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                         struct rte_security_ctx *sec_ctx =
  47                                 (struct rte_security_ctx *)
  48                                         rte_cryptodev_get_sec_ctx(ctx->dev_id);
  49                         rte_security_session_destroy(
  50                                 sec_ctx,
  51                                 (struct rte_security_session *)ctx->sess);
  52                 }
  53 #endif
  54                 else {
  55                         rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
  56                         rte_cryptodev_sym_session_free(ctx->sess);
  57                 }
  58         }
  59         if (ctx->pool)
  60                 rte_mempool_free(ctx->pool);
  61
  62         rte_free(ctx);
  63 }
  64
  65 void *
  66 cperf_throughput_test_constructor(struct rte_mempool *sess_mp,
  67                 struct rte_mempool *sess_priv_mp,
  68                 uint8_t dev_id, uint16_t qp_id,
  69                 const struct cperf_options *options,
  70                 const struct cperf_test_vector *test_vector,
  71                 const struct cperf_op_fns *op_fns)
  72 {
  73         struct cperf_throughput_ctx *ctx = NULL;
  74
  75         ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0);
  76         if (ctx == NULL)
  77                 goto err;
  78
  79         ctx->dev_id = dev_id;
  80         ctx->qp_id = qp_id;
  81
  82         ctx->populate_ops = op_fns->populate_ops;
  83         ctx->options = options;
  84         ctx->test_vector = test_vector;
  85
  86         /* IV goes at the end of the crypto operation */
  87         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
  88                 sizeof(struct rte_crypto_sym_op);
  89
  90         ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
  91                         test_vector, iv_offset);
  92         if (ctx->sess == NULL)
  93                 goto err;
  94
  95         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
  96                         &ctx->src_buf_offset, &ctx->dst_buf_offset,
  97                         &ctx->pool) < 0)
  98                 goto err;
  99
 100         return ctx;
 101 err:
 102         cperf_throughput_test_free(ctx);
 103
 104         return NULL;
 105 }
 106
 107 int
 108 cperf_throughput_test_runner(void *test_ctx)
 109 {
 110         struct cperf_throughput_ctx *ctx = test_ctx;
 111         uint16_t test_burst_size;
 112         uint8_t burst_size_idx = 0;
 113         uint32_t imix_idx = 0;
 114
 115         static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
 116
 117         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 118         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 119         uint64_t i;
 120
 121         uint32_t lcore = rte_lcore_id();
 122
 123 #ifdef CPERF_LINEARIZATION_ENABLE
 124         struct rte_cryptodev_info dev_info;
 125         int linearize = 0;
 126
 127         /* Check if source mbufs require coalescing */
 128         if ((ctx->options->op_type != CPERF_ASYM_MODEX) &&
 129             (ctx->options->segment_sz < ctx->options->max_buffer_size)) {
 130                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 131                 if ((dev_info.feature_flags &
 132                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 133                         linearize = 1;
 134         }
 135 #endif /* CPERF_LINEARIZATION_ENABLE */
 136
 137         ctx->lcore_id = lcore;
 138
 139         /* Warm up the host CPU before starting the test */
 140         for (i = 0; i < ctx->options->total_ops; i++)
 141                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 142
 143         /* Get first size from range or list */
 144         if (ctx->options->inc_burst_size != 0)
 145                 test_burst_size = ctx->options->min_burst_size;
 146         else
 147                 test_burst_size = ctx->options->burst_size_list[0];
 148
 149         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 150                 sizeof(struct rte_crypto_sym_op);
 151
 152         while (test_burst_size <= ctx->options->max_burst_size) {
 153                 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0;
 154                 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0;
 155
 156                 uint64_t tsc_start, tsc_end, tsc_duration;
 157
 158                 uint16_t ops_unused = 0;
 159
 160                 tsc_start = rte_rdtsc_precise();
 161
 162                 while (ops_enqd_total < ctx->options->total_ops) {
 163
 164                         uint16_t burst_size = ((ops_enqd_total + test_burst_size)
 165                                         <= ctx->options->total_ops) ?
 166                                                         test_burst_size :
 167                                                         ctx->options->total_ops -
 168                                                         ops_enqd_total;
 169
 170                         uint16_t ops_needed = burst_size - ops_unused;
 171
 172                         /* Allocate objects containing crypto operations and mbufs */
 173                         if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
 174                                                 ops_needed) != 0) {
 175                                 RTE_LOG(ERR, USER1,
 176                                         "Failed to allocate more crypto operations "
 177                                         "from the crypto operation pool.\n"
 178                                         "Consider increasing the pool size "
 179                                         "with --pool-sz\n");
 180                                 return -1;
 181                         }
 182
 183                         /* Setup crypto op, attach mbuf etc */
 184                         (ctx->populate_ops)(ops, ctx->src_buf_offset,
 185                                         ctx->dst_buf_offset,
 186                                         ops_needed, ctx->sess,
 187                                         ctx->options, ctx->test_vector,
 188                                         iv_offset, &imix_idx);
 189
 190                         /**
 191                          * When ops_needed is smaller than ops_enqd, the
 192                          * unused ops need to be moved to the front for
 193                          * next round use.
 194                          */
 195                         if (unlikely(ops_enqd > ops_needed)) {
 196                                 size_t nb_b_to_mov = ops_unused * sizeof(
 197                                                 struct rte_crypto_op *);
 198
 199                                 memmove(&ops[ops_needed], &ops[ops_enqd],
 200                                         nb_b_to_mov);
 201                         }
 202
 203 #ifdef CPERF_LINEARIZATION_ENABLE
 204                         if (linearize) {
 205                                 /* PMD doesn't support scatter-gather and source buffer
 206                                  * is segmented.
 207                                  * We need to linearize it before enqueuing.
 208                                  */
 209                                 for (i = 0; i < burst_size; i++)
 210                                         rte_pktmbuf_linearize(
 211                                                 ops[i]->sym->m_src);
 212                         }
 213 #endif /* CPERF_LINEARIZATION_ENABLE */
 214
 215                         /* Enqueue burst of ops on crypto device */
 216                         ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
 217                                         ops, burst_size);
 218                         if (ops_enqd < burst_size)
 219                                 ops_enqd_failed++;
 220
 221                         /**
 222                          * Calculate number of ops not enqueued (mainly for hw
 223                          * accelerators whose ingress queue can fill up).
 224                          */
 225                         ops_unused = burst_size - ops_enqd;
 226                         ops_enqd_total += ops_enqd;
 227
 228
 229                         /* Dequeue processed burst of ops from crypto device */
 230                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 231                                         ops_processed, test_burst_size);
 232
 233                         if (likely(ops_deqd))  {
 234                                 /* Free crypto ops so they can be reused. */
 235                                 rte_mempool_put_bulk(ctx->pool,
 236                                                 (void **)ops_processed, ops_deqd);
 237
 238                                 ops_deqd_total += ops_deqd;
 239                         } else {
 240                                 /**
 241                                  * Count dequeue polls which didn't return any
 242                                  * processed operations. This statistic is mainly
 243                                  * relevant to hw accelerators.
 244                                  */
 245                                 ops_deqd_failed++;
 246                         }
 247
 248                 }
 249
 250                 /* Dequeue any operations still in the crypto device */
 251
 252                 while (ops_deqd_total < ctx->options->total_ops) {
 253                         /* Sending 0 length burst to flush sw crypto device */
 254                         rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 255
 256                         /* dequeue burst */
 257                         ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
 258                                         ops_processed, test_burst_size);
 259                         if (ops_deqd == 0)
 260                                 ops_deqd_failed++;
 261                         else {
 262                                 rte_mempool_put_bulk(ctx->pool,
 263                                                 (void **)ops_processed, ops_deqd);
 264                                 ops_deqd_total += ops_deqd;
 265                         }
 266                 }
 267
 268                 tsc_end = rte_rdtsc_precise();
 269                 tsc_duration = (tsc_end - tsc_start);
 270
 271                 /* Calculate average operations processed per second */
 272                 double ops_per_second = ((double)ctx->options->total_ops /
 273                                 tsc_duration) * rte_get_tsc_hz();
 274
 275                 /* Calculate average throughput (Gbps) in bits per second */
 276                 double throughput_gbps = ((ops_per_second *
 277                                 ctx->options->test_buffer_size * 8) / 1000000000);
 278
 279                 /* Calculate average cycles per packet */
 280                 double cycles_per_packet = ((double)tsc_duration /
 281                                 ctx->options->total_ops);
 282
 283                 if (!ctx->options->csv) {
 284                         if (rte_atomic16_test_and_set(&display_once))
 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 (rte_atomic16_test_and_set(&display_once))
 305                                 printf("#lcore id,Buffer Size(B),"
 306                                         "Burst Size,Enqueued,Dequeued,Failed Enq,"
 307                                         "Failed Deq,Ops(Millions),Throughput(Gbps),"
 308                                         "Cycles/Buf\n\n");
 309
 310                         printf("%u,%u,%u,%"PRIu64",%"PRIu64",%"PRIu64",%"PRIu64","
 311                                         "%.3f,%.3f,%.3f\n",
 312                                         ctx->lcore_id,
 313                                         ctx->options->test_buffer_size,
 314                                         test_burst_size,
 315                                         ops_enqd_total,
 316                                         ops_deqd_total,
 317                                         ops_enqd_failed,
 318                                         ops_deqd_failed,
 319                                         ops_per_second/1000000,
 320                                         throughput_gbps,
 321                                         cycles_per_packet);
 322                 }
 323
 324                 /* Get next size from range or list */
 325                 if (ctx->options->inc_burst_size != 0)
 326                         test_burst_size += ctx->options->inc_burst_size;
 327                 else {
 328                         if (++burst_size_idx == ctx->options->burst_size_count)
 329                                 break;
 330                         test_burst_size = ctx->options->burst_size_list[burst_size_idx];
 331                 }
 332
 333         }
 334
 335         return 0;
 336 }
 337
 338
 339 void
 340 cperf_throughput_test_destructor(void *arg)
 341 {
 342         struct cperf_throughput_ctx *ctx = arg;
 343
 344         if (ctx == NULL)
 345                 return;
 346
 347         cperf_throughput_test_free(ctx);
 348 }