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 #ifdef RTE_LIBRTE_SECURITY
  39                 if (ctx->options->op_type == CPERF_PDCP) {
  40                         struct rte_security_ctx *sec_ctx =
  41                                 (struct rte_security_ctx *)
  42                                 rte_cryptodev_get_sec_ctx(ctx->dev_id);
  43                         rte_security_session_destroy(sec_ctx,
  44                                 (struct rte_security_session *)ctx->sess);
  45                 } else
  46 #endif
  47                 {
  48                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
  49                         rte_cryptodev_sym_session_free(ctx->sess);
  50                 }
  51         }
  52         if (ctx->pool)
  53                 rte_mempool_free(ctx->pool);
  54
  55         rte_free(ctx);
  56 }
  57
  58 void *
  59 cperf_throughput_test_constructor(struct rte_mempool *sess_mp,
  60                 struct rte_mempool *sess_priv_mp,
  61                 uint8_t dev_id, uint16_t qp_id,
  62                 const struct cperf_options *options,
  63                 const struct cperf_test_vector *test_vector,
  64                 const struct cperf_op_fns *op_fns)
  65 {
  66         struct cperf_throughput_ctx *ctx = NULL;
  67
  68         ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0);
  69         if (ctx == NULL)
  70                 goto err;
  71
  72         ctx->dev_id = dev_id;
  73         ctx->qp_id = qp_id;
  74
  75         ctx->populate_ops = op_fns->populate_ops;
  76         ctx->options = options;
  77         ctx->test_vector = test_vector;
  78
  79         /* IV goes at the end of the crypto operation */
  80         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
  81                 sizeof(struct rte_crypto_sym_op);
  82
  83         ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
  84                         test_vector, iv_offset);
  85         if (ctx->sess == NULL)
  86                 goto err;
  87
  88         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
  89                         &ctx->src_buf_offset, &ctx->dst_buf_offset,
  90                         &ctx->pool) < 0)
  91                 goto err;
  92
  93         return ctx;
  94 err:
  95         cperf_throughput_test_free(ctx);
  96
  97         return NULL;
  98 }
  99
 100 int
 101 cperf_throughput_test_runner(void *test_ctx)
 102 {
 103         struct cperf_throughput_ctx *ctx = test_ctx;
 104         uint16_t test_burst_size;
 105         uint8_t burst_size_idx = 0;
 106         uint32_t imix_idx = 0;
 107
 108         static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
 109
 110         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 111         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 112         uint64_t i;
 113
 114         uint32_t lcore = rte_lcore_id();
 115
 116 #ifdef CPERF_LINEARIZATION_ENABLE
 117         struct rte_cryptodev_info dev_info;
 118         int linearize = 0;
 119
 120         /* Check if source mbufs require coalescing */
 121         if (ctx->options->segment_sz < ctx->options->max_buffer_size) {
 122                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 123                 if ((dev_info.feature_flags &
 124                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 125                         linearize = 1;
 126         }
 127 #endif /* CPERF_LINEARIZATION_ENABLE */
 128
 129         ctx->lcore_id = lcore;
 130
 131         /* Warm up the host CPU before starting the test */
 132         for (i = 0; i < ctx->options->total_ops; i++)
 133                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 134
 135         /* Get first size from range or list */
 136         if (ctx->options->inc_burst_size != 0)
 137                 test_burst_size = ctx->options->min_burst_size;
 138         else
 139                 test_burst_size = ctx->options->burst_size_list[0];
 140
 141         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 142                 sizeof(struct rte_crypto_sym_op);
 143
 144         while (test_burst_size <= ctx->options->max_burst_size) {
 145                 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 146                 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 147
 148                 uint64_t tsc_start, tsc_end, tsc_duration;
 149
 150                 uint16_t ops_unused = 0;
 151
 152                 tsc_start = rte_rdtsc_precise();
 153
 154                 while (ops_enqd_total < ctx->options->total_ops) {
 155
 156                         uint16_t burst_size = ((ops_enqd_total + test_burst_size)
 157                                         <= ctx->options->total_ops) ?
 158                                                         test_burst_size :
 159                                                         ctx->options->total_ops -
 160                                                         ops_enqd_total;
 161
 162                         uint16_t ops_needed = burst_size - ops_unused;
 163
 164                         /* Allocate objects containing crypto operations and mbufs */
 165                         if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
 166                                                 ops_needed) != 0) {
 167                                 RTE_LOG(ERR, USER1,
 168                                         "Failed to allocate more crypto operations "
 169                                         "from the crypto operation pool.\n"
 170                                         "Consider increasing the pool size "
 171                                         "with --pool-sz\n");
 172                                 return -1;
 173                         }
 174
 175                         /* Setup crypto op, attach mbuf etc */
 176                         (ctx->populate_ops)(ops, ctx->src_buf_offset,
 177                                         ctx->dst_buf_offset,
 178                                         ops_needed, ctx->sess,
 179                                         ctx->options, ctx->test_vector,
 180                                         iv_offset, &imix_idx);
 181
 182                         /**
 183                          * When ops_needed is smaller than ops_enqd, the
 184                          * unused ops need to be moved to the front for
 185                          * next round use.
 186                          */
 187                         if (unlikely(ops_enqd > ops_needed)) {
 188                                 size_t nb_b_to_mov = ops_unused * sizeof(
 189                                                 struct rte_crypto_op *);
 190
 191                                 memmove(&ops[ops_needed], &ops[ops_enqd],
 192                                         nb_b_to_mov);
 193                         }
 194
 195 #ifdef CPERF_LINEARIZATION_ENABLE
 196                         if (linearize) {
 197                                 /* PMD doesn't support scatter-gather and source buffer
 198                                  * is segmented.
 199                                  * We need to linearize it before enqueuing.
 200                                  */
 201                                 for (i = 0; i < burst_size; i++)
 202                                         rte_pktmbuf_linearize(ops[i]->sym->m_src);
 203                         }
 204 #endif /* CPERF_LINEARIZATION_ENABLE */
 205
 206                         /* Enqueue burst of ops on crypto device */
 207                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 208                                         ops, burst_size);
 209                         if (ops_enqd < burst_size)
 210                                 ops_enqd_failed++;
 211
 212                         /**
 213                          * Calculate number of ops not enqueued (mainly for hw
 214                          * accelerators whose ingress queue can fill up).
 215                          */
 216                         ops_unused = burst_size - ops_enqd;
 217                         ops_enqd_total += ops_enqd;
 218
 219
 220                         /* Dequeue processed burst of ops from crypto device */
 221                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 222                                         ops_processed, test_burst_size);
 223
 224                         if (likely(ops_deqd))  {
 225                                 /* Free crypto ops so they can be reused. */
 226                                 rte_mempool_put_bulk(ctx->pool,
 227                                                 (void **)ops_processed, ops_deqd);
 228
 229                                 ops_deqd_total += ops_deqd;
 230                         } else {
 231                                 /**
 232                                  * Count dequeue polls which didn't return any
 233                                  * processed operations. This statistic is mainly
 234                                  * relevant to hw accelerators.
 235                                  */
 236                                 ops_deqd_failed++;
 237                         }
 238
 239                 }
 240
 241                 /* Dequeue any operations still in the crypto device */
 242
 243                 while (ops_deqd_total < ctx->options->total_ops) {
 244                         /* Sending 0 length burst to flush sw crypto device */
 245                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 246
 247                         /* dequeue burst */
 248                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 249                                         ops_processed, test_burst_size);
 250                         if (ops_deqd == 0)
 251                                 ops_deqd_failed++;
 252                         else {
 253                                 rte_mempool_put_bulk(ctx->pool,
 254                                                 (void **)ops_processed, ops_deqd);
 255                                 ops_deqd_total += ops_deqd;
 256                         }
 257                 }
 258
 259                 tsc_end = rte_rdtsc_precise();
 260                 tsc_duration = (tsc_end - tsc_start);
 261
 262                 /* Calculate average operations processed per second */
 263                 double ops_per_second = ((double)ctx->options->total_ops /
 264                                 tsc_duration) * rte_get_tsc_hz();
 265
 266                 /* Calculate average throughput (Gbps) in bits per second */
 267                 double throughput_gbps = ((ops_per_second *
 268                                 ctx->options->test_buffer_size * 8) / 1000000000);
 269
 270                 /* Calculate average cycles per packet */
 271                 double cycles_per_packet = ((double)tsc_duration /
 272                                 ctx->options->total_ops);
 273
 274                 if (!ctx->options->csv) {
 275                         if (rte_atomic16_test_and_set(&display_once))
 276                                 printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n",
 277                                         "lcore id", "Buf Size", "Burst Size",
 278                                         "Enqueued", "Dequeued", "Failed Enq",
 279                                         "Failed Deq", "MOps", "Gbps",
 280                                         "Cycles/Buf");
 281
 282                         printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64
 283                                         "%12"PRIu64"%12.4f%12.4f%12.2f\n",
 284                                         ctx->lcore_id,
 285                                         ctx->options->test_buffer_size,
 286                                         test_burst_size,
 287                                         ops_enqd_total,
 288                                         ops_deqd_total,
 289                                         ops_enqd_failed,
 290                                         ops_deqd_failed,
 291                                         ops_per_second/1000000,
 292                                         throughput_gbps,
 293                                         cycles_per_packet);
 294                 } else {
 295                         if (rte_atomic16_test_and_set(&display_once))
 296                                 printf("#lcore id,Buffer Size(B),"
 297                                         "Burst Size,Enqueued,Dequeued,Failed Enq,"
 298                                         "Failed Deq,Ops(Millions),Throughput(Gbps),"
 299                                         "Cycles/Buf\n\n");
 300
 301                         printf("%u;%u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";"
 302                                         "%.3f;%.3f;%.3f\n",
 303                                         ctx->lcore_id,
 304                                         ctx->options->test_buffer_size,
 305                                         test_burst_size,
 306                                         ops_enqd_total,
 307                                         ops_deqd_total,
 308                                         ops_enqd_failed,
 309                                         ops_deqd_failed,
 310                                         ops_per_second/1000000,
 311                                         throughput_gbps,
 312                                         cycles_per_packet);
 313                 }
 314
 315                 /* Get next size from range or list */
 316                 if (ctx->options->inc_burst_size != 0)
 317                         test_burst_size += ctx->options->inc_burst_size;
 318                 else {
 319                         if (++burst_size_idx == ctx->options->burst_size_count)
 320                                 break;
 321                         test_burst_size = ctx->options->burst_size_list[burst_size_idx];
 322                 }
 323
 324         }
 325
 326         return 0;
 327 }
 328
 329
 330 void
 331 cperf_throughput_test_destructor(void *arg)
 332 {
 333         struct cperf_throughput_ctx *ctx = arg;
 334
 335         if (ctx == NULL)
 336                 return;
 337
 338         cperf_throughput_test_free(ctx);
 339 }