]>
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 | ||
144 | static struct bucket_table *bucket_table_alloc(struct rhashtable *ht, | |
a5b6846f | 145 | size_t nbuckets, u32 hash_rnd) |
7e1e7763 | 146 | { |
eb6d1abf | 147 | struct bucket_table *tbl = NULL; |
7e1e7763 | 148 | size_t size; |
f89bd6f8 | 149 | int i; |
7e1e7763 TG |
150 | |
151 | size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]); | |
eb6d1abf DB |
152 | if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) |
153 | tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); | |
7e1e7763 TG |
154 | if (tbl == NULL) |
155 | tbl = vzalloc(size); | |
7e1e7763 TG |
156 | if (tbl == NULL) |
157 | return NULL; | |
158 | ||
159 | tbl->size = nbuckets; | |
a5b6846f DB |
160 | tbl->shift = ilog2(nbuckets); |
161 | tbl->hash_rnd = hash_rnd; | |
7e1e7763 | 162 | |
97defe1e TG |
163 | if (alloc_bucket_locks(ht, tbl) < 0) { |
164 | bucket_table_free(tbl); | |
165 | return NULL; | |
166 | } | |
7e1e7763 | 167 | |
eddee5ba HX |
168 | INIT_LIST_HEAD(&tbl->walkers); |
169 | ||
f89bd6f8 TG |
170 | for (i = 0; i < nbuckets; i++) |
171 | INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i); | |
172 | ||
97defe1e | 173 | return tbl; |
7e1e7763 TG |
174 | } |
175 | ||
176 | /** | |
177 | * rht_grow_above_75 - returns true if nelems > 0.75 * table-size | |
178 | * @ht: hash table | |
a5b6846f | 179 | * @tbl: current table |
7e1e7763 | 180 | */ |
a5b6846f DB |
181 | static bool rht_grow_above_75(const struct rhashtable *ht, |
182 | const struct bucket_table *tbl) | |
7e1e7763 TG |
183 | { |
184 | /* Expand table when exceeding 75% load */ | |
a5b6846f DB |
185 | return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) && |
186 | (!ht->p.max_shift || tbl->shift < ht->p.max_shift); | |
7e1e7763 | 187 | } |
7e1e7763 TG |
188 | |
189 | /** | |
190 | * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size | |
191 | * @ht: hash table | |
a5b6846f | 192 | * @tbl: current table |
7e1e7763 | 193 | */ |
a5b6846f DB |
194 | static bool rht_shrink_below_30(const struct rhashtable *ht, |
195 | const struct bucket_table *tbl) | |
7e1e7763 TG |
196 | { |
197 | /* Shrink table beneath 30% load */ | |
a5b6846f DB |
198 | return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) && |
199 | tbl->shift > ht->p.min_shift; | |
7e1e7763 | 200 | } |
7e1e7763 | 201 | |
aa34a6cb | 202 | static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash) |
a5ec68e3 | 203 | { |
aa34a6cb HX |
204 | struct bucket_table *new_tbl = rht_dereference(ht->future_tbl, ht); |
205 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
206 | struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash]; | |
207 | int err = -ENOENT; | |
208 | struct rhash_head *head, *next, *entry; | |
209 | spinlock_t *new_bucket_lock; | |
210 | unsigned new_hash; | |
211 | ||
212 | rht_for_each(entry, old_tbl, old_hash) { | |
213 | err = 0; | |
214 | next = rht_dereference_bucket(entry->next, old_tbl, old_hash); | |
215 | ||
216 | if (rht_is_a_nulls(next)) | |
217 | break; | |
a5ec68e3 | 218 | |
aa34a6cb HX |
219 | pprev = &entry->next; |
220 | } | |
a5ec68e3 | 221 | |
aa34a6cb HX |
222 | if (err) |
223 | goto out; | |
97defe1e | 224 | |
aa34a6cb | 225 | new_hash = head_hashfn(ht, new_tbl, entry); |
7e1e7763 | 226 | |
aa34a6cb | 227 | new_bucket_lock = bucket_lock(new_tbl, new_hash); |
7e1e7763 | 228 | |
8f2484bd | 229 | spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING); |
aa34a6cb HX |
230 | head = rht_dereference_bucket(new_tbl->buckets[new_hash], |
231 | new_tbl, new_hash); | |
97defe1e | 232 | |
aa34a6cb HX |
233 | if (rht_is_a_nulls(head)) |
234 | INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash); | |
235 | else | |
236 | RCU_INIT_POINTER(entry->next, head); | |
a5ec68e3 | 237 | |
aa34a6cb HX |
238 | rcu_assign_pointer(new_tbl->buckets[new_hash], entry); |
239 | spin_unlock(new_bucket_lock); | |
97defe1e | 240 | |
aa34a6cb | 241 | rcu_assign_pointer(*pprev, next); |
7e1e7763 | 242 | |
aa34a6cb HX |
243 | out: |
244 | return err; | |
245 | } | |
97defe1e | 246 | |
aa34a6cb HX |
247 | static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash) |
248 | { | |
249 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
250 | spinlock_t *old_bucket_lock; | |
251 | ||
252 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
a5ec68e3 | 253 | |
aa34a6cb HX |
254 | spin_lock_bh(old_bucket_lock); |
255 | while (!rhashtable_rehash_one(ht, old_hash)) | |
256 | ; | |
257 | spin_unlock_bh(old_bucket_lock); | |
97defe1e TG |
258 | } |
259 | ||
aa34a6cb HX |
260 | static void rhashtable_rehash(struct rhashtable *ht, |
261 | struct bucket_table *new_tbl) | |
97defe1e | 262 | { |
aa34a6cb | 263 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
eddee5ba | 264 | struct rhashtable_walker *walker; |
aa34a6cb | 265 | unsigned old_hash; |
7cd10db8 | 266 | |
aa34a6cb HX |
267 | get_random_bytes(&new_tbl->hash_rnd, sizeof(new_tbl->hash_rnd)); |
268 | ||
269 | /* Make insertions go into the new, empty table right away. Deletions | |
270 | * and lookups will be attempted in both tables until we synchronize. | |
271 | * The synchronize_rcu() guarantees for the new table to be picked up | |
272 | * so no new additions go into the old table while we relink. | |
273 | */ | |
274 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
275 | ||
9497df88 HX |
276 | /* Ensure the new table is visible to readers. */ |
277 | smp_wmb(); | |
278 | ||
aa34a6cb HX |
279 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) |
280 | rhashtable_rehash_chain(ht, old_hash); | |
281 | ||
282 | /* Publish the new table pointer. */ | |
283 | rcu_assign_pointer(ht->tbl, new_tbl); | |
284 | ||
eddee5ba HX |
285 | list_for_each_entry(walker, &old_tbl->walkers, list) |
286 | walker->tbl = NULL; | |
287 | ||
aa34a6cb HX |
288 | /* Wait for readers. All new readers will see the new |
289 | * table, and thus no references to the old table will | |
290 | * remain. | |
291 | */ | |
292 | synchronize_rcu(); | |
293 | ||
294 | bucket_table_free(old_tbl); | |
7e1e7763 TG |
295 | } |
296 | ||
297 | /** | |
298 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
299 | * @ht: the hash table to expand | |
7e1e7763 | 300 | * |
aa34a6cb | 301 | * A secondary bucket array is allocated and the hash entries are migrated. |
7e1e7763 TG |
302 | * |
303 | * This function may only be called in a context where it is safe to call | |
304 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
305 | * | |
97defe1e TG |
306 | * The caller must ensure that no concurrent resizing occurs by holding |
307 | * ht->mutex. | |
308 | * | |
309 | * It is valid to have concurrent insertions and deletions protected by per | |
310 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 311 | */ |
6eba8224 | 312 | int rhashtable_expand(struct rhashtable *ht) |
7e1e7763 TG |
313 | { |
314 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
7e1e7763 TG |
315 | |
316 | ASSERT_RHT_MUTEX(ht); | |
317 | ||
a5b6846f | 318 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, old_tbl->hash_rnd); |
7e1e7763 TG |
319 | if (new_tbl == NULL) |
320 | return -ENOMEM; | |
321 | ||
aa34a6cb | 322 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
323 | return 0; |
324 | } | |
325 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
326 | ||
327 | /** | |
328 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
329 | * @ht: the hash table to shrink | |
7e1e7763 TG |
330 | * |
331 | * This function may only be called in a context where it is safe to call | |
332 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
333 | * | |
97defe1e TG |
334 | * The caller must ensure that no concurrent resizing occurs by holding |
335 | * ht->mutex. | |
336 | * | |
7e1e7763 TG |
337 | * The caller must ensure that no concurrent table mutations take place. |
338 | * It is however valid to have concurrent lookups if they are RCU protected. | |
97defe1e TG |
339 | * |
340 | * It is valid to have concurrent insertions and deletions protected by per | |
341 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 342 | */ |
6eba8224 | 343 | int rhashtable_shrink(struct rhashtable *ht) |
7e1e7763 | 344 | { |
a5b6846f | 345 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); |
7e1e7763 TG |
346 | |
347 | ASSERT_RHT_MUTEX(ht); | |
348 | ||
a5b6846f | 349 | new_tbl = bucket_table_alloc(ht, old_tbl->size / 2, old_tbl->hash_rnd); |
97defe1e | 350 | if (new_tbl == NULL) |
7e1e7763 TG |
351 | return -ENOMEM; |
352 | ||
aa34a6cb | 353 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
354 | return 0; |
355 | } | |
356 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
357 | ||
97defe1e TG |
358 | static void rht_deferred_worker(struct work_struct *work) |
359 | { | |
360 | struct rhashtable *ht; | |
361 | struct bucket_table *tbl; | |
362 | ||
57699a40 | 363 | ht = container_of(work, struct rhashtable, run_work); |
97defe1e | 364 | mutex_lock(&ht->mutex); |
28134a53 HX |
365 | if (ht->being_destroyed) |
366 | goto unlock; | |
367 | ||
97defe1e TG |
368 | tbl = rht_dereference(ht->tbl, ht); |
369 | ||
a5b6846f | 370 | if (rht_grow_above_75(ht, tbl)) |
97defe1e | 371 | rhashtable_expand(ht); |
a5b6846f | 372 | else if (rht_shrink_below_30(ht, tbl)) |
97defe1e | 373 | rhashtable_shrink(ht); |
28134a53 | 374 | unlock: |
97defe1e TG |
375 | mutex_unlock(&ht->mutex); |
376 | } | |
377 | ||
aa34a6cb HX |
378 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
379 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 380 | { |
aa34a6cb | 381 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 382 | struct rhash_head *head; |
aa34a6cb HX |
383 | bool no_resize_running; |
384 | unsigned hash; | |
385 | bool success = true; | |
386 | ||
387 | rcu_read_lock(); | |
388 | ||
389 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
eca84933 | 390 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
391 | |
392 | spin_lock_bh(bucket_lock(old_tbl, hash)); | |
393 | ||
394 | /* Because we have already taken the bucket lock in old_tbl, | |
395 | * if we find that future_tbl is not yet visible then that | |
396 | * guarantees all other insertions of the same entry will | |
397 | * also grab the bucket lock in old_tbl because until the | |
398 | * rehash completes ht->tbl won't be changed. | |
399 | */ | |
400 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
401 | if (tbl != old_tbl) { | |
eca84933 | 402 | hash = head_hashfn(ht, tbl, obj); |
8f2484bd | 403 | spin_lock_nested(bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING); |
aa34a6cb HX |
404 | } |
405 | ||
406 | if (compare && | |
407 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
408 | compare, arg)) { | |
409 | success = false; | |
410 | goto exit; | |
411 | } | |
412 | ||
413 | no_resize_running = tbl == old_tbl; | |
020219a6 | 414 | |
020219a6 | 415 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); |
db304854 YX |
416 | |
417 | if (rht_is_a_nulls(head)) | |
418 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
419 | else | |
420 | RCU_INIT_POINTER(obj->next, head); | |
421 | ||
422 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
423 | ||
424 | atomic_inc(&ht->nelems); | |
a5b6846f | 425 | if (no_resize_running && rht_grow_above_75(ht, tbl)) |
4c4b52d9 | 426 | schedule_work(&ht->run_work); |
aa34a6cb HX |
427 | |
428 | exit: | |
429 | if (tbl != old_tbl) { | |
eca84933 | 430 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb HX |
431 | spin_unlock(bucket_lock(tbl, hash)); |
432 | } | |
433 | ||
eca84933 | 434 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
435 | spin_unlock_bh(bucket_lock(old_tbl, hash)); |
436 | ||
437 | rcu_read_unlock(); | |
438 | ||
439 | return success; | |
db304854 YX |
440 | } |
441 | ||
7e1e7763 | 442 | /** |
db304854 | 443 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
444 | * @ht: hash table |
445 | * @obj: pointer to hash head inside object | |
7e1e7763 | 446 | * |
97defe1e TG |
447 | * Will take a per bucket spinlock to protect against mutual mutations |
448 | * on the same bucket. Multiple insertions may occur in parallel unless | |
449 | * they map to the same bucket lock. | |
7e1e7763 | 450 | * |
97defe1e TG |
451 | * It is safe to call this function from atomic context. |
452 | * | |
453 | * Will trigger an automatic deferred table resizing if the size grows | |
454 | * beyond the watermark indicated by grow_decision() which can be passed | |
455 | * to rhashtable_init(). | |
7e1e7763 | 456 | */ |
6eba8224 | 457 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 458 | { |
aa34a6cb HX |
459 | __rhashtable_insert(ht, obj, NULL, NULL); |
460 | } | |
461 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
462 | ||
463 | static bool __rhashtable_remove(struct rhashtable *ht, | |
464 | struct bucket_table *tbl, | |
465 | struct rhash_head *obj) | |
466 | { | |
467 | struct rhash_head __rcu **pprev; | |
468 | struct rhash_head *he; | |
469 | spinlock_t * lock; | |
97defe1e | 470 | unsigned hash; |
aa34a6cb | 471 | bool ret = false; |
7e1e7763 | 472 | |
eca84933 | 473 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb | 474 | lock = bucket_lock(tbl, hash); |
7e1e7763 | 475 | |
aa34a6cb | 476 | spin_lock_bh(lock); |
97defe1e | 477 | |
aa34a6cb HX |
478 | pprev = &tbl->buckets[hash]; |
479 | rht_for_each(he, tbl, hash) { | |
480 | if (he != obj) { | |
481 | pprev = &he->next; | |
482 | continue; | |
483 | } | |
7e1e7763 | 484 | |
aa34a6cb HX |
485 | rcu_assign_pointer(*pprev, obj->next); |
486 | ret = true; | |
487 | break; | |
488 | } | |
489 | ||
490 | spin_unlock_bh(lock); | |
491 | ||
492 | return ret; | |
7e1e7763 | 493 | } |
7e1e7763 | 494 | |
7e1e7763 TG |
495 | /** |
496 | * rhashtable_remove - remove object from hash table | |
497 | * @ht: hash table | |
498 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
499 | * |
500 | * Since the hash chain is single linked, the removal operation needs to | |
501 | * walk the bucket chain upon removal. The removal operation is thus | |
502 | * considerable slow if the hash table is not correctly sized. | |
503 | * | |
db304854 | 504 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
505 | * shrink_decision function specified at rhashtable_init() returns true. |
506 | * | |
507 | * The caller must ensure that no concurrent table mutations occur. It is | |
508 | * however valid to have concurrent lookups if they are RCU protected. | |
509 | */ | |
6eba8224 | 510 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 511 | { |
aa34a6cb HX |
512 | struct bucket_table *tbl, *old_tbl; |
513 | bool ret; | |
7e1e7763 | 514 | |
97defe1e | 515 | rcu_read_lock(); |
7e1e7763 | 516 | |
aa34a6cb HX |
517 | old_tbl = rht_dereference_rcu(ht->tbl, ht); |
518 | ret = __rhashtable_remove(ht, old_tbl, obj); | |
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 | |
523 | * old_tbl if it exists. | |
fe6a043c | 524 | */ |
aa34a6cb HX |
525 | tbl = rht_dereference_rcu(ht->future_tbl, ht); |
526 | if (!ret && old_tbl != tbl) | |
527 | ret = __rhashtable_remove(ht, tbl, obj); | |
fe6a043c TG |
528 | |
529 | if (ret) { | |
aa34a6cb | 530 | bool no_resize_running = tbl == old_tbl; |
4c4b52d9 | 531 | |
fe6a043c | 532 | atomic_dec(&ht->nelems); |
a5b6846f | 533 | if (no_resize_running && rht_shrink_below_30(ht, tbl)) |
4c4b52d9 | 534 | schedule_work(&ht->run_work); |
fe6a043c TG |
535 | } |
536 | ||
97defe1e TG |
537 | rcu_read_unlock(); |
538 | ||
fe6a043c | 539 | return ret; |
7e1e7763 TG |
540 | } |
541 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
542 | ||
efb975a6 YX |
543 | struct rhashtable_compare_arg { |
544 | struct rhashtable *ht; | |
545 | const void *key; | |
546 | }; | |
547 | ||
548 | static bool rhashtable_compare(void *ptr, void *arg) | |
549 | { | |
550 | struct rhashtable_compare_arg *x = arg; | |
551 | struct rhashtable *ht = x->ht; | |
552 | ||
553 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
554 | } | |
555 | ||
7e1e7763 TG |
556 | /** |
557 | * rhashtable_lookup - lookup key in hash table | |
558 | * @ht: hash table | |
559 | * @key: pointer to key | |
560 | * | |
561 | * Computes the hash value for the key and traverses the bucket chain looking | |
562 | * for a entry with an identical key. The first matching entry is returned. | |
563 | * | |
564 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 565 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 566 | * |
97defe1e | 567 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 568 | */ |
97defe1e | 569 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 570 | { |
efb975a6 YX |
571 | struct rhashtable_compare_arg arg = { |
572 | .ht = ht, | |
573 | .key = key, | |
574 | }; | |
7e1e7763 TG |
575 | |
576 | BUG_ON(!ht->p.key_len); | |
577 | ||
efb975a6 | 578 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
579 | } |
580 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
581 | ||
582 | /** | |
583 | * rhashtable_lookup_compare - search hash table with compare function | |
584 | * @ht: hash table | |
8d24c0b4 | 585 | * @key: the pointer to the key |
7e1e7763 TG |
586 | * @compare: compare function, must return true on match |
587 | * @arg: argument passed on to compare function | |
588 | * | |
589 | * Traverses the bucket chain behind the provided hash value and calls the | |
590 | * specified compare function for each entry. | |
591 | * | |
97defe1e | 592 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
593 | * |
594 | * Returns the first entry on which the compare function returned true. | |
595 | */ | |
97defe1e | 596 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
597 | bool (*compare)(void *, void *), void *arg) |
598 | { | |
97defe1e | 599 | const struct bucket_table *tbl, *old_tbl; |
7e1e7763 | 600 | struct rhash_head *he; |
8d24c0b4 | 601 | u32 hash; |
7e1e7763 | 602 | |
97defe1e TG |
603 | rcu_read_lock(); |
604 | ||
aa34a6cb | 605 | tbl = rht_dereference_rcu(ht->tbl, ht); |
97defe1e | 606 | restart: |
39361947 | 607 | hash = key_hashfn(ht, tbl, key); |
8d2b1879 | 608 | rht_for_each_rcu(he, tbl, hash) { |
7e1e7763 TG |
609 | if (!compare(rht_obj(ht, he), arg)) |
610 | continue; | |
97defe1e | 611 | rcu_read_unlock(); |
a4b18cda | 612 | return rht_obj(ht, he); |
7e1e7763 TG |
613 | } |
614 | ||
9497df88 HX |
615 | /* Ensure we see any new tables. */ |
616 | smp_rmb(); | |
617 | ||
aa34a6cb HX |
618 | old_tbl = tbl; |
619 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
620 | if (unlikely(tbl != old_tbl)) | |
97defe1e | 621 | goto restart; |
97defe1e TG |
622 | rcu_read_unlock(); |
623 | ||
7e1e7763 TG |
624 | return NULL; |
625 | } | |
626 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
627 | ||
db304854 YX |
628 | /** |
629 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
630 | * @ht: hash table | |
631 | * @obj: pointer to hash head inside object | |
632 | * | |
633 | * Locks down the bucket chain in both the old and new table if a resize | |
634 | * is in progress to ensure that writers can't remove from the old table | |
635 | * and can't insert to the new table during the atomic operation of search | |
636 | * and insertion. Searches for duplicates in both the old and new table if | |
637 | * a resize is in progress. | |
638 | * | |
639 | * This lookup function may only be used for fixed key hash table (key_len | |
640 | * parameter set). It will BUG() if used inappropriately. | |
641 | * | |
642 | * It is safe to call this function from atomic context. | |
643 | * | |
644 | * Will trigger an automatic deferred table resizing if the size grows | |
645 | * beyond the watermark indicated by grow_decision() which can be passed | |
646 | * to rhashtable_init(). | |
647 | */ | |
648 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
7a868d1e YX |
649 | { |
650 | struct rhashtable_compare_arg arg = { | |
651 | .ht = ht, | |
652 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
653 | }; | |
654 | ||
655 | BUG_ON(!ht->p.key_len); | |
656 | ||
657 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
658 | &arg); | |
659 | } | |
660 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
661 | ||
662 | /** | |
663 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
664 | * with compare function | |
665 | * @ht: hash table | |
666 | * @obj: pointer to hash head inside object | |
667 | * @compare: compare function, must return true on match | |
668 | * @arg: argument passed on to compare function | |
669 | * | |
670 | * Locks down the bucket chain in both the old and new table if a resize | |
671 | * is in progress to ensure that writers can't remove from the old table | |
672 | * and can't insert to the new table during the atomic operation of search | |
673 | * and insertion. Searches for duplicates in both the old and new table if | |
674 | * a resize is in progress. | |
675 | * | |
676 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
677 | * | |
678 | * Will trigger an automatic deferred table resizing if the size grows | |
679 | * beyond the watermark indicated by grow_decision() which can be passed | |
680 | * to rhashtable_init(). | |
681 | */ | |
682 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
683 | struct rhash_head *obj, | |
684 | bool (*compare)(void *, void *), | |
685 | void *arg) | |
db304854 | 686 | { |
db304854 YX |
687 | BUG_ON(!ht->p.key_len); |
688 | ||
aa34a6cb | 689 | return __rhashtable_insert(ht, obj, compare, arg); |
db304854 | 690 | } |
7a868d1e | 691 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); |
db304854 | 692 | |
f2dba9c6 HX |
693 | /** |
694 | * rhashtable_walk_init - Initialise an iterator | |
695 | * @ht: Table to walk over | |
696 | * @iter: Hash table Iterator | |
697 | * | |
698 | * This function prepares a hash table walk. | |
699 | * | |
700 | * Note that if you restart a walk after rhashtable_walk_stop you | |
701 | * may see the same object twice. Also, you may miss objects if | |
702 | * there are removals in between rhashtable_walk_stop and the next | |
703 | * call to rhashtable_walk_start. | |
704 | * | |
705 | * For a completely stable walk you should construct your own data | |
706 | * structure outside the hash table. | |
707 | * | |
708 | * This function may sleep so you must not call it from interrupt | |
709 | * context or with spin locks held. | |
710 | * | |
711 | * You must call rhashtable_walk_exit if this function returns | |
712 | * successfully. | |
713 | */ | |
714 | int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter) | |
715 | { | |
716 | iter->ht = ht; | |
717 | iter->p = NULL; | |
718 | iter->slot = 0; | |
719 | iter->skip = 0; | |
720 | ||
721 | iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL); | |
722 | if (!iter->walker) | |
723 | return -ENOMEM; | |
724 | ||
725 | mutex_lock(&ht->mutex); | |
eddee5ba HX |
726 | iter->walker->tbl = rht_dereference(ht->tbl, ht); |
727 | list_add(&iter->walker->list, &iter->walker->tbl->walkers); | |
f2dba9c6 HX |
728 | mutex_unlock(&ht->mutex); |
729 | ||
730 | return 0; | |
731 | } | |
732 | EXPORT_SYMBOL_GPL(rhashtable_walk_init); | |
733 | ||
734 | /** | |
735 | * rhashtable_walk_exit - Free an iterator | |
736 | * @iter: Hash table Iterator | |
737 | * | |
738 | * This function frees resources allocated by rhashtable_walk_init. | |
739 | */ | |
740 | void rhashtable_walk_exit(struct rhashtable_iter *iter) | |
741 | { | |
742 | mutex_lock(&iter->ht->mutex); | |
eddee5ba HX |
743 | if (iter->walker->tbl) |
744 | list_del(&iter->walker->list); | |
f2dba9c6 HX |
745 | mutex_unlock(&iter->ht->mutex); |
746 | kfree(iter->walker); | |
747 | } | |
748 | EXPORT_SYMBOL_GPL(rhashtable_walk_exit); | |
749 | ||
750 | /** | |
751 | * rhashtable_walk_start - Start a hash table walk | |
752 | * @iter: Hash table iterator | |
753 | * | |
754 | * Start a hash table walk. Note that we take the RCU lock in all | |
755 | * cases including when we return an error. So you must always call | |
756 | * rhashtable_walk_stop to clean up. | |
757 | * | |
758 | * Returns zero if successful. | |
759 | * | |
760 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
761 | * will rewind back to the beginning and you may use it immediately | |
762 | * by calling rhashtable_walk_next. | |
763 | */ | |
764 | int rhashtable_walk_start(struct rhashtable_iter *iter) | |
765 | { | |
eddee5ba HX |
766 | struct rhashtable *ht = iter->ht; |
767 | ||
768 | mutex_lock(&ht->mutex); | |
769 | ||
770 | if (iter->walker->tbl) | |
771 | list_del(&iter->walker->list); | |
772 | ||
f2dba9c6 HX |
773 | rcu_read_lock(); |
774 | ||
eddee5ba HX |
775 | mutex_unlock(&ht->mutex); |
776 | ||
777 | if (!iter->walker->tbl) { | |
778 | iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht); | |
f2dba9c6 HX |
779 | return -EAGAIN; |
780 | } | |
781 | ||
782 | return 0; | |
783 | } | |
784 | EXPORT_SYMBOL_GPL(rhashtable_walk_start); | |
785 | ||
786 | /** | |
787 | * rhashtable_walk_next - Return the next object and advance the iterator | |
788 | * @iter: Hash table iterator | |
789 | * | |
790 | * Note that you must call rhashtable_walk_stop when you are finished | |
791 | * with the walk. | |
792 | * | |
793 | * Returns the next object or NULL when the end of the table is reached. | |
794 | * | |
795 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
796 | * will rewind back to the beginning and you may continue to use it. | |
797 | */ | |
798 | void *rhashtable_walk_next(struct rhashtable_iter *iter) | |
799 | { | |
eddee5ba | 800 | struct bucket_table *tbl = iter->walker->tbl; |
f2dba9c6 HX |
801 | struct rhashtable *ht = iter->ht; |
802 | struct rhash_head *p = iter->p; | |
803 | void *obj = NULL; | |
804 | ||
f2dba9c6 HX |
805 | if (p) { |
806 | p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot); | |
807 | goto next; | |
808 | } | |
809 | ||
810 | for (; iter->slot < tbl->size; iter->slot++) { | |
811 | int skip = iter->skip; | |
812 | ||
813 | rht_for_each_rcu(p, tbl, iter->slot) { | |
814 | if (!skip) | |
815 | break; | |
816 | skip--; | |
817 | } | |
818 | ||
819 | next: | |
820 | if (!rht_is_a_nulls(p)) { | |
821 | iter->skip++; | |
822 | iter->p = p; | |
823 | obj = rht_obj(ht, p); | |
824 | goto out; | |
825 | } | |
826 | ||
827 | iter->skip = 0; | |
828 | } | |
829 | ||
eddee5ba HX |
830 | iter->walker->tbl = rht_dereference_rcu(ht->future_tbl, ht); |
831 | if (iter->walker->tbl != tbl) { | |
f2dba9c6 HX |
832 | iter->slot = 0; |
833 | iter->skip = 0; | |
f2dba9c6 HX |
834 | return ERR_PTR(-EAGAIN); |
835 | } | |
836 | ||
eddee5ba HX |
837 | iter->walker->tbl = NULL; |
838 | iter->p = NULL; | |
839 | ||
840 | out: | |
841 | ||
f2dba9c6 HX |
842 | return obj; |
843 | } | |
844 | EXPORT_SYMBOL_GPL(rhashtable_walk_next); | |
845 | ||
846 | /** | |
847 | * rhashtable_walk_stop - Finish a hash table walk | |
848 | * @iter: Hash table iterator | |
849 | * | |
850 | * Finish a hash table walk. | |
851 | */ | |
852 | void rhashtable_walk_stop(struct rhashtable_iter *iter) | |
853 | { | |
eddee5ba HX |
854 | struct rhashtable *ht; |
855 | struct bucket_table *tbl = iter->walker->tbl; | |
856 | ||
f2dba9c6 | 857 | rcu_read_unlock(); |
eddee5ba HX |
858 | |
859 | if (!tbl) | |
860 | return; | |
861 | ||
862 | ht = iter->ht; | |
863 | ||
864 | mutex_lock(&ht->mutex); | |
865 | if (rht_dereference(ht->tbl, ht) == tbl || | |
866 | rht_dereference(ht->future_tbl, ht) == tbl) | |
867 | list_add(&iter->walker->list, &tbl->walkers); | |
868 | else | |
869 | iter->walker->tbl = NULL; | |
870 | mutex_unlock(&ht->mutex); | |
871 | ||
f2dba9c6 HX |
872 | iter->p = NULL; |
873 | } | |
874 | EXPORT_SYMBOL_GPL(rhashtable_walk_stop); | |
875 | ||
94000176 | 876 | static size_t rounded_hashtable_size(struct rhashtable_params *params) |
7e1e7763 | 877 | { |
94000176 YX |
878 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), |
879 | 1UL << params->min_shift); | |
7e1e7763 TG |
880 | } |
881 | ||
882 | /** | |
883 | * rhashtable_init - initialize a new hash table | |
884 | * @ht: hash table to be initialized | |
885 | * @params: configuration parameters | |
886 | * | |
887 | * Initializes a new hash table based on the provided configuration | |
888 | * parameters. A table can be configured either with a variable or | |
889 | * fixed length key: | |
890 | * | |
891 | * Configuration Example 1: Fixed length keys | |
892 | * struct test_obj { | |
893 | * int key; | |
894 | * void * my_member; | |
895 | * struct rhash_head node; | |
896 | * }; | |
897 | * | |
898 | * struct rhashtable_params params = { | |
899 | * .head_offset = offsetof(struct test_obj, node), | |
900 | * .key_offset = offsetof(struct test_obj, key), | |
901 | * .key_len = sizeof(int), | |
87545899 | 902 | * .hashfn = jhash, |
f89bd6f8 | 903 | * .nulls_base = (1U << RHT_BASE_SHIFT), |
7e1e7763 TG |
904 | * }; |
905 | * | |
906 | * Configuration Example 2: Variable length keys | |
907 | * struct test_obj { | |
908 | * [...] | |
909 | * struct rhash_head node; | |
910 | * }; | |
911 | * | |
912 | * u32 my_hash_fn(const void *data, u32 seed) | |
913 | * { | |
914 | * struct test_obj *obj = data; | |
915 | * | |
916 | * return [... hash ...]; | |
917 | * } | |
918 | * | |
919 | * struct rhashtable_params params = { | |
920 | * .head_offset = offsetof(struct test_obj, node), | |
87545899 | 921 | * .hashfn = jhash, |
7e1e7763 | 922 | * .obj_hashfn = my_hash_fn, |
7e1e7763 TG |
923 | * }; |
924 | */ | |
925 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
926 | { | |
927 | struct bucket_table *tbl; | |
928 | size_t size; | |
a5b6846f | 929 | u32 hash_rnd; |
7e1e7763 TG |
930 | |
931 | size = HASH_DEFAULT_SIZE; | |
932 | ||
933 | if ((params->key_len && !params->hashfn) || | |
934 | (!params->key_len && !params->obj_hashfn)) | |
935 | return -EINVAL; | |
936 | ||
f89bd6f8 TG |
937 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
938 | return -EINVAL; | |
939 | ||
94000176 YX |
940 | params->min_shift = max_t(size_t, params->min_shift, |
941 | ilog2(HASH_MIN_SIZE)); | |
942 | ||
7e1e7763 | 943 | if (params->nelem_hint) |
94000176 | 944 | size = rounded_hashtable_size(params); |
7e1e7763 | 945 | |
97defe1e TG |
946 | memset(ht, 0, sizeof(*ht)); |
947 | mutex_init(&ht->mutex); | |
948 | memcpy(&ht->p, params, sizeof(*params)); | |
949 | ||
950 | if (params->locks_mul) | |
951 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
952 | else | |
953 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
954 | ||
a5b6846f DB |
955 | get_random_bytes(&hash_rnd, sizeof(hash_rnd)); |
956 | ||
957 | tbl = bucket_table_alloc(ht, size, hash_rnd); | |
7e1e7763 TG |
958 | if (tbl == NULL) |
959 | return -ENOMEM; | |
960 | ||
545a148e | 961 | atomic_set(&ht->nelems, 0); |
a5b6846f | 962 | |
7e1e7763 | 963 | RCU_INIT_POINTER(ht->tbl, tbl); |
97defe1e | 964 | RCU_INIT_POINTER(ht->future_tbl, tbl); |
7e1e7763 | 965 | |
4c4b52d9 | 966 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 967 | |
7e1e7763 TG |
968 | return 0; |
969 | } | |
970 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
971 | ||
972 | /** | |
973 | * rhashtable_destroy - destroy hash table | |
974 | * @ht: the hash table to destroy | |
975 | * | |
ae82ddcf PNA |
976 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
977 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
978 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 979 | */ |
97defe1e | 980 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 981 | { |
97defe1e TG |
982 | ht->being_destroyed = true; |
983 | ||
4c4b52d9 | 984 | cancel_work_sync(&ht->run_work); |
97defe1e | 985 | |
57699a40 | 986 | mutex_lock(&ht->mutex); |
97defe1e | 987 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 988 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
989 | } |
990 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |