net/virtio: fix incorrect cast of void *
[dpdk.git] / table / rte_table_hash_lru.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2017 Intel Corporation
3  */
4
5 #include <string.h>
6 #include <stdio.h>
7
8 #include <rte_common.h>
9 #include <rte_mbuf.h>
10 #include <rte_memory.h>
11 #include <rte_malloc.h>
12 #include <rte_log.h>
13
14 #include "rte_table_hash.h"
15 #include "rte_lru.h"
16
17 #define KEYS_PER_BUCKET 4
18
19 #ifdef RTE_TABLE_STATS_COLLECT
20
21 #define RTE_TABLE_HASH_LRU_STATS_PKTS_IN_ADD(table, val) \
22         table->stats.n_pkts_in += val
23 #define RTE_TABLE_HASH_LRU_STATS_PKTS_LOOKUP_MISS(table, val) \
24         table->stats.n_pkts_lookup_miss += val
25
26 #else
27
28 #define RTE_TABLE_HASH_LRU_STATS_PKTS_IN_ADD(table, val)
29 #define RTE_TABLE_HASH_LRU_STATS_PKTS_LOOKUP_MISS(table, val)
30
31 #endif
32
33 struct bucket {
34         union {
35                 struct bucket *next;
36                 uint64_t lru_list;
37         };
38         uint16_t sig[KEYS_PER_BUCKET];
39         uint32_t key_pos[KEYS_PER_BUCKET];
40 };
41
42 struct grinder {
43         struct bucket *bkt;
44         uint64_t sig;
45         uint64_t match;
46         uint64_t match_pos;
47         uint32_t key_index;
48 };
49
50 struct rte_table_hash {
51         struct rte_table_stats stats;
52
53         /* Input parameters */
54         uint32_t key_size;
55         uint32_t entry_size;
56         uint32_t n_keys;
57         uint32_t n_buckets;
58         rte_table_hash_op_hash f_hash;
59         uint64_t seed;
60         uint32_t key_offset;
61
62         /* Internal */
63         uint64_t bucket_mask;
64         uint32_t key_size_shl;
65         uint32_t data_size_shl;
66         uint32_t key_stack_tos;
67
68         /* Grinder */
69         struct grinder grinders[RTE_PORT_IN_BURST_SIZE_MAX];
70
71         /* Tables */
72         uint64_t *key_mask;
73         struct bucket *buckets;
74         uint8_t *key_mem;
75         uint8_t *data_mem;
76         uint32_t *key_stack;
77
78         /* Table memory */
79         uint8_t memory[0] __rte_cache_aligned;
80 };
81
82 static int
83 keycmp(void *a, void *b, void *b_mask, uint32_t n_bytes)
84 {
85         uint64_t *a64 = a, *b64 = b, *b_mask64 = b_mask;
86         uint32_t i;
87
88         for (i = 0; i < n_bytes / sizeof(uint64_t); i++)
89                 if (a64[i] != (b64[i] & b_mask64[i]))
90                         return 1;
91
92         return 0;
93 }
94
95 static void
96 keycpy(void *dst, void *src, void *src_mask, uint32_t n_bytes)
97 {
98         uint64_t *dst64 = dst, *src64 = src, *src_mask64 = src_mask;
99         uint32_t i;
100
101         for (i = 0; i < n_bytes / sizeof(uint64_t); i++)
102                 dst64[i] = src64[i] & src_mask64[i];
103 }
104
105 static int
106 check_params_create(struct rte_table_hash_params *params)
107 {
108         /* name */
109         if (params->name == NULL) {
110                 RTE_LOG(ERR, TABLE, "%s: name invalid value\n", __func__);
111                 return -EINVAL;
112         }
113
114         /* key_size */
115         if ((params->key_size < sizeof(uint64_t)) ||
116                 (!rte_is_power_of_2(params->key_size))) {
117                 RTE_LOG(ERR, TABLE, "%s: key_size invalid value\n", __func__);
118                 return -EINVAL;
119         }
120
121         /* n_keys */
122         if (params->n_keys == 0) {
123                 RTE_LOG(ERR, TABLE, "%s: n_keys invalid value\n", __func__);
124                 return -EINVAL;
125         }
126
127         /* n_buckets */
128         if ((params->n_buckets == 0) ||
129                 (!rte_is_power_of_2(params->n_buckets))) {
130                 RTE_LOG(ERR, TABLE, "%s: n_buckets invalid value\n", __func__);
131                 return -EINVAL;
132         }
133
134         /* f_hash */
135         if (params->f_hash == NULL) {
136                 RTE_LOG(ERR, TABLE, "%s: f_hash invalid value\n", __func__);
137                 return -EINVAL;
138         }
139
140         return 0;
141 }
142
143 static void *
144 rte_table_hash_lru_create(void *params, int socket_id, uint32_t entry_size)
145 {
146         struct rte_table_hash_params *p = params;
147         struct rte_table_hash *t;
148         uint64_t table_meta_sz, key_mask_sz, bucket_sz, key_sz, key_stack_sz;
149         uint64_t data_sz, total_size;
150         uint64_t key_mask_offset, bucket_offset, key_offset, key_stack_offset;
151         uint64_t data_offset;
152         uint32_t n_buckets, i;
153
154         /* Check input parameters */
155         if ((check_params_create(p) != 0) ||
156                 (!rte_is_power_of_2(entry_size)) ||
157                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
158                 (sizeof(struct bucket) != (RTE_CACHE_LINE_SIZE / 2))) {
159                 return NULL;
160         }
161
162         /*
163          * Table dimensioning
164          *
165          * Objective: Pick the number of buckets (n_buckets) so that there a chance
166          * to store n_keys keys in the table.
167          *
168          * Note: Since the buckets do not get extended, it is not possible to
169          * guarantee that n_keys keys can be stored in the table at any time. In the
170          * worst case scenario when all the n_keys fall into the same bucket, only
171          * a maximum of KEYS_PER_BUCKET keys will be stored in the table. This case
172          * defeats the purpose of the hash table. It indicates unsuitable f_hash or
173          * n_keys to n_buckets ratio.
174          *
175          * MIN(n_buckets) = (n_keys + KEYS_PER_BUCKET - 1) / KEYS_PER_BUCKET
176          */
177         n_buckets = rte_align32pow2(
178                 (p->n_keys + KEYS_PER_BUCKET - 1) / KEYS_PER_BUCKET);
179         n_buckets = RTE_MAX(n_buckets, p->n_buckets);
180
181         /* Memory allocation */
182         table_meta_sz = RTE_CACHE_LINE_ROUNDUP(sizeof(struct rte_table_hash));
183         key_mask_sz = RTE_CACHE_LINE_ROUNDUP(p->key_size);
184         bucket_sz = RTE_CACHE_LINE_ROUNDUP(n_buckets * sizeof(struct bucket));
185         key_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * p->key_size);
186         key_stack_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * sizeof(uint32_t));
187         data_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * entry_size);
188         total_size = table_meta_sz + key_mask_sz + bucket_sz + key_sz +
189                 key_stack_sz + data_sz;
190
191         if (total_size > SIZE_MAX) {
192                 RTE_LOG(ERR, TABLE,
193                         "%s: Cannot allocate %" PRIu64 " bytes for hash "
194                         "table %s\n",
195                         __func__, total_size, p->name);
196                 return NULL;
197         }
198
199         t = rte_zmalloc_socket(p->name,
200                 (size_t)total_size,
201                 RTE_CACHE_LINE_SIZE,
202                 socket_id);
203         if (t == NULL) {
204                 RTE_LOG(ERR, TABLE,
205                         "%s: Cannot allocate %" PRIu64 " bytes for hash "
206                         "table %s\n",
207                         __func__, total_size, p->name);
208                 return NULL;
209         }
210         RTE_LOG(INFO, TABLE, "%s (%u-byte key): Hash table %s memory footprint"
211                 " is %" PRIu64 " bytes\n",
212                 __func__, p->key_size, p->name, total_size);
213
214         /* Memory initialization */
215         t->key_size = p->key_size;
216         t->entry_size = entry_size;
217         t->n_keys = p->n_keys;
218         t->n_buckets = n_buckets;
219         t->f_hash = p->f_hash;
220         t->seed = p->seed;
221         t->key_offset = p->key_offset;
222
223         /* Internal */
224         t->bucket_mask = t->n_buckets - 1;
225         t->key_size_shl = __builtin_ctzl(p->key_size);
226         t->data_size_shl = __builtin_ctzl(entry_size);
227
228         /* Tables */
229         key_mask_offset = 0;
230         bucket_offset = key_mask_offset + key_mask_sz;
231         key_offset = bucket_offset + bucket_sz;
232         key_stack_offset = key_offset + key_sz;
233         data_offset = key_stack_offset + key_stack_sz;
234
235         t->key_mask = (uint64_t *) &t->memory[key_mask_offset];
236         t->buckets = (struct bucket *) &t->memory[bucket_offset];
237         t->key_mem = &t->memory[key_offset];
238         t->key_stack = (uint32_t *) &t->memory[key_stack_offset];
239         t->data_mem = &t->memory[data_offset];
240
241         /* Key mask */
242         if (p->key_mask == NULL)
243                 memset(t->key_mask, 0xFF, p->key_size);
244         else
245                 memcpy(t->key_mask, p->key_mask, p->key_size);
246
247         /* Key stack */
248         for (i = 0; i < t->n_keys; i++)
249                 t->key_stack[i] = t->n_keys - 1 - i;
250         t->key_stack_tos = t->n_keys;
251
252         /* LRU */
253         for (i = 0; i < t->n_buckets; i++) {
254                 struct bucket *bkt = &t->buckets[i];
255
256                 lru_init(bkt);
257         }
258
259         return t;
260 }
261
262 static int
263 rte_table_hash_lru_free(void *table)
264 {
265         struct rte_table_hash *t = table;
266
267         /* Check input parameters */
268         if (t == NULL)
269                 return -EINVAL;
270
271         rte_free(t);
272         return 0;
273 }
274
275 static int
276 rte_table_hash_lru_entry_add(void *table, void *key, void *entry,
277         int *key_found, void **entry_ptr)
278 {
279         struct rte_table_hash *t = table;
280         struct bucket *bkt;
281         uint64_t sig;
282         uint32_t bkt_index, i;
283
284         sig = t->f_hash(key, t->key_mask, t->key_size, t->seed);
285         bkt_index = sig & t->bucket_mask;
286         bkt = &t->buckets[bkt_index];
287         sig = (sig >> 16) | 1LLU;
288
289         /* Key is present in the bucket */
290         for (i = 0; i < KEYS_PER_BUCKET; i++) {
291                 uint64_t bkt_sig = (uint64_t) bkt->sig[i];
292                 uint32_t bkt_key_index = bkt->key_pos[i];
293                 uint8_t *bkt_key = &t->key_mem[bkt_key_index <<
294                         t->key_size_shl];
295
296                 if ((sig == bkt_sig) && (keycmp(bkt_key, key, t->key_mask,
297                         t->key_size) == 0)) {
298                         uint8_t *data = &t->data_mem[bkt_key_index <<
299                                 t->data_size_shl];
300
301                         memcpy(data, entry, t->entry_size);
302                         lru_update(bkt, i);
303                         *key_found = 1;
304                         *entry_ptr = (void *) data;
305                         return 0;
306                 }
307         }
308
309         /* Key is not present in the bucket */
310         for (i = 0; i < KEYS_PER_BUCKET; i++) {
311                 uint64_t bkt_sig = (uint64_t) bkt->sig[i];
312
313                 if (bkt_sig == 0) {
314                         uint32_t bkt_key_index;
315                         uint8_t *bkt_key, *data;
316
317                         /* Allocate new key */
318                         if (t->key_stack_tos == 0) {
319                                 /* No keys available */
320                                 return -ENOSPC;
321                         }
322                         bkt_key_index = t->key_stack[--t->key_stack_tos];
323
324                         /* Install new key */
325                         bkt_key = &t->key_mem[bkt_key_index << t->key_size_shl];
326                         data = &t->data_mem[bkt_key_index << t->data_size_shl];
327
328                         bkt->sig[i] = (uint16_t) sig;
329                         bkt->key_pos[i] = bkt_key_index;
330                         keycpy(bkt_key, key, t->key_mask, t->key_size);
331                         memcpy(data, entry, t->entry_size);
332                         lru_update(bkt, i);
333
334                         *key_found = 0;
335                         *entry_ptr = (void *) data;
336                         return 0;
337                 }
338         }
339
340         /* Bucket full */
341         {
342                 uint64_t pos = lru_pos(bkt);
343                 uint32_t bkt_key_index = bkt->key_pos[pos];
344                 uint8_t *bkt_key = &t->key_mem[bkt_key_index <<
345                         t->key_size_shl];
346                 uint8_t *data = &t->data_mem[bkt_key_index << t->data_size_shl];
347
348                 bkt->sig[pos] = (uint16_t) sig;
349                 keycpy(bkt_key, key, t->key_mask, t->key_size);
350                 memcpy(data, entry, t->entry_size);
351                 lru_update(bkt, pos);
352
353                 *key_found = 0;
354                 *entry_ptr = (void *) data;
355                 return 0;
356         }
357 }
358
359 static int
360 rte_table_hash_lru_entry_delete(void *table, void *key, int *key_found,
361         void *entry)
362 {
363         struct rte_table_hash *t = table;
364         struct bucket *bkt;
365         uint64_t sig;
366         uint32_t bkt_index, i;
367
368         sig = t->f_hash(key, t->key_mask, t->key_size, t->seed);
369         bkt_index = sig & t->bucket_mask;
370         bkt = &t->buckets[bkt_index];
371         sig = (sig >> 16) | 1LLU;
372
373         /* Key is present in the bucket */
374         for (i = 0; i < KEYS_PER_BUCKET; i++) {
375                 uint64_t bkt_sig = (uint64_t) bkt->sig[i];
376                 uint32_t bkt_key_index = bkt->key_pos[i];
377                 uint8_t *bkt_key = &t->key_mem[bkt_key_index <<
378                         t->key_size_shl];
379
380                 if ((sig == bkt_sig) &&
381                         (keycmp(bkt_key, key, t->key_mask, t->key_size) == 0)) {
382                         uint8_t *data = &t->data_mem[bkt_key_index <<
383                                 t->data_size_shl];
384
385                         bkt->sig[i] = 0;
386                         t->key_stack[t->key_stack_tos++] = bkt_key_index;
387                         *key_found = 1;
388                         if (entry)
389                                 memcpy(entry, data, t->entry_size);
390                         return 0;
391                 }
392         }
393
394         /* Key is not present in the bucket */
395         *key_found = 0;
396         return 0;
397 }
398
399 static int rte_table_hash_lru_lookup_unoptimized(
400         void *table,
401         struct rte_mbuf **pkts,
402         uint64_t pkts_mask,
403         uint64_t *lookup_hit_mask,
404         void **entries)
405 {
406         struct rte_table_hash *t = (struct rte_table_hash *) table;
407         uint64_t pkts_mask_out = 0;
408
409         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
410         RTE_TABLE_HASH_LRU_STATS_PKTS_IN_ADD(t, n_pkts_in);
411
412         for ( ; pkts_mask; ) {
413                 struct bucket *bkt;
414                 struct rte_mbuf *pkt;
415                 uint8_t *key;
416                 uint64_t pkt_mask, sig;
417                 uint32_t pkt_index, bkt_index, i;
418
419                 pkt_index = __builtin_ctzll(pkts_mask);
420                 pkt_mask = 1LLU << pkt_index;
421                 pkts_mask &= ~pkt_mask;
422
423                 pkt = pkts[pkt_index];
424                 key = RTE_MBUF_METADATA_UINT8_PTR(pkt, t->key_offset);
425                 sig = (uint64_t) t->f_hash(key, t->key_mask, t->key_size, t->seed);
426
427                 bkt_index = sig & t->bucket_mask;
428                 bkt = &t->buckets[bkt_index];
429                 sig = (sig >> 16) | 1LLU;
430
431                 /* Key is present in the bucket */
432                 for (i = 0; i < KEYS_PER_BUCKET; i++) {
433                         uint64_t bkt_sig = (uint64_t) bkt->sig[i];
434                         uint32_t bkt_key_index = bkt->key_pos[i];
435                         uint8_t *bkt_key = &t->key_mem[bkt_key_index <<
436                                 t->key_size_shl];
437
438                         if ((sig == bkt_sig) && (keycmp(bkt_key, key, t->key_mask,
439                                 t->key_size) == 0)) {
440                                 uint8_t *data = &t->data_mem[bkt_key_index <<
441                                         t->data_size_shl];
442
443                                 lru_update(bkt, i);
