hash: modify lookup bulk pipeline
[dpdk.git] / lib / librte_hash / rte_cuckoo_hash.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33
34 #include <string.h>
35 #include <stdint.h>
36 #include <errno.h>
37 #include <stdio.h>
38 #include <stdarg.h>
39 #include <sys/queue.h>
40
41 #include <rte_common.h>
42 #include <rte_memory.h>         /* for definition of RTE_CACHE_LINE_SIZE */
43 #include <rte_log.h>
44 #include <rte_memcpy.h>
45 #include <rte_prefetch.h>
46 #include <rte_branch_prediction.h>
47 #include <rte_memzone.h>
48 #include <rte_malloc.h>
49 #include <rte_eal.h>
50 #include <rte_eal_memconfig.h>
51 #include <rte_per_lcore.h>
52 #include <rte_errno.h>
53 #include <rte_string_fns.h>
54 #include <rte_cpuflags.h>
55 #include <rte_log.h>
56 #include <rte_rwlock.h>
57 #include <rte_spinlock.h>
58 #include <rte_ring.h>
59 #include <rte_compat.h>
60
61 #include "rte_hash.h"
62 #include "rte_cuckoo_hash.h"
63
64 #if defined(RTE_ARCH_X86)
65 #include "rte_cuckoo_hash_x86.h"
66 #endif
67
68 TAILQ_HEAD(rte_hash_list, rte_tailq_entry);
69
70 static struct rte_tailq_elem rte_hash_tailq = {
71         .name = "RTE_HASH",
72 };
73 EAL_REGISTER_TAILQ(rte_hash_tailq)
74
75 struct rte_hash *
76 rte_hash_find_existing(const char *name)
77 {
78         struct rte_hash *h = NULL;
79         struct rte_tailq_entry *te;
80         struct rte_hash_list *hash_list;
81
82         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
83
84         rte_rwlock_read_lock(RTE_EAL_TAILQ_RWLOCK);
85         TAILQ_FOREACH(te, hash_list, next) {
86                 h = (struct rte_hash *) te->data;
87                 if (strncmp(name, h->name, RTE_HASH_NAMESIZE) == 0)
88                         break;
89         }
90         rte_rwlock_read_unlock(RTE_EAL_TAILQ_RWLOCK);
91
92         if (te == NULL) {
93                 rte_errno = ENOENT;
94                 return NULL;
95         }
96         return h;
97 }
98
99 void rte_hash_set_cmp_func(struct rte_hash *h, rte_hash_cmp_eq_t func)
100 {
101         h->cmp_jump_table_idx = KEY_CUSTOM;
102         h->rte_hash_custom_cmp_eq = func;
103 }
104
105 static inline int
106 rte_hash_cmp_eq(const void *key1, const void *key2, const struct rte_hash *h)
107 {
108         if (h->cmp_jump_table_idx == KEY_CUSTOM)
109                 return h->rte_hash_custom_cmp_eq(key1, key2, h->key_len);
110         else
111                 return cmp_jump_table[h->cmp_jump_table_idx](key1, key2, h->key_len);
112 }
113
114 struct rte_hash *
115 rte_hash_create(const struct rte_hash_parameters *params)
116 {
117         struct rte_hash *h = NULL;
118         struct rte_tailq_entry *te = NULL;
119         struct rte_hash_list *hash_list;
120         struct rte_ring *r = NULL;
121         char hash_name[RTE_HASH_NAMESIZE];
122         void *k = NULL;
123         void *buckets = NULL;
124         char ring_name[RTE_RING_NAMESIZE];
125         unsigned num_key_slots;
126         unsigned hw_trans_mem_support = 0;
127         unsigned i;
128
129         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
130
131         if (params == NULL) {
132                 RTE_LOG(ERR, HASH, "rte_hash_create has no parameters\n");
133                 return NULL;
134         }
135
136         /* Check for valid parameters */
137         if ((params->entries > RTE_HASH_ENTRIES_MAX) ||
138                         (params->entries < RTE_HASH_BUCKET_ENTRIES) ||
139                         !rte_is_power_of_2(RTE_HASH_BUCKET_ENTRIES) ||
140                         (params->key_len == 0)) {
141                 rte_errno = EINVAL;
142                 RTE_LOG(ERR, HASH, "rte_hash_create has invalid parameters\n");
143                 return NULL;
144         }
145
146         /* Check extra flags field to check extra options. */
147         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT)
148                 hw_trans_mem_support = 1;
149
150         /* Store all keys and leave the first entry as a dummy entry for lookup_bulk */
151         if (hw_trans_mem_support)
152                 /*
153                  * Increase number of slots by total number of indices
154                  * that can be stored in the lcore caches
155                  * except for the first cache
156                  */
157                 num_key_slots = params->entries + (RTE_MAX_LCORE - 1) *
158                                         LCORE_CACHE_SIZE + 1;
159         else
160                 num_key_slots = params->entries + 1;
161
162         snprintf(ring_name, sizeof(ring_name), "HT_%s", params->name);
163         /* Create ring (Dummy slot index is not enqueued) */
164         r = rte_ring_create(ring_name, rte_align32pow2(num_key_slots - 1),
165                         params->socket_id, 0);
166         if (r == NULL) {
167                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
168                 goto err;
169         }
170
171         snprintf(hash_name, sizeof(hash_name), "HT_%s", params->name);
172
173         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
174
175         /* guarantee there's no existing: this is normally already checked
176          * by ring creation above */
177         TAILQ_FOREACH(te, hash_list, next) {
178                 h = (struct rte_hash *) te->data;
179                 if (strncmp(params->name, h->name, RTE_HASH_NAMESIZE) == 0)
180                         break;
181         }
182         h = NULL;
183         if (te != NULL) {
184                 rte_errno = EEXIST;
185                 te = NULL;
186                 goto err_unlock;
187         }
188
189         te = rte_zmalloc("HASH_TAILQ_ENTRY", sizeof(*te), 0);
190         if (te == NULL) {
191                 RTE_LOG(ERR, HASH, "tailq entry allocation failed\n");
192                 goto err_unlock;
193         }
194
195         h = (struct rte_hash *)rte_zmalloc_socket(hash_name, sizeof(struct rte_hash),
196                                         RTE_CACHE_LINE_SIZE, params->socket_id);
197
198         if (h == NULL) {
199                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
200                 goto err_unlock;
201         }
202
203         const uint32_t num_buckets = rte_align32pow2(params->entries)
204                                         / RTE_HASH_BUCKET_ENTRIES;
205
206         buckets = rte_zmalloc_socket(NULL,
207                                 num_buckets * sizeof(struct rte_hash_bucket),
208                                 RTE_CACHE_LINE_SIZE, params->socket_id);
209
210         if (buckets == NULL) {
211                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
212                 goto err_unlock;
213         }
214
215         const uint32_t key_entry_size = sizeof(struct rte_hash_key) + params->key_len;
216         const uint64_t key_tbl_size = (uint64_t) key_entry_size * num_key_slots;
217
218         k = rte_zmalloc_socket(NULL, key_tbl_size,
219                         RTE_CACHE_LINE_SIZE, params->socket_id);
220
221         if (k == NULL) {
222                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
223                 goto err_unlock;
224         }
225
226 /*
227  * If x86 architecture is used, select appropriate compare function,
228  * which may use x86 intrinsics, otherwise use memcmp
229  */
230 #if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
231         /* Select function to compare keys */
232         switch (params->key_len) {
233         case 16:
234                 h->cmp_jump_table_idx = KEY_16_BYTES;
235                 break;
236         case 32:
237                 h->cmp_jump_table_idx = KEY_32_BYTES;
238                 break;
239         case 48:
240                 h->cmp_jump_table_idx = KEY_48_BYTES;
241                 break;
242         case 64:
243                 h->cmp_jump_table_idx = KEY_64_BYTES;
244                 break;
245         case 80:
246                 h->cmp_jump_table_idx = KEY_80_BYTES;
247                 break;
248         case 96:
249                 h->cmp_jump_table_idx = KEY_96_BYTES;
250                 break;
251         case 112:
252                 h->cmp_jump_table_idx = KEY_112_BYTES;
253                 break;
254         case 128:
255                 h->cmp_jump_table_idx = KEY_128_BYTES;
256                 break;
257         default:
258                 /* If key is not multiple of 16, use generic memcmp */
259                 h->cmp_jump_table_idx = KEY_OTHER_BYTES;
260         }
261 #else
262         h->cmp_jump_table_idx = KEY_OTHER_BYTES;
263 #endif
264
265         if (hw_trans_mem_support) {
266                 h->local_free_slots = rte_zmalloc_socket(NULL,
267                                 sizeof(struct lcore_cache) * RTE_MAX_LCORE,
268                                 RTE_CACHE_LINE_SIZE, params->socket_id);
269         }
270
271         /* Setup hash context */
272         snprintf(h->name, sizeof(h->name), "%s", params->name);
273         h->entries = params->entries;
274         h->key_len = params->key_len;
275         h->key_entry_size = key_entry_size;
276         h->hash_func_init_val = params->hash_func_init_val;
277
278         h->num_buckets = num_buckets;
279         h->bucket_bitmask = h->num_buckets - 1;
280         h->buckets = buckets;
281         h->hash_func = (params->hash_func == NULL) ?
282                 DEFAULT_HASH_FUNC : params->hash_func;
283         h->key_store = k;
284         h->free_slots = r;
285         h->hw_trans_mem_support = hw_trans_mem_support;
286
287 #if defined(RTE_ARCH_X86)
288         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
289                 h->sig_cmp_fn = RTE_HASH_COMPARE_AVX2;
290         else if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE2))
291                 h->sig_cmp_fn = RTE_HASH_COMPARE_SSE;
292         else
293 #endif
294                 h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
295
296         /* Turn on multi-writer only with explicit flat from user and TM
297          * support.
298          */
299         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
300                 if (h->hw_trans_mem_support) {
301                         h->add_key = ADD_KEY_MULTIWRITER_TM;
302                 } else {
303                         h->add_key = ADD_KEY_MULTIWRITER;
304                         h->multiwriter_lock = rte_malloc(NULL,
305                                                         sizeof(rte_spinlock_t),
306                                                         LCORE_CACHE_SIZE);
307                         rte_spinlock_init(h->multiwriter_lock);
308                 }
309         } else
310                 h->add_key = ADD_KEY_SINGLEWRITER;
311
312         /* Populate free slots ring. Entry zero is reserved for key misses. */
313         for (i = 1; i < params->entries + 1; i++)
314                 rte_ring_sp_enqueue(r, (void *)((uintptr_t) i));
315
316         te->data = (void *) h;
317         TAILQ_INSERT_TAIL(hash_list, te, next);
318         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
319
320         return h;
321 err_unlock:
322         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
323 err:
324         rte_ring_free(r);
325         rte_free(te);
326         rte_free(h);
327         rte_free(buckets);
328         rte_free(k);
329         return NULL;
330 }
331
332 void
333 rte_hash_free(struct rte_hash *h)
334 {
335         struct rte_tailq_entry *te;
336         struct rte_hash_list *hash_list;
337
338         if (h == NULL)
339                 return;
340
341         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
342
343         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
344
345         /* find out tailq entry */
346         TAILQ_FOREACH(te, hash_list, next) {
347                 if (te->data == (void *) h)
348                         break;
349         }
350
351         if (te == NULL) {
352                 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
353                 return;
354         }
355
356         TAILQ_REMOVE(hash_list, te, next);
357
358         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
359
360         if (h->hw_trans_mem_support)
361                 rte_free(h->local_free_slots);
362
363         if (h->add_key == ADD_KEY_MULTIWRITER)
364                 rte_free(h->multiwriter_lock);
365         rte_ring_free(h->free_slots);
366         rte_free(h->key_store);
367         rte_free(h->buckets);
368         rte_free(h);
369         rte_free(te);
370 }
371
372 hash_sig_t
373 rte_hash_hash(const struct rte_hash *h, const void *key)
374 {
375         /* calc hash result by key */
376         return h->hash_func(key, h->key_len, h->hash_func_init_val);
377 }
378
379 /* Calc the secondary hash value from the primary hash value of a given key */
380 static inline hash_sig_t
381 rte_hash_secondary_hash(const hash_sig_t primary_hash)
382 {
383         static const unsigned all_bits_shift = 12;
384         static const unsigned alt_bits_xor = 0x5bd1e995;
385
386         uint32_t tag = primary_hash >> all_bits_shift;
387
388         return primary_hash ^ ((tag + 1) * alt_bits_xor);
389 }
390
391 void
392 rte_hash_reset(struct rte_hash *h)
393 {
394         void *ptr;
395         unsigned i;
396
397         if (h == NULL)
398                 return;
399
400         memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
401         memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
402
403         /* clear the free ring */
404         while (rte_ring_dequeue(h->free_slots, &ptr) == 0)
405                 rte_pause();
406
407         /* Repopulate the free slots ring. Entry zero is reserved for key misses */
408         for (i = 1; i < h->entries + 1; i++)
409                 rte_ring_sp_enqueue(h->free_slots, (void *)((uintptr_t) i));
410
411         if (h->hw_trans_mem_support) {
412                 /* Reset local caches per lcore */
413                 for (i = 0; i < RTE_MAX_LCORE; i++)
414                         h->local_free_slots[i].len = 0;
415         }
416 }
417
418 /* Search for an entry that can be pushed to its alternative location */
419 static inline int
420 make_space_bucket(const struct rte_hash *h, struct rte_hash_bucket *bkt)
421 {
422         unsigned i, j;
423         int ret;
424         uint32_t next_bucket_idx;
425         struct rte_hash_bucket *next_bkt[RTE_HASH_BUCKET_ENTRIES];
426
427         /*
428          * Push existing item (search for bucket with space in
429          * alternative locations) to its alternative location
430          */
431         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
432                 /* Search for space in alternative locations */
433                 next_bucket_idx = bkt->sig_alt[i] & h->bucket_bitmask;
434                 next_bkt[i] = &h->buckets[next_bucket_idx];
435                 for (j = 0; j < RTE_HASH_BUCKET_ENTRIES; j++) {
436                         if (next_bkt[i]->key_idx[j] == EMPTY_SLOT)
437                                 break;
438                 }
439
440                 if (j != RTE_HASH_BUCKET_ENTRIES)
441                         break;
442         }
443
444         /* Alternative location has spare room (end of recursive function) */
445         if (i != RTE_HASH_BUCKET_ENTRIES) {
446                 next_bkt[i]->sig_alt[j] = bkt->sig_current[i];
447                 next_bkt[i]->sig_current[j] = bkt->sig_alt[i];
448                 next_bkt[i]->key_idx[j] = bkt->key_idx[i];
449                 return i;
450         }
451
452         /* Pick entry that has not been pushed yet */
453         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++)
454                 if (bkt->flag[i] == 0)
455                         break;
456
457         /* All entries have been pushed, so entry cannot be added */
458         if (i == RTE_HASH_BUCKET_ENTRIES)
459                 return -ENOSPC;
460
461         /* Set flag to indicate that this entry is going to be pushed */
462         bkt->flag[i] = 1;
463         /* Need room in alternative bucket to insert the pushed entry */
464         ret = make_space_bucket(h, next_bkt[i]);
465         /*
466          * After recursive function.
467          * Clear flags and insert the pushed entry
468          * in its alternative location if successful,
469          * or return error
470          */
471         bkt->flag[i] = 0;
472         if (ret >= 0) {
473                 next_bkt[i]->sig_alt[ret] = bkt->sig_current[i];
474                 next_bkt[i]->sig_current[ret] = bkt->sig_alt[i];
475                 next_bkt[i]->key_idx[ret] = bkt->key_idx[i];
476                 return i;
477         } else
478                 return ret;
479
480 }
481
482 /*
483  * Function called to enqueue back an index in the cache/ring,
484  * as slot has not being used and it can be used in the
485  * next addition attempt.
486  */
487 static inline void
488 enqueue_slot_back(const struct rte_hash *h,
489                 struct lcore_cache *cached_free_slots,
490                 void *slot_id)
491 {
492         if (h->hw_trans_mem_support) {
493                 cached_free_slots->objs[cached_free_slots->len] = slot_id;
494                 cached_free_slots->len++;
495         } else
496                 rte_ring_sp_enqueue(h->free_slots, slot_id);
497 }
498
499 static inline int32_t
500 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
501                                                 hash_sig_t sig, void *data)
502 {
503         hash_sig_t alt_hash;
504         uint32_t prim_bucket_idx, sec_bucket_idx;
505         unsigned i;
506         struct rte_hash_bucket *prim_bkt, *sec_bkt;
507         struct rte_hash_key *new_k, *k, *keys = h->key_store;
508         void *slot_id = NULL;
509         uint32_t new_idx;
510         int ret;
511         unsigned n_slots;
512         unsigned lcore_id;
513         struct lcore_cache *cached_free_slots = NULL;
514
515         if (h->add_key == ADD_KEY_MULTIWRITER)
516                 rte_spinlock_lock(h->multiwriter_lock);
517
518         prim_bucket_idx = sig & h->bucket_bitmask;
519         prim_bkt = &h->buckets[prim_bucket_idx];
520         rte_prefetch0(prim_bkt);
521
522         alt_hash = rte_hash_secondary_hash(sig);
523         sec_bucket_idx = alt_hash & h->bucket_bitmask;
524         sec_bkt = &h->buckets[sec_bucket_idx];
525         rte_prefetch0(sec_bkt);
526
527         /* Get a new slot for storing the new key */
528         if (h->hw_trans_mem_support) {
529                 lcore_id = rte_lcore_id();
530                 cached_free_slots = &h->local_free_slots[lcore_id];
531                 /* Try to get a free slot from the local cache */
532                 if (cached_free_slots->len == 0) {
533                         /* Need to get another burst of free slots from global ring */
534                         n_slots = rte_ring_mc_dequeue_burst(h->free_slots,
535                                         cached_free_slots->objs, LCORE_CACHE_SIZE);
536                         if (n_slots == 0)
537                                 return -ENOSPC;
538
539                         cached_free_slots->len += n_slots;
540                 }
541
542                 /* Get a free slot from the local cache */
543                 cached_free_slots->len--;
544                 slot_id = cached_free_slots->objs[cached_free_slots->len];
545         } else {
546                 if (rte_ring_sc_dequeue(h->free_slots, &slot_id) != 0)
547                         return -ENOSPC;
548         }
549
550         new_k = RTE_PTR_ADD(keys, (uintptr_t)slot_id * h->key_entry_size);
551         rte_prefetch0(new_k);
552         new_idx = (uint32_t)((uintptr_t) slot_id);
553
554         /* Check if key is already inserted in primary location */
555         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
556                 if (prim_bkt->sig_current[i] == sig &&
557                                 prim_bkt->sig_alt[i] == alt_hash) {
558                         k = (struct rte_hash_key *) ((char *)keys +
559                                         prim_bkt->key_idx[i] * h->key_entry_size);
560                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
561                                 /* Enqueue index of free slot back in the ring. */
562                                 enqueue_slot_back(h, cached_free_slots, slot_id);
563                                 /* Update data */
564                                 k->pdata = data;
565                                 /*
566                                  * Return index where key is stored,
567                                  * substracting the first dummy index
568                                  */
569                                 return prim_bkt->key_idx[i] - 1;
570                         }
571                 }
572         }
573
574         /* Check if key is already inserted in secondary location */
575         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
576                 if (sec_bkt->sig_alt[i] == sig &&
577                                 sec_bkt->sig_current[i] == alt_hash) {
578                         k = (struct rte_hash_key *) ((char *)keys +
579                                         sec_bkt->key_idx[i] * h->key_entry_size);
580                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
581                                 /* Enqueue index of free slot back in the ring. */
582                                 enqueue_slot_back(h, cached_free_slots, slot_id);
583                                 /* Update data */
584                                 k->pdata = data;
585                                 /*
586                                  * Return index where key is stored,
587                                  * substracting the first dummy index
588                                  */
589                                 return sec_bkt->key_idx[i] - 1;
590                         }
591                 }
592         }
593
594         /* Copy key */
595         rte_memcpy(new_k->key, key, h->key_len);
596         new_k->pdata = data;
597
598 #if defined(RTE_ARCH_X86) /* currently only x86 support HTM */
599         if (h->add_key == ADD_KEY_MULTIWRITER_TM) {
600                 ret = rte_hash_cuckoo_insert_mw_tm(prim_bkt,
601                                 sig, alt_hash, new_idx);
602                 if (ret >= 0)
603                         return new_idx - 1;
604
605                 /* Primary bucket full, need to make space for new entry */
606                 ret = rte_hash_cuckoo_make_space_mw_tm(h, prim_bkt, sig,
607                                                         alt_hash, new_idx);
608
609                 if (ret >= 0)
610                         return new_idx - 1;
611
612                 /* Also search secondary bucket to get better occupancy */
613                 ret = rte_hash_cuckoo_make_space_mw_tm(h, sec_bkt, sig,
614                                                         alt_hash, new_idx);
615
616                 if (ret >= 0)
617                         return new_idx - 1;
618         } else {
619 #endif
620                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
621                         /* Check if slot is available */
622                         if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
623                                 prim_bkt->sig_current[i] = sig;
624                                 prim_bkt->sig_alt[i] = alt_hash;
625                                 prim_bkt->key_idx[i] = new_idx;
626                                 break;
627                         }
628                 }
629
630                 if (i != RTE_HASH_BUCKET_ENTRIES) {
631                         if (h->add_key == ADD_KEY_MULTIWRITER)
632                                 rte_spinlock_unlock(h->multiwriter_lock);
633                         return new_idx - 1;
634                 }
635
636                 /* Primary bucket full, need to make space for new entry
637                  * After recursive function.
638                  * Insert the new entry in the position of the pushed entry
639                  * if successful or return error and
640                  * store the new slot back in the ring
641                  */
642                 ret = make_space_bucket(h, prim_bkt);
643                 if (ret >= 0) {
644                         prim_bkt->sig_current[ret] = sig;
645                         prim_bkt->sig_alt[ret] = alt_hash;
646                         prim_bkt->key_idx[ret] = new_idx;
647                         if (h->add_key == ADD_KEY_MULTIWRITER)
648                                 rte_spinlock_unlock(h->multiwriter_lock);
649                         return new_idx - 1;
650                 }
651 #if defined(RTE_ARCH_X86)
652         }
653 #endif
654         /* Error in addition, store new slot back in the ring and return error */
655         enqueue_slot_back(h, cached_free_slots, (void *)((uintptr_t) new_idx));
656
657         if (h->add_key == ADD_KEY_MULTIWRITER)
658                 rte_spinlock_unlock(h->multiwriter_lock);
659         return ret;
660 }
661
662 int32_t
663 rte_hash_add_key_with_hash(const struct rte_hash *h,
664                         const void *key, hash_sig_t sig)
665 {
666         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
667         return __rte_hash_add_key_with_hash(h, key, sig, 0);
668 }
669
670 int32_t
671 rte_hash_add_key(const struct rte_hash *h, const void *key)
672 {
673         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
674         return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
675 }
676
677 int
678 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
679                         const void *key, hash_sig_t sig, void *data)
680 {
681         int ret;
682
683         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
684         ret = __rte_hash_add_key_with_hash(h, key, sig, data);
685         if (ret >= 0)
686                 return 0;
687         else
688                 return ret;
689 }
690
691 int
692 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
693 {
694         int ret;
695
696         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
697
698         ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
699         if (ret >= 0)
700                 return 0;
701         else
702                 return ret;
703 }
704 static inline int32_t
705 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
706                                         hash_sig_t sig, void **data)
707 {
708         uint32_t bucket_idx;
709         hash_sig_t alt_hash;
710         unsigned i;
711         struct rte_hash_bucket *bkt;
712         struct rte_hash_key *k, *keys = h->key_store;
713
714         bucket_idx = sig & h->bucket_bitmask;
715         bkt = &h->buckets[bucket_idx];
716
717         /* Check if key is in primary location */
718         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
719                 if (bkt->sig_current[i] == sig &&
720                                 bkt->key_idx[i] != EMPTY_SLOT) {
721                         k = (struct rte_hash_key *) ((char *)keys +
722                                         bkt->key_idx[i] * h->key_entry_size);
723                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
724                                 if (data != NULL)
725                                         *data = k->pdata;
726                                 /*
727                                  * Return index where key is stored,
728                                  * substracting the first dummy index
729                                  */
730                                 return bkt->key_idx[i] - 1;
731                         }
732                 }
733         }
734
735         /* Calculate secondary hash */
736         alt_hash = rte_hash_secondary_hash(sig);
737         bucket_idx = alt_hash & h->bucket_bitmask;
738         bkt = &h->buckets[bucket_idx];
739
740         /* Check if key is in secondary location */
741         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
742                 if (bkt->sig_current[i] == alt_hash &&
743                                 bkt->sig_alt[i] == sig) {
744                         k = (struct rte_hash_key *) ((char *)keys +
745                                         bkt->key_idx[i] * h->key_entry_size);
746                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
747                                 if (data != NULL)
748                                         *data = k->pdata;
749                                 /*
750                                  * Return index where key is stored,
751                                  * substracting the first dummy index
752                                  */
753                                 return bkt->key_idx[i] - 1;
754                         }
755                 }
756         }
757
758         return -ENOENT;
759 }
760
761 int32_t
762 rte_hash_lookup_with_hash(const struct rte_hash *h,
763                         const void *key, hash_sig_t sig)
764 {
765         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
766         return __rte_hash_lookup_with_hash(h, key, sig, NULL);
767 }
768
769 int32_t
770 rte_hash_lookup(const struct rte_hash *h, const void *key)
771 {
772         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
773         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
774 }
775
776 int
777 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
778                         const void *key, hash_sig_t sig, void **data)
779 {
780         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
781         return __rte_hash_lookup_with_hash(h, key, sig, data);
782 }
783
784 int
785 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
786 {
787         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
788         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
789 }
790
791 static inline void
792 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
793 {
794         unsigned lcore_id, n_slots;
795         struct lcore_cache *cached_free_slots;
796
797         bkt->sig_current[i] = NULL_SIGNATURE;
798         bkt->sig_alt[i] = NULL_SIGNATURE;
799         if (h->hw_trans_mem_support) {
800                 lcore_id = rte_lcore_id();
801                 cached_free_slots = &h->local_free_slots[lcore_id];
802                 /* Cache full, need to free it. */
803                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
804                         /* Need to enqueue the free slots in global ring. */
805                         n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
806                                                 cached_free_slots->objs,
807                                                 LCORE_CACHE_SIZE);
808                         cached_free_slots->len -= n_slots;
809                 }
810                 /* Put index of new free slot in cache. */
811                 cached_free_slots->objs[cached_free_slots->len] =
812                                 (void *)((uintptr_t)bkt->key_idx[i]);
813                 cached_free_slots->len++;
814         } else {
815                 rte_ring_sp_enqueue(h->free_slots,
816                                 (void *)((uintptr_t)bkt->key_idx[i]));
817         }
818 }
819
820 static inline int32_t
821 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
822                                                 hash_sig_t sig)
823 {
824         uint32_t bucket_idx;
825         hash_sig_t alt_hash;
826         unsigned i;
827         struct rte_hash_bucket *bkt;
828         struct rte_hash_key *k, *keys = h->key_store;
829         int32_t ret;
830
831         bucket_idx = sig & h->bucket_bitmask;
832         bkt = &h->buckets[bucket_idx];
833
834         /* Check if key is in primary location */
835         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
836                 if (bkt->sig_current[i] == sig &&
837                                 bkt->key_idx[i] != EMPTY_SLOT) {
838                         k = (struct rte_hash_key *) ((char *)keys +
839                                         bkt->key_idx[i] * h->key_entry_size);
840                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
841                                 remove_entry(h, bkt, i);
842
843                                 /*
844                                  * Return index where key is stored,
845                                  * substracting the first dummy index
846                                  */
847                                 ret = bkt->key_idx[i] - 1;
848                                 bkt->key_idx[i] = EMPTY_SLOT;
849                                 return ret;
850                         }
851                 }
852         }
853
854         /* Calculate secondary hash */
855         alt_hash = rte_hash_secondary_hash(sig);
856         bucket_idx = alt_hash & h->bucket_bitmask;
857         bkt = &h->buckets[bucket_idx];
858
859         /* Check if key is in secondary location */
860         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
861                 if (bkt->sig_current[i] == alt_hash &&
862                                 bkt->key_idx[i] != EMPTY_SLOT) {
863                         k = (struct rte_hash_key *) ((char *)keys +
864                                         bkt->key_idx[i] * h->key_entry_size);
865                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
866                                 remove_entry(h, bkt, i);
867
868                                 /*
869                                  * Return index where key is stored,
870                                  * substracting the first dummy index
871                                  */
872                                 ret = bkt->key_idx[i] - 1;
873                                 bkt->key_idx[i] = EMPTY_SLOT;
874                                 return ret;
875                         }
876                 }
877         }
878
879         return -ENOENT;
880 }
881
882 int32_t
883 rte_hash_del_key_with_hash(const struct rte_hash *h,
884                         const void *key, hash_sig_t sig)
885 {
886         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
887         return __rte_hash_del_key_with_hash(h, key, sig);
888 }
889
890 int32_t
891 rte_hash_del_key(const struct rte_hash *h, const void *key)
892 {
893         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
894         return __rte_hash_del_key_with_hash(h, key, rte_hash_hash(h, key));
895 }
896
897 int
898 rte_hash_get_key_with_position(const struct rte_hash *h, const int32_t position,
899                                void **key)
900 {
901         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
902
903         struct rte_hash_key *k, *keys = h->key_store;
904         k = (struct rte_hash_key *) ((char *) keys + (position + 1) *
905                                      h->key_entry_size);
906         *key = k->key;
907
908         if (position !=
909             __rte_hash_lookup_with_hash(h, *key, rte_hash_hash(h, *key),
910                                         NULL)) {
911                 return -ENOENT;
912         }
913
914         return 0;
915 }
916
917 static inline void
918 compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
919                         const struct rte_hash_bucket *prim_bkt,
920                         const struct rte_hash_bucket *sec_bkt,
921                         hash_sig_t prim_hash, hash_sig_t sec_hash,
922                         enum rte_hash_sig_compare_function sig_cmp_fn)
923 {
924         unsigned int i;
925
926         switch (sig_cmp_fn) {
927 #ifdef RTE_MACHINE_CPUFLAG_AVX2
928         case RTE_HASH_COMPARE_AVX2:
929                 *prim_hash_matches = _mm256_movemask_ps((__m256)_mm256_cmpeq_epi32(
930                                 _mm256_load_si256(
931                                         (__m256i const *)prim_bkt->sig_current),
932                                 _mm256_set1_epi32(prim_hash)));
933                 *sec_hash_matches = _mm256_movemask_ps((__m256)_mm256_cmpeq_epi32(
934                                 _mm256_load_si256(
935                                         (__m256i const *)sec_bkt->sig_current),
936                                 _mm256_set1_epi32(sec_hash)));
937                 break;
938 #endif
939 #ifdef RTE_MACHINE_CPUFLAG_SSE2
940         case RTE_HASH_COMPARE_SSE:
941                 /* Compare the first 4 signatures in the bucket */
942                 *prim_hash_matches = _mm_movemask_ps((__m128)_mm_cmpeq_epi16(
943                                 _mm_load_si128(
944                                         (__m128i const *)prim_bkt->sig_current),
945                                 _mm_set1_epi32(prim_hash)));
946                 *prim_hash_matches |= (_mm_movemask_ps((__m128)_mm_cmpeq_epi16(
947                                 _mm_load_si128(
948                                         (__m128i const *)&prim_bkt->sig_current[4]),
949                                 _mm_set1_epi32(prim_hash)))) << 4;
950                 /* Compare the first 4 signatures in the bucket */
951                 *sec_hash_matches = _mm_movemask_ps((__m128)_mm_cmpeq_epi16(
952                                 _mm_load_si128(
953                                         (__m128i const *)sec_bkt->sig_current),
954                                 _mm_set1_epi32(sec_hash)));
955                 *sec_hash_matches |= (_mm_movemask_ps((__m128)_mm_cmpeq_epi16(
956                                 _mm_load_si128(
957                                         (__m128i const *)&sec_bkt->sig_current[4]),
958                                 _mm_set1_epi32(sec_hash)))) << 4;
959                 break;
960 #endif
961         default:
962                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
963                         *prim_hash_matches |=
964                                 ((prim_hash == prim_bkt->sig_current[i]) << i);
965                         *sec_hash_matches |=
966                                 ((sec_hash == sec_bkt->sig_current[i]) << i);
967                 }
968         }
969
970 }
971
972 #define PREFETCH_OFFSET 4
973 static inline void
974 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
975                         int32_t num_keys, int32_t *positions,
976                         uint64_t *hit_mask, void *data[])
977 {
978         uint64_t hits = 0;
979         int32_t i;
980         uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
981         uint32_t sec_hash[RTE_HASH_LOOKUP_BULK_MAX];
982         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
983         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
984         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
985         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
986
987         /* Prefetch first keys */
988         for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
989                 rte_prefetch0(keys[i]);
990
991         /*
992          * Prefetch rest of the keys, calculate primary and
993          * secondary bucket and prefetch them
994          */
995         for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
996                 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
997
998                 prim_hash[i] = rte_hash_hash(h, keys[i]);
999                 sec_hash[i] = rte_hash_secondary_hash(prim_hash[i]);
1000
1001                 primary_bkt[i] = &h->buckets[prim_hash[i] & h->bucket_bitmask];
1002                 secondary_bkt[i] = &h->buckets[sec_hash[i] & h->bucket_bitmask];
1003
1004                 rte_prefetch0(primary_bkt[i]);
1005                 rte_prefetch0(secondary_bkt[i]);
1006         }
1007
1008         /* Calculate and prefetch rest of the buckets */
1009         for (; i < num_keys; i++) {
1010                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1011                 sec_hash[i] = rte_hash_secondary_hash(prim_hash[i]);
1012
1013                 primary_bkt[i] = &h->buckets[prim_hash[i] & h->bucket_bitmask];
1014                 secondary_bkt[i] = &h->buckets[sec_hash[i] & h->bucket_bitmask];
1015
1016                 rte_prefetch0(primary_bkt[i]);
1017                 rte_prefetch0(secondary_bkt[i]);
1018         }
1019
1020         /* Compare signatures and prefetch key slot of first hit */
1021         for (i = 0; i < num_keys; i++) {
1022                 compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1023                                 primary_bkt[i], secondary_bkt[i],
1024                                 prim_hash[i], sec_hash[i], h->sig_cmp_fn);
1025
1026                 if (prim_hitmask[i]) {
1027                         uint32_t first_hit = __builtin_ctzl(prim_hitmask[i]);
1028                         uint32_t key_idx = primary_bkt[i]->key_idx[first_hit];
1029                         const struct rte_hash_key *key_slot =
1030                                 (const struct rte_hash_key *)(
1031                                 (const char *)h->key_store +
1032                                 key_idx * h->key_entry_size);
1033                         rte_prefetch0(key_slot);
1034                         continue;
1035                 }
1036
1037                 if (sec_hitmask[i]) {
1038                         uint32_t first_hit = __builtin_ctzl(sec_hitmask[i]);
1039                         uint32_t key_idx = secondary_bkt[i]->key_idx[first_hit];
1040                         const struct rte_hash_key *key_slot =
1041                                 (const struct rte_hash_key *)(
1042                                 (const char *)h->key_store +
1043                                 key_idx * h->key_entry_size);
1044                         rte_prefetch0(key_slot);
1045                 }
1046         }
1047
1048         /* Compare keys, first hits in primary first */
1049         for (i = 0; i < num_keys; i++) {
1050                 positions[i] = -ENOENT;
1051                 while (prim_hitmask[i]) {
1052                         uint32_t hit_index = __builtin_ctzl(prim_hitmask[i]);
1053
1054                         uint32_t key_idx = primary_bkt[i]->key_idx[hit_index];
1055                         const struct rte_hash_key *key_slot =
1056                                 (const struct rte_hash_key *)(
1057                                 (const char *)h->key_store +
1058                                 key_idx * h->key_entry_size);
1059                         /*
1060                          * If key index is 0, do not compare key,
1061                          * as it is checking the dummy slot
1062                          */
1063                         if (!!key_idx & !rte_hash_cmp_eq(key_slot->key, keys[i], h)) {
1064                                 if (data != NULL)
1065                                         data[i] = key_slot->pdata;
1066
1067                                 hits |= 1ULL << i;
1068                                 positions[i] = key_idx - 1;
1069                                 goto next_key;
1070                         }
1071                         prim_hitmask[i] &= ~(1 << (hit_index));
1072                 }
1073
1074                 while (sec_hitmask[i]) {
1075                         uint32_t hit_index = __builtin_ctzl(sec_hitmask[i]);
1076
1077                         uint32_t key_idx = secondary_bkt[i]->key_idx[hit_index];
1078                         const struct rte_hash_key *key_slot =
1079                                 (const struct rte_hash_key *)(
1080                                 (const char *)h->key_store +
1081                                 key_idx * h->key_entry_size);
1082                         /*
1083                          * If key index is 0, do not compare key,
1084                          * as it is checking the dummy slot
1085                          */
1086
1087                         if (!!key_idx & !rte_hash_cmp_eq(key_slot->key, keys[i], h)) {
1088                                 if (data != NULL)
1089                                         data[i] = key_slot->pdata;
1090
1091                                 hits |= 1ULL << i;
1092                                 positions[i] = key_idx - 1;
1093                                 goto next_key;
1094                         }
1095                         sec_hitmask[i] &= ~(1 << (hit_index));
1096                 }
1097
1098 next_key:
1099                 continue;
1100         }
1101
1102         if (hit_mask != NULL)
1103                 *hit_mask = hits;
1104 }
1105
1106 int
1107 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
1108                       uint32_t num_keys, int32_t *positions)
1109 {
1110         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
1111                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
1112                         (positions == NULL)), -EINVAL);
1113
1114         __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
1115         return 0;
1116 }
1117
1118 int
1119 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
1120                       uint32_t num_keys, uint64_t *hit_mask, void *data[])
1121 {
1122         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
1123                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
1124                         (hit_mask == NULL)), -EINVAL);
1125
1126         int32_t positions[num_keys];
1127
1128         __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
1129
1130         /* Return number of hits */
1131         return __builtin_popcountl(*hit_mask);
1132 }
1133
1134 int32_t
1135 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
1136 {
1137         uint32_t bucket_idx, idx, position;
1138         struct rte_hash_key *next_key;
1139
1140         RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
1141
1142         const uint32_t total_entries = h->num_buckets * RTE_HASH_BUCKET_ENTRIES;
1143         /* Out of bounds */
1144         if (*next >= total_entries)
1145                 return -ENOENT;
1146
1147         /* Calculate bucket and index of current iterator */
1148         bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
1149         idx = *next % RTE_HASH_BUCKET_ENTRIES;
1150
1151         /* If current position is empty, go to the next one */
1152         while (h->buckets[bucket_idx].key_idx[idx] == EMPTY_SLOT) {
1153                 (*next)++;
1154                 /* End of table */
1155                 if (*next == total_entries)
1156                         return -ENOENT;
1157                 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
1158                 idx = *next % RTE_HASH_BUCKET_ENTRIES;
1159         }
1160
1161         /* Get position of entry in key table */
1162         position = h->buckets[bucket_idx].key_idx[idx];
1163         next_key = (struct rte_hash_key *) ((char *)h->key_store +
1164                                 position * h->key_entry_size);
1165         /* Return key and data */
1166         *key = next_key->key;
1167         *data = next_key->pdata;
1168
1169         /* Increment iterator */
1170         (*next)++;
1171
1172         return position - 1;
1173 }