglock);
559 
560 	trace_xlog_iclog_release(iclog, _RET_IP_);
561 	/*
562 	 * Grabbing the current log tail needs to be atomic w.r.t. the writing
563 	 * of the tail LSN into the iclog so we guarantee that the log tail does
564 	 * not move between deciding if a cache flush is required and writing
565 	 * the LSN into the iclog below.
566 	 */
567 	if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
568 		tail_lsn = xlog_assign_tail_lsn(log->l_mp);
569 
570 		if (old_tail_lsn && tail_lsn != old_tail_lsn)
571 			iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
572 
573 		if ((iclog->ic_flags & XLOG_ICL_NEED_FUA) &&
574 		    !iclog->ic_header.h_tail_lsn)
575 			iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
576 	}
577 
578 	last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
579 
580 	if (xlog_is_shutdown(log)) {
581 		/*
582 		 * If there are no more references to this iclog, process the
583 		 * pending iclog callbacks that were waiting on the release of
584 		 * this iclog.
585 		 */
586 		if (last_ref) {
587 			spin_unlock(&log->l_icloglock);
588 			xlog_state_shutdown_callbacks(log);
589 			spin_lock(&log->l_icloglock);
590 		}
591 		return -EIO;
592 	}
593 
594 	if (!last_ref)
595 		return 0;
596 
597 	if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
598 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
599 		return 0;
600 	}
601 
602 	iclog->ic_state = XLOG_STATE_SYNCING;
603 	if (!iclog->ic_header.h_tail_lsn)
604 		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
605 	xlog_verify_tail_lsn(log, iclog);
606 	trace_xlog_iclog_syncing(iclog, _RET_IP_);
607 
608 	spin_unlock(&log->l_icloglock);
609 	xlog_sync(log, iclog);
610 	spin_lock(&log->l_icloglock);
611 	return 0;
612 }
613 
614 /*
615  * Mount a log filesystem
616  *
617  * mp		- ubiquitous xfs mount point structure
618  * log_target	- buftarg of on-disk log device
619  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
620  * num_bblocks	- Number of BBSIZE blocks in on-disk log
621  *
622  * Return error or zero.
623  */
624 int
625 xfs_log_mount(
626 	xfs_mount_t	*mp,
627 	xfs_buftarg_t	*log_target,
628 	xfs_daddr_t	blk_offset,
629 	int		num_bblks)
630 {
631 	struct xlog	*log;
632 	bool		fatal = xfs_has_crc(mp);
633 	int		error = 0;
634 	int		min_logfsbs;
635 
636 	if (!xfs_has_norecovery(mp)) {
637 		xfs_notice(mp, "Mounting V%d Filesystem",
638 			   XFS_SB_VERSION_NUM(&mp->m_sb));
639 	} else {
640 		xfs_notice(mp,
641 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
642 			   XFS_SB_VERSION_NUM(&mp->m_sb));
643 		ASSERT(xfs_is_readonly(mp));
644 	}
645 
646 	log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
647 	if (IS_ERR(log)) {
648 		error = PTR_ERR(log);
649 		goto out;
650 	}
651 	mp->m_log = log;
652 
653 	/*
654 	 * Validate the given log space and drop a critical message via syslog
655 	 * if the log size is too small that would lead to some unexpected
656 	 * situations in transaction log space reservation stage.
657 	 *
658 	 * Note: we can't just reject the mount if the validation fails.  This
659 	 * would mean that people would have to downgrade their kernel just to
660 	 * remedy the situation as there is no way to grow the log (short of
661 	 * black magic surgery with xfs_db).
662 	 *
663 	 * We can, however, reject mounts for CRC format filesystems, as the
664 	 * mkfs binary being used to make the filesystem should never create a
665 	 * filesystem with a log that is too small.
666 	 */
667 	min_logfsbs = xfs_log_calc_minimum_size(mp);
668 
669 	if (mp->m_sb.sb_logblocks < min_logfsbs) {
670 		xfs_warn(mp,
671 		"Log size %d blocks too small, minimum size is %d blocks",
672 			 mp->m_sb.sb_logblocks, min_logfsbs);
673 		error = -EINVAL;
674 	} else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
675 		xfs_warn(mp,
676 		"Log size %d blocks too large, maximum size is %lld blocks",
677 			 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
678 		error = -EINVAL;
679 	} else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
680 		xfs_warn(mp,
681 		"log size %lld bytes too large, maximum size is %lld bytes",
682 			 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
683 			 XFS_MAX_LOG_BYTES);
684 		error = -EINVAL;
685 	} else if (mp->m_sb.sb_logsunit > 1 &&
686 		   mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
687 		xfs_warn(mp,
688 		"log stripe unit %u bytes must be a multiple of block size",
689 			 mp->m_sb.sb_logsunit);
690 		error = -EINVAL;
691 		fatal = true;
692 	}
693 	if (error) {
694 		/*
695 		 * Log check errors are always fatal on v5; or whenever bad
696 		 * metadata leads to a crash.
697 		 */
698 		if (fatal) {
699 			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
700 			ASSERT(0);
701 			goto out_free_log;
702 		}
703 		xfs_crit(mp, "Log size out of supported range.");
704 		xfs_crit(mp,
705 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
706 	}
707 
708 	/*
709 	 * Initialize the AIL now we have a log.
710 	 */
711 	error = xfs_trans_ail_init(mp);
712 	if (error) {
713 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
714 		goto out_free_log;
715 	}
716 	log->l_ailp = mp->m_ail;
717 
718 	/*
719 	 * skip log recovery on a norecovery mount.  pretend it all
720 	 * just worked.
721 	 */
722 	if (!xfs_has_norecovery(mp)) {
723 		/*
724 		 * log recovery ignores readonly state and so we need to clear
725 		 * mount-based read only state so it can write to disk.
726 		 */
727 		bool	readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
728 						&mp->m_opstate);
729 		error = xlog_recover(log);
730 		if (readonly)
731 			set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
732 		if (error) {
733 			xfs_warn(mp, "log mount/recovery failed: error %d",
734 				error);
735 			xlog_recover_cancel(log);
736 			goto out_destroy_ail;
737 		}
738 	}
739 
740 	error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
741 			       "log");
742 	if (error)
743 		goto out_destroy_ail;
744 
745 	/* Normal transactions can now occur */
746 	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
747 
748 	/*
749 	 * Now the log has been fully initialised and we know were our
750 	 * space grant counters are, we can initialise the permanent ticket
751 	 * needed for delayed logging to work.
752 	 */
753 	xlog_cil_init_post_recovery(log);
754 
755 	return 0;
756 
757 out_destroy_ail:
758 	xfs_trans_ail_destroy(mp);
759 out_free_log:
760 	xlog_dealloc_log(log);
761 out:
762 	return error;
763 }
764 
765 /*
766  * Finish the recovery of the file system.  This is separate from the
767  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
768  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
769  * here.
770  *
771  * If we finish recovery successfully, start the background log work. If we are
772  * not doing recovery, then we have a RO filesystem and we don't need to start
773  * it.
774  */
775 int
776 xfs_log_mount_finish(
777 	struct xfs_mount	*mp)
778 {
779 	struct xlog		*log = mp->m_log;
780 	bool			readonly;
781 	int			error = 0;
782 
783 	if (xfs_has_norecovery(mp)) {
784 		ASSERT(xfs_is_readonly(mp));
785 		return 0;
786 	}
787 
788 	/*
789 	 * log recovery ignores readonly state and so we need to clear
790 	 * mount-based read only state so it can write to disk.
791 	 */
792 	readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
793 
794 	/*
795 	 * During the second phase of log recovery, we need iget and
796 	 * iput to behave like they do for an active filesystem.
797 	 * xfs_fs_drop_inode needs to be able to prevent the deletion
798 	 * of inodes before we're done replaying log items on those
799 	 * inodes.  Turn it off immediately after recovery finishes
800 	 * so that we don't leak the quota inodes if subsequent mount
801 	 * activities fail.
802 	 *
803 	 * We let all inodes involved in redo item processing end up on
804 	 * the LRU instead of being evicted immediately so that if we do
805 	 * something to an unlinked inode, the irele won't cause
806 	 * premature truncation and freeing of the inode, which results
807 	 * in log recovery failure.  We have to evict the unreferenced
808 	 * lru inodes after clearing SB_ACTIVE because we don't
809 	 * otherwise clean up the lru if there's a subsequent failure in
810 	 * xfs_mountfs, which leads to us leaking the inodes if nothing
811 	 * else (e.g. quotacheck) references the inodes before the
812 	 * mount failure occurs.
813 	 */
814 	mp->m_super->s_flags |= SB_ACTIVE;
815 	xfs_log_work_queue(mp);
816 	if (xlog_recovery_needed(log))
817 		error = xlog_recover_finish(log);
818 	mp->m_super->s_flags &= ~SB_ACTIVE;
819 	evict_inodes(mp->m_super);
820 
821 	/*
822 	 * Drain the buffer LRU after log recovery. This is required for v4
823 	 * filesystems to avoid leaving around buffers with NULL verifier ops,
824 	 * but we do it unconditionally to make sure we're always in a clean
825 	 * cache state after mount.
826 	 *
827 	 * Don't push in the error case because the AIL may have pending intents
828 	 * that aren't removed until recovery is cancelled.
829 	 */
830 	if (xlog_recovery_needed(log)) {
831 		if (!error) {
832 			xfs_log_force(mp, XFS_LOG_SYNC);
833 			xfs_ail_push_all_sync(mp->m_ail);
834 		}
835 		xfs_notice(mp, "Ending recovery (logdev: %s)",
836 				mp->m_logname ? mp->m_logname : "internal");
837 	} else {
838 		xfs_info(mp, "Ending clean mount");
839 	}
840 	xfs_buftarg_drain(mp->m_ddev_targp);
841 
842 	clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
843 	if (readonly)
844 		set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
845 
846 	/* Make sure the log is dead if we're returning failure. */
847 	ASSERT(!error || xlog_is_shutdown(log));
848 
849 	return error;
850 }
851 
852 /*
853  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
854  * the log.
855  */
856 void
857 xfs_log_mount_cancel(
858 	struct xfs_mount	*mp)
859 {
860 	xlog_recover_cancel(mp->m_log);
861 	xfs_log_unmount(mp);
862 }
863 
864 /*
865  * Flush out the iclog to disk ensuring that device caches are flushed and
866  * the iclog hits stable storage before any completion waiters are woken.
867  */
868 static inline int
869 xlog_force_iclog(
870 	struct xlog_in_core	*iclog)
871 {
872 	atomic_inc(&iclog->ic_refcnt);
873 	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
874 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
875 		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
876 	return xlog_state_release_iclog(iclog->ic_log, iclog, 0);
877 }
878 
879 /*
880  * Wait for the iclog and all prior iclogs to be written disk as required by the
881  * log force state machine. Waiting on ic_force_wait ensures iclog completions
882  * have been ordered and callbacks run before we are woken here, hence
883  * guaranteeing that all the iclogs up to this one are on stable storage.
884  */
885 int
886 xlog_wait_on_iclog(
887 	struct xlog_in_core	*iclog)
888 		__releases(iclog->ic_log->l_icloglock)
889 {
890 	struct xlog		*log = iclog->ic_log;
891 
892 	trace_xlog_iclog_wait_on(iclog, _RET_IP_);
893 	if (!xlog_is_shutdown(log) &&
894 	    iclog->ic_state != XLOG_STATE_ACTIVE &&
895 	    iclog->ic_state != XLOG_STATE_DIRTY) {
896 		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
897 		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
898 	} else {
899 		spin_unlock(&log->l_icloglock);
900 	}
901 
902 	if (xlog_is_shutdown(log))
903 		return -EIO;
904 	return 0;
905 }
906 
907 /*
908  * Write out an unmount record using the ticket provided. We have to account for
909  * the data space used in the unmount ticket as this write is not done from a
910  * transaction context that has already done the accounting for us.
911  */
912 static int
913 xlog_write_unmount_record(
914 	struct xlog		*log,
915 	struct xlog_ticket	*ticket)
916 {
917 	struct xfs_unmount_log_format ulf = {
918 		.magic = XLOG_UNMOUNT_TYPE,
919 	};
920 	struct xfs_log_iovec reg = {
921 		.i_addr = &ulf,
922 		.i_len = sizeof(ulf),
923 		.i_type = XLOG_REG_TYPE_UNMOUNT,
924 	};
925 	struct xfs_log_vec vec = {
926 		.lv_niovecs = 1,
927 		.lv_iovecp = &reg,
928 	};
929 
930 	/* account for space used by record data */
931 	ticket->t_curr_res -= sizeof(ulf);
932 
933 	return xlog_write(log, NULL, &vec, ticket, XLOG_UNMOUNT_TRANS);
934 }
935 
936 /*
937  * Mark the filesystem clean by writing an unmount record to the head of the
938  * log.
939  */
940 static void
941 xlog_unmount_write(
942 	struct xlog		*log)
943 {
944 	struct xfs_mount	*mp = log->l_mp;
945 	struct xlog_in_core	*iclog;
946 	struct xlog_ticket	*tic = NULL;
947 	int			error;
948 
949 	error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
950 	if (error)
951 		goto out_err;
952 
953 	error = xlog_write_unmount_record(log, tic);
954 	/*
955 	 * At this point, we're umounting anyway, so there's no point in
956 	 * transitioning log state to shutdown. Just continue...
957 	 */
958 out_err:
959 	if (error)
960 		xfs_alert(mp, "%s: unmount record failed", __func__);
961 
962 	spin_lock(&log->l_icloglock);
963 	iclog = log->l_iclog;
964 	error = xlog_force_iclog(iclog);
965 	xlog_wait_on_iclog(iclog);
966 
967 	if (tic) {
968 		trace_xfs_log_umount_write(log, tic);
969 		xfs_log_ticket_ungrant(log, tic);
970 	}
971 }
972 
973 static void
974 xfs_log_unmount_verify_iclog(
975 	struct xlog		*log)
976 {
977 	struct xlog_in_core	*iclog = log->l_iclog;
978 
979 	do {
980 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
981 		ASSERT(iclog->ic_offset == 0);
982 	} while ((iclog = iclog->ic_next) != log->l_iclog);
983 }
984 
985 /*
986  * Unmount record used to have a string "Unmount filesystem--" in the
987  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
988  * We just write the magic number now since that particular field isn't
989  * currently architecture converted and "Unmount" is a bit foo.
990  * As far as I know, there weren't any dependencies on the old behaviour.
991  */
992 static void
993 xfs_log_unmount_write(
994 	struct xfs_mount	*mp)
995 {
996 	struct xlog		*log = mp->m_log;
997 
998 	if (!xfs_log_writable(mp))
999 		return;
1000 
1001 	xfs_log_force(mp, XFS_LOG_SYNC);
1002 
1003 	if (xlog_is_shutdown(log))
1004 		return;
1005 
1006 	/*
1007 	 * If we think the summary counters are bad, avoid writing the unmount
1008 	 * record to force log recovery at next mount, after which the summary
1009 	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
1010 	 * more details.
1011 	 */
1012 	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1013 			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1014 		xfs_alert(mp, "%s: will fix summary counters at next mount",
1015 				__func__);
1016 		return;
1017 	}
1018 
1019 	xfs_log_unmount_verify_iclog(log);
1020 	xlog_unmount_write(log);
1021 }
1022 
1023 /*
1024  * Empty the log for unmount/freeze.
1025  *
1026  * To do this, we first need to shut down the background log work so it is not
1027  * trying to cover the log as we clean up. We then need to unpin all objects in
1028  * the log so we can then flush them out. Once they have completed their IO and
1029  * run the callbacks removing themselves from the AIL, we can cover the log.
1030  */
1031 int
1032 xfs_log_quiesce(
1033 	struct xfs_mount	*mp)
1034 {
1035 	/*
1036 	 * Clear log incompat features since we're quiescing the log.  Report
1037 	 * failures, though it's not fatal to have a higher log feature
1038 	 * protection level than the log contents actually require.
1039 	 */
1040 	if (xfs_clear_incompat_log_features(mp)) {
1041 		int error;
1042 
1043 		error = xfs_sync_sb(mp, false);
1044 		if (error)
1045 			xfs_warn(mp,
1046 	"Failed to clear log incompat features on quiesce");
1047 	}
1048 
1049 	cancel_delayed_work_sync(&mp->m_log->l_work);
1050 	xfs_log_force(mp, XFS_LOG_SYNC);
1051 
1052 	/*
1053 	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1054 	 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1055 	 * xfs_buf_iowait() cannot be used because it was pushed with the
1056 	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1057 	 * the IO to complete.
1058 	 */
1059 	xfs_ail_push_all_sync(mp->m_ail);
1060 	xfs_buftarg_wait(mp->m_ddev_targp);
1061 	xfs_buf_lock(mp->m_sb_bp);
1062 	xfs_buf_unlock(mp->m_sb_bp);
1063 
1064 	return xfs_log_cover(mp);
1065 }
1066 
1067 void
1068 xfs_log_clean(
1069 	struct xfs_mount	*mp)
1070 {
1071 	xfs_log_quiesce(mp);
1072 	xfs_log_unmount_write(mp);
1073 }
1074 
1075 /*
1076  * Shut down and release the AIL and Log.
1077  *
1078  * During unmount, we need to ensure we flush all the dirty metadata objects
1079  * from the AIL so that the log is empty before we write the unmount record to
1080  * the log. Once this is done, we can tear down the AIL and the log.
1081  */
1082 void
1083 xfs_log_unmount(
1084 	struct xfs_mount	*mp)
1085 {
1086 	xfs_log_clean(mp);
1087 
1088 	xfs_buftarg_drain(mp->m_ddev_targp);
1089 
1090 	xfs_trans_ail_destroy(mp);
1091 
1092 	xfs_sysfs_del(&mp->m_log->l_kobj);
1093 
1094 	xlog_dealloc_log(mp->m_log);
1095 }
1096 
1097 void
1098 xfs_log_item_init(
1099 	struct xfs_mount	*mp,
1100 	struct xfs_log_item	*item,
1101 	int			type,
1102 	const struct xfs_item_ops *ops)
1103 {
1104 	item->li_log = mp->m_log;
1105 	item->li_ailp = mp->m_ail;
1106 	item->li_type = type;
1107 	item->li_ops = ops;
1108 	item->li_lv = NULL;
1109 
1110 	INIT_LIST_HEAD(&item->li_ail);
1111 	INIT_LIST_HEAD(&item->li_cil);
1112 	INIT_LIST_HEAD(&item->li_bio_list);
1113 	INIT_LIST_HEAD(&item->li_trans);
1114 }
1115 
1116 /*
1117  * Wake up processes waiting for log space after we have moved the log tail.
1118  */
1119 void
1120 xfs_log_space_wake(
1121 	struct xfs_mount	*mp)
1122 {
1123 	struct xlog		*log = mp->m_log;
1124 	int			free_bytes;
1125 
1126 	if (xlog_is_shutdown(log))
1127 		return;
1128 
1129 	if (!list_empty_careful(&log->l_write_head.waiters)) {
1130 		ASSERT(!xlog_in_recovery(log));
1131 
1132 		spin_lock(&log->l_write_head.lock);
1133 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1134 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1135 		spin_unlock(&log->l_write_head.lock);
1136 	}
1137 
1138 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1139 		ASSERT(!xlog_in_recovery(log));
1140 
1141 		spin_lock(&log->l_reserve_head.lock);
1142 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1143 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1144 		spin_unlock(&log->l_reserve_head.lock);
1145 	}
1146 }
1147 
1148 /*
1149  * Determine if we have a transaction that has gone to disk that needs to be
1150  * covered. To begin the transition to the idle state firstly the log needs to
1151  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1152  * we start attempting to cover the log.
1153  *
1154  * Only if we are then in a state where covering is needed, the caller is
1155  * informed that dummy transactions are required to move the log into the idle
1156  * state.
1157  *
1158  * If there are any items in the AIl or CIL, then we do not want to attempt to
1159  * cover the log as we may be in a situation where there isn't log space
1160  * available to run a dummy transaction and this can lead to deadlocks when the
1161  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1162  * there's no point in running a dummy transaction at this point because we
1163  * can't start trying to idle the log until both the CIL and AIL are empty.
1164  */
1165 static bool
1166 xfs_log_need_covered(
1167 	struct xfs_mount	*mp)
1168 {
1169 	struct xlog		*log = mp->m_log;
1170 	bool			needed = false;
1171 
1172 	if (!xlog_cil_empty(log))
1173 		return false;
1174 
1175 	spin_lock(&log->l_icloglock);
1176 	switch (log->l_covered_state) {
1177 	case XLOG_STATE_COVER_DONE:
1178 	case XLOG_STATE_COVER_DONE2:
1179 	case XLOG_STATE_COVER_IDLE:
1180 		break;
1181 	case XLOG_STATE_COVER_NEED:
1182 	case XLOG_STATE_COVER_NEED2:
1183 		if (xfs_ail_min_lsn(log->l_ailp))
1184 			break;
1185 		if (!xlog_iclogs_empty(log))
1186 			break;
1187 
1188 		needed = true;
1189 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1190 			log->l_covered_state = XLOG_STATE_COVER_DONE;
1191 		else
1192 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1193 		break;
1194 	default:
1195 		needed = true;
1196 		break;
1197 	}
1198 	spin_unlock(&log->l_icloglock);
1199 	return needed;
1200 }
1201 
1202 /*
1203  * Explicitly cover the log. This is similar to background log covering but
1204  * intended for usage in quiesce codepaths. The caller is responsible to ensure
1205  * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1206  * must all be empty.
1207  */
1208 static int
1209 xfs_log_cover(
1210 	struct xfs_mount	*mp)
1211 {
1212 	int			error = 0;
1213 	bool			need_covered;
1214 
1215 	ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1216 	        !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1217 		xlog_is_shutdown(mp->m_log));
1218 
1219 	if (!xfs_log_writable(mp))
1220 		return 0;
1221 
1222 	/*
1223 	 * xfs_log_need_covered() is not idempotent because it progresses the
1224 	 * state machine if the log requires covering. Therefore, we must call
1225 	 * this function once and use the result until we've issued an sb sync.
1226 	 * Do so first to make that abundantly clear.
1227 	 *
1228 	 * Fall into the covering sequence if the log needs covering or the
1229 	 * mount has lazy superblock accounting to sync to disk. The sb sync
1230 	 * used for covering accumulates the in-core counters, so covering
1231 	 * handles this for us.
1232 	 */
1233 	need_covered = xfs_log_need_covered(mp);
1234 	if (!need_covered && !xfs_has_lazysbcount(mp))
1235 		return 0;
1236 
1237 	/*
1238 	 * To cover the log, commit the superblock twice (at most) in
1239 	 * independent checkpoints. The first serves as a reference for the
1240 	 * tail pointer. The sync transaction and AIL push empties the AIL and
1241 	 * updates the in-core tail to the LSN of the first checkpoint. The
1242 	 * second commit updates the on-disk tail with the in-core LSN,
1243 	 * covering the log. Push the AIL one more time to leave it empty, as
1244 	 * we found it.
1245 	 */
1246 	do {
1247 		error = xfs_sync_sb(mp, true);
1248 		if (error)
1249 			break;
1250 		xfs_ail_push_all_sync(mp->m_ail);
1251 	} while (xfs_log_need_covered(mp));
1252 
1253 	return error;
1254 }
1255 
1256 /*
1257  * We may be holding the log iclog lock upon entering this routine.
1258  */
1259 xfs_lsn_t
1260 xlog_assign_tail_lsn_locked(
1261 	struct xfs_mount	*mp)
1262 {
1263 	struct xlog		*log = mp->m_log;
1264 	struct xfs_log_item	*lip;
1265 	xfs_lsn_t		tail_lsn;
1266 
1267 	assert_spin_locked(&mp->m_ail->ail_lock);
1268 
1269 	/*
1270 	 * To make sure we always have a valid LSN for the log tail we keep
1271 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1272 	 * and use that when the AIL was empty.
1273 	 */
1274 	lip = xfs_ail_min(mp->m_ail);
1275 	if (lip)
1276 		tail_lsn = lip->li_lsn;
1277 	else
1278 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1279 	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1280 	atomic64_set(&log->l_tail_lsn, tail_lsn);
1281 	return tail_lsn;
1282 }
1283 
1284 xfs_lsn_t
1285 xlog_assign_tail_lsn(
1286 	struct xfs_mount	*mp)
1287 {
1288 	xfs_lsn_t		tail_lsn;
1289 
1290 	spin_lock(&mp->m_ail->ail_lock);
1291 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1292 	spin_unlock(&mp->m_ail->ail_lock);
1293 
1294 	return tail_lsn;
1295 }
1296 
1297 /*
1298  * Return the space in the log between the tail and the head.  The head
1299  * is passed in the cycle/bytes formal parms.  In the special case where
1300  * the reserve head has wrapped passed the tail, this calculation is no
1301  * longer valid.  In this case, just return 0 which means there is no space
1302  * in the log.  This works for all places where this function is called
1303  * with the reserve head.  Of course, if the write head were to ever
1304  * wrap the tail, we should blow up.  Rather than catch this case here,
1305  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1306  *
1307  * If reservation head is behind the tail, we have a problem. Warn about it,
1308  * but then treat it as if the log is empty.
1309  *
1310  * If the log is shut down, the head and tail may be invalid or out of whack, so
1311  * shortcut invalidity asserts in this case so that we don't trigger them
1312  * falsely.
1313  */
1314 STATIC int
1315 xlog_space_left(
1316 	struct xlog	*log,
1317 	atomic64_t	*head)
1318 {
1319 	int		tail_bytes;
1320 	int		tail_cycle;
1321 	int		head_cycle;
1322 	int		head_bytes;
1323 
1324 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1325 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1326 	tail_bytes = BBTOB(tail_bytes);
1327 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1328 		return log->l_logsize - (head_bytes - tail_bytes);
1329 	if (tail_cycle + 1 < head_cycle)
1330 		return 0;
1331 
1332 	/* Ignore potential inconsistency when shutdown. */
1333 	if (xlog_is_shutdown(log))
1334 		return log->l_logsize;
1335 
1336 	if (tail_cycle < head_cycle) {
1337 		ASSERT(tail_cycle == (head_cycle - 1));
1338 		return tail_bytes - head_bytes;
1339 	}
1340 
1341 	/*
1342 	 * The reservation head is behind the tail. In this case we just want to
1343 	 * return the size of the log as the amount of space left.
1344 	 */
1345 	xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1346 	xfs_alert(log->l_mp, "  tail_cycle = %d, tail_bytes = %d",
1347 		  tail_cycle, tail_bytes);
1348 	xfs_alert(log->l_mp, "  GH   cycle = %d, GH   bytes = %d",
1349 		  head_cycle, head_bytes);
1350 	ASSERT(0);
1351 	return log->l_logsize;
1352 }
1353 
1354 
1355 static void
1356 xlog_ioend_work(
1357 	struct work_struct	*work)
1358 {
1359 	struct xlog_in_core     *iclog =
1360 		container_of(work, struct xlog_in_core, ic_end_io_work);
1361 	struct xlog		*log = iclog->ic_log;
1362 	int			error;
1363 
1364 	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1365 #ifdef DEBUG
1366 	/* treat writes with injected CRC errors as failed */
1367 	if (iclog->ic_fail_crc)
1368 		error = -EIO;
1369 #endif
1370 
1371 	/*
1372 	 * Race to shutdown the filesystem if we see an error.
1373 	 */
1374 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1375 		xfs_alert(log->l_mp, "log I/O error %d", error);
1376 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1377 	}
1378 
1379 	xlog_state_done_syncing(iclog);
1380 	bio_uninit(&iclog->ic_bio);
1381 
1382 	/*
1383 	 * Drop the lock to signal that we are done. Nothing references the
1384 	 * iclog after this, so an unmount waiting on this lock can now tear it
1385 	 * down safely. As such, it is unsafe to reference the iclog after the
1386 	 * unlock as we could race with it being freed.
1387 	 */
1388 	up(&iclog->ic_sema);
1389 }
1390 
1391 /*
1392  * Return size of each in-core log record buffer.
1393  *
1394  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1395  *
1396  * If the filesystem blocksize is too large, we may need to choose a
1397  * larger size since the directory code currently logs entire blocks.
1398  */
1399 STATIC void
1400 xlog_get_iclog_buffer_size(
1401 	struct xfs_mount	*mp,
1402 	struct xlog		*log)
1403 {
1404 	if (mp->m_logbufs <= 0)
1405 		mp->m_logbufs = XLOG_MAX_ICLOGS;
1406 	if (mp->m_logbsize <= 0)
1407 		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1408 
1409 	log->l_iclog_bufs = mp->m_logbufs;
1410 	log->l_iclog_size = mp->m_logbsize;
1411 
1412 	/*
1413 	 * # headers = size / 32k - one header holds cycles from 32k of data.
1414 	 */
1415 	log->l_iclog_heads =
1416 		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1417 	log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1418 }
1419 
1420 void
1421 xfs_log_work_queue(
1422 	struct xfs_mount        *mp)
1423 {
1424 	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1425 				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1426 }
1427 
1428 /*
1429  * Clear the log incompat flags if we have the opportunity.
