4 * Copyright(c) 2017 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37 * RTE Membership Library
39 * The Membership Library is an extension and generalization of a traditional
40 * filter (for example Bloom Filter and cuckoo filter) structure that has
41 * multiple usages in a variety of workloads and applications. The library is
42 * used to test if a key belongs to certain sets. Two types of such
43 * "set-summary" structures are implemented: hash-table based (HT) and vector
44 * bloom filter (vBF). For HT setsummary, two subtypes or modes are available,
45 * cache and non-cache modes. The table below summarize some properties of
46 * the different implementations.
49 * @b EXPERIMENTAL: this API may change without prior notice
54 * +==========+=====================+================+=========================+
55 * | type | vbf | HT-cache | HT-non-cache |
56 * +==========+=====================+==========================================+
57 * |structure | bloom-filter array | hash-table like without storing key |
58 * +----------+---------------------+------------------------------------------+
59 * |set id | limited by bf count | [1, 0x7fff] |
61 * +----------+---------------------+------------------------------------------+
62 * |usages & | small set range, | can delete, | cache most recent keys, |
63 * |properties| user-specified | big set range, | have both false-positive|
64 * | | false-positive rate,| small false | and false-negative |
65 * | | no deletion support.| positive depend| depend on table size, |
66 * | | | on table size, | automatic overwritten. |
67 * | | | new key does | |
68 * | | | not overwrite | |
69 * | | | existing key. | |
70 * +----------+---------------------+----------------+-------------------------+
74 #ifndef _RTE_MEMBER_H_
75 #define _RTE_MEMBER_H_
83 /** The set ID type that stored internally in hash table based set summary. */
84 typedef uint16_t member_set_t;
85 /** Invalid set ID used to mean no match found. */
86 #define RTE_MEMBER_NO_MATCH 0
87 /** Maximum size of hash table that can be created. */
88 #define RTE_MEMBER_ENTRIES_MAX (1 << 30)
89 /** Maximum number of keys that can be searched as a bulk */
90 #define RTE_MEMBER_LOOKUP_BULK_MAX 64
91 /** Entry count per bucket in hash table based mode. */
92 #define RTE_MEMBER_BUCKET_ENTRIES 16
93 /** Maximum number of characters in setsum name. */
94 #define RTE_MEMBER_NAMESIZE 32
96 /** @internal Hash function used by membership library. */
97 #if defined(RTE_ARCH_X86) || defined(RTE_MACHINE_CPUFLAG_CRC32)
98 #include <rte_hash_crc.h>
99 #define MEMBER_HASH_FUNC rte_hash_crc
101 #include <rte_jhash.h>
102 #define MEMBER_HASH_FUNC rte_jhash
105 extern int librte_member_logtype;
107 #define RTE_MEMBER_LOG(level, fmt, args...) \
108 rte_log(RTE_LOG_ ## level, librte_member_logtype, "%s(): " fmt, \
111 /** @internal setsummary structure. */
112 struct rte_member_setsum;
116 * @b EXPERIMENTAL: this API may change without prior notice
118 * Parameter struct used to create set summary
120 struct rte_member_parameters;
124 * @b EXPERIMENTAL: this API may change without prior notice
126 * Define different set summary types
128 enum rte_member_setsum_type {
129 RTE_MEMBER_TYPE_HT = 0, /**< Hash table based set summary. */
130 RTE_MEMBER_TYPE_VBF, /**< Vector of bloom filters. */
134 /** @internal compare function for different arch. */
135 enum rte_member_sig_compare_function {
136 RTE_MEMBER_COMPARE_SCALAR = 0,
137 RTE_MEMBER_COMPARE_NUM
140 /** @internal setsummary structure. */
141 struct rte_member_setsum {
142 enum rte_member_setsum_type type; /* Type of the set summary. */
143 uint32_t key_len; /* Length of key. */
144 uint32_t prim_hash_seed; /* Primary hash function seed. */
145 uint32_t sec_hash_seed; /* Secondary hash function seed. */
147 /* Hash table based. */
148 uint32_t bucket_cnt; /* Number of buckets. */
149 uint32_t bucket_mask; /* Bit mask to get bucket index. */
150 /* For runtime selecting AVX, scalar, etc for signature comparison. */
151 enum rte_member_sig_compare_function sig_cmp_fn;
152 uint8_t cache; /* If it is cache mode for ht based. */
154 /* Vector bloom filter. */
155 uint32_t num_set; /* Number of set (bf) in vbf. */
156 uint32_t bits; /* Number of bits in each bf. */
157 uint32_t bit_mask; /* Bit mask to get bit location in bf. */
158 uint32_t num_hashes; /* Number of hash values to index bf. */
160 uint32_t mul_shift; /* vbf internal variable used during bit test. */
161 uint32_t div_shift; /* vbf internal variable used during bit test. */
163 void *table; /* This is the handler of hash table or vBF array. */
166 /* Second cache line should start here. */
167 uint32_t socket_id; /* NUMA Socket ID for memory. */
168 char name[RTE_MEMBER_NAMESIZE]; /* Name of this set summary. */
169 } __rte_cache_aligned;
173 * @b EXPERIMENTAL: this API may change without prior notice
175 * Parameters used when create the set summary table. Currently user can
176 * specify two types of setsummary: HT based and vBF. For HT based, user can
177 * specify cache or non-cache mode. Here is a table to describe some differences
180 struct rte_member_parameters {
181 const char *name; /**< Name of the hash. */
184 * User to specify the type of the setsummary from one of
185 * rte_member_setsum_type.
187 * HT based setsummary is implemented like a hash table. User should use
188 * this type when there are many sets.
190 * vBF setsummary is a vector of bloom filters. It is used when number
191 * of sets is not big (less than 32 for current implementation).
193 enum rte_member_setsum_type type;
196 * is_cache is only used for HT based setsummary.
198 * If it is HT based setsummary, user to specify the subtype or mode
199 * of the setsummary. It could be cache, or non-cache mode.
200 * Set is_cache to be 1 if to use as cache mode.
202 * For cache mode, keys can be evicted out of the HT setsummary. Keys
203 * with the same signature and map to the same bucket
204 * will overwrite each other in the setsummary table.
205 * This mode is useful for the case that the set-summary only
206 * needs to keep record of the recently inserted keys. Both
207 * false-negative and false-positive could happen.
209 * For non-cache mode, keys cannot be evicted out of the cache. So for
210 * this mode the setsummary will become full eventually. Keys with the
211 * same signature but map to the same bucket will still occupy multiple
212 * entries. This mode does not give false-negative result.
217 * For HT setsummary, num_keys equals to the number of entries of the
218 * table. When the number of keys inserted in the HT setsummary
219 * approaches this number, eviction could happen. For cache mode,
220 * keys could be evicted out of the table. For non-cache mode, keys will
221 * be evicted to other buckets like cuckoo hash. The table will also
222 * likely to become full before the number of inserted keys equal to the
223 * total number of entries.
225 * For vBF, num_keys equal to the expected number of keys that will
226 * be inserted into the vBF. The implementation assumes the keys are
227 * evenly distributed to each BF in vBF. This is used to calculate the
228 * number of bits we need for each BF. User does not specify the size of
229 * each BF directly because the optimal size depends on the num_keys
230 * and false positive rate.
235 * The length of key is used for hash calculation. Since key is not
236 * stored in set-summary, large key does not require more memory space.
241 * num_set is only used for vBF, but not used for HT setsummary.
243 * num_set is equal to the number of BFs in vBF. For current
244 * implementation, it only supports 1,2,4,8,16,32 BFs in one vBF set
245 * summary. If other number of sets are needed, for example 5, the user
246 * should allocate the minimum available value that larger than 5,
252 * false_positive_rate is only used for vBF, but not used for HT
255 * For vBF, false_positive_rate is the user-defined false positive rate
256 * given expected number of inserted keys (num_keys). It is used to
257 * calculate the total number of bits for each BF, and the number of
258 * hash values used during lookup and insertion. For details please
259 * refer to vBF implementation and membership library documentation.
261 * For HT, This parameter is not directly set by users.
262 * HT setsummary's false positive rate is in the order of:
263 * false_pos = (1/bucket_count)*(1/2^16), since we use 16-bit signature.
264 * This is because two keys needs to map to same bucket and same
265 * signature to have a collision (false positive). bucket_count is equal
266 * to number of entries (num_keys) divided by entry count per bucket
267 * (RTE_MEMBER_BUCKET_ENTRIES). Thus, the false_positive_rate is not
268 * directly set by users for HT mode.
270 float false_positive_rate;
273 * We use two seeds to calculate two independent hashes for each key.
275 * For HT type, one hash is used as signature, and the other is used
276 * for bucket location.
277 * For vBF type, these two hashes and their combinations are used as
278 * hash locations to index the bit array.
280 uint32_t prim_hash_seed;
283 * The secondary seed should be a different value from the primary seed.
285 uint32_t sec_hash_seed;
287 int socket_id; /**< NUMA Socket ID for memory. */
292 * @b EXPERIMENTAL: this API may change without prior notice
294 * Find an existing set-summary and return a pointer to it.
297 * Name of the set-summary.
299 * Pointer to the set-summary or NULL if object not found
300 * with rte_errno set appropriately. Possible rte_errno values include:
301 * - ENOENT - value not available for return
303 struct rte_member_setsum *
304 rte_member_find_existing(const char *name);
308 * @b EXPERIMENTAL: this API may change without prior notice
310 * Create set-summary (SS).
313 * Parameters to initialize the setsummary.
315 * Return the pointer to the setsummary.
316 * Return value is NULL if the creation failed.
318 struct rte_member_setsum *
319 rte_member_create(const struct rte_member_parameters *params);
323 * @b EXPERIMENTAL: this API may change without prior notice
325 * Lookup key in set-summary (SS).
326 * Single key lookup and return as soon as the first match found
329 * Pointer of a setsummary.
331 * Pointer of the key to be looked up.
333 * Output the set id matches the key.
335 * Return 1 for found a match and 0 for not found a match.
338 rte_member_lookup(const struct rte_member_setsum *setsum, const void *key,
339 member_set_t *set_id);
343 * @b EXPERIMENTAL: this API may change without prior notice
345 * Lookup bulk of keys in set-summary (SS).
346 * Each key lookup returns as soon as the first match found
349 * Pointer of a setsummary.
351 * Pointer of the bulk of keys to be looked up.
353 * Number of keys that will be lookup.
355 * Output set ids for all the keys to this array.
356 * User should preallocate array that can contain all results, which size is
359 * The number of keys that found a match.
362 rte_member_lookup_bulk(const struct rte_member_setsum *setsum,
363 const void **keys, uint32_t num_keys,
364 member_set_t *set_ids);
368 * @b EXPERIMENTAL: this API may change without prior notice
370 * Lookup a key in set-summary (SS) for multiple matches.
371 * The key lookup will find all matched entries (multiple match).
372 * Note that for cache mode of HT, each key can have at most one match. This is
373 * because keys with same signature that maps to same bucket will overwrite
374 * each other. So multi-match lookup should be used for vBF and non-cache HT.
377 * Pointer of a set-summary.
379 * Pointer of the key that to be looked up.
380 * @param max_match_per_key
381 * User specified maximum number of matches for each key. The function returns
382 * as soon as this number of matches found for the key.
384 * Output set ids for all the matches of the key. User needs to preallocate
385 * the array that can contain max_match_per_key number of results.
387 * The number of matches that found for the key.
388 * For cache mode HT set-summary, the number should be at most 1.
391 rte_member_lookup_multi(const struct rte_member_setsum *setsum,
392 const void *key, uint32_t max_match_per_key,
393 member_set_t *set_id);
397 * @b EXPERIMENTAL: this API may change without prior notice
399 * Lookup a bulk of keys in set-summary (SS) for multiple matches each key.
400 * Each key lookup will find all matched entries (multiple match).
401 * Note that for cache mode HT, each key can have at most one match. So
402 * multi-match function is mainly used for vBF and non-cache mode HT.
405 * Pointer of a setsummary.
407 * Pointer of the keys to be looked up.
409 * The number of keys that will be lookup.
410 * @param max_match_per_key
411 * The possible maximum number of matches for each key.
413 * Output the number of matches for each key in an array.
415 * Return set ids for all the matches of all keys. Users pass in a
416 * preallocated 2D array with first dimension as key index and second
417 * dimension as match index. For example set_ids[bulk_size][max_match_per_key]
419 * The number of keys that found one or more matches in the set-summary.
422 rte_member_lookup_multi_bulk(const struct rte_member_setsum *setsum,
423 const void **keys, uint32_t num_keys,
424 uint32_t max_match_per_key,
425 uint32_t *match_count,
426 member_set_t *set_ids);
430 * @b EXPERIMENTAL: this API may change without prior notice
432 * Insert key into set-summary (SS).
435 * Pointer of a set-summary.
437 * Pointer of the key to be added.
439 * The set id associated with the key that needs to be added. Different mode
440 * supports different set_id ranges. 0 cannot be used as set_id since
441 * RTE_MEMBER_NO_MATCH by default is set as 0.
442 * For HT mode, the set_id has range as [1, 0x7FFF], MSB is reserved.
443 * For vBF mode the set id is limited by the num_set parameter when create
446 * HT (cache mode) and vBF should never fail unless the set_id is not in the
447 * valid range. In such case -EINVAL is returned.
448 * For HT (non-cache mode) it could fail with -ENOSPC error code when table is
450 * For success it returns different values for different modes to provide
451 * extra information for users.
452 * Return 0 for HT (cache mode) if the add does not cause
453 * eviction, return 1 otherwise. Return 0 for non-cache mode if success,
454 * -ENOSPC for full, and 1 if cuckoo eviction happens.
455 * Always returns 0 for vBF mode.
458 rte_member_add(const struct rte_member_setsum *setsum, const void *key,
459 member_set_t set_id);
463 * @b EXPERIMENTAL: this API may change without prior notice
465 * De-allocate memory used by set-summary.
468 * Pointer to the set summary.
471 rte_member_free(struct rte_member_setsum *setsum);
475 * @b EXPERIMENTAL: this API may change without prior notice
477 * Reset the set-summary tables. E.g. reset bits to be 0 in BF,
478 * reset set_id in each entry to be RTE_MEMBER_NO_MATCH in HT based SS.
481 * Pointer to the set-summary.
484 rte_member_reset(const struct rte_member_setsum *setsum);
488 * @b EXPERIMENTAL: this API may change without prior notice
490 * Delete items from the set-summary. Note that vBF does not support deletion
491 * in current implementation. For vBF, error code of -EINVAL will be returned.
494 * Pointer to the set-summary.
496 * Pointer of the key to be deleted.
498 * For HT mode, we need both key and its corresponding set_id to
499 * properly delete the key. Without set_id, we may delete other keys with the
502 * If no entry found to delete, an error code of -ENOENT could be returned.
505 rte_member_delete(const struct rte_member_setsum *setsum, const void *key,
506 member_set_t set_id);
512 #endif /* _RTE_MEMBER_H_ */