hash: fix gcc 10 maybe-uninitialized warning
[dpdk.git] / lib / librte_hash / rte_cuckoo_hash.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2016 Intel Corporation
3  * Copyright(c) 2018 Arm Limited
4  */
5
6 #include <string.h>
7 #include <stdint.h>
8 #include <errno.h>
9 #include <stdio.h>
10 #include <stdarg.h>
11 #include <sys/queue.h>
12
13 #include <rte_common.h>
14 #include <rte_memory.h>         /* for definition of RTE_CACHE_LINE_SIZE */
15 #include <rte_log.h>
16 #include <rte_prefetch.h>
17 #include <rte_branch_prediction.h>
18 #include <rte_malloc.h>
19 #include <rte_eal.h>
20 #include <rte_eal_memconfig.h>
21 #include <rte_per_lcore.h>
22 #include <rte_errno.h>
23 #include <rte_string_fns.h>
24 #include <rte_cpuflags.h>
25 #include <rte_rwlock.h>
26 #include <rte_spinlock.h>
27 #include <rte_ring_elem.h>
28 #include <rte_compat.h>
29 #include <rte_vect.h>
30 #include <rte_tailq.h>
31
32 #include "rte_hash.h"
33 #include "rte_cuckoo_hash.h"
34
35 #define FOR_EACH_BUCKET(CURRENT_BKT, START_BUCKET)                            \
36         for (CURRENT_BKT = START_BUCKET;                                      \
37                 CURRENT_BKT != NULL;                                          \
38                 CURRENT_BKT = CURRENT_BKT->next)
39
40 TAILQ_HEAD(rte_hash_list, rte_tailq_entry);
41
42 static struct rte_tailq_elem rte_hash_tailq = {
43         .name = "RTE_HASH",
44 };
45 EAL_REGISTER_TAILQ(rte_hash_tailq)
46
47 struct rte_hash *
48 rte_hash_find_existing(const char *name)
49 {
50         struct rte_hash *h = NULL;
51         struct rte_tailq_entry *te;
52         struct rte_hash_list *hash_list;
53
54         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
55
56         rte_mcfg_tailq_read_lock();
57         TAILQ_FOREACH(te, hash_list, next) {
58                 h = (struct rte_hash *) te->data;
59                 if (strncmp(name, h->name, RTE_HASH_NAMESIZE) == 0)
60                         break;
61         }
62         rte_mcfg_tailq_read_unlock();
63
64         if (te == NULL) {
65                 rte_errno = ENOENT;
66                 return NULL;
67         }
68         return h;
69 }
70
71 static inline struct rte_hash_bucket *
72 rte_hash_get_last_bkt(struct rte_hash_bucket *lst_bkt)
73 {
74         while (lst_bkt->next != NULL)
75                 lst_bkt = lst_bkt->next;
76         return lst_bkt;
77 }
78
79 void rte_hash_set_cmp_func(struct rte_hash *h, rte_hash_cmp_eq_t func)
80 {
81         h->cmp_jump_table_idx = KEY_CUSTOM;
82         h->rte_hash_custom_cmp_eq = func;
83 }
84
85 static inline int
86 rte_hash_cmp_eq(const void *key1, const void *key2, const struct rte_hash *h)
87 {
88         if (h->cmp_jump_table_idx == KEY_CUSTOM)
89                 return h->rte_hash_custom_cmp_eq(key1, key2, h->key_len);
90         else
91                 return cmp_jump_table[h->cmp_jump_table_idx](key1, key2, h->key_len);
92 }
93
94 /*
95  * We use higher 16 bits of hash as the signature value stored in table.
96  * We use the lower bits for the primary bucket
97  * location. Then we XOR primary bucket location and the signature
98  * to get the secondary bucket location. This is same as
99  * proposed in Bin Fan, et al's paper
100  * "MemC3: Compact and Concurrent MemCache with Dumber Caching and
101  * Smarter Hashing". The benefit to use
102  * XOR is that one could derive the alternative bucket location
103  * by only using the current bucket location and the signature.
104  */
105 static inline uint16_t
106 get_short_sig(const hash_sig_t hash)
107 {
108         return hash >> 16;
109 }
110
111 static inline uint32_t
112 get_prim_bucket_index(const struct rte_hash *h, const hash_sig_t hash)
113 {
114         return hash & h->bucket_bitmask;
115 }
116
117 static inline uint32_t
118 get_alt_bucket_index(const struct rte_hash *h,
119                         uint32_t cur_bkt_idx, uint16_t sig)
120 {
121         return (cur_bkt_idx ^ sig) & h->bucket_bitmask;
122 }
123
124 struct rte_hash *
125 rte_hash_create(const struct rte_hash_parameters *params)
126 {
127         struct rte_hash *h = NULL;
128         struct rte_tailq_entry *te = NULL;
129         struct rte_hash_list *hash_list;
130         struct rte_ring *r = NULL;
131         struct rte_ring *r_ext = NULL;
132         char hash_name[RTE_HASH_NAMESIZE];
133         void *k = NULL;
134         void *buckets = NULL;
135         void *buckets_ext = NULL;
136         char ring_name[RTE_RING_NAMESIZE];
137         char ext_ring_name[RTE_RING_NAMESIZE];
138         unsigned num_key_slots;
139         unsigned int hw_trans_mem_support = 0, use_local_cache = 0;
140         unsigned int ext_table_support = 0;
141         unsigned int readwrite_concur_support = 0;
142         unsigned int writer_takes_lock = 0;
143         unsigned int no_free_on_del = 0;
144         uint32_t *ext_bkt_to_free = NULL;
145         uint32_t *tbl_chng_cnt = NULL;
146         unsigned int readwrite_concur_lf_support = 0;
147         uint32_t i;
148
149         rte_hash_function default_hash_func = (rte_hash_function)rte_jhash;
150
151         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
152
153         if (params == NULL) {
154                 RTE_LOG(ERR, HASH, "rte_hash_create has no parameters\n");
155                 return NULL;
156         }
157
158         /* Check for valid parameters */
159         if ((params->entries > RTE_HASH_ENTRIES_MAX) ||
160                         (params->entries < RTE_HASH_BUCKET_ENTRIES) ||
161                         (params->key_len == 0)) {
162                 rte_errno = EINVAL;
163                 RTE_LOG(ERR, HASH, "rte_hash_create has invalid parameters\n");
164                 return NULL;
165         }
166
167         /* Validate correct usage of extra options */
168         if ((params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) &&
169             (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY_LF)) {
170                 rte_errno = EINVAL;
171                 RTE_LOG(ERR, HASH, "rte_hash_create: choose rw concurrency or "
172                         "rw concurrency lock free\n");
173                 return NULL;
174         }
175
176         /* Check extra flags field to check extra options. */
177         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT)
178                 hw_trans_mem_support = 1;
179
180         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
181                 use_local_cache = 1;
182                 writer_takes_lock = 1;
183         }
184
185         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) {
186                 readwrite_concur_support = 1;
187                 writer_takes_lock = 1;
188         }
189
190         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_EXT_TABLE)
191                 ext_table_support = 1;
192
193         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_NO_FREE_ON_DEL)
194                 no_free_on_del = 1;
195
196         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY_LF) {
197                 readwrite_concur_lf_support = 1;
198                 /* Enable not freeing internal memory/index on delete */
199                 no_free_on_del = 1;
200         }
201
202         /* Store all keys and leave the first entry as a dummy entry for lookup_bulk */
203         if (use_local_cache)
204                 /*
205                  * Increase number of slots by total number of indices
206                  * that can be stored in the lcore caches
207                  * except for the first cache
208                  */
209                 num_key_slots = params->entries + (RTE_MAX_LCORE - 1) *
210                                         (LCORE_CACHE_SIZE - 1) + 1;
211         else
212                 num_key_slots = params->entries + 1;
213
214         snprintf(ring_name, sizeof(ring_name), "HT_%s", params->name);
215         /* Create ring (Dummy slot index is not enqueued) */
216         r = rte_ring_create_elem(ring_name, sizeof(uint32_t),
217                         rte_align32pow2(num_key_slots), params->socket_id, 0);
218         if (r == NULL) {
219                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
220                 goto err;
221         }
222
223         const uint32_t num_buckets = rte_align32pow2(params->entries) /
224                                                 RTE_HASH_BUCKET_ENTRIES;
225
226         /* Create ring for extendable buckets. */
227         if (ext_table_support) {
228                 snprintf(ext_ring_name, sizeof(ext_ring_name), "HT_EXT_%s",
229                                                                 params->name);
230                 r_ext = rte_ring_create_elem(ext_ring_name, sizeof(uint32_t),
231                                 rte_align32pow2(num_buckets + 1),
232                                 params->socket_id, 0);
233
234                 if (r_ext == NULL) {
235                         RTE_LOG(ERR, HASH, "ext buckets memory allocation "
236                                                                 "failed\n");
237                         goto err;
238                 }
239         }
240
241         snprintf(hash_name, sizeof(hash_name), "HT_%s", params->name);
242
243         rte_mcfg_tailq_write_lock();
244
245         /* guarantee there's no existing: this is normally already checked
246          * by ring creation above */
247         TAILQ_FOREACH(te, hash_list, next) {
248                 h = (struct rte_hash *) te->data;
249                 if (strncmp(params->name, h->name, RTE_HASH_NAMESIZE) == 0)
250                         break;
251         }
252         h = NULL;
253         if (te != NULL) {
254                 rte_errno = EEXIST;
255                 te = NULL;
256                 goto err_unlock;
257         }
258
259         te = rte_zmalloc("HASH_TAILQ_ENTRY", sizeof(*te), 0);
260         if (te == NULL) {
261                 RTE_LOG(ERR, HASH, "tailq entry allocation failed\n");
262                 goto err_unlock;
263         }
264
265         h = (struct rte_hash *)rte_zmalloc_socket(hash_name, sizeof(struct rte_hash),
266                                         RTE_CACHE_LINE_SIZE, params->socket_id);
267
268         if (h == NULL) {
269                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
270                 goto err_unlock;
271         }
272
273         buckets = rte_zmalloc_socket(NULL,
274                                 num_buckets * sizeof(struct rte_hash_bucket),
275                                 RTE_CACHE_LINE_SIZE, params->socket_id);
276
277         if (buckets == NULL) {
278                 RTE_LOG(ERR, HASH, "buckets memory allocation failed\n");
279                 goto err_unlock;
280         }
281
282         /* Allocate same number of extendable buckets */
283         if (ext_table_support) {
284                 buckets_ext = rte_zmalloc_socket(NULL,
285                                 num_buckets * sizeof(struct rte_hash_bucket),
286                                 RTE_CACHE_LINE_SIZE, params->socket_id);
287                 if (buckets_ext == NULL) {
288                         RTE_LOG(ERR, HASH, "ext buckets memory allocation "
289                                                         "failed\n");
290                         goto err_unlock;
291                 }
292                 /* Populate ext bkt ring. We reserve 0 similar to the
293                  * key-data slot, just in case in future we want to
294                  * use bucket index for the linked list and 0 means NULL
295                  * for next bucket
296                  */
297                 for (i = 1; i <= num_buckets; i++)
298                         rte_ring_sp_enqueue_elem(r_ext, &i, sizeof(uint32_t));
299
300                 if (readwrite_concur_lf_support) {
301                         ext_bkt_to_free = rte_zmalloc(NULL, sizeof(uint32_t) *
302                                                                 num_key_slots, 0);
303                         if (ext_bkt_to_free == NULL) {
304                                 RTE_LOG(ERR, HASH, "ext bkt to free memory allocation "
305                                                                 "failed\n");
306                                 goto err_unlock;
307                         }
308                 }
309         }
310
311         const uint32_t key_entry_size =
312                 RTE_ALIGN(sizeof(struct rte_hash_key) + params->key_len,
313                           KEY_ALIGNMENT);
314         const uint64_t key_tbl_size = (uint64_t) key_entry_size * num_key_slots;
315
316         k = rte_zmalloc_socket(NULL, key_tbl_size,
317                         RTE_CACHE_LINE_SIZE, params->socket_id);
318
319         if (k == NULL) {
320                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
321                 goto err_unlock;
322         }
323
324         tbl_chng_cnt = rte_zmalloc_socket(NULL, sizeof(uint32_t),
325                         RTE_CACHE_LINE_SIZE, params->socket_id);
326
327         if (tbl_chng_cnt == NULL) {
328                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
329                 goto err_unlock;
330         }
331
332 /*
333  * If x86 architecture is used, select appropriate compare function,
334  * which may use x86 intrinsics, otherwise use memcmp
335  */
336 #if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
337         /* Select function to compare keys */
338         switch (params->key_len) {
339         case 16:
340                 h->cmp_jump_table_idx = KEY_16_BYTES;
341                 break;
342         case 32:
343                 h->cmp_jump_table_idx = KEY_32_BYTES;
344                 break;
345         case 48:
346                 h->cmp_jump_table_idx = KEY_48_BYTES;
347                 break;
348         case 64:
349                 h->cmp_jump_table_idx = KEY_64_BYTES;
350                 break;
351         case 80:
352                 h->cmp_jump_table_idx = KEY_80_BYTES;
353                 break;
354         case 96:
355                 h->cmp_jump_table_idx = KEY_96_BYTES;
356                 break;
357         case 112:
358                 h->cmp_jump_table_idx = KEY_112_BYTES;
359                 break;
360         case 128:
361                 h->cmp_jump_table_idx = KEY_128_BYTES;
362                 break;
363         default:
364                 /* If key is not multiple of 16, use generic memcmp */
365                 h->cmp_jump_table_idx = KEY_OTHER_BYTES;
366         }
367 #else
368         h->cmp_jump_table_idx = KEY_OTHER_BYTES;
369 #endif
370
371         if (use_local_cache) {
372                 h->local_free_slots = rte_zmalloc_socket(NULL,
373                                 sizeof(struct lcore_cache) * RTE_MAX_LCORE,
374                                 RTE_CACHE_LINE_SIZE, params->socket_id);
375         }
376
377         /* Default hash function */
378 #if defined(RTE_ARCH_X86)
379         default_hash_func = (rte_hash_function)rte_hash_crc;
380 #elif defined(RTE_ARCH_ARM64)
381         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_CRC32))
382                 default_hash_func = (rte_hash_function)rte_hash_crc;
383 #endif
384         /* Setup hash context */
385         strlcpy(h->name, params->name, sizeof(h->name));
386         h->entries = params->entries;
387         h->key_len = params->key_len;
388         h->key_entry_size = key_entry_size;
389         h->hash_func_init_val = params->hash_func_init_val;
390
391         h->num_buckets = num_buckets;
392         h->bucket_bitmask = h->num_buckets - 1;
393         h->buckets = buckets;
394         h->buckets_ext = buckets_ext;
395         h->free_ext_bkts = r_ext;
396         h->hash_func = (params->hash_func == NULL) ?
397                 default_hash_func : params->hash_func;
398         h->key_store = k;
399         h->free_slots = r;
400         h->ext_bkt_to_free = ext_bkt_to_free;
401         h->tbl_chng_cnt = tbl_chng_cnt;
402         *h->tbl_chng_cnt = 0;
403         h->hw_trans_mem_support = hw_trans_mem_support;
404         h->use_local_cache = use_local_cache;
405         h->readwrite_concur_support = readwrite_concur_support;
406         h->ext_table_support = ext_table_support;
407         h->writer_takes_lock = writer_takes_lock;
408         h->no_free_on_del = no_free_on_del;
409         h->readwrite_concur_lf_support = readwrite_concur_lf_support;
410
411 #if defined(RTE_ARCH_X86)
412         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE2))
413                 h->sig_cmp_fn = RTE_HASH_COMPARE_SSE;
414         else
415 #elif defined(RTE_ARCH_ARM64)
416         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
417                 h->sig_cmp_fn = RTE_HASH_COMPARE_NEON;
418         else
419 #endif
420                 h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
421
422         /* Writer threads need to take the lock when:
423          * 1) RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY is enabled OR
424          * 2) RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD is enabled
425          */
426         if (h->writer_takes_lock) {
427                 h->readwrite_lock = rte_malloc(NULL, sizeof(rte_rwlock_t),
428                                                 RTE_CACHE_LINE_SIZE);
429                 if (h->readwrite_lock == NULL)
430                         goto err_unlock;
431
432                 rte_rwlock_init(h->readwrite_lock);
433         }
434
435         /* Populate free slots ring. Entry zero is reserved for key misses. */
436         for (i = 1; i < num_key_slots; i++)
437                 rte_ring_sp_enqueue_elem(r, &i, sizeof(uint32_t));
438
439         te->data = (void *) h;
440         TAILQ_INSERT_TAIL(hash_list, te, next);
441         rte_mcfg_tailq_write_unlock();
442
443         return h;
444 err_unlock:
445         rte_mcfg_tailq_write_unlock();
446 err:
447         rte_ring_free(r);
448         rte_ring_free(r_ext);
449         rte_free(te);
450         rte_free(h);
451         rte_free(buckets);
452         rte_free(buckets_ext);
453         rte_free(k);
454         rte_free(tbl_chng_cnt);
455         rte_free(ext_bkt_to_free);
456         return NULL;
457 }
458
459 void
460 rte_hash_free(struct rte_hash *h)
461 {
462         struct rte_tailq_entry *te;
463         struct rte_hash_list *hash_list;
464
465         if (h == NULL)
466                 return;
467
468         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
469
470         rte_mcfg_tailq_write_lock();
471
472         /* find out tailq entry */
473         TAILQ_FOREACH(te, hash_list, next) {
474                 if (te->data == (void *) h)
475                         break;
476         }
477
478         if (te == NULL) {
479                 rte_mcfg_tailq_write_unlock();
480                 return;
481         }
482
483         TAILQ_REMOVE(hash_list, te, next);
484
485         rte_mcfg_tailq_write_unlock();
486
487         if (h->use_local_cache)
488                 rte_free(h->local_free_slots);
489         if (h->writer_takes_lock)
490                 rte_free(h->readwrite_lock);
491         rte_ring_free(h->free_slots);
492         rte_ring_free(h->free_ext_bkts);
493         rte_free(h->key_store);
494         rte_free(h->buckets);
495         rte_free(h->buckets_ext);
496         rte_free(h->tbl_chng_cnt);
497         rte_free(h->ext_bkt_to_free);
498         rte_free(h);
499         rte_free(te);
500 }
501
502 hash_sig_t
503 rte_hash_hash(const struct rte_hash *h, const void *key)
504 {
505         /* calc hash result by key */
506         return h->hash_func(key, h->key_len, h->hash_func_init_val);
507 }
508
509 int32_t
510 rte_hash_max_key_id(const struct rte_hash *h)
511 {
512         RETURN_IF_TRUE((h == NULL), -EINVAL);
513         if (h->use_local_cache)
514                 /*
515                  * Increase number of slots by total number of indices
516                  * that can be stored in the lcore caches
517                  */
518                 return (h->entries + ((RTE_MAX_LCORE - 1) *
519                                         (LCORE_CACHE_SIZE - 1)));
520         else
521                 return h->entries;
522 }
523
524 int32_t
525 rte_hash_count(const struct rte_hash *h)
526 {
527         uint32_t tot_ring_cnt, cached_cnt = 0;
528         uint32_t i, ret;
529
530         if (h == NULL)
531                 return -EINVAL;
532
533         if (h->use_local_cache) {
534                 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
535                                         (LCORE_CACHE_SIZE - 1);
536                 for (i = 0; i < RTE_MAX_LCORE; i++)
537                         cached_cnt += h->local_free_slots[i].len;
538
539                 ret = tot_ring_cnt - rte_ring_count(h->free_slots) -
540                                                                 cached_cnt;
541         } else {
542                 tot_ring_cnt = h->entries;
543                 ret = tot_ring_cnt - rte_ring_count(h->free_slots);
544         }
545         return ret;
546 }
547
548 /* Read write locks implemented using rte_rwlock */
549 static inline void
550 __hash_rw_writer_lock(const struct rte_hash *h)
551 {
552         if (h->writer_takes_lock && h->hw_trans_mem_support)
553                 rte_rwlock_write_lock_tm(h->readwrite_lock);
554         else if (h->writer_takes_lock)
555                 rte_rwlock_write_lock(h->readwrite_lock);
556 }
557
558 static inline void
559 __hash_rw_reader_lock(const struct rte_hash *h)
560 {
561         if (h->readwrite_concur_support && h->hw_trans_mem_support)
562                 rte_rwlock_read_lock_tm(h->readwrite_lock);
563         else if (h->readwrite_concur_support)
564                 rte_rwlock_read_lock(h->readwrite_lock);
565 }
566
567 static inline void
568 __hash_rw_writer_unlock(const struct rte_hash *h)
569 {
570         if (h->writer_takes_lock && h->hw_trans_mem_support)
571                 rte_rwlock_write_unlock_tm(h->readwrite_lock);
572         else if (h->writer_takes_lock)
573                 rte_rwlock_write_unlock(h->readwrite_lock);
574 }
575
576 static inline void
577 __hash_rw_reader_unlock(const struct rte_hash *h)
578 {
579         if (h->readwrite_concur_support && h->hw_trans_mem_support)
580                 rte_rwlock_read_unlock_tm(h->readwrite_lock);
581         else if (h->readwrite_concur_support)
582                 rte_rwlock_read_unlock(h->readwrite_lock);
583 }
584
585 void
586 rte_hash_reset(struct rte_hash *h)
587 {
588         uint32_t tot_ring_cnt, i;
589
590         if (h == NULL)
591                 return;
592
593         __hash_rw_writer_lock(h);
594         memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
595         memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
596         *h->tbl_chng_cnt = 0;
597
598         /* reset the free ring */
599         rte_ring_reset(h->free_slots);
600
601         /* flush free extendable bucket ring and memory */
602         if (h->ext_table_support) {
603                 memset(h->buckets_ext, 0, h->num_buckets *
604                                                 sizeof(struct rte_hash_bucket));
605                 rte_ring_reset(h->free_ext_bkts);
606         }
607
608         /* Repopulate the free slots ring. Entry zero is reserved for key misses */
609         if (h->use_local_cache)
610                 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
611                                         (LCORE_CACHE_SIZE - 1);
612         else
613                 tot_ring_cnt = h->entries;
614
615         for (i = 1; i < tot_ring_cnt + 1; i++)
616                 rte_ring_sp_enqueue_elem(h->free_slots, &i, sizeof(uint32_t));
617
618         /* Repopulate the free ext bkt ring. */
619         if (h->ext_table_support) {
620                 for (i = 1; i <= h->num_buckets; i++)
621                         rte_ring_sp_enqueue_elem(h->free_ext_bkts, &i,
622                                                         sizeof(uint32_t));
623         }
624
625         if (h->use_local_cache) {
626                 /* Reset local caches per lcore */
627                 for (i = 0; i < RTE_MAX_LCORE; i++)
628                         h->local_free_slots[i].len = 0;
629         }
630         __hash_rw_writer_unlock(h);
631 }
632
633 /*
634  * Function called to enqueue back an index in the cache/ring,
635  * as slot has not being used and it can be used in the
636  * next addition attempt.
637  */
638 static inline void
639 enqueue_slot_back(const struct rte_hash *h,
640                 struct lcore_cache *cached_free_slots,
641                 uint32_t slot_id)
642 {
643         if (h->use_local_cache) {
644                 cached_free_slots->objs[cached_free_slots->len] = slot_id;
645                 cached_free_slots->len++;
646         } else
647                 rte_ring_sp_enqueue_elem(h->free_slots, &slot_id,
648                                                 sizeof(uint32_t));
649 }
650
651 /* Search a key from bucket and update its data.
652  * Writer holds the lock before calling this.
653  */
654 static inline int32_t
655 search_and_update(const struct rte_hash *h, void *data, const void *key,
656         struct rte_hash_bucket *bkt, uint16_t sig)
657 {
658         int i;
659         struct rte_hash_key *k, *keys = h->key_store;
660
661         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
662                 if (bkt->sig_current[i] == sig) {
663                         k = (struct rte_hash_key *) ((char *)keys +
664                                         bkt->key_idx[i] * h->key_entry_size);
665                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
666                                 /* The store to application data at *data
667                                  * should not leak after the store to pdata
668                                  * in the key store. i.e. pdata is the guard
669                                  * variable. Release the application data
670                                  * to the readers.
671                                  */
672                                 __atomic_store_n(&k->pdata,
673                                         data,
674                                         __ATOMIC_RELEASE);
675                                 /*
676                                  * Return index where key is stored,
677                                  * subtracting the first dummy index
678                                  */
679                                 return bkt->key_idx[i] - 1;
680                         }
681                 }
682         }
683         return -1;
684 }
685
686 /* Only tries to insert at one bucket (@prim_bkt) without trying to push
687  * buckets around.
688  * return 1 if matching existing key, return 0 if succeeds, return -1 for no
689  * empty entry.
690  */
691 static inline int32_t
692 rte_hash_cuckoo_insert_mw(const struct rte_hash *h,
693                 struct rte_hash_bucket *prim_bkt,
694                 struct rte_hash_bucket *sec_bkt,
695                 const struct rte_hash_key *key, void *data,
696                 uint16_t sig, uint32_t new_idx,
697                 int32_t *ret_val)
698 {
699         unsigned int i;
700         struct rte_hash_bucket *cur_bkt;
701         int32_t ret;
702
703         __hash_rw_writer_lock(h);
704         /* Check if key was inserted after last check but before this
705          * protected region in case of inserting duplicated keys.
706          */
707         ret = search_and_update(h, data, key, prim_bkt, sig);
708         if (ret != -1) {
709                 __hash_rw_writer_unlock(h);
710                 *ret_val = ret;
711                 return 1;
712         }
713
714         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
715                 ret = search_and_update(h, data, key, cur_bkt, sig);
716                 if (ret != -1) {
717                         __hash_rw_writer_unlock(h);
718                         *ret_val = ret;
719                         return 1;
720                 }
721         }
722
723         /* Insert new entry if there is room in the primary
724          * bucket.
725          */
726         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
727                 /* Check if slot is available */
728                 if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
729                         prim_bkt->sig_current[i] = sig;
730                         /* Store to signature and key should not
731                          * leak after the store to key_idx. i.e.
732                          * key_idx is the guard variable for signature
733                          * and key.
734                          */
735                         __atomic_store_n(&prim_bkt->key_idx[i],
736                                          new_idx,
737                                          __ATOMIC_RELEASE);
738                         break;
739                 }
740         }
741         __hash_rw_writer_unlock(h);
742
743         if (i != RTE_HASH_BUCKET_ENTRIES)
744                 return 0;
745
746         /* no empty entry */
747         return -1;
748 }
749
750 /* Shift buckets along provided cuckoo_path (@leaf and @leaf_slot) and fill
751  * the path head with new entry (sig, alt_hash, new_idx)
752  * return 1 if matched key found, return -1 if cuckoo path invalided and fail,
753  * return 0 if succeeds.
754  */
755 static inline int
756 rte_hash_cuckoo_move_insert_mw(const struct rte_hash *h,
757                         struct rte_hash_bucket *bkt,
758                         struct rte_hash_bucket *alt_bkt,
759                         const struct rte_hash_key *key, void *data,
760                         struct queue_node *leaf, uint32_t leaf_slot,
761                         uint16_t sig, uint32_t new_idx,
762                         int32_t *ret_val)
763 {
764         uint32_t prev_alt_bkt_idx;
765         struct rte_hash_bucket *cur_bkt;
766         struct queue_node *prev_node, *curr_node = leaf;
767         struct rte_hash_bucket *prev_bkt, *curr_bkt = leaf->bkt;
768         uint32_t prev_slot, curr_slot = leaf_slot;
769         int32_t ret;
770
771         __hash_rw_writer_lock(h);
772
773         /* In case empty slot was gone before entering protected region */
774         if (curr_bkt->key_idx[curr_slot] != EMPTY_SLOT) {
775                 __hash_rw_writer_unlock(h);
776                 return -1;
777         }
778
779         /* Check if key was inserted after last check but before this
780          * protected region.
781          */
782         ret = search_and_update(h, data, key, bkt, sig);
783         if (ret != -1) {
784                 __hash_rw_writer_unlock(h);
785                 *ret_val = ret;
786                 return 1;
787         }
788
789         FOR_EACH_BUCKET(cur_bkt, alt_bkt) {
790                 ret = search_and_update(h, data, key, cur_bkt, sig);
791                 if (ret != -1) {
792                         __hash_rw_writer_unlock(h);
793                         *ret_val = ret;
794                         return 1;
795                 }
796         }
797
798         while (likely(curr_node->prev != NULL)) {
799                 prev_node = curr_node->prev;
800                 prev_bkt = prev_node->bkt;
801                 prev_slot = curr_node->prev_slot;
802
803                 prev_alt_bkt_idx = get_alt_bucket_index(h,
804                                         prev_node->cur_bkt_idx,
805                                         prev_bkt->sig_current[prev_slot]);
806
807                 if (unlikely(&h->buckets[prev_alt_bkt_idx]
808                                 != curr_bkt)) {
809                         /* revert it to empty, otherwise duplicated keys */
810                         __atomic_store_n(&curr_bkt->key_idx[curr_slot],
811                                 EMPTY_SLOT,
812                                 __ATOMIC_RELEASE);
813                         __hash_rw_writer_unlock(h);
814                         return -1;
815                 }
816
817                 if (h->readwrite_concur_lf_support) {
818                         /* Inform the previous move. The current move need
819                          * not be informed now as the current bucket entry
820                          * is present in both primary and secondary.
821                          * Since there is one writer, load acquires on
822                          * tbl_chng_cnt are not required.
823                          */
824                         __atomic_store_n(h->tbl_chng_cnt,
825                                          *h->tbl_chng_cnt + 1,
826                                          __ATOMIC_RELEASE);
827                         /* The store to sig_current should not
828                          * move above the store to tbl_chng_cnt.
829                          */
830                         __atomic_thread_fence(__ATOMIC_RELEASE);
831                 }
832
833                 /* Need to swap current/alt sig to allow later
834                  * Cuckoo insert to move elements back to its
835                  * primary bucket if available
836                  */
837                 curr_bkt->sig_current[curr_slot] =
838                         prev_bkt->sig_current[prev_slot];
839                 /* Release the updated bucket entry */
840                 __atomic_store_n(&curr_bkt->key_idx[curr_slot],
841                         prev_bkt->key_idx[prev_slot],
842                         __ATOMIC_RELEASE);
843
844                 curr_slot = prev_slot;
845                 curr_node = prev_node;
846                 curr_bkt = curr_node->bkt;
847         }
848
849         if (h->readwrite_concur_lf_support) {
850                 /* Inform the previous move. The current move need
851                  * not be informed now as the current bucket entry
852                  * is present in both primary and secondary.
853                  * Since there is one writer, load acquires on
854                  * tbl_chng_cnt are not required.
855                  */
856                 __atomic_store_n(h->tbl_chng_cnt,
857                                  *h->tbl_chng_cnt + 1,
858                                  __ATOMIC_RELEASE);
859                 /* The store to sig_current should not
860                  * move above the store to tbl_chng_cnt.
861                  */
862                 __atomic_thread_fence(__ATOMIC_RELEASE);
863         }
864
865         curr_bkt->sig_current[curr_slot] = sig;
866         /* Release the new bucket entry */
867         __atomic_store_n(&curr_bkt->key_idx[curr_slot],
868                          new_idx,
869                          __ATOMIC_RELEASE);
870
871         __hash_rw_writer_unlock(h);
872
873         return 0;
874
875 }
876
877 /*
878  * Make space for new key, using bfs Cuckoo Search and Multi-Writer safe
879  * Cuckoo
880  */
881 static inline int
882 rte_hash_cuckoo_make_space_mw(const struct rte_hash *h,
883                         struct rte_hash_bucket *bkt,
884                         struct rte_hash_bucket *sec_bkt,
885                         const struct rte_hash_key *key, void *data,
886                         uint16_t sig, uint32_t bucket_idx,
887                         uint32_t new_idx, int32_t *ret_val)
888 {
889         unsigned int i;
890         struct queue_node queue[RTE_HASH_BFS_QUEUE_MAX_LEN];
891         struct queue_node *tail, *head;
892         struct rte_hash_bucket *curr_bkt, *alt_bkt;
893         uint32_t cur_idx, alt_idx;
894
895         tail = queue;
896         head = queue + 1;
897         tail->bkt = bkt;
898         tail->prev = NULL;
899         tail->prev_slot = -1;
900         tail->cur_bkt_idx = bucket_idx;
901
902         /* Cuckoo bfs Search */
903         while (likely(tail != head && head <
904                                         queue + RTE_HASH_BFS_QUEUE_MAX_LEN -
905                                         RTE_HASH_BUCKET_ENTRIES)) {
906                 curr_bkt = tail->bkt;
907                 cur_idx = tail->cur_bkt_idx;
908                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
909                         if (curr_bkt->key_idx[i] == EMPTY_SLOT) {
910                                 int32_t ret = rte_hash_cuckoo_move_insert_mw(h,
911                                                 bkt, sec_bkt, key, data,
912                                                 tail, i, sig,
913                                                 new_idx, ret_val);
914                                 if (likely(ret != -1))
915                                         return ret;
916                         }
917
918                         /* Enqueue new node and keep prev node info */
919                         alt_idx = get_alt_bucket_index(h, cur_idx,
920                                                 curr_bkt->sig_current[i]);
921                         alt_bkt = &(h->buckets[alt_idx]);
922                         head->bkt = alt_bkt;
923                         head->cur_bkt_idx = alt_idx;
924                         head->prev = tail;
925                         head->prev_slot = i;
926                         head++;
927                 }
928                 tail++;
929         }
930
931         return -ENOSPC;
932 }
933
934 static inline int32_t
935 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
936                                                 hash_sig_t sig, void *data)
937 {
938         uint16_t short_sig;
939         uint32_t prim_bucket_idx, sec_bucket_idx;
940         struct rte_hash_bucket *prim_bkt, *sec_bkt, *cur_bkt;
941         struct rte_hash_key *new_k, *keys = h->key_store;
942         uint32_t ext_bkt_id = 0;
943         uint32_t slot_id;
944         int ret;
945         unsigned n_slots;
946         unsigned lcore_id;
947         unsigned int i;
948         struct lcore_cache *cached_free_slots = NULL;
949         int32_t ret_val;
950         struct rte_hash_bucket *last;
951
952         short_sig = get_short_sig(sig);
953         prim_bucket_idx = get_prim_bucket_index(h, sig);
954         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
955         prim_bkt = &h->buckets[prim_bucket_idx];
956         sec_bkt = &h->buckets[sec_bucket_idx];
957         rte_prefetch0(prim_bkt);
958         rte_prefetch0(sec_bkt);
959
960         /* Check if key is already inserted in primary location */
961         __hash_rw_writer_lock(h);
962         ret = search_and_update(h, data, key, prim_bkt, short_sig);
963         if (ret != -1) {
964                 __hash_rw_writer_unlock(h);
965                 return ret;
966         }
967
968         /* Check if key is already inserted in secondary location */
969         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
970                 ret = search_and_update(h, data, key, cur_bkt, short_sig);
971                 if (ret != -1) {
972                         __hash_rw_writer_unlock(h);
973                         return ret;
974                 }
975         }
976
977         __hash_rw_writer_unlock(h);
978
979         /* Did not find a match, so get a new slot for storing the new key */
980         if (h->use_local_cache) {
981                 lcore_id = rte_lcore_id();
982                 cached_free_slots = &h->local_free_slots[lcore_id];
983                 /* Try to get a free slot from the local cache */
984                 if (cached_free_slots->len == 0) {
985                         /* Need to get another burst of free slots from global ring */
986                         n_slots = rte_ring_mc_dequeue_burst_elem(h->free_slots,
987                                         cached_free_slots->objs,
988                                         sizeof(uint32_t),
989                                         LCORE_CACHE_SIZE, NULL);
990                         if (n_slots == 0) {
991                                 return -ENOSPC;
992                         }
993
994                         cached_free_slots->len += n_slots;
995                 }
996
997                 /* Get a free slot from the local cache */
998                 cached_free_slots->len--;
999                 slot_id = cached_free_slots->objs[cached_free_slots->len];
1000         } else {
1001                 if (rte_ring_sc_dequeue_elem(h->free_slots, &slot_id,
1002                                                 sizeof(uint32_t)) != 0) {
1003                         return -ENOSPC;
1004                 }
1005         }
1006
1007         new_k = RTE_PTR_ADD(keys, slot_id * h->key_entry_size);
1008         /* The store to application data (by the application) at *data should
1009          * not leak after the store of pdata in the key store. i.e. pdata is
1010          * the guard variable. Release the application data to the readers.
1011          */
1012         __atomic_store_n(&new_k->pdata,
1013                 data,
1014                 __ATOMIC_RELEASE);
1015         /* Copy key */
1016         memcpy(new_k->key, key, h->key_len);
1017
1018         /* Find an empty slot and insert */
1019         ret = rte_hash_cuckoo_insert_mw(h, prim_bkt, sec_bkt, key, data,
1020                                         short_sig, slot_id, &ret_val);
1021         if (ret == 0)
1022                 return slot_id - 1;
1023         else if (ret == 1) {
1024                 enqueue_slot_back(h, cached_free_slots, slot_id);
1025                 return ret_val;
1026         }
1027
1028         /* Primary bucket full, need to make space for new entry */
1029         ret = rte_hash_cuckoo_make_space_mw(h, prim_bkt, sec_bkt, key, data,
1030                                 short_sig, prim_bucket_idx, slot_id, &ret_val);
1031         if (ret == 0)
1032                 return slot_id - 1;
1033         else if (ret == 1) {
1034                 enqueue_slot_back(h, cached_free_slots, slot_id);
1035                 return ret_val;
1036         }
1037
1038         /* Also search secondary bucket to get better occupancy */
1039         ret = rte_hash_cuckoo_make_space_mw(h, sec_bkt, prim_bkt, key, data,
1040                                 short_sig, sec_bucket_idx, slot_id, &ret_val);
1041
1042         if (ret == 0)
1043                 return slot_id - 1;
1044         else if (ret == 1) {
1045                 enqueue_slot_back(h, cached_free_slots, slot_id);
1046                 return ret_val;
1047         }
1048
1049         /* if ext table not enabled, we failed the insertion */
1050         if (!h->ext_table_support) {
1051                 enqueue_slot_back(h, cached_free_slots, slot_id);
1052                 return ret;
1053         }
1054
1055         /* Now we need to go through the extendable bucket. Protection is needed
1056          * to protect all extendable bucket processes.
1057          */
1058         __hash_rw_writer_lock(h);
1059         /* We check for duplicates again since could be inserted before the lock */
1060         ret = search_and_update(h, data, key, prim_bkt, short_sig);
1061         if (ret != -1) {
1062                 enqueue_slot_back(h, cached_free_slots, slot_id);
1063                 goto failure;
1064         }
1065
1066         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1067                 ret = search_and_update(h, data, key, cur_bkt, short_sig);
1068                 if (ret != -1) {
1069                         enqueue_slot_back(h, cached_free_slots, slot_id);
1070                         goto failure;
1071                 }
1072         }
1073
1074         /* Search sec and ext buckets to find an empty entry to insert. */
1075         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1076                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1077                         /* Check if slot is available */
1078                         if (likely(cur_bkt->key_idx[i] == EMPTY_SLOT)) {
1079                                 cur_bkt->sig_current[i] = short_sig;
1080                                 /* Store to signature and key should not
1081                                  * leak after the store to key_idx. i.e.
1082                                  * key_idx is the guard variable for signature
1083                                  * and key.
1084                                  */
1085                                 __atomic_store_n(&cur_bkt->key_idx[i],
1086                                                  slot_id,
1087                                                  __ATOMIC_RELEASE);
1088                                 __hash_rw_writer_unlock(h);
1089                                 return slot_id - 1;
1090                         }
1091                 }
1092         }
1093
1094         /* Failed to get an empty entry from extendable buckets. Link a new
1095          * extendable bucket. We first get a free bucket from ring.
1096          */
1097         if (rte_ring_sc_dequeue_elem(h->free_ext_bkts, &ext_bkt_id,
1098                                                 sizeof(uint32_t)) != 0 ||
1099                                         ext_bkt_id == 0) {
1100                 ret = -ENOSPC;
1101                 goto failure;
1102         }
1103
1104         /* Use the first location of the new bucket */
1105         (h->buckets_ext[ext_bkt_id - 1]).sig_current[0] = short_sig;
1106         /* Store to signature and key should not leak after
1107          * the store to key_idx. i.e. key_idx is the guard variable
1108          * for signature and key.
1109          */
1110         __atomic_store_n(&(h->buckets_ext[ext_bkt_id - 1]).key_idx[0],
1111                          slot_id,
1112                          __ATOMIC_RELEASE);
1113         /* Link the new bucket to sec bucket linked list */
1114         last = rte_hash_get_last_bkt(sec_bkt);
1115         last->next = &h->buckets_ext[ext_bkt_id - 1];
1116         __hash_rw_writer_unlock(h);
1117         return slot_id - 1;
1118
1119 failure:
1120         __hash_rw_writer_unlock(h);
1121         return ret;
1122
1123 }
1124
1125 int32_t
1126 rte_hash_add_key_with_hash(const struct rte_hash *h,
1127                         const void *key, hash_sig_t sig)
1128 {
1129         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1130         return __rte_hash_add_key_with_hash(h, key, sig, 0);
1131 }
1132
1133 int32_t
1134 rte_hash_add_key(const struct rte_hash *h, const void *key)
1135 {
1136         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1137         return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
1138 }
1139
1140 int
1141 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
1142                         const void *key, hash_sig_t sig, void *data)
1143 {
1144         int ret;
1145
1146         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1147         ret = __rte_hash_add_key_with_hash(h, key, sig, data);
1148         if (ret >= 0)
1149                 return 0;
1150         else
1151                 return ret;
1152 }
1153
1154 int
1155 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
1156 {
1157         int ret;
1158
1159         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1160
1161         ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
1162         if (ret >= 0)
1163                 return 0;
1164         else
1165                 return ret;
1166 }
1167
1168 /* Search one bucket to find the match key - uses rw lock */
1169 static inline int32_t
1170 search_one_bucket_l(const struct rte_hash *h, const void *key,
1171                 uint16_t sig, void **data,
1172                 const struct rte_hash_bucket *bkt)
1173 {
1174         int i;
1175         struct rte_hash_key *k, *keys = h->key_store;
1176
1177         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1178                 if (bkt->sig_current[i] == sig &&
1179                                 bkt->key_idx[i] != EMPTY_SLOT) {
1180                         k = (struct rte_hash_key *) ((char *)keys +
1181                                         bkt->key_idx[i] * h->key_entry_size);
1182
1183                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1184                                 if (data != NULL)
1185                                         *data = k->pdata;
1186                                 /*
1187                                  * Return index where key is stored,
1188                                  * subtracting the first dummy index
1189                                  */
1190                                 return bkt->key_idx[i] - 1;
1191                         }
1192                 }
1193         }
1194         return -1;
1195 }
1196
1197 /* Search one bucket to find the match key */
1198 static inline int32_t
1199 search_one_bucket_lf(const struct rte_hash *h, const void *key, uint16_t sig,
1200                         void **data, const struct rte_hash_bucket *bkt)
1201 {
1202         int i;
1203         uint32_t key_idx;
1204         struct rte_hash_key *k, *keys = h->key_store;
1205
1206         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1207                 /* Signature comparison is done before the acquire-load
1208                  * of the key index to achieve better performance.
1209                  * This can result in the reader loading old signature
1210                  * (which matches), while the key_idx is updated to a
1211                  * value that belongs to a new key. However, the full
1212                  * key comparison will ensure that the lookup fails.
1213                  */
1214                 if (bkt->sig_current[i] == sig) {
1215                         key_idx = __atomic_load_n(&bkt->key_idx[i],
1216                                           __ATOMIC_ACQUIRE);
1217                         if (key_idx != EMPTY_SLOT) {
1218                                 k = (struct rte_hash_key *) ((char *)keys +
1219                                                 key_idx * h->key_entry_size);
1220
1221                                 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1222                                         if (data != NULL) {
1223                                                 *data = __atomic_load_n(
1224                                                         &k->pdata,
1225                                                         __ATOMIC_ACQUIRE);
1226                                         }
1227                                         /*
1228                                          * Return index where key is stored,
1229                                          * subtracting the first dummy index
1230                                          */
1231                                         return key_idx - 1;
1232                                 }
1233                         }
1234                 }
1235         }
1236         return -1;
1237 }
1238
1239 static inline int32_t
1240 __rte_hash_lookup_with_hash_l(const struct rte_hash *h, const void *key,
1241                                 hash_sig_t sig, void **data)
1242 {
1243         uint32_t prim_bucket_idx, sec_bucket_idx;
1244         struct rte_hash_bucket *bkt, *cur_bkt;
1245         int ret;
1246         uint16_t short_sig;
1247
1248         short_sig = get_short_sig(sig);
1249         prim_bucket_idx = get_prim_bucket_index(h, sig);
1250         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1251
1252         bkt = &h->buckets[prim_bucket_idx];
1253
1254         __hash_rw_reader_lock(h);
1255
1256         /* Check if key is in primary location */
1257         ret = search_one_bucket_l(h, key, short_sig, data, bkt);
1258         if (ret != -1) {
1259                 __hash_rw_reader_unlock(h);
1260                 return ret;
1261         }
1262         /* Calculate secondary hash */
1263         bkt = &h->buckets[sec_bucket_idx];
1264
1265         /* Check if key is in secondary location */
1266         FOR_EACH_BUCKET(cur_bkt, bkt) {
1267                 ret = search_one_bucket_l(h, key, short_sig,
1268                                         data, cur_bkt);
1269                 if (ret != -1) {
1270                         __hash_rw_reader_unlock(h);
1271                         return ret;
1272                 }
1273         }
1274
1275         __hash_rw_reader_unlock(h);
1276
1277         return -ENOENT;
1278 }
1279
1280 static inline int32_t
1281 __rte_hash_lookup_with_hash_lf(const struct rte_hash *h, const void *key,
1282                                         hash_sig_t sig, void **data)
1283 {
1284         uint32_t prim_bucket_idx, sec_bucket_idx;
1285         struct rte_hash_bucket *bkt, *cur_bkt;
1286         uint32_t cnt_b, cnt_a;
1287         int ret;
1288         uint16_t short_sig;
1289
1290         short_sig = get_short_sig(sig);
1291         prim_bucket_idx = get_prim_bucket_index(h, sig);
1292         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1293
1294         do {
1295                 /* Load the table change counter before the lookup
1296                  * starts. Acquire semantics will make sure that
1297                  * loads in search_one_bucket are not hoisted.
1298                  */
1299                 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1300                                 __ATOMIC_ACQUIRE);
1301
1302                 /* Check if key is in primary location */
1303                 bkt = &h->buckets[prim_bucket_idx];
1304                 ret = search_one_bucket_lf(h, key, short_sig, data, bkt);
1305                 if (ret != -1)
1306                         return ret;
1307                 /* Calculate secondary hash */
1308                 bkt = &h->buckets[sec_bucket_idx];
1309
1310                 /* Check if key is in secondary location */
1311                 FOR_EACH_BUCKET(cur_bkt, bkt) {
1312                         ret = search_one_bucket_lf(h, key, short_sig,
1313                                                 data, cur_bkt);
1314                         if (ret != -1)
1315                                 return ret;
1316                 }
1317
1318                 /* The loads of sig_current in search_one_bucket
1319                  * should not move below the load from tbl_chng_cnt.
1320                  */
1321                 __atomic_thread_fence(__ATOMIC_ACQUIRE);
1322                 /* Re-read the table change counter to check if the
1323                  * table has changed during search. If yes, re-do
1324                  * the search.
1325                  * This load should not get hoisted. The load
1326                  * acquires on cnt_b, key index in primary bucket
1327                  * and key index in secondary bucket will make sure
1328                  * that it does not get hoisted.
1329                  */
1330                 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
1331                                         __ATOMIC_ACQUIRE);
1332         } while (cnt_b != cnt_a);
1333
1334         return -ENOENT;
1335 }
1336
1337 static inline int32_t
1338 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
1339                                         hash_sig_t sig, void **data)
1340 {
1341         if (h->readwrite_concur_lf_support)
1342                 return __rte_hash_lookup_with_hash_lf(h, key, sig, data);
1343         else
1344                 return __rte_hash_lookup_with_hash_l(h, key, sig, data);
1345 }
1346
1347 int32_t
1348 rte_hash_lookup_with_hash(const struct rte_hash *h,
1349                         const void *key, hash_sig_t sig)
1350 {
1351         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1352         return __rte_hash_lookup_with_hash(h, key, sig, NULL);
1353 }
1354
1355 int32_t
1356 rte_hash_lookup(const struct rte_hash *h, const void *key)
1357 {
1358         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1359         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
1360 }
1361
1362 int
1363 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
1364                         const void *key, hash_sig_t sig, void **data)
1365 {
1366         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1367         return __rte_hash_lookup_with_hash(h, key, sig, data);
1368 }
1369
1370 int
1371 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
1372 {
1373         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1374         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
1375 }
1376
1377 static inline void
1378 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
1379 {
1380         unsigned lcore_id, n_slots;
1381         struct lcore_cache *cached_free_slots;
1382
1383         if (h->use_local_cache) {
1384                 lcore_id = rte_lcore_id();
1385                 cached_free_slots = &h->local_free_slots[lcore_id];
1386                 /* Cache full, need to free it. */
1387                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1388                         /* Need to enqueue the free slots in global ring. */
1389                         n_slots = rte_ring_mp_enqueue_burst_elem(h->free_slots,
1390                                                 cached_free_slots->objs,
1391                                                 sizeof(uint32_t),
1392                                                 LCORE_CACHE_SIZE, NULL);
1393                         ERR_IF_TRUE((n_slots == 0),
1394                                 "%s: could not enqueue free slots in global ring\n",
1395                                 __func__);
1396                         cached_free_slots->len -= n_slots;
1397                 }
1398                 /* Put index of new free slot in cache. */
1399                 cached_free_slots->objs[cached_free_slots->len] =
1400                                                         bkt->key_idx[i];
1401                 cached_free_slots->len++;
1402         } else {
1403                 rte_ring_sp_enqueue_elem(h->free_slots,
1404                                 &bkt->key_idx[i], sizeof(uint32_t));
1405         }
1406 }
1407
1408 /* Compact the linked list by moving key from last entry in linked list to the
1409  * empty slot.
1410  */
1411 static inline void
1412 __rte_hash_compact_ll(const struct rte_hash *h,
1413                         struct rte_hash_bucket *cur_bkt, int pos) {
1414         int i;
1415         struct rte_hash_bucket *last_bkt;
1416
1417         if (!cur_bkt->next)
1418                 return;
1419
1420         last_bkt = rte_hash_get_last_bkt(cur_bkt);
1421
1422         for (i = RTE_HASH_BUCKET_ENTRIES - 1; i >= 0; i--) {
1423                 if (last_bkt->key_idx[i] != EMPTY_SLOT) {
1424                         cur_bkt->sig_current[pos] = last_bkt->sig_current[i];
1425                         __atomic_store_n(&cur_bkt->key_idx[pos],
1426                                          last_bkt->key_idx[i],
1427                                          __ATOMIC_RELEASE);
1428                         if (h->readwrite_concur_lf_support) {
1429                                 /* Inform the readers that the table has changed
1430                                  * Since there is one writer, load acquire on
1431                                  * tbl_chng_cnt is not required.
1432                                  */
1433                                 __atomic_store_n(h->tbl_chng_cnt,
1434                                          *h->tbl_chng_cnt + 1,
1435                                          __ATOMIC_RELEASE);
1436                                 /* The store to sig_current should
1437                                  * not move above the store to tbl_chng_cnt.
1438                                  */
1439                                 __atomic_thread_fence(__ATOMIC_RELEASE);
1440                         }
1441                         last_bkt->sig_current[i] = NULL_SIGNATURE;
1442                         __atomic_store_n(&last_bkt->key_idx[i],
1443                                          EMPTY_SLOT,
1444                                          __ATOMIC_RELEASE);
1445                         return;
1446                 }
1447         }
1448 }
1449
1450 /* Search one bucket and remove the matched key.
1451  * Writer is expected to hold the lock while calling this
1452  * function.
1453  */
1454 static inline int32_t
1455 search_and_remove(const struct rte_hash *h, const void *key,
1456                         struct rte_hash_bucket *bkt, uint16_t sig, int *pos)
1457 {
1458         struct rte_hash_key *k, *keys = h->key_store;
1459         unsigned int i;
1460         uint32_t key_idx;
1461
1462         /* Check if key is in bucket */
1463         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1464                 key_idx = __atomic_load_n(&bkt->key_idx[i],
1465                                           __ATOMIC_ACQUIRE);
1466                 if (bkt->sig_current[i] == sig && key_idx != EMPTY_SLOT) {
1467                         k = (struct rte_hash_key *) ((char *)keys +
1468                                         key_idx * h->key_entry_size);
1469                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1470                                 bkt->sig_current[i] = NULL_SIGNATURE;
1471                                 /* Free the key store index if
1472                                  * no_free_on_del is disabled.
1473                                  */
1474                                 if (!h->no_free_on_del)
1475                                         remove_entry(h, bkt, i);
1476
1477                                 __atomic_store_n(&bkt->key_idx[i],
1478                                                  EMPTY_SLOT,
1479                                                  __ATOMIC_RELEASE);
1480
1481                                 *pos = i;
1482                                 /*
1483                                  * Return index where key is stored,
1484                                  * subtracting the first dummy index
1485                                  */
1486                                 return key_idx - 1;
1487                         }
1488                 }
1489         }
1490         return -1;
1491 }
1492
1493 static inline int32_t
1494 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
1495                                                 hash_sig_t sig)
1496 {
1497         uint32_t prim_bucket_idx, sec_bucket_idx;
1498         struct rte_hash_bucket *prim_bkt, *sec_bkt, *prev_bkt, *last_bkt;
1499         struct rte_hash_bucket *cur_bkt;
1500         int pos;
1501         int32_t ret, i;
1502         uint16_t short_sig;
1503
1504         short_sig = get_short_sig(sig);
1505         prim_bucket_idx = get_prim_bucket_index(h, sig);
1506         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1507         prim_bkt = &h->buckets[prim_bucket_idx];
1508
1509         __hash_rw_writer_lock(h);
1510         /* look for key in primary bucket */
1511         ret = search_and_remove(h, key, prim_bkt, short_sig, &pos);
1512         if (ret != -1) {
1513                 __rte_hash_compact_ll(h, prim_bkt, pos);
1514                 last_bkt = prim_bkt->next;
1515                 prev_bkt = prim_bkt;
1516                 goto return_bkt;
1517         }
1518
1519         /* Calculate secondary hash */
1520         sec_bkt = &h->buckets[sec_bucket_idx];
1521
1522         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1523                 ret = search_and_remove(h, key, cur_bkt, short_sig, &pos);
1524                 if (ret != -1) {
1525                         __rte_hash_compact_ll(h, cur_bkt, pos);
1526                         last_bkt = sec_bkt->next;
1527                         prev_bkt = sec_bkt;
1528                         goto return_bkt;
1529                 }
1530         }
1531
1532         __hash_rw_writer_unlock(h);
1533         return -ENOENT;
1534
1535 /* Search last bucket to see if empty to be recycled */
1536 return_bkt:
1537         if (!last_bkt) {
1538                 __hash_rw_writer_unlock(h);
1539                 return ret;
1540         }
1541         while (last_bkt->next) {
1542                 prev_bkt = last_bkt;
1543                 last_bkt = last_bkt->next;
1544         }
1545
1546         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1547                 if (last_bkt->key_idx[i] != EMPTY_SLOT)
1548                         break;
1549         }
1550         /* found empty bucket and recycle */
1551         if (i == RTE_HASH_BUCKET_ENTRIES) {
1552                 prev_bkt->next = NULL;
1553                 uint32_t index = last_bkt - h->buckets_ext + 1;
1554                 /* Recycle the empty bkt if
1555                  * no_free_on_del is disabled.
1556                  */
1557                 if (h->no_free_on_del)
1558                         /* Store index of an empty ext bkt to be recycled
1559                          * on calling rte_hash_del_xxx APIs.
1560                          * When lock free read-write concurrency is enabled,
1561                          * an empty ext bkt cannot be put into free list
1562                          * immediately (as readers might be using it still).
1563                          * Hence freeing of the ext bkt is piggy-backed to
1564                          * freeing of the key index.
1565                          */
1566                         h->ext_bkt_to_free[ret] = index;
1567                 else
1568                         rte_ring_sp_enqueue_elem(h->free_ext_bkts, &index,
1569                                                         sizeof(uint32_t));
1570         }
1571         __hash_rw_writer_unlock(h);
1572         return ret;
1573 }
1574
1575 int32_t
1576 rte_hash_del_key_with_hash(const struct rte_hash *h,
1577                         const void *key, hash_sig_t sig)
1578 {
1579         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1580         return __rte_hash_del_key_with_hash(h, key, sig);
1581 }
1582
1583 int32_t
1584 rte_hash_del_key(const struct rte_hash *h, const void *key)
1585 {
1586         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1587         return __rte_hash_del_key_with_hash(h, key, rte_hash_hash(h, key));
1588 }
1589
1590 int
1591 rte_hash_get_key_with_position(const struct rte_hash *h, const int32_t position,
1592                                void **key)
1593 {
1594         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1595
1596         struct rte_hash_key *k, *keys = h->key_store;
1597         k = (struct rte_hash_key *) ((char *) keys + (position + 1) *
1598                                      h->key_entry_size);
1599         *key = k->key;
1600
1601         if (position !=
1602             __rte_hash_lookup_with_hash(h, *key, rte_hash_hash(h, *key),
1603                                         NULL)) {
1604                 return -ENOENT;
1605         }
1606
1607         return 0;
1608 }
1609
1610 int
1611 rte_hash_free_key_with_position(const struct rte_hash *h,
1612                                 const int32_t position)
1613 {
1614         /* Key index where key is stored, adding the first dummy index */
1615         uint32_t key_idx = position + 1;
1616
1617         RETURN_IF_TRUE(((h == NULL) || (key_idx == EMPTY_SLOT)), -EINVAL);
1618
1619         unsigned int lcore_id, n_slots;
1620         struct lcore_cache *cached_free_slots;
1621         const uint32_t total_entries = h->use_local_cache ?
1622                 h->entries + (RTE_MAX_LCORE - 1) * (LCORE_CACHE_SIZE - 1) + 1
1623                                                         : h->entries + 1;
1624
1625         /* Out of bounds */
1626         if (key_idx >= total_entries)
1627                 return -EINVAL;
1628         if (h->ext_table_support && h->readwrite_concur_lf_support) {
1629                 uint32_t index = h->ext_bkt_to_free[position];
1630                 if (index) {
1631                         /* Recycle empty ext bkt to free list. */
1632                         rte_ring_sp_enqueue_elem(h->free_ext_bkts, &index,
1633                                                         sizeof(uint32_t));
1634                         h->ext_bkt_to_free[position] = 0;
1635                 }
1636         }
1637
1638         if (h->use_local_cache) {
1639                 lcore_id = rte_lcore_id();
1640                 cached_free_slots = &h->local_free_slots[lcore_id];
1641                 /* Cache full, need to free it. */
1642                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1643                         /* Need to enqueue the free slots in global ring. */
1644                         n_slots = rte_ring_mp_enqueue_burst_elem(h->free_slots,
1645                                                 cached_free_slots->objs,
1646                                                 sizeof(uint32_t),
1647                                                 LCORE_CACHE_SIZE, NULL);
1648                         RETURN_IF_TRUE((n_slots == 0), -EFAULT);
1649                         cached_free_slots->len -= n_slots;
1650                 }
1651                 /* Put index of new free slot in cache. */
1652                 cached_free_slots->objs[cached_free_slots->len] = key_idx;
1653                 cached_free_slots->len++;
1654         } else {
1655                 rte_ring_sp_enqueue_elem(h->free_slots, &key_idx,
1656                                                 sizeof(uint32_t));
1657         }
1658
1659         return 0;
1660 }
1661
1662 static inline void
1663 compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
1664                         const struct rte_hash_bucket *prim_bkt,
1665                         const struct rte_hash_bucket *sec_bkt,
1666                         uint16_t sig,
1667                         enum rte_hash_sig_compare_function sig_cmp_fn)
1668 {
1669         unsigned int i;
1670
1671         /* For match mask the first bit of every two bits indicates the match */
1672         switch (sig_cmp_fn) {
1673 #if defined(RTE_MACHINE_CPUFLAG_SSE2)
1674         case RTE_HASH_COMPARE_SSE:
1675                 /* Compare all signatures in the bucket */
1676                 *prim_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1677                                 _mm_load_si128(
1678                                         (__m128i const *)prim_bkt->sig_current),
1679                                 _mm_set1_epi16(sig)));
1680                 /* Compare all signatures in the bucket */
1681                 *sec_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1682                                 _mm_load_si128(
1683                                         (__m128i const *)sec_bkt->sig_current),
1684                                 _mm_set1_epi16(sig)));
1685                 break;
1686 #elif defined(RTE_MACHINE_CPUFLAG_NEON)
1687         case RTE_HASH_COMPARE_NEON: {
1688                 uint16x8_t vmat, vsig, x;
1689                 int16x8_t shift = {-15, -13, -11, -9, -7, -5, -3, -1};
1690
1691                 vsig = vld1q_dup_u16((uint16_t const *)&sig);
1692                 /* Compare all signatures in the primary bucket */
1693                 vmat = vceqq_u16(vsig,
1694                         vld1q_u16((uint16_t const *)prim_bkt->sig_current));
1695                 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1696                 *prim_hash_matches = (uint32_t)(vaddvq_u16(x));
1697                 /* Compare all signatures in the secondary bucket */
1698                 vmat = vceqq_u16(vsig,
1699                         vld1q_u16((uint16_t const *)sec_bkt->sig_current));
1700                 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1701                 *sec_hash_matches = (uint32_t)(vaddvq_u16(x));
1702                 }
1703                 break;
1704 #endif
1705         default:
1706                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1707                         *prim_hash_matches |=
1708                                 ((sig == prim_bkt->sig_current[i]) << (i << 1));
1709                         *sec_hash_matches |=
1710                                 ((sig == sec_bkt->sig_current[i]) << (i << 1));
1711                 }
1712         }
1713 }
1714
1715 static inline void
1716 __bulk_lookup_l(const struct rte_hash *h, const void **keys,
1717                 const struct rte_hash_bucket **primary_bkt,
1718                 const struct rte_hash_bucket **secondary_bkt,
1719                 uint16_t *sig, int32_t num_keys, int32_t *positions,
1720                 uint64_t *hit_mask, void *data[])
1721 {
1722         uint64_t hits = 0;
1723         int32_t i;
1724         int32_t ret;
1725         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1726         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1727         struct rte_hash_bucket *cur_bkt, *next_bkt;
1728
1729         __hash_rw_reader_lock(h);
1730
1731         /* Compare signatures and prefetch key slot of first hit */
1732         for (i = 0; i < num_keys; i++) {
1733                 compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1734                         primary_bkt[i], secondary_bkt[i],
1735                         sig[i], h->sig_cmp_fn);
1736
1737                 if (prim_hitmask[i]) {
1738                         uint32_t first_hit =
1739                                         __builtin_ctzl(prim_hitmask[i])
1740                                         >> 1;
1741                         uint32_t key_idx =
1742                                 primary_bkt[i]->key_idx[first_hit];
1743                         const struct rte_hash_key *key_slot =
1744                                 (const struct rte_hash_key *)(
1745                                 (const char *)h->key_store +
1746                                 key_idx * h->key_entry_size);
1747                         rte_prefetch0(key_slot);
1748                         continue;
1749                 }
1750
1751                 if (sec_hitmask[i]) {
1752                         uint32_t first_hit =
1753                                         __builtin_ctzl(sec_hitmask[i])
1754                                         >> 1;
1755                         uint32_t key_idx =
1756                                 secondary_bkt[i]->key_idx[first_hit];
1757                         const struct rte_hash_key *key_slot =
1758                                 (const struct rte_hash_key *)(
1759                                 (const char *)h->key_store +
1760                                 key_idx * h->key_entry_size);
1761                         rte_prefetch0(key_slot);
1762                 }
1763         }
1764
1765         /* Compare keys, first hits in primary first */
1766         for (i = 0; i < num_keys; i++) {
1767                 positions[i] = -ENOENT;
1768                 while (prim_hitmask[i]) {
1769                         uint32_t hit_index =
1770                                         __builtin_ctzl(prim_hitmask[i])
1771                                         >> 1;
1772                         uint32_t key_idx =
1773                                 primary_bkt[i]->key_idx[hit_index];
1774                         const struct rte_hash_key *key_slot =
1775                                 (const struct rte_hash_key *)(
1776                                 (const char *)h->key_store +
1777                                 key_idx * h->key_entry_size);
1778
1779                         /*
1780                          * If key index is 0, do not compare key,
1781                          * as it is checking the dummy slot
1782                          */
1783                         if (!!key_idx &
1784                                 !rte_hash_cmp_eq(
1785                                         key_slot->key, keys[i], h)) {
1786                                 if (data != NULL)
1787                                         data[i] = key_slot->pdata;
1788
1789                                 hits |= 1ULL << i;
1790                                 positions[i] = key_idx - 1;
1791                                 goto next_key;
1792                         }
1793                         prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
1794                 }
1795
1796                 while (sec_hitmask[i]) {
1797                         uint32_t hit_index =
1798                                         __builtin_ctzl(sec_hitmask[i])
1799                                         >> 1;
1800                         uint32_t key_idx =
1801                                 secondary_bkt[i]->key_idx[hit_index];
1802                         const struct rte_hash_key *key_slot =
1803                                 (const struct rte_hash_key *)(
1804                                 (const char *)h->key_store +
1805                                 key_idx * h->key_entry_size);
1806
1807                         /*
1808                          * If key index is 0, do not compare key,
1809                          * as it is checking the dummy slot
1810                          */
1811
1812                         if (!!key_idx &
1813                                 !rte_hash_cmp_eq(
1814                                         key_slot->key, keys[i], h)) {
1815                                 if (data != NULL)
1816                                         data[i] = key_slot->pdata;
1817
1818                                 hits |= 1ULL << i;
1819                                 positions[i] = key_idx - 1;
1820                                 goto next_key;
1821                         }
1822                         sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
1823                 }
1824 next_key:
1825                 continue;
1826         }
1827
1828         /* all found, do not need to go through ext bkt */
1829         if ((hits == ((1ULL << num_keys) - 1)) || !h->ext_table_support) {
1830                 if (hit_mask != NULL)
1831                         *hit_mask = hits;
1832                 __hash_rw_reader_unlock(h);
1833                 return;
1834         }
1835
1836         /* need to check ext buckets for match */
1837         for (i = 0; i < num_keys; i++) {
1838                 if ((hits & (1ULL << i)) != 0)
1839                         continue;
1840                 next_bkt = secondary_bkt[i]->next;
1841                 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
1842                         if (data != NULL)
1843                                 ret = search_one_bucket_l(h, keys[i],
1844                                                 sig[i], &data[i], cur_bkt);
1845                         else
1846                                 ret = search_one_bucket_l(h, keys[i],
1847                                                 sig[i], NULL, cur_bkt);
1848                         if (ret != -1) {
1849                                 positions[i] = ret;
1850                                 hits |= 1ULL << i;
1851                                 break;
1852                         }
1853                 }
1854         }
1855
1856         __hash_rw_reader_unlock(h);
1857
1858         if (hit_mask != NULL)
1859                 *hit_mask = hits;
1860 }
1861
1862 static inline void
1863 __bulk_lookup_lf(const struct rte_hash *h, const void **keys,
1864                 const struct rte_hash_bucket **primary_bkt,
1865                 const struct rte_hash_bucket **secondary_bkt,
1866                 uint16_t *sig, int32_t num_keys, int32_t *positions,
1867                 uint64_t *hit_mask, void *data[])
1868 {
1869         uint64_t hits = 0;
1870         int32_t i;
1871         int32_t ret;
1872         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1873         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1874         struct rte_hash_bucket *cur_bkt, *next_bkt;
1875         uint32_t cnt_b, cnt_a;
1876
1877         for (i = 0; i < num_keys; i++)
1878                 positions[i] = -ENOENT;
1879
1880         do {
1881                 /* Load the table change counter before the lookup
1882                  * starts. Acquire semantics will make sure that
1883                  * loads in compare_signatures are not hoisted.
1884                  */
1885                 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1886                                         __ATOMIC_ACQUIRE);
1887
1888                 /* Compare signatures and prefetch key slot of first hit */
1889                 for (i = 0; i < num_keys; i++) {
1890                         compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1891                                 primary_bkt[i], secondary_bkt[i],
1892                                 sig[i], h->sig_cmp_fn);
1893
1894                         if (prim_hitmask[i]) {
1895                                 uint32_t first_hit =
1896                                                 __builtin_ctzl(prim_hitmask[i])
1897                                                 >> 1;
1898                                 uint32_t key_idx =
1899                                         primary_bkt[i]->key_idx[first_hit];
1900                                 const struct rte_hash_key *key_slot =
1901                                         (const struct rte_hash_key *)(
1902                                         (const char *)h->key_store +
1903                                         key_idx * h->key_entry_size);
1904                                 rte_prefetch0(key_slot);
1905                                 continue;
1906                         }
1907
1908                         if (sec_hitmask[i]) {
1909                                 uint32_t first_hit =
1910                                                 __builtin_ctzl(sec_hitmask[i])
1911                                                 >> 1;
1912                                 uint32_t key_idx =
1913                                         secondary_bkt[i]->key_idx[first_hit];
1914                                 const struct rte_hash_key *key_slot =
1915                                         (const struct rte_hash_key *)(
1916                                         (const char *)h->key_store +
1917                                         key_idx * h->key_entry_size);
1918                                 rte_prefetch0(key_slot);
1919                         }
1920                 }
1921
1922                 /* Compare keys, first hits in primary first */
1923                 for (i = 0; i < num_keys; i++) {
1924                         while (prim_hitmask[i]) {
1925                                 uint32_t hit_index =
1926                                                 __builtin_ctzl(prim_hitmask[i])
1927                                                 >> 1;
1928                                 uint32_t key_idx =
1929                                 __atomic_load_n(
1930                                         &primary_bkt[i]->key_idx[hit_index],
1931                                         __ATOMIC_ACQUIRE);
1932                                 const struct rte_hash_key *key_slot =
1933                                         (const struct rte_hash_key *)(
1934                                         (const char *)h->key_store +
1935                                         key_idx * h->key_entry_size);
1936
1937                                 /*
1938                                  * If key index is 0, do not compare key,
1939                                  * as it is checking the dummy slot
1940                                  */
1941                                 if (!!key_idx &
1942                                         !rte_hash_cmp_eq(
1943                                                 key_slot->key, keys[i], h)) {
1944                                         if (data != NULL)
1945                                                 data[i] = __atomic_load_n(
1946                                                         &key_slot->pdata,
1947                                                         __ATOMIC_ACQUIRE);
1948
1949                                         hits |= 1ULL << i;
1950                                         positions[i] = key_idx - 1;
1951                                         goto next_key;
1952                                 }
1953                                 prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
1954                         }
1955
1956                         while (sec_hitmask[i]) {
1957                                 uint32_t hit_index =
1958                                                 __builtin_ctzl(sec_hitmask[i])
1959                                                 >> 1;
1960                                 uint32_t key_idx =
1961                                 __atomic_load_n(
1962                                         &secondary_bkt[i]->key_idx[hit_index],
1963                                         __ATOMIC_ACQUIRE);
1964                                 const struct rte_hash_key *key_slot =
1965                                         (const struct rte_hash_key *)(
1966                                         (const char *)h->key_store +
1967                                         key_idx * h->key_entry_size);
1968
1969                                 /*
1970                                  * If key index is 0, do not compare key,
1971                                  * as it is checking the dummy slot
1972                                  */
1973
1974                                 if (!!key_idx &
1975                                         !rte_hash_cmp_eq(
1976                                                 key_slot->key, keys[i], h)) {
1977                                         if (data != NULL)
1978                                                 data[i] = __atomic_load_n(
1979                                                         &key_slot->pdata,
1980                                                         __ATOMIC_ACQUIRE);
1981
1982                                         hits |= 1ULL << i;
1983                                         positions[i] = key_idx - 1;
1984                                         goto next_key;
1985                                 }
1986                                 sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
1987                         }
1988 next_key:
1989                         continue;
1990                 }
1991
1992                 /* all found, do not need to go through ext bkt */
1993                 if (hits == ((1ULL << num_keys) - 1)) {
1994                         if (hit_mask != NULL)
1995                                 *hit_mask = hits;
1996                         return;
1997                 }
1998                 /* need to check ext buckets for match */
1999                 if (h->ext_table_support) {
2000                         for (i = 0; i < num_keys; i++) {
2001                                 if ((hits & (1ULL << i)) != 0)
2002                                         continue;
2003                                 next_bkt = secondary_bkt[i]->next;
2004                                 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
2005                                         if (data != NULL)
2006                                                 ret = search_one_bucket_lf(h,
2007                                                         keys[i], sig[i],
2008                                                         &data[i], cur_bkt);
2009                                         else
2010                                                 ret = search_one_bucket_lf(h,
2011                                                                 keys[i], sig[i],
2012                                                                 NULL, cur_bkt);
2013                                         if (ret != -1) {
2014                                                 positions[i] = ret;
2015                                                 hits |= 1ULL << i;
2016                                                 break;
2017                                         }
2018                                 }
2019                         }
2020                 }
2021                 /* The loads of sig_current in compare_signatures
2022                  * should not move below the load from tbl_chng_cnt.
2023                  */
2024                 __atomic_thread_fence(__ATOMIC_ACQUIRE);
2025                 /* Re-read the table change counter to check if the
2026                  * table has changed during search. If yes, re-do
2027                  * the search.
2028                  * This load should not get hoisted. The load
2029                  * acquires on cnt_b, primary key index and secondary
2030                  * key index will make sure that it does not get
2031                  * hoisted.
2032                  */
2033                 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
2034                                         __ATOMIC_ACQUIRE);
2035         } while (cnt_b != cnt_a);
2036
2037         if (hit_mask != NULL)
2038                 *hit_mask = hits;
2039 }
2040
2041 #define PREFETCH_OFFSET 4
2042 static inline void
2043 __bulk_lookup_prefetching_loop(const struct rte_hash *h,
2044         const void **keys, int32_t num_keys,
2045         uint16_t *sig,
2046         const struct rte_hash_bucket **primary_bkt,
2047         const struct rte_hash_bucket **secondary_bkt)
2048 {
2049         int32_t i;
2050         uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
2051         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
2052         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
2053
2054         /* Prefetch first keys */
2055         for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
2056                 rte_prefetch0(keys[i]);
2057
2058         /*
2059          * Prefetch rest of the keys, calculate primary and
2060          * secondary bucket and prefetch them
2061          */
2062         for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
2063                 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
2064
2065                 prim_hash[i] = rte_hash_hash(h, keys[i]);
2066
2067                 sig[i] = get_short_sig(prim_hash[i]);
2068                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
2069                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
2070
2071                 primary_bkt[i] = &h->buckets[prim_index[i]];
2072                 secondary_bkt[i] = &h->buckets[sec_index[i]];
2073
2074                 rte_prefetch0(primary_bkt[i]);
2075                 rte_prefetch0(secondary_bkt[i]);
2076         }
2077
2078         /* Calculate and prefetch rest of the buckets */
2079         for (; i < num_keys; i++) {
2080                 prim_hash[i] = rte_hash_hash(h, keys[i]);
2081
2082                 sig[i] = get_short_sig(prim_hash[i]);
2083                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
2084                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
2085
2086                 primary_bkt[i] = &h->buckets[prim_index[i]];
2087                 secondary_bkt[i] = &h->buckets[sec_index[i]];
2088
2089                 rte_prefetch0(primary_bkt[i]);
2090                 rte_prefetch0(secondary_bkt[i]);
2091         }
2092 }
2093
2094
2095 static inline void
2096 __rte_hash_lookup_bulk_l(const struct rte_hash *h, const void **keys,
2097                         int32_t num_keys, int32_t *positions,
2098                         uint64_t *hit_mask, void *data[])
2099 {
2100         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
2101         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2102         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2103
2104         __bulk_lookup_prefetching_loop(h, keys, num_keys, sig,
2105                 primary_bkt, secondary_bkt);
2106
2107         __bulk_lookup_l(h, keys, primary_bkt, secondary_bkt, sig, num_keys,
2108                 positions, hit_mask, data);
2109 }
2110
2111 static inline void
2112 __rte_hash_lookup_bulk_lf(const struct rte_hash *h, const void **keys,
2113                         int32_t num_keys, int32_t *positions,
2114                         uint64_t *hit_mask, void *data[])
2115 {
2116         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
2117         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2118         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2119
2120         __bulk_lookup_prefetching_loop(h, keys, num_keys, sig,
2121                 primary_bkt, secondary_bkt);
2122
2123         __bulk_lookup_lf(h, keys, primary_bkt, secondary_bkt, sig, num_keys,
2124                 positions, hit_mask, data);
2125 }
2126
2127 static inline void
2128 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2129                         int32_t num_keys, int32_t *positions,
2130                         uint64_t *hit_mask, void *data[])
2131 {
2132         if (h->readwrite_concur_lf_support)
2133                 __rte_hash_lookup_bulk_lf(h, keys, num_keys, positions,
2134                                           hit_mask, data);
2135         else
2136                 __rte_hash_lookup_bulk_l(h, keys, num_keys, positions,
2137                                          hit_mask, data);
2138 }
2139
2140 int
2141 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2142                       uint32_t num_keys, int32_t *positions)
2143 {
2144         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2145                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2146                         (positions == NULL)), -EINVAL);
2147
2148         __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
2149         return 0;
2150 }
2151
2152 int
2153 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
2154                       uint32_t num_keys, uint64_t *hit_mask, void *data[])
2155 {
2156         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2157                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2158                         (hit_mask == NULL)), -EINVAL);
2159
2160         int32_t positions[num_keys];
2161
2162         __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
2163
2164         /* Return number of hits */
2165         return __builtin_popcountl(*hit_mask);
2166 }
2167
2168
2169 static inline void
2170 __rte_hash_lookup_with_hash_bulk_l(const struct rte_hash *h,
2171                         const void **keys, hash_sig_t *prim_hash,
2172                         int32_t num_keys, int32_t *positions,
2173                         uint64_t *hit_mask, void *data[])
2174 {
2175         int32_t i;
2176         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
2177         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
2178         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
2179         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2180         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2181
2182         /*
2183          * Prefetch keys, calculate primary and
2184          * secondary bucket and prefetch them
2185          */
2186         for (i = 0; i < num_keys; i++) {
2187                 rte_prefetch0(keys[i]);
2188
2189                 sig[i] = get_short_sig(prim_hash[i]);
2190                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
2191                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
2192
2193                 primary_bkt[i] = &h->buckets[prim_index[i]];
2194                 secondary_bkt[i] = &h->buckets[sec_index[i]];
2195
2196                 rte_prefetch0(primary_bkt[i]);
2197                 rte_prefetch0(secondary_bkt[i]);
2198         }
2199
2200         __bulk_lookup_l(h, keys, primary_bkt, secondary_bkt, sig, num_keys,
2201                 positions, hit_mask, data);
2202 }
2203
2204 static inline void
2205 __rte_hash_lookup_with_hash_bulk_lf(const struct rte_hash *h,
2206                         const void **keys, hash_sig_t *prim_hash,
2207                         int32_t num_keys, int32_t *positions,
2208                         uint64_t *hit_mask, void *data[])
2209 {
2210         int32_t i;
2211         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
2212         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
2213         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
2214         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2215         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
2216
2217         /*
2218          * Prefetch keys, calculate primary and
2219          * secondary bucket and prefetch them
2220          */
2221         for (i = 0; i < num_keys; i++) {
2222                 rte_prefetch0(keys[i]);
2223
2224                 sig[i] = get_short_sig(prim_hash[i]);
2225                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
2226                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
2227
2228                 primary_bkt[i] = &h->buckets[prim_index[i]];
2229                 secondary_bkt[i] = &h->buckets[sec_index[i]];
2230
2231                 rte_prefetch0(primary_bkt[i]);
2232                 rte_prefetch0(secondary_bkt[i]);
2233         }
2234
2235         __bulk_lookup_lf(h, keys, primary_bkt, secondary_bkt, sig, num_keys,
2236                 positions, hit_mask, data);
2237 }
2238
2239 static inline void
2240 __rte_hash_lookup_with_hash_bulk(const struct rte_hash *h, const void **keys,
2241                         hash_sig_t *prim_hash, int32_t num_keys,
2242                         int32_t *positions, uint64_t *hit_mask, void *data[])
2243 {
2244         if (h->readwrite_concur_lf_support)
2245                 __rte_hash_lookup_with_hash_bulk_lf(h, keys, prim_hash,
2246                                 num_keys, positions, hit_mask, data);
2247         else
2248                 __rte_hash_lookup_with_hash_bulk_l(h, keys, prim_hash,
2249                                 num_keys, positions, hit_mask, data);
2250 }
2251
2252 int
2253 rte_hash_lookup_with_hash_bulk(const struct rte_hash *h, const void **keys,
2254                 hash_sig_t *sig, uint32_t num_keys, int32_t *positions)
2255 {
2256         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) ||
2257                         (sig == NULL) || (num_keys == 0) ||
2258                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2259                         (positions == NULL)), -EINVAL);
2260
2261         __rte_hash_lookup_with_hash_bulk(h, keys, sig, num_keys,
2262                 positions, NULL, NULL);
2263         return 0;
2264 }
2265
2266 int
2267 rte_hash_lookup_with_hash_bulk_data(const struct rte_hash *h,
2268                 const void **keys, hash_sig_t *sig,
2269                 uint32_t num_keys, uint64_t *hit_mask, void *data[])
2270 {
2271         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) ||
2272                         (sig == NULL) || (num_keys == 0) ||
2273                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2274                         (hit_mask == NULL)), -EINVAL);
2275
2276         int32_t positions[num_keys];
2277
2278         __rte_hash_lookup_with_hash_bulk(h, keys, sig, num_keys,
2279                         positions, hit_mask, data);
2280
2281         /* Return number of hits */
2282         return __builtin_popcountl(*hit_mask);
2283 }
2284
2285 int32_t
2286 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
2287 {
2288         uint32_t bucket_idx, idx, position;
2289         struct rte_hash_key *next_key;
2290
2291         RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
2292
2293         const uint32_t total_entries_main = h->num_buckets *
2294                                                         RTE_HASH_BUCKET_ENTRIES;
2295         const uint32_t total_entries = total_entries_main << 1;
2296
2297         /* Out of bounds of all buckets (both main table and ext table) */
2298         if (*next >= total_entries_main)
2299                 goto extend_table;
2300
2301         /* Calculate bucket and index of current iterator */
2302         bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2303         idx = *next % RTE_HASH_BUCKET_ENTRIES;
2304
2305         /* If current position is empty, go to the next one */
2306         while ((position = __atomic_load_n(&h->buckets[bucket_idx].key_idx[idx],
2307                                         __ATOMIC_ACQUIRE)) == EMPTY_SLOT) {
2308                 (*next)++;
2309                 /* End of table */
2310                 if (*next == total_entries_main)
2311                         goto extend_table;
2312                 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2313                 idx = *next % RTE_HASH_BUCKET_ENTRIES;
2314         }
2315
2316         __hash_rw_reader_lock(h);
2317         next_key = (struct rte_hash_key *) ((char *)h->key_store +
2318                                 position * h->key_entry_size);
2319         /* Return key and data */
2320         *key = next_key->key;
2321         *data = next_key->pdata;
2322
2323         __hash_rw_reader_unlock(h);
2324
2325         /* Increment iterator */
2326         (*next)++;
2327
2328         return position - 1;
2329
2330 /* Begin to iterate extendable buckets */
2331 extend_table:
2332         /* Out of total bound or if ext bucket feature is not enabled */
2333         if (*next >= total_entries || !h->ext_table_support)
2334                 return -ENOENT;
2335
2336         bucket_idx = (*next - total_entries_main) / RTE_HASH_BUCKET_ENTRIES;
2337         idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2338
2339         while ((position = h->buckets_ext[bucket_idx].key_idx[idx]) == EMPTY_SLOT) {
2340                 (*next)++;
2341                 if (*next == total_entries)
2342                         return -ENOENT;
2343                 bucket_idx = (*next - total_entries_main) /
2344                                                 RTE_HASH_BUCKET_ENTRIES;
2345                 idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2346         }
2347         __hash_rw_reader_lock(h);
2348         next_key = (struct rte_hash_key *) ((char *)h->key_store +
2349                                 position * h->key_entry_size);
2350         /* Return key and data */
2351         *key = next_key->key;
2352         *data = next_key->pdata;
2353
2354         __hash_rw_reader_unlock(h);
2355
2356         /* Increment iterator */
2357         (*next)++;
2358         return position - 1;
2359 }