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
  13 static struct workqueue_struct *dirty_wq;
  14
  15 static void read_dirty(struct closure *);
  16
  17 struct dirty_io {
  18         struct closure          cl;
  19         struct cached_dev       *dc;
  20         struct bio              bio;
  21 };
  22
  23 /* Rate limiting */
  24
  25 static void __update_writeback_rate(struct cached_dev *dc)
  26 {
  27         struct cache_set *c = dc->disk.c;
  28         uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size;
  29         uint64_t cache_dirty_target =
  30                 div_u64(cache_sectors * dc->writeback_percent, 100);
  31
  32         int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev),
  33                                    c->cached_dev_sectors);
  34
  35         /* PD controller */
  36
  37         int change = 0;
  38         int64_t error;
  39         int64_t dirty = atomic_long_read(&dc->disk.sectors_dirty);
  40         int64_t derivative = dirty - dc->disk.sectors_dirty_last;
  41
  42         dc->disk.sectors_dirty_last = dirty;
  43
  44         derivative *= dc->writeback_rate_d_term;
  45         derivative = clamp(derivative, -dirty, dirty);
  46
  47         derivative = ewma_add(dc->disk.sectors_dirty_derivative, derivative,
  48                               dc->writeback_rate_d_smooth, 0);
  49
  50         /* Avoid divide by zero */
  51         if (!target)
  52                 goto out;
  53
  54         error = div64_s64((dirty + derivative - target) << 8, target);
  55
  56         change = div_s64((dc->writeback_rate.rate * error) >> 8,
  57                          dc->writeback_rate_p_term_inverse);
  58
  59         /* Don't increase writeback rate if the device isn't keeping up */
  60         if (change > 0 &&
  61             time_after64(local_clock(),
  62                          dc->writeback_rate.next + 10 * NSEC_PER_MSEC))
  63                 change = 0;
  64
  65         dc->writeback_rate.rate =
  66                 clamp_t(int64_t, dc->writeback_rate.rate + change,
  67                         1, NSEC_PER_MSEC);
  68 out:
  69         dc->writeback_rate_derivative = derivative;
  70         dc->writeback_rate_change = change;
  71         dc->writeback_rate_target = target;
  72
  73         schedule_delayed_work(&dc->writeback_rate_update,
  74                               dc->writeback_rate_update_seconds * HZ);
  75 }
  76
  77 static void update_writeback_rate(struct work_struct *work)
  78 {
  79         struct cached_dev *dc = container_of(to_delayed_work(work),
  80                                              struct cached_dev,
  81                                              writeback_rate_update);
  82
  83         down_read(&dc->writeback_lock);
  84
  85         if (atomic_read(&dc->has_dirty) &&
  86             dc->writeback_percent)
  87                 __update_writeback_rate(dc);
  88
  89         up_read(&dc->writeback_lock);
  90 }
  91
  92 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
  93 {
  94         if (atomic_read(&dc->disk.detaching) ||
  95             !dc->writeback_percent)
  96                 return 0;
  97
  98         return bch_next_delay(&dc->writeback_rate, sectors * 10000000ULL);
  99 }
 100
 101 /* Background writeback */
 102
 103 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 104 {
 105         return KEY_DIRTY(k);
 106 }
 107
 108 static void dirty_init(struct keybuf_key *w)
 109 {
 110         struct dirty_io *io = w->private;
 111         struct bio *bio = &io->bio;
 112
 113         bio_init(bio);
 114         if (!io->dc->writeback_percent)
 115                 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
 116
 117         bio->bi_size            = KEY_SIZE(&w->key) << 9;
 118         bio->bi_max_vecs        = DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS);
 119         bio->bi_private         = w;
 120         bio->bi_io_vec          = bio->bi_inline_vecs;
 121         bch_bio_map(bio, NULL);
 122 }
 123
 124 static void refill_dirty(struct closure *cl)
 125 {
 126         struct cached_dev *dc = container_of(cl, struct cached_dev,
 127                                              writeback.cl);
 128         struct keybuf *buf = &dc->writeback_keys;
 129         bool searched_from_start = false;
 130         struct bkey end = MAX_KEY;
 131         SET_KEY_INODE(&end, dc->disk.id);
 132
 133         if (!atomic_read(&dc->disk.detaching) &&
 134             !dc->writeback_running)
 135                 closure_return(cl);
 136
 137         down_write(&dc->writeback_lock);
 138
 139         if (!atomic_read(&dc->has_dirty)) {
 140                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 141                 bch_write_bdev_super(dc, NULL);
 142
 143                 up_write(&dc->writeback_lock);
 144                 closure_return(cl);
 145         }
 146
 147         if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
 148                 buf->last_scanned = KEY(dc->disk.id, 0, 0);
 149                 searched_from_start = true;
 150         }
 151
 152         bch_refill_keybuf(dc->disk.c, buf, &end);
 153
 154         if (bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start) {
 155                 /* Searched the entire btree  - delay awhile */
 156
 157                 if (RB_EMPTY_ROOT(&buf->keys)) {
 158                         atomic_set(&dc->has_dirty, 0);
 159                         cached_dev_put(dc);
 160                 }
 161
 162                 if (!atomic_read(&dc->disk.detaching))
 163                         closure_delay(&dc->writeback, dc->writeback_delay * HZ);
 164         }
 165
 166         up_write(&dc->writeback_lock);
 167
 168         ratelimit_reset(&dc->writeback_rate);
 169
 170         /* Punt to workqueue only so we don't recurse and blow the stack */
 171         continue_at(cl, read_dirty, dirty_wq);
 172 }
 173
 174 void bch_writeback_queue(struct cached_dev *dc)
 175 {
 176         if (closure_trylock(&dc->writeback.cl, &dc->disk.cl)) {
 177                 if (!atomic_read(&dc->disk.detaching))
 178                         closure_delay(&dc->writeback, dc->writeback_delay * HZ);
 179
 180                 continue_at(&dc->writeback.cl, refill_dirty, dirty_wq);
 181         }
 182 }
 183
 184 void bch_writeback_add(struct cached_dev *dc, unsigned sectors)
 185 {
 186         atomic_long_add(sectors, &dc->disk.sectors_dirty);
 187
 188         if (!atomic_read(&dc->has_dirty) &&
 189             !atomic_xchg(&dc->has_dirty, 1)) {
 190                 atomic_inc(&dc->count);
 191
 192                 if (BDEV_STATE(&dc->sb) != BDEV_STATE_DIRTY) {
 193                         SET_BDEV_STATE(&dc->sb, BDEV_STATE_DIRTY);
 194                         /* XXX: should do this synchronously */
 195                         bch_write_bdev_super(dc, NULL);
 196                 }
 197
 198                 bch_writeback_queue(dc);
 199
 200                 if (dc->writeback_percent)
 201                         schedule_delayed_work(&dc->writeback_rate_update,
 202                                       dc->writeback_rate_update_seconds * HZ);
 203         }
 204 }
 205
 206 /* Background writeback - IO loop */
 207
 208 static void dirty_io_destructor(struct closure *cl)
 209 {
 210         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 211         kfree(io);
 212 }
 213
 214 static void write_dirty_finish(struct closure *cl)
 215 {
 216         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 217         struct keybuf_key *w = io->bio.bi_private;
 218         struct cached_dev *dc = io->dc;
 219         struct bio_vec *bv = bio_iovec_idx(&io->bio, io->bio.bi_vcnt);
 220
 221         while (bv-- != io->bio.bi_io_vec)
 222                 __free_page(bv->bv_page);
 223
 224         /* This is kind of a dumb way of signalling errors. */
 225         if (KEY_DIRTY(&w->key)) {
 226                 unsigned i;
 227                 struct btree_op op;
 228                 bch_btree_op_init_stack(&op);
 229
 230                 op.type = BTREE_REPLACE;
 231                 bkey_copy(&op.replace, &w->key);
 232
 233                 SET_KEY_DIRTY(&w->key, false);
 234                 bch_keylist_add(&op.keys, &w->key);
 235
 236                 for (i = 0; i < KEY_PTRS(&w->key); i++)
 237                         atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 238
 239                 pr_debug("clearing %s", pkey(&w->key));
 240                 bch_btree_insert(&op, dc->disk.c);
 241                 closure_sync(&op.cl);
 242
 243                 atomic_long_inc(op.insert_collision
 244                                 ? &dc->disk.c->writeback_keys_failed
 245                                 : &dc->disk.c->writeback_keys_done);
 246         }
 247
 248         bch_keybuf_del(&dc->writeback_keys, w);
 249         atomic_dec_bug(&dc->in_flight);
 250
 251         closure_wake_up(&dc->writeback_wait);
 252
 253         closure_return_with_destructor(cl, dirty_io_destructor);
 254 }
 255
 256 static void dirty_endio(struct bio *bio, int error)
 257 {
 258         struct keybuf_key *w = bio->bi_private;
 259         struct dirty_io *io = w->private;
 260
 261         if (error)
 262                 SET_KEY_DIRTY(&w->key, false);
 263
 264         closure_put(&io->cl);
 265 }
 266
 267 static void write_dirty(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
 272         dirty_init(w);
 273         io->bio.bi_rw           = WRITE;
 274         io->bio.bi_sector       = KEY_START(&w->key);
 275         io->bio.bi_bdev         = io->dc->bdev;
 276         io->bio.bi_end_io       = dirty_endio;
 277
 278         trace_bcache_write_dirty(&io->bio);
 279         closure_bio_submit(&io->bio, cl, &io->dc->disk);
 280
 281         continue_at(cl, write_dirty_finish, dirty_wq);
 282 }
 283
 284 static void read_dirty_endio(struct bio *bio, int error)
 285 {
 286         struct keybuf_key *w = bio->bi_private;
 287         struct dirty_io *io = w->private;
 288
 289         bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
 290                             error, "reading dirty data from cache");
 291
 292         dirty_endio(bio, error);
 293 }
 294
 295 static void read_dirty_submit(struct closure *cl)
 296 {
 297         struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 298
 299         trace_bcache_read_dirty(&io->bio);
 300         closure_bio_submit(&io->bio, cl, &io->dc->disk);
 301
 302         continue_at(cl, write_dirty, dirty_wq);
 303 }
 304
 305 static void read_dirty(struct closure *cl)
 306 {
 307         struct cached_dev *dc = container_of(cl, struct cached_dev,
 308                                              writeback.cl);
 309         unsigned delay = writeback_delay(dc, 0);
 310         struct keybuf_key *w;
 311         struct dirty_io *io;
 312
 313         /*
 314          * XXX: if we error, background writeback just spins. Should use some
 315          * mempools.
 316          */
 317
 318         while (1) {
 319                 w = bch_keybuf_next(&dc->writeback_keys);
 320                 if (!w)
 321                         break;
 322
 323                 BUG_ON(ptr_stale(dc->disk.c, &w->key, 0));
 324
 325                 if (delay > 0 &&
 326                     (KEY_START(&w->key) != dc->last_read ||
 327                      jiffies_to_msecs(delay) > 50)) {
 328                         w->private = NULL;
 329
 330                         closure_delay(&dc->writeback, delay);
 331                         continue_at(cl, read_dirty, dirty_wq);
 332                 }
 333
 334                 dc->last_read   = KEY_OFFSET(&w->key);
 335
 336                 io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec)
 337                              * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
 338                              GFP_KERNEL);
 339                 if (!io)
 340                         goto err;
 341
 342                 w->private      = io;
 343                 io->dc          = dc;
 344
 345                 dirty_init(w);
 346                 io->bio.bi_sector       = PTR_OFFSET(&w->key, 0);
 347                 io->bio.bi_bdev         = PTR_CACHE(dc->disk.c,
 348                                                     &w->key, 0)->bdev;
 349                 io->bio.bi_rw           = READ;
 350                 io->bio.bi_end_io       = read_dirty_endio;
 351
 352                 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
 353                         goto err_free;
 354
 355                 pr_debug("%s", pkey(&w->key));
 356
 357                 closure_call(&io->cl, read_dirty_submit, NULL, &dc->disk.cl);
 358
 359                 delay = writeback_delay(dc, KEY_SIZE(&w->key));
 360
 361                 atomic_inc(&dc->in_flight);
 362
 363                 if (!closure_wait_event(&dc->writeback_wait, cl,
 364                                         atomic_read(&dc->in_flight) < 64))
 365                         continue_at(cl, read_dirty, dirty_wq);
 366         }
 367
 368         if (0) {
 369 err_free:
 370                 kfree(w->private);
 371 err:
 372                 bch_keybuf_del(&dc->writeback_keys, w);
 373         }
 374
 375         refill_dirty(cl);
 376 }
 377
 378 void bch_cached_dev_writeback_init(struct cached_dev *dc)
 379 {
 380         closure_init_unlocked(&dc->writeback);
 381         init_rwsem(&dc->writeback_lock);
 382
 383         bch_keybuf_init(&dc->writeback_keys, dirty_pred);
 384
 385         dc->writeback_metadata          = true;
 386         dc->writeback_running           = true;
 387         dc->writeback_percent           = 10;
 388         dc->writeback_delay             = 30;
 389         dc->writeback_rate.rate         = 1024;
 390
 391         dc->writeback_rate_update_seconds = 30;
 392         dc->writeback_rate_d_term       = 16;
 393         dc->writeback_rate_p_term_inverse = 64;
 394         dc->writeback_rate_d_smooth     = 8;
 395
 396         INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
 397         schedule_delayed_work(&dc->writeback_rate_update,
 398                               dc->writeback_rate_update_seconds * HZ);
 399 }
 400
 401 void bch_writeback_exit(void)
 402 {
 403         if (dirty_wq)
 404                 destroy_workqueue(dirty_wq);
 405 }
 406
 407 int __init bch_writeback_init(void)
 408 {
 409         dirty_wq = create_singlethread_workqueue("bcache_writeback");
 410         if (!dirty_wq)
 411                 return -ENOMEM;
 412
 413         return 0;
 414 }