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