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 static void update_gc_after_writeback(struct cache_set *c)
  21 {
  22         if (c->gc_after_writeback != (BCH_ENABLE_AUTO_GC) ||
  23             c->gc_stats.in_use < BCH_AUTO_GC_DIRTY_THRESHOLD)
  24                 return;
  25
  26         c->gc_after_writeback |= BCH_DO_AUTO_GC;
  27 }
  28
  29 /* Rate limiting */
  30 static uint64_t __calc_target_rate(struct cached_dev *dc)
  31 {
  32         struct cache_set *c = dc->disk.c;
  33
  34         /*
  35          * This is the size of the cache, minus the amount used for
  36          * flash-only devices
  37          */
  38         uint64_t cache_sectors = c->nbuckets * c->cache->sb.bucket_size -
  39                                 atomic_long_read(&c->flash_dev_dirty_sectors);
  40
  41         /*
  42          * Unfortunately there is no control of global dirty data.  If the
  43          * user states that they want 10% dirty data in the cache, and has,
  44          * e.g., 5 backing volumes of equal size, we try and ensure each
  45          * backing volume uses about 2% of the cache for dirty data.
  46          */
  47         uint32_t bdev_share =
  48                 div64_u64(bdev_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
  49                                 c->cached_dev_sectors);
  50
  51         uint64_t cache_dirty_target =
  52                 div_u64(cache_sectors * dc->writeback_percent, 100);
  53
  54         /* Ensure each backing dev gets at least one dirty share */
  55         if (bdev_share < 1)
  56                 bdev_share = 1;
  57
  58         return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
  59 }
  60
  61 static void __update_writeback_rate(struct cached_dev *dc)
  62 {
  63         /*
  64          * PI controller:
  65          * Figures out the amount that should be written per second.
  66          *
  67          * First, the error (number of sectors that are dirty beyond our
  68          * target) is calculated.  The error is accumulated (numerically
  69          * integrated).
  70          *
  71          * Then, the proportional value and integral value are scaled
  72          * based on configured values.  These are stored as inverses to
  73          * avoid fixed point math and to make configuration easy-- e.g.
  74          * the default value of 40 for writeback_rate_p_term_inverse
  75          * attempts to write at a rate that would retire all the dirty
  76          * blocks in 40 seconds.
  77          *
  78          * The writeback_rate_i_inverse value of 10000 means that 1/10000th
  79          * of the error is accumulated in the integral term per second.
  80          * This acts as a slow, long-term average that is not subject to
  81          * variations in usage like the p term.
  82          */
  83         int64_t target = __calc_target_rate(dc);
  84         int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
  85         int64_t error = dirty - target;
  86         int64_t proportional_scaled =
  87                 div_s64(error, dc->writeback_rate_p_term_inverse);
  88         int64_t integral_scaled;
  89         uint32_t new_rate;
  90
  91         if ((error < 0 && dc->writeback_rate_integral > 0) ||
  92             (error > 0 && time_before64(local_clock(),
  93                          dc->writeback_rate.next + NSEC_PER_MSEC))) {
  94                 /*
  95                  * Only decrease the integral term if it's more than
  96                  * zero.  Only increase the integral term if the device
  97                  * is keeping up.  (Don't wind up the integral
  98                  * ineffectively in either case).
  99                  *
 100                  * It's necessary to scale this by
 101                  * writeback_rate_update_seconds to keep the integral
 102                  * term dimensioned properly.
 103                  */
 104                 dc->writeback_rate_integral += error *
 105                         dc->writeback_rate_update_seconds;
 106         }
 107
 108         integral_scaled = div_s64(dc->writeback_rate_integral,
 109                         dc->writeback_rate_i_term_inverse);
 110
 111         new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
 112                         dc->writeback_rate_minimum, NSEC_PER_SEC);
 113
 114         dc->writeback_rate_proportional = proportional_scaled;
 115         dc->writeback_rate_integral_scaled = integral_scaled;
 116         dc->writeback_rate_change = new_rate -
 117                         atomic_long_read(&dc->writeback_rate.rate);
 118         atomic_long_set(&dc->writeback_rate.rate, new_rate);
 119         dc->writeback_rate_target = target;
 120 }
 121
 122 static bool idle_counter_exceeded(struct cache_set *c)
 123 {
 124         int counter, dev_nr;
 125
 126         /*
 127          * If c->idle_counter is overflow (idel for really long time),
 128          * reset as 0 and not set maximum rate this time for code
 129          * simplicity.
 130          */
 131         counter = atomic_inc_return(&c->idle_counter);
 132         if (counter <= 0) {
 133                 atomic_set(&c->idle_counter, 0);
 134                 return false;
 135         }
 136
 137         dev_nr = atomic_read(&c->attached_dev_nr);
 138         if (dev_nr == 0)
 139                 return false;
 140
 141         /*
 142          * c->idle_counter is increased by writeback thread of all
 143          * attached backing devices, in order to represent a rough
 144          * time period, counter should be divided by dev_nr.
 145          * Otherwise the idle time cannot be larger with more backing
 146          * device attached.
 147          * The following calculation equals to checking
 148          *      (counter / dev_nr) < (dev_nr * 6)
 149          */
 150         if (counter < (dev_nr * dev_nr * 6))
 151                 return false;
 152
 153         return true;
 154 }
 155
 156 /*
 157  * Idle_counter is increased every time when update_writeback_rate() is
 158  * called. If all backing devices attached to the same cache set have
 159  * identical dc->writeback_rate_update_seconds values, it is about 6
 160  * rounds of update_writeback_rate() on each backing device before
 161  * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
 162  * to each dc->writeback_rate.rate.
 163  * In order to avoid extra locking cost for counting exact dirty cached
 164  * devices number, c->attached_dev_nr is used to calculate the idle
 165  * throushold. It might be bigger if not all cached device are in write-
 166  * back mode, but it still works well with limited extra rounds of
 167  * update_writeback_rate().
 168  */
 169 static bool set_at_max_writeback_rate(struct cache_set *c,
 170                                        struct cached_dev *dc)
 171 {
 172         /* Don't sst max writeback rate if it is disabled */
 173         if (!c->idle_max_writeback_rate_enabled)
 174                 return false;
 175
 176         /* Don't set max writeback rate if gc is running */
 177         if (!c->gc_mark_valid)
 178                 return false;
 179
 180         if (!idle_counter_exceeded(c))
 181                 return false;
 182
 183         if (atomic_read(&c->at_max_writeback_rate) != 1)
 184                 atomic_set(&c->at_max_writeback_rate, 1);
 185
 186         atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
 187
 188         /* keep writeback_rate_target as existing value */
 189         dc->writeback_rate_proportional = 0;
 190         dc->writeback_rate_integral_scaled = 0;
 191         dc->writeback_rate_change = 0;
 192
 193         /*
 194          * In case new I/O arrives during before
 195          * set_at_max_writeback_rate() returns.
 196          */
 197         if (!idle_counter_exceeded(c) ||
 198             !atomic_read(&c->at_max_writeback_rate))
 199                 return false;
 200
 201         return true;
 202 }
 203
 204 static void update_writeback_rate(struct work_struct *work)
 205 {
 206         struct cached_dev *dc = container_of(to_delayed_work(work),
 207                                              struct cached_dev,
 208                                              writeback_rate_update);
 209         struct cache_set *c = dc->disk.c;
 210
 211         /*
 212          * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 213          * cancel_delayed_work_sync().
 214          */
 215         set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 216         /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 217         smp_mb__after_atomic();
 218
 219         /*
 220          * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 221          * check it here too.
 222          */
 223         if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
 224             test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 225                 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 226                 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 227                 smp_mb__after_atomic();
 228                 return;
 229         }
 230
 231         if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
 232                 /*
 233                  * If the whole cache set is idle, set_at_max_writeback_rate()
 234                  * will set writeback rate to a max number. Then it is
 235                  * unncessary to update writeback rate for an idle cache set
 236                  * in maximum writeback rate number(s).
 237                  */
 238                 if (!set_at_max_writeback_rate(c, dc)) {
 239                         down_read(&dc->writeback_lock);
 240                         __update_writeback_rate(dc);
 241                         update_gc_after_writeback(c);
 242                         up_read(&dc->writeback_lock);
 243                 }
 244         }
 245
 246
 247         /*
 248          * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 249          * check it here too.
 250          */
 251         if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
 252             !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 253                 schedule_delayed_work(&dc->writeback_rate_update,
 254                               dc->writeback_rate_update_seconds * HZ);
 255         }
 256
 257         /*
 258          * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 259          * cancel_delayed_work_sync().
 260          */
 261         clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 262         /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 263         smp_mb__after_atomic();
 264 }
 265
 266 static unsigned int writeback_delay(struct cached_dev *dc,
 267                                     unsigned int sectors)
 268 {
 269         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 270             !dc->writeback_percent)
 271                 return 0;
 272
 273         return bch_next_delay(&dc->writeback_rate, sectors);
 274 }
 275
 276 struct dirty_io {
 277         struct closure          cl;
 278         struct cached_dev       *dc;
 279         uint16_t                sequence;
 280         struct bio              bio;
 281 };
 282
 283 static void dirty_init(struct keybuf_key *w)
 284 {
 285         struct dirty_io *io = w->private;
 286         struct bio *bio = &io->bio;
 287
 288         bio_init(bio, bio->bi_inline_vecs,
 289                  DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
 290         if (!io->dc->writeback_percent)
 291                 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 292
 293         bio->bi_iter.bi_size    = KEY_SIZE(&w->key) << 9;
 294         bio->bi_private         = w;
 295         bch_bio_map(bio, NULL);
 296 }
 297
 298 static void dirty_io_destructor(struct closure *cl)
 299 {
 300         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 301
 302         kfree(io);
 303 }
 304
 305 static void write_dirty_finish(struct closure *cl)
 306 {
 307         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 308         struct keybuf_key *w = io->bio.bi_private;
 309         struct cached_dev *dc = io->dc;
 310
 311         bio_free_pages(&io->bio);
 312
 313         /* This is kind of a dumb way of signalling errors. */
 314         if (KEY_DIRTY(&w->key)) {
 315                 int ret;
 316                 unsigned int i;
 317                 struct keylist keys;
 318
 319                 bch_keylist_init(&keys);
 320
 321                 bkey_copy(keys.top, &w->key);
 322                 SET_KEY_DIRTY(keys.top, false);
 323                 bch_keylist_push(&keys);
 324
 325                 for (i = 0; i < KEY_PTRS(&w->key); i++)
 326                         atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 327
 328                 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
 329
 330                 if (ret)
 331                         trace_bcache_writeback_collision(&w->key);
 332
 333                 atomic_long_inc(ret
 334                                 ? &dc->disk.c->writeback_keys_failed
 335                                 : &dc->disk.c->writeback_keys_done);
 336         }
 337
 338         bch_keybuf_del(&dc->writeback_keys, w);
 339         up(&dc->in_flight);
 340
 341         closure_return_with_destructor(cl, dirty_io_destructor);
 342 }
 343
 344 static void dirty_endio(struct bio *bio)
 345 {
 346         struct keybuf_key *w = bio->bi_private;
 347         struct dirty_io *io = w->private;
 348
 349         if (bio->bi_status) {
 350                 SET_KEY_DIRTY(&w->key, false);
 351                 bch_count_backing_io_errors(io->dc, bio);
 352         }
 353
 354         closure_put(&io->cl);
 355 }
 356
 357 static void write_dirty(struct closure *cl)
 358 {
 359         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 360         struct keybuf_key *w = io->bio.bi_private;
 361         struct cached_dev *dc = io->dc;
 362
 363         uint16_t next_sequence;
 364
 365         if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
 366                 /* Not our turn to write; wait for a write to complete */
 367                 closure_wait(&dc->writeback_ordering_wait, cl);
 368
 369                 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
 370                         /*
 371                          * Edge case-- it happened in indeterminate order
 372                          * relative to when we were added to wait list..
