1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
3 * Copyright(c) 2018 Arm Limited
11 #include <sys/queue.h>
13 #include <rte_common.h>
14 #include <rte_memory.h> /* for definition of RTE_CACHE_LINE_SIZE */
16 #include <rte_prefetch.h>
17 #include <rte_branch_prediction.h>
18 #include <rte_malloc.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>
28 #include <rte_compat.h>
30 #include <rte_tailq.h>
33 #include "rte_cuckoo_hash.h"
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)
40 TAILQ_HEAD(rte_hash_list, rte_tailq_entry);
42 static struct rte_tailq_elem rte_hash_tailq = {
45 EAL_REGISTER_TAILQ(rte_hash_tailq)
48 rte_hash_find_existing(const char *name)
50 struct rte_hash *h = NULL;
51 struct rte_tailq_entry *te;
52 struct rte_hash_list *hash_list;
54 hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
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)
62 rte_mcfg_tailq_read_unlock();
71 static inline struct rte_hash_bucket *
72 rte_hash_get_last_bkt(struct rte_hash_bucket *lst_bkt)
74 while (lst_bkt->next != NULL)
75 lst_bkt = lst_bkt->next;
79 void rte_hash_set_cmp_func(struct rte_hash *h, rte_hash_cmp_eq_t func)
81 h->cmp_jump_table_idx = KEY_CUSTOM;
82 h->rte_hash_custom_cmp_eq = func;
86 rte_hash_cmp_eq(const void *key1, const void *key2, const struct rte_hash *h)
88 if (h->cmp_jump_table_idx == KEY_CUSTOM)
89 return h->rte_hash_custom_cmp_eq(key1, key2, h->key_len);
91 return cmp_jump_table[h->cmp_jump_table_idx](key1, key2, h->key_len);
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.
105 static inline uint16_t
106 get_short_sig(const hash_sig_t hash)
111 static inline uint32_t
112 get_prim_bucket_index(const struct rte_hash *h, const hash_sig_t hash)
114 return hash & h->bucket_bitmask;
117 static inline uint32_t
118 get_alt_bucket_index(const struct rte_hash *h,
119 uint32_t cur_bkt_idx, uint16_t sig)
121 return (cur_bkt_idx ^ sig) & h->bucket_bitmask;
125 rte_hash_create(const struct rte_hash_parameters *params)
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];
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;
140 unsigned int hw_trans_mem_support = 0, use_local_cache = 0;
141 unsigned int ext_table_support = 0;
142 unsigned int readwrite_concur_support = 0;
143 unsigned int writer_takes_lock = 0;
144 unsigned int no_free_on_del = 0;
145 uint32_t *ext_bkt_to_free = NULL;
146 uint32_t *tbl_chng_cnt = NULL;
147 unsigned int readwrite_concur_lf_support = 0;
149 rte_hash_function default_hash_func = (rte_hash_function)rte_jhash;
151 hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
153 if (params == NULL) {
154 RTE_LOG(ERR, HASH, "rte_hash_create has no parameters\n");
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)) {
163 RTE_LOG(ERR, HASH, "rte_hash_create has invalid parameters\n");
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)) {
171 RTE_LOG(ERR, HASH, "rte_hash_create: choose rw concurrency or "
172 "rw concurrency lock free\n");
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;
180 if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
182 writer_takes_lock = 1;
185 if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) {
186 readwrite_concur_support = 1;
187 writer_takes_lock = 1;
190 if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_EXT_TABLE)
191 ext_table_support = 1;
193 if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_NO_FREE_ON_DEL)
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 */
202 /* Store all keys and leave the first entry as a dummy entry for lookup_bulk */
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
209 num_key_slots = params->entries + (RTE_MAX_LCORE - 1) *
210 (LCORE_CACHE_SIZE - 1) + 1;
212 num_key_slots = params->entries + 1;
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(ring_name, rte_align32pow2(num_key_slots),
217 params->socket_id, 0);
219 RTE_LOG(ERR, HASH, "memory allocation failed\n");
223 const uint32_t num_buckets = rte_align32pow2(params->entries) /
224 RTE_HASH_BUCKET_ENTRIES;
226 /* Create ring for extendable buckets. */
227 if (ext_table_support) {
228 snprintf(ext_ring_name, sizeof(ext_ring_name), "HT_EXT_%s",
230 r_ext = rte_ring_create(ext_ring_name,
231 rte_align32pow2(num_buckets + 1),
232 params->socket_id, 0);
235 RTE_LOG(ERR, HASH, "ext buckets memory allocation "
241 snprintf(hash_name, sizeof(hash_name), "HT_%s", params->name);
243 rte_mcfg_tailq_write_lock();
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)
259 te = rte_zmalloc("HASH_TAILQ_ENTRY", sizeof(*te), 0);
261 RTE_LOG(ERR, HASH, "tailq entry allocation failed\n");
265 h = (struct rte_hash *)rte_zmalloc_socket(hash_name, sizeof(struct rte_hash),
266 RTE_CACHE_LINE_SIZE, params->socket_id);
269 RTE_LOG(ERR, HASH, "memory allocation failed\n");
273 buckets = rte_zmalloc_socket(NULL,
274 num_buckets * sizeof(struct rte_hash_bucket),
275 RTE_CACHE_LINE_SIZE, params->socket_id);
277 if (buckets == NULL) {
278 RTE_LOG(ERR, HASH, "buckets memory allocation failed\n");
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 "
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
297 for (i = 1; i <= num_buckets; i++)
298 rte_ring_sp_enqueue(r_ext, (void *)((uintptr_t) i));
300 if (readwrite_concur_lf_support) {
301 ext_bkt_to_free = rte_zmalloc(NULL, sizeof(uint32_t) *
303 if (ext_bkt_to_free == NULL) {
304 RTE_LOG(ERR, HASH, "ext bkt to free memory allocation "
311 const uint32_t key_entry_size =
312 RTE_ALIGN(sizeof(struct rte_hash_key) + params->key_len,
314 const uint64_t key_tbl_size = (uint64_t) key_entry_size * num_key_slots;
316 k = rte_zmalloc_socket(NULL, key_tbl_size,
317 RTE_CACHE_LINE_SIZE, params->socket_id);
320 RTE_LOG(ERR, HASH, "memory allocation failed\n");
324 tbl_chng_cnt = rte_zmalloc_socket(NULL, sizeof(uint32_t),
325 RTE_CACHE_LINE_SIZE, params->socket_id);
327 if (tbl_chng_cnt == NULL) {
328 RTE_LOG(ERR, HASH, "memory allocation failed\n");
333 * If x86 architecture is used, select appropriate compare function,
334 * which may use x86 intrinsics, otherwise use memcmp
336 #if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
337 /* Select function to compare keys */
338 switch (params->key_len) {
340 h->cmp_jump_table_idx = KEY_16_BYTES;
343 h->cmp_jump_table_idx = KEY_32_BYTES;
346 h->cmp_jump_table_idx = KEY_48_BYTES;
349 h->cmp_jump_table_idx = KEY_64_BYTES;
352 h->cmp_jump_table_idx = KEY_80_BYTES;
355 h->cmp_jump_table_idx = KEY_96_BYTES;
358 h->cmp_jump_table_idx = KEY_112_BYTES;
361 h->cmp_jump_table_idx = KEY_128_BYTES;
364 /* If key is not multiple of 16, use generic memcmp */
365 h->cmp_jump_table_idx = KEY_OTHER_BYTES;
368 h->cmp_jump_table_idx = KEY_OTHER_BYTES;
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);
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;
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;
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;
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;
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;
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;
420 h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
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
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)
432 rte_rwlock_init(h->readwrite_lock);
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(r, (void *)((uintptr_t) i));
439 te->data = (void *) h;
440 TAILQ_INSERT_TAIL(hash_list, te, next);
441 rte_mcfg_tailq_write_unlock();
445 rte_mcfg_tailq_write_unlock();
448 rte_ring_free(r_ext);
452 rte_free(buckets_ext);
454 rte_free(tbl_chng_cnt);
455 rte_free(ext_bkt_to_free);
460 rte_hash_free(struct rte_hash *h)
462 struct rte_tailq_entry *te;
463 struct rte_hash_list *hash_list;
468 hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
470 rte_mcfg_tailq_write_lock();
472 /* find out tailq entry */
473 TAILQ_FOREACH(te, hash_list, next) {
474 if (te->data == (void *) h)
479 rte_mcfg_tailq_write_unlock();
483 TAILQ_REMOVE(hash_list, te, next);
485 rte_mcfg_tailq_write_unlock();
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);
503 rte_hash_hash(const struct rte_hash *h, const void *key)
505 /* calc hash result by key */
506 return h->hash_func(key, h->key_len, h->hash_func_init_val);
510 rte_hash_count(const struct rte_hash *h)
512 uint32_t tot_ring_cnt, cached_cnt = 0;
518 if (h->use_local_cache) {
519 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
520 (LCORE_CACHE_SIZE - 1);
521 for (i = 0; i < RTE_MAX_LCORE; i++)
522 cached_cnt += h->local_free_slots[i].len;
524 ret = tot_ring_cnt - rte_ring_count(h->free_slots) -
527 tot_ring_cnt = h->entries;
528 ret = tot_ring_cnt - rte_ring_count(h->free_slots);
533 /* Read write locks implemented using rte_rwlock */
535 __hash_rw_writer_lock(const struct rte_hash *h)
537 if (h->writer_takes_lock && h->hw_trans_mem_support)
538 rte_rwlock_write_lock_tm(h->readwrite_lock);
539 else if (h->writer_takes_lock)
540 rte_rwlock_write_lock(h->readwrite_lock);
544 __hash_rw_reader_lock(const struct rte_hash *h)
546 if (h->readwrite_concur_support && h->hw_trans_mem_support)
547 rte_rwlock_read_lock_tm(h->readwrite_lock);
548 else if (h->readwrite_concur_support)
549 rte_rwlock_read_lock(h->readwrite_lock);
553 __hash_rw_writer_unlock(const struct rte_hash *h)
555 if (h->writer_takes_lock && h->hw_trans_mem_support)
556 rte_rwlock_write_unlock_tm(h->readwrite_lock);
557 else if (h->writer_takes_lock)
558 rte_rwlock_write_unlock(h->readwrite_lock);
562 __hash_rw_reader_unlock(const struct rte_hash *h)
564 if (h->readwrite_concur_support && h->hw_trans_mem_support)
565 rte_rwlock_read_unlock_tm(h->readwrite_lock);
566 else if (h->readwrite_concur_support)
567 rte_rwlock_read_unlock(h->readwrite_lock);
571 rte_hash_reset(struct rte_hash *h)
573 uint32_t tot_ring_cnt, i;
578 __hash_rw_writer_lock(h);
579 memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
580 memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
581 *h->tbl_chng_cnt = 0;
583 /* reset the free ring */
584 rte_ring_reset(h->free_slots);
586 /* flush free extendable bucket ring and memory */
587 if (h->ext_table_support) {
588 memset(h->buckets_ext, 0, h->num_buckets *
589 sizeof(struct rte_hash_bucket));
590 rte_ring_reset(h->free_ext_bkts);
593 /* Repopulate the free slots ring. Entry zero is reserved for key misses */
594 if (h->use_local_cache)
595 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
596 (LCORE_CACHE_SIZE - 1);
598 tot_ring_cnt = h->entries;
600 for (i = 1; i < tot_ring_cnt + 1; i++)
601 rte_ring_sp_enqueue(h->free_slots, (void *)((uintptr_t) i));
603 /* Repopulate the free ext bkt ring. */
604 if (h->ext_table_support) {
605 for (i = 1; i <= h->num_buckets; i++)
606 rte_ring_sp_enqueue(h->free_ext_bkts,
607 (void *)((uintptr_t) i));
610 if (h->use_local_cache) {
611 /* Reset local caches per lcore */
612 for (i = 0; i < RTE_MAX_LCORE; i++)
613 h->local_free_slots[i].len = 0;
615 __hash_rw_writer_unlock(h);
619 * Function called to enqueue back an index in the cache/ring,
620 * as slot has not being used and it can be used in the
621 * next addition attempt.
624 enqueue_slot_back(const struct rte_hash *h,
625 struct lcore_cache *cached_free_slots,
628 if (h->use_local_cache) {
629 cached_free_slots->objs[cached_free_slots->len] = slot_id;
630 cached_free_slots->len++;
632 rte_ring_sp_enqueue(h->free_slots, slot_id);
635 /* Search a key from bucket and update its data.
636 * Writer holds the lock before calling this.
638 static inline int32_t
639 search_and_update(const struct rte_hash *h, void *data, const void *key,
640 struct rte_hash_bucket *bkt, uint16_t sig)
643 struct rte_hash_key *k, *keys = h->key_store;
645 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
646 if (bkt->sig_current[i] == sig) {
647 k = (struct rte_hash_key *) ((char *)keys +
648 bkt->key_idx[i] * h->key_entry_size);
649 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
650 /* The store to application data at *data
651 * should not leak after the store to pdata
652 * in the key store. i.e. pdata is the guard
653 * variable. Release the application data
656 __atomic_store_n(&k->pdata,
660 * Return index where key is stored,
661 * subtracting the first dummy index
663 return bkt->key_idx[i] - 1;
670 /* Only tries to insert at one bucket (@prim_bkt) without trying to push
672 * return 1 if matching existing key, return 0 if succeeds, return -1 for no
675 static inline int32_t
676 rte_hash_cuckoo_insert_mw(const struct rte_hash *h,
677 struct rte_hash_bucket *prim_bkt,
678 struct rte_hash_bucket *sec_bkt,
679 const struct rte_hash_key *key, void *data,
680 uint16_t sig, uint32_t new_idx,
684 struct rte_hash_bucket *cur_bkt;
687 __hash_rw_writer_lock(h);
688 /* Check if key was inserted after last check but before this
689 * protected region in case of inserting duplicated keys.
691 ret = search_and_update(h, data, key, prim_bkt, sig);
693 __hash_rw_writer_unlock(h);
698 FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
699 ret = search_and_update(h, data, key, cur_bkt, sig);
701 __hash_rw_writer_unlock(h);
707 /* Insert new entry if there is room in the primary
710 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
711 /* Check if slot is available */
712 if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
713 prim_bkt->sig_current[i] = sig;
714 /* Store to signature and key should not
715 * leak after the store to key_idx. i.e.
716 * key_idx is the guard variable for signature
719 __atomic_store_n(&prim_bkt->key_idx[i],
725 __hash_rw_writer_unlock(h);
727 if (i != RTE_HASH_BUCKET_ENTRIES)
734 /* Shift buckets along provided cuckoo_path (@leaf and @leaf_slot) and fill
735 * the path head with new entry (sig, alt_hash, new_idx)
736 * return 1 if matched key found, return -1 if cuckoo path invalided and fail,
737 * return 0 if succeeds.
740 rte_hash_cuckoo_move_insert_mw(const struct rte_hash *h,
741 struct rte_hash_bucket *bkt,
742 struct rte_hash_bucket *alt_bkt,
743 const struct rte_hash_key *key, void *data,
744 struct queue_node *leaf, uint32_t leaf_slot,
745 uint16_t sig, uint32_t new_idx,
748 uint32_t prev_alt_bkt_idx;
749 struct rte_hash_bucket *cur_bkt;
750 struct queue_node *prev_node, *curr_node = leaf;
751 struct rte_hash_bucket *prev_bkt, *curr_bkt = leaf->bkt;
752 uint32_t prev_slot, curr_slot = leaf_slot;
755 __hash_rw_writer_lock(h);
757 /* In case empty slot was gone before entering protected region */
758 if (curr_bkt->key_idx[curr_slot] != EMPTY_SLOT) {
759 __hash_rw_writer_unlock(h);
763 /* Check if key was inserted after last check but before this
766 ret = search_and_update(h, data, key, bkt, sig);
768 __hash_rw_writer_unlock(h);
773 FOR_EACH_BUCKET(cur_bkt, alt_bkt) {
774 ret = search_and_update(h, data, key, cur_bkt, sig);
776 __hash_rw_writer_unlock(h);
782 while (likely(curr_node->prev != NULL)) {
783 prev_node = curr_node->prev;
784 prev_bkt = prev_node->bkt;
785 prev_slot = curr_node->prev_slot;
787 prev_alt_bkt_idx = get_alt_bucket_index(h,
788 prev_node->cur_bkt_idx,
789 prev_bkt->sig_current[prev_slot]);
791 if (unlikely(&h->buckets[prev_alt_bkt_idx]
793 /* revert it to empty, otherwise duplicated keys */
794 __atomic_store_n(&curr_bkt->key_idx[curr_slot],
797 __hash_rw_writer_unlock(h);
801 if (h->readwrite_concur_lf_support) {
802 /* Inform the previous move. The current move need
803 * not be informed now as the current bucket entry
804 * is present in both primary and secondary.
805 * Since there is one writer, load acquires on
806 * tbl_chng_cnt are not required.
808 __atomic_store_n(h->tbl_chng_cnt,
809 *h->tbl_chng_cnt + 1,
811 /* The store to sig_current should not
812 * move above the store to tbl_chng_cnt.
814 __atomic_thread_fence(__ATOMIC_RELEASE);
817 /* Need to swap current/alt sig to allow later
818 * Cuckoo insert to move elements back to its
819 * primary bucket if available
821 curr_bkt->sig_current[curr_slot] =
822 prev_bkt->sig_current[prev_slot];
823 /* Release the updated bucket entry */
824 __atomic_store_n(&curr_bkt->key_idx[curr_slot],
825 prev_bkt->key_idx[prev_slot],
828 curr_slot = prev_slot;
829 curr_node = prev_node;
830 curr_bkt = curr_node->bkt;
833 if (h->readwrite_concur_lf_support) {
834 /* Inform the previous move. The current move need
835 * not be informed now as the current bucket entry
836 * is present in both primary and secondary.
837 * Since there is one writer, load acquires on
838 * tbl_chng_cnt are not required.
840 __atomic_store_n(h->tbl_chng_cnt,
841 *h->tbl_chng_cnt + 1,
843 /* The store to sig_current should not
844 * move above the store to tbl_chng_cnt.
846 __atomic_thread_fence(__ATOMIC_RELEASE);
849 curr_bkt->sig_current[curr_slot] = sig;
850 /* Release the new bucket entry */
851 __atomic_store_n(&curr_bkt->key_idx[curr_slot],
855 __hash_rw_writer_unlock(h);
862 * Make space for new key, using bfs Cuckoo Search and Multi-Writer safe
866 rte_hash_cuckoo_make_space_mw(const struct rte_hash *h,
867 struct rte_hash_bucket *bkt,
868 struct rte_hash_bucket *sec_bkt,
869 const struct rte_hash_key *key, void *data,
870 uint16_t sig, uint32_t bucket_idx,
871 uint32_t new_idx, int32_t *ret_val)
874 struct queue_node queue[RTE_HASH_BFS_QUEUE_MAX_LEN];
875 struct queue_node *tail, *head;
876 struct rte_hash_bucket *curr_bkt, *alt_bkt;
877 uint32_t cur_idx, alt_idx;
883 tail->prev_slot = -1;
884 tail->cur_bkt_idx = bucket_idx;
886 /* Cuckoo bfs Search */
887 while (likely(tail != head && head <
888 queue + RTE_HASH_BFS_QUEUE_MAX_LEN -
889 RTE_HASH_BUCKET_ENTRIES)) {
890 curr_bkt = tail->bkt;
891 cur_idx = tail->cur_bkt_idx;
892 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
893 if (curr_bkt->key_idx[i] == EMPTY_SLOT) {
894 int32_t ret = rte_hash_cuckoo_move_insert_mw(h,
895 bkt, sec_bkt, key, data,
898 if (likely(ret != -1))
902 /* Enqueue new node and keep prev node info */
903 alt_idx = get_alt_bucket_index(h, cur_idx,
904 curr_bkt->sig_current[i]);
905 alt_bkt = &(h->buckets[alt_idx]);
907 head->cur_bkt_idx = alt_idx;
918 static inline int32_t
919 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
920 hash_sig_t sig, void *data)
923 uint32_t prim_bucket_idx, sec_bucket_idx;
924 struct rte_hash_bucket *prim_bkt, *sec_bkt, *cur_bkt;
925 struct rte_hash_key *new_k, *keys = h->key_store;
926 void *slot_id = NULL;
927 void *ext_bkt_id = NULL;
928 uint32_t new_idx, bkt_id;
933 struct lcore_cache *cached_free_slots = NULL;
935 struct rte_hash_bucket *last;
937 short_sig = get_short_sig(sig);
938 prim_bucket_idx = get_prim_bucket_index(h, sig);
939 sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
940 prim_bkt = &h->buckets[prim_bucket_idx];
941 sec_bkt = &h->buckets[sec_bucket_idx];
942 rte_prefetch0(prim_bkt);
943 rte_prefetch0(sec_bkt);
945 /* Check if key is already inserted in primary location */
946 __hash_rw_writer_lock(h);
947 ret = search_and_update(h, data, key, prim_bkt, short_sig);
949 __hash_rw_writer_unlock(h);
953 /* Check if key is already inserted in secondary location */
954 FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
955 ret = search_and_update(h, data, key, cur_bkt, short_sig);
957 __hash_rw_writer_unlock(h);
962 __hash_rw_writer_unlock(h);
964 /* Did not find a match, so get a new slot for storing the new key */
965 if (h->use_local_cache) {
966 lcore_id = rte_lcore_id();
967 cached_free_slots = &h->local_free_slots[lcore_id];
968 /* Try to get a free slot from the local cache */
969 if (cached_free_slots->len == 0) {
970 /* Need to get another burst of free slots from global ring */
971 n_slots = rte_ring_mc_dequeue_burst(h->free_slots,
972 cached_free_slots->objs,
973 LCORE_CACHE_SIZE, NULL);
978 cached_free_slots->len += n_slots;
981 /* Get a free slot from the local cache */
982 cached_free_slots->len--;
983 slot_id = cached_free_slots->objs[cached_free_slots->len];
985 if (rte_ring_sc_dequeue(h->free_slots, &slot_id) != 0) {
990 new_k = RTE_PTR_ADD(keys, (uintptr_t)slot_id * h->key_entry_size);
991 new_idx = (uint32_t)((uintptr_t) slot_id);
992 /* The store to application data (by the application) at *data should
993 * not leak after the store of pdata in the key store. i.e. pdata is
994 * the guard variable. Release the application data to the readers.
996 __atomic_store_n(&new_k->pdata,
1000 memcpy(new_k->key, key, h->key_len);
1002 /* Find an empty slot and insert */
1003 ret = rte_hash_cuckoo_insert_mw(h, prim_bkt, sec_bkt, key, data,
1004 short_sig, new_idx, &ret_val);
1007 else if (ret == 1) {
1008 enqueue_slot_back(h, cached_free_slots, slot_id);
1012 /* Primary bucket full, need to make space for new entry */
1013 ret = rte_hash_cuckoo_make_space_mw(h, prim_bkt, sec_bkt, key, data,
1014 short_sig, prim_bucket_idx, new_idx, &ret_val);
1017 else if (ret == 1) {
1018 enqueue_slot_back(h, cached_free_slots, slot_id);
1022 /* Also search secondary bucket to get better occupancy */
1023 ret = rte_hash_cuckoo_make_space_mw(h, sec_bkt, prim_bkt, key, data,
1024 short_sig, sec_bucket_idx, new_idx, &ret_val);
1028 else if (ret == 1) {
1029 enqueue_slot_back(h, cached_free_slots, slot_id);
1033 /* if ext table not enabled, we failed the insertion */
1034 if (!h->ext_table_support) {
1035 enqueue_slot_back(h, cached_free_slots, slot_id);
1039 /* Now we need to go through the extendable bucket. Protection is needed
1040 * to protect all extendable bucket processes.
1042 __hash_rw_writer_lock(h);
1043 /* We check for duplicates again since could be inserted before the lock */
1044 ret = search_and_update(h, data, key, prim_bkt, short_sig);
1046 enqueue_slot_back(h, cached_free_slots, slot_id);
1050 FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1051 ret = search_and_update(h, data, key, cur_bkt, short_sig);
1053 enqueue_slot_back(h, cached_free_slots, slot_id);
1058 /* Search sec and ext buckets to find an empty entry to insert. */
1059 FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1060 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1061 /* Check if slot is available */
1062 if (likely(cur_bkt->key_idx[i] == EMPTY_SLOT)) {
1063 cur_bkt->sig_current[i] = short_sig;
1064 /* Store to signature and key should not
1065 * leak after the store to key_idx. i.e.
1066 * key_idx is the guard variable for signature
1069 __atomic_store_n(&cur_bkt->key_idx[i],
1072 __hash_rw_writer_unlock(h);
1078 /* Failed to get an empty entry from extendable buckets. Link a new
1079 * extendable bucket. We first get a free bucket from ring.
1081 if (rte_ring_sc_dequeue(h->free_ext_bkts, &ext_bkt_id) != 0) {
1086 bkt_id = (uint32_t)((uintptr_t)ext_bkt_id) - 1;
1087 /* Use the first location of the new bucket */
1088 (h->buckets_ext[bkt_id]).sig_current[0] = short_sig;
1089 /* Store to signature and key should not leak after
1090 * the store to key_idx. i.e. key_idx is the guard variable
1091 * for signature and key.
1093 __atomic_store_n(&(h->buckets_ext[bkt_id]).key_idx[0],
1096 /* Link the new bucket to sec bucket linked list */
1097 last = rte_hash_get_last_bkt(sec_bkt);
1098 last->next = &h->buckets_ext[bkt_id];
1099 __hash_rw_writer_unlock(h);
1103 __hash_rw_writer_unlock(h);
1109 rte_hash_add_key_with_hash(const struct rte_hash *h,
1110 const void *key, hash_sig_t sig)
1112 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1113 return __rte_hash_add_key_with_hash(h, key, sig, 0);
1117 rte_hash_add_key(const struct rte_hash *h, const void *key)
1119 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1120 return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
1124 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
1125 const void *key, hash_sig_t sig, void *data)
1129 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1130 ret = __rte_hash_add_key_with_hash(h, key, sig, data);
1138 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
1142 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1144 ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
1151 /* Search one bucket to find the match key - uses rw lock */
1152 static inline int32_t
1153 search_one_bucket_l(const struct rte_hash *h, const void *key,
1154 uint16_t sig, void **data,
1155 const struct rte_hash_bucket *bkt)
1158 struct rte_hash_key *k, *keys = h->key_store;
1160 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1161 if (bkt->sig_current[i] == sig &&
1162 bkt->key_idx[i] != EMPTY_SLOT) {
1163 k = (struct rte_hash_key *) ((char *)keys +
1164 bkt->key_idx[i] * h->key_entry_size);
1166 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1170 * Return index where key is stored,
1171 * subtracting the first dummy index
1173 return bkt->key_idx[i] - 1;
1180 /* Search one bucket to find the match key */
1181 static inline int32_t
1182 search_one_bucket_lf(const struct rte_hash *h, const void *key, uint16_t sig,
1183 void **data, const struct rte_hash_bucket *bkt)
1187 struct rte_hash_key *k, *keys = h->key_store;
1189 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1190 /* Signature comparison is done before the acquire-load
1191 * of the key index to achieve better performance.
1192 * This can result in the reader loading old signature
1193 * (which matches), while the key_idx is updated to a
1194 * value that belongs to a new key. However, the full
1195 * key comparison will ensure that the lookup fails.
1197 if (bkt->sig_current[i] == sig) {
1198 key_idx = __atomic_load_n(&bkt->key_idx[i],
1200 if (key_idx != EMPTY_SLOT) {
1201 k = (struct rte_hash_key *) ((char *)keys +
1202 key_idx * h->key_entry_size);
1204 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1206 *data = __atomic_load_n(
1211 * Return index where key is stored,
1212 * subtracting the first dummy index
1222 static inline int32_t
1223 __rte_hash_lookup_with_hash_l(const struct rte_hash *h, const void *key,
1224 hash_sig_t sig, void **data)
1226 uint32_t prim_bucket_idx, sec_bucket_idx;
1227 struct rte_hash_bucket *bkt, *cur_bkt;
1231 short_sig = get_short_sig(sig);
1232 prim_bucket_idx = get_prim_bucket_index(h, sig);
1233 sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1235 bkt = &h->buckets[prim_bucket_idx];
1237 __hash_rw_reader_lock(h);
1239 /* Check if key is in primary location */
1240 ret = search_one_bucket_l(h, key, short_sig, data, bkt);
1242 __hash_rw_reader_unlock(h);
1245 /* Calculate secondary hash */
1246 bkt = &h->buckets[sec_bucket_idx];
1248 /* Check if key is in secondary location */
1249 FOR_EACH_BUCKET(cur_bkt, bkt) {
1250 ret = search_one_bucket_l(h, key, short_sig,
1253 __hash_rw_reader_unlock(h);
1258 __hash_rw_reader_unlock(h);
1263 static inline int32_t
1264 __rte_hash_lookup_with_hash_lf(const struct rte_hash *h, const void *key,
1265 hash_sig_t sig, void **data)
1267 uint32_t prim_bucket_idx, sec_bucket_idx;
1268 struct rte_hash_bucket *bkt, *cur_bkt;
1269 uint32_t cnt_b, cnt_a;
1273 short_sig = get_short_sig(sig);
1274 prim_bucket_idx = get_prim_bucket_index(h, sig);
1275 sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1278 /* Load the table change counter before the lookup
1279 * starts. Acquire semantics will make sure that
1280 * loads in search_one_bucket are not hoisted.
1282 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1285 /* Check if key is in primary location */
1286 bkt = &h->buckets[prim_bucket_idx];
1287 ret = search_one_bucket_lf(h, key, short_sig, data, bkt);
1289 __hash_rw_reader_unlock(h);
1292 /* Calculate secondary hash */
1293 bkt = &h->buckets[sec_bucket_idx];
1295 /* Check if key is in secondary location */
1296 FOR_EACH_BUCKET(cur_bkt, bkt) {
1297 ret = search_one_bucket_lf(h, key, short_sig,
1300 __hash_rw_reader_unlock(h);
1305 /* The loads of sig_current in search_one_bucket
1306 * should not move below the load from tbl_chng_cnt.
1308 __atomic_thread_fence(__ATOMIC_ACQUIRE);
1309 /* Re-read the table change counter to check if the
1310 * table has changed during search. If yes, re-do
1312 * This load should not get hoisted. The load
1313 * acquires on cnt_b, key index in primary bucket
1314 * and key index in secondary bucket will make sure
1315 * that it does not get hoisted.
1317 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
1319 } while (cnt_b != cnt_a);
1324 static inline int32_t
1325 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
1326 hash_sig_t sig, void **data)
1328 if (h->readwrite_concur_lf_support)
1329 return __rte_hash_lookup_with_hash_lf(h, key, sig, data);
1331 return __rte_hash_lookup_with_hash_l(h, key, sig, data);
1335 rte_hash_lookup_with_hash(const struct rte_hash *h,
1336 const void *key, hash_sig_t sig)
1338 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1339 return __rte_hash_lookup_with_hash(h, key, sig, NULL);
1343 rte_hash_lookup(const struct rte_hash *h, const void *key)
1345 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1346 return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
1350 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
1351 const void *key, hash_sig_t sig, void **data)
1353 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1354 return __rte_hash_lookup_with_hash(h, key, sig, data);
1358 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
1360 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1361 return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
1365 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
1367 unsigned lcore_id, n_slots;
1368 struct lcore_cache *cached_free_slots;
1370 if (h->use_local_cache) {
1371 lcore_id = rte_lcore_id();
1372 cached_free_slots = &h->local_free_slots[lcore_id];
1373 /* Cache full, need to free it. */
1374 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1375 /* Need to enqueue the free slots in global ring. */
1376 n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
1377 cached_free_slots->objs,
1378 LCORE_CACHE_SIZE, NULL);
1379 ERR_IF_TRUE((n_slots == 0),
1380 "%s: could not enqueue free slots in global ring\n",
1382 cached_free_slots->len -= n_slots;
1384 /* Put index of new free slot in cache. */
1385 cached_free_slots->objs[cached_free_slots->len] =
1386 (void *)((uintptr_t)bkt->key_idx[i]);
1387 cached_free_slots->len++;
1389 rte_ring_sp_enqueue(h->free_slots,
1390 (void *)((uintptr_t)bkt->key_idx[i]));
1394 /* Compact the linked list by moving key from last entry in linked list to the
1398 __rte_hash_compact_ll(const struct rte_hash *h,
1399 struct rte_hash_bucket *cur_bkt, int pos) {
1401 struct rte_hash_bucket *last_bkt;
1406 last_bkt = rte_hash_get_last_bkt(cur_bkt);
1408 for (i = RTE_HASH_BUCKET_ENTRIES - 1; i >= 0; i--) {
1409 if (last_bkt->key_idx[i] != EMPTY_SLOT) {
1410 cur_bkt->sig_current[pos] = last_bkt->sig_current[i];
1411 __atomic_store_n(&cur_bkt->key_idx[pos],
1412 last_bkt->key_idx[i],
1414 if (h->readwrite_concur_lf_support) {
1415 /* Inform the readers that the table has changed
1416 * Since there is one writer, load acquire on
1417 * tbl_chng_cnt is not required.
1419 __atomic_store_n(h->tbl_chng_cnt,
1420 *h->tbl_chng_cnt + 1,
1422 /* The store to sig_current should
1423 * not move above the store to tbl_chng_cnt.
1425 __atomic_thread_fence(__ATOMIC_RELEASE);
1427 last_bkt->sig_current[i] = NULL_SIGNATURE;
1428 __atomic_store_n(&last_bkt->key_idx[i],
1436 /* Search one bucket and remove the matched key.
1437 * Writer is expected to hold the lock while calling this
1440 static inline int32_t
1441 search_and_remove(const struct rte_hash *h, const void *key,
1442 struct rte_hash_bucket *bkt, uint16_t sig, int *pos)
1444 struct rte_hash_key *k, *keys = h->key_store;
1448 /* Check if key is in bucket */
1449 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1450 key_idx = __atomic_load_n(&bkt->key_idx[i],
1452 if (bkt->sig_current[i] == sig && key_idx != EMPTY_SLOT) {
1453 k = (struct rte_hash_key *) ((char *)keys +
1454 key_idx * h->key_entry_size);
1455 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1456 bkt->sig_current[i] = NULL_SIGNATURE;
1457 /* Free the key store index if
1458 * no_free_on_del is disabled.
1460 if (!h->no_free_on_del)
1461 remove_entry(h, bkt, i);
1463 __atomic_store_n(&bkt->key_idx[i],
1469 * Return index where key is stored,
1470 * subtracting the first dummy index
1479 static inline int32_t
1480 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
1483 uint32_t prim_bucket_idx, sec_bucket_idx;
1484 struct rte_hash_bucket *prim_bkt, *sec_bkt, *prev_bkt, *last_bkt;
1485 struct rte_hash_bucket *cur_bkt;
1490 short_sig = get_short_sig(sig);
1491 prim_bucket_idx = get_prim_bucket_index(h, sig);
1492 sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1493 prim_bkt = &h->buckets[prim_bucket_idx];
1495 __hash_rw_writer_lock(h);
1496 /* look for key in primary bucket */
1497 ret = search_and_remove(h, key, prim_bkt, short_sig, &pos);
1499 __rte_hash_compact_ll(h, prim_bkt, pos);
1500 last_bkt = prim_bkt->next;
1501 prev_bkt = prim_bkt;
1505 /* Calculate secondary hash */
1506 sec_bkt = &h->buckets[sec_bucket_idx];
1508 FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1509 ret = search_and_remove(h, key, cur_bkt, short_sig, &pos);
1511 __rte_hash_compact_ll(h, cur_bkt, pos);
1512 last_bkt = sec_bkt->next;
1518 __hash_rw_writer_unlock(h);
1521 /* Search last bucket to see if empty to be recycled */
1524 __hash_rw_writer_unlock(h);
1527 while (last_bkt->next) {
1528 prev_bkt = last_bkt;
1529 last_bkt = last_bkt->next;
1532 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1533 if (last_bkt->key_idx[i] != EMPTY_SLOT)
1536 /* found empty bucket and recycle */
1537 if (i == RTE_HASH_BUCKET_ENTRIES) {
1538 prev_bkt->next = NULL;
1539 uint32_t index = last_bkt - h->buckets_ext + 1;
1540 /* Recycle the empty bkt if
1541 * no_free_on_del is disabled.
1543 if (h->no_free_on_del)
1544 /* Store index of an empty ext bkt to be recycled
1545 * on calling rte_hash_del_xxx APIs.
1546 * When lock free read-write concurrency is enabled,
1547 * an empty ext bkt cannot be put into free list
1548 * immediately (as readers might be using it still).
1549 * Hence freeing of the ext bkt is piggy-backed to
1550 * freeing of the key index.
1552 h->ext_bkt_to_free[ret] = index;
1554 rte_ring_sp_enqueue(h->free_ext_bkts, (void *)(uintptr_t)index);
1556 __hash_rw_writer_unlock(h);
1561 rte_hash_del_key_with_hash(const struct rte_hash *h,
1562 const void *key, hash_sig_t sig)
1564 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1565 return __rte_hash_del_key_with_hash(h, key, sig);
1569 rte_hash_del_key(const struct rte_hash *h, const void *key)
1571 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1572 return __rte_hash_del_key_with_hash(h, key, rte_hash_hash(h, key));
1576 rte_hash_get_key_with_position(const struct rte_hash *h, const int32_t position,
1579 RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1581 struct rte_hash_key *k, *keys = h->key_store;
1582 k = (struct rte_hash_key *) ((char *) keys + (position + 1) *
1587 __rte_hash_lookup_with_hash(h, *key, rte_hash_hash(h, *key),
1596 rte_hash_free_key_with_position(const struct rte_hash *h,
1597 const int32_t position)
1599 /* Key index where key is stored, adding the first dummy index */
1600 uint32_t key_idx = position + 1;
1602 RETURN_IF_TRUE(((h == NULL) || (key_idx == EMPTY_SLOT)), -EINVAL);
1604 unsigned int lcore_id, n_slots;
1605 struct lcore_cache *cached_free_slots;
1606 const uint32_t total_entries = h->use_local_cache ?
1607 h->entries + (RTE_MAX_LCORE - 1) * (LCORE_CACHE_SIZE - 1) + 1
1611 if (key_idx >= total_entries)
1613 if (h->ext_table_support && h->readwrite_concur_lf_support) {
1614 uint32_t index = h->ext_bkt_to_free[position];
1616 /* Recycle empty ext bkt to free list. */
1617 rte_ring_sp_enqueue(h->free_ext_bkts, (void *)(uintptr_t)index);
1618 h->ext_bkt_to_free[position] = 0;
1622 if (h->use_local_cache) {
1623 lcore_id = rte_lcore_id();
1624 cached_free_slots = &h->local_free_slots[lcore_id];
1625 /* Cache full, need to free it. */
1626 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1627 /* Need to enqueue the free slots in global ring. */
1628 n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
1629 cached_free_slots->objs,
1630 LCORE_CACHE_SIZE, NULL);
1631 RETURN_IF_TRUE((n_slots == 0), -EFAULT);
1632 cached_free_slots->len -= n_slots;
1634 /* Put index of new free slot in cache. */
1635 cached_free_slots->objs[cached_free_slots->len] =
1636 (void *)((uintptr_t)key_idx);
1637 cached_free_slots->len++;
1639 rte_ring_sp_enqueue(h->free_slots,
1640 (void *)((uintptr_t)key_idx));
1647 compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
1648 const struct rte_hash_bucket *prim_bkt,
1649 const struct rte_hash_bucket *sec_bkt,
1651 enum rte_hash_sig_compare_function sig_cmp_fn)
1655 /* For match mask the first bit of every two bits indicates the match */
1656 switch (sig_cmp_fn) {
1657 #if defined(RTE_MACHINE_CPUFLAG_SSE2)
1658 case RTE_HASH_COMPARE_SSE:
1659 /* Compare all signatures in the bucket */
1660 *prim_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1662 (__m128i const *)prim_bkt->sig_current),
1663 _mm_set1_epi16(sig)));
1664 /* Compare all signatures in the bucket */
1665 *sec_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1667 (__m128i const *)sec_bkt->sig_current),
1668 _mm_set1_epi16(sig)));
1670 #elif defined(RTE_MACHINE_CPUFLAG_NEON)
1671 case RTE_HASH_COMPARE_NEON: {
1672 uint16x8_t vmat, vsig, x;
1673 int16x8_t shift = {-15, -13, -11, -9, -7, -5, -3, -1};
1675 vsig = vld1q_dup_u16((uint16_t const *)&sig);
1676 /* Compare all signatures in the primary bucket */
1677 vmat = vceqq_u16(vsig,
1678 vld1q_u16((uint16_t const *)prim_bkt->sig_current));
1679 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1680 *prim_hash_matches = (uint32_t)(vaddvq_u16(x));
1681 /* Compare all signatures in the secondary bucket */
1682 vmat = vceqq_u16(vsig,
1683 vld1q_u16((uint16_t const *)sec_bkt->sig_current));
1684 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1685 *sec_hash_matches = (uint32_t)(vaddvq_u16(x));
1690 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1691 *prim_hash_matches |=
1692 ((sig == prim_bkt->sig_current[i]) << (i << 1));
1693 *sec_hash_matches |=
1694 ((sig == sec_bkt->sig_current[i]) << (i << 1));
1699 #define PREFETCH_OFFSET 4
1701 __rte_hash_lookup_bulk_l(const struct rte_hash *h, const void **keys,
1702 int32_t num_keys, int32_t *positions,
1703 uint64_t *hit_mask, void *data[])
1708 uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
1709 uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
1710 uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
1711 uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
1712 const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1713 const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1714 uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1715 uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1716 struct rte_hash_bucket *cur_bkt, *next_bkt;
1718 /* Prefetch first keys */
1719 for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
1720 rte_prefetch0(keys[i]);
1723 * Prefetch rest of the keys, calculate primary and
1724 * secondary bucket and prefetch them
1726 for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
1727 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
1729 prim_hash[i] = rte_hash_hash(h, keys[i]);
1731 sig[i] = get_short_sig(prim_hash[i]);
1732 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1733 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1735 primary_bkt[i] = &h->buckets[prim_index[i]];
1736 secondary_bkt[i] = &h->buckets[sec_index[i]];
1738 rte_prefetch0(primary_bkt[i]);
1739 rte_prefetch0(secondary_bkt[i]);
1742 /* Calculate and prefetch rest of the buckets */
1743 for (; i < num_keys; i++) {
1744 prim_hash[i] = rte_hash_hash(h, keys[i]);
1746 sig[i] = get_short_sig(prim_hash[i]);
1747 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1748 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1750 primary_bkt[i] = &h->buckets[prim_index[i]];
1751 secondary_bkt[i] = &h->buckets[sec_index[i]];
1753 rte_prefetch0(primary_bkt[i]);
1754 rte_prefetch0(secondary_bkt[i]);
1757 __hash_rw_reader_lock(h);
1759 /* Compare signatures and prefetch key slot of first hit */
1760 for (i = 0; i < num_keys; i++) {
1761 compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1762 primary_bkt[i], secondary_bkt[i],
1763 sig[i], h->sig_cmp_fn);
1765 if (prim_hitmask[i]) {
1766 uint32_t first_hit =
1767 __builtin_ctzl(prim_hitmask[i])
1770 primary_bkt[i]->key_idx[first_hit];
1771 const struct rte_hash_key *key_slot =
1772 (const struct rte_hash_key *)(
1773 (const char *)h->key_store +
1774 key_idx * h->key_entry_size);
1775 rte_prefetch0(key_slot);
1779 if (sec_hitmask[i]) {
1780 uint32_t first_hit =
1781 __builtin_ctzl(sec_hitmask[i])
1784 secondary_bkt[i]->key_idx[first_hit];
1785 const struct rte_hash_key *key_slot =
1786 (const struct rte_hash_key *)(
1787 (const char *)h->key_store +
1788 key_idx * h->key_entry_size);
1789 rte_prefetch0(key_slot);
1793 /* Compare keys, first hits in primary first */
1794 for (i = 0; i < num_keys; i++) {
1795 positions[i] = -ENOENT;
1796 while (prim_hitmask[i]) {
1797 uint32_t hit_index =
1798 __builtin_ctzl(prim_hitmask[i])
1801 primary_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);
1808 * If key index is 0, do not compare key,
1809 * as it is checking the dummy slot
1813 key_slot->key, keys[i], h)) {
1815 data[i] = key_slot->pdata;
1818 positions[i] = key_idx - 1;
1821 prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
1824 while (sec_hitmask[i]) {
1825 uint32_t hit_index =
1826 __builtin_ctzl(sec_hitmask[i])
1829 secondary_bkt[i]->key_idx[hit_index];
1830 const struct rte_hash_key *key_slot =
1831 (const struct rte_hash_key *)(
1832 (const char *)h->key_store +
1833 key_idx * h->key_entry_size);
1836 * If key index is 0, do not compare key,
1837 * as it is checking the dummy slot
1842 key_slot->key, keys[i], h)) {
1844 data[i] = key_slot->pdata;
1847 positions[i] = key_idx - 1;
1850 sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
1856 /* all found, do not need to go through ext bkt */
1857 if ((hits == ((1ULL << num_keys) - 1)) || !h->ext_table_support) {
1858 if (hit_mask != NULL)
1860 __hash_rw_reader_unlock(h);
1864 /* need to check ext buckets for match */
1865 for (i = 0; i < num_keys; i++) {
1866 if ((hits & (1ULL << i)) != 0)
1868 next_bkt = secondary_bkt[i]->next;
1869 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
1871 ret = search_one_bucket_l(h, keys[i],
1872 sig[i], &data[i], cur_bkt);
1874 ret = search_one_bucket_l(h, keys[i],
1875 sig[i], NULL, cur_bkt);
1884 __hash_rw_reader_unlock(h);
1886 if (hit_mask != NULL)
1891 __rte_hash_lookup_bulk_lf(const struct rte_hash *h, const void **keys,
1892 int32_t num_keys, int32_t *positions,
1893 uint64_t *hit_mask, void *data[])
1898 uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
1899 uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
1900 uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
1901 uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
1902 const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1903 const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1904 uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1905 uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1906 struct rte_hash_bucket *cur_bkt, *next_bkt;
1907 uint32_t cnt_b, cnt_a;
1909 /* Prefetch first keys */
1910 for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
1911 rte_prefetch0(keys[i]);
1914 * Prefetch rest of the keys, calculate primary and
1915 * secondary bucket and prefetch them
1917 for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
1918 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
1920 prim_hash[i] = rte_hash_hash(h, keys[i]);
1922 sig[i] = get_short_sig(prim_hash[i]);
1923 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1924 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1926 primary_bkt[i] = &h->buckets[prim_index[i]];
1927 secondary_bkt[i] = &h->buckets[sec_index[i]];
1929 rte_prefetch0(primary_bkt[i]);
1930 rte_prefetch0(secondary_bkt[i]);
1933 /* Calculate and prefetch rest of the buckets */
1934 for (; i < num_keys; i++) {
1935 prim_hash[i] = rte_hash_hash(h, keys[i]);
1937 sig[i] = get_short_sig(prim_hash[i]);
1938 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1939 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1941 primary_bkt[i] = &h->buckets[prim_index[i]];
1942 secondary_bkt[i] = &h->buckets[sec_index[i]];
1944 rte_prefetch0(primary_bkt[i]);
1945 rte_prefetch0(secondary_bkt[i]);
1948 for (i = 0; i < num_keys; i++)
1949 positions[i] = -ENOENT;
1952 /* Load the table change counter before the lookup
1953 * starts. Acquire semantics will make sure that
1954 * loads in compare_signatures are not hoisted.
1956 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1959 /* Compare signatures and prefetch key slot of first hit */
1960 for (i = 0; i < num_keys; i++) {
1961 compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1962 primary_bkt[i], secondary_bkt[i],
1963 sig[i], h->sig_cmp_fn);
1965 if (prim_hitmask[i]) {
1966 uint32_t first_hit =
1967 __builtin_ctzl(prim_hitmask[i])
1970 primary_bkt[i]->key_idx[first_hit];
1971 const struct rte_hash_key *key_slot =
1972 (const struct rte_hash_key *)(
1973 (const char *)h->key_store +
1974 key_idx * h->key_entry_size);
1975 rte_prefetch0(key_slot);
1979 if (sec_hitmask[i]) {
1980 uint32_t first_hit =
1981 __builtin_ctzl(sec_hitmask[i])
1984 secondary_bkt[i]->key_idx[first_hit];
1985 const struct rte_hash_key *key_slot =
1986 (const struct rte_hash_key *)(
1987 (const char *)h->key_store +
1988 key_idx * h->key_entry_size);
1989 rte_prefetch0(key_slot);
1993 /* Compare keys, first hits in primary first */
1994 for (i = 0; i < num_keys; i++) {
1995 while (prim_hitmask[i]) {
1996 uint32_t hit_index =
1997 __builtin_ctzl(prim_hitmask[i])
2001 &primary_bkt[i]->key_idx[hit_index],
2003 const struct rte_hash_key *key_slot =
2004 (const struct rte_hash_key *)(
2005 (const char *)h->key_store +
2006 key_idx * h->key_entry_size);
2009 * If key index is 0, do not compare key,
2010 * as it is checking the dummy slot
2014 key_slot->key, keys[i], h)) {
2016 data[i] = __atomic_load_n(
2021 positions[i] = key_idx - 1;
2024 prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
2027 while (sec_hitmask[i]) {
2028 uint32_t hit_index =
2029 __builtin_ctzl(sec_hitmask[i])
2033 &secondary_bkt[i]->key_idx[hit_index],
2035 const struct rte_hash_key *key_slot =
2036 (const struct rte_hash_key *)(
2037 (const char *)h->key_store +
2038 key_idx * h->key_entry_size);
2041 * If key index is 0, do not compare key,
2042 * as it is checking the dummy slot
2047 key_slot->key, keys[i], h)) {
2049 data[i] = __atomic_load_n(
2054 positions[i] = key_idx - 1;
2057 sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
2063 /* all found, do not need to go through ext bkt */
2064 if (hits == ((1ULL << num_keys) - 1)) {
2065 if (hit_mask != NULL)
2069 /* need to check ext buckets for match */
2070 if (h->ext_table_support) {
2071 for (i = 0; i < num_keys; i++) {
2072 if ((hits & (1ULL << i)) != 0)
2074 next_bkt = secondary_bkt[i]->next;
2075 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
2077 ret = search_one_bucket_lf(h,
2081 ret = search_one_bucket_lf(h,
2092 /* The loads of sig_current in compare_signatures
2093 * should not move below the load from tbl_chng_cnt.
2095 __atomic_thread_fence(__ATOMIC_ACQUIRE);
2096 /* Re-read the table change counter to check if the
2097 * table has changed during search. If yes, re-do
2099 * This load should not get hoisted. The load
2100 * acquires on cnt_b, primary key index and secondary
2101 * key index will make sure that it does not get
2104 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
2106 } while (cnt_b != cnt_a);
2108 if (hit_mask != NULL)
2113 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2114 int32_t num_keys, int32_t *positions,
2115 uint64_t *hit_mask, void *data[])
2117 if (h->readwrite_concur_lf_support)
2118 __rte_hash_lookup_bulk_lf(h, keys, num_keys, positions,
2121 __rte_hash_lookup_bulk_l(h, keys, num_keys, positions,
2126 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2127 uint32_t num_keys, int32_t *positions)
2129 RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2130 (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2131 (positions == NULL)), -EINVAL);
2133 __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
2138 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
2139 uint32_t num_keys, uint64_t *hit_mask, void *data[])
2141 RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2142 (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2143 (hit_mask == NULL)), -EINVAL);
2145 int32_t positions[num_keys];
2147 __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
2149 /* Return number of hits */
2150 return __builtin_popcountl(*hit_mask);
2154 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
2156 uint32_t bucket_idx, idx, position;
2157 struct rte_hash_key *next_key;
2159 RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
2161 const uint32_t total_entries_main = h->num_buckets *
2162 RTE_HASH_BUCKET_ENTRIES;
2163 const uint32_t total_entries = total_entries_main << 1;
2165 /* Out of bounds of all buckets (both main table and ext table) */
2166 if (*next >= total_entries_main)
2169 /* Calculate bucket and index of current iterator */
2170 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2171 idx = *next % RTE_HASH_BUCKET_ENTRIES;
2173 /* If current position is empty, go to the next one */
2174 while ((position = __atomic_load_n(&h->buckets[bucket_idx].key_idx[idx],
2175 __ATOMIC_ACQUIRE)) == EMPTY_SLOT) {
2178 if (*next == total_entries_main)
2180 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2181 idx = *next % RTE_HASH_BUCKET_ENTRIES;
2184 __hash_rw_reader_lock(h);
2185 next_key = (struct rte_hash_key *) ((char *)h->key_store +
2186 position * h->key_entry_size);
2187 /* Return key and data */
2188 *key = next_key->key;
2189 *data = next_key->pdata;
2191 __hash_rw_reader_unlock(h);
2193 /* Increment iterator */
2196 return position - 1;
2198 /* Begin to iterate extendable buckets */
2200 /* Out of total bound or if ext bucket feature is not enabled */
2201 if (*next >= total_entries || !h->ext_table_support)
2204 bucket_idx = (*next - total_entries_main) / RTE_HASH_BUCKET_ENTRIES;
2205 idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2207 while ((position = h->buckets_ext[bucket_idx].key_idx[idx]) == EMPTY_SLOT) {
2209 if (*next == total_entries)
2211 bucket_idx = (*next - total_entries_main) /
2212 RTE_HASH_BUCKET_ENTRIES;
2213 idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2215 __hash_rw_reader_lock(h);
2216 next_key = (struct rte_hash_key *) ((char *)h->key_store +
2217 position * h->key_entry_size);
2218 /* Return key and data */
2219 *key = next_key->key;
2220 *data = next_key->pdata;
2222 __hash_rw_reader_unlock(h);
2224 /* Increment iterator */
2226 return position - 1;