xref: /linux/drivers/md/raid5-cache.c (revision ec63e2a4897075e427c121d863bd89c44578094f)
1 /*
2  * Copyright (C) 2015 Shaohua Li <shli@fb.com>
3  * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  */
15 #include <linux/kernel.h>
16 #include <linux/wait.h>
17 #include <linux/blkdev.h>
18 #include <linux/slab.h>
19 #include <linux/raid/md_p.h>
20 #include <linux/crc32c.h>
21 #include <linux/random.h>
22 #include <linux/kthread.h>
23 #include <linux/types.h>
24 #include "md.h"
25 #include "raid5.h"
26 #include "md-bitmap.h"
27 #include "raid5-log.h"
28 
29 /*
30  * metadata/data stored in disk with 4k size unit (a block) regardless
31  * underneath hardware sector size. only works with PAGE_SIZE == 4096
32  */
33 #define BLOCK_SECTORS (8)
34 #define BLOCK_SECTOR_SHIFT (3)
35 
36 /*
37  * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
38  *
39  * In write through mode, the reclaim runs every log->max_free_space.
40  * This can prevent the recovery scans for too long
41  */
42 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
43 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
44 
45 /* wake up reclaim thread periodically */
46 #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
47 /* start flush with these full stripes */
48 #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
49 /* reclaim stripes in groups */
50 #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
51 
52 /*
53  * We only need 2 bios per I/O unit to make progress, but ensure we
54  * have a few more available to not get too tight.
55  */
56 #define R5L_POOL_SIZE	4
57 
58 static char *r5c_journal_mode_str[] = {"write-through",
59 				       "write-back"};
60 /*
61  * raid5 cache state machine
62  *
63  * With the RAID cache, each stripe works in two phases:
64  *	- caching phase
65  *	- writing-out phase
66  *
67  * These two phases are controlled by bit STRIPE_R5C_CACHING:
68  *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
69  *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
70  *
71  * When there is no journal, or the journal is in write-through mode,
72  * the stripe is always in writing-out phase.
73  *
74  * For write-back journal, the stripe is sent to caching phase on write
75  * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
76  * the write-out phase by clearing STRIPE_R5C_CACHING.
77  *
78  * Stripes in caching phase do not write the raid disks. Instead, all
79  * writes are committed from the log device. Therefore, a stripe in
80  * caching phase handles writes as:
81  *	- write to log device
82  *	- return IO
83  *
84  * Stripes in writing-out phase handle writes as:
85  *	- calculate parity
86  *	- write pending data and parity to journal
87  *	- write data and parity to raid disks
88  *	- return IO for pending writes
89  */
90 
91 struct r5l_log {
92 	struct md_rdev *rdev;
93 
94 	u32 uuid_checksum;
95 
96 	sector_t device_size;		/* log device size, round to
97 					 * BLOCK_SECTORS */
98 	sector_t max_free_space;	/* reclaim run if free space is at
99 					 * this size */
100 
101 	sector_t last_checkpoint;	/* log tail. where recovery scan
102 					 * starts from */
103 	u64 last_cp_seq;		/* log tail sequence */
104 
105 	sector_t log_start;		/* log head. where new data appends */
106 	u64 seq;			/* log head sequence */
107 
108 	sector_t next_checkpoint;
109 
110 	struct mutex io_mutex;
111 	struct r5l_io_unit *current_io;	/* current io_unit accepting new data */
112 
113 	spinlock_t io_list_lock;
114 	struct list_head running_ios;	/* io_units which are still running,
115 					 * and have not yet been completely
116 					 * written to the log */
117 	struct list_head io_end_ios;	/* io_units which have been completely
118 					 * written to the log but not yet written
119 					 * to the RAID */
120 	struct list_head flushing_ios;	/* io_units which are waiting for log
121 					 * cache flush */
122 	struct list_head finished_ios;	/* io_units which settle down in log disk */
123 	struct bio flush_bio;
124 
125 	struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
126 
127 	struct kmem_cache *io_kc;
128 	mempool_t io_pool;
129 	struct bio_set bs;
130 	mempool_t meta_pool;
131 
132 	struct md_thread *reclaim_thread;
133 	unsigned long reclaim_target;	/* number of space that need to be
134 					 * reclaimed.  if it's 0, reclaim spaces
135 					 * used by io_units which are in
136 					 * IO_UNIT_STRIPE_END state (eg, reclaim
137 					 * dones't wait for specific io_unit
138 					 * switching to IO_UNIT_STRIPE_END
139 					 * state) */
140 	wait_queue_head_t iounit_wait;
141 
142 	struct list_head no_space_stripes; /* pending stripes, log has no space */
143 	spinlock_t no_space_stripes_lock;
144 
145 	bool need_cache_flush;
146 
147 	/* for r5c_cache */
148 	enum r5c_journal_mode r5c_journal_mode;
149 
150 	/* all stripes in r5cache, in the order of seq at sh->log_start */
151 	struct list_head stripe_in_journal_list;
152 
153 	spinlock_t stripe_in_journal_lock;
154 	atomic_t stripe_in_journal_count;
155 
156 	/* to submit async io_units, to fulfill ordering of flush */
157 	struct work_struct deferred_io_work;
158 	/* to disable write back during in degraded mode */
159 	struct work_struct disable_writeback_work;
160 
161 	/* to for chunk_aligned_read in writeback mode, details below */
162 	spinlock_t tree_lock;
163 	struct radix_tree_root big_stripe_tree;
164 };
165 
166 /*
167  * Enable chunk_aligned_read() with write back cache.
168  *
169  * Each chunk may contain more than one stripe (for example, a 256kB
170  * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
171  * chunk_aligned_read, these stripes are grouped into one "big_stripe".
172  * For each big_stripe, we count how many stripes of this big_stripe
173  * are in the write back cache. These data are tracked in a radix tree
174  * (big_stripe_tree). We use radix_tree item pointer as the counter.
175  * r5c_tree_index() is used to calculate keys for the radix tree.
176  *
177  * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
178  * big_stripe of each chunk in the tree. If this big_stripe is in the
179  * tree, chunk_aligned_read() aborts. This look up is protected by
180  * rcu_read_lock().
181  *
182  * It is necessary to remember whether a stripe is counted in
183  * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
184  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
185  * two flags are set, the stripe is counted in big_stripe_tree. This
186  * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
187  * r5c_try_caching_write(); and moving clear_bit of
188  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
189  * r5c_finish_stripe_write_out().
190  */
191 
192 /*
193  * radix tree requests lowest 2 bits of data pointer to be 2b'00.
194  * So it is necessary to left shift the counter by 2 bits before using it
195  * as data pointer of the tree.
196  */
197 #define R5C_RADIX_COUNT_SHIFT 2
198 
199 /*
200  * calculate key for big_stripe_tree
201  *
202  * sect: align_bi->bi_iter.bi_sector or sh->sector
203  */
204 static inline sector_t r5c_tree_index(struct r5conf *conf,
205 				      sector_t sect)
206 {
207 	sector_t offset;
208 
209 	offset = sector_div(sect, conf->chunk_sectors);
210 	return sect;
211 }
212 
213 /*
214  * an IO range starts from a meta data block and end at the next meta data
215  * block. The io unit's the meta data block tracks data/parity followed it. io
216  * unit is written to log disk with normal write, as we always flush log disk
217  * first and then start move data to raid disks, there is no requirement to
218  * write io unit with FLUSH/FUA
219  */
220 struct r5l_io_unit {
221 	struct r5l_log *log;
222 
223 	struct page *meta_page;	/* store meta block */
224 	int meta_offset;	/* current offset in meta_page */
225 
226 	struct bio *current_bio;/* current_bio accepting new data */
227 
228 	atomic_t pending_stripe;/* how many stripes not flushed to raid */
229 	u64 seq;		/* seq number of the metablock */
230 	sector_t log_start;	/* where the io_unit starts */
231 	sector_t log_end;	/* where the io_unit ends */
232 	struct list_head log_sibling; /* log->running_ios */
233 	struct list_head stripe_list; /* stripes added to the io_unit */
234 
235 	int state;
236 	bool need_split_bio;
237 	struct bio *split_bio;
238 
239 	unsigned int has_flush:1;		/* include flush request */
240 	unsigned int has_fua:1;			/* include fua request */
241 	unsigned int has_null_flush:1;		/* include null flush request */
242 	unsigned int has_flush_payload:1;	/* include flush payload  */
243 	/*
244 	 * io isn't sent yet, flush/fua request can only be submitted till it's
245 	 * the first IO in running_ios list
246 	 */
247 	unsigned int io_deferred:1;
248 
249 	struct bio_list flush_barriers;   /* size == 0 flush bios */
250 };
251 
252 /* r5l_io_unit state */
253 enum r5l_io_unit_state {
254 	IO_UNIT_RUNNING = 0,	/* accepting new IO */
255 	IO_UNIT_IO_START = 1,	/* io_unit bio start writing to log,
256 				 * don't accepting new bio */
257 	IO_UNIT_IO_END = 2,	/* io_unit bio finish writing to log */
258 	IO_UNIT_STRIPE_END = 3,	/* stripes data finished writing to raid */
259 };
260 
261 bool r5c_is_writeback(struct r5l_log *log)
262 {
263 	return (log != NULL &&
264 		log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
265 }
266 
267 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
268 {
269 	start += inc;
270 	if (start >= log->device_size)
271 		start = start - log->device_size;
272 	return start;
273 }
274 
275 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
276 				  sector_t end)
277 {
278 	if (end >= start)
279 		return end - start;
280 	else
281 		return end + log->device_size - start;
282 }
283 
284 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
285 {
286 	sector_t used_size;
287 
288 	used_size = r5l_ring_distance(log, log->last_checkpoint,
289 					log->log_start);
290 
291 	return log->device_size > used_size + size;
292 }
293 
294 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
295 				    enum r5l_io_unit_state state)
296 {
297 	if (WARN_ON(io->state >= state))
298 		return;
299 	io->state = state;
300 }
301 
302 static void
303 r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
304 {
305 	struct bio *wbi, *wbi2;
306 
307 	wbi = dev->written;
308 	dev->written = NULL;
309 	while (wbi && wbi->bi_iter.bi_sector <
310 	       dev->sector + STRIPE_SECTORS) {
311 		wbi2 = r5_next_bio(wbi, dev->sector);
312 		md_write_end(conf->mddev);
313 		bio_endio(wbi);
314 		wbi = wbi2;
315 	}
316 }
317 
318 void r5c_handle_cached_data_endio(struct r5conf *conf,
319 				  struct stripe_head *sh, int disks)
320 {
321 	int i;
322 
323 	for (i = sh->disks; i--; ) {
324 		if (sh->dev[i].written) {
325 			set_bit(R5_UPTODATE, &sh->dev[i].flags);
326 			r5c_return_dev_pending_writes(conf, &sh->dev[i]);
327 			md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
328 					   STRIPE_SECTORS,
329 					   !test_bit(STRIPE_DEGRADED, &sh->state),
330 					   0);
331 		}
332 	}
333 }
334 
335 void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
336 
337 /* Check whether we should flush some stripes to free up stripe cache */
338 void r5c_check_stripe_cache_usage(struct r5conf *conf)
339 {
340 	int total_cached;
341 
342 	if (!r5c_is_writeback(conf->log))
343 		return;
344 
345 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
346 		atomic_read(&conf->r5c_cached_full_stripes);
347 
348 	/*
349 	 * The following condition is true for either of the following:
350 	 *   - stripe cache pressure high:
351 	 *          total_cached > 3/4 min_nr_stripes ||
352 	 *          empty_inactive_list_nr > 0
353 	 *   - stripe cache pressure moderate:
354 	 *          total_cached > 1/2 min_nr_stripes
355 	 */
356 	if (total_cached > conf->min_nr_stripes * 1 / 2 ||
357 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
358 		r5l_wake_reclaim(conf->log, 0);
359 }
360 
361 /*
362  * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
363  * stripes in the cache
364  */
365 void r5c_check_cached_full_stripe(struct r5conf *conf)
366 {
367 	if (!r5c_is_writeback(conf->log))
368 		return;
369 
370 	/*
371 	 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
372 	 * or a full stripe (chunk size / 4k stripes).
373 	 */
374 	if (atomic_read(&conf->r5c_cached_full_stripes) >=
375 	    min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
376 		conf->chunk_sectors >> STRIPE_SHIFT))
377 		r5l_wake_reclaim(conf->log, 0);
378 }
379 
380 /*
381  * Total log space (in sectors) needed to flush all data in cache
382  *
383  * To avoid deadlock due to log space, it is necessary to reserve log
384  * space to flush critical stripes (stripes that occupying log space near
385  * last_checkpoint). This function helps check how much log space is
386  * required to flush all cached stripes.
387  *
388  * To reduce log space requirements, two mechanisms are used to give cache
389  * flush higher priorities:
390  *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
391  *       stripes ALREADY in journal can be flushed w/o pending writes;
392  *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
393  *       can be delayed (r5l_add_no_space_stripe).
394  *
395  * In cache flush, the stripe goes through 1 and then 2. For a stripe that
396  * already passed 1, flushing it requires at most (conf->max_degraded + 1)
397  * pages of journal space. For stripes that has not passed 1, flushing it
398  * requires (conf->raid_disks + 1) pages of journal space. There are at
399  * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
400  * required to flush all cached stripes (in pages) is:
401  *
402  *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
403  *     (group_cnt + 1) * (raid_disks + 1)
404  * or
405  *     (stripe_in_journal_count) * (max_degraded + 1) +
406  *     (group_cnt + 1) * (raid_disks - max_degraded)
407  */
408 static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
409 {
410 	struct r5l_log *log = conf->log;
411 
412 	if (!r5c_is_writeback(log))
413 		return 0;
414 
415 	return BLOCK_SECTORS *
416 		((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
417 		 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
418 }
419 
420 /*
421  * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
422  *
423  * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
424  * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
425  * device is less than 2x of reclaim_required_space.
426  */
427 static inline void r5c_update_log_state(struct r5l_log *log)
428 {
429 	struct r5conf *conf = log->rdev->mddev->private;
430 	sector_t free_space;
431 	sector_t reclaim_space;
432 	bool wake_reclaim = false;
433 
434 	if (!r5c_is_writeback(log))
435 		return;
436 
437 	free_space = r5l_ring_distance(log, log->log_start,
438 				       log->last_checkpoint);
439 	reclaim_space = r5c_log_required_to_flush_cache(conf);
440 	if (free_space < 2 * reclaim_space)
441 		set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
442 	else {
443 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
444 			wake_reclaim = true;
445 		clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
446 	}
447 	if (free_space < 3 * reclaim_space)
448 		set_bit(R5C_LOG_TIGHT, &conf->cache_state);
449 	else
450 		clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
451 
452 	if (wake_reclaim)
453 		r5l_wake_reclaim(log, 0);
454 }
455 
456 /*
457  * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
458  * This function should only be called in write-back mode.
459  */
460 void r5c_make_stripe_write_out(struct stripe_head *sh)
461 {
462 	struct r5conf *conf = sh->raid_conf;
463 	struct r5l_log *log = conf->log;
464 
465 	BUG_ON(!r5c_is_writeback(log));
466 
467 	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
468 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
469 
470 	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
471 		atomic_inc(&conf->preread_active_stripes);
472 }
473 
474 static void r5c_handle_data_cached(struct stripe_head *sh)
475 {
476 	int i;
477 
478 	for (i = sh->disks; i--; )
479 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
480 			set_bit(R5_InJournal, &sh->dev[i].flags);
481 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
482 		}
483 	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
484 }
485 
486 /*
487  * this journal write must contain full parity,
488  * it may also contain some data pages
489  */
490 static void r5c_handle_parity_cached(struct stripe_head *sh)
491 {
492 	int i;
493 
494 	for (i = sh->disks; i--; )
495 		if (test_bit(R5_InJournal, &sh->dev[i].flags))
496 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
497 }
498 
499 /*
500  * Setting proper flags after writing (or flushing) data and/or parity to the
501  * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
502  */
503 static void r5c_finish_cache_stripe(struct stripe_head *sh)
504 {
505 	struct r5l_log *log = sh->raid_conf->log;
506 
507 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
508 		BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
509 		/*
510 		 * Set R5_InJournal for parity dev[pd_idx]. This means
511 		 * all data AND parity in the journal. For RAID 6, it is
512 		 * NOT necessary to set the flag for dev[qd_idx], as the
513 		 * two parities are written out together.
514 		 */
515 		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
516 	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
517 		r5c_handle_data_cached(sh);
518 	} else {
519 		r5c_handle_parity_cached(sh);
520 		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
521 	}
522 }
523 
524 static void r5l_io_run_stripes(struct r5l_io_unit *io)
525 {
526 	struct stripe_head *sh, *next;
527 
528 	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
529 		list_del_init(&sh->log_list);
530 
531 		r5c_finish_cache_stripe(sh);
532 
533 		set_bit(STRIPE_HANDLE, &sh->state);
534 		raid5_release_stripe(sh);
535 	}
536 }
537 
538 static void r5l_log_run_stripes(struct r5l_log *log)
539 {
540 	struct r5l_io_unit *io, *next;
541 
542 	lockdep_assert_held(&log->io_list_lock);
543 
544 	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
545 		/* don't change list order */
546 		if (io->state < IO_UNIT_IO_END)
547 			break;
548 
549 		list_move_tail(&io->log_sibling, &log->finished_ios);
550 		r5l_io_run_stripes(io);
551 	}
552 }
553 
554 static void r5l_move_to_end_ios(struct r5l_log *log)
555 {
556 	struct r5l_io_unit *io, *next;
557 
558 	lockdep_assert_held(&log->io_list_lock);
559 
560 	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
561 		/* don't change list order */
562 		if (io->state < IO_UNIT_IO_END)
563 			break;
564 		list_move_tail(&io->log_sibling, &log->io_end_ios);
565 	}
566 }
567 
568 static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
569 static void r5l_log_endio(struct bio *bio)
570 {
571 	struct r5l_io_unit *io = bio->bi_private;
572 	struct r5l_io_unit *io_deferred;
573 	struct r5l_log *log = io->log;
574 	unsigned long flags;
575 	bool has_null_flush;
576 	bool has_flush_payload;
577 
578 	if (bio->bi_status)
579 		md_error(log->rdev->mddev, log->rdev);
580 
581 	bio_put(bio);
582 	mempool_free(io->meta_page, &log->meta_pool);
583 
584 	spin_lock_irqsave(&log->io_list_lock, flags);
585 	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
586 
587 	/*
588 	 * if the io doesn't not have null_flush or flush payload,
589 	 * it is not safe to access it after releasing io_list_lock.
590 	 * Therefore, it is necessary to check the condition with
591 	 * the lock held.
592 	 */
593 	has_null_flush = io->has_null_flush;
594 	has_flush_payload = io->has_flush_payload;
595 
596 	if (log->need_cache_flush && !list_empty(&io->stripe_list))
597 		r5l_move_to_end_ios(log);
598 	else
599 		r5l_log_run_stripes(log);
600 	if (!list_empty(&log->running_ios)) {
601 		/*
602 		 * FLUSH/FUA io_unit is deferred because of ordering, now we
603 		 * can dispatch it
604 		 */
605 		io_deferred = list_first_entry(&log->running_ios,
606 					       struct r5l_io_unit, log_sibling);
607 		if (io_deferred->io_deferred)
608 			schedule_work(&log->deferred_io_work);
609 	}
610 
611 	spin_unlock_irqrestore(&log->io_list_lock, flags);
612 
613 	if (log->need_cache_flush)
614 		md_wakeup_thread(log->rdev->mddev->thread);
615 
616 	/* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
617 	if (has_null_flush) {
618 		struct bio *bi;
619 
620 		WARN_ON(bio_list_empty(&io->flush_barriers));
621 		while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
622 			bio_endio(bi);
623 			if (atomic_dec_and_test(&io->pending_stripe)) {
624 				__r5l_stripe_write_finished(io);
625 				return;
626 			}
627 		}
628 	}
629 	/* decrease pending_stripe for flush payload */
630 	if (has_flush_payload)
631 		if (atomic_dec_and_test(&io->pending_stripe))
632 			__r5l_stripe_write_finished(io);
633 }
634 
635 static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
636 {
637 	unsigned long flags;
638 
639 	spin_lock_irqsave(&log->io_list_lock, flags);
640 	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
641 	spin_unlock_irqrestore(&log->io_list_lock, flags);
642 
643 	/*
644 	 * In case of journal device failures, submit_bio will get error
645 	 * and calls endio, then active stripes will continue write
646 	 * process. Therefore, it is not necessary to check Faulty bit
647 	 * of journal device here.
648 	 *
649 	 * We can't check split_bio after current_bio is submitted. If
650 	 * io->split_bio is null, after current_bio is submitted, current_bio
651 	 * might already be completed and the io_unit is freed. We submit
652 	 * split_bio first to avoid the issue.
653 	 */
654 	if (io->split_bio) {
655 		if (io->has_flush)
656 			io->split_bio->bi_opf |= REQ_PREFLUSH;
657 		if (io->has_fua)
658 			io->split_bio->bi_opf |= REQ_FUA;
659 		submit_bio(io->split_bio);
660 	}
661 
662 	if (io->has_flush)
663 		io->current_bio->bi_opf |= REQ_PREFLUSH;
664 	if (io->has_fua)
665 		io->current_bio->bi_opf |= REQ_FUA;
666 	submit_bio(io->current_bio);
667 }
668 
669 /* deferred io_unit will be dispatched here */
670 static void r5l_submit_io_async(struct work_struct *work)
671 {
672 	struct r5l_log *log = container_of(work, struct r5l_log,
673 					   deferred_io_work);
674 	struct r5l_io_unit *io = NULL;
675 	unsigned long flags;
676 
677 	spin_lock_irqsave(&log->io_list_lock, flags);
678 	if (!list_empty(&log->running_ios)) {
679 		io = list_first_entry(&log->running_ios, struct r5l_io_unit,
680 				      log_sibling);
681 		if (!io->io_deferred)
682 			io = NULL;
683 		else
684 			io->io_deferred = 0;
685 	}
686 	spin_unlock_irqrestore(&log->io_list_lock, flags);
687 	if (io)
688 		r5l_do_submit_io(log, io);
689 }
690 
691 static void r5c_disable_writeback_async(struct work_struct *work)
692 {
693 	struct r5l_log *log = container_of(work, struct r5l_log,
694 					   disable_writeback_work);
695 	struct mddev *mddev = log->rdev->mddev;
696 	struct r5conf *conf = mddev->private;
697 	int locked = 0;
698 
699 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
700 		return;
701 	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
702 		mdname(mddev));
703 
704 	/* wait superblock change before suspend */
705 	wait_event(mddev->sb_wait,
706 		   conf->log == NULL ||
707 		   (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
708 		    (locked = mddev_trylock(mddev))));
709 	if (locked) {
710 		mddev_suspend(mddev);
711 		log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
712 		mddev_resume(mddev);
713 		mddev_unlock(mddev);
714 	}
715 }
716 
717 static void r5l_submit_current_io(struct r5l_log *log)
718 {
719 	struct r5l_io_unit *io = log->current_io;
720 	struct r5l_meta_block *block;
721 	unsigned long flags;
722 	u32 crc;
723 	bool do_submit = true;
724 
725 	if (!io)
726 		return;
727 
728 	block = page_address(io->meta_page);
729 	block->meta_size = cpu_to_le32(io->meta_offset);
730 	crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
731 	block->checksum = cpu_to_le32(crc);
732 
733 	log->current_io = NULL;
734 	spin_lock_irqsave(&log->io_list_lock, flags);
735 	if (io->has_flush || io->has_fua) {
736 		if (io != list_first_entry(&log->running_ios,
737 					   struct r5l_io_unit, log_sibling)) {
738 			io->io_deferred = 1;
739 			do_submit = false;
740 		}
741 	}
742 	spin_unlock_irqrestore(&log->io_list_lock, flags);
743 	if (do_submit)
744 		r5l_do_submit_io(log, io);
745 }
746 
747 static struct bio *r5l_bio_alloc(struct r5l_log *log)
748 {
749 	struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, &log->bs);
750 
751 	bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
752 	bio_set_dev(bio, log->rdev->bdev);
753 	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
754 
755 	return bio;
756 }
757 
758 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
759 {
760 	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
761 
762 	r5c_update_log_state(log);
763 	/*
764 	 * If we filled up the log device start from the beginning again,
765 	 * which will require a new bio.
766 	 *
767 	 * Note: for this to work properly the log size needs to me a multiple
768 	 * of BLOCK_SECTORS.
769 	 */
770 	if (log->log_start == 0)
771 		io->need_split_bio = true;
772 
773 	io->log_end = log->log_start;
774 }
775 
776 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
777 {
778 	struct r5l_io_unit *io;
779 	struct r5l_meta_block *block;
780 
781 	io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
782 	if (!io)
783 		return NULL;
784 	memset(io, 0, sizeof(*io));
785 
786 	io->log = log;
787 	INIT_LIST_HEAD(&io->log_sibling);
788 	INIT_LIST_HEAD(&io->stripe_list);
789 	bio_list_init(&io->flush_barriers);
790 	io->state = IO_UNIT_RUNNING;
791 
792 	io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
793 	block = page_address(io->meta_page);
794 	clear_page(block);
795 	block->magic = cpu_to_le32(R5LOG_MAGIC);
796 	block->version = R5LOG_VERSION;
797 	block->seq = cpu_to_le64(log->seq);
798 	block->position = cpu_to_le64(log->log_start);
799 
800 	io->log_start = log->log_start;
801 	io->meta_offset = sizeof(struct r5l_meta_block);
802 	io->seq = log->seq++;
803 
804 	io->current_bio = r5l_bio_alloc(log);
805 	io->current_bio->bi_end_io = r5l_log_endio;
806 	io->current_bio->bi_private = io;
807 	bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
808 
809 	r5_reserve_log_entry(log, io);
810 
811 	spin_lock_irq(&log->io_list_lock);
812 	list_add_tail(&io->log_sibling, &log->running_ios);
813 	spin_unlock_irq(&log->io_list_lock);
814 
815 	return io;
816 }
817 
818 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
819 {
820 	if (log->current_io &&
821 	    log->current_io->meta_offset + payload_size > PAGE_SIZE)
822 		r5l_submit_current_io(log);
823 
824 	if (!log->current_io) {
825 		log->current_io = r5l_new_meta(log);
826 		if (!log->current_io)
827 			return -ENOMEM;
828 	}
829 
830 	return 0;
831 }
832 
833 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
834 				    sector_t location,
835 				    u32 checksum1, u32 checksum2,
836 				    bool checksum2_valid)
837 {
838 	struct r5l_io_unit *io = log->current_io;
839 	struct r5l_payload_data_parity *payload;
840 
841 	payload = page_address(io->meta_page) + io->meta_offset;
842 	payload->header.type = cpu_to_le16(type);
843 	payload->header.flags = cpu_to_le16(0);
844 	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
845 				    (PAGE_SHIFT - 9));
846 	payload->location = cpu_to_le64(location);
847 	payload->checksum[0] = cpu_to_le32(checksum1);
848 	if (checksum2_valid)
849 		payload->checksum[1] = cpu_to_le32(checksum2);
850 
851 	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
852 		sizeof(__le32) * (1 + !!checksum2_valid);
853 }
854 
855 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
856 {
857 	struct r5l_io_unit *io = log->current_io;
858 
859 	if (io->need_split_bio) {
860 		BUG_ON(io->split_bio);
861 		io->split_bio = io->current_bio;
862 		io->current_bio = r5l_bio_alloc(log);
863 		bio_chain(io->current_bio, io->split_bio);
864 		io->need_split_bio = false;
865 	}
866 
867 	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
868 		BUG();
869 
870 	r5_reserve_log_entry(log, io);
871 }
872 
873 static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
874 {
875 	struct mddev *mddev = log->rdev->mddev;
876 	struct r5conf *conf = mddev->private;
877 	struct r5l_io_unit *io;
878 	struct r5l_payload_flush *payload;
879 	int meta_size;
880 
881 	/*
882 	 * payload_flush requires extra writes to the journal.
883 	 * To avoid handling the extra IO in quiesce, just skip
884 	 * flush_payload
885 	 */
886 	if (conf->quiesce)
887 		return;
888 
889 	mutex_lock(&log->io_mutex);
890 	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
891 
892 	if (r5l_get_meta(log, meta_size)) {
893 		mutex_unlock(&log->io_mutex);
894 		return;
895 	}
896 
897 	/* current implementation is one stripe per flush payload */
898 	io = log->current_io;
899 	payload = page_address(io->meta_page) + io->meta_offset;
900 	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
901 	payload->header.flags = cpu_to_le16(0);
902 	payload->size = cpu_to_le32(sizeof(__le64));
903 	payload->flush_stripes[0] = cpu_to_le64(sect);
904 	io->meta_offset += meta_size;
905 	/* multiple flush payloads count as one pending_stripe */
906 	if (!io->has_flush_payload) {
907 		io->has_flush_payload = 1;
908 		atomic_inc(&io->pending_stripe);
909 	}
910 	mutex_unlock(&log->io_mutex);
911 }
912 
913 static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
914 			   int data_pages, int parity_pages)
915 {
916 	int i;
917 	int meta_size;
918 	int ret;
919 	struct r5l_io_unit *io;
920 
921 	meta_size =
922 		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
923 		 * data_pages) +
924 		sizeof(struct r5l_payload_data_parity) +
925 		sizeof(__le32) * parity_pages;
926 
927 	ret = r5l_get_meta(log, meta_size);
928 	if (ret)
929 		return ret;
930 
931 	io = log->current_io;
932 
933 	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
934 		io->has_flush = 1;
935 
936 	for (i = 0; i < sh->disks; i++) {
937 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
938 		    test_bit(R5_InJournal, &sh->dev[i].flags))
939 			continue;
940 		if (i == sh->pd_idx || i == sh->qd_idx)
941 			continue;
942 		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
943 		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
944 			io->has_fua = 1;
945 			/*
946 			 * we need to flush journal to make sure recovery can
947 			 * reach the data with fua flag
948 			 */
949 			io->has_flush = 1;
950 		}
951 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
952 					raid5_compute_blocknr(sh, i, 0),
953 					sh->dev[i].log_checksum, 0, false);
954 		r5l_append_payload_page(log, sh->dev[i].page);
955 	}
956 
957 	if (parity_pages == 2) {
958 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
959 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
960 					sh->dev[sh->qd_idx].log_checksum, true);
961 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
962 		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
963 	} else if (parity_pages == 1) {
964 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
965 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
966 					0, false);
967 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
968 	} else  /* Just writing data, not parity, in caching phase */
969 		BUG_ON(parity_pages != 0);
970 
971 	list_add_tail(&sh->log_list, &io->stripe_list);
972 	atomic_inc(&io->pending_stripe);
973 	sh->log_io = io;
974 
975 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
976 		return 0;
977 
978 	if (sh->log_start == MaxSector) {
979 		BUG_ON(!list_empty(&sh->r5c));
980 		sh->log_start = io->log_start;
981 		spin_lock_irq(&log->stripe_in_journal_lock);
982 		list_add_tail(&sh->r5c,
983 			      &log->stripe_in_journal_list);
984 		spin_unlock_irq(&log->stripe_in_journal_lock);
985 		atomic_inc(&log->stripe_in_journal_count);
986 	}
987 	return 0;
988 }
989 
990 /* add stripe to no_space_stripes, and then wake up reclaim */
991 static inline void r5l_add_no_space_stripe(struct r5l_log *log,
992 					   struct stripe_head *sh)
993 {
994 	spin_lock(&log->no_space_stripes_lock);
995 	list_add_tail(&sh->log_list, &log->no_space_stripes);
996 	spin_unlock(&log->no_space_stripes_lock);
997 }
998 
999 /*
1000  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
1001  * data from log to raid disks), so we shouldn't wait for reclaim here
1002  */
1003 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
1004 {
1005 	struct r5conf *conf = sh->raid_conf;
1006 	int write_disks = 0;
1007 	int data_pages, parity_pages;
1008 	int reserve;
1009 	int i;
1010 	int ret = 0;
1011 	bool wake_reclaim = false;
1012 
1013 	if (!log)
1014 		return -EAGAIN;
1015 	/* Don't support stripe batch */
1016 	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1017 	    test_bit(STRIPE_SYNCING, &sh->state)) {
1018 		/* the stripe is written to log, we start writing it to raid */
1019 		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1020 		return -EAGAIN;
1021 	}
1022 
1023 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1024 
1025 	for (i = 0; i < sh->disks; i++) {
1026 		void *addr;
1027 
1028 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1029 		    test_bit(R5_InJournal, &sh->dev[i].flags))
1030 			continue;
1031 
1032 		write_disks++;
1033 		/* checksum is already calculated in last run */
1034 		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1035 			continue;
1036 		addr = kmap_atomic(sh->dev[i].page);
1037 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1038 						    addr, PAGE_SIZE);
1039 		kunmap_atomic(addr);
1040 	}
1041 	parity_pages = 1 + !!(sh->qd_idx >= 0);
1042 	data_pages = write_disks - parity_pages;
1043 
1044 	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1045 	/*
1046 	 * The stripe must enter state machine again to finish the write, so
1047 	 * don't delay.
1048 	 */
1049 	clear_bit(STRIPE_DELAYED, &sh->state);
1050 	atomic_inc(&sh->count);
1051 
1052 	mutex_lock(&log->io_mutex);
1053 	/* meta + data */
1054 	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1055 
1056 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1057 		if (!r5l_has_free_space(log, reserve)) {
1058 			r5l_add_no_space_stripe(log, sh);
1059 			wake_reclaim = true;
1060 		} else {
1061 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1062 			if (ret) {
1063 				spin_lock_irq(&log->io_list_lock);
1064 				list_add_tail(&sh->log_list,
1065 					      &log->no_mem_stripes);
1066 				spin_unlock_irq(&log->io_list_lock);
1067 			}
1068 		}
1069 	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1070 		/*
1071 		 * log space critical, do not process stripes that are
1072 		 * not in cache yet (sh->log_start == MaxSector).
1073 		 */
1074 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1075 		    sh->log_start == MaxSector) {
1076 			r5l_add_no_space_stripe(log, sh);
1077 			wake_reclaim = true;
1078 			reserve = 0;
1079 		} else if (!r5l_has_free_space(log, reserve)) {
1080 			if (sh->log_start == log->last_checkpoint)
1081 				BUG();
1082 			else
1083 				r5l_add_no_space_stripe(log, sh);
1084 		} else {
1085 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1086 			if (ret) {
1087 				spin_lock_irq(&log->io_list_lock);
1088 				list_add_tail(&sh->log_list,
1089 					      &log->no_mem_stripes);
1090 				spin_unlock_irq(&log->io_list_lock);
1091 			}
1092 		}
1093 	}
1094 
1095 	mutex_unlock(&log->io_mutex);
1096 	if (wake_reclaim)
1097 		r5l_wake_reclaim(log, reserve);
1098 	return 0;
1099 }
1100 
1101 void r5l_write_stripe_run(struct r5l_log *log)
1102 {
1103 	if (!log)
1104 		return;
1105 	mutex_lock(&log->io_mutex);
1106 	r5l_submit_current_io(log);
1107 	mutex_unlock(&log->io_mutex);
1108 }
1109 
1110 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1111 {
1112 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1113 		/*
1114 		 * in write through (journal only)
1115 		 * we flush log disk cache first, then write stripe data to
1116 		 * raid disks. So if bio is finished, the log disk cache is
1117 		 * flushed already. The recovery guarantees we can recovery
1118 		 * the bio from log disk, so we don't need to flush again
1119 		 */
1120 		if (bio->bi_iter.bi_size == 0) {
1121 			bio_endio(bio);
1122 			return 0;
1123 		}
1124 		bio->bi_opf &= ~REQ_PREFLUSH;
1125 	} else {
1126 		/* write back (with cache) */
1127 		if (bio->bi_iter.bi_size == 0) {
1128 			mutex_lock(&log->io_mutex);
1129 			r5l_get_meta(log, 0);
1130 			bio_list_add(&log->current_io->flush_barriers, bio);
1131 			log->current_io->has_flush = 1;
1132 			log->current_io->has_null_flush = 1;
1133 			atomic_inc(&log->current_io->pending_stripe);
1134 			r5l_submit_current_io(log);
1135 			mutex_unlock(&log->io_mutex);
1136 			return 0;
1137 		}
1138 	}
1139 	return -EAGAIN;
1140 }
1141 
1142 /* This will run after log space is reclaimed */
1143 static void r5l_run_no_space_stripes(struct r5l_log *log)
1144 {
1145 	struct stripe_head *sh;
1146 
1147 	spin_lock(&log->no_space_stripes_lock);
1148 	while (!list_empty(&log->no_space_stripes)) {
1149 		sh = list_first_entry(&log->no_space_stripes,
1150 				      struct stripe_head, log_list);
1151 		list_del_init(&sh->log_list);
1152 		set_bit(STRIPE_HANDLE, &sh->state);
1153 		raid5_release_stripe(sh);
1154 	}
1155 	spin_unlock(&log->no_space_stripes_lock);
1156 }
1157 
1158 /*
1159  * calculate new last_checkpoint
1160  * for write through mode, returns log->next_checkpoint
1161  * for write back, returns log_start of first sh in stripe_in_journal_list
1162  */
1163 static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1164 {
1165 	struct stripe_head *sh;
1166 	struct r5l_log *log = conf->log;
1167 	sector_t new_cp;
1168 	unsigned long flags;
1169 
1170 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1171 		return log->next_checkpoint;
1172 
1173 	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1174 	if (list_empty(&conf->log->stripe_in_journal_list)) {
1175 		/* all stripes flushed */
1176 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1177 		return log->next_checkpoint;
1178 	}
1179 	sh = list_first_entry(&conf->log->stripe_in_journal_list,
1180 			      struct stripe_head, r5c);
1181 	new_cp = sh->log_start;
1182 	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1183 	return new_cp;
1184 }
1185 
1186 static sector_t r5l_reclaimable_space(struct r5l_log *log)
1187 {
1188 	struct r5conf *conf = log->rdev->mddev->private;
1189 
1190 	return r5l_ring_distance(log, log->last_checkpoint,
1191 				 r5c_calculate_new_cp(conf));
1192 }
1193 
1194 static void r5l_run_no_mem_stripe(struct r5l_log *log)
1195 {
1196 	struct stripe_head *sh;
1197 
1198 	lockdep_assert_held(&log->io_list_lock);
1199 
1200 	if (!list_empty(&log->no_mem_stripes)) {
1201 		sh = list_first_entry(&log->no_mem_stripes,
1202 				      struct stripe_head, log_list);
1203 		list_del_init(&sh->log_list);
1204 		set_bit(STRIPE_HANDLE, &sh->state);
1205 		raid5_release_stripe(sh);
1206 	}
1207 }
1208 
1209 static bool r5l_complete_finished_ios(struct r5l_log *log)
1210 {
1211 	struct r5l_io_unit *io, *next;
1212 	bool found = false;
1213 
1214 	lockdep_assert_held(&log->io_list_lock);
1215 
1216 	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1217 		/* don't change list order */
1218 		if (io->state < IO_UNIT_STRIPE_END)
1219 			break;
1220 
1221 		log->next_checkpoint = io->log_start;
1222 
1223 		list_del(&io->log_sibling);
1224 		mempool_free(io, &log->io_pool);
1225 		r5l_run_no_mem_stripe(log);
1226 
1227 		found = true;
1228 	}
1229 
1230 	return found;
1231 }
1232 
1233 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1234 {
1235 	struct r5l_log *log = io->log;
1236 	struct r5conf *conf = log->rdev->mddev->private;
1237 	unsigned long flags;
1238 
1239 	spin_lock_irqsave(&log->io_list_lock, flags);
1240 	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1241 
1242 	if (!r5l_complete_finished_ios(log)) {
1243 		spin_unlock_irqrestore(&log->io_list_lock, flags);
1244 		return;
1245 	}
1246 
1247 	if (r5l_reclaimable_space(log) > log->max_free_space ||
1248 	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1249 		r5l_wake_reclaim(log, 0);
1250 
1251 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1252 	wake_up(&log->iounit_wait);
1253 }
1254 
1255 void r5l_stripe_write_finished(struct stripe_head *sh)
1256 {
1257 	struct r5l_io_unit *io;
1258 
1259 	io = sh->log_io;
1260 	sh->log_io = NULL;
1261 
1262 	if (io && atomic_dec_and_test(&io->pending_stripe))
1263 		__r5l_stripe_write_finished(io);
1264 }
1265 
1266 static void r5l_log_flush_endio(struct bio *bio)
1267 {
1268 	struct r5l_log *log = container_of(bio, struct r5l_log,
1269 		flush_bio);
1270 	unsigned long flags;
1271 	struct r5l_io_unit *io;
1272 
1273 	if (bio->bi_status)
1274 		md_error(log->rdev->mddev, log->rdev);
1275 
1276 	spin_lock_irqsave(&log->io_list_lock, flags);
1277 	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1278 		r5l_io_run_stripes(io);
1279 	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1280 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1281 }
1282 
1283 /*
1284  * Starting dispatch IO to raid.
1285  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1286  * broken meta in the middle of a log causes recovery can't find meta at the
1287  * head of log. If operations require meta at the head persistent in log, we
1288  * must make sure meta before it persistent in log too. A case is:
1289  *
1290  * stripe data/parity is in log, we start write stripe to raid disks. stripe
1291  * data/parity must be persistent in log before we do the write to raid disks.
1292  *
1293  * The solution is we restrictly maintain io_unit list order. In this case, we
1294  * only write stripes of an io_unit to raid disks till the io_unit is the first
1295  * one whose data/parity is in log.
1296  */
1297 void r5l_flush_stripe_to_raid(struct r5l_log *log)
1298 {
1299 	bool do_flush;
1300 
1301 	if (!log || !log->need_cache_flush)
1302 		return;
1303 
1304 	spin_lock_irq(&log->io_list_lock);
1305 	/* flush bio is running */
1306 	if (!list_empty(&log->flushing_ios)) {
1307 		spin_unlock_irq(&log->io_list_lock);
1308 		return;
1309 	}
1310 	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1311 	do_flush = !list_empty(&log->flushing_ios);
1312 	spin_unlock_irq(&log->io_list_lock);
1313 
1314 	if (!do_flush)
1315 		return;
1316 	bio_reset(&log->flush_bio);
1317 	bio_set_dev(&log->flush_bio, log->rdev->bdev);
1318 	log->flush_bio.bi_end_io = r5l_log_flush_endio;
1319 	log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1320 	submit_bio(&log->flush_bio);
1321 }
1322 
1323 static void r5l_write_super(struct r5l_log *log, sector_t cp);
1324 static void r5l_write_super_and_discard_space(struct r5l_log *log,
1325 	sector_t end)
1326 {
1327 	struct block_device *bdev = log->rdev->bdev;
1328 	struct mddev *mddev;
1329 
1330 	r5l_write_super(log, end);
1331 
1332 	if (!blk_queue_discard(bdev_get_queue(bdev)))
1333 		return;
1334 
1335 	mddev = log->rdev->mddev;
1336 	/*
1337 	 * Discard could zero data, so before discard we must make sure
1338 	 * superblock is updated to new log tail. Updating superblock (either
1339 	 * directly call md_update_sb() or depend on md thread) must hold
1340 	 * reconfig mutex. On the other hand, raid5_quiesce is called with
1341 	 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1342 	 * for all IO finish, hence waitting for reclaim thread, while reclaim
1343 	 * thread is calling this function and waitting for reconfig mutex. So
1344 	 * there is a deadlock. We workaround this issue with a trylock.
1345 	 * FIXME: we could miss discard if we can't take reconfig mutex
1346 	 */
1347 	set_mask_bits(&mddev->sb_flags, 0,
1348 		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1349 	if (!mddev_trylock(mddev))
1350 		return;
1351 	md_update_sb(mddev, 1);
1352 	mddev_unlock(mddev);
1353 
1354 	/* discard IO error really doesn't matter, ignore it */
1355 	if (log->last_checkpoint < end) {
1356 		blkdev_issue_discard(bdev,
1357 				log->last_checkpoint + log->rdev->data_offset,
1358 				end - log->last_checkpoint, GFP_NOIO, 0);
1359 	} else {
1360 		blkdev_issue_discard(bdev,
1361 				log->last_checkpoint + log->rdev->data_offset,
1362 				log->device_size - log->last_checkpoint,
1363 				GFP_NOIO, 0);
1364 		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1365 				GFP_NOIO, 0);
1366 	}
1367 }
1368 
1369 /*
1370  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1371  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1372  *
1373  * must hold conf->device_lock
1374  */
1375 static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1376 {
1377 	BUG_ON(list_empty(&sh->lru));
1378 	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1379 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1380 
1381 	/*
1382 	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1383 	 * raid5_release_stripe() while holding conf->device_lock
1384 	 */
1385 	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1386 	lockdep_assert_held(&conf->device_lock);
1387 
1388 	list_del_init(&sh->lru);
1389 	atomic_inc(&sh->count);
1390 
1391 	set_bit(STRIPE_HANDLE, &sh->state);
1392 	atomic_inc(&conf->active_stripes);
1393 	r5c_make_stripe_write_out(sh);
1394 
1395 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1396 		atomic_inc(&conf->r5c_flushing_partial_stripes);
1397 	else
1398 		atomic_inc(&conf->r5c_flushing_full_stripes);
1399 	raid5_release_stripe(sh);
1400 }
1401 
1402 /*
1403  * if num == 0, flush all full stripes
1404  * if num > 0, flush all full stripes. If less than num full stripes are
1405  *             flushed, flush some partial stripes until totally num stripes are
1406  *             flushed or there is no more cached stripes.
1407  */
1408 void r5c_flush_cache(struct r5conf *conf, int num)
1409 {
1410 	int count;
1411 	struct stripe_head *sh, *next;
1412 
1413 	lockdep_assert_held(&conf->device_lock);
1414 	if (!conf->log)
1415 		return;
1416 
1417 	count = 0;
1418 	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1419 		r5c_flush_stripe(conf, sh);
1420 		count++;
1421 	}
1422 
1423 	if (count >= num)
1424 		return;
1425 	list_for_each_entry_safe(sh, next,
1426 				 &conf->r5c_partial_stripe_list, lru) {
1427 		r5c_flush_stripe(conf, sh);
1428 		if (++count >= num)
1429 			break;
1430 	}
1431 }
1432 
1433 static void r5c_do_reclaim(struct r5conf *conf)
1434 {
1435 	struct r5l_log *log = conf->log;
1436 	struct stripe_head *sh;
1437 	int count = 0;
1438 	unsigned long flags;
1439 	int total_cached;
1440 	int stripes_to_flush;
1441 	int flushing_partial, flushing_full;
1442 
1443 	if (!r5c_is_writeback(log))
1444 		return;
1445 
1446 	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1447 	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1448 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1449 		atomic_read(&conf->r5c_cached_full_stripes) -
1450 		flushing_full - flushing_partial;
1451 
1452 	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1453 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1454 		/*
1455 		 * if stripe cache pressure high, flush all full stripes and
1456 		 * some partial stripes
1457 		 */
1458 		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1459 	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1460 		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1461 		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1462 		/*
1463 		 * if stripe cache pressure moderate, or if there is many full
1464 		 * stripes,flush all full stripes
1465 		 */
1466 		stripes_to_flush = 0;
1467 	else
1468 		/* no need to flush */
1469 		stripes_to_flush = -1;
1470 
1471 	if (stripes_to_flush >= 0) {
1472 		spin_lock_irqsave(&conf->device_lock, flags);
1473 		r5c_flush_cache(conf, stripes_to_flush);
1474 		spin_unlock_irqrestore(&conf->device_lock, flags);
1475 	}
1476 
1477 	/* if log space is tight, flush stripes on stripe_in_journal_list */
1478 	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1479 		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1480 		spin_lock(&conf->device_lock);
1481 		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1482 			/*
1483 			 * stripes on stripe_in_journal_list could be in any
1484 			 * state of the stripe_cache state machine. In this
1485 			 * case, we only want to flush stripe on
1486 			 * r5c_cached_full/partial_stripes. The following
1487 			 * condition makes sure the stripe is on one of the
1488 			 * two lists.
1489 			 */
1490 			if (!list_empty(&sh->lru) &&
1491 			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1492 			    atomic_read(&sh->count) == 0) {
1493 				r5c_flush_stripe(conf, sh);
1494 				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1495 					break;
1496 			}
1497 		}
1498 		spin_unlock(&conf->device_lock);
1499 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1500 	}
1501 
1502 	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1503 		r5l_run_no_space_stripes(log);
1504 
1505 	md_wakeup_thread(conf->mddev->thread);
1506 }
1507 
1508 static void r5l_do_reclaim(struct r5l_log *log)
1509 {
1510 	struct r5conf *conf = log->rdev->mddev->private;
1511 	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1512 	sector_t reclaimable;
1513 	sector_t next_checkpoint;
1514 	bool write_super;
1515 
1516 	spin_lock_irq(&log->io_list_lock);
1517 	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1518 		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1519 	/*
1520 	 * move proper io_unit to reclaim list. We should not change the order.
1521 	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1522 	 * shouldn't reuse space of an unreclaimable io_unit
1523 	 */
1524 	while (1) {
1525 		reclaimable = r5l_reclaimable_space(log);
1526 		if (reclaimable >= reclaim_target ||
1527 		    (list_empty(&log->running_ios) &&
1528 		     list_empty(&log->io_end_ios) &&
1529 		     list_empty(&log->flushing_ios) &&
1530 		     list_empty(&log->finished_ios)))
1531 			break;
1532 
1533 		md_wakeup_thread(log->rdev->mddev->thread);
1534 		wait_event_lock_irq(log->iounit_wait,
1535 				    r5l_reclaimable_space(log) > reclaimable,
1536 				    log->io_list_lock);
1537 	}
1538 
1539 	next_checkpoint = r5c_calculate_new_cp(conf);
1540 	spin_unlock_irq(&log->io_list_lock);
1541 
1542 	if (reclaimable == 0 || !write_super)
1543 		return;
1544 
1545 	/*
1546 	 * write_super will flush cache of each raid disk. We must write super
1547 	 * here, because the log area might be reused soon and we don't want to
1548 	 * confuse recovery
1549 	 */
1550 	r5l_write_super_and_discard_space(log, next_checkpoint);
1551 
1552 	mutex_lock(&log->io_mutex);
1553 	log->last_checkpoint = next_checkpoint;
1554 	r5c_update_log_state(log);
1555 	mutex_unlock(&log->io_mutex);
1556 
1557 	r5l_run_no_space_stripes(log);
1558 }
1559 
1560 static void r5l_reclaim_thread(struct md_thread *thread)
1561 {
1562 	struct mddev *mddev = thread->mddev;
1563 	struct r5conf *conf = mddev->private;
1564 	struct r5l_log *log = conf->log;
1565 
1566 	if (!log)
1567 		return;
1568 	r5c_do_reclaim(conf);
1569 	r5l_do_reclaim(log);
1570 }
1571 
1572 void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1573 {
1574 	unsigned long target;
1575 	unsigned long new = (unsigned long)space; /* overflow in theory */
1576 
1577 	if (!log)
1578 		return;
1579 	do {
1580 		target = log->reclaim_target;
1581 		if (new < target)
1582 			return;
1583 	} while (cmpxchg(&log->reclaim_target, target, new) != target);
1584 	md_wakeup_thread(log->reclaim_thread);
1585 }
1586 
1587 void r5l_quiesce(struct r5l_log *log, int quiesce)
1588 {
1589 	struct mddev *mddev;
1590 
1591 	if (quiesce) {
1592 		/* make sure r5l_write_super_and_discard_space exits */
1593 		mddev = log->rdev->mddev;
1594 		wake_up(&mddev->sb_wait);
1595 		kthread_park(log->reclaim_thread->tsk);
1596 		r5l_wake_reclaim(log, MaxSector);
1597 		r5l_do_reclaim(log);
1598 	} else
1599 		kthread_unpark(log->reclaim_thread->tsk);
1600 }
1601 
1602 bool r5l_log_disk_error(struct r5conf *conf)
1603 {
1604 	struct r5l_log *log;
1605 	bool ret;
1606 	/* don't allow write if journal disk is missing */
1607 	rcu_read_lock();
1608 	log = rcu_dereference(conf->log);
1609 
1610 	if (!log)
1611 		ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1612 	else
1613 		ret = test_bit(Faulty, &log->rdev->flags);
1614 	rcu_read_unlock();
1615 	return ret;
1616 }
1617 
1618 #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1619 
1620 struct r5l_recovery_ctx {
1621 	struct page *meta_page;		/* current meta */
1622 	sector_t meta_total_blocks;	/* total size of current meta and data */
1623 	sector_t pos;			/* recovery position */
1624 	u64 seq;			/* recovery position seq */
1625 	int data_parity_stripes;	/* number of data_parity stripes */
1626 	int data_only_stripes;		/* number of data_only stripes */
1627 	struct list_head cached_list;
1628 
1629 	/*
1630 	 * read ahead page pool (ra_pool)
1631 	 * in recovery, log is read sequentially. It is not efficient to
1632 	 * read every page with sync_page_io(). The read ahead page pool
1633 	 * reads multiple pages with one IO, so further log read can
1634 	 * just copy data from the pool.
1635 	 */
1636 	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1637 	sector_t pool_offset;	/* offset of first page in the pool */
1638 	int total_pages;	/* total allocated pages */
1639 	int valid_pages;	/* pages with valid data */
1640 	struct bio *ra_bio;	/* bio to do the read ahead */
1641 };
1642 
1643 static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1644 					    struct r5l_recovery_ctx *ctx)
1645 {
1646 	struct page *page;
1647 
1648 	ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, &log->bs);
1649 	if (!ctx->ra_bio)
1650 		return -ENOMEM;
1651 
1652 	ctx->valid_pages = 0;
1653 	ctx->total_pages = 0;
1654 	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1655 		page = alloc_page(GFP_KERNEL);
1656 
1657 		if (!page)
1658 			break;
1659 		ctx->ra_pool[ctx->total_pages] = page;
1660 		ctx->total_pages += 1;
1661 	}
1662 
1663 	if (ctx->total_pages == 0) {
1664 		bio_put(ctx->ra_bio);
1665 		return -ENOMEM;
1666 	}
1667 
1668 	ctx->pool_offset = 0;
1669 	return 0;
1670 }
1671 
1672 static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1673 					struct r5l_recovery_ctx *ctx)
1674 {
1675 	int i;
1676 
1677 	for (i = 0; i < ctx->total_pages; ++i)
1678 		put_page(ctx->ra_pool[i]);
1679 	bio_put(ctx->ra_bio);
1680 }
1681 
1682 /*
1683  * fetch ctx->valid_pages pages from offset
1684  * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1685  * However, if the offset is close to the end of the journal device,
1686  * ctx->valid_pages could be smaller than ctx->total_pages
1687  */
1688 static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1689 				      struct r5l_recovery_ctx *ctx,
1690 				      sector_t offset)
1691 {
1692 	bio_reset(ctx->ra_bio);
1693 	bio_set_dev(ctx->ra_bio, log->rdev->bdev);
1694 	bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1695 	ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1696 
1697 	ctx->valid_pages = 0;
1698 	ctx->pool_offset = offset;
1699 
1700 	while (ctx->valid_pages < ctx->total_pages) {
1701 		bio_add_page(ctx->ra_bio,
1702 			     ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1703 		ctx->valid_pages += 1;
1704 
1705 		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1706 
1707 		if (offset == 0)  /* reached end of the device */
1708 			break;
1709 	}
1710 
1711 	return submit_bio_wait(ctx->ra_bio);
1712 }
1713 
1714 /*
1715  * try read a page from the read ahead page pool, if the page is not in the
1716  * pool, call r5l_recovery_fetch_ra_pool
1717  */
1718 static int r5l_recovery_read_page(struct r5l_log *log,
1719 				  struct r5l_recovery_ctx *ctx,
1720 				  struct page *page,
1721 				  sector_t offset)
1722 {
1723 	int ret;
1724 
1725 	if (offset < ctx->pool_offset ||
1726 	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1727 		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1728 		if (ret)
1729 			return ret;
1730 	}
1731 
1732 	BUG_ON(offset < ctx->pool_offset ||
1733 	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1734 
1735 	memcpy(page_address(page),
1736 	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1737 					 BLOCK_SECTOR_SHIFT]),
1738 	       PAGE_SIZE);
1739 	return 0;
1740 }
1741 
1742 static int r5l_recovery_read_meta_block(struct r5l_log *log,
1743 					struct r5l_recovery_ctx *ctx)
1744 {
1745 	struct page *page = ctx->meta_page;
1746 	struct r5l_meta_block *mb;
1747 	u32 crc, stored_crc;
1748 	int ret;
1749 
1750 	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1751 	if (ret != 0)
1752 		return ret;
1753 
1754 	mb = page_address(page);
1755 	stored_crc = le32_to_cpu(mb->checksum);
1756 	mb->checksum = 0;
1757 
1758 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1759 	    le64_to_cpu(mb->seq) != ctx->seq ||
1760 	    mb->version != R5LOG_VERSION ||
1761 	    le64_to_cpu(mb->position) != ctx->pos)
1762 		return -EINVAL;
1763 
1764 	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1765 	if (stored_crc != crc)
1766 		return -EINVAL;
1767 
1768 	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1769 		return -EINVAL;
1770 
1771 	ctx->meta_total_blocks = BLOCK_SECTORS;
1772 
1773 	return 0;
1774 }
1775 
1776 static void
1777 r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1778 				     struct page *page,
1779 				     sector_t pos, u64 seq)
1780 {
1781 	struct r5l_meta_block *mb;
1782 
1783 	mb = page_address(page);
1784 	clear_page(mb);
1785 	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1786 	mb->version = R5LOG_VERSION;
1787 	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1788 	mb->seq = cpu_to_le64(seq);
1789 	mb->position = cpu_to_le64(pos);
1790 }
1791 
1792 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1793 					  u64 seq)
1794 {
1795 	struct page *page;
1796 	struct r5l_meta_block *mb;
1797 
1798 	page = alloc_page(GFP_KERNEL);
1799 	if (!page)
1800 		return -ENOMEM;
1801 	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1802 	mb = page_address(page);
1803 	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1804 					     mb, PAGE_SIZE));
1805 	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1806 			  REQ_SYNC | REQ_FUA, false)) {
1807 		__free_page(page);
1808 		return -EIO;
1809 	}
1810 	__free_page(page);
1811 	return 0;
1812 }
1813 
1814 /*
1815  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1816  * to mark valid (potentially not flushed) data in the journal.
1817  *
1818  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1819  * so there should not be any mismatch here.
1820  */
1821 static void r5l_recovery_load_data(struct r5l_log *log,
1822 				   struct stripe_head *sh,
1823 				   struct r5l_recovery_ctx *ctx,
1824 				   struct r5l_payload_data_parity *payload,
1825 				   sector_t log_offset)
1826 {
1827 	struct mddev *mddev = log->rdev->mddev;
1828 	struct r5conf *conf = mddev->private;
1829 	int dd_idx;
1830 
1831 	raid5_compute_sector(conf,
1832 			     le64_to_cpu(payload->location), 0,
1833 			     &dd_idx, sh);
1834 	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1835 	sh->dev[dd_idx].log_checksum =
1836 		le32_to_cpu(payload->checksum[0]);
1837 	ctx->meta_total_blocks += BLOCK_SECTORS;
1838 
1839 	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1840 	set_bit(STRIPE_R5C_CACHING, &sh->state);
1841 }
1842 
1843 static void r5l_recovery_load_parity(struct r5l_log *log,
1844 				     struct stripe_head *sh,
1845 				     struct r5l_recovery_ctx *ctx,
1846 				     struct r5l_payload_data_parity *payload,
1847 				     sector_t log_offset)
1848 {
1849 	struct mddev *mddev = log->rdev->mddev;
1850 	struct r5conf *conf = mddev->private;
1851 
1852 	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1853 	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1854 	sh->dev[sh->pd_idx].log_checksum =
1855 		le32_to_cpu(payload->checksum[0]);
1856 	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1857 
1858 	if (sh->qd_idx >= 0) {
1859 		r5l_recovery_read_page(
1860 			log, ctx, sh->dev[sh->qd_idx].page,
1861 			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1862 		sh->dev[sh->qd_idx].log_checksum =
1863 			le32_to_cpu(payload->checksum[1]);
1864 		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1865 	}
1866 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1867 }
1868 
1869 static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1870 {
1871 	int i;
1872 
1873 	sh->state = 0;
1874 	sh->log_start = MaxSector;
1875 	for (i = sh->disks; i--; )
1876 		sh->dev[i].flags = 0;
1877 }
1878 
1879 static void
1880 r5l_recovery_replay_one_stripe(struct r5conf *conf,
1881 			       struct stripe_head *sh,
1882 			       struct r5l_recovery_ctx *ctx)
1883 {
1884 	struct md_rdev *rdev, *rrdev;
1885 	int disk_index;
1886 	int data_count = 0;
1887 
1888 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1889 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1890 			continue;
1891 		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1892 			continue;
1893 		data_count++;
1894 	}
1895 
1896 	/*
1897 	 * stripes that only have parity must have been flushed
1898 	 * before the crash that we are now recovering from, so
1899 	 * there is nothing more to recovery.
1900 	 */
1901 	if (data_count == 0)
1902 		goto out;
1903 
1904 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1905 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1906 			continue;
1907 
1908 		/* in case device is broken */
1909 		rcu_read_lock();
1910 		rdev = rcu_dereference(conf->disks[disk_index].rdev);
1911 		if (rdev) {
1912 			atomic_inc(&rdev->nr_pending);
1913 			rcu_read_unlock();
1914 			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1915 				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1916 				     false);
1917 			rdev_dec_pending(rdev, rdev->mddev);
1918 			rcu_read_lock();
1919 		}
1920 		rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1921 		if (rrdev) {
1922 			atomic_inc(&rrdev->nr_pending);
1923 			rcu_read_unlock();
1924 			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1925 				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1926 				     false);
1927 			rdev_dec_pending(rrdev, rrdev->mddev);
1928 			rcu_read_lock();
1929 		}
1930 		rcu_read_unlock();
1931 	}
1932 	ctx->data_parity_stripes++;
1933 out:
1934 	r5l_recovery_reset_stripe(sh);
1935 }
1936 
1937 static struct stripe_head *
1938 r5c_recovery_alloc_stripe(
1939 		struct r5conf *conf,
1940 		sector_t stripe_sect,
1941 		int noblock)
1942 {
1943 	struct stripe_head *sh;
1944 
1945 	sh = raid5_get_active_stripe(conf, stripe_sect, 0, noblock, 0);
1946 	if (!sh)
1947 		return NULL;  /* no more stripe available */
1948 
1949 	r5l_recovery_reset_stripe(sh);
1950 
1951 	return sh;
1952 }
1953 
1954 static struct stripe_head *
1955 r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1956 {
1957 	struct stripe_head *sh;
1958 
1959 	list_for_each_entry(sh, list, lru)
1960 		if (sh->sector == sect)
1961 			return sh;
1962 	return NULL;
1963 }
1964 
1965 static void
1966 r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1967 			  struct r5l_recovery_ctx *ctx)
1968 {
1969 	struct stripe_head *sh, *next;
1970 
1971 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1972 		r5l_recovery_reset_stripe(sh);
1973 		list_del_init(&sh->lru);
1974 		raid5_release_stripe(sh);
1975 	}
1976 }
1977 
1978 static void
1979 r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1980 			    struct r5l_recovery_ctx *ctx)
1981 {
1982 	struct stripe_head *sh, *next;
1983 
1984 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1985 		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1986 			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1987 			list_del_init(&sh->lru);
1988 			raid5_release_stripe(sh);
1989 		}
1990 }
1991 
1992 /* if matches return 0; otherwise return -EINVAL */
1993 static int
1994 r5l_recovery_verify_data_checksum(struct r5l_log *log,
1995 				  struct r5l_recovery_ctx *ctx,
1996 				  struct page *page,
1997 				  sector_t log_offset, __le32 log_checksum)
1998 {
1999 	void *addr;
2000 	u32 checksum;
2001 
2002 	r5l_recovery_read_page(log, ctx, page, log_offset);
2003 	addr = kmap_atomic(page);
2004 	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
2005 	kunmap_atomic(addr);
2006 	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
2007 }
2008 
2009 /*
2010  * before loading data to stripe cache, we need verify checksum for all data,
2011  * if there is mismatch for any data page, we drop all data in the mata block
2012  */
2013 static int
2014 r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
2015 					 struct r5l_recovery_ctx *ctx)
2016 {
2017 	struct mddev *mddev = log->rdev->mddev;
2018 	struct r5conf *conf = mddev->private;
2019 	struct r5l_meta_block *mb = page_address(ctx->meta_page);
2020 	sector_t mb_offset = sizeof(struct r5l_meta_block);
2021 	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2022 	struct page *page;
2023 	struct r5l_payload_data_parity *payload;
2024 	struct r5l_payload_flush *payload_flush;
2025 
2026 	page = alloc_page(GFP_KERNEL);
2027 	if (!page)
2028 		return -ENOMEM;
2029 
2030 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2031 		payload = (void *)mb + mb_offset;
2032 		payload_flush = (void *)mb + mb_offset;
2033 
2034 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2035 			if (r5l_recovery_verify_data_checksum(
2036 				    log, ctx, page, log_offset,
2037 				    payload->checksum[0]) < 0)
2038 				goto mismatch;
2039 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2040 			if (r5l_recovery_verify_data_checksum(
2041 				    log, ctx, page, log_offset,
2042 				    payload->checksum[0]) < 0)
2043 				goto mismatch;
2044 			if (conf->max_degraded == 2 && /* q for RAID 6 */
2045 			    r5l_recovery_verify_data_checksum(
2046 				    log, ctx, page,
2047 				    r5l_ring_add(log, log_offset,
2048 						 BLOCK_SECTORS),
2049 				    payload->checksum[1]) < 0)
2050 				goto mismatch;
2051 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2052 			/* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2053 		} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2054 			goto mismatch;
2055 
2056 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2057 			mb_offset += sizeof(struct r5l_payload_flush) +
2058 				le32_to_cpu(payload_flush->size);
2059 		} else {
2060 			/* DATA or PARITY payload */
2061 			log_offset = r5l_ring_add(log, log_offset,
2062 						  le32_to_cpu(payload->size));
2063 			mb_offset += sizeof(struct r5l_payload_data_parity) +
2064 				sizeof(__le32) *
2065 				(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2066 		}
2067 
2068 	}
2069 
2070 	put_page(page);
2071 	return 0;
2072 
2073 mismatch:
2074 	put_page(page);
2075 	return -EINVAL;
2076 }
2077 
2078 /*
2079  * Analyze all data/parity pages in one meta block
2080  * Returns:
2081  * 0 for success
2082  * -EINVAL for unknown playload type
2083  * -EAGAIN for checksum mismatch of data page
2084  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2085  */
2086 static int
2087 r5c_recovery_analyze_meta_block(struct r5l_log *log,
2088 				struct r5l_recovery_ctx *ctx,
2089 				struct list_head *cached_stripe_list)
2090 {
2091 	struct mddev *mddev = log->rdev->mddev;
2092 	struct r5conf *conf = mddev->private;
2093 	struct r5l_meta_block *mb;
2094 	struct r5l_payload_data_parity *payload;
2095 	struct r5l_payload_flush *payload_flush;
2096 	int mb_offset;
2097 	sector_t log_offset;
2098 	sector_t stripe_sect;
2099 	struct stripe_head *sh;
2100 	int ret;
2101 
2102 	/*
2103 	 * for mismatch in data blocks, we will drop all data in this mb, but
2104 	 * we will still read next mb for other data with FLUSH flag, as
2105 	 * io_unit could finish out of order.
2106 	 */
2107 	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2108 	if (ret == -EINVAL)
2109 		return -EAGAIN;
2110 	else if (ret)
2111 		return ret;   /* -ENOMEM duo to alloc_page() failed */
2112 
2113 	mb = page_address(ctx->meta_page);
2114 	mb_offset = sizeof(struct r5l_meta_block);
2115 	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2116 
2117 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2118 		int dd;
2119 
2120 		payload = (void *)mb + mb_offset;
2121 		payload_flush = (void *)mb + mb_offset;
2122 
2123 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2124 			int i, count;
2125 
2126 			count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2127 			for (i = 0; i < count; ++i) {
2128 				stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2129 				sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2130 								stripe_sect);
2131 				if (sh) {
2132 					WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2133 					r5l_recovery_reset_stripe(sh);
2134 					list_del_init(&sh->lru);
2135 					raid5_release_stripe(sh);
2136 				}
2137 			}
2138 
2139 			mb_offset += sizeof(struct r5l_payload_flush) +
2140 				le32_to_cpu(payload_flush->size);
2141 			continue;
2142 		}
2143 
2144 		/* DATA or PARITY payload */
2145 		stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2146 			raid5_compute_sector(
2147 				conf, le64_to_cpu(payload->location), 0, &dd,
2148 				NULL)
2149 			: le64_to_cpu(payload->location);
2150 
2151 		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2152 						stripe_sect);
2153 
2154 		if (!sh) {
2155 			sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2156 			/*
2157 			 * cannot get stripe from raid5_get_active_stripe
2158 			 * try replay some stripes
2159 			 */
2160 			if (!sh) {
2161 				r5c_recovery_replay_stripes(
2162 					cached_stripe_list, ctx);
2163 				sh = r5c_recovery_alloc_stripe(
2164 					conf, stripe_sect, 1);
2165 			}
2166 			if (!sh) {
2167 				int new_size = conf->min_nr_stripes * 2;
2168 				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2169 					mdname(mddev),
2170 					new_size);
2171 				ret = raid5_set_cache_size(mddev, new_size);
2172 				if (conf->min_nr_stripes <= new_size / 2) {
2173 					pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2174 						mdname(mddev),
2175 						ret,
2176 						new_size,
2177 						conf->min_nr_stripes,
2178 						conf->max_nr_stripes);
2179 					return -ENOMEM;
2180 				}
2181 				sh = r5c_recovery_alloc_stripe(
2182 					conf, stripe_sect, 0);
2183 			}
2184 			if (!sh) {
2185 				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2186 					mdname(mddev));
2187 				return -ENOMEM;
2188 			}
2189 			list_add_tail(&sh->lru, cached_stripe_list);
2190 		}
2191 
2192 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2193 			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2194 			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2195 				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2196 				list_move_tail(&sh->lru, cached_stripe_list);
2197 			}
2198 			r5l_recovery_load_data(log, sh, ctx, payload,
2199 					       log_offset);
2200 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2201 			r5l_recovery_load_parity(log, sh, ctx, payload,
2202 						 log_offset);
2203 		else
2204 			return -EINVAL;
2205 
2206 		log_offset = r5l_ring_add(log, log_offset,
2207 					  le32_to_cpu(payload->size));
2208 
2209 		mb_offset += sizeof(struct r5l_payload_data_parity) +
2210 			sizeof(__le32) *
2211 			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2212 	}
2213 
2214 	return 0;
2215 }
2216 
2217 /*
2218  * Load the stripe into cache. The stripe will be written out later by
2219  * the stripe cache state machine.
2220  */
2221 static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2222 					 struct stripe_head *sh)
2223 {
2224 	struct r5dev *dev;
2225 	int i;
2226 
2227 	for (i = sh->disks; i--; ) {
2228 		dev = sh->dev + i;
2229 		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2230 			set_bit(R5_InJournal, &dev->flags);
2231 			set_bit(R5_UPTODATE, &dev->flags);
2232 		}
2233 	}
2234 }
2235 
2236 /*
2237  * Scan through the log for all to-be-flushed data
2238  *
2239  * For stripes with data and parity, namely Data-Parity stripe
2240  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2241  *
2242  * For stripes with only data, namely Data-Only stripe
2243  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2244  *
2245  * For a stripe, if we see data after parity, we should discard all previous
2246  * data and parity for this stripe, as these data are already flushed to
2247  * the array.
2248  *
2249  * At the end of the scan, we return the new journal_tail, which points to
2250  * first data-only stripe on the journal device, or next invalid meta block.
2251  */
2252 static int r5c_recovery_flush_log(struct r5l_log *log,
2253 				  struct r5l_recovery_ctx *ctx)
2254 {
2255 	struct stripe_head *sh;
2256 	int ret = 0;
2257 
2258 	/* scan through the log */
2259 	while (1) {
2260 		if (r5l_recovery_read_meta_block(log, ctx))
2261 			break;
2262 
2263 		ret = r5c_recovery_analyze_meta_block(log, ctx,
2264 						      &ctx->cached_list);
2265 		/*
2266 		 * -EAGAIN means mismatch in data block, in this case, we still
2267 		 * try scan the next metablock
2268 		 */
2269 		if (ret && ret != -EAGAIN)
2270 			break;   /* ret == -EINVAL or -ENOMEM */
2271 		ctx->seq++;
2272 		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2273 	}
2274 
2275 	if (ret == -ENOMEM) {
2276 		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2277 		return ret;
2278 	}
2279 
2280 	/* replay data-parity stripes */
2281 	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2282 
2283 	/* load data-only stripes to stripe cache */
2284 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2285 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2286 		r5c_recovery_load_one_stripe(log, sh);
2287 		ctx->data_only_stripes++;
2288 	}
2289 
2290 	return 0;
2291 }
2292 
2293 /*
2294  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2295  * log will start here. but we can't let superblock point to last valid
2296  * meta block. The log might looks like:
2297  * | meta 1| meta 2| meta 3|
2298  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2299  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2300  * happens again, new recovery will start from meta 1. Since meta 2n is
2301  * valid now, recovery will think meta 3 is valid, which is wrong.
2302  * The solution is we create a new meta in meta2 with its seq == meta
2303  * 1's seq + 10000 and let superblock points to meta2. The same recovery
2304  * will not think meta 3 is a valid meta, because its seq doesn't match
2305  */
2306 
2307 /*
2308  * Before recovery, the log looks like the following
2309  *
2310  *   ---------------------------------------------
2311  *   |           valid log        | invalid log  |
2312  *   ---------------------------------------------
2313  *   ^
2314  *   |- log->last_checkpoint
2315  *   |- log->last_cp_seq
2316  *
2317  * Now we scan through the log until we see invalid entry
2318  *
2319  *   ---------------------------------------------
2320  *   |           valid log        | invalid log  |
2321  *   ---------------------------------------------
2322  *   ^                            ^
2323  *   |- log->last_checkpoint      |- ctx->pos
2324  *   |- log->last_cp_seq          |- ctx->seq
2325  *
2326  * From this point, we need to increase seq number by 10 to avoid
2327  * confusing next recovery.
2328  *
2329  *   ---------------------------------------------
2330  *   |           valid log        | invalid log  |
2331  *   ---------------------------------------------
2332  *   ^                              ^
2333  *   |- log->last_checkpoint        |- ctx->pos+1
2334  *   |- log->last_cp_seq            |- ctx->seq+10001
2335  *
2336  * However, it is not safe to start the state machine yet, because data only
2337  * parities are not yet secured in RAID. To save these data only parities, we
2338  * rewrite them from seq+11.
2339  *
2340  *   -----------------------------------------------------------------
2341  *   |           valid log        | data only stripes | invalid log  |
2342  *   -----------------------------------------------------------------
2343  *   ^                                                ^
2344  *   |- log->last_checkpoint                          |- ctx->pos+n
2345  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2346  *
2347  * If failure happens again during this process, the recovery can safe start
2348  * again from log->last_checkpoint.
2349  *
2350  * Once data only stripes are rewritten to journal, we move log_tail
2351  *
2352  *   -----------------------------------------------------------------
2353  *   |     old log        |    data only stripes    | invalid log  |
2354  *   -----------------------------------------------------------------
2355  *                        ^                         ^
2356  *                        |- log->last_checkpoint   |- ctx->pos+n
2357  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2358  *
2359  * Then we can safely start the state machine. If failure happens from this
2360  * point on, the recovery will start from new log->last_checkpoint.
