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 #include "cperf_test_common.h"
  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)
  78 {
  79         if (ctx) {
  80                 if (ctx->sess) {
  81                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
  82                         rte_cryptodev_sym_session_free(ctx->sess);
  83                 }
  84
  85                 cperf_free_common_memory(ctx->options,
  86                                 ctx->pkt_mbuf_pool_in,
  87                                 ctx->pkt_mbuf_pool_out,
  88                                 ctx->mbufs_in, ctx->mbufs_out,
  89                                 ctx->crypto_op_pool);
  90
  91                 rte_free(ctx->res);
  92                 rte_free(ctx);
  93         }
  94 }
  95
  96 void *
  97 cperf_latency_test_constructor(struct rte_mempool *sess_mp,
  98                 uint8_t dev_id, uint16_t qp_id,
  99                 const struct cperf_options *options,
 100                 const struct cperf_test_vector *test_vector,
 101                 const struct cperf_op_fns *op_fns)
 102 {
 103         struct cperf_latency_ctx *ctx = NULL;
 104         size_t extra_op_priv_size = sizeof(struct priv_op_data);
 105
 106         ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
 107         if (ctx == NULL)
 108                 goto err;
 109
 110         ctx->dev_id = dev_id;
 111         ctx->qp_id = qp_id;
 112
 113         ctx->populate_ops = op_fns->populate_ops;
 114         ctx->options = options;
 115         ctx->test_vector = test_vector;
 116
 117         /* IV goes at the end of the crypto operation */
 118         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 119                 sizeof(struct rte_crypto_sym_op) +
 120                 sizeof(struct cperf_op_result *);
 121
 122         ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, test_vector,
 123                         iv_offset);
 124         if (ctx->sess == NULL)
 125                 goto err;
 126
 127         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id,
 128                         extra_op_priv_size,
 129                         &ctx->pkt_mbuf_pool_in, &ctx->pkt_mbuf_pool_out,
 130                         &ctx->mbufs_in, &ctx->mbufs_out,
 131                         &ctx->crypto_op_pool) < 0)
 132                 goto err;
 133
 134         ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
 135                         ctx->options->total_ops, 0);
 136
 137         if (ctx->res == NULL)
 138                 goto err;
 139
 140         return ctx;
 141 err:
 142         cperf_latency_test_free(ctx);
 143
 144         return NULL;
 145 }
 146
 147 static inline void
 148 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp)
 149 {
 150         struct priv_op_data *priv_data;
 151
 152         priv_data = (struct priv_op_data *) (op->sym + 1);
 153         priv_data->result->status = op->status;
 154         priv_data->result->tsc_end = timestamp;
 155 }
 156
 157 int
 158 cperf_latency_test_runner(void *arg)
 159 {
 160         struct cperf_latency_ctx *ctx = arg;
 161         uint16_t test_burst_size;
 162         uint8_t burst_size_idx = 0;
 163
 164         static int only_once;
 165
 166         if (ctx == NULL)
 167                 return 0;
 168
 169         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 170         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 171         uint64_t i;
 172         struct priv_op_data *priv_data;
 173
 174         uint32_t lcore = rte_lcore_id();
 175
 176 #ifdef CPERF_LINEARIZATION_ENABLE
 177         struct rte_cryptodev_info dev_info;
 178         int linearize = 0;
 179
 180         /* Check if source mbufs require coalescing */
 181         if (ctx->options->segment_sz < ctx->options->max_buffer_size) {
 182                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 183                 if ((dev_info.feature_flags &
 184                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 185                         linearize = 1;
 186         }
 187 #endif /* CPERF_LINEARIZATION_ENABLE */
 188
 189         ctx->lcore_id = lcore;
 190
 191         /* Warm up the host CPU before starting the test */
 192         for (i = 0; i < ctx->options->total_ops; i++)
 193                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 194
 195         /* Get first size from range or list */
 196         if (ctx->options->inc_burst_size != 0)
 197                 test_burst_size = ctx->options->min_burst_size;
 198         else
 199                 test_burst_size = ctx->options->burst_size_list[0];
 200
 201         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 202                 sizeof(struct rte_crypto_sym_op) +
 203                 sizeof(struct cperf_op_result *);
 204
 205         while (test_burst_size <= ctx->options->max_burst_size) {
 206                 uint64_t ops_enqd = 0, ops_deqd = 0;
 207                 uint64_t m_idx = 0, b_idx = 0;
 208
 209                 uint64_t tsc_val, tsc_end, tsc_start;
 210                 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
 211                 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
 212                 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
 213
 214                 while (enqd_tot < ctx->options->total_ops) {
 215
 216                         uint16_t burst_size = ((enqd_tot + test_burst_size)
 217                                         <= ctx->options->total_ops) ?
 218                                                         test_burst_size :
 219                                                         ctx->options->total_ops -
 220                                                         enqd_tot;
 221
 222                         /* Allocate crypto ops from pool */
 223                         if (burst_size != rte_crypto_op_bulk_alloc(
 224                                         ctx->crypto_op_pool,
 225                                         RTE_CRYPTO_OP_TYPE_SYMMETRIC,
 226                                         ops, burst_size)) {
 227                                 RTE_LOG(ERR, USER1,
 228                                         "Failed to allocate more crypto operations "
 229                                         "from the the crypto operation pool.\n"
 230                                         "Consider increasing the pool size "
 231                                         "with --pool-sz\n");
 232                                 return -1;
 233                         }
 234
 235                         /* Setup crypto op, attach mbuf etc */
 236                         (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx],
 237                                         &ctx->mbufs_out[m_idx],
 238                                         burst_size, ctx->sess, ctx->options,
 239                                         ctx->test_vector, iv_offset);
 240
 241                         tsc_start = rte_rdtsc_precise();
 242
 243 #ifdef CPERF_LINEARIZATION_ENABLE
 244                         if (linearize) {
 245                                 /* PMD doesn't support scatter-gather and source buffer
 246                                  * is segmented.
 247                                  * We need to linearize it before enqueuing.
 248                                  */
 249                                 for (i = 0; i < burst_size; i++)
 250                                         rte_pktmbuf_linearize(ops[i]->sym->m_src);
 251                         }
 252 #endif /* CPERF_LINEARIZATION_ENABLE */
 253
 254                         /* Enqueue burst of ops on crypto device */
 255                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 256                                         ops, burst_size);
 257
 258                         /* Dequeue processed burst of ops from crypto device */
 259                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 260                                         ops_processed, test_burst_size);
 261
 262                         tsc_end = rte_rdtsc_precise();
 263
 264                         /* Free memory for not enqueued operations */
 265                         if (ops_enqd != burst_size)
 266                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 267                                                 (void **)&ops[ops_enqd],
 268                                                 burst_size - ops_enqd);
 269
 270                         for (i = 0; i < ops_enqd; i++) {
 271                                 ctx->res[tsc_idx].tsc_start = tsc_start;
 272                                 /*
 273                                  * Private data structure starts after the end of the
 274                                  * rte_crypto_sym_op structure.
 275                                  */
 276                                 priv_data = (struct priv_op_data *) (ops[i]->sym + 1);
 277                                 priv_data->result = (void *)&ctx->res[tsc_idx];
 278                                 tsc_idx++;
 279                         }
 280
 281                         if (likely(ops_deqd))  {
 282                                 /*
 283                                  * free crypto ops so they can be reused. We don't free
 284                                  * the mbufs here as we don't want to reuse them as
 285                                  * the crypto operation will change the data and cause
 286                                  * failures.
 287                                  */
 288                                 for (i = 0; i < ops_deqd; i++)
 289                                         store_timestamp(ops_processed[i], tsc_end);
 290
 291                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 292                                                 (void **)ops_processed, ops_deqd);
 293
 294                                 deqd_tot += ops_deqd;
 295                                 deqd_max = max(ops_deqd, deqd_max);
 296                                 deqd_min = min(ops_deqd, deqd_min);
 297                         }
 298
 299                         enqd_tot += ops_enqd;
 300                         enqd_max = max(ops_enqd, enqd_max);
 301                         enqd_min = min(ops_enqd, enqd_min);
 302
 303                         m_idx += ops_enqd;
 304                         m_idx = m_idx + test_burst_size > ctx->options->pool_sz ?
 305                                         0 : m_idx;
 306                         b_idx++;
 307                 }
 308
 309                 /* Dequeue any operations still in the crypto device */
 310                 while (deqd_tot < ctx->options->total_ops) {
 311                         /* Sending 0 length burst to flush sw crypto device */
 312                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 313
 314                         /* dequeue burst */
 315                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 316                                         ops_processed, test_burst_size);
 317
 318                         tsc_end = rte_rdtsc_precise();
 319
 320                         if (ops_deqd != 0) {
 321                                 for (i = 0; i < ops_deqd; i++)
 322                                         store_timestamp(ops_processed[i], tsc_end);
 323
 324                                 rte_mempool_put_bulk(ctx->crypto_op_pool,
 325                                                 (void **)ops_processed, ops_deqd);
 326
 327                                 deqd_tot += ops_deqd;
 328                                 deqd_max = max(ops_deqd, deqd_max);
 329                                 deqd_min = min(ops_deqd, deqd_min);
 330                         }
 331                 }
 332
 333                 for (i = 0; i < tsc_idx; i++) {
 334                         tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start;
 335                         tsc_max = max(tsc_val, tsc_max);
 336                         tsc_min = min(tsc_val, tsc_min);
 337                         tsc_tot += tsc_val;
 338                 }
 339
 340                 double time_tot, time_avg, time_max, time_min;
 341
 342                 const uint64_t tunit = 1000000; /* us */
 343                 const uint64_t tsc_hz = rte_get_tsc_hz();
 344
 345                 uint64_t enqd_avg = enqd_tot / b_idx;
 346                 uint64_t deqd_avg = deqd_tot / b_idx;
 347                 uint64_t tsc_avg = tsc_tot / tsc_idx;
 348
 349                 time_tot = tunit*(double)(tsc_tot) / tsc_hz;
 350                 time_avg = tunit*(double)(tsc_avg) / tsc_hz;
 351                 time_max = tunit*(double)(tsc_max) / tsc_hz;
 352                 time_min = tunit*(double)(tsc_min) / tsc_hz;
 353
 354                 if (ctx->options->csv) {
 355                         if (!only_once)
 356                                 printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, "
 357                                                 "Packet Size, cycles, time (us)");
 358
 359                         for (i = 0; i < ctx->options->total_ops; i++) {
 360
 361                                 printf("\n%u;%u;%u;%"PRIu64";%"PRIu64";%.3f",
 362                                         ctx->lcore_id, ctx->options->test_buffer_size,
 363                                         test_burst_size, i + 1,
 364                                         ctx->res[i].tsc_end - ctx->res[i].tsc_start,
 365                                         tunit * (double) (ctx->res[i].tsc_end
 366                                                         - ctx->res[i].tsc_start)
 367                                                 / tsc_hz);
 368
 369                         }
 370                         only_once = 1;
 371                 } else {
 372                         printf("\n# Device %d on lcore %u\n", ctx->dev_id,
 373                                 ctx->lcore_id);
 374                         printf("\n# total operations: %u", ctx->options->total_ops);
 375                         printf("\n# Buffer size: %u", ctx->options->test_buffer_size);
 376                         printf("\n# Burst size: %u", test_burst_size);
 377                         printf("\n#     Number of bursts: %"PRIu64,
 378                                         b_idx);
 379
 380                         printf("\n#");
 381                         printf("\n#          \t       Total\t   Average\t   "
 382                                         "Maximum\t   Minimum");
 383                         printf("\n#  enqueued\t%12"PRIu64"\t%10"PRIu64"\t"
 384                                         "%10"PRIu64"\t%10"PRIu64, enqd_tot,
 385                                         enqd_avg, enqd_max, enqd_min);
 386                         printf("\n#  dequeued\t%12"PRIu64"\t%10"PRIu64"\t"
 387                                         "%10"PRIu64"\t%10"PRIu64, deqd_tot,
 388                                         deqd_avg, deqd_max, deqd_min);
 389                         printf("\n#    cycles\t%12"PRIu64"\t%10"PRIu64"\t"
 390                                         "%10"PRIu64"\t%10"PRIu64, tsc_tot,
 391                                         tsc_avg, tsc_max, tsc_min);
 392                         printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f",
 393                                         time_tot, time_avg, time_max, time_min);
 394                         printf("\n\n");
 395
 396                 }
 397
 398                 /* Get next size from range or list */
 399                 if (ctx->options->inc_burst_size != 0)
 400                         test_burst_size += ctx->options->inc_burst_size;
 401                 else {
 402                         if (++burst_size_idx == ctx->options->burst_size_count)
 403                                 break;
 404                         test_burst_size =
 405                                 ctx->options->burst_size_list[burst_size_idx];
 406                 }
 407         }
 408
 409         return 0;
 410 }
 411
 412 void
 413 cperf_latency_test_destructor(void *arg)
 414 {
 415         struct cperf_latency_ctx *ctx = arg;
 416
 417         if (ctx == NULL)
 418                 return;
 419
 420         cperf_latency_test_free(ctx);
 421 }