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->buffer_sz / segments_nb;
 107         uint32_t last_sz = options->buffer_sz % 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         ctx->sess = op_fns->sess_create(dev_id, options, test_vector);
 199         if (ctx->sess == NULL)
 200                 goto err;
 201
 202         snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d",
 203                         dev_id);
 204
 205         ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name,
 206                         options->pool_sz * options->segments_nb, 0, 0,
 207                         RTE_PKTMBUF_HEADROOM +
 208                         RTE_CACHE_LINE_ROUNDUP(
 209                                 (options->buffer_sz / options->segments_nb) +
 210                                 (options->buffer_sz % options->segments_nb) +
 211                                         options->auth_digest_sz),
 212                         rte_socket_id());
 213
 214         if (ctx->pkt_mbuf_pool_in == NULL)
 215                 goto err;
 216
 217         /* Generate mbufs_in with plaintext populated for test */
 218         ctx->mbufs_in = rte_malloc(NULL,
 219                         (sizeof(struct rte_mbuf *) * ctx->options->pool_sz), 0);
 220
 221         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 222                 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create(
 223                                 ctx->pkt_mbuf_pool_in, options->segments_nb,
 224                                 options, test_vector);
 225                 if (ctx->mbufs_in[mbuf_idx] == NULL)
 226                         goto err;
 227         }
 228
 229         if (options->out_of_place == 1) {
 230
 231                 snprintf(pool_name, sizeof(pool_name), "cperf_pool_out_cdev_%d",
 232                                 dev_id);
 233
 234                 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create(
 235                                 pool_name, options->pool_sz, 0, 0,
 236                                 RTE_PKTMBUF_HEADROOM +
 237                                 RTE_CACHE_LINE_ROUNDUP(
 238                                         options->buffer_sz +
 239                                         options->auth_digest_sz),
 240                                 rte_socket_id());
 241
 242                 if (ctx->pkt_mbuf_pool_out == NULL)
 243                         goto err;
 244         }
 245
 246         ctx->mbufs_out = rte_malloc(NULL,
 247                         (sizeof(struct rte_mbuf *) *
 248                         ctx->options->pool_sz), 0);
 249
 250         for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) {
 251                 if (options->out_of_place == 1) {
 252                         ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create(
 253                                         ctx->pkt_mbuf_pool_out, 1,
 254                                         options, test_vector);
 255                         if (ctx->mbufs_out[mbuf_idx] == NULL)
 256                                 goto err;
 257                 } else {
 258                         ctx->mbufs_out[mbuf_idx] = NULL;
 259                 }
 260         }
 261
 262         snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d",
 263                         dev_id);
 264
 265         ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name,
 266                         RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 0, 0,
 267                         rte_socket_id());
 268         if (ctx->crypto_op_pool == NULL)
 269                 goto err;
 270
 271         return ctx;
 272 err:
 273         cperf_throughput_test_free(ctx, mbuf_idx);
 274
 275         return NULL;
 276 }
 277
 278 int
 279 cperf_throughput_test_runner(void *test_ctx)
 280 {
 281         struct cperf_throughput_ctx *ctx = test_ctx;
 282
 283         static int only_once;
 284
 285         struct rte_crypto_op *ops[ctx->options->burst_sz];
 286         struct rte_crypto_op *ops_processed[ctx->options->burst_sz];
 287         uint64_t i;
 288
 289         uint32_t lcore = rte_lcore_id();
 290
 291 #ifdef CPERF_LINEARIZATION_ENABLE
 292         struct rte_cryptodev_info dev_info;
 293         int linearize = 0;
 294
 295         /* Check if source mbufs require coalescing */
 296         if (ctx->options->segments_nb > 1) {
 297                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 298                 if ((dev_info.feature_flags &
 299                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 300                         linearize = 1;
 301         }
 302 #endif /* CPERF_LINEARIZATION_ENABLE */
 303
 304         ctx->lcore_id = lcore;
 305
 306         /* Warm up the host CPU before starting the test */
 307         for (i = 0; i < ctx->options->total_ops; i++)
 308                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 309
 310         uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 311         uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 312         uint64_t m_idx = 0, tsc_start, tsc_end, tsc_duration;
 313
 314         tsc_start = rte_rdtsc_precise();
 315         while (ops_enqd_total < ctx->options->total_ops) {
 316
 317                 uint16_t ops_unused = 0;
 318
 319                 uint16_t burst_size = ((ops_enqd_total + ctx->options->burst_sz)
 320                                 <= ctx->options->total_ops) ?
 321                                                 ctx->options->burst_sz :
 322                                                 ctx->options->total_ops -
 323                                                 ops_enqd_total;
 324
 325                 uint16_t ops_needed = burst_size - ops_unused;
 326
 327                 /* Allocate crypto ops from pool */
 328                 if (ops_needed != rte_crypto_op_bulk_alloc(
 329                                 ctx->crypto_op_pool,
 330                                 RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 331                                 ops, ops_needed))
 332                         return -1;
 333
 334                 /* Setup crypto op, attach mbuf etc */
 335                 (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 336                                 &ctx->mbufs_out[m_idx],
 337                                 ops_needed, ctx->sess, ctx->options,
 338                                 ctx->test_vector);
 339
 340 #ifdef CPERF_LINEARIZATION_ENABLE
 341                 if (linearize) {
 342                         /* PMD doesn't support scatter-gather and source buffer
 343                          * is segmented.
 344                          * We need to linearize it before enqueuing.
 345                          */
 346                         for (i = 0; i < burst_size; i++)
 347                                 rte_pktmbuf_linearize(ops[i]->sym->m_src);
 348                 }
 349 #endif /* CPERF_LINEARIZATION_ENABLE */
 350
 351                 /* Enqueue burst of ops on crypto device */
 352                 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 353                                 ops, burst_size);
 354                 if (ops_enqd < burst_size)
 355                         ops_enqd_failed++;
 356
 357                 /**
 358                  * Calculate number of ops not enqueued (mainly for hw
 359                  * accelerators whose ingress queue can fill up).
 360                  */
 361                 ops_unused = burst_size - ops_enqd;
 362                 ops_enqd_total += ops_enqd;
 363
 364
 365                 /* Dequeue processed burst of ops from crypto device */
 366                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 367                                 ops_processed, ctx->options->burst_sz);
 368
 369                 if (likely(ops_deqd))  {
 370                         /* free crypto ops so they can be reused. We don't free
 371                          * the mbufs here as we don't want to reuse them as
 372                          * the crypto operation will change the data and cause
 373                          * failures.
 374                          */
 375                         for (i = 0; i < ops_deqd; i++)
 376                                 rte_crypto_op_free(ops_processed[i]);
 377
 378                         ops_deqd_total += ops_deqd;
 379                 } else {
 380                         /**
 381                          * Count dequeue polls which didn't return any
 382                          * processed operations. This statistic is mainly
 383                          * relevant to hw accelerators.
 384                          */
 385                         ops_deqd_failed++;
 386                 }
 387
 388                 m_idx += ops_needed;
 389                 m_idx = m_idx + ctx->options->burst_sz > ctx->options->pool_sz ?
 390                                 0 : m_idx;
 391         }
 392
 393         /* Dequeue any operations still in the crypto device */
 394
 395         while (ops_deqd_total < ctx->options->total_ops) {
 396                 /* Sending 0 length burst to flush sw crypto device */
 397                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 398
 399                 /* dequeue burst */
 400                 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 401                                 ops_processed, ctx->options->burst_sz);
 402                 if (ops_deqd == 0)
 403                         ops_deqd_failed++;
 404                 else {
 405                         for (i = 0; i < ops_deqd; i++)
 406                                 rte_crypto_op_free(ops_processed[i]);
 407
 408                         ops_deqd_total += ops_deqd;
 409                 }
 410         }
 411
 412         tsc_end = rte_rdtsc_precise();
 413         tsc_duration = (tsc_end - tsc_start);
 414
 415         /* Calculate average operations processed per second */
 416         double ops_per_second = ((double)ctx->options->total_ops /
 417                         tsc_duration) * rte_get_tsc_hz();
 418
 419         /* Calculate average throughput (Gbps) in bits per second */
 420         double throughput_gbps = ((ops_per_second *
 421                         ctx->options->buffer_sz * 8) / 1000000000);
 422
 423         /* Calculate average cycles per packet */
 424         double cycles_per_packet = ((double)tsc_duration /
 425                         ctx->options->total_ops);
 426
 427         if (!ctx->options->csv) {
 428                 if (!only_once)
 429                         printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n",
 430                                 "lcore id", "Buf Size", "Burst Size",
 431                                 "Enqueued", "Dequeued", "Failed Enq",
 432                                 "Failed Deq", "MOps", "Gbps",
 433                                 "Cycles/Buf");
 434                 only_once = 1;
 435
 436                 printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64
 437                                 "%12"PRIu64"%12.4f%12.4f%12.2f\n",
 438                                 ctx->lcore_id,
 439                                 ctx->options->buffer_sz,
 440                                 ctx->options->burst_sz,
 441                                 ops_enqd_total,
 442                                 ops_deqd_total,
 443                                 ops_enqd_failed,
 444                                 ops_deqd_failed,
 445                                 ops_per_second/1000000,
 446                                 throughput_gbps,
 447                                 cycles_per_packet);
 448         } else {
 449                 if (!only_once)
 450                         printf("# lcore id, Buffer Size(B),"
 451                                 "Burst Size,Enqueued,Dequeued,Failed Enq,"
 452                                 "Failed Deq,Ops(Millions),Throughput(Gbps),"
 453                                 "Cycles/Buf\n\n");
 454                 only_once = 1;
 455
 456                 printf("%10u;%10u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";"
 457                                 "%.f3;%.f3;%.f3\n",
 458                                 ctx->lcore_id,
 459                                 ctx->options->buffer_sz,
 460                                 ctx->options->burst_sz,
 461                                 ops_enqd_total,
 462                                 ops_deqd_total,
 463                                 ops_enqd_failed,
 464                                 ops_deqd_failed,
 465                                 ops_per_second/1000000,
 466                                 throughput_gbps,
 467                                 cycles_per_packet);
 468         }
 469
 470         return 0;
 471 }
 472
 473
 474 void
 475 cperf_throughput_test_destructor(void *arg)
 476 {
 477         struct cperf_throughput_ctx *ctx = arg;
 478
 479         if (ctx == NULL)
 480                 return;
 481
 482         cperf_throughput_test_free(ctx, ctx->options->pool_sz);
 483 }