1 /* 2 * background writeback - scan btree for dirty data and write it to the backing 3 * device 4 * 5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com> 6 * Copyright 2012 Google, Inc. 7 */ 8 9 #include "bcache.h" 10 #include "btree.h" 11 #include "debug.h" 12 #include "writeback.h" 13 14 #include <linux/delay.h> 15 #include <linux/kthread.h> 16 #include <linux/sched/clock.h> 17 #include <trace/events/bcache.h> 18 19 /* Rate limiting */ 20 21 static void __update_writeback_rate(struct cached_dev *dc) 22 { 23 struct cache_set *c = dc->disk.c; 24 uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size - 25 bcache_flash_devs_sectors_dirty(c); 26 uint64_t cache_dirty_target = 27 div_u64(cache_sectors * dc->writeback_percent, 100); 28 29 int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev), 30 c->cached_dev_sectors); 31 32 /* PD controller */ 33 34 int64_t dirty = bcache_dev_sectors_dirty(&dc->disk); 35 int64_t derivative = dirty - dc->disk.sectors_dirty_last; 36 int64_t proportional = dirty - target; 37 int64_t change; 38 39 dc->disk.sectors_dirty_last = dirty; 40 41 /* Scale to sectors per second */ 42 43 proportional *= dc->writeback_rate_update_seconds; 44 proportional = div_s64(proportional, dc->writeback_rate_p_term_inverse); 45 46 derivative = div_s64(derivative, dc->writeback_rate_update_seconds); 47 48 derivative = ewma_add(dc->disk.sectors_dirty_derivative, derivative, 49 (dc->writeback_rate_d_term / 50 dc->writeback_rate_update_seconds) ?: 1, 0); 51 52 derivative *= dc->writeback_rate_d_term; 53 derivative = div_s64(derivative, dc->writeback_rate_p_term_inverse); 54 55 change = proportional + derivative; 56 57 /* Don't increase writeback rate if the device isn't keeping up */ 58 if (change > 0 && 59 time_after64(local_clock(), 60 dc->writeback_rate.next + NSEC_PER_MSEC)) 61 change = 0; 62 63 dc->writeback_rate.rate = 64 clamp_t(int64_t, (int64_t) dc->writeback_rate.rate + change, 65 1, NSEC_PER_MSEC); 66 67 dc->writeback_rate_proportional = proportional; 68 dc->writeback_rate_derivative = derivative; 69 dc->writeback_rate_change = change; 70 dc->writeback_rate_target = target; 71 } 72 73 static void update_writeback_rate(struct work_struct *work) 74 { 75 struct cached_dev *dc = container_of(to_delayed_work(work), 76 struct cached_dev, 77 writeback_rate_update); 78 79 down_read(&dc->writeback_lock); 80 81 if (atomic_read(&dc->has_dirty) && 82 dc->writeback_percent) 83 __update_writeback_rate(dc); 84 85 up_read(&dc->writeback_lock); 86 87 schedule_delayed_work(&dc->writeback_rate_update, 88 dc->writeback_rate_update_seconds * HZ); 89 } 90 91 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors) 92 { 93 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) || 94 !dc->writeback_percent) 95 return 0; 96 97 return bch_next_delay(&dc->writeback_rate, sectors); 98 } 99 100 struct dirty_io { 101 struct closure cl; 102 struct cached_dev *dc; 103 struct bio bio; 104 }; 105 106 static void dirty_init(struct keybuf_key *w) 107 { 108 struct dirty_io *io = w->private; 109 struct bio *bio = &io->bio; 110 111 bio_init(bio, bio->bi_inline_vecs, 112 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)); 113 if (!io->dc->writeback_percent) 114 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0)); 115 116 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9; 117 bio->bi_private = w; 118 bch_bio_map(bio, NULL); 119 } 120 121 static void dirty_io_destructor(struct closure *cl) 122 { 123 struct dirty_io *io = container_of(cl, struct dirty_io, cl); 124 kfree(io); 125 } 126 127 static void write_dirty_finish(struct closure *cl) 128 { 129 struct dirty_io *io = container_of(cl, struct dirty_io, cl); 130 struct keybuf_key *w = io->bio.bi_private; 131 struct cached_dev *dc = io->dc; 132 133 bio_free_pages(&io->bio); 134 135 /* This is kind of a dumb way of signalling errors. */ 136 if (KEY_DIRTY(&w->key)) { 137 int ret; 138 unsigned i; 139 struct keylist keys; 140 141 bch_keylist_init(&keys); 142 143 bkey_copy(keys.top, &w->key); 144 SET_KEY_DIRTY(keys.top, false); 145 bch_keylist_push(&keys); 146 147 for (i = 0; i < KEY_PTRS(&w->key); i++) 148 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin); 149 150 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key); 151 152 if (ret) 153 trace_bcache_writeback_collision(&w->key); 154 155 atomic_long_inc(ret 156 ? &dc->disk.c->writeback_keys_failed 157 : &dc->disk.c->writeback_keys_done); 158 } 159 160 bch_keybuf_del(&dc->writeback_keys, w); 161 up(&dc->in_flight); 162 163 closure_return_with_destructor(cl, dirty_io_destructor); 164 } 165 166 static void dirty_endio(struct bio *bio) 167 { 168 struct keybuf_key *w = bio->bi_private; 169 struct dirty_io *io = w->private; 170 171 if (bio->bi_status) 172 SET_KEY_DIRTY(&w->key, false); 173 174 closure_put(&io->cl); 175 } 176 177 static void write_dirty(struct closure *cl) 178 { 179 struct dirty_io *io = container_of(cl, struct dirty_io, cl); 180 struct keybuf_key *w = io->bio.bi_private; 181 182 dirty_init(w); 183 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0); 184 io->bio.bi_iter.bi_sector = KEY_START(&w->key); 185 bio_set_dev(&io->bio, io->dc->bdev); 186 io->bio.bi_end_io = dirty_endio; 187 188 closure_bio_submit(&io->bio, cl); 189 190 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq); 191 } 192 193 static void read_dirty_endio(struct bio *bio) 194 { 195 struct keybuf_key *w = bio->bi_private; 196 struct dirty_io *io = w->private; 197 198 bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0), 199 bio->bi_status, "reading dirty data from cache"); 200 201 dirty_endio(bio); 202 } 203 204 static void read_dirty_submit(struct closure *cl) 205 { 206 struct dirty_io *io = container_of(cl, struct dirty_io, cl); 207 208 closure_bio_submit(&io->bio, cl); 209 210 continue_at(cl, write_dirty, io->dc->writeback_write_wq); 211 } 212 213 static void read_dirty(struct cached_dev *dc) 214 { 215 unsigned delay = 0; 216 struct keybuf_key *w; 217 struct dirty_io *io; 218 struct closure cl; 219 220 closure_init_stack(&cl); 221 222 /* 223 * XXX: if we error, background writeback just spins. Should use some 224 * mempools. 225 */ 226 227 while (!kthread_should_stop()) { 228 229 w = bch_keybuf_next(&dc->writeback_keys); 230 if (!w) 231 break; 232 233 BUG_ON(ptr_stale(dc->disk.c, &w->key, 0)); 234 235 if (KEY_START(&w->key) != dc->last_read || 236 jiffies_to_msecs(delay) > 50) 237 while (!kthread_should_stop() && delay) 238 delay = schedule_timeout_interruptible(delay); 239 240 dc->last_read = KEY_OFFSET(&w->key); 241 242 io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec) 243 * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS), 244 GFP_KERNEL); 245 if (!io) 246 goto err; 247 248 w->private = io; 249 io->dc = dc; 250 251 dirty_init(w); 252 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0); 253 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0); 254 bio_set_dev(&io->bio, PTR_CACHE(dc->disk.c, &w->key, 0)->bdev); 255 io->bio.bi_end_io = read_dirty_endio; 256 257 if (bio_alloc_pages(&io->bio, GFP_KERNEL)) 