xref: /linux/drivers/md/raid5-cache.c (revision effe6ee7523aa50d0517bd7da141e112b44d89fa)
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)
33*effe6ee7SSong 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 
572ded3703SSong Liu /*
582ded3703SSong Liu  * r5c journal modes of the array: write-back or write-through.
592ded3703SSong Liu  * write-through mode has identical behavior as existing log only
602ded3703SSong Liu  * implementation.
612ded3703SSong Liu  */
622ded3703SSong Liu enum r5c_journal_mode {
632ded3703SSong Liu 	R5C_JOURNAL_MODE_WRITE_THROUGH = 0,
642ded3703SSong Liu 	R5C_JOURNAL_MODE_WRITE_BACK = 1,
652ded3703SSong Liu };
662ded3703SSong Liu 
672c7da14bSSong Liu static char *r5c_journal_mode_str[] = {"write-through",
682c7da14bSSong Liu 				       "write-back"};
692ded3703SSong Liu /*
702ded3703SSong Liu  * raid5 cache state machine
712ded3703SSong Liu  *
729b69173eSJackieLiu  * With the RAID cache, each stripe works in two phases:
732ded3703SSong Liu  *	- caching phase
742ded3703SSong Liu  *	- writing-out phase
752ded3703SSong Liu  *
762ded3703SSong Liu  * These two phases are controlled by bit STRIPE_R5C_CACHING:
772ded3703SSong Liu  *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
782ded3703SSong Liu  *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
792ded3703SSong Liu  *
802ded3703SSong Liu  * When there is no journal, or the journal is in write-through mode,
812ded3703SSong Liu  * the stripe is always in writing-out phase.
822ded3703SSong Liu  *
832ded3703SSong Liu  * For write-back journal, the stripe is sent to caching phase on write
842ded3703SSong Liu  * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
852ded3703SSong Liu  * the write-out phase by clearing STRIPE_R5C_CACHING.
862ded3703SSong Liu  *
872ded3703SSong Liu  * Stripes in caching phase do not write the raid disks. Instead, all
882ded3703SSong Liu  * writes are committed from the log device. Therefore, a stripe in
892ded3703SSong Liu  * caching phase handles writes as:
902ded3703SSong Liu  *	- write to log device
912ded3703SSong Liu  *	- return IO
922ded3703SSong Liu  *
932ded3703SSong Liu  * Stripes in writing-out phase handle writes as:
942ded3703SSong Liu  *	- calculate parity
952ded3703SSong Liu  *	- write pending data and parity to journal
962ded3703SSong Liu  *	- write data and parity to raid disks
972ded3703SSong Liu  *	- return IO for pending writes
982ded3703SSong Liu  */
992ded3703SSong Liu 
100f6bed0efSShaohua Li struct r5l_log {
101f6bed0efSShaohua Li 	struct md_rdev *rdev;
102f6bed0efSShaohua Li 
103f6bed0efSShaohua Li 	u32 uuid_checksum;
104f6bed0efSShaohua Li 
105f6bed0efSShaohua Li 	sector_t device_size;		/* log device size, round to
106f6bed0efSShaohua Li 					 * BLOCK_SECTORS */
1070576b1c6SShaohua Li 	sector_t max_free_space;	/* reclaim run if free space is at
1080576b1c6SShaohua Li 					 * this size */
109f6bed0efSShaohua Li 
110f6bed0efSShaohua Li 	sector_t last_checkpoint;	/* log tail. where recovery scan
111f6bed0efSShaohua Li 					 * starts from */
112f6bed0efSShaohua Li 	u64 last_cp_seq;		/* log tail sequence */
113f6bed0efSShaohua Li 
114f6bed0efSShaohua Li 	sector_t log_start;		/* log head. where new data appends */
115f6bed0efSShaohua Li 	u64 seq;			/* log head sequence */
116f6bed0efSShaohua Li 
11717036461SChristoph Hellwig 	sector_t next_checkpoint;
11817036461SChristoph Hellwig 
119f6bed0efSShaohua Li 	struct mutex io_mutex;
120f6bed0efSShaohua Li 	struct r5l_io_unit *current_io;	/* current io_unit accepting new data */
121f6bed0efSShaohua Li 
122f6bed0efSShaohua Li 	spinlock_t io_list_lock;
123f6bed0efSShaohua Li 	struct list_head running_ios;	/* io_units which are still running,
124f6bed0efSShaohua Li 					 * and have not yet been completely
125f6bed0efSShaohua Li 					 * written to the log */
126f6bed0efSShaohua Li 	struct list_head io_end_ios;	/* io_units which have been completely
127f6bed0efSShaohua Li 					 * written to the log but not yet written
128f6bed0efSShaohua Li 					 * to the RAID */
129a8c34f91SShaohua Li 	struct list_head flushing_ios;	/* io_units which are waiting for log
130a8c34f91SShaohua Li 					 * cache flush */
13104732f74SChristoph Hellwig 	struct list_head finished_ios;	/* io_units which settle down in log disk */
132a8c34f91SShaohua Li 	struct bio flush_bio;
133f6bed0efSShaohua Li 
1345036c390SChristoph Hellwig 	struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
1355036c390SChristoph Hellwig 
136f6bed0efSShaohua Li 	struct kmem_cache *io_kc;
1375036c390SChristoph Hellwig 	mempool_t *io_pool;
138c38d29b3SChristoph Hellwig 	struct bio_set *bs;
139e8deb638SChristoph Hellwig 	mempool_t *meta_pool;
140f6bed0efSShaohua Li 
1410576b1c6SShaohua Li 	struct md_thread *reclaim_thread;
1420576b1c6SShaohua Li 	unsigned long reclaim_target;	/* number of space that need to be
1430576b1c6SShaohua Li 					 * reclaimed.  if it's 0, reclaim spaces
1440576b1c6SShaohua Li 					 * used by io_units which are in
1450576b1c6SShaohua Li 					 * IO_UNIT_STRIPE_END state (eg, reclaim
1460576b1c6SShaohua Li 					 * dones't wait for specific io_unit
1470576b1c6SShaohua Li 					 * switching to IO_UNIT_STRIPE_END
1480576b1c6SShaohua Li 					 * state) */
1490fd22b45SShaohua Li 	wait_queue_head_t iounit_wait;
1500576b1c6SShaohua Li 
151f6bed0efSShaohua Li 	struct list_head no_space_stripes; /* pending stripes, log has no space */
152f6bed0efSShaohua Li 	spinlock_t no_space_stripes_lock;
15356fef7c6SChristoph Hellwig 
15456fef7c6SChristoph Hellwig 	bool need_cache_flush;
1552ded3703SSong Liu 
1562ded3703SSong Liu 	/* for r5c_cache */
1572ded3703SSong Liu 	enum r5c_journal_mode r5c_journal_mode;
158a39f7afdSSong Liu 
159a39f7afdSSong Liu 	/* all stripes in r5cache, in the order of seq at sh->log_start */
160a39f7afdSSong Liu 	struct list_head stripe_in_journal_list;
161a39f7afdSSong Liu 
162a39f7afdSSong Liu 	spinlock_t stripe_in_journal_lock;
163a39f7afdSSong Liu 	atomic_t stripe_in_journal_count;
1643bddb7f8SSong Liu 
1653bddb7f8SSong Liu 	/* to submit async io_units, to fulfill ordering of flush */
1663bddb7f8SSong Liu 	struct work_struct deferred_io_work;
1672e38a37fSSong Liu 	/* to disable write back during in degraded mode */
1682e38a37fSSong Liu 	struct work_struct disable_writeback_work;
16903b047f4SSong Liu 
17003b047f4SSong Liu 	/* to for chunk_aligned_read in writeback mode, details below */
17103b047f4SSong Liu 	spinlock_t tree_lock;
17203b047f4SSong Liu 	struct radix_tree_root big_stripe_tree;
173f6bed0efSShaohua Li };
174f6bed0efSShaohua Li 
175f6bed0efSShaohua Li /*
17603b047f4SSong Liu  * Enable chunk_aligned_read() with write back cache.
17703b047f4SSong Liu  *
17803b047f4SSong Liu  * Each chunk may contain more than one stripe (for example, a 256kB
17903b047f4SSong Liu  * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
18003b047f4SSong Liu  * chunk_aligned_read, these stripes are grouped into one "big_stripe".
18103b047f4SSong Liu  * For each big_stripe, we count how many stripes of this big_stripe
18203b047f4SSong Liu  * are in the write back cache. These data are tracked in a radix tree
18303b047f4SSong Liu  * (big_stripe_tree). We use radix_tree item pointer as the counter.
18403b047f4SSong Liu  * r5c_tree_index() is used to calculate keys for the radix tree.
18503b047f4SSong Liu  *
18603b047f4SSong Liu  * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
18703b047f4SSong Liu  * big_stripe of each chunk in the tree. If this big_stripe is in the
18803b047f4SSong Liu  * tree, chunk_aligned_read() aborts. This look up is protected by
18903b047f4SSong Liu  * rcu_read_lock().
19003b047f4SSong Liu  *
19103b047f4SSong Liu  * It is necessary to remember whether a stripe is counted in
19203b047f4SSong Liu  * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
19303b047f4SSong Liu  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
19403b047f4SSong Liu  * two flags are set, the stripe is counted in big_stripe_tree. This
19503b047f4SSong Liu  * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
19603b047f4SSong Liu  * r5c_try_caching_write(); and moving clear_bit of
19703b047f4SSong Liu  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
19803b047f4SSong Liu  * r5c_finish_stripe_write_out().
19903b047f4SSong Liu  */
20003b047f4SSong Liu 
20103b047f4SSong Liu /*
20203b047f4SSong Liu  * radix tree requests lowest 2 bits of data pointer to be 2b'00.
20303b047f4SSong Liu  * So it is necessary to left shift the counter by 2 bits before using it
20403b047f4SSong Liu  * as data pointer of the tree.
20503b047f4SSong Liu  */
20603b047f4SSong Liu #define R5C_RADIX_COUNT_SHIFT 2
20703b047f4SSong Liu 
20803b047f4SSong Liu /*
20903b047f4SSong Liu  * calculate key for big_stripe_tree
21003b047f4SSong Liu  *
21103b047f4SSong Liu  * sect: align_bi->bi_iter.bi_sector or sh->sector
21203b047f4SSong Liu  */
21303b047f4SSong Liu static inline sector_t r5c_tree_index(struct r5conf *conf,
21403b047f4SSong Liu 				      sector_t sect)
21503b047f4SSong Liu {
21603b047f4SSong Liu 	sector_t offset;
21703b047f4SSong Liu 
21803b047f4SSong Liu 	offset = sector_div(sect, conf->chunk_sectors);
21903b047f4SSong Liu 	return sect;
22003b047f4SSong Liu }
22103b047f4SSong Liu 
22203b047f4SSong Liu /*
223f6bed0efSShaohua Li  * an IO range starts from a meta data block and end at the next meta data
224f6bed0efSShaohua Li  * block. The io unit's the meta data block tracks data/parity followed it. io
225f6bed0efSShaohua Li  * unit is written to log disk with normal write, as we always flush log disk
226f6bed0efSShaohua Li  * first and then start move data to raid disks, there is no requirement to
227f6bed0efSShaohua Li  * write io unit with FLUSH/FUA
228f6bed0efSShaohua Li  */
229f6bed0efSShaohua Li struct r5l_io_unit {
230f6bed0efSShaohua Li 	struct r5l_log *log;
231f6bed0efSShaohua Li 
232f6bed0efSShaohua Li 	struct page *meta_page;	/* store meta block */
233f6bed0efSShaohua Li 	int meta_offset;	/* current offset in meta_page */
234f6bed0efSShaohua Li 
235f6bed0efSShaohua Li 	struct bio *current_bio;/* current_bio accepting new data */
236f6bed0efSShaohua Li 
237f6bed0efSShaohua Li 	atomic_t pending_stripe;/* how many stripes not flushed to raid */
238f6bed0efSShaohua Li 	u64 seq;		/* seq number of the metablock */
239f6bed0efSShaohua Li 	sector_t log_start;	/* where the io_unit starts */
240f6bed0efSShaohua Li 	sector_t log_end;	/* where the io_unit ends */
241f6bed0efSShaohua Li 	struct list_head log_sibling; /* log->running_ios */
242f6bed0efSShaohua Li 	struct list_head stripe_list; /* stripes added to the io_unit */
243f6bed0efSShaohua Li 
244f6bed0efSShaohua Li 	int state;
2456143e2ceSChristoph Hellwig 	bool need_split_bio;
2463bddb7f8SSong Liu 	struct bio *split_bio;
2473bddb7f8SSong Liu 
2483bddb7f8SSong Liu 	unsigned int has_flush:1;      /* include flush request */
2493bddb7f8SSong Liu 	unsigned int has_fua:1;        /* include fua request */
2503bddb7f8SSong Liu 	unsigned int has_null_flush:1; /* include empty flush request */
2513bddb7f8SSong Liu 	/*
2523bddb7f8SSong Liu 	 * io isn't sent yet, flush/fua request can only be submitted till it's
2533bddb7f8SSong Liu 	 * the first IO in running_ios list
2543bddb7f8SSong Liu 	 */
2553bddb7f8SSong Liu 	unsigned int io_deferred:1;
2563bddb7f8SSong Liu 
2573bddb7f8SSong Liu 	struct bio_list flush_barriers;   /* size == 0 flush bios */
258f6bed0efSShaohua Li };
259f6bed0efSShaohua Li 
260f6bed0efSShaohua Li /* r5l_io_unit state */
261f6bed0efSShaohua Li enum r5l_io_unit_state {
262f6bed0efSShaohua Li 	IO_UNIT_RUNNING = 0,	/* accepting new IO */
263f6bed0efSShaohua Li 	IO_UNIT_IO_START = 1,	/* io_unit bio start writing to log,
264f6bed0efSShaohua Li 				 * don't accepting new bio */
265f6bed0efSShaohua Li 	IO_UNIT_IO_END = 2,	/* io_unit bio finish writing to log */
266a8c34f91SShaohua Li 	IO_UNIT_STRIPE_END = 3,	/* stripes data finished writing to raid */
267f6bed0efSShaohua Li };
268f6bed0efSShaohua Li 
2692ded3703SSong Liu bool r5c_is_writeback(struct r5l_log *log)
2702ded3703SSong Liu {
2712ded3703SSong Liu 	return (log != NULL &&
2722ded3703SSong Liu 		log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
2732ded3703SSong Liu }
2742ded3703SSong Liu 
275f6bed0efSShaohua Li static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
276f6bed0efSShaohua Li {
277f6bed0efSShaohua Li 	start += inc;
278f6bed0efSShaohua Li 	if (start >= log->device_size)
279f6bed0efSShaohua Li 		start = start - log->device_size;
280f6bed0efSShaohua Li 	return start;
281f6bed0efSShaohua Li }
282f6bed0efSShaohua Li 
283f6bed0efSShaohua Li static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
284f6bed0efSShaohua Li 				  sector_t end)
285f6bed0efSShaohua Li {
286f6bed0efSShaohua Li 	if (end >= start)
287f6bed0efSShaohua Li 		return end - start;
288f6bed0efSShaohua Li 	else
289f6bed0efSShaohua Li 		return end + log->device_size - start;
290f6bed0efSShaohua Li }
291f6bed0efSShaohua Li 
292f6bed0efSShaohua Li static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
293f6bed0efSShaohua Li {
294f6bed0efSShaohua Li 	sector_t used_size;
295f6bed0efSShaohua Li 
296f6bed0efSShaohua Li 	used_size = r5l_ring_distance(log, log->last_checkpoint,
297f6bed0efSShaohua Li 					log->log_start);
298f6bed0efSShaohua Li 
299f6bed0efSShaohua Li 	return log->device_size > used_size + size;
300f6bed0efSShaohua Li }
301f6bed0efSShaohua Li 
302f6bed0efSShaohua Li static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
303f6bed0efSShaohua Li 				    enum r5l_io_unit_state state)
304f6bed0efSShaohua Li {
305f6bed0efSShaohua Li 	if (WARN_ON(io->state >= state))
306f6bed0efSShaohua Li 		return;
307f6bed0efSShaohua Li 	io->state = state;
308f6bed0efSShaohua Li }
309f6bed0efSShaohua Li 
3101e6d690bSSong Liu static void
3111e6d690bSSong Liu r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev,
3121e6d690bSSong Liu 			      struct bio_list *return_bi)
3131e6d690bSSong Liu {
3141e6d690bSSong Liu 	struct bio *wbi, *wbi2;
3151e6d690bSSong Liu 
3161e6d690bSSong Liu 	wbi = dev->written;
3171e6d690bSSong Liu 	dev->written = NULL;
3181e6d690bSSong Liu 	while (wbi && wbi->bi_iter.bi_sector <
3191e6d690bSSong Liu 	       dev->sector + STRIPE_SECTORS) {
3201e6d690bSSong Liu 		wbi2 = r5_next_bio(wbi, dev->sector);
3211e6d690bSSong Liu 		if (!raid5_dec_bi_active_stripes(wbi)) {
3221e6d690bSSong Liu 			md_write_end(conf->mddev);
3231e6d690bSSong Liu 			bio_list_add(return_bi, wbi);
3241e6d690bSSong Liu 		}
3251e6d690bSSong Liu 		wbi = wbi2;
3261e6d690bSSong Liu 	}
3271e6d690bSSong Liu }
3281e6d690bSSong Liu 
3291e6d690bSSong Liu void r5c_handle_cached_data_endio(struct r5conf *conf,
3301e6d690bSSong Liu 	  struct stripe_head *sh, int disks, struct bio_list *return_bi)
3311e6d690bSSong Liu {
3321e6d690bSSong Liu 	int i;
3331e6d690bSSong Liu 
3341e6d690bSSong Liu 	for (i = sh->disks; i--; ) {
3351e6d690bSSong Liu 		if (sh->dev[i].written) {
3361e6d690bSSong Liu 			set_bit(R5_UPTODATE, &sh->dev[i].flags);
3371e6d690bSSong Liu 			r5c_return_dev_pending_writes(conf, &sh->dev[i],
3381e6d690bSSong Liu 						      return_bi);
3391e6d690bSSong Liu 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3401e6d690bSSong Liu 					STRIPE_SECTORS,
3411e6d690bSSong Liu 					!test_bit(STRIPE_DEGRADED, &sh->state),
3421e6d690bSSong Liu 					0);
3431e6d690bSSong Liu 		}
3441e6d690bSSong Liu 	}
3451e6d690bSSong Liu }
3461e6d690bSSong Liu 
347a39f7afdSSong Liu /* Check whether we should flush some stripes to free up stripe cache */
348a39f7afdSSong Liu void r5c_check_stripe_cache_usage(struct r5conf *conf)
349a39f7afdSSong Liu {
350a39f7afdSSong Liu 	int total_cached;
351a39f7afdSSong Liu 
352a39f7afdSSong Liu 	if (!r5c_is_writeback(conf->log))
353a39f7afdSSong Liu 		return;
354a39f7afdSSong Liu 
355a39f7afdSSong Liu 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
356a39f7afdSSong Liu 		atomic_read(&conf->r5c_cached_full_stripes);
357a39f7afdSSong Liu 
358a39f7afdSSong Liu 	/*
359a39f7afdSSong Liu 	 * The following condition is true for either of the following:
360a39f7afdSSong Liu 	 *   - stripe cache pressure high:
361a39f7afdSSong Liu 	 *          total_cached > 3/4 min_nr_stripes ||
362a39f7afdSSong Liu 	 *          empty_inactive_list_nr > 0
363a39f7afdSSong Liu 	 *   - stripe cache pressure moderate:
364a39f7afdSSong Liu 	 *          total_cached > 1/2 min_nr_stripes
365a39f7afdSSong Liu 	 */
366a39f7afdSSong Liu 	if (total_cached > conf->min_nr_stripes * 1 / 2 ||
367a39f7afdSSong Liu 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
368a39f7afdSSong Liu 		r5l_wake_reclaim(conf->log, 0);
369a39f7afdSSong Liu }
370a39f7afdSSong Liu 
371a39f7afdSSong Liu /*
372a39f7afdSSong Liu  * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
373a39f7afdSSong Liu  * stripes in the cache
374a39f7afdSSong Liu  */
375a39f7afdSSong Liu void r5c_check_cached_full_stripe(struct r5conf *conf)
376a39f7afdSSong Liu {
377a39f7afdSSong Liu 	if (!r5c_is_writeback(conf->log))
378a39f7afdSSong Liu 		return;
379a39f7afdSSong Liu 
380a39f7afdSSong Liu 	/*
381a39f7afdSSong Liu 	 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
382a39f7afdSSong Liu 	 * or a full stripe (chunk size / 4k stripes).
383a39f7afdSSong Liu 	 */
384a39f7afdSSong Liu 	if (atomic_read(&conf->r5c_cached_full_stripes) >=
38584890c03SShaohua Li 	    min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
386a39f7afdSSong Liu 		conf->chunk_sectors >> STRIPE_SHIFT))
387a39f7afdSSong Liu 		r5l_wake_reclaim(conf->log, 0);
388a39f7afdSSong Liu }
389a39f7afdSSong Liu 
390a39f7afdSSong Liu /*
391a39f7afdSSong Liu  * Total log space (in sectors) needed to flush all data in cache
392a39f7afdSSong Liu  *
39339b99586SSong Liu  * To avoid deadlock due to log space, it is necessary to reserve log
39439b99586SSong Liu  * space to flush critical stripes (stripes that occupying log space near
39539b99586SSong Liu  * last_checkpoint). This function helps check how much log space is
39639b99586SSong Liu  * required to flush all cached stripes.
397a39f7afdSSong Liu  *
39839b99586SSong Liu  * To reduce log space requirements, two mechanisms are used to give cache
39939b99586SSong Liu  * flush higher priorities:
40039b99586SSong Liu  *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
40139b99586SSong Liu  *       stripes ALREADY in journal can be flushed w/o pending writes;
40239b99586SSong Liu  *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
40339b99586SSong Liu  *       can be delayed (r5l_add_no_space_stripe).
404a39f7afdSSong Liu  *
40539b99586SSong Liu  * In cache flush, the stripe goes through 1 and then 2. For a stripe that
40639b99586SSong Liu  * already passed 1, flushing it requires at most (conf->max_degraded + 1)
40739b99586SSong Liu  * pages of journal space. For stripes that has not passed 1, flushing it
40839b99586SSong Liu  * requires (conf->raid_disks + 1) pages of journal space. There are at
40939b99586SSong Liu  * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
41039b99586SSong Liu  * required to flush all cached stripes (in pages) is:
41139b99586SSong Liu  *
41239b99586SSong Liu  *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
41339b99586SSong Liu  *     (group_cnt + 1) * (raid_disks + 1)
41439b99586SSong Liu  * or
41539b99586SSong Liu  *     (stripe_in_journal_count) * (max_degraded + 1) +
41639b99586SSong Liu  *     (group_cnt + 1) * (raid_disks - max_degraded)
417a39f7afdSSong Liu  */
418a39f7afdSSong Liu static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
419a39f7afdSSong Liu {
420a39f7afdSSong Liu 	struct r5l_log *log = conf->log;
421a39f7afdSSong Liu 
422a39f7afdSSong Liu 	if (!r5c_is_writeback(log))
423a39f7afdSSong Liu 		return 0;
424a39f7afdSSong Liu 
42539b99586SSong Liu 	return BLOCK_SECTORS *
42639b99586SSong Liu 		((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
42739b99586SSong Liu 		 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
428a39f7afdSSong Liu }
429a39f7afdSSong Liu 
430a39f7afdSSong Liu /*
431a39f7afdSSong Liu  * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
432a39f7afdSSong Liu  *
433a39f7afdSSong Liu  * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
434a39f7afdSSong Liu  * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
435a39f7afdSSong Liu  * device is less than 2x of reclaim_required_space.
436a39f7afdSSong Liu  */
437a39f7afdSSong Liu static inline void r5c_update_log_state(struct r5l_log *log)
438a39f7afdSSong Liu {
439a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
440a39f7afdSSong Liu 	sector_t free_space;
441a39f7afdSSong Liu 	sector_t reclaim_space;
442f687a33eSSong Liu 	bool wake_reclaim = false;
443a39f7afdSSong Liu 
444a39f7afdSSong Liu 	if (!r5c_is_writeback(log))
445a39f7afdSSong Liu 		return;
446a39f7afdSSong Liu 
447a39f7afdSSong Liu 	free_space = r5l_ring_distance(log, log->log_start,
448a39f7afdSSong Liu 				       log->last_checkpoint);
449a39f7afdSSong Liu 	reclaim_space = r5c_log_required_to_flush_cache(conf);
450a39f7afdSSong Liu 	if (free_space < 2 * reclaim_space)
451a39f7afdSSong Liu 		set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
452f687a33eSSong Liu 	else {
453f687a33eSSong Liu 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
454f687a33eSSong Liu 			wake_reclaim = true;
455a39f7afdSSong Liu 		clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
456f687a33eSSong Liu 	}
457a39f7afdSSong Liu 	if (free_space < 3 * reclaim_space)
458a39f7afdSSong Liu 		set_bit(R5C_LOG_TIGHT, &conf->cache_state);
459a39f7afdSSong Liu 	else
460a39f7afdSSong Liu 		clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
461f687a33eSSong Liu 
462f687a33eSSong Liu 	if (wake_reclaim)
463f687a33eSSong Liu 		r5l_wake_reclaim(log, 0);
464a39f7afdSSong Liu }
465a39f7afdSSong Liu 
4662ded3703SSong Liu /*
4672ded3703SSong Liu  * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
4682ded3703SSong Liu  * This function should only be called in write-back mode.
4692ded3703SSong Liu  */
470a39f7afdSSong Liu void r5c_make_stripe_write_out(struct stripe_head *sh)
4712ded3703SSong Liu {
4722ded3703SSong Liu 	struct r5conf *conf = sh->raid_conf;
4732ded3703SSong Liu 	struct r5l_log *log = conf->log;
4742ded3703SSong Liu 
4752ded3703SSong Liu 	BUG_ON(!r5c_is_writeback(log));
4762ded3703SSong Liu 
4772ded3703SSong Liu 	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
4782ded3703SSong Liu 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
4791e6d690bSSong Liu 
4801e6d690bSSong Liu 	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4811e6d690bSSong Liu 		atomic_inc(&conf->preread_active_stripes);
4821e6d690bSSong Liu }
4831e6d690bSSong Liu 
4841e6d690bSSong Liu static void r5c_handle_data_cached(struct stripe_head *sh)
4851e6d690bSSong Liu {
4861e6d690bSSong Liu 	int i;
4871e6d690bSSong Liu 
4881e6d690bSSong Liu 	for (i = sh->disks; i--; )
4891e6d690bSSong Liu 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
4901e6d690bSSong Liu 			set_bit(R5_InJournal, &sh->dev[i].flags);
4911e6d690bSSong Liu 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
4921e6d690bSSong Liu 		}
4931e6d690bSSong Liu 	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
4941e6d690bSSong Liu }
4951e6d690bSSong Liu 
4961e6d690bSSong Liu /*
4971e6d690bSSong Liu  * this journal write must contain full parity,
4981e6d690bSSong Liu  * it may also contain some data pages
4991e6d690bSSong Liu  */
5001e6d690bSSong Liu static void r5c_handle_parity_cached(struct stripe_head *sh)
5011e6d690bSSong Liu {
5021e6d690bSSong Liu 	int i;
5031e6d690bSSong Liu 
5041e6d690bSSong Liu 	for (i = sh->disks; i--; )
5051e6d690bSSong Liu 		if (test_bit(R5_InJournal, &sh->dev[i].flags))
5061e6d690bSSong Liu 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
5072ded3703SSong Liu }
5082ded3703SSong Liu 
5092ded3703SSong Liu /*
5102ded3703SSong Liu  * Setting proper flags after writing (or flushing) data and/or parity to the
5112ded3703SSong Liu  * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
5122ded3703SSong Liu  */
5132ded3703SSong Liu static void r5c_finish_cache_stripe(struct stripe_head *sh)
5142ded3703SSong Liu {
5152ded3703SSong Liu 	struct r5l_log *log = sh->raid_conf->log;
5162ded3703SSong Liu 
5172ded3703SSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
5182ded3703SSong Liu 		BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
5192ded3703SSong Liu 		/*
5202ded3703SSong Liu 		 * Set R5_InJournal for parity dev[pd_idx]. This means
5212ded3703SSong Liu 		 * all data AND parity in the journal. For RAID 6, it is
5222ded3703SSong Liu 		 * NOT necessary to set the flag for dev[qd_idx], as the
5232ded3703SSong Liu 		 * two parities are written out together.
5242ded3703SSong Liu 		 */
5252ded3703SSong Liu 		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
5261e6d690bSSong Liu 	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
5271e6d690bSSong Liu 		r5c_handle_data_cached(sh);
5281e6d690bSSong Liu 	} else {
5291e6d690bSSong Liu 		r5c_handle_parity_cached(sh);
5301e6d690bSSong Liu 		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
5311e6d690bSSong Liu 	}
5322ded3703SSong Liu }
5332ded3703SSong Liu 
534d8858f43SChristoph Hellwig static void r5l_io_run_stripes(struct r5l_io_unit *io)
535d8858f43SChristoph Hellwig {
536d8858f43SChristoph Hellwig 	struct stripe_head *sh, *next;
537d8858f43SChristoph Hellwig 
538d8858f43SChristoph Hellwig 	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
539d8858f43SChristoph Hellwig 		list_del_init(&sh->log_list);
5402ded3703SSong Liu 
5412ded3703SSong Liu 		r5c_finish_cache_stripe(sh);
5422ded3703SSong Liu 
543d8858f43SChristoph Hellwig 		set_bit(STRIPE_HANDLE, &sh->state);
544d8858f43SChristoph Hellwig 		raid5_release_stripe(sh);
545d8858f43SChristoph Hellwig 	}
546d8858f43SChristoph Hellwig }
547d8858f43SChristoph Hellwig 
54856fef7c6SChristoph Hellwig static void r5l_log_run_stripes(struct r5l_log *log)
54956fef7c6SChristoph Hellwig {
55056fef7c6SChristoph Hellwig 	struct r5l_io_unit *io, *next;
55156fef7c6SChristoph Hellwig 
55256fef7c6SChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
55356fef7c6SChristoph Hellwig 
55456fef7c6SChristoph Hellwig 	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
55556fef7c6SChristoph Hellwig 		/* don't change list order */
55656fef7c6SChristoph Hellwig 		if (io->state < IO_UNIT_IO_END)
55756fef7c6SChristoph Hellwig 			break;
55856fef7c6SChristoph Hellwig 
55956fef7c6SChristoph Hellwig 		list_move_tail(&io->log_sibling, &log->finished_ios);
56056fef7c6SChristoph Hellwig 		r5l_io_run_stripes(io);
56156fef7c6SChristoph Hellwig 	}
56256fef7c6SChristoph Hellwig }
56356fef7c6SChristoph Hellwig 
5643848c0bcSChristoph Hellwig static void r5l_move_to_end_ios(struct r5l_log *log)
5653848c0bcSChristoph Hellwig {
5663848c0bcSChristoph Hellwig 	struct r5l_io_unit *io, *next;
5673848c0bcSChristoph Hellwig 
5683848c0bcSChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
5693848c0bcSChristoph Hellwig 
5703848c0bcSChristoph Hellwig 	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
5713848c0bcSChristoph Hellwig 		/* don't change list order */
5723848c0bcSChristoph Hellwig 		if (io->state < IO_UNIT_IO_END)
5733848c0bcSChristoph Hellwig 			break;
5743848c0bcSChristoph Hellwig 		list_move_tail(&io->log_sibling, &log->io_end_ios);
5753848c0bcSChristoph Hellwig 	}
5763848c0bcSChristoph Hellwig }
5773848c0bcSChristoph Hellwig 
5783bddb7f8SSong Liu static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
579f6bed0efSShaohua Li static void r5l_log_endio(struct bio *bio)
580f6bed0efSShaohua Li {
581f6bed0efSShaohua Li 	struct r5l_io_unit *io = bio->bi_private;
5823bddb7f8SSong Liu 	struct r5l_io_unit *io_deferred;
583f6bed0efSShaohua Li 	struct r5l_log *log = io->log;
584509ffec7SChristoph Hellwig 	unsigned long flags;
585f6bed0efSShaohua Li 
5866e74a9cfSShaohua Li 	if (bio->bi_error)
5876e74a9cfSShaohua Li 		md_error(log->rdev->mddev, log->rdev);
5886e74a9cfSShaohua Li 
589f6bed0efSShaohua Li 	bio_put(bio);
590e8deb638SChristoph Hellwig 	mempool_free(io->meta_page, log->meta_pool);
591f6bed0efSShaohua Li 
592509ffec7SChristoph Hellwig 	spin_lock_irqsave(&log->io_list_lock, flags);
593509ffec7SChristoph Hellwig 	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
59456fef7c6SChristoph Hellwig 	if (log->need_cache_flush)
5953848c0bcSChristoph Hellwig 		r5l_move_to_end_ios(log);
59656fef7c6SChristoph Hellwig 	else
59756fef7c6SChristoph Hellwig 		r5l_log_run_stripes(log);
5983bddb7f8SSong Liu 	if (!list_empty(&log->running_ios)) {
5993bddb7f8SSong Liu 		/*
6003bddb7f8SSong Liu 		 * FLUSH/FUA io_unit is deferred because of ordering, now we
6013bddb7f8SSong Liu 		 * can dispatch it
6023bddb7f8SSong Liu 		 */
6033bddb7f8SSong Liu 		io_deferred = list_first_entry(&log->running_ios,
6043bddb7f8SSong Liu 					       struct r5l_io_unit, log_sibling);
6053bddb7f8SSong Liu 		if (io_deferred->io_deferred)
6063bddb7f8SSong Liu 			schedule_work(&log->deferred_io_work);
6073bddb7f8SSong Liu 	}
6083bddb7f8SSong Liu 
609509ffec7SChristoph Hellwig 	spin_unlock_irqrestore(&log->io_list_lock, flags);
610509ffec7SChristoph Hellwig 
61156fef7c6SChristoph Hellwig 	if (log->need_cache_flush)
612f6bed0efSShaohua Li 		md_wakeup_thread(log->rdev->mddev->thread);
6133bddb7f8SSong Liu 
6143bddb7f8SSong Liu 	if (io->has_null_flush) {
6153bddb7f8SSong Liu 		struct bio *bi;
6163bddb7f8SSong Liu 
6173bddb7f8SSong Liu 		WARN_ON(bio_list_empty(&io->flush_barriers));
6183bddb7f8SSong Liu 		while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
6193bddb7f8SSong Liu 			bio_endio(bi);
6203bddb7f8SSong Liu 			atomic_dec(&io->pending_stripe);
6213bddb7f8SSong Liu 		}
6223bddb7f8SSong Liu 		if (atomic_read(&io->pending_stripe) == 0)
6233bddb7f8SSong Liu 			__r5l_stripe_write_finished(io);
6243bddb7f8SSong Liu 	}
6253bddb7f8SSong Liu }
6263bddb7f8SSong Liu 
6273bddb7f8SSong Liu static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
6283bddb7f8SSong Liu {
6293bddb7f8SSong Liu 	unsigned long flags;
6303bddb7f8SSong Liu 
6313bddb7f8SSong Liu 	spin_lock_irqsave(&log->io_list_lock, flags);
6323bddb7f8SSong Liu 	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
6333bddb7f8SSong Liu 	spin_unlock_irqrestore(&log->io_list_lock, flags);
6343bddb7f8SSong Liu 
6353bddb7f8SSong Liu 	if (io->has_flush)
63620737738SShaohua Li 		io->current_bio->bi_opf |= REQ_PREFLUSH;
6373bddb7f8SSong Liu 	if (io->has_fua)
63820737738SShaohua Li 		io->current_bio->bi_opf |= REQ_FUA;
6393bddb7f8SSong Liu 	submit_bio(io->current_bio);
6403bddb7f8SSong Liu 
6413bddb7f8SSong Liu 	if (!io->split_bio)
6423bddb7f8SSong Liu 		return;
6433bddb7f8SSong Liu 
6443bddb7f8SSong Liu 	if (io->has_flush)
64520737738SShaohua Li 		io->split_bio->bi_opf |= REQ_PREFLUSH;
6463bddb7f8SSong Liu 	if (io->has_fua)
64720737738SShaohua Li 		io->split_bio->bi_opf |= REQ_FUA;
6483bddb7f8SSong Liu 	submit_bio(io->split_bio);
6493bddb7f8SSong Liu }
6503bddb7f8SSong 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 
8465036c390SChristoph Hellwig static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
847f6bed0efSShaohua Li 			   int data_pages, int parity_pages)
848f6bed0efSShaohua Li {
849f6bed0efSShaohua Li 	int i;
850f6bed0efSShaohua Li 	int meta_size;
8515036c390SChristoph Hellwig 	int ret;
852f6bed0efSShaohua Li 	struct r5l_io_unit *io;
853f6bed0efSShaohua Li 
854f6bed0efSShaohua Li 	meta_size =
855f6bed0efSShaohua Li 		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
856f6bed0efSShaohua Li 		 * data_pages) +
857f6bed0efSShaohua Li 		sizeof(struct r5l_payload_data_parity) +
858f6bed0efSShaohua Li 		sizeof(__le32) * parity_pages;
859f6bed0efSShaohua Li 
8605036c390SChristoph Hellwig 	ret = r5l_get_meta(log, meta_size);
8615036c390SChristoph Hellwig 	if (ret)
8625036c390SChristoph Hellwig 		return ret;
8635036c390SChristoph Hellwig 
864f6bed0efSShaohua Li 	io = log->current_io;
865f6bed0efSShaohua Li 
8663bddb7f8SSong Liu 	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
8673bddb7f8SSong Liu 		io->has_flush = 1;
8683bddb7f8SSong Liu 
869f6bed0efSShaohua Li 	for (i = 0; i < sh->disks; i++) {
8701e6d690bSSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
8711e6d690bSSong Liu 		    test_bit(R5_InJournal, &sh->dev[i].flags))
872f6bed0efSShaohua Li 			continue;
873f6bed0efSShaohua Li 		if (i == sh->pd_idx || i == sh->qd_idx)
874f6bed0efSShaohua Li 			continue;
8753bddb7f8SSong Liu 		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
8763bddb7f8SSong Liu 		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
8773bddb7f8SSong Liu 			io->has_fua = 1;
8783bddb7f8SSong Liu 			/*
8793bddb7f8SSong Liu 			 * we need to flush journal to make sure recovery can
8803bddb7f8SSong Liu 			 * reach the data with fua flag
8813bddb7f8SSong Liu 			 */
8823bddb7f8SSong Liu 			io->has_flush = 1;
8833bddb7f8SSong Liu 		}
884f6bed0efSShaohua Li 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
885f6bed0efSShaohua Li 					raid5_compute_blocknr(sh, i, 0),
886f6bed0efSShaohua Li 					sh->dev[i].log_checksum, 0, false);
887f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[i].page);
888f6bed0efSShaohua Li 	}
889f6bed0efSShaohua Li 
8902ded3703SSong Liu 	if (parity_pages == 2) {
891f6bed0efSShaohua Li 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
892f6bed0efSShaohua Li 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
893f6bed0efSShaohua Li 					sh->dev[sh->qd_idx].log_checksum, true);
894f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
895f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
8962ded3703SSong Liu 	} else if (parity_pages == 1) {
897f6bed0efSShaohua Li 		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
898f6bed0efSShaohua Li 					sh->sector, sh->dev[sh->pd_idx].log_checksum,
899f6bed0efSShaohua Li 					0, false);
900f6bed0efSShaohua Li 		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
9012ded3703SSong Liu 	} else  /* Just writing data, not parity, in caching phase */
9022ded3703SSong Liu 		BUG_ON(parity_pages != 0);
903f6bed0efSShaohua Li 
904f6bed0efSShaohua Li 	list_add_tail(&sh->log_list, &io->stripe_list);
905f6bed0efSShaohua Li 	atomic_inc(&io->pending_stripe);
906f6bed0efSShaohua Li 	sh->log_io = io;
9075036c390SChristoph Hellwig 
908a39f7afdSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
909a39f7afdSSong Liu 		return 0;
910a39f7afdSSong Liu 
911a39f7afdSSong Liu 	if (sh->log_start == MaxSector) {
912a39f7afdSSong Liu 		BUG_ON(!list_empty(&sh->r5c));
913a39f7afdSSong Liu 		sh->log_start = io->log_start;
914a39f7afdSSong Liu 		spin_lock_irq(&log->stripe_in_journal_lock);
915a39f7afdSSong Liu 		list_add_tail(&sh->r5c,
916a39f7afdSSong Liu 			      &log->stripe_in_journal_list);
917a39f7afdSSong Liu 		spin_unlock_irq(&log->stripe_in_journal_lock);
918a39f7afdSSong Liu 		atomic_inc(&log->stripe_in_journal_count);
919a39f7afdSSong Liu 	}
9205036c390SChristoph Hellwig 	return 0;
921f6bed0efSShaohua Li }
922f6bed0efSShaohua Li 
923a39f7afdSSong Liu /* add stripe to no_space_stripes, and then wake up reclaim */
924a39f7afdSSong Liu static inline void r5l_add_no_space_stripe(struct r5l_log *log,
925a39f7afdSSong Liu 					   struct stripe_head *sh)
926a39f7afdSSong Liu {
927a39f7afdSSong Liu 	spin_lock(&log->no_space_stripes_lock);
928a39f7afdSSong Liu 	list_add_tail(&sh->log_list, &log->no_space_stripes);
929a39f7afdSSong Liu 	spin_unlock(&log->no_space_stripes_lock);
930a39f7afdSSong Liu }
931a39f7afdSSong Liu 
932f6bed0efSShaohua Li /*
933f6bed0efSShaohua Li  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
934f6bed0efSShaohua Li  * data from log to raid disks), so we shouldn't wait for reclaim here
935f6bed0efSShaohua Li  */
936f6bed0efSShaohua Li int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
937f6bed0efSShaohua Li {
938a39f7afdSSong Liu 	struct r5conf *conf = sh->raid_conf;
939f6bed0efSShaohua Li 	int write_disks = 0;
940f6bed0efSShaohua Li 	int data_pages, parity_pages;
941f6bed0efSShaohua Li 	int reserve;
942f6bed0efSShaohua Li 	int i;
9435036c390SChristoph Hellwig 	int ret = 0;
944a39f7afdSSong Liu 	bool wake_reclaim = false;
945f6bed0efSShaohua Li 
946f6bed0efSShaohua Li 	if (!log)
947f6bed0efSShaohua Li 		return -EAGAIN;
948f6bed0efSShaohua Li 	/* Don't support stripe batch */
949f6bed0efSShaohua Li 	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
950f6bed0efSShaohua Li 	    test_bit(STRIPE_SYNCING, &sh->state)) {
951f6bed0efSShaohua Li 		/* the stripe is written to log, we start writing it to raid */
952f6bed0efSShaohua Li 		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
953f6bed0efSShaohua Li 		return -EAGAIN;
954f6bed0efSShaohua Li 	}
955f6bed0efSShaohua Li 
9562ded3703SSong Liu 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
9572ded3703SSong Liu 
958f6bed0efSShaohua Li 	for (i = 0; i < sh->disks; i++) {
959f6bed0efSShaohua Li 		void *addr;
960f6bed0efSShaohua Li 
9611e6d690bSSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
9621e6d690bSSong Liu 		    test_bit(R5_InJournal, &sh->dev[i].flags))
963f6bed0efSShaohua Li 			continue;
9641e6d690bSSong Liu 
965f6bed0efSShaohua Li 		write_disks++;
966f6bed0efSShaohua Li 		/* checksum is already calculated in last run */
967f6bed0efSShaohua Li 		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
968f6bed0efSShaohua Li 			continue;
969f6bed0efSShaohua Li 		addr = kmap_atomic(sh->dev[i].page);
9705cb2fbd6SShaohua Li 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
971f6bed0efSShaohua Li 						    addr, PAGE_SIZE);
972f6bed0efSShaohua Li 		kunmap_atomic(addr);
973f6bed0efSShaohua Li 	}
974f6bed0efSShaohua Li 	parity_pages = 1 + !!(sh->qd_idx >= 0);
975f6bed0efSShaohua Li 	data_pages = write_disks - parity_pages;
976f6bed0efSShaohua Li 
977f6bed0efSShaohua Li 	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
978253f9fd4SShaohua Li 	/*
979253f9fd4SShaohua Li 	 * The stripe must enter state machine again to finish the write, so
980253f9fd4SShaohua Li 	 * don't delay.
