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
  84         const struct cperf_options *options;
  85         const struct cperf_test_vector *test_vector;
  86         struct cperf_op_result *res;
  87         struct cperf_latency_results results;
  88 };
  89
  90 #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b)
  91 #define min(a, b) (a < b ? (uint64_t)a : (uint64_t)b)
  92
  93 static void
  94 cperf_latency_test_free(struct cperf_latency_ctx *ctx, uint32_t mbuf_nb)
  95 {
  96         uint32_t i;
  97
  98         if (ctx) {
  99                 if (ctx->sess)
 100                         rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
 101
 102                 if (ctx->mbufs_in) {
 103                         for (i = 0; i < mbuf_nb; i++)
 104                                 rte_pktmbuf_free(ctx->mbufs_in[i]);
 105
 106                         rte_free(ctx->mbufs_in);
 107                 }
 108
 109                 if (ctx->mbufs_out) {
 110                         for (i = 0; i < mbuf_nb; i++) {
 111                                 if (ctx->mbufs_out[i] != NULL)
 112                                         rte_pktmbuf_free(ctx->mbufs_out[i]);
 113                         }
 114
 115                         rte_free(ctx->mbufs_out);
 116                 }
 117
 118                 if (ctx->pkt_mbuf_pool_in)
 119                         rte_mempool_free(ctx->pkt_mbuf_pool_in);
 120
 121                 if (ctx->pkt_mbuf_pool_out)
 122                         rte_mempool_free(ctx->pkt_mbuf_pool_out);
 123
 124                 if (ctx->crypto_op_pool)
 125                         rte_mempool_free(ctx->crypto_op_pool);
 126
 127                 rte_free(ctx->res);
 128                 rte_free(ctx);
 129         }
 130 }
 131
 132 static struct rte_mbuf *
 133 cperf_mbuf_create(struct rte_mempool *mempool,
 134                 uint32_t segments_nb,
 135                 const struct cperf_options *options,
 136                 const struct cperf_test_vector *test_vector)
 137 {
 138         struct rte_mbuf *mbuf;
 139         uint32_t segment_sz = options->buffer_sz / segments_nb;
 140         uint32_t last_sz = options->buffer_sz % segments_nb;
 141         uint8_t *mbuf_data;
 142         uint8_t *test_data =
 143                         (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ?
 144                                         test_vector->plaintext.data :
 145                                         test_vector->ciphertext.data;
 146
 147         mbuf = rte_pktmbuf_alloc(mempool);
 148         if (mbuf == NULL)
 149                 goto error;
 150
 151         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 152         if (mbuf_data == NULL)
 153                 goto error;
 154
 155         memcpy(mbuf_data, test_data, segment_sz);
 156         test_data += segment_sz;
 157         segments_nb--;
 158
 159         while (segments_nb) {
 160                 struct rte_mbuf *m;
 161
 162                 m = rte_pktmbuf_alloc(mempool);
 163                 if (m == NULL)
 164                         goto error;
 165
 166                 rte_pktmbuf_chain(mbuf, m);
 167
 168                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 169                 if (mbuf_data == NULL)
 170                         goto error;
 171
 172                 memcpy(mbuf_data, test_data, segment_sz);
 173                 test_data += segment_sz;
 174                 segments_nb--;
 175         }
 176
 177         if (last_sz) {
 178                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz);
 179                 if (mbuf_data == NULL)
 180                         goto error;
 181
 182                 memcpy(mbuf_data, test_data, last_sz);
 183         }
 184
 185         if (options->op_type != CPERF_CIPHER_ONLY) {
 186                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf,
 187                         options->auth_digest_sz);
 188                 if (mbuf_data == NULL)
 189                         goto error;
 190         }
 191
 192         if (options->op_type == CPERF_AEAD) {
 193                 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf,
 194                         RTE_ALIGN_CEIL(options->auth_aad_sz, 16));
 195
 196                 if (aead == NULL)
 197                         goto error;
 198
 199                 memcpy(aead, test_vector->aad.data, test_vector->aad.length);
 200         }
 201
 202         return mbuf;
 203 error:
 204         if (mbuf != NULL)
 205                 rte_pktmbuf_free(mbuf);
 206
 207         return NULL;
 208 }
 209
 210 void *
 211 cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id,
 212                 const struct cperf_options *options,
 213                 const struct cperf_test_vector *test_vector,
 214                 const struct cperf_op_fns *op_fns)
 215 {
 216         struct cperf_latency_ctx *ctx = NULL;
 217         unsigned int mbuf_idx = 0;
 218         char pool_name[32] = "";
 219
 220         ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
 221         if (ctx == NULL)
 222                 goto err;
 223
 224         ctx->dev_id = dev_id;
 225         ctx->qp_id = qp_id;
 226
 227         ctx->populate_ops = op_fns->populate_ops;
 228         ctx->options = options;
 229         ctx->test_vector = test_vector;
 230
 231         ctx->sess = op_fns->sess_create(dev_id, options, test_vector);
 232         if (ctx->sess == NULL)
 233                 goto err;
 234
 235         snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d",
 236                                 dev_id);
 237
 238         ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name,
 239                         options->pool_sz * options->segments_nb, 0, 0,
 240                         RTE_PKTMBUF_HEADROOM +
 241                         RTE_CACHE_LINE_ROUNDUP(
 242                                 (options->buffer_sz / options->segments_nb) +
 243                                 (options->buffer_sz % options->segments_nb) +
 244                                         options->auth_digest_sz),
 245                         rte_socket_id());
 246
 247         if (ctx->pkt_mbuf_pool_in == NULL)
 248                 goto err;
 249
 250         /* Generate mbufs_in with plaintext populated for test */
 251         if (ctx->options->pool_sz % ctx->options->burst_sz)
 252                 goto err;
 253
 254         ctx->mbufs_in = rte_malloc(NULL,
 255                         (sizeof(struct rte_mbuf *) *
 256                         ctx->options->pool_sz), 0);
 257
 258         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 259                 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create(
 260                                 ctx->pkt_mbuf_pool_in, options->segments_nb,
 261                                 options, test_vector);
 262                 if (ctx->mbufs_in[mbuf_idx] == NULL)
 263                         goto err;
 264         }
 265
 266         if (options->out_of_place == 1) {
 267
 268                 snprintf(pool_name, sizeof(pool_name),
 269                                 "cperf_pool_out_cdev_%d",
 270                                 dev_id);
 271
 272                 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create(
 273                                 pool_name, options->pool_sz, 0, 0,
 274                                 RTE_PKTMBUF_HEADROOM +
 275                                 RTE_CACHE_LINE_ROUNDUP(
 276                                         options->buffer_sz +
 277                                         options->auth_digest_sz),
 278                                 rte_socket_id());
 279
 280                 if (ctx->pkt_mbuf_pool_out == NULL)
 281                         goto err;
 282         }
 283
 284         ctx->mbufs_out = rte_malloc(NULL,
 285                         (sizeof(struct rte_mbuf *) *
 286                         ctx->options->pool_sz), 0);
 287
 288         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 289                 if (options->out_of_place == 1) {
 290                         ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create(
 291                                         ctx->pkt_mbuf_pool_out, 1,
 292                                         options, test_vector);
 293                         if (ctx->mbufs_out[mbuf_idx] == NULL)
 294                                 goto err;
 295                 } else {
 296                         ctx->mbufs_out[mbuf_idx] = NULL;
 297                 }
 298         }
 299
 300         snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d",
 301                         dev_id);
 302
 303         ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name,
 304                         RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 0, 0,
 305                         rte_socket_id());
 306         if (ctx->crypto_op_pool == NULL)
 307                 goto err;
 308
 309         ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
 310                         ctx->options->total_ops, 0);
 311
 312         if (ctx->res == NULL)
 313                 goto err;
 314
 315         return ctx;
 316 err:
 317         cperf_latency_test_free(ctx, mbuf_idx);
 318
 319         return NULL;
 320 }
 321
 322 int
 323 cperf_latency_test_runner(void *arg)
 324 {
 325         struct cperf_latency_ctx *ctx = arg;
 326         struct cperf_op_result *pres;
 327
 328         if (ctx == NULL)
 329                 return 0;
 330
 331         struct rte_crypto_op *ops[ctx->options->burst_sz];
 332         struct rte_crypto_op *ops_processed[ctx->options->burst_sz];
 333         uint64_t ops_enqd = 0, ops_deqd = 0;
 334         uint64_t m_idx = 0, b_idx = 0, i;
 335
 336         uint64_t tsc_val, tsc_end, tsc_start;
 337         uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
 338         uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
 339         uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
 340
 341         uint32_t lcore = rte_lcore_id();
 342
 343 #ifdef CPERF_LINEARIZATION_ENABLE
 344         struct rte_cryptodev_info dev_info;
 345         int linearize = 0;
 346
 347         /* Check if source mbufs require coalescing */
 348         if (ctx->options->segments_nb > 1) {
 349                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 350                 if ((dev_info.feature_flags &
 351                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 352                         linearize = 1;
 353         }
 354 #endif /* CPERF_LINEARIZATION_ENABLE */
 355
 356         ctx->lcore_id = lcore;
 357
 358         /* Warm up the host CPU before starting the test */
 359         for (i = 0; i < ctx->options->total_ops; i++)
 360                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 361
 362         while (enqd_tot < ctx->options->total_ops) {
 363
 364                 uint16_t burst_size = ((enqd_tot + ctx->options->burst_sz)
 365                                 <= ctx->options->total_ops) ?
 366                                                 ctx->options->burst_sz :
 367                                                 ctx->options->total_ops -
 368                                                 enqd_tot;
 369
 370                 /* Allocate crypto ops from pool */
 371                 if (burst_size != rte_crypto_op_bulk_alloc(
 372                                 ctx->crypto_op_pool,
 373                                 RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 374                                 ops, burst_size))
 375                         return -1;
 376
 377                 /* Setup crypto op, attach mbuf etc */
 378                 (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 379                                 &ctx->mbufs_out[m_idx],
 380                                 burst_size, ctx->sess, ctx->options,
 381                                 ctx->test_vector);
 382
 383                 tsc_start = rte_rdtsc_precise();
 384
 385 #ifdef CPERF_LINEARIZATION_ENABLE
 386                 if (linearize) {
 387                         /* PMD doesn't support scatter-gather and source buffer
 388                          * is segmented.
 389                          * We need to linearize it before enqueuing.
 390                          */
 391                         for (i = 0; i < burst_size; i++)
 392                                 rte_pktmbuf_linearize(ops[i]->sym->m_src);
 393                 }
 394 #endif /* CPERF_LINEARIZATION_ENABLE */
 395
 396                 /* Enqueue burst of ops on crypto device */
 397                 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 398                                 ops, burst_size);
 399
 400                 /* Dequeue processed burst of ops from crypto device */
 401                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 402                                 ops_processed, ctx->options->burst_sz);
 403
 404                 tsc_end = rte_rdtsc_precise();
 405
 406                 for (i = 0; i < ops_enqd; i++) {
 407                         ctx->res[tsc_idx].tsc_start = tsc_start;
 408                         ops[i]->opaque_data = (void *)&ctx->res[tsc_idx];
 409                         tsc_idx++;
 410                 }
 411
 412                 /* Free memory for not enqueued operations */
 413                 for (i = ops_enqd; i < burst_size; i++)
 414                         rte_crypto_op_free(ops[i]);
 415
 416                 if (likely(ops_deqd))  {
 417                         /*
 418                          * free crypto ops so they can be reused. We don't free
 419                          * the mbufs here as we don't want to reuse them as
 420                          * the crypto operation will change the data and cause
 421                          * failures.
 422                          */
 423                         for (i = 0; i < ops_deqd; i++) {
 424                                 pres = (struct cperf_op_result *)
 425                                                 (ops_processed[i]->opaque_data);
 426                                 pres->status = ops_processed[i]->status;
 427                                 pres->tsc_end = tsc_end;
 428
 429                                 rte_crypto_op_free(ops_processed[i]);
 430                         }
 431
 432                         deqd_tot += ops_deqd;
 433                         deqd_max = max(ops_deqd, deqd_max);
 434                         deqd_min = min(ops_deqd, deqd_min);
 435                 }
 436
 437                 enqd_tot += ops_enqd;
 438                 enqd_max = max(ops_enqd, enqd_max);
 439                 enqd_min = min(ops_enqd, enqd_min);
 440
 441                 m_idx += ops_enqd;
 442                 m_idx = m_idx + ctx->options->burst_sz > ctx->options->pool_sz ?
 443                                 0 : m_idx;
 444                 b_idx++;
 445         }
 446
 447         /* Dequeue any operations still in the crypto device */
 448         while (deqd_tot < ctx->options->total_ops) {
 449                 /* Sending 0 length burst to flush sw crypto device */
 450                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 451
 452                 /* dequeue burst */
 453                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 454                                 ops_processed, ctx->options->burst_sz);
 455
 456                 tsc_end = rte_rdtsc_precise();
 457
 458                 if (ops_deqd != 0) {
 459                         for (i = 0; i < ops_deqd; i++) {
 460                                 pres = (struct cperf_op_result *)
 461                                                 (ops_processed[i]->opaque_data);
 462                                 pres->status = ops_processed[i]->status;
 463                                 pres->tsc_end = tsc_end;
 464
 465                                 rte_crypto_op_free(ops_processed[i]);
 466                         }
 467
 468                         deqd_tot += ops_deqd;
 469                         deqd_max = max(ops_deqd, deqd_max);
 470                         deqd_min = min(ops_deqd, deqd_min);
 471                 }
 472         }
 473
 474         for (i = 0; i < tsc_idx; i++) {
 475                 tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start;
 476                 tsc_max = max(tsc_val, tsc_max);
 477                 tsc_min = min(tsc_val, tsc_min);
 478                 tsc_tot += tsc_val;
 479         }
 480
 481         ctx->results.enqd_tot = enqd_tot;
 482         ctx->results.enqd_max = enqd_max;
 483         ctx->results.enqd_min = enqd_min;
 484
 485         ctx->results.deqd_tot = deqd_tot;
 486         ctx->results.deqd_max = deqd_max;
 487         ctx->results.deqd_min = deqd_min;
 488
 489         ctx->results.cycles_tot = tsc_tot;
 490         ctx->results.cycles_max = tsc_max;
 491         ctx->results.cycles_min = tsc_min;
 492
 493         ctx->results.burst_num = b_idx;
 494         ctx->results.num = tsc_idx;
 495
 496         return 0;
 497 }
 498
 499 void
 500 cperf_latency_test_destructor(void *arg)
 501 {
 502         struct cperf_latency_ctx *ctx = arg;
 503         uint64_t i;
 504         if (ctx == NULL)
 505                 return;
 506         static int only_once;
 507         uint64_t etot, eavg, emax, emin;
 508         uint64_t dtot, davg, dmax, dmin;
 509         uint64_t ctot, cavg, cmax, cmin;
 510         double ttot, tavg, tmax, tmin;
 511
 512         const uint64_t tunit = 1000000; /* us */
 513         const uint64_t tsc_hz = rte_get_tsc_hz();
 514
 515         etot = ctx->results.enqd_tot;
 516         eavg = ctx->results.enqd_tot / ctx->results.burst_num;
 517         emax = ctx->results.enqd_max;
 518         emin = ctx->results.enqd_min;
 519
 520         dtot = ctx->results.deqd_tot;
 521         davg = ctx->results.deqd_tot / ctx->results.burst_num;
 522         dmax = ctx->results.deqd_max;
 523         dmin = ctx->results.deqd_min;
 524
 525         ctot = ctx->results.cycles_tot;
 526         cavg = ctx->results.cycles_tot / ctx->results.num;
 527         cmax = ctx->results.cycles_max;
 528         cmin = ctx->results.cycles_min;
 529
 530         ttot = tunit*(double)(ctot) / tsc_hz;
 531         tavg = tunit*(double)(cavg) / tsc_hz;
 532         tmax = tunit*(double)(cmax) / tsc_hz;
 533         tmin = tunit*(double)(cmin) / tsc_hz;
 534
 535         if (ctx->options->csv) {
 536                 if (!only_once)
 537                         printf("\n# lcore, Pakt Seq #, Packet Size, cycles,"
 538                                         " time (us)");
 539
 540                 for (i = 0; i < ctx->options->total_ops; i++) {
 541
 542                         printf("\n%u;%"PRIu64";%"PRIu64";%.3f",
 543                                 ctx->lcore_id, i + 1,
 544                                 ctx->res[i].tsc_end - ctx->res[i].tsc_start,
 545                                 tunit * (double) (ctx->res[i].tsc_end
 546                                                 - ctx->res[i].tsc_start)
 547                                         / tsc_hz);
 548
 549                 }
 550                 only_once = 1;
 551         } else {
 552                 printf("\n# Device %d on lcore %u\n", ctx->dev_id,
 553                         ctx->lcore_id);
 554                 printf("\n# total operations: %u", ctx->options->total_ops);
 555                 printf("\n#     burst number: %"PRIu64,
 556                                 ctx->results.burst_num);
 557                 printf("\n#");
 558                 printf("\n#          \t       Total\t   Average\t   Maximum\t "
 559                                 "  Minimum");
 560                 printf("\n#  enqueued\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 561                                 "%10"PRIu64, etot, eavg, emax, emin);
 562                 printf("\n#  dequeued\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 563                                 "%10"PRIu64, dtot, davg, dmax, dmin);
 564                 printf("\n#    cycles\t%12"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 565                                 "%10"PRIu64, ctot, cavg, cmax, cmin);
 566                 printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f", ttot,
 567                         tavg, tmax, tmin);
 568                 printf("\n\n");
 569
 570         }
 571         cperf_latency_test_free(ctx, ctx->options->pool_sz);
 572
 573 }