app/test-crypto-perf/cperf_test_throughput.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_throughput.h"
  39 #include "cperf_ops.h"
  40
  41 struct cperf_throughput_results {
  42         uint64_t ops_enqueued;
  43         uint64_t ops_dequeued;
  44
  45         uint64_t ops_enqueued_failed;
  46         uint64_t ops_dequeued_failed;
  47
  48         double ops_per_second;
  49         double throughput_gbps;
  50         double cycles_per_byte;
  51 };
  52
  53 struct cperf_throughput_ctx {
  54         uint8_t dev_id;
  55         uint16_t qp_id;
  56         uint8_t lcore_id;
  57
  58         struct rte_mempool *pkt_mbuf_pool_in;
  59         struct rte_mempool *pkt_mbuf_pool_out;
  60         struct rte_mbuf **mbufs_in;
  61         struct rte_mbuf **mbufs_out;
  62
  63         struct rte_mempool *crypto_op_pool;
  64
  65         struct rte_cryptodev_sym_session *sess;
  66
  67         cperf_populate_ops_t populate_ops;
  68
  69         const struct cperf_options *options;
  70         const struct cperf_test_vector *test_vector;
  71         struct cperf_throughput_results results;
  72
  73 };
  74
  75 static void
  76 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx, uint32_t mbuf_nb)
  77 {
  78         uint32_t i;
  79
  80         if (ctx) {
  81                 if (ctx->sess)
  82                         rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
  83
  84                 if (ctx->mbufs_in) {
  85                         for (i = 0; i < mbuf_nb; i++)
  86                                 rte_pktmbuf_free(ctx->mbufs_in[i]);
  87
  88                         rte_free(ctx->mbufs_in);
  89                 }
  90
  91                 if (ctx->mbufs_out) {
  92                         for (i = 0; i < mbuf_nb; i++) {
  93                                 if (ctx->mbufs_out[i] != NULL)
  94                                         rte_pktmbuf_free(ctx->mbufs_out[i]);
  95                         }
  96
  97                         rte_free(ctx->mbufs_out);
  98                 }
  99
 100                 if (ctx->pkt_mbuf_pool_in)
 101                         rte_mempool_free(ctx->pkt_mbuf_pool_in);
 102
 103                 if (ctx->pkt_mbuf_pool_out)
 104                         rte_mempool_free(ctx->pkt_mbuf_pool_out);
 105
 106                 if (ctx->crypto_op_pool)
 107                         rte_mempool_free(ctx->crypto_op_pool);
 108
 109                 rte_free(ctx);
 110         }
 111 }
 112
 113 static struct rte_mbuf *
 114 cperf_mbuf_create(struct rte_mempool *mempool,
 115                 uint32_t segments_nb,
 116                 const struct cperf_options *options,
 117                 const struct cperf_test_vector *test_vector)
 118 {
 119         struct rte_mbuf *mbuf;
 120         uint32_t segment_sz = options->buffer_sz / segments_nb;
 121         uint32_t last_sz = options->buffer_sz % segments_nb;
 122         uint8_t *mbuf_data;
 123         uint8_t *test_data =
 124                         (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ?
 125                                         test_vector->plaintext.data :
 126                                         test_vector->ciphertext.data;
 127
 128         mbuf = rte_pktmbuf_alloc(mempool);
 129         if (mbuf == NULL)
 130                 goto error;
 131
 132         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 133         if (mbuf_data == NULL)
 134                 goto error;
 135
 136         memcpy(mbuf_data, test_data, segment_sz);
 137         test_data += segment_sz;
 138         segments_nb--;
 139
 140         while (segments_nb) {
 141                 struct rte_mbuf *m;
 142
 143                 m = rte_pktmbuf_alloc(mempool);
 144                 if (m == NULL)
 145                         goto error;
 146
 147                 rte_pktmbuf_chain(mbuf, m);
 148
 149                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 150                 if (mbuf_data == NULL)
 151                         goto error;
 152
 153                 memcpy(mbuf_data, test_data, segment_sz);
 154                 test_data += segment_sz;
 155                 segments_nb--;
 156         }
 157
 158         if (last_sz) {
 159                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz);
 160                 if (mbuf_data == NULL)
 161                         goto error;
 162
 163                 memcpy(mbuf_data, test_data, last_sz);
 164         }
 165
 166         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf,
 167                         options->auth_digest_sz);
 168         if (mbuf_data == NULL)
 169                 goto error;
 170
 171         if (options->op_type == CPERF_AEAD) {
 172                 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf,
 173                         RTE_ALIGN_CEIL(options->auth_aad_sz, 16));
 174
 175                 if (aead == NULL)
 176                         goto error;
 177
 178                 memcpy(aead, test_vector->aad.data, test_vector->aad.length);
 179         }
 180
 181         return mbuf;
 182 error:
 183         if (mbuf != NULL)
 184                 rte_pktmbuf_free(mbuf);
 185
 186         return NULL;
 187 }
 188
 189 void *
 190 cperf_throughput_test_constructor(uint8_t dev_id, uint16_t qp_id,
 191                 const struct cperf_options *options,
 192                 const struct cperf_test_vector *test_vector,
 193                 const struct cperf_op_fns *op_fns)
 194 {
 195         struct cperf_throughput_ctx *ctx = NULL;
 196         unsigned int mbuf_idx = 0;
 197         char pool_name[32] = "";
 198
 199         ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0);
 200         if (ctx == NULL)
 201                 goto err;
 202
 203         ctx->dev_id = dev_id;
 204         ctx->qp_id = qp_id;
 205
 206         ctx->populate_ops = op_fns->populate_ops;
 207         ctx->options = options;
 208         ctx->test_vector = test_vector;
 209
 210         ctx->sess = op_fns->sess_create(dev_id, options, test_vector);
 211         if (ctx->sess == NULL)
 212                 goto err;
 213
 214         snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d",
 215                         dev_id);
 216
 217         ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name,
 218                         options->pool_sz * options->segments_nb, 0, 0,
 219                         RTE_PKTMBUF_HEADROOM +
 220                         RTE_CACHE_LINE_ROUNDUP(
 221                                 (options->buffer_sz / options->segments_nb) +
 222                                 (options->buffer_sz % options->segments_nb) +
 223                                         options->auth_digest_sz),
 224                         rte_socket_id());
 225
 226         if (ctx->pkt_mbuf_pool_in == NULL)
 227                 goto err;
 228
 229         /* Generate mbufs_in with plaintext populated for test */
 230         if (ctx->options->pool_sz % ctx->options->burst_sz)
 231                 goto err;
 232
 233         ctx->mbufs_in = rte_malloc(NULL,
 234                         (sizeof(struct rte_mbuf *) * ctx->options->pool_sz), 0);
 235
 236         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 237                 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create(
 238                                 ctx->pkt_mbuf_pool_in, options->segments_nb,
 239                                 options, test_vector);
 240                 if (ctx->mbufs_in[mbuf_idx] == NULL)
 241                         goto err;
 242         }
 243
 244         if (options->out_of_place == 1) {
 245
 246                 snprintf(pool_name, sizeof(pool_name), "cperf_pool_out_cdev_%d",
 247                                 dev_id);
 248
 249                 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create(
 250                                 pool_name, options->pool_sz, 0, 0,
 251                                 RTE_PKTMBUF_HEADROOM +
 252                                 RTE_CACHE_LINE_ROUNDUP(
 253                                         options->buffer_sz +
 254                                         options->auth_digest_sz),
 255                                 rte_socket_id());
 256
 257                 if (ctx->pkt_mbuf_pool_out == NULL)
 258                         goto err;
 259         }
 260
 261         ctx->mbufs_out = rte_malloc(NULL,
 262                         (sizeof(struct rte_mbuf *) *
 263                         ctx->options->pool_sz), 0);
 264
 265         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 266                 if (options->out_of_place == 1) {
 267                         ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create(
 268                                         ctx->pkt_mbuf_pool_out, 1,
 269                                         options, test_vector);
 270                         if (ctx->mbufs_out[mbuf_idx] == NULL)
 271                                 goto err;
 272                 } else {
 273                         ctx->mbufs_out[mbuf_idx] = NULL;
 274                 }
 275         }
 276
 277         snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d",
 278                         dev_id);
 279
 280         ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name,
 281                         RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 0, 0,
 282                         rte_socket_id());
 283         if (ctx->crypto_op_pool == NULL)
 284                 goto err;
 285
 286         return ctx;
 287 err:
 288         cperf_throughput_test_free(ctx, mbuf_idx);
 289
 290         return NULL;
 291 }
 292
 293 int
 294 cperf_throughput_test_runner(void *test_ctx)
 295 {
 296         struct cperf_throughput_ctx *ctx = test_ctx;
 297
 298         uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 299         uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 300
 301         uint64_t i, m_idx = 0, tsc_start, tsc_end, tsc_duration;
 302
 303         uint16_t ops_unused = 0;
 304
 305         struct rte_crypto_op *ops[ctx->options->burst_sz];
 306         struct rte_crypto_op *ops_processed[ctx->options->burst_sz];
 307
 308         uint32_t lcore = rte_lcore_id();
 309
 310 #ifdef CPERF_LINEARIZATION_ENABLE
 311         struct rte_cryptodev_info dev_info;
 312         int linearize = 0;
 313
 314         /* Check if source mbufs require coalescing */
 315         if (ctx->options->segments_nb > 1) {
 316                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 317                 if ((dev_info.feature_flags &
 318                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 319                         linearize = 1;
 320         }
 321 #endif /* CPERF_LINEARIZATION_ENABLE */
 322
 323         ctx->lcore_id = lcore;
 324
 325         if (!ctx->options->csv)
 326                 printf("\n# Running throughput test on device: %u, lcore: %u\n",
 327                         ctx->dev_id, lcore);
 328
 329         /* Warm up the host CPU before starting the test */
 330         for (i = 0; i < ctx->options->total_ops; i++)
 331                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 332
 333         tsc_start = rte_rdtsc_precise();
 334
 335         while (ops_enqd_total < ctx->options->total_ops) {
 336
 337                 uint16_t burst_size = ((ops_enqd_total + ctx->options->burst_sz)
 338                                 <= ctx->options->total_ops) ?
 339                                                 ctx->options->burst_sz :
 340                                                 ctx->options->total_ops -
 341                                                 ops_enqd_total;
 342
 343                 uint16_t ops_needed = burst_size - ops_unused;
 344
 345                 /* Allocate crypto ops from pool */
 346                 if (ops_needed != rte_crypto_op_bulk_alloc(
 347                                 ctx->crypto_op_pool,
 348                                 RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 349                                 ops, ops_needed))
 350                         return -1;
 351
 352                 /* Setup crypto op, attach mbuf etc */
 353                 (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 354                                 &ctx->mbufs_out[m_idx],
 355                                 ops_needed, ctx->sess, ctx->options,
 356                                 ctx->test_vector);
 357
 358 #ifdef CPERF_LINEARIZATION_ENABLE
 359                 if (linearize) {
 360                         /* PMD doesn't support scatter-gather and source buffer
 361                          * is segmented.
 362                          * We need to linearize it before enqueuing.
 363                          */
 364                         for (i = 0; i < burst_size; i++)
 365                                 rte_pktmbuf_linearize(ops[i]->sym->m_src);
 366                 }
 367 #endif /* CPERF_LINEARIZATION_ENABLE */
 368
 369                 /* Enqueue burst of ops on crypto device */
 370                 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 371                                 ops, burst_size);
 372                 if (ops_enqd < burst_size)
 373                         ops_enqd_failed++;
 374
 375                 /**
 376                  * Calculate number of ops not enqueued (mainly for hw
 377                  * accelerators whose ingress queue can fill up).
 378                  */
 379                 ops_unused = burst_size - ops_enqd;
 380                 ops_enqd_total += ops_enqd;
 381
 382
 383                 /* Dequeue processed burst of ops from crypto device */
 384                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 385                                 ops_processed, ctx->options->burst_sz);
 386
 387                 if (likely(ops_deqd))  {
 388                         /* free crypto ops so they can be reused. We don't free
 389                          * the mbufs here as we don't want to reuse them as
 390                          * the crypto operation will change the data and cause
 391                          * failures.
 392                          */
 393                         for (i = 0; i < ops_deqd; i++)
 394                                 rte_crypto_op_free(ops_processed[i]);
 395
 396                         ops_deqd_total += ops_deqd;
 397                 } else {
 398                         /**
 399                          * Count dequeue polls which didn't return any
 400                          * processed operations. This statistic is mainly
 401                          * relevant to hw accelerators.
 402                          */
 403                         ops_deqd_failed++;
 404                 }
 405
 406                 m_idx += ops_needed;
 407                 m_idx = m_idx + ctx->options->burst_sz > ctx->options->pool_sz ?
 408                                 0 : m_idx;
 409         }
 410
 411         /* Dequeue any operations still in the crypto device */
 412
 413         while (ops_deqd_total < ctx->options->total_ops) {
 414                 /* Sending 0 length burst to flush sw crypto device */
 415                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 416
 417                 /* dequeue burst */
 418                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 419                                 ops_processed, ctx->options->burst_sz);
 420                 if (ops_deqd == 0)
 421                         ops_deqd_failed++;
 422                 else {
 423                         for (i = 0; i < ops_deqd; i++)
 424                                 rte_crypto_op_free(ops_processed[i]);
 425
 426                         ops_deqd_total += ops_deqd;
 427                 }
 428         }
 429
 430         tsc_end = rte_rdtsc_precise();
 431         tsc_duration = (tsc_end - tsc_start);
 432
 433         /* Calculate average operations processed per second */
 434         ctx->results.ops_per_second = ((double)ctx->options->total_ops /
 435                         tsc_duration) * rte_get_tsc_hz();
 436
 437         /* Calculate average throughput (Gbps) in bits per second */
 438         ctx->results.throughput_gbps = ((ctx->results.ops_per_second *
 439                         ctx->options->buffer_sz * 8) / 1000000000);
 440
 441         /* Calculate average cycles per byte */
 442         ctx->results.cycles_per_byte =  ((double)tsc_duration /
 443                         ctx->options->total_ops) / ctx->options->buffer_sz;
 444
 445         ctx->results.ops_enqueued = ops_enqd_total;
 446         ctx->results.ops_dequeued = ops_deqd_total;
 447
 448         ctx->results.ops_enqueued_failed = ops_enqd_failed;
 449         ctx->results.ops_dequeued_failed = ops_deqd_failed;
 450
 451         return 0;
 452 }
 453
 454
 455 void
 456 cperf_throughput_test_destructor(void *arg)
 457 {
 458         struct cperf_throughput_ctx *ctx = arg;
 459         struct cperf_throughput_results *results = &ctx->results;
 460         static int only_once;
 461
 462         if (ctx == NULL)
 463                 return;
 464
 465         if (!ctx->options->csv) {
 466                 printf("\n# Device %d on lcore %u\n",
 467                                 ctx->dev_id, ctx->lcore_id);
 468                 printf("# Buffer Size(B)\t  Enqueued\t  Dequeued\tFailed Enq"
 469                                 "\tFailed Deq\tOps(Millions)\tThroughput(Gbps)"
 470                                 "\tCycles Per Byte\n");
 471
 472                 printf("\n%16u\t%10"PRIu64"\t%10"PRIu64"\t%10"PRIu64"\t"
 473                                 "%10"PRIu64"\t%16.4f\t%16.4f\t%15.2f\n",
 474                                 ctx->options->buffer_sz,
 475                                 results->ops_enqueued,
 476                                 results->ops_dequeued,
 477                                 results->ops_enqueued_failed,
 478                                 results->ops_dequeued_failed,
 479                                 results->ops_per_second/1000000,
 480                                 results->throughput_gbps,
 481                                 results->cycles_per_byte);
 482         } else {
 483                 if (!only_once)
 484                         printf("\n# CPU lcore id, Burst Size(B), "
 485                                 "Buffer Size(B),Enqueued,Dequeued,Failed Enq,"
 486                                 "Failed Deq,Ops(Millions),Throughput(Gbps),"
 487                                 "Cycles Per Byte\n");
 488                 only_once = 1;
 489
 490                 printf("%u;%u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";"
 491                                 "%.f3;%.f3;%.f3\n",
 492                                 ctx->lcore_id,
 493                                 ctx->options->burst_sz,
 494                                 ctx->options->buffer_sz,
 495                                 results->ops_enqueued,
 496                                 results->ops_dequeued,
 497                                 results->ops_enqueued_failed,
 498                                 results->ops_dequeued_failed,
 499                                 results->ops_per_second/1000000,
 500                                 results->throughput_gbps,
 501                                 results->cycles_per_byte);
 502         }
 503
 504         cperf_throughput_test_free(ctx, ctx->options->pool_sz);
 505 }