981253f9fd4SShaohua Li 	 */
982253f9fd4SShaohua Li 	clear_bit(STRIPE_DELAYED, &sh->state);
983f6bed0efSShaohua Li 	atomic_inc(&sh->count);
984f6bed0efSShaohua Li 
985f6bed0efSShaohua Li 	mutex_lock(&log->io_mutex);
986f6bed0efSShaohua Li 	/* meta + data */
987f6bed0efSShaohua Li 	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
988f6bed0efSShaohua Li 
989a39f7afdSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
990a39f7afdSSong Liu 		if (!r5l_has_free_space(log, reserve)) {
991a39f7afdSSong Liu 			r5l_add_no_space_stripe(log, sh);
992a39f7afdSSong Liu 			wake_reclaim = true;
9935036c390SChristoph Hellwig 		} else {
9945036c390SChristoph Hellwig 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
9955036c390SChristoph Hellwig 			if (ret) {
9965036c390SChristoph Hellwig 				spin_lock_irq(&log->io_list_lock);
997a39f7afdSSong Liu 				list_add_tail(&sh->log_list,
998a39f7afdSSong Liu 					      &log->no_mem_stripes);
9995036c390SChristoph Hellwig 				spin_unlock_irq(&log->io_list_lock);
1000f6bed0efSShaohua Li 			}
10015036c390SChristoph Hellwig 		}
1002a39f7afdSSong Liu 	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1003a39f7afdSSong Liu 		/*
1004a39f7afdSSong Liu 		 * log space critical, do not process stripes that are
1005a39f7afdSSong Liu 		 * not in cache yet (sh->log_start == MaxSector).
1006a39f7afdSSong Liu 		 */
1007a39f7afdSSong Liu 		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1008a39f7afdSSong Liu 		    sh->log_start == MaxSector) {
1009a39f7afdSSong Liu 			r5l_add_no_space_stripe(log, sh);
1010a39f7afdSSong Liu 			wake_reclaim = true;
1011a39f7afdSSong Liu 			reserve = 0;
1012a39f7afdSSong Liu 		} else if (!r5l_has_free_space(log, reserve)) {
1013a39f7afdSSong Liu 			if (sh->log_start == log->last_checkpoint)
1014a39f7afdSSong Liu 				BUG();
1015a39f7afdSSong Liu 			else
1016a39f7afdSSong Liu 				r5l_add_no_space_stripe(log, sh);
1017a39f7afdSSong Liu 		} else {
1018a39f7afdSSong Liu 			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1019a39f7afdSSong Liu 			if (ret) {
1020a39f7afdSSong Liu 				spin_lock_irq(&log->io_list_lock);
1021a39f7afdSSong Liu 				list_add_tail(&sh->log_list,
1022a39f7afdSSong Liu 					      &log->no_mem_stripes);
1023a39f7afdSSong Liu 				spin_unlock_irq(&log->io_list_lock);
1024a39f7afdSSong Liu 			}
1025a39f7afdSSong Liu 		}
1026a39f7afdSSong Liu 	}
1027f6bed0efSShaohua Li 
10285036c390SChristoph Hellwig 	mutex_unlock(&log->io_mutex);
1029a39f7afdSSong Liu 	if (wake_reclaim)
1030a39f7afdSSong Liu 		r5l_wake_reclaim(log, reserve);
1031f6bed0efSShaohua Li 	return 0;
1032f6bed0efSShaohua Li }
1033f6bed0efSShaohua Li 
1034f6bed0efSShaohua Li void r5l_write_stripe_run(struct r5l_log *log)
1035f6bed0efSShaohua Li {
1036f6bed0efSShaohua Li 	if (!log)
1037f6bed0efSShaohua Li 		return;
1038f6bed0efSShaohua Li 	mutex_lock(&log->io_mutex);
1039f6bed0efSShaohua Li 	r5l_submit_current_io(log);
1040f6bed0efSShaohua Li 	mutex_unlock(&log->io_mutex);
1041f6bed0efSShaohua Li }
1042f6bed0efSShaohua Li 
1043828cbe98SShaohua Li int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1044828cbe98SShaohua Li {
1045828cbe98SShaohua Li 	if (!log)
1046828cbe98SShaohua Li 		return -ENODEV;
10473bddb7f8SSong Liu 
10483bddb7f8SSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1049828cbe98SShaohua Li 		/*
10503bddb7f8SSong Liu 		 * in write through (journal only)
10513bddb7f8SSong Liu 		 * we flush log disk cache first, then write stripe data to
10523bddb7f8SSong Liu 		 * raid disks. So if bio is finished, the log disk cache is
10533bddb7f8SSong Liu 		 * flushed already. The recovery guarantees we can recovery
10543bddb7f8SSong Liu 		 * the bio from log disk, so we don't need to flush again
1055828cbe98SShaohua Li 		 */
1056828cbe98SShaohua Li 		if (bio->bi_iter.bi_size == 0) {
1057828cbe98SShaohua Li 			bio_endio(bio);
1058828cbe98SShaohua Li 			return 0;
1059828cbe98SShaohua Li 		}
10601eff9d32SJens Axboe 		bio->bi_opf &= ~REQ_PREFLUSH;
10613bddb7f8SSong Liu 	} else {
10623bddb7f8SSong Liu 		/* write back (with cache) */
10633bddb7f8SSong Liu 		if (bio->bi_iter.bi_size == 0) {
10643bddb7f8SSong Liu 			mutex_lock(&log->io_mutex);
10653bddb7f8SSong Liu 			r5l_get_meta(log, 0);
10663bddb7f8SSong Liu 			bio_list_add(&log->current_io->flush_barriers, bio);
10673bddb7f8SSong Liu 			log->current_io->has_flush = 1;
10683bddb7f8SSong Liu 			log->current_io->has_null_flush = 1;
10693bddb7f8SSong Liu 			atomic_inc(&log->current_io->pending_stripe);
10703bddb7f8SSong Liu 			r5l_submit_current_io(log);
10713bddb7f8SSong Liu 			mutex_unlock(&log->io_mutex);
10723bddb7f8SSong Liu 			return 0;
10733bddb7f8SSong Liu 		}
10743bddb7f8SSong Liu 	}
1075828cbe98SShaohua Li 	return -EAGAIN;
1076828cbe98SShaohua Li }
1077828cbe98SShaohua Li 
1078f6bed0efSShaohua Li /* This will run after log space is reclaimed */
1079f6bed0efSShaohua Li static void r5l_run_no_space_stripes(struct r5l_log *log)
1080f6bed0efSShaohua Li {
1081f6bed0efSShaohua Li 	struct stripe_head *sh;
1082f6bed0efSShaohua Li 
1083f6bed0efSShaohua Li 	spin_lock(&log->no_space_stripes_lock);
1084f6bed0efSShaohua Li 	while (!list_empty(&log->no_space_stripes)) {
1085f6bed0efSShaohua Li 		sh = list_first_entry(&log->no_space_stripes,
1086f6bed0efSShaohua Li 				      struct stripe_head, log_list);
1087f6bed0efSShaohua Li 		list_del_init(&sh->log_list);
1088f6bed0efSShaohua Li 		set_bit(STRIPE_HANDLE, &sh->state);
1089f6bed0efSShaohua Li 		raid5_release_stripe(sh);
1090f6bed0efSShaohua Li 	}
1091f6bed0efSShaohua Li 	spin_unlock(&log->no_space_stripes_lock);
1092f6bed0efSShaohua Li }
1093f6bed0efSShaohua Li 
1094a39f7afdSSong Liu /*
1095a39f7afdSSong Liu  * calculate new last_checkpoint
1096a39f7afdSSong Liu  * for write through mode, returns log->next_checkpoint
1097a39f7afdSSong Liu  * for write back, returns log_start of first sh in stripe_in_journal_list
1098a39f7afdSSong Liu  */
1099a39f7afdSSong Liu static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1100a39f7afdSSong Liu {
1101a39f7afdSSong Liu 	struct stripe_head *sh;
1102a39f7afdSSong Liu 	struct r5l_log *log = conf->log;
1103a39f7afdSSong Liu 	sector_t new_cp;
1104a39f7afdSSong Liu 	unsigned long flags;
1105a39f7afdSSong Liu 
1106a39f7afdSSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1107a39f7afdSSong Liu 		return log->next_checkpoint;
1108a39f7afdSSong Liu 
1109a39f7afdSSong Liu 	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1110a39f7afdSSong Liu 	if (list_empty(&conf->log->stripe_in_journal_list)) {
1111a39f7afdSSong Liu 		/* all stripes flushed */
1112d3014e21SDan Carpenter 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1113a39f7afdSSong Liu 		return log->next_checkpoint;
1114a39f7afdSSong Liu 	}
1115a39f7afdSSong Liu 	sh = list_first_entry(&conf->log->stripe_in_journal_list,
1116a39f7afdSSong Liu 			      struct stripe_head, r5c);
1117a39f7afdSSong Liu 	new_cp = sh->log_start;
1118a39f7afdSSong Liu 	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1119a39f7afdSSong Liu 	return new_cp;
1120a39f7afdSSong Liu }
1121a39f7afdSSong Liu 
112217036461SChristoph Hellwig static sector_t r5l_reclaimable_space(struct r5l_log *log)
112317036461SChristoph Hellwig {
1124a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
1125a39f7afdSSong Liu 
112617036461SChristoph Hellwig 	return r5l_ring_distance(log, log->last_checkpoint,
1127a39f7afdSSong Liu 				 r5c_calculate_new_cp(conf));
112817036461SChristoph Hellwig }
112917036461SChristoph Hellwig 
11305036c390SChristoph Hellwig static void r5l_run_no_mem_stripe(struct r5l_log *log)
11315036c390SChristoph Hellwig {
11325036c390SChristoph Hellwig 	struct stripe_head *sh;
11335036c390SChristoph Hellwig 
11345036c390SChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
11355036c390SChristoph Hellwig 
11365036c390SChristoph Hellwig 	if (!list_empty(&log->no_mem_stripes)) {
11375036c390SChristoph Hellwig 		sh = list_first_entry(&log->no_mem_stripes,
11385036c390SChristoph Hellwig 				      struct stripe_head, log_list);
11395036c390SChristoph Hellwig 		list_del_init(&sh->log_list);
11405036c390SChristoph Hellwig 		set_bit(STRIPE_HANDLE, &sh->state);
11415036c390SChristoph Hellwig 		raid5_release_stripe(sh);
11425036c390SChristoph Hellwig 	}
11435036c390SChristoph Hellwig }
11445036c390SChristoph Hellwig 
114504732f74SChristoph Hellwig static bool r5l_complete_finished_ios(struct r5l_log *log)
114617036461SChristoph Hellwig {
114717036461SChristoph Hellwig 	struct r5l_io_unit *io, *next;
114817036461SChristoph Hellwig 	bool found = false;
114917036461SChristoph Hellwig 
115017036461SChristoph Hellwig 	assert_spin_locked(&log->io_list_lock);
115117036461SChristoph Hellwig 
115204732f74SChristoph Hellwig 	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
115317036461SChristoph Hellwig 		/* don't change list order */
115417036461SChristoph Hellwig 		if (io->state < IO_UNIT_STRIPE_END)
115517036461SChristoph Hellwig 			break;
115617036461SChristoph Hellwig 
115717036461SChristoph Hellwig 		log->next_checkpoint = io->log_start;
115817036461SChristoph Hellwig 
115917036461SChristoph Hellwig 		list_del(&io->log_sibling);
11605036c390SChristoph Hellwig 		mempool_free(io, log->io_pool);
11615036c390SChristoph Hellwig 		r5l_run_no_mem_stripe(log);
116217036461SChristoph Hellwig 
116317036461SChristoph Hellwig 		found = true;
116417036461SChristoph Hellwig 	}
116517036461SChristoph Hellwig 
116617036461SChristoph Hellwig 	return found;
116717036461SChristoph Hellwig }
116817036461SChristoph Hellwig 
1169509ffec7SChristoph Hellwig static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1170509ffec7SChristoph Hellwig {
1171509ffec7SChristoph Hellwig 	struct r5l_log *log = io->log;
1172a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
1173509ffec7SChristoph Hellwig 	unsigned long flags;
1174509ffec7SChristoph Hellwig 
1175509ffec7SChristoph Hellwig 	spin_lock_irqsave(&log->io_list_lock, flags);
1176509ffec7SChristoph Hellwig 	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
117717036461SChristoph Hellwig 
117804732f74SChristoph Hellwig 	if (!r5l_complete_finished_ios(log)) {
117985f2f9a4SShaohua Li 		spin_unlock_irqrestore(&log->io_list_lock, flags);
118085f2f9a4SShaohua Li 		return;
118185f2f9a4SShaohua Li 	}
1182509ffec7SChristoph Hellwig 
1183a39f7afdSSong Liu 	if (r5l_reclaimable_space(log) > log->max_free_space ||
1184a39f7afdSSong Liu 	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1185509ffec7SChristoph Hellwig 		r5l_wake_reclaim(log, 0);
1186509ffec7SChristoph Hellwig 
1187509ffec7SChristoph Hellwig 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1188509ffec7SChristoph Hellwig 	wake_up(&log->iounit_wait);
1189509ffec7SChristoph Hellwig }
1190509ffec7SChristoph Hellwig 
11910576b1c6SShaohua Li void r5l_stripe_write_finished(struct stripe_head *sh)
11920576b1c6SShaohua Li {
11930576b1c6SShaohua Li 	struct r5l_io_unit *io;
11940576b1c6SShaohua Li 
11950576b1c6SShaohua Li 	io = sh->log_io;
11960576b1c6SShaohua Li 	sh->log_io = NULL;
11970576b1c6SShaohua Li 
1198509ffec7SChristoph Hellwig 	if (io && atomic_dec_and_test(&io->pending_stripe))
1199509ffec7SChristoph Hellwig 		__r5l_stripe_write_finished(io);
12000576b1c6SShaohua Li }
12010576b1c6SShaohua Li 
1202a8c34f91SShaohua Li static void r5l_log_flush_endio(struct bio *bio)
1203a8c34f91SShaohua Li {
1204a8c34f91SShaohua Li 	struct r5l_log *log = container_of(bio, struct r5l_log,
1205a8c34f91SShaohua Li 		flush_bio);
1206a8c34f91SShaohua Li 	unsigned long flags;
1207a8c34f91SShaohua Li 	struct r5l_io_unit *io;
1208a8c34f91SShaohua Li 
12096e74a9cfSShaohua Li 	if (bio->bi_error)
12106e74a9cfSShaohua Li 		md_error(log->rdev->mddev, log->rdev);
12116e74a9cfSShaohua Li 
1212a8c34f91SShaohua Li 	spin_lock_irqsave(&log->io_list_lock, flags);
1213d8858f43SChristoph Hellwig 	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1214d8858f43SChristoph Hellwig 		r5l_io_run_stripes(io);
121504732f74SChristoph Hellwig 	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1216a8c34f91SShaohua Li 	spin_unlock_irqrestore(&log->io_list_lock, flags);
1217a8c34f91SShaohua Li }
1218a8c34f91SShaohua Li 
12190576b1c6SShaohua Li /*
12200576b1c6SShaohua Li  * Starting dispatch IO to raid.
12210576b1c6SShaohua Li  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
12220576b1c6SShaohua Li  * broken meta in the middle of a log causes recovery can't find meta at the
12230576b1c6SShaohua Li  * head of log. If operations require meta at the head persistent in log, we
12240576b1c6SShaohua Li  * must make sure meta before it persistent in log too. A case is:
12250576b1c6SShaohua Li  *
12260576b1c6SShaohua Li  * stripe data/parity is in log, we start write stripe to raid disks. stripe
12270576b1c6SShaohua Li  * data/parity must be persistent in log before we do the write to raid disks.
12280576b1c6SShaohua Li  *
12290576b1c6SShaohua Li  * The solution is we restrictly maintain io_unit list order. In this case, we
12300576b1c6SShaohua Li  * only write stripes of an io_unit to raid disks till the io_unit is the first
12310576b1c6SShaohua Li  * one whose data/parity is in log.
12320576b1c6SShaohua Li  */
12330576b1c6SShaohua Li void r5l_flush_stripe_to_raid(struct r5l_log *log)
12340576b1c6SShaohua Li {
1235a8c34f91SShaohua Li 	bool do_flush;
123656fef7c6SChristoph Hellwig 
123756fef7c6SChristoph Hellwig 	if (!log || !log->need_cache_flush)
12380576b1c6SShaohua Li 		return;
12390576b1c6SShaohua Li 
1240a8c34f91SShaohua Li 	spin_lock_irq(&log->io_list_lock);
1241a8c34f91SShaohua Li 	/* flush bio is running */
1242a8c34f91SShaohua Li 	if (!list_empty(&log->flushing_ios)) {
1243a8c34f91SShaohua Li 		spin_unlock_irq(&log->io_list_lock);
12440576b1c6SShaohua Li 		return;
12450576b1c6SShaohua Li 	}
1246a8c34f91SShaohua Li 	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1247a8c34f91SShaohua Li 	do_flush = !list_empty(&log->flushing_ios);
12480576b1c6SShaohua Li 	spin_unlock_irq(&log->io_list_lock);
1249a8c34f91SShaohua Li 
1250a8c34f91SShaohua Li 	if (!do_flush)
1251a8c34f91SShaohua Li 		return;
1252a8c34f91SShaohua Li 	bio_reset(&log->flush_bio);
1253a8c34f91SShaohua Li 	log->flush_bio.bi_bdev = log->rdev->bdev;
1254a8c34f91SShaohua Li 	log->flush_bio.bi_end_io = r5l_log_flush_endio;
125570fd7614SChristoph Hellwig 	log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
12564e49ea4aSMike Christie 	submit_bio(&log->flush_bio);
12570576b1c6SShaohua Li }
12580576b1c6SShaohua Li 
12590576b1c6SShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp);
12604b482044SShaohua Li static void r5l_write_super_and_discard_space(struct r5l_log *log,
12614b482044SShaohua Li 	sector_t end)
12624b482044SShaohua Li {
12634b482044SShaohua Li 	struct block_device *bdev = log->rdev->bdev;
12644b482044SShaohua Li 	struct mddev *mddev;
12654b482044SShaohua Li 
12664b482044SShaohua Li 	r5l_write_super(log, end);
12674b482044SShaohua Li 
12684b482044SShaohua Li 	if (!blk_queue_discard(bdev_get_queue(bdev)))
12694b482044SShaohua Li 		return;
12704b482044SShaohua Li 
12714b482044SShaohua Li 	mddev = log->rdev->mddev;
12724b482044SShaohua Li 	/*
12738e018c21SShaohua Li 	 * Discard could zero data, so before discard we must make sure
12748e018c21SShaohua Li 	 * superblock is updated to new log tail. Updating superblock (either
12758e018c21SShaohua Li 	 * directly call md_update_sb() or depend on md thread) must hold
12768e018c21SShaohua Li 	 * reconfig mutex. On the other hand, raid5_quiesce is called with
12778e018c21SShaohua Li 	 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
12788e018c21SShaohua Li 	 * for all IO finish, hence waitting for reclaim thread, while reclaim
12798e018c21SShaohua Li 	 * thread is calling this function and waitting for reconfig mutex. So
12808e018c21SShaohua Li 	 * there is a deadlock. We workaround this issue with a trylock.