258 goto err_free; 259 260 trace_bcache_writeback(&w->key); 261 262 down(&dc->in_flight); 263 closure_call(&io->cl, read_dirty_submit, NULL, &cl); 264 265 delay = writeback_delay(dc, KEY_SIZE(&w->key)); 266 } 267 268 if (0) { 269 err_free: 270 kfree(w->private); 271 err: 272 bch_keybuf_del(&dc->writeback_keys, w); 273 } 274 275 /* 276 * Wait for outstanding writeback IOs to finish (and keybuf slots to be 277 * freed) before refilling again 278 */ 279 closure_sync(&cl); 280 } 281 282 /* Scan for dirty data */ 283 284 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode, 285 uint64_t offset, int nr_sectors) 286 { 287 struct bcache_device *d = c->devices[inode]; 288 unsigned stripe_offset, stripe, sectors_dirty; 289 290 if (!d) 291 return; 292 293 stripe = offset_to_stripe(d, offset); 294 stripe_offset = offset & (d->stripe_size - 1); 295 296 while (nr_sectors) { 297 int s = min_t(unsigned, abs(nr_sectors), 298 d->stripe_size - stripe_offset); 299 300 if (nr_sectors < 0) 301 s = -s; 302 303 if (stripe >= d->nr_stripes) 304 return; 305 306 sectors_dirty = atomic_add_return(s, 307 d->stripe_sectors_dirty + stripe); 308 if (sectors_dirty == d->stripe_size) 309 set_bit(stripe, d->full_dirty_stripes); 310 else 311 clear_bit(stripe, d->full_dirty_stripes); 312 313 nr_sectors -= s; 314 stripe_offset = 0; 315 stripe++; 316 } 317 } 318 319 static bool dirty_pred(struct keybuf *buf, struct bkey *k) 320 { 321 struct cached_dev *dc = container_of(buf, struct cached_dev, writeback_keys); 322 323 BUG_ON(KEY_INODE(k) != dc->disk.id); 324 325 return KEY_DIRTY(k); 326 } 327 328 static void refill_full_stripes(struct cached_dev *dc) 329 { 330 struct keybuf *buf = &dc->writeback_keys; 331 unsigned start_stripe, stripe, next_stripe; 332 bool wrapped = false; 333 334 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned)); 335 336 if (stripe >= dc->disk.nr_stripes) 337 stripe = 0; 338 339 start_stripe = stripe; 340 341 while (1) { 342 stripe = find_next_bit(dc->disk.full_dirty_stripes, 343 dc->disk.nr_stripes, stripe); 344 345 if (stripe == dc->disk.nr_stripes) 346 goto next; 347 348 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes, 349 dc->disk.nr_stripes, stripe); 350 351 buf->last_scanned = KEY(dc->disk.id, 352 stripe * dc->disk.stripe_size, 0); 353 354 bch_refill_keybuf(dc->disk.c, buf, 355 &KEY(dc->disk.id, 356 next_stripe * dc->disk.stripe_size, 0), 357 dirty_pred); 358 359 if (array_freelist_empty(&buf->freelist)) 360 return; 361 362 stripe = next_stripe; 363 next: 364 if (wrapped && stripe > start_stripe) 365 return; 366 367 if (stripe == dc->disk.nr_stripes) { 368 stripe = 0; 369 wrapped = true; 370 } 371 } 372 } 373 374 /* 375 * Returns true if we scanned the entire disk 376 */ 377 static bool refill_dirty(struct cached_dev *dc) 378 { 379 struct keybuf *buf = &dc->writeback_keys; 380 struct bkey start = KEY(dc->disk.id, 0, 0); 381 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0); 382 struct bkey start_pos; 383 384 /* 385 * make sure keybuf pos is inside the range for this disk - at bringup 386 * we might not be attached yet so this disk's inode nr isn't 387 * initialized then 388 */ 389 if (bkey_cmp(&buf->last_scanned, &start) < 0 || 390 bkey_cmp(&buf->last_scanned, &end) > 0) 391 buf->last_scanned = start; 392 393 if (dc->partial_stripes_expensive) { 394 refill_full_stripes(dc); 395 if (array_freelist_empty(&buf->freelist)) 396 return false; 397 } 398 399 start_pos = buf->last_scanned; 400 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred); 