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 | ||
9497df88 HX |
274 | /* Ensure the new table is visible to readers. */ |
275 | smp_wmb(); | |
276 | ||
aa34a6cb HX |
277 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) |
278 | rhashtable_rehash_chain(ht, old_hash); | |
279 | ||
280 | /* Publish the new table pointer. */ | |
281 | rcu_assign_pointer(ht->tbl, new_tbl); | |
282 | ||
283 | /* Wait for readers. All new readers will see the new | |
284 | * table, and thus no references to the old table will | |
285 | * remain. | |
286 | */ | |
287 | synchronize_rcu(); | |
288 | ||
289 | bucket_table_free(old_tbl); | |
7e1e7763 TG |
290 | } |
291 | ||
292 | /** | |
293 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
294 | * @ht: the hash table to expand | |
7e1e7763 | 295 | * |
aa34a6cb | 296 | * A secondary bucket array is allocated and the hash entries are migrated. |
7e1e7763 TG |
297 | * |
298 | * This function may only be called in a context where it is safe to call | |
299 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
300 | * | |
97defe1e TG |
301 | * The caller must ensure that no concurrent resizing occurs by holding |
302 | * ht->mutex. | |
303 | * | |
304 | * It is valid to have concurrent insertions and deletions protected by per | |
305 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 306 | */ |
6eba8224 | 307 | int rhashtable_expand(struct rhashtable *ht) |
7e1e7763 TG |
308 | { |
309 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
7e1e7763 TG |
310 | |
311 | ASSERT_RHT_MUTEX(ht); | |
312 | ||
97defe1e | 313 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2); |
7e1e7763 TG |
314 | if (new_tbl == NULL) |
315 | return -ENOMEM; | |
316 | ||
988dfbd7 HX |
317 | new_tbl->hash_rnd = old_tbl->hash_rnd; |
318 | ||
c0c09bfd | 319 | atomic_inc(&ht->shift); |
7e1e7763 | 320 | |
aa34a6cb | 321 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 | 322 | |
7e1e7763 TG |
323 | return 0; |
324 | } | |
325 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
326 | ||
327 | /** | |
328 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
329 | * @ht: the hash table to shrink | |
7e1e7763 TG |
330 | * |
331 | * This function may only be called in a context where it is safe to call | |
332 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
333 | * | |
97defe1e TG |
334 | * The caller must ensure that no concurrent resizing occurs by holding |
335 | * ht->mutex. | |
336 | * | |
7e1e7763 TG |
337 | * The caller must ensure that no concurrent table mutations take place. |
338 | * It is however valid to have concurrent lookups if they are RCU protected. | |
97defe1e TG |
339 | * |
340 | * It is valid to have concurrent insertions and deletions protected by per | |
341 | * bucket locks or concurrent RCU protected lookups and traversals. | |
7e1e7763 | 342 | */ |
6eba8224 | 343 | int rhashtable_shrink(struct rhashtable *ht) |
7e1e7763 | 344 | { |
97defe1e | 345 | struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht); |
7e1e7763 TG |
346 | |
347 | ASSERT_RHT_MUTEX(ht); | |
348 | ||
97defe1e TG |
349 | new_tbl = bucket_table_alloc(ht, tbl->size / 2); |
350 | if (new_tbl == NULL) | |
7e1e7763 TG |
351 | return -ENOMEM; |
352 | ||
988dfbd7 HX |
353 | new_tbl->hash_rnd = tbl->hash_rnd; |
354 | ||
c0c09bfd | 355 | atomic_dec(&ht->shift); |
7e1e7763 | 356 | |
aa34a6cb | 357 | rhashtable_rehash(ht, new_tbl); |
7e1e7763 TG |
358 | |
359 | return 0; | |
360 | } | |
361 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
362 | ||
97defe1e TG |
363 | static void rht_deferred_worker(struct work_struct *work) |
364 | { | |
365 | struct rhashtable *ht; | |
366 | struct bucket_table *tbl; | |
f2dba9c6 | 367 | struct rhashtable_walker *walker; |
97defe1e | 368 | |
57699a40 | 369 | ht = container_of(work, struct rhashtable, run_work); |
97defe1e | 370 | mutex_lock(&ht->mutex); |
28134a53 HX |
371 | if (ht->being_destroyed) |
372 | goto unlock; | |
373 | ||
97defe1e TG |
374 | tbl = rht_dereference(ht->tbl, ht); |
375 | ||
f2dba9c6 HX |
376 | list_for_each_entry(walker, &ht->walkers, list) |
377 | walker->resize = true; | |
378 | ||
4c4b52d9 | 379 | if (rht_grow_above_75(ht, tbl->size)) |
97defe1e | 380 | rhashtable_expand(ht); |
4c4b52d9 | 381 | else if (rht_shrink_below_30(ht, tbl->size)) |
97defe1e | 382 | rhashtable_shrink(ht); |
28134a53 | 383 | unlock: |
97defe1e TG |
384 | mutex_unlock(&ht->mutex); |
385 | } | |
386 | ||
aa34a6cb HX |
387 | static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, |
388 | bool (*compare)(void *, void *), void *arg) | |
db304854 | 389 | { |
aa34a6cb | 390 | struct bucket_table *tbl, *old_tbl; |
020219a6 | 391 | struct rhash_head *head; |
aa34a6cb HX |
392 | bool no_resize_running; |
393 | unsigned hash; | |
394 | bool success = true; | |
395 | ||
396 | rcu_read_lock(); | |
397 | ||
398 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
eca84933 | 399 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
400 | |
401 | spin_lock_bh(bucket_lock(old_tbl, hash)); | |
402 | ||
403 | /* Because we have already taken the bucket lock in old_tbl, | |
404 | * if we find that future_tbl is not yet visible then that | |
405 | * guarantees all other insertions of the same entry will | |
406 | * also grab the bucket lock in old_tbl because until the | |
407 | * rehash completes ht->tbl won't be changed. | |
408 | */ | |
409 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
410 | if (tbl != old_tbl) { | |
eca84933 | 411 | hash = head_hashfn(ht, tbl, obj); |
84ed82b7 | 412 | spin_lock_nested(bucket_lock(tbl, hash), RHT_LOCK_NESTED); |
aa34a6cb HX |
413 | } |
414 | ||
415 | if (compare && | |
416 | rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
417 | compare, arg)) { | |
418 | success = false; | |
419 | goto exit; | |
420 | } | |
421 | ||
422 | no_resize_running = tbl == old_tbl; | |
020219a6 | 423 | |
020219a6 | 424 | head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash); |
db304854 YX |
425 | |
426 | if (rht_is_a_nulls(head)) | |
427 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
428 | else | |
429 | RCU_INIT_POINTER(obj->next, head); | |
430 | ||
431 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
432 | ||
433 | atomic_inc(&ht->nelems); | |
4c4b52d9 DB |
434 | if (no_resize_running && rht_grow_above_75(ht, tbl->size)) |
435 | schedule_work(&ht->run_work); | |
aa34a6cb HX |
436 | |
437 | exit: | |
438 | if (tbl != old_tbl) { | |
eca84933 | 439 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb HX |
440 | spin_unlock(bucket_lock(tbl, hash)); |
441 | } | |
442 | ||
eca84933 | 443 | hash = head_hashfn(ht, old_tbl, obj); |
aa34a6cb HX |
444 | spin_unlock_bh(bucket_lock(old_tbl, hash)); |
445 | ||
446 | rcu_read_unlock(); | |
447 | ||
448 | return success; | |
db304854 YX |
449 | } |
450 | ||
7e1e7763 | 451 | /** |
db304854 | 452 | * rhashtable_insert - insert object into hash table |
7e1e7763 TG |
453 | * @ht: hash table |
454 | * @obj: pointer to hash head inside object | |
7e1e7763 | 455 | * |
97defe1e TG |
456 | * Will take a per bucket spinlock to protect against mutual mutations |
457 | * on the same bucket. Multiple insertions may occur in parallel unless | |
458 | * they map to the same bucket lock. | |
7e1e7763 | 459 | * |
97defe1e TG |
460 | * It is safe to call this function from atomic context. |
461 | * | |
462 | * Will trigger an automatic deferred table resizing if the size grows | |
463 | * beyond the watermark indicated by grow_decision() which can be passed | |
464 | * to rhashtable_init(). | |
7e1e7763 | 465 | */ |
6eba8224 | 466 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 467 | { |
aa34a6cb HX |
468 | __rhashtable_insert(ht, obj, NULL, NULL); |
469 | } | |
470 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
471 | ||
472 | static bool __rhashtable_remove(struct rhashtable *ht, | |
473 | struct bucket_table *tbl, | |
474 | struct rhash_head *obj) | |
475 | { | |
476 | struct rhash_head __rcu **pprev; | |
477 | struct rhash_head *he; | |
478 | spinlock_t * lock; | |
97defe1e | 479 | unsigned hash; |
aa34a6cb | 480 | bool ret = false; |
7e1e7763 | 481 | |
eca84933 | 482 | hash = head_hashfn(ht, tbl, obj); |
aa34a6cb | 483 | lock = bucket_lock(tbl, hash); |
7e1e7763 | 484 | |
aa34a6cb | 485 | spin_lock_bh(lock); |
97defe1e | 486 | |
aa34a6cb HX |
487 | pprev = &tbl->buckets[hash]; |
488 | rht_for_each(he, tbl, hash) { | |
489 | if (he != obj) { | |
490 | pprev = &he->next; | |
491 | continue; | |
492 | } | |
7e1e7763 | 493 | |
aa34a6cb HX |
494 | rcu_assign_pointer(*pprev, obj->next); |
495 | ret = true; | |
496 | break; | |
497 | } | |
498 | ||
499 | spin_unlock_bh(lock); | |
500 | ||
501 | return ret; | |
7e1e7763 | 502 | } |
7e1e7763 | 503 | |
7e1e7763 TG |
504 | /** |
505 | * rhashtable_remove - remove object from hash table | |
506 | * @ht: hash table | |
507 | * @obj: pointer to hash head inside object | |
7e1e7763 TG |
508 | * |
509 | * Since the hash chain is single linked, the removal operation needs to | |
510 | * walk the bucket chain upon removal. The removal operation is thus | |
511 | * considerable slow if the hash table is not correctly sized. | |
512 | * | |
db304854 | 513 | * Will automatically shrink the table via rhashtable_expand() if the |
7e1e7763 TG |
514 | * shrink_decision function specified at rhashtable_init() returns true. |
515 | * | |
516 | * The caller must ensure that no concurrent table mutations occur. It is | |
517 | * however valid to have concurrent lookups if they are RCU protected. | |
518 | */ | |
6eba8224 | 519 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) |
7e1e7763 | 520 | { |
aa34a6cb HX |
521 | struct bucket_table *tbl, *old_tbl; |
522 | bool ret; | |
7e1e7763 | 523 | |
97defe1e | 524 | rcu_read_lock(); |
7e1e7763 | 525 | |
aa34a6cb HX |
526 | old_tbl = rht_dereference_rcu(ht->tbl, ht); |
527 | ret = __rhashtable_remove(ht, old_tbl, obj); | |
7e1e7763 | 528 | |
aa34a6cb HX |
529 | /* Because we have already taken (and released) the bucket |
530 | * lock in old_tbl, if we find that future_tbl is not yet | |
531 | * visible then that guarantees the entry to still be in | |
532 | * old_tbl if it exists. | |
fe6a043c | 533 | */ |
aa34a6cb HX |
534 | tbl = rht_dereference_rcu(ht->future_tbl, ht); |
535 | if (!ret && old_tbl != tbl) | |
536 | ret = __rhashtable_remove(ht, tbl, obj); | |
fe6a043c TG |
537 | |
538 | if (ret) { | |
aa34a6cb | 539 | bool no_resize_running = tbl == old_tbl; |
4c4b52d9 | 540 | |
fe6a043c | 541 | atomic_dec(&ht->nelems); |
aa34a6cb | 542 | if (no_resize_running && rht_shrink_below_30(ht, tbl->size)) |
4c4b52d9 | 543 | schedule_work(&ht->run_work); |
fe6a043c TG |
544 | } |
545 | ||
97defe1e TG |
546 | rcu_read_unlock(); |
547 | ||
fe6a043c | 548 | return ret; |
7e1e7763 TG |
549 | } |
550 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
551 | ||
efb975a6 YX |
552 | struct rhashtable_compare_arg { |
553 | struct rhashtable *ht; | |
554 | const void *key; | |
555 | }; | |
556 | ||
557 | static bool rhashtable_compare(void *ptr, void *arg) | |
558 | { | |
559 | struct rhashtable_compare_arg *x = arg; | |
560 | struct rhashtable *ht = x->ht; | |
561 | ||
562 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
563 | } | |
564 | ||
7e1e7763 TG |
565 | /** |
566 | * rhashtable_lookup - lookup key in hash table | |
567 | * @ht: hash table | |
568 | * @key: pointer to key | |
569 | * | |
570 | * Computes the hash value for the key and traverses the bucket chain looking | |
571 | * for a entry with an identical key. The first matching entry is returned. | |
572 | * | |
573 | * This lookup function may only be used for fixed key hash table (key_len | |
db304854 | 574 | * parameter set). It will BUG() if used inappropriately. |
7e1e7763 | 575 | * |
97defe1e | 576 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 | 577 | */ |
97defe1e | 578 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) |
7e1e7763 | 579 | { |
efb975a6 YX |
580 | struct rhashtable_compare_arg arg = { |
581 | .ht = ht, | |
582 | .key = key, | |
583 | }; | |
7e1e7763 TG |
584 | |
585 | BUG_ON(!ht->p.key_len); | |
586 | ||
efb975a6 | 587 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); |
7e1e7763 TG |
588 | } |
589 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
590 | ||
591 | /** | |
592 | * rhashtable_lookup_compare - search hash table with compare function | |
593 | * @ht: hash table | |
8d24c0b4 | 594 | * @key: the pointer to the key |
7e1e7763 TG |
595 | * @compare: compare function, must return true on match |
596 | * @arg: argument passed on to compare function | |
597 | * | |
598 | * Traverses the bucket chain behind the provided hash value and calls the | |
599 | * specified compare function for each entry. | |
600 | * | |
97defe1e | 601 | * Lookups may occur in parallel with hashtable mutations and resizing. |
7e1e7763 TG |
602 | * |
603 | * Returns the first entry on which the compare function returned true. | |
604 | */ | |
97defe1e | 605 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, |
7e1e7763 TG |
606 | bool (*compare)(void *, void *), void *arg) |
607 | { | |
97defe1e | 608 | const struct bucket_table *tbl, *old_tbl; |
7e1e7763 | 609 | struct rhash_head *he; |
8d24c0b4 | 610 | u32 hash; |
7e1e7763 | 611 | |
97defe1e TG |
612 | rcu_read_lock(); |
613 | ||
aa34a6cb | 614 | tbl = rht_dereference_rcu(ht->tbl, ht); |
cffaa9cb | 615 | hash = key_hashfn(ht, tbl, key); |
97defe1e | 616 | restart: |
8d2b1879 | 617 | rht_for_each_rcu(he, tbl, hash) { |
7e1e7763 TG |
618 | if (!compare(rht_obj(ht, he), arg)) |
619 | continue; | |
97defe1e | 620 | rcu_read_unlock(); |
a4b18cda | 621 | return rht_obj(ht, he); |
7e1e7763 TG |
622 | } |
623 | ||
9497df88 HX |
624 | /* Ensure we see any new tables. */ |
625 | smp_rmb(); | |
626 | ||
aa34a6cb HX |
627 | old_tbl = tbl; |
628 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
629 | if (unlikely(tbl != old_tbl)) | |
97defe1e | 630 | goto restart; |
97defe1e TG |
631 | rcu_read_unlock(); |
632 | ||
7e1e7763 TG |
633 | return NULL; |
634 | } | |
635 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
636 | ||
db304854 YX |
637 | /** |
638 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
639 | * @ht: hash table | |
640 | * @obj: pointer to hash head inside object | |
641 | * | |
642 | * Locks down the bucket chain in both the old and new table if a resize | |
643 | * is in progress to ensure that writers can't remove from the old table | |
644 | * and can't insert to the new table during the atomic operation of search | |
645 | * and insertion. Searches for duplicates in both the old and new table if | |
646 | * a resize is in progress. | |
647 | * | |
648 | * This lookup function may only be used for fixed key hash table (key_len | |
649 | * parameter set). It will BUG() if used inappropriately. | |
650 | * | |
651 | * It is safe to call this function from atomic context. | |
652 | * | |
653 | * Will trigger an automatic deferred table resizing if the size grows | |
654 | * beyond the watermark indicated by grow_decision() which can be passed | |
655 | * to rhashtable_init(). | |
656 | */ | |
657 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
7a868d1e YX |
658 | { |
659 | struct rhashtable_compare_arg arg = { | |
660 | .ht = ht, | |
661 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
662 | }; | |
663 | ||
664 | BUG_ON(!ht->p.key_len); | |
665 | ||
666 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
667 | &arg); | |
668 | } | |
669 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
670 | ||
671 | /** | |
672 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
673 | * with compare function | |
674 | * @ht: hash table | |
675 | * @obj: pointer to hash head inside object | |
676 | * @compare: compare function, must return true on match | |
677 | * @arg: argument passed on to compare function | |
678 | * | |
679 | * Locks down the bucket chain in both the old and new table if a resize | |
680 | * is in progress to ensure that writers can't remove from the old table | |
681 | * and can't insert to the new table during the atomic operation of search | |
682 | * and insertion. Searches for duplicates in both the old and new table if | |
683 | * a resize is in progress. | |
684 | * | |
685 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
686 | * | |
687 | * Will trigger an automatic deferred table resizing if the size grows | |
688 | * beyond the watermark indicated by grow_decision() which can be passed | |
689 | * to rhashtable_init(). | |
690 | */ | |
691 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
692 | struct rhash_head *obj, | |
693 | bool (*compare)(void *, void *), | |
694 | void *arg) | |
db304854 | 695 | { |
db304854 YX |
696 | BUG_ON(!ht->p.key_len); |
697 | ||
aa34a6cb | 698 | return __rhashtable_insert(ht, obj, compare, arg); |
db304854 | 699 | } |
7a868d1e | 700 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); |
db304854 | 701 | |
f2dba9c6 HX |
702 | /** |
703 | * rhashtable_walk_init - Initialise an iterator | |
704 | * @ht: Table to walk over | |
705 | * @iter: Hash table Iterator | |
706 | * | |
707 | * This function prepares a hash table walk. | |
708 | * | |
709 | * Note that if you restart a walk after rhashtable_walk_stop you | |
710 | * may see the same object twice. Also, you may miss objects if | |
711 | * there are removals in between rhashtable_walk_stop and the next | |
712 | * call to rhashtable_walk_start. | |
713 | * | |
714 | * For a completely stable walk you should construct your own data | |
715 | * structure outside the hash table. | |
716 | * | |
717 | * This function may sleep so you must not call it from interrupt | |
718 | * context or with spin locks held. | |
719 | * | |
720 | * You must call rhashtable_walk_exit if this function returns | |
721 | * successfully. | |
722 | */ | |
723 | int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter) | |
724 | { | |
725 | iter->ht = ht; | |
726 | iter->p = NULL; | |
727 | iter->slot = 0; | |
728 | iter->skip = 0; | |
729 | ||
730 | iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL); | |
731 | if (!iter->walker) | |
732 | return -ENOMEM; | |
733 | ||
71bb0012 SL |
734 | INIT_LIST_HEAD(&iter->walker->list); |
735 | iter->walker->resize = false; | |
736 | ||
f2dba9c6 HX |
737 | mutex_lock(&ht->mutex); |
738 | list_add(&iter->walker->list, &ht->walkers); | |
739 | mutex_unlock(&ht->mutex); | |
740 | ||
741 | return 0; | |
742 | } | |
743 | EXPORT_SYMBOL_GPL(rhashtable_walk_init); | |
744 | ||
745 | /** | |
746 | * rhashtable_walk_exit - Free an iterator | |
747 | * @iter: Hash table Iterator | |
748 | * | |
749 | * This function frees resources allocated by rhashtable_walk_init. | |
750 | */ | |
751 | void rhashtable_walk_exit(struct rhashtable_iter *iter) | |
752 | { | |
753 | mutex_lock(&iter->ht->mutex); | |
754 | list_del(&iter->walker->list); | |
755 | mutex_unlock(&iter->ht->mutex); | |
756 | kfree(iter->walker); | |
757 | } | |
758 | EXPORT_SYMBOL_GPL(rhashtable_walk_exit); | |
759 | ||
760 | /** | |
761 | * rhashtable_walk_start - Start a hash table walk | |
762 | * @iter: Hash table iterator | |
763 | * | |
764 | * Start a hash table walk. Note that we take the RCU lock in all | |
765 | * cases including when we return an error. So you must always call | |
766 | * rhashtable_walk_stop to clean up. | |
767 | * | |
768 | * Returns zero if successful. | |
769 | * | |
770 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
771 | * will rewind back to the beginning and you may use it immediately | |
772 | * by calling rhashtable_walk_next. | |
773 | */ | |
774 | int rhashtable_walk_start(struct rhashtable_iter *iter) | |
775 | { | |
776 | rcu_read_lock(); | |
777 | ||
778 | if (iter->walker->resize) { | |
779 | iter->slot = 0; | |
780 | iter->skip = 0; | |
781 | iter->walker->resize = false; | |
782 | return -EAGAIN; | |
783 | } | |
784 | ||
785 | return 0; | |
786 | } | |
787 | EXPORT_SYMBOL_GPL(rhashtable_walk_start); | |
788 | ||
789 | /** | |
790 | * rhashtable_walk_next - Return the next object and advance the iterator | |
791 | * @iter: Hash table iterator | |
792 | * | |
793 | * Note that you must call rhashtable_walk_stop when you are finished | |
794 | * with the walk. | |
795 | * | |
796 | * Returns the next object or NULL when the end of the table is reached. | |
797 | * | |
798 | * Returns -EAGAIN if resize event occured. Note that the iterator | |
799 | * will rewind back to the beginning and you may continue to use it. | |
800 | */ | |
801 | void *rhashtable_walk_next(struct rhashtable_iter *iter) | |
802 | { | |
803 | const struct bucket_table *tbl; | |
804 | struct rhashtable *ht = iter->ht; | |
805 | struct rhash_head *p = iter->p; | |
806 | void *obj = NULL; | |
807 | ||
808 | tbl = rht_dereference_rcu(ht->tbl, ht); | |
809 | ||
810 | if (p) { | |
811 | p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot); | |
812 | goto next; | |
813 | } | |
814 | ||
815 | for (; iter->slot < tbl->size; iter->slot++) { | |
816 | int skip = iter->skip; | |
817 | ||
818 | rht_for_each_rcu(p, tbl, iter->slot) { | |
819 | if (!skip) | |
820 | break; | |
821 | skip--; | |
822 | } | |
823 | ||
824 | next: | |
825 | if (!rht_is_a_nulls(p)) { | |
826 | iter->skip++; | |
827 | iter->p = p; | |
828 | obj = rht_obj(ht, p); | |
829 | goto out; | |
830 | } | |
831 | ||
832 | iter->skip = 0; | |
833 | } | |
834 | ||
835 | iter->p = NULL; | |
836 | ||
837 | out: | |
838 | if (iter->walker->resize) { | |
839 | iter->p = NULL; | |
840 | iter->slot = 0; | |
841 | iter->skip = 0; | |
842 | iter->walker->resize = false; | |
843 | return ERR_PTR(-EAGAIN); | |
844 | } | |
845 | ||
846 | return obj; | |
847 | } | |
848 | EXPORT_SYMBOL_GPL(rhashtable_walk_next); | |
849 | ||
850 | /** | |
851 | * rhashtable_walk_stop - Finish a hash table walk | |
852 | * @iter: Hash table iterator | |
853 | * | |
854 | * Finish a hash table walk. | |
855 | */ | |
856 | void rhashtable_walk_stop(struct rhashtable_iter *iter) | |
857 | { | |
858 | rcu_read_unlock(); | |
859 | iter->p = NULL; | |
860 | } | |
861 | EXPORT_SYMBOL_GPL(rhashtable_walk_stop); | |
862 | ||
94000176 | 863 | static size_t rounded_hashtable_size(struct rhashtable_params *params) |
7e1e7763 | 864 | { |
94000176 YX |
865 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), |
866 | 1UL << params->min_shift); | |
7e1e7763 TG |
867 | } |
868 | ||
869 | /** | |
870 | * rhashtable_init - initialize a new hash table | |
871 | * @ht: hash table to be initialized | |
872 | * @params: configuration parameters | |
873 | * | |
874 | * Initializes a new hash table based on the provided configuration | |
875 | * parameters. A table can be configured either with a variable or | |
876 | * fixed length key: | |
877 | * | |
878 | * Configuration Example 1: Fixed length keys | |
879 | * struct test_obj { | |
880 | * int key; | |
881 | * void * my_member; | |
882 | * struct rhash_head node; | |
883 | * }; | |
884 | * | |
885 | * struct rhashtable_params params = { | |
886 | * .head_offset = offsetof(struct test_obj, node), | |
887 | * .key_offset = offsetof(struct test_obj, key), | |
888 | * .key_len = sizeof(int), | |
87545899 | 889 | * .hashfn = jhash, |
f89bd6f8 | 890 | * .nulls_base = (1U << RHT_BASE_SHIFT), |
7e1e7763 TG |
891 | * }; |
892 | * | |
893 | * Configuration Example 2: Variable length keys | |
894 | * struct test_obj { | |
895 | * [...] | |
896 | * struct rhash_head node; | |
897 | * }; | |
898 | * | |
899 | * u32 my_hash_fn(const void *data, u32 seed) | |
900 | * { | |
901 | * struct test_obj *obj = data; | |
902 | * | |
903 | * return [... hash ...]; | |
904 | * } | |
905 | * | |
906 | * struct rhashtable_params params = { | |
907 | * .head_offset = offsetof(struct test_obj, node), | |
87545899 | 908 | * .hashfn = jhash, |
7e1e7763 | 909 | * .obj_hashfn = my_hash_fn, |
7e1e7763 TG |
910 | * }; |
911 | */ | |
912 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
913 | { | |
914 | struct bucket_table *tbl; | |
915 | size_t size; | |
916 | ||
917 | size = HASH_DEFAULT_SIZE; | |
918 | ||
919 | if ((params->key_len && !params->hashfn) || | |
920 | (!params->key_len && !params->obj_hashfn)) | |
921 | return -EINVAL; | |
922 | ||
f89bd6f8 TG |
923 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) |
924 | return -EINVAL; | |
925 | ||
94000176 YX |
926 | params->min_shift = max_t(size_t, params->min_shift, |
927 | ilog2(HASH_MIN_SIZE)); | |
928 | ||
7e1e7763 | 929 | if (params->nelem_hint) |
94000176 | 930 | size = rounded_hashtable_size(params); |
7e1e7763 | 931 | |
97defe1e TG |
932 | memset(ht, 0, sizeof(*ht)); |
933 | mutex_init(&ht->mutex); | |
934 | memcpy(&ht->p, params, sizeof(*params)); | |
f2dba9c6 | 935 | INIT_LIST_HEAD(&ht->walkers); |
97defe1e TG |
936 | |
937 | if (params->locks_mul) | |
938 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
939 | else | |
940 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
941 | ||
942 | tbl = bucket_table_alloc(ht, size); | |
7e1e7763 TG |
943 | if (tbl == NULL) |
944 | return -ENOMEM; | |
945 | ||
988dfbd7 HX |
946 | get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd)); |
947 | ||
545a148e | 948 | atomic_set(&ht->nelems, 0); |
c0c09bfd | 949 | atomic_set(&ht->shift, ilog2(tbl->size)); |
7e1e7763 | 950 | RCU_INIT_POINTER(ht->tbl, tbl); |
97defe1e | 951 | RCU_INIT_POINTER(ht->future_tbl, tbl); |
7e1e7763 | 952 | |
4c4b52d9 | 953 | INIT_WORK(&ht->run_work, rht_deferred_worker); |
97defe1e | 954 | |
7e1e7763 TG |
955 | return 0; |
956 | } | |
957 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
958 | ||
959 | /** | |
960 | * rhashtable_destroy - destroy hash table | |
961 | * @ht: the hash table to destroy | |
962 | * | |
ae82ddcf PNA |
963 | * Frees the bucket array. This function is not rcu safe, therefore the caller |
964 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
965 | * and waiting for the quiescent cycle before releasing the bucket array. | |
7e1e7763 | 966 | */ |
97defe1e | 967 | void rhashtable_destroy(struct rhashtable *ht) |
7e1e7763 | 968 | { |
97defe1e TG |
969 | ht->being_destroyed = true; |
970 | ||
4c4b52d9 | 971 | cancel_work_sync(&ht->run_work); |
97defe1e | 972 | |
57699a40 | 973 | mutex_lock(&ht->mutex); |
97defe1e | 974 | bucket_table_free(rht_dereference(ht->tbl, ht)); |
97defe1e | 975 | mutex_unlock(&ht->mutex); |
7e1e7763 TG |
976 | } |
977 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |