blk-mq: don't count completed flush data request as inflight in case of quiesce
[linux-block.git] / fs / mbcache.c
CommitLineData
09c434b8 1// SPDX-License-Identifier: GPL-2.0-only
f9a61eb4
JK
2#include <linux/spinlock.h>
3#include <linux/slab.h>
4#include <linux/list.h>
5#include <linux/list_bl.h>
6#include <linux/module.h>
7#include <linux/sched.h>
c2f3140f 8#include <linux/workqueue.h>
7a2508e1 9#include <linux/mbcache.h>
f9a61eb4
JK
10
11/*
12 * Mbcache is a simple key-value store. Keys need not be unique, however
13 * key-value pairs are expected to be unique (we use this fact in
3dc96bba 14 * mb_cache_entry_delete_or_get()).
f9a61eb4
JK
15 *
16 * Ext2 and ext4 use this cache for deduplication of extended attribute blocks.
dec214d0
TE
17 * Ext4 also uses it for deduplication of xattr values stored in inodes.
18 * They use hash of data as a key and provide a value that may represent a
19 * block or inode number. That's why keys need not be unique (hash of different
20 * data may be the same). However user provided value always uniquely
21 * identifies a cache entry.
f9a61eb4
JK
22 *
23 * We provide functions for creation and removal of entries, search by key,
24 * and a special "delete entry with given key-value pair" operation. Fixed
25 * size hash table is used for fast key lookups.
26 */
27
7a2508e1 28struct mb_cache {
f9a61eb4
JK
29 /* Hash table of entries */
30 struct hlist_bl_head *c_hash;
31 /* log2 of hash table size */
32 int c_bucket_bits;
c2f3140f 33 /* Maximum entries in cache to avoid degrading hash too much */
132d4e2d 34 unsigned long c_max_entries;
f0c8b462
JK
35 /* Protects c_list, c_entry_count */
36 spinlock_t c_list_lock;
37 struct list_head c_list;
f9a61eb4
JK
38 /* Number of entries in cache */
39 unsigned long c_entry_count;
714e5bde 40 struct shrinker *c_shrink;
c2f3140f
JK
41 /* Work for shrinking when the cache has too many entries */
42 struct work_struct c_shrink_work;
f9a61eb4
JK
43};
44
7a2508e1 45static struct kmem_cache *mb_entry_cache;
f9a61eb4 46
7a2508e1 47static unsigned long mb_cache_shrink(struct mb_cache *cache,
132d4e2d 48 unsigned long nr_to_scan);
c2f3140f 49
dc8d5e56
AG
50static inline struct hlist_bl_head *mb_cache_entry_head(struct mb_cache *cache,
51 u32 key)
f0c8b462 52{
dc8d5e56 53 return &cache->c_hash[hash_32(key, cache->c_bucket_bits)];
f0c8b462
JK
54}
55
c2f3140f
JK
56/*
57 * Number of entries to reclaim synchronously when there are too many entries
58 * in cache
59 */
60#define SYNC_SHRINK_BATCH 64
61
f9a61eb4 62/*
7a2508e1 63 * mb_cache_entry_create - create entry in cache
f9a61eb4
JK
64 * @cache - cache where the entry should be created
65 * @mask - gfp mask with which the entry should be allocated
66 * @key - key of the entry
c07dfcb4
TE
67 * @value - value of the entry
68 * @reusable - is the entry reusable by others?
f9a61eb4 69 *
c07dfcb4
TE
70 * Creates entry in @cache with key @key and value @value. The function returns
71 * -EBUSY if entry with the same key and value already exists in cache.
72 * Otherwise 0 is returned.
f9a61eb4 73 */
7a2508e1 74int mb_cache_entry_create(struct mb_cache *cache, gfp_t mask, u32 key,
c07dfcb4 75 u64 value, bool reusable)
f9a61eb4 76{
7a2508e1 77 struct mb_cache_entry *entry, *dup;
f9a61eb4
JK
78 struct hlist_bl_node *dup_node;
79 struct hlist_bl_head *head;
80
c2f3140f
JK
81 /* Schedule background reclaim if there are too many entries */
82 if (cache->c_entry_count >= cache->c_max_entries)
83 schedule_work(&cache->c_shrink_work);
84 /* Do some sync reclaim if background reclaim cannot keep up */
85 if (cache->c_entry_count >= 2*cache->c_max_entries)
7a2508e1 86 mb_cache_shrink(cache, SYNC_SHRINK_BATCH);
c2f3140f 87
7a2508e1 88 entry = kmem_cache_alloc(mb_entry_cache, mask);
f9a61eb4
JK
89 if (!entry)
90 return -ENOMEM;
91
f0c8b462 92 INIT_LIST_HEAD(&entry->e_list);
5fc4cbd9
JK
93 /*
94 * We create entry with two references. One reference is kept by the
95 * hash table, the other reference is used to protect us from
96 * mb_cache_entry_delete_or_get() until the entry is fully setup. This
97 * avoids nesting of cache->c_list_lock into hash table bit locks which
98 * is problematic for RT.
99 */
100 atomic_set(&entry->e_refcnt, 2);
f9a61eb4 101 entry->e_key = key;
c07dfcb4 102 entry->e_value = value;
a44e84a9
JK
103 entry->e_flags = 0;
104 if (reusable)
105 set_bit(MBE_REUSABLE_B, &entry->e_flags);
dc8d5e56 106 head = mb_cache_entry_head(cache, key);
f9a61eb4
JK
107 hlist_bl_lock(head);
108 hlist_bl_for_each_entry(dup, dup_node, head, e_hash_list) {
c07dfcb4 109 if (dup->e_key == key && dup->e_value == value) {
f9a61eb4 110 hlist_bl_unlock(head);
7a2508e1 111 kmem_cache_free(mb_entry_cache, entry);
f9a61eb4
JK
112 return -EBUSY;
113 }
114 }
115 hlist_bl_add_head(&entry->e_hash_list, head);
5fc4cbd9 116 hlist_bl_unlock(head);
f0c8b462
JK
117 spin_lock(&cache->c_list_lock);
118 list_add_tail(&entry->e_list, &cache->c_list);
f9a61eb4 119 cache->c_entry_count++;
f0c8b462 120 spin_unlock(&cache->c_list_lock);
5fc4cbd9 121 mb_cache_entry_put(cache, entry);
f9a61eb4
JK
122
123 return 0;
124}
7a2508e1 125EXPORT_SYMBOL(mb_cache_entry_create);
f9a61eb4 126
307af6c8 127void __mb_cache_entry_free(struct mb_cache *cache, struct mb_cache_entry *entry)
f9a61eb4 128{
307af6c8
JK
129 struct hlist_bl_head *head;
130
131 head = mb_cache_entry_head(cache, entry->e_key);
132 hlist_bl_lock(head);
133 hlist_bl_del(&entry->e_hash_list);
134 hlist_bl_unlock(head);
7a2508e1 135 kmem_cache_free(mb_entry_cache, entry);
f9a61eb4 136}
7a2508e1 137EXPORT_SYMBOL(__mb_cache_entry_free);
f9a61eb4 138
3dc96bba
JK
139/*
140 * mb_cache_entry_wait_unused - wait to be the last user of the entry
141 *
142 * @entry - entry to work on
143 *
144 * Wait to be the last user of the entry.
