bcache: improve multithreaded bch_sectors_dirty_init()
[linux-2.6-microblaze.git] / drivers / md / bcache / writeback.h
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _BCACHE_WRITEBACK_H
3 #define _BCACHE_WRITEBACK_H
4
5 #define CUTOFF_WRITEBACK        40
6 #define CUTOFF_WRITEBACK_SYNC   70
7
8 #define CUTOFF_WRITEBACK_MAX            70
9 #define CUTOFF_WRITEBACK_SYNC_MAX       90
10
11 #define MAX_WRITEBACKS_IN_PASS  5
12 #define MAX_WRITESIZE_IN_PASS   5000    /* *512b */
13
14 #define WRITEBACK_RATE_UPDATE_SECS_MAX          60
15 #define WRITEBACK_RATE_UPDATE_SECS_DEFAULT      5
16
17 #define BCH_AUTO_GC_DIRTY_THRESHOLD     50
18
19 #define BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW 50
20 #define BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID 57
21 #define BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH 64
22
23 #define BCH_DIRTY_INIT_THRD_MAX 12
24 /*
25  * 14 (16384ths) is chosen here as something that each backing device
26  * should be a reasonable fraction of the share, and not to blow up
27  * until individual backing devices are a petabyte.
28  */
29 #define WRITEBACK_SHARE_SHIFT   14
30
31 struct bch_dirty_init_state;
32 struct dirty_init_thrd_info {
33         struct bch_dirty_init_state     *state;
34         struct task_struct              *thread;
35 };
36
37 struct bch_dirty_init_state {
38         struct cache_set                *c;
39         struct bcache_device            *d;
40         int                             total_threads;
41         int                             key_idx;
42         spinlock_t                      idx_lock;
43         atomic_t                        started;
44         atomic_t                        enough;
45         wait_queue_head_t               wait;
46         struct dirty_init_thrd_info     infos[BCH_DIRTY_INIT_THRD_MAX];
47 };
48
49 static inline uint64_t bcache_dev_sectors_dirty(struct bcache_device *d)
50 {
51         uint64_t i, ret = 0;
52
53         for (i = 0; i < d->nr_stripes; i++)
54                 ret += atomic_read(d->stripe_sectors_dirty + i);
55
56         return ret;
57 }
58
59 static inline int offset_to_stripe(struct bcache_device *d,
60                                         uint64_t offset)
61 {
62         do_div(offset, d->stripe_size);
63
64         /* d->nr_stripes is in range [1, INT_MAX] */
65         if (unlikely(offset >= d->nr_stripes)) {
66                 pr_err("Invalid stripe %llu (>= nr_stripes %d).\n",
67                         offset, d->nr_stripes);
68                 return -EINVAL;
69         }
70
71         /*
72          * Here offset is definitly smaller than INT_MAX,
73          * return it as int will never overflow.
74          */
75         return offset;
76 }
77
78 static inline bool bcache_dev_stripe_dirty(struct cached_dev *dc,
79                                            uint64_t offset,
80                                            unsigned int nr_sectors)
81 {
82         int stripe = offset_to_stripe(&dc->disk, offset);
83
84         if (stripe < 0)
85                 return false;
86
87         while (1) {
88                 if (atomic_read(dc->disk.stripe_sectors_dirty + stripe))
89                         return true;
90
91                 if (nr_sectors <= dc->disk.stripe_size)
92                         return false;
93
94                 nr_sectors -= dc->disk.stripe_size;
95                 stripe++;
96         }
97 }
98
99 extern unsigned int bch_cutoff_writeback;
100 extern unsigned int bch_cutoff_writeback_sync;
101
102 static inline bool should_writeback(struct cached_dev *dc, struct bio *bio,
103                                     unsigned int cache_mode, bool would_skip)
104 {
105         unsigned int in_use = dc->disk.c->gc_stats.in_use;
106
107         if (cache_mode != CACHE_MODE_WRITEBACK ||
108             test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
109             in_use > bch_cutoff_writeback_sync)
110                 return false;
111
112         if (bio_op(bio) == REQ_OP_DISCARD)
113                 return false;
114
115         if (dc->partial_stripes_expensive &&
116             bcache_dev_stripe_dirty(dc, bio->bi_iter.bi_sector,
117                                     bio_sectors(bio)))
118                 return true;
119
120         if (would_skip)
121                 return false;
122
123         return (op_is_sync(bio->bi_opf) ||
124                 bio->bi_opf & (REQ_META|REQ_PRIO) ||
125                 in_use <= bch_cutoff_writeback);
126 }
127
128 static inline void bch_writeback_queue(struct cached_dev *dc)
129 {
130         if (!IS_ERR_OR_NULL(dc->writeback_thread))
131                 wake_up_process(dc->writeback_thread);
132 }
133
134 static inline void bch_writeback_add(struct cached_dev *dc)
135 {
136         if (!atomic_read(&dc->has_dirty) &&
137             !atomic_xchg(&dc->has_dirty, 1)) {
138                 if (BDEV_STATE(&dc->sb) != BDEV_STATE_DIRTY) {
139                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_DIRTY);
140                         /* XXX: should do this synchronously */
141                         bch_write_bdev_super(dc, NULL);
142                 }
143
144                 bch_writeback_queue(dc);
145         }
146 }
147
148 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
149                                   uint64_t offset, int nr_sectors);
150
151 void bch_sectors_dirty_init(struct bcache_device *d);
152 void bch_cached_dev_writeback_init(struct cached_dev *dc);
153 int bch_cached_dev_writeback_start(struct cached_dev *dc);
154
155 #endif