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(array_size(nr_entries, sizeof(struct entry)));
  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_head(struct entry_space *es, struct ilist *l)
 217 {
 218         struct entry *e;
 219
 220         for (e = l_head(es, l); e; e = l_next(es, e))
 221                 if (!e->sentinel) {
 222                         l_del(es, l, e);
 223                         return e;
 224                 }
 225
 226         return NULL;
 227 }
 228
 229 static struct entry *l_pop_tail(struct entry_space *es, struct ilist *l)
 230 {
 231         struct entry *e;
 232
 233         for (e = l_tail(es, l); e; e = l_prev(es, e))
 234                 if (!e->sentinel) {
 235                         l_del(es, l, e);
 236                         return e;
 237                 }
 238
 239         return NULL;
 240 }
 241
 242 /*----------------------------------------------------------------*/
 243
 244 /*
 245  * The stochastic-multi-queue is a set of lru lists stacked into levels.
 246  * Entries are moved up levels when they are used, which loosely orders the
 247  * most accessed entries in the top levels and least in the bottom.  This
 248  * structure is *much* better than a single lru list.
 249  */
 250 #define MAX_LEVELS 64u
 251
 252 struct queue {
 253         struct entry_space *es;
 254
 255         unsigned nr_elts;
 256         unsigned nr_levels;
 257         struct ilist qs[MAX_LEVELS];
 258
 259         /*
 260          * We maintain a count of the number of entries we would like in each
 261          * level.
 262          */
 263         unsigned last_target_nr_elts;
 264         unsigned nr_top_levels;
 265         unsigned nr_in_top_levels;
 266         unsigned target_count[MAX_LEVELS];
 267 };
 268
 269 static void q_init(struct queue *q, struct entry_space *es, unsigned nr_levels)
 270 {
 271         unsigned i;
 272
 273         q->es = es;
 274         q->nr_elts = 0;
 275         q->nr_levels = nr_levels;
 276
 277         for (i = 0; i < q->nr_levels; i++) {
 278                 l_init(q->qs + i);
 279                 q->target_count[i] = 0u;
 280         }
 281
 282         q->last_target_nr_elts = 0u;
 283         q->nr_top_levels = 0u;
 284         q->nr_in_top_levels = 0u;
 285 }
 286
 287 static unsigned q_size(struct queue *q)
 288 {
 289         return q->nr_elts;
 290 }
 291
 292 /*
 293  * Insert an entry to the back of the given level.
 294  */
 295 static void q_push(struct queue *q, struct entry *e)
 296 {
 297         BUG_ON(e->pending_work);
 298
 299         if (!e->sentinel)
 300                 q->nr_elts++;
 301
 302         l_add_tail(q->es, q->qs + e->level, e);
 303 }
 304
 305 static void q_push_front(struct queue *q, struct entry *e)
 306 {
 307         BUG_ON(e->pending_work);
 308
 309         if (!e->sentinel)
 310                 q->nr_elts++;
 311
 312         l_add_head(q->es, q->qs + e->level, e);
 313 }
 314
 315 static void q_push_before(struct queue *q, struct entry *old, struct entry *e)
 316 {
 317         BUG_ON(e->pending_work);
 318
 319         if (!e->sentinel)
 320                 q->nr_elts++;
 321
 322         l_add_before(q->es, q->qs + e->level, old, e);
 323 }
 324
 325 static void q_del(struct queue *q, struct entry *e)
 326 {
 327         l_del(q->es, q->qs + e->level, e);
 328         if (!e->sentinel)
 329                 q->nr_elts--;
 330 }
 331
 332 /*
 333  * Return the oldest entry of the lowest populated level.
 334  */
 335 static struct entry *q_peek(struct queue *q, unsigned max_level, bool can_cross_sentinel)
 336 {
 337         unsigned level;
 338         struct entry *e;
 339
 340         max_level = min(max_level, q->nr_levels);
 341
 342         for (level = 0; level < max_level; level++)
 343                 for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e)) {
 344                         if (e->sentinel) {
 345                                 if (can_cross_sentinel)
 346                                         continue;
 347                                 else
 348                                         break;
 349                         }
 350
 351                         return e;
 352                 }
 353
 354         return NULL;
 355 }
 356
 357 static struct entry *q_pop(struct queue *q)
 358 {
 359         struct entry *e = q_peek(q, q->nr_levels, true);
 360
 361         if (e)
 362                 q_del(q, e);
 363
 364         return e;
 365 }
 366
 367 /*
 368  * This function assumes there is a non-sentinel entry to pop.  It's only
 369  * used by redistribute, so we know this is true.  It also doesn't adjust
 370  * the q->nr_elts count.
 371  */
 372 static struct entry *__redist_pop_from(struct queue *q, unsigned level)
 373 {
 374         struct entry *e;
 375
 376         for (; level < q->nr_levels; level++)
 377                 for (e = l_head(q->es, q->qs + level); e; e = l_next(q->es, e))
 378                         if (!e->sentinel) {
 379                                 l_del(q->es, q->qs + e->level, e);
 380                                 return e;
 381                         }
 382
 383         return NULL;
 384 }
 385
 386 static void q_set_targets_subrange_(struct queue *q, unsigned nr_elts, unsigned lbegin, unsigned lend)
 387 {
 388         unsigned level, nr_levels, entries_per_level, remainder;
 389
 390         BUG_ON(lbegin > lend);
 391         BUG_ON(lend > q->nr_levels);
 392         nr_levels = lend - lbegin;
 393         entries_per_level = safe_div(nr_elts, nr_levels);
 394         remainder = safe_mod(nr_elts, nr_levels);
 395
 396         for (level = lbegin; level < lend; level++)
 397                 q->target_count[level] =
 398                         (level < (lbegin + remainder)) ? entries_per_level + 1u : entries_per_level;
 399 }
 400
 401 /*
 402  * Typically we have fewer elements in the top few levels which allows us
 403  * to adjust the promote threshold nicely.
 404  */
 405 static void q_set_targets(struct queue *q)
 406 {
 407         if (q->last_target_nr_elts == q->nr_elts)
 408                 return;
 409
 410         q->last_target_nr_elts = q->nr_elts;
 411
 412         if (q->nr_top_levels > q->nr_levels)
 413                 q_set_targets_subrange_(q, q->nr_elts, 0, q->nr_levels);
 414
 415         else {
 416                 q_set_targets_subrange_(q, q->nr_in_top_levels,
 417                                         q->nr_levels - q->nr_top_levels, q->nr_levels);
 418
 419                 if (q->nr_in_top_levels < q->nr_elts)
 420                         q_set_targets_subrange_(q, q->nr_elts - q->nr_in_top_levels,
 421                                                 0, q->nr_levels - q->nr_top_levels);
 422                 else
 423                         q_set_targets_subrange_(q, 0, 0, q->nr_levels - q->nr_top_levels);
 424         }
 425 }
 426
 427 static void q_redistribute(struct queue *q)
 428 {
 429         unsigned target, level;
 430         struct ilist *l, *l_above;
 431         struct entry *e;
 432
 433         q_set_targets(q);
 434
 435         for (level = 0u; level < q->nr_levels - 1u; level++) {
 436                 l = q->qs + level;
 437                 target = q->target_count[level];
 438
 439                 /*
 440                  * Pull down some entries from the level above.
 441                  */
 442                 while (l->nr_elts < target) {
 443                         e = __redist_pop_from(q, level + 1u);
 444                         if (!e) {
 445                                 /* bug in nr_elts */
 446                                 break;
 447                         }
 448
 449                         e->level = level;
 450                         l_add_tail(q->es, l, e);
 451                 }
 452
 453                 /*
 454                  * Push some entries up.
 455                  */
 456                 l_above = q->qs + level + 1u;
 457                 while (l->nr_elts > target) {
 458                         e = l_pop_tail(q->es, l);
 459
 460                         if (!e)
 461                                 /* bug in nr_elts */
 462                                 break;
 463
 464                         e->level = level + 1u;
 465                         l_add_tail(q->es, l_above, e);
 466                 }
 467         }
 468 }
 469
 470 static void q_requeue(struct queue *q, struct entry *e, unsigned extra_levels,
 471                       struct entry *s1, struct entry *s2)
 472 {
 473         struct entry *de;
 474         unsigned sentinels_passed = 0;
 475         unsigned new_level = min(q->nr_levels - 1u, e->level + extra_levels);
 476
 477         /* try and find an entry to swap with */
 478         if (extra_levels && (e->level < q->nr_levels - 1u)) {
 479                 for (de = l_head(q->es, q->qs + new_level); de && de->sentinel; de = l_next(q->es, de))
 480                         sentinels_passed++;
 481
 482                 if (de) {
 483                         q_del(q, de);
 484                         de->level = e->level;
 485                         if (s1) {
 486                                 switch (sentinels_passed) {
 487                                 case 0:
 488                                         q_push_before(q, s1, de);
 489                                         break;
 490
 491                                 case 1:
 492                                         q_push_before(q, s2, de);
 493                                         break;
 494
 495                                 default:
 496                                         q_push(q, de);
 497                                 }
 498                         } else
 499                                 q_push(q, de);
 500                 }
 501         }
 502
 503         q_del(q, e);
 504         e->level = new_level;
 505         q_push(q, e);
 506 }
 507
 508 /*----------------------------------------------------------------*/
 509
 510 #define FP_SHIFT 8
 511 #define SIXTEENTH (1u << (FP_SHIFT - 4u))
 512 #define EIGHTH (1u << (FP_SHIFT - 3u))
 513
 514 struct stats {
 515         unsigned hit_threshold;
 516         unsigned hits;
 517         unsigned misses;
 518 };
 519
 520 enum performance {
 521         Q_POOR,
 522         Q_FAIR,
 523         Q_WELL
 524 };
 525
 526 static void stats_init(struct stats *s, unsigned nr_levels)
 527 {
 528         s->hit_threshold = (nr_levels * 3u) / 4u;
 529         s->hits = 0u;
 530         s->misses = 0u;
 531 }
 532
 533 static void stats_reset(struct stats *s)
 534 {
 535         s->hits = s->misses = 0u;
 536 }
 537
 538 static void stats_level_accessed(struct stats *s, unsigned level)
 539 {
 540         if (level >= s->hit_threshold)
 541                 s->hits++;
 542         else
 543                 s->misses++;
 544 }
 545
 546 static void stats_miss(struct stats *s)
 547 {
 548         s->misses++;
 549 }
 550
 551 /*
 552  * There are times when we don't have any confidence in the hotspot queue.
 553  * Such as when a fresh cache is created and the blocks have been spread
 554  * out across the levels, or if an io load changes.  We detect this by
 555  * seeing how often a lookup is in the top levels of the hotspot queue.
 556  */
 557 static enum performance stats_assess(struct stats *s)
 558 {
 559         unsigned confidence = safe_div(s->hits << FP_SHIFT, s->hits + s->misses);
 560
 561         if (confidence < SIXTEENTH)
 562                 return Q_POOR;
 563
 564         else if (confidence < EIGHTH)
 565                 return Q_FAIR;
 566
 567         else
 568                 return Q_WELL;
 569 }
 570
 571 /*----------------------------------------------------------------*/
 572
 573 struct smq_hash_table {
 574         struct entry_space *es;
 575         unsigned long long hash_bits;
 576         unsigned *buckets;
 577 };
 578
 579 /*
 580  * All cache entries are stored in a chained hash table.  To save space we
 581  * use indexing again, and only store indexes to the next entry.
 582  */
 583 static int h_init(struct smq_hash_table *ht, struct entry_space *es, unsigned nr_entries)
 584 {
 585         unsigned i, nr_buckets;
 586
 587         ht->es = es;
 588         nr_buckets = roundup_pow_of_two(max(nr_entries / 4u, 16u));
 589         ht->hash_bits = __ffs(nr_buckets);
 590
 591         ht->buckets = vmalloc(array_size(nr_buckets, sizeof(*ht->buckets)));
 592         if (!ht->buckets)
 593                 return -ENOMEM;
 594
 595         for (i = 0; i < nr_buckets; i++)
 596                 ht->buckets[i] = INDEXER_NULL;
 597
 598         return 0;
 599 }
 600
 601 static void h_exit(struct smq_hash_table *ht)
 602 {
 603         vfree(ht->buckets);
 604 }
 605
 606 static struct entry *h_head(struct smq_hash_table *ht, unsigned bucket)
 607 {
 608         return to_entry(ht->es, ht->buckets[bucket]);
 609 }
 610
 611 static struct entry *h_next(struct smq_hash_table *ht, struct entry *e)
 612 {
 613         return to_entry(ht->es, e->hash_next);
 614 }
 615
 616 static void __h_insert(struct smq_hash_table *ht, unsigned bucket, struct entry *e)
 617 {
 618         e->hash_next = ht->buckets[bucket];
 619         ht->buckets[bucket] = to_index(ht->es, e);
 620 }
 621
 622 static void h_insert(struct smq_hash_table *ht, struct entry *e)
 623 {
 624         unsigned h = hash_64(from_oblock(e->oblock), ht->hash_bits);
 625         __h_insert(ht, h, e);
 626 }
 627
 628 static struct entry *__h_lookup(struct smq_hash_table *ht, unsigned h, dm_oblock_t oblock,
 629                                 struct entry **prev)
 630 {
 631         struct entry *e;
 632
 633         *prev = NULL;
 634         for (e = h_head(ht, h); e; e = h_next(ht, e)) {
 635                 if (e->oblock == oblock)
 636                         return e;
 637
 638                 *prev = e;
 639         }
 640
 641         return NULL;
 642 }
 643
 644 static void __h_unlink(struct smq_hash_table *ht, unsigned h,
 645                        struct entry *e, struct entry *prev)
 646 {
 647         if (prev)
 648                 prev->hash_next = e->hash_next;
 649         else
 650                 ht->buckets[h] = e->hash_next;
 651 }
 652
 653 /*
 654  * Also moves each entry to the front of the bucket.
 655  */
 656 static struct entry *h_lookup(struct smq_hash_table *ht, dm_oblock_t oblock)
 657 {
 658         struct entry *e, *prev;
 659         unsigned h = hash_64(from_oblock(oblock), ht->hash_bits);
 660
 661         e = __h_lookup(ht, h, oblock, &prev);
 662         if (e && prev) {
 663                 /*
 664                  * Move to the front because this entry is likely
 665                  * to be hit again.
 666                  */
 667                 __h_unlink(ht, h, e, prev);
 668                 __h_insert(ht, h, e);
 669         }
 670
 671         return e;
 672 }
 673
 674 static void h_remove(struct smq_hash_table *ht, struct entry *e)
 675 {
 676         unsigned h = hash_64(from_oblock(e->oblock), ht->hash_bits);
 677         struct entry *prev;
 678
 679         /*
 680          * The down side of using a singly linked list is we have to
 681          * iterate the bucket to remove an item.
 682          */
 683         e = __h_lookup(ht, h, e->oblock, &prev);
 684         if (e)
 685                 __h_unlink(ht, h, e, prev);
 686 }
 687
 688 /*----------------------------------------------------------------*/
 689
 690 struct entry_alloc {
 691         struct entry_space *es;
 692         unsigned begin;
 693
 694         unsigned nr_allocated;
 695         struct ilist free;
 696 };
 697
 698 static void init_allocator(struct entry_alloc *ea, struct entry_space *es,
 699                            unsigned begin, unsigned end)
 700 {
 701         unsigned i;
 702
 703         ea->es = es;
 704         ea->nr_allocated = 0u;
 705         ea->begin = begin;
 706
 707         l_init(&ea->free);
 708         for (i = begin; i != end; i++)
 709                 l_add_tail(ea->es, &ea->free, __get_entry(ea->es, i));
 710 }
 711
 712 static void init_entry(struct entry *e)
 713 {
 714         /*
 715          * We can't memset because that would clear the hotspot and
 716          * sentinel bits which remain constant.
 717          */
 718         e->hash_next = INDEXER_NULL;
 719         e->next = INDEXER_NULL;
 720         e->prev = INDEXER_NULL;
 721         e->level = 0u;
 722         e->dirty = true;        /* FIXME: audit */
 723         e->allocated = true;
 724         e->sentinel = false;
 725         e->pending_work = false;
 726 }
 727
 728 static struct entry *alloc_entry(struct entry_alloc *ea)
 729 {
 730         struct entry *e;
 731
 732         if (l_empty(&ea->free))
 733                 return NULL;
 734
 735         e = l_pop_head(ea->es, &ea->free);
 736         init_entry(e);
 737         ea->nr_allocated++;
 738
 739         return e;
 740 }
 741
 742 /*
 743  * This assumes the cblock hasn't already been allocated.
 744  */
 745 static struct entry *alloc_particular_entry(struct entry_alloc *ea, unsigned i)
 746 {
 747         struct entry *e = __get_entry(ea->es, ea->begin + i);
 748
 749         BUG_ON(e->allocated);
 750
 751         l_del(ea->es, &ea->free, e);
 752         init_entry(e);
 753         ea->nr_allocated++;
 754
 755         return e;
 756 }
 757
 758 static void free_entry(struct entry_alloc *ea, struct entry *e)
 759 {
 760         BUG_ON(!ea->nr_allocated);
 761         BUG_ON(!e->allocated);
 762
 763         ea->nr_allocated--;
 764         e->allocated = false;
 765         l_add_tail(ea->es, &ea->free, e);
 766 }
 767
 768 static bool allocator_empty(struct entry_alloc *ea)
 769 {
 770         return l_empty(&ea->free);
 771 }
 772
 773 static unsigned get_index(struct entry_alloc *ea, struct entry *e)
 774 {
 775         return to_index(ea->es, e) - ea->begin;
 776 }
 777
 778 static struct entry *get_entry(struct entry_alloc *ea, unsigned index)
 779 {
 780         return __get_entry(ea->es, ea->begin + index);
 781 }
 782
 783 /*----------------------------------------------------------------*/
 784
 785 #define NR_HOTSPOT_LEVELS 64u
 786 #define NR_CACHE_LEVELS 64u
 787
 788 #define WRITEBACK_PERIOD (10ul * HZ)
 789 #define DEMOTE_PERIOD (60ul * HZ)
 790
 791 #define HOTSPOT_UPDATE_PERIOD (HZ)
 792 #define CACHE_UPDATE_PERIOD (60ul * HZ)
 793
 794 struct smq_policy {
 795         struct dm_cache_policy policy;
 796
 797         /* protects everything */
 798         spinlock_t lock;
 799         dm_cblock_t cache_size;
 800         sector_t cache_block_size;
 801
 802         sector_t hotspot_block_size;
 803         unsigned nr_hotspot_blocks;
 804         unsigned cache_blocks_per_hotspot_block;
 805         unsigned hotspot_level_jump;
 806
 807         struct entry_space es;
 808         struct entry_alloc writeback_sentinel_alloc;
 809         struct entry_alloc demote_sentinel_alloc;
 810         struct entry_alloc hotspot_alloc;
 811         struct entry_alloc cache_alloc;
 812
 813         unsigned long *hotspot_hit_bits;
 814         unsigned long *cache_hit_bits;
 815
 816         /*
 817          * We maintain three queues of entries.  The cache proper,
 818          * consisting of a clean and dirty queue, containing the currently
 819          * active mappings.  The hotspot queue uses a larger block size to
 820          * track blocks that are being hit frequently and potential
 821          * candidates for promotion to the cache.
 822          */
 823         struct queue hotspot;
 824         struct queue clean;
 825         struct queue dirty;
 826
 827         struct stats hotspot_stats;
 828         struct stats cache_stats;
 829
 830         /*
 831          * Keeps track of time, incremented by the core.  We use this to
 832          * avoid attributing multiple hits within the same tick.
 833          */
 834         unsigned tick;
 835
 836         /*
 837          * The hash tables allows us to quickly find an entry by origin
 838          * block.
 839          */
 840         struct smq_hash_table table;
 841         struct smq_hash_table hotspot_table;
 842
 843         bool current_writeback_sentinels;
 844         unsigned long next_writeback_period;
 845
 846         bool current_demote_sentinels;
 847         unsigned long next_demote_period;
 848
 849         unsigned write_promote_level;
 850         unsigned read_promote_level;
 851
 852         unsigned long next_hotspot_period;
 853         unsigned long next_cache_period;
 854
 855         struct background_tracker *bg_work;
 856
 857         bool migrations_allowed;
 858 };
 859
 860 /*----------------------------------------------------------------*/
 861
 862 static struct entry *get_sentinel(struct entry_alloc *ea, unsigned level, bool which)
 863 {
 864         return get_entry(ea, which ? level : NR_CACHE_LEVELS + level);
 865 }
 866
 867 static struct entry *writeback_sentinel(struct smq_policy *mq, unsigned level)
 868 {
 869         return get_sentinel(&mq->writeback_sentinel_alloc, level, mq->current_writeback_sentinels);
 870 }
 871
 872 static struct entry *demote_sentinel(struct smq_policy *mq, unsigned level)
 873 {
 874         return get_sentinel(&mq->demote_sentinel_alloc, level, mq->current_demote_sentinels);
 875 }
 876
 877 static void __update_writeback_sentinels(struct smq_policy *mq)
 878 {
 879         unsigned level;
 880         struct queue *q = &mq->dirty;
 881         struct entry *sentinel;
 882
 883         for (level = 0; level < q->nr_levels; level++) {
 884                 sentinel = writeback_sentinel(mq, level);
 885                 q_del(q, sentinel);
 886                 q_push(q, sentinel);
 887         }
 888 }
 889
 890 static void __update_demote_sentinels(struct smq_policy *mq)
 891 {
 892         unsigned level;
 893         struct queue *q = &mq->clean;
 894         struct entry *sentinel;
 895
 896         for (level = 0; level < q->nr_levels; level++) {
 897                 sentinel = demote_sentinel(mq, level);
 898                 q_del(q, sentinel);
 899                 q_push(q, sentinel);
 900         }
 901 }
 902
 903 static void update_sentinels(struct smq_policy *mq)
 904 {
 905         if (time_after(jiffies, mq->next_writeback_period)) {
 906                 mq->next_writeback_period = jiffies + WRITEBACK_PERIOD;
 907                 mq->current_writeback_sentinels = !mq->current_writeback_sentinels;
 908                 __update_writeback_sentinels(mq);
 909         }
 910
 911         if (time_after(jiffies, mq->next_demote_period)) {
 912                 mq->next_demote_period = jiffies + DEMOTE_PERIOD;
 913                 mq->current_demote_sentinels = !mq->current_demote_sentinels;
 914                 __update_demote_sentinels(mq);
 915         }
 916 }
 917
 918 static void __sentinels_init(struct smq_policy *mq)
 919 {
 920         unsigned level;
 921         struct entry *sentinel;
 922
 923         for (level = 0; level < NR_CACHE_LEVELS; level++) {
 924                 sentinel = writeback_sentinel(mq, level);
 925                 sentinel->level = level;
 926                 q_push(&mq->dirty, sentinel);
 927
 928                 sentinel = demote_sentinel(mq, level);
 929                 sentinel->level = level;
 930                 q_push(&mq->clean, sentinel);
 931         }
 932 }
 933
 934 static void sentinels_init(struct smq_policy *mq)
 935 {
 936         mq->next_writeback_period = jiffies + WRITEBACK_PERIOD;
 937         mq->next_demote_period = jiffies + DEMOTE_PERIOD;
 938
 939         mq->current_writeback_sentinels = false;
 940         mq->current_demote_sentinels = false;
 941         __sentinels_init(mq);
 942
 943         mq->current_writeback_sentinels = !mq->current_writeback_sentinels;
 944         mq->current_demote_sentinels = !mq->current_demote_sentinels;
 945         __sentinels_init(mq);
 946 }
 947
 948 /*----------------------------------------------------------------*/
 949
 950 static void del_queue(struct smq_policy *mq, struct entry *e)
 951 {
 952         q_del(e->dirty ? &mq->dirty : &mq->clean, e);
 953 }
 954
 955 static void push_queue(struct smq_policy *mq, struct entry *e)
 956 {
 957         if (e->dirty)
 958                 q_push(&mq->dirty, e);
 959         else
 960                 q_push(&mq->clean, e);
 961 }
 962
 963 // !h, !q, a -> h, q, a
 964 static void push(struct smq_policy *mq, struct entry *e)
 965 {
 966         h_insert(&mq->table, e);
 967         if (!e->pending_work)
 968                 push_queue(mq, e);
 969 }
 970
 971 static void push_queue_front(struct smq_policy *mq, struct entry *e)
 972 {
 973         if (e->dirty)
 974                 q_push_front(&mq->dirty, e);
 975         else
 976                 q_push_front(&mq->clean, e);
 977 }
 978
 979 static void push_front(struct smq_policy *mq, struct entry *e)
 980 {
 981         h_insert(&mq->table, e);
 982         if (!e->pending_work)
 983                 push_queue_front(mq, e);
 984 }
 985
 986 static dm_cblock_t infer_cblock(struct smq_policy *mq, struct entry *e)
 987 {
 988         return to_cblock(get_index(&mq->cache_alloc, e));
 989 }
 990
 991 static void requeue(struct smq_policy *mq, struct entry *e)
 992 {
 993         /*
 994          * Pending work has temporarily been taken out of the queues.
 995          */
 996         if (e->pending_work)
 997                 return;
 998
 999         if (!test_and_set_bit(from_cblock(infer_cblock(mq, e)), mq->cache_hit_bits)) {
1000                 if (!e->dirty) {
1001                         q_requeue(&mq->clean, e, 1u, NULL, NULL);
1002                         return;
1003                 }
1004
1005                 q_requeue(&mq->dirty, e, 1u,
1006                           get_sentinel(&mq->writeback_sentinel_alloc, e->level, !mq->current_writeback_sentinels),
1007                           get_sentinel(&mq->writeback_sentinel_alloc, e->level, mq->current_writeback_sentinels));
1008         }
1009 }
1010
1011 static unsigned default_promote_level(struct smq_policy *mq)
1012 {
1013         /*
1014          * The promote level depends on the current performance of the
1015          * cache.
1016          *
1017          * If the cache is performing badly, then we can't afford
1018          * to promote much without causing performance to drop below that
1019          * of the origin device.
1020          *
1021          * If the cache is performing well, then we don't need to promote
1022          * much.  If it isn't broken, don't fix it.
1023          *
1024          * If the cache is middling then we promote more.
1025          *
1026          * This scheme reminds me of a graph of entropy vs probability of a
1027          * binary variable.
1028          */
1029         static const unsigned int table[] = {
1030                 1, 1, 1, 2, 4, 6, 7, 8, 7, 6, 4, 4, 3, 3, 2, 2, 1
1031         };
1032
1033         unsigned hits = mq->cache_stats.hits;
1034         unsigned misses = mq->cache_stats.misses;
1035         unsigned index = safe_div(hits << 4u, hits + misses);
1036         return table[index];
1037 }
1038
1039 static void update_promote_levels(struct smq_policy *mq)
1040 {
1041         /*
1042          * If there are unused cache entries then we want to be really
1043          * eager to promote.
1044          */
1045         unsigned threshold_level = allocator_empty(&mq->cache_alloc) ?
1046                 default_promote_level(mq) : (NR_HOTSPOT_LEVELS / 2u);
1047
1048         threshold_level = max(threshold_level, NR_HOTSPOT_LEVELS);
1049
1050         /*
1051          * If the hotspot queue is performing badly then we have little
1052          * confidence that we know which blocks to promote.  So we cut down
1053          * the amount of promotions.
1054          */
1055         switch (stats_assess(&mq->hotspot_stats)) {
1056         case Q_POOR:
1057                 threshold_level /= 4u;
1058                 break;
1059
1060         case Q_FAIR:
1061                 threshold_level /= 2u;
1062                 break;
1063
1064         case Q_WELL:
1065                 break;
1066         }
1067
1068         mq->read_promote_level = NR_HOTSPOT_LEVELS - threshold_level;
1069         mq->write_promote_level = (NR_HOTSPOT_LEVELS - threshold_level);
1070 }
1071
1072 /*
1073  * If the hotspot queue is performing badly, then we try and move entries
1074  * around more quickly.
1075  */
1076 static void update_level_jump(struct smq_policy *mq)
1077 {
1078         switch (stats_assess(&mq->hotspot_stats)) {
1079         case Q_POOR:
1080                 mq->hotspot_level_jump = 4u;
1081                 break;
1082
1083         case Q_FAIR:
1084                 mq->hotspot_level_jump = 2u;
1085                 break;
1086
1087         case Q_WELL:
1088                 mq->hotspot_level_jump = 1u;
1089                 break;
1090         }
1091 }
1092
1093 static void end_hotspot_period(struct smq_policy *mq)
1094 {
1095         clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
1096         update_promote_levels(mq);
1097
1098         if (time_after(jiffies, mq->next_hotspot_period)) {
1099                 update_level_jump(mq);
1100                 q_redistribute(&mq->hotspot);
1101                 stats_reset(&mq->hotspot_stats);
1102                 mq->next_hotspot_period = jiffies + HOTSPOT_UPDATE_PERIOD;
1103         }
1104 }
1105
1106 static void end_cache_period(struct smq_policy *mq)
1107 {
1108         if (time_after(jiffies, mq->next_cache_period)) {
1109                 clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
1110
1111                 q_redistribute(&mq->dirty);
1112                 q_redistribute(&mq->clean);
1113                 stats_reset(&mq->cache_stats);
1114
1115                 mq->next_cache_period = jiffies + CACHE_UPDATE_PERIOD;
1116         }
1117 }
1118
1119 /*----------------------------------------------------------------*/
1120
1121 /*
1122  * Targets are given as a percentage.
1123  */
1124 #define CLEAN_TARGET 25u
1125 #define FREE_TARGET 25u
1126
1127 static unsigned percent_to_target(struct smq_policy *mq, unsigned p)
1128 {
1129         return from_cblock(mq->cache_size) * p / 100u;
1130 }
1131
1132 static bool clean_target_met(struct smq_policy *mq, bool idle)
1133 {
1134         /*
1135          * Cache entries may not be populated.  So we cannot rely on the
1136          * size of the clean queue.
1137          */
1138         if (idle) {
1139                 /*
1140                  * We'd like to clean everything.
1141                  */
1142                 return q_size(&mq->dirty) == 0u;
1143         }
1144
1145         /*
1146          * If we're busy we don't worry about cleaning at all.
1147          */
1148         return true;
1149 }
1150
1151 static bool free_target_met(struct smq_policy *mq)
1152 {
1153         unsigned nr_free;
1154
1155         nr_free = from_cblock(mq->cache_size) - mq->cache_alloc.nr_allocated;
1156         return (nr_free + btracker_nr_demotions_queued(mq->bg_work)) >=
1157                 percent_to_target(mq, FREE_TARGET);
1158 }
1159
1160 /*----------------------------------------------------------------*/
1161
1162 static void mark_pending(struct smq_policy *mq, struct entry *e)
1163 {
1164         BUG_ON(e->sentinel);
1165         BUG_ON(!e->allocated);
1166         BUG_ON(e->pending_work);
1167         e->pending_work = true;
1168 }
1169
1170 static void clear_pending(struct smq_policy *mq, struct entry *e)
1171 {
1172         BUG_ON(!e->pending_work);
1173         e->pending_work = false;
1174 }
1175
1176 static void queue_writeback(struct smq_policy *mq, bool idle)
1177 {
1178         int r;
1179         struct policy_work work;
1180         struct entry *e;
1181
1182         e = q_peek(&mq->dirty, mq->dirty.nr_levels, idle);
1183         if (e) {
1184                 mark_pending(mq, e);
1185                 q_del(&mq->dirty, e);
1186
1187                 work.op = POLICY_WRITEBACK;
1188                 work.oblock = e->oblock;
1189                 work.cblock = infer_cblock(mq, e);
1190
1191                 r = btracker_queue(mq->bg_work, &work, NULL);
1192                 if (r) {
1193                         clear_pending(mq, e);
1194                         q_push_front(&mq->dirty, e);
1195                 }
1196         }
1197 }
1198
1199 static void queue_demotion(struct smq_policy *mq)
1200 {
1201         int r;
1202         struct policy_work work;
1203         struct entry *e;
1204
1205         if (WARN_ON_ONCE(!mq->migrations_allowed))
1206                 return;
1207
1208         e = q_peek(&mq->clean, mq->clean.nr_levels / 2, true);
1209         if (!e) {
1210                 if (!clean_target_met(mq, true))
1211                         queue_writeback(mq, false);
1212                 return;
1213         }
1214
1215         mark_pending(mq, e);
1216         q_del(&mq->clean, e);
1217
1218         work.op = POLICY_DEMOTE;
1219         work.oblock = e->oblock;
1220         work.cblock = infer_cblock(mq, e);
1221         r = btracker_queue(mq->bg_work, &work, NULL);
1222         if (r) {
1223                 clear_pending(mq, e);
1224                 q_push_front(&mq->clean, e);
1225         }
1226 }
1227
1228 static void queue_promotion(struct smq_policy *mq, dm_oblock_t oblock,
1229                             struct policy_work **workp)
1230 {
1231         int r;
1232         struct entry *e;
1233         struct policy_work work;
1234
1235         if (!mq->migrations_allowed)
1236                 return;
1237
1238         if (allocator_empty(&mq->cache_alloc)) {
1239                 /*
1240                  * We always claim to be 'idle' to ensure some demotions happen
1241                  * with continuous loads.
1242                  */
1243                 if (!free_target_met(mq))
1244                         queue_demotion(mq);
1245                 return;
1246         }
1247
1248         if (btracker_promotion_already_present(mq->bg_work, oblock))
1249                 return;
1250
1251         /*
1252          * We allocate the entry now to reserve the cblock.  If the
1253          * background work is aborted we must remember to free it.
1254          */
1255         e = alloc_entry(&mq->cache_alloc);
1256         BUG_ON(!e);
1257         e->pending_work = true;
1258         work.op = POLICY_PROMOTE;
1259         work.oblock = oblock;
1260         work.cblock = infer_cblock(mq, e);
1261         r = btracker_queue(mq->bg_work, &work, workp);
1262         if (r)
1263                 free_entry(&mq->cache_alloc, e);
1264 }
1265
1266 /*----------------------------------------------------------------*/
1267
1268 enum promote_result {
1269         PROMOTE_NOT,
1270         PROMOTE_TEMPORARY,
1271         PROMOTE_PERMANENT
1272 };
1273
1274 /*
1275  * Converts a boolean into a promote result.
1276  */
1277 static enum promote_result maybe_promote(bool promote)
1278 {
1279         return promote ? PROMOTE_PERMANENT : PROMOTE_NOT;
1280 }
1281
1282 static enum promote_result should_promote(struct smq_policy *mq, struct entry *hs_e,
1283                                           int data_dir, bool fast_promote)
1284 {
1285         if (data_dir == WRITE) {
1286                 if (!allocator_empty(&mq->cache_alloc) && fast_promote)
1287                         return PROMOTE_TEMPORARY;
1288
1289                 return maybe_promote(hs_e->level >= mq->write_promote_level);
1290         } else
1291                 return maybe_promote(hs_e->level >= mq->read_promote_level);
1292 }
1293
1294 static dm_oblock_t to_hblock(struct smq_policy *mq, dm_oblock_t b)
1295 {
1296         sector_t r = from_oblock(b);
1297         (void) sector_div(r, mq->cache_blocks_per_hotspot_block);
1298         return to_oblock(r);
1299 }
1300
1301 static struct entry *update_hotspot_queue(struct smq_policy *mq, dm_oblock_t b)
1302 {
1303         unsigned hi;
1304         dm_oblock_t hb = to_hblock(mq, b);
1305         struct entry *e = h_lookup(&mq->hotspot_table, hb);
1306
1307         if (e) {
1308                 stats_level_accessed(&mq->hotspot_stats, e->level);
1309
1310                 hi = get_index(&mq->hotspot_alloc, e);
1311                 q_requeue(&mq->hotspot, e,
1312                           test_and_set_bit(hi, mq->hotspot_hit_bits) ?
1313                           0u : mq->hotspot_level_jump,
1314                           NULL, NULL);
1315
1316         } else {
1317                 stats_miss(&mq->hotspot_stats);
1318
1319                 e = alloc_entry(&mq->hotspot_alloc);
1320                 if (!e) {
1321                         e = q_pop(&mq->hotspot);
1322                         if (e) {
1323                                 h_remove(&mq->hotspot_table, e);
1324                                 hi = get_index(&mq->hotspot_alloc, e);
1325                                 clear_bit(hi, mq->hotspot_hit_bits);
1326                         }
1327
1328                 }
1329
1330                 if (e) {
1331                         e->oblock = hb;
1332                         q_push(&mq->hotspot, e);
1333                         h_insert(&mq->hotspot_table, e);
1334                 }
1335         }
1336
1337         return e;
1338 }
1339
1340 /*----------------------------------------------------------------*/
1341
1342 /*
1343  * Public interface, via the policy struct.  See dm-cache-policy.h for a
1344  * description of these.
1345  */
1346
1347 static struct smq_policy *to_smq_policy(struct dm_cache_policy *p)
1348 {
1349         return container_of(p, struct smq_policy, policy);
1350 }
1351
1352 static void smq_destroy(struct dm_cache_policy *p)
1353 {
1354         struct smq_policy *mq = to_smq_policy(p);
1355
1356         btracker_destroy(mq->bg_work);
1357         h_exit(&mq->hotspot_table);
1358         h_exit(&mq->table);
1359         free_bitset(mq->hotspot_hit_bits);
1360         free_bitset(mq->cache_hit_bits);
1361         space_exit(&mq->es);
1362         kfree(mq);
1363 }
1364
1365 /*----------------------------------------------------------------*/
1366
1367 static int __lookup(struct smq_policy *mq, dm_oblock_t oblock, dm_cblock_t *cblock,
1368                     int data_dir, bool fast_copy,
1369                     struct policy_work **work, bool *background_work)
1370 {
1371         struct entry *e, *hs_e;
1372         enum promote_result pr;
1373
1374         *background_work = false;
1375
1376         e = h_lookup(&mq->table, oblock);
1377         if (e) {
1378                 stats_level_accessed(&mq->cache_stats, e->level);
1379
1380                 requeue(mq, e);
1381                 *cblock = infer_cblock(mq, e);
1382                 return 0;
1383
1384         } else {
1385                 stats_miss(&mq->cache_stats);
1386
1387                 /*
1388                  * The hotspot queue only gets updated with misses.
1389                  */
1390                 hs_e = update_hotspot_queue(mq, oblock);
1391
1392                 pr = should_promote(mq, hs_e, data_dir, fast_copy);
1393                 if (pr != PROMOTE_NOT) {
1394                         queue_promotion(mq, oblock, work);
1395                         *background_work = true;
1396                 }
1397
1398                 return -ENOENT;
1399         }
1400 }
1401
1402 static int smq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock,
1403                       int data_dir, bool fast_copy,
1404                       bool *background_work)
1405 {
1406         int r;
1407         unsigned long flags;
1408         struct smq_policy *mq = to_smq_policy(p);
1409
1410         spin_lock_irqsave(&mq->lock, flags);
1411         r = __lookup(mq, oblock, cblock,
1412                      data_dir, fast_copy,
1413                      NULL, background_work);
1414         spin_unlock_irqrestore(&mq->lock, flags);
1415
1416         return r;
1417 }
1418
1419 static int smq_lookup_with_work(struct dm_cache_policy *p,
1420                                 dm_oblock_t oblock, dm_cblock_t *cblock,
1421                                 int data_dir, bool fast_copy,
1422                                 struct policy_work **work)
1423 {
1424         int r;
1425         bool background_queued;
1426         unsigned long flags;
1427         struct smq_policy *mq = to_smq_policy(p);
1428
1429         spin_lock_irqsave(&mq->lock, flags);
1430         r = __lookup(mq, oblock, cblock, data_dir, fast_copy, work, &background_queued);
1431         spin_unlock_irqrestore(&mq->lock, flags);
1432
1433         return r;
1434 }
1435
1436 static int smq_get_background_work(struct dm_cache_policy *p, bool idle,
1437                                    struct policy_work **result)
1438 {
1439         int r;
1440         unsigned long flags;
1441         struct smq_policy *mq = to_smq_policy(p);
1442
1443         spin_lock_irqsave(&mq->lock, flags);
1444         r = btracker_issue(mq->bg_work, result);
1445         if (r == -ENODATA) {
1446                 if (!clean_target_met(mq, idle)) {
1447                         queue_writeback(mq, idle);
1448                         r = btracker_issue(mq->bg_work, result);
1449                 }
1450         }
1451         spin_unlock_irqrestore(&mq->lock, flags);
1452
1453         return r;
1454 }
1455
1456 /*
1457  * We need to clear any pending work flags that have been set, and in the
1458  * case of promotion free the entry for the destination cblock.
1459  */
1460 static void __complete_background_work(struct smq_policy *mq,
1461                                        struct policy_work *work,
1462                                        bool success)
1463 {
1464         struct entry *e = get_entry(&mq->cache_alloc,
1465                                     from_cblock(work->cblock));
1466
1467         switch (work->op) {
1468         case POLICY_PROMOTE:
1469                 // !h, !q, a
1470                 clear_pending(mq, e);
1471                 if (success) {
1472                         e->oblock = work->oblock;
1473                         e->level = NR_CACHE_LEVELS - 1;
1474                         push(mq, e);
1475                         // h, q, a
1476                 } else {
1477                         free_entry(&mq->cache_alloc, e);
1478                         // !h, !q, !a
1479                 }
1480                 break;
1481
1482         case POLICY_DEMOTE:
1483                 // h, !q, a
1484                 if (success) {
1485                         h_remove(&mq->table, e);
1486                         free_entry(&mq->cache_alloc, e);
1487                         // !h, !q, !a
1488                 } else {
1489                         clear_pending(mq, e);
1490                         push_queue(mq, e);
1491                         // h, q, a
1492                 }
1493                 break;
1494
1495         case POLICY_WRITEBACK:
1496                 // h, !q, a
1497                 clear_pending(mq, e);
1498                 push_queue(mq, e);
1499                 // h, q, a
1500                 break;
1501         }
1502
1503         btracker_complete(mq->bg_work, work);
1504 }
1505
1506 static void smq_complete_background_work(struct dm_cache_policy *p,
1507                                          struct policy_work *work,
1508                                          bool success)
1509 {
1510         unsigned long flags;
1511         struct smq_policy *mq = to_smq_policy(p);
1512
1513         spin_lock_irqsave(&mq->lock, flags);
1514         __complete_background_work(mq, work, success);
1515         spin_unlock_irqrestore(&mq->lock, flags);
1516 }
1517
1518 // in_hash(oblock) -> in_hash(oblock)
1519 static void __smq_set_clear_dirty(struct smq_policy *mq, dm_cblock_t cblock, bool set)
1520 {
1521         struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1522
1523         if (e->pending_work)
1524                 e->dirty = set;
1525         else {
1526                 del_queue(mq, e);
1527                 e->dirty = set;
1528                 push_queue(mq, e);
1529         }
1530 }
1531
1532 static void smq_set_dirty(struct dm_cache_policy *p, dm_cblock_t cblock)
1533 {
1534         unsigned long flags;
1535         struct smq_policy *mq = to_smq_policy(p);
1536
1537         spin_lock_irqsave(&mq->lock, flags);
1538         __smq_set_clear_dirty(mq, cblock, true);
1539         spin_unlock_irqrestore(&mq->lock, flags);
1540 }
1541
1542 static void smq_clear_dirty(struct dm_cache_policy *p, dm_cblock_t cblock)
1543 {
1544         struct smq_policy *mq = to_smq_policy(p);
1545         unsigned long flags;
1546
1547         spin_lock_irqsave(&mq->lock, flags);
1548         __smq_set_clear_dirty(mq, cblock, false);
1549         spin_unlock_irqrestore(&mq->lock, flags);
1550 }
1551
1552 static unsigned random_level(dm_cblock_t cblock)
1553 {
1554         return hash_32(from_cblock(cblock), 9) & (NR_CACHE_LEVELS - 1);
1555 }
1556
1557 static int smq_load_mapping(struct dm_cache_policy *p,
1558                             dm_oblock_t oblock, dm_cblock_t cblock,
1559                             bool dirty, uint32_t hint, bool hint_valid)
1560 {
1561         struct smq_policy *mq = to_smq_policy(p);
1562         struct entry *e;
1563
1564         e = alloc_particular_entry(&mq->cache_alloc, from_cblock(cblock));
1565         e->oblock = oblock;
1566         e->dirty = dirty;
1567         e->level = hint_valid ? min(hint, NR_CACHE_LEVELS - 1) : random_level(cblock);
1568         e->pending_work = false;
1569
1570         /*
1571          * When we load mappings we push ahead of both sentinels in order to
1572          * allow demotions and cleaning to occur immediately.
1573          */
1574         push_front(mq, e);
1575
1576         return 0;
1577 }
1578
1579 static int smq_invalidate_mapping(struct dm_cache_policy *p, dm_cblock_t cblock)
1580 {
1581         struct smq_policy *mq = to_smq_policy(p);
1582         struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1583
1584         if (!e->allocated)
1585                 return -ENODATA;
1586
1587         // FIXME: what if this block has pending background work?
1588         del_queue(mq, e);
1589         h_remove(&mq->table, e);
1590         free_entry(&mq->cache_alloc, e);
1591         return 0;
1592 }
1593
1594 static uint32_t smq_get_hint(struct dm_cache_policy *p, dm_cblock_t cblock)
1595 {
1596         struct smq_policy *mq = to_smq_policy(p);
1597         struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1598
1599         if (!e->allocated)
1600                 return 0;
1601
1602         return e->level;
1603 }
1604
1605 static dm_cblock_t smq_residency(struct dm_cache_policy *p)
1606 {
1607         dm_cblock_t r;
1608         unsigned long flags;
1609         struct smq_policy *mq = to_smq_policy(p);
1610
1611         spin_lock_irqsave(&mq->lock, flags);
1612         r = to_cblock(mq->cache_alloc.nr_allocated);
1613         spin_unlock_irqrestore(&mq->lock, flags);
1614
1615         return r;
1616 }
1617
1618 static void smq_tick(struct dm_cache_policy *p, bool can_block)
1619 {
1620         struct smq_policy *mq = to_smq_policy(p);
1621         unsigned long flags;
1622
1623         spin_lock_irqsave(&mq->lock, flags);
1624         mq->tick++;
1625         update_sentinels(mq);
1626         end_hotspot_period(mq);
1627         end_cache_period(mq);
1628         spin_unlock_irqrestore(&mq->lock, flags);
1629 }
1630
1631 static void smq_allow_migrations(struct dm_cache_policy *p, bool allow)
1632 {
1633         struct smq_policy *mq = to_smq_policy(p);
1634         mq->migrations_allowed = allow;
1635 }
1636
1637 /*
1638  * smq has no config values, but the old mq policy did.  To avoid breaking
1639  * software we continue to accept these configurables for the mq policy,
1640  * but they have no effect.
1641  */
1642 static int mq_set_config_value(struct dm_cache_policy *p,
1643                                const char *key, const char *value)
1644 {
1645         unsigned long tmp;
1646
1647         if (kstrtoul(value, 10, &tmp))
1648                 return -EINVAL;
1649
1650         if (!strcasecmp(key, "random_threshold") ||
1651             !strcasecmp(key, "sequential_threshold") ||
1652             !strcasecmp(key, "discard_promote_adjustment") ||
1653             !strcasecmp(key, "read_promote_adjustment") ||
1654             !strcasecmp(key, "write_promote_adjustment")) {
1655                 DMWARN("tunable '%s' no longer has any effect, mq policy is now an alias for smq", key);
1656                 return 0;
1657         }
1658
1659         return -EINVAL;
1660 }
1661
1662 static int mq_emit_config_values(struct dm_cache_policy *p, char *result,
1663                                  unsigned maxlen, ssize_t *sz_ptr)
1664 {
1665         ssize_t sz = *sz_ptr;
1666
1667         DMEMIT("10 random_threshold 0 "
1668                "sequential_threshold 0 "
1669                "discard_promote_adjustment 0 "
1670                "read_promote_adjustment 0 "
1671                "write_promote_adjustment 0 ");
1672
1673         *sz_ptr = sz;
1674         return 0;
1675 }
1676
1677 /* Init the policy plugin interface function pointers. */
1678 static void init_policy_functions(struct smq_policy *mq, bool mimic_mq)
1679 {
1680         mq->policy.destroy = smq_destroy;
1681         mq->policy.lookup = smq_lookup;
1682         mq->policy.lookup_with_work = smq_lookup_with_work;
1683         mq->policy.get_background_work = smq_get_background_work;
1684         mq->policy.complete_background_work = smq_complete_background_work;
1685         mq->policy.set_dirty = smq_set_dirty;
1686         mq->policy.clear_dirty = smq_clear_dirty;
1687         mq->policy.load_mapping = smq_load_mapping;
1688         mq->policy.invalidate_mapping = smq_invalidate_mapping;
1689         mq->policy.get_hint = smq_get_hint;
1690         mq->policy.residency = smq_residency;
1691         mq->policy.tick = smq_tick;
1692         mq->policy.allow_migrations = smq_allow_migrations;
1693
1694         if (mimic_mq) {
1695                 mq->policy.set_config_value = mq_set_config_value;
1696                 mq->policy.emit_config_values = mq_emit_config_values;
1697         }
1698 }
1699
1700 static bool too_many_hotspot_blocks(sector_t origin_size,
1701                                     sector_t hotspot_block_size,
1702                                     unsigned nr_hotspot_blocks)
1703 {
1704         return (hotspot_block_size * nr_hotspot_blocks) > origin_size;
1705 }
1706
1707 static void calc_hotspot_params(sector_t origin_size,
1708                                 sector_t cache_block_size,
1709                                 unsigned nr_cache_blocks,
1710                                 sector_t *hotspot_block_size,
1711                                 unsigned *nr_hotspot_blocks)
1712 {
1713         *hotspot_block_size = cache_block_size * 16u;
1714         *nr_hotspot_blocks = max(nr_cache_blocks / 4u, 1024u);
1715
1716         while ((*hotspot_block_size > cache_block_size) &&
1717                too_many_hotspot_blocks(origin_size, *hotspot_block_size, *nr_hotspot_blocks))
1718                 *hotspot_block_size /= 2u;
1719 }
1720
1721 static struct dm_cache_policy *__smq_create(dm_cblock_t cache_size,
1722                                             sector_t origin_size,
1723                                             sector_t cache_block_size,
1724                                             bool mimic_mq,
1725                                             bool migrations_allowed)
1726 {
1727         unsigned i;
1728         unsigned nr_sentinels_per_queue = 2u * NR_CACHE_LEVELS;
1729         unsigned total_sentinels = 2u * nr_sentinels_per_queue;
1730         struct smq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL);
1731
1732         if (!mq)
1733                 return NULL;
1734
1735         init_policy_functions(mq, mimic_mq);
1736         mq->cache_size = cache_size;
1737         mq->cache_block_size = cache_block_size;
1738
1739         calc_hotspot_params(origin_size, cache_block_size, from_cblock(cache_size),
1740                             &mq->hotspot_block_size, &mq->nr_hotspot_blocks);
1741
1742         mq->cache_blocks_per_hotspot_block = div64_u64(mq->hotspot_block_size, mq->cache_block_size);
1743         mq->hotspot_level_jump = 1u;
1744         if (space_init(&mq->es, total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size))) {
1745                 DMERR("couldn't initialize entry space");
1746                 goto bad_pool_init;
1747         }
1748
1749         init_allocator(&mq->writeback_sentinel_alloc, &mq->es, 0, nr_sentinels_per_queue);
1750         for (i = 0; i < nr_sentinels_per_queue; i++)
1751                 get_entry(&mq->writeback_sentinel_alloc, i)->sentinel = true;
1752
1753         init_allocator(&mq->demote_sentinel_alloc, &mq->es, nr_sentinels_per_queue, total_sentinels);
1754         for (i = 0; i < nr_sentinels_per_queue; i++)
1755                 get_entry(&mq->demote_sentinel_alloc, i)->sentinel = true;
1756
1757         init_allocator(&mq->hotspot_alloc, &mq->es, total_sentinels,
1758                        total_sentinels + mq->nr_hotspot_blocks);
1759
1760         init_allocator(&mq->cache_alloc, &mq->es,
1761                        total_sentinels + mq->nr_hotspot_blocks,
1762                        total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size));
1763
1764         mq->hotspot_hit_bits = alloc_bitset(mq->nr_hotspot_blocks);
1765         if (!mq->hotspot_hit_bits) {
1766                 DMERR("couldn't allocate hotspot hit bitset");
1767                 goto bad_hotspot_hit_bits;
1768         }
1769         clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
1770
1771         if (from_cblock(cache_size)) {
1772                 mq->cache_hit_bits = alloc_bitset(from_cblock(cache_size));
1773                 if (!mq->cache_hit_bits) {
1774                         DMERR("couldn't allocate cache hit bitset");
1775                         goto bad_cache_hit_bits;
1776                 }
1777                 clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
1778         } else
1779                 mq->cache_hit_bits = NULL;
1780
1781         mq->tick = 0;
1782         spin_lock_init(&mq->lock);
1783
1784         q_init(&mq->hotspot, &mq->es, NR_HOTSPOT_LEVELS);
1785         mq->hotspot.nr_top_levels = 8;
1786         mq->hotspot.nr_in_top_levels = min(mq->nr_hotspot_blocks / NR_HOTSPOT_LEVELS,
1787                                            from_cblock(mq->cache_size) / mq->cache_blocks_per_hotspot_block);
1788
1789         q_init(&mq->clean, &mq->es, NR_CACHE_LEVELS);
1790         q_init(&mq->dirty, &mq->es, NR_CACHE_LEVELS);
1791
1792         stats_init(&mq->hotspot_stats, NR_HOTSPOT_LEVELS);
1793         stats_init(&mq->cache_stats, NR_CACHE_LEVELS);
1794
1795         if (h_init(&mq->table, &mq->es, from_cblock(cache_size)))
1796                 goto bad_alloc_table;
1797
1798         if (h_init(&mq->hotspot_table, &mq->es, mq->nr_hotspot_blocks))
1799                 goto bad_alloc_hotspot_table;
1800
1801         sentinels_init(mq);
1802         mq->write_promote_level = mq->read_promote_level = NR_HOTSPOT_LEVELS;
1803
1804         mq->next_hotspot_period = jiffies;
1805         mq->next_cache_period = jiffies;
1806
1807         mq->bg_work = btracker_create(4096); /* FIXME: hard coded value */
1808         if (!mq->bg_work)
1809                 goto bad_btracker;
1810
1811         mq->migrations_allowed = migrations_allowed;
1812
1813         return &mq->policy;
1814
1815 bad_btracker:
1816         h_exit(&mq->hotspot_table);
1817 bad_alloc_hotspot_table:
1818         h_exit(&mq->table);
1819 bad_alloc_table:
1820         free_bitset(mq->cache_hit_bits);
1821 bad_cache_hit_bits:
1822         free_bitset(mq->hotspot_hit_bits);
1823 bad_hotspot_hit_bits:
1824         space_exit(&mq->es);
1825 bad_pool_init:
1826         kfree(mq);
1827
1828         return NULL;
1829 }
1830
1831 static struct dm_cache_policy *smq_create(dm_cblock_t cache_size,
1832                                           sector_t origin_size,
1833                                           sector_t cache_block_size)
1834 {
1835         return __smq_create(cache_size, origin_size, cache_block_size, false, true);
1836 }
1837
1838 static struct dm_cache_policy *mq_create(dm_cblock_t cache_size,
1839                                          sector_t origin_size,
1840                                          sector_t cache_block_size)
1841 {
1842         return __smq_create(cache_size, origin_size, cache_block_size, true, true);
1843 }
1844
1845 static struct dm_cache_policy *cleaner_create(dm_cblock_t cache_size,
1846                                               sector_t origin_size,
1847                                               sector_t cache_block_size)
1848 {
1849         return __smq_create(cache_size, origin_size, cache_block_size, false, false);
1850 }
1851
1852 /*----------------------------------------------------------------*/
1853
1854 static struct dm_cache_policy_type smq_policy_type = {
1855         .name = "smq",
1856         .version = {2, 0, 0},
1857         .hint_size = 4,
1858         .owner = THIS_MODULE,
1859         .create = smq_create
1860 };
1861
1862 static struct dm_cache_policy_type mq_policy_type = {
1863         .name = "mq",
1864         .version = {2, 0, 0},
1865         .hint_size = 4,
1866         .owner = THIS_MODULE,
1867         .create = mq_create,
1868 };
1869
1870 static struct dm_cache_policy_type cleaner_policy_type = {
1871         .name = "cleaner",
1872         .version = {2, 0, 0},
1873         .hint_size = 4,
1874         .owner = THIS_MODULE,
1875         .create = cleaner_create,
1876 };
1877
1878 static struct dm_cache_policy_type default_policy_type = {
1879         .name = "default",
1880         .version = {2, 0, 0},
1881         .hint_size = 4,
1882         .owner = THIS_MODULE,
1883         .create = smq_create,
1884         .real = &smq_policy_type
1885 };
1886
1887 static int __init smq_init(void)
1888 {
1889         int r;
1890
1891         r = dm_cache_policy_register(&smq_policy_type);
1892         if (r) {
1893                 DMERR("register failed %d", r);
1894                 return -ENOMEM;
1895         }
1896
1897         r = dm_cache_policy_register(&mq_policy_type);
1898         if (r) {
1899                 DMERR("register failed (as mq) %d", r);
1900                 goto out_mq;
1901         }
1902
1903         r = dm_cache_policy_register(&cleaner_policy_type);
1904         if (r) {
1905                 DMERR("register failed (as cleaner) %d", r);
1906                 goto out_cleaner;
1907         }
1908
1909         r = dm_cache_policy_register(&default_policy_type);
1910         if (r) {
1911                 DMERR("register failed (as default) %d", r);
1912                 goto out_default;
1913         }
1914
1915         return 0;
1916
1917 out_default:
1918         dm_cache_policy_unregister(&cleaner_policy_type);
1919 out_cleaner:
1920         dm_cache_policy_unregister(&mq_policy_type);
1921 out_mq:
1922         dm_cache_policy_unregister(&smq_policy_type);
1923
1924         return -ENOMEM;
1925 }
1926
1927 static void __exit smq_exit(void)
1928 {
1929         dm_cache_policy_unregister(&cleaner_policy_type);
1930         dm_cache_policy_unregister(&smq_policy_type);
1931         dm_cache_policy_unregister(&mq_policy_type);
1932         dm_cache_policy_unregister(&default_policy_type);
1933 }
1934
1935 module_init(smq_init);
1936 module_exit(smq_exit);
1937
1938 MODULE_AUTHOR("Joe Thornber <[email protected]>");
1939 MODULE_LICENSE("GPL");
1940 MODULE_DESCRIPTION("smq cache policy");
1941
1942 MODULE_ALIAS("dm-cache-default");
1943 MODULE_ALIAS("dm-cache-mq");
1944 MODULE_ALIAS("dm-cache-cleaner");