drivers/md/dm-cache-policy-smq.c

   1 /*
   2  * Copyright (C) 2015 Red Hat. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-cache-background-tracker.h"
   8 #include "dm-cache-policy-internal.h"
   9 #include "dm-cache-policy.h"
  10 #include "dm.h"
  11
  12 #include <linux/hash.h>
  13 #include <linux/jiffies.h>
  14 #include <linux/module.h>
  15 #include <linux/mutex.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/math64.h>
  18
  19 #define DM_MSG_PREFIX "cache-policy-smq"
  20
  21 /*----------------------------------------------------------------*/
  22
  23 /*
  24  * Safe division functions that return zero on divide by zero.
  25  */
  26 static unsigned safe_div(unsigned n, unsigned d)
  27 {
  28         return d ? n / d : 0u;
  29 }
  30
  31 static unsigned safe_mod(unsigned n, unsigned d)
  32 {
  33         return d ? n % d : 0u;
  34 }
  35
  36 /*----------------------------------------------------------------*/
  37
  38 struct entry {
  39         unsigned hash_next:28;
  40         unsigned prev:28;
  41         unsigned next:28;
  42         unsigned level:6;
  43         bool dirty:1;
  44         bool allocated:1;
  45         bool sentinel:1;
  46         bool pending_work:1;
  47
  48         dm_oblock_t oblock;
  49 };
  50
  51 /*----------------------------------------------------------------*/
  52
  53 #define INDEXER_NULL ((1u << 28u) - 1u)
  54
  55 /*
  56  * An entry_space manages a set of entries that we use for the queues.
  57  * The clean and dirty queues share entries, so this object is separate
  58  * from the queue itself.
  59  */
  60 struct entry_space {
  61         struct entry *begin;
  62         struct entry *end;
  63 };
  64
  65 static int space_init(struct entry_space *es, unsigned nr_entries)
  66 {
  67         if (!nr_entries) {
  68                 es->begin = es->end = NULL;
  69                 return 0;
  70         }
  71
  72         es->begin = vzalloc(sizeof(struct entry) * nr_entries);
  73         if (!es->begin)
  74                 return -ENOMEM;
  75
  76         es->end = es->begin + nr_entries;
  77         return 0;
  78 }
  79
  80 static void space_exit(struct entry_space *es)
  81 {
  82         vfree(es->begin);
  83 }
  84
  85 static struct entry *__get_entry(struct entry_space *es, unsigned block)
  86 {
  87         struct entry *e;
  88
  89         e = es->begin + block;
  90         BUG_ON(e >= es->end);
  91
  92         return e;
  93 }
  94
  95 static unsigned to_index(struct entry_space *es, struct entry *e)
  96 {
  97         BUG_ON(e < es->begin || e >= es->end);
  98         return e - es->begin;
  99 }
 100
 101 static struct entry *to_entry(struct entry_space *es, unsigned block)
 102 {
 103         if (block == INDEXER_NULL)
 104                 return NULL;
 105
 106         return __get_entry(es, block);
 107 }
 108
 109 /*----------------------------------------------------------------*/
 110
 111 struct ilist {
 112         unsigned nr_elts;       /* excluding sentinel entries */
 113         unsigned head, tail;
 114 };
 115
 116 static void l_init(struct ilist *l)
 117 {
 118         l->nr_elts = 0;
 119         l->head = l->tail = INDEXER_NULL;
 120 }
 121
 122 static struct entry *l_head(struct entry_space *es, struct ilist *l)
 123 {
 124         return to_entry(es, l->head);
 125 }
 126
 127 static struct entry *l_tail(struct entry_space *es, struct ilist *l)
 128 {
 129         return to_entry(es, l->tail);
 130 }
 131
 132 static struct entry *l_next(struct entry_space *es, struct entry *e)
 133 {
 134         return to_entry(es, e->next);
 135 }
 136
 137 static struct entry *l_prev(struct entry_space *es, struct entry *e)
 138 {
 139         return to_entry(es, e->prev);
 140 }
 141
 142 static bool l_empty(struct ilist *l)
 143 {
 144         return l->head == INDEXER_NULL;
 145 }
 146
 147 static void l_add_head(struct entry_space *es, struct ilist *l, struct entry *e)
 148 {
 149         struct entry *head = l_head(es, l);
 150
 151         e->next = l->head;
 152         e->prev = INDEXER_NULL;
 153
 154         if (head)
 155                 head->prev = l->head = to_index(es, e);
 156         else
 157                 l->head = l->tail = to_index(es, e);
 158
 159         if (!e->sentinel)
 160                 l->nr_elts++;
 161 }
 162
 163 static void l_add_tail(struct entry_space *es, struct ilist *l, struct entry *e)
 164 {
 165         struct entry *tail = l_tail(es, l);
 166
 167         e->next = INDEXER_NULL;
 168         e->prev = l->tail;
 169
 170         if (tail)
 171                 tail->next = l->tail = to_index(es, e);
 172         else
 173                 l->head = l->tail = to_index(es, e);
 174
 175         if (!e->sentinel)
 176                 l->nr_elts++;
 177 }
 178
 179 static void l_add_before(struct entry_space *es, struct ilist *l,
 180                          struct entry *old, struct entry *e)
 181 {
 182         struct entry *prev = l_prev(es, old);
 183
 184         if (!prev)
 185                 l_add_head(es, l, e);
 186
 187         else {
 188                 e->prev = old->prev;
 189                 e->next = to_index(es, old);
 190                 prev->next = old->prev = to_index(es, e);
 191
 192                 if (!e->sentinel)
 193                         l->nr_elts++;
 194         }
 195 }
 196
 197 static void l_del(struct entry_space *es, struct ilist *l, struct entry *e)
 198 {
 199         struct entry *prev = l_prev(es, e);
 200         struct entry *next = l_next(es, e);
 201
 202         if (prev)
 203                 prev->next = e->next;
 204         else
 205                 l->head = e->next;
 206
 207         if (next)
 208                 next->prev = e->prev;
 209         else
 210                 l->tail = e->prev;
 211
 212         if (!e->sentinel)
 213                 l->nr_elts--;
 214 }
 215
 216 static struct entry *l_pop_tail(struct entry_space *es, struct ilist *l)
 217 {
 218         struct entry *e;
 219
 220         for (e = l_tail(es, l); e; e = l_prev(es, e))
 221                 if (!e->sentinel) {
 222                         l_del(es, l, e);
 223                         return e;
 224                 }
 225
 226         return NULL;
 227 }
 228
 229 /*----------------------------------------------------------------*/
 230
 231 /*
 232  * The stochastic-multi-queue is a set of lru lists stacked into levels.
 233  * Entries are moved up levels when they are used, which loosely orders the
 234  * most accessed entries in the top levels and least in the bottom.  This
 235  * structure is *much* better than a single lru list.
 236  */
 237 #define MAX_LEVELS 64u
 238
 239 struct queue {
 240         struct entry_space *es;
 241
 242         unsigned nr_elts;
 243         unsigned nr_levels;
 244         struct ilist qs[MAX_LEVELS];
 245
 246         /*
 247          * We maintain a count of the number of entries we would like in each
 248          * level.
 249          */
 250         unsigned last_target_nr_elts;
 251         unsigned nr_top_levels;
 252         unsigned nr_in_top_levels;
 253         unsigned target_count[MAX_LEVELS];
 254 };
 255
 256 static void q_init(struct queue *q, struct entry_space *es, unsigned nr_levels)
 257 {
 258         unsigned i;
 259
 260         q->es = es;
 261         q->nr_elts = 0;
 262         q->nr_levels = nr_levels;
 263
 264         for (i = 0; i < q->nr_levels; i++) {
 265                 l_init(q->qs + i);
 266                 q->target_count[i] = 0u;
 267         }
 268
 269         q->last_target_nr_elts = 0u;
 270         q->nr_top_levels = 0u;
 271         q->nr_in_top_levels = 0u;
 272 }
 273
 274 static unsigned q_size(struct queue *q)
 275 {
 276         return q->nr_elts;
 277 }
 278
 279 /*
 280  * Insert an entry to the back of the given level.
 281  */
 282 static void q_push(struct queue *q, struct entry *e)
 283 {
 284         BUG_ON(e->pending_work);
 285
 286         if (!e->sentinel)
 287                 q->nr_elts++;
 288
 289         l_add_tail(q->es, q->qs + e->level, e);
 290 }
 291
 292 static void q_push_front(struct queue *q, struct entry *e)
 293 {
 294         BUG_ON(e->pending_work);
 295
 296         if (!e->sentinel)
 297                 q->nr_elts++;
 298
 299         l_add_head(q->es, q->qs + e->level, e);
 300 }
 301
 302 static void q_push_before(struct queue *q, struct entry *old, struct entry *e)
 303 {
 304         BUG_ON(e->pending_work);
 305
 306         if (!e->sentinel)
 307                 q->nr_elts++;
 308
 309         l_add_before(q->es, q->qs + e->level, old, e);
 310 }
 311
 312 static void q_del(struct queue *q, struct entry *e)
 313 {
 314         l_del(q->es, q->qs + e->level, e);
 315         if (!e->sentinel)
 316                 q->nr_elts--;
 317 }
 318
 319 /*
 320  * Return the oldest entry of the lowest populated level.
 321  */
 322 static struct entry *q_peek(struct queue *q, unsigned max_level, bool can_cross_sentinel)
 323 {
 324         unsigned level;
 325         struct entry *e;
 326
 327         max_level = min(max_level, q->nr_levels);
 328
 329         for (level = 0; level < max_level; level++)
 330                 for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e)) {
 331                         if (e->sentinel) {
 332                                 if (can_cross_sentinel)
 333                                         continue;
 334                                 else
 335                                         break;
 336                         }
 337
 338                         return e;
 339                 }
 340
 341         return NULL;
 342 }
 343
 344 static struct entry *q_pop(struct queue *q)
 345 {
 346         struct entry *e = q_peek(q, q->nr_levels, true);
 347
 348         if (e)
 349                 q_del(q, e);
 350
 351         return e;
 352 }
 353
 354 /*
 355  * This function assumes there is a non-sentinel entry to pop.  It's only
 356  * used by redistribute, so we know this is true.  It also doesn't adjust
 357  * the q->nr_elts count.
 358  */
 359 static struct entry *__redist_pop_from(struct queue *q, unsigned level)
 360 {
 361         struct entry *e;
 362
 363         for (; level < q->nr_levels; level++)
 364                 for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e))
 365                         if (!e->sentinel) {
 366                                 l_del(q->es, q->qs + e->level, e);
 367                                 return e;
 368                         }
 369
 370         return NULL;
 371 }
 372
 373 static void q_set_targets_subrange_(struct queue *q, unsigned nr_elts, unsigned lbegin, unsigned lend)
 374 {
 375         unsigned level, nr_levels, entries_per_level, remainder;
 376
 377         BUG_ON(lbegin > lend);
 378         BUG_ON(lend > q->nr_levels);
 379         nr_levels = lend - lbegin;
 380         entries_per_level = safe_div(nr_elts, nr_levels);
 381         remainder = safe_mod(nr_elts, nr_levels);
 382
 383         for (level = lbegin; level < lend; level++)
 384                 q->target_count[level] =
 385                         (level < (lbegin + remainder)) ? entries_per_level + 1u : entries_per_level;
 386 }
 387
 388 /*
 389  * Typically we have fewer elements in the top few levels which allows us
 390  * to adjust the promote threshold nicely.
 391  */
 392 static void q_set_targets(struct queue *q)
 393 {
 394         if (q->last_target_nr_elts == q->nr_elts)
 395                 return;
 396
 397         q->last_target_nr_elts = q->nr_elts;
 398
 399         if (q->nr_top_levels > q->nr_levels)
 400                 q_set_targets_subrange_(q, q->nr_elts, 0, q->nr_levels);
 401
 402         else {
 403                 q_set_targets_subrange_(q, q->nr_in_top_levels,
 404                                         q->nr_levels - q->nr_top_levels, q->nr_levels);
 405
 406                 if (q->nr_in_top_levels < q->nr_elts)
 407                         q_set_targets_subrange_(q, q->nr_elts - q->nr_in_top_levels,
 408                                                 0, q->nr_levels - q->nr_top_levels);
 409                 else
 410                         q_set_targets_subrange_(q, 0, 0, q->nr_levels - q->nr_top_levels);
 411         }
 412 }
 413
 414 static void q_redistribute(struct queue *q)
 415 {
 416         unsigned target, level;
 417         struct ilist *l, *l_above;
 418         struct entry *e;
 419
 420         q_set_targets(q);
 421
 422         for (level = 0u; level < q->nr_levels - 1u; level++) {
 423                 l = q->qs + level;
 424                 target = q->target_count[level];
 425
 426                 /*
 427                  * Pull down some entries from the level above.
 428                  */
 429                 while (l->nr_elts < target) {
 430                         e = __redist_pop_from(q, level + 1u);
 431                         if (!e) {
 432                                 /* bug in nr_elts */
 433                                 break;
 434                         }
 435
 436                         e->level = level;
 437                         l_add_tail(q->es, l, e);
 438                 }
 439
 440                 /*
 441                  * Push some entries up.
 442                  */
 443                 l_above = q->qs + level + 1u;
 444                 while (l->nr_elts > target) {
 445                         e = l_pop_tail(q->es, l);
 446
 447                         if (!e)
 448                                 /* bug in nr_elts */
 449                                 break;
 450
 451                         e->level = level + 1u;
 452                         l_add_tail(q->es, l_above, e);
 453                 }
 454         }
 455 }
 456
 457 static void q_requeue(struct queue *q, struct entry *e, unsigned extra_levels,
 458                       struct entry *s1, struct entry *s2)
 459 {
 460         struct entry *de;
 461         unsigned sentinels_passed = 0;
 462         unsigned new_level = min(q->nr_levels - 1u, e->level + extra_levels);
 463
 464         /* try and find an entry to swap with */
 465         if (extra_levels && (e->level < q->nr_levels - 1u)) {
 466                 for (de = l_head(q->es, q->qs + new_level); de && de->sentinel; de = l_next(q->es, de))
 467                         sentinels_passed++;
 468
 469                 if (de) {
 470                         q_del(q, de);
 471                         de->level = e->level;
 472                         if (s1) {
 473                                 switch (sentinels_passed) {
 474                                 case 0:
 475                                         q_push_before(q, s1, de);
 476                                         break;
 477
 478                                 case 1:
 479                                         q_push_before(q, s2, de);
 480                                         break;
 481
 482                                 default:
 483                                         q_push(q, de);
 484                                 }
 485                         } else
 486                                 q_push(q, de);
 487                 }
 488         }
 489
 490         q_del(q, e);
 491         e->level = new_level;
 492         q_push(q, e);
 493 }
 494
 495 /*----------------------------------------------------------------*/
 496
 497 #define FP_SHIFT 8
 498 #define SIXTEENTH (1u << (FP_SHIFT - 4u))
 499 #define EIGHTH (1u << (FP_SHIFT - 3u))
 500
 501 struct stats {
 502         unsigned hit_threshold;
 503         unsigned hits;
 504         unsigned misses;
 505 };
 506
 507 enum performance {
 508         Q_POOR,
 509         Q_FAIR,
 510         Q_WELL
 511 };
 512
 513 static void stats_init(struct stats *s, unsigned nr_levels)
 514 {
 515         s->hit_threshold = (nr_levels * 3u) / 4u;
 516         s->hits = 0u;
 517         s->misses = 0u;
 518 }
 519
 520 static void stats_reset(struct stats *s)
 521 {
 522         s->hits = s->misses = 0u;
 523 }
 524
 525 static void stats_level_accessed(struct stats *s, unsigned level)
 526 {
 527         if (level >= s->hit_threshold)
 528                 s->hits++;
 529         else
 530                 s->misses++;
 531 }
 532
 533 static void stats_miss(struct stats *s)
 534 {
 535         s->misses++;
 536 }
 537
 538 /*
 539  * There are times when we don't have any confidence in the hotspot queue.
 540  * Such as when a fresh cache is created and the blocks have been spread
 541  * out across the levels, or if an io load changes.  We detect this by
 542  * seeing how often a lookup is in the top levels of the hotspot queue.
 543  */
 544 static enum performance stats_assess(struct stats *s)
 545 {
 546         unsigned confidence = safe_div(s->hits << FP_SHIFT, s->hits + s->misses);
 547
 548         if (confidence < SIXTEENTH)
 549                 return Q_POOR;
 550
 551         else if (confidence < EIGHTH)
 552                 return Q_FAIR;
 553
 554         else
 555                 return Q_WELL;
 556 }
 557
 558 /*----------------------------------------------------------------*/
 559
 560 struct smq_hash_table {
 561         struct entry_space *es;
 562         unsigned long long hash_bits;
 563         unsigned *buckets;
 564 };
 565
 566 /*
 567  * All cache entries are stored in a chained hash table.  To save space we
 568  * use indexing again, and only store indexes to the next entry.
 569  */
 570 static int h_init(struct smq_hash_table *ht, struct entry_space *es, unsigned nr_entries)
 571 {
 572         unsigned i, nr_buckets;
 573
 574         ht->es = es;
 575         nr_buckets = roundup_pow_of_two(max(nr_entries / 4u, 16u));
 576         ht->hash_bits = __ffs(nr_buckets);
 577
 578         ht->buckets = vmalloc(sizeof(*ht->buckets) * nr_buckets);
 579         if (!ht->buckets)
 580                 return -ENOMEM;
 581
 582         for (i = 0; i < nr_buckets; i++)
 583                 ht->buckets[i] = INDEXER_NULL;
 584
 585         return 0;
 586 }
 587
 588 static void h_exit(struct smq_hash_table *ht)
 589 {
 590         vfree(ht->buckets);
 591 }
 592
 593 static struct entry *h_head(struct smq_hash_table *ht, unsigned bucket)
 594 {
 595         return to_entry(ht->es, ht->buckets[bucket]);
 596 }
 597
 598 static struct entry *h_next(struct smq_hash_table *ht, struct entry *e)
 599 {
 600         return to_entry(ht->es, e->hash_next);
 601 }
 602
 603 static void __h_insert(struct smq_hash_table *ht, unsigned bucket, struct entry *e)
 604 {
 605         e->hash_next = ht->buckets[bucket];
 606         ht->buckets[bucket] = to_index(ht->es, e);
 607 }
 608
 609 static void h_insert(struct smq_hash_table *ht, struct entry *e)
 610 {
 611         unsigned h = hash_64(from_oblock(e->oblock), ht->hash_bits);
 612         __h_insert(ht, h, e);
 613 }
 614
 615 static struct entry *__h_lookup(struct smq_hash_table *ht, unsigned h, dm_oblock_t oblock,
 616                                 struct entry **prev)
 617 {
 618         struct entry *e;
 619
 620         *prev = NULL;
 621         for (e = h_head(ht, h); e; e = h_next(ht, e)) {
 622                 if (e->oblock == oblock)
 623                         return e;
 624
 625                 *prev = e;
 626         }
 627
 628         return NULL;
 629 }
 630
 631 static void __h_unlink(struct smq_hash_table *ht, unsigned h,
 632                        struct entry *e, struct entry *prev)
 633 {
 634         if (prev)
 635                 prev->hash_next = e->hash_next;
 636         else
 637                 ht->buckets[h] = e->hash_next;
 638 }
 639
 640 /*
 641  * Also moves each entry to the front of the bucket.
 642  */
 643 static struct entry *h_lookup(struct smq_hash_table *ht, dm_oblock_t oblock)
 644 {
 645         struct entry *e, *prev;
 646         unsigned h = hash_64(from_oblock(oblock), ht->hash_bits);
 647
 648         e = __h_lookup(ht, h, oblock, &prev);
 649         if (e && prev) {
 650                 /*
 651                  * Move to the front because this entry is likely
 652                  * to be hit again.
 653                  */
 654                 __h_unlink(ht, h, e, prev);
 655                 __h_insert(ht, h, e);
 656         }
 657
 658         return e;
 659 }
 660
 661 static void h_remove(struct smq_hash_table *ht, struct entry *e)
 662 {
 663         unsigned h = hash_64(from_oblock(e->oblock), ht->hash_bits);
 664         struct entry *prev;
 665
 666         /*
 667          * The down side of using a singly linked list is we have to
 668          * iterate the bucket to remove an item.
 669          */
 670         e = __h_lookup(ht, h, e->oblock, &prev);
 671         if (e)
 672                 __h_unlink(ht, h, e, prev);
 673 }
 674
 675 /*----------------------------------------------------------------*/
 676
 677 struct entry_alloc {
 678         struct entry_space *es;
 679         unsigned begin;
 680
 681         unsigned nr_allocated;
 682         struct ilist free;
 683 };
 684
 685 static void init_allocator(struct entry_alloc *ea, struct entry_space *es,
 686                            unsigned begin, unsigned end)
 687 {
 688         unsigned i;
 689
 690         ea->es = es;
 691         ea->nr_allocated = 0u;
 692         ea->begin = begin;
 693
 694         l_init(&ea->free);
 695         for (i = begin; i != end; i++)
 696                 l_add_tail(ea->es, &ea->free, __get_entry(ea->es, i));
 697 }
 698
 699 static void init_entry(struct entry *e)
 700 {
 701         /*
 702          * We can't memset because that would clear the hotspot and
 703          * sentinel bits which remain constant.
 704          */
 705         e->hash_next = INDEXER_NULL;
 706         e->next = INDEXER_NULL;
 707         e->prev = INDEXER_NULL;
 708         e->level = 0u;
 709         e->dirty = true;        /* FIXME: audit */
 710         e->allocated = true;
 711         e->sentinel = false;
 712         e->pending_work = false;
 713 }
 714
 715 static struct entry *alloc_entry(struct entry_alloc *ea)
 716 {
 717         struct entry *e;
 718
 719         if (l_empty(&ea->free))
 720                 return NULL;
 721
 722         e = l_pop_tail(ea->es, &ea->free);
 723         init_entry(e);
 724         ea->nr_allocated++;
 725
 726         return e;
 727 }
 728
 729 /*
 730  * This assumes the cblock hasn't already been allocated.
 731  */
 732 static struct entry *alloc_particular_entry(struct entry_alloc *ea, unsigned i)
 733 {
 734         struct entry *e = __get_entry(ea->es, ea->begin + i);
 735
 736         BUG_ON(e->allocated);
 737
 738         l_del(ea->es, &ea->free, e);
 739         init_entry(e);
 740         ea->nr_allocated++;
 741
 742         return e;
 743 }
 744
 745 static void free_entry(struct entry_alloc *ea, struct entry *e)
 746 {
 747         BUG_ON(!ea->nr_allocated);
 748         BUG_ON(!e->allocated);
 749
 750         ea->nr_allocated--;
 751         e->allocated = false;
 752         l_add_tail(ea->es, &ea->free, e);
 753 }
 754
 755 static bool allocator_empty(struct entry_alloc *ea)
 756 {
 757         return l_empty(&ea->free);
 758 }
 759
 760 static unsigned get_index(struct entry_alloc *ea, struct entry *e)
 761 {
 762         return to_index(ea->es, e) - ea->begin;
 763 }
 764
 765 static struct entry *get_entry(struct entry_alloc *ea, unsigned index)
 766 {
 767         return __get_entry(ea->es, ea->begin + index);
 768 }
 769
 770 /*----------------------------------------------------------------*/
 771
 772 #define NR_HOTSPOT_LEVELS 64u
 773 #define NR_CACHE_LEVELS 64u
 774
 775 #define WRITEBACK_PERIOD (10ul * HZ)
 776 #define DEMOTE_PERIOD (60ul * HZ)
 777
 778 #define HOTSPOT_UPDATE_PERIOD (HZ)
 779 #define CACHE_UPDATE_PERIOD (60ul * HZ)
 780
 781 struct smq_policy {
 782         struct dm_cache_policy policy;
 783
 784         /* protects everything */
 785         spinlock_t lock;
 786         dm_cblock_t cache_size;
 787         sector_t cache_block_size;
 788
 789         sector_t hotspot_block_size;
 790         unsigned nr_hotspot_blocks;
 791         unsigned cache_blocks_per_hotspot_block;
 792         unsigned hotspot_level_jump;
 793
 794         struct entry_space es;
 795         struct entry_alloc writeback_sentinel_alloc;
 796         struct entry_alloc demote_sentinel_alloc;
 797         struct entry_alloc hotspot_alloc;
 798         struct entry_alloc cache_alloc;
 799
 800         unsigned long *hotspot_hit_bits;
 801         unsigned long *cache_hit_bits;
 802
 803         /*
 804          * We maintain three queues of entries.  The cache proper,
 805          * consisting of a clean and dirty queue, containing the currently
 806          * active mappings.  The hotspot queue uses a larger block size to
 807          * track blocks that are being hit frequently and potential
 808          * candidates for promotion to the cache.
 809          */
 810         struct queue hotspot;
 811         struct queue clean;
 812         struct queue dirty;
 813
 814         struct stats hotspot_stats;
 815         struct stats cache_stats;
 816
 817         /*
 818          * Keeps track of time, incremented by the core.  We use this to
 819          * avoid attributing multiple hits within the same tick.
 820          */
 821         unsigned tick;
 822
 823         /*
 824          * The hash tables allows us to quickly find an entry by origin
 825          * block.
 826          */
 827         struct smq_hash_table table;
 828         struct smq_hash_table hotspot_table;
 829
 830         bool current_writeback_sentinels;
 831         unsigned long next_writeback_period;
 832
 833         bool current_demote_sentinels;
 834         unsigned long next_demote_period;
 835
 836         unsigned write_promote_level;
 837         unsigned read_promote_level;
 838
 839         unsigned long next_hotspot_period;
 840         unsigned long next_cache_period;
 841
 842         struct background_tracker *bg_work;
 843
 844         bool migrations_allowed;
 845 };
 846
 847 /*----------------------------------------------------------------*/
 848
 849 static struct entry *get_sentinel(struct entry_alloc *ea, unsigned level, bool which)
 850 {
 851         return get_entry(ea, which ? level : NR_CACHE_LEVELS + level);
 852 }
 853
 854 static struct entry *writeback_sentinel(struct smq_policy *mq, unsigned level)
 855 {
 856         return get_sentinel(&mq->writeback_sentinel_alloc, level, mq->current_writeback_sentinels);
 857 }
 858
 859 static struct entry *demote_sentinel(struct smq_policy *mq, unsigned level)
 860 {
 861         return get_sentinel(&mq->demote_sentinel_alloc, level, mq->current_demote_sentinels);
 862 }
 863
 864 static void __update_writeback_sentinels(struct smq_policy *mq)
 865 {
 866         unsigned level;
 867         struct queue *q = &mq->dirty;
 868         struct entry *sentinel;
 869
 870         for (level = 0; level < q->nr_levels; level++) {
 871                 sentinel = writeback_sentinel(mq, level);
 872                 q_del(q, sentinel);
 873                 q_push(q, sentinel);
 874         }
 875 }
 876
 877 static void __update_demote_sentinels(struct smq_policy *mq)
 878 {
 879         unsigned level;
 880         struct queue *q = &mq->clean;
 881         struct entry *sentinel;
 882
 883         for (level = 0; level < q->nr_levels; level++) {
 884                 sentinel = demote_sentinel(mq, level);
 885                 q_del(q, sentinel);
 886                 q_push(q, sentinel);
 887         }
 888 }
 889
 890 static void update_sentinels(struct smq_policy *mq)
 891 {
 892         if (time_after(jiffies, mq->next_writeback_period)) {
 893                 mq->next_writeback_period = jiffies + WRITEBACK_PERIOD;
 894                 mq->current_writeback_sentinels = !mq->current_writeback_sentinels;
 895                 __update_writeback_sentinels(mq);
 896         }
 897
 898         if (time_after(jiffies, mq->next_demote_period)) {
 899                 mq->next_demote_period = jiffies + DEMOTE_PERIOD;
 900                 mq->current_demote_sentinels = !mq->current_demote_sentinels;
 901                 __update_demote_sentinels(mq);
 902         }
 903 }
 904
 905 static void __sentinels_init(struct smq_policy *mq)
 906 {
 907         unsigned level;
 908         struct entry *sentinel;
 909
 910         for (level = 0; level < NR_CACHE_LEVELS; level++) {
 911                 sentinel = writeback_sentinel(mq, level);
 912                 sentinel->level = level;
 913                 q_push(&mq->dirty, sentinel);
 914
 915                 sentinel = demote_sentinel(mq, level);
 916                 sentinel->level = level;
 917                 q_push(&mq->clean, sentinel);
 918         }
 919 }
 920
 921 static void sentinels_init(struct smq_policy *mq)
 922 {
 923         mq->next_writeback_period = jiffies + WRITEBACK_PERIOD;
 924         mq->next_demote_period = jiffies + DEMOTE_PERIOD;
 925
 926         mq->current_writeback_sentinels = false;
 927         mq->current_demote_sentinels = false;
 928         __sentinels_init(mq);
 929
 930         mq->current_writeback_sentinels = !mq->current_writeback_sentinels;
 931         mq->current_demote_sentinels = !mq->current_demote_sentinels;
 932         __sentinels_init(mq);
 933 }
 934
 935 /*----------------------------------------------------------------*/
 936
 937 static void del_queue(struct smq_policy *mq, struct entry *e)
 938 {
 939         q_del(e->dirty ? &mq->dirty : &mq->clean, e);
 940 }
 941
 942 static void push_queue(struct smq_policy *mq, struct entry *e)
 943 {
 944         if (e->dirty)
 945                 q_push(&mq->dirty, e);
 946         else
 947                 q_push(&mq->clean, e);
 948 }
 949
 950 // !h, !q, a -> h, q, a
 951 static void push(struct smq_policy *mq, struct entry *e)
 952 {
 953         h_insert(&mq->table, e);
 954         if (!e->pending_work)
 955                 push_queue(mq, e);
 956 }
 957
 958 static void push_queue_front(struct smq_policy *mq, struct entry *e)
 959 {
 960         if (e->dirty)
 961                 q_push_front(&mq->dirty, e);
 962         else
 963                 q_push_front(&mq->clean, e);
 964 }
 965
 966 static void push_front(struct smq_policy *mq, struct entry *e)
 967 {
 968         h_insert(&mq->table, e);
 969         if (!e->pending_work)
 970                 push_queue_front(mq, e);
 971 }
 972
 973 static dm_cblock_t infer_cblock(struct smq_policy *mq, struct entry *e)
 974 {
 975         return to_cblock(get_index(&mq->cache_alloc, e));
 976 }
 977
 978 static void requeue(struct smq_policy *mq, struct entry *e)
 979 {
 980         /*
 981          * Pending work has temporarily been taken out of the queues.
 982          */
 983         if (e->pending_work)
 984                 return;
 985
 986         if (!test_and_set_bit(from_cblock(infer_cblock(mq, e)), mq->cache_hit_bits)) {
 987                 if (!e->dirty) {
 988                         q_requeue(&mq->clean, e, 1u, NULL, NULL);
 989                         return;
 990                 }
 991
 992                 q_requeue(&mq->dirty, e, 1u,
 993                           get_sentinel(&mq->writeback_sentinel_alloc, e->level, !mq->current_writeback_sentinels),
 994                           get_sentinel(&mq->writeback_sentinel_alloc, e->level, mq->current_writeback_sentinels));
 995         }
 996 }
 997
 998 static unsigned default_promote_level(struct smq_policy *mq)
 999 {
1000         /*
1001          * The promote level depends on the current performance of the
1002          * cache.
1003          *
1004          * If the cache is performing badly, then we can't afford
1005          * to promote much without causing performance to drop below that
1006          * of the origin device.
1007          *
1008          * If the cache is performing well, then we don't need to promote
1009          * much.  If it isn't broken, don't fix it.
1010          *
1011          * If the cache is middling then we promote more.
1012          *
1013          * This scheme reminds me of a graph of entropy vs probability of a
1014          * binary variable.
1015          */
1016         static unsigned table[] = {1, 1, 1, 2, 4, 6, 7, 8, 7, 6, 4, 4, 3, 3, 2, 2, 1};
1017
1018         unsigned hits = mq->cache_stats.hits;
1019         unsigned misses = mq->cache_stats.misses;
1020         unsigned index = safe_div(hits << 4u, hits + misses);
1021         return table[index];
1022 }
1023
1024 static void update_promote_levels(struct smq_policy *mq)
1025 {
1026         /*
1027          * If there are unused cache entries then we want to be really
1028          * eager to promote.
1029          */
1030         unsigned threshold_level = allocator_empty(&mq->cache_alloc) ?
1031                 default_promote_level(mq) : (NR_HOTSPOT_LEVELS / 2u);
1032
1033         threshold_level = max(threshold_level, NR_HOTSPOT_LEVELS);
1034
1035         /*
1036          * If the hotspot queue is performing badly then we have little
1037          * confidence that we know which blocks to promote.  So we cut down
1038          * the amount of promotions.
1039          */
1040         switch (stats_assess(&mq->hotspot_stats)) {
1041         case Q_POOR:
1042                 threshold_level /= 4u;
1043                 break;
1044
1045         case Q_FAIR:
1046                 threshold_level /= 2u;
1047                 break;
1048
1049         case Q_WELL:
1050                 break;
1051         }
1052
1053         mq->read_promote_level = NR_HOTSPOT_LEVELS - threshold_level;
1054         mq->write_promote_level = (NR_HOTSPOT_LEVELS - threshold_level);
1055 }
1056
1057 /*
1058  * If the hotspot queue is performing badly, then we try and move entries
1059  * around more quickly.
1060  */
1061 static void update_level_jump(struct smq_policy *mq)
1062 {
1063         switch (stats_assess(&mq->hotspot_stats)) {
1064         case Q_POOR:
1065                 mq->hotspot_level_jump = 4u;
1066                 break;
1067
1068         case Q_FAIR:
1069                 mq->hotspot_level_jump = 2u;
1070                 break;
1071
1072         case Q_WELL:
1073                 mq->hotspot_level_jump = 1u;
1074                 break;
1075         }
1076 }
1077
1078 static void end_hotspot_period(struct smq_policy *mq)
1079 {
1080         clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
1081         update_promote_levels(mq);
1082
1083         if (time_after(jiffies, mq->next_hotspot_period)) {
1084                 update_level_jump(mq);
1085                 q_redistribute(&mq->hotspot);
1086                 stats_reset(&mq->hotspot_stats);
1087                 mq->next_hotspot_period = jiffies + HOTSPOT_UPDATE_PERIOD;
1088         }
1089 }
1090
1091 static void end_cache_period(struct smq_policy *mq)
1092 {
1093         if (time_after(jiffies, mq->next_cache_period)) {
1094                 clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
1095
1096                 q_redistribute(&mq->dirty);
1097                 q_redistribute(&mq->clean);
1098                 stats_reset(&mq->cache_stats);
1099
1100                 mq->next_cache_period = jiffies + CACHE_UPDATE_PERIOD;
1101         }
1102 }
1103
1104 /*----------------------------------------------------------------*/
1105
1106 /*
1107  * Targets are given as a percentage.
1108  */
1109 #define CLEAN_TARGET 25u
1110 #define FREE_TARGET 25u
1111
1112 static unsigned percent_to_target(struct smq_policy *mq, unsigned p)
1113 {
1114         return from_cblock(mq->cache_size) * p / 100u;
1115 }
1116
1117 static bool clean_target_met(struct smq_policy *mq, bool idle)
1118 {
1119         /*
1120          * Cache entries may not be populated.  So we cannot rely on the
1121          * size of the clean queue.
1122          */
1123         if (idle) {
1124                 /*
1125                  * We'd like to clean everything.
1126                  */
1127                 return q_size(&mq->dirty) == 0u;
1128         }
1129
1130         /*
1131          * If we're busy we don't worry about cleaning at all.
1132          */
1133         return true;
1134 }
1135
1136 static bool free_target_met(struct smq_policy *mq)
1137 {
1138         unsigned nr_free;
1139
1140         nr_free = from_cblock(mq->cache_size) - mq->cache_alloc.nr_allocated;
1141         return (nr_free + btracker_nr_demotions_queued(mq->bg_work)) >=
1142                 percent_to_target(mq, FREE_TARGET);
1143 }
1144
1145 /*----------------------------------------------------------------*/
1146
1147 static void mark_pending(struct smq_policy *mq, struct entry *e)
1148 {
1149         BUG_ON(e->sentinel);
1150         BUG_ON(!e->allocated);
1151         BUG_ON(e->pending_work);
1152         e->pending_work = true;
1153 }
1154
1155 static void clear_pending(struct smq_policy *mq, struct entry *e)
1156 {
1157         BUG_ON(!e->pending_work);
1158         e->pending_work = false;
1159 }
1160
1161 static void queue_writeback(struct smq_policy *mq)
1162 {
1163         int r;
1164         struct policy_work work;
1165         struct entry *e;
1166
1167         e = q_peek(&mq->dirty, mq->dirty.nr_levels, !mq->migrations_allowed);
1168         if (e) {
1169                 mark_pending(mq, e);
1170                 q_del(&mq->dirty, e);
1171
1172                 work.op = POLICY_WRITEBACK;
1173                 work.oblock = e->oblock;
1174                 work.cblock = infer_cblock(mq, e);
1175
1176                 r = btracker_queue(mq->bg_work, &work, NULL);
1177                 WARN_ON_ONCE(r); // FIXME: finish, I think we have to get rid of this race.
1178         }
1179 }
1180
1181 static void queue_demotion(struct smq_policy *mq)
1182 {
1183         struct policy_work work;
1184         struct entry *e;
1185
1186         if (unlikely(WARN_ON_ONCE(!mq->migrations_allowed)))
1187                 return;
1188
1189         e = q_peek(&mq->clean, mq->clean.nr_levels / 2, true);
1190         if (!e) {
1191                 if (!clean_target_met(mq, true))
1192                         queue_writeback(mq);
1193                 return;
1194         }
1195
1196         mark_pending(mq, e);
1197         q_del(&mq->clean, e);
1198
1199         work.op = POLICY_DEMOTE;
1200         work.oblock = e->oblock;
1201         work.cblock = infer_cblock(mq, e);
1202         btracker_queue(mq->bg_work, &work, NULL);
1203 }
1204
1205 static void queue_promotion(struct smq_policy *mq, dm_oblock_t oblock,
1206                             struct policy_work **workp)
1207 {
1208         struct entry *e;
1209         struct policy_work work;
1210
1211         if (!mq->migrations_allowed)
1212                 return;
1213
1214         if (allocator_empty(&mq->cache_alloc)) {
1215                 /*
1216                  * We always claim to be 'idle' to ensure some demotions happen
1217                  * with continuous loads.
1218                  */
1219                 if (!free_target_met(mq))
1220                         queue_demotion(mq);
1221                 return;
1222         }
1223
1224         if (btracker_promotion_already_present(mq->bg_work, oblock))
1225                 return;
1226
1227         /*
1228          * We allocate the entry now to reserve the cblock.  If the
1229          * background work is aborted we must remember to free it.
1230          */
1231         e = alloc_entry(&mq->cache_alloc);
1232         BUG_ON(!e);
1233         e->pending_work = true;
1234         work.op = POLICY_PROMOTE;
1235         work.oblock = oblock;
1236         work.cblock = infer_cblock(mq, e);
1237         btracker_queue(mq->bg_work, &work, workp);
1238 }
1239
1240 /*----------------------------------------------------------------*/
1241
1242 enum promote_result {
1243         PROMOTE_NOT,
1244         PROMOTE_TEMPORARY,
1245         PROMOTE_PERMANENT
1246 };
1247
1248 /*
1249  * Converts a boolean into a promote result.
1250  */
1251 static enum promote_result maybe_promote(bool promote)
1252 {
1253         return promote ? PROMOTE_PERMANENT : PROMOTE_NOT;
1254 }
1255
1256 static enum promote_result should_promote(struct smq_policy *mq, struct entry *hs_e,
1257                                           int data_dir, bool fast_promote)
1258 {
1259         if (data_dir == WRITE) {
1260                 if (!allocator_empty(&mq->cache_alloc) && fast_promote)
1261                         return PROMOTE_TEMPORARY;
1262
1263                 return maybe_promote(hs_e->level >= mq->write_promote_level);
1264         } else
1265                 return maybe_promote(hs_e->level >= mq->read_promote_level);
1266 }
1267
1268 static dm_oblock_t to_hblock(struct smq_policy *mq, dm_oblock_t b)
1269 {
1270         sector_t r = from_oblock(b);
1271         (void) sector_div(r, mq->cache_blocks_per_hotspot_block);
1272         return to_oblock(r);
1273 }
1274
1275 static struct entry *update_hotspot_queue(struct smq_policy *mq, dm_oblock_t b)
1276 {
1277         unsigned hi;
1278         dm_oblock_t hb = to_hblock(mq, b);
1279         struct entry *e = h_lookup(&mq->hotspot_table, hb);
1280
1281         if (e) {
1282                 stats_level_accessed(&mq->hotspot_stats, e->level);
1283
1284                 hi = get_index(&mq->hotspot_alloc, e);
1285                 q_requeue(&mq->hotspot, e,
1286                           test_and_set_bit(hi, mq->hotspot_hit_bits) ?
1287                           0u : mq->hotspot_level_jump,
1288                           NULL, NULL);
1289
1290         } else {
1291                 stats_miss(&mq->hotspot_stats);
1292
1293                 e = alloc_entry(&mq->hotspot_alloc);
1294                 if (!e) {
1295                         e = q_pop(&mq->hotspot);
1296                         if (e) {
1297                                 h_remove(&mq->hotspot_table, e);
1298                                 hi = get_index(&mq->hotspot_alloc, e);
1299                                 clear_bit(hi, mq->hotspot_hit_bits);
1300                         }
1301
1302                 }
1303
1304                 if (e) {
1305                         e->oblock = hb;
1306                         q_push(&mq->hotspot, e);
1307                         h_insert(&mq->hotspot_table, e);
1308                 }
1309         }
1310
1311         return e;
1312 }
1313
1314 /*----------------------------------------------------------------*/
1315
1316 /*
1317  * Public interface, via the policy struct.  See dm-cache-policy.h for a
1318  * description of these.
1319  */
1320
1321 static struct smq_policy *to_smq_policy(struct dm_cache_policy *p)
1322 {
1323         return container_of(p, struct smq_policy, policy);
1324 }
1325
1326 static void smq_destroy(struct dm_cache_policy *p)
1327 {
1328         struct smq_policy *mq = to_smq_policy(p);
1329
1330         btracker_destroy(mq->bg_work);
1331         h_exit(&mq->hotspot_table);
1332         h_exit(&mq->table);
1333         free_bitset(mq->hotspot_hit_bits);
1334         free_bitset(mq->cache_hit_bits);
1335         space_exit(&mq->es);
1336         kfree(mq);
1337 }
1338
1339 /*----------------------------------------------------------------*/
1340
1341 static int __lookup(struct smq_policy *mq, dm_oblock_t oblock, dm_cblock_t *cblock,
1342                     int data_dir, bool fast_copy,
1343                     struct policy_work **work, bool *background_work)
1344 {
1345         struct entry *e, *hs_e;
1346         enum promote_result pr;
1347
1348         *background_work = false;
1349
1350         e = h_lookup(&mq->table, oblock);
1351         if (e) {
1352                 stats_level_accessed(&mq->cache_stats, e->level);
1353
1354                 requeue(mq, e);
1355                 *cblock = infer_cblock(mq, e);
1356                 return 0;
1357
1358         } else {
1359                 stats_miss(&mq->cache_stats);
1360
1361                 /*
1362                  * The hotspot queue only gets updated with misses.
1363                  */
1364                 hs_e = update_hotspot_queue(mq, oblock);
1365
1366                 pr = should_promote(mq, hs_e, data_dir, fast_copy);
1367                 if (pr != PROMOTE_NOT) {
1368                         queue_promotion(mq, oblock, work);
1369                         *background_work = true;
1370                 }
1371
1372                 return -ENOENT;
1373         }
1374 }
1375
1376 static int smq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock,
1377                       int data_dir, bool fast_copy,
1378                       bool *background_work)
1379 {
1380         int r;
1381         unsigned long flags;
1382         struct smq_policy *mq = to_smq_policy(p);
1383
1384         spin_lock_irqsave(&mq->lock, flags);
1385         r = __lookup(mq, oblock, cblock,
1386                      data_dir, fast_copy,
1387                      NULL, background_work);
1388         spin_unlock_irqrestore(&mq->lock, flags);
1389
1390         return r;
1391 }
1392
1393 static int smq_lookup_with_work(struct dm_cache_policy *p,
1394                                 dm_oblock_t oblock, dm_cblock_t *cblock,
1395                                 int data_dir, bool fast_copy,
1396                                 struct policy_work **work)
1397 {
1398         int r;
1399         bool background_queued;
1400         unsigned long flags;
1401         struct smq_policy *mq = to_smq_policy(p);
1402
1403         spin_lock_irqsave(&mq->lock, flags);
1404         r = __lookup(mq, oblock, cblock, data_dir, fast_copy, work, &background_queued);
1405         spin_unlock_irqrestore(&mq->lock, flags);
1406
1407         return r;
1408 }
1409
1410 static int smq_get_background_work(struct dm_cache_policy *p, bool idle,
1411                                    struct policy_work **result)
1412 {
1413         int r;
1414         unsigned long flags;
1415         struct smq_policy *mq = to_smq_policy(p);
1416
1417         spin_lock_irqsave(&mq->lock, flags);
1418         r = btracker_issue(mq->bg_work, result);
1419         if (r == -ENODATA) {
1420                 if (!clean_target_met(mq, idle)) {
1421                         queue_writeback(mq);
1422                         r = btracker_issue(mq->bg_work, result);
1423                 }
1424         }
1425         spin_unlock_irqrestore(&mq->lock, flags);
1426
1427         return r;
1428 }
1429
1430 /*
1431  * We need to clear any pending work flags that have been set, and in the
1432  * case of promotion free the entry for the destination cblock.
1433  */
1434 static void __complete_background_work(struct smq_policy *mq,
1435                                        struct policy_work *work,
1436                                        bool success)
1437 {
1438         struct entry *e = get_entry(&mq->cache_alloc,
1439                                     from_cblock(work->cblock));
1440
1441         switch (work->op) {
1442         case POLICY_PROMOTE:
1443                 // !h, !q, a
1444                 clear_pending(mq, e);
1445                 if (success) {
1446                         e->oblock = work->oblock;
1447                         e->level = NR_CACHE_LEVELS - 1;
1448                         push(mq, e);
1449                         // h, q, a
1450                 } else {
1451                         free_entry(&mq->cache_alloc, e);
1452                         // !h, !q, !a
1453                 }
1454                 break;
1455
1456         case POLICY_DEMOTE:
1457                 // h, !q, a
1458                 if (success) {
1459                         h_remove(&mq->table, e);
1460                         free_entry(&mq->cache_alloc, e);
1461                         // !h, !q, !a
1462                 } else {
1463                         clear_pending(mq, e);
1464                         push_queue(mq, e);
1465                         // h, q, a
1466                 }
1467                 break;
1468
1469         case POLICY_WRITEBACK:
1470                 // h, !q, a
1471                 clear_pending(mq, e);
1472                 push_queue(mq, e);
1473                 // h, q, a
1474                 break;
1475         }
1476
1477         btracker_complete(mq->bg_work, work);
1478 }
1479
1480 static void smq_complete_background_work(struct dm_cache_policy *p,
1481                                          struct policy_work *work,
1482                                          bool success)
1483 {
1484         unsigned long flags;
1485         struct smq_policy *mq = to_smq_policy(p);
1486
1487         spin_lock_irqsave(&mq->lock, flags);
1488         __complete_background_work(mq, work, success);
1489         spin_unlock_irqrestore(&mq->lock, flags);
1490 }
1491
1492 // in_hash(oblock) -> in_hash(oblock)
1493 static void __smq_set_clear_dirty(struct smq_policy *mq, dm_cblock_t cblock, bool set)
1494 {
1495         struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1496
1497         if (e->pending_work)
1498                 e->dirty = set;
1499         else {
1500                 del_queue(mq, e);
1501                 e->dirty = set;
1502                 push_queue(mq, e);
1503         }
1504 }
1505
1506 static void smq_set_dirty(struct dm_cache_policy *p, dm_cblock_t cblock)
1507 {
1508         unsigned long flags;
1509         struct smq_policy *mq = to_smq_policy(p);
1510
1511         spin_lock_irqsave(&mq->lock, flags);
1512         __smq_set_clear_dirty(mq, cblock, true);
1513         spin_unlock_irqrestore(&mq->lock, flags);
1514 }
1515
1516 static void smq_clear_dirty(struct dm_cache_policy *p, dm_cblock_t cblock)
1517 {
1518         struct smq_policy *mq = to_smq_policy(p);
1519         unsigned long flags;
1520
1521         spin_lock_irqsave(&mq->lock, flags);
1522         __smq_set_clear_dirty(mq, cblock, false);
1523         spin_unlock_irqrestore(&mq->lock, flags);
1524 }
1525
1526 static unsigned random_level(dm_cblock_t cblock)
1527 {
1528         return hash_32(from_cblock(cblock), 9) & (NR_CACHE_LEVELS - 1);
1529 }
1530
1531 static int smq_load_mapping(struct dm_cache_policy *p,
1532                             dm_oblock_t oblock, dm_cblock_t cblock,
1533                             bool dirty, uint32_t hint, bool hint_valid)
1534 {
1535         struct smq_policy *mq = to_smq_policy(p);
1536         struct entry *e;
1537
1538         e = alloc_particular_entry(&mq->cache_alloc, from_cblock(cblock));
1539         e->oblock = oblock;
1540         e->dirty = dirty;
1541         e->level = hint_valid ? min(hint, NR_CACHE_LEVELS - 1) : random_level(cblock);
1542         e->pending_work = false;
1543
1544         /*
1545          * When we load mappings we push ahead of both sentinels in order to
1546          * allow demotions and cleaning to occur immediately.
1547          */
1548         push_front(mq, e);
1549
1550         return 0;
1551 }
1552
1553 static int smq_invalidate_mapping(struct dm_cache_policy *p, dm_cblock_t cblock)
1554 {
1555         struct smq_policy *mq = to_smq_policy(p);
1556         struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1557
1558         if (!e->allocated)
1559                 return -ENODATA;
1560
1561         // FIXME: what if this block has pending background work?
1562         del_queue(mq, e);
1563         h_remove(&mq->table, e);
1564         free_entry(&mq->cache_alloc, e);
1565         return 0;
1566 }
1567
1568 static uint32_t smq_get_hint(struct dm_cache_policy *p, dm_cblock_t cblock)
1569 {
1570         struct smq_policy *mq = to_smq_policy(p);
1571         struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1572
1573         if (!e->allocated)
1574                 return 0;
1575
1576         return e->level;
1577 }
1578
1579 static dm_cblock_t smq_residency(struct dm_cache_policy *p)
1580 {
1581         dm_cblock_t r;
1582         unsigned long flags;
1583         struct smq_policy *mq = to_smq_policy(p);
1584
1585         spin_lock_irqsave(&mq->lock, flags);
1586         r = to_cblock(mq->cache_alloc.nr_allocated);
1587         spin_unlock_irqrestore(&mq->lock, flags);
1588
1589         return r;
1590 }
1591
1592 static void smq_tick(struct dm_cache_policy *p, bool can_block)
1593 {
1594         struct smq_policy *mq = to_smq_policy(p);
1595         unsigned long flags;
1596
1597         spin_lock_irqsave(&mq->lock, flags);
1598         mq->tick++;
1599         update_sentinels(mq);
1600         end_hotspot_period(mq);
1601         end_cache_period(mq);
1602         spin_unlock_irqrestore(&mq->lock, flags);
1603 }
1604
1605 static void smq_allow_migrations(struct dm_cache_policy *p, bool allow)
1606 {
1607         struct smq_policy *mq = to_smq_policy(p);
1608         mq->migrations_allowed = allow;
1609 }
1610
1611 /*
1612  * smq has no config values, but the old mq policy did.  To avoid breaking
1613  * software we continue to accept these configurables for the mq policy,
1614  * but they have no effect.
1615  */
1616 static int mq_set_config_value(struct dm_cache_policy *p,
1617                                const char *key, const char *value)
1618 {
1619         unsigned long tmp;
1620
1621         if (kstrtoul(value, 10, &tmp))
1622                 return -EINVAL;
1623
1624         if (!strcasecmp(key, "random_threshold") ||
1625             !strcasecmp(key, "sequential_threshold") ||
1626             !strcasecmp(key, "discard_promote_adjustment") ||
1627             !strcasecmp(key, "read_promote_adjustment") ||
1628             !strcasecmp(key, "write_promote_adjustment")) {
1629                 DMWARN("tunable '%s' no longer has any effect, mq policy is now an alias for smq", key);
1630                 return 0;
1631         }
1632
1633         return -EINVAL;
1634 }
1635
1636 static int mq_emit_config_values(struct dm_cache_policy *p, char *result,
1637                                  unsigned maxlen, ssize_t *sz_ptr)
1638 {
1639         ssize_t sz = *sz_ptr;
1640
1641         DMEMIT("10 random_threshold 0 "
1642                "sequential_threshold 0 "
1643                "discard_promote_adjustment 0 "
1644                "read_promote_adjustment 0 "
1645                "write_promote_adjustment 0 ");
1646
1647         *sz_ptr = sz;
1648         return 0;
1649 }
1650
1651 /* Init the policy plugin interface function pointers. */
1652 static void init_policy_functions(struct smq_policy *mq, bool mimic_mq)
1653 {
1654         mq->policy.destroy = smq_destroy;
1655         mq->policy.lookup = smq_lookup;
1656         mq->policy.lookup_with_work = smq_lookup_with_work;
1657         mq->policy.get_background_work = smq_get_background_work;
1658         mq->policy.complete_background_work = smq_complete_background_work;
1659         mq->policy.set_dirty = smq_set_dirty;
1660         mq->policy.clear_dirty = smq_clear_dirty;
1661         mq->policy.load_mapping = smq_load_mapping;
1662         mq->policy.invalidate_mapping = smq_invalidate_mapping;
1663         mq->policy.get_hint = smq_get_hint;
1664         mq->policy.residency = smq_residency;
1665         mq->policy.tick = smq_tick;
1666         mq->policy.allow_migrations = smq_allow_migrations;
1667
1668         if (mimic_mq) {
1669                 mq->policy.set_config_value = mq_set_config_value;
1670                 mq->policy.emit_config_values = mq_emit_config_values;
1671         }
1672 }
1673
1674 static bool too_many_hotspot_blocks(sector_t origin_size,
1675                                     sector_t hotspot_block_size,
1676                                     unsigned nr_hotspot_blocks)
1677 {
1678         return (hotspot_block_size * nr_hotspot_blocks) > origin_size;
1679 }
1680
1681 static void calc_hotspot_params(sector_t origin_size,
1682                                 sector_t cache_block_size,
1683                                 unsigned nr_cache_blocks,
1684                                 sector_t *hotspot_block_size,
1685                                 unsigned *nr_hotspot_blocks)
1686 {
1687         *hotspot_block_size = cache_block_size * 16u;
1688         *nr_hotspot_blocks = max(nr_cache_blocks / 4u, 1024u);
1689
1690         while ((*hotspot_block_size > cache_block_size) &&
1691                too_many_hotspot_blocks(origin_size, *hotspot_block_size, *nr_hotspot_blocks))
1692                 *hotspot_block_size /= 2u;
1693 }
1694
1695 static struct dm_cache_policy *__smq_create(dm_cblock_t cache_size,
1696                                             sector_t origin_size,
1697                                             sector_t cache_block_size,
1698                                             bool mimic_mq,
1699                                             bool migrations_allowed)
1700 {
1701         unsigned i;
1702         unsigned nr_sentinels_per_queue = 2u * NR_CACHE_LEVELS;
1703         unsigned total_sentinels = 2u * nr_sentinels_per_queue;
1704         struct smq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL);
1705
1706         if (!mq)
1707                 return NULL;
1708
1709         init_policy_functions(mq, mimic_mq);
1710         mq->cache_size = cache_size;
1711         mq->cache_block_size = cache_block_size;
1712
1713         calc_hotspot_params(origin_size, cache_block_size, from_cblock(cache_size),
1714                             &mq->hotspot_block_size, &mq->nr_hotspot_blocks);
1715
1716         mq->cache_blocks_per_hotspot_block = div64_u64(mq->hotspot_block_size, mq->cache_block_size);
1717         mq->hotspot_level_jump = 1u;
1718         if (space_init(&mq->es, total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size))) {
1719                 DMERR("couldn't initialize entry space");
1720                 goto bad_pool_init;
1721         }
1722
1723         init_allocator(&mq->writeback_sentinel_alloc, &mq->es, 0, nr_sentinels_per_queue);
1724         for (i = 0; i < nr_sentinels_per_queue; i++)
1725                 get_entry(&mq->writeback_sentinel_alloc, i)->sentinel = true;
1726
1727         init_allocator(&mq->demote_sentinel_alloc, &mq->es, nr_sentinels_per_queue, total_sentinels);
1728         for (i = 0; i < nr_sentinels_per_queue; i++)
1729                 get_entry(&mq->demote_sentinel_alloc, i)->sentinel = true;
1730
1731         init_allocator(&mq->hotspot_alloc, &mq->es, total_sentinels,
1732                        total_sentinels + mq->nr_hotspot_blocks);
1733
1734         init_allocator(&mq->cache_alloc, &mq->es,
1735                        total_sentinels + mq->nr_hotspot_blocks,
1736                        total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size));
1737
1738         mq->hotspot_hit_bits = alloc_bitset(mq->nr_hotspot_blocks);
1739         if (!mq->hotspot_hit_bits) {
1740                 DMERR("couldn't allocate hotspot hit bitset");
1741                 goto bad_hotspot_hit_bits;
1742         }
1743         clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
1744
1745         if (from_cblock(cache_size)) {
1746                 mq->cache_hit_bits = alloc_bitset(from_cblock(cache_size));
1747                 if (!mq->cache_hit_bits) {
1748                         DMERR("couldn't allocate cache hit bitset");
1749                         goto bad_cache_hit_bits;
1750                 }
1751                 clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
1752         } else
1753                 mq->cache_hit_bits = NULL;
1754
1755         mq->tick = 0;
1756         spin_lock_init(&mq->lock);
1757
1758         q_init(&mq->hotspot, &mq->es, NR_HOTSPOT_LEVELS);
1759         mq->hotspot.nr_top_levels = 8;
1760         mq->hotspot.nr_in_top_levels = min(mq->nr_hotspot_blocks / NR_HOTSPOT_LEVELS,
1761                                            from_cblock(mq->cache_size) / mq->cache_blocks_per_hotspot_block);
1762
1763         q_init(&mq->clean, &mq->es, NR_CACHE_LEVELS);
1764         q_init(&mq->dirty, &mq->es, NR_CACHE_LEVELS);
1765
1766         stats_init(&mq->hotspot_stats, NR_HOTSPOT_LEVELS);
1767         stats_init(&mq->cache_stats, NR_CACHE_LEVELS);
1768
1769         if (h_init(&mq->table, &mq->es, from_cblock(cache_size)))
1770                 goto bad_alloc_table;
1771
1772         if (h_init(&mq->hotspot_table, &mq->es, mq->nr_hotspot_blocks))
1773                 goto bad_alloc_hotspot_table;
1774
1775         sentinels_init(mq);
1776         mq->write_promote_level = mq->read_promote_level = NR_HOTSPOT_LEVELS;
1777
1778         mq->next_hotspot_period = jiffies;
1779         mq->next_cache_period = jiffies;
1780
1781         mq->bg_work = btracker_create(10240); /* FIXME: hard coded value */
1782         if (!mq->bg_work)
1783                 goto bad_btracker;
1784
1785         mq->migrations_allowed = migrations_allowed;
1786
1787         return &mq->policy;
1788
1789 bad_btracker:
1790         h_exit(&mq->hotspot_table);
1791 bad_alloc_hotspot_table:
1792         h_exit(&mq->table);
1793 bad_alloc_table:
1794         free_bitset(mq->cache_hit_bits);
1795 bad_cache_hit_bits:
1796         free_bitset(mq->hotspot_hit_bits);
1797 bad_hotspot_hit_bits:
1798         space_exit(&mq->es);
1799 bad_pool_init:
1800         kfree(mq);
1801
1802         return NULL;
1803 }
1804
1805 static struct dm_cache_policy *smq_create(dm_cblock_t cache_size,
1806                                           sector_t origin_size,
1807                                           sector_t cache_block_size)
1808 {
1809         return __smq_create(cache_size, origin_size, cache_block_size, false, true);
1810 }
1811
1812 static struct dm_cache_policy *mq_create(dm_cblock_t cache_size,
1813                                          sector_t origin_size,
1814                                          sector_t cache_block_size)
1815 {
1816         return __smq_create(cache_size, origin_size, cache_block_size, true, true);
1817 }
1818
1819 static struct dm_cache_policy *cleaner_create(dm_cblock_t cache_size,
1820                                               sector_t origin_size,
1821                                               sector_t cache_block_size)
1822 {
1823         return __smq_create(cache_size, origin_size, cache_block_size, false, false);
1824 }
1825
1826 /*----------------------------------------------------------------*/
1827
1828 static struct dm_cache_policy_type smq_policy_type = {
1829         .name = "smq",
1830         .version = {2, 0, 0},
1831         .hint_size = 4,
1832         .owner = THIS_MODULE,
1833         .create = smq_create
1834 };
1835
1836 static struct dm_cache_policy_type mq_policy_type = {
1837         .name = "mq",
1838         .version = {2, 0, 0},
1839         .hint_size = 4,
1840         .owner = THIS_MODULE,
1841         .create = mq_create,
1842 };
1843
1844 static struct dm_cache_policy_type cleaner_policy_type = {
1845         .name = "cleaner",
1846         .version = {2, 0, 0},
1847         .hint_size = 4,
1848         .owner = THIS_MODULE,
1849         .create = cleaner_create,
1850 };
1851
1852 static struct dm_cache_policy_type default_policy_type = {
1853         .name = "default",
1854         .version = {2, 0, 0},
1855         .hint_size = 4,
1856         .owner = THIS_MODULE,
1857         .create = smq_create,
1858         .real = &smq_policy_type
1859 };
1860
1861 static int __init smq_init(void)
1862 {
1863         int r;
1864
1865         r = dm_cache_policy_register(&smq_policy_type);
1866         if (r) {
1867                 DMERR("register failed %d", r);
1868                 return -ENOMEM;
1869         }
1870
1871         r = dm_cache_policy_register(&mq_policy_type);
1872         if (r) {
1873                 DMERR("register failed (as mq) %d", r);
1874                 goto out_mq;
1875         }
1876
1877         r = dm_cache_policy_register(&cleaner_policy_type);
1878         if (r) {
1879                 DMERR("register failed (as cleaner) %d", r);
1880                 goto out_cleaner;
1881         }
1882
1883         r = dm_cache_policy_register(&default_policy_type);
1884         if (r) {
1885                 DMERR("register failed (as default) %d", r);
1886                 goto out_default;
1887         }
1888
1889         return 0;
1890
1891 out_default:
1892         dm_cache_policy_unregister(&cleaner_policy_type);
1893 out_cleaner:
1894         dm_cache_policy_unregister(&mq_policy_type);
1895 out_mq:
1896         dm_cache_policy_unregister(&smq_policy_type);
1897
1898         return -ENOMEM;
1899 }
1900
1901 static void __exit smq_exit(void)
1902 {
1903         dm_cache_policy_unregister(&cleaner_policy_type);
1904         dm_cache_policy_unregister(&smq_policy_type);
1905         dm_cache_policy_unregister(&mq_policy_type);
1906         dm_cache_policy_unregister(&default_policy_type);
1907 }
1908
1909 module_init(smq_init);
1910 module_exit(smq_exit);
1911
1912 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
1913 MODULE_LICENSE("GPL");
1914 MODULE_DESCRIPTION("smq cache policy");
1915
1916 MODULE_ALIAS("dm-cache-default");
1917 MODULE_ALIAS("dm-cache-mq");
1918 MODULE_ALIAS("dm-cache-cleaner");