]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Resizable, Scalable, Concurrent Hash Table | |
3 | * | |
4 | * Copyright (c) 2014 Thomas Graf <tgraf@suug.ch> | |
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> | |
20 | #include <linux/slab.h> | |
21 | #include <linux/vmalloc.h> | |
22 | #include <linux/mm.h> | |
23 | #include <linux/jhash.h> | |
24 | #include <linux/random.h> | |
25 | #include <linux/rhashtable.h> | |
26 | ||
27 | #define HASH_DEFAULT_SIZE 64UL | |
28 | #define HASH_MIN_SIZE 4UL | |
29 | #define BUCKET_LOCKS_PER_CPU 128UL | |
30 | ||
31 | /* Base bits plus 1 bit for nulls marker */ | |
32 | #define HASH_RESERVED_SPACE (RHT_BASE_BITS + 1) | |
33 | ||
34 | enum { | |
35 | RHT_LOCK_NORMAL, | |
36 | RHT_LOCK_NESTED, | |
37 | RHT_LOCK_NESTED2, | |
38 | }; | |
39 | ||
40 | /* The bucket lock is selected based on the hash and protects mutations | |
41 | * on a group of hash buckets. | |
42 | * | |
43 | * IMPORTANT: When holding the bucket lock of both the old and new table | |
44 | * during expansions and shrinking, the old bucket lock must always be | |
45 | * acquired first. | |
46 | */ | |
47 | static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash) | |
48 | { | |
49 | return &tbl->locks[hash & tbl->locks_mask]; | |
50 | } | |
51 | ||
52 | #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT)) | |
53 | #define ASSERT_BUCKET_LOCK(TBL, HASH) \ | |
54 | BUG_ON(!lockdep_rht_bucket_is_held(TBL, HASH)) | |
55 | ||
56 | #ifdef CONFIG_PROVE_LOCKING | |
57 | int lockdep_rht_mutex_is_held(struct rhashtable *ht) | |
58 | { | |
59 | return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1; | |
60 | } | |
61 | EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held); | |
62 | ||
63 | int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash) | |
64 | { | |
65 | spinlock_t *lock = bucket_lock(tbl, hash); | |
66 | ||
67 | return (debug_locks) ? lockdep_is_held(lock) : 1; | |
68 | } | |
69 | EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held); | |
70 | #endif | |
71 | ||
72 | static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he) | |
73 | { | |
74 | return (void *) he - ht->p.head_offset; | |
75 | } | |
76 | ||
77 | static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash) | |
78 | { | |
79 | return hash & (tbl->size - 1); | |
80 | } | |
81 | ||
82 | static u32 obj_raw_hashfn(const struct rhashtable *ht, const void *ptr) | |
83 | { | |
84 | u32 hash; | |
85 | ||
86 | if (unlikely(!ht->p.key_len)) | |
87 | hash = ht->p.obj_hashfn(ptr, ht->p.hash_rnd); | |
88 | else | |
89 | hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len, | |
90 | ht->p.hash_rnd); | |
91 | ||
92 | return hash >> HASH_RESERVED_SPACE; | |
93 | } | |
94 | ||
95 | static u32 key_hashfn(struct rhashtable *ht, const void *key, u32 len) | |
96 | { | |
97 | struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht); | |
98 | u32 hash; | |
99 | ||
100 | hash = ht->p.hashfn(key, len, ht->p.hash_rnd); | |
101 | hash >>= HASH_RESERVED_SPACE; | |
102 | ||
103 | return rht_bucket_index(tbl, hash); | |
104 | } | |
105 | ||
106 | static u32 head_hashfn(const struct rhashtable *ht, | |
107 | const struct bucket_table *tbl, | |
108 | const struct rhash_head *he) | |
109 | { | |
110 | return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he))); | |
111 | } | |
112 | ||
113 | static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n) | |
114 | { | |
115 | struct rhash_head __rcu **pprev; | |
116 | ||
117 | for (pprev = &tbl->buckets[n]; | |
118 | !rht_is_a_nulls(rht_dereference_bucket(*pprev, tbl, n)); | |
119 | pprev = &rht_dereference_bucket(*pprev, tbl, n)->next) | |
120 | ; | |
121 | ||
122 | return pprev; | |
123 | } | |
124 | ||
125 | static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl) | |
126 | { | |
127 | unsigned int i, size; | |
128 | #if defined(CONFIG_PROVE_LOCKING) | |
129 | unsigned int nr_pcpus = 2; | |
130 | #else | |
131 | unsigned int nr_pcpus = num_possible_cpus(); | |
132 | #endif | |
133 | ||
134 | nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL); | |
135 | size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul); | |
136 | ||
137 | /* Never allocate more than one lock per bucket */ | |
138 | size = min_t(unsigned int, size, tbl->size); | |
139 | ||
140 | if (sizeof(spinlock_t) != 0) { | |
141 | #ifdef CONFIG_NUMA | |
142 | if (size * sizeof(spinlock_t) > PAGE_SIZE) | |
143 | tbl->locks = vmalloc(size * sizeof(spinlock_t)); | |
144 | else | |
145 | #endif | |
146 | tbl->locks = kmalloc_array(size, sizeof(spinlock_t), | |
147 | GFP_KERNEL); | |
148 | if (!tbl->locks) | |
149 | return -ENOMEM; | |
150 | for (i = 0; i < size; i++) | |
151 | spin_lock_init(&tbl->locks[i]); | |
152 | } | |
153 | tbl->locks_mask = size - 1; | |
154 | ||
155 | return 0; | |
156 | } | |
157 | ||
158 | static void bucket_table_free(const struct bucket_table *tbl) | |
159 | { | |
160 | if (tbl) | |
161 | kvfree(tbl->locks); | |
162 | ||
163 | kvfree(tbl); | |
164 | } | |
165 | ||
166 | static struct bucket_table *bucket_table_alloc(struct rhashtable *ht, | |
167 | size_t nbuckets) | |
168 | { | |
169 | struct bucket_table *tbl; | |
170 | size_t size; | |
171 | int i; | |
172 | ||
173 | size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]); | |
174 | tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
175 | if (tbl == NULL) | |
176 | tbl = vzalloc(size); | |
177 | ||
178 | if (tbl == NULL) | |
179 | return NULL; | |
180 | ||
181 | tbl->size = nbuckets; | |
182 | ||
183 | if (alloc_bucket_locks(ht, tbl) < 0) { | |
184 | bucket_table_free(tbl); | |
185 | return NULL; | |
186 | } | |
187 | ||
188 | for (i = 0; i < nbuckets; i++) | |
189 | INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i); | |
190 | ||
191 | return tbl; | |
192 | } | |
193 | ||
194 | /** | |
195 | * rht_grow_above_75 - returns true if nelems > 0.75 * table-size | |
196 | * @ht: hash table | |
197 | * @new_size: new table size | |
198 | */ | |
199 | bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size) | |
200 | { | |
201 | /* Expand table when exceeding 75% load */ | |
202 | return atomic_read(&ht->nelems) > (new_size / 4 * 3) && | |
203 | (ht->p.max_shift && atomic_read(&ht->shift) < ht->p.max_shift); | |
204 | } | |
205 | EXPORT_SYMBOL_GPL(rht_grow_above_75); | |
206 | ||
207 | /** | |
208 | * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size | |
209 | * @ht: hash table | |
210 | * @new_size: new table size | |
211 | */ | |
212 | bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size) | |
213 | { | |
214 | /* Shrink table beneath 30% load */ | |
215 | return atomic_read(&ht->nelems) < (new_size * 3 / 10) && | |
216 | (atomic_read(&ht->shift) > ht->p.min_shift); | |
217 | } | |
218 | EXPORT_SYMBOL_GPL(rht_shrink_below_30); | |
219 | ||
220 | static void hashtable_chain_unzip(const struct rhashtable *ht, | |
221 | const struct bucket_table *new_tbl, | |
222 | struct bucket_table *old_tbl, | |
223 | size_t old_hash) | |
224 | { | |
225 | struct rhash_head *he, *p, *next; | |
226 | spinlock_t *new_bucket_lock, *new_bucket_lock2 = NULL; | |
227 | unsigned int new_hash, new_hash2; | |
228 | ||
229 | ASSERT_BUCKET_LOCK(old_tbl, old_hash); | |
230 | ||
231 | /* Old bucket empty, no work needed. */ | |
232 | p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl, | |
233 | old_hash); | |
234 | if (rht_is_a_nulls(p)) | |
235 | return; | |
236 | ||
237 | new_hash = new_hash2 = head_hashfn(ht, new_tbl, p); | |
238 | new_bucket_lock = bucket_lock(new_tbl, new_hash); | |
239 | ||
240 | /* Advance the old bucket pointer one or more times until it | |
241 | * reaches a node that doesn't hash to the same bucket as the | |
242 | * previous node p. Call the previous node p; | |
243 | */ | |
244 | rht_for_each_continue(he, p->next, old_tbl, old_hash) { | |
245 | new_hash2 = head_hashfn(ht, new_tbl, he); | |
246 | if (new_hash != new_hash2) | |
247 | break; | |
248 | p = he; | |
249 | } | |
250 | rcu_assign_pointer(old_tbl->buckets[old_hash], p->next); | |
251 | ||
252 | spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED); | |
253 | ||
254 | /* If we have encountered an entry that maps to a different bucket in | |
255 | * the new table, lock down that bucket as well as we might cut off | |
256 | * the end of the chain. | |
257 | */ | |
258 | new_bucket_lock2 = bucket_lock(new_tbl, new_hash); | |
259 | if (new_bucket_lock != new_bucket_lock2) | |
260 | spin_lock_bh_nested(new_bucket_lock2, RHT_LOCK_NESTED2); | |
261 | ||
262 | /* Find the subsequent node which does hash to the same | |
263 | * bucket as node P, or NULL if no such node exists. | |
264 | */ | |
265 | INIT_RHT_NULLS_HEAD(next, ht, old_hash); | |
266 | if (!rht_is_a_nulls(he)) { | |
267 | rht_for_each_continue(he, he->next, old_tbl, old_hash) { | |
268 | if (head_hashfn(ht, new_tbl, he) == new_hash) { | |
269 | next = he; | |
270 | break; | |
271 | } | |
272 | } | |
273 | } | |
274 | ||
275 | /* Set p's next pointer to that subsequent node pointer, | |
276 | * bypassing the nodes which do not hash to p's bucket | |
277 | */ | |
278 | rcu_assign_pointer(p->next, next); | |
279 | ||
280 | if (new_bucket_lock != new_bucket_lock2) | |
281 | spin_unlock_bh(new_bucket_lock2); | |
282 | spin_unlock_bh(new_bucket_lock); | |
283 | } | |
284 | ||
285 | static void link_old_to_new(struct bucket_table *new_tbl, | |
286 | unsigned int new_hash, struct rhash_head *entry) | |
287 | { | |
288 | spinlock_t *new_bucket_lock; | |
289 | ||
290 | new_bucket_lock = bucket_lock(new_tbl, new_hash); | |
291 | ||
292 | spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED); | |
293 | rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry); | |
294 | spin_unlock_bh(new_bucket_lock); | |
295 | } | |
296 | ||
297 | /** | |
298 | * rhashtable_expand - Expand hash table while allowing concurrent lookups | |
299 | * @ht: the hash table to expand | |
300 | * | |
301 | * A secondary bucket array is allocated and the hash entries are migrated | |
302 | * while keeping them on both lists until the end of the RCU grace period. | |
303 | * | |
304 | * This function may only be called in a context where it is safe to call | |
305 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
306 | * | |
307 | * The caller must ensure that no concurrent resizing occurs by holding | |
308 | * ht->mutex. | |
309 | * | |
310 | * It is valid to have concurrent insertions and deletions protected by per | |
311 | * bucket locks or concurrent RCU protected lookups and traversals. | |
312 | */ | |
313 | int rhashtable_expand(struct rhashtable *ht) | |
314 | { | |
315 | struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht); | |
316 | struct rhash_head *he; | |
317 | spinlock_t *old_bucket_lock; | |
318 | unsigned int new_hash, old_hash; | |
319 | bool complete = false; | |
320 | ||
321 | ASSERT_RHT_MUTEX(ht); | |
322 | ||
323 | new_tbl = bucket_table_alloc(ht, old_tbl->size * 2); | |
324 | if (new_tbl == NULL) | |
325 | return -ENOMEM; | |
326 | ||
327 | atomic_inc(&ht->shift); | |
328 | ||
329 | /* Make insertions go into the new, empty table right away. Deletions | |
330 | * and lookups will be attempted in both tables until we synchronize. | |
331 | * The synchronize_rcu() guarantees for the new table to be picked up | |
332 | * so no new additions go into the old table while we relink. | |
333 | */ | |
334 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
335 | synchronize_rcu(); | |
336 | ||
337 | /* For each new bucket, search the corresponding old bucket for the | |
338 | * first entry that hashes to the new bucket, and link the end of | |
339 | * newly formed bucket chain (containing entries added to future | |
340 | * table) to that entry. Since all the entries which will end up in | |
341 | * the new bucket appear in the same old bucket, this constructs an | |
342 | * entirely valid new hash table, but with multiple buckets | |
343 | * "zipped" together into a single imprecise chain. | |
344 | */ | |
345 | for (new_hash = 0; new_hash < new_tbl->size; new_hash++) { | |
346 | old_hash = rht_bucket_index(old_tbl, new_hash); | |
347 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
348 | ||
349 | spin_lock_bh(old_bucket_lock); | |
350 | rht_for_each(he, old_tbl, old_hash) { | |
351 | if (head_hashfn(ht, new_tbl, he) == new_hash) { | |
352 | link_old_to_new(new_tbl, new_hash, he); | |
353 | break; | |
354 | } | |
355 | } | |
356 | spin_unlock_bh(old_bucket_lock); | |
357 | } | |
358 | ||
359 | /* Publish the new table pointer. Lookups may now traverse | |
360 | * the new table, but they will not benefit from any | |
361 | * additional efficiency until later steps unzip the buckets. | |
362 | */ | |
363 | rcu_assign_pointer(ht->tbl, new_tbl); | |
364 | ||
365 | /* Unzip interleaved hash chains */ | |
366 | while (!complete && !ht->being_destroyed) { | |
367 | /* Wait for readers. All new readers will see the new | |
368 | * table, and thus no references to the old table will | |
369 | * remain. | |
370 | */ | |
371 | synchronize_rcu(); | |
372 | ||
373 | /* For each bucket in the old table (each of which | |
374 | * contains items from multiple buckets of the new | |
375 | * table): ... | |
376 | */ | |
377 | complete = true; | |
378 | for (old_hash = 0; old_hash < old_tbl->size; old_hash++) { | |
379 | struct rhash_head *head; | |
380 | ||
381 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
382 | spin_lock_bh(old_bucket_lock); | |
383 | ||
384 | hashtable_chain_unzip(ht, new_tbl, old_tbl, old_hash); | |
385 | head = rht_dereference_bucket(old_tbl->buckets[old_hash], | |
386 | old_tbl, old_hash); | |
387 | if (!rht_is_a_nulls(head)) | |
388 | complete = false; | |
389 | ||
390 | spin_unlock_bh(old_bucket_lock); | |
391 | } | |
392 | } | |
393 | ||
394 | bucket_table_free(old_tbl); | |
395 | return 0; | |
396 | } | |
397 | EXPORT_SYMBOL_GPL(rhashtable_expand); | |
398 | ||
399 | /** | |
400 | * rhashtable_shrink - Shrink hash table while allowing concurrent lookups | |
401 | * @ht: the hash table to shrink | |
402 | * | |
403 | * This function may only be called in a context where it is safe to call | |
404 | * synchronize_rcu(), e.g. not within a rcu_read_lock() section. | |
405 | * | |
406 | * The caller must ensure that no concurrent resizing occurs by holding | |
407 | * ht->mutex. | |
408 | * | |
409 | * The caller must ensure that no concurrent table mutations take place. | |
410 | * It is however valid to have concurrent lookups if they are RCU protected. | |
411 | * | |
412 | * It is valid to have concurrent insertions and deletions protected by per | |
413 | * bucket locks or concurrent RCU protected lookups and traversals. | |
414 | */ | |
415 | int rhashtable_shrink(struct rhashtable *ht) | |
416 | { | |
417 | struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht); | |
418 | spinlock_t *new_bucket_lock, *old_bucket_lock1, *old_bucket_lock2; | |
419 | unsigned int new_hash; | |
420 | ||
421 | ASSERT_RHT_MUTEX(ht); | |
422 | ||
423 | new_tbl = bucket_table_alloc(ht, tbl->size / 2); | |
424 | if (new_tbl == NULL) | |
425 | return -ENOMEM; | |
426 | ||
427 | rcu_assign_pointer(ht->future_tbl, new_tbl); | |
428 | synchronize_rcu(); | |
429 | ||
430 | /* Link the first entry in the old bucket to the end of the | |
431 | * bucket in the new table. As entries are concurrently being | |
432 | * added to the new table, lock down the new bucket. As we | |
433 | * always divide the size in half when shrinking, each bucket | |
434 | * in the new table maps to exactly two buckets in the old | |
435 | * table. | |
436 | * | |
437 | * As removals can occur concurrently on the old table, we need | |
438 | * to lock down both matching buckets in the old table. | |
439 | */ | |
440 | for (new_hash = 0; new_hash < new_tbl->size; new_hash++) { | |
441 | old_bucket_lock1 = bucket_lock(tbl, new_hash); | |
442 | old_bucket_lock2 = bucket_lock(tbl, new_hash + new_tbl->size); | |
443 | new_bucket_lock = bucket_lock(new_tbl, new_hash); | |
444 | ||
445 | spin_lock_bh(old_bucket_lock1); | |
446 | ||
447 | /* Depending on the lock per buckets mapping, the bucket in | |
448 | * the lower and upper region may map to the same lock. | |
449 | */ | |
450 | if (old_bucket_lock1 != old_bucket_lock2) { | |
451 | spin_lock_bh_nested(old_bucket_lock2, RHT_LOCK_NESTED); | |
452 | spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED2); | |
453 | } else { | |
454 | spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED); | |
455 | } | |
456 | ||
457 | rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), | |
458 | tbl->buckets[new_hash]); | |
459 | rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), | |
460 | tbl->buckets[new_hash + new_tbl->size]); | |
461 | ||
462 | spin_unlock_bh(new_bucket_lock); | |
463 | if (old_bucket_lock1 != old_bucket_lock2) | |
464 | spin_unlock_bh(old_bucket_lock2); | |
465 | spin_unlock_bh(old_bucket_lock1); | |
466 | } | |
467 | ||
468 | /* Publish the new, valid hash table */ | |
469 | rcu_assign_pointer(ht->tbl, new_tbl); | |
470 | atomic_dec(&ht->shift); | |
471 | ||
472 | /* Wait for readers. No new readers will have references to the | |
473 | * old hash table. | |
474 | */ | |
475 | synchronize_rcu(); | |
476 | ||
477 | bucket_table_free(tbl); | |
478 | ||
479 | return 0; | |
480 | } | |
481 | EXPORT_SYMBOL_GPL(rhashtable_shrink); | |
482 | ||
483 | static void rht_deferred_worker(struct work_struct *work) | |
484 | { | |
485 | struct rhashtable *ht; | |
486 | struct bucket_table *tbl; | |
487 | ||
488 | ht = container_of(work, struct rhashtable, run_work); | |
489 | mutex_lock(&ht->mutex); | |
490 | tbl = rht_dereference(ht->tbl, ht); | |
491 | ||
492 | if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size)) | |
493 | rhashtable_expand(ht); | |
494 | else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size)) | |
495 | rhashtable_shrink(ht); | |
496 | ||
497 | mutex_unlock(&ht->mutex); | |
498 | } | |
499 | ||
500 | static void rhashtable_wakeup_worker(struct rhashtable *ht) | |
501 | { | |
502 | struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht); | |
503 | struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
504 | size_t size = tbl->size; | |
505 | ||
506 | /* Only adjust the table if no resizing is currently in progress. */ | |
507 | if (tbl == new_tbl && | |
508 | ((ht->p.grow_decision && ht->p.grow_decision(ht, size)) || | |
509 | (ht->p.shrink_decision && ht->p.shrink_decision(ht, size)))) | |
510 | schedule_work(&ht->run_work); | |
511 | } | |
512 | ||
513 | static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj, | |
514 | struct bucket_table *tbl, u32 hash) | |
515 | { | |
516 | struct rhash_head *head = rht_dereference_bucket(tbl->buckets[hash], | |
517 | tbl, hash); | |
518 | ||
519 | if (rht_is_a_nulls(head)) | |
520 | INIT_RHT_NULLS_HEAD(obj->next, ht, hash); | |
521 | else | |
522 | RCU_INIT_POINTER(obj->next, head); | |
523 | ||
524 | rcu_assign_pointer(tbl->buckets[hash], obj); | |
525 | ||
526 | atomic_inc(&ht->nelems); | |
527 | ||
528 | rhashtable_wakeup_worker(ht); | |
529 | } | |
530 | ||
531 | /** | |
532 | * rhashtable_insert - insert object into hash table | |
533 | * @ht: hash table | |
534 | * @obj: pointer to hash head inside object | |
535 | * | |
536 | * Will take a per bucket spinlock to protect against mutual mutations | |
537 | * on the same bucket. Multiple insertions may occur in parallel unless | |
538 | * they map to the same bucket lock. | |
539 | * | |
540 | * It is safe to call this function from atomic context. | |
541 | * | |
542 | * Will trigger an automatic deferred table resizing if the size grows | |
543 | * beyond the watermark indicated by grow_decision() which can be passed | |
544 | * to rhashtable_init(). | |
545 | */ | |
546 | void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj) | |
547 | { | |
548 | struct bucket_table *tbl; | |
549 | spinlock_t *lock; | |
550 | unsigned hash; | |
551 | ||
552 | rcu_read_lock(); | |
553 | ||
554 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
555 | hash = head_hashfn(ht, tbl, obj); | |
556 | lock = bucket_lock(tbl, hash); | |
557 | ||
558 | spin_lock_bh(lock); | |
559 | __rhashtable_insert(ht, obj, tbl, hash); | |
560 | spin_unlock_bh(lock); | |
561 | ||
562 | rcu_read_unlock(); | |
563 | } | |
564 | EXPORT_SYMBOL_GPL(rhashtable_insert); | |
565 | ||
566 | /** | |
567 | * rhashtable_remove - remove object from hash table | |
568 | * @ht: hash table | |
569 | * @obj: pointer to hash head inside object | |
570 | * | |
571 | * Since the hash chain is single linked, the removal operation needs to | |
572 | * walk the bucket chain upon removal. The removal operation is thus | |
573 | * considerable slow if the hash table is not correctly sized. | |
574 | * | |
575 | * Will automatically shrink the table via rhashtable_expand() if the | |
576 | * shrink_decision function specified at rhashtable_init() returns true. | |
577 | * | |
578 | * The caller must ensure that no concurrent table mutations occur. It is | |
579 | * however valid to have concurrent lookups if they are RCU protected. | |
580 | */ | |
581 | bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj) | |
582 | { | |
583 | struct bucket_table *tbl; | |
584 | struct rhash_head __rcu **pprev; | |
585 | struct rhash_head *he; | |
586 | spinlock_t *lock; | |
587 | unsigned int hash; | |
588 | bool ret = false; | |
589 | ||
590 | rcu_read_lock(); | |
591 | tbl = rht_dereference_rcu(ht->tbl, ht); | |
592 | hash = head_hashfn(ht, tbl, obj); | |
593 | ||
594 | lock = bucket_lock(tbl, hash); | |
595 | spin_lock_bh(lock); | |
596 | ||
597 | restart: | |
598 | pprev = &tbl->buckets[hash]; | |
599 | rht_for_each(he, tbl, hash) { | |
600 | if (he != obj) { | |
601 | pprev = &he->next; | |
602 | continue; | |
603 | } | |
604 | ||
605 | rcu_assign_pointer(*pprev, obj->next); | |
606 | ||
607 | ret = true; | |
608 | break; | |
609 | } | |
610 | ||
611 | /* The entry may be linked in either 'tbl', 'future_tbl', or both. | |
612 | * 'future_tbl' only exists for a short period of time during | |
613 | * resizing. Thus traversing both is fine and the added cost is | |
614 | * very rare. | |
615 | */ | |
616 | if (tbl != rht_dereference_rcu(ht->future_tbl, ht)) { | |
617 | spin_unlock_bh(lock); | |
618 | ||
619 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
620 | hash = head_hashfn(ht, tbl, obj); | |
621 | ||
622 | lock = bucket_lock(tbl, hash); | |
623 | spin_lock_bh(lock); | |
624 | goto restart; | |
625 | } | |
626 | ||
627 | spin_unlock_bh(lock); | |
628 | ||
629 | if (ret) { | |
630 | atomic_dec(&ht->nelems); | |
631 | rhashtable_wakeup_worker(ht); | |
632 | } | |
633 | ||
634 | rcu_read_unlock(); | |
635 | ||
636 | return ret; | |
637 | } | |
638 | EXPORT_SYMBOL_GPL(rhashtable_remove); | |
639 | ||
640 | struct rhashtable_compare_arg { | |
641 | struct rhashtable *ht; | |
642 | const void *key; | |
643 | }; | |
644 | ||
645 | static bool rhashtable_compare(void *ptr, void *arg) | |
646 | { | |
647 | struct rhashtable_compare_arg *x = arg; | |
648 | struct rhashtable *ht = x->ht; | |
649 | ||
650 | return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len); | |
651 | } | |
652 | ||
653 | /** | |
654 | * rhashtable_lookup - lookup key in hash table | |
655 | * @ht: hash table | |
656 | * @key: pointer to key | |
657 | * | |
658 | * Computes the hash value for the key and traverses the bucket chain looking | |
659 | * for a entry with an identical key. The first matching entry is returned. | |
660 | * | |
661 | * This lookup function may only be used for fixed key hash table (key_len | |
662 | * parameter set). It will BUG() if used inappropriately. | |
663 | * | |
664 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
665 | */ | |
666 | void *rhashtable_lookup(struct rhashtable *ht, const void *key) | |
667 | { | |
668 | struct rhashtable_compare_arg arg = { | |
669 | .ht = ht, | |
670 | .key = key, | |
671 | }; | |
672 | ||
673 | BUG_ON(!ht->p.key_len); | |
674 | ||
675 | return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg); | |
676 | } | |
677 | EXPORT_SYMBOL_GPL(rhashtable_lookup); | |
678 | ||
679 | /** | |
680 | * rhashtable_lookup_compare - search hash table with compare function | |
681 | * @ht: hash table | |
682 | * @key: the pointer to the key | |
683 | * @compare: compare function, must return true on match | |
684 | * @arg: argument passed on to compare function | |
685 | * | |
686 | * Traverses the bucket chain behind the provided hash value and calls the | |
687 | * specified compare function for each entry. | |
688 | * | |
689 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
690 | * | |
691 | * Returns the first entry on which the compare function returned true. | |
692 | */ | |
693 | void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key, | |
694 | bool (*compare)(void *, void *), void *arg) | |
695 | { | |
696 | const struct bucket_table *tbl, *old_tbl; | |
697 | struct rhash_head *he; | |
698 | u32 hash; | |
699 | ||
700 | rcu_read_lock(); | |
701 | ||
702 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
703 | tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
704 | hash = key_hashfn(ht, key, ht->p.key_len); | |
705 | restart: | |
706 | rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) { | |
707 | if (!compare(rht_obj(ht, he), arg)) | |
708 | continue; | |
709 | rcu_read_unlock(); | |
710 | return rht_obj(ht, he); | |
711 | } | |
712 | ||
713 | if (unlikely(tbl != old_tbl)) { | |
714 | tbl = old_tbl; | |
715 | goto restart; | |
716 | } | |
717 | rcu_read_unlock(); | |
718 | ||
719 | return NULL; | |
720 | } | |
721 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare); | |
722 | ||
723 | /** | |
724 | * rhashtable_lookup_insert - lookup and insert object into hash table | |
725 | * @ht: hash table | |
726 | * @obj: pointer to hash head inside object | |
727 | * | |
728 | * Locks down the bucket chain in both the old and new table if a resize | |
729 | * is in progress to ensure that writers can't remove from the old table | |
730 | * and can't insert to the new table during the atomic operation of search | |
731 | * and insertion. Searches for duplicates in both the old and new table if | |
732 | * a resize is in progress. | |
733 | * | |
734 | * This lookup function may only be used for fixed key hash table (key_len | |
735 | * parameter set). It will BUG() if used inappropriately. | |
736 | * | |
737 | * It is safe to call this function from atomic context. | |
738 | * | |
739 | * Will trigger an automatic deferred table resizing if the size grows | |
740 | * beyond the watermark indicated by grow_decision() which can be passed | |
741 | * to rhashtable_init(). | |
742 | */ | |
743 | bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj) | |
744 | { | |
745 | struct rhashtable_compare_arg arg = { | |
746 | .ht = ht, | |
747 | .key = rht_obj(ht, obj) + ht->p.key_offset, | |
748 | }; | |
749 | ||
750 | BUG_ON(!ht->p.key_len); | |
751 | ||
752 | return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare, | |
753 | &arg); | |
754 | } | |
755 | EXPORT_SYMBOL_GPL(rhashtable_lookup_insert); | |
756 | ||
757 | /** | |
758 | * rhashtable_lookup_compare_insert - search and insert object to hash table | |
759 | * with compare function | |
760 | * @ht: hash table | |
761 | * @obj: pointer to hash head inside object | |
762 | * @compare: compare function, must return true on match | |
763 | * @arg: argument passed on to compare function | |
764 | * | |
765 | * Locks down the bucket chain in both the old and new table if a resize | |
766 | * is in progress to ensure that writers can't remove from the old table | |
767 | * and can't insert to the new table during the atomic operation of search | |
768 | * and insertion. Searches for duplicates in both the old and new table if | |
769 | * a resize is in progress. | |
770 | * | |
771 | * Lookups may occur in parallel with hashtable mutations and resizing. | |
772 | * | |
773 | * Will trigger an automatic deferred table resizing if the size grows | |
774 | * beyond the watermark indicated by grow_decision() which can be passed | |
775 | * to rhashtable_init(). | |
776 | */ | |
777 | bool rhashtable_lookup_compare_insert(struct rhashtable *ht, | |
778 | struct rhash_head *obj, | |
779 | bool (*compare)(void *, void *), | |
780 | void *arg) | |
781 | { | |
782 | struct bucket_table *new_tbl, *old_tbl; | |
783 | spinlock_t *new_bucket_lock, *old_bucket_lock; | |
784 | u32 new_hash, old_hash; | |
785 | bool success = true; | |
786 | ||
787 | BUG_ON(!ht->p.key_len); | |
788 | ||
789 | rcu_read_lock(); | |
790 | ||
791 | old_tbl = rht_dereference_rcu(ht->tbl, ht); | |
792 | old_hash = head_hashfn(ht, old_tbl, obj); | |
793 | old_bucket_lock = bucket_lock(old_tbl, old_hash); | |
794 | spin_lock_bh(old_bucket_lock); | |
795 | ||
796 | new_tbl = rht_dereference_rcu(ht->future_tbl, ht); | |
797 | new_hash = head_hashfn(ht, new_tbl, obj); | |
798 | new_bucket_lock = bucket_lock(new_tbl, new_hash); | |
799 | if (unlikely(old_tbl != new_tbl)) | |
800 | spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED); | |
801 | ||
802 | if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset, | |
803 | compare, arg)) { | |
804 | success = false; | |
805 | goto exit; | |
806 | } | |
807 | ||
808 | __rhashtable_insert(ht, obj, new_tbl, new_hash); | |
809 | ||
810 | exit: | |
811 | if (unlikely(old_tbl != new_tbl)) | |
812 | spin_unlock_bh(new_bucket_lock); | |
813 | spin_unlock_bh(old_bucket_lock); | |
814 | ||
815 | rcu_read_unlock(); | |
816 | ||
817 | return success; | |
818 | } | |
819 | EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert); | |
820 | ||
821 | static size_t rounded_hashtable_size(struct rhashtable_params *params) | |
822 | { | |
823 | return max(roundup_pow_of_two(params->nelem_hint * 4 / 3), | |
824 | 1UL << params->min_shift); | |
825 | } | |
826 | ||
827 | /** | |
828 | * rhashtable_init - initialize a new hash table | |
829 | * @ht: hash table to be initialized | |
830 | * @params: configuration parameters | |
831 | * | |
832 | * Initializes a new hash table based on the provided configuration | |
833 | * parameters. A table can be configured either with a variable or | |
834 | * fixed length key: | |
835 | * | |
836 | * Configuration Example 1: Fixed length keys | |
837 | * struct test_obj { | |
838 | * int key; | |
839 | * void * my_member; | |
840 | * struct rhash_head node; | |
841 | * }; | |
842 | * | |
843 | * struct rhashtable_params params = { | |
844 | * .head_offset = offsetof(struct test_obj, node), | |
845 | * .key_offset = offsetof(struct test_obj, key), | |
846 | * .key_len = sizeof(int), | |
847 | * .hashfn = jhash, | |
848 | * .nulls_base = (1U << RHT_BASE_SHIFT), | |
849 | * }; | |
850 | * | |
851 | * Configuration Example 2: Variable length keys | |
852 | * struct test_obj { | |
853 | * [...] | |
854 | * struct rhash_head node; | |
855 | * }; | |
856 | * | |
857 | * u32 my_hash_fn(const void *data, u32 seed) | |
858 | * { | |
859 | * struct test_obj *obj = data; | |
860 | * | |
861 | * return [... hash ...]; | |
862 | * } | |
863 | * | |
864 | * struct rhashtable_params params = { | |
865 | * .head_offset = offsetof(struct test_obj, node), | |
866 | * .hashfn = jhash, | |
867 | * .obj_hashfn = my_hash_fn, | |
868 | * }; | |
869 | */ | |
870 | int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params) | |
871 | { | |
872 | struct bucket_table *tbl; | |
873 | size_t size; | |
874 | ||
875 | size = HASH_DEFAULT_SIZE; | |
876 | ||
877 | if ((params->key_len && !params->hashfn) || | |
878 | (!params->key_len && !params->obj_hashfn)) | |
879 | return -EINVAL; | |
880 | ||
881 | if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT)) | |
882 | return -EINVAL; | |
883 | ||
884 | params->min_shift = max_t(size_t, params->min_shift, | |
885 | ilog2(HASH_MIN_SIZE)); | |
886 | ||
887 | if (params->nelem_hint) | |
888 | size = rounded_hashtable_size(params); | |
889 | ||
890 | memset(ht, 0, sizeof(*ht)); | |
891 | mutex_init(&ht->mutex); | |
892 | memcpy(&ht->p, params, sizeof(*params)); | |
893 | ||
894 | if (params->locks_mul) | |
895 | ht->p.locks_mul = roundup_pow_of_two(params->locks_mul); | |
896 | else | |
897 | ht->p.locks_mul = BUCKET_LOCKS_PER_CPU; | |
898 | ||
899 | tbl = bucket_table_alloc(ht, size); | |
900 | if (tbl == NULL) | |
901 | return -ENOMEM; | |
902 | ||
903 | atomic_set(&ht->nelems, 0); | |
904 | atomic_set(&ht->shift, ilog2(tbl->size)); | |
905 | RCU_INIT_POINTER(ht->tbl, tbl); | |
906 | RCU_INIT_POINTER(ht->future_tbl, tbl); | |
907 | ||
908 | if (!ht->p.hash_rnd) | |
909 | get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd)); | |
910 | ||
911 | if (ht->p.grow_decision || ht->p.shrink_decision) | |
912 | INIT_WORK(&ht->run_work, rht_deferred_worker); | |
913 | ||
914 | return 0; | |
915 | } | |
916 | EXPORT_SYMBOL_GPL(rhashtable_init); | |
917 | ||
918 | /** | |
919 | * rhashtable_destroy - destroy hash table | |
920 | * @ht: the hash table to destroy | |
921 | * | |
922 | * Frees the bucket array. This function is not rcu safe, therefore the caller | |
923 | * has to make sure that no resizing may happen by unpublishing the hashtable | |
924 | * and waiting for the quiescent cycle before releasing the bucket array. | |
925 | */ | |
926 | void rhashtable_destroy(struct rhashtable *ht) | |
927 | { | |
928 | ht->being_destroyed = true; | |
929 | ||
930 | if (ht->p.grow_decision || ht->p.shrink_decision) | |
931 | cancel_work_sync(&ht->run_work); | |
932 | ||
933 | mutex_lock(&ht->mutex); | |
934 | bucket_table_free(rht_dereference(ht->tbl, ht)); | |
935 | mutex_unlock(&ht->mutex); | |
936 | } | |
937 | EXPORT_SYMBOL_GPL(rhashtable_destroy); |