]>
Commit | Line | Data |
---|---|---|
7e1e7763 TG |
1 | /* |
2 | * Resizable, Scalable, Concurrent Hash Table | |
3 | * | |
a5ec68e3 | 4 | * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch> |
7e1e7763 TG |
5 | * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net> |
6 | * | |
7 | * Based on the following paper: | |
8 | * https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf | |
9 | * | |
10 | * Code partially derived from nft_hash | |
11 | * | |
12 | * This program is free software; you can redistribute it and/or modify | |
13 | * it under the terms of the GNU General Public License version 2 as | |
14 | * published by the Free Software Foundation. | |
15 | */ | |
16 | ||
17 | #include <linux/kernel.h> | |
18 | #include <linux/init.h> | |
19 | #include <linux/log2.h> | |
5beb5c90 | 20 | #include <linux/sched.h> |
7e1e7763 TG |
21 | #include <linux/slab.h> |
22 | #include <linux/vmalloc.h> | |
23 | #include <linux/mm.h> | |
87545899 | 24 | #include <linux/jhash.h> |
7e1e7763 TG |
25 | #include <linux/random.h> |
26 | #include <linux/rhashtable.h> | |
61d7b097 | 27 | #include <linux/err.h> |
7e1e7763 TG |
28 | |
29 | #define HASH_DEFAULT_SIZE 64UL | |
30 | #define HASH_MIN_SIZE 4UL | |
97defe1e TG |
31 | #define BUCKET_LOCKS_PER_CPU 128UL |
32 | ||
f89bd6f8 TG |
33 | /* Base bits plus 1 bit for nulls marker */ |
34 | #define HASH_RESERVED_SPACE (RHT_BASE_BITS + 1) | |
35 | ||
97defe1e TG |
36 | enum { |
37 | RHT_LOCK_NORMAL, | |
38 | RHT_LOCK_NESTED, | |
97defe1e TG |
39 | }; |
40 | ||
41 | /* The bucket lock is selected based on the hash and protects mutations | |
42 | * on a group of hash buckets. | |
43 | * | |
a5ec68e3 TG |
44 | * A maximum of tbl->size/2 bucket locks is allocated. This ensures that |
45 | * a single lock always covers both buckets which may both contains | |
46 | * entries which link to the same bucket of the old table during resizing. | |
47 | * This allows to simplify the locking as locking the bucket in both | |
48 | * tables during resize always guarantee protection. | |
49 | * | |
97defe1e TG |
50 | * IMPORTANT: When holding the bucket lock of both the old and new table |
51 | * during expansions and shrinking, the old bucket lock must always be | |
52 | * acquired first. | |
53 | */ | |
54 | static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash) | |
55 | { | |
56 | return &tbl->locks[hash & tbl->locks_mask]; | |
57 | } | |
7e1e7763 | 58 | |
c91eee56 | 59 | static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he) |
7e1e7763 TG |
60 | { |
61 | return (void *) he - ht->p.head_offset; | |
62 | } | |
7e1e7763 | 63 | |
8d24c0b4 | 64 | static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash) |
7e1e7763 | 65 | { |
8d24c0b4 | 66 | return hash & (tbl->size - 1); |
7e1e7763 | 67 | } |
7e1e7763 | 68 | |
aa34a6cb HX |
69 | static u32 obj_raw_hashfn(struct rhashtable *ht, |
70 | const struct bucket_table *tbl, const void *ptr) | |
7e1e7763 | 71 | { |
8d24c0b4 | 72 | u32 hash; |
7e1e7763 | 73 | |
8d24c0b4 | 74 | if (unlikely(!ht->p.key_len)) |
988dfbd7 | 75 | hash = ht->p.obj_hashfn(ptr, tbl->hash_rnd); |
8d24c0b4 TG |
76 | else |
77 | hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len, | |
988dfbd7 | 78 | tbl->hash_rnd); |
7e1e7763 | 79 | |
f89bd6f8 | 80 | return hash >> HASH_RESERVED_SPACE; |
7e1e7763 TG |
81 | } |
82 | ||
aa34a6cb HX |
83 | static u32 key_hashfn(struct rhashtable *ht, const struct bucket_table *tbl, |
84 | const void *key, u32 len) | |
7e1e7763 | 85 | { |
988dfbd7 | 86 | return ht->p.hashfn(key, len, tbl->hash_rnd) >> HASH_RESERVED_SPACE; |
7e1e7763 | 87 | } |
7e1e7763 | 88 | |
988dfbd7 | 89 | static u32 head_hashfn(struct rhashtable *ht, |
8d24c0b4 TG |
90 | const struct bucket_table *tbl, |
91 | const struct rhash_head *he) | |
7e1e7763 | 92 | { |
aa34a6cb | 93 | return rht_bucket_index(tbl, obj_raw_hashfn(ht, tbl, rht_obj(ht, he))); |
7e1e7763 TG |
94 | } |
95 | ||
a03eaec0 | 96 | #ifdef CONFIG_PROVE_LOCKING |
a03eaec0 | 97 | #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT)) |
a03eaec0 TG |
98 | |
99 | int lockdep_rht_mutex_is_held(struct rhashtable *ht) | |
100 | { | |
101 | return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1; | |
102 | } | |
103 | EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held); | |
104 | ||
105 | int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash) | |
106 | { | |
107 | spinlock_t *lock = bucket_lock(tbl, hash); | |
108 | ||
109 | return (debug_locks) ? lockdep_is_held(lock) : 1; | |
110 | } | |
111 | EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held); | |
112 | #else | |
113 | #define ASSERT_RHT_MUTEX(HT) | |
a03eaec0 TG |
114 | #endif |
115 | ||
116 | ||
97defe1e TG |
117 | static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl) |
118 | { | |
119 | unsigned int i, size; | |
120 | #if defined(CONFIG_PROVE_LOCKING) | |
121 | unsigned int nr_pcpus = 2; | |
122 | #else | |
123 | unsigned int nr_pcpus = num_possible_cpus(); | |
124 | #endif | |
125 | ||
126 | nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL); | |
127 | size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul); | |
128 | ||
a5ec68e3 TG |
129 | /* Never allocate more than 0.5 locks per bucket */ |
130 | size = min_t(unsigned int, size, tbl->size >> 1); | |
97defe1e TG |
131 | |
132 | if (sizeof(spinlock_t) != 0) { | |
133 | #ifdef CONFIG_NUMA | |
134 | if (size * sizeof(spinlock_t) > PAGE_SIZE) | |
135 | tbl->locks = vmalloc(size * sizeof(spinlock_t)); | |
136 | else | |
137 | #endif | |
138 | tbl->locks = kmalloc_array(size, sizeof(spinlock_t), | |
139 | GFP_KERNEL); | |
140 | if (!tbl->locks) | |
141 | return -ENOMEM; | |
142 | for (i = 0; i < size; i++) | |
143 | spin_lock_init(&tbl->locks[i]); | |
144 | } | |
145 | tbl->locks_mask = size - 1; | |
146 | ||
147 | return 0; | |
148 | } | |
149 | ||
150 | static void bucket_table_free(const struct bucket_table *tbl) | |
151 | { | |
152 | if (tbl) | |
153 | kvfree(tbl->locks); | |
154 | ||
155 | kvfree(tbl); | |
156 | } | |
157 | ||
158 | static struct bucket_table *bucket_table_alloc(struct rhashtable *ht, | |
159 | size_t nbuckets) | |
7e1e7763 | 160 | { |
eb6d1abf | 161 | struct bucket_table *tbl = NULL; |
7e1e7763 | 162 | size_t size; |
f89bd6f8 | 163 | int i; |
7e1e7763 TG |
164 | |
165 | size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]); | |
eb6d1abf DB |
166 | if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) |
167 | tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); | |
7e1e7763 TG |
168 | if (tbl == NULL) |
169 | tbl = vzalloc(size); | |
7e1e7763 TG |
170 | if (tbl == NULL) |
171 | return NULL; | |
172 | ||
173 | tbl->size = nbuckets; | |
174 | ||
97defe1e TG |
175 | if (alloc_bucket_locks(ht, tbl) < 0) { |
176 | bucket_table_free(tbl); | |
177 | return NULL; | |
178 | } | |
7e1e7763 | 179 | |
f89bd6f8 TG |
180 | for (i = 0; i < nbuckets; i++) |
181 | INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i); | |
182 | ||
97defe1e | 183 | return tbl; |
7e1e7763 TG |
184 | } |
185 | ||
186 | /** | |
187 | * rht_grow_above_75 - returns true if nelems > 0.75 * table-size | |
188 | * @ht: hash table | |
189 | * @new_size: new table size | |
190 | */ | |
4c4b52d9 | 191 | static bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size) |
7e1e7763 TG |
192 | { |
193 | /* Expand table when exceeding 75% load */ | |
c0c09bfd | 194 | return atomic_read(&ht->nelems) > (new_size / 4 * 3) && |
8331de75 | 195 | (!ht->p.max_shift || atomic_read(&ht->shift) < ht->p.max_shift); |
7e1e7763 | 196 | } |
7e1e7763 TG |
197 | |
198 | /** | |
199 | * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size | |
200 | * @ht: hash table | |
201 | * @new_size: new table size | |
202 | */ | |
4c4b52d9 | 203 | static bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size) |
7e1e7763 TG |
204 | { |
205 | /* Shrink table beneath 30% load */ | |
c0c09bfd YX |
206 | return atomic_read(&ht->nelems) < (new_size * 3 / 10) && |
207 | (atomic_read(&ht->shift) > ht->p.min_shift); | |
7e1e7763 | 208 | } |
7e1e7763 | 209 | |
aa34a6cb | 210 | static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash) |
a5ec68e3 | 211 | { |
aa34a6cb HX |
212 | struct bucket_table *new_tbl = rht_dereference(ht->future_tbl, ht); |
213 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
214 | struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash]; | |
215 | int err = -ENOENT; | |
216 | struct rhash_head *head, *next, *entry; | |
217 | spinlock_t *new_bucket_lock; | |
218 | unsigned new_hash; | |
219 | ||
220 | rht_for_each(entry, old_tbl, old_hash) { | |
221 | err = 0; | |
222 | next = rht_dereference_bucket(entry->next, old_tbl, old_hash); | |
223 | ||
224 | if (rht_is_a_nulls(next)) | |
225 | break; | |
a5ec68e3 | 226 | |
aa34a6cb HX |
227 | pprev = &entry->next; |
228 | } | |
a5ec68e3 | 229 | |
aa34a6cb HX |
230 | if (err) |
231 | goto out; | |
97defe1e | 232 | |
aa34a6cb | 233 | new_hash = head_hashfn(ht, new_tbl, entry); |
7e1e7763 | 234 | |
aa34a6cb | 235 | new_bucket_lock = bucket_lock(new_tbl, new_hash); |
7e1e7763 | 236 | |
84ed82b7 | 237 | spin_lock_nested(new_bucket_lock, RHT_LOCK_NESTED); |
aa34a6cb HX |
238 | head = rht_dereference_bucket(new_tbl->buckets[new_hash], |
239 | new_tbl, new_hash); | |
97defe1e | 240 | |
aa34a6cb HX |
241 | if (rht_is_a_nulls(head)) |
242 | INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash); | |
243 | else | |
244 | RCU_INIT_POINTER(entry->next, head); | |
a5ec68e3 | 245 | |
aa34a6cb HX |
246 | rcu_assign_pointer(new_tbl->buckets[new_hash], entry); |
247 | spin_unlock(new_bucket_lock); | |
97defe1e | 248 | |
aa34a6cb | 249 | rcu_assign_pointer(*pprev, next); |
7e1e7763 | 250 | |
aa34a6cb HX |
251 | out: |
252 | return err; | |
253 | } | |
97defe1e | 254 | |
aa34a6cb HX |
255 | static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash) |
256 | { | |
257 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
258 | spinlock_t *old_bucket_lock; | |
259 | ||
260 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
a5ec68e3 | 261 | |
aa34a6cb HX |
262 | spin_lock_bh(old_bucket_lock); |
263 | while (!rhashtable_rehash_one(ht, old_hash)) | |
264 | ; | |
265 | spin_unlock_bh(old_bucket_lock); | |
97defe1e TG |
266 | } |
267 | ||
aa34a6cb HX |
268 | static void rhashtable_rehash(struct rhashtable *ht, |
269 | struct bucket_table *new_tbl) | |
97defe1e | 270 | { |
aa34a6cb HX |
271 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
272 | unsigned old_hash; | |
7cd10db8 | 273 | |
aa34a6cb HX |
274 | get_random_bytes(&new_tbl->hash_rnd, sizeof(new_tbl->hash_rnd)); |
275 | ||
276 | /* Make insertions go into the new, empty table right away. Deletions | |
277 | * and lookups will be attempted in both tables until we synchronize. | |
278 | * The synchronize_rcu() guarantees for the new table to be picked up | |
279 | * so no new additions go into the old table while we relink. | |
280 | */ | |
281 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
282 | ||
283 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) | |
284 | rhashtable_rehash_chain(ht, old_hash); | |
285 | ||
286 | /* Publish the new table pointer. */ | |
287 | rcu_assign_pointer(ht->tbl, new_tbl); | |
288 | ||
289 | /* Wait for readers. All new readers will see the new | |
290 | * table, and thus no references to the old table will | |
291 | * remain. | |
292 | */ | |
293 | synchronize_rcu(); | |
294 | ||
295 | bucket_table_free(old_tbl); | |
7e1e7763 TG |
296 | } |
297 | ||
298 | /** | |
299 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
300 | * @ht: the hash table to expand | |
7e1e7763 | 301 | * |
aa34a6cb | 302 | * A secondary bucket array is allocated and the hash entries are migrated. |
7e1e7763 TG |
303 | * |
304 | * This function may only be called in a context where it is safe to call | |
305 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
306 | * | |
97defe1e TG |
307 | * The caller must ensure that no concurrent resizing occurs by holding |
308 | * ht->mutex. | |
309 | * | |
310 | * It is valid to have concurrent insertions and deletions protected by per | |
311 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 312 | */ |
6eba8224 | 313 | int rhashtable_expand(struct rhashtable *ht) |
7e1e7763 TG |
314 | { |
315 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
7e1e7763 TG |
316 | |
317 | ASSERT_RHT_MUTEX(ht); | |
318 | ||
97defe1e | 319 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2); |
7e1e7763 TG |
320 | if (new_tbl == NULL) |
321 | return -ENOMEM; | |
322 | ||
988dfbd7 HX |
323 | new_tbl->hash_rnd = old_tbl->hash_rnd; |
324 | ||
c0c09bfd | 325 | atomic_inc(&ht->shift); |
7e1e7763 | 326 | |
aa34a6cb | 327 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 | 328 | |
7e1e7763 TG |
329 | return 0; |
330 | } | |
331 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
332 | ||
333 | /** | |
334 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
335 | * @ht: the hash table to shrink | |
7e1e7763 TG |
336 | * |
337 | * This function may only be called in a context where it is safe to call | |
338 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
339 | * | |
97defe1e TG |
340 | * The caller must ensure that no concurrent resizing occurs by holding |
341 | * ht->mutex. | |
342 | * | |
7e1e7763 TG |
343 | * The caller must ensure that no concurrent table mutations take place. |
344 | * It is however valid to have concurrent lookups if they are RCU protected. | |
97defe1e TG |
345 | * |
346 | * It is valid to have concurrent insertions and deletions protected by per | |
347 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 348 | */ |
6eba8224 | 349 | int rhashtable_shrink(struct rhashtable *ht) |
7e1e7763 | 350 | { |
97defe1e | 351 | struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht); |
7e1e7763 TG |
352 | |
353 | ASSERT_RHT_MUTEX(ht); | |
354 | ||
97defe1e TG |
355 | new_tbl = bucket_table_alloc(ht, tbl->size / 2); |
356 | if (new_tbl == NULL) | |
7e1e7763 TG |
357 | return -ENOMEM; |
358 | ||
988dfbd7 HX |
359 | new_tbl->hash_rnd = tbl->hash_rnd; |
360 | ||
c0c09bfd | 361 | atomic_dec(&ht->shift); |
7e1e7763 | 362 | |
aa34a6cb | 363 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
364 | |
365 | return 0; | |
366 | } | |
367 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
368 | ||
97defe1e TG |
369 | static void rht_deferred_worker(struct work_struct *work) |
370 | { | |
371 | struct rhashtable *ht; | |
372 | struct bucket_table *tbl; | |
f2dba9c6 | 373 | struct rhashtable_walker *walker; |
97defe1e | 374 | |
57699a40 | 375 | ht = container_of(work, struct rhashtable, run_work); |
97defe1e | 376 | mutex_lock(&ht->mutex); |
28134a53 HX |
377 | if (ht->being_destroyed) |
378 | goto unlock; | |
379 | ||
97defe1e TG |
380 | tbl = rht_dereference(ht->tbl, ht); |
381 | ||
f2dba9c6 HX |
382 | list_for_each_entry(walker, &ht->walkers, list) |
383 | walker->resize = true; | |
384 | ||
4c4b52d9 | 385 | if (rht_grow_above_75(ht, tbl->size)) |
97defe1e | 386 | rhashtable_expand(ht); |
4c4b52d9 | 387 | else if (rht_shrink_below_30(ht, tbl->size)) |
97defe1e | 388 | rhashtable_shrink(ht); |
28134a53 | 389 | unlock: |
97defe1e TG |
390 | mutex_unlock(&ht->mutex); |
391 | } | |
392 | ||
aa34a6cb HX |
393 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
394 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 395 | { |
aa34a6cb | 396 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 397 | struct rhash_head *head; |
aa34a6cb HX |
398 | bool no_resize_running; |
399 | unsigned hash; | |
400 | bool success = true; | |
401 | ||
402 | rcu_read_lock(); | |
403 | ||
404 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
405 | hash = obj_raw_hashfn(ht, old_tbl, rht_obj(ht, obj)); | |
406 | ||
407 | spin_lock_bh(bucket_lock(old_tbl, hash)); | |
408 | ||
409 | /* Because we have already taken the bucket lock in old_tbl, | |
410 | * if we find that future_tbl is not yet visible then that | |
411 | * guarantees all other insertions of the same entry will | |
412 | * also grab the bucket lock in old_tbl because until the | |
413 | * rehash completes ht->tbl won't be changed. | |
414 | */ | |
415 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
416 | if (tbl != old_tbl) { | |
417 | hash = obj_raw_hashfn(ht, tbl, rht_obj(ht, obj)); | |
84ed82b7 | 418 | spin_lock_nested(bucket_lock(tbl, hash), RHT_LOCK_NESTED); |
aa34a6cb HX |
419 | } |
420 | ||
421 | if (compare && | |
422 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
423 | compare, arg)) { | |
424 | success = false; | |
425 | goto exit; | |
426 | } | |
427 | ||
428 | no_resize_running = tbl == old_tbl; | |
020219a6 TG |
429 | |
430 | hash = rht_bucket_index(tbl, hash); | |
431 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); | |
db304854 YX |
432 | |
433 | if (rht_is_a_nulls(head)) | |
434 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
435 | else | |
436 | RCU_INIT_POINTER(obj->next, head); | |
437 | ||
438 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
439 | ||
440 | atomic_inc(&ht->nelems); | |
4c4b52d9 DB |
441 | if (no_resize_running && rht_grow_above_75(ht, tbl->size)) |
442 | schedule_work(&ht->run_work); | |
aa34a6cb HX |
443 | |
444 | exit: | |
445 | if (tbl != old_tbl) { | |
446 | hash = obj_raw_hashfn(ht, tbl, rht_obj(ht, obj)); | |
447 | spin_unlock(bucket_lock(tbl, hash)); | |
448 | } | |
449 | ||
450 | hash = obj_raw_hashfn(ht, old_tbl, rht_obj(ht, obj)); | |
451 | spin_unlock_bh(bucket_lock(old_tbl, hash)); | |
452 | ||
453 | rcu_read_unlock(); | |
454 | ||
455 | return success; | |
db304854 YX |
456 | } |
457 | ||
7e1e7763 | 458 | /** |
db304854 | 459 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
460 | * @ht: hash table |
461 | * @obj: pointer to hash head inside object | |
7e1e7763 | 462 | * |
97defe1e TG |
463 | * Will take a per bucket spinlock to protect against mutual mutations |
464 | * on the same bucket. Multiple insertions may occur in parallel unless | |
465 | * they map to the same bucket lock. | |
7e1e7763 | 466 | * |
97defe1e TG |
467 | * It is safe to call this function from atomic context. |
468 | * | |
469 | * Will trigger an automatic deferred table resizing if the size grows | |
470 | * beyond the watermark indicated by grow_decision() which can be passed | |
471 | * to rhashtable_init(). | |
7e1e7763 | 472 | */ |
6eba8224 | 473 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 474 | { |
aa34a6cb HX |
