ceph/src/spdk/dpdk/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 static void
  44 cperf_latency_test_free(struct cperf_latency_ctx *ctx)
  45 {
  46         if (ctx) {
  47                 if (ctx->sess) {
  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->res);
  56                 rte_free(ctx);
  57         }
  58 }
  59
  60 void *
  61 cperf_latency_test_constructor(struct rte_mempool *sess_mp,
  62                 struct rte_mempool *sess_priv_mp,
  63                 uint8_t dev_id, uint16_t qp_id,
  64                 const struct cperf_options *options,
  65                 const struct cperf_test_vector *test_vector,
  66                 const struct cperf_op_fns *op_fns)
  67 {
  68         struct cperf_latency_ctx *ctx = NULL;
  69         size_t extra_op_priv_size = sizeof(struct priv_op_data);
  70
  71         ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
  72         if (ctx == NULL)
  73                 goto err;
  74
  75         ctx->dev_id = dev_id;
  76         ctx->qp_id = qp_id;
  77
  78         ctx->populate_ops = op_fns->populate_ops;
  79         ctx->options = options;
  80         ctx->test_vector = test_vector;
  81
  82         /* IV goes at the end of the crypto operation */
  83         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
  84                 sizeof(struct rte_crypto_sym_op) +
  85                 sizeof(struct cperf_op_result *);
  86
  87         ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
  88                         test_vector, iv_offset);
  89         if (ctx->sess == NULL)
  90                 goto err;
  91
  92         if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id,
  93                         extra_op_priv_size,
  94                         &ctx->src_buf_offset, &ctx->dst_buf_offset,
  95                         &ctx->pool) < 0)
  96                 goto err;
  97
  98         ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
  99                         ctx->options->total_ops, 0);
 100
 101         if (ctx->res == NULL)
 102                 goto err;
 103
 104         return ctx;
 105 err:
 106         cperf_latency_test_free(ctx);
 107
 108         return NULL;
 109 }
 110
 111 static inline void
 112 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp)
 113 {
 114         struct priv_op_data *priv_data;
 115
 116         priv_data = (struct priv_op_data *) (op->sym + 1);
 117         priv_data->result->status = op->status;
 118         priv_data->result->tsc_end = timestamp;
 119 }
 120
 121 int
 122 cperf_latency_test_runner(void *arg)
 123 {
 124         struct cperf_latency_ctx *ctx = arg;
 125         uint16_t test_burst_size;
 126         uint8_t burst_size_idx = 0;
 127         uint32_t imix_idx = 0;
 128
 129         static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
 130
 131         if (ctx == NULL)
 132                 return 0;
 133
 134         struct rte_crypto_op *ops[ctx->options->max_burst_size];
 135         struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
 136         uint64_t i;
 137         struct priv_op_data *priv_data;
 138
 139         uint32_t lcore = rte_lcore_id();
 140
 141 #ifdef CPERF_LINEARIZATION_ENABLE
 142         struct rte_cryptodev_info dev_info;
 143         int linearize = 0;
 144
 145         /* Check if source mbufs require coalescing */
 146         if (ctx->options->segment_sz < ctx->options->max_buffer_size) {
 147                 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
 148                 if ((dev_info.feature_flags &
 149                                 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
 150                         linearize = 1;
 151         }
 152 #endif /* CPERF_LINEARIZATION_ENABLE */
 153
 154         ctx->lcore_id = lcore;
 155
 156         /* Warm up the host CPU before starting the test */
 157         for (i = 0; i < ctx->options->total_ops; i++)
 158                 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
 159
 160         /* Get first size from range or list */
 161         if (ctx->options->inc_burst_size != 0)
 162                 test_burst_size = ctx->options->min_burst_size;
 163         else
 164                 test_burst_size = ctx->options->burst_size_list[0];
 165
 166         uint16_t iv_offset = sizeof(struct rte_crypto_op) +
 167                 sizeof(struct rte_crypto_sym_op) +
 168                 sizeof(struct cperf_op_result *);
 169
 170         while (test_burst_size <= ctx->options->max_burst_size) {
 171                 uint64_t ops_enqd = 0, ops_deqd = 0;
 172                 uint64_t b_idx = 0;
 173
 174                 uint64_t tsc_val, tsc_end, tsc_start;
 175                 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
 176                 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
 177                 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
 178
 179                 while (enqd_tot < ctx->options->total_ops) {
 180
 181                         uint16_t burst_size = ((enqd_tot + test_burst_size)
 182                                         <= ctx->options->total_ops) ?
