]>
Commit | Line | Data |
---|---|---|
7e1e7763 TG |
1 | /* |
2 | * Resizable, Scalable, Concurrent Hash Table | |
3 | * | |
a5ec68e3 | 4 | * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch> |
7e1e7763 TG |
5 | * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net> |
6 | * | |
7 | * Based on the following paper: | |
8 | * https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf | |
9 | * | |
10 | * Code partially derived from nft_hash | |
11 | * | |
12 | * This program is free software; you can redistribute it and/or modify | |
13 | * it under the terms of the GNU General Public License version 2 as | |
14 | * published by the Free Software Foundation. | |
15 | */ | |
16 | ||
17 | #include <linux/kernel.h> | |
18 | #include <linux/init.h> | |
19 | #include <linux/log2.h> | |
5beb5c90 | 20 | #include <linux/sched.h> |
7e1e7763 TG |
21 | #include <linux/slab.h> |
22 | #include <linux/vmalloc.h> | |
23 | #include <linux/mm.h> | |
87545899 | 24 | #include <linux/jhash.h> |
7e1e7763 TG |
25 | #include <linux/random.h> |
26 | #include <linux/rhashtable.h> | |
61d7b097 | 27 | #include <linux/err.h> |
7e1e7763 TG |
28 | |
29 | #define HASH_DEFAULT_SIZE 64UL | |
30 | #define HASH_MIN_SIZE 4UL | |
97defe1e TG |
31 | #define BUCKET_LOCKS_PER_CPU 128UL |
32 | ||
f89bd6f8 TG |
33 | /* Base bits plus 1 bit for nulls marker */ |
34 | #define HASH_RESERVED_SPACE (RHT_BASE_BITS + 1) | |
35 | ||
97defe1e TG |
36 | enum { |
37 | RHT_LOCK_NORMAL, | |
38 | RHT_LOCK_NESTED, | |
97defe1e TG |
39 | }; |
40 | ||
41 | /* The bucket lock is selected based on the hash and protects mutations | |
42 | * on a group of hash buckets. | |
43 | * | |
a5ec68e3 TG |
44 | * A maximum of tbl->size/2 bucket locks is allocated. This ensures that |
45 | * a single lock always covers both buckets which may both contains | |
46 | * entries which link to the same bucket of the old table during resizing. | |
47 | * This allows to simplify the locking as locking the bucket in both | |
48 | * tables during resize always guarantee protection. | |
49 | * | |
97defe1e TG |
50 | * IMPORTANT: When holding the bucket lock of both the old and new table |
51 | * during expansions and shrinking, the old bucket lock must always be | |
52 | * acquired first. | |
53 | */ | |
54 | static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash) | |
55 | { | |
56 | return &tbl->locks[hash & tbl->locks_mask]; | |
57 | } | |
7e1e7763 | 58 | |
c91eee56 | 59 | static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he) |
7e1e7763 TG |
60 | { |
61 | return (void *) he - ht->p.head_offset; | |
62 | } | |
7e1e7763 | 63 | |
8d24c0b4 | 64 | static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash) |
7e1e7763 | 65 | { |
ec9f71c5 | 66 | return (hash >> HASH_RESERVED_SPACE) & (tbl->size - 1); |
7e1e7763 TG |
67 | } |
68 | ||
aa34a6cb | 69 | static u32 key_hashfn(struct rhashtable *ht, const struct bucket_table *tbl, |
cffaa9cb | 70 | const void *key) |
7e1e7763 | 71 | { |
cffaa9cb | 72 | return rht_bucket_index(tbl, ht->p.hashfn(key, ht->p.key_len, |
ec9f71c5 | 73 | tbl->hash_rnd)); |
7e1e7763 | 74 | } |
7e1e7763 | 75 | |
988dfbd7 | 76 | static u32 head_hashfn(struct rhashtable *ht, |
8d24c0b4 TG |
77 | const struct bucket_table *tbl, |
78 | const struct rhash_head *he) | |
7e1e7763 | 79 | { |
ec9f71c5 HX |
80 | const char *ptr = rht_obj(ht, he); |
81 | ||
82 | return likely(ht->p.key_len) ? | |
83 | key_hashfn(ht, tbl, ptr + ht->p.key_offset) : | |
84 | rht_bucket_index(tbl, ht->p.obj_hashfn(ptr, tbl->hash_rnd)); | |
7e1e7763 TG |
85 | } |
86 | ||
a03eaec0 | 87 | #ifdef CONFIG_PROVE_LOCKING |
a03eaec0 | 88 | #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT)) |
a03eaec0 TG |
89 | |
90 | int lockdep_rht_mutex_is_held(struct rhashtable *ht) | |
91 | { | |
92 | return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1; | |
93 | } | |
94 | EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held); | |
95 | ||
96 | int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash) | |
97 | { | |
98 | spinlock_t *lock = bucket_lock(tbl, hash); | |
99 | ||
100 | return (debug_locks) ? lockdep_is_held(lock) : 1; | |
101 | } | |
102 | EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held); | |
103 | #else | |
104 | #define ASSERT_RHT_MUTEX(HT) | |
a03eaec0 TG |
105 | #endif |
106 | ||
107 | ||
97defe1e TG |
108 | static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl) |
109 | { | |
110 | unsigned int i, size; | |
111 | #if defined(CONFIG_PROVE_LOCKING) | |
112 | unsigned int nr_pcpus = 2; | |
113 | #else | |
114 | unsigned int nr_pcpus = num_possible_cpus(); | |
115 | #endif | |
116 | ||
117 | nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL); | |
118 | size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul); | |
119 | ||
a5ec68e3 TG |
120 | /* Never allocate more than 0.5 locks per bucket */ |
121 | size = min_t(unsigned int, size, tbl->size >> 1); | |
97defe1e TG |
122 | |
123 | if (sizeof(spinlock_t) != 0) { | |
124 | #ifdef CONFIG_NUMA | |
125 | if (size * sizeof(spinlock_t) > PAGE_SIZE) | |
126 | tbl->locks = vmalloc(size * sizeof(spinlock_t)); | |
127 | else | |
128 | #endif | |
129 | tbl->locks = kmalloc_array(size, sizeof(spinlock_t), | |
130 | GFP_KERNEL); | |
131 | if (!tbl->locks) | |
132 | return -ENOMEM; | |
133 | for (i = 0; i < size; i++) | |
134 | spin_lock_init(&tbl->locks[i]); | |
135 | } | |
136 | tbl->locks_mask = size - 1; | |
137 | ||
138 | return 0; | |
139 | } | |
140 | ||
141 | static void bucket_table_free(const struct bucket_table *tbl) | |
142 | { | |
143 | if (tbl) | |
144 | kvfree(tbl->locks); | |
145 | ||
146 | kvfree(tbl); | |
147 | } | |
148 | ||
149 | static struct bucket_table *bucket_table_alloc(struct rhashtable *ht, | |
a5b6846f | 150 | size_t nbuckets, u32 hash_rnd) |
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 DB |
165 | tbl->shift = ilog2(nbuckets); |
166 | tbl->hash_rnd = hash_rnd; | |
7e1e7763 | 167 | |
97defe1e TG |
168 | if (alloc_bucket_locks(ht, tbl) < 0) { |
169 | bucket_table_free(tbl); | |
170 | return NULL; | |
171 | } | |
7e1e7763 | 172 | |
f89bd6f8 TG |
173 | for (i = 0; i < nbuckets; i++) |
174 | INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i); | |
175 | ||
97defe1e | 176 | return tbl; |
7e1e7763 TG |
177 | } |
178 | ||
179 | /** | |
180 | * rht_grow_above_75 - returns true if nelems > 0.75 * table-size | |
181 | * @ht: hash table | |
a5b6846f | 182 | * @tbl: current table |
7e1e7763 | 183 | */ |
a5b6846f DB |
184 | static bool rht_grow_above_75(const struct rhashtable *ht, |
185 | const struct bucket_table *tbl) | |
7e1e7763 TG |
186 | { |
187 | /* Expand table when exceeding 75% load */ | |
a5b6846f DB |
188 | return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) && |
189 | (!ht->p.max_shift || tbl->shift < ht->p.max_shift); | |
7e1e7763 | 190 | } |
7e1e7763 TG |
191 | |
192 | /** | |
193 | * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size | |
194 | * @ht: hash table | |
a5b6846f | 195 | * @tbl: current table |
7e1e7763 | 196 | */ |
a5b6846f DB |
197 | static bool rht_shrink_below_30(const struct rhashtable *ht, |
198 | const struct bucket_table *tbl) | |
7e1e7763 TG |
199 | { |
200 | /* Shrink table beneath 30% load */ | |
a5b6846f DB |
201 | return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) && |
202 | tbl->shift > ht->p.min_shift; | |
7e1e7763 | 203 | } |
7e1e7763 | 204 | |
aa34a6cb | 205 | static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash) |
a5ec68e3 | 206 | { |
aa34a6cb HX |
207 | struct bucket_table *new_tbl = rht_dereference(ht->future_tbl, ht); |
208 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
209 | struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash]; | |
210 | int err = -ENOENT; | |
211 | struct rhash_head *head, *next, *entry; | |
212 | spinlock_t *new_bucket_lock; | |
213 | unsigned new_hash; | |
214 | ||
215 | rht_for_each(entry, old_tbl, old_hash) { | |
216 | err = 0; | |
217 | next = rht_dereference_bucket(entry->next, old_tbl, old_hash); | |
218 | ||
219 | if (rht_is_a_nulls(next)) | |
220 | break; | |
a5ec68e3 | 221 | |
aa34a6cb HX |
222 | pprev = &entry->next; |
223 | } | |
a5ec68e3 | 224 | |
aa34a6cb HX |
225 | if (err) |
226 | goto out; | |
97defe1e | 227 | |
aa34a6cb | 228 | new_hash = head_hashfn(ht, new_tbl, entry); |
7e1e7763 | 229 | |
aa34a6cb | 230 | new_bucket_lock = bucket_lock(new_tbl, new_hash); |
7e1e7763 | 231 | |
84ed82b7 | 232 | spin_lock_nested(new_bucket_lock, RHT_LOCK_NESTED); |
aa34a6cb HX |
233 | head = rht_dereference_bucket(new_tbl->buckets[new_hash], |
234 | new_tbl, new_hash); | |
97defe1e | 235 | |
aa34a6cb HX |
236 | if (rht_is_a_nulls(head)) |
237 | INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash); | |
238 | else | |
239 | RCU_INIT_POINTER(entry->next, head); | |
a5ec68e3 | 240 | |
aa34a6cb HX |
241 | rcu_assign_pointer(new_tbl->buckets[new_hash], entry); |
242 | spin_unlock(new_bucket_lock); | |
97defe1e | 243 | |
aa34a6cb | 244 | rcu_assign_pointer(*pprev, next); |
7e1e7763 | 245 | |
aa34a6cb HX |
246 | out: |
247 | return err; | |
248 | } | |
97defe1e | 249 | |
aa34a6cb HX |
250 | static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash) |
251 | { | |
252 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
253 | spinlock_t *old_bucket_lock; | |
254 | ||
255 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
a5ec68e3 | 256 | |
aa34a6cb HX |
257 | spin_lock_bh(old_bucket_lock); |
258 | while (!rhashtable_rehash_one(ht, old_hash)) | |
259 | ; | |
260 | spin_unlock_bh(old_bucket_lock); | |
97defe1e TG |
261 | } |
262 | ||
aa34a6cb HX |
263 | static void rhashtable_rehash(struct rhashtable *ht, |
264 | struct bucket_table *new_tbl) | |
97defe1e | 265 | { |
aa34a6cb HX |
266 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
267 | unsigned old_hash; | |
7cd10db8 | 268 | |
aa34a6cb HX |
269 | get_random_bytes(&new_tbl->hash_rnd, sizeof(new_tbl->hash_rnd)); |
270 | ||
271 | /* Make insertions go into the new, empty table right away. Deletions | |
272 | * and lookups will be attempted in both tables until we synchronize. | |
273 | * The synchronize_rcu() guarantees for the new table to be picked up | |
274 | * so no new additions go into the old table while we relink. | |
275 | */ | |
276 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
277 | ||
9497df88 HX |
278 | /* Ensure the new table is visible to readers. */ |
279 | smp_wmb(); | |
280 | ||
aa34a6cb HX |
281 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) |
282 | rhashtable_rehash_chain(ht, old_hash); | |
283 | ||
284 | /* Publish the new table pointer. */ | |
285 | rcu_assign_pointer(ht->tbl, new_tbl); | |
286 | ||
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 | ||
a5b6846f | 317 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, old_tbl->hash_rnd); |
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 | ||
a5b6846f | 348 | new_tbl = bucket_table_alloc(ht, old_tbl->size / 2, old_tbl->hash_rnd); |
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; | |
f2dba9c6 | 361 | struct rhashtable_walker *walker; |
97defe1e | 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 | ||
f2dba9c6 HX |
370 | list_for_each_entry(walker, &ht->walkers, list) |
371 | walker->resize = true; | |
372 | ||
a5b6846f | 373 | if (rht_grow_above_75(ht, tbl)) |
97defe1e | 374 | rhashtable_expand(ht); |
a5b6846f | 375 | else if (rht_shrink_below_30(ht, tbl)) |
97defe1e | 376 | rhashtable_shrink(ht); |
28134a53 | 377 | unlock: |
97defe1e TG |
378 | mutex_unlock(&ht->mutex); |
379 | } | |
380 | ||
aa34a6cb HX |
381 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
382 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 383 | { |
aa34a6cb | 384 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 385 | struct rhash_head *head; |
aa34a6cb HX |
386 | bool no_resize_running; |
387 | unsigned hash; | |
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); |
aa34a6cb HX |
394 | |
395 | spin_lock_bh(bucket_lock(old_tbl, hash)); | |
396 | ||
397 | /* Because we have already taken the bucket lock in old_tbl, | |
398 | * if we find that future_tbl is not yet visible then that | |
399 | * guarantees all other insertions of the same entry will | |
400 | * also grab the bucket lock in old_tbl because until the | |
401 | * rehash completes ht->tbl won't be changed. | |
402 | */ | |
403 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
404 | if (tbl != old_tbl) { | |
eca84933 | 405 | hash = head_hashfn(ht, tbl, obj); |
84ed82b7 | 406 | spin_lock_nested(bucket_lock(tbl, hash), RHT_LOCK_NESTED); |
aa34a6cb HX |
407 | } |
408 | ||
409 | if (compare && | |
410 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
411 | compare, arg)) { | |
412 | success = false; | |
413 | goto exit; | |
414 | } | |
415 | ||
416 | no_resize_running = tbl == old_tbl; | |
020219a6 | 417 | |
020219a6 | 418 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); |
db304854 YX |
419 | |
420 | if (rht_is_a_nulls(head)) | |
421 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
422 | else | |
423 | RCU_INIT_POINTER(obj->next, head); | |
424 | ||
425 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
426 | ||
427 | atomic_inc(&ht->nelems); | |
a5b6846f | 428 | if (no_resize_running && rht_grow_above_75(ht, tbl)) |
4c4b52d9 | 429 | schedule_work(&ht->run_work); |
aa34a6cb HX |
430 | |
431 | exit: | |
432 | if (tbl != old_tbl) { | |
eca84933 | 433 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb HX |
434 | spin_unlock(bucket_lock(tbl, hash)); |
435 | } | |
436 | ||
eca84933 | 437 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
438 | spin_unlock_bh(bucket_lock(old_tbl, hash)); |
439 | ||
440 | rcu_read_unlock(); | |
441 | ||
442 | return success; | |
db304854 YX |
443 | } |
444 | ||
7e1e7763 | 445 | /** |
db304854 | 446 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
447 | * @ht: hash table |
448 | * @obj: pointer to hash head inside object | |
7e1e7763 | 449 | * |
97defe1e TG |
450 | * Will take a per bucket spinlock to protect against mutual mutations |
451 | * on the same bucket. Multiple insertions may occur in parallel unless | |
452 | * they map to the same bucket lock. | |
7e1e7763 | 453 | * |
97defe1e TG |
454 | * It is safe to call this function from atomic context. |
455 | * | |
456 | * Will trigger an automatic deferred table resizing if the size grows | |
457 | * beyond the watermark indicated by grow_decision() which can be passed | |
458 | * to rhashtable_init(). | |
7e1e7763 | 459 | */ |
6eba8224 | 460 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 461 | { |
aa34a6cb HX |
462 | __rhashtable_insert(ht, obj, NULL, NULL); |
463 | } | |
464 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
465 | ||
466 | static bool __rhashtable_remove(struct rhashtable *ht, | |
467 | struct bucket_table *tbl, | |
468 | struct rhash_head *obj) | |
469 | { | |
470 | struct rhash_head __rcu **pprev; | |
471 | struct rhash_head *he; | |
472 | spinlock_t * lock; | |
97defe1e | 473 | unsigned hash; |
aa34a6cb | 474 | bool ret = false; |
7e1e7763 | 475 | |
eca84933 | 476 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb | 477 | lock = bucket_lock(tbl, hash); |
7e1e7763 | 478 | |
aa34a6cb | 479 | spin_lock_bh(lock); |
97defe1e | 480 | |
aa34a6cb HX |
481 | pprev = &tbl->buckets[hash]; |
482 | rht_for_each(he, tbl, hash) { | |
483 | if (he != obj) { | |
484 | pprev = &he->next; | |
485 | continue; | |
486 | } | |
7e1e7763 | 487 | |
aa34a6cb HX |
488 | rcu_assign_pointer(*pprev, obj->next); |
489 | ret = true; | |
490 | break; | |
491 | } | |
492 | ||
493 | spin_unlock_bh(lock); | |
494 | ||
495 | return ret; | |
7e1e7763 | 496 | } |
7e1e7763 | 497 | |
7e1e7763 TG |
498 | /** |
499 | * rhashtable_remove - remove object from hash table | |
500 | * @ht: hash table | |
501 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
502 | * |
503 | * Since the hash chain is single linked, the removal operation needs to | |
504 | * walk the bucket chain upon removal. The removal operation is thus | |
505 | * considerable slow if the hash table is not correctly sized. | |
506 | * | |
db304854 | 507 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
508 | * shrink_decision function specified at rhashtable_init() returns true. |
509 | * | |
510 | * The caller must ensure that no concurrent table mutations occur. It is | |
511 | * however valid to have concurrent lookups if they are RCU protected. | |
512 | */ | |
6eba8224 | 513 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 514 | { |
aa34a6cb HX |
515 | struct bucket_table *tbl, *old_tbl; |
516 | bool ret; | |
7e1e7763 | 517 | |
97defe1e | 518 | rcu_read_lock(); |
7e1e7763 | 519 | |
aa34a6cb HX |
520 | old_tbl = rht_dereference_rcu(ht->tbl, ht); |
521 | ret = __rhashtable_remove(ht, old_tbl, obj); | |
7e1e7763 | 522 | |
aa34a6cb HX |
523 | /* Because we have already taken (and released) the bucket |
524 | * lock in old_tbl, if we find that future_tbl is not yet | |
525 | * visible then that guarantees the entry to still be in | |
526 | * old_tbl if it exists. | |
fe6a043c | 527 | */ |
aa34a6cb HX |
528 | tbl = rht_dereference_rcu(ht->future_tbl, ht); |
529 | if (!ret && old_tbl != tbl) | |
530 | ret = __rhashtable_remove(ht, tbl, obj); | |
fe6a043c TG |
531 | |
532 | if (ret) { | |
aa34a6cb | 533 | bool no_resize_running = tbl == old_tbl; |
4c4b52d9 | 534 | |
fe6a043c | 535 | atomic_dec(&ht->nelems); |
a5b6846f | 536 | if (no_resize_running && rht_shrink_below_30(ht, tbl)) |
4c4b52d9 | 537 | schedule_work(&ht->run_work); |
fe6a043c TG |
538 | } |
539 | ||
97defe1e TG |
540 | rcu_read_unlock(); |
541 | ||
fe6a043c | 542 | return ret; |
7e1e7763 TG |
543 | } |
544 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
545 | ||
efb975a6 YX |
546 | struct rhashtable_compare_arg { |
547 | struct rhashtable *ht; | |
548 | const void *key; | |
549 | }; | |
550 | ||
551 | static bool rhashtable_compare(void *ptr, void *arg) | |
552 | { | |
553 | struct rhashtable_compare_arg *x = arg; | |
554 | struct rhashtable *ht = x->ht; | |
555 | ||
556 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
557 | } | |
558 | ||
7e1e7763 TG |
559 | /** |
560 | * rhashtable_lookup - lookup key in hash table | |
561 | * @ht: hash table | |
562 | * @key: pointer to key | |
563 | * | |
564 | * Computes the hash value for the key and traverses the bucket chain looking | |
565 | * for a entry with an identical key. The first matching entry is returned. | |
566 | * | |
567 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 568 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 569 | * |
97defe1e | 570 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 571 | */ |
97defe1e | 572 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 573 | { |
efb975a6 YX |
574 | struct rhashtable_compare_arg arg = { |
575 | .ht = ht, | |
576 | .key = key, | |
577 | }; | |
7e1e7763 TG |
578 | |
579 | BUG_ON(!ht->p.key_len); | |
580 | ||
efb975a6 | 581 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
582 | } |
583 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
584 | ||
585 | /** | |
586 | * rhashtable_lookup_compare - search hash table with compare function | |
587 | * @ht: hash table | |
8d24c0b4 | 588 | * @key: the pointer to the key |
7e1e7763 TG |
589 | * @compare: compare function, must return true on match |
590 | * @arg: argument passed on to compare function | |
591 | * | |
592 | * Traverses the bucket chain behind the provided hash value and calls the | |
593 | * specified compare function for each entry. | |
594 | * | |
97defe1e | 595 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
596 | * |
597 | * Returns the first entry on which the compare function returned true. | |
598 | */ | |
97defe1e | 599 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
600 | bool (*compare)(void *, void *), void *arg) |
601 | { | |
97defe1e | 602 | const struct bucket_table *tbl, *old_tbl; |
7e1e7763 | 603 | struct rhash_head *he; |
8d24c0b4 | 604 | u32 hash; |
7e1e7763 | 605 | |
97defe1e TG |
606 | rcu_read_lock(); |
607 | ||
aa34a6cb | 608 | tbl = rht_dereference_rcu(ht->tbl, ht); |
97defe1e | 609 | restart: |
39361947 | 610 | hash = key_hashfn(ht, tbl, key); |
8d2b1879 | 611 | rht_for_each_rcu(he, tbl, hash) { |
7e1e7763 TG |
612 | if (!compare(rht_obj(ht, he), arg)) |
613 | continue; | |
97defe1e | 614 | rcu_read_unlock(); |
a4b18cda | 615 | return rht_obj(ht, he); |
7e1e7763 TG |
616 | } |
617 | ||
9497df88 HX |
618 | /* Ensure we see any new tables. */ |
619 | smp_rmb(); | |
620 | ||
aa34a6cb HX |
621 | old_tbl = tbl; |
622 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
623 | if (unlikely(tbl != old_tbl)) | |
97defe1e | 624 | goto restart; |
97defe1e TG |
625 | rcu_read_unlock(); |
626 | ||
7e1e7763 TG |
627 | return NULL; |
628 | } | |
629 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
630 | ||
db304854 YX |
631 | /** |
632 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
633 | * @ht: hash table | |
634 | * @obj: pointer to hash head inside object | |
635 | * | |
636 | * Locks down the bucket chain in both the old and new table if a resize | |
637 | * is in progress to ensure that writers can't remove from the old table | |
638 | * and can't insert to the new table during the atomic operation of search | |
639 | * and insertion. Searches for duplicates in both the old and new table if | |
640 | * a resize is in progress. | |
641 | * | |
642 | * This lookup function may only be used for fixed key hash table (key_len | |
643 | * parameter set). It will BUG() if used inappropriately. | |
644 | * | |
645 | * It is safe to call this function from atomic context. | |
646 | * | |
647 | * Will trigger an automatic deferred table resizing if the size grows | |
648 | * beyond the watermark indicated by grow_decision() which can be passed | |
649 | * to rhashtable_init(). | |
650 | */ | |
651 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
7a868d1e YX |
652 | { |
653 | struct rhashtable_compare_arg arg = { | |
654 | .ht = ht, | |
655 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
656 | }; | |
657 | ||
658 | BUG_ON(!ht->p.key_len); | |
659 | ||
660 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
661 | &arg); | |
662 | } | |
663 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
664 | ||
665 | /** | |
666 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
667 | * with compare function | |
668 | * @ht: hash table | |
669 | * @obj: pointer to hash head inside object | |
670 | * @compare: compare function, must return true on match | |
671 | * @arg: argument passed on to compare function | |
672 | * | |
673 | * Locks down the bucket chain in both the old and new table if a resize | |
674 | * is in progress to ensure that writers can't remove from the old table | |
675 | * and can't insert to the new table during the atomic operation of search | |
676 | * and insertion. Searches for duplicates in both the old and new table if | |
677 | * a resize is in progress. | |
678 | * | |
679 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
680 | * | |
681 | * Will trigger an automatic deferred table resizing if the size grows | |
682 | * beyond the watermark indicated by grow_decision() which can be passed | |
683 | * to rhashtable_init(). | |
684 | */ | |
685 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
686 | struct rhash_head *obj, | |
687 | bool (*compare)(void *, void *), | |
688 | void *arg) | |
db304854 | 689 | { |
db304854 YX |
690 | BUG_ON(!ht->p.key_len); |
691 | ||
aa34a6cb | 692 | return __rhashtable_insert(ht, obj, compare, arg); |
db304854 | 693 | } |
7a868d1e | 694 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); |
db304854 | 695 | |
f2dba9c6 HX |
696 | /** |
697 | * rhashtable_walk_init - Initialise an iterator | |
698 | * @ht: Table to walk over | |
699 | * @iter: Hash table Iterator | |
700 | * | |
701 | * This function prepares a hash table walk. | |
702 | * | |
703 | * Note that if you restart a walk after rhashtable_walk_stop you | |
704 | * may see the same object twice. Also, you may miss objects if | |
705 | * there are removals in between rhashtable_walk_stop and the next | |
706 | * call to rhashtable_walk_start. | |
707 | * | |
708 | * For a completely stable walk you should construct your own data | |
709 | * structure outside the hash table. | |
710 | * | |
711 | * This function may sleep so you must not call it from interrupt | |
712 | * context or with spin locks held. | |
713 | * | |
714 | * You must call rhashtable_walk_exit if this function returns | |
715 | * successfully. | |
716 | */ | |
717 | int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter) | |
718 | { | |
719 | iter->ht = ht; | |
720 | iter->p = NULL; | |
721 | iter->slot = 0; | |
722 | iter->skip = 0; | |
723 | ||
724 | iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL); | |
725 | if (!iter->walker) | |
726 | return -ENOMEM; | |
727 | ||
71bb0012 SL |
728 | INIT_LIST_HEAD(&iter->walker->list); |
729 | iter->walker->resize = false; | |
730 | ||
f2dba9c6 HX |
731 | mutex_lock(&ht->mutex); |
732 | list_add(&iter->walker->list, &ht->walkers); | |
733 | mutex_unlock(&ht->mutex); | |
734 | ||
735 | return 0; | |
736 | } | |
737 | EXPORT_SYMBOL_GPL(rhashtable_walk_init); | |
738 | ||
739 | /** | |
740 | * rhashtable_walk_exit - Free an iterator | |
741 | * @iter: Hash table Iterator | |
742 | * | |
743 | * This function frees resources allocated by rhashtable_walk_init. | |
744 | */ | |
745 | void rhashtable_walk_exit(struct rhashtable_iter *iter) | |
746 | { | |
747 | mutex_lock(&iter->ht->mutex); | |
748 | list_del(&iter->walker->list); | |
749 | mutex_unlock(&iter->ht->mutex); | |
750 | kfree(iter->walker); | |
751 | } | |
752 | EXPORT_SYMBOL_GPL(rhashtable_walk_exit); | |
753 | ||
754 | /** | |
755 | * rhashtable_walk_start - Start a hash table walk | |
756 | * @iter: Hash table iterator | |
757 | * | |
758 | * Start a hash table walk. Note that we take the RCU lock in all | |
759 | * cases including when we return an error. So you must always call | |
760 | * rhashtable_walk_stop to clean up. | |
761 | * | |
762 | * Returns zero if successful. | |
763 | * | |
764 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
765 | * will rewind back to the beginning and you may use it immediately | |
766 | * by calling rhashtable_walk_next. | |
767 | */ | |
768 | int rhashtable_walk_start(struct rhashtable_iter *iter) | |
769 | { | |
770 | rcu_read_lock(); | |
771 | ||
772 | if (iter->walker->resize) { | |
773 | iter->slot = 0; | |
774 | iter->skip = 0; | |
775 | iter->walker->resize = false; | |
776 | return -EAGAIN; | |
777 | } | |
778 | ||
779 | return 0; | |
780 | } | |
781 | EXPORT_SYMBOL_GPL(rhashtable_walk_start); | |
782 | ||
783 | /** | |
784 | * rhashtable_walk_next - Return the next object and advance the iterator | |
785 | * @iter: Hash table iterator | |
786 | * | |
787 | * Note that you must call rhashtable_walk_stop when you are finished | |
788 | * with the walk. | |
789 | * | |
790 | * Returns the next object or NULL when the end of the table is reached. | |
791 | * | |
792 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
793 | * will rewind back to the beginning and you may continue to use it. | |
794 | */ | |
795 | void *rhashtable_walk_next(struct rhashtable_iter *iter) | |
796 | { | |
797 | const struct bucket_table *tbl; | |
798 | struct rhashtable *ht = iter->ht; | |
799 | struct rhash_head *p = iter->p; | |
800 | void *obj = NULL; | |
801 | ||
802 | tbl = rht_dereference_rcu(ht->tbl, ht); | |
803 | ||
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 | ||
829 | iter->p = NULL; | |
830 | ||
831 | out: | |
832 | if (iter->walker->resize) { | |
833 | iter->p = NULL; | |
834 | iter->slot = 0; | |
835 | iter->skip = 0; | |
836 | iter->walker->resize = false; | |
837 | return ERR_PTR(-EAGAIN); | |
838 | } | |
839 | ||
840 | return obj; | |
841 | } | |
842 | EXPORT_SYMBOL_GPL(rhashtable_walk_next); | |
843 | ||
844 | /** | |
845 | * rhashtable_walk_stop - Finish a hash table walk | |
846 | * @iter: Hash table iterator | |
847 | * | |
848 | * Finish a hash table walk. | |
849 | */ | |
850 | void rhashtable_walk_stop(struct rhashtable_iter *iter) | |
851 | { | |
852 | rcu_read_unlock(); | |
853 | iter->p = NULL; | |
854 | } | |
855 | EXPORT_SYMBOL_GPL(rhashtable_walk_stop); | |
856 | ||
94000176 | 857 | static size_t rounded_hashtable_size(struct rhashtable_params *params) |
7e1e7763 | 858 | { |
94000176 YX |
859 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), |
860 | 1UL << params->min_shift); | |
7e1e7763 TG |
861 | } |
862 | ||
863 | /** | |
864 | * rhashtable_init - initialize a new hash table | |
865 | * @ht: hash table to be initialized | |
866 | * @params: configuration parameters | |
867 | * | |
868 | * Initializes a new hash table based on the provided configuration | |
869 | * parameters. A table can be configured either with a variable or | |
870 | * fixed length key: | |
871 | * | |
872 | * Configuration Example 1: Fixed length keys | |
873 | * struct test_obj { | |
874 | * int key; | |
875 | * void * my_member; | |
876 | * struct rhash_head node; | |
877 | * }; | |
878 | * | |
879 | * struct rhashtable_params params = { | |
880 | * .head_offset = offsetof(struct test_obj, node), | |
881 | * .key_offset = offsetof(struct test_obj, key), | |
882 | * .key_len = sizeof(int), | |
87545899 | 883 | * .hashfn = jhash, |
f89bd6f8 | 884 | * .nulls_base = (1U << RHT_BASE_SHIFT), |
7e1e7763 TG |
885 | * }; |
886 | * | |
887 | * Configuration Example 2: Variable length keys | |
888 | * struct test_obj { | |
889 | * [...] | |
890 | * struct rhash_head node; | |
891 | * }; | |
892 | * | |
893 | * u32 my_hash_fn(const void *data, u32 seed) | |
894 | * { | |
895 | * struct test_obj *obj = data; | |
896 | * | |
897 | * return [... hash ...]; | |
898 | * } | |
899 | * | |
900 | * struct rhashtable_params params = { | |
901 | * .head_offset = offsetof(struct test_obj, node), | |
87545899 | 902 | * .hashfn = jhash, |
7e1e7763 | 903 | * .obj_hashfn = my_hash_fn, |
7e1e7763 TG |
904 | * }; |
905 | */ | |
906 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
907 | { | |
908 | struct bucket_table *tbl; | |
909 | size_t size; | |
a5b6846f | 910 | u32 hash_rnd; |
7e1e7763 TG |
911 | |
912 | size = HASH_DEFAULT_SIZE; | |
913 | ||
914 | if ((params->key_len && !params->hashfn) || | |
915 | (!params->key_len && !params->obj_hashfn)) | |
916 | return -EINVAL; | |
917 | ||
f89bd6f8 TG |
918 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
919 | return -EINVAL; | |
920 | ||
94000176 YX |
921 | params->min_shift = max_t(size_t, params->min_shift, |
922 | ilog2(HASH_MIN_SIZE)); | |
923 | ||
7e1e7763 | 924 | if (params->nelem_hint) |
94000176 | 925 | size = rounded_hashtable_size(params); |
7e1e7763 | 926 | |
97defe1e TG |
927 | memset(ht, 0, sizeof(*ht)); |
928 | mutex_init(&ht->mutex); | |
929 | memcpy(&ht->p, params, sizeof(*params)); | |
f2dba9c6 | 930 | INIT_LIST_HEAD(&ht->walkers); |
97defe1e TG |
931 | |
932 | if (params->locks_mul) | |
933 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
934 | else | |
935 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
936 | ||
a5b6846f DB |
937 | get_random_bytes(&hash_rnd, sizeof(hash_rnd)); |
938 | ||
939 | tbl = bucket_table_alloc(ht, size, hash_rnd); | |
7e1e7763 TG |
940 | if (tbl == NULL) |
941 | return -ENOMEM; | |
942 | ||
545a148e | 943 | atomic_set(&ht->nelems, 0); |
a5b6846f | 944 | |
7e1e7763 | 945 | RCU_INIT_POINTER(ht->tbl, tbl); |
97defe1e | 946 | RCU_INIT_POINTER(ht->future_tbl, tbl); |
7e1e7763 | 947 | |
4c4b52d9 | 948 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 949 | |
7e1e7763 TG |
950 | return 0; |
951 | } | |
952 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
953 | ||
954 | /** | |
955 | * rhashtable_destroy - destroy hash table | |
956 | * @ht: the hash table to destroy | |
957 | * | |
ae82ddcf PNA |
958 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
959 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
960 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 961 | */ |
97defe1e | 962 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 963 | { |
97defe1e TG |
964 | ht->being_destroyed = true; |
965 | ||
4c4b52d9 | 966 | cancel_work_sync(&ht->run_work); |
97defe1e | 967 | |
57699a40 | 968 | mutex_lock(&ht->mutex); |
97defe1e | 969 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 970 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
971 | } |
972 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |