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->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 set_at_max_writeback_rate(struct cache_set *c,
 123                                        struct cached_dev *dc)
 124 {
 125         /*
 126          * Idle_counter is increased everytime when update_writeback_rate() is
 127          * called. If all backing devices attached to the same cache set have
 128          * identical dc->writeback_rate_update_seconds values, it is about 6
 129          * rounds of update_writeback_rate() on each backing device before
 130          * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
 131          * to each dc->writeback_rate.rate.
 132          * In order to avoid extra locking cost for counting exact dirty cached
 133          * devices number, c->attached_dev_nr is used to calculate the idle
 134          * throushold. It might be bigger if not all cached device are in write-
 135          * back mode, but it still works well with limited extra rounds of
 136          * update_writeback_rate().
 137          */
 138         if (atomic_inc_return(&c->idle_counter) <
 139             atomic_read(&c->attached_dev_nr) * 6)
 140                 return false;
 141
 142         if (atomic_read(&c->at_max_writeback_rate) != 1)
 143                 atomic_set(&c->at_max_writeback_rate, 1);
 144
 145         atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
 146
 147         /* keep writeback_rate_target as existing value */
 148         dc->writeback_rate_proportional = 0;
 149         dc->writeback_rate_integral_scaled = 0;
 150         dc->writeback_rate_change = 0;
 151
 152         /*
 153          * Check c->idle_counter and c->at_max_writeback_rate agagain in case
 154          * new I/O arrives during before set_at_max_writeback_rate() returns.
 155          * Then the writeback rate is set to 1, and its new value should be
 156          * decided via __update_writeback_rate().
 157          */
 158         if ((atomic_read(&c->idle_counter) <
 159              atomic_read(&c->attached_dev_nr) * 6) ||
 160             !atomic_read(&c->at_max_writeback_rate))
 161                 return false;
 162
 163         return true;
 164 }
 165
 166 static void update_writeback_rate(struct work_struct *work)
 167 {
 168         struct cached_dev *dc = container_of(to_delayed_work(work),
 169                                              struct cached_dev,
 170                                              writeback_rate_update);
 171         struct cache_set *c = dc->disk.c;
 172
 173         /*
 174          * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 175          * cancel_delayed_work_sync().
 176          */
 177         set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 178         /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 179         smp_mb();
 180
 181         /*
 182          * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 183          * check it here too.
 184          */
 185         if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
 186             test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 187                 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 188                 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 189                 smp_mb();
 190                 return;
 191         }
 192
 193         if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
 194                 /*
 195                  * If the whole cache set is idle, set_at_max_writeback_rate()
 196                  * will set writeback rate to a max number. Then it is
 197                  * unncessary to update writeback rate for an idle cache set
 198                  * in maximum writeback rate number(s).
 199                  */
 200                 if (!set_at_max_writeback_rate(c, dc)) {
 201                         down_read(&dc->writeback_lock);
 202                         __update_writeback_rate(dc);
 203                         update_gc_after_writeback(c);
 204                         up_read(&dc->writeback_lock);
 205                 }
 206         }
 207
 208
 209         /*
 210          * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 211          * check it here too.
 212          */
 213         if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
 214             !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 215                 schedule_delayed_work(&dc->writeback_rate_update,
 216                               dc->writeback_rate_update_seconds * HZ);
 217         }
 218
 219         /*
 220          * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 221          * cancel_delayed_work_sync().
 222          */
 223         clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 224         /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 225         smp_mb();
 226 }
 227
 228 static unsigned int writeback_delay(struct cached_dev *dc,
 229                                     unsigned int sectors)
 230 {
 231         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 232             !dc->writeback_percent)
 233                 return 0;
 234
 235         return bch_next_delay(&dc->writeback_rate, sectors);
 236 }
 237
 238 struct dirty_io {
 239         struct closure          cl;
 240         struct cached_dev       *dc;
 241         uint16_t                sequence;
 242         struct bio              bio;
 243 };
 244
 245 static void dirty_init(struct keybuf_key *w)
 246 {
 247         struct dirty_io *io = w->private;
 248         struct bio *bio = &io->bio;
 249
 250         bio_init(bio, bio->bi_inline_vecs,
 251                  DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
 252         if (!io->dc->writeback_percent)
 253                 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 254
 255         bio->bi_iter.bi_size    = KEY_SIZE(&w->key) << 9;
 256         bio->bi_private         = w;
 257         bch_bio_map(bio, NULL);
 258 }
 259
 260 static void dirty_io_destructor(struct closure *cl)
 261 {
 262         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 263
 264         kfree(io);
 265 }
 266
 267 static void write_dirty_finish(struct closure *cl)
 268 {
 269         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 270         struct keybuf_key *w = io->bio.bi_private;
 271         struct cached_dev *dc = io->dc;
 272
 273         bio_free_pages(&io->bio);
 274
 275         /* This is kind of a dumb way of signalling errors. */
 276         if (KEY_DIRTY(&w->key)) {
 277                 int ret;
 278                 unsigned int i;
 279                 struct keylist keys;
 280
 281                 bch_keylist_init(&keys);
 282
 283                 bkey_copy(keys.top, &w->key);
 284                 SET_KEY_DIRTY(keys.top, false);
 285                 bch_keylist_push(&keys);
 286
 287                 for (i = 0; i < KEY_PTRS(&w->key); i++)
 288                         atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 289
 290                 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
 291
 292                 if (ret)
 293                         trace_bcache_writeback_collision(&w->key);
 294
 295                 atomic_long_inc(ret
 296                                 ? &dc->disk.c->writeback_keys_failed
 297                                 : &dc->disk.c->writeback_keys_done);
 298         }
 299
 300         bch_keybuf_del(&dc->writeback_keys, w);
 301         up(&dc->in_flight);
 302
 303         closure_return_with_destructor(cl, dirty_io_destructor);
 304 }
 305
 306 static void dirty_endio(struct bio *bio)
 307 {
 308         struct keybuf_key *w = bio->bi_private;
 309         struct dirty_io *io = w->private;
 310
 311         if (bio->bi_status) {
 312                 SET_KEY_DIRTY(&w->key, false);
 313                 bch_count_backing_io_errors(io->dc, bio);
 314         }
 315
 316         closure_put(&io->cl);
 317 }
 318
 319 static void write_dirty(struct closure *cl)
 320 {
 321         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 322         struct keybuf_key *w = io->bio.bi_private;
 323         struct cached_dev *dc = io->dc;
 324
 325         uint16_t next_sequence;
 326
 327         if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
 328                 /* Not our turn to write; wait for a write to complete */
 329                 closure_wait(&dc->writeback_ordering_wait, cl);
 330
 331                 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
 332                         /*
 333                          * Edge case-- it happened in indeterminate order
 334                          * relative to when we were added to wait list..
 335                          */
 336                         closure_wake_up(&dc->writeback_ordering_wait);
 337                 }
 338
 339                 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 340                 return;
 341         }
 342
 343         next_sequence = io->sequence + 1;
 344
 345         /*
 346          * IO errors are signalled using the dirty bit on the key.
 347          * If we failed to read, we should not attempt to write to the
 348          * backing device.  Instead, immediately go to write_dirty_finish
 349          * to clean up.
 350          */
 351         if (KEY_DIRTY(&w->key)) {
 352                 dirty_init(w);
 353                 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
 354                 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
 355                 bio_set_dev(&io->bio, io->dc->bdev);
 356                 io->bio.bi_end_io       = dirty_endio;
 357
 358                 /* I/O request sent to backing device */
 359                 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 360         }
 361
 362         atomic_set(&dc->writeback_sequence_next, next_sequence);
 363         closure_wake_up(&dc->writeback_ordering_wait);
 364
 365         continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
 366 }
 367
 368 static void read_dirty_endio(struct bio *bio)
 369 {
 370         struct keybuf_key *w = bio->bi_private;
 371         struct dirty_io *io = w->private;
 372
 373         /* is_read = 1 */
 374         bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
 375                             bio->bi_status, 1,
 376                             "reading dirty data from cache");
 377
 378         dirty_endio(bio);
 379 }
 380
 381 static void read_dirty_submit(struct closure *cl)
 382 {
 383         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 384
 385         closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 386
 387         continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 388 }
 389
 390 static void read_dirty(struct cached_dev *dc)
 391 {
 392         unsigned int delay = 0;
 393         struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
 394         size_t size;
 395         int nk, i;
 396         struct dirty_io *io;
 397         struct closure cl;
 398         uint16_t sequence = 0;
 399
 400         BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
 401         atomic_set(&dc->writeback_sequence_next, sequence);
 402         closure_init_stack(&cl);
 403
 404         /*
 405          * XXX: if we error, background writeback just spins. Should use some
 406          * mempools.
 407          */
 408
 409         next = bch_keybuf_next(&dc->writeback_keys);
 410
 411         while (!kthread_should_stop() &&
 412                !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 413                next) {
 414                 size = 0;
 415                 nk = 0;
 416
 417                 do {
 418                         BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
 419
 420                         /*
 421                          * Don't combine too many operations, even if they
 422                          * are all small.
