xref: /linux/fs/xfs/xfs_log.c (revision aa23aa55166c2865ac430168c4b9d405cf8c6980)
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_log_recover.h"
20 #include "xfs_inode.h"
21 #include "xfs_trace.h"
22 #include "xfs_fsops.h"
23 #include "xfs_cksum.h"
24 #include "xfs_sysfs.h"
25 #include "xfs_sb.h"
26 #include "xfs_health.h"
27 
28 kmem_zone_t	*xfs_log_ticket_zone;
29 
30 /* Local miscellaneous function prototypes */
31 STATIC int
32 xlog_commit_record(
33 	struct xlog		*log,
34 	struct xlog_ticket	*ticket,
35 	struct xlog_in_core	**iclog,
36 	xfs_lsn_t		*commitlsnp);
37 
38 STATIC struct xlog *
39 xlog_alloc_log(
40 	struct xfs_mount	*mp,
41 	struct xfs_buftarg	*log_target,
42 	xfs_daddr_t		blk_offset,
43 	int			num_bblks);
44 STATIC int
45 xlog_space_left(
46 	struct xlog		*log,
47 	atomic64_t		*head);
48 STATIC int
49 xlog_sync(
50 	struct xlog		*log,
51 	struct xlog_in_core	*iclog);
52 STATIC void
53 xlog_dealloc_log(
54 	struct xlog		*log);
55 
56 /* local state machine functions */
57 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
58 STATIC void
59 xlog_state_do_callback(
60 	struct xlog		*log,
61 	int			aborted,
62 	struct xlog_in_core	*iclog);
63 STATIC int
64 xlog_state_get_iclog_space(
65 	struct xlog		*log,
66 	int			len,
67 	struct xlog_in_core	**iclog,
68 	struct xlog_ticket	*ticket,
69 	int			*continued_write,
70 	int			*logoffsetp);
71 STATIC int
72 xlog_state_release_iclog(
73 	struct xlog		*log,
74 	struct xlog_in_core	*iclog);
75 STATIC void
76 xlog_state_switch_iclogs(
77 	struct xlog		*log,
78 	struct xlog_in_core	*iclog,
79 	int			eventual_size);
80 STATIC void
81 xlog_state_want_sync(
82 	struct xlog		*log,
83 	struct xlog_in_core	*iclog);
84 
85 STATIC void
86 xlog_grant_push_ail(
87 	struct xlog		*log,
88 	int			need_bytes);
89 STATIC void
90 xlog_regrant_reserve_log_space(
91 	struct xlog		*log,
92 	struct xlog_ticket	*ticket);
93 STATIC void
94 xlog_ungrant_log_space(
95 	struct xlog		*log,
96 	struct xlog_ticket	*ticket);
97 
98 #if defined(DEBUG)
99 STATIC void
100 xlog_verify_dest_ptr(
101 	struct xlog		*log,
102 	void			*ptr);
103 STATIC void
104 xlog_verify_grant_tail(
105 	struct xlog *log);
106 STATIC void
107 xlog_verify_iclog(
108 	struct xlog		*log,
109 	struct xlog_in_core	*iclog,
110 	int			count,
111 	bool                    syncing);
112 STATIC void
113 xlog_verify_tail_lsn(
114 	struct xlog		*log,
115 	struct xlog_in_core	*iclog,
116 	xfs_lsn_t		tail_lsn);
117 #else
118 #define xlog_verify_dest_ptr(a,b)
119 #define xlog_verify_grant_tail(a)
120 #define xlog_verify_iclog(a,b,c,d)
121 #define xlog_verify_tail_lsn(a,b,c)
122 #endif
123 
124 STATIC int
125 xlog_iclogs_empty(
126 	struct xlog		*log);
127 
128 static void
129 xlog_grant_sub_space(
130 	struct xlog		*log,
131 	atomic64_t		*head,
132 	int			bytes)
133 {
134 	int64_t	head_val = atomic64_read(head);
135 	int64_t new, old;
136 
137 	do {
138 		int	cycle, space;
139 
140 		xlog_crack_grant_head_val(head_val, &cycle, &space);
141 
142 		space -= bytes;
143 		if (space < 0) {
144 			space += log->l_logsize;
145 			cycle--;
146 		}
147 
148 		old = head_val;
149 		new = xlog_assign_grant_head_val(cycle, space);
150 		head_val = atomic64_cmpxchg(head, old, new);
151 	} while (head_val != old);
152 }
153 
154 static void
155 xlog_grant_add_space(
156 	struct xlog		*log,
157 	atomic64_t		*head,
158 	int			bytes)
159 {
160 	int64_t	head_val = atomic64_read(head);
161 	int64_t new, old;
162 
163 	do {
164 		int		tmp;
165 		int		cycle, space;
166 
167 		xlog_crack_grant_head_val(head_val, &cycle, &space);
168 
169 		tmp = log->l_logsize - space;
170 		if (tmp > bytes)
171 			space += bytes;
172 		else {
173 			space = bytes - tmp;
174 			cycle++;
175 		}
176 
177 		old = head_val;
178 		new = xlog_assign_grant_head_val(cycle, space);
179 		head_val = atomic64_cmpxchg(head, old, new);
180 	} while (head_val != old);
181 }
182 
183 STATIC void
184 xlog_grant_head_init(
185 	struct xlog_grant_head	*head)
186 {
187 	xlog_assign_grant_head(&head->grant, 1, 0);
188 	INIT_LIST_HEAD(&head->waiters);
189 	spin_lock_init(&head->lock);
190 }
191 
192 STATIC void
193 xlog_grant_head_wake_all(
194 	struct xlog_grant_head	*head)
195 {
196 	struct xlog_ticket	*tic;
197 
198 	spin_lock(&head->lock);
199 	list_for_each_entry(tic, &head->waiters, t_queue)
200 		wake_up_process(tic->t_task);
201 	spin_unlock(&head->lock);
202 }
203 
204 static inline int
205 xlog_ticket_reservation(
206 	struct xlog		*log,
207 	struct xlog_grant_head	*head,
208 	struct xlog_ticket	*tic)
209 {
210 	if (head == &log->l_write_head) {
211 		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
212 		return tic->t_unit_res;
213 	} else {
214 		if (tic->t_flags & XLOG_TIC_PERM_RESERV)
215 			return tic->t_unit_res * tic->t_cnt;
216 		else
217 			return tic->t_unit_res;
218 	}
219 }
220 
221 STATIC bool
222 xlog_grant_head_wake(
223 	struct xlog		*log,
224 	struct xlog_grant_head	*head,
225 	int			*free_bytes)
226 {
227 	struct xlog_ticket	*tic;
228 	int			need_bytes;
229 
230 	list_for_each_entry(tic, &head->waiters, t_queue) {
231 		need_bytes = xlog_ticket_reservation(log, head, tic);
232 		if (*free_bytes < need_bytes)
233 			return false;
234 
235 		*free_bytes -= need_bytes;
236 		trace_xfs_log_grant_wake_up(log, tic);
237 		wake_up_process(tic->t_task);
238 	}
239 
240 	return true;
241 }
242 
243 STATIC int
244 xlog_grant_head_wait(
245 	struct xlog		*log,
246 	struct xlog_grant_head	*head,
247 	struct xlog_ticket	*tic,
248 	int			need_bytes) __releases(&head->lock)
249 					    __acquires(&head->lock)
250 {
251 	list_add_tail(&tic->t_queue, &head->waiters);
252 
253 	do {
254 		if (XLOG_FORCED_SHUTDOWN(log))
255 			goto shutdown;
256 		xlog_grant_push_ail(log, need_bytes);
257 
258 		__set_current_state(TASK_UNINTERRUPTIBLE);
259 		spin_unlock(&head->lock);
260 
261 		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
262 
263 		trace_xfs_log_grant_sleep(log, tic);
264 		schedule();
265 		trace_xfs_log_grant_wake(log, tic);
266 
267 		spin_lock(&head->lock);
268 		if (XLOG_FORCED_SHUTDOWN(log))
269 			goto shutdown;
270 	} while (xlog_space_left(log, &head->grant) < need_bytes);
271 
272 	list_del_init(&tic->t_queue);
273 	return 0;
274 shutdown:
275 	list_del_init(&tic->t_queue);
276 	return -EIO;
277 }
278 
279 /*
280  * Atomically get the log space required for a log ticket.
281  *
282  * Once a ticket gets put onto head->waiters, it will only return after the
283  * needed reservation is satisfied.
284  *
285  * This function is structured so that it has a lock free fast path. This is
286  * necessary because every new transaction reservation will come through this
287  * path. Hence any lock will be globally hot if we take it unconditionally on
288  * every pass.
289  *
290  * As tickets are only ever moved on and off head->waiters under head->lock, we
291  * only need to take that lock if we are going to add the ticket to the queue
292  * and sleep. We can avoid taking the lock if the ticket was never added to
293  * head->waiters because the t_queue list head will be empty and we hold the
294  * only reference to it so it can safely be checked unlocked.
295  */
296 STATIC int
297 xlog_grant_head_check(
298 	struct xlog		*log,
299 	struct xlog_grant_head	*head,
300 	struct xlog_ticket	*tic,
301 	int			*need_bytes)
302 {
303 	int			free_bytes;
304 	int			error = 0;
305 
306 	ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
307 
308 	/*
309 	 * If there are other waiters on the queue then give them a chance at
310 	 * logspace before us.  Wake up the first waiters, if we do not wake
311 	 * up all the waiters then go to sleep waiting for more free space,
312 	 * otherwise try to get some space for this transaction.
313 	 */
314 	*need_bytes = xlog_ticket_reservation(log, head, tic);
315 	free_bytes = xlog_space_left(log, &head->grant);
316 	if (!list_empty_careful(&head->waiters)) {
317 		spin_lock(&head->lock);
318 		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
319 		    free_bytes < *need_bytes) {
320 			error = xlog_grant_head_wait(log, head, tic,
321 						     *need_bytes);
322 		}
323 		spin_unlock(&head->lock);
324 	} else if (free_bytes < *need_bytes) {
325 		spin_lock(&head->lock);
326 		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
327 		spin_unlock(&head->lock);
328 	}
329 
330 	return error;
331 }
332 
333 static void
334 xlog_tic_reset_res(xlog_ticket_t *tic)
335 {
336 	tic->t_res_num = 0;
337 	tic->t_res_arr_sum = 0;
338 	tic->t_res_num_ophdrs = 0;
339 }
340 
341 static void
342 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
343 {
344 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
345 		/* add to overflow and start again */
346 		tic->t_res_o_flow += tic->t_res_arr_sum;
347 		tic->t_res_num = 0;
348 		tic->t_res_arr_sum = 0;
349 	}
350 
351 	tic->t_res_arr[tic->t_res_num].r_len = len;
352 	tic->t_res_arr[tic->t_res_num].r_type = type;
353 	tic->t_res_arr_sum += len;
354 	tic->t_res_num++;
355 }
356 
357 /*
358  * Replenish the byte reservation required by moving the grant write head.
359  */
360 int
361 xfs_log_regrant(
362 	struct xfs_mount	*mp,
363 	struct xlog_ticket	*tic)
364 {
365 	struct xlog		*log = mp->m_log;
366 	int			need_bytes;
367 	int			error = 0;
368 
369 	if (XLOG_FORCED_SHUTDOWN(log))
370 		return -EIO;
371 
372 	XFS_STATS_INC(mp, xs_try_logspace);
373 
374 	/*
375 	 * This is a new transaction on the ticket, so we need to change the
376 	 * transaction ID so that the next transaction has a different TID in
377 	 * the log. Just add one to the existing tid so that we can see chains
378 	 * of rolling transactions in the log easily.
379 	 */
380 	tic->t_tid++;
381 
382 	xlog_grant_push_ail(log, tic->t_unit_res);
383 
384 	tic->t_curr_res = tic->t_unit_res;
385 	xlog_tic_reset_res(tic);
386 
387 	if (tic->t_cnt > 0)
388 		return 0;
389 
390 	trace_xfs_log_regrant(log, tic);
391 
392 	error = xlog_grant_head_check(log, &log->l_write_head, tic,
393 				      &need_bytes);
394 	if (error)
395 		goto out_error;
396 
397 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
398 	trace_xfs_log_regrant_exit(log, tic);
399 	xlog_verify_grant_tail(log);
400 	return 0;
401 
402 out_error:
403 	/*
404 	 * If we are failing, make sure the ticket doesn't have any current
405 	 * reservations.  We don't want to add this back when the ticket/
406 	 * transaction gets cancelled.
407 	 */
408 	tic->t_curr_res = 0;
409 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
410 	return error;
411 }
412 
413 /*
414  * Reserve log space and return a ticket corresponding to the reservation.
415  *
416  * Each reservation is going to reserve extra space for a log record header.
417  * When writes happen to the on-disk log, we don't subtract the length of the
418  * log record header from any reservation.  By wasting space in each
419  * reservation, we prevent over allocation problems.
420  */
421 int
422 xfs_log_reserve(
423 	struct xfs_mount	*mp,
424 	int		 	unit_bytes,
425 	int		 	cnt,
426 	struct xlog_ticket	**ticp,
427 	uint8_t		 	client,
428 	bool			permanent)
429 {
430 	struct xlog		*log = mp->m_log;
431 	struct xlog_ticket	*tic;
432 	int			need_bytes;
433 	int			error = 0;
434 
435 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
436 
437 	if (XLOG_FORCED_SHUTDOWN(log))
438 		return -EIO;
439 
440 	XFS_STATS_INC(mp, xs_try_logspace);
441 
442 	ASSERT(*ticp == NULL);
443 	tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
444 				KM_SLEEP | KM_MAYFAIL);
445 	if (!tic)
446 		return -ENOMEM;
447 
448 	*ticp = tic;
449 
450 	xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
451 					    : tic->t_unit_res);
452 
453 	trace_xfs_log_reserve(log, tic);
454 
455 	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
456 				      &need_bytes);
457 	if (error)
458 		goto out_error;
459 
460 	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
461 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
462 	trace_xfs_log_reserve_exit(log, tic);
463 	xlog_verify_grant_tail(log);
464 	return 0;
465 
466 out_error:
467 	/*
468 	 * If we are failing, make sure the ticket doesn't have any current
469 	 * reservations.  We don't want to add this back when the ticket/
470 	 * transaction gets cancelled.
471 	 */
472 	tic->t_curr_res = 0;
473 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
474 	return error;
475 }
476 
477 
478 /*
479  * NOTES:
480  *
481  *	1. currblock field gets updated at startup and after in-core logs
482  *		marked as with WANT_SYNC.
483  */
484 
485 /*
486  * This routine is called when a user of a log manager ticket is done with
487  * the reservation.  If the ticket was ever used, then a commit record for
488  * the associated transaction is written out as a log operation header with
489  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
490  * a given ticket.  If the ticket was one with a permanent reservation, then
491  * a few operations are done differently.  Permanent reservation tickets by
492  * default don't release the reservation.  They just commit the current
493  * transaction with the belief that the reservation is still needed.  A flag
494  * must be passed in before permanent reservations are actually released.
495  * When these type of tickets are not released, they need to be set into
496  * the inited state again.  By doing this, a start record will be written
497  * out when the next write occurs.
498  */
499 xfs_lsn_t
500 xfs_log_done(
501 	struct xfs_mount	*mp,
502 	struct xlog_ticket	*ticket,
503 	struct xlog_in_core	**iclog,
504 	bool			regrant)
505 {
506 	struct xlog		*log = mp->m_log;
507 	xfs_lsn_t		lsn = 0;
508 
509 	if (XLOG_FORCED_SHUTDOWN(log) ||
510 	    /*
511 	     * If nothing was ever written, don't write out commit record.
512 	     * If we get an error, just continue and give back the log ticket.
513 	     */
514 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
515 	     (xlog_commit_record(log, ticket, iclog, &lsn)))) {
516 		lsn = (xfs_lsn_t) -1;
517 		regrant = false;
518 	}
519 
520 
521 	if (!regrant) {
522 		trace_xfs_log_done_nonperm(log, ticket);
523 
524 		/*
525 		 * Release ticket if not permanent reservation or a specific
526 		 * request has been made to release a permanent reservation.
527 		 */
528 		xlog_ungrant_log_space(log, ticket);
529 	} else {
530 		trace_xfs_log_done_perm(log, ticket);
531 
532 		xlog_regrant_reserve_log_space(log, ticket);
533 		/* If this ticket was a permanent reservation and we aren't
534 		 * trying to release it, reset the inited flags; so next time
535 		 * we write, a start record will be written out.
536 		 */
537 		ticket->t_flags |= XLOG_TIC_INITED;
538 	}
539 
540 	xfs_log_ticket_put(ticket);
541 	return lsn;
542 }
543 
544 /*
545  * Attaches a new iclog I/O completion callback routine during
546  * transaction commit.  If the log is in error state, a non-zero
547  * return code is handed back and the caller is responsible for
548  * executing the callback at an appropriate time.
549  */
550 int
551 xfs_log_notify(
552 	struct xlog_in_core	*iclog,
553 	xfs_log_callback_t	*cb)
554 {
555 	int	abortflg;
556 
557 	spin_lock(&iclog->ic_callback_lock);
558 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
559 	if (!abortflg) {
560 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
561 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
562 		cb->cb_next = NULL;
563 		*(iclog->ic_callback_tail) = cb;
564 		iclog->ic_callback_tail = &(cb->cb_next);
565 	}
566 	spin_unlock(&iclog->ic_callback_lock);
567 	return abortflg;
568 }
569 
570 int
571 xfs_log_release_iclog(
572 	struct xfs_mount	*mp,
573 	struct xlog_in_core	*iclog)
574 {
575 	if (xlog_state_release_iclog(mp->m_log, iclog)) {
576 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
577 		return -EIO;
578 	}
579 
580 	return 0;
581 }
582 
583 /*
584  * Mount a log filesystem
585  *
586  * mp		- ubiquitous xfs mount point structure
587  * log_target	- buftarg of on-disk log device
588  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
589  * num_bblocks	- Number of BBSIZE blocks in on-disk log
590  *
591  * Return error or zero.
592  */
593 int
594 xfs_log_mount(
595 	xfs_mount_t	*mp,
596 	xfs_buftarg_t	*log_target,
597 	xfs_daddr_t	blk_offset,
598 	int		num_bblks)
599 {
600 	bool		fatal = xfs_sb_version_hascrc(&mp->m_sb);
601 	int		error = 0;
602 	int		min_logfsbs;
603 
604 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
605 		xfs_notice(mp, "Mounting V%d Filesystem",
606 			   XFS_SB_VERSION_NUM(&mp->m_sb));
607 	} else {
608 		xfs_notice(mp,
609 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
610 			   XFS_SB_VERSION_NUM(&mp->m_sb));
611 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
612 	}
613 
614 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
615 	if (IS_ERR(mp->m_log)) {
616 		error = PTR_ERR(mp->m_log);
617 		goto out;
618 	}
619 
620 	/*
621 	 * Validate the given log space and drop a critical message via syslog
622 	 * if the log size is too small that would lead to some unexpected
623 	 * situations in transaction log space reservation stage.
624 	 *
625 	 * Note: we can't just reject the mount if the validation fails.  This
626 	 * would mean that people would have to downgrade their kernel just to
627 	 * remedy the situation as there is no way to grow the log (short of
628 	 * black magic surgery with xfs_db).
629 	 *
630 	 * We can, however, reject mounts for CRC format filesystems, as the
631 	 * mkfs binary being used to make the filesystem should never create a
632 	 * filesystem with a log that is too small.
633 	 */
634 	min_logfsbs = xfs_log_calc_minimum_size(mp);
635 
636 	if (mp->m_sb.sb_logblocks < min_logfsbs) {
637 		xfs_warn(mp,
638 		"Log size %d blocks too small, minimum size is %d blocks",
639 			 mp->m_sb.sb_logblocks, min_logfsbs);
640 		error = -EINVAL;
641 	} else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
642 		xfs_warn(mp,
643 		"Log size %d blocks too large, maximum size is %lld blocks",
644 			 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
645 		error = -EINVAL;
646 	} else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
647 		xfs_warn(mp,
648 		"log size %lld bytes too large, maximum size is %lld bytes",
649 			 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
650 			 XFS_MAX_LOG_BYTES);
651 		error = -EINVAL;
652 	} else if (mp->m_sb.sb_logsunit > 1 &&
653 		   mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
654 		xfs_warn(mp,
655 		"log stripe unit %u bytes must be a multiple of block size",
656 			 mp->m_sb.sb_logsunit);
657 		error = -EINVAL;
658 		fatal = true;
659 	}
660 	if (error) {
661 		/*
662 		 * Log check errors are always fatal on v5; or whenever bad
663 		 * metadata leads to a crash.
664 		 */
665 		if (fatal) {
666 			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
667 			ASSERT(0);
668 			goto out_free_log;
669 		}
670 		xfs_crit(mp, "Log size out of supported range.");
671 		xfs_crit(mp,
672 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
673 	}
674 
675 	/*
676 	 * Initialize the AIL now we have a log.
677 	 */
678 	error = xfs_trans_ail_init(mp);
679 	if (error) {
680 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
681 		goto out_free_log;
682 	}
683 	mp->m_log->l_ailp = mp->m_ail;
684 
685 	/*
686 	 * skip log recovery on a norecovery mount.  pretend it all
687 	 * just worked.
688 	 */
689 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
690 		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
691 
692 		if (readonly)
693 			mp->m_flags &= ~XFS_MOUNT_RDONLY;
694 
695 		error = xlog_recover(mp->m_log);
696 
697 		if (readonly)
698 			mp->m_flags |= XFS_MOUNT_RDONLY;
699 		if (error) {
700 			xfs_warn(mp, "log mount/recovery failed: error %d",
701 				error);
702 			xlog_recover_cancel(mp->m_log);
703 			goto out_destroy_ail;
704 		}
705 	}
706 
707 	error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
708 			       "log");
709 	if (error)
710 		goto out_destroy_ail;
711 
712 	/* Normal transactions can now occur */
713 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
714 
715 	/*
716 	 * Now the log has been fully initialised and we know were our
717 	 * space grant counters are, we can initialise the permanent ticket
718 	 * needed for delayed logging to work.
719 	 */
720 	xlog_cil_init_post_recovery(mp->m_log);
721 
722 	return 0;
723 
724 out_destroy_ail:
725 	xfs_trans_ail_destroy(mp);
726 out_free_log:
727 	xlog_dealloc_log(mp->m_log);
728 out:
729 	return error;
730 }
731 
732 /*
733  * Finish the recovery of the file system.  This is separate from the
734  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
735  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
736  * here.
737  *
738  * If we finish recovery successfully, start the background log work. If we are
739  * not doing recovery, then we have a RO filesystem and we don't need to start
740  * it.
741  */
742 int
743 xfs_log_mount_finish(
744 	struct xfs_mount	*mp)
745 {
746 	int	error = 0;
747 	bool	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
748 	bool	recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
749 
750 	if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
751 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
752 		return 0;
753 	} else if (readonly) {
754 		/* Allow unlinked processing to proceed */
755 		mp->m_flags &= ~XFS_MOUNT_RDONLY;
756 	}
757 
758 	/*
759 	 * During the second phase of log recovery, we need iget and
760 	 * iput to behave like they do for an active filesystem.
761 	 * xfs_fs_drop_inode needs to be able to prevent the deletion
762 	 * of inodes before we're done replaying log items on those
763 	 * inodes.  Turn it off immediately after recovery finishes
764 	 * so that we don't leak the quota inodes if subsequent mount
765 	 * activities fail.
766 	 *
767 	 * We let all inodes involved in redo item processing end up on
768 	 * the LRU instead of being evicted immediately so that if we do
769 	 * something to an unlinked inode, the irele won't cause
770 	 * premature truncation and freeing of the inode, which results
771 	 * in log recovery failure.  We have to evict the unreferenced
772 	 * lru inodes after clearing SB_ACTIVE because we don't
773 	 * otherwise clean up the lru if there's a subsequent failure in
774 	 * xfs_mountfs, which leads to us leaking the inodes if nothing
775 	 * else (e.g. quotacheck) references the inodes before the
776 	 * mount failure occurs.
777 	 */
778 	mp->m_super->s_flags |= SB_ACTIVE;
779 	error = xlog_recover_finish(mp->m_log);
780 	if (!error)
781 		xfs_log_work_queue(mp);
782 	mp->m_super->s_flags &= ~SB_ACTIVE;
783 	evict_inodes(mp->m_super);
784 
785 	/*
786 	 * Drain the buffer LRU after log recovery. This is required for v4
787 	 * filesystems to avoid leaving around buffers with NULL verifier ops,
788 	 * but we do it unconditionally to make sure we're always in a clean
789 	 * cache state after mount.
790 	 *
791 	 * Don't push in the error case because the AIL may have pending intents
792 	 * that aren't removed until recovery is cancelled.
793 	 */
794 	if (!error && recovered) {
795 		xfs_log_force(mp, XFS_LOG_SYNC);
796 		xfs_ail_push_all_sync(mp->m_ail);
797 	}
798 	xfs_wait_buftarg(mp->m_ddev_targp);
799 
800 	if (readonly)
801 		mp->m_flags |= XFS_MOUNT_RDONLY;
802 
803 	return error;
804 }
805 
806 /*
807  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
808  * the log.
809  */
810 int
811 xfs_log_mount_cancel(
812 	struct xfs_mount	*mp)
813 {
814 	int			error;
815 
816 	error = xlog_recover_cancel(mp->m_log);
817 	xfs_log_unmount(mp);
818 
819 	return error;
820 }
821 
822 /*
823  * Final log writes as part of unmount.
824  *
825  * Mark the filesystem clean as unmount happens.  Note that during relocation
826  * this routine needs to be executed as part of source-bag while the
827  * deallocation must not be done until source-end.
828  */
829 
830 /* Actually write the unmount record to disk. */
831 static void
832 xfs_log_write_unmount_record(
833 	struct xfs_mount	*mp)
834 {
835 	/* the data section must be 32 bit size aligned */
836 	struct xfs_unmount_log_format magic = {
837 		.magic = XLOG_UNMOUNT_TYPE,
838 	};
839 	struct xfs_log_iovec reg = {
840 		.i_addr = &magic,
841 		.i_len = sizeof(magic),
842 		.i_type = XLOG_REG_TYPE_UNMOUNT,
843 	};
844 	struct xfs_log_vec vec = {
845 		.lv_niovecs = 1,
846 		.lv_iovecp = &reg,
847 	};
848 	struct xlog		*log = mp->m_log;
849 	struct xlog_in_core	*iclog;
850 	struct xlog_ticket	*tic = NULL;
851 	xfs_lsn_t		lsn;
852 	uint			flags = XLOG_UNMOUNT_TRANS;
853 	int			error;
854 
855 	error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
856 	if (error)
857 		goto out_err;
858 
859 	/*
860 	 * If we think the summary counters are bad, clear the unmount header
861 	 * flag in the unmount record so that the summary counters will be
862 	 * recalculated during log recovery at next mount.  Refer to
863 	 * xlog_check_unmount_rec for more details.
864 	 */
865 	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
866 			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
867 		xfs_alert(mp, "%s: will fix summary counters at next mount",
868 				__func__);
869 		flags &= ~XLOG_UNMOUNT_TRANS;
870 	}
871 
872 	/* remove inited flag, and account for space used */
873 	tic->t_flags = 0;
874 	tic->t_curr_res -= sizeof(magic);
875 	error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
876 	/*
877 	 * At this point, we're umounting anyway, so there's no point in
878 	 * transitioning log state to IOERROR. Just continue...
879 	 */
880 out_err:
881 	if (error)
882 		xfs_alert(mp, "%s: unmount record failed", __func__);
883 
884 	spin_lock(&log->l_icloglock);
885 	iclog = log->l_iclog;
886 	atomic_inc(&iclog->ic_refcnt);
887 	xlog_state_want_sync(log, iclog);
888 	spin_unlock(&log->l_icloglock);
889 	error = xlog_state_release_iclog(log, iclog);
890 
891 	spin_lock(&log->l_icloglock);
892 	switch (iclog->ic_state) {
893 	default:
894 		if (!XLOG_FORCED_SHUTDOWN(log)) {
895 			xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
896 			break;
897 		}
898 		/* fall through */
899 	case XLOG_STATE_ACTIVE:
900 	case XLOG_STATE_DIRTY:
901 		spin_unlock(&log->l_icloglock);
902 		break;
903 	}
904 
905 	if (tic) {
906 		trace_xfs_log_umount_write(log, tic);
907 		xlog_ungrant_log_space(log, tic);
908 		xfs_log_ticket_put(tic);
909 	}
910 }
911 
912 /*
913  * Unmount record used to have a string "Unmount filesystem--" in the
914  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
915  * We just write the magic number now since that particular field isn't
916  * currently architecture converted and "Unmount" is a bit foo.
917  * As far as I know, there weren't any dependencies on the old behaviour.
918  */
919 
920 static int
921 xfs_log_unmount_write(xfs_mount_t *mp)
922 {
923 	struct xlog	 *log = mp->m_log;
924 	xlog_in_core_t	 *iclog;
925 #ifdef DEBUG
926 	xlog_in_core_t	 *first_iclog;
927 #endif
928 	int		 error;
929 
930 	/*
931 	 * Don't write out unmount record on norecovery mounts or ro devices.
932 	 * Or, if we are doing a forced umount (typically because of IO errors).
933 	 */
934 	if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
935 	    xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
936 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
937 		return 0;
938 	}
939 
940 	error = xfs_log_force(mp, XFS_LOG_SYNC);
941 	ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
942 
943 #ifdef DEBUG
944 	first_iclog = iclog = log->l_iclog;
945 	do {
946 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
947 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
948 			ASSERT(iclog->ic_offset == 0);
949 		}
950 		iclog = iclog->ic_next;
951 	} while (iclog != first_iclog);
952 #endif
953 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
954 		xfs_log_write_unmount_record(mp);
955 	} else {
956 		/*
957 		 * We're already in forced_shutdown mode, couldn't
958 		 * even attempt to write out the unmount transaction.
959 		 *
960 		 * Go through the motions of sync'ing and releasing
961 		 * the iclog, even though no I/O will actually happen,
962 		 * we need to wait for other log I/Os that may already
963 		 * be in progress.  Do this as a separate section of
964 		 * code so we'll know if we ever get stuck here that
965 		 * we're in this odd situation of trying to unmount
966 		 * a file system that went into forced_shutdown as
967 		 * the result of an unmount..
968 		 */
969 		spin_lock(&log->l_icloglock);
970 		iclog = log->l_iclog;
971 		atomic_inc(&iclog->ic_refcnt);
972 
973 		xlog_state_want_sync(log, iclog);
974 		spin_unlock(&log->l_icloglock);
975 		error =  xlog_state_release_iclog(log, iclog);
976 
977 		spin_lock(&log->l_icloglock);
978 
979 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
980 			|| iclog->ic_state == XLOG_STATE_DIRTY
981 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
982 
983 				xlog_wait(&iclog->ic_force_wait,
984 							&log->l_icloglock);
985 		} else {
986 			spin_unlock(&log->l_icloglock);
987 		}
988 	}
989 
990 	return error;
991 }	/* xfs_log_unmount_write */
992 
993 /*
994  * Empty the log for unmount/freeze.
995  *
996  * To do this, we first need to shut down the background log work so it is not
997  * trying to cover the log as we clean up. We then need to unpin all objects in
998  * the log so we can then flush them out. Once they have completed their IO and
999  * run the callbacks removing themselves from the AIL, we can write the unmount
1000  * record.
1001  */
1002 void
1003 xfs_log_quiesce(
1004 	struct xfs_mount	*mp)
1005 {
1006 	cancel_delayed_work_sync(&mp->m_log->l_work);
1007 	xfs_log_force(mp, XFS_LOG_SYNC);
1008 
1009 	/*
1010 	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1011 	 * will push it, xfs_wait_buftarg() will not wait for it. Further,
1012 	 * xfs_buf_iowait() cannot be used because it was pushed with the
1013 	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1014 	 * the IO to complete.
1015 	 */
1016 	xfs_ail_push_all_sync(mp->m_ail);
1017 	xfs_wait_buftarg(mp->m_ddev_targp);
1018 	xfs_buf_lock(mp->m_sb_bp);
1019 	xfs_buf_unlock(mp->m_sb_bp);
1020 
1021 	xfs_log_unmount_write(mp);
1022 }
1023 
1024 /*
1025  * Shut down and release the AIL and Log.
1026  *
1027  * During unmount, we need to ensure we flush all the dirty metadata objects
1028  * from the AIL so that the log is empty before we write the unmount record to
1029  * the log. Once this is done, we can tear down the AIL and the log.
1030  */
1031 void
1032 xfs_log_unmount(
1033 	struct xfs_mount	*mp)
1034 {
1035 	xfs_log_quiesce(mp);
1036 
1037 	xfs_trans_ail_destroy(mp);
1038 
1039 	xfs_sysfs_del(&mp->m_log->l_kobj);
1040 
1041 	xlog_dealloc_log(mp->m_log);
1042 }
1043 
1044 void
1045 xfs_log_item_init(
1046 	struct xfs_mount	*mp,
1047 	struct xfs_log_item	*item,
1048 	int			type,
1049 	const struct xfs_item_ops *ops)
1050 {
1051 	item->li_mountp = mp;
1052 	item->li_ailp = mp->m_ail;
1053 	item->li_type = type;
1054 	item->li_ops = ops;
1055 	item->li_lv = NULL;
1056 
1057 	INIT_LIST_HEAD(&item->li_ail);
1058 	INIT_LIST_HEAD(&item->li_cil);
1059 	INIT_LIST_HEAD(&item->li_bio_list);
1060 	INIT_LIST_HEAD(&item->li_trans);
1061 }
1062 
1063 /*
1064  * Wake up processes waiting for log space after we have moved the log tail.
1065  */
1066 void
1067 xfs_log_space_wake(
1068 	struct xfs_mount	*mp)
1069 {
1070 	struct xlog		*log = mp->m_log;
1071 	int			free_bytes;
1072 
1073 	if (XLOG_FORCED_SHUTDOWN(log))
1074 		return;
1075 
1076 	if (!list_empty_careful(&log->l_write_head.waiters)) {
1077 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1078 
1079 		spin_lock(&log->l_write_head.lock);
1080 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1081 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1082 		spin_unlock(&log->l_write_head.lock);
1083 	}
1084 
1085 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1086 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1087 
1088 		spin_lock(&log->l_reserve_head.lock);
1089 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1090 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1091 		spin_unlock(&log->l_reserve_head.lock);
1092 	}
1093 }
1094 
1095 /*
1096  * Determine if we have a transaction that has gone to disk that needs to be
1097  * covered. To begin the transition to the idle state firstly the log needs to
1098  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1099  * we start attempting to cover the log.
1100  *
1101  * Only if we are then in a state where covering is needed, the caller is
1102  * informed that dummy transactions are required to move the log into the idle
1103  * state.
1104  *
1105  * If there are any items in the AIl or CIL, then we do not want to attempt to
1106  * cover the log as we may be in a situation where there isn't log space
1107  * available to run a dummy transaction and this can lead to deadlocks when the
1108  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1109  * there's no point in running a dummy transaction at this point because we
1110  * can't start trying to idle the log until both the CIL and AIL are empty.
1111  */
1112 static int
1113 xfs_log_need_covered(xfs_mount_t *mp)
1114 {
1115 	struct xlog	*log = mp->m_log;
1116 	int		needed = 0;
1117 
1118 	if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1119 		return 0;
1120 
1121 	if (!xlog_cil_empty(log))
1122 		return 0;
1123 
1124 	spin_lock(&log->l_icloglock);
1125 	switch (log->l_covered_state) {
1126 	case XLOG_STATE_COVER_DONE:
1127 	case XLOG_STATE_COVER_DONE2:
1128 	case XLOG_STATE_COVER_IDLE:
1129 		break;
1130 	case XLOG_STATE_COVER_NEED:
1131 	case XLOG_STATE_COVER_NEED2:
1132 		if (xfs_ail_min_lsn(log->l_ailp))
1133 			break;
1134 		if (!xlog_iclogs_empty(log))
1135 			break;
1136 
1137 		needed = 1;
1138 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1139 			log->l_covered_state = XLOG_STATE_COVER_DONE;
1140 		else
1141 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1142 		break;
1143 	default:
1144 		needed = 1;
1145 		break;
1146 	}
1147 	spin_unlock(&log->l_icloglock);
1148 	return needed;
1149 }
1150 
1151 /*
1152  * We may be holding the log iclog lock upon entering this routine.
1153  */
1154 xfs_lsn_t
1155 xlog_assign_tail_lsn_locked(
1156 	struct xfs_mount	*mp)
1157 {
1158 	struct xlog		*log = mp->m_log;
1159 	struct xfs_log_item	*lip;
1160 	xfs_lsn_t		tail_lsn;
1161 
1162 	assert_spin_locked(&mp->m_ail->ail_lock);
1163 
1164 	/*
1165 	 * To make sure we always have a valid LSN for the log tail we keep
1166 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1167 	 * and use that when the AIL was empty.
1168 	 */
1169 	lip = xfs_ail_min(mp->m_ail);
1170 	if (lip)
1171 		tail_lsn = lip->li_lsn;
1172 	else
1173 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1174 	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1175 	atomic64_set(&log->l_tail_lsn, tail_lsn);
1176 	return tail_lsn;
1177 }
1178 
1179 xfs_lsn_t
1180 xlog_assign_tail_lsn(
1181 	struct xfs_mount	*mp)
1182 {
1183 	xfs_lsn_t		tail_lsn;
1184 
1185 	spin_lock(&mp->m_ail->ail_lock);
1186 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1187 	spin_unlock(&mp->m_ail->ail_lock);
1188 
1189 	return tail_lsn;
1190 }
1191 
1192 /*
1193  * Return the space in the log between the tail and the head.  The head
1194  * is passed in the cycle/bytes formal parms.  In the special case where
1195  * the reserve head has wrapped passed the tail, this calculation is no
1196  * longer valid.  In this case, just return 0 which means there is no space
1197  * in the log.  This works for all places where this function is called
1198  * with the reserve head.  Of course, if the write head were to ever
1199  * wrap the tail, we should blow up.  Rather than catch this case here,
1200  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1201  *
1202  * This code also handles the case where the reservation head is behind
1203  * the tail.  The details of this case are described below, but the end
1204  * result is that we return the size of the log as the amount of space left.
1205  */
1206 STATIC int
1207 xlog_space_left(
1208 	struct xlog	*log,
1209 	atomic64_t	*head)
1210 {
1211 	int		free_bytes;
1212 	int		tail_bytes;
1213 	int		tail_cycle;
1214 	int		head_cycle;
1215 	int		head_bytes;
1216 
1217 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1218 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1219 	tail_bytes = BBTOB(tail_bytes);
1220 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1221 		free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1222 	else if (tail_cycle + 1 < head_cycle)
1223 		return 0;
1224 	else if (tail_cycle < head_cycle) {
1225 		ASSERT(tail_cycle == (head_cycle - 1));
1226 		free_bytes = tail_bytes - head_bytes;
1227 	} else {
1228 		/*
1229 		 * The reservation head is behind the tail.
1230 		 * In this case we just want to return the size of the
1231 		 * log as the amount of space left.
1232 		 */
1233 		xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1234 		xfs_alert(log->l_mp,
1235 			  "  tail_cycle = %d, tail_bytes = %d",
1236 			  tail_cycle, tail_bytes);
1237 		xfs_alert(log->l_mp,
1238 			  "  GH   cycle = %d, GH   bytes = %d",
1239 			  head_cycle, head_bytes);
1240 		ASSERT(0);
1241 		free_bytes = log->l_logsize;
1242 	}
1243 	return free_bytes;
1244 }
1245 
1246 
1247 /*
1248  * Log function which is called when an io completes.
1249  *
1250  * The log manager needs its own routine, in order to control what
1251  * happens with the buffer after the write completes.
1252  */
1253 static void
1254 xlog_iodone(xfs_buf_t *bp)
1255 {
1256 	struct xlog_in_core	*iclog = bp->b_log_item;
1257 	struct xlog		*l = iclog->ic_log;
1258 	int			aborted = 0;
1259 
1260 	/*
1261 	 * Race to shutdown the filesystem if we see an error or the iclog is in
1262 	 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1263 	 * CRC errors into log recovery.
1264 	 */
1265 	if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR) ||
1266 	    iclog->ic_state & XLOG_STATE_IOABORT) {
1267 		if (iclog->ic_state & XLOG_STATE_IOABORT)
1268 			iclog->ic_state &= ~XLOG_STATE_IOABORT;
1269 
1270 		xfs_buf_ioerror_alert(bp, __func__);
1271 		xfs_buf_stale(bp);
1272 		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1273 		/*
1274 		 * This flag will be propagated to the trans-committed
1275 		 * callback routines to let them know that the log-commit
1276 		 * didn't succeed.
1277 		 */
1278 		aborted = XFS_LI_ABORTED;
1279 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1280 		aborted = XFS_LI_ABORTED;
1281 	}
1282 
1283 	/* log I/O is always issued ASYNC */
1284 	ASSERT(bp->b_flags & XBF_ASYNC);
1285 	xlog_state_done_syncing(iclog, aborted);
1286 
1287 	/*
1288 	 * drop the buffer lock now that we are done. Nothing references
1289 	 * the buffer after this, so an unmount waiting on this lock can now
1290 	 * tear it down safely. As such, it is unsafe to reference the buffer
1291 	 * (bp) after the unlock as we could race with it being freed.
1292 	 */
1293 	xfs_buf_unlock(bp);
1294 }
1295 
1296 /*
1297  * Return size of each in-core log record buffer.
1298  *
1299  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1300  *
1301  * If the filesystem blocksize is too large, we may need to choose a
1302  * larger size since the directory code currently logs entire blocks.
1303  */
1304 
1305 STATIC void
1306 xlog_get_iclog_buffer_size(
1307 	struct xfs_mount	*mp,
1308 	struct xlog		*log)
1309 {
1310 	int size;
1311 	int xhdrs;
1312 
1313 	if (mp->m_logbufs <= 0)
1314 		log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1315 	else
1316 		log->l_iclog_bufs = mp->m_logbufs;
1317 
1318 	/*
1319 	 * Buffer size passed in from mount system call.
1320 	 */
1321 	if (mp->m_logbsize > 0) {
1322 		size = log->l_iclog_size = mp->m_logbsize;
1323 		log->l_iclog_size_log = 0;
1324 		while (size != 1) {
1325 			log->l_iclog_size_log++;
1326 			size >>= 1;
1327 		}
1328 
1329 		if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1330 			/* # headers = size / 32k
1331 			 * one header holds cycles from 32k of data
1332 			 */
1333 
1334 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1335 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1336 				xhdrs++;
1337 			log->l_iclog_hsize = xhdrs << BBSHIFT;
1338 			log->l_iclog_heads = xhdrs;
1339 		} else {
1340 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1341 			log->l_iclog_hsize = BBSIZE;
1342 			log->l_iclog_heads = 1;
1343 		}
1344 		goto done;
1345 	}
1346 
1347 	/* All machines use 32kB buffers by default. */
1348 	log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1349 	log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1350 
1351 	/* the default log size is 16k or 32k which is one header sector */
1352 	log->l_iclog_hsize = BBSIZE;
1353 	log->l_iclog_heads = 1;
1354 
1355 done:
1356 	/* are we being asked to make the sizes selected above visible? */
1357 	if (mp->m_logbufs == 0)
1358 		mp->m_logbufs = log->l_iclog_bufs;
1359 	if (mp->m_logbsize == 0)
1360 		mp->m_logbsize = log->l_iclog_size;
1361 }	/* xlog_get_iclog_buffer_size */
1362 
1363 
1364 void
1365 xfs_log_work_queue(
1366 	struct xfs_mount        *mp)
1367 {
1368 	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1369 				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1370 }
1371 
1372 /*
1373  * Every sync period we need to unpin all items in the AIL and push them to
1374  * disk. If there is nothing dirty, then we might need to cover the log to
1375  * indicate that the filesystem is idle.
1376  */
1377 static void
1378 xfs_log_worker(
1379 	struct work_struct	*work)
1380 {
1381 	struct xlog		*log = container_of(to_delayed_work(work),
1382 						struct xlog, l_work);
1383 	struct xfs_mount	*mp = log->l_mp;
1384 
1385 	/* dgc: errors ignored - not fatal and nowhere to report them */
1386 	if (xfs_log_need_covered(mp)) {
1387 		/*
1388 		 * Dump a transaction into the log that contains no real change.
1389 		 * This is needed to stamp the current tail LSN into the log
1390 		 * during the covering operation.
1391 		 *
1392 		 * We cannot use an inode here for this - that will push dirty
1393 		 * state back up into the VFS and then periodic inode flushing
1394 		 * will prevent log covering from making progress. Hence we
1395 		 * synchronously log the superblock instead to ensure the
1396 		 * superblock is immediately unpinned and can be written back.
1397 		 */
1398 		xfs_sync_sb(mp, true);
1399 	} else
1400 		xfs_log_force(mp, 0);
1401 
1402 	/* start pushing all the metadata that is currently dirty */
1403 	xfs_ail_push_all(mp->m_ail);
1404 
1405 	/* queue us up again */
1406 	xfs_log_work_queue(mp);
1407 }
1408 
1409 /*
1410  * This routine initializes some of the log structure for a given mount point.
1411  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1412  * some other stuff may be filled in too.
1413  */
1414 STATIC struct xlog *
1415 xlog_alloc_log(
1416 	struct xfs_mount	*mp,
1417 	struct xfs_buftarg	*log_target,
1418 	xfs_daddr_t		blk_offset,
1419 	int			num_bblks)
1420 {
1421 	struct xlog		*log;
1422 	xlog_rec_header_t	*head;
1423 	xlog_in_core_t		**iclogp;
1424 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1425 	xfs_buf_t		*bp;
1426 	int			i;
1427 	int			error = -ENOMEM;
1428 	uint			log2_size = 0;
1429 
1430 	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1431 	if (!log) {
1432 		xfs_warn(mp, "Log allocation failed: No memory!");
1433 		goto out;
1434 	}
1435 
1436 	log->l_mp	   = mp;
1437 	log->l_targ	   = log_target;
1438 	log->l_logsize     = BBTOB(num_bblks);
1439 	log->l_logBBstart  = blk_offset;
1440 	log->l_logBBsize   = num_bblks;
1441 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1442 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1443 	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1444 
1445 	log->l_prev_block  = -1;
1446 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1447 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1448 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1449 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1450 
1451 	xlog_grant_head_init(&log->l_reserve_head);
1452 	xlog_grant_head_init(&log->l_write_head);
1453 
1454 	error = -EFSCORRUPTED;
1455 	if (xfs_sb_version_hassector(&mp->m_sb)) {
1456 	        log2_size = mp->m_sb.sb_logsectlog;
1457 		if (log2_size < BBSHIFT) {
1458 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1459 				log2_size, BBSHIFT);
1460 			goto out_free_log;
1461 		}
1462 
1463 	        log2_size -= BBSHIFT;
1464 		if (log2_size > mp->m_sectbb_log) {
1465 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1466 				log2_size, mp->m_sectbb_log);
1467 			goto out_free_log;
1468 		}
1469 
1470 		/* for larger sector sizes, must have v2 or external log */
1471 		if (log2_size && log->l_logBBstart > 0 &&
1472 			    !xfs_sb_version_haslogv2(&mp->m_sb)) {
1473 			xfs_warn(mp,
1474 		"log sector size (0x%x) invalid for configuration.",
1475 				log2_size);
1476 			goto out_free_log;
1477 		}
1478 	}
1479 	log->l_sectBBsize = 1 << log2_size;
1480 
1481 	xlog_get_iclog_buffer_size(mp, log);
1482 
1483 	/*
1484 	 * Use a NULL block for the extra log buffer used during splits so that
1485 	 * it will trigger errors if we ever try to do IO on it without first
1486 	 * having set it up properly.
1487 	 */
1488 	error = -ENOMEM;
1489 	bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1490 			   BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1491 	if (!bp)
1492 		goto out_free_log;
1493 
1494 	/*
1495 	 * The iclogbuf buffer locks are held over IO but we are not going to do
1496 	 * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1497 	 * when appropriately.
1498 	 */
1499 	ASSERT(xfs_buf_islocked(bp));
1500 	xfs_buf_unlock(bp);
1501 
1502 	/* use high priority wq for log I/O completion */
1503 	bp->b_ioend_wq = mp->m_log_workqueue;
1504 	bp->b_iodone = xlog_iodone;
1505 	log->l_xbuf = bp;
1506 
1507 	spin_lock_init(&log->l_icloglock);
1508 	init_waitqueue_head(&log->l_flush_wait);
1509 
1510 	iclogp = &log->l_iclog;
1511 	/*
1512 	 * The amount of memory to allocate for the iclog structure is
1513 	 * rather funky due to the way the structure is defined.  It is
1514 	 * done this way so that we can use different sizes for machines
1515 	 * with different amounts of memory.  See the definition of
1516 	 * xlog_in_core_t in xfs_log_priv.h for details.
1517 	 */
1518 	ASSERT(log->l_iclog_size >= 4096);
1519 	for (i=0; i < log->l_iclog_bufs; i++) {
1520 		*iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1521 		if (!*iclogp)
1522 			goto out_free_iclog;
1523 
1524 		iclog = *iclogp;
1525 		iclog->ic_prev = prev_iclog;
1526 		prev_iclog = iclog;
1527 
1528 		bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1529 					  BTOBB(log->l_iclog_size),
1530 					  XBF_NO_IOACCT);
1531 		if (!bp)
1532 			goto out_free_iclog;
1533 
1534 		ASSERT(xfs_buf_islocked(bp));
1535 		xfs_buf_unlock(bp);
1536 
1537 		/* use high priority wq for log I/O completion */
1538 		bp->b_ioend_wq = mp->m_log_workqueue;
1539 		bp->b_iodone = xlog_iodone;
1540 		iclog->ic_bp = bp;
1541 		iclog->ic_data = bp->b_addr;
1542 #ifdef DEBUG
1543 		log->l_iclog_bak[i] = &iclog->ic_header;
1544 #endif
1545 		head = &iclog->ic_header;
1546 		memset(head, 0, sizeof(xlog_rec_header_t));
1547 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1548 		head->h_version = cpu_to_be32(
1549 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1550 		head->h_size = cpu_to_be32(log->l_iclog_size);
1551 		/* new fields */
1552 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1553 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1554 
1555 		iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1556 		iclog->ic_state = XLOG_STATE_ACTIVE;
1557 		iclog->ic_log = log;
1558 		atomic_set(&iclog->ic_refcnt, 0);
1559 		spin_lock_init(&iclog->ic_callback_lock);
1560 		iclog->ic_callback_tail = &(iclog->ic_callback);
1561 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1562 
1563 		init_waitqueue_head(&iclog->ic_force_wait);
1564 		init_waitqueue_head(&iclog->ic_write_wait);
1565 
1566 		iclogp = &iclog->ic_next;
1567 	}
1568 	*iclogp = log->l_iclog;			/* complete ring */
1569 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1570 
1571 	error = xlog_cil_init(log);
1572 	if (error)
1573 		goto out_free_iclog;
1574 	return log;
1575 
1576 out_free_iclog:
1577 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1578 		prev_iclog = iclog->ic_next;
1579 		if (iclog->ic_bp)
1580 			xfs_buf_free(iclog->ic_bp);
1581 		kmem_free(iclog);
1582 	}
1583 	spinlock_destroy(&log->l_icloglock);
1584 	xfs_buf_free(log->l_xbuf);
1585 out_free_log:
1586 	kmem_free(log);
1587 out:
1588 	return ERR_PTR(error);
1589 }	/* xlog_alloc_log */
1590 
1591 
1592 /*
1593  * Write out the commit record of a transaction associated with the given
1594  * ticket.  Return the lsn of the commit record.
1595  */
1596 STATIC int
1597 xlog_commit_record(
1598 	struct xlog		*log,
1599 	struct xlog_ticket	*ticket,
1600 	struct xlog_in_core	**iclog,
1601 	xfs_lsn_t		*commitlsnp)
1602 {
1603 	struct xfs_mount *mp = log->l_mp;
1604 	int	error;
1605 	struct xfs_log_iovec reg = {
1606 		.i_addr = NULL,
1607 		.i_len = 0,
1608 		.i_type = XLOG_REG_TYPE_COMMIT,
1609 	};
1610 	struct xfs_log_vec vec = {
1611 		.lv_niovecs = 1,
1612 		.lv_iovecp = &reg,
1613 	};
1614 
1615 	ASSERT_ALWAYS(iclog);
1616 	error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1617 					XLOG_COMMIT_TRANS);
1618 	if (error)
1619 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1620 	return error;
1621 }
1622 
1623 /*
1624  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1625  * log space.  This code pushes on the lsn which would supposedly free up
1626  * the 25% which we want to leave free.  We may need to adopt a policy which
1627  * pushes on an lsn which is further along in the log once we reach the high
1628  * water mark.  In this manner, we would be creating a low water mark.
1629  */
1630 STATIC void
1631 xlog_grant_push_ail(
1632 	struct xlog	*log,
1633 	int		need_bytes)
1634 {
1635 	xfs_lsn_t	threshold_lsn = 0;
1636 	xfs_lsn_t	last_sync_lsn;
1637 	int		free_blocks;
1638 	int		free_bytes;
1639 	int		threshold_block;
1640 	int		threshold_cycle;
1641 	int		free_threshold;
1642 
1643 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1644 
1645 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1646 	free_blocks = BTOBBT(free_bytes);
1647 
1648 	/*
1649 	 * Set the threshold for the minimum number of free blocks in the
1650 	 * log to the maximum of what the caller needs, one quarter of the
1651 	 * log, and 256 blocks.
1652 	 */
1653 	free_threshold = BTOBB(need_bytes);
1654 	free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1655 	free_threshold = max(free_threshold, 256);
1656 	if (free_blocks >= free_threshold)
1657 		return;
1658 
1659 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1660 						&threshold_block);
1661 	threshold_block += free_threshold;
1662 	if (threshold_block >= log->l_logBBsize) {
1663 		threshold_block -= log->l_logBBsize;
1664 		threshold_cycle += 1;
1665 	}
1666 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1667 					threshold_block);
1668 	/*
1669 	 * Don't pass in an lsn greater than the lsn of the last
1670 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1671 	 * so that it doesn't change between the compare and the set.
1672 	 */
1673 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1674 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1675 		threshold_lsn = last_sync_lsn;
1676 
1677 	/*
1678 	 * Get the transaction layer to kick the dirty buffers out to
1679 	 * disk asynchronously. No point in trying to do this if
1680 	 * the filesystem is shutting down.
1681 	 */
1682 	if (!XLOG_FORCED_SHUTDOWN(log))
1683 		xfs_ail_push(log->l_ailp, threshold_lsn);
1684 }
1685 
1686 /*
1687  * Stamp cycle number in every block
1688  */
1689 STATIC void
1690 xlog_pack_data(
1691 	struct xlog		*log,
1692 	struct xlog_in_core	*iclog,
1693 	int			roundoff)
1694 {
1695 	int			i, j, k;
1696 	int			size = iclog->ic_offset + roundoff;
1697 	__be32			cycle_lsn;
1698 	char			*dp;
1699 
1700 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1701 
1702 	dp = iclog->ic_datap;
1703 	for (i = 0; i < BTOBB(size); i++) {
1704 		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1705 			break;
1706 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1707 		*(__be32 *)dp = cycle_lsn;
1708 		dp += BBSIZE;
1709 	}
1710 
1711 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1712 		xlog_in_core_2_t *xhdr = iclog->ic_data;
1713 
1714 		for ( ; i < BTOBB(size); i++) {
1715 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1716 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1717 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1718 			*(__be32 *)dp = cycle_lsn;
1719 			dp += BBSIZE;
1720 		}
1721 
1722 		for (i = 1; i < log->l_iclog_heads; i++)
1723 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1724 	}
1725 }
1726 
1727 /*
1728  * Calculate the checksum for a log buffer.
1729  *
1730  * This is a little more complicated than it should be because the various
1731  * headers and the actual data are non-contiguous.
1732  */
1733 __le32
1734 xlog_cksum(
1735 	struct xlog		*log,
1736 	struct xlog_rec_header	*rhead,
1737 	char			*dp,
1738 	int			size)
1739 {
1740 	uint32_t		crc;
1741 
1742 	/* first generate the crc for the record header ... */
1743 	crc = xfs_start_cksum_update((char *)rhead,
1744 			      sizeof(struct xlog_rec_header),
1745 			      offsetof(struct xlog_rec_header, h_crc));
1746 
1747 	/* ... then for additional cycle data for v2 logs ... */
1748 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1749 		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1750 		int		i;
1751 		int		xheads;
1752 
1753 		xheads = size / XLOG_HEADER_CYCLE_SIZE;
1754 		if (size % XLOG_HEADER_CYCLE_SIZE)
1755 			xheads++;
1756 
1757 		for (i = 1; i < xheads; i++) {
1758 			crc = crc32c(crc, &xhdr[i].hic_xheader,
1759 				     sizeof(struct xlog_rec_ext_header));
1760 		}
1761 	}
1762 
1763 	/* ... and finally for the payload */
1764 	crc = crc32c(crc, dp, size);
1765 
1766 	return xfs_end_cksum(crc);
1767 }
1768 
1769 /*
1770  * The bdstrat callback function for log bufs. This gives us a central
1771  * place to trap bufs in case we get hit by a log I/O error and need to
1772  * shutdown. Actually, in practice, even when we didn't get a log error,
1773  * we transition the iclogs to IOERROR state *after* flushing all existing
1774  * iclogs to disk. This is because we don't want anymore new transactions to be
1775  * started or completed afterwards.
1776  *
1777  * We lock the iclogbufs here so that we can serialise against IO completion
1778  * during unmount. We might be processing a shutdown triggered during unmount,
1779  * and that can occur asynchronously to the unmount thread, and hence we need to
1780  * ensure that completes before tearing down the iclogbufs. Hence we need to
1781  * hold the buffer lock across the log IO to acheive that.
1782  */
1783 STATIC int
1784 xlog_bdstrat(
1785 	struct xfs_buf		*bp)
1786 {
1787 	struct xlog_in_core	*iclog = bp->b_log_item;
1788 
1789 	xfs_buf_lock(bp);
1790 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
1791 		xfs_buf_ioerror(bp, -EIO);
1792 		xfs_buf_stale(bp);
1793 		xfs_buf_ioend(bp);
1794 		/*
1795 		 * It would seem logical to return EIO here, but we rely on
1796 		 * the log state machine to propagate I/O errors instead of
1797 		 * doing it here. Similarly, IO completion will unlock the
1798 		 * buffer, so we don't do it here.
1799 		 */
1800 		return 0;
1801 	}
1802 
1803 	xfs_buf_submit(bp);
1804 	return 0;
1805 }
1806 
1807 /*
1808  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1809  * fashion.  Previously, we should have moved the current iclog
1810  * ptr in the log to point to the next available iclog.  This allows further
1811  * write to continue while this code syncs out an iclog ready to go.
1812  * Before an in-core log can be written out, the data section must be scanned
1813  * to save away the 1st word of each BBSIZE block into the header.  We replace
1814  * it with the current cycle count.  Each BBSIZE block is tagged with the
1815  * cycle count because there in an implicit assumption that drives will
1816  * guarantee that entire 512 byte blocks get written at once.  In other words,
1817  * we can't have part of a 512 byte block written and part not written.  By
1818  * tagging each block, we will know which blocks are valid when recovering
1819  * after an unclean shutdown.
1820  *
1821  * This routine is single threaded on the iclog.  No other thread can be in
1822  * this routine with the same iclog.  Changing contents of iclog can there-
1823  * fore be done without grabbing the state machine lock.  Updating the global
1824  * log will require grabbing the lock though.
1825  *
1826  * The entire log manager uses a logical block numbering scheme.  Only
1827  * log_sync (and then only bwrite()) know about the fact that the log may
1828  * not start with block zero on a given device.  The log block start offset
1829  * is added immediately before calling bwrite().
1830  */
1831 
1832 STATIC int
1833 xlog_sync(
1834 	struct xlog		*log,
1835 	struct xlog_in_core	*iclog)
1836 {
1837 	xfs_buf_t	*bp;
1838 	int		i;
1839 	uint		count;		/* byte count of bwrite */
1840 	uint		count_init;	/* initial count before roundup */
1841 	int		roundoff;       /* roundoff to BB or stripe */
1842 	int		split = 0;	/* split write into two regions */
1843 	int		error;
1844 	int		v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1845 	int		size;
1846 
1847 	XFS_STATS_INC(log->l_mp, xs_log_writes);
1848 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1849 
1850 	/* Add for LR header */
1851 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1852 
1853 	/* Round out the log write size */
1854 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1855 		/* we have a v2 stripe unit to use */
1856 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1857 	} else {
1858 		count = BBTOB(BTOBB(count_init));
1859 	}
1860 	roundoff = count - count_init;
1861 	ASSERT(roundoff >= 0);
1862 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1863                 roundoff < log->l_mp->m_sb.sb_logsunit)
1864 		||
1865 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1866 		 roundoff < BBTOB(1)));
1867 
1868 	/* move grant heads by roundoff in sync */
1869 	xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1870 	xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1871 
1872 	/* put cycle number in every block */
1873 	xlog_pack_data(log, iclog, roundoff);
1874 
1875 	/* real byte length */
1876 	size = iclog->ic_offset;
1877 	if (v2)
1878 		size += roundoff;
1879 	iclog->ic_header.h_len = cpu_to_be32(size);
1880 
1881 	bp = iclog->ic_bp;
1882 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1883 
1884 	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1885 
1886 	/* Do we need to split this write into 2 parts? */
1887 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1888 		char		*dptr;
1889 
1890 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1891 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1892 		iclog->ic_bwritecnt = 2;
1893 
1894 		/*
1895 		 * Bump the cycle numbers at the start of each block in the
1896 		 * part of the iclog that ends up in the buffer that gets
1897 		 * written to the start of the log.
1898 		 *
1899 		 * Watch out for the header magic number case, though.
1900 		 */
1901 		dptr = (char *)&iclog->ic_header + count;
1902 		for (i = 0; i < split; i += BBSIZE) {
1903 			uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1904 			if (++cycle == XLOG_HEADER_MAGIC_NUM)
1905 				cycle++;
1906 			*(__be32 *)dptr = cpu_to_be32(cycle);
1907 
1908 			dptr += BBSIZE;
1909 		}
1910 	} else {
1911 		iclog->ic_bwritecnt = 1;
1912 	}
1913 
1914 	/* calculcate the checksum */
1915 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1916 					    iclog->ic_datap, size);
1917 	/*
1918 	 * Intentionally corrupt the log record CRC based on the error injection
1919 	 * frequency, if defined. This facilitates testing log recovery in the
1920 	 * event of torn writes. Hence, set the IOABORT state to abort the log
1921 	 * write on I/O completion and shutdown the fs. The subsequent mount
1922 	 * detects the bad CRC and attempts to recover.
1923 	 */
1924 	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1925 		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1926 		iclog->ic_state |= XLOG_STATE_IOABORT;
1927 		xfs_warn(log->l_mp,
1928 	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1929 			 be64_to_cpu(iclog->ic_header.h_lsn));
1930 	}
1931 
1932 	bp->b_io_length = BTOBB(count);
1933 	bp->b_log_item = iclog;
1934 	bp->b_flags &= ~XBF_FLUSH;
1935 	bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1936 
1937 	/*
1938 	 * Flush the data device before flushing the log to make sure all meta
1939 	 * data written back from the AIL actually made it to disk before
1940 	 * stamping the new log tail LSN into the log buffer.  For an external
1941 	 * log we need to issue the flush explicitly, and unfortunately
1942 	 * synchronously here; for an internal log we can simply use the block
1943 	 * layer state machine for preflushes.
1944 	 */
1945 	if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1946 		xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1947 	else
1948 		bp->b_flags |= XBF_FLUSH;
1949 
1950 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1951 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1952 
1953 	xlog_verify_iclog(log, iclog, count, true);
1954 
1955 	/* account for log which doesn't start at block #0 */
1956 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1957 
1958 	/*
1959 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1960 	 * is shutting down.
1961 	 */
1962 	error = xlog_bdstrat(bp);
1963 	if (error) {
1964 		xfs_buf_ioerror_alert(bp, "xlog_sync");
1965 		return error;
1966 	}
1967 	if (split) {
1968 		bp = iclog->ic_log->l_xbuf;
1969 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1970 		xfs_buf_associate_memory(bp,
1971 				(char *)&iclog->ic_header + count, split);
1972 		bp->b_log_item = iclog;
1973 		bp->b_flags &= ~XBF_FLUSH;
1974 		bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1975 
1976 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1977 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1978 
1979 		/* account for internal log which doesn't start at block #0 */
1980 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1981 		error = xlog_bdstrat(bp);
1982 		if (error) {
1983 			xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1984 			return error;
1985 		}
1986 	}
1987 	return 0;
1988 }	/* xlog_sync */
1989 
1990 /*
1991  * Deallocate a log structure
1992  */
1993 STATIC void
1994 xlog_dealloc_log(
1995 	struct xlog	*log)
1996 {
1997 	xlog_in_core_t	*iclog, *next_iclog;
1998 	int		i;
1999 
2000 	xlog_cil_destroy(log);
2001 
2002 	/*
2003 	 * Cycle all the iclogbuf locks to make sure all log IO completion
2004 	 * is done before we tear down these buffers.
2005 	 */
2006 	iclog = log->l_iclog;
2007 	for (i = 0; i < log->l_iclog_bufs; i++) {
2008 		xfs_buf_lock(iclog->ic_bp);
2009 		xfs_buf_unlock(iclog->ic_bp);
2010 		iclog = iclog->ic_next;
2011 	}
2012 
2013 	/*
2014 	 * Always need to ensure that the extra buffer does not point to memory
2015 	 * owned by another log buffer before we free it. Also, cycle the lock
2016 	 * first to ensure we've completed IO on it.
2017 	 */
2018 	xfs_buf_lock(log->l_xbuf);
2019 	xfs_buf_unlock(log->l_xbuf);
2020 	xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
2021 	xfs_buf_free(log->l_xbuf);
2022 
2023 	iclog = log->l_iclog;
2024 	for (i = 0; i < log->l_iclog_bufs; i++) {
2025 		xfs_buf_free(iclog->ic_bp);
2026 		next_iclog = iclog->ic_next;
2027 		kmem_free(iclog);
2028 		iclog = next_iclog;
2029 	}
2030 	spinlock_destroy(&log->l_icloglock);
2031 
2032 	log->l_mp->m_log = NULL;
2033 	kmem_free(log);
2034 }	/* xlog_dealloc_log */
2035 
2036 /*
2037  * Update counters atomically now that memcpy is done.
2038  */
2039 /* ARGSUSED */
2040 static inline void
2041 xlog_state_finish_copy(
2042 	struct xlog		*log,
2043 	struct xlog_in_core	*iclog,
2044 	int			record_cnt,
2045 	int			copy_bytes)
2046 {
2047 	spin_lock(&log->l_icloglock);
2048 
2049 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2050 	iclog->ic_offset += copy_bytes;
2051 
2052 	spin_unlock(&log->l_icloglock);
2053 }	/* xlog_state_finish_copy */
2054 
2055 
2056 
2057 
2058 /*
2059  * print out info relating to regions written which consume
2060  * the reservation
2061  */
2062 void
2063 xlog_print_tic_res(
2064 	struct xfs_mount	*mp,
2065 	struct xlog_ticket	*ticket)
2066 {
2067 	uint i;
2068 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2069 
2070 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
2071 #define REG_TYPE_STR(type, str)	[XLOG_REG_TYPE_##type] = str
2072 	static char *res_type_str[] = {
2073 	    REG_TYPE_STR(BFORMAT, "bformat"),
2074 	    REG_TYPE_STR(BCHUNK, "bchunk"),
2075 	    REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2076 	    REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2077 	    REG_TYPE_STR(IFORMAT, "iformat"),
2078 	    REG_TYPE_STR(ICORE, "icore"),
2079 	    REG_TYPE_STR(IEXT, "iext"),
2080 	    REG_TYPE_STR(IBROOT, "ibroot"),
2081 	    REG_TYPE_STR(ILOCAL, "ilocal"),
2082 	    REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2083 	    REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2084 	    REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2085 	    REG_TYPE_STR(QFORMAT, "qformat"),
2086 	    REG_TYPE_STR(DQUOT, "dquot"),
2087 	    REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2088 	    REG_TYPE_STR(LRHEADER, "LR header"),
2089 	    REG_TYPE_STR(UNMOUNT, "unmount"),
2090 	    REG_TYPE_STR(COMMIT, "commit"),
2091 	    REG_TYPE_STR(TRANSHDR, "trans header"),
2092 	    REG_TYPE_STR(ICREATE, "inode create"),
2093 	    REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2094 	    REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2095 	    REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2096 	    REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2097 	    REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2098 	    REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2099 	};
2100 	BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2101 #undef REG_TYPE_STR
2102 
2103 	xfs_warn(mp, "ticket reservation summary:");
2104 	xfs_warn(mp, "  unit res    = %d bytes",
2105 		 ticket->t_unit_res);
2106 	xfs_warn(mp, "  current res = %d bytes",
2107 		 ticket->t_curr_res);
2108 	xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2109 		 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2110 	xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2111 		 ticket->t_res_num_ophdrs, ophdr_spc);
2112 	xfs_warn(mp, "  ophdr + reg = %u bytes",
2113 		 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2114 	xfs_warn(mp, "  num regions = %u",
2115 		 ticket->t_res_num);
2116 
2117 	for (i = 0; i < ticket->t_res_num; i++) {
2118 		uint r_type = ticket->t_res_arr[i].r_type;
2119 		xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2120 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2121 			    "bad-rtype" : res_type_str[r_type]),
2122 			    ticket->t_res_arr[i].r_len);
2123 	}
2124 }
2125 
2126 /*
2127  * Print a summary of the transaction.
2128  */
2129 void
2130 xlog_print_trans(
2131 	struct xfs_trans	*tp)
2132 {
2133 	struct xfs_mount	*mp = tp->t_mountp;
2134 	struct xfs_log_item	*lip;
2135 
2136 	/* dump core transaction and ticket info */
2137 	xfs_warn(mp, "transaction summary:");
2138 	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2139 	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2140 	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2141 
2142 	xlog_print_tic_res(mp, tp->t_ticket);
2143 
2144 	/* dump each log item */
2145 	list_for_each_entry(lip, &tp->t_items, li_trans) {
2146 		struct xfs_log_vec	*lv = lip->li_lv;
2147 		struct xfs_log_iovec	*vec;
2148 		int			i;
2149 
2150 		xfs_warn(mp, "log item: ");
2151 		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2152 		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
2153 		if (!lv)
2154 			continue;
2155 		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2156 		xfs_warn(mp, "  size	= %d", lv->lv_size);
2157 		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2158 		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2159 
2160 		/* dump each iovec for the log item */
2161 		vec = lv->lv_iovecp;
2162 		for (i = 0; i < lv->lv_niovecs; i++) {
2163 			int dumplen = min(vec->i_len, 32);
2164 
2165 			xfs_warn(mp, "  iovec[%d]", i);
2166 			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2167 			xfs_warn(mp, "    len	= %d", vec->i_len);
2168 			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2169 			xfs_hex_dump(vec->i_addr, dumplen);
2170 
2171 			vec++;
2172 		}
2173 	}
2174 }
2175 
2176 /*
2177  * Calculate the potential space needed by the log vector.  Each region gets
2178  * its own xlog_op_header_t and may need to be double word aligned.
2179  */
2180 static int
2181 xlog_write_calc_vec_length(
2182 	struct xlog_ticket	*ticket,
2183 	struct xfs_log_vec	*log_vector)
2184 {
2185 	struct xfs_log_vec	*lv;
2186 	int			headers = 0;
2187 	int			len = 0;
2188 	int			i;
2189 
2190 	/* acct for start rec of xact */
2191 	if (ticket->t_flags & XLOG_TIC_INITED)
2192 		headers++;
2193 
2194 	for (lv = log_vector; lv; lv = lv->lv_next) {
2195 		/* we don't write ordered log vectors */
2196 		if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2197 			continue;
2198 
2199 		headers += lv->lv_niovecs;
2200 
2201 		for (i = 0; i < lv->lv_niovecs; i++) {
2202 			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
2203 
2204 			len += vecp->i_len;
2205 			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2206 		}
2207 	}
2208 
2209 	ticket->t_res_num_ophdrs += headers;
2210 	len += headers * sizeof(struct xlog_op_header);
2211 
2212 	return len;
2213 }
2214 
2215 /*
2216  * If first write for transaction, insert start record  We can't be trying to
2217  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2218  */
2219 static int
2220 xlog_write_start_rec(
2221 	struct xlog_op_header	*ophdr,
2222 	struct xlog_ticket	*ticket)
2223 {
2224 	if (!(ticket->t_flags & XLOG_TIC_INITED))
2225 		return 0;
2226 
2227 	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
2228 	ophdr->oh_clientid = ticket->t_clientid;
2229 	ophdr->oh_len = 0;
2230 	ophdr->oh_flags = XLOG_START_TRANS;
2231 	ophdr->oh_res2 = 0;
2232 
2233 	ticket->t_flags &= ~XLOG_TIC_INITED;
2234 
2235 	return sizeof(struct xlog_op_header);
2236 }
2237 
2238 static xlog_op_header_t *
2239 xlog_write_setup_ophdr(
2240 	struct xlog		*log,
2241 	struct xlog_op_header	*ophdr,
2242 	struct xlog_ticket	*ticket,
2243 	uint			flags)
2244 {
2245 	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2246 	ophdr->oh_clientid = ticket->t_clientid;
2247 	ophdr->oh_res2 = 0;
2248 
2249 	/* are we copying a commit or unmount record? */
2250 	ophdr->oh_flags = flags;
2251 
2252 	/*
2253 	 * We've seen logs corrupted with bad transaction client ids.  This
2254 	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2255 	 * and shut down the filesystem.
2256 	 */
2257 	switch (ophdr->oh_clientid)  {
2258 	case XFS_TRANSACTION:
2259 	case XFS_VOLUME:
2260 	case XFS_LOG:
2261 		break;
2262 	default:
2263 		xfs_warn(log->l_mp,
2264 			"Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2265 			ophdr->oh_clientid, ticket);
2266 		return NULL;
2267 	}
2268 
2269 	return ophdr;
2270 }
2271 
2272 /*
2273  * Set up the parameters of the region copy into the log. This has
2274  * to handle region write split across multiple log buffers - this
2275  * state is kept external to this function so that this code can
2276  * be written in an obvious, self documenting manner.
2277  */
2278 static int
2279 xlog_write_setup_copy(
2280 	struct xlog_ticket	*ticket,
2281 	struct xlog_op_header	*ophdr,
2282 	int			space_available,
2283 	int			space_required,
2284 	int			*copy_off,
2285 	int			*copy_len,
2286 	int			*last_was_partial_copy,
2287 	int			*bytes_consumed)
2288 {
2289 	int			still_to_copy;
2290 
2291 	still_to_copy = space_required - *bytes_consumed;
2292 	*copy_off = *bytes_consumed;
2293 
2294 	if (still_to_copy <= space_available) {
2295 		/* write of region completes here */
2296 		*copy_len = still_to_copy;
2297 		ophdr->oh_len = cpu_to_be32(*copy_len);
2298 		if (*last_was_partial_copy)
2299 			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2300 		*last_was_partial_copy = 0;
2301 		*bytes_consumed = 0;
2302 		return 0;
2303 	}
2304 
2305 	/* partial write of region, needs extra log op header reservation */
2306 	*copy_len = space_available;
2307 	ophdr->oh_len = cpu_to_be32(*copy_len);
2308 	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2309 	if (*last_was_partial_copy)
2310 		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2311 	*bytes_consumed += *copy_len;
2312 	(*last_was_partial_copy)++;
2313 
2314 	/* account for new log op header */
2315 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2316 	ticket->t_res_num_ophdrs++;
2317 
2318 	return sizeof(struct xlog_op_header);
2319 }
2320 
2321 static int
2322 xlog_write_copy_finish(
2323 	struct xlog		*log,
2324 	struct xlog_in_core	*iclog,
2325 	uint			flags,
2326 	int			*record_cnt,
2327 	int			*data_cnt,
2328 	int			*partial_copy,
2329 	int			*partial_copy_len,
2330 	int			log_offset,
2331 	struct xlog_in_core	**commit_iclog)
2332 {
2333 	if (*partial_copy) {
2334 		/*
2335 		 * This iclog has already been marked WANT_SYNC by
2336 		 * xlog_state_get_iclog_space.
2337 		 */
2338 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2339 		*record_cnt = 0;
2340 		*data_cnt = 0;
2341 		return xlog_state_release_iclog(log, iclog);
2342 	}
2343 
2344 	*partial_copy = 0;
2345 	*partial_copy_len = 0;
2346 
2347 	if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2348 		/* no more space in this iclog - push it. */
2349 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2350 		*record_cnt = 0;
2351 		*data_cnt = 0;
2352 
2353 		spin_lock(&log->l_icloglock);
2354 		xlog_state_want_sync(log, iclog);
2355 		spin_unlock(&log->l_icloglock);
2356 
2357 		if (!commit_iclog)
2358 			return xlog_state_release_iclog(log, iclog);
2359 		ASSERT(flags & XLOG_COMMIT_TRANS);
2360 		*commit_iclog = iclog;
2361 	}
2362 
2363 	return 0;
2364 }
2365 
2366 /*
2367  * Write some region out to in-core log
2368  *
2369  * This will be called when writing externally provided regions or when
2370  * writing out a commit record for a given transaction.
2371  *
2372  * General algorithm:
2373  *	1. Find total length of this write.  This may include adding to the
2374  *		lengths passed in.
2375  *	2. Check whether we violate the tickets reservation.
2376  *	3. While writing to this iclog
2377  *	    A. Reserve as much space in this iclog as can get
2378  *	    B. If this is first write, save away start lsn
2379  *	    C. While writing this region:
2380  *		1. If first write of transaction, write start record
2381  *		2. Write log operation header (header per region)
2382  *		3. Find out if we can fit entire region into this iclog
2383  *		4. Potentially, verify destination memcpy ptr
2384  *		5. Memcpy (partial) region
2385  *		6. If partial copy, release iclog; otherwise, continue
2386  *			copying more regions into current iclog
2387  *	4. Mark want sync bit (in simulation mode)
2388  *	5. Release iclog for potential flush to on-disk log.
2389  *
2390  * ERRORS:
2391  * 1.	Panic if reservation is overrun.  This should never happen since
2392  *	reservation amounts are generated internal to the filesystem.
2393  * NOTES:
2394  * 1. Tickets are single threaded data structures.
2395  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2396  *	syncing routine.  When a single log_write region needs to span
2397  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2398  *	on all log operation writes which don't contain the end of the
2399  *	region.  The XLOG_END_TRANS bit is used for the in-core log
2400  *	operation which contains the end of the continued log_write region.
2401  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2402  *	we don't really know exactly how much space will be used.  As a result,
2403  *	we don't update ic_offset until the end when we know exactly how many
2404  *	bytes have been written out.
2405  */
2406 int
2407 xlog_write(
2408 	struct xlog		*log,
2409 	struct xfs_log_vec	*log_vector,
2410 	struct xlog_ticket	*ticket,
2411 	xfs_lsn_t		*start_lsn,
2412 	struct xlog_in_core	**commit_iclog,
2413 	uint			flags)
2414 {
2415 	struct xlog_in_core	*iclog = NULL;
2416 	struct xfs_log_iovec	*vecp;
2417 	struct xfs_log_vec	*lv;
2418 	int			len;
2419 	int			index;
2420 	int			partial_copy = 0;
2421 	int			partial_copy_len = 0;
2422 	int			contwr = 0;
2423 	int			record_cnt = 0;
2424 	int			data_cnt = 0;
2425 	int			error;
2426 
2427 	*start_lsn = 0;
2428 
2429 	len = xlog_write_calc_vec_length(ticket, log_vector);
2430 
2431 	/*
2432 	 * Region headers and bytes are already accounted for.
2433 	 * We only need to take into account start records and
2434 	 * split regions in this function.
2435 	 */
2436 	if (ticket->t_flags & XLOG_TIC_INITED)
2437 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
2438 
2439 	/*
2440 	 * Commit record headers need to be accounted for. These
2441 	 * come in as separate writes so are easy to detect.
2442 	 */
2443 	if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2444 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
2445 
2446 	if (ticket->t_curr_res < 0) {
2447 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2448 		     "ctx ticket reservation ran out. Need to up reservation");
2449 		xlog_print_tic_res(log->l_mp, ticket);
2450 		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2451 	}
2452 
2453 	index = 0;
2454 	lv = log_vector;
2455 	vecp = lv->lv_iovecp;
2456 	while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2457 		void		*ptr;
2458 		int		log_offset;
2459 
2460 		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2461 						   &contwr, &log_offset);
2462 		if (error)
2463 			return error;
2464 
2465 		ASSERT(log_offset <= iclog->ic_size - 1);
2466 		ptr = iclog->ic_datap + log_offset;
2467 
2468 		/* start_lsn is the first lsn written to. That's all we need. */
2469 		if (!*start_lsn)
2470 			*start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2471 
2472 		/*
2473 		 * This loop writes out as many regions as can fit in the amount
2474 		 * of space which was allocated by xlog_state_get_iclog_space().
2475 		 */
2476 		while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2477 			struct xfs_log_iovec	*reg;
2478 			struct xlog_op_header	*ophdr;
2479 			int			start_rec_copy;
2480 			int			copy_len;
2481 			int			copy_off;
2482 			bool			ordered = false;
2483 
2484 			/* ordered log vectors have no regions to write */
2485 			if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2486 				ASSERT(lv->lv_niovecs == 0);
2487 				ordered = true;
2488 				goto next_lv;
2489 			}
2490 
2491 			reg = &vecp[index];
2492 			ASSERT(reg->i_len % sizeof(int32_t) == 0);
2493 			ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2494 
2495 			start_rec_copy = xlog_write_start_rec(ptr, ticket);
2496 			if (start_rec_copy) {
2497 				record_cnt++;
2498 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2499 						   start_rec_copy);
2500 			}
2501 
2502 			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2503 			if (!ophdr)
2504 				return -EIO;
2505 
2506 			xlog_write_adv_cnt(&ptr, &len, &log_offset,
2507 					   sizeof(struct xlog_op_header));
2508 
2509 			len += xlog_write_setup_copy(ticket, ophdr,
2510 						     iclog->ic_size-log_offset,
2511 						     reg->i_len,
2512 						     &copy_off, &copy_len,
2513 						     &partial_copy,
2514 						     &partial_copy_len);
2515 			xlog_verify_dest_ptr(log, ptr);
2516 
2517 			/*
2518 			 * Copy region.
2519 			 *
2520 			 * Unmount records just log an opheader, so can have
2521 			 * empty payloads with no data region to copy. Hence we
2522 			 * only copy the payload if the vector says it has data
2523 			 * to copy.
2524 			 */
2525 			ASSERT(copy_len >= 0);
2526 			if (copy_len > 0) {
2527 				memcpy(ptr, reg->i_addr + copy_off, copy_len);
2528 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2529 						   copy_len);
2530 			}
2531 			copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2532 			record_cnt++;
2533 			data_cnt += contwr ? copy_len : 0;
2534 
2535 			error = xlog_write_copy_finish(log, iclog, flags,
2536 						       &record_cnt, &data_cnt,
2537 						       &partial_copy,
2538 						       &partial_copy_len,
2539 						       log_offset,
2540 						       commit_iclog);
2541 			if (error)
2542 				return error;
2543 
2544 			/*
2545 			 * if we had a partial copy, we need to get more iclog
2546 			 * space but we don't want to increment the region
2547 			 * index because there is still more is this region to
2548 			 * write.
2549 			 *
2550 			 * If we completed writing this region, and we flushed
2551 			 * the iclog (indicated by resetting of the record
2552 			 * count), then we also need to get more log space. If
2553 			 * this was the last record, though, we are done and
2554 			 * can just return.
2555 			 */
2556 			if (partial_copy)
2557 				break;
2558 
2559 			if (++index == lv->lv_niovecs) {
2560 next_lv:
2561 				lv = lv->lv_next;
2562 				index = 0;
2563 				if (lv)
2564 					vecp = lv->lv_iovecp;
2565 			}
2566 			if (record_cnt == 0 && !ordered) {
2567 				if (!lv)
2568 					return 0;
2569 				break;
2570 			}
2571 		}
2572 	}
2573 
2574 	ASSERT(len == 0);
2575 
2576 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2577 	if (!commit_iclog)
2578 		return xlog_state_release_iclog(log, iclog);
2579 
2580 	ASSERT(flags & XLOG_COMMIT_TRANS);
2581 	*commit_iclog = iclog;
2582 	return 0;
2583 }
2584 
2585 
2586 /*****************************************************************************
2587  *
2588  *		State Machine functions
2589  *
2590  *****************************************************************************
2591  */
2592 
2593 /* Clean iclogs starting from the head.  This ordering must be
2594  * maintained, so an iclog doesn't become ACTIVE beyond one that
2595  * is SYNCING.  This is also required to maintain the notion that we use
2596  * a ordered wait queue to hold off would be writers to the log when every
2597  * iclog is trying to sync to disk.
2598  *
2599  * State Change: DIRTY -> ACTIVE
2600  */
2601 STATIC void
2602 xlog_state_clean_log(
2603 	struct xlog *log)
2604 {
2605 	xlog_in_core_t	*iclog;
2606 	int changed = 0;
2607 
2608 	iclog = log->l_iclog;
2609 	do {
2610 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
2611 			iclog->ic_state	= XLOG_STATE_ACTIVE;
2612 			iclog->ic_offset       = 0;
2613 			ASSERT(iclog->ic_callback == NULL);
2614 			/*
2615 			 * If the number of ops in this iclog indicate it just
2616 			 * contains the dummy transaction, we can
2617 			 * change state into IDLE (the second time around).
2618 			 * Otherwise we should change the state into
2619 			 * NEED a dummy.
2620 			 * We don't need to cover the dummy.
2621 			 */
2622 			if (!changed &&
2623 			   (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2624 			   		XLOG_COVER_OPS)) {
2625 				changed = 1;
2626 			} else {
2627 				/*
2628 				 * We have two dirty iclogs so start over
2629 				 * This could also be num of ops indicates
2630 				 * this is not the dummy going out.
2631 				 */
2632 				changed = 2;
2633 			}
2634 			iclog->ic_header.h_num_logops = 0;
2635 			memset(iclog->ic_header.h_cycle_data, 0,
2636 			      sizeof(iclog->ic_header.h_cycle_data));
2637 			iclog->ic_header.h_lsn = 0;
2638 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2639 			/* do nothing */;
2640 		else
2641 			break;	/* stop cleaning */
2642 		iclog = iclog->ic_next;
2643 	} while (iclog != log->l_iclog);
2644 
2645 	/* log is locked when we are called */
2646 	/*
2647 	 * Change state for the dummy log recording.
2648 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
2649 	 * we are done writing the dummy record.
2650 	 * If we are done with the second dummy recored (DONE2), then
2651 	 * we go to IDLE.
2652 	 */
2653 	if (changed) {
2654 		switch (log->l_covered_state) {
2655 		case XLOG_STATE_COVER_IDLE:
2656 		case XLOG_STATE_COVER_NEED:
2657 		case XLOG_STATE_COVER_NEED2:
2658 			log->l_covered_state = XLOG_STATE_COVER_NEED;
2659 			break;
2660 
2661 		case XLOG_STATE_COVER_DONE:
2662 			if (changed == 1)
2663 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
2664 			else
2665 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2666 			break;
2667 
2668 		case XLOG_STATE_COVER_DONE2:
2669 			if (changed == 1)
2670 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2671 			else
2672 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2673 			break;
2674 
2675 		default:
2676 			ASSERT(0);
2677 		}
2678 	}
2679 }	/* xlog_state_clean_log */
2680 
2681 STATIC xfs_lsn_t
2682 xlog_get_lowest_lsn(
2683 	struct xlog	*log)
2684 {
2685 	xlog_in_core_t  *lsn_log;
2686 	xfs_lsn_t	lowest_lsn, lsn;
2687 
2688 	lsn_log = log->l_iclog;
2689 	lowest_lsn = 0;
2690 	do {
2691 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2692 		lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2693 		if ((lsn && !lowest_lsn) ||
2694 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2695 			lowest_lsn = lsn;
2696 		}
2697 	    }
2698 	    lsn_log = lsn_log->ic_next;
2699 	} while (lsn_log != log->l_iclog);
2700 	return lowest_lsn;
2701 }
2702 
2703 
2704 STATIC void
2705 xlog_state_do_callback(
2706 	struct xlog		*log,
2707 	int			aborted,
2708 	struct xlog_in_core	*ciclog)
2709 {
2710 	xlog_in_core_t	   *iclog;
2711 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2712 						 * processed all iclogs once */
2713 	xfs_log_callback_t *cb, *cb_next;
2714 	int		   flushcnt = 0;
2715 	xfs_lsn_t	   lowest_lsn;
2716 	int		   ioerrors;	/* counter: iclogs with errors */
2717 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2718 	int		   funcdidcallbacks; /* flag: function did callbacks */
2719 	int		   repeats;	/* for issuing console warnings if
2720 					 * looping too many times */
2721 	int		   wake = 0;
2722 
2723 	spin_lock(&log->l_icloglock);
2724 	first_iclog = iclog = log->l_iclog;
2725 	ioerrors = 0;
2726 	funcdidcallbacks = 0;
2727 	repeats = 0;
2728 
2729 	do {
2730 		/*
2731 		 * Scan all iclogs starting with the one pointed to by the
2732 		 * log.  Reset this starting point each time the log is
2733 		 * unlocked (during callbacks).
2734 		 *
2735 		 * Keep looping through iclogs until one full pass is made
2736 		 * without running any callbacks.
2737 		 */
2738 		first_iclog = log->l_iclog;
2739 		iclog = log->l_iclog;
2740 		loopdidcallbacks = 0;
2741 		repeats++;
2742 
2743 		do {
2744 
2745 			/* skip all iclogs in the ACTIVE & DIRTY states */
2746 			if (iclog->ic_state &
2747 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2748 				iclog = iclog->ic_next;
2749 				continue;
2750 			}
2751 
2752 			/*
2753 			 * Between marking a filesystem SHUTDOWN and stopping
2754 			 * the log, we do flush all iclogs to disk (if there
2755 			 * wasn't a log I/O error). So, we do want things to
2756 			 * go smoothly in case of just a SHUTDOWN  w/o a
2757 			 * LOG_IO_ERROR.
2758 			 */
2759 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2760 				/*
2761 				 * Can only perform callbacks in order.  Since
2762 				 * this iclog is not in the DONE_SYNC/
2763 				 * DO_CALLBACK state, we skip the rest and
2764 				 * just try to clean up.  If we set our iclog
2765 				 * to DO_CALLBACK, we will not process it when
2766 				 * we retry since a previous iclog is in the
2767 				 * CALLBACK and the state cannot change since
2768 				 * we are holding the l_icloglock.
2769 				 */
2770 				if (!(iclog->ic_state &
2771 					(XLOG_STATE_DONE_SYNC |
2772 						 XLOG_STATE_DO_CALLBACK))) {
2773 					if (ciclog && (ciclog->ic_state ==
2774 							XLOG_STATE_DONE_SYNC)) {
2775 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2776 					}
2777 					break;
2778 				}
2779 				/*
2780 				 * We now have an iclog that is in either the
2781 				 * DO_CALLBACK or DONE_SYNC states. The other
2782 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2783 				 * caught by the above if and are going to
2784 				 * clean (i.e. we aren't doing their callbacks)
2785 				 * see the above if.
2786 				 */
2787 
2788 				/*
2789 				 * We will do one more check here to see if we
2790 				 * have chased our tail around.
2791 				 */
2792 
2793 				lowest_lsn = xlog_get_lowest_lsn(log);
2794 				if (lowest_lsn &&
2795 				    XFS_LSN_CMP(lowest_lsn,
2796 						be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2797 					iclog = iclog->ic_next;
2798 					continue; /* Leave this iclog for
2799 						   * another thread */
2800 				}
2801 
2802 				iclog->ic_state = XLOG_STATE_CALLBACK;
2803 
2804 
2805 				/*
2806 				 * Completion of a iclog IO does not imply that
2807 				 * a transaction has completed, as transactions
2808 				 * can be large enough to span many iclogs. We
2809 				 * cannot change the tail of the log half way
2810 				 * through a transaction as this may be the only
2811 				 * transaction in the log and moving th etail to
2812 				 * point to the middle of it will prevent
2813 				 * recovery from finding the start of the
2814 				 * transaction. Hence we should only update the
2815 				 * last_sync_lsn if this iclog contains
2816 				 * transaction completion callbacks on it.
2817 				 *
2818 				 * We have to do this before we drop the
2819 				 * icloglock to ensure we are the only one that
2820 				 * can update it.
2821 				 */
2822 				ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2823 					be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2824 				if (iclog->ic_callback)
2825 					atomic64_set(&log->l_last_sync_lsn,
2826 						be64_to_cpu(iclog->ic_header.h_lsn));
2827 
2828 			} else
2829 				ioerrors++;
2830 
2831 			spin_unlock(&log->l_icloglock);
2832 
2833 			/*
2834 			 * Keep processing entries in the callback list until
2835 			 * we come around and it is empty.  We need to
2836 			 * atomically see that the list is empty and change the
2837 			 * state to DIRTY so that we don't miss any more
2838 			 * callbacks being added.
2839 			 */
2840 			spin_lock(&iclog->ic_callback_lock);
2841 			cb = iclog->ic_callback;
2842 			while (cb) {
2843 				iclog->ic_callback_tail = &(iclog->ic_callback);
2844 				iclog->ic_callback = NULL;
2845 				spin_unlock(&iclog->ic_callback_lock);
2846 
2847 				/* perform callbacks in the order given */
2848 				for (; cb; cb = cb_next) {
2849 					cb_next = cb->cb_next;
2850 					cb->cb_func(cb->cb_arg, aborted);
2851 				}
2852 				spin_lock(&iclog->ic_callback_lock);
2853 				cb = iclog->ic_callback;
2854 			}
2855 
2856 			loopdidcallbacks++;
2857 			funcdidcallbacks++;
2858 
2859 			spin_lock(&log->l_icloglock);
2860 			ASSERT(iclog->ic_callback == NULL);
2861 			spin_unlock(&iclog->ic_callback_lock);
2862 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2863 				iclog->ic_state = XLOG_STATE_DIRTY;
2864 
2865 			/*
2866 			 * Transition from DIRTY to ACTIVE if applicable.
2867 			 * NOP if STATE_IOERROR.
2868 			 */
2869 			xlog_state_clean_log(log);
2870 
2871 			/* wake up threads waiting in xfs_log_force() */
2872 			wake_up_all(&iclog->ic_force_wait);
2873 
2874 			iclog = iclog->ic_next;
2875 		} while (first_iclog != iclog);
2876 
2877 		if (repeats > 5000) {
2878 			flushcnt += repeats;
2879 			repeats = 0;
2880 			xfs_warn(log->l_mp,
2881 				"%s: possible infinite loop (%d iterations)",
2882 				__func__, flushcnt);
2883 		}
2884 	} while (!ioerrors && loopdidcallbacks);
2885 
2886 #ifdef DEBUG
2887 	/*
2888 	 * Make one last gasp attempt to see if iclogs are being left in limbo.
2889 	 * If the above loop finds an iclog earlier than the current iclog and
2890 	 * in one of the syncing states, the current iclog is put into
2891 	 * DO_CALLBACK and the callbacks are deferred to the completion of the
2892 	 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2893 	 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2894 	 * states.
2895 	 *
2896 	 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2897 	 * for ic_state == SYNCING.
2898 	 */
2899 	if (funcdidcallbacks) {
2900 		first_iclog = iclog = log->l_iclog;
2901 		do {
2902 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2903 			/*
2904 			 * Terminate the loop if iclogs are found in states
2905 			 * which will cause other threads to clean up iclogs.
2906 			 *
2907 			 * SYNCING - i/o completion will go through logs
2908 			 * DONE_SYNC - interrupt thread should be waiting for
2909 			 *              l_icloglock
2910 			 * IOERROR - give up hope all ye who enter here
2911 			 */
2912 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2913 			    iclog->ic_state & XLOG_STATE_SYNCING ||
2914 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2915 			    iclog->ic_state == XLOG_STATE_IOERROR )
2916 				break;
2917 			iclog = iclog->ic_next;
2918 		} while (first_iclog != iclog);
2919 	}
2920 #endif
2921 
2922 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2923 		wake = 1;
2924 	spin_unlock(&log->l_icloglock);
2925 
2926 	if (wake)
2927 		wake_up_all(&log->l_flush_wait);
2928 }
2929 
2930 
2931 /*
2932  * Finish transitioning this iclog to the dirty state.
2933  *
2934  * Make sure that we completely execute this routine only when this is
2935  * the last call to the iclog.  There is a good chance that iclog flushes,
2936  * when we reach the end of the physical log, get turned into 2 separate
2937  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2938  * routine.  By using the reference count bwritecnt, we guarantee that only
2939  * the second completion goes through.
2940  *
2941  * Callbacks could take time, so they are done outside the scope of the
2942  * global state machine log lock.
2943  */
2944 STATIC void
2945 xlog_state_done_syncing(
2946 	xlog_in_core_t	*iclog,
2947 	int		aborted)
2948 {
2949 	struct xlog	   *log = iclog->ic_log;
2950 
2951 	spin_lock(&log->l_icloglock);
2952 
2953 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2954 	       iclog->ic_state == XLOG_STATE_IOERROR);
2955 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2956 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2957 
2958 
2959 	/*
2960 	 * If we got an error, either on the first buffer, or in the case of
2961 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2962 	 * and none should ever be attempted to be written to disk
2963 	 * again.
2964 	 */
2965 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2966 		if (--iclog->ic_bwritecnt == 1) {
2967 			spin_unlock(&log->l_icloglock);
2968 			return;
2969 		}
2970 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2971 	}
2972 
2973 	/*
2974 	 * Someone could be sleeping prior to writing out the next
2975 	 * iclog buffer, we wake them all, one will get to do the
2976 	 * I/O, the others get to wait for the result.
2977 	 */
2978 	wake_up_all(&iclog->ic_write_wait);
2979 	spin_unlock(&log->l_icloglock);
2980 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2981 }	/* xlog_state_done_syncing */
2982 
2983 
2984 /*
2985  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2986  * sleep.  We wait on the flush queue on the head iclog as that should be
2987  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2988  * we will wait here and all new writes will sleep until a sync completes.
2989  *
2990  * The in-core logs are used in a circular fashion. They are not used
2991  * out-of-order even when an iclog past the head is free.
2992  *
2993  * return:
2994  *	* log_offset where xlog_write() can start writing into the in-core
2995  *		log's data space.
2996  *	* in-core log pointer to which xlog_write() should write.
2997  *	* boolean indicating this is a continued write to an in-core log.
2998  *		If this is the last write, then the in-core log's offset field
2999  *		needs to be incremented, depending on the amount of data which
3000  *		is copied.
3001  */
3002 STATIC int
3003 xlog_state_get_iclog_space(
3004 	struct xlog		*log,
3005 	int			len,
3006 	struct xlog_in_core	**iclogp,
3007 	struct xlog_ticket	*ticket,
3008 	int			*continued_write,
3009 	int			*logoffsetp)
3010 {
3011 	int		  log_offset;
3012 	xlog_rec_header_t *head;
3013 	xlog_in_core_t	  *iclog;
3014 	int		  error;
3015 
3016 restart:
3017 	spin_lock(&log->l_icloglock);
3018 	if (XLOG_FORCED_SHUTDOWN(log)) {
3019 		spin_unlock(&log->l_icloglock);
3020 		return -EIO;
3021 	}
3022 
3023 	iclog = log->l_iclog;
3024 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3025 		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3026 
3027 		/* Wait for log writes to have flushed */
3028 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3029 		goto restart;
3030 	}
3031 
3032 	head = &iclog->ic_header;
3033 
3034 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
3035 	log_offset = iclog->ic_offset;
3036 
3037 	/* On the 1st write to an iclog, figure out lsn.  This works
3038 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3039 	 * committing to.  If the offset is set, that's how many blocks
3040 	 * must be written.
3041 	 */
3042 	if (log_offset == 0) {
3043 		ticket->t_curr_res -= log->l_iclog_hsize;
3044 		xlog_tic_add_region(ticket,
3045 				    log->l_iclog_hsize,
3046 				    XLOG_REG_TYPE_LRHEADER);
3047 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3048 		head->h_lsn = cpu_to_be64(
3049 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3050 		ASSERT(log->l_curr_block >= 0);
3051 	}
3052 
3053 	/* If there is enough room to write everything, then do it.  Otherwise,
3054 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3055 	 * bit is on, so this will get flushed out.  Don't update ic_offset
3056 	 * until you know exactly how many bytes get copied.  Therefore, wait
3057 	 * until later to update ic_offset.
3058 	 *
3059 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3060 	 * can fit into remaining data section.
3061 	 */
3062 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3063 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3064 
3065 		/*
3066 		 * If I'm the only one writing to this iclog, sync it to disk.
3067 		 * We need to do an atomic compare and decrement here to avoid
3068 		 * racing with concurrent atomic_dec_and_lock() calls in
3069 		 * xlog_state_release_iclog() when there is more than one
3070 		 * reference to the iclog.
3071 		 */
3072 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3073 			/* we are the only one */
3074 			spin_unlock(&log->l_icloglock);
3075 			error = xlog_state_release_iclog(log, iclog);
3076 			if (error)
3077 				return error;
3078 		} else {
3079 			spin_unlock(&log->l_icloglock);
3080 		}
3081 		goto restart;
3082 	}
3083 
3084 	/* Do we have enough room to write the full amount in the remainder
3085 	 * of this iclog?  Or must we continue a write on the next iclog and
3086 	 * mark this iclog as completely taken?  In the case where we switch
3087 	 * iclogs (to mark it taken), this particular iclog will release/sync
3088 	 * to disk in xlog_write().
3089 	 */
3090 	if (len <= iclog->ic_size - iclog->ic_offset) {
3091 		*continued_write = 0;
3092 		iclog->ic_offset += len;
3093 	} else {
3094 		*continued_write = 1;
3095 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3096 	}
3097 	*iclogp = iclog;
3098 
3099 	ASSERT(iclog->ic_offset <= iclog->ic_size);
3100 	spin_unlock(&log->l_icloglock);
3101 
3102 	*logoffsetp = log_offset;
3103 	return 0;
3104 }	/* xlog_state_get_iclog_space */
3105 
3106 /* The first cnt-1 times through here we don't need to
3107  * move the grant write head because the permanent
3108  * reservation has reserved cnt times the unit amount.
3109  * Release part of current permanent unit reservation and
3110  * reset current reservation to be one units worth.  Also
3111  * move grant reservation head forward.
3112  */
3113 STATIC void
3114 xlog_regrant_reserve_log_space(
3115 	struct xlog		*log,
3116 	struct xlog_ticket	*ticket)
3117 {
3118 	trace_xfs_log_regrant_reserve_enter(log, ticket);
3119 
3120 	if (ticket->t_cnt > 0)
3121 		ticket->t_cnt--;
3122 
3123 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3124 					ticket->t_curr_res);
3125 	xlog_grant_sub_space(log, &log->l_write_head.grant,
3126 					ticket->t_curr_res);
3127 	ticket->t_curr_res = ticket->t_unit_res;
3128 	xlog_tic_reset_res(ticket);
3129 
3130 	trace_xfs_log_regrant_reserve_sub(log, ticket);
3131 
3132 	/* just return if we still have some of the pre-reserved space */
3133 	if (ticket->t_cnt > 0)
3134 		return;
3135 
3136 	xlog_grant_add_space(log, &log->l_reserve_head.grant,
3137 					ticket->t_unit_res);
3138 
3139 	trace_xfs_log_regrant_reserve_exit(log, ticket);
3140 
3141 	ticket->t_curr_res = ticket->t_unit_res;
3142 	xlog_tic_reset_res(ticket);
3143 }	/* xlog_regrant_reserve_log_space */
3144 
3145 
3146 /*
3147  * Give back the space left from a reservation.
3148  *
3149  * All the information we need to make a correct determination of space left
3150  * is present.  For non-permanent reservations, things are quite easy.  The
3151  * count should have been decremented to zero.  We only need to deal with the
3152  * space remaining in the current reservation part of the ticket.  If the
3153  * ticket contains a permanent reservation, there may be left over space which
3154  * needs to be released.  A count of N means that N-1 refills of the current
3155  * reservation can be done before we need to ask for more space.  The first
3156  * one goes to fill up the first current reservation.  Once we run out of
3157  * space, the count will stay at zero and the only space remaining will be
3158  * in the current reservation field.
3159  */
3160 STATIC void
3161 xlog_ungrant_log_space(
3162 	struct xlog		*log,
3163 	struct xlog_ticket	*ticket)
3164 {
3165 	int	bytes;
3166 
3167 	if (ticket->t_cnt > 0)
3168 		ticket->t_cnt--;
3169 
3170 	trace_xfs_log_ungrant_enter(log, ticket);
3171 	trace_xfs_log_ungrant_sub(log, ticket);
3172 
3173 	/*
3174 	 * If this is a permanent reservation ticket, we may be able to free
3175 	 * up more space based on the remaining count.
3176 	 */
3177 	bytes = ticket->t_curr_res;
3178 	if (ticket->t_cnt > 0) {
3179 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3180 		bytes += ticket->t_unit_res*ticket->t_cnt;
3181 	}
3182 
3183 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3184 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3185 
3186 	trace_xfs_log_ungrant_exit(log, ticket);
3187 
3188 	xfs_log_space_wake(log->l_mp);
3189 }
3190 
3191 /*
3192  * Flush iclog to disk if this is the last reference to the given iclog and
3193  * the WANT_SYNC bit is set.
3194  *
3195  * When this function is entered, the iclog is not necessarily in the
3196  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3197  *
3198  *
3199  */
3200 STATIC int
3201 xlog_state_release_iclog(
3202 	struct xlog		*log,
3203 	struct xlog_in_core	*iclog)
3204 {
3205 	int		sync = 0;	/* do we sync? */
3206 
3207 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3208 		return -EIO;
3209 
3210 	ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3211 	if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3212 		return 0;
3213 
3214 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
3215 		spin_unlock(&log->l_icloglock);
3216 		return -EIO;
3217 	}
3218 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3219 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
3220 
3221 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3222 		/* update tail before writing to iclog */
3223 		xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3224 		sync++;
3225 		iclog->ic_state = XLOG_STATE_SYNCING;
3226 		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3227 		xlog_verify_tail_lsn(log, iclog, tail_lsn);
3228 		/* cycle incremented when incrementing curr_block */
3229 	}
3230 	spin_unlock(&log->l_icloglock);
3231 
3232 	/*
3233 	 * We let the log lock go, so it's possible that we hit a log I/O
3234 	 * error or some other SHUTDOWN condition that marks the iclog
3235 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3236 	 * this iclog has consistent data, so we ignore IOERROR
3237 	 * flags after this point.
3238 	 */
3239 	if (sync)
3240 		return xlog_sync(log, iclog);
3241 	return 0;
3242 }	/* xlog_state_release_iclog */
3243 
3244 
3245 /*
3246  * This routine will mark the current iclog in the ring as WANT_SYNC
3247  * and move the current iclog pointer to the next iclog in the ring.
3248  * When this routine is called from xlog_state_get_iclog_space(), the
3249  * exact size of the iclog has not yet been determined.  All we know is
3250  * that every data block.  We have run out of space in this log record.
3251  */
3252 STATIC void
3253 xlog_state_switch_iclogs(
3254 	struct xlog		*log,
3255 	struct xlog_in_core	*iclog,
3256 	int			eventual_size)
3257 {
3258 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3259 	if (!eventual_size)
3260 		eventual_size = iclog->ic_offset;
3261 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3262 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3263 	log->l_prev_block = log->l_curr_block;
3264 	log->l_prev_cycle = log->l_curr_cycle;
3265 
3266 	/* roll log?: ic_offset changed later */
3267 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3268 
3269 	/* Round up to next log-sunit */
3270 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3271 	    log->l_mp->m_sb.sb_logsunit > 1) {
3272 		uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3273 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3274 	}
3275 
3276 	if (log->l_curr_block >= log->l_logBBsize) {
3277 		/*
3278 		 * Rewind the current block before the cycle is bumped to make
3279 		 * sure that the combined LSN never transiently moves forward
3280 		 * when the log wraps to the next cycle. This is to support the
3281 		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3282 		 * other cases should acquire l_icloglock.
3283 		 */
3284 		log->l_curr_block -= log->l_logBBsize;
3285 		ASSERT(log->l_curr_block >= 0);
3286 		smp_wmb();
3287 		log->l_curr_cycle++;
3288 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3289 			log->l_curr_cycle++;
3290 	}
3291 	ASSERT(iclog == log->l_iclog);
3292 	log->l_iclog = iclog->ic_next;
3293 }	/* xlog_state_switch_iclogs */
3294 
3295 /*
3296  * Write out all data in the in-core log as of this exact moment in time.
3297  *
3298  * Data may be written to the in-core log during this call.  However,
3299  * we don't guarantee this data will be written out.  A change from past
3300  * implementation means this routine will *not* write out zero length LRs.
3301  *
3302  * Basically, we try and perform an intelligent scan of the in-core logs.
3303  * If we determine there is no flushable data, we just return.  There is no
3304  * flushable data if:
3305  *
3306  *	1. the current iclog is active and has no data; the previous iclog
3307  *		is in the active or dirty state.
3308  *	2. the current iclog is drity, and the previous iclog is in the
3309  *		active or dirty state.
3310  *
3311  * We may sleep if:
3312  *
3313  *	1. the current iclog is not in the active nor dirty state.
3314  *	2. the current iclog dirty, and the previous iclog is not in the
3315  *		active nor dirty state.
3316  *	3. the current iclog is active, and there is another thread writing
3317  *		to this particular iclog.
3318  *	4. a) the current iclog is active and has no other writers
3319  *	   b) when we return from flushing out this iclog, it is still
3320  *		not in the active nor dirty state.
3321  */
3322 int
3323 xfs_log_force(
3324 	struct xfs_mount	*mp,
3325 	uint			flags)
3326 {
3327 	struct xlog		*log = mp->m_log;
3328 	struct xlog_in_core	*iclog;
3329 	xfs_lsn_t		lsn;
3330 
3331 	XFS_STATS_INC(mp, xs_log_force);
3332 	trace_xfs_log_force(mp, 0, _RET_IP_);
3333 
3334 	xlog_cil_force(log);
3335 
3336 	spin_lock(&log->l_icloglock);
3337 	iclog = log->l_iclog;
3338 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3339 		goto out_error;
3340 
3341 	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3342 	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3343 	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3344 		/*
3345 		 * If the head is dirty or (active and empty), then we need to
3346 		 * look at the previous iclog.
3347 		 *
3348 		 * If the previous iclog is active or dirty we are done.  There
3349 		 * is nothing to sync out. Otherwise, we attach ourselves to the
3350 		 * previous iclog and go to sleep.
3351 		 */
3352 		iclog = iclog->ic_prev;
3353 		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3354 		    iclog->ic_state == XLOG_STATE_DIRTY)
3355 			goto out_unlock;
3356 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3357 		if (atomic_read(&iclog->ic_refcnt) == 0) {
3358 			/*
3359 			 * We are the only one with access to this iclog.
3360 			 *
3361 			 * Flush it out now.  There should be a roundoff of zero
3362 			 * to show that someone has already taken care of the
3363 			 * roundoff from the previous sync.
3364 			 */
3365 			atomic_inc(&iclog->ic_refcnt);
3366 			lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3367 			xlog_state_switch_iclogs(log, iclog, 0);
3368 			spin_unlock(&log->l_icloglock);
3369 
3370 			if (xlog_state_release_iclog(log, iclog))
3371 				return -EIO;
3372 
3373 			spin_lock(&log->l_icloglock);
3374 			if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3375 			    iclog->ic_state == XLOG_STATE_DIRTY)
3376 				goto out_unlock;
3377 		} else {
3378 			/*
3379 			 * Someone else is writing to this iclog.
3380 			 *
3381 			 * Use its call to flush out the data.  However, the
3382 			 * other thread may not force out this LR, so we mark
3383 			 * it WANT_SYNC.
3384 			 */
3385 			xlog_state_switch_iclogs(log, iclog, 0);
3386 		}
3387 	} else {
3388 		/*
3389 		 * If the head iclog is not active nor dirty, we just attach
3390 		 * ourselves to the head and go to sleep if necessary.
3391 		 */
3392 		;
3393 	}
3394 
3395 	if (!(flags & XFS_LOG_SYNC))
3396 		goto out_unlock;
3397 
3398 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3399 		goto out_error;
3400 	XFS_STATS_INC(mp, xs_log_force_sleep);
3401 	xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3402 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3403 		return -EIO;
3404 	return 0;
3405 
3406 out_unlock:
3407 	spin_unlock(&log->l_icloglock);
3408 	return 0;
3409 out_error:
3410 	spin_unlock(&log->l_icloglock);
3411 	return -EIO;
3412 }
3413 
3414 static int
3415 __xfs_log_force_lsn(
3416 	struct xfs_mount	*mp,
3417 	xfs_lsn_t		lsn,
3418 	uint			flags,
3419 	int			*log_flushed,
3420 	bool			already_slept)
3421 {
3422 	struct xlog		*log = mp->m_log;
3423 	struct xlog_in_core	*iclog;
3424 
3425 	spin_lock(&log->l_icloglock);
3426 	iclog = log->l_iclog;
3427 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3428 		goto out_error;
3429 
3430 	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3431 		iclog = iclog->ic_next;
3432 		if (iclog == log->l_iclog)
3433 			goto out_unlock;
3434 	}
3435 
3436 	if (iclog->ic_state == XLOG_STATE_DIRTY)
3437 		goto out_unlock;
3438 
3439 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3440 		/*
3441 		 * We sleep here if we haven't already slept (e.g. this is the
3442 		 * first time we've looked at the correct iclog buf) and the
3443 		 * buffer before us is going to be sync'ed.  The reason for this
3444 		 * is that if we are doing sync transactions here, by waiting
3445 		 * for the previous I/O to complete, we can allow a few more
3446 		 * transactions into this iclog before we close it down.
3447 		 *
3448 		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3449 		 * refcnt so we can release the log (which drops the ref count).
3450 		 * The state switch keeps new transaction commits from using
3451 		 * this buffer.  When the current commits finish writing into
3452 		 * the buffer, the refcount will drop to zero and the buffer
3453 		 * will go out then.
3454 		 */
3455 		if (!already_slept &&
3456 		    (iclog->ic_prev->ic_state &
3457 		     (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3458 			ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3459 
3460 			XFS_STATS_INC(mp, xs_log_force_sleep);
3461 
3462 			xlog_wait(&iclog->ic_prev->ic_write_wait,
3463 					&log->l_icloglock);
3464 			return -EAGAIN;
3465 		}
3466 		atomic_inc(&iclog->ic_refcnt);
3467 		xlog_state_switch_iclogs(log, iclog, 0);
3468 		spin_unlock(&log->l_icloglock);
3469 		if (xlog_state_release_iclog(log, iclog))
3470 			return -EIO;
3471 		if (log_flushed)
3472 			*log_flushed = 1;
3473 		spin_lock(&log->l_icloglock);
3474 	}
3475 
3476 	if (!(flags & XFS_LOG_SYNC) ||
3477 	    (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3478 		goto out_unlock;
3479 
3480 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3481 		goto out_error;
3482 
3483 	XFS_STATS_INC(mp, xs_log_force_sleep);
3484 	xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3485 	if (iclog->ic_state & XLOG_STATE_IOERROR)
3486 		return -EIO;
3487 	return 0;
3488 
3489 out_unlock:
3490 	spin_unlock(&log->l_icloglock);
3491 	return 0;
3492 out_error:
3493 	spin_unlock(&log->l_icloglock);
3494 	return -EIO;
3495 }
3496 
3497 /*
3498  * Force the in-core log to disk for a specific LSN.
3499  *
3500  * Find in-core log with lsn.
3501  *	If it is in the DIRTY state, just return.
3502  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3503  *		state and go to sleep or return.
3504  *	If it is in any other state, go to sleep or return.
3505  *
3506  * Synchronous forces are implemented with a wait queue.  All callers trying
3507  * to force a given lsn to disk must wait on the queue attached to the
3508  * specific in-core log.  When given in-core log finally completes its write
3509  * to disk, that thread will wake up all threads waiting on the queue.
3510  */
3511 int
3512 xfs_log_force_lsn(
3513 	struct xfs_mount	*mp,
3514 	xfs_lsn_t		lsn,
3515 	uint			flags,
3516 	int			*log_flushed)
3517 {
3518 	int			ret;
3519 	ASSERT(lsn != 0);
3520 
3521 	XFS_STATS_INC(mp, xs_log_force);
3522 	trace_xfs_log_force(mp, lsn, _RET_IP_);
3523 
3524 	lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3525 	if (lsn == NULLCOMMITLSN)
3526 		return 0;
3527 
3528 	ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3529 	if (ret == -EAGAIN)
3530 		ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3531 	return ret;
3532 }
3533 
3534 /*
3535  * Called when we want to mark the current iclog as being ready to sync to
3536  * disk.
3537  */
3538 STATIC void
3539 xlog_state_want_sync(
3540 	struct xlog		*log,
3541 	struct xlog_in_core	*iclog)
3542 {
3543 	assert_spin_locked(&log->l_icloglock);
3544 
3545 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3546 		xlog_state_switch_iclogs(log, iclog, 0);
3547 	} else {
3548 		ASSERT(iclog->ic_state &
3549 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3550 	}
3551 }
3552 
3553 
3554 /*****************************************************************************
3555  *
3556  *		TICKET functions
3557  *
3558  *****************************************************************************
3559  */
3560 
3561 /*
3562  * Free a used ticket when its refcount falls to zero.
3563  */
3564 void
3565 xfs_log_ticket_put(
3566 	xlog_ticket_t	*ticket)
3567 {
3568 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3569 	if (atomic_dec_and_test(&ticket->t_ref))
3570 		kmem_zone_free(xfs_log_ticket_zone, ticket);
3571 }
3572 
3573 xlog_ticket_t *
3574 xfs_log_ticket_get(
3575 	xlog_ticket_t	*ticket)
3576 {
3577 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3578 	atomic_inc(&ticket->t_ref);
3579 	return ticket;
3580 }
3581 
3582 /*
3583  * Figure out the total log space unit (in bytes) that would be
3584  * required for a log ticket.
3585  */
3586 int
3587 xfs_log_calc_unit_res(
3588 	struct xfs_mount	*mp,
3589 	int			unit_bytes)
3590 {
3591 	struct xlog		*log = mp->m_log;
3592 	int			iclog_space;
3593 	uint			num_headers;
3594 
3595 	/*
3596 	 * Permanent reservations have up to 'cnt'-1 active log operations
3597 	 * in the log.  A unit in this case is the amount of space for one
3598 	 * of these log operations.  Normal reservations have a cnt of 1
3599 	 * and their unit amount is the total amount of space required.
