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