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	#include "xfs_zone_alloc.h"
24
25	struct kmem_cache *xfs_log_ticket_cache;
26
27	/* Local miscellaneous function prototypes */
28	STATIC struct xlog *
29	xlog_alloc_log(
30	struct xfs_mount *mp,
31	struct xfs_buftarg *log_target,
32	xfs_daddr_t blk_offset,
33	int num_bblks);
34	STATIC void
35	xlog_dealloc_log(
36	struct xlog *log);
37
38	/* local state machine functions */
39	STATIC void xlog_state_done_syncing(
40	struct xlog_in_core *iclog);
41	STATIC void xlog_state_do_callback(
42	struct xlog *log);
43	STATIC int
44	xlog_state_get_iclog_space(
45	struct xlog *log,
46	int len,
47	struct xlog_in_core **iclog,
48	struct xlog_ticket *ticket,
49	int *logoffsetp);
50	STATIC void
51	xlog_sync(
52	struct xlog *log,
53	struct xlog_in_core *iclog,
54	struct xlog_ticket *ticket);
55	#if defined(DEBUG)
56	STATIC void
57	xlog_verify_iclog(
58	struct xlog *log,
59	struct xlog_in_core *iclog,
60	int count);
61	STATIC void
62	xlog_verify_tail_lsn(
63	struct xlog *log,
64	struct xlog_in_core *iclog);
65	#else
66	#define xlog_verify_iclog(a,b,c)
67	#define xlog_verify_tail_lsn(a,b)
68	#endif
69
70	STATIC int
71	xlog_iclogs_empty(
72	struct xlog *log);
73
74	static int
75	xfs_log_cover(struct xfs_mount *);
76
77	/*
78	* We need to make sure the buffer pointer returned is naturally aligned for the
79	* biggest basic data type we put into it. We have already accounted for this
80	* padding when sizing the buffer.
81	*
82	* However, this padding does not get written into the log, and hence we have to
83	* track the space used by the log vectors separately to prevent log space hangs
84	* due to inaccurate accounting (i.e. a leak) of the used log space through the
85	* CIL context ticket.
86	*
87	* We also add space for the xlog_op_header that describes this region in the
88	* log. This prepends the data region we return to the caller to copy their data
89	* into, so do all the static initialisation of the ophdr now. Because the ophdr
90	* is not 8 byte aligned, we have to be careful to ensure that we align the
91	* start of the buffer such that the region we return to the call is 8 byte
92	* aligned and packed against the tail of the ophdr.
93	*/
94	void *
95	xlog_prepare_iovec(
96	struct xfs_log_vec *lv,
97	struct xfs_log_iovec **vecp,
98	uint type)
99	{
100	struct xfs_log_iovec *vec = *vecp;
101	struct xlog_op_header *oph;
102	uint32_t len;
103	void *buf;
104
105	if (vec) {
106	ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
107	vec++;
108	} else {
109	vec = &lv->lv_iovecp[0];
110	}
111
112	len = lv->lv_buf_used + sizeof(struct xlog_op_header);
113	if (!IS_ALIGNED(len, sizeof(uint64_t))) {
114	lv->lv_buf_used = round_up(len, sizeof(uint64_t)) -
115	sizeof(struct xlog_op_header);
116	}
117
118	vec->i_type = type;
119	vec->i_addr = lv->lv_buf + lv->lv_buf_used;
120
121	oph = vec->i_addr;
122	oph->oh_clientid = XFS_TRANSACTION;
123	oph->oh_res2 = 0;
124	oph->oh_flags = 0;
125
126	buf = vec->i_addr + sizeof(struct xlog_op_header);
127	ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
128
129	*vecp = vec;
130	return buf;
131	}
132
133	static inline void
134	xlog_grant_sub_space(
135	struct xlog_grant_head *head,
136	int64_t bytes)
137	{
138	atomic64_sub(i: bytes, v: &head->grant);
139	}
140
141	static inline void
142	xlog_grant_add_space(
143	struct xlog_grant_head *head,
144	int64_t bytes)
145	{
146	atomic64_add(i: bytes, v: &head->grant);
147	}
148
149	static void
150	xlog_grant_head_init(
151	struct xlog_grant_head *head)
152	{
153	atomic64_set(v: &head->grant, i: 0);
154	INIT_LIST_HEAD(list: &head->waiters);
155	spin_lock_init(&head->lock);
156	}
157
158	void
159	xlog_grant_return_space(
160	struct xlog *log,
161	xfs_lsn_t old_head,
162	xfs_lsn_t new_head)
163	{
164	int64_t diff = xlog_lsn_sub(log, new_head, old_head);
165
166	xlog_grant_sub_space(head: &log->l_reserve_head, bytes: diff);
167	xlog_grant_sub_space(head: &log->l_write_head, bytes: diff);
168	}
169
170	/*
171	* Return the space in the log between the tail and the head. In the case where
172	* we have overrun available reservation space, return 0. The memory barrier
173	* pairs with the smp_wmb() in xlog_cil_ail_insert() to ensure that grant head
174	* vs tail space updates are seen in the correct order and hence avoid
175	* transients as space is transferred from the grant heads to the AIL on commit
176	* completion.
177	*/
178	static uint64_t
179	xlog_grant_space_left(
180	struct xlog *log,
181	struct xlog_grant_head *head)
182	{
183	int64_t free_bytes;
184
185	smp_rmb(); /* paired with smp_wmb in xlog_cil_ail_insert() */
186	free_bytes = log->l_logsize - READ_ONCE(log->l_tail_space) -
187	atomic64_read(v: &head->grant);
188	if (free_bytes > 0)
189	return free_bytes;
190	return 0;
191	}
192
193	STATIC void
194	xlog_grant_head_wake_all(
195	struct xlog_grant_head *head)
196	{
197	struct xlog_ticket *tic;
198
199	spin_lock(lock: &head->lock);
200	list_for_each_entry(tic, &head->waiters, t_queue)
201	wake_up_process(tsk: tic->t_task);
202	spin_unlock(lock: &head->lock);
203	}
204
205	static inline int
206	xlog_ticket_reservation(
207	struct xlog *log,
208	struct xlog_grant_head *head,
209	struct xlog_ticket *tic)
210	{
211	if (head == &log->l_write_head) {
212	ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
213	return tic->t_unit_res;
214	}
215
216	if (tic->t_flags & XLOG_TIC_PERM_RESERV)
217	return tic->t_unit_res * tic->t_cnt;
218
219	return tic->t_unit_res;
220	}
221
222	STATIC bool
223	xlog_grant_head_wake(
224	struct xlog *log,
225	struct xlog_grant_head *head,
226	int *free_bytes)
227	{
228	struct xlog_ticket *tic;
229	int need_bytes;
230
231	list_for_each_entry(tic, &head->waiters, t_queue) {
232	need_bytes = xlog_ticket_reservation(log, head, tic);
233	if (*free_bytes < need_bytes)
234	return false;
235
236	*free_bytes -= need_bytes;
237	trace_xfs_log_grant_wake_up(log, tic);
238	wake_up_process(tsk: tic->t_task);
239	}
240
241	return true;
242	}
243
244	STATIC int
245	xlog_grant_head_wait(
246	struct xlog *log,
247	struct xlog_grant_head *head,
248	struct xlog_ticket *tic,
249	int need_bytes) __releases(&head->lock)
250	__acquires(&head->lock)
251	{
252	list_add_tail(new: &tic->t_queue, head: &head->waiters);
253
254	do {
255	if (xlog_is_shutdown(log))
256	goto shutdown;
257
258	__set_current_state(TASK_UNINTERRUPTIBLE);
259	spin_unlock(lock: &head->lock);
260
261	XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
262
263	/* Push on the AIL to free up all the log space. */
264	xfs_ail_push_all(ailp: log->l_ailp);
265
266	trace_xfs_log_grant_sleep(log, tic);
267	schedule();
268	trace_xfs_log_grant_wake(log, tic);
269
270	spin_lock(lock: &head->lock);
271	if (xlog_is_shutdown(log))
272	goto shutdown;
273	} while (xlog_grant_space_left(log, head) < need_bytes);
274
275	list_del_init(entry: &tic->t_queue);
276	return 0;
277	shutdown:
278	list_del_init(entry: &tic->t_queue);
279	return -EIO;
280	}
281
282	/*
283	* Atomically get the log space required for a log ticket.
284	*
285	* Once a ticket gets put onto head->waiters, it will only return after the
286	* needed reservation is satisfied.
287	*
288	* This function is structured so that it has a lock free fast path. This is
289	* necessary because every new transaction reservation will come through this
290	* path. Hence any lock will be globally hot if we take it unconditionally on
291	* every pass.
292	*
293	* As tickets are only ever moved on and off head->waiters under head->lock, we
294	* only need to take that lock if we are going to add the ticket to the queue
295	* and sleep. We can avoid taking the lock if the ticket was never added to
296	* head->waiters because the t_queue list head will be empty and we hold the
297	* only reference to it so it can safely be checked unlocked.
298	*/
299	STATIC int
300	xlog_grant_head_check(
301	struct xlog *log,
302	struct xlog_grant_head *head,
303	struct xlog_ticket *tic,
304	int *need_bytes)
305	{
306	int free_bytes;
307	int error = 0;
308
309	ASSERT(!xlog_in_recovery(log));
310
311	/*
312	* If there are other waiters on the queue then give them a chance at
313	* logspace before us. Wake up the first waiters, if we do not wake
314	* up all the waiters then go to sleep waiting for more free space,
315	* otherwise try to get some space for this transaction.
316	*/
317	*need_bytes = xlog_ticket_reservation(log, head, tic);
318	free_bytes = xlog_grant_space_left(log, head);
319	if (!list_empty_careful(head: &head->waiters)) {
320	spin_lock(lock: &head->lock);
321	if (!xlog_grant_head_wake(log, head, free_bytes: &free_bytes) ||
322	free_bytes < *need_bytes) {
323	error = xlog_grant_head_wait(log, head, tic,
324	need_bytes: *need_bytes);
325	}
326	spin_unlock(lock: &head->lock);
327	} else if (free_bytes < *need_bytes) {
328	spin_lock(lock: &head->lock);
329	error = xlog_grant_head_wait(log, head, tic, need_bytes: *need_bytes);
330	spin_unlock(lock: &head->lock);
331	}
332
333	return error;
334	}
335
336	bool
337	xfs_log_writable(
338	struct xfs_mount *mp)
339	{
340	/*
341	* Do not write to the log on norecovery mounts, if the data or log
342	* devices are read-only, or if the filesystem is shutdown. Read-only
343	* mounts allow internal writes for log recovery and unmount purposes,
344	* so don't restrict that case.
345	*/
346	if (xfs_has_norecovery(mp))
347	return false;
348	if (xfs_readonly_buftarg(mp->m_ddev_targp))
349	return false;
350	if (xfs_readonly_buftarg(mp->m_log->l_targ))
351	return false;
352	if (xlog_is_shutdown(log: mp->m_log))
353	return false;
354	return true;
355	}
356
357	/*
358	* Replenish the byte reservation required by moving the grant write head.
359	*/
360	int
361	xfs_log_regrant(
362	struct xfs_mount *mp,
363	struct xlog_ticket *tic)
364	{
365	struct xlog *log = mp->m_log;
366	int need_bytes;
367	int error = 0;
368
369	if (xlog_is_shutdown(log))
370	return -EIO;
371
372	XFS_STATS_INC(mp, xs_try_logspace);
373
374	/*
375	* This is a new transaction on the ticket, so we need to change the
376	* transaction ID so that the next transaction has a different TID in
377	* the log. Just add one to the existing tid so that we can see chains
378	* of rolling transactions in the log easily.
379	*/
380	tic->t_tid++;
381	tic->t_curr_res = tic->t_unit_res;
382	if (tic->t_cnt > 0)
383	return 0;
384
385	trace_xfs_log_regrant(log, tic);
386
387	error = xlog_grant_head_check(log, head: &log->l_write_head, tic,
388	need_bytes: &need_bytes);
389	if (error)
390	goto out_error;
391
392	xlog_grant_add_space(head: &log->l_write_head, bytes: need_bytes);
393	trace_xfs_log_regrant_exit(log, tic);
394	return 0;
395
396	out_error:
397	/*
398	* If we are failing, make sure the ticket doesn't have any current
399	* reservations. We don't want to add this back when the ticket/
400	* transaction gets cancelled.
401	*/
402	tic->t_curr_res = 0;
403	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
404	return error;
405	}
406
407	/*
408	* Reserve log space and return a ticket corresponding to the reservation.
409	*
410	* Each reservation is going to reserve extra space for a log record header.
411	* When writes happen to the on-disk log, we don't subtract the length of the
412	* log record header from any reservation. By wasting space in each
413	* reservation, we prevent over allocation problems.
414	*/
415	int
416	xfs_log_reserve(
417	struct xfs_mount *mp,
418	int unit_bytes,
419	int cnt,
420	struct xlog_ticket **ticp,
421	bool permanent)
422	{
423	struct xlog *log = mp->m_log;
424	struct xlog_ticket *tic;
425	int need_bytes;
426	int error = 0;
427
428	if (xlog_is_shutdown(log))
429	return -EIO;
430
431	XFS_STATS_INC(mp, xs_try_logspace);
432
433	ASSERT(*ticp == NULL);
434	tic = xlog_ticket_alloc(log, unit_bytes, count: cnt, permanent);
435	*ticp = tic;
436	trace_xfs_log_reserve(log, tic);
437	error = xlog_grant_head_check(log, head: &log->l_reserve_head, tic,
438	need_bytes: &need_bytes);
439	if (error)
440	goto out_error;
441
442	xlog_grant_add_space(head: &log->l_reserve_head, bytes: need_bytes);
443	xlog_grant_add_space(head: &log->l_write_head, bytes: need_bytes);
444	trace_xfs_log_reserve_exit(log, tic);
445	return 0;
446
447	out_error:
448	/*
449	* If we are failing, make sure the ticket doesn't have any current
450	* reservations. We don't want to add this back when the ticket/
451	* transaction gets cancelled.
452	*/
453	tic->t_curr_res = 0;
454	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
455	return error;
456	}
457
458	/*
459	* Run all the pending iclog callbacks and wake log force waiters and iclog
460	* space waiters so they can process the newly set shutdown state. We really
461	* don't care what order we process callbacks here because the log is shut down
462	* and so state cannot change on disk anymore. However, we cannot wake waiters
463	* until the callbacks have been processed because we may be in unmount and
464	* we must ensure that all AIL operations the callbacks perform have completed
465	* before we tear down the AIL.
466	*
467	* We avoid processing actively referenced iclogs so that we don't run callbacks
468	* while the iclog owner might still be preparing the iclog for IO submssion.
469	* These will be caught by xlog_state_iclog_release() and call this function
470	* again to process any callbacks that may have been added to that iclog.
471	*/
472	static void
473	xlog_state_shutdown_callbacks(
474	struct xlog *log)
475	{
476	struct xlog_in_core *iclog;
477	LIST_HEAD(cb_list);
478
479	iclog = log->l_iclog;
480	do {
481	if (atomic_read(v: &iclog->ic_refcnt)) {
482	/* Reference holder will re-run iclog callbacks. */
483	continue;
484	}
485	list_splice_init(list: &iclog->ic_callbacks, head: &cb_list);
486	spin_unlock(lock: &log->l_icloglock);
487
488	xlog_cil_process_committed(list: &cb_list);
489
490	spin_lock(lock: &log->l_icloglock);
491	wake_up_all(&iclog->ic_write_wait);
492	wake_up_all(&iclog->ic_force_wait);
493	} while ((iclog = iclog->ic_next) != log->l_iclog);
494
495	wake_up_all(&log->l_flush_wait);
496	}
497
498	/*
499	* Flush iclog to disk if this is the last reference to the given iclog and the
500	* it is in the WANT_SYNC state.
