drivers/md/bcache/writeback.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * background writeback - scan btree for dirty data and write it to the backing
   4  * device
   5  *
   6  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   7  * Copyright 2012 Google, Inc.
   8  */
   9
  10 #include "bcache.h"
  11 #include "btree.h"
  12 #include "debug.h"
  13 #include "writeback.h"
  14
  15 #include <linux/delay.h>
  16 #include <linux/kthread.h>
  17 #include <linux/sched/clock.h>
  18 #include <trace/events/bcache.h>
  19
  20 /* Rate limiting */
  21
  22 static void __update_writeback_rate(struct cached_dev *dc)
  23 {
  24         struct cache_set *c = dc->disk.c;
  25         uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size -
  26                                 bcache_flash_devs_sectors_dirty(c);
  27         uint64_t cache_dirty_target =
  28                 div_u64(cache_sectors * dc->writeback_percent, 100);
  29
  30         int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev),
  31                                    c->cached_dev_sectors);
  32
  33         /* PD controller */
  34
  35         int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
  36         int64_t derivative = dirty - dc->disk.sectors_dirty_last;
  37         int64_t proportional = dirty - target;
  38         int64_t change;
  39
  40         dc->disk.sectors_dirty_last = dirty;
  41
  42         /* Scale to sectors per second */
  43
  44         proportional *= dc->writeback_rate_update_seconds;
  45         proportional = div_s64(proportional, dc->writeback_rate_p_term_inverse);
  46
  47         derivative = div_s64(derivative, dc->writeback_rate_update_seconds);
  48
  49         derivative = ewma_add(dc->disk.sectors_dirty_derivative, derivative,
  50                               (dc->writeback_rate_d_term /
  51                                dc->writeback_rate_update_seconds) ?: 1, 0);
  52
  53         derivative *= dc->writeback_rate_d_term;
  54         derivative = div_s64(derivative, dc->writeback_rate_p_term_inverse);
  55
  56         change = proportional + derivative;
  57
  58         /* Don't increase writeback rate if the device isn't keeping up */
  59         if (change > 0 &&
  60             time_after64(local_clock(),
  61                          dc->writeback_rate.next + NSEC_PER_MSEC))
  62                 change = 0;
  63
  64         dc->writeback_rate.rate =
  65                 clamp_t(int64_t, (int64_t) dc->writeback_rate.rate + change,
  66                         1, NSEC_PER_MSEC);
  67
  68         dc->writeback_rate_proportional = proportional;
  69         dc->writeback_rate_derivative = derivative;
  70         dc->writeback_rate_change = change;
  71         dc->writeback_rate_target = target;
  72 }
  73
  74 static void update_writeback_rate(struct work_struct *work)
  75 {
  76         struct cached_dev *dc = container_of(to_delayed_work(work),
  77                                              struct cached_dev,
  78                                              writeback_rate_update);
  79
  80         down_read(&dc->writeback_lock);
  81
  82         if (atomic_read(&dc->has_dirty) &&
  83             dc->writeback_percent)
  84                 __update_writeback_rate(dc);
  85
  86         up_read(&dc->writeback_lock);
  87
  88         schedule_delayed_work(&dc->writeback_rate_update,
  89                               dc->writeback_rate_update_seconds * HZ);
  90 }
  91
  92 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
  93 {
  94         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
  95             !dc->writeback_percent)
  96                 return 0;
  97
  98         return bch_next_delay(&dc->writeback_rate, sectors);
  99 }
 100
 101 struct dirty_io {
 102         struct closure          cl;
 103         struct cached_dev       *dc;
 104         struct bio              bio;
 105 };
 106
 107 static void dirty_init(struct keybuf_key *w)
 108 {
 109         struct dirty_io *io = w->private;
 110         struct bio *bio = &io->bio;
 111
 112         bio_init(bio, bio->bi_inline_vecs,
 113                  DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
 114         if (!io->dc->writeback_percent)
 115                 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 116
 117         bio->bi_iter.bi_size    = KEY_SIZE(&w->key) << 9;
 118         bio->bi_private         = w;
 119         bch_bio_map(bio, NULL);
 120 }
 121
 122 static void dirty_io_destructor(struct closure *cl)
 123 {
 124         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 125         kfree(io);
 126 }
 127
 128 static void write_dirty_finish(struct closure *cl)
 129 {
 130         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 131         struct keybuf_key *w = io->bio.bi_private;
 132         struct cached_dev *dc = io->dc;
 133
 134         bio_free_pages(&io->bio);
 135
 136         /* This is kind of a dumb way of signalling errors. */
 137         if (KEY_DIRTY(&w->key)) {
 138                 int ret;
 139                 unsigned i;
 140                 struct keylist keys;
 141
 142                 bch_keylist_init(&keys);
 143
 144                 bkey_copy(keys.top, &w->key);
 145                 SET_KEY_DIRTY(keys.top, false);
 146                 bch_keylist_push(&keys);
 147
 148                 for (i = 0; i < KEY_PTRS(&w->key); i++)
 149                         atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 150
 151                 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
 152
 153                 if (ret)
 154                         trace_bcache_writeback_collision(&w->key);
 155
 156                 atomic_long_inc(ret
 157                                 ? &dc->disk.c->writeback_keys_failed
 158                                 : &dc->disk.c->writeback_keys_done);
 159         }
 160
 161         bch_keybuf_del(&dc->writeback_keys, w);
 162         up(&dc->in_flight);
 163
 164         closure_return_with_destructor(cl, dirty_io_destructor);
 165 }
 166
 167 static void dirty_endio(struct bio *bio)
 168 {
 169         struct keybuf_key *w = bio->bi_private;
 170         struct dirty_io *io = w->private;
 171
 172         if (bio->bi_status)
 173                 SET_KEY_DIRTY(&w->key, false);
 174
 175         closure_put(&io->cl);
 176 }
 177
 178 static void write_dirty(struct closure *cl)
 179 {
 180         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 181         struct keybuf_key *w = io->bio.bi_private;
 182
 183         dirty_init(w);
 184         bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
 185         io->bio.bi_iter.bi_sector = KEY_START(&w->key);
 186         bio_set_dev(&io->bio, io->dc->bdev);
 187         io->bio.bi_end_io       = dirty_endio;
 188
 189         closure_bio_submit(&io->bio, cl);
 190
 191         continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
 192 }
 193
 194 static void read_dirty_endio(struct bio *bio)
 195 {
 196         struct keybuf_key *w = bio->bi_private;
 197         struct dirty_io *io = w->private;
 198
 199         bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
 200                             bio->bi_status, "reading dirty data from cache");
 201
 202         dirty_endio(bio);
 203 }
 204
 205 static void read_dirty_submit(struct closure *cl)
 206 {
 207         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 208
 209         closure_bio_submit(&io->bio, cl);
 210
 211         continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 212 }
 213
 214 static void read_dirty(struct cached_dev *dc)
 215 {
 216         unsigned delay = 0;
 217         struct keybuf_key *w;
 218         struct dirty_io *io;
 219         struct closure cl;
 220
 221         closure_init_stack(&cl);
 222
 223         /*
 224          * XXX: if we error, background writeback just spins. Should use some
 225          * mempools.
 226          */
 227
 228         while (!kthread_should_stop()) {
 229
 230                 w = bch_keybuf_next(&dc->writeback_keys);
 231                 if (!w)
 232                         break;
 233
 234                 BUG_ON(ptr_stale(dc->disk.c, &w->key, 0));
 235
 236                 if (KEY_START(&w->key) != dc->last_read ||
 237                     jiffies_to_msecs(delay) > 50)
 238                         while (!kthread_should_stop() && delay)
 239                                 delay = schedule_timeout_interruptible(delay);
 240
 241                 dc->last_read   = KEY_OFFSET(&w->key);
 242
 243                 io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec)
 244                              * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
 245                              GFP_KERNEL);
 246                 if (!io)
 247                         goto err;
 248
 249                 w->private      = io;
 250                 io->dc          = dc;
 251
 252                 dirty_init(w);
 253                 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
 254                 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
 255                 bio_set_dev(&io->bio, PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
 256                 io->bio.bi_end_io       = read_dirty_endio;
 257
 258                 if (bio_alloc_pages(&io->bio, GFP_KERNEL))
 259                         goto err_free;
 260
 261                 trace_bcache_writeback(&w->key);
 262
 263                 down(&dc->in_flight);
 264                 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
 265
 266                 delay = writeback_delay(dc, KEY_SIZE(&w->key));
 267         }
 268
 269         if (0) {
 270 err_free:
 271                 kfree(w->private);
 272 err:
 273                 bch_keybuf_del(&dc->writeback_keys, w);
 274         }
 275
 276         /*
 277          * Wait for outstanding writeback IOs to finish (and keybuf slots to be
 278          * freed) before refilling again
 279          */
 280         closure_sync(&cl);
 281 }
 282
 283 /* Scan for dirty data */
 284
 285 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
 286                                   uint64_t offset, int nr_sectors)
 287 {
 288         struct bcache_device *d = c->devices[inode];
 289         unsigned stripe_offset, stripe, sectors_dirty;
 290
 291         if (!d)
 292                 return;
 293
 294         stripe = offset_to_stripe(d, offset);
 295         stripe_offset = offset & (d->stripe_size - 1);
 296
 297         while (nr_sectors) {
 298                 int s = min_t(unsigned, abs(nr_sectors),
 299                               d->stripe_size - stripe_offset);
 300
 301                 if (nr_sectors < 0)
 302                         s = -s;
 303
 304                 if (stripe >= d->nr_stripes)
 305                         return;
 306
 307                 sectors_dirty = atomic_add_return(s,
 308                                         d->stripe_sectors_dirty + stripe);
 309                 if (sectors_dirty == d->stripe_size)
 310                         set_bit(stripe, d->full_dirty_stripes);
 311                 else
 312                         clear_bit(stripe, d->full_dirty_stripes);
 313
 314                 nr_sectors -= s;
 315                 stripe_offset = 0;
 316                 stripe++;
 317         }
 318 }
 319
 320 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 321 {
 322         struct cached_dev *dc = container_of(buf, struct cached_dev, writeback_keys);
 323
 324         BUG_ON(KEY_INODE(k) != dc->disk.id);
 325
 326         return KEY_DIRTY(k);
 327 }
 328
 329 static void refill_full_stripes(struct cached_dev *dc)
 330 {
 331         struct keybuf *buf = &dc->writeback_keys;
 332         unsigned start_stripe, stripe, next_stripe;
 333         bool wrapped = false;
 334
 335         stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
 336
 337         if (stripe >= dc->disk.nr_stripes)
 338                 stripe = 0;
 339
 340         start_stripe = stripe;
 341
 342         while (1) {
 343                 stripe = find_next_bit(dc->disk.full_dirty_stripes,
 344                                        dc->disk.nr_stripes, stripe);
 345
 346                 if (stripe == dc->disk.nr_stripes)
 347                         goto next;
 348
 349                 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
 350                                                  dc->disk.nr_stripes, stripe);
 351
 352                 buf->last_scanned = KEY(dc->disk.id,
 353                                         stripe * dc->disk.stripe_size, 0);
 354
 355                 bch_refill_keybuf(dc->disk.c, buf,
 356                                   &KEY(dc->disk.id,
 357                                        next_stripe * dc->disk.stripe_size, 0),
 358                                   dirty_pred);
 359
 360                 if (array_freelist_empty(&buf->freelist))
 361                         return;
 362
 363                 stripe = next_stripe;
 364 next:
 365                 if (wrapped && stripe > start_stripe)
 366                         return;
 367
 368                 if (stripe == dc->disk.nr_stripes) {
 369                         stripe = 0;
 370                         wrapped = true;
 371                 }
 372         }
 373 }
 374
 375 /*
 376  * Returns true if we scanned the entire disk
 377  */
 378 static bool refill_dirty(struct cached_dev *dc)
 379 {
 380         struct keybuf *buf = &dc->writeback_keys;
 381         struct bkey start = KEY(dc->disk.id, 0, 0);
 382         struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 383         struct bkey start_pos;
 384
 385         /*
 386          * make sure keybuf pos is inside the range for this disk - at bringup
 387          * we might not be attached yet so this disk's inode nr isn't
 388          * initialized then
 389          */
 390         if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
 391             bkey_cmp(&buf->last_scanned, &end) > 0)
 392                 buf->last_scanned = start;
 393
 394         if (dc->partial_stripes_expensive) {
 395                 refill_full_stripes(dc);
 396                 if (array_freelist_empty(&buf->freelist))
 397                         return false;
 398         }
 399
 400         start_pos = buf->last_scanned;
 401         bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 402
 403         if (bkey_cmp(&buf->last_scanned, &end) < 0)
 404                 return false;
 405
 406         /*
 407          * If we get to the end start scanning again from the beginning, and
 408          * only scan up to where we initially started scanning from:
 409          */
 410         buf->last_scanned = start;
 411         bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
 412
 413         return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 414 }
 415
 416 static int bch_writeback_thread(void *arg)
 417 {
 418         struct cached_dev *dc = arg;
 419         bool searched_full_index;
 420
 421         while (!kthread_should_stop()) {
 422                 down_write(&dc->writeback_lock);
 423                 set_current_state(TASK_INTERRUPTIBLE);
 424                 /*
 425                  * If the bache device is detaching, skip here and continue
 426                  * to perform writeback. Otherwise, if no dirty data on cache,
 427                  * or there is dirty data on cache but writeback is disabled,
 428                  * the writeback thread should sleep here and wait for others
 429                  * to wake up it.
 430                  */
 431                 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
 432                     (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
 433                         up_write(&dc->writeback_lock);
 434
 435                         if (kthread_should_stop()) {
 436                                 set_current_state(TASK_RUNNING);
 437                                 return 0;
 438                         }
 439
 440                         schedule();
 441                         continue;
 442                 }
 443                 set_current_state(TASK_RUNNING);
 444
 445                 searched_full_index = refill_dirty(dc);
 446
 447                 if (searched_full_index &&
 448                     RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
 449                         atomic_set(&dc->has_dirty, 0);
 450                         cached_dev_put(dc);
 451                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 452                         bch_write_bdev_super(dc, NULL);
 453                         /*
 454                          * If bcache device is detaching via sysfs interface,
 455                          * writeback thread should stop after there is no dirty
 456                          * data on cache. BCACHE_DEV_DETACHING flag is set in
 457                          * bch_cached_dev_detach().
 458                          */
 459                         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
 460                                 up_write(&dc->writeback_lock);
 461                                 break;
 462                         }
 463                 }
 464
 465                 up_write(&dc->writeback_lock);
 466
 467                 bch_ratelimit_reset(&dc->writeback_rate);
 468                 read_dirty(dc);
 469
 470                 if (searched_full_index) {
 471                         unsigned delay = dc->writeback_delay * HZ;
 472
 473                         while (delay &&
 474                                !kthread_should_stop() &&
 475                                !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
 476                                 delay = schedule_timeout_interruptible(delay);
 477                 }
 478         }
 479
 480         return 0;
 481 }
 482
 483 /* Init */
 484
 485 struct sectors_dirty_init {
 486         struct btree_op op;
 487         unsigned        inode;
 488 };
 489
 490 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
 491                                  struct bkey *k)
 492 {
 493         struct sectors_dirty_init *op = container_of(_op,
 494                                                 struct sectors_dirty_init, op);
 495         if (KEY_INODE(k) > op->inode)
 496                 return MAP_DONE;
 497
 498         if (KEY_DIRTY(k))
 499                 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
 500                                              KEY_START(k), KEY_SIZE(k));
 501
 502         return MAP_CONTINUE;
 503 }
 504
 505 void bch_sectors_dirty_init(struct bcache_device *d)
 506 {
 507         struct sectors_dirty_init op;
 508
 509         bch_btree_op_init(&op.op, -1);
 510         op.inode = d->id;
 511
 512         bch_btree_map_keys(&op.op, d->c, &KEY(op.inode, 0, 0),
 513                            sectors_dirty_init_fn, 0);
 514
 515         d->sectors_dirty_last = bcache_dev_sectors_dirty(d);
 516 }
 517
 518 void bch_cached_dev_writeback_init(struct cached_dev *dc)
 519 {
 520         sema_init(&dc->in_flight, 64);
 521         init_rwsem(&dc->writeback_lock);
 522         bch_keybuf_init(&dc->writeback_keys);
 523
 524         dc->writeback_metadata          = true;
 525         dc->writeback_running           = true;
 526         dc->writeback_percent           = 10;
 527         dc->writeback_delay             = 30;
 528         dc->writeback_rate.rate         = 1024;
 529
 530         dc->writeback_rate_update_seconds = 5;
 531         dc->writeback_rate_d_term       = 30;
 532         dc->writeback_rate_p_term_inverse = 6000;
 533
 534         INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
 535 }
 536
 537 int bch_cached_dev_writeback_start(struct cached_dev *dc)
 538 {
 539         dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
 540                                                 WQ_MEM_RECLAIM, 0);
 541         if (!dc->writeback_write_wq)
 542                 return -ENOMEM;
 543
 544         dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
 545                                               "bcache_writeback");
 546         if (IS_ERR(dc->writeback_thread))
 547                 return PTR_ERR(dc->writeback_thread);
 548
 549         schedule_delayed_work(&dc->writeback_rate_update,
 550                               dc->writeback_rate_update_seconds * HZ);
 551
 552         bch_writeback_queue(dc);
 553
 554         return 0;
 555 }