444                                 pkts_mask_out |= pkt_mask;
445                                 entries[pkt_index] = (void *) data;
446                                 break;
447                         }
448                 }
449         }
450
451         *lookup_hit_mask = pkts_mask_out;
452         RTE_TABLE_HASH_LRU_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - __builtin_popcountll(pkts_mask_out));
453         return 0;
454 }
455
456 /***
457 *
458 * mask = match bitmask
459 * match = at least one match
460 * match_many = more than one match
461 * match_pos = position of first match
462 *
463 * ----------------------------------------
464 * mask           match   match_many       match_pos
465 * ----------------------------------------
466 * 0000           0               0                        00
467 * 0001           1               0                        00
468 * 0010           1               0                        01
469 * 0011           1               1                        00
470 * ----------------------------------------
471 * 0100           1               0                        10
472 * 0101           1               1                        00
473 * 0110           1               1                        01
474 * 0111           1               1                        00
475 * ----------------------------------------
476 * 1000           1               0                        11
477 * 1001           1               1                        00
478 * 1010           1               1                        01
479 * 1011           1               1                        00
480 * ----------------------------------------
481 * 1100           1               1                        10
482 * 1101           1               1                        00
483 * 1110           1               1                        01
484 * 1111           1               1                        00
485 * ----------------------------------------
486 *
487 * match = 1111_1111_1111_1110
488 * match_many = 1111_1110_1110_1000
489 * match_pos = 0001_0010_0001_0011__0001_0010_0001_0000
490 *
491 * match = 0xFFFELLU
492 * match_many = 0xFEE8LLU
493 * match_pos = 0x12131210LLU
494 *
495 ***/
496
497 #define LUT_MATCH                                               0xFFFELLU
498 #define LUT_MATCH_MANY                                          0xFEE8LLU
499 #define LUT_MATCH_POS                                           0x12131210LLU
500
501 #define lookup_cmp_sig(mbuf_sig, bucket, match, match_many, match_pos)\
502 {                                                               \
503         uint64_t bucket_sig[4], mask[4], mask_all;              \
504                                                                 \
505         bucket_sig[0] = bucket->sig[0];                         \
506         bucket_sig[1] = bucket->sig[1];                         \
507         bucket_sig[2] = bucket->sig[2];                         \
508         bucket_sig[3] = bucket->sig[3];                         \
509                                                                 \
510         bucket_sig[0] ^= mbuf_sig;                              \
511         bucket_sig[1] ^= mbuf_sig;                              \
512         bucket_sig[2] ^= mbuf_sig;                              \
513         bucket_sig[3] ^= mbuf_sig;                              \
514                                                                 \
515         mask[0] = 0;                                            \
516         mask[1] = 0;                                            \
517         mask[2] = 0;                                            \
518         mask[3] = 0;                                            \
519                                                                 \
520         if (bucket_sig[0] == 0)                                 \
521                 mask[0] = 1;                                    \
522         if (bucket_sig[1] == 0)                                 \
523                 mask[1] = 2;                                    \
524         if (bucket_sig[2] == 0)                                 \
525                 mask[2] = 4;                                    \
526         if (bucket_sig[3] == 0)                                 \
527                 mask[3] = 8;                                    \
528                                                                 \
529         mask_all = (mask[0] | mask[1]) | (mask[2] | mask[3]);   \
530                                                                 \
531         match = (LUT_MATCH >> mask_all) & 1;                    \
532         match_many = (LUT_MATCH_MANY >> mask_all) & 1;          \
533         match_pos = (LUT_MATCH_POS >> (mask_all << 1)) & 3;     \
534 }
535
536 #define lookup_cmp_key(mbuf, key, match_key, f)                         \
537 {                                                                       \
538         uint64_t *pkt_key = RTE_MBUF_METADATA_UINT64_PTR(mbuf, f->key_offset);\
539         uint64_t *bkt_key = (uint64_t *) key;                           \
540         uint64_t *key_mask = f->key_mask;                                       \
541                                                                         \
542         switch (f->key_size) {                                          \
543         case 8:                                                         \
544         {                                                               \
545                 uint64_t xor = (pkt_key[0] & key_mask[0]) ^ bkt_key[0]; \
546                 match_key = 0;                                          \
547                 if (xor == 0)                                           \
548                         match_key = 1;                                  \
549         }                                                               \
550         break;                                                          \
551                                                                         \
552         case 16:                                                        \
553         {                                                               \
554                 uint64_t xor[2], or;                                    \
555                                                                         \
556                 xor[0] = (pkt_key[0] & key_mask[0]) ^ bkt_key[0];               \
557                 xor[1] = (pkt_key[1] & key_mask[1]) ^ bkt_key[1];               \
558                 or = xor[0] | xor[1];                                   \
559                 match_key = 0;                                          \
560                 if (or == 0)                                            \
561                         match_key = 1;                                  \
562         }                                                               \
563         break;                                                          \
564                                                                         \
565         case 32:                                                        \
566         {                                                               \
567                 uint64_t xor[4], or;                                    \
568                                                                         \
569                 xor[0] = (pkt_key[0] & key_mask[0]) ^ bkt_key[0];               \
570                 xor[1] = (pkt_key[1] & key_mask[1]) ^ bkt_key[1];               \
571                 xor[2] = (pkt_key[2] & key_mask[2]) ^ bkt_key[2];               \
572                 xor[3] = (pkt_key[3] & key_mask[3]) ^ bkt_key[3];               \
573                 or = xor[0] | xor[1] | xor[2] | xor[3];                 \
574                 match_key = 0;                                          \
575                 if (or == 0)                                            \
576                         match_key = 1;                                  \
577         }                                                               \
578         break;                                                          \
579                                                                         \
580         case 64:                                                        \
581         {                                                               \
582                 uint64_t xor[8], or;                                    \
583                                                                         \
584                 xor[0] = (pkt_key[0] & key_mask[0]) ^ bkt_key[0];               \
585                 xor[1] = (pkt_key[1] & key_mask[1]) ^ bkt_key[1];               \
586                 xor[2] = (pkt_key[2] & key_mask[2]) ^ bkt_key[2];               \
587                 xor[3] = (pkt_key[3] & key_mask[3]) ^ bkt_key[3];               \
588                 xor[4] = (pkt_key[4] & key_mask[4]) ^ bkt_key[4];               \
589                 xor[5] = (pkt_key[5] & key_mask[5]) ^ bkt_key[5];               \
590                 xor[6] = (pkt_key[6] & key_mask[6]) ^ bkt_key[6];               \
591                 xor[7] = (pkt_key[7] & key_mask[7]) ^ bkt_key[7];               \
592                 or = xor[0] | xor[1] | xor[2] | xor[3] |                \
593                         xor[4] | xor[5] | xor[6] | xor[7];              \
594                 match_key = 0;                                          \
595                 if (or == 0)                                            \
596                         match_key = 1;                                  \
597         }                                                               \
598         break;                                                          \
599                                                                         \
600         default:                                                        \
601                 match_key = 0;                                          \
602                 if (keycmp(bkt_key, pkt_key, key_mask, f->key_size) == 0)       \
603                         match_key = 1;                                  \
604         }                                                               \
605 }
606
607 #define lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index)\
608 {                                                               \
609         uint64_t pkt00_mask, pkt01_mask;                        \
610         struct rte_mbuf *mbuf00, *mbuf01;                       \
611         uint32_t key_offset = t->key_offset;            \
612                                                                 \
613         pkt00_index = __builtin_ctzll(pkts_mask);               \
614         pkt00_mask = 1LLU << pkt00_index;                       \
615         pkts_mask &= ~pkt00_mask;                               \
616         mbuf00 = pkts[pkt00_index];                             \
617                                                                 \
618         pkt01_index = __builtin_ctzll(pkts_mask);               \
619         pkt01_mask = 1LLU << pkt01_index;                       \
620         pkts_mask &= ~pkt01_mask;                               \
621         mbuf01 = pkts[pkt01_index];                             \
622                                                                 \
623         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, key_offset));\
624         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, key_offset));\
625 }
626
627 #define lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask, pkt00_index, \
628         pkt01_index)                                            \
629 {                                                               \
630         uint64_t pkt00_mask, pkt01_mask;                        \
631         struct rte_mbuf *mbuf00, *mbuf01;                       \
632         uint32_t key_offset = t->key_offset;            \
633                                                                 \
634         pkt00_index = __builtin_ctzll(pkts_mask);               \
635         pkt00_mask = 1LLU << pkt00_index;                       \
636         pkts_mask &= ~pkt00_mask;                               \
637         mbuf00 = pkts[pkt00_index];                             \
638                                                                 \
639         pkt01_index = __builtin_ctzll(pkts_mask);               \
640         if (pkts_mask == 0)                                     \
641                 pkt01_index = pkt00_index;                      \
642                                                                 \
643         pkt01_mask = 1LLU << pkt01_index;                       \
644         pkts_mask &= ~pkt01_mask;                               \
645         mbuf01 = pkts[pkt01_index];                             \
646                                                                 \
647         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, key_offset));\
648         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, key_offset));\
649 }
650
651 #define lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index)\
652 {                                                               \
653         struct grinder *g10, *g11;                              \
654         uint64_t sig10, sig11, bkt10_index, bkt11_index;        \
655         struct rte_mbuf *mbuf10, *mbuf11;                       \
656         struct bucket *bkt10, *bkt11, *buckets = t->buckets;    \
657         uint8_t *key10, *key11;                                 \
658         uint64_t bucket_mask = t->bucket_mask;                  \
659         rte_table_hash_op_hash f_hash = t->f_hash;              \
660         uint64_t seed = t->seed;                                \
661         uint32_t key_size = t->key_size;                        \
662         uint32_t key_offset = t->key_offset;                    \
663                                                                 \
664         mbuf10 = pkts[pkt10_index];                             \
665         key10 = RTE_MBUF_METADATA_UINT8_PTR(mbuf10, key_offset);\
666         sig10 = (uint64_t) f_hash(key10, t->key_mask, key_size, seed);\
667         bkt10_index = sig10 & bucket_mask;                      \
668         bkt10 = &buckets[bkt10_index];                          \
669                                                                 \
670         mbuf11 = pkts[pkt11_index];                             \
671         key11 = RTE_MBUF_METADATA_UINT8_PTR(mbuf11, key_offset);\
672         sig11 = (uint64_t) f_hash(key11, t->key_mask, key_size, seed);\
673         bkt11_index = sig11 & bucket_mask;                      \
674         bkt11 = &buckets[bkt11_index];                          \
675                                                                 \
676         rte_prefetch0(bkt10);                                   \
677         rte_prefetch0(bkt11);                                   \
678                                                                 \
679         g10 = &g[pkt10_index];                                  \
680         g10->sig = sig10;                                       \
681         g10->bkt = bkt10;                                       \
682                                                                 \
683         g11 = &g[pkt11_index];                                  \
684         g11->sig = sig11;                                       \
685         g11->bkt = bkt11;                                       \
686 }
687
688 #define lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many)\
689 {                                                               \
690         struct grinder *g20, *g21;                              \
691         uint64_t sig20, sig21;                                  \
692         struct bucket *bkt20, *bkt21;                           \
693         uint8_t *key20, *key21, *key_mem = t->key_mem;          \
694         uint64_t match20, match21, match_many20, match_many21;  \
695         uint64_t match_pos20, match_pos21;                      \
696         uint32_t key20_index, key21_index, key_size_shl = t->key_size_shl;\
697                                                                 \
698         g20 = &g[pkt20_index];                                  \
699         sig20 = g20->sig;                                       \
700         bkt20 = g20->bkt;                                       \
701         sig20 = (sig20 >> 16) | 1LLU;                           \
702         lookup_cmp_sig(sig20, bkt20, match20, match_many20, match_pos20);\
703         match20 <<= pkt20_index;                                \
704         match_many20 <<= pkt20_index;                           \
705         key20_index = bkt20->key_pos[match_pos20];              \
706         key20 = &key_mem[key20_index << key_size_shl];          \
