drivers/md/bcache/writeback.c

   1 /*
   2  * background writeback - scan btree for dirty data and write it to the backing
   3  * device
   4  *
   5  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcache.h"
  10 #include "btree.h"
  11 #include "debug.h"
  12 #include "writeback.h"
  13
  14 #include <linux/delay.h>
  15 #include <linux/freezer.h>
  16 #include <linux/kthread.h>
  17 #include <trace/events/bcache.h>
  18
  19 /* Rate limiting */
  20
  21 static void __update_writeback_rate(struct cached_dev *dc)
  22 {
  23         struct cache_set *c = dc->disk.c;
  24         uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size;
  25         uint64_t cache_dirty_target =
  26                 div_u64(cache_sectors * dc->writeback_percent, 100);
  27
  28         int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev),
  29                                    c->cached_dev_sectors);
  30
  31         /* PD controller */
  32
  33         int change = 0;
  34         int64_t error;
  35         int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
  36         int64_t derivative = dirty - dc->disk.sectors_dirty_last;
  37
  38         dc->disk.sectors_dirty_last = dirty;
  39
  40         derivative *= dc->writeback_rate_d_term;
  41         derivative = clamp(derivative, -dirty, dirty);
  42
  43         derivative = ewma_add(dc->disk.sectors_dirty_derivative, derivative,
  44                               dc->writeback_rate_d_smooth, 0);
  45
  46         /* Avoid divide by zero */
  47         if (!target)
  48                 goto out;
  49
  50         error = div64_s64((dirty + derivative - target) << 8, target);
  51
  52         change = div_s64((dc->writeback_rate.rate * error) >> 8,
  53                          dc->writeback_rate_p_term_inverse);
  54
  55         /* Don't increase writeback rate if the device isn't keeping up */
  56         if (change > 0 &&
  57             time_after64(local_clock(),
  58                          dc->writeback_rate.next + 10 * NSEC_PER_MSEC))
  59                 change = 0;
  60
  61         dc->writeback_rate.rate =
  62                 clamp_t(int64_t, dc->writeback_rate.rate + change,
  63                         1, NSEC_PER_MSEC);
  64 out:
  65         dc->writeback_rate_derivative = derivative;
  66         dc->writeback_rate_change = change;
  67         dc->writeback_rate_target = target;
  68 }
  69
  70 static void update_writeback_rate(struct work_struct *work)
  71 {
  72         struct cached_dev *dc = container_of(to_delayed_work(work),
  73                                              struct cached_dev,
  74                                              writeback_rate_update);
  75
  76         down_read(&dc->writeback_lock);
  77
  78         if (atomic_read(&dc->has_dirty) &&
  79             dc->writeback_percent)
  80                 __update_writeback_rate(dc);
  81
  82         up_read(&dc->writeback_lock);
  83
  84         schedule_delayed_work(&dc->writeback_rate_update,
  85                               dc->writeback_rate_update_seconds * HZ);
  86 }
  87
  88 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
  89 {
  90         uint64_t ret;
  91
  92         if (atomic_read(&dc->disk.detaching) ||
  93             !dc->writeback_percent)
  94                 return 0;
  95
  96         ret = bch_next_delay(&dc->writeback_rate, sectors * 10000000ULL);
  97
  98         return min_t(uint64_t, ret, HZ);
  99 }
 100
 101 struct dirty_io {
 102         struct closure          cl;
 103         struct cached_dev       *dc;
 104         struct bio              bio;
 105 };
 106
 107 static void dirty_init(struct keybuf_key *w)
 108 {
 109         struct dirty_io *io = w->private;
 110         struct bio *bio = &io->bio;
 111
 112         bio_init(bio);
 113         if (!io->dc->writeback_percent)
 114                 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 115
 116         bio->bi_size            = KEY_SIZE(&w->key) << 9;
 117         bio->bi_max_vecs        = DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS);
 118         bio->bi_private         = w;
 119         bio->bi_io_vec          = bio->bi_inline_vecs;
 120         bch_bio_map(bio, NULL);
 121 }
 122
 123 static void dirty_io_destructor(struct closure *cl)
 124 {
 125         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 126         kfree(io);
 127 }
 128
 129 static void write_dirty_finish(struct closure *cl)
 130 {
 131         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 132         struct keybuf_key *w = io->bio.bi_private;
 133         struct cached_dev *dc = io->dc;
 134         struct bio_vec *bv;
 135         int i;
 136
 137         bio_for_each_segment_all(bv, &io->bio, i)
 138                 __free_page(bv->bv_page);
 139
 140         /* This is kind of a dumb way of signalling errors. */
 141         if (KEY_DIRTY(&w->key)) {
 142                 unsigned i;
 143                 struct btree_op op;
 144                 struct keylist keys;
 145
 146                 bch_btree_op_init_stack(&op);
 147                 bch_keylist_init(&keys);
 148
 149                 op.type = BTREE_REPLACE;
 150                 bkey_copy(&op.replace, &w->key);
 151
 152                 SET_KEY_DIRTY(&w->key, false);
 153                 bch_keylist_add(&keys, &w->key);
 154
 155                 for (i = 0; i < KEY_PTRS(&w->key); i++)
 156                         atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 157
 158                 bch_btree_insert(&op, dc->disk.c, &keys);
 159                 closure_sync(&op.cl);
 160
 161                 if (op.insert_collision)
 162                         trace_bcache_writeback_collision(&w->key);
 163
 164                 atomic_long_inc(op.insert_collision
 165                                 ? &dc->disk.c->writeback_keys_failed
 166                                 : &dc->disk.c->writeback_keys_done);
 167         }
 168
 169         bch_keybuf_del(&dc->writeback_keys, w);
 170         up(&dc->in_flight);
 171
 172         closure_return_with_destructor(cl, dirty_io_destructor);
 173 }
 174
 175 static void dirty_endio(struct bio *bio, int error)
 176 {
 177         struct keybuf_key *w = bio->bi_private;
 178         struct dirty_io *io = w->private;
 179
 180         if (error)
 181                 SET_KEY_DIRTY(&w->key, false);
 182
 183         closure_put(&io->cl);
 184 }
 185
 186 static void write_dirty(struct closure *cl)
 187 {
 188         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 189         struct keybuf_key *w = io->bio.bi_private;
 190
 191         dirty_init(w);
 192         io->bio.bi_rw           = WRITE;
 193         io->bio.bi_sector       = KEY_START(&w->key);
 194         io->bio.bi_bdev         = io->dc->bdev;
 195         io->bio.bi_end_io       = dirty_endio;
 196
 197         closure_bio_submit(&io->bio, cl, &io->dc->disk);
 198
 199         continue_at(cl, write_dirty_finish, system_wq);
 200 }
 201
 202 static void read_dirty_endio(struct bio *bio, int error)
 203 {
 204         struct keybuf_key *w = bio->bi_private;
 205         struct dirty_io *io = w->private;
 206
 207         bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
 208                             error, "reading dirty data from cache");
 209
 210         dirty_endio(bio, error);
 211 }
 212
 213 static void read_dirty_submit(struct closure *cl)
 214 {
 215         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 216
 217         closure_bio_submit(&io->bio, cl, &io->dc->disk);
 218
 219         continue_at(cl, write_dirty, system_wq);
 220 }
 221
 222 static void read_dirty(struct cached_dev *dc)
 223 {
 224         unsigned delay = 0;
 225         struct keybuf_key *w;
 226         struct dirty_io *io;
 227         struct closure cl;
 228
 229         closure_init_stack(&cl);
 230
 231         /*
 232          * XXX: if we error, background writeback just spins. Should use some
 233          * mempools.
 234          */
 235
 236         while (!kthread_should_stop()) {
 237                 try_to_freeze();
 238
 239                 w = bch_keybuf_next(&dc->writeback_keys);
 240                 if (!w)
 241                         break;
 242
 243                 BUG_ON(ptr_stale(dc->disk.c, &w->key, 0));
 244
 245                 if (KEY_START(&w->key) != dc->last_read ||
 246                     jiffies_to_msecs(delay) > 50)
 247                         while (!kthread_should_stop() && delay)
 248                                 delay = schedule_timeout_interruptible(delay);
 249
 250                 dc->last_read   = KEY_OFFSET(&w->key);
 251
 252                 io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec)
 253                              * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
 254                              GFP_KERNEL);
 255                 if (!io)
 256                         goto err;
 257
 258                 w->private      = io;
 259                 io->dc          = dc;
 260
 261                 dirty_init(w);
 262                 io->bio.bi_sector       = PTR_OFFSET(&w->key, 0);
 263                 io->bio.bi_bdev         = PTR_CACHE(dc->disk.c,
 264                                                     &w->key, 0)->bdev;
 265                 io->bio.bi_rw           = READ;
 266                 io->bio.bi_end_io       = read_dirty_endio;
 267
 268                 if (bio_alloc_pages(&io->bio, GFP_KERNEL))
 269                         goto err_free;
 270
 271                 trace_bcache_writeback(&w->key);
 272
 273                 down(&dc->in_flight);
 274                 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
 275
 276                 delay = writeback_delay(dc, KEY_SIZE(&w->key));
 277         }
 278
 279         if (0) {
 280 err_free:
 281                 kfree(w->private);
 282 err:
 283                 bch_keybuf_del(&dc->writeback_keys, w);
 284         }
 285
 286         /*
 287          * Wait for outstanding writeback IOs to finish (and keybuf slots to be
 288          * freed) before refilling again
 289          */
 290         closure_sync(&cl);
 291 }
 292
 293 /* Scan for dirty data */
 294
 295 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
 296                                   uint64_t offset, int nr_sectors)
 297 {
 298         struct bcache_device *d = c->devices[inode];
 299         unsigned stripe_offset;
 300         uint64_t stripe = offset;
 301
 302         if (!d)
 303                 return;
 304
 305         do_div(stripe, d->stripe_size);
 306
 307         stripe_offset = offset & (d->stripe_size - 1);
 308
 309         while (nr_sectors) {
 310                 int s = min_t(unsigned, abs(nr_sectors),
 311                               d->stripe_size - stripe_offset);
 312
 313                 if (nr_sectors < 0)
 314                         s = -s;
 315
 316                 atomic_add(s, d->stripe_sectors_dirty + stripe);
 317                 nr_sectors -= s;
 318                 stripe_offset = 0;
 319                 stripe++;
 320         }
 321 }
 322
 323 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 324 {
 325         return KEY_DIRTY(k);
 326 }
 327
 328 static bool dirty_full_stripe_pred(struct keybuf *buf, struct bkey *k)
 329 {
 330         uint64_t stripe = KEY_START(k);
 331         unsigned nr_sectors = KEY_SIZE(k);
 332         struct cached_dev *dc = container_of(buf, struct cached_dev,
 333                                              writeback_keys);
 334
 335         if (!KEY_DIRTY(k))
 336                 return false;
 337
 338         do_div(stripe, dc->disk.stripe_size);
 339
 340         while (1) {
 341                 if (atomic_read(dc->disk.stripe_sectors_dirty + stripe) ==
 342                     dc->disk.stripe_size)
 343                         return true;
 344
 345                 if (nr_sectors <= dc->disk.stripe_size)
 346                         return false;
 347
 348                 nr_sectors -= dc->disk.stripe_size;
 349                 stripe++;
 350         }
 351 }
 352
 353 static bool refill_dirty(struct cached_dev *dc)
 354 {
 355         struct keybuf *buf = &dc->writeback_keys;
 356         bool searched_from_start = false;
 357         struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 358
 359         if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
 360                 buf->last_scanned = KEY(dc->disk.id, 0, 0);
 361                 searched_from_start = true;
 362         }
 363
 364         if (dc->partial_stripes_expensive) {
 365                 uint64_t i;
 366
 367                 for (i = 0; i < dc->disk.nr_stripes; i++)
 368                         if (atomic_read(dc->disk.stripe_sectors_dirty + i) ==
 369                             dc->disk.stripe_size)
 370                                 goto full_stripes;
 371
 372                 goto normal_refill;
 373 full_stripes:
 374                 searched_from_start = false;    /* not searching entire btree */
 375                 bch_refill_keybuf(dc->disk.c, buf, &end,
 376                                   dirty_full_stripe_pred);
 377         } else {
 378 normal_refill:
 379                 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 380         }
 381
 382         return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
 383 }
 384
 385 static int bch_writeback_thread(void *arg)
 386 {
 387         struct cached_dev *dc = arg;
 388         bool searched_full_index;
 389
 390         while (!kthread_should_stop()) {
 391                 down_write(&dc->writeback_lock);
 392                 if (!atomic_read(&dc->has_dirty) ||
 393                     (!atomic_read(&dc->disk.detaching) &&
 394                      !dc->writeback_running)) {
 395                         up_write(&dc->writeback_lock);
 396                         set_current_state(TASK_INTERRUPTIBLE);
 397
 398                         if (kthread_should_stop())
 399                                 return 0;
 400
 401                         try_to_freeze();
 402                         schedule();
 403                         continue;
 404                 }
 405
 406                 searched_full_index = refill_dirty(dc);
 407
 408                 if (searched_full_index &&
 409                     RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
 410                         atomic_set(&dc->has_dirty, 0);
 411                         cached_dev_put(dc);
 412                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 413                         bch_write_bdev_super(dc, NULL);
 414                 }
 415
 416                 up_write(&dc->writeback_lock);
 417
 418                 bch_ratelimit_reset(&dc->writeback_rate);
 419                 read_dirty(dc);
 420
 421                 if (searched_full_index) {
 422                         unsigned delay = dc->writeback_delay * HZ;
 423
 424                         while (delay &&
 425                                !kthread_should_stop() &&
 426                                !atomic_read(&dc->disk.detaching))
 427                                 delay = schedule_timeout_interruptible(delay);
 428                 }
 429         }
 430
 431         return 0;
 432 }
 433
 434 /* Init */
 435
 436 static int bch_btree_sectors_dirty_init(struct btree *b, struct btree_op *op,
 437                                         struct cached_dev *dc)
 438 {
 439         struct bkey *k;
 440         struct btree_iter iter;
 441
 442         bch_btree_iter_init(b, &iter, &KEY(dc->disk.id, 0, 0));
 443         while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad)))
 444                 if (!b->level) {
 445                         if (KEY_INODE(k) > dc->disk.id)
 446                                 break;
 447
 448                         if (KEY_DIRTY(k))
 449                                 bcache_dev_sectors_dirty_add(b->c, dc->disk.id,
 450                                                              KEY_START(k),
 451                                                              KEY_SIZE(k));
 452                 } else {
 453                         btree(sectors_dirty_init, k, b, op, dc);
 454                         if (KEY_INODE(k) > dc->disk.id)
 455                                 break;
 456
 457                         cond_resched();
 458                 }
 459
 460         return 0;
 461 }
 462
 463 void bch_sectors_dirty_init(struct cached_dev *dc)
 464 {
 465         struct btree_op op;
 466
 467         bch_btree_op_init_stack(&op);
 468         btree_root(sectors_dirty_init, dc->disk.c, &op, dc);
 469 }
 470
 471 int bch_cached_dev_writeback_init(struct cached_dev *dc)
 472 {
 473         sema_init(&dc->in_flight, 64);
 474         init_rwsem(&dc->writeback_lock);
 475         bch_keybuf_init(&dc->writeback_keys);
 476
 477         dc->writeback_metadata          = true;
 478         dc->writeback_running           = true;
 479         dc->writeback_percent           = 10;
 480         dc->writeback_delay             = 30;
 481         dc->writeback_rate.rate         = 1024;
 482
 483         dc->writeback_rate_update_seconds = 30;
 484         dc->writeback_rate_d_term       = 16;
 485         dc->writeback_rate_p_term_inverse = 64;
 486         dc->writeback_rate_d_smooth     = 8;
 487
 488         dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
 489                                               "bcache_writeback");
 490         if (IS_ERR(dc->writeback_thread))
 491                 return PTR_ERR(dc->writeback_thread);
 492
 493         set_task_state(dc->writeback_thread, TASK_INTERRUPTIBLE);
 494
 495         INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
 496         schedule_delayed_work(&dc->writeback_rate_update,
 497                               dc->writeback_rate_update_seconds * HZ);
 498
 499         return 0;
 500 }