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 #include "cperf_test_common.h"
  41
  42 struct cperf_throughput_ctx {
  43         uint8_t dev_id;
  44         uint16_t qp_id;
  45         uint8_t lcore_id;
  46
  47         struct rte_mempool *pkt_mbuf_pool_in;
  48         struct rte_mempool *pkt_mbuf_pool_out;
  49         struct rte_mbuf **mbufs_in;
  50         struct rte_mbuf **mbufs_out;
  51
  52         struct rte_mempool *crypto_op_pool;
  53
  54         struct rte_cryptodev_sym_session *sess;
  55
  56         cperf_populate_ops_t populate_ops;
  57
  58         const struct cperf_options *options;
  59         const struct cperf_test_vector *test_vector;
  60 };
  61
  62 static void
  63 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx)
  64 {
  65         if (ctx) {
  66                 if (ctx->sess) {
  67                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
  68                         rte_cryptodev_sym_session_free(ctx->sess);
  69                 }
  70
  71                 cperf_free_common_memory(ctx->options,
  72                                 ctx->pkt_mbuf_pool_in,
  73                                 ctx->pkt_mbuf_pool_out,
  74                                 ctx->mbufs_in, ctx->mbufs_out,
  75                                 ctx->crypto_op_pool);
  76
  77                 rte_free(ctx);
  78         }
  79 }
  80
  81 void *
  82 cperf_throughput_test_constructor(struct rte_mempool *sess_mp,
  83                 uint8_t dev_id, uint16_t qp_id,
  84                 const struct cperf_options *options,
  85                 const struct cperf_test_vector *test_vector,
  86                 const struct cperf_op_fns *op_fns)
  87 {
  88         struct cperf_throughput_ctx *ctx = NULL;
  89
  90         ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0);
  91         if (ctx == NULL)
  92                 goto err;
  93
  94         ctx->dev_id = dev_id;
  95         ctx->qp_id = qp_id;
  96
  97         ctx->populate_ops = op_fns->populate_ops;
  98         ctx->options = options;
  99         ctx->test_vector = test_vector;
 100
 101         /* IV goes at the end of the crypto operation */
 102         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 103                 sizeof(struct rte_crypto_sym_op);
 104
 105         ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, test_vector,
 106                                         iv_offset);
 107         if (ctx->sess == NULL)
 108                 goto err;
 109
 110         if (cperf_alloc_common_memory(options, test_vector, dev_id, 0,
 111                         &ctx->pkt_mbuf_pool_in, &ctx->pkt_mbuf_pool_out,
 112                         &ctx->mbufs_in, &ctx->mbufs_out,
 113                         &ctx->crypto_op_pool) < 0)
 114                 goto err;
 115
 116         return ctx;
 117 err:
 118         cperf_throughput_test_free(ctx);
 119
 120         return NULL;
 121 }
 122
 123 int
 124 cperf_throughput_test_runner(void *test_ctx)
 125 {
 126         struct cperf_throughput_ctx *ctx = test_ctx;
 127         uint16_t test_burst_size;
 128         uint8_t burst_size_idx = 0;
 129
 130         static int only_once;
 131
 132         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 133         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 134         uint64_t i;
 135
 136         uint32_t lcore = rte_lcore_id();
 137
 138 #ifdef CPERF_LINEARIZATION_ENABLE
 139         struct rte_cryptodev_info dev_info;
 140         int linearize = 0;
 141
 142         /* Check if source mbufs require coalescing */
 143         if (ctx->options->segment_sz < ctx->options->max_buffer_size) {
 144                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 145                 if ((dev_info.feature_flags &
 146                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 147                         linearize = 1;
 148         }
 149 #endif /* CPERF_LINEARIZATION_ENABLE */
 150
 151         ctx->lcore_id = lcore;
 152
 153         /* Warm up the host CPU before starting the test */
 154         for (i = 0; i < ctx->options->total_ops; i++)
 155                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 156
 157         /* Get first size from range or list */
 158         if (ctx->options->inc_burst_size != 0)
 159                 test_burst_size = ctx->options->min_burst_size;
 160         else
 161                 test_burst_size = ctx->options->burst_size_list[0];
 162
 163         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 164                 sizeof(struct rte_crypto_sym_op);
 165
 166         while (test_burst_size <= ctx->options->max_burst_size) {
 167                 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 168                 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 169
 170                 uint64_t m_idx = 0, tsc_start, tsc_end, tsc_duration;
 171
 172                 uint16_t ops_unused = 0;
 173
 174                 tsc_start = rte_rdtsc_precise();
 175
 176                 while (ops_enqd_total < ctx->options->total_ops) {
 177
 178                         uint16_t burst_size = ((ops_enqd_total + test_burst_size)
 179                                         <= ctx->options->total_ops) ?
 180                                                         test_burst_size :
 181                                                         ctx->options->total_ops -
 182                                                         ops_enqd_total;
 183
 184                         uint16_t ops_needed = burst_size - ops_unused;
 185
 186                         /* Allocate crypto ops from pool */
 187                         if (ops_needed != rte_crypto_op_bulk_alloc(
 188                                         ctx->crypto_op_pool,
 189                                         RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 190                                         ops, ops_needed)) {
 191                                 RTE_LOG(ERR, USER1,
 192                                         "Failed to allocate more crypto operations "
 193                                         "from the the crypto operation pool.\n"
 194                                         "Consider increasing the pool size "
 195                                         "with --pool-sz\n");
 196                                 return -1;
 197                         }
 198
 199                         /* Setup crypto op, attach mbuf etc */
 200                         (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 201                                         &ctx->mbufs_out[m_idx],
 202                                         ops_needed, ctx->sess, ctx->options,
 203                                         ctx->test_vector, iv_offset);
 204
 205                         /**
 206                          * When ops_needed is smaller than ops_enqd, the
 207                          * unused ops need to be moved to the front for
 208                          * next round use.
 209                          */
 210                         if (unlikely(ops_enqd > ops_needed)) {
 211                                 size_t nb_b_to_mov = ops_unused * sizeof(
 212                                                 struct rte_crypto_op *);
 213
 214                                 memmove(&ops[ops_needed], &ops[ops_enqd],
 215                                         nb_b_to_mov);
 216                         }
 217
 218 #ifdef CPERF_LINEARIZATION_ENABLE
 219                         if (linearize) {
 220                                 /* PMD doesn't support scatter-gather and source buffer
 221                                  * is segmented.
 222                                  * We need to linearize it before enqueuing.
 223                                  */
 224                                 for (i = 0; i < burst_size; i++)
 225                                         rte_pktmbuf_linearize(ops[i]->sym->m_src);
 226                         }
 227 #endif /* CPERF_LINEARIZATION_ENABLE */
 228
 229                         /* Enqueue burst of ops on crypto device */
 230                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 231                                         ops, burst_size);
 232                         if (ops_enqd < burst_size)
 233                                 ops_enqd_failed++;
 234
 235                         /**
 236                          * Calculate number of ops not enqueued (mainly for hw
 237                          * accelerators whose ingress queue can fill up).
 238                          */
 239                         ops_unused = burst_size - ops_enqd;
 240                         ops_enqd_total += ops_enqd;
 241
 242
 243                         /* Dequeue processed burst of ops from crypto device */
 244                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 245                                         ops_processed, test_burst_size);
 246
 247                         if (likely(ops_deqd))  {
 248                                 /* free crypto ops so they can be reused. We don't free
 249                                  * the mbufs here as we don't want to reuse them as
 250                                  * the crypto operation will change the data and cause
 251                                  * failures.
 252                                  */
 253                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 254                                                 (void **)ops_processed, ops_deqd);
 255
 256                                 ops_deqd_total += ops_deqd;
 257                         } else {
 258                                 /**
 259                                  * Count dequeue polls which didn't return any
 260                                  * processed operations. This statistic is mainly
 261                                  * relevant to hw accelerators.
 262                                  */
 263                                 ops_deqd_failed++;
 264                         }
 265
 266                         m_idx += ops_needed;
 267                         m_idx = m_idx + test_burst_size > ctx->options->pool_sz ?
 268                                         0 : m_idx;
 269                 }
 270
 271                 /* Dequeue any operations still in the crypto device */
 272
 273                 while (ops_deqd_total < ctx->options->total_ops) {
 274                         /* Sending 0 length burst to flush sw crypto device */
 275                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 276
 277                         /* dequeue burst */
 278                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 279                                         ops_processed, test_burst_size);
 280                         if (ops_deqd == 0)
 281                                 ops_deqd_failed++;
 282                         else {
 283                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 284                                                 (void **)ops_processed, ops_deqd);
 285
 286                                 ops_deqd_total += ops_deqd;
 287                         }
 288                 }
 289
 290                 tsc_end = rte_rdtsc_precise();
 291                 tsc_duration = (tsc_end - tsc_start);
 292
 293                 /* Calculate average operations processed per second */
 294                 double ops_per_second = ((double)ctx->options->total_ops /
 295                                 tsc_duration) * rte_get_tsc_hz();
 296
 297                 /* Calculate average throughput (Gbps) in bits per second */
 298                 double throughput_gbps = ((ops_per_second *
 299                                 ctx->options->test_buffer_size * 8) / 1000000000);
 300
 301                 /* Calculate average cycles per packet */
 302                 double cycles_per_packet = ((double)tsc_duration /
 303                                 ctx->options->total_ops);
 304
 305                 if (!ctx->options->csv) {
 306                         if (!only_once)
 307                                 printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n",
 308                                         "lcore id", "Buf Size", "Burst Size",
 309                                         "Enqueued", "Dequeued", "Failed Enq",
 310                                         "Failed Deq", "MOps", "Gbps",
 311                                         "Cycles/Buf");
 312                         only_once = 1;
 313
 314                         printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64
 315                                         "%12"PRIu64"%12.4f%12.4f%12.2f\n",
 316                                         ctx->lcore_id,
 317                                         ctx->options->test_buffer_size,
 318                                         test_burst_size,
 319                                         ops_enqd_total,
 320                                         ops_deqd_total,
 321                                         ops_enqd_failed,
 322                                         ops_deqd_failed,
 323                                         ops_per_second/1000000,
 324                                         throughput_gbps,
 325                                         cycles_per_packet);
 326                 } else {
 327                         if (!only_once)
 328                                 printf("#lcore id,Buffer Size(B),"
 329                                         "Burst Size,Enqueued,Dequeued,Failed Enq,"
 330                                         "Failed Deq,Ops(Millions),Throughput(Gbps),"
 331                                         "Cycles/Buf\n\n");
 332                         only_once = 1;
 333
 334                         printf("%u;%u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";"
 335                                         "%.3f;%.3f;%.3f\n",
 336                                         ctx->lcore_id,
 337                                         ctx->options->test_buffer_size,
 338                                         test_burst_size,
 339                                         ops_enqd_total,
 340                                         ops_deqd_total,
 341                                         ops_enqd_failed,
 342                                         ops_deqd_failed,
 343                                         ops_per_second/1000000,
 344                                         throughput_gbps,
 345                                         cycles_per_packet);
 346                 }
 347
 348                 /* Get next size from range or list */
 349                 if (ctx->options->inc_burst_size != 0)
 350                         test_burst_size += ctx->options->inc_burst_size;
 351                 else {
 352                         if (++burst_size_idx == ctx->options->burst_size_count)
 353                                 break;
 354                         test_burst_size = ctx->options->burst_size_list[burst_size_idx];
 355                 }
 356
 357         }
 358
 359         return 0;
 360 }
 361
 362
 363 void
 364 cperf_throughput_test_destructor(void *arg)
 365 {
 366         struct cperf_throughput_ctx *ctx = arg;
 367
 368         if (ctx == NULL)
 369                 return;
 370
 371         rte_cryptodev_stop(ctx->dev_id);
 372
 373         cperf_throughput_test_free(ctx);
 374 }