1 /*- 2 * BSD LICENSE 3 * 4 * Copyright(c) 2016-2017 Intel Corporation. All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 10 * * Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * * Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in 14 * the documentation and/or other materials provided with the 15 * distribution. 16 * * Neither the name of Intel Corporation nor the names of its 17 * contributors may be used to endorse or promote products derived 18 * from this software without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 23 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 24 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 25 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 26 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 27 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 28 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 29 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 30 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 31 */ 32 33 #include <rte_malloc.h> 34 #include <rte_cycles.h> 35 #include <rte_crypto.h> 36 #include <rte_cryptodev.h> 37 38 #include "cperf_test_latency.h" 39 #include "cperf_ops.h" 40 41 42 struct cperf_op_result { 43 uint64_t tsc_start; 44 uint64_t tsc_end; 45 enum rte_crypto_op_status status; 46 }; 47 48 struct cperf_latency_ctx { 49 uint8_t dev_id; 50 uint16_t qp_id; 51 uint8_t lcore_id; 52 53 struct rte_mempool *pkt_mbuf_pool_in; 54 struct rte_mempool *pkt_mbuf_pool_out; 55 struct rte_mbuf **mbufs_in; 56 struct rte_mbuf **mbufs_out; 57 58 struct rte_mempool *crypto_op_pool; 59 60 struct rte_cryptodev_sym_session *sess; 61 62 cperf_populate_ops_t populate_ops; 63 64 const struct cperf_options *options; 65 const struct cperf_test_vector *test_vector; 66 struct cperf_op_result *res; 67 }; 68 69 struct priv_op_data { 70 struct cperf_op_result *result; 71 }; 72 73 #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b) 74 #define min(a, b) (a < b ? (uint64_t)a : (uint64_t)b) 75 76 static void 77 cperf_latency_test_free(struct cperf_latency_ctx *ctx, uint32_t mbuf_nb) 78 { 79 uint32_t i; 80 81 if (ctx) { 82 if (ctx->sess) 83 rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess); 84 85 if (ctx->mbufs_in) { 86 for (i = 0; i < mbuf_nb; i++) 87 rte_pktmbuf_free(ctx->mbufs_in[i]); 88 89 rte_free(ctx->mbufs_in); 90 } 91 92 if (ctx->mbufs_out) { 93 for (i = 0; i < mbuf_nb; i++) { 94 if (ctx->mbufs_out[i] != NULL) 95 rte_pktmbuf_free(ctx->mbufs_out[i]); 96 } 97 98 rte_free(ctx->mbufs_out); 99 } 100 101 if (ctx->pkt_mbuf_pool_in) 102 rte_mempool_free(ctx->pkt_mbuf_pool_in); 103 104 if (ctx->pkt_mbuf_pool_out) 105 rte_mempool_free(ctx->pkt_mbuf_pool_out); 106 107 if (ctx->crypto_op_pool) 108 rte_mempool_free(ctx->crypto_op_pool); 109 110 rte_free(ctx->res); 111 rte_free(ctx); 112 } 113 } 114 115 static struct rte_mbuf * 116 cperf_mbuf_create(struct rte_mempool *mempool, 117 uint32_t segments_nb, 118 const struct cperf_options *options, 119 const struct cperf_test_vector *test_vector) 120 { 121 struct rte_mbuf *mbuf; 122 uint32_t segment_sz = options->max_buffer_size / segments_nb; 123 uint32_t last_sz = options->max_buffer_size % segments_nb; 124 uint8_t *mbuf_data; 125 uint8_t *test_data = 126 (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ? 127 test_vector->plaintext.data : 128 test_vector->ciphertext.data; 129 130 mbuf = rte_pktmbuf_alloc(mempool); 131 if (mbuf == NULL) 132 goto error; 133 134 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz); 135 if (mbuf_data == NULL) 136 goto error; 137 138 memcpy(mbuf_data, test_data, segment_sz); 139 test_data += segment_sz; 140 segments_nb--; 141 142 while (segments_nb) { 143 struct rte_mbuf *m; 144 145 m = rte_pktmbuf_alloc(mempool); 146 if (m == NULL) 147 goto error; 148 149 rte_pktmbuf_chain(mbuf, m); 150 151 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz); 152 if (mbuf_data == NULL) 153 goto error; 154 155 memcpy(mbuf_data, test_data, segment_sz); 156 test_data += segment_sz; 157 segments_nb--; 158 } 159 160 if (last_sz) { 161 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz); 162 if (mbuf_data == NULL) 163 goto error; 164 165 memcpy(mbuf_data, test_data, last_sz); 166 } 167 168 if (options->op_type != CPERF_CIPHER_ONLY) { 169 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, 170 options->auth_digest_sz); 171 if (mbuf_data == NULL) 172 goto error; 173 } 174 175 if (options->op_type == CPERF_AEAD) { 176 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf, 177 RTE_ALIGN_CEIL(options->auth_aad_sz, 16)); 178 179 if (aead == NULL) 180 goto error; 181 182 memcpy(aead, test_vector->aad.data, test_vector->aad.length); 183 } 184 185 return mbuf; 186 error: 187 if (mbuf != NULL) 188 rte_pktmbuf_free(mbuf); 189 190 return NULL; 191 } 192 193 void * 194 cperf_latency_test_constructor(uint8_t dev_id, uint16_t qp_id, 195 const struct cperf_options *options, 196 const struct cperf_test_vector *test_vector, 197 const struct cperf_op_fns *op_fns) 198 { 199 struct cperf_latency_ctx *ctx = NULL; 200 unsigned int mbuf_idx = 0; 201 char pool_name[32] = ""; 202 203 ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0); 204 if (ctx == NULL) 205 goto err; 206 207 ctx->dev_id = dev_id; 208 ctx->qp_id = qp_id; 209 210 ctx->populate_ops = op_fns->populate_ops; 211 ctx->options = options; 212 ctx->test_vector = test_vector; 213 214 /* IV goes at the end of the crypto operation */ 215 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 216 sizeof(struct rte_crypto_sym_op) + 217 sizeof(struct cperf_op_result *); 218 219 ctx->sess = op_fns->sess_create(dev_id, options, test_vector, iv_offset); 220 if (ctx->sess == NULL) 221 goto err; 222 223 snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d", 224 dev_id); 225 226 ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name, 227 options->pool_sz * options->segments_nb, 0, 0, 228 RTE_PKTMBUF_HEADROOM + 229 RTE_CACHE_LINE_ROUNDUP( 230 (options->max_buffer_size / options->segments_nb) + 231 (options->max_buffer_size % options->segments_nb) + 232 options->auth_digest_sz), 233 rte_socket_id()); 234 235 if (ctx->pkt_mbuf_pool_in == NULL) 236 goto err; 237 238 /* Generate mbufs_in with plaintext populated for test */ 239 ctx->mbufs_in = rte_malloc(NULL, 240 (sizeof(struct rte_mbuf *) * 241 ctx->options->pool_sz), 0); 242 243 for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) { 244 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create( 245 ctx->pkt_mbuf_pool_in, options->segments_nb, 246 options, test_vector); 247 if (ctx->mbufs_in[mbuf_idx] == NULL) 248 goto err; 249 } 250 251 if (options->out_of_place == 1) { 252 253 snprintf(pool_name, sizeof(pool_name), 254 "cperf_pool_out_cdev_%d", 255 dev_id); 256 257 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create( 258 pool_name, options->pool_sz, 0, 0, 259 RTE_PKTMBUF_HEADROOM + 260 RTE_CACHE_LINE_ROUNDUP( 261 options->max_buffer_size + 262 options->auth_digest_sz), 263 rte_socket_id()); 264 265 if (ctx->pkt_mbuf_pool_out == NULL) 266 goto err; 267 } 268 269 ctx->mbufs_out = rte_malloc(NULL, 270 (sizeof(struct rte_mbuf *) * 271 ctx->options->pool_sz), 0); 272 273 for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) { 274 if (options->out_of_place == 1) { 275 ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create( 276 ctx->pkt_mbuf_pool_out, 1, 277 options, test_vector); 278 if (ctx->mbufs_out[mbuf_idx] == NULL) 279 goto err; 280 } else { 281 ctx->mbufs_out[mbuf_idx] = NULL; 282 } 283 } 284 285 snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d", 286 dev_id); 287 288 uint16_t priv_size = sizeof(struct priv_op_data) + test_vector->iv.length; 289 ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name, 290 RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 291 512, priv_size, rte_socket_id()); 292 293 if (ctx->crypto_op_pool == NULL) 294 goto err; 295 296 ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) * 297 ctx->options->total_ops, 0); 298 299 if (ctx->res == NULL) 300 goto err; 301 302 return ctx; 303 err: 304 cperf_latency_test_free(ctx, mbuf_idx); 305 306 return NULL; 307 } 308 309 static inline void 310 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp) 311 { 312 struct priv_op_data *priv_data; 313 314 priv_data = (struct priv_op_data *) (op->sym + 1); 315 priv_data->result->status = op->status; 316 priv_data->result->tsc_end = timestamp; 317 } 318 319 int 320 cperf_latency_test_runner(void *arg) 321 { 322 struct cperf_latency_ctx *ctx = arg; 323 uint16_t test_burst_size; 324 uint8_t burst_size_idx = 0; 325 326 static int only_once; 327 328 if (ctx == NULL) 329 return 0; 330 331 struct rte_crypto_op *ops[ctx->options->max_burst_size]; 332 struct rte_crypto_op *ops_processed[ctx->options->max_burst_size]; 333 uint64_t i; 334 struct priv_op_data *priv_data; 335 336 uint32_t lcore = rte_lcore_id(); 337 338 #ifdef CPERF_LINEARIZATION_ENABLE 339 struct rte_cryptodev_info dev_info; 340 int linearize = 0; 341 342 /* Check if source mbufs require coalescing */ 343 if (ctx->options->segments_nb > 1) { 344 rte_cryptodev_info_get(ctx->dev_id, &dev_info); 345 if ((dev_info.feature_flags & 346 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0) 347 linearize = 1; 348 } 349 #endif /* CPERF_LINEARIZATION_ENABLE */ 350 351 ctx->lcore_id = lcore; 352 353 /* Warm up the host CPU before starting the test */ 354 for (i = 0; i < ctx->options->total_ops; i++) 355 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 356 357 /* Get first size from range or list */ 358 if (ctx->options->inc_burst_size != 0) 359 test_burst_size = ctx->options->min_burst_size; 360 else 361 test_burst_size = ctx->options->burst_size_list[0]; 362 363 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 364 sizeof(struct rte_crypto_sym_op) + 365 sizeof(struct cperf_op_result *); 366 367 while (test_burst_size <= ctx->options->max_burst_size) { 368 uint64_t ops_enqd = 0, ops_deqd = 0; 369 uint64_t m_idx = 0, b_idx = 0; 370 371 uint64_t tsc_val, tsc_end, tsc_start; 372 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0; 373 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0; 374 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0; 375 376 while (enqd_tot < ctx->options->total_ops) { 377 378 uint16_t burst_size = ((enqd_tot + test_burst_size) 379 <= ctx->options->total_ops) ? 380 test_burst_size : 381 ctx->options->total_ops - 382 enqd_tot; 383 384 /* Allocate crypto ops from pool */ 385 if (burst_size != rte_crypto_op_bulk_alloc( 386 ctx->crypto_op_pool, 387 RTE_CRYPTO_OP_TYPE_SYMMETRIC, 388 ops, burst_size)) 389 return -1; 390 391 /* Setup crypto op, attach mbuf etc */ 392 (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx], 393 &ctx->mbufs_out[m_idx], 394 burst_size, ctx->sess, ctx->options, 395 ctx->test_vector, iv_offset); 396 397 tsc_start = rte_rdtsc_precise(); 398 399 #ifdef CPERF_LINEARIZATION_ENABLE 400 if (linearize) { 401 /* PMD doesn't support scatter-gather and source buffer 402 * is segmented. 403 * We need to linearize it before enqueuing. 404 */ 405 for (i = 0; i < burst_size; i++) 406 rte_pktmbuf_linearize(ops[i]->sym->m_src); 407 } 408 #endif /* CPERF_LINEARIZATION_ENABLE */ 409 410 /* Enqueue burst of ops on crypto device */ 411 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, 412 ops, burst_size); 413 414 /* Dequeue processed burst of ops from crypto device */ 415 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 416 ops_processed, test_burst_size); 417 418 tsc_end = rte_rdtsc_precise(); 419 420 /* Free memory for not enqueued operations */ 421 if (ops_enqd != burst_size) 422 rte_mempool_put_bulk(ctx->crypto_op_pool, 423 (void **)&ops_processed[ops_enqd], 424 burst_size - ops_enqd); 425 426 for (i = 0; i < ops_enqd; i++) { 427 ctx->res[tsc_idx].tsc_start = tsc_start; 428 /* 429 * Private data structure starts after the end of the 430 * rte_crypto_sym_op structure. 431 */ 432 priv_data = (struct priv_op_data *) (ops[i]->sym + 1); 433 priv_data->result = (void *)&ctx->res[tsc_idx]; 434 tsc_idx++; 435 } 436 437 if (likely(ops_deqd)) { 438 /* 439 * free crypto ops so they can be reused. We don't free 440 * the mbufs here as we don't want to reuse them as 441 * the crypto operation will change the data and cause 442 * failures. 443 */ 444 for (i = 0; i < ops_deqd; i++) 445 store_timestamp(ops_processed[i], tsc_end); 446 447 rte_mempool_put_bulk(ctx->crypto_op_pool, 448 (void **)ops_processed, ops_deqd); 449 450 deqd_tot += ops_deqd; 451 deqd_max = max(ops_deqd, deqd_max); 452 deqd_min = min(ops_deqd, deqd_min); 453 } 454 455 enqd_tot += ops_enqd; 456 enqd_max = max(ops_enqd, enqd_max); 457 enqd_min = min(ops_enqd, enqd_min); 458 459 m_idx += ops_enqd; 460 m_idx = m_idx + test_burst_size > ctx->options->pool_sz ? 461 0 : m_idx; 462 b_idx++; 463 } 464 465 /* Dequeue any operations still in the crypto device */ 466 while (deqd_tot < ctx->options->total_ops) { 467 /* Sending 0 length burst to flush sw crypto device */ 468 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 469 470 /* dequeue burst */ 471 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 472 ops_processed, test_burst_size); 473 474 tsc_end = rte_rdtsc_precise(); 475 476 if (ops_deqd != 0) { 477 for (i = 0; i < ops_deqd; i++) 478 store_timestamp(ops_processed[i], tsc_end); 479 480 rte_mempool_put_bulk(ctx->crypto_op_pool, 481 (void **)ops_processed, ops_deqd); 482 483 deqd_tot += ops_deqd; 484 deqd_max = max(ops_deqd, deqd_max); 485 deqd_min = min(ops_deqd, deqd_min); 486 } 487 } 488 489 for (i = 0; i < tsc_idx; i++) { 490 tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start; 491 tsc_max = max(tsc_val, tsc_max); 492 tsc_min = min(tsc_val, tsc_min); 493 tsc_tot += tsc_val; 494 } 495 496 double time_tot, time_avg, time_max, time_min; 497 498 const uint64_t tunit = 1000000; /* us */ 499 const uint64_t tsc_hz = rte_get_tsc_hz(); 500 501 uint64_t enqd_avg = enqd_tot / b_idx; 502 uint64_t deqd_avg = deqd_tot / b_idx; 503 uint64_t tsc_avg = tsc_tot / tsc_idx; 504 505 time_tot = tunit*(double)(tsc_tot) / tsc_hz; 506 time_avg = tunit*(double)(tsc_avg) / tsc_hz; 507 time_max = tunit*(double)(tsc_max) / tsc_hz; 508 time_min = tunit*(double)(tsc_min) / tsc_hz; 509 510 if (ctx->options->csv) { 511 if (!only_once) 512 printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, " 513 "Packet Size, cycles, time (us)"); 514 515 for (i = 0; i < ctx->options->total_ops; i++) { 516 517 printf("\n%u;%u;%u;%"PRIu64";%"PRIu64";%.3f", 518 ctx->lcore_id, ctx->options->test_buffer_size, 519 test_burst_size, i + 1, 520 ctx->res[i].tsc_end - ctx->res[i].tsc_start, 521 tunit * (double) (ctx->res[i].tsc_end 522 - ctx->res[i].tsc_start) 523 / tsc_hz); 524 525 } 526 only_once = 1; 527 } else { 528 printf("\n# Device %d on lcore %u\n", ctx->dev_id, 529 ctx->lcore_id); 530 printf("\n# total operations: %u", ctx->options->total_ops); 531 printf("\n# Buffer size: %u", ctx->options->test_buffer_size); 532 printf("\n# Burst size: %u", test_burst_size); 533 printf("\n# Number of bursts: %"PRIu64, 534 b_idx); 535 536 printf("\n#"); 537 printf("\n# \t Total\t Average\t " 538 "Maximum\t Minimum"); 539 printf("\n# enqueued\t%12"PRIu64"\t%10"PRIu64"\t" 540 "%10"PRIu64"\t%10"PRIu64, enqd_tot, 541 enqd_avg, enqd_max, enqd_min); 542 printf("\n# dequeued\t%12"PRIu64"\t%10"PRIu64"\t" 543 "%10"PRIu64"\t%10"PRIu64, deqd_tot, 544 deqd_avg, deqd_max, deqd_min); 545 printf("\n# cycles\t%12"PRIu64"\t%10"PRIu64"\t" 546 "%10"PRIu64"\t%10"PRIu64, tsc_tot, 547 tsc_avg, tsc_max, tsc_min); 548 printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f", 549 time_tot, time_avg, time_max, time_min); 550 printf("\n\n"); 551 552 } 553 554 /* Get next size from range or list */ 555 if (ctx->options->inc_burst_size != 0) 556 test_burst_size += ctx->options->inc_burst_size; 557 else { 558 if (++burst_size_idx == ctx->options->burst_size_count) 559 break; 560 test_burst_size = 561 ctx->options->burst_size_list[burst_size_idx]; 562 } 563 } 564 565 return 0; 566 } 567 568 void 569 cperf_latency_test_destructor(void *arg) 570 { 571 struct cperf_latency_ctx *ctx = arg; 572 573 if (ctx == NULL) 574 return; 575 576 cperf_latency_test_free(ctx, ctx->options->pool_sz); 577 578 } 579