drivers/md/bcache/writeback.c

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