707                                                                 \
708         g21 = &g[pkt21_index];                                  \
709         sig21 = g21->sig;                                       \
710         bkt21 = g21->bkt;                                       \
711         sig21 = (sig21 >> 16) | 1LLU;                           \
712         lookup_cmp_sig(sig21, bkt21, match21, match_many21, match_pos21);\
713         match21 <<= pkt21_index;                                \
714         match_many21 <<= pkt21_index;                           \
715         key21_index = bkt21->key_pos[match_pos21];              \
716         key21 = &key_mem[key21_index << key_size_shl];          \
717                                                                 \
718         rte_prefetch0(key20);                                   \
719         rte_prefetch0(key21);                                   \
720                                                                 \
721         pkts_mask_match_many |= match_many20 | match_many21;    \
722                                                                 \
723         g20->match = match20;                                   \
724         g20->match_pos = match_pos20;                           \
725         g20->key_index = key20_index;                           \
726                                                                 \
727         g21->match = match21;                                   \
728         g21->match_pos = match_pos21;                           \
729         g21->key_index = key21_index;                           \
730 }
731
732 #define lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out, \
733         entries)                                                \
734 {                                                               \
735         struct grinder *g30, *g31;                              \
736         struct rte_mbuf *mbuf30, *mbuf31;                       \
737         struct bucket *bkt30, *bkt31;                           \
738         uint8_t *key30, *key31, *key_mem = t->key_mem;          \
739         uint8_t *data30, *data31, *data_mem = t->data_mem;      \
740         uint64_t match30, match31, match_pos30, match_pos31;    \
741         uint64_t match_key30, match_key31, match_keys;          \
742         uint32_t key30_index, key31_index;                      \
743         uint32_t key_size_shl = t->key_size_shl;                \
744         uint32_t data_size_shl = t->data_size_shl;              \
745                                                                 \
746         mbuf30 = pkts[pkt30_index];                             \
747         g30 = &g[pkt30_index];                                  \
748         bkt30 = g30->bkt;                                       \
749         match30 = g30->match;                                   \
750         match_pos30 = g30->match_pos;                           \
751         key30_index = g30->key_index;                           \
752         key30 = &key_mem[key30_index << key_size_shl];          \
753         lookup_cmp_key(mbuf30, key30, match_key30, t);          \
754         match_key30 <<= pkt30_index;                            \
755         match_key30 &= match30;                                 \
756         data30 = &data_mem[key30_index << data_size_shl];       \
757         entries[pkt30_index] = data30;                          \
758                                                                 \
759         mbuf31 = pkts[pkt31_index];                             \
760         g31 = &g[pkt31_index];                                  \
761         bkt31 = g31->bkt;                                       \
762         match31 = g31->match;                                   \
763         match_pos31 = g31->match_pos;                           \
764         key31_index = g31->key_index;                           \
765         key31 = &key_mem[key31_index << key_size_shl];          \
766         lookup_cmp_key(mbuf31, key31, match_key31, t);          \
767         match_key31 <<= pkt31_index;                            \
768         match_key31 &= match31;                                 \
769         data31 = &data_mem[key31_index << data_size_shl];       \
770         entries[pkt31_index] = data31;                          \
771                                                                 \
772         rte_prefetch0(data30);                                  \
773         rte_prefetch0(data31);                                  \
774                                                                 \
775         match_keys = match_key30 | match_key31;                 \
776         pkts_mask_out |= match_keys;                            \
777                                                                 \
778         if (match_key30 == 0)                                   \
779                 match_pos30 = 4;                                \
780         lru_update(bkt30, match_pos30);                         \
781                                                                 \
782         if (match_key31 == 0)                                   \
783                 match_pos31 = 4;                                \
784         lru_update(bkt31, match_pos31);                         \
785 }
786
787 /***
788 * The lookup function implements a 4-stage pipeline, with each stage processing
789 * two different packets. The purpose of pipelined implementation is to hide the
790 * latency of prefetching the data structures and loosen the data dependency
791 * between instructions.
792 *
793 *   p00  _______   p10  _______   p20  _______   p30  _______
794 * ----->|       |----->|       |----->|       |----->|       |----->
795 *       |   0   |      |   1   |      |   2   |      |   3   |
796 * ----->|_______|----->|_______|----->|_______|----->|_______|----->
797 *   p01            p11            p21            p31
798 *
799 * The naming convention is:
800 *         pXY = packet Y of stage X, X = 0 .. 3, Y = 0 .. 1
801 *
802 ***/
803 static int rte_table_hash_lru_lookup(
804         void *table,
805         struct rte_mbuf **pkts,
806         uint64_t pkts_mask,
807         uint64_t *lookup_hit_mask,
808         void **entries)
809 {
810         struct rte_table_hash *t = (struct rte_table_hash *) table;
811         struct grinder *g = t->grinders;
812         uint64_t pkt00_index, pkt01_index, pkt10_index, pkt11_index;
813         uint64_t pkt20_index, pkt21_index, pkt30_index, pkt31_index;
814         uint64_t pkts_mask_out = 0, pkts_mask_match_many = 0;
815         int status = 0;
816
817         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
818         RTE_TABLE_HASH_LRU_STATS_PKTS_IN_ADD(t, n_pkts_in);
819
820         /* Cannot run the pipeline with less than 7 packets */
821         if (__builtin_popcountll(pkts_mask) < 7)
822                 return rte_table_hash_lru_lookup_unoptimized(table, pkts,
823                         pkts_mask, lookup_hit_mask, entries);
824
825         /* Pipeline stage 0 */
826         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
827
828         /* Pipeline feed */
829         pkt10_index = pkt00_index;
830         pkt11_index = pkt01_index;
831
832         /* Pipeline stage 0 */
833         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
834
835         /* Pipeline stage 1 */
836         lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
837
838         /* Pipeline feed */
839         pkt20_index = pkt10_index;
840         pkt21_index = pkt11_index;
841         pkt10_index = pkt00_index;
842         pkt11_index = pkt01_index;
843
844         /* Pipeline stage 0 */
845         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
846
847         /* Pipeline stage 1 */
848         lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
849
850         /* Pipeline stage 2 */
851         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
852
853         /*
854         * Pipeline run
855         *
856         */
857         for ( ; pkts_mask; ) {
858                 /* Pipeline feed */
859                 pkt30_index = pkt20_index;
860                 pkt31_index = pkt21_index;
861                 pkt20_index = pkt10_index;
862                 pkt21_index = pkt11_index;
863                 pkt10_index = pkt00_index;
864                 pkt11_index = pkt01_index;
865
866                 /* Pipeline stage 0 */
867                 lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask,
868                         pkt00_index, pkt01_index);
869
870                 /* Pipeline stage 1 */
871                 lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
872
873                 /* Pipeline stage 2 */
874                 lookup2_stage2(t, g, pkt20_index, pkt21_index,
875                         pkts_mask_match_many);
876
877                 /* Pipeline stage 3 */
878                 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index,
879                         pkts_mask_out, entries);
880         }
881
882         /* Pipeline feed */
883         pkt30_index = pkt20_index;
884         pkt31_index = pkt21_index;
885         pkt20_index = pkt10_index;
886         pkt21_index = pkt11_index;
887         pkt10_index = pkt00_index;
888         pkt11_index = pkt01_index;
889
890         /* Pipeline stage 1 */
891         lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
892
893         /* Pipeline stage 2 */
894         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
895
896         /* Pipeline stage 3 */
897         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
898                 entries);
899
900         /* Pipeline feed */
901         pkt30_index = pkt20_index;
902         pkt31_index = pkt21_index;
903         pkt20_index = pkt10_index;
904         pkt21_index = pkt11_index;
905
906         /* Pipeline stage 2 */
907         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
908
909         /* Pipeline stage 3 */
910         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
911                 entries);
912
913         /* Pipeline feed */
914         pkt30_index = pkt20_index;
915         pkt31_index = pkt21_index;
916
917         /* Pipeline stage 3 */
918         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
919                 entries);
920
921         /* Slow path */
922         pkts_mask_match_many &= ~pkts_mask_out;
923         if (pkts_mask_match_many) {
924                 uint64_t pkts_mask_out_slow = 0;
925
926                 status = rte_table_hash_lru_lookup_unoptimized(table, pkts,
927                         pkts_mask_match_many, &pkts_mask_out_slow, entries);
928                 pkts_mask_out |= pkts_mask_out_slow;
929         }
930
931         *lookup_hit_mask = pkts_mask_out;
932         RTE_TABLE_HASH_LRU_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - __builtin_popcountll(pkts_mask_out));
933         return status;
934 }
935
936 static int
937 rte_table_hash_lru_stats_read(void *table, struct rte_table_stats *stats, int clear)
938 {
939         struct rte_table_hash *t = table;
940
941         if (stats != NULL)
942                 memcpy(stats, &t->stats, sizeof(t->stats));
943
944         if (clear)
945                 memset(&t->stats, 0, sizeof(t->stats));
946
947         return 0;
948 }
949
950 struct rte_table_ops rte_table_hash_lru_ops = {
951         .f_create = rte_table_hash_lru_create,
952         .f_free = rte_table_hash_lru_free,
953         .f_add = rte_table_hash_lru_entry_add,
954         .f_delete = rte_table_hash_lru_entry_delete,
955         .f_add_bulk = NULL,
956         .f_delete_bulk = NULL,
957         .f_lookup = rte_table_hash_lru_lookup,
958         .f_stats = rte_table_hash_lru_stats_read,
959 };