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_nr_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         /*
  92          * We need to consider the number of dirty buckets as well
  93          * when calculating the proportional_scaled, Otherwise we might
  94          * have an unreasonable small writeback rate at a highly fragmented situation
  95          * when very few dirty sectors consumed a lot dirty buckets, the
  96          * worst case is when dirty buckets reached cutoff_writeback_sync and
  97          * dirty data is still not even reached to writeback percent, so the rate
  98          * still will be at the minimum value, which will cause the write
  99          * stuck at a non-writeback mode.
 100          */
 101         struct cache_set *c = dc->disk.c;
 102
 103         int64_t dirty_buckets = c->nbuckets - c->avail_nbuckets;
 104
 105         if (dc->writeback_consider_fragment &&
 106                 c->gc_stats.in_use > BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW && dirty > 0) {
 107                 int64_t fragment =
 108                         div_s64((dirty_buckets *  c->cache->sb.bucket_size), dirty);
 109                 int64_t fp_term;
 110                 int64_t fps;
 111
 112                 if (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID) {
 113                         fp_term = (int64_t)dc->writeback_rate_fp_term_low *
 114                         (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW);
 115                 } else if (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH) {
 116                         fp_term = (int64_t)dc->writeback_rate_fp_term_mid *
 117                         (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID);
 118                 } else {
 119                         fp_term = (int64_t)dc->writeback_rate_fp_term_high *
 120                         (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH);
 121                 }
 122                 fps = div_s64(dirty, dirty_buckets) * fp_term;
 123                 if (fragment > 3 && fps > proportional_scaled) {
 124                         /* Only overrite the p when fragment > 3 */
 125                         proportional_scaled = fps;
 126                 }
 127         }
 128
 129         if ((error < 0 && dc->writeback_rate_integral > 0) ||
 130             (error > 0 && time_before64(local_clock(),
 131                          dc->writeback_rate.next + NSEC_PER_MSEC))) {
 132                 /*
 133                  * Only decrease the integral term if it's more than
 134                  * zero.  Only increase the integral term if the device
 135                  * is keeping up.  (Don't wind up the integral
 136                  * ineffectively in either case).
 137                  *
 138                  * It's necessary to scale this by
 139                  * writeback_rate_update_seconds to keep the integral
 140                  * term dimensioned properly.
 141                  */
 142                 dc->writeback_rate_integral += error *
 143                         dc->writeback_rate_update_seconds;
 144         }
 145
 146         integral_scaled = div_s64(dc->writeback_rate_integral,
 147                         dc->writeback_rate_i_term_inverse);
 148
 149         new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
 150                         dc->writeback_rate_minimum, NSEC_PER_SEC);
 151
 152         dc->writeback_rate_proportional = proportional_scaled;
 153         dc->writeback_rate_integral_scaled = integral_scaled;
 154         dc->writeback_rate_change = new_rate -
 155                         atomic_long_read(&dc->writeback_rate.rate);
 156         atomic_long_set(&dc->writeback_rate.rate, new_rate);
 157         dc->writeback_rate_target = target;
 158 }
 159
 160 static bool set_at_max_writeback_rate(struct cache_set *c,
 161                                        struct cached_dev *dc)
 162 {
 163         /* Don't sst max writeback rate if it is disabled */
 164         if (!c->idle_max_writeback_rate_enabled)
 165                 return false;
 166
 167         /* Don't set max writeback rate if gc is running */
 168         if (!c->gc_mark_valid)
 169                 return false;
 170         /*
 171          * Idle_counter is increased everytime when update_writeback_rate() is
 172          * called. If all backing devices attached to the same cache set have
 173          * identical dc->writeback_rate_update_seconds values, it is about 6
 174          * rounds of update_writeback_rate() on each backing device before
 175          * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
 176          * to each dc->writeback_rate.rate.
 177          * In order to avoid extra locking cost for counting exact dirty cached
 178          * devices number, c->attached_dev_nr is used to calculate the idle
 179          * throushold. It might be bigger if not all cached device are in write-
 180          * back mode, but it still works well with limited extra rounds of
 181          * update_writeback_rate().
 182          */
 183         if (atomic_inc_return(&c->idle_counter) <
 184             atomic_read(&c->attached_dev_nr) * 6)
 185                 return false;
 186
 187         if (atomic_read(&c->at_max_writeback_rate) != 1)
 188                 atomic_set(&c->at_max_writeback_rate, 1);
 189
 190         atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
 191
 192         /* keep writeback_rate_target as existing value */
 193         dc->writeback_rate_proportional = 0;
 194         dc->writeback_rate_integral_scaled = 0;
 195         dc->writeback_rate_change = 0;
 196
 197         /*
 198          * Check c->idle_counter and c->at_max_writeback_rate agagain in case
 199          * new I/O arrives during before set_at_max_writeback_rate() returns.
 200          * Then the writeback rate is set to 1, and its new value should be
 201          * decided via __update_writeback_rate().
 202          */
 203         if ((atomic_read(&c->idle_counter) <
 204              atomic_read(&c->attached_dev_nr) * 6) ||
 205             !atomic_read(&c->at_max_writeback_rate))
 206                 return false;
 207
 208         return true;
 209 }
 210
 211 static void update_writeback_rate(struct work_struct *work)
 212 {
 213         struct cached_dev *dc = container_of(to_delayed_work(work),
 214                                              struct cached_dev,
 215                                              writeback_rate_update);
 216         struct cache_set *c = dc->disk.c;
 217
 218         /*
 219          * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 220          * cancel_delayed_work_sync().
 221          */
 222         set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 223         /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 224         smp_mb__after_atomic();
 225
 226         /*
 227          * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 228          * check it here too.
 229          */
 230         if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
 231             test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 232                 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 233                 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 234                 smp_mb__after_atomic();
 235                 return;
 236         }
 237
 238         if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
 239                 /*
 240                  * If the whole cache set is idle, set_at_max_writeback_rate()
 241                  * will set writeback rate to a max number. Then it is
 242                  * unncessary to update writeback rate for an idle cache set
 243                  * in maximum writeback rate number(s).
 244                  */
 245                 if (!set_at_max_writeback_rate(c, dc)) {
 246                         down_read(&dc->writeback_lock);
 247                         __update_writeback_rate(dc);
 248                         update_gc_after_writeback(c);
 249                         up_read(&dc->writeback_lock);
 250                 }
 251         }
 252
 253
 254         /*
 255          * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 256          * check it here too.
 257          */
 258         if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
 259             !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 260                 schedule_delayed_work(&dc->writeback_rate_update,
 261                               dc->writeback_rate_update_seconds * HZ);
 262         }
 263
 264         /*
 265          * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 266          * cancel_delayed_work_sync().
 267          */
 268         clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 269         /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 270         smp_mb__after_atomic();
 271 }
 272
 273 static unsigned int writeback_delay(struct cached_dev *dc,
 274                                     unsigned int sectors)
 275 {
 276         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 277             !dc->writeback_percent)
 278                 return 0;
 279
 280         return bch_next_delay(&dc->writeback_rate, sectors);
 281 }
 282
 283 struct dirty_io {
 284         struct closure          cl;
 285         struct cached_dev       *dc;
 286         uint16_t                sequence;
 287         struct bio              bio;
 288 };
 289
 290 static void dirty_init(struct keybuf_key *w)
 291 {
 292         struct dirty_io *io = w->private;
 293         struct bio *bio = &io->bio;
 294
 295         bio_init(bio, NULL, bio->bi_inline_vecs,
 296                  DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS), 0);
 297         if (!io->dc->writeback_percent)
 298                 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 299
 300         bio->bi_iter.bi_size    = KEY_SIZE(&w->key) << 9;
 301         bio->bi_private         = w;
 302         bch_bio_map(bio, NULL);
 303 }
 304
 305 static void dirty_io_destructor(struct closure *cl)
 306 {
 307         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 308
 309         kfree(io);
 310 }
 311
 312 static void write_dirty_finish(struct closure *cl)
 313 {
 314         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 315         struct keybuf_key *w = io->bio.bi_private;
 316         struct cached_dev *dc = io->dc;
 317
 318         bio_free_pages(&io->bio);
 319
 320         /* This is kind of a dumb way of signalling errors. */
 321         if (KEY_DIRTY(&w->key)) {
 322                 int ret;
 323                 unsigned int i;
 324                 struct keylist keys;
 325
 326                 bch_keylist_init(&keys);
 327
 328                 bkey_copy(keys.top, &w->key);
 329                 SET_KEY_DIRTY(keys.top, false);
 330                 bch_keylist_push(&keys);
 331
 332                 for (i = 0; i < KEY_PTRS(&w->key); i++)
 333                         atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 334
 335                 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
 336
 337                 if (ret)
 338                         trace_bcache_writeback_collision(&w->key);
 339
 340                 atomic_long_inc(ret
 341                                 ? &dc->disk.c->writeback_keys_failed
 342                                 : &dc->disk.c->writeback_keys_done);
 343         }
 344
 345         bch_keybuf_del(&dc->writeback_keys, w);
 346         up(&dc->in_flight);
 347
 348         closure_return_with_destructor(cl, dirty_io_destructor);
 349 }
 350
 351 static void dirty_endio(struct bio *bio)
 352 {
 353         struct keybuf_key *w = bio->bi_private;
 354         struct dirty_io *io = w->private;
 355
 356         if (bio->bi_status) {
 357                 SET_KEY_DIRTY(&w->key, false);
 358                 bch_count_backing_io_errors(io->dc, bio);
 359         }
 360
 361         closure_put(&io->cl);
 362 }
 363
 364 static void write_dirty(struct closure *cl)
 365 {
 366         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 367         struct keybuf_key *w = io->bio.bi_private;
 368         struct cached_dev *dc = io->dc;
 369
 370         uint16_t next_sequence;
 371
 372         if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
 373                 /* Not our turn to write; wait for a write to complete */
 374                 closure_wait(&dc->writeback_ordering_wait, cl);
 375
 376                 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
 377                         /*
 378                          * Edge case-- it happened in indeterminate order
 379                          * relative to when we were added to wait list..
 380                          */
 381                         closure_wake_up(&dc->writeback_ordering_wait);
 382                 }
 383
 384                 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 385                 return;
 386         }
 387
 388         next_sequence = io->sequence + 1;
 389
 390         /*
 391          * IO errors are signalled using the dirty bit on the key.
 392          * If we failed to read, we should not attempt to write to the
 393          * backing device.  Instead, immediately go to write_dirty_finish
 394          * to clean up.
 395          */
 396         if (KEY_DIRTY(&w->key)) {
 397                 dirty_init(w);
 398                 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
 399                 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
 400                 bio_set_dev(&io->bio, io->dc->bdev);
 401                 io->bio.bi_end_io       = dirty_endio;
 402
 403                 /* I/O request sent to backing device */
 404                 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 405         }
 406
 407         atomic_set(&dc->writeback_sequence_next, next_sequence);
 408         closure_wake_up(&dc->writeback_ordering_wait);
 409
 410         continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
 411 }
 412
 413 static void read_dirty_endio(struct bio *bio)
 414 {
 415         struct keybuf_key *w = bio->bi_private;
 416         struct dirty_io *io = w->private;
 417
 418         /* is_read = 1 */
 419         bch_count_io_errors(io->dc->disk.c->cache,
 420                             bio->bi_status, 1,
 421                             "reading dirty data from cache");
 422
 423         dirty_endio(bio);
 424 }
 425
 426 static void read_dirty_submit(struct closure *cl)
 427 {
 428         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 429
 430         closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 431
 432         continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 433 }
 434
 435 static void read_dirty(struct cached_dev *dc)
 436 {
 437         unsigned int delay = 0;
 438         struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
 439         size_t size;
 440         int nk, i;
 441         struct dirty_io *io;
 442         struct closure cl;
 443         uint16_t sequence = 0;
 444
 445         BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
 446         atomic_set(&dc->writeback_sequence_next, sequence);
 447         closure_init_stack(&cl);
 448
 449         /*
 450          * XXX: if we error, background writeback just spins. Should use some
 451          * mempools.
