6790594944b548e998f8a9c64b9269525c1c26c7
[dpdk.git] / lib / librte_table / rte_table_hash_key32.c
1 /*-
2  *       BSD LICENSE
3  *
4  *       Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5  *       All rights reserved.
6  *
7  *       Redistribution and use in source and binary forms, with or without
8  *       modification, are permitted provided that the following conditions
9  *       are met:
10  *
11  *      * Redistributions of source code must retain the above copyright
12  *               notice, this list of conditions and the following disclaimer.
13  *      * Redistributions in binary form must reproduce the above copyright
14  *               notice, this list of conditions and the following disclaimer in
15  *               the documentation and/or other materials provided with the
16  *               distribution.
17  *      * Neither the name of Intel Corporation nor the names of its
18  *               contributors may be used to endorse or promote products derived
19  *               from this software without specific prior written permission.
20  *
21  *       THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *       "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *       LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *       A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *       OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *       SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *       LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *       DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *       THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *       (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *       OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 #include <string.h>
34 #include <stdio.h>
35
36 #include <rte_common.h>
37 #include <rte_mbuf.h>
38 #include <rte_memory.h>
39 #include <rte_malloc.h>
40 #include <rte_log.h>
41
42 #include "rte_table_hash.h"
43 #include "rte_lru.h"
44
45 #define RTE_TABLE_HASH_KEY_SIZE                                         32
46
47 #define RTE_BUCKET_ENTRY_VALID                                          0x1LLU
48
49 struct rte_bucket_4_32 {
50         /* Cache line 0 */
51         uint64_t signature[4 + 1];
52         uint64_t lru_list;
53         struct rte_bucket_4_32 *next;
54         uint64_t next_valid;
55
56         /* Cache lines 1 and 2 */
57         uint64_t key[4][4];
58
59         /* Cache line 3 */
60         uint8_t data[0];
61 };
62
63 struct rte_table_hash {
64         /* Input parameters */
65         uint32_t n_buckets;
66         uint32_t n_entries_per_bucket;
67         uint32_t key_size;
68         uint32_t entry_size;
69         uint32_t bucket_size;
70         uint32_t signature_offset;
71         uint32_t key_offset;
72         rte_table_hash_op_hash f_hash;
73         uint64_t seed;
74
75         /* Extendible buckets */
76         uint32_t n_buckets_ext;
77         uint32_t stack_pos;
78         uint32_t *stack;
79
80         /* Lookup table */
81         uint8_t memory[0] __rte_cache_aligned;
82 };
83
84 static int
85 check_params_create_lru(struct rte_table_hash_key32_lru_params *params) {
86         /* n_entries */
87         if (params->n_entries == 0) {
88                 RTE_LOG(ERR, TABLE, "%s: n_entries is zero\n", __func__);
89                 return -EINVAL;
90         }
91
92         /* signature offset */
93         if ((params->signature_offset & 0x3) != 0) {
94                 RTE_LOG(ERR, TABLE, "%s: invalid signature offset\n", __func__);
95                 return -EINVAL;
96         }
97
98         /* key offset */
99         if ((params->key_offset & 0x7) != 0) {
100                 RTE_LOG(ERR, TABLE, "%s: invalid key offset\n", __func__);
101                 return -EINVAL;
102         }
103
104         /* f_hash */
105         if (params->f_hash == NULL) {
106                 RTE_LOG(ERR, TABLE, "%s: f_hash function pointer is NULL\n",
107                         __func__);
108                 return -EINVAL;
109         }
110
111         return 0;
112 }
113
114 static void *
115 rte_table_hash_create_key32_lru(void *params,
116                 int socket_id,
117                 uint32_t entry_size)
118 {
119         struct rte_table_hash_key32_lru_params *p =
120                 (struct rte_table_hash_key32_lru_params *) params;
121         struct rte_table_hash *f;
122         uint32_t n_buckets, n_entries_per_bucket, key_size, bucket_size_cl;
123         uint32_t total_size, i;
124
125         /* Check input parameters */
126         if ((check_params_create_lru(p) != 0) ||
127                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
128                 ((sizeof(struct rte_bucket_4_32) % RTE_CACHE_LINE_SIZE) != 0)) {
129                 return NULL;
130         }
131         n_entries_per_bucket = 4;
132         key_size = 32;
133
134         /* Memory allocation */
135         n_buckets = rte_align32pow2((p->n_entries + n_entries_per_bucket - 1) /
136                 n_entries_per_bucket);
137         bucket_size_cl = (sizeof(struct rte_bucket_4_32) + n_entries_per_bucket
138                 * entry_size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
139         total_size = sizeof(struct rte_table_hash) + n_buckets *
140                 bucket_size_cl * RTE_CACHE_LINE_SIZE;
141
142         f = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
143         if (f == NULL) {
144                 RTE_LOG(ERR, TABLE,
145                         "%s: Cannot allocate %u bytes for hash table\n",
146                         __func__, total_size);
147                 return NULL;
148         }
149         RTE_LOG(INFO, TABLE,
150                 "%s: Hash table memory footprint is %u bytes\n", __func__,
151                 total_size);
152
153         /* Memory initialization */
154         f->n_buckets = n_buckets;
155         f->n_entries_per_bucket = n_entries_per_bucket;
156         f->key_size = key_size;
157         f->entry_size = entry_size;
158         f->bucket_size = bucket_size_cl * RTE_CACHE_LINE_SIZE;
159         f->signature_offset = p->signature_offset;
160         f->key_offset = p->key_offset;
161         f->f_hash = p->f_hash;
162         f->seed = p->seed;
163
164         for (i = 0; i < n_buckets; i++) {
165                 struct rte_bucket_4_32 *bucket;
166
167                 bucket = (struct rte_bucket_4_32 *) &f->memory[i *
168                         f->bucket_size];
169                 bucket->lru_list = 0x0000000100020003LLU;
170         }
171
172         return f;
173 }
174
175 static int
176 rte_table_hash_free_key32_lru(void *table)
177 {
178         struct rte_table_hash *f = (struct rte_table_hash *) table;
179
180         /* Check input parameters */
181         if (f == NULL) {
182                 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
183                 return -EINVAL;
184         }
185
186         rte_free(f);
187         return 0;
188 }
189
190 static int
191 rte_table_hash_entry_add_key32_lru(
192         void *table,
193         void *key,
194         void *entry,
195         int *key_found,
196         void **entry_ptr)
197 {
198         struct rte_table_hash *f = (struct rte_table_hash *) table;
199         struct rte_bucket_4_32 *bucket;
200         uint64_t signature, pos;
201         uint32_t bucket_index, i;
202
203         signature = f->f_hash(key, f->key_size, f->seed);
204         bucket_index = signature & (f->n_buckets - 1);
205         bucket = (struct rte_bucket_4_32 *)
206                 &f->memory[bucket_index * f->bucket_size];
207         signature |= RTE_BUCKET_ENTRY_VALID;
208
209         /* Key is present in the bucket */
210         for (i = 0; i < 4; i++) {
211                 uint64_t bucket_signature = bucket->signature[i];
212                 uint8_t *bucket_key = (uint8_t *) bucket->key[i];
213
214                 if ((bucket_signature == signature) &&
215                         (memcmp(key, bucket_key, f->key_size) == 0)) {
216                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
217
218                         memcpy(bucket_data, entry, f->entry_size);
219                         lru_update(bucket, i);
220                         *key_found = 1;
221                         *entry_ptr = (void *) bucket_data;
222                         return 0;
223                 }
224         }
225
226         /* Key is not present in the bucket */
227         for (i = 0; i < 4; i++) {
228                 uint64_t bucket_signature = bucket->signature[i];
229                 uint8_t *bucket_key = (uint8_t *) bucket->key[i];
230
231                 if (bucket_signature == 0) {
232                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
233
234                         bucket->signature[i] = signature;
235                         memcpy(bucket_key, key, f->key_size);
236                         memcpy(bucket_data, entry, f->entry_size);
237                         lru_update(bucket, i);
238                         *key_found = 0;
239                         *entry_ptr = (void *) bucket_data;
240
241                         return 0;
242                 }
243         }
244
245         /* Bucket full: replace LRU entry */
246         pos = lru_pos(bucket);
247         bucket->signature[pos] = signature;
248         memcpy(bucket->key[pos], key, f->key_size);
249         memcpy(&bucket->data[pos * f->entry_size], entry, f->entry_size);
250         lru_update(bucket, pos);
251         *key_found      = 0;
252         *entry_ptr = (void *) &bucket->data[pos * f->entry_size];
253
254         return 0;
255 }
256
257 static int
258 rte_table_hash_entry_delete_key32_lru(
259         void *table,
260         void *key,
261         int *key_found,
262         void *entry)
263 {
264         struct rte_table_hash *f = (struct rte_table_hash *) table;
265         struct rte_bucket_4_32 *bucket;
266         uint64_t signature;
267         uint32_t bucket_index, i;
268
269         signature = f->f_hash(key, f->key_size, f->seed);
270         bucket_index = signature & (f->n_buckets - 1);
271         bucket = (struct rte_bucket_4_32 *)
272                 &f->memory[bucket_index * f->bucket_size];
273         signature |= RTE_BUCKET_ENTRY_VALID;
274
275         /* Key is present in the bucket */
276         for (i = 0; i < 4; i++) {
277                 uint64_t bucket_signature = bucket->signature[i];
278                 uint8_t *bucket_key = (uint8_t *) bucket->key[i];
279
280                 if ((bucket_signature == signature) &&
281                         (memcmp(key, bucket_key, f->key_size) == 0)) {
282                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
283
284                         bucket->signature[i] = 0;
285                         *key_found = 1;
286                         if (entry)
287                                 memcpy(entry, bucket_data, f->entry_size);
288
289                         return 0;
290                 }
291         }
292
293         /* Key is not present in the bucket */
294         *key_found = 0;
295         return 0;
296 }
297
298 static int
299 check_params_create_ext(struct rte_table_hash_key32_ext_params *params) {
300         /* n_entries */
301         if (params->n_entries == 0) {
302                 RTE_LOG(ERR, TABLE, "%s: n_entries is zero\n", __func__);
303                 return -EINVAL;
304         }
305
306         /* n_entries_ext */
307         if (params->n_entries_ext == 0) {
308                 RTE_LOG(ERR, TABLE, "%s: n_entries_ext is zero\n", __func__);
309                 return -EINVAL;
310         }
311
312         /* signature offset */
313         if ((params->signature_offset & 0x3) != 0) {
314                 RTE_LOG(ERR, TABLE, "%s: invalid signature offset\n", __func__);
315                 return -EINVAL;
316         }
317
318         /* key offset */
319         if ((params->key_offset & 0x7) != 0) {
320                 RTE_LOG(ERR, TABLE, "%s: invalid key offset\n", __func__);
321                 return -EINVAL;
322         }
323
324         /* f_hash */
325         if (params->f_hash == NULL) {
326                 RTE_LOG(ERR, TABLE, "%s: f_hash function pointer is NULL\n",
327                         __func__);
328                 return -EINVAL;
329         }
330
331         return 0;
332 }
333
334 static void *
335 rte_table_hash_create_key32_ext(void *params,
336         int socket_id,
337         uint32_t entry_size)
338 {
339         struct rte_table_hash_key32_ext_params *p =
340                         (struct rte_table_hash_key32_ext_params *) params;
341         struct rte_table_hash *f;
342         uint32_t n_buckets, n_buckets_ext, n_entries_per_bucket;