2361  */
2362 static int
2363 r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2364 				       struct r5l_recovery_ctx *ctx)
2365 {
2366 	struct stripe_head *sh;
2367 	struct mddev *mddev = log->rdev->mddev;
2368 	struct page *page;
2369 	sector_t next_checkpoint = MaxSector;
2370 
2371 	page = alloc_page(GFP_KERNEL);
2372 	if (!page) {
2373 		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2374 		       mdname(mddev));
2375 		return -ENOMEM;
2376 	}
2377 
2378 	WARN_ON(list_empty(&ctx->cached_list));
2379 
2380 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2381 		struct r5l_meta_block *mb;
2382 		int i;
2383 		int offset;
2384 		sector_t write_pos;
2385 
2386 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2387 		r5l_recovery_create_empty_meta_block(log, page,
2388 						     ctx->pos, ctx->seq);
2389 		mb = page_address(page);
2390 		offset = le32_to_cpu(mb->meta_size);
2391 		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2392 
2393 		for (i = sh->disks; i--; ) {
2394 			struct r5dev *dev = &sh->dev[i];
2395 			struct r5l_payload_data_parity *payload;
2396 			void *addr;
2397 
2398 			if (test_bit(R5_InJournal, &dev->flags)) {
2399 				payload = (void *)mb + offset;
2400 				payload->header.type = cpu_to_le16(
2401 					R5LOG_PAYLOAD_DATA);
2402 				payload->size = cpu_to_le32(BLOCK_SECTORS);
2403 				payload->location = cpu_to_le64(
2404 					raid5_compute_blocknr(sh, i, 0));
2405 				addr = kmap_atomic(dev->page);
2406 				payload->checksum[0] = cpu_to_le32(
2407 					crc32c_le(log->uuid_checksum, addr,
2408 						  PAGE_SIZE));
2409 				kunmap_atomic(addr);
2410 				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2411 					     dev->page, REQ_OP_WRITE, 0, false);
2412 				write_pos = r5l_ring_add(log, write_pos,
2413 							 BLOCK_SECTORS);
2414 				offset += sizeof(__le32) +
2415 					sizeof(struct r5l_payload_data_parity);
2416 
2417 			}
2418 		}
2419 		mb->meta_size = cpu_to_le32(offset);
2420 		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2421 						     mb, PAGE_SIZE));
2422 		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2423 			     REQ_OP_WRITE, REQ_SYNC | REQ_FUA, false);
2424 		sh->log_start = ctx->pos;
2425 		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2426 		atomic_inc(&log->stripe_in_journal_count);
2427 		ctx->pos = write_pos;
2428 		ctx->seq += 1;
2429 		next_checkpoint = sh->log_start;
2430 	}
2431 	log->next_checkpoint = next_checkpoint;
2432 	__free_page(page);
2433 	return 0;
2434 }
2435 
2436 static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2437 						 struct r5l_recovery_ctx *ctx)
2438 {
2439 	struct mddev *mddev = log->rdev->mddev;
2440 	struct r5conf *conf = mddev->private;
2441 	struct stripe_head *sh, *next;
2442 
2443 	if (ctx->data_only_stripes == 0)
2444 		return;
2445 
2446 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2447 
2448 	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2449 		r5c_make_stripe_write_out(sh);
2450 		set_bit(STRIPE_HANDLE, &sh->state);
2451 		list_del_init(&sh->lru);
2452 		raid5_release_stripe(sh);
2453 	}
2454 
2455 	/* reuse conf->wait_for_quiescent in recovery */
2456 	wait_event(conf->wait_for_quiescent,
2457 		   atomic_read(&conf->active_stripes) == 0);
2458 
2459 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2460 }
2461 
2462 static int r5l_recovery_log(struct r5l_log *log)
2463 {
2464 	struct mddev *mddev = log->rdev->mddev;
2465 	struct r5l_recovery_ctx *ctx;
2466 	int ret;
2467 	sector_t pos;
2468 
2469 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2470 	if (!ctx)
2471 		return -ENOMEM;
2472 
2473 	ctx->pos = log->last_checkpoint;
2474 	ctx->seq = log->last_cp_seq;
2475 	INIT_LIST_HEAD(&ctx->cached_list);
2476 	ctx->meta_page = alloc_page(GFP_KERNEL);
2477 
2478 	if (!ctx->meta_page) {
2479 		ret =  -ENOMEM;
2480 		goto meta_page;
2481 	}
2482 
2483 	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2484 		ret = -ENOMEM;
2485 		goto ra_pool;
2486 	}
2487 
2488 	ret = r5c_recovery_flush_log(log, ctx);
2489 
2490 	if (ret)
2491 		goto error;
2492 
2493 	pos = ctx->pos;
2494 	ctx->seq += 10000;
2495 
2496 	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2497 		pr_info("md/raid:%s: starting from clean shutdown\n",
2498 			 mdname(mddev));
2499 	else
2500 		pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2501 			 mdname(mddev), ctx->data_only_stripes,
2502 			 ctx->data_parity_stripes);
2503 
2504 	if (ctx->data_only_stripes == 0) {
2505 		log->next_checkpoint = ctx->pos;
2506 		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2507 		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2508 	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2509 		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2510 		       mdname(mddev));
2511 		ret =  -EIO;
2512 		goto error;
2513 	}
2514 
2515 	log->log_start = ctx->pos;
2516 	log->seq = ctx->seq;
2517 	log->last_checkpoint = pos;
2518 	r5l_write_super(log, pos);
2519 
2520 	r5c_recovery_flush_data_only_stripes(log, ctx);
2521 	ret = 0;
2522 error:
2523 	r5l_recovery_free_ra_pool(log, ctx);
2524 ra_pool:
2525 	__free_page(ctx->meta_page);
2526 meta_page:
2527 	kfree(ctx);
2528 	return ret;
2529 }
2530 
2531 static void r5l_write_super(struct r5l_log *log, sector_t cp)
2532 {
2533 	struct mddev *mddev = log->rdev->mddev;
2534 
2535 	log->rdev->journal_tail = cp;
2536 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2537 }
2538 
2539 static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2540 {
2541 	struct r5conf *conf;
2542 	int ret;
2543 
2544 	ret = mddev_lock(mddev);
2545 	if (ret)
2546 		return ret;
2547 
2548 	conf = mddev->private;
2549 	if (!conf || !conf->log) {
2550 		mddev_unlock(mddev);
2551 		return 0;
2552 	}
2553 
2554 	switch (conf->log->r5c_journal_mode) {
2555 	case R5C_JOURNAL_MODE_WRITE_THROUGH:
2556 		ret = snprintf(
2557 			page, PAGE_SIZE, "[%s] %s\n",
2558 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2559 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2560 		break;
2561 	case R5C_JOURNAL_MODE_WRITE_BACK:
2562 		ret = snprintf(
2563 			page, PAGE_SIZE, "%s [%s]\n",
2564 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2565 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2566 		break;
2567 	default:
2568 		ret = 0;
2569 	}
2570 	mddev_unlock(mddev);
2571 	return ret;
2572 }
2573 
2574 /*
2575  * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2576  *
2577  * @mode as defined in 'enum r5c_journal_mode'.
2578  *
2579  */
2580 int r5c_journal_mode_set(struct mddev *mddev, int mode)
2581 {
2582 	struct r5conf *conf;
2583 
2584 	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2585 	    mode > R5C_JOURNAL_MODE_WRITE_BACK)
2586 		return -EINVAL;
2587 
2588 	conf = mddev->private;
2589 	if (!conf || !conf->log)
2590 		return -ENODEV;
2591 
2592 	if (raid5_calc_degraded(conf) > 0 &&
2593 	    mode == R5C_JOURNAL_MODE_WRITE_BACK)
2594 		return -EINVAL;
2595 
2596 	mddev_suspend(mddev);
2597 	conf->log->r5c_journal_mode = mode;
2598 	mddev_resume(mddev);
2599 
2600 	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2601 		 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2602 	return 0;
2603 }
2604 EXPORT_SYMBOL(r5c_journal_mode_set);
2605 
2606 static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2607 				      const char *page, size_t length)
2608 {
2609 	int mode = ARRAY_SIZE(r5c_journal_mode_str);
2610 	size_t len = length;
2611 	int ret;
2612 
2613 	if (len < 2)
2614 		return -EINVAL;
2615 
2616 	if (page[len - 1] == '\n')
2617 		len--;
2618 
2619 	while (mode--)
2620 		if (strlen(r5c_journal_mode_str[mode]) == len &&
2621 		    !strncmp(page, r5c_journal_mode_str[mode], len))
2622 			break;
2623 	ret = mddev_lock(mddev);
2624 	if (ret)
2625 		return ret;
2626 	ret = r5c_journal_mode_set(mddev, mode);
2627 	mddev_unlock(mddev);
2628 	return ret ?: length;
2629 }
2630 
2631 struct md_sysfs_entry
2632 r5c_journal_mode = __ATTR(journal_mode, 0644,
2633 			  r5c_journal_mode_show, r5c_journal_mode_store);
2634 
2635 /*
2636  * Try handle write operation in caching phase. This function should only
2637  * be called in write-back mode.
2638  *
2639  * If all outstanding writes can be handled in caching phase, returns 0
2640  * If writes requires write-out phase, call r5c_make_stripe_write_out()
2641  * and returns -EAGAIN
2642  */
2643 int r5c_try_caching_write(struct r5conf *conf,
2644 			  struct stripe_head *sh,
2645 			  struct stripe_head_state *s,
2646 			  int disks)
2647 {
2648 	struct r5l_log *log = conf->log;
2649 	int i;
2650 	struct r5dev *dev;
2651 	int to_cache = 0;
2652 	void **pslot;
2653 	sector_t tree_index;
2654 	int ret;
2655 	uintptr_t refcount;
2656 
2657 	BUG_ON(!r5c_is_writeback(log));
2658 
2659 	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2660 		/*
2661 		 * There are two different scenarios here:
2662 		 *  1. The stripe has some data cached, and it is sent to
2663 		 *     write-out phase for reclaim
2664 		 *  2. The stripe is clean, and this is the first write
2665 		 *
2666 		 * For 1, return -EAGAIN, so we continue with
2667 		 * handle_stripe_dirtying().
2668 		 *
2669 		 * For 2, set STRIPE_R5C_CACHING and continue with caching
2670 		 * write.
2671 		 */
2672 
2673 		/* case 1: anything injournal or anything in written */
2674 		if (s->injournal > 0 || s->written > 0)
2675 			return -EAGAIN;
2676 		/* case 2 */
2677 		set_bit(STRIPE_R5C_CACHING, &sh->state);
2678 	}
2679 
2680 	/*
2681 	 * When run in degraded mode, array is set to write-through mode.
2682 	 * This check helps drain pending write safely in the transition to
2683 	 * write-through mode.
2684 	 *
2685 	 * When a stripe is syncing, the write is also handled in write
2686 	 * through mode.
2687 	 */
2688 	if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2689 		r5c_make_stripe_write_out(sh);
2690 		return -EAGAIN;
2691 	}
2692 
2693 	for (i = disks; i--; ) {
2694 		dev = &sh->dev[i];
2695 		/* if non-overwrite, use writing-out phase */
2696 		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2697 		    !test_bit(R5_InJournal, &dev->flags)) {
2698 			r5c_make_stripe_write_out(sh);
2699 			return -EAGAIN;
2700 		}
2701 	}
2702 
2703 	/* if the stripe is not counted in big_stripe_tree, add it now */
2704 	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2705 	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2706 		tree_index = r5c_tree_index(conf, sh->sector);
2707 		spin_lock(&log->tree_lock);
2708 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2709 					       tree_index);
2710 		if (pslot) {
2711 			refcount = (uintptr_t)radix_tree_deref_slot_protected(
2712 				pslot, &log->tree_lock) >>
2713 				R5C_RADIX_COUNT_SHIFT;
2714 			radix_tree_replace_slot(
2715 				&log->big_stripe_tree, pslot,
2716 				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2717 		} else {
2718 			/*
2719 			 * this radix_tree_insert can fail safely, so no
2720 			 * need to call radix_tree_preload()
2721 			 */
2722 			ret = radix_tree_insert(
2723 				&log->big_stripe_tree, tree_index,
2724 				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
2725 			if (ret) {
2726 				spin_unlock(&log->tree_lock);
2727 				r5c_make_stripe_write_out(sh);
2728 				return -EAGAIN;
2729 			}
2730 		}
2731 		spin_unlock(&log->tree_lock);
2732 
2733 		/*
2734 		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2735 		 * counted in the radix tree
2736 		 */
2737 		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2738 		atomic_inc(&conf->r5c_cached_partial_stripes);
2739 	}
2740 
2741 	for (i = disks; i--; ) {
2742 		dev = &sh->dev[i];
2743 		if (dev->towrite) {
2744 			set_bit(R5_Wantwrite, &dev->flags);
2745 			set_bit(R5_Wantdrain, &dev->flags);
2746 			set_bit(R5_LOCKED, &dev->flags);
2747 			to_cache++;
2748 		}
2749 	}
2750 
2751 	if (to_cache) {
2752 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2753 		/*
2754 		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2755 		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2756 		 * r5c_handle_data_cached()
2757 		 */
2758 		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2759 	}
2760 
2761 	return 0;
2762 }
2763 
2764 /*
2765  * free extra pages (orig_page) we allocated for prexor
2766  */
2767 void r5c_release_extra_page(struct stripe_head *sh)
2768 {
2769 	struct r5conf *conf = sh->raid_conf;
2770 	int i;
2771 	bool using_disk_info_extra_page;
2772 
2773 	using_disk_info_extra_page =
2774 		sh->dev[0].orig_page == conf->disks[0].extra_page;
2775 
2776 	for (i = sh->disks; i--; )
2777 		if (sh->dev[i].page != sh->dev[i].orig_page) {
2778 			struct page *p = sh->dev[i].orig_page;
2779 
2780 			sh->dev[i].orig_page = sh->dev[i].page;
2781 			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2782 
2783 			if (!using_disk_info_extra_page)
2784 				put_page(p);
2785 		}
2786 
2787 	if (using_disk_info_extra_page) {
2788 		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2789 		md_wakeup_thread(conf->mddev->thread);
2790 	}
2791 }
2792 
2793 void r5c_use_extra_page(struct stripe_head *sh)
2794 {
2795 	struct r5conf *conf = sh->raid_conf;
2796 	int i;
2797 	struct r5dev *dev;
2798 
2799 	for (i = sh->disks; i--; ) {
2800 		dev = &sh->dev[i];
2801 		if (dev->orig_page != dev->page)
2802 			put_page(dev->orig_page);
2803 		dev->orig_page = conf->disks[i].extra_page;
2804 	}
2805 }
2806 
2807 /*
2808  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2809  * stripe is committed to RAID disks.
2810  */
2811 void r5c_finish_stripe_write_out(struct r5conf *conf,
2812 				 struct stripe_head *sh,
2813 				 struct stripe_head_state *s)
2814 {
2815 	struct r5l_log *log = conf->log;
2816 	int i;
2817 	int do_wakeup = 0;
2818 	sector_t tree_index;
2819 	void **pslot;
2820 	uintptr_t refcount;
2821 
2822 	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2823 		return;
2824 
2825 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2826 	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2827 
2828 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2829 		return;
2830 
2831 	for (i = sh->disks; i--; ) {
2832 		clear_bit(R5_InJournal, &sh->dev[i].flags);
2833 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2834 			do_wakeup = 1;
2835 	}
2836 
2837 	/*
2838 	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2839 	 * We updated R5_InJournal, so we also update s->injournal.
2840 	 */
2841 	s->injournal = 0;
2842 
2843 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2844 		if (atomic_dec_and_test(&conf->pending_full_writes))
2845 			md_wakeup_thread(conf->mddev->thread);
2846 
2847 	if (do_wakeup)
2848 		wake_up(&conf->wait_for_overlap);
2849 
2850 	spin_lock_irq(&log->stripe_in_journal_lock);
2851 	list_del_init(&sh->r5c);
2852 	spin_unlock_irq(&log->stripe_in_journal_lock);
2853 	sh->log_start = MaxSector;
2854 
2855 	atomic_dec(&log->stripe_in_journal_count);
2856 	r5c_update_log_state(log);
2857 
2858 	/* stop counting this stripe in big_stripe_tree */
2859 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2860 	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2861 		tree_index = r5c_tree_index(conf, sh->sector);
2862 		spin_lock(&log->tree_lock);
2863 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2864 					       tree_index);
2865 		BUG_ON(pslot == NULL);
2866 		refcount = (uintptr_t)radix_tree_deref_slot_protected(
2867 			pslot, &log->tree_lock) >>
2868 			R5C_RADIX_COUNT_SHIFT;
2869 		if (refcount == 1)
2870 			radix_tree_delete(&log->big_stripe_tree, tree_index);
2871 		else
2872 			radix_tree_replace_slot(
2873 				&log->big_stripe_tree, pslot,
2874 				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2875 		spin_unlock(&log->tree_lock);
2876 	}
2877 
2878 	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2879 		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2880 		atomic_dec(&conf->r5c_flushing_partial_stripes);
2881 		atomic_dec(&conf->r5c_cached_partial_stripes);
2882 	}
2883 
2884 	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2885 		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2886 		atomic_dec(&conf->r5c_flushing_full_stripes);
2887 		atomic_dec(&conf->r5c_cached_full_stripes);
2888 	}
2889 
2890 	r5l_append_flush_payload(log, sh->sector);
2891 	/* stripe is flused to raid disks, we can do resync now */
2892 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2893 		set_bit(STRIPE_HANDLE, &sh->state);
2894 }
2895 
2896 int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2897 {
2898 	struct r5conf *conf = sh->raid_conf;
2899 	int pages = 0;
2900 	int reserve;
2901 	int i;
2902 	int ret = 0;
2903 
2904 	BUG_ON(!log);
2905 
2906 	for (i = 0; i < sh->disks; i++) {
2907 		void *addr;
2908 
2909 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2910 			continue;
2911 		addr = kmap_atomic(sh->dev[i].page);
2912 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2913 						    addr, PAGE_SIZE);
2914 		kunmap_atomic(addr);
2915 		pages++;
2916 	}
2917 	WARN_ON(pages == 0);
2918 
2919 	/*
2920 	 * The stripe must enter state machine again to call endio, so
2921 	 * don't delay.
2922 	 */
2923 	clear_bit(STRIPE_DELAYED, &sh->state);
2924 	atomic_inc(&sh->count);
2925 
2926 	mutex_lock(&log->io_mutex);
2927 	/* meta + data */
2928 	reserve = (1 + pages) << (PAGE_SHIFT - 9);
2929 
2930 	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2931 	    sh->log_start == MaxSector)
2932 		r5l_add_no_space_stripe(log, sh);
2933 	else if (!r5l_has_free_space(log, reserve)) {
2934 		if (sh->log_start == log->last_checkpoint)
2935 			BUG();
2936 		else
2937 			r5l_add_no_space_stripe(log, sh);
2938 	} else {
2939 		ret = r5l_log_stripe(log, sh, pages, 0);
2940 		if (ret) {
2941 			spin_lock_irq(&log->io_list_lock);
2942 			list_add_tail(&sh->log_list, &log->no_mem_stripes);
2943 			spin_unlock_irq(&log->io_list_lock);
2944 		}
2945 	}
2946 
2947 	mutex_unlock(&log->io_mutex);
2948 	return 0;
2949 }
2950 
2951 /* check whether this big stripe is in write back cache. */
2952 bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2953 {
2954 	struct r5l_log *log = conf->log;
2955 	sector_t tree_index;
2956 	void *slot;
2957 
2958 	if (!log)
2959 		return false;
2960 
2961 	WARN_ON_ONCE(!rcu_read_lock_held());
2962 	tree_index = r5c_tree_index(conf, sect);
2963 	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2964 	return slot != NULL;
2965 }
2966 
2967 static int r5l_load_log(struct r5l_log *log)
2968 {
2969 	struct md_rdev *rdev = log->rdev;
2970 	struct page *page;
2971 	struct r5l_meta_block *mb;
2972 	sector_t cp = log->rdev->journal_tail;
2973 	u32 stored_crc, expected_crc;
2974 	bool create_super = false;
2975 	int ret = 0;
2976 
2977 	/* Make sure it's valid */
2978 	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2979 		cp = 0;
2980 	page = alloc_page(GFP_KERNEL);
2981 	if (!page)
2982 		return -ENOMEM;
2983 
2984 	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2985 		ret = -EIO;
2986 		goto ioerr;
2987 	}
2988 	mb = page_address(page);
2989 
2990 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2991 	    mb->version != R5LOG_VERSION) {
2992 		create_super = true;
2993 		goto create;
2994 	}
2995 	stored_crc = le32_to_cpu(mb->checksum);
2996 	mb->checksum = 0;
2997 	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2998 	if (stored_crc != expected_crc) {
2999 		create_super = true;
3000 		goto create;
3001 	}
3002 	if (le64_to_cpu(mb->position) != cp) {
3003 		create_super = true;
3004 		goto create;
3005 	}
3006 create:
3007 	if (create_super) {
3008 		log->last_cp_seq = prandom_u32();
3009 		cp = 0;
3010 		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3011 		/*
3012 		 * Make sure super points to correct address. Log might have
3013 		 * data very soon. If super hasn't correct log tail address,
3014 		 * recovery can't find the log
3015 		 */
3016 		r5l_write_super(log, cp);
3017 	} else
3018 		log->last_cp_seq = le64_to_cpu(mb->seq);
3019 
3020 	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3021 	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3022 	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3023 		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3024 	log->last_checkpoint = cp;
3025 
3026 	__free_page(page);
3027 
3028 	if (create_super) {
3029 		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3030 		log->seq = log->last_cp_seq + 1;
3031 		log->next_checkpoint = cp;
3032 	} else
3033 		ret = r5l_recovery_log(log);
3034 
3035 	r5c_update_log_state(log);
3036 	return ret;
3037 ioerr:
3038 	__free_page(page);
3039 	return ret;
3040 }
3041 
3042 int r5l_start(struct r5l_log *log)
3043 {
3044 	int ret;
3045 
3046 	if (!log)
3047 		return 0;
3048 
3049 	ret = r5l_load_log(log);
3050 	if (ret) {
3051 		struct mddev *mddev = log->rdev->mddev;
3052 		struct r5conf *conf = mddev->private;
3053 
3054 		r5l_exit_log(conf);
3055 	}
3056 	return ret;
3057 }
3058 
3059 void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3060 {
3061 	struct r5conf *conf = mddev->private;
3062 	struct r5l_log *log = conf->log;
3063 
3064 	if (!log)
3065 		return;
3066 
3067 	if ((raid5_calc_degraded(conf) > 0 ||
3068 	     test_bit(Journal, &rdev->flags)) &&
3069 	    conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3070 		schedule_work(&log->disable_writeback_work);
3071 }
3072 
3073 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3074 {
3075 	struct request_queue *q = bdev_get_queue(rdev->bdev);
3076 	struct r5l_log *log;
3077 	char b[BDEVNAME_SIZE];
3078 	int ret;
3079 
3080 	pr_debug("md/raid:%s: using device %s as journal\n",
3081 		 mdname(conf->mddev), bdevname(rdev->bdev, b));
3082 
3083 	if (PAGE_SIZE != 4096)
3084 		return -EINVAL;
3085 
3086 	/*
3087 	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3088 	 * raid_disks r5l_payload_data_parity.
3089 	 *
3090 	 * Write journal and cache does not work for very big array
3091 	 * (raid_disks > 203)
3092 	 */
3093 	if (sizeof(struct r5l_meta_block) +
3094 	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3095 	     conf->raid_disks) > PAGE_SIZE) {
3096 		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3097 		       mdname(conf->mddev), conf->raid_disks);
3098 		return -EINVAL;
3099 	}
3100 
3101 	log = kzalloc(sizeof(*log), GFP_KERNEL);
3102 	if (!log)
3103 		return -ENOMEM;
3104 	log->rdev = rdev;
3105 
3106 	log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
3107 
3108 	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3109 				       sizeof(rdev->mddev->uuid));
3110 
3111 	mutex_init(&log->io_mutex);
3112 
3113 	spin_lock_init(&log->io_list_lock);
3114 	INIT_LIST_HEAD(&log->running_ios);
3115 	INIT_LIST_HEAD(&log->io_end_ios);
3116 	INIT_LIST_HEAD(&log->flushing_ios);
3117 	INIT_LIST_HEAD(&log->finished_ios);
3118 	bio_init(&log->flush_bio, NULL, 0);
3119 
3120 	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3121 	if (!log->io_kc)
3122 		goto io_kc;
3123 
3124 	ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3125 	if (ret)
3126 		goto io_pool;
3127 
3128 	ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3129 	if (ret)
3130 		goto io_bs;
3131 
3132 	ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3133 	if (ret)
3134 		goto out_mempool;
3135 
3136 	spin_lock_init(&log->tree_lock);
3137 	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3138 
3139 	log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3140 						 log->rdev->mddev, "reclaim");
3141 	if (!log->reclaim_thread)
3142 		goto reclaim_thread;
3143 	log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3144 
3145 	init_waitqueue_head(&log->iounit_wait);
3146 
3147 	INIT_LIST_HEAD(&log->no_mem_stripes);
3148 
3149 	INIT_LIST_HEAD(&log->no_space_stripes);
3150 	spin_lock_init(&log->no_space_stripes_lock);
3151 
3152 	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3153 	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3154 
3155 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3156 	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3157 	spin_lock_init(&log->stripe_in_journal_lock);
3158 	atomic_set(&log->stripe_in_journal_count, 0);
3159 
3160 	rcu_assign_pointer(conf->log, log);
3161 
3162 	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3163 	return 0;
3164 
3165 reclaim_thread:
3166 	mempool_exit(&log->meta_pool);
3167 out_mempool:
3168 	bioset_exit(&log->bs);
3169 io_bs:
3170 	mempool_exit(&log->io_pool);
3171 io_pool:
3172 	kmem_cache_destroy(log->io_kc);
3173 io_kc:
3174 	kfree(log);
3175 	return -EINVAL;
3176 }
3177 
3178 void r5l_exit_log(struct r5conf *conf)
3179 {
3180 	struct r5l_log *log = conf->log;
3181 
3182 	conf->log = NULL;
3183 	synchronize_rcu();
3184 
3185 	/* Ensure disable_writeback_work wakes up and exits */
3186 	wake_up(&conf->mddev->sb_wait);
3187 	flush_work(&log->disable_writeback_work);
3188 	md_unregister_thread(&log->reclaim_thread);
3189 	mempool_exit(&log->meta_pool);
3190 	bioset_exit(&log->bs);
3191 	mempool_exit(&log->io_pool);
3192 	kmem_cache_destroy(log->io_kc);
3193 	kfree(log);
3194 }
3195