 452          */
 453
 454         next = bch_keybuf_next(&dc->writeback_keys);
 455
 456         while (!kthread_should_stop() &&
 457                !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 458                next) {
 459                 size = 0;
 460                 nk = 0;
 461
 462                 do {
 463                         BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
 464
 465                         /*
 466                          * Don't combine too many operations, even if they
 467                          * are all small.
 468                          */
 469                         if (nk >= MAX_WRITEBACKS_IN_PASS)
 470                                 break;
 471
 472                         /*
 473                          * If the current operation is very large, don't
 474                          * further combine operations.
 475                          */
 476                         if (size >= MAX_WRITESIZE_IN_PASS)
 477                                 break;
 478
 479                         /*
 480                          * Operations are only eligible to be combined
 481                          * if they are contiguous.
 482                          *
 483                          * TODO: add a heuristic willing to fire a
 484                          * certain amount of non-contiguous IO per pass,
 485                          * so that we can benefit from backing device
 486                          * command queueing.
 487                          */
 488                         if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
 489                                                 &START_KEY(&next->key)))
 490                                 break;
 491
 492                         size += KEY_SIZE(&next->key);
 493                         keys[nk++] = next;
 494                 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
 495
 496                 /* Now we have gathered a set of 1..5 keys to write back. */
 497                 for (i = 0; i < nk; i++) {
 498                         w = keys[i];
 499
 500                         io = kzalloc(struct_size(io, bio.bi_inline_vecs,
 501                                                 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
 502                                      GFP_KERNEL);
 503                         if (!io)
 504                                 goto err;
 505
 506                         w->private      = io;
 507                         io->dc          = dc;
 508                         io->sequence    = sequence++;
 509
 510                         dirty_init(w);
 511                         bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
 512                         io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
 513                         bio_set_dev(&io->bio, dc->disk.c->cache->bdev);
 514                         io->bio.bi_end_io       = read_dirty_endio;
 515
 516                         if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
 517                                 goto err_free;
 518
 519                         trace_bcache_writeback(&w->key);
 520
 521                         down(&dc->in_flight);
 522
 523                         /*
 524                          * We've acquired a semaphore for the maximum
 525                          * simultaneous number of writebacks; from here
 526                          * everything happens asynchronously.
 527                          */
 528                         closure_call(&io->cl, read_dirty_submit, NULL, &cl);
 529                 }
 530
 531                 delay = writeback_delay(dc, size);
 532
 533                 while (!kthread_should_stop() &&
 534                        !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 535                        delay) {
 536                         schedule_timeout_interruptible(delay);
 537                         delay = writeback_delay(dc, 0);
 538                 }
 539         }
 540
 541         if (0) {
 542 err_free:
 543                 kfree(w->private);
 544 err:
 545                 bch_keybuf_del(&dc->writeback_keys, w);
 546         }
 547
 548         /*
 549          * Wait for outstanding writeback IOs to finish (and keybuf slots to be
 550          * freed) before refilling again
 551          */
 552         closure_sync(&cl);
 553 }
 554
 555 /* Scan for dirty data */
 556
 557 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
 558                                   uint64_t offset, int nr_sectors)
 559 {
 560         struct bcache_device *d = c->devices[inode];
 561         unsigned int stripe_offset, sectors_dirty;
 562         int stripe;
 563
 564         if (!d)
 565                 return;
 566
 567         stripe = offset_to_stripe(d, offset);
 568         if (stripe < 0)
 569                 return;
 570
 571         if (UUID_FLASH_ONLY(&c->uuids[inode]))
 572                 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
 573
 574         stripe_offset = offset & (d->stripe_size - 1);
 575
 576         while (nr_sectors) {
 577                 int s = min_t(unsigned int, abs(nr_sectors),
 578                               d->stripe_size - stripe_offset);
 579
 580                 if (nr_sectors < 0)
 581                         s = -s;
 582
 583                 if (stripe >= d->nr_stripes)
 584                         return;
 585
 586                 sectors_dirty = atomic_add_return(s,
 587                                         d->stripe_sectors_dirty + stripe);
 588                 if (sectors_dirty == d->stripe_size) {
 589                         if (!test_bit(stripe, d->full_dirty_stripes))
 590                                 set_bit(stripe, d->full_dirty_stripes);
 591                 } else {
 592                         if (test_bit(stripe, d->full_dirty_stripes))
 593                                 clear_bit(stripe, d->full_dirty_stripes);
 594                 }
 595
 596                 nr_sectors -= s;
 597                 stripe_offset = 0;
 598                 stripe++;
 599         }
 600 }
 601
 602 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 603 {
 604         struct cached_dev *dc = container_of(buf,
 605                                              struct cached_dev,
 606                                              writeback_keys);
 607
 608         BUG_ON(KEY_INODE(k) != dc->disk.id);
 609
 610         return KEY_DIRTY(k);
 611 }
 612
 613 static void refill_full_stripes(struct cached_dev *dc)
 614 {
 615         struct keybuf *buf = &dc->writeback_keys;
 616         unsigned int start_stripe, next_stripe;
 617         int stripe;
 618         bool wrapped = false;
 619
 620         stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
 621         if (stripe < 0)
 622                 stripe = 0;
 623
 624         start_stripe = stripe;
 625
 626         while (1) {
 627                 stripe = find_next_bit(dc->disk.full_dirty_stripes,
 628                                        dc->disk.nr_stripes, stripe);
 629
 630                 if (stripe == dc->disk.nr_stripes)
 631                         goto next;
 632
 633                 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
 634                                                  dc->disk.nr_stripes, stripe);
 635
 636                 buf->last_scanned = KEY(dc->disk.id,
 637                                         stripe * dc->disk.stripe_size, 0);
 638
 639                 bch_refill_keybuf(dc->disk.c, buf,
 640                                   &KEY(dc->disk.id,
 641                                        next_stripe * dc->disk.stripe_size, 0),
 642                                   dirty_pred);
 643
 644                 if (array_freelist_empty(&buf->freelist))
 645                         return;
 646
 647                 stripe = next_stripe;
 648 next:
 649                 if (wrapped && stripe > start_stripe)
 650                         return;
 651
 652                 if (stripe == dc->disk.nr_stripes) {
 653                         stripe = 0;
 654                         wrapped = true;
 655                 }
 656         }
 657 }
 658
 659 /*
 660  * Returns true if we scanned the entire disk
 661  */
 662 static bool refill_dirty(struct cached_dev *dc)
 663 {
 664         struct keybuf *buf = &dc->writeback_keys;
 665         struct bkey start = KEY(dc->disk.id, 0, 0);
 666         struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 667         struct bkey start_pos;
 668
 669         /*
 670          * make sure keybuf pos is inside the range for this disk - at bringup
 671          * we might not be attached yet so this disk's inode nr isn't
 672          * initialized then
 673          */
 674         if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
 675             bkey_cmp(&buf->last_scanned, &end) > 0)
 676                 buf->last_scanned = start;
 677
 678         if (dc->partial_stripes_expensive) {
 679                 refill_full_stripes(dc);
 680                 if (array_freelist_empty(&buf->freelist))
 681                         return false;
 682         }
 683
 684         start_pos = buf->last_scanned;
 685         bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 686
 687         if (bkey_cmp(&buf->last_scanned, &end) < 0)
 688                 return false;
 689
 690         /*
 691          * If we get to the end start scanning again from the beginning, and
 692          * only scan up to where we initially started scanning from:
 693          */
 694         buf->last_scanned = start;
 695         bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
 696
 697         return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 698 }
 699
 700 static int bch_writeback_thread(void *arg)
 701 {
 702         struct cached_dev *dc = arg;
 703         struct cache_set *c = dc->disk.c;
 704         bool searched_full_index;
 705
 706         bch_ratelimit_reset(&dc->writeback_rate);
 707
 708         while (!kthread_should_stop() &&
 709                !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 710                 down_write(&dc->writeback_lock);
 711                 set_current_state(TASK_INTERRUPTIBLE);
 712                 /*
 713                  * If the bache device is detaching, skip here and continue
 714                  * to perform writeback. Otherwise, if no dirty data on cache,
 715                  * or there is dirty data on cache but writeback is disabled,
 716                  * the writeback thread should sleep here and wait for others
 717                  * to wake up it.
 718                  */
 719                 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
 720                     (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
 721                         up_write(&dc->writeback_lock);
 722
 723                         if (kthread_should_stop() ||
 724                             test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 725                                 set_current_state(TASK_RUNNING);
 726                                 break;
 727                         }
 728
 729                         schedule();
 730                         continue;
 731                 }
 732                 set_current_state(TASK_RUNNING);
 733
 734                 searched_full_index = refill_dirty(dc);
 735
 736                 if (searched_full_index &&
 737                     RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
 738                         atomic_set(&dc->has_dirty, 0);
 739                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 740                         bch_write_bdev_super(dc, NULL);
 741                         /*
 742                          * If bcache device is detaching via sysfs interface,
 743                          * writeback thread should stop after there is no dirty
 744                          * data on cache. BCACHE_DEV_DETACHING flag is set in
 745                          * bch_cached_dev_detach().
 746                          */
 747                         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
 748                                 struct closure cl;
 749
 750                                 closure_init_stack(&cl);
 751                                 memset(&dc->sb.set_uuid, 0, 16);
 752                                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
 753
 754                                 bch_write_bdev_super(dc, &cl);
 755                                 closure_sync(&cl);
 756
 757                                 up_write(&dc->writeback_lock);
 758                                 break;
 759                         }
 760
 761                         /*
 762                          * When dirty data rate is high (e.g. 50%+), there might
 763                          * be heavy buckets fragmentation after writeback
 764                          * finished, which hurts following write performance.
 765                          * If users really care about write performance they
 766                          * may set BCH_ENABLE_AUTO_GC via sysfs, then when
 767                          * BCH_DO_AUTO_GC is set, garbage collection thread
 768                          * will be wake up here. After moving gc, the shrunk
 769                          * btree and discarded free buckets SSD space may be
 770                          * helpful for following write requests.
 771                          */
 772                         if (c->gc_after_writeback ==
 773                             (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
 774                                 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
 775                                 force_wake_up_gc(c);
 776                         }
 777                 }
 778
 779                 up_write(&dc->writeback_lock);
 780
 781                 read_dirty(dc);
 782
 783                 if (searched_full_index) {
 784                         unsigned int delay = dc->writeback_delay * HZ;
 785
 786                         while (delay &&
 787                                !kthread_should_stop() &&
 788                                !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
 789                                !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
 790                                 delay = schedule_timeout_interruptible(delay);
 791
 792                         bch_ratelimit_reset(&dc->writeback_rate);
 793                 }
 794         }
 795
 796         if (dc->writeback_write_wq) {
 797                 flush_workqueue(dc->writeback_write_wq);
 798                 destroy_workqueue(dc->writeback_write_wq);
 799         }
 800         cached_dev_put(dc);
 801         wait_for_kthread_stop();
 802
 803         return 0;
 804 }
 805
 806 /* Init */
 807 #define INIT_KEYS_EACH_TIME     500000
 808 #define INIT_KEYS_SLEEP_MS      100
 809
 810 struct sectors_dirty_init {
 811         struct btree_op op;
 812         unsigned int    inode;
 813         size_t          count;
 814         struct bkey     start;
 815 };
 816
 817 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
 818                                  struct bkey *k)
 819 {
 820         struct sectors_dirty_init *op = container_of(_op,
 821                                                 struct sectors_dirty_init, op);
 822         if (KEY_INODE(k) > op->inode)
 823                 return MAP_DONE;
 824
 825         if (KEY_DIRTY(k))
 826                 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
 827                                              KEY_START(k), KEY_SIZE(k));
 828
 829         op->count++;
 830         if (atomic_read(&b->c->search_inflight) &&
 831             !(op->count % INIT_KEYS_EACH_TIME)) {
 832                 bkey_copy_key(&op->start, k);
 833                 return -EAGAIN;
 834         }
 835
 836         return MAP_CONTINUE;
 837 }
 838
 839 static int bch_root_node_dirty_init(struct cache_set *c,
 840                                      struct bcache_device *d,
 841                                      struct bkey *k)
 842 {
 843         struct sectors_dirty_init op;
 844         int ret;
 845
 846         bch_btree_op_init(&op.op, -1);
 847         op.inode = d->id;
 848         op.count = 0;
 849         op.start = KEY(op.inode, 0, 0);
 850
 851         do {
 852                 ret = bcache_btree(map_keys_recurse,
 853                                    k,
 854                                    c->root,
 855                                    &op.op,
 856                                    &op.start,
 857                                    sectors_dirty_init_fn,
 858                                    0);
 859                 if (ret == -EAGAIN)
 860                         schedule_timeout_interruptible(
 861                                 msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
 862                 else if (ret < 0) {
 863                         pr_warn("sectors dirty init failed, ret=%d!\n", ret);
 864                         break;
 865                 }
 866         } while (ret == -EAGAIN);
 867
 868         return ret;
 869 }
 870
 871 static int bch_dirty_init_thread(void *arg)
 872 {
 873         struct dirty_init_thrd_info *info = arg;
 874         struct bch_dirty_init_state *state = info->state;
 875         struct cache_set *c = state->c;
 876         struct btree_iter iter;
 877         struct bkey *k, *p;
 878         int cur_idx, prev_idx, skip_nr;
 879
 880         k = p = NULL;
 881         cur_idx = prev_idx = 0;
 882
 883         bch_btree_iter_init(&c->root->keys, &iter, NULL);
 884         k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
 885         BUG_ON(!k);
 886
 887         p = k;
 888
 889         while (k) {
 890                 spin_lock(&state->idx_lock);
 891                 cur_idx = state->key_idx;
 892                 state->key_idx++;
 893                 spin_unlock(&state->idx_lock);
 894
 895                 skip_nr = cur_idx - prev_idx;
 896
 897                 while (skip_nr) {
 898                         k = bch_btree_iter_next_filter(&iter,
 899                                                        &c->root->keys,
 900                                                        bch_ptr_bad);
 901                         if (k)
 902                                 p = k;
 903                         else {
 904                                 atomic_set(&state->enough, 1);
 905                                 /* Update state->enough earlier */
 906                                 smp_mb__after_atomic();
 907                                 goto out;
 908                         }
 909                         skip_nr--;
 910                         cond_resched();
 911                 }
 912
 913                 if (p) {
 914                         if (bch_root_node_dirty_init(c, state->d, p) < 0)
 915                                 goto out;
 916                 }
 917
 918                 p = NULL;
 919                 prev_idx = cur_idx;
 920                 cond_resched();
 921         }
 922
 923 out:
 924         /* In order to wake up state->wait in time */
 925         smp_mb__before_atomic();
 926         if (atomic_dec_and_test(&state->started))
 927                 wake_up(&state->wait);
 928
 929         return 0;
 930 }
 931
 932 static int bch_btre_dirty_init_thread_nr(void)
 933 {
 934         int n = num_online_cpus()/2;
 935
 936         if (n == 0)
 937                 n = 1;
 938         else if (n > BCH_DIRTY_INIT_THRD_MAX)
 939                 n = BCH_DIRTY_INIT_THRD_MAX;
 940
 941         return n;
 942 }
 943
 944 void bch_sectors_dirty_init(struct bcache_device *d)
 945 {
 946         int i;
 947         struct bkey *k = NULL;
 948         struct btree_iter iter;
 949         struct sectors_dirty_init op;
 950         struct cache_set *c = d->c;
 951         struct bch_dirty_init_state *state;
 952         char name[32];
 953
 954         /* Just count root keys if no leaf node */
 955         if (c->root->level == 0) {
 956                 bch_btree_op_init(&op.op, -1);
 957                 op.inode = d->id;
 958                 op.count = 0;
 959                 op.start = KEY(op.inode, 0, 0);
 960
 961                 for_each_key_filter(&c->root->keys,
 962                                     k, &iter, bch_ptr_invalid)
 963                         sectors_dirty_init_fn(&op.op, c->root, k);
 964                 return;
 965         }
 966
 967         state = kzalloc(sizeof(struct bch_dirty_init_state), GFP_KERNEL);
 968         if (!state) {
 969                 pr_warn("sectors dirty init failed: cannot allocate memory\n");
 970                 return;
 971         }
 972
 973         state->c = c;
 974         state->d = d;
 975         state->total_threads = bch_btre_dirty_init_thread_nr();
 976         state->key_idx = 0;
 977         spin_lock_init(&state->idx_lock);
 978         atomic_set(&state->started, 0);
 979         atomic_set(&state->enough, 0);
 980         init_waitqueue_head(&state->wait);
 981
 982         for (i = 0; i < state->total_threads; i++) {
 983                 /* Fetch latest state->enough earlier */
 984                 smp_mb__before_atomic();
 985                 if (atomic_read(&state->enough))
 986                         break;
 987
 988                 state->infos[i].state = state;
 989                 atomic_inc(&state->started);
 990                 snprintf(name, sizeof(name), "bch_dirty_init[%d]", i);
 991
 992                 state->infos[i].thread =
 993                         kthread_run(bch_dirty_init_thread,
 994                                     &state->infos[i],
 995                                     name);
 996                 if (IS_ERR(state->infos[i].thread)) {
 997                         pr_err("fails to run thread bch_dirty_init[%d]\n", i);
 998                         for (--i; i >= 0; i--)
 999                                 kthread_stop(state->infos[i].thread);
1000                         goto out;
1001                 }
1002         }
1003
1004         /*
1005          * Must wait for all threads to stop.
1006          */
1007         wait_event_interruptible(state->wait,
1008                  atomic_read(&state->started) == 0);
1009
1010 out:
1011         kfree(state);
1012 }
1013
1014 void bch_cached_dev_writeback_init(struct cached_dev *dc)
1015 {
1016         sema_init(&dc->in_flight, 64);
1017         init_rwsem(&dc->writeback_lock);
1018         bch_keybuf_init(&dc->writeback_keys);
1019
1020         dc->writeback_metadata          = true;
1021         dc->writeback_running           = false;
1022         dc->writeback_consider_fragment = true;
1023         dc->writeback_percent           = 10;
1024         dc->writeback_delay             = 30;
1025         atomic_long_set(&dc->writeback_rate.rate, 1024);
1026         dc->writeback_rate_minimum      = 8;
1027
1028         dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1029         dc->writeback_rate_p_term_inverse = 40;
1030         dc->writeback_rate_fp_term_low = 1;
1031         dc->writeback_rate_fp_term_mid = 10;
1032         dc->writeback_rate_fp_term_high = 1000;
1033         dc->writeback_rate_i_term_inverse = 10000;
1034
1035         WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1036         INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1037 }
1038
1039 int bch_cached_dev_writeback_start(struct cached_dev *dc)
1040 {
1041         dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1042                                                 WQ_MEM_RECLAIM, 0);
1043         if (!dc->writeback_write_wq)
1044                 return -ENOMEM;
1045
1046         cached_dev_get(dc);
1047         dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1048                                               "bcache_writeback");
1049         if (IS_ERR(dc->writeback_thread)) {
1050                 cached_dev_put(dc);
1051                 destroy_workqueue(dc->writeback_write_wq);
1052                 return PTR_ERR(dc->writeback_thread);
1053         }
1054         dc->writeback_running = true;
1055
1056         WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1057         schedule_delayed_work(&dc->writeback_rate_update,
1058                               dc->writeback_rate_update_seconds * HZ);
1059
1060         bch_writeback_queue(dc);
1061
1062         return 0;
1063 }