1430  *
1431  * This only happens if we're about to log the second dummy transaction as part
1432  * of covering the log and we can get the log incompat feature usage lock.
1433  */
1434 static inline void
1435 xlog_clear_incompat(
1436 	struct xlog		*log)
1437 {
1438 	struct xfs_mount	*mp = log->l_mp;
1439 
1440 	if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1441 				XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1442 		return;
1443 
1444 	if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1445 		return;
1446 
1447 	if (!down_write_trylock(&log->l_incompat_users))
1448 		return;
1449 
1450 	xfs_clear_incompat_log_features(mp);
1451 	up_write(&log->l_incompat_users);
1452 }
1453 
1454 /*
1455  * Every sync period we need to unpin all items in the AIL and push them to
1456  * disk. If there is nothing dirty, then we might need to cover the log to
1457  * indicate that the filesystem is idle.
1458  */
1459 static void
1460 xfs_log_worker(
1461 	struct work_struct	*work)
1462 {
1463 	struct xlog		*log = container_of(to_delayed_work(work),
1464 						struct xlog, l_work);
1465 	struct xfs_mount	*mp = log->l_mp;
1466 
1467 	/* dgc: errors ignored - not fatal and nowhere to report them */
1468 	if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1469 		/*
1470 		 * Dump a transaction into the log that contains no real change.
1471 		 * This is needed to stamp the current tail LSN into the log
1472 		 * during the covering operation.
1473 		 *
1474 		 * We cannot use an inode here for this - that will push dirty
1475 		 * state back up into the VFS and then periodic inode flushing
1476 		 * will prevent log covering from making progress. Hence we
1477 		 * synchronously log the superblock instead to ensure the
1478 		 * superblock is immediately unpinned and can be written back.
1479 		 */
1480 		xlog_clear_incompat(log);
1481 		xfs_sync_sb(mp, true);
1482 	} else
1483 		xfs_log_force(mp, 0);
1484 
1485 	/* start pushing all the metadata that is currently dirty */
1486 	xfs_ail_push_all(mp->m_ail);
1487 
1488 	/* queue us up again */
1489 	xfs_log_work_queue(mp);
1490 }
1491 
1492 /*
1493  * This routine initializes some of the log structure for a given mount point.
1494  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1495  * some other stuff may be filled in too.
1496  */
1497 STATIC struct xlog *
1498 xlog_alloc_log(
1499 	struct xfs_mount	*mp,
1500 	struct xfs_buftarg	*log_target,
1501 	xfs_daddr_t		blk_offset,
1502 	int			num_bblks)
1503 {
1504 	struct xlog		*log;
1505 	xlog_rec_header_t	*head;
1506 	xlog_in_core_t		**iclogp;
1507 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1508 	int			i;
1509 	int			error = -ENOMEM;
1510 	uint			log2_size = 0;
1511 
1512 	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1513 	if (!log) {
1514 		xfs_warn(mp, "Log allocation failed: No memory!");
1515 		goto out;
1516 	}
1517 
1518 	log->l_mp	   = mp;
1519 	log->l_targ	   = log_target;
1520 	log->l_logsize     = BBTOB(num_bblks);
1521 	log->l_logBBstart  = blk_offset;
1522 	log->l_logBBsize   = num_bblks;
1523 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1524 	set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1525 	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1526 
1527 	log->l_prev_block  = -1;
1528 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1529 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1530 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1531 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1532 
1533 	if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1534 		log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1535 	else
1536 		log->l_iclog_roundoff = BBSIZE;
1537 
1538 	xlog_grant_head_init(&log->l_reserve_head);
1539 	xlog_grant_head_init(&log->l_write_head);
1540 
1541 	error = -EFSCORRUPTED;
1542 	if (xfs_has_sector(mp)) {
1543 	        log2_size = mp->m_sb.sb_logsectlog;
1544 		if (log2_size < BBSHIFT) {
1545 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1546 				log2_size, BBSHIFT);
1547 			goto out_free_log;
1548 		}
1549 
1550 	        log2_size -= BBSHIFT;
1551 		if (log2_size > mp->m_sectbb_log) {
1552 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1553 				log2_size, mp->m_sectbb_log);
1554 			goto out_free_log;
1555 		}
1556 
1557 		/* for larger sector sizes, must have v2 or external log */
1558 		if (log2_size && log->l_logBBstart > 0 &&
1559 			    !xfs_has_logv2(mp)) {
1560 			xfs_warn(mp,
1561 		"log sector size (0x%x) invalid for configuration.",
1562 				log2_size);
1563 			goto out_free_log;
1564 		}
1565 	}
1566 	log->l_sectBBsize = 1 << log2_size;
1567 
1568 	init_rwsem(&log->l_incompat_users);
1569 
1570 	xlog_get_iclog_buffer_size(mp, log);
1571 
1572 	spin_lock_init(&log->l_icloglock);
1573 	init_waitqueue_head(&log->l_flush_wait);
1574 
1575 	iclogp = &log->l_iclog;
1576 	/*
1577 	 * The amount of memory to allocate for the iclog structure is
1578 	 * rather funky due to the way the structure is defined.  It is
1579 	 * done this way so that we can use different sizes for machines
1580 	 * with different amounts of memory.  See the definition of
1581 	 * xlog_in_core_t in xfs_log_priv.h for details.
1582 	 */
1583 	ASSERT(log->l_iclog_size >= 4096);
1584 	for (i = 0; i < log->l_iclog_bufs; i++) {
1585 		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1586 				sizeof(struct bio_vec);
1587 
1588 		iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1589 		if (!iclog)
1590 			goto out_free_iclog;
1591 
1592 		*iclogp = iclog;
1593 		iclog->ic_prev = prev_iclog;
1594 		prev_iclog = iclog;
1595 
1596 		iclog->ic_data = kvzalloc(log->l_iclog_size,
1597 				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1598 		if (!iclog->ic_data)
1599 			goto out_free_iclog;
1600 #ifdef DEBUG
1601 		log->l_iclog_bak[i] = &iclog->ic_header;
1602 #endif
1603 		head = &iclog->ic_header;
1604 		memset(head, 0, sizeof(xlog_rec_header_t));
1605 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1606 		head->h_version = cpu_to_be32(
1607 			xfs_has_logv2(log->l_mp) ? 2 : 1);
1608 		head->h_size = cpu_to_be32(log->l_iclog_size);
1609 		/* new fields */
1610 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1611 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1612 
1613 		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1614 		iclog->ic_state = XLOG_STATE_ACTIVE;
1615 		iclog->ic_log = log;
1616 		atomic_set(&iclog->ic_refcnt, 0);
1617 		INIT_LIST_HEAD(&iclog->ic_callbacks);
1618 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1619 
1620 		init_waitqueue_head(&iclog->ic_force_wait);
1621 		init_waitqueue_head(&iclog->ic_write_wait);
1622 		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1623 		sema_init(&iclog->ic_sema, 1);
1624 
1625 		iclogp = &iclog->ic_next;
1626 	}
1627 	*iclogp = log->l_iclog;			/* complete ring */
1628 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1629 
1630 	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1631 			XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1632 				    WQ_HIGHPRI),
1633 			0, mp->m_super->s_id);
1634 	if (!log->l_ioend_workqueue)
1635 		goto out_free_iclog;
1636 
1637 	error = xlog_cil_init(log);
1638 	if (error)
1639 		goto out_destroy_workqueue;
1640 	return log;
1641 
1642 out_destroy_workqueue:
1643 	destroy_workqueue(log->l_ioend_workqueue);
1644 out_free_iclog:
1645 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1646 		prev_iclog = iclog->ic_next;
1647 		kmem_free(iclog->ic_data);
1648 		kmem_free(iclog);
1649 		if (prev_iclog == log->l_iclog)
1650 			break;
1651 	}
1652 out_free_log:
1653 	kmem_free(log);
1654 out:
1655 	return ERR_PTR(error);
1656 }	/* xlog_alloc_log */
1657 
1658 /*
1659  * Compute the LSN that we'd need to push the log tail towards in order to have
1660  * (a) enough on-disk log space to log the number of bytes specified, (b) at
1661  * least 25% of the log space free, and (c) at least 256 blocks free.  If the
1662  * log free space already meets all three thresholds, this function returns
1663  * NULLCOMMITLSN.
1664  */
1665 xfs_lsn_t
1666 xlog_grant_push_threshold(
1667 	struct xlog	*log,
1668 	int		need_bytes)
1669 {
1670 	xfs_lsn_t	threshold_lsn = 0;
1671 	xfs_lsn_t	last_sync_lsn;
1672 	int		free_blocks;
1673 	int		free_bytes;
1674 	int		threshold_block;
1675 	int		threshold_cycle;
1676 	int		free_threshold;
1677 
1678 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1679 
1680 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1681 	free_blocks = BTOBBT(free_bytes);
1682 
1683 	/*
1684 	 * Set the threshold for the minimum number of free blocks in the
1685 	 * log to the maximum of what the caller needs, one quarter of the
1686 	 * log, and 256 blocks.
1687 	 */
1688 	free_threshold = BTOBB(need_bytes);
1689 	free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1690 	free_threshold = max(free_threshold, 256);
1691 	if (free_blocks >= free_threshold)
1692 		return NULLCOMMITLSN;
1693 
1694 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1695 						&threshold_block);
1696 	threshold_block += free_threshold;
1697 	if (threshold_block >= log->l_logBBsize) {
1698 		threshold_block -= log->l_logBBsize;
1699 		threshold_cycle += 1;
1700 	}
1701 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1702 					threshold_block);
1703 	/*
1704 	 * Don't pass in an lsn greater than the lsn of the last
1705 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1706 	 * so that it doesn't change between the compare and the set.
1707 	 */
1708 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1709 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1710 		threshold_lsn = last_sync_lsn;
1711 
1712 	return threshold_lsn;
1713 }
1714 
1715 /*
1716  * Push the tail of the log if we need to do so to maintain the free log space
1717  * thresholds set out by xlog_grant_push_threshold.  We may need to adopt a
1718  * policy which pushes on an lsn which is further along in the log once we
1719  * reach the high water mark.  In this manner, we would be creating a low water
1720  * mark.
1721  */
1722 STATIC void
1723 xlog_grant_push_ail(
1724 	struct xlog	*log,
1725 	int		need_bytes)
1726 {
1727 	xfs_lsn_t	threshold_lsn;
1728 
1729 	threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1730 	if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1731 		return;
1732 
1733 	/*
1734 	 * Get the transaction layer to kick the dirty buffers out to
1735 	 * disk asynchronously. No point in trying to do this if
1736 	 * the filesystem is shutting down.
1737 	 */
1738 	xfs_ail_push(log->l_ailp, threshold_lsn);
1739 }
1740 
1741 /*
1742  * Stamp cycle number in every block
1743  */
1744 STATIC void
1745 xlog_pack_data(
1746 	struct xlog		*log,
1747 	struct xlog_in_core	*iclog,
1748 	int			roundoff)
1749 {
1750 	int			i, j, k;
1751 	int			size = iclog->ic_offset + roundoff;
1752 	__be32			cycle_lsn;
1753 	char			*dp;
1754 
1755 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1756 
1757 	dp = iclog->ic_datap;
1758 	for (i = 0; i < BTOBB(size); i++) {
1759 		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1760 			break;
1761 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1762 		*(__be32 *)dp = cycle_lsn;
1763 		dp += BBSIZE;
1764 	}
1765 
1766 	if (xfs_has_logv2(log->l_mp)) {
1767 		xlog_in_core_2_t *xhdr = iclog->ic_data;
1768 
1769 		for ( ; i < BTOBB(size); i++) {
1770 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1771 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1772 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1773 			*(__be32 *)dp = cycle_lsn;
1774 			dp += BBSIZE;
1775 		}
1776 
1777 		for (i = 1; i < log->l_iclog_heads; i++)
1778 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1779 	}
1780 }
1781 
1782 /*
1783  * Calculate the checksum for a log buffer.
1784  *
1785  * This is a little more complicated than it should be because the various
1786  * headers and the actual data are non-contiguous.
1787  */
1788 __le32
1789 xlog_cksum(
1790 	struct xlog		*log,
1791 	struct xlog_rec_header	*rhead,
1792 	char			*dp,
1793 	int			size)
1794 {
1795 	uint32_t		crc;
1796 
1797 	/* first generate the crc for the record header ... */
1798 	crc = xfs_start_cksum_update((char *)rhead,
1799 			      sizeof(struct xlog_rec_header),
1800 			      offsetof(struct xlog_rec_header, h_crc));
1801 
1802 	/* ... then for additional cycle data for v2 logs ... */
1803 	if (xfs_has_logv2(log->l_mp)) {
1804 		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1805 		int		i;
1806 		int		xheads;
1807 
1808 		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1809 
1810 		for (i = 1; i < xheads; i++) {
1811 			crc = crc32c(crc, &xhdr[i].hic_xheader,
1812 				     sizeof(struct xlog_rec_ext_header));
1813 		}
1814 	}
1815 
1816 	/* ... and finally for the payload */
1817 	crc = crc32c(crc, dp, size);
1818 
1819 	return xfs_end_cksum(crc);
1820 }
1821 
1822 static void
1823 xlog_bio_end_io(
1824 	struct bio		*bio)
1825 {
1826 	struct xlog_in_core	*iclog = bio->bi_private;
1827 
1828 	queue_work(iclog->ic_log->l_ioend_workqueue,
1829 		   &iclog->ic_end_io_work);
1830 }
1831 
1832 static int
1833 xlog_map_iclog_data(
1834 	struct bio		*bio,
1835 	void			*data,
1836 	size_t			count)
1837 {
1838 	do {
1839 		struct page	*page = kmem_to_page(data);
1840 		unsigned int	off = offset_in_page(data);
1841 		size_t		len = min_t(size_t, count, PAGE_SIZE - off);
1842 
1843 		if (bio_add_page(bio, page, len, off) != len)
1844 			return -EIO;
1845 
1846 		data += len;
1847 		count -= len;
1848 	} while (count);
1849 
1850 	return 0;
1851 }
1852 
1853 STATIC void
1854 xlog_write_iclog(
1855 	struct xlog		*log,
1856 	struct xlog_in_core	*iclog,
1857 	uint64_t		bno,
1858 	unsigned int		count)
1859 {
1860 	ASSERT(bno < log->l_logBBsize);
1861 	trace_xlog_iclog_write(iclog, _RET_IP_);
1862 
1863 	/*
1864 	 * We lock the iclogbufs here so that we can serialise against I/O
1865 	 * completion during unmount.  We might be processing a shutdown
1866 	 * triggered during unmount, and that can occur asynchronously to the
1867 	 * unmount thread, and hence we need to ensure that completes before
1868 	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1869 	 * across the log IO to archieve that.