475 | __rhashtable_insert(ht, obj, NULL, NULL); |
476 | } | |
477 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
478 | ||
479 | static bool __rhashtable_remove(struct rhashtable *ht, | |
480 | struct bucket_table *tbl, | |
481 | struct rhash_head *obj) | |
482 | { | |
483 | struct rhash_head __rcu **pprev; | |
484 | struct rhash_head *he; | |
485 | spinlock_t * lock; | |
97defe1e | 486 | unsigned hash; |
aa34a6cb | 487 | bool ret = false; |
7e1e7763 | 488 | |
aa34a6cb HX |
489 | hash = obj_raw_hashfn(ht, tbl, rht_obj(ht, obj)); |
490 | lock = bucket_lock(tbl, hash); | |
491 | hash = rht_bucket_index(tbl, hash); | |
7e1e7763 | 492 | |
aa34a6cb | 493 | spin_lock_bh(lock); |
97defe1e | 494 | |
aa34a6cb HX |
495 | pprev = &tbl->buckets[hash]; |
496 | rht_for_each(he, tbl, hash) { | |
497 | if (he != obj) { | |
498 | pprev = &he->next; | |
499 | continue; | |
500 | } | |
7e1e7763 | 501 | |
aa34a6cb HX |
502 | rcu_assign_pointer(*pprev, obj->next); |
503 | ret = true; | |
504 | break; | |
505 | } | |
506 | ||
507 | spin_unlock_bh(lock); | |
508 | ||
509 | return ret; | |
7e1e7763 | 510 | } |
7e1e7763 | 511 | |
7e1e7763 TG |
512 | /** |
513 | * rhashtable_remove - remove object from hash table | |
514 | * @ht: hash table | |
515 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
516 | * |
517 | * Since the hash chain is single linked, the removal operation needs to | |
518 | * walk the bucket chain upon removal. The removal operation is thus | |
519 | * considerable slow if the hash table is not correctly sized. | |
520 | * | |
db304854 | 521 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
522 | * shrink_decision function specified at rhashtable_init() returns true. |
523 | * | |
524 | * The caller must ensure that no concurrent table mutations occur. It is | |
525 | * however valid to have concurrent lookups if they are RCU protected. | |
526 | */ | |
6eba8224 | 527 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 528 | { |
aa34a6cb HX |
529 | struct bucket_table *tbl, *old_tbl; |
530 | bool ret; | |
7e1e7763 | 531 | |
97defe1e | 532 | rcu_read_lock(); |
7e1e7763 | 533 | |
aa34a6cb HX |
534 | old_tbl = rht_dereference_rcu(ht->tbl, ht); |
535 | ret = __rhashtable_remove(ht, old_tbl, obj); | |
7e1e7763 | 536 | |
aa34a6cb HX |
537 | /* Because we have already taken (and released) the bucket |
538 | * lock in old_tbl, if we find that future_tbl is not yet | |
539 | * visible then that guarantees the entry to still be in | |
540 | * old_tbl if it exists. | |
fe6a043c | 541 | */ |
aa34a6cb HX |
542 | tbl = rht_dereference_rcu(ht->future_tbl, ht); |
543 | if (!ret && old_tbl != tbl) | |
544 | ret = __rhashtable_remove(ht, tbl, obj); | |
fe6a043c TG |
545 | |
546 | if (ret) { | |
aa34a6cb | 547 | bool no_resize_running = tbl == old_tbl; |
4c4b52d9 | 548 | |
fe6a043c | 549 | atomic_dec(&ht->nelems); |
aa34a6cb | 550 | if (no_resize_running && rht_shrink_below_30(ht, tbl->size)) |
4c4b52d9 | 551 | schedule_work(&ht->run_work); |
fe6a043c TG |
552 | } |
553 | ||
97defe1e TG |
554 | rcu_read_unlock(); |
555 | ||
fe6a043c | 556 | return ret; |
7e1e7763 TG |
557 | } |
558 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
559 | ||
efb975a6 YX |
560 | struct rhashtable_compare_arg { |
561 | struct rhashtable *ht; | |
562 | const void *key; | |
563 | }; | |
564 | ||
565 | static bool rhashtable_compare(void *ptr, void *arg) | |
566 | { | |
567 | struct rhashtable_compare_arg *x = arg; | |
568 | struct rhashtable *ht = x->ht; | |
569 | ||
570 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
571 | } | |
572 | ||
7e1e7763 TG |
573 | /** |
574 | * rhashtable_lookup - lookup key in hash table | |
575 | * @ht: hash table | |
576 | * @key: pointer to key | |
577 | * | |
578 | * Computes the hash value for the key and traverses the bucket chain looking | |
579 | * for a entry with an identical key. The first matching entry is returned. | |
580 | * | |
581 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 582 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 583 | * |
97defe1e | 584 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 585 | */ |
97defe1e | 586 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 587 | { |
efb975a6 YX |
588 | struct rhashtable_compare_arg arg = { |
589 | .ht = ht, | |
590 | .key = key, | |
591 | }; | |
7e1e7763 TG |
592 | |
593 | BUG_ON(!ht->p.key_len); | |
594 | ||
efb975a6 | 595 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
596 | } |
597 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
598 | ||
599 | /** | |
600 | * rhashtable_lookup_compare - search hash table with compare function | |
601 | * @ht: hash table | |
8d24c0b4 | 602 | * @key: the pointer to the key |
7e1e7763 TG |
603 | * @compare: compare function, must return true on match |
604 | * @arg: argument passed on to compare function | |
605 | * | |
606 | * Traverses the bucket chain behind the provided hash value and calls the | |
607 | * specified compare function for each entry. | |
608 | * | |
97defe1e | 609 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
610 | * |
611 | * Returns the first entry on which the compare function returned true. | |
612 | */ | |
97defe1e | 613 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
614 | bool (*compare)(void *, void *), void *arg) |
615 | { | |
97defe1e | 616 | const struct bucket_table *tbl, *old_tbl; |
7e1e7763 | 617 | struct rhash_head *he; |
8d24c0b4 | 618 | u32 hash; |
7e1e7763 | 619 | |
97defe1e TG |
620 | rcu_read_lock(); |
621 | ||
aa34a6cb HX |
622 | tbl = rht_dereference_rcu(ht->tbl, ht); |
623 | hash = key_hashfn(ht, tbl, key, ht->p.key_len); | |
97defe1e TG |
624 | restart: |
625 | rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) { | |
7e1e7763 TG |
626 | if (!compare(rht_obj(ht, he), arg)) |
627 | continue; | |
97defe1e | 628 | rcu_read_unlock(); |
a4b18cda | 629 | return rht_obj(ht, he); |
7e1e7763 TG |
630 | } |
631 | ||
aa34a6cb HX |
632 | old_tbl = tbl; |
633 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
634 | if (unlikely(tbl != old_tbl)) | |
97defe1e | 635 | goto restart; |
97defe1e TG |
636 | rcu_read_unlock(); |
637 | ||
7e1e7763 TG |
638 | return NULL; |
639 | } | |
640 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
641 | ||
db304854 YX |
642 | /** |
643 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
644 | * @ht: hash table | |
645 | * @obj: pointer to hash head inside object | |
646 | * | |
647 | * Locks down the bucket chain in both the old and new table if a resize | |
648 | * is in progress to ensure that writers can't remove from the old table | |
649 | * and can't insert to the new table during the atomic operation of search | |
650 | * and insertion. Searches for duplicates in both the old and new table if | |
651 | * a resize is in progress. | |
652 | * | |
653 | * This lookup function may only be used for fixed key hash table (key_len | |
654 | * parameter set). It will BUG() if used inappropriately. | |
655 | * | |
656 | * It is safe to call this function from atomic context. | |
657 | * | |
658 | * Will trigger an automatic deferred table resizing if the size grows | |
659 | * beyond the watermark indicated by grow_decision() which can be passed | |
660 | * to rhashtable_init(). | |
661 | */ | |
662 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
7a868d1e YX |
663 | { |
664 | struct rhashtable_compare_arg arg = { | |
665 | .ht = ht, | |
666 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
667 | }; | |
668 | ||
669 | BUG_ON(!ht->p.key_len); | |
670 | ||
671 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
672 | &arg); | |
673 | } | |
674 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
675 | ||
676 | /** | |
677 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
678 | * with compare function | |
679 | * @ht: hash table | |
680 | * @obj: pointer to hash head inside object | |
681 | * @compare: compare function, must return true on match | |
682 | * @arg: argument passed on to compare function | |
683 | * | |
684 | * Locks down the bucket chain in both the old and new table if a resize | |
685 | * is in progress to ensure that writers can't remove from the old table | |
686 | * and can't insert to the new table during the atomic operation of search | |
687 | * and insertion. Searches for duplicates in both the old and new table if | |
688 | * a resize is in progress. | |
689 | * | |
690 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
691 | * | |
692 | * Will trigger an automatic deferred table resizing if the size grows | |
693 | * beyond the watermark indicated by grow_decision() which can be passed | |
694 | * to rhashtable_init(). | |
695 | */ | |
696 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
697 | struct rhash_head *obj, | |
698 | bool (*compare)(void *, void *), | |
699 | void *arg) | |
db304854 | 700 | { |
db304854 YX |
701 | BUG_ON(!ht->p.key_len); |
702 | ||
aa34a6cb | 703 | return __rhashtable_insert(ht, obj, compare, arg); |
db304854 | 704 | } |
7a868d1e | 705 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); |
db304854 | 706 | |
f2dba9c6 HX |
707 | /** |
708 | * rhashtable_walk_init - Initialise an iterator | |
709 | * @ht: Table to walk over | |
710 | * @iter: Hash table Iterator | |
711 | * | |
712 | * This function prepares a hash table walk. | |
713 | * | |
714 | * Note that if you restart a walk after rhashtable_walk_stop you | |
715 | * may see the same object twice. Also, you may miss objects if | |
716 | * there are removals in between rhashtable_walk_stop and the next | |
717 | * call to rhashtable_walk_start. | |
718 | * | |
719 | * For a completely stable walk you should construct your own data | |
720 | * structure outside the hash table. | |
721 | * | |
722 | * This function may sleep so you must not call it from interrupt | |
723 | * context or with spin locks held. | |
724 | * | |
725 | * You must call rhashtable_walk_exit if this function returns | |
726 | * successfully. | |
727 | */ | |
728 | int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter) | |
729 | { | |
730 | iter->ht = ht; | |
731 | iter->p = NULL; | |
732 | iter->slot = 0; | |
733 | iter->skip = 0; | |
734 | ||
735 | iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL); | |
736 | if (!iter->walker) | |
737 | return -ENOMEM; | |
738 | ||
71bb0012 SL |
739 | INIT_LIST_HEAD(&iter->walker->list); |
740 | iter->walker->resize = false; | |
741 | ||
f2dba9c6 HX |
742 | mutex_lock(&ht->mutex); |
743 | list_add(&iter->walker->list, &ht->walkers); | |
744 | mutex_unlock(&ht->mutex); | |
745 | ||
746 | return 0; | |
747 | } | |
748 | EXPORT_SYMBOL_GPL(rhashtable_walk_init); | |
749 | ||
750 | /** | |
751 | * rhashtable_walk_exit - Free an iterator | |
752 | * @iter: Hash table Iterator | |
753 | * | |
754 | * This function frees resources allocated by rhashtable_walk_init. | |
755 | */ | |
756 | void rhashtable_walk_exit(struct rhashtable_iter *iter) | |
757 | { | |
758 | mutex_lock(&iter->ht->mutex); | |
759 | list_del(&iter->walker->list); | |
760 | mutex_unlock(&iter->ht->mutex); | |
761 | kfree(iter->walker); | |
762 | } | |
763 | EXPORT_SYMBOL_GPL(rhashtable_walk_exit); | |
764 | ||
765 | /** | |
766 | * rhashtable_walk_start - Start a hash table walk | |
767 | * @iter: Hash table iterator | |
768 | * | |
769 | * Start a hash table walk. Note that we take the RCU lock in all | |
770 | * cases including when we return an error. So you must always call | |
771 | * rhashtable_walk_stop to clean up. | |
772 | * | |
773 | * Returns zero if successful. | |
774 | * | |
775 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
776 | * will rewind back to the beginning and you may use it immediately | |
777 | * by calling rhashtable_walk_next. | |
778 | */ | |
779 | int rhashtable_walk_start(struct rhashtable_iter *iter) | |
780 | { | |
781 | rcu_read_lock(); | |
782 | ||
783 | if (iter->walker->resize) { | |
784 | iter->slot = 0; | |
785 | iter->skip = 0; | |
786 | iter->walker->resize = false; | |
787 | return -EAGAIN; | |
788 | } | |
789 | ||
790 | return 0; | |
791 | } | |
792 | EXPORT_SYMBOL_GPL(rhashtable_walk_start); | |
793 | ||
794 | /** | |
795 | * rhashtable_walk_next - Return the next object and advance the iterator | |
796 | * @iter: Hash table iterator | |
797 | * | |
798 | * Note that you must call rhashtable_walk_stop when you are finished | |
799 | * with the walk. | |
800 | * | |
801 | * Returns the next object or NULL when the end of the table is reached. | |
802 | * | |
803 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
804 | * will rewind back to the beginning and you may continue to use it. | |
805 | */ | |
806 | void *rhashtable_walk_next(struct rhashtable_iter *iter) | |
807 | { | |
808 | const struct bucket_table *tbl; | |
809 | struct rhashtable *ht = iter->ht; | |
810 | struct rhash_head *p = iter->p; | |
811 | void *obj = NULL; | |
812 | ||
813 | tbl = rht_dereference_rcu(ht->tbl, ht); | |
814 | ||
815 | if (p) { | |
816 | p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot); | |
817 | goto next; | |
818 | } | |
819 | ||
820 | for (; iter->slot < tbl->size; iter->slot++) { | |
821 | int skip = iter->skip; | |
822 | ||
823 | rht_for_each_rcu(p, tbl, iter->slot) { | |
824 | if (!skip) | |
825 | break; | |
826 | skip--; | |
827 | } | |
828 | ||
829 | next: | |
830 | if (!rht_is_a_nulls(p)) { | |
831 | iter->skip++; | |
832 | iter->p = p; | |
833 | obj = rht_obj(ht, p); | |
834 | goto out; | |
835 | } | |
836 | ||
837 | iter->skip = 0; | |
838 | } | |
839 | ||
840 | iter->p = NULL; | |
841 | ||
842 | out: | |
843 | if (iter->walker->resize) { | |
844 | iter->p = NULL; | |
845 | iter->slot = 0; | |
846 | iter->skip = 0; | |
847 | iter->walker->resize = false; | |
848 | return ERR_PTR(-EAGAIN); | |
849 | } | |
850 | ||
851 | return obj; | |
852 | } | |
853 | EXPORT_SYMBOL_GPL(rhashtable_walk_next); | |
854 | ||
855 | /** | |
856 | * rhashtable_walk_stop - Finish a hash table walk | |
857 | * @iter: Hash table iterator | |
858 | * | |
859 | * Finish a hash table walk. | |
860 | */ | |
861 | void rhashtable_walk_stop(struct rhashtable_iter *iter) | |
862 | { | |
863 | rcu_read_unlock(); | |
864 | iter->p = NULL; | |
865 | } | |
866 | EXPORT_SYMBOL_GPL(rhashtable_walk_stop); | |
867 | ||
94000176 | 868 | static size_t rounded_hashtable_size(struct rhashtable_params *params) |
7e1e7763 | 869 | { |
94000176 YX |
870 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), |
871 | 1UL << params->min_shift); | |
7e1e7763 TG |
872 | } |
873 | ||
874 | /** | |
875 | * rhashtable_init - initialize a new hash table | |
876 | * @ht: hash table to be initialized | |
877 | * @params: configuration parameters | |
878 | * | |
879 | * Initializes a new hash table based on the provided configuration | |
880 | * parameters. A table can be configured either with a variable or | |
881 | * fixed length key: | |
882 | * | |
883 | * Configuration Example 1: Fixed length keys | |
884 | * struct test_obj { | |
885 | * int key; | |
886 | * void * my_member; | |
887 | * struct rhash_head node; | |
888 | * }; | |
889 | * | |
890 | * struct rhashtable_params params = { | |
891 | * .head_offset = offsetof(struct test_obj, node), | |
892 | * .key_offset = offsetof(struct test_obj, key), | |
893 | * .key_len = sizeof(int), | |
87545899 | 894 | * .hashfn = jhash, |
f89bd6f8 | 895 | * .nulls_base = (1U << RHT_BASE_SHIFT), |
7e1e7763 TG |
896 | * }; |
897 | * | |
898 | * Configuration Example 2: Variable length keys | |
899 | * struct test_obj { | |
900 | * [...] | |
901 | * struct rhash_head node; | |
902 | * }; | |
903 | * | |
904 | * u32 my_hash_fn(const void *data, u32 seed) | |
905 | * { | |
906 | * struct test_obj *obj = data; | |
907 | * | |
908 | * return [... hash ...]; | |
909 | * } | |
910 | * | |
911 | * struct rhashtable_params params = { | |
912 | * .head_offset = offsetof(struct test_obj, node), | |
87545899 | 913 | * .hashfn = jhash, |
7e1e7763 | 914 | * .obj_hashfn = my_hash_fn, |
7e1e7763 TG |
915 | * }; |
916 | */ | |
917 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
918 | { | |
919 | struct bucket_table *tbl; | |
920 | size_t size; | |
921 | ||
922 | size = HASH_DEFAULT_SIZE; | |
923 | ||
924 | if ((params->key_len && !params->hashfn) || | |
925 | (!params->key_len && !params->obj_hashfn)) | |
926 | return -EINVAL; | |
927 | ||
f89bd6f8 TG |
928 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
929 | return -EINVAL; | |
930 | ||
94000176 YX |
931 | params->min_shift = max_t(size_t, params->min_shift, |
932 | ilog2(HASH_MIN_SIZE)); | |
933 | ||
7e1e7763 | 934 | if (params->nelem_hint) |
94000176 | 935 | size = rounded_hashtable_size(params); |
7e1e7763 | 936 | |
97defe1e TG |
937 | memset(ht, 0, sizeof(*ht)); |
938 | mutex_init(&ht->mutex); | |
939 | memcpy(&ht->p, params, sizeof(*params)); | |
f2dba9c6 | 940 | INIT_LIST_HEAD(&ht->walkers); |
97defe1e TG |
941 | |
942 | if (params->locks_mul) | |
943 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
944 | else | |
945 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
946 | ||
947 | tbl = bucket_table_alloc(ht, size); | |
7e1e7763 TG |
948 | if (tbl == NULL) |
949 | return -ENOMEM; | |
950 | ||
988dfbd7 HX |
951 | get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd)); |
952 | ||
545a148e | 953 | atomic_set(&ht->nelems, 0); |
c0c09bfd | 954 | atomic_set(&ht->shift, ilog2(tbl->size)); |
7e1e7763 | 955 | RCU_INIT_POINTER(ht->tbl, tbl); |
97defe1e | 956 | RCU_INIT_POINTER(ht->future_tbl, tbl); |
7e1e7763 | 957 | |
4c4b52d9 | 958 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 959 | |
7e1e7763 TG |
960 | return 0; |
961 | } | |
962 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
963 | ||
964 | /** | |
965 | * rhashtable_destroy - destroy hash table | |
966 | * @ht: the hash table to destroy | |
967 | * | |
ae82ddcf PNA |
968 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
969 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
970 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 971 | */ |
97defe1e | 972 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 973 | { |
97defe1e TG |
974 | ht->being_destroyed = true; |
975 | ||
4c4b52d9 | 976 | cancel_work_sync(&ht->run_work); |
97defe1e | 977 | |
57699a40 | 978 | mutex_lock(&ht->mutex); |
97defe1e | 979 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 980 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
981 | } |
982 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |