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 	int		error = 0;
643 	int		min_logfsbs;
644 
645 	if (!xfs_has_norecovery(mp)) {
646 		xfs_notice(mp, "Mounting V%d Filesystem %pU",
647 			   XFS_SB_VERSION_NUM(&mp->m_sb),
648 			   &mp->m_sb.sb_uuid);
649 	} else {
650 		xfs_notice(mp,
651 "Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
652 			   XFS_SB_VERSION_NUM(&mp->m_sb),
653 			   &mp->m_sb.sb_uuid);
654 		ASSERT(xfs_is_readonly(mp));
655 	}
656 
657 	log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
658 	if (IS_ERR(log)) {
659 		error = PTR_ERR(log);
660 		goto out;
661 	}
662 	mp->m_log = log;
663 
664 	/*
665 	 * Now that we have set up the log and it's internal geometry
666 	 * parameters, we can validate the given log space and drop a critical
667 	 * message via syslog if the log size is too small. A log that is too
668 	 * small can lead to unexpected situations in transaction log space
669 	 * reservation stage. The superblock verifier has already validated all
670 	 * the other log geometry constraints, so we don't have to check those
671 	 * here.
672 	 *
673 	 * Note: For v4 filesystems, we can't just reject the mount if the
674 	 * validation fails.  This would mean that people would have to
675 	 * downgrade their kernel just to remedy the situation as there is no
676 	 * way to grow the log (short of black magic surgery with xfs_db).
677 	 *
678 	 * We can, however, reject mounts for V5 format filesystems, as the
679 	 * mkfs binary being used to make the filesystem should never create a
680 	 * filesystem with a log that is too small.
681 	 */
682 	min_logfsbs = xfs_log_calc_minimum_size(mp);
683 	if (mp->m_sb.sb_logblocks < min_logfsbs) {
684 		xfs_warn(mp,
685 		"Log size %d blocks too small, minimum size is %d blocks",
686 			 mp->m_sb.sb_logblocks, min_logfsbs);
687 
688 		/*
689 		 * Log check errors are always fatal on v5; or whenever bad
690 		 * metadata leads to a crash.
691 		 */
692 		if (xfs_has_crc(mp)) {
693 			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
694 			ASSERT(0);
695 			error = -EINVAL;
696 			goto out_free_log;
697 		}
698 		xfs_crit(mp, "Log size out of supported range.");
699 		xfs_crit(mp,
700 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
701 	}
702 
703 	/*
704 	 * Initialize the AIL now we have a log.
705 	 */
706 	error = xfs_trans_ail_init(mp);
707 	if (error) {
708 		xfs_warn(mp, "AIL initialisation failed: error %d", error);
709 		goto out_free_log;
710 	}
711 	log->l_ailp = mp->m_ail;
712 
713 	/*
714 	 * skip log recovery on a norecovery mount.  pretend it all
715 	 * just worked.
716 	 */
717 	if (!xfs_has_norecovery(mp)) {
718 		error = xlog_recover(log);
719 		if (error) {
720 			xfs_warn(mp, "log mount/recovery failed: error %d",
721 				error);
722 			xlog_recover_cancel(log);
723 			goto out_destroy_ail;
724 		}
725 	}
726 
727 	error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
728 			       "log");
729 	if (error)
730 		goto out_destroy_ail;
731 
732 	/* Normal transactions can now occur */
733 	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
734 
735 	/*
736 	 * Now the log has been fully initialised and we know were our
737 	 * space grant counters are, we can initialise the permanent ticket
738 	 * needed for delayed logging to work.
739 	 */
740 	xlog_cil_init_post_recovery(log);
741 
742 	return 0;
743 
744 out_destroy_ail:
745 	xfs_trans_ail_destroy(mp);
746 out_free_log:
747 	xlog_dealloc_log(log);
748 out:
749 	return error;
750 }
751 
752 /*
753  * Finish the recovery of the file system.  This is separate from the
754  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
755  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
756  * here.
757  *
758  * If we finish recovery successfully, start the background log work. If we are
759  * not doing recovery, then we have a RO filesystem and we don't need to start
760  * it.
761  */
762 int
xfs_log_mount_finish(struct xfs_mount * mp)763 xfs_log_mount_finish(
764 	struct xfs_mount	*mp)
765 {
766 	struct xlog		*log = mp->m_log;
767 	int			error = 0;
768 
769 	if (xfs_has_norecovery(mp)) {
770 		ASSERT(xfs_is_readonly(mp));
771 		return 0;
772 	}
773 
774 	/*
775 	 * During the second phase of log recovery, we need iget and
776 	 * iput to behave like they do for an active filesystem.
777 	 * xfs_fs_drop_inode needs to be able to prevent the deletion
778 	 * of inodes before we're done replaying log items on those
779 	 * inodes.  Turn it off immediately after recovery finishes
780 	 * so that we don't leak the quota inodes if subsequent mount
781 	 * activities fail.
782 	 *
783 	 * We let all inodes involved in redo item processing end up on
784 	 * the LRU instead of being evicted immediately so that if we do
785 	 * something to an unlinked inode, the irele won't cause
786 	 * premature truncation and freeing of the inode, which results
787 	 * in log recovery failure.  We have to evict the unreferenced
788 	 * lru inodes after clearing SB_ACTIVE because we don't
789 	 * otherwise clean up the lru if there's a subsequent failure in
790 	 * xfs_mountfs, which leads to us leaking the inodes if nothing
791 	 * else (e.g. quotacheck) references the inodes before the
792 	 * mount failure occurs.
793 	 */
794 	mp->m_super->s_flags |= SB_ACTIVE;
795 	xfs_log_work_queue(mp);
796 	if (xlog_recovery_needed(log))
797 		error = xlog_recover_finish(log);
798 	mp->m_super->s_flags &= ~SB_ACTIVE;
799 	evict_inodes(mp->m_super);
800 
801 	/*
802 	 * Drain the buffer LRU after log recovery. This is required for v4
803 	 * filesystems to avoid leaving around buffers with NULL verifier ops,
804 	 * but we do it unconditionally to make sure we're always in a clean
805 	 * cache state after mount.
806 	 *
807 	 * Don't push in the error case because the AIL may have pending intents
808 	 * that aren't removed until recovery is cancelled.
809 	 */
810 	if (xlog_recovery_needed(log)) {
811 		if (!error) {
812 			xfs_log_force(mp, XFS_LOG_SYNC);
813 			xfs_ail_push_all_sync(mp->m_ail);
814 		}
815 		xfs_notice(mp, "Ending recovery (logdev: %s)",
816 				mp->m_logname ? mp->m_logname : "internal");
817 	} else {
818 		xfs_info(mp, "Ending clean mount");
819 	}
820 	xfs_buftarg_drain(mp->m_ddev_targp);
821 
822 	clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
823 
824 	/* Make sure the log is dead if we're returning failure. */
825 	ASSERT(!error || xlog_is_shutdown(log));
826 
827 	return error;
828 }
829 
830 /*
831  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
832  * the log.
833  */
834 void
xfs_log_mount_cancel(struct xfs_mount * mp)835 xfs_log_mount_cancel(
836 	struct xfs_mount	*mp)
837 {
838 	xlog_recover_cancel(mp->m_log);
839 	xfs_log_unmount(mp);
840 }
841 
842 /*
843  * Flush out the iclog to disk ensuring that device caches are flushed and
844  * the iclog hits stable storage before any completion waiters are woken.
845  */
846 static inline int
xlog_force_iclog(struct xlog_in_core * iclog)847 xlog_force_iclog(
848 	struct xlog_in_core	*iclog)
849 {
850 	atomic_inc(&iclog->ic_refcnt);
851 	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
852 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
853 		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
854 	return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
855 }
856 
857 /*
858  * Cycle all the iclogbuf locks to make sure all log IO completion
859  * is done before we tear down these buffers.
860  */
861 static void
xlog_wait_iclog_completion(struct xlog * log)862 xlog_wait_iclog_completion(struct xlog *log)
863 {
864 	int		i;
865 	struct xlog_in_core	*iclog = log->l_iclog;
866 
867 	for (i = 0; i < log->l_iclog_bufs; i++) {
868 		down(&iclog->ic_sema);
869 		up(&iclog->ic_sema);
870 		iclog = iclog->ic_next;
871 	}
872 }
873 
874 /*
875  * Wait for the iclog and all prior iclogs to be written disk as required by the
876  * log force state machine. Waiting on ic_force_wait ensures iclog completions
877  * have been ordered and callbacks run before we are woken here, hence
878  * guaranteeing that all the iclogs up to this one are on stable storage.
879  */
880 int
xlog_wait_on_iclog(struct xlog_in_core * iclog)881 xlog_wait_on_iclog(
882 	struct xlog_in_core	*iclog)
883 		__releases(iclog->ic_log->l_icloglock)
884 {
885 	struct xlog		*log = iclog->ic_log;
886 
887 	trace_xlog_iclog_wait_on(iclog, _RET_IP_);
888 	if (!xlog_is_shutdown(log) &&
889 	    iclog->ic_state != XLOG_STATE_ACTIVE &&
890 	    iclog->ic_state != XLOG_STATE_DIRTY) {
891 		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
892 		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
893 	} else {
894 		spin_unlock(&log->l_icloglock);
895 	}
896 
897 	if (xlog_is_shutdown(log))
898 		return -EIO;
899 	return 0;
900 }
901 
902 /*
903  * Write out an unmount record using the ticket provided. We have to account for
904  * the data space used in the unmount ticket as this write is not done from a
905  * transaction context that has already done the accounting for us.
906  */
907 static int
xlog_write_unmount_record(struct xlog * log,struct xlog_ticket * ticket)908 xlog_write_unmount_record(
909 	struct xlog		*log,
910 	struct xlog_ticket	*ticket)
911 {
912 	struct  {
913 		struct xlog_op_header ophdr;
914 		struct xfs_unmount_log_format ulf;
915 	} unmount_rec = {
916 		.ophdr = {
917 			.oh_clientid = XFS_LOG,
918 			.oh_tid = cpu_to_be32(ticket->t_tid),
919 			.oh_flags = XLOG_UNMOUNT_TRANS,
920 		},
921 		.ulf = {
922 			.magic = XLOG_UNMOUNT_TYPE,
923 		},
924 	};
925 	struct xfs_log_iovec reg = {
926 		.i_addr = &unmount_rec,
927 		.i_len = sizeof(unmount_rec),
928 		.i_type = XLOG_REG_TYPE_UNMOUNT,
929 	};
930 	struct xfs_log_vec vec = {
931 		.lv_niovecs = 1,
932 		.lv_iovecp = &reg,
933 	};
934 	LIST_HEAD(lv_chain);
935 	list_add(&vec.lv_list, &lv_chain);
936 
937 	BUILD_BUG_ON((sizeof(struct xlog_op_header) +
938 		      sizeof(struct xfs_unmount_log_format)) !=
939 							sizeof(unmount_rec));
940 
941 	/* account for space used by record data */
942 	ticket->t_curr_res -= sizeof(unmount_rec);
943 
944 	return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
945 }
946 
947 /*
948  * Mark the filesystem clean by writing an unmount record to the head of the
949  * log.
950  */
951 static void
xlog_unmount_write(struct xlog * log)952 xlog_unmount_write(
953 	struct xlog		*log)
954 {
955 	struct xfs_mount	*mp = log->l_mp;
956 	struct xlog_in_core	*iclog;
957 	struct xlog_ticket	*tic = NULL;
958 	int			error;
959 
960 	error = xfs_log_reserve(mp, 600, 1, &tic, 0);
961 	if (error)
962 		goto out_err;
963 
964 	error = xlog_write_unmount_record(log, tic);
965 	/*
966 	 * At this point, we're umounting anyway, so there's no point in
967 	 * transitioning log state to shutdown. Just continue...
968 	 */
969 out_err:
970 	if (error)
971 		xfs_alert(mp, "%s: unmount record failed", __func__);
972 
973 	spin_lock(&log->l_icloglock);
974 	iclog = log->l_iclog;
975 	error = xlog_force_iclog(iclog);
976 	xlog_wait_on_iclog(iclog);
977 
978 	if (tic) {
979 		trace_xfs_log_umount_write(log, tic);
980 		xfs_log_ticket_ungrant(log, tic);
981 	}
982 }
983 
984 static void
xfs_log_unmount_verify_iclog(struct xlog * log)985 xfs_log_unmount_verify_iclog(
986 	struct xlog		*log)
987 {
988 	struct xlog_in_core	*iclog = log->l_iclog;
989 
990 	do {
991 		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
992 		ASSERT(iclog->ic_offset == 0);
993 	} while ((iclog = iclog->ic_next) != log->l_iclog);
994 }
995 
996 /*
997  * Unmount record used to have a string "Unmount filesystem--" in the
998  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
999  * We just write the magic number now since that particular field isn't
1000  * currently architecture converted and "Unmount" is a bit foo.
1001  * As far as I know, there weren't any dependencies on the old behaviour.
1002  */
1003 static void
xfs_log_unmount_write(struct xfs_mount * mp)1004 xfs_log_unmount_write(
1005 	struct xfs_mount	*mp)
1006 {
1007 	struct xlog		*log = mp->m_log;
1008 
1009 	if (!xfs_log_writable(mp))
1010 		return;
1011 
1012 	xfs_log_force(mp, XFS_LOG_SYNC);
1013 
1014 	if (xlog_is_shutdown(log))
1015 		return;
1016 
1017 	/*
1018 	 * If we think the summary counters are bad, avoid writing the unmount
1019 	 * record to force log recovery at next mount, after which the summary
1020 	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
1021 	 * more details.
1022 	 */
1023 	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1024 			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1025 		xfs_alert(mp, "%s: will fix summary counters at next mount",
1026 				__func__);
1027 		return;
1028 	}
1029 
1030 	xfs_log_unmount_verify_iclog(log);
1031 	xlog_unmount_write(log);
1032 }
1033 
1034 /*
1035  * Empty the log for unmount/freeze.
1036  *
1037  * To do this, we first need to shut down the background log work so it is not
1038  * trying to cover the log as we clean up. We then need to unpin all objects in
1039  * the log so we can then flush them out. Once they have completed their IO and
1040  * run the callbacks removing themselves from the AIL, we can cover the log.
1041  */
1042 int
xfs_log_quiesce(struct xfs_mount * mp)1043 xfs_log_quiesce(
1044 	struct xfs_mount	*mp)
1045 {
1046 	/*
1047 	 * Clear log incompat features since we're quiescing the log.  Report
1048 	 * failures, though it's not fatal to have a higher log feature
1049 	 * protection level than the log contents actually require.
