xref: /linux/drivers/md/raid5-cache.c (revision bb3338d3474e0329918fda9dae2c52751731eb58)
1f6bed0efSShaohua Li /*
2f6bed0efSShaohua Li  * Copyright (C) 2015 Shaohua Li <shli@fb.com>
3b4c625c6SSong Liu  * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
4f6bed0efSShaohua Li  *
5f6bed0efSShaohua Li  * This program is free software; you can redistribute it and/or modify it
6f6bed0efSShaohua Li  * under the terms and conditions of the GNU General Public License,
7f6bed0efSShaohua Li  * version 2, as published by the Free Software Foundation.
8f6bed0efSShaohua Li  *
9f6bed0efSShaohua Li  * This program is distributed in the hope it will be useful, but WITHOUT
10f6bed0efSShaohua Li  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11f6bed0efSShaohua Li  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12f6bed0efSShaohua Li  * more details.
13f6bed0efSShaohua Li  *
14f6bed0efSShaohua Li  */
15f6bed0efSShaohua Li #include <linux/kernel.h>
16f6bed0efSShaohua Li #include <linux/wait.h>
17f6bed0efSShaohua Li #include <linux/blkdev.h>
18f6bed0efSShaohua Li #include <linux/slab.h>
19f6bed0efSShaohua Li #include <linux/raid/md_p.h>
205cb2fbd6SShaohua Li #include <linux/crc32c.h>
21f6bed0efSShaohua Li #include <linux/random.h>
22ce1ccd07SShaohua Li #include <linux/kthread.h>
2303b047f4SSong Liu #include <linux/types.h>
24f6bed0efSShaohua Li #include "md.h"
25f6bed0efSShaohua Li #include "raid5.h"
261e6d690bSSong Liu #include "bitmap.h"
27f6bed0efSShaohua Li 
28f6bed0efSShaohua Li /*
29f6bed0efSShaohua Li  * metadata/data stored in disk with 4k size unit (a block) regardless
30f6bed0efSShaohua Li  * underneath hardware sector size. only works with PAGE_SIZE == 4096
31f6bed0efSShaohua Li  */
32f6bed0efSShaohua Li #define BLOCK_SECTORS (8)
33effe6ee7SSong Liu #define BLOCK_SECTOR_SHIFT (3)
34f6bed0efSShaohua Li 
350576b1c6SShaohua Li /*
36a39f7afdSSong Liu  * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
37a39f7afdSSong Liu  *
38a39f7afdSSong Liu  * In write through mode, the reclaim runs every log->max_free_space.
39a39f7afdSSong Liu  * This can prevent the recovery scans for too long
400576b1c6SShaohua Li  */
410576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
420576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
430576b1c6SShaohua Li 
44a39f7afdSSong Liu /* wake up reclaim thread periodically */
45a39f7afdSSong Liu #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
46a39f7afdSSong Liu /* start flush with these full stripes */
4784890c03SShaohua Li #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
48a39f7afdSSong Liu /* reclaim stripes in groups */
49a39f7afdSSong Liu #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
50a39f7afdSSong Liu 
51c38d29b3SChristoph Hellwig /*
52c38d29b3SChristoph Hellwig  * We only need 2 bios per I/O unit to make progress, but ensure we
53c38d29b3SChristoph Hellwig  * have a few more available to not get too tight.
54c38d29b3SChristoph Hellwig  */
55c38d29b3SChristoph Hellwig #define R5L_POOL_SIZE	4
56c38d29b3SChristoph Hellwig 
572c7da14bSSong Liu static char *r5c_journal_mode_str[] = {"write-through",
582c7da14bSSong Liu 				       "write-back"};
592ded3703SSong Liu /*
602ded3703SSong Liu  * raid5 cache state machine
612ded3703SSong Liu  *
629b69173eSJackieLiu  * With the RAID cache, each stripe works in two phases:
632ded3703SSong Liu  *	- caching phase
642ded3703SSong Liu  *	- writing-out phase
652ded3703SSong Liu  *
662ded3703SSong Liu  * These two phases are controlled by bit STRIPE_R5C_CACHING:
672ded3703SSong Liu  *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
682ded3703SSong Liu  *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
692ded3703SSong Liu  *
702ded3703SSong Liu  * When there is no journal, or the journal is in write-through mode,
712ded3703SSong Liu  * the stripe is always in writing-out phase.
722ded3703SSong Liu  *
732ded3703SSong Liu  * For write-back journal, the stripe is sent to caching phase on write
742ded3703SSong Liu  * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
752ded3703SSong Liu  * the write-out phase by clearing STRIPE_R5C_CACHING.
762ded3703SSong Liu  *
772ded3703SSong Liu  * Stripes in caching phase do not write the raid disks. Instead, all
782ded3703SSong Liu  * writes are committed from the log device. Therefore, a stripe in
792ded3703SSong Liu  * caching phase handles writes as:
802ded3703SSong Liu  *	- write to log device
812ded3703SSong Liu  *	- return IO
822ded3703SSong Liu  *
832ded3703SSong Liu  * Stripes in writing-out phase handle writes as:
842ded3703SSong Liu  *	- calculate parity
852ded3703SSong Liu  *	- write pending data and parity to journal
862ded3703SSong Liu  *	- write data and parity to raid disks
872ded3703SSong Liu  *	- return IO for pending writes
882ded3703SSong Liu  */
892ded3703SSong Liu 
90f6bed0efSShaohua Li struct r5l_log {
91f6bed0efSShaohua Li 	struct md_rdev *rdev;
92f6bed0efSShaohua Li 
93f6bed0efSShaohua Li 	u32 uuid_checksum;
94f6bed0efSShaohua Li 
95f6bed0efSShaohua Li 	sector_t device_size;		/* log device size, round to
96f6bed0efSShaohua Li 					 * BLOCK_SECTORS */
970576b1c6SShaohua Li 	sector_t max_free_space;	/* reclaim run if free space is at
980576b1c6SShaohua Li 					 * this size */
99f6bed0efSShaohua Li 
100f6bed0efSShaohua Li 	sector_t last_checkpoint;	/* log tail. where recovery scan
101f6bed0efSShaohua Li 					 * starts from */
102f6bed0efSShaohua Li 	u64 last_cp_seq;		/* log tail sequence */
103f6bed0efSShaohua Li 
104f6bed0efSShaohua Li 	sector_t log_start;		/* log head. where new data appends */
105f6bed0efSShaohua Li 	u64 seq;			/* log head sequence */
106f6bed0efSShaohua Li 
10717036461SChristoph Hellwig 	sector_t next_checkpoint;
10817036461SChristoph Hellwig 
109f6bed0efSShaohua Li 	struct mutex io_mutex;
110f6bed0efSShaohua Li 	struct r5l_io_unit *current_io;	/* current io_unit accepting new data */
111f6bed0efSShaohua Li 
112f6bed0efSShaohua Li 	spinlock_t io_list_lock;
113f6bed0efSShaohua Li 	struct list_head running_ios;	/* io_units which are still running,
114f6bed0efSShaohua Li 					 * and have not yet been completely
115f6bed0efSShaohua Li 					 * written to the log */
116f6bed0efSShaohua Li 	struct list_head io_end_ios;	/* io_units which have been completely
117f6bed0efSShaohua Li 					 * written to the log but not yet written
118f6bed0efSShaohua Li 					 * to the RAID */
119a8c34f91SShaohua Li 	struct list_head flushing_ios;	/* io_units which are waiting for log
120a8c34f91SShaohua Li 					 * cache flush */
12104732f74SChristoph Hellwig 	struct list_head finished_ios;	/* io_units which settle down in log disk */
122a8c34f91SShaohua Li 	struct bio flush_bio;
123f6bed0efSShaohua Li 
1245036c390SChristoph Hellwig 	struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
1255036c390SChristoph Hellwig 
126f6bed0efSShaohua Li 	struct kmem_cache *io_kc;
1275036c390SChristoph Hellwig 	mempool_t *io_pool;
128c38d29b3SChristoph Hellwig 	struct bio_set *bs;
129e8deb638SChristoph Hellwig 	mempool_t *meta_pool;
130f6bed0efSShaohua Li 
1310576b1c6SShaohua Li 	struct md_thread *reclaim_thread;
1320576b1c6SShaohua Li 	unsigned long reclaim_target;	/* number of space that need to be
1330576b1c6SShaohua Li 					 * reclaimed.  if it's 0, reclaim spaces
1340576b1c6SShaohua Li 					 * used by io_units which are in
1350576b1c6SShaohua Li 					 * IO_UNIT_STRIPE_END state (eg, reclaim
1360576b1c6SShaohua Li 					 * dones't wait for specific io_unit
1370576b1c6SShaohua Li 					 * switching to IO_UNIT_STRIPE_END
1380576b1c6SShaohua Li 					 * state) */
1390fd22b45SShaohua Li 	wait_queue_head_t iounit_wait;
1400576b1c6SShaohua Li 
141f6bed0efSShaohua Li 	struct list_head no_space_stripes; /* pending stripes, log has no space */
142f6bed0efSShaohua Li 	spinlock_t no_space_stripes_lock;
14356fef7c6SChristoph Hellwig 
14456fef7c6SChristoph Hellwig 	bool need_cache_flush;
1452ded3703SSong Liu 
1462ded3703SSong Liu 	/* for r5c_cache */
1472ded3703SSong Liu 	enum r5c_journal_mode r5c_journal_mode;
148a39f7afdSSong Liu 
149a39f7afdSSong Liu 	/* all stripes in r5cache, in the order of seq at sh->log_start */
150a39f7afdSSong Liu 	struct list_head stripe_in_journal_list;
151a39f7afdSSong Liu 
152a39f7afdSSong Liu 	spinlock_t stripe_in_journal_lock;
153a39f7afdSSong Liu 	atomic_t stripe_in_journal_count;
1543bddb7f8SSong Liu 
1553bddb7f8SSong Liu 	/* to submit async io_units, to fulfill ordering of flush */
1563bddb7f8SSong Liu 	struct work_struct deferred_io_work;
1572e38a37fSSong Liu 	/* to disable write back during in degraded mode */
1582e38a37fSSong Liu 	struct work_struct disable_writeback_work;
15903b047f4SSong Liu 
16003b047f4SSong Liu 	/* to for chunk_aligned_read in writeback mode, details below */
16103b047f4SSong Liu 	spinlock_t tree_lock;
16203b047f4SSong Liu 	struct radix_tree_root big_stripe_tree;
163f6bed0efSShaohua Li };
164f6bed0efSShaohua Li 
165f6bed0efSShaohua Li /*
16603b047f4SSong Liu  * Enable chunk_aligned_read() with write back cache.
16703b047f4SSong Liu  *
16803b047f4SSong Liu  * Each chunk may contain more than one stripe (for example, a 256kB
16903b047f4SSong Liu  * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
17003b047f4SSong Liu  * chunk_aligned_read, these stripes are grouped into one "big_stripe".
17103b047f4SSong Liu  * For each big_stripe, we count how many stripes of this big_stripe
17203b047f4SSong Liu  * are in the write back cache. These data are tracked in a radix tree
17303b047f4SSong Liu  * (big_stripe_tree). We use radix_tree item pointer as the counter.
17403b047f4SSong Liu  * r5c_tree_index() is used to calculate keys for the radix tree.
17503b047f4SSong Liu  *
17603b047f4SSong Liu  * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
17703b047f4SSong Liu  * big_stripe of each chunk in the tree. If this big_stripe is in the
17803b047f4SSong Liu  * tree, chunk_aligned_read() aborts. This look up is protected by
17903b047f4SSong Liu  * rcu_read_lock().
18003b047f4SSong Liu  *
18103b047f4SSong Liu  * It is necessary to remember whether a stripe is counted in
18203b047f4SSong Liu  * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
18303b047f4SSong Liu  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
18403b047f4SSong Liu  * two flags are set, the stripe is counted in big_stripe_tree. This
18503b047f4SSong Liu  * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
18603b047f4SSong Liu  * r5c_try_caching_write(); and moving clear_bit of
18703b047f4SSong Liu  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
18803b047f4SSong Liu  * r5c_finish_stripe_write_out().
18903b047f4SSong Liu  */
19003b047f4SSong Liu 
19103b047f4SSong Liu /*
19203b047f4SSong Liu  * radix tree requests lowest 2 bits of data pointer to be 2b'00.
19303b047f4SSong Liu  * So it is necessary to left shift the counter by 2 bits before using it
19403b047f4SSong Liu  * as data pointer of the tree.
19503b047f4SSong Liu  */
19603b047f4SSong Liu #define R5C_RADIX_COUNT_SHIFT 2
19703b047f4SSong Liu 
19803b047f4SSong Liu /*
19903b047f4SSong Liu  * calculate key for big_stripe_tree
20003b047f4SSong Liu  *
20103b047f4SSong Liu  * sect: align_bi->bi_iter.bi_sector or sh->sector
20203b047f4SSong Liu  */
20303b047f4SSong Liu static inline sector_t r5c_tree_index(struct r5conf *conf,
20403b047f4SSong Liu 				      sector_t sect)
20503b047f4SSong Liu {
20603b047f4SSong Liu 	sector_t offset;
20703b047f4SSong Liu 
20803b047f4SSong Liu 	offset = sector_div(sect, conf->chunk_sectors);
20903b047f4SSong Liu 	return sect;
21003b047f4SSong Liu }
21103b047f4SSong Liu 
21203b047f4SSong Liu /*
213f6bed0efSShaohua Li  * an IO range starts from a meta data block and end at the next meta data
214f6bed0efSShaohua Li  * block. The io unit's the meta data block tracks data/parity followed it. io
215f6bed0efSShaohua Li  * unit is written to log disk with normal write, as we always flush log disk
216f6bed0efSShaohua Li  * first and then start move data to raid disks, there is no requirement to
217f6bed0efSShaohua Li  * write io unit with FLUSH/FUA
218f6bed0efSShaohua Li  */
219f6bed0efSShaohua Li struct r5l_io_unit {
220f6bed0efSShaohua Li 	struct r5l_log *log;
221f6bed0efSShaohua Li 
222f6bed0efSShaohua Li 	struct page *meta_page;	/* store meta block */
223f6bed0efSShaohua Li 	int meta_offset;	/* current offset in meta_page */
224f6bed0efSShaohua Li 
225f6bed0efSShaohua Li 	struct bio *current_bio;/* current_bio accepting new data */
226f6bed0efSShaohua Li 
227f6bed0efSShaohua Li 	atomic_t pending_stripe;/* how many stripes not flushed to raid */
228f6bed0efSShaohua Li 	u64 seq;		/* seq number of the metablock */
229f6bed0efSShaohua Li 	sector_t log_start;	/* where the io_unit starts */
230f6bed0efSShaohua Li 	sector_t log_end;	/* where the io_unit ends */
231f6bed0efSShaohua Li 	struct list_head log_sibling; /* log->running_ios */
232f6bed0efSShaohua Li 	struct list_head stripe_list; /* stripes added to the io_unit */
233f6bed0efSShaohua Li 
234f6bed0efSShaohua Li 	int state;
2356143e2ceSChristoph Hellwig 	bool need_split_bio;
2363bddb7f8SSong Liu 	struct bio *split_bio;
2373bddb7f8SSong Liu 
2383bddb7f8SSong Liu 	unsigned int has_flush:1;      /* include flush request */
2393bddb7f8SSong Liu 	unsigned int has_fua:1;        /* include fua request */
2403bddb7f8SSong Liu 	unsigned int has_null_flush:1; /* include empty flush request */
2413bddb7f8SSong Liu 	/*
2423bddb7f8SSong Liu 	 * io isn't sent yet, flush/fua request can only be submitted till it's
2433bddb7f8SSong Liu 	 * the first IO in running_ios list
2443bddb7f8SSong Liu 	 */
2453bddb7f8SSong Liu 	unsigned int io_deferred:1;
2463bddb7f8SSong Liu 
2473bddb7f8SSong Liu 	struct bio_list flush_barriers;   /* size == 0 flush bios */
248f6bed0efSShaohua Li };
249f6bed0efSShaohua Li 
250f6bed0efSShaohua Li /* r5l_io_unit state */
251f6bed0efSShaohua Li enum r5l_io_unit_state {
252f6bed0efSShaohua Li 	IO_UNIT_RUNNING = 0,	/* accepting new IO */
253f6bed0efSShaohua Li 	IO_UNIT_IO_START = 1,	/* io_unit bio start writing to log,
254f6bed0efSShaohua Li 				 * don't accepting new bio */
255f6bed0efSShaohua Li 	IO_UNIT_IO_END = 2,	/* io_unit bio finish writing to log */
256a8c34f91SShaohua Li 	IO_UNIT_STRIPE_END = 3,	/* stripes data finished writing to raid */
257f6bed0efSShaohua Li };
258f6bed0efSShaohua Li 
2592ded3703SSong Liu bool r5c_is_writeback(struct r5l_log *log)
2602ded3703SSong Liu {
2612ded3703SSong Liu 	return (log != NULL &&
2622ded3703SSong Liu 		log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
2632ded3703SSong Liu }
2642ded3703SSong Liu 
265f6bed0efSShaohua Li static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
266f6bed0efSShaohua Li {
267f6bed0efSShaohua Li 	start += inc;
268f6bed0efSShaohua Li 	if (start >= log->device_size)
269f6bed0efSShaohua Li 		start = start - log->device_size;
270f6bed0efSShaohua Li 	return start;
271f6bed0efSShaohua Li }
272f6bed0efSShaohua Li 
273f6bed0efSShaohua Li static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
274f6bed0efSShaohua Li 				  sector_t end)
275f6bed0efSShaohua Li {
276f6bed0efSShaohua Li 	if (end >= start)
277f6bed0efSShaohua Li 		return end - start;
278f6bed0efSShaohua Li 	else
279f6bed0efSShaohua Li 		return end + log->device_size - start;
280f6bed0efSShaohua Li }
281f6bed0efSShaohua Li 
282f6bed0efSShaohua Li static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
283f6bed0efSShaohua Li {
284f6bed0efSShaohua Li 	sector_t used_size;
285f6bed0efSShaohua Li 
286f6bed0efSShaohua Li 	used_size = r5l_ring_distance(log, log->last_checkpoint,
287f6bed0efSShaohua Li 					log->log_start);
288f6bed0efSShaohua Li 
289f6bed0efSShaohua Li 	return log->device_size > used_size + size;
290f6bed0efSShaohua Li }
291f6bed0efSShaohua Li 
292f6bed0efSShaohua Li static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
293f6bed0efSShaohua Li 				    enum r5l_io_unit_state state)
294f6bed0efSShaohua Li {
295f6bed0efSShaohua Li 	if (WARN_ON(io->state >= state))
296f6bed0efSShaohua Li 		return;
297f6bed0efSShaohua Li 	io->state = state;
298f6bed0efSShaohua Li }
299f6bed0efSShaohua Li 
3001e6d690bSSong Liu static void
301bd83d0a2SNeilBrown r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
3021e6d690bSSong Liu {
3031e6d690bSSong Liu 	struct bio *wbi, *wbi2;
3041e6d690bSSong Liu 
3051e6d690bSSong Liu 	wbi = dev->written;
3061e6d690bSSong Liu 	dev->written = NULL;
3071e6d690bSSong Liu 	while (wbi && wbi->bi_iter.bi_sector <
3081e6d690bSSong Liu 	       dev->sector + STRIPE_SECTORS) {
3091e6d690bSSong Liu 		wbi2 = r5_next_bio(wbi, dev->sector);
3101e6d690bSSong Liu 		md_write_end(conf->mddev);
311bd83d0a2SNeilBrown 		bio_endio(wbi);
3121e6d690bSSong Liu 		wbi = wbi2;
3131e6d690bSSong Liu 	}
3141e6d690bSSong Liu }
3151e6d690bSSong Liu 
3161e6d690bSSong Liu void r5c_handle_cached_data_endio(struct r5conf *conf,
317bd83d0a2SNeilBrown 				  struct stripe_head *sh, int disks)
3181e6d690bSSong Liu {
3191e6d690bSSong Liu 	int i;
3201e6d690bSSong Liu 
3211e6d690bSSong Liu 	for (i = sh->disks; i--; ) {
3221e6d690bSSong Liu 		if (sh->dev[i].written) {
3231e6d690bSSong Liu 			set_bit(R5_UPTODATE, &sh->dev[i].flags);
324bd83d0a2SNeilBrown 			r5c_return_dev_pending_writes(conf, &sh->dev[i]);
3251e6d690bSSong Liu 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3261e6d690bSSong Liu 					STRIPE_SECTORS,
3271e6d690bSSong Liu 					!test_bit(STRIPE_DEGRADED, &sh->state),
3281e6d690bSSong Liu 					0);
3291e6d690bSSong Liu 		}
3301e6d690bSSong Liu 	}
3311e6d690bSSong Liu }
3321e6d690bSSong Liu 
333ff875738SArtur Paszkiewicz void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
334ff875738SArtur Paszkiewicz 
335a39f7afdSSong Liu /* Check whether we should flush some stripes to free up stripe cache */
336a39f7afdSSong Liu void r5c_check_stripe_cache_usage(struct r5conf *conf)
337a39f7afdSSong Liu {
338a39f7afdSSong Liu 	int total_cached;
339a39f7afdSSong Liu 
340a39f7afdSSong Liu 	if (!r5c_is_writeback(conf->log))
341a39f7afdSSong Liu 		return;
342a39f7afdSSong Liu 
343a39f7afdSSong Liu 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
344a39f7afdSSong Liu 		atomic_read(&conf->r5c_cached_full_stripes);
345a39f7afdSSong Liu 
346a39f7afdSSong Liu 	/*
347a39f7afdSSong Liu 	 * The following condition is true for either of the following:
348a39f7afdSSong Liu 	 *   - stripe cache pressure high:
349a39f7afdSSong Liu 	 *          total_cached > 3/4 min_nr_stripes ||
350a39f7afdSSong Liu 	 *          empty_inactive_list_nr > 0
351a39f7afdSSong Liu 	 *   - stripe cache pressure moderate:
352a39f7afdSSong Liu 	 *          total_cached > 1/2 min_nr_stripes
353a39f7afdSSong Liu 	 */
354a39f7afdSSong Liu 	if (total_cached > conf->min_nr_stripes * 1 / 2 ||
355a39f7afdSSong Liu 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
356a39f7afdSSong Liu 		r5l_wake_reclaim(conf->log, 0);
357a39f7afdSSong Liu }
358a39f7afdSSong Liu 
359a39f7afdSSong Liu /*
360a39f7afdSSong Liu  * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
361a39f7afdSSong Liu  * stripes in the cache
362a39f7afdSSong Liu  */
363a39f7afdSSong Liu void r5c_check_cached_full_stripe(struct r5conf *conf)
364a39f7afdSSong Liu {
365a39f7afdSSong Liu 	if (!r5c_is_writeback(conf->log))
366a39f7afdSSong Liu 		return;
367a39f7afdSSong Liu 
368a39f7afdSSong Liu 	/*
369a39f7afdSSong Liu 	 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
370a39f7afdSSong Liu 	 * or a full stripe (chunk size / 4k stripes).
371a39f7afdSSong Liu 	 */
372a39f7afdSSong Liu 	if (atomic_read(&conf->r5c_cached_full_stripes) >=
37384890c03SShaohua Li 	    min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
374a39f7afdSSong Liu 		conf->chunk_sectors >> STRIPE_SHIFT))
375a39f7afdSSong Liu 		r5l_wake_reclaim(conf->log, 0);
376a39f7afdSSong Liu }
377a39f7afdSSong Liu 
378a39f7afdSSong Liu /*
379a39f7afdSSong Liu  * Total log space (in sectors) needed to flush all data in cache
380a39f7afdSSong Liu  *
38139b99586SSong Liu  * To avoid deadlock due to log space, it is necessary to reserve log
38239b99586SSong Liu  * space to flush critical stripes (stripes that occupying log space near
38339b99586SSong Liu  * last_checkpoint). This function helps check how much log space is
38439b99586SSong Liu  * required to flush all cached stripes.
385a39f7afdSSong Liu  *
38639b99586SSong Liu  * To reduce log space requirements, two mechanisms are used to give cache
38739b99586SSong Liu  * flush higher priorities:
38839b99586SSong Liu  *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
38939b99586SSong Liu  *       stripes ALREADY in journal can be flushed w/o pending writes;
39039b99586SSong Liu  *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
39139b99586SSong Liu  *       can be delayed (r5l_add_no_space_stripe).
392a39f7afdSSong Liu  *
39339b99586SSong Liu  * In cache flush, the stripe goes through 1 and then 2. For a stripe that
39439b99586SSong Liu  * already passed 1, flushing it requires at most (conf->max_degraded + 1)
39539b99586SSong Liu  * pages of journal space. For stripes that has not passed 1, flushing it
39639b99586SSong Liu  * requires (conf->raid_disks + 1) pages of journal space. There are at
39739b99586SSong Liu  * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
39839b99586SSong Liu  * required to flush all cached stripes (in pages) is:
39939b99586SSong Liu  *
40039b99586SSong Liu  *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
40139b99586SSong Liu  *     (group_cnt + 1) * (raid_disks + 1)
40239b99586SSong Liu  * or
40339b99586SSong Liu  *     (stripe_in_journal_count) * (max_degraded + 1) +
40439b99586SSong Liu  *     (group_cnt + 1) * (raid_disks - max_degraded)
405a39f7afdSSong Liu  */
406a39f7afdSSong Liu static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
407a39f7afdSSong Liu {
408a39f7afdSSong Liu 	struct r5l_log *log = conf->log;
409a39f7afdSSong Liu 
410a39f7afdSSong Liu 	if (!r5c_is_writeback(log))
411a39f7afdSSong Liu 		return 0;
412a39f7afdSSong Liu 
41339b99586SSong Liu 	return BLOCK_SECTORS *
41439b99586SSong Liu 		((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
41539b99586SSong Liu 		 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
416a39f7afdSSong Liu }
417a39f7afdSSong Liu 
418a39f7afdSSong Liu /*
419a39f7afdSSong Liu  * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
420a39f7afdSSong Liu  *
421a39f7afdSSong Liu  * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
422a39f7afdSSong Liu  * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
423a39f7afdSSong Liu  * device is less than 2x of reclaim_required_space.
424a39f7afdSSong Liu  */
425a39f7afdSSong Liu static inline void r5c_update_log_state(struct r5l_log *log)
426a39f7afdSSong Liu {
427a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
428a39f7afdSSong Liu 	sector_t free_space;
429a39f7afdSSong Liu 	sector_t reclaim_space;
430f687a33eSSong Liu 	bool wake_reclaim = false;
431a39f7afdSSong Liu 
432a39f7afdSSong Liu 	if (!r5c_is_writeback(log))
433a39f7afdSSong Liu 		return;
434a39f7afdSSong Liu 
435a39f7afdSSong Liu 	free_space = r5l_ring_distance(log, log->log_start,
436a39f7afdSSong Liu 				       log->last_checkpoint);
437a39f7afdSSong Liu 	reclaim_space = r5c_log_required_to_flush_cache(conf);
438a39f7afdSSong Liu 	if (free_space < 2 * reclaim_space)
439a39f7afdSSong Liu 		set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
440f687a33eSSong Liu 	else {
441f687a33eSSong Liu 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
442f687a33eSSong Liu 			wake_reclaim = true;
443a39f7afdSSong Liu 		clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
444f687a33eSSong Liu 	}
445a39f7afdSSong Liu 	if (free_space < 3 * reclaim_space)
446a39f7afdSSong Liu 		set_bit(R5C_LOG_TIGHT, &conf->cache_state);
447a39f7afdSSong Liu 	else
448a39f7afdSSong Liu 		clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
449f687a33eSSong Liu 
450f687a33eSSong Liu 	if (wake_reclaim)
451f687a33eSSong Liu 		r5l_wake_reclaim(log, 0);
452a39f7afdSSong Liu }
453a39f7afdSSong Liu 
4542ded3703SSong Liu /*
4552ded3703SSong Liu  * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
4562ded3703SSong Liu  * This function should only be called in write-back mode.
4572ded3703SSong Liu  */
458a39f7afdSSong Liu void r5c_make_stripe_write_out(struct stripe_head *sh)
4592ded3703SSong Liu {
4602ded3703SSong Liu 	struct r5conf *conf = sh->raid_conf;
4612ded3703SSong Liu 	struct r5l_log *log = conf->log;
4622ded3703SSong Liu 
4632ded3703SSong Liu 	BUG_ON(!r5c_is_writeback(log));
4642ded3703SSong Liu 
4652ded3703SSong Liu 	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
4662ded3703SSong Liu 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
4671e6d690bSSong Liu 
4681e6d690bSSong Liu 	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4691e6d690bSSong Liu 		atomic_inc(&conf->preread_active_stripes);
4701e6d690bSSong Liu }
4711e6d690bSSong Liu 
4721e6d690bSSong Liu static void r5c_handle_data_cached(struct stripe_head *sh)
4731e6d690bSSong Liu {
4741e6d690bSSong Liu 	int i;
4751e6d690bSSong Liu 
4761e6d690bSSong Liu 	for (i = sh->disks; i--; )
4771e6d690bSSong Liu 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
4781e6d690bSSong Liu 			set_bit(R5_InJournal, &sh->dev[i].flags);
4791e6d690bSSong Liu 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
4801e6d690bSSong Liu 		}
4811e6d690bSSong Liu 	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
4821e6d690bSSong Liu }
4831e6d690bSSong Liu 
4841e6d690bSSong Liu /*
4851e6d690bSSong Liu  * this journal write must contain full parity,
4861e6d690bSSong Liu  * it may also contain some data pages
4871e6d690bSSong Liu  */
4881e6d690bSSong Liu static void r5c_handle_parity_cached(struct stripe_head *sh)
4891e6d690bSSong Liu {
4901e6d690bSSong Liu 	int i;
4911e6d690bSSong Liu 
4921e6d690bSSong Liu 	for (i = sh->disks; i--; )
4931e6d690bSSong Liu 		if (test_bit(R5_InJournal, &sh->dev[i].flags))
4941e6d690bSSong Liu 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
4952ded3703SSong Liu }
4962ded3703SSong Liu 
4972ded3703SSong Liu /*
4982ded3703SSong Liu  * Setting proper flags after writing (or flushing) data and/or parity to the
4992ded3703SSong Liu  * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
5002ded3703SSong Liu  */
5012ded3703SSong Liu static void r5c_finish_cache_stripe(struct stripe_head *sh)
5022ded3703SSong Liu {
5032ded3703SSong Liu 	struct r5l_log *log = sh->raid_conf->log;
5042ded3703SSong Liu 
5052ded3703SSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
5062ded3703SSong Liu 		BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
5072ded3703SSong Liu 		/*
5082ded3703SSong Liu 		 * Set R5_InJournal for parity dev[pd_idx]. This means
5092ded3703SSong Liu 		 * all data AND parity in the journal. For RAID 6, it is
5102ded3703SSong Liu 		 * NOT necessary to set the flag for dev[qd_idx], as the
5112ded3703SSong Liu 		 * two parities are written out together.
5122ded3703SSong Liu 		 */
5132ded3703SSong Liu 		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
5141e6d690bSSong Liu 	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
5151e6d690bSSong Liu 		r5c_handle_data_cached(sh);
5161e6d690bSSong Liu 	} else {
5171e6d690bSSong Liu 		r5c_handle_parity_cached(sh);
5181e6d690bSSong Liu 		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
5191e6d690bSSong Liu 	}
5202ded3703SSong Liu }
5212ded3703SSong Liu 
522d8858f43SChristoph Hellwig static void r5l_io_run_stripes(struct r5l_io_unit *io)
523d8858f43SChristoph Hellwig {
524d8858f43SChristoph Hellwig 	struct stripe_head *sh, *next;
525d8858f43SChristoph Hellwig 
526d8858f43SChristoph Hellwig 	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
527d8858f43SChristoph Hellwig 		list_del_init(&sh->log_list);
5282ded3703SSong Liu 
5292ded3703SSong Liu 		r5c_finish_cache_stripe(sh);
5302ded3703SSong Liu 
531d8858f43SChristoph Hellwig 		set_bit(STRIPE_HANDLE, &sh->state);
532d8858f43SChristoph Hellwig 		raid5_release_stripe(sh);
533d8858f43SChristoph Hellwig 	}
534d8858f43SChristoph Hellwig }
535d8858f43SChristoph Hellwig 
53656fef7c6SChristoph Hellwig static void r5l_log_run_stripes(struct r5l_log *log)
53756fef7c6SChristoph Hellwig {
53856fef7c6SChristoph Hellwig 	struct r5l_io_unit *io, *next;
53956fef7c6SChristoph Hellwig 
54056fef7c6SChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
54156fef7c6SChristoph Hellwig 
54256fef7c6SChristoph Hellwig 	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
54356fef7c6SChristoph Hellwig 		/* don't change list order */
54456fef7c6SChristoph Hellwig 		if (io->state < IO_UNIT_IO_END)
54556fef7c6SChristoph Hellwig 			break;
54656fef7c6SChristoph Hellwig 
54756fef7c6SChristoph Hellwig 		list_move_tail(&io->log_sibling, &log->finished_ios);
54856fef7c6SChristoph Hellwig 		r5l_io_run_stripes(io);
54956fef7c6SChristoph Hellwig 	}
55056fef7c6SChristoph Hellwig }
55156fef7c6SChristoph Hellwig 
5523848c0bcSChristoph Hellwig static void r5l_move_to_end_ios(struct r5l_log *log)
5533848c0bcSChristoph Hellwig {
5543848c0bcSChristoph Hellwig 	struct r5l_io_unit *io, *next;
5553848c0bcSChristoph Hellwig 
5563848c0bcSChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
5573848c0bcSChristoph Hellwig 
5583848c0bcSChristoph Hellwig 	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
5593848c0bcSChristoph Hellwig 		/* don't change list order */
5603848c0bcSChristoph Hellwig 		if (io->state < IO_UNIT_IO_END)
5613848c0bcSChristoph Hellwig 			break;
5623848c0bcSChristoph Hellwig 		list_move_tail(&io->log_sibling, &log->io_end_ios);
5633848c0bcSChristoph Hellwig 	}
5643848c0bcSChristoph Hellwig }
5653848c0bcSChristoph Hellwig 
5663bddb7f8SSong Liu static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
567f6bed0efSShaohua Li static void r5l_log_endio(struct bio *bio)
568f6bed0efSShaohua Li {
569f6bed0efSShaohua Li 	struct r5l_io_unit *io = bio->bi_private;
5703bddb7f8SSong Liu 	struct r5l_io_unit *io_deferred;
571f6bed0efSShaohua Li 	struct r5l_log *log = io->log;
572509ffec7SChristoph Hellwig 	unsigned long flags;
573f6bed0efSShaohua Li 
5746e74a9cfSShaohua Li 	if (bio->bi_error)
5756e74a9cfSShaohua Li 		md_error(log->rdev->mddev, log->rdev);
5766e74a9cfSShaohua Li 
577f6bed0efSShaohua Li 	bio_put(bio);
578e8deb638SChristoph Hellwig 	mempool_free(io->meta_page, log->meta_pool);
579f6bed0efSShaohua Li 
580509ffec7SChristoph Hellwig 	spin_lock_irqsave(&log->io_list_lock, flags);
581509ffec7SChristoph Hellwig 	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
582ea17481fSSong Liu 	if (log->need_cache_flush && !list_empty(&io->stripe_list))
5833848c0bcSChristoph Hellwig 		r5l_move_to_end_ios(log);
58456fef7c6SChristoph Hellwig 	else
58556fef7c6SChristoph Hellwig 		r5l_log_run_stripes(log);
5863bddb7f8SSong Liu 	if (!list_empty(&log->running_ios)) {
5873bddb7f8SSong Liu 		/*
5883bddb7f8SSong Liu 		 * FLUSH/FUA io_unit is deferred because of ordering, now we
5893bddb7f8SSong Liu 		 * can dispatch it
5903bddb7f8SSong Liu 		 */
5913bddb7f8SSong Liu 		io_deferred = list_first_entry(&log->running_ios,
5923bddb7f8SSong Liu 					       struct r5l_io_unit, log_sibling);
5933bddb7f8SSong Liu 		if (io_deferred->io_deferred)
5943bddb7f8SSong Liu 			schedule_work(&log->deferred_io_work);
5953bddb7f8SSong Liu 	}
5963bddb7f8SSong Liu 
597509ffec7SChristoph Hellwig 	spin_unlock_irqrestore(&log->io_list_lock, flags);
598509ffec7SChristoph Hellwig 
59956fef7c6SChristoph Hellwig 	if (log->need_cache_flush)
600f6bed0efSShaohua Li 		md_wakeup_thread(log->rdev->mddev->thread);
6013bddb7f8SSong Liu 
6023bddb7f8SSong Liu 	if (io->has_null_flush) {
6033bddb7f8SSong Liu 		struct bio *bi;
6043bddb7f8SSong Liu 
6053bddb7f8SSong Liu 		WARN_ON(bio_list_empty(&io->flush_barriers));
6063bddb7f8SSong Liu 		while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
6073bddb7f8SSong Liu 			bio_endio(bi);
6083bddb7f8SSong Liu 			atomic_dec(&io->pending_stripe);
6093bddb7f8SSong Liu 		}
610ea17481fSSong Liu 	}
611ea17481fSSong Liu 
612ea17481fSSong Liu 	/* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
6133bddb7f8SSong Liu 	if (atomic_read(&io->pending_stripe) == 0)
6143bddb7f8SSong Liu 		__r5l_stripe_write_finished(io);
6153bddb7f8SSong Liu }
6163bddb7f8SSong Liu 
6173bddb7f8SSong Liu static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
6183bddb7f8SSong Liu {
6193bddb7f8SSong Liu 	unsigned long flags;
6203bddb7f8SSong Liu 
6213bddb7f8SSong Liu 	spin_lock_irqsave(&log->io_list_lock, flags);
6223bddb7f8SSong Liu 	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
6233bddb7f8SSong Liu 	spin_unlock_irqrestore(&log->io_list_lock, flags);
6243bddb7f8SSong Liu 
625*bb3338d3SSong Liu 	/*
626*bb3338d3SSong Liu 	 * In case of journal device failures, submit_bio will get error
627*bb3338d3SSong Liu 	 * and calls endio, then active stripes will continue write
628*bb3338d3SSong Liu 	 * process. Therefore, it is not necessary to check Faulty bit
629*bb3338d3SSong Liu 	 * of journal device here.
630*bb3338d3SSong Liu 	 *
631*bb3338d3SSong Liu 	 * We can't check split_bio after current_bio is submitted. If
632*bb3338d3SSong Liu 	 * io->split_bio is null, after current_bio is submitted, current_bio
633*bb3338d3SSong Liu 	 * might already be completed and the io_unit is freed. We submit
634*bb3338d3SSong Liu 	 * split_bio first to avoid the issue.
635*bb3338d3SSong Liu 	 */
636*bb3338d3SSong Liu 	if (io->split_bio) {
6373bddb7f8SSong Liu 		if (io->has_flush)
63820737738SShaohua Li 			io->split_bio->bi_opf |= REQ_PREFLUSH;
6393bddb7f8SSong Liu 		if (io->has_fua)
64020737738SShaohua Li 			io->split_bio->bi_opf |= REQ_FUA;
6413bddb7f8SSong Liu 		submit_bio(io->split_bio);
6423bddb7f8SSong Liu 	}
6433bddb7f8SSong Liu 
644*bb3338d3SSong Liu 	if (io->has_flush)
645*bb3338d3SSong Liu 		io->current_bio->bi_opf |= REQ_PREFLUSH;
646*bb3338d3SSong Liu 	if (io->has_fua)
647*bb3338d3SSong Liu 		io->current_bio->bi_opf |= REQ_FUA;
648*bb3338d3SSong Liu 	submit_bio(io->current_bio);
649*bb3338d3SSong Liu }
650*bb3338d3SSong Liu 
6513bddb7f8SSong Liu /* deferred io_unit will be dispatched here */
6523bddb7f8SSong Liu static void r5l_submit_io_async(struct work_struct *work)
6533bddb7f8SSong Liu {
6543bddb7f8SSong Liu 	struct r5l_log *log = container_of(work, struct r5l_log,
6553bddb7f8SSong Liu 					   deferred_io_work);
6563bddb7f8SSong Liu 	struct r5l_io_unit *io = NULL;
6573bddb7f8SSong Liu 	unsigned long flags;
6583bddb7f8SSong Liu 
6593bddb7f8SSong Liu 	spin_lock_irqsave(&log->io_list_lock, flags);
6603bddb7f8SSong Liu 	if (!list_empty(&log->running_ios)) {
6613bddb7f8SSong Liu 		io = list_first_entry(&log->running_ios, struct r5l_io_unit,
6623bddb7f8SSong Liu 				      log_sibling);
6633bddb7f8SSong Liu 		if (!io->io_deferred)
6643bddb7f8SSong Liu 			io = NULL;
6653bddb7f8SSong Liu 		else
6663bddb7f8SSong Liu 			io->io_deferred = 0;
6673bddb7f8SSong Liu 	}
6683bddb7f8SSong Liu 	spin_unlock_irqrestore(&log->io_list_lock, flags);
6693bddb7f8SSong Liu 	if (io)
6703bddb7f8SSong Liu 		r5l_do_submit_io(log, io);
671f6bed0efSShaohua Li }
672f6bed0efSShaohua Li 
6732e38a37fSSong Liu static void r5c_disable_writeback_async(struct work_struct *work)
6742e38a37fSSong Liu {
6752e38a37fSSong Liu 	struct r5l_log *log = container_of(work, struct r5l_log,
6762e38a37fSSong Liu 					   disable_writeback_work);
6772e38a37fSSong Liu 	struct mddev *mddev = log->rdev->mddev;
6782e38a37fSSong Liu 
6792e38a37fSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
6802e38a37fSSong Liu 		return;
6812e38a37fSSong Liu 	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
6822e38a37fSSong Liu 		mdname(mddev));
6832e38a37fSSong Liu 	mddev_suspend(mddev);
6842e38a37fSSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
6852e38a37fSSong Liu 	mddev_resume(mddev);
6862e38a37fSSong Liu }
6872e38a37fSSong Liu 
688f6bed0efSShaohua Li static void r5l_submit_current_io(struct r5l_log *log)
689f6bed0efSShaohua Li {
690f6bed0efSShaohua Li 	struct r5l_io_unit *io = log->current_io;
6913bddb7f8SSong Liu 	struct bio *bio;
692f6bed0efSShaohua Li 	struct r5l_meta_block *block;
693509ffec7SChristoph Hellwig 	unsigned long flags;
694f6bed0efSShaohua Li 	u32 crc;
6953bddb7f8SSong Liu 	bool do_submit = true;
696f6bed0efSShaohua Li 
697f6bed0efSShaohua Li 	if (!io)
698f6bed0efSShaohua Li 		return;
699f6bed0efSShaohua Li 
700f6bed0efSShaohua Li 	block = page_address(io->meta_page);
701f6bed0efSShaohua Li 	block->meta_size = cpu_to_le32(io->meta_offset);
7025cb2fbd6SShaohua Li 	crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
703f6bed0efSShaohua Li 	block->checksum = cpu_to_le32(crc);
7043bddb7f8SSong Liu 	bio = io->current_bio;
705f6bed0efSShaohua Li 
706f6bed0efSShaohua Li 	log->current_io = NULL;
707509ffec7SChristoph Hellwig 	spin_lock_irqsave(&log->io_list_lock, flags);
7083bddb7f8SSong Liu 	if (io->has_flush || io->has_fua) {
7093bddb7f8SSong Liu 		if (io != list_first_entry(&log->running_ios,
7103bddb7f8SSong Liu 					   struct r5l_io_unit, log_sibling)) {
7113bddb7f8SSong Liu 			io->io_deferred = 1;
7123bddb7f8SSong Liu 			do_submit = false;
7133bddb7f8SSong Liu 		}
7143bddb7f8SSong Liu 	}
715509ffec7SChristoph Hellwig 	spin_unlock_irqrestore(&log->io_list_lock, flags);
7163bddb7f8SSong Liu 	if (do_submit)
7173bddb7f8SSong Liu 		r5l_do_submit_io(log, io);
718f6bed0efSShaohua Li }
719f6bed0efSShaohua Li 
7206143e2ceSChristoph Hellwig static struct bio *r5l_bio_alloc(struct r5l_log *log)
721b349feb3SChristoph Hellwig {
722c38d29b3SChristoph Hellwig 	struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
723b349feb3SChristoph Hellwig 
724796a5cf0SMike Christie 	bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
725b349feb3SChristoph Hellwig 	bio->bi_bdev = log->rdev->bdev;
7261e932a37SChristoph Hellwig 	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
727b349feb3SChristoph Hellwig 
728b349feb3SChristoph Hellwig 	return bio;
729b349feb3SChristoph Hellwig }
730b349feb3SChristoph Hellwig 
731c1b99198SChristoph Hellwig static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
732c1b99198SChristoph Hellwig {
733c1b99198SChristoph Hellwig 	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
734c1b99198SChristoph Hellwig 
735a39f7afdSSong Liu 	r5c_update_log_state(log);
736c1b99198SChristoph Hellwig 	/*
737c1b99198SChristoph Hellwig 	 * If we filled up the log device start from the beginning again,
738c1b99198SChristoph Hellwig 	 * which will require a new bio.
739c1b99198SChristoph Hellwig 	 *
740c1b99198SChristoph Hellwig 	 * Note: for this to work properly the log size needs to me a multiple
741c1b99198SChristoph Hellwig 	 * of BLOCK_SECTORS.
742c1b99198SChristoph Hellwig 	 */
743c1b99198SChristoph Hellwig 	if (log->log_start == 0)
7446143e2ceSChristoph Hellwig 		io->need_split_bio = true;
745c1b99198SChristoph Hellwig 
746c1b99198SChristoph Hellwig 	io->log_end = log->log_start;
747c1b99198SChristoph Hellwig }
748c1b99198SChristoph Hellwig 
749f6bed0efSShaohua Li static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
750f6bed0efSShaohua Li {
751f6bed0efSShaohua Li 	struct r5l_io_unit *io;
752f6bed0efSShaohua Li 	struct r5l_meta_block *block;
753f6bed0efSShaohua Li 
7545036c390SChristoph Hellwig 	io = mempool_alloc(log->io_pool, GFP_ATOMIC);
7555036c390SChristoph Hellwig 	if (!io)
7565036c390SChristoph Hellwig 		return NULL;
7575036c390SChristoph Hellwig 	memset(io, 0, sizeof(*io));
7585036c390SChristoph Hellwig 
75951039cd0SChristoph Hellwig 	io->log = log;
76051039cd0SChristoph Hellwig 	INIT_LIST_HEAD(&io->log_sibling);
76151039cd0SChristoph Hellwig 	INIT_LIST_HEAD(&io->stripe_list);
7623bddb7f8SSong Liu 	bio_list_init(&io->flush_barriers);
76351039cd0SChristoph Hellwig 	io->state = IO_UNIT_RUNNING;
764f6bed0efSShaohua Li 
765e8deb638SChristoph Hellwig 	io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
766f6bed0efSShaohua Li 	block = page_address(io->meta_page);
767e8deb638SChristoph Hellwig 	clear_page(block);
768f6bed0efSShaohua Li 	block->magic = cpu_to_le32(R5LOG_MAGIC);
769f6bed0efSShaohua Li 	block->version = R5LOG_VERSION;
770f6bed0efSShaohua Li 	block->seq = cpu_to_le64(log->seq);
771f6bed0efSShaohua Li 	block->position = cpu_to_le64(log->log_start);
772f6bed0efSShaohua Li 
773f6bed0efSShaohua Li 	io->log_start = log->log_start;
774f6bed0efSShaohua Li 	io->meta_offset = sizeof(struct r5l_meta_block);
7752b8ef16eSChristoph Hellwig 	io->seq = log->seq++;
776f6bed0efSShaohua Li 
7776143e2ceSChristoph Hellwig 	io->current_bio = r5l_bio_alloc(log);
7786143e2ceSChristoph Hellwig 	io->current_bio->bi_end_io = r5l_log_endio;
7796143e2ceSChristoph Hellwig 	io->current_bio->bi_private = io;
780b349feb3SChristoph Hellwig 	bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
781f6bed0efSShaohua Li 
782c1b99198SChristoph Hellwig 	r5_reserve_log_entry(log, io);
783f6bed0efSShaohua Li 
784f6bed0efSShaohua Li 	spin_lock_irq(&log->io_list_lock);
785f6bed0efSShaohua Li 	list_add_tail(&io->log_sibling, &log->running_ios);
786f6bed0efSShaohua Li 	spin_unlock_irq(&log->io_list_lock);
787f6bed0efSShaohua Li 
788f6bed0efSShaohua Li 	return io;
789f6bed0efSShaohua Li }
790f6bed0efSShaohua Li 
791f6bed0efSShaohua Li static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
792f6bed0efSShaohua Li {
79322581f58SChristoph Hellwig 	if (log->current_io &&
79422581f58SChristoph Hellwig 	    log->current_io->meta_offset + payload_size > PAGE_SIZE)
795f6bed0efSShaohua Li 		r5l_submit_current_io(log);
796f6bed0efSShaohua Li 
7975036c390SChristoph Hellwig 	if (!log->current_io) {
798f6bed0efSShaohua Li 		log->current_io = r5l_new_meta(log);
7995036c390SChristoph Hellwig 		if (!log->current_io)
8005036c390SChristoph Hellwig 			return -ENOMEM;
8015036c390SChristoph Hellwig 	}
8025036c390SChristoph Hellwig 
803f6bed0efSShaohua Li 	return 0;
804f6bed0efSShaohua Li }
805f6bed0efSShaohua Li 
806f6bed0efSShaohua Li static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
807f6bed0efSShaohua Li 				    sector_t location,
808f6bed0efSShaohua Li 				    u32 checksum1, u32 checksum2,
809f6bed0efSShaohua Li 				    bool checksum2_valid)
810f6bed0efSShaohua Li {
811f6bed0efSShaohua Li 	struct r5l_io_unit *io = log->current_io;
812f6bed0efSShaohua Li 	struct r5l_payload_data_parity *payload;
813f6bed0efSShaohua Li 
814f6bed0efSShaohua Li 	payload = page_address(io->meta_page) + io->meta_offset;
815f6bed0efSShaohua Li 	payload->header.type = cpu_to_le16(type);
816f6bed0efSShaohua Li 	payload->header.flags = cpu_to_le16(0);
817f6bed0efSShaohua Li 	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
818f6bed0efSShaohua Li 				    (PAGE_SHIFT - 9));
819f6bed0efSShaohua Li 	payload->location = cpu_to_le64(location);
820f6bed0efSShaohua Li 	payload->checksum[0] = cpu_to_le32(checksum1);
821f6bed0efSShaohua Li 	if (checksum2_valid)
822f6bed0efSShaohua Li 		payload->checksum[1] = cpu_to_le32(checksum2);
823f6bed0efSShaohua Li 
824f6bed0efSShaohua Li 	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
825f6bed0efSShaohua Li 		sizeof(__le32) * (1 + !!checksum2_valid);
826f6bed0efSShaohua Li }
827f6bed0efSShaohua Li 
828f6bed0efSShaohua Li static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
829f6bed0efSShaohua Li {
830f6bed0efSShaohua Li 	struct r5l_io_unit *io = log->current_io;
831f6bed0efSShaohua Li 
8326143e2ceSChristoph Hellwig 	if (io->need_split_bio) {
8333bddb7f8SSong Liu 		BUG_ON(io->split_bio);
8343bddb7f8SSong Liu 		io->split_bio = io->current_bio;
8356143e2ceSChristoph Hellwig 		io->current_bio = r5l_bio_alloc(log);
8363bddb7f8SSong Liu 		bio_chain(io->current_bio, io->split_bio);
8373bddb7f8SSong Liu 		io->need_split_bio = false;
838f6bed0efSShaohua Li 	}
839f6bed0efSShaohua Li 
8406143e2ceSChristoph Hellwig 	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
8416143e2ceSChristoph Hellwig 		BUG();
8426143e2ceSChristoph Hellwig 
843c1b99198SChristoph Hellwig 	r5_reserve_log_entry(log, io);
844f6bed0efSShaohua Li }
845f6bed0efSShaohua Li 
846ea17481fSSong Liu static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
847ea17481fSSong Liu {
848ea17481fSSong Liu 	struct mddev *mddev = log->rdev->mddev;
849ea17481fSSong Liu 	struct r5conf *conf = mddev->private;
850ea17481fSSong Liu 	struct r5l_io_unit *io;
851ea17481fSSong Liu 	struct r5l_payload_flush *payload;
852ea17481fSSong Liu 	int meta_size;
853ea17481fSSong Liu 
854ea17481fSSong Liu 	/*
855ea17481fSSong Liu 	 * payload_flush requires extra writes to the journal.
856ea17481fSSong Liu 	 * To avoid handling the extra IO in quiesce, just skip
857ea17481fSSong Liu 	 * flush_payload
858ea17481fSSong Liu 	 */
859ea17481fSSong Liu 	if (conf->quiesce)
860ea17481fSSong Liu 		return;
861ea17481fSSong Liu 
862ea17481fSSong Liu 	mutex_lock(&log->io_mutex);
863ea17481fSSong Liu 	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
864ea17481fSSong Liu 
865ea17481fSSong Liu 	if (r5l_get_meta(log, meta_size)) {
866ea17481fSSong Liu 		mutex_unlock(&log->io_mutex);
867ea17481fSSong Liu 		return;
868ea17481fSSong Liu 	}
869ea17481fSSong Liu 
870ea17481fSSong Liu 	/* current implementation is one stripe per flush payload */
871ea17481fSSong Liu 	io = log->current_io;
872ea17481fSSong Liu 	payload = page_address(io->meta_page) + io->meta_offset;
873ea17481fSSong Liu 	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
874ea17481fSSong Liu 	payload->header.flags = cpu_to_le16(0);
875ea17481fSSong Liu 	payload->size = cpu_to_le32(sizeof(__le64));
876ea17481fSSong Liu 	payload->flush_stripes[0] = cpu_to_le64(sect);
877ea17481fSSong Liu 	io->meta_offset += meta_size;
878ea17481fSSong Liu 	mutex_unlock(&log->io_mutex);
879ea17481fSSong Liu }
880ea17481fSSong Liu 
8815036c390SChristoph Hellwig static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
882f6bed0efSShaohua Li 			   int data_pages, int parity_pages)
883f6bed0efSShaohua Li {
884f6bed0efSShaohua Li 	int i;
885f6bed0efSShaohua Li 	int meta_size;
8865036c390SChristoph Hellwig 	int ret;
887f6bed0efSShaohua Li 	struct r5l_io_unit *io;
888f6bed0efSShaohua Li 
889f6bed0efSShaohua Li 	meta_size =
890f6bed0efSShaohua Li 		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
891f6bed0efSShaohua Li 		 * data_pages) +
892f6bed0efSShaohua Li 		sizeof(struct r5l_payload_data_parity) +
893f6bed0efSShaohua Li 		sizeof(__le32) * parity_pages;
894f6bed0efSShaohua Li 
8955036c390SChristoph Hellwig 	ret = r5l_get_meta(log, meta_size);
8965036c390SChristoph Hellwig 	if (ret)
8975036c390SChristoph Hellwig 		return ret;
8985036c390SChristoph Hellwig 
899f6bed0efSShaohua Li 	io = log->current_io;
900f6bed0efSShaohua Li 
9013bddb7f8SSong Liu 	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
9023bddb7f8SSong Liu 		io->has_flush = 1;
9033bddb7f8SSong Liu 
904f6bed0efSShaohua Li 	for (i = 0; i < sh->disks; i++) {
9051e6d690bSSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
9061e6d690bSSong Liu 		    test_bit(R5_InJournal, &sh->dev[i].flags))
907f6bed0efSShaohua Li 			continue;
908f6bed0efSShaohua Li 		if (i == sh->pd_idx || i == sh->qd_idx)
909f6bed0efSShaohua Li 			continue;
9103bddb7f8SSong Liu 		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
9113bddb7f8SSong Liu 		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
9123bddb7f8SSong Liu 			io->has_fua = 1;
9133bddb7f8SSong Liu 			/*
9143bddb7f8SSong Liu 			 * we need to flush journal to make sure recovery can
9153bddb7f8SSong Liu 			 * reach the data with fua flag
9163bddb7f8SSong Liu 			 */
9173bddb7f8SSong Liu 			io->has_flush = 1;
9183bddb7f8SSong Liu 		}
919f6bed0efSShaohua Li 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
920f6bed0efSShaohua Li 					raid5_compute_blocknr(sh, i, 0),
921f6bed0efSShaohua Li 					sh->dev[i].log_checksum, 0, false);
922f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[i].page);
923f6bed0efSShaohua Li 	}
924f6bed0efSShaohua Li 
9252ded3703SSong Liu 	if (parity_pages == 2) {
926f6bed0efSShaohua Li 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
927f6bed0efSShaohua Li 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
928f6bed0efSShaohua Li 					sh->dev[sh->qd_idx].log_checksum, true);
929f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
930f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
9312ded3703SSong Liu 	} else if (parity_pages == 1) {
932f6bed0efSShaohua Li 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
933f6bed0efSShaohua Li 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
934f6bed0efSShaohua Li 					0, false);
935f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
9362ded3703SSong Liu 	} else  /* Just writing data, not parity, in caching phase */
9372ded3703SSong Liu 		BUG_ON(parity_pages != 0);
938f6bed0efSShaohua Li 
939f6bed0efSShaohua Li 	list_add_tail(&sh->log_list, &io->stripe_list);
940f6bed0efSShaohua Li 	atomic_inc(&io->pending_stripe);
941f6bed0efSShaohua Li 	sh->log_io = io;
9425036c390SChristoph Hellwig 
943a39f7afdSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
944a39f7afdSSong Liu 		return 0;
945a39f7afdSSong Liu 
946a39f7afdSSong Liu 	if (sh->log_start == MaxSector) {
947a39f7afdSSong Liu 		BUG_ON(!list_empty(&sh->r5c));
948a39f7afdSSong Liu 		sh->log_start = io->log_start;
949a39f7afdSSong Liu 		spin_lock_irq(&log->stripe_in_journal_lock);
950a39f7afdSSong Liu 		list_add_tail(&sh->r5c,
951a39f7afdSSong Liu 			      &log->stripe_in_journal_list);
952a39f7afdSSong Liu 		spin_unlock_irq(&log->stripe_in_journal_lock);
953a39f7afdSSong Liu 		atomic_inc(&log->stripe_in_journal_count);
954a39f7afdSSong Liu 	}
9555036c390SChristoph Hellwig 	return 0;
956f6bed0efSShaohua Li }
957f6bed0efSShaohua Li 
958a39f7afdSSong Liu /* add stripe to no_space_stripes, and then wake up reclaim */
959a39f7afdSSong Liu static inline void r5l_add_no_space_stripe(struct r5l_log *log,
960a39f7afdSSong Liu 					   struct stripe_head *sh)
961a39f7afdSSong Liu {
962a39f7afdSSong Liu 	spin_lock(&log->no_space_stripes_lock);
963a39f7afdSSong Liu 	list_add_tail(&sh->log_list, &log->no_space_stripes);
964a39f7afdSSong Liu 	spin_unlock(&log->no_space_stripes_lock);
965a39f7afdSSong Liu }
966a39f7afdSSong Liu 
967f6bed0efSShaohua Li /*
968f6bed0efSShaohua Li  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
969f6bed0efSShaohua Li  * data from log to raid disks), so we shouldn't wait for reclaim here
970f6bed0efSShaohua Li  */
971f6bed0efSShaohua Li int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
972f6bed0efSShaohua Li {
973a39f7afdSSong Liu 	struct r5conf *conf = sh->raid_conf;
974f6bed0efSShaohua Li 	int write_disks = 0;
975f6bed0efSShaohua Li 	int data_pages, parity_pages;
976f6bed0efSShaohua Li 	int reserve;
977f6bed0efSShaohua Li 	int i;
9785036c390SChristoph Hellwig 	int ret = 0;
979a39f7afdSSong Liu 	bool wake_reclaim = false;
980f6bed0efSShaohua Li 
981f6bed0efSShaohua Li 	if (!log)
982f6bed0efSShaohua Li 		return -EAGAIN;
983f6bed0efSShaohua Li 	/* Don't support stripe batch */
984f6bed0efSShaohua Li 	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
985f6bed0efSShaohua Li 	    test_bit(STRIPE_SYNCING, &sh->state)) {
986f6bed0efSShaohua Li 		/* the stripe is written to log, we start writing it to raid */
987f6bed0efSShaohua Li 		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
988f6bed0efSShaohua Li 		return -EAGAIN;
989f6bed0efSShaohua Li 	}
990f6bed0efSShaohua Li 
9912ded3703SSong Liu 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
9922ded3703SSong Liu 
993f6bed0efSShaohua Li 	for (i = 0; i < sh->disks; i++) {
994f6bed0efSShaohua Li 		void *addr;
995f6bed0efSShaohua Li 
9961e6d690bSSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
9971e6d690bSSong Liu 		    test_bit(R5_InJournal, &sh->dev[i].flags))
998f6bed0efSShaohua Li 			continue;
9991e6d690bSSong Liu 
1000f6bed0efSShaohua Li 		write_disks++;
1001f6bed0efSShaohua Li 		/* checksum is already calculated in last run */
1002f6bed0efSShaohua Li 		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1003f6bed0efSShaohua Li 			continue;
1004f6bed0efSShaohua Li 		addr = kmap_atomic(sh->dev[i].page);
10055cb2fbd6SShaohua Li 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1006f6bed0efSShaohua Li 						    addr, PAGE_SIZE);
1007f6bed0efSShaohua Li 		kunmap_atomic(addr);
1008f6bed0efSShaohua Li 	}
1009f6bed0efSShaohua Li 	parity_pages = 1 + !!(sh->qd_idx >= 0);
1010f6bed0efSShaohua Li 	data_pages = write_disks - parity_pages;
1011f6bed0efSShaohua Li 
1012f6bed0efSShaohua Li 	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1013253f9fd4SShaohua Li 	/*
1014253f9fd4SShaohua Li 	 * The stripe must enter state machine again to finish the write, so
1015253f9fd4SShaohua Li 	 * don't delay.
1016253f9fd4SShaohua Li 	 */
1017253f9fd4SShaohua Li 	clear_bit(STRIPE_DELAYED, &sh->state);
1018f6bed0efSShaohua Li 	atomic_inc(&sh->count);
1019f6bed0efSShaohua Li 
1020f6bed0efSShaohua Li 	mutex_lock(&log->io_mutex);
1021f6bed0efSShaohua Li 	/* meta + data */
1022f6bed0efSShaohua Li 	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1023f6bed0efSShaohua Li 
1024a39f7afdSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1025a39f7afdSSong Liu 		if (!r5l_has_free_space(log, reserve)) {
1026a39f7afdSSong Liu 			r5l_add_no_space_stripe(log, sh);
1027a39f7afdSSong Liu 			wake_reclaim = true;
10285036c390SChristoph Hellwig 		} else {
10295036c390SChristoph Hellwig 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
10305036c390SChristoph Hellwig 			if (ret) {
10315036c390SChristoph Hellwig 				spin_lock_irq(&log->io_list_lock);
1032a39f7afdSSong Liu 				list_add_tail(&sh->log_list,
1033a39f7afdSSong Liu 					      &log->no_mem_stripes);
10345036c390SChristoph Hellwig 				spin_unlock_irq(&log->io_list_lock);
1035f6bed0efSShaohua Li 			}
10365036c390SChristoph Hellwig 		}
1037a39f7afdSSong Liu 	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1038a39f7afdSSong Liu 		/*
1039a39f7afdSSong Liu 		 * log space critical, do not process stripes that are
1040a39f7afdSSong Liu 		 * not in cache yet (sh->log_start == MaxSector).
1041a39f7afdSSong Liu 		 */
1042a39f7afdSSong Liu 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1043a39f7afdSSong Liu 		    sh->log_start == MaxSector) {
1044a39f7afdSSong Liu 			r5l_add_no_space_stripe(log, sh);
1045a39f7afdSSong Liu 			wake_reclaim = true;
1046a39f7afdSSong Liu 			reserve = 0;
1047a39f7afdSSong Liu 		} else if (!r5l_has_free_space(log, reserve)) {
1048a39f7afdSSong Liu 			if (sh->log_start == log->last_checkpoint)
1049a39f7afdSSong Liu 				BUG();
1050a39f7afdSSong Liu 			else
1051a39f7afdSSong Liu 				r5l_add_no_space_stripe(log, sh);
1052a39f7afdSSong Liu 		} else {
1053a39f7afdSSong Liu 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1054a39f7afdSSong Liu 			if (ret) {
1055a39f7afdSSong Liu 				spin_lock_irq(&log->io_list_lock);
1056a39f7afdSSong Liu 				list_add_tail(&sh->log_list,
1057a39f7afdSSong Liu 					      &log->no_mem_stripes);
1058a39f7afdSSong Liu 				spin_unlock_irq(&log->io_list_lock);
1059a39f7afdSSong Liu 			}
1060a39f7afdSSong Liu 		}
1061a39f7afdSSong Liu 	}
1062f6bed0efSShaohua Li 
10635036c390SChristoph Hellwig 	mutex_unlock(&log->io_mutex);
1064a39f7afdSSong Liu 	if (wake_reclaim)
1065a39f7afdSSong Liu 		r5l_wake_reclaim(log, reserve);
1066f6bed0efSShaohua Li 	return 0;
1067f6bed0efSShaohua Li }
1068f6bed0efSShaohua Li 
1069f6bed0efSShaohua Li void r5l_write_stripe_run(struct r5l_log *log)
1070f6bed0efSShaohua Li {
1071f6bed0efSShaohua Li 	if (!log)
1072f6bed0efSShaohua Li 		return;
1073f6bed0efSShaohua Li 	mutex_lock(&log->io_mutex);
1074f6bed0efSShaohua Li 	r5l_submit_current_io(log);
1075f6bed0efSShaohua Li 	mutex_unlock(&log->io_mutex);
1076f6bed0efSShaohua Li }
1077f6bed0efSShaohua Li 
1078828cbe98SShaohua Li int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1079828cbe98SShaohua Li {
1080828cbe98SShaohua Li 	if (!log)
1081828cbe98SShaohua Li 		return -ENODEV;
10823bddb7f8SSong Liu 
10833bddb7f8SSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1084828cbe98SShaohua Li 		/*
10853bddb7f8SSong Liu 		 * in write through (journal only)
10863bddb7f8SSong Liu 		 * we flush log disk cache first, then write stripe data to
10873bddb7f8SSong Liu 		 * raid disks. So if bio is finished, the log disk cache is
10883bddb7f8SSong Liu 		 * flushed already. The recovery guarantees we can recovery
10893bddb7f8SSong Liu 		 * the bio from log disk, so we don't need to flush again
1090828cbe98SShaohua Li 		 */
1091828cbe98SShaohua Li 		if (bio->bi_iter.bi_size == 0) {
1092828cbe98SShaohua Li 			bio_endio(bio);
1093828cbe98SShaohua Li 			return 0;
1094828cbe98SShaohua Li 		}
10951eff9d32SJens Axboe 		bio->bi_opf &= ~REQ_PREFLUSH;
10963bddb7f8SSong Liu 	} else {
10973bddb7f8SSong Liu 		/* write back (with cache) */
10983bddb7f8SSong Liu 		if (bio->bi_iter.bi_size == 0) {
10993bddb7f8SSong Liu 			mutex_lock(&log->io_mutex);
11003bddb7f8SSong Liu 			r5l_get_meta(log, 0);
11013bddb7f8SSong Liu 			bio_list_add(&log->current_io->flush_barriers, bio);
11023bddb7f8SSong Liu 			log->current_io->has_flush = 1;
11033bddb7f8SSong Liu 			log->current_io->has_null_flush = 1;
11043bddb7f8SSong Liu 			atomic_inc(&log->current_io->pending_stripe);
11053bddb7f8SSong Liu 			r5l_submit_current_io(log);
11063bddb7f8SSong Liu 			mutex_unlock(&log->io_mutex);
11073bddb7f8SSong Liu 			return 0;
11083bddb7f8SSong Liu 		}
11093bddb7f8SSong Liu 	}
1110828cbe98SShaohua Li 	return -EAGAIN;
1111828cbe98SShaohua Li }
1112828cbe98SShaohua Li 
1113f6bed0efSShaohua Li /* This will run after log space is reclaimed */
1114f6bed0efSShaohua Li static void r5l_run_no_space_stripes(struct r5l_log *log)
1115f6bed0efSShaohua Li {
1116f6bed0efSShaohua Li 	struct stripe_head *sh;
1117f6bed0efSShaohua Li 
1118f6bed0efSShaohua Li 	spin_lock(&log->no_space_stripes_lock);
1119f6bed0efSShaohua Li 	while (!list_empty(&log->no_space_stripes)) {
1120f6bed0efSShaohua Li 		sh = list_first_entry(&log->no_space_stripes,
1121f6bed0efSShaohua Li 				      struct stripe_head, log_list);
1122f6bed0efSShaohua Li 		list_del_init(&sh->log_list);
1123f6bed0efSShaohua Li 		set_bit(STRIPE_HANDLE, &sh->state);
1124f6bed0efSShaohua Li 		raid5_release_stripe(sh);
1125f6bed0efSShaohua Li 	}
1126f6bed0efSShaohua Li 	spin_unlock(&log->no_space_stripes_lock);
1127f6bed0efSShaohua Li }
1128f6bed0efSShaohua Li 
1129a39f7afdSSong Liu /*
1130a39f7afdSSong Liu  * calculate new last_checkpoint
1131a39f7afdSSong Liu  * for write through mode, returns log->next_checkpoint
1132a39f7afdSSong Liu  * for write back, returns log_start of first sh in stripe_in_journal_list
1133a39f7afdSSong Liu  */
1134a39f7afdSSong Liu static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1135a39f7afdSSong Liu {
1136a39f7afdSSong Liu 	struct stripe_head *sh;
1137a39f7afdSSong Liu 	struct r5l_log *log = conf->log;
1138a39f7afdSSong Liu 	sector_t new_cp;
1139a39f7afdSSong Liu 	unsigned long flags;
1140a39f7afdSSong Liu 
1141a39f7afdSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1142a39f7afdSSong Liu 		return log->next_checkpoint;
1143a39f7afdSSong Liu 
1144a39f7afdSSong Liu 	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1145a39f7afdSSong Liu 	if (list_empty(&conf->log->stripe_in_journal_list)) {
1146a39f7afdSSong Liu 		/* all stripes flushed */
1147d3014e21SDan Carpenter 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1148a39f7afdSSong Liu 		return log->next_checkpoint;
1149a39f7afdSSong Liu 	}
1150a39f7afdSSong Liu 	sh = list_first_entry(&conf->log->stripe_in_journal_list,
1151a39f7afdSSong Liu 			      struct stripe_head, r5c);
1152a39f7afdSSong Liu 	new_cp = sh->log_start;
1153a39f7afdSSong Liu 	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1154a39f7afdSSong Liu 	return new_cp;
1155a39f7afdSSong Liu }
1156a39f7afdSSong Liu 
115717036461SChristoph Hellwig static sector_t r5l_reclaimable_space(struct r5l_log *log)
115817036461SChristoph Hellwig {
1159a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
1160a39f7afdSSong Liu 
116117036461SChristoph Hellwig 	return r5l_ring_distance(log, log->last_checkpoint,
1162a39f7afdSSong Liu 				 r5c_calculate_new_cp(conf));
116317036461SChristoph Hellwig }
116417036461SChristoph Hellwig 
11655036c390SChristoph Hellwig static void r5l_run_no_mem_stripe(struct r5l_log *log)
11665036c390SChristoph Hellwig {
11675036c390SChristoph Hellwig 	struct stripe_head *sh;
11685036c390SChristoph Hellwig 
11695036c390SChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
11705036c390SChristoph Hellwig 
11715036c390SChristoph Hellwig 	if (!list_empty(&log->no_mem_stripes)) {
11725036c390SChristoph Hellwig 		sh = list_first_entry(&log->no_mem_stripes,
11735036c390SChristoph Hellwig 				      struct stripe_head, log_list);
11745036c390SChristoph Hellwig 		list_del_init(&sh->log_list);
11755036c390SChristoph Hellwig 		set_bit(STRIPE_HANDLE, &sh->state);
11765036c390SChristoph Hellwig 		raid5_release_stripe(sh);
11775036c390SChristoph Hellwig 	}
11785036c390SChristoph Hellwig }
11795036c390SChristoph Hellwig 
118004732f74SChristoph Hellwig static bool r5l_complete_finished_ios(struct r5l_log *log)
118117036461SChristoph Hellwig {
118217036461SChristoph Hellwig 	struct r5l_io_unit *io, *next;
118317036461SChristoph Hellwig 	bool found = false;
118417036461SChristoph Hellwig 
118517036461SChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
118617036461SChristoph Hellwig 
118704732f74SChristoph Hellwig 	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
118817036461SChristoph Hellwig 		/* don't change list order */
118917036461SChristoph Hellwig 		if (io->state < IO_UNIT_STRIPE_END)
119017036461SChristoph Hellwig 			break;
119117036461SChristoph Hellwig 
119217036461SChristoph Hellwig 		log->next_checkpoint = io->log_start;
119317036461SChristoph Hellwig 
119417036461SChristoph Hellwig 		list_del(&io->log_sibling);
11955036c390SChristoph Hellwig 		mempool_free(io, log->io_pool);
11965036c390SChristoph Hellwig 		r5l_run_no_mem_stripe(log);
119717036461SChristoph Hellwig 
119817036461SChristoph Hellwig 		found = true;
119917036461SChristoph Hellwig 	}
120017036461SChristoph Hellwig 
120117036461SChristoph Hellwig 	return found;
120217036461SChristoph Hellwig }
120317036461SChristoph Hellwig 
1204509ffec7SChristoph Hellwig static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1205509ffec7SChristoph Hellwig {
1206509ffec7SChristoph Hellwig 	struct r5l_log *log = io->log;
1207a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
1208509ffec7SChristoph Hellwig 	unsigned long flags;
1209509ffec7SChristoph Hellwig 
1210509ffec7SChristoph Hellwig 	spin_lock_irqsave(&log->io_list_lock, flags);
1211509ffec7SChristoph Hellwig 	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
121217036461SChristoph Hellwig 
121304732f74SChristoph Hellwig 	if (!r5l_complete_finished_ios(log)) {
121485f2f9a4SShaohua Li 		spin_unlock_irqrestore(&log->io_list_lock, flags);
121585f2f9a4SShaohua Li 		return;
121685f2f9a4SShaohua Li 	}
1217509ffec7SChristoph Hellwig 
1218a39f7afdSSong Liu 	if (r5l_reclaimable_space(log) > log->max_free_space ||
1219a39f7afdSSong Liu 	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1220509ffec7SChristoph Hellwig 		r5l_wake_reclaim(log, 0);
1221509ffec7SChristoph Hellwig 
1222509ffec7SChristoph Hellwig 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1223509ffec7SChristoph Hellwig 	wake_up(&log->iounit_wait);
1224509ffec7SChristoph Hellwig }
1225509ffec7SChristoph Hellwig 
12260576b1c6SShaohua Li void r5l_stripe_write_finished(struct stripe_head *sh)
12270576b1c6SShaohua Li {
12280576b1c6SShaohua Li 	struct r5l_io_unit *io;
12290576b1c6SShaohua Li 
12300576b1c6SShaohua Li 	io = sh->log_io;
12310576b1c6SShaohua Li 	sh->log_io = NULL;
12320576b1c6SShaohua Li 
1233509ffec7SChristoph Hellwig 	if (io && atomic_dec_and_test(&io->pending_stripe))
1234509ffec7SChristoph Hellwig 		__r5l_stripe_write_finished(io);
12350576b1c6SShaohua Li }
12360576b1c6SShaohua Li 
1237a8c34f91SShaohua Li static void r5l_log_flush_endio(struct bio *bio)
1238a8c34f91SShaohua Li {
1239a8c34f91SShaohua Li 	struct r5l_log *log = container_of(bio, struct r5l_log,
1240a8c34f91SShaohua Li 		flush_bio);
1241a8c34f91SShaohua Li 	unsigned long flags;
1242a8c34f91SShaohua Li 	struct r5l_io_unit *io;
1243a8c34f91SShaohua Li 
12446e74a9cfSShaohua Li 	if (bio->bi_error)
12456e74a9cfSShaohua Li 		md_error(log->rdev->mddev, log->rdev);
12466e74a9cfSShaohua Li 
1247a8c34f91SShaohua Li 	spin_lock_irqsave(&log->io_list_lock, flags);
1248d8858f43SChristoph Hellwig 	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1249d8858f43SChristoph Hellwig 		r5l_io_run_stripes(io);
125004732f74SChristoph Hellwig 	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1251a8c34f91SShaohua Li 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1252a8c34f91SShaohua Li }
1253a8c34f91SShaohua Li 
12540576b1c6SShaohua Li /*
12550576b1c6SShaohua Li  * Starting dispatch IO to raid.