401 402 if (bkey_cmp(&buf->last_scanned, &end) < 0) 403 return false; 404 405 /* 406 * If we get to the end start scanning again from the beginning, and 407 * only scan up to where we initially started scanning from: 408 */ 409 buf->last_scanned = start; 410 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred); 411 412 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0; 413 } 414 415 static int bch_writeback_thread(void *arg) 416 { 417 struct cached_dev *dc = arg; 418 bool searched_full_index; 419 420 while (!kthread_should_stop()) { 421 down_write(&dc->writeback_lock); 422 if (!atomic_read(&dc->has_dirty) || 423 (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) && 424 !dc->writeback_running)) { 425 up_write(&dc->writeback_lock); 426 set_current_state(TASK_INTERRUPTIBLE); 427 428 if (kthread_should_stop()) 429 return 0; 430 431 schedule(); 432 continue; 433 } 434 435 searched_full_index = refill_dirty(dc); 436 437 if (searched_full_index && 438 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) { 439 atomic_set(&dc->has_dirty, 0); 440 cached_dev_put(dc); 441 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN); 442 bch_write_bdev_super(dc, NULL); 443 } 444 445 up_write(&dc->writeback_lock); 446 447 bch_ratelimit_reset(&dc->writeback_rate); 448 read_dirty(dc); 449 450 if (searched_full_index) { 451 unsigned delay = dc->writeback_delay * HZ; 452 453 while (delay && 454 !kthread_should_stop() && 455 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) 456 delay = schedule_timeout_interruptible(delay); 457 } 458 } 459 460 return 0; 461 } 462 463 /* Init */ 464 465 struct sectors_dirty_init { 466 struct btree_op op; 467 unsigned inode; 468 }; 469 470 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b, 471 struct bkey *k) 472 { 473 struct sectors_dirty_init *op = container_of(_op, 474 struct sectors_dirty_init, op); 475 if (KEY_INODE(k) > op->inode) 476 return MAP_DONE; 477 478 if (KEY_DIRTY(k)) 479 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 480 KEY_START(k), KEY_SIZE(k)); 481 482 return MAP_CONTINUE; 483 } 484 485 void bch_sectors_dirty_init(struct bcache_device *d) 486 { 487 struct sectors_dirty_init op; 488 489 bch_btree_op_init(&op.op, -1); 490 op.inode = d->id; 491 492 bch_btree_map_keys(&op.op, d->c, &KEY(op.inode, 0, 0), 493 sectors_dirty_init_fn, 0); 494 495 d->sectors_dirty_last = bcache_dev_sectors_dirty(d); 496 } 497 498 void bch_cached_dev_writeback_init(struct cached_dev *dc) 499 { 500 sema_init(&dc->in_flight, 64); 501 init_rwsem(&dc->writeback_lock); 502 bch_keybuf_init(&dc->writeback_keys); 503 504 dc->writeback_metadata = true; 505 dc->writeback_running = true; 506 dc->writeback_percent = 10; 507 dc->writeback_delay = 30; 508 dc->writeback_rate.rate = 1024; 509 510 dc->writeback_rate_update_seconds = 5; 511 dc->writeback_rate_d_term = 30; 512 dc->writeback_rate_p_term_inverse = 6000; 513 514 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate); 515 } 516 517 int bch_cached_dev_writeback_start(struct cached_dev *dc) 518 { 519 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq", 520 WQ_MEM_RECLAIM, 0); 521 if (!dc->writeback_write_wq) 522 return -ENOMEM; 523 524 dc->writeback_thread = kthread_create(bch_writeback_thread, dc, 525 "bcache_writeback"); 526 if (IS_ERR(dc->writeback_thread)) 527 return PTR_ERR(dc->writeback_thread); 528 529 schedule_delayed_work(&dc->writeback_rate_update, 530 dc->writeback_rate_update_seconds * HZ); 531 532 bch_writeback_queue(dc); 533 534 return 0; 535 } 536