1050 	 */
1051 	if (xfs_clear_incompat_log_features(mp)) {
1052 		int error;
1053 
1054 		error = xfs_sync_sb(mp, false);
1055 		if (error)
1056 			xfs_warn(mp,
1057 	"Failed to clear log incompat features on quiesce");
1058 	}
1059 
1060 	cancel_delayed_work_sync(&mp->m_log->l_work);
1061 	xfs_log_force(mp, XFS_LOG_SYNC);
1062 
1063 	/*
1064 	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1065 	 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1066 	 * xfs_buf_iowait() cannot be used because it was pushed with the
1067 	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1068 	 * the IO to complete.
1069 	 */
1070 	xfs_ail_push_all_sync(mp->m_ail);
1071 	xfs_buftarg_wait(mp->m_ddev_targp);
1072 	xfs_buf_lock(mp->m_sb_bp);
1073 	xfs_buf_unlock(mp->m_sb_bp);
1074 
1075 	return xfs_log_cover(mp);
1076 }
1077 
1078 void
xfs_log_clean(struct xfs_mount * mp)1079 xfs_log_clean(
1080 	struct xfs_mount	*mp)
1081 {
1082 	xfs_log_quiesce(mp);
1083 	xfs_log_unmount_write(mp);
1084 }
1085 
1086 /*
1087  * Shut down and release the AIL and Log.
1088  *
1089  * During unmount, we need to ensure we flush all the dirty metadata objects
1090  * from the AIL so that the log is empty before we write the unmount record to
1091  * the log. Once this is done, we can tear down the AIL and the log.
1092  */
1093 void
xfs_log_unmount(struct xfs_mount * mp)1094 xfs_log_unmount(
1095 	struct xfs_mount	*mp)
1096 {
1097 	xfs_log_clean(mp);
1098 
1099 	/*
1100 	 * If shutdown has come from iclog IO context, the log
1101 	 * cleaning will have been skipped and so we need to wait
1102 	 * for the iclog to complete shutdown processing before we
1103 	 * tear anything down.
1104 	 */
1105 	xlog_wait_iclog_completion(mp->m_log);
1106 
1107 	xfs_buftarg_drain(mp->m_ddev_targp);
1108 
1109 	xfs_trans_ail_destroy(mp);
1110 
1111 	xfs_sysfs_del(&mp->m_log->l_kobj);
1112 
1113 	xlog_dealloc_log(mp->m_log);
1114 }
1115 
1116 void
xfs_log_item_init(struct xfs_mount * mp,struct xfs_log_item * item,int type,const struct xfs_item_ops * ops)1117 xfs_log_item_init(
1118 	struct xfs_mount	*mp,
1119 	struct xfs_log_item	*item,
1120 	int			type,
1121 	const struct xfs_item_ops *ops)
1122 {
1123 	item->li_log = mp->m_log;
1124 	item->li_ailp = mp->m_ail;
1125 	item->li_type = type;
1126 	item->li_ops = ops;
1127 	item->li_lv = NULL;
1128 
1129 	INIT_LIST_HEAD(&item->li_ail);
1130 	INIT_LIST_HEAD(&item->li_cil);
1131 	INIT_LIST_HEAD(&item->li_bio_list);
1132 	INIT_LIST_HEAD(&item->li_trans);
1133 }
1134 
1135 /*
1136  * Wake up processes waiting for log space after we have moved the log tail.
1137  */
1138 void
xfs_log_space_wake(struct xfs_mount * mp)1139 xfs_log_space_wake(
1140 	struct xfs_mount	*mp)
1141 {
1142 	struct xlog		*log = mp->m_log;
1143 	int			free_bytes;
1144 
1145 	if (xlog_is_shutdown(log))
1146 		return;
1147 
1148 	if (!list_empty_careful(&log->l_write_head.waiters)) {
1149 		ASSERT(!xlog_in_recovery(log));
1150 
1151 		spin_lock(&log->l_write_head.lock);
1152 		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1153 		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1154 		spin_unlock(&log->l_write_head.lock);
1155 	}
1156 
1157 	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1158 		ASSERT(!xlog_in_recovery(log));
1159 
1160 		spin_lock(&log->l_reserve_head.lock);
1161 		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1162 		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1163 		spin_unlock(&log->l_reserve_head.lock);
1164 	}
1165 }
1166 
1167 /*
1168  * Determine if we have a transaction that has gone to disk that needs to be
1169  * covered. To begin the transition to the idle state firstly the log needs to
1170  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1171  * we start attempting to cover the log.
1172  *
1173  * Only if we are then in a state where covering is needed, the caller is
1174  * informed that dummy transactions are required to move the log into the idle
1175  * state.
1176  *
1177  * If there are any items in the AIl or CIL, then we do not want to attempt to
1178  * cover the log as we may be in a situation where there isn't log space
1179  * available to run a dummy transaction and this can lead to deadlocks when the
1180  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1181  * there's no point in running a dummy transaction at this point because we
1182  * can't start trying to idle the log until both the CIL and AIL are empty.
1183  */
1184 static bool
xfs_log_need_covered(struct xfs_mount * mp)1185 xfs_log_need_covered(
1186 	struct xfs_mount	*mp)
1187 {
1188 	struct xlog		*log = mp->m_log;
1189 	bool			needed = false;
1190 
1191 	if (!xlog_cil_empty(log))
1192 		return false;
1193 
1194 	spin_lock(&log->l_icloglock);
1195 	switch (log->l_covered_state) {
1196 	case XLOG_STATE_COVER_DONE:
1197 	case XLOG_STATE_COVER_DONE2:
1198 	case XLOG_STATE_COVER_IDLE:
1199 		break;
1200 	case XLOG_STATE_COVER_NEED:
1201 	case XLOG_STATE_COVER_NEED2:
1202 		if (xfs_ail_min_lsn(log->l_ailp))
1203 			break;
1204 		if (!xlog_iclogs_empty(log))
1205 			break;
1206 
1207 		needed = true;
1208 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1209 			log->l_covered_state = XLOG_STATE_COVER_DONE;
1210 		else
1211 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1212 		break;
1213 	default:
1214 		needed = true;
1215 		break;
1216 	}
1217 	spin_unlock(&log->l_icloglock);
1218 	return needed;
1219 }
1220 
1221 /*
1222  * Explicitly cover the log. This is similar to background log covering but
1223  * intended for usage in quiesce codepaths. The caller is responsible to ensure
1224  * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1225  * must all be empty.
1226  */
1227 static int
xfs_log_cover(struct xfs_mount * mp)1228 xfs_log_cover(
1229 	struct xfs_mount	*mp)
1230 {
1231 	int			error = 0;
1232 	bool			need_covered;
1233 
1234 	ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1235 	        !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1236 		xlog_is_shutdown(mp->m_log));
1237 
1238 	if (!xfs_log_writable(mp))
1239 		return 0;
1240 
1241 	/*
1242 	 * xfs_log_need_covered() is not idempotent because it progresses the
1243 	 * state machine if the log requires covering. Therefore, we must call
1244 	 * this function once and use the result until we've issued an sb sync.
1245 	 * Do so first to make that abundantly clear.
1246 	 *
1247 	 * Fall into the covering sequence if the log needs covering or the
1248 	 * mount has lazy superblock accounting to sync to disk. The sb sync
1249 	 * used for covering accumulates the in-core counters, so covering
1250 	 * handles this for us.
1251 	 */
1252 	need_covered = xfs_log_need_covered(mp);
1253 	if (!need_covered && !xfs_has_lazysbcount(mp))
1254 		return 0;
1255 
1256 	/*
1257 	 * To cover the log, commit the superblock twice (at most) in
1258 	 * independent checkpoints. The first serves as a reference for the
1259 	 * tail pointer. The sync transaction and AIL push empties the AIL and
1260 	 * updates the in-core tail to the LSN of the first checkpoint. The
1261 	 * second commit updates the on-disk tail with the in-core LSN,
1262 	 * covering the log. Push the AIL one more time to leave it empty, as
1263 	 * we found it.
1264 	 */
1265 	do {
1266 		error = xfs_sync_sb(mp, true);
1267 		if (error)
1268 			break;
1269 		xfs_ail_push_all_sync(mp->m_ail);
1270 	} while (xfs_log_need_covered(mp));
1271 
1272 	return error;
1273 }
1274 
1275 /*
1276  * We may be holding the log iclog lock upon entering this routine.
1277  */
1278 xfs_lsn_t
xlog_assign_tail_lsn_locked(struct xfs_mount * mp)1279 xlog_assign_tail_lsn_locked(
1280 	struct xfs_mount	*mp)
1281 {
1282 	struct xlog		*log = mp->m_log;
1283 	struct xfs_log_item	*lip;
1284 	xfs_lsn_t		tail_lsn;
1285 
1286 	assert_spin_locked(&mp->m_ail->ail_lock);
1287 
1288 	/*
1289 	 * To make sure we always have a valid LSN for the log tail we keep
1290 	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1291 	 * and use that when the AIL was empty.
1292 	 */
1293 	lip = xfs_ail_min(mp->m_ail);
1294 	if (lip)
1295 		tail_lsn = lip->li_lsn;
1296 	else
1297 		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1298 	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1299 	atomic64_set(&log->l_tail_lsn, tail_lsn);
1300 	return tail_lsn;
1301 }
1302 
1303 xfs_lsn_t
xlog_assign_tail_lsn(struct xfs_mount * mp)1304 xlog_assign_tail_lsn(
1305 	struct xfs_mount	*mp)
1306 {
1307 	xfs_lsn_t		tail_lsn;
1308 
1309 	spin_lock(&mp->m_ail->ail_lock);
1310 	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1311 	spin_unlock(&mp->m_ail->ail_lock);
1312 
1313 	return tail_lsn;
1314 }
1315 
1316 /*
1317  * Return the space in the log between the tail and the head.  The head
1318  * is passed in the cycle/bytes formal parms.  In the special case where
1319  * the reserve head has wrapped passed the tail, this calculation is no
1320  * longer valid.  In this case, just return 0 which means there is no space
1321  * in the log.  This works for all places where this function is called
1322  * with the reserve head.  Of course, if the write head were to ever
1323  * wrap the tail, we should blow up.  Rather than catch this case here,
1324  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1325  *
1326  * If reservation head is behind the tail, we have a problem. Warn about it,
1327  * but then treat it as if the log is empty.
1328  *
1329  * If the log is shut down, the head and tail may be invalid or out of whack, so
1330  * shortcut invalidity asserts in this case so that we don't trigger them
1331  * falsely.
1332  */
1333 STATIC int
xlog_space_left(struct xlog * log,atomic64_t * head)1334 xlog_space_left(
1335 	struct xlog	*log,
1336 	atomic64_t	*head)
1337 {
1338 	int		tail_bytes;
1339 	int		tail_cycle;
1340 	int		head_cycle;
1341 	int		head_bytes;
1342 
1343 	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1344 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1345 	tail_bytes = BBTOB(tail_bytes);
1346 	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1347 		return log->l_logsize - (head_bytes - tail_bytes);
1348 	if (tail_cycle + 1 < head_cycle)
1349 		return 0;
1350 
1351 	/* Ignore potential inconsistency when shutdown. */
1352 	if (xlog_is_shutdown(log))
1353 		return log->l_logsize;
1354 
1355 	if (tail_cycle < head_cycle) {
1356 		ASSERT(tail_cycle == (head_cycle - 1));
1357 		return tail_bytes - head_bytes;
1358 	}
1359 
1360 	/*
1361 	 * The reservation head is behind the tail. In this case we just want to
1362 	 * return the size of the log as the amount of space left.
1363 	 */
1364 	xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1365 	xfs_alert(log->l_mp, "  tail_cycle = %d, tail_bytes = %d",
1366 		  tail_cycle, tail_bytes);
1367 	xfs_alert(log->l_mp, "  GH   cycle = %d, GH   bytes = %d",
1368 		  head_cycle, head_bytes);
1369 	ASSERT(0);
1370 	return log->l_logsize;
1371 }
1372 
1373 
1374 static void
xlog_ioend_work(struct work_struct * work)1375 xlog_ioend_work(
1376 	struct work_struct	*work)
1377 {
1378 	struct xlog_in_core     *iclog =
1379 		container_of(work, struct xlog_in_core, ic_end_io_work);
1380 	struct xlog		*log = iclog->ic_log;
1381 	int			error;
1382 
1383 	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1384 #ifdef DEBUG
1385 	/* treat writes with injected CRC errors as failed */
1386 	if (iclog->ic_fail_crc)
1387 		error = -EIO;
1388 #endif
1389 
1390 	/*
1391 	 * Race to shutdown the filesystem if we see an error.
1392 	 */
1393 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1394 		xfs_alert(log->l_mp, "log I/O error %d", error);
1395 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1396 	}
1397 
1398 	xlog_state_done_syncing(iclog);
1399 	bio_uninit(&iclog->ic_bio);
1400 
1401 	/*
1402 	 * Drop the lock to signal that we are done. Nothing references the
1403 	 * iclog after this, so an unmount waiting on this lock can now tear it
1404 	 * down safely. As such, it is unsafe to reference the iclog after the
1405 	 * unlock as we could race with it being freed.
1406 	 */
1407 	up(&iclog->ic_sema);
1408 }
1409 
1410 /*
1411  * Return size of each in-core log record buffer.
1412  *
1413  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1414  *
1415  * If the filesystem blocksize is too large, we may need to choose a
1416  * larger size since the directory code currently logs entire blocks.
1417  */
1418 STATIC void
xlog_get_iclog_buffer_size(struct xfs_mount * mp,struct xlog * log)1419 xlog_get_iclog_buffer_size(
1420 	struct xfs_mount	*mp,
1421 	struct xlog		*log)
1422 {
1423 	if (mp->m_logbufs <= 0)
1424 		mp->m_logbufs = XLOG_MAX_ICLOGS;
1425 	if (mp->m_logbsize <= 0)
1426 		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1427 
1428 	log->l_iclog_bufs = mp->m_logbufs;
1429 	log->l_iclog_size = mp->m_logbsize;
1430 
1431 	/*
1432 	 * # headers = size / 32k - one header holds cycles from 32k of data.
1433 	 */
1434 	log->l_iclog_heads =
1435 		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1436 	log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1437 }
1438 
1439 void
xfs_log_work_queue(struct xfs_mount * mp)1440 xfs_log_work_queue(
1441 	struct xfs_mount        *mp)
1442 {
1443 	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1444 				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1445 }
1446 
1447 /*
1448  * Clear the log incompat flags if we have the opportunity.
1449  *
1450  * This only happens if we're about to log the second dummy transaction as part
1451  * of covering the log and we can get the log incompat feature usage lock.
