1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_errortag.h"
14 #include "xfs_error.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
17 #include "xfs_log.h"
18 #include "xfs_log_priv.h"
19 #include "xfs_trace.h"
20 #include "xfs_sysfs.h"
21 #include "xfs_sb.h"
22 #include "xfs_health.h"
23 
24 struct kmem_cache	*xfs_log_ticket_cache;
25 
26 /* Local miscellaneous function prototypes */
27 STATIC struct xlog *
28 xlog_alloc_log(
29 	struct xfs_mount	*mp,
30 	struct xfs_buftarg	*log_target,
31 	xfs_daddr_t		blk_offset,
32 	int			num_bblks);
33 STATIC int
34 xlog_space_left(
35 	struct xlog		*log,
36 	atomic64_t		*head);
37 STATIC void
38 xlog_dealloc_log(
39 	struct xlog		*log);
40 
41 /* local state machine functions */
42 STATIC void xlog_state_done_syncing(
43 	struct xlog_in_core	*iclog);
44 STATIC void xlog_state_do_callback(
45 	struct xlog		*log);
46 STATIC int
47 xlog_state_get_iclog_space(
48 	struct xlog		*log,
49 	int			len,
50 	struct xlog_in_core	**iclog,
51 	struct xlog_ticket	*ticket,
52 	int			*logoffsetp);
53 STATIC void
54 xlog_grant_push_ail(
55 	struct xlog		*log,
56 	int			need_bytes);
57 STATIC void
58 xlog_sync(
59 	struct xlog		*log,
60 	struct xlog_in_core	*iclog,
61 	struct xlog_ticket	*ticket);
62 #if defined(DEBUG)
63 STATIC void
64 xlog_verify_grant_tail(
65 	struct xlog *log);
66 STATIC void
67 xlog_verify_iclog(
68 	struct xlog		*log,
69 	struct xlog_in_core	*iclog,
70 	int			count);
71 STATIC void
72 xlog_verify_tail_lsn(
73 	struct xlog		*log,
74 	struct xlog_in_core	*iclog);
75 #else
76 #define xlog_verify_grant_tail(a)
77 #define xlog_verify_iclog(a,b,c)
78 #define xlog_verify_tail_lsn(a,b)
79 #endif
80 
81 STATIC int
82 xlog_iclogs_empty(
83 	struct xlog		*log);
84 
85 static int
86 xfs_log_cover(struct xfs_mount *);
87 
88 /*
89  * We need to make sure the buffer pointer returned is naturally aligned for the
90  * biggest basic data type we put into it. We have already accounted for this
91  * padding when sizing the buffer.
92  *
93  * However, this padding does not get written into the log, and hence we have to
94  * track the space used by the log vectors separately to prevent log space hangs
95  * due to inaccurate accounting (i.e. a leak) of the used log space through the
96  * CIL context ticket.
97  *
98  * We also add space for the xlog_op_header that describes this region in the
99  * log. This prepends the data region we return to the caller to copy their data
100  * into, so do all the static initialisation of the ophdr now. Because the ophdr
101  * is not 8 byte aligned, we have to be careful to ensure that we align the
102  * start of the buffer such that the region we return to the call is 8 byte
103  * aligned and packed against the tail of the ophdr.
104  */
105 void *
xlog_prepare_iovec(struct xfs_log_vec * lv,struct xfs_log_iovec ** vecp,uint type)106 xlog_prepare_iovec(
107 	struct xfs_log_vec	*lv,
108 	struct xfs_log_iovec	**vecp,
109 	uint			type)
110 {
111 	struct xfs_log_iovec	*vec = *vecp;
112 	struct xlog_op_header	*oph;
113 	uint32_t		len;
114 	void			*buf;
115 
116 	if (vec) {
117 		ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
118 		vec++;
119 	} else {
120 		vec = &lv->lv_iovecp[0];
121 	}
122 
123 	len = lv->lv_buf_len + sizeof(struct xlog_op_header);
124 	if (!IS_ALIGNED(len, sizeof(uint64_t))) {
125 		lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
126 					sizeof(struct xlog_op_header);
127 	}
128 
129 	vec->i_type = type;
130 	vec->i_addr = lv->lv_buf + lv->lv_buf_len;
131 
132 	oph = vec->i_addr;
133 	oph->oh_clientid = XFS_TRANSACTION;
134 	oph->oh_res2 = 0;
135 	oph->oh_flags = 0;
136 
137 	buf = vec->i_addr + sizeof(struct xlog_op_header);
138 	ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
139 
140 	*vecp = vec;
141 	return buf;
142 }
143 
144 static void
xlog_grant_sub_space(struct xlog * log,atomic64_t * head,int bytes)145 xlog_grant_sub_space(
146 	struct xlog		*log,
147 	atomic64_t		*head,
148 	int			bytes)
149 {
150 	int64_t	head_val = atomic64_read(head);
151 	int64_t new, old;
152 
153 	do {
154 		int	cycle, space;
155 
156 		xlog_crack_grant_head_val(head_val, &cycle, &space);
157 
158 		space -= bytes;
159 		if (space < 0) {
160 			space += log->l_logsize;
161 			cycle--;
162 		}
163 
164 		old = head_val;
165 		new = xlog_assign_grant_head_val(cycle, space);
166 		head_val = atomic64_cmpxchg(head, old, new);
167 	} while (head_val != old);
168 }
169 
170 static void
xlog_grant_add_space(struct xlog * log,atomic64_t * head,int bytes)171 xlog_grant_add_space(
172 	struct xlog		*log,
173 	atomic64_t		*head,
174 	int			bytes)
175 {
176 	int64_t	head_val = atomic64_read(head);
177 	int64_t new, old;
178 
179 	do {
180 		int		tmp;
181 		int		cycle, space;
182 
183 		xlog_crack_grant_head_val(head_val, &cycle, &space);
184 
185 		tmp = log->l_logsize - space;
186 		if (tmp > bytes)
187 			space += bytes;
188 		else {
189 			space = bytes - tmp;
190 			cycle++;
191 		}
192 
193 		old = head_val;
194 		new = xlog_assign_grant_head_val(cycle, space);
195 		head_val = atomic64_cmpxchg(head, old, new);
196 	} while (head_val != old);
197 }
198 
199 STATIC void
xlog_grant_head_init(struct xlog_grant_head * head)200 xlog_grant_head_init(
201 	struct xlog_grant_head	*head)
202 {
203 	xlog_assign_grant_head(&head->grant, 1, 0);
204 	INIT_LIST_HEAD(&head->waiters);
205 	spin_lock_init(&head->lock);
206 }
207 
208 STATIC void
xlog_grant_head_wake_all(struct xlog_grant_head * head)209 xlog_grant_head_wake_all(
210 	struct xlog_grant_head	*head)
211 {
212 	struct xlog_ticket	*tic;
213 
214 	spin_lock(&head->lock);
215 	list_for_each_entry(tic, &head->waiters, t_queue)
216 		wake_up_process(tic->t_task);
217 	spin_unlock(&head->lock);
218 }
219 
220 static inline int
xlog_ticket_reservation(struct xlog * log,struct xlog_grant_head * head,struct xlog_ticket * tic)221 xlog_ticket_reservation(
222 	struct xlog		*log,
223 	struct xlog_grant_head	*head,
224 	struct xlog_ticket	*tic)
225 {
226 	if (head == &log->l_write_head) {
227 		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
228 		return tic->t_unit_res;
229 	}
230 
231 	if (tic->t_flags & XLOG_TIC_PERM_RESERV)
232 		return tic->t_unit_res * tic->t_cnt;
233 
234 	return tic->t_unit_res;
235 }
236 
237 STATIC bool
xlog_grant_head_wake(struct xlog * log,struct xlog_grant_head * head,int * free_bytes)238 xlog_grant_head_wake(
239 	struct xlog		*log,
240 	struct xlog_grant_head	*head,
241 	int			*free_bytes)
242 {
243 	struct xlog_ticket	*tic;
244 	int			need_bytes;
245 	bool			woken_task = false;
246 
247 	list_for_each_entry(tic, &head->waiters, t_queue) {
248 
249 		/*
250 		 * There is a chance that the size of the CIL checkpoints in
251 		 * progress at the last AIL push target calculation resulted in
252 		 * limiting the target to the log head (l_last_sync_lsn) at the
253 		 * time. This may not reflect where the log head is now as the
254 		 * CIL checkpoints may have completed.
255 		 *
256 		 * Hence when we are woken here, it may be that the head of the
257 		 * log that has moved rather than the tail. As the tail didn't
258 		 * move, there still won't be space available for the
259 		 * reservation we require.  However, if the AIL has already
260 		 * pushed to the target defined by the old log head location, we
261 		 * will hang here waiting for something else to update the AIL
262 		 * push target.
263 		 *
264 		 * Therefore, if there isn't space to wake the first waiter on
265 		 * the grant head, we need to push the AIL again to ensure the
266 		 * target reflects both the current log tail and log head
267 		 * position before we wait for the tail to move again.
268 		 */
269 
270 		need_bytes = xlog_ticket_reservation(log, head, tic);
271 		if (*free_bytes < need_bytes) {
272 			if (!woken_task)
273 				xlog_grant_push_ail(log, need_bytes);
274 			return false;
275 		}
276 
277 		*free_bytes -= need_bytes;
278 		trace_xfs_log_grant_wake_up(log, tic);
279 		wake_up_process(tic->t_task);
280 		woken_task = true;
281 	}
282 
283 	return true;
284 }
285 
286 STATIC int
xlog_grant_head_wait(struct xlog * log,struct xlog_grant_head * head,struct xlog_ticket * tic,int need_bytes)287 xlog_grant_head_wait(
288 	struct xlog		*log,
289 	struct xlog_grant_head	*head,
290 	struct xlog_ticket	*tic,
291 	int			need_bytes) __releases(&head->lock)
292 					    __acquires(&head->lock)
293 {
294 	list_add_tail(&tic->t_queue, &head->waiters);
295 
296 	do {
297 		if (xlog_is_shutdown(log))
298 			goto shutdown;
299 		xlog_grant_push_ail(log, need_bytes);
300 
301 		__set_current_state(TASK_UNINTERRUPTIBLE);
302 		spin_unlock(&head->lock);
303 
304 		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
305 
306 		trace_xfs_log_grant_sleep(log, tic);
307 		schedule();
308 		trace_xfs_log_grant_wake(log, tic);
309 
310 		spin_lock(&head->lock);
311 		if (xlog_is_shutdown(log))
312 			goto shutdown;
313 	} while (xlog_space_left(log, &head->grant) < need_bytes);
314 
315 	list_del_init(&tic->t_queue);
316 	return 0;
317 shutdown:
318 	list_del_init(&tic->t_queue);
319 	return -EIO;
320 }
321 
322 /*
323  * Atomically get the log space required for a log ticket.
324  *
325  * Once a ticket gets put onto head->waiters, it will only return after the
326  * needed reservation is satisfied.
327  *
328  * This function is structured so that it has a lock free fast path. This is
329  * necessary because every new transaction reservation will come through this
330  * path. Hence any lock will be globally hot if we take it unconditionally on
331  * every pass.
332  *
333  * As tickets are only ever moved on and off head->waiters under head->lock, we
334  * only need to take that lock if we are going to add the ticket to the queue
335  * and sleep. We can avoid taking the lock if the ticket was never added to
336  * head->waiters because the t_queue list head will be empty and we hold the
337  * only reference to it so it can safely be checked unlocked.
338  */
339 STATIC int
xlog_grant_head_check(struct xlog * log,struct xlog_grant_head * head,struct xlog_ticket * tic,int * need_bytes)340 xlog_grant_head_check(
341 	struct xlog		*log,
342 	struct xlog_grant_head	*head,
343 	struct xlog_ticket	*tic,
344 	int			*need_bytes)
345 {
346 	int			free_bytes;
347 	int			error = 0;
348 
349 	ASSERT(!xlog_in_recovery(log));
350 
351 	/*
352 	 * If there are other waiters on the queue then give them a chance at
353 	 * logspace before us.  Wake up the first waiters, if we do not wake
354 	 * up all the waiters then go to sleep waiting for more free space,
355 	 * otherwise try to get some space for this transaction.
356 	 */
357 	*need_bytes = xlog_ticket_reservation(log, head, tic);
358 	free_bytes = xlog_space_left(log, &head->grant);
359 	if (!list_empty_careful(&head->waiters)) {
360 		spin_lock(&head->lock);
361 		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
362 		    free_bytes < *need_bytes) {
363 			error = xlog_grant_head_wait(log, head, tic,
364 						     *need_bytes);
365 		}
366 		spin_unlock(&head->lock);
367 	} else if (free_bytes < *need_bytes) {
368 		spin_lock(&head->lock);
369 		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
370 		spin_unlock(&head->lock);
371 	}
372 
373 	return error;
374 }
375 
376 bool
xfs_log_writable(struct xfs_mount * mp)377 xfs_log_writable(
378 	struct xfs_mount	*mp)
379 {
380 	/*
381 	 * Do not write to the log on norecovery mounts, if the data or log
382 	 * devices are read-only, or if the filesystem is shutdown. Read-only
383 	 * mounts allow internal writes for log recovery and unmount purposes,
384 	 * so don't restrict that case.
385 	 */
386 	if (xfs_has_norecovery(mp))
387 		return false;
388 	if (xfs_readonly_buftarg(mp->m_ddev_targp))
389 		return false;
390 	if (xfs_readonly_buftarg(mp->m_log->l_targ))
391 		return false;
392 	if (xlog_is_shutdown(mp->m_log))
393 		return false;
394 	return true;
395 }
396 
397 /*
398  * Replenish the byte reservation required by moving the grant write head.
399  */
400 int
xfs_log_regrant(struct xfs_mount * mp,struct xlog_ticket * tic)401 xfs_log_regrant(
402 	struct xfs_mount	*mp,
403 	struct xlog_ticket	*tic)
404 {
405 	struct xlog		*log = mp->m_log;
406 	int			need_bytes;
407 	int			error = 0;
408 
409 	if (xlog_is_shutdown(log))
410 		return -EIO;
411 
412 	XFS_STATS_INC(mp, xs_try_logspace);
413 
414 	/*
415 	 * This is a new transaction on the ticket, so we need to change the
416 	 * transaction ID so that the next transaction has a different TID in
417 	 * the log. Just add one to the existing tid so that we can see chains
418 	 * of rolling transactions in the log easily.
419 	 */
420 	tic->t_tid++;
421 
422 	xlog_grant_push_ail(log, tic->t_unit_res);
423 
424 	tic->t_curr_res = tic->t_unit_res;
425 	if (tic->t_cnt > 0)
426 		return 0;
427 
428 	trace_xfs_log_regrant(log, tic);
429 
430 	error = xlog_grant_head_check(log, &log->l_write_head, tic,
431 				      &need_bytes);
432 	if (error)
433 		goto out_error;
434 
435 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
436 	trace_xfs_log_regrant_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  * Reserve log space and return a ticket corresponding to the reservation.
453  *
454  * Each reservation is going to reserve extra space for a log record header.
455  * When writes happen to the on-disk log, we don't subtract the length of the
456  * log record header from any reservation.  By wasting space in each
457  * reservation, we prevent over allocation problems.
458  */
459 int
xfs_log_reserve(struct xfs_mount * mp,int unit_bytes,int cnt,struct xlog_ticket ** ticp,bool permanent)460 xfs_log_reserve(
461 	struct xfs_mount	*mp,
462 	int			unit_bytes,
463 	int			cnt,
464 	struct xlog_ticket	**ticp,
465 	bool			permanent)
466 {
467 	struct xlog		*log = mp->m_log;
468 	struct xlog_ticket	*tic;
469 	int			need_bytes;
470 	int			error = 0;
471 
472 	if (xlog_is_shutdown(log))
473 		return -EIO;
474 
475 	XFS_STATS_INC(mp, xs_try_logspace);
476 
477 	ASSERT(*ticp == NULL);
478 	tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
479 	*ticp = tic;
480 
481 	xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
482 					    : tic->t_unit_res);
483 
484 	trace_xfs_log_reserve(log, tic);
485 
486 	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
487 				      &need_bytes);
488 	if (error)
489 		goto out_error;
490 
491 	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
492 	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
493 	trace_xfs_log_reserve_exit(log, tic);
494 	xlog_verify_grant_tail(log);
495 	return 0;
496 
497 out_error:
498 	/*
499 	 * If we are failing, make sure the ticket doesn't have any current
500 	 * reservations.  We don't want to add this back when the ticket/
501 	 * transaction gets cancelled.
502 	 */
503 	tic->t_curr_res = 0;
504 	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
505 	return error;
506 }
507 
508 /*
509  * Run all the pending iclog callbacks and wake log force waiters and iclog
510  * space waiters so they can process the newly set shutdown state. We really
511  * don't care what order we process callbacks here because the log is shut down
512  * and so state cannot change on disk anymore. However, we cannot wake waiters
513  * until the callbacks have been processed because we may be in unmount and
514  * we must ensure that all AIL operations the callbacks perform have completed
515  * before we tear down the AIL.
516  *
517  * We avoid processing actively referenced iclogs so that we don't run callbacks
518  * while the iclog owner might still be preparing the iclog for IO submssion.
519  * These will be caught by xlog_state_iclog_release() and call this function
520  * again to process any callbacks that may have been added to that iclog.
