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 | /* The bucket lock is selected based on the hash and protects mutations |
37 | * on a group of hash buckets. | |
38 | * | |
a5ec68e3 TG |
39 | * A maximum of tbl->size/2 bucket locks is allocated. This ensures that |
40 | * a single lock always covers both buckets which may both contains | |
41 | * entries which link to the same bucket of the old table during resizing. | |
42 | * This allows to simplify the locking as locking the bucket in both | |
43 | * tables during resize always guarantee protection. | |
44 | * | |
97defe1e TG |
45 | * IMPORTANT: When holding the bucket lock of both the old and new table |
46 | * during expansions and shrinking, the old bucket lock must always be | |
47 | * acquired first. | |
48 | */ | |
49 | static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash) | |
50 | { | |
51 | return &tbl->locks[hash & tbl->locks_mask]; | |
52 | } | |
7e1e7763 | 53 | |
c91eee56 | 54 | static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he) |
7e1e7763 TG |
55 | { |
56 | return (void *) he - ht->p.head_offset; | |
57 | } | |
7e1e7763 | 58 | |
8d24c0b4 | 59 | static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash) |
7e1e7763 | 60 | { |
ec9f71c5 | 61 | return (hash >> HASH_RESERVED_SPACE) & (tbl->size - 1); |
7e1e7763 TG |
62 | } |
63 | ||
aa34a6cb | 64 | static u32 key_hashfn(struct rhashtable *ht, const struct bucket_table *tbl, |
cffaa9cb | 65 | const void *key) |
7e1e7763 | 66 | { |
cffaa9cb | 67 | return rht_bucket_index(tbl, ht->p.hashfn(key, ht->p.key_len, |
ec9f71c5 | 68 | tbl->hash_rnd)); |
7e1e7763 | 69 | } |
7e1e7763 | 70 | |
988dfbd7 | 71 | static u32 head_hashfn(struct rhashtable *ht, |
8d24c0b4 TG |
72 | const struct bucket_table *tbl, |
73 | const struct rhash_head *he) | |
7e1e7763 | 74 | { |
ec9f71c5 HX |
75 | const char *ptr = rht_obj(ht, he); |
76 | ||
77 | return likely(ht->p.key_len) ? | |
78 | key_hashfn(ht, tbl, ptr + ht->p.key_offset) : | |
79 | rht_bucket_index(tbl, ht->p.obj_hashfn(ptr, tbl->hash_rnd)); | |
7e1e7763 TG |
80 | } |
81 | ||
a03eaec0 | 82 | #ifdef CONFIG_PROVE_LOCKING |
a03eaec0 | 83 | #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT)) |
a03eaec0 TG |
84 | |
85 | int lockdep_rht_mutex_is_held(struct rhashtable *ht) | |
86 | { | |
87 | return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1; | |
88 | } | |
89 | EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held); | |
90 | ||
91 | int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash) | |
92 | { | |
93 | spinlock_t *lock = bucket_lock(tbl, hash); | |
94 | ||
95 | return (debug_locks) ? lockdep_is_held(lock) : 1; | |
96 | } | |
97 | EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held); | |
98 | #else | |
99 | #define ASSERT_RHT_MUTEX(HT) | |
a03eaec0 TG |
100 | #endif |
101 | ||
102 | ||
97defe1e TG |
103 | static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl) |
104 | { | |
105 | unsigned int i, size; | |
106 | #if defined(CONFIG_PROVE_LOCKING) | |
107 | unsigned int nr_pcpus = 2; | |
108 | #else | |
109 | unsigned int nr_pcpus = num_possible_cpus(); | |
110 | #endif | |
111 | ||
112 | nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL); | |
113 | size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul); | |
114 | ||
a5ec68e3 TG |
115 | /* Never allocate more than 0.5 locks per bucket */ |
116 | size = min_t(unsigned int, size, tbl->size >> 1); | |
97defe1e TG |
117 | |
118 | if (sizeof(spinlock_t) != 0) { | |
119 | #ifdef CONFIG_NUMA | |
120 | if (size * sizeof(spinlock_t) > PAGE_SIZE) | |
121 | tbl->locks = vmalloc(size * sizeof(spinlock_t)); | |
122 | else | |
123 | #endif | |
124 | tbl->locks = kmalloc_array(size, sizeof(spinlock_t), | |
125 | GFP_KERNEL); | |
126 | if (!tbl->locks) | |
127 | return -ENOMEM; | |
128 | for (i = 0; i < size; i++) | |
129 | spin_lock_init(&tbl->locks[i]); | |
130 | } | |
131 | tbl->locks_mask = size - 1; | |
132 | ||
133 | return 0; | |
134 | } | |
135 | ||
136 | static void bucket_table_free(const struct bucket_table *tbl) | |
137 | { | |
138 | if (tbl) | |
139 | kvfree(tbl->locks); | |
140 | ||
141 | kvfree(tbl); | |
142 | } | |
143 | ||
9d901bc0 HX |
144 | static void bucket_table_free_rcu(struct rcu_head *head) |
145 | { | |
146 | bucket_table_free(container_of(head, struct bucket_table, rcu)); | |
147 | } | |
148 | ||
97defe1e | 149 | static struct bucket_table *bucket_table_alloc(struct rhashtable *ht, |
