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