521  */
522 static void
xlog_state_shutdown_callbacks(struct xlog * log)523 xlog_state_shutdown_callbacks(
524 	struct xlog		*log)
525 {
526 	struct xlog_in_core	*iclog;
527 	LIST_HEAD(cb_list);
528 
529 	iclog = log->l_iclog;
530 	do {
531 		if (atomic_read(&iclog->ic_refcnt)) {
532 			/* Reference holder will re-run iclog callbacks. */
533 			continue;
534 		}
535 		list_splice_init(&iclog->ic_callbacks, &cb_list);
536 		spin_unlock(&log->l_icloglock);
537 
538 		xlog_cil_process_committed(&cb_list);
539 
540 		spin_lock(&log->l_icloglock);
541 		wake_up_all(&iclog->ic_write_wait);
542 		wake_up_all(&iclog->ic_force_wait);
543 	} while ((iclog = iclog->ic_next) != log->l_iclog);
544 
545 	wake_up_all(&log->l_flush_wait);
546 }
547 
548 /*
549  * Flush iclog to disk if this is the last reference to the given iclog and the
550  * it is in the WANT_SYNC state.
551  *
552  * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
553  * log tail is updated correctly. NEED_FUA indicates that the iclog will be
554  * written to stable storage, and implies that a commit record is contained
555  * within the iclog. We need to ensure that the log tail does not move beyond
556  * the tail that the first commit record in the iclog ordered against, otherwise
557  * correct recovery of that checkpoint becomes dependent on future operations
558  * performed on this iclog.
559  *
560  * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
561  * current tail into iclog. Once the iclog tail is set, future operations must
562  * not modify it, otherwise they potentially violate ordering constraints for
563  * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
564  * the iclog will get zeroed on activation of the iclog after sync, so we
565  * always capture the tail lsn on the iclog on the first NEED_FUA release
566  * regardless of the number of active reference counts on this iclog.
567  */
568 int
xlog_state_release_iclog(struct xlog * log,struct xlog_in_core * iclog,struct xlog_ticket * ticket)569 xlog_state_release_iclog(
570 	struct xlog		*log,
571 	struct xlog_in_core	*iclog,
572 	struct xlog_ticket	*ticket)
573 {
574 	xfs_lsn_t		tail_lsn;
575 	bool			last_ref;
576 
577 	lockdep_assert_held(&log->l_icloglock);
578 
579 	trace_xlog_iclog_release(iclog, _RET_IP_);
580 	/*
581 	 * Grabbing the current log tail needs to be atomic w.r.t. the writing
582 	 * of the tail LSN into the iclog so we guarantee that the log tail does
583 	 * not move between the first time we know that the iclog needs to be
584 	 * made stable and when we eventually submit it.
585 	 */
586 	if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
587 	     (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
588 	    !iclog->ic_header.h_tail_lsn) {
589 		tail_lsn = xlog_assign_tail_lsn(log->l_mp);
590 		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
591 	}
592 
593 	last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
594 
595 	if (xlog_is_shutdown(log)) {
596 		/*
597 		 * If there are no more references to this iclog, process the
598 		 * pending iclog callbacks that were waiting on the release of
599 		 * this iclog.
600 		 */
601 		if (last_ref)
602 			xlog_state_shutdown_callbacks(log);
603 		return -EIO;
604 	}
605 
606 	if (!last_ref)
607 		return 0;
608 
609 	if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
610 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
611 		return 0;
612 	}
613 
614 	iclog->ic_state = XLOG_STATE_SYNCING;
615 	xlog_verify_tail_lsn(log, iclog);
616 	trace_xlog_iclog_syncing(iclog, _RET_IP_);
617 
618 	spin_unlock(&log->l_icloglock);
619 	xlog_sync(log, iclog, ticket);
620 	spin_lock(&log->l_icloglock);
621 	return 0;
622 }
623 
624 /*
625  * Mount a log filesystem
626  *
627  * mp		- ubiquitous xfs mount point structure
628  * log_target	- buftarg of on-disk log device
629  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
630  * num_bblocks	- Number of BBSIZE blocks in on-disk log
631  *
632  * Return error or zero.
633  */
634 int
xfs_log_mount(xfs_mount_t * mp,xfs_buftarg_t * log_target,xfs_daddr_t blk_offset,int num_bblks)635 xfs_log_mount(
636 	xfs_mount_t	*mp,
637 	xfs_buftarg_t	*log_target,
638 	xfs_daddr_t	blk_offset,
639 	int		num_bblks)
640 {
641 	struct xlog	*log;
642 	bool		fatal = xfs_has_crc(mp);
643 	int		error = 0;
644 	int		min_logfsbs;
645 
646 	if (!xfs_has_norecovery(mp)) {
647 		xfs_notice(mp, "Mounting V%d Filesystem",
648 			   XFS_SB_VERSION_NUM(&mp->m_sb));
649 	} else {
650 		xfs_notice(mp,
651 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
652 			   XFS_SB_VERSION_NUM(&mp->m_sb));
653 		ASSERT(xfs_is_readonly(mp));
654 	}
655 
656 	log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
657 	if (IS_ERR(log)) {
658 		error = PTR_ERR(log);
659 		goto out;
660 	}
661 	mp->m_log = log;
662 
663 	/*
664 	 * Validate the given log space and drop a critical message via syslog
665 	 * if the log size is too small that would lead to some unexpected
666 	 * situations in transaction log space reservation stage.
667 	 *
668 	 * Note: we can't just reject the mount if the validation fails.  This
669 	 * would mean that people would have to downgrade their kernel just to
670 	 * remedy the situation as there is no way to grow the log (short of
671 	 * black magic surgery with xfs_db).
672 	 *
673 	 * We can, however, reject mounts for CRC format filesystems, as the
674 	 * mkfs binary being used to make the filesystem should never create a
675 	 * filesystem with a log that is too small.
676 	 */
677 	min_logfsbs = xfs_log_calc_minimum_size(mp);
678 
679 	if (mp->m_sb.sb_logblocks < min_logfsbs) {
680 		xfs_warn(mp,
681 		"Log size %d blocks too small, minimum size is %d blocks",
682 			 mp->m_sb.sb_logblocks, min_logfsbs);
683 		error = -EINVAL;
684 	} else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
685 		xfs_warn(mp,
686 		"Log size %d blocks too large, maximum size is %lld blocks",
687 			 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
688 		error = -EINVAL;
689 	} else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
690 		xfs_warn(mp,
691 		"log size %lld bytes too large, maximum size is %lld bytes",
692 			 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
693 			 XFS_MAX_LOG_BYTES);
694 		error = -EINVAL;
695 	} else if (mp->m_sb.sb_logsunit > 1 &&
696 		   mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
697 		xfs_warn(mp,
698 		"log stripe unit %u bytes must be a multiple of block size",
699 			 mp->m_sb.sb_logsunit);
700 		error = -EINVAL;
701 		fatal = true;
702 	}
703 	if (error) {
704 		/*
705 		 * Log check errors are always fatal on v5; or whenever bad
706 		 * metadata leads to a crash.
707 		 */
708 		if (fatal) {
709 			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
710 			ASSERT(0);
711 			goto out_free_log;
712 		}
713 		xfs_crit(mp, "Log size out of supported range.");
714 		xfs_crit(mp,
715 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
716 	}
717 
718 	/*
719 	 * Initialize the AIL now we have a log.
720 	 */
721 	error = xfs_trans_ail_init(mp);
722 	if (error) {
723 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
724 		goto out_free_log;
725 	}
726 	log->l_ailp = mp->m_ail;
727 
728 	/*
729 	 * skip log recovery on a norecovery mount.  pretend it all
730 	 * just worked.
731 	 */
732 	if (!xfs_has_norecovery(mp)) {
733 		/*
734 		 * log recovery ignores readonly state and so we need to clear
735 		 * mount-based read only state so it can write to disk.
736 		 */
737 		bool	readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
738 						&mp->m_opstate);
739 		error = xlog_recover(log);
740 		if (readonly)
741 			set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
742 		if (error) {
743 			xfs_warn(mp, "log mount/recovery failed: error %d",
744 				error);
745 			xlog_recover_cancel(log);
746 			goto out_destroy_ail;
747 		}
748 	}
749 
750 	error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
751 			       "log");
752 	if (error)
753 		goto out_destroy_ail;
754 
755 	/* Normal transactions can now occur */
756 	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
757 
758 	/*
759 	 * Now the log has been fully initialised and we know were our
760 	 * space grant counters are, we can initialise the permanent ticket
761 	 * needed for delayed logging to work.
762 	 */
763 	xlog_cil_init_post_recovery(log);
764 
765 	return 0;
766 
767 out_destroy_ail:
768 	xfs_trans_ail_destroy(mp);
769 out_free_log:
770 	xlog_dealloc_log(log);
771 out:
772 	return error;
773 }
774 
775 /*
776  * Finish the recovery of the file system.  This is separate from the
777  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
778  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
779  * here.
780  *
781  * If we finish recovery successfully, start the background log work. If we are
782  * not doing recovery, then we have a RO filesystem and we don't need to start
783  * it.
784  */
785 int
xfs_log_mount_finish(struct xfs_mount * mp)786 xfs_log_mount_finish(
787 	struct xfs_mount	*mp)
788 {
789 	struct xlog		*log = mp->m_log;
790 	bool			readonly;
791 	int			error = 0;
792 
793 	if (xfs_has_norecovery(mp)) {
794 		ASSERT(xfs_is_readonly(mp));
795 		return 0;
796 	}
797 
798 	/*
799 	 * log recovery ignores readonly state and so we need to clear
800 	 * mount-based read only state so it can write to disk.
801 	 */
802 	readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
803 
804 	/*
805 	 * During the second phase of log recovery, we need iget and
806 	 * iput to behave like they do for an active filesystem.
807 	 * xfs_fs_drop_inode needs to be able to prevent the deletion
808 	 * of inodes before we're done replaying log items on those
809 	 * inodes.  Turn it off immediately after recovery finishes
810 	 * so that we don't leak the quota inodes if subsequent mount
811 	 * activities fail.
812 	 *
813 	 * We let all inodes involved in redo item processing end up on
814 	 * the LRU instead of being evicted immediately so that if we do
815 	 * something to an unlinked inode, the irele won't cause
816 	 * premature truncation and freeing of the inode, which results
817 	 * in log recovery failure.  We have to evict the unreferenced
818 	 * lru inodes after clearing SB_ACTIVE because we don't
819 	 * otherwise clean up the lru if there's a subsequent failure in
820 	 * xfs_mountfs, which leads to us leaking the inodes if nothing
821 	 * else (e.g. quotacheck) references the inodes before the
822 	 * mount failure occurs.
823 	 */
824 	mp->m_super->s_flags |= SB_ACTIVE;
825 	xfs_log_work_queue(mp);
826 	if (xlog_recovery_needed(log))
827 		error = xlog_recover_finish(log);
828 	mp->m_super->s_flags &= ~SB_ACTIVE;
829 	evict_inodes(mp->m_super);
830 
831 	/*
832 	 * Drain the buffer LRU after log recovery. This is required for v4
833 	 * filesystems to avoid leaving around buffers with NULL verifier ops,
834 	 * but we do it unconditionally to make sure we're always in a clean
835 	 * cache state after mount.
836 	 *
837 	 * Don't push in the error case because the AIL may have pending intents
838 	 * that aren't removed until recovery is cancelled.
839 	 */
840 	if (xlog_recovery_needed(log)) {
841 		if (!error) {
842 			xfs_log_force(mp, XFS_LOG_SYNC);
843 			xfs_ail_push_all_sync(mp->m_ail);
844 		}
845 		xfs_notice(mp, "Ending recovery (logdev: %s)",
846 				mp->m_logname ? mp->m_logname : "internal");
847 	} else {
848 		xfs_info(mp, "Ending clean mount");
849 	}
850 	xfs_buftarg_drain(mp->m_ddev_targp);
851 
852 	clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
853 	if (readonly)
854 		set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
855 
856 	/* Make sure the log is dead if we're returning failure. */
857 	ASSERT(!error || xlog_is_shutdown(log));
858 
859 	return error;
860 }
861 
862 /*
863  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
864  * the log.
865  */
866 void
xfs_log_mount_cancel(struct xfs_mount * mp)867 xfs_log_mount_cancel(
868 	struct xfs_mount	*mp)
869 {
870 	xlog_recover_cancel(mp->m_log);
871 	xfs_log_unmount(mp);
872 }
873 
874 /*
875  * Flush out the iclog to disk ensuring that device caches are flushed and
876  * the iclog hits stable storage before any completion waiters are woken.
877  */
878 static inline int
xlog_force_iclog(struct xlog_in_core * iclog)879 xlog_force_iclog(
880 	struct xlog_in_core	*iclog)
881 {
882 	atomic_inc(&iclog->ic_refcnt);
883 	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
884 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
885 		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
886 	return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
887 }
888 
889 /*
890  * Wait for the iclog and all prior iclogs to be written disk as required by the
891  * log force state machine. Waiting on ic_force_wait ensures iclog completions
892  * have been ordered and callbacks run before we are woken here, hence
893  * guaranteeing that all the iclogs up to this one are on stable storage.
894  */
895 int
xlog_wait_on_iclog(struct xlog_in_core * iclog)896 xlog_wait_on_iclog(
897 	struct xlog_in_core	*iclog)
898 		__releases(iclog->ic_log->l_icloglock)
899 {
900 	struct xlog		*log = iclog->ic_log;
901 
902 	trace_xlog_iclog_wait_on(iclog, _RET_IP_);
903 	if (!xlog_is_shutdown(log) &&
904 	    iclog->ic_state != XLOG_STATE_ACTIVE &&
905 	    iclog->ic_state != XLOG_STATE_DIRTY) {
906 		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
907 		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
908 	} else {
909 		spin_unlock(&log->l_icloglock);
910 	}
911 
912 	if (xlog_is_shutdown(log))
913 		return -EIO;
914 	return 0;
915 }
916 
917 /*
918  * Write out an unmount record using the ticket provided. We have to account for
919  * the data space used in the unmount ticket as this write is not done from a
920  * transaction context that has already done the accounting for us.
921  */
922 static int
xlog_write_unmount_record(struct xlog * log,struct xlog_ticket * ticket)923 xlog_write_unmount_record(
924 	struct xlog		*log,
925 	struct xlog_ticket	*ticket)
926 {
927 	struct  {
928 		struct xlog_op_header ophdr;
929 		struct xfs_unmount_log_format ulf;
930 	} unmount_rec = {
931 		.ophdr = {
932 			.oh_clientid = XFS_LOG,
933 			.oh_tid = cpu_to_be32(ticket->t_tid),
934 			.oh_flags = XLOG_UNMOUNT_TRANS,
935 		},
936 		.ulf = {
937 			.magic = XLOG_UNMOUNT_TYPE,
938 		},
939 	};
940 	struct xfs_log_iovec reg = {
941 		.i_addr = &unmount_rec,
942 		.i_len = sizeof(unmount_rec),
943 		.i_type = XLOG_REG_TYPE_UNMOUNT,
944 	};
945 	struct xfs_log_vec vec = {
946 		.lv_niovecs = 1,
947 		.lv_iovecp = &reg,
948 	};
949 	LIST_HEAD(lv_chain);
950 	list_add(&vec.lv_list, &lv_chain);
951 
952 	BUILD_BUG_ON((sizeof(struct xlog_op_header) +
953 		      sizeof(struct xfs_unmount_log_format)) !=
954 							sizeof(unmount_rec));
955 
956 	/* account for space used by record data */
957 	ticket->t_curr_res -= sizeof(unmount_rec);
958 
959 	return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
960 }
961 
962 /*
963  * Mark the filesystem clean by writing an unmount record to the head of the
964  * log.
965  */
966 static void
xlog_unmount_write(struct xlog * log)967 xlog_unmount_write(
968 	struct xlog		*log)
969 {
970 	struct xfs_mount	*mp = log->l_mp;
971 	struct xlog_in_core	*iclog;
972 	struct xlog_ticket	*tic = NULL;
973 	int			error;
974 
975 	error = xfs_log_reserve(mp, 600, 1, &tic, 0);
976 	if (error)
977 		goto out_err;
978 
979 	error = xlog_write_unmount_record(log, tic);
980 	/*
981 	 * At this point, we're umounting anyway, so there's no point in
982 	 * transitioning log state to shutdown. Just continue...
983 	 */
984 out_err:
985 	if (error)
986 		xfs_alert(mp, "%s: unmount record failed", __func__);
987 
988 	spin_lock(&log->l_icloglock);
989 	iclog = log->l_iclog;
990 	error = xlog_force_iclog(iclog);
991 	xlog_wait_on_iclog(iclog);
992 
993 	if (tic) {
994 		trace_xfs_log_umount_write(log, tic);
995 		xfs_log_ticket_ungrant(log, tic);
996 	}
997 }
998 
999 static void
xfs_log_unmount_verify_iclog(struct xlog * log)1000 xfs_log_unmount_verify_iclog(
1001 	struct xlog		*log)
1002 {
1003 	struct xlog_in_core	*iclog = log->l_iclog;
1004 
1005 	do {
1006 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
1007 		ASSERT(iclog->ic_offset == 0);
1008 	} while ((iclog = iclog->ic_next) != log->l_iclog);
1009 }
1010 
1011 /*
1012  * Unmount record used to have a string "Unmount filesystem--" in the
1013  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
1014  * We just write the magic number now since that particular field isn't
1015  * currently architecture converted and "Unmount" is a bit foo.
1016  * As far as I know, there weren't any dependencies on the old behaviour.
1017  */
1018 static void
xfs_log_unmount_write(struct xfs_mount * mp)1019 xfs_log_unmount_write(
1020 	struct xfs_mount	*mp)
1021 {
1022 	struct xlog		*log = mp->m_log;
1023 
1024 	if (!xfs_log_writable(mp))
1025 		return;
1026 
1027 	xfs_log_force(mp, XFS_LOG_SYNC);
1028 
1029 	if (xlog_is_shutdown(log))
1030 		return;
1031 
1032 	/*
1033 	 * If we think the summary counters are bad, avoid writing the unmount
1034 	 * record to force log recovery at next mount, after which the summary
1035 	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
1036 	 * more details.
1037 	 */
1038 	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1039 			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1040 		xfs_alert(mp, "%s: will fix summary counters at next mount",
1041 				__func__);
1042 		return;
1043 	}
1044 
1045 	xfs_log_unmount_verify_iclog(log);
1046 	xlog_unmount_write(log);
1047 }
1048 
1049 /*
1050  * Empty the log for unmount/freeze.
1051  *
1052  * To do this, we first need to shut down the background log work so it is not
1053  * trying to cover the log as we clean up. We then need to unpin all objects in
1054  * the log so we can then flush them out. Once they have completed their IO and
1055  * run the callbacks removing themselves from the AIL, we can cover the log.
1056  */
1057 int
xfs_log_quiesce(struct xfs_mount * mp)1058 xfs_log_quiesce(
1059 	struct xfs_mount	*mp)
1060 {
1061 	/*
1062 	 * Clear log incompat features since we're quiescing the log.  Report
1063 	 * failures, though it's not fatal to have a higher log feature
1064 	 * protection level than the log contents actually require.
1065 	 */
1066 	if (xfs_clear_incompat_log_features(mp)) {
1067 		int error;
1068 
1069 		error = xfs_sync_sb(mp, false);
1070 		if (error)
1071 			xfs_warn(mp,
1072 	"Failed to clear log incompat features on quiesce");
1073 	}
1074 
1075 	cancel_delayed_work_sync(&mp->m_log->l_work);
1076 	xfs_log_force(mp, XFS_LOG_SYNC);
1077 
1078 	/*
1079 	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1080 	 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1081 	 * xfs_buf_iowait() cannot be used because it was pushed with the
1082 	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1083 	 * the IO to complete.
1084 	 */
1085 	xfs_ail_push_all_sync(mp->m_ail);
1086 	xfs_buftarg_wait(mp->m_ddev_targp);
1087 	xfs_buf_lock(mp->m_sb_bp);
1088 	xfs_buf_unlock(mp->m_sb_bp);
1089 
1090 	return xfs_log_cover(mp);
1091 }
1092 
1093 void
xfs_log_clean(struct xfs_mount * mp)1094 xfs_log_clean(
1095 	struct xfs_mount	*mp)
1096 {
1097 	xfs_log_quiesce(mp);
1098 	xfs_log_unmount_write(mp);
1099 }
1100 
1101 /*
1102  * Shut down and release the AIL and Log.
1103  *
1104  * During unmount, we need to ensure we flush all the dirty metadata objects
1105  * from the AIL so that the log is empty before we write the unmount record to
1106  * the log. Once this is done, we can tear down the AIL and the log.
1107  */
1108 void
xfs_log_unmount(struct xfs_mount * mp)1109 xfs_log_unmount(
1110 	struct xfs_mount	*mp)
1111 {
1112 	xfs_log_clean(mp);
1113 
1114 	xfs_buftarg_drain(mp->m_ddev_targp);
1115 
1116 	xfs_trans_ail_destroy(mp);
1117 
1118 	xfs_sysfs_del(&mp->m_log->l_kobj);
1119 
1120 	xlog_dealloc_log(mp->m_log);
1121 }
1122 
1123 void
xfs_log_item_init(struct xfs_mount * mp,struct xfs_log_item * item,int type,const struct xfs_item_ops * ops)1124 xfs_log_item_init(
1125 	struct xfs_mount	*mp,
1126 	struct xfs_log_item	*item,
1127 	int			type,
1128 	const struct xfs_item_ops *ops)
1129 {
1130 	item->li_log = mp->m_log;
1131 	item->li_ailp = mp->m_ail;
1132 	item->li_type = type;
1133 	item->li_ops = ops;
1134 	item->li_lv = NULL;
1135 
1136 	INIT_LIST_HEAD(&item->li_ail);
1137 	INIT_LIST_HEAD(&item->li_cil);
1138 	INIT_LIST_HEAD(&item->li_bio_list);
1139 	INIT_LIST_HEAD(&item->li_trans);
1140 }
1141 
1142 /*
1143  * Wake up processes waiting for log space after we have moved the log tail.
1144  */
1145 void
xfs_log_space_wake(struct xfs_mount * mp)1146 xfs_log_space_wake(
1147 	struct xfs_mount	*mp)
1148 {
1149 	struct xlog		*log = mp->m_log;
1150 	int			free_bytes;
1151 
1152 	if (xlog_is_shutdown(log))
1153 		return;
1154 
1155 	if (!list_empty_careful(&log->l_write_head.waiters)) {
1156 		ASSERT(!xlog_in_recovery(log));
1157 
1158 		spin_lock(&log->l_write_head.lock);
1159 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1160 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1161 		spin_unlock(&log->l_write_head.lock);
1162 	}
1163 
1164 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1165 		ASSERT(!xlog_in_recovery(log));
1166 
1167 		spin_lock(&log->l_reserve_head.lock);
1168 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1169 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1170 		spin_unlock(&log->l_reserve_head.lock);
1171 	}
1172 }
1173 
1174 /*
1175  * Determine if we have a transaction that has gone to disk that needs to be
1176  * covered. To begin the transition to the idle state firstly the log needs to
1177  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1178  * we start attempting to cover the log.
1179  *
1180  * Only if we are then in a state where covering is needed, the caller is
1181  * informed that dummy transactions are required to move the log into the idle
1182  * state.
1183  *
1184  * If there are any items in the AIl or CIL, then we do not want to attempt to
1185  * cover the log as we may be in a situation where there isn't log space
1186  * available to run a dummy transaction and this can lead to deadlocks when the
1187  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1188  * there's no point in running a dummy transaction at this point because we
1189  * can't start trying to idle the log until both the CIL and AIL are empty.
1190  */
1191 static bool
xfs_log_need_covered(struct xfs_mount * mp)1192 xfs_log_need_covered(
1193 	struct xfs_mount	*mp)
1194 {
1195 	struct xlog		*log = mp->m_log;
1196 	bool			needed = false;
1197 
1198 	if (!xlog_cil_empty(log))
1199 		return false;
1200 
1201 	spin_lock(&log->l_icloglock);
1202 	switch (log->l_covered_state) {
1203 	case XLOG_STATE_COVER_DONE:
1204 	case XLOG_STATE_COVER_DONE2:
1205 	case XLOG_STATE_COVER_IDLE:
1206 		break;
1207 	case XLOG_STATE_COVER_NEED:
1208 	case XLOG_STATE_COVER_NEED2:
1209 		if (xfs_ail_min_lsn(log->l_ailp))
1210 			break;
1211 		if (!xlog_iclogs_empty(log))
1212 			break;
1213 
1214 		needed = true;
1215 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1216 			log->l_covered_state = XLOG_STATE_COVER_DONE;
1217 		else
1218 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1219 		break;
1220 	default:
1221 		needed = true;
1222 		break;
1223 	}
1224 	spin_unlock(&log->l_icloglock);
1225 	return needed;
1226 }
1227 
1228 /*
1229  * Explicitly cover the log. This is similar to background log covering but
1230  * intended for usage in quiesce codepaths. The caller is responsible to ensure
1231  * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1232  * must all be empty.
1233  */
1234 static int
xfs_log_cover(struct xfs_mount * mp)1235 xfs_log_cover(
1236 	struct xfs_mount	*mp)
1237 {
1238 	int			error = 0;
1239 	bool			need_covered;
1240 
1241 	ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1242 	        !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1243 		xlog_is_shutdown(mp->m_log));
1244 
1245 	if (!xfs_log_writable(mp))
1246 		return 0;
1247 
1248 	/*
1249 	 * xfs_log_need_covered() is not idempotent because it progresses the
1250 	 * state machine if the log requires covering. Therefore, we must call
1251 	 * this function once and use the result until we've issued an sb sync.
1252 	 * Do so first to make that abundantly clear.
1253 	 *
1254 	 * Fall into the covering sequence if the log needs covering or the
1255 	 * mount has lazy superblock accounting to sync to disk. The sb sync
1256 	 * used for covering accumulates the in-core counters, so covering
1257 	 * handles this for us.
1258 	 */
1259 	need_covered = xfs_log_need_covered(mp);
1260 	if (!need_covered && !xfs_has_lazysbcount(mp))
1261 		return 0;
1262 
1263 	/*
1264 	 * To cover the log, commit the superblock twice (at most) in
1265 	 * independent checkpoints. The first serves as a reference for the
1266 	 * tail pointer. The sync transaction and AIL push empties the AIL and
1267 	 * updates the in-core tail to the LSN of the first checkpoint. The
1268 	 * second commit updates the on-disk tail with the in-core LSN,
1269 	 * covering the log. Push the AIL one more time to leave it empty, as
1270 	 * we found it.
1271 	 */
1272 	do {
1273 		error = xfs_sync_sb(mp, true);
1274 		if (error)
1275 			break;
1276 		xfs_ail_push_all_sync(mp->m_ail);
1277 	} while (xfs_log_need_covered(mp));
1278 
1279 	return error;
1280 }
1281 
1282 /*
1283  * We may be holding the log iclog lock upon entering this routine.
1284  */
1285 xfs_lsn_t
xlog_assign_tail_lsn_locked(struct xfs_mount * mp)1286 xlog_assign_tail_lsn_locked(
1287 	struct xfs_mount	*mp)
1288 {
1289 	struct xlog		*log = mp->m_log;
1290 	struct xfs_log_item	*lip;
1291 	xfs_lsn_t		tail_lsn;
1292 
1293 	assert_spin_locked(&mp->m_ail->ail_lock);
1294 
1295 	/*
1296 	 * To make sure we always have a valid LSN for the log tail we keep
1297 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1298 	 * and use that when the AIL was empty.
1299 	 */
1300 	lip = xfs_ail_min(mp->m_ail);
1301 	if (lip)
1302 		tail_lsn = lip->li_lsn;
1303 	else
1304 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1305 	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1306 	atomic64_set(&log->l_tail_lsn, tail_lsn);
1307 	return tail_lsn;
1308 }
1309 
1310 xfs_lsn_t
xlog_assign_tail_lsn(struct xfs_mount * mp)1311 xlog_assign_tail_lsn(
1312 	struct xfs_mount	*mp)
1313 {
1314 	xfs_lsn_t		tail_lsn;
1315 
1316 	spin_lock(&mp->m_ail->ail_lock);
1317 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1318 	spin_unlock(&mp->m_ail->ail_lock);
1319 
1320 	return tail_lsn;
1321 }
1322 
1323 /*
1324  * Return the space in the log between the tail and the head.  The head
1325  * is passed in the cycle/bytes formal parms.  In the special case where
1326  * the reserve head has wrapped passed the tail, this calculation is no
1327  * longer valid.  In this case, just return 0 which means there is no space
1328  * in the log.  This works for all places where this function is called
1329  * with the reserve head.  Of course, if the write head were to ever
1330  * wrap the tail, we should blow up.  Rather than catch this case here,
1331  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1332  *
1333  * If reservation head is behind the tail, we have a problem. Warn about it,
1334  * but then treat it as if the log is empty.
1335  *
1336  * If the log is shut down, the head and tail may be invalid or out of whack, so
1337  * shortcut invalidity asserts in this case so that we don't trigger them
1338  * falsely.
1339  */
1340 STATIC int
xlog_space_left(struct xlog * log,atomic64_t * head)1341 xlog_space_left(
1342 	struct xlog	*log,
1343 	atomic64_t	*head)
1344 {
1345 	int		tail_bytes;
1346 	int		tail_cycle;
1347 	int		head_cycle;
1348 	int		head_bytes;
1349 
1350 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1351 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1352 	tail_bytes = BBTOB(tail_bytes);
1353 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1354 		return log->l_logsize - (head_bytes - tail_bytes);
1355 	if (tail_cycle + 1 < head_cycle)
1356 		return 0;
1357 
1358 	/* Ignore potential inconsistency when shutdown. */
1359 	if (xlog_is_shutdown(log))
1360 		return log->l_logsize;
1361 
1362 	if (tail_cycle < head_cycle) {
1363 		ASSERT(tail_cycle == (head_cycle - 1));
1364 		return tail_bytes - head_bytes;
1365 	}
1366 
1367 	/*
1368 	 * The reservation head is behind the tail. In this case we just want to
1369 	 * return the size of the log as the amount of space left.
1370 	 */
1371 	xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1372 	xfs_alert(log->l_mp, "  tail_cycle = %d, tail_bytes = %d",
1373 		  tail_cycle, tail_bytes);
1374 	xfs_alert(log->l_mp, "  GH   cycle = %d, GH   bytes = %d",
1375 		  head_cycle, head_bytes);
1376 	ASSERT(0);
1377 	return log->l_logsize;
1378 }
1379 
1380 
1381 static void
xlog_ioend_work(struct work_struct * work)1382 xlog_ioend_work(
1383 	struct work_struct	*work)
1384 {
1385 	struct xlog_in_core     *iclog =
1386 		container_of(work, struct xlog_in_core, ic_end_io_work);
1387 	struct xlog		*log = iclog->ic_log;
1388 	int			error;
1389 
1390 	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1391 #ifdef DEBUG
1392 	/* treat writes with injected CRC errors as failed */
1393 	if (iclog->ic_fail_crc)
1394 		error = -EIO;
1395 #endif
1396 
1397 	/*
1398 	 * Race to shutdown the filesystem if we see an error.
1399 	 */
1400 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1401 		xfs_alert(log->l_mp, "log I/O error %d", error);
1402 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1403 	}
1404 
1405 	xlog_state_done_syncing(iclog);
1406 	bio_uninit(&iclog->ic_bio);
1407 
1408 	/*
1409 	 * Drop the lock to signal that we are done. Nothing references the
1410 	 * iclog after this, so an unmount waiting on this lock can now tear it
1411 	 * down safely. As such, it is unsafe to reference the iclog after the
1412 	 * unlock as we could race with it being freed.
1413 	 */
1414 	up(&iclog->ic_sema);
1415 }
1416 
1417 /*
1418  * Return size of each in-core log record buffer.
1419  *
1420  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1421  *
1422  * If the filesystem blocksize is too large, we may need to choose a
1423  * larger size since the directory code currently logs entire blocks.
1424  */
1425 STATIC void
xlog_get_iclog_buffer_size(struct xfs_mount * mp,struct xlog * log)1426 xlog_get_iclog_buffer_size(
1427 	struct xfs_mount	*mp,
1428 	struct xlog		*log)
1429 {
1430 	if (mp->m_logbufs <= 0)
1431 		mp->m_logbufs = XLOG_MAX_ICLOGS;
1432 	if (mp->m_logbsize <= 0)
1433 		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1434 
1435 	log->l_iclog_bufs = mp->m_logbufs;
1436 	log->l_iclog_size = mp->m_logbsize;
1437 
1438 	/*
1439 	 * # headers = size / 32k - one header holds cycles from 32k of data.
1440 	 */
1441 	log->l_iclog_heads =
1442 		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1443 	log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1444 }
1445 
1446 void
xfs_log_work_queue(struct xfs_mount * mp)1447 xfs_log_work_queue(
1448 	struct xfs_mount        *mp)
1449 {
1450 	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1451 				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1452 }
1453 
1454 /*
1455  * Clear the log incompat flags if we have the opportunity.
1456  *
1457  * This only happens if we're about to log the second dummy transaction as part
1458  * of covering the log and we can get the log incompat feature usage lock.
1459  */
1460 static inline void
xlog_clear_incompat(struct xlog * log)1461 xlog_clear_incompat(
1462 	struct xlog		*log)
1463 {
1464 	struct xfs_mount	*mp = log->l_mp;
1465 
1466 	if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1467 				XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1468 		return;
1469 
1470 	if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1471 		return;
1472 
1473 	if (!down_write_trylock(&log->l_incompat_users))
1474 		return;
1475 
1476 	xfs_clear_incompat_log_features(mp);
1477 	up_write(&log->l_incompat_users);
1478 }
1479 
1480 /*
1481  * Every sync period we need to unpin all items in the AIL and push them to
1482  * disk. If there is nothing dirty, then we might need to cover the log to
1483  * indicate that the filesystem is idle.
1484  */
1485 static void
xfs_log_worker(struct work_struct * work)1486 xfs_log_worker(
1487 	struct work_struct	*work)
1488 {
1489 	struct xlog		*log = container_of(to_delayed_work(work),
1490 						struct xlog, l_work);
1491 	struct xfs_mount	*mp = log->l_mp;
1492 
1493 	/* dgc: errors ignored - not fatal and nowhere to report them */
1494 	if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1495 		/*
1496 		 * Dump a transaction into the log that contains no real change.
1497 		 * This is needed to stamp the current tail LSN into the log
1498 		 * during the covering operation.
1499 		 *
1500 		 * We cannot use an inode here for this - that will push dirty
1501 		 * state back up into the VFS and then periodic inode flushing
1502 		 * will prevent log covering from making progress. Hence we
1503 		 * synchronously log the superblock instead to ensure the
1504 		 * superblock is immediately unpinned and can be written back.
1505 		 */
1506 		xlog_clear_incompat(log);
1507 		xfs_sync_sb(mp, true);
1508 	} else
1509 		xfs_log_force(mp, 0);
1510 
1511 	/* start pushing all the metadata that is currently dirty */
1512 	xfs_ail_push_all(mp->m_ail);
1513 
1514 	/* queue us up again */
1515 	xfs_log_work_queue(mp);
1516 }
1517 
1518 /*
1519  * This routine initializes some of the log structure for a given mount point.
1520  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1521  * some other stuff may be filled in too.
1522  */
1523 STATIC struct xlog *
xlog_alloc_log(struct xfs_mount * mp,struct xfs_buftarg * log_target,xfs_daddr_t blk_offset,int num_bblks)1524 xlog_alloc_log(
1525 	struct xfs_mount	*mp,
1526 	struct xfs_buftarg	*log_target,
1527 	xfs_daddr_t		blk_offset,
1528 	int			num_bblks)
1529 {
1530 	struct xlog		*log;
1531 	xlog_rec_header_t	*head;
1532 	xlog_in_core_t		**iclogp;
1533 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1534 	int			i;
1535 	int			error = -ENOMEM;
1536 	uint			log2_size = 0;
1537 
1538 	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1539 	if (!log) {
1540 		xfs_warn(mp, "Log allocation failed: No memory!");
1541 		goto out;
1542 	}
1543 
1544 	log->l_mp	   = mp;
1545 	log->l_targ	   = log_target;
1546 	log->l_logsize     = BBTOB(num_bblks);
1547 	log->l_logBBstart  = blk_offset;
1548 	log->l_logBBsize   = num_bblks;
1549 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1550 	set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1551 	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1552 
1553 	log->l_prev_block  = -1;
1554 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1555 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1556 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1557 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1558 
1559 	if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1560 		log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1561 	else
1562 		log->l_iclog_roundoff = BBSIZE;
1563 
1564 	xlog_grant_head_init(&log->l_reserve_head);
1565 	xlog_grant_head_init(&log->l_write_head);
1566 
1567 	error = -EFSCORRUPTED;
1568 	if (xfs_has_sector(mp)) {
1569 	        log2_size = mp->m_sb.sb_logsectlog;
1570 		if (log2_size < BBSHIFT) {
1571 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1572 				log2_size, BBSHIFT);
1573 			goto out_free_log;
1574 		}
1575 
1576 	        log2_size -= BBSHIFT;
1577 		if (log2_size > mp->m_sectbb_log) {
1578 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1579 				log2_size, mp->m_sectbb_log);
1580 			goto out_free_log;
1581 		}
1582 
1583 		/* for larger sector sizes, must have v2 or external log */
1584 		if (log2_size && log->l_logBBstart > 0 &&
1585 			    !xfs_has_logv2(mp)) {
1586 			xfs_warn(mp,
1587 		"log sector size (0x%x) invalid for configuration.",
1588 				log2_size);
1589 			goto out_free_log;
1590 		}
1591 	}
1592 	log->l_sectBBsize = 1 << log2_size;
1593 
1594 	init_rwsem(&log->l_incompat_users);
1595 
1596 	xlog_get_iclog_buffer_size(mp, log);
1597 
1598 	spin_lock_init(&log->l_icloglock);
1599 	init_waitqueue_head(&log->l_flush_wait);
1600 
1601 	iclogp = &log->l_iclog;
1602 	/*
1603 	 * The amount of memory to allocate for the iclog structure is
1604 	 * rather funky due to the way the structure is defined.  It is
1605 	 * done this way so that we can use different sizes for machines
1606 	 * with different amounts of memory.  See the definition of
1607 	 * xlog_in_core_t in xfs_log_priv.h for details.
1608 	 */
1609 	ASSERT(log->l_iclog_size >= 4096);
1610 	for (i = 0; i < log->l_iclog_bufs; i++) {
1611 		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1612 				sizeof(struct bio_vec);
1613 
1614 		iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1615 		if (!iclog)
1616 			goto out_free_iclog;
1617 
1618 		*iclogp = iclog;
1619 		iclog->ic_prev = prev_iclog;
1620 		prev_iclog = iclog;
1621 
1622 		iclog->ic_data = kvzalloc(log->l_iclog_size,
1623 				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1624 		if (!iclog->ic_data)
1625 			goto out_free_iclog;
1626 		head = &iclog->ic_header;
1627 		memset(head, 0, sizeof(xlog_rec_header_t));
1628 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1629 		head->h_version = cpu_to_be32(
1630 			xfs_has_logv2(log->l_mp) ? 2 : 1);
1631 		head->h_size = cpu_to_be32(log->l_iclog_size);
1632 		/* new fields */
1633 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1634 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1635 
1636 		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1637 		iclog->ic_state = XLOG_STATE_ACTIVE;
1638 		iclog->ic_log = log;
1639 		atomic_set(&iclog->ic_refcnt, 0);
1640 		INIT_LIST_HEAD(&iclog->ic_callbacks);
1641 		iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1642 
1643 		init_waitqueue_head(&iclog->ic_force_wait);
1644 		init_waitqueue_head(&iclog->ic_write_wait);
1645 		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1646 		sema_init(&iclog->ic_sema, 1);
1647 
1648 		iclogp = &iclog->ic_next;
1649 	}
1650 	*iclogp = log->l_iclog;			/* complete ring */
1651 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1652 
1653 	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1654 			XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1655 				    WQ_HIGHPRI),
1656 			0, mp->m_super->s_id);
1657 	if (!log->l_ioend_workqueue)
1658 		goto out_free_iclog;
1659 
1660 	error = xlog_cil_init(log);
1661 	if (error)
1662 		goto out_destroy_workqueue;
1663 	return log;
1664 
1665 out_destroy_workqueue:
1666 	destroy_workqueue(log->l_ioend_workqueue);
1667 out_free_iclog:
1668 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1669 		prev_iclog = iclog->ic_next;
1670 		kmem_free(iclog->ic_data);
1671 		kmem_free(iclog);
1672 		if (prev_iclog == log->l_iclog)
1673 			break;
1674 	}
1675 out_free_log:
1676 	kmem_free(log);
1677 out:
1678 	return ERR_PTR(error);
1679 }	/* xlog_alloc_log */
1680 
1681 /*
1682  * Compute the LSN that we'd need to push the log tail towards in order to have
1683  * (a) enough on-disk log space to log the number of bytes specified, (b) at
1684  * least 25% of the log space free, and (c) at least 256 blocks free.  If the
1685  * log free space already meets all three thresholds, this function returns
1686  * NULLCOMMITLSN.
1687  */
1688 xfs_lsn_t
xlog_grant_push_threshold(struct xlog * log,int need_bytes)1689 xlog_grant_push_threshold(
1690 	struct xlog	*log,
1691 	int		need_bytes)
1692 {
1693 	xfs_lsn_t	threshold_lsn = 0;
1694 	xfs_lsn_t	last_sync_lsn;
1695 	int		free_blocks;
1696 	int		free_bytes;
1697 	int		threshold_block;
1698 	int		threshold_cycle;
1699 	int		free_threshold;
1700 
1701 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1702 
1703 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1704 	free_blocks = BTOBBT(free_bytes);
1705 
1706 	/*
1707 	 * Set the threshold for the minimum number of free blocks in the
1708 	 * log to the maximum of what the caller needs, one quarter of the
1709 	 * log, and 256 blocks.
1710 	 */
1711 	free_threshold = BTOBB(need_bytes);
1712 	free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1713 	free_threshold = max(free_threshold, 256);
1714 	if (free_blocks >= free_threshold)
1715 		return NULLCOMMITLSN;
1716 
1717 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1718 						&threshold_block);
1719 	threshold_block += free_threshold;
1720 	if (threshold_block >= log->l_logBBsize) {
1721 		threshold_block -= log->l_logBBsize;
1722 		threshold_cycle += 1;
1723 	}
1724 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1725 					threshold_block);
1726 	/*
1727 	 * Don't pass in an lsn greater than the lsn of the last
1728 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1729 	 * so that it doesn't change between the compare and the set.
1730 	 */
1731 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1732 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1733 		threshold_lsn = last_sync_lsn;
1734 
1735 	return threshold_lsn;
1736 }
1737 
1738 /*
1739  * Push the tail of the log if we need to do so to maintain the free log space
1740  * thresholds set out by xlog_grant_push_threshold.  We may need to adopt a
1741  * policy which pushes on an lsn which is further along in the log once we
1742  * reach the high water mark.  In this manner, we would be creating a low water
1743  * mark.
1744  */
1745 STATIC void
xlog_grant_push_ail(struct xlog * log,int need_bytes)1746 xlog_grant_push_ail(
1747 	struct xlog	*log,
1748 	int		need_bytes)
1749 {
1750 	xfs_lsn_t	threshold_lsn;
1751 
1752 	threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1753 	if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1754 		return;
1755 
1756 	/*
1757 	 * Get the transaction layer to kick the dirty buffers out to
1758 	 * disk asynchronously. No point in trying to do this if
1759 	 * the filesystem is shutting down.
1760 	 */
1761 	xfs_ail_push(log->l_ailp, threshold_lsn);
1762 }
1763 
1764 /*
1765  * Stamp cycle number in every block
1766  */
1767 STATIC void
xlog_pack_data(struct xlog * log,struct xlog_in_core * iclog,int roundoff)1768 xlog_pack_data(
1769 	struct xlog		*log,
1770 	struct xlog_in_core	*iclog,
1771 	int			roundoff)
1772 {
1773 	int			i, j, k;
1774 	int			size = iclog->ic_offset + roundoff;
1775 	__be32			cycle_lsn;
1776 	char			*dp;
1777 
1778 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1779 
1780 	dp = iclog->ic_datap;
1781 	for (i = 0; i < BTOBB(size); i++) {
1782 		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1783 			break;
1784 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1785 		*(__be32 *)dp = cycle_lsn;
1786 		dp += BBSIZE;
1787 	}
1788 
1789 	if (xfs_has_logv2(log->l_mp)) {
1790 		xlog_in_core_2_t *xhdr = iclog->ic_data;
1791 
1792 		for ( ; i < BTOBB(size); i++) {
1793 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1794 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1795 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1796 			*(__be32 *)dp = cycle_lsn;
1797 			dp += BBSIZE;
1798 		}
1799 
1800 		for (i = 1; i < log->l_iclog_heads; i++)
1801 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1802 	}
1803 }
1804 
1805 /*
1806  * Calculate the checksum for a log buffer.
1807  *
1808  * This is a little more complicated than it should be because the various
1809  * headers and the actual data are non-contiguous.
1810  */
1811 __le32
xlog_cksum(struct xlog * log,struct xlog_rec_header * rhead,char * dp,int size)1812 xlog_cksum(
1813 	struct xlog		*log,
1814 	struct xlog_rec_header	*rhead,
1815 	char			*dp,
1816 	int			size)
1817 {
1818 	uint32_t		crc;
1819 
1820 	/* first generate the crc for the record header ... */
1821 	crc = xfs_start_cksum_update((char *)rhead,
1822 			      sizeof(struct xlog_rec_header),
1823 			      offsetof(struct xlog_rec_header, h_crc));
1824 
1825 	/* ... then for additional cycle data for v2 logs ... */
1826 	if (xfs_has_logv2(log->l_mp)) {
1827 		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1828 		int		i;
1829 		int		xheads;
1830 
1831 		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1832 
1833 		for (i = 1; i < xheads; i++) {
1834 			crc = crc32c(crc, &xhdr[i].hic_xheader,
1835 				     sizeof(struct xlog_rec_ext_header));
1836 		}
1837 	}
1838 
1839 	/* ... and finally for the payload */
1840 	crc = crc32c(crc, dp, size);
1841 
1842 	return xfs_end_cksum(crc);
1843 }
1844 
1845 static void
xlog_bio_end_io(struct bio * bio)1846 xlog_bio_end_io(
1847 	struct bio		*bio)
1848 {
1849 	struct xlog_in_core	*iclog = bio->bi_private;
1850 
1851 	queue_work(iclog->ic_log->l_ioend_workqueue,
1852 		   &iclog->ic_end_io_work);
1853 }
1854 
1855 static int
xlog_map_iclog_data(struct bio * bio,void * data,size_t count)1856 xlog_map_iclog_data(
1857 	struct bio		*bio,
1858 	void			*data,
1859 	size_t			count)
1860 {
1861 	do {
1862 		struct page	*page = kmem_to_page(data);
1863 		unsigned int	off = offset_in_page(data);
1864 		size_t		len = min_t(size_t, count, PAGE_SIZE - off);
1865 
1866 		if (bio_add_page(bio, page, len, off) != len)
1867 			return -EIO;
1868 
1869 		data += len;
1870 		count -= len;
1871 	} while (count);
1872 
1873 	return 0;
1874 }
1875 
1876 STATIC void
xlog_write_iclog(struct xlog * log,struct xlog_in_core * iclog,uint64_t bno,unsigned int count)1877 xlog_write_iclog(
1878 	struct xlog		*log,
1879 	struct xlog_in_core	*iclog,
1880 	uint64_t		bno,
1881 	unsigned int		count)
1882 {
1883 	ASSERT(bno < log->l_logBBsize);
1884 	trace_xlog_iclog_write(iclog, _RET_IP_);
1885 
1886 	/*
1887 	 * We lock the iclogbufs here so that we can serialise against I/O
1888 	 * completion during unmount.  We might be processing a shutdown
1889 	 * triggered during unmount, and that can occur asynchronously to the
1890 	 * unmount thread, and hence we need to ensure that completes before
1891 	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1892 	 * across the log IO to archieve that.
1893 	 */
1894 	down(&iclog->ic_sema);
1895 	if (xlog_is_shutdown(log)) {
1896 		/*
1897 		 * It would seem logical to return EIO here, but we rely on
1898 		 * the log state machine to propagate I/O errors instead of
1899 		 * doing it here.  We kick of the state machine and unlock
1900 		 * the buffer manually, the code needs to be kept in sync
1901 		 * with the I/O completion path.
1902 		 */
1903 		xlog_state_done_syncing(iclog);
1904 		up(&iclog->ic_sema);
1905 		return;
1906 	}
1907 
1908 	/*
1909 	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1910 	 * IOs coming immediately after this one. This prevents the block layer
1911 	 * writeback throttle from throttling log writes behind background
1912 	 * metadata writeback and causing priority inversions.
1913 	 */
1914 	bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1915 		 howmany(count, PAGE_SIZE),
1916 		 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1917 	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1918 	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1919 	iclog->ic_bio.bi_private = iclog;
1920 
1921 	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1922 		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1923 		/*
1924 		 * For external log devices, we also need to flush the data
1925 		 * device cache first to ensure all metadata writeback covered
1926 		 * by the LSN in this iclog is on stable storage. This is slow,
1927 		 * but it *must* complete before we issue the external log IO.
1928 		 *
1929 		 * If the flush fails, we cannot conclude that past metadata
1930 		 * writeback from the log succeeded.  Repeating the flush is
1931 		 * not possible, hence we must shut down with log IO error to
1932 		 * avoid shutdown re-entering this path and erroring out again.
1933 		 */
1934 		if (log->l_targ != log->l_mp->m_ddev_targp &&
1935 		    blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev)) {
1936 			xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1937 			return;
1938 		}
1939 	}
1940 	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1941 		iclog->ic_bio.bi_opf |= REQ_FUA;
1942 
1943 	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1944 
1945 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1946 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1947 		return;
1948 	}
1949 	if (is_vmalloc_addr(iclog->ic_data))
1950 		flush_kernel_vmap_range(iclog->ic_data, count);
1951 
1952 	/*
1953 	 * If this log buffer would straddle the end of the log we will have
1954 	 * to split it up into two bios, so that we can continue at the start.
1955 	 */
1956 	if (bno + BTOBB(count) > log->l_logBBsize) {
1957 		struct bio *split;
1958 
1959 		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1960 				  GFP_NOIO, &fs_bio_set);
1961 		bio_chain(split, &iclog->ic_bio);
1962 		submit_bio(split);
1963 
1964 		/* restart at logical offset zero for the remainder */
1965 		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1966 	}
1967 
1968 	submit_bio(&iclog->ic_bio);
1969 }
1970 
1971 /*
1972  * We need to bump cycle number for the part of the iclog that is
1973  * written to the start of the log. Watch out for the header magic
1974  * number case, though.
1975  */
1976 static void
xlog_split_iclog(struct xlog * log,void * data,uint64_t bno,unsigned int count)1977 xlog_split_iclog(
1978 	struct xlog		*log,
1979 	void			*data,
1980 	uint64_t		bno,
1981 	unsigned int		count)
1982 {
1983 	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1984 	unsigned int		i;
1985 
1986 	for (i = split_offset; i < count; i += BBSIZE) {
1987 		uint32_t cycle = get_unaligned_be32(data + i);
1988 
1989 		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1990 			cycle++;
1991 		put_unaligned_be32(cycle, data + i);
1992 	}
1993 }
1994 
1995 static int
xlog_calc_iclog_size(struct xlog * log,struct xlog_in_core * iclog,uint32_t * roundoff)1996 xlog_calc_iclog_size(
1997 	struct xlog		*log,
1998 	struct xlog_in_core	*iclog,
1999 	uint32_t		*roundoff)
2000 {
2001 	uint32_t		count_init, count;
2002 
2003 	/* Add for LR header */
2004 	count_init = log->l_iclog_hsize + iclog->ic_offset;
2005 	count = roundup(count_init, log->l_iclog_roundoff);
2006 
2007 	*roundoff = count - count_init;
2008 
2009 	ASSERT(count >= count_init);
2010 	ASSERT(*roundoff < log->l_iclog_roundoff);
2011 	return count;
2012 }
2013 
2014 /*
2015  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2016  * fashion.  Previously, we should have moved the current iclog
2017  * ptr in the log to point to the next available iclog.  This allows further
2018  * write to continue while this code syncs out an iclog ready to go.
2019  * Before an in-core log can be written out, the data section must be scanned
2020  * to save away the 1st word of each BBSIZE block into the header.  We replace
2021  * it with the current cycle count.  Each BBSIZE block is tagged with the
2022  * cycle count because there in an implicit assumption that drives will
2023  * guarantee that entire 512 byte blocks get written at once.  In other words,
2024  * we can't have part of a 512 byte block written and part not written.  By
2025  * tagging each block, we will know which blocks are valid when recovering
2026  * after an unclean shutdown.
2027  *
2028  * This routine is single threaded on the iclog.  No other thread can be in
2029  * this routine with the same iclog.  Changing contents of iclog can there-
2030  * fore be done without grabbing the state machine lock.  Updating the global
2031  * log will require grabbing the lock though.
2032  *
2033  * The entire log manager uses a logical block numbering scheme.  Only
2034  * xlog_write_iclog knows about the fact that the log may not start with
2035  * block zero on a given device.
2036  */
2037 STATIC void
xlog_sync(struct xlog * log,struct xlog_in_core * iclog,struct xlog_ticket * ticket)2038 xlog_sync(
2039 	struct xlog		*log,
2040 	struct xlog_in_core	*iclog,
2041 	struct xlog_ticket	*ticket)
2042 {
2043 	unsigned int		count;		/* byte count of bwrite */
2044 	unsigned int		roundoff;       /* roundoff to BB or stripe */
2045 	uint64_t		bno;
2046 	unsigned int		size;
2047 
2048 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2049 	trace_xlog_iclog_sync(iclog, _RET_IP_);
2050 
2051 	count = xlog_calc_iclog_size(log, iclog, &roundoff);
2052 
2053 	/*
2054 	 * If we have a ticket, account for the roundoff via the ticket
2055 	 * reservation to avoid touching the hot grant heads needlessly.
2056 	 * Otherwise, we have to move grant heads directly.
2057 	 */
2058 	if (ticket) {
2059 		ticket->t_curr_res -= roundoff;
2060 	} else {
2061 		xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2062 		xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2063 	}
2064 
2065 	/* put cycle number in every block */
2066 	xlog_pack_data(log, iclog, roundoff);
2067 
2068 	/* real byte length */
2069 	size = iclog->ic_offset;
2070 	if (xfs_has_logv2(log->l_mp))
2071 		size += roundoff;
2072 	iclog->ic_header.h_len = cpu_to_be32(size);
2073 
2074 	XFS_STATS_INC(log->l_mp, xs_log_writes);
2075 	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2076 
2077 	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2078 
2079 	/* Do we need to split this write into 2 parts? */
2080 	if (bno + BTOBB(count) > log->l_logBBsize)
2081 		xlog_split_iclog(log, &iclog->ic_header, bno, count);
2082 
2083 	/* calculcate the checksum */
2084 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2085 					    iclog->ic_datap, size);
2086 	/*
2087 	 * Intentionally corrupt the log record CRC based on the error injection
2088 	 * frequency, if defined. This facilitates testing log recovery in the
2089 	 * event of torn writes. Hence, set the IOABORT state to abort the log
2090 	 * write on I/O completion and shutdown the fs. The subsequent mount
2091 	 * detects the bad CRC and attempts to recover.
2092 	 */
2093 #ifdef DEBUG
2094 	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2095 		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2096 		iclog->ic_fail_crc = true;
2097 		xfs_warn(log->l_mp,
2098 	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2099 			 be64_to_cpu(iclog->ic_header.h_lsn));
2100 	}
2101 #endif
2102 	xlog_verify_iclog(log, iclog, count);
2103 	xlog_write_iclog(log, iclog, bno, count);
2104 }
2105 
2106 /*
2107  * Deallocate a log structure
2108  */
2109 STATIC void
xlog_dealloc_log(struct xlog * log)2110 xlog_dealloc_log(
2111 	struct xlog	*log)
2112 {
2113 	xlog_in_core_t	*iclog, *next_iclog;
2114 	int		i;
2115 
2116 	/*
2117 	 * Cycle all the iclogbuf locks to make sure all log IO completion
2118 	 * is done before we tear down these buffers.
2119 	 */
2120 	iclog = log->l_iclog;
2121 	for (i = 0; i < log->l_iclog_bufs; i++) {
2122 		down(&iclog->ic_sema);
2123 		up(&iclog->ic_sema);
2124 		iclog = iclog->ic_next;
2125 	}
2126 
2127 	/*
2128 	 * Destroy the CIL after waiting for iclog IO completion because an
2129 	 * iclog EIO error will try to shut down the log, which accesses the
2130 	 * CIL to wake up the waiters.
2131 	 */
2132 	xlog_cil_destroy(log);
2133 
2134 	iclog = log->l_iclog;
2135 	for (i = 0; i < log->l_iclog_bufs; i++) {
2136 		next_iclog = iclog->ic_next;
2137 		kmem_free(iclog->ic_data);
2138 		kmem_free(iclog);
2139 		iclog = next_iclog;
2140 	}
2141 
2142 	log->l_mp->m_log = NULL;
2143 	destroy_workqueue(log->l_ioend_workqueue);
2144 	kmem_free(log);
2145 }
2146 
2147 /*
2148  * Update counters atomically now that memcpy is done.
2149  */
2150 static inline void
xlog_state_finish_copy(struct xlog * log,struct xlog_in_core * iclog,int record_cnt,int copy_bytes)2151 xlog_state_finish_copy(
2152 	struct xlog		*log,
2153 	struct xlog_in_core	*iclog,
2154 	int			record_cnt,
2155 	int			copy_bytes)
2156 {
2157 	lockdep_assert_held(&log->l_icloglock);
2158 
2159 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2160 	iclog->ic_offset += copy_bytes;
2161 }
2162 
2163 /*
2164  * print out info relating to regions written which consume
2165  * the reservation
2166  */
2167 void
xlog_print_tic_res(struct xfs_mount * mp,struct xlog_ticket * ticket)2168 xlog_print_tic_res(
2169 	struct xfs_mount	*mp,
2170 	struct xlog_ticket	*ticket)
2171 {
2172 	xfs_warn(mp, "ticket reservation summary:");
2173 	xfs_warn(mp, "  unit res    = %d bytes", ticket->t_unit_res);
2174 	xfs_warn(mp, "  current res = %d bytes", ticket->t_curr_res);
2175 	xfs_warn(mp, "  original count  = %d", ticket->t_ocnt);
2176 	xfs_warn(mp, "  remaining count = %d", ticket->t_cnt);
2177 }
2178 
2179 /*
2180  * Print a summary of the transaction.
2181  */
2182 void
xlog_print_trans(struct xfs_trans * tp)2183 xlog_print_trans(
2184 	struct xfs_trans	*tp)
2185 {
2186 	struct xfs_mount	*mp = tp->t_mountp;
2187 	struct xfs_log_item	*lip;
2188 
2189 	/* dump core transaction and ticket info */
2190 	xfs_warn(mp, "transaction summary:");
2191 	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2192 	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2193 	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2194 
2195 	xlog_print_tic_res(mp, tp->t_ticket);
2196 
2197 	/* dump each log item */
2198 	list_for_each_entry(lip, &tp->t_items, li_trans) {
2199 		struct xfs_log_vec	*lv = lip->li_lv;
2200 		struct xfs_log_iovec	*vec;
2201 		int			i;
2202 
2203 		xfs_warn(mp, "log item: ");
2204 		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2205 		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
2206 		if (!lv)
2207 			continue;
2208 		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2209 		xfs_warn(mp, "  size	= %d", lv->lv_size);
2210 		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2211 		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2212 
2213 		/* dump each iovec for the log item */
2214 		vec = lv->lv_iovecp;
2215 		for (i = 0; i < lv->lv_niovecs; i++) {
2216 			int dumplen = min(vec->i_len, 32);
2217 
2218 			xfs_warn(mp, "  iovec[%d]", i);
2219 			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2220 			xfs_warn(mp, "    len	= %d", vec->i_len);
2221 			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2222 			xfs_hex_dump(vec->i_addr, dumplen);
2223 
2224 			vec++;
2225 		}
2226 	}
2227 }
2228 
2229 static inline void
xlog_write_iovec(struct xlog_in_core * iclog,uint32_t * log_offset,void * data,uint32_t write_len,int * bytes_left,uint32_t * record_cnt,uint32_t * data_cnt)2230 xlog_write_iovec(
2231 	struct xlog_in_core	*iclog,
2232 	uint32_t		*log_offset,
2233 	void			*data,
2234 	uint32_t		write_len,
2235 	int			*bytes_left,
2236 	uint32_t		*record_cnt,
2237 	uint32_t		*data_cnt)
2238 {
2239 	ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2240 	ASSERT(*log_offset % sizeof(int32_t) == 0);
2241 	ASSERT(write_len % sizeof(int32_t) == 0);
2242 
2243 	memcpy(iclog->ic_datap + *log_offset, data, write_len);
2244 	*log_offset += write_len;
2245 	*bytes_left -= write_len;
2246 	(*record_cnt)++;
2247 	*data_cnt += write_len;
2248 }
2249 
2250 /*
2251  * Write log vectors into a single iclog which is guaranteed by the caller
2252  * to have enough space to write the entire log vector into.
2253  */
2254 static void
xlog_write_full(struct xfs_log_vec * lv,struct xlog_ticket * ticket,struct xlog_in_core * iclog,uint32_t * log_offset,uint32_t * len,uint32_t * record_cnt,uint32_t * data_cnt)2255 xlog_write_full(
2256 	struct xfs_log_vec	*lv,
2257 	struct xlog_ticket	*ticket,
2258 	struct xlog_in_core	*iclog,
2259 	uint32_t		*log_offset,
2260 	uint32_t		*len,
2261 	uint32_t		*record_cnt,
2262 	uint32_t		*data_cnt)
2263 {
2264 	int			index;
2265 
2266 	ASSERT(*log_offset + *len <= iclog->ic_size ||
2267 		iclog->ic_state == XLOG_STATE_WANT_SYNC);
2268 
2269 	/*
2270 	 * Ordered log vectors have no regions to write so this
2271 	 * loop will naturally skip them.
2272 	 */
2273 	for (index = 0; index < lv->lv_niovecs; index++) {
2274 		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2275 		struct xlog_op_header	*ophdr = reg->i_addr;
2276 
2277 		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2278 		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2279 				reg->i_len, len, record_cnt, data_cnt);
2280 	}
2281 }
2282 
2283 static int
xlog_write_get_more_iclog_space(struct xlog_ticket * ticket,struct xlog_in_core ** iclogp,uint32_t * log_offset,uint32_t len,uint32_t * record_cnt,uint32_t * data_cnt)2284 xlog_write_get_more_iclog_space(
2285 	struct xlog_ticket	*ticket,
2286 	struct xlog_in_core	**iclogp,
2287 	uint32_t		*log_offset,
2288 	uint32_t		len,
2289 	uint32_t		*record_cnt,
2290 	uint32_t		*data_cnt)
2291 {
2292 	struct xlog_in_core	*iclog = *iclogp;
2293 	struct xlog		*log = iclog->ic_log;
2294 	int			error;
2295 
2296 	spin_lock(&log->l_icloglock);
2297 	ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2298 	xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2299 	error = xlog_state_release_iclog(log, iclog, ticket);
2300 	spin_unlock(&log->l_icloglock);
2301 	if (error)
2302 		return error;
2303 
2304 	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2305 					log_offset);
2306 	if (error)
2307 		return error;
2308 	*record_cnt = 0;
2309 	*data_cnt = 0;
2310 	*iclogp = iclog;
2311 	return 0;
2312 }
2313 
2314 /*
2315  * Write log vectors into a single iclog which is smaller than the current chain
2316  * length. We write until we cannot fit a full record into the remaining space
2317  * and then stop. We return the log vector that is to be written that cannot
2318  * wholly fit in the iclog.
2319  */
2320 static int
xlog_write_partial(struct xfs_log_vec * lv,struct xlog_ticket * ticket,struct xlog_in_core ** iclogp,uint32_t * log_offset,uint32_t * len,uint32_t * record_cnt,uint32_t * data_cnt)2321 xlog_write_partial(
2322 	struct xfs_log_vec	*lv,
2323 	struct xlog_ticket	*ticket,
2324 	struct xlog_in_core	**iclogp,
2325 	uint32_t		*log_offset,
2326 	uint32_t		*len,
2327 	uint32_t		*record_cnt,
2328 	uint32_t		*data_cnt)
2329 {
2330 	struct xlog_in_core	*iclog = *iclogp;
2331 	struct xlog_op_header	*ophdr;
2332 	int			index = 0;
2333 	uint32_t		rlen;
2334 	int			error;
2335 
2336 	/* walk the logvec, copying until we run out of space in the iclog */
2337 	for (index = 0; index < lv->lv_niovecs; index++) {
2338 		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2339 		uint32_t		reg_offset = 0;
2340 
2341 		/*
2342 		 * The first region of a continuation must have a non-zero
2343 		 * length otherwise log recovery will just skip over it and
2344 		 * start recovering from the next opheader it finds. Because we
2345 		 * mark the next opheader as a continuation, recovery will then
2346 		 * incorrectly add the continuation to the previous region and
2347 		 * that breaks stuff.
2348 		 *
2349 		 * Hence if there isn't space for region data after the
2350 		 * opheader, then we need to start afresh with a new iclog.
2351 		 */
2352 		if (iclog->ic_size - *log_offset <=
2353 					sizeof(struct xlog_op_header)) {
2354 			error = xlog_write_get_more_iclog_space(ticket,
2355 					&iclog, log_offset, *len, record_cnt,
2356 					data_cnt);
2357 			if (error)
2358 				return error;
2359 		}
2360 
2361 		ophdr = reg->i_addr;
2362 		rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2363 
2364 		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2365 		ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2366 		if (rlen != reg->i_len)
2367 			ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2368 
2369 		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2370 				rlen, len, record_cnt, data_cnt);
2371 
2372 		/* If we wrote the whole region, move to the next. */
2373 		if (rlen == reg->i_len)
2374 			continue;
2375 
2376 		/*
2377 		 * We now have a partially written iovec, but it can span
2378 		 * multiple iclogs so we loop here. First we release the iclog
2379 		 * we currently have, then we get a new iclog and add a new
2380 		 * opheader. Then we continue copying from where we were until
2381 		 * we either complete the iovec or fill the iclog. If we
2382 		 * complete the iovec, then we increment the index and go right
2383 		 * back to the top of the outer loop. if we fill the iclog, we
2384 		 * run the inner loop again.
2385 		 *
2386 		 * This is complicated by the tail of a region using all the
2387 		 * space in an iclog and hence requiring us to release the iclog
2388 		 * and get a new one before returning to the outer loop. We must
2389 		 * always guarantee that we exit this inner loop with at least
2390 		 * space for log transaction opheaders left in the current
2391 		 * iclog, hence we cannot just terminate the loop at the end
2392 		 * of the of the continuation. So we loop while there is no
2393 		 * space left in the current iclog, and check for the end of the
2394 		 * continuation after getting a new iclog.
2395 		 */
2396 		do {
2397 			/*
2398 			 * Ensure we include the continuation opheader in the
2399 			 * space we need in the new iclog by adding that size
2400 			 * to the length we require. This continuation opheader
2401 			 * needs to be accounted to the ticket as the space it
2402 			 * consumes hasn't been accounted to the lv we are
2403 			 * writing.
2404 			 */
2405 			error = xlog_write_get_more_iclog_space(ticket,
2406 					&iclog, log_offset,
2407 					*len + sizeof(struct xlog_op_header),
2408 					record_cnt, data_cnt);
2409 			if (error)
2410 				return error;
2411 
2412 			ophdr = iclog->ic_datap + *log_offset;
2413 			ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2414 			ophdr->oh_clientid = XFS_TRANSACTION;
2415 			ophdr->oh_res2 = 0;
2416 			ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2417 
2418 			ticket->t_curr_res -= sizeof(struct xlog_op_header);
2419 			*log_offset += sizeof(struct xlog_op_header);
2420 			*data_cnt += sizeof(struct xlog_op_header);
2421 
2422 			/*
2423 			 * If rlen fits in the iclog, then end the region
2424 			 * continuation. Otherwise we're going around again.
2425 			 */
2426 			reg_offset += rlen;
2427 			rlen = reg->i_len - reg_offset;
2428 			if (rlen <= iclog->ic_size - *log_offset)
2429 				ophdr->oh_flags |= XLOG_END_TRANS;
2430 			else
2431 				ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2432 
2433 			rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2434 			ophdr->oh_len = cpu_to_be32(rlen);
2435 
2436 			xlog_write_iovec(iclog, log_offset,
2437 					reg->i_addr + reg_offset,
2438 					rlen, len, record_cnt, data_cnt);
2439 
2440 		} while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2441 	}
2442 
2443 	/*
2444 	 * No more iovecs remain in this logvec so return the next log vec to
2445 	 * the caller so it can go back to fast path copying.
2446 	 */
2447 	*iclogp = iclog;
2448 	return 0;
2449 }
2450 
2451 /*
2452  * Write some region out to in-core log
2453  *
2454  * This will be called when writing externally provided regions or when
2455  * writing out a commit record for a given transaction.
2456  *
2457  * General algorithm:
2458  *	1. Find total length of this write.  This may include adding to the
2459  *		lengths passed in.
2460  *	2. Check whether we violate the tickets reservation.
2461  *	3. While writing to this iclog
2462  *	    A. Reserve as much space in this iclog as can get
2463  *	    B. If this is first write, save away start lsn
2464  *	    C. While writing this region:
2465  *		1. If first write of transaction, write start record
2466  *		2. Write log operation header (header per region)
2467  *		3. Find out if we can fit entire region into this iclog
2468  *		4. Potentially, verify destination memcpy ptr
2469  *		5. Memcpy (partial) region
2470  *		6. If partial copy, release iclog; otherwise, continue
2471  *			copying more regions into current iclog
2472  *	4. Mark want sync bit (in simulation mode)
2473  *	5. Release iclog for potential flush to on-disk log.
2474  *
2475  * ERRORS:
2476  * 1.	Panic if reservation is overrun.  This should never happen since
2477  *	reservation amounts are generated internal to the filesystem.
2478  * NOTES:
2479  * 1. Tickets are single threaded data structures.
2480  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2481  *	syncing routine.  When a single log_write region needs to span
2482  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2483  *	on all log operation writes which don't contain the end of the
2484  *	region.  The XLOG_END_TRANS bit is used for the in-core log
2485  *	operation which contains the end of the continued log_write region.
2486  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2487  *	we don't really know exactly how much space will be used.  As a result,
2488  *	we don't update ic_offset until the end when we know exactly how many
2489  *	bytes have been written out.
2490  */
2491 int
xlog_write(struct xlog * log,struct xfs_cil_ctx * ctx,struct list_head * lv_chain,struct xlog_ticket * ticket,uint32_t len)2492 xlog_write(
2493 	struct xlog		*log,
2494 	struct xfs_cil_ctx	*ctx,
2495 	struct list_head	*lv_chain,
2496 	struct xlog_ticket	*ticket,
2497 	uint32_t		len)
2498 
2499 {
2500 	struct xlog_in_core	*iclog = NULL;
2501 	struct xfs_log_vec	*lv;
2502 	uint32_t		record_cnt = 0;
2503 	uint32_t		data_cnt = 0;
2504 	int			error = 0;
2505 	int			log_offset;
2506 
2507 	if (ticket->t_curr_res < 0) {
2508 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2509 		     "ctx ticket reservation ran out. Need to up reservation");
2510 		xlog_print_tic_res(log->l_mp, ticket);
2511 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2512 	}
2513 
2514 	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2515 					   &log_offset);
2516 	if (error)
2517 		return error;
2518 
2519 	ASSERT(log_offset <= iclog->ic_size - 1);
2520 
2521 	/*
2522 	 * If we have a context pointer, pass it the first iclog we are
2523 	 * writing to so it can record state needed for iclog write
2524 	 * ordering.
2525 	 */
2526 	if (ctx)
2527 		xlog_cil_set_ctx_write_state(ctx, iclog);
2528 
2529 	list_for_each_entry(lv, lv_chain, lv_list) {
2530 		/*
2531 		 * If the entire log vec does not fit in the iclog, punt it to
2532 		 * the partial copy loop which can handle this case.
2533 		 */
2534 		if (lv->lv_niovecs &&
2535 		    lv->lv_bytes > iclog->ic_size - log_offset) {
2536 			error = xlog_write_partial(lv, ticket, &iclog,
2537 					&log_offset, &len, &record_cnt,
2538 					&data_cnt);
2539 			if (error) {
2540 				/*
2541 				 * We have no iclog to release, so just return
2542 				 * the error immediately.
2543 				 */
2544 				return error;
2545 			}
2546 		} else {
2547 			xlog_write_full(lv, ticket, iclog, &log_offset,
2548 					 &len, &record_cnt, &data_cnt);
2549 		}
2550 	}
2551 	ASSERT(len == 0);
2552 
2553 	/*
2554 	 * We've already been guaranteed that the last writes will fit inside
2555 	 * the current iclog, and hence it will already have the space used by
2556 	 * those writes accounted to it. Hence we do not need to update the
2557 	 * iclog with the number of bytes written here.
2558 	 */
2559 	spin_lock(&log->l_icloglock);
2560 	xlog_state_finish_copy(log, iclog, record_cnt, 0);
2561 	error = xlog_state_release_iclog(log, iclog, ticket);
2562 	spin_unlock(&log->l_icloglock);
2563 
2564 	return error;
2565 }
2566 
2567 static void
xlog_state_activate_iclog(struct xlog_in_core * iclog,int * iclogs_changed)2568 xlog_state_activate_iclog(
2569 	struct xlog_in_core	*iclog,
2570 	int			*iclogs_changed)
2571 {
2572 	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2573 	trace_xlog_iclog_activate(iclog, _RET_IP_);
2574 
2575 	/*
2576 	 * If the number of ops in this iclog indicate it just contains the
2577 	 * dummy transaction, we can change state into IDLE (the second time
2578 	 * around). Otherwise we should change the state into NEED a dummy.
2579 	 * We don't need to cover the dummy.
2580 	 */
2581 	if (*iclogs_changed == 0 &&
2582 	    iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2583 		*iclogs_changed = 1;
2584 	} else {
2585 		/*
2586 		 * We have two dirty iclogs so start over.  This could also be
2587 		 * num of ops indicating this is not the dummy going out.
2588 		 */
2589 		*iclogs_changed = 2;
2590 	}
2591 
2592 	iclog->ic_state	= XLOG_STATE_ACTIVE;
2593 	iclog->ic_offset = 0;
2594 	iclog->ic_header.h_num_logops = 0;
2595 	memset(iclog->ic_header.h_cycle_data, 0,
2596 		sizeof(iclog->ic_header.h_cycle_data));
2597 	iclog->ic_header.h_lsn = 0;
2598 	iclog->ic_header.h_tail_lsn = 0;
2599 }
2600 
2601 /*
2602  * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2603  * ACTIVE after iclog I/O has completed.
2604  */
2605 static void
xlog_state_activate_iclogs(struct xlog * log,int * iclogs_changed)2606 xlog_state_activate_iclogs(
2607 	struct xlog		*log,
2608 	int			*iclogs_changed)
2609 {
2610 	struct xlog_in_core	*iclog = log->l_iclog;
2611 
2612 	do {
2613 		if (iclog->ic_state == XLOG_STATE_DIRTY)
2614 			xlog_state_activate_iclog(iclog, iclogs_changed);
2615 		/*
2616 		 * The ordering of marking iclogs ACTIVE must be maintained, so
2617 		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2618 		 */
2619 		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2620 			break;
2621 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2622 }
2623 
2624 static int
xlog_covered_state(int prev_state,int iclogs_changed)2625 xlog_covered_state(
2626 	int			prev_state,
2627 	int			iclogs_changed)
2628 {
2629 	/*
2630 	 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2631 	 * wrote the first covering record (DONE). We go to IDLE if we just
2632 	 * wrote the second covering record (DONE2) and remain in IDLE until a
2633 	 * non-covering write occurs.
2634 	 */
2635 	switch (prev_state) {
2636 	case XLOG_STATE_COVER_IDLE:
2637 		if (iclogs_changed == 1)
2638 			return XLOG_STATE_COVER_IDLE;
2639 		fallthrough;
2640 	case XLOG_STATE_COVER_NEED:
2641 	case XLOG_STATE_COVER_NEED2:
2642 		break;
2643 	case XLOG_STATE_COVER_DONE:
2644 		if (iclogs_changed == 1)
2645 			return XLOG_STATE_COVER_NEED2;
2646 		break;
2647 	case XLOG_STATE_COVER_DONE2:
2648 		if (iclogs_changed == 1)
2649 			return XLOG_STATE_COVER_IDLE;
2650 		break;
2651 	default:
2652 		ASSERT(0);
2653 	}
2654 
2655 	return XLOG_STATE_COVER_NEED;
2656 }
2657 
2658 STATIC void
xlog_state_clean_iclog(struct xlog * log,struct xlog_in_core * dirty_iclog)2659 xlog_state_clean_iclog(
2660 	struct xlog		*log,
2661 	struct xlog_in_core	*dirty_iclog)
2662 {
2663 	int			iclogs_changed = 0;
2664 
2665 	trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2666 
2667 	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2668 
2669 	xlog_state_activate_iclogs(log, &iclogs_changed);
2670 	wake_up_all(&dirty_iclog->ic_force_wait);
2671 
2672 	if (iclogs_changed) {
2673 		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2674 				iclogs_changed);
2675 	}
2676 }
2677 
2678 STATIC xfs_lsn_t
xlog_get_lowest_lsn(struct xlog * log)2679 xlog_get_lowest_lsn(
2680 	struct xlog		*log)
2681 {
2682 	struct xlog_in_core	*iclog = log->l_iclog;
2683 	xfs_lsn_t		lowest_lsn = 0, lsn;
2684 
2685 	do {
2686 		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2687 		    iclog->ic_state == XLOG_STATE_DIRTY)
2688 			continue;
2689 
2690 		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2691 		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2692 			lowest_lsn = lsn;
2693 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2694 
2695 	return lowest_lsn;
2696 }
2697 
2698 /*
2699  * Completion of a iclog IO does not imply that a transaction has completed, as
2700  * transactions can be large enough to span many iclogs. We cannot change the
2701  * tail of the log half way through a transaction as this may be the only
2702  * transaction in the log and moving the tail to point to the middle of it
2703  * will prevent recovery from finding the start of the transaction. Hence we
2704  * should only update the last_sync_lsn if this iclog contains transaction
2705  * completion callbacks on it.
2706  *
2707  * We have to do this before we drop the icloglock to ensure we are the only one
2708  * that can update it.
2709  *
2710  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2711  * the reservation grant head pushing. This is due to the fact that the push
2712  * target is bound by the current last_sync_lsn value. Hence if we have a large
2713  * amount of log space bound up in this committing transaction then the
2714  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2715  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2716  * should push the AIL to ensure the push target (and hence the grant head) is
2717  * no longer bound by the old log head location and can move forwards and make
2718  * progress again.
2719  */
2720 static void
xlog_state_set_callback(struct xlog * log,struct xlog_in_core * iclog,xfs_lsn_t header_lsn)2721 xlog_state_set_callback(
2722 	struct xlog		*log,
2723 	struct xlog_in_core	*iclog,
2724 	xfs_lsn_t		header_lsn)
2725 {
2726 	trace_xlog_iclog_callback(iclog, _RET_IP_);
2727 	iclog->ic_state = XLOG_STATE_CALLBACK;
2728 
2729 	ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2730 			   header_lsn) <= 0);
2731 
2732 	if (list_empty_careful(&iclog->ic_callbacks))
2733 		return;
2734 
2735 	atomic64_set(&log->l_last_sync_lsn, header_lsn);
2736 	xlog_grant_push_ail(log, 0);
2737 }
2738 
2739 /*
2740  * Return true if we need to stop processing, false to continue to the next
2741  * iclog. The caller will need to run callbacks if the iclog is returned in the
2742  * XLOG_STATE_CALLBACK state.
2743  */
2744 static bool
xlog_state_iodone_process_iclog(struct xlog * log,struct xlog_in_core * iclog)2745 xlog_state_iodone_process_iclog(
2746 	struct xlog		*log,
2747 	struct xlog_in_core	*iclog)
2748 {
2749 	xfs_lsn_t		lowest_lsn;
2750 	xfs_lsn_t		header_lsn;
2751 
2752 	switch (iclog->ic_state) {
2753 	case XLOG_STATE_ACTIVE:
2754 	case XLOG_STATE_DIRTY:
2755 		/*
2756 		 * Skip all iclogs in the ACTIVE & DIRTY states:
2757 		 */
2758 		return false;
2759 	case XLOG_STATE_DONE_SYNC:
2760 		/*
2761 		 * Now that we have an iclog that is in the DONE_SYNC state, do
2762 		 * one more check here to see if we have chased our tail around.
2763 		 * If this is not the lowest lsn iclog, then we will leave it
2764 		 * for another completion to process.
2765 		 */
2766 		header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2767 		lowest_lsn = xlog_get_lowest_lsn(log);
2768 		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2769 			return false;
2770 		xlog_state_set_callback(log, iclog, header_lsn);
2771 		return false;
2772 	default:
2773 		/*
2774 		 * Can only perform callbacks in order.  Since this iclog is not
2775 		 * in the DONE_SYNC state, we skip the rest and just try to
2776 		 * clean up.
2777 		 */
2778 		return true;
2779 	}
2780 }
2781 
2782 /*
2783  * Loop over all the iclogs, running attached callbacks on them. Return true if
2784  * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2785  * to handle transient shutdown state here at all because
2786  * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2787  * cleanup of the callbacks.
2788  */
2789 static bool
xlog_state_do_iclog_callbacks(struct xlog * log)2790 xlog_state_do_iclog_callbacks(
2791 	struct xlog		*log)
2792 		__releases(&log->l_icloglock)
2793 		__acquires(&log->l_icloglock)
2794 {
2795 	struct xlog_in_core	*first_iclog = log->l_iclog;
2796 	struct xlog_in_core	*iclog = first_iclog;
2797 	bool			ran_callback = false;
2798 
2799 	do {
2800 		LIST_HEAD(cb_list);
2801 
2802 		if (xlog_state_iodone_process_iclog(log, iclog))
2803 			break;
2804 		if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2805 			iclog = iclog->ic_next;
2806 			continue;
2807 		}
2808 		list_splice_init(&iclog->ic_callbacks, &cb_list);
2809 		spin_unlock(&log->l_icloglock);
2810 
2811 		trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2812 		xlog_cil_process_committed(&cb_list);
2813 		trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2814 		ran_callback = true;
2815 
2816 		spin_lock(&log->l_icloglock);
2817 		xlog_state_clean_iclog(log, iclog);
2818 		iclog = iclog->ic_next;
2819 	} while (iclog != first_iclog);
2820 
2821 	return ran_callback;
2822 }
2823 
2824 
2825 /*
2826  * Loop running iclog completion callbacks until there are no more iclogs in a
2827  * state that can run callbacks.
2828  */
2829 STATIC void
xlog_state_do_callback(struct xlog * log)2830 xlog_state_do_callback(
2831 	struct xlog		*log)
2832 {
2833 	int			flushcnt = 0;
2834 	int			repeats = 0;
2835 
2836 	spin_lock(&log->l_icloglock);
2837 	while (xlog_state_do_iclog_callbacks(log)) {
2838 		if (xlog_is_shutdown(log))
2839 			break;
2840 
2841 		if (++repeats > 5000) {
2842 			flushcnt += repeats;
2843 			repeats = 0;
2844 			xfs_warn(log->l_mp,
2845 				"%s: possible infinite loop (%d iterations)",
2846 				__func__, flushcnt);
2847 		}
2848 	}
2849 
2850 	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2851 		wake_up_all(&log->l_flush_wait);
2852 
2853 	spin_unlock(&log->l_icloglock);
2854 }
2855 
2856 
2857 /*
2858  * Finish transitioning this iclog to the dirty state.
2859  *
2860  * Callbacks could take time, so they are done outside the scope of the
2861  * global state machine log lock.
2862  */
2863 STATIC void
xlog_state_done_syncing(struct xlog_in_core * iclog)2864 xlog_state_done_syncing(
2865 	struct xlog_in_core	*iclog)
2866 {
2867 	struct xlog		*log = iclog->ic_log;
2868 
2869 	spin_lock(&log->l_icloglock);
2870 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2871 	trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2872 
2873 	/*
2874 	 * If we got an error, either on the first buffer, or in the case of
2875 	 * split log writes, on the second, we shut down the file system and
2876 	 * no iclogs should ever be attempted to be written to disk again.
2877 	 */
2878 	if (!xlog_is_shutdown(log)) {
2879 		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2880 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2881 	}
2882 
2883 	/*
2884 	 * Someone could be sleeping prior to writing out the next
2885 	 * iclog buffer, we wake them all, one will get to do the
2886 	 * I/O, the others get to wait for the result.
2887 	 */
2888 	wake_up_all(&iclog->ic_write_wait);
2889 	spin_unlock(&log->l_icloglock);
2890 	xlog_state_do_callback(log);
2891 }
2892 
2893 /*
2894  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2895  * sleep.  We wait on the flush queue on the head iclog as that should be
2896  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2897  * we will wait here and all new writes will sleep until a sync completes.
2898  *
2899  * The in-core logs are used in a circular fashion. They are not used
2900  * out-of-order even when an iclog past the head is free.
2901  *
2902  * return:
2903  *	* log_offset where xlog_write() can start writing into the in-core
2904  *		log's data space.
2905  *	* in-core log pointer to which xlog_write() should write.
2906  *	* boolean indicating this is a continued write to an in-core log.
2907  *		If this is the last write, then the in-core log's offset field
2908  *		needs to be incremented, depending on the amount of data which
2909  *		is copied.
2910  */
2911 STATIC int
xlog_state_get_iclog_space(struct xlog * log,int len,struct xlog_in_core ** iclogp,struct xlog_ticket * ticket,int * logoffsetp)2912 xlog_state_get_iclog_space(
2913 	struct xlog		*log,
2914 	int			len,
2915 	struct xlog_in_core	**iclogp,
2916 	struct xlog_ticket	*ticket,
2917 	int			*logoffsetp)
2918 {
2919 	int		  log_offset;
2920 	xlog_rec_header_t *head;
2921 	xlog_in_core_t	  *iclog;
2922 
2923 restart:
2924 	spin_lock(&log->l_icloglock);
2925 	if (xlog_is_shutdown(log)) {
2926 		spin_unlock(&log->l_icloglock);
2927 		return -EIO;
2928 	}
2929 
2930 	iclog = log->l_iclog;
2931 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2932 		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2933 
2934 		/* Wait for log writes to have flushed */
2935 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2936 		goto restart;
2937 	}
2938 
2939 	head = &iclog->ic_header;
2940 
2941 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2942 	log_offset = iclog->ic_offset;
2943 
2944 	trace_xlog_iclog_get_space(iclog, _RET_IP_);
2945 
2946 	/* On the 1st write to an iclog, figure out lsn.  This works
2947 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2948 	 * committing to.  If the offset is set, that's how many blocks
2949 	 * must be written.
2950 	 */
2951 	if (log_offset == 0) {
2952 		ticket->t_curr_res -= log->l_iclog_hsize;
2953 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2954 		head->h_lsn = cpu_to_be64(
2955 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2956 		ASSERT(log->l_curr_block >= 0);
2957 	}
2958 
2959 	/* If there is enough room to write everything, then do it.  Otherwise,
2960 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2961 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2962 	 * until you know exactly how many bytes get copied.  Therefore, wait
2963 	 * until later to update ic_offset.
2964 	 *
2965 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2966 	 * can fit into remaining data section.
2967 	 */
2968 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2969 		int		error = 0;
2970 
2971 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2972 
2973 		/*
2974 		 * If we are the only one writing to this iclog, sync it to
2975 		 * disk.  We need to do an atomic compare and decrement here to
2976 		 * avoid racing with concurrent atomic_dec_and_lock() calls in
2977 		 * xlog_state_release_iclog() when there is more than one
2978 		 * reference to the iclog.
2979 		 */
2980 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2981 			error = xlog_state_release_iclog(log, iclog, ticket);
2982 		spin_unlock(&log->l_icloglock);
2983 		if (error)
2984 			return error;
2985 		goto restart;
2986 	}
2987 
2988 	/* Do we have enough room to write the full amount in the remainder
2989 	 * of this iclog?  Or must we continue a write on the next iclog and
2990 	 * mark this iclog as completely taken?  In the case where we switch
2991 	 * iclogs (to mark it taken), this particular iclog will release/sync
2992 	 * to disk in xlog_write().
2993 	 */
2994 	if (len <= iclog->ic_size - iclog->ic_offset)
2995 		iclog->ic_offset += len;
2996 	else
2997 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2998 	*iclogp = iclog;
2999 
3000 	ASSERT(iclog->ic_offset <= iclog->ic_size);
3001 	spin_unlock(&log->l_icloglock);
3002 
3003 	*logoffsetp = log_offset;
3004 	return 0;
3005 }
3006 
3007 /*
3008  * The first cnt-1 times a ticket goes through here we don't need to move the
3009  * grant write head because the permanent reservation has reserved cnt times the
3010  * unit amount.  Release part of current permanent unit reservation and reset
3011  * current reservation to be one units worth.  Also move grant reservation head
3012  * forward.
3013  */
3014 void
xfs_log_ticket_regrant(struct xlog * log,struct xlog_ticket * ticket)3015 xfs_log_ticket_regrant(
3016 	struct xlog		*log,
3017 	struct xlog_ticket	*ticket)
3018 {
3019 	trace_xfs_log_ticket_regrant(log, ticket);
3020 
3021 	if (ticket->t_cnt > 0)
3022 		ticket->t_cnt--;
3023 
3024 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3025 					ticket->t_curr_res);
3026 	xlog_grant_sub_space(log, &log->l_write_head.grant,
3027 					ticket->t_curr_res);
3028 	ticket->t_curr_res = ticket->t_unit_res;
3029 
3030 	trace_xfs_log_ticket_regrant_sub(log, ticket);
3031 
3032 	/* just return if we still have some of the pre-reserved space */
3033 	if (!ticket->t_cnt) {
3034 		xlog_grant_add_space(log, &log->l_reserve_head.grant,
3035 				     ticket->t_unit_res);
3036 		trace_xfs_log_ticket_regrant_exit(log, ticket);
3037 
3038 		ticket->t_curr_res = ticket->t_unit_res;
3039 	}
3040 
3041 	xfs_log_ticket_put(ticket);
3042 }
3043 
3044 /*
3045  * Give back the space left from a reservation.
3046  *
3047  * All the information we need to make a correct determination of space left
3048  * is present.  For non-permanent reservations, things are quite easy.  The
3049  * count should have been decremented to zero.  We only need to deal with the
3050  * space remaining in the current reservation part of the ticket.  If the
3051  * ticket contains a permanent reservation, there may be left over space which
3052  * needs to be released.  A count of N means that N-1 refills of the current
3053  * reservation can be done before we need to ask for more space.  The first
3054  * one goes to fill up the first current reservation.  Once we run out of
3055  * space, the count will stay at zero and the only space remaining will be
3056  * in the current reservation field.
3057  */
3058 void
xfs_log_ticket_ungrant(struct xlog * log,struct xlog_ticket * ticket)3059 xfs_log_ticket_ungrant(
3060 	struct xlog		*log,
3061 	struct xlog_ticket	*ticket)
3062 {
3063 	int			bytes;
3064 
3065 	trace_xfs_log_ticket_ungrant(log, ticket);
3066 
3067 	if (ticket->t_cnt > 0)
3068 		ticket->t_cnt--;
3069 
3070 	trace_xfs_log_ticket_ungrant_sub(log, ticket);
3071 
3072 	/*
3073 	 * If this is a permanent reservation ticket, we may be able to free
3074 	 * up more space based on the remaining count.
3075 	 */
3076 	bytes = ticket->t_curr_res;
3077 	if (ticket->t_cnt > 0) {
3078 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3079 		bytes += ticket->t_unit_res*ticket->t_cnt;
3080 	}
3081 
3082 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3083 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3084 
3085 	trace_xfs_log_ticket_ungrant_exit(log, ticket);
3086 
3087 	xfs_log_space_wake(log->l_mp);
3088 	xfs_log_ticket_put(ticket);
3089 }
3090 
3091 /*
3092  * This routine will mark the current iclog in the ring as WANT_SYNC and move
3093  * the current iclog pointer to the next iclog in the ring.
3094  */
3095 void
xlog_state_switch_iclogs(struct xlog * log,struct xlog_in_core * iclog,int eventual_size)3096 xlog_state_switch_iclogs(
3097 	struct xlog		*log,
3098 	struct xlog_in_core	*iclog,
3099 	int			eventual_size)
3100 {
3101 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3102 	assert_spin_locked(&log->l_icloglock);
3103 	trace_xlog_iclog_switch(iclog, _RET_IP_);
3104 
3105 	if (!eventual_size)
3106 		eventual_size = iclog->ic_offset;
3107 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3108 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3109 	log->l_prev_block = log->l_curr_block;
3110 	log->l_prev_cycle = log->l_curr_cycle;
3111 
3112 	/* roll log?: ic_offset changed later */
3113 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3114 
3115 	/* Round up to next log-sunit */
3116 	if (log->l_iclog_roundoff > BBSIZE) {
3117 		uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3118 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3119 	}
3120 
3121 	if (log->l_curr_block >= log->l_logBBsize) {
3122 		/*
3123 		 * Rewind the current block before the cycle is bumped to make
3124 		 * sure that the combined LSN never transiently moves forward
3125 		 * when the log wraps to the next cycle. This is to support the
3126 		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3127 		 * other cases should acquire l_icloglock.
3128 		 */
3129 		log->l_curr_block -= log->l_logBBsize;
3130 		ASSERT(log->l_curr_block >= 0);
3131 		smp_wmb();
3132 		log->l_curr_cycle++;
3133 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3134 			log->l_curr_cycle++;
3135 	}
3136 	ASSERT(iclog == log->l_iclog);
3137 	log->l_iclog = iclog->ic_next;
3138 }
3139 
3140 /*
3141  * Force the iclog to disk and check if the iclog has been completed before
3142  * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3143  * pmem) or fast async storage because we drop the icloglock to issue the IO.
3144  * If completion has already occurred, tell the caller so that it can avoid an
3145  * unnecessary wait on the iclog.
3146  */
3147 static int
xlog_force_and_check_iclog(struct xlog_in_core * iclog,bool * completed)3148 xlog_force_and_check_iclog(
3149 	struct xlog_in_core	*iclog,
3150 	bool			*completed)
3151 {
3152 	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3153 	int			error;
3154 
3155 	*completed = false;
3156 	error = xlog_force_iclog(iclog);
3157 	if (error)
3158 		return error;
3159 
3160 	/*
3161 	 * If the iclog has already been completed and reused the header LSN
3162 	 * will have been rewritten by completion
3163 	 */
3164 	if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3165 		*completed = true;
3166 	return 0;
3167 }
3168 
3169 /*
3170  * Write out all data in the in-core log as of this exact moment in time.
3171  *
3172  * Data may be written to the in-core log during this call.  However,
3173  * we don't guarantee this data will be written out.  A change from past
3174  * implementation means this routine will *not* write out zero length LRs.
3175  *
3176  * Basically, we try and perform an intelligent scan of the in-core logs.
3177  * If we determine there is no flushable data, we just return.  There is no
3178  * flushable data if:
3179  *
3180  *	1. the current iclog is active and has no data; the previous iclog
3181  *		is in the active or dirty state.
3182  *	2. the current iclog is drity, and the previous iclog is in the
3183  *		active or dirty state.
3184  *
3185  * We may sleep if:
3186  *
3187  *	1. the current iclog is not in the active nor dirty state.
3188  *	2. the current iclog dirty, and the previous iclog is not in the
3189  *		active nor dirty state.
3190  *	3. the current iclog is active, and there is another thread writing
3191  *		to this particular iclog.
3192  *	4. a) the current iclog is active and has no other writers
3193  *	   b) when we return from flushing out this iclog, it is still
3194  *		not in the active nor dirty state.
3195  */
3196 int
xfs_log_force(struct xfs_mount * mp,uint flags)3197 xfs_log_force(
3198 	struct xfs_mount	*mp,
3199 	uint			flags)
3200 {
3201 	struct xlog		*log = mp->m_log;
3202 	struct xlog_in_core	*iclog;
3203 
3204 	XFS_STATS_INC(mp, xs_log_force);
3205 	trace_xfs_log_force(mp, 0, _RET_IP_);
3206 
3207 	xlog_cil_force(log);
3208 
3209 	spin_lock(&log->l_icloglock);
3210 	if (xlog_is_shutdown(log))
3211 		goto out_error;
3212 
3213 	iclog = log->l_iclog;
3214 	trace_xlog_iclog_force(iclog, _RET_IP_);
3215 
3216 	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3217 	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3218 	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3219 		/*
3220 		 * If the head is dirty or (active and empty), then we need to
3221 		 * look at the previous iclog.
3222 		 *
3223 		 * If the previous iclog is active or dirty we are done.  There
3224 		 * is nothing to sync out. Otherwise, we attach ourselves to the
3225 		 * previous iclog and go to sleep.
3226 		 */
3227 		iclog = iclog->ic_prev;
3228 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3229 		if (atomic_read(&iclog->ic_refcnt) == 0) {
3230 			/* We have exclusive access to this iclog. */
3231 			bool	completed;
3232 
3233 			if (xlog_force_and_check_iclog(iclog, &completed))
3234 				goto out_error;
3235 
3236 			if (completed)
3237 				goto out_unlock;
3238 		} else {
3239 			/*
3240 			 * Someone else is still writing to this iclog, so we
3241 			 * need to ensure that when they release the iclog it
3242 			 * gets synced immediately as we may be waiting on it.
3243 			 */
3244 			xlog_state_switch_iclogs(log, iclog, 0);
3245 		}
3246 	}
3247 
3248 	/*
3249 	 * The iclog we are about to wait on may contain the checkpoint pushed
3250 	 * by the above xlog_cil_force() call, but it may not have been pushed
3251 	 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3252 	 * are flushed when this iclog is written.
3253 	 */
3254 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3255 		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3256 
3257 	if (flags & XFS_LOG_SYNC)
3258 		return xlog_wait_on_iclog(iclog);
3259 out_unlock:
3260 	spin_unlock(&log->l_icloglock);
3261 	return 0;
3262 out_error:
3263 	spin_unlock(&log->l_icloglock);
3264 	return -EIO;
3265 }
3266 
3267 /*
3268  * Force the log to a specific LSN.
3269  *
3270  * If an iclog with that lsn can be found:
3271  *	If it is in the DIRTY state, just return.
3272  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3273  *		state and go to sleep or return.
3274  *	If it is in any other state, go to sleep or return.
3275  *
3276  * Synchronous forces are implemented with a wait queue.  All callers trying
3277  * to force a given lsn to disk must wait on the queue attached to the
3278  * specific in-core log.  When given in-core log finally completes its write
3279  * to disk, that thread will wake up all threads waiting on the queue.
3280  */
3281 static int
xlog_force_lsn(struct xlog * log,xfs_lsn_t lsn,uint flags,int * log_flushed,bool already_slept)3282 xlog_force_lsn(
3283 	struct xlog		*log,
3284 	xfs_lsn_t		lsn,
3285 	uint			flags,
3286 	int			*log_flushed,
3287 	bool			already_slept)
3288 {
3289 	struct xlog_in_core	*iclog;
3290 	bool			completed;
3291 
3292 	spin_lock(&log->l_icloglock);
3293 	if (xlog_is_shutdown(log))
3294 		goto out_error;
3295 
3296 	iclog = log->l_iclog;
3297 	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3298 		trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3299 		iclog = iclog->ic_next;
3300 		if (iclog == log->l_iclog)
3301 			goto out_unlock;
3302 	}
3303 
3304 	switch (iclog->ic_state) {
3305 	case XLOG_STATE_ACTIVE:
3306 		/*
3307 		 * We sleep here if we haven't already slept (e.g. this is the
3308 		 * first time we've looked at the correct iclog buf) and the
3309 		 * buffer before us is going to be sync'ed.  The reason for this
3310 		 * is that if we are doing sync transactions here, by waiting
3311 		 * for the previous I/O to complete, we can allow a few more
3312 		 * transactions into this iclog before we close it down.
3313 		 *
3314 		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3315 		 * refcnt so we can release the log (which drops the ref count).
3316 		 * The state switch keeps new transaction commits from using
3317 		 * this buffer.  When the current commits finish writing into
3318 		 * the buffer, the refcount will drop to zero and the buffer
3319 		 * will go out then.
3320 		 */
3321 		if (!already_slept &&
3322 		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3323 		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3324 			xlog_wait(&iclog->ic_prev->ic_write_wait,
3325 					&log->l_icloglock);
3326 			return -EAGAIN;
3327 		}
3328 		if (xlog_force_and_check_iclog(iclog, &completed))
3329 			goto out_error;
3330 		if (log_flushed)
3331 			*log_flushed = 1;
3332 		if (completed)
3333 			goto out_unlock;
3334 		break;
3335 	case XLOG_STATE_WANT_SYNC:
3336 		/*
3337 		 * This iclog may contain the checkpoint pushed by the
3338 		 * xlog_cil_force_seq() call, but there are other writers still
3339 		 * accessing it so it hasn't been pushed to disk yet. Like the
3340 		 * ACTIVE case above, we need to make sure caches are flushed
3341 		 * when this iclog is written.
3342 		 */
3343 		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3344 		break;
3345 	default:
3346 		/*
3347 		 * The entire checkpoint was written by the CIL force and is on
3348 		 * its way to disk already. It will be stable when it
3349 		 * completes, so we don't need to manipulate caches here at all.
3350 		 * We just need to wait for completion if necessary.
3351 		 */
3352 		break;
3353 	}
3354 
3355 	if (flags & XFS_LOG_SYNC)
3356 		return xlog_wait_on_iclog(iclog);
3357 out_unlock:
3358 	spin_unlock(&log->l_icloglock);
3359 	return 0;
3360 out_error:
3361 	spin_unlock(&log->l_icloglock);
3362 	return -EIO;
3363 }
3364 
3365 /*
3366  * Force the log to a specific checkpoint sequence.
3367  *
3368  * First force the CIL so that all the required changes have been flushed to the
3369  * iclogs. If the CIL force completed it will return a commit LSN that indicates
3370  * the iclog that needs to be flushed to stable storage. If the caller needs
3371  * a synchronous log force, we will wait on the iclog with the LSN returned by
3372  * xlog_cil_force_seq() to be completed.
3373  */
3374 int
xfs_log_force_seq(struct xfs_mount * mp,xfs_csn_t seq,uint flags,int * log_flushed)3375 xfs_log_force_seq(
3376 	struct xfs_mount	*mp,
3377 	xfs_csn_t		seq,
3378 	uint			flags,
3379 	int			*log_flushed)
3380 {
3381 	struct xlog		*log = mp->m_log;
3382 	xfs_lsn_t		lsn;
3383 	int			ret;
3384 	ASSERT(seq != 0);
3385 
3386 	XFS_STATS_INC(mp, xs_log_force);
3387 	trace_xfs_log_force(mp, seq, _RET_IP_);
3388 
3389 	lsn = xlog_cil_force_seq(log, seq);
3390 	if (lsn == NULLCOMMITLSN)
3391 		return 0;
3392 
3393 	ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3394 	if (ret == -EAGAIN) {
3395 		XFS_STATS_INC(mp, xs_log_force_sleep);
3396 		ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3397 	}
3398 	return ret;
3399 }
3400 
3401 /*
3402  * Free a used ticket when its refcount falls to zero.
3403  */
3404 void
xfs_log_ticket_put(xlog_ticket_t * ticket)3405 xfs_log_ticket_put(
3406 	xlog_ticket_t	*ticket)
3407 {
3408 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3409 	if (atomic_dec_and_test(&ticket->t_ref))
3410 		kmem_cache_free(xfs_log_ticket_cache, ticket);
3411 }
3412 
3413 xlog_ticket_t *
xfs_log_ticket_get(xlog_ticket_t * ticket)3414 xfs_log_ticket_get(
3415 	xlog_ticket_t	*ticket)
3416 {
3417 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3418 	atomic_inc(&ticket->t_ref);
3419 	return ticket;
3420 }
3421 
3422 /*
3423  * Figure out the total log space unit (in bytes) that would be
3424  * required for a log ticket.
3425  */
3426 static int
xlog_calc_unit_res(struct xlog * log,int unit_bytes,int * niclogs)3427 xlog_calc_unit_res(
3428 	struct xlog		*log,
3429 	int			unit_bytes,
3430 	int			*niclogs)
3431 {
3432 	int			iclog_space;
3433 	uint			num_headers;
3434 
3435 	/*
3436 	 * Permanent reservations have up to 'cnt'-1 active log operations
3437 	 * in the log.  A unit in this case is the amount of space for one
3438 	 * of these log operations.  Normal reservations have a cnt of 1
3439 	 * and their unit amount is the total amount of space required.
3440 	 *
3441 	 * The following lines of code account for non-transaction data
3442 	 * which occupy space in the on-disk log.
3443 	 *
3444 	 * Normal form of a transaction is:
3445 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3446 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3447 	 *
3448 	 * We need to account for all the leadup data and trailer data
3449 	 * around the transaction data.
3450 	 * And then we need to account for the worst case in terms of using
3451 	 * more space.
3452 	 * The worst case will happen if:
3453 	 * - the placement of the transaction happens to be such that the
3454 	 *   roundoff is at its maximum
3455 	 * - the transaction data is synced before the commit record is synced
3456 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3457 	 *   Therefore the commit record is in its own Log Record.
3458 	 *   This can happen as the commit record is called with its
3459 	 *   own region to xlog_write().
3460 	 *   This then means that in the worst case, roundoff can happen for
3461 	 *   the commit-rec as well.
3462 	 *   The commit-rec is smaller than padding in this scenario and so it is
3463 	 *   not added separately.
3464 	 */
3465 
3466 	/* for trans header */
3467 	unit_bytes += sizeof(xlog_op_header_t);
3468 	unit_bytes += sizeof(xfs_trans_header_t);
3469 
3470 	/* for start-rec */
3471 	unit_bytes += sizeof(xlog_op_header_t);
3472 
3473 	/*
3474 	 * for LR headers - the space for data in an iclog is the size minus
3475 	 * the space used for the headers. If we use the iclog size, then we
3476 	 * undercalculate the number of headers required.
3477 	 *
3478 	 * Furthermore - the addition of op headers for split-recs might
3479 	 * increase the space required enough to require more log and op
3480 	 * headers, so take that into account too.
3481 	 *
3482 	 * IMPORTANT: This reservation makes the assumption that if this
3483 	 * transaction is the first in an iclog and hence has the LR headers
3484 	 * accounted to it, then the remaining space in the iclog is
3485 	 * exclusively for this transaction.  i.e. if the transaction is larger
3486 	 * than the iclog, it will be the only thing in that iclog.
3487 	 * Fundamentally, this means we must pass the entire log vector to
3488 	 * xlog_write to guarantee this.
3489 	 */
3490 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3491 	num_headers = howmany(unit_bytes, iclog_space);
3492 
3493 	/* for split-recs - ophdrs added when data split over LRs */
3494 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3495 
3496 	/* add extra header reservations if we overrun */
3497 	while (!num_headers ||
3498 	       howmany(unit_bytes, iclog_space) > num_headers) {
3499 		unit_bytes += sizeof(xlog_op_header_t);
3500 		num_headers++;
3501 	}
3502 	unit_bytes += log->l_iclog_hsize * num_headers;
3503 
3504 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3505 	unit_bytes += log->l_iclog_hsize;
3506 
3507 	/* roundoff padding for transaction data and one for commit record */
3508 	unit_bytes += 2 * log->l_iclog_roundoff;
3509 
3510 	if (niclogs)
3511 		*niclogs = num_headers;
3512 	return unit_bytes;
3513 }
3514 
3515 int
xfs_log_calc_unit_res(struct xfs_mount * mp,int unit_bytes)3516 xfs_log_calc_unit_res(
3517 	struct xfs_mount	*mp,
3518 	int			unit_bytes)
3519 {
3520 	return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3521 }
3522 
3523 /*
3524  * Allocate and initialise a new log ticket.
3525  */
3526 struct xlog_ticket *
xlog_ticket_alloc(struct xlog * log,int unit_bytes,int cnt,bool permanent)3527 xlog_ticket_alloc(
3528 	struct xlog		*log,
3529 	int			unit_bytes,
3530 	int			cnt,
3531 	bool			permanent)
3532 {
3533 	struct xlog_ticket	*tic;
3534 	int			unit_res;
3535 
3536 	tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3537 
3538 	unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3539 
3540 	atomic_set(&tic->t_ref, 1);
3541 	tic->t_task		= current;
3542 	INIT_LIST_HEAD(&tic->t_queue);
3543 	tic->t_unit_res		= unit_res;
3544 	tic->t_curr_res		= unit_res;
3545 	tic->t_cnt		= cnt;
3546 	tic->t_ocnt		= cnt;
3547 	tic->t_tid		= get_random_u32();
3548 	if (permanent)
3549 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3550 
3551 	return tic;
3552 }
3553 
3554 #if defined(DEBUG)
3555 /*
3556  * Check to make sure the grant write head didn't just over lap the tail.  If
3557  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3558  * the cycles differ by exactly one and check the byte count.
3559  *
3560  * This check is run unlocked, so can give false positives. Rather than assert
3561  * on failures, use a warn-once flag and a panic tag to allow the admin to
3562  * determine if they want to panic the machine when such an error occurs. For
3563  * debug kernels this will have the same effect as using an assert but, unlinke
3564  * an assert, it can be turned off at runtime.
3565  */
3566 STATIC void
xlog_verify_grant_tail(struct xlog * log)3567 xlog_verify_grant_tail(
3568 	struct xlog	*log)
3569 {
3570 	int		tail_cycle, tail_blocks;
3571 	int		cycle, space;
3572 
3573 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3574 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3575 	if (tail_cycle != cycle) {
3576 		if (cycle - 1 != tail_cycle &&
3577 		    !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3578 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3579 				"%s: cycle - 1 != tail_cycle", __func__);
3580 		}
3581 
3582 		if (space > BBTOB(tail_blocks) &&
3583 		    !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3584 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3585 				"%s: space > BBTOB(tail_blocks)", __func__);
3586 		}
3587 	}
3588 }
3589 
3590 /* check if it will fit */
3591 STATIC void
xlog_verify_tail_lsn(struct xlog * log,struct xlog_in_core * iclog)3592 xlog_verify_tail_lsn(
3593 	struct xlog		*log,
3594 	struct xlog_in_core	*iclog)
3595 {
3596 	xfs_lsn_t	tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3597 	int		blocks;
3598 
3599     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3600 	blocks =
3601 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3602 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3603 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3604     } else {
3605 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3606 
3607 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3608 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3609 
3610 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3611 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3612 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3613     }
3614 }
3615 
3616 /*
3617  * Perform a number of checks on the iclog before writing to disk.
3618  *
3619  * 1. Make sure the iclogs are still circular
3620  * 2. Make sure we have a good magic number
3621  * 3. Make sure we don't have magic numbers in the data
3622  * 4. Check fields of each log operation header for:
3623  *	A. Valid client identifier
3624  *	B. tid ptr value falls in valid ptr space (user space code)
3625  *	C. Length in log record header is correct according to the
3626  *		individual operation headers within record.
3627  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3628  *	log, check the preceding blocks of the physical log to make sure all
3629  *	the cycle numbers agree with the current cycle number.
3630  */
3631 STATIC void
xlog_verify_iclog(struct xlog * log,struct xlog_in_core * iclog,int count)3632 xlog_verify_iclog(
3633 	struct xlog		*log,
3634 	struct xlog_in_core	*iclog,
3635 	int			count)
3636 {
3637 	xlog_op_header_t	*ophead;
3638 	xlog_in_core_t		*icptr;
3639 	xlog_in_core_2_t	*xhdr;
3640 	void			*base_ptr, *ptr, *p;
3641 	ptrdiff_t		field_offset;
3642 	uint8_t			clientid;
3643 	int			len, i, j, k, op_len;
3644 	int			idx;
3645 
3646 	/* check validity of iclog pointers */
3647 	spin_lock(&log->l_icloglock);
3648 	icptr = log->l_iclog;
3649 	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3650 		ASSERT(icptr);
3651 
3652 	if (icptr != log->l_iclog)
3653 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3654 	spin_unlock(&log->l_icloglock);
3655 
3656 	/* check log magic numbers */
3657 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3658 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3659 
3660 	base_ptr = ptr = &iclog->ic_header;
3661 	p = &iclog->ic_header;
3662 	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3663 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3664 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3665 				__func__);
3666 	}
3667 
3668 	/* check fields */
3669 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3670 	base_ptr = ptr = iclog->ic_datap;
3671 	ophead = ptr;
3672 	xhdr = iclog->ic_data;
3673 	for (i = 0; i < len; i++) {
3674 		ophead = ptr;
3675 
3676 		/* clientid is only 1 byte */
3677 		p = &ophead->oh_clientid;
3678 		field_offset = p - base_ptr;
3679 		if (field_offset & 0x1ff) {
3680 			clientid = ophead->oh_clientid;
3681 		} else {
3682 			idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3683 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3684 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3685 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3686 				clientid = xlog_get_client_id(
3687 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3688 			} else {
3689 				clientid = xlog_get_client_id(
3690 					iclog->ic_header.h_cycle_data[idx]);
3691 			}
3692 		}
3693 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3694 			xfs_warn(log->l_mp,
3695 				"%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3696 				__func__, i, clientid, ophead,
3697 				(unsigned long)field_offset);
3698 		}
3699 
3700 		/* check length */
3701 		p = &ophead->oh_len;
3702 		field_offset = p - base_ptr;
3703 		if (field_offset & 0x1ff) {
3704 			op_len = be32_to_cpu(ophead->oh_len);
3705 		} else {
3706 			idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3707 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3708 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3709 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3710 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3711 			} else {
3712 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3713 			}
3714 		}
3715 		ptr += sizeof(xlog_op_header_t) + op_len;
3716 	}
3717 }
3718 #endif
3719 
3720 /*
3721  * Perform a forced shutdown on the log.
3722  *
3723  * This can be called from low level log code to trigger a shutdown, or from the
3724  * high level mount shutdown code when the mount shuts down.
3725  *
3726  * Our main objectives here are to make sure that:
3727  *	a. if the shutdown was not due to a log IO error, flush the logs to
3728  *	   disk. Anything modified after this is ignored.
3729  *	b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3730  *	   parties to find out. Nothing new gets queued after this is done.
3731  *	c. Tasks sleeping on log reservations, pinned objects and
3732  *	   other resources get woken up.
3733  *	d. The mount is also marked as shut down so that log triggered shutdowns
3734  *	   still behave the same as if they called xfs_forced_shutdown().
3735  *
3736  * Return true if the shutdown cause was a log IO error and we actually shut the
3737  * log down.
3738  */
3739 bool
xlog_force_shutdown(struct xlog * log,uint32_t shutdown_flags)3740 xlog_force_shutdown(
3741 	struct xlog	*log,
3742 	uint32_t	shutdown_flags)
3743 {
3744 	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3745 
3746 	if (!log)
3747 		return false;
3748 
3749 	/*
3750 	 * Flush all the completed transactions to disk before marking the log
3751 	 * being shut down. We need to do this first as shutting down the log
3752 	 * before the force will prevent the log force from flushing the iclogs
3753 	 * to disk.
3754 	 *
3755 	 * When we are in recovery, there are no transactions to flush, and
3756 	 * we don't want to touch the log because we don't want to perturb the
3757 	 * current head/tail for future recovery attempts. Hence we need to
3758 	 * avoid a log force in this case.
3759 	 *
3760 	 * If we are shutting down due to a log IO error, then we must avoid
3761 	 * trying to write the log as that may just result in more IO errors and
3762 	 * an endless shutdown/force loop.
3763 	 */
3764 	if (!log_error && !xlog_in_recovery(log))
3765 		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3766 
3767 	/*
3768 	 * Atomically set the shutdown state. If the shutdown state is already
3769 	 * set, there someone else is performing the shutdown and so we are done
3770 	 * here. This should never happen because we should only ever get called
3771 	 * once by the first shutdown caller.
3772 	 *
3773 	 * Much of the log state machine transitions assume that shutdown state
3774 	 * cannot change once they hold the log->l_icloglock. Hence we need to
3775 	 * hold that lock here, even though we use the atomic test_and_set_bit()
3776 	 * operation to set the shutdown state.
3777 	 */
3778 	spin_lock(&log->l_icloglock);
3779 	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3780 		spin_unlock(&log->l_icloglock);
3781 		return false;
3782 	}
3783 	spin_unlock(&log->l_icloglock);
3784 
3785 	/*
3786 	 * If this log shutdown also sets the mount shutdown state, issue a
3787 	 * shutdown warning message.
3788 	 */
3789 	if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3790 		xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3791 "Filesystem has been shut down due to log error (0x%x).",
3792 				shutdown_flags);
3793 		xfs_alert(log->l_mp,
3794 "Please unmount the filesystem and rectify the problem(s).");
3795 		if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3796 			xfs_stack_trace();
3797 	}
3798 
3799 	/*
3800 	 * We don't want anybody waiting for log reservations after this. That
3801 	 * means we have to wake up everybody queued up on reserveq as well as
3802 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3803 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3804 	 * action is protected by the grant locks.
3805 	 */
3806 	xlog_grant_head_wake_all(&log->l_reserve_head);
3807 	xlog_grant_head_wake_all(&log->l_write_head);
3808 
3809 	/*
3810 	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3811 	 * as if the log writes were completed. The abort handling in the log
3812 	 * item committed callback functions will do this again under lock to
3813 	 * avoid races.
3814 	 */
3815 	spin_lock(&log->l_cilp->xc_push_lock);
3816 	wake_up_all(&log->l_cilp->xc_start_wait);
3817 	wake_up_all(&log->l_cilp->xc_commit_wait);
3818 	spin_unlock(&log->l_cilp->xc_push_lock);
3819 
3820 	spin_lock(&log->l_icloglock);
3821 	xlog_state_shutdown_callbacks(log);
3822 	spin_unlock(&log->l_icloglock);
3823 
3824 	wake_up_var(&log->l_opstate);
3825 	return log_error;
3826 }
3827 
3828 STATIC int
xlog_iclogs_empty(struct xlog * log)3829 xlog_iclogs_empty(
3830 	struct xlog	*log)
3831 {
3832 	xlog_in_core_t	*iclog;
3833 
3834 	iclog = log->l_iclog;
3835 	do {
3836 		/* endianness does not matter here, zero is zero in
3837 		 * any language.
3838 		 */
3839 		if (iclog->ic_header.h_num_logops)
3840 			return 0;
3841 		iclog = iclog->ic_next;
3842 	} while (iclog != log->l_iclog);
3843 	return 1;
3844 }
3845 
3846 /*
3847  * Verify that an LSN stamped into a piece of metadata is valid. This is
3848  * intended for use in read verifiers on v5 superblocks.
3849  */
3850 bool
xfs_log_check_lsn(struct xfs_mount * mp,xfs_lsn_t lsn)3851 xfs_log_check_lsn(
3852 	struct xfs_mount	*mp,
3853 	xfs_lsn_t		lsn)
3854 {
3855 	struct xlog		*log = mp->m_log;
3856 	bool			valid;
3857 
3858 	/*
3859 	 * norecovery mode skips mount-time log processing and unconditionally
3860 	 * resets the in-core LSN. We can't validate in this mode, but
3861 	 * modifications are not allowed anyways so just return true.
3862 	 */
3863 	if (xfs_has_norecovery(mp))
3864 		return true;
3865 
3866 	/*
3867 	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3868 	 * handled by recovery and thus safe to ignore here.
3869 	 */
3870 	if (lsn == NULLCOMMITLSN)
3871 		return true;
3872 
3873 	valid = xlog_valid_lsn(mp->m_log, lsn);
3874 
3875 	/* warn the user about what's gone wrong before verifier failure */
3876 	if (!valid) {
3877 		spin_lock(&log->l_icloglock);
3878 		xfs_warn(mp,
3879 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3880 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3881 			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3882 			 log->l_curr_cycle, log->l_curr_block);
3883 		spin_unlock(&log->l_icloglock);
3884 	}
3885 
3886 	return valid;
3887 }
3888 
3889 /*
3890  * Notify the log that we're about to start using a feature that is protected
3891  * by a log incompat feature flag.  This will prevent log covering from
3892  * clearing those flags.
3893  */
3894 void
xlog_use_incompat_feat(struct xlog * log)3895 xlog_use_incompat_feat(
3896 	struct xlog		*log)
3897 {
3898 	down_read(&log->l_incompat_users);
3899 }
3900 
3901 /* Notify the log that we've finished using log incompat features. */
3902 void
xlog_drop_incompat_feat(struct xlog * log)3903 xlog_drop_incompat_feat(
3904 	struct xlog		*log)
3905 {
3906 	up_read(&log->l_incompat_users);
3907 }
3908