2 * Moving/copying garbage collector
4 * Copyright 2012 Google, Inc.
12 #include <trace/events/bcache.h>
17 struct data_insert_op op;
21 static bool moving_pred(struct keybuf *buf, struct bkey *k)
23 struct cache_set *c = container_of(buf, struct cache_set,
27 for (i = 0; i < KEY_PTRS(k); i++) {
28 struct bucket *g = PTR_BUCKET(c, k, i);
37 /* Moving GC - IO loop */
39 static void moving_io_destructor(struct closure *cl)
41 struct moving_io *io = container_of(cl, struct moving_io, cl);
45 static void write_moving_finish(struct closure *cl)
47 struct moving_io *io = container_of(cl, struct moving_io, cl);
48 struct bio *bio = &io->bio.bio;
52 bio_for_each_segment_all(bv, bio, i)
53 __free_page(bv->bv_page);
55 if (io->op.replace_collision)
56 trace_bcache_gc_copy_collision(&io->w->key);
58 bch_keybuf_del(&io->op.c->moving_gc_keys, io->w);
60 up(&io->op.c->moving_in_flight);
62 closure_return_with_destructor(cl, moving_io_destructor);
65 static void read_moving_endio(struct bio *bio, int error)
67 struct bbio *b = container_of(bio, struct bbio, bio);
68 struct moving_io *io = container_of(bio->bi_private,
69 struct moving_io, cl);
73 else if (!KEY_DIRTY(&b->key) &&
74 ptr_stale(io->op.c, &b->key, 0)) {
75 io->op.error = -EINTR;
78 bch_bbio_endio(io->op.c, bio, error, "reading data to move");
81 static void moving_init(struct moving_io *io)
83 struct bio *bio = &io->bio.bio;
87 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
89 bio->bi_size = KEY_SIZE(&io->w->key) << 9;
90 bio->bi_max_vecs = DIV_ROUND_UP(KEY_SIZE(&io->w->key),
92 bio->bi_private = &io->cl;
93 bio->bi_io_vec = bio->bi_inline_vecs;
94 bch_bio_map(bio, NULL);
97 static void write_moving(struct closure *cl)
99 struct moving_io *io = container_of(cl, struct moving_io, cl);
100 struct data_insert_op *op = &io->op;
105 io->bio.bio.bi_sector = KEY_START(&io->w->key);
107 op->bio = &io->bio.bio;
109 op->writeback = KEY_DIRTY(&io->w->key);
110 op->csum = KEY_CSUM(&io->w->key);
112 bkey_copy(&op->replace_key, &io->w->key);
115 closure_call(&op->cl, bch_data_insert, NULL, cl);
118 continue_at(cl, write_moving_finish, system_wq);
121 static void read_moving_submit(struct closure *cl)
123 struct moving_io *io = container_of(cl, struct moving_io, cl);
124 struct bio *bio = &io->bio.bio;
126 bch_submit_bbio(bio, io->op.c, &io->w->key, 0);
128 continue_at(cl, write_moving, system_wq);
131 static void read_moving(struct cache_set *c)
133 struct keybuf_key *w;
134 struct moving_io *io;
138 closure_init_stack(&cl);
140 /* XXX: if we error, background writeback could stall indefinitely */
142 while (!test_bit(CACHE_SET_STOPPING, &c->flags)) {
143 w = bch_keybuf_next_rescan(c, &c->moving_gc_keys,
144 &MAX_KEY, moving_pred);
148 if (ptr_stale(c, &w->key, 0)) {
149 bch_keybuf_del(&c->moving_gc_keys, w);
153 io = kzalloc(sizeof(struct moving_io) + sizeof(struct bio_vec)
154 * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
161 io->op.inode = KEY_INODE(&w->key);
168 bio->bi_end_io = read_moving_endio;
170 if (bio_alloc_pages(bio, GFP_KERNEL))
173 trace_bcache_gc_copy(&w->key);
175 down(&c->moving_in_flight);
176 closure_call(&io->cl, read_moving_submit, NULL, &cl);
180 err: if (!IS_ERR_OR_NULL(w->private))
183 bch_keybuf_del(&c->moving_gc_keys, w);
189 static bool bucket_cmp(struct bucket *l, struct bucket *r)
191 return GC_SECTORS_USED(l) < GC_SECTORS_USED(r);
194 static unsigned bucket_heap_top(struct cache *ca)
197 return (b = heap_peek(&ca->heap)) ? GC_SECTORS_USED(b) : 0;
200 void bch_moving_gc(struct cache_set *c)
206 if (!c->copy_gc_enabled)
209 mutex_lock(&c->bucket_lock);
211 for_each_cache(ca, c, i) {
212 unsigned sectors_to_move = 0;
213 unsigned reserve_sectors = ca->sb.bucket_size *
214 min(fifo_used(&ca->free), ca->free.size / 2);
218 for_each_bucket(b, ca) {
219 if (!GC_SECTORS_USED(b))
222 if (!heap_full(&ca->heap)) {
223 sectors_to_move += GC_SECTORS_USED(b);
224 heap_add(&ca->heap, b, bucket_cmp);
225 } else if (bucket_cmp(b, heap_peek(&ca->heap))) {
226 sectors_to_move -= bucket_heap_top(ca);
227 sectors_to_move += GC_SECTORS_USED(b);
229 ca->heap.data[0] = b;
230 heap_sift(&ca->heap, 0, bucket_cmp);
234 while (sectors_to_move > reserve_sectors) {
235 heap_pop(&ca->heap, b, bucket_cmp);
236 sectors_to_move -= GC_SECTORS_USED(b);
239 while (heap_pop(&ca->heap, b, bucket_cmp))
243 mutex_unlock(&c->bucket_lock);
245 c->moving_gc_keys.last_scanned = ZERO_KEY;
250 void bch_moving_init_cache_set(struct cache_set *c)
252 bch_keybuf_init(&c->moving_gc_keys);
253 sema_init(&c->moving_in_flight, 64);