12560576b1c6SShaohua Li  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
12570576b1c6SShaohua Li  * broken meta in the middle of a log causes recovery can't find meta at the
12580576b1c6SShaohua Li  * head of log. If operations require meta at the head persistent in log, we
12590576b1c6SShaohua Li  * must make sure meta before it persistent in log too. A case is:
12600576b1c6SShaohua Li  *
12610576b1c6SShaohua Li  * stripe data/parity is in log, we start write stripe to raid disks. stripe
12620576b1c6SShaohua Li  * data/parity must be persistent in log before we do the write to raid disks.
12630576b1c6SShaohua Li  *
12640576b1c6SShaohua Li  * The solution is we restrictly maintain io_unit list order. In this case, we
12650576b1c6SShaohua Li  * only write stripes of an io_unit to raid disks till the io_unit is the first
12660576b1c6SShaohua Li  * one whose data/parity is in log.
12670576b1c6SShaohua Li  */
12680576b1c6SShaohua Li void r5l_flush_stripe_to_raid(struct r5l_log *log)
12690576b1c6SShaohua Li {
1270a8c34f91SShaohua Li 	bool do_flush;
127156fef7c6SChristoph Hellwig 
127256fef7c6SChristoph Hellwig 	if (!log || !log->need_cache_flush)
12730576b1c6SShaohua Li 		return;
12740576b1c6SShaohua Li 
1275a8c34f91SShaohua Li 	spin_lock_irq(&log->io_list_lock);
1276a8c34f91SShaohua Li 	/* flush bio is running */
1277a8c34f91SShaohua Li 	if (!list_empty(&log->flushing_ios)) {
1278a8c34f91SShaohua Li 		spin_unlock_irq(&log->io_list_lock);
12790576b1c6SShaohua Li 		return;
12800576b1c6SShaohua Li 	}
1281a8c34f91SShaohua Li 	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1282a8c34f91SShaohua Li 	do_flush = !list_empty(&log->flushing_ios);
12830576b1c6SShaohua Li 	spin_unlock_irq(&log->io_list_lock);
1284a8c34f91SShaohua Li 
1285a8c34f91SShaohua Li 	if (!do_flush)
1286a8c34f91SShaohua Li 		return;
1287a8c34f91SShaohua Li 	bio_reset(&log->flush_bio);
1288a8c34f91SShaohua Li 	log->flush_bio.bi_bdev = log->rdev->bdev;
1289a8c34f91SShaohua Li 	log->flush_bio.bi_end_io = r5l_log_flush_endio;
129070fd7614SChristoph Hellwig 	log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
12914e49ea4aSMike Christie 	submit_bio(&log->flush_bio);
12920576b1c6SShaohua Li }
12930576b1c6SShaohua Li 
12940576b1c6SShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp);
12954b482044SShaohua Li static void r5l_write_super_and_discard_space(struct r5l_log *log,
12964b482044SShaohua Li 	sector_t end)
12974b482044SShaohua Li {
12984b482044SShaohua Li 	struct block_device *bdev = log->rdev->bdev;
12994b482044SShaohua Li 	struct mddev *mddev;
13004b482044SShaohua Li 
13014b482044SShaohua Li 	r5l_write_super(log, end);
13024b482044SShaohua Li 
13034b482044SShaohua Li 	if (!blk_queue_discard(bdev_get_queue(bdev)))
13044b482044SShaohua Li 		return;
13054b482044SShaohua Li 
13064b482044SShaohua Li 	mddev = log->rdev->mddev;
13074b482044SShaohua Li 	/*
13088e018c21SShaohua Li 	 * Discard could zero data, so before discard we must make sure
13098e018c21SShaohua Li 	 * superblock is updated to new log tail. Updating superblock (either
13108e018c21SShaohua Li 	 * directly call md_update_sb() or depend on md thread) must hold
13118e018c21SShaohua Li 	 * reconfig mutex. On the other hand, raid5_quiesce is called with
13128e018c21SShaohua Li 	 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
13138e018c21SShaohua Li 	 * for all IO finish, hence waitting for reclaim thread, while reclaim
13148e018c21SShaohua Li 	 * thread is calling this function and waitting for reconfig mutex. So
13158e018c21SShaohua Li 	 * there is a deadlock. We workaround this issue with a trylock.
13168e018c21SShaohua Li 	 * FIXME: we could miss discard if we can't take reconfig mutex
13174b482044SShaohua Li 	 */
13182953079cSShaohua Li 	set_mask_bits(&mddev->sb_flags, 0,
13192953079cSShaohua Li 		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
13208e018c21SShaohua Li 	if (!mddev_trylock(mddev))
13218e018c21SShaohua Li 		return;
13224b482044SShaohua Li 	md_update_sb(mddev, 1);
13238e018c21SShaohua Li 	mddev_unlock(mddev);
13244b482044SShaohua Li 
13256e74a9cfSShaohua Li 	/* discard IO error really doesn't matter, ignore it */
13264b482044SShaohua Li 	if (log->last_checkpoint < end) {
13274b482044SShaohua Li 		blkdev_issue_discard(bdev,
13284b482044SShaohua Li 				log->last_checkpoint + log->rdev->data_offset,
13294b482044SShaohua Li 				end - log->last_checkpoint, GFP_NOIO, 0);
13304b482044SShaohua Li 	} else {
13314b482044SShaohua Li 		blkdev_issue_discard(bdev,
13324b482044SShaohua Li 				log->last_checkpoint + log->rdev->data_offset,
13334b482044SShaohua Li 				log->device_size - log->last_checkpoint,
13344b482044SShaohua Li 				GFP_NOIO, 0);
13354b482044SShaohua Li 		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
13364b482044SShaohua Li 				GFP_NOIO, 0);
13374b482044SShaohua Li 	}
13384b482044SShaohua Li }
13394b482044SShaohua Li 
1340a39f7afdSSong Liu /*
1341a39f7afdSSong Liu  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1342a39f7afdSSong Liu  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1343a39f7afdSSong Liu  *
1344a39f7afdSSong Liu  * must hold conf->device_lock
1345a39f7afdSSong Liu  */
1346a39f7afdSSong Liu static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1347a39f7afdSSong Liu {
1348a39f7afdSSong Liu 	BUG_ON(list_empty(&sh->lru));
1349a39f7afdSSong Liu 	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1350a39f7afdSSong Liu 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1351a39f7afdSSong Liu 
1352a39f7afdSSong Liu 	/*
1353a39f7afdSSong Liu 	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1354a39f7afdSSong Liu 	 * raid5_release_stripe() while holding conf->device_lock
1355a39f7afdSSong Liu 	 */
1356a39f7afdSSong Liu 	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1357a39f7afdSSong Liu 	assert_spin_locked(&conf->device_lock);
1358a39f7afdSSong Liu 
1359a39f7afdSSong Liu 	list_del_init(&sh->lru);
1360a39f7afdSSong Liu 	atomic_inc(&sh->count);
1361a39f7afdSSong Liu 
1362a39f7afdSSong Liu 	set_bit(STRIPE_HANDLE, &sh->state);
1363a39f7afdSSong Liu 	atomic_inc(&conf->active_stripes);
1364a39f7afdSSong Liu 	r5c_make_stripe_write_out(sh);
1365a39f7afdSSong Liu 
1366e33fbb9cSShaohua Li 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1367e33fbb9cSShaohua Li 		atomic_inc(&conf->r5c_flushing_partial_stripes);
1368e33fbb9cSShaohua Li 	else
1369e33fbb9cSShaohua Li 		atomic_inc(&conf->r5c_flushing_full_stripes);
1370a39f7afdSSong Liu 	raid5_release_stripe(sh);
1371a39f7afdSSong Liu }
1372a39f7afdSSong Liu 
1373a39f7afdSSong Liu /*
1374a39f7afdSSong Liu  * if num == 0, flush all full stripes
1375a39f7afdSSong Liu  * if num > 0, flush all full stripes. If less than num full stripes are
1376a39f7afdSSong Liu  *             flushed, flush some partial stripes until totally num stripes are
1377a39f7afdSSong Liu  *             flushed or there is no more cached stripes.
1378a39f7afdSSong Liu  */
1379a39f7afdSSong Liu void r5c_flush_cache(struct r5conf *conf, int num)
1380a39f7afdSSong Liu {
1381a39f7afdSSong Liu 	int count;
1382a39f7afdSSong Liu 	struct stripe_head *sh, *next;
1383a39f7afdSSong Liu 
1384a39f7afdSSong Liu 	assert_spin_locked(&conf->device_lock);
1385a39f7afdSSong Liu 	if (!conf->log)
1386a39f7afdSSong Liu 		return;
1387a39f7afdSSong Liu 
1388a39f7afdSSong Liu 	count = 0;
1389a39f7afdSSong Liu 	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1390a39f7afdSSong Liu 		r5c_flush_stripe(conf, sh);
1391a39f7afdSSong Liu 		count++;
1392a39f7afdSSong Liu 	}
1393a39f7afdSSong Liu 
1394a39f7afdSSong Liu 	if (count >= num)
1395a39f7afdSSong Liu 		return;
1396a39f7afdSSong Liu 	list_for_each_entry_safe(sh, next,
1397a39f7afdSSong Liu 				 &conf->r5c_partial_stripe_list, lru) {
1398a39f7afdSSong Liu 		r5c_flush_stripe(conf, sh);
1399a39f7afdSSong Liu 		if (++count >= num)
1400a39f7afdSSong Liu 			break;
1401a39f7afdSSong Liu 	}
1402a39f7afdSSong Liu }
1403a39f7afdSSong Liu 
1404a39f7afdSSong Liu static void r5c_do_reclaim(struct r5conf *conf)
1405a39f7afdSSong Liu {
1406a39f7afdSSong Liu 	struct r5l_log *log = conf->log;
1407a39f7afdSSong Liu 	struct stripe_head *sh;
1408a39f7afdSSong Liu 	int count = 0;
1409a39f7afdSSong Liu 	unsigned long flags;
1410a39f7afdSSong Liu 	int total_cached;
1411a39f7afdSSong Liu 	int stripes_to_flush;
1412e33fbb9cSShaohua Li 	int flushing_partial, flushing_full;
1413a39f7afdSSong Liu 
1414a39f7afdSSong Liu 	if (!r5c_is_writeback(log))
1415a39f7afdSSong Liu 		return;
1416a39f7afdSSong Liu 
1417e33fbb9cSShaohua Li 	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1418e33fbb9cSShaohua Li 	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1419a39f7afdSSong Liu 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1420e33fbb9cSShaohua Li 		atomic_read(&conf->r5c_cached_full_stripes) -
1421e33fbb9cSShaohua Li 		flushing_full - flushing_partial;
1422a39f7afdSSong Liu 
1423a39f7afdSSong Liu 	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1424a39f7afdSSong Liu 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1425a39f7afdSSong Liu 		/*
1426a39f7afdSSong Liu 		 * if stripe cache pressure high, flush all full stripes and
1427a39f7afdSSong Liu 		 * some partial stripes
1428a39f7afdSSong Liu 		 */
1429a39f7afdSSong Liu 		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1430a39f7afdSSong Liu 	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1431e33fbb9cSShaohua Li 		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
143284890c03SShaohua Li 		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1433a39f7afdSSong Liu 		/*
1434a39f7afdSSong Liu 		 * if stripe cache pressure moderate, or if there is many full
1435a39f7afdSSong Liu 		 * stripes,flush all full stripes
1436a39f7afdSSong Liu 		 */
1437a39f7afdSSong Liu 		stripes_to_flush = 0;
1438a39f7afdSSong Liu 	else
1439a39f7afdSSong Liu 		/* no need to flush */
1440a39f7afdSSong Liu 		stripes_to_flush = -1;
1441a39f7afdSSong Liu 
1442a39f7afdSSong Liu 	if (stripes_to_flush >= 0) {
1443a39f7afdSSong Liu 		spin_lock_irqsave(&conf->device_lock, flags);
1444a39f7afdSSong Liu 		r5c_flush_cache(conf, stripes_to_flush);
1445a39f7afdSSong Liu 		spin_unlock_irqrestore(&conf->device_lock, flags);
1446a39f7afdSSong Liu 	}
1447a39f7afdSSong Liu 
1448a39f7afdSSong Liu 	/* if log space is tight, flush stripes on stripe_in_journal_list */
1449a39f7afdSSong Liu 	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1450a39f7afdSSong Liu 		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1451a39f7afdSSong Liu 		spin_lock(&conf->device_lock);
1452a39f7afdSSong Liu 		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1453a39f7afdSSong Liu 			/*
1454a39f7afdSSong Liu 			 * stripes on stripe_in_journal_list could be in any
1455a39f7afdSSong Liu 			 * state of the stripe_cache state machine. In this
1456a39f7afdSSong Liu 			 * case, we only want to flush stripe on
1457a39f7afdSSong Liu 			 * r5c_cached_full/partial_stripes. The following
1458a39f7afdSSong Liu 			 * condition makes sure the stripe is on one of the
1459a39f7afdSSong Liu 			 * two lists.
1460a39f7afdSSong Liu 			 */
1461a39f7afdSSong Liu 			if (!list_empty(&sh->lru) &&
1462a39f7afdSSong Liu 			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1463a39f7afdSSong Liu 			    atomic_read(&sh->count) == 0) {
1464a39f7afdSSong Liu 				r5c_flush_stripe(conf, sh);
1465a39f7afdSSong Liu 				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1466a39f7afdSSong Liu 					break;
1467a39f7afdSSong Liu 			}
1468e8fd52eeSShaohua Li 		}
1469a39f7afdSSong Liu 		spin_unlock(&conf->device_lock);
1470a39f7afdSSong Liu 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1471a39f7afdSSong Liu 	}
1472f687a33eSSong Liu 
1473f687a33eSSong Liu 	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1474f687a33eSSong Liu 		r5l_run_no_space_stripes(log);
1475f687a33eSSong Liu 
1476a39f7afdSSong Liu 	md_wakeup_thread(conf->mddev->thread);
1477a39f7afdSSong Liu }
14784b482044SShaohua Li 
14790576b1c6SShaohua Li static void r5l_do_reclaim(struct r5l_log *log)
14800576b1c6SShaohua Li {
1481a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
14820576b1c6SShaohua Li 	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
148317036461SChristoph Hellwig 	sector_t reclaimable;
148417036461SChristoph Hellwig 	sector_t next_checkpoint;
1485a39f7afdSSong Liu 	bool write_super;
14860576b1c6SShaohua Li 
14870576b1c6SShaohua Li 	spin_lock_irq(&log->io_list_lock);
1488a39f7afdSSong Liu 	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1489a39f7afdSSong Liu 		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
14900576b1c6SShaohua Li 	/*
14910576b1c6SShaohua Li 	 * move proper io_unit to reclaim list. We should not change the order.
14920576b1c6SShaohua Li 	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
14930576b1c6SShaohua Li 	 * shouldn't reuse space of an unreclaimable io_unit
14940576b1c6SShaohua Li 	 */
14950576b1c6SShaohua Li 	while (1) {
149617036461SChristoph Hellwig 		reclaimable = r5l_reclaimable_space(log);
149717036461SChristoph Hellwig 		if (reclaimable >= reclaim_target ||
14980576b1c6SShaohua Li 		    (list_empty(&log->running_ios) &&
14990576b1c6SShaohua Li 		     list_empty(&log->io_end_ios) &&
1500a8c34f91SShaohua Li 		     list_empty(&log->flushing_ios) &&
150104732f74SChristoph Hellwig 		     list_empty(&log->finished_ios)))
15020576b1c6SShaohua Li 			break;
15030576b1c6SShaohua Li 
150417036461SChristoph Hellwig 		md_wakeup_thread(log->rdev->mddev->thread);
150517036461SChristoph Hellwig 		wait_event_lock_irq(log->iounit_wait,
150617036461SChristoph Hellwig 				    r5l_reclaimable_space(log) > reclaimable,
150717036461SChristoph Hellwig 				    log->io_list_lock);
15080576b1c6SShaohua Li 	}
150917036461SChristoph Hellwig 
1510a39f7afdSSong Liu 	next_checkpoint = r5c_calculate_new_cp(conf);
15110576b1c6SShaohua Li 	spin_unlock_irq(&log->io_list_lock);
15120576b1c6SShaohua Li 
1513a39f7afdSSong Liu 	if (reclaimable == 0 || !write_super)
15140576b1c6SShaohua Li 		return;
15150576b1c6SShaohua Li 
15160576b1c6SShaohua Li 	/*
15170576b1c6SShaohua Li 	 * write_super will flush cache of each raid disk. We must write super
15180576b1c6SShaohua Li 	 * here, because the log area might be reused soon and we don't want to
15190576b1c6SShaohua Li 	 * confuse recovery
15200576b1c6SShaohua Li 	 */
15214b482044SShaohua Li 	r5l_write_super_and_discard_space(log, next_checkpoint);
15220576b1c6SShaohua Li 
15230576b1c6SShaohua Li 	mutex_lock(&log->io_mutex);
152417036461SChristoph Hellwig 	log->last_checkpoint = next_checkpoint;
1525a39f7afdSSong Liu 	r5c_update_log_state(log);
15260576b1c6SShaohua Li 	mutex_unlock(&log->io_mutex);
15270576b1c6SShaohua Li 
152817036461SChristoph Hellwig 	r5l_run_no_space_stripes(log);
15290576b1c6SShaohua Li }
15300576b1c6SShaohua Li 
15310576b1c6SShaohua Li static void r5l_reclaim_thread(struct md_thread *thread)
15320576b1c6SShaohua Li {
15330576b1c6SShaohua Li 	struct mddev *mddev = thread->mddev;
15340576b1c6SShaohua Li 	struct r5conf *conf = mddev->private;
15350576b1c6SShaohua Li 	struct r5l_log *log = conf->log;
15360576b1c6SShaohua Li 
15370576b1c6SShaohua Li 	if (!log)
15380576b1c6SShaohua Li 		return;
1539a39f7afdSSong Liu 	r5c_do_reclaim(conf);
15400576b1c6SShaohua Li 	r5l_do_reclaim(log);
15410576b1c6SShaohua Li }
15420576b1c6SShaohua Li 
1543a39f7afdSSong Liu void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1544f6bed0efSShaohua Li {
15450576b1c6SShaohua Li 	unsigned long target;
15460576b1c6SShaohua Li 	unsigned long new = (unsigned long)space; /* overflow in theory */
15470576b1c6SShaohua Li 
1548a39f7afdSSong Liu 	if (!log)
1549a39f7afdSSong Liu 		return;
15500576b1c6SShaohua Li 	do {
15510576b1c6SShaohua Li 		target = log->reclaim_target;
15520576b1c6SShaohua Li 		if (new < target)
15530576b1c6SShaohua Li 			return;
15540576b1c6SShaohua Li 	} while (cmpxchg(&log->reclaim_target, target, new) != target);
15550576b1c6SShaohua Li 	md_wakeup_thread(log->reclaim_thread);
1556f6bed0efSShaohua Li }
1557f6bed0efSShaohua Li 
1558e6c033f7SShaohua Li void r5l_quiesce(struct r5l_log *log, int state)
1559e6c033f7SShaohua Li {
15604b482044SShaohua Li 	struct mddev *mddev;
1561e6c033f7SShaohua Li 	if (!log || state == 2)
1562e6c033f7SShaohua Li 		return;
1563ce1ccd07SShaohua Li 	if (state == 0)
1564ce1ccd07SShaohua Li 		kthread_unpark(log->reclaim_thread->tsk);
1565ce1ccd07SShaohua Li 	else if (state == 1) {
15664b482044SShaohua Li 		/* make sure r5l_write_super_and_discard_space exits */
15674b482044SShaohua Li 		mddev = log->rdev->mddev;
15684b482044SShaohua Li 		wake_up(&mddev->sb_wait);
1569ce1ccd07SShaohua Li 		kthread_park(log->reclaim_thread->tsk);
1570a39f7afdSSong Liu 		r5l_wake_reclaim(log, MaxSector);
1571e6c033f7SShaohua Li 		r5l_do_reclaim(log);
1572e6c033f7SShaohua Li 	}
1573e6c033f7SShaohua Li }
1574e6c033f7SShaohua Li 
15756e74a9cfSShaohua Li bool r5l_log_disk_error(struct r5conf *conf)
15766e74a9cfSShaohua Li {
1577f6b6ec5cSShaohua Li 	struct r5l_log *log;
1578f6b6ec5cSShaohua Li 	bool ret;
15797dde2ad3SShaohua Li 	/* don't allow write if journal disk is missing */
1580f6b6ec5cSShaohua Li 	rcu_read_lock();
1581f6b6ec5cSShaohua Li 	log = rcu_dereference(conf->log);
1582f6b6ec5cSShaohua Li 
1583f6b6ec5cSShaohua Li 	if (!log)
1584f6b6ec5cSShaohua Li 		ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1585f6b6ec5cSShaohua Li 	else
1586f6b6ec5cSShaohua Li 		ret = test_bit(Faulty, &log->rdev->flags);
1587f6b6ec5cSShaohua Li 	rcu_read_unlock();
1588f6b6ec5cSShaohua Li 	return ret;
15896e74a9cfSShaohua Li }
15906e74a9cfSShaohua Li 
1591effe6ee7SSong Liu #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1592effe6ee7SSong Liu 
1593355810d1SShaohua Li struct r5l_recovery_ctx {
1594355810d1SShaohua Li 	struct page *meta_page;		/* current meta */
1595355810d1SShaohua Li 	sector_t meta_total_blocks;	/* total size of current meta and data */
1596355810d1SShaohua Li 	sector_t pos;			/* recovery position */
1597355810d1SShaohua Li 	u64 seq;			/* recovery position seq */
1598b4c625c6SSong Liu 	int data_parity_stripes;	/* number of data_parity stripes */
1599b4c625c6SSong Liu 	int data_only_stripes;		/* number of data_only stripes */
1600b4c625c6SSong Liu 	struct list_head cached_list;
1601effe6ee7SSong Liu 
1602effe6ee7SSong Liu 	/*
1603effe6ee7SSong Liu 	 * read ahead page pool (ra_pool)
1604effe6ee7SSong Liu 	 * in recovery, log is read sequentially. It is not efficient to
1605effe6ee7SSong Liu 	 * read every page with sync_page_io(). The read ahead page pool
1606effe6ee7SSong Liu 	 * reads multiple pages with one IO, so further log read can
1607effe6ee7SSong Liu 	 * just copy data from the pool.
1608effe6ee7SSong Liu 	 */
1609effe6ee7SSong Liu 	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1610effe6ee7SSong Liu 	sector_t pool_offset;	/* offset of first page in the pool */
1611effe6ee7SSong Liu 	int total_pages;	/* total allocated pages */
1612effe6ee7SSong Liu 	int valid_pages;	/* pages with valid data */
1613effe6ee7SSong Liu 	struct bio *ra_bio;	/* bio to do the read ahead */
1614355810d1SShaohua Li };
1615355810d1SShaohua Li 
1616effe6ee7SSong Liu static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1617effe6ee7SSong Liu 					    struct r5l_recovery_ctx *ctx)
1618effe6ee7SSong Liu {
1619effe6ee7SSong Liu 	struct page *page;
1620effe6ee7SSong Liu 
1621effe6ee7SSong Liu 	ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1622effe6ee7SSong Liu 	if (!ctx->ra_bio)
1623effe6ee7SSong Liu 		return -ENOMEM;
1624effe6ee7SSong Liu 
1625effe6ee7SSong Liu 	ctx->valid_pages = 0;
1626effe6ee7SSong Liu 	ctx->total_pages = 0;
1627effe6ee7SSong Liu 	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1628effe6ee7SSong Liu 		page = alloc_page(GFP_KERNEL);
1629effe6ee7SSong Liu 
1630effe6ee7SSong Liu 		if (!page)
1631effe6ee7SSong Liu 			break;
1632effe6ee7SSong Liu 		ctx->ra_pool[ctx->total_pages] = page;
1633effe6ee7SSong Liu 		ctx->total_pages += 1;
1634effe6ee7SSong Liu 	}
1635effe6ee7SSong Liu 
1636effe6ee7SSong Liu 	if (ctx->total_pages == 0) {
1637effe6ee7SSong Liu 		bio_put(ctx->ra_bio);
1638effe6ee7SSong Liu 		return -ENOMEM;
1639effe6ee7SSong Liu 	}
1640effe6ee7SSong Liu 
1641effe6ee7SSong Liu 	ctx->pool_offset = 0;
1642effe6ee7SSong Liu 	return 0;
1643effe6ee7SSong Liu }
1644effe6ee7SSong Liu 
1645effe6ee7SSong Liu static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1646effe6ee7SSong Liu 					struct r5l_recovery_ctx *ctx)
1647effe6ee7SSong Liu {
1648effe6ee7SSong Liu 	int i;
1649effe6ee7SSong Liu 
1650effe6ee7SSong Liu 	for (i = 0; i < ctx->total_pages; ++i)
1651effe6ee7SSong Liu 		put_page(ctx->ra_pool[i]);
1652effe6ee7SSong Liu 	bio_put(ctx->ra_bio);
1653effe6ee7SSong Liu }
1654effe6ee7SSong Liu 
1655effe6ee7SSong Liu /*
1656effe6ee7SSong Liu  * fetch ctx->valid_pages pages from offset
1657effe6ee7SSong Liu  * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1658effe6ee7SSong Liu  * However, if the offset is close to the end of the journal device,
1659effe6ee7SSong Liu  * ctx->valid_pages could be smaller than ctx->total_pages
1660effe6ee7SSong Liu  */
1661effe6ee7SSong Liu static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1662effe6ee7SSong Liu 				      struct r5l_recovery_ctx *ctx,
1663effe6ee7SSong Liu 				      sector_t offset)
1664effe6ee7SSong Liu {
1665effe6ee7SSong Liu 	bio_reset(ctx->ra_bio);
1666effe6ee7SSong Liu 	ctx->ra_bio->bi_bdev = log->rdev->bdev;
1667effe6ee7SSong Liu 	bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1668effe6ee7SSong Liu 	ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1669effe6ee7SSong Liu 
1670effe6ee7SSong Liu 	ctx->valid_pages = 0;
1671effe6ee7SSong Liu 	ctx->pool_offset = offset;
1672effe6ee7SSong Liu 
1673effe6ee7SSong Liu 	while (ctx->valid_pages < ctx->total_pages) {
1674effe6ee7SSong Liu 		bio_add_page(ctx->ra_bio,
1675effe6ee7SSong Liu 			     ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1676effe6ee7SSong Liu 		ctx->valid_pages += 1;
1677effe6ee7SSong Liu 
1678effe6ee7SSong Liu 		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1679effe6ee7SSong Liu 
1680effe6ee7SSong Liu 		if (offset == 0)  /* reached end of the device */
1681effe6ee7SSong Liu 			break;
1682effe6ee7SSong Liu 	}
1683effe6ee7SSong Liu 
1684effe6ee7SSong Liu 	return submit_bio_wait(ctx->ra_bio);
1685effe6ee7SSong Liu }
1686effe6ee7SSong Liu 
1687effe6ee7SSong Liu /*
1688effe6ee7SSong Liu  * try read a page from the read ahead page pool, if the page is not in the
1689effe6ee7SSong Liu  * pool, call r5l_recovery_fetch_ra_pool
1690effe6ee7SSong Liu  */
1691effe6ee7SSong Liu static int r5l_recovery_read_page(struct r5l_log *log,
1692effe6ee7SSong Liu 				  struct r5l_recovery_ctx *ctx,
1693effe6ee7SSong Liu 				  struct page *page,
1694effe6ee7SSong Liu 				  sector_t offset)
1695effe6ee7SSong Liu {
1696effe6ee7SSong Liu 	int ret;
1697effe6ee7SSong Liu 
1698effe6ee7SSong Liu 	if (offset < ctx->pool_offset ||
1699effe6ee7SSong Liu 	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1700effe6ee7SSong Liu 		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1701effe6ee7SSong Liu 		if (ret)
1702effe6ee7SSong Liu 			return ret;
1703effe6ee7SSong Liu 	}
1704effe6ee7SSong Liu 
1705effe6ee7SSong Liu 	BUG_ON(offset < ctx->pool_offset ||
1706effe6ee7SSong Liu 	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1707effe6ee7SSong Liu 
1708effe6ee7SSong Liu 	memcpy(page_address(page),
1709effe6ee7SSong Liu 	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1710effe6ee7SSong Liu 					 BLOCK_SECTOR_SHIFT]),
1711effe6ee7SSong Liu 	       PAGE_SIZE);
1712effe6ee7SSong Liu 	return 0;
1713effe6ee7SSong Liu }
1714effe6ee7SSong Liu 
17159ed988f5SSong Liu static int r5l_recovery_read_meta_block(struct r5l_log *log,
1716355810d1SShaohua Li 					struct r5l_recovery_ctx *ctx)
1717355810d1SShaohua Li {
1718355810d1SShaohua Li 	struct page *page = ctx->meta_page;
1719355810d1SShaohua Li 	struct r5l_meta_block *mb;
1720355810d1SShaohua Li 	u32 crc, stored_crc;
1721effe6ee7SSong Liu 	int ret;
1722355810d1SShaohua Li 
1723effe6ee7SSong Liu 	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1724effe6ee7SSong Liu 	if (ret != 0)
1725effe6ee7SSong Liu 		return ret;
1726355810d1SShaohua Li 
1727355810d1SShaohua Li 	mb = page_address(page);
1728355810d1SShaohua Li 	stored_crc = le32_to_cpu(mb->checksum);
1729355810d1SShaohua Li 	mb->checksum = 0;
1730355810d1SShaohua Li 
1731355810d1SShaohua Li 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1732355810d1SShaohua Li 	    le64_to_cpu(mb->seq) != ctx->seq ||
1733355810d1SShaohua Li 	    mb->version != R5LOG_VERSION ||
1734355810d1SShaohua Li 	    le64_to_cpu(mb->position) != ctx->pos)
1735355810d1SShaohua Li 		return -EINVAL;
1736355810d1SShaohua Li 
17375cb2fbd6SShaohua Li 	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1738355810d1SShaohua Li 	if (stored_crc != crc)
1739355810d1SShaohua Li 		return -EINVAL;
1740355810d1SShaohua Li 
1741355810d1SShaohua Li 	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1742355810d1SShaohua Li 		return -EINVAL;
1743355810d1SShaohua Li 
1744355810d1SShaohua Li 	ctx->meta_total_blocks = BLOCK_SECTORS;
1745355810d1SShaohua Li 
1746355810d1SShaohua Li 	return 0;
1747355810d1SShaohua Li }
1748355810d1SShaohua Li 
17499ed988f5SSong Liu static void
17509ed988f5SSong Liu r5l_recovery_create_empty_meta_block(struct r5l_log *log,
17519ed988f5SSong Liu 				     struct page *page,
17529ed988f5SSong Liu 				     sector_t pos, u64 seq)
1753355810d1SShaohua Li {
1754355810d1SShaohua Li 	struct r5l_meta_block *mb;
1755355810d1SShaohua Li 
1756355810d1SShaohua Li 	mb = page_address(page);
17579ed988f5SSong Liu 	clear_page(mb);
1758355810d1SShaohua Li 	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1759355810d1SShaohua Li 	mb->version = R5LOG_VERSION;
1760355810d1SShaohua Li 	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1761355810d1SShaohua Li 	mb->seq = cpu_to_le64(seq);
1762355810d1SShaohua Li 	mb->position = cpu_to_le64(pos);
1763355810d1SShaohua Li }
1764355810d1SShaohua Li 
1765355810d1SShaohua Li static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1766355810d1SShaohua Li 					  u64 seq)
1767355810d1SShaohua Li {
1768355810d1SShaohua Li 	struct page *page;
1769355810d1SShaohua Li 	struct r5l_meta_block *mb;
1770355810d1SShaohua Li 
17719ed988f5SSong Liu 	page = alloc_page(GFP_KERNEL);
1772355810d1SShaohua Li 	if (!page)
1773355810d1SShaohua Li 		return -ENOMEM;
17749ed988f5SSong Liu 	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1775355810d1SShaohua Li 	mb = page_address(page);
17765c88f403SSong Liu 	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
17775c88f403SSong Liu 					     mb, PAGE_SIZE));
1778796a5cf0SMike Christie 	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
177970fd7614SChristoph Hellwig 			  REQ_FUA, false)) {
1780355810d1SShaohua Li 		__free_page(page);
1781355810d1SShaohua Li 		return -EIO;
1782355810d1SShaohua Li 	}
1783355810d1SShaohua Li 	__free_page(page);
1784355810d1SShaohua Li 	return 0;
1785355810d1SShaohua Li }
1786355810d1SShaohua Li 
1787b4c625c6SSong Liu /*
1788b4c625c6SSong Liu  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1789b4c625c6SSong Liu  * to mark valid (potentially not flushed) data in the journal.
1790b4c625c6SSong Liu  *
1791b4c625c6SSong Liu  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1792b4c625c6SSong Liu  * so there should not be any mismatch here.
1793b4c625c6SSong Liu  */
1794b4c625c6SSong Liu static void r5l_recovery_load_data(struct r5l_log *log,
1795b4c625c6SSong Liu 				   struct stripe_head *sh,
1796b4c625c6SSong Liu 				   struct r5l_recovery_ctx *ctx,
1797b4c625c6SSong Liu 				   struct r5l_payload_data_parity *payload,
1798b4c625c6SSong Liu 				   sector_t log_offset)
1799f6bed0efSShaohua Li {
1800b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
1801b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
1802b4c625c6SSong Liu 	int dd_idx;
1803355810d1SShaohua Li 
1804b4c625c6SSong Liu 	raid5_compute_sector(conf,
1805b4c625c6SSong Liu 			     le64_to_cpu(payload->location), 0,
1806b4c625c6SSong Liu 			     &dd_idx, sh);
1807effe6ee7SSong Liu 	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1808b4c625c6SSong Liu 	sh->dev[dd_idx].log_checksum =
1809b4c625c6SSong Liu 		le32_to_cpu(payload->checksum[0]);
1810b4c625c6SSong Liu 	ctx->meta_total_blocks += BLOCK_SECTORS;
1811b4c625c6SSong Liu 
1812b4c625c6SSong Liu 	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1813b4c625c6SSong Liu 	set_bit(STRIPE_R5C_CACHING, &sh->state);
1814b4c625c6SSong Liu }
1815b4c625c6SSong Liu 
1816b4c625c6SSong Liu static void r5l_recovery_load_parity(struct r5l_log *log,
1817b4c625c6SSong Liu 				     struct stripe_head *sh,
1818b4c625c6SSong Liu 				     struct r5l_recovery_ctx *ctx,
1819b4c625c6SSong Liu 				     struct r5l_payload_data_parity *payload,
1820b4c625c6SSong Liu 				     sector_t log_offset)
1821b4c625c6SSong Liu {
1822b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
1823b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
1824b4c625c6SSong Liu 
1825b4c625c6SSong Liu 	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1826effe6ee7SSong Liu 	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1827b4c625c6SSong Liu 	sh->dev[sh->pd_idx].log_checksum =
1828b4c625c6SSong Liu 		le32_to_cpu(payload->checksum[0]);
1829b4c625c6SSong Liu 	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1830b4c625c6SSong Liu 
1831b4c625c6SSong Liu 	if (sh->qd_idx >= 0) {
1832effe6ee7SSong Liu 		r5l_recovery_read_page(
1833effe6ee7SSong Liu 			log, ctx, sh->dev[sh->qd_idx].page,
1834effe6ee7SSong Liu 			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1835b4c625c6SSong Liu 		sh->dev[sh->qd_idx].log_checksum =
1836b4c625c6SSong Liu 			le32_to_cpu(payload->checksum[1]);
1837b4c625c6SSong Liu 		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1838b4c625c6SSong Liu 	}
1839b4c625c6SSong Liu 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1840b4c625c6SSong Liu }
1841b4c625c6SSong Liu 
1842b4c625c6SSong Liu static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1843b4c625c6SSong Liu {
1844b4c625c6SSong Liu 	int i;
1845b4c625c6SSong Liu 
1846b4c625c6SSong Liu 	sh->state = 0;
1847b4c625c6SSong Liu 	sh->log_start = MaxSector;
1848b4c625c6SSong Liu 	for (i = sh->disks; i--; )
1849b4c625c6SSong Liu 		sh->dev[i].flags = 0;
1850b4c625c6SSong Liu }
1851b4c625c6SSong Liu 
1852b4c625c6SSong Liu static void
1853b4c625c6SSong Liu r5l_recovery_replay_one_stripe(struct r5conf *conf,
1854b4c625c6SSong Liu 			       struct stripe_head *sh,
1855b4c625c6SSong Liu 			       struct r5l_recovery_ctx *ctx)
1856b4c625c6SSong Liu {
1857b4c625c6SSong Liu 	struct md_rdev *rdev, *rrdev;
1858b4c625c6SSong Liu 	int disk_index;
1859b4c625c6SSong Liu 	int data_count = 0;
1860b4c625c6SSong Liu 
1861b4c625c6SSong Liu 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1862b4c625c6SSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1863b4c625c6SSong Liu 			continue;
1864b4c625c6SSong Liu 		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1865b4c625c6SSong Liu 			continue;
1866b4c625c6SSong Liu 		data_count++;
1867b4c625c6SSong Liu 	}
1868b4c625c6SSong Liu 
1869b4c625c6SSong Liu 	/*
1870b4c625c6SSong Liu 	 * stripes that only have parity must have been flushed
1871b4c625c6SSong Liu 	 * before the crash that we are now recovering from, so
1872b4c625c6SSong Liu 	 * there is nothing more to recovery.
1873b4c625c6SSong Liu 	 */
1874b4c625c6SSong Liu 	if (data_count == 0)
1875b4c625c6SSong Liu 		goto out;
1876b4c625c6SSong Liu 
1877b4c625c6SSong Liu 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1878b4c625c6SSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1879b4c625c6SSong Liu 			continue;
1880b4c625c6SSong Liu 
1881b4c625c6SSong Liu 		/* in case device is broken */
1882b4c625c6SSong Liu 		rcu_read_lock();
1883b4c625c6SSong Liu 		rdev = rcu_dereference(conf->disks[disk_index].rdev);
1884b4c625c6SSong Liu 		if (rdev) {
1885b4c625c6SSong Liu 			atomic_inc(&rdev->nr_pending);
1886b4c625c6SSong Liu 			rcu_read_unlock();
1887b4c625c6SSong Liu 			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1888b4c625c6SSong Liu 				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1889b4c625c6SSong Liu 				     false);
1890b4c625c6SSong Liu 			rdev_dec_pending(rdev, rdev->mddev);
1891b4c625c6SSong Liu 			rcu_read_lock();
1892b4c625c6SSong Liu 		}
1893b4c625c6SSong Liu 		rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1894b4c625c6SSong Liu 		if (rrdev) {
1895b4c625c6SSong Liu 			atomic_inc(&rrdev->nr_pending);
1896b4c625c6SSong Liu 			rcu_read_unlock();
1897b4c625c6SSong Liu 			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1898b4c625c6SSong Liu 				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1899b4c625c6SSong Liu 				     false);
1900b4c625c6SSong Liu 			rdev_dec_pending(rrdev, rrdev->mddev);
1901b4c625c6SSong Liu 			rcu_read_lock();
1902b4c625c6SSong Liu 		}
1903b4c625c6SSong Liu 		rcu_read_unlock();
1904b4c625c6SSong Liu 	}
1905b4c625c6SSong Liu 	ctx->data_parity_stripes++;
1906b4c625c6SSong Liu out:
1907b4c625c6SSong Liu 	r5l_recovery_reset_stripe(sh);
1908b4c625c6SSong Liu }
1909b4c625c6SSong Liu 
1910b4c625c6SSong Liu static struct stripe_head *
1911b4c625c6SSong Liu r5c_recovery_alloc_stripe(struct r5conf *conf,
19123c66abbaSSong Liu 			  sector_t stripe_sect)
1913b4c625c6SSong Liu {
1914b4c625c6SSong Liu 	struct stripe_head *sh;
1915b4c625c6SSong Liu 
1916b4c625c6SSong Liu 	sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1917b4c625c6SSong Liu 	if (!sh)
1918b4c625c6SSong Liu 		return NULL;  /* no more stripe available */
1919b4c625c6SSong Liu 
1920b4c625c6SSong Liu 	r5l_recovery_reset_stripe(sh);
1921b4c625c6SSong Liu 
1922b4c625c6SSong Liu 	return sh;
1923b4c625c6SSong Liu }
1924b4c625c6SSong Liu 
1925b4c625c6SSong Liu static struct stripe_head *
1926b4c625c6SSong Liu r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1927b4c625c6SSong Liu {
1928b4c625c6SSong Liu 	struct stripe_head *sh;
1929b4c625c6SSong Liu 
1930b4c625c6SSong Liu 	list_for_each_entry(sh, list, lru)
1931b4c625c6SSong Liu 		if (sh->sector == sect)
1932b4c625c6SSong Liu 			return sh;
1933b4c625c6SSong Liu 	return NULL;
1934b4c625c6SSong Liu }
1935b4c625c6SSong Liu 
1936b4c625c6SSong Liu static void
1937b4c625c6SSong Liu r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1938b4c625c6SSong Liu 			  struct r5l_recovery_ctx *ctx)
1939b4c625c6SSong Liu {
1940b4c625c6SSong Liu 	struct stripe_head *sh, *next;
1941b4c625c6SSong Liu 
1942b4c625c6SSong Liu 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1943b4c625c6SSong Liu 		r5l_recovery_reset_stripe(sh);
1944b4c625c6SSong Liu 		list_del_init(&sh->lru);
1945b4c625c6SSong Liu 		raid5_release_stripe(sh);
1946b4c625c6SSong Liu 	}
1947b4c625c6SSong Liu }
1948b4c625c6SSong Liu 
1949b4c625c6SSong Liu static void
1950b4c625c6SSong Liu r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1951b4c625c6SSong Liu 			    struct r5l_recovery_ctx *ctx)
1952b4c625c6SSong Liu {
1953b4c625c6SSong Liu 	struct stripe_head *sh, *next;
1954b4c625c6SSong Liu 
1955b4c625c6SSong Liu 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1956b4c625c6SSong Liu 		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1957b4c625c6SSong Liu 			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1958b4c625c6SSong Liu 			list_del_init(&sh->lru);
1959b4c625c6SSong Liu 			raid5_release_stripe(sh);
1960b4c625c6SSong Liu 		}
1961b4c625c6SSong Liu }
1962b4c625c6SSong Liu 
1963b4c625c6SSong Liu /* if matches return 0; otherwise return -EINVAL */
1964b4c625c6SSong Liu static int
1965effe6ee7SSong Liu r5l_recovery_verify_data_checksum(struct r5l_log *log,
1966effe6ee7SSong Liu 				  struct r5l_recovery_ctx *ctx,
1967effe6ee7SSong Liu 				  struct page *page,
1968b4c625c6SSong Liu 				  sector_t log_offset, __le32 log_checksum)
1969b4c625c6SSong Liu {
1970b4c625c6SSong Liu 	void *addr;
1971b4c625c6SSong Liu 	u32 checksum;
1972b4c625c6SSong Liu 
1973effe6ee7SSong Liu 	r5l_recovery_read_page(log, ctx, page, log_offset);
1974b4c625c6SSong Liu 	addr = kmap_atomic(page);
1975b4c625c6SSong Liu 	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1976b4c625c6SSong Liu 	kunmap_atomic(addr);
1977b4c625c6SSong Liu 	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1978b4c625c6SSong Liu }
1979b4c625c6SSong Liu 
1980b4c625c6SSong Liu /*
1981b4c625c6SSong Liu  * before loading data to stripe cache, we need verify checksum for all data,
1982b4c625c6SSong Liu  * if there is mismatch for any data page, we drop all data in the mata block
1983b4c625c6SSong Liu  */
1984b4c625c6SSong Liu static int
1985b4c625c6SSong Liu r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1986b4c625c6SSong Liu 					 struct r5l_recovery_ctx *ctx)
1987b4c625c6SSong Liu {
1988b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
1989b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
1990b4c625c6SSong Liu 	struct r5l_meta_block *mb = page_address(ctx->meta_page);
1991b4c625c6SSong Liu 	sector_t mb_offset = sizeof(struct r5l_meta_block);
1992b4c625c6SSong Liu 	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1993b4c625c6SSong Liu 	struct page *page;
1994b4c625c6SSong Liu 	struct r5l_payload_data_parity *payload;
19952d4f4687SSong Liu 	struct r5l_payload_flush *payload_flush;
1996b4c625c6SSong Liu 
1997b4c625c6SSong Liu 	page = alloc_page(GFP_KERNEL);
1998b4c625c6SSong Liu 	if (!page)
1999355810d1SShaohua Li 		return -ENOMEM;
2000355810d1SShaohua Li 
2001b4c625c6SSong Liu 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2002b4c625c6SSong Liu 		payload = (void *)mb + mb_offset;
20032d4f4687SSong Liu 		payload_flush = (void *)mb + mb_offset;
2004b4c625c6SSong Liu 
20051ad45a9bSJason Yan 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2006b4c625c6SSong Liu 			if (r5l_recovery_verify_data_checksum(
2007effe6ee7SSong Liu 				    log, ctx, page, log_offset,
2008b4c625c6SSong Liu 				    payload->checksum[0]) < 0)
2009b4c625c6SSong Liu 				goto mismatch;
20101ad45a9bSJason Yan 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2011b4c625c6SSong Liu 			if (r5l_recovery_verify_data_checksum(
2012effe6ee7SSong Liu 				    log, ctx, page, log_offset,
2013b4c625c6SSong Liu 				    payload->checksum[0]) < 0)
2014b4c625c6SSong Liu 				goto mismatch;
2015b4c625c6SSong Liu 			if (conf->max_degraded == 2 && /* q for RAID 6 */
2016b4c625c6SSong Liu 			    r5l_recovery_verify_data_checksum(
2017effe6ee7SSong Liu 				    log, ctx, page,
2018b4c625c6SSong Liu 				    r5l_ring_add(log, log_offset,
2019b4c625c6SSong Liu 						 BLOCK_SECTORS),
2020b4c625c6SSong Liu 				    payload->checksum[1]) < 0)
2021b4c625c6SSong Liu 				goto mismatch;
20221ad45a9bSJason Yan 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
20232d4f4687SSong Liu 			/* nothing to do for R5LOG_PAYLOAD_FLUSH here */
20242d4f4687SSong Liu 		} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2025b4c625c6SSong Liu 			goto mismatch;
2026b4c625c6SSong Liu 
20271ad45a9bSJason Yan 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
20282d4f4687SSong Liu 			mb_offset += sizeof(struct r5l_payload_flush) +
20292d4f4687SSong Liu 				le32_to_cpu(payload_flush->size);
20302d4f4687SSong Liu 		} else {
20312d4f4687SSong Liu 			/* DATA or PARITY payload */
2032b4c625c6SSong Liu 			log_offset = r5l_ring_add(log, log_offset,
2033b4c625c6SSong Liu 						  le32_to_cpu(payload->size));
2034b4c625c6SSong Liu 			mb_offset += sizeof(struct r5l_payload_data_parity) +
2035b4c625c6SSong Liu 				sizeof(__le32) *
2036b4c625c6SSong Liu 				(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2037b4c625c6SSong Liu 		}
2038b4c625c6SSong Liu 
20392d4f4687SSong Liu 	}
20402d4f4687SSong Liu 
2041b4c625c6SSong Liu 	put_page(page);
2042b4c625c6SSong Liu 	return 0;
2043b4c625c6SSong Liu 
2044b4c625c6SSong Liu mismatch:
2045b4c625c6SSong Liu 	put_page(page);
2046b4c625c6SSong Liu 	return -EINVAL;
2047b4c625c6SSong Liu }
2048b4c625c6SSong Liu 
2049b4c625c6SSong Liu /*
2050b4c625c6SSong Liu  * Analyze all data/parity pages in one meta block
2051b4c625c6SSong Liu  * Returns:
2052b4c625c6SSong Liu  * 0 for success
2053b4c625c6SSong Liu  * -EINVAL for unknown playload type
2054b4c625c6SSong Liu  * -EAGAIN for checksum mismatch of data page
2055b4c625c6SSong Liu  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2056b4c625c6SSong Liu  */
2057b4c625c6SSong Liu static int
2058b4c625c6SSong Liu r5c_recovery_analyze_meta_block(struct r5l_log *log,
2059b4c625c6SSong Liu 				struct r5l_recovery_ctx *ctx,
2060b4c625c6SSong Liu 				struct list_head *cached_stripe_list)
2061b4c625c6SSong Liu {
2062b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2063b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
2064b4c625c6SSong Liu 	struct r5l_meta_block *mb;
2065b4c625c6SSong Liu 	struct r5l_payload_data_parity *payload;
20662d4f4687SSong Liu 	struct r5l_payload_flush *payload_flush;
2067b4c625c6SSong Liu 	int mb_offset;
2068b4c625c6SSong Liu 	sector_t log_offset;
2069b4c625c6SSong Liu 	sector_t stripe_sect;
2070b4c625c6SSong Liu 	struct stripe_head *sh;
2071b4c625c6SSong Liu 	int ret;
2072b4c625c6SSong Liu 
2073b4c625c6SSong Liu 	/*
2074b4c625c6SSong Liu 	 * for mismatch in data blocks, we will drop all data in this mb, but
2075b4c625c6SSong Liu 	 * we will still read next mb for other data with FLUSH flag, as
2076b4c625c6SSong Liu 	 * io_unit could finish out of order.
2077b4c625c6SSong Liu 	 */
2078b4c625c6SSong Liu 	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2079b4c625c6SSong Liu 	if (ret == -EINVAL)
2080b4c625c6SSong Liu 		return -EAGAIN;
2081b4c625c6SSong Liu 	else if (ret)
2082b4c625c6SSong Liu 		return ret;   /* -ENOMEM duo to alloc_page() failed */
2083b4c625c6SSong Liu 
2084b4c625c6SSong Liu 	mb = page_address(ctx->meta_page);
2085b4c625c6SSong Liu 	mb_offset = sizeof(struct r5l_meta_block);
2086b4c625c6SSong Liu 	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2087b4c625c6SSong Liu 
2088b4c625c6SSong Liu 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2089b4c625c6SSong Liu 		int dd;
2090b4c625c6SSong Liu 
2091b4c625c6SSong Liu 		payload = (void *)mb + mb_offset;
20922d4f4687SSong Liu 		payload_flush = (void *)mb + mb_offset;
20932d4f4687SSong Liu 
20941ad45a9bSJason Yan 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
20952d4f4687SSong Liu 			int i, count;
20962d4f4687SSong Liu 
20972d4f4687SSong Liu 			count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
20982d4f4687SSong Liu 			for (i = 0; i < count; ++i) {
20992d4f4687SSong Liu 				stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
21002d4f4687SSong Liu 				sh = r5c_recovery_lookup_stripe(cached_stripe_list,
21012d4f4687SSong Liu 								stripe_sect);
21022d4f4687SSong Liu 				if (sh) {
21032d4f4687SSong Liu 					WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
21042d4f4687SSong Liu 					r5l_recovery_reset_stripe(sh);
21052d4f4687SSong Liu 					list_del_init(&sh->lru);
21062d4f4687SSong Liu 					raid5_release_stripe(sh);
21072d4f4687SSong Liu 				}
21082d4f4687SSong Liu 			}
21092d4f4687SSong Liu 
21102d4f4687SSong Liu 			mb_offset += sizeof(struct r5l_payload_flush) +
21112d4f4687SSong Liu 				le32_to_cpu(payload_flush->size);
21122d4f4687SSong Liu 			continue;
21132d4f4687SSong Liu 		}
21142d4f4687SSong Liu 
21152d4f4687SSong Liu 		/* DATA or PARITY payload */
21161ad45a9bSJason Yan 		stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2117b4c625c6SSong Liu 			raid5_compute_sector(
2118b4c625c6SSong Liu 				conf, le64_to_cpu(payload->location), 0, &dd,
2119b4c625c6SSong Liu 				NULL)
2120b4c625c6SSong Liu 			: le64_to_cpu(payload->location);
2121b4c625c6SSong Liu 
2122b4c625c6SSong Liu 		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2123b4c625c6SSong Liu 						stripe_sect);
2124b4c625c6SSong Liu 
2125b4c625c6SSong Liu 		if (!sh) {
21263c66abbaSSong Liu 			sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
2127b4c625c6SSong Liu 			/*
2128b4c625c6SSong Liu 			 * cannot get stripe from raid5_get_active_stripe
2129b4c625c6SSong Liu 			 * try replay some stripes
2130b4c625c6SSong Liu 			 */
2131b4c625c6SSong Liu 			if (!sh) {
2132b4c625c6SSong Liu 				r5c_recovery_replay_stripes(
2133b4c625c6SSong Liu 					cached_stripe_list, ctx);
2134b4c625c6SSong Liu 				sh = r5c_recovery_alloc_stripe(
21353c66abbaSSong Liu 					conf, stripe_sect);
2136b4c625c6SSong Liu 			}
2137b4c625c6SSong Liu 			if (!sh) {
2138b4c625c6SSong Liu 				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2139b4c625c6SSong Liu 					mdname(mddev),
2140b4c625c6SSong Liu 					conf->min_nr_stripes * 2);
2141b4c625c6SSong Liu 				raid5_set_cache_size(mddev,
2142b4c625c6SSong Liu 						     conf->min_nr_stripes * 2);
21433c66abbaSSong Liu 				sh = r5c_recovery_alloc_stripe(conf,
21443c66abbaSSong Liu 							       stripe_sect);
2145b4c625c6SSong Liu 			}
2146b4c625c6SSong Liu 			if (!sh) {
2147b4c625c6SSong Liu 				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2148b4c625c6SSong Liu 				       mdname(mddev));
2149b4c625c6SSong Liu 				return -ENOMEM;
2150b4c625c6SSong Liu 			}
2151b4c625c6SSong Liu 			list_add_tail(&sh->lru, cached_stripe_list);
2152b4c625c6SSong Liu 		}
2153b4c625c6SSong Liu 
21541ad45a9bSJason Yan 		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2155f7b7bee7SZhengyuan Liu 			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2156f7b7bee7SZhengyuan Liu 			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2157b4c625c6SSong Liu 				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2158b4c625c6SSong Liu 				list_move_tail(&sh->lru, cached_stripe_list);
2159b4c625c6SSong Liu 			}
2160b4c625c6SSong Liu 			r5l_recovery_load_data(log, sh, ctx, payload,
2161b4c625c6SSong Liu 					       log_offset);
21621ad45a9bSJason Yan 		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2163b4c625c6SSong Liu 			r5l_recovery_load_parity(log, sh, ctx, payload,
2164b4c625c6SSong Liu 						 log_offset);
2165b4c625c6SSong Liu 		else
2166b4c625c6SSong Liu 			return -EINVAL;
2167b4c625c6SSong Liu 
2168b4c625c6SSong Liu 		log_offset = r5l_ring_add(log, log_offset,
2169b4c625c6SSong Liu 					  le32_to_cpu(payload->size));
2170b4c625c6SSong Liu 
2171b4c625c6SSong Liu 		mb_offset += sizeof(struct r5l_payload_data_parity) +
2172b4c625c6SSong Liu 			sizeof(__le32) *
2173b4c625c6SSong Liu 			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2174b4c625c6SSong Liu 	}
2175b4c625c6SSong Liu 
2176b4c625c6SSong Liu 	return 0;
2177b4c625c6SSong Liu }
2178b4c625c6SSong Liu 
2179b4c625c6SSong Liu /*
2180b4c625c6SSong Liu  * Load the stripe into cache. The stripe will be written out later by
2181b4c625c6SSong Liu  * the stripe cache state machine.
2182b4c625c6SSong Liu  */
2183b4c625c6SSong Liu static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2184b4c625c6SSong Liu 					 struct stripe_head *sh)
2185b4c625c6SSong Liu {
2186b4c625c6SSong Liu 	struct r5dev *dev;
2187b4c625c6SSong Liu 	int i;
2188b4c625c6SSong Liu 
2189b4c625c6SSong Liu 	for (i = sh->disks; i--; ) {
2190b4c625c6SSong Liu 		dev = sh->dev + i;
2191b4c625c6SSong Liu 		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2192b4c625c6SSong Liu 			set_bit(R5_InJournal, &dev->flags);
2193b4c625c6SSong Liu 			set_bit(R5_UPTODATE, &dev->flags);
2194b4c625c6SSong Liu 		}
2195b4c625c6SSong Liu 	}
2196b4c625c6SSong Liu }
2197b4c625c6SSong Liu 
2198b4c625c6SSong Liu /*
2199b4c625c6SSong Liu  * Scan through the log for all to-be-flushed data
2200b4c625c6SSong Liu  *
2201b4c625c6SSong Liu  * For stripes with data and parity, namely Data-Parity stripe
2202b4c625c6SSong Liu  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2203b4c625c6SSong Liu  *
2204b4c625c6SSong Liu  * For stripes with only data, namely Data-Only stripe
2205b4c625c6SSong Liu  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2206b4c625c6SSong Liu  *
2207b4c625c6SSong Liu  * For a stripe, if we see data after parity, we should discard all previous
2208b4c625c6SSong Liu  * data and parity for this stripe, as these data are already flushed to
2209b4c625c6SSong Liu  * the array.
2210b4c625c6SSong Liu  *
2211b4c625c6SSong Liu  * At the end of the scan, we return the new journal_tail, which points to
2212b4c625c6SSong Liu  * first data-only stripe on the journal device, or next invalid meta block.
2213b4c625c6SSong Liu  */
2214b4c625c6SSong Liu static int r5c_recovery_flush_log(struct r5l_log *log,
2215b4c625c6SSong Liu 				  struct r5l_recovery_ctx *ctx)
2216b4c625c6SSong Liu {
2217bc8f167fSJackieLiu 	struct stripe_head *sh;
2218b4c625c6SSong Liu 	int ret = 0;
2219b4c625c6SSong Liu 
2220b4c625c6SSong Liu 	/* scan through the log */
2221b4c625c6SSong Liu 	while (1) {
2222b4c625c6SSong Liu 		if (r5l_recovery_read_meta_block(log, ctx))
2223b4c625c6SSong Liu 			break;
2224b4c625c6SSong Liu 
2225b4c625c6SSong Liu 		ret = r5c_recovery_analyze_meta_block(log, ctx,
2226b4c625c6SSong Liu 						      &ctx->cached_list);
2227b4c625c6SSong Liu 		/*
2228b4c625c6SSong Liu 		 * -EAGAIN means mismatch in data block, in this case, we still
2229b4c625c6SSong Liu 		 * try scan the next metablock
2230b4c625c6SSong Liu 		 */
2231b4c625c6SSong Liu 		if (ret && ret != -EAGAIN)
2232b4c625c6SSong Liu 			break;   /* ret == -EINVAL or -ENOMEM */
2233b4c625c6SSong Liu 		ctx->seq++;
2234b4c625c6SSong Liu 		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2235b4c625c6SSong Liu 	}
2236b4c625c6SSong Liu 
2237b4c625c6SSong Liu 	if (ret == -ENOMEM) {
2238b4c625c6SSong Liu 		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2239b4c625c6SSong Liu 		return ret;
2240b4c625c6SSong Liu 	}
2241b4c625c6SSong Liu 
2242b4c625c6SSong Liu 	/* replay data-parity stripes */
2243b4c625c6SSong Liu 	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2244b4c625c6SSong Liu 
2245b4c625c6SSong Liu 	/* load data-only stripes to stripe cache */
2246bc8f167fSJackieLiu 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2247b4c625c6SSong Liu 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2248b4c625c6SSong Liu 		r5c_recovery_load_one_stripe(log, sh);
2249b4c625c6SSong Liu 		ctx->data_only_stripes++;
2250b4c625c6SSong Liu 	}
2251b4c625c6SSong Liu 
2252b4c625c6SSong Liu 	return 0;
2253b4c625c6SSong Liu }
2254355810d1SShaohua Li 
2255355810d1SShaohua Li /*
2256355810d1SShaohua Li  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2257355810d1SShaohua Li  * log will start here. but we can't let superblock point to last valid
2258355810d1SShaohua Li  * meta block. The log might looks like:
2259355810d1SShaohua Li  * | meta 1| meta 2| meta 3|
2260355810d1SShaohua Li  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2261355810d1SShaohua Li  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2262355810d1SShaohua Li  * happens again, new recovery will start from meta 1. Since meta 2n is
2263355810d1SShaohua Li  * valid now, recovery will think meta 3 is valid, which is wrong.
2264355810d1SShaohua Li  * The solution is we create a new meta in meta2 with its seq == meta
22653c6edc66SSong Liu  * 1's seq + 10000 and let superblock points to meta2. The same recovery
22663c6edc66SSong Liu  * will not think meta 3 is a valid meta, because its seq doesn't match
2267355810d1SShaohua Li  */
2268355810d1SShaohua Li 
2269b4c625c6SSong Liu /*
2270b4c625c6SSong Liu  * Before recovery, the log looks like the following
2271b4c625c6SSong Liu  *
2272b4c625c6SSong Liu  *   ---------------------------------------------
2273b4c625c6SSong Liu  *   |           valid log        | invalid log  |
2274b4c625c6SSong Liu  *   ---------------------------------------------
2275b4c625c6SSong Liu  *   ^
2276b4c625c6SSong Liu  *   |- log->last_checkpoint
2277b4c625c6SSong Liu  *   |- log->last_cp_seq
2278b4c625c6SSong Liu  *
2279b4c625c6SSong Liu  * Now we scan through the log until we see invalid entry
2280b4c625c6SSong Liu  *
2281b4c625c6SSong Liu  *   ---------------------------------------------
2282b4c625c6SSong Liu  *   |           valid log        | invalid log  |
2283b4c625c6SSong Liu  *   ---------------------------------------------
2284b4c625c6SSong Liu  *   ^                            ^
2285b4c625c6SSong Liu  *   |- log->last_checkpoint      |- ctx->pos
2286b4c625c6SSong Liu  *   |- log->last_cp_seq          |- ctx->seq
2287b4c625c6SSong Liu  *
2288b4c625c6SSong Liu  * From this point, we need to increase seq number by 10 to avoid
2289b4c625c6SSong Liu  * confusing next recovery.
2290b4c625c6SSong Liu  *
2291b4c625c6SSong Liu  *   ---------------------------------------------
2292b4c625c6SSong Liu  *   |           valid log        | invalid log  |
2293b4c625c6SSong Liu  *   ---------------------------------------------
2294b4c625c6SSong Liu  *   ^                              ^
2295b4c625c6SSong Liu  *   |- log->last_checkpoint        |- ctx->pos+1
22963c6edc66SSong Liu  *   |- log->last_cp_seq            |- ctx->seq+10001
2297b4c625c6SSong Liu  *
2298b4c625c6SSong Liu  * However, it is not safe to start the state machine yet, because data only
2299b4c625c6SSong Liu  * parities are not yet secured in RAID. To save these data only parities, we
2300b4c625c6SSong Liu  * rewrite them from seq+11.
2301b4c625c6SSong Liu  *
2302b4c625c6SSong Liu  *   -----------------------------------------------------------------
2303b4c625c6SSong Liu  *   |           valid log        | data only stripes | invalid log  |
2304b4c625c6SSong Liu  *   -----------------------------------------------------------------
2305b4c625c6SSong Liu  *   ^                                                ^
2306b4c625c6SSong Liu  *   |- log->last_checkpoint                          |- ctx->pos+n
23073c6edc66SSong Liu  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2308b4c625c6SSong Liu  *
2309b4c625c6SSong Liu  * If failure happens again during this process, the recovery can safe start
2310b4c625c6SSong Liu  * again from log->last_checkpoint.
2311b4c625c6SSong Liu  *
2312b4c625c6SSong Liu  * Once data only stripes are rewritten to journal, we move log_tail
2313b4c625c6SSong Liu  *
2314b4c625c6SSong Liu  *   -----------------------------------------------------------------
2315b4c625c6SSong Liu  *   |     old log        |    data only stripes    | invalid log  |
2316b4c625c6SSong Liu  *   -----------------------------------------------------------------
2317b4c625c6SSong Liu  *                        ^                         ^
2318b4c625c6SSong Liu  *                        |- log->last_checkpoint   |- ctx->pos+n
23193c6edc66SSong Liu  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2320b4c625c6SSong Liu  *
2321b4c625c6SSong Liu  * Then we can safely start the state machine. If failure happens from this
2322b4c625c6SSong Liu  * point on, the recovery will start from new log->last_checkpoint.
2323b4c625c6SSong Liu  */
2324b4c625c6SSong Liu static int
2325b4c625c6SSong Liu r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2326b4c625c6SSong Liu 				       struct r5l_recovery_ctx *ctx)
2327b4c625c6SSong Liu {
2328a85dd7b8SSong Liu 	struct stripe_head *sh;
2329b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2330b4c625c6SSong Liu 	struct page *page;
23313c66abbaSSong Liu 	sector_t next_checkpoint = MaxSector;
2332b4c625c6SSong Liu 
2333b4c625c6SSong Liu 	page = alloc_page(GFP_KERNEL);
2334b4c625c6SSong Liu 	if (!page) {
2335b4c625c6SSong Liu 		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2336b4c625c6SSong Liu 		       mdname(mddev));
2337b4c625c6SSong Liu 		return -ENOMEM;
2338b4c625c6SSong Liu 	}
2339b4c625c6SSong Liu 
23403c66abbaSSong Liu 	WARN_ON(list_empty(&ctx->cached_list));
23413c66abbaSSong Liu 
2342a85dd7b8SSong Liu 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2343b4c625c6SSong Liu 		struct r5l_meta_block *mb;
2344b4c625c6SSong Liu 		int i;
2345b4c625c6SSong Liu 		int offset;
2346b4c625c6SSong Liu 		sector_t write_pos;
2347b4c625c6SSong Liu 
2348b4c625c6SSong Liu 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2349b4c625c6SSong Liu 		r5l_recovery_create_empty_meta_block(log, page,
2350b4c625c6SSong Liu 						     ctx->pos, ctx->seq);
2351b4c625c6SSong Liu 		mb = page_address(page);
2352b4c625c6SSong Liu 		offset = le32_to_cpu(mb->meta_size);
2353fc833c2aSJackieLiu 		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2354b4c625c6SSong Liu 
2355b4c625c6SSong Liu 		for (i = sh->disks; i--; ) {
2356b4c625c6SSong Liu 			struct r5dev *dev = &sh->dev[i];
2357b4c625c6SSong Liu 			struct r5l_payload_data_parity *payload;
2358b4c625c6SSong Liu 			void *addr;
2359b4c625c6SSong Liu 
2360b4c625c6SSong Liu 			if (test_bit(R5_InJournal, &dev->flags)) {
2361b4c625c6SSong Liu 				payload = (void *)mb + offset;
2362b4c625c6SSong Liu 				payload->header.type = cpu_to_le16(
2363b4c625c6SSong Liu 					R5LOG_PAYLOAD_DATA);
23641ad45a9bSJason Yan 				payload->size = cpu_to_le32(BLOCK_SECTORS);
2365b4c625c6SSong Liu 				payload->location = cpu_to_le64(
2366b4c625c6SSong Liu 					raid5_compute_blocknr(sh, i, 0));
2367b4c625c6SSong Liu 				addr = kmap_atomic(dev->page);
2368b4c625c6SSong Liu 				payload->checksum[0] = cpu_to_le32(
2369b4c625c6SSong Liu 					crc32c_le(log->uuid_checksum, addr,
2370b4c625c6SSong Liu 						  PAGE_SIZE));
2371b4c625c6SSong Liu 				kunmap_atomic(addr);
2372b4c625c6SSong Liu 				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2373b4c625c6SSong Liu 					     dev->page, REQ_OP_WRITE, 0, false);
2374b4c625c6SSong Liu 				write_pos = r5l_ring_add(log, write_pos,
2375b4c625c6SSong Liu 							 BLOCK_SECTORS);
2376b4c625c6SSong Liu 				offset += sizeof(__le32) +
2377b4c625c6SSong Liu 					sizeof(struct r5l_payload_data_parity);
2378b4c625c6SSong Liu 
2379b4c625c6SSong Liu 			}
2380b4c625c6SSong Liu 		}
2381b4c625c6SSong Liu 		mb->meta_size = cpu_to_le32(offset);
23825c88f403SSong Liu 		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
23835c88f403SSong Liu 						     mb, PAGE_SIZE));
2384b4c625c6SSong Liu 		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
238520737738SShaohua Li 			     REQ_OP_WRITE, REQ_FUA, false);
2386b4c625c6SSong Liu 		sh->log_start = ctx->pos;
23873c66abbaSSong Liu 		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
23883c66abbaSSong Liu 		atomic_inc(&log->stripe_in_journal_count);
2389b4c625c6SSong Liu 		ctx->pos = write_pos;
2390b4c625c6SSong Liu 		ctx->seq += 1;
23913c66abbaSSong Liu 		next_checkpoint = sh->log_start;
2392b4c625c6SSong Liu 	}
23933c66abbaSSong Liu 	log->next_checkpoint = next_checkpoint;
2394b4c625c6SSong Liu 	__free_page(page);
2395b4c625c6SSong Liu 	return 0;
2396b4c625c6SSong Liu }
2397b4c625c6SSong Liu 
2398a85dd7b8SSong Liu static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2399a85dd7b8SSong Liu 						 struct r5l_recovery_ctx *ctx)
2400a85dd7b8SSong Liu {
2401a85dd7b8SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2402a85dd7b8SSong Liu 	struct r5conf *conf = mddev->private;
2403a85dd7b8SSong Liu 	struct stripe_head *sh, *next;
2404a85dd7b8SSong Liu 
2405a85dd7b8SSong Liu 	if (ctx->data_only_stripes == 0)
2406a85dd7b8SSong Liu 		return;
2407a85dd7b8SSong Liu 
2408a85dd7b8SSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2409a85dd7b8SSong Liu 
2410a85dd7b8SSong Liu 	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2411a85dd7b8SSong Liu 		r5c_make_stripe_write_out(sh);
2412a85dd7b8SSong Liu 		set_bit(STRIPE_HANDLE, &sh->state);
2413a85dd7b8SSong Liu 		list_del_init(&sh->lru);
2414a85dd7b8SSong Liu 		raid5_release_stripe(sh);
2415a85dd7b8SSong Liu 	}
2416a85dd7b8SSong Liu 
2417a85dd7b8SSong Liu 	md_wakeup_thread(conf->mddev->thread);
2418a85dd7b8SSong Liu 	/* reuse conf->wait_for_quiescent in recovery */
2419a85dd7b8SSong Liu 	wait_event(conf->wait_for_quiescent,
2420a85dd7b8SSong Liu 		   atomic_read(&conf->active_stripes) == 0);
2421a85dd7b8SSong Liu 
2422a85dd7b8SSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2423a85dd7b8SSong Liu }
2424a85dd7b8SSong Liu 
2425f6bed0efSShaohua Li static int r5l_recovery_log(struct r5l_log *log)
2426f6bed0efSShaohua Li {
24275aabf7c4SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2428effe6ee7SSong Liu 	struct r5l_recovery_ctx *ctx;
24295aabf7c4SSong Liu 	int ret;
243043b96748SJackieLiu 	sector_t pos;
2431355810d1SShaohua Li 
2432effe6ee7SSong Liu 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2433effe6ee7SSong Liu 	if (!ctx)
2434355810d1SShaohua Li 		return -ENOMEM;
2435355810d1SShaohua Li 
2436effe6ee7SSong Liu 	ctx->pos = log->last_checkpoint;
2437effe6ee7SSong Liu 	ctx->seq = log->last_cp_seq;
2438effe6ee7SSong Liu 	INIT_LIST_HEAD(&ctx->cached_list);
2439effe6ee7SSong Liu 	ctx->meta_page = alloc_page(GFP_KERNEL);
2440effe6ee7SSong Liu 
2441effe6ee7SSong Liu 	if (!ctx->meta_page) {
2442effe6ee7SSong Liu 		ret =  -ENOMEM;
2443effe6ee7SSong Liu 		goto meta_page;
2444effe6ee7SSong Liu 	}
2445effe6ee7SSong Liu 
2446effe6ee7SSong Liu 	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2447effe6ee7SSong Liu 		ret = -ENOMEM;
2448effe6ee7SSong Liu 		goto ra_pool;
2449effe6ee7SSong Liu 	}
2450effe6ee7SSong Liu 
2451effe6ee7SSong Liu 	ret = r5c_recovery_flush_log(log, ctx);
2452355810d1SShaohua Li 
2453355810d1SShaohua Li 	if (ret)
2454effe6ee7SSong Liu 		goto error;
24555aabf7c4SSong Liu 
2456effe6ee7SSong Liu 	pos = ctx->pos;
2457effe6ee7SSong Liu 	ctx->seq += 10000;
245843b96748SJackieLiu 
2459effe6ee7SSong Liu 	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
24605aabf7c4SSong Liu 		pr_debug("md/raid:%s: starting from clean shutdown\n",
24615aabf7c4SSong Liu 			 mdname(mddev));
2462a85dd7b8SSong Liu 	else
246399f17890SColin Ian King 		pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2464effe6ee7SSong Liu 			 mdname(mddev), ctx->data_only_stripes,
2465effe6ee7SSong Liu 			 ctx->data_parity_stripes);
24665aabf7c4SSong Liu 
2467effe6ee7SSong Liu 	if (ctx->data_only_stripes == 0) {
2468effe6ee7SSong Liu 		log->next_checkpoint = ctx->pos;
2469effe6ee7SSong Liu 		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2470effe6ee7SSong Liu 		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2471effe6ee7SSong Liu 	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
24725aabf7c4SSong Liu 		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
24735aabf7c4SSong Liu 		       mdname(mddev));
2474effe6ee7SSong Liu 		ret =  -EIO;
2475effe6ee7SSong Liu 		goto error;
24765aabf7c4SSong Liu 	}
24775aabf7c4SSong Liu 
2478effe6ee7SSong Liu 	log->log_start = ctx->pos;
2479effe6ee7SSong Liu 	log->seq = ctx->seq;
248043b96748SJackieLiu 	log->last_checkpoint = pos;
248143b96748SJackieLiu 	r5l_write_super(log, pos);
2482a85dd7b8SSong Liu 
2483effe6ee7SSong Liu 	r5c_recovery_flush_data_only_stripes(log, ctx);
2484effe6ee7SSong Liu 	ret = 0;
2485effe6ee7SSong Liu error:
2486effe6ee7SSong Liu 	r5l_recovery_free_ra_pool(log, ctx);
2487effe6ee7SSong Liu ra_pool:
2488effe6ee7SSong Liu 	__free_page(ctx->meta_page);
2489effe6ee7SSong Liu meta_page:
2490effe6ee7SSong Liu 	kfree(ctx);
2491effe6ee7SSong Liu 	return ret;
2492f6bed0efSShaohua Li }
2493f6bed0efSShaohua Li 
2494f6bed0efSShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp)
2495f6bed0efSShaohua Li {
2496f6bed0efSShaohua Li 	struct mddev *mddev = log->rdev->mddev;
2497f6bed0efSShaohua Li 
2498f6bed0efSShaohua Li 	log->rdev->journal_tail = cp;
24992953079cSShaohua Li 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2500f6bed0efSShaohua Li }
2501f6bed0efSShaohua Li 
25022c7da14bSSong Liu static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
25032c7da14bSSong Liu {
25042c7da14bSSong Liu 	struct r5conf *conf = mddev->private;
25052c7da14bSSong Liu 	int ret;
25062c7da14bSSong Liu 
25072c7da14bSSong Liu 	if (!conf->log)
25082c7da14bSSong Liu 		return 0;
25092c7da14bSSong Liu 
25102c7da14bSSong Liu 	switch (conf->log->r5c_journal_mode) {
25112c7da14bSSong Liu 	case R5C_JOURNAL_MODE_WRITE_THROUGH:
25122c7da14bSSong Liu 		ret = snprintf(
25132c7da14bSSong Liu 			page, PAGE_SIZE, "[%s] %s\n",
25142c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
25152c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
25162c7da14bSSong Liu 		break;
25172c7da14bSSong Liu 	case R5C_JOURNAL_MODE_WRITE_BACK:
25182c7da14bSSong Liu 		ret = snprintf(
25192c7da14bSSong Liu 			page, PAGE_SIZE, "%s [%s]\n",
25202c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
25212c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
25222c7da14bSSong Liu 		break;
25232c7da14bSSong Liu 	default:
25242c7da14bSSong Liu 		ret = 0;
25252c7da14bSSong Liu 	}
25262c7da14bSSong Liu 	return ret;
25272c7da14bSSong Liu }
25282c7da14bSSong Liu 
252978e470c2SHeinz Mauelshagen /*
253078e470c2SHeinz Mauelshagen  * Set journal cache mode on @mddev (external API initially needed by dm-raid).
253178e470c2SHeinz Mauelshagen  *
253278e470c2SHeinz Mauelshagen  * @mode as defined in 'enum r5c_journal_mode'.
253378e470c2SHeinz Mauelshagen  *
253478e470c2SHeinz Mauelshagen  */
253578e470c2SHeinz Mauelshagen int r5c_journal_mode_set(struct mddev *mddev, int mode)
25362c7da14bSSong Liu {
25372c7da14bSSong Liu 	struct r5conf *conf = mddev->private;
25382c7da14bSSong Liu 	struct r5l_log *log = conf->log;
25392c7da14bSSong Liu 
25402c7da14bSSong Liu 	if (!log)
25412c7da14bSSong Liu 		return -ENODEV;
25422c7da14bSSong Liu 
254378e470c2SHeinz Mauelshagen 	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
254478e470c2SHeinz Mauelshagen 	    mode > R5C_JOURNAL_MODE_WRITE_BACK)
25452c7da14bSSong Liu 		return -EINVAL;
25462c7da14bSSong Liu 
25472e38a37fSSong Liu 	if (raid5_calc_degraded(conf) > 0 &&
254878e470c2SHeinz Mauelshagen 	    mode == R5C_JOURNAL_MODE_WRITE_BACK)
25492e38a37fSSong Liu 		return -EINVAL;
25502e38a37fSSong Liu 
25512c7da14bSSong Liu 	mddev_suspend(mddev);
255278e470c2SHeinz Mauelshagen 	conf->log->r5c_journal_mode = mode;
25532c7da14bSSong Liu 	mddev_resume(mddev);
25542c7da14bSSong Liu 
25552c7da14bSSong Liu 	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
255678e470c2SHeinz Mauelshagen 		 mdname(mddev), mode, r5c_journal_mode_str[mode]);
255778e470c2SHeinz Mauelshagen 	return 0;
255878e470c2SHeinz Mauelshagen }
255978e470c2SHeinz Mauelshagen EXPORT_SYMBOL(r5c_journal_mode_set);
256078e470c2SHeinz Mauelshagen 
256178e470c2SHeinz Mauelshagen static ssize_t r5c_journal_mode_store(struct mddev *mddev,
256278e470c2SHeinz Mauelshagen 				      const char *page, size_t length)
256378e470c2SHeinz Mauelshagen {
256478e470c2SHeinz Mauelshagen 	int mode = ARRAY_SIZE(r5c_journal_mode_str);
256578e470c2SHeinz Mauelshagen 	size_t len = length;
256678e470c2SHeinz Mauelshagen 
256778e470c2SHeinz Mauelshagen 	if (len < 2)
256878e470c2SHeinz Mauelshagen 		return -EINVAL;
256978e470c2SHeinz Mauelshagen 
257078e470c2SHeinz Mauelshagen 	if (page[len - 1] == '\n')
257178e470c2SHeinz Mauelshagen 		len--;
257278e470c2SHeinz Mauelshagen 
257378e470c2SHeinz Mauelshagen 	while (mode--)
257478e470c2SHeinz Mauelshagen 		if (strlen(r5c_journal_mode_str[mode]) == len &&
257578e470c2SHeinz Mauelshagen 		    !strncmp(page, r5c_journal_mode_str[mode], len))
257678e470c2SHeinz Mauelshagen 			break;
257778e470c2SHeinz Mauelshagen 
257878e470c2SHeinz Mauelshagen 	return r5c_journal_mode_set(mddev, mode) ?: length;
25792c7da14bSSong Liu }
25802c7da14bSSong Liu 
25812c7da14bSSong Liu struct md_sysfs_entry
25822c7da14bSSong Liu r5c_journal_mode = __ATTR(journal_mode, 0644,
25832c7da14bSSong Liu 			  r5c_journal_mode_show, r5c_journal_mode_store);
25842c7da14bSSong Liu 
25852ded3703SSong Liu /*
25862ded3703SSong Liu  * Try handle write operation in caching phase. This function should only
25872ded3703SSong Liu  * be called in write-back mode.
25882ded3703SSong Liu  *
25892ded3703SSong Liu  * If all outstanding writes can be handled in caching phase, returns 0
25902ded3703SSong Liu  * If writes requires write-out phase, call r5c_make_stripe_write_out()
25912ded3703SSong Liu  * and returns -EAGAIN
25922ded3703SSong Liu  */
25932ded3703SSong Liu int r5c_try_caching_write(struct r5conf *conf,
25942ded3703SSong Liu 			  struct stripe_head *sh,
25952ded3703SSong Liu 			  struct stripe_head_state *s,
25962ded3703SSong Liu 			  int disks)
25972ded3703SSong Liu {
25982ded3703SSong Liu 	struct r5l_log *log = conf->log;
25991e6d690bSSong Liu 	int i;
26001e6d690bSSong Liu 	struct r5dev *dev;
26011e6d690bSSong Liu 	int to_cache = 0;
260203b047f4SSong Liu 	void **pslot;
260303b047f4SSong Liu 	sector_t tree_index;
260403b047f4SSong Liu 	int ret;
260503b047f4SSong Liu 	uintptr_t refcount;
26062ded3703SSong Liu 
26072ded3703SSong Liu 	BUG_ON(!r5c_is_writeback(log));
26082ded3703SSong Liu 
26091e6d690bSSong Liu 	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
26101e6d690bSSong Liu 		/*
26111e6d690bSSong Liu 		 * There are two different scenarios here:
26121e6d690bSSong Liu 		 *  1. The stripe has some data cached, and it is sent to
26131e6d690bSSong Liu 		 *     write-out phase for reclaim
26141e6d690bSSong Liu 		 *  2. The stripe is clean, and this is the first write
26151e6d690bSSong Liu 		 *
26161e6d690bSSong Liu 		 * For 1, return -EAGAIN, so we continue with
26171e6d690bSSong Liu 		 * handle_stripe_dirtying().
26181e6d690bSSong Liu 		 *
26191e6d690bSSong Liu 		 * For 2, set STRIPE_R5C_CACHING and continue with caching
26201e6d690bSSong Liu 		 * write.
26211e6d690bSSong Liu 		 */
26221e6d690bSSong Liu 
26231e6d690bSSong Liu 		/* case 1: anything injournal or anything in written */
26241e6d690bSSong Liu 		if (s->injournal > 0 || s->written > 0)
26251e6d690bSSong Liu 			return -EAGAIN;
26261e6d690bSSong Liu 		/* case 2 */
26271e6d690bSSong Liu 		set_bit(STRIPE_R5C_CACHING, &sh->state);
26281e6d690bSSong Liu 	}
26291e6d690bSSong Liu 
26302e38a37fSSong Liu 	/*
26312e38a37fSSong Liu 	 * When run in degraded mode, array is set to write-through mode.
26322e38a37fSSong Liu 	 * This check helps drain pending write safely in the transition to
26332e38a37fSSong Liu 	 * write-through mode.
26342e38a37fSSong Liu 	 */
26352e38a37fSSong Liu 	if (s->failed) {
26362e38a37fSSong Liu 		r5c_make_stripe_write_out(sh);
26372e38a37fSSong Liu 		return -EAGAIN;
26382e38a37fSSong Liu 	}
26392e38a37fSSong Liu 
26401e6d690bSSong Liu 	for (i = disks; i--; ) {
26411e6d690bSSong Liu 		dev = &sh->dev[i];
26421e6d690bSSong Liu 		/* if non-overwrite, use writing-out phase */
26431e6d690bSSong Liu 		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
26441e6d690bSSong Liu 		    !test_bit(R5_InJournal, &dev->flags)) {
26452ded3703SSong Liu 			r5c_make_stripe_write_out(sh);
26462ded3703SSong Liu 			return -EAGAIN;
26472ded3703SSong Liu 		}
26481e6d690bSSong Liu 	}
26491e6d690bSSong Liu 
265003b047f4SSong Liu 	/* if the stripe is not counted in big_stripe_tree, add it now */
265103b047f4SSong Liu 	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
265203b047f4SSong Liu 	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
265303b047f4SSong Liu 		tree_index = r5c_tree_index(conf, sh->sector);
265403b047f4SSong Liu 		spin_lock(&log->tree_lock);
265503b047f4SSong Liu 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
265603b047f4SSong Liu 					       tree_index);
265703b047f4SSong Liu 		if (pslot) {
265803b047f4SSong Liu 			refcount = (uintptr_t)radix_tree_deref_slot_protected(
265903b047f4SSong Liu 				pslot, &log->tree_lock) >>
266003b047f4SSong Liu 				R5C_RADIX_COUNT_SHIFT;
266103b047f4SSong Liu 			radix_tree_replace_slot(
266203b047f4SSong Liu 				&log->big_stripe_tree, pslot,
266303b047f4SSong Liu 				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
266403b047f4SSong Liu 		} else {
266503b047f4SSong Liu 			/*
266603b047f4SSong Liu 			 * this radix_tree_insert can fail safely, so no
266703b047f4SSong Liu 			 * need to call radix_tree_preload()
266803b047f4SSong Liu 			 */
266903b047f4SSong Liu 			ret = radix_tree_insert(
267003b047f4SSong Liu 				&log->big_stripe_tree, tree_index,
267103b047f4SSong Liu 				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
267203b047f4SSong Liu 			if (ret) {
267303b047f4SSong Liu 				spin_unlock(&log->tree_lock);
267403b047f4SSong Liu 				r5c_make_stripe_write_out(sh);
267503b047f4SSong Liu 				return -EAGAIN;
267603b047f4SSong Liu 			}
267703b047f4SSong Liu 		}
267803b047f4SSong Liu 		spin_unlock(&log->tree_lock);
267903b047f4SSong Liu 
268003b047f4SSong Liu 		/*
268103b047f4SSong Liu 		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
268203b047f4SSong Liu 		 * counted in the radix tree
268303b047f4SSong Liu 		 */
268403b047f4SSong Liu 		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
268503b047f4SSong Liu 		atomic_inc(&conf->r5c_cached_partial_stripes);
268603b047f4SSong Liu 	}
268703b047f4SSong Liu 
26881e6d690bSSong Liu 	for (i = disks; i--; ) {
26891e6d690bSSong Liu 		dev = &sh->dev[i];
26901e6d690bSSong Liu 		if (dev->towrite) {
26911e6d690bSSong Liu 			set_bit(R5_Wantwrite, &dev->flags);
26921e6d690bSSong Liu 			set_bit(R5_Wantdrain, &dev->flags);
26931e6d690bSSong Liu 			set_bit(R5_LOCKED, &dev->flags);
26941e6d690bSSong Liu 			to_cache++;
26951e6d690bSSong Liu 		}
26961e6d690bSSong Liu 	}
26971e6d690bSSong Liu 
26981e6d690bSSong Liu 	if (to_cache) {
26991e6d690bSSong Liu 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
27001e6d690bSSong Liu 		/*
27011e6d690bSSong Liu 		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
27021e6d690bSSong Liu 		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
27031e6d690bSSong Liu 		 * r5c_handle_data_cached()
27041e6d690bSSong Liu 		 */
27051e6d690bSSong Liu 		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
27061e6d690bSSong Liu 	}
27071e6d690bSSong Liu 
27081e6d690bSSong Liu 	return 0;
27091e6d690bSSong Liu }
27101e6d690bSSong Liu 
27111e6d690bSSong Liu /*
27121e6d690bSSong Liu  * free extra pages (orig_page) we allocated for prexor
27131e6d690bSSong Liu  */
27141e6d690bSSong Liu void r5c_release_extra_page(struct stripe_head *sh)
27151e6d690bSSong Liu {
2716d7bd398eSSong Liu 	struct r5conf *conf = sh->raid_conf;
27171e6d690bSSong Liu 	int i;
2718d7bd398eSSong Liu 	bool using_disk_info_extra_page;
2719d7bd398eSSong Liu 
2720d7bd398eSSong Liu 	using_disk_info_extra_page =
2721d7bd398eSSong Liu 		sh->dev[0].orig_page == conf->disks[0].extra_page;
27221e6d690bSSong Liu 
27231e6d690bSSong Liu 	for (i = sh->disks; i--; )
27241e6d690bSSong Liu 		if (sh->dev[i].page != sh->dev[i].orig_page) {
27251e6d690bSSong Liu 			struct page *p = sh->dev[i].orig_page;
27261e6d690bSSong Liu 
27271e6d690bSSong Liu 			sh->dev[i].orig_page = sh->dev[i].page;
272886aa1397SSong Liu 			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
272986aa1397SSong Liu 
2730d7bd398eSSong Liu 			if (!using_disk_info_extra_page)
27311e6d690bSSong Liu 				put_page(p);
27321e6d690bSSong Liu 		}
2733d7bd398eSSong Liu 
2734d7bd398eSSong Liu 	if (using_disk_info_extra_page) {
2735d7bd398eSSong Liu 		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2736d7bd398eSSong Liu 		md_wakeup_thread(conf->mddev->thread);
2737d7bd398eSSong Liu 	}
2738d7bd398eSSong Liu }
2739d7bd398eSSong Liu 
2740d7bd398eSSong Liu void r5c_use_extra_page(struct stripe_head *sh)
2741d7bd398eSSong Liu {
2742d7bd398eSSong Liu 	struct r5conf *conf = sh->raid_conf;
2743d7bd398eSSong Liu 	int i;
2744d7bd398eSSong Liu 	struct r5dev *dev;
2745d7bd398eSSong Liu 
2746d7bd398eSSong Liu 	for (i = sh->disks; i--; ) {
2747d7bd398eSSong Liu 		dev = &sh->dev[i];
2748d7bd398eSSong Liu 		if (dev->orig_page != dev->page)
2749d7bd398eSSong Liu 			put_page(dev->orig_page);
2750d7bd398eSSong Liu 		dev->orig_page = conf->disks[i].extra_page;
2751d7bd398eSSong Liu 	}
27521e6d690bSSong Liu }
27532ded3703SSong Liu 
27542ded3703SSong Liu /*
27552ded3703SSong Liu  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
27562ded3703SSong Liu  * stripe is committed to RAID disks.
27572ded3703SSong Liu  */
27582ded3703SSong Liu void r5c_finish_stripe_write_out(struct r5conf *conf,
27592ded3703SSong Liu 				 struct stripe_head *sh,
27602ded3703SSong Liu 				 struct stripe_head_state *s)
27612ded3703SSong Liu {
276203b047f4SSong Liu 	struct r5l_log *log = conf->log;
27631e6d690bSSong Liu 	int i;
27641e6d690bSSong Liu 	int do_wakeup = 0;
276503b047f4SSong Liu 	sector_t tree_index;
276603b047f4SSong Liu 	void **pslot;
276703b047f4SSong Liu 	uintptr_t refcount;
27681e6d690bSSong Liu 
276903b047f4SSong Liu 	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
27702ded3703SSong Liu 		return;
27712ded3703SSong Liu 
27722ded3703SSong Liu 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
27732ded3703SSong Liu 	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
27742ded3703SSong Liu 
277503b047f4SSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
27762ded3703SSong Liu 		return;
27771e6d690bSSong Liu 
27781e6d690bSSong Liu 	for (i = sh->disks; i--; ) {
27791e6d690bSSong Liu 		clear_bit(R5_InJournal, &sh->dev[i].flags);
27801e6d690bSSong Liu 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
27811e6d690bSSong Liu 			do_wakeup = 1;
27821e6d690bSSong Liu 	}
27831e6d690bSSong Liu 
27841e6d690bSSong Liu 	/*
27851e6d690bSSong Liu 	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
27861e6d690bSSong Liu 	 * We updated R5_InJournal, so we also update s->injournal.
27871e6d690bSSong Liu 	 */
27881e6d690bSSong Liu 	s->injournal = 0;
27891e6d690bSSong Liu 
27901e6d690bSSong Liu 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
27911e6d690bSSong Liu 		if (atomic_dec_and_test(&conf->pending_full_writes))
27921e6d690bSSong Liu 			md_wakeup_thread(conf->mddev->thread);
27931e6d690bSSong Liu 
27941e6d690bSSong Liu 	if (do_wakeup)
27951e6d690bSSong Liu 		wake_up(&conf->wait_for_overlap);
2796a39f7afdSSong Liu 
279703b047f4SSong Liu 	spin_lock_irq(&log->stripe_in_journal_lock);
2798a39f7afdSSong Liu 	list_del_init(&sh->r5c);
279903b047f4SSong Liu 	spin_unlock_irq(&log->stripe_in_journal_lock);
2800a39f7afdSSong Liu 	sh->log_start = MaxSector;
280103b047f4SSong Liu 
280203b047f4SSong Liu 	atomic_dec(&log->stripe_in_journal_count);
280303b047f4SSong Liu 	r5c_update_log_state(log);
280403b047f4SSong Liu 
280503b047f4SSong Liu 	/* stop counting this stripe in big_stripe_tree */
280603b047f4SSong Liu 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
280703b047f4SSong Liu 	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
280803b047f4SSong Liu 		tree_index = r5c_tree_index(conf, sh->sector);
280903b047f4SSong Liu 		spin_lock(&log->tree_lock);
281003b047f4SSong Liu 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
281103b047f4SSong Liu 					       tree_index);
281203b047f4SSong Liu 		BUG_ON(pslot == NULL);
281303b047f4SSong Liu 		refcount = (uintptr_t)radix_tree_deref_slot_protected(
281403b047f4SSong Liu 			pslot, &log->tree_lock) >>
281503b047f4SSong Liu 			R5C_RADIX_COUNT_SHIFT;
281603b047f4SSong Liu 		if (refcount == 1)
281703b047f4SSong Liu 			radix_tree_delete(&log->big_stripe_tree, tree_index);
281803b047f4SSong Liu 		else
281903b047f4SSong Liu 			radix_tree_replace_slot(
282003b047f4SSong Liu 				&log->big_stripe_tree, pslot,
282103b047f4SSong Liu 				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
282203b047f4SSong Liu 		spin_unlock(&log->tree_lock);
282303b047f4SSong Liu 	}
282403b047f4SSong Liu 
282503b047f4SSong Liu 	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
282603b047f4SSong Liu 		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2827e33fbb9cSShaohua Li 		atomic_dec(&conf->r5c_flushing_partial_stripes);
282803b047f4SSong Liu 		atomic_dec(&conf->r5c_cached_partial_stripes);
282903b047f4SSong Liu 	}
283003b047f4SSong Liu 
283103b047f4SSong Liu 	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
283203b047f4SSong Liu 		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2833e33fbb9cSShaohua Li 		atomic_dec(&conf->r5c_flushing_full_stripes);
283403b047f4SSong Liu 		atomic_dec(&conf->r5c_cached_full_stripes);
283503b047f4SSong Liu 	}
2836ea17481fSSong Liu 
2837ea17481fSSong Liu 	r5l_append_flush_payload(log, sh->sector);
28381e6d690bSSong Liu }
28391e6d690bSSong Liu 
2840ff875738SArtur Paszkiewicz int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
28411e6d690bSSong Liu {
2842a39f7afdSSong Liu 	struct r5conf *conf = sh->raid_conf;
28431e6d690bSSong Liu 	int pages = 0;
28441e6d690bSSong Liu 	int reserve;
28451e6d690bSSong Liu 	int i;
28461e6d690bSSong Liu 	int ret = 0;
28471e6d690bSSong Liu 
28481e6d690bSSong Liu 	BUG_ON(!log);
28491e6d690bSSong Liu 
28501e6d690bSSong Liu 	for (i = 0; i < sh->disks; i++) {
28511e6d690bSSong Liu 		void *addr;
28521e6d690bSSong Liu 
28531e6d690bSSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
28541e6d690bSSong Liu 			continue;
28551e6d690bSSong Liu 		addr = kmap_atomic(sh->dev[i].page);
28561e6d690bSSong Liu 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
28571e6d690bSSong Liu 						    addr, PAGE_SIZE);
28581e6d690bSSong Liu 		kunmap_atomic(addr);
28591e6d690bSSong Liu 		pages++;
28601e6d690bSSong Liu 	}
28611e6d690bSSong Liu 	WARN_ON(pages == 0);
28621e6d690bSSong Liu 
28631e6d690bSSong Liu 	/*
28641e6d690bSSong Liu 	 * The stripe must enter state machine again to call endio, so
28651e6d690bSSong Liu 	 * don't delay.
28661e6d690bSSong Liu 	 */
28671e6d690bSSong Liu 	clear_bit(STRIPE_DELAYED, &sh->state);
28681e6d690bSSong Liu 	atomic_inc(&sh->count);
28691e6d690bSSong Liu 
28701e6d690bSSong Liu 	mutex_lock(&log->io_mutex);
28711e6d690bSSong Liu 	/* meta + data */
28721e6d690bSSong Liu 	reserve = (1 + pages) << (PAGE_SHIFT - 9);
28731e6d690bSSong Liu 
2874a39f7afdSSong Liu 	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2875a39f7afdSSong Liu 	    sh->log_start == MaxSector)
2876a39f7afdSSong Liu 		r5l_add_no_space_stripe(log, sh);
2877a39f7afdSSong Liu 	else if (!r5l_has_free_space(log, reserve)) {
2878a39f7afdSSong Liu 		if (sh->log_start == log->last_checkpoint)
2879a39f7afdSSong Liu 			BUG();
2880a39f7afdSSong Liu 		else
2881a39f7afdSSong Liu 			r5l_add_no_space_stripe(log, sh);
28821e6d690bSSong Liu 	} else {
28831e6d690bSSong Liu 		ret = r5l_log_stripe(log, sh, pages, 0);
28841e6d690bSSong Liu 		if (ret) {
28851e6d690bSSong Liu 			spin_lock_irq(&log->io_list_lock);
28861e6d690bSSong Liu 			list_add_tail(&sh->log_list, &log->no_mem_stripes);
28871e6d690bSSong Liu 			spin_unlock_irq(&log->io_list_lock);
28881e6d690bSSong Liu 		}
28891e6d690bSSong Liu 	}
28901e6d690bSSong Liu 
28911e6d690bSSong Liu 	mutex_unlock(&log->io_mutex);
28921e6d690bSSong Liu 	return 0;
2893f6bed0efSShaohua Li }
2894f6bed0efSShaohua Li 
289503b047f4SSong Liu /* check whether this big stripe is in write back cache. */
289603b047f4SSong Liu bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
289703b047f4SSong Liu {
289803b047f4SSong Liu 	struct r5l_log *log = conf->log;
289903b047f4SSong Liu 	sector_t tree_index;
290003b047f4SSong Liu 	void *slot;
290103b047f4SSong Liu 
290203b047f4SSong Liu 	if (!log)
290303b047f4SSong Liu 		return false;
290403b047f4SSong Liu 
290503b047f4SSong Liu 	WARN_ON_ONCE(!rcu_read_lock_held());
290603b047f4SSong Liu 	tree_index = r5c_tree_index(conf, sect);
290703b047f4SSong Liu 	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
290803b047f4SSong Liu 	return slot != NULL;
290903b047f4SSong Liu }
291003b047f4SSong Liu 
2911f6bed0efSShaohua Li static int r5l_load_log(struct r5l_log *log)
2912f6bed0efSShaohua Li {
2913f6bed0efSShaohua Li 	struct md_rdev *rdev = log->rdev;
2914f6bed0efSShaohua Li 	struct page *page;
2915f6bed0efSShaohua Li 	struct r5l_meta_block *mb;
2916f6bed0efSShaohua Li 	sector_t cp = log->rdev->journal_tail;
2917f6bed0efSShaohua Li 	u32 stored_crc, expected_crc;
2918f6bed0efSShaohua Li 	bool create_super = false;
2919d30dfeb9SJackieLiu 	int ret = 0;
2920f6bed0efSShaohua Li 
2921f6bed0efSShaohua Li 	/* Make sure it's valid */
2922f6bed0efSShaohua Li 	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2923f6bed0efSShaohua Li 		cp = 0;
2924f6bed0efSShaohua Li 	page = alloc_page(GFP_KERNEL);
2925f6bed0efSShaohua Li 	if (!page)
2926f6bed0efSShaohua Li 		return -ENOMEM;
2927f6bed0efSShaohua Li 
2928796a5cf0SMike Christie 	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2929f6bed0efSShaohua Li 		ret = -EIO;
2930f6bed0efSShaohua Li 		goto ioerr;
2931f6bed0efSShaohua Li 	}
2932f6bed0efSShaohua Li 	mb = page_address(page);
2933f6bed0efSShaohua Li 
2934f6bed0efSShaohua Li 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2935f6bed0efSShaohua Li 	    mb->version != R5LOG_VERSION) {
2936f6bed0efSShaohua Li 		create_super = true;
2937f6bed0efSShaohua Li 		goto create;
2938f6bed0efSShaohua Li 	}
2939f6bed0efSShaohua Li 	stored_crc = le32_to_cpu(mb->checksum);
2940f6bed0efSShaohua Li 	mb->checksum = 0;
29415cb2fbd6SShaohua Li 	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2942f6bed0efSShaohua Li 	if (stored_crc != expected_crc) {
2943f6bed0efSShaohua Li 		create_super = true;
2944f6bed0efSShaohua Li 		goto create;
2945f6bed0efSShaohua Li 	}
2946f6bed0efSShaohua Li 	if (le64_to_cpu(mb->position) != cp) {
2947f6bed0efSShaohua Li 		create_super = true;
2948f6bed0efSShaohua Li 		goto create;
2949f6bed0efSShaohua Li 	}
2950f6bed0efSShaohua Li create:
2951f6bed0efSShaohua Li 	if (create_super) {
2952f6bed0efSShaohua Li 		log->last_cp_seq = prandom_u32();
2953f6bed0efSShaohua Li 		cp = 0;
295456056c2eSZhengyuan Liu 		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2955f6bed0efSShaohua Li 		/*
2956f6bed0efSShaohua Li 		 * Make sure super points to correct address. Log might have
2957f6bed0efSShaohua Li 		 * data very soon. If super hasn't correct log tail address,
2958f6bed0efSShaohua Li 		 * recovery can't find the log
2959f6bed0efSShaohua Li 		 */
2960f6bed0efSShaohua Li 		r5l_write_super(log, cp);
2961f6bed0efSShaohua Li 	} else
2962f6bed0efSShaohua Li 		log->last_cp_seq = le64_to_cpu(mb->seq);
2963f6bed0efSShaohua Li 
2964f6bed0efSShaohua Li 	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
29650576b1c6SShaohua Li 	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
29660576b1c6SShaohua Li 	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
29670576b1c6SShaohua Li 		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2968f6bed0efSShaohua Li 	log->last_checkpoint = cp;
2969f6bed0efSShaohua Li 
2970f6bed0efSShaohua Li 	__free_page(page);
2971f6bed0efSShaohua Li 
2972d30dfeb9SJackieLiu 	if (create_super) {
2973d30dfeb9SJackieLiu 		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
2974d30dfeb9SJackieLiu 		log->seq = log->last_cp_seq + 1;
2975d30dfeb9SJackieLiu 		log->next_checkpoint = cp;
2976d30dfeb9SJackieLiu 	} else
29773d7e7e1dSZhengyuan Liu 		ret = r5l_recovery_log(log);
2978d30dfeb9SJackieLiu 
29793d7e7e1dSZhengyuan Liu 	r5c_update_log_state(log);
29803d7e7e1dSZhengyuan Liu 	return ret;
2981f6bed0efSShaohua Li ioerr:
2982f6bed0efSShaohua Li 	__free_page(page);
2983f6bed0efSShaohua Li 	return ret;
2984f6bed0efSShaohua Li }
2985f6bed0efSShaohua Li 
29862e38a37fSSong Liu void r5c_update_on_rdev_error(struct mddev *mddev)
29872e38a37fSSong Liu {
29882e38a37fSSong Liu 	struct r5conf *conf = mddev->private;
29892e38a37fSSong Liu 	struct r5l_log *log = conf->log;
29902e38a37fSSong Liu 
29912e38a37fSSong Liu 	if (!log)
29922e38a37fSSong Liu 		return;
29932e38a37fSSong Liu 
29942e38a37fSSong Liu 	if (raid5_calc_degraded(conf) > 0 &&
29952e38a37fSSong Liu 	    conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
29962e38a37fSSong Liu 		schedule_work(&log->disable_writeback_work);
29972e38a37fSSong Liu }
29982e38a37fSSong Liu 
2999f6bed0efSShaohua Li int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3000f6bed0efSShaohua Li {
3001c888a8f9SJens Axboe 	struct request_queue *q = bdev_get_queue(rdev->bdev);
3002f6bed0efSShaohua Li 	struct r5l_log *log;
3003ff875738SArtur Paszkiewicz 	char b[BDEVNAME_SIZE];
3004ff875738SArtur Paszkiewicz 
3005ff875738SArtur Paszkiewicz 	pr_debug("md/raid:%s: using device %s as journal\n",
3006ff875738SArtur Paszkiewicz 		 mdname(conf->mddev), bdevname(rdev->bdev, b));
3007f6bed0efSShaohua Li 
3008f6bed0efSShaohua Li 	if (PAGE_SIZE != 4096)
3009f6bed0efSShaohua Li 		return -EINVAL;
3010c757ec95SSong Liu 
3011c757ec95SSong Liu 	/*
3012c757ec95SSong Liu 	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3013c757ec95SSong Liu 	 * raid_disks r5l_payload_data_parity.
3014c757ec95SSong Liu 	 *
3015c757ec95SSong Liu 	 * Write journal and cache does not work for very big array
3016c757ec95SSong Liu 	 * (raid_disks > 203)
3017c757ec95SSong Liu 	 */
3018c757ec95SSong Liu 	if (sizeof(struct r5l_meta_block) +
3019c757ec95SSong Liu 	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3020c757ec95SSong Liu 	     conf->raid_disks) > PAGE_SIZE) {
3021c757ec95SSong Liu 		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3022c757ec95SSong Liu 		       mdname(conf->mddev), conf->raid_disks);
3023c757ec95SSong Liu 		return -EINVAL;
3024c757ec95SSong Liu 	}
3025c757ec95SSong Liu 
3026f6bed0efSShaohua Li 	log = kzalloc(sizeof(*log), GFP_KERNEL);
3027f6bed0efSShaohua Li 	if (!log)
3028f6bed0efSShaohua Li 		return -ENOMEM;
3029f6bed0efSShaohua Li 	log->rdev = rdev;
3030f6bed0efSShaohua Li 
3031c888a8f9SJens Axboe 	log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
303256fef7c6SChristoph Hellwig 
30335cb2fbd6SShaohua Li 	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3034f6bed0efSShaohua Li 				       sizeof(rdev->mddev->uuid));
3035f6bed0efSShaohua Li 
3036f6bed0efSShaohua Li 	mutex_init(&log->io_mutex);
3037f6bed0efSShaohua Li 
3038f6bed0efSShaohua Li 	spin_lock_init(&log->io_list_lock);
3039f6bed0efSShaohua Li 	INIT_LIST_HEAD(&log->running_ios);
30400576b1c6SShaohua Li 	INIT_LIST_HEAD(&log->io_end_ios);
3041a8c34f91SShaohua Li 	INIT_LIST_HEAD(&log->flushing_ios);
304204732f74SChristoph Hellwig 	INIT_LIST_HEAD(&log->finished_ios);
30433a83f467SMing Lei 	bio_init(&log->flush_bio, NULL, 0);
3044f6bed0efSShaohua Li 
3045f6bed0efSShaohua Li 	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3046f6bed0efSShaohua Li 	if (!log->io_kc)
3047f6bed0efSShaohua Li 		goto io_kc;
3048f6bed0efSShaohua Li 
30495036c390SChristoph Hellwig 	log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
30505036c390SChristoph Hellwig 	if (!log->io_pool)
30515036c390SChristoph Hellwig 		goto io_pool;
30525036c390SChristoph Hellwig 
3053c38d29b3SChristoph Hellwig 	log->bs = bioset_create(R5L_POOL_SIZE, 0);
3054c38d29b3SChristoph Hellwig 	if (!log->bs)
3055c38d29b3SChristoph Hellwig 		goto io_bs;
3056c38d29b3SChristoph Hellwig 
3057e8deb638SChristoph Hellwig 	log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
3058e8deb638SChristoph Hellwig 	if (!log->meta_pool)
3059e8deb638SChristoph Hellwig 		goto out_mempool;
3060e8deb638SChristoph Hellwig 
306103b047f4SSong Liu 	spin_lock_init(&log->tree_lock);
306203b047f4SSong Liu 	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
306303b047f4SSong Liu 
30640576b1c6SShaohua Li 	log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
30650576b1c6SShaohua Li 						 log->rdev->mddev, "reclaim");
30660576b1c6SShaohua Li 	if (!log->reclaim_thread)
30670576b1c6SShaohua Li 		goto reclaim_thread;
3068a39f7afdSSong Liu 	log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3069a39f7afdSSong Liu 
30700fd22b45SShaohua Li 	init_waitqueue_head(&log->iounit_wait);
30710576b1c6SShaohua Li 
30725036c390SChristoph Hellwig 	INIT_LIST_HEAD(&log->no_mem_stripes);
30735036c390SChristoph Hellwig 
3074f6bed0efSShaohua Li 	INIT_LIST_HEAD(&log->no_space_stripes);
3075f6bed0efSShaohua Li 	spin_lock_init(&log->no_space_stripes_lock);
3076f6bed0efSShaohua Li 
30773bddb7f8SSong Liu 	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
30782e38a37fSSong Liu 	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
30793bddb7f8SSong Liu 
30802ded3703SSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3081a39f7afdSSong Liu 	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3082a39f7afdSSong Liu 	spin_lock_init(&log->stripe_in_journal_lock);
3083a39f7afdSSong Liu 	atomic_set(&log->stripe_in_journal_count, 0);
30842ded3703SSong Liu 
3085d2250f10SSong Liu 	rcu_assign_pointer(conf->log, log);
3086d2250f10SSong Liu 
3087f6bed0efSShaohua Li 	if (r5l_load_log(log))
3088f6bed0efSShaohua Li 		goto error;
3089f6bed0efSShaohua Li 
3090a62ab49eSShaohua Li 	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3091f6bed0efSShaohua Li 	return 0;
3092e8deb638SChristoph Hellwig 
3093f6bed0efSShaohua Li error:
3094d2250f10SSong Liu 	rcu_assign_pointer(conf->log, NULL);
30950576b1c6SShaohua Li 	md_unregister_thread(&log->reclaim_thread);
30960576b1c6SShaohua Li reclaim_thread:
3097e8deb638SChristoph Hellwig 	mempool_destroy(log->meta_pool);
3098e8deb638SChristoph Hellwig out_mempool:
3099c38d29b3SChristoph Hellwig 	bioset_free(log->bs);
3100c38d29b3SChristoph Hellwig io_bs:
31015036c390SChristoph Hellwig 	mempool_destroy(log->io_pool);
31025036c390SChristoph Hellwig io_pool:
3103f6bed0efSShaohua Li 	kmem_cache_destroy(log->io_kc);
3104f6bed0efSShaohua Li io_kc:
3105f6bed0efSShaohua Li 	kfree(log);
3106f6bed0efSShaohua Li 	return -EINVAL;
3107f6bed0efSShaohua Li }
3108f6bed0efSShaohua Li 
3109ff875738SArtur Paszkiewicz void r5l_exit_log(struct r5conf *conf)
3110f6bed0efSShaohua Li {
3111ff875738SArtur Paszkiewicz 	struct r5l_log *log = conf->log;
3112ff875738SArtur Paszkiewicz 
3113ff875738SArtur Paszkiewicz 	conf->log = NULL;
3114ff875738SArtur Paszkiewicz 	synchronize_rcu();
3115ff875738SArtur Paszkiewicz 
31162e38a37fSSong Liu 	flush_work(&log->disable_writeback_work);
31170576b1c6SShaohua Li 	md_unregister_thread(&log->reclaim_thread);
3118e8deb638SChristoph Hellwig 	mempool_destroy(log->meta_pool);
3119c38d29b3SChristoph Hellwig 	bioset_free(log->bs);
31205036c390SChristoph Hellwig 	mempool_destroy(log->io_pool);
3121f6bed0efSShaohua Li 	kmem_cache_destroy(log->io_kc);
3122f6bed0efSShaohua Li 	kfree(log);
3123f6bed0efSShaohua Li }
3124