1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Moving/copying garbage collector
4 *
5 * Copyright 2012 Google, Inc.
6 */
7
8 #include "bcache.h"
9 #include "btree.h"
10 #include "debug.h"
11 #include "request.h"
12
13 #include <trace/events/bcache.h>
14
15 struct moving_io {
16 struct closure cl;
17 struct keybuf_key *w;
18 struct data_insert_op op;
19 struct bbio bio;
20 };
21
moving_pred(struct keybuf * buf,struct bkey * k)22 static bool moving_pred(struct keybuf *buf, struct bkey *k)
23 {
24 struct cache_set *c = container_of(buf, struct cache_set,
25 moving_gc_keys);
26 unsigned int i;
27
28 for (i = 0; i < KEY_PTRS(k); i++)
29 if (ptr_available(c, k, i) &&
30 GC_MOVE(PTR_BUCKET(c, k, i)))
31 return true;
32
33 return false;
34 }
35
36 /* Moving GC - IO loop */
37
CLOSURE_CALLBACK(moving_io_destructor)38 static CLOSURE_CALLBACK(moving_io_destructor)
39 {
40 closure_type(io, struct moving_io, cl);
41
42 kfree(io);
43 }
44
CLOSURE_CALLBACK(write_moving_finish)45 static CLOSURE_CALLBACK(write_moving_finish)
46 {
47 closure_type(io, struct moving_io, cl);
48 struct bio *bio = &io->bio.bio;
49
50 bio_free_pages(bio);
51
52 if (io->op.replace_collision)
53 trace_bcache_gc_copy_collision(&io->w->key);
54
55 bch_keybuf_del(&io->op.c->moving_gc_keys, io->w);
56
57 up(&io->op.c->moving_in_flight);
58
59 closure_return_with_destructor(cl, moving_io_destructor);
60 }
61
read_moving_endio(struct bio * bio)62 static void read_moving_endio(struct bio *bio)
63 {
64 struct bbio *b = container_of(bio, struct bbio, bio);
65 struct moving_io *io = container_of(bio->bi_private,
66 struct moving_io, cl);
67
68 if (bio->bi_status)
69 io->op.status = bio->bi_status;
70 else if (!KEY_DIRTY(&b->key) &&
71 ptr_stale(io->op.c, &b->key, 0)) {
72 io->op.status = BLK_STS_IOERR;
73 }
74
75 bch_bbio_endio(io->op.c, bio, bio->bi_status, "reading data to move");
76 }
77
moving_init(struct moving_io * io)78 static void moving_init(struct moving_io *io)
79 {
80 struct bio *bio = &io->bio.bio;
81
82 bio_init(bio, NULL, bio->bi_inline_vecs,
83 DIV_ROUND_UP(KEY_SIZE(&io->w->key), PAGE_SECTORS), 0);
84 bio_get(bio);
85 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
86
87 bio->bi_iter.bi_size = KEY_SIZE(&io->w->key) << 9;
88 bio->bi_private = &io->cl;
89 bch_bio_map(bio, NULL);
90 }
91
CLOSURE_CALLBACK(write_moving)92 static CLOSURE_CALLBACK(write_moving)
93 {
94 closure_type(io, struct moving_io, cl);
95 struct data_insert_op *op = &io->op;
96
97 if (!op->status) {
98 moving_init(io);
99
100 io->bio.bio.bi_iter.bi_sector = KEY_START(&io->w->key);
101 op->write_prio = 1;
102 op->bio = &io->bio.bio;
103
104 op->writeback = KEY_DIRTY(&io->w->key);
105 op->csum = KEY_CSUM(&io->w->key);
106
107 bkey_copy(&op->replace_key, &io->w->key);
108 op->replace = true;
109
110 closure_call(&op->cl, bch_data_insert, NULL, cl);
111 }
112
113 continue_at(cl, write_moving_finish, op->wq);
114 }
115
CLOSURE_CALLBACK(read_moving_submit)116 static CLOSURE_CALLBACK(read_moving_submit)
117 {
118 closure_type(io, struct moving_io, cl);
119 struct bio *bio = &io->bio.bio;
120
121 bch_submit_bbio(bio, io->op.c, &io->w->key, 0);
122
123 continue_at(cl, write_moving, io->op.wq);
124 }
125
read_moving(struct cache_set * c)126 static void read_moving(struct cache_set *c)
127 {
128 struct keybuf_key *w;
129 struct moving_io *io;
130 struct bio *bio;
131 struct closure cl;
132
133 closure_init_stack(&cl);
134
135 /* XXX: if we error, background writeback could stall indefinitely */
136
137 while (!test_bit(CACHE_SET_STOPPING, &c->flags)) {
138 w = bch_keybuf_next_rescan(c, &c->moving_gc_keys,
139 &MAX_KEY, moving_pred);
140 if (!w)
141 break;
142
143 if (ptr_stale(c, &w->key, 0)) {
144 bch_keybuf_del(&c->moving_gc_keys, w);
145 continue;
146 }
147
148 io = kzalloc(struct_size(io, bio.bio.bi_inline_vecs,
149 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
150 GFP_KERNEL);
151 if (!io)
152 goto err;
153
154 w->private = io;
155 io->w = w;
156 io->op.inode = KEY_INODE(&w->key);
157 io->op.c = c;
158 io->op.wq = c->moving_gc_wq;
159
160 moving_init(io);
161 bio = &io->bio.bio;
162
163 bio->bi_opf = REQ_OP_READ;
164 bio->bi_end_io = read_moving_endio;
165
166 if (bch_bio_alloc_pages(bio, GFP_KERNEL))
167 goto err;
168
169 trace_bcache_gc_copy(&w->key);
170
171 down(&c->moving_in_flight);
172 closure_call(&io->cl, read_moving_submit, NULL, &cl);
173 }
174
175 if (0) {
176 err: if (!IS_ERR_OR_NULL(w->private))
177 kfree(w->private);
178
179 bch_keybuf_del(&c->moving_gc_keys, w);
180 }
181
182 closure_sync(&cl);
183 }
184
new_bucket_cmp(const void * l,const void * r,void __always_unused * args)185 static bool new_bucket_cmp(const void *l, const void *r, void __always_unused *args)
186 {
187 struct bucket **_l = (struct bucket **)l;
188 struct bucket **_r = (struct bucket **)r;
189
190 return GC_SECTORS_USED(*_l) >= GC_SECTORS_USED(*_r);
191 }
192
new_bucket_swap(void * l,void * r,void __always_unused * args)193 static void new_bucket_swap(void *l, void *r, void __always_unused *args)
194 {
195 struct bucket **_l = l;
196 struct bucket **_r = r;
197
198 swap(*_l, *_r);
199 }
200
bucket_heap_top(struct cache * ca)201 static unsigned int bucket_heap_top(struct cache *ca)
202 {
203 struct bucket *b;
204
205 return (b = min_heap_peek(&ca->heap)[0]) ? GC_SECTORS_USED(b) : 0;
206 }
207
bch_moving_gc(struct cache_set * c)208 void bch_moving_gc(struct cache_set *c)
209 {
210 struct cache *ca = c->cache;
211 struct bucket *b;
212 unsigned long sectors_to_move, reserve_sectors;
213 const struct min_heap_callbacks callbacks = {
214 .less = new_bucket_cmp,
215 .swp = new_bucket_swap,
216 };
217
218 if (!c->copy_gc_enabled)
219 return;
220
221 mutex_lock(&c->bucket_lock);
222
223 sectors_to_move = 0;
224 reserve_sectors = ca->sb.bucket_size *
225 fifo_used(&ca->free[RESERVE_MOVINGGC]);
226
227 ca->heap.nr = 0;
228
229 for_each_bucket(b, ca) {
230 if (GC_MARK(b) == GC_MARK_METADATA ||
231 !GC_SECTORS_USED(b) ||
232 GC_SECTORS_USED(b) == ca->sb.bucket_size ||
233 atomic_read(&b->pin))
234 continue;
235
236 if (!min_heap_full(&ca->heap)) {
237 sectors_to_move += GC_SECTORS_USED(b);
238 min_heap_push(&ca->heap, &b, &callbacks, NULL);
239 } else if (!new_bucket_cmp(&b, min_heap_peek(&ca->heap), ca)) {
240 sectors_to_move -= bucket_heap_top(ca);
241 sectors_to_move += GC_SECTORS_USED(b);
242
243 ca->heap.data[0] = b;
244 min_heap_sift_down(&ca->heap, 0, &callbacks, NULL);
245 }
246 }
247
248 while (sectors_to_move > reserve_sectors) {
249 if (ca->heap.nr) {
250 b = min_heap_peek(&ca->heap)[0];
251 min_heap_pop(&ca->heap, &callbacks, NULL);
252 }
253 sectors_to_move -= GC_SECTORS_USED(b);
254 }
255
256 while (ca->heap.nr) {
257 b = min_heap_peek(&ca->heap)[0];
258 min_heap_pop(&ca->heap, &callbacks, NULL);
259 SET_GC_MOVE(b, 1);
260 }
261
262 mutex_unlock(&c->bucket_lock);
263
264 c->moving_gc_keys.last_scanned = ZERO_KEY;
265
266 read_moving(c);
267 }
268
bch_moving_init_cache_set(struct cache_set * c)269 void bch_moving_init_cache_set(struct cache_set *c)
270 {
271 bch_keybuf_init(&c->moving_gc_keys);
272 sema_init(&c->moving_in_flight, 64);
273 }
274