3600 	 *
3601 	 * The following lines of code account for non-transaction data
3602 	 * which occupy space in the on-disk log.
3603 	 *
3604 	 * Normal form of a transaction is:
3605 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3606 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3607 	 *
3608 	 * We need to account for all the leadup data and trailer data
3609 	 * around the transaction data.
3610 	 * And then we need to account for the worst case in terms of using
3611 	 * more space.
3612 	 * The worst case will happen if:
3613 	 * - the placement of the transaction happens to be such that the
3614 	 *   roundoff is at its maximum
3615 	 * - the transaction data is synced before the commit record is synced
3616 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3617 	 *   Therefore the commit record is in its own Log Record.
3618 	 *   This can happen as the commit record is called with its
3619 	 *   own region to xlog_write().
3620 	 *   This then means that in the worst case, roundoff can happen for
3621 	 *   the commit-rec as well.
3622 	 *   The commit-rec is smaller than padding in this scenario and so it is
3623 	 *   not added separately.
3624 	 */
3625 
3626 	/* for trans header */
3627 	unit_bytes += sizeof(xlog_op_header_t);
3628 	unit_bytes += sizeof(xfs_trans_header_t);
3629 
3630 	/* for start-rec */
3631 	unit_bytes += sizeof(xlog_op_header_t);
3632 
3633 	/*
3634 	 * for LR headers - the space for data in an iclog is the size minus
3635 	 * the space used for the headers. If we use the iclog size, then we
3636 	 * undercalculate the number of headers required.
3637 	 *
3638 	 * Furthermore - the addition of op headers for split-recs might
3639 	 * increase the space required enough to require more log and op
3640 	 * headers, so take that into account too.
3641 	 *
3642 	 * IMPORTANT: This reservation makes the assumption that if this
3643 	 * transaction is the first in an iclog and hence has the LR headers
3644 	 * accounted to it, then the remaining space in the iclog is
3645 	 * exclusively for this transaction.  i.e. if the transaction is larger
3646 	 * than the iclog, it will be the only thing in that iclog.
3647 	 * Fundamentally, this means we must pass the entire log vector to
3648 	 * xlog_write to guarantee this.
3649 	 */
3650 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3651 	num_headers = howmany(unit_bytes, iclog_space);
3652 
3653 	/* for split-recs - ophdrs added when data split over LRs */
3654 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3655 
3656 	/* add extra header reservations if we overrun */
3657 	while (!num_headers ||
3658 	       howmany(unit_bytes, iclog_space) > num_headers) {
3659 		unit_bytes += sizeof(xlog_op_header_t);
3660 		num_headers++;
3661 	}
3662 	unit_bytes += log->l_iclog_hsize * num_headers;
3663 
3664 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3665 	unit_bytes += log->l_iclog_hsize;
3666 
3667 	/* for roundoff padding for transaction data and one for commit record */
3668 	if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3669 		/* log su roundoff */
3670 		unit_bytes += 2 * mp->m_sb.sb_logsunit;
3671 	} else {
3672 		/* BB roundoff */
3673 		unit_bytes += 2 * BBSIZE;
3674         }
3675 
3676 	return unit_bytes;
3677 }
3678 
3679 /*
3680  * Allocate and initialise a new log ticket.
3681  */
3682 struct xlog_ticket *
3683 xlog_ticket_alloc(
3684 	struct xlog		*log,
3685 	int			unit_bytes,
3686 	int			cnt,
3687 	char			client,
3688 	bool			permanent,
3689 	xfs_km_flags_t		alloc_flags)
3690 {
3691 	struct xlog_ticket	*tic;
3692 	int			unit_res;
3693 
3694 	tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3695 	if (!tic)
3696 		return NULL;
3697 
3698 	unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3699 
3700 	atomic_set(&tic->t_ref, 1);
3701 	tic->t_task		= current;
3702 	INIT_LIST_HEAD(&tic->t_queue);
3703 	tic->t_unit_res		= unit_res;
3704 	tic->t_curr_res		= unit_res;
3705 	tic->t_cnt		= cnt;
3706 	tic->t_ocnt		= cnt;
3707 	tic->t_tid		= prandom_u32();
3708 	tic->t_clientid		= client;
3709 	tic->t_flags		= XLOG_TIC_INITED;
3710 	if (permanent)
3711 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3712 
3713 	xlog_tic_reset_res(tic);
3714 
3715 	return tic;
3716 }
3717 
3718 
3719 /******************************************************************************
3720  *
3721  *		Log debug routines
3722  *
3723  ******************************************************************************
3724  */
3725 #if defined(DEBUG)
3726 /*
3727  * Make sure that the destination ptr is within the valid data region of
3728  * one of the iclogs.  This uses backup pointers stored in a different
3729  * part of the log in case we trash the log structure.
3730  */
3731 STATIC void
3732 xlog_verify_dest_ptr(
3733 	struct xlog	*log,
3734 	void		*ptr)
3735 {
3736 	int i;
3737 	int good_ptr = 0;
3738 
3739 	for (i = 0; i < log->l_iclog_bufs; i++) {
3740 		if (ptr >= log->l_iclog_bak[i] &&
3741 		    ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3742 			good_ptr++;
3743 	}
3744 
3745 	if (!good_ptr)
3746 		xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3747 }
3748 
3749 /*
3750  * Check to make sure the grant write head didn't just over lap the tail.  If
3751  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3752  * the cycles differ by exactly one and check the byte count.
3753  *
3754  * This check is run unlocked, so can give false positives. Rather than assert
3755  * on failures, use a warn-once flag and a panic tag to allow the admin to
3756  * determine if they want to panic the machine when such an error occurs. For
3757  * debug kernels this will have the same effect as using an assert but, unlinke
3758  * an assert, it can be turned off at runtime.
3759  */
3760 STATIC void
3761 xlog_verify_grant_tail(
3762 	struct xlog	*log)
3763 {
3764 	int		tail_cycle, tail_blocks;
3765 	int		cycle, space;
3766 
3767 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3768 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3769 	if (tail_cycle != cycle) {
3770 		if (cycle - 1 != tail_cycle &&
3771 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3772 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3773 				"%s: cycle - 1 != tail_cycle", __func__);
3774 			log->l_flags |= XLOG_TAIL_WARN;
3775 		}
3776 
3777 		if (space > BBTOB(tail_blocks) &&
3778 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3779 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3780 				"%s: space > BBTOB(tail_blocks)", __func__);
3781 			log->l_flags |= XLOG_TAIL_WARN;
3782 		}
3783 	}
3784 }
3785 
3786 /* check if it will fit */
3787 STATIC void
3788 xlog_verify_tail_lsn(
3789 	struct xlog		*log,
3790 	struct xlog_in_core	*iclog,
3791 	xfs_lsn_t		tail_lsn)
3792 {
3793     int blocks;
3794 
3795     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3796 	blocks =
3797 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3798 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3799 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3800     } else {
3801 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3802 
3803 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3804 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3805 
3806 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3807 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3808 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3809     }
3810 }	/* xlog_verify_tail_lsn */
3811 
3812 /*
3813  * Perform a number of checks on the iclog before writing to disk.
3814  *
3815  * 1. Make sure the iclogs are still circular
3816  * 2. Make sure we have a good magic number
3817  * 3. Make sure we don't have magic numbers in the data
3818  * 4. Check fields of each log operation header for:
3819  *	A. Valid client identifier
3820  *	B. tid ptr value falls in valid ptr space (user space code)
3821  *	C. Length in log record header is correct according to the
3822  *		individual operation headers within record.
3823  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3824  *	log, check the preceding blocks of the physical log to make sure all
3825  *	the cycle numbers agree with the current cycle number.
3826  */
3827 STATIC void
3828 xlog_verify_iclog(
3829 	struct xlog		*log,
3830 	struct xlog_in_core	*iclog,
3831 	int			count,
3832 	bool                    syncing)
3833 {
3834 	xlog_op_header_t	*ophead;
3835 	xlog_in_core_t		*icptr;
3836 	xlog_in_core_2_t	*xhdr;
3837 	void			*base_ptr, *ptr, *p;
3838 	ptrdiff_t		field_offset;
3839 	uint8_t			clientid;
3840 	int			len, i, j, k, op_len;
3841 	int			idx;
3842 
3843 	/* check validity of iclog pointers */
3844 	spin_lock(&log->l_icloglock);
3845 	icptr = log->l_iclog;
3846 	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3847 		ASSERT(icptr);
3848 
3849 	if (icptr != log->l_iclog)
3850 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3851 	spin_unlock(&log->l_icloglock);
3852 
3853 	/* check log magic numbers */
3854 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3855 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3856 
3857 	base_ptr = ptr = &iclog->ic_header;
3858 	p = &iclog->ic_header;
3859 	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3860 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3861 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3862 				__func__);
3863 	}
3864 
3865 	/* check fields */
3866 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3867 	base_ptr = ptr = iclog->ic_datap;
3868 	ophead = ptr;
3869 	xhdr = iclog->ic_data;
3870 	for (i = 0; i < len; i++) {
3871 		ophead = ptr;
3872 
3873 		/* clientid is only 1 byte */
3874 		p = &ophead->oh_clientid;
3875 		field_offset = p - base_ptr;
3876 		if (!syncing || (field_offset & 0x1ff)) {
3877 			clientid = ophead->oh_clientid;
3878 		} else {
3879 			idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3880 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3881 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3882 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3883 				clientid = xlog_get_client_id(
3884 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3885 			} else {
3886 				clientid = xlog_get_client_id(
3887 					iclog->ic_header.h_cycle_data[idx]);
3888 			}
3889 		}
3890 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3891 			xfs_warn(log->l_mp,
3892 				"%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3893 				__func__, clientid, ophead,
3894 				(unsigned long)field_offset);
3895 
3896 		/* check length */
3897 		p = &ophead->oh_len;
3898 		field_offset = p - base_ptr;
3899 		if (!syncing || (field_offset & 0x1ff)) {
3900 			op_len = be32_to_cpu(ophead->oh_len);
3901 		} else {
3902 			idx = BTOBBT((uintptr_t)&ophead->oh_len -
3903 				    (uintptr_t)iclog->ic_datap);
3904 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3905 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3906 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3907 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3908 			} else {
3909 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3910 			}
3911 		}
3912 		ptr += sizeof(xlog_op_header_t) + op_len;
3913 	}
3914 }	/* xlog_verify_iclog */
3915 #endif
3916 
3917 /*
3918  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3919  */
3920 STATIC int
3921 xlog_state_ioerror(
3922 	struct xlog	*log)
3923 {
3924 	xlog_in_core_t	*iclog, *ic;
3925 
3926 	iclog = log->l_iclog;
3927 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3928 		/*
3929 		 * Mark all the incore logs IOERROR.
3930 		 * From now on, no log flushes will result.
3931 		 */
3932 		ic = iclog;
3933 		do {
3934 			ic->ic_state = XLOG_STATE_IOERROR;
3935 			ic = ic->ic_next;
3936 		} while (ic != iclog);
3937 		return 0;
3938 	}
3939 	/*
3940 	 * Return non-zero, if state transition has already happened.
3941 	 */
3942 	return 1;
3943 }
3944 
3945 /*
3946  * This is called from xfs_force_shutdown, when we're forcibly
3947  * shutting down the filesystem, typically because of an IO error.
3948  * Our main objectives here are to make sure that:
3949  *	a. if !logerror, flush the logs to disk. Anything modified
3950  *	   after this is ignored.
3951  *	b. the filesystem gets marked 'SHUTDOWN' for all interested
3952  *	   parties to find out, 'atomically'.
3953  *	c. those who're sleeping on log reservations, pinned objects and
3954  *	    other resources get woken up, and be told the bad news.
3955  *	d. nothing new gets queued up after (b) and (c) are done.
3956  *
3957  * Note: for the !logerror case we need to flush the regions held in memory out
3958  * to disk first. This needs to be done before the log is marked as shutdown,
3959  * otherwise the iclog writes will fail.
3960  */
3961 int
3962 xfs_log_force_umount(
3963 	struct xfs_mount	*mp,
3964 	int			logerror)
3965 {
3966 	struct xlog	*log;
3967 	int		retval;
3968 
3969 	log = mp->m_log;
3970 
3971 	/*
3972 	 * If this happens during log recovery, don't worry about
3973 	 * locking; the log isn't open for business yet.
3974 	 */
3975 	if (!log ||
3976 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3977 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3978 		if (mp->m_sb_bp)
3979 			mp->m_sb_bp->b_flags |= XBF_DONE;
3980 		return 0;
3981 	}
3982 
3983 	/*
3984 	 * Somebody could've already done the hard work for us.
3985 	 * No need to get locks for this.
3986 	 */
3987 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3988 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3989 		return 1;
3990 	}
3991 
3992 	/*
3993 	 * Flush all the completed transactions to disk before marking the log
3994 	 * being shut down. We need to do it in this order to ensure that
3995 	 * completed operations are safely on disk before we shut down, and that
3996 	 * we don't have to issue any buffer IO after the shutdown flags are set
3997 	 * to guarantee this.
3998 	 */
3999 	if (!logerror)
4000 		xfs_log_force(mp, XFS_LOG_SYNC);
4001 
4002 	/*
4003 	 * mark the filesystem and the as in a shutdown state and wake
4004 	 * everybody up to tell them the bad news.
4005 	 */
4006 	spin_lock(&log->l_icloglock);
4007 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4008 	if (mp->m_sb_bp)
4009 		mp->m_sb_bp->b_flags |= XBF_DONE;
4010 
4011 	/*
4012 	 * Mark the log and the iclogs with IO error flags to prevent any
4013 	 * further log IO from being issued or completed.
4014 	 */
4015 	log->l_flags |= XLOG_IO_ERROR;
4016 	retval = xlog_state_ioerror(log);
4017 	spin_unlock(&log->l_icloglock);
4018 
4019 	/*
4020 	 * We don't want anybody waiting for log reservations after this. That
4021 	 * means we have to wake up everybody queued up on reserveq as well as
4022 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
4023 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
4024 	 * action is protected by the grant locks.
4025 	 */
4026 	xlog_grant_head_wake_all(&log->l_reserve_head);
4027 	xlog_grant_head_wake_all(&log->l_write_head);
4028 
4029 	/*
4030 	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4031 	 * as if the log writes were completed. The abort handling in the log
4032 	 * item committed callback functions will do this again under lock to
4033 	 * avoid races.
4034 	 */
4035 	wake_up_all(&log->l_cilp->xc_commit_wait);
4036 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4037 
4038 #ifdef XFSERRORDEBUG
4039 	{
4040 		xlog_in_core_t	*iclog;
4041 
4042 		spin_lock(&log->l_icloglock);
4043 		iclog = log->l_iclog;
4044 		do {
4045 			ASSERT(iclog->ic_callback == 0);
4046 			iclog = iclog->ic_next;
4047 		} while (iclog != log->l_iclog);
4048 		spin_unlock(&log->l_icloglock);
4049 	}
4050 #endif
4051 	/* return non-zero if log IOERROR transition had already happened */
4052 	return retval;
4053 }
4054 
4055 STATIC int
4056 xlog_iclogs_empty(
4057 	struct xlog	*log)
4058 {
4059 	xlog_in_core_t	*iclog;
4060 
4061 	iclog = log->l_iclog;
4062 	do {
4063 		/* endianness does not matter here, zero is zero in
4064 		 * any language.
4065 		 */
4066 		if (iclog->ic_header.h_num_logops)
4067 			return 0;
4068 		iclog = iclog->ic_next;
4069 	} while (iclog != log->l_iclog);
4070 	return 1;
4071 }
4072 
4073 /*
4074  * Verify that an LSN stamped into a piece of metadata is valid. This is
4075  * intended for use in read verifiers on v5 superblocks.
4076  */
4077 bool
4078 xfs_log_check_lsn(
4079 	struct xfs_mount	*mp,
4080 	xfs_lsn_t		lsn)
4081 {
4082 	struct xlog		*log = mp->m_log;
4083 	bool			valid;
4084 
4085 	/*
4086 	 * norecovery mode skips mount-time log processing and unconditionally
4087 	 * resets the in-core LSN. We can't validate in this mode, but
4088 	 * modifications are not allowed anyways so just return true.
4089 	 */
4090 	if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4091 		return true;
4092 
4093 	/*
4094 	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4095 	 * handled by recovery and thus safe to ignore here.
4096 	 */
4097 	if (lsn == NULLCOMMITLSN)
4098 		return true;
4099 
4100 	valid = xlog_valid_lsn(mp->m_log, lsn);
4101 
4102 	/* warn the user about what's gone wrong before verifier failure */
4103 	if (!valid) {
4104 		spin_lock(&log->l_icloglock);
4105 		xfs_warn(mp,
4106 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4107 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4108 			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4109 			 log->l_curr_cycle, log->l_curr_block);
4110 		spin_unlock(&log->l_icloglock);
4111 	}
4112 
4113 	return valid;
4114 }
4115 
4116 bool
4117 xfs_log_in_recovery(
4118 	struct xfs_mount	*mp)
4119 {
4120 	struct xlog		*log = mp->m_log;
4121 
4122 	return log->l_flags & XLOG_ACTIVE_RECOVERY;
4123 }
4124