]>
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, | |
5269b53d | 145 | size_t nbuckets) |
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 | 160 | tbl->shift = ilog2(nbuckets); |
7e1e7763 | 161 | |
97defe1e TG |
162 | if (alloc_bucket_locks(ht, tbl) < 0) { |
163 | bucket_table_free(tbl); | |
164 | return NULL; | |
165 | } | |
7e1e7763 | 166 | |
eddee5ba HX |
167 | INIT_LIST_HEAD(&tbl->walkers); |
168 | ||
5269b53d HX |
169 | get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd)); |
170 | ||
f89bd6f8 TG |
171 | for (i = 0; i < nbuckets; i++) |
172 | INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i); | |
173 | ||
97defe1e | 174 | return tbl; |
7e1e7763 TG |
175 | } |
176 | ||
177 | /** | |
178 | * rht_grow_above_75 - returns true if nelems > 0.75 * table-size | |
179 | * @ht: hash table | |
a5b6846f | 180 | * @tbl: current table |
7e1e7763 | 181 | */ |
a5b6846f DB |
182 | static bool rht_grow_above_75(const struct rhashtable *ht, |
183 | const struct bucket_table *tbl) | |
7e1e7763 TG |
184 | { |
185 | /* Expand table when exceeding 75% load */ | |
a5b6846f DB |
186 | return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) && |
187 | (!ht->p.max_shift || tbl->shift < ht->p.max_shift); | |
7e1e7763 | 188 | } |
7e1e7763 TG |
189 | |
190 | /** | |
191 | * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size | |
192 | * @ht: hash table | |
a5b6846f | 193 | * @tbl: current table |
7e1e7763 | 194 | */ |
a5b6846f DB |
195 | static bool rht_shrink_below_30(const struct rhashtable *ht, |
196 | const struct bucket_table *tbl) | |
7e1e7763 TG |
197 | { |
198 | /* Shrink table beneath 30% load */ | |
a5b6846f DB |
199 | return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) && |
200 | tbl->shift > ht->p.min_shift; | |
7e1e7763 | 201 | } |
7e1e7763 | 202 | |
aa34a6cb | 203 | static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash) |
a5ec68e3 | 204 | { |
aa34a6cb HX |
205 | struct bucket_table *new_tbl = rht_dereference(ht->future_tbl, ht); |
206 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
207 | struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash]; | |
208 | int err = -ENOENT; | |
209 | struct rhash_head *head, *next, *entry; | |
210 | spinlock_t *new_bucket_lock; | |
211 | unsigned new_hash; | |
212 | ||
213 | rht_for_each(entry, old_tbl, old_hash) { | |
214 | err = 0; | |
215 | next = rht_dereference_bucket(entry->next, old_tbl, old_hash); | |
216 | ||
217 | if (rht_is_a_nulls(next)) | |
218 | break; | |
a5ec68e3 | 219 | |
aa34a6cb HX |
220 | pprev = &entry->next; |
221 | } | |
a5ec68e3 | 222 | |
aa34a6cb HX |
223 | if (err) |
224 | goto out; | |
97defe1e | 225 | |
aa34a6cb | 226 | new_hash = head_hashfn(ht, new_tbl, entry); |
7e1e7763 | 227 | |
aa34a6cb | 228 | new_bucket_lock = bucket_lock(new_tbl, new_hash); |
7e1e7763 | 229 | |
8f2484bd | 230 | spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING); |
aa34a6cb HX |
231 | head = rht_dereference_bucket(new_tbl->buckets[new_hash], |
232 | new_tbl, new_hash); | |
97defe1e | 233 | |
aa34a6cb HX |
234 | if (rht_is_a_nulls(head)) |
235 | INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash); | |
236 | else | |
237 | RCU_INIT_POINTER(entry->next, head); | |
a5ec68e3 | 238 | |
aa34a6cb HX |
239 | rcu_assign_pointer(new_tbl->buckets[new_hash], entry); |
240 | spin_unlock(new_bucket_lock); | |
97defe1e | 241 | |
aa34a6cb | 242 | rcu_assign_pointer(*pprev, next); |
7e1e7763 | 243 | |
aa34a6cb HX |
244 | out: |
245 | return err; | |
246 | } | |
97defe1e | 247 | |
aa34a6cb HX |
248 | static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash) |
249 | { | |
250 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
251 | spinlock_t *old_bucket_lock; | |
252 | ||
253 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
a5ec68e3 | 254 | |
aa34a6cb HX |
255 | spin_lock_bh(old_bucket_lock); |
256 | while (!rhashtable_rehash_one(ht, old_hash)) | |
257 | ; | |
258 | spin_unlock_bh(old_bucket_lock); | |
97defe1e TG |
259 | } |
260 | ||
aa34a6cb HX |
261 | static void rhashtable_rehash(struct rhashtable *ht, |
262 | struct bucket_table *new_tbl) | |
97defe1e | 263 | { |
aa34a6cb | 264 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
eddee5ba | 265 | struct rhashtable_walker *walker; |
aa34a6cb | 266 | unsigned old_hash; |
7cd10db8 | 267 | |
aa34a6cb HX |
268 | /* Make insertions go into the new, empty table right away. Deletions |
269 | * and lookups will be attempted in both tables until we synchronize. | |
270 | * The synchronize_rcu() guarantees for the new table to be picked up | |
271 | * so no new additions go into the old table while we relink. | |
272 | */ | |
273 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
274 | ||
9497df88 HX |
275 | /* Ensure the new table is visible to readers. */ |
276 | smp_wmb(); | |
277 | ||
aa34a6cb HX |
278 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) |
279 | rhashtable_rehash_chain(ht, old_hash); | |
280 | ||
281 | /* Publish the new table pointer. */ | |
282 | rcu_assign_pointer(ht->tbl, new_tbl); | |
283 | ||
eddee5ba HX |
284 | list_for_each_entry(walker, &old_tbl->walkers, list) |
285 | walker->tbl = NULL; | |
286 | ||
aa34a6cb HX |
287 | /* Wait for readers. All new readers will see the new |
288 | * table, and thus no references to the old table will | |
289 | * remain. | |
290 | */ | |
291 | synchronize_rcu(); | |
292 | ||
293 | bucket_table_free(old_tbl); | |
7e1e7763 TG |
294 | } |
295 | ||
296 | /** | |
297 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
298 | * @ht: the hash table to expand | |
7e1e7763 | 299 | * |
aa34a6cb | 300 | * A secondary bucket array is allocated and the hash entries are migrated. |
7e1e7763 TG |
301 | * |
302 | * This function may only be called in a context where it is safe to call | |
303 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
304 | * | |
97defe1e TG |
305 | * The caller must ensure that no concurrent resizing occurs by holding |
306 | * ht->mutex. | |
307 | * | |
308 | * It is valid to have concurrent insertions and deletions protected by per | |
309 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 310 | */ |
6eba8224 | 311 | int rhashtable_expand(struct rhashtable *ht) |
7e1e7763 TG |
312 | { |
313 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
7e1e7763 TG |
314 | |
315 | ASSERT_RHT_MUTEX(ht); | |
316 | ||
5269b53d | 317 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2); |
7e1e7763 TG |
318 | if (new_tbl == NULL) |
319 | return -ENOMEM; | |
320 | ||
aa34a6cb | 321 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
322 | return 0; |
323 | } | |
324 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
325 | ||
326 | /** | |
327 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
328 | * @ht: the hash table to shrink | |
7e1e7763 TG |
329 | * |
330 | * This function may only be called in a context where it is safe to call | |
331 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
332 | * | |
97defe1e TG |
333 | * The caller must ensure that no concurrent resizing occurs by holding |
334 | * ht->mutex. | |
335 | * | |
7e1e7763 TG |
336 | * The caller must ensure that no concurrent table mutations take place. |
337 | * It is however valid to have concurrent lookups if they are RCU protected. | |
97defe1e TG |
338 | * |
339 | * It is valid to have concurrent insertions and deletions protected by per | |
340 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 341 | */ |
6eba8224 | 342 | int rhashtable_shrink(struct rhashtable *ht) |
7e1e7763 | 343 | { |
a5b6846f | 344 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); |
7e1e7763 TG |
345 | |
346 | ASSERT_RHT_MUTEX(ht); | |
347 | ||
5269b53d | 348 | new_tbl = bucket_table_alloc(ht, old_tbl->size / 2); |
97defe1e | 349 | if (new_tbl == NULL) |
7e1e7763 TG |
350 | return -ENOMEM; |
351 | ||
aa34a6cb | 352 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
353 | return 0; |
354 | } | |
355 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
356 | ||
97defe1e TG |
357 | static void rht_deferred_worker(struct work_struct *work) |
358 | { | |
359 | struct rhashtable *ht; | |
360 | struct bucket_table *tbl; | |
361 | ||
57699a40 | 362 | ht = container_of(work, struct rhashtable, run_work); |
97defe1e | 363 | mutex_lock(&ht->mutex); |
28134a53 HX |
364 | if (ht->being_destroyed) |
365 | goto unlock; | |
366 | ||
97defe1e TG |
367 | tbl = rht_dereference(ht->tbl, ht); |
368 | ||
a5b6846f | 369 | if (rht_grow_above_75(ht, tbl)) |
97defe1e | 370 | rhashtable_expand(ht); |
a5b6846f | 371 | else if (rht_shrink_below_30(ht, tbl)) |
97defe1e | 372 | rhashtable_shrink(ht); |
28134a53 | 373 | unlock: |
97defe1e TG |
374 | mutex_unlock(&ht->mutex); |
375 | } | |
376 | ||
aa34a6cb HX |
377 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
378 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 379 | { |
aa34a6cb | 380 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 381 | struct rhash_head *head; |
aa34a6cb HX |
382 | bool no_resize_running; |
383 | unsigned hash; | |
384 | bool success = true; | |
385 | ||
386 | rcu_read_lock(); | |
387 | ||
388 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
eca84933 | 389 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
390 | |
391 | spin_lock_bh(bucket_lock(old_tbl, hash)); | |
392 | ||
393 | /* Because we have already taken the bucket lock in old_tbl, | |
394 | * if we find that future_tbl is not yet visible then that | |
395 | * guarantees all other insertions of the same entry will | |
396 | * also grab the bucket lock in old_tbl because until the | |
397 | * rehash completes ht->tbl won't be changed. | |
398 | */ | |
399 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
400 | if (tbl != old_tbl) { | |
eca84933 | 401 | hash = head_hashfn(ht, tbl, obj); |
8f2484bd | 402 | spin_lock_nested(bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING); |
aa34a6cb HX |
403 | } |
404 | ||
405 | if (compare && | |
406 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
407 | compare, arg)) { | |
408 | success = false; | |
409 | goto exit; | |
410 | } | |
411 | ||
412 | no_resize_running = tbl == old_tbl; | |
020219a6 | 413 | |
020219a6 | 414 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); |
db304854 YX |
415 | |
416 | if (rht_is_a_nulls(head)) | |
417 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
418 | else | |
419 | RCU_INIT_POINTER(obj->next, head); | |
420 | ||
421 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
422 | ||
423 | atomic_inc(&ht->nelems); | |
a5b6846f | 424 | if (no_resize_running && rht_grow_above_75(ht, tbl)) |
4c4b52d9 | 425 | schedule_work(&ht->run_work); |
aa34a6cb HX |
426 | |
427 | exit: | |
428 | if (tbl != old_tbl) { | |
eca84933 | 429 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb HX |
430 | spin_unlock(bucket_lock(tbl, hash)); |
431 | } | |
432 | ||
eca84933 | 433 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
434 | spin_unlock_bh(bucket_lock(old_tbl, hash)); |
435 | ||
436 | rcu_read_unlock(); | |
437 | ||
438 | return success; | |
db304854 YX |
439 | } |
440 | ||
7e1e7763 | 441 | /** |
db304854 | 442 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
443 | * @ht: hash table |
444 | * @obj: pointer to hash head inside object | |
7e1e7763 | 445 | * |
97defe1e TG |
446 | * Will take a per bucket spinlock to protect against mutual mutations |
447 | * on the same bucket. Multiple insertions may occur in parallel unless | |
448 | * they map to the same bucket lock. | |
7e1e7763 | 449 | * |
97defe1e TG |
450 | * It is safe to call this function from atomic context. |
451 | * | |
452 | * Will trigger an automatic deferred table resizing if the size grows | |
453 | * beyond the watermark indicated by grow_decision() which can be passed | |
454 | * to rhashtable_init(). | |
7e1e7763 | 455 | */ |
6eba8224 | 456 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 457 | { |
aa34a6cb HX |
458 | __rhashtable_insert(ht, obj, NULL, NULL); |
459 | } | |
460 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
461 | ||
462 | static bool __rhashtable_remove(struct rhashtable *ht, | |
463 | struct bucket_table *tbl, | |
464 | struct rhash_head *obj) | |
465 | { | |
466 | struct rhash_head __rcu **pprev; | |
467 | struct rhash_head *he; | |
468 | spinlock_t * lock; | |
97defe1e | 469 | unsigned hash; |
aa34a6cb | 470 | bool ret = false; |
7e1e7763 | 471 | |
eca84933 | 472 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb | 473 | lock = bucket_lock(tbl, hash); |
7e1e7763 | 474 | |
aa34a6cb | 475 | spin_lock_bh(lock); |
97defe1e | 476 | |
aa34a6cb HX |
477 | pprev = &tbl->buckets[hash]; |
478 | rht_for_each(he, tbl, hash) { | |
479 | if (he != obj) { | |
480 | pprev = &he->next; | |
481 | continue; | |
482 | } | |
7e1e7763 | 483 | |
aa34a6cb HX |
484 | rcu_assign_pointer(*pprev, obj->next); |
485 | ret = true; | |
486 | break; | |
487 | } | |
488 | ||
489 | spin_unlock_bh(lock); | |
490 | ||
491 | return ret; | |
7e1e7763 | 492 | } |
7e1e7763 | 493 | |
7e1e7763 TG |
494 | /** |
495 | * rhashtable_remove - remove object from hash table | |
496 | * @ht: hash table | |
497 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
498 | * |
499 | * Since the hash chain is single linked, the removal operation needs to | |
500 | * walk the bucket chain upon removal. The removal operation is thus | |
501 | * considerable slow if the hash table is not correctly sized. | |
502 | * | |
db304854 | 503 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
504 | * shrink_decision function specified at rhashtable_init() returns true. |
505 | * | |
506 | * The caller must ensure that no concurrent table mutations occur. It is | |
507 | * however valid to have concurrent lookups if they are RCU protected. | |
508 | */ | |
6eba8224 | 509 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 510 | { |
aa34a6cb HX |
511 | struct bucket_table *tbl, *old_tbl; |
512 | bool ret; | |
7e1e7763 | 513 | |
97defe1e | 514 | rcu_read_lock(); |
7e1e7763 | 515 | |
aa34a6cb HX |
516 | old_tbl = rht_dereference_rcu(ht->tbl, ht); |
517 | ret = __rhashtable_remove(ht, old_tbl, obj); | |
7e1e7763 | 518 | |
aa34a6cb HX |
519 | /* Because we have already taken (and released) the bucket |
520 | * lock in old_tbl, if we find that future_tbl is not yet | |
521 | * visible then that guarantees the entry to still be in | |
522 | * old_tbl if it exists. | |
fe6a043c | 523 | */ |
aa34a6cb HX |
524 | tbl = rht_dereference_rcu(ht->future_tbl, ht); |
525 | if (!ret && old_tbl != tbl) | |
526 | ret = __rhashtable_remove(ht, tbl, obj); | |
fe6a043c TG |
527 | |
528 | if (ret) { | |
aa34a6cb | 529 | bool no_resize_running = tbl == old_tbl; |
4c4b52d9 | 530 | |
fe6a043c | 531 | atomic_dec(&ht->nelems); |
a5b6846f | 532 | if (no_resize_running && rht_shrink_below_30(ht, tbl)) |
4c4b52d9 | 533 | schedule_work(&ht->run_work); |
fe6a043c TG |
534 | } |
535 | ||
97defe1e TG |
536 | rcu_read_unlock(); |
537 | ||
fe6a043c | 538 | return ret; |
7e1e7763 TG |
539 | } |
540 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
541 | ||
efb975a6 YX |
542 | struct rhashtable_compare_arg { |
543 | struct rhashtable *ht; | |
544 | const void *key; | |
545 | }; | |
546 | ||
547 | static bool rhashtable_compare(void *ptr, void *arg) | |
548 | { | |
549 | struct rhashtable_compare_arg *x = arg; | |
550 | struct rhashtable *ht = x->ht; | |
551 | ||
552 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
553 | } | |
554 | ||
7e1e7763 TG |
555 | /** |
556 | * rhashtable_lookup - lookup key in hash table | |
557 | * @ht: hash table | |
558 | * @key: pointer to key | |
559 | * | |
560 | * Computes the hash value for the key and traverses the bucket chain looking | |
561 | * for a entry with an identical key. The first matching entry is returned. | |
562 | * | |
563 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 564 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 565 | * |
97defe1e | 566 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 567 | */ |
97defe1e | 568 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 569 | { |
efb975a6 YX |
570 | struct rhashtable_compare_arg arg = { |
571 | .ht = ht, | |
572 | .key = key, | |
573 | }; | |
7e1e7763 TG |
574 | |
575 | BUG_ON(!ht->p.key_len); | |
576 | ||
efb975a6 | 577 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
578 | } |
579 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
580 | ||
581 | /** | |
582 | * rhashtable_lookup_compare - search hash table with compare function | |
583 | * @ht: hash table | |
8d24c0b4 | 584 | * @key: the pointer to the key |
7e1e7763 TG |
585 | * @compare: compare function, must return true on match |
586 | * @arg: argument passed on to compare function | |
587 | * | |
588 | * Traverses the bucket chain behind the provided hash value and calls the | |
589 | * specified compare function for each entry. | |
590 | * | |
97defe1e | 591 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
592 | * |
593 | * Returns the first entry on which the compare function returned true. | |
594 | */ | |
97defe1e | 595 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
596 | bool (*compare)(void *, void *), void *arg) |
597 | { | |
97defe1e | 598 | const struct bucket_table *tbl, *old_tbl; |
7e1e7763 | 599 | struct rhash_head *he; |
8d24c0b4 | 600 | u32 hash; |
7e1e7763 | 601 | |
97defe1e TG |
602 | rcu_read_lock(); |
603 | ||
aa34a6cb | 604 | tbl = rht_dereference_rcu(ht->tbl, ht); |
97defe1e | 605 | restart: |
39361947 | 606 | hash = key_hashfn(ht, tbl, key); |
8d2b1879 | 607 | rht_for_each_rcu(he, tbl, hash) { |
7e1e7763 TG |
608 | if (!compare(rht_obj(ht, he), arg)) |
609 | continue; | |
97defe1e | 610 | rcu_read_unlock(); |
a4b18cda | 611 | return rht_obj(ht, he); |
7e1e7763 TG |
612 | } |
613 | ||
9497df88 HX |
614 | /* Ensure we see any new tables. */ |
615 | smp_rmb(); | |
616 | ||
aa34a6cb HX |
617 | old_tbl = tbl; |
618 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
619 | if (unlikely(tbl != old_tbl)) | |
97defe1e | 620 | goto restart; |
97defe1e TG |
621 | rcu_read_unlock(); |
622 | ||
7e1e7763 TG |
623 | return NULL; |
624 | } | |
625 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
626 | ||
db304854 YX |
627 | /** |
628 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
629 | * @ht: hash table | |
630 | * @obj: pointer to hash head inside object | |
631 | * | |
632 | * Locks down the bucket chain in both the old and new table if a resize | |
633 | * is in progress to ensure that writers can't remove from the old table | |
634 | * and can't insert to the new table during the atomic operation of search | |
635 | * and insertion. Searches for duplicates in both the old and new table if | |
636 | * a resize is in progress. | |
637 | * | |
638 | * This lookup function may only be used for fixed key hash table (key_len | |
639 | * parameter set). It will BUG() if used inappropriately. | |
640 | * | |
641 | * It is safe to call this function from atomic context. | |
642 | * | |
643 | * Will trigger an automatic deferred table resizing if the size grows | |
644 | * beyond the watermark indicated by grow_decision() which can be passed | |
645 | * to rhashtable_init(). | |
646 | */ | |
647 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
7a868d1e YX |
648 | { |
649 | struct rhashtable_compare_arg arg = { | |
650 | .ht = ht, | |
651 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
652 | }; | |
653 | ||
654 | BUG_ON(!ht->p.key_len); | |
655 | ||
656 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
657 | &arg); | |
658 | } | |
659 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
660 | ||
661 | /** | |
662 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
663 | * with compare function | |
664 | * @ht: hash table | |
665 | * @obj: pointer to hash head inside object | |
666 | * @compare: compare function, must return true on match | |
667 | * @arg: argument passed on to compare function | |
668 | * | |
669 | * Locks down the bucket chain in both the old and new table if a resize | |
670 | * is in progress to ensure that writers can't remove from the old table | |
671 | * and can't insert to the new table during the atomic operation of search | |
672 | * and insertion. Searches for duplicates in both the old and new table if | |
673 | * a resize is in progress. | |
674 | * | |
675 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
676 | * | |
677 | * Will trigger an automatic deferred table resizing if the size grows | |
678 | * beyond the watermark indicated by grow_decision() which can be passed | |
679 | * to rhashtable_init(). | |
680 | */ | |
681 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
682 | struct rhash_head *obj, | |
683 | bool (*compare)(void *, void *), | |
684 | void *arg) | |
db304854 | 685 | { |
db304854 YX |
686 | BUG_ON(!ht->p.key_len); |
687 | ||
aa34a6cb | 688 | return __rhashtable_insert(ht, obj, compare, arg); |
db304854 | 689 | } |
7a868d1e | 690 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); |
db304854 | 691 | |
f2dba9c6 HX |
692 | /** |
693 | * rhashtable_walk_init - Initialise an iterator | |
694 | * @ht: Table to walk over | |
695 | * @iter: Hash table Iterator | |
696 | * | |
697 | * This function prepares a hash table walk. | |
698 | * | |
699 | * Note that if you restart a walk after rhashtable_walk_stop you | |
700 | * may see the same object twice. Also, you may miss objects if | |
701 | * there are removals in between rhashtable_walk_stop and the next | |
702 | * call to rhashtable_walk_start. | |
703 | * | |
704 | * For a completely stable walk you should construct your own data | |
705 | * structure outside the hash table. | |
706 | * | |
707 | * This function may sleep so you must not call it from interrupt | |
708 | * context or with spin locks held. | |
709 | * | |
710 | * You must call rhashtable_walk_exit if this function returns | |
711 | * successfully. | |
712 | */ | |
713 | int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter) | |
714 | { | |
715 | iter->ht = ht; | |
716 | iter->p = NULL; | |
717 | iter->slot = 0; | |
718 | iter->skip = 0; | |
719 | ||
720 | iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL); | |
721 | if (!iter->walker) | |
722 | return -ENOMEM; | |
723 | ||
724 | mutex_lock(&ht->mutex); | |
eddee5ba HX |
725 | iter->walker->tbl = rht_dereference(ht->tbl, ht); |
726 | list_add(&iter->walker->list, &iter->walker->tbl->walkers); | |
f2dba9c6 HX |
727 | mutex_unlock(&ht->mutex); |
728 | ||
729 | return 0; | |
730 | } | |
731 | EXPORT_SYMBOL_GPL(rhashtable_walk_init); | |
732 | ||
733 | /** | |
734 | * rhashtable_walk_exit - Free an iterator | |
735 | * @iter: Hash table Iterator | |
736 | * | |
737 | * This function frees resources allocated by rhashtable_walk_init. | |
738 | */ | |
739 | void rhashtable_walk_exit(struct rhashtable_iter *iter) | |
740 | { | |
741 | mutex_lock(&iter->ht->mutex); | |
eddee5ba HX |
742 | if (iter->walker->tbl) |
743 | list_del(&iter->walker->list); | |
f2dba9c6 HX |
744 | mutex_unlock(&iter->ht->mutex); |
745 | kfree(iter->walker); | |
746 | } | |
747 | EXPORT_SYMBOL_GPL(rhashtable_walk_exit); | |
748 | ||
749 | /** | |
750 | * rhashtable_walk_start - Start a hash table walk | |
751 | * @iter: Hash table iterator | |
752 | * | |
753 | * Start a hash table walk. Note that we take the RCU lock in all | |
754 | * cases including when we return an error. So you must always call | |
755 | * rhashtable_walk_stop to clean up. | |
756 | * | |
757 | * Returns zero if successful. | |
758 | * | |
759 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
760 | * will rewind back to the beginning and you may use it immediately | |
761 | * by calling rhashtable_walk_next. | |
762 | */ | |
763 | int rhashtable_walk_start(struct rhashtable_iter *iter) | |
764 | { | |
eddee5ba HX |
765 | struct rhashtable *ht = iter->ht; |
766 | ||
767 | mutex_lock(&ht->mutex); | |
768 | ||
769 | if (iter->walker->tbl) | |
770 | list_del(&iter->walker->list); | |
771 | ||
f2dba9c6 HX |
772 | rcu_read_lock(); |
773 | ||
eddee5ba HX |
774 | mutex_unlock(&ht->mutex); |
775 | ||
776 | if (!iter->walker->tbl) { | |
777 | iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht); | |
f2dba9c6 HX |
778 | return -EAGAIN; |
779 | } | |
780 | ||
781 | return 0; | |
782 | } | |
783 | EXPORT_SYMBOL_GPL(rhashtable_walk_start); | |
784 | ||
785 | /** | |
786 | * rhashtable_walk_next - Return the next object and advance the iterator | |
787 | * @iter: Hash table iterator | |
788 | * | |
789 | * Note that you must call rhashtable_walk_stop when you are finished | |
790 | * with the walk. | |
791 | * | |
792 | * Returns the next object or NULL when the end of the table is reached. | |
793 | * | |
794 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
795 | * will rewind back to the beginning and you may continue to use it. | |
796 | */ | |
797 | void *rhashtable_walk_next(struct rhashtable_iter *iter) | |
798 | { | |
eddee5ba | 799 | struct bucket_table *tbl = iter->walker->tbl; |
f2dba9c6 HX |
800 | struct rhashtable *ht = iter->ht; |
801 | struct rhash_head *p = iter->p; | |
802 | void *obj = NULL; | |
803 | ||
f2dba9c6 HX |
804 | if (p) { |
805 | p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot); | |
806 | goto next; | |
807 | } | |
808 | ||
809 | for (; iter->slot < tbl->size; iter->slot++) { | |
810 | int skip = iter->skip; | |
811 | ||
812 | rht_for_each_rcu(p, tbl, iter->slot) { | |
813 | if (!skip) | |
814 | break; | |
815 | skip--; | |
816 | } | |
817 | ||
818 | next: | |
819 | if (!rht_is_a_nulls(p)) { | |
820 | iter->skip++; | |
821 | iter->p = p; | |
822 | obj = rht_obj(ht, p); | |
823 | goto out; | |
824 | } | |
825 | ||
826 | iter->skip = 0; | |
827 | } | |
828 | ||
eddee5ba HX |
829 | iter->walker->tbl = rht_dereference_rcu(ht->future_tbl, ht); |
830 | if (iter->walker->tbl != tbl) { | |
f2dba9c6 HX |
831 | iter->slot = 0; |
832 | iter->skip = 0; | |
f2dba9c6 HX |
833 | return ERR_PTR(-EAGAIN); |
834 | } | |
835 | ||
eddee5ba HX |
836 | iter->walker->tbl = NULL; |
837 | iter->p = NULL; | |
838 | ||
839 | out: | |
840 | ||
f2dba9c6 HX |
841 | return obj; |
842 | } | |
843 | EXPORT_SYMBOL_GPL(rhashtable_walk_next); | |
844 | ||
845 | /** | |
846 | * rhashtable_walk_stop - Finish a hash table walk | |
847 | * @iter: Hash table iterator | |
848 | * | |
849 | * Finish a hash table walk. | |
850 | */ | |
851 | void rhashtable_walk_stop(struct rhashtable_iter *iter) | |
852 | { | |
eddee5ba HX |
853 | struct rhashtable *ht; |
854 | struct bucket_table *tbl = iter->walker->tbl; | |
855 | ||
f2dba9c6 | 856 | rcu_read_unlock(); |
eddee5ba HX |
857 | |
858 | if (!tbl) | |
859 | return; | |
860 | ||
861 | ht = iter->ht; | |
862 | ||
863 | mutex_lock(&ht->mutex); | |
864 | if (rht_dereference(ht->tbl, ht) == tbl || | |
865 | rht_dereference(ht->future_tbl, ht) == tbl) | |
866 | list_add(&iter->walker->list, &tbl->walkers); | |
867 | else | |
868 | iter->walker->tbl = NULL; | |
869 | mutex_unlock(&ht->mutex); | |
870 | ||
f2dba9c6 HX |
871 | iter->p = NULL; |
872 | } | |
873 | EXPORT_SYMBOL_GPL(rhashtable_walk_stop); | |
874 | ||
94000176 | 875 | static size_t rounded_hashtable_size(struct rhashtable_params *params) |
7e1e7763 | 876 | { |
94000176 YX |
877 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), |
878 | 1UL << params->min_shift); | |
7e1e7763 TG |
879 | } |
880 | ||
881 | /** | |
882 | * rhashtable_init - initialize a new hash table | |
883 | * @ht: hash table to be initialized | |
884 | * @params: configuration parameters | |
885 | * | |
886 | * Initializes a new hash table based on the provided configuration | |
887 | * parameters. A table can be configured either with a variable or | |
888 | * fixed length key: | |
889 | * | |
890 | * Configuration Example 1: Fixed length keys | |
891 | * struct test_obj { | |
892 | * int key; | |
893 | * void * my_member; | |
894 | * struct rhash_head node; | |
895 | * }; | |
896 | * | |
897 | * struct rhashtable_params params = { | |
898 | * .head_offset = offsetof(struct test_obj, node), | |
899 | * .key_offset = offsetof(struct test_obj, key), | |
900 | * .key_len = sizeof(int), | |
87545899 | 901 | * .hashfn = jhash, |
f89bd6f8 | 902 | * .nulls_base = (1U << RHT_BASE_SHIFT), |
7e1e7763 TG |
903 | * }; |
904 | * | |
905 | * Configuration Example 2: Variable length keys | |
906 | * struct test_obj { | |
907 | * [...] | |
908 | * struct rhash_head node; | |
909 | * }; | |
910 | * | |
911 | * u32 my_hash_fn(const void *data, u32 seed) | |
912 | * { | |
913 | * struct test_obj *obj = data; | |
914 | * | |
915 | * return [... hash ...]; | |
916 | * } | |
917 | * | |
918 | * struct rhashtable_params params = { | |
919 | * .head_offset = offsetof(struct test_obj, node), | |
87545899 | 920 | * .hashfn = jhash, |
7e1e7763 | 921 | * .obj_hashfn = my_hash_fn, |
7e1e7763 TG |
922 | * }; |
923 | */ | |
924 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
925 | { | |
926 | struct bucket_table *tbl; | |
927 | size_t size; | |
928 | ||
929 | size = HASH_DEFAULT_SIZE; | |
930 | ||
931 | if ((params->key_len && !params->hashfn) || | |
932 | (!params->key_len && !params->obj_hashfn)) | |
933 | return -EINVAL; | |
934 | ||
f89bd6f8 TG |
935 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
936 | return -EINVAL; | |
937 | ||
94000176 YX |
938 | params->min_shift = max_t(size_t, params->min_shift, |
939 | ilog2(HASH_MIN_SIZE)); | |
940 | ||
7e1e7763 | 941 | if (params->nelem_hint) |
94000176 | 942 | size = rounded_hashtable_size(params); |
7e1e7763 | 943 | |
97defe1e TG |
944 | memset(ht, 0, sizeof(*ht)); |
945 | mutex_init(&ht->mutex); | |
946 | memcpy(&ht->p, params, sizeof(*params)); | |
947 | ||
948 | if (params->locks_mul) | |
949 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
950 | else | |
951 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
952 | ||
5269b53d | 953 | tbl = bucket_table_alloc(ht, size); |
7e1e7763 TG |
954 | if (tbl == NULL) |
955 | return -ENOMEM; | |
956 | ||
545a148e | 957 | atomic_set(&ht->nelems, 0); |
a5b6846f | 958 | |
7e1e7763 | 959 | RCU_INIT_POINTER(ht->tbl, tbl); |
97defe1e | 960 | RCU_INIT_POINTER(ht->future_tbl, tbl); |
7e1e7763 | 961 | |
4c4b52d9 | 962 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 963 | |
7e1e7763 TG |
964 | return 0; |
965 | } | |
966 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
967 | ||
968 | /** | |
969 | * rhashtable_destroy - destroy hash table | |
970 | * @ht: the hash table to destroy | |
971 | * | |
ae82ddcf PNA |
972 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
973 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
974 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 975 | */ |
97defe1e | 976 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 977 | { |
97defe1e TG |
978 | ht->being_destroyed = true; |
979 | ||
4c4b52d9 | 980 | cancel_work_sync(&ht->run_work); |
97defe1e | 981 | |
57699a40 | 982 | mutex_lock(&ht->mutex); |
97defe1e | 983 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 984 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
985 | } |
986 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |