app/test-crypto-perf/cperf_test_throughput.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_throughput.h"
  11 #include "cperf_ops.h"
  12 #include "cperf_test_common.h"
  13
  14 struct cperf_throughput_ctx {
  15         uint8_t dev_id;
  16         uint16_t qp_id;
  17         uint8_t lcore_id;
  18
  19         struct rte_mempool *pool;
  20
  21         struct rte_cryptodev_sym_session *sess;
  22
  23         cperf_populate_ops_t populate_ops;
  24
  25         uint32_t src_buf_offset;
  26         uint32_t dst_buf_offset;
  27
  28         const struct cperf_options *options;
  29         const struct cperf_test_vector *test_vector;
  30 };
  31
  32 static void
  33 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx)
  34 {
  35         if (!ctx)
  36                 return;
  37         if (ctx->sess) {
  38                 if (ctx->options->op_type == CPERF_ASYM_MODEX)
  39                         rte_cryptodev_asym_session_free(ctx->dev_id,
  40                                         (void *)ctx->sess);
  41 #ifdef RTE_LIB_SECURITY
  42                 else if (ctx->options->op_type == CPERF_PDCP ||
  43                          ctx->options->op_type == CPERF_DOCSIS ||
  44                          ctx->options->op_type == CPERF_IPSEC) {
  45                         struct rte_security_ctx *sec_ctx =
  46                                 (struct rte_security_ctx *)
  47                                         rte_cryptodev_get_sec_ctx(ctx->dev_id);
  48                         rte_security_session_destroy(
  49                                 sec_ctx,
  50                                 (struct rte_security_session *)ctx->sess);
  51                 }
  52 #endif
  53                 else {
  54                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
  55                         rte_cryptodev_sym_session_free(ctx->sess);
  56                 }
  57         }
  58         rte_mempool_free(ctx->pool);
  59
  60         rte_free(ctx);
  61 }
  62
  63 void *
  64 cperf_throughput_test_constructor(struct rte_mempool *sess_mp,
  65                 struct rte_mempool *sess_priv_mp,
  66                 uint8_t dev_id, uint16_t qp_id,
  67                 const struct cperf_options *options,
  68                 const struct cperf_test_vector *test_vector,
  69                 const struct cperf_op_fns *op_fns)
  70 {
  71         struct cperf_throughput_ctx *ctx = NULL;
  72
  73         ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_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
  88         ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
  89                         test_vector, iv_offset);
  90         if (ctx->sess == NULL)
  91                 goto err;
  92
  93         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
  94                         &ctx->src_buf_offset, &ctx->dst_buf_offset,
  95                         &ctx->pool) < 0)
  96                 goto err;
  97
  98         return ctx;
  99 err:
 100         cperf_throughput_test_free(ctx);
 101
 102         return NULL;
 103 }
 104
 105 int
 106 cperf_throughput_test_runner(void *test_ctx)
 107 {
 108         struct cperf_throughput_ctx *ctx = test_ctx;
 109         uint16_t test_burst_size;
 110         uint8_t burst_size_idx = 0;
 111         uint32_t imix_idx = 0;
 112
 113         static uint16_t display_once;
 114
 115         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 116         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 117         uint64_t i;
 118
 119         uint32_t lcore = rte_lcore_id();
 120
 121 #ifdef CPERF_LINEARIZATION_ENABLE
 122         struct rte_cryptodev_info dev_info;
 123         int linearize = 0;
 124
 125         /* Check if source mbufs require coalescing */
 126         if ((ctx->options->op_type != CPERF_ASYM_MODEX) &&
 127             (ctx->options->segment_sz < ctx->options->max_buffer_size)) {
 128                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 129                 if ((dev_info.feature_flags &
 130                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 131                         linearize = 1;
 132         }
 133 #endif /* CPERF_LINEARIZATION_ENABLE */
 134
 135         ctx->lcore_id = lcore;
 136
 137         /* Warm up the host CPU before starting the test */
 138         for (i = 0; i < ctx->options->total_ops; i++)
 139                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 140
 141         /* Get first size from range or list */
 142         if (ctx->options->inc_burst_size != 0)
 143                 test_burst_size = ctx->options->min_burst_size;
 144         else
 145                 test_burst_size = ctx->options->burst_size_list[0];
 146
 147         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 148                 sizeof(struct rte_crypto_sym_op);
 149
 150         while (test_burst_size <= ctx->options->max_burst_size) {
 151                 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 152                 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 153
 154                 uint64_t tsc_start, tsc_end, tsc_duration;
 155
 156                 uint16_t ops_unused = 0;
 157
 158                 tsc_start = rte_rdtsc_precise();
 159
 160                 while (ops_enqd_total < ctx->options->total_ops) {
 161
 162                         uint16_t burst_size = ((ops_enqd_total + test_burst_size)
 163                                         <= ctx->options->total_ops) ?
 164                                                         test_burst_size :
 165                                                         ctx->options->total_ops -
 166                                                         ops_enqd_total;
 167
 168                         uint16_t ops_needed = burst_size - ops_unused;
 169
 170                         /* Allocate objects containing crypto operations and mbufs */
 171                         if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
 172                                                 ops_needed) != 0) {
 173                                 RTE_LOG(ERR, USER1,
 174                                         "Failed to allocate more crypto operations "
 175                                         "from the crypto operation pool.\n"
 176                                         "Consider increasing the pool size "
 177                                         "with --pool-sz\n");
 178                                 return -1;
 179                         }
 180
 181                         /* Setup crypto op, attach mbuf etc */
 182                         (ctx->populate_ops)(ops, ctx->src_buf_offset,
 183                                         ctx->dst_buf_offset,
 184                                         ops_needed, ctx->sess,
 185                                         ctx->options, ctx->test_vector,
 186                                         iv_offset, &imix_idx, &tsc_start);
 187
 188                         /**
 189                          * When ops_needed is smaller than ops_enqd, the
 190                          * unused ops need to be moved to the front for
 191                          * next round use.
 192                          */
 193                         if (unlikely(ops_enqd > ops_needed)) {
 194                                 size_t nb_b_to_mov = ops_unused * sizeof(
 195                                                 struct rte_crypto_op *);
 196
 197                                 memmove(&ops[ops_needed], &ops[ops_enqd],
 198                                         nb_b_to_mov);
 199                         }
 200
 201 #ifdef CPERF_LINEARIZATION_ENABLE
 202                         if (linearize) {
 203                                 /* PMD doesn't support scatter-gather and source buffer
 204                                  * is segmented.
 205                                  * We need to linearize it before enqueuing.
 206                                  */
 207                                 for (i = 0; i < burst_size; i++)
 208                                         rte_pktmbuf_linearize(
 209                                                 ops[i]->sym->m_src);
 210                         }
 211 #endif /* CPERF_LINEARIZATION_ENABLE */
 212
 213                         /* Enqueue burst of ops on crypto device */
 214                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 215                                         ops, burst_size);
 216                         if (ops_enqd < burst_size)
 217                                 ops_enqd_failed++;
 218
 219                         /**
 220                          * Calculate number of ops not enqueued (mainly for hw
 221                          * accelerators whose ingress queue can fill up).
 222                          */
 223                         ops_unused = burst_size - ops_enqd;
 224                         ops_enqd_total += ops_enqd;
 225
 226
 227                         /* Dequeue processed burst of ops from crypto device */
 228                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 229                                         ops_processed, test_burst_size);
 230
 231                         if (likely(ops_deqd))  {
 232                                 /* Free crypto ops so they can be reused. */
 233                                 rte_mempool_put_bulk(ctx->pool,
 234                                                 (void **)ops_processed, ops_deqd);
 235
 236                                 ops_deqd_total += ops_deqd;
 237                         } else {
 238                                 /**
 239                                  * Count dequeue polls which didn't return any
 240                                  * processed operations. This statistic is mainly
 241                                  * relevant to hw accelerators.
 242                                  */
 243                                 ops_deqd_failed++;
 244                         }
 245
 246                 }
 247
 248                 /* Dequeue any operations still in the crypto device */
 249
 250                 while (ops_deqd_total < ctx->options->total_ops) {
 251                         /* Sending 0 length burst to flush sw crypto device */
 252                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 253
 254                         /* dequeue burst */
 255                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 256                                         ops_processed, test_burst_size);
 257                         if (ops_deqd == 0)
 258                                 ops_deqd_failed++;
 259                         else {
 260                                 rte_mempool_put_bulk(ctx->pool,
 261                                                 (void **)ops_processed, ops_deqd);
 262                                 ops_deqd_total += ops_deqd;
 263                         }
 264                 }
 265
 266                 tsc_end = rte_rdtsc_precise();
 267                 tsc_duration = (tsc_end - tsc_start);
 268
 269                 /* Calculate average operations processed per second */
 270                 double ops_per_second = ((double)ctx->options->total_ops /
 271                                 tsc_duration) * rte_get_tsc_hz();
 272
 273                 /* Calculate average throughput (Gbps) in bits per second */
 274                 double throughput_gbps = ((ops_per_second *
 275                                 ctx->options->test_buffer_size * 8) / 1000000000);
 276
 277                 /* Calculate average cycles per packet */
 278                 double cycles_per_packet = ((double)tsc_duration /
 279                                 ctx->options->total_ops);
 280
 281                 uint16_t exp = 0;
 282                 if (!ctx->options->csv) {
 283                         if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0,
 284                                         __ATOMIC_RELAXED, __ATOMIC_RELAXED))
 285                                 printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n",
 286                                         "lcore id", "Buf Size", "Burst Size",
 287                                         "Enqueued", "Dequeued", "Failed Enq",
 288                                         "Failed Deq", "MOps", "Gbps",
 289                                         "Cycles/Buf");
 290
 291                         printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64
 292                                         "%12"PRIu64"%12.4f%12.4f%12.2f\n",
 293                                         ctx->lcore_id,
 294                                         ctx->options->test_buffer_size,
 295                                         test_burst_size,
 296                                         ops_enqd_total,
 297                                         ops_deqd_total,
 298                                         ops_enqd_failed,
 299                                         ops_deqd_failed,
 300                                         ops_per_second/1000000,
 301                                         throughput_gbps,
 302                                         cycles_per_packet);
 303                 } else {
 304                         if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0,
 305                                         __ATOMIC_RELAXED, __ATOMIC_RELAXED))
 306                                 printf("#lcore id,Buffer Size(B),"
 307                                         "Burst Size,Enqueued,Dequeued,Failed Enq,"
 308                                         "Failed Deq,Ops(Millions),Throughput(Gbps),"
 309                                         "Cycles/Buf\n\n");
 310
 311                         printf("%u,%u,%u,%"PRIu64",%"PRIu64",%"PRIu64",%"PRIu64","
 312                                         "%.3f,%.3f,%.3f\n",
 313                                         ctx->lcore_id,
 314                                         ctx->options->test_buffer_size,
 315                                         test_burst_size,
 316                                         ops_enqd_total,
 317                                         ops_deqd_total,
 318                                         ops_enqd_failed,
 319                                         ops_deqd_failed,
 320                                         ops_per_second/1000000,
 321                                         throughput_gbps,
 322                                         cycles_per_packet);
 323                 }
 324
 325                 /* Get next size from range or list */
 326                 if (ctx->options->inc_burst_size != 0)
 327                         test_burst_size += ctx->options->inc_burst_size;
 328                 else {
 329                         if (++burst_size_idx == ctx->options->burst_size_count)
 330                                 break;
 331                         test_burst_size = ctx->options->burst_size_list[burst_size_idx];
 332                 }
 333
 334         }
 335
 336         return 0;
 337 }
 338
 339
 340 void
 341 cperf_throughput_test_destructor(void *arg)
 342 {
 343         struct cperf_throughput_ctx *ctx = arg;
 344
 345         if (ctx == NULL)
 346                 return;
 347
 348         cperf_throughput_test_free(ctx);
 349 }