5269b53d | 150 | size_t nbuckets) |
7e1e7763 | 151 | { |
eb6d1abf | 152 | struct bucket_table *tbl = NULL; |
7e1e7763 | 153 | size_t size; |
f89bd6f8 | 154 | int i; |
7e1e7763 TG |
155 | |
156 | size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]); | |
eb6d1abf DB |
157 | if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) |
158 | tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); | |
7e1e7763 TG |
159 | if (tbl == NULL) |
160 | tbl = vzalloc(size); | |
7e1e7763 TG |
161 | if (tbl == NULL) |
162 | return NULL; | |
163 | ||
164 | tbl->size = nbuckets; | |
a5b6846f | 165 | tbl->shift = ilog2(nbuckets); |
7e1e7763 | 166 | |
97defe1e TG |
167 | if (alloc_bucket_locks(ht, tbl) < 0) { |
168 | bucket_table_free(tbl); | |
169 | return NULL; | |
170 | } | |
7e1e7763 | 171 | |
eddee5ba HX |
172 | INIT_LIST_HEAD(&tbl->walkers); |
173 | ||
5269b53d HX |
174 | get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd)); |
175 | ||
f89bd6f8 TG |
176 | for (i = 0; i < nbuckets; i++) |
177 | INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i); | |
178 | ||
97defe1e | 179 | return tbl; |
7e1e7763 TG |
180 | } |
181 | ||
182 | /** | |
183 | * rht_grow_above_75 - returns true if nelems > 0.75 * table-size | |
184 | * @ht: hash table | |
a5b6846f | 185 | * @tbl: current table |
7e1e7763 | 186 | */ |
a5b6846f DB |
187 | static bool rht_grow_above_75(const struct rhashtable *ht, |
188 | const struct bucket_table *tbl) | |
7e1e7763 TG |
189 | { |
190 | /* Expand table when exceeding 75% load */ | |
a5b6846f DB |
191 | return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) && |
192 | (!ht->p.max_shift || tbl->shift < ht->p.max_shift); | |
7e1e7763 | 193 | } |
7e1e7763 TG |
194 | |
195 | /** | |
196 | * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size | |
197 | * @ht: hash table | |
a5b6846f | 198 | * @tbl: current table |
7e1e7763 | 199 | */ |
a5b6846f DB |
200 | static bool rht_shrink_below_30(const struct rhashtable *ht, |
201 | const struct bucket_table *tbl) | |
7e1e7763 TG |
202 | { |
203 | /* Shrink table beneath 30% load */ | |
a5b6846f DB |
204 | return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) && |
205 | tbl->shift > ht->p.min_shift; | |
7e1e7763 | 206 | } |
7e1e7763 | 207 | |
aa34a6cb | 208 | static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash) |
a5ec68e3 | 209 | { |
aa34a6cb | 210 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
c4db8848 HX |
211 | struct bucket_table *new_tbl = |
212 | rht_dereference(old_tbl->future_tbl, ht) ?: old_tbl; | |
aa34a6cb HX |
213 | struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash]; |
214 | int err = -ENOENT; | |
215 | struct rhash_head *head, *next, *entry; | |
216 | spinlock_t *new_bucket_lock; | |
217 | unsigned new_hash; | |
218 | ||
219 | rht_for_each(entry, old_tbl, old_hash) { | |
220 | err = 0; | |
221 | next = rht_dereference_bucket(entry->next, old_tbl, old_hash); | |
222 | ||
223 | if (rht_is_a_nulls(next)) | |
224 | break; | |
a5ec68e3 | 225 | |
aa34a6cb HX |
226 | pprev = &entry->next; |
227 | } | |
a5ec68e3 | 228 | |
aa34a6cb HX |
229 | if (err) |
230 | goto out; | |
97defe1e | 231 | |
aa34a6cb | 232 | new_hash = head_hashfn(ht, new_tbl, entry); |
7e1e7763 | 233 | |
aa34a6cb | 234 | new_bucket_lock = bucket_lock(new_tbl, new_hash); |
7e1e7763 | 235 | |
8f2484bd | 236 | spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING); |
aa34a6cb HX |
237 | head = rht_dereference_bucket(new_tbl->buckets[new_hash], |
238 | new_tbl, new_hash); | |
97defe1e | 239 | |
aa34a6cb HX |
240 | if (rht_is_a_nulls(head)) |
241 | INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash); | |
242 | else | |
243 | RCU_INIT_POINTER(entry->next, head); | |
a5ec68e3 | 244 | |
aa34a6cb HX |
245 | rcu_assign_pointer(new_tbl->buckets[new_hash], entry); |
246 | spin_unlock(new_bucket_lock); | |
97defe1e | 247 | |
aa34a6cb | 248 | rcu_assign_pointer(*pprev, next); |
7e1e7763 | 249 | |
aa34a6cb HX |
250 | out: |
251 | return err; | |
252 | } | |
97defe1e | 253 | |
aa34a6cb HX |
254 | static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash) |
255 | { | |
256 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
257 | spinlock_t *old_bucket_lock; | |
258 | ||
259 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
a5ec68e3 | 260 | |
aa34a6cb HX |
261 | spin_lock_bh(old_bucket_lock); |
262 | while (!rhashtable_rehash_one(ht, old_hash)) | |
263 | ; | |
63d512d0 | 264 | old_tbl->rehash++; |
aa34a6cb | 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 | 271 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
eddee5ba | 272 | struct rhashtable_walker *walker; |
aa34a6cb | 273 | unsigned old_hash; |
7cd10db8 | 274 | |
aa34a6cb HX |
275 | /* Make insertions go into the new, empty table right away. Deletions |
276 | * and lookups will be attempted in both tables until we synchronize. | |
aa34a6cb | 277 | */ |
c4db8848 | 278 | rcu_assign_pointer(old_tbl->future_tbl, new_tbl); |
aa34a6cb | 279 | |
9497df88 HX |
280 | /* Ensure the new table is visible to readers. */ |
281 | smp_wmb(); | |
282 | ||
aa34a6cb HX |
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 | ||
eddee5ba HX |
289 | list_for_each_entry(walker, &old_tbl->walkers, list) |
290 | walker->tbl = NULL; | |
291 | ||
aa34a6cb HX |
292 | /* Wait for readers. All new readers will see the new |
293 | * table, and thus no references to the old table will | |
294 | * remain. | |
295 | */ | |
9d901bc0 | 296 | call_rcu(&old_tbl->rcu, bucket_table_free_rcu); |
7e1e7763 TG |
297 | } |
298 | ||
299 | /** | |
300 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
301 | * @ht: the hash table to expand | |
7e1e7763 | 302 | * |
aa34a6cb | 303 | * A secondary bucket array is allocated and the hash entries are migrated. |
7e1e7763 TG |
304 | * |
305 | * This function may only be called in a context where it is safe to call | |
306 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
307 | * | |
97defe1e TG |
308 | * The caller must ensure that no concurrent resizing occurs by holding |
309 | * ht->mutex. | |
310 | * | |
311 | * It is valid to have concurrent insertions and deletions protected by per | |
312 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 313 | */ |
6eba8224 | 314 | int rhashtable_expand(struct rhashtable *ht) |
7e1e7763 TG |
315 | { |
316 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
7e1e7763 TG |
317 | |
318 | ASSERT_RHT_MUTEX(ht); | |
319 | ||
5269b53d | 320 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2); |
7e1e7763 TG |
321 | if (new_tbl == NULL) |
322 | return -ENOMEM; | |
323 | ||
aa34a6cb | 324 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
325 | return 0; |
326 | } | |
327 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
328 | ||
329 | /** | |
330 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
331 | * @ht: the hash table to shrink | |
7e1e7763 TG |
332 | * |
333 | * This function may only be called in a context where it is safe to call | |
334 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
335 | * | |
97defe1e TG |
336 | * The caller must ensure that no concurrent resizing occurs by holding |
337 | * ht->mutex. | |
338 | * | |
7e1e7763 TG |
339 | * The caller must ensure that no concurrent table mutations take place. |
340 | * It is however valid to have concurrent lookups if they are RCU protected. | |
97defe1e TG |
341 | * |
342 | * It is valid to have concurrent insertions and deletions protected by per | |
343 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 344 | */ |
6eba8224 | 345 | int rhashtable_shrink(struct rhashtable *ht) |
7e1e7763 | 346 | { |
a5b6846f | 347 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); |
7e1e7763 TG |
348 | |
349 | ASSERT_RHT_MUTEX(ht); | |
350 | ||
5269b53d | 351 | new_tbl = bucket_table_alloc(ht, old_tbl->size / 2); |
97defe1e | 352 | if (new_tbl == NULL) |
7e1e7763 TG |
353 | return -ENOMEM; |
354 | ||
aa34a6cb | 355 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
356 | return 0; |
357 | } | |
358 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
359 | ||
97defe1e TG |
360 | static void rht_deferred_worker(struct work_struct *work) |
361 | { | |
362 | struct rhashtable *ht; | |
363 | struct bucket_table *tbl; | |
364 | ||
57699a40 | 365 | ht = container_of(work, struct rhashtable, run_work); |
97defe1e | 366 | mutex_lock(&ht->mutex); |
28134a53 HX |
367 | if (ht->being_destroyed) |
368 | goto unlock; | |
369 | ||
97defe1e TG |
370 | tbl = rht_dereference(ht->tbl, ht); |
371 | ||
a5b6846f | 372 | if (rht_grow_above_75(ht, tbl)) |
97defe1e | 373 | rhashtable_expand(ht); |
a5b6846f | 374 | else if (rht_shrink_below_30(ht, tbl)) |
97defe1e | 375 | rhashtable_shrink(ht); |
28134a53 | 376 | unlock: |
97defe1e TG |
377 | mutex_unlock(&ht->mutex); |
378 | } | |
379 | ||
aa34a6cb HX |
380 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
381 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 382 | { |
aa34a6cb | 383 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 384 | struct rhash_head *head; |
aa34a6cb HX |
385 | bool no_resize_running; |
386 | unsigned hash; | |
617011e7 | 387 | spinlock_t *old_lock; |
aa34a6cb HX |
388 | bool success = true; |
389 | ||
390 | rcu_read_lock(); | |
391 | ||
392 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
eca84933 | 393 | hash = head_hashfn(ht, old_tbl, obj); |
617011e7 | 394 | old_lock = bucket_lock(old_tbl, hash); |
aa34a6cb | 395 | |
617011e7 | 396 | spin_lock_bh(old_lock); |
aa34a6cb HX |
397 | |
398 | /* Because we have already taken the bucket lock in old_tbl, | |
399 | * if we find that future_tbl is not yet visible then that | |
400 | * guarantees all other insertions of the same entry will | |
401 | * also grab the bucket lock in old_tbl because until the | |
402 | * rehash completes ht->tbl won't be changed. | |
403 | */ | |
c4db8848 | 404 | tbl = rht_dereference_rcu(old_tbl->future_tbl, ht) ?: old_tbl; |
aa34a6cb | 405 | if (tbl != old_tbl) { |
eca84933 | 406 | hash = head_hashfn(ht, tbl, obj); |
8f2484bd | 407 | spin_lock_nested(bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING); |
aa34a6cb HX |
408 | } |
409 | ||
410 | if (compare && | |
411 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
412 | compare, arg)) { | |
413 | success = false; | |
414 | goto exit; | |
415 | } | |
416 | ||
417 | no_resize_running = tbl == old_tbl; | |
020219a6 | 418 | |
020219a6 | 419 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); |
db304854 YX |
420 | |
421 | if (rht_is_a_nulls(head)) | |
422 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
423 | else | |
424 | RCU_INIT_POINTER(obj->next, head); | |
425 | ||
426 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
427 | ||
428 | atomic_inc(&ht->nelems); | |
a5b6846f | 429 | if (no_resize_running && rht_grow_above_75(ht, tbl)) |
4c4b52d9 | 430 | schedule_work(&ht->run_work); |
aa34a6cb HX |
431 | |
432 | exit: | |
617011e7 | 433 | if (tbl != old_tbl) |
aa34a6cb | 434 | spin_unlock(bucket_lock(tbl, hash)); |
aa34a6cb | 435 | |
617011e7 | 436 | spin_unlock_bh(old_lock); |
aa34a6cb HX |
437 | |
438 | rcu_read_unlock(); | |
439 | ||
440 | return success; | |
db304854 YX |
441 | } |
442 | ||
7e1e7763 | 443 | /** |
db304854 | 444 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
445 | * @ht: hash table |
446 | * @obj: pointer to hash head inside object | |
7e1e7763 | 447 | * |
97defe1e TG |
448 | * Will take a per bucket spinlock to protect against mutual mutations |
449 | * on the same bucket. Multiple insertions may occur in parallel unless | |
450 | * they map to the same bucket lock. | |
7e1e7763 | 451 | * |
97defe1e TG |
452 | * It is safe to call this function from atomic context. |
453 | * | |
454 | * Will trigger an automatic deferred table resizing if the size grows | |
455 | * beyond the watermark indicated by grow_decision() which can be passed | |
456 | * to rhashtable_init(). | |
7e1e7763 | 457 | */ |
6eba8224 | 458 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 459 | { |
aa34a6cb HX |
460 | __rhashtable_insert(ht, obj, NULL, NULL); |
461 | } | |
462 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
463 | ||
464 | static bool __rhashtable_remove(struct rhashtable *ht, | |
465 | struct bucket_table *tbl, | |
466 | struct rhash_head *obj) | |
467 | { | |
468 | struct rhash_head __rcu **pprev; | |
469 | struct rhash_head *he; | |
470 | spinlock_t * lock; | |
97defe1e | 471 | unsigned hash; |
aa34a6cb | 472 | bool ret = false; |
7e1e7763 | 473 | |
eca84933 | 474 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb | 475 | lock = bucket_lock(tbl, hash); |
7e1e7763 | 476 | |
aa34a6cb | 477 | spin_lock_bh(lock); |
97defe1e | 478 | |
aa34a6cb HX |
479 | pprev = &tbl->buckets[hash]; |
480 | rht_for_each(he, tbl, hash) { | |
481 | if (he != obj) { | |
482 | pprev = &he->next; | |
483 | continue; | |
484 | } | |
7e1e7763 | 485 | |
aa34a6cb HX |
486 | rcu_assign_pointer(*pprev, obj->next); |
487 | ret = true; | |
488 | break; | |
489 | } | |
490 | ||
491 | spin_unlock_bh(lock); | |
492 | ||
493 | return ret; | |
7e1e7763 | 494 | } |
7e1e7763 | 495 | |
7e1e7763 TG |
496 | /** |
497 | * rhashtable_remove - remove object from hash table | |
498 | * @ht: hash table | |
499 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