12818e018c21SShaohua Li 	 * FIXME: we could miss discard if we can't take reconfig mutex
12824b482044SShaohua Li 	 */
12832953079cSShaohua Li 	set_mask_bits(&mddev->sb_flags, 0,
12842953079cSShaohua Li 		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
12858e018c21SShaohua Li 	if (!mddev_trylock(mddev))
12868e018c21SShaohua Li 		return;
12874b482044SShaohua Li 	md_update_sb(mddev, 1);
12888e018c21SShaohua Li 	mddev_unlock(mddev);
12894b482044SShaohua Li 
12906e74a9cfSShaohua Li 	/* discard IO error really doesn't matter, ignore it */
12914b482044SShaohua Li 	if (log->last_checkpoint < end) {
12924b482044SShaohua Li 		blkdev_issue_discard(bdev,
12934b482044SShaohua Li 				log->last_checkpoint + log->rdev->data_offset,
12944b482044SShaohua Li 				end - log->last_checkpoint, GFP_NOIO, 0);
12954b482044SShaohua Li 	} else {
12964b482044SShaohua Li 		blkdev_issue_discard(bdev,
12974b482044SShaohua Li 				log->last_checkpoint + log->rdev->data_offset,
12984b482044SShaohua Li 				log->device_size - log->last_checkpoint,
12994b482044SShaohua Li 				GFP_NOIO, 0);
13004b482044SShaohua Li 		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
13014b482044SShaohua Li 				GFP_NOIO, 0);
13024b482044SShaohua Li 	}
13034b482044SShaohua Li }
13044b482044SShaohua Li 
1305a39f7afdSSong Liu /*
1306a39f7afdSSong Liu  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1307a39f7afdSSong Liu  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1308a39f7afdSSong Liu  *
1309a39f7afdSSong Liu  * must hold conf->device_lock
1310a39f7afdSSong Liu  */
1311a39f7afdSSong Liu static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1312a39f7afdSSong Liu {
1313a39f7afdSSong Liu 	BUG_ON(list_empty(&sh->lru));
1314a39f7afdSSong Liu 	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1315a39f7afdSSong Liu 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1316a39f7afdSSong Liu 
1317a39f7afdSSong Liu 	/*
1318a39f7afdSSong Liu 	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1319a39f7afdSSong Liu 	 * raid5_release_stripe() while holding conf->device_lock
1320a39f7afdSSong Liu 	 */
1321a39f7afdSSong Liu 	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1322a39f7afdSSong Liu 	assert_spin_locked(&conf->device_lock);
1323a39f7afdSSong Liu 
1324a39f7afdSSong Liu 	list_del_init(&sh->lru);
1325a39f7afdSSong Liu 	atomic_inc(&sh->count);
1326a39f7afdSSong Liu 
1327a39f7afdSSong Liu 	set_bit(STRIPE_HANDLE, &sh->state);
1328a39f7afdSSong Liu 	atomic_inc(&conf->active_stripes);
1329a39f7afdSSong Liu 	r5c_make_stripe_write_out(sh);
1330a39f7afdSSong Liu 
1331e33fbb9cSShaohua Li 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1332e33fbb9cSShaohua Li 		atomic_inc(&conf->r5c_flushing_partial_stripes);
1333e33fbb9cSShaohua Li 	else
1334e33fbb9cSShaohua Li 		atomic_inc(&conf->r5c_flushing_full_stripes);
1335a39f7afdSSong Liu 	raid5_release_stripe(sh);
1336a39f7afdSSong Liu }
1337a39f7afdSSong Liu 
1338a39f7afdSSong Liu /*
1339a39f7afdSSong Liu  * if num == 0, flush all full stripes
1340a39f7afdSSong Liu  * if num > 0, flush all full stripes. If less than num full stripes are
1341a39f7afdSSong Liu  *             flushed, flush some partial stripes until totally num stripes are
1342a39f7afdSSong Liu  *             flushed or there is no more cached stripes.
1343a39f7afdSSong Liu  */
1344a39f7afdSSong Liu void r5c_flush_cache(struct r5conf *conf, int num)
1345a39f7afdSSong Liu {
1346a39f7afdSSong Liu 	int count;
1347a39f7afdSSong Liu 	struct stripe_head *sh, *next;
1348a39f7afdSSong Liu 
1349a39f7afdSSong Liu 	assert_spin_locked(&conf->device_lock);
1350a39f7afdSSong Liu 	if (!conf->log)
1351a39f7afdSSong Liu 		return;
1352a39f7afdSSong Liu 
1353a39f7afdSSong Liu 	count = 0;
1354a39f7afdSSong Liu 	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1355a39f7afdSSong Liu 		r5c_flush_stripe(conf, sh);
1356a39f7afdSSong Liu 		count++;
1357a39f7afdSSong Liu 	}
1358a39f7afdSSong Liu 
1359a39f7afdSSong Liu 	if (count >= num)
1360a39f7afdSSong Liu 		return;
1361a39f7afdSSong Liu 	list_for_each_entry_safe(sh, next,
1362a39f7afdSSong Liu 				 &conf->r5c_partial_stripe_list, lru) {
1363a39f7afdSSong Liu 		r5c_flush_stripe(conf, sh);
1364a39f7afdSSong Liu 		if (++count >= num)
1365a39f7afdSSong Liu 			break;
1366a39f7afdSSong Liu 	}
1367a39f7afdSSong Liu }
1368a39f7afdSSong Liu 
1369a39f7afdSSong Liu static void r5c_do_reclaim(struct r5conf *conf)
1370a39f7afdSSong Liu {
1371a39f7afdSSong Liu 	struct r5l_log *log = conf->log;
1372a39f7afdSSong Liu 	struct stripe_head *sh;
1373a39f7afdSSong Liu 	int count = 0;
1374a39f7afdSSong Liu 	unsigned long flags;
1375a39f7afdSSong Liu 	int total_cached;
1376a39f7afdSSong Liu 	int stripes_to_flush;
1377e33fbb9cSShaohua Li 	int flushing_partial, flushing_full;
1378a39f7afdSSong Liu 
1379a39f7afdSSong Liu 	if (!r5c_is_writeback(log))
1380a39f7afdSSong Liu 		return;
1381a39f7afdSSong Liu 
1382e33fbb9cSShaohua Li 	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1383e33fbb9cSShaohua Li 	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1384a39f7afdSSong Liu 	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1385e33fbb9cSShaohua Li 		atomic_read(&conf->r5c_cached_full_stripes) -
1386e33fbb9cSShaohua Li 		flushing_full - flushing_partial;
1387a39f7afdSSong Liu 
1388a39f7afdSSong Liu 	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1389a39f7afdSSong Liu 	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1390a39f7afdSSong Liu 		/*
1391a39f7afdSSong Liu 		 * if stripe cache pressure high, flush all full stripes and
1392a39f7afdSSong Liu 		 * some partial stripes
1393a39f7afdSSong Liu 		 */
1394a39f7afdSSong Liu 		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1395a39f7afdSSong Liu 	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1396e33fbb9cSShaohua Li 		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
139784890c03SShaohua Li 		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1398a39f7afdSSong Liu 		/*
1399a39f7afdSSong Liu 		 * if stripe cache pressure moderate, or if there is many full
1400a39f7afdSSong Liu 		 * stripes,flush all full stripes
1401a39f7afdSSong Liu 		 */
1402a39f7afdSSong Liu 		stripes_to_flush = 0;
1403a39f7afdSSong Liu 	else
1404a39f7afdSSong Liu 		/* no need to flush */
1405a39f7afdSSong Liu 		stripes_to_flush = -1;
1406a39f7afdSSong Liu 
1407a39f7afdSSong Liu 	if (stripes_to_flush >= 0) {
1408a39f7afdSSong Liu 		spin_lock_irqsave(&conf->device_lock, flags);
1409a39f7afdSSong Liu 		r5c_flush_cache(conf, stripes_to_flush);
1410a39f7afdSSong Liu 		spin_unlock_irqrestore(&conf->device_lock, flags);
1411a39f7afdSSong Liu 	}
1412a39f7afdSSong Liu 
1413a39f7afdSSong Liu 	/* if log space is tight, flush stripes on stripe_in_journal_list */
1414a39f7afdSSong Liu 	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1415a39f7afdSSong Liu 		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1416a39f7afdSSong Liu 		spin_lock(&conf->device_lock);
1417a39f7afdSSong Liu 		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1418a39f7afdSSong Liu 			/*
1419a39f7afdSSong Liu 			 * stripes on stripe_in_journal_list could be in any
1420a39f7afdSSong Liu 			 * state of the stripe_cache state machine. In this
1421a39f7afdSSong Liu 			 * case, we only want to flush stripe on
1422a39f7afdSSong Liu 			 * r5c_cached_full/partial_stripes. The following
1423a39f7afdSSong Liu 			 * condition makes sure the stripe is on one of the
1424a39f7afdSSong Liu 			 * two lists.
1425a39f7afdSSong Liu 			 */
1426a39f7afdSSong Liu 			if (!list_empty(&sh->lru) &&
1427a39f7afdSSong Liu 			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1428a39f7afdSSong Liu 			    atomic_read(&sh->count) == 0) {
1429a39f7afdSSong Liu 				r5c_flush_stripe(conf, sh);
1430a39f7afdSSong Liu 				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1431a39f7afdSSong Liu 					break;
1432a39f7afdSSong Liu 			}
1433e8fd52eeSShaohua Li 		}
1434a39f7afdSSong Liu 		spin_unlock(&conf->device_lock);
1435a39f7afdSSong Liu 		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1436a39f7afdSSong Liu 	}
1437f687a33eSSong Liu 
1438f687a33eSSong Liu 	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1439f687a33eSSong Liu 		r5l_run_no_space_stripes(log);
1440f687a33eSSong Liu 
1441a39f7afdSSong Liu 	md_wakeup_thread(conf->mddev->thread);
1442a39f7afdSSong Liu }
14434b482044SShaohua Li 
14440576b1c6SShaohua Li static void r5l_do_reclaim(struct r5l_log *log)
14450576b1c6SShaohua Li {
1446a39f7afdSSong Liu 	struct r5conf *conf = log->rdev->mddev->private;
14470576b1c6SShaohua Li 	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
144817036461SChristoph Hellwig 	sector_t reclaimable;
144917036461SChristoph Hellwig 	sector_t next_checkpoint;
1450a39f7afdSSong Liu 	bool write_super;
14510576b1c6SShaohua Li 
14520576b1c6SShaohua Li 	spin_lock_irq(&log->io_list_lock);
1453a39f7afdSSong Liu 	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1454a39f7afdSSong Liu 		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
14550576b1c6SShaohua Li 	/*
14560576b1c6SShaohua Li 	 * move proper io_unit to reclaim list. We should not change the order.
14570576b1c6SShaohua Li 	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
14580576b1c6SShaohua Li 	 * shouldn't reuse space of an unreclaimable io_unit
14590576b1c6SShaohua Li 	 */
14600576b1c6SShaohua Li 	while (1) {
146117036461SChristoph Hellwig 		reclaimable = r5l_reclaimable_space(log);
146217036461SChristoph Hellwig 		if (reclaimable >= reclaim_target ||
14630576b1c6SShaohua Li 		    (list_empty(&log->running_ios) &&
14640576b1c6SShaohua Li 		     list_empty(&log->io_end_ios) &&
1465a8c34f91SShaohua Li 		     list_empty(&log->flushing_ios) &&
146604732f74SChristoph Hellwig 		     list_empty(&log->finished_ios)))
14670576b1c6SShaohua Li 			break;
14680576b1c6SShaohua Li 
146917036461SChristoph Hellwig 		md_wakeup_thread(log->rdev->mddev->thread);
147017036461SChristoph Hellwig 		wait_event_lock_irq(log->iounit_wait,
147117036461SChristoph Hellwig 				    r5l_reclaimable_space(log) > reclaimable,
147217036461SChristoph Hellwig 				    log->io_list_lock);
14730576b1c6SShaohua Li 	}
147417036461SChristoph Hellwig 
1475a39f7afdSSong Liu 	next_checkpoint = r5c_calculate_new_cp(conf);
14760576b1c6SShaohua Li 	spin_unlock_irq(&log->io_list_lock);
14770576b1c6SShaohua Li 
1478a39f7afdSSong Liu 	if (reclaimable == 0 || !write_super)
14790576b1c6SShaohua Li 		return;
14800576b1c6SShaohua Li 
14810576b1c6SShaohua Li 	/*
14820576b1c6SShaohua Li 	 * write_super will flush cache of each raid disk. We must write super
14830576b1c6SShaohua Li 	 * here, because the log area might be reused soon and we don't want to
14840576b1c6SShaohua Li 	 * confuse recovery
14850576b1c6SShaohua Li 	 */
14864b482044SShaohua Li 	r5l_write_super_and_discard_space(log, next_checkpoint);
14870576b1c6SShaohua Li 
14880576b1c6SShaohua Li 	mutex_lock(&log->io_mutex);
148917036461SChristoph Hellwig 	log->last_checkpoint = next_checkpoint;
1490a39f7afdSSong Liu 	r5c_update_log_state(log);
14910576b1c6SShaohua Li 	mutex_unlock(&log->io_mutex);
14920576b1c6SShaohua Li 
149317036461SChristoph Hellwig 	r5l_run_no_space_stripes(log);
14940576b1c6SShaohua Li }
14950576b1c6SShaohua Li 
14960576b1c6SShaohua Li static void r5l_reclaim_thread(struct md_thread *thread)
14970576b1c6SShaohua Li {
14980576b1c6SShaohua Li 	struct mddev *mddev = thread->mddev;
14990576b1c6SShaohua Li 	struct r5conf *conf = mddev->private;
15000576b1c6SShaohua Li 	struct r5l_log *log = conf->log;
15010576b1c6SShaohua Li 
15020576b1c6SShaohua Li 	if (!log)
15030576b1c6SShaohua Li 		return;
1504a39f7afdSSong Liu 	r5c_do_reclaim(conf);
15050576b1c6SShaohua Li 	r5l_do_reclaim(log);
15060576b1c6SShaohua Li }
15070576b1c6SShaohua Li 
1508a39f7afdSSong Liu void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1509f6bed0efSShaohua Li {
15100576b1c6SShaohua Li 	unsigned long target;
15110576b1c6SShaohua Li 	unsigned long new = (unsigned long)space; /* overflow in theory */
15120576b1c6SShaohua Li 
1513a39f7afdSSong Liu 	if (!log)
1514a39f7afdSSong Liu 		return;
15150576b1c6SShaohua Li 	do {
15160576b1c6SShaohua Li 		target = log->reclaim_target;
15170576b1c6SShaohua Li 		if (new < target)
15180576b1c6SShaohua Li 			return;
15190576b1c6SShaohua Li 	} while (cmpxchg(&log->reclaim_target, target, new) != target);
15200576b1c6SShaohua Li 	md_wakeup_thread(log->reclaim_thread);
1521f6bed0efSShaohua Li }
1522f6bed0efSShaohua Li 
1523e6c033f7SShaohua Li void r5l_quiesce(struct r5l_log *log, int state)
1524e6c033f7SShaohua Li {
15254b482044SShaohua Li 	struct mddev *mddev;
1526e6c033f7SShaohua Li 	if (!log || state == 2)
1527e6c033f7SShaohua Li 		return;
1528ce1ccd07SShaohua Li 	if (state == 0)
1529ce1ccd07SShaohua Li 		kthread_unpark(log->reclaim_thread->tsk);
1530ce1ccd07SShaohua Li 	else if (state == 1) {
15314b482044SShaohua Li 		/* make sure r5l_write_super_and_discard_space exits */
15324b482044SShaohua Li 		mddev = log->rdev->mddev;
15334b482044SShaohua Li 		wake_up(&mddev->sb_wait);
1534ce1ccd07SShaohua Li 		kthread_park(log->reclaim_thread->tsk);
1535a39f7afdSSong Liu 		r5l_wake_reclaim(log, MaxSector);
1536e6c033f7SShaohua Li 		r5l_do_reclaim(log);
1537e6c033f7SShaohua Li 	}
1538e6c033f7SShaohua Li }
1539e6c033f7SShaohua Li 
15406e74a9cfSShaohua Li bool r5l_log_disk_error(struct r5conf *conf)
15416e74a9cfSShaohua Li {
1542f6b6ec5cSShaohua Li 	struct r5l_log *log;
1543f6b6ec5cSShaohua Li 	bool ret;
15447dde2ad3SShaohua Li 	/* don't allow write if journal disk is missing */
1545f6b6ec5cSShaohua Li 	rcu_read_lock();
1546f6b6ec5cSShaohua Li 	log = rcu_dereference(conf->log);
1547f6b6ec5cSShaohua Li 
1548f6b6ec5cSShaohua Li 	if (!log)
1549f6b6ec5cSShaohua Li 		ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1550f6b6ec5cSShaohua Li 	else
1551f6b6ec5cSShaohua Li 		ret = test_bit(Faulty, &log->rdev->flags);
1552f6b6ec5cSShaohua Li 	rcu_read_unlock();
1553f6b6ec5cSShaohua Li 	return ret;
15546e74a9cfSShaohua Li }
15556e74a9cfSShaohua Li 
1556*effe6ee7SSong Liu #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1557*effe6ee7SSong Liu 
1558355810d1SShaohua Li struct r5l_recovery_ctx {
1559355810d1SShaohua Li 	struct page *meta_page;		/* current meta */
1560355810d1SShaohua Li 	sector_t meta_total_blocks;	/* total size of current meta and data */
1561355810d1SShaohua Li 	sector_t pos;			/* recovery position */
1562355810d1SShaohua Li 	u64 seq;			/* recovery position seq */
1563b4c625c6SSong Liu 	int data_parity_stripes;	/* number of data_parity stripes */
1564b4c625c6SSong Liu 	int data_only_stripes;		/* number of data_only stripes */
1565b4c625c6SSong Liu 	struct list_head cached_list;
1566*effe6ee7SSong Liu 
1567*effe6ee7SSong Liu 	/*
1568*effe6ee7SSong Liu 	 * read ahead page pool (ra_pool)
1569*effe6ee7SSong Liu 	 * in recovery, log is read sequentially. It is not efficient to
1570*effe6ee7SSong Liu 	 * read every page with sync_page_io(). The read ahead page pool
1571*effe6ee7SSong Liu 	 * reads multiple pages with one IO, so further log read can
1572*effe6ee7SSong Liu 	 * just copy data from the pool.
1573*effe6ee7SSong Liu 	 */
1574*effe6ee7SSong Liu 	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1575*effe6ee7SSong Liu 	sector_t pool_offset;	/* offset of first page in the pool */
1576*effe6ee7SSong Liu 	int total_pages;	/* total allocated pages */
1577*effe6ee7SSong Liu 	int valid_pages;	/* pages with valid data */
1578*effe6ee7SSong Liu 	struct bio *ra_bio;	/* bio to do the read ahead */
1579355810d1SShaohua Li };
1580355810d1SShaohua Li 
1581*effe6ee7SSong Liu static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1582*effe6ee7SSong Liu 					    struct r5l_recovery_ctx *ctx)
1583*effe6ee7SSong Liu {
1584*effe6ee7SSong Liu 	struct page *page;
1585*effe6ee7SSong Liu 
1586*effe6ee7SSong Liu 	ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1587*effe6ee7SSong Liu 	if (!ctx->ra_bio)
1588*effe6ee7SSong Liu 		return -ENOMEM;
1589*effe6ee7SSong Liu 
1590*effe6ee7SSong Liu 	ctx->valid_pages = 0;
1591*effe6ee7SSong Liu 	ctx->total_pages = 0;
1592*effe6ee7SSong Liu 	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1593*effe6ee7SSong Liu 		page = alloc_page(GFP_KERNEL);
1594*effe6ee7SSong Liu 
1595*effe6ee7SSong Liu 		if (!page)
1596*effe6ee7SSong Liu 			break;
1597*effe6ee7SSong Liu 		ctx->ra_pool[ctx->total_pages] = page;
1598*effe6ee7SSong Liu 		ctx->total_pages += 1;
1599*effe6ee7SSong Liu 	}
1600*effe6ee7SSong Liu 
1601*effe6ee7SSong Liu 	if (ctx->total_pages == 0) {
1602*effe6ee7SSong Liu 		bio_put(ctx->ra_bio);
1603*effe6ee7SSong Liu 		return -ENOMEM;
1604*effe6ee7SSong Liu 	}
1605*effe6ee7SSong Liu 
1606*effe6ee7SSong Liu 	ctx->pool_offset = 0;
1607*effe6ee7SSong Liu 	return 0;
1608*effe6ee7SSong Liu }
1609*effe6ee7SSong Liu 
1610*effe6ee7SSong Liu static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1611*effe6ee7SSong Liu 					struct r5l_recovery_ctx *ctx)
1612*effe6ee7SSong Liu {
1613*effe6ee7SSong Liu 	int i;
1614*effe6ee7SSong Liu 
1615*effe6ee7SSong Liu 	for (i = 0; i < ctx->total_pages; ++i)
1616*effe6ee7SSong Liu 		put_page(ctx->ra_pool[i]);
1617*effe6ee7SSong Liu 	bio_put(ctx->ra_bio);
1618*effe6ee7SSong Liu }
1619*effe6ee7SSong Liu 
1620*effe6ee7SSong Liu /*
1621*effe6ee7SSong Liu  * fetch ctx->valid_pages pages from offset
1622*effe6ee7SSong Liu  * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1623*effe6ee7SSong Liu  * However, if the offset is close to the end of the journal device,
1624*effe6ee7SSong Liu  * ctx->valid_pages could be smaller than ctx->total_pages
1625*effe6ee7SSong Liu  */
1626*effe6ee7SSong Liu static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1627*effe6ee7SSong Liu 				      struct r5l_recovery_ctx *ctx,
1628*effe6ee7SSong Liu 				      sector_t offset)
1629*effe6ee7SSong Liu {
1630*effe6ee7SSong Liu 	bio_reset(ctx->ra_bio);
1631*effe6ee7SSong Liu 	ctx->ra_bio->bi_bdev = log->rdev->bdev;
1632*effe6ee7SSong Liu 	bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1633*effe6ee7SSong Liu 	ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1634*effe6ee7SSong Liu 
1635*effe6ee7SSong Liu 	ctx->valid_pages = 0;
1636*effe6ee7SSong Liu 	ctx->pool_offset = offset;
1637*effe6ee7SSong Liu 
1638*effe6ee7SSong Liu 	while (ctx->valid_pages < ctx->total_pages) {
1639*effe6ee7SSong Liu 		bio_add_page(ctx->ra_bio,
1640*effe6ee7SSong Liu 			     ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1641*effe6ee7SSong Liu 		ctx->valid_pages += 1;
1642*effe6ee7SSong Liu 
1643*effe6ee7SSong Liu 		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1644*effe6ee7SSong Liu 
1645*effe6ee7SSong Liu 		if (offset == 0)  /* reached end of the device */
1646*effe6ee7SSong Liu 			break;
1647*effe6ee7SSong Liu 	}
1648*effe6ee7SSong Liu 
1649*effe6ee7SSong Liu 	return submit_bio_wait(ctx->ra_bio);
1650*effe6ee7SSong Liu }
1651*effe6ee7SSong Liu 
1652*effe6ee7SSong Liu /*
1653*effe6ee7SSong Liu  * try read a page from the read ahead page pool, if the page is not in the
1654*effe6ee7SSong Liu  * pool, call r5l_recovery_fetch_ra_pool
1655*effe6ee7SSong Liu  */
1656*effe6ee7SSong Liu static int r5l_recovery_read_page(struct r5l_log *log,
1657*effe6ee7SSong Liu 				  struct r5l_recovery_ctx *ctx,
1658*effe6ee7SSong Liu 				  struct page *page,
1659*effe6ee7SSong Liu 				  sector_t offset)
1660*effe6ee7SSong Liu {
1661*effe6ee7SSong Liu 	int ret;
1662*effe6ee7SSong Liu 
1663*effe6ee7SSong Liu 	if (offset < ctx->pool_offset ||
1664*effe6ee7SSong Liu 	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1665*effe6ee7SSong Liu 		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1666*effe6ee7SSong Liu 		if (ret)
1667*effe6ee7SSong Liu 			return ret;
1668*effe6ee7SSong Liu 	}
1669*effe6ee7SSong Liu 
1670*effe6ee7SSong Liu 	BUG_ON(offset < ctx->pool_offset ||
1671*effe6ee7SSong Liu 	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1672*effe6ee7SSong Liu 
1673*effe6ee7SSong Liu 	memcpy(page_address(page),
1674*effe6ee7SSong Liu 	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1675*effe6ee7SSong Liu 					 BLOCK_SECTOR_SHIFT]),
1676*effe6ee7SSong Liu 	       PAGE_SIZE);
1677*effe6ee7SSong Liu 	return 0;
1678*effe6ee7SSong Liu }
1679*effe6ee7SSong Liu 
16809ed988f5SSong Liu static int r5l_recovery_read_meta_block(struct r5l_log *log,
1681355810d1SShaohua Li 					struct r5l_recovery_ctx *ctx)
1682355810d1SShaohua Li {
1683355810d1SShaohua Li 	struct page *page = ctx->meta_page;
1684355810d1SShaohua Li 	struct r5l_meta_block *mb;
1685355810d1SShaohua Li 	u32 crc, stored_crc;
1686*effe6ee7SSong Liu 	int ret;
1687355810d1SShaohua Li 
1688*effe6ee7SSong Liu 	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1689*effe6ee7SSong Liu 	if (ret != 0)
1690*effe6ee7SSong Liu 		return ret;
1691355810d1SShaohua Li 
1692355810d1SShaohua Li 	mb = page_address(page);
1693355810d1SShaohua Li 	stored_crc = le32_to_cpu(mb->checksum);
1694355810d1SShaohua Li 	mb->checksum = 0;
1695355810d1SShaohua Li 
1696355810d1SShaohua Li 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1697355810d1SShaohua Li 	    le64_to_cpu(mb->seq) != ctx->seq ||
1698355810d1SShaohua Li 	    mb->version != R5LOG_VERSION ||
1699355810d1SShaohua Li 	    le64_to_cpu(mb->position) != ctx->pos)
1700355810d1SShaohua Li 		return -EINVAL;
1701355810d1SShaohua Li 
17025cb2fbd6SShaohua Li 	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1703355810d1SShaohua Li 	if (stored_crc != crc)
1704355810d1SShaohua Li 		return -EINVAL;
1705355810d1SShaohua Li 
1706355810d1SShaohua Li 	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1707355810d1SShaohua Li 		return -EINVAL;
1708355810d1SShaohua Li 
1709355810d1SShaohua Li 	ctx->meta_total_blocks = BLOCK_SECTORS;
1710355810d1SShaohua Li 
1711355810d1SShaohua Li 	return 0;
1712355810d1SShaohua Li }
1713355810d1SShaohua Li 
17149ed988f5SSong Liu static void
17159ed988f5SSong Liu r5l_recovery_create_empty_meta_block(struct r5l_log *log,
17169ed988f5SSong Liu 				     struct page *page,
17179ed988f5SSong Liu 				     sector_t pos, u64 seq)
1718355810d1SShaohua Li {
1719355810d1SShaohua Li 	struct r5l_meta_block *mb;
1720355810d1SShaohua Li 
1721355810d1SShaohua Li 	mb = page_address(page);
17229ed988f5SSong Liu 	clear_page(mb);
1723355810d1SShaohua Li 	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1724355810d1SShaohua Li 	mb->version = R5LOG_VERSION;
1725355810d1SShaohua Li 	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1726355810d1SShaohua Li 	mb->seq = cpu_to_le64(seq);
1727355810d1SShaohua Li 	mb->position = cpu_to_le64(pos);
1728355810d1SShaohua Li }
1729355810d1SShaohua Li 
1730355810d1SShaohua Li static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1731355810d1SShaohua Li 					  u64 seq)
1732355810d1SShaohua Li {
1733355810d1SShaohua Li 	struct page *page;
1734355810d1SShaohua Li 	struct r5l_meta_block *mb;
1735355810d1SShaohua Li 
17369ed988f5SSong Liu 	page = alloc_page(GFP_KERNEL);
1737355810d1SShaohua Li 	if (!page)
1738355810d1SShaohua Li 		return -ENOMEM;
17399ed988f5SSong Liu 	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1740355810d1SShaohua Li 	mb = page_address(page);
17415c88f403SSong Liu 	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
17425c88f403SSong Liu 					     mb, PAGE_SIZE));
1743796a5cf0SMike Christie 	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
174470fd7614SChristoph Hellwig 			  REQ_FUA, false)) {
1745355810d1SShaohua Li 		__free_page(page);
1746355810d1SShaohua Li 		return -EIO;
1747355810d1SShaohua Li 	}
1748355810d1SShaohua Li 	__free_page(page);
1749355810d1SShaohua Li 	return 0;
1750355810d1SShaohua Li }
1751355810d1SShaohua Li 
1752b4c625c6SSong Liu /*
1753b4c625c6SSong Liu  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1754b4c625c6SSong Liu  * to mark valid (potentially not flushed) data in the journal.
1755b4c625c6SSong Liu  *
1756b4c625c6SSong Liu  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1757b4c625c6SSong Liu  * so there should not be any mismatch here.
1758b4c625c6SSong Liu  */
1759b4c625c6SSong Liu static void r5l_recovery_load_data(struct r5l_log *log,
1760b4c625c6SSong Liu 				   struct stripe_head *sh,
1761b4c625c6SSong Liu 				   struct r5l_recovery_ctx *ctx,
1762b4c625c6SSong Liu 				   struct r5l_payload_data_parity *payload,
1763b4c625c6SSong Liu 				   sector_t log_offset)
1764f6bed0efSShaohua Li {
1765b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
1766b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
1767b4c625c6SSong Liu 	int dd_idx;
1768355810d1SShaohua Li 
1769b4c625c6SSong Liu 	raid5_compute_sector(conf,
1770b4c625c6SSong Liu 			     le64_to_cpu(payload->location), 0,
1771b4c625c6SSong Liu 			     &dd_idx, sh);
1772*effe6ee7SSong Liu 	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1773b4c625c6SSong Liu 	sh->dev[dd_idx].log_checksum =
1774b4c625c6SSong Liu 		le32_to_cpu(payload->checksum[0]);
1775b4c625c6SSong Liu 	ctx->meta_total_blocks += BLOCK_SECTORS;
1776b4c625c6SSong Liu 
1777b4c625c6SSong Liu 	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1778b4c625c6SSong Liu 	set_bit(STRIPE_R5C_CACHING, &sh->state);
1779b4c625c6SSong Liu }
1780b4c625c6SSong Liu 
1781b4c625c6SSong Liu static void r5l_recovery_load_parity(struct r5l_log *log,
1782b4c625c6SSong Liu 				     struct stripe_head *sh,
1783b4c625c6SSong Liu 				     struct r5l_recovery_ctx *ctx,
1784b4c625c6SSong Liu 				     struct r5l_payload_data_parity *payload,
1785b4c625c6SSong Liu 				     sector_t log_offset)
1786b4c625c6SSong Liu {
1787b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
1788b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
1789b4c625c6SSong Liu 
1790b4c625c6SSong Liu 	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1791*effe6ee7SSong Liu 	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1792b4c625c6SSong Liu 	sh->dev[sh->pd_idx].log_checksum =
1793b4c625c6SSong Liu 		le32_to_cpu(payload->checksum[0]);
1794b4c625c6SSong Liu 	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1795b4c625c6SSong Liu 
1796b4c625c6SSong Liu 	if (sh->qd_idx >= 0) {
1797*effe6ee7SSong Liu 		r5l_recovery_read_page(
1798*effe6ee7SSong Liu 			log, ctx, sh->dev[sh->qd_idx].page,
1799*effe6ee7SSong Liu 			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1800b4c625c6SSong Liu 		sh->dev[sh->qd_idx].log_checksum =
1801b4c625c6SSong Liu 			le32_to_cpu(payload->checksum[1]);
1802b4c625c6SSong Liu 		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1803b4c625c6SSong Liu 	}
1804b4c625c6SSong Liu 	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1805b4c625c6SSong Liu }
1806b4c625c6SSong Liu 
1807b4c625c6SSong Liu static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1808b4c625c6SSong Liu {
1809b4c625c6SSong Liu 	int i;
1810b4c625c6SSong Liu 
1811b4c625c6SSong Liu 	sh->state = 0;
1812b4c625c6SSong Liu 	sh->log_start = MaxSector;
1813b4c625c6SSong Liu 	for (i = sh->disks; i--; )
1814b4c625c6SSong Liu 		sh->dev[i].flags = 0;
1815b4c625c6SSong Liu }
1816b4c625c6SSong Liu 
1817b4c625c6SSong Liu static void
1818b4c625c6SSong Liu r5l_recovery_replay_one_stripe(struct r5conf *conf,
1819b4c625c6SSong Liu 			       struct stripe_head *sh,
1820b4c625c6SSong Liu 			       struct r5l_recovery_ctx *ctx)
1821b4c625c6SSong Liu {
1822b4c625c6SSong Liu 	struct md_rdev *rdev, *rrdev;
1823b4c625c6SSong Liu 	int disk_index;
1824b4c625c6SSong Liu 	int data_count = 0;
1825b4c625c6SSong Liu 
1826b4c625c6SSong Liu 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1827b4c625c6SSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1828b4c625c6SSong Liu 			continue;
1829b4c625c6SSong Liu 		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1830b4c625c6SSong Liu 			continue;
1831b4c625c6SSong Liu 		data_count++;
1832b4c625c6SSong Liu 	}
1833b4c625c6SSong Liu 
1834b4c625c6SSong Liu 	/*
1835b4c625c6SSong Liu 	 * stripes that only have parity must have been flushed
1836b4c625c6SSong Liu 	 * before the crash that we are now recovering from, so
1837b4c625c6SSong Liu 	 * there is nothing more to recovery.
1838b4c625c6SSong Liu 	 */
1839b4c625c6SSong Liu 	if (data_count == 0)
1840b4c625c6SSong Liu 		goto out;
1841b4c625c6SSong Liu 
1842b4c625c6SSong Liu 	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1843b4c625c6SSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1844b4c625c6SSong Liu 			continue;
1845b4c625c6SSong Liu 
1846b4c625c6SSong Liu 		/* in case device is broken */
1847b4c625c6SSong Liu 		rcu_read_lock();
1848b4c625c6SSong Liu 		rdev = rcu_dereference(conf->disks[disk_index].rdev);
1849b4c625c6SSong Liu 		if (rdev) {
1850b4c625c6SSong Liu 			atomic_inc(&rdev->nr_pending);
1851b4c625c6SSong Liu 			rcu_read_unlock();
1852b4c625c6SSong Liu 			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1853b4c625c6SSong Liu 				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1854b4c625c6SSong Liu 				     false);
1855b4c625c6SSong Liu 			rdev_dec_pending(rdev, rdev->mddev);
1856b4c625c6SSong Liu 			rcu_read_lock();
1857b4c625c6SSong Liu 		}
1858b4c625c6SSong Liu 		rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1859b4c625c6SSong Liu 		if (rrdev) {
1860b4c625c6SSong Liu 			atomic_inc(&rrdev->nr_pending);
1861b4c625c6SSong Liu 			rcu_read_unlock();
1862b4c625c6SSong Liu 			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1863b4c625c6SSong Liu 				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1864b4c625c6SSong Liu 				     false);
1865b4c625c6SSong Liu 			rdev_dec_pending(rrdev, rrdev->mddev);
1866b4c625c6SSong Liu 			rcu_read_lock();
1867b4c625c6SSong Liu 		}
1868b4c625c6SSong Liu 		rcu_read_unlock();
1869b4c625c6SSong Liu 	}
1870b4c625c6SSong Liu 	ctx->data_parity_stripes++;
1871b4c625c6SSong Liu out:
1872b4c625c6SSong Liu 	r5l_recovery_reset_stripe(sh);
1873b4c625c6SSong Liu }
1874b4c625c6SSong Liu 
1875b4c625c6SSong Liu static struct stripe_head *
1876b4c625c6SSong Liu r5c_recovery_alloc_stripe(struct r5conf *conf,
18773c66abbaSSong Liu 			  sector_t stripe_sect)
1878b4c625c6SSong Liu {
1879b4c625c6SSong Liu 	struct stripe_head *sh;
1880b4c625c6SSong Liu 
1881b4c625c6SSong Liu 	sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1882b4c625c6SSong Liu 	if (!sh)
1883b4c625c6SSong Liu 		return NULL;  /* no more stripe available */
1884b4c625c6SSong Liu 
1885b4c625c6SSong Liu 	r5l_recovery_reset_stripe(sh);
1886b4c625c6SSong Liu 
1887b4c625c6SSong Liu 	return sh;
1888b4c625c6SSong Liu }
1889b4c625c6SSong Liu 
1890b4c625c6SSong Liu static struct stripe_head *
1891b4c625c6SSong Liu r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1892b4c625c6SSong Liu {
1893b4c625c6SSong Liu 	struct stripe_head *sh;
1894b4c625c6SSong Liu 
1895b4c625c6SSong Liu 	list_for_each_entry(sh, list, lru)
1896b4c625c6SSong Liu 		if (sh->sector == sect)
1897b4c625c6SSong Liu 			return sh;
1898b4c625c6SSong Liu 	return NULL;
1899b4c625c6SSong Liu }
1900b4c625c6SSong Liu 
1901b4c625c6SSong Liu static void
1902b4c625c6SSong Liu r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1903b4c625c6SSong Liu 			  struct r5l_recovery_ctx *ctx)
1904b4c625c6SSong Liu {
1905b4c625c6SSong Liu 	struct stripe_head *sh, *next;
1906b4c625c6SSong Liu 
1907b4c625c6SSong Liu 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1908b4c625c6SSong Liu 		r5l_recovery_reset_stripe(sh);
1909b4c625c6SSong Liu 		list_del_init(&sh->lru);
1910b4c625c6SSong Liu 		raid5_release_stripe(sh);
1911b4c625c6SSong Liu 	}
1912b4c625c6SSong Liu }
1913b4c625c6SSong Liu 
1914b4c625c6SSong Liu static void
1915b4c625c6SSong Liu r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1916b4c625c6SSong Liu 			    struct r5l_recovery_ctx *ctx)
1917b4c625c6SSong Liu {
1918b4c625c6SSong Liu 	struct stripe_head *sh, *next;
1919b4c625c6SSong Liu 
1920b4c625c6SSong Liu 	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1921b4c625c6SSong Liu 		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1922b4c625c6SSong Liu 			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1923b4c625c6SSong Liu 			list_del_init(&sh->lru);
1924b4c625c6SSong Liu 			raid5_release_stripe(sh);
1925b4c625c6SSong Liu 		}
1926b4c625c6SSong Liu }
1927b4c625c6SSong Liu 
1928b4c625c6SSong Liu /* if matches return 0; otherwise return -EINVAL */
1929b4c625c6SSong Liu static int
1930*effe6ee7SSong Liu r5l_recovery_verify_data_checksum(struct r5l_log *log,
1931*effe6ee7SSong Liu 				  struct r5l_recovery_ctx *ctx,
1932*effe6ee7SSong Liu 				  struct page *page,
1933b4c625c6SSong Liu 				  sector_t log_offset, __le32 log_checksum)
1934b4c625c6SSong Liu {
1935b4c625c6SSong Liu 	void *addr;
1936b4c625c6SSong Liu 	u32 checksum;
1937b4c625c6SSong Liu 
1938*effe6ee7SSong Liu 	r5l_recovery_read_page(log, ctx, page, log_offset);
1939b4c625c6SSong Liu 	addr = kmap_atomic(page);
1940b4c625c6SSong Liu 	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1941b4c625c6SSong Liu 	kunmap_atomic(addr);
1942b4c625c6SSong Liu 	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1943b4c625c6SSong Liu }
1944b4c625c6SSong Liu 
1945b4c625c6SSong Liu /*
1946b4c625c6SSong Liu  * before loading data to stripe cache, we need verify checksum for all data,
1947b4c625c6SSong Liu  * if there is mismatch for any data page, we drop all data in the mata block
1948b4c625c6SSong Liu  */
1949b4c625c6SSong Liu static int
1950b4c625c6SSong Liu r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1951b4c625c6SSong Liu 					 struct r5l_recovery_ctx *ctx)
1952b4c625c6SSong Liu {
1953b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
1954b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
1955b4c625c6SSong Liu 	struct r5l_meta_block *mb = page_address(ctx->meta_page);
1956b4c625c6SSong Liu 	sector_t mb_offset = sizeof(struct r5l_meta_block);
1957b4c625c6SSong Liu 	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1958b4c625c6SSong Liu 	struct page *page;
1959b4c625c6SSong Liu 	struct r5l_payload_data_parity *payload;
1960b4c625c6SSong Liu 
1961b4c625c6SSong Liu 	page = alloc_page(GFP_KERNEL);
1962b4c625c6SSong Liu 	if (!page)
1963355810d1SShaohua Li 		return -ENOMEM;
1964355810d1SShaohua Li 
1965b4c625c6SSong Liu 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
1966b4c625c6SSong Liu 		payload = (void *)mb + mb_offset;
1967b4c625c6SSong Liu 
1968b4c625c6SSong Liu 		if (payload->header.type == R5LOG_PAYLOAD_DATA) {
1969b4c625c6SSong Liu 			if (r5l_recovery_verify_data_checksum(
1970*effe6ee7SSong Liu 				    log, ctx, page, log_offset,
1971b4c625c6SSong Liu 				    payload->checksum[0]) < 0)
1972b4c625c6SSong Liu 				goto mismatch;
1973b4c625c6SSong Liu 		} else if (payload->header.type == R5LOG_PAYLOAD_PARITY) {
1974b4c625c6SSong Liu 			if (r5l_recovery_verify_data_checksum(
1975*effe6ee7SSong Liu 				    log, ctx, page, log_offset,
1976b4c625c6SSong Liu 				    payload->checksum[0]) < 0)
1977b4c625c6SSong Liu 				goto mismatch;
1978b4c625c6SSong Liu 			if (conf->max_degraded == 2 && /* q for RAID 6 */
1979b4c625c6SSong Liu 			    r5l_recovery_verify_data_checksum(
1980*effe6ee7SSong Liu 				    log, ctx, page,
1981b4c625c6SSong Liu 				    r5l_ring_add(log, log_offset,
1982b4c625c6SSong Liu 						 BLOCK_SECTORS),
1983b4c625c6SSong Liu 				    payload->checksum[1]) < 0)
1984b4c625c6SSong Liu 				goto mismatch;
1985b4c625c6SSong Liu 		} else /* not R5LOG_PAYLOAD_DATA or R5LOG_PAYLOAD_PARITY */
1986b4c625c6SSong Liu 			goto mismatch;
1987b4c625c6SSong Liu 
1988b4c625c6SSong Liu 		log_offset = r5l_ring_add(log, log_offset,
1989b4c625c6SSong Liu 					  le32_to_cpu(payload->size));
1990b4c625c6SSong Liu 
1991b4c625c6SSong Liu 		mb_offset += sizeof(struct r5l_payload_data_parity) +
1992b4c625c6SSong Liu 			sizeof(__le32) *
1993b4c625c6SSong Liu 			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
1994b4c625c6SSong Liu 	}
1995b4c625c6SSong Liu 
1996b4c625c6SSong Liu 	put_page(page);
1997b4c625c6SSong Liu 	return 0;
1998b4c625c6SSong Liu 
1999b4c625c6SSong Liu mismatch:
2000b4c625c6SSong Liu 	put_page(page);
2001b4c625c6SSong Liu 	return -EINVAL;
2002b4c625c6SSong Liu }
2003b4c625c6SSong Liu 
2004b4c625c6SSong Liu /*
2005b4c625c6SSong Liu  * Analyze all data/parity pages in one meta block
2006b4c625c6SSong Liu  * Returns:
2007b4c625c6SSong Liu  * 0 for success
2008b4c625c6SSong Liu  * -EINVAL for unknown playload type
2009b4c625c6SSong Liu  * -EAGAIN for checksum mismatch of data page
2010b4c625c6SSong Liu  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2011b4c625c6SSong Liu  */
2012b4c625c6SSong Liu static int
2013b4c625c6SSong Liu r5c_recovery_analyze_meta_block(struct r5l_log *log,
2014b4c625c6SSong Liu 				struct r5l_recovery_ctx *ctx,
2015b4c625c6SSong Liu 				struct list_head *cached_stripe_list)
2016b4c625c6SSong Liu {
2017b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2018b4c625c6SSong Liu 	struct r5conf *conf = mddev->private;
2019b4c625c6SSong Liu 	struct r5l_meta_block *mb;
2020b4c625c6SSong Liu 	struct r5l_payload_data_parity *payload;
2021b4c625c6SSong Liu 	int mb_offset;
2022b4c625c6SSong Liu 	sector_t log_offset;
2023b4c625c6SSong Liu 	sector_t stripe_sect;
2024b4c625c6SSong Liu 	struct stripe_head *sh;
2025b4c625c6SSong Liu 	int ret;
2026b4c625c6SSong Liu 
2027b4c625c6SSong Liu 	/*
2028b4c625c6SSong Liu 	 * for mismatch in data blocks, we will drop all data in this mb, but
2029b4c625c6SSong Liu 	 * we will still read next mb for other data with FLUSH flag, as
2030b4c625c6SSong Liu 	 * io_unit could finish out of order.
2031b4c625c6SSong Liu 	 */
2032b4c625c6SSong Liu 	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2033b4c625c6SSong Liu 	if (ret == -EINVAL)
2034b4c625c6SSong Liu 		return -EAGAIN;
2035b4c625c6SSong Liu 	else if (ret)
2036b4c625c6SSong Liu 		return ret;   /* -ENOMEM duo to alloc_page() failed */
2037b4c625c6SSong Liu 
2038b4c625c6SSong Liu 	mb = page_address(ctx->meta_page);
2039b4c625c6SSong Liu 	mb_offset = sizeof(struct r5l_meta_block);
2040b4c625c6SSong Liu 	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2041b4c625c6SSong Liu 
2042b4c625c6SSong Liu 	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2043b4c625c6SSong Liu 		int dd;
2044b4c625c6SSong Liu 
2045b4c625c6SSong Liu 		payload = (void *)mb + mb_offset;
2046b4c625c6SSong Liu 		stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ?
2047b4c625c6SSong Liu 			raid5_compute_sector(
2048b4c625c6SSong Liu 				conf, le64_to_cpu(payload->location), 0, &dd,
2049b4c625c6SSong Liu 				NULL)
2050b4c625c6SSong Liu 			: le64_to_cpu(payload->location);
2051b4c625c6SSong Liu 
2052b4c625c6SSong Liu 		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2053b4c625c6SSong Liu 						stripe_sect);
2054b4c625c6SSong Liu 
2055b4c625c6SSong Liu 		if (!sh) {
20563c66abbaSSong Liu 			sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
2057b4c625c6SSong Liu 			/*
2058b4c625c6SSong Liu 			 * cannot get stripe from raid5_get_active_stripe
2059b4c625c6SSong Liu 			 * try replay some stripes
2060b4c625c6SSong Liu 			 */
2061b4c625c6SSong Liu 			if (!sh) {
2062b4c625c6SSong Liu 				r5c_recovery_replay_stripes(
2063b4c625c6SSong Liu 					cached_stripe_list, ctx);
2064b4c625c6SSong Liu 				sh = r5c_recovery_alloc_stripe(
20653c66abbaSSong Liu 					conf, stripe_sect);
2066b4c625c6SSong Liu 			}
2067b4c625c6SSong Liu 			if (!sh) {
2068b4c625c6SSong Liu 				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2069b4c625c6SSong Liu 					mdname(mddev),
2070b4c625c6SSong Liu 					conf->min_nr_stripes * 2);
2071b4c625c6SSong Liu 				raid5_set_cache_size(mddev,
2072b4c625c6SSong Liu 						     conf->min_nr_stripes * 2);
20733c66abbaSSong Liu 				sh = r5c_recovery_alloc_stripe(conf,
20743c66abbaSSong Liu 							       stripe_sect);
2075b4c625c6SSong Liu 			}
2076b4c625c6SSong Liu 			if (!sh) {
2077b4c625c6SSong Liu 				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2078b4c625c6SSong Liu 				       mdname(mddev));
2079b4c625c6SSong Liu 				return -ENOMEM;
2080b4c625c6SSong Liu 			}
2081b4c625c6SSong Liu 			list_add_tail(&sh->lru, cached_stripe_list);
2082b4c625c6SSong Liu 		}
2083b4c625c6SSong Liu 
2084b4c625c6SSong Liu 		if (payload->header.type == R5LOG_PAYLOAD_DATA) {
2085f7b7bee7SZhengyuan Liu 			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2086f7b7bee7SZhengyuan Liu 			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2087b4c625c6SSong Liu 				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2088b4c625c6SSong Liu 				list_move_tail(&sh->lru, cached_stripe_list);
2089b4c625c6SSong Liu 			}
2090b4c625c6SSong Liu 			r5l_recovery_load_data(log, sh, ctx, payload,
2091b4c625c6SSong Liu 					       log_offset);
2092b4c625c6SSong Liu 		} else if (payload->header.type == R5LOG_PAYLOAD_PARITY)
2093b4c625c6SSong Liu 			r5l_recovery_load_parity(log, sh, ctx, payload,
2094b4c625c6SSong Liu 						 log_offset);
2095b4c625c6SSong Liu 		else
2096b4c625c6SSong Liu 			return -EINVAL;
2097b4c625c6SSong Liu 
2098b4c625c6SSong Liu 		log_offset = r5l_ring_add(log, log_offset,
2099b4c625c6SSong Liu 					  le32_to_cpu(payload->size));
2100b4c625c6SSong Liu 
2101b4c625c6SSong Liu 		mb_offset += sizeof(struct r5l_payload_data_parity) +
2102b4c625c6SSong Liu 			sizeof(__le32) *
2103b4c625c6SSong Liu 			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2104b4c625c6SSong Liu 	}
2105b4c625c6SSong Liu 
2106b4c625c6SSong Liu 	return 0;
2107b4c625c6SSong Liu }
2108b4c625c6SSong Liu 
2109b4c625c6SSong Liu /*
2110b4c625c6SSong Liu  * Load the stripe into cache. The stripe will be written out later by
2111b4c625c6SSong Liu  * the stripe cache state machine.
2112b4c625c6SSong Liu  */
2113b4c625c6SSong Liu static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2114b4c625c6SSong Liu 					 struct stripe_head *sh)
2115b4c625c6SSong Liu {
2116b4c625c6SSong Liu 	struct r5dev *dev;
2117b4c625c6SSong Liu 	int i;
2118b4c625c6SSong Liu 
2119b4c625c6SSong Liu 	for (i = sh->disks; i--; ) {
2120b4c625c6SSong Liu 		dev = sh->dev + i;
2121b4c625c6SSong Liu 		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2122b4c625c6SSong Liu 			set_bit(R5_InJournal, &dev->flags);
2123b4c625c6SSong Liu 			set_bit(R5_UPTODATE, &dev->flags);
2124b4c625c6SSong Liu 		}
2125b4c625c6SSong Liu 	}
2126b4c625c6SSong Liu }
2127b4c625c6SSong Liu 
2128b4c625c6SSong Liu /*
2129b4c625c6SSong Liu  * Scan through the log for all to-be-flushed data
2130b4c625c6SSong Liu  *
2131b4c625c6SSong Liu  * For stripes with data and parity, namely Data-Parity stripe
2132b4c625c6SSong Liu  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2133b4c625c6SSong Liu  *
2134b4c625c6SSong Liu  * For stripes with only data, namely Data-Only stripe
2135b4c625c6SSong Liu  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2136b4c625c6SSong Liu  *
2137b4c625c6SSong Liu  * For a stripe, if we see data after parity, we should discard all previous
2138b4c625c6SSong Liu  * data and parity for this stripe, as these data are already flushed to
2139b4c625c6SSong Liu  * the array.
2140b4c625c6SSong Liu  *
2141b4c625c6SSong Liu  * At the end of the scan, we return the new journal_tail, which points to
2142b4c625c6SSong Liu  * first data-only stripe on the journal device, or next invalid meta block.
2143b4c625c6SSong Liu  */
2144b4c625c6SSong Liu static int r5c_recovery_flush_log(struct r5l_log *log,
2145b4c625c6SSong Liu 				  struct r5l_recovery_ctx *ctx)
2146b4c625c6SSong Liu {
2147bc8f167fSJackieLiu 	struct stripe_head *sh;
2148b4c625c6SSong Liu 	int ret = 0;
2149b4c625c6SSong Liu 
2150b4c625c6SSong Liu 	/* scan through the log */
2151b4c625c6SSong Liu 	while (1) {
2152b4c625c6SSong Liu 		if (r5l_recovery_read_meta_block(log, ctx))
2153b4c625c6SSong Liu 			break;
2154b4c625c6SSong Liu 
2155b4c625c6SSong Liu 		ret = r5c_recovery_analyze_meta_block(log, ctx,
2156b4c625c6SSong Liu 						      &ctx->cached_list);
2157b4c625c6SSong Liu 		/*
2158b4c625c6SSong Liu 		 * -EAGAIN means mismatch in data block, in this case, we still
2159b4c625c6SSong Liu 		 * try scan the next metablock
2160b4c625c6SSong Liu 		 */
2161b4c625c6SSong Liu 		if (ret && ret != -EAGAIN)
2162b4c625c6SSong Liu 			break;   /* ret == -EINVAL or -ENOMEM */
2163b4c625c6SSong Liu 		ctx->seq++;
2164b4c625c6SSong Liu 		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2165b4c625c6SSong Liu 	}
2166b4c625c6SSong Liu 
2167b4c625c6SSong Liu 	if (ret == -ENOMEM) {
2168b4c625c6SSong Liu 		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2169b4c625c6SSong Liu 		return ret;
2170b4c625c6SSong Liu 	}
2171b4c625c6SSong Liu 
2172b4c625c6SSong Liu 	/* replay data-parity stripes */
2173b4c625c6SSong Liu 	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2174b4c625c6SSong Liu 
2175b4c625c6SSong Liu 	/* load data-only stripes to stripe cache */
2176bc8f167fSJackieLiu 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2177b4c625c6SSong Liu 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2178b4c625c6SSong Liu 		r5c_recovery_load_one_stripe(log, sh);
2179b4c625c6SSong Liu 		ctx->data_only_stripes++;
2180b4c625c6SSong Liu 	}
2181b4c625c6SSong Liu 
2182b4c625c6SSong Liu 	return 0;
2183b4c625c6SSong Liu }
2184355810d1SShaohua Li 
2185355810d1SShaohua Li /*
2186355810d1SShaohua Li  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2187355810d1SShaohua Li  * log will start here. but we can't let superblock point to last valid
2188355810d1SShaohua Li  * meta block. The log might looks like:
2189355810d1SShaohua Li  * | meta 1| meta 2| meta 3|
2190355810d1SShaohua Li  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2191355810d1SShaohua Li  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2192355810d1SShaohua Li  * happens again, new recovery will start from meta 1. Since meta 2n is
2193355810d1SShaohua Li  * valid now, recovery will think meta 3 is valid, which is wrong.
2194355810d1SShaohua Li  * The solution is we create a new meta in meta2 with its seq == meta
21953c6edc66SSong Liu  * 1's seq + 10000 and let superblock points to meta2. The same recovery
21963c6edc66SSong Liu  * will not think meta 3 is a valid meta, because its seq doesn't match
2197355810d1SShaohua Li  */
2198355810d1SShaohua Li 
2199b4c625c6SSong Liu /*
2200b4c625c6SSong Liu  * Before recovery, the log looks like the following
2201b4c625c6SSong Liu  *
2202b4c625c6SSong Liu  *   ---------------------------------------------
2203b4c625c6SSong Liu  *   |           valid log        | invalid log  |
2204b4c625c6SSong Liu  *   ---------------------------------------------
2205b4c625c6SSong Liu  *   ^
2206b4c625c6SSong Liu  *   |- log->last_checkpoint
2207b4c625c6SSong Liu  *   |- log->last_cp_seq
2208b4c625c6SSong Liu  *
2209b4c625c6SSong Liu  * Now we scan through the log until we see invalid entry
2210b4c625c6SSong Liu  *
2211b4c625c6SSong Liu  *   ---------------------------------------------
2212b4c625c6SSong Liu  *   |           valid log        | invalid log  |
2213b4c625c6SSong Liu  *   ---------------------------------------------
2214b4c625c6SSong Liu  *   ^                            ^
2215b4c625c6SSong Liu  *   |- log->last_checkpoint      |- ctx->pos
2216b4c625c6SSong Liu  *   |- log->last_cp_seq          |- ctx->seq
2217b4c625c6SSong Liu  *
2218b4c625c6SSong Liu  * From this point, we need to increase seq number by 10 to avoid
2219b4c625c6SSong Liu  * confusing next recovery.
2220b4c625c6SSong Liu  *
2221b4c625c6SSong Liu  *   ---------------------------------------------
2222b4c625c6SSong Liu  *   |           valid log        | invalid log  |
2223b4c625c6SSong Liu  *   ---------------------------------------------
2224b4c625c6SSong Liu  *   ^                              ^
2225b4c625c6SSong Liu  *   |- log->last_checkpoint        |- ctx->pos+1
22263c6edc66SSong Liu  *   |- log->last_cp_seq            |- ctx->seq+10001
2227b4c625c6SSong Liu  *
2228b4c625c6SSong Liu  * However, it is not safe to start the state machine yet, because data only
2229b4c625c6SSong Liu  * parities are not yet secured in RAID. To save these data only parities, we
2230b4c625c6SSong Liu  * rewrite them from seq+11.
2231b4c625c6SSong Liu  *
2232b4c625c6SSong Liu  *   -----------------------------------------------------------------
2233b4c625c6SSong Liu  *   |           valid log        | data only stripes | invalid log  |
2234b4c625c6SSong Liu  *   -----------------------------------------------------------------
2235b4c625c6SSong Liu  *   ^                                                ^
2236b4c625c6SSong Liu  *   |- log->last_checkpoint                          |- ctx->pos+n
22373c6edc66SSong Liu  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2238b4c625c6SSong Liu  *
2239b4c625c6SSong Liu  * If failure happens again during this process, the recovery can safe start
2240b4c625c6SSong Liu  * again from log->last_checkpoint.
2241b4c625c6SSong Liu  *
2242b4c625c6SSong Liu  * Once data only stripes are rewritten to journal, we move log_tail
2243b4c625c6SSong Liu  *
2244b4c625c6SSong Liu  *   -----------------------------------------------------------------
2245b4c625c6SSong Liu  *   |     old log        |    data only stripes    | invalid log  |
2246b4c625c6SSong Liu  *   -----------------------------------------------------------------
2247b4c625c6SSong Liu  *                        ^                         ^
2248b4c625c6SSong Liu  *                        |- log->last_checkpoint   |- ctx->pos+n
22493c6edc66SSong Liu  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2250b4c625c6SSong Liu  *
2251b4c625c6SSong Liu  * Then we can safely start the state machine. If failure happens from this
2252b4c625c6SSong Liu  * point on, the recovery will start from new log->last_checkpoint.
2253b4c625c6SSong Liu  */
2254b4c625c6SSong Liu static int
2255b4c625c6SSong Liu r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2256b4c625c6SSong Liu 				       struct r5l_recovery_ctx *ctx)
2257b4c625c6SSong Liu {
2258a85dd7b8SSong Liu 	struct stripe_head *sh;
2259b4c625c6SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2260b4c625c6SSong Liu 	struct page *page;
22613c66abbaSSong Liu 	sector_t next_checkpoint = MaxSector;
2262b4c625c6SSong Liu 
2263b4c625c6SSong Liu 	page = alloc_page(GFP_KERNEL);
2264b4c625c6SSong Liu 	if (!page) {
2265b4c625c6SSong Liu 		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2266b4c625c6SSong Liu 		       mdname(mddev));
2267b4c625c6SSong Liu 		return -ENOMEM;
2268b4c625c6SSong Liu 	}
2269b4c625c6SSong Liu 
22703c66abbaSSong Liu 	WARN_ON(list_empty(&ctx->cached_list));
22713c66abbaSSong Liu 
2272a85dd7b8SSong Liu 	list_for_each_entry(sh, &ctx->cached_list, lru) {
2273b4c625c6SSong Liu 		struct r5l_meta_block *mb;
2274b4c625c6SSong Liu 		int i;
2275b4c625c6SSong Liu 		int offset;
2276b4c625c6SSong Liu 		sector_t write_pos;
2277b4c625c6SSong Liu 
2278b4c625c6SSong Liu 		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2279b4c625c6SSong Liu 		r5l_recovery_create_empty_meta_block(log, page,
2280b4c625c6SSong Liu 						     ctx->pos, ctx->seq);
2281b4c625c6SSong Liu 		mb = page_address(page);
2282b4c625c6SSong Liu 		offset = le32_to_cpu(mb->meta_size);
2283fc833c2aSJackieLiu 		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2284b4c625c6SSong Liu 
2285b4c625c6SSong Liu 		for (i = sh->disks; i--; ) {
2286b4c625c6SSong Liu 			struct r5dev *dev = &sh->dev[i];
2287b4c625c6SSong Liu 			struct r5l_payload_data_parity *payload;
2288b4c625c6SSong Liu 			void *addr;
2289b4c625c6SSong Liu 
2290b4c625c6SSong Liu 			if (test_bit(R5_InJournal, &dev->flags)) {
2291b4c625c6SSong Liu 				payload = (void *)mb + offset;
2292b4c625c6SSong Liu 				payload->header.type = cpu_to_le16(
2293b4c625c6SSong Liu 					R5LOG_PAYLOAD_DATA);
2294b4c625c6SSong Liu 				payload->size = BLOCK_SECTORS;
2295b4c625c6SSong Liu 				payload->location = cpu_to_le64(
2296b4c625c6SSong Liu 					raid5_compute_blocknr(sh, i, 0));
2297b4c625c6SSong Liu 				addr = kmap_atomic(dev->page);
2298b4c625c6SSong Liu 				payload->checksum[0] = cpu_to_le32(
2299b4c625c6SSong Liu 					crc32c_le(log->uuid_checksum, addr,
2300b4c625c6SSong Liu 						  PAGE_SIZE));
2301b4c625c6SSong Liu 				kunmap_atomic(addr);
2302b4c625c6SSong Liu 				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2303b4c625c6SSong Liu 					     dev->page, REQ_OP_WRITE, 0, false);
2304b4c625c6SSong Liu 				write_pos = r5l_ring_add(log, write_pos,
2305b4c625c6SSong Liu 							 BLOCK_SECTORS);
2306b4c625c6SSong Liu 				offset += sizeof(__le32) +
2307b4c625c6SSong Liu 					sizeof(struct r5l_payload_data_parity);
2308b4c625c6SSong Liu 
2309b4c625c6SSong Liu 			}
2310b4c625c6SSong Liu 		}
2311b4c625c6SSong Liu 		mb->meta_size = cpu_to_le32(offset);
23125c88f403SSong Liu 		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
23135c88f403SSong Liu 						     mb, PAGE_SIZE));
2314b4c625c6SSong Liu 		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
231520737738SShaohua Li 			     REQ_OP_WRITE, REQ_FUA, false);
2316b4c625c6SSong Liu 		sh->log_start = ctx->pos;
23173c66abbaSSong Liu 		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
23183c66abbaSSong Liu 		atomic_inc(&log->stripe_in_journal_count);
2319b4c625c6SSong Liu 		ctx->pos = write_pos;
2320b4c625c6SSong Liu 		ctx->seq += 1;
23213c66abbaSSong Liu 		next_checkpoint = sh->log_start;
2322b4c625c6SSong Liu 	}
23233c66abbaSSong Liu 	log->next_checkpoint = next_checkpoint;
2324b4c625c6SSong Liu 	__free_page(page);
2325b4c625c6SSong Liu 	return 0;
2326b4c625c6SSong Liu }
2327b4c625c6SSong Liu 
2328a85dd7b8SSong Liu static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2329a85dd7b8SSong Liu 						 struct r5l_recovery_ctx *ctx)
2330a85dd7b8SSong Liu {
2331a85dd7b8SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2332a85dd7b8SSong Liu 	struct r5conf *conf = mddev->private;
2333a85dd7b8SSong Liu 	struct stripe_head *sh, *next;
2334a85dd7b8SSong Liu 
2335a85dd7b8SSong Liu 	if (ctx->data_only_stripes == 0)
2336a85dd7b8SSong Liu 		return;
2337a85dd7b8SSong Liu 
2338a85dd7b8SSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2339a85dd7b8SSong Liu 
2340a85dd7b8SSong Liu 	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2341a85dd7b8SSong Liu 		r5c_make_stripe_write_out(sh);
2342a85dd7b8SSong Liu 		set_bit(STRIPE_HANDLE, &sh->state);
2343a85dd7b8SSong Liu 		list_del_init(&sh->lru);
2344a85dd7b8SSong Liu 		raid5_release_stripe(sh);
2345a85dd7b8SSong Liu 	}
2346a85dd7b8SSong Liu 
2347a85dd7b8SSong Liu 	md_wakeup_thread(conf->mddev->thread);
2348a85dd7b8SSong Liu 	/* reuse conf->wait_for_quiescent in recovery */
2349a85dd7b8SSong Liu 	wait_event(conf->wait_for_quiescent,
2350a85dd7b8SSong Liu 		   atomic_read(&conf->active_stripes) == 0);
2351a85dd7b8SSong Liu 
2352a85dd7b8SSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2353a85dd7b8SSong Liu }
2354a85dd7b8SSong Liu 
2355f6bed0efSShaohua Li static int r5l_recovery_log(struct r5l_log *log)
2356f6bed0efSShaohua Li {
23575aabf7c4SSong Liu 	struct mddev *mddev = log->rdev->mddev;
2358*effe6ee7SSong Liu 	struct r5l_recovery_ctx *ctx;
23595aabf7c4SSong Liu 	int ret;
236043b96748SJackieLiu 	sector_t pos;
2361355810d1SShaohua Li 
2362*effe6ee7SSong Liu 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2363*effe6ee7SSong Liu 	if (!ctx)
2364355810d1SShaohua Li 		return -ENOMEM;
2365355810d1SShaohua Li 
2366*effe6ee7SSong Liu 	ctx->pos = log->last_checkpoint;
2367*effe6ee7SSong Liu 	ctx->seq = log->last_cp_seq;
2368*effe6ee7SSong Liu 	INIT_LIST_HEAD(&ctx->cached_list);
2369*effe6ee7SSong Liu 	ctx->meta_page = alloc_page(GFP_KERNEL);
2370*effe6ee7SSong Liu 
2371*effe6ee7SSong Liu 	if (!ctx->meta_page) {
2372*effe6ee7SSong Liu 		ret =  -ENOMEM;
2373*effe6ee7SSong Liu 		goto meta_page;
2374*effe6ee7SSong Liu 	}
2375*effe6ee7SSong Liu 
2376*effe6ee7SSong Liu 	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2377*effe6ee7SSong Liu 		ret = -ENOMEM;
2378*effe6ee7SSong Liu 		goto ra_pool;
2379*effe6ee7SSong Liu 	}
2380*effe6ee7SSong Liu 
2381*effe6ee7SSong Liu 	ret = r5c_recovery_flush_log(log, ctx);
2382355810d1SShaohua Li 
2383355810d1SShaohua Li 	if (ret)
2384*effe6ee7SSong Liu 		goto error;
23855aabf7c4SSong Liu 
2386*effe6ee7SSong Liu 	pos = ctx->pos;
2387*effe6ee7SSong Liu 	ctx->seq += 10000;
238843b96748SJackieLiu 
2389*effe6ee7SSong Liu 	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
23905aabf7c4SSong Liu 		pr_debug("md/raid:%s: starting from clean shutdown\n",
23915aabf7c4SSong Liu 			 mdname(mddev));
2392a85dd7b8SSong Liu 	else
239399f17890SColin Ian King 		pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2394*effe6ee7SSong Liu 			 mdname(mddev), ctx->data_only_stripes,
2395*effe6ee7SSong Liu 			 ctx->data_parity_stripes);
23965aabf7c4SSong Liu 
2397*effe6ee7SSong Liu 	if (ctx->data_only_stripes == 0) {
2398*effe6ee7SSong Liu 		log->next_checkpoint = ctx->pos;
2399*effe6ee7SSong Liu 		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2400*effe6ee7SSong Liu 		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2401*effe6ee7SSong Liu 	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
24025aabf7c4SSong Liu 		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
24035aabf7c4SSong Liu 		       mdname(mddev));
2404*effe6ee7SSong Liu 		ret =  -EIO;
2405*effe6ee7SSong Liu 		goto error;
24065aabf7c4SSong Liu 	}
24075aabf7c4SSong Liu 
2408*effe6ee7SSong Liu 	log->log_start = ctx->pos;
2409*effe6ee7SSong Liu 	log->seq = ctx->seq;
241043b96748SJackieLiu 	log->last_checkpoint = pos;
241143b96748SJackieLiu 	r5l_write_super(log, pos);
2412a85dd7b8SSong Liu 
2413*effe6ee7SSong Liu 	r5c_recovery_flush_data_only_stripes(log, ctx);
2414*effe6ee7SSong Liu 	ret = 0;
2415*effe6ee7SSong Liu error:
2416*effe6ee7SSong Liu 	r5l_recovery_free_ra_pool(log, ctx);
2417*effe6ee7SSong Liu ra_pool:
2418*effe6ee7SSong Liu 	__free_page(ctx->meta_page);
2419*effe6ee7SSong Liu meta_page:
2420*effe6ee7SSong Liu 	kfree(ctx);
2421*effe6ee7SSong Liu 	return ret;
2422f6bed0efSShaohua Li }
2423f6bed0efSShaohua Li 
2424f6bed0efSShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp)
2425f6bed0efSShaohua Li {
2426f6bed0efSShaohua Li 	struct mddev *mddev = log->rdev->mddev;
2427f6bed0efSShaohua Li 
2428f6bed0efSShaohua Li 	log->rdev->journal_tail = cp;
24292953079cSShaohua Li 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2430f6bed0efSShaohua Li }
2431f6bed0efSShaohua Li 
24322c7da14bSSong Liu static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
24332c7da14bSSong Liu {
24342c7da14bSSong Liu 	struct r5conf *conf = mddev->private;
24352c7da14bSSong Liu 	int ret;
24362c7da14bSSong Liu 
24372c7da14bSSong Liu 	if (!conf->log)
24382c7da14bSSong Liu 		return 0;
24392c7da14bSSong Liu 
24402c7da14bSSong Liu 	switch (conf->log->r5c_journal_mode) {
24412c7da14bSSong Liu 	case R5C_JOURNAL_MODE_WRITE_THROUGH:
24422c7da14bSSong Liu 		ret = snprintf(
24432c7da14bSSong Liu 			page, PAGE_SIZE, "[%s] %s\n",
24442c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
24452c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
24462c7da14bSSong Liu 		break;
24472c7da14bSSong Liu 	case R5C_JOURNAL_MODE_WRITE_BACK:
24482c7da14bSSong Liu 		ret = snprintf(
24492c7da14bSSong Liu 			page, PAGE_SIZE, "%s [%s]\n",
24502c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
24512c7da14bSSong Liu 			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
24522c7da14bSSong Liu 		break;
24532c7da14bSSong Liu 	default:
24542c7da14bSSong Liu 		ret = 0;
24552c7da14bSSong Liu 	}
24562c7da14bSSong Liu 	return ret;
24572c7da14bSSong Liu }
24582c7da14bSSong Liu 
24592c7da14bSSong Liu static ssize_t r5c_journal_mode_store(struct mddev *mddev,
24602c7da14bSSong Liu 				      const char *page, size_t length)
24612c7da14bSSong Liu {
24622c7da14bSSong Liu 	struct r5conf *conf = mddev->private;
24632c7da14bSSong Liu 	struct r5l_log *log = conf->log;
24642c7da14bSSong Liu 	int val = -1, i;
24652c7da14bSSong Liu 	int len = length;
24662c7da14bSSong Liu 
24672c7da14bSSong Liu 	if (!log)
24682c7da14bSSong Liu 		return -ENODEV;
24692c7da14bSSong Liu 
24702c7da14bSSong Liu 	if (len && page[len - 1] == '\n')
24712c7da14bSSong Liu 		len -= 1;
24722c7da14bSSong Liu 	for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++)
24732c7da14bSSong Liu 		if (strlen(r5c_journal_mode_str[i]) == len &&
24742c7da14bSSong Liu 		    strncmp(page, r5c_journal_mode_str[i], len) == 0) {
24752c7da14bSSong Liu 			val = i;
24762c7da14bSSong Liu 			break;
24772c7da14bSSong Liu 		}
24782c7da14bSSong Liu 	if (val < R5C_JOURNAL_MODE_WRITE_THROUGH ||
24792c7da14bSSong Liu 	    val > R5C_JOURNAL_MODE_WRITE_BACK)
24802c7da14bSSong Liu 		return -EINVAL;
24812c7da14bSSong Liu 
24822e38a37fSSong Liu 	if (raid5_calc_degraded(conf) > 0 &&
24832e38a37fSSong Liu 	    val == R5C_JOURNAL_MODE_WRITE_BACK)
24842e38a37fSSong Liu 		return -EINVAL;
24852e38a37fSSong Liu 
24862c7da14bSSong Liu 	mddev_suspend(mddev);
24872c7da14bSSong Liu 	conf->log->r5c_journal_mode = val;
24882c7da14bSSong Liu 	mddev_resume(mddev);
24892c7da14bSSong Liu 
24902c7da14bSSong Liu 	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
24912c7da14bSSong Liu 		 mdname(mddev), val, r5c_journal_mode_str[val]);
24922c7da14bSSong Liu 	return length;
24932c7da14bSSong Liu }
24942c7da14bSSong Liu 
24952c7da14bSSong Liu struct md_sysfs_entry
24962c7da14bSSong Liu r5c_journal_mode = __ATTR(journal_mode, 0644,
24972c7da14bSSong Liu 			  r5c_journal_mode_show, r5c_journal_mode_store);
24982c7da14bSSong Liu 
24992ded3703SSong Liu /*
25002ded3703SSong Liu  * Try handle write operation in caching phase. This function should only
25012ded3703SSong Liu  * be called in write-back mode.
25022ded3703SSong Liu  *
25032ded3703SSong Liu  * If all outstanding writes can be handled in caching phase, returns 0
25042ded3703SSong Liu  * If writes requires write-out phase, call r5c_make_stripe_write_out()
25052ded3703SSong Liu  * and returns -EAGAIN
25062ded3703SSong Liu  */
25072ded3703SSong Liu int r5c_try_caching_write(struct r5conf *conf,
25082ded3703SSong Liu 			  struct stripe_head *sh,
25092ded3703SSong Liu 			  struct stripe_head_state *s,
25102ded3703SSong Liu 			  int disks)
25112ded3703SSong Liu {
25122ded3703SSong Liu 	struct r5l_log *log = conf->log;
25131e6d690bSSong Liu 	int i;
25141e6d690bSSong Liu 	struct r5dev *dev;
25151e6d690bSSong Liu 	int to_cache = 0;
251603b047f4SSong Liu 	void **pslot;
251703b047f4SSong Liu 	sector_t tree_index;
251803b047f4SSong Liu 	int ret;
251903b047f4SSong Liu 	uintptr_t refcount;
25202ded3703SSong Liu 
25212ded3703SSong Liu 	BUG_ON(!r5c_is_writeback(log));
25222ded3703SSong Liu 
25231e6d690bSSong Liu 	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
25241e6d690bSSong Liu 		/*
25251e6d690bSSong Liu 		 * There are two different scenarios here:
25261e6d690bSSong Liu 		 *  1. The stripe has some data cached, and it is sent to
25271e6d690bSSong Liu 		 *     write-out phase for reclaim
25281e6d690bSSong Liu 		 *  2. The stripe is clean, and this is the first write
25291e6d690bSSong Liu 		 *
25301e6d690bSSong Liu 		 * For 1, return -EAGAIN, so we continue with
25311e6d690bSSong Liu 		 * handle_stripe_dirtying().
25321e6d690bSSong Liu 		 *
25331e6d690bSSong Liu 		 * For 2, set STRIPE_R5C_CACHING and continue with caching
25341e6d690bSSong Liu 		 * write.
25351e6d690bSSong Liu 		 */
25361e6d690bSSong Liu 
25371e6d690bSSong Liu 		/* case 1: anything injournal or anything in written */
25381e6d690bSSong Liu 		if (s->injournal > 0 || s->written > 0)
25391e6d690bSSong Liu 			return -EAGAIN;
25401e6d690bSSong Liu 		/* case 2 */
25411e6d690bSSong Liu 		set_bit(STRIPE_R5C_CACHING, &sh->state);
25421e6d690bSSong Liu 	}
25431e6d690bSSong Liu 
25442e38a37fSSong Liu 	/*
25452e38a37fSSong Liu 	 * When run in degraded mode, array is set to write-through mode.
25462e38a37fSSong Liu 	 * This check helps drain pending write safely in the transition to
25472e38a37fSSong Liu 	 * write-through mode.
25482e38a37fSSong Liu 	 */
25492e38a37fSSong Liu 	if (s->failed) {
25502e38a37fSSong Liu 		r5c_make_stripe_write_out(sh);
25512e38a37fSSong Liu 		return -EAGAIN;
25522e38a37fSSong Liu 	}
25532e38a37fSSong Liu 
25541e6d690bSSong Liu 	for (i = disks; i--; ) {
25551e6d690bSSong Liu 		dev = &sh->dev[i];
25561e6d690bSSong Liu 		/* if non-overwrite, use writing-out phase */
25571e6d690bSSong Liu 		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
25581e6d690bSSong Liu 		    !test_bit(R5_InJournal, &dev->flags)) {
25592ded3703SSong Liu 			r5c_make_stripe_write_out(sh);
25602ded3703SSong Liu 			return -EAGAIN;
25612ded3703SSong Liu 		}
25621e6d690bSSong Liu 	}
25631e6d690bSSong Liu 
256403b047f4SSong Liu 	/* if the stripe is not counted in big_stripe_tree, add it now */
256503b047f4SSong Liu 	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
256603b047f4SSong Liu 	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
256703b047f4SSong Liu 		tree_index = r5c_tree_index(conf, sh->sector);
256803b047f4SSong Liu 		spin_lock(&log->tree_lock);
256903b047f4SSong Liu 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
257003b047f4SSong Liu 					       tree_index);
257103b047f4SSong Liu 		if (pslot) {
257203b047f4SSong Liu 			refcount = (uintptr_t)radix_tree_deref_slot_protected(
257303b047f4SSong Liu 				pslot, &log->tree_lock) >>
257403b047f4SSong Liu 				R5C_RADIX_COUNT_SHIFT;
257503b047f4SSong Liu 			radix_tree_replace_slot(
257603b047f4SSong Liu 				&log->big_stripe_tree, pslot,
257703b047f4SSong Liu 				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
257803b047f4SSong Liu 		} else {
257903b047f4SSong Liu 			/*
258003b047f4SSong Liu 			 * this radix_tree_insert can fail safely, so no
258103b047f4SSong Liu 			 * need to call radix_tree_preload()
258203b047f4SSong Liu 			 */
258303b047f4SSong Liu 			ret = radix_tree_insert(
258403b047f4SSong Liu 				&log->big_stripe_tree, tree_index,
258503b047f4SSong Liu 				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
258603b047f4SSong Liu 			if (ret) {
258703b047f4SSong Liu 				spin_unlock(&log->tree_lock);
258803b047f4SSong Liu 				r5c_make_stripe_write_out(sh);
258903b047f4SSong Liu 				return -EAGAIN;
259003b047f4SSong Liu 			}
259103b047f4SSong Liu 		}
259203b047f4SSong Liu 		spin_unlock(&log->tree_lock);
259303b047f4SSong Liu 
259403b047f4SSong Liu 		/*
259503b047f4SSong Liu 		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
259603b047f4SSong Liu 		 * counted in the radix tree
259703b047f4SSong Liu 		 */
259803b047f4SSong Liu 		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
259903b047f4SSong Liu 		atomic_inc(&conf->r5c_cached_partial_stripes);
260003b047f4SSong Liu 	}
260103b047f4SSong Liu 
26021e6d690bSSong Liu 	for (i = disks; i--; ) {
26031e6d690bSSong Liu 		dev = &sh->dev[i];
26041e6d690bSSong Liu 		if (dev->towrite) {
26051e6d690bSSong Liu 			set_bit(R5_Wantwrite, &dev->flags);
26061e6d690bSSong Liu 			set_bit(R5_Wantdrain, &dev->flags);
26071e6d690bSSong Liu 			set_bit(R5_LOCKED, &dev->flags);
26081e6d690bSSong Liu 			to_cache++;
26091e6d690bSSong Liu 		}
26101e6d690bSSong Liu 	}
26111e6d690bSSong Liu 
26121e6d690bSSong Liu 	if (to_cache) {
26131e6d690bSSong Liu 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
26141e6d690bSSong Liu 		/*
26151e6d690bSSong Liu 		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
26161e6d690bSSong Liu 		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
26171e6d690bSSong Liu 		 * r5c_handle_data_cached()
26181e6d690bSSong Liu 		 */
26191e6d690bSSong Liu 		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
26201e6d690bSSong Liu 	}
26211e6d690bSSong Liu 
26221e6d690bSSong Liu 	return 0;
26231e6d690bSSong Liu }
26241e6d690bSSong Liu 
26251e6d690bSSong Liu /*
26261e6d690bSSong Liu  * free extra pages (orig_page) we allocated for prexor
26271e6d690bSSong Liu  */
26281e6d690bSSong Liu void r5c_release_extra_page(struct stripe_head *sh)
26291e6d690bSSong Liu {
2630d7bd398eSSong Liu 	struct r5conf *conf = sh->raid_conf;
26311e6d690bSSong Liu 	int i;
2632d7bd398eSSong Liu 	bool using_disk_info_extra_page;
2633d7bd398eSSong Liu 
2634d7bd398eSSong Liu 	using_disk_info_extra_page =
2635d7bd398eSSong Liu 		sh->dev[0].orig_page == conf->disks[0].extra_page;
26361e6d690bSSong Liu 
26371e6d690bSSong Liu 	for (i = sh->disks; i--; )
26381e6d690bSSong Liu 		if (sh->dev[i].page != sh->dev[i].orig_page) {
26391e6d690bSSong Liu 			struct page *p = sh->dev[i].orig_page;
26401e6d690bSSong Liu 
26411e6d690bSSong Liu 			sh->dev[i].orig_page = sh->dev[i].page;
264286aa1397SSong Liu 			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
264386aa1397SSong Liu 
2644d7bd398eSSong Liu 			if (!using_disk_info_extra_page)
26451e6d690bSSong Liu 				put_page(p);
26461e6d690bSSong Liu 		}
2647d7bd398eSSong Liu 
2648d7bd398eSSong Liu 	if (using_disk_info_extra_page) {
2649d7bd398eSSong Liu 		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2650d7bd398eSSong Liu 		md_wakeup_thread(conf->mddev->thread);
2651d7bd398eSSong Liu 	}
2652d7bd398eSSong Liu }
2653d7bd398eSSong Liu 
2654d7bd398eSSong Liu void r5c_use_extra_page(struct stripe_head *sh)
2655d7bd398eSSong Liu {
2656d7bd398eSSong Liu 	struct r5conf *conf = sh->raid_conf;
2657d7bd398eSSong Liu 	int i;
2658d7bd398eSSong Liu 	struct r5dev *dev;
2659d7bd398eSSong Liu 
2660d7bd398eSSong Liu 	for (i = sh->disks; i--; ) {
2661d7bd398eSSong Liu 		dev = &sh->dev[i];
2662d7bd398eSSong Liu 		if (dev->orig_page != dev->page)
2663d7bd398eSSong Liu 			put_page(dev->orig_page);
2664d7bd398eSSong Liu 		dev->orig_page = conf->disks[i].extra_page;
2665d7bd398eSSong Liu 	}
26661e6d690bSSong Liu }
26672ded3703SSong Liu 
26682ded3703SSong Liu /*
26692ded3703SSong Liu  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
26702ded3703SSong Liu  * stripe is committed to RAID disks.
26712ded3703SSong Liu  */
26722ded3703SSong Liu void r5c_finish_stripe_write_out(struct r5conf *conf,
26732ded3703SSong Liu 				 struct stripe_head *sh,
26742ded3703SSong Liu 				 struct stripe_head_state *s)
26752ded3703SSong Liu {
267603b047f4SSong Liu 	struct r5l_log *log = conf->log;
26771e6d690bSSong Liu 	int i;
26781e6d690bSSong Liu 	int do_wakeup = 0;
267903b047f4SSong Liu 	sector_t tree_index;
268003b047f4SSong Liu 	void **pslot;
268103b047f4SSong Liu 	uintptr_t refcount;
26821e6d690bSSong Liu 
268303b047f4SSong Liu 	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
26842ded3703SSong Liu 		return;
26852ded3703SSong Liu 
26862ded3703SSong Liu 	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
26872ded3703SSong Liu 	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
26882ded3703SSong Liu 
268903b047f4SSong Liu 	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
26902ded3703SSong Liu 		return;
26911e6d690bSSong Liu 
26921e6d690bSSong Liu 	for (i = sh->disks; i--; ) {
26931e6d690bSSong Liu 		clear_bit(R5_InJournal, &sh->dev[i].flags);
26941e6d690bSSong Liu 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
26951e6d690bSSong Liu 			do_wakeup = 1;
26961e6d690bSSong Liu 	}
26971e6d690bSSong Liu 
26981e6d690bSSong Liu 	/*
26991e6d690bSSong Liu 	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
27001e6d690bSSong Liu 	 * We updated R5_InJournal, so we also update s->injournal.
27011e6d690bSSong Liu 	 */
27021e6d690bSSong Liu 	s->injournal = 0;
27031e6d690bSSong Liu 
27041e6d690bSSong Liu 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
27051e6d690bSSong Liu 		if (atomic_dec_and_test(&conf->pending_full_writes))
27061e6d690bSSong Liu 			md_wakeup_thread(conf->mddev->thread);
27071e6d690bSSong Liu 
27081e6d690bSSong Liu 	if (do_wakeup)
27091e6d690bSSong Liu 		wake_up(&conf->wait_for_overlap);
2710a39f7afdSSong Liu 
271103b047f4SSong Liu 	spin_lock_irq(&log->stripe_in_journal_lock);
2712a39f7afdSSong Liu 	list_del_init(&sh->r5c);
271303b047f4SSong Liu 	spin_unlock_irq(&log->stripe_in_journal_lock);
2714a39f7afdSSong Liu 	sh->log_start = MaxSector;
271503b047f4SSong Liu 
271603b047f4SSong Liu 	atomic_dec(&log->stripe_in_journal_count);
271703b047f4SSong Liu 	r5c_update_log_state(log);
271803b047f4SSong Liu 
271903b047f4SSong Liu 	/* stop counting this stripe in big_stripe_tree */
272003b047f4SSong Liu 	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
272103b047f4SSong Liu 	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
272203b047f4SSong Liu 		tree_index = r5c_tree_index(conf, sh->sector);
272303b047f4SSong Liu 		spin_lock(&log->tree_lock);
272403b047f4SSong Liu 		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
272503b047f4SSong Liu 					       tree_index);
272603b047f4SSong Liu 		BUG_ON(pslot == NULL);
272703b047f4SSong Liu 		refcount = (uintptr_t)radix_tree_deref_slot_protected(
272803b047f4SSong Liu 			pslot, &log->tree_lock) >>
272903b047f4SSong Liu 			R5C_RADIX_COUNT_SHIFT;
273003b047f4SSong Liu 		if (refcount == 1)
273103b047f4SSong Liu 			radix_tree_delete(&log->big_stripe_tree, tree_index);
273203b047f4SSong Liu 		else
273303b047f4SSong Liu 			radix_tree_replace_slot(
273403b047f4SSong Liu 				&log->big_stripe_tree, pslot,
273503b047f4SSong Liu 				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
273603b047f4SSong Liu 		spin_unlock(&log->tree_lock);
273703b047f4SSong Liu 	}
273803b047f4SSong Liu 
273903b047f4SSong Liu 	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
274003b047f4SSong Liu 		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2741e33fbb9cSShaohua Li 		atomic_dec(&conf->r5c_flushing_partial_stripes);
274203b047f4SSong Liu 		atomic_dec(&conf->r5c_cached_partial_stripes);
274303b047f4SSong Liu 	}
274403b047f4SSong Liu 
274503b047f4SSong Liu 	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
274603b047f4SSong Liu 		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2747e33fbb9cSShaohua Li 		atomic_dec(&conf->r5c_flushing_full_stripes);
274803b047f4SSong Liu 		atomic_dec(&conf->r5c_cached_full_stripes);
274903b047f4SSong Liu 	}
27501e6d690bSSong Liu }
27511e6d690bSSong Liu 
27521e6d690bSSong Liu int
27531e6d690bSSong Liu r5c_cache_data(struct r5l_log *log, struct stripe_head *sh,
27541e6d690bSSong Liu 	       struct stripe_head_state *s)
27551e6d690bSSong Liu {
2756a39f7afdSSong Liu 	struct r5conf *conf = sh->raid_conf;
27571e6d690bSSong Liu 	int pages = 0;
27581e6d690bSSong Liu 	int reserve;
27591e6d690bSSong Liu 	int i;
27601e6d690bSSong Liu 	int ret = 0;
27611e6d690bSSong Liu 
27621e6d690bSSong Liu 	BUG_ON(!log);
27631e6d690bSSong Liu 
27641e6d690bSSong Liu 	for (i = 0; i < sh->disks; i++) {
27651e6d690bSSong Liu 		void *addr;
27661e6d690bSSong Liu 
27671e6d690bSSong Liu 		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
27681e6d690bSSong Liu 			continue;
27691e6d690bSSong Liu 		addr = kmap_atomic(sh->dev[i].page);
27701e6d690bSSong Liu 		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
27711e6d690bSSong Liu 						    addr, PAGE_SIZE);
27721e6d690bSSong Liu 		kunmap_atomic(addr);
27731e6d690bSSong Liu 		pages++;
27741e6d690bSSong Liu 	}
27751e6d690bSSong Liu 	WARN_ON(pages == 0);
27761e6d690bSSong Liu 
27771e6d690bSSong Liu 	/*
27781e6d690bSSong Liu 	 * The stripe must enter state machine again to call endio, so
27791e6d690bSSong Liu 	 * don't delay.
27801e6d690bSSong Liu 	 */
27811e6d690bSSong Liu 	clear_bit(STRIPE_DELAYED, &sh->state);
27821e6d690bSSong Liu 	atomic_inc(&sh->count);
27831e6d690bSSong Liu 
27841e6d690bSSong Liu 	mutex_lock(&log->io_mutex);
27851e6d690bSSong Liu 	/* meta + data */
27861e6d690bSSong Liu 	reserve = (1 + pages) << (PAGE_SHIFT - 9);
27871e6d690bSSong Liu 
2788a39f7afdSSong Liu 	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2789a39f7afdSSong Liu 	    sh->log_start == MaxSector)
2790a39f7afdSSong Liu 		r5l_add_no_space_stripe(log, sh);
2791a39f7afdSSong Liu 	else if (!r5l_has_free_space(log, reserve)) {
2792a39f7afdSSong Liu 		if (sh->log_start == log->last_checkpoint)
2793a39f7afdSSong Liu 			BUG();
2794a39f7afdSSong Liu 		else
2795a39f7afdSSong Liu 			r5l_add_no_space_stripe(log, sh);
27961e6d690bSSong Liu 	} else {
27971e6d690bSSong Liu 		ret = r5l_log_stripe(log, sh, pages, 0);
27981e6d690bSSong Liu 		if (ret) {
27991e6d690bSSong Liu 			spin_lock_irq(&log->io_list_lock);
28001e6d690bSSong Liu 			list_add_tail(&sh->log_list, &log->no_mem_stripes);
28011e6d690bSSong Liu 			spin_unlock_irq(&log->io_list_lock);
28021e6d690bSSong Liu 		}
28031e6d690bSSong Liu 	}
28041e6d690bSSong Liu 
28051e6d690bSSong Liu 	mutex_unlock(&log->io_mutex);
28061e6d690bSSong Liu 	return 0;
2807f6bed0efSShaohua Li }
2808f6bed0efSShaohua Li 
280903b047f4SSong Liu /* check whether this big stripe is in write back cache. */
281003b047f4SSong Liu bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
281103b047f4SSong Liu {
281203b047f4SSong Liu 	struct r5l_log *log = conf->log;
281303b047f4SSong Liu 	sector_t tree_index;
281403b047f4SSong Liu 	void *slot;
281503b047f4SSong Liu 
281603b047f4SSong Liu 	if (!log)
281703b047f4SSong Liu 		return false;
281803b047f4SSong Liu 
281903b047f4SSong Liu 	WARN_ON_ONCE(!rcu_read_lock_held());
282003b047f4SSong Liu 	tree_index = r5c_tree_index(conf, sect);
282103b047f4SSong Liu 	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
282203b047f4SSong Liu 	return slot != NULL;
282303b047f4SSong Liu }
282403b047f4SSong Liu 
2825f6bed0efSShaohua Li static int r5l_load_log(struct r5l_log *log)
2826f6bed0efSShaohua Li {
2827f6bed0efSShaohua Li 	struct md_rdev *rdev = log->rdev;
2828f6bed0efSShaohua Li 	struct page *page;
2829f6bed0efSShaohua Li 	struct r5l_meta_block *mb;
2830f6bed0efSShaohua Li 	sector_t cp = log->rdev->journal_tail;
2831f6bed0efSShaohua Li 	u32 stored_crc, expected_crc;
2832f6bed0efSShaohua Li 	bool create_super = false;
2833d30dfeb9SJackieLiu 	int ret = 0;
2834f6bed0efSShaohua Li 
2835f6bed0efSShaohua Li 	/* Make sure it's valid */
2836f6bed0efSShaohua Li 	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2837f6bed0efSShaohua Li 		cp = 0;
2838f6bed0efSShaohua Li 	page = alloc_page(GFP_KERNEL);
2839f6bed0efSShaohua Li 	if (!page)
2840f6bed0efSShaohua Li 		return -ENOMEM;
2841f6bed0efSShaohua Li 
2842796a5cf0SMike Christie 	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2843f6bed0efSShaohua Li 		ret = -EIO;
2844f6bed0efSShaohua Li 		goto ioerr;
2845f6bed0efSShaohua Li 	}
2846f6bed0efSShaohua Li 	mb = page_address(page);
2847f6bed0efSShaohua Li 
2848f6bed0efSShaohua Li 	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2849f6bed0efSShaohua Li 	    mb->version != R5LOG_VERSION) {
2850f6bed0efSShaohua Li 		create_super = true;
2851f6bed0efSShaohua Li 		goto create;
2852f6bed0efSShaohua Li 	}
2853f6bed0efSShaohua Li 	stored_crc = le32_to_cpu(mb->checksum);
2854f6bed0efSShaohua Li 	mb->checksum = 0;
28555cb2fbd6SShaohua Li 	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2856f6bed0efSShaohua Li 	if (stored_crc != expected_crc) {
2857f6bed0efSShaohua Li 		create_super = true;
2858f6bed0efSShaohua Li 		goto create;
2859f6bed0efSShaohua Li 	}
2860f6bed0efSShaohua Li 	if (le64_to_cpu(mb->position) != cp) {
2861f6bed0efSShaohua Li 		create_super = true;
2862f6bed0efSShaohua Li 		goto create;
2863f6bed0efSShaohua Li 	}
2864f6bed0efSShaohua Li create:
2865f6bed0efSShaohua Li 	if (create_super) {
2866f6bed0efSShaohua Li 		log->last_cp_seq = prandom_u32();
2867f6bed0efSShaohua Li 		cp = 0;
286856056c2eSZhengyuan Liu 		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2869f6bed0efSShaohua Li 		/*
2870f6bed0efSShaohua Li 		 * Make sure super points to correct address. Log might have
2871f6bed0efSShaohua Li 		 * data very soon. If super hasn't correct log tail address,
2872f6bed0efSShaohua Li 		 * recovery can't find the log
2873f6bed0efSShaohua Li 		 */
2874f6bed0efSShaohua Li 		r5l_write_super(log, cp);
2875f6bed0efSShaohua Li 	} else
2876f6bed0efSShaohua Li 		log->last_cp_seq = le64_to_cpu(mb->seq);
2877f6bed0efSShaohua Li 
2878f6bed0efSShaohua Li 	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
28790576b1c6SShaohua Li 	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
28800576b1c6SShaohua Li 	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
28810576b1c6SShaohua Li 		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2882f6bed0efSShaohua Li 	log->last_checkpoint = cp;
2883f6bed0efSShaohua Li 
2884f6bed0efSShaohua Li 	__free_page(page);
2885f6bed0efSShaohua Li 
2886d30dfeb9SJackieLiu 	if (create_super) {
2887d30dfeb9SJackieLiu 		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
2888d30dfeb9SJackieLiu 		log->seq = log->last_cp_seq + 1;
2889d30dfeb9SJackieLiu 		log->next_checkpoint = cp;
2890d30dfeb9SJackieLiu 	} else
28913d7e7e1dSZhengyuan Liu 		ret = r5l_recovery_log(log);
2892d30dfeb9SJackieLiu 
28933d7e7e1dSZhengyuan Liu 	r5c_update_log_state(log);
28943d7e7e1dSZhengyuan Liu 	return ret;
2895f6bed0efSShaohua Li ioerr:
2896f6bed0efSShaohua Li 	__free_page(page);
2897f6bed0efSShaohua Li 	return ret;
2898f6bed0efSShaohua Li }
2899f6bed0efSShaohua Li 
29002e38a37fSSong Liu void r5c_update_on_rdev_error(struct mddev *mddev)
29012e38a37fSSong Liu {
29022e38a37fSSong Liu 	struct r5conf *conf = mddev->private;
29032e38a37fSSong Liu 	struct r5l_log *log = conf->log;
29042e38a37fSSong Liu 
29052e38a37fSSong Liu 	if (!log)
29062e38a37fSSong Liu 		return;
29072e38a37fSSong Liu 
29082e38a37fSSong Liu 	if (raid5_calc_degraded(conf) > 0 &&
29092e38a37fSSong Liu 	    conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
29102e38a37fSSong Liu 		schedule_work(&log->disable_writeback_work);
29112e38a37fSSong Liu }
29122e38a37fSSong Liu 
2913f6bed0efSShaohua Li int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
2914f6bed0efSShaohua Li {
2915c888a8f9SJens Axboe 	struct request_queue *q = bdev_get_queue(rdev->bdev);
2916f6bed0efSShaohua Li 	struct r5l_log *log;
2917f6bed0efSShaohua Li 
2918f6bed0efSShaohua Li 	if (PAGE_SIZE != 4096)
2919f6bed0efSShaohua Li 		return -EINVAL;
2920c757ec95SSong Liu 
2921c757ec95SSong Liu 	/*
2922c757ec95SSong Liu 	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
2923c757ec95SSong Liu 	 * raid_disks r5l_payload_data_parity.
2924c757ec95SSong Liu 	 *
2925c757ec95SSong Liu 	 * Write journal and cache does not work for very big array
2926c757ec95SSong Liu 	 * (raid_disks > 203)
2927c757ec95SSong Liu 	 */
2928c757ec95SSong Liu 	if (sizeof(struct r5l_meta_block) +
2929c757ec95SSong Liu 	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
2930c757ec95SSong Liu 	     conf->raid_disks) > PAGE_SIZE) {
2931c757ec95SSong Liu 		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
2932c757ec95SSong Liu 		       mdname(conf->mddev), conf->raid_disks);
2933c757ec95SSong Liu 		return -EINVAL;
2934c757ec95SSong Liu 	}
2935c757ec95SSong Liu 
2936f6bed0efSShaohua Li 	log = kzalloc(sizeof(*log), GFP_KERNEL);
2937f6bed0efSShaohua Li 	if (!log)
2938f6bed0efSShaohua Li 		return -ENOMEM;
2939f6bed0efSShaohua Li 	log->rdev = rdev;
2940f6bed0efSShaohua Li 
2941c888a8f9SJens Axboe 	log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
294256fef7c6SChristoph Hellwig 
29435cb2fbd6SShaohua Li 	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
2944f6bed0efSShaohua Li 				       sizeof(rdev->mddev->uuid));
2945f6bed0efSShaohua Li 
2946f6bed0efSShaohua Li 	mutex_init(&log->io_mutex);
2947f6bed0efSShaohua Li 
2948f6bed0efSShaohua Li 	spin_lock_init(&log->io_list_lock);
2949f6bed0efSShaohua Li 	INIT_LIST_HEAD(&log->running_ios);
29500576b1c6SShaohua Li 	INIT_LIST_HEAD(&log->io_end_ios);
2951a8c34f91SShaohua Li 	INIT_LIST_HEAD(&log->flushing_ios);
295204732f74SChristoph Hellwig 	INIT_LIST_HEAD(&log->finished_ios);
29533a83f467SMing Lei 	bio_init(&log->flush_bio, NULL, 0);
2954f6bed0efSShaohua Li 
2955f6bed0efSShaohua Li 	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
2956f6bed0efSShaohua Li 	if (!log->io_kc)
2957f6bed0efSShaohua Li 		goto io_kc;
2958f6bed0efSShaohua Li 
29595036c390SChristoph Hellwig 	log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
29605036c390SChristoph Hellwig 	if (!log->io_pool)
29615036c390SChristoph Hellwig 		goto io_pool;
29625036c390SChristoph Hellwig 
2963c38d29b3SChristoph Hellwig 	log->bs = bioset_create(R5L_POOL_SIZE, 0);
2964c38d29b3SChristoph Hellwig 	if (!log->bs)
2965c38d29b3SChristoph Hellwig 		goto io_bs;
2966c38d29b3SChristoph Hellwig 
2967e8deb638SChristoph Hellwig 	log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
2968e8deb638SChristoph Hellwig 	if (!log->meta_pool)
2969e8deb638SChristoph Hellwig 		goto out_mempool;
2970e8deb638SChristoph Hellwig 
297103b047f4SSong Liu 	spin_lock_init(&log->tree_lock);
297203b047f4SSong Liu 	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
297303b047f4SSong Liu 
29740576b1c6SShaohua Li 	log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
29750576b1c6SShaohua Li 						 log->rdev->mddev, "reclaim");
29760576b1c6SShaohua Li 	if (!log->reclaim_thread)
29770576b1c6SShaohua Li 		goto reclaim_thread;
2978a39f7afdSSong Liu 	log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
2979a39f7afdSSong Liu 
29800fd22b45SShaohua Li 	init_waitqueue_head(&log->iounit_wait);
29810576b1c6SShaohua Li 
29825036c390SChristoph Hellwig 	INIT_LIST_HEAD(&log->no_mem_stripes);
29835036c390SChristoph Hellwig 
2984f6bed0efSShaohua Li 	INIT_LIST_HEAD(&log->no_space_stripes);
2985f6bed0efSShaohua Li 	spin_lock_init(&log->no_space_stripes_lock);
2986f6bed0efSShaohua Li 
29873bddb7f8SSong Liu 	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
29882e38a37fSSong Liu 	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
29893bddb7f8SSong Liu 
29902ded3703SSong Liu 	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2991a39f7afdSSong Liu 	INIT_LIST_HEAD(&log->stripe_in_journal_list);
2992a39f7afdSSong Liu 	spin_lock_init(&log->stripe_in_journal_lock);
2993a39f7afdSSong Liu 	atomic_set(&log->stripe_in_journal_count, 0);
29942ded3703SSong Liu 
2995d2250f10SSong Liu 	rcu_assign_pointer(conf->log, log);
2996d2250f10SSong Liu 
2997f6bed0efSShaohua Li 	if (r5l_load_log(log))
2998f6bed0efSShaohua Li 		goto error;
2999f6bed0efSShaohua Li 
3000a62ab49eSShaohua Li 	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3001f6bed0efSShaohua Li 	return 0;
3002e8deb638SChristoph Hellwig 
3003f6bed0efSShaohua Li error:
3004d2250f10SSong Liu 	rcu_assign_pointer(conf->log, NULL);
30050576b1c6SShaohua Li 	md_unregister_thread(&log->reclaim_thread);
30060576b1c6SShaohua Li reclaim_thread:
3007e8deb638SChristoph Hellwig 	mempool_destroy(log->meta_pool);
3008e8deb638SChristoph Hellwig out_mempool:
3009c38d29b3SChristoph Hellwig 	bioset_free(log->bs);
3010c38d29b3SChristoph Hellwig io_bs:
30115036c390SChristoph Hellwig 	mempool_destroy(log->io_pool);
30125036c390SChristoph Hellwig io_pool:
3013f6bed0efSShaohua Li 	kmem_cache_destroy(log->io_kc);
3014f6bed0efSShaohua Li io_kc:
3015f6bed0efSShaohua Li 	kfree(log);
3016f6bed0efSShaohua Li 	return -EINVAL;
3017f6bed0efSShaohua Li }
3018f6bed0efSShaohua Li 
3019f6bed0efSShaohua Li void r5l_exit_log(struct r5l_log *log)
3020f6bed0efSShaohua Li {
30212e38a37fSSong Liu 	flush_work(&log->disable_writeback_work);
30220576b1c6SShaohua Li 	md_unregister_thread(&log->reclaim_thread);
3023e8deb638SChristoph Hellwig 	mempool_destroy(log->meta_pool);
3024c38d29b3SChristoph Hellwig 	bioset_free(log->bs);
30255036c390SChristoph Hellwig 	mempool_destroy(log->io_pool);
3026f6bed0efSShaohua Li 	kmem_cache_destroy(log->io_kc);
3027f6bed0efSShaohua Li 	kfree(log);
3028f6bed0efSShaohua Li }
3029