1870 	 */
1871 	down(&iclog->ic_sema);
1872 	if (xlog_is_shutdown(log)) {
1873 		/*
1874 		 * It would seem logical to return EIO here, but we rely on
1875 		 * the log state machine to propagate I/O errors instead of
1876 		 * doing it here.  We kick of the state machine and unlock
1877 		 * the buffer manually, the code needs to be kept in sync
1878 		 * with the I/O completion path.
1879 		 */
1880 		xlog_state_done_syncing(iclog);
1881 		up(&iclog->ic_sema);
1882 		return;
1883 	}
1884 
1885 	/*
1886 	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1887 	 * IOs coming immediately after this one. This prevents the block layer
1888 	 * writeback throttle from throttling log writes behind background
1889 	 * metadata writeback and causing priority inversions.
1890 	 */
1891 	bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1892 		 howmany(count, PAGE_SIZE),
1893 		 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1894 	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1895 	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1896 	iclog->ic_bio.bi_private = iclog;
1897 
1898 	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1899 		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1900 		/*
1901 		 * For external log devices, we also need to flush the data
1902 		 * device cache first to ensure all metadata writeback covered
1903 		 * by the LSN in this iclog is on stable storage. This is slow,
1904 		 * but it *must* complete before we issue the external log IO.
1905 		 */
1906 		if (log->l_targ != log->l_mp->m_ddev_targp)
1907 			blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
1908 	}
1909 	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1910 		iclog->ic_bio.bi_opf |= REQ_FUA;
1911 
1912 	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1913 
1914 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1915 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1916 		return;
1917 	}
1918 	if (is_vmalloc_addr(iclog->ic_data))
1919 		flush_kernel_vmap_range(iclog->ic_data, count);
1920 
1921 	/*
1922 	 * If this log buffer would straddle the end of the log we will have
1923 	 * to split it up into two bios, so that we can continue at the start.
1924 	 */
1925 	if (bno + BTOBB(count) > log->l_logBBsize) {
1926 		struct bio *split;
1927 
1928 		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1929 				  GFP_NOIO, &fs_bio_set);
1930 		bio_chain(split, &iclog->ic_bio);
1931 		submit_bio(split);
1932 
1933 		/* restart at logical offset zero for the remainder */
1934 		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1935 	}
1936 
1937 	submit_bio(&iclog->ic_bio);
1938 }
1939 
1940 /*
1941  * We need to bump cycle number for the part of the iclog that is
1942  * written to the start of the log. Watch out for the header magic
1943  * number case, though.
1944  */
1945 static void
1946 xlog_split_iclog(
1947 	struct xlog		*log,
1948 	void			*data,
1949 	uint64_t		bno,
1950 	unsigned int		count)
1951 {
1952 	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1953 	unsigned int		i;
1954 
1955 	for (i = split_offset; i < count; i += BBSIZE) {
1956 		uint32_t cycle = get_unaligned_be32(data + i);
1957 
1958 		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1959 			cycle++;
1960 		put_unaligned_be32(cycle, data + i);
1961 	}
1962 }
1963 
1964 static int
1965 xlog_calc_iclog_size(
1966 	struct xlog		*log,
1967 	struct xlog_in_core	*iclog,
1968 	uint32_t		*roundoff)
1969 {
1970 	uint32_t		count_init, count;
1971 
1972 	/* Add for LR header */
1973 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1974 	count = roundup(count_init, log->l_iclog_roundoff);
1975 
1976 	*roundoff = count - count_init;
1977 
1978 	ASSERT(count >= count_init);
1979 	ASSERT(*roundoff < log->l_iclog_roundoff);
1980 	return count;
1981 }
1982 
1983 /*
1984  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1985  * fashion.  Previously, we should have moved the current iclog
1986  * ptr in the log to point to the next available iclog.  This allows further
1987  * write to continue while this code syncs out an iclog ready to go.
1988  * Before an in-core log can be written out, the data section must be scanned
1989  * to save away the 1st word of each BBSIZE block into the header.  We replace
1990  * it with the current cycle count.  Each BBSIZE block is tagged with the
1991  * cycle count because there in an implicit assumption that drives will
1992  * guarantee that entire 512 byte blocks get written at once.  In other words,
1993  * we can't have part of a 512 byte block written and part not written.  By
1994  * tagging each block, we will know which blocks are valid when recovering
1995  * after an unclean shutdown.
1996  *
1997  * This routine is single threaded on the iclog.  No other thread can be in
1998  * this routine with the same iclog.  Changing contents of iclog can there-
1999  * fore be done without grabbing the state machine lock.  Updating the global
2000  * log will require grabbing the lock though.
2001  *
2002  * The entire log manager uses a logical block numbering scheme.  Only
2003  * xlog_write_iclog knows about the fact that the log may not start with
2004  * block zero on a given device.
2005  */
2006 STATIC void
2007 xlog_sync(
2008 	struct xlog		*log,
2009 	struct xlog_in_core	*iclog)
2010 {
2011 	unsigned int		count;		/* byte count of bwrite */
2012 	unsigned int		roundoff;       /* roundoff to BB or stripe */
2013 	uint64_t		bno;
2014 	unsigned int		size;
2015 
2016 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2017 	trace_xlog_iclog_sync(iclog, _RET_IP_);
2018 
2019 	count = xlog_calc_iclog_size(log, iclog, &roundoff);
2020 
2021 	/* move grant heads by roundoff in sync */
2022 	xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2023 	xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2024 
2025 	/* put cycle number in every block */
2026 	xlog_pack_data(log, iclog, roundoff);
2027 
2028 	/* real byte length */
2029 	size = iclog->ic_offset;
2030 	if (xfs_has_logv2(log->l_mp))
2031 		size += roundoff;
2032 	iclog->ic_header.h_len = cpu_to_be32(size);
2033 
2034 	XFS_STATS_INC(log->l_mp, xs_log_writes);
2035 	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2036 
2037 	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2038 
2039 	/* Do we need to split this write into 2 parts? */
2040 	if (bno + BTOBB(count) > log->l_logBBsize)
2041 		xlog_split_iclog(log, &iclog->ic_header, bno, count);
2042 
2043 	/* calculcate the checksum */
2044 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2045 					    iclog->ic_datap, size);
2046 	/*
2047 	 * Intentionally corrupt the log record CRC based on the error injection
2048 	 * frequency, if defined. This facilitates testing log recovery in the
2049 	 * event of torn writes. Hence, set the IOABORT state to abort the log
2050 	 * write on I/O completion and shutdown the fs. The subsequent mount
2051 	 * detects the bad CRC and attempts to recover.
2052 	 */
2053 #ifdef DEBUG
2054 	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2055 		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2056 		iclog->ic_fail_crc = true;
2057 		xfs_warn(log->l_mp,
2058 	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2059 			 be64_to_cpu(iclog->ic_header.h_lsn));
2060 	}
2061 #endif
2062 	xlog_verify_iclog(log, iclog, count);
2063 	xlog_write_iclog(log, iclog, bno, count);
2064 }
2065 
2066 /*
2067  * Deallocate a log structure
2068  */
2069 STATIC void
2070 xlog_dealloc_log(
2071 	struct xlog	*log)
2072 {
2073 	xlog_in_core_t	*iclog, *next_iclog;
2074 	int		i;
2075 
2076 	xlog_cil_destroy(log);
2077 
2078 	/*
2079 	 * Cycle all the iclogbuf locks to make sure all log IO completion
2080 	 * is done before we tear down these buffers.
2081 	 */
2082 	iclog = log->l_iclog;
2083 	for (i = 0; i < log->l_iclog_bufs; i++) {
2084 		down(&iclog->ic_sema);
2085 		up(&iclog->ic_sema);
2086 		iclog = iclog->ic_next;
2087 	}
2088 
2089 	iclog = log->l_iclog;
2090 	for (i = 0; i < log->l_iclog_bufs; i++) {
2091 		next_iclog = iclog->ic_next;
2092 		kmem_free(iclog->ic_data);
2093 		kmem_free(iclog);
2094 		iclog = next_iclog;
2095 	}
2096 
2097 	log->l_mp->m_log = NULL;
2098 	destroy_workqueue(log->l_ioend_workqueue);
2099 	kmem_free(log);
2100 }
2101 
2102 /*
2103  * Update counters atomically now that memcpy is done.
2104  */
2105 static inline void
2106 xlog_state_finish_copy(
2107 	struct xlog		*log,
2108 	struct xlog_in_core	*iclog,
2109 	int			record_cnt,
2110 	int			copy_bytes)
2111 {
2112 	lockdep_assert_held(&log->l_icloglock);
2113 
2114 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2115 	iclog->ic_offset += copy_bytes;
2116 }
2117 
2118 /*
2119  * print out info relating to regions written which consume
2120  * the reservation
2121  */
2122 void
2123 xlog_print_tic_res(
2124 	struct xfs_mount	*mp,
2125 	struct xlog_ticket	*ticket)
2126 {
2127 	uint i;
2128 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2129 
2130 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
2131 #define REG_TYPE_STR(type, str)	[XLOG_REG_TYPE_##type] = str
2132 	static char *res_type_str[] = {
2133 	    REG_TYPE_STR(BFORMAT, "bformat"),
2134 	    REG_TYPE_STR(BCHUNK, "bchunk"),
2135 	    REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2136 	    REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2137 	    REG_TYPE_STR(IFORMAT, "iformat"),
2138 	    REG_TYPE_STR(ICORE, "icore"),
2139 	    REG_TYPE_STR(IEXT, "iext"),
2140 	    REG_TYPE_STR(IBROOT, "ibroot"),
2141 	    REG_TYPE_STR(ILOCAL, "ilocal"),
2142 	    REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2143 	    REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2144 	    REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2145 	    REG_TYPE_STR(QFORMAT, "qformat"),
2146 	    REG_TYPE_STR(DQUOT, "dquot"),
2147 	    REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2148 	    REG_TYPE_STR(LRHEADER, "LR header"),
2149 	    REG_TYPE_STR(UNMOUNT, "unmount"),
2150 	    REG_TYPE_STR(COMMIT, "commit"),
2151 	    REG_TYPE_STR(TRANSHDR, "trans header"),
2152 	    REG_TYPE_STR(ICREATE, "inode create"),
2153 	    REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2154 	    REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2155 	    REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2156 	    REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2157 	    REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2158 	    REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2159 	};
2160 	BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2161 #undef REG_TYPE_STR
2162 
2163 	xfs_warn(mp, "ticket reservation summary:");
2164 	xfs_warn(mp, "  unit res    = %d bytes",
2165 		 ticket->t_unit_res);
2166 	xfs_warn(mp, "  current res = %d bytes",
2167 		 ticket->t_curr_res);
2168 	xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2169 		 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2170 	xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2171 		 ticket->t_res_num_ophdrs, ophdr_spc);
2172 	xfs_warn(mp, "  ophdr + reg = %u bytes",
2173 		 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2174 	xfs_warn(mp, "  num regions = %u",
2175 		 ticket->t_res_num);
2176 
2177 	for (i = 0; i < ticket->t_res_num; i++) {
2178 		uint r_type = ticket->t_res_arr[i].r_type;
2179 		xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2180 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2181 			    "bad-rtype" : res_type_str[r_type]),
2182 			    ticket->t_res_arr[i].r_len);
2183 	}
2184 }
2185 
2186 /*
2187  * Print a summary of the transaction.
2188  */
2189 void
2190 xlog_print_trans(
2191 	struct xfs_trans	*tp)
2192 {
2193 	struct xfs_mount	*mp = tp->t_mountp;
2194 	struct xfs_log_item	*lip;
2195 
2196 	/* dump core transaction and ticket info */
2197 	xfs_warn(mp, "transaction summary:");
2198 	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2199 	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2200 	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2201 
2202 	xlog_print_tic_res(mp, tp->t_ticket);
2203 
2204 	/* dump each log item */
2205 	list_for_each_entry(lip, &tp->t_items, li_trans) {
2206 		struct xfs_log_vec	*lv = lip->li_lv;
2207 		struct xfs_log_iovec	*vec;
2208 		int			i;
2209 
2210 		xfs_warn(mp, "log item: ");
2211 		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2212 		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
2213 		if (!lv)
2214 			continue;
2215 		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2216 		xfs_warn(mp, "  size	= %d", lv->lv_size);
2217 		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2218 		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2219 
2220 		/* dump each iovec for the log item */
2221 		vec = lv->lv_iovecp;
2222 		for (i = 0; i < lv->lv_niovecs; i++) {
2223 			int dumplen = min(vec->i_len, 32);
2224 
2225 			xfs_warn(mp, "  iovec[%d]", i);
2226 			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2227 			xfs_warn(mp, "    len	= %d", vec->i_len);
2228 			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2229 			xfs_hex_dump(vec->i_addr, dumplen);
2230 
2231 			vec++;
2232 		}
2233 	}
2234 }
2235 
2236 /*
2237  * Calculate the potential space needed by the log vector.  We may need a start
2238  * record, and each region gets its own struct xlog_op_header and may need to be
2239  * double word aligned.
2240  */
2241 static int
2242 xlog_write_calc_vec_length(
2243 	struct xlog_ticket	*ticket,
2244 	struct xfs_log_vec	*log_vector,
2245 	uint			optype)
2246 {
2247 	struct xfs_log_vec	*lv;
2248 	int			headers = 0;
2249 	int			len = 0;
2250 	int			i;
2251 
2252 	if (optype & XLOG_START_TRANS)
2253 		headers++;
2254 
2255 	for (lv = log_vector; lv; lv = lv->lv_next) {
2256 		/* we don't write ordered log vectors */
2257 		if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2258 			continue;
2259 
2260 		headers += lv->lv_niovecs;
2261 
2262 		for (i = 0; i < lv->lv_niovecs; i++) {
2263 			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
2264 
2265 			len += vecp->i_len;
2266 			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2267 		}
2268 	}
2269 
2270 	ticket->t_res_num_ophdrs += headers;
2271 	len += headers * sizeof(struct xlog_op_header);
2272 
2273 	return len;
2274 }
2275 
2276 static void
2277 xlog_write_start_rec(
2278 	struct xlog_op_header	*ophdr,
2279 	struct xlog_ticket	*ticket)
2280 {
2281 	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
2282 	ophdr->oh_clientid = ticket->t_clientid;
2283 	ophdr->oh_len = 0;
2284 	ophdr->oh_flags = XLOG_START_TRANS;
2285 	ophdr->oh_res2 = 0;
2286 }
2287 
2288 static xlog_op_header_t *
2289 xlog_write_setup_ophdr(
2290 	struct xlog		*log,
2291 	struct xlog_op_header	*ophdr,
2292 	struct xlog_ticket	*ticket,
2293 	uint			flags)
2294 {
2295 	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2296 	ophdr->oh_clientid = ticket->t_clientid;
2297 	ophdr->oh_res2 = 0;
2298 
2299 	/* are we copying a commit or unmount record? */
2300 	ophdr->oh_flags = flags;
2301 
2302 	/*
2303 	 * We've seen logs corrupted with bad transaction client ids.  This
2304 	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2305 	 * and shut down the filesystem.
2306 	 */
2307 	switch (ophdr->oh_clientid)  {
2308 	case XFS_TRANSACTION:
2309 	case XFS_VOLUME:
2310 	case XFS_LOG:
2311 		break;
2312 	default:
2313 		xfs_warn(log->l_mp,
2314 			"Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2315 			ophdr->oh_clientid, ticket);
2316 		return NULL;
2317 	}
2318 
2319 	return ophdr;
2320 }
2321 
2322 /*
2323  * Set up the parameters of the region copy into the log. This has
2324  * to handle region write split across multiple log buffers - this
2325  * state is kept external to this function so that this code can
2326  * be written in an obvious, self documenting manner.
2327  */
2328 static int
2329 xlog_write_setup_copy(
2330 	struct xlog_ticket	*ticket,
2331 	struct xlog_op_header	*ophdr,
2332 	int			space_available,
2333 	int			space_required,
2334 	int			*copy_off,
2335 	int			*copy_len,
2336 	int			*last_was_partial_copy,
2337 	int			*bytes_consumed)
2338 {
2339 	int			still_to_copy;
2340 
2341 	still_to_copy = space_required - *bytes_consumed;
2342 	*copy_off = *bytes_consumed;
2343 
2344 	if (still_to_copy <= space_available) {
2345 		/* write of region completes here */
2346 		*copy_len = still_to_copy;
2347 		ophdr->oh_len = cpu_to_be32(*copy_len);
2348 		if (*last_was_partial_copy)
2349 			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2350 		*last_was_partial_copy = 0;
2351 		*bytes_consumed = 0;
2352 		return 0;
2353 	}
2354 
2355 	/* partial write of region, needs extra log op header reservation */
2356 	*copy_len = space_available;
2357 	ophdr->oh_len = cpu_to_be32(*copy_len);
2358 	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2359 	if (*last_was_partial_copy)
2360 		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2361 	*bytes_consumed += *copy_len;
2362 	(*last_was_partial_copy)++;
2363 
2364 	/* account for new log op header */
2365 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2366 	ticket->t_res_num_ophdrs++;
2367 
2368 	return sizeof(struct xlog_op_header);
2369 }
2370 
2371 static int
2372 xlog_write_copy_finish(
2373 	struct xlog		*log,
2374 	struct xlog_in_core	*iclog,
2375 	uint			flags,
2376 	int			*record_cnt,
2377 	int			*data_cnt,
2378 	int			*partial_copy,
2379 	int			*partial_copy_len,
2380 	int			log_offset)
2381 {
2382 	int			error;
2383 
2384 	if (*partial_copy) {
2385 		/*
2386 		 * This iclog has already been marked WANT_SYNC by
2387 		 * xlog_state_get_iclog_space.
2388 		 */
2389 		spin_lock(&log->l_icloglock);
2390 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2391 		*record_cnt = 0;
2392 		*data_cnt = 0;
2393 		goto release_iclog;
2394 	}
2395 
2396 	*partial_copy = 0;
2397 	*partial_copy_len = 0;
2398 
2399 	if (iclog->ic_size - log_offset > sizeof(xlog_op_header_t))
2400 		return 0;
2401 
2402 	/* no more space in this iclog - push it. */
2403 	spin_lock(&log->l_icloglock);
2404 	xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2405 	*record_cnt = 0;
2406 	*data_cnt = 0;
2407 
2408 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
2409 		xlog_state_switch_iclogs(log, iclog, 0);
2410 	else
2411 		ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2412 			xlog_is_shutdown(log));
2413 release_iclog:
2414 	error = xlog_state_release_iclog(log, iclog, 0);
2415 	spin_unlock(&log->l_icloglock);
2416 	return error;
2417 }
2418 
2419 /*
2420  * Write some region out to in-core log
2421  *
2422  * This will be called when writing externally provided regions or when
2423  * writing out a commit record for a given transaction.
2424  *
2425  * General algorithm:
2426  *	1. Find total length of this write.  This may include adding to the
2427  *		lengths passed in.
2428  *	2. Check whether we violate the tickets reservation.
2429  *	3. While writing to this iclog
2430  *	    A. Reserve as much space in this iclog as can get
2431  *	    B. If this is first write, save away start lsn
2432  *	    C. While writing this region:
2433  *		1. If first write of transaction, write start record
2434  *		2. Write log operation header (header per region)
2435  *		3. Find out if we can fit entire region into this iclog
2436  *		4. Potentially, verify destination memcpy ptr
2437  *		5. Memcpy (partial) region
2438  *		6. If partial copy, release iclog; otherwise, continue
2439  *			copying more regions into current iclog
2440  *	4. Mark want sync bit (in simulation mode)
2441  *	5. Release iclog for potential flush to on-disk log.
2442  *
2443  * ERRORS:
2444  * 1.	Panic if reservation is overrun.  This should never happen since
2445  *	reservation amounts are generated internal to the filesystem.
2446  * NOTES:
2447  * 1. Tickets are single threaded data structures.
2448  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2449  *	syncing routine.  When a single log_write region needs to span
2450  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2451  *	on all log operation writes which don't contain the end of the
2452  *	region.  The XLOG_END_TRANS bit is used for the in-core log
2453  *	operation which contains the end of the continued log_write region.
2454  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2455  *	we don't really know exactly how much space will be used.  As a result,
2456  *	we don't update ic_offset until the end when we know exactly how many
2457  *	bytes have been written out.
2458  */
2459 int
2460 xlog_write(
2461 	struct xlog		*log,
2462 	struct xfs_cil_ctx	*ctx,
2463 	struct xfs_log_vec	*log_vector,
2464 	struct xlog_ticket	*ticket,
2465 	uint			optype)
2466 {
2467 	struct xlog_in_core	*iclog = NULL;
2468 	struct xfs_log_vec	*lv = log_vector;
2469 	struct xfs_log_iovec	*vecp = lv->lv_iovecp;
2470 	int			index = 0;
2471 	int			len;
2472 	int			partial_copy = 0;
2473 	int			partial_copy_len = 0;
2474 	int			contwr = 0;
2475 	int			record_cnt = 0;
2476 	int			data_cnt = 0;
2477 	int			error = 0;
2478 
2479 	/*
2480 	 * If this is a commit or unmount transaction, we don't need a start
2481 	 * record to be written.  We do, however, have to account for the
2482 	 * commit or unmount header that gets written. Hence we always have
2483 	 * to account for an extra xlog_op_header here.
2484 	 */
2485 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2486 	if (ticket->t_curr_res < 0) {
2487 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2488 		     "ctx ticket reservation ran out. Need to up reservation");
2489 		xlog_print_tic_res(log->l_mp, ticket);
2490 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2491 	}
2492 
2493 	len = xlog_write_calc_vec_length(ticket, log_vector, optype);
2494 	while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2495 		void		*ptr;
2496 		int		log_offset;
2497 
2498 		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2499 						   &contwr, &log_offset);
2500 		if (error)
2501 			return error;
2502 
2503 		ASSERT(log_offset <= iclog->ic_size - 1);
2504 		ptr = iclog->ic_datap + log_offset;
2505 
2506 		/*
2507 		 * If we have a context pointer, pass it the first iclog we are
2508 		 * writing to so it can record state needed for iclog write
2509 		 * ordering.
2510 		 */
2511 		if (ctx) {
2512 			xlog_cil_set_ctx_write_state(ctx, iclog);
2513 			ctx = NULL;
2514 		}
2515 
2516 		/*
2517 		 * This loop writes out as many regions as can fit in the amount
2518 		 * of space which was allocated by xlog_state_get_iclog_space().
2519 		 */
2520 		while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2521 			struct xfs_log_iovec	*reg;
2522 			struct xlog_op_header	*ophdr;
2523 			int			copy_len;
2524 			int			copy_off;
2525 			bool			ordered = false;
2526 			bool			wrote_start_rec = false;
2527 
2528 			/* ordered log vectors have no regions to write */
2529 			if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2530 				ASSERT(lv->lv_niovecs == 0);
2531 				ordered = true;
2532 				goto next_lv;
2533 			}
2534 
2535 			reg = &vecp[index];
2536 			ASSERT(reg->i_len % sizeof(int32_t) == 0);
2537 			ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2538 
2539 			/*
2540 			 * Before we start formatting log vectors, we need to
2541 			 * write a start record. Only do this for the first
2542 			 * iclog we write to.
2543 			 */
2544 			if (optype & XLOG_START_TRANS) {
2545 				xlog_write_start_rec(ptr, ticket);
2546 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2547 						sizeof(struct xlog_op_header));
2548 				optype &= ~XLOG_START_TRANS;
2549 				wrote_start_rec = true;
2550 			}
2551 
2552 			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, optype);
2553 			if (!ophdr)
2554 				return -EIO;
2555 
2556 			xlog_write_adv_cnt(&ptr, &len, &log_offset,
2557 					   sizeof(struct xlog_op_header));
2558 
2559 			len += xlog_write_setup_copy(ticket, ophdr,
2560 						     iclog->ic_size-log_offset,
2561 						     reg->i_len,
2562 						     &copy_off, &copy_len,
2563 						     &partial_copy,
2564 						     &partial_copy_len);
2565 			xlog_verify_dest_ptr(log, ptr);
2566 
2567 			/*
2568 			 * Copy region.
2569 			 *
2570 			 * Unmount records just log an opheader, so can have
2571 			 * empty payloads with no data region to copy. Hence we
2572 			 * only copy the payload if the vector says it has data
2573 			 * to copy.
2574 			 */
2575 			ASSERT(copy_len >= 0);
2576 			if (copy_len > 0) {
2577 				memcpy(ptr, reg->i_addr + copy_off, copy_len);
2578 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2579 						   copy_len);
2580 			}
2581 			copy_len += sizeof(struct xlog_op_header);
2582 			record_cnt++;
2583 			if (wrote_start_rec) {
2584 				copy_len += sizeof(struct xlog_op_header);
2585 				record_cnt++;
2586 			}
2587 			data_cnt += contwr ? copy_len : 0;
2588 
2589 			error = xlog_write_copy_finish(log, iclog, optype,
2590 						       &record_cnt, &data_cnt,
2591 						       &partial_copy,
2592 						       &partial_copy_len,
2593 						       log_offset);
2594 			if (error)
2595 				return error;
2596 
2597 			/*
2598 			 * if we had a partial copy, we need to get more iclog
2599 			 * space but we don't want to increment the region
2600 			 * index because there is still more is this region to
2601 			 * write.
2602 			 *
2603 			 * If we completed writing this region, and we flushed
2604 			 * the iclog (indicated by resetting of the record
2605 			 * count), then we also need to get more log space. If
2606 			 * this was the last record, though, we are done and
2607 			 * can just return.
2608 			 */
2609 			if (partial_copy)
2610 				break;
2611 
2612 			if (++index == lv->lv_niovecs) {
2613 next_lv:
2614 				lv = lv->lv_next;
2615 				index = 0;
2616 				if (lv)
2617 					vecp = lv->lv_iovecp;
2618 			}
2619 			if (record_cnt == 0 && !ordered) {
2620 				if (!lv)
2621 					return 0;
2622 				break;
2623 			}
2624 		}
2625 	}
2626 
2627 	ASSERT(len == 0);
2628 
2629 	spin_lock(&log->l_icloglock);
2630 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2631 	error = xlog_state_release_iclog(log, iclog, 0);
2632 	spin_unlock(&log->l_icloglock);
2633 
2634 	return error;
2635 }
2636 
2637 static void
2638 xlog_state_activate_iclog(
2639 	struct xlog_in_core	*iclog,
2640 	int			*iclogs_changed)
2641 {
2642 	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2643 	trace_xlog_iclog_activate(iclog, _RET_IP_);
2644 
2645 	/*
2646 	 * If the number of ops in this iclog indicate it just contains the
2647 	 * dummy transaction, we can change state into IDLE (the second time
2648 	 * around). Otherwise we should change the state into NEED a dummy.
2649 	 * We don't need to cover the dummy.
2650 	 */
2651 	if (*iclogs_changed == 0 &&
2652 	    iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2653 		*iclogs_changed = 1;
2654 	} else {
2655 		/*
2656 		 * We have two dirty iclogs so start over.  This could also be
2657 		 * num of ops indicating this is not the dummy going out.
2658 		 */
2659 		*iclogs_changed = 2;
2660 	}
2661 
2662 	iclog->ic_state	= XLOG_STATE_ACTIVE;
2663 	iclog->ic_offset = 0;
2664 	iclog->ic_header.h_num_logops = 0;
2665 	memset(iclog->ic_header.h_cycle_data, 0,
2666 		sizeof(iclog->ic_header.h_cycle_data));
2667 	iclog->ic_header.h_lsn = 0;
2668 	iclog->ic_header.h_tail_lsn = 0;
2669 }
2670 
2671 /*
2672  * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2673  * ACTIVE after iclog I/O has completed.
2674  */
2675 static void
2676 xlog_state_activate_iclogs(
2677 	struct xlog		*log,
2678 	int			*iclogs_changed)
2679 {
2680 	struct xlog_in_core	*iclog = log->l_iclog;
2681 
2682 	do {
2683 		if (iclog->ic_state == XLOG_STATE_DIRTY)
2684 			xlog_state_activate_iclog(iclog, iclogs_changed);
2685 		/*
2686 		 * The ordering of marking iclogs ACTIVE must be maintained, so
2687 		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2688 		 */
2689 		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2690 			break;
2691 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2692 }
2693 
2694 static int
2695 xlog_covered_state(
2696 	int			prev_state,
2697 	int			iclogs_changed)
2698 {
2699 	/*
2700 	 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2701 	 * wrote the first covering record (DONE). We go to IDLE if we just
2702 	 * wrote the second covering record (DONE2) and remain in IDLE until a
2703 	 * non-covering write occurs.
2704 	 */
2705 	switch (prev_state) {
2706 	case XLOG_STATE_COVER_IDLE:
2707 		if (iclogs_changed == 1)
2708 			return XLOG_STATE_COVER_IDLE;
2709 		fallthrough;
2710 	case XLOG_STATE_COVER_NEED:
2711 	case XLOG_STATE_COVER_NEED2:
2712 		break;
2713 	case XLOG_STATE_COVER_DONE:
2714 		if (iclogs_changed == 1)
2715 			return XLOG_STATE_COVER_NEED2;
2716 		break;
2717 	case XLOG_STATE_COVER_DONE2:
2718 		if (iclogs_changed == 1)
2719 			return XLOG_STATE_COVER_IDLE;
2720 		break;
2721 	default:
2722 		ASSERT(0);
2723 	}
2724 
2725 	return XLOG_STATE_COVER_NEED;
2726 }
2727 
2728 STATIC void
2729 xlog_state_clean_iclog(
2730 	struct xlog		*log,
2731 	struct xlog_in_core	*dirty_iclog)
2732 {
2733 	int			iclogs_changed = 0;
2734 
2735 	trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2736 
2737 	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2738 
2739 	xlog_state_activate_iclogs(log, &iclogs_changed);
2740 	wake_up_all(&dirty_iclog->ic_force_wait);
2741 
2742 	if (iclogs_changed) {
2743 		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2744 				iclogs_changed);
2745 	}
2746 }
2747 
2748 STATIC xfs_lsn_t
2749 xlog_get_lowest_lsn(
2750 	struct xlog		*log)
2751 {
2752 	struct xlog_in_core	*iclog = log->l_iclog;
2753 	xfs_lsn_t		lowest_lsn = 0, lsn;
2754 
2755 	do {
2756 		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2757 		    iclog->ic_state == XLOG_STATE_DIRTY)
2758 			continue;
2759 
2760 		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2761 		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2762 			lowest_lsn = lsn;
2763 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2764 
2765 	return lowest_lsn;
2766 }
2767 
2768 /*
2769  * Completion of a iclog IO does not imply that a transaction has completed, as
2770  * transactions can be large enough to span many iclogs. We cannot change the
2771  * tail of the log half way through a transaction as this may be the only
2772  * transaction in the log and moving the tail to point to the middle of it
2773  * will prevent recovery from finding the start of the transaction. Hence we
2774  * should only update the last_sync_lsn if this iclog contains transaction
2775  * completion callbacks on it.
2776  *
2777  * We have to do this before we drop the icloglock to ensure we are the only one
2778  * that can update it.
2779  *
2780  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2781  * the reservation grant head pushing. This is due to the fact that the push
2782  * target is bound by the current last_sync_lsn value. Hence if we have a large
2783  * amount of log space bound up in this committing transaction then the
2784  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2785  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2786  * should push the AIL to ensure the push target (and hence the grant head) is
2787  * no longer bound by the old log head location and can move forwards and make
2788  * progress again.
2789  */
2790 static void
2791 xlog_state_set_callback(
2792 	struct xlog		*log,
2793 	struct xlog_in_core	*iclog,
2794 	xfs_lsn_t		header_lsn)
2795 {
2796 	trace_xlog_iclog_callback(iclog, _RET_IP_);
2797 	iclog->ic_state = XLOG_STATE_CALLBACK;
2798 
2799 	ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2800 			   header_lsn) <= 0);
2801 
2802 	if (list_empty_careful(&iclog->ic_callbacks))
2803 		return;
2804 
2805 	atomic64_set(&log->l_last_sync_lsn, header_lsn);
2806 	xlog_grant_push_ail(log, 0);
2807 }
2808 
2809 /*
2810  * Return true if we need to stop processing, false to continue to the next
2811  * iclog. The caller will need to run callbacks if the iclog is returned in the
2812  * XLOG_STATE_CALLBACK state.
2813  */
2814 static bool
2815 xlog_state_iodone_process_iclog(
2816 	struct xlog		*log,
2817 	struct xlog_in_core	*iclog)
2818 {
2819 	xfs_lsn_t		lowest_lsn;
2820 	xfs_lsn_t		header_lsn;
2821 
2822 	switch (iclog->ic_state) {
2823 	case XLOG_STATE_ACTIVE:
2824 	case XLOG_STATE_DIRTY:
2825 		/*
2826 		 * Skip all iclogs in the ACTIVE & DIRTY states:
2827 		 */
2828 		return false;
2829 	case XLOG_STATE_DONE_SYNC:
2830 		/*
2831 		 * Now that we have an iclog that is in the DONE_SYNC state, do
2832 		 * one more check here to see if we have chased our tail around.
2833 		 * If this is not the lowest lsn iclog, then we will leave it
2834 		 * for another completion to process.
2835 		 */
2836 		header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2837 		lowest_lsn = xlog_get_lowest_lsn(log);
2838 		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2839 			return false;
2840 		xlog_state_set_callback(log, iclog, header_lsn);
2841 		return false;
2842 	default:
2843 		/*
2844 		 * Can only perform callbacks in order.  Since this iclog is not
2845 		 * in the DONE_SYNC state, we skip the rest and just try to
2846 		 * clean up.
2847 		 */
2848 		return true;
2849 	}
2850 }
2851 
2852 /*
2853  * Loop over all the iclogs, running attached callbacks on them. Return true if
2854  * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2855  * to handle transient shutdown state here at all because
2856  * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2857  * cleanup of the callbacks.
2858  */
2859 static bool
2860 xlog_state_do_iclog_callbacks(
2861 	struct xlog		*log)
2862 		__releases(&log->l_icloglock)
2863 		__acquires(&log->l_icloglock)
2864 {
2865 	struct xlog_in_core	*first_iclog = log->l_iclog;
2866 	struct xlog_in_core	*iclog = first_iclog;
2867 	bool			ran_callback = false;
2868 
2869 	do {
2870 		LIST_HEAD(cb_list);
2871 
2872 		if (xlog_state_iodone_process_iclog(log, iclog))
2873 			break;
2874 		if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2875 			iclog = iclog->ic_next;
2876 			continue;
2877 		}
2878 		list_splice_init(&iclog->ic_callbacks, &cb_list);
2879 		spin_unlock(&log->l_icloglock);
2880 
2881 		trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2882 		xlog_cil_process_committed(&cb_list);
2883 		trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2884 		ran_callback = true;
2885 
2886 		spin_lock(&log->l_icloglock);
2887 		xlog_state_clean_iclog(log, iclog);
2888 		iclog = iclog->ic_next;
2889 	} while (iclog != first_iclog);
2890 
2891 	return ran_callback;
2892 }
2893 
2894 
2895 /*
2896  * Loop running iclog completion callbacks until there are no more iclogs in a
2897  * state that can run callbacks.
2898  */
2899 STATIC void
2900 xlog_state_do_callback(
2901 	struct xlog		*log)
2902 {
2903 	int			flushcnt = 0;
2904 	int			repeats = 0;
2905 
2906 	spin_lock(&log->l_icloglock);
2907 	while (xlog_state_do_iclog_callbacks(log)) {
2908 		if (xlog_is_shutdown(log))
2909 			break;
2910 
2911 		if (++repeats > 5000) {
2912 			flushcnt += repeats;
2913 			repeats = 0;
2914 			xfs_warn(log->l_mp,
2915 				"%s: possible infinite loop (%d iterations)",
2916 				__func__, flushcnt);
2917 		}
2918 	}
2919 
2920 	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2921 		wake_up_all(&log->l_flush_wait);
2922 
2923 	spin_unlock(&log->l_icloglock);
2924 }
2925 
2926 
2927 /*
2928  * Finish transitioning this iclog to the dirty state.
2929  *
2930  * Callbacks could take time, so they are done outside the scope of the
2931  * global state machine log lock.
2932  */
2933 STATIC void
2934 xlog_state_done_syncing(
2935 	struct xlog_in_core	*iclog)
2936 {
2937 	struct xlog		*log = iclog->ic_log;
2938 
2939 	spin_lock(&log->l_icloglock);
2940 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2941 	trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2942 
2943 	/*
2944 	 * If we got an error, either on the first buffer, or in the case of
2945 	 * split log writes, on the second, we shut down the file system and
2946 	 * no iclogs should ever be attempted to be written to disk again.
2947 	 */
2948 	if (!xlog_is_shutdown(log)) {
2949 		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2950 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2951 	}
2952 
2953 	/*
2954 	 * Someone could be sleeping prior to writing out the next
2955 	 * iclog buffer, we wake them all, one will get to do the
2956 	 * I/O, the others get to wait for the result.
2957 	 */
2958 	wake_up_all(&iclog->ic_write_wait);
2959 	spin_unlock(&log->l_icloglock);
2960 	xlog_state_do_callback(log);
2961 }
2962 
2963 /*
2964  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2965  * sleep.  We wait on the flush queue on the head iclog as that should be
2966  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2967  * we will wait here and all new writes will sleep until a sync completes.
2968  *
2969  * The in-core logs are used in a circular fashion. They are not used
2970  * out-of-order even when an iclog past the head is free.
2971  *
2972  * return:
2973  *	* log_offset where xlog_write() can start writing into the in-core
2974  *		log's data space.
2975  *	* in-core log pointer to which xlog_write() should write.
2976  *	* boolean indicating this is a continued write to an in-core log.
2977  *		If this is the last write, then the in-core log's offset field
2978  *		needs to be incremented, depending on the amount of data which
2979  *		is copied.
2980  */
2981 STATIC int
2982 xlog_state_get_iclog_space(
2983 	struct xlog		*log,
2984 	int			len,
2985 	struct xlog_in_core	**iclogp,
2986 	struct xlog_ticket	*ticket,
2987 	int			*continued_write,
2988 	int			*logoffsetp)
2989 {
2990 	int		  log_offset;
2991 	xlog_rec_header_t *head;
2992 	xlog_in_core_t	  *iclog;
2993 
2994 restart:
2995 	spin_lock(&log->l_icloglock);
2996 	if (xlog_is_shutdown(log)) {
2997 		spin_unlock(&log->l_icloglock);
2998 		return -EIO;
2999 	}
3000 
3001 	iclog = log->l_iclog;
3002 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3003 		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3004 
3005 		/* Wait for log writes to have flushed */
3006 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3007 		goto restart;
3008 	}
3009 
3010 	head = &iclog->ic_header;
3011 
3012 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
3013 	log_offset = iclog->ic_offset;
3014 
3015 	trace_xlog_iclog_get_space(iclog, _RET_IP_);
3016 
3017 	/* On the 1st write to an iclog, figure out lsn.  This works
3018 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3019 	 * committing to.  If the offset is set, that's how many blocks
3020 	 * must be written.
3021 	 */
3022 	if (log_offset == 0) {
3023 		ticket->t_curr_res -= log->l_iclog_hsize;
3024 		xlog_tic_add_region(ticket,
3025 				    log->l_iclog_hsize,
3026 				    XLOG_REG_TYPE_LRHEADER);
3027 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3028 		head->h_lsn = cpu_to_be64(
3029 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3030 		ASSERT(log->l_curr_block >= 0);
3031 	}
3032 
3033 	/* If there is enough room to write everything, then do it.  Otherwise,
3034 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3035 	 * bit is on, so this will get flushed out.  Don't update ic_offset
3036 	 * until you know exactly how many bytes get copied.  Therefore, wait
3037 	 * until later to update ic_offset.
3038 	 *
3039 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3040 	 * can fit into remaining data section.
3041 	 */
3042 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3043 		int		error = 0;
3044 
3045 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3046 
3047 		/*
3048 		 * If we are the only one writing to this iclog, sync it to
3049 		 * disk.  We need to do an atomic compare and decrement here to
3050 		 * avoid racing with concurrent atomic_dec_and_lock() calls in
3051 		 * xlog_state_release_iclog() when there is more than one
3052 		 * reference to the iclog.
3053 		 */
3054 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3055 			error = xlog_state_release_iclog(log, iclog, 0);
3056 		spin_unlock(&log->l_icloglock);
3057 		if (error)
3058 			return error;
3059 		goto restart;
3060 	}
3061 
3062 	/* Do we have enough room to write the full amount in the remainder
3063 	 * of this iclog?  Or must we continue a write on the next iclog and
3064 	 * mark this iclog as completely taken?  In the case where we switch
3065 	 * iclogs (to mark it taken), this particular iclog will release/sync
3066 	 * to disk in xlog_write().
3067 	 */
3068 	if (len <= iclog->ic_size - iclog->ic_offset) {
3069 		*continued_write = 0;
3070 		iclog->ic_offset += len;
3071 	} else {
3072 		*continued_write = 1;
3073 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3074 	}
3075 	*iclogp = iclog;
3076 
3077 	ASSERT(iclog->ic_offset <= iclog->ic_size);
3078 	spin_unlock(&log->l_icloglock);
3079 
3080 	*logoffsetp = log_offset;
3081 	return 0;
3082 }
3083 
3084 /*
3085  * The first cnt-1 times a ticket goes through here we don't need to move the
3086  * grant write head because the permanent reservation has reserved cnt times the
3087  * unit amount.  Release part of current permanent unit reservation and reset
3088  * current reservation to be one units worth.  Also move grant reservation head
3089  * forward.
3090  */
3091 void
3092 xfs_log_ticket_regrant(
3093 	struct xlog		*log,
3094 	struct xlog_ticket	*ticket)
3095 {
3096 	trace_xfs_log_ticket_regrant(log, ticket);
3097 
3098 	if (ticket->t_cnt > 0)
3099 		ticket->t_cnt--;
3100 
3101 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3102 					ticket->t_curr_res);
3103 	xlog_grant_sub_space(log, &log->l_write_head.grant,
3104 					ticket->t_curr_res);
3105 	ticket->t_curr_res = ticket->t_unit_res;
3106 	xlog_tic_reset_res(ticket);
3107 
3108 	trace_xfs_log_ticket_regrant_sub(log, ticket);
3109 
3110 	/* just return if we still have some of the pre-reserved space */
3111 	if (!ticket->t_cnt) {
3112 		xlog_grant_add_space(log, &log->l_reserve_head.grant,
3113 				     ticket->t_unit_res);
3114 		trace_xfs_log_ticket_regrant_exit(log, ticket);
3115 
3116 		ticket->t_curr_res = ticket->t_unit_res;
3117 		xlog_tic_reset_res(ticket);
3118 	}
3119 
3120 	xfs_log_ticket_put(ticket);
3121 }
3122 
3123 /*
3124  * Give back the space left from a reservation.
3125  *
3126  * All the information we need to make a correct determination of space left
3127  * is present.  For non-permanent reservations, things are quite easy.  The
3128  * count should have been decremented to zero.  We only need to deal with the
3129  * space remaining in the current reservation part of the ticket.  If the
3130  * ticket contains a permanent reservation, there may be left over space which
3131  * needs to be released.  A count of N means that N-1 refills of the current
3132  * reservation can be done before we need to ask for more space.  The first
3133  * one goes to fill up the first current reservation.  Once we run out of
3134  * space, the count will stay at zero and the only space remaining will be
3135  * in the current reservation field.
3136  */
3137 void
3138 xfs_log_ticket_ungrant(
3139 	struct xlog		*log,
3140 	struct xlog_ticket	*ticket)
3141 {
3142 	int			bytes;
3143 
3144 	trace_xfs_log_ticket_ungrant(log, ticket);
3145 
3146 	if (ticket->t_cnt > 0)
3147 		ticket->t_cnt--;
3148 
3149 	trace_xfs_log_ticket_ungrant_sub(log, ticket);
3150 
3151 	/*
3152 	 * If this is a permanent reservation ticket, we may be able to free
3153 	 * up more space based on the remaining count.
3154 	 */
3155 	bytes = ticket->t_curr_res;
3156 	if (ticket->t_cnt > 0) {
3157 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3158 		bytes += ticket->t_unit_res*ticket->t_cnt;
3159 	}
3160 
3161 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3162 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3163 
3164 	trace_xfs_log_ticket_ungrant_exit(log, ticket);
3165 
3166 	xfs_log_space_wake(log->l_mp);
3167 	xfs_log_ticket_put(ticket);
3168 }
3169 
3170 /*
3171  * This routine will mark the current iclog in the ring as WANT_SYNC and move
3172  * the current iclog pointer to the next iclog in the ring.
3173  */
3174 void
3175 xlog_state_switch_iclogs(
3176 	struct xlog		*log,
3177 	struct xlog_in_core	*iclog,
3178 	int			eventual_size)
3179 {
3180 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3181 	assert_spin_locked(&log->l_icloglock);
3182 	trace_xlog_iclog_switch(iclog, _RET_IP_);
3183 
3184 	if (!eventual_size)
3185 		eventual_size = iclog->ic_offset;
3186 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3187 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3188 	log->l_prev_block = log->l_curr_block;
3189 	log->l_prev_cycle = log->l_curr_cycle;
3190 
3191 	/* roll log?: ic_offset changed later */
3192 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3193 
3194 	/* Round up to next log-sunit */
3195 	if (log->l_iclog_roundoff > BBSIZE) {
3196 		uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3197 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3198 	}
3199 
3200 	if (log->l_curr_block >= log->l_logBBsize) {
3201 		/*
3202 		 * Rewind the current block before the cycle is bumped to make
3203 		 * sure that the combined LSN never transiently moves forward
3204 		 * when the log wraps to the next cycle. This is to support the
3205 		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3206 		 * other cases should acquire l_icloglock.
3207 		 */
3208 		log->l_curr_block -= log->l_logBBsize;
3209 		ASSERT(log->l_curr_block >= 0);
3210 		smp_wmb();
3211 		log->l_curr_cycle++;
3212 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3213 			log->l_curr_cycle++;
3214 	}
3215 	ASSERT(iclog == log->l_iclog);
3216 	log->l_iclog = iclog->ic_next;
3217 }
3218 
3219 /*
3220  * Force the iclog to disk and check if the iclog has been completed before
3221  * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3222  * pmem) or fast async storage because we drop the icloglock to issue the IO.
3223  * If completion has already occurred, tell the caller so that it can avoid an
3224  * unnecessary wait on the iclog.
3225  */
3226 static int
3227 xlog_force_and_check_iclog(
3228 	struct xlog_in_core	*iclog,
3229 	bool			*completed)
3230 {
3231 	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3232 	int			error;
3233 
3234 	*completed = false;
3235 	error = xlog_force_iclog(iclog);
3236 	if (error)
3237 		return error;
3238 
3239 	/*
3240 	 * If the iclog has already been completed and reused the header LSN
3241 	 * will have been rewritten by completion
3242 	 */
3243 	if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3244 		*completed = true;
3245 	return 0;
3246 }
3247 
3248 /*
3249  * Write out all data in the in-core log as of this exact moment in time.
3250  *
3251  * Data may be written to the in-core log during this call.  However,
3252  * we don't guarantee this data will be written out.  A change from past
3253  * implementation means this routine will *not* write out zero length LRs.
3254  *
3255  * Basically, we try and perform an intelligent scan of the in-core logs.
3256  * If we determine there is no flushable data, we just return.  There is no
3257  * flushable data if:
3258  *
3259  *	1. the current iclog is active and has no data; the previous iclog
3260  *		is in the active or dirty state.
3261  *	2. the current iclog is drity, and the previous iclog is in the
3262  *		active or dirty state.
3263  *
3264  * We may sleep if:
3265  *
3266  *	1. the current iclog is not in the active nor dirty state.
3267  *	2. the current iclog dirty, and the previous iclog is not in the
3268  *		active nor dirty state.
3269  *	3. the current iclog is active, and there is another thread writing
3270  *		to this particular iclog.
3271  *	4. a) the current iclog is active and has no other writers
3272  *	   b) when we return from flushing out this iclog, it is still
3273  *		not in the active nor dirty state.
3274  */
3275 int
3276 xfs_log_force(
3277 	struct xfs_mount	*mp,
3278 	uint			flags)
3279 {
3280 	struct xlog		*log = mp->m_log;
3281 	struct xlog_in_core	*iclog;
3282 
3283 	XFS_STATS_INC(mp, xs_log_force);
3284 	trace_xfs_log_force(mp, 0, _RET_IP_);
3285 
3286 	xlog_cil_force(log);
3287 
3288 	spin_lock(&log->l_icloglock);
3289 	if (xlog_is_shutdown(log))
3290 		goto out_error;
3291 
3292 	iclog = log->l_iclog;
3293 	trace_xlog_iclog_force(iclog, _RET_IP_);
3294 
3295 	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3296 	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3297 	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3298 		/*
3299 		 * If the head is dirty or (active and empty), then we need to
3300 		 * look at the previous iclog.
3301 		 *
3302 		 * If the previous iclog is active or dirty we are done.  There
3303 		 * is nothing to sync out. Otherwise, we attach ourselves to the
3304 		 * previous iclog and go to sleep.
3305 		 */
3306 		iclog = iclog->ic_prev;
3307 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3308 		if (atomic_read(&iclog->ic_refcnt) == 0) {
3309 			/* We have exclusive access to this iclog. */
3310 			bool	completed;
3311 
3312 			if (xlog_force_and_check_iclog(iclog, &completed))
3313 				goto out_error;
3314 
3315 			if (completed)
3316 				goto out_unlock;
3317 		} else {
3318 			/*
3319 			 * Someone else is still writing to this iclog, so we
3320 			 * need to ensure that when they release the iclog it
3321 			 * gets synced immediately as we may be waiting on it.
3322 			 */
3323 			xlog_state_switch_iclogs(log, iclog, 0);
3324 		}
3325 	}
3326 
3327 	/*
3328 	 * The iclog we are about to wait on may contain the checkpoint pushed
3329 	 * by the above xlog_cil_force() call, but it may not have been pushed
3330 	 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3331 	 * are flushed when this iclog is written.
3332 	 */
3333 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3334 		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3335 
3336 	if (flags & XFS_LOG_SYNC)
3337 		return xlog_wait_on_iclog(iclog);
3338 out_unlock:
3339 	spin_unlock(&log->l_icloglock);
3340 	return 0;
3341 out_error:
3342 	spin_unlock(&log->l_icloglock);
3343 	return -EIO;
3344 }
3345 
3346 /*
3347  * Force the log to a specific LSN.
3348  *
3349  * If an iclog with that lsn can be found:
3350  *	If it is in the DIRTY state, just return.
3351  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3352  *		state and go to sleep or return.
3353  *	If it is in any other state, go to sleep or return.
3354  *
3355  * Synchronous forces are implemented with a wait queue.  All callers trying
3356  * to force a given lsn to disk must wait on the queue attached to the
3357  * specific in-core log.  When given in-core log finally completes its write
3358  * to disk, that thread will wake up all threads waiting on the queue.
3359  */
3360 static int
3361 xlog_force_lsn(
3362 	struct xlog		*log,
3363 	xfs_lsn_t		lsn,
3364 	uint			flags,
3365 	int			*log_flushed,
3366 	bool			already_slept)
3367 {
3368 	struct xlog_in_core	*iclog;
3369 	bool			completed;
3370 
3371 	spin_lock(&log->l_icloglock);
3372 	if (xlog_is_shutdown(log))
3373 		goto out_error;
3374 
3375 	iclog = log->l_iclog;
3376 	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3377 		trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3378 		iclog = iclog->ic_next;
3379 		if (iclog == log->l_iclog)
3380 			goto out_unlock;
3381 	}
3382 
3383 	switch (iclog->ic_state) {
3384 	case XLOG_STATE_ACTIVE:
3385 		/*
3386 		 * We sleep here if we haven't already slept (e.g. this is the
3387 		 * first time we've looked at the correct iclog buf) and the
3388 		 * buffer before us is going to be sync'ed.  The reason for this
3389 		 * is that if we are doing sync transactions here, by waiting
3390 		 * for the previous I/O to complete, we can allow a few more
3391 		 * transactions into this iclog before we close it down.
3392 		 *
3393 		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3394 		 * refcnt so we can release the log (which drops the ref count).
3395 		 * The state switch keeps new transaction commits from using
3396 		 * this buffer.  When the current commits finish writing into
3397 		 * the buffer, the refcount will drop to zero and the buffer
3398 		 * will go out then.
3399 		 */
3400 		if (!already_slept &&
3401 		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3402 		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3403 			xlog_wait(&iclog->ic_prev->ic_write_wait,
3404 					&log->l_icloglock);
3405 			return -EAGAIN;
3406 		}
3407 		if (xlog_force_and_check_iclog(iclog, &completed))
3408 			goto out_error;
3409 		if (log_flushed)
3410 			*log_flushed = 1;
3411 		if (completed)
3412 			goto out_unlock;
3413 		break;
3414 	case XLOG_STATE_WANT_SYNC:
3415 		/*
3416 		 * This iclog may contain the checkpoint pushed by the
3417 		 * xlog_cil_force_seq() call, but there are other writers still
3418 		 * accessing it so it hasn't been pushed to disk yet. Like the
3419 		 * ACTIVE case above, we need to make sure caches are flushed
3420 		 * when this iclog is written.
3421 		 */
3422 		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3423 		break;
3424 	default:
3425 		/*
3426 		 * The entire checkpoint was written by the CIL force and is on
3427 		 * its way to disk already. It will be stable when it
3428 		 * completes, so we don't need to manipulate caches here at all.
3429 		 * We just need to wait for completion if necessary.
3430 		 */
3431 		break;
3432 	}
3433 
3434 	if (flags & XFS_LOG_SYNC)
3435 		return xlog_wait_on_iclog(iclog);
3436 out_unlock:
3437 	spin_unlock(&log->l_icloglock);
3438 	return 0;
3439 out_error:
3440 	spin_unlock(&log->l_icloglock);
3441 	return -EIO;
3442 }
3443 
3444 /*
3445  * Force the log to a specific checkpoint sequence.
3446  *
3447  * First force the CIL so that all the required changes have been flushed to the
3448  * iclogs. If the CIL force completed it will return a commit LSN that indicates
3449  * the iclog that needs to be flushed to stable storage. If the caller needs
3450  * a synchronous log force, we will wait on the iclog with the LSN returned by
3451  * xlog_cil_force_seq() to be completed.
3452  */
3453 int
3454 xfs_log_force_seq(
3455 	struct xfs_mount	*mp,
3456 	xfs_csn_t		seq,
3457 	uint			flags,
3458 	int			*log_flushed)
3459 {
3460 	struct xlog		*log = mp->m_log;
3461 	xfs_lsn_t		lsn;
3462 	int			ret;
3463 	ASSERT(seq != 0);
3464 
3465 	XFS_STATS_INC(mp, xs_log_force);
3466 	trace_xfs_log_force(mp, seq, _RET_IP_);
3467 
3468 	lsn = xlog_cil_force_seq(log, seq);
3469 	if (lsn == NULLCOMMITLSN)
3470 		return 0;
3471 
3472 	ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3473 	if (ret == -EAGAIN) {
3474 		XFS_STATS_INC(mp, xs_log_force_sleep);
3475 		ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3476 	}
3477 	return ret;
3478 }
3479 
3480 /*
3481  * Free a used ticket when its refcount falls to zero.
3482  */
3483 void
3484 xfs_log_ticket_put(
3485 	xlog_ticket_t	*ticket)
3486 {
3487 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3488 	if (atomic_dec_and_test(&ticket->t_ref))
3489 		kmem_cache_free(xfs_log_ticket_cache, ticket);
3490 }
3491 
3492 xlog_ticket_t *
3493 xfs_log_ticket_get(
3494 	xlog_ticket_t	*ticket)
3495 {
3496 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3497 	atomic_inc(&ticket->t_ref);
3498 	return ticket;
3499 }
3500 
3501 /*
3502  * Figure out the total log space unit (in bytes) that would be
3503  * required for a log ticket.
3504  */
3505 static int
3506 xlog_calc_unit_res(
3507 	struct xlog		*log,
3508 	int			unit_bytes)
3509 {
3510 	int			iclog_space;
3511 	uint			num_headers;
3512 
3513 	/*
3514 	 * Permanent reservations have up to 'cnt'-1 active log operations
3515 	 * in the log.  A unit in this case is the amount of space for one
3516 	 * of these log operations.  Normal reservations have a cnt of 1
3517 	 * and their unit amount is the total amount of space required.
3518 	 *
3519 	 * The following lines of code account for non-transaction data
3520 	 * which occupy space in the on-disk log.
3521 	 *
3522 	 * Normal form of a transaction is:
3523 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3524 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3525 	 *
3526 	 * We need to account for all the leadup data and trailer data
3527 	 * around the transaction data.
3528 	 * And then we need to account for the worst case in terms of using
3529 	 * more space.
3530 	 * The worst case will happen if:
3531 	 * - the placement of the transaction happens to be such that the
3532 	 *   roundoff is at its maximum
3533 	 * - the transaction data is synced before the commit record is synced
3534 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3535 	 *   Therefore the commit record is in its own Log Record.
3536 	 *   This can happen as the commit record is called with its
3537 	 *   own region to xlog_write().
3538 	 *   This then means that in the worst case, roundoff can happen for
3539 	 *   the commit-rec as well.
3540 	 *   The commit-rec is smaller than padding in this scenario and so it is
3541 	 *   not added separately.
3542 	 */
3543 
3544 	/* for trans header */
3545 	unit_bytes += sizeof(xlog_op_header_t);
3546 	unit_bytes += sizeof(xfs_trans_header_t);
3547 
3548 	/* for start-rec */
3549 	unit_bytes += sizeof(xlog_op_header_t);
3550 
3551 	/*
3552 	 * for LR headers - the space for data in an iclog is the size minus
3553 	 * the space used for the headers. If we use the iclog size, then we
3554 	 * undercalculate the number of headers required.
3555 	 *
3556 	 * Furthermore - the addition of op headers for split-recs might
3557 	 * increase the space required enough to require more log and op
3558 	 * headers, so take that into account too.
3559 	 *
3560 	 * IMPORTANT: This reservation makes the assumption that if this
3561 	 * transaction is the first in an iclog and hence has the LR headers
3562 	 * accounted to it, then the remaining space in the iclog is
3563 	 * exclusively for this transaction.  i.e. if the transaction is larger
3564 	 * than the iclog, it will be the only thing in that iclog.
3565 	 * Fundamentally, this means we must pass the entire log vector to
3566 	 * xlog_write to guarantee this.
3567 	 */
3568 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3569 	num_headers = howmany(unit_bytes, iclog_space);
3570 
3571 	/* for split-recs - ophdrs added when data split over LRs */
3572 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3573 
3574 	/* add extra header reservations if we overrun */
3575 	while (!num_headers ||
3576 	       howmany(unit_bytes, iclog_space) > num_headers) {
3577 		unit_bytes += sizeof(xlog_op_header_t);
3578 		num_headers++;
3579 	}
3580 	unit_bytes += log->l_iclog_hsize * num_headers;
3581 
3582 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3583 	unit_bytes += log->l_iclog_hsize;
3584 
3585 	/* roundoff padding for transaction data and one for commit record */
3586 	unit_bytes += 2 * log->l_iclog_roundoff;
3587 
3588 	return unit_bytes;
3589 }
3590 
3591 int
3592 xfs_log_calc_unit_res(
3593 	struct xfs_mount	*mp,
3594 	int			unit_bytes)
3595 {
3596 	return xlog_calc_unit_res(mp->m_log, unit_bytes);
3597 }
3598 
3599 /*
3600  * Allocate and initialise a new log ticket.
3601  */
3602 struct xlog_ticket *
3603 xlog_ticket_alloc(
3604 	struct xlog		*log,
3605 	int			unit_bytes,
3606 	int			cnt,
3607 	char			client,
3608 	bool			permanent)
3609 {
3610 	struct xlog_ticket	*tic;
3611 	int			unit_res;
3612 
3613 	tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3614 
3615 	unit_res = xlog_calc_unit_res(log, unit_bytes);
3616 
3617 	atomic_set(&tic->t_ref, 1);
3618 	tic->t_task		= current;
3619 	INIT_LIST_HEAD(&tic->t_queue);
3620 	tic->t_unit_res		= unit_res;
3621 	tic->t_curr_res		= unit_res;
3622 	tic->t_cnt		= cnt;
3623 	tic->t_ocnt		= cnt;
3624 	tic->t_tid		= prandom_u32();
3625 	tic->t_clientid		= client;
3626 	if (permanent)
3627 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3628 
3629 	xlog_tic_reset_res(tic);
3630 
3631 	return tic;
3632 }
3633 
3634 #if defined(DEBUG)
3635 /*
3636  * Make sure that the destination ptr is within the valid data region of
3637  * one of the iclogs.  This uses backup pointers stored in a different
3638  * part of the log in case we trash the log structure.
3639  */
3640 STATIC void
3641 xlog_verify_dest_ptr(
3642 	struct xlog	*log,
3643 	void		*ptr)
3644 {
3645 	int i;
3646 	int good_ptr = 0;
3647 
3648 	for (i = 0; i < log->l_iclog_bufs; i++) {
3649 		if (ptr >= log->l_iclog_bak[i] &&
3650 		    ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3651 			good_ptr++;
3652 	}
3653 
3654 	if (!good_ptr)
3655 		xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3656 }
3657 
3658 /*
3659  * Check to make sure the grant write head didn't just over lap the tail.  If
3660  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3661  * the cycles differ by exactly one and check the byte count.
3662  *
3663  * This check is run unlocked, so can give false positives. Rather than assert
3664  * on failures, use a warn-once flag and a panic tag to allow the admin to
3665  * determine if they want to panic the machine when such an error occurs. For
3666  * debug kernels this will have the same effect as using an assert but, unlinke
3667  * an assert, it can be turned off at runtime.
3668  */
3669 STATIC void
3670 xlog_verify_grant_tail(
3671 	struct xlog	*log)
3672 {
3673 	int		tail_cycle, tail_blocks;
3674 	int		cycle, space;
3675 
3676 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3677 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3678 	if (tail_cycle != cycle) {
3679 		if (cycle - 1 != tail_cycle &&
3680 		    !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3681 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3682 				"%s: cycle - 1 != tail_cycle", __func__);
3683 		}
3684 
3685 		if (space > BBTOB(tail_blocks) &&
3686 		    !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3687 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3688 				"%s: space > BBTOB(tail_blocks)", __func__);
3689 		}
3690 	}
3691 }
3692 
3693 /* check if it will fit */
3694 STATIC void
3695 xlog_verify_tail_lsn(
3696 	struct xlog		*log,
3697 	struct xlog_in_core	*iclog)
3698 {
3699 	xfs_lsn_t	tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3700 	int		blocks;
3701 
3702     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3703 	blocks =
3704 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3705 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3706 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3707     } else {
3708 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3709 
3710 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3711 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3712 
3713 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3714 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3715 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3716     }
3717 }
3718 
3719 /*
3720  * Perform a number of checks on the iclog before writing to disk.
3721  *
3722  * 1. Make sure the iclogs are still circular
3723  * 2. Make sure we have a good magic number
3724  * 3. Make sure we don't have magic numbers in the data
3725  * 4. Check fields of each log operation header for:
3726  *	A. Valid client identifier
3727  *	B. tid ptr value falls in valid ptr space (user space code)
3728  *	C. Length in log record header is correct according to the
3729  *		individual operation headers within record.
3730  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3731  *	log, check the preceding blocks of the physical log to make sure all
3732  *	the cycle numbers agree with the current cycle number.
3733  */
3734 STATIC void
3735 xlog_verify_iclog(
3736 	struct xlog		*log,
3737 	struct xlog_in_core	*iclog,
3738 	int			count)
3739 {
3740 	xlog_op_header_t	*ophead;
3741 	xlog_in_core_t		*icptr;
3742 	xlog_in_core_2_t	*xhdr;
3743 	void			*base_ptr, *ptr, *p;
3744 	ptrdiff_t		field_offset;
3745 	uint8_t			clientid;
3746 	int			len, i, j, k, op_len;
3747 	int			idx;
3748 
3749 	/* check validity of iclog pointers */
3750 	spin_lock(&log->l_icloglock);
3751 	icptr = log->l_iclog;
3752 	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3753 		ASSERT(icptr);
3754 
3755 	if (icptr != log->l_iclog)
3756 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3757 	spin_unlock(&log->l_icloglock);
3758 
3759 	/* check log magic numbers */
3760 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3761 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3762 
3763 	base_ptr = ptr = &iclog->ic_header;
3764 	p = &iclog->ic_header;
3765 	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3766 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3767 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3768 				__func__);
3769 	}
3770 
3771 	/* check fields */
3772 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3773 	base_ptr = ptr = iclog->ic_datap;
3774 	ophead = ptr;
3775 	xhdr = iclog->ic_data;
3776 	for (i = 0; i < len; i++) {
3777 		ophead = ptr;
3778 
3779 		/* clientid is only 1 byte */
3780 		p = &ophead->oh_clientid;
3781 		field_offset = p - base_ptr;
3782 		if (field_offset & 0x1ff) {
3783 			clientid = ophead->oh_clientid;
3784 		} else {
3785 			idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3786 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3787 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3788 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3789 				clientid = xlog_get_client_id(
3790 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3791 			} else {
3792 				clientid = xlog_get_client_id(
3793 					iclog->ic_header.h_cycle_data[idx]);
3794 			}
3795 		}
3796 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3797 			xfs_warn(log->l_mp,
3798 				"%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3799 				__func__, clientid, ophead,
3800 				(unsigned long)field_offset);
3801 
3802 		/* check length */
3803 		p = &ophead->oh_len;
3804 		field_offset = p - base_ptr;
3805 		if (field_offset & 0x1ff) {
3806 			op_len = be32_to_cpu(ophead->oh_len);
3807 		} else {
3808 			idx = BTOBBT((uintptr_t)&ophead->oh_len -
3809 				    (uintptr_t)iclog->ic_datap);
3810 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3811 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3812 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3813 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3814 			} else {
3815 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3816 			}
3817 		}
3818 		ptr += sizeof(xlog_op_header_t) + op_len;
3819 	}
3820 }
3821 #endif
3822 
3823 /*
3824  * Perform a forced shutdown on the log. This should be called once and once
3825  * only by the high level filesystem shutdown code to shut the log subsystem
3826  * down cleanly.
3827  *
3828  * Our main objectives here are to make sure that:
3829  *	a. if the shutdown was not due to a log IO error, flush the logs to
3830  *	   disk. Anything modified after this is ignored.
3831  *	b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3832  *	   parties to find out. Nothing new gets queued after this is done.
3833  *	c. Tasks sleeping on log reservations, pinned objects and
3834  *	   other resources get woken up.
3835  *
3836  * Return true if the shutdown cause was a log IO error and we actually shut the
3837  * log down.
3838  */
3839 bool
3840 xlog_force_shutdown(
3841 	struct xlog	*log,
3842 	int		shutdown_flags)
3843 {
3844 	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3845 
3846 	/*
3847 	 * If this happens during log recovery then we aren't using the runtime
3848 	 * log mechanisms yet so there's nothing to shut down.
3849 	 */
3850 	if (!log || xlog_in_recovery(log))
3851 		return false;
3852 
3853 	ASSERT(!xlog_is_shutdown(log));
3854 
3855 	/*
3856 	 * Flush all the completed transactions to disk before marking the log
3857 	 * being shut down. We need to do this first as shutting down the log
3858 	 * before the force will prevent the log force from flushing the iclogs
3859 	 * to disk.
3860 	 *
3861 	 * Re-entry due to a log IO error shutdown during the log force is
3862 	 * prevented by the atomicity of higher level shutdown code.
3863 	 */
3864 	if (!log_error)
3865 		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3866 
3867 	/*
3868 	 * Atomically set the shutdown state. If the shutdown state is already
3869 	 * set, there someone else is performing the shutdown and so we are done
3870 	 * here. This should never happen because we should only ever get called
3871 	 * once by the first shutdown caller.
3872 	 *
3873 	 * Much of the log state machine transitions assume that shutdown state
3874 	 * cannot change once they hold the log->l_icloglock. Hence we need to
3875 	 * hold that lock here, even though we use the atomic test_and_set_bit()
3876 	 * operation to set the shutdown state.
3877 	 */
3878 	spin_lock(&log->l_icloglock);
3879 	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3880 		spin_unlock(&log->l_icloglock);
3881 		ASSERT(0);
3882 		return false;
3883 	}
3884 	spin_unlock(&log->l_icloglock);
3885 
3886 	/*
3887 	 * We don't want anybody waiting for log reservations after this. That
3888 	 * means we have to wake up everybody queued up on reserveq as well as
3889 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3890 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3891 	 * action is protected by the grant locks.
3892 	 */
3893 	xlog_grant_head_wake_all(&log->l_reserve_head);
3894 	xlog_grant_head_wake_all(&log->l_write_head);
3895 
3896 	/*
3897 	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3898 	 * as if the log writes were completed. The abort handling in the log
3899 	 * item committed callback functions will do this again under lock to
3900 	 * avoid races.
3901 	 */
3902 	spin_lock(&log->l_cilp->xc_push_lock);
3903 	wake_up_all(&log->l_cilp->xc_start_wait);
3904 	wake_up_all(&log->l_cilp->xc_commit_wait);
3905 	spin_unlock(&log->l_cilp->xc_push_lock);
3906 	xlog_state_shutdown_callbacks(log);
3907 
3908 	return log_error;
3909 }
3910 
3911 STATIC int
3912 xlog_iclogs_empty(
3913 	struct xlog	*log)
3914 {
3915 	xlog_in_core_t	*iclog;
3916 
3917 	iclog = log->l_iclog;
3918 	do {
3919 		/* endianness does not matter here, zero is zero in
3920 		 * any language.
3921 		 */
3922 		if (iclog->ic_header.h_num_logops)
3923 			return 0;
3924 		iclog = iclog->ic_next;
3925 	} while (iclog != log->l_iclog);
3926 	return 1;
3927 }
3928 
3929 /*
3930  * Verify that an LSN stamped into a piece of metadata is valid. This is
3931  * intended for use in read verifiers on v5 superblocks.
3932  */
3933 bool
3934 xfs_log_check_lsn(
3935 	struct xfs_mount	*mp,
3936 	xfs_lsn_t		lsn)
3937 {
3938 	struct xlog		*log = mp->m_log;
3939 	bool			valid;
3940 
3941 	/*
3942 	 * norecovery mode skips mount-time log processing and unconditionally
3943 	 * resets the in-core LSN. We can't validate in this mode, but
3944 	 * modifications are not allowed anyways so just return true.
3945 	 */
3946 	if (xfs_has_norecovery(mp))
3947 		return true;
3948 
3949 	/*
3950 	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3951 	 * handled by recovery and thus safe to ignore here.
3952 	 */
3953 	if (lsn == NULLCOMMITLSN)
3954 		return true;
3955 
3956 	valid = xlog_valid_lsn(mp->m_log, lsn);
3957 
3958 	/* warn the user about what's gone wrong before verifier failure */
3959 	if (!valid) {
3960 		spin_lock(&log->l_icloglock);
3961 		xfs_warn(mp,
3962 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3963 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3964 			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3965 			 log->l_curr_cycle, log->l_curr_block);
3966 		spin_unlock(&log->l_icloglock);
3967 	}
3968 
3969 	return valid;
3970 }
3971 
3972 /*
3973  * Notify the log that we're about to start using a feature that is protected
3974  * by a log incompat feature flag.  This will prevent log covering from
3975  * clearing those flags.
3976  */
3977 void
3978 xlog_use_incompat_feat(
3979 	struct xlog		*log)
3980 {
3981 	down_read(&log->l_incompat_users);
3982 }
3983 
3984 /* Notify the log that we've finished using log incompat features. */
3985 void
3986 xlog_drop_incompat_feat(
3987 	struct xlog		*log)
3988 {
3989 	up_read(&log->l_incompat_users);
3990 }
3991