1452  */
1453 static inline void
xlog_clear_incompat(struct xlog * log)1454 xlog_clear_incompat(
1455 	struct xlog		*log)
1456 {
1457 	struct xfs_mount	*mp = log->l_mp;
1458 
1459 	if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1460 				XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1461 		return;
1462 
1463 	if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1464 		return;
1465 
1466 	if (!down_write_trylock(&log->l_incompat_users))
1467 		return;
1468 
1469 	xfs_clear_incompat_log_features(mp);
1470 	up_write(&log->l_incompat_users);
1471 }
1472 
1473 /*
1474  * Every sync period we need to unpin all items in the AIL and push them to
1475  * disk. If there is nothing dirty, then we might need to cover the log to
1476  * indicate that the filesystem is idle.
1477  */
1478 static void
xfs_log_worker(struct work_struct * work)1479 xfs_log_worker(
1480 	struct work_struct	*work)
1481 {
1482 	struct xlog		*log = container_of(to_delayed_work(work),
1483 						struct xlog, l_work);
1484 	struct xfs_mount	*mp = log->l_mp;
1485 
1486 	/* dgc: errors ignored - not fatal and nowhere to report them */
1487 	if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1488 		/*
1489 		 * Dump a transaction into the log that contains no real change.
1490 		 * This is needed to stamp the current tail LSN into the log
1491 		 * during the covering operation.
1492 		 *
1493 		 * We cannot use an inode here for this - that will push dirty
1494 		 * state back up into the VFS and then periodic inode flushing
1495 		 * will prevent log covering from making progress. Hence we
1496 		 * synchronously log the superblock instead to ensure the
1497 		 * superblock is immediately unpinned and can be written back.
1498 		 */
1499 		xlog_clear_incompat(log);
1500 		xfs_sync_sb(mp, true);
1501 	} else
1502 		xfs_log_force(mp, 0);
1503 
1504 	/* start pushing all the metadata that is currently dirty */
1505 	xfs_ail_push_all(mp->m_ail);
1506 
1507 	/* queue us up again */
1508 	xfs_log_work_queue(mp);
1509 }
1510 
1511 /*
1512  * This routine initializes some of the log structure for a given mount point.
1513  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1514  * some other stuff may be filled in too.
1515  */
1516 STATIC struct xlog *
xlog_alloc_log(struct xfs_mount * mp,struct xfs_buftarg * log_target,xfs_daddr_t blk_offset,int num_bblks)1517 xlog_alloc_log(
1518 	struct xfs_mount	*mp,
1519 	struct xfs_buftarg	*log_target,
1520 	xfs_daddr_t		blk_offset,
1521 	int			num_bblks)
1522 {
1523 	struct xlog		*log;
1524 	xlog_rec_header_t	*head;
1525 	xlog_in_core_t		**iclogp;
1526 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1527 	int			i;
1528 	int			error = -ENOMEM;
1529 	uint			log2_size = 0;
1530 
1531 	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1532 	if (!log) {
1533 		xfs_warn(mp, "Log allocation failed: No memory!");
1534 		goto out;
1535 	}
1536 
1537 	log->l_mp	   = mp;
1538 	log->l_targ	   = log_target;
1539 	log->l_logsize     = BBTOB(num_bblks);
1540 	log->l_logBBstart  = blk_offset;
1541 	log->l_logBBsize   = num_bblks;
1542 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1543 	set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1544 	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1545 
1546 	log->l_prev_block  = -1;
1547 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1548 	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1549 	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1550 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1551 
1552 	if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1553 		log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1554 	else
1555 		log->l_iclog_roundoff = BBSIZE;
1556 
1557 	xlog_grant_head_init(&log->l_reserve_head);
1558 	xlog_grant_head_init(&log->l_write_head);
1559 
1560 	error = -EFSCORRUPTED;
1561 	if (xfs_has_sector(mp)) {
1562 	        log2_size = mp->m_sb.sb_logsectlog;
1563 		if (log2_size < BBSHIFT) {
1564 			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1565 				log2_size, BBSHIFT);
1566 			goto out_free_log;
1567 		}
1568 
1569 	        log2_size -= BBSHIFT;
1570 		if (log2_size > mp->m_sectbb_log) {
1571 			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1572 				log2_size, mp->m_sectbb_log);
1573 			goto out_free_log;
1574 		}
1575 
1576 		/* for larger sector sizes, must have v2 or external log */
1577 		if (log2_size && log->l_logBBstart > 0 &&
1578 			    !xfs_has_logv2(mp)) {
1579 			xfs_warn(mp,
1580 		"log sector size (0x%x) invalid for configuration.",
1581 				log2_size);
1582 			goto out_free_log;
1583 		}
1584 	}
1585 	log->l_sectBBsize = 1 << log2_size;
1586 
1587 	init_rwsem(&log->l_incompat_users);
1588 
1589 	xlog_get_iclog_buffer_size(mp, log);
1590 
1591 	spin_lock_init(&log->l_icloglock);
1592 	init_waitqueue_head(&log->l_flush_wait);
1593 
1594 	iclogp = &log->l_iclog;
1595 	/*
1596 	 * The amount of memory to allocate for the iclog structure is
1597 	 * rather funky due to the way the structure is defined.  It is
1598 	 * done this way so that we can use different sizes for machines
1599 	 * with different amounts of memory.  See the definition of
1600 	 * xlog_in_core_t in xfs_log_priv.h for details.
1601 	 */
1602 	ASSERT(log->l_iclog_size >= 4096);
1603 	for (i = 0; i < log->l_iclog_bufs; i++) {
1604 		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1605 				sizeof(struct bio_vec);
1606 
1607 		iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1608 		if (!iclog)
1609 			goto out_free_iclog;
1610 
1611 		*iclogp = iclog;
1612 		iclog->ic_prev = prev_iclog;
1613 		prev_iclog = iclog;
1614 
1615 		iclog->ic_data = kvzalloc(log->l_iclog_size,
1616 				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1617 		if (!iclog->ic_data)
1618 			goto out_free_iclog;
1619 		head = &iclog->ic_header;
1620 		memset(head, 0, sizeof(xlog_rec_header_t));
1621 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1622 		head->h_version = cpu_to_be32(
1623 			xfs_has_logv2(log->l_mp) ? 2 : 1);
1624 		head->h_size = cpu_to_be32(log->l_iclog_size);
1625 		/* new fields */
1626 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1627 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1628 
1629 		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1630 		iclog->ic_state = XLOG_STATE_ACTIVE;
1631 		iclog->ic_log = log;
1632 		atomic_set(&iclog->ic_refcnt, 0);
1633 		INIT_LIST_HEAD(&iclog->ic_callbacks);
1634 		iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1635 
1636 		init_waitqueue_head(&iclog->ic_force_wait);
1637 		init_waitqueue_head(&iclog->ic_write_wait);
1638 		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1639 		sema_init(&iclog->ic_sema, 1);
1640 
1641 		iclogp = &iclog->ic_next;
1642 	}
1643 	*iclogp = log->l_iclog;			/* complete ring */
1644 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1645 
1646 	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1647 			XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1648 				    WQ_HIGHPRI),
1649 			0, mp->m_super->s_id);
1650 	if (!log->l_ioend_workqueue)
1651 		goto out_free_iclog;
1652 
1653 	error = xlog_cil_init(log);
1654 	if (error)
1655 		goto out_destroy_workqueue;
1656 	return log;
1657 
1658 out_destroy_workqueue:
1659 	destroy_workqueue(log->l_ioend_workqueue);
1660 out_free_iclog:
1661 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1662 		prev_iclog = iclog->ic_next;
1663 		kmem_free(iclog->ic_data);
1664 		kmem_free(iclog);
1665 		if (prev_iclog == log->l_iclog)
1666 			break;
1667 	}
1668 out_free_log:
1669 	kmem_free(log);
1670 out:
1671 	return ERR_PTR(error);
1672 }	/* xlog_alloc_log */
1673 
1674 /*
1675  * Compute the LSN that we'd need to push the log tail towards in order to have
1676  * (a) enough on-disk log space to log the number of bytes specified, (b) at
1677  * least 25% of the log space free, and (c) at least 256 blocks free.  If the
1678  * log free space already meets all three thresholds, this function returns
1679  * NULLCOMMITLSN.
1680  */
1681 xfs_lsn_t
xlog_grant_push_threshold(struct xlog * log,int need_bytes)1682 xlog_grant_push_threshold(
1683 	struct xlog	*log,
1684 	int		need_bytes)
1685 {
1686 	xfs_lsn_t	threshold_lsn = 0;
1687 	xfs_lsn_t	last_sync_lsn;
1688 	int		free_blocks;
1689 	int		free_bytes;
1690 	int		threshold_block;
1691 	int		threshold_cycle;
1692 	int		free_threshold;
1693 
1694 	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1695 
1696 	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1697 	free_blocks = BTOBBT(free_bytes);
1698 
1699 	/*
1700 	 * Set the threshold for the minimum number of free blocks in the
1701 	 * log to the maximum of what the caller needs, one quarter of the
1702 	 * log, and 256 blocks.
1703 	 */
1704 	free_threshold = BTOBB(need_bytes);
1705 	free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1706 	free_threshold = max(free_threshold, 256);
1707 	if (free_blocks >= free_threshold)
1708 		return NULLCOMMITLSN;
1709 
1710 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1711 						&threshold_block);
1712 	threshold_block += free_threshold;
1713 	if (threshold_block >= log->l_logBBsize) {
1714 		threshold_block -= log->l_logBBsize;
1715 		threshold_cycle += 1;
1716 	}
1717 	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1718 					threshold_block);
1719 	/*
1720 	 * Don't pass in an lsn greater than the lsn of the last
1721 	 * log record known to be on disk. Use a snapshot of the last sync lsn
1722 	 * so that it doesn't change between the compare and the set.
1723 	 */
1724 	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1725 	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1726 		threshold_lsn = last_sync_lsn;
1727 
1728 	return threshold_lsn;
1729 }
1730 
1731 /*
1732  * Push the tail of the log if we need to do so to maintain the free log space
1733  * thresholds set out by xlog_grant_push_threshold.  We may need to adopt a
1734  * policy which pushes on an lsn which is further along in the log once we
1735  * reach the high water mark.  In this manner, we would be creating a low water
1736  * mark.
1737  */
1738 STATIC void
xlog_grant_push_ail(struct xlog * log,int need_bytes)1739 xlog_grant_push_ail(
1740 	struct xlog	*log,
1741 	int		need_bytes)
1742 {
1743 	xfs_lsn_t	threshold_lsn;
1744 
1745 	threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1746 	if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1747 		return;
1748 
1749 	/*
1750 	 * Get the transaction layer to kick the dirty buffers out to
1751 	 * disk asynchronously. No point in trying to do this if
1752 	 * the filesystem is shutting down.
1753 	 */
1754 	xfs_ail_push(log->l_ailp, threshold_lsn);
1755 }
1756 
1757 /*
1758  * Stamp cycle number in every block
1759  */
1760 STATIC void
xlog_pack_data(struct xlog * log,struct xlog_in_core * iclog,int roundoff)1761 xlog_pack_data(
1762 	struct xlog		*log,
1763 	struct xlog_in_core	*iclog,
1764 	int			roundoff)
1765 {
1766 	int			i, j, k;
1767 	int			size = iclog->ic_offset + roundoff;
1768 	__be32			cycle_lsn;
1769 	char			*dp;
1770 
1771 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1772 
1773 	dp = iclog->ic_datap;
1774 	for (i = 0; i < BTOBB(size); i++) {
1775 		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1776 			break;
1777 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1778 		*(__be32 *)dp = cycle_lsn;
1779 		dp += BBSIZE;
1780 	}
1781 
1782 	if (xfs_has_logv2(log->l_mp)) {
1783 		xlog_in_core_2_t *xhdr = iclog->ic_data;
1784 
1785 		for ( ; i < BTOBB(size); i++) {
1786 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1787 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1788 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1789 			*(__be32 *)dp = cycle_lsn;
1790 			dp += BBSIZE;
1791 		}
1792 
1793 		for (i = 1; i < log->l_iclog_heads; i++)
1794 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1795 	}
1796 }
1797 
1798 /*
1799  * Calculate the checksum for a log buffer.
1800  *
1801  * This is a little more complicated than it should be because the various
1802  * headers and the actual data are non-contiguous.
1803  */
1804 __le32
xlog_cksum(struct xlog * log,struct xlog_rec_header * rhead,char * dp,int size)1805 xlog_cksum(
1806 	struct xlog		*log,
1807 	struct xlog_rec_header	*rhead,
1808 	char			*dp,
1809 	int			size)
1810 {
1811 	uint32_t		crc;
1812 
1813 	/* first generate the crc for the record header ... */
1814 	crc = xfs_start_cksum_update((char *)rhead,
1815 			      sizeof(struct xlog_rec_header),
1816 			      offsetof(struct xlog_rec_header, h_crc));
1817 
1818 	/* ... then for additional cycle data for v2 logs ... */
1819 	if (xfs_has_logv2(log->l_mp)) {
1820 		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1821 		int		i;
1822 		int		xheads;
1823 
1824 		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1825 
1826 		for (i = 1; i < xheads; i++) {
1827 			crc = crc32c(crc, &xhdr[i].hic_xheader,
1828 				     sizeof(struct xlog_rec_ext_header));
1829 		}
1830 	}
1831 
1832 	/* ... and finally for the payload */
1833 	crc = crc32c(crc, dp, size);
1834 
1835 	return xfs_end_cksum(crc);
1836 }
1837 
1838 static void
xlog_bio_end_io(struct bio * bio)1839 xlog_bio_end_io(
1840 	struct bio		*bio)
1841 {
1842 	struct xlog_in_core	*iclog = bio->bi_private;
1843 
1844 	queue_work(iclog->ic_log->l_ioend_workqueue,
1845 		   &iclog->ic_end_io_work);
1846 }
1847 
1848 static int
xlog_map_iclog_data(struct bio * bio,void * data,size_t count)1849 xlog_map_iclog_data(
1850 	struct bio		*bio,
1851 	void			*data,
1852 	size_t			count)
1853 {
1854 	do {
1855 		struct page	*page = kmem_to_page(data);
1856 		unsigned int	off = offset_in_page(data);
1857 		size_t		len = min_t(size_t, count, PAGE_SIZE - off);
1858 
1859 		if (bio_add_page(bio, page, len, off) != len)
1860 			return -EIO;
1861 
1862 		data += len;
1863 		count -= len;
1864 	} while (count);
1865 
1866 	return 0;
1867 }
1868 
1869 STATIC void
xlog_write_iclog(struct xlog * log,struct xlog_in_core * iclog,uint64_t bno,unsigned int count)1870 xlog_write_iclog(
1871 	struct xlog		*log,
1872 	struct xlog_in_core	*iclog,
1873 	uint64_t		bno,
1874 	unsigned int		count)
1875 {
1876 	ASSERT(bno < log->l_logBBsize);
1877 	trace_xlog_iclog_write(iclog, _RET_IP_);
1878 
1879 	/*
1880 	 * We lock the iclogbufs here so that we can serialise against I/O
1881 	 * completion during unmount.  We might be processing a shutdown
1882 	 * triggered during unmount, and that can occur asynchronously to the
1883 	 * unmount thread, and hence we need to ensure that completes before
1884 	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1885 	 * across the log IO to archieve that.