500 | * |
501 | * Since the hash chain is single linked, the removal operation needs to | |
502 | * walk the bucket chain upon removal. The removal operation is thus | |
503 | * considerable slow if the hash table is not correctly sized. | |
504 | * | |
db304854 | 505 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
506 | * shrink_decision function specified at rhashtable_init() returns true. |
507 | * | |
508 | * The caller must ensure that no concurrent table mutations occur. It is | |
509 | * however valid to have concurrent lookups if they are RCU protected. | |
510 | */ | |
6eba8224 | 511 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 512 | { |
565e8640 | 513 | struct bucket_table *tbl; |
aa34a6cb | 514 | bool ret; |
7e1e7763 | 515 | |
97defe1e | 516 | rcu_read_lock(); |
7e1e7763 | 517 | |
565e8640 | 518 | tbl = rht_dereference_rcu(ht->tbl, ht); |
7e1e7763 | 519 | |
aa34a6cb HX |
520 | /* Because we have already taken (and released) the bucket |
521 | * lock in old_tbl, if we find that future_tbl is not yet | |
522 | * visible then that guarantees the entry to still be in | |
565e8640 | 523 | * the old tbl if it exists. |
fe6a043c | 524 | */ |
565e8640 HX |
525 | while (!(ret = __rhashtable_remove(ht, tbl, obj)) && |
526 | (tbl = rht_dereference_rcu(tbl->future_tbl, ht))) | |
527 | ; | |
fe6a043c TG |
528 | |
529 | if (ret) { | |
530 | atomic_dec(&ht->nelems); | |
565e8640 | 531 | if (rht_shrink_below_30(ht, tbl)) |
4c4b52d9 | 532 | schedule_work(&ht->run_work); |
fe6a043c TG |
533 | } |
534 | ||
97defe1e TG |
535 | rcu_read_unlock(); |
536 | ||
fe6a043c | 537 | return ret; |
7e1e7763 TG |
538 | } |
539 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
540 | ||
efb975a6 YX |
541 | struct rhashtable_compare_arg { |
542 | struct rhashtable *ht; | |
543 | const void *key; | |
544 | }; | |
545 | ||
546 | static bool rhashtable_compare(void *ptr, void *arg) | |
547 | { | |
548 | struct rhashtable_compare_arg *x = arg; | |
549 | struct rhashtable *ht = x->ht; | |
550 | ||
551 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
552 | } | |
553 | ||
7e1e7763 TG |
554 | /** |
555 | * rhashtable_lookup - lookup key in hash table | |
556 | * @ht: hash table | |
557 | * @key: pointer to key | |
558 | * | |
559 | * Computes the hash value for the key and traverses the bucket chain looking | |
560 | * for a entry with an identical key. The first matching entry is returned. | |
561 | * | |
562 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 563 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 564 | * |
97defe1e | 565 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 566 | */ |
97defe1e | 567 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 568 | { |
efb975a6 YX |
569 | struct rhashtable_compare_arg arg = { |
570 | .ht = ht, | |
571 | .key = key, | |
572 | }; | |
7e1e7763 TG |
573 | |
574 | BUG_ON(!ht->p.key_len); | |
575 | ||
efb975a6 | 576 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
577 | } |
578 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
579 | ||
580 | /** | |
581 | * rhashtable_lookup_compare - search hash table with compare function | |
582 | * @ht: hash table | |
8d24c0b4 | 583 | * @key: the pointer to the key |
7e1e7763 TG |
584 | * @compare: compare function, must return true on match |
585 | * @arg: argument passed on to compare function | |
586 | * | |
587 | * Traverses the bucket chain behind the provided hash value and calls the | |
588 | * specified compare function for each entry. | |
589 | * | |
97defe1e | 590 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
591 | * |
592 | * Returns the first entry on which the compare function returned true. | |
593 | */ | |
97defe1e | 594 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
595 | bool (*compare)(void *, void *), void *arg) |
596 | { | |
c4db8848 | 597 | const struct bucket_table *tbl; |
7e1e7763 | 598 | struct rhash_head *he; |
8d24c0b4 | 599 | u32 hash; |
7e1e7763 | 600 | |
97defe1e TG |
601 | rcu_read_lock(); |
602 | ||
aa34a6cb | 603 | tbl = rht_dereference_rcu(ht->tbl, ht); |
97defe1e | 604 | restart: |
39361947 | 605 | hash = key_hashfn(ht, tbl, key); |
8d2b1879 | 606 | rht_for_each_rcu(he, tbl, hash) { |
7e1e7763 TG |
607 | if (!compare(rht_obj(ht, he), arg)) |
608 | continue; | |
97defe1e | 609 | rcu_read_unlock(); |
a4b18cda | 610 | return rht_obj(ht, he); |
7e1e7763 TG |
611 | } |
612 | ||
9497df88 HX |
613 | /* Ensure we see any new tables. */ |
614 | smp_rmb(); | |
615 | ||
c4db8848 HX |
616 | tbl = rht_dereference_rcu(tbl->future_tbl, ht); |
617 | if (unlikely(tbl)) | |
97defe1e | 618 | goto restart; |
97defe1e TG |
619 | rcu_read_unlock(); |
620 | ||
7e1e7763 TG |
621 | return NULL; |
622 | } | |
623 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
624 | ||
db304854 YX |
625 | /** |
626 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
627 | * @ht: hash table | |
628 | * @obj: pointer to hash head inside object | |
629 | * | |
630 | * Locks down the bucket chain in both the old and new table if a resize | |
631 | * is in progress to ensure that writers can't remove from the old table | |
632 | * and can't insert to the new table during the atomic operation of search | |
633 | * and insertion. Searches for duplicates in both the old and new table if | |
634 | * a resize is in progress. | |
635 | * | |
636 | * This lookup function may only be used for fixed key hash table (key_len | |
637 | * parameter set). It will BUG() if used inappropriately. | |
638 | * | |
639 | * It is safe to call this function from atomic context. | |
640 | * | |
641 | * Will trigger an automatic deferred table resizing if the size grows | |
642 | * beyond the watermark indicated by grow_decision() which can be passed | |
643 | * to rhashtable_init(). | |
644 | */ | |
645 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
7a868d1e YX |
646 | { |
647 | struct rhashtable_compare_arg arg = { | |
648 | .ht = ht, | |
649 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
650 | }; | |
651 | ||
652 | BUG_ON(!ht->p.key_len); | |
653 | ||
654 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
655 | &arg); | |
656 | } | |
657 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
658 | ||
659 | /** | |
660 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
661 | * with compare function | |
662 | * @ht: hash table | |
663 | * @obj: pointer to hash head inside object | |
664 | * @compare: compare function, must return true on match | |
665 | * @arg: argument passed on to compare function | |
666 | * | |
667 | * Locks down the bucket chain in both the old and new table if a resize | |
668 | * is in progress to ensure that writers can't remove from the old table | |
669 | * and can't insert to the new table during the atomic operation of search | |
670 | * and insertion. Searches for duplicates in both the old and new table if | |
671 | * a resize is in progress. | |
672 | * | |
673 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
674 | * | |
675 | * Will trigger an automatic deferred table resizing if the size grows | |
676 | * beyond the watermark indicated by grow_decision() which can be passed | |
677 | * to rhashtable_init(). | |
678 | */ | |
679 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
680 | struct rhash_head *obj, | |
681 | bool (*compare)(void *, void *), | |
682 | void *arg) | |
db304854 | 683 | { |
db304854 YX |
684 | BUG_ON(!ht->p.key_len); |
685 | ||
aa34a6cb | 686 | return __rhashtable_insert(ht, obj, compare, arg); |
db304854 | 687 | } |
7a868d1e | 688 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); |
db304854 | 689 | |
f2dba9c6 HX |
690 | /** |
691 | * rhashtable_walk_init - Initialise an iterator | |
692 | * @ht: Table to walk over | |
693 | * @iter: Hash table Iterator | |
694 | * | |
695 | * This function prepares a hash table walk. | |
696 | * | |
697 | * Note that if you restart a walk after rhashtable_walk_stop you | |
698 | * may see the same object twice. Also, you may miss objects if | |
699 | * there are removals in between rhashtable_walk_stop and the next | |
700 | * call to rhashtable_walk_start. | |
701 | * | |
702 | * For a completely stable walk you should construct your own data | |
703 | * structure outside the hash table. | |
704 | * | |
705 | * This function may sleep so you must not call it from interrupt | |
706 | * context or with spin locks held. | |
707 | * | |
708 | * You must call rhashtable_walk_exit if this function returns | |
709 | * successfully. | |
710 | */ | |
711 | int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter) | |
712 | { | |
713 | iter->ht = ht; | |
714 | iter->p = NULL; | |
715 | iter->slot = 0; | |
716 | iter->skip = 0; | |
717 | ||
718 | iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL); | |
719 | if (!iter->walker) | |
720 | return -ENOMEM; | |
721 | ||
722 | mutex_lock(&ht->mutex); | |
eddee5ba HX |
723 | iter->walker->tbl = rht_dereference(ht->tbl, ht); |
724 | list_add(&iter->walker->list, &iter->walker->tbl->walkers); | |
f2dba9c6 HX |
725 | mutex_unlock(&ht->mutex); |
726 | ||
727 | return 0; | |
728 | } | |
729 | EXPORT_SYMBOL_GPL(rhashtable_walk_init); | |
730 | ||
731 | /** | |
732 | * rhashtable_walk_exit - Free an iterator | |
733 | * @iter: Hash table Iterator | |
734 | * | |
735 | * This function frees resources allocated by rhashtable_walk_init. | |
736 | */ | |
737 | void rhashtable_walk_exit(struct rhashtable_iter *iter) | |
738 | { | |
739 | mutex_lock(&iter->ht->mutex); | |
eddee5ba HX |
740 | if (iter->walker->tbl) |
741 | list_del(&iter->walker->list); | |
f2dba9c6 HX |
742 | mutex_unlock(&iter->ht->mutex); |
743 | kfree(iter->walker); | |
744 | } | |
745 | EXPORT_SYMBOL_GPL(rhashtable_walk_exit); | |
746 | ||
747 | /** | |
748 | * rhashtable_walk_start - Start a hash table walk | |
749 | * @iter: Hash table iterator | |
750 | * | |
751 | * Start a hash table walk. Note that we take the RCU lock in all | |
752 | * cases including when we return an error. So you must always call | |
753 | * rhashtable_walk_stop to clean up. | |
754 | * | |
755 | * Returns zero if successful. | |
756 | * | |
757 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
758 | * will rewind back to the beginning and you may use it immediately | |
759 | * by calling rhashtable_walk_next. | |
760 | */ | |
761 | int rhashtable_walk_start(struct rhashtable_iter *iter) | |
db4374f4 | 762 | __acquires(RCU) |
f2dba9c6 | 763 | { |
eddee5ba HX |
764 | struct rhashtable *ht = iter->ht; |
765 | ||
766 | mutex_lock(&ht->mutex); | |
767 | ||
768 | if (iter->walker->tbl) | |
769 | list_del(&iter->walker->list); | |
770 | ||
f2dba9c6 HX |
771 | rcu_read_lock(); |
772 | ||
eddee5ba HX |
773 | mutex_unlock(&ht->mutex); |
774 | ||
775 | if (!iter->walker->tbl) { | |
776 | iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht); | |
f2dba9c6 HX |
777 | return -EAGAIN; |
778 | } | |
779 | ||
780 | return 0; | |
781 | } | |
782 | EXPORT_SYMBOL_GPL(rhashtable_walk_start); | |
783 | ||
784 | /** | |
785 | * rhashtable_walk_next - Return the next object and advance the iterator | |
786 | * @iter: Hash table iterator | |
787 | * | |
788 | * Note that you must call rhashtable_walk_stop when you are finished | |
789 | * with the walk. | |
790 | * | |
791 | * Returns the next object or NULL when the end of the table is reached. | |
792 | * | |
793 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
794 | * will rewind back to the beginning and you may continue to use it. | |
795 | */ | |
796 | void *rhashtable_walk_next(struct rhashtable_iter *iter) | |
797 | { | |
eddee5ba | 798 | struct bucket_table *tbl = iter->walker->tbl; |
f2dba9c6 HX |
799 | struct rhashtable *ht = iter->ht; |
800 | struct rhash_head *p = iter->p; | |
801 | void *obj = NULL; | |
802 | ||
f2dba9c6 HX |
803 | if (p) { |
804 | p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot); | |
805 | goto next; | |
806 | } | |
807 | ||
808 | for (; iter->slot < tbl->size; iter->slot++) { | |
809 | int skip = iter->skip; | |
810 | ||
811 | rht_for_each_rcu(p, tbl, iter->slot) { | |
812 | if (!skip) | |
813 | break; | |
814 | skip--; | |
815 | } | |
816 | ||
817 | next: | |
818 | if (!rht_is_a_nulls(p)) { | |
819 | iter->skip++; | |
820 | iter->p = p; | |
821 | obj = rht_obj(ht, p); | |
822 | goto out; | |
823 | } | |
824 | ||
825 | iter->skip = 0; | |
826 | } | |
827 | ||
c4db8848 HX |
828 | iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht); |
829 | if (iter->walker->tbl) { | |
f2dba9c6 HX |
830 | iter->slot = 0; |
831 | iter->skip = 0; | |
f2dba9c6 HX |
832 | return ERR_PTR(-EAGAIN); |
833 | } | |
834 | ||
eddee5ba HX |
835 | iter->p = NULL; |
836 | ||
837 | out: | |
838 | ||
f2dba9c6 HX |
839 | return obj; |
840 | } | |
841 | EXPORT_SYMBOL_GPL(rhashtable_walk_next); | |
842 | ||
843 | /** | |
844 | * rhashtable_walk_stop - Finish a hash table walk | |
845 | * @iter: Hash table iterator | |
846 | * | |
847 | * Finish a hash table walk. | |
848 | */ | |
849 | void rhashtable_walk_stop(struct rhashtable_iter *iter) | |
db4374f4 | 850 | __releases(RCU) |
f2dba9c6 | 851 | { |
eddee5ba HX |
852 | struct rhashtable *ht; |
853 | struct bucket_table *tbl = iter->walker->tbl; | |
854 | ||
eddee5ba | 855 | if (!tbl) |
963ecbd4 | 856 | goto out; |
eddee5ba HX |
857 | |
858 | ht = iter->ht; | |
859 | ||
860 | mutex_lock(&ht->mutex); | |
c4db8848 | 861 | if (tbl->rehash < tbl->size) |
eddee5ba HX |
862 | list_add(&iter->walker->list, &tbl->walkers); |
863 | else | |
864 | iter->walker->tbl = NULL; | |
865 | mutex_unlock(&ht->mutex); | |
866 | ||
f2dba9c6 | 867 | iter->p = NULL; |
963ecbd4 HX |
868 | |
869 | out: | |
870 | rcu_read_unlock(); | |
f2dba9c6 HX |
871 | } |
872 | EXPORT_SYMBOL_GPL(rhashtable_walk_stop); | |
873 | ||
94000176 | 874 | static size_t rounded_hashtable_size(struct rhashtable_params *params) |
7e1e7763 | 875 | { |
94000176 YX |
876 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), |
877 | 1UL << params->min_shift); | |
7e1e7763 TG |
878 | } |
879 | ||
880 | /** | |
881 | * rhashtable_init - initialize a new hash table | |
882 | * @ht: hash table to be initialized | |
883 | * @params: configuration parameters | |
884 | * | |
885 | * Initializes a new hash table based on the provided configuration | |
886 | * parameters. A table can be configured either with a variable or | |
887 | * fixed length key: | |
888 | * | |
889 | * Configuration Example 1: Fixed length keys | |
890 | * struct test_obj { | |
891 | * int key; | |
892 | * void * my_member; | |
893 | * struct rhash_head node; | |
894 | * }; | |
895 | * | |
896 | * struct rhashtable_params params = { | |
897 | * .head_offset = offsetof(struct test_obj, node), | |
898 | * .key_offset = offsetof(struct test_obj, key), | |
899 | * .key_len = sizeof(int), | |
87545899 | 900 | * .hashfn = jhash, |
f89bd6f8 | 901 | * .nulls_base = (1U << RHT_BASE_SHIFT), |
7e1e7763 TG |
902 | * }; |
903 | * | |
904 | * Configuration Example 2: Variable length keys | |
905 | * struct test_obj { | |
906 | * [...] | |
907 | * struct rhash_head node; | |
908 | * }; | |
909 | * | |
910 | * u32 my_hash_fn(const void *data, u32 seed) | |
911 | * { | |
912 | * struct test_obj *obj = data; | |
913 | * | |
914 | * return [... hash ...]; | |
915 | * } | |
916 | * | |
917 | * struct rhashtable_params params = { | |
918 | * .head_offset = offsetof(struct test_obj, node), | |
87545899 | 919 | * .hashfn = jhash, |
7e1e7763 | 920 | * .obj_hashfn = my_hash_fn, |
7e1e7763 TG |
921 | * }; |
922 | */ | |
923 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
924 | { | |
925 | struct bucket_table *tbl; | |
926 | size_t size; | |
927 | ||
928 | size = HASH_DEFAULT_SIZE; | |
929 | ||
930 | if ((params->key_len && !params->hashfn) || | |
931 | (!params->key_len && !params->obj_hashfn)) | |
932 | return -EINVAL; | |
933 | ||
f89bd6f8 TG |
934 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
935 | return -EINVAL; | |
936 | ||
94000176 YX |
937 | params->min_shift = max_t(size_t, params->min_shift, |
938 | ilog2(HASH_MIN_SIZE)); | |
939 | ||
7e1e7763 | 940 | if (params->nelem_hint) |
94000176 | 941 | size = rounded_hashtable_size(params); |
7e1e7763 | 942 | |
97defe1e TG |
943 | memset(ht, 0, sizeof(*ht)); |
944 | mutex_init(&ht->mutex); | |
945 | memcpy(&ht->p, params, sizeof(*params)); | |
946 | ||
947 | if (params->locks_mul) | |
948 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
949 | else | |
950 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
951 | ||
5269b53d | 952 | tbl = bucket_table_alloc(ht, size); |
7e1e7763 TG |
953 | if (tbl == NULL) |
954 | return -ENOMEM; | |
955 | ||
545a148e | 956 | atomic_set(&ht->nelems, 0); |
a5b6846f | 957 | |
7e1e7763 TG |
958 | RCU_INIT_POINTER(ht->tbl, tbl); |
959 | ||
4c4b52d9 | 960 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 961 | |
7e1e7763 TG |
962 | return 0; |
963 | } | |
964 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
965 | ||
966 | /** | |
967 | * rhashtable_destroy - destroy hash table | |
968 | * @ht: the hash table to destroy | |
969 | * | |
ae82ddcf PNA |
970 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
971 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
972 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 973 | */ |
97defe1e | 974 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 975 | { |
97defe1e TG |
976 | ht->being_destroyed = true; |
977 | ||
4c4b52d9 | 978 | cancel_work_sync(&ht->run_work); |
97defe1e | 979 | |
57699a40 | 980 | mutex_lock(&ht->mutex); |
97defe1e | 981 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 982 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
983 | } |
984 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |