xref: /linux/fs/xfs/xfs_log.c (revision 26b0d14106954ae46d2f4f7eec3481828a210f7d)
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_log.h"
22 #include "xfs_trans.h"
23 #include "xfs_sb.h"
24 #include "xfs_ag.h"
25 #include "xfs_mount.h"
26 #include "xfs_error.h"
27 #include "xfs_log_priv.h"
28 #include "xfs_buf_item.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_log_recover.h"
33 #include "xfs_trans_priv.h"
34 #include "xfs_dinode.h"
35 #include "xfs_inode.h"
36 #include "xfs_trace.h"
37 
38 kmem_zone_t	*xfs_log_ticket_zone;
39 
40 /* Local miscellaneous function prototypes */
41 STATIC int
42 xlog_commit_record(
43 	struct xlog		*log,
44 	struct xlog_ticket	*ticket,
45 	struct xlog_in_core	**iclog,
46 	xfs_lsn_t		*commitlsnp);
47 
48 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t	*mp,
49 				xfs_buftarg_t	*log_target,
50 				xfs_daddr_t	blk_offset,
51 				int		num_bblks);
52 STATIC int
53 xlog_space_left(
54 	struct xlog		*log,
55 	atomic64_t		*head);
56 STATIC int	 xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
57 STATIC void	 xlog_dealloc_log(xlog_t *log);
58 
59 /* local state machine functions */
60 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
61 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
62 STATIC int  xlog_state_get_iclog_space(xlog_t		*log,
63 				       int		len,
64 				       xlog_in_core_t	**iclog,
65 				       xlog_ticket_t	*ticket,
66 				       int		*continued_write,
67 				       int		*logoffsetp);
68 STATIC int  xlog_state_release_iclog(xlog_t		*log,
69 				     xlog_in_core_t	*iclog);
70 STATIC void xlog_state_switch_iclogs(xlog_t		*log,
71 				     xlog_in_core_t *iclog,
72 				     int		eventual_size);
73 STATIC void xlog_state_want_sync(xlog_t	*log, xlog_in_core_t *iclog);
74 
75 STATIC void
76 xlog_grant_push_ail(
77 	struct xlog	*log,
78 	int		need_bytes);
79 STATIC void xlog_regrant_reserve_log_space(xlog_t	 *log,
80 					   xlog_ticket_t *ticket);
81 STATIC void xlog_ungrant_log_space(xlog_t	 *log,
82 				   xlog_ticket_t *ticket);
83 
84 #if defined(DEBUG)
85 STATIC void	xlog_verify_dest_ptr(xlog_t *log, char *ptr);
86 STATIC void
87 xlog_verify_grant_tail(
88 	struct xlog	*log);
89 STATIC void	xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
90 				  int count, boolean_t syncing);
91 STATIC void	xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
92 				     xfs_lsn_t tail_lsn);
93 #else
94 #define xlog_verify_dest_ptr(a,b)
95 #define xlog_verify_grant_tail(a)
96 #define xlog_verify_iclog(a,b,c,d)
97 #define xlog_verify_tail_lsn(a,b,c)
98 #endif
99 
100 STATIC int	xlog_iclogs_empty(xlog_t *log);
101 
102 static void
103 xlog_grant_sub_space(
104 	struct xlog		*log,
105 	atomic64_t		*head,
106 	int			bytes)
107 {
108 	int64_t	head_val = atomic64_read(head);
109 	int64_t new, old;
110 
111 	do {
112 		int	cycle, space;
113 
114 		xlog_crack_grant_head_val(head_val, &cycle, &space);
115 
116 		space -= bytes;
117 		if (space < 0) {
118 			space += log->l_logsize;
119 			cycle--;
120 		}
121 
122 		old = head_val;
123 		new = xlog_assign_grant_head_val(cycle, space);
124 		head_val = atomic64_cmpxchg(head, old, new);
125 	} while (head_val != old);
126 }
127 
128 static void
129 xlog_grant_add_space(
130 	struct xlog		*log,
131 	atomic64_t		*head,
132 	int			bytes)
133 {
134 	int64_t	head_val = atomic64_read(head);
135 	int64_t new, old;
136 
137 	do {
138 		int		tmp;
139 		int		cycle, space;
140 
141 		xlog_crack_grant_head_val(head_val, &cycle, &space);
142 
143 		tmp = log->l_logsize - space;
144 		if (tmp > bytes)
145 			space += bytes;
146 		else {
147 			space = bytes - tmp;
148 			cycle++;
149 		}
150 
151 		old = head_val;
152 		new = xlog_assign_grant_head_val(cycle, space);
153 		head_val = atomic64_cmpxchg(head, old, new);
154 	} while (head_val != old);
155 }
156 
157 STATIC void
158 xlog_grant_head_init(
159 	struct xlog_grant_head	*head)
160 {
161 	xlog_assign_grant_head(&head->grant, 1, 0);
162 	INIT_LIST_HEAD(&head->waiters);
163 	spin_lock_init(&head->lock);
164 }
165 
166 STATIC void
167 xlog_grant_head_wake_all(
168 	struct xlog_grant_head	*head)
169 {
170 	struct xlog_ticket	*tic;
171 
172 	spin_lock(&head->lock);
173 	list_for_each_entry(tic, &head->waiters, t_queue)
174 		wake_up_process(tic->t_task);
175 	spin_unlock(&head->lock);
176 }
177 
178 static inline int
179 xlog_ticket_reservation(
180 	struct xlog		*log,
181 	struct xlog_grant_head	*head,
182 	struct xlog_ticket	*tic)
183 {
184 	if (head == &log->l_write_head) {
185 		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
186 		return tic->t_unit_res;
187 	} else {
188 		if (tic->t_flags & XLOG_TIC_PERM_RESERV)
189 			return tic->t_unit_res * tic->t_cnt;
190 		else
191 			return tic->t_unit_res;
192 	}
193 }
194 
195 STATIC bool
196 xlog_grant_head_wake(
197 	struct xlog		*log,
198 	struct xlog_grant_head	*head,
199 	int			*free_bytes)
200 {
201 	struct xlog_ticket	*tic;
202 	int			need_bytes;
203 
204 	list_for_each_entry(tic, &head->waiters, t_queue) {
205 		need_bytes = xlog_ticket_reservation(log, head, tic);
206 		if (*free_bytes < need_bytes)
207 			return false;
208 
209 		*free_bytes -= need_bytes;
210 		trace_xfs_log_grant_wake_up(log, tic);
211 		wake_up_process(tic->t_task);
212 	}
213 
214 	return true;
215 }
216 
217 STATIC int
218 xlog_grant_head_wait(
219 	struct xlog		*log,
220 	struct xlog_grant_head	*head,
221 	struct xlog_ticket	*tic,
222 	int			need_bytes)
223 {
224 	list_add_tail(&tic->t_queue, &head->waiters);
225 
226 	do {
227 		if (XLOG_FORCED_SHUTDOWN(log))
228 			goto shutdown;
229 		xlog_grant_push_ail(log, need_bytes);
230 
231 		__set_current_state(TASK_UNINTERRUPTIBLE);
232 		spin_unlock(&head->lock);
233 
234 		XFS_STATS_INC(xs_sleep_logspace);
235 
236 		trace_xfs_log_grant_sleep(log, tic);
237 		schedule();
238 		trace_xfs_log_grant_wake(log, tic);
239 
240 		spin_lock(&head->lock);
241 		if (XLOG_FORCED_SHUTDOWN(log))
242 			goto shutdown;
243 	} while (xlog_space_left(log, &head->grant) < need_bytes);
244 
245 	list_del_init(&tic->t_queue);
246 	return 0;
247 shutdown:
248 	list_del_init(&tic->t_queue);
249 	return XFS_ERROR(EIO);
250 }
251 
252 /*
253  * Atomically get the log space required for a log ticket.
254  *
255  * Once a ticket gets put onto head->waiters, it will only return after the
256  * needed reservation is satisfied.
257  *
258  * This function is structured so that it has a lock free fast path. This is
259  * necessary because every new transaction reservation will come through this
260  * path. Hence any lock will be globally hot if we take it unconditionally on
261  * every pass.
262  *
263  * As tickets are only ever moved on and off head->waiters under head->lock, we
264  * only need to take that lock if we are going to add the ticket to the queue
265  * and sleep. We can avoid taking the lock if the ticket was never added to
266  * head->waiters because the t_queue list head will be empty and we hold the
267  * only reference to it so it can safely be checked unlocked.
268  */
269 STATIC int
270 xlog_grant_head_check(
271 	struct xlog		*log,
272 	struct xlog_grant_head	*head,
273 	struct xlog_ticket	*tic,
274 	int			*need_bytes)
275 {
276 	int			free_bytes;
277 	int			error = 0;
278 
279 	ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
280 
281 	/*
282 	 * If there are other waiters on the queue then give them a chance at
283 	 * logspace before us.  Wake up the first waiters, if we do not wake
284 	 * up all the waiters then go to sleep waiting for more free space,
285 	 * otherwise try to get some space for this transaction.
286 	 */
287 	*need_bytes = xlog_ticket_reservation(log, head, tic);
288 	free_bytes = xlog_space_left(log, &head->grant);
289 	if (!list_empty_careful(&head->waiters)) {
290 		spin_lock(&head->lock);
291 		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
292 		    free_bytes < *need_bytes) {
293 			error = xlog_grant_head_wait(log, head, tic,
294 						     *need_bytes);
295 		}
296 		spin_unlock(&head->lock);
297 	} else if (free_bytes < *need_bytes) {
298 		spin_lock(&head->lock);
299 		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
300 		spin_unlock(&head->lock);
301 	}
302 
303 	return error;
304 }
305 
306 static void
307 xlog_tic_reset_res(xlog_ticket_t *tic)
308 {
309 	tic->t_res_num = 0;
310 	tic->t_res_arr_sum = 0;
311 	tic->t_res_num_ophdrs = 0;
312 }
313 
314 static void
315 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
316 {
317 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
318 		/* add to overflow and start again */
319 		tic->t_res_o_flow += tic->t_res_arr_sum;
320 		tic->t_res_num = 0;
321 		tic->t_res_arr_sum = 0;
322 	}
323 
324 	tic->t_res_arr[tic->t_res_num].r_len = len;
325 	tic->t_res_arr[tic->t_res_num].r_type = type;
326 	tic->t_res_arr_sum += len;
327 	tic->t_res_num++;
328 }
329 
330 /*
331  * Replenish the byte reservation required by moving the grant write head.
332  */
333 int
334 xfs_log_regrant(
335 	struct xfs_mount	*mp,
336 	struct xlog_ticket	*tic)
337 {
338 	struct xlog		*log = mp->m_log;
339 	int			need_bytes;
340 	int			error = 0;
341 
342 	if (XLOG_FORCED_SHUTDOWN(log))
343 		return XFS_ERROR(EIO);
344 
345 	XFS_STATS_INC(xs_try_logspace);
346 
347 	/*
348 	 * This is a new transaction on the ticket, so we need to change the
349 	 * transaction ID so that the next transaction has a different TID in
350 	 * the log. Just add one to the existing tid so that we can see chains
351 	 * of rolling transactions in the log easily.
352 	 */
353 	tic->t_tid++;
354 
355 	xlog_grant_push_ail(log, tic->t_unit_res);
356 
357 	tic->t_curr_res = tic->t_unit_res;
358 	xlog_tic_reset_res(tic);
359 
360 	if (tic->t_cnt > 0)
361 		return 0;
362 
363 	trace_xfs_log_regrant(log, tic);
364 
365 	error = xlog_grant_head_check(log, &log->l_write_head, tic,
366 				      &need_bytes);
367 	if (error)
368 		goto out_error;
369 
370 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
371 	trace_xfs_log_regrant_exit(log, tic);
372 	xlog_verify_grant_tail(log);
373 	return 0;
374 
375 out_error:
376 	/*
377 	 * If we are failing, make sure the ticket doesn't have any current
378 	 * reservations.  We don't want to add this back when the ticket/
379 	 * transaction gets cancelled.
380 	 */
381 	tic->t_curr_res = 0;
382 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
383 	return error;
384 }
385 
386 /*
387  * Reserve log space and return a ticket corresponding the reservation.
388  *
389  * Each reservation is going to reserve extra space for a log record header.
390  * When writes happen to the on-disk log, we don't subtract the length of the
391  * log record header from any reservation.  By wasting space in each
392  * reservation, we prevent over allocation problems.
393  */
394 int
395 xfs_log_reserve(
396 	struct xfs_mount	*mp,
397 	int		 	unit_bytes,
398 	int		 	cnt,
399 	struct xlog_ticket	**ticp,
400 	__uint8_t	 	client,
401 	bool			permanent,
402 	uint		 	t_type)
403 {
404 	struct xlog		*log = mp->m_log;
405 	struct xlog_ticket	*tic;
406 	int			need_bytes;
407 	int			error = 0;
408 
409 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
410 
411 	if (XLOG_FORCED_SHUTDOWN(log))
412 		return XFS_ERROR(EIO);
413 
414 	XFS_STATS_INC(xs_try_logspace);
415 
416 	ASSERT(*ticp == NULL);
417 	tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
418 				KM_SLEEP | KM_MAYFAIL);
419 	if (!tic)
420 		return XFS_ERROR(ENOMEM);
421 
422 	tic->t_trans_type = t_type;
423 	*ticp = tic;
424 
425 	xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
426 
427 	trace_xfs_log_reserve(log, tic);
428 
429 	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
430 				      &need_bytes);
431 	if (error)
432 		goto out_error;
433 
434 	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
435 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
436 	trace_xfs_log_reserve_exit(log, tic);
437 	xlog_verify_grant_tail(log);
438 	return 0;
439 
440 out_error:
441 	/*
442 	 * If we are failing, make sure the ticket doesn't have any current
443 	 * reservations.  We don't want to add this back when the ticket/
444 	 * transaction gets cancelled.
445 	 */
446 	tic->t_curr_res = 0;
447 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
448 	return error;
449 }
450 
451 
452 /*
453  * NOTES:
454  *
455  *	1. currblock field gets updated at startup and after in-core logs
456  *		marked as with WANT_SYNC.
457  */
458 
459 /*
460  * This routine is called when a user of a log manager ticket is done with
461  * the reservation.  If the ticket was ever used, then a commit record for
462  * the associated transaction is written out as a log operation header with
463  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
464  * a given ticket.  If the ticket was one with a permanent reservation, then
465  * a few operations are done differently.  Permanent reservation tickets by
466  * default don't release the reservation.  They just commit the current
467  * transaction with the belief that the reservation is still needed.  A flag
468  * must be passed in before permanent reservations are actually released.
469  * When these type of tickets are not released, they need to be set into
470  * the inited state again.  By doing this, a start record will be written
471  * out when the next write occurs.
472  */
473 xfs_lsn_t
474 xfs_log_done(
475 	struct xfs_mount	*mp,
476 	struct xlog_ticket	*ticket,
477 	struct xlog_in_core	**iclog,
478 	uint			flags)
479 {
480 	struct xlog		*log = mp->m_log;
481 	xfs_lsn_t		lsn = 0;
482 
483 	if (XLOG_FORCED_SHUTDOWN(log) ||
484 	    /*
485 	     * If nothing was ever written, don't write out commit record.
486 	     * If we get an error, just continue and give back the log ticket.
487 	     */
488 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
489 	     (xlog_commit_record(log, ticket, iclog, &lsn)))) {
490 		lsn = (xfs_lsn_t) -1;
491 		if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
492 			flags |= XFS_LOG_REL_PERM_RESERV;
493 		}
494 	}
495 
496 
497 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
498 	    (flags & XFS_LOG_REL_PERM_RESERV)) {
499 		trace_xfs_log_done_nonperm(log, ticket);
500 
501 		/*
502 		 * Release ticket if not permanent reservation or a specific
503 		 * request has been made to release a permanent reservation.
504 		 */
505 		xlog_ungrant_log_space(log, ticket);
506 		xfs_log_ticket_put(ticket);
507 	} else {
508 		trace_xfs_log_done_perm(log, ticket);
509 
510 		xlog_regrant_reserve_log_space(log, ticket);
511 		/* If this ticket was a permanent reservation and we aren't
512 		 * trying to release it, reset the inited flags; so next time
513 		 * we write, a start record will be written out.
514 		 */
515 		ticket->t_flags |= XLOG_TIC_INITED;
516 	}
517 
518 	return lsn;
519 }
520 
521 /*
522  * Attaches a new iclog I/O completion callback routine during
523  * transaction commit.  If the log is in error state, a non-zero
524  * return code is handed back and the caller is responsible for
525  * executing the callback at an appropriate time.
526  */
527 int
528 xfs_log_notify(
529 	struct xfs_mount	*mp,
530 	struct xlog_in_core	*iclog,
531 	xfs_log_callback_t	*cb)
532 {
533 	int	abortflg;
534 
535 	spin_lock(&iclog->ic_callback_lock);
536 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
537 	if (!abortflg) {
538 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
539 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
540 		cb->cb_next = NULL;
541 		*(iclog->ic_callback_tail) = cb;
542 		iclog->ic_callback_tail = &(cb->cb_next);
543 	}
544 	spin_unlock(&iclog->ic_callback_lock);
545 	return abortflg;
546 }
547 
548 int
549 xfs_log_release_iclog(
550 	struct xfs_mount	*mp,
551 	struct xlog_in_core	*iclog)
552 {
553 	if (xlog_state_release_iclog(mp->m_log, iclog)) {
554 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
555 		return EIO;
556 	}
557 
558 	return 0;
559 }
560 
561 /*
562  * Mount a log filesystem
563  *
564  * mp		- ubiquitous xfs mount point structure
565  * log_target	- buftarg of on-disk log device
566  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
567  * num_bblocks	- Number of BBSIZE blocks in on-disk log
568  *
569  * Return error or zero.
570  */
571 int
572 xfs_log_mount(
573 	xfs_mount_t	*mp,
574 	xfs_buftarg_t	*log_target,
575 	xfs_daddr_t	blk_offset,
576 	int		num_bblks)
577 {
578 	int		error;
579 
580 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
581 		xfs_notice(mp, "Mounting Filesystem");
582 	else {
583 		xfs_notice(mp,
584 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
585 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
586 	}
587 
588 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
589 	if (IS_ERR(mp->m_log)) {
590 		error = -PTR_ERR(mp->m_log);
591 		goto out;
592 	}
593 
594 	/*
595 	 * Initialize the AIL now we have a log.
596 	 */
597 	error = xfs_trans_ail_init(mp);
598 	if (error) {
599 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
600 		goto out_free_log;
601 	}
602 	mp->m_log->l_ailp = mp->m_ail;
603 
604 	/*
605 	 * skip log recovery on a norecovery mount.  pretend it all
606 	 * just worked.
607 	 */
608 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
609 		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
610 
611 		if (readonly)
612 			mp->m_flags &= ~XFS_MOUNT_RDONLY;
613 
614 		error = xlog_recover(mp->m_log);
615 
616 		if (readonly)
617 			mp->m_flags |= XFS_MOUNT_RDONLY;
618 		if (error) {
619 			xfs_warn(mp, "log mount/recovery failed: error %d",
620 				error);
621 			goto out_destroy_ail;
622 		}
623 	}
624 
625 	/* Normal transactions can now occur */
626 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
627 
628 	/*
629 	 * Now the log has been fully initialised and we know were our
630 	 * space grant counters are, we can initialise the permanent ticket
631 	 * needed for delayed logging to work.
632 	 */
633 	xlog_cil_init_post_recovery(mp->m_log);
634 
635 	return 0;
636 
637 out_destroy_ail:
638 	xfs_trans_ail_destroy(mp);
639 out_free_log:
640 	xlog_dealloc_log(mp->m_log);
641 out:
642 	return error;
643 }
644 
645 /*
646  * Finish the recovery of the file system.  This is separate from
647  * the xfs_log_mount() call, because it depends on the code in
648  * xfs_mountfs() to read in the root and real-time bitmap inodes
649  * between calling xfs_log_mount() and here.
650  *
651  * mp		- ubiquitous xfs mount point structure
652  */
653 int
654 xfs_log_mount_finish(xfs_mount_t *mp)
655 {
656 	int	error;
657 
658 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
659 		error = xlog_recover_finish(mp->m_log);
660 	else {
661 		error = 0;
662 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
663 	}
664 
665 	return error;
666 }
667 
668 /*
669  * Final log writes as part of unmount.
670  *
671  * Mark the filesystem clean as unmount happens.  Note that during relocation
672  * this routine needs to be executed as part of source-bag while the
673  * deallocation must not be done until source-end.
674  */
675 
676 /*
677  * Unmount record used to have a string "Unmount filesystem--" in the
678  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
679  * We just write the magic number now since that particular field isn't
680  * currently architecture converted and "nUmount" is a bit foo.
681  * As far as I know, there weren't any dependencies on the old behaviour.
682  */
683 
684 int
685 xfs_log_unmount_write(xfs_mount_t *mp)
686 {
687 	xlog_t		 *log = mp->m_log;
688 	xlog_in_core_t	 *iclog;
689 #ifdef DEBUG
690 	xlog_in_core_t	 *first_iclog;
691 #endif
692 	xlog_ticket_t	*tic = NULL;
693 	xfs_lsn_t	 lsn;
694 	int		 error;
695 
696 	/*
697 	 * Don't write out unmount record on read-only mounts.
698 	 * Or, if we are doing a forced umount (typically because of IO errors).
699 	 */
700 	if (mp->m_flags & XFS_MOUNT_RDONLY)
701 		return 0;
702 
703 	error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
704 	ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
705 
706 #ifdef DEBUG
707 	first_iclog = iclog = log->l_iclog;
708 	do {
709 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
710 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
711 			ASSERT(iclog->ic_offset == 0);
712 		}
713 		iclog = iclog->ic_next;
714 	} while (iclog != first_iclog);
715 #endif
716 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
717 		error = xfs_log_reserve(mp, 600, 1, &tic,
718 					XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
719 		if (!error) {
720 			/* the data section must be 32 bit size aligned */
721 			struct {
722 			    __uint16_t magic;
723 			    __uint16_t pad1;
724 			    __uint32_t pad2; /* may as well make it 64 bits */
725 			} magic = {
726 				.magic = XLOG_UNMOUNT_TYPE,
727 			};
728 			struct xfs_log_iovec reg = {
729 				.i_addr = &magic,
730 				.i_len = sizeof(magic),
731 				.i_type = XLOG_REG_TYPE_UNMOUNT,
732 			};
733 			struct xfs_log_vec vec = {
734 				.lv_niovecs = 1,
735 				.lv_iovecp = &reg,
736 			};
737 
738 			/* remove inited flag, and account for space used */
739 			tic->t_flags = 0;
740 			tic->t_curr_res -= sizeof(magic);
741 			error = xlog_write(log, &vec, tic, &lsn,
742 					   NULL, XLOG_UNMOUNT_TRANS);
743 			/*
744 			 * At this point, we're umounting anyway,
745 			 * so there's no point in transitioning log state
746 			 * to IOERROR. Just continue...
747 			 */
748 		}
749 
750 		if (error)
751 			xfs_alert(mp, "%s: unmount record failed", __func__);
752 
753 
754 		spin_lock(&log->l_icloglock);
755 		iclog = log->l_iclog;
756 		atomic_inc(&iclog->ic_refcnt);
757 		xlog_state_want_sync(log, iclog);
758 		spin_unlock(&log->l_icloglock);
759 		error = xlog_state_release_iclog(log, iclog);
760 
761 		spin_lock(&log->l_icloglock);
762 		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
763 		      iclog->ic_state == XLOG_STATE_DIRTY)) {
764 			if (!XLOG_FORCED_SHUTDOWN(log)) {
765 				xlog_wait(&iclog->ic_force_wait,
766 							&log->l_icloglock);
767 			} else {
768 				spin_unlock(&log->l_icloglock);
769 			}
770 		} else {
771 			spin_unlock(&log->l_icloglock);
772 		}
773 		if (tic) {
774 			trace_xfs_log_umount_write(log, tic);
775 			xlog_ungrant_log_space(log, tic);
776 			xfs_log_ticket_put(tic);
777 		}
778 	} else {
779 		/*
780 		 * We're already in forced_shutdown mode, couldn't
781 		 * even attempt to write out the unmount transaction.
782 		 *
783 		 * Go through the motions of sync'ing and releasing
784 		 * the iclog, even though no I/O will actually happen,
785 		 * we need to wait for other log I/Os that may already
786 		 * be in progress.  Do this as a separate section of
787 		 * code so we'll know if we ever get stuck here that
788 		 * we're in this odd situation of trying to unmount
789 		 * a file system that went into forced_shutdown as
790 		 * the result of an unmount..
791 		 */
792 		spin_lock(&log->l_icloglock);
793 		iclog = log->l_iclog;
794 		atomic_inc(&iclog->ic_refcnt);
795 
796 		xlog_state_want_sync(log, iclog);
797 		spin_unlock(&log->l_icloglock);
798 		error =  xlog_state_release_iclog(log, iclog);
799 
800 		spin_lock(&log->l_icloglock);
801 
802 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
803 			|| iclog->ic_state == XLOG_STATE_DIRTY
804 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
805 
806 				xlog_wait(&iclog->ic_force_wait,
807 							&log->l_icloglock);
808 		} else {
809 			spin_unlock(&log->l_icloglock);
810 		}
811 	}
812 
813 	return error;
814 }	/* xfs_log_unmount_write */
815 
816 /*
817  * Deallocate log structures for unmount/relocation.
818  *
819  * We need to stop the aild from running before we destroy
820  * and deallocate the log as the aild references the log.
821  */
822 void
823 xfs_log_unmount(xfs_mount_t *mp)
824 {
825 	cancel_delayed_work_sync(&mp->m_sync_work);
826 	xfs_trans_ail_destroy(mp);
827 	xlog_dealloc_log(mp->m_log);
828 }
829 
830 void
831 xfs_log_item_init(
832 	struct xfs_mount	*mp,
833 	struct xfs_log_item	*item,
834 	int			type,
835 	const struct xfs_item_ops *ops)
836 {
837 	item->li_mountp = mp;
838 	item->li_ailp = mp->m_ail;
839 	item->li_type = type;
840 	item->li_ops = ops;
841 	item->li_lv = NULL;
842 
843 	INIT_LIST_HEAD(&item->li_ail);
844 	INIT_LIST_HEAD(&item->li_cil);
845 }
846 
847 /*
848  * Wake up processes waiting for log space after we have moved the log tail.
849  */
850 void
851 xfs_log_space_wake(
852 	struct xfs_mount	*mp)
853 {
854 	struct xlog		*log = mp->m_log;
855 	int			free_bytes;
856 
857 	if (XLOG_FORCED_SHUTDOWN(log))
858 		return;
859 
860 	if (!list_empty_careful(&log->l_write_head.waiters)) {
861 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
862 
863 		spin_lock(&log->l_write_head.lock);
864 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
865 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
866 		spin_unlock(&log->l_write_head.lock);
867 	}
868 
869 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
870 		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
871 
872 		spin_lock(&log->l_reserve_head.lock);
873 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
874 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
875 		spin_unlock(&log->l_reserve_head.lock);
876 	}
877 }
878 
879 /*
880  * Determine if we have a transaction that has gone to disk
881  * that needs to be covered. To begin the transition to the idle state
882  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
883  * If we are then in a state where covering is needed, the caller is informed
884  * that dummy transactions are required to move the log into the idle state.
885  *
886  * Because this is called as part of the sync process, we should also indicate
887  * that dummy transactions should be issued in anything but the covered or
888  * idle states. This ensures that the log tail is accurately reflected in
889  * the log at the end of the sync, hence if a crash occurrs avoids replay
890  * of transactions where the metadata is already on disk.
891  */
892 int
893 xfs_log_need_covered(xfs_mount_t *mp)
894 {
895 	int		needed = 0;
896 	xlog_t		*log = mp->m_log;
897 
898 	if (!xfs_fs_writable(mp))
899 		return 0;
900 
901 	spin_lock(&log->l_icloglock);
902 	switch (log->l_covered_state) {
903 	case XLOG_STATE_COVER_DONE:
904 	case XLOG_STATE_COVER_DONE2:
905 	case XLOG_STATE_COVER_IDLE:
906 		break;
907 	case XLOG_STATE_COVER_NEED:
908 	case XLOG_STATE_COVER_NEED2:
909 		if (!xfs_ail_min_lsn(log->l_ailp) &&
910 		    xlog_iclogs_empty(log)) {
911 			if (log->l_covered_state == XLOG_STATE_COVER_NEED)
912 				log->l_covered_state = XLOG_STATE_COVER_DONE;
913 			else
914 				log->l_covered_state = XLOG_STATE_COVER_DONE2;
915 		}
916 		/* FALLTHRU */
917 	default:
918 		needed = 1;
919 		break;
920 	}
921 	spin_unlock(&log->l_icloglock);
922 	return needed;
923 }
924 
925 /*
926  * We may be holding the log iclog lock upon entering this routine.
927  */
928 xfs_lsn_t
929 xlog_assign_tail_lsn_locked(
930 	struct xfs_mount	*mp)
931 {
932 	struct xlog		*log = mp->m_log;
933 	struct xfs_log_item	*lip;
934 	xfs_lsn_t		tail_lsn;
935 
936 	assert_spin_locked(&mp->m_ail->xa_lock);
937 
938 	/*
939 	 * To make sure we always have a valid LSN for the log tail we keep
940 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
941 	 * and use that when the AIL was empty.
942 	 */
943 	lip = xfs_ail_min(mp->m_ail);
944 	if (lip)
945 		tail_lsn = lip->li_lsn;
946 	else
947 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
948 	atomic64_set(&log->l_tail_lsn, tail_lsn);
949 	return tail_lsn;
950 }
951 
952 xfs_lsn_t
953 xlog_assign_tail_lsn(
954 	struct xfs_mount	*mp)
955 {
956 	xfs_lsn_t		tail_lsn;
957 
958 	spin_lock(&mp->m_ail->xa_lock);
959 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
960 	spin_unlock(&mp->m_ail->xa_lock);
961 
962 	return tail_lsn;
963 }
964 
965 /*
966  * Return the space in the log between the tail and the head.  The head
967  * is passed in the cycle/bytes formal parms.  In the special case where
968  * the reserve head has wrapped passed the tail, this calculation is no
969  * longer valid.  In this case, just return 0 which means there is no space
970  * in the log.  This works for all places where this function is called
971  * with the reserve head.  Of course, if the write head were to ever
972  * wrap the tail, we should blow up.  Rather than catch this case here,
973  * we depend on other ASSERTions in other parts of the code.   XXXmiken
974  *
975  * This code also handles the case where the reservation head is behind
976  * the tail.  The details of this case are described below, but the end
977  * result is that we return the size of the log as the amount of space left.
978  */
979 STATIC int
980 xlog_space_left(
981 	struct xlog	*log,
982 	atomic64_t	*head)
983 {
984 	int		free_bytes;
985 	int		tail_bytes;
986 	int		tail_cycle;
987 	int		head_cycle;
988 	int		head_bytes;
989 
990 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
991 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
992 	tail_bytes = BBTOB(tail_bytes);
993 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
994 		free_bytes = log->l_logsize - (head_bytes - tail_bytes);
995 	else if (tail_cycle + 1 < head_cycle)
996 		return 0;
997 	else if (tail_cycle < head_cycle) {
998 		ASSERT(tail_cycle == (head_cycle - 1));
999 		free_bytes = tail_bytes - head_bytes;
1000 	} else {
1001 		/*
1002 		 * The reservation head is behind the tail.
1003 		 * In this case we just want to return the size of the
1004 		 * log as the amount of space left.
1005 		 */
1006 		xfs_alert(log->l_mp,
1007 			"xlog_space_left: head behind tail\n"
1008 			"  tail_cycle = %d, tail_bytes = %d\n"
1009 			"  GH   cycle = %d, GH   bytes = %d",
1010 			tail_cycle, tail_bytes, head_cycle, head_bytes);
1011 		ASSERT(0);
1012 		free_bytes = log->l_logsize;
1013 	}
1014 	return free_bytes;
1015 }
1016 
1017 
1018 /*
1019  * Log function which is called when an io completes.
1020  *
1021  * The log manager needs its own routine, in order to control what
1022  * happens with the buffer after the write completes.
1023  */
1024 void
1025 xlog_iodone(xfs_buf_t *bp)
1026 {
1027 	xlog_in_core_t	*iclog = bp->b_fspriv;
1028 	xlog_t		*l = iclog->ic_log;
1029 	int		aborted = 0;
1030 
1031 	/*
1032 	 * Race to shutdown the filesystem if we see an error.
1033 	 */
1034 	if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1035 			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1036 		xfs_buf_ioerror_alert(bp, __func__);
1037 		xfs_buf_stale(bp);
1038 		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1039 		/*
1040 		 * This flag will be propagated to the trans-committed
1041 		 * callback routines to let them know that the log-commit
1042 		 * didn't succeed.
1043 		 */
1044 		aborted = XFS_LI_ABORTED;
1045 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1046 		aborted = XFS_LI_ABORTED;
1047 	}
1048 
1049 	/* log I/O is always issued ASYNC */
1050 	ASSERT(XFS_BUF_ISASYNC(bp));
1051 	xlog_state_done_syncing(iclog, aborted);
1052 	/*
1053 	 * do not reference the buffer (bp) here as we could race
1054 	 * with it being freed after writing the unmount record to the
1055 	 * log.
1056 	 */
1057 
1058 }	/* xlog_iodone */
1059 
1060 /*
1061  * Return size of each in-core log record buffer.
1062  *
1063  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1064  *
1065  * If the filesystem blocksize is too large, we may need to choose a
1066  * larger size since the directory code currently logs entire blocks.
1067  */
1068 
1069 STATIC void
1070 xlog_get_iclog_buffer_size(xfs_mount_t	*mp,
1071 			   xlog_t	*log)
1072 {
1073 	int size;
1074 	int xhdrs;
1075 
1076 	if (mp->m_logbufs <= 0)
1077 		log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1078 	else
1079 		log->l_iclog_bufs = mp->m_logbufs;
1080 
1081 	/*
1082 	 * Buffer size passed in from mount system call.
1083 	 */
1084 	if (mp->m_logbsize > 0) {
1085 		size = log->l_iclog_size = mp->m_logbsize;
1086 		log->l_iclog_size_log = 0;
1087 		while (size != 1) {
1088 			log->l_iclog_size_log++;
1089 			size >>= 1;
1090 		}
1091 
1092 		if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1093 			/* # headers = size / 32k
1094 			 * one header holds cycles from 32k of data
1095 			 */
1096 
1097 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1098 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1099 				xhdrs++;
1100 			log->l_iclog_hsize = xhdrs << BBSHIFT;
1101 			log->l_iclog_heads = xhdrs;
1102 		} else {
1103 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1104 			log->l_iclog_hsize = BBSIZE;
1105 			log->l_iclog_heads = 1;
1106 		}
1107 		goto done;
1108 	}
1109 
1110 	/* All machines use 32kB buffers by default. */
1111 	log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1112 	log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1113 
1114 	/* the default log size is 16k or 32k which is one header sector */
1115 	log->l_iclog_hsize = BBSIZE;
1116 	log->l_iclog_heads = 1;
1117 
1118 done:
1119 	/* are we being asked to make the sizes selected above visible? */
1120 	if (mp->m_logbufs == 0)
1121 		mp->m_logbufs = log->l_iclog_bufs;
1122 	if (mp->m_logbsize == 0)
1123 		mp->m_logbsize = log->l_iclog_size;
1124 }	/* xlog_get_iclog_buffer_size */
1125 
1126 
1127 /*
1128  * This routine initializes some of the log structure for a given mount point.
1129  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1130  * some other stuff may be filled in too.
1131  */
1132 STATIC xlog_t *
1133 xlog_alloc_log(xfs_mount_t	*mp,
1134 	       xfs_buftarg_t	*log_target,
1135 	       xfs_daddr_t	blk_offset,
1136 	       int		num_bblks)
1137 {
1138 	xlog_t			*log;
1139 	xlog_rec_header_t	*head;
1140 	xlog_in_core_t		**iclogp;
1141 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1142 	xfs_buf_t		*bp;
1143 	int			i;
1144 	int			error = ENOMEM;
1145 	uint			log2_size = 0;
1146 
1147 	log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1148 	if (!log) {
1149 		xfs_warn(mp, "Log allocation failed: No memory!");
1150 		goto out;
1151 	}
1152 
1153 	log->l_mp	   = mp;
1154 	log->l_targ	   = log_target;
1155 	log->l_logsize     = BBTOB(num_bblks);
1156 	log->l_logBBstart  = blk_offset;
1157 	log->l_logBBsize   = num_bblks;
1158 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1159 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1160 
1161 	log->l_prev_block  = -1;
1162 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1163 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1164 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1165 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1166 
1167 	xlog_grant_head_init(&log->l_reserve_head);
1168 	xlog_grant_head_init(&log->l_write_head);
1169 
1170 	error = EFSCORRUPTED;
1171 	if (xfs_sb_version_hassector(&mp->m_sb)) {
1172 	        log2_size = mp->m_sb.sb_logsectlog;
1173 		if (log2_size < BBSHIFT) {
1174 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1175 				log2_size, BBSHIFT);
1176 			goto out_free_log;
1177 		}
1178 
1179 	        log2_size -= BBSHIFT;
1180 		if (log2_size > mp->m_sectbb_log) {
1181 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1182 				log2_size, mp->m_sectbb_log);
1183 			goto out_free_log;
1184 		}
1185 
1186 		/* for larger sector sizes, must have v2 or external log */
1187 		if (log2_size && log->l_logBBstart > 0 &&
1188 			    !xfs_sb_version_haslogv2(&mp->m_sb)) {
1189 			xfs_warn(mp,
1190 		"log sector size (0x%x) invalid for configuration.",
1191 				log2_size);
1192 			goto out_free_log;
1193 		}
1194 	}
1195 	log->l_sectBBsize = 1 << log2_size;
1196 
1197 	xlog_get_iclog_buffer_size(mp, log);
1198 
1199 	error = ENOMEM;
1200 	bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1201 	if (!bp)
1202 		goto out_free_log;
1203 	bp->b_iodone = xlog_iodone;
1204 	ASSERT(xfs_buf_islocked(bp));
1205 	log->l_xbuf = bp;
1206 
1207 	spin_lock_init(&log->l_icloglock);
1208 	init_waitqueue_head(&log->l_flush_wait);
1209 
1210 	iclogp = &log->l_iclog;
1211 	/*
1212 	 * The amount of memory to allocate for the iclog structure is
1213 	 * rather funky due to the way the structure is defined.  It is
1214 	 * done this way so that we can use different sizes for machines
1215 	 * with different amounts of memory.  See the definition of
1216 	 * xlog_in_core_t in xfs_log_priv.h for details.
1217 	 */
1218 	ASSERT(log->l_iclog_size >= 4096);
1219 	for (i=0; i < log->l_iclog_bufs; i++) {
1220 		*iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1221 		if (!*iclogp)
1222 			goto out_free_iclog;
1223 
1224 		iclog = *iclogp;
1225 		iclog->ic_prev = prev_iclog;
1226 		prev_iclog = iclog;
1227 
1228 		bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1229 						BTOBB(log->l_iclog_size), 0);
1230 		if (!bp)
1231 			goto out_free_iclog;
1232 
1233 		bp->b_iodone = xlog_iodone;
1234 		iclog->ic_bp = bp;
1235 		iclog->ic_data = bp->b_addr;
1236 #ifdef DEBUG
1237 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1238 #endif
1239 		head = &iclog->ic_header;
1240 		memset(head, 0, sizeof(xlog_rec_header_t));
1241 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1242 		head->h_version = cpu_to_be32(
1243 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1244 		head->h_size = cpu_to_be32(log->l_iclog_size);
1245 		/* new fields */
1246 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1247 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1248 
1249 		iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1250 		iclog->ic_state = XLOG_STATE_ACTIVE;
1251 		iclog->ic_log = log;
1252 		atomic_set(&iclog->ic_refcnt, 0);
1253 		spin_lock_init(&iclog->ic_callback_lock);
1254 		iclog->ic_callback_tail = &(iclog->ic_callback);
1255 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1256 
1257 		ASSERT(xfs_buf_islocked(iclog->ic_bp));
1258 		init_waitqueue_head(&iclog->ic_force_wait);
1259 		init_waitqueue_head(&iclog->ic_write_wait);
1260 
1261 		iclogp = &iclog->ic_next;
1262 	}
1263 	*iclogp = log->l_iclog;			/* complete ring */
1264 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1265 
1266 	error = xlog_cil_init(log);
1267 	if (error)
1268 		goto out_free_iclog;
1269 	return log;
1270 
1271 out_free_iclog:
1272 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1273 		prev_iclog = iclog->ic_next;
1274 		if (iclog->ic_bp)
1275 			xfs_buf_free(iclog->ic_bp);
1276 		kmem_free(iclog);
1277 	}
1278 	spinlock_destroy(&log->l_icloglock);
1279 	xfs_buf_free(log->l_xbuf);
1280 out_free_log:
1281 	kmem_free(log);
1282 out:
1283 	return ERR_PTR(-error);
1284 }	/* xlog_alloc_log */
1285 
1286 
1287 /*
1288  * Write out the commit record of a transaction associated with the given
1289  * ticket.  Return the lsn of the commit record.
1290  */
1291 STATIC int
1292 xlog_commit_record(
1293 	struct xlog		*log,
1294 	struct xlog_ticket	*ticket,
1295 	struct xlog_in_core	**iclog,
1296 	xfs_lsn_t		*commitlsnp)
1297 {
1298 	struct xfs_mount *mp = log->l_mp;
1299 	int	error;
1300 	struct xfs_log_iovec reg = {
1301 		.i_addr = NULL,
1302 		.i_len = 0,
1303 		.i_type = XLOG_REG_TYPE_COMMIT,
1304 	};
1305 	struct xfs_log_vec vec = {
1306 		.lv_niovecs = 1,
1307 		.lv_iovecp = &reg,
1308 	};
1309 
1310 	ASSERT_ALWAYS(iclog);
1311 	error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1312 					XLOG_COMMIT_TRANS);
1313 	if (error)
1314 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1315 	return error;
1316 }
1317 
1318 /*
1319  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1320  * log space.  This code pushes on the lsn which would supposedly free up
1321  * the 25% which we want to leave free.  We may need to adopt a policy which
1322  * pushes on an lsn which is further along in the log once we reach the high
1323  * water mark.  In this manner, we would be creating a low water mark.
1324  */
1325 STATIC void
1326 xlog_grant_push_ail(
1327 	struct xlog	*log,
1328 	int		need_bytes)
1329 {
1330 	xfs_lsn_t	threshold_lsn = 0;
1331 	xfs_lsn_t	last_sync_lsn;
1332 	int		free_blocks;
1333 	int		free_bytes;
1334 	int		threshold_block;
1335 	int		threshold_cycle;
1336 	int		free_threshold;
1337 
1338 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1339 
1340 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1341 	free_blocks = BTOBBT(free_bytes);
1342 
1343 	/*
1344 	 * Set the threshold for the minimum number of free blocks in the
1345 	 * log to the maximum of what the caller needs, one quarter of the
1346 	 * log, and 256 blocks.
1347 	 */
1348 	free_threshold = BTOBB(need_bytes);
1349 	free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1350 	free_threshold = MAX(free_threshold, 256);
1351 	if (free_blocks >= free_threshold)
1352 		return;
1353 
1354 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1355 						&threshold_block);
1356 	threshold_block += free_threshold;
1357 	if (threshold_block >= log->l_logBBsize) {
1358 		threshold_block -= log->l_logBBsize;
1359 		threshold_cycle += 1;
1360 	}
1361 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1362 					threshold_block);
1363 	/*
1364 	 * Don't pass in an lsn greater than the lsn of the last
1365 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1366 	 * so that it doesn't change between the compare and the set.
1367 	 */
1368 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1369 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1370 		threshold_lsn = last_sync_lsn;
1371 
1372 	/*
1373 	 * Get the transaction layer to kick the dirty buffers out to
1374 	 * disk asynchronously. No point in trying to do this if
1375 	 * the filesystem is shutting down.
1376 	 */
1377 	if (!XLOG_FORCED_SHUTDOWN(log))
1378 		xfs_ail_push(log->l_ailp, threshold_lsn);
1379 }
1380 
1381 /*
1382  * The bdstrat callback function for log bufs. This gives us a central
1383  * place to trap bufs in case we get hit by a log I/O error and need to
1384  * shutdown. Actually, in practice, even when we didn't get a log error,
1385  * we transition the iclogs to IOERROR state *after* flushing all existing
1386  * iclogs to disk. This is because we don't want anymore new transactions to be
1387  * started or completed afterwards.
1388  */
1389 STATIC int
1390 xlog_bdstrat(
1391 	struct xfs_buf		*bp)
1392 {
1393 	struct xlog_in_core	*iclog = bp->b_fspriv;
1394 
1395 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
1396 		xfs_buf_ioerror(bp, EIO);
1397 		xfs_buf_stale(bp);
1398 		xfs_buf_ioend(bp, 0);
1399 		/*
1400 		 * It would seem logical to return EIO here, but we rely on
1401 		 * the log state machine to propagate I/O errors instead of
1402 		 * doing it here.
1403 		 */
1404 		return 0;
1405 	}
1406 
1407 	xfs_buf_iorequest(bp);
1408 	return 0;
1409 }
1410 
1411 /*
1412  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1413  * fashion.  Previously, we should have moved the current iclog
1414  * ptr in the log to point to the next available iclog.  This allows further
1415  * write to continue while this code syncs out an iclog ready to go.
1416  * Before an in-core log can be written out, the data section must be scanned
1417  * to save away the 1st word of each BBSIZE block into the header.  We replace
1418  * it with the current cycle count.  Each BBSIZE block is tagged with the
1419  * cycle count because there in an implicit assumption that drives will
1420  * guarantee that entire 512 byte blocks get written at once.  In other words,
1421  * we can't have part of a 512 byte block written and part not written.  By
1422  * tagging each block, we will know which blocks are valid when recovering
1423  * after an unclean shutdown.
1424  *
1425  * This routine is single threaded on the iclog.  No other thread can be in
1426  * this routine with the same iclog.  Changing contents of iclog can there-
1427  * fore be done without grabbing the state machine lock.  Updating the global
1428  * log will require grabbing the lock though.
1429  *
1430  * The entire log manager uses a logical block numbering scheme.  Only
1431  * log_sync (and then only bwrite()) know about the fact that the log may
1432  * not start with block zero on a given device.  The log block start offset
1433  * is added immediately before calling bwrite().
1434  */
1435 
1436 STATIC int
1437 xlog_sync(xlog_t		*log,
1438 	  xlog_in_core_t	*iclog)
1439 {
1440 	xfs_caddr_t	dptr;		/* pointer to byte sized element */
1441 	xfs_buf_t	*bp;
1442 	int		i;
1443 	uint		count;		/* byte count of bwrite */
1444 	uint		count_init;	/* initial count before roundup */
1445 	int		roundoff;       /* roundoff to BB or stripe */
1446 	int		split = 0;	/* split write into two regions */
1447 	int		error;
1448 	int		v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1449 
1450 	XFS_STATS_INC(xs_log_writes);
1451 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1452 
1453 	/* Add for LR header */
1454 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1455 
1456 	/* Round out the log write size */
1457 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1458 		/* we have a v2 stripe unit to use */
1459 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1460 	} else {
1461 		count = BBTOB(BTOBB(count_init));
1462 	}
1463 	roundoff = count - count_init;
1464 	ASSERT(roundoff >= 0);
1465 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1466                 roundoff < log->l_mp->m_sb.sb_logsunit)
1467 		||
1468 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1469 		 roundoff < BBTOB(1)));
1470 
1471 	/* move grant heads by roundoff in sync */
1472 	xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1473 	xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1474 
1475 	/* put cycle number in every block */
1476 	xlog_pack_data(log, iclog, roundoff);
1477 
1478 	/* real byte length */
1479 	if (v2) {
1480 		iclog->ic_header.h_len =
1481 			cpu_to_be32(iclog->ic_offset + roundoff);
1482 	} else {
1483 		iclog->ic_header.h_len =
1484 			cpu_to_be32(iclog->ic_offset);
1485 	}
1486 
1487 	bp = iclog->ic_bp;
1488 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1489 
1490 	XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1491 
1492 	/* Do we need to split this write into 2 parts? */
1493 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1494 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1495 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1496 		iclog->ic_bwritecnt = 2;	/* split into 2 writes */
1497 	} else {
1498 		iclog->ic_bwritecnt = 1;
1499 	}
1500 	bp->b_io_length = BTOBB(count);
1501 	bp->b_fspriv = iclog;
1502 	XFS_BUF_ZEROFLAGS(bp);
1503 	XFS_BUF_ASYNC(bp);
1504 	bp->b_flags |= XBF_SYNCIO;
1505 
1506 	if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1507 		bp->b_flags |= XBF_FUA;
1508 
1509 		/*
1510 		 * Flush the data device before flushing the log to make
1511 		 * sure all meta data written back from the AIL actually made
1512 		 * it to disk before stamping the new log tail LSN into the
1513 		 * log buffer.  For an external log we need to issue the
1514 		 * flush explicitly, and unfortunately synchronously here;
1515 		 * for an internal log we can simply use the block layer
1516 		 * state machine for preflushes.
1517 		 */
1518 		if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1519 			xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1520 		else
1521 			bp->b_flags |= XBF_FLUSH;
1522 	}
1523 
1524 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1525 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1526 
1527 	xlog_verify_iclog(log, iclog, count, B_TRUE);
1528 
1529 	/* account for log which doesn't start at block #0 */
1530 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1531 	/*
1532 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1533 	 * is shutting down.
1534 	 */
1535 	XFS_BUF_WRITE(bp);
1536 
1537 	error = xlog_bdstrat(bp);
1538 	if (error) {
1539 		xfs_buf_ioerror_alert(bp, "xlog_sync");
1540 		return error;
1541 	}
1542 	if (split) {
1543 		bp = iclog->ic_log->l_xbuf;
1544 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1545 		xfs_buf_associate_memory(bp,
1546 				(char *)&iclog->ic_header + count, split);
1547 		bp->b_fspriv = iclog;
1548 		XFS_BUF_ZEROFLAGS(bp);
1549 		XFS_BUF_ASYNC(bp);
1550 		bp->b_flags |= XBF_SYNCIO;
1551 		if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1552 			bp->b_flags |= XBF_FUA;
1553 		dptr = bp->b_addr;
1554 		/*
1555 		 * Bump the cycle numbers at the start of each block
1556 		 * since this part of the buffer is at the start of
1557 		 * a new cycle.  Watch out for the header magic number
1558 		 * case, though.
1559 		 */
1560 		for (i = 0; i < split; i += BBSIZE) {
1561 			be32_add_cpu((__be32 *)dptr, 1);
1562 			if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1563 				be32_add_cpu((__be32 *)dptr, 1);
1564 			dptr += BBSIZE;
1565 		}
1566 
1567 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1568 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1569 
1570 		/* account for internal log which doesn't start at block #0 */
1571 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1572 		XFS_BUF_WRITE(bp);
1573 		error = xlog_bdstrat(bp);
1574 		if (error) {
1575 			xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1576 			return error;
1577 		}
1578 	}
1579 	return 0;
1580 }	/* xlog_sync */
1581 
1582 
1583 /*
1584  * Deallocate a log structure
1585  */
1586 STATIC void
1587 xlog_dealloc_log(xlog_t *log)
1588 {
1589 	xlog_in_core_t	*iclog, *next_iclog;
1590 	int		i;
1591 
1592 	xlog_cil_destroy(log);
1593 
1594 	/*
1595 	 * always need to ensure that the extra buffer does not point to memory
1596 	 * owned by another log buffer before we free it.
1597 	 */
1598 	xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1599 	xfs_buf_free(log->l_xbuf);
1600 
1601 	iclog = log->l_iclog;
1602 	for (i=0; i<log->l_iclog_bufs; i++) {
1603 		xfs_buf_free(iclog->ic_bp);
1604 		next_iclog = iclog->ic_next;
1605 		kmem_free(iclog);
1606 		iclog = next_iclog;
1607 	}
1608 	spinlock_destroy(&log->l_icloglock);
1609 
1610 	log->l_mp->m_log = NULL;
1611 	kmem_free(log);
1612 }	/* xlog_dealloc_log */
1613 
1614 /*
1615  * Update counters atomically now that memcpy is done.
1616  */
1617 /* ARGSUSED */
1618 static inline void
1619 xlog_state_finish_copy(xlog_t		*log,
1620 		       xlog_in_core_t	*iclog,
1621 		       int		record_cnt,
1622 		       int		copy_bytes)
1623 {
1624 	spin_lock(&log->l_icloglock);
1625 
1626 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1627 	iclog->ic_offset += copy_bytes;
1628 
1629 	spin_unlock(&log->l_icloglock);
1630 }	/* xlog_state_finish_copy */
1631 
1632 
1633 
1634 
1635 /*
1636  * print out info relating to regions written which consume
1637  * the reservation
1638  */
1639 void
1640 xlog_print_tic_res(
1641 	struct xfs_mount	*mp,
1642 	struct xlog_ticket	*ticket)
1643 {
1644 	uint i;
1645 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1646 
1647 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1648 	static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1649 	    "bformat",
1650 	    "bchunk",
1651 	    "efi_format",
1652 	    "efd_format",
1653 	    "iformat",
1654 	    "icore",
1655 	    "iext",
1656 	    "ibroot",
1657 	    "ilocal",
1658 	    "iattr_ext",
1659 	    "iattr_broot",
1660 	    "iattr_local",
1661 	    "qformat",
1662 	    "dquot",
1663 	    "quotaoff",
1664 	    "LR header",
1665 	    "unmount",
1666 	    "commit",
1667 	    "trans header"
1668 	};
1669 	static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1670 	    "SETATTR_NOT_SIZE",
1671 	    "SETATTR_SIZE",
1672 	    "INACTIVE",
1673 	    "CREATE",
1674 	    "CREATE_TRUNC",
1675 	    "TRUNCATE_FILE",
1676 	    "REMOVE",
1677 	    "LINK",
1678 	    "RENAME",
1679 	    "MKDIR",
1680 	    "RMDIR",
1681 	    "SYMLINK",
1682 	    "SET_DMATTRS",
1683 	    "GROWFS",
1684 	    "STRAT_WRITE",
1685 	    "DIOSTRAT",
1686 	    "WRITE_SYNC",
1687 	    "WRITEID",
1688 	    "ADDAFORK",
1689 	    "ATTRINVAL",
1690 	    "ATRUNCATE",
1691 	    "ATTR_SET",
1692 	    "ATTR_RM",
1693 	    "ATTR_FLAG",
1694 	    "CLEAR_AGI_BUCKET",
1695 	    "QM_SBCHANGE",
1696 	    "DUMMY1",
1697 	    "DUMMY2",
1698 	    "QM_QUOTAOFF",
1699 	    "QM_DQALLOC",
1700 	    "QM_SETQLIM",
1701 	    "QM_DQCLUSTER",
1702 	    "QM_QINOCREATE",
1703 	    "QM_QUOTAOFF_END",
1704 	    "SB_UNIT",
1705 	    "FSYNC_TS",
1706 	    "GROWFSRT_ALLOC",
1707 	    "GROWFSRT_ZERO",
1708 	    "GROWFSRT_FREE",
1709 	    "SWAPEXT"
1710 	};
1711 
1712 	xfs_warn(mp,
1713 		"xlog_write: reservation summary:\n"
1714 		"  trans type  = %s (%u)\n"
1715 		"  unit res    = %d bytes\n"
1716 		"  current res = %d bytes\n"
1717 		"  total reg   = %u bytes (o/flow = %u bytes)\n"
1718 		"  ophdrs      = %u (ophdr space = %u bytes)\n"
1719 		"  ophdr + reg = %u bytes\n"
1720 		"  num regions = %u\n",
1721 		((ticket->t_trans_type <= 0 ||
1722 		  ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1723 		  "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1724 		ticket->t_trans_type,
1725 		ticket->t_unit_res,
1726 		ticket->t_curr_res,
1727 		ticket->t_res_arr_sum, ticket->t_res_o_flow,
1728 		ticket->t_res_num_ophdrs, ophdr_spc,
1729 		ticket->t_res_arr_sum +
1730 		ticket->t_res_o_flow + ophdr_spc,
1731 		ticket->t_res_num);
1732 
1733 	for (i = 0; i < ticket->t_res_num; i++) {
1734 		uint r_type = ticket->t_res_arr[i].r_type;
1735 		xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1736 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1737 			    "bad-rtype" : res_type_str[r_type-1]),
1738 			    ticket->t_res_arr[i].r_len);
1739 	}
1740 
1741 	xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1742 		"xlog_write: reservation ran out. Need to up reservation");
1743 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1744 }
1745 
1746 /*
1747  * Calculate the potential space needed by the log vector.  Each region gets
1748  * its own xlog_op_header_t and may need to be double word aligned.
1749  */
1750 static int
1751 xlog_write_calc_vec_length(
1752 	struct xlog_ticket	*ticket,
1753 	struct xfs_log_vec	*log_vector)
1754 {
1755 	struct xfs_log_vec	*lv;
1756 	int			headers = 0;
1757 	int			len = 0;
1758 	int			i;
1759 
1760 	/* acct for start rec of xact */
1761 	if (ticket->t_flags & XLOG_TIC_INITED)
1762 		headers++;
1763 
1764 	for (lv = log_vector; lv; lv = lv->lv_next) {
1765 		headers += lv->lv_niovecs;
1766 
1767 		for (i = 0; i < lv->lv_niovecs; i++) {
1768 			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
1769 
1770 			len += vecp->i_len;
1771 			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1772 		}
1773 	}
1774 
1775 	ticket->t_res_num_ophdrs += headers;
1776 	len += headers * sizeof(struct xlog_op_header);
1777 
1778 	return len;
1779 }
1780 
1781 /*
1782  * If first write for transaction, insert start record  We can't be trying to
1783  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1784  */
1785 static int
1786 xlog_write_start_rec(
1787 	struct xlog_op_header	*ophdr,
1788 	struct xlog_ticket	*ticket)
1789 {
1790 	if (!(ticket->t_flags & XLOG_TIC_INITED))
1791 		return 0;
1792 
1793 	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
1794 	ophdr->oh_clientid = ticket->t_clientid;
1795 	ophdr->oh_len = 0;
1796 	ophdr->oh_flags = XLOG_START_TRANS;
1797 	ophdr->oh_res2 = 0;
1798 
1799 	ticket->t_flags &= ~XLOG_TIC_INITED;
1800 
1801 	return sizeof(struct xlog_op_header);
1802 }
1803 
1804 static xlog_op_header_t *
1805 xlog_write_setup_ophdr(
1806 	struct xlog		*log,
1807 	struct xlog_op_header	*ophdr,
1808 	struct xlog_ticket	*ticket,
1809 	uint			flags)
1810 {
1811 	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1812 	ophdr->oh_clientid = ticket->t_clientid;
1813 	ophdr->oh_res2 = 0;
1814 
1815 	/* are we copying a commit or unmount record? */
1816 	ophdr->oh_flags = flags;
1817 
1818 	/*
1819 	 * We've seen logs corrupted with bad transaction client ids.  This
1820 	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1821 	 * and shut down the filesystem.
1822 	 */
1823 	switch (ophdr->oh_clientid)  {
1824 	case XFS_TRANSACTION:
1825 	case XFS_VOLUME:
1826 	case XFS_LOG:
1827 		break;
1828 	default:
1829 		xfs_warn(log->l_mp,
1830 			"Bad XFS transaction clientid 0x%x in ticket 0x%p",
1831 			ophdr->oh_clientid, ticket);
1832 		return NULL;
1833 	}
1834 
1835 	return ophdr;
1836 }
1837 
1838 /*
1839  * Set up the parameters of the region copy into the log. This has
1840  * to handle region write split across multiple log buffers - this
1841  * state is kept external to this function so that this code can
1842  * can be written in an obvious, self documenting manner.
1843  */
1844 static int
1845 xlog_write_setup_copy(
1846 	struct xlog_ticket	*ticket,
1847 	struct xlog_op_header	*ophdr,
1848 	int			space_available,
1849 	int			space_required,
1850 	int			*copy_off,
1851 	int			*copy_len,
1852 	int			*last_was_partial_copy,
1853 	int			*bytes_consumed)
1854 {
1855 	int			still_to_copy;
1856 
1857 	still_to_copy = space_required - *bytes_consumed;
1858 	*copy_off = *bytes_consumed;
1859 
1860 	if (still_to_copy <= space_available) {
1861 		/* write of region completes here */
1862 		*copy_len = still_to_copy;
1863 		ophdr->oh_len = cpu_to_be32(*copy_len);
1864 		if (*last_was_partial_copy)
1865 			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1866 		*last_was_partial_copy = 0;
1867 		*bytes_consumed = 0;
1868 		return 0;
1869 	}
1870 
1871 	/* partial write of region, needs extra log op header reservation */
1872 	*copy_len = space_available;
1873 	ophdr->oh_len = cpu_to_be32(*copy_len);
1874 	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1875 	if (*last_was_partial_copy)
1876 		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1877 	*bytes_consumed += *copy_len;
1878 	(*last_was_partial_copy)++;
1879 
1880 	/* account for new log op header */
1881 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
1882 	ticket->t_res_num_ophdrs++;
1883 
1884 	return sizeof(struct xlog_op_header);
1885 }
1886 
1887 static int
1888 xlog_write_copy_finish(
1889 	struct xlog		*log,
1890 	struct xlog_in_core	*iclog,
1891 	uint			flags,
1892 	int			*record_cnt,
1893 	int			*data_cnt,
1894 	int			*partial_copy,
1895 	int			*partial_copy_len,
1896 	int			log_offset,
1897 	struct xlog_in_core	**commit_iclog)
1898 {
1899 	if (*partial_copy) {
1900 		/*
1901 		 * This iclog has already been marked WANT_SYNC by
1902 		 * xlog_state_get_iclog_space.
1903 		 */
1904 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1905 		*record_cnt = 0;
1906 		*data_cnt = 0;
1907 		return xlog_state_release_iclog(log, iclog);
1908 	}
1909 
1910 	*partial_copy = 0;
1911 	*partial_copy_len = 0;
1912 
1913 	if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1914 		/* no more space in this iclog - push it. */
1915 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1916 		*record_cnt = 0;
1917 		*data_cnt = 0;
1918 
1919 		spin_lock(&log->l_icloglock);
1920 		xlog_state_want_sync(log, iclog);
1921 		spin_unlock(&log->l_icloglock);
1922 
1923 		if (!commit_iclog)
1924 			return xlog_state_release_iclog(log, iclog);
1925 		ASSERT(flags & XLOG_COMMIT_TRANS);
1926 		*commit_iclog = iclog;
1927 	}
1928 
1929 	return 0;
1930 }
1931 
1932 /*
1933  * Write some region out to in-core log
1934  *
1935  * This will be called when writing externally provided regions or when
1936  * writing out a commit record for a given transaction.
1937  *
1938  * General algorithm:
1939  *	1. Find total length of this write.  This may include adding to the
1940  *		lengths passed in.
1941  *	2. Check whether we violate the tickets reservation.
1942  *	3. While writing to this iclog
1943  *	    A. Reserve as much space in this iclog as can get
1944  *	    B. If this is first write, save away start lsn
1945  *	    C. While writing this region:
1946  *		1. If first write of transaction, write start record
1947  *		2. Write log operation header (header per region)
1948  *		3. Find out if we can fit entire region into this iclog
1949  *		4. Potentially, verify destination memcpy ptr
1950  *		5. Memcpy (partial) region
1951  *		6. If partial copy, release iclog; otherwise, continue
1952  *			copying more regions into current iclog
1953  *	4. Mark want sync bit (in simulation mode)
1954  *	5. Release iclog for potential flush to on-disk log.
1955  *
1956  * ERRORS:
1957  * 1.	Panic if reservation is overrun.  This should never happen since
1958  *	reservation amounts are generated internal to the filesystem.
1959  * NOTES:
1960  * 1. Tickets are single threaded data structures.
1961  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1962  *	syncing routine.  When a single log_write region needs to span
1963  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1964  *	on all log operation writes which don't contain the end of the
1965  *	region.  The XLOG_END_TRANS bit is used for the in-core log
1966  *	operation which contains the end of the continued log_write region.
1967  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1968  *	we don't really know exactly how much space will be used.  As a result,
1969  *	we don't update ic_offset until the end when we know exactly how many
1970  *	bytes have been written out.
1971  */
1972 int
1973 xlog_write(
1974 	struct xlog		*log,
1975 	struct xfs_log_vec	*log_vector,
1976 	struct xlog_ticket	*ticket,
1977 	xfs_lsn_t		*start_lsn,
1978 	struct xlog_in_core	**commit_iclog,
1979 	uint			flags)
1980 {
1981 	struct xlog_in_core	*iclog = NULL;
1982 	struct xfs_log_iovec	*vecp;
1983 	struct xfs_log_vec	*lv;
1984 	int			len;
1985 	int			index;
1986 	int			partial_copy = 0;
1987 	int			partial_copy_len = 0;
1988 	int			contwr = 0;
1989 	int			record_cnt = 0;
1990 	int			data_cnt = 0;
1991 	int			error;
1992 
1993 	*start_lsn = 0;
1994 
1995 	len = xlog_write_calc_vec_length(ticket, log_vector);
1996 
1997 	/*
1998 	 * Region headers and bytes are already accounted for.
1999 	 * We only need to take into account start records and
2000 	 * split regions in this function.
2001 	 */
2002 	if (ticket->t_flags & XLOG_TIC_INITED)
2003 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
2004 
2005 	/*
2006 	 * Commit record headers need to be accounted for. These
2007 	 * come in as separate writes so are easy to detect.
2008 	 */
2009 	if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2010 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
2011 
2012 	if (ticket->t_curr_res < 0)
2013 		xlog_print_tic_res(log->l_mp, ticket);
2014 
2015 	index = 0;
2016 	lv = log_vector;
2017 	vecp = lv->lv_iovecp;
2018 	while (lv && index < lv->lv_niovecs) {
2019 		void		*ptr;
2020 		int		log_offset;
2021 
2022 		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2023 						   &contwr, &log_offset);
2024 		if (error)
2025 			return error;
2026 
2027 		ASSERT(log_offset <= iclog->ic_size - 1);
2028 		ptr = iclog->ic_datap + log_offset;
2029 
2030 		/* start_lsn is the first lsn written to. That's all we need. */
2031 		if (!*start_lsn)
2032 			*start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2033 
2034 		/*
2035 		 * This loop writes out as many regions as can fit in the amount
2036 		 * of space which was allocated by xlog_state_get_iclog_space().
2037 		 */
2038 		while (lv && index < lv->lv_niovecs) {
2039 			struct xfs_log_iovec	*reg = &vecp[index];
2040 			struct xlog_op_header	*ophdr;
2041 			int			start_rec_copy;
2042 			int			copy_len;
2043 			int			copy_off;
2044 
2045 			ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2046 			ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2047 
2048 			start_rec_copy = xlog_write_start_rec(ptr, ticket);
2049 			if (start_rec_copy) {
2050 				record_cnt++;
2051 				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2052 						   start_rec_copy);
2053 			}
2054 
2055 			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2056 			if (!ophdr)
2057 				return XFS_ERROR(EIO);
2058 
2059 			xlog_write_adv_cnt(&ptr, &len, &log_offset,
2060 					   sizeof(struct xlog_op_header));
2061 
2062 			len += xlog_write_setup_copy(ticket, ophdr,
2063 						     iclog->ic_size-log_offset,
2064 						     reg->i_len,
2065 						     &copy_off, &copy_len,
2066 						     &partial_copy,
2067 						     &partial_copy_len);
2068 			xlog_verify_dest_ptr(log, ptr);
2069 
2070 			/* copy region */
2071 			ASSERT(copy_len >= 0);
2072 			memcpy(ptr, reg->i_addr + copy_off, copy_len);
2073 			xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2074 
2075 			copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2076 			record_cnt++;
2077 			data_cnt += contwr ? copy_len : 0;
2078 
2079 			error = xlog_write_copy_finish(log, iclog, flags,
2080 						       &record_cnt, &data_cnt,
2081 						       &partial_copy,
2082 						       &partial_copy_len,
2083 						       log_offset,
2084 						       commit_iclog);
2085 			if (error)
2086 				return error;
2087 
2088 			/*
2089 			 * if we had a partial copy, we need to get more iclog
2090 			 * space but we don't want to increment the region
2091 			 * index because there is still more is this region to
2092 			 * write.
2093 			 *
2094 			 * If we completed writing this region, and we flushed
2095 			 * the iclog (indicated by resetting of the record
2096 			 * count), then we also need to get more log space. If
2097 			 * this was the last record, though, we are done and
2098 			 * can just return.
2099 			 */
2100 			if (partial_copy)
2101 				break;
2102 
2103 			if (++index == lv->lv_niovecs) {
2104 				lv = lv->lv_next;
2105 				index = 0;
2106 				if (lv)
2107 					vecp = lv->lv_iovecp;
2108 			}
2109 			if (record_cnt == 0) {
2110 				if (!lv)
2111 					return 0;
2112 				break;
2113 			}
2114 		}
2115 	}
2116 
2117 	ASSERT(len == 0);
2118 
2119 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2120 	if (!commit_iclog)
2121 		return xlog_state_release_iclog(log, iclog);
2122 
2123 	ASSERT(flags & XLOG_COMMIT_TRANS);
2124 	*commit_iclog = iclog;
2125 	return 0;
2126 }
2127 
2128 
2129 /*****************************************************************************
2130  *
2131  *		State Machine functions
2132  *
2133  *****************************************************************************
2134  */
2135 
2136 /* Clean iclogs starting from the head.  This ordering must be
2137  * maintained, so an iclog doesn't become ACTIVE beyond one that
2138  * is SYNCING.  This is also required to maintain the notion that we use
2139  * a ordered wait queue to hold off would be writers to the log when every
2140  * iclog is trying to sync to disk.
2141  *
2142  * State Change: DIRTY -> ACTIVE
2143  */
2144 STATIC void
2145 xlog_state_clean_log(xlog_t *log)
2146 {
2147 	xlog_in_core_t	*iclog;
2148 	int changed = 0;
2149 
2150 	iclog = log->l_iclog;
2151 	do {
2152 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
2153 			iclog->ic_state	= XLOG_STATE_ACTIVE;
2154 			iclog->ic_offset       = 0;
2155 			ASSERT(iclog->ic_callback == NULL);
2156 			/*
2157 			 * If the number of ops in this iclog indicate it just
2158 			 * contains the dummy transaction, we can
2159 			 * change state into IDLE (the second time around).
2160 			 * Otherwise we should change the state into
2161 			 * NEED a dummy.
2162 			 * We don't need to cover the dummy.
2163 			 */
2164 			if (!changed &&
2165 			   (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2166 			   		XLOG_COVER_OPS)) {
2167 				changed = 1;
2168 			} else {
2169 				/*
2170 				 * We have two dirty iclogs so start over
2171 				 * This could also be num of ops indicates
2172 				 * this is not the dummy going out.
2173 				 */
2174 				changed = 2;
2175 			}
2176 			iclog->ic_header.h_num_logops = 0;
2177 			memset(iclog->ic_header.h_cycle_data, 0,
2178 			      sizeof(iclog->ic_header.h_cycle_data));
2179 			iclog->ic_header.h_lsn = 0;
2180 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2181 			/* do nothing */;
2182 		else
2183 			break;	/* stop cleaning */
2184 		iclog = iclog->ic_next;
2185 	} while (iclog != log->l_iclog);
2186 
2187 	/* log is locked when we are called */
2188 	/*
2189 	 * Change state for the dummy log recording.
2190 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
2191 	 * we are done writing the dummy record.
2192 	 * If we are done with the second dummy recored (DONE2), then
2193 	 * we go to IDLE.
2194 	 */
2195 	if (changed) {
2196 		switch (log->l_covered_state) {
2197 		case XLOG_STATE_COVER_IDLE:
2198 		case XLOG_STATE_COVER_NEED:
2199 		case XLOG_STATE_COVER_NEED2:
2200 			log->l_covered_state = XLOG_STATE_COVER_NEED;
2201 			break;
2202 
2203 		case XLOG_STATE_COVER_DONE:
2204 			if (changed == 1)
2205 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
2206 			else
2207 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2208 			break;
2209 
2210 		case XLOG_STATE_COVER_DONE2:
2211 			if (changed == 1)
2212 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2213 			else
2214 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2215 			break;
2216 
2217 		default:
2218 			ASSERT(0);
2219 		}
2220 	}
2221 }	/* xlog_state_clean_log */
2222 
2223 STATIC xfs_lsn_t
2224 xlog_get_lowest_lsn(
2225 	xlog_t		*log)
2226 {
2227 	xlog_in_core_t  *lsn_log;
2228 	xfs_lsn_t	lowest_lsn, lsn;
2229 
2230 	lsn_log = log->l_iclog;
2231 	lowest_lsn = 0;
2232 	do {
2233 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2234 		lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2235 		if ((lsn && !lowest_lsn) ||
2236 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2237 			lowest_lsn = lsn;
2238 		}
2239 	    }
2240 	    lsn_log = lsn_log->ic_next;
2241 	} while (lsn_log != log->l_iclog);
2242 	return lowest_lsn;
2243 }
2244 
2245 
2246 STATIC void
2247 xlog_state_do_callback(
2248 	xlog_t		*log,
2249 	int		aborted,
2250 	xlog_in_core_t	*ciclog)
2251 {
2252 	xlog_in_core_t	   *iclog;
2253 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2254 						 * processed all iclogs once */
2255 	xfs_log_callback_t *cb, *cb_next;
2256 	int		   flushcnt = 0;
2257 	xfs_lsn_t	   lowest_lsn;
2258 	int		   ioerrors;	/* counter: iclogs with errors */
2259 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2260 	int		   funcdidcallbacks; /* flag: function did callbacks */
2261 	int		   repeats;	/* for issuing console warnings if
2262 					 * looping too many times */
2263 	int		   wake = 0;
2264 
2265 	spin_lock(&log->l_icloglock);
2266 	first_iclog = iclog = log->l_iclog;
2267 	ioerrors = 0;
2268 	funcdidcallbacks = 0;
2269 	repeats = 0;
2270 
2271 	do {
2272 		/*
2273 		 * Scan all iclogs starting with the one pointed to by the
2274 		 * log.  Reset this starting point each time the log is
2275 		 * unlocked (during callbacks).
2276 		 *
2277 		 * Keep looping through iclogs until one full pass is made
2278 		 * without running any callbacks.
2279 		 */
2280 		first_iclog = log->l_iclog;
2281 		iclog = log->l_iclog;
2282 		loopdidcallbacks = 0;
2283 		repeats++;
2284 
2285 		do {
2286 
2287 			/* skip all iclogs in the ACTIVE & DIRTY states */
2288 			if (iclog->ic_state &
2289 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2290 				iclog = iclog->ic_next;
2291 				continue;
2292 			}
2293 
2294 			/*
2295 			 * Between marking a filesystem SHUTDOWN and stopping
2296 			 * the log, we do flush all iclogs to disk (if there
2297 			 * wasn't a log I/O error). So, we do want things to
2298 			 * go smoothly in case of just a SHUTDOWN  w/o a
2299 			 * LOG_IO_ERROR.
2300 			 */
2301 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2302 				/*
2303 				 * Can only perform callbacks in order.  Since
2304 				 * this iclog is not in the DONE_SYNC/
2305 				 * DO_CALLBACK state, we skip the rest and
2306 				 * just try to clean up.  If we set our iclog
2307 				 * to DO_CALLBACK, we will not process it when
2308 				 * we retry since a previous iclog is in the
2309 				 * CALLBACK and the state cannot change since
2310 				 * we are holding the l_icloglock.
2311 				 */
2312 				if (!(iclog->ic_state &
2313 					(XLOG_STATE_DONE_SYNC |
2314 						 XLOG_STATE_DO_CALLBACK))) {
2315 					if (ciclog && (ciclog->ic_state ==
2316 							XLOG_STATE_DONE_SYNC)) {
2317 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2318 					}
2319 					break;
2320 				}
2321 				/*
2322 				 * We now have an iclog that is in either the
2323 				 * DO_CALLBACK or DONE_SYNC states. The other
2324 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2325 				 * caught by the above if and are going to
2326 				 * clean (i.e. we aren't doing their callbacks)
2327 				 * see the above if.
2328 				 */
2329 
2330 				/*
2331 				 * We will do one more check here to see if we
2332 				 * have chased our tail around.
2333 				 */
2334 
2335 				lowest_lsn = xlog_get_lowest_lsn(log);
2336 				if (lowest_lsn &&
2337 				    XFS_LSN_CMP(lowest_lsn,
2338 						be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2339 					iclog = iclog->ic_next;
2340 					continue; /* Leave this iclog for
2341 						   * another thread */
2342 				}
2343 
2344 				iclog->ic_state = XLOG_STATE_CALLBACK;
2345 
2346 
2347 				/*
2348 				 * update the last_sync_lsn before we drop the
2349 				 * icloglock to ensure we are the only one that
2350 				 * can update it.
2351 				 */
2352 				ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2353 					be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2354 				atomic64_set(&log->l_last_sync_lsn,
2355 					be64_to_cpu(iclog->ic_header.h_lsn));
2356 
2357 			} else
2358 				ioerrors++;
2359 
2360 			spin_unlock(&log->l_icloglock);
2361 
2362 			/*
2363 			 * Keep processing entries in the callback list until
2364 			 * we come around and it is empty.  We need to
2365 			 * atomically see that the list is empty and change the
2366 			 * state to DIRTY so that we don't miss any more
2367 			 * callbacks being added.
2368 			 */
2369 			spin_lock(&iclog->ic_callback_lock);
2370 			cb = iclog->ic_callback;
2371 			while (cb) {
2372 				iclog->ic_callback_tail = &(iclog->ic_callback);
2373 				iclog->ic_callback = NULL;
2374 				spin_unlock(&iclog->ic_callback_lock);
2375 
2376 				/* perform callbacks in the order given */
2377 				for (; cb; cb = cb_next) {
2378 					cb_next = cb->cb_next;
2379 					cb->cb_func(cb->cb_arg, aborted);
2380 				}
2381 				spin_lock(&iclog->ic_callback_lock);
2382 				cb = iclog->ic_callback;
2383 			}
2384 
2385 			loopdidcallbacks++;
2386 			funcdidcallbacks++;
2387 
2388 			spin_lock(&log->l_icloglock);
2389 			ASSERT(iclog->ic_callback == NULL);
2390 			spin_unlock(&iclog->ic_callback_lock);
2391 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2392 				iclog->ic_state = XLOG_STATE_DIRTY;
2393 
2394 			/*
2395 			 * Transition from DIRTY to ACTIVE if applicable.
2396 			 * NOP if STATE_IOERROR.
2397 			 */
2398 			xlog_state_clean_log(log);
2399 
2400 			/* wake up threads waiting in xfs_log_force() */
2401 			wake_up_all(&iclog->ic_force_wait);
2402 
2403 			iclog = iclog->ic_next;
2404 		} while (first_iclog != iclog);
2405 
2406 		if (repeats > 5000) {
2407 			flushcnt += repeats;
2408 			repeats = 0;
2409 			xfs_warn(log->l_mp,
2410 				"%s: possible infinite loop (%d iterations)",
2411 				__func__, flushcnt);
2412 		}
2413 	} while (!ioerrors && loopdidcallbacks);
2414 
2415 	/*
2416 	 * make one last gasp attempt to see if iclogs are being left in
2417 	 * limbo..
2418 	 */
2419 #ifdef DEBUG
2420 	if (funcdidcallbacks) {
2421 		first_iclog = iclog = log->l_iclog;
2422 		do {
2423 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2424 			/*
2425 			 * Terminate the loop if iclogs are found in states
2426 			 * which will cause other threads to clean up iclogs.
2427 			 *
2428 			 * SYNCING - i/o completion will go through logs
2429 			 * DONE_SYNC - interrupt thread should be waiting for
2430 			 *              l_icloglock
2431 			 * IOERROR - give up hope all ye who enter here
2432 			 */
2433 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2434 			    iclog->ic_state == XLOG_STATE_SYNCING ||
2435 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2436 			    iclog->ic_state == XLOG_STATE_IOERROR )
2437 				break;
2438 			iclog = iclog->ic_next;
2439 		} while (first_iclog != iclog);
2440 	}
2441 #endif
2442 
2443 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2444 		wake = 1;
2445 	spin_unlock(&log->l_icloglock);
2446 
2447 	if (wake)
2448 		wake_up_all(&log->l_flush_wait);
2449 }
2450 
2451 
2452 /*
2453  * Finish transitioning this iclog to the dirty state.
2454  *
2455  * Make sure that we completely execute this routine only when this is
2456  * the last call to the iclog.  There is a good chance that iclog flushes,
2457  * when we reach the end of the physical log, get turned into 2 separate
2458  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2459  * routine.  By using the reference count bwritecnt, we guarantee that only
2460  * the second completion goes through.
2461  *
2462  * Callbacks could take time, so they are done outside the scope of the
2463  * global state machine log lock.
2464  */
2465 STATIC void
2466 xlog_state_done_syncing(
2467 	xlog_in_core_t	*iclog,
2468 	int		aborted)
2469 {
2470 	xlog_t		   *log = iclog->ic_log;
2471 
2472 	spin_lock(&log->l_icloglock);
2473 
2474 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2475 	       iclog->ic_state == XLOG_STATE_IOERROR);
2476 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2477 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2478 
2479 
2480 	/*
2481 	 * If we got an error, either on the first buffer, or in the case of
2482 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2483 	 * and none should ever be attempted to be written to disk
2484 	 * again.
2485 	 */
2486 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2487 		if (--iclog->ic_bwritecnt == 1) {
2488 			spin_unlock(&log->l_icloglock);
2489 			return;
2490 		}
2491 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2492 	}
2493 
2494 	/*
2495 	 * Someone could be sleeping prior to writing out the next
2496 	 * iclog buffer, we wake them all, one will get to do the
2497 	 * I/O, the others get to wait for the result.
2498 	 */
2499 	wake_up_all(&iclog->ic_write_wait);
2500 	spin_unlock(&log->l_icloglock);
2501 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2502 }	/* xlog_state_done_syncing */
2503 
2504 
2505 /*
2506  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2507  * sleep.  We wait on the flush queue on the head iclog as that should be
2508  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2509  * we will wait here and all new writes will sleep until a sync completes.
2510  *
2511  * The in-core logs are used in a circular fashion. They are not used
2512  * out-of-order even when an iclog past the head is free.
2513  *
2514  * return:
2515  *	* log_offset where xlog_write() can start writing into the in-core
2516  *		log's data space.
2517  *	* in-core log pointer to which xlog_write() should write.
2518  *	* boolean indicating this is a continued write to an in-core log.
2519  *		If this is the last write, then the in-core log's offset field
2520  *		needs to be incremented, depending on the amount of data which
2521  *		is copied.
2522  */
2523 STATIC int
2524 xlog_state_get_iclog_space(xlog_t	  *log,
2525 			   int		  len,
2526 			   xlog_in_core_t **iclogp,
2527 			   xlog_ticket_t  *ticket,
2528 			   int		  *continued_write,
2529 			   int		  *logoffsetp)
2530 {
2531 	int		  log_offset;
2532 	xlog_rec_header_t *head;
2533 	xlog_in_core_t	  *iclog;
2534 	int		  error;
2535 
2536 restart:
2537 	spin_lock(&log->l_icloglock);
2538 	if (XLOG_FORCED_SHUTDOWN(log)) {
2539 		spin_unlock(&log->l_icloglock);
2540 		return XFS_ERROR(EIO);
2541 	}
2542 
2543 	iclog = log->l_iclog;
2544 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2545 		XFS_STATS_INC(xs_log_noiclogs);
2546 
2547 		/* Wait for log writes to have flushed */
2548 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2549 		goto restart;
2550 	}
2551 
2552 	head = &iclog->ic_header;
2553 
2554 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2555 	log_offset = iclog->ic_offset;
2556 
2557 	/* On the 1st write to an iclog, figure out lsn.  This works
2558 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2559 	 * committing to.  If the offset is set, that's how many blocks
2560 	 * must be written.
2561 	 */
2562 	if (log_offset == 0) {
2563 		ticket->t_curr_res -= log->l_iclog_hsize;
2564 		xlog_tic_add_region(ticket,
2565 				    log->l_iclog_hsize,
2566 				    XLOG_REG_TYPE_LRHEADER);
2567 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2568 		head->h_lsn = cpu_to_be64(
2569 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2570 		ASSERT(log->l_curr_block >= 0);
2571 	}
2572 
2573 	/* If there is enough room to write everything, then do it.  Otherwise,
2574 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2575 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2576 	 * until you know exactly how many bytes get copied.  Therefore, wait
2577 	 * until later to update ic_offset.
2578 	 *
2579 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2580 	 * can fit into remaining data section.
2581 	 */
2582 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2583 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2584 
2585 		/*
2586 		 * If I'm the only one writing to this iclog, sync it to disk.
2587 		 * We need to do an atomic compare and decrement here to avoid
2588 		 * racing with concurrent atomic_dec_and_lock() calls in
2589 		 * xlog_state_release_iclog() when there is more than one
2590 		 * reference to the iclog.
2591 		 */
2592 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2593 			/* we are the only one */
2594 			spin_unlock(&log->l_icloglock);
2595 			error = xlog_state_release_iclog(log, iclog);
2596 			if (error)
2597 				return error;
2598 		} else {
2599 			spin_unlock(&log->l_icloglock);
2600 		}
2601 		goto restart;
2602 	}
2603 
2604 	/* Do we have enough room to write the full amount in the remainder
2605 	 * of this iclog?  Or must we continue a write on the next iclog and
2606 	 * mark this iclog as completely taken?  In the case where we switch
2607 	 * iclogs (to mark it taken), this particular iclog will release/sync
2608 	 * to disk in xlog_write().
2609 	 */
2610 	if (len <= iclog->ic_size - iclog->ic_offset) {
2611 		*continued_write = 0;
2612 		iclog->ic_offset += len;
2613 	} else {
2614 		*continued_write = 1;
2615 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2616 	}
2617 	*iclogp = iclog;
2618 
2619 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2620 	spin_unlock(&log->l_icloglock);
2621 
2622 	*logoffsetp = log_offset;
2623 	return 0;
2624 }	/* xlog_state_get_iclog_space */
2625 
2626 /* The first cnt-1 times through here we don't need to
2627  * move the grant write head because the permanent
2628  * reservation has reserved cnt times the unit amount.
2629  * Release part of current permanent unit reservation and
2630  * reset current reservation to be one units worth.  Also
2631  * move grant reservation head forward.
2632  */
2633 STATIC void
2634 xlog_regrant_reserve_log_space(xlog_t	     *log,
2635 			       xlog_ticket_t *ticket)
2636 {
2637 	trace_xfs_log_regrant_reserve_enter(log, ticket);
2638 
2639 	if (ticket->t_cnt > 0)
2640 		ticket->t_cnt--;
2641 
2642 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2643 					ticket->t_curr_res);
2644 	xlog_grant_sub_space(log, &log->l_write_head.grant,
2645 					ticket->t_curr_res);
2646 	ticket->t_curr_res = ticket->t_unit_res;
2647 	xlog_tic_reset_res(ticket);
2648 
2649 	trace_xfs_log_regrant_reserve_sub(log, ticket);
2650 
2651 	/* just return if we still have some of the pre-reserved space */
2652 	if (ticket->t_cnt > 0)
2653 		return;
2654 
2655 	xlog_grant_add_space(log, &log->l_reserve_head.grant,
2656 					ticket->t_unit_res);
2657 
2658 	trace_xfs_log_regrant_reserve_exit(log, ticket);
2659 
2660 	ticket->t_curr_res = ticket->t_unit_res;
2661 	xlog_tic_reset_res(ticket);
2662 }	/* xlog_regrant_reserve_log_space */
2663 
2664 
2665 /*
2666  * Give back the space left from a reservation.
2667  *
2668  * All the information we need to make a correct determination of space left
2669  * is present.  For non-permanent reservations, things are quite easy.  The
2670  * count should have been decremented to zero.  We only need to deal with the
2671  * space remaining in the current reservation part of the ticket.  If the
2672  * ticket contains a permanent reservation, there may be left over space which
2673  * needs to be released.  A count of N means that N-1 refills of the current
2674  * reservation can be done before we need to ask for more space.  The first
2675  * one goes to fill up the first current reservation.  Once we run out of
2676  * space, the count will stay at zero and the only space remaining will be
2677  * in the current reservation field.
2678  */
2679 STATIC void
2680 xlog_ungrant_log_space(xlog_t	     *log,
2681 		       xlog_ticket_t *ticket)
2682 {
2683 	int	bytes;
2684 
2685 	if (ticket->t_cnt > 0)
2686 		ticket->t_cnt--;
2687 
2688 	trace_xfs_log_ungrant_enter(log, ticket);
2689 	trace_xfs_log_ungrant_sub(log, ticket);
2690 
2691 	/*
2692 	 * If this is a permanent reservation ticket, we may be able to free
2693 	 * up more space based on the remaining count.
2694 	 */
2695 	bytes = ticket->t_curr_res;
2696 	if (ticket->t_cnt > 0) {
2697 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2698 		bytes += ticket->t_unit_res*ticket->t_cnt;
2699 	}
2700 
2701 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2702 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2703 
2704 	trace_xfs_log_ungrant_exit(log, ticket);
2705 
2706 	xfs_log_space_wake(log->l_mp);
2707 }
2708 
2709 /*
2710  * Flush iclog to disk if this is the last reference to the given iclog and
2711  * the WANT_SYNC bit is set.
2712  *
2713  * When this function is entered, the iclog is not necessarily in the
2714  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2715  *
2716  *
2717  */
2718 STATIC int
2719 xlog_state_release_iclog(
2720 	xlog_t		*log,
2721 	xlog_in_core_t	*iclog)
2722 {
2723 	int		sync = 0;	/* do we sync? */
2724 
2725 	if (iclog->ic_state & XLOG_STATE_IOERROR)
2726 		return XFS_ERROR(EIO);
2727 
2728 	ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2729 	if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2730 		return 0;
2731 
2732 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2733 		spin_unlock(&log->l_icloglock);
2734 		return XFS_ERROR(EIO);
2735 	}
2736 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2737 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
2738 
2739 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2740 		/* update tail before writing to iclog */
2741 		xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2742 		sync++;
2743 		iclog->ic_state = XLOG_STATE_SYNCING;
2744 		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2745 		xlog_verify_tail_lsn(log, iclog, tail_lsn);
2746 		/* cycle incremented when incrementing curr_block */
2747 	}
2748 	spin_unlock(&log->l_icloglock);
2749 
2750 	/*
2751 	 * We let the log lock go, so it's possible that we hit a log I/O
2752 	 * error or some other SHUTDOWN condition that marks the iclog
2753 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2754 	 * this iclog has consistent data, so we ignore IOERROR
2755 	 * flags after this point.
2756 	 */
2757 	if (sync)
2758 		return xlog_sync(log, iclog);
2759 	return 0;
2760 }	/* xlog_state_release_iclog */
2761 
2762 
2763 /*
2764  * This routine will mark the current iclog in the ring as WANT_SYNC
2765  * and move the current iclog pointer to the next iclog in the ring.
2766  * When this routine is called from xlog_state_get_iclog_space(), the
2767  * exact size of the iclog has not yet been determined.  All we know is
2768  * that every data block.  We have run out of space in this log record.
2769  */
2770 STATIC void
2771 xlog_state_switch_iclogs(xlog_t		*log,
2772 			 xlog_in_core_t *iclog,
2773 			 int		eventual_size)
2774 {
2775 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2776 	if (!eventual_size)
2777 		eventual_size = iclog->ic_offset;
2778 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2779 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2780 	log->l_prev_block = log->l_curr_block;
2781 	log->l_prev_cycle = log->l_curr_cycle;
2782 
2783 	/* roll log?: ic_offset changed later */
2784 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2785 
2786 	/* Round up to next log-sunit */
2787 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2788 	    log->l_mp->m_sb.sb_logsunit > 1) {
2789 		__uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2790 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2791 	}
2792 
2793 	if (log->l_curr_block >= log->l_logBBsize) {
2794 		log->l_curr_cycle++;
2795 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2796 			log->l_curr_cycle++;
2797 		log->l_curr_block -= log->l_logBBsize;
2798 		ASSERT(log->l_curr_block >= 0);
2799 	}
2800 	ASSERT(iclog == log->l_iclog);
2801 	log->l_iclog = iclog->ic_next;
2802 }	/* xlog_state_switch_iclogs */
2803 
2804 /*
2805  * Write out all data in the in-core log as of this exact moment in time.
2806  *
2807  * Data may be written to the in-core log during this call.  However,
2808  * we don't guarantee this data will be written out.  A change from past
2809  * implementation means this routine will *not* write out zero length LRs.
2810  *
2811  * Basically, we try and perform an intelligent scan of the in-core logs.
2812  * If we determine there is no flushable data, we just return.  There is no
2813  * flushable data if:
2814  *
2815  *	1. the current iclog is active and has no data; the previous iclog
2816  *		is in the active or dirty state.
2817  *	2. the current iclog is drity, and the previous iclog is in the
2818  *		active or dirty state.
2819  *
2820  * We may sleep if:
2821  *
2822  *	1. the current iclog is not in the active nor dirty state.
2823  *	2. the current iclog dirty, and the previous iclog is not in the
2824  *		active nor dirty state.
2825  *	3. the current iclog is active, and there is another thread writing
2826  *		to this particular iclog.
2827  *	4. a) the current iclog is active and has no other writers
2828  *	   b) when we return from flushing out this iclog, it is still
2829  *		not in the active nor dirty state.
2830  */
2831 int
2832 _xfs_log_force(
2833 	struct xfs_mount	*mp,
2834 	uint			flags,
2835 	int			*log_flushed)
2836 {
2837 	struct xlog		*log = mp->m_log;
2838 	struct xlog_in_core	*iclog;
2839 	xfs_lsn_t		lsn;
2840 
2841 	XFS_STATS_INC(xs_log_force);
2842 
2843 	xlog_cil_force(log);
2844 
2845 	spin_lock(&log->l_icloglock);
2846 
2847 	iclog = log->l_iclog;
2848 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2849 		spin_unlock(&log->l_icloglock);
2850 		return XFS_ERROR(EIO);
2851 	}
2852 
2853 	/* If the head iclog is not active nor dirty, we just attach
2854 	 * ourselves to the head and go to sleep.
2855 	 */
2856 	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2857 	    iclog->ic_state == XLOG_STATE_DIRTY) {
2858 		/*
2859 		 * If the head is dirty or (active and empty), then
2860 		 * we need to look at the previous iclog.  If the previous
2861 		 * iclog is active or dirty we are done.  There is nothing
2862 		 * to sync out.  Otherwise, we attach ourselves to the
2863 		 * previous iclog and go to sleep.
2864 		 */
2865 		if (iclog->ic_state == XLOG_STATE_DIRTY ||
2866 		    (atomic_read(&iclog->ic_refcnt) == 0
2867 		     && iclog->ic_offset == 0)) {
2868 			iclog = iclog->ic_prev;
2869 			if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2870 			    iclog->ic_state == XLOG_STATE_DIRTY)
2871 				goto no_sleep;
2872 			else
2873 				goto maybe_sleep;
2874 		} else {
2875 			if (atomic_read(&iclog->ic_refcnt) == 0) {
2876 				/* We are the only one with access to this
2877 				 * iclog.  Flush it out now.  There should
2878 				 * be a roundoff of zero to show that someone
2879 				 * has already taken care of the roundoff from
2880 				 * the previous sync.
2881 				 */
2882 				atomic_inc(&iclog->ic_refcnt);
2883 				lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2884 				xlog_state_switch_iclogs(log, iclog, 0);
2885 				spin_unlock(&log->l_icloglock);
2886 
2887 				if (xlog_state_release_iclog(log, iclog))
2888 					return XFS_ERROR(EIO);
2889 
2890 				if (log_flushed)
2891 					*log_flushed = 1;
2892 				spin_lock(&log->l_icloglock);
2893 				if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2894 				    iclog->ic_state != XLOG_STATE_DIRTY)
2895 					goto maybe_sleep;
2896 				else
2897 					goto no_sleep;
2898 			} else {
2899 				/* Someone else is writing to this iclog.
2900 				 * Use its call to flush out the data.  However,
2901 				 * the other thread may not force out this LR,
2902 				 * so we mark it WANT_SYNC.
2903 				 */
2904 				xlog_state_switch_iclogs(log, iclog, 0);
2905 				goto maybe_sleep;
2906 			}
2907 		}
2908 	}
2909 
2910 	/* By the time we come around again, the iclog could've been filled
2911 	 * which would give it another lsn.  If we have a new lsn, just
2912 	 * return because the relevant data has been flushed.
2913 	 */
2914 maybe_sleep:
2915 	if (flags & XFS_LOG_SYNC) {
2916 		/*
2917 		 * We must check if we're shutting down here, before
2918 		 * we wait, while we're holding the l_icloglock.
2919 		 * Then we check again after waking up, in case our
2920 		 * sleep was disturbed by a bad news.
2921 		 */
2922 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
2923 			spin_unlock(&log->l_icloglock);
2924 			return XFS_ERROR(EIO);
2925 		}
2926 		XFS_STATS_INC(xs_log_force_sleep);
2927 		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2928 		/*
2929 		 * No need to grab the log lock here since we're
2930 		 * only deciding whether or not to return EIO
2931 		 * and the memory read should be atomic.
2932 		 */
2933 		if (iclog->ic_state & XLOG_STATE_IOERROR)
2934 			return XFS_ERROR(EIO);
2935 		if (log_flushed)
2936 			*log_flushed = 1;
2937 	} else {
2938 
2939 no_sleep:
2940 		spin_unlock(&log->l_icloglock);
2941 	}
2942 	return 0;
2943 }
2944 
2945 /*
2946  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2947  * about errors or whether the log was flushed or not. This is the normal
2948  * interface to use when trying to unpin items or move the log forward.
2949  */
2950 void
2951 xfs_log_force(
2952 	xfs_mount_t	*mp,
2953 	uint		flags)
2954 {
2955 	int	error;
2956 
2957 	trace_xfs_log_force(mp, 0);
2958 	error = _xfs_log_force(mp, flags, NULL);
2959 	if (error)
2960 		xfs_warn(mp, "%s: error %d returned.", __func__, error);
2961 }
2962 
2963 /*
2964  * Force the in-core log to disk for a specific LSN.
2965  *
2966  * Find in-core log with lsn.
2967  *	If it is in the DIRTY state, just return.
2968  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2969  *		state and go to sleep or return.
2970  *	If it is in any other state, go to sleep or return.
2971  *
2972  * Synchronous forces are implemented with a signal variable. All callers
2973  * to force a given lsn to disk will wait on a the sv attached to the
2974  * specific in-core log.  When given in-core log finally completes its
2975  * write to disk, that thread will wake up all threads waiting on the
2976  * sv.
2977  */
2978 int
2979 _xfs_log_force_lsn(
2980 	struct xfs_mount	*mp,
2981 	xfs_lsn_t		lsn,
2982 	uint			flags,
2983 	int			*log_flushed)
2984 {
2985 	struct xlog		*log = mp->m_log;
2986 	struct xlog_in_core	*iclog;
2987 	int			already_slept = 0;
2988 
2989 	ASSERT(lsn != 0);
2990 
2991 	XFS_STATS_INC(xs_log_force);
2992 
2993 	lsn = xlog_cil_force_lsn(log, lsn);
2994 	if (lsn == NULLCOMMITLSN)
2995 		return 0;
2996 
2997 try_again:
2998 	spin_lock(&log->l_icloglock);
2999 	iclog = log->l_iclog;
3000 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
3001 		spin_unlock(&log->l_icloglock);
3002 		return XFS_ERROR(EIO);
3003 	}
3004 
3005 	do {
3006 		if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3007 			iclog = iclog->ic_next;
3008 			continue;
3009 		}
3010 
3011 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
3012 			spin_unlock(&log->l_icloglock);
3013 			return 0;
3014 		}
3015 
3016 		if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3017 			/*
3018 			 * We sleep here if we haven't already slept (e.g.
3019 			 * this is the first time we've looked at the correct
3020 			 * iclog buf) and the buffer before us is going to
3021 			 * be sync'ed. The reason for this is that if we
3022 			 * are doing sync transactions here, by waiting for
3023 			 * the previous I/O to complete, we can allow a few
3024 			 * more transactions into this iclog before we close
3025 			 * it down.
3026 			 *
3027 			 * Otherwise, we mark the buffer WANT_SYNC, and bump
3028 			 * up the refcnt so we can release the log (which
3029 			 * drops the ref count).  The state switch keeps new
3030 			 * transaction commits from using this buffer.  When
3031 			 * the current commits finish writing into the buffer,
3032 			 * the refcount will drop to zero and the buffer will
3033 			 * go out then.
3034 			 */
3035 			if (!already_slept &&
3036 			    (iclog->ic_prev->ic_state &
3037 			     (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3038 				ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3039 
3040 				XFS_STATS_INC(xs_log_force_sleep);
3041 
3042 				xlog_wait(&iclog->ic_prev->ic_write_wait,
3043 							&log->l_icloglock);
3044 				if (log_flushed)
3045 					*log_flushed = 1;
3046 				already_slept = 1;
3047 				goto try_again;
3048 			}
3049 			atomic_inc(&iclog->ic_refcnt);
3050 			xlog_state_switch_iclogs(log, iclog, 0);
3051 			spin_unlock(&log->l_icloglock);
3052 			if (xlog_state_release_iclog(log, iclog))
3053 				return XFS_ERROR(EIO);
3054 			if (log_flushed)
3055 				*log_flushed = 1;
3056 			spin_lock(&log->l_icloglock);
3057 		}
3058 
3059 		if ((flags & XFS_LOG_SYNC) && /* sleep */
3060 		    !(iclog->ic_state &
3061 		      (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3062 			/*
3063 			 * Don't wait on completion if we know that we've
3064 			 * gotten a log write error.
3065 			 */
3066 			if (iclog->ic_state & XLOG_STATE_IOERROR) {
3067 				spin_unlock(&log->l_icloglock);
3068 				return XFS_ERROR(EIO);
3069 			}
3070 			XFS_STATS_INC(xs_log_force_sleep);
3071 			xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3072 			/*
3073 			 * No need to grab the log lock here since we're
3074 			 * only deciding whether or not to return EIO
3075 			 * and the memory read should be atomic.
3076 			 */
3077 			if (iclog->ic_state & XLOG_STATE_IOERROR)
3078 				return XFS_ERROR(EIO);
3079 
3080 			if (log_flushed)
3081 				*log_flushed = 1;
3082 		} else {		/* just return */
3083 			spin_unlock(&log->l_icloglock);
3084 		}
3085 
3086 		return 0;
3087 	} while (iclog != log->l_iclog);
3088 
3089 	spin_unlock(&log->l_icloglock);
3090 	return 0;
3091 }
3092 
3093 /*
3094  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3095  * about errors or whether the log was flushed or not. This is the normal
3096  * interface to use when trying to unpin items or move the log forward.
3097  */
3098 void
3099 xfs_log_force_lsn(
3100 	xfs_mount_t	*mp,
3101 	xfs_lsn_t	lsn,
3102 	uint		flags)
3103 {
3104 	int	error;
3105 
3106 	trace_xfs_log_force(mp, lsn);
3107 	error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3108 	if (error)
3109 		xfs_warn(mp, "%s: error %d returned.", __func__, error);
3110 }
3111 
3112 /*
3113  * Called when we want to mark the current iclog as being ready to sync to
3114  * disk.
3115  */
3116 STATIC void
3117 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3118 {
3119 	assert_spin_locked(&log->l_icloglock);
3120 
3121 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3122 		xlog_state_switch_iclogs(log, iclog, 0);
3123 	} else {
3124 		ASSERT(iclog->ic_state &
3125 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3126 	}
3127 }
3128 
3129 
3130 /*****************************************************************************
3131  *
3132  *		TICKET functions
3133  *
3134  *****************************************************************************
3135  */
3136 
3137 /*
3138  * Free a used ticket when its refcount falls to zero.
3139  */
3140 void
3141 xfs_log_ticket_put(
3142 	xlog_ticket_t	*ticket)
3143 {
3144 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3145 	if (atomic_dec_and_test(&ticket->t_ref))
3146 		kmem_zone_free(xfs_log_ticket_zone, ticket);
3147 }
3148 
3149 xlog_ticket_t *
3150 xfs_log_ticket_get(
3151 	xlog_ticket_t	*ticket)
3152 {
3153 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3154 	atomic_inc(&ticket->t_ref);
3155 	return ticket;
3156 }
3157 
3158 /*
3159  * Allocate and initialise a new log ticket.
3160  */
3161 xlog_ticket_t *
3162 xlog_ticket_alloc(
3163 	struct xlog	*log,
3164 	int		unit_bytes,
3165 	int		cnt,
3166 	char		client,
3167 	bool		permanent,
3168 	xfs_km_flags_t	alloc_flags)
3169 {
3170 	struct xlog_ticket *tic;
3171 	uint		num_headers;
3172 	int		iclog_space;
3173 
3174 	tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3175 	if (!tic)
3176 		return NULL;
3177 
3178 	/*
3179 	 * Permanent reservations have up to 'cnt'-1 active log operations
3180 	 * in the log.  A unit in this case is the amount of space for one
3181 	 * of these log operations.  Normal reservations have a cnt of 1
3182 	 * and their unit amount is the total amount of space required.
3183 	 *
3184 	 * The following lines of code account for non-transaction data
3185 	 * which occupy space in the on-disk log.
3186 	 *
3187 	 * Normal form of a transaction is:
3188 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3189 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3190 	 *
3191 	 * We need to account for all the leadup data and trailer data
3192 	 * around the transaction data.
3193 	 * And then we need to account for the worst case in terms of using
3194 	 * more space.
3195 	 * The worst case will happen if:
3196 	 * - the placement of the transaction happens to be such that the
3197 	 *   roundoff is at its maximum
3198 	 * - the transaction data is synced before the commit record is synced
3199 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3200 	 *   Therefore the commit record is in its own Log Record.
3201 	 *   This can happen as the commit record is called with its
3202 	 *   own region to xlog_write().
3203 	 *   This then means that in the worst case, roundoff can happen for
3204 	 *   the commit-rec as well.
3205 	 *   The commit-rec is smaller than padding in this scenario and so it is
3206 	 *   not added separately.
3207 	 */
3208 
3209 	/* for trans header */
3210 	unit_bytes += sizeof(xlog_op_header_t);
3211 	unit_bytes += sizeof(xfs_trans_header_t);
3212 
3213 	/* for start-rec */
3214 	unit_bytes += sizeof(xlog_op_header_t);
3215 
3216 	/*
3217 	 * for LR headers - the space for data in an iclog is the size minus
3218 	 * the space used for the headers. If we use the iclog size, then we
3219 	 * undercalculate the number of headers required.
3220 	 *
3221 	 * Furthermore - the addition of op headers for split-recs might
3222 	 * increase the space required enough to require more log and op
3223 	 * headers, so take that into account too.
3224 	 *
3225 	 * IMPORTANT: This reservation makes the assumption that if this
3226 	 * transaction is the first in an iclog and hence has the LR headers
3227 	 * accounted to it, then the remaining space in the iclog is
3228 	 * exclusively for this transaction.  i.e. if the transaction is larger
3229 	 * than the iclog, it will be the only thing in that iclog.
3230 	 * Fundamentally, this means we must pass the entire log vector to
3231 	 * xlog_write to guarantee this.
3232 	 */
3233 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3234 	num_headers = howmany(unit_bytes, iclog_space);
3235 
3236 	/* for split-recs - ophdrs added when data split over LRs */
3237 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3238 
3239 	/* add extra header reservations if we overrun */
3240 	while (!num_headers ||
3241 	       howmany(unit_bytes, iclog_space) > num_headers) {
3242 		unit_bytes += sizeof(xlog_op_header_t);
3243 		num_headers++;
3244 	}
3245 	unit_bytes += log->l_iclog_hsize * num_headers;
3246 
3247 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3248 	unit_bytes += log->l_iclog_hsize;
3249 
3250 	/* for roundoff padding for transaction data and one for commit record */
3251 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3252 	    log->l_mp->m_sb.sb_logsunit > 1) {
3253 		/* log su roundoff */
3254 		unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3255 	} else {
3256 		/* BB roundoff */
3257 		unit_bytes += 2*BBSIZE;
3258         }
3259 
3260 	atomic_set(&tic->t_ref, 1);
3261 	tic->t_task		= current;
3262 	INIT_LIST_HEAD(&tic->t_queue);
3263 	tic->t_unit_res		= unit_bytes;
3264 	tic->t_curr_res		= unit_bytes;
3265 	tic->t_cnt		= cnt;
3266 	tic->t_ocnt		= cnt;
3267 	tic->t_tid		= random32();
3268 	tic->t_clientid		= client;
3269 	tic->t_flags		= XLOG_TIC_INITED;
3270 	tic->t_trans_type	= 0;
3271 	if (permanent)
3272 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3273 
3274 	xlog_tic_reset_res(tic);
3275 
3276 	return tic;
3277 }
3278 
3279 
3280 /******************************************************************************
3281  *
3282  *		Log debug routines
3283  *
3284  ******************************************************************************
3285  */
3286 #if defined(DEBUG)
3287 /*
3288  * Make sure that the destination ptr is within the valid data region of
3289  * one of the iclogs.  This uses backup pointers stored in a different
3290  * part of the log in case we trash the log structure.
3291  */
3292 void
3293 xlog_verify_dest_ptr(
3294 	struct xlog	*log,
3295 	char		*ptr)
3296 {
3297 	int i;
3298 	int good_ptr = 0;
3299 
3300 	for (i = 0; i < log->l_iclog_bufs; i++) {
3301 		if (ptr >= log->l_iclog_bak[i] &&
3302 		    ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3303 			good_ptr++;
3304 	}
3305 
3306 	if (!good_ptr)
3307 		xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3308 }
3309 
3310 /*
3311  * Check to make sure the grant write head didn't just over lap the tail.  If
3312  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3313  * the cycles differ by exactly one and check the byte count.
3314  *
3315  * This check is run unlocked, so can give false positives. Rather than assert
3316  * on failures, use a warn-once flag and a panic tag to allow the admin to
3317  * determine if they want to panic the machine when such an error occurs. For
3318  * debug kernels this will have the same effect as using an assert but, unlinke
3319  * an assert, it can be turned off at runtime.
3320  */
3321 STATIC void
3322 xlog_verify_grant_tail(
3323 	struct xlog	*log)
3324 {
3325 	int		tail_cycle, tail_blocks;
3326 	int		cycle, space;
3327 
3328 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3329 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3330 	if (tail_cycle != cycle) {
3331 		if (cycle - 1 != tail_cycle &&
3332 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3333 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3334 				"%s: cycle - 1 != tail_cycle", __func__);
3335 			log->l_flags |= XLOG_TAIL_WARN;
3336 		}
3337 
3338 		if (space > BBTOB(tail_blocks) &&
3339 		    !(log->l_flags & XLOG_TAIL_WARN)) {
3340 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3341 				"%s: space > BBTOB(tail_blocks)", __func__);
3342 			log->l_flags |= XLOG_TAIL_WARN;
3343 		}
3344 	}
3345 }
3346 
3347 /* check if it will fit */
3348 STATIC void
3349 xlog_verify_tail_lsn(xlog_t	    *log,
3350 		     xlog_in_core_t *iclog,
3351 		     xfs_lsn_t	    tail_lsn)
3352 {
3353     int blocks;
3354 
3355     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3356 	blocks =
3357 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3358 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3359 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3360     } else {
3361 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3362 
3363 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3364 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3365 
3366 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3367 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3368 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3369     }
3370 }	/* xlog_verify_tail_lsn */
3371 
3372 /*
3373  * Perform a number of checks on the iclog before writing to disk.
3374  *
3375  * 1. Make sure the iclogs are still circular
3376  * 2. Make sure we have a good magic number
3377  * 3. Make sure we don't have magic numbers in the data
3378  * 4. Check fields of each log operation header for:
3379  *	A. Valid client identifier
3380  *	B. tid ptr value falls in valid ptr space (user space code)
3381  *	C. Length in log record header is correct according to the
3382  *		individual operation headers within record.
3383  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3384  *	log, check the preceding blocks of the physical log to make sure all
3385  *	the cycle numbers agree with the current cycle number.
3386  */
3387 STATIC void
3388 xlog_verify_iclog(xlog_t	 *log,
3389 		  xlog_in_core_t *iclog,
3390 		  int		 count,
3391 		  boolean_t	 syncing)
3392 {
3393 	xlog_op_header_t	*ophead;
3394 	xlog_in_core_t		*icptr;
3395 	xlog_in_core_2_t	*xhdr;
3396 	xfs_caddr_t		ptr;
3397 	xfs_caddr_t		base_ptr;
3398 	__psint_t		field_offset;
3399 	__uint8_t		clientid;
3400 	int			len, i, j, k, op_len;
3401 	int			idx;
3402 
3403 	/* check validity of iclog pointers */
3404 	spin_lock(&log->l_icloglock);
3405 	icptr = log->l_iclog;
3406 	for (i=0; i < log->l_iclog_bufs; i++) {
3407 		if (icptr == NULL)
3408 			xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3409 		icptr = icptr->ic_next;
3410 	}
3411 	if (icptr != log->l_iclog)
3412 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3413 	spin_unlock(&log->l_icloglock);
3414 
3415 	/* check log magic numbers */
3416 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3417 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3418 
3419 	ptr = (xfs_caddr_t) &iclog->ic_header;
3420 	for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3421 	     ptr += BBSIZE) {
3422 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3423 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3424 				__func__);
3425 	}
3426 
3427 	/* check fields */
3428 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3429 	ptr = iclog->ic_datap;
3430 	base_ptr = ptr;
3431 	ophead = (xlog_op_header_t *)ptr;
3432 	xhdr = iclog->ic_data;
3433 	for (i = 0; i < len; i++) {
3434 		ophead = (xlog_op_header_t *)ptr;
3435 
3436 		/* clientid is only 1 byte */
3437 		field_offset = (__psint_t)
3438 			       ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3439 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3440 			clientid = ophead->oh_clientid;
3441 		} else {
3442 			idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3443 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3444 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3445 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3446 				clientid = xlog_get_client_id(
3447 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3448 			} else {
3449 				clientid = xlog_get_client_id(
3450 					iclog->ic_header.h_cycle_data[idx]);
3451 			}
3452 		}
3453 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3454 			xfs_warn(log->l_mp,
3455 				"%s: invalid clientid %d op 0x%p offset 0x%lx",
3456 				__func__, clientid, ophead,
3457 				(unsigned long)field_offset);
3458 
3459 		/* check length */
3460 		field_offset = (__psint_t)
3461 			       ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3462 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3463 			op_len = be32_to_cpu(ophead->oh_len);
3464 		} else {
3465 			idx = BTOBBT((__psint_t)&ophead->oh_len -
3466 				    (__psint_t)iclog->ic_datap);
3467 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3468 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3469 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3470 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3471 			} else {
3472 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3473 			}
3474 		}
3475 		ptr += sizeof(xlog_op_header_t) + op_len;
3476 	}
3477 }	/* xlog_verify_iclog */
3478 #endif
3479 
3480 /*
3481  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3482  */
3483 STATIC int
3484 xlog_state_ioerror(
3485 	xlog_t	*log)
3486 {
3487 	xlog_in_core_t	*iclog, *ic;
3488 
3489 	iclog = log->l_iclog;
3490 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3491 		/*
3492 		 * Mark all the incore logs IOERROR.
3493 		 * From now on, no log flushes will result.
3494 		 */
3495 		ic = iclog;
3496 		do {
3497 			ic->ic_state = XLOG_STATE_IOERROR;
3498 			ic = ic->ic_next;
3499 		} while (ic != iclog);
3500 		return 0;
3501 	}
3502 	/*
3503 	 * Return non-zero, if state transition has already happened.
3504 	 */
3505 	return 1;
3506 }
3507 
3508 /*
3509  * This is called from xfs_force_shutdown, when we're forcibly
3510  * shutting down the filesystem, typically because of an IO error.
3511  * Our main objectives here are to make sure that:
3512  *	a. the filesystem gets marked 'SHUTDOWN' for all interested
3513  *	   parties to find out, 'atomically'.
3514  *	b. those who're sleeping on log reservations, pinned objects and
3515  *	    other resources get woken up, and be told the bad news.
3516  *	c. nothing new gets queued up after (a) and (b) are done.
3517  *	d. if !logerror, flush the iclogs to disk, then seal them off
3518  *	   for business.
3519  *
3520  * Note: for delayed logging the !logerror case needs to flush the regions
3521  * held in memory out to the iclogs before flushing them to disk. This needs
3522  * to be done before the log is marked as shutdown, otherwise the flush to the
3523  * iclogs will fail.
3524  */
3525 int
3526 xfs_log_force_umount(
3527 	struct xfs_mount	*mp,
3528 	int			logerror)
3529 {
3530 	xlog_t		*log;
3531 	int		retval;
3532 
3533 	log = mp->m_log;
3534 
3535 	/*
3536 	 * If this happens during log recovery, don't worry about
3537 	 * locking; the log isn't open for business yet.
3538 	 */
3539 	if (!log ||
3540 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3541 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3542 		if (mp->m_sb_bp)
3543 			XFS_BUF_DONE(mp->m_sb_bp);
3544 		return 0;
3545 	}
3546 
3547 	/*
3548 	 * Somebody could've already done the hard work for us.
3549 	 * No need to get locks for this.
3550 	 */
3551 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3552 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3553 		return 1;
3554 	}
3555 	retval = 0;
3556 
3557 	/*
3558 	 * Flush the in memory commit item list before marking the log as
3559 	 * being shut down. We need to do it in this order to ensure all the
3560 	 * completed transactions are flushed to disk with the xfs_log_force()
3561 	 * call below.
3562 	 */
3563 	if (!logerror)
3564 		xlog_cil_force(log);
3565 
3566 	/*
3567 	 * mark the filesystem and the as in a shutdown state and wake
3568 	 * everybody up to tell them the bad news.
3569 	 */
3570 	spin_lock(&log->l_icloglock);
3571 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3572 	if (mp->m_sb_bp)
3573 		XFS_BUF_DONE(mp->m_sb_bp);
3574 
3575 	/*
3576 	 * This flag is sort of redundant because of the mount flag, but
3577 	 * it's good to maintain the separation between the log and the rest
3578 	 * of XFS.
3579 	 */
3580 	log->l_flags |= XLOG_IO_ERROR;
3581 
3582 	/*
3583 	 * If we hit a log error, we want to mark all the iclogs IOERROR
3584 	 * while we're still holding the loglock.
3585 	 */
3586 	if (logerror)
3587 		retval = xlog_state_ioerror(log);
3588 	spin_unlock(&log->l_icloglock);
3589 
3590 	/*
3591 	 * We don't want anybody waiting for log reservations after this. That
3592 	 * means we have to wake up everybody queued up on reserveq as well as
3593 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3594 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3595 	 * action is protected by the grant locks.
3596 	 */
3597 	xlog_grant_head_wake_all(&log->l_reserve_head);
3598 	xlog_grant_head_wake_all(&log->l_write_head);
3599 
3600 	if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3601 		ASSERT(!logerror);
3602 		/*
3603 		 * Force the incore logs to disk before shutting the
3604 		 * log down completely.
3605 		 */
3606 		_xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3607 
3608 		spin_lock(&log->l_icloglock);
3609 		retval = xlog_state_ioerror(log);
3610 		spin_unlock(&log->l_icloglock);
3611 	}
3612 	/*
3613 	 * Wake up everybody waiting on xfs_log_force.
3614 	 * Callback all log item committed functions as if the
3615 	 * log writes were completed.
3616 	 */
3617 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3618 
3619 #ifdef XFSERRORDEBUG
3620 	{
3621 		xlog_in_core_t	*iclog;
3622 
3623 		spin_lock(&log->l_icloglock);
3624 		iclog = log->l_iclog;
3625 		do {
3626 			ASSERT(iclog->ic_callback == 0);
3627 			iclog = iclog->ic_next;
3628 		} while (iclog != log->l_iclog);
3629 		spin_unlock(&log->l_icloglock);
3630 	}
3631 #endif
3632 	/* return non-zero if log IOERROR transition had already happened */
3633 	return retval;
3634 }
3635 
3636 STATIC int
3637 xlog_iclogs_empty(xlog_t *log)
3638 {
3639 	xlog_in_core_t	*iclog;
3640 
3641 	iclog = log->l_iclog;
3642 	do {
3643 		/* endianness does not matter here, zero is zero in
3644 		 * any language.
3645 		 */
3646 		if (iclog->ic_header.h_num_logops)
3647 			return 0;
3648 		iclog = iclog->ic_next;
3649 	} while (iclog != log->l_iclog);
3650 	return 1;
3651 }
3652