 183                                                         test_burst_size :
 184                                                         ctx->options->total_ops -
 185                                                         enqd_tot;
 186
 187                         /* Allocate objects containing crypto operations and mbufs */
 188                         if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
 189                                                 burst_size) != 0) {
 190                                 RTE_LOG(ERR, USER1,
 191                                         "Failed to allocate more crypto operations "
 192                                         "from the crypto operation pool.\n"
 193                                         "Consider increasing the pool size "
 194                                         "with --pool-sz\n");
 195                                 return -1;
 196                         }
 197
 198                         /* Setup crypto op, attach mbuf etc */
 199                         (ctx->populate_ops)(ops, ctx->src_buf_offset,
 200                                         ctx->dst_buf_offset,
 201                                         burst_size, ctx->sess, ctx->options,
 202                                         ctx->test_vector, iv_offset,
 203                                         &imix_idx);
 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 = RTE_MAX(ops_deqd, deqd_max);
 255                                 deqd_min = RTE_MIN(ops_deqd, deqd_min);
 256                         }
 257
 258                         enqd_tot += ops_enqd;
 259                         enqd_max = RTE_MAX(ops_enqd, enqd_max);
 260                         enqd_min = RTE_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 = RTE_MAX(ops_deqd, deqd_max);
 285                                 deqd_min = RTE_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 = RTE_MAX(tsc_val, tsc_max);
 292                         tsc_min = RTE_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 (rte_atomic16_test_and_set(&display_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                 } else {
 327                         printf("\n# Device %d on lcore %u\n", ctx->dev_id,
 328                                 ctx->lcore_id);
 329                         printf("\n# total operations: %u", ctx->options->total_ops);
 330                         printf("\n# Buffer size: %u", ctx->options->test_buffer_size);
 331                         printf("\n# Burst size: %u", test_burst_size);
 332                         printf("\n#     Number of bursts: %"PRIu64,
 333                                         b_idx);
 334
 335                         printf("\n#");
 336                         printf("\n#          \t       Total\t   Average\t   "
 337                                         "Maximum\t   Minimum");
 338                         printf("\n#  enqueued\t%12"PRIu64"\t%10"PRIu64"\t"
 339                                         "%10"PRIu64"\t%10"PRIu64, enqd_tot,
 340                                         enqd_avg, enqd_max, enqd_min);
 341                         printf("\n#  dequeued\t%12"PRIu64"\t%10"PRIu64"\t"
 342                                         "%10"PRIu64"\t%10"PRIu64, deqd_tot,
 343                                         deqd_avg, deqd_max, deqd_min);
 344                         printf("\n#    cycles\t%12"PRIu64"\t%10"PRIu64"\t"
 345                                         "%10"PRIu64"\t%10"PRIu64, tsc_tot,
 346                                         tsc_avg, tsc_max, tsc_min);
 347                         printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f",
 348                                         time_tot, time_avg, time_max, time_min);
 349                         printf("\n\n");
 350
 351                 }
 352
 353                 /* Get next size from range or list */
 354                 if (ctx->options->inc_burst_size != 0)
 355                         test_burst_size += ctx->options->inc_burst_size;
 356                 else {
 357                         if (++burst_size_idx == ctx->options->burst_size_count)
 358                                 break;
 359                         test_burst_size =
 360                                 ctx->options->burst_size_list[burst_size_idx];
 361                 }
 362         }
 363
 364         return 0;
 365 }
 366
 367 void
 368 cperf_latency_test_destructor(void *arg)
 369 {
 370         struct cperf_latency_ctx *ctx = arg;
 371
 372         if (ctx == NULL)
 373                 return;
 374
 375         cperf_latency_test_free(ctx);
 376 }