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