 373                          */
 374                         closure_wake_up(&dc->writeback_ordering_wait);
 375                 }
 376
 377                 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 378                 return;
 379         }
 380
 381         next_sequence = io->sequence + 1;
 382
 383         /*
 384          * IO errors are signalled using the dirty bit on the key.
 385          * If we failed to read, we should not attempt to write to the
 386          * backing device.  Instead, immediately go to write_dirty_finish
 387          * to clean up.
 388          */
 389         if (KEY_DIRTY(&w->key)) {
 390                 dirty_init(w);
 391                 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
 392                 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
 393                 bio_set_dev(&io->bio, io->dc->bdev);
 394                 io->bio.bi_end_io       = dirty_endio;
 395
 396                 /* I/O request sent to backing device */
 397                 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 398         }
 399
 400         atomic_set(&dc->writeback_sequence_next, next_sequence);
 401         closure_wake_up(&dc->writeback_ordering_wait);
 402
 403         continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
 404 }
 405
 406 static void read_dirty_endio(struct bio *bio)
 407 {
 408         struct keybuf_key *w = bio->bi_private;
 409         struct dirty_io *io = w->private;
 410
 411         /* is_read = 1 */
 412         bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
 413                             bio->bi_status, 1,
 414                             "reading dirty data from cache");
 415
 416         dirty_endio(bio);
 417 }
 418
 419 static void read_dirty_submit(struct closure *cl)
 420 {
 421         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 422
 423         closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 424
 425         continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 426 }
 427
 428 static void read_dirty(struct cached_dev *dc)
 429 {
 430         unsigned int delay = 0;
 431         struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
 432         size_t size;
 433         int nk, i;
 434         struct dirty_io *io;
 435         struct closure cl;
 436         uint16_t sequence = 0;
 437
 438         BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
 439         atomic_set(&dc->writeback_sequence_next, sequence);
 440         closure_init_stack(&cl);
 441
 442         /*
 443          * XXX: if we error, background writeback just spins. Should use some
 444          * mempools.
 445          */
 446
 447         next = bch_keybuf_next(&dc->writeback_keys);
 448
 449         while (!kthread_should_stop() &&
 450                !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 451                next) {
 452                 size = 0;
 453                 nk = 0;
 454
 455                 do {
 456                         BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
 457
 458                         /*
 459                          * Don't combine too many operations, even if they
 460                          * are all small.
 461                          */
 462                         if (nk >= MAX_WRITEBACKS_IN_PASS)
 463                                 break;
 464
 465                         /*
 466                          * If the current operation is very large, don't
 467                          * further combine operations.
 468                          */
 469                         if (size >= MAX_WRITESIZE_IN_PASS)
 470                                 break;
 471
 472                         /*
 473                          * Operations are only eligible to be combined
 474                          * if they are contiguous.
 475                          *
 476                          * TODO: add a heuristic willing to fire a
 477                          * certain amount of non-contiguous IO per pass,
 478                          * so that we can benefit from backing device
 479                          * command queueing.
 480                          */
 481                         if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
 482                                                 &START_KEY(&next->key)))
 483                                 break;
 484
 485                         size += KEY_SIZE(&next->key);
 486                         keys[nk++] = next;
 487                 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
 488
 489                 /* Now we have gathered a set of 1..5 keys to write back. */
 490                 for (i = 0; i < nk; i++) {
 491                         w = keys[i];
 492
 493                         io = kzalloc(struct_size(io, bio.bi_inline_vecs,
 494                                                 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
 495                                      GFP_KERNEL);
 496                         if (!io)
 497                                 goto err;
 498
 499                         w->private      = io;
 500                         io->dc          = dc;
 501                         io->sequence    = sequence++;
 502
 503                         dirty_init(w);
 504                         bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
 505                         io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
 506                         bio_set_dev(&io->bio,
 507                                     PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
 508                         io->bio.bi_end_io       = read_dirty_endio;
 509
 510                         if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
 511                                 goto err_free;
 512
 513                         trace_bcache_writeback(&w->key);
 514
 515                         down(&dc->in_flight);
 516
 517                         /*
 518                          * We've acquired a semaphore for the maximum
 519                          * simultaneous number of writebacks; from here
 520                          * everything happens asynchronously.
 521                          */
 522                         closure_call(&io->cl, read_dirty_submit, NULL, &cl);
 523                 }
 524
 525                 delay = writeback_delay(dc, size);
 526
 527                 while (!kthread_should_stop() &&
 528                        !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 529                        delay) {
 530                         schedule_timeout_interruptible(delay);
 531                         delay = writeback_delay(dc, 0);
 532                 }
 533         }
 534
 535         if (0) {
 536 err_free:
 537                 kfree(w->private);
 538 err:
 539                 bch_keybuf_del(&dc->writeback_keys, w);
 540         }
 541
 542         /*
 543          * Wait for outstanding writeback IOs to finish (and keybuf slots to be
 544          * freed) before refilling again
 545          */
 546         closure_sync(&cl);
 547 }
 548
 549 /* Scan for dirty data */
 550
 551 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
 552                                   uint64_t offset, int nr_sectors)
 553 {
 554         struct bcache_device *d = c->devices[inode];
 555         unsigned int stripe_offset, sectors_dirty;
 556         int stripe;
 557
 558         if (!d)
 559                 return;
 560
 561         stripe = offset_to_stripe(d, offset);
 562         if (stripe < 0)
 563                 return;
 564
 565         if (UUID_FLASH_ONLY(&c->uuids[inode]))
 566                 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
 567
 568         stripe_offset = offset & (d->stripe_size - 1);
 569
 570         while (nr_sectors) {
 571                 int s = min_t(unsigned int, abs(nr_sectors),
 572                               d->stripe_size - stripe_offset);
 573
 574                 if (nr_sectors < 0)
 575                         s = -s;
 576
 577                 if (stripe >= d->nr_stripes)
 578                         return;
 579
 580                 sectors_dirty = atomic_add_return(s,
 581                                         d->stripe_sectors_dirty + stripe);
 582                 if (sectors_dirty == d->stripe_size)
 583                         set_bit(stripe, d->full_dirty_stripes);
 584                 else
 585                         clear_bit(stripe, d->full_dirty_stripes);
 586
 587                 nr_sectors -= s;
 588                 stripe_offset = 0;
 589                 stripe++;
 590         }
 591 }
 592
 593 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 594 {
 595         struct cached_dev *dc = container_of(buf,
 596                                              struct cached_dev,
 597                                              writeback_keys);
 598
 599         BUG_ON(KEY_INODE(k) != dc->disk.id);
 600
 601         return KEY_DIRTY(k);
 602 }
 603
 604 static void refill_full_stripes(struct cached_dev *dc)
 605 {
 606         struct keybuf *buf = &dc->writeback_keys;
 607         unsigned int start_stripe, next_stripe;
 608         int stripe;
 609         bool wrapped = false;
 610
 611         stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
 612         if (stripe < 0)
 613                 stripe = 0;
 614
 615         start_stripe = stripe;
 616
 617         while (1) {
 618                 stripe = find_next_bit(dc->disk.full_dirty_stripes,
 619                                        dc->disk.nr_stripes, stripe);
 620
 621                 if (stripe == dc->disk.nr_stripes)
 622                         goto next;
 623
 624                 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
 625                                                  dc->disk.nr_stripes, stripe);
 626
 627                 buf->last_scanned = KEY(dc->disk.id,
 628                                         stripe * dc->disk.stripe_size, 0);
 629
 630                 bch_refill_keybuf(dc->disk.c, buf,
 631                                   &KEY(dc->disk.id,
 632                                        next_stripe * dc->disk.stripe_size, 0),
 633                                   dirty_pred);
 634
 635                 if (array_freelist_empty(&buf->freelist))
 636                         return;
 637
 638                 stripe = next_stripe;
 639 next:
 640                 if (wrapped && stripe > start_stripe)
 641                         return;
 642
 643                 if (stripe == dc->disk.nr_stripes) {
 644                         stripe = 0;
 645                         wrapped = true;
 646                 }
 647         }
 648 }
 649
 650 /*
 651  * Returns true if we scanned the entire disk
 652  */
 653 static bool refill_dirty(struct cached_dev *dc)
 654 {
 655         struct keybuf *buf = &dc->writeback_keys;
 656         struct bkey start = KEY(dc->disk.id, 0, 0);
 657         struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 658         struct bkey start_pos;
 659
 660         /*
 661          * make sure keybuf pos is inside the range for this disk - at bringup
 662          * we might not be attached yet so this disk's inode nr isn't
 663          * initialized then
 664          */
 665         if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
 666             bkey_cmp(&buf->last_scanned, &end) > 0)
 667                 buf->last_scanned = start;
 668
 669         if (dc->partial_stripes_expensive) {
 670                 refill_full_stripes(dc);
 671                 if (array_freelist_empty(&buf->freelist))
 672                         return false;
 673         }
 674
 675         start_pos = buf->last_scanned;
 676         bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 677
 678         if (bkey_cmp(&buf->last_scanned, &end) < 0)
 679                 return false;
 680
 681         /*
 682          * If we get to the end start scanning again from the beginning, and
 683          * only scan up to where we initially started scanning from:
 684          */
 685         buf->last_scanned = start;
 686         bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
 687
 688         return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 689 }
 690
 691 static int bch_writeback_thread(void *arg)
 692 {
 693         struct cached_dev *dc = arg;
 694         struct cache_set *c = dc->disk.c;
 695         bool searched_full_index;
 696
 697         bch_ratelimit_reset(&dc->writeback_rate);
 698
 699         while (!kthread_should_stop() &&
 700                !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 701                 down_write(&dc->writeback_lock);
 702                 set_current_state(TASK_INTERRUPTIBLE);
 703                 /*
 704                  * If the bache device is detaching, skip here and continue
 705                  * to perform writeback. Otherwise, if no dirty data on cache,
 706                  * or there is dirty data on cache but writeback is disabled,
 707                  * the writeback thread should sleep here and wait for others
 708                  * to wake up it.
 709                  */
 710                 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
 711                     (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
 712                         up_write(&dc->writeback_lock);
 713
 714                         if (kthread_should_stop() ||
 715                             test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 716                                 set_current_state(TASK_RUNNING);
 717                                 break;
 718                         }
 719
 720                         schedule();
 721                         continue;
 722                 }
 723                 set_current_state(TASK_RUNNING);
 724
 725                 searched_full_index = refill_dirty(dc);
 726
 727                 if (searched_full_index &&
 728                     RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
 729                         atomic_set(&dc->has_dirty, 0);
 730                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 731                         bch_write_bdev_super(dc, NULL);
 732                         /*
 733                          * If bcache device is detaching via sysfs interface,
 734                          * writeback thread should stop after there is no dirty
 735                          * data on cache. BCACHE_DEV_DETACHING flag is set in
 736                          * bch_cached_dev_detach().
 737                          */
 738                         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
 739                                 up_write(&dc->writeback_lock);
 740                                 break;
 741                         }
 742
 743                         /*
 744                          * When dirty data rate is high (e.g. 50%+), there might
 745                          * be heavy buckets fragmentation after writeback
 746                          * finished, which hurts following write performance.
 747                          * If users really care about write performance they
 748                          * may set BCH_ENABLE_AUTO_GC via sysfs, then when
 749                          * BCH_DO_AUTO_GC is set, garbage collection thread
 750                          * will be wake up here. After moving gc, the shrunk
 751                          * btree and discarded free buckets SSD space may be
 752                          * helpful for following write requests.
 753                          */
 754                         if (c->gc_after_writeback ==
 755                             (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
 756                                 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
 757                                 force_wake_up_gc(c);
 758                         }
 759                 }
 760
 761                 up_write(&dc->writeback_lock);
 762
 763                 read_dirty(dc);
 764
 765                 if (searched_full_index) {
 766                         unsigned int delay = dc->writeback_delay * HZ;
 767
 768                         while (delay &&
 769                                !kthread_should_stop() &&
 770                                !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
 771                                !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
 772                                 delay = schedule_timeout_interruptible(delay);
 773
 774                         bch_ratelimit_reset(&dc->writeback_rate);
 775                 }
 776         }
 777
 778         if (dc->writeback_write_wq) {
 779                 flush_workqueue(dc->writeback_write_wq);
 780                 destroy_workqueue(dc->writeback_write_wq);
 781         }
 782         cached_dev_put(dc);
 783         wait_for_kthread_stop();
 784
 785         return 0;
 786 }
 787
 788 /* Init */
 789 #define INIT_KEYS_EACH_TIME     500000
 790
 791 struct sectors_dirty_init {
 792         struct btree_op op;
 793         unsigned int    inode;
 794         size_t          count;
 795 };
 796
 797 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
 798                                  struct bkey *k)
 799 {
 800         struct sectors_dirty_init *op = container_of(_op,
 801                                                 struct sectors_dirty_init, op);
 802         if (KEY_INODE(k) > op->inode)
 803                 return MAP_DONE;
 804
 805         if (KEY_DIRTY(k))
 806                 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
 807                                              KEY_START(k), KEY_SIZE(k));
 808
 809         op->count++;
 810         if (!(op->count % INIT_KEYS_EACH_TIME))
 811                 cond_resched();
 812
 813         return MAP_CONTINUE;
 814 }
 815
 816 static int bch_root_node_dirty_init(struct cache_set *c,
 817                                      struct bcache_device *d,
 818                                      struct bkey *k)
 819 {
 820         struct sectors_dirty_init op;
 821         int ret;
 822
 823         bch_btree_op_init(&op.op, -1);
 824         op.inode = d->id;
 825         op.count = 0;
 826
 827         ret = bcache_btree(map_keys_recurse,
 828                            k,
 829                            c->root,
 830                            &op.op,
 831                            &KEY(op.inode, 0, 0),
 832                            sectors_dirty_init_fn,
 833                            0);
 834         if (ret < 0)
 835                 pr_warn("sectors dirty init failed, ret=%d!\n", ret);
 836
 837         return ret;
 838 }
 839
 840 static int bch_dirty_init_thread(void *arg)
 841 {
 842         struct dirty_init_thrd_info *info = arg;
 843         struct bch_dirty_init_state *state = info->state;
 844         struct cache_set *c = state->c;
 845         struct btree_iter iter;
 846         struct bkey *k, *p;
 847         int cur_idx, prev_idx, skip_nr;
 848
 849         k = p = NULL;
 850         cur_idx = prev_idx = 0;
 851
 852         bch_btree_iter_init(&c->root->keys, &iter, NULL);
 853         k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
 854         BUG_ON(!k);
 855
 856         p = k;
 857
 858         while (k) {
 859                 spin_lock(&state->idx_lock);
 860                 cur_idx = state->key_idx;
 861                 state->key_idx++;
 862                 spin_unlock(&state->idx_lock);
 863
 864                 skip_nr = cur_idx - prev_idx;
 865
 866                 while (skip_nr) {
 867                         k = bch_btree_iter_next_filter(&iter,
 868                                                        &c->root->keys,
 869                                                        bch_ptr_bad);
 870                         if (k)
 871                                 p = k;
 872                         else {
 873                                 atomic_set(&state->enough, 1);
 874                                 /* Update state->enough earlier */
 875                                 smp_mb__after_atomic();
 876                                 goto out;
 877                         }
 878                         skip_nr--;
 879                 }
 880
 881                 if (p) {
 882                         if (bch_root_node_dirty_init(c, state->d, p) < 0)
 883                                 goto out;
 884                 }