 423                          */
 424                         if (nk >= MAX_WRITEBACKS_IN_PASS)
 425                                 break;
 426
 427                         /*
 428                          * If the current operation is very large, don't
 429                          * further combine operations.
 430                          */
 431                         if (size >= MAX_WRITESIZE_IN_PASS)
 432                                 break;
 433
 434                         /*
 435                          * Operations are only eligible to be combined
 436                          * if they are contiguous.
 437                          *
 438                          * TODO: add a heuristic willing to fire a
 439                          * certain amount of non-contiguous IO per pass,
 440                          * so that we can benefit from backing device
 441                          * command queueing.
 442                          */
 443                         if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
 444                                                 &START_KEY(&next->key)))
 445                                 break;
 446
 447                         size += KEY_SIZE(&next->key);
 448                         keys[nk++] = next;
 449                 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
 450
 451                 /* Now we have gathered a set of 1..5 keys to write back. */
 452                 for (i = 0; i < nk; i++) {
 453                         w = keys[i];
 454
 455                         io = kzalloc(sizeof(struct dirty_io) +
 456                                      sizeof(struct bio_vec) *
 457                                      DIV_ROUND_UP(KEY_SIZE(&w->key),
 458                                                   PAGE_SECTORS),
 459                                      GFP_KERNEL);
 460                         if (!io)
 461                                 goto err;
 462
 463                         w->private      = io;
 464                         io->dc          = dc;
 465                         io->sequence    = sequence++;
 466
 467                         dirty_init(w);
 468                         bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
 469                         io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
 470                         bio_set_dev(&io->bio,
 471                                     PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
 472                         io->bio.bi_end_io       = read_dirty_endio;
 473
 474                         if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
 475                                 goto err_free;
 476
 477                         trace_bcache_writeback(&w->key);
 478
 479                         down(&dc->in_flight);
 480
 481                         /*
 482                          * We've acquired a semaphore for the maximum
 483                          * simultaneous number of writebacks; from here
 484                          * everything happens asynchronously.
 485                          */
 486                         closure_call(&io->cl, read_dirty_submit, NULL, &cl);
 487                 }
 488
 489                 delay = writeback_delay(dc, size);
 490
 491                 while (!kthread_should_stop() &&
 492                        !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 493                        delay) {
 494                         schedule_timeout_interruptible(delay);
 495                         delay = writeback_delay(dc, 0);
 496                 }
 497         }
 498
 499         if (0) {
 500 err_free:
 501                 kfree(w->private);
 502 err:
 503                 bch_keybuf_del(&dc->writeback_keys, w);
 504         }
 505
 506         /*
 507          * Wait for outstanding writeback IOs to finish (and keybuf slots to be
 508          * freed) before refilling again
 509          */
 510         closure_sync(&cl);
 511 }
 512
 513 /* Scan for dirty data */
 514
 515 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
 516                                   uint64_t offset, int nr_sectors)
 517 {
 518         struct bcache_device *d = c->devices[inode];
 519         unsigned int stripe_offset, stripe, sectors_dirty;
 520
 521         if (!d)
 522                 return;
 523
 524         if (UUID_FLASH_ONLY(&c->uuids[inode]))
 525                 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
 526
 527         stripe = offset_to_stripe(d, offset);
 528         stripe_offset = offset & (d->stripe_size - 1);
 529
 530         while (nr_sectors) {
 531                 int s = min_t(unsigned int, abs(nr_sectors),
 532                               d->stripe_size - stripe_offset);
 533
 534                 if (nr_sectors < 0)
 535                         s = -s;
 536
 537                 if (stripe >= d->nr_stripes)
 538                         return;
 539
 540                 sectors_dirty = atomic_add_return(s,
 541                                         d->stripe_sectors_dirty + stripe);
 542                 if (sectors_dirty == d->stripe_size)
 543                         set_bit(stripe, d->full_dirty_stripes);
 544                 else
 545                         clear_bit(stripe, d->full_dirty_stripes);
 546
 547                 nr_sectors -= s;
 548                 stripe_offset = 0;
 549                 stripe++;
 550         }
 551 }
 552
 553 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 554 {
 555         struct cached_dev *dc = container_of(buf,
 556                                              struct cached_dev,
 557                                              writeback_keys);
 558
 559         BUG_ON(KEY_INODE(k) != dc->disk.id);
 560
 561         return KEY_DIRTY(k);
 562 }
 563
 564 static void refill_full_stripes(struct cached_dev *dc)
 565 {
 566         struct keybuf *buf = &dc->writeback_keys;
 567         unsigned int start_stripe, stripe, next_stripe;
 568         bool wrapped = false;
 569
 570         stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
 571
 572         if (stripe >= dc->disk.nr_stripes)
 573                 stripe = 0;
 574
 575         start_stripe = stripe;
 576
 577         while (1) {
 578                 stripe = find_next_bit(dc->disk.full_dirty_stripes,
 579                                        dc->disk.nr_stripes, stripe);
 580
 581                 if (stripe == dc->disk.nr_stripes)
 582                         goto next;
 583
 584                 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
 585                                                  dc->disk.nr_stripes, stripe);
 586
 587                 buf->last_scanned = KEY(dc->disk.id,
 588                                         stripe * dc->disk.stripe_size, 0);
 589
 590                 bch_refill_keybuf(dc->disk.c, buf,
 591                                   &KEY(dc->disk.id,
 592                                        next_stripe * dc->disk.stripe_size, 0),
 593                                   dirty_pred);
 594
 595                 if (array_freelist_empty(&buf->freelist))
 596                         return;
 597
 598                 stripe = next_stripe;
 599 next:
 600                 if (wrapped && stripe > start_stripe)
 601                         return;
 602
 603                 if (stripe == dc->disk.nr_stripes) {
 604                         stripe = 0;
 605                         wrapped = true;
 606                 }
 607         }
 608 }
 609
 610 /*
 611  * Returns true if we scanned the entire disk
 612  */
 613 static bool refill_dirty(struct cached_dev *dc)
 614 {
 615         struct keybuf *buf = &dc->writeback_keys;
 616         struct bkey start = KEY(dc->disk.id, 0, 0);
 617         struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 618         struct bkey start_pos;
 619
 620         /*
 621          * make sure keybuf pos is inside the range for this disk - at bringup
 622          * we might not be attached yet so this disk's inode nr isn't
 623          * initialized then
 624          */
 625         if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
 626             bkey_cmp(&buf->last_scanned, &end) > 0)
 627                 buf->last_scanned = start;
 628
 629         if (dc->partial_stripes_expensive) {
 630                 refill_full_stripes(dc);
 631                 if (array_freelist_empty(&buf->freelist))
 632                         return false;
 633         }
 634
 635         start_pos = buf->last_scanned;
 636         bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 637
 638         if (bkey_cmp(&buf->last_scanned, &end) < 0)
 639                 return false;
 640
 641         /*
 642          * If we get to the end start scanning again from the beginning, and
 643          * only scan up to where we initially started scanning from:
 644          */
 645         buf->last_scanned = start;
 646         bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
 647
 648         return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 649 }
 650
 651 static int bch_writeback_thread(void *arg)
 652 {
 653         struct cached_dev *dc = arg;
 654         struct cache_set *c = dc->disk.c;
 655         bool searched_full_index;
 656
 657         bch_ratelimit_reset(&dc->writeback_rate);
 658
 659         while (!kthread_should_stop() &&
 660                !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 661                 down_write(&dc->writeback_lock);
 662                 set_current_state(TASK_INTERRUPTIBLE);
 663                 /*
 664                  * If the bache device is detaching, skip here and continue
 665                  * to perform writeback. Otherwise, if no dirty data on cache,
 666                  * or there is dirty data on cache but writeback is disabled,
 667                  * the writeback thread should sleep here and wait for others
 668                  * to wake up it.
 669                  */
 670                 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
 671                     (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
 672                         up_write(&dc->writeback_lock);
 673
 674                         if (kthread_should_stop() ||
 675                             test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 676                                 set_current_state(TASK_RUNNING);
 677                                 break;
 678                         }
 679
 680                         schedule();
 681                         continue;
 682                 }
 683                 set_current_state(TASK_RUNNING);
 684
 685                 searched_full_index = refill_dirty(dc);
 686
 687                 if (searched_full_index &&
 688                     RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
 689                         atomic_set(&dc->has_dirty, 0);
 690                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 691                         bch_write_bdev_super(dc, NULL);
 692                         /*
 693                          * If bcache device is detaching via sysfs interface,
 694                          * writeback thread should stop after there is no dirty
 695                          * data on cache. BCACHE_DEV_DETACHING flag is set in
 696                          * bch_cached_dev_detach().
 697                          */
 698                         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
 699                                 up_write(&dc->writeback_lock);
 700                                 break;
 701                         }
 702
 703                         /*
 704                          * When dirty data rate is high (e.g. 50%+), there might
 705                          * be heavy buckets fragmentation after writeback
 706                          * finished, which hurts following write performance.
 707                          * If users really care about write performance they
 708                          * may set BCH_ENABLE_AUTO_GC via sysfs, then when
 709                          * BCH_DO_AUTO_GC is set, garbage collection thread
 710                          * will be wake up here. After moving gc, the shrunk
 711                          * btree and discarded free buckets SSD space may be
 712                          * helpful for following write requests.
 713                          */
 714                         if (c->gc_after_writeback ==
 715                             (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
 716                                 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
 717                                 force_wake_up_gc(c);
 718                         }
 719                 }
 720
 721                 up_write(&dc->writeback_lock);
 722
 723                 read_dirty(dc);
 724
 725                 if (searched_full_index) {
 726                         unsigned int delay = dc->writeback_delay * HZ;
 727
 728                         while (delay &&
 729                                !kthread_should_stop() &&
 730                                !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
 731                                !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
 732                                 delay = schedule_timeout_interruptible(delay);
 733
 734                         bch_ratelimit_reset(&dc->writeback_rate);
 735                 }
 736         }
 737
 738         cached_dev_put(dc);
 739         wait_for_kthread_stop();
 740
 741         return 0;
 742 }
 743
 744 /* Init */
 745 #define INIT_KEYS_EACH_TIME     500000
 746 #define INIT_KEYS_SLEEP_MS      100
 747
 748 struct sectors_dirty_init {
 749         struct btree_op op;
 750         unsigned int    inode;
 751         size_t          count;
 752         struct bkey     start;
 753 };
 754
 755 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
 756                                  struct bkey *k)
 757 {
 758         struct sectors_dirty_init *op = container_of(_op,
 759                                                 struct sectors_dirty_init, op);
 760         if (KEY_INODE(k) > op->inode)
 761                 return MAP_DONE;
 762
 763         if (KEY_DIRTY(k))
 764                 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
 765                                              KEY_START(k), KEY_SIZE(k));
 766
 767         op->count++;
 768         if (atomic_read(&b->c->search_inflight) &&
 769             !(op->count % INIT_KEYS_EACH_TIME)) {
 770                 bkey_copy_key(&op->start, k);
 771                 return -EAGAIN;
 772         }
 773
 774         return MAP_CONTINUE;
 775 }
 776
 777 void bch_sectors_dirty_init(struct bcache_device *d)
 778 {
 779         struct sectors_dirty_init op;
 780         int ret;
 781
 782         bch_btree_op_init(&op.op, -1);
 783         op.inode = d->id;
 784         op.count = 0;
 785         op.start = KEY(op.inode, 0, 0);
 786
 787         do {
 788                 ret = bch_btree_map_keys(&op.op, d->c, &op.start,
 789                                          sectors_dirty_init_fn, 0);
 790                 if (ret == -EAGAIN)
 791                         schedule_timeout_interruptible(
 792                                 msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
 793                 else if (ret < 0) {
 794                         pr_warn("sectors dirty init failed, ret=%d!", ret);
 795                         break;
 796                 }
 797         } while (ret == -EAGAIN);
 798 }
 799
 800 void bch_cached_dev_writeback_init(struct cached_dev *dc)
 801 {
 802         sema_init(&dc->in_flight, 64);
 803         init_rwsem(&dc->writeback_lock);
 804         bch_keybuf_init(&dc->writeback_keys);
 805
 806         dc->writeback_metadata          = true;
 807         dc->writeback_running           = false;
 808         dc->writeback_percent           = 10;
 809         dc->writeback_delay             = 30;
 810         atomic_long_set(&dc->writeback_rate.rate, 1024);
 811         dc->writeback_rate_minimum      = 8;
 812
 813         dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
 814         dc->writeback_rate_p_term_inverse = 40;
 815         dc->writeback_rate_i_term_inverse = 10000;
 816
 817         WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
 818         INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
 819 }
 820
 821 int bch_cached_dev_writeback_start(struct cached_dev *dc)
 822 {
 823         dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
 824                                                 WQ_MEM_RECLAIM, 0);
 825         if (!dc->writeback_write_wq)
 826                 return -ENOMEM;
 827
 828         cached_dev_get(dc);
 829         dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
 830                                               "bcache_writeback");
 831         if (IS_ERR(dc->writeback_thread)) {
 832                 cached_dev_put(dc);
 833                 return PTR_ERR(dc->writeback_thread);
 834         }
 835         dc->writeback_running = true;
 836
 837         WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
 838         schedule_delayed_work(&dc->writeback_rate_update,
 839                               dc->writeback_rate_update_seconds * HZ);
 840
 841         bch_writeback_queue(dc);
 842
 843         return 0;
 844 }