343         uint32_t key_size, bucket_size_cl, stack_size_cl, total_size, i;
344
345         /* Check input parameters */
346         if ((check_params_create_ext(p) != 0) ||
347                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
348                 ((sizeof(struct rte_bucket_4_32) % RTE_CACHE_LINE_SIZE) != 0))
349                 return NULL;
350
351         n_entries_per_bucket = 4;
352         key_size = 32;
353
354         /* Memory allocation */
355         n_buckets = rte_align32pow2((p->n_entries + n_entries_per_bucket - 1) /
356                 n_entries_per_bucket);
357         n_buckets_ext = (p->n_entries_ext + n_entries_per_bucket - 1) /
358                 n_entries_per_bucket;
359         bucket_size_cl = (sizeof(struct rte_bucket_4_32) + n_entries_per_bucket
360                 * entry_size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
361         stack_size_cl = (n_buckets_ext * sizeof(uint32_t) + RTE_CACHE_LINE_SIZE - 1)
362                 / RTE_CACHE_LINE_SIZE;
363         total_size = sizeof(struct rte_table_hash) +
364                 ((n_buckets + n_buckets_ext) * bucket_size_cl + stack_size_cl) *
365                 RTE_CACHE_LINE_SIZE;
366
367         f = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
368         if (f == NULL) {
369                 RTE_LOG(ERR, TABLE,
370                         "%s: Cannot allocate %u bytes for hash table\n",
371                         __func__, total_size);
372                 return NULL;
373         }
374         RTE_LOG(INFO, TABLE,
375                 "%s: Hash table memory footprint is %u bytes\n", __func__,
376                 total_size);
377
378         /* Memory initialization */
379         f->n_buckets = n_buckets;
380         f->n_entries_per_bucket = n_entries_per_bucket;
381         f->key_size = key_size;
382         f->entry_size = entry_size;
383         f->bucket_size = bucket_size_cl * RTE_CACHE_LINE_SIZE;
384         f->signature_offset = p->signature_offset;
385         f->key_offset = p->key_offset;
386         f->f_hash = p->f_hash;
387         f->seed = p->seed;
388
389         f->n_buckets_ext = n_buckets_ext;
390         f->stack_pos = n_buckets_ext;
391         f->stack = (uint32_t *)
392                 &f->memory[(n_buckets + n_buckets_ext) * f->bucket_size];
393
394         for (i = 0; i < n_buckets_ext; i++)
395                 f->stack[i] = i;
396
397         return f;
398 }
399
400 static int
401 rte_table_hash_free_key32_ext(void *table)
402 {
403         struct rte_table_hash *f = (struct rte_table_hash *) table;
404
405         /* Check input parameters */
406         if (f == NULL) {
407                 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
408                 return -EINVAL;
409         }
410
411         rte_free(f);
412         return 0;
413 }
414
415 static int
416 rte_table_hash_entry_add_key32_ext(
417         void *table,
418         void *key,
419         void *entry,
420         int *key_found,
421         void **entry_ptr)
422 {
423         struct rte_table_hash *f = (struct rte_table_hash *) table;
424         struct rte_bucket_4_32 *bucket0, *bucket, *bucket_prev;
425         uint64_t signature;
426         uint32_t bucket_index, i;
427
428         signature = f->f_hash(key, f->key_size, f->seed);
429         bucket_index = signature & (f->n_buckets - 1);
430         bucket0 = (struct rte_bucket_4_32 *)
431                         &f->memory[bucket_index * f->bucket_size];
432         signature |= RTE_BUCKET_ENTRY_VALID;
433
434         /* Key is present in the bucket */
435         for (bucket = bucket0; bucket != NULL; bucket = bucket->next) {
436                 for (i = 0; i < 4; i++) {
437                         uint64_t bucket_signature = bucket->signature[i];
438                         uint8_t *bucket_key = (uint8_t *) bucket->key[i];
439
440                         if ((bucket_signature == signature) &&
441                                 (memcmp(key, bucket_key, f->key_size) == 0)) {
442                                 uint8_t *bucket_data = &bucket->data[i *
443                                         f->entry_size];
444
445                                 memcpy(bucket_data, entry, f->entry_size);
446                                 *key_found = 1;
447                                 *entry_ptr = (void *) bucket_data;
448
449                                 return 0;
450                         }
451                 }
452         }
453
454         /* Key is not present in the bucket */
455         for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
456                 bucket_prev = bucket, bucket = bucket->next)
457                 for (i = 0; i < 4; i++) {
458                         uint64_t bucket_signature = bucket->signature[i];
459                         uint8_t *bucket_key = (uint8_t *) bucket->key[i];
460
461                         if (bucket_signature == 0) {
462                                 uint8_t *bucket_data = &bucket->data[i *
463                                         f->entry_size];
464
465                                 bucket->signature[i] = signature;
466                                 memcpy(bucket_key, key, f->key_size);
467                                 memcpy(bucket_data, entry, f->entry_size);
468                                 *key_found = 0;
469                                 *entry_ptr = (void *) bucket_data;
470
471                                 return 0;
472                         }
473                 }
474
475         /* Bucket full: extend bucket */
476         if (f->stack_pos > 0) {
477                 bucket_index = f->stack[--f->stack_pos];
478
479                 bucket = (struct rte_bucket_4_32 *)
480                         &f->memory[(f->n_buckets + bucket_index) *
481                         f->bucket_size];
482                 bucket_prev->next = bucket;
483                 bucket_prev->next_valid = 1;
484
485                 bucket->signature[0] = signature;
486                 memcpy(bucket->key[0], key, f->key_size);
487                 memcpy(&bucket->data[0], entry, f->entry_size);
488                 *key_found = 0;
489                 *entry_ptr = (void *) &bucket->data[0];
490                 return 0;
491         }
492
493         return -ENOSPC;
494 }
495
496 static int
497 rte_table_hash_entry_delete_key32_ext(
498         void *table,
499         void *key,
500         int *key_found,
501         void *entry)
502 {
503         struct rte_table_hash *f = (struct rte_table_hash *) table;
504         struct rte_bucket_4_32 *bucket0, *bucket, *bucket_prev;
505         uint64_t signature;
506         uint32_t bucket_index, i;
507
508         signature = f->f_hash(key, f->key_size, f->seed);
509         bucket_index = signature & (f->n_buckets - 1);
510         bucket0 = (struct rte_bucket_4_32 *)
511                 &f->memory[bucket_index * f->bucket_size];
512         signature |= RTE_BUCKET_ENTRY_VALID;
513
514         /* Key is present in the bucket */
515         for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
516                 bucket_prev = bucket, bucket = bucket->next)
517                 for (i = 0; i < 4; i++) {
518                         uint64_t bucket_signature = bucket->signature[i];
519                         uint8_t *bucket_key = (uint8_t *) bucket->key[i];
520
521                         if ((bucket_signature == signature) &&
522                                 (memcmp(key, bucket_key, f->key_size) == 0)) {
523                                 uint8_t *bucket_data = &bucket->data[i *
524                                         f->entry_size];
525
526                                 bucket->signature[i] = 0;
527                                 *key_found = 1;
528                                 if (entry)
529                                         memcpy(entry, bucket_data,
530                                                 f->entry_size);
531
532                                 if ((bucket->signature[0] == 0) &&
533                                                 (bucket->signature[1] == 0) &&
534                                                 (bucket->signature[2] == 0) &&
535                                                 (bucket->signature[3] == 0) &&
536                                                 (bucket_prev != NULL)) {
537                                         bucket_prev->next = bucket->next;
538                                         bucket_prev->next_valid =
539                                                 bucket->next_valid;
540
541                                         memset(bucket, 0,
542                                                 sizeof(struct rte_bucket_4_32));
543                                         bucket_index = (((uint8_t *)bucket -
544                                                 (uint8_t *)f->memory)/f->bucket_size) - f->n_buckets;
545                                         f->stack[f->stack_pos++] = bucket_index;
546                                 }
547
548                                 return 0;
549                         }
550                 }
551
552         /* Key is not present in the bucket */
553         *key_found = 0;
554         return 0;
555 }
556
557 #define lookup_key32_cmp(key_in, bucket, pos)                   \
558 {                                                               \
559         uint64_t xor[4][4], or[4], signature[4];                \
560                                                                 \
561         signature[0] = ((~bucket->signature[0]) & 1);           \
562         signature[1] = ((~bucket->signature[1]) & 1);           \
563         signature[2] = ((~bucket->signature[2]) & 1);           \
564         signature[3] = ((~bucket->signature[3]) & 1);           \
565                                                                 \
566         xor[0][0] = key_in[0] ^  bucket->key[0][0];             \
567         xor[0][1] = key_in[1] ^  bucket->key[0][1];             \
568         xor[0][2] = key_in[2] ^  bucket->key[0][2];             \
569         xor[0][3] = key_in[3] ^  bucket->key[0][3];             \
570                                                                 \
571         xor[1][0] = key_in[0] ^  bucket->key[1][0];             \
572         xor[1][1] = key_in[1] ^  bucket->key[1][1];             \
573         xor[1][2] = key_in[2] ^  bucket->key[1][2];             \
574         xor[1][3] = key_in[3] ^  bucket->key[1][3];             \
575                                                                 \
576         xor[2][0] = key_in[0] ^  bucket->key[2][0];             \
577         xor[2][1] = key_in[1] ^  bucket->key[2][1];             \
578         xor[2][2] = key_in[2] ^  bucket->key[2][2];             \
579         xor[2][3] = key_in[3] ^  bucket->key[2][3];             \
580                                                                 \
581         xor[3][0] = key_in[0] ^  bucket->key[3][0];             \
582         xor[3][1] = key_in[1] ^  bucket->key[3][1];             \
583         xor[3][2] = key_in[2] ^  bucket->key[3][2];             \
584         xor[3][3] = key_in[3] ^  bucket->key[3][3];             \
585                                                                 \
586         or[0] = xor[0][0] | xor[0][1] | xor[0][2] | xor[0][3] | signature[0];\
587         or[1] = xor[1][0] | xor[1][1] | xor[1][2] | xor[1][3] | signature[1];\
588         or[2] = xor[2][0] | xor[2][1] | xor[2][2] | xor[2][3] | signature[2];\
589         or[3] = xor[3][0] | xor[3][1] | xor[3][2] | xor[3][3] | signature[3];\
590                                                                 \
591         pos = 4;                                                \
592         if (or[0] == 0)                                         \
593                 pos = 0;                                        \
594         if (or[1] == 0)                                         \
595                 pos = 1;                                        \
596         if (or[2] == 0)                                         \
597                 pos = 2;                                        \
598         if (or[3] == 0)                                         \
599                 pos = 3;                                        \
600 }
601
602 #define lookup1_stage0(pkt0_index, mbuf0, pkts, pkts_mask)      \
603 {                                                               \
604         uint64_t pkt_mask;                                      \
605                                                                 \
606         pkt0_index = __builtin_ctzll(pkts_mask);                \
607         pkt_mask = 1LLU << pkt0_index;                          \
608         pkts_mask &= ~pkt_mask;                                 \
609                                                                 \
610         mbuf0 = pkts[pkt0_index];                               \
611         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf0, 0));   \
612 }
613
614 #define lookup1_stage1(mbuf1, bucket1, f)                       \
615 {                                                               \
616         uint64_t signature;                                     \
617         uint32_t bucket_index;                                  \
618                                                                 \
619         signature = RTE_MBUF_METADATA_UINT32(mbuf1, f->signature_offset);\
620         bucket_index = signature & (f->n_buckets - 1);          \
621         bucket1 = (struct rte_bucket_4_32 *)                    \
622                 &f->memory[bucket_index * f->bucket_size];      \
623         rte_prefetch0(bucket1);                                 \
624         rte_prefetch0((void *)(((uintptr_t) bucket1) + RTE_CACHE_LINE_SIZE));\
625         rte_prefetch0((void *)(((uintptr_t) bucket1) + 2 * RTE_CACHE_LINE_SIZE));\
626 }
627
628 #define lookup1_stage2_lru(pkt2_index, mbuf2, bucket2,          \
629         pkts_mask_out, entries, f)                              \
630 {                                                               \
631         void *a;                                                \
632         uint64_t pkt_mask;                                      \
633         uint64_t *key;                                          \
634         uint32_t pos;                                           \
635                                                                 \
636         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
637                                                                 \
638         lookup_key32_cmp(key, bucket2, pos);                    \
639                                                                 \
640         pkt_mask = (bucket2->signature[pos] & 1LLU) << pkt2_index;\
641         pkts_mask_out |= pkt_mask;                              \
642                                                                 \
643         a = (void *) &bucket2->data[pos * f->entry_size];       \
644         rte_prefetch0(a);                                       \
645         entries[pkt2_index] = a;                                \
646         lru_update(bucket2, pos);                               \
647 }
648
649 #define lookup1_stage2_ext(pkt2_index, mbuf2, bucket2, pkts_mask_out,\
650         entries, buckets_mask, buckets, keys, f)                \
651 {                                                               \
652         struct rte_bucket_4_32 *bucket_next;                    \
653         void *a;                                                \
654         uint64_t pkt_mask, bucket_mask;                         \
655         uint64_t *key;                                          \
656         uint32_t pos;                                           \
657                                                                 \
658         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
659                                                                 \
660         lookup_key32_cmp(key, bucket2, pos);                    \
661                                                                 \
662         pkt_mask = (bucket2->signature[pos] & 1LLU) << pkt2_index;\
663         pkts_mask_out |= pkt_mask;                              \
664                                                                 \
665         a = (void *) &bucket2->data[pos * f->entry_size];       \
666         rte_prefetch0(a);                                       \
667         entries[pkt2_index] = a;                                \
668                                                                 \
669         bucket_mask = (~pkt_mask) & (bucket2->next_valid << pkt2_index);\
670         buckets_mask |= bucket_mask;                            \
671         bucket_next = bucket2->next;                            \
672         buckets[pkt2_index] = bucket_next;                      \
673         keys[pkt2_index] = key;                                 \
674 }
675
676 #define lookup_grinder(pkt_index, buckets, keys, pkts_mask_out, \
677         entries, buckets_mask, f)                               \
678 {                                                               \
679         struct rte_bucket_4_32 *bucket, *bucket_next;           \
680         void *a;                                                \
681         uint64_t pkt_mask, bucket_mask;                         \
682         uint64_t *key;                                          \
683         uint32_t pos;                                           \
684                                                                 \
685         bucket = buckets[pkt_index];                            \
686         key = keys[pkt_index];                                  \
687                                                                 \
688         lookup_key32_cmp(key, bucket, pos);                     \
689                                                                 \
690         pkt_mask = (bucket->signature[pos] & 1LLU) << pkt_index;\
691         pkts_mask_out |= pkt_mask;                              \
692                                                                 \
693         a = (void *) &bucket->data[pos * f->entry_size];        \
694         rte_prefetch0(a);                                       \
695         entries[pkt_index] = a;                                 \
696                                                                 \
697         bucket_mask = (~pkt_mask) & (bucket->next_valid << pkt_index);\
698         buckets_mask |= bucket_mask;                            \
699         bucket_next = bucket->next;                             \
700         rte_prefetch0(bucket_next);                             \
701         rte_prefetch0((void *)(((uintptr_t) bucket_next) + RTE_CACHE_LINE_SIZE));\
702         rte_prefetch0((void *)(((uintptr_t) bucket_next) +      \
703                 2 * RTE_CACHE_LINE_SIZE));                              \
704         buckets[pkt_index] = bucket_next;                       \
705         keys[pkt_index] = key;                                  \
706 }
707
708 #define lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01,\
709         pkts, pkts_mask)                                        \
710 {                                                               \
711         uint64_t pkt00_mask, pkt01_mask;                        \
712                                                                 \
713         pkt00_index = __builtin_ctzll(pkts_mask);               \
714         pkt00_mask = 1LLU << pkt00_index;                       \
715         pkts_mask &= ~pkt00_mask;                               \
716                                                                 \
717         mbuf00 = pkts[pkt00_index];                             \
718         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));  \
719                                                                 \
720         pkt01_index = __builtin_ctzll(pkts_mask);               \
721         pkt01_mask = 1LLU << pkt01_index;                       \
722         pkts_mask &= ~pkt01_mask;                               \
723                                                                 \
724         mbuf01 = pkts[pkt01_index];                             \
725         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));  \
726 }
727
728 #define lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,\
729         mbuf00, mbuf01, pkts, pkts_mask)                        \
730 {                                                               \
731         uint64_t pkt00_mask, pkt01_mask;                        \
732                                                                 \
733         pkt00_index = __builtin_ctzll(pkts_mask);               \
734         pkt00_mask = 1LLU << pkt00_index;                       \
735         pkts_mask &= ~pkt00_mask;                               \
736                                                                 \
737         mbuf00 = pkts[pkt00_index];                             \
738         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));  \
739                                                                 \
740         pkt01_index = __builtin_ctzll(pkts_mask);               \
741         if (pkts_mask == 0)                                     \
742                 pkt01_index = pkt00_index;                      \
743                                                                 \
744         pkt01_mask = 1LLU << pkt01_index;                       \
745         pkts_mask &= ~pkt01_mask;                               \
746                                                                 \
747         mbuf01 = pkts[pkt01_index];                             \
748         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));  \
749 }
750
751 #define lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f)   \
752 {                                                               \
753         uint64_t signature10, signature11;                      \
754         uint32_t bucket10_index, bucket11_index;                \
755                                                                 \
756         signature10 = RTE_MBUF_METADATA_UINT32(mbuf10, f->signature_offset);\
757         bucket10_index = signature10 & (f->n_buckets - 1);      \
758         bucket10 = (struct rte_bucket_4_32 *)                   \
759                 &f->memory[bucket10_index * f->bucket_size];    \
760         rte_prefetch0(bucket10);                                \
761         rte_prefetch0((void *)(((uintptr_t) bucket10) + RTE_CACHE_LINE_SIZE));\
762         rte_prefetch0((void *)(((uintptr_t) bucket10) + 2 * RTE_CACHE_LINE_SIZE));\
763                                                                 \
764         signature11 = RTE_MBUF_METADATA_UINT32(mbuf11, f->signature_offset);\
765         bucket11_index = signature11 & (f->n_buckets - 1);      \
766         bucket11 = (struct rte_bucket_4_32 *)                   \
767                 &f->memory[bucket11_index * f->bucket_size];    \
768         rte_prefetch0(bucket11);                                \
769         rte_prefetch0((void *)(((uintptr_t) bucket11) + RTE_CACHE_LINE_SIZE));\
770         rte_prefetch0((void *)(((uintptr_t) bucket11) + 2 * RTE_CACHE_LINE_SIZE));\
771 }
772
773 #define lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,\
774         bucket20, bucket21, pkts_mask_out, entries, f)          \
775 {                                                               \
776         void *a20, *a21;                                        \
777         uint64_t pkt20_mask, pkt21_mask;                        \
778         uint64_t *key20, *key21;                                \
779         uint32_t pos20, pos21;                                  \
780                                                                 \
781         key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
782         key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
783                                                                 \
784         lookup_key32_cmp(key20, bucket20, pos20);               \
785         lookup_key32_cmp(key21, bucket21, pos21);               \
786                                                                 \
787         pkt20_mask = (bucket20->signature[pos20] & 1LLU) << pkt20_index;\
788         pkt21_mask = (bucket21->signature[pos21] & 1LLU) << pkt21_index;\
789         pkts_mask_out |= pkt20_mask | pkt21_mask;               \
790                                                                 \
791         a20 = (void *) &bucket20->data[pos20 * f->entry_size];  \
792         a21 = (void *) &bucket21->data[pos21 * f->entry_size];  \
793         rte_prefetch0(a20);                                     \
794         rte_prefetch0(a21);                                     \
795         entries[pkt20_index] = a20;                             \
796         entries[pkt21_index] = a21;                             \
797         lru_update(bucket20, pos20);                            \
798         lru_update(bucket21, pos21);                            \
799 }
800
801 #define lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21, bucket20, \
802         bucket21, pkts_mask_out, entries, buckets_mask, buckets, keys, f)\
803 {                                                               \
804         struct rte_bucket_4_32 *bucket20_next, *bucket21_next;  \
805         void *a20, *a21;                                        \
806         uint64_t pkt20_mask, pkt21_mask, bucket20_mask, bucket21_mask;\
807         uint64_t *key20, *key21;                                \
808         uint32_t pos20, pos21;                                  \
809                                                                 \
810         key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
811         key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
812                                                                 \
813         lookup_key32_cmp(key20, bucket20, pos20);               \
814         lookup_key32_cmp(key21, bucket21, pos21);               \
815                                                                 \
816         pkt20_mask = (bucket20->signature[pos20] & 1LLU) << pkt20_index;\
817         pkt21_mask = (bucket21->signature[pos21] & 1LLU) << pkt21_index;\
818         pkts_mask_out |= pkt20_mask | pkt21_mask;               \
819                                                                 \
820         a20 = (void *) &bucket20->data[pos20 * f->entry_size];  \
821         a21 = (void *) &bucket21->data[pos21 * f->entry_size];  \
822         rte_prefetch0(a20);                                     \
823         rte_prefetch0(a21);                                     \
824         entries[pkt20_index] = a20;                             \
825         entries[pkt21_index] = a21;                             \
826                                                                 \
827         bucket20_mask = (~pkt20_mask) & (bucket20->next_valid << pkt20_index);\
828         bucket21_mask = (~pkt21_mask) & (bucket21->next_valid << pkt21_index);\
829         buckets_mask |= bucket20_mask | bucket21_mask;          \
830         bucket20_next = bucket20->next;                         \
831         bucket21_next = bucket21->next;                         \
832         buckets[pkt20_index] = bucket20_next;                   \
833         buckets[pkt21_index] = bucket21_next;                   \
834         keys[pkt20_index] = key20;                              \
835         keys[pkt21_index] = key21;                              \
836 }
837
838 static int
839 rte_table_hash_lookup_key32_lru(
840         void *table,
841         struct rte_mbuf **pkts,
842         uint64_t pkts_mask,
843         uint64_t *lookup_hit_mask,
844         void **entries)
845 {
846         struct rte_table_hash *f = (struct rte_table_hash *) table;
847         struct rte_bucket_4_32 *bucket10, *bucket11, *bucket20, *bucket21;
848         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
849         uint32_t pkt00_index, pkt01_index, pkt10_index;
850         uint32_t pkt11_index, pkt20_index, pkt21_index;
851         uint64_t pkts_mask_out = 0;
852
853         /* Cannot run the pipeline with less than 5 packets */
854         if (__builtin_popcountll(pkts_mask) < 5) {
855                 for ( ; pkts_mask; ) {
856                         struct rte_bucket_4_32 *bucket;
857                         struct rte_mbuf *mbuf;
858                         uint32_t pkt_index;
859
860                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
861                         lookup1_stage1(mbuf, bucket, f);
862                         lookup1_stage2_lru(pkt_index, mbuf, bucket,
863                                         pkts_mask_out, entries, f);
864                 }
865
866                 *lookup_hit_mask = pkts_mask_out;
867                 return 0;
868         }
869
870         /*
871          * Pipeline fill
872          *
873          */
874         /* Pipeline stage 0 */
875         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
876                 pkts_mask);
877
878         /* Pipeline feed */
879         mbuf10 = mbuf00;
880         mbuf11 = mbuf01;
881         pkt10_index = pkt00_index;
882         pkt11_index = pkt01_index;
883
884         /* Pipeline stage 0 */
885         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
886                 pkts_mask);
887
888         /* Pipeline stage 1 */
889         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
890
891         /*
892          * Pipeline run
893          *
894          */
895         for ( ; pkts_mask; ) {
896                 /* Pipeline feed */
897                 bucket20 = bucket10;
898                 bucket21 = bucket11;
899                 mbuf20 = mbuf10;
900                 mbuf21 = mbuf11;
901                 mbuf10 = mbuf00;
902                 mbuf11 = mbuf01;
903                 pkt20_index = pkt10_index;
904                 pkt21_index = pkt11_index;
905                 pkt10_index = pkt00_index;
906                 pkt11_index = pkt01_index;
907
908                 /* Pipeline stage 0 */
909                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
910                         mbuf00, mbuf01, pkts, pkts_mask);
911
912                 /* Pipeline stage 1 */
913                 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
914
915                 /* Pipeline stage 2 */
916                 lookup2_stage2_lru(pkt20_index, pkt21_index,
917                         mbuf20, mbuf21, bucket20, bucket21, pkts_mask_out,
918                         entries, f);
919         }
920
921         /*
922          * Pipeline flush
923          *
924          */
925         /* Pipeline feed */
926         bucket20 = bucket10;
927         bucket21 = bucket11;
928         mbuf20 = mbuf10;
929         mbuf21 = mbuf11;
930         mbuf10 = mbuf00;
931         mbuf11 = mbuf01;
932         pkt20_index = pkt10_index;
933         pkt21_index = pkt11_index;
934         pkt10_index = pkt00_index;
935         pkt11_index = pkt01_index;
936
937         /* Pipeline stage 1 */
938         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
939
940         /* Pipeline stage 2 */
941         lookup2_stage2_lru(pkt20_index, pkt21_index,
942                 mbuf20, mbuf21, bucket20, bucket21, pkts_mask_out, entries, f);
943
944         /* Pipeline feed */
945         bucket20 = bucket10;
946         bucket21 = bucket11;
947         mbuf20 = mbuf10;
948         mbuf21 = mbuf11;
949         pkt20_index = pkt10_index;
950         pkt21_index = pkt11_index;
951
952         /* Pipeline stage 2 */
953         lookup2_stage2_lru(pkt20_index, pkt21_index,
954                 mbuf20, mbuf21, bucket20, bucket21, pkts_mask_out, entries, f);
955
956         *lookup_hit_mask = pkts_mask_out;
957         return 0;
958 } /* rte_table_hash_lookup_key32_lru() */
959
960 static int
961 rte_table_hash_lookup_key32_ext(
962         void *table,
963         struct rte_mbuf **pkts,
964         uint64_t pkts_mask,
965         uint64_t *lookup_hit_mask,
966         void **entries)
967 {
968         struct rte_table_hash *f = (struct rte_table_hash *) table;
969         struct rte_bucket_4_32 *bucket10, *bucket11, *bucket20, *bucket21;
970         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
971         uint32_t pkt00_index, pkt01_index, pkt10_index;
972         uint32_t pkt11_index, pkt20_index, pkt21_index;
973         uint64_t pkts_mask_out = 0, buckets_mask = 0;
974         struct rte_bucket_4_32 *buckets[RTE_PORT_IN_BURST_SIZE_MAX];
975         uint64_t *keys[RTE_PORT_IN_BURST_SIZE_MAX];
976
977         /* Cannot run the pipeline with less than 5 packets */
978         if (__builtin_popcountll(pkts_mask) < 5) {
979                 for ( ; pkts_mask; ) {
980                         struct rte_bucket_4_32 *bucket;
981                         struct rte_mbuf *mbuf;
982                         uint32_t pkt_index;
983
984                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
985                         lookup1_stage1(mbuf, bucket, f);
986                         lookup1_stage2_ext(pkt_index, mbuf, bucket,
987                                 pkts_mask_out, entries, buckets_mask, buckets,
988                                 keys, f);
989                 }
990
991                 goto grind_next_buckets;
992         }
993
994         /*
995          * Pipeline fill
996          *
997          */
998         /* Pipeline stage 0 */
999         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1000                 pkts_mask);
1001
1002         /* Pipeline feed */
1003         mbuf10 = mbuf00;
1004         mbuf11 = mbuf01;
1005         pkt10_index = pkt00_index;
1006         pkt11_index = pkt01_index;
1007
1008         /* Pipeline stage 0 */
1009         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1010                 pkts_mask);
1011
1012         /* Pipeline stage 1 */
1013         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1014
1015         /*
1016          * Pipeline run
1017          *
1018          */
1019         for ( ; pkts_mask; ) {
1020                 /* Pipeline feed */
1021                 bucket20 = bucket10;
1022                 bucket21 = bucket11;
1023                 mbuf20 = mbuf10;
1024                 mbuf21 = mbuf11;
1025                 mbuf10 = mbuf00;
1026                 mbuf11 = mbuf01;
1027                 pkt20_index = pkt10_index;
1028                 pkt21_index = pkt11_index;
1029                 pkt10_index = pkt00_index;
1030                 pkt11_index = pkt01_index;
1031
1032                 /* Pipeline stage 0 */
1033                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1034                         mbuf00, mbuf01, pkts, pkts_mask);
1035
1036                 /* Pipeline stage 1 */
1037                 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1038
1039                 /* Pipeline stage 2 */
1040                 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1041                         bucket20, bucket21, pkts_mask_out, entries,
1042                         buckets_mask, buckets, keys, f);
1043         }
1044
1045         /*
1046          * Pipeline flush
1047          *
1048          */
1049         /* Pipeline feed */
1050         bucket20 = bucket10;
1051         bucket21 = bucket11;
1052         mbuf20 = mbuf10;
1053         mbuf21 = mbuf11;
1054         mbuf10 = mbuf00;
1055         mbuf11 = mbuf01;
1056         pkt20_index = pkt10_index;
1057         pkt21_index = pkt11_index;
1058         pkt10_index = pkt00_index;
1059         pkt11_index = pkt01_index;
1060
1061         /* Pipeline stage 1 */
1062         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1063
1064         /* Pipeline stage 2 */
1065         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1066                 bucket20, bucket21, pkts_mask_out, entries,
1067                 buckets_mask, buckets, keys, f);
1068
1069         /* Pipeline feed */
1070         bucket20 = bucket10;
1071         bucket21 = bucket11;
1072         mbuf20 = mbuf10;
1073         mbuf21 = mbuf11;
1074         pkt20_index = pkt10_index;
1075         pkt21_index = pkt11_index;
1076
1077         /* Pipeline stage 2 */
1078         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1079                 bucket20, bucket21, pkts_mask_out, entries,
1080                 buckets_mask, buckets, keys, f);
1081
1082 grind_next_buckets:
1083         /* Grind next buckets */
1084         for ( ; buckets_mask; ) {
1085                 uint64_t buckets_mask_next = 0;
1086
1087                 for ( ; buckets_mask; ) {
1088                         uint64_t pkt_mask;
1089                         uint32_t pkt_index;
1090
1091                         pkt_index = __builtin_ctzll(buckets_mask);
1092                         pkt_mask = 1LLU << pkt_index;
1093                         buckets_mask &= ~pkt_mask;
1094
1095                         lookup_grinder(pkt_index, buckets, keys, pkts_mask_out,
1096                                 entries, buckets_mask_next, f);
1097                 }
1098
1099                 buckets_mask = buckets_mask_next;
1100         }
1101
1102         *lookup_hit_mask = pkts_mask_out;
1103         return 0;
1104 } /* rte_table_hash_lookup_key32_ext() */
1105
1106 struct rte_table_ops rte_table_hash_key32_lru_ops = {
1107         .f_create = rte_table_hash_create_key32_lru,
1108         .f_free = rte_table_hash_free_key32_lru,
1109         .f_add = rte_table_hash_entry_add_key32_lru,
1110         .f_delete = rte_table_hash_entry_delete_key32_lru,
1111         .f_lookup = rte_table_hash_lookup_key32_lru,
1112 };
1113
1114 struct rte_table_ops rte_table_hash_key32_ext_ops = {
1115         .f_create = rte_table_hash_create_key32_ext,
1116         .f_free = rte_table_hash_free_key32_ext,
1117         .f_add = rte_table_hash_entry_add_key32_ext,
1118         .f_delete = rte_table_hash_entry_delete_key32_ext,
1119         .f_lookup = rte_table_hash_lookup_key32_ext,
1120 };