xref: /linux/drivers/md/dm-vdo/indexer/sparse-cache.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright 2023 Red Hat
4  */
5 
6 #include "sparse-cache.h"
7 
8 #include <linux/cache.h>
9 #include <linux/delay.h>
10 #include <linux/dm-bufio.h>
11 
12 #include "logger.h"
13 #include "memory-alloc.h"
14 #include "permassert.h"
15 
16 #include "chapter-index.h"
17 #include "config.h"
18 #include "index.h"
19 
20 /*
21  * Since the cache is small, it is implemented as a simple array of cache entries. Searching for a
22  * specific virtual chapter is implemented as a linear search. The cache replacement policy is
23  * least-recently-used (LRU). Again, the small size of the cache allows the LRU order to be
24  * maintained by shifting entries in an array list.
25  *
26  * Changing the contents of the cache requires the coordinated participation of all zone threads
27  * via the careful use of barrier messages sent to all the index zones by the triage queue worker
28  * thread. The critical invariant for coordination is that the cache membership must not change
29  * between updates, so that all calls to uds_sparse_cache_contains() from the zone threads must all
30  * receive the same results for every virtual chapter number. To ensure that critical invariant,
31  * state changes such as "that virtual chapter is no longer in the volume" and "skip searching that
32  * chapter because it has had too many cache misses" are represented separately from the cache
33  * membership information (the virtual chapter number).
34  *
35  * As a result of this invariant, we have the guarantee that every zone thread will call
36  * uds_update_sparse_cache() once and exactly once to request a chapter that is not in the cache,
37  * and the serialization of the barrier requests from the triage queue ensures they will all
38  * request the same chapter number. This means the only synchronization we need can be provided by
39  * a pair of thread barriers used only in the uds_update_sparse_cache() call, providing a critical
40  * section where a single zone thread can drive the cache update while all the other zone threads
41  * are known to be blocked, waiting in the second barrier. Outside that critical section, all the
42  * zone threads implicitly hold a shared lock. Inside it, the thread for zone zero holds an
43  * exclusive lock. No other threads may access or modify the cache entries.
44  *
45  * Chapter statistics must only be modified by a single thread, which is also the zone zero thread.
46  * All fields that might be frequently updated by that thread are kept in separate cache-aligned
47  * structures so they will not cause cache contention via "false sharing" with the fields that are
48  * frequently accessed by all of the zone threads.
49  *
50  * The LRU order is managed independently by each zone thread, and each zone uses its own list for
51  * searching and cache membership queries. The zone zero list is used to decide which chapter to
52  * evict when the cache is updated, and its search list is copied to the other threads at that
53  * time.
54  *
55  * The virtual chapter number field of the cache entry is the single field indicating whether a
56  * chapter is a member of the cache or not. The value NO_CHAPTER is used to represent a null or
57  * undefined chapter number. When present in the virtual chapter number field of a
58  * cached_chapter_index, it indicates that the cache entry is dead, and all the other fields of
59  * that entry (other than immutable pointers to cache memory) are undefined and irrelevant. Any
60  * cache entry that is not marked as dead is fully defined and a member of the cache, and
61  * uds_sparse_cache_contains() will always return true for any virtual chapter number that appears
62  * in any of the cache entries.
63  *
64  * A chapter index that is a member of the cache may be excluded from searches between calls to
65  * uds_update_sparse_cache() in two different ways. First, when a chapter falls off the end of the
66  * volume, its virtual chapter number will be less that the oldest virtual chapter number. Since
67  * that chapter is no longer part of the volume, there's no point in continuing to search that
68  * chapter index. Once invalidated, that virtual chapter will still be considered a member of the
69  * cache, but it will no longer be searched for matching names.
70  *
71  * The second mechanism is a heuristic based on keeping track of the number of consecutive search
72  * misses in a given chapter index. Once that count exceeds a threshold, the skip_search flag will
73  * be set to true, causing the chapter to be skipped when searching the entire cache, but still
74  * allowing it to be found when searching for a hook in that specific chapter. Finding a hook will
75  * clear the skip_search flag, once again allowing the non-hook searches to use that cache entry.
76  * Again, regardless of the state of the skip_search flag, the virtual chapter must still
77  * considered to be a member of the cache for uds_sparse_cache_contains().
78  */
79 
80 #define SKIP_SEARCH_THRESHOLD 20000
81 #define ZONE_ZERO 0
82 
83 /*
84  * These counters are essentially fields of the struct cached_chapter_index, but are segregated
85  * into this structure because they are frequently modified. They are grouped and aligned to keep
86  * them on different cache lines from the chapter fields that are accessed far more often than they
87  * are updated.
88  */
89 struct __aligned(L1_CACHE_BYTES) cached_index_counters {
90 	u64 consecutive_misses;
91 };
92 
93 struct __aligned(L1_CACHE_BYTES) cached_chapter_index {
94 	/*
95 	 * The virtual chapter number of the cached chapter index. NO_CHAPTER means this cache
96 	 * entry is unused. This field must only be modified in the critical section in
97 	 * uds_update_sparse_cache().
98 	 */
99 	u64 virtual_chapter;
100 
101 	u32 index_pages_count;
102 
103 	/*
104 	 * These pointers are immutable during the life of the cache. The contents of the arrays
105 	 * change when the cache entry is replaced.
106 	 */
107 	struct delta_index_page *index_pages;
108 	struct dm_buffer **page_buffers;
109 
110 	/*
111 	 * If set, skip the chapter when searching the entire cache. This flag is just a
112 	 * performance optimization. This flag is mutable between cache updates, but it rarely
113 	 * changes and is frequently accessed, so it groups with the immutable fields.
114 	 */
115 	bool skip_search;
116 
117 	/*
118 	 * The cache-aligned counters change often and are placed at the end of the structure to
119 	 * prevent false sharing with the more stable fields above.
120 	 */
121 	struct cached_index_counters counters;
122 };
123 
124 /*
125  * A search_list represents an ordering of the sparse chapter index cache entry array, from most
126  * recently accessed to least recently accessed, which is the order in which the indexes should be
127  * searched and the reverse order in which they should be evicted from the cache.
128  *
129  * Cache entries that are dead or empty are kept at the end of the list, avoiding the need to even
130  * iterate over them to search, and ensuring that dead entries are replaced before any live entries
131  * are evicted.
132  *
133  * The search list is instantiated for each zone thread, avoiding any need for synchronization. The
134  * structure is allocated on a cache boundary to avoid false sharing of memory cache lines between
135  * zone threads.
136  */
137 struct search_list {
138 	u8 capacity;
139 	u8 first_dead_entry;
140 	struct cached_chapter_index *entries[];
141 };
142 
143 struct threads_barrier {
144 	/* Lock for this barrier object */
145 	struct semaphore lock;
146 	/* Semaphore for threads waiting at this barrier */
147 	struct semaphore wait;
148 	/* Number of threads which have arrived */
149 	int arrived;
150 	/* Total number of threads using this barrier */
151 	int thread_count;
152 };
153 
154 struct sparse_cache {
155 	const struct index_geometry *geometry;
156 	unsigned int capacity;
157 	unsigned int zone_count;
158 
159 	unsigned int skip_threshold;
160 	struct search_list *search_lists[MAX_ZONES];
161 	struct cached_chapter_index **scratch_entries;
162 
163 	struct threads_barrier begin_update_barrier;
164 	struct threads_barrier end_update_barrier;
165 
166 	struct cached_chapter_index chapters[];
167 };
168 
169 static void initialize_threads_barrier(struct threads_barrier *barrier,
170 				       unsigned int thread_count)
171 {
172 	sema_init(&barrier->lock, 1);
173 	barrier->arrived = 0;
174 	barrier->thread_count = thread_count;
175 	sema_init(&barrier->wait, 0);
176 }
177 
178 static inline void __down(struct semaphore *semaphore)
179 {
180 	/*
181 	 * Do not use down(semaphore). Instead use down_interruptible so that
182 	 * we do not get 120 second stall messages in kern.log.
183 	 */
184 	while (down_interruptible(semaphore) != 0) {
185 		/*
186 		 * If we're called from a user-mode process (e.g., "dmsetup
187 		 * remove") while waiting for an operation that may take a
188 		 * while (e.g., UDS index save), and a signal is sent (SIGINT,
189 		 * SIGUSR2), then down_interruptible will not block. If that
190 		 * happens, sleep briefly to avoid keeping the CPU locked up in
191 		 * this loop. We could just call cond_resched, but then we'd
192 		 * still keep consuming CPU time slices and swamp other threads
193 		 * trying to do computational work.
194 		 */
195 		fsleep(1000);
196 	}
197 }
198 
199 static void enter_threads_barrier(struct threads_barrier *barrier)
200 {
201 	__down(&barrier->lock);
202 	if (++barrier->arrived == barrier->thread_count) {
203 		/* last thread */
204 		int i;
205 
206 		for (i = 1; i < barrier->thread_count; i++)
207 			up(&barrier->wait);
208 
209 		barrier->arrived = 0;
210 		up(&barrier->lock);
211 	} else {
212 		up(&barrier->lock);
213 		__down(&barrier->wait);
214 	}
215 }
216 
217 static int __must_check initialize_cached_chapter_index(struct cached_chapter_index *chapter,
218 							const struct index_geometry *geometry)
219 {
220 	int result;
221 
222 	chapter->virtual_chapter = NO_CHAPTER;
223 	chapter->index_pages_count = geometry->index_pages_per_chapter;
224 
225 	result = vdo_allocate(chapter->index_pages_count, struct delta_index_page,
226 			      __func__, &chapter->index_pages);
227 	if (result != VDO_SUCCESS)
228 		return result;
229 
230 	return vdo_allocate(chapter->index_pages_count, struct dm_buffer *,
231 			    "sparse index volume pages", &chapter->page_buffers);
232 }
233 
234 static int __must_check make_search_list(struct sparse_cache *cache,
235 					 struct search_list **list_ptr)
236 {
237 	struct search_list *list;
238 	unsigned int bytes;
239 	u8 i;
240 	int result;
241 
242 	bytes = (sizeof(struct search_list) +
243 		 (cache->capacity * sizeof(struct cached_chapter_index *)));
244 	result = vdo_allocate_cache_aligned(bytes, "search list", &list);
245 	if (result != VDO_SUCCESS)
246 		return result;
247 
248 	list->capacity = cache->capacity;
249 	list->first_dead_entry = 0;
250 
251 	for (i = 0; i < list->capacity; i++)
252 		list->entries[i] = &cache->chapters[i];
253 
254 	*list_ptr = list;
255 	return UDS_SUCCESS;
256 }
257 
258 int uds_make_sparse_cache(const struct index_geometry *geometry, unsigned int capacity,
259 			  unsigned int zone_count, struct sparse_cache **cache_ptr)
260 {
261 	int result;
262 	unsigned int i;
263 	struct sparse_cache *cache;
264 	unsigned int bytes;
265 
266 	bytes = (sizeof(struct sparse_cache) + (capacity * sizeof(struct cached_chapter_index)));
267 	result = vdo_allocate_cache_aligned(bytes, "sparse cache", &cache);
268 	if (result != VDO_SUCCESS)
269 		return result;
270 
271 	cache->geometry = geometry;
272 	cache->capacity = capacity;
273 	cache->zone_count = zone_count;
274 
275 	/*
276 	 * Scale down the skip threshold since the cache only counts cache misses in zone zero, but
277 	 * requests are being handled in all zones.
278 	 */
279 	cache->skip_threshold = (SKIP_SEARCH_THRESHOLD / zone_count);
280 
281 	initialize_threads_barrier(&cache->begin_update_barrier, zone_count);
282 	initialize_threads_barrier(&cache->end_update_barrier, zone_count);
283 
284 	for (i = 0; i < capacity; i++) {
285 		result = initialize_cached_chapter_index(&cache->chapters[i], geometry);
286 		if (result != UDS_SUCCESS)
287 			goto out;
288 	}
289 
290 	for (i = 0; i < zone_count; i++) {
291 		result = make_search_list(cache, &cache->search_lists[i]);
292 		if (result != UDS_SUCCESS)
293 			goto out;
294 	}
295 
296 	/* purge_search_list() needs some temporary lists for sorting. */
297 	result = vdo_allocate(capacity * 2, struct cached_chapter_index *,
298 			      "scratch entries", &cache->scratch_entries);
299 	if (result != VDO_SUCCESS)
300 		goto out;
301 
302 	*cache_ptr = cache;
303 	return UDS_SUCCESS;
304 out:
305 	uds_free_sparse_cache(cache);
306 	return result;
307 }
308 
309 static inline void set_skip_search(struct cached_chapter_index *chapter,
310 				   bool skip_search)
311 {
312 	/* Check before setting to reduce cache line contention. */
313 	if (READ_ONCE(chapter->skip_search) != skip_search)
314 		WRITE_ONCE(chapter->skip_search, skip_search);
315 }
316 
317 static void score_search_hit(struct cached_chapter_index *chapter)
318 {
319 	chapter->counters.consecutive_misses = 0;
320 	set_skip_search(chapter, false);
321 }
322 
323 static void score_search_miss(struct sparse_cache *cache,
324 			      struct cached_chapter_index *chapter)
325 {
326 	chapter->counters.consecutive_misses++;
327 	if (chapter->counters.consecutive_misses > cache->skip_threshold)
328 		set_skip_search(chapter, true);
329 }
330 
331 static void release_cached_chapter_index(struct cached_chapter_index *chapter)
332 {
333 	unsigned int i;
334 
335 	chapter->virtual_chapter = NO_CHAPTER;
336 	if (chapter->page_buffers == NULL)
337 		return;
338 
339 	for (i = 0; i < chapter->index_pages_count; i++) {
340 		if (chapter->page_buffers[i] != NULL)
341 			dm_bufio_release(vdo_forget(chapter->page_buffers[i]));
342 	}
343 }
344 
345 void uds_free_sparse_cache(struct sparse_cache *cache)
346 {
347 	unsigned int i;
348 
349 	if (cache == NULL)
350 		return;
351 
352 	vdo_free(cache->scratch_entries);
353 
354 	for (i = 0; i < cache->zone_count; i++)
355 		vdo_free(cache->search_lists[i]);
356 
357 	for (i = 0; i < cache->capacity; i++) {
358 		release_cached_chapter_index(&cache->chapters[i]);
359 		vdo_free(cache->chapters[i].index_pages);
360 		vdo_free(cache->chapters[i].page_buffers);
361 	}
362 
363 	vdo_free(cache);
364 }
365 
366 /*
367  * Take the indicated element of the search list and move it to the start, pushing the pointers
368  * previously before it back down the list.
369  */
370 static inline void set_newest_entry(struct search_list *search_list, u8 index)
371 {
372 	struct cached_chapter_index *newest;
373 
374 	if (index > 0) {
375 		newest = search_list->entries[index];
376 		memmove(&search_list->entries[1], &search_list->entries[0],
377 			index * sizeof(struct cached_chapter_index *));
378 		search_list->entries[0] = newest;
379 	}
380 
381 	/*
382 	 * This function may have moved a dead chapter to the front of the list for reuse, in which
383 	 * case the set of dead chapters becomes smaller.
384 	 */
385 	if (search_list->first_dead_entry <= index)
386 		search_list->first_dead_entry++;
387 }
388 
389 bool uds_sparse_cache_contains(struct sparse_cache *cache, u64 virtual_chapter,
390 			       unsigned int zone_number)
391 {
392 	struct search_list *search_list;
393 	struct cached_chapter_index *chapter;
394 	u8 i;
395 
396 	/*
397 	 * The correctness of the barriers depends on the invariant that between calls to
398 	 * uds_update_sparse_cache(), the answers this function returns must never vary: the result
399 	 * for a given chapter must be identical across zones. That invariant must be maintained
400 	 * even if the chapter falls off the end of the volume, or if searching it is disabled
401 	 * because of too many search misses.
402 	 */
403 	search_list = cache->search_lists[zone_number];
404 	for (i = 0; i < search_list->first_dead_entry; i++) {
405 		chapter = search_list->entries[i];
406 
407 		if (virtual_chapter == chapter->virtual_chapter) {
408 			if (zone_number == ZONE_ZERO)
409 				score_search_hit(chapter);
410 
411 			set_newest_entry(search_list, i);
412 			return true;
413 		}
414 	}
415 
416 	return false;
417 }
418 
419 /*
420  * Re-sort cache entries into three sets (active, skippable, and dead) while maintaining the LRU
421  * ordering that already existed. This operation must only be called during the critical section in
422  * uds_update_sparse_cache().
423  */
424 static void purge_search_list(struct search_list *search_list,
425 			      struct sparse_cache *cache, u64 oldest_virtual_chapter)
426 {
427 	struct cached_chapter_index **entries;
428 	struct cached_chapter_index **skipped;
429 	struct cached_chapter_index **dead;
430 	struct cached_chapter_index *chapter;
431 	unsigned int next_alive = 0;
432 	unsigned int next_skipped = 0;
433 	unsigned int next_dead = 0;
434 	unsigned int i;
435 
436 	entries = &search_list->entries[0];
437 	skipped = &cache->scratch_entries[0];
438 	dead = &cache->scratch_entries[search_list->capacity];
439 
440 	for (i = 0; i < search_list->first_dead_entry; i++) {
441 		chapter = search_list->entries[i];
442 		if ((chapter->virtual_chapter < oldest_virtual_chapter) ||
443 		    (chapter->virtual_chapter == NO_CHAPTER))
444 			dead[next_dead++] = chapter;
445 		else if (chapter->skip_search)
446 			skipped[next_skipped++] = chapter;
447 		else
448 			entries[next_alive++] = chapter;
449 	}
450 
451 	memcpy(&entries[next_alive], skipped,
452 	       next_skipped * sizeof(struct cached_chapter_index *));
453 	memcpy(&entries[next_alive + next_skipped], dead,
454 	       next_dead * sizeof(struct cached_chapter_index *));
455 	search_list->first_dead_entry = next_alive + next_skipped;
456 }
457 
458 static int __must_check cache_chapter_index(struct cached_chapter_index *chapter,
459 					    u64 virtual_chapter,
460 					    const struct volume *volume)
461 {
462 	int result;
463 
464 	release_cached_chapter_index(chapter);
465 
466 	result = uds_read_chapter_index_from_volume(volume, virtual_chapter,
467 						    chapter->page_buffers,
468 						    chapter->index_pages);
469 	if (result != UDS_SUCCESS)
470 		return result;
471 
472 	chapter->counters.consecutive_misses = 0;
473 	chapter->virtual_chapter = virtual_chapter;
474 	chapter->skip_search = false;
475 
476 	return UDS_SUCCESS;
477 }
478 
479 static inline void copy_search_list(const struct search_list *source,
480 				    struct search_list *target)
481 {
482 	*target = *source;
483 	memcpy(target->entries, source->entries,
484 	       source->capacity * sizeof(struct cached_chapter_index *));
485 }
486 
487 /*
488  * Update the sparse cache to contain a chapter index. This function must be called by all the zone
489  * threads with the same chapter number to correctly enter the thread barriers used to synchronize
490  * the cache updates.
491  */
492 int uds_update_sparse_cache(struct index_zone *zone, u64 virtual_chapter)
493 {
494 	int result = UDS_SUCCESS;
495 	const struct uds_index *index = zone->index;
496 	struct sparse_cache *cache = index->volume->sparse_cache;
497 
498 	if (uds_sparse_cache_contains(cache, virtual_chapter, zone->id))
499 		return UDS_SUCCESS;
500 
501 	/*
502 	 * Wait for every zone thread to reach its corresponding barrier request and invoke this
503 	 * function before starting to modify the cache.
504 	 */
505 	enter_threads_barrier(&cache->begin_update_barrier);
506 
507 	/*
508 	 * This is the start of the critical section: the zone zero thread is captain, effectively
509 	 * holding an exclusive lock on the sparse cache. All the other zone threads must do
510 	 * nothing between the two barriers. They will wait at the end_update_barrier again for the
511 	 * captain to finish the update.
512 	 */
513 
514 	if (zone->id == ZONE_ZERO) {
515 		unsigned int z;
516 		struct search_list *list = cache->search_lists[ZONE_ZERO];
517 
518 		purge_search_list(list, cache, zone->oldest_virtual_chapter);
519 
520 		if (virtual_chapter >= index->oldest_virtual_chapter) {
521 			set_newest_entry(list, list->capacity - 1);
522 			result = cache_chapter_index(list->entries[0], virtual_chapter,
523 						     index->volume);
524 		}
525 
526 		for (z = 1; z < cache->zone_count; z++)
527 			copy_search_list(list, cache->search_lists[z]);
528 	}
529 
530 	/*
531 	 * This is the end of the critical section. All cache invariants must have been restored.
532 	 */
533 	enter_threads_barrier(&cache->end_update_barrier);
534 	return result;
535 }
536 
537 void uds_invalidate_sparse_cache(struct sparse_cache *cache)
538 {
539 	unsigned int i;
540 
541 	for (i = 0; i < cache->capacity; i++)
542 		release_cached_chapter_index(&cache->chapters[i]);
543 }
544 
545 static inline bool should_skip_chapter(struct cached_chapter_index *chapter,
546 				       u64 oldest_chapter, u64 requested_chapter)
547 {
548 	if ((chapter->virtual_chapter == NO_CHAPTER) ||
549 	    (chapter->virtual_chapter < oldest_chapter))
550 		return true;
551 
552 	if (requested_chapter != NO_CHAPTER)
553 		return requested_chapter != chapter->virtual_chapter;
554 	else
555 		return READ_ONCE(chapter->skip_search);
556 }
557 
558 static int __must_check search_cached_chapter_index(struct cached_chapter_index *chapter,
559 						    const struct index_geometry *geometry,
560 						    const struct index_page_map *index_page_map,
561 						    const struct uds_record_name *name,
562 						    u16 *record_page_ptr)
563 {
564 	u32 physical_chapter =
565 		uds_map_to_physical_chapter(geometry, chapter->virtual_chapter);
566 	u32 index_page_number =
567 		uds_find_index_page_number(index_page_map, name, physical_chapter);
568 	struct delta_index_page *index_page =
569 		&chapter->index_pages[index_page_number];
570 
571 	return uds_search_chapter_index_page(index_page, geometry, name,
572 					     record_page_ptr);
573 }
574 
575 int uds_search_sparse_cache(struct index_zone *zone, const struct uds_record_name *name,
576 			    u64 *virtual_chapter_ptr, u16 *record_page_ptr)
577 {
578 	int result;
579 	struct volume *volume = zone->index->volume;
580 	struct sparse_cache *cache = volume->sparse_cache;
581 	struct cached_chapter_index *chapter;
582 	struct search_list *search_list;
583 	u8 i;
584 	/* Search the entire cache unless a specific chapter was requested. */
585 	bool search_one = (*virtual_chapter_ptr != NO_CHAPTER);
586 
587 	*record_page_ptr = NO_CHAPTER_INDEX_ENTRY;
588 	search_list = cache->search_lists[zone->id];
589 	for (i = 0; i < search_list->first_dead_entry; i++) {
590 		chapter = search_list->entries[i];
591 
592 		if (should_skip_chapter(chapter, zone->oldest_virtual_chapter,
593 					*virtual_chapter_ptr))
594 			continue;
595 
596 		result = search_cached_chapter_index(chapter, cache->geometry,
597 						     volume->index_page_map, name,
598 						     record_page_ptr);
599 		if (result != UDS_SUCCESS)
600 			return result;
601 
602 		if (*record_page_ptr != NO_CHAPTER_INDEX_ENTRY) {
603 			/*
604 			 * In theory, this might be a false match while a true match exists in
605 			 * another chapter, but that's a very rare case and not worth the extra
606 			 * search complexity.
607 			 */
608 			set_newest_entry(search_list, i);
609 			if (zone->id == ZONE_ZERO)
610 				score_search_hit(chapter);
611 
612 			*virtual_chapter_ptr = chapter->virtual_chapter;
613 			return UDS_SUCCESS;
614 		}
615 
616 		if (zone->id == ZONE_ZERO)
617 			score_search_miss(cache, chapter);
618 
619 		if (search_one)
620 			break;
621 	}
622 
623 	return UDS_SUCCESS;
624 }
625