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
 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_error)
 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         io->bio.bi_bdev         = 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_error, "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                 io->bio.bi_bdev         = PTR_CACHE(dc->disk.c,
 254                                                     &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 cached_dev *dc)
 486 {
 487         struct sectors_dirty_init op;
 488
 489         bch_btree_op_init(&op.op, -1);
 490         op.inode = dc->disk.id;
 491
 492         bch_btree_map_keys(&op.op, dc->disk.c, &KEY(op.inode, 0, 0),
 493                            sectors_dirty_init_fn, 0);
 494
 495         dc->disk.sectors_dirty_last = bcache_dev_sectors_dirty(&dc->disk);
 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_thread = kthread_create(bch_writeback_thread, dc,
 520                                               "bcache_writeback");
 521         if (IS_ERR(dc->writeback_thread))
 522                 return PTR_ERR(dc->writeback_thread);
 523
 524         schedule_delayed_work(&dc->writeback_rate_update,
 525                               dc->writeback_rate_update_seconds * HZ);
 526
 527         bch_writeback_queue(dc);
 528
 529         return 0;
 530 }