1886 	 */
1887 	down(&iclog->ic_sema);
1888 	if (xlog_is_shutdown(log)) {
1889 		/*
1890 		 * It would seem logical to return EIO here, but we rely on
1891 		 * the log state machine to propagate I/O errors instead of
1892 		 * doing it here.  We kick of the state machine and unlock
1893 		 * the buffer manually, the code needs to be kept in sync
1894 		 * with the I/O completion path.
1895 		 */
1896 		goto sync;
1897 	}
1898 
1899 	/*
1900 	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1901 	 * IOs coming immediately after this one. This prevents the block layer
1902 	 * writeback throttle from throttling log writes behind background
1903 	 * metadata writeback and causing priority inversions.
1904 	 */
1905 	bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1906 		 howmany(count, PAGE_SIZE),
1907 		 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1908 	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1909 	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1910 	iclog->ic_bio.bi_private = iclog;
1911 
1912 	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1913 		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1914 		/*
1915 		 * For external log devices, we also need to flush the data
1916 		 * device cache first to ensure all metadata writeback covered
1917 		 * by the LSN in this iclog is on stable storage. This is slow,
1918 		 * but it *must* complete before we issue the external log IO.
1919 		 *
1920 		 * If the flush fails, we cannot conclude that past metadata
1921 		 * writeback from the log succeeded.  Repeating the flush is
1922 		 * not possible, hence we must shut down with log IO error to
1923 		 * avoid shutdown re-entering this path and erroring out again.
1924 		 */
1925 		if (log->l_targ != log->l_mp->m_ddev_targp &&
1926 		    blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev))
1927 			goto shutdown;
1928 	}
1929 	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1930 		iclog->ic_bio.bi_opf |= REQ_FUA;
1931 
1932 	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1933 
1934 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count))
1935 		goto shutdown;
1936 
1937 	if (is_vmalloc_addr(iclog->ic_data))
1938 		flush_kernel_vmap_range(iclog->ic_data, count);
1939 
1940 	/*
1941 	 * If this log buffer would straddle the end of the log we will have
1942 	 * to split it up into two bios, so that we can continue at the start.
1943 	 */
1944 	if (bno + BTOBB(count) > log->l_logBBsize) {
1945 		struct bio *split;
1946 
1947 		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1948 				  GFP_NOIO, &fs_bio_set);
1949 		bio_chain(split, &iclog->ic_bio);
1950 		submit_bio(split);
1951 
1952 		/* restart at logical offset zero for the remainder */
1953 		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1954 	}
1955 
1956 	submit_bio(&iclog->ic_bio);
1957 	return;
1958 shutdown:
1959 	xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1960 sync:
1961 	xlog_state_done_syncing(iclog);
1962 	up(&iclog->ic_sema);
1963 }
1964 
1965 /*
1966  * We need to bump cycle number for the part of the iclog that is
1967  * written to the start of the log. Watch out for the header magic
1968  * number case, though.
1969  */
1970 static void
xlog_split_iclog(struct xlog * log,void * data,uint64_t bno,unsigned int count)1971 xlog_split_iclog(
1972 	struct xlog		*log,
1973 	void			*data,
1974 	uint64_t		bno,
1975 	unsigned int		count)
1976 {
1977 	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1978 	unsigned int		i;
1979 
1980 	for (i = split_offset; i < count; i += BBSIZE) {
1981 		uint32_t cycle = get_unaligned_be32(data + i);
1982 
1983 		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1984 			cycle++;
1985 		put_unaligned_be32(cycle, data + i);
1986 	}
1987 }
1988 
1989 static int
xlog_calc_iclog_size(struct xlog * log,struct xlog_in_core * iclog,uint32_t * roundoff)1990 xlog_calc_iclog_size(
1991 	struct xlog		*log,
1992 	struct xlog_in_core	*iclog,
1993 	uint32_t		*roundoff)
1994 {
1995 	uint32_t		count_init, count;
1996 
1997 	/* Add for LR header */
1998 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1999 	count = roundup(count_init, log->l_iclog_roundoff);
2000 
2001 	*roundoff = count - count_init;
2002 
2003 	ASSERT(count >= count_init);
2004 	ASSERT(*roundoff < log->l_iclog_roundoff);
2005 	return count;
2006 }
2007 
2008 /*
2009  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2010  * fashion.  Previously, we should have moved the current iclog
2011  * ptr in the log to point to the next available iclog.  This allows further
2012  * write to continue while this code syncs out an iclog ready to go.
2013  * Before an in-core log can be written out, the data section must be scanned
2014  * to save away the 1st word of each BBSIZE block into the header.  We replace
2015  * it with the current cycle count.  Each BBSIZE block is tagged with the
2016  * cycle count because there in an implicit assumption that drives will
2017  * guarantee that entire 512 byte blocks get written at once.  In other words,
2018  * we can't have part of a 512 byte block written and part not written.  By
2019  * tagging each block, we will know which blocks are valid when recovering
2020  * after an unclean shutdown.
2021  *
2022  * This routine is single threaded on the iclog.  No other thread can be in
2023  * this routine with the same iclog.  Changing contents of iclog can there-
2024  * fore be done without grabbing the state machine lock.  Updating the global
2025  * log will require grabbing the lock though.
2026  *
2027  * The entire log manager uses a logical block numbering scheme.  Only
2028  * xlog_write_iclog knows about the fact that the log may not start with
2029  * block zero on a given device.
2030  */
2031 STATIC void
xlog_sync(struct xlog * log,struct xlog_in_core * iclog,struct xlog_ticket * ticket)2032 xlog_sync(
2033 	struct xlog		*log,
2034 	struct xlog_in_core	*iclog,
2035 	struct xlog_ticket	*ticket)
2036 {
2037 	unsigned int		count;		/* byte count of bwrite */
2038 	unsigned int		roundoff;       /* roundoff to BB or stripe */
2039 	uint64_t		bno;
2040 	unsigned int		size;
2041 
2042 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2043 	trace_xlog_iclog_sync(iclog, _RET_IP_);
2044 
2045 	count = xlog_calc_iclog_size(log, iclog, &roundoff);
2046 
2047 	/*
2048 	 * If we have a ticket, account for the roundoff via the ticket
2049 	 * reservation to avoid touching the hot grant heads needlessly.
2050 	 * Otherwise, we have to move grant heads directly.
2051 	 */
2052 	if (ticket) {
2053 		ticket->t_curr_res -= roundoff;
2054 	} else {
2055 		xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2056 		xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2057 	}
2058 
2059 	/* put cycle number in every block */
2060 	xlog_pack_data(log, iclog, roundoff);
2061 
2062 	/* real byte length */
2063 	size = iclog->ic_offset;
2064 	if (xfs_has_logv2(log->l_mp))
2065 		size += roundoff;
2066 	iclog->ic_header.h_len = cpu_to_be32(size);
2067 
2068 	XFS_STATS_INC(log->l_mp, xs_log_writes);
2069 	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2070 
2071 	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2072 
2073 	/* Do we need to split this write into 2 parts? */
2074 	if (bno + BTOBB(count) > log->l_logBBsize)
2075 		xlog_split_iclog(log, &iclog->ic_header, bno, count);
2076 
2077 	/* calculcate the checksum */
2078 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2079 					    iclog->ic_datap, size);
2080 	/*
2081 	 * Intentionally corrupt the log record CRC based on the error injection
2082 	 * frequency, if defined. This facilitates testing log recovery in the
2083 	 * event of torn writes. Hence, set the IOABORT state to abort the log
2084 	 * write on I/O completion and shutdown the fs. The subsequent mount
2085 	 * detects the bad CRC and attempts to recover.
2086 	 */
2087 #ifdef DEBUG
2088 	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2089 		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2090 		iclog->ic_fail_crc = true;
2091 		xfs_warn(log->l_mp,
2092 	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2093 			 be64_to_cpu(iclog->ic_header.h_lsn));
2094 	}
2095 #endif
2096 	xlog_verify_iclog(log, iclog, count);
2097 	xlog_write_iclog(log, iclog, bno, count);
2098 }
2099 
2100 /*
2101  * Deallocate a log structure
2102  */
2103 STATIC void
xlog_dealloc_log(struct xlog * log)2104 xlog_dealloc_log(
2105 	struct xlog	*log)
2106 {
2107 	xlog_in_core_t	*iclog, *next_iclog;
2108 	int		i;
2109 
2110 	/*
2111 	 * Destroy the CIL after waiting for iclog IO completion because an
2112 	 * iclog EIO error will try to shut down the log, which accesses the
2113 	 * CIL to wake up the waiters.
2114 	 */
2115 	xlog_cil_destroy(log);
2116 
2117 	iclog = log->l_iclog;
2118 	for (i = 0; i < log->l_iclog_bufs; i++) {
2119 		next_iclog = iclog->ic_next;
2120 		kmem_free(iclog->ic_data);
2121 		kmem_free(iclog);
2122 		iclog = next_iclog;
2123 	}
2124 
2125 	log->l_mp->m_log = NULL;
2126 	destroy_workqueue(log->l_ioend_workqueue);
2127 	kmem_free(log);
2128 }
2129 
2130 /*
2131  * Update counters atomically now that memcpy is done.
2132  */
2133 static inline void
xlog_state_finish_copy(struct xlog * log,struct xlog_in_core * iclog,int record_cnt,int copy_bytes)2134 xlog_state_finish_copy(
2135 	struct xlog		*log,
2136 	struct xlog_in_core	*iclog,
2137 	int			record_cnt,
2138 	int			copy_bytes)
2139 {
2140 	lockdep_assert_held(&log->l_icloglock);
2141 
2142 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2143 	iclog->ic_offset += copy_bytes;
2144 }
2145 
2146 /*
2147  * print out info relating to regions written which consume
2148  * the reservation
2149  */
2150 void
xlog_print_tic_res(struct xfs_mount * mp,struct xlog_ticket * ticket)2151 xlog_print_tic_res(
2152 	struct xfs_mount	*mp,
2153 	struct xlog_ticket	*ticket)
2154 {
2155 	xfs_warn(mp, "ticket reservation summary:");
2156 	xfs_warn(mp, "  unit res    = %d bytes", ticket->t_unit_res);
2157 	xfs_warn(mp, "  current res = %d bytes", ticket->t_curr_res);
2158 	xfs_warn(mp, "  original count  = %d", ticket->t_ocnt);
2159 	xfs_warn(mp, "  remaining count = %d", ticket->t_cnt);
2160 }
2161 
2162 /*
2163  * Print a summary of the transaction.
2164  */
2165 void
xlog_print_trans(struct xfs_trans * tp)2166 xlog_print_trans(
2167 	struct xfs_trans	*tp)
2168 {
2169 	struct xfs_mount	*mp = tp->t_mountp;
2170 	struct xfs_log_item	*lip;
2171 
2172 	/* dump core transaction and ticket info */
2173 	xfs_warn(mp, "transaction summary:");
2174 	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2175 	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2176 	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2177 
2178 	xlog_print_tic_res(mp, tp->t_ticket);
2179 
2180 	/* dump each log item */
2181 	list_for_each_entry(lip, &tp->t_items, li_trans) {
2182 		struct xfs_log_vec	*lv = lip->li_lv;
2183 		struct xfs_log_iovec	*vec;
2184 		int			i;
2185 
2186 		xfs_warn(mp, "log item: ");
2187 		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2188 		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
2189 		if (!lv)
2190 			continue;
2191 		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2192 		xfs_warn(mp, "  size	= %d", lv->lv_size);
2193 		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2194 		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2195 
2196 		/* dump each iovec for the log item */
2197 		vec = lv->lv_iovecp;
2198 		for (i = 0; i < lv->lv_niovecs; i++) {
2199 			int dumplen = min(vec->i_len, 32);
2200 
2201 			xfs_warn(mp, "  iovec[%d]", i);
2202 			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2203 			xfs_warn(mp, "    len	= %d", vec->i_len);
2204 			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2205 			xfs_hex_dump(vec->i_addr, dumplen);
2206 
2207 			vec++;
2208 		}
2209 	}
2210 }
2211 
2212 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)2213 xlog_write_iovec(
2214 	struct xlog_in_core	*iclog,
2215 	uint32_t		*log_offset,
2216 	void			*data,
2217 	uint32_t		write_len,
2218 	int			*bytes_left,
2219 	uint32_t		*record_cnt,
2220 	uint32_t		*data_cnt)
2221 {
2222 	ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2223 	ASSERT(*log_offset % sizeof(int32_t) == 0);
2224 	ASSERT(write_len % sizeof(int32_t) == 0);
2225 
2226 	memcpy(iclog->ic_datap + *log_offset, data, write_len);
2227 	*log_offset += write_len;
2228 	*bytes_left -= write_len;
2229 	(*record_cnt)++;
2230 	*data_cnt += write_len;
2231 }
2232 
2233 /*
2234  * Write log vectors into a single iclog which is guaranteed by the caller
2235  * to have enough space to write the entire log vector into.
2236  */
2237 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)2238 xlog_write_full(
2239 	struct xfs_log_vec	*lv,
2240 	struct xlog_ticket	*ticket,
2241 	struct xlog_in_core	*iclog,
2242 	uint32_t		*log_offset,
2243 	uint32_t		*len,
2244 	uint32_t		*record_cnt,
2245 	uint32_t		*data_cnt)
2246 {
2247 	int			index;
2248 
2249 	ASSERT(*log_offset + *len <= iclog->ic_size ||
2250 		iclog->ic_state == XLOG_STATE_WANT_SYNC);
2251 
2252 	/*
2253 	 * Ordered log vectors have no regions to write so this
2254 	 * loop will naturally skip them.
2255 	 */
2256 	for (index = 0; index < lv->lv_niovecs; index++) {
2257 		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2258 		struct xlog_op_header	*ophdr = reg->i_addr;
2259 
2260 		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2261 		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2262 				reg->i_len, len, record_cnt, data_cnt);
2263 	}
2264 }
2265 
2266 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)2267 xlog_write_get_more_iclog_space(
2268 	struct xlog_ticket	*ticket,
2269 	struct xlog_in_core	**iclogp,
2270 	uint32_t		*log_offset,
2271 	uint32_t		len,
2272 	uint32_t		*record_cnt,
2273 	uint32_t		*data_cnt)
2274 {
2275 	struct xlog_in_core	*iclog = *iclogp;
2276 	struct xlog		*log = iclog->ic_log;
2277 	int			error;
2278 
2279 	spin_lock(&log->l_icloglock);
2280 	ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2281 	xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2282 	error = xlog_state_release_iclog(log, iclog, ticket);
2283 	spin_unlock(&log->l_icloglock);
2284 	if (error)
2285 		return error;
2286 
2287 	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2288 					log_offset);
2289 	if (error)
2290 		return error;
2291 	*record_cnt = 0;
2292 	*data_cnt = 0;
2293 	*iclogp = iclog;
2294 	return 0;
2295 }
2296 
2297 /*
2298  * Write log vectors into a single iclog which is smaller than the current chain
2299  * length. We write until we cannot fit a full record into the remaining space
2300  * and then stop. We return the log vector that is to be written that cannot
2301  * wholly fit in the iclog.
2302  */
2303 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)2304 xlog_write_partial(
2305 	struct xfs_log_vec	*lv,
2306 	struct xlog_ticket	*ticket,
2307 	struct xlog_in_core	**iclogp,
2308 	uint32_t		*log_offset,
2309 	uint32_t		*len,
2310 	uint32_t		*record_cnt,
2311 	uint32_t		*data_cnt)
2312 {
2313 	struct xlog_in_core	*iclog = *iclogp;
2314 	struct xlog_op_header	*ophdr;
2315 	int			index = 0;
2316 	uint32_t		rlen;
2317 	int			error;
2318 
2319 	/* walk the logvec, copying until we run out of space in the iclog */
2320 	for (index = 0; index < lv->lv_niovecs; index++) {
2321 		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2322 		uint32_t		reg_offset = 0;
2323 
2324 		/*
2325 		 * The first region of a continuation must have a non-zero
2326 		 * length otherwise log recovery will just skip over it and
2327 		 * start recovering from the next opheader it finds. Because we
2328 		 * mark the next opheader as a continuation, recovery will then
2329 		 * incorrectly add the continuation to the previous region and
2330 		 * that breaks stuff.
2331 		 *
2332 		 * Hence if there isn't space for region data after the
2333 		 * opheader, then we need to start afresh with a new iclog.
2334 		 */
2335 		if (iclog->ic_size - *log_offset <=
2336 					sizeof(struct xlog_op_header)) {
2337 			error = xlog_write_get_more_iclog_space(ticket,
2338 					&iclog, log_offset, *len, record_cnt,
2339 					data_cnt);
2340 			if (error)
2341 				return error;
2342 		}
2343 
2344 		ophdr = reg->i_addr;
2345 		rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2346 
2347 		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2348 		ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2349 		if (rlen != reg->i_len)
2350 			ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2351 
2352 		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2353 				rlen, len, record_cnt, data_cnt);
2354 
2355 		/* If we wrote the whole region, move to the next. */
2356 		if (rlen == reg->i_len)
2357 			continue;
2358 
2359 		/*
2360 		 * We now have a partially written iovec, but it can span
2361 		 * multiple iclogs so we loop here. First we release the iclog
2362 		 * we currently have, then we get a new iclog and add a new
2363 		 * opheader. Then we continue copying from where we were until
2364 		 * we either complete the iovec or fill the iclog. If we
2365 		 * complete the iovec, then we increment the index and go right
2366 		 * back to the top of the outer loop. if we fill the iclog, we
2367 		 * run the inner loop again.
2368 		 *
2369 		 * This is complicated by the tail of a region using all the
2370 		 * space in an iclog and hence requiring us to release the iclog
2371 		 * and get a new one before returning to the outer loop. We must
2372 		 * always guarantee that we exit this inner loop with at least
2373 		 * space for log transaction opheaders left in the current
2374 		 * iclog, hence we cannot just terminate the loop at the end
2375 		 * of the of the continuation. So we loop while there is no
2376 		 * space left in the current iclog, and check for the end of the
2377 		 * continuation after getting a new iclog.
2378 		 */
2379 		do {
2380 			/*
2381 			 * Ensure we include the continuation opheader in the
2382 			 * space we need in the new iclog by adding that size
2383 			 * to the length we require. This continuation opheader
2384 			 * needs to be accounted to the ticket as the space it
2385 			 * consumes hasn't been accounted to the lv we are
2386 			 * writing.
2387 			 */
2388 			error = xlog_write_get_more_iclog_space(ticket,
2389 					&iclog, log_offset,
2390 					*len + sizeof(struct xlog_op_header),
2391 					record_cnt, data_cnt);
2392 			if (error)
2393 				return error;
2394 
2395 			ophdr = iclog->ic_datap + *log_offset;
2396 			ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2397 			ophdr->oh_clientid = XFS_TRANSACTION;
2398 			ophdr->oh_res2 = 0;
2399 			ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2400 
2401 			ticket->t_curr_res -= sizeof(struct xlog_op_header);
2402 			*log_offset += sizeof(struct xlog_op_header);
2403 			*data_cnt += sizeof(struct xlog_op_header);
2404 
2405 			/*
2406 			 * If rlen fits in the iclog, then end the region
2407 			 * continuation. Otherwise we're going around again.
2408 			 */
2409 			reg_offset += rlen;
2410 			rlen = reg->i_len - reg_offset;
2411 			if (rlen <= iclog->ic_size - *log_offset)
2412 				ophdr->oh_flags |= XLOG_END_TRANS;
2413 			else
2414 				ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2415 
2416 			rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2417 			ophdr->oh_len = cpu_to_be32(rlen);
2418 
2419 			xlog_write_iovec(iclog, log_offset,
2420 					reg->i_addr + reg_offset,
2421 					rlen, len, record_cnt, data_cnt);
2422 
2423 		} while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2424 	}
2425 
2426 	/*
2427 	 * No more iovecs remain in this logvec so return the next log vec to
2428 	 * the caller so it can go back to fast path copying.
2429 	 */
2430 	*iclogp = iclog;
2431 	return 0;
2432 }
2433 
2434 /*
2435  * Write some region out to in-core log
2436  *
2437  * This will be called when writing externally provided regions or when
2438  * writing out a commit record for a given transaction.
2439  *
2440  * General algorithm:
2441  *	1. Find total length of this write.  This may include adding to the
2442  *		lengths passed in.
2443  *	2. Check whether we violate the tickets reservation.
2444  *	3. While writing to this iclog
2445  *	    A. Reserve as much space in this iclog as can get
2446  *	    B. If this is first write, save away start lsn
2447  *	    C. While writing this region:
2448  *		1. If first write of transaction, write start record
2449  *		2. Write log operation header (header per region)
2450  *		3. Find out if we can fit entire region into this iclog
2451  *		4. Potentially, verify destination memcpy ptr
2452  *		5. Memcpy (partial) region
2453  *		6. If partial copy, release iclog; otherwise, continue
2454  *			copying more regions into current iclog
2455  *	4. Mark want sync bit (in simulation mode)
2456  *	5. Release iclog for potential flush to on-disk log.
2457  *
2458  * ERRORS:
2459  * 1.	Panic if reservation is overrun.  This should never happen since
2460  *	reservation amounts are generated internal to the filesystem.
2461  * NOTES:
2462  * 1. Tickets are single threaded data structures.
2463  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2464  *	syncing routine.  When a single log_write region needs to span
2465  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2466  *	on all log operation writes which don't contain the end of the
2467  *	region.  The XLOG_END_TRANS bit is used for the in-core log
2468  *	operation which contains the end of the continued log_write region.
2469  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2470  *	we don't really know exactly how much space will be used.  As a result,
2471  *	we don't update ic_offset until the end when we know exactly how many
2472  *	bytes have been written out.
2473  */
2474 int
xlog_write(struct xlog * log,struct xfs_cil_ctx * ctx,struct list_head * lv_chain,struct xlog_ticket * ticket,uint32_t len)2475 xlog_write(
2476 	struct xlog		*log,
2477 	struct xfs_cil_ctx	*ctx,
2478 	struct list_head	*lv_chain,
2479 	struct xlog_ticket	*ticket,
2480 	uint32_t		len)
2481 
2482 {
2483 	struct xlog_in_core	*iclog = NULL;
2484 	struct xfs_log_vec	*lv;
2485 	uint32_t		record_cnt = 0;
2486 	uint32_t		data_cnt = 0;
2487 	int			error = 0;
2488 	int			log_offset;
2489 
2490 	if (ticket->t_curr_res < 0) {
2491 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2492 		     "ctx ticket reservation ran out. Need to up reservation");
2493 		xlog_print_tic_res(log->l_mp, ticket);
2494 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2495 	}
2496 
2497 	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2498 					   &log_offset);
2499 	if (error)
2500 		return error;
2501 
2502 	ASSERT(log_offset <= iclog->ic_size - 1);
2503 
2504 	/*
2505 	 * If we have a context pointer, pass it the first iclog we are
2506 	 * writing to so it can record state needed for iclog write
2507 	 * ordering.
2508 	 */
2509 	if (ctx)
2510 		xlog_cil_set_ctx_write_state(ctx, iclog);
2511 
2512 	list_for_each_entry(lv, lv_chain, lv_list) {
2513 		/*
2514 		 * If the entire log vec does not fit in the iclog, punt it to
2515 		 * the partial copy loop which can handle this case.
2516 		 */
2517 		if (lv->lv_niovecs &&
2518 		    lv->lv_bytes > iclog->ic_size - log_offset) {
2519 			error = xlog_write_partial(lv, ticket, &iclog,
2520 					&log_offset, &len, &record_cnt,
2521 					&data_cnt);
2522 			if (error) {
2523 				/*
2524 				 * We have no iclog to release, so just return
2525 				 * the error immediately.
2526 				 */
2527 				return error;
2528 			}
2529 		} else {
2530 			xlog_write_full(lv, ticket, iclog, &log_offset,
2531 					 &len, &record_cnt, &data_cnt);
2532 		}
2533 	}
2534 	ASSERT(len == 0);
2535 
2536 	/*
2537 	 * We've already been guaranteed that the last writes will fit inside
2538 	 * the current iclog, and hence it will already have the space used by
2539 	 * those writes accounted to it. Hence we do not need to update the
2540 	 * iclog with the number of bytes written here.
2541 	 */
2542 	spin_lock(&log->l_icloglock);
2543 	xlog_state_finish_copy(log, iclog, record_cnt, 0);
2544 	error = xlog_state_release_iclog(log, iclog, ticket);
2545 	spin_unlock(&log->l_icloglock);
2546 
2547 	return error;
2548 }
2549 
2550 static void
xlog_state_activate_iclog(struct xlog_in_core * iclog,int * iclogs_changed)2551 xlog_state_activate_iclog(
2552 	struct xlog_in_core	*iclog,
2553 	int			*iclogs_changed)
2554 {
2555 	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2556 	trace_xlog_iclog_activate(iclog, _RET_IP_);
2557 
2558 	/*
2559 	 * If the number of ops in this iclog indicate it just contains the
2560 	 * dummy transaction, we can change state into IDLE (the second time
2561 	 * around). Otherwise we should change the state into NEED a dummy.
2562 	 * We don't need to cover the dummy.
2563 	 */
2564 	if (*iclogs_changed == 0 &&
2565 	    iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2566 		*iclogs_changed = 1;
2567 	} else {
2568 		/*
2569 		 * We have two dirty iclogs so start over.  This could also be
2570 		 * num of ops indicating this is not the dummy going out.
2571 		 */
2572 		*iclogs_changed = 2;
2573 	}
2574 
2575 	iclog->ic_state	= XLOG_STATE_ACTIVE;
2576 	iclog->ic_offset = 0;
2577 	iclog->ic_header.h_num_logops = 0;
2578 	memset(iclog->ic_header.h_cycle_data, 0,
2579 		sizeof(iclog->ic_header.h_cycle_data));
2580 	iclog->ic_header.h_lsn = 0;
2581 	iclog->ic_header.h_tail_lsn = 0;
2582 }
2583 
2584 /*
2585  * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2586  * ACTIVE after iclog I/O has completed.
2587  */
2588 static void
xlog_state_activate_iclogs(struct xlog * log,int * iclogs_changed)2589 xlog_state_activate_iclogs(
2590 	struct xlog		*log,
2591 	int			*iclogs_changed)
2592 {
2593 	struct xlog_in_core	*iclog = log->l_iclog;
2594 
2595 	do {
2596 		if (iclog->ic_state == XLOG_STATE_DIRTY)
2597 			xlog_state_activate_iclog(iclog, iclogs_changed);
2598 		/*
2599 		 * The ordering of marking iclogs ACTIVE must be maintained, so
2600 		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2601 		 */
2602 		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2603 			break;
2604 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2605 }
2606 
2607 static int
xlog_covered_state(int prev_state,int iclogs_changed)2608 xlog_covered_state(
2609 	int			prev_state,
2610 	int			iclogs_changed)
2611 {
2612 	/*
2613 	 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2614 	 * wrote the first covering record (DONE). We go to IDLE if we just
2615 	 * wrote the second covering record (DONE2) and remain in IDLE until a
2616 	 * non-covering write occurs.
2617 	 */
2618 	switch (prev_state) {
2619 	case XLOG_STATE_COVER_IDLE:
2620 		if (iclogs_changed == 1)
2621 			return XLOG_STATE_COVER_IDLE;
2622 		fallthrough;
2623 	case XLOG_STATE_COVER_NEED:
2624 	case XLOG_STATE_COVER_NEED2:
2625 		break;
2626 	case XLOG_STATE_COVER_DONE:
2627 		if (iclogs_changed == 1)
2628 			return XLOG_STATE_COVER_NEED2;
2629 		break;
2630 	case XLOG_STATE_COVER_DONE2:
2631 		if (iclogs_changed == 1)
2632 			return XLOG_STATE_COVER_IDLE;
2633 		break;
2634 	default:
2635 		ASSERT(0);
2636 	}
2637 
2638 	return XLOG_STATE_COVER_NEED;
2639 }
2640 
2641 STATIC void
xlog_state_clean_iclog(struct xlog * log,struct xlog_in_core * dirty_iclog)2642 xlog_state_clean_iclog(
2643 	struct xlog		*log,
2644 	struct xlog_in_core	*dirty_iclog)
2645 {
2646 	int			iclogs_changed = 0;
2647 
2648 	trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2649 
2650 	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2651 
2652 	xlog_state_activate_iclogs(log, &iclogs_changed);
2653 	wake_up_all(&dirty_iclog->ic_force_wait);
2654 
2655 	if (iclogs_changed) {
2656 		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2657 				iclogs_changed);
2658 	}
2659 }
2660 
2661 STATIC xfs_lsn_t
xlog_get_lowest_lsn(struct xlog * log)2662 xlog_get_lowest_lsn(
2663 	struct xlog		*log)
2664 {
2665 	struct xlog_in_core	*iclog = log->l_iclog;
2666 	xfs_lsn_t		lowest_lsn = 0, lsn;
2667 
2668 	do {
2669 		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2670 		    iclog->ic_state == XLOG_STATE_DIRTY)
2671 			continue;
2672 
2673 		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2674 		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2675 			lowest_lsn = lsn;
2676 	} while ((iclog = iclog->ic_next) != log->l_iclog);
2677 
2678 	return lowest_lsn;
2679 }
2680 
2681 /*
2682  * Completion of a iclog IO does not imply that a transaction has completed, as
2683  * transactions can be large enough to span many iclogs. We cannot change the
2684  * tail of the log half way through a transaction as this may be the only
2685  * transaction in the log and moving the tail to point to the middle of it
2686  * will prevent recovery from finding the start of the transaction. Hence we
2687  * should only update the last_sync_lsn if this iclog contains transaction
2688  * completion callbacks on it.
2689  *
2690  * We have to do this before we drop the icloglock to ensure we are the only one
2691  * that can update it.
2692  *
2693  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2694  * the reservation grant head pushing. This is due to the fact that the push
2695  * target is bound by the current last_sync_lsn value. Hence if we have a large
2696  * amount of log space bound up in this committing transaction then the
2697  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2698  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2699  * should push the AIL to ensure the push target (and hence the grant head) is
2700  * no longer bound by the old log head location and can move forwards and make
2701  * progress again.
2702  */
2703 static void
xlog_state_set_callback(struct xlog * log,struct xlog_in_core * iclog,xfs_lsn_t header_lsn)2704 xlog_state_set_callback(
2705 	struct xlog		*log,
2706 	struct xlog_in_core	*iclog,
2707 	xfs_lsn_t		header_lsn)
2708 {
2709 	trace_xlog_iclog_callback(iclog, _RET_IP_);
2710 	iclog->ic_state = XLOG_STATE_CALLBACK;
2711 
2712 	ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2713 			   header_lsn) <= 0);
2714 
2715 	if (list_empty_careful(&iclog->ic_callbacks))
2716 		return;
2717 
2718 	atomic64_set(&log->l_last_sync_lsn, header_lsn);
2719 	xlog_grant_push_ail(log, 0);
2720 }
2721 
2722 /*
2723  * Return true if we need to stop processing, false to continue to the next
2724  * iclog. The caller will need to run callbacks if the iclog is returned in the
2725  * XLOG_STATE_CALLBACK state.
2726  */
2727 static bool
xlog_state_iodone_process_iclog(struct xlog * log,struct xlog_in_core * iclog)2728 xlog_state_iodone_process_iclog(
2729 	struct xlog		*log,
2730 	struct xlog_in_core	*iclog)
2731 {
2732 	xfs_lsn_t		lowest_lsn;
2733 	xfs_lsn_t		header_lsn;
2734 
2735 	switch (iclog->ic_state) {
2736 	case XLOG_STATE_ACTIVE:
2737 	case XLOG_STATE_DIRTY:
2738 		/*
2739 		 * Skip all iclogs in the ACTIVE & DIRTY states:
2740 		 */
2741 		return false;
2742 	case XLOG_STATE_DONE_SYNC:
2743 		/*
2744 		 * Now that we have an iclog that is in the DONE_SYNC state, do
2745 		 * one more check here to see if we have chased our tail around.
2746 		 * If this is not the lowest lsn iclog, then we will leave it
2747 		 * for another completion to process.
2748 		 */
2749 		header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2750 		lowest_lsn = xlog_get_lowest_lsn(log);
2751 		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2752 			return false;
2753 		xlog_state_set_callback(log, iclog, header_lsn);
2754 		return false;
2755 	default:
2756 		/*
2757 		 * Can only perform callbacks in order.  Since this iclog is not
2758 		 * in the DONE_SYNC state, we skip the rest and just try to
2759 		 * clean up.
2760 		 */
2761 		return true;
2762 	}
2763 }
2764 
2765 /*
2766  * Loop over all the iclogs, running attached callbacks on them. Return true if
2767  * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2768  * to handle transient shutdown state here at all because
2769  * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2770  * cleanup of the callbacks.
2771  */
2772 static bool
xlog_state_do_iclog_callbacks(struct xlog * log)2773 xlog_state_do_iclog_callbacks(
2774 	struct xlog		*log)
2775 		__releases(&log->l_icloglock)
2776 		__acquires(&log->l_icloglock)
2777 {
2778 	struct xlog_in_core	*first_iclog = log->l_iclog;
2779 	struct xlog_in_core	*iclog = first_iclog;
2780 	bool			ran_callback = false;
2781 
2782 	do {
2783 		LIST_HEAD(cb_list);
2784 
2785 		if (xlog_state_iodone_process_iclog(log, iclog))
2786 			break;
2787 		if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2788 			iclog = iclog->ic_next;
2789 			continue;
2790 		}
2791 		list_splice_init(&iclog->ic_callbacks, &cb_list);
2792 		spin_unlock(&log->l_icloglock);
2793 
2794 		trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2795 		xlog_cil_process_committed(&cb_list);
2796 		trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2797 		ran_callback = true;
2798 
2799 		spin_lock(&log->l_icloglock);
2800 		xlog_state_clean_iclog(log, iclog);
2801 		iclog = iclog->ic_next;
2802 	} while (iclog != first_iclog);
2803 
2804 	return ran_callback;
2805 }
2806 
2807 
2808 /*
2809  * Loop running iclog completion callbacks until there are no more iclogs in a
2810  * state that can run callbacks.
2811  */
2812 STATIC void
xlog_state_do_callback(struct xlog * log)2813 xlog_state_do_callback(
2814 	struct xlog		*log)
2815 {
2816 	int			flushcnt = 0;
2817 	int			repeats = 0;
2818 
2819 	spin_lock(&log->l_icloglock);
2820 	while (xlog_state_do_iclog_callbacks(log)) {
2821 		if (xlog_is_shutdown(log))
2822 			break;
2823 
2824 		if (++repeats > 5000) {
2825 			flushcnt += repeats;
2826 			repeats = 0;
2827 			xfs_warn(log->l_mp,
2828 				"%s: possible infinite loop (%d iterations)",
2829 				__func__, flushcnt);
2830 		}
2831 	}
2832 
2833 	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2834 		wake_up_all(&log->l_flush_wait);
2835 
2836 	spin_unlock(&log->l_icloglock);
2837 }
2838 
2839 
2840 /*
2841  * Finish transitioning this iclog to the dirty state.
2842  *
2843  * Callbacks could take time, so they are done outside the scope of the
2844  * global state machine log lock.
2845  */
2846 STATIC void
xlog_state_done_syncing(struct xlog_in_core * iclog)2847 xlog_state_done_syncing(
2848 	struct xlog_in_core	*iclog)
2849 {
2850 	struct xlog		*log = iclog->ic_log;
2851 
2852 	spin_lock(&log->l_icloglock);
2853 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2854 	trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2855 
2856 	/*
2857 	 * If we got an error, either on the first buffer, or in the case of
2858 	 * split log writes, on the second, we shut down the file system and
2859 	 * no iclogs should ever be attempted to be written to disk again.
2860 	 */
2861 	if (!xlog_is_shutdown(log)) {
2862 		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2863 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2864 	}
2865 
2866 	/*
2867 	 * Someone could be sleeping prior to writing out the next
2868 	 * iclog buffer, we wake them all, one will get to do the
2869 	 * I/O, the others get to wait for the result.
2870 	 */
2871 	wake_up_all(&iclog->ic_write_wait);
2872 	spin_unlock(&log->l_icloglock);
2873 	xlog_state_do_callback(log);
2874 }
2875 
2876 /*
2877  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2878  * sleep.  We wait on the flush queue on the head iclog as that should be
2879  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2880  * we will wait here and all new writes will sleep until a sync completes.
2881  *
2882  * The in-core logs are used in a circular fashion. They are not used
2883  * out-of-order even when an iclog past the head is free.
2884  *
2885  * return:
2886  *	* log_offset where xlog_write() can start writing into the in-core
2887  *		log's data space.
2888  *	* in-core log pointer to which xlog_write() should write.
2889  *	* boolean indicating this is a continued write to an in-core log.
2890  *		If this is the last write, then the in-core log's offset field
2891  *		needs to be incremented, depending on the amount of data which
2892  *		is copied.
2893  */
2894 STATIC int
xlog_state_get_iclog_space(struct xlog * log,int len,struct xlog_in_core ** iclogp,struct xlog_ticket * ticket,int * logoffsetp)2895 xlog_state_get_iclog_space(
2896 	struct xlog		*log,
2897 	int			len,
2898 	struct xlog_in_core	**iclogp,
2899 	struct xlog_ticket	*ticket,
2900 	int			*logoffsetp)
2901 {
2902 	int		  log_offset;
2903 	xlog_rec_header_t *head;
2904 	xlog_in_core_t	  *iclog;
2905 
2906 restart:
2907 	spin_lock(&log->l_icloglock);
2908 	if (xlog_is_shutdown(log)) {
2909 		spin_unlock(&log->l_icloglock);
2910 		return -EIO;
2911 	}
2912 
2913 	iclog = log->l_iclog;
2914 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2915 		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2916 
2917 		/* Wait for log writes to have flushed */
2918 		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2919 		goto restart;
2920 	}
2921 
2922 	head = &iclog->ic_header;
2923 
2924 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2925 	log_offset = iclog->ic_offset;
2926 
2927 	trace_xlog_iclog_get_space(iclog, _RET_IP_);
2928 
2929 	/* On the 1st write to an iclog, figure out lsn.  This works
2930 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2931 	 * committing to.  If the offset is set, that's how many blocks
2932 	 * must be written.
2933 	 */
2934 	if (log_offset == 0) {
2935 		ticket->t_curr_res -= log->l_iclog_hsize;
2936 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2937 		head->h_lsn = cpu_to_be64(
2938 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2939 		ASSERT(log->l_curr_block >= 0);
2940 	}
2941 
2942 	/* If there is enough room to write everything, then do it.  Otherwise,
2943 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2944 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2945 	 * until you know exactly how many bytes get copied.  Therefore, wait
2946 	 * until later to update ic_offset.
2947 	 *
2948 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2949 	 * can fit into remaining data section.
2950 	 */
2951 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2952 		int		error = 0;
2953 
2954 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2955 
2956 		/*
2957 		 * If we are the only one writing to this iclog, sync it to
2958 		 * disk.  We need to do an atomic compare and decrement here to
2959 		 * avoid racing with concurrent atomic_dec_and_lock() calls in
2960 		 * xlog_state_release_iclog() when there is more than one
2961 		 * reference to the iclog.
2962 		 */
2963 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2964 			error = xlog_state_release_iclog(log, iclog, ticket);
2965 		spin_unlock(&log->l_icloglock);
2966 		if (error)
2967 			return error;
2968 		goto restart;
2969 	}
2970 
2971 	/* Do we have enough room to write the full amount in the remainder
2972 	 * of this iclog?  Or must we continue a write on the next iclog and
2973 	 * mark this iclog as completely taken?  In the case where we switch
2974 	 * iclogs (to mark it taken), this particular iclog will release/sync
2975 	 * to disk in xlog_write().
2976 	 */
2977 	if (len <= iclog->ic_size - iclog->ic_offset)
2978 		iclog->ic_offset += len;
2979 	else
2980 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2981 	*iclogp = iclog;
2982 
2983 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2984 	spin_unlock(&log->l_icloglock);
2985 
2986 	*logoffsetp = log_offset;
2987 	return 0;
2988 }
2989 
2990 /*
2991  * The first cnt-1 times a ticket goes through here we don't need to move the
2992  * grant write head because the permanent reservation has reserved cnt times the
2993  * unit amount.  Release part of current permanent unit reservation and reset
2994  * current reservation to be one units worth.  Also move grant reservation head
2995  * forward.
2996  */
2997 void
xfs_log_ticket_regrant(struct xlog * log,struct xlog_ticket * ticket)2998 xfs_log_ticket_regrant(
2999 	struct xlog		*log,
3000 	struct xlog_ticket	*ticket)
3001 {
3002 	trace_xfs_log_ticket_regrant(log, ticket);
3003 
3004 	if (ticket->t_cnt > 0)
3005 		ticket->t_cnt--;
3006 
3007 	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3008 					ticket->t_curr_res);
3009 	xlog_grant_sub_space(log, &log->l_write_head.grant,
3010 					ticket->t_curr_res);
3011 	ticket->t_curr_res = ticket->t_unit_res;
3012 
3013 	trace_xfs_log_ticket_regrant_sub(log, ticket);
3014 
3015 	/* just return if we still have some of the pre-reserved space */
3016 	if (!ticket->t_cnt) {
3017 		xlog_grant_add_space(log, &log->l_reserve_head.grant,
3018 				     ticket->t_unit_res);
3019 		trace_xfs_log_ticket_regrant_exit(log, ticket);
3020 
3021 		ticket->t_curr_res = ticket->t_unit_res;
3022 	}
3023 
3024 	xfs_log_ticket_put(ticket);
3025 }
3026 
3027 /*
3028  * Give back the space left from a reservation.
3029  *
3030  * All the information we need to make a correct determination of space left
3031  * is present.  For non-permanent reservations, things are quite easy.  The
3032  * count should have been decremented to zero.  We only need to deal with the
3033  * space remaining in the current reservation part of the ticket.  If the
3034  * ticket contains a permanent reservation, there may be left over space which
3035  * needs to be released.  A count of N means that N-1 refills of the current
3036  * reservation can be done before we need to ask for more space.  The first
3037  * one goes to fill up the first current reservation.  Once we run out of
3038  * space, the count will stay at zero and the only space remaining will be
3039  * in the current reservation field.
3040  */
3041 void
xfs_log_ticket_ungrant(struct xlog * log,struct xlog_ticket * ticket)3042 xfs_log_ticket_ungrant(
3043 	struct xlog		*log,
3044 	struct xlog_ticket	*ticket)
3045 {
3046 	int			bytes;
3047 
3048 	trace_xfs_log_ticket_ungrant(log, ticket);
3049 
3050 	if (ticket->t_cnt > 0)
3051 		ticket->t_cnt--;
3052 
3053 	trace_xfs_log_ticket_ungrant_sub(log, ticket);
3054 
3055 	/*
3056 	 * If this is a permanent reservation ticket, we may be able to free
3057 	 * up more space based on the remaining count.
3058 	 */
3059 	bytes = ticket->t_curr_res;
3060 	if (ticket->t_cnt > 0) {
3061 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3062 		bytes += ticket->t_unit_res*ticket->t_cnt;
3063 	}
3064 
3065 	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3066 	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3067 
3068 	trace_xfs_log_ticket_ungrant_exit(log, ticket);
3069 
3070 	xfs_log_space_wake(log->l_mp);
3071 	xfs_log_ticket_put(ticket);
3072 }
3073 
3074 /*
3075  * This routine will mark the current iclog in the ring as WANT_SYNC and move
3076  * the current iclog pointer to the next iclog in the ring.
3077  */
3078 void
xlog_state_switch_iclogs(struct xlog * log,struct xlog_in_core * iclog,int eventual_size)3079 xlog_state_switch_iclogs(
3080 	struct xlog		*log,
3081 	struct xlog_in_core	*iclog,
3082 	int			eventual_size)
3083 {
3084 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3085 	assert_spin_locked(&log->l_icloglock);
3086 	trace_xlog_iclog_switch(iclog, _RET_IP_);
3087 
3088 	if (!eventual_size)
3089 		eventual_size = iclog->ic_offset;
3090 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3091 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3092 	log->l_prev_block = log->l_curr_block;
3093 	log->l_prev_cycle = log->l_curr_cycle;
3094 
3095 	/* roll log?: ic_offset changed later */
3096 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3097 
3098 	/* Round up to next log-sunit */
3099 	if (log->l_iclog_roundoff > BBSIZE) {
3100 		uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3101 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3102 	}
3103 
3104 	if (log->l_curr_block >= log->l_logBBsize) {
3105 		/*
3106 		 * Rewind the current block before the cycle is bumped to make
3107 		 * sure that the combined LSN never transiently moves forward
3108 		 * when the log wraps to the next cycle. This is to support the
3109 		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3110 		 * other cases should acquire l_icloglock.
3111 		 */
3112 		log->l_curr_block -= log->l_logBBsize;
3113 		ASSERT(log->l_curr_block >= 0);
3114 		smp_wmb();
3115 		log->l_curr_cycle++;
3116 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3117 			log->l_curr_cycle++;
3118 	}
3119 	ASSERT(iclog == log->l_iclog);
3120 	log->l_iclog = iclog->ic_next;
3121 }
3122 
3123 /*
3124  * Force the iclog to disk and check if the iclog has been completed before
3125  * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3126  * pmem) or fast async storage because we drop the icloglock to issue the IO.
3127  * If completion has already occurred, tell the caller so that it can avoid an
3128  * unnecessary wait on the iclog.
3129  */
3130 static int
xlog_force_and_check_iclog(struct xlog_in_core * iclog,bool * completed)3131 xlog_force_and_check_iclog(
3132 	struct xlog_in_core	*iclog,
3133 	bool			*completed)
3134 {
3135 	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3136 	int			error;
3137 
3138 	*completed = false;
3139 	error = xlog_force_iclog(iclog);
3140 	if (error)
3141 		return error;
3142 
3143 	/*
3144 	 * If the iclog has already been completed and reused the header LSN
3145 	 * will have been rewritten by completion
3146 	 */
3147 	if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3148 		*completed = true;
3149 	return 0;
3150 }
3151 
3152 /*
3153  * Write out all data in the in-core log as of this exact moment in time.
3154  *
3155  * Data may be written to the in-core log during this call.  However,
3156  * we don't guarantee this data will be written out.  A change from past
3157  * implementation means this routine will *not* write out zero length LRs.
3158  *
3159  * Basically, we try and perform an intelligent scan of the in-core logs.
3160  * If we determine there is no flushable data, we just return.  There is no
3161  * flushable data if:
3162  *
3163  *	1. the current iclog is active and has no data; the previous iclog
3164  *		is in the active or dirty state.
3165  *	2. the current iclog is drity, and the previous iclog is in the
3166  *		active or dirty state.
3167  *
3168  * We may sleep if:
3169  *
3170  *	1. the current iclog is not in the active nor dirty state.
3171  *	2. the current iclog dirty, and the previous iclog is not in the
3172  *		active nor dirty state.
3173  *	3. the current iclog is active, and there is another thread writing
3174  *		to this particular iclog.
3175  *	4. a) the current iclog is active and has no other writers
3176  *	   b) when we return from flushing out this iclog, it is still
3177  *		not in the active nor dirty state.
3178  */
3179 int
xfs_log_force(struct xfs_mount * mp,uint flags)3180 xfs_log_force(
3181 	struct xfs_mount	*mp,
3182 	uint			flags)
3183 {
3184 	struct xlog		*log = mp->m_log;
3185 	struct xlog_in_core	*iclog;
3186 
3187 	XFS_STATS_INC(mp, xs_log_force);
3188 	trace_xfs_log_force(mp, 0, _RET_IP_);
3189 
3190 	xlog_cil_force(log);
3191 
3192 	spin_lock(&log->l_icloglock);
3193 	if (xlog_is_shutdown(log))
3194 		goto out_error;
3195 
3196 	iclog = log->l_iclog;
3197 	trace_xlog_iclog_force(iclog, _RET_IP_);
3198 
3199 	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3200 	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3201 	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3202 		/*
3203 		 * If the head is dirty or (active and empty), then we need to
3204 		 * look at the previous iclog.
3205 		 *
3206 		 * If the previous iclog is active or dirty we are done.  There
3207 		 * is nothing to sync out. Otherwise, we attach ourselves to the
3208 		 * previous iclog and go to sleep.
3209 		 */
3210 		iclog = iclog->ic_prev;
3211 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3212 		if (atomic_read(&iclog->ic_refcnt) == 0) {
3213 			/* We have exclusive access to this iclog. */
3214 			bool	completed;
3215 
3216 			if (xlog_force_and_check_iclog(iclog, &completed))
3217 				goto out_error;
3218 
3219 			if (completed)
3220 				goto out_unlock;
3221 		} else {
3222 			/*
3223 			 * Someone else is still writing to this iclog, so we
3224 			 * need to ensure that when they release the iclog it
3225 			 * gets synced immediately as we may be waiting on it.
3226 			 */
3227 			xlog_state_switch_iclogs(log, iclog, 0);
3228 		}
3229 	}
3230 
3231 	/*
3232 	 * The iclog we are about to wait on may contain the checkpoint pushed
3233 	 * by the above xlog_cil_force() call, but it may not have been pushed
3234 	 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3235 	 * are flushed when this iclog is written.
3236 	 */
3237 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3238 		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3239 
3240 	if (flags & XFS_LOG_SYNC)
3241 		return xlog_wait_on_iclog(iclog);
3242 out_unlock:
3243 	spin_unlock(&log->l_icloglock);
3244 	return 0;
3245 out_error:
3246 	spin_unlock(&log->l_icloglock);
3247 	return -EIO;
3248 }
3249 
3250 /*
3251  * Force the log to a specific LSN.
3252  *
3253  * If an iclog with that lsn can be found:
3254  *	If it is in the DIRTY state, just return.
3255  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3256  *		state and go to sleep or return.
3257  *	If it is in any other state, go to sleep or return.
3258  *
3259  * Synchronous forces are implemented with a wait queue.  All callers trying
3260  * to force a given lsn to disk must wait on the queue attached to the
3261  * specific in-core log.  When given in-core log finally completes its write
3262  * to disk, that thread will wake up all threads waiting on the queue.
3263  */
3264 static int
xlog_force_lsn(struct xlog * log,xfs_lsn_t lsn,uint flags,int * log_flushed,bool already_slept)3265 xlog_force_lsn(
3266 	struct xlog		*log,
3267 	xfs_lsn_t		lsn,
3268 	uint			flags,
3269 	int			*log_flushed,
3270 	bool			already_slept)
3271 {
3272 	struct xlog_in_core	*iclog;
3273 	bool			completed;
3274 
3275 	spin_lock(&log->l_icloglock);
3276 	if (xlog_is_shutdown(log))
3277 		goto out_error;
3278 
3279 	iclog = log->l_iclog;
3280 	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3281 		trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3282 		iclog = iclog->ic_next;
3283 		if (iclog == log->l_iclog)
3284 			goto out_unlock;
3285 	}
3286 
3287 	switch (iclog->ic_state) {
3288 	case XLOG_STATE_ACTIVE:
3289 		/*
3290 		 * We sleep here if we haven't already slept (e.g. this is the
3291 		 * first time we've looked at the correct iclog buf) and the
3292 		 * buffer before us is going to be sync'ed.  The reason for this
3293 		 * is that if we are doing sync transactions here, by waiting
3294 		 * for the previous I/O to complete, we can allow a few more
3295 		 * transactions into this iclog before we close it down.
3296 		 *
3297 		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3298 		 * refcnt so we can release the log (which drops the ref count).
3299 		 * The state switch keeps new transaction commits from using
3300 		 * this buffer.  When the current commits finish writing into
3301 		 * the buffer, the refcount will drop to zero and the buffer
3302 		 * will go out then.
3303 		 */
3304 		if (!already_slept &&
3305 		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3306 		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3307 			xlog_wait(&iclog->ic_prev->ic_write_wait,
3308 					&log->l_icloglock);
3309 			return -EAGAIN;
3310 		}
3311 		if (xlog_force_and_check_iclog(iclog, &completed))
3312 			goto out_error;
3313 		if (log_flushed)
3314 			*log_flushed = 1;
3315 		if (completed)
3316 			goto out_unlock;
3317 		break;
3318 	case XLOG_STATE_WANT_SYNC:
3319 		/*
3320 		 * This iclog may contain the checkpoint pushed by the
3321 		 * xlog_cil_force_seq() call, but there are other writers still
3322 		 * accessing it so it hasn't been pushed to disk yet. Like the
3323 		 * ACTIVE case above, we need to make sure caches are flushed
3324 		 * when this iclog is written.
3325 		 */
3326 		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3327 		break;
3328 	default:
3329 		/*
3330 		 * The entire checkpoint was written by the CIL force and is on
3331 		 * its way to disk already. It will be stable when it
3332 		 * completes, so we don't need to manipulate caches here at all.
3333 		 * We just need to wait for completion if necessary.
3334 		 */
3335 		break;
3336 	}
3337 
3338 	if (flags & XFS_LOG_SYNC)
3339 		return xlog_wait_on_iclog(iclog);
3340 out_unlock:
3341 	spin_unlock(&log->l_icloglock);
3342 	return 0;
3343 out_error:
3344 	spin_unlock(&log->l_icloglock);
3345 	return -EIO;
3346 }
3347 
3348 /*
3349  * Force the log to a specific checkpoint sequence.
3350  *
3351  * First force the CIL so that all the required changes have been flushed to the
3352  * iclogs. If the CIL force completed it will return a commit LSN that indicates
3353  * the iclog that needs to be flushed to stable storage. If the caller needs
3354  * a synchronous log force, we will wait on the iclog with the LSN returned by
3355  * xlog_cil_force_seq() to be completed.
3356  */
3357 int
xfs_log_force_seq(struct xfs_mount * mp,xfs_csn_t seq,uint flags,int * log_flushed)3358 xfs_log_force_seq(
3359 	struct xfs_mount	*mp,
3360 	xfs_csn_t		seq,
3361 	uint			flags,
3362 	int			*log_flushed)
3363 {
3364 	struct xlog		*log = mp->m_log;
3365 	xfs_lsn_t		lsn;
3366 	int			ret;
3367 	ASSERT(seq != 0);
3368 
3369 	XFS_STATS_INC(mp, xs_log_force);
3370 	trace_xfs_log_force(mp, seq, _RET_IP_);
3371 
3372 	lsn = xlog_cil_force_seq(log, seq);
3373 	if (lsn == NULLCOMMITLSN)
3374 		return 0;
3375 
3376 	ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3377 	if (ret == -EAGAIN) {
3378 		XFS_STATS_INC(mp, xs_log_force_sleep);
3379 		ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3380 	}
3381 	return ret;
3382 }
3383 
3384 /*
3385  * Free a used ticket when its refcount falls to zero.
3386  */
3387 void
xfs_log_ticket_put(xlog_ticket_t * ticket)3388 xfs_log_ticket_put(
3389 	xlog_ticket_t	*ticket)
3390 {
3391 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3392 	if (atomic_dec_and_test(&ticket->t_ref))
3393 		kmem_cache_free(xfs_log_ticket_cache, ticket);
3394 }
3395 
3396 xlog_ticket_t *
xfs_log_ticket_get(xlog_ticket_t * ticket)3397 xfs_log_ticket_get(
3398 	xlog_ticket_t	*ticket)
3399 {
3400 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3401 	atomic_inc(&ticket->t_ref);
3402 	return ticket;
3403 }
3404 
3405 /*
3406  * Figure out the total log space unit (in bytes) that would be
3407  * required for a log ticket.
3408  */
3409 static int
xlog_calc_unit_res(struct xlog * log,int unit_bytes,int * niclogs)3410 xlog_calc_unit_res(
3411 	struct xlog		*log,
3412 	int			unit_bytes,
3413 	int			*niclogs)
3414 {
3415 	int			iclog_space;
3416 	uint			num_headers;
3417 
3418 	/*
3419 	 * Permanent reservations have up to 'cnt'-1 active log operations
3420 	 * in the log.  A unit in this case is the amount of space for one
3421 	 * of these log operations.  Normal reservations have a cnt of 1
3422 	 * and their unit amount is the total amount of space required.
3423 	 *
3424 	 * The following lines of code account for non-transaction data
3425 	 * which occupy space in the on-disk log.
3426 	 *
3427 	 * Normal form of a transaction is:
3428 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3429 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3430 	 *
3431 	 * We need to account for all the leadup data and trailer data
3432 	 * around the transaction data.
3433 	 * And then we need to account for the worst case in terms of using
3434 	 * more space.
3435 	 * The worst case will happen if:
3436 	 * - the placement of the transaction happens to be such that the
3437 	 *   roundoff is at its maximum
3438 	 * - the transaction data is synced before the commit record is synced
3439 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3440 	 *   Therefore the commit record is in its own Log Record.
3441 	 *   This can happen as the commit record is called with its
3442 	 *   own region to xlog_write().
3443 	 *   This then means that in the worst case, roundoff can happen for
3444 	 *   the commit-rec as well.
3445 	 *   The commit-rec is smaller than padding in this scenario and so it is
3446 	 *   not added separately.
3447 	 */
3448 
3449 	/* for trans header */
3450 	unit_bytes += sizeof(xlog_op_header_t);
3451 	unit_bytes += sizeof(xfs_trans_header_t);
3452 
3453 	/* for start-rec */
3454 	unit_bytes += sizeof(xlog_op_header_t);
3455 
3456 	/*
3457 	 * for LR headers - the space for data in an iclog is the size minus
3458 	 * the space used for the headers. If we use the iclog size, then we
3459 	 * undercalculate the number of headers required.
3460 	 *
3461 	 * Furthermore - the addition of op headers for split-recs might
3462 	 * increase the space required enough to require more log and op
3463 	 * headers, so take that into account too.
3464 	 *
3465 	 * IMPORTANT: This reservation makes the assumption that if this
3466 	 * transaction is the first in an iclog and hence has the LR headers
3467 	 * accounted to it, then the remaining space in the iclog is
3468 	 * exclusively for this transaction.  i.e. if the transaction is larger
3469 	 * than the iclog, it will be the only thing in that iclog.
3470 	 * Fundamentally, this means we must pass the entire log vector to
3471 	 * xlog_write to guarantee this.
3472 	 */
3473 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3474 	num_headers = howmany(unit_bytes, iclog_space);
3475 
3476 	/* for split-recs - ophdrs added when data split over LRs */
3477 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3478 
3479 	/* add extra header reservations if we overrun */
3480 	while (!num_headers ||
3481 	       howmany(unit_bytes, iclog_space) > num_headers) {
3482 		unit_bytes += sizeof(xlog_op_header_t);
3483 		num_headers++;
3484 	}
3485 	unit_bytes += log->l_iclog_hsize * num_headers;
3486 
3487 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3488 	unit_bytes += log->l_iclog_hsize;
3489 
3490 	/* roundoff padding for transaction data and one for commit record */
3491 	unit_bytes += 2 * log->l_iclog_roundoff;
3492 
3493 	if (niclogs)
3494 		*niclogs = num_headers;
3495 	return unit_bytes;
3496 }
3497 
3498 int
xfs_log_calc_unit_res(struct xfs_mount * mp,int unit_bytes)3499 xfs_log_calc_unit_res(
3500 	struct xfs_mount	*mp,
3501 	int			unit_bytes)
3502 {
3503 	return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3504 }
3505 
3506 /*
3507  * Allocate and initialise a new log ticket.
3508  */
3509 struct xlog_ticket *
xlog_ticket_alloc(struct xlog * log,int unit_bytes,int cnt,bool permanent)3510 xlog_ticket_alloc(
3511 	struct xlog		*log,
3512 	int			unit_bytes,
3513 	int			cnt,
3514 	bool			permanent)
3515 {
3516 	struct xlog_ticket	*tic;
3517 	int			unit_res;
3518 
3519 	tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3520 
3521 	unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3522 
3523 	atomic_set(&tic->t_ref, 1);
3524 	tic->t_task		= current;
3525 	INIT_LIST_HEAD(&tic->t_queue);
3526 	tic->t_unit_res		= unit_res;
3527 	tic->t_curr_res		= unit_res;
3528 	tic->t_cnt		= cnt;
3529 	tic->t_ocnt		= cnt;
3530 	tic->t_tid		= get_random_u32();
3531 	if (permanent)
3532 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3533 
3534 	return tic;
3535 }
3536 
3537 #if defined(DEBUG)
3538 /*
3539  * Check to make sure the grant write head didn't just over lap the tail.  If
3540  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3541  * the cycles differ by exactly one and check the byte count.
3542  *
3543  * This check is run unlocked, so can give false positives. Rather than assert
3544  * on failures, use a warn-once flag and a panic tag to allow the admin to
3545  * determine if they want to panic the machine when such an error occurs. For
3546  * debug kernels this will have the same effect as using an assert but, unlinke
3547  * an assert, it can be turned off at runtime.
3548  */
3549 STATIC void
xlog_verify_grant_tail(struct xlog * log)3550 xlog_verify_grant_tail(
3551 	struct xlog	*log)
3552 {
3553 	int		tail_cycle, tail_blocks;
3554 	int		cycle, space;
3555 
3556 	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3557 	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3558 	if (tail_cycle != cycle) {
3559 		if (cycle - 1 != tail_cycle &&
3560 		    !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3561 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3562 				"%s: cycle - 1 != tail_cycle", __func__);
3563 		}
3564 
3565 		if (space > BBTOB(tail_blocks) &&
3566 		    !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3567 			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3568 				"%s: space > BBTOB(tail_blocks)", __func__);
3569 		}
3570 	}
3571 }
3572 
3573 /* check if it will fit */
3574 STATIC void
xlog_verify_tail_lsn(struct xlog * log,struct xlog_in_core * iclog)3575 xlog_verify_tail_lsn(
3576 	struct xlog		*log,
3577 	struct xlog_in_core	*iclog)
3578 {
3579 	xfs_lsn_t	tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3580 	int		blocks;
3581 
3582     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3583 	blocks =
3584 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3585 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3586 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3587     } else {
3588 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3589 
3590 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3591 		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3592 
3593 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3594 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3595 		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3596     }
3597 }
3598 
3599 /*
3600  * Perform a number of checks on the iclog before writing to disk.
3601  *
3602  * 1. Make sure the iclogs are still circular
3603  * 2. Make sure we have a good magic number
3604  * 3. Make sure we don't have magic numbers in the data
3605  * 4. Check fields of each log operation header for:
3606  *	A. Valid client identifier
3607  *	B. tid ptr value falls in valid ptr space (user space code)
3608  *	C. Length in log record header is correct according to the
3609  *		individual operation headers within record.
3610  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3611  *	log, check the preceding blocks of the physical log to make sure all
3612  *	the cycle numbers agree with the current cycle number.
3613  */
3614 STATIC void
xlog_verify_iclog(struct xlog * log,struct xlog_in_core * iclog,int count)3615 xlog_verify_iclog(
3616 	struct xlog		*log,
3617 	struct xlog_in_core	*iclog,
3618 	int			count)
3619 {
3620 	xlog_op_header_t	*ophead;
3621 	xlog_in_core_t		*icptr;
3622 	xlog_in_core_2_t	*xhdr;
3623 	void			*base_ptr, *ptr, *p;
3624 	ptrdiff_t		field_offset;
3625 	uint8_t			clientid;
3626 	int			len, i, j, k, op_len;
3627 	int			idx;
3628 
3629 	/* check validity of iclog pointers */
3630 	spin_lock(&log->l_icloglock);
3631 	icptr = log->l_iclog;
3632 	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3633 		ASSERT(icptr);
3634 
3635 	if (icptr != log->l_iclog)
3636 		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3637 	spin_unlock(&log->l_icloglock);
3638 
3639 	/* check log magic numbers */
3640 	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3641 		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3642 
3643 	base_ptr = ptr = &iclog->ic_header;
3644 	p = &iclog->ic_header;
3645 	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3646 		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3647 			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3648 				__func__);
3649 	}
3650 
3651 	/* check fields */
3652 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3653 	base_ptr = ptr = iclog->ic_datap;
3654 	ophead = ptr;
3655 	xhdr = iclog->ic_data;
3656 	for (i = 0; i < len; i++) {
3657 		ophead = ptr;
3658 
3659 		/* clientid is only 1 byte */
3660 		p = &ophead->oh_clientid;
3661 		field_offset = p - base_ptr;
3662 		if (field_offset & 0x1ff) {
3663 			clientid = ophead->oh_clientid;
3664 		} else {
3665 			idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3666 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3667 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3668 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3669 				clientid = xlog_get_client_id(
3670 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3671 			} else {
3672 				clientid = xlog_get_client_id(
3673 					iclog->ic_header.h_cycle_data[idx]);
3674 			}
3675 		}
3676 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3677 			xfs_warn(log->l_mp,
3678 				"%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3679 				__func__, i, clientid, ophead,
3680 				(unsigned long)field_offset);
3681 		}
3682 
3683 		/* check length */
3684 		p = &ophead->oh_len;
3685 		field_offset = p - base_ptr;
3686 		if (field_offset & 0x1ff) {
3687 			op_len = be32_to_cpu(ophead->oh_len);
3688 		} else {
3689 			idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3690 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3691 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3692 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3693 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3694 			} else {
3695 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3696 			}
3697 		}
3698 		ptr += sizeof(xlog_op_header_t) + op_len;
3699 	}
3700 }
3701 #endif
3702 
3703 /*
3704  * Perform a forced shutdown on the log.
3705  *
3706  * This can be called from low level log code to trigger a shutdown, or from the
3707  * high level mount shutdown code when the mount shuts down.
3708  *
3709  * Our main objectives here are to make sure that:
3710  *	a. if the shutdown was not due to a log IO error, flush the logs to
3711  *	   disk. Anything modified after this is ignored.
3712  *	b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3713  *	   parties to find out. Nothing new gets queued after this is done.
3714  *	c. Tasks sleeping on log reservations, pinned objects and
3715  *	   other resources get woken up.
3716  *	d. The mount is also marked as shut down so that log triggered shutdowns
3717  *	   still behave the same as if they called xfs_forced_shutdown().
3718  *
3719  * Return true if the shutdown cause was a log IO error and we actually shut the
3720  * log down.
3721  */
3722 bool
xlog_force_shutdown(struct xlog * log,uint32_t shutdown_flags)3723 xlog_force_shutdown(
3724 	struct xlog	*log,
3725 	uint32_t	shutdown_flags)
3726 {
3727 	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3728 
3729 	if (!log)
3730 		return false;
3731 
3732 	/*
3733 	 * Flush all the completed transactions to disk before marking the log
3734 	 * being shut down. We need to do this first as shutting down the log
3735 	 * before the force will prevent the log force from flushing the iclogs
3736 	 * to disk.
3737 	 *
3738 	 * When we are in recovery, there are no transactions to flush, and
3739 	 * we don't want to touch the log because we don't want to perturb the
3740 	 * current head/tail for future recovery attempts. Hence we need to
3741 	 * avoid a log force in this case.
3742 	 *
3743 	 * If we are shutting down due to a log IO error, then we must avoid
3744 	 * trying to write the log as that may just result in more IO errors and
3745 	 * an endless shutdown/force loop.
3746 	 */
3747 	if (!log_error && !xlog_in_recovery(log))
3748 		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3749 
3750 	/*
3751 	 * Atomically set the shutdown state. If the shutdown state is already
3752 	 * set, there someone else is performing the shutdown and so we are done
3753 	 * here. This should never happen because we should only ever get called
3754 	 * once by the first shutdown caller.
3755 	 *
3756 	 * Much of the log state machine transitions assume that shutdown state
3757 	 * cannot change once they hold the log->l_icloglock. Hence we need to
3758 	 * hold that lock here, even though we use the atomic test_and_set_bit()
3759 	 * operation to set the shutdown state.
3760 	 */
3761 	spin_lock(&log->l_icloglock);
3762 	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3763 		spin_unlock(&log->l_icloglock);
3764 		return false;
3765 	}
3766 	spin_unlock(&log->l_icloglock);
3767 
3768 	/*
3769 	 * If this log shutdown also sets the mount shutdown state, issue a
3770 	 * shutdown warning message.
3771 	 */
3772 	if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3773 		xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3774 "Filesystem has been shut down due to log error (0x%x).",
3775 				shutdown_flags);
3776 		xfs_alert(log->l_mp,
3777 "Please unmount the filesystem and rectify the problem(s).");
3778 		if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3779 			xfs_stack_trace();
3780 	}
3781 
3782 	/*
3783 	 * We don't want anybody waiting for log reservations after this. That
3784 	 * means we have to wake up everybody queued up on reserveq as well as
3785 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3786 	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3787 	 * action is protected by the grant locks.
3788 	 */
3789 	xlog_grant_head_wake_all(&log->l_reserve_head);
3790 	xlog_grant_head_wake_all(&log->l_write_head);
3791 
3792 	/*
3793 	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3794 	 * as if the log writes were completed. The abort handling in the log
3795 	 * item committed callback functions will do this again under lock to
3796 	 * avoid races.
3797 	 */
3798 	spin_lock(&log->l_cilp->xc_push_lock);
3799 	wake_up_all(&log->l_cilp->xc_start_wait);
3800 	wake_up_all(&log->l_cilp->xc_commit_wait);
3801 	spin_unlock(&log->l_cilp->xc_push_lock);
3802 
3803 	spin_lock(&log->l_icloglock);
3804 	xlog_state_shutdown_callbacks(log);
3805 	spin_unlock(&log->l_icloglock);
3806 
3807 	wake_up_var(&log->l_opstate);
3808 	return log_error;
3809 }
3810 
3811 STATIC int
xlog_iclogs_empty(struct xlog * log)3812 xlog_iclogs_empty(
3813 	struct xlog	*log)
3814 {
3815 	xlog_in_core_t	*iclog;
3816 
3817 	iclog = log->l_iclog;
3818 	do {
3819 		/* endianness does not matter here, zero is zero in
3820 		 * any language.
3821 		 */
3822 		if (iclog->ic_header.h_num_logops)
3823 			return 0;
3824 		iclog = iclog->ic_next;
3825 	} while (iclog != log->l_iclog);
3826 	return 1;
3827 }
3828 
3829 /*
3830  * Verify that an LSN stamped into a piece of metadata is valid. This is
3831  * intended for use in read verifiers on v5 superblocks.
3832  */
3833 bool
xfs_log_check_lsn(struct xfs_mount * mp,xfs_lsn_t lsn)3834 xfs_log_check_lsn(
3835 	struct xfs_mount	*mp,
3836 	xfs_lsn_t		lsn)
3837 {
3838 	struct xlog		*log = mp->m_log;
3839 	bool			valid;
3840 
3841 	/*
3842 	 * norecovery mode skips mount-time log processing and unconditionally
3843 	 * resets the in-core LSN. We can't validate in this mode, but
3844 	 * modifications are not allowed anyways so just return true.
3845 	 */
3846 	if (xfs_has_norecovery(mp))
3847 		return true;
3848 
3849 	/*
3850 	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3851 	 * handled by recovery and thus safe to ignore here.
3852 	 */
3853 	if (lsn == NULLCOMMITLSN)
3854 		return true;
3855 
3856 	valid = xlog_valid_lsn(mp->m_log, lsn);
3857 
3858 	/* warn the user about what's gone wrong before verifier failure */
3859 	if (!valid) {
3860 		spin_lock(&log->l_icloglock);
3861 		xfs_warn(mp,
3862 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3863 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3864 			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3865 			 log->l_curr_cycle, log->l_curr_block);
3866 		spin_unlock(&log->l_icloglock);
3867 	}
3868 
3869 	return valid;
3870 }
3871 
3872 /*
3873  * Notify the log that we're about to start using a feature that is protected
3874  * by a log incompat feature flag.  This will prevent log covering from
3875  * clearing those flags.
3876  */
3877 void
xlog_use_incompat_feat(struct xlog * log)3878 xlog_use_incompat_feat(
3879 	struct xlog		*log)
3880 {
3881 	down_read(&log->l_incompat_users);
3882 }
3883 
3884 /* Notify the log that we've finished using log incompat features. */
3885 void
xlog_drop_incompat_feat(struct xlog * log)3886 xlog_drop_incompat_feat(
3887 	struct xlog		*log)
3888 {
3889 	up_read(&log->l_incompat_users);
3890 }
3891