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, | |
150 | size_t nbuckets) | |
7e1e7763 | 151 | { |
eb6d1abf | 152 | struct bucket_table *tbl = NULL; |
7e1e7763 | 153 | size_t size; |
f89bd6f8 | 154 | int i; |
7e1e7763 TG |
155 | |
156 | size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]); | |
eb6d1abf DB |
157 | if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) |
158 | tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); | |
7e1e7763 TG |
159 | if (tbl == NULL) |
160 | tbl = vzalloc(size); | |
7e1e7763 TG |
161 | if (tbl == NULL) |
162 | return NULL; | |
163 | ||
164 | tbl->size = nbuckets; | |
165 | ||
97defe1e TG |
166 | if (alloc_bucket_locks(ht, tbl) < 0) { |
167 | bucket_table_free(tbl); | |
168 | return NULL; | |
169 | } | |
7e1e7763 | 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 | |
180 | * @new_size: new table size | |
181 | */ | |
4c4b52d9 | 182 | static bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size) |
7e1e7763 TG |
183 | { |
184 | /* Expand table when exceeding 75% load */ | |
c0c09bfd | 185 | return atomic_read(&ht->nelems) > (new_size / 4 * 3) && |
8331de75 | 186 | (!ht->p.max_shift || atomic_read(&ht->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 | |
192 | * @new_size: new table size | |
193 | */ | |
4c4b52d9 | 194 | static bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size) |
7e1e7763 TG |
195 | { |
196 | /* Shrink table beneath 30% load */ | |
c0c09bfd YX |
197 | return atomic_read(&ht->nelems) < (new_size * 3 / 10) && |
198 | (atomic_read(&ht->shift) > ht->p.min_shift); | |
7e1e7763 | 199 | } |
7e1e7763 | 200 | |
aa34a6cb | 201 | static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash) |
a5ec68e3 | 202 | { |
aa34a6cb HX |
203 | struct bucket_table *new_tbl = rht_dereference(ht->future_tbl, ht); |
204 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
205 | struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash]; | |
206 | int err = -ENOENT; | |
207 | struct rhash_head *head, *next, *entry; | |
208 | spinlock_t *new_bucket_lock; | |
209 | unsigned new_hash; | |
210 | ||
211 | rht_for_each(entry, old_tbl, old_hash) { | |
212 | err = 0; | |
213 | next = rht_dereference_bucket(entry->next, old_tbl, old_hash); | |
214 | ||
215 | if (rht_is_a_nulls(next)) | |
216 | break; | |
a5ec68e3 | 217 | |
aa34a6cb HX |
218 | pprev = &entry->next; |
219 | } | |
a5ec68e3 | 220 | |
aa34a6cb HX |
221 | if (err) |
222 | goto out; | |
97defe1e | 223 | |
aa34a6cb | 224 | new_hash = head_hashfn(ht, new_tbl, entry); |
7e1e7763 | 225 | |
aa34a6cb | 226 | new_bucket_lock = bucket_lock(new_tbl, new_hash); |
7e1e7763 | 227 | |
84ed82b7 | 228 | spin_lock_nested(new_bucket_lock, RHT_LOCK_NESTED); |
aa34a6cb HX |
229 | head = rht_dereference_bucket(new_tbl->buckets[new_hash], |
230 | new_tbl, new_hash); | |
97defe1e | 231 | |
aa34a6cb HX |
232 | if (rht_is_a_nulls(head)) |
233 | INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash); | |
234 | else | |
235 | RCU_INIT_POINTER(entry->next, head); | |
a5ec68e3 | 236 | |
aa34a6cb HX |
237 | rcu_assign_pointer(new_tbl->buckets[new_hash], entry); |
238 | spin_unlock(new_bucket_lock); | |
97defe1e | 239 | |
aa34a6cb | 240 | rcu_assign_pointer(*pprev, next); |
7e1e7763 | 241 | |
aa34a6cb HX |
242 | out: |
243 | return err; | |
244 | } | |
97defe1e | 245 | |
aa34a6cb HX |
246 | static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash) |
247 | { | |
248 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); | |
249 | spinlock_t *old_bucket_lock; | |
250 | ||
251 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
a5ec68e3 | 252 | |
aa34a6cb HX |
253 | spin_lock_bh(old_bucket_lock); |
254 | while (!rhashtable_rehash_one(ht, old_hash)) | |
255 | ; | |
256 | spin_unlock_bh(old_bucket_lock); | |
97defe1e TG |
257 | } |
258 | ||
aa34a6cb HX |
259 | static void rhashtable_rehash(struct rhashtable *ht, |
260 | struct bucket_table *new_tbl) | |
97defe1e | 261 | { |
aa34a6cb HX |
262 | struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht); |
263 | unsigned old_hash; | |
7cd10db8 | 264 | |
aa34a6cb HX |
265 | get_random_bytes(&new_tbl->hash_rnd, sizeof(new_tbl->hash_rnd)); |
266 | ||
267 | /* Make insertions go into the new, empty table right away. Deletions | |
268 | * and lookups will be attempted in both tables until we synchronize. | |
269 | * The synchronize_rcu() guarantees for the new table to be picked up | |
270 | * so no new additions go into the old table while we relink. | |
271 | */ | |
272 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
273 | ||
274 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) | |
275 | rhashtable_rehash_chain(ht, old_hash); | |
276 | ||
277 | /* Publish the new table pointer. */ | |
278 | rcu_assign_pointer(ht->tbl, new_tbl); | |
279 | ||
280 | /* Wait for readers. All new readers will see the new | |
281 | * table, and thus no references to the old table will | |
282 | * remain. | |
283 | */ | |
284 | synchronize_rcu(); | |
285 | ||
286 | bucket_table_free(old_tbl); | |
7e1e7763 TG |
287 | } |
288 | ||
289 | /** | |
290 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
291 | * @ht: the hash table to expand | |
7e1e7763 | 292 | * |
aa34a6cb | 293 | * A secondary bucket array is allocated and the hash entries are migrated. |
7e1e7763 TG |
294 | * |
295 | * This function may only be called in a context where it is safe to call | |
296 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
297 | * | |
97defe1e TG |
298 | * The caller must ensure that no concurrent resizing occurs by holding |
299 | * ht->mutex. | |
300 | * | |
301 | * It is valid to have concurrent insertions and deletions protected by per | |
302 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 303 | */ |
6eba8224 | 304 | int rhashtable_expand(struct rhashtable *ht) |
7e1e7763 TG |
305 | { |
306 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
7e1e7763 TG |
307 | |
308 | ASSERT_RHT_MUTEX(ht); | |
309 | ||
97defe1e | 310 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2); |
7e1e7763 TG |
311 | if (new_tbl == NULL) |
312 | return -ENOMEM; | |
313 | ||
988dfbd7 HX |
314 | new_tbl->hash_rnd = old_tbl->hash_rnd; |
315 | ||
c0c09bfd | 316 | atomic_inc(&ht->shift); |
7e1e7763 | 317 | |
aa34a6cb | 318 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 | 319 | |
7e1e7763 TG |
320 | return 0; |
321 | } | |
322 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
323 | ||
324 | /** | |
325 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
326 | * @ht: the hash table to shrink | |
7e1e7763 TG |
327 | * |
328 | * This function may only be called in a context where it is safe to call | |
329 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
330 | * | |
97defe1e TG |
331 | * The caller must ensure that no concurrent resizing occurs by holding |
332 | * ht->mutex. | |
333 | * | |
7e1e7763 TG |
334 | * The caller must ensure that no concurrent table mutations take place. |
335 | * It is however valid to have concurrent lookups if they are RCU protected. | |
97defe1e TG |
336 | * |
337 | * It is valid to have concurrent insertions and deletions protected by per | |
338 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 339 | */ |
6eba8224 | 340 | int rhashtable_shrink(struct rhashtable *ht) |
7e1e7763 | 341 | { |
97defe1e | 342 | struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht); |
7e1e7763 TG |
343 | |
344 | ASSERT_RHT_MUTEX(ht); | |
345 | ||
97defe1e TG |
346 | new_tbl = bucket_table_alloc(ht, tbl->size / 2); |
347 | if (new_tbl == NULL) | |
7e1e7763 TG |
348 | return -ENOMEM; |
349 | ||
988dfbd7 HX |
350 | new_tbl->hash_rnd = tbl->hash_rnd; |
351 | ||
c0c09bfd | 352 | atomic_dec(&ht->shift); |
7e1e7763 | 353 | |
aa34a6cb | 354 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
355 | |
356 | return 0; | |
357 | } | |
358 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
359 | ||
97defe1e TG |
360 | static void rht_deferred_worker(struct work_struct *work) |
361 | { | |
362 | struct rhashtable *ht; | |
363 | struct bucket_table *tbl; | |
f2dba9c6 | 364 | struct rhashtable_walker *walker; |
97defe1e | 365 | |
57699a40 | 366 | ht = container_of(work, struct rhashtable, run_work); |
97defe1e | 367 | mutex_lock(&ht->mutex); |
28134a53 HX |
368 | if (ht->being_destroyed) |
369 | goto unlock; | |
370 | ||
97defe1e TG |
371 | tbl = rht_dereference(ht->tbl, ht); |
372 | ||
f2dba9c6 HX |
373 | list_for_each_entry(walker, &ht->walkers, list) |
374 | walker->resize = true; | |
375 | ||
4c4b52d9 | 376 | if (rht_grow_above_75(ht, tbl->size)) |
97defe1e | 377 | rhashtable_expand(ht); |
4c4b52d9 | 378 | else if (rht_shrink_below_30(ht, tbl->size)) |
97defe1e | 379 | rhashtable_shrink(ht); |
28134a53 | 380 | unlock: |
97defe1e TG |
381 | mutex_unlock(&ht->mutex); |
382 | } | |
383 | ||
aa34a6cb HX |
384 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
385 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 386 | { |
aa34a6cb | 387 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 388 | struct rhash_head *head; |
aa34a6cb HX |
389 | bool no_resize_running; |
390 | unsigned hash; | |
391 | bool success = true; | |
392 | ||
393 | rcu_read_lock(); | |
394 | ||
395 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
eca84933 | 396 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
397 | |
398 | spin_lock_bh(bucket_lock(old_tbl, hash)); | |
399 | ||
400 | /* Because we have already taken the bucket lock in old_tbl, | |
401 | * if we find that future_tbl is not yet visible then that | |
402 | * guarantees all other insertions of the same entry will | |
403 | * also grab the bucket lock in old_tbl because until the | |
404 | * rehash completes ht->tbl won't be changed. | |
405 | */ | |
406 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
407 | if (tbl != old_tbl) { | |
eca84933 | 408 | hash = head_hashfn(ht, tbl, obj); |
84ed82b7 | 409 | spin_lock_nested(bucket_lock(tbl, hash), RHT_LOCK_NESTED); |
aa34a6cb HX |
410 | } |
411 | ||
412 | if (compare && | |
413 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
414 | compare, arg)) { | |
415 | success = false; | |
416 | goto exit; | |
417 | } | |
418 | ||
419 | no_resize_running = tbl == old_tbl; | |
020219a6 | 420 | |
020219a6 | 421 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); |
db304854 YX |
422 | |
423 | if (rht_is_a_nulls(head)) | |
424 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
425 | else | |
426 | RCU_INIT_POINTER(obj->next, head); | |
427 | ||
428 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
429 | ||
430 | atomic_inc(&ht->nelems); | |
4c4b52d9 DB |
431 | if (no_resize_running && rht_grow_above_75(ht, tbl->size)) |
432 | schedule_work(&ht->run_work); | |
aa34a6cb HX |
433 | |
434 | exit: | |
435 | if (tbl != old_tbl) { | |
eca84933 | 436 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb HX |
437 | spin_unlock(bucket_lock(tbl, hash)); |
438 | } | |
439 | ||
eca84933 | 440 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
441 | spin_unlock_bh(bucket_lock(old_tbl, hash)); |
442 | ||
443 | rcu_read_unlock(); | |
444 | ||
445 | return success; | |
db304854 YX |
446 | } |
447 | ||
7e1e7763 | 448 | /** |
db304854 | 449 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
450 | * @ht: hash table |
451 | * @obj: pointer to hash head inside object | |
7e1e7763 | 452 | * |
97defe1e TG |
453 | * Will take a per bucket spinlock to protect against mutual mutations |
454 | * on the same bucket. Multiple insertions may occur in parallel unless | |
455 | * they map to the same bucket lock. | |
7e1e7763 | 456 | * |
97defe1e TG |
457 | * It is safe to call this function from atomic context. |
458 | * | |
459 | * Will trigger an automatic deferred table resizing if the size grows | |
460 | * beyond the watermark indicated by grow_decision() which can be passed | |
461 | * to rhashtable_init(). | |
7e1e7763 | 462 | */ |
6eba8224 | 463 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 464 | { |
aa34a6cb HX |
465 | __rhashtable_insert(ht, obj, NULL, NULL); |
466 | } | |
467 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
468 | ||
469 | static bool __rhashtable_remove(struct rhashtable *ht, | |
470 | struct bucket_table *tbl, | |
471 | struct rhash_head *obj) | |
472 | { | |
473 | struct rhash_head __rcu **pprev; | |
474 | struct rhash_head *he; | |
475 | spinlock_t * lock; | |
97defe1e | 476 | unsigned hash; |
aa34a6cb | 477 | bool ret = false; |
7e1e7763 | 478 | |
eca84933 | 479 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb | 480 | lock = bucket_lock(tbl, hash); |
7e1e7763 | 481 | |
aa34a6cb | 482 | spin_lock_bh(lock); |
97defe1e | 483 | |
aa34a6cb HX |
484 | pprev = &tbl->buckets[hash]; |
485 | rht_for_each(he, tbl, hash) { | |
486 | if (he != obj) { | |
487 | pprev = &he->next; | |
488 | continue; | |
489 | } | |
7e1e7763 | 490 | |
aa34a6cb HX |
491 | rcu_assign_pointer(*pprev, obj->next); |
492 | ret = true; | |
493 | break; | |
494 | } | |
495 | ||
496 | spin_unlock_bh(lock); | |
497 | ||
498 | return ret; | |
7e1e7763 | 499 | } |
7e1e7763 | 500 | |
7e1e7763 TG |
501 | /** |
502 | * rhashtable_remove - remove object from hash table | |
503 | * @ht: hash table | |
504 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
505 | * |
506 | * Since the hash chain is single linked, the removal operation needs to | |
507 | * walk the bucket chain upon removal. The removal operation is thus | |
508 | * considerable slow if the hash table is not correctly sized. | |
509 | * | |
db304854 | 510 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
511 | * shrink_decision function specified at rhashtable_init() returns true. |
512 | * | |
513 | * The caller must ensure that no concurrent table mutations occur. It is | |
514 | * however valid to have concurrent lookups if they are RCU protected. | |
515 | */ | |
6eba8224 | 516 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 517 | { |
aa34a6cb HX |
518 | struct bucket_table *tbl, *old_tbl; |
519 | bool ret; | |
7e1e7763 | 520 | |
97defe1e | 521 | rcu_read_lock(); |
7e1e7763 | 522 | |
aa34a6cb HX |
523 | old_tbl = rht_dereference_rcu(ht->tbl, ht); |
524 | ret = __rhashtable_remove(ht, old_tbl, obj); | |
7e1e7763 | 525 | |
aa34a6cb HX |
526 | /* Because we have already taken (and released) the bucket |
527 | * lock in old_tbl, if we find that future_tbl is not yet | |
528 | * visible then that guarantees the entry to still be in | |
529 | * old_tbl if it exists. | |
fe6a043c | 530 | */ |
aa34a6cb HX |
531 | tbl = rht_dereference_rcu(ht->future_tbl, ht); |
532 | if (!ret && old_tbl != tbl) | |
533 | ret = __rhashtable_remove(ht, tbl, obj); | |
fe6a043c TG |
534 | |
535 | if (ret) { | |
aa34a6cb | 536 | bool no_resize_running = tbl == old_tbl; |
4c4b52d9 | 537 | |
fe6a043c | 538 | atomic_dec(&ht->nelems); |
aa34a6cb | 539 | if (no_resize_running && rht_shrink_below_30(ht, tbl->size)) |
4c4b52d9 | 540 | schedule_work(&ht->run_work); |
fe6a043c TG |
541 | } |
542 | ||
97defe1e TG |
543 | rcu_read_unlock(); |
544 | ||
fe6a043c | 545 | return ret; |
7e1e7763 TG |
546 | } |
547 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
548 | ||
efb975a6 YX |
549 | struct rhashtable_compare_arg { |
550 | struct rhashtable *ht; | |
551 | const void *key; | |
552 | }; | |
553 | ||
554 | static bool rhashtable_compare(void *ptr, void *arg) | |
555 | { | |
556 | struct rhashtable_compare_arg *x = arg; | |
557 | struct rhashtable *ht = x->ht; | |
558 | ||
559 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
560 | } | |
561 | ||
7e1e7763 TG |
562 | /** |
563 | * rhashtable_lookup - lookup key in hash table | |
564 | * @ht: hash table | |
565 | * @key: pointer to key | |
566 | * | |
567 | * Computes the hash value for the key and traverses the bucket chain looking | |
568 | * for a entry with an identical key. The first matching entry is returned. | |
569 | * | |
570 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 571 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 572 | * |
97defe1e | 573 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 574 | */ |
97defe1e | 575 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 576 | { |
efb975a6 YX |
577 | struct rhashtable_compare_arg arg = { |
578 | .ht = ht, | |
579 | .key = key, | |
580 | }; | |
7e1e7763 TG |
581 | |
582 | BUG_ON(!ht->p.key_len); | |
583 | ||
efb975a6 | 584 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
585 | } |
586 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
587 | ||
588 | /** | |
589 | * rhashtable_lookup_compare - search hash table with compare function | |
590 | * @ht: hash table | |
8d24c0b4 | 591 | * @key: the pointer to the key |
7e1e7763 TG |
592 | * @compare: compare function, must return true on match |
593 | * @arg: argument passed on to compare function | |
594 | * | |
595 | * Traverses the bucket chain behind the provided hash value and calls the | |
596 | * specified compare function for each entry. | |
597 | * | |
97defe1e | 598 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
599 | * |
600 | * Returns the first entry on which the compare function returned true. | |
601 | */ | |
97defe1e | 602 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
603 | bool (*compare)(void *, void *), void *arg) |
604 | { | |
97defe1e | 605 | const struct bucket_table *tbl, *old_tbl; |
7e1e7763 | 606 | struct rhash_head *he; |
8d24c0b4 | 607 | u32 hash; |
7e1e7763 | 608 | |
97defe1e TG |
609 | rcu_read_lock(); |
610 | ||
aa34a6cb | 611 | tbl = rht_dereference_rcu(ht->tbl, ht); |
cffaa9cb | 612 | hash = key_hashfn(ht, tbl, key); |
97defe1e | 613 | restart: |
8d2b1879 | 614 | rht_for_each_rcu(he, tbl, hash) { |
7e1e7763 TG |
615 | if (!compare(rht_obj(ht, he), arg)) |
616 | continue; | |
97defe1e | 617 | rcu_read_unlock(); |
a4b18cda | 618 | return rht_obj(ht, he); |
7e1e7763 TG |
619 | } |
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; | |
910 | ||
911 | size = HASH_DEFAULT_SIZE; | |
912 | ||
913 | if ((params->key_len && !params->hashfn) || | |
914 | (!params->key_len && !params->obj_hashfn)) | |
915 | return -EINVAL; | |
916 | ||
f89bd6f8 TG |
917 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
918 | return -EINVAL; | |
919 | ||
94000176 YX |
920 | params->min_shift = max_t(size_t, params->min_shift, |
921 | ilog2(HASH_MIN_SIZE)); | |
922 | ||
7e1e7763 | 923 | if (params->nelem_hint) |
94000176 | 924 | size = rounded_hashtable_size(params); |
7e1e7763 | 925 | |
97defe1e TG |
926 | memset(ht, 0, sizeof(*ht)); |
927 | mutex_init(&ht->mutex); | |
928 | memcpy(&ht->p, params, sizeof(*params)); | |
f2dba9c6 | 929 | INIT_LIST_HEAD(&ht->walkers); |
97defe1e TG |
930 | |
931 | if (params->locks_mul) | |
932 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
933 | else | |
934 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
935 | ||
936 | tbl = bucket_table_alloc(ht, size); | |
7e1e7763 TG |
937 | if (tbl == NULL) |
938 | return -ENOMEM; | |
939 | ||
988dfbd7 HX |
940 | get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd)); |
941 | ||
545a148e | 942 | atomic_set(&ht->nelems, 0); |
c0c09bfd | 943 | atomic_set(&ht->shift, ilog2(tbl->size)); |
7e1e7763 | 944 | RCU_INIT_POINTER(ht->tbl, tbl); |
97defe1e | 945 | RCU_INIT_POINTER(ht->future_tbl, tbl); |
7e1e7763 | 946 | |
4c4b52d9 | 947 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 948 | |
7e1e7763 TG |
949 | return 0; |
950 | } | |
951 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
952 | ||
953 | /** | |
954 | * rhashtable_destroy - destroy hash table | |
955 | * @ht: the hash table to destroy | |
956 | * | |
ae82ddcf PNA |
957 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
958 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
959 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 960 | */ |
97defe1e | 961 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 962 | { |
97defe1e TG |
963 | ht->being_destroyed = true; |
964 | ||
4c4b52d9 | 965 | cancel_work_sync(&ht->run_work); |
97defe1e | 966 | |
57699a40 | 967 | mutex_lock(&ht->mutex); |
97defe1e | 968 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 969 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
970 | } |
971 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |