1
2
3
4#include <string.h>
5#include <stdio.h>
6
7#include <rte_common.h>
8#include <rte_malloc.h>
9#include <rte_log.h>
10
11#include "rte_table_hash.h"
12#include "rte_lru.h"
13
14#define KEY_SIZE 32
15
16#define KEYS_PER_BUCKET 4
17
18#define RTE_BUCKET_ENTRY_VALID 0x1LLU
19
20#ifdef RTE_TABLE_STATS_COLLECT
21
22#define RTE_TABLE_HASH_KEY32_STATS_PKTS_IN_ADD(table, val) \
23 table->stats.n_pkts_in += val
24#define RTE_TABLE_HASH_KEY32_STATS_PKTS_LOOKUP_MISS(table, val) \
25 table->stats.n_pkts_lookup_miss += val
26
27#else
28
29#define RTE_TABLE_HASH_KEY32_STATS_PKTS_IN_ADD(table, val)
30#define RTE_TABLE_HASH_KEY32_STATS_PKTS_LOOKUP_MISS(table, val)
31
32#endif
33
34#ifdef RTE_ARCH_64
35struct rte_bucket_4_32 {
36
37 uint64_t signature[4 + 1];
38 uint64_t lru_list;
39 struct rte_bucket_4_32 *next;
40 uint64_t next_valid;
41
42
43 uint64_t key[4][4];
44
45
46 uint8_t data[];
47};
48#else
49struct rte_bucket_4_32 {
50
51 uint64_t signature[4 + 1];
52 uint64_t lru_list;
53 struct rte_bucket_4_32 *next;
54 uint32_t pad;
55 uint64_t next_valid;
56
57
58 uint64_t key[4][4];
59
60
61 uint8_t data[];
62};
63#endif
64
65struct rte_table_hash {
66 struct rte_table_stats stats;
67
68
69 uint32_t n_buckets;
70 uint32_t key_size;
71 uint32_t entry_size;
72 uint32_t bucket_size;
73 uint32_t key_offset;
74 uint64_t key_mask[4];
75 rte_table_hash_op_hash f_hash;
76 uint64_t seed;
77
78
79 uint32_t n_buckets_ext;
80 uint32_t stack_pos;
81 uint32_t *stack;
82
83
84 uint8_t memory[0] __rte_cache_aligned;
85};
86
87static int
88keycmp(void *a, void *b, void *b_mask)
89{
90 uint64_t *a64 = a, *b64 = b, *b_mask64 = b_mask;
91
92 return (a64[0] != (b64[0] & b_mask64[0])) ||
93 (a64[1] != (b64[1] & b_mask64[1])) ||
94 (a64[2] != (b64[2] & b_mask64[2])) ||
95 (a64[3] != (b64[3] & b_mask64[3]));
96}
97
98static void
99keycpy(void *dst, void *src, void *src_mask)
100{
101 uint64_t *dst64 = dst, *src64 = src, *src_mask64 = src_mask;
102
103 dst64[0] = src64[0] & src_mask64[0];
104 dst64[1] = src64[1] & src_mask64[1];
105 dst64[2] = src64[2] & src_mask64[2];
106 dst64[3] = src64[3] & src_mask64[3];
107}
108
109static int
110check_params_create(struct rte_table_hash_params *params)
111{
112
113 if (params->name == NULL) {
114 RTE_LOG(ERR, TABLE, "%s: name invalid value\n", __func__);
115 return -EINVAL;
116 }
117
118
119 if (params->key_size != KEY_SIZE) {
120 RTE_LOG(ERR, TABLE, "%s: key_size invalid value\n", __func__);
121 return -EINVAL;
122 }
123
124
125 if (params->n_keys == 0) {
126 RTE_LOG(ERR, TABLE, "%s: n_keys is zero\n", __func__);
127 return -EINVAL;
128 }
129
130
131 if ((params->n_buckets == 0) ||
132 (!rte_is_power_of_2(params->n_buckets))) {
133 RTE_LOG(ERR, TABLE, "%s: n_buckets invalid value\n", __func__);
134 return -EINVAL;
135 }
136
137
138 if (params->f_hash == NULL) {
139 RTE_LOG(ERR, TABLE, "%s: f_hash function pointer is NULL\n",
140 __func__);
141 return -EINVAL;
142 }
143
144 return 0;
145}
146
147static void *
148rte_table_hash_create_key32_lru(void *params,
149 int socket_id,
150 uint32_t entry_size)
151{
152 struct rte_table_hash_params *p = params;
153 struct rte_table_hash *f;
154 uint64_t bucket_size, total_size;
155 uint32_t n_buckets, i;
156
157
158 if ((check_params_create(p) != 0) ||
159 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
160 ((sizeof(struct rte_bucket_4_32) % 64) != 0))
161 return NULL;
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178 n_buckets = rte_align32pow2(
179 (p->n_keys + KEYS_PER_BUCKET - 1) / KEYS_PER_BUCKET);
180 n_buckets = RTE_MAX(n_buckets, p->n_buckets);
181
182
183 bucket_size = RTE_CACHE_LINE_ROUNDUP(sizeof(struct rte_bucket_4_32) +
184 KEYS_PER_BUCKET * entry_size);
185 total_size = sizeof(struct rte_table_hash) + n_buckets * bucket_size;
186 if (total_size > SIZE_MAX) {
187 RTE_LOG(ERR, TABLE, "%s: Cannot allocate %" PRIu64 " bytes "
188 "for hash table %s\n",
189 __func__, total_size, p->name);
190 return NULL;
191 }
192
193 f = rte_zmalloc_socket(p->name,
194 (size_t)total_size,
195 RTE_CACHE_LINE_SIZE,
196 socket_id);
197 if (f == NULL) {
198 RTE_LOG(ERR, TABLE, "%s: Cannot allocate %" PRIu64 " bytes "
199 "for hash table %s\n",
200 __func__, total_size, p->name);
201 return NULL;
202 }
203 RTE_LOG(INFO, TABLE,
204 "%s: Hash table %s memory footprint "
205 "is %" PRIu64 " bytes\n",
206 __func__, p->name, total_size);
207
208
209 f->n_buckets = n_buckets;
210 f->key_size = KEY_SIZE;
211 f->entry_size = entry_size;
212 f->bucket_size = bucket_size;
213 f->key_offset = p->key_offset;
214 f->f_hash = p->f_hash;
215 f->seed = p->seed;
216
217 if (p->key_mask != NULL) {
218 f->key_mask[0] = ((uint64_t *)p->key_mask)[0];
219 f->key_mask[1] = ((uint64_t *)p->key_mask)[1];
220 f->key_mask[2] = ((uint64_t *)p->key_mask)[2];
221 f->key_mask[3] = ((uint64_t *)p->key_mask)[3];
222 } else {
223 f->key_mask[0] = 0xFFFFFFFFFFFFFFFFLLU;
224 f->key_mask[1] = 0xFFFFFFFFFFFFFFFFLLU;
225 f->key_mask[2] = 0xFFFFFFFFFFFFFFFFLLU;
226 f->key_mask[3] = 0xFFFFFFFFFFFFFFFFLLU;
227 }
228
229 for (i = 0; i < n_buckets; i++) {
230 struct rte_bucket_4_32 *bucket;
231
232 bucket = (struct rte_bucket_4_32 *) &f->memory[i *
233 f->bucket_size];
234 bucket->lru_list = 0x0000000100020003LLU;
235 }
236
237 return f;
238}
239
240static int
241rte_table_hash_free_key32_lru(void *table)
242{
243 struct rte_table_hash *f = table;
244
245
246 if (f == NULL) {
247 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
248 return -EINVAL;
249 }
250
251 rte_free(f);
252 return 0;
253}
254
255static int
256rte_table_hash_entry_add_key32_lru(
257 void *table,
258 void *key,
259 void *entry,
260 int *key_found,
261 void **entry_ptr)
262{
263 struct rte_table_hash *f = table;
264 struct rte_bucket_4_32 *bucket;
265 uint64_t signature, pos;
266 uint32_t bucket_index, i;
267
268 signature = f->f_hash(key, f->key_mask, f->key_size, f->seed);
269 bucket_index = signature & (f->n_buckets - 1);
270 bucket = (struct rte_bucket_4_32 *)
271 &f->memory[bucket_index * f->bucket_size];
272 signature |= RTE_BUCKET_ENTRY_VALID;
273
274
275 for (i = 0; i < 4; i++) {
276 uint64_t bucket_signature = bucket->signature[i];
277 uint8_t *bucket_key = (uint8_t *) &bucket->key[i];
278
279 if ((bucket_signature == signature) &&
280 (keycmp(bucket_key, key, f->key_mask) == 0)) {
281 uint8_t *bucket_data = &bucket->data[i * f->entry_size];
282
283 memcpy(bucket_data, entry, f->entry_size);
284 lru_update(bucket, i);
285 *key_found = 1;
286 *entry_ptr = (void *) bucket_data;
287 return 0;
288 }
289 }
290
291
292 for (i = 0; i < 4; i++) {
293 uint64_t bucket_signature = bucket->signature[i];
294 uint8_t *bucket_key = (uint8_t *) &bucket->key[i];
295
296 if (bucket_signature == 0) {
297 uint8_t *bucket_data = &bucket->data[i * f->entry_size];
298
299 bucket->signature[i] = signature;
300 keycpy(bucket_key, key, f->key_mask);
301 memcpy(bucket_data, entry, f->entry_size);
302 lru_update(bucket, i);
303 *key_found = 0;
304 *entry_ptr = (void *) bucket_data;
305
306 return 0;
307 }
308 }
309
310
311 pos = lru_pos(bucket);
312 bucket->signature[pos] = signature;
313 keycpy(&bucket->key[pos], key, f->key_mask);
314 memcpy(&bucket->data[pos * f->entry_size], entry, f->entry_size);
315 lru_update(bucket, pos);
316 *key_found = 0;
317 *entry_ptr = (void *) &bucket->data[pos * f->entry_size];
318
319 return 0;
320}
321
322static int
323rte_table_hash_entry_delete_key32_lru(
324 void *table,
325 void *key,
326 int *key_found,
327 void *entry)
328{
329 struct rte_table_hash *f = table;
330 struct rte_bucket_4_32 *bucket;
331 uint64_t signature;
332 uint32_t bucket_index, i;
333
334 signature = f->f_hash(key, f->key_mask, f->key_size, f->seed);
335 bucket_index = signature & (f->n_buckets - 1);
336 bucket = (struct rte_bucket_4_32 *)
337 &f->memory[bucket_index * f->bucket_size];
338 signature |= RTE_BUCKET_ENTRY_VALID;
339
340
341 for (i = 0; i < 4; i++) {
342 uint64_t bucket_signature = bucket->signature[i];
343 uint8_t *bucket_key = (uint8_t *) &bucket->key[i];
344
345 if ((bucket_signature == signature) &&
346 (keycmp(bucket_key, key, f->key_mask) == 0)) {
347 uint8_t *bucket_data = &bucket->data[i * f->entry_size];
348
349 bucket->signature[i] = 0;
350 *key_found = 1;
351 if (entry)
352 memcpy(entry, bucket_data, f->entry_size);
353
354 return 0;
355 }
356 }
357
358
359 *key_found = 0;
360 return 0;
361}
362
363static void *
364rte_table_hash_create_key32_ext(void *params,
365 int socket_id,
366 uint32_t entry_size)
367{
368 struct rte_table_hash_params *p = params;
369 struct rte_table_hash *f;
370 uint64_t bucket_size, stack_size, total_size;
371 uint32_t n_buckets_ext, i;
372
373
374 if ((check_params_create(p) != 0) ||
375 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
376 ((sizeof(struct rte_bucket_4_32) % 64) != 0))
377 return NULL;
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394 n_buckets_ext = p->n_keys / KEYS_PER_BUCKET + KEYS_PER_BUCKET - 1;
395
396
397 bucket_size = RTE_CACHE_LINE_ROUNDUP(sizeof(struct rte_bucket_4_32) +
398 KEYS_PER_BUCKET * entry_size);
399 stack_size = RTE_CACHE_LINE_ROUNDUP(n_buckets_ext * sizeof(uint32_t));
400 total_size = sizeof(struct rte_table_hash) +
401 (p->n_buckets + n_buckets_ext) * bucket_size + stack_size;
402 if (total_size > SIZE_MAX) {
403 RTE_LOG(ERR, TABLE, "%s: Cannot allocate %" PRIu64 " bytes "
404 "for hash table %s\n",
405 __func__, total_size, p->name);
406 return NULL;
407 }
408
409 f = rte_zmalloc_socket(p->name,
410 (size_t)total_size,
411 RTE_CACHE_LINE_SIZE,
412 socket_id);
413 if (f == NULL) {
414 RTE_LOG(ERR, TABLE, "%s: Cannot allocate %" PRIu64 " bytes "
415 "for hash table %s\n",
416 __func__, total_size, p->name);
417 return NULL;
418 }
419 RTE_LOG(INFO, TABLE,
420 "%s: Hash table %s memory footprint "
421 "is %" PRIu64" bytes\n",
422 __func__, p->name, total_size);
423
424
425 f->n_buckets = p->n_buckets;
426 f->key_size = KEY_SIZE;
427 f->entry_size = entry_size;
428 f->bucket_size = bucket_size;
429 f->key_offset = p->key_offset;
430 f->f_hash = p->f_hash;
431 f->seed = p->seed;
432
433 f->n_buckets_ext = n_buckets_ext;
434 f->stack_pos = n_buckets_ext;
435 f->stack = (uint32_t *)
436 &f->memory[(p->n_buckets + n_buckets_ext) * f->bucket_size];
437
438 if (p->key_mask != NULL) {
439 f->key_mask[0] = (((uint64_t *)p->key_mask)[0]);
440 f->key_mask[1] = (((uint64_t *)p->key_mask)[1]);
441 f->key_mask[2] = (((uint64_t *)p->key_mask)[2]);
442 f->key_mask[3] = (((uint64_t *)p->key_mask)[3]);
443 } else {
444 f->key_mask[0] = 0xFFFFFFFFFFFFFFFFLLU;
445 f->key_mask[1] = 0xFFFFFFFFFFFFFFFFLLU;
446 f->key_mask[2] = 0xFFFFFFFFFFFFFFFFLLU;
447 f->key_mask[3] = 0xFFFFFFFFFFFFFFFFLLU;
448 }
449
450 for (i = 0; i < n_buckets_ext; i++)
451 f->stack[i] = i;
452
453 return f;
454}
455
456static int
457rte_table_hash_free_key32_ext(void *table)
458{
459 struct rte_table_hash *f = table;
460
461
462 if (f == NULL) {
463 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
464 return -EINVAL;
465 }
466
467 rte_free(f);
468 return 0;
469}
470
471static int
472rte_table_hash_entry_add_key32_ext(
473 void *table,
474 void *key,
475 void *entry,
476 int *key_found,
477 void **entry_ptr)
478{
479 struct rte_table_hash *f = table;
480 struct rte_bucket_4_32 *bucket0, *bucket, *bucket_prev;
481 uint64_t signature;
482 uint32_t bucket_index, i;
483
484 signature = f->f_hash(key, f->key_mask, f->key_size, f->seed);
485 bucket_index = signature & (f->n_buckets - 1);
486 bucket0 = (struct rte_bucket_4_32 *)
487 &f->memory[bucket_index * f->bucket_size];
488 signature |= RTE_BUCKET_ENTRY_VALID;
489
490
491 for (bucket = bucket0; bucket != NULL; bucket = bucket->next) {
492 for (i = 0; i < 4; i++) {
493 uint64_t bucket_signature = bucket->signature[i];
494 uint8_t *bucket_key = (uint8_t *) &bucket->key[i];
495
496 if ((bucket_signature == signature) &&
497 (keycmp(bucket_key, key, f->key_mask) == 0)) {
498 uint8_t *bucket_data = &bucket->data[i *
499 f->entry_size];
500
501 memcpy(bucket_data, entry, f->entry_size);
502 *key_found = 1;
503 *entry_ptr = (void *) bucket_data;
504
505 return 0;
506 }
507 }
508 }
509
510
511 for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
512 bucket_prev = bucket, bucket = bucket->next)
513 for (i = 0; i < 4; i++) {
514 uint64_t bucket_signature = bucket->signature[i];
515 uint8_t *bucket_key = (uint8_t *) &bucket->key[i];
516
517 if (bucket_signature == 0) {
518 uint8_t *bucket_data = &bucket->data[i *
519 f->entry_size];
520
521 bucket->signature[i] = signature;
522 keycpy(bucket_key, key, f->key_mask);
523 memcpy(bucket_data, entry, f->entry_size);
524 *key_found = 0;
525 *entry_ptr = (void *) bucket_data;
526
527 return 0;
528 }
529 }
530
531
532 if (f->stack_pos > 0) {
533 bucket_index = f->stack[--f->stack_pos];
534
535 bucket = (struct rte_bucket_4_32 *)
536 &f->memory[(f->n_buckets + bucket_index) *
537 f->bucket_size];
538 bucket_prev->next = bucket;
539 bucket_prev->next_valid = 1;
540
541 bucket->signature[0] = signature;
542 keycpy(&bucket->key[0], key, f->key_mask);
543 memcpy(&bucket->data[0], entry, f->entry_size);
544 *key_found = 0;
545 *entry_ptr = (void *) &bucket->data[0];
546 return 0;
547 }
548
549 return -ENOSPC;
550}
551
552static int
553rte_table_hash_entry_delete_key32_ext(
554 void *table,
555 void *key,
556 int *key_found,
557 void *entry)
558{
559 struct rte_table_hash *f = table;
560 struct rte_bucket_4_32 *bucket0, *bucket, *bucket_prev;
561 uint64_t signature;
562 uint32_t bucket_index, i;
563
564 signature = f->f_hash(key, f->key_mask, f->key_size, f->seed);
565 bucket_index = signature & (f->n_buckets - 1);
566 bucket0 = (struct rte_bucket_4_32 *)
567 &f->memory[bucket_index * f->bucket_size];
568 signature |= RTE_BUCKET_ENTRY_VALID;
569
570
571 for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
572 bucket_prev = bucket, bucket = bucket->next)
573 for (i = 0; i < 4; i++) {
574 uint64_t bucket_signature = bucket->signature[i];
575 uint8_t *bucket_key = (uint8_t *) &bucket->key[i];
576
577 if ((bucket_signature == signature) &&
578 (keycmp(bucket_key, key, f->key_mask) == 0)) {
579 uint8_t *bucket_data = &bucket->data[i *
580 f->entry_size];
581
582 bucket->signature[i] = 0;
583 *key_found = 1;
584 if (entry)
585 memcpy(entry, bucket_data, f->entry_size);
586
587 if ((bucket->signature[0] == 0) &&
588 (bucket->signature[1] == 0) &&
589 (bucket->signature[2] == 0) &&
590 (bucket->signature[3] == 0) &&
591 (bucket_prev != NULL)) {
592 bucket_prev->next = bucket->next;
593 bucket_prev->next_valid =
594 bucket->next_valid;
595
596 memset(bucket, 0,
597 sizeof(struct rte_bucket_4_32));
598 bucket_index = (((uint8_t *)bucket -
599 (uint8_t *)f->memory)/f->bucket_size) - f->n_buckets;
600 f->stack[f->stack_pos++] = bucket_index;
601 }
602
603 return 0;
604 }
605 }
606
607
608 *key_found = 0;
609 return 0;
610}
611
612#define lookup_key32_cmp(key_in, bucket, pos, f) \
613{ \
614 uint64_t xor[4][4], or[4], signature[4], k[4]; \
615 \
616 k[0] = key_in[0] & f->key_mask[0]; \
617 k[1] = key_in[1] & f->key_mask[1]; \
618 k[2] = key_in[2] & f->key_mask[2]; \
619 k[3] = key_in[3] & f->key_mask[3]; \
620 \
621 signature[0] = ((~bucket->signature[0]) & 1); \
622 signature[1] = ((~bucket->signature[1]) & 1); \
623 signature[2] = ((~bucket->signature[2]) & 1); \
624 signature[3] = ((~bucket->signature[3]) & 1); \
625 \
626 xor[0][0] = k[0] ^ bucket->key[0][0]; \
627 xor[0][1] = k[1] ^ bucket->key[0][1]; \
628 xor[0][2] = k[2] ^ bucket->key[0][2]; \
629 xor[0][3] = k[3] ^ bucket->key[0][3]; \
630 \
631 xor[1][0] = k[0] ^ bucket->key[1][0]; \
632 xor[1][1] = k[1] ^ bucket->key[1][1]; \
633 xor[1][2] = k[2] ^ bucket->key[1][2]; \
634 xor[1][3] = k[3] ^ bucket->key[1][3]; \
635 \
636 xor[2][0] = k[0] ^ bucket->key[2][0]; \
637 xor[2][1] = k[1] ^ bucket->key[2][1]; \
638 xor[2][2] = k[2] ^ bucket->key[2][2]; \
639 xor[2][3] = k[3] ^ bucket->key[2][3]; \
640 \
641 xor[3][0] = k[0] ^ bucket->key[3][0]; \
642 xor[3][1] = k[1] ^ bucket->key[3][1]; \
643 xor[3][2] = k[2] ^ bucket->key[3][2]; \
644 xor[3][3] = k[3] ^ bucket->key[3][3]; \
645 \
646 or[0] = xor[0][0] | xor[0][1] | xor[0][2] | xor[0][3] | signature[0];\
647 or[1] = xor[1][0] | xor[1][1] | xor[1][2] | xor[1][3] | signature[1];\
648 or[2] = xor[2][0] | xor[2][1] | xor[2][2] | xor[2][3] | signature[2];\
649 or[3] = xor[3][0] | xor[3][1] | xor[3][2] | xor[3][3] | signature[3];\
650 \
651 pos = 4; \
652 if (or[0] == 0) \
653 pos = 0; \
654 if (or[1] == 0) \
655 pos = 1; \
656 if (or[2] == 0) \
657 pos = 2; \
658 if (or[3] == 0) \
659 pos = 3; \
660}
661
662#define lookup1_stage0(pkt0_index, mbuf0, pkts, pkts_mask, f) \
663{ \
664 uint64_t pkt_mask; \
665 uint32_t key_offset = f->key_offset; \
666 \
667 pkt0_index = __builtin_ctzll(pkts_mask); \
668 pkt_mask = 1LLU << pkt0_index; \
669 pkts_mask &= ~pkt_mask; \
670 \
671 mbuf0 = pkts[pkt0_index]; \
672 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf0, key_offset));\
673}
674
675#define lookup1_stage1(mbuf1, bucket1, f) \
676{ \
677 uint64_t *key; \
678 uint64_t signature; \
679 uint32_t bucket_index; \
680 \
681 key = RTE_MBUF_METADATA_UINT64_PTR(mbuf1, f->key_offset); \
682 signature = f->f_hash(key, f->key_mask, KEY_SIZE, f->seed); \
683 \
684 bucket_index = signature & (f->n_buckets - 1); \
685 bucket1 = (struct rte_bucket_4_32 *) \
686 &f->memory[bucket_index * f->bucket_size]; \
687 rte_prefetch0(bucket1); \
688 rte_prefetch0((void *)(((uintptr_t) bucket1) + RTE_CACHE_LINE_SIZE));\
689 rte_prefetch0((void *)(((uintptr_t) bucket1) + 2 * RTE_CACHE_LINE_SIZE));\
690}
691
692#define lookup1_stage2_lru(pkt2_index, mbuf2, bucket2, \
693 pkts_mask_out, entries, f) \
694{ \
695 void *a; \
696 uint64_t pkt_mask; \
697 uint64_t *key; \
698 uint32_t pos; \
699 \
700 key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
701 lookup_key32_cmp(key, bucket2, pos, f); \
702 \
703 pkt_mask = (bucket2->signature[pos] & 1LLU) << pkt2_index;\
704 pkts_mask_out |= pkt_mask; \
705 \
706 a = (void *) &bucket2->data[pos * f->entry_size]; \
707 rte_prefetch0(a); \
708 entries[pkt2_index] = a; \
709 lru_update(bucket2, pos); \
710}
711
712#define lookup1_stage2_ext(pkt2_index, mbuf2, bucket2, pkts_mask_out,\
713 entries, buckets_mask, buckets, keys, f) \
714{ \
715 struct rte_bucket_4_32 *bucket_next; \
716 void *a; \
717 uint64_t pkt_mask, bucket_mask; \
718 uint64_t *key; \
719 uint32_t pos; \
720 \
721 key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
722 lookup_key32_cmp(key, bucket2, pos, f); \
723 \
724 pkt_mask = (bucket2->signature[pos] & 1LLU) << pkt2_index;\
725 pkts_mask_out |= pkt_mask; \
726 \
727 a = (void *) &bucket2->data[pos * f->entry_size]; \
728 rte_prefetch0(a); \
729 entries[pkt2_index] = a; \
730 \
731 bucket_mask = (~pkt_mask) & (bucket2->next_valid << pkt2_index);\
732 buckets_mask |= bucket_mask; \
733 bucket_next = bucket2->next; \
734 buckets[pkt2_index] = bucket_next; \
735 keys[pkt2_index] = key; \
736}
737
738#define lookup_grinder(pkt_index, buckets, keys, pkts_mask_out, \
739 entries, buckets_mask, f) \
740{ \
741 struct rte_bucket_4_32 *bucket, *bucket_next; \
742 void *a; \
743 uint64_t pkt_mask, bucket_mask; \
744 uint64_t *key; \
745 uint32_t pos; \
746 \
747 bucket = buckets[pkt_index]; \
748 key = keys[pkt_index]; \
749 \
750 lookup_key32_cmp(key, bucket, pos, f); \
751 \
752 pkt_mask = (bucket->signature[pos] & 1LLU) << pkt_index;\
753 pkts_mask_out |= pkt_mask; \
754 \
755 a = (void *) &bucket->data[pos * f->entry_size]; \
756 rte_prefetch0(a); \
757 entries[pkt_index] = a; \
758 \
759 bucket_mask = (~pkt_mask) & (bucket->next_valid << pkt_index);\
760 buckets_mask |= bucket_mask; \
761 bucket_next = bucket->next; \
762 rte_prefetch0(bucket_next); \
763 rte_prefetch0((void *)(((uintptr_t) bucket_next) + RTE_CACHE_LINE_SIZE));\
764 rte_prefetch0((void *)(((uintptr_t) bucket_next) + \
765 2 * RTE_CACHE_LINE_SIZE)); \
766 buckets[pkt_index] = bucket_next; \
767 keys[pkt_index] = key; \
768}
769
770#define lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01,\
771 pkts, pkts_mask, f) \
772{ \
773 uint64_t pkt00_mask, pkt01_mask; \
774 uint32_t key_offset = f->key_offset; \
775 \
776 pkt00_index = __builtin_ctzll(pkts_mask); \
777 pkt00_mask = 1LLU << pkt00_index; \
778 pkts_mask &= ~pkt00_mask; \
779 \
780 mbuf00 = pkts[pkt00_index]; \
781 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, key_offset));\
782 \
783 pkt01_index = __builtin_ctzll(pkts_mask); \
784 pkt01_mask = 1LLU << pkt01_index; \
785 pkts_mask &= ~pkt01_mask; \
786 \
787 mbuf01 = pkts[pkt01_index]; \
788 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, key_offset));\
789}
790
791#define lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,\
792 mbuf00, mbuf01, pkts, pkts_mask, f) \
793{ \
794 uint64_t pkt00_mask, pkt01_mask; \
795 uint32_t key_offset = f->key_offset; \
796 \
797 pkt00_index = __builtin_ctzll(pkts_mask); \
798 pkt00_mask = 1LLU << pkt00_index; \
799 pkts_mask &= ~pkt00_mask; \
800 \
801 mbuf00 = pkts[pkt00_index]; \
802 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, key_offset)); \
803 \
804 pkt01_index = __builtin_ctzll(pkts_mask); \
805 if (pkts_mask == 0) \
806 pkt01_index = pkt00_index; \
807 \
808 pkt01_mask = 1LLU << pkt01_index; \
809 pkts_mask &= ~pkt01_mask; \
810 \
811 mbuf01 = pkts[pkt01_index]; \
812 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, key_offset)); \
813}
814
815#define lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f) \
816{ \
817 uint64_t *key10, *key11; \
818 uint64_t signature10, signature11; \
819 uint32_t bucket10_index, bucket11_index; \
820 \
821 key10 = RTE_MBUF_METADATA_UINT64_PTR(mbuf10, f->key_offset); \
822 signature10 = f->f_hash(key10, f->key_mask, KEY_SIZE, f->seed); \
823 \
824 bucket10_index = signature10 & (f->n_buckets - 1); \
825 bucket10 = (struct rte_bucket_4_32 *) \
826 &f->memory[bucket10_index * f->bucket_size]; \
827 rte_prefetch0(bucket10); \
828 rte_prefetch0((void *)(((uintptr_t) bucket10) + RTE_CACHE_LINE_SIZE));\
829 rte_prefetch0((void *)(((uintptr_t) bucket10) + 2 * RTE_CACHE_LINE_SIZE));\
830 \
831 key11 = RTE_MBUF_METADATA_UINT64_PTR(mbuf11, f->key_offset); \
832 signature11 = f->f_hash(key11, f->key_mask, KEY_SIZE, f->seed);\
833 \
834 bucket11_index = signature11 & (f->n_buckets - 1); \
835 bucket11 = (struct rte_bucket_4_32 *) \
836 &f->memory[bucket11_index * f->bucket_size]; \
837 rte_prefetch0(bucket11); \
838 rte_prefetch0((void *)(((uintptr_t) bucket11) + RTE_CACHE_LINE_SIZE));\
839 rte_prefetch0((void *)(((uintptr_t) bucket11) + 2 * RTE_CACHE_LINE_SIZE));\
840}
841
842#define lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,\
843 bucket20, bucket21, pkts_mask_out, entries, f) \
844{ \
845 void *a20, *a21; \
846 uint64_t pkt20_mask, pkt21_mask; \
847 uint64_t *key20, *key21; \
848 uint32_t pos20, pos21; \
849 \
850 key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
851 key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
852 \
853 lookup_key32_cmp(key20, bucket20, pos20, f); \
854 lookup_key32_cmp(key21, bucket21, pos21, f); \
855 \
856 pkt20_mask = (bucket20->signature[pos20] & 1LLU) << pkt20_index;\
857 pkt21_mask = (bucket21->signature[pos21] & 1LLU) << pkt21_index;\
858 pkts_mask_out |= pkt20_mask | pkt21_mask; \
859 \
860 a20 = (void *) &bucket20->data[pos20 * f->entry_size]; \
861 a21 = (void *) &bucket21->data[pos21 * f->entry_size]; \
862 rte_prefetch0(a20); \
863 rte_prefetch0(a21); \
864 entries[pkt20_index] = a20; \
865 entries[pkt21_index] = a21; \
866 lru_update(bucket20, pos20); \
867 lru_update(bucket21, pos21); \
868}
869
870#define lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21, bucket20, \
871 bucket21, pkts_mask_out, entries, buckets_mask, buckets, keys, f)\
872{ \
873 struct rte_bucket_4_32 *bucket20_next, *bucket21_next; \
874 void *a20, *a21; \
875 uint64_t pkt20_mask, pkt21_mask, bucket20_mask, bucket21_mask;\
876 uint64_t *key20, *key21; \
877 uint32_t pos20, pos21; \
878 \
879 key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
880 key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
881 \
882 lookup_key32_cmp(key20, bucket20, pos20, f); \
883 lookup_key32_cmp(key21, bucket21, pos21, f); \
884 \
885 pkt20_mask = (bucket20->signature[pos20] & 1LLU) << pkt20_index;\
886 pkt21_mask = (bucket21->signature[pos21] & 1LLU) << pkt21_index;\
887 pkts_mask_out |= pkt20_mask | pkt21_mask; \
888 \
889 a20 = (void *) &bucket20->data[pos20 * f->entry_size]; \
890 a21 = (void *) &bucket21->data[pos21 * f->entry_size]; \
891 rte_prefetch0(a20); \
892 rte_prefetch0(a21); \
893 entries[pkt20_index] = a20; \
894 entries[pkt21_index] = a21; \
895 \
896 bucket20_mask = (~pkt20_mask) & (bucket20->next_valid << pkt20_index);\
897 bucket21_mask = (~pkt21_mask) & (bucket21->next_valid << pkt21_index);\
898 buckets_mask |= bucket20_mask | bucket21_mask; \
899 bucket20_next = bucket20->next; \
900 bucket21_next = bucket21->next; \
901 buckets[pkt20_index] = bucket20_next; \
902 buckets[pkt21_index] = bucket21_next; \
903 keys[pkt20_index] = key20; \
904 keys[pkt21_index] = key21; \
905}
906
907static int
908rte_table_hash_lookup_key32_lru(
909 void *table,
910 struct rte_mbuf **pkts,
911 uint64_t pkts_mask,
912 uint64_t *lookup_hit_mask,
913 void **entries)
914{
915 struct rte_table_hash *f = (struct rte_table_hash *) table;
916 struct rte_bucket_4_32 *bucket10, *bucket11, *bucket20, *bucket21;
917 struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
918 uint32_t pkt00_index, pkt01_index, pkt10_index;
919 uint32_t pkt11_index, pkt20_index, pkt21_index;
920 uint64_t pkts_mask_out = 0;
921
922 __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
923 RTE_TABLE_HASH_KEY32_STATS_PKTS_IN_ADD(f, n_pkts_in);
924
925
926 if (__builtin_popcountll(pkts_mask) < 5) {
927 for ( ; pkts_mask; ) {
928 struct rte_bucket_4_32 *bucket;
929 struct rte_mbuf *mbuf;
930 uint32_t pkt_index;
931
932 lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask, f);
933 lookup1_stage1(mbuf, bucket, f);
934 lookup1_stage2_lru(pkt_index, mbuf, bucket,
935 pkts_mask_out, entries, f);
936 }
937
938 *lookup_hit_mask = pkts_mask_out;
939 RTE_TABLE_HASH_KEY32_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
940 return 0;
941 }
942
943
944
945
946
947
948 lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
949 pkts_mask, f);
950
951
952 mbuf10 = mbuf00;
953 mbuf11 = mbuf01;
954 pkt10_index = pkt00_index;
955 pkt11_index = pkt01_index;
956
957
958 lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
959 pkts_mask, f);
960
961
962 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
963
964
965
966
967
968 for ( ; pkts_mask; ) {
969
970 bucket20 = bucket10;
971 bucket21 = bucket11;
972 mbuf20 = mbuf10;
973 mbuf21 = mbuf11;
974 mbuf10 = mbuf00;
975 mbuf11 = mbuf01;
976 pkt20_index = pkt10_index;
977 pkt21_index = pkt11_index;
978 pkt10_index = pkt00_index;
979 pkt11_index = pkt01_index;
980
981
982 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
983 mbuf00, mbuf01, pkts, pkts_mask, f);
984
985
986 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
987
988
989 lookup2_stage2_lru(pkt20_index, pkt21_index,
990 mbuf20, mbuf21, bucket20, bucket21, pkts_mask_out,
991 entries, f);
992 }
993
994
995
996
997
998
999 bucket20 = bucket10;
1000 bucket21 = bucket11;
1001 mbuf20 = mbuf10;
1002 mbuf21 = mbuf11;
1003 mbuf10 = mbuf00;
1004 mbuf11 = mbuf01;
1005 pkt20_index = pkt10_index;
1006 pkt21_index = pkt11_index;
1007 pkt10_index = pkt00_index;
1008 pkt11_index = pkt01_index;
1009
1010
1011 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1012
1013
1014 lookup2_stage2_lru(pkt20_index, pkt21_index,
1015 mbuf20, mbuf21, bucket20, bucket21, pkts_mask_out, entries, f);
1016
1017
1018 bucket20 = bucket10;
1019 bucket21 = bucket11;
1020 mbuf20 = mbuf10;
1021 mbuf21 = mbuf11;
1022 pkt20_index = pkt10_index;
1023 pkt21_index = pkt11_index;
1024
1025
1026 lookup2_stage2_lru(pkt20_index, pkt21_index,
1027 mbuf20, mbuf21, bucket20, bucket21, pkts_mask_out, entries, f);
1028
1029 *lookup_hit_mask = pkts_mask_out;
1030 RTE_TABLE_HASH_KEY32_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1031 return 0;
1032}
1033
1034static int
1035rte_table_hash_lookup_key32_ext(
1036 void *table,
1037 struct rte_mbuf **pkts,
1038 uint64_t pkts_mask,
1039 uint64_t *lookup_hit_mask,
1040 void **entries)
1041{
1042 struct rte_table_hash *f = (struct rte_table_hash *) table;
1043 struct rte_bucket_4_32 *bucket10, *bucket11, *bucket20, *bucket21;
1044 struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
1045 uint32_t pkt00_index, pkt01_index, pkt10_index;
1046 uint32_t pkt11_index, pkt20_index, pkt21_index;
1047 uint64_t pkts_mask_out = 0, buckets_mask = 0;
1048 struct rte_bucket_4_32 *buckets[RTE_PORT_IN_BURST_SIZE_MAX];
1049 uint64_t *keys[RTE_PORT_IN_BURST_SIZE_MAX];
1050
1051 __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1052 RTE_TABLE_HASH_KEY32_STATS_PKTS_IN_ADD(f, n_pkts_in);
1053
1054
1055 if (__builtin_popcountll(pkts_mask) < 5) {
1056 for ( ; pkts_mask; ) {
1057 struct rte_bucket_4_32 *bucket;
1058 struct rte_mbuf *mbuf;
1059 uint32_t pkt_index;
1060
1061 lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask, f);
1062 lookup1_stage1(mbuf, bucket, f);
1063 lookup1_stage2_ext(pkt_index, mbuf, bucket,
1064 pkts_mask_out, entries, buckets_mask, buckets,
1065 keys, f);
1066 }
1067
1068 goto grind_next_buckets;
1069 }
1070
1071
1072
1073
1074
1075
1076 lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1077 pkts_mask, f);
1078
1079
1080 mbuf10 = mbuf00;
1081 mbuf11 = mbuf01;
1082 pkt10_index = pkt00_index;
1083 pkt11_index = pkt01_index;
1084
1085
1086 lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1087 pkts_mask, f);
1088
1089
1090 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1091
1092
1093
1094
1095
1096 for ( ; pkts_mask; ) {
1097
1098 bucket20 = bucket10;
1099 bucket21 = bucket11;
1100 mbuf20 = mbuf10;
1101 mbuf21 = mbuf11;
1102 mbuf10 = mbuf00;
1103 mbuf11 = mbuf01;
1104 pkt20_index = pkt10_index;
1105 pkt21_index = pkt11_index;
1106 pkt10_index = pkt00_index;
1107 pkt11_index = pkt01_index;
1108
1109
1110 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1111 mbuf00, mbuf01, pkts, pkts_mask, f);
1112
1113
1114 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1115
1116
1117 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1118 bucket20, bucket21, pkts_mask_out, entries,
1119 buckets_mask, buckets, keys, f);
1120 }
1121
1122
1123
1124
1125
1126
1127 bucket20 = bucket10;
1128 bucket21 = bucket11;
1129 mbuf20 = mbuf10;
1130 mbuf21 = mbuf11;
1131 mbuf10 = mbuf00;
1132 mbuf11 = mbuf01;
1133 pkt20_index = pkt10_index;
1134 pkt21_index = pkt11_index;
1135 pkt10_index = pkt00_index;
1136 pkt11_index = pkt01_index;
1137
1138
1139 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1140
1141
1142 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1143 bucket20, bucket21, pkts_mask_out, entries,
1144 buckets_mask, buckets, keys, f);
1145
1146
1147 bucket20 = bucket10;
1148 bucket21 = bucket11;
1149 mbuf20 = mbuf10;
1150 mbuf21 = mbuf11;
1151 pkt20_index = pkt10_index;
1152 pkt21_index = pkt11_index;
1153
1154
1155 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1156 bucket20, bucket21, pkts_mask_out, entries,
1157 buckets_mask, buckets, keys, f);
1158
1159grind_next_buckets:
1160
1161 for ( ; buckets_mask; ) {
1162 uint64_t buckets_mask_next = 0;
1163
1164 for ( ; buckets_mask; ) {
1165 uint64_t pkt_mask;
1166 uint32_t pkt_index;
1167
1168 pkt_index = __builtin_ctzll(buckets_mask);
1169 pkt_mask = 1LLU << pkt_index;
1170 buckets_mask &= ~pkt_mask;
1171
1172 lookup_grinder(pkt_index, buckets, keys, pkts_mask_out,
1173 entries, buckets_mask_next, f);
1174 }
1175
1176 buckets_mask = buckets_mask_next;
1177 }
1178
1179 *lookup_hit_mask = pkts_mask_out;
1180 RTE_TABLE_HASH_KEY32_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1181 return 0;
1182}
1183
1184static int
1185rte_table_hash_key32_stats_read(void *table, struct rte_table_stats *stats, int clear)
1186{
1187 struct rte_table_hash *t = table;
1188
1189 if (stats != NULL)
1190 memcpy(stats, &t->stats, sizeof(t->stats));
1191
1192 if (clear)
1193 memset(&t->stats, 0, sizeof(t->stats));
1194
1195 return 0;
1196}
1197
1198struct rte_table_ops rte_table_hash_key32_lru_ops = {
1199 .f_create = rte_table_hash_create_key32_lru,
1200 .f_free = rte_table_hash_free_key32_lru,
1201 .f_add = rte_table_hash_entry_add_key32_lru,
1202 .f_delete = rte_table_hash_entry_delete_key32_lru,
1203 .f_add_bulk = NULL,
1204 .f_delete_bulk = NULL,
1205 .f_lookup = rte_table_hash_lookup_key32_lru,
1206 .f_stats = rte_table_hash_key32_stats_read,
1207};
1208
1209struct rte_table_ops rte_table_hash_key32_ext_ops = {
1210 .f_create = rte_table_hash_create_key32_ext,
1211 .f_free = rte_table_hash_free_key32_ext,
1212 .f_add = rte_table_hash_entry_add_key32_ext,
1213 .f_delete = rte_table_hash_entry_delete_key32_ext,
1214 .f_add_bulk = NULL,
1215 .f_delete_bulk = NULL,
1216 .f_lookup = rte_table_hash_lookup_key32_ext,
1217 .f_stats = rte_table_hash_key32_stats_read,
1218};
1219
