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 void *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 == NULL) 47 return; 48 49 if (ctx->sess != NULL) { 50 if (ctx->options->op_type == CPERF_ASYM_MODEX) 51 rte_cryptodev_asym_session_free(ctx->dev_id, ctx->sess); 52 #ifdef RTE_LIB_SECURITY 53 else if (ctx->options->op_type == CPERF_PDCP || 54 ctx->options->op_type == CPERF_DOCSIS || 55 ctx->options->op_type == CPERF_IPSEC) { 56 struct rte_security_ctx *sec_ctx = 57 rte_cryptodev_get_sec_ctx(ctx->dev_id); 58 rte_security_session_destroy(sec_ctx, ctx->sess); 59 } 60 #endif 61 else 62 rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess); 63 } 64 65 rte_mempool_free(ctx->pool); 66 rte_free(ctx->res); 67 rte_free(ctx); 68 } 69 70 void * 71 cperf_latency_test_constructor(struct rte_mempool *sess_mp, 72 uint8_t dev_id, uint16_t qp_id, 73 const struct cperf_options *options, 74 const struct cperf_test_vector *test_vector, 75 const struct cperf_op_fns *op_fns) 76 { 77 struct cperf_latency_ctx *ctx = NULL; 78 size_t extra_op_priv_size = sizeof(struct priv_op_data); 79 80 ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0); 81 if (ctx == NULL) 82 goto err; 83 84 ctx->dev_id = dev_id; 85 ctx->qp_id = qp_id; 86 87 ctx->populate_ops = op_fns->populate_ops; 88 ctx->options = options; 89 ctx->test_vector = test_vector; 90 91 /* IV goes at the end of the crypto operation */ 92 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 93 sizeof(struct rte_crypto_sym_op) + 94 sizeof(struct cperf_op_result *); 95 96 ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, 97 test_vector, iv_offset); 98 if (ctx->sess == NULL) 99 goto err; 100 101 if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 102 extra_op_priv_size, 103 &ctx->src_buf_offset, &ctx->dst_buf_offset, 104 &ctx->pool) < 0) 105 goto err; 106 107 ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) * 108 ctx->options->total_ops, 0); 109 110 if (ctx->res == NULL) 111 goto err; 112 113 return ctx; 114 err: 115 cperf_latency_test_free(ctx); 116 117 return NULL; 118 } 119 120 static inline void 121 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp) 122 { 123 struct priv_op_data *priv_data; 124 125 priv_data = (struct priv_op_data *) (op->sym + 1); 126 priv_data->result->status = op->status; 127 priv_data->result->tsc_end = timestamp; 128 } 129 130 int 131 cperf_latency_test_runner(void *arg) 132 { 133 struct cperf_latency_ctx *ctx = arg; 134 uint16_t test_burst_size; 135 uint8_t burst_size_idx = 0; 136 uint32_t imix_idx = 0; 137 138 static uint16_t display_once; 139 140 if (ctx == NULL) 141 return 0; 142 143 struct rte_crypto_op *ops[ctx->options->max_burst_size]; 144 struct rte_crypto_op *ops_processed[ctx->options->max_burst_size]; 145 uint64_t i; 146 struct priv_op_data *priv_data; 147 148 uint32_t lcore = rte_lcore_id(); 149 150 #ifdef CPERF_LINEARIZATION_ENABLE 151 struct rte_cryptodev_info dev_info; 152 int linearize = 0; 153 154 /* Check if source mbufs require coalescing */ 155 if (ctx->options->segment_sz < ctx->options->max_buffer_size) { 156 rte_cryptodev_info_get(ctx->dev_id, &dev_info); 157 if ((dev_info.feature_flags & 158 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0) 159 linearize = 1; 160 } 161 #endif /* CPERF_LINEARIZATION_ENABLE */ 162 163 ctx->lcore_id = lcore; 164 165 /* Warm up the host CPU before starting the test */ 166 for (i = 0; i < ctx->options->total_ops; i++) 167 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 168 169 /* Get first size from range or list */ 170 if (ctx->options->inc_burst_size != 0) 171 test_burst_size = ctx->options->min_burst_size; 172 else 173 test_burst_size = ctx->options->burst_size_list[0]; 174 175 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 176 sizeof(struct rte_crypto_sym_op) + 177 sizeof(struct cperf_op_result *); 178 179 while (test_burst_size <= ctx->options->max_burst_size) { 180 uint64_t ops_enqd = 0, ops_deqd = 0; 181 uint64_t b_idx = 0; 182 183 uint64_t tsc_val, tsc_end, tsc_start; 184 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0; 185 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0; 186 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0; 187 188 while (enqd_tot < ctx->options->total_ops) { 189 190 uint16_t burst_size = ((enqd_tot + test_burst_size) 191 <= ctx->options->total_ops) ? 192 test_burst_size : 193 ctx->options->total_ops - 194 enqd_tot; 195 196 /* Allocate objects containing crypto operations and mbufs */ 197 if (rte_mempool_get_bulk(ctx->pool, (void **)ops, 198 burst_size) != 0) { 199 RTE_LOG(ERR, USER1, 200 "Failed to allocate more crypto operations " 201 "from the crypto operation pool.\n" 202 "Consider increasing the pool size " 203 "with --pool-sz\n"); 204 return -1; 205 } 206 207 /* Setup crypto op, attach mbuf etc */ 208 (ctx->populate_ops)(ops, ctx->src_buf_offset, 209 ctx->dst_buf_offset, 210 burst_size, ctx->sess, ctx->options, 211 ctx->test_vector, iv_offset, 212 &imix_idx, &tsc_start); 213 214 /* Populate the mbuf with the test vector */ 215 for (i = 0; i < burst_size; i++) 216 cperf_mbuf_set(ops[i]->sym->m_src, 217 ctx->options, 218 ctx->test_vector); 219 220 tsc_start = rte_rdtsc_precise(); 221 222 #ifdef CPERF_LINEARIZATION_ENABLE 223 if (linearize) { 224 /* PMD doesn't support scatter-gather and source buffer 225 * is segmented. 226 * We need to linearize it before enqueuing. 227 */ 228 for (i = 0; i < burst_size; i++) 229 rte_pktmbuf_linearize(ops[i]->sym->m_src); 230 } 231 #endif /* CPERF_LINEARIZATION_ENABLE */ 232 233 /* Enqueue burst of ops on crypto device */ 234 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, 235 ops, burst_size); 236 237 /* Dequeue processed burst of ops from crypto device */ 238 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 239 ops_processed, test_burst_size); 240 241 tsc_end = rte_rdtsc_precise(); 242 243 /* Free memory for not enqueued operations */ 244 if (ops_enqd != burst_size) 245 rte_mempool_put_bulk(ctx->pool, 246 (void **)&ops[ops_enqd], 247 burst_size - ops_enqd); 248 249 for (i = 0; i < ops_enqd; i++) { 250 ctx->res[tsc_idx].tsc_start = tsc_start; 251 /* 252 * Private data structure starts after the end of the 253 * rte_crypto_sym_op structure. 254 */ 255 priv_data = (struct priv_op_data *) (ops[i]->sym + 1); 256 priv_data->result = (void *)&ctx->res[tsc_idx]; 257 tsc_idx++; 258 } 259 260 if (likely(ops_deqd)) { 261 /* Free crypto ops so they can be reused. */ 262 for (i = 0; i < ops_deqd; i++) 263 store_timestamp(ops_processed[i], tsc_end); 264 265 rte_mempool_put_bulk(ctx->pool, 266 (void **)ops_processed, ops_deqd); 267 268 deqd_tot += ops_deqd; 269 deqd_max = RTE_MAX(ops_deqd, deqd_max); 270 deqd_min = RTE_MIN(ops_deqd, deqd_min); 271 } 272 273 enqd_tot += ops_enqd; 274 enqd_max = RTE_MAX(ops_enqd, enqd_max); 275 enqd_min = RTE_MIN(ops_enqd, enqd_min); 276 277 b_idx++; 278 } 279 280 /* Dequeue any operations still in the crypto device */ 281 while (deqd_tot < ctx->options->total_ops) { 282 /* Sending 0 length burst to flush sw crypto device */ 283 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 284 285 /* dequeue burst */ 286 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 287 ops_processed, test_burst_size); 288 289 tsc_end = rte_rdtsc_precise(); 290 291 if (ops_deqd != 0) { 292 for (i = 0; i < ops_deqd; i++) 293 store_timestamp(ops_processed[i], tsc_end); 294 295 rte_mempool_put_bulk(ctx->pool, 296 (void **)ops_processed, ops_deqd); 297 298 deqd_tot += ops_deqd; 299 deqd_max = RTE_MAX(ops_deqd, deqd_max); 300 deqd_min = RTE_MIN(ops_deqd, deqd_min); 301 } 302 } 303 304 for (i = 0; i < tsc_idx; i++) { 305 tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start; 306 tsc_max = RTE_MAX(tsc_val, tsc_max); 307 tsc_min = RTE_MIN(tsc_val, tsc_min); 308 tsc_tot += tsc_val; 309 } 310 311 double time_tot, time_avg, time_max, time_min; 312 313 const uint64_t tunit = 1000000; /* us */ 314 const uint64_t tsc_hz = rte_get_tsc_hz(); 315 316 uint64_t enqd_avg = enqd_tot / b_idx; 317 uint64_t deqd_avg = deqd_tot / b_idx; 318 uint64_t tsc_avg = tsc_tot / tsc_idx; 319 320 time_tot = tunit*(double)(tsc_tot) / tsc_hz; 321 time_avg = tunit*(double)(tsc_avg) / tsc_hz; 322 time_max = tunit*(double)(tsc_max) / tsc_hz; 323 time_min = tunit*(double)(tsc_min) / tsc_hz; 324 325 uint16_t exp = 0; 326 if (ctx->options->csv) { 327 if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0, 328 __ATOMIC_RELAXED, __ATOMIC_RELAXED)) 329 printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, " 330 "cycles, time (us)"); 331 332 for (i = 0; i < ctx->options->total_ops; i++) { 333 334 printf("\n%u,%u,%u,%"PRIu64",%"PRIu64",%.3f", 335 ctx->lcore_id, ctx->options->test_buffer_size, 336 test_burst_size, i + 1, 337 ctx->res[i].tsc_end - ctx->res[i].tsc_start, 338 tunit * (double) (ctx->res[i].tsc_end 339 - ctx->res[i].tsc_start) 340 / tsc_hz); 341 342 } 343 } else { 344 printf("\n# Device %d on lcore %u\n", ctx->dev_id, 345 ctx->lcore_id); 346 printf("\n# total operations: %u", ctx->options->total_ops); 347 printf("\n# Buffer size: %u", ctx->options->test_buffer_size); 348 printf("\n# Burst size: %u", test_burst_size); 349 printf("\n# Number of bursts: %"PRIu64, 350 b_idx); 351 352 printf("\n#"); 353 printf("\n# \t Total\t Average\t " 354 "Maximum\t Minimum"); 355 printf("\n# enqueued\t%12"PRIu64"\t%10"PRIu64"\t" 356 "%10"PRIu64"\t%10"PRIu64, enqd_tot, 357 enqd_avg, enqd_max, enqd_min); 358 printf("\n# dequeued\t%12"PRIu64"\t%10"PRIu64"\t" 359 "%10"PRIu64"\t%10"PRIu64, deqd_tot, 360 deqd_avg, deqd_max, deqd_min); 361 printf("\n# cycles\t%12"PRIu64"\t%10"PRIu64"\t" 362 "%10"PRIu64"\t%10"PRIu64, tsc_tot, 363 tsc_avg, tsc_max, tsc_min); 364 printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f", 365 time_tot, time_avg, time_max, time_min); 366 printf("\n\n"); 367 368 } 369 370 /* Get next size from range or list */ 371 if (ctx->options->inc_burst_size != 0) 372 test_burst_size += ctx->options->inc_burst_size; 373 else { 374 if (++burst_size_idx == ctx->options->burst_size_count) 375 break; 376 test_burst_size = 377 ctx->options->burst_size_list[burst_size_idx]; 378 } 379 } 380 381 return 0; 382 } 383 384 void 385 cperf_latency_test_destructor(void *arg) 386 { 387 struct cperf_latency_ctx *ctx = arg; 388 389 if (ctx == NULL) 390 return; 391 392 cperf_latency_test_free(ctx); 393 } 394