501	*
502	* If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
503	* log tail is updated correctly. NEED_FUA indicates that the iclog will be
504	* written to stable storage, and implies that a commit record is contained
505	* within the iclog. We need to ensure that the log tail does not move beyond
506	* the tail that the first commit record in the iclog ordered against, otherwise
507	* correct recovery of that checkpoint becomes dependent on future operations
508	* performed on this iclog.
509	*
510	* Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
511	* current tail into iclog. Once the iclog tail is set, future operations must
512	* not modify it, otherwise they potentially violate ordering constraints for
513	* the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
514	* the iclog will get zeroed on activation of the iclog after sync, so we
515	* always capture the tail lsn on the iclog on the first NEED_FUA release
516	* regardless of the number of active reference counts on this iclog.
517	*/
518	int
519	xlog_state_release_iclog(
520	struct xlog *log,
521	struct xlog_in_core *iclog,
522	struct xlog_ticket *ticket)
523	{
524	bool last_ref;
525
526	lockdep_assert_held(&log->l_icloglock);
527
528	trace_xlog_iclog_release(iclog, _RET_IP_);
529	/*
530	* Grabbing the current log tail needs to be atomic w.r.t. the writing
531	* of the tail LSN into the iclog so we guarantee that the log tail does
532	* not move between the first time we know that the iclog needs to be
533	* made stable and when we eventually submit it.
534	*/
535	if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
536	(iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
537	!iclog->ic_header->h_tail_lsn) {
538	iclog->ic_header->h_tail_lsn =
539	cpu_to_be64(atomic64_read(&log->l_tail_lsn));
540	}
541
542	last_ref = atomic_dec_and_test(v: &iclog->ic_refcnt);
543
544	if (xlog_is_shutdown(log)) {
545	/*
546	* If there are no more references to this iclog, process the
547	* pending iclog callbacks that were waiting on the release of
548	* this iclog.
549	*/
550	if (last_ref)
551	xlog_state_shutdown_callbacks(log);
552	return -EIO;
553	}
554
555	if (!last_ref)
556	return 0;
557
558	if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
559	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
560	return 0;
561	}
562
563	iclog->ic_state = XLOG_STATE_SYNCING;
564	xlog_verify_tail_lsn(log, iclog);
565	trace_xlog_iclog_syncing(iclog, _RET_IP_);
566
567	spin_unlock(lock: &log->l_icloglock);
568	xlog_sync(log, iclog, ticket);
569	spin_lock(lock: &log->l_icloglock);
570	return 0;
571	}
572
573	/*
574	* Mount a log filesystem
575	*
576	* mp - ubiquitous xfs mount point structure
577	* log_target - buftarg of on-disk log device
578	* blk_offset - Start block # where block size is 512 bytes (BBSIZE)
579	* num_bblocks - Number of BBSIZE blocks in on-disk log
580	*
581	* Return error or zero.
582	*/
583	int
584	xfs_log_mount(
585	xfs_mount_t *mp,
586	struct xfs_buftarg *log_target,
587	xfs_daddr_t blk_offset,
588	int num_bblks)
589	{
590	struct xlog *log;
591	int error = 0;
592	int min_logfsbs;
593
594	if (!xfs_has_norecovery(mp)) {
595	xfs_notice(mp, "Mounting V%d Filesystem %pU",
596	XFS_SB_VERSION_NUM(&mp->m_sb),
597	&mp->m_sb.sb_uuid);
598	} else {
599	xfs_notice(mp,
600	"Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
601	XFS_SB_VERSION_NUM(&mp->m_sb),
602	&mp->m_sb.sb_uuid);
603	ASSERT(xfs_is_readonly(mp));
604	}
605
606	log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
607	if (IS_ERR(ptr: log)) {
608	error = PTR_ERR(ptr: log);
609	goto out;
610	}
611	mp->m_log = log;
612
613	/*
614	* Now that we have set up the log and it's internal geometry
615	* parameters, we can validate the given log space and drop a critical
616	* message via syslog if the log size is too small. A log that is too
617	* small can lead to unexpected situations in transaction log space
618	* reservation stage. The superblock verifier has already validated all
619	* the other log geometry constraints, so we don't have to check those
620	* here.
621	*
622	* Note: For v4 filesystems, we can't just reject the mount if the
623	* validation fails. This would mean that people would have to
624	* downgrade their kernel just to remedy the situation as there is no
625	* way to grow the log (short of black magic surgery with xfs_db).
626	*
627	* We can, however, reject mounts for V5 format filesystems, as the
628	* mkfs binary being used to make the filesystem should never create a
629	* filesystem with a log that is too small.
630	*/
631	min_logfsbs = xfs_log_calc_minimum_size(mp);
632	if (mp->m_sb.sb_logblocks < min_logfsbs) {
633	xfs_warn(mp,
634	"Log size %d blocks too small, minimum size is %d blocks",
635	mp->m_sb.sb_logblocks, min_logfsbs);
636
637	/*
638	* Log check errors are always fatal on v5; or whenever bad
639	* metadata leads to a crash.
640	*/
641	if (xfs_has_crc(mp)) {
642	xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
643	ASSERT(0);
644	error = -EINVAL;
645	goto out_free_log;
646	}
647	xfs_crit(mp, "Log size out of supported range.");
648	xfs_crit(mp,
649	"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
650	}
651
652	/*
653	* Initialize the AIL now we have a log.
654	*/
655	error = xfs_trans_ail_init(mp);
656	if (error) {
657	xfs_warn(mp, "AIL initialisation failed: error %d", error);
658	goto out_free_log;
659	}
660	log->l_ailp = mp->m_ail;
661
662	/*
663	* skip log recovery on a norecovery mount. pretend it all
664	* just worked.
665	*/
666	if (!xfs_has_norecovery(mp)) {
667	error = xlog_recover(log);
668	if (error) {
669	xfs_warn(mp, "log mount/recovery failed: error %d",
670	error);
671	xlog_recover_cancel(log);
672	goto out_destroy_ail;
673	}
674	}
675
676	error = xfs_sysfs_init(kobj: &log->l_kobj, ktype: &xfs_log_ktype, parent_kobj: &mp->m_kobj,
677	name: "log");
678	if (error)
679	goto out_destroy_ail;
680
681	/* Normal transactions can now occur */
682	clear_bit(XLOG_ACTIVE_RECOVERY, addr: &log->l_opstate);
683
684	/*
685	* Now the log has been fully initialised and we know were our
686	* space grant counters are, we can initialise the permanent ticket
687	* needed for delayed logging to work.
688	*/
689	xlog_cil_init_post_recovery(log);
690
691	return 0;
692
693	out_destroy_ail:
694	xfs_trans_ail_destroy(mp);
695	out_free_log:
696	xlog_dealloc_log(log);
697	out:
698	return error;
699	}
700
701	/*
702	* Finish the recovery of the file system. This is separate from the
703	* xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
704	* in the root and real-time bitmap inodes between calling xfs_log_mount() and
705	* here.
706	*
707	* If we finish recovery successfully, start the background log work. If we are
708	* not doing recovery, then we have a RO filesystem and we don't need to start
709	* it.
710	*/
711	int
712	xfs_log_mount_finish(
713	struct xfs_mount *mp)
714	{
715	struct xlog *log = mp->m_log;
716	int error = 0;
717
718	if (xfs_has_norecovery(mp)) {
719	ASSERT(xfs_is_readonly(mp));
720	return 0;
721	}
722
723	/*
724	* During the second phase of log recovery, we need iget and
725	* iput to behave like they do for an active filesystem.
726	* xfs_fs_drop_inode needs to be able to prevent the deletion
727	* of inodes before we're done replaying log items on those
728	* inodes. Turn it off immediately after recovery finishes
729	* so that we don't leak the quota inodes if subsequent mount
730	* activities fail.
731	*
732	* We let all inodes involved in redo item processing end up on
733	* the LRU instead of being evicted immediately so that if we do
734	* something to an unlinked inode, the irele won't cause
735	* premature truncation and freeing of the inode, which results
736	* in log recovery failure. We have to evict the unreferenced
737	* lru inodes after clearing SB_ACTIVE because we don't
738	* otherwise clean up the lru if there's a subsequent failure in
739	* xfs_mountfs, which leads to us leaking the inodes if nothing
740	* else (e.g. quotacheck) references the inodes before the
741	* mount failure occurs.
742	*/
743	mp->m_super->s_flags |= SB_ACTIVE;
744	xfs_log_work_queue(mp);
745	if (xlog_recovery_needed(log))
746	error = xlog_recover_finish(log);
747	mp->m_super->s_flags &= ~SB_ACTIVE;
748	evict_inodes(sb: mp->m_super);
749
750	/*
751	* Drain the buffer LRU after log recovery. This is required for v4
752	* filesystems to avoid leaving around buffers with NULL verifier ops,
753	* but we do it unconditionally to make sure we're always in a clean
754	* cache state after mount.
755	*
756	* Don't push in the error case because the AIL may have pending intents
757	* that aren't removed until recovery is cancelled.
758	*/
759	if (xlog_recovery_needed(log)) {
760	if (!error) {
761	xfs_log_force(mp, XFS_LOG_SYNC);
762	xfs_ail_push_all_sync(ailp: mp->m_ail);
763	}
764	xfs_notice(mp, "Ending recovery (logdev: %s)",
765	mp->m_logname ? mp->m_logname : "internal");
766	} else {
767	xfs_info(mp, "Ending clean mount");
768	}
769	xfs_buftarg_drain(mp->m_ddev_targp);
770
771	clear_bit(XLOG_RECOVERY_NEEDED, addr: &log->l_opstate);
772
773	/* Make sure the log is dead if we're returning failure. */
774	ASSERT(!error || xlog_is_shutdown(log));
775
776	return error;
777	}
778
779	/*
780	* The mount has failed. Cancel the recovery if it hasn't completed and destroy
781	* the log.
782	*/
783	void
784	xfs_log_mount_cancel(
785	struct xfs_mount *mp)
786	{
787	xlog_recover_cancel(mp->m_log);
788	xfs_log_unmount(mp);
789	}
790
791	/*
792	* Flush out the iclog to disk ensuring that device caches are flushed and
793	* the iclog hits stable storage before any completion waiters are woken.
794	*/
795	static inline int
796	xlog_force_iclog(
797	struct xlog_in_core *iclog)
798	{
799	atomic_inc(v: &iclog->ic_refcnt);
800	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
801	if (iclog->ic_state == XLOG_STATE_ACTIVE)
802	xlog_state_switch_iclogs(log: iclog->ic_log, iclog, eventual_size: 0);
803	return xlog_state_release_iclog(log: iclog->ic_log, iclog, NULL);
804	}
805
806	/*
807	* Cycle all the iclogbuf locks to make sure all log IO completion
808	* is done before we tear down these buffers.
809	*/
810	static void
811	xlog_wait_iclog_completion(struct xlog *log)
812	{
813	int i;
814	struct xlog_in_core *iclog = log->l_iclog;
815
816	for (i = 0; i < log->l_iclog_bufs; i++) {
817	down(sem: &iclog->ic_sema);
818	up(sem: &iclog->ic_sema);
819	iclog = iclog->ic_next;
820	}
821	}
822
823	/*
824	* Wait for the iclog and all prior iclogs to be written disk as required by the
825	* log force state machine. Waiting on ic_force_wait ensures iclog completions
826	* have been ordered and callbacks run before we are woken here, hence
827	* guaranteeing that all the iclogs up to this one are on stable storage.
828	*/
829	int
830	xlog_wait_on_iclog(
831	struct xlog_in_core *iclog)
832	__releases(iclog->ic_log->l_icloglock)
833	{
834	struct xlog *log = iclog->ic_log;
835
836	trace_xlog_iclog_wait_on(iclog, _RET_IP_);
837	if (!xlog_is_shutdown(log) &&
838	iclog->ic_state != XLOG_STATE_ACTIVE &&
839	iclog->ic_state != XLOG_STATE_DIRTY) {
840	XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
841	xlog_wait(wq: &iclog->ic_force_wait, lock: &log->l_icloglock);
842	} else {
843	spin_unlock(lock: &log->l_icloglock);
844	}
845
846	if (xlog_is_shutdown(log))
847	return -EIO;
848	return 0;
849	}
850
851	/*
852	* Write out an unmount record using the ticket provided. We have to account for
853	* the data space used in the unmount ticket as this write is not done from a
854	* transaction context that has already done the accounting for us.
855	*/
856	static int
857	xlog_write_unmount_record(
858	struct xlog *log,
859	struct xlog_ticket *ticket)
860	{
861	struct {
862	struct xlog_op_header ophdr;
863	struct xfs_unmount_log_format ulf;
864	} unmount_rec = {
865	.ophdr = {
866	.oh_clientid = XFS_LOG,
867	.oh_tid = cpu_to_be32(ticket->t_tid),
868	.oh_flags = XLOG_UNMOUNT_TRANS,
869	},
870	.ulf = {
871	.magic = XLOG_UNMOUNT_TYPE,
872	},
873	};
874	struct xfs_log_iovec reg = {
875	.i_addr = &unmount_rec,
876	.i_len = sizeof(unmount_rec),
877	.i_type = XLOG_REG_TYPE_UNMOUNT,
878	};
879	struct xfs_log_vec vec = {
880	.lv_niovecs = 1,
881	.lv_iovecp = &reg,
882	};
883	LIST_HEAD(lv_chain);
884	list_add(new: &vec.lv_list, head: &lv_chain);
885
886	BUILD_BUG_ON((sizeof(struct xlog_op_header) +
887	sizeof(struct xfs_unmount_log_format)) !=
888	sizeof(unmount_rec));
889
890	/* account for space used by record data */
891	ticket->t_curr_res -= sizeof(unmount_rec);
892
893	return xlog_write(log, NULL, lv_chain: &lv_chain, tic: ticket, len: reg.i_len);
894	}
895
896	/*
897	* Mark the filesystem clean by writing an unmount record to the head of the
898	* log.
899	*/
900	static void
901	xlog_unmount_write(
902	struct xlog *log)
903	{
904	struct xfs_mount *mp = log->l_mp;
905	struct xlog_in_core *iclog;
906	struct xlog_ticket *tic = NULL;
907	int error;
908
909	error = xfs_log_reserve(mp, unit_bytes: 600, cnt: 1, ticp: &tic, permanent: 0);
910	if (error)
911	goto out_err;
912
913	error = xlog_write_unmount_record(log, ticket: tic);
914	/*
915	* At this point, we're umounting anyway, so there's no point in
916	* transitioning log state to shutdown. Just continue...
917	*/
918	out_err:
919	if (error)
920	xfs_alert(mp, "%s: unmount record failed", __func__);
921
922	spin_lock(lock: &log->l_icloglock);
923	iclog = log->l_iclog;
924	error = xlog_force_iclog(iclog);
925	xlog_wait_on_iclog(iclog);
926
927	if (tic) {
928	trace_xfs_log_umount_write(log, tic);
929	xfs_log_ticket_ungrant(log, ticket: tic);
930	}
931	}
932
933	static void
934	xfs_log_unmount_verify_iclog(
935	struct xlog *log)
936	{
937	struct xlog_in_core *iclog = log->l_iclog;
938
939	do {
940	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
941	ASSERT(iclog->ic_offset == 0);
942	} while ((iclog = iclog->ic_next) != log->l_iclog);
943	}
944
945	/*
946	* Unmount record used to have a string "Unmount filesystem--" in the
947	* data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
948	* We just write the magic number now since that particular field isn't
949	* currently architecture converted and "Unmount" is a bit foo.
