linux - drivers/md/raid5-cache.c

3187 lines · 87.8 KiB

1	// SPDX-License-Identifier: GPL-2.0-only
2	/*
3	* Copyright (C) 2015 Shaohua Li <[email protected]>
4	* Copyright (C) 2016 Song Liu <[email protected]>
5	*/
6	#include <linux/kernel.h>
7	#include <linux/wait.h>
8	#include <linux/blkdev.h>
9	#include <linux/slab.h>
10	#include <linux/raid/md_p.h>
11	#include <linux/crc32c.h>
12	#include <linux/random.h>
13	#include <linux/kthread.h>
14	#include <linux/types.h>
15	#include "md.h"
16	#include "raid5.h"
17	#include "md-bitmap.h"
18	#include "raid5-log.h"
19
20	/*
21	* metadata/data stored in disk with 4k size unit (a block) regardless
22	* underneath hardware sector size. only works with PAGE_SIZE == 4096
23	*/
24	#define BLOCK_SECTORS (8)
25	#define BLOCK_SECTOR_SHIFT (3)
26
27	/*
28	* log->max_free_space is min(1/4 disk size, 10G reclaimable space).
29	*
30	* In write through mode, the reclaim runs every log->max_free_space.
31	* This can prevent the recovery scans for too long
32	*/
33	#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
34	#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
35
36	/* wake up reclaim thread periodically */
37	#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
38	/* start flush with these full stripes */
39	#define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
40	/* reclaim stripes in groups */
41	#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
42
43	/*
44	* We only need 2 bios per I/O unit to make progress, but ensure we
45	* have a few more available to not get too tight.
46	*/
47	#define R5L_POOL_SIZE 4
48
49	static char *r5c_journal_mode_str[] = {"write-through",
50	"write-back"};
51	/*
52	* raid5 cache state machine
53	*
54	* With the RAID cache, each stripe works in two phases:
55	* - caching phase
56	* - writing-out phase
57	*
58	* These two phases are controlled by bit STRIPE_R5C_CACHING:
59	* if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
60	* if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
61	*
62	* When there is no journal, or the journal is in write-through mode,
63	* the stripe is always in writing-out phase.
64	*
65	* For write-back journal, the stripe is sent to caching phase on write
66	* (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
67	* the write-out phase by clearing STRIPE_R5C_CACHING.
68	*
69	* Stripes in caching phase do not write the raid disks. Instead, all
70	* writes are committed from the log device. Therefore, a stripe in
71	* caching phase handles writes as:
72	* - write to log device
73	* - return IO
74	*
75	* Stripes in writing-out phase handle writes as:
76	* - calculate parity
77	* - write pending data and parity to journal
78	* - write data and parity to raid disks
79	* - return IO for pending writes
80	*/
81
82	struct r5l_log {
83	struct md_rdev *rdev;
84
85	u32 uuid_checksum;
86
87	sector_t device_size; /* log device size, round to
88	* BLOCK_SECTORS */
89	sector_t max_free_space; /* reclaim run if free space is at
90	* this size */
91
92	sector_t last_checkpoint; /* log tail. where recovery scan
93	* starts from */
94	u64 last_cp_seq; /* log tail sequence */
95
96	sector_t log_start; /* log head. where new data appends */
97	u64 seq; /* log head sequence */
98
99	sector_t next_checkpoint;
100
101	struct mutex io_mutex;
102	struct r5l_io_unit current_io; / current io_unit accepting new data */
103
104	spinlock_t io_list_lock;
105	struct list_head running_ios; /* io_units which are still running,
106	* and have not yet been completely
107	* written to the log */
108	struct list_head io_end_ios; /* io_units which have been completely
109	* written to the log but not yet written
110	* to the RAID */
111	struct list_head flushing_ios; /* io_units which are waiting for log
112	* cache flush */
113	struct list_head finished_ios; /* io_units which settle down in log disk */
114	struct bio flush_bio;
115
116	struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
117
118	struct kmem_cache *io_kc;
119	mempool_t io_pool;
120	struct bio_set bs;
121	mempool_t meta_pool;
122
123	struct md_thread __rcu *reclaim_thread;
124	unsigned long reclaim_target; /* number of space that need to be
125	* reclaimed. if it's 0, reclaim spaces
126	* used by io_units which are in
127	* IO_UNIT_STRIPE_END state (eg, reclaim
128	* doesn't wait for specific io_unit
129	* switching to IO_UNIT_STRIPE_END
130	* state) */
131	wait_queue_head_t iounit_wait;
132
133	struct list_head no_space_stripes; /* pending stripes, log has no space */
134	spinlock_t no_space_stripes_lock;
135
136	bool need_cache_flush;
137
138	/* for r5c_cache */
139	enum r5c_journal_mode r5c_journal_mode;
140
141	/* all stripes in r5cache, in the order of seq at sh->log_start */
142	struct list_head stripe_in_journal_list;
143
144	spinlock_t stripe_in_journal_lock;
145	atomic_t stripe_in_journal_count;
146
147	/* to submit async io_units, to fulfill ordering of flush */
148	struct work_struct deferred_io_work;
149	/* to disable write back during in degraded mode */
150	struct work_struct disable_writeback_work;
151
152	/* to for chunk_aligned_read in writeback mode, details below */
153	spinlock_t tree_lock;
154	struct radix_tree_root big_stripe_tree;
155	};
156
157	/*
158	* Enable chunk_aligned_read() with write back cache.
159	*
160	* Each chunk may contain more than one stripe (for example, a 256kB
161	* chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
162	* chunk_aligned_read, these stripes are grouped into one "big_stripe".
163	* For each big_stripe, we count how many stripes of this big_stripe
164	* are in the write back cache. These data are tracked in a radix tree
165	* (big_stripe_tree). We use radix_tree item pointer as the counter.
166	* r5c_tree_index() is used to calculate keys for the radix tree.
167	*
168	* chunk_aligned_read() calls r5c_big_stripe_cached() to look up
169	* big_stripe of each chunk in the tree. If this big_stripe is in the
170	* tree, chunk_aligned_read() aborts. This look up is protected by
171	* rcu_read_lock().
172	*
173	* It is necessary to remember whether a stripe is counted in
174	* big_stripe_tree. Instead of adding new flag, we reuses existing flags:
175	* STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
176	* two flags are set, the stripe is counted in big_stripe_tree. This
177	* requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
178	* r5c_try_caching_write(); and moving clear_bit of
179	* STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
180	* r5c_finish_stripe_write_out().
181	*/
182
183	/*
184	* radix tree requests lowest 2 bits of data pointer to be 2b'00.
185	* So it is necessary to left shift the counter by 2 bits before using it
186	* as data pointer of the tree.
187	*/
188	#define R5C_RADIX_COUNT_SHIFT 2
189
190	/*
191	* calculate key for big_stripe_tree
192	*
193	* sect: align_bi->bi_iter.bi_sector or sh->sector
194	*/
195	static inline sector_t r5c_tree_index(struct r5conf *conf,
196	sector_t sect)
197	{
198	sector_div(sect, conf->chunk_sectors);
199	return sect;
200	}
201
202	/*
203	* an IO range starts from a meta data block and end at the next meta data
204	* block. The io unit's the meta data block tracks data/parity followed it. io
205	* unit is written to log disk with normal write, as we always flush log disk
206	* first and then start move data to raid disks, there is no requirement to
207	* write io unit with FLUSH/FUA
208	*/
209	struct r5l_io_unit {
210	struct r5l_log *log;
211
212	struct page meta_page; / store meta block */
213	int meta_offset; /* current offset in meta_page */
214
215	struct bio current_bio;/ current_bio accepting new data */
216
217	atomic_t pending_stripe;/* how many stripes not flushed to raid */
218	u64 seq; /* seq number of the metablock */
219	sector_t log_start; /* where the io_unit starts */
220	sector_t log_end; /* where the io_unit ends */
221	struct list_head log_sibling; /* log->running_ios */
222	struct list_head stripe_list; /* stripes added to the io_unit */
223
224	int state;
225	bool need_split_bio;
226	struct bio *split_bio;
227
228	unsigned int has_flush:1; /* include flush request */
229	unsigned int has_fua:1; /* include fua request */
230	unsigned int has_null_flush:1; /* include null flush request */
231	unsigned int has_flush_payload:1; /* include flush payload */
232	/*
233	* io isn't sent yet, flush/fua request can only be submitted till it's
234	* the first IO in running_ios list
235	*/
236	unsigned int io_deferred:1;
237
238	struct bio_list flush_barriers; /* size == 0 flush bios */
239	};
240
241	/* r5l_io_unit state */
242	enum r5l_io_unit_state {
243	IO_UNIT_RUNNING = 0, /* accepting new IO */
244	IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
245	* don't accepting new bio */
246	IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
247	IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
248	};
249
250	bool r5c_is_writeback(struct r5l_log *log)
251	{
252	return (log != NULL &&
253	log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
254	}
255
256	static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
257	{
258	start += inc;
259	if (start >= log->device_size)
260	start = start - log->device_size;
261	return start;
262	}
263
264	static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
265	sector_t end)
266	{
267	if (end >= start)
268	return end - start;
269	else
270	return end + log->device_size - start;
271	}
272
273	static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
274	{
275	sector_t used_size;
276
277	used_size = r5l_ring_distance(log, log->last_checkpoint,
278	log->log_start);
279
280	return log->device_size > used_size + size;
281	}
282
283	static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
284	enum r5l_io_unit_state state)
285	{
286	if (WARN_ON(io->state >= state))
287	return;
288	io->state = state;
289	}
290
291	static void
292	r5c_return_dev_pending_writes(struct r5conf conf, struct r5dev dev)
293	{
294	struct bio wbi, wbi2;
295
296	wbi = dev->written;
297	dev->written = NULL;
298	while (wbi && wbi->bi_iter.bi_sector <
299	dev->sector + RAID5_STRIPE_SECTORS(conf)) {
300	wbi2 = r5_next_bio(conf, wbi, dev->sector);
301	md_write_end(conf->mddev);
302	bio_endio(wbi);
303	wbi = wbi2;
304	}
305	}
306
307	void r5c_handle_cached_data_endio(struct r5conf *conf,
308	struct stripe_head *sh, int disks)
309	{
310	int i;
311
312	for (i = sh->disks; i--; ) {
313	if (sh->dev[i].written) {
314	set_bit(R5_UPTODATE, &sh->dev[i].flags);
315	r5c_return_dev_pending_writes(conf, &sh->dev[i]);
316	}
317	}
318	}
319
320	void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
321
322	/* Check whether we should flush some stripes to free up stripe cache */
323	void r5c_check_stripe_cache_usage(struct r5conf *conf)
324	{
325	int total_cached;
326	struct r5l_log *log = READ_ONCE(conf->log);
327
328	if (!r5c_is_writeback(log))
329	return;
330
331	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
332	atomic_read(&conf->r5c_cached_full_stripes);
333
334	/*
335	* The following condition is true for either of the following:
336	* - stripe cache pressure high:
337	* total_cached > 3/4 min_nr_stripes \|\|
338	* empty_inactive_list_nr > 0
339	* - stripe cache pressure moderate:
340	* total_cached > 1/2 min_nr_stripes
341	*/
342	if (total_cached > conf->min_nr_stripes * 1 / 2 \|\|
343	atomic_read(&conf->empty_inactive_list_nr) > 0)
344	r5l_wake_reclaim(log, 0);
345	}
346
347	/*
348	* flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
349	* stripes in the cache
350	*/
351	void r5c_check_cached_full_stripe(struct r5conf *conf)
352	{
353	struct r5l_log *log = READ_ONCE(conf->log);
354
355	if (!r5c_is_writeback(log))
356	return;
357
358	/*
359	* wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
360	* or a full stripe (chunk size / 4k stripes).
361	*/
362	if (atomic_read(&conf->r5c_cached_full_stripes) >=
363	min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
364	conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
365	r5l_wake_reclaim(log, 0);
366	}
367
368	/*
369	* Total log space (in sectors) needed to flush all data in cache
370	*
371	* To avoid deadlock due to log space, it is necessary to reserve log
372	* space to flush critical stripes (stripes that occupying log space near
373	* last_checkpoint). This function helps check how much log space is
374	* required to flush all cached stripes.
375	*
376	* To reduce log space requirements, two mechanisms are used to give cache
377	* flush higher priorities:
378	* 1. In handle_stripe_dirtying() and schedule_reconstruction(),
379	* stripes ALREADY in journal can be flushed w/o pending writes;
380	* 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
381	* can be delayed (r5l_add_no_space_stripe).
382	*
383	* In cache flush, the stripe goes through 1 and then 2. For a stripe that
384	* already passed 1, flushing it requires at most (conf->max_degraded + 1)
385	* pages of journal space. For stripes that has not passed 1, flushing it
386	* requires (conf->raid_disks + 1) pages of journal space. There are at
387	* most (conf->group_cnt + 1) stripe that passed 1. So total journal space
388	* required to flush all cached stripes (in pages) is:
389	*
390	* (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
391	* (group_cnt + 1) * (raid_disks + 1)
392	* or
393	* (stripe_in_journal_count) * (max_degraded + 1) +
394	* (group_cnt + 1) * (raid_disks - max_degraded)
395	*/
396	static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
397	{
398	struct r5l_log *log = READ_ONCE(conf->log);
399
400	if (!r5c_is_writeback(log))
401	return 0;
402
403	return BLOCK_SECTORS *
404	((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
405	(conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
406	}
407
408	/*
409	* evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
410	*
411	* R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
412	* reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
413	* device is less than 2x of reclaim_required_space.
414	*/
415	static inline void r5c_update_log_state(struct r5l_log *log)
416	{
417	struct r5conf *conf = log->rdev->mddev->private;
418	sector_t free_space;
419	sector_t reclaim_space;
420	bool wake_reclaim = false;
421
422	if (!r5c_is_writeback(log))
423	return;
424
425	free_space = r5l_ring_distance(log, log->log_start,
426	log->last_checkpoint);
427	reclaim_space = r5c_log_required_to_flush_cache(conf);
428	if (free_space < 2 * reclaim_space)
429	set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
430	else {
431	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
432	wake_reclaim = true;
433	clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
434	}
435	if (free_space < 3 * reclaim_space)
436	set_bit(R5C_LOG_TIGHT, &conf->cache_state);
437	else
438	clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
439
440	if (wake_reclaim)
441	r5l_wake_reclaim(log, 0);
442	}
443
444	/*
445	* Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
446	* This function should only be called in write-back mode.
447	*/
448	void r5c_make_stripe_write_out(struct stripe_head *sh)
449	{
450	struct r5conf *conf = sh->raid_conf;
451	struct r5l_log *log = READ_ONCE(conf->log);
452
453	BUG_ON(!r5c_is_writeback(log));
454
455	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
456	clear_bit(STRIPE_R5C_CACHING, &sh->state);
457
458	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
459	atomic_inc(&conf->preread_active_stripes);
460	}
461
462	static void r5c_handle_data_cached(struct stripe_head *sh)
463	{
464	int i;
465
466	for (i = sh->disks; i--; )
467	if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
468	set_bit(R5_InJournal, &sh->dev[i].flags);
469	clear_bit(R5_LOCKED, &sh->dev[i].flags);
470	}
471	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
472	}
473
474	/*
475	* this journal write must contain full parity,
476	* it may also contain some data pages
477	*/
478	static void r5c_handle_parity_cached(struct stripe_head *sh)
479	{
480	int i;
481
482	for (i = sh->disks; i--; )
483	if (test_bit(R5_InJournal, &sh->dev[i].flags))
484	set_bit(R5_Wantwrite, &sh->dev[i].flags);
485	}
486
487	/*
488	* Setting proper flags after writing (or flushing) data and/or parity to the
489	* log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
490	*/
491	static void r5c_finish_cache_stripe(struct stripe_head *sh)
492	{
493	struct r5l_log *log = READ_ONCE(sh->raid_conf->log);
494
495	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
496	BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
497	/*
498	* Set R5_InJournal for parity dev[pd_idx]. This means
499	* all data AND parity in the journal. For RAID 6, it is
500	* NOT necessary to set the flag for dev[qd_idx], as the
501	* two parities are written out together.
502	*/
503	set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
504	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
505	r5c_handle_data_cached(sh);
506	} else {
507	r5c_handle_parity_cached(sh);
508	set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
509	}
510	}
511
512	static void r5l_io_run_stripes(struct r5l_io_unit *io)
513	{
514	struct stripe_head sh, next;
515
516	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
517	list_del_init(&sh->log_list);
518
519	r5c_finish_cache_stripe(sh);
520
521	set_bit(STRIPE_HANDLE, &sh->state);
522	raid5_release_stripe(sh);
523	}
524	}
525
526	static void r5l_log_run_stripes(struct r5l_log *log)
527	{
528	struct r5l_io_unit io, next;
529
530	lockdep_assert_held(&log->io_list_lock);
531
532	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
533	/* don't change list order */
534	if (io->state < IO_UNIT_IO_END)
535	break;
536
537	list_move_tail(&io->log_sibling, &log->finished_ios);
538	r5l_io_run_stripes(io);
539	}
540	}
541
542	static void r5l_move_to_end_ios(struct r5l_log *log)
543	{
544	struct r5l_io_unit io, next;
545
546	lockdep_assert_held(&log->io_list_lock);
547
548	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
549	/* don't change list order */
550	if (io->state < IO_UNIT_IO_END)
551	break;
552	list_move_tail(&io->log_sibling, &log->io_end_ios);
553	}
554	}
555
556	static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
557	static void r5l_log_endio(struct bio *bio)
558	{
559	struct r5l_io_unit *io = bio->bi_private;
560	struct r5l_io_unit *io_deferred;
561	struct r5l_log *log = io->log;
562	unsigned long flags;
563	bool has_null_flush;
564	bool has_flush_payload;
565
566	if (bio->bi_status)
567	md_error(log->rdev->mddev, log->rdev);
568
569	bio_put(bio);
570	mempool_free(io->meta_page, &log->meta_pool);
571
572	spin_lock_irqsave(&log->io_list_lock, flags);
573	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
574
575	/*
576	* if the io doesn't not have null_flush or flush payload,
577	* it is not safe to access it after releasing io_list_lock.
578	* Therefore, it is necessary to check the condition with
579	* the lock held.
580	*/
581	has_null_flush = io->has_null_flush;
582	has_flush_payload = io->has_flush_payload;
583
584	if (log->need_cache_flush && !list_empty(&io->stripe_list))
585	r5l_move_to_end_ios(log);
586	else
587	r5l_log_run_stripes(log);
588	if (!list_empty(&log->running_ios)) {
589	/*
590	* FLUSH/FUA io_unit is deferred because of ordering, now we
591	* can dispatch it
592	*/
593	io_deferred = list_first_entry(&log->running_ios,
594	struct r5l_io_unit, log_sibling);
595	if (io_deferred->io_deferred)
596	schedule_work(&log->deferred_io_work);
597	}
598
599	spin_unlock_irqrestore(&log->io_list_lock, flags);
600
601	if (log->need_cache_flush)
602	md_wakeup_thread(log->rdev->mddev->thread);
603
604	/* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
605	if (has_null_flush) {
606	struct bio *bi;
607
608	WARN_ON(bio_list_empty(&io->flush_barriers));
609	while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
610	bio_endio(bi);
611	if (atomic_dec_and_test(&io->pending_stripe)) {
612	__r5l_stripe_write_finished(io);
613	return;
614	}
615	}
616	}
617	/* decrease pending_stripe for flush payload */
618	if (has_flush_payload)
619	if (atomic_dec_and_test(&io->pending_stripe))
620	__r5l_stripe_write_finished(io);
621	}
622
623	static void r5l_do_submit_io(struct r5l_log log, struct r5l_io_unit io)
624	{
625	unsigned long flags;
626
627	spin_lock_irqsave(&log->io_list_lock, flags);
628	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
629	spin_unlock_irqrestore(&log->io_list_lock, flags);
630
631	/*
632	* In case of journal device failures, submit_bio will get error
633	* and calls endio, then active stripes will continue write
634	* process. Therefore, it is not necessary to check Faulty bit
635	* of journal device here.
636	*
637	* We can't check split_bio after current_bio is submitted. If
638	* io->split_bio is null, after current_bio is submitted, current_bio
639	* might already be completed and the io_unit is freed. We submit
640	* split_bio first to avoid the issue.
641	*/
642	if (io->split_bio) {
643	if (io->has_flush)
644	io->split_bio->bi_opf \|= REQ_PREFLUSH;
645	if (io->has_fua)
646	io->split_bio->bi_opf \|= REQ_FUA;
647	submit_bio(io->split_bio);
648	}
649
650	if (io->has_flush)
651	io->current_bio->bi_opf \|= REQ_PREFLUSH;
652	if (io->has_fua)
653	io->current_bio->bi_opf \|= REQ_FUA;
654	submit_bio(io->current_bio);
655	}
656
657	/* deferred io_unit will be dispatched here */
658	static void r5l_submit_io_async(struct work_struct *work)
659	{
660	struct r5l_log *log = container_of(work, struct r5l_log,
661	deferred_io_work);
662	struct r5l_io_unit *io = NULL;
663	unsigned long flags;
664
665	spin_lock_irqsave(&log->io_list_lock, flags);
666	if (!list_empty(&log->running_ios)) {
667	io = list_first_entry(&log->running_ios, struct r5l_io_unit,
668	log_sibling);
669	if (!io->io_deferred)
670	io = NULL;
671	else
672	io->io_deferred = 0;
673	}
674	spin_unlock_irqrestore(&log->io_list_lock, flags);
675	if (io)
676	r5l_do_submit_io(log, io);
677	}
678
679	static void r5c_disable_writeback_async(struct work_struct *work)
680	{
681	struct r5l_log *log = container_of(work, struct r5l_log,
682	disable_writeback_work);
683	struct mddev *mddev = log->rdev->mddev;
684	struct r5conf *conf = mddev->private;
685
686	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
687	return;
688	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
689	mdname(mddev));
690
691	/* wait superblock change before suspend */
692	wait_event(mddev->sb_wait,
693	!READ_ONCE(conf->log) \|\|
694	!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
695
696	log = READ_ONCE(conf->log);
697	if (log) {
698	mddev_suspend(mddev, false);
699	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
700	mddev_resume(mddev);
701	}
702	}
703
704	static void r5l_submit_current_io(struct r5l_log *log)
705	{
706	struct r5l_io_unit *io = log->current_io;
707	struct r5l_meta_block *block;
708	unsigned long flags;
709	u32 crc;
710	bool do_submit = true;
711
712	if (!io)
713	return;
714
715	block = page_address(io->meta_page);
716	block->meta_size = cpu_to_le32(io->meta_offset);
717	crc = crc32c(log->uuid_checksum, block, PAGE_SIZE);
718	block->checksum = cpu_to_le32(crc);
719
720	log->current_io = NULL;
721	spin_lock_irqsave(&log->io_list_lock, flags);
722	if (io->has_flush \|\| io->has_fua) {
723	if (io != list_first_entry(&log->running_ios,
724	struct r5l_io_unit, log_sibling)) {
725	io->io_deferred = 1;
726	do_submit = false;
727	}
728	}
729	spin_unlock_irqrestore(&log->io_list_lock, flags);
730	if (do_submit)
731	r5l_do_submit_io(log, io);
732	}
733
734	static struct bio r5l_bio_alloc(struct r5l_log log)
735	{
736	struct bio *bio = bio_alloc_bioset(log->rdev->bdev, BIO_MAX_VECS,
737	REQ_OP_WRITE, GFP_NOIO, &log->bs);
738
739	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
740
741	return bio;
742	}
743
744	static void r5_reserve_log_entry(struct r5l_log log, struct r5l_io_unit io)
745	{
746	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
747
748	r5c_update_log_state(log);
749	/*
750	* If we filled up the log device start from the beginning again,
751	* which will require a new bio.
752	*
753	* Note: for this to work properly the log size needs to me a multiple
754	* of BLOCK_SECTORS.
755	*/
756	if (log->log_start == 0)
757	io->need_split_bio = true;
758
759	io->log_end = log->log_start;
760	}
761
762	static struct r5l_io_unit r5l_new_meta(struct r5l_log log)
763	{
764	struct r5l_io_unit *io;
765	struct r5l_meta_block *block;
766
767	io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
768	if (!io)
769	return NULL;
770	memset(io, 0, sizeof(*io));
771
772	io->log = log;
773	INIT_LIST_HEAD(&io->log_sibling);
774	INIT_LIST_HEAD(&io->stripe_list);
775	bio_list_init(&io->flush_barriers);
776	io->state = IO_UNIT_RUNNING;
777
778	io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
779	block = page_address(io->meta_page);
780	clear_page(block);
781	block->magic = cpu_to_le32(R5LOG_MAGIC);
782	block->version = R5LOG_VERSION;
783	block->seq = cpu_to_le64(log->seq);
784	block->position = cpu_to_le64(log->log_start);
785
786	io->log_start = log->log_start;
787	io->meta_offset = sizeof(struct r5l_meta_block);
788	io->seq = log->seq++;
789
790	io->current_bio = r5l_bio_alloc(log);
791	io->current_bio->bi_end_io = r5l_log_endio;
792	io->current_bio->bi_private = io;
793	__bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
794
795	r5_reserve_log_entry(log, io);
796
797	spin_lock_irq(&log->io_list_lock);
798	list_add_tail(&io->log_sibling, &log->running_ios);
799	spin_unlock_irq(&log->io_list_lock);
800
801	return io;
802	}
803
804	static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
805	{
806	if (log->current_io &&
807	log->current_io->meta_offset + payload_size > PAGE_SIZE)
808	r5l_submit_current_io(log);
809
810	if (!log->current_io) {
811	log->current_io = r5l_new_meta(log);
812	if (!log->current_io)
813	return -ENOMEM;
814	}
815
816	return 0;
817	}
818
819	static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
820	sector_t location,
821	u32 checksum1, u32 checksum2,
822	bool checksum2_valid)
823	{
824	struct r5l_io_unit *io = log->current_io;
825	struct r5l_payload_data_parity *payload;
826
827	payload = page_address(io->meta_page) + io->meta_offset;
828	payload->header.type = cpu_to_le16(type);
829	payload->header.flags = cpu_to_le16(0);
830	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
831	(PAGE_SHIFT - 9));
832	payload->location = cpu_to_le64(location);
833	payload->checksum[0] = cpu_to_le32(checksum1);
834	if (checksum2_valid)
835	payload->checksum[1] = cpu_to_le32(checksum2);
836
837	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
838	sizeof(__le32) * (1 + !!checksum2_valid);
839	}
840
841	static void r5l_append_payload_page(struct r5l_log log, struct page page)
842	{
843	struct r5l_io_unit *io = log->current_io;
844
845	if (io->need_split_bio) {
846	BUG_ON(io->split_bio);
847	io->split_bio = io->current_bio;
848	io->current_bio = r5l_bio_alloc(log);
849	bio_chain(io->current_bio, io->split_bio);
850	io->need_split_bio = false;
851	}
852
853	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
854	BUG();
855
856	r5_reserve_log_entry(log, io);
857	}
858
859	static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
860	{
861	struct mddev *mddev = log->rdev->mddev;
862	struct r5conf *conf = mddev->private;
863	struct r5l_io_unit *io;
864	struct r5l_payload_flush *payload;
865	int meta_size;
866
867	/*
868	* payload_flush requires extra writes to the journal.
869	* To avoid handling the extra IO in quiesce, just skip
870	* flush_payload
871	*/
872	if (conf->quiesce)
873	return;
874
875	mutex_lock(&log->io_mutex);
876	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
877
878	if (r5l_get_meta(log, meta_size)) {
879	mutex_unlock(&log->io_mutex);
880	return;
881	}
882
883	/* current implementation is one stripe per flush payload */
884	io = log->current_io;
885	payload = page_address(io->meta_page) + io->meta_offset;
886	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
887	payload->header.flags = cpu_to_le16(0);
888	payload->size = cpu_to_le32(sizeof(__le64));
889	payload->flush_stripes[0] = cpu_to_le64(sect);
890	io->meta_offset += meta_size;
891	/* multiple flush payloads count as one pending_stripe */
892	if (!io->has_flush_payload) {
893	io->has_flush_payload = 1;
894	atomic_inc(&io->pending_stripe);
895	}
896	mutex_unlock(&log->io_mutex);
897	}
898
899	static int r5l_log_stripe(struct r5l_log log, struct stripe_head sh,
900	int data_pages, int parity_pages)
901	{
902	int i;
903	int meta_size;
904	int ret;
905	struct r5l_io_unit *io;
906
907	meta_size =
908	((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
909	* data_pages) +
910	sizeof(struct r5l_payload_data_parity) +
911	sizeof(__le32) * parity_pages;
912
913	ret = r5l_get_meta(log, meta_size);
914	if (ret)
915	return ret;
916
917	io = log->current_io;
918
919	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
920	io->has_flush = 1;
921
922	for (i = 0; i < sh->disks; i++) {
923	if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) \|\|
924	test_bit(R5_InJournal, &sh->dev[i].flags))
925	continue;
926	if (i == sh->pd_idx \|\| i == sh->qd_idx)
927	continue;
928	if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
929	log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
930	io->has_fua = 1;
931	/*
932	* we need to flush journal to make sure recovery can
933	* reach the data with fua flag
934	*/
935	io->has_flush = 1;
936	}
937	r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
938	raid5_compute_blocknr(sh, i, 0),
939	sh->dev[i].log_checksum, 0, false);
940	r5l_append_payload_page(log, sh->dev[i].page);
941	}
942
943	if (parity_pages == 2) {
944	r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
945	sh->sector, sh->dev[sh->pd_idx].log_checksum,
946	sh->dev[sh->qd_idx].log_checksum, true);
947	r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
948	r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
949	} else if (parity_pages == 1) {
950	r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
951	sh->sector, sh->dev[sh->pd_idx].log_checksum,
952	0, false);
953	r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
954	} else /* Just writing data, not parity, in caching phase */
955	BUG_ON(parity_pages != 0);
956
957	list_add_tail(&sh->log_list, &io->stripe_list);
958	atomic_inc(&io->pending_stripe);
959	sh->log_io = io;
960
961	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
962	return 0;
963
964	if (sh->log_start == MaxSector) {
965	BUG_ON(!list_empty(&sh->r5c));
966	sh->log_start = io->log_start;
967	spin_lock_irq(&log->stripe_in_journal_lock);
968	list_add_tail(&sh->r5c,
969	&log->stripe_in_journal_list);
970	spin_unlock_irq(&log->stripe_in_journal_lock);
971	atomic_inc(&log->stripe_in_journal_count);
972	}
973	return 0;
974	}
975
976	/* add stripe to no_space_stripes, and then wake up reclaim */
977	static inline void r5l_add_no_space_stripe(struct r5l_log *log,
978	struct stripe_head *sh)
979	{
980	spin_lock(&log->no_space_stripes_lock);
981	list_add_tail(&sh->log_list, &log->no_space_stripes);
982	spin_unlock(&log->no_space_stripes_lock);
983	}
984
985	/*
986	* running in raid5d, where reclaim could wait for raid5d too (when it flushes
987	* data from log to raid disks), so we shouldn't wait for reclaim here
988	*/
989	int r5l_write_stripe(struct r5l_log log, struct stripe_head sh)
990	{
991	struct r5conf *conf = sh->raid_conf;
992	int write_disks = 0;
993	int data_pages, parity_pages;
994	int reserve;
995	int i;
996	int ret = 0;
997	bool wake_reclaim = false;
998
999	if (!log)
1000	return -EAGAIN;
1001	/* Don't support stripe batch */
1002	if (sh->log_io \|\| !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) \|\|
1003	test_bit(STRIPE_SYNCING, &sh->state)) {
1004	/* the stripe is written to log, we start writing it to raid */
1005	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1006	return -EAGAIN;
1007	}
1008
1009	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1010
1011	for (i = 0; i < sh->disks; i++) {
1012	void *addr;
1013
1014	if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) \|\|
1015	test_bit(R5_InJournal, &sh->dev[i].flags))
1016	continue;
1017
1018	write_disks++;
1019	/* checksum is already calculated in last run */
1020	if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1021	continue;
1022	addr = kmap_local_page(sh->dev[i].page);
1023	sh->dev[i].log_checksum = crc32c(log->uuid_checksum,
1024	addr, PAGE_SIZE);
1025	kunmap_local(addr);
1026	}
1027	parity_pages = 1 + !!(sh->qd_idx >= 0);
1028	data_pages = write_disks - parity_pages;
1029
1030	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1031	/*
1032	* The stripe must enter state machine again to finish the write, so
1033	* don't delay.
1034	*/
1035	clear_bit(STRIPE_DELAYED, &sh->state);
1036	atomic_inc(&sh->count);
1037
1038	mutex_lock(&log->io_mutex);
1039	/* meta + data */
1040	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1041
1042	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1043	if (!r5l_has_free_space(log, reserve)) {
1044	r5l_add_no_space_stripe(log, sh);
1045	wake_reclaim = true;
1046	} else {
1047	ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1048	if (ret) {
1049	spin_lock_irq(&log->io_list_lock);
1050	list_add_tail(&sh->log_list,
1051	&log->no_mem_stripes);
1052	spin_unlock_irq(&log->io_list_lock);
1053	}
1054	}
1055	} else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1056	/*
1057	* log space critical, do not process stripes that are
1058	* not in cache yet (sh->log_start == MaxSector).
1059	*/
1060	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1061	sh->log_start == MaxSector) {
1062	r5l_add_no_space_stripe(log, sh);
1063	wake_reclaim = true;
1064	reserve = 0;
1065	} else if (!r5l_has_free_space(log, reserve)) {
1066	if (sh->log_start == log->last_checkpoint)
1067	BUG();
1068	else
1069	r5l_add_no_space_stripe(log, sh);
1070	} else {
1071	ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1072	if (ret) {
1073	spin_lock_irq(&log->io_list_lock);
1074	list_add_tail(&sh->log_list,
1075	&log->no_mem_stripes);
1076	spin_unlock_irq(&log->io_list_lock);
1077	}
1078	}
1079	}
1080
1081	mutex_unlock(&log->io_mutex);
1082	if (wake_reclaim)
1083	r5l_wake_reclaim(log, reserve);
1084	return 0;
1085	}
1086
1087	void r5l_write_stripe_run(struct r5l_log *log)
1088	{
1089	if (!log)
1090	return;
1091	mutex_lock(&log->io_mutex);
1092	r5l_submit_current_io(log);
1093	mutex_unlock(&log->io_mutex);
1094	}
1095
1096	int r5l_handle_flush_request(struct r5l_log log, struct bio bio)
1097	{
1098	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1099	/*
1100	* in write through (journal only)
1101	* we flush log disk cache first, then write stripe data to
1102	* raid disks. So if bio is finished, the log disk cache is
1103	* flushed already. The recovery guarantees we can recovery
1104	* the bio from log disk, so we don't need to flush again
1105	*/
1106	if (bio->bi_iter.bi_size == 0) {
1107	bio_endio(bio);
1108	return 0;
1109	}
1110	bio->bi_opf &= ~REQ_PREFLUSH;
1111	} else {
1112	/* write back (with cache) */
1113	if (bio->bi_iter.bi_size == 0) {
1114	mutex_lock(&log->io_mutex);
1115	r5l_get_meta(log, 0);
1116	bio_list_add(&log->current_io->flush_barriers, bio);
1117	log->current_io->has_flush = 1;
1118	log->current_io->has_null_flush = 1;
1119	atomic_inc(&log->current_io->pending_stripe);
1120	r5l_submit_current_io(log);
1121	mutex_unlock(&log->io_mutex);
1122	return 0;
1123	}
1124	}
1125	return -EAGAIN;
1126	}
1127
1128	/* This will run after log space is reclaimed */
1129	static void r5l_run_no_space_stripes(struct r5l_log *log)
1130	{
1131	struct stripe_head *sh;
1132
1133	spin_lock(&log->no_space_stripes_lock);
1134	while (!list_empty(&log->no_space_stripes)) {
1135	sh = list_first_entry(&log->no_space_stripes,
1136	struct stripe_head, log_list);
1137	list_del_init(&sh->log_list);
1138	set_bit(STRIPE_HANDLE, &sh->state);
1139	raid5_release_stripe(sh);
1140	}
1141	spin_unlock(&log->no_space_stripes_lock);
1142	}
1143
1144	/*
1145	* calculate new last_checkpoint
1146	* for write through mode, returns log->next_checkpoint
1147	* for write back, returns log_start of first sh in stripe_in_journal_list
1148	*/
1149	static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1150	{
1151	struct stripe_head *sh;
1152	struct r5l_log *log = READ_ONCE(conf->log);
1153	sector_t new_cp;
1154	unsigned long flags;
1155
1156	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1157	return log->next_checkpoint;
1158
1159	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1160	if (list_empty(&log->stripe_in_journal_list)) {
1161	/* all stripes flushed */
1162	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1163	return log->next_checkpoint;
1164	}
1165	sh = list_first_entry(&log->stripe_in_journal_list,
1166	struct stripe_head, r5c);
1167	new_cp = sh->log_start;
1168	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1169	return new_cp;
1170	}
1171
1172	static sector_t r5l_reclaimable_space(struct r5l_log *log)
1173	{
1174	struct r5conf *conf = log->rdev->mddev->private;
1175
1176	return r5l_ring_distance(log, log->last_checkpoint,
1177	r5c_calculate_new_cp(conf));
1178	}
1179
1180	static void r5l_run_no_mem_stripe(struct r5l_log *log)
1181	{
1182	struct stripe_head *sh;
1183
1184	lockdep_assert_held(&log->io_list_lock);
1185
1186	if (!list_empty(&log->no_mem_stripes)) {
1187	sh = list_first_entry(&log->no_mem_stripes,
1188	struct stripe_head, log_list);
1189	list_del_init(&sh->log_list);
1190	set_bit(STRIPE_HANDLE, &sh->state);
1191	raid5_release_stripe(sh);
1192	}
1193	}
1194
1195	static bool r5l_complete_finished_ios(struct r5l_log *log)
1196	{
1197	struct r5l_io_unit io, next;
1198	bool found = false;
1199
1200	lockdep_assert_held(&log->io_list_lock);
1201
1202	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1203	/* don't change list order */
1204	if (io->state < IO_UNIT_STRIPE_END)
1205	break;
1206
1207	log->next_checkpoint = io->log_start;
1208
1209	list_del(&io->log_sibling);
1210	mempool_free(io, &log->io_pool);
1211	r5l_run_no_mem_stripe(log);
1212
1213	found = true;
1214	}
1215
1216	return found;
1217	}
1218
1219	static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1220	{
1221	struct r5l_log *log = io->log;
1222	struct r5conf *conf = log->rdev->mddev->private;
1223	unsigned long flags;
1224
1225	spin_lock_irqsave(&log->io_list_lock, flags);
1226	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1227
1228	if (!r5l_complete_finished_ios(log)) {
1229	spin_unlock_irqrestore(&log->io_list_lock, flags);
1230	return;
1231	}
1232
1233	if (r5l_reclaimable_space(log) > log->max_free_space \|\|
1234	test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1235	r5l_wake_reclaim(log, 0);
1236
1237	spin_unlock_irqrestore(&log->io_list_lock, flags);
1238	wake_up(&log->iounit_wait);
1239	}
1240
1241	void r5l_stripe_write_finished(struct stripe_head *sh)
1242	{
1243	struct r5l_io_unit *io;
1244
1245	io = sh->log_io;
1246	sh->log_io = NULL;
1247
1248	if (io && atomic_dec_and_test(&io->pending_stripe))
1249	__r5l_stripe_write_finished(io);
1250	}
1251
1252	static void r5l_log_flush_endio(struct bio *bio)
1253	{
1254	struct r5l_log *log = container_of(bio, struct r5l_log,
1255	flush_bio);
1256	unsigned long flags;
1257	struct r5l_io_unit *io;
1258
1259	if (bio->bi_status)
1260	md_error(log->rdev->mddev, log->rdev);
1261	bio_uninit(bio);
1262
1263	spin_lock_irqsave(&log->io_list_lock, flags);
1264	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1265	r5l_io_run_stripes(io);
1266	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1267	spin_unlock_irqrestore(&log->io_list_lock, flags);
1268	}
1269
1270	/*
1271	* Starting dispatch IO to raid.
1272	* io_unit(meta) consists of a log. There is one situation we want to avoid. A
1273	* broken meta in the middle of a log causes recovery can't find meta at the
1274	* head of log. If operations require meta at the head persistent in log, we
1275	* must make sure meta before it persistent in log too. A case is:
1276	*
1277	* stripe data/parity is in log, we start write stripe to raid disks. stripe
1278	* data/parity must be persistent in log before we do the write to raid disks.
1279	*
1280	* The solution is we restrictly maintain io_unit list order. In this case, we
1281	* only write stripes of an io_unit to raid disks till the io_unit is the first
1282	* one whose data/parity is in log.
1283	*/
1284	void r5l_flush_stripe_to_raid(struct r5l_log *log)
1285	{
1286	bool do_flush;
1287
1288	if (!log \|\| !log->need_cache_flush)
1289	return;
1290
1291	spin_lock_irq(&log->io_list_lock);
1292	/* flush bio is running */
1293	if (!list_empty(&log->flushing_ios)) {
1294	spin_unlock_irq(&log->io_list_lock);
1295	return;
1296	}
1297	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1298	do_flush = !list_empty(&log->flushing_ios);
1299	spin_unlock_irq(&log->io_list_lock);
1300
1301	if (!do_flush)
1302	return;
1303	bio_init(&log->flush_bio, log->rdev->bdev, NULL, 0,
1304	REQ_OP_WRITE \| REQ_PREFLUSH);
1305	log->flush_bio.bi_end_io = r5l_log_flush_endio;
1306	submit_bio(&log->flush_bio);
1307	}
1308
1309	static void r5l_write_super(struct r5l_log *log, sector_t cp);
1310	static void r5l_write_super_and_discard_space(struct r5l_log *log,
1311	sector_t end)
1312	{
1313	struct block_device *bdev = log->rdev->bdev;
1314	struct mddev *mddev;
1315
1316	r5l_write_super(log, end);
1317
1318	if (!bdev_max_discard_sectors(bdev))
1319	return;
1320
1321	mddev = log->rdev->mddev;
1322	/*
1323	* Discard could zero data, so before discard we must make sure
1324	* superblock is updated to new log tail. Updating superblock (either
1325	* directly call md_update_sb() or depend on md thread) must hold
1326	* reconfig mutex. On the other hand, raid5_quiesce is called with
1327	* reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1328	* for all IO finish, hence waiting for reclaim thread, while reclaim
1329	* thread is calling this function and waiting for reconfig mutex. So
1330	* there is a deadlock. We workaround this issue with a trylock.
1331	* FIXME: we could miss discard if we can't take reconfig mutex
1332	*/
1333	set_mask_bits(&mddev->sb_flags, 0,
1334	BIT(MD_SB_CHANGE_DEVS) \| BIT(MD_SB_CHANGE_PENDING));
1335	if (!mddev_trylock(mddev))
1336	return;
1337	md_update_sb(mddev, 1);
1338	mddev_unlock(mddev);
1339
1340	/* discard IO error really doesn't matter, ignore it */
1341	if (log->last_checkpoint < end) {
1342	blkdev_issue_discard(bdev,
1343	log->last_checkpoint + log->rdev->data_offset,
1344	end - log->last_checkpoint, GFP_NOIO);
1345	} else {
1346	blkdev_issue_discard(bdev,
1347	log->last_checkpoint + log->rdev->data_offset,
1348	log->device_size - log->last_checkpoint,
1349	GFP_NOIO);
1350	blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1351	GFP_NOIO);
1352	}
1353	}
1354
1355	/*
1356	* r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1357	* the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1358	*
1359	* must hold conf->device_lock
1360	*/
1361	static void r5c_flush_stripe(struct r5conf conf, struct stripe_head sh)
1362	{
1363	BUG_ON(list_empty(&sh->lru));
1364	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1365	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1366
1367	/*
1368	* The stripe is not ON_RELEASE_LIST, so it is safe to call
1369	* raid5_release_stripe() while holding conf->device_lock
1370	*/
1371	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1372	lockdep_assert_held(&conf->device_lock);
1373
1374	list_del_init(&sh->lru);
1375	atomic_inc(&sh->count);
1376
1377	set_bit(STRIPE_HANDLE, &sh->state);
1378	atomic_inc(&conf->active_stripes);
1379	r5c_make_stripe_write_out(sh);
1380
1381	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1382	atomic_inc(&conf->r5c_flushing_partial_stripes);
1383	else
1384	atomic_inc(&conf->r5c_flushing_full_stripes);
1385	raid5_release_stripe(sh);
1386	}
1387
1388	/*
1389	* if num == 0, flush all full stripes
1390	* if num > 0, flush all full stripes. If less than num full stripes are
1391	* flushed, flush some partial stripes until totally num stripes are
1392	* flushed or there is no more cached stripes.
1393	*/
1394	void r5c_flush_cache(struct r5conf *conf, int num)
1395	{
1396	int count;
1397	struct stripe_head sh, next;
1398
1399	lockdep_assert_held(&conf->device_lock);
1400	if (!READ_ONCE(conf->log))
1401	return;
1402
1403	count = 0;
1404	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1405	r5c_flush_stripe(conf, sh);
1406	count++;
1407	}
1408
1409	if (count >= num)
1410	return;
1411	list_for_each_entry_safe(sh, next,
1412	&conf->r5c_partial_stripe_list, lru) {
1413	r5c_flush_stripe(conf, sh);
1414	if (++count >= num)
1415	break;
1416	}
1417	}
1418
1419	static void r5c_do_reclaim(struct r5conf *conf)
1420	{
1421	struct r5l_log *log = READ_ONCE(conf->log);
1422	struct stripe_head *sh;
1423	int count = 0;
1424	unsigned long flags;
1425	int total_cached;
1426	int stripes_to_flush;
1427	int flushing_partial, flushing_full;
1428
1429	if (!r5c_is_writeback(log))
1430	return;
1431
1432	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1433	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1434	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1435	atomic_read(&conf->r5c_cached_full_stripes) -
1436	flushing_full - flushing_partial;
1437
1438	if (total_cached > conf->min_nr_stripes * 3 / 4 \|\|
1439	atomic_read(&conf->empty_inactive_list_nr) > 0)
1440	/*
1441	* if stripe cache pressure high, flush all full stripes and
1442	* some partial stripes
1443	*/
1444	stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1445	else if (total_cached > conf->min_nr_stripes * 1 / 2 \|\|
1446	atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1447	R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1448	/*
1449	* if stripe cache pressure moderate, or if there is many full
1450	* stripes,flush all full stripes
1451	*/
1452	stripes_to_flush = 0;
1453	else
1454	/* no need to flush */
1455	stripes_to_flush = -1;
1456
1457	if (stripes_to_flush >= 0) {
1458	spin_lock_irqsave(&conf->device_lock, flags);
1459	r5c_flush_cache(conf, stripes_to_flush);
1460	spin_unlock_irqrestore(&conf->device_lock, flags);
1461	}
1462
1463	/* if log space is tight, flush stripes on stripe_in_journal_list */
1464	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1465	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1466	spin_lock(&conf->device_lock);
1467	list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1468	/*
1469	* stripes on stripe_in_journal_list could be in any
1470	* state of the stripe_cache state machine. In this
1471	* case, we only want to flush stripe on
1472	* r5c_cached_full/partial_stripes. The following
1473	* condition makes sure the stripe is on one of the
1474	* two lists.
1475	*/
1476	if (!list_empty(&sh->lru) &&
1477	!test_bit(STRIPE_HANDLE, &sh->state) &&
1478	atomic_read(&sh->count) == 0) {
1479	r5c_flush_stripe(conf, sh);
1480	if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1481	break;
1482	}
1483	}
1484	spin_unlock(&conf->device_lock);
1485	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1486	}
1487
1488	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1489	r5l_run_no_space_stripes(log);
1490
1491	md_wakeup_thread(conf->mddev->thread);
1492	}
1493
1494	static void r5l_do_reclaim(struct r5l_log *log)
1495	{
1496	struct r5conf *conf = log->rdev->mddev->private;
1497	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1498	sector_t reclaimable;
1499	sector_t next_checkpoint;
1500	bool write_super;
1501
1502	spin_lock_irq(&log->io_list_lock);
1503	write_super = r5l_reclaimable_space(log) > log->max_free_space \|\|
1504	reclaim_target != 0 \|\| !list_empty(&log->no_space_stripes);
1505	/*
1506	* move proper io_unit to reclaim list. We should not change the order.
1507	* reclaimable/unreclaimable io_unit can be mixed in the list, we
1508	* shouldn't reuse space of an unreclaimable io_unit
1509	*/
1510	while (1) {
1511	reclaimable = r5l_reclaimable_space(log);
1512	if (reclaimable >= reclaim_target \|\|
1513	(list_empty(&log->running_ios) &&
1514	list_empty(&log->io_end_ios) &&
1515	list_empty(&log->flushing_ios) &&
1516	list_empty(&log->finished_ios)))
1517	break;
1518
1519	md_wakeup_thread(log->rdev->mddev->thread);
1520	wait_event_lock_irq(log->iounit_wait,
1521	r5l_reclaimable_space(log) > reclaimable,
1522	log->io_list_lock);
1523	}
1524
1525	next_checkpoint = r5c_calculate_new_cp(conf);
1526	spin_unlock_irq(&log->io_list_lock);
1527
1528	if (reclaimable == 0 \|\| !write_super)
1529	return;
1530
1531	/*
1532	* write_super will flush cache of each raid disk. We must write super
1533	* here, because the log area might be reused soon and we don't want to
1534	* confuse recovery
1535	*/
1536	r5l_write_super_and_discard_space(log, next_checkpoint);
1537
1538	mutex_lock(&log->io_mutex);
1539	log->last_checkpoint = next_checkpoint;
1540	r5c_update_log_state(log);
1541	mutex_unlock(&log->io_mutex);
1542
1543	r5l_run_no_space_stripes(log);
1544	}
1545
1546	static void r5l_reclaim_thread(struct md_thread *thread)
1547	{
1548	struct mddev *mddev = thread->mddev;
1549	struct r5conf *conf = mddev->private;
1550	struct r5l_log *log = READ_ONCE(conf->log);
1551
1552	if (!log)
1553	return;
1554	r5c_do_reclaim(conf);
1555	r5l_do_reclaim(log);
1556	}
1557
1558	void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1559	{
1560	unsigned long target;
1561	unsigned long new = (unsigned long)space; /* overflow in theory */
1562
1563	if (!log)
1564	return;
1565
1566	target = READ_ONCE(log->reclaim_target);
1567	do {
1568	if (new < target)
1569	return;
1570	} while (!try_cmpxchg(&log->reclaim_target, &target, new));
1571	md_wakeup_thread(log->reclaim_thread);
1572	}
1573
1574	void r5l_quiesce(struct r5l_log *log, int quiesce)
1575	{
1576	struct mddev *mddev = log->rdev->mddev;
1577	struct md_thread *thread = rcu_dereference_protected(
1578	log->reclaim_thread, lockdep_is_held(&mddev->reconfig_mutex));
1579
1580	if (quiesce) {
1581	/* make sure r5l_write_super_and_discard_space exits */
1582	wake_up(&mddev->sb_wait);
1583	kthread_park(thread->tsk);
1584	r5l_wake_reclaim(log, MaxSector);
1585	r5l_do_reclaim(log);
1586	} else
1587	kthread_unpark(thread->tsk);
1588	}
1589
1590	bool r5l_log_disk_error(struct r5conf *conf)
1591	{
1592	struct r5l_log *log = READ_ONCE(conf->log);
1593
1594	/* don't allow write if journal disk is missing */
1595	if (!log)
1596	return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1597	else
1598	return test_bit(Faulty, &log->rdev->flags);
1599	}
1600
1601	#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1602
1603	struct r5l_recovery_ctx {
1604	struct page meta_page; / current meta */
1605	sector_t meta_total_blocks; /* total size of current meta and data */
1606	sector_t pos; /* recovery position */
1607	u64 seq; /* recovery position seq */
1608	int data_parity_stripes; /* number of data_parity stripes */
1609	int data_only_stripes; /* number of data_only stripes */
1610	struct list_head cached_list;
1611
1612	/*
1613	* read ahead page pool (ra_pool)
1614	* in recovery, log is read sequentially. It is not efficient to
1615	* read every page with sync_page_io(). The read ahead page pool
1616	* reads multiple pages with one IO, so further log read can
1617	* just copy data from the pool.
1618	*/
1619	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1620	struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1621	sector_t pool_offset; /* offset of first page in the pool */
1622	int total_pages; /* total allocated pages */
1623	int valid_pages; /* pages with valid data */
1624	};
1625
1626	static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1627	struct r5l_recovery_ctx *ctx)
1628	{
1629	struct page *page;
1630
1631	ctx->valid_pages = 0;
1632	ctx->total_pages = 0;
1633	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1634	page = alloc_page(GFP_KERNEL);
1635
1636	if (!page)
1637	break;
1638	ctx->ra_pool[ctx->total_pages] = page;
1639	ctx->total_pages += 1;
1640	}
1641
1642	if (ctx->total_pages == 0)
1643	return -ENOMEM;
1644
1645	ctx->pool_offset = 0;
1646	return 0;
1647	}
1648
1649	static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1650	struct r5l_recovery_ctx *ctx)
1651	{
1652	int i;
1653
1654	for (i = 0; i < ctx->total_pages; ++i)
1655	put_page(ctx->ra_pool[i]);
1656	}
1657
1658	/*
1659	* fetch ctx->valid_pages pages from offset
1660	* In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1661	* However, if the offset is close to the end of the journal device,
1662	* ctx->valid_pages could be smaller than ctx->total_pages
1663	*/
1664	static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1665	struct r5l_recovery_ctx *ctx,
1666	sector_t offset)
1667	{
1668	struct bio bio;
1669	int ret;
1670
1671	bio_init(&bio, log->rdev->bdev, ctx->ra_bvec,
1672	R5L_RECOVERY_PAGE_POOL_SIZE, REQ_OP_READ);
1673	bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1674
1675	ctx->valid_pages = 0;
1676	ctx->pool_offset = offset;
1677
1678	while (ctx->valid_pages < ctx->total_pages) {
1679	__bio_add_page(&bio, ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1680	0);
1681	ctx->valid_pages += 1;
1682
1683	offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1684
1685	if (offset == 0) /* reached end of the device */
1686	break;
1687	}
1688
1689	ret = submit_bio_wait(&bio);
1690	bio_uninit(&bio);
1691	return ret;
1692	}
1693
1694	/*
1695	* try read a page from the read ahead page pool, if the page is not in the
1696	* pool, call r5l_recovery_fetch_ra_pool
1697	*/
1698	static int r5l_recovery_read_page(struct r5l_log *log,
1699	struct r5l_recovery_ctx *ctx,
1700	struct page *page,
1701	sector_t offset)
1702	{
1703	int ret;
1704
1705	if (offset < ctx->pool_offset \|\|
1706	offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1707	ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1708	if (ret)
1709	return ret;
1710	}
1711
1712	BUG_ON(offset < ctx->pool_offset \|\|
1713	offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1714
1715	memcpy(page_address(page),
1716	page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1717	BLOCK_SECTOR_SHIFT]),
1718	PAGE_SIZE);
1719	return 0;
1720	}
1721
1722	static int r5l_recovery_read_meta_block(struct r5l_log *log,
1723	struct r5l_recovery_ctx *ctx)
1724	{
1725	struct page *page = ctx->meta_page;
1726	struct r5l_meta_block *mb;
1727	u32 crc, stored_crc;
1728	int ret;
1729
1730	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1731	if (ret != 0)
1732	return ret;
1733
1734	mb = page_address(page);
1735	stored_crc = le32_to_cpu(mb->checksum);
1736	mb->checksum = 0;
1737
1738	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC \|\|
1739	le64_to_cpu(mb->seq) != ctx->seq \|\|
1740	mb->version != R5LOG_VERSION \|\|
1741	le64_to_cpu(mb->position) != ctx->pos)
1742	return -EINVAL;
1743
1744	crc = crc32c(log->uuid_checksum, mb, PAGE_SIZE);
1745	if (stored_crc != crc)
1746	return -EINVAL;
1747
1748	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1749	return -EINVAL;
1750
1751	ctx->meta_total_blocks = BLOCK_SECTORS;
1752
1753	return 0;
1754	}
1755
1756	static void
1757	r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1758	struct page *page,
1759	sector_t pos, u64 seq)
1760	{
1761	struct r5l_meta_block *mb;
1762
1763	mb = page_address(page);
1764	clear_page(mb);
1765	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1766	mb->version = R5LOG_VERSION;
1767	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1768	mb->seq = cpu_to_le64(seq);
1769	mb->position = cpu_to_le64(pos);
1770	}
1771
1772	static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1773	u64 seq)
1774	{
1775	struct page *page;
1776	struct r5l_meta_block *mb;
1777
1778	page = alloc_page(GFP_KERNEL);
1779	if (!page)
1780	return -ENOMEM;
1781	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1782	mb = page_address(page);
1783	mb->checksum = cpu_to_le32(crc32c(log->uuid_checksum, mb, PAGE_SIZE));
1784	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE \|
1785	REQ_SYNC \| REQ_FUA, false)) {
1786	__free_page(page);
1787	return -EIO;
1788	}
1789	__free_page(page);
1790	return 0;
1791	}
1792
1793	/*
1794	* r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1795	* to mark valid (potentially not flushed) data in the journal.
1796	*
1797	* We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1798	* so there should not be any mismatch here.
1799	*/
1800	static void r5l_recovery_load_data(struct r5l_log *log,
1801	struct stripe_head *sh,
1802	struct r5l_recovery_ctx *ctx,
1803	struct r5l_payload_data_parity *payload,
1804	sector_t log_offset)
1805	{
1806	struct mddev *mddev = log->rdev->mddev;
1807	struct r5conf *conf = mddev->private;
1808	int dd_idx;
1809
1810	raid5_compute_sector(conf,
1811	le64_to_cpu(payload->location), 0,
1812	&dd_idx, sh);
1813	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1814	sh->dev[dd_idx].log_checksum =
1815	le32_to_cpu(payload->checksum[0]);
1816	ctx->meta_total_blocks += BLOCK_SECTORS;
1817
1818	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1819	set_bit(STRIPE_R5C_CACHING, &sh->state);
1820	}
1821
1822	static void r5l_recovery_load_parity(struct r5l_log *log,
1823	struct stripe_head *sh,
1824	struct r5l_recovery_ctx *ctx,
1825	struct r5l_payload_data_parity *payload,
1826	sector_t log_offset)
1827	{
1828	struct mddev *mddev = log->rdev->mddev;
1829	struct r5conf *conf = mddev->private;
1830
1831	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1832	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1833	sh->dev[sh->pd_idx].log_checksum =
1834	le32_to_cpu(payload->checksum[0]);
1835	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1836
1837	if (sh->qd_idx >= 0) {
1838	r5l_recovery_read_page(
1839	log, ctx, sh->dev[sh->qd_idx].page,
1840	r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1841	sh->dev[sh->qd_idx].log_checksum =
1842	le32_to_cpu(payload->checksum[1]);
1843	set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1844	}
1845	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1846	}
1847
1848	static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1849	{
1850	int i;
1851
1852	sh->state = 0;
1853	sh->log_start = MaxSector;
1854	for (i = sh->disks; i--; )
1855	sh->dev[i].flags = 0;
1856	}
1857
1858	static void
1859	r5l_recovery_replay_one_stripe(struct r5conf *conf,
1860	struct stripe_head *sh,
1861	struct r5l_recovery_ctx *ctx)
1862	{
1863	struct md_rdev rdev, rrdev;
1864	int disk_index;
1865	int data_count = 0;
1866
1867	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1868	if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1869	continue;
1870	if (disk_index == sh->qd_idx \|\| disk_index == sh->pd_idx)
1871	continue;
1872	data_count++;
1873	}
1874
1875	/*
1876	* stripes that only have parity must have been flushed
1877	* before the crash that we are now recovering from, so
1878	* there is nothing more to recovery.
1879	*/
1880	if (data_count == 0)
1881	goto out;
1882
1883	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1884	if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1885	continue;
1886
1887	/* in case device is broken */
1888	rdev = conf->disks[disk_index].rdev;
1889	if (rdev) {
1890	atomic_inc(&rdev->nr_pending);
1891	sync_page_io(rdev, sh->sector, PAGE_SIZE,
1892	sh->dev[disk_index].page, REQ_OP_WRITE,
1893	false);
1894	rdev_dec_pending(rdev, rdev->mddev);
1895	}
1896	rrdev = conf->disks[disk_index].replacement;
1897	if (rrdev) {
1898	atomic_inc(&rrdev->nr_pending);
1899	sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1900	sh->dev[disk_index].page, REQ_OP_WRITE,
1901	false);
1902	rdev_dec_pending(rrdev, rrdev->mddev);
1903	}
1904	}
1905	ctx->data_parity_stripes++;
1906	out:
1907	r5l_recovery_reset_stripe(sh);
1908	}
1909
1910	static struct stripe_head *
1911	r5c_recovery_alloc_stripe(
1912	struct r5conf *conf,
1913	sector_t stripe_sect,
1914	int noblock)
1915	{
1916	struct stripe_head *sh;
1917
1918	sh = raid5_get_active_stripe(conf, NULL, stripe_sect,
1919	noblock ? R5_GAS_NOBLOCK : 0);
1920	if (!sh)
1921	return NULL; /* no more stripe available */
1922
1923	r5l_recovery_reset_stripe(sh);
1924
1925	return sh;
1926	}
1927
1928	static struct stripe_head *
1929	r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1930	{
1931	struct stripe_head *sh;
1932
1933	list_for_each_entry(sh, list, lru)
1934	if (sh->sector == sect)
1935	return sh;
1936	return NULL;
1937	}
1938
1939	static void
1940	r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1941	struct r5l_recovery_ctx *ctx)
1942	{
1943	struct stripe_head sh, next;
1944
1945	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1946	r5l_recovery_reset_stripe(sh);
1947	list_del_init(&sh->lru);
1948	raid5_release_stripe(sh);
1949	}
1950	}
1951
1952	static void
1953	r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1954	struct r5l_recovery_ctx *ctx)
1955	{
1956	struct stripe_head sh, next;
1957
1958	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1959	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1960	r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1961	list_del_init(&sh->lru);
1962	raid5_release_stripe(sh);
1963	}
1964	}
1965
1966	/* if matches return 0; otherwise return -EINVAL */
1967	static int
1968	r5l_recovery_verify_data_checksum(struct r5l_log *log,
1969	struct r5l_recovery_ctx *ctx,
1970	struct page *page,
1971	sector_t log_offset, __le32 log_checksum)
1972	{
1973	void *addr;
1974	u32 checksum;
1975
1976	r5l_recovery_read_page(log, ctx, page, log_offset);
1977	addr = kmap_local_page(page);
1978	checksum = crc32c(log->uuid_checksum, addr, PAGE_SIZE);
1979	kunmap_local(addr);
1980	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1981	}
1982
1983	/*
1984	* before loading data to stripe cache, we need verify checksum for all data,
1985	* if there is mismatch for any data page, we drop all data in the mata block
1986	*/
1987	static int
1988	r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1989	struct r5l_recovery_ctx *ctx)
1990	{
1991	struct mddev *mddev = log->rdev->mddev;
1992	struct r5conf *conf = mddev->private;
1993	struct r5l_meta_block *mb = page_address(ctx->meta_page);
1994	sector_t mb_offset = sizeof(struct r5l_meta_block);
1995	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1996	struct page *page;
1997	struct r5l_payload_data_parity *payload;
1998	struct r5l_payload_flush *payload_flush;
1999
2000	page = alloc_page(GFP_KERNEL);
2001	if (!page)
2002	return -ENOMEM;
2003
2004	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2005	sector_t payload_len;
2006
2007	payload = (void *)mb + mb_offset;
2008	payload_flush = (void *)mb + mb_offset;
2009
2010	if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2011	payload_len = sizeof(struct r5l_payload_data_parity) +
2012	(sector_t)sizeof(__le32) *
2013	(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2014	if (mb_offset + payload_len > le32_to_cpu(mb->meta_size))
2015	goto mismatch;
2016	if (r5l_recovery_verify_data_checksum(
2017	log, ctx, page, log_offset,
2018	payload->checksum[0]) < 0)
2019	goto mismatch;
2020	} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2021	payload_len = sizeof(struct r5l_payload_data_parity) +
2022	(sector_t)sizeof(__le32) *
2023	(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2024	if (mb_offset + payload_len > le32_to_cpu(mb->meta_size))
2025	goto mismatch;
2026	if (r5l_recovery_verify_data_checksum(
2027	log, ctx, page, log_offset,
2028	payload->checksum[0]) < 0)
2029	goto mismatch;
2030	if (conf->max_degraded == 2 && /* q for RAID 6 */
2031	r5l_recovery_verify_data_checksum(
2032	log, ctx, page,
2033	r5l_ring_add(log, log_offset,
2034	BLOCK_SECTORS),
2035	payload->checksum[1]) < 0)
2036	goto mismatch;
2037	} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2038	payload_len = sizeof(struct r5l_payload_flush) +
2039	(sector_t)le32_to_cpu(payload_flush->size);
2040	if (mb_offset + payload_len > le32_to_cpu(mb->meta_size))
2041	goto mismatch;
2042	} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2043	goto mismatch;
2044
2045	if (le16_to_cpu(payload->header.type) != R5LOG_PAYLOAD_FLUSH) {
2046	log_offset = r5l_ring_add(log, log_offset,
2047	le32_to_cpu(payload->size));
2048	}
2049	mb_offset += payload_len;
2050	}
2051
2052	put_page(page);
2053	return 0;
2054
2055	mismatch:
2056	put_page(page);
2057	return -EINVAL;
2058	}
2059
2060	/*
2061	* Analyze all data/parity pages in one meta block
2062	* Returns:
2063	* 0 for success
2064	* -EINVAL for unknown playload type
2065	* -EAGAIN for checksum mismatch of data page
2066	* -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2067	*/
2068	static int
2069	r5c_recovery_analyze_meta_block(struct r5l_log *log,
2070	struct r5l_recovery_ctx *ctx,
2071	struct list_head *cached_stripe_list)
2072	{
2073	struct mddev *mddev = log->rdev->mddev;
2074	struct r5conf *conf = mddev->private;
2075	struct r5l_meta_block *mb;
2076	struct r5l_payload_data_parity *payload;
2077	struct r5l_payload_flush *payload_flush;
2078	int mb_offset;
2079	sector_t log_offset;
2080	sector_t stripe_sect;
2081	struct stripe_head *sh;
2082	int ret;
2083
2084	/*
2085	* for mismatch in data blocks, we will drop all data in this mb, but
2086	* we will still read next mb for other data with FLUSH flag, as
2087	* io_unit could finish out of order.
2088	*/
2089	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2090	if (ret == -EINVAL)
2091	return -EAGAIN;
2092	else if (ret)
2093	return ret; /* -ENOMEM duo to alloc_page() failed */
2094
2095	mb = page_address(ctx->meta_page);
2096	mb_offset = sizeof(struct r5l_meta_block);
2097	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2098
2099	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2100	sector_t payload_len;
2101	int dd;
2102
2103	payload = (void *)mb + mb_offset;
2104	payload_flush = (void *)mb + mb_offset;
2105
2106	if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2107	int i, count;
2108
2109	payload_len = sizeof(struct r5l_payload_flush) +
2110	(sector_t)le32_to_cpu(payload_flush->size);
2111	if (mb_offset + payload_len >
2112	le32_to_cpu(mb->meta_size))
2113	return -EINVAL;
2114
2115	count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2116	for (i = 0; i < count; ++i) {
2117	stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2118	sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2119	stripe_sect);
2120	if (sh) {
2121	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2122	r5l_recovery_reset_stripe(sh);
2123	list_del_init(&sh->lru);
2124	raid5_release_stripe(sh);
2125	}
2126	}
2127
2128	mb_offset += payload_len;
2129	continue;
2130	}
2131
2132	/* DATA or PARITY payload */
2133	payload_len = sizeof(struct r5l_payload_data_parity) +
2134	(sector_t)sizeof(__le32) *
2135	(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2136	if (mb_offset + payload_len > le32_to_cpu(mb->meta_size))
2137	return -EINVAL;
2138
2139	stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2140	raid5_compute_sector(
2141	conf, le64_to_cpu(payload->location), 0, &dd,
2142	NULL)
2143	: le64_to_cpu(payload->location);
2144
2145	sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2146	stripe_sect);
2147
2148	if (!sh) {
2149	sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2150	/*
2151	* cannot get stripe from raid5_get_active_stripe
2152	* try replay some stripes
2153	*/
2154	if (!sh) {
2155	r5c_recovery_replay_stripes(
2156	cached_stripe_list, ctx);
2157	sh = r5c_recovery_alloc_stripe(
2158	conf, stripe_sect, 1);
2159	}
2160	if (!sh) {
2161	int new_size = conf->min_nr_stripes * 2;
2162	pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2163	mdname(mddev),
2164	new_size);
2165	ret = raid5_set_cache_size(mddev, new_size);
2166	if (conf->min_nr_stripes <= new_size / 2) {
2167	pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2168	mdname(mddev),
2169	ret,
2170	new_size,
2171	conf->min_nr_stripes,
2172	conf->max_nr_stripes);
2173	return -ENOMEM;
2174	}
2175	sh = r5c_recovery_alloc_stripe(
2176	conf, stripe_sect, 0);
2177	}
2178	if (!sh) {
2179	pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2180	mdname(mddev));
2181	return -ENOMEM;
2182	}
2183	list_add_tail(&sh->lru, cached_stripe_list);
2184	}
2185
2186	if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2187	if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2188	test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2189	r5l_recovery_replay_one_stripe(conf, sh, ctx);
2190	list_move_tail(&sh->lru, cached_stripe_list);
2191	}
2192	r5l_recovery_load_data(log, sh, ctx, payload,
2193	log_offset);
2194	} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2195	r5l_recovery_load_parity(log, sh, ctx, payload,
2196	log_offset);
2197	else
2198	return -EINVAL;
2199
2200	log_offset = r5l_ring_add(log, log_offset,
2201	le32_to_cpu(payload->size));
2202
2203	mb_offset += payload_len;
2204	}
2205
2206	return 0;
2207	}
2208
2209	/*
2210	* Load the stripe into cache. The stripe will be written out later by
2211	* the stripe cache state machine.
2212	*/
2213	static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2214	struct stripe_head *sh)
2215	{
2216	struct r5dev *dev;
2217	int i;
2218
2219	for (i = sh->disks; i--; ) {
2220	dev = sh->dev + i;
2221	if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2222	set_bit(R5_InJournal, &dev->flags);
2223	set_bit(R5_UPTODATE, &dev->flags);
2224	}
2225	}
2226	}
2227
2228	/*
2229	* Scan through the log for all to-be-flushed data
2230	*
2231	* For stripes with data and parity, namely Data-Parity stripe
2232	* (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2233	*
2234	* For stripes with only data, namely Data-Only stripe
2235	* (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2236	*
2237	* For a stripe, if we see data after parity, we should discard all previous
2238	* data and parity for this stripe, as these data are already flushed to
2239	* the array.
2240	*
2241	* At the end of the scan, we return the new journal_tail, which points to
2242	* first data-only stripe on the journal device, or next invalid meta block.
2243	*/
2244	static int r5c_recovery_flush_log(struct r5l_log *log,
2245	struct r5l_recovery_ctx *ctx)
2246	{
2247	struct stripe_head *sh;
2248	int ret = 0;
2249
2250	/* scan through the log */
2251	while (1) {
2252	if (r5l_recovery_read_meta_block(log, ctx))
2253	break;
2254
2255	ret = r5c_recovery_analyze_meta_block(log, ctx,
2256	&ctx->cached_list);
2257	/*
2258	* -EAGAIN means mismatch in data block, in this case, we still
2259	* try scan the next metablock
2260	*/
2261	if (ret && ret != -EAGAIN)
2262	break; /* ret == -EINVAL or -ENOMEM */
2263	ctx->seq++;
2264	ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2265	}
2266
2267	if (ret == -ENOMEM) {
2268	r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2269	return ret;
2270	}
2271
2272	/* replay data-parity stripes */
2273	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2274
2275	/* load data-only stripes to stripe cache */
2276	list_for_each_entry(sh, &ctx->cached_list, lru) {
2277	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2278	r5c_recovery_load_one_stripe(log, sh);
2279	ctx->data_only_stripes++;
2280	}
2281
2282	return 0;
2283	}
2284
2285	/*
2286	* we did a recovery. Now ctx.pos points to an invalid meta block. New
2287	* log will start here. but we can't let superblock point to last valid
2288	* meta block. The log might looks like:
2289	* \| meta 1\| meta 2\| meta 3\|
2290	* meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2291	* superblock points to meta 1, we write a new valid meta 2n. if crash
2292	* happens again, new recovery will start from meta 1. Since meta 2n is
2293	* valid now, recovery will think meta 3 is valid, which is wrong.
2294	* The solution is we create a new meta in meta2 with its seq == meta
2295	* 1's seq + 10000 and let superblock points to meta2. The same recovery
2296	* will not think meta 3 is a valid meta, because its seq doesn't match
2297	*/
2298
2299	/*
2300	* Before recovery, the log looks like the following
2301	*
2302	* ---------------------------------------------
2303	* \| valid log \| invalid log \|
2304	* ---------------------------------------------
2305	* ^
2306	* \|- log->last_checkpoint
2307	* \|- log->last_cp_seq
2308	*
2309	* Now we scan through the log until we see invalid entry
2310	*
2311	* ---------------------------------------------
2312	* \| valid log \| invalid log \|
2313	* ---------------------------------------------
2314	* ^ ^
2315	* \|- log->last_checkpoint \|- ctx->pos
2316	* \|- log->last_cp_seq \|- ctx->seq
2317	*
2318	* From this point, we need to increase seq number by 10 to avoid
2319	* confusing next recovery.
2320	*
2321	* ---------------------------------------------
2322	* \| valid log \| invalid log \|
2323	* ---------------------------------------------
2324	* ^ ^
2325	* \|- log->last_checkpoint \|- ctx->pos+1
2326	* \|- log->last_cp_seq \|- ctx->seq+10001
2327	*
2328	* However, it is not safe to start the state machine yet, because data only
2329	* parities are not yet secured in RAID. To save these data only parities, we
2330	* rewrite them from seq+11.
2331	*
2332	* -----------------------------------------------------------------
2333	* \| valid log \| data only stripes \| invalid log \|
2334	* -----------------------------------------------------------------
2335	* ^ ^
2336	* \|- log->last_checkpoint \|- ctx->pos+n
2337	* \|- log->last_cp_seq \|- ctx->seq+10000+n
2338	*
2339	* If failure happens again during this process, the recovery can safe start
2340	* again from log->last_checkpoint.
2341	*
2342	* Once data only stripes are rewritten to journal, we move log_tail
2343	*
2344	* -----------------------------------------------------------------
2345	* \| old log \| data only stripes \| invalid log \|
2346	* -----------------------------------------------------------------
2347	* ^ ^
2348	* \|- log->last_checkpoint \|- ctx->pos+n
2349	* \|- log->last_cp_seq \|- ctx->seq+10000+n
2350	*
2351	* Then we can safely start the state machine. If failure happens from this
2352	* point on, the recovery will start from new log->last_checkpoint.
2353	*/
2354	static int
2355	r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2356	struct r5l_recovery_ctx *ctx)
2357	{
2358	struct stripe_head *sh;
2359	struct mddev *mddev = log->rdev->mddev;
2360	struct page *page;
2361	sector_t next_checkpoint = MaxSector;
2362
2363	page = alloc_page(GFP_KERNEL);
2364	if (!page) {
2365	pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2366	mdname(mddev));
2367	return -ENOMEM;
2368	}
2369
2370	WARN_ON(list_empty(&ctx->cached_list));
2371
2372	list_for_each_entry(sh, &ctx->cached_list, lru) {
2373	struct r5l_meta_block *mb;
2374	int i;
2375	int offset;
2376	sector_t write_pos;
2377
2378	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2379	r5l_recovery_create_empty_meta_block(log, page,
2380	ctx->pos, ctx->seq);
2381	mb = page_address(page);
2382	offset = le32_to_cpu(mb->meta_size);
2383	write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2384
2385	for (i = sh->disks; i--; ) {
2386	struct r5dev *dev = &sh->dev[i];
2387	struct r5l_payload_data_parity *payload;
2388	void *addr;
2389
2390	if (test_bit(R5_InJournal, &dev->flags)) {
2391	payload = (void *)mb + offset;
2392	payload->header.type = cpu_to_le16(
2393	R5LOG_PAYLOAD_DATA);
2394	payload->size = cpu_to_le32(BLOCK_SECTORS);
2395	payload->location = cpu_to_le64(
2396	raid5_compute_blocknr(sh, i, 0));
2397	addr = kmap_local_page(dev->page);
2398	payload->checksum[0] = cpu_to_le32(
2399	crc32c(log->uuid_checksum, addr,
2400	PAGE_SIZE));
2401	kunmap_local(addr);
2402	sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2403	dev->page, REQ_OP_WRITE, false);
2404	write_pos = r5l_ring_add(log, write_pos,
2405	BLOCK_SECTORS);
2406	offset += sizeof(__le32) +
2407	sizeof(struct r5l_payload_data_parity);
2408
2409	}
2410	}
2411	mb->meta_size = cpu_to_le32(offset);
2412	mb->checksum = cpu_to_le32(crc32c(log->uuid_checksum,
2413	mb, PAGE_SIZE));
2414	sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2415	REQ_OP_WRITE \| REQ_SYNC \| REQ_FUA, false);
2416	sh->log_start = ctx->pos;
2417	list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2418	atomic_inc(&log->stripe_in_journal_count);
2419	ctx->pos = write_pos;
2420	ctx->seq += 1;
2421	next_checkpoint = sh->log_start;
2422	}
2423	log->next_checkpoint = next_checkpoint;
2424	__free_page(page);
2425	return 0;
2426	}
2427
2428	static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2429	struct r5l_recovery_ctx *ctx)
2430	{
2431	struct mddev *mddev = log->rdev->mddev;
2432	struct r5conf *conf = mddev->private;
2433	struct stripe_head sh, next;
2434	bool cleared_pending = false;
2435
2436	if (ctx->data_only_stripes == 0)
2437	return;
2438
2439	if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2440	cleared_pending = true;
2441	clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2442	}
2443	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2444
2445	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2446	r5c_make_stripe_write_out(sh);
2447	set_bit(STRIPE_HANDLE, &sh->state);
2448	list_del_init(&sh->lru);
2449	raid5_release_stripe(sh);
2450	}
2451
2452	/* reuse conf->wait_for_quiescent in recovery */
2453	wait_event(conf->wait_for_quiescent,
2454	atomic_read(&conf->active_stripes) == 0);
2455
2456	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2457	if (cleared_pending)
2458	set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2459	}
2460
2461	static int r5l_recovery_log(struct r5l_log *log)
2462	{
2463	struct mddev *mddev = log->rdev->mddev;
2464	struct r5l_recovery_ctx *ctx;
2465	int ret;
2466	sector_t pos;
2467
2468	ctx = kzalloc_obj(*ctx);
2469	if (!ctx)
2470	return -ENOMEM;
2471
2472	ctx->pos = log->last_checkpoint;
2473	ctx->seq = log->last_cp_seq;
2474	INIT_LIST_HEAD(&ctx->cached_list);
2475	ctx->meta_page = alloc_page(GFP_KERNEL);
2476
2477	if (!ctx->meta_page) {
2478	ret = -ENOMEM;
2479	goto meta_page;
2480	}
2481
2482	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2483	ret = -ENOMEM;
2484	goto ra_pool;
2485	}
2486
2487	ret = r5c_recovery_flush_log(log, ctx);
2488
2489	if (ret)
2490	goto error;
2491
2492	pos = ctx->pos;
2493	ctx->seq += 10000;
2494
2495	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2496	pr_info("md/raid:%s: starting from clean shutdown\n",
2497	mdname(mddev));
2498	else
2499	pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2500	mdname(mddev), ctx->data_only_stripes,
2501	ctx->data_parity_stripes);
2502
2503	if (ctx->data_only_stripes == 0) {
2504	log->next_checkpoint = ctx->pos;
2505	r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2506	ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2507	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2508	pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2509	mdname(mddev));
2510	ret = -EIO;
2511	goto error;
2512	}
2513
2514	log->log_start = ctx->pos;
2515	log->seq = ctx->seq;
2516	log->last_checkpoint = pos;
2517	r5l_write_super(log, pos);
2518
2519	r5c_recovery_flush_data_only_stripes(log, ctx);
2520	ret = 0;
2521	error:
2522	r5l_recovery_free_ra_pool(log, ctx);
2523	ra_pool:
2524	__free_page(ctx->meta_page);
2525	meta_page:
2526	kfree(ctx);
2527	return ret;
2528	}
2529
2530	static void r5l_write_super(struct r5l_log *log, sector_t cp)
2531	{
2532	struct mddev *mddev = log->rdev->mddev;
2533
2534	log->rdev->journal_tail = cp;
2535	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2536	}
2537
2538	static ssize_t r5c_journal_mode_show(struct mddev mddev, char page)
2539	{
2540	struct r5conf *conf;
2541	int ret;
2542
2543	ret = mddev_lock(mddev);
2544	if (ret)
2545	return ret;
2546
2547	conf = mddev->private;
2548	if (!conf \|\| !conf->log)
2549	goto out_unlock;
2550
2551	switch (conf->log->r5c_journal_mode) {
2552	case R5C_JOURNAL_MODE_WRITE_THROUGH:
2553	ret = snprintf(
2554	page, PAGE_SIZE, "[%s] %s\n",
2555	r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2556	r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2557	break;
2558	case R5C_JOURNAL_MODE_WRITE_BACK:
2559	ret = snprintf(
2560	page, PAGE_SIZE, "%s [%s]\n",
2561	r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2562	r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2563	break;
2564	default:
2565	ret = 0;
2566	}
2567
2568	out_unlock:
2569	mddev_unlock(mddev);
2570	return ret;
2571	}
2572
2573	/*
2574	* Set journal cache mode on @mddev (external API initially needed by dm-raid).
2575	*
2576	* @mode as defined in 'enum r5c_journal_mode'.
2577	*
2578	*/
2579	int r5c_journal_mode_set(struct mddev *mddev, int mode)
2580	{
2581	struct r5conf *conf;
2582
2583	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH \|\|
2584	mode > R5C_JOURNAL_MODE_WRITE_BACK)
2585	return -EINVAL;
2586
2587	conf = mddev->private;
2588	if (!conf \|\| !conf->log)
2589	return -ENODEV;
2590
2591	if (raid5_calc_degraded(conf) > 0 &&
2592	mode == R5C_JOURNAL_MODE_WRITE_BACK)
2593	return -EINVAL;
2594
2595	conf->log->r5c_journal_mode = mode;
2596
2597	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2598	mdname(mddev), mode, r5c_journal_mode_str[mode]);
2599	return 0;
2600	}
2601	EXPORT_SYMBOL(r5c_journal_mode_set);
2602
2603	static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2604	const char *page, size_t length)
2605	{
2606	int mode = ARRAY_SIZE(r5c_journal_mode_str);
2607	size_t len = length;
2608	int ret;
2609
2610	if (len < 2)
2611	return -EINVAL;
2612
2613	if (page[len - 1] == '\n')
2614	len--;
2615
2616	while (mode--)
2617	if (strlen(r5c_journal_mode_str[mode]) == len &&
2618	!strncmp(page, r5c_journal_mode_str[mode], len))
2619	break;
2620	ret = mddev_suspend_and_lock(mddev);
2621	if (ret)
2622	return ret;
2623	ret = r5c_journal_mode_set(mddev, mode);
2624	mddev_unlock_and_resume(mddev);
2625	return ret ?: length;
2626	}
2627
2628	struct md_sysfs_entry
2629	r5c_journal_mode = __ATTR(journal_mode, 0644,
2630	r5c_journal_mode_show, r5c_journal_mode_store);
2631
2632	/*
2633	* Try handle write operation in caching phase. This function should only
2634	* be called in write-back mode.
2635	*
2636	* If all outstanding writes can be handled in caching phase, returns 0
2637	* If writes requires write-out phase, call r5c_make_stripe_write_out()
2638	* and returns -EAGAIN
2639	*/
2640	int r5c_try_caching_write(struct r5conf *conf,
2641	struct stripe_head *sh,
2642	struct stripe_head_state *s,
2643	int disks)
2644	{
2645	struct r5l_log *log = READ_ONCE(conf->log);
2646	int i;
2647	struct r5dev *dev;
2648	int to_cache = 0;
2649	void __rcu **pslot;
2650	sector_t tree_index;
2651	int ret;
2652	uintptr_t refcount;
2653
2654	BUG_ON(!r5c_is_writeback(log));
2655
2656	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2657	/*
2658	* There are two different scenarios here:
2659	* 1. The stripe has some data cached, and it is sent to
2660	* write-out phase for reclaim
2661	* 2. The stripe is clean, and this is the first write
2662	*
2663	* For 1, return -EAGAIN, so we continue with
2664	* handle_stripe_dirtying().
2665	*
2666	* For 2, set STRIPE_R5C_CACHING and continue with caching
2667	* write.
2668	*/
2669
2670	/* case 1: anything injournal or anything in written */
2671	if (s->injournal > 0 \|\| s->written > 0)
2672	return -EAGAIN;
2673	/* case 2 */
2674	set_bit(STRIPE_R5C_CACHING, &sh->state);
2675	}
2676
2677	/*
2678	* When run in degraded mode, array is set to write-through mode.
2679	* This check helps drain pending write safely in the transition to
2680	* write-through mode.
2681	*
2682	* When a stripe is syncing, the write is also handled in write
2683	* through mode.
2684	*/
2685	if (s->failed \|\| test_bit(STRIPE_SYNCING, &sh->state)) {
2686	r5c_make_stripe_write_out(sh);
2687	return -EAGAIN;
2688	}
2689
2690	for (i = disks; i--; ) {
2691	dev = &sh->dev[i];
2692	/* if non-overwrite, use writing-out phase */
2693	if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2694	!test_bit(R5_InJournal, &dev->flags)) {
2695	r5c_make_stripe_write_out(sh);
2696	return -EAGAIN;
2697	}
2698	}
2699
2700	/* if the stripe is not counted in big_stripe_tree, add it now */
2701	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2702	!test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2703	tree_index = r5c_tree_index(conf, sh->sector);
2704	spin_lock(&log->tree_lock);
2705	pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2706	tree_index);
2707	if (pslot) {
2708	refcount = (uintptr_t)radix_tree_deref_slot_protected(
2709	pslot, &log->tree_lock) >>
2710	R5C_RADIX_COUNT_SHIFT;
2711	radix_tree_replace_slot(
2712	&log->big_stripe_tree, pslot,
2713	(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2714	} else {
2715	/*
2716	* this radix_tree_insert can fail safely, so no
2717	* need to call radix_tree_preload()
2718	*/
2719	ret = radix_tree_insert(
2720	&log->big_stripe_tree, tree_index,
2721	(void *)(1 << R5C_RADIX_COUNT_SHIFT));
2722	if (ret) {
2723	spin_unlock(&log->tree_lock);
2724	r5c_make_stripe_write_out(sh);
2725	return -EAGAIN;
2726	}
2727	}
2728	spin_unlock(&log->tree_lock);
2729
2730	/*
2731	* set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2732	* counted in the radix tree
2733	*/
2734	set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2735	atomic_inc(&conf->r5c_cached_partial_stripes);
2736	}
2737
2738	for (i = disks; i--; ) {
2739	dev = &sh->dev[i];
2740	if (dev->towrite) {
2741	set_bit(R5_Wantwrite, &dev->flags);
2742	set_bit(R5_Wantdrain, &dev->flags);
2743	set_bit(R5_LOCKED, &dev->flags);
2744	to_cache++;
2745	}
2746	}
2747
2748	if (to_cache) {
2749	set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2750	/*
2751	* set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2752	* in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2753	* r5c_handle_data_cached()
2754	*/
2755	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2756	}
2757
2758	return 0;
2759	}
2760
2761	/*
2762	* free extra pages (orig_page) we allocated for prexor
2763	*/
2764	void r5c_release_extra_page(struct stripe_head *sh)
2765	{
2766	struct r5conf *conf = sh->raid_conf;
2767	int i;
2768	bool using_disk_info_extra_page;
2769
2770	using_disk_info_extra_page =
2771	sh->dev[0].orig_page == conf->disks[0].extra_page;
2772
2773	for (i = sh->disks; i--; )
2774	if (sh->dev[i].page != sh->dev[i].orig_page) {
2775	struct page *p = sh->dev[i].orig_page;
2776
2777	sh->dev[i].orig_page = sh->dev[i].page;
2778	clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2779
2780	if (!using_disk_info_extra_page)
2781	put_page(p);
2782	}
2783
2784	if (using_disk_info_extra_page) {
2785	clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2786	md_wakeup_thread(conf->mddev->thread);
2787	}
2788	}
2789
2790	void r5c_use_extra_page(struct stripe_head *sh)
2791	{
2792	struct r5conf *conf = sh->raid_conf;
2793	int i;
2794	struct r5dev *dev;
2795
2796	for (i = sh->disks; i--; ) {
2797	dev = &sh->dev[i];
2798	if (dev->orig_page != dev->page)
2799	put_page(dev->orig_page);
2800	dev->orig_page = conf->disks[i].extra_page;
2801	}
2802	}
2803
2804	/*
2805	* clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2806	* stripe is committed to RAID disks.
2807	*/
2808	void r5c_finish_stripe_write_out(struct r5conf *conf,
2809	struct stripe_head *sh,
2810	struct stripe_head_state *s)
2811	{
2812	struct r5l_log *log = READ_ONCE(conf->log);
2813	int i;
2814	sector_t tree_index;
2815	void __rcu **pslot;
2816	uintptr_t refcount;
2817
2818	if (!log \|\| !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2819	return;
2820
2821	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2822	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2823
2824	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2825	return;
2826
2827	for (i = sh->disks; i--; ) {
2828	clear_bit(R5_InJournal, &sh->dev[i].flags);
2829	if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2830	wake_up_bit(&sh->dev[i].flags, R5_Overlap);
2831	}
2832
2833	/*
2834	* analyse_stripe() runs before r5c_finish_stripe_write_out(),
2835	* We updated R5_InJournal, so we also update s->injournal.
2836	*/
2837	s->injournal = 0;
2838
2839	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2840	if (atomic_dec_and_test(&conf->pending_full_writes))
2841	md_wakeup_thread(conf->mddev->thread);
2842
2843	spin_lock_irq(&log->stripe_in_journal_lock);
2844	list_del_init(&sh->r5c);
2845	spin_unlock_irq(&log->stripe_in_journal_lock);
2846	sh->log_start = MaxSector;
2847
2848	atomic_dec(&log->stripe_in_journal_count);
2849	r5c_update_log_state(log);
2850
2851	/* stop counting this stripe in big_stripe_tree */
2852	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) \|\|
2853	test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2854	tree_index = r5c_tree_index(conf, sh->sector);
2855	spin_lock(&log->tree_lock);
2856	pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2857	tree_index);
2858	BUG_ON(pslot == NULL);
2859	refcount = (uintptr_t)radix_tree_deref_slot_protected(
2860	pslot, &log->tree_lock) >>
2861	R5C_RADIX_COUNT_SHIFT;
2862	if (refcount == 1)
2863	radix_tree_delete(&log->big_stripe_tree, tree_index);
2864	else
2865	radix_tree_replace_slot(
2866	&log->big_stripe_tree, pslot,
2867	(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2868	spin_unlock(&log->tree_lock);
2869	}
2870
2871	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2872	BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2873	atomic_dec(&conf->r5c_flushing_partial_stripes);
2874	atomic_dec(&conf->r5c_cached_partial_stripes);
2875	}
2876
2877	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2878	BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2879	atomic_dec(&conf->r5c_flushing_full_stripes);
2880	atomic_dec(&conf->r5c_cached_full_stripes);
2881	}
2882
2883	r5l_append_flush_payload(log, sh->sector);
2884	/* stripe is flused to raid disks, we can do resync now */
2885	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2886	set_bit(STRIPE_HANDLE, &sh->state);
2887	}
2888
2889	int r5c_cache_data(struct r5l_log log, struct stripe_head sh)
2890	{
2891	struct r5conf *conf = sh->raid_conf;
2892	int pages = 0;
2893	int reserve;
2894	int i;
2895	int ret = 0;
2896
2897	BUG_ON(!log);
2898
2899	for (i = 0; i < sh->disks; i++) {
2900	void *addr;
2901
2902	if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2903	continue;
2904	addr = kmap_local_page(sh->dev[i].page);
2905	sh->dev[i].log_checksum = crc32c(log->uuid_checksum,
2906	addr, PAGE_SIZE);
2907	kunmap_local(addr);
2908	pages++;
2909	}
2910	WARN_ON(pages == 0);
2911
2912	/*
2913	* The stripe must enter state machine again to call endio, so
2914	* don't delay.
2915	*/
2916	clear_bit(STRIPE_DELAYED, &sh->state);
2917	atomic_inc(&sh->count);
2918
2919	mutex_lock(&log->io_mutex);
2920	/* meta + data */
2921	reserve = (1 + pages) << (PAGE_SHIFT - 9);
2922
2923	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2924	sh->log_start == MaxSector)
2925	r5l_add_no_space_stripe(log, sh);
2926	else if (!r5l_has_free_space(log, reserve)) {
2927	if (sh->log_start == log->last_checkpoint)
2928	BUG();
2929	else
2930	r5l_add_no_space_stripe(log, sh);
2931	} else {
2932	ret = r5l_log_stripe(log, sh, pages, 0);
2933	if (ret) {
2934	spin_lock_irq(&log->io_list_lock);
2935	list_add_tail(&sh->log_list, &log->no_mem_stripes);
2936	spin_unlock_irq(&log->io_list_lock);
2937	}
2938	}
2939
2940	mutex_unlock(&log->io_mutex);
2941	return 0;
2942	}
2943
2944	/* check whether this big stripe is in write back cache. */
2945	bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2946	{
2947	struct r5l_log *log = READ_ONCE(conf->log);
2948	sector_t tree_index;
2949	void *slot;
2950
2951	if (!log)
2952	return false;
2953
2954	tree_index = r5c_tree_index(conf, sect);
2955	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2956	return slot != NULL;
2957	}
2958
2959	static int r5l_load_log(struct r5l_log *log)
2960	{
2961	struct md_rdev *rdev = log->rdev;
2962	struct page *page;
2963	struct r5l_meta_block *mb;
2964	sector_t cp = log->rdev->journal_tail;
2965	u32 stored_crc, expected_crc;
2966	bool create_super = false;
2967	int ret = 0;
2968
2969	/* Make sure it's valid */
2970	if (cp >= rdev->sectors \|\| round_down(cp, BLOCK_SECTORS) != cp)
2971	cp = 0;
2972	page = alloc_page(GFP_KERNEL);
2973	if (!page)
2974	return -ENOMEM;
2975
2976	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, false)) {
2977	ret = -EIO;
2978	goto ioerr;
2979	}
2980	mb = page_address(page);
2981
2982	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC \|\|
2983	mb->version != R5LOG_VERSION) {
2984	create_super = true;
2985	goto create;
2986	}
2987	stored_crc = le32_to_cpu(mb->checksum);
2988	mb->checksum = 0;
2989	expected_crc = crc32c(log->uuid_checksum, mb, PAGE_SIZE);
2990	if (stored_crc != expected_crc) {
2991	create_super = true;
2992	goto create;
2993	}
2994	if (le64_to_cpu(mb->position) != cp) {
2995	create_super = true;
2996	goto create;
2997	}
2998	create:
2999	if (create_super) {
3000	log->last_cp_seq = get_random_u32();
3001	cp = 0;
3002	r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3003	/*
3004	* Make sure super points to correct address. Log might have
3005	* data very soon. If super hasn't correct log tail address,
3006	* recovery can't find the log
3007	*/
3008	r5l_write_super(log, cp);
3009	} else
3010	log->last_cp_seq = le64_to_cpu(mb->seq);
3011
3012	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3013	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3014	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3015	log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3016	log->last_checkpoint = cp;
3017
3018	__free_page(page);
3019
3020	if (create_super) {
3021	log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3022	log->seq = log->last_cp_seq + 1;
3023	log->next_checkpoint = cp;
3024	} else
3025	ret = r5l_recovery_log(log);
3026
3027	r5c_update_log_state(log);
3028	return ret;
3029	ioerr:
3030	__free_page(page);
3031	return ret;
3032	}
3033
3034	int r5l_start(struct r5l_log *log)
3035	{
3036	int ret;
3037
3038	if (!log)
3039	return 0;
3040
3041	ret = r5l_load_log(log);
3042	if (ret) {
3043	struct mddev *mddev = log->rdev->mddev;
3044	struct r5conf *conf = mddev->private;
3045
3046	r5l_exit_log(conf);
3047	}
3048	return ret;
3049	}
3050
3051	void r5c_update_on_rdev_error(struct mddev mddev, struct md_rdev rdev)
3052	{
3053	struct r5conf *conf = mddev->private;
3054	struct r5l_log *log = READ_ONCE(conf->log);
3055
3056	if (!log)
3057	return;
3058
3059	if ((raid5_calc_degraded(conf) > 0 \|\|
3060	test_bit(Journal, &rdev->flags)) &&
3061	log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3062	schedule_work(&log->disable_writeback_work);
3063	}
3064
3065	int r5l_init_log(struct r5conf conf, struct md_rdev rdev)
3066	{
3067	struct r5l_log *log;
3068	struct md_thread *thread;
3069	int ret;
3070
3071	pr_debug("md/raid:%s: using device %pg as journal\n",
3072	mdname(conf->mddev), rdev->bdev);
3073
3074	if (PAGE_SIZE != 4096)
3075	return -EINVAL;
3076
3077	/*
3078	* The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3079	* raid_disks r5l_payload_data_parity.
3080	*
3081	* Write journal and cache does not work for very big array
3082	* (raid_disks > 203)
3083	*/
3084	if (sizeof(struct r5l_meta_block) +
3085	((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3086	conf->raid_disks) > PAGE_SIZE) {
3087	pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3088	mdname(conf->mddev), conf->raid_disks);
3089	return -EINVAL;
3090	}
3091
3092	log = kzalloc_obj(*log);
3093	if (!log)
3094	return -ENOMEM;
3095	log->rdev = rdev;
3096	log->need_cache_flush = bdev_write_cache(rdev->bdev);
3097	log->uuid_checksum = crc32c(~0, rdev->mddev->uuid,
3098	sizeof(rdev->mddev->uuid));
3099
3100	mutex_init(&log->io_mutex);
3101
3102	spin_lock_init(&log->io_list_lock);
3103	INIT_LIST_HEAD(&log->running_ios);
3104	INIT_LIST_HEAD(&log->io_end_ios);
3105	INIT_LIST_HEAD(&log->flushing_ios);
3106	INIT_LIST_HEAD(&log->finished_ios);
3107
3108	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3109	if (!log->io_kc)
3110	goto io_kc;
3111
3112	ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3113	if (ret)
3114	goto io_pool;
3115
3116	ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3117	if (ret)
3118	goto io_bs;
3119
3120	ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3121	if (ret)
3122	goto out_mempool;
3123
3124	spin_lock_init(&log->tree_lock);
3125	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT);
3126
3127	thread = md_register_thread(r5l_reclaim_thread, log->rdev->mddev,
3128	"reclaim");
3129	if (!thread)
3130	goto reclaim_thread;
3131
3132	thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3133	rcu_assign_pointer(log->reclaim_thread, thread);
3134
3135	init_waitqueue_head(&log->iounit_wait);
3136
3137	INIT_LIST_HEAD(&log->no_mem_stripes);
3138
3139	INIT_LIST_HEAD(&log->no_space_stripes);
3140	spin_lock_init(&log->no_space_stripes_lock);
3141
3142	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3143	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3144
3145	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3146	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3147	spin_lock_init(&log->stripe_in_journal_lock);
3148	atomic_set(&log->stripe_in_journal_count, 0);
3149
3150	WRITE_ONCE(conf->log, log);
3151
3152	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3153	return 0;
3154
3155	reclaim_thread:
3156	mempool_exit(&log->meta_pool);
3157	out_mempool:
3158	bioset_exit(&log->bs);
3159	io_bs:
3160	mempool_exit(&log->io_pool);
3161	io_pool:
3162	kmem_cache_destroy(log->io_kc);
3163	io_kc:
3164	kfree(log);
3165	return -EINVAL;
3166	}
3167
3168	void r5l_exit_log(struct r5conf *conf)
3169	{
3170	struct r5l_log *log = conf->log;
3171
3172	md_unregister_thread(conf->mddev, &log->reclaim_thread);
3173
3174	/*
3175	* 'reconfig_mutex' is held by caller, set 'confg->log' to NULL to
3176	* ensure disable_writeback_work wakes up and exits.
3177	*/
3178	WRITE_ONCE(conf->log, NULL);
3179	wake_up(&conf->mddev->sb_wait);
3180	flush_work(&log->disable_writeback_work);
3181
3182	mempool_exit(&log->meta_pool);
3183	bioset_exit(&log->bs);
3184	mempool_exit(&log->io_pool);
3185	kmem_cache_destroy(log->io_kc);
3186	kfree(log);
3187	}