xref: /linux/drivers/md/dm-cache-policy-smq.c (revision fcc8487d477a3452a1d0ccbdd4c5e0e1e3cb8bed)
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 	unsigned nr_clean;
1124 
1125 	if (idle) {
1126 		/*
1127 		 * We'd like to clean everything.
1128 		 */
1129 		return q_size(&mq->dirty) == 0u;
1130 	}
1131 
1132 	nr_clean = from_cblock(mq->cache_size) - q_size(&mq->dirty);
1133 	return (nr_clean + btracker_nr_writebacks_queued(mq->bg_work)) >=
1134 		percent_to_target(mq, CLEAN_TARGET);
1135 }
1136 
1137 static bool free_target_met(struct smq_policy *mq, bool idle)
1138 {
1139 	unsigned nr_free;
1140 
1141 	if (!idle)
1142 		return true;
1143 
1144 	nr_free = from_cblock(mq->cache_size) - mq->cache_alloc.nr_allocated;
1145 	return (nr_free + btracker_nr_demotions_queued(mq->bg_work)) >=
1146 		percent_to_target(mq, FREE_TARGET);
1147 }
1148 
1149 /*----------------------------------------------------------------*/
1150 
1151 static void mark_pending(struct smq_policy *mq, struct entry *e)
1152 {
1153 	BUG_ON(e->sentinel);
1154 	BUG_ON(!e->allocated);
1155 	BUG_ON(e->pending_work);
1156 	e->pending_work = true;
1157 }
1158 
1159 static void clear_pending(struct smq_policy *mq, struct entry *e)
1160 {
1161 	BUG_ON(!e->pending_work);
1162 	e->pending_work = false;
1163 }
1164 
1165 static void queue_writeback(struct smq_policy *mq)
1166 {
1167 	int r;
1168 	struct policy_work work;
1169 	struct entry *e;
1170 
1171 	e = q_peek(&mq->dirty, mq->dirty.nr_levels, !mq->migrations_allowed);
1172 	if (e) {
1173 		mark_pending(mq, e);
1174 		q_del(&mq->dirty, e);
1175 
1176 		work.op = POLICY_WRITEBACK;
1177 		work.oblock = e->oblock;
1178 		work.cblock = infer_cblock(mq, e);
1179 
1180 		r = btracker_queue(mq->bg_work, &work, NULL);
1181 		WARN_ON_ONCE(r); // FIXME: finish, I think we have to get rid of this race.
1182 	}
1183 }
1184 
1185 static void queue_demotion(struct smq_policy *mq)
1186 {
1187 	struct policy_work work;
1188 	struct entry *e;
1189 
1190 	if (unlikely(WARN_ON_ONCE(!mq->migrations_allowed)))
1191 		return;
1192 
1193 	e = q_peek(&mq->clean, mq->clean.nr_levels, true);
1194 	if (!e) {
1195 		if (!clean_target_met(mq, false))
1196 			queue_writeback(mq);
1197 		return;
1198 	}
1199 
1200 	mark_pending(mq, e);
1201 	q_del(&mq->clean, e);
1202 
1203 	work.op = POLICY_DEMOTE;
1204 	work.oblock = e->oblock;
1205 	work.cblock = infer_cblock(mq, e);
1206 	btracker_queue(mq->bg_work, &work, NULL);
1207 }
1208 
1209 static void queue_promotion(struct smq_policy *mq, dm_oblock_t oblock,
1210 			    struct policy_work **workp)
1211 {
1212 	struct entry *e;
1213 	struct policy_work work;
1214 
1215 	if (!mq->migrations_allowed)
1216 		return;
1217 
1218 	if (allocator_empty(&mq->cache_alloc)) {
1219 		/*
1220 		 * We always claim to be 'idle' to ensure some demotions happen
1221 		 * with continuous loads.
1222 		 */
1223 		if (!free_target_met(mq, true))
1224 			queue_demotion(mq);
1225 		return;
1226 	}
1227 
1228 	if (btracker_promotion_already_present(mq->bg_work, oblock))
1229 		return;
1230 
1231 	/*
1232 	 * We allocate the entry now to reserve the cblock.  If the
1233 	 * background work is aborted we must remember to free it.
1234 	 */
1235 	e = alloc_entry(&mq->cache_alloc);
1236 	BUG_ON(!e);
1237 	e->pending_work = true;
1238 	work.op = POLICY_PROMOTE;
1239 	work.oblock = oblock;
1240 	work.cblock = infer_cblock(mq, e);
1241 	btracker_queue(mq->bg_work, &work, workp);
1242 }
1243 
1244 /*----------------------------------------------------------------*/
1245 
1246 enum promote_result {
1247 	PROMOTE_NOT,
1248 	PROMOTE_TEMPORARY,
1249 	PROMOTE_PERMANENT
1250 };
1251 
1252 /*
1253  * Converts a boolean into a promote result.
1254  */
1255 static enum promote_result maybe_promote(bool promote)
1256 {
1257 	return promote ? PROMOTE_PERMANENT : PROMOTE_NOT;
1258 }
1259 
1260 static enum promote_result should_promote(struct smq_policy *mq, struct entry *hs_e,
1261 					  int data_dir, bool fast_promote)
1262 {
1263 	if (data_dir == WRITE) {
1264 		if (!allocator_empty(&mq->cache_alloc) && fast_promote)
1265 			return PROMOTE_TEMPORARY;
1266 
1267 		return maybe_promote(hs_e->level >= mq->write_promote_level);
1268 	} else
1269 		return maybe_promote(hs_e->level >= mq->read_promote_level);
1270 }
1271 
1272 static dm_oblock_t to_hblock(struct smq_policy *mq, dm_oblock_t b)
1273 {
1274 	sector_t r = from_oblock(b);
1275 	(void) sector_div(r, mq->cache_blocks_per_hotspot_block);
1276 	return to_oblock(r);
1277 }
1278 
1279 static struct entry *update_hotspot_queue(struct smq_policy *mq, dm_oblock_t b)
1280 {
1281 	unsigned hi;
1282 	dm_oblock_t hb = to_hblock(mq, b);
1283 	struct entry *e = h_lookup(&mq->hotspot_table, hb);
1284 
1285 	if (e) {
1286 		stats_level_accessed(&mq->hotspot_stats, e->level);
1287 
1288 		hi = get_index(&mq->hotspot_alloc, e);
1289 		q_requeue(&mq->hotspot, e,
1290 			  test_and_set_bit(hi, mq->hotspot_hit_bits) ?
1291 			  0u : mq->hotspot_level_jump,
1292 			  NULL, NULL);
1293 
1294 	} else {
1295 		stats_miss(&mq->hotspot_stats);
1296 
1297 		e = alloc_entry(&mq->hotspot_alloc);
1298 		if (!e) {
1299 			e = q_pop(&mq->hotspot);
1300 			if (e) {
1301 				h_remove(&mq->hotspot_table, e);
1302 				hi = get_index(&mq->hotspot_alloc, e);
1303 				clear_bit(hi, mq->hotspot_hit_bits);
1304 			}
1305 
1306 		}
1307 
1308 		if (e) {
1309 			e->oblock = hb;
1310 			q_push(&mq->hotspot, e);
1311 			h_insert(&mq->hotspot_table, e);
1312 		}
1313 	}
1314 
1315 	return e;
1316 }
1317 
1318 /*----------------------------------------------------------------*/
1319 
1320 /*
1321  * Public interface, via the policy struct.  See dm-cache-policy.h for a
1322  * description of these.
1323  */
1324 
1325 static struct smq_policy *to_smq_policy(struct dm_cache_policy *p)
1326 {
1327 	return container_of(p, struct smq_policy, policy);
1328 }
1329 
1330 static void smq_destroy(struct dm_cache_policy *p)
1331 {
1332 	struct smq_policy *mq = to_smq_policy(p);
1333 
1334 	btracker_destroy(mq->bg_work);
1335 	h_exit(&mq->hotspot_table);
1336 	h_exit(&mq->table);
1337 	free_bitset(mq->hotspot_hit_bits);
1338 	free_bitset(mq->cache_hit_bits);
1339 	space_exit(&mq->es);
1340 	kfree(mq);
1341 }
1342 
1343 /*----------------------------------------------------------------*/
1344 
1345 static int __lookup(struct smq_policy *mq, dm_oblock_t oblock, dm_cblock_t *cblock,
1346 		    int data_dir, bool fast_copy,
1347 		    struct policy_work **work, bool *background_work)
1348 {
1349 	struct entry *e, *hs_e;
1350 	enum promote_result pr;
1351 
1352 	*background_work = false;
1353 
1354 	e = h_lookup(&mq->table, oblock);
1355 	if (e) {
1356 		stats_level_accessed(&mq->cache_stats, e->level);
1357 
1358 		requeue(mq, e);
1359 		*cblock = infer_cblock(mq, e);
1360 		return 0;
1361 
1362 	} else {
1363 		stats_miss(&mq->cache_stats);
1364 
1365 		/*
1366 		 * The hotspot queue only gets updated with misses.
1367 		 */
1368 		hs_e = update_hotspot_queue(mq, oblock);
1369 
1370 		pr = should_promote(mq, hs_e, data_dir, fast_copy);
1371 		if (pr != PROMOTE_NOT) {
1372 			queue_promotion(mq, oblock, work);
1373 			*background_work = true;
1374 		}
1375 
1376 		return -ENOENT;
1377 	}
1378 }
1379 
1380 static int smq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock,
1381 		      int data_dir, bool fast_copy,
1382 		      bool *background_work)
1383 {
1384 	int r;
1385 	unsigned long flags;
1386 	struct smq_policy *mq = to_smq_policy(p);
1387 
1388 	spin_lock_irqsave(&mq->lock, flags);
1389 	r = __lookup(mq, oblock, cblock,
1390 		     data_dir, fast_copy,
1391 		     NULL, background_work);
1392 	spin_unlock_irqrestore(&mq->lock, flags);
1393 
1394 	return r;
1395 }
1396 
1397 static int smq_lookup_with_work(struct dm_cache_policy *p,
1398 				dm_oblock_t oblock, dm_cblock_t *cblock,
1399 				int data_dir, bool fast_copy,
1400 				struct policy_work **work)
1401 {
1402 	int r;
1403 	bool background_queued;
1404 	unsigned long flags;
1405 	struct smq_policy *mq = to_smq_policy(p);
1406 
1407 	spin_lock_irqsave(&mq->lock, flags);
1408 	r = __lookup(mq, oblock, cblock, data_dir, fast_copy, work, &background_queued);
1409 	spin_unlock_irqrestore(&mq->lock, flags);
1410 
1411 	return r;
1412 }
1413 
1414 static int smq_get_background_work(struct dm_cache_policy *p, bool idle,
1415 				   struct policy_work **result)
1416 {
1417 	int r;
1418 	unsigned long flags;
1419 	struct smq_policy *mq = to_smq_policy(p);
1420 
1421 	spin_lock_irqsave(&mq->lock, flags);
1422 	r = btracker_issue(mq->bg_work, result);
1423 	if (r == -ENODATA) {
1424 		/* find some writeback work to do */
1425 		if (mq->migrations_allowed && !free_target_met(mq, idle))
1426 			queue_demotion(mq);
1427 
1428 		else if (!clean_target_met(mq, idle))
1429 			queue_writeback(mq);
1430 
1431 		r = btracker_issue(mq->bg_work, result);
1432 	}
1433 	spin_unlock_irqrestore(&mq->lock, flags);
1434 
1435 	return r;
1436 }
1437 
1438 /*
1439  * We need to clear any pending work flags that have been set, and in the
1440  * case of promotion free the entry for the destination cblock.
1441  */
1442 static void __complete_background_work(struct smq_policy *mq,
1443 				       struct policy_work *work,
1444 				       bool success)
1445 {
1446 	struct entry *e = get_entry(&mq->cache_alloc,
1447 				    from_cblock(work->cblock));
1448 
1449 	switch (work->op) {
1450 	case POLICY_PROMOTE:
1451 		// !h, !q, a
1452 		clear_pending(mq, e);
1453 		if (success) {
1454 			e->oblock = work->oblock;
1455 			push(mq, e);
1456 			// h, q, a
1457 		} else {
1458 			free_entry(&mq->cache_alloc, e);
1459 			// !h, !q, !a
1460 		}
1461 		break;
1462 
1463 	case POLICY_DEMOTE:
1464 		// h, !q, a
1465 		if (success) {
1466 			h_remove(&mq->table, e);
1467 			free_entry(&mq->cache_alloc, e);
1468 			// !h, !q, !a
1469 		} else {
1470 			clear_pending(mq, e);
1471 			push_queue(mq, e);
1472 			// h, q, a
1473 		}
1474 		break;
1475 
1476 	case POLICY_WRITEBACK:
1477 		// h, !q, a
1478 		clear_pending(mq, e);
1479 		push_queue(mq, e);
1480 		// h, q, a
1481 		break;
1482 	}
1483 
1484 	btracker_complete(mq->bg_work, work);
1485 }
1486 
1487 static void smq_complete_background_work(struct dm_cache_policy *p,
1488 					 struct policy_work *work,
1489 					 bool success)
1490 {
1491 	unsigned long flags;
1492 	struct smq_policy *mq = to_smq_policy(p);
1493 
1494 	spin_lock_irqsave(&mq->lock, flags);
1495 	__complete_background_work(mq, work, success);
1496 	spin_unlock_irqrestore(&mq->lock, flags);
1497 }
1498 
1499 // in_hash(oblock) -> in_hash(oblock)
1500 static void __smq_set_clear_dirty(struct smq_policy *mq, dm_cblock_t cblock, bool set)
1501 {
1502 	struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1503 
1504 	if (e->pending_work)
1505 		e->dirty = set;
1506 	else {
1507 		del_queue(mq, e);
1508 		e->dirty = set;
1509 		push_queue(mq, e);
1510 	}
1511 }
1512 
1513 static void smq_set_dirty(struct dm_cache_policy *p, dm_cblock_t cblock)
1514 {
1515 	unsigned long flags;
1516 	struct smq_policy *mq = to_smq_policy(p);
1517 
1518 	spin_lock_irqsave(&mq->lock, flags);
1519 	__smq_set_clear_dirty(mq, cblock, true);
1520 	spin_unlock_irqrestore(&mq->lock, flags);
1521 }
1522 
1523 static void smq_clear_dirty(struct dm_cache_policy *p, dm_cblock_t cblock)
1524 {
1525 	struct smq_policy *mq = to_smq_policy(p);
1526 	unsigned long flags;
1527 
1528 	spin_lock_irqsave(&mq->lock, flags);
1529 	__smq_set_clear_dirty(mq, cblock, false);
1530 	spin_unlock_irqrestore(&mq->lock, flags);
1531 }
1532 
1533 static unsigned random_level(dm_cblock_t cblock)
1534 {
1535 	return hash_32(from_cblock(cblock), 9) & (NR_CACHE_LEVELS - 1);
1536 }
1537 
1538 static int smq_load_mapping(struct dm_cache_policy *p,
1539 			    dm_oblock_t oblock, dm_cblock_t cblock,
1540 			    bool dirty, uint32_t hint, bool hint_valid)
1541 {
1542 	struct smq_policy *mq = to_smq_policy(p);
1543 	struct entry *e;
1544 
1545 	e = alloc_particular_entry(&mq->cache_alloc, from_cblock(cblock));
1546 	e->oblock = oblock;
1547 	e->dirty = dirty;
1548 	e->level = hint_valid ? min(hint, NR_CACHE_LEVELS - 1) : random_level(cblock);
1549 	e->pending_work = false;
1550 
1551 	/*
1552 	 * When we load mappings we push ahead of both sentinels in order to
1553 	 * allow demotions and cleaning to occur immediately.
1554 	 */
1555 	push_front(mq, e);
1556 
1557 	return 0;
1558 }
1559 
1560 static int smq_invalidate_mapping(struct dm_cache_policy *p, dm_cblock_t cblock)
1561 {
1562 	struct smq_policy *mq = to_smq_policy(p);
1563 	struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1564 
1565 	if (!e->allocated)
1566 		return -ENODATA;
1567 
1568 	// FIXME: what if this block has pending background work?
1569 	del_queue(mq, e);
1570 	h_remove(&mq->table, e);
1571 	free_entry(&mq->cache_alloc, e);
1572 	return 0;
1573 }
1574 
1575 static uint32_t smq_get_hint(struct dm_cache_policy *p, dm_cblock_t cblock)
1576 {
1577 	struct smq_policy *mq = to_smq_policy(p);
1578 	struct entry *e = get_entry(&mq->cache_alloc, from_cblock(cblock));
1579 
1580 	if (!e->allocated)
1581 		return 0;
1582 
1583 	return e->level;
1584 }
1585 
1586 static dm_cblock_t smq_residency(struct dm_cache_policy *p)
1587 {
1588 	dm_cblock_t r;
1589 	unsigned long flags;
1590 	struct smq_policy *mq = to_smq_policy(p);
1591 
1592 	spin_lock_irqsave(&mq->lock, flags);
1593 	r = to_cblock(mq->cache_alloc.nr_allocated);
1594 	spin_unlock_irqrestore(&mq->lock, flags);
1595 
1596 	return r;
1597 }
1598 
1599 static void smq_tick(struct dm_cache_policy *p, bool can_block)
1600 {
1601 	struct smq_policy *mq = to_smq_policy(p);
1602 	unsigned long flags;
1603 
1604 	spin_lock_irqsave(&mq->lock, flags);
1605 	mq->tick++;
1606 	update_sentinels(mq);
1607 	end_hotspot_period(mq);
1608 	end_cache_period(mq);
1609 	spin_unlock_irqrestore(&mq->lock, flags);
1610 }
1611 
1612 static void smq_allow_migrations(struct dm_cache_policy *p, bool allow)
1613 {
1614 	struct smq_policy *mq = to_smq_policy(p);
1615 	mq->migrations_allowed = allow;
1616 }
1617 
1618 /*
1619  * smq has no config values, but the old mq policy did.  To avoid breaking
1620  * software we continue to accept these configurables for the mq policy,
1621  * but they have no effect.
1622  */
1623 static int mq_set_config_value(struct dm_cache_policy *p,
1624 			       const char *key, const char *value)
1625 {
1626 	unsigned long tmp;
1627 
1628 	if (kstrtoul(value, 10, &tmp))
1629 		return -EINVAL;
1630 
1631 	if (!strcasecmp(key, "random_threshold") ||
1632 	    !strcasecmp(key, "sequential_threshold") ||
1633 	    !strcasecmp(key, "discard_promote_adjustment") ||
1634 	    !strcasecmp(key, "read_promote_adjustment") ||
1635 	    !strcasecmp(key, "write_promote_adjustment")) {
1636 		DMWARN("tunable '%s' no longer has any effect, mq policy is now an alias for smq", key);
1637 		return 0;
1638 	}
1639 
1640 	return -EINVAL;
1641 }
1642 
1643 static int mq_emit_config_values(struct dm_cache_policy *p, char *result,
1644 				 unsigned maxlen, ssize_t *sz_ptr)
1645 {
1646 	ssize_t sz = *sz_ptr;
1647 
1648 	DMEMIT("10 random_threshold 0 "
1649 	       "sequential_threshold 0 "
1650 	       "discard_promote_adjustment 0 "
1651 	       "read_promote_adjustment 0 "
1652 	       "write_promote_adjustment 0 ");
1653 
1654 	*sz_ptr = sz;
1655 	return 0;
1656 }
1657 
1658 /* Init the policy plugin interface function pointers. */
1659 static void init_policy_functions(struct smq_policy *mq, bool mimic_mq)
1660 {
1661 	mq->policy.destroy = smq_destroy;
1662 	mq->policy.lookup = smq_lookup;
1663 	mq->policy.lookup_with_work = smq_lookup_with_work;
1664 	mq->policy.get_background_work = smq_get_background_work;
1665 	mq->policy.complete_background_work = smq_complete_background_work;
1666 	mq->policy.set_dirty = smq_set_dirty;
1667 	mq->policy.clear_dirty = smq_clear_dirty;
1668 	mq->policy.load_mapping = smq_load_mapping;
1669 	mq->policy.invalidate_mapping = smq_invalidate_mapping;
1670 	mq->policy.get_hint = smq_get_hint;
1671 	mq->policy.residency = smq_residency;
1672 	mq->policy.tick = smq_tick;
1673 	mq->policy.allow_migrations = smq_allow_migrations;
1674 
1675 	if (mimic_mq) {
1676 		mq->policy.set_config_value = mq_set_config_value;
1677 		mq->policy.emit_config_values = mq_emit_config_values;
1678 	}
1679 }
1680 
1681 static bool too_many_hotspot_blocks(sector_t origin_size,
1682 				    sector_t hotspot_block_size,
1683 				    unsigned nr_hotspot_blocks)
1684 {
1685 	return (hotspot_block_size * nr_hotspot_blocks) > origin_size;
1686 }
1687 
1688 static void calc_hotspot_params(sector_t origin_size,
1689 				sector_t cache_block_size,
1690 				unsigned nr_cache_blocks,
1691 				sector_t *hotspot_block_size,
1692 				unsigned *nr_hotspot_blocks)
1693 {
1694 	*hotspot_block_size = cache_block_size * 16u;
1695 	*nr_hotspot_blocks = max(nr_cache_blocks / 4u, 1024u);
1696 
1697 	while ((*hotspot_block_size > cache_block_size) &&
1698 	       too_many_hotspot_blocks(origin_size, *hotspot_block_size, *nr_hotspot_blocks))
1699 		*hotspot_block_size /= 2u;
1700 }
1701 
1702 static struct dm_cache_policy *__smq_create(dm_cblock_t cache_size,
1703 					    sector_t origin_size,
1704 					    sector_t cache_block_size,
1705 					    bool mimic_mq,
1706 					    bool migrations_allowed)
1707 {
1708 	unsigned i;
1709 	unsigned nr_sentinels_per_queue = 2u * NR_CACHE_LEVELS;
1710 	unsigned total_sentinels = 2u * nr_sentinels_per_queue;
1711 	struct smq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL);
1712 
1713 	if (!mq)
1714 		return NULL;
1715 
1716 	init_policy_functions(mq, mimic_mq);
1717 	mq->cache_size = cache_size;
1718 	mq->cache_block_size = cache_block_size;
1719 
1720 	calc_hotspot_params(origin_size, cache_block_size, from_cblock(cache_size),
1721 			    &mq->hotspot_block_size, &mq->nr_hotspot_blocks);
1722 
1723 	mq->cache_blocks_per_hotspot_block = div64_u64(mq->hotspot_block_size, mq->cache_block_size);
1724 	mq->hotspot_level_jump = 1u;
1725 	if (space_init(&mq->es, total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size))) {
1726 		DMERR("couldn't initialize entry space");
1727 		goto bad_pool_init;
1728 	}
1729 
1730 	init_allocator(&mq->writeback_sentinel_alloc, &mq->es, 0, nr_sentinels_per_queue);
1731 	for (i = 0; i < nr_sentinels_per_queue; i++)
1732 		get_entry(&mq->writeback_sentinel_alloc, i)->sentinel = true;
1733 
1734 	init_allocator(&mq->demote_sentinel_alloc, &mq->es, nr_sentinels_per_queue, total_sentinels);
1735 	for (i = 0; i < nr_sentinels_per_queue; i++)
1736 		get_entry(&mq->demote_sentinel_alloc, i)->sentinel = true;
1737 
1738 	init_allocator(&mq->hotspot_alloc, &mq->es, total_sentinels,
1739 		       total_sentinels + mq->nr_hotspot_blocks);
1740 
1741 	init_allocator(&mq->cache_alloc, &mq->es,
1742 		       total_sentinels + mq->nr_hotspot_blocks,
1743 		       total_sentinels + mq->nr_hotspot_blocks + from_cblock(cache_size));
1744 
1745 	mq->hotspot_hit_bits = alloc_bitset(mq->nr_hotspot_blocks);
1746 	if (!mq->hotspot_hit_bits) {
1747 		DMERR("couldn't allocate hotspot hit bitset");
1748 		goto bad_hotspot_hit_bits;
1749 	}
1750 	clear_bitset(mq->hotspot_hit_bits, mq->nr_hotspot_blocks);
1751 
1752 	if (from_cblock(cache_size)) {
1753 		mq->cache_hit_bits = alloc_bitset(from_cblock(cache_size));
1754 		if (!mq->cache_hit_bits) {
1755 			DMERR("couldn't allocate cache hit bitset");
1756 			goto bad_cache_hit_bits;
1757 		}
1758 		clear_bitset(mq->cache_hit_bits, from_cblock(mq->cache_size));
1759 	} else
1760 		mq->cache_hit_bits = NULL;
1761 
1762 	mq->tick = 0;
1763 	spin_lock_init(&mq->lock);
1764 
1765 	q_init(&mq->hotspot, &mq->es, NR_HOTSPOT_LEVELS);
1766 	mq->hotspot.nr_top_levels = 8;
1767 	mq->hotspot.nr_in_top_levels = min(mq->nr_hotspot_blocks / NR_HOTSPOT_LEVELS,
1768 					   from_cblock(mq->cache_size) / mq->cache_blocks_per_hotspot_block);
1769 
1770 	q_init(&mq->clean, &mq->es, NR_CACHE_LEVELS);
1771 	q_init(&mq->dirty, &mq->es, NR_CACHE_LEVELS);
1772 
1773 	stats_init(&mq->hotspot_stats, NR_HOTSPOT_LEVELS);
1774 	stats_init(&mq->cache_stats, NR_CACHE_LEVELS);
1775 
1776 	if (h_init(&mq->table, &mq->es, from_cblock(cache_size)))
1777 		goto bad_alloc_table;
1778 
1779 	if (h_init(&mq->hotspot_table, &mq->es, mq->nr_hotspot_blocks))
1780 		goto bad_alloc_hotspot_table;
1781 
1782 	sentinels_init(mq);
1783 	mq->write_promote_level = mq->read_promote_level = NR_HOTSPOT_LEVELS;
1784 
1785 	mq->next_hotspot_period = jiffies;
1786 	mq->next_cache_period = jiffies;
1787 
1788 	mq->bg_work = btracker_create(10240); /* FIXME: hard coded value */
1789 	if (!mq->bg_work)
1790 		goto bad_btracker;
1791 
1792 	mq->migrations_allowed = migrations_allowed;
1793 
1794 	return &mq->policy;
1795 
1796 bad_btracker:
1797 	h_exit(&mq->hotspot_table);
1798 bad_alloc_hotspot_table:
1799 	h_exit(&mq->table);
1800 bad_alloc_table:
1801 	free_bitset(mq->cache_hit_bits);
1802 bad_cache_hit_bits:
1803 	free_bitset(mq->hotspot_hit_bits);
1804 bad_hotspot_hit_bits:
1805 	space_exit(&mq->es);
1806 bad_pool_init:
1807 	kfree(mq);
1808 
1809 	return NULL;
1810 }
1811 
1812 static struct dm_cache_policy *smq_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, false, true);
1817 }
1818 
1819 static struct dm_cache_policy *mq_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, true, true);
1824 }
1825 
1826 static struct dm_cache_policy *cleaner_create(dm_cblock_t cache_size,
1827 					      sector_t origin_size,
1828 					      sector_t cache_block_size)
1829 {
1830 	return __smq_create(cache_size, origin_size, cache_block_size, false, false);
1831 }
1832 
1833 /*----------------------------------------------------------------*/
1834 
1835 static struct dm_cache_policy_type smq_policy_type = {
1836 	.name = "smq",
1837 	.version = {2, 0, 0},
1838 	.hint_size = 4,
1839 	.owner = THIS_MODULE,
1840 	.create = smq_create
1841 };
1842 
1843 static struct dm_cache_policy_type mq_policy_type = {
1844 	.name = "mq",
1845 	.version = {2, 0, 0},
1846 	.hint_size = 4,
1847 	.owner = THIS_MODULE,
1848 	.create = mq_create,
1849 };
1850 
1851 static struct dm_cache_policy_type cleaner_policy_type = {
1852 	.name = "cleaner",
1853 	.version = {2, 0, 0},
1854 	.hint_size = 4,
1855 	.owner = THIS_MODULE,
1856 	.create = cleaner_create,
1857 };
1858 
1859 static struct dm_cache_policy_type default_policy_type = {
1860 	.name = "default",
1861 	.version = {2, 0, 0},
1862 	.hint_size = 4,
1863 	.owner = THIS_MODULE,
1864 	.create = smq_create,
1865 	.real = &smq_policy_type
1866 };
1867 
1868 static int __init smq_init(void)
1869 {
1870 	int r;
1871 
1872 	r = dm_cache_policy_register(&smq_policy_type);
1873 	if (r) {
1874 		DMERR("register failed %d", r);
1875 		return -ENOMEM;
1876 	}
1877 
1878 	r = dm_cache_policy_register(&mq_policy_type);
1879 	if (r) {
1880 		DMERR("register failed (as mq) %d", r);
1881 		goto out_mq;
1882 	}
1883 
1884 	r = dm_cache_policy_register(&cleaner_policy_type);
1885 	if (r) {
1886 		DMERR("register failed (as cleaner) %d", r);
1887 		goto out_cleaner;
1888 	}
1889 
1890 	r = dm_cache_policy_register(&default_policy_type);
1891 	if (r) {
1892 		DMERR("register failed (as default) %d", r);
1893 		goto out_default;
1894 	}
1895 
1896 	return 0;
1897 
1898 out_default:
1899 	dm_cache_policy_unregister(&cleaner_policy_type);
1900 out_cleaner:
1901 	dm_cache_policy_unregister(&mq_policy_type);
1902 out_mq:
1903 	dm_cache_policy_unregister(&smq_policy_type);
1904 
1905 	return -ENOMEM;
1906 }
1907 
1908 static void __exit smq_exit(void)
1909 {
1910 	dm_cache_policy_unregister(&cleaner_policy_type);
1911 	dm_cache_policy_unregister(&smq_policy_type);
1912 	dm_cache_policy_unregister(&mq_policy_type);
1913 	dm_cache_policy_unregister(&default_policy_type);
1914 }
1915 
1916 module_init(smq_init);
1917 module_exit(smq_exit);
1918 
1919 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
1920 MODULE_LICENSE("GPL");
1921 MODULE_DESCRIPTION("smq cache policy");
1922 
1923 MODULE_ALIAS("dm-cache-default");
1924 MODULE_ALIAS("dm-cache-mq");
1925 MODULE_ALIAS("dm-cache-cleaner");
1926