 885
 886                 p = NULL;
 887                 prev_idx = cur_idx;
 888         }
 889
 890 out:
 891         /* In order to wake up state->wait in time */
 892         smp_mb__before_atomic();
 893         if (atomic_dec_and_test(&state->started))
 894                 wake_up(&state->wait);
 895
 896         return 0;
 897 }
 898
 899 static int bch_btre_dirty_init_thread_nr(void)
 900 {
 901         int n = num_online_cpus()/2;
 902
 903         if (n == 0)
 904                 n = 1;
 905         else if (n > BCH_DIRTY_INIT_THRD_MAX)
 906                 n = BCH_DIRTY_INIT_THRD_MAX;
 907
 908         return n;
 909 }
 910
 911 void bch_sectors_dirty_init(struct bcache_device *d)
 912 {
 913         int i;
 914         struct bkey *k = NULL;
 915         struct btree_iter iter;
 916         struct sectors_dirty_init op;
 917         struct cache_set *c = d->c;
 918         struct bch_dirty_init_state state;
 919
 920         /* Just count root keys if no leaf node */
 921         rw_lock(0, c->root, c->root->level);
 922         if (c->root->level == 0) {
 923                 bch_btree_op_init(&op.op, -1);
 924                 op.inode = d->id;
 925                 op.count = 0;
 926
 927                 for_each_key_filter(&c->root->keys,
 928                                     k, &iter, bch_ptr_invalid)
 929                         sectors_dirty_init_fn(&op.op, c->root, k);
 930
 931                 rw_unlock(0, c->root);
 932                 return;
 933         }
 934
 935         memset(&state, 0, sizeof(struct bch_dirty_init_state));
 936         state.c = c;
 937         state.d = d;
 938         state.total_threads = bch_btre_dirty_init_thread_nr();
 939         state.key_idx = 0;
 940         spin_lock_init(&state.idx_lock);
 941         atomic_set(&state.started, 0);
 942         atomic_set(&state.enough, 0);
 943         init_waitqueue_head(&state.wait);
 944
 945         for (i = 0; i < state.total_threads; i++) {
 946                 /* Fetch latest state.enough earlier */
 947                 smp_mb__before_atomic();
 948                 if (atomic_read(&state.enough))
 949                         break;
 950
 951                 state.infos[i].state = &state;
 952                 state.infos[i].thread =
 953                         kthread_run(bch_dirty_init_thread, &state.infos[i],
 954                                     "bch_dirtcnt[%d]", i);
 955                 if (IS_ERR(state.infos[i].thread)) {
 956                         pr_err("fails to run thread bch_dirty_init[%d]\n", i);
 957                         for (--i; i >= 0; i--)
 958                                 kthread_stop(state.infos[i].thread);
 959                         goto out;
 960                 }
 961                 atomic_inc(&state.started);
 962         }
 963
 964 out:
 965         /* Must wait for all threads to stop. */
 966         wait_event(state.wait, atomic_read(&state.started) == 0);
 967         rw_unlock(0, c->root);
 968 }
 969
 970 void bch_cached_dev_writeback_init(struct cached_dev *dc)
 971 {
 972         sema_init(&dc->in_flight, 64);
 973         init_rwsem(&dc->writeback_lock);
 974         bch_keybuf_init(&dc->writeback_keys);
 975
 976         dc->writeback_metadata          = true;
 977         dc->writeback_running           = false;
 978         dc->writeback_percent           = 10;
 979         dc->writeback_delay             = 30;
 980         atomic_long_set(&dc->writeback_rate.rate, 1024);
 981         dc->writeback_rate_minimum      = 8;
 982
 983         dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
 984         dc->writeback_rate_p_term_inverse = 40;
 985         dc->writeback_rate_i_term_inverse = 10000;
 986
 987         WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
 988         INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
 989 }
 990
 991 int bch_cached_dev_writeback_start(struct cached_dev *dc)
 992 {
 993         dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
 994                                                 WQ_MEM_RECLAIM, 0);
 995         if (!dc->writeback_write_wq)
 996                 return -ENOMEM;
 997
 998         cached_dev_get(dc);
 999         dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1000                                               "bcache_writeback");
1001         if (IS_ERR(dc->writeback_thread)) {
1002                 cached_dev_put(dc);
1003                 destroy_workqueue(dc->writeback_write_wq);
1004                 return PTR_ERR(dc->writeback_thread);
1005         }
1006         dc->writeback_running = true;
1007
1008         WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1009         schedule_delayed_work(&dc->writeback_rate_update,
1010                               dc->writeback_rate_update_seconds * HZ);
1011
1012         bch_writeback_queue(dc);
1013
1014         return 0;
1015 }