950	* As far as I know, there weren't any dependencies on the old behaviour.
951	*/
952	static void
953	xfs_log_unmount_write(
954	struct xfs_mount *mp)
955	{
956	struct xlog *log = mp->m_log;
957
958	if (!xfs_log_writable(mp))
959	return;
960
961	xfs_log_force(mp, XFS_LOG_SYNC);
962
963	if (xlog_is_shutdown(log))
964	return;
965
966	/*
967	* If we think the summary counters are bad, avoid writing the unmount
968	* record to force log recovery at next mount, after which the summary
969	* counters will be recalculated. Refer to xlog_check_unmount_rec for
970	* more details.
971	*/
972	if (xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS) ||
973	XFS_TEST_ERROR(mp, XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
974	xfs_alert(mp, "%s: will fix summary counters at next mount",
975	__func__);
976	return;
977	}
978
979	xfs_log_unmount_verify_iclog(log);
980	xlog_unmount_write(log);
981	}
982
983	/*
984	* Empty the log for unmount/freeze.
985	*
986	* To do this, we first need to shut down the background log work so it is not
987	* trying to cover the log as we clean up. We then need to unpin all objects in
988	* the log so we can then flush them out. Once they have completed their IO and
989	* run the callbacks removing themselves from the AIL, we can cover the log.
990	*/
991	int
992	xfs_log_quiesce(
993	struct xfs_mount *mp)
994	{
995	/*
996	* Clear log incompat features since we're quiescing the log. Report
997	* failures, though it's not fatal to have a higher log feature
998	* protection level than the log contents actually require.
999	*/
1000	if (xfs_clear_incompat_log_features(mp)) {
1001	int error;
1002
1003	error = xfs_sync_sb(mp, false);
1004	if (error)
1005	xfs_warn(mp,
1006	"Failed to clear log incompat features on quiesce");
1007	}
1008
1009	cancel_delayed_work_sync(dwork: &mp->m_log->l_work);
1010	xfs_log_force(mp, XFS_LOG_SYNC);
1011
1012	/*
1013	* The superblock buffer is uncached and while xfs_ail_push_all_sync()
1014	* will push it, xfs_buftarg_wait() will not wait for it. Further,
1015	* xfs_buf_iowait() cannot be used because it was pushed with the
1016	* XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1017	* the IO to complete.
1018	*/
1019	xfs_ail_push_all_sync(ailp: mp->m_ail);
1020	xfs_buftarg_wait(mp->m_ddev_targp);
1021	xfs_buf_lock(mp->m_sb_bp);
1022	xfs_buf_unlock(mp->m_sb_bp);
1023
1024	return xfs_log_cover(mp);
1025	}
1026
1027	void
1028	xfs_log_clean(
1029	struct xfs_mount *mp)
1030	{
1031	xfs_log_quiesce(mp);
1032	xfs_log_unmount_write(mp);
1033	}
1034
1035	/*
1036	* Shut down and release the AIL and Log.
1037	*
1038	* During unmount, we need to ensure we flush all the dirty metadata objects
1039	* from the AIL so that the log is empty before we write the unmount record to
1040	* the log. Once this is done, we can tear down the AIL and the log.
1041	*/
1042	void
1043	xfs_log_unmount(
1044	struct xfs_mount *mp)
1045	{
1046	xfs_log_clean(mp);
1047
1048	/*
1049	* If shutdown has come from iclog IO context, the log
1050	* cleaning will have been skipped and so we need to wait
1051	* for the iclog to complete shutdown processing before we
1052	* tear anything down.
1053	*/
1054	xlog_wait_iclog_completion(log: mp->m_log);
1055
1056	xfs_buftarg_drain(mp->m_ddev_targp);
1057
1058	xfs_trans_ail_destroy(mp);
1059
1060	xfs_sysfs_del(kobj: &mp->m_log->l_kobj);
1061
1062	xlog_dealloc_log(log: mp->m_log);
1063	}
1064
1065	void
1066	xfs_log_item_init(
1067	struct xfs_mount *mp,
1068	struct xfs_log_item *item,
1069	int type,
1070	const struct xfs_item_ops *ops)
1071	{
1072	item->li_log = mp->m_log;
1073	item->li_ailp = mp->m_ail;
1074	item->li_type = type;
1075	item->li_ops = ops;
1076	item->li_lv = NULL;
1077
1078	INIT_LIST_HEAD(list: &item->li_ail);
1079	INIT_LIST_HEAD(list: &item->li_cil);
1080	INIT_LIST_HEAD(list: &item->li_bio_list);
1081	INIT_LIST_HEAD(list: &item->li_trans);
1082	}
1083
1084	/*
1085	* Wake up processes waiting for log space after we have moved the log tail.
1086	*/
1087	void
1088	xfs_log_space_wake(
1089	struct xfs_mount *mp)
1090	{
1091	struct xlog *log = mp->m_log;
1092	int free_bytes;
1093
1094	if (xlog_is_shutdown(log))
1095	return;
1096
1097	if (!list_empty_careful(head: &log->l_write_head.waiters)) {
1098	ASSERT(!xlog_in_recovery(log));
1099
1100	spin_lock(lock: &log->l_write_head.lock);
1101	free_bytes = xlog_grant_space_left(log, head: &log->l_write_head);
1102	xlog_grant_head_wake(log, head: &log->l_write_head, free_bytes: &free_bytes);
1103	spin_unlock(lock: &log->l_write_head.lock);
1104	}
1105
1106	if (!list_empty_careful(head: &log->l_reserve_head.waiters)) {
1107	ASSERT(!xlog_in_recovery(log));
1108
1109	spin_lock(lock: &log->l_reserve_head.lock);
1110	free_bytes = xlog_grant_space_left(log, head: &log->l_reserve_head);
1111	xlog_grant_head_wake(log, head: &log->l_reserve_head, free_bytes: &free_bytes);
1112	spin_unlock(lock: &log->l_reserve_head.lock);
1113	}
1114	}
1115
1116	/*
1117	* Determine if we have a transaction that has gone to disk that needs to be
1118	* covered. To begin the transition to the idle state firstly the log needs to
1119	* be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1120	* we start attempting to cover the log.
1121	*
1122	* Only if we are then in a state where covering is needed, the caller is
1123	* informed that dummy transactions are required to move the log into the idle
1124	* state.
1125	*
1126	* If there are any items in the AIl or CIL, then we do not want to attempt to
1127	* cover the log as we may be in a situation where there isn't log space
1128	* available to run a dummy transaction and this can lead to deadlocks when the
1129	* tail of the log is pinned by an item that is modified in the CIL. Hence
1130	* there's no point in running a dummy transaction at this point because we
1131	* can't start trying to idle the log until both the CIL and AIL are empty.
1132	*/
1133	static bool
1134	xfs_log_need_covered(
1135	struct xfs_mount *mp)
1136	{
1137	struct xlog *log = mp->m_log;
1138	bool needed = false;
1139
1140	if (!xlog_cil_empty(log))
1141	return false;
1142
1143	spin_lock(lock: &log->l_icloglock);
1144	switch (log->l_covered_state) {
1145	case XLOG_STATE_COVER_DONE:
1146	case XLOG_STATE_COVER_DONE2:
1147	case XLOG_STATE_COVER_IDLE:
1148	break;
1149	case XLOG_STATE_COVER_NEED:
1150	case XLOG_STATE_COVER_NEED2:
1151	if (xfs_ail_min_lsn(log->l_ailp))
1152	break;
1153	if (!xlog_iclogs_empty(log))
1154	break;
1155
1156	needed = true;
1157	if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1158	log->l_covered_state = XLOG_STATE_COVER_DONE;
1159	else
1160	log->l_covered_state = XLOG_STATE_COVER_DONE2;
1161	break;
1162	default:
1163	needed = true;
1164	break;
1165	}
1166	spin_unlock(lock: &log->l_icloglock);
1167	return needed;
1168	}
1169
1170	/*
1171	* Explicitly cover the log. This is similar to background log covering but
1172	* intended for usage in quiesce codepaths. The caller is responsible to ensure
1173	* the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1174	* must all be empty.
1175	*/
1176	static int
1177	xfs_log_cover(
1178	struct xfs_mount *mp)
1179	{
1180	int error = 0;
1181	bool need_covered;
1182
1183	if (!xlog_is_shutdown(log: mp->m_log)) {
1184	ASSERT(xlog_cil_empty(mp->m_log));
1185	ASSERT(xlog_iclogs_empty(mp->m_log));
1186	ASSERT(!xfs_ail_min_lsn(mp->m_log->l_ailp));
1187	}
1188
1189	if (!xfs_log_writable(mp))
1190	return 0;
1191
1192	/*
1193	* xfs_log_need_covered() is not idempotent because it progresses the
1194	* state machine if the log requires covering. Therefore, we must call
1195	* this function once and use the result until we've issued an sb sync.
1196	* Do so first to make that abundantly clear.
1197	*
1198	* Fall into the covering sequence if the log needs covering or the
1199	* mount has lazy superblock accounting to sync to disk. The sb sync
1200	* used for covering accumulates the in-core counters, so covering
1201	* handles this for us.
1202	*/
1203	need_covered = xfs_log_need_covered(mp);
1204	if (!need_covered && !xfs_has_lazysbcount(mp))
1205	return 0;
1206
1207	/*
1208	* To cover the log, commit the superblock twice (at most) in
1209	* independent checkpoints. The first serves as a reference for the
1210	* tail pointer. The sync transaction and AIL push empties the AIL and
1211	* updates the in-core tail to the LSN of the first checkpoint. The
1212	* second commit updates the on-disk tail with the in-core LSN,
1213	* covering the log. Push the AIL one more time to leave it empty, as
1214	* we found it.
1215	*/
1216	do {
1217	error = xfs_sync_sb(mp, true);
1218	if (error)
1219	break;
1220	xfs_ail_push_all_sync(ailp: mp->m_ail);
1221	} while (xfs_log_need_covered(mp));
1222
1223	return error;
1224	}
1225
1226	static void
1227	xlog_ioend_work(
1228	struct work_struct *work)
1229	{
1230	struct xlog_in_core *iclog =
1231	container_of(work, struct xlog_in_core, ic_end_io_work);
1232	struct xlog *log = iclog->ic_log;
1233	int error;
1234
1235	error = blk_status_to_errno(status: iclog->ic_bio.bi_status);
1236	#ifdef DEBUG
1237	/* treat writes with injected CRC errors as failed */
1238	if (iclog->ic_fail_crc)
1239	error = -EIO;
1240	#endif
1241
1242	/*
1243	* Race to shutdown the filesystem if we see an error.
1244	*/
1245	if (error || XFS_TEST_ERROR(log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1246	xfs_alert(log->l_mp, "log I/O error %d", error);
1247	xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1248	}
1249
1250	xlog_state_done_syncing(iclog);
1251	bio_uninit(&iclog->ic_bio);
1252
1253	/*
1254	* Drop the lock to signal that we are done. Nothing references the
1255	* iclog after this, so an unmount waiting on this lock can now tear it
1256	* down safely. As such, it is unsafe to reference the iclog after the
1257	* unlock as we could race with it being freed.
1258	*/
1259	up(sem: &iclog->ic_sema);
1260	}
1261
1262	/*
1263	* Return size of each in-core log record buffer.
1264	*
1265	* All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1266	*
1267	* If the filesystem blocksize is too large, we may need to choose a
1268	* larger size since the directory code currently logs entire blocks.
1269	*/
1270	STATIC void
1271	xlog_get_iclog_buffer_size(
1272	struct xfs_mount *mp,
1273	struct xlog *log)
1274	{
1275	if (mp->m_logbufs <= 0)
1276	mp->m_logbufs = XLOG_MAX_ICLOGS;
1277	if (mp->m_logbsize <= 0)
1278	mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1279
1280	log->l_iclog_bufs = mp->m_logbufs;
1281	log->l_iclog_size = mp->m_logbsize;
1282
1283	/*
1284	* Combined size of the log record headers. The first 32k cycles
1285	* are stored directly in the xlog_rec_header, the rest in the
1286	* variable number of xlog_rec_ext_headers at its end.
1287	*/
1288	log->l_iclog_hsize = struct_size(log->l_iclog->ic_header, h_ext,
1289	DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE) - 1);
1290	}
1291
1292	void
1293	xfs_log_work_queue(
1294	struct xfs_mount *mp)
1295	{
1296	queue_delayed_work(wq: mp->m_sync_workqueue, dwork: &mp->m_log->l_work,
1297	delay: msecs_to_jiffies(xfs_syncd_centisecs * 10));
1298	}
1299
1300	/*
1301	* Clear the log incompat flags if we have the opportunity.
1302	*
1303	* This only happens if we're about to log the second dummy transaction as part
1304	* of covering the log.
1305	*/
1306	static inline void
1307	xlog_clear_incompat(
1308	struct xlog *log)
1309	{
1310	struct xfs_mount *mp = log->l_mp;
1311
1312	if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1313	XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1314	return;
1315
1316	if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1317	return;
1318
1319	xfs_clear_incompat_log_features(mp);
1320	}
1321
1322	/*
1323	* Every sync period we need to unpin all items in the AIL and push them to
1324	* disk. If there is nothing dirty, then we might need to cover the log to
1325	* indicate that the filesystem is idle.
1326	*/
1327	static void
1328	xfs_log_worker(
1329	struct work_struct *work)
1330	{
1331	struct xlog *log = container_of(to_delayed_work(work),
1332	struct xlog, l_work);
1333	struct xfs_mount *mp = log->l_mp;
1334
1335	/* dgc: errors ignored - not fatal and nowhere to report them */
1336	if (xfs_fs_writable(mp, level: SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1337	/*
1338	* Dump a transaction into the log that contains no real change.
1339	* This is needed to stamp the current tail LSN into the log
1340	* during the covering operation.
1341	*
1342	* We cannot use an inode here for this - that will push dirty
1343	* state back up into the VFS and then periodic inode flushing
1344	* will prevent log covering from making progress. Hence we
1345	* synchronously log the superblock instead to ensure the
1346	* superblock is immediately unpinned and can be written back.
1347	*/
1348	xlog_clear_incompat(log);
1349	xfs_sync_sb(mp, true);
1350	} else
1351	xfs_log_force(mp, flags: 0);
1352
1353	/* start pushing all the metadata that is currently dirty */
1354	xfs_ail_push_all(ailp: mp->m_ail);
1355
1356	/* queue us up again */
1357	xfs_log_work_queue(mp);
1358	}
1359
1360	/*
1361	* This routine initializes some of the log structure for a given mount point.
1362	* Its primary purpose is to fill in enough, so recovery can occur. However,
1363	* some other stuff may be filled in too.
1364	*/
1365	STATIC struct xlog *
1366	xlog_alloc_log(
1367	struct xfs_mount *mp,
1368	struct xfs_buftarg *log_target,
1369	xfs_daddr_t blk_offset,
1370	int num_bblks)
1371	{
1372	struct xlog *log;
1373	struct xlog_in_core **iclogp;
1374	struct xlog_in_core *iclog, *prev_iclog = NULL;
1375	int i;
1376	int error = -ENOMEM;
1377	uint log2_size = 0;
1378
1379	log = kzalloc(sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1380	if (!log) {
1381	xfs_warn(mp, "Log allocation failed: No memory!");
1382	goto out;
1383	}
1384
1385	log->l_mp = mp;
1386	log->l_targ = log_target;
1387	log->l_logsize = BBTOB(num_bblks);
1388	log->l_logBBstart = blk_offset;
1389	log->l_logBBsize = num_bblks;
1390	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1391	set_bit(XLOG_ACTIVE_RECOVERY, addr: &log->l_opstate);
1392	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1393	INIT_LIST_HEAD(list: &log->r_dfops);
1394
1395	log->l_prev_block = -1;
1396	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1397	xlog_assign_atomic_lsn(lsn: &log->l_tail_lsn, cycle: 1, block: 0);
1398	log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1399
1400	if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1401	log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1402	else
1403	log->l_iclog_roundoff = BBSIZE;
1404
1405	xlog_grant_head_init(head: &log->l_reserve_head);
1406	xlog_grant_head_init(head: &log->l_write_head);
1407
1408	error = -EFSCORRUPTED;
1409	if (xfs_has_sector(mp)) {
1410	log2_size = mp->m_sb.sb_logsectlog;
1411	if (log2_size < BBSHIFT) {
1412	xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1413	log2_size, BBSHIFT);
1414	goto out_free_log;
1415	}
1416
1417	log2_size -= BBSHIFT;
1418	if (log2_size > mp->m_sectbb_log) {
1419	xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1420	log2_size, mp->m_sectbb_log);
1421	goto out_free_log;
1422	}
1423
1424	/* for larger sector sizes, must have v2 or external log */
1425	if (log2_size && log->l_logBBstart > 0 &&
1426	!xfs_has_logv2(mp)) {
1427	xfs_warn(mp,
1428	"log sector size (0x%x) invalid for configuration.",
1429	log2_size);
1430	goto out_free_log;
1431	}
1432	}
1433	log->l_sectBBsize = 1 << log2_size;
1434
1435	xlog_get_iclog_buffer_size(mp, log);
1436
1437	spin_lock_init(&log->l_icloglock);
1438	init_waitqueue_head(&log->l_flush_wait);
1439
1440	iclogp = &log->l_iclog;
1441	ASSERT(log->l_iclog_size >= 4096);
1442	for (i = 0; i < log->l_iclog_bufs; i++) {
1443	size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1444	sizeof(struct bio_vec);
1445
1446	iclog = kzalloc(sizeof(*iclog) + bvec_size,
1447	GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1448	if (!iclog)
1449	goto out_free_iclog;
1450
1451	*iclogp = iclog;
1452	iclog->ic_prev = prev_iclog;
1453	prev_iclog = iclog;
1454
1455	iclog->ic_header = kvzalloc(log->l_iclog_size,
1456	GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1457	if (!iclog->ic_header)
1458	goto out_free_iclog;
1459	iclog->ic_header->h_magicno =
1460	cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1461	iclog->ic_header->h_version = cpu_to_be32(
1462	xfs_has_logv2(log->l_mp) ? 2 : 1);
1463	iclog->ic_header->h_size = cpu_to_be32(log->l_iclog_size);
1464	iclog->ic_header->h_fmt = cpu_to_be32(XLOG_FMT);
1465	memcpy(&iclog->ic_header->h_fs_uuid, &mp->m_sb.sb_uuid,
1466	sizeof(iclog->ic_header->h_fs_uuid));
1467
1468	iclog->ic_datap = (void *)iclog->ic_header + log->l_iclog_hsize;
1469	iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1470	iclog->ic_state = XLOG_STATE_ACTIVE;
1471	iclog->ic_log = log;
1472	atomic_set(v: &iclog->ic_refcnt, i: 0);
1473	INIT_LIST_HEAD(list: &iclog->ic_callbacks);
1474
1475	init_waitqueue_head(&iclog->ic_force_wait);
1476	init_waitqueue_head(&iclog->ic_write_wait);
1477	INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1478	sema_init(sem: &iclog->ic_sema, val: 1);
1479
1480	iclogp = &iclog->ic_next;
1481	}
1482	*iclogp = log->l_iclog; /* complete ring */
1483	log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1484
1485	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1486	XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_PERCPU),
1487	0, mp->m_super->s_id);
1488	if (!log->l_ioend_workqueue)
1489	goto out_free_iclog;
1490
1491	error = xlog_cil_init(log);
1492	if (error)
1493	goto out_destroy_workqueue;
1494	return log;
1495
1496	out_destroy_workqueue:
1497	destroy_workqueue(wq: log->l_ioend_workqueue);
1498	out_free_iclog:
1499	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1500	prev_iclog = iclog->ic_next;
1501	kvfree(addr: iclog->ic_header);
1502	kfree(objp: iclog);
1503	if (prev_iclog == log->l_iclog)
1504	break;
1505	}
1506	out_free_log:
1507	kfree(objp: log);
1508	out:
1509	return ERR_PTR(error);
1510	} /* xlog_alloc_log */
1511
1512	/*
1513	* Stamp cycle number in every block
1514	*/
1515	STATIC void
1516	xlog_pack_data(
1517	struct xlog *log,
1518	struct xlog_in_core *iclog,
1519	int roundoff)
1520	{
1521	struct xlog_rec_header *rhead = iclog->ic_header;
1522	__be32 cycle_lsn = CYCLE_LSN_DISK(rhead->h_lsn);
1523	char *dp = iclog->ic_datap;
1524	int i;
1525
1526	for (i = 0; i < BTOBB(iclog->ic_offset + roundoff); i++) {
1527	*xlog_cycle_data(rhead, i) = *(__be32 *)dp;
1528	*(__be32 *)dp = cycle_lsn;
1529	dp += BBSIZE;
1530	}
1531
1532	for (i = 0; i < (log->l_iclog_hsize >> BBSHIFT) - 1; i++)
1533	rhead->h_ext[i].xh_cycle = cycle_lsn;
1534	}
1535
1536	/*
1537	* Calculate the checksum for a log buffer.
1538	*
1539	* This is a little more complicated than it should be because the various
1540	* headers and the actual data are non-contiguous.
1541	*/
1542	__le32
1543	xlog_cksum(
1544	struct xlog *log,
1545	struct xlog_rec_header *rhead,
1546	char *dp,
1547	unsigned int hdrsize,
1548	unsigned int size)
1549	{
1550	uint32_t crc;
1551
1552	/* first generate the crc for the record header ... */
1553	crc = xfs_start_cksum_update((char *)rhead, hdrsize,
1554	offsetof(struct xlog_rec_header, h_crc));
1555
1556	/* ... then for additional cycle data for v2 logs ... */
1557	if (xfs_has_logv2(mp: log->l_mp)) {
1558	int xheads, i;
1559
1560	xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE) - 1;
1561	for (i = 0; i < xheads; i++)
1562	crc = crc32c(crc, &rhead->h_ext[i], XLOG_REC_EXT_SIZE);
1563	}
1564
1565	/* ... and finally for the payload */
1566	crc = crc32c(crc, p: dp, len: size);
1567
1568	return xfs_end_cksum(crc);
1569	}
1570
1571	static void
1572	xlog_bio_end_io(
1573	struct bio *bio)
1574	{
1575	struct xlog_in_core *iclog = bio->bi_private;
1576
1577	queue_work(wq: iclog->ic_log->l_ioend_workqueue,
1578	work: &iclog->ic_end_io_work);
1579	}
1580
1581	STATIC void
1582	xlog_write_iclog(
1583	struct xlog *log,
1584	struct xlog_in_core *iclog,
1585	uint64_t bno,
1586	unsigned int count)
1587	{
1588	ASSERT(bno < log->l_logBBsize);
1589	trace_xlog_iclog_write(iclog, _RET_IP_);
1590
1591	/*
1592	* We lock the iclogbufs here so that we can serialise against I/O
1593	* completion during unmount. We might be processing a shutdown
1594	* triggered during unmount, and that can occur asynchronously to the
1595	* unmount thread, and hence we need to ensure that completes before
1596	* tearing down the iclogbufs. Hence we need to hold the buffer lock
1597	* across the log IO to archieve that.
1598	*/
1599	down(sem: &iclog->ic_sema);
1600	if (xlog_is_shutdown(log)) {
1601	/*
1602	* It would seem logical to return EIO here, but we rely on
1603	* the log state machine to propagate I/O errors instead of
1604	* doing it here. We kick of the state machine and unlock
1605	* the buffer manually, the code needs to be kept in sync
1606	* with the I/O completion path.
1607	*/
1608	goto sync;
1609	}
1610
1611	/*
1612	* We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1613	* IOs coming immediately after this one. This prevents the block layer
1614	* writeback throttle from throttling log writes behind background
1615	* metadata writeback and causing priority inversions.
1616	*/
1617	bio_init(bio: &iclog->ic_bio, bdev: log->l_targ->bt_bdev, table: iclog->ic_bvec,
1618	howmany(count, PAGE_SIZE),
1619	opf: REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1620	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1621	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1622	iclog->ic_bio.bi_private = iclog;
1623
1624	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1625	iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1626	/*
1627	* For external log devices, we also need to flush the data
1628	* device cache first to ensure all metadata writeback covered
1629	* by the LSN in this iclog is on stable storage. This is slow,
1630	* but it *must* complete before we issue the external log IO.
1631	*
1632	* If the flush fails, we cannot conclude that past metadata
1633	* writeback from the log succeeded. Repeating the flush is
1634	* not possible, hence we must shut down with log IO error to
1635	* avoid shutdown re-entering this path and erroring out again.
1636	*/
1637	if (log->l_targ != log->l_mp->m_ddev_targp &&
1638	blkdev_issue_flush(bdev: log->l_mp->m_ddev_targp->bt_bdev))
1639	goto shutdown;
1640	}
1641	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1642	iclog->ic_bio.bi_opf |= REQ_FUA;
1643
1644	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1645
1646	if (is_vmalloc_addr(x: iclog->ic_header)) {
1647	if (!bio_add_vmalloc(bio: &iclog->ic_bio, vaddr: iclog->ic_header, len: count))
1648	goto shutdown;
1649	} else {
1650	bio_add_virt_nofail(bio: &iclog->ic_bio, vaddr: iclog->ic_header, len: count);
1651	}
1652
1653	/*
1654	* If this log buffer would straddle the end of the log we will have
1655	* to split it up into two bios, so that we can continue at the start.
1656	*/
1657	if (bno + BTOBB(count) > log->l_logBBsize) {
1658	struct bio *split;
1659
1660	split = bio_split(bio: &iclog->ic_bio, sectors: log->l_logBBsize - bno,
1661	GFP_NOIO, bs: &fs_bio_set);
1662	bio_chain(split, &iclog->ic_bio);
1663	submit_bio(bio: split);
1664
1665	/* restart at logical offset zero for the remainder */
1666	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1667	}
1668
1669	submit_bio(bio: &iclog->ic_bio);
1670	return;
1671	shutdown:
1672	xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1673	sync:
1674	xlog_state_done_syncing(iclog);
1675	up(sem: &iclog->ic_sema);
1676	}
1677
1678	/*
1679	* We need to bump cycle number for the part of the iclog that is
1680	* written to the start of the log. Watch out for the header magic
1681	* number case, though.
1682	*/
1683	static void
1684	xlog_split_iclog(
1685	struct xlog *log,
1686	void *data,
1687	uint64_t bno,
1688	unsigned int count)
1689	{
1690	unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1691	unsigned int i;
1692
1693	for (i = split_offset; i < count; i += BBSIZE) {
1694	uint32_t cycle = get_unaligned_be32(p: data + i);
1695
1696	if (++cycle == XLOG_HEADER_MAGIC_NUM)
1697	cycle++;
1698	put_unaligned_be32(val: cycle, p: data + i);
1699	}
1700	}
1701
1702	static int
1703	xlog_calc_iclog_size(
1704	struct xlog *log,
1705	struct xlog_in_core *iclog,
1706	uint32_t *roundoff)
1707	{
1708	uint32_t count_init, count;
1709
1710	/* Add for LR header */
1711	count_init = log->l_iclog_hsize + iclog->ic_offset;
1712	count = roundup(count_init, log->l_iclog_roundoff);
1713
1714	*roundoff = count - count_init;
1715
1716	ASSERT(count >= count_init);
1717	ASSERT(*roundoff < log->l_iclog_roundoff);
1718	return count;
1719	}
1720
1721	/*
1722	* Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1723	* fashion. Previously, we should have moved the current iclog
1724	* ptr in the log to point to the next available iclog. This allows further
1725	* write to continue while this code syncs out an iclog ready to go.
1726	* Before an in-core log can be written out, the data section must be scanned
1727	* to save away the 1st word of each BBSIZE block into the header. We replace
1728	* it with the current cycle count. Each BBSIZE block is tagged with the
1729	* cycle count because there in an implicit assumption that drives will
1730	* guarantee that entire 512 byte blocks get written at once. In other words,
1731	* we can't have part of a 512 byte block written and part not written. By
1732	* tagging each block, we will know which blocks are valid when recovering
1733	* after an unclean shutdown.
1734	*
1735	* This routine is single threaded on the iclog. No other thread can be in
1736	* this routine with the same iclog. Changing contents of iclog can there-
1737	* fore be done without grabbing the state machine lock. Updating the global
1738	* log will require grabbing the lock though.
1739	*
1740	* The entire log manager uses a logical block numbering scheme. Only
1741	* xlog_write_iclog knows about the fact that the log may not start with
1742	* block zero on a given device.
1743	*/
1744	STATIC void
1745	xlog_sync(
1746	struct xlog *log,
1747	struct xlog_in_core *iclog,
1748	struct xlog_ticket *ticket)
1749	{
1750	unsigned int count; /* byte count of bwrite */
1751	unsigned int roundoff; /* roundoff to BB or stripe */
1752	uint64_t bno;
1753	unsigned int size;
1754
1755	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1756	trace_xlog_iclog_sync(iclog, _RET_IP_);
1757
1758	count = xlog_calc_iclog_size(log, iclog, roundoff: &roundoff);
1759
1760	/*
1761	* If we have a ticket, account for the roundoff via the ticket
1762	* reservation to avoid touching the hot grant heads needlessly.
1763	* Otherwise, we have to move grant heads directly.
1764	*/
1765	if (ticket) {
1766	ticket->t_curr_res -= roundoff;
1767	} else {
1768	xlog_grant_add_space(head: &log->l_reserve_head, bytes: roundoff);
1769	xlog_grant_add_space(head: &log->l_write_head, bytes: roundoff);
1770	}
1771
1772	/* put cycle number in every block */
1773	xlog_pack_data(log, iclog, roundoff);
1774
1775	/* real byte length */
1776	size = iclog->ic_offset;
1777	if (xfs_has_logv2(mp: log->l_mp))
1778	size += roundoff;
1779	iclog->ic_header->h_len = cpu_to_be32(size);
1780
1781	XFS_STATS_INC(log->l_mp, xs_log_writes);
1782	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1783
1784	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header->h_lsn));
1785
1786	/* Do we need to split this write into 2 parts? */
1787	if (bno + BTOBB(count) > log->l_logBBsize)
1788	xlog_split_iclog(log, data: iclog->ic_header, bno, count);
1789
1790	/* calculcate the checksum */
1791	iclog->ic_header->h_crc = xlog_cksum(log, iclog->ic_header,
1792	iclog->ic_datap, XLOG_REC_SIZE, size);
1793	/*
1794	* Intentionally corrupt the log record CRC based on the error injection
1795	* frequency, if defined. This facilitates testing log recovery in the
1796	* event of torn writes. Hence, set the IOABORT state to abort the log
1797	* write on I/O completion and shutdown the fs. The subsequent mount
1798	* detects the bad CRC and attempts to recover.
1799	*/
1800	#ifdef DEBUG
1801	if (XFS_TEST_ERROR(log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1802	iclog->ic_header->h_crc &= cpu_to_le32(0xAAAAAAAA);
1803	iclog->ic_fail_crc = true;
1804	xfs_warn(log->l_mp,
1805	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1806	be64_to_cpu(iclog->ic_header->h_lsn));
1807	}
1808	#endif
1809	xlog_verify_iclog(log, iclog, count);
1810	xlog_write_iclog(log, iclog, bno, count);
1811	}
1812
1813	/*
1814	* Deallocate a log structure
1815	*/
1816	STATIC void
1817	xlog_dealloc_log(
1818	struct xlog *log)
1819	{
1820	struct xlog_in_core *iclog, *next_iclog;
1821	int i;
1822
1823	/*
1824	* Destroy the CIL after waiting for iclog IO completion because an
1825	* iclog EIO error will try to shut down the log, which accesses the
1826	* CIL to wake up the waiters.
1827	*/
1828	xlog_cil_destroy(log);
1829
1830	iclog = log->l_iclog;
1831	for (i = 0; i < log->l_iclog_bufs; i++) {
1832	next_iclog = iclog->ic_next;
1833	kvfree(addr: iclog->ic_header);
1834	kfree(objp: iclog);
1835	iclog = next_iclog;
1836	}
1837
1838	log->l_mp->m_log = NULL;
1839	destroy_workqueue(wq: log->l_ioend_workqueue);
1840	kfree(objp: log);
1841	}
1842
1843	/*
1844	* Update counters atomically now that memcpy is done.
1845	*/
1846	static inline void
1847	xlog_state_finish_copy(
1848	struct xlog *log,
1849	struct xlog_in_core *iclog,
1850	int record_cnt,
1851	int copy_bytes)
1852	{
1853	lockdep_assert_held(&log->l_icloglock);
1854
1855	be32_add_cpu(var: &iclog->ic_header->h_num_logops, val: record_cnt);
1856	iclog->ic_offset += copy_bytes;
1857	}
1858
1859	/*
1860	* print out info relating to regions written which consume
1861	* the reservation
1862	*/
1863	void
1864	xlog_print_tic_res(
1865	struct xfs_mount *mp,
1866	struct xlog_ticket *ticket)
1867	{
1868	xfs_warn(mp, "ticket reservation summary:");
1869	xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
1870	xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
1871	xfs_warn(mp, " original count = %d", ticket->t_ocnt);
1872	xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
1873	}
1874
1875	/*
1876	* Print a summary of the transaction.
1877	*/
1878	void
1879	xlog_print_trans(
1880	struct xfs_trans *tp)
1881	{
1882	struct xfs_mount *mp = tp->t_mountp;
1883	struct xfs_log_item *lip;
1884
1885	/* dump core transaction and ticket info */
1886	xfs_warn(mp, "transaction summary:");
1887	xfs_warn(mp, " log res = %d", tp->t_log_res);
1888	xfs_warn(mp, " log count = %d", tp->t_log_count);
1889	xfs_warn(mp, " flags = 0x%x", tp->t_flags);
1890
1891	xlog_print_tic_res(mp, ticket: tp->t_ticket);
1892
1893	/* dump each log item */
1894	list_for_each_entry(lip, &tp->t_items, li_trans) {
1895	struct xfs_log_vec *lv = lip->li_lv;
1896	struct xfs_log_iovec *vec;
1897	int i;
1898
1899	xfs_warn(mp, "log item: ");
1900	xfs_warn(mp, " type = 0x%x", lip->li_type);
1901	xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
1902	if (!lv)
1903	continue;
1904	xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
1905	xfs_warn(mp, " alloc_size = %d", lv->lv_alloc_size);
1906	xfs_warn(mp, " bytes = %d", lv->lv_bytes);
1907	xfs_warn(mp, " buf used= %d", lv->lv_buf_used);
1908
1909	/* dump each iovec for the log item */
1910	vec = lv->lv_iovecp;
1911	for (i = 0; i < lv->lv_niovecs; i++) {
1912	int dumplen = min(vec->i_len, 32);
1913
1914	xfs_warn(mp, " iovec[%d]", i);
1915	xfs_warn(mp, " type = 0x%x", vec->i_type);
1916	xfs_warn(mp, " len = %d", vec->i_len);
1917	xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
1918	xfs_hex_dump(p: vec->i_addr, length: dumplen);
1919
1920	vec++;
1921	}
1922	}
1923	}
1924
1925	static inline void
1926	xlog_write_iovec(
1927	struct xlog_in_core *iclog,
1928	uint32_t *log_offset,
1929	void *data,
1930	uint32_t write_len,
1931	int *bytes_left,
1932	uint32_t *record_cnt,
1933	uint32_t *data_cnt)
1934	{
1935	ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
1936	ASSERT(*log_offset % sizeof(int32_t) == 0);
1937	ASSERT(write_len % sizeof(int32_t) == 0);
1938
1939	memcpy(iclog->ic_datap + *log_offset, data, write_len);
1940	*log_offset += write_len;
1941	*bytes_left -= write_len;
1942	(*record_cnt)++;
1943	*data_cnt += write_len;
1944	}
1945
1946	/*
1947	* Write log vectors into a single iclog which is guaranteed by the caller
1948	* to have enough space to write the entire log vector into.
1949	*/
1950	static void
1951	xlog_write_full(
1952	struct xfs_log_vec *lv,
1953	struct xlog_ticket *ticket,
1954	struct xlog_in_core *iclog,
1955	uint32_t *log_offset,
1956	uint32_t *len,
1957	uint32_t *record_cnt,
1958	uint32_t *data_cnt)
1959	{
1960	int index;
1961
1962	ASSERT(*log_offset + *len <= iclog->ic_size ||
1963	iclog->ic_state == XLOG_STATE_WANT_SYNC);
1964
1965	/*
1966	* Ordered log vectors have no regions to write so this
1967	* loop will naturally skip them.
1968	*/
1969	for (index = 0; index < lv->lv_niovecs; index++) {
1970	struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
1971	struct xlog_op_header *ophdr = reg->i_addr;
1972
1973	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1974	xlog_write_iovec(iclog, log_offset, data: reg->i_addr,
1975	write_len: reg->i_len, bytes_left: len, record_cnt, data_cnt);
1976	}
1977	}
1978
1979	static int
1980	xlog_write_get_more_iclog_space(
1981	struct xlog_ticket *ticket,
1982	struct xlog_in_core **iclogp,
1983	uint32_t *log_offset,
1984	uint32_t len,
1985	uint32_t *record_cnt,
1986	uint32_t *data_cnt)
1987	{
1988	struct xlog_in_core *iclog = *iclogp;
1989	struct xlog *log = iclog->ic_log;
1990	int error;
1991
1992	spin_lock(lock: &log->l_icloglock);
1993	ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
1994	xlog_state_finish_copy(log, iclog, record_cnt: *record_cnt, copy_bytes: *data_cnt);
1995	error = xlog_state_release_iclog(log, iclog, ticket);
1996	spin_unlock(lock: &log->l_icloglock);
1997	if (error)
1998	return error;
1999
2000	error = xlog_state_get_iclog_space(log, len, iclog: &iclog, ticket,
2001	logoffsetp: log_offset);
2002	if (error)
2003	return error;
2004	*record_cnt = 0;
2005	*data_cnt = 0;
2006	*iclogp = iclog;
2007	return 0;
2008	}
2009
2010	/*
2011	* Write log vectors into a single iclog which is smaller than the current chain
2012	* length. We write until we cannot fit a full record into the remaining space
2013	* and then stop. We return the log vector that is to be written that cannot
2014	* wholly fit in the iclog.
2015	*/
2016	static int
2017	xlog_write_partial(
2018	struct xfs_log_vec *lv,
2019	struct xlog_ticket *ticket,
2020	struct xlog_in_core **iclogp,
2021	uint32_t *log_offset,
2022	uint32_t *len,
2023	uint32_t *record_cnt,
2024	uint32_t *data_cnt)
2025	{
2026	struct xlog_in_core *iclog = *iclogp;
2027	struct xlog_op_header *ophdr;
2028	int index = 0;
2029	uint32_t rlen;
2030	int error;
2031
2032	/* walk the logvec, copying until we run out of space in the iclog */
2033	for (index = 0; index < lv->lv_niovecs; index++) {
2034	struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2035	uint32_t reg_offset = 0;
2036
2037	/*
2038	* The first region of a continuation must have a non-zero
2039	* length otherwise log recovery will just skip over it and
2040	* start recovering from the next opheader it finds. Because we
2041	* mark the next opheader as a continuation, recovery will then
2042	* incorrectly add the continuation to the previous region and
2043	* that breaks stuff.
2044	*
2045	* Hence if there isn't space for region data after the
2046	* opheader, then we need to start afresh with a new iclog.
2047	*/
2048	if (iclog->ic_size - *log_offset <=
2049	sizeof(struct xlog_op_header)) {
2050	error = xlog_write_get_more_iclog_space(ticket,
2051	iclogp: &iclog, log_offset, len: *len, record_cnt,
2052	data_cnt);
2053	if (error)
2054	return error;
2055	}
2056
2057	ophdr = reg->i_addr;
2058	rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2059
2060	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2061	ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2062	if (rlen != reg->i_len)
2063	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2064
2065	xlog_write_iovec(iclog, log_offset, data: reg->i_addr,
2066	write_len: rlen, bytes_left: len, record_cnt, data_cnt);
2067
2068	/* If we wrote the whole region, move to the next. */
2069	if (rlen == reg->i_len)
2070	continue;
2071
2072	/*
2073	* We now have a partially written iovec, but it can span
2074	* multiple iclogs so we loop here. First we release the iclog
2075	* we currently have, then we get a new iclog and add a new
2076	* opheader. Then we continue copying from where we were until
2077	* we either complete the iovec or fill the iclog. If we
2078	* complete the iovec, then we increment the index and go right
2079	* back to the top of the outer loop. if we fill the iclog, we
2080	* run the inner loop again.
2081	*
2082	* This is complicated by the tail of a region using all the
2083	* space in an iclog and hence requiring us to release the iclog
2084	* and get a new one before returning to the outer loop. We must
2085	* always guarantee that we exit this inner loop with at least
2086	* space for log transaction opheaders left in the current
2087	* iclog, hence we cannot just terminate the loop at the end
2088	* of the of the continuation. So we loop while there is no
2089	* space left in the current iclog, and check for the end of the
2090	* continuation after getting a new iclog.
2091	*/
2092	do {
2093	/*
2094	* Ensure we include the continuation opheader in the
2095	* space we need in the new iclog by adding that size
2096	* to the length we require. This continuation opheader
2097	* needs to be accounted to the ticket as the space it
2098	* consumes hasn't been accounted to the lv we are
2099	* writing.
2100	*/
2101	error = xlog_write_get_more_iclog_space(ticket,
2102	&iclog, log_offset,
2103	*len + sizeof(struct xlog_op_header),
2104	record_cnt, data_cnt);
2105	if (error)
2106	return error;
2107
2108	ophdr = iclog->ic_datap + *log_offset;
2109	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2110	ophdr->oh_clientid = XFS_TRANSACTION;
2111	ophdr->oh_res2 = 0;
2112	ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2113
2114	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2115	*log_offset += sizeof(struct xlog_op_header);
2116	*data_cnt += sizeof(struct xlog_op_header);
2117
2118	/*
2119	* If rlen fits in the iclog, then end the region
2120	* continuation. Otherwise we're going around again.
2121	*/
2122	reg_offset += rlen;
2123	rlen = reg->i_len - reg_offset;
2124	if (rlen <= iclog->ic_size - *log_offset)
2125	ophdr->oh_flags |= XLOG_END_TRANS;
2126	else
2127	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2128
2129	rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2130	ophdr->oh_len = cpu_to_be32(rlen);
2131
2132	xlog_write_iovec(iclog, log_offset,
2133	data: reg->i_addr + reg_offset,
2134	write_len: rlen, bytes_left: len, record_cnt, data_cnt);
2135
2136	} while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2137	}
2138
2139	/*
2140	* No more iovecs remain in this logvec so return the next log vec to
2141	* the caller so it can go back to fast path copying.
2142	*/
2143	*iclogp = iclog;
2144	return 0;
2145	}
2146
2147	/*
2148	* Write some region out to in-core log
2149	*
2150	* This will be called when writing externally provided regions or when
2151	* writing out a commit record for a given transaction.
2152	*
2153	* General algorithm:
2154	* 1. Find total length of this write. This may include adding to the
2155	* lengths passed in.
2156	* 2. Check whether we violate the tickets reservation.
2157	* 3. While writing to this iclog
2158	* A. Reserve as much space in this iclog as can get
2159	* B. If this is first write, save away start lsn
2160	* C. While writing this region:
2161	* 1. If first write of transaction, write start record
2162	* 2. Write log operation header (header per region)
2163	* 3. Find out if we can fit entire region into this iclog
2164	* 4. Potentially, verify destination memcpy ptr
2165	* 5. Memcpy (partial) region
2166	* 6. If partial copy, release iclog; otherwise, continue
2167	* copying more regions into current iclog
2168	* 4. Mark want sync bit (in simulation mode)
2169	* 5. Release iclog for potential flush to on-disk log.
2170	*
2171	* ERRORS:
2172	* 1. Panic if reservation is overrun. This should never happen since
2173	* reservation amounts are generated internal to the filesystem.
2174	* NOTES:
2175	* 1. Tickets are single threaded data structures.
2176	* 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2177	* syncing routine. When a single log_write region needs to span
2178	* multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2179	* on all log operation writes which don't contain the end of the
2180	* region. The XLOG_END_TRANS bit is used for the in-core log
2181	* operation which contains the end of the continued log_write region.
2182	* 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2183	* we don't really know exactly how much space will be used. As a result,
2184	* we don't update ic_offset until the end when we know exactly how many
2185	* bytes have been written out.
2186	*/
2187	int
2188	xlog_write(
2189	struct xlog *log,
2190	struct xfs_cil_ctx *ctx,
2191	struct list_head *lv_chain,
2192	struct xlog_ticket *ticket,
2193	uint32_t len)
2194
2195	{
2196	struct xlog_in_core *iclog = NULL;
2197	struct xfs_log_vec *lv;
2198	uint32_t record_cnt = 0;
2199	uint32_t data_cnt = 0;
2200	int error = 0;
2201	int log_offset;
2202
2203	if (ticket->t_curr_res < 0) {
2204	xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2205	"ctx ticket reservation ran out. Need to up reservation");
2206	xlog_print_tic_res(mp: log->l_mp, ticket);
2207	xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2208	}
2209
2210	error = xlog_state_get_iclog_space(log, len, iclog: &iclog, ticket,
2211	logoffsetp: &log_offset);
2212	if (error)
2213	return error;
2214
2215	ASSERT(log_offset <= iclog->ic_size - 1);
2216
2217	/*
2218	* If we have a context pointer, pass it the first iclog we are
2219	* writing to so it can record state needed for iclog write
2220	* ordering.
2221	*/
2222	if (ctx)
2223	xlog_cil_set_ctx_write_state(ctx, iclog);
2224
2225	list_for_each_entry(lv, lv_chain, lv_list) {
2226	/*
2227	* If the entire log vec does not fit in the iclog, punt it to
2228	* the partial copy loop which can handle this case.
2229	*/
2230	if (lv->lv_niovecs &&
2231	lv->lv_bytes > iclog->ic_size - log_offset) {
2232	error = xlog_write_partial(lv, ticket, iclogp: &iclog,
2233	log_offset: &log_offset, len: &len, record_cnt: &record_cnt,
2234	data_cnt: &data_cnt);
2235	if (error) {
2236	/*
2237	* We have no iclog to release, so just return
2238	* the error immediately.
2239	*/
2240	return error;
2241	}
2242	} else {
2243	xlog_write_full(lv, ticket, iclog, log_offset: &log_offset,
2244	len: &len, record_cnt: &record_cnt, data_cnt: &data_cnt);
2245	}
2246	}
2247	ASSERT(len == 0);
2248
2249	/*
2250	* We've already been guaranteed that the last writes will fit inside
2251	* the current iclog, and hence it will already have the space used by
2252	* those writes accounted to it. Hence we do not need to update the
2253	* iclog with the number of bytes written here.
2254	*/
2255	spin_lock(lock: &log->l_icloglock);
2256	xlog_state_finish_copy(log, iclog, record_cnt, copy_bytes: 0);
2257	error = xlog_state_release_iclog(log, iclog, ticket);
2258	spin_unlock(lock: &log->l_icloglock);
2259
2260	return error;
2261	}
2262
2263	static void
2264	xlog_state_activate_iclog(
2265	struct xlog_in_core *iclog,
2266	int *iclogs_changed)
2267	{
2268	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2269	trace_xlog_iclog_activate(iclog, _RET_IP_);
2270
2271	/*
2272	* If the number of ops in this iclog indicate it just contains the
2273	* dummy transaction, we can change state into IDLE (the second time
2274	* around). Otherwise we should change the state into NEED a dummy.
2275	* We don't need to cover the dummy.
2276	*/
2277	if (*iclogs_changed == 0 &&
2278	iclog->ic_header->h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2279	*iclogs_changed = 1;
2280	} else {
2281	/*
2282	* We have two dirty iclogs so start over. This could also be
2283	* num of ops indicating this is not the dummy going out.
2284	*/
2285	*iclogs_changed = 2;
2286	}
2287
2288	iclog->ic_state = XLOG_STATE_ACTIVE;
2289	iclog->ic_offset = 0;
2290	iclog->ic_header->h_num_logops = 0;
2291	memset(iclog->ic_header->h_cycle_data, 0,
2292	sizeof(iclog->ic_header->h_cycle_data));
2293	iclog->ic_header->h_lsn = 0;
2294	iclog->ic_header->h_tail_lsn = 0;
2295	}
2296
2297	/*
2298	* Loop through all iclogs and mark all iclogs currently marked DIRTY as
2299	* ACTIVE after iclog I/O has completed.
2300	*/
2301	static void
2302	xlog_state_activate_iclogs(
2303	struct xlog *log,
2304	int *iclogs_changed)
2305	{
2306	struct xlog_in_core *iclog = log->l_iclog;
2307
2308	do {
2309	if (iclog->ic_state == XLOG_STATE_DIRTY)
2310	xlog_state_activate_iclog(iclog, iclogs_changed);
2311	/*
2312	* The ordering of marking iclogs ACTIVE must be maintained, so
2313	* an iclog doesn't become ACTIVE beyond one that is SYNCING.
2314	*/
2315	else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2316	break;
2317	} while ((iclog = iclog->ic_next) != log->l_iclog);
2318	}
2319
2320	static int
2321	xlog_covered_state(
2322	int prev_state,
2323	int iclogs_changed)
2324	{
2325	/*
2326	* We go to NEED for any non-covering writes. We go to NEED2 if we just
2327	* wrote the first covering record (DONE). We go to IDLE if we just
2328	* wrote the second covering record (DONE2) and remain in IDLE until a
2329	* non-covering write occurs.
2330	*/
2331	switch (prev_state) {
2332	case XLOG_STATE_COVER_IDLE:
2333	if (iclogs_changed == 1)
2334	return XLOG_STATE_COVER_IDLE;
2335	fallthrough;
2336	case XLOG_STATE_COVER_NEED:
2337	case XLOG_STATE_COVER_NEED2:
2338	break;
2339	case XLOG_STATE_COVER_DONE:
2340	if (iclogs_changed == 1)
2341	return XLOG_STATE_COVER_NEED2;
2342	break;
2343	case XLOG_STATE_COVER_DONE2:
2344	if (iclogs_changed == 1)
2345	return XLOG_STATE_COVER_IDLE;
2346	break;
2347	default:
2348	ASSERT(0);
2349	}
2350
2351	return XLOG_STATE_COVER_NEED;
2352	}
2353
2354	STATIC void
2355	xlog_state_clean_iclog(
2356	struct xlog *log,
2357	struct xlog_in_core *dirty_iclog)
2358	{
2359	int iclogs_changed = 0;
2360
2361	trace_xlog_iclog_clean(iclog: dirty_iclog, _RET_IP_);
2362
2363	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2364
2365	xlog_state_activate_iclogs(log, iclogs_changed: &iclogs_changed);
2366	wake_up_all(&dirty_iclog->ic_force_wait);
2367
2368	if (iclogs_changed) {
2369	log->l_covered_state = xlog_covered_state(prev_state: log->l_covered_state,
2370	iclogs_changed);
2371	}
2372	}
2373
2374	STATIC xfs_lsn_t
2375	xlog_get_lowest_lsn(
2376	struct xlog *log)
2377	{
2378	struct xlog_in_core *iclog = log->l_iclog;
2379	xfs_lsn_t lowest_lsn = 0, lsn;
2380
2381	do {
2382	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2383	iclog->ic_state == XLOG_STATE_DIRTY)
2384	continue;
2385
2386	lsn = be64_to_cpu(iclog->ic_header->h_lsn);
2387	if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2388	lowest_lsn = lsn;
2389	} while ((iclog = iclog->ic_next) != log->l_iclog);
2390
2391	return lowest_lsn;
2392	}
2393
2394	/*
2395	* Return true if we need to stop processing, false to continue to the next
2396	* iclog. The caller will need to run callbacks if the iclog is returned in the
2397	* XLOG_STATE_CALLBACK state.
2398	*/
2399	static bool
2400	xlog_state_iodone_process_iclog(
2401	struct xlog *log,
2402	struct xlog_in_core *iclog)
2403	{
2404	xfs_lsn_t lowest_lsn;
2405	xfs_lsn_t header_lsn;
2406
2407	switch (iclog->ic_state) {
2408	case XLOG_STATE_ACTIVE:
2409	case XLOG_STATE_DIRTY:
2410	/*
2411	* Skip all iclogs in the ACTIVE & DIRTY states:
2412	*/
2413	return false;
2414	case XLOG_STATE_DONE_SYNC:
2415	/*
2416	* Now that we have an iclog that is in the DONE_SYNC state, do
2417	* one more check here to see if we have chased our tail around.
2418	* If this is not the lowest lsn iclog, then we will leave it
2419	* for another completion to process.
2420	*/
2421	header_lsn = be64_to_cpu(iclog->ic_header->h_lsn);
2422	lowest_lsn = xlog_get_lowest_lsn(log);
2423	if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2424	return false;
2425	/*
2426	* If there are no callbacks on this iclog, we can mark it clean
2427	* immediately and return. Otherwise we need to run the
2428	* callbacks.
2429	*/
2430	if (list_empty(head: &iclog->ic_callbacks)) {
2431	xlog_state_clean_iclog(log, dirty_iclog: iclog);
2432	return false;
2433	}
2434	trace_xlog_iclog_callback(iclog, _RET_IP_);
2435	iclog->ic_state = XLOG_STATE_CALLBACK;
2436	return false;
2437	default:
2438	/*
2439	* Can only perform callbacks in order. Since this iclog is not
2440	* in the DONE_SYNC state, we skip the rest and just try to
2441	* clean up.
2442	*/
2443	return true;
2444	}
2445	}
2446
2447	/*
2448	* Loop over all the iclogs, running attached callbacks on them. Return true if
2449	* we ran any callbacks, indicating that we dropped the icloglock. We don't need
2450	* to handle transient shutdown state here at all because
2451	* xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2452	* cleanup of the callbacks.
2453	*/
2454	static bool
2455	xlog_state_do_iclog_callbacks(
2456	struct xlog *log)
2457	__releases(&log->l_icloglock)
2458	__acquires(&log->l_icloglock)
2459	{
2460	struct xlog_in_core *first_iclog = log->l_iclog;
2461	struct xlog_in_core *iclog = first_iclog;
2462	bool ran_callback = false;
2463
2464	do {
2465	LIST_HEAD(cb_list);
2466
2467	if (xlog_state_iodone_process_iclog(log, iclog))
2468	break;
2469	if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2470	iclog = iclog->ic_next;
2471	continue;
2472	}
2473	list_splice_init(list: &iclog->ic_callbacks, head: &cb_list);
2474	spin_unlock(lock: &log->l_icloglock);
2475
2476	trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2477	xlog_cil_process_committed(list: &cb_list);
2478	trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2479	ran_callback = true;
2480
2481	spin_lock(lock: &log->l_icloglock);
2482	xlog_state_clean_iclog(log, dirty_iclog: iclog);
2483	iclog = iclog->ic_next;
2484	} while (iclog != first_iclog);
2485
2486	return ran_callback;
2487	}
2488
2489
2490	/*
2491	* Loop running iclog completion callbacks until there are no more iclogs in a
2492	* state that can run callbacks.
2493	*/
2494	STATIC void
2495	xlog_state_do_callback(
2496	struct xlog *log)
2497	{
2498	int flushcnt = 0;
2499	int repeats = 0;
2500
2501	spin_lock(lock: &log->l_icloglock);
2502	while (xlog_state_do_iclog_callbacks(log)) {
2503	if (xlog_is_shutdown(log))
2504	break;
2505
2506	if (++repeats > 5000) {
2507	flushcnt += repeats;
2508	repeats = 0;
2509	xfs_warn(log->l_mp,
2510	"%s: possible infinite loop (%d iterations)",
2511	__func__, flushcnt);
2512	}
2513	}
2514
2515	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2516	wake_up_all(&log->l_flush_wait);
2517
2518	spin_unlock(lock: &log->l_icloglock);
2519	}
2520
2521
2522	/*
2523	* Finish transitioning this iclog to the dirty state.
2524	*
2525	* Callbacks could take time, so they are done outside the scope of the
2526	* global state machine log lock.
2527	*/
2528	STATIC void
2529	xlog_state_done_syncing(
2530	struct xlog_in_core *iclog)
2531	{
2532	struct xlog *log = iclog->ic_log;
2533
2534	spin_lock(lock: &log->l_icloglock);
2535	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2536	trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2537
2538	/*
2539	* If we got an error, either on the first buffer, or in the case of
2540	* split log writes, on the second, we shut down the file system and
2541	* no iclogs should ever be attempted to be written to disk again.
2542	*/
2543	if (!xlog_is_shutdown(log)) {
2544	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2545	iclog->ic_state = XLOG_STATE_DONE_SYNC;
2546	}
2547
2548	/*
2549	* Someone could be sleeping prior to writing out the next
2550	* iclog buffer, we wake them all, one will get to do the
2551	* I/O, the others get to wait for the result.
2552	*/
2553	wake_up_all(&iclog->ic_write_wait);
2554	spin_unlock(lock: &log->l_icloglock);
2555	xlog_state_do_callback(log);
2556	}
2557
2558	/*
2559	* If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2560	* sleep. We wait on the flush queue on the head iclog as that should be
2561	* the first iclog to complete flushing. Hence if all iclogs are syncing,
2562	* we will wait here and all new writes will sleep until a sync completes.
2563	*
2564	* The in-core logs are used in a circular fashion. They are not used
2565	* out-of-order even when an iclog past the head is free.
2566	*
2567	* return:
2568	* * log_offset where xlog_write() can start writing into the in-core
2569	* log's data space.
2570	* * in-core log pointer to which xlog_write() should write.
2571	* * boolean indicating this is a continued write to an in-core log.
2572	* If this is the last write, then the in-core log's offset field
2573	* needs to be incremented, depending on the amount of data which
2574	* is copied.
2575	*/
2576	STATIC int
2577	xlog_state_get_iclog_space(
2578	struct xlog *log,
2579	int len,
2580	struct xlog_in_core **iclogp,
2581	struct xlog_ticket *ticket,
2582	int *logoffsetp)
2583	{
2584	int log_offset;
2585	struct xlog_rec_header *head;
2586	struct xlog_in_core *iclog;
2587
2588	restart:
2589	spin_lock(lock: &log->l_icloglock);
2590	if (xlog_is_shutdown(log)) {
2591	spin_unlock(lock: &log->l_icloglock);
2592	return -EIO;
2593	}
2594
2595	iclog = log->l_iclog;
2596	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2597	XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2598
2599	/* Wait for log writes to have flushed */
2600	xlog_wait(wq: &log->l_flush_wait, lock: &log->l_icloglock);
2601	goto restart;
2602	}
2603
2604	head = iclog->ic_header;
2605
2606	atomic_inc(v: &iclog->ic_refcnt); /* prevents sync */
2607	log_offset = iclog->ic_offset;
2608
2609	trace_xlog_iclog_get_space(iclog, _RET_IP_);
2610
2611	/* On the 1st write to an iclog, figure out lsn. This works
2612	* if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2613	* committing to. If the offset is set, that's how many blocks
2614	* must be written.
2615	*/
2616	if (log_offset == 0) {
2617	ticket->t_curr_res -= log->l_iclog_hsize;
2618	head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2619	head->h_lsn = cpu_to_be64(
2620	xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2621	ASSERT(log->l_curr_block >= 0);
2622	}
2623
2624	/* If there is enough room to write everything, then do it. Otherwise,
2625	* claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2626	* bit is on, so this will get flushed out. Don't update ic_offset
2627	* until you know exactly how many bytes get copied. Therefore, wait
2628	* until later to update ic_offset.
2629	*
2630	* xlog_write() algorithm assumes that at least 2 xlog_op_header's
2631	* can fit into remaining data section.
2632	*/
2633	if (iclog->ic_size - iclog->ic_offset <
2634	2 * sizeof(struct xlog_op_header)) {
2635	int error = 0;
2636
2637	xlog_state_switch_iclogs(log, iclog, eventual_size: iclog->ic_size);
2638
2639	/*
2640	* If we are the only one writing to this iclog, sync it to
2641	* disk. We need to do an atomic compare and decrement here to
2642	* avoid racing with concurrent atomic_dec_and_lock() calls in
2643	* xlog_state_release_iclog() when there is more than one
2644	* reference to the iclog.
2645	*/
2646	if (!atomic_add_unless(v: &iclog->ic_refcnt, a: -1, u: 1))
2647	error = xlog_state_release_iclog(log, iclog, ticket);
2648	spin_unlock(lock: &log->l_icloglock);
2649	if (error)
2650	return error;
2651	goto restart;
2652	}
2653
2654	/* Do we have enough room to write the full amount in the remainder
2655	* of this iclog? Or must we continue a write on the next iclog and
2656	* mark this iclog as completely taken? In the case where we switch
2657	* iclogs (to mark it taken), this particular iclog will release/sync
2658	* to disk in xlog_write().
2659	*/
2660	if (len <= iclog->ic_size - iclog->ic_offset)
2661	iclog->ic_offset += len;
2662	else
2663	xlog_state_switch_iclogs(log, iclog, eventual_size: iclog->ic_size);
2664	*iclogp = iclog;
2665
2666	ASSERT(iclog->ic_offset <= iclog->ic_size);
2667	spin_unlock(lock: &log->l_icloglock);
2668
2669	*logoffsetp = log_offset;
2670	return 0;
2671	}
2672
2673	/*
2674	* The first cnt-1 times a ticket goes through here we don't need to move the
2675	* grant write head because the permanent reservation has reserved cnt times the
2676	* unit amount. Release part of current permanent unit reservation and reset
2677	* current reservation to be one units worth. Also move grant reservation head
2678	* forward.
2679	*/
2680	void
2681	xfs_log_ticket_regrant(
2682	struct xlog *log,
2683	struct xlog_ticket *ticket)
2684	{
2685	trace_xfs_log_ticket_regrant(log, tic: ticket);
2686
2687	if (ticket->t_cnt > 0)
2688	ticket->t_cnt--;
2689
2690	xlog_grant_sub_space(head: &log->l_reserve_head, bytes: ticket->t_curr_res);
2691	xlog_grant_sub_space(head: &log->l_write_head, bytes: ticket->t_curr_res);
2692	ticket->t_curr_res = ticket->t_unit_res;
2693
2694	trace_xfs_log_ticket_regrant_sub(log, tic: ticket);
2695
2696	/* just return if we still have some of the pre-reserved space */
2697	if (!ticket->t_cnt) {
2698	xlog_grant_add_space(head: &log->l_reserve_head, bytes: ticket->t_unit_res);
2699	trace_xfs_log_ticket_regrant_exit(log, tic: ticket);
2700	}
2701
2702	xfs_log_ticket_put(ticket);
2703	}
2704
2705	/*
2706	* Give back the space left from a reservation.
2707	*
2708	* All the information we need to make a correct determination of space left
2709	* is present. For non-permanent reservations, things are quite easy. The
2710	* count should have been decremented to zero. We only need to deal with the
2711	* space remaining in the current reservation part of the ticket. If the
2712	* ticket contains a permanent reservation, there may be left over space which
2713	* needs to be released. A count of N means that N-1 refills of the current
2714	* reservation can be done before we need to ask for more space. The first
2715	* one goes to fill up the first current reservation. Once we run out of
2716	* space, the count will stay at zero and the only space remaining will be
2717	* in the current reservation field.
2718	*/
2719	void
2720	xfs_log_ticket_ungrant(
2721	struct xlog *log,
2722	struct xlog_ticket *ticket)
2723	{
2724	int bytes;
2725
2726	trace_xfs_log_ticket_ungrant(log, tic: ticket);
2727
2728	if (ticket->t_cnt > 0)
2729	ticket->t_cnt--;
2730
2731	trace_xfs_log_ticket_ungrant_sub(log, tic: ticket);
2732
2733	/*
2734	* If this is a permanent reservation ticket, we may be able to free
2735	* up more space based on the remaining count.
2736	*/
2737	bytes = ticket->t_curr_res;
2738	if (ticket->t_cnt > 0) {
2739	ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2740	bytes += ticket->t_unit_res*ticket->t_cnt;
2741	}
2742
2743	xlog_grant_sub_space(head: &log->l_reserve_head, bytes);
2744	xlog_grant_sub_space(head: &log->l_write_head, bytes);
2745
2746	trace_xfs_log_ticket_ungrant_exit(log, tic: ticket);
2747
2748	xfs_log_space_wake(mp: log->l_mp);
2749	xfs_log_ticket_put(ticket);
2750	}
2751
2752	/*
2753	* This routine will mark the current iclog in the ring as WANT_SYNC and move
2754	* the current iclog pointer to the next iclog in the ring.
2755	*/
2756	void
2757	xlog_state_switch_iclogs(
2758	struct xlog *log,
2759	struct xlog_in_core *iclog,
2760	int eventual_size)
2761	{
2762	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2763	assert_spin_locked(&log->l_icloglock);
2764	trace_xlog_iclog_switch(iclog, _RET_IP_);
2765
2766	if (!eventual_size)
2767	eventual_size = iclog->ic_offset;
2768	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2769	iclog->ic_header->h_prev_block = cpu_to_be32(log->l_prev_block);
2770	log->l_prev_block = log->l_curr_block;
2771	log->l_prev_cycle = log->l_curr_cycle;
2772
2773	/* roll log?: ic_offset changed later */
2774	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2775
2776	/* Round up to next log-sunit */
2777	if (log->l_iclog_roundoff > BBSIZE) {
2778	uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
2779	log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2780	}
2781
2782	if (log->l_curr_block >= log->l_logBBsize) {
2783	/*
2784	* Rewind the current block before the cycle is bumped to make
2785	* sure that the combined LSN never transiently moves forward
2786	* when the log wraps to the next cycle. This is to support the
2787	* unlocked sample of these fields from xlog_valid_lsn(). Most
2788	* other cases should acquire l_icloglock.
2789	*/
2790	log->l_curr_block -= log->l_logBBsize;
2791	ASSERT(log->l_curr_block >= 0);
2792	smp_wmb();
2793	log->l_curr_cycle++;
2794	if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2795	log->l_curr_cycle++;
2796	}
2797	ASSERT(iclog == log->l_iclog);
2798	log->l_iclog = iclog->ic_next;
2799	}
2800
2801	/*
2802	* Force the iclog to disk and check if the iclog has been completed before
2803	* xlog_force_iclog() returns. This can happen on synchronous (e.g.
2804	* pmem) or fast async storage because we drop the icloglock to issue the IO.
2805	* If completion has already occurred, tell the caller so that it can avoid an
2806	* unnecessary wait on the iclog.
2807	*/
2808	static int
2809	xlog_force_and_check_iclog(
2810	struct xlog_in_core *iclog,
2811	bool *completed)
2812	{
2813	xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header->h_lsn);
2814	int error;
2815
2816	*completed = false;
2817	error = xlog_force_iclog(iclog);
2818	if (error)
2819	return error;
2820
2821	/*
2822	* If the iclog has already been completed and reused the header LSN
2823	* will have been rewritten by completion
2824	*/
2825	if (be64_to_cpu(iclog->ic_header->h_lsn) != lsn)
2826	*completed = true;
2827	return 0;
2828	}
2829
2830	/*
2831	* Write out all data in the in-core log as of this exact moment in time.
2832	*
2833	* Data may be written to the in-core log during this call. However,
2834	* we don't guarantee this data will be written out. A change from past
2835	* implementation means this routine will *not* write out zero length LRs.
2836	*
2837	* Basically, we try and perform an intelligent scan of the in-core logs.
2838	* If we determine there is no flushable data, we just return. There is no
2839	* flushable data if:
2840	*
2841	* 1. the current iclog is active and has no data; the previous iclog
2842	* is in the active or dirty state.
2843	* 2. the current iclog is dirty, and the previous iclog is in the
2844	* active or dirty state.
2845	*
2846	* We may sleep if:
2847	*
2848	* 1. the current iclog is not in the active nor dirty state.
2849	* 2. the current iclog dirty, and the previous iclog is not in the
2850	* active nor dirty state.
2851	* 3. the current iclog is active, and there is another thread writing
2852	* to this particular iclog.
2853	* 4. a) the current iclog is active and has no other writers
2854	* b) when we return from flushing out this iclog, it is still
2855	* not in the active nor dirty state.
2856	*/
2857	int
2858	xfs_log_force(
2859	struct xfs_mount *mp,
2860	uint flags)
2861	{
2862	struct xlog *log = mp->m_log;
2863	struct xlog_in_core *iclog;
2864
2865	XFS_STATS_INC(mp, xs_log_force);
2866	trace_xfs_log_force(mp, 0, _RET_IP_);
2867
2868	xlog_cil_force(log);
2869
2870	spin_lock(lock: &log->l_icloglock);
2871	if (xlog_is_shutdown(log))
2872	goto out_error;
2873
2874	iclog = log->l_iclog;
2875	trace_xlog_iclog_force(iclog, _RET_IP_);
2876
2877	if (iclog->ic_state == XLOG_STATE_DIRTY ||
2878	(iclog->ic_state == XLOG_STATE_ACTIVE &&
2879	atomic_read(v: &iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
2880	/*
2881	* If the head is dirty or (active and empty), then we need to
2882	* look at the previous iclog.
2883	*
2884	* If the previous iclog is active or dirty we are done. There
2885	* is nothing to sync out. Otherwise, we attach ourselves to the
2886	* previous iclog and go to sleep.
2887	*/
2888	iclog = iclog->ic_prev;
2889	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2890	if (atomic_read(v: &iclog->ic_refcnt) == 0) {
2891	/* We have exclusive access to this iclog. */
2892	bool completed;
2893
2894	if (xlog_force_and_check_iclog(iclog, completed: &completed))
2895	goto out_error;
2896
2897	if (completed)
2898	goto out_unlock;
2899	} else {
2900	/*
2901	* Someone else is still writing to this iclog, so we
2902	* need to ensure that when they release the iclog it
2903	* gets synced immediately as we may be waiting on it.
2904	*/
2905	xlog_state_switch_iclogs(log, iclog, eventual_size: 0);
2906	}
2907	}
2908
2909	/*
2910	* The iclog we are about to wait on may contain the checkpoint pushed
2911	* by the above xlog_cil_force() call, but it may not have been pushed
2912	* to disk yet. Like the ACTIVE case above, we need to make sure caches
2913	* are flushed when this iclog is written.
2914	*/
2915	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
2916	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
2917
2918	if (flags & XFS_LOG_SYNC)
2919	return xlog_wait_on_iclog(iclog);
2920	out_unlock:
2921	spin_unlock(lock: &log->l_icloglock);
2922	return 0;
2923	out_error:
2924	spin_unlock(lock: &log->l_icloglock);
2925	return -EIO;
2926	}
2927
2928	/*
2929	* Force the log to a specific LSN.
2930	*
2931	* If an iclog with that lsn can be found:
2932	* If it is in the DIRTY state, just return.
2933	* If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2934	* state and go to sleep or return.
2935	* If it is in any other state, go to sleep or return.
2936	*
2937	* Synchronous forces are implemented with a wait queue. All callers trying
2938	* to force a given lsn to disk must wait on the queue attached to the
2939	* specific in-core log. When given in-core log finally completes its write
2940	* to disk, that thread will wake up all threads waiting on the queue.
2941	*/
2942	static int
2943	xlog_force_lsn(
2944	struct xlog *log,
2945	xfs_lsn_t lsn,
2946	uint flags,
2947	int *log_flushed,
2948	bool already_slept)
2949	{
2950	struct xlog_in_core *iclog;
2951	bool completed;
2952
2953	spin_lock(lock: &log->l_icloglock);
2954	if (xlog_is_shutdown(log))
2955	goto out_error;
2956
2957	iclog = log->l_iclog;
2958	while (be64_to_cpu(iclog->ic_header->h_lsn) != lsn) {
2959	trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
2960	iclog = iclog->ic_next;
2961	if (iclog == log->l_iclog)
2962	goto out_unlock;
2963	}
2964
2965	switch (iclog->ic_state) {
2966	case XLOG_STATE_ACTIVE:
2967	/*
2968	* We sleep here if we haven't already slept (e.g. this is the
2969	* first time we've looked at the correct iclog buf) and the
2970	* buffer before us is going to be sync'ed. The reason for this
2971	* is that if we are doing sync transactions here, by waiting
2972	* for the previous I/O to complete, we can allow a few more
2973	* transactions into this iclog before we close it down.
2974	*
2975	* Otherwise, we mark the buffer WANT_SYNC, and bump up the
2976	* refcnt so we can release the log (which drops the ref count).
2977	* The state switch keeps new transaction commits from using
2978	* this buffer. When the current commits finish writing into
2979	* the buffer, the refcount will drop to zero and the buffer
2980	* will go out then.
2981	*/
2982	if (!already_slept &&
2983	(iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
2984	iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
2985	xlog_wait(wq: &iclog->ic_prev->ic_write_wait,
2986	lock: &log->l_icloglock);
2987	return -EAGAIN;
2988	}
2989	if (xlog_force_and_check_iclog(iclog, completed: &completed))
2990	goto out_error;
2991	if (log_flushed)
2992	*log_flushed = 1;
2993	if (completed)
2994	goto out_unlock;
2995	break;
2996	case XLOG_STATE_WANT_SYNC:
2997	/*
2998	* This iclog may contain the checkpoint pushed by the
2999	* xlog_cil_force_seq() call, but there are other writers still
3000	* accessing it so it hasn't been pushed to disk yet. Like the
3001	* ACTIVE case above, we need to make sure caches are flushed
3002	* when this iclog is written.
3003	*/
3004	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3005	break;
3006	default:
3007	/*
3008	* The entire checkpoint was written by the CIL force and is on
3009	* its way to disk already. It will be stable when it
3010	* completes, so we don't need to manipulate caches here at all.
3011	* We just need to wait for completion if necessary.
3012	*/
3013	break;
3014	}
3015
3016	if (flags & XFS_LOG_SYNC)
3017	return xlog_wait_on_iclog(iclog);
3018	out_unlock:
3019	spin_unlock(lock: &log->l_icloglock);
3020	return 0;
3021	out_error:
3022	spin_unlock(lock: &log->l_icloglock);
3023	return -EIO;
3024	}
3025
3026	/*
3027	* Force the log to a specific checkpoint sequence.
3028	*
3029	* First force the CIL so that all the required changes have been flushed to the
3030	* iclogs. If the CIL force completed it will return a commit LSN that indicates
3031	* the iclog that needs to be flushed to stable storage. If the caller needs
3032	* a synchronous log force, we will wait on the iclog with the LSN returned by
3033	* xlog_cil_force_seq() to be completed.
3034	*/
3035	int
3036	xfs_log_force_seq(
3037	struct xfs_mount *mp,
3038	xfs_csn_t seq,
3039	uint flags,
3040	int *log_flushed)
3041	{
3042	struct xlog *log = mp->m_log;
3043	xfs_lsn_t lsn;
3044	int ret;
3045	ASSERT(seq != 0);
3046
3047	XFS_STATS_INC(mp, xs_log_force);
3048	trace_xfs_log_force(mp, seq, _RET_IP_);
3049
3050	lsn = xlog_cil_force_seq(log, seq);
3051	if (lsn == NULLCOMMITLSN)
3052	return 0;
3053
3054	ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3055	if (ret == -EAGAIN) {
3056	XFS_STATS_INC(mp, xs_log_force_sleep);
3057	ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3058	}
3059	return ret;
3060	}
3061
3062	/*
3063	* Free a used ticket when its refcount falls to zero.
3064	*/
3065	void
3066	xfs_log_ticket_put(
3067	struct xlog_ticket *ticket)
3068	{
3069	ASSERT(atomic_read(&ticket->t_ref) > 0);
3070	if (atomic_dec_and_test(v: &ticket->t_ref))
3071	kmem_cache_free(s: xfs_log_ticket_cache, objp: ticket);
3072	}
3073
3074	struct xlog_ticket *
3075	xfs_log_ticket_get(
3076	struct xlog_ticket *ticket)
3077	{
3078	ASSERT(atomic_read(&ticket->t_ref) > 0);
3079	atomic_inc(v: &ticket->t_ref);
3080	return ticket;
3081	}
3082
3083	/*
3084	* Figure out the total log space unit (in bytes) that would be
3085	* required for a log ticket.
3086	*/
3087	static int
3088	xlog_calc_unit_res(
3089	struct xlog *log,
3090	int unit_bytes,
3091	int *niclogs)
3092	{
3093	int iclog_space;
3094	uint num_headers;
3095
3096	/*
3097	* Permanent reservations have up to 'cnt'-1 active log operations
3098	* in the log. A unit in this case is the amount of space for one
3099	* of these log operations. Normal reservations have a cnt of 1
3100	* and their unit amount is the total amount of space required.
3101	*
3102	* The following lines of code account for non-transaction data
3103	* which occupy space in the on-disk log.
3104	*
3105	* Normal form of a transaction is:
3106	* <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3107	* and then there are LR hdrs, split-recs and roundoff at end of syncs.
3108	*
3109	* We need to account for all the leadup data and trailer data
3110	* around the transaction data.
3111	* And then we need to account for the worst case in terms of using
3112	* more space.
3113	* The worst case will happen if:
3114	* - the placement of the transaction happens to be such that the
3115	* roundoff is at its maximum
3116	* - the transaction data is synced before the commit record is synced
3117	* i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3118	* Therefore the commit record is in its own Log Record.
3119	* This can happen as the commit record is called with its
3120	* own region to xlog_write().
3121	* This then means that in the worst case, roundoff can happen for
3122	* the commit-rec as well.
3123	* The commit-rec is smaller than padding in this scenario and so it is
3124	* not added separately.
3125	*/
3126
3127	/* for trans header */
3128	unit_bytes += sizeof(struct xlog_op_header);
3129	unit_bytes += sizeof(struct xfs_trans_header);
3130
3131	/* for start-rec */
3132	unit_bytes += sizeof(struct xlog_op_header);
3133
3134	/*
3135	* for LR headers - the space for data in an iclog is the size minus
3136	* the space used for the headers. If we use the iclog size, then we
3137	* undercalculate the number of headers required.
3138	*
3139	* Furthermore - the addition of op headers for split-recs might
3140	* increase the space required enough to require more log and op
3141	* headers, so take that into account too.
3142	*
3143	* IMPORTANT: This reservation makes the assumption that if this
3144	* transaction is the first in an iclog and hence has the LR headers
3145	* accounted to it, then the remaining space in the iclog is
3146	* exclusively for this transaction. i.e. if the transaction is larger
3147	* than the iclog, it will be the only thing in that iclog.
3148	* Fundamentally, this means we must pass the entire log vector to
3149	* xlog_write to guarantee this.
3150	*/
3151	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3152	num_headers = howmany(unit_bytes, iclog_space);
3153
3154	/* for split-recs - ophdrs added when data split over LRs */
3155	unit_bytes += sizeof(struct xlog_op_header) * num_headers;
3156
3157	/* add extra header reservations if we overrun */
3158	while (!num_headers ||
3159	howmany(unit_bytes, iclog_space) > num_headers) {
3160	unit_bytes += sizeof(struct xlog_op_header);
3161	num_headers++;
3162	}
3163	unit_bytes += log->l_iclog_hsize * num_headers;
3164
3165	/* for commit-rec LR header - note: padding will subsume the ophdr */
3166	unit_bytes += log->l_iclog_hsize;
3167
3168	/* roundoff padding for transaction data and one for commit record */
3169	unit_bytes += 2 * log->l_iclog_roundoff;
3170
3171	if (niclogs)
3172	*niclogs = num_headers;
3173	return unit_bytes;
3174	}
3175
3176	int
3177	xfs_log_calc_unit_res(
3178	struct xfs_mount *mp,
3179	int unit_bytes)
3180	{
3181	return xlog_calc_unit_res(log: mp->m_log, unit_bytes, NULL);
3182	}
3183
3184	/*
3185	* Allocate and initialise a new log ticket.
3186	*/
3187	struct xlog_ticket *
3188	xlog_ticket_alloc(
3189	struct xlog *log,
3190	int unit_bytes,
3191	int cnt,
3192	bool permanent)
3193	{
3194	struct xlog_ticket *tic;
3195	int unit_res;
3196
3197	tic = kmem_cache_zalloc(xfs_log_ticket_cache,
3198	GFP_KERNEL | __GFP_NOFAIL);
3199
3200	unit_res = xlog_calc_unit_res(log, unit_bytes, niclogs: &tic->t_iclog_hdrs);
3201
3202	atomic_set(v: &tic->t_ref, i: 1);
3203	tic->t_task = current;
3204	INIT_LIST_HEAD(list: &tic->t_queue);
3205	tic->t_unit_res = unit_res;
3206	tic->t_curr_res = unit_res;
3207	tic->t_cnt = cnt;
3208	tic->t_ocnt = cnt;
3209	tic->t_tid = get_random_u32();
3210	if (permanent)
3211	tic->t_flags |= XLOG_TIC_PERM_RESERV;
3212
3213	return tic;
3214	}
3215
3216	#if defined(DEBUG)
3217	static void
3218	xlog_verify_dump_tail(
3219	struct xlog *log,
3220	struct xlog_in_core *iclog)
3221	{
3222	xfs_alert(log->l_mp,
3223	"ran out of log space tail 0x%llx/0x%llx, head lsn 0x%llx, head 0x%x/0x%x, prev head 0x%x/0x%x",
3224	iclog ? be64_to_cpu(iclog->ic_header->h_tail_lsn) : -1,
3225	atomic64_read(&log->l_tail_lsn),
3226	log->l_ailp->ail_head_lsn,
3227	log->l_curr_cycle, log->l_curr_block,
3228	log->l_prev_cycle, log->l_prev_block);
3229	xfs_alert(log->l_mp,
3230	"write grant 0x%llx, reserve grant 0x%llx, tail_space 0x%llx, size 0x%x, iclog flags 0x%x",
3231	atomic64_read(&log->l_write_head.grant),
3232	atomic64_read(&log->l_reserve_head.grant),
3233	log->l_tail_space, log->l_logsize,
3234	iclog ? iclog->ic_flags : -1);
3235	}
3236
3237	/* Check if the new iclog will fit in the log. */
3238	STATIC void
3239	xlog_verify_tail_lsn(
3240	struct xlog *log,
3241	struct xlog_in_core *iclog)
3242	{
3243	xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header->h_tail_lsn);
3244	int blocks;
3245
3246	if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3247	blocks = log->l_logBBsize -
3248	(log->l_prev_block - BLOCK_LSN(tail_lsn));
3249	if (blocks < BTOBB(iclog->ic_offset) +
3250	BTOBB(log->l_iclog_hsize)) {
3251	xfs_emerg(log->l_mp,
3252	"%s: ran out of log space", __func__);
3253	xlog_verify_dump_tail(log, iclog);
3254	}
3255	return;
3256	}
3257
3258	if (CYCLE_LSN(tail_lsn) + 1 != log->l_prev_cycle) {
3259	xfs_emerg(log->l_mp, "%s: head has wrapped tail.", __func__);
3260	xlog_verify_dump_tail(log, iclog);
3261	return;
3262	}
3263	if (BLOCK_LSN(tail_lsn) == log->l_prev_block) {
3264	xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3265	xlog_verify_dump_tail(log, iclog);
3266	return;
3267	}
3268
3269	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3270	if (blocks < BTOBB(iclog->ic_offset) + 1) {
3271	xfs_emerg(log->l_mp, "%s: ran out of iclog space", __func__);
3272	xlog_verify_dump_tail(log, iclog);
3273	}
3274	}
3275
3276	/*
3277	* Perform a number of checks on the iclog before writing to disk.
3278	*
3279	* 1. Make sure the iclogs are still circular
3280	* 2. Make sure we have a good magic number
3281	* 3. Make sure we don't have magic numbers in the data
3282	* 4. Check fields of each log operation header for:
3283	* A. Valid client identifier
3284	* B. tid ptr value falls in valid ptr space (user space code)
3285	* C. Length in log record header is correct according to the
3286	* individual operation headers within record.
3287	* 5. When a bwrite will occur within 5 blocks of the front of the physical
3288	* log, check the preceding blocks of the physical log to make sure all
3289	* the cycle numbers agree with the current cycle number.
3290	*/
3291	STATIC void
3292	xlog_verify_iclog(
3293	struct xlog *log,
3294	struct xlog_in_core *iclog,
3295	int count)
3296	{
3297	struct xlog_rec_header *rhead = iclog->ic_header;
3298	struct xlog_in_core *icptr;
3299	void *base_ptr, *ptr;
3300	ptrdiff_t field_offset;
3301	uint8_t clientid;
3302	int len, i, op_len;
3303	int idx;
3304
3305	/* check validity of iclog pointers */
3306	spin_lock(lock: &log->l_icloglock);
3307	icptr = log->l_iclog;
3308	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3309	ASSERT(icptr);
3310
3311	if (icptr != log->l_iclog)
3312	xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3313	spin_unlock(lock: &log->l_icloglock);
3314
3315	/* check log magic numbers */
3316	if (rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3317	xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3318
3319	base_ptr = ptr = rhead;
3320	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3321	if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3322	xfs_emerg(log->l_mp, "%s: unexpected magic num",
3323	__func__);
3324	}
3325
3326	/* check fields */
3327	len = be32_to_cpu(rhead->h_num_logops);
3328	base_ptr = ptr = iclog->ic_datap;
3329	for (i = 0; i < len; i++) {
3330	struct xlog_op_header *ophead = ptr;
3331	void *p = &ophead->oh_clientid;
3332
3333	/* clientid is only 1 byte */
3334	field_offset = p - base_ptr;
3335	if (field_offset & 0x1ff) {
3336	clientid = ophead->oh_clientid;
3337	} else {
3338	idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3339	clientid = xlog_get_client_id(i: *xlog_cycle_data(rhead, i: idx));
3340	}
3341	if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3342	xfs_warn(log->l_mp,
3343	"%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3344	__func__, i, clientid, ophead,
3345	(unsigned long)field_offset);
3346	}
3347
3348	/* check length */
3349	p = &ophead->oh_len;
3350	field_offset = p - base_ptr;
3351	if (field_offset & 0x1ff) {
3352	op_len = be32_to_cpu(ophead->oh_len);
3353	} else {
3354	idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3355	op_len = be32_to_cpu(*xlog_cycle_data(rhead, idx));
3356	}
3357	ptr += sizeof(struct xlog_op_header) + op_len;
3358	}
3359	}
3360	#endif
3361
3362	/*
3363	* Perform a forced shutdown on the log.
3364	*
3365	* This can be called from low level log code to trigger a shutdown, or from the
3366	* high level mount shutdown code when the mount shuts down.
3367	*
3368	* Our main objectives here are to make sure that:
3369	* a. if the shutdown was not due to a log IO error, flush the logs to
3370	* disk. Anything modified after this is ignored.
3371	* b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3372	* parties to find out. Nothing new gets queued after this is done.
3373	* c. Tasks sleeping on log reservations, pinned objects and
3374	* other resources get woken up.
3375	* d. The mount is also marked as shut down so that log triggered shutdowns
3376	* still behave the same as if they called xfs_forced_shutdown().
3377	*
3378	* Return true if the shutdown cause was a log IO error and we actually shut the
3379	* log down.
3380	*/
3381	bool
3382	xlog_force_shutdown(
3383	struct xlog *log,
3384	uint32_t shutdown_flags)
3385	{
3386	bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3387
3388	if (!log)
3389	return false;
3390
3391	/*
3392	* Ensure that there is only ever one log shutdown being processed.
3393	* If we allow the log force below on a second pass after shutting
3394	* down the log, we risk deadlocking the CIL push as it may require
3395	* locks on objects the current shutdown context holds (e.g. taking
3396	* buffer locks to abort buffers on last unpin of buf log items).
3397	*/
3398	if (test_and_set_bit(XLOG_SHUTDOWN_STARTED, addr: &log->l_opstate))
3399	return false;
3400
3401	/*
3402	* Flush all the completed transactions to disk before marking the log
3403	* being shut down. We need to do this first as shutting down the log
3404	* before the force will prevent the log force from flushing the iclogs
3405	* to disk.
3406	*
3407	* When we are in recovery, there are no transactions to flush, and
3408	* we don't want to touch the log because we don't want to perturb the
3409	* current head/tail for future recovery attempts. Hence we need to
3410	* avoid a log force in this case.
3411	*
3412	* If we are shutting down due to a log IO error, then we must avoid
3413	* trying to write the log as that may just result in more IO errors and
3414	* an endless shutdown/force loop.
3415	*/
3416	if (!log_error && !xlog_in_recovery(log))
3417	xfs_log_force(mp: log->l_mp, XFS_LOG_SYNC);
3418
3419	/*
3420	* Atomically set the shutdown state. If the shutdown state is already
3421	* set, there someone else is performing the shutdown and so we are done
3422	* here. This should never happen because we should only ever get called
3423	* once by the first shutdown caller.
3424	*
3425	* Much of the log state machine transitions assume that shutdown state
3426	* cannot change once they hold the log->l_icloglock. Hence we need to
3427	* hold that lock here, even though we use the atomic test_and_set_bit()
3428	* operation to set the shutdown state.
3429	*/
3430	spin_lock(lock: &log->l_icloglock);
3431	if (test_and_set_bit(XLOG_IO_ERROR, addr: &log->l_opstate)) {
3432	spin_unlock(lock: &log->l_icloglock);
3433	ASSERT(0);
3434	return false;
3435	}
3436	spin_unlock(lock: &log->l_icloglock);
3437
3438	/*
3439	* If this log shutdown also sets the mount shutdown state, issue a
3440	* shutdown warning message.
3441	*/
3442	if (!xfs_set_shutdown(mp: log->l_mp)) {
3443	xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3444	"Filesystem has been shut down due to log error (0x%x).",
3445	shutdown_flags);
3446	xfs_alert(log->l_mp,
3447	"Please unmount the filesystem and rectify the problem(s).");
3448	if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3449	xfs_stack_trace();
3450	}
3451
3452	/*
3453	* We don't want anybody waiting for log reservations after this. That
3454	* means we have to wake up everybody queued up on reserveq as well as
3455	* writeq. In addition, we make sure in xlog_{re}grant_log_space that
3456	* we don't enqueue anything once the SHUTDOWN flag is set, and this
3457	* action is protected by the grant locks.
3458	*/
3459	xlog_grant_head_wake_all(head: &log->l_reserve_head);
3460	xlog_grant_head_wake_all(head: &log->l_write_head);
3461
3462	/*
3463	* Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3464	* as if the log writes were completed. The abort handling in the log
3465	* item committed callback functions will do this again under lock to
3466	* avoid races.
3467	*/
3468	spin_lock(lock: &log->l_cilp->xc_push_lock);
3469	wake_up_all(&log->l_cilp->xc_start_wait);
3470	wake_up_all(&log->l_cilp->xc_commit_wait);
3471	spin_unlock(lock: &log->l_cilp->xc_push_lock);
3472
3473	spin_lock(lock: &log->l_icloglock);
3474	xlog_state_shutdown_callbacks(log);
3475	spin_unlock(lock: &log->l_icloglock);
3476
3477	wake_up_var(var: &log->l_opstate);
3478	if (IS_ENABLED(CONFIG_XFS_RT) && xfs_has_zoned(mp: log->l_mp))
3479	xfs_zoned_wake_all(mp: log->l_mp);
3480
3481	return log_error;
3482	}
3483
3484	STATIC int
3485	xlog_iclogs_empty(
3486	struct xlog *log)
3487	{
3488	struct xlog_in_core *iclog = log->l_iclog;
3489
3490	do {
3491	/* endianness does not matter here, zero is zero in
3492	* any language.
3493	*/
3494	if (iclog->ic_header->h_num_logops)
3495	return 0;
3496	iclog = iclog->ic_next;
3497	} while (iclog != log->l_iclog);
3498
3499	return 1;
3500	}
3501
3502	/*
3503	* Verify that an LSN stamped into a piece of metadata is valid. This is
3504	* intended for use in read verifiers on v5 superblocks.
3505	*/
3506	bool
3507	xfs_log_check_lsn(
3508	struct xfs_mount *mp,
3509	xfs_lsn_t lsn)
3510	{
3511	struct xlog *log = mp->m_log;
3512	bool valid;
3513
3514	/*
3515	* norecovery mode skips mount-time log processing and unconditionally
3516	* resets the in-core LSN. We can't validate in this mode, but
3517	* modifications are not allowed anyways so just return true.
3518	*/
3519	if (xfs_has_norecovery(mp))
3520	return true;
3521
3522	/*
3523	* Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3524	* handled by recovery and thus safe to ignore here.
3525	*/
3526	if (lsn == NULLCOMMITLSN)
3527	return true;
3528
3529	valid = xlog_valid_lsn(mp->m_log, lsn);
3530
3531	/* warn the user about what's gone wrong before verifier failure */
3532	if (!valid) {
3533	spin_lock(lock: &log->l_icloglock);
3534	xfs_warn(mp,
3535	"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3536	"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3537	CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3538	log->l_curr_cycle, log->l_curr_block);
3539	spin_unlock(lock: &log->l_icloglock);
3540	}
3541
3542	return valid;
3543	}
3544