145 */
146void mb_cache_entry_wait_unused(struct mb_cache_entry *entry)
147{
307af6c8 148 wait_var_event(&entry->e_refcnt, atomic_read(&entry->e_refcnt) <= 2);
3dc96bba
JK
149}
150EXPORT_SYMBOL(mb_cache_entry_wait_unused);
151
7a2508e1
JK
152static struct mb_cache_entry *__entry_find(struct mb_cache *cache,
153 struct mb_cache_entry *entry,
154 u32 key)
f9a61eb4 155{
7a2508e1 156 struct mb_cache_entry *old_entry = entry;
f9a61eb4
JK
157 struct hlist_bl_node *node;
158 struct hlist_bl_head *head;
159
dc8d5e56 160 head = mb_cache_entry_head(cache, key);
f9a61eb4
JK
161 hlist_bl_lock(head);
162 if (entry && !hlist_bl_unhashed(&entry->e_hash_list))
163 node = entry->e_hash_list.next;
164 else
165 node = hlist_bl_first(head);
166 while (node) {
7a2508e1 167 entry = hlist_bl_entry(node, struct mb_cache_entry,
f9a61eb4 168 e_hash_list);
a44e84a9
JK
169 if (entry->e_key == key &&
170 test_bit(MBE_REUSABLE_B, &entry->e_flags) &&
307af6c8 171 atomic_inc_not_zero(&entry->e_refcnt))
f9a61eb4 172 goto out;
f9a61eb4
JK
173 node = node->next;
174 }
175 entry = NULL;
176out:
177 hlist_bl_unlock(head);
178 if (old_entry)
7a2508e1 179 mb_cache_entry_put(cache, old_entry);
f9a61eb4
JK
180
181 return entry;
182}
183
184/*
b649668c 185 * mb_cache_entry_find_first - find the first reusable entry with the given key
f9a61eb4
JK
186 * @cache: cache where we should search
187 * @key: key to look for
188 *
b649668c
EB
189 * Search in @cache for a reusable entry with key @key. Grabs reference to the
190 * first reusable entry found and returns the entry.
f9a61eb4 191 */
7a2508e1
JK
192struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
193 u32 key)
f9a61eb4
JK
194{
195 return __entry_find(cache, NULL, key);
196}
7a2508e1 197EXPORT_SYMBOL(mb_cache_entry_find_first);
f9a61eb4
JK
198
199/*
b649668c 200 * mb_cache_entry_find_next - find next reusable entry with the same key
f9a61eb4
JK
201 * @cache: cache where we should search
202 * @entry: entry to start search from
203 *
b649668c
EB
204 * Finds next reusable entry in the hash chain which has the same key as @entry.
205 * If @entry is unhashed (which can happen when deletion of entry races with the
206 * search), finds the first reusable entry in the hash chain. The function drops
207 * reference to @entry and returns with a reference to the found entry.
f9a61eb4 208 */
7a2508e1
JK
209struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache,
210 struct mb_cache_entry *entry)
f9a61eb4
JK
211{
212 return __entry_find(cache, entry, entry->e_key);
213}
7a2508e1 214EXPORT_SYMBOL(mb_cache_entry_find_next);
f9a61eb4 215
6048c64b 216/*
c07dfcb4 217 * mb_cache_entry_get - get a cache entry by value (and key)
6048c64b 218 * @cache - cache we work with
c07dfcb4
TE
219 * @key - key
220 * @value - value
6048c64b
AG
221 */
222struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key,
c07dfcb4 223 u64 value)
6048c64b
AG
224{
225 struct hlist_bl_node *node;
226 struct hlist_bl_head *head;
227 struct mb_cache_entry *entry;
228
229 head = mb_cache_entry_head(cache, key);
230 hlist_bl_lock(head);
231 hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
307af6c8
JK
232 if (entry->e_key == key && entry->e_value == value &&
233 atomic_inc_not_zero(&entry->e_refcnt))
6048c64b 234 goto out;
6048c64b
AG
235 }
236 entry = NULL;
237out:
238 hlist_bl_unlock(head);
239 return entry;
240}
241EXPORT_SYMBOL(mb_cache_entry_get);
242
3dc96bba
JK
243/* mb_cache_entry_delete_or_get - remove a cache entry if it has no users
244 * @cache - cache we work with
245 * @key - key
246 * @value - value
247 *
248 * Remove entry from cache @cache with key @key and value @value. The removal
249 * happens only if the entry is unused. The function returns NULL in case the
250 * entry was successfully removed or there's no entry in cache. Otherwise the
251 * function grabs reference of the entry that we failed to delete because it
252 * still has users and return it.
253 */
254struct mb_cache_entry *mb_cache_entry_delete_or_get(struct mb_cache *cache,
255 u32 key, u64 value)
256{
3dc96bba
JK
257 struct mb_cache_entry *entry;
258
307af6c8
JK
259 entry = mb_cache_entry_get(cache, key, value);
260 if (!entry)
261 return NULL;
262
263 /*
264 * Drop the ref we got from mb_cache_entry_get() and the initial hash
265 * ref if we are the last user
266 */
267 if (atomic_cmpxchg(&entry->e_refcnt, 2, 0) != 2)
268 return entry;
3dc96bba 269
307af6c8
JK
270 spin_lock(&cache->c_list_lock);
271 if (!list_empty(&entry->e_list))
272 list_del_init(&entry->e_list);
273 cache->c_entry_count--;
274 spin_unlock(&cache->c_list_lock);
275 __mb_cache_entry_free(cache, entry);
3dc96bba
JK
276 return NULL;
277}
278EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
279
7a2508e1 280/* mb_cache_entry_touch - cache entry got used
f9a61eb4
JK
281 * @cache - cache the entry belongs to
282 * @entry - entry that got used
283 *
f0c8b462 284 * Marks entry as used to give hit higher chances of surviving in cache.
f9a61eb4 285 */
7a2508e1
JK
286void mb_cache_entry_touch(struct mb_cache *cache,
287 struct mb_cache_entry *entry)
f9a61eb4 288{
a44e84a9 289 set_bit(MBE_REFERENCED_B, &entry->e_flags);
f9a61eb4 290}
7a2508e1 291EXPORT_SYMBOL(mb_cache_entry_touch);
f9a61eb4 292
7a2508e1
JK
293static unsigned long mb_cache_count(struct shrinker *shrink,
294 struct shrink_control *sc)
f9a61eb4 295{
714e5bde 296 struct mb_cache *cache = shrink->private_data;
f9a61eb4
JK
297
298 return cache->c_entry_count;
299}
300
301/* Shrink number of entries in cache */
7a2508e1 302static unsigned long mb_cache_shrink(struct mb_cache *cache,
132d4e2d 303 unsigned long nr_to_scan)
f9a61eb4 304{
7a2508e1 305 struct mb_cache_entry *entry;
132d4e2d 306 unsigned long shrunk = 0;
f9a61eb4 307
f0c8b462
JK
308 spin_lock(&cache->c_list_lock);
309 while (nr_to_scan-- && !list_empty(&cache->c_list)) {
310 entry = list_first_entry(&cache->c_list,
7a2508e1 311 struct mb_cache_entry, e_list);
307af6c8 312 /* Drop initial hash reference if there is no user */
a44e84a9 313 if (test_bit(MBE_REFERENCED_B, &entry->e_flags) ||
307af6c8 314 atomic_cmpxchg(&entry->e_refcnt, 1, 0) != 1) {
a44e84a9 315 clear_bit(MBE_REFERENCED_B, &entry->e_flags);
918b7306 316 list_move_tail(&entry->e_list, &cache->c_list);
f0c8b462
JK
317 continue;
318 }
319 list_del_init(&entry->e_list);
f9a61eb4 320 cache->c_entry_count--;
f0c8b462 321 spin_unlock(&cache->c_list_lock);
307af6c8
JK
322 __mb_cache_entry_free(cache, entry);
323 shrunk++;
f9a61eb4 324 cond_resched();
f0c8b462 325 spin_lock(&cache->c_list_lock);
f9a61eb4 326 }
f0c8b462 327 spin_unlock(&cache->c_list_lock);
f9a61eb4
JK
328
329 return shrunk;
330}
331
7a2508e1
JK
332static unsigned long mb_cache_scan(struct shrinker *shrink,
333 struct shrink_control *sc)
c2f3140f 334{
714e5bde 335 struct mb_cache *cache = shrink->private_data;
132d4e2d 336 return mb_cache_shrink(cache, sc->nr_to_scan);
c2f3140f
JK
337}
338
339/* We shrink 1/X of the cache when we have too many entries in it */
340#define SHRINK_DIVISOR 16
341
7a2508e1 342static void mb_cache_shrink_worker(struct work_struct *work)
c2f3140f 343{
7a2508e1
JK
344 struct mb_cache *cache = container_of(work, struct mb_cache,
345 c_shrink_work);
346 mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR);
c2f3140f
JK
347}
348
f9a61eb4 349/*
7a2508e1 350 * mb_cache_create - create cache
f9a61eb4
JK
351 * @bucket_bits: log2 of the hash table size
352 *
353 * Create cache for keys with 2^bucket_bits hash entries.
354 */
7a2508e1 355struct mb_cache *mb_cache_create(int bucket_bits)
f9a61eb4 356{
7a2508e1 357 struct mb_cache *cache;
132d4e2d
EB
358 unsigned long bucket_count = 1UL << bucket_bits;
359 unsigned long i;
f9a61eb4 360
7a2508e1 361 cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL);
f9a61eb4
JK
362 if (!cache)
363 goto err_out;
364 cache->c_bucket_bits = bucket_bits;
c2f3140f 365 cache->c_max_entries = bucket_count << 4;
f0c8b462
JK
366 INIT_LIST_HEAD(&cache->c_list);
367 spin_lock_init(&cache->c_list_lock);
6da2ec56
KC
368 cache->c_hash = kmalloc_array(bucket_count,
369 sizeof(struct hlist_bl_head),
370 GFP_KERNEL);
f9a61eb4
JK
371 if (!cache->c_hash) {
372 kfree(cache);
373 goto err_out;
374 }
375 for (i = 0; i < bucket_count; i++)
376 INIT_HLIST_BL_HEAD(&cache->c_hash[i]);
377
714e5bde
QZ
378 cache->c_shrink = shrinker_alloc(0, "mbcache-shrinker");
379 if (!cache->c_shrink) {
8913f343
CY
380 kfree(cache->c_hash);
381 kfree(cache);
382 goto err_out;
383 }
f9a61eb4 384
714e5bde
QZ
385 cache->c_shrink->count_objects = mb_cache_count;
386 cache->c_shrink->scan_objects = mb_cache_scan;
387 cache->c_shrink->private_data = cache;
388
389 shrinker_register(cache->c_shrink);
390
7a2508e1 391 INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
c2f3140f 392
f9a61eb4
JK
393 return cache;
394
395err_out:
f9a61eb4
JK
396 return NULL;
397}
7a2508e1 398EXPORT_SYMBOL(mb_cache_create);
f9a61eb4
JK
399
400/*
7a2508e1 401 * mb_cache_destroy - destroy cache
f9a61eb4
JK
402 * @cache: the cache to destroy
403 *
404 * Free all entries in cache and cache itself. Caller must make sure nobody
405 * (except shrinker) can reach @cache when calling this.
406 */
7a2508e1 407void mb_cache_destroy(struct mb_cache *cache)
f9a61eb4 408{
7a2508e1 409 struct mb_cache_entry *entry, *next;
f9a61eb4 410
714e5bde 411 shrinker_free(cache->c_shrink);
f9a61eb4
JK
412
413 /*
414 * We don't bother with any locking. Cache must not be used at this
415 * point.
416 */
f0c8b462 417 list_for_each_entry_safe(entry, next, &cache->c_list, e_list) {
f0c8b462 418 list_del(&entry->e_list);
f9a61eb4 419 WARN_ON(atomic_read(&entry->e_refcnt) != 1);
7a2508e1 420 mb_cache_entry_put(cache, entry);
f9a61eb4
JK
421 }
422 kfree(cache->c_hash);
423 kfree(cache);
f9a61eb4 424}
7a2508e1 425EXPORT_SYMBOL(mb_cache_destroy);
f9a61eb4 426
7a2508e1 427static int __init mbcache_init(void)
f9a61eb4 428{
7a2508e1
JK
429 mb_entry_cache = kmem_cache_create("mbcache",
430 sizeof(struct mb_cache_entry), 0,
f9a61eb4 431 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
21d0f4fa
EB
432 if (!mb_entry_cache)
433 return -ENOMEM;
f9a61eb4
JK
434 return 0;
435}
436
7a2508e1 437static void __exit mbcache_exit(void)
f9a61eb4 438{
7a2508e1 439 kmem_cache_destroy(mb_entry_cache);
f9a61eb4
JK
440}
441
7a2508e1
JK
442module_init(mbcache_init)
443module_exit(mbcache_exit)
f9a61eb4
JK
444
445MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
446MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
447MODULE_LICENSE("GPL");