1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017 Intel Corporation
8 * RTE Membership Library
10 * The Membership Library is an extension and generalization of a traditional
11 * filter (for example Bloom Filter and cuckoo filter) structure that has
12 * multiple usages in a variety of workloads and applications. The library is
13 * used to test if a key belongs to certain sets. Two types of such
14 * "set-summary" structures are implemented: hash-table based (HT) and vector
15 * bloom filter (vBF). For HT setsummary, two subtypes or modes are available,
16 * cache and non-cache modes. The table below summarize some properties of
17 * the different implementations.
20 * @b EXPERIMENTAL: this API may change without prior notice
25 * +==========+=====================+================+=========================+
26 * | type | vbf | HT-cache | HT-non-cache |
27 * +==========+=====================+==========================================+
28 * |structure | bloom-filter array | hash-table like without storing key |
29 * +----------+---------------------+------------------------------------------+
30 * |set id | limited by bf count | [1, 0x7fff] |
32 * +----------+---------------------+------------------------------------------+
33 * |usages & | small set range, | can delete, | cache most recent keys, |
34 * |properties| user-specified | big set range, | have both false-positive|
35 * | | false-positive rate,| small false | and false-negative |
36 * | | no deletion support.| positive depend| depend on table size, |
37 * | | | on table size, | automatic overwritten. |
38 * | | | new key does | |
39 * | | | not overwrite | |
40 * | | | existing key. | |
41 * +----------+---------------------+----------------+-------------------------+
45 #ifndef _RTE_MEMBER_H_
46 #define _RTE_MEMBER_H_
54 #include <rte_common.h>
56 /** The set ID type that stored internally in hash table based set summary. */
57 typedef uint16_t member_set_t;
58 /** Invalid set ID used to mean no match found. */
59 #define RTE_MEMBER_NO_MATCH 0
60 /** Maximum size of hash table that can be created. */
61 #define RTE_MEMBER_ENTRIES_MAX (1 << 30)
62 /** Maximum number of keys that can be searched as a bulk */
63 #define RTE_MEMBER_LOOKUP_BULK_MAX 64
64 /** Entry count per bucket in hash table based mode. */
65 #define RTE_MEMBER_BUCKET_ENTRIES 16
66 /** Maximum number of characters in setsum name. */
67 #define RTE_MEMBER_NAMESIZE 32
69 /** @internal Hash function used by membership library. */
70 #if defined(RTE_ARCH_X86) || defined(__ARM_FEATURE_CRC32)
71 #include <rte_hash_crc.h>
72 #define MEMBER_HASH_FUNC rte_hash_crc
74 #include <rte_jhash.h>
75 #define MEMBER_HASH_FUNC rte_jhash
78 extern int librte_member_logtype;
80 #define RTE_MEMBER_LOG(level, ...) \
81 rte_log(RTE_LOG_ ## level, \
82 librte_member_logtype, \
83 RTE_FMT("%s(): " RTE_FMT_HEAD(__VA_ARGS__,), \
85 RTE_FMT_TAIL(__VA_ARGS__,)))
87 /** @internal setsummary structure. */
88 struct rte_member_setsum;
92 * @b EXPERIMENTAL: this API may change without prior notice
94 * Parameter struct used to create set summary
96 struct rte_member_parameters;
100 * @b EXPERIMENTAL: this API may change without prior notice
102 * Define different set summary types
104 enum rte_member_setsum_type {
105 RTE_MEMBER_TYPE_HT = 0, /**< Hash table based set summary. */
106 RTE_MEMBER_TYPE_VBF, /**< Vector of bloom filters. */
110 /** @internal compare function for different arch. */
111 enum rte_member_sig_compare_function {
112 RTE_MEMBER_COMPARE_SCALAR = 0,
113 RTE_MEMBER_COMPARE_AVX2,
114 RTE_MEMBER_COMPARE_NUM
117 /** @internal setsummary structure. */
118 struct rte_member_setsum {
119 enum rte_member_setsum_type type; /* Type of the set summary. */
120 uint32_t key_len; /* Length of key. */
121 uint32_t prim_hash_seed; /* Primary hash function seed. */
122 uint32_t sec_hash_seed; /* Secondary hash function seed. */
124 /* Hash table based. */
125 uint32_t bucket_cnt; /* Number of buckets. */
126 uint32_t bucket_mask; /* Bit mask to get bucket index. */
127 /* For runtime selecting AVX, scalar, etc for signature comparison. */
128 enum rte_member_sig_compare_function sig_cmp_fn;
129 uint8_t cache; /* If it is cache mode for ht based. */
131 /* Vector bloom filter. */
132 uint32_t num_set; /* Number of set (bf) in vbf. */
133 uint32_t bits; /* Number of bits in each bf. */
134 uint32_t bit_mask; /* Bit mask to get bit location in bf. */
135 uint32_t num_hashes; /* Number of hash values to index bf. */
137 uint32_t mul_shift; /* vbf internal variable used during bit test. */
138 uint32_t div_shift; /* vbf internal variable used during bit test. */
140 void *table; /* This is the handler of hash table or vBF array. */
143 /* Second cache line should start here. */
144 uint32_t socket_id; /* NUMA Socket ID for memory. */
145 char name[RTE_MEMBER_NAMESIZE]; /* Name of this set summary. */
146 } __rte_cache_aligned;
150 * @b EXPERIMENTAL: this API may change without prior notice
152 * Parameters used when create the set summary table. Currently user can
153 * specify two types of setsummary: HT based and vBF. For HT based, user can
154 * specify cache or non-cache mode. Here is a table to describe some differences
157 struct rte_member_parameters {
158 const char *name; /**< Name of the hash. */
161 * User to specify the type of the setsummary from one of
162 * rte_member_setsum_type.
164 * HT based setsummary is implemented like a hash table. User should use
165 * this type when there are many sets.
167 * vBF setsummary is a vector of bloom filters. It is used when number
168 * of sets is not big (less than 32 for current implementation).
170 enum rte_member_setsum_type type;
173 * is_cache is only used for HT based setsummary.
175 * If it is HT based setsummary, user to specify the subtype or mode
176 * of the setsummary. It could be cache, or non-cache mode.
177 * Set is_cache to be 1 if to use as cache mode.
179 * For cache mode, keys can be evicted out of the HT setsummary. Keys
180 * with the same signature and map to the same bucket
181 * will overwrite each other in the setsummary table.
182 * This mode is useful for the case that the set-summary only
183 * needs to keep record of the recently inserted keys. Both
184 * false-negative and false-positive could happen.
186 * For non-cache mode, keys cannot be evicted out of the cache. So for
187 * this mode the setsummary will become full eventually. Keys with the
188 * same signature but map to the same bucket will still occupy multiple
189 * entries. This mode does not give false-negative result.
194 * For HT setsummary, num_keys equals to the number of entries of the
195 * table. When the number of keys inserted in the HT setsummary
196 * approaches this number, eviction could happen. For cache mode,
197 * keys could be evicted out of the table. For non-cache mode, keys will
198 * be evicted to other buckets like cuckoo hash. The table will also
199 * likely to become full before the number of inserted keys equal to the
200 * total number of entries.
202 * For vBF, num_keys equal to the expected number of keys that will
203 * be inserted into the vBF. The implementation assumes the keys are
204 * evenly distributed to each BF in vBF. This is used to calculate the
205 * number of bits we need for each BF. User does not specify the size of
206 * each BF directly because the optimal size depends on the num_keys
207 * and false positive rate.
212 * The length of key is used for hash calculation. Since key is not
213 * stored in set-summary, large key does not require more memory space.
218 * num_set is only used for vBF, but not used for HT setsummary.
220 * num_set is equal to the number of BFs in vBF. For current
221 * implementation, it only supports 1,2,4,8,16,32 BFs in one vBF set
222 * summary. If other number of sets are needed, for example 5, the user
223 * should allocate the minimum available value that larger than 5,
229 * false_positive_rate is only used for vBF, but not used for HT
232 * For vBF, false_positive_rate is the user-defined false positive rate
233 * given expected number of inserted keys (num_keys). It is used to
234 * calculate the total number of bits for each BF, and the number of
235 * hash values used during lookup and insertion. For details please
236 * refer to vBF implementation and membership library documentation.
238 * For HT, This parameter is not directly set by users.
239 * HT setsummary's false positive rate is in the order of:
240 * false_pos = (1/bucket_count)*(1/2^16), since we use 16-bit signature.
241 * This is because two keys needs to map to same bucket and same
242 * signature to have a collision (false positive). bucket_count is equal
243 * to number of entries (num_keys) divided by entry count per bucket
244 * (RTE_MEMBER_BUCKET_ENTRIES). Thus, the false_positive_rate is not
245 * directly set by users for HT mode.
247 float false_positive_rate;
250 * We use two seeds to calculate two independent hashes for each key.
252 * For HT type, one hash is used as signature, and the other is used
253 * for bucket location.
254 * For vBF type, these two hashes and their combinations are used as
255 * hash locations to index the bit array.
257 uint32_t prim_hash_seed;
260 * The secondary seed should be a different value from the primary seed.
262 uint32_t sec_hash_seed;
264 int socket_id; /**< NUMA Socket ID for memory. */
269 * @b EXPERIMENTAL: this API may change without prior notice
271 * Find an existing set-summary and return a pointer to it.
274 * Name of the set-summary.
276 * Pointer to the set-summary or NULL if object not found
277 * with rte_errno set appropriately. Possible rte_errno values include:
278 * - ENOENT - value not available for return
280 struct rte_member_setsum *
281 rte_member_find_existing(const char *name);
285 * @b EXPERIMENTAL: this API may change without prior notice
287 * Create set-summary (SS).
290 * Parameters to initialize the setsummary.
292 * Return the pointer to the setsummary.
293 * Return value is NULL if the creation failed.
295 struct rte_member_setsum *
296 rte_member_create(const struct rte_member_parameters *params);
300 * @b EXPERIMENTAL: this API may change without prior notice
302 * Lookup key in set-summary (SS).
303 * Single key lookup and return as soon as the first match found
306 * Pointer of a setsummary.
308 * Pointer of the key to be looked up.
310 * Output the set id matches the key.
312 * Return 1 for found a match and 0 for not found a match.
315 rte_member_lookup(const struct rte_member_setsum *setsum, const void *key,
316 member_set_t *set_id);
320 * @b EXPERIMENTAL: this API may change without prior notice
322 * Lookup bulk of keys in set-summary (SS).
323 * Each key lookup returns as soon as the first match found
326 * Pointer of a setsummary.
328 * Pointer of the bulk of keys to be looked up.
330 * Number of keys that will be lookup.
332 * Output set ids for all the keys to this array.
333 * User should preallocate array that can contain all results, which size is
336 * The number of keys that found a match.
339 rte_member_lookup_bulk(const struct rte_member_setsum *setsum,
340 const void **keys, uint32_t num_keys,
341 member_set_t *set_ids);
345 * @b EXPERIMENTAL: this API may change without prior notice
347 * Lookup a key in set-summary (SS) for multiple matches.
348 * The key lookup will find all matched entries (multiple match).
349 * Note that for cache mode of HT, each key can have at most one match. This is
350 * because keys with same signature that maps to same bucket will overwrite
351 * each other. So multi-match lookup should be used for vBF and non-cache HT.
354 * Pointer of a set-summary.
356 * Pointer of the key that to be looked up.
357 * @param max_match_per_key
358 * User specified maximum number of matches for each key. The function returns
359 * as soon as this number of matches found for the key.
361 * Output set ids for all the matches of the key. User needs to preallocate
362 * the array that can contain max_match_per_key number of results.
364 * The number of matches that found for the key.
365 * For cache mode HT set-summary, the number should be at most 1.
368 rte_member_lookup_multi(const struct rte_member_setsum *setsum,
369 const void *key, uint32_t max_match_per_key,
370 member_set_t *set_id);
374 * @b EXPERIMENTAL: this API may change without prior notice
376 * Lookup a bulk of keys in set-summary (SS) for multiple matches each key.
377 * Each key lookup will find all matched entries (multiple match).
378 * Note that for cache mode HT, each key can have at most one match. So
379 * multi-match function is mainly used for vBF and non-cache mode HT.
382 * Pointer of a setsummary.
384 * Pointer of the keys to be looked up.
386 * The number of keys that will be lookup.
387 * @param max_match_per_key
388 * The possible maximum number of matches for each key.
390 * Output the number of matches for each key in an array.
392 * Return set ids for all the matches of all keys. Users pass in a
393 * preallocated 2D array with first dimension as key index and second
394 * dimension as match index. For example set_ids[bulk_size][max_match_per_key]
396 * The number of keys that found one or more matches in the set-summary.
399 rte_member_lookup_multi_bulk(const struct rte_member_setsum *setsum,
400 const void **keys, uint32_t num_keys,
401 uint32_t max_match_per_key,
402 uint32_t *match_count,
403 member_set_t *set_ids);
407 * @b EXPERIMENTAL: this API may change without prior notice
409 * Insert key into set-summary (SS).
412 * Pointer of a set-summary.
414 * Pointer of the key to be added.
416 * The set id associated with the key that needs to be added. Different mode
417 * supports different set_id ranges. 0 cannot be used as set_id since
418 * RTE_MEMBER_NO_MATCH by default is set as 0.
419 * For HT mode, the set_id has range as [1, 0x7FFF], MSB is reserved.
420 * For vBF mode the set id is limited by the num_set parameter when create
423 * HT (cache mode) and vBF should never fail unless the set_id is not in the
424 * valid range. In such case -EINVAL is returned.
425 * For HT (non-cache mode) it could fail with -ENOSPC error code when table is
427 * For success it returns different values for different modes to provide
428 * extra information for users.
429 * Return 0 for HT (cache mode) if the add does not cause
430 * eviction, return 1 otherwise. Return 0 for non-cache mode if success,
431 * -ENOSPC for full, and 1 if cuckoo eviction happens.
432 * Always returns 0 for vBF mode.
435 rte_member_add(const struct rte_member_setsum *setsum, const void *key,
436 member_set_t set_id);
440 * @b EXPERIMENTAL: this API may change without prior notice
442 * De-allocate memory used by set-summary.
445 * Pointer to the set summary.
448 rte_member_free(struct rte_member_setsum *setsum);
452 * @b EXPERIMENTAL: this API may change without prior notice
454 * Reset the set-summary tables. E.g. reset bits to be 0 in BF,
455 * reset set_id in each entry to be RTE_MEMBER_NO_MATCH in HT based SS.
458 * Pointer to the set-summary.
461 rte_member_reset(const struct rte_member_setsum *setsum);
465 * @b EXPERIMENTAL: this API may change without prior notice
467 * Delete items from the set-summary. Note that vBF does not support deletion
468 * in current implementation. For vBF, error code of -EINVAL will be returned.
471 * Pointer to the set-summary.
473 * Pointer of the key to be deleted.
475 * For HT mode, we need both key and its corresponding set_id to
476 * properly delete the key. Without set_id, we may delete other keys with the
479 * If no entry found to delete, an error code of -ENOENT could be returned.
482 rte_member_delete(const struct rte_member_setsum *setsum, const void *key,
483 member_set_t set_id);
489 #endif /* _RTE_MEMBER_H_ */