Loading...
1// SPDX-License-Identifier: GPL-2.0-only
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>
8#include <linux/workqueue.h>
9#include <linux/mbcache.h>
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
14 * mb_cache_entry_delete_or_get()).
15 *
16 * Ext2 and ext4 use this cache for deduplication of extended attribute blocks.
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.
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
28struct mb_cache {
29 /* Hash table of entries */
30 struct hlist_bl_head *c_hash;
31 /* log2 of hash table size */
32 int c_bucket_bits;
33 /* Maximum entries in cache to avoid degrading hash too much */
34 unsigned long c_max_entries;
35 /* Protects c_list, c_entry_count */
36 spinlock_t c_list_lock;
37 struct list_head c_list;
38 /* Number of entries in cache */
39 unsigned long c_entry_count;
40 struct shrinker *c_shrink;
41 /* Work for shrinking when the cache has too many entries */
42 struct work_struct c_shrink_work;
43};
44
45static struct kmem_cache *mb_entry_cache;
46
47static unsigned long mb_cache_shrink(struct mb_cache *cache,
48 unsigned long nr_to_scan);
49
50static inline struct hlist_bl_head *mb_cache_entry_head(struct mb_cache *cache,
51 u32 key)
52{
53 return &cache->c_hash[hash_32(key, cache->c_bucket_bits)];
54}
55
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
62/*
63 * mb_cache_entry_create - create entry in cache
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
67 * @value - value of the entry
68 * @reusable - is the entry reusable by others?
69 *
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.
73 */
74int mb_cache_entry_create(struct mb_cache *cache, gfp_t mask, u32 key,
75 u64 value, bool reusable)
76{
77 struct mb_cache_entry *entry, *dup;
78 struct hlist_bl_node *dup_node;
79 struct hlist_bl_head *head;
80
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)
86 mb_cache_shrink(cache, SYNC_SHRINK_BATCH);
87
88 entry = kmem_cache_alloc(mb_entry_cache, mask);
89 if (!entry)
90 return -ENOMEM;
91
92 INIT_LIST_HEAD(&entry->e_list);
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);
101 entry->e_key = key;
102 entry->e_value = value;
103 entry->e_flags = 0;
104 if (reusable)
105 set_bit(MBE_REUSABLE_B, &entry->e_flags);
106 head = mb_cache_entry_head(cache, key);
107 hlist_bl_lock(head);
108 hlist_bl_for_each_entry(dup, dup_node, head, e_hash_list) {
109 if (dup->e_key == key && dup->e_value == value) {
110 hlist_bl_unlock(head);
111 kmem_cache_free(mb_entry_cache, entry);
112 return -EBUSY;
113 }
114 }
115 hlist_bl_add_head(&entry->e_hash_list, head);
116 hlist_bl_unlock(head);
117 spin_lock(&cache->c_list_lock);
118 list_add_tail(&entry->e_list, &cache->c_list);
119 cache->c_entry_count++;
120 spin_unlock(&cache->c_list_lock);
121 mb_cache_entry_put(cache, entry);
122
123 return 0;
124}
125EXPORT_SYMBOL(mb_cache_entry_create);
126
127void __mb_cache_entry_free(struct mb_cache *cache, struct mb_cache_entry *entry)
128{
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);
135 kmem_cache_free(mb_entry_cache, entry);
136}
137EXPORT_SYMBOL(__mb_cache_entry_free);
138
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{
148 wait_var_event(&entry->e_refcnt, atomic_read(&entry->e_refcnt) <= 2);
149}
150EXPORT_SYMBOL(mb_cache_entry_wait_unused);
151
152static struct mb_cache_entry *__entry_find(struct mb_cache *cache,
153 struct mb_cache_entry *entry,
154 u32 key)
155{
156 struct mb_cache_entry *old_entry = entry;
157 struct hlist_bl_node *node;
158 struct hlist_bl_head *head;
159
160 head = mb_cache_entry_head(cache, key);
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) {
167 entry = hlist_bl_entry(node, struct mb_cache_entry,
168 e_hash_list);
169 if (entry->e_key == key &&
170 test_bit(MBE_REUSABLE_B, &entry->e_flags) &&
171 atomic_inc_not_zero(&entry->e_refcnt))
172 goto out;
173 node = node->next;
174 }
175 entry = NULL;
176out:
177 hlist_bl_unlock(head);
178 if (old_entry)
179 mb_cache_entry_put(cache, old_entry);
180
181 return entry;
182}
183
184/*
185 * mb_cache_entry_find_first - find the first reusable entry with the given key
186 * @cache: cache where we should search
187 * @key: key to look for
188 *
189 * Search in @cache for a reusable entry with key @key. Grabs reference to the
190 * first reusable entry found and returns the entry.
191 */
192struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
193 u32 key)
194{
195 return __entry_find(cache, NULL, key);
196}
197EXPORT_SYMBOL(mb_cache_entry_find_first);
198
199/*
200 * mb_cache_entry_find_next - find next reusable entry with the same key
201 * @cache: cache where we should search
202 * @entry: entry to start search from
203 *
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.
208 */
209struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache,
210 struct mb_cache_entry *entry)
211{
212 return __entry_find(cache, entry, entry->e_key);
213}
214EXPORT_SYMBOL(mb_cache_entry_find_next);
215
216/*
217 * mb_cache_entry_get - get a cache entry by value (and key)
218 * @cache - cache we work with
219 * @key - key
220 * @value - value
221 */
222struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key,
223 u64 value)
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) {
232 if (entry->e_key == key && entry->e_value == value &&
233 atomic_inc_not_zero(&entry->e_refcnt))
234 goto out;
235 }
236 entry = NULL;
237out:
238 hlist_bl_unlock(head);
239 return entry;
240}
241EXPORT_SYMBOL(mb_cache_entry_get);
242
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{
257 struct mb_cache_entry *entry;
258
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;
269
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);
276 return NULL;
277}
278EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
279
280/* mb_cache_entry_touch - cache entry got used
281 * @cache - cache the entry belongs to
282 * @entry - entry that got used
283 *
284 * Marks entry as used to give hit higher chances of surviving in cache.
285 */
286void mb_cache_entry_touch(struct mb_cache *cache,
287 struct mb_cache_entry *entry)
288{
289 set_bit(MBE_REFERENCED_B, &entry->e_flags);
290}
291EXPORT_SYMBOL(mb_cache_entry_touch);
292
293static unsigned long mb_cache_count(struct shrinker *shrink,
294 struct shrink_control *sc)
295{
296 struct mb_cache *cache = shrink->private_data;
297
298 return cache->c_entry_count;
299}
300
301/* Shrink number of entries in cache */
302static unsigned long mb_cache_shrink(struct mb_cache *cache,
303 unsigned long nr_to_scan)
304{
305 struct mb_cache_entry *entry;
306 unsigned long shrunk = 0;
307
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,
311 struct mb_cache_entry, e_list);
312 /* Drop initial hash reference if there is no user */
313 if (test_bit(MBE_REFERENCED_B, &entry->e_flags) ||
314 atomic_cmpxchg(&entry->e_refcnt, 1, 0) != 1) {
315 clear_bit(MBE_REFERENCED_B, &entry->e_flags);
316 list_move_tail(&entry->e_list, &cache->c_list);
317 continue;
318 }
319 list_del_init(&entry->e_list);
320 cache->c_entry_count--;
321 spin_unlock(&cache->c_list_lock);
322 __mb_cache_entry_free(cache, entry);
323 shrunk++;
324 cond_resched();
325 spin_lock(&cache->c_list_lock);
326 }
327 spin_unlock(&cache->c_list_lock);
328
329 return shrunk;
330}
331
332static unsigned long mb_cache_scan(struct shrinker *shrink,
333 struct shrink_control *sc)
334{
335 struct mb_cache *cache = shrink->private_data;
336 return mb_cache_shrink(cache, sc->nr_to_scan);
337}
338
339/* We shrink 1/X of the cache when we have too many entries in it */
340#define SHRINK_DIVISOR 16
341
342static void mb_cache_shrink_worker(struct work_struct *work)
343{
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);
347}
348
349/*
350 * mb_cache_create - create cache
351 * @bucket_bits: log2 of the hash table size
352 *
353 * Create cache for keys with 2^bucket_bits hash entries.
354 */
355struct mb_cache *mb_cache_create(int bucket_bits)
356{
357 struct mb_cache *cache;
358 unsigned long bucket_count = 1UL << bucket_bits;
359 unsigned long i;
360
361 cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL);
362 if (!cache)
363 goto err_out;
364 cache->c_bucket_bits = bucket_bits;
365 cache->c_max_entries = bucket_count << 4;
366 INIT_LIST_HEAD(&cache->c_list);
367 spin_lock_init(&cache->c_list_lock);
368 cache->c_hash = kmalloc_array(bucket_count,
369 sizeof(struct hlist_bl_head),
370 GFP_KERNEL);
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
378 cache->c_shrink = shrinker_alloc(0, "mbcache-shrinker");
379 if (!cache->c_shrink) {
380 kfree(cache->c_hash);
381 kfree(cache);
382 goto err_out;
383 }
384
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
391 INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
392
393 return cache;
394
395err_out:
396 return NULL;
397}
398EXPORT_SYMBOL(mb_cache_create);
399
400/*
401 * mb_cache_destroy - destroy cache
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 */
407void mb_cache_destroy(struct mb_cache *cache)
408{
409 struct mb_cache_entry *entry, *next;
410
411 shrinker_free(cache->c_shrink);
412
413 /*
414 * We don't bother with any locking. Cache must not be used at this
415 * point.
416 */
417 list_for_each_entry_safe(entry, next, &cache->c_list, e_list) {
418 list_del(&entry->e_list);
419 WARN_ON(atomic_read(&entry->e_refcnt) != 1);
420 mb_cache_entry_put(cache, entry);
421 }
422 kfree(cache->c_hash);
423 kfree(cache);
424}
425EXPORT_SYMBOL(mb_cache_destroy);
426
427static int __init mbcache_init(void)
428{
429 mb_entry_cache = KMEM_CACHE(mb_cache_entry, SLAB_RECLAIM_ACCOUNT);
430 if (!mb_entry_cache)
431 return -ENOMEM;
432 return 0;
433}
434
435static void __exit mbcache_exit(void)
436{
437 kmem_cache_destroy(mb_entry_cache);
438}
439
440module_init(mbcache_init)
441module_exit(mbcache_exit)
442
443MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
444MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
445MODULE_LICENSE("GPL");
1/*
2 * linux/fs/mbcache.c
3 * (C) 2001-2002 Andreas Gruenbacher, <a.gruenbacher@computer.org>
4 */
5
6/*
7 * Filesystem Meta Information Block Cache (mbcache)
8 *
9 * The mbcache caches blocks of block devices that need to be located
10 * by their device/block number, as well as by other criteria (such
11 * as the block's contents).
12 *
13 * There can only be one cache entry in a cache per device and block number.
14 * Additional indexes need not be unique in this sense. The number of
15 * additional indexes (=other criteria) can be hardwired at compile time
16 * or specified at cache create time.
17 *
18 * Each cache entry is of fixed size. An entry may be `valid' or `invalid'
19 * in the cache. A valid entry is in the main hash tables of the cache,
20 * and may also be in the lru list. An invalid entry is not in any hashes
21 * or lists.
22 *
23 * A valid cache entry is only in the lru list if no handles refer to it.
24 * Invalid cache entries will be freed when the last handle to the cache
25 * entry is released. Entries that cannot be freed immediately are put
26 * back on the lru list.
27 */
28
29/*
30 * Lock descriptions and usage:
31 *
32 * Each hash chain of both the block and index hash tables now contains
33 * a built-in lock used to serialize accesses to the hash chain.
34 *
35 * Accesses to global data structures mb_cache_list and mb_cache_lru_list
36 * are serialized via the global spinlock mb_cache_spinlock.
37 *
38 * Each mb_cache_entry contains a spinlock, e_entry_lock, to serialize
39 * accesses to its local data, such as e_used and e_queued.
40 *
41 * Lock ordering:
42 *
43 * Each block hash chain's lock has the highest lock order, followed by an
44 * index hash chain's lock, mb_cache_bg_lock (used to implement mb_cache_entry's
45 * lock), and mb_cach_spinlock, with the lowest order. While holding
46 * either a block or index hash chain lock, a thread can acquire an
47 * mc_cache_bg_lock, which in turn can also acquire mb_cache_spinlock.
48 *
49 * Synchronization:
50 *
51 * Since both mb_cache_entry_get and mb_cache_entry_find scan the block and
52 * index hash chian, it needs to lock the corresponding hash chain. For each
53 * mb_cache_entry within the chain, it needs to lock the mb_cache_entry to
54 * prevent either any simultaneous release or free on the entry and also
55 * to serialize accesses to either the e_used or e_queued member of the entry.
56 *
57 * To avoid having a dangling reference to an already freed
58 * mb_cache_entry, an mb_cache_entry is only freed when it is not on a
59 * block hash chain and also no longer being referenced, both e_used,
60 * and e_queued are 0's. When an mb_cache_entry is explicitly freed it is
61 * first removed from a block hash chain.
62 */
63
64#include <linux/kernel.h>
65#include <linux/module.h>
66
67#include <linux/hash.h>
68#include <linux/fs.h>
69#include <linux/mm.h>
70#include <linux/slab.h>
71#include <linux/sched.h>
72#include <linux/list_bl.h>
73#include <linux/mbcache.h>
74#include <linux/init.h>
75#include <linux/blockgroup_lock.h>
76
77#ifdef MB_CACHE_DEBUG
78# define mb_debug(f...) do { \
79 printk(KERN_DEBUG f); \
80 printk("\n"); \
81 } while (0)
82#define mb_assert(c) do { if (!(c)) \
83 printk(KERN_ERR "assertion " #c " failed\n"); \
84 } while(0)
85#else
86# define mb_debug(f...) do { } while(0)
87# define mb_assert(c) do { } while(0)
88#endif
89#define mb_error(f...) do { \
90 printk(KERN_ERR f); \
91 printk("\n"); \
92 } while(0)
93
94#define MB_CACHE_WRITER ((unsigned short)~0U >> 1)
95
96#define MB_CACHE_ENTRY_LOCK_BITS __builtin_log2(NR_BG_LOCKS)
97#define MB_CACHE_ENTRY_LOCK_INDEX(ce) \
98 (hash_long((unsigned long)ce, MB_CACHE_ENTRY_LOCK_BITS))
99
100static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
101static struct blockgroup_lock *mb_cache_bg_lock;
102static struct kmem_cache *mb_cache_kmem_cache;
103
104MODULE_AUTHOR("Andreas Gruenbacher <a.gruenbacher@computer.org>");
105MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
106MODULE_LICENSE("GPL");
107
108EXPORT_SYMBOL(mb_cache_create);
109EXPORT_SYMBOL(mb_cache_shrink);
110EXPORT_SYMBOL(mb_cache_destroy);
111EXPORT_SYMBOL(mb_cache_entry_alloc);
112EXPORT_SYMBOL(mb_cache_entry_insert);
113EXPORT_SYMBOL(mb_cache_entry_release);
114EXPORT_SYMBOL(mb_cache_entry_free);
115EXPORT_SYMBOL(mb_cache_entry_get);
116#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
117EXPORT_SYMBOL(mb_cache_entry_find_first);
118EXPORT_SYMBOL(mb_cache_entry_find_next);
119#endif
120
121/*
122 * Global data: list of all mbcache's, lru list, and a spinlock for
123 * accessing cache data structures on SMP machines. The lru list is
124 * global across all mbcaches.
125 */
126
127static LIST_HEAD(mb_cache_list);
128static LIST_HEAD(mb_cache_lru_list);
129static DEFINE_SPINLOCK(mb_cache_spinlock);
130
131static inline void
132__spin_lock_mb_cache_entry(struct mb_cache_entry *ce)
133{
134 spin_lock(bgl_lock_ptr(mb_cache_bg_lock,
135 MB_CACHE_ENTRY_LOCK_INDEX(ce)));
136}
137
138static inline void
139__spin_unlock_mb_cache_entry(struct mb_cache_entry *ce)
140{
141 spin_unlock(bgl_lock_ptr(mb_cache_bg_lock,
142 MB_CACHE_ENTRY_LOCK_INDEX(ce)));
143}
144
145static inline int
146__mb_cache_entry_is_block_hashed(struct mb_cache_entry *ce)
147{
148 return !hlist_bl_unhashed(&ce->e_block_list);
149}
150
151
152static inline void
153__mb_cache_entry_unhash_block(struct mb_cache_entry *ce)
154{
155 if (__mb_cache_entry_is_block_hashed(ce))
156 hlist_bl_del_init(&ce->e_block_list);
157}
158
159static inline int
160__mb_cache_entry_is_index_hashed(struct mb_cache_entry *ce)
161{
162 return !hlist_bl_unhashed(&ce->e_index.o_list);
163}
164
165static inline void
166__mb_cache_entry_unhash_index(struct mb_cache_entry *ce)
167{
168 if (__mb_cache_entry_is_index_hashed(ce))
169 hlist_bl_del_init(&ce->e_index.o_list);
170}
171
172/*
173 * __mb_cache_entry_unhash_unlock()
174 *
175 * This function is called to unhash both the block and index hash
176 * chain.
177 * It assumes both the block and index hash chain is locked upon entry.
178 * It also unlock both hash chains both exit
179 */
180static inline void
181__mb_cache_entry_unhash_unlock(struct mb_cache_entry *ce)
182{
183 __mb_cache_entry_unhash_index(ce);
184 hlist_bl_unlock(ce->e_index_hash_p);
185 __mb_cache_entry_unhash_block(ce);
186 hlist_bl_unlock(ce->e_block_hash_p);
187}
188
189static void
190__mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
191{
192 struct mb_cache *cache = ce->e_cache;
193
194 mb_assert(!(ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt)));
195 kmem_cache_free(cache->c_entry_cache, ce);
196 atomic_dec(&cache->c_entry_count);
197}
198
199static void
200__mb_cache_entry_release(struct mb_cache_entry *ce)
201{
202 /* First lock the entry to serialize access to its local data. */
203 __spin_lock_mb_cache_entry(ce);
204 /* Wake up all processes queuing for this cache entry. */
205 if (ce->e_queued)
206 wake_up_all(&mb_cache_queue);
207 if (ce->e_used >= MB_CACHE_WRITER)
208 ce->e_used -= MB_CACHE_WRITER;
209 /*
210 * Make sure that all cache entries on lru_list have
211 * both e_used and e_qued of 0s.
212 */
213 ce->e_used--;
214 if (!(ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt))) {
215 if (!__mb_cache_entry_is_block_hashed(ce)) {
216 __spin_unlock_mb_cache_entry(ce);
217 goto forget;
218 }
219 /*
220 * Need access to lru list, first drop entry lock,
221 * then reacquire the lock in the proper order.
222 */
223 spin_lock(&mb_cache_spinlock);
224 if (list_empty(&ce->e_lru_list))
225 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
226 spin_unlock(&mb_cache_spinlock);
227 }
228 __spin_unlock_mb_cache_entry(ce);
229 return;
230forget:
231 mb_assert(list_empty(&ce->e_lru_list));
232 __mb_cache_entry_forget(ce, GFP_KERNEL);
233}
234
235/*
236 * mb_cache_shrink_scan() memory pressure callback
237 *
238 * This function is called by the kernel memory management when memory
239 * gets low.
240 *
241 * @shrink: (ignored)
242 * @sc: shrink_control passed from reclaim
243 *
244 * Returns the number of objects freed.
245 */
246static unsigned long
247mb_cache_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
248{
249 LIST_HEAD(free_list);
250 struct mb_cache_entry *entry, *tmp;
251 int nr_to_scan = sc->nr_to_scan;
252 gfp_t gfp_mask = sc->gfp_mask;
253 unsigned long freed = 0;
254
255 mb_debug("trying to free %d entries", nr_to_scan);
256 spin_lock(&mb_cache_spinlock);
257 while ((nr_to_scan-- > 0) && !list_empty(&mb_cache_lru_list)) {
258 struct mb_cache_entry *ce =
259 list_entry(mb_cache_lru_list.next,
260 struct mb_cache_entry, e_lru_list);
261 list_del_init(&ce->e_lru_list);
262 if (ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt))
263 continue;
264 spin_unlock(&mb_cache_spinlock);
265 /* Prevent any find or get operation on the entry */
266 hlist_bl_lock(ce->e_block_hash_p);
267 hlist_bl_lock(ce->e_index_hash_p);
268 /* Ignore if it is touched by a find/get */
269 if (ce->e_used || ce->e_queued || atomic_read(&ce->e_refcnt) ||
270 !list_empty(&ce->e_lru_list)) {
271 hlist_bl_unlock(ce->e_index_hash_p);
272 hlist_bl_unlock(ce->e_block_hash_p);
273 spin_lock(&mb_cache_spinlock);
274 continue;
275 }
276 __mb_cache_entry_unhash_unlock(ce);
277 list_add_tail(&ce->e_lru_list, &free_list);
278 spin_lock(&mb_cache_spinlock);
279 }
280 spin_unlock(&mb_cache_spinlock);
281
282 list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
283 __mb_cache_entry_forget(entry, gfp_mask);
284 freed++;
285 }
286 return freed;
287}
288
289static unsigned long
290mb_cache_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
291{
292 struct mb_cache *cache;
293 unsigned long count = 0;
294
295 spin_lock(&mb_cache_spinlock);
296 list_for_each_entry(cache, &mb_cache_list, c_cache_list) {
297 mb_debug("cache %s (%d)", cache->c_name,
298 atomic_read(&cache->c_entry_count));
299 count += atomic_read(&cache->c_entry_count);
300 }
301 spin_unlock(&mb_cache_spinlock);
302
303 return vfs_pressure_ratio(count);
304}
305
306static struct shrinker mb_cache_shrinker = {
307 .count_objects = mb_cache_shrink_count,
308 .scan_objects = mb_cache_shrink_scan,
309 .seeks = DEFAULT_SEEKS,
310};
311
312/*
313 * mb_cache_create() create a new cache
314 *
315 * All entries in one cache are equal size. Cache entries may be from
316 * multiple devices. If this is the first mbcache created, registers
317 * the cache with kernel memory management. Returns NULL if no more
318 * memory was available.
319 *
320 * @name: name of the cache (informal)
321 * @bucket_bits: log2(number of hash buckets)
322 */
323struct mb_cache *
324mb_cache_create(const char *name, int bucket_bits)
325{
326 int n, bucket_count = 1 << bucket_bits;
327 struct mb_cache *cache = NULL;
328
329 if (!mb_cache_bg_lock) {
330 mb_cache_bg_lock = kmalloc(sizeof(struct blockgroup_lock),
331 GFP_KERNEL);
332 if (!mb_cache_bg_lock)
333 return NULL;
334 bgl_lock_init(mb_cache_bg_lock);
335 }
336
337 cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
338 if (!cache)
339 return NULL;
340 cache->c_name = name;
341 atomic_set(&cache->c_entry_count, 0);
342 cache->c_bucket_bits = bucket_bits;
343 cache->c_block_hash = kmalloc(bucket_count *
344 sizeof(struct hlist_bl_head), GFP_KERNEL);
345 if (!cache->c_block_hash)
346 goto fail;
347 for (n=0; n<bucket_count; n++)
348 INIT_HLIST_BL_HEAD(&cache->c_block_hash[n]);
349 cache->c_index_hash = kmalloc(bucket_count *
350 sizeof(struct hlist_bl_head), GFP_KERNEL);
351 if (!cache->c_index_hash)
352 goto fail;
353 for (n=0; n<bucket_count; n++)
354 INIT_HLIST_BL_HEAD(&cache->c_index_hash[n]);
355 if (!mb_cache_kmem_cache) {
356 mb_cache_kmem_cache = kmem_cache_create(name,
357 sizeof(struct mb_cache_entry), 0,
358 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
359 if (!mb_cache_kmem_cache)
360 goto fail2;
361 }
362 cache->c_entry_cache = mb_cache_kmem_cache;
363
364 /*
365 * Set an upper limit on the number of cache entries so that the hash
366 * chains won't grow too long.
367 */
368 cache->c_max_entries = bucket_count << 4;
369
370 spin_lock(&mb_cache_spinlock);
371 list_add(&cache->c_cache_list, &mb_cache_list);
372 spin_unlock(&mb_cache_spinlock);
373 return cache;
374
375fail2:
376 kfree(cache->c_index_hash);
377
378fail:
379 kfree(cache->c_block_hash);
380 kfree(cache);
381 return NULL;
382}
383
384
385/*
386 * mb_cache_shrink()
387 *
388 * Removes all cache entries of a device from the cache. All cache entries
389 * currently in use cannot be freed, and thus remain in the cache. All others
390 * are freed.
391 *
392 * @bdev: which device's cache entries to shrink
393 */
394void
395mb_cache_shrink(struct block_device *bdev)
396{
397 LIST_HEAD(free_list);
398 struct list_head *l;
399 struct mb_cache_entry *ce, *tmp;
400
401 l = &mb_cache_lru_list;
402 spin_lock(&mb_cache_spinlock);
403 while (!list_is_last(l, &mb_cache_lru_list)) {
404 l = l->next;
405 ce = list_entry(l, struct mb_cache_entry, e_lru_list);
406 if (ce->e_bdev == bdev) {
407 list_del_init(&ce->e_lru_list);
408 if (ce->e_used || ce->e_queued ||
409 atomic_read(&ce->e_refcnt))
410 continue;
411 spin_unlock(&mb_cache_spinlock);
412 /*
413 * Prevent any find or get operation on the entry.
