app/test-crypto-perf/cperf_test_latency.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2017 Intel Corporation
   3  */
   4
   5 #include <rte_malloc.h>
   6 #include <rte_cycles.h>
   7 #include <rte_crypto.h>
   8 #include <rte_cryptodev.h>
   9
  10 #include "cperf_test_latency.h"
  11 #include "cperf_ops.h"
  12 #include "cperf_test_common.h"
  13
  14 struct cperf_op_result {
  15         uint64_t tsc_start;
  16         uint64_t tsc_end;
  17         enum rte_crypto_op_status status;
  18 };
  19
  20 struct cperf_latency_ctx {
  21         uint8_t dev_id;
  22         uint16_t qp_id;
  23         uint8_t lcore_id;
  24
  25         struct rte_mempool *pool;
  26
  27         struct rte_cryptodev_sym_session *sess;
  28
  29         cperf_populate_ops_t populate_ops;
  30
  31         uint32_t src_buf_offset;
  32         uint32_t dst_buf_offset;
  33
  34         const struct cperf_options *options;
  35         const struct cperf_test_vector *test_vector;
  36         struct cperf_op_result *res;
  37 };
  38
  39 struct priv_op_data {
  40         struct cperf_op_result *result;
  41 };
  42
  43 #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b)
  44 #define min(a, b) (a < b ? (uint64_t)a : (uint64_t)b)
  45
  46 static void
  47 cperf_latency_test_free(struct cperf_latency_ctx *ctx)
  48 {
  49         if (ctx) {
  50                 if (ctx->sess) {
  51                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
  52                         rte_cryptodev_sym_session_free(ctx->sess);
  53                 }
  54
  55                 if (ctx->pool)
  56                         rte_mempool_free(ctx->pool);
  57
  58                 rte_free(ctx->res);
  59                 rte_free(ctx);
  60         }
  61 }
  62
  63 void *
  64 cperf_latency_test_constructor(struct rte_mempool *sess_mp,
  65                 uint8_t dev_id, uint16_t qp_id,
  66                 const struct cperf_options *options,
  67                 const struct cperf_test_vector *test_vector,
  68                 const struct cperf_op_fns *op_fns)
  69 {
  70         struct cperf_latency_ctx *ctx = NULL;
  71         size_t extra_op_priv_size = sizeof(struct priv_op_data);
  72
  73         ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
  74         if (ctx == NULL)
  75                 goto err;
  76
  77         ctx->dev_id = dev_id;
  78         ctx->qp_id = qp_id;
  79
  80         ctx->populate_ops = op_fns->populate_ops;
  81         ctx->options = options;
  82         ctx->test_vector = test_vector;
  83
  84         /* IV goes at the end of the crypto operation */
  85         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
  86                 sizeof(struct rte_crypto_sym_op) +
  87                 sizeof(struct cperf_op_result *);
  88
  89         ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, test_vector,
  90                         iv_offset);
  91         if (ctx->sess == NULL)
  92                 goto err;
  93
  94         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id,
  95                         extra_op_priv_size,
  96                         &ctx->src_buf_offset, &ctx->dst_buf_offset,
  97                         &ctx->pool) < 0)
  98                 goto err;
  99
 100         ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
 101                         ctx->options->total_ops, 0);
 102
 103         if (ctx->res == NULL)
 104                 goto err;
 105
 106         return ctx;
 107 err:
 108         cperf_latency_test_free(ctx);
 109
 110         return NULL;
 111 }
 112
 113 static inline void
 114 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp)
 115 {
 116         struct priv_op_data *priv_data;
 117
 118         priv_data = (struct priv_op_data *) (op->sym + 1);
 119         priv_data->result->status = op->status;
 120         priv_data->result->tsc_end = timestamp;
 121 }
 122
 123 int
 124 cperf_latency_test_runner(void *arg)
 125 {
 126         struct cperf_latency_ctx *ctx = arg;
 127         uint16_t test_burst_size;
 128         uint8_t burst_size_idx = 0;
 129
 130         static int only_once;
 131
 132         if (ctx == NULL)
 133                 return 0;
 134
 135         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 136         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 137         uint64_t i;
 138         struct priv_op_data *priv_data;
 139
 140         uint32_t lcore = rte_lcore_id();
 141
 142 #ifdef CPERF_LINEARIZATION_ENABLE
 143         struct rte_cryptodev_info dev_info;
 144         int linearize = 0;
 145
 146         /* Check if source mbufs require coalescing */
 147         if (ctx->options->segment_sz < ctx->options->max_buffer_size) {
 148                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 149                 if ((dev_info.feature_flags &
 150                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 151                         linearize = 1;
 152         }
 153 #endif /* CPERF_LINEARIZATION_ENABLE */
 154
 155         ctx->lcore_id = lcore;
 156
 157         /* Warm up the host CPU before starting the test */
 158         for (i = 0; i < ctx->options->total_ops; i++)
 159                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 160
 161         /* Get first size from range or list */
 162         if (ctx->options->inc_burst_size != 0)
 163                 test_burst_size = ctx->options->min_burst_size;
 164         else
 165                 test_burst_size = ctx->options->burst_size_list[0];
 166
 167         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 168                 sizeof(struct rte_crypto_sym_op) +
 169                 sizeof(struct cperf_op_result *);
 170
 171         while (test_burst_size <= ctx->options->max_burst_size) {
 172                 uint64_t ops_enqd = 0, ops_deqd = 0;
 173                 uint64_t b_idx = 0;
 174
 175                 uint64_t tsc_val, tsc_end, tsc_start;
 176                 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
 177                 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
 178                 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
 179
 180                 while (enqd_tot < ctx->options->total_ops) {
 181
 182                         uint16_t burst_size = ((enqd_tot + test_burst_size)
 183                                         <= ctx->options->total_ops) ?
 184                                                         test_burst_size :
 185                                                         ctx->options->total_ops -
 186                                                         enqd_tot;
 187
 188                         /* Allocate objects containing crypto operations and mbufs */
 189                         if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
 190                                                 burst_size) != 0) {
 191                                 RTE_LOG(ERR, USER1,
 192                                         "Failed to allocate more crypto operations "
 193                                         "from 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->src_buf_offset,
 201                                         ctx->dst_buf_offset,
 202                                         burst_size, ctx->sess, ctx->options,
 203                                         ctx->test_vector, iv_offset);
 204
 205                         tsc_start = rte_rdtsc_precise();
 206
 207 #ifdef CPERF_LINEARIZATION_ENABLE
 208                         if (linearize) {
 209                                 /* PMD doesn't support scatter-gather and source buffer
 210                                  * is segmented.
 211                                  * We need to linearize it before enqueuing.
 212                                  */
 213                                 for (i = 0; i < burst_size; i++)
 214                                         rte_pktmbuf_linearize(ops[i]->sym->m_src);
 215                         }
 216 #endif /* CPERF_LINEARIZATION_ENABLE */
 217
 218                         /* Enqueue burst of ops on crypto device */
 219                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 220                                         ops, burst_size);
 221
 222                         /* Dequeue processed burst of ops from crypto device */
 223                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 224                                         ops_processed, test_burst_size);
 225
 226                         tsc_end = rte_rdtsc_precise();
 227
 228                         /* Free memory for not enqueued operations */
 229                         if (ops_enqd != burst_size)
 230                                 rte_mempool_put_bulk(ctx->pool,
 231                                                 (void **)&ops[ops_enqd],
 232                                                 burst_size - ops_enqd);
 233
 234                         for (i = 0; i < ops_enqd; i++) {
 235                                 ctx->res[tsc_idx].tsc_start = tsc_start;
 236                                 /*
 237                                  * Private data structure starts after the end of the
 238                                  * rte_crypto_sym_op structure.
 239                                  */
 240                                 priv_data = (struct priv_op_data *) (ops[i]->sym + 1);
 241                                 priv_data->result = (void *)&ctx->res[tsc_idx];
 242                                 tsc_idx++;
 243                         }
 244
 245                         if (likely(ops_deqd))  {
 246                                 /* Free crypto ops so they can be reused. */
 247                                 for (i = 0; i < ops_deqd; i++)
 248                                         store_timestamp(ops_processed[i], tsc_end);
 249
 250                                 rte_mempool_put_bulk(ctx->pool,
 251                                                 (void **)ops_processed, ops_deqd);
 252
 253                                 deqd_tot += ops_deqd;
 254                                 deqd_max = max(ops_deqd, deqd_max);
 255                                 deqd_min = min(ops_deqd, deqd_min);
 256                         }
 257
 258                         enqd_tot += ops_enqd;
 259                         enqd_max = max(ops_enqd, enqd_max);
 260                         enqd_min = min(ops_enqd, enqd_min);
 261
 262                         b_idx++;
 263                 }
 264
 265                 /* Dequeue any operations still in the crypto device */
 266                 while (deqd_tot < ctx->options->total_ops) {
 267                         /* Sending 0 length burst to flush sw crypto device */
 268                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 269
 270                         /* dequeue burst */
 271                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 272                                         ops_processed, test_burst_size);
 273
 274                         tsc_end = rte_rdtsc_precise();
 275
 276                         if (ops_deqd != 0) {
 277                                 for (i = 0; i < ops_deqd; i++)
 278                                         store_timestamp(ops_processed[i], tsc_end);
 279
 280                                 rte_mempool_put_bulk(ctx->pool,
 281                                                 (void **)ops_processed, ops_deqd);
 282
 283                                 deqd_tot += ops_deqd;
 284                                 deqd_max = max(ops_deqd, deqd_max);
 285                                 deqd_min = min(ops_deqd, deqd_min);
 286                         }
 287                 }
 288
 289                 for (i = 0; i < tsc_idx; i++) {
 290                         tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start;
 291                         tsc_max = max(tsc_val, tsc_max);
 292                         tsc_min = min(tsc_val, tsc_min);
 293                         tsc_tot += tsc_val;
 294                 }
 295
 296                 double time_tot, time_avg, time_max, time_min;
 297
 298                 const uint64_t tunit = 1000000; /* us */
 299                 const uint64_t tsc_hz = rte_get_tsc_hz();
 300
 301                 uint64_t enqd_avg = enqd_tot / b_idx;
 302                 uint64_t deqd_avg = deqd_tot / b_idx;
 303                 uint64_t tsc_avg = tsc_tot / tsc_idx;
 304
 305                 time_tot = tunit*(double)(tsc_tot) / tsc_hz;
 306                 time_avg = tunit*(double)(tsc_avg) / tsc_hz;
 307                 time_max = tunit*(double)(tsc_max) / tsc_hz;
 308                 time_min = tunit*(double)(tsc_min) / tsc_hz;
 309
 310                 if (ctx->options->csv) {
 311                         if (!only_once)
 312                                 printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, "
 313                                                 "Packet Size, cycles, time (us)");
 314
 315                         for (i = 0; i < ctx->options->total_ops; i++) {
 316
 317                                 printf("\n%u;%u;%u;%"PRIu64";%"PRIu64";%.3f",
 318                                         ctx->lcore_id, ctx->options->test_buffer_size,
 319                                         test_burst_size, i + 1,
 320                                         ctx->res[i].tsc_end - ctx->res[i].tsc_start,
 321                                         tunit * (double) (ctx->res[i].tsc_end
 322                                                         - ctx->res[i].tsc_start)
 323                                                 / tsc_hz);
 324
 325                         }
 326                         only_once = 1;
 327                 } else {
 328                         printf("\n# Device %d on lcore %u\n", ctx->dev_id,
 329                                 ctx->lcore_id);
 330                         printf("\n# total operations: %u", ctx->options->total_ops);
 331                         printf("\n# Buffer size: %u", ctx->options->test_buffer_size);
 332                         printf("\n# Burst size: %u", test_burst_size);
 333                         printf("\n#     Number of bursts: %"PRIu64,
 334                                         b_idx);
 335
 336                         printf("\n#");
 337                         printf("\n#          \t       Total\t   Average\t   "
 338                                         "Maximum\t   Minimum");
 339                         printf("\n#  enqueued\t%12"PRIu64"\t%10"PRIu64"\t"
 340                                         "%10"PRIu64"\t%10"PRIu64, enqd_tot,
 341                                         enqd_avg, enqd_max, enqd_min);
 342                         printf("\n#  dequeued\t%12"PRIu64"\t%10"PRIu64"\t"
 343                                         "%10"PRIu64"\t%10"PRIu64, deqd_tot,
 344                                         deqd_avg, deqd_max, deqd_min);
 345                         printf("\n#    cycles\t%12"PRIu64"\t%10"PRIu64"\t"
 346                                         "%10"PRIu64"\t%10"PRIu64, tsc_tot,
 347                                         tsc_avg, tsc_max, tsc_min);
 348                         printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f",
 349                                         time_tot, time_avg, time_max, time_min);
 350                         printf("\n\n");
 351
 352                 }
 353
 354                 /* Get next size from range or list */
 355                 if (ctx->options->inc_burst_size != 0)
 356                         test_burst_size += ctx->options->inc_burst_size;
 357                 else {
 358                         if (++burst_size_idx == ctx->options->burst_size_count)
 359                                 break;
 360                         test_burst_size =
 361                                 ctx->options->burst_size_list[burst_size_idx];
 362                 }
 363         }
 364
 365         return 0;
 366 }
 367
 368 void
 369 cperf_latency_test_destructor(void *arg)
 370 {
 371         struct cperf_latency_ctx *ctx = arg;
 372
 373         if (ctx == NULL)
 374                 return;
 375
 376         cperf_latency_test_free(ctx);
 377 }