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_ctx {
  42         uint8_t dev_id;
  43         uint16_t qp_id;
  44         uint8_t lcore_id;
  45
  46         struct rte_mempool *pkt_mbuf_pool_in;
  47         struct rte_mempool *pkt_mbuf_pool_out;
  48         struct rte_mbuf **mbufs_in;
  49         struct rte_mbuf **mbufs_out;
  50
  51         struct rte_mempool *crypto_op_pool;
  52
  53         struct rte_cryptodev_sym_session *sess;
  54
  55         cperf_populate_ops_t populate_ops;
  56
  57         const struct cperf_options *options;
  58         const struct cperf_test_vector *test_vector;
  59 };
  60
  61 static void
  62 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx, uint32_t mbuf_nb)
  63 {
  64         uint32_t i;
  65
  66         if (ctx) {
  67                 if (ctx->sess)
  68                         rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
  69
  70                 if (ctx->mbufs_in) {
  71                         for (i = 0; i < mbuf_nb; i++)
  72                                 rte_pktmbuf_free(ctx->mbufs_in[i]);
  73
  74                         rte_free(ctx->mbufs_in);
  75                 }
  76
  77                 if (ctx->mbufs_out) {
  78                         for (i = 0; i < mbuf_nb; i++) {
  79                                 if (ctx->mbufs_out[i] != NULL)
  80                                         rte_pktmbuf_free(ctx->mbufs_out[i]);
  81                         }
  82
  83                         rte_free(ctx->mbufs_out);
  84                 }
  85
  86                 if (ctx->pkt_mbuf_pool_in)
  87                         rte_mempool_free(ctx->pkt_mbuf_pool_in);
  88
  89                 if (ctx->pkt_mbuf_pool_out)
  90                         rte_mempool_free(ctx->pkt_mbuf_pool_out);
  91
  92                 if (ctx->crypto_op_pool)
  93                         rte_mempool_free(ctx->crypto_op_pool);
  94
  95                 rte_free(ctx);
  96         }
  97 }
  98
  99 static struct rte_mbuf *
 100 cperf_mbuf_create(struct rte_mempool *mempool,
 101                 uint32_t segments_nb,
 102                 const struct cperf_options *options,
 103                 const struct cperf_test_vector *test_vector)
 104 {
 105         struct rte_mbuf *mbuf;
 106         uint32_t segment_sz = options->max_buffer_size / segments_nb;
 107         uint32_t last_sz = options->max_buffer_size % segments_nb;
 108         uint8_t *mbuf_data;
 109         uint8_t *test_data =
 110                         (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ?
 111                                         test_vector->plaintext.data :
 112                                         test_vector->ciphertext.data;
 113
 114         mbuf = rte_pktmbuf_alloc(mempool);
 115         if (mbuf == NULL)
 116                 goto error;
 117
 118         mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 119         if (mbuf_data == NULL)
 120                 goto error;
 121
 122         memcpy(mbuf_data, test_data, segment_sz);
 123         test_data += segment_sz;
 124         segments_nb--;
 125
 126         while (segments_nb) {
 127                 struct rte_mbuf *m;
 128
 129                 m = rte_pktmbuf_alloc(mempool);
 130                 if (m == NULL)
 131                         goto error;
 132
 133                 rte_pktmbuf_chain(mbuf, m);
 134
 135                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz);
 136                 if (mbuf_data == NULL)
 137                         goto error;
 138
 139                 memcpy(mbuf_data, test_data, segment_sz);
 140                 test_data += segment_sz;
 141                 segments_nb--;
 142         }
 143
 144         if (last_sz) {
 145                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz);
 146                 if (mbuf_data == NULL)
 147                         goto error;
 148
 149                 memcpy(mbuf_data, test_data, last_sz);
 150         }
 151
 152         if (options->op_type != CPERF_CIPHER_ONLY) {
 153                 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf,
 154                                 options->auth_digest_sz);
 155                 if (mbuf_data == NULL)
 156                         goto error;
 157         }
 158
 159         if (options->op_type == CPERF_AEAD) {
 160                 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf,
 161                         RTE_ALIGN_CEIL(options->auth_aad_sz, 16));
 162
 163                 if (aead == NULL)
 164                         goto error;
 165
 166                 memcpy(aead, test_vector->aad.data, test_vector->aad.length);
 167         }
 168
 169         return mbuf;
 170 error:
 171         if (mbuf != NULL)
 172                 rte_pktmbuf_free(mbuf);
 173
 174         return NULL;
 175 }
 176
 177 void *
 178 cperf_throughput_test_constructor(uint8_t dev_id, uint16_t qp_id,
 179                 const struct cperf_options *options,
 180                 const struct cperf_test_vector *test_vector,
 181                 const struct cperf_op_fns *op_fns)
 182 {
 183         struct cperf_throughput_ctx *ctx = NULL;
 184         unsigned int mbuf_idx = 0;
 185         char pool_name[32] = "";
 186
 187         ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0);
 188         if (ctx == NULL)
 189                 goto err;
 190
 191         ctx->dev_id = dev_id;
 192         ctx->qp_id = qp_id;
 193
 194         ctx->populate_ops = op_fns->populate_ops;
 195         ctx->options = options;
 196         ctx->test_vector = test_vector;
 197
 198         /* IV goes at the end of the cryptop operation */
 199         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 200                 sizeof(struct rte_crypto_sym_op);
 201
 202         ctx->sess = op_fns->sess_create(dev_id, options, test_vector, iv_offset);
 203         if (ctx->sess == NULL)
 204                 goto err;
 205
 206         snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d",
 207                         dev_id);
 208
 209         ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name,
 210                         options->pool_sz * options->segments_nb, 0, 0,
 211                         RTE_PKTMBUF_HEADROOM +
 212                         RTE_CACHE_LINE_ROUNDUP(
 213                                 (options->max_buffer_size / options->segments_nb) +
 214                                 (options->max_buffer_size % options->segments_nb) +
 215                                         options->auth_digest_sz),
 216                         rte_socket_id());
 217
 218         if (ctx->pkt_mbuf_pool_in == NULL)
 219                 goto err;
 220
 221         /* Generate mbufs_in with plaintext populated for test */
 222         ctx->mbufs_in = rte_malloc(NULL,
 223                         (sizeof(struct rte_mbuf *) * ctx->options->pool_sz), 0);
 224
 225         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 226                 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create(
 227                                 ctx->pkt_mbuf_pool_in, options->segments_nb,
 228                                 options, test_vector);
 229                 if (ctx->mbufs_in[mbuf_idx] == NULL)
 230                         goto err;
 231         }
 232
 233         if (options->out_of_place == 1) {
 234
 235                 snprintf(pool_name, sizeof(pool_name), "cperf_pool_out_cdev_%d",
 236                                 dev_id);
 237
 238                 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create(
 239                                 pool_name, options->pool_sz, 0, 0,
 240                                 RTE_PKTMBUF_HEADROOM +
 241                                 RTE_CACHE_LINE_ROUNDUP(
 242                                         options->max_buffer_size +
 243                                         options->auth_digest_sz),
 244                                 rte_socket_id());
 245
 246                 if (ctx->pkt_mbuf_pool_out == NULL)
 247                         goto err;
 248         }
 249
 250         ctx->mbufs_out = rte_malloc(NULL,
 251                         (sizeof(struct rte_mbuf *) *
 252                         ctx->options->pool_sz), 0);
 253
 254         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 255                 if (options->out_of_place == 1) {
 256                         ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create(
 257                                         ctx->pkt_mbuf_pool_out, 1,
 258                                         options, test_vector);
 259                         if (ctx->mbufs_out[mbuf_idx] == NULL)
 260                                 goto err;
 261                 } else {
 262                         ctx->mbufs_out[mbuf_idx] = NULL;
 263                 }
 264         }
 265
 266         snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d",
 267                         dev_id);
 268
 269         uint16_t priv_size = test_vector->iv.length;
 270
 271         ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name,
 272                         RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz,
 273                         512, priv_size, rte_socket_id());
 274         if (ctx->crypto_op_pool == NULL)
 275                 goto err;
 276
 277         return ctx;
 278 err:
 279         cperf_throughput_test_free(ctx, mbuf_idx);
 280
 281         return NULL;
 282 }
 283
 284 int
 285 cperf_throughput_test_runner(void *test_ctx)
 286 {
 287         struct cperf_throughput_ctx *ctx = test_ctx;
 288         uint16_t test_burst_size;
 289         uint8_t burst_size_idx = 0;
 290
 291         static int only_once;
 292
 293         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 294         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 295         uint64_t i;
 296
 297         uint32_t lcore = rte_lcore_id();
 298
 299 #ifdef CPERF_LINEARIZATION_ENABLE
 300         struct rte_cryptodev_info dev_info;
 301         int linearize = 0;
 302
 303         /* Check if source mbufs require coalescing */
 304         if (ctx->options->segments_nb > 1) {
 305                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 306                 if ((dev_info.feature_flags &
 307                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 308                         linearize = 1;
 309         }
 310 #endif /* CPERF_LINEARIZATION_ENABLE */
 311
 312         ctx->lcore_id = lcore;
 313
 314         /* Warm up the host CPU before starting the test */
 315         for (i = 0; i < ctx->options->total_ops; i++)
 316                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 317
 318         /* Get first size from range or list */
 319         if (ctx->options->inc_burst_size != 0)
 320                 test_burst_size = ctx->options->min_burst_size;
 321         else
 322                 test_burst_size = ctx->options->burst_size_list[0];
 323
 324         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 325                 sizeof(struct rte_crypto_sym_op);
 326
 327         while (test_burst_size <= ctx->options->max_burst_size) {
 328                 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 329                 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 330
 331                 uint64_t m_idx = 0, tsc_start, tsc_end, tsc_duration;
 332
 333                 uint16_t ops_unused = 0;
 334
 335                 tsc_start = rte_rdtsc_precise();
 336
 337                 while (ops_enqd_total < ctx->options->total_ops) {
 338
 339                         uint16_t burst_size = ((ops_enqd_total + test_burst_size)
 340                                         <= ctx->options->total_ops) ?
 341                                                         test_burst_size :
 342                                                         ctx->options->total_ops -
 343                                                         ops_enqd_total;
 344
 345                         uint16_t ops_needed = burst_size - ops_unused;
 346
 347                         /* Allocate crypto ops from pool */
 348                         if (ops_needed != rte_crypto_op_bulk_alloc(
 349                                         ctx->crypto_op_pool,
 350                                         RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 351                                         ops, ops_needed))
 352                                 return -1;
 353
 354                         /* Setup crypto op, attach mbuf etc */
 355                         (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 356                                         &ctx->mbufs_out[m_idx],
 357                                         ops_needed, ctx->sess, ctx->options,
 358                                         ctx->test_vector, iv_offset);
 359
 360                         /**
 361                          * When ops_needed is smaller than ops_enqd, the
 362                          * unused ops need to be moved to the front for
 363                          * next round use.
 364                          */
 365                         if (unlikely(ops_enqd > ops_needed)) {
 366                                 size_t nb_b_to_mov = ops_unused * sizeof(
 367                                                 struct rte_crypto_op *);
 368
 369                                 memmove(&ops[ops_needed], &ops[ops_enqd],
 370                                         nb_b_to_mov);
 371                         }
 372
 373 #ifdef CPERF_LINEARIZATION_ENABLE
 374                         if (linearize) {
 375                                 /* PMD doesn't support scatter-gather and source buffer
 376                                  * is segmented.
 377                                  * We need to linearize it before enqueuing.
 378                                  */
 379                                 for (i = 0; i < burst_size; i++)
 380                                         rte_pktmbuf_linearize(ops[i]->sym->m_src);
 381                         }
 382 #endif /* CPERF_LINEARIZATION_ENABLE */
 383
 384                         /* Enqueue burst of ops on crypto device */
 385                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 386                                         ops, burst_size);
 387                         if (ops_enqd < burst_size)
 388                                 ops_enqd_failed++;
 389
 390                         /**
 391                          * Calculate number of ops not enqueued (mainly for hw
 392                          * accelerators whose ingress queue can fill up).
 393                          */
 394                         ops_unused = burst_size - ops_enqd;
 395                         ops_enqd_total += ops_enqd;
 396
 397
 398                         /* Dequeue processed burst of ops from crypto device */
 399                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 400                                         ops_processed, test_burst_size);
 401
 402                         if (likely(ops_deqd))  {
 403                                 /* free crypto ops so they can be reused. We don't free
 404                                  * the mbufs here as we don't want to reuse them as
 405                                  * the crypto operation will change the data and cause
 406                                  * failures.
 407                                  */
 408                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 409                                                 (void **)ops_processed, ops_deqd);
 410
 411                                 ops_deqd_total += ops_deqd;
 412                         } else {
 413                                 /**
 414                                  * Count dequeue polls which didn't return any
 415                                  * processed operations. This statistic is mainly
 416                                  * relevant to hw accelerators.
 417                                  */
 418                                 ops_deqd_failed++;
 419                         }
 420
 421                         m_idx += ops_needed;
 422                         m_idx = m_idx + test_burst_size > ctx->options->pool_sz ?
 423                                         0 : m_idx;
 424                 }
 425
 426                 /* Dequeue any operations still in the crypto device */
 427
 428                 while (ops_deqd_total < ctx->options->total_ops) {
 429                         /* Sending 0 length burst to flush sw crypto device */
 430                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 431
 432                         /* dequeue burst */
 433                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 434                                         ops_processed, test_burst_size);
 435                         if (ops_deqd == 0)
 436                                 ops_deqd_failed++;
 437                         else {
 438                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 439                                                 (void **)ops_processed, ops_deqd);
 440
 441                                 ops_deqd_total += ops_deqd;
 442                         }
 443                 }
 444
 445                 tsc_end = rte_rdtsc_precise();
 446                 tsc_duration = (tsc_end - tsc_start);
 447
 448                 /* Calculate average operations processed per second */
 449                 double ops_per_second = ((double)ctx->options->total_ops /
 450                                 tsc_duration) * rte_get_tsc_hz();
 451
 452                 /* Calculate average throughput (Gbps) in bits per second */
 453                 double throughput_gbps = ((ops_per_second *
 454                                 ctx->options->test_buffer_size * 8) / 1000000000);
 455
 456                 /* Calculate average cycles per packet */
 457                 double cycles_per_packet = ((double)tsc_duration /
 458                                 ctx->options->total_ops);
 459
 460                 if (!ctx->options->csv) {
 461                         if (!only_once)
 462                                 printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n",
 463                                         "lcore id", "Buf Size", "Burst Size",
 464                                         "Enqueued", "Dequeued", "Failed Enq",
 465                                         "Failed Deq", "MOps", "Gbps",
 466                                         "Cycles/Buf");
 467                         only_once = 1;
 468
 469                         printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64
 470                                         "%12"PRIu64"%12.4f%12.4f%12.2f\n",
 471                                         ctx->lcore_id,
 472                                         ctx->options->test_buffer_size,
 473                                         test_burst_size,
 474                                         ops_enqd_total,
 475                                         ops_deqd_total,
 476                                         ops_enqd_failed,
 477                                         ops_deqd_failed,
 478                                         ops_per_second/1000000,
 479                                         throughput_gbps,
 480                                         cycles_per_packet);
 481                 } else {
 482                         if (!only_once)
 483                                 printf("# lcore id, Buffer Size(B),"
 484                                         "Burst Size,Enqueued,Dequeued,Failed Enq,"
 485                                         "Failed Deq,Ops(Millions),Throughput(Gbps),"
 486                                         "Cycles/Buf\n\n");
 487                         only_once = 1;
 488
 489                         printf("%10u;%10u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";"
 490                                         "%.f3;%.f3;%.f3\n",
 491                                         ctx->lcore_id,
 492                                         ctx->options->test_buffer_size,
 493                                         test_burst_size,
 494                                         ops_enqd_total,
 495                                         ops_deqd_total,
 496                                         ops_enqd_failed,
 497                                         ops_deqd_failed,
 498                                         ops_per_second/1000000,
 499                                         throughput_gbps,
 500                                         cycles_per_packet);
 501                 }
 502
 503                 /* Get next size from range or list */
 504                 if (ctx->options->inc_burst_size != 0)
 505                         test_burst_size += ctx->options->inc_burst_size;
 506                 else {
 507                         if (++burst_size_idx == ctx->options->burst_size_count)
 508                                 break;
 509                         test_burst_size = ctx->options->burst_size_list[burst_size_idx];
 510                 }
 511
 512         }
 513
 514         return 0;
 515 }
 516
 517
 518 void
 519 cperf_throughput_test_destructor(void *arg)
 520 {
 521         struct cperf_throughput_ctx *ctx = arg;
 522
 523         if (ctx == NULL)
 524                 return;
 525
 526         cperf_throughput_test_free(ctx, ctx->options->pool_sz);
 527 }