414 */
415 hlist_bl_lock(ce->e_block_hash_p);
416 hlist_bl_lock(ce->e_index_hash_p);
417 /* Ignore if it is touched by a find/get */
418 if (ce->e_used || ce->e_queued ||
419 atomic_read(&ce->e_refcnt) ||
420 !list_empty(&ce->e_lru_list)) {
421 hlist_bl_unlock(ce->e_index_hash_p);
422 hlist_bl_unlock(ce->e_block_hash_p);
423 l = &mb_cache_lru_list;
424 spin_lock(&mb_cache_spinlock);
425 continue;
426 }
427 __mb_cache_entry_unhash_unlock(ce);
428 mb_assert(!(ce->e_used || ce->e_queued ||
429 atomic_read(&ce->e_refcnt)));
430 list_add_tail(&ce->e_lru_list, &free_list);
431 l = &mb_cache_lru_list;
432 spin_lock(&mb_cache_spinlock);
433 }
434 }
435 spin_unlock(&mb_cache_spinlock);
436
437 list_for_each_entry_safe(ce, tmp, &free_list, e_lru_list) {
438 __mb_cache_entry_forget(ce, GFP_KERNEL);
439 }
440}
441
442
443/*
444 * mb_cache_destroy()
445 *
446 * Shrinks the cache to its minimum possible size (hopefully 0 entries),
447 * and then destroys it. If this was the last mbcache, un-registers the
448 * mbcache from kernel memory management.
449 */
450void
451mb_cache_destroy(struct mb_cache *cache)
452{
453 LIST_HEAD(free_list);
454 struct mb_cache_entry *ce, *tmp;
455
456 spin_lock(&mb_cache_spinlock);
457 list_for_each_entry_safe(ce, tmp, &mb_cache_lru_list, e_lru_list) {
458 if (ce->e_cache == cache)
459 list_move_tail(&ce->e_lru_list, &free_list);
460 }
461 list_del(&cache->c_cache_list);
462 spin_unlock(&mb_cache_spinlock);
463
464 list_for_each_entry_safe(ce, tmp, &free_list, e_lru_list) {
465 list_del_init(&ce->e_lru_list);
466 /*
467 * Prevent any find or get operation on the entry.
468 */
469 hlist_bl_lock(ce->e_block_hash_p);
470 hlist_bl_lock(ce->e_index_hash_p);
471 mb_assert(!(ce->e_used || ce->e_queued ||
472 atomic_read(&ce->e_refcnt)));
473 __mb_cache_entry_unhash_unlock(ce);
474 __mb_cache_entry_forget(ce, GFP_KERNEL);
475 }
476
477 if (atomic_read(&cache->c_entry_count) > 0) {
478 mb_error("cache %s: %d orphaned entries",
479 cache->c_name,
480 atomic_read(&cache->c_entry_count));
481 }
482
483 if (list_empty(&mb_cache_list)) {
484 kmem_cache_destroy(mb_cache_kmem_cache);
485 mb_cache_kmem_cache = NULL;
486 }
487 kfree(cache->c_index_hash);
488 kfree(cache->c_block_hash);
489 kfree(cache);
490}
491
492/*
493 * mb_cache_entry_alloc()
494 *
495 * Allocates a new cache entry. The new entry will not be valid initially,
496 * and thus cannot be looked up yet. It should be filled with data, and
497 * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
498 * if no more memory was available.
499 */
500struct mb_cache_entry *
501mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
502{
503 struct mb_cache_entry *ce;
504
505 if (atomic_read(&cache->c_entry_count) >= cache->c_max_entries) {
506 struct list_head *l;
507
508 l = &mb_cache_lru_list;
509 spin_lock(&mb_cache_spinlock);
510 while (!list_is_last(l, &mb_cache_lru_list)) {
511 l = l->next;
512 ce = list_entry(l, struct mb_cache_entry, e_lru_list);
513 if (ce->e_cache == cache) {
514 list_del_init(&ce->e_lru_list);
515 if (ce->e_used || ce->e_queued ||
516 atomic_read(&ce->e_refcnt))
517 continue;
518 spin_unlock(&mb_cache_spinlock);
519 /*
520 * Prevent any find or get operation on the
521 * entry.
522 */
523 hlist_bl_lock(ce->e_block_hash_p);
524 hlist_bl_lock(ce->e_index_hash_p);
525 /* Ignore if it is touched by a find/get */
526 if (ce->e_used || ce->e_queued ||
527 atomic_read(&ce->e_refcnt) ||
528 !list_empty(&ce->e_lru_list)) {
529 hlist_bl_unlock(ce->e_index_hash_p);
530 hlist_bl_unlock(ce->e_block_hash_p);
531 l = &mb_cache_lru_list;
532 spin_lock(&mb_cache_spinlock);
533 continue;
534 }
535 mb_assert(list_empty(&ce->e_lru_list));
536 mb_assert(!(ce->e_used || ce->e_queued ||
537 atomic_read(&ce->e_refcnt)));
538 __mb_cache_entry_unhash_unlock(ce);
539 goto found;
540 }
541 }
542 spin_unlock(&mb_cache_spinlock);
543 }
544
545 ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
546 if (!ce)
547 return NULL;
548 atomic_inc(&cache->c_entry_count);
549 INIT_LIST_HEAD(&ce->e_lru_list);
550 INIT_HLIST_BL_NODE(&ce->e_block_list);
551 INIT_HLIST_BL_NODE(&ce->e_index.o_list);
552 ce->e_cache = cache;
553 ce->e_queued = 0;
554 atomic_set(&ce->e_refcnt, 0);
555found:
556 ce->e_block_hash_p = &cache->c_block_hash[0];
557 ce->e_index_hash_p = &cache->c_index_hash[0];
558 ce->e_used = 1 + MB_CACHE_WRITER;
559 return ce;
560}
561
562
563/*
564 * mb_cache_entry_insert()
565 *
566 * Inserts an entry that was allocated using mb_cache_entry_alloc() into
567 * the cache. After this, the cache entry can be looked up, but is not yet
568 * in the lru list as the caller still holds a handle to it. Returns 0 on
569 * success, or -EBUSY if a cache entry for that device + inode exists
570 * already (this may happen after a failed lookup, but when another process
571 * has inserted the same cache entry in the meantime).
572 *
573 * @bdev: device the cache entry belongs to
574 * @block: block number
575 * @key: lookup key
576 */
577int
578mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
579 sector_t block, unsigned int key)
580{
581 struct mb_cache *cache = ce->e_cache;
582 unsigned int bucket;
583 struct hlist_bl_node *l;
584 struct hlist_bl_head *block_hash_p;
585 struct hlist_bl_head *index_hash_p;
586 struct mb_cache_entry *lce;
587
588 mb_assert(ce);
589 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
590 cache->c_bucket_bits);
591 block_hash_p = &cache->c_block_hash[bucket];
592 hlist_bl_lock(block_hash_p);
593 hlist_bl_for_each_entry(lce, l, block_hash_p, e_block_list) {
594 if (lce->e_bdev == bdev && lce->e_block == block) {
595 hlist_bl_unlock(block_hash_p);
596 return -EBUSY;
597 }
598 }
599 mb_assert(!__mb_cache_entry_is_block_hashed(ce));
600 __mb_cache_entry_unhash_block(ce);
601 __mb_cache_entry_unhash_index(ce);
602 ce->e_bdev = bdev;
603 ce->e_block = block;
604 ce->e_block_hash_p = block_hash_p;
605 ce->e_index.o_key = key;
606 hlist_bl_add_head(&ce->e_block_list, block_hash_p);
607 hlist_bl_unlock(block_hash_p);
608 bucket = hash_long(key, cache->c_bucket_bits);
609 index_hash_p = &cache->c_index_hash[bucket];
610 hlist_bl_lock(index_hash_p);
611 ce->e_index_hash_p = index_hash_p;
612 hlist_bl_add_head(&ce->e_index.o_list, index_hash_p);
613 hlist_bl_unlock(index_hash_p);
614 return 0;
615}
616
617
618/*
619 * mb_cache_entry_release()
620 *
621 * Release a handle to a cache entry. When the last handle to a cache entry
622 * is released it is either freed (if it is invalid) or otherwise inserted
623 * in to the lru list.
624 */
625void
626mb_cache_entry_release(struct mb_cache_entry *ce)
627{
628 __mb_cache_entry_release(ce);
629}
630
631
632/*
633 * mb_cache_entry_free()
634 *
635 */
636void
637mb_cache_entry_free(struct mb_cache_entry *ce)
638{
639 mb_assert(ce);
640 mb_assert(list_empty(&ce->e_lru_list));
641 hlist_bl_lock(ce->e_index_hash_p);
642 __mb_cache_entry_unhash_index(ce);
643 hlist_bl_unlock(ce->e_index_hash_p);
644 hlist_bl_lock(ce->e_block_hash_p);
645 __mb_cache_entry_unhash_block(ce);
646 hlist_bl_unlock(ce->e_block_hash_p);
647 __mb_cache_entry_release(ce);
648}
649
650
651/*
652 * mb_cache_entry_get()
653 *
654 * Get a cache entry by device / block number. (There can only be one entry
655 * in the cache per device and block.) Returns NULL if no such cache entry
656 * exists. The returned cache entry is locked for exclusive access ("single
657 * writer").
658 */
659struct mb_cache_entry *
660mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
661 sector_t block)
662{
663 unsigned int bucket;
664 struct hlist_bl_node *l;
665 struct mb_cache_entry *ce;
666 struct hlist_bl_head *block_hash_p;
667
668 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
669 cache->c_bucket_bits);
670 block_hash_p = &cache->c_block_hash[bucket];
671 /* First serialize access to the block corresponding hash chain. */
672 hlist_bl_lock(block_hash_p);
673 hlist_bl_for_each_entry(ce, l, block_hash_p, e_block_list) {
674 mb_assert(ce->e_block_hash_p == block_hash_p);
675 if (ce->e_bdev == bdev && ce->e_block == block) {
676 /*
677 * Prevent a free from removing the entry.
678 */
679 atomic_inc(&ce->e_refcnt);
680 hlist_bl_unlock(block_hash_p);
681 __spin_lock_mb_cache_entry(ce);
682 atomic_dec(&ce->e_refcnt);
683 if (ce->e_used > 0) {
684 DEFINE_WAIT(wait);
685 while (ce->e_used > 0) {
686 ce->e_queued++;
687 prepare_to_wait(&mb_cache_queue, &wait,
688 TASK_UNINTERRUPTIBLE);
689 __spin_unlock_mb_cache_entry(ce);
690 schedule();
691 __spin_lock_mb_cache_entry(ce);
692 ce->e_queued--;
693 }
694 finish_wait(&mb_cache_queue, &wait);
695 }
696 ce->e_used += 1 + MB_CACHE_WRITER;
697 __spin_unlock_mb_cache_entry(ce);
698
699 if (!list_empty(&ce->e_lru_list)) {
700 spin_lock(&mb_cache_spinlock);
701 list_del_init(&ce->e_lru_list);
702 spin_unlock(&mb_cache_spinlock);
703 }
704 if (!__mb_cache_entry_is_block_hashed(ce)) {
705 __mb_cache_entry_release(ce);
706 return NULL;
707 }
708 return ce;
709 }
710 }
711 hlist_bl_unlock(block_hash_p);
712 return NULL;
713}
714
715#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
716
717static struct mb_cache_entry *
718__mb_cache_entry_find(struct hlist_bl_node *l, struct hlist_bl_head *head,
719 struct block_device *bdev, unsigned int key)
720{
721
722 /* The index hash chain is alredy acquire by caller. */
723 while (l != NULL) {
724 struct mb_cache_entry *ce =
725 hlist_bl_entry(l, struct mb_cache_entry,
726 e_index.o_list);
727 mb_assert(ce->e_index_hash_p == head);
728 if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
729 /*
730 * Prevent a free from removing the entry.
731 */
732 atomic_inc(&ce->e_refcnt);
733 hlist_bl_unlock(head);
734 __spin_lock_mb_cache_entry(ce);
735 atomic_dec(&ce->e_refcnt);
736 ce->e_used++;
737 /* Incrementing before holding the lock gives readers
738 priority over writers. */
739 if (ce->e_used >= MB_CACHE_WRITER) {
740 DEFINE_WAIT(wait);
741
742 while (ce->e_used >= MB_CACHE_WRITER) {
743 ce->e_queued++;
744 prepare_to_wait(&mb_cache_queue, &wait,
745 TASK_UNINTERRUPTIBLE);
746 __spin_unlock_mb_cache_entry(ce);
747 schedule();
748 __spin_lock_mb_cache_entry(ce);
749 ce->e_queued--;
750 }
751 finish_wait(&mb_cache_queue, &wait);
752 }
753 __spin_unlock_mb_cache_entry(ce);
754 if (!list_empty(&ce->e_lru_list)) {
755 spin_lock(&mb_cache_spinlock);
756 list_del_init(&ce->e_lru_list);
757 spin_unlock(&mb_cache_spinlock);
758 }
759 if (!__mb_cache_entry_is_block_hashed(ce)) {
760 __mb_cache_entry_release(ce);
761 return ERR_PTR(-EAGAIN);
762 }
763 return ce;
764 }
765 l = l->next;
766 }
767 hlist_bl_unlock(head);
768 return NULL;
769}
770
771
772/*
773 * mb_cache_entry_find_first()
774 *
775 * Find the first cache entry on a given device with a certain key in
776 * an additional index. Additional matches can be found with
777 * mb_cache_entry_find_next(). Returns NULL if no match was found. The
778 * returned cache entry is locked for shared access ("multiple readers").
779 *
780 * @cache: the cache to search
781 * @bdev: the device the cache entry should belong to
782 * @key: the key in the index
783 */
784struct mb_cache_entry *
785mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
786 unsigned int key)
787{
788 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
789 struct hlist_bl_node *l;
790 struct mb_cache_entry *ce = NULL;
791 struct hlist_bl_head *index_hash_p;
792
793 index_hash_p = &cache->c_index_hash[bucket];
794 hlist_bl_lock(index_hash_p);
795 if (!hlist_bl_empty(index_hash_p)) {
796 l = hlist_bl_first(index_hash_p);
797 ce = __mb_cache_entry_find(l, index_hash_p, bdev, key);
798 } else
799 hlist_bl_unlock(index_hash_p);
800 return ce;
801}
802
803
804/*
805 * mb_cache_entry_find_next()
806 *
807 * Find the next cache entry on a given device with a certain key in an
808 * additional index. Returns NULL if no match could be found. The previous
809 * entry is atomatically released, so that mb_cache_entry_find_next() can
810 * be called like this:
811 *
812 * entry = mb_cache_entry_find_first();
813 * while (entry) {
814 * ...
815 * entry = mb_cache_entry_find_next(entry, ...);
816 * }
817 *
818 * @prev: The previous match
819 * @bdev: the device the cache entry should belong to
820 * @key: the key in the index
821 */
822struct mb_cache_entry *
823mb_cache_entry_find_next(struct mb_cache_entry *prev,
824 struct block_device *bdev, unsigned int key)
825{
826 struct mb_cache *cache = prev->e_cache;
827 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
828 struct hlist_bl_node *l;
829 struct mb_cache_entry *ce;
830 struct hlist_bl_head *index_hash_p;
831
832 index_hash_p = &cache->c_index_hash[bucket];
833 mb_assert(prev->e_index_hash_p == index_hash_p);
834 hlist_bl_lock(index_hash_p);
835 mb_assert(!hlist_bl_empty(index_hash_p));
836 l = prev->e_index.o_list.next;
837 ce = __mb_cache_entry_find(l, index_hash_p, bdev, key);
838 __mb_cache_entry_release(prev);
839 return ce;
840}
841
842#endif /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
843
844static int __init init_mbcache(void)
845{
846 register_shrinker(&mb_cache_shrinker);
847 return 0;
848}
849
850static void __exit exit_mbcache(void)
851{
852 unregister_shrinker(&mb_cache_shrinker);
853}
854
855module_init(init_mbcache)
856module_exit(exit_mbcache)
857