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_op_result {
  43         uint64_t tsc_start;
  44         uint64_t tsc_end;
  45         enum rte_crypto_op_status status;
  46 };
  47
  48 struct cperf_latency_ctx {
  49         uint8_t dev_id;
  50         uint16_t qp_id;
  51         uint8_t lcore_id;
  52
  53         struct rte_mempool *pkt_mbuf_pool_in;
  54         struct rte_mempool *pkt_mbuf_pool_out;
  55         struct rte_mbuf **mbufs_in;
  56         struct rte_mbuf **mbufs_out;
  57
  58         struct rte_mempool *crypto_op_pool;
  59
  60         struct rte_cryptodev_sym_session *sess;
  61
  62         cperf_populate_ops_t populate_ops;
  63
  64         const struct cperf_options *options;
  65         const struct cperf_test_vector *test_vector;
  66         struct cperf_op_result *res;
  67 };
  68
  69 struct priv_op_data {
  70         struct cperf_op_result *result;
  71 };
  72
  73 #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b)
  74 #define min(a, b) (a < b ? (uint64_t)a : (uint64_t)b)
  75
  76 static void
  77 cperf_latency_test_free(struct cperf_latency_ctx *ctx, uint32_t mbuf_nb)
  78 {
  79         uint32_t i;
  80
  81         if (ctx) {
  82                 if (ctx->sess)
  83                         rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
  84
  85                 if (ctx->mbufs_in) {
  86                         for (i = 0; i < mbuf_nb; i++)
  87                                 rte_pktmbuf_free(ctx->mbufs_in[i]);
  88
  89                         rte_free(ctx->mbufs_in);
  90                 }
  91
  92                 if (ctx->mbufs_out) {
  93                         for (i = 0; i < mbuf_nb; i++) {
  94                                 if (ctx->mbufs_out[i] != NULL)
  95                                         rte_pktmbuf_free(ctx->mbufs_out[i]);
  96                         }
  97
  98                         rte_free(ctx->mbufs_out);
  99                 }
 100
 101                 if (ctx->pkt_mbuf_pool_in)
 102                         rte_mempool_free(ctx->pkt_mbuf_pool_in);
 103
 104                 if (ctx->pkt_mbuf_pool_out)
 105                         rte_mempool_free(ctx->pkt_mbuf_pool_out);
 106
 107                 if (ctx->crypto_op_pool)
 108                         rte_mempool_free(ctx->crypto_op_pool);
 109
 110                 rte_free(ctx->res);
 111                 rte_free(ctx);
 112         }
 113 }
 114
 115 static struct rte_mbuf *
 116 cperf_mbuf_create(struct rte_mempool *mempool,
 117                 uint32_t segments_nb,
 118                 const struct cperf_options *options,
 119                 const struct cperf_test_vector *test_vector)
 120 {
 121         struct rte_mbuf *mbuf;
 122         uint32_t segment_sz = options->max_buffer_size / segments_nb;
 123         uint32_t last_sz = options->max_buffer_size % segments_nb;
 124         uint8_t *mbuf_data;
 125         uint8_t *test_data =
 126                         (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ?
 127                                         test_vector->plaintext.data :
 128                                         test_vector->ciphertext.data;
 129
 130         mbuf = rte_pktmbuf_alloc(mempool);
 131         if (mbuf == NULL)
 132                 goto error;
 133
 134         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 135         if (mbuf_data == NULL)
 136                 goto error;
 137
 138         memcpy(mbuf_data, test_data, segment_sz);
 139         test_data += segment_sz;
 140         segments_nb--;
 141
 142         while (segments_nb) {
 143                 struct rte_mbuf *m;
 144
 145                 m = rte_pktmbuf_alloc(mempool);
 146                 if (m == NULL)
 147                         goto error;
 148
 149                 rte_pktmbuf_chain(mbuf, m);
 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
 160         if (last_sz) {
 161                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz);
 162                 if (mbuf_data == NULL)
 163                         goto error;
 164
 165                 memcpy(mbuf_data, test_data, last_sz);
 166         }
 167
 168         if (options->op_type != CPERF_CIPHER_ONLY) {
 169                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf,
 170                         options->auth_digest_sz);
 171                 if (mbuf_data == NULL)
 172                         goto error;
 173         }
 174
 175         if (options->op_type == CPERF_AEAD) {
 176                 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf,
 177                         RTE_ALIGN_CEIL(options->auth_aad_sz, 16));
 178
 179                 if (aead == NULL)
 180                         goto error;
 181
 182                 memcpy(aead, test_vector->aad.data, test_vector->aad.length);
 183         }
 184
 185         return mbuf;
 186 error:
 187         if (mbuf != NULL)
 188                 rte_pktmbuf_free(mbuf);
 189
 190         return NULL;
 191 }
 192
 193 void *
 194 cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id,
 195                 const struct cperf_options *options,
 196                 const struct cperf_test_vector *test_vector,
 197                 const struct cperf_op_fns *op_fns)
 198 {
 199         struct cperf_latency_ctx *ctx = NULL;
 200         unsigned int mbuf_idx = 0;
 201         char pool_name[32] = "";
 202
 203         ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
 204         if (ctx == NULL)
 205                 goto err;
 206
 207         ctx->dev_id = dev_id;
 208         ctx->qp_id = qp_id;
 209
 210         ctx->populate_ops = op_fns->populate_ops;
 211         ctx->options = options;
 212         ctx->test_vector = test_vector;
 213
 214         ctx->sess = op_fns->sess_create(dev_id, options, test_vector);
 215         if (ctx->sess == NULL)
 216                 goto err;
 217
 218         snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d",
 219                                 dev_id);
 220
 221         ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name,
 222                         options->pool_sz * options->segments_nb, 0, 0,
 223                         RTE_PKTMBUF_HEADROOM +
 224                         RTE_CACHE_LINE_ROUNDUP(
 225                                 (options->max_buffer_size / options->segments_nb) +
 226                                 (options->max_buffer_size % options->segments_nb) +
 227                                         options->auth_digest_sz),
 228                         rte_socket_id());
 229
 230         if (ctx->pkt_mbuf_pool_in == NULL)
 231                 goto err;
 232
 233         /* Generate mbufs_in with plaintext populated for test */
 234         ctx->mbufs_in = rte_malloc(NULL,
 235                         (sizeof(struct rte_mbuf *) *
 236                         ctx->options->pool_sz), 0);
 237
 238         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 239                 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create(
 240                                 ctx->pkt_mbuf_pool_in, options->segments_nb,
 241                                 options, test_vector);
 242                 if (ctx->mbufs_in[mbuf_idx] == NULL)
 243                         goto err;
 244         }
 245
 246         if (options->out_of_place == 1) {
 247
 248                 snprintf(pool_name, sizeof(pool_name),
 249                                 "cperf_pool_out_cdev_%d",
 250                                 dev_id);
 251
 252                 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create(
 253                                 pool_name, options->pool_sz, 0, 0,
 254                                 RTE_PKTMBUF_HEADROOM +
 255                                 RTE_CACHE_LINE_ROUNDUP(
 256                                         options->max_buffer_size +
 257                                         options->auth_digest_sz),
 258                                 rte_socket_id());
 259
 260                 if (ctx->pkt_mbuf_pool_out == NULL)
 261                         goto err;
 262         }
 263
 264         ctx->mbufs_out = rte_malloc(NULL,
 265                         (sizeof(struct rte_mbuf *) *
 266                         ctx->options->pool_sz), 0);
 267
 268         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 269                 if (options->out_of_place == 1) {
 270                         ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create(
 271                                         ctx->pkt_mbuf_pool_out, 1,
 272                                         options, test_vector);
 273                         if (ctx->mbufs_out[mbuf_idx] == NULL)
 274                                 goto err;
 275                 } else {
 276                         ctx->mbufs_out[mbuf_idx] = NULL;
 277                 }
 278         }
 279
 280         snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d",
 281                         dev_id);
 282
 283         uint16_t priv_size = sizeof(struct priv_op_data);
 284         ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name,
 285                         RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz,
 286                         512, priv_size, rte_socket_id());
 287
 288         if (ctx->crypto_op_pool == NULL)
 289                 goto err;
 290
 291         ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
 292                         ctx->options->total_ops, 0);
 293
 294         if (ctx->res == NULL)
 295                 goto err;
 296
 297         return ctx;
 298 err:
 299         cperf_latency_test_free(ctx, mbuf_idx);
 300
 301         return NULL;
 302 }
 303
 304 static inline void
 305 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp)
 306 {
 307         struct priv_op_data *priv_data;
 308
 309         priv_data = (struct priv_op_data *) (op->sym + 1);
 310         priv_data->result->status = op->status;
 311         priv_data->result->tsc_end = timestamp;
 312 }
 313
 314 int
 315 cperf_latency_test_runner(void *arg)
 316 {
 317         struct cperf_latency_ctx *ctx = arg;
 318         uint16_t test_burst_size;
 319         uint8_t burst_size_idx = 0;
 320
 321         static int only_once;
 322
 323         if (ctx == NULL)
 324                 return 0;
 325
 326         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 327         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 328         uint64_t i;
 329         struct priv_op_data *priv_data;
 330
 331         uint32_t lcore = rte_lcore_id();
 332
 333 #ifdef CPERF_LINEARIZATION_ENABLE
 334         struct rte_cryptodev_info dev_info;
 335         int linearize = 0;
 336
 337         /* Check if source mbufs require coalescing */
 338         if (ctx->options->segments_nb > 1) {
 339                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 340                 if ((dev_info.feature_flags &
 341                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 342                         linearize = 1;
 343         }
 344 #endif /* CPERF_LINEARIZATION_ENABLE */
 345
 346         ctx->lcore_id = lcore;
 347
 348         /* Warm up the host CPU before starting the test */
 349         for (i = 0; i < ctx->options->total_ops; i++)
 350                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 351
 352         /* Get first size from range or list */
 353         if (ctx->options->inc_burst_size != 0)
 354                 test_burst_size = ctx->options->min_burst_size;
 355         else
 356                 test_burst_size = ctx->options->burst_size_list[0];
 357
 358         while (test_burst_size <= ctx->options->max_burst_size) {
 359                 uint64_t ops_enqd = 0, ops_deqd = 0;
 360                 uint64_t m_idx = 0, b_idx = 0;
 361
 362                 uint64_t tsc_val, tsc_end, tsc_start;
 363                 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
 364                 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
 365                 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
 366
 367                 while (enqd_tot < ctx->options->total_ops) {
 368
 369                         uint16_t burst_size = ((enqd_tot + test_burst_size)
 370                                         <= ctx->options->total_ops) ?
 371                                                         test_burst_size :
 372                                                         ctx->options->total_ops -
 373                                                         enqd_tot;
 374
 375                         /* Allocate crypto ops from pool */
 376                         if (burst_size != rte_crypto_op_bulk_alloc(
 377                                         ctx->crypto_op_pool,
 378                                         RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 379                                         ops, burst_size))
 380                                 return -1;
 381
 382                         /* Setup crypto op, attach mbuf etc */
 383                         (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 384                                         &ctx->mbufs_out[m_idx],
 385                                         burst_size, ctx->sess, ctx->options,
 386                                         ctx->test_vector);
 387
 388                         tsc_start = rte_rdtsc_precise();
 389
 390 #ifdef CPERF_LINEARIZATION_ENABLE
 391                         if (linearize) {
 392                                 /* PMD doesn't support scatter-gather and source buffer
 393                                  * is segmented.
 394                                  * We need to linearize it before enqueuing.
 395                                  */
 396                                 for (i = 0; i < burst_size; i++)
 397                                         rte_pktmbuf_linearize(ops[i]->sym->m_src);
 398                         }
 399 #endif /* CPERF_LINEARIZATION_ENABLE */
 400
 401                         /* Enqueue burst of ops on crypto device */
 402                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 403                                         ops, burst_size);
 404
 405                         /* Dequeue processed burst of ops from crypto device */
 406                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 407                                         ops_processed, test_burst_size);
 408
 409                         tsc_end = rte_rdtsc_precise();
 410
 411                         /* Free memory for not enqueued operations */
 412                         if (ops_enqd != burst_size)
 413                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 414                                                 (void **)&ops_processed[ops_enqd],
 415                                                 burst_size - ops_enqd);
 416
 417                         for (i = 0; i < ops_enqd; i++) {
 418                                 ctx->res[tsc_idx].tsc_start = tsc_start;
 419                                 /*
 420                                  * Private data structure starts after the end of the
 421                                  * rte_crypto_sym_op structure.
 422                                  */
 423                                 priv_data = (struct priv_op_data *) (ops[i]->sym + 1);
 424                                 priv_data->result = (void *)&ctx->res[tsc_idx];
 425                                 tsc_idx++;
 426                         }
 427
 428                         if (likely(ops_deqd))  {
 429                                 /*
 430                                  * free crypto ops so they can be reused. We don't free
 431                                  * the mbufs here as we don't want to reuse them as
 432                                  * the crypto operation will change the data and cause
 433                                  * failures.
 434                                  */
 435                                 for (i = 0; i < ops_deqd; i++)
 436                                         store_timestamp(ops_processed[i], tsc_end);
 437
 438                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 439                                                 (void **)ops_processed, ops_deqd);
 440
 441                                 deqd_tot += ops_deqd;
 442                                 deqd_max = max(ops_deqd, deqd_max);
 443                                 deqd_min = min(ops_deqd, deqd_min);
 444                         }
 445
 446                         enqd_tot += ops_enqd;
 447                         enqd_max = max(ops_enqd, enqd_max);
 448                         enqd_min = min(ops_enqd, enqd_min);
 449
 450                         m_idx += ops_enqd;
 451                         m_idx = m_idx + test_burst_size > ctx->options->pool_sz ?
 452                                         0 : m_idx;
 453                         b_idx++;
 454                 }
 455
 456                 /* Dequeue any operations still in the crypto device */
 457                 while (deqd_tot < ctx->options->total_ops) {
 458                         /* Sending 0 length burst to flush sw crypto device */
 459                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 460
 461                         /* dequeue burst */
 462                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 463                                         ops_processed, test_burst_size);
 464
 465                         tsc_end = rte_rdtsc_precise();
 466
 467                         if (ops_deqd != 0) {
 468                                 for (i = 0; i < ops_deqd; i++)
 469                                         store_timestamp(ops_processed[i], tsc_end);
 470
 471                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 472                                                 (void **)ops_processed, ops_deqd);
 473
 474                                 deqd_tot += ops_deqd;
 475                                 deqd_max = max(ops_deqd, deqd_max);
 476                                 deqd_min = min(ops_deqd, deqd_min);
 477                         }
 478                 }
 479
 480                 for (i = 0; i < tsc_idx; i++) {
 481                         tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start;
 482                         tsc_max = max(tsc_val, tsc_max);
 483                         tsc_min = min(tsc_val, tsc_min);
 484                         tsc_tot += tsc_val;
 485                 }
 486
 487                 double time_tot, time_avg, time_max, time_min;
 488
 489                 const uint64_t tunit = 1000000; /* us */
 490                 const uint64_t tsc_hz = rte_get_tsc_hz();
 491
 492                 uint64_t enqd_avg = enqd_tot / b_idx;
 493                 uint64_t deqd_avg = deqd_tot / b_idx;
 494                 uint64_t tsc_avg = tsc_tot / tsc_idx;
 495
 496                 time_tot = tunit*(double)(tsc_tot) / tsc_hz;
 497                 time_avg = tunit*(double)(tsc_avg) / tsc_hz;
 498                 time_max = tunit*(double)(tsc_max) / tsc_hz;
 499                 time_min = tunit*(double)(tsc_min) / tsc_hz;
 500
 501                 if (ctx->options->csv) {
 502                         if (!only_once)
 503                                 printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, "
 504                                                 "Packet Size, cycles, time (us)");
 505
 506                         for (i = 0; i < ctx->options->total_ops; i++) {
 507
 508                                 printf("\n%u;%u;%u;%"PRIu64";%"PRIu64";%.3f",
 509                                         ctx->lcore_id, ctx->options->test_buffer_size,
 510                                         test_burst_size, i + 1,
 511                                         ctx->res[i].tsc_end - ctx->res[i].tsc_start,
 512                                         tunit * (double) (ctx->res[i].tsc_end
 513                                                         - ctx->res[i].tsc_start)
 514                                                 / tsc_hz);
 515
 516                         }
 517                         only_once = 1;
 518                 } else {
 519                         printf("\n# Device %d on lcore %u\n", ctx->dev_id,
 520                                 ctx->lcore_id);
 521                         printf("\n# total operations: %u", ctx->options->total_ops);
 522                         printf("\n# Buffer size: %u", ctx->options->test_buffer_size);
 523                         printf("\n# Burst size: %u", test_burst_size);
 524                         printf("\n#     Number of bursts: %"PRIu64,
 525                                         b_idx);
 526
 527                         printf("\n#");
 528                         printf("\n#          \t       Total\t   Average\t   "
 529                                         "Maximum\t   Minimum");
 530                         printf("\n#  enqueued\t%12"PRIu64"\t%10"PRIu64"\t"
 531                                         "%10"PRIu64"\t%10"PRIu64, enqd_tot,
 532                                         enqd_avg, enqd_max, enqd_min);
 533                         printf("\n#  dequeued\t%12"PRIu64"\t%10"PRIu64"\t"
 534                                         "%10"PRIu64"\t%10"PRIu64, deqd_tot,
 535                                         deqd_avg, deqd_max, deqd_min);
 536                         printf("\n#    cycles\t%12"PRIu64"\t%10"PRIu64"\t"
 537                                         "%10"PRIu64"\t%10"PRIu64, tsc_tot,
 538                                         tsc_avg, tsc_max, tsc_min);
 539                         printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f",
 540                                         time_tot, time_avg, time_max, time_min);
 541                         printf("\n\n");
 542
 543                 }
 544
 545                 /* Get next size from range or list */
 546                 if (ctx->options->inc_burst_size != 0)
 547                         test_burst_size += ctx->options->inc_burst_size;
 548                 else {
 549                         if (++burst_size_idx == ctx->options->burst_size_count)
 550                                 break;
 551                         test_burst_size =
 552                                 ctx->options->burst_size_list[burst_size_idx];
 553                 }
 554         }
 555
 556         return 0;
 557 }
 558
 559 void
 560 cperf_latency_test_destructor(void *arg)
 561 {
 562         struct cperf_latency_ctx *ctx = arg;
 563
 564         if (ctx == NULL)
 565                 return;
 566
 567         cperf_latency_test_free(ctx, ctx->options->pool_sz);
 568
 569 }