app/test-crypto-perf/cperf_test_latency.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2016-2017 Intel Corporation. All rights reserved.
   5  *
   6  *   Redistribution and use in source and binary forms, with or without
   7  *   modification, are permitted provided that the following conditions
   8  *   are met:
   9  *
  10  *     * Redistributions of source code must retain the above copyright
  11  *       notice, this list of conditions and the following disclaimer.
  12  *     * Redistributions in binary form must reproduce the above copyright
  13  *       notice, this list of conditions and the following disclaimer in
  14  *       the documentation and/or other materials provided with the
  15  *       distribution.
  16  *     * Neither the name of Intel Corporation nor the names of its
  17  *       contributors may be used to endorse or promote products derived
  18  *       from this software without specific prior written permission.
  19  *
  20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31  */
  32
  33 #include <rte_malloc.h>
  34 #include <rte_cycles.h>
  35 #include <rte_crypto.h>
  36 #include <rte_cryptodev.h>
  37
  38 #include "cperf_test_latency.h"
  39 #include "cperf_ops.h"
  40
  41
  42 struct cperf_latency_results {
  43
  44         uint64_t ops_failed;
  45
  46         uint64_t enqd_tot;
  47         uint64_t enqd_max;
  48         uint64_t enqd_min;
  49
  50         uint64_t deqd_tot;
  51         uint64_t deqd_max;
  52         uint64_t deqd_min;
  53
  54         uint64_t cycles_tot;
  55         uint64_t cycles_max;
  56         uint64_t cycles_min;
  57
  58         uint64_t burst_num;
  59         uint64_t num;
  60 };
  61
  62 struct cperf_op_result {
  63         uint64_t tsc_start;
  64         uint64_t tsc_end;
  65         enum rte_crypto_op_status status;
  66 };
  67
  68 struct cperf_latency_ctx {
  69         uint8_t dev_id;
  70         uint16_t qp_id;
  71         uint8_t lcore_id;
  72
  73         struct rte_mempool *pkt_mbuf_pool_in;
  74         struct rte_mempool *pkt_mbuf_pool_out;
  75         struct rte_mbuf **mbufs_in;
  76         struct rte_mbuf **mbufs_out;
  77
  78         struct rte_mempool *crypto_op_pool;
  79
  80         struct rte_cryptodev_sym_session *sess;
  81
  82         cperf_populate_ops_t populate_ops;
  83         cperf_verify_crypto_op_t verify_op_output;
  84
  85         const struct cperf_options *options;
  86         const struct cperf_test_vector *test_vector;
  87         struct cperf_op_result *res;
  88         struct cperf_latency_results results;
  89 };
  90
  91 #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b)
  92 #define min(a, b) (a < b ? (uint64_t)a : (uint64_t)b)
  93
  94 static void
  95 cperf_latency_test_free(struct cperf_latency_ctx *ctx, uint32_t mbuf_nb)
  96 {
  97         uint32_t i;
  98
  99         if (ctx) {
 100                 if (ctx->sess)
 101                         rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
 102
 103                 if (ctx->mbufs_in) {
 104                         for (i = 0; i < mbuf_nb; i++)
 105                                 rte_pktmbuf_free(ctx->mbufs_in[i]);
 106
 107                         rte_free(ctx->mbufs_in);
 108                 }
 109
 110                 if (ctx->mbufs_out) {
 111                         for (i = 0; i < mbuf_nb; i++) {
 112                                 if (ctx->mbufs_out[i] != NULL)
 113                                         rte_pktmbuf_free(ctx->mbufs_out[i]);
 114                         }
 115
 116                         rte_free(ctx->mbufs_out);
 117                 }
 118
 119                 if (ctx->pkt_mbuf_pool_in)
 120                         rte_mempool_free(ctx->pkt_mbuf_pool_in);
 121
 122                 if (ctx->pkt_mbuf_pool_out)
 123                         rte_mempool_free(ctx->pkt_mbuf_pool_out);
 124
 125                 if (ctx->crypto_op_pool)
 126                         rte_mempool_free(ctx->crypto_op_pool);
 127
 128                 rte_free(ctx->res);
 129                 rte_free(ctx);
 130         }
 131 }
 132
 133 static struct rte_mbuf *
 134 cperf_mbuf_create(struct rte_mempool *mempool,
 135                 uint32_t segments_nb,
 136                 const struct cperf_options *options,
 137                 const struct cperf_test_vector *test_vector)
 138 {
 139         struct rte_mbuf *mbuf;
 140         uint32_t segment_sz = options->buffer_sz / segments_nb;
 141         uint32_t last_sz = options->buffer_sz % segments_nb;
 142         uint8_t *mbuf_data;
 143         uint8_t *test_data =
 144                         (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ?
 145                                         test_vector->plaintext.data :
 146                                         test_vector->ciphertext.data;
 147
 148         mbuf = rte_pktmbuf_alloc(mempool);
 149         if (mbuf == NULL)
 150                 goto error;
 151
 152         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 153         if (mbuf_data == NULL)
 154                 goto error;
 155
 156         memcpy(mbuf_data, test_data, segment_sz);
 157         test_data += segment_sz;
 158         segments_nb--;
 159
 160         while (segments_nb) {
 161                 struct rte_mbuf *m;
 162
 163                 m = rte_pktmbuf_alloc(mempool);
 164                 if (m == NULL)
 165                         goto error;
 166
 167                 rte_pktmbuf_chain(mbuf, m);
 168
 169                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 170                 if (mbuf_data == NULL)
 171                         goto error;
 172
 173                 memcpy(mbuf_data, test_data, segment_sz);
 174                 test_data += segment_sz;
 175                 segments_nb--;
 176         }
 177
 178         if (last_sz) {
 179                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz);
 180                 if (mbuf_data == NULL)
 181                         goto error;
 182
 183                 memcpy(mbuf_data, test_data, last_sz);
 184         }
 185
 186         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf,
 187                         options->auth_digest_sz);
 188         if (mbuf_data == NULL)
 189                 goto error;
 190
 191         if (options->op_type == CPERF_AEAD) {
 192                 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf,
 193                         RTE_ALIGN_CEIL(options->auth_aad_sz, 16));
 194
 195                 if (aead == NULL)
 196                         goto error;
 197
 198                 memcpy(aead, test_vector->aad.data, test_vector->aad.length);
 199         }
 200
 201         return mbuf;
 202 error:
 203         if (mbuf != NULL)
 204                 rte_pktmbuf_free(mbuf);
 205
 206         return NULL;
 207 }
 208
 209 void *
 210 cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id,
 211                 const struct cperf_options *options,
 212                 const struct cperf_test_vector *test_vector,
 213                 const struct cperf_op_fns *op_fns)
 214 {
 215         struct cperf_latency_ctx *ctx = NULL;
 216         unsigned int mbuf_idx = 0;
 217         char pool_name[32] = "";
 218
 219         ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
 220         if (ctx == NULL)
 221                 goto err;
 222
 223         ctx->dev_id = dev_id;
 224         ctx->qp_id = qp_id;
 225
 226         ctx->populate_ops = op_fns->populate_ops;
 227         ctx->options = options;
 228         ctx->test_vector = test_vector;
 229
 230         ctx->sess = op_fns->sess_create(dev_id, options, test_vector);
 231         if (ctx->sess == NULL)
 232                 goto err;
 233
 234         snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d",
 235                                 dev_id);
 236
 237         ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name,
 238                         options->pool_sz * options->segments_nb, 0, 0,
 239                         RTE_PKTMBUF_HEADROOM +
 240                         RTE_CACHE_LINE_ROUNDUP(
 241                                 (options->buffer_sz / options->segments_nb) +
 242                                 (options->buffer_sz % options->segments_nb) +
 243                                         options->auth_digest_sz),
 244                         rte_socket_id());
 245
 246         if (ctx->pkt_mbuf_pool_in == NULL)
 247                 goto err;
 248
 249         /* Generate mbufs_in with plaintext populated for test */
 250         if (ctx->options->pool_sz % ctx->options->burst_sz)
 251                 goto err;
 252
 253         ctx->mbufs_in = rte_malloc(NULL,
 254                         (sizeof(struct rte_mbuf *) *
 255                         ctx->options->pool_sz), 0);
 256
 257         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 258                 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create(
 259                                 ctx->pkt_mbuf_pool_in, options->segments_nb,
 260                                 options, test_vector);
 261                 if (ctx->mbufs_in[mbuf_idx] == NULL)
 262                         goto err;
 263         }
 264
 265         if (options->out_of_place == 1) {
 266
 267                 snprintf(pool_name, sizeof(pool_name),
 268                                 "cperf_pool_out_cdev_%d",
 269                                 dev_id);
 270
 271                 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create(
 272                                 pool_name, options->pool_sz, 0, 0,
 273                                 RTE_PKTMBUF_HEADROOM +
 274                                 RTE_CACHE_LINE_ROUNDUP(
 275                                         options->buffer_sz +
 276                                         options->auth_digest_sz),
 277                                 rte_socket_id());
 278
 279                 if (ctx->pkt_mbuf_pool_out == NULL)
 280                         goto err;
 281         }
 282
 283         ctx->mbufs_out = rte_malloc(NULL,
 284                         (sizeof(struct rte_mbuf *) *
 285                         ctx->options->pool_sz), 0);
 286
 287         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 288                 if (options->out_of_place == 1) {
 289                         ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create(
 290                                         ctx->pkt_mbuf_pool_out, 1,
 291                                         options, test_vector);
 292                         if (ctx->mbufs_out[mbuf_idx] == NULL)
 293                                 goto err;
 294                 } else {
 295                         ctx->mbufs_out[mbuf_idx] = NULL;
 296                 }
 297         }
 298
 299         snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d",
 300                         dev_id);
 301
 302         ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name,
 303                         RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 0, 0,
 304                         rte_socket_id());
 305         if (ctx->crypto_op_pool == NULL)
 306                 goto err;
 307
 308         ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
 309                         ctx->options->total_ops, 0);
 310
 311         if (ctx->res == NULL)
 312                 goto err;
 313
 314         return ctx;
 315 err:
 316         cperf_latency_test_free(ctx, mbuf_idx);
 317
 318         return NULL;
 319 }
 320
 321 static int
 322 cperf_latency_test_verifier(struct rte_mbuf *mbuf,
 323                 const struct cperf_options *options,
 324                 const struct cperf_test_vector *vector)
 325 {
 326         const struct rte_mbuf *m;
 327         uint32_t len;
 328         uint16_t nb_segs;
 329         uint8_t *data;
 330         uint32_t cipher_offset, auth_offset;
 331         uint8_t cipher, auth;
 332         int res = 0;
 333
 334         m = mbuf;
 335         nb_segs = m->nb_segs;
 336         len = 0;
 337         while (m && nb_segs != 0) {
 338                 len += m->data_len;
 339                 m = m->next;
 340                 nb_segs--;
 341         }
 342
 343         data = rte_malloc(NULL, len, 0);
 344         if (data == NULL)
 345                 return 1;
 346
 347         m = mbuf;
 348         nb_segs = m->nb_segs;
 349         len = 0;
 350         while (m && nb_segs != 0) {
 351                 memcpy(data + len, rte_pktmbuf_mtod(m, uint8_t *),
 352                                 m->data_len);
 353                 len += m->data_len;
 354                 m = m->next;
 355                 nb_segs--;
 356         }
 357
 358         switch (options->op_type) {
 359         case CPERF_CIPHER_ONLY:
 360                 cipher = 1;
 361                 cipher_offset = 0;
 362                 auth = 0;
 363                 auth_offset = 0;
 364                 break;
 365         case CPERF_CIPHER_THEN_AUTH:
 366                 cipher = 1;
 367                 cipher_offset = 0;
 368                 auth = 1;
 369                 auth_offset = vector->plaintext.length;
 370                 break;
 371         case CPERF_AUTH_ONLY:
 372                 cipher = 0;
 373                 cipher_offset = 0;
 374                 auth = 1;
 375                 auth_offset = vector->plaintext.length;
 376                 break;
 377         case CPERF_AUTH_THEN_CIPHER:
 378                 cipher = 1;
 379                 cipher_offset = 0;
 380                 auth = 1;
 381                 auth_offset = vector->plaintext.length;
 382                 break;
 383         case CPERF_AEAD:
 384                 cipher = 1;
 385                 cipher_offset = vector->aad.length;
 386                 auth = 1;
 387                 auth_offset = vector->aad.length + vector->plaintext.length;
 388                 break;
 389         }
 390
 391         if (cipher == 1) {
 392                 if (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT)
 393                         res += memcmp(data + cipher_offset,
 394                                         vector->ciphertext.data,
 395                                         vector->ciphertext.length);
 396                 else
 397                         res += memcmp(data + cipher_offset,
 398                                         vector->plaintext.data,
 399                                         vector->plaintext.length);
 400         }
 401
 402         if (auth == 1) {
 403                 if (options->auth_op == RTE_CRYPTO_AUTH_OP_GENERATE)
 404                         res += memcmp(data + auth_offset,
 405                                         vector->digest.data,
 406                                         vector->digest.length);
 407         }
 408
 409         if (res != 0)
 410                 res = 1;
 411
 412         return res;
 413 }
 414
 415 int
 416 cperf_latency_test_runner(void *arg)
 417 {
 418         struct cperf_latency_ctx *ctx = arg;
 419         struct cperf_op_result *pres;
 420
 421         if (ctx == NULL)
 422                 return 0;
 423
 424         struct rte_crypto_op *ops[ctx->options->burst_sz];
 425         struct rte_crypto_op *ops_processed[ctx->options->burst_sz];
 426         uint64_t ops_enqd = 0, ops_deqd = 0;
 427         uint64_t m_idx = 0, b_idx = 0, i;
 428
 429         uint64_t tsc_val, tsc_end, tsc_start;
 430         uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
 431         uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
 432         uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
 433
 434         uint32_t lcore = rte_lcore_id();
 435
 436 #ifdef CPERF_LINEARIZATION_ENABLE
 437         struct rte_cryptodev_info dev_info;
 438         int linearize = 0;
 439
 440         /* Check if source mbufs require coalescing */
 441         if (ctx->options->segments_nb > 1) {
 442                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 443                 if ((dev_info.feature_flags &
 444                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 445                         linearize = 1;
 446         }
 447 #endif /* CPERF_LINEARIZATION_ENABLE */
 448
 449         ctx->lcore_id = lcore;
 450
 451         /* Warm up the host CPU before starting the test */
 452         for (i = 0; i < ctx->options->total_ops; i++)
 453                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 454
 455         while (enqd_tot < ctx->options->total_ops) {
 456
 457                 uint16_t burst_size = ((enqd_tot + ctx->options->burst_sz)
 458                                 <= ctx->options->total_ops) ?
 459                                                 ctx->options->burst_sz :
 460                                                 ctx->options->total_ops -
 461                                                 enqd_tot;
 462
 463                 /* Allocate crypto ops from pool */
 464                 if (burst_size != rte_crypto_op_bulk_alloc(
 465                                 ctx->crypto_op_pool,
 466                                 RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 467                                 ops, burst_size))
 468                         return -1;
 469
 470                 /* Setup crypto op, attach mbuf etc */
 471                 (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 472                                 &ctx->mbufs_out[m_idx],
 473                                 burst_size, ctx->sess, ctx->options,
 474                                 ctx->test_vector);
 475
 476                 tsc_start = rte_rdtsc_precise();
 477
 478 #ifdef CPERF_LINEARIZATION_ENABLE
 479                 if (linearize) {
 480                         /* PMD doesn't support scatter-gather and source buffer
 481                          * is segmented.
 482                          * We need to linearize it before enqueuing.
 483                          */
 484                         for (i = 0; i < burst_size; i++)
 485                                 rte_pktmbuf_linearize(ops[i]->sym->m_src);
 486                 }
 487 #endif /* CPERF_LINEARIZATION_ENABLE */
 488
 489                 /* Enqueue burst of ops on crypto device */
 490                 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 491                                 ops, burst_size);
 492
 493                 /* Dequeue processed burst of ops from crypto device */
 494                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 495                                 ops_processed, ctx->options->burst_sz);
 496
 497                 tsc_end = rte_rdtsc_precise();
 498
 499                 for (i = 0; i < ops_enqd; i++) {
 500                         ctx->res[tsc_idx].tsc_start = tsc_start;
 501                         ops[i]->opaque_data = (void *)&ctx->res[tsc_idx];
 502                         tsc_idx++;
 503                 }
 504
 505                 /* Free memory for not enqueued operations */
 506                 for (i = ops_enqd; i < burst_size; i++)
 507                         rte_crypto_op_free(ops[i]);
 508
 509                 if (likely(ops_deqd))  {
 510                         /*
 511                          * free crypto ops so they can be reused. We don't free
 512                          * the mbufs here as we don't want to reuse them as
 513                          * the crypto operation will change the data and cause
 514                          * failures.
 515                          */
 516                         for (i = 0; i < ops_deqd; i++) {
 517                                 pres = (struct cperf_op_result *)
 518                                                 (ops_processed[i]->opaque_data);
 519                                 pres->status = ops_processed[i]->status;
 520                                 pres->tsc_end = tsc_end;
 521
 522                                 rte_crypto_op_free(ops_processed[i]);
 523                         }
 524
 525                         deqd_tot += ops_deqd;
 526                         deqd_max = max(ops_deqd, deqd_max);
 527                         deqd_min = min(ops_deqd, deqd_min);
 528                 }
 529
 530                 enqd_tot += ops_enqd;
 531                 enqd_max = max(ops_enqd, enqd_max);
 532                 enqd_min = min(ops_enqd, enqd_min);
 533
 534                 m_idx += ops_enqd;
 535                 m_idx = m_idx + ctx->options->burst_sz > ctx->options->pool_sz ?
 536                                 0 : m_idx;
 537                 b_idx++;
 538         }
 539
 540         /* Dequeue any operations still in the crypto device */
 541         while (deqd_tot < ctx->options->total_ops) {
 542                 /* Sending 0 length burst to flush sw crypto device */
 543                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 544
 545                 /* dequeue burst */
 546                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 547                                 ops_processed, ctx->options->burst_sz);
 548
 549                 tsc_end = rte_rdtsc_precise();
 550
 551                 if (ops_deqd != 0) {
 552                         for (i = 0; i < ops_deqd; i++) {
 553                                 pres = (struct cperf_op_result *)
 554                                                 (ops_processed[i]->opaque_data);
 555                                 pres->status = ops_processed[i]->status;
 556                                 pres->tsc_end = tsc_end;
 557
 558                                 rte_crypto_op_free(ops_processed[i]);
 559                         }
 560
 561                         deqd_tot += ops_deqd;
 562                         deqd_max = max(ops_deqd, deqd_max);
 563                         deqd_min = min(ops_deqd, deqd_min);
 564                 }
 565         }
 566
 567         for (i = 0; i < tsc_idx; i++) {
 568                 tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start;
 569                 tsc_max = max(tsc_val, tsc_max);
 570                 tsc_min = min(tsc_val, tsc_min);
 571                 tsc_tot += tsc_val;
 572         }
 573
 574         if (ctx->options->verify) {
 575                 struct rte_mbuf **mbufs;
 576
 577                 if (ctx->options->out_of_place == 1)
 578                         mbufs = ctx->mbufs_out;
 579                 else
 580                         mbufs = ctx->mbufs_in;
 581
 582                 for (i = 0; i < ctx->options->total_ops; i++) {
 583
 584                         if (ctx->res[i].status != RTE_CRYPTO_OP_STATUS_SUCCESS
 585                                         || cperf_latency_test_verifier(mbufs[i],
 586                                                         ctx->options,
 587                                                         ctx->test_vector)) {
 588
 589                                 ctx->results.ops_failed++;
 590                         }
 591                 }
 592         }
 593
 594         ctx->results.enqd_tot = enqd_tot;
 595         ctx->results.enqd_max = enqd_max;
 596         ctx->results.enqd_min = enqd_min;
 597
 598         ctx->results.deqd_tot = deqd_tot;
 599         ctx->results.deqd_max = deqd_max;
 600         ctx->results.deqd_min = deqd_min;
 601
 602         ctx->results.cycles_tot = tsc_tot;
 603         ctx->results.cycles_max = tsc_max;
 604         ctx->results.cycles_min = tsc_min;
 605
 606         ctx->results.burst_num = b_idx;
 607         ctx->results.num = tsc_idx;
 608
 609         return 0;
 610 }
 611
 612 void
 613 cperf_latency_test_destructor(void *arg)
 614 {
 615         struct cperf_latency_ctx *ctx = arg;
 616         uint64_t i;
 617         if (ctx == NULL)
 618                 return;
 619         static int only_once;
 620         uint64_t etot, eavg, emax, emin;
 621         uint64_t dtot, davg, dmax, dmin;
 622         uint64_t ctot, cavg, cmax, cmin;
 623         double ttot, tavg, tmax, tmin;
 624
 625         const uint64_t tunit = 1000000; /* us */
 626         const uint64_t tsc_hz = rte_get_tsc_hz();
 627
 628         etot = ctx->results.enqd_tot;
 629         eavg = ctx->results.enqd_tot / ctx->results.burst_num;
 630         emax = ctx->results.enqd_max;
 631         emin = ctx->results.enqd_min;
 632
 633         dtot = ctx->results.deqd_tot;
 634         davg = ctx->results.deqd_tot / ctx->results.burst_num;
 635         dmax = ctx->results.deqd_max;
 636         dmin = ctx->results.deqd_min;
 637
 638         ctot = ctx->results.cycles_tot;
 639         cavg = ctx->results.cycles_tot / ctx->results.num;
 640         cmax = ctx->results.cycles_max;
 641         cmin = ctx->results.cycles_min;
 642
 643         ttot = tunit*(double)(ctot) / tsc_hz;
 644         tavg = tunit*(double)(cavg) / tsc_hz;
 645         tmax = tunit*(double)(cmax) / tsc_hz;
 646         tmin = tunit*(double)(cmin) / tsc_hz;
 647
 648         if (ctx->options->csv) {
 649                 if (!only_once)
 650                         printf("\n# lcore, Pakt Seq #, Packet Size, cycles,"
 651                                         " time (us)");
 652
 653                 for (i = 0; i < ctx->options->total_ops; i++) {
 654
 655                         printf("\n%u;%"PRIu64";%"PRIu64";%.3f",
 656                                 ctx->lcore_id, i + 1,
 657                                 ctx->res[i].tsc_end - ctx->res[i].tsc_start,
 658                                 tunit * (double) (ctx->res[i].tsc_end
 659                                                 - ctx->res[i].tsc_start)
 660                                         / tsc_hz);
 661
 662                 }
 663                 only_once = 1;
 664         } else {
 665                 printf("\n# Device %d on lcore %u\n", ctx->dev_id,
 666                         ctx->lcore_id);
 667                 printf("\n# total operations: %u", ctx->options->total_ops);
 668                 printf("\n#  verified failed: %"PRIu64,
 669                                 ctx->results.ops_failed);
 670                 printf("\n#     burst number: %"PRIu64,
 671                                 ctx->results.burst_num);
 672                 printf("\n#");
 673                 printf("\n#          \t       Total\t   Average\t   Maximum\t "
 674                                 "  Minimum");
 675                 printf("\n#  enqueued\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 676                                 "%10"PRIu64, etot, eavg, emax, emin);
 677                 printf("\n#  dequeued\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 678                                 "%10"PRIu64, dtot, davg, dmax, dmin);
 679                 printf("\n#    cycles\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 680                                 "%10"PRIu64, ctot, cavg, cmax, cmin);
 681                 printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f", ttot,
 682                         tavg, tmax, tmin);
 683                 printf("\n\n");
 684
 685         }
 686         cperf_latency_test_free(ctx, ctx->options->pool_sz);
 687
 688 }