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_throughput.h" 39 #include "cperf_ops.h" 40 41 struct cperf_throughput_ctx { 42 uint8_t dev_id; 43 uint16_t qp_id; 44 uint8_t lcore_id; 45 46 struct rte_mempool *pkt_mbuf_pool_in; 47 struct rte_mempool *pkt_mbuf_pool_out; 48 struct rte_mbuf **mbufs_in; 49 struct rte_mbuf **mbufs_out; 50 51 struct rte_mempool *crypto_op_pool; 52 53 struct rte_cryptodev_sym_session *sess; 54 55 cperf_populate_ops_t populate_ops; 56 57 const struct cperf_options *options; 58 const struct cperf_test_vector *test_vector; 59 }; 60 61 static void 62 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx, uint32_t mbuf_nb) 63 { 64 uint32_t i; 65 66 if (ctx) { 67 if (ctx->sess) 68 rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess); 69 70 if (ctx->mbufs_in) { 71 for (i = 0; i < mbuf_nb; i++) 72 rte_pktmbuf_free(ctx->mbufs_in[i]); 73 74 rte_free(ctx->mbufs_in); 75 } 76 77 if (ctx->mbufs_out) { 78 for (i = 0; i < mbuf_nb; i++) { 79 if (ctx->mbufs_out[i] != NULL) 80 rte_pktmbuf_free(ctx->mbufs_out[i]); 81 } 82 83 rte_free(ctx->mbufs_out); 84 } 85 86 if (ctx->pkt_mbuf_pool_in) 87 rte_mempool_free(ctx->pkt_mbuf_pool_in); 88 89 if (ctx->pkt_mbuf_pool_out) 90 rte_mempool_free(ctx->pkt_mbuf_pool_out); 91 92 if (ctx->crypto_op_pool) 93 rte_mempool_free(ctx->crypto_op_pool); 94 95 rte_free(ctx); 96 } 97 } 98 99 static struct rte_mbuf * 100 cperf_mbuf_create(struct rte_mempool *mempool, 101 uint32_t segments_nb, 102 const struct cperf_options *options, 103 const struct cperf_test_vector *test_vector) 104 { 105 struct rte_mbuf *mbuf; 106 uint32_t segment_sz = options->max_buffer_size / segments_nb; 107 uint32_t last_sz = options->max_buffer_size % segments_nb; 108 uint8_t *mbuf_data; 109 uint8_t *test_data = 110 (options->cipher_op == RTE_CRYPTO_CIPHER_OP_ENCRYPT) ? 111 test_vector->plaintext.data : 112 test_vector->ciphertext.data; 113 114 mbuf = rte_pktmbuf_alloc(mempool); 115 if (mbuf == NULL) 116 goto error; 117 118 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz); 119 if (mbuf_data == NULL) 120 goto error; 121 122 memcpy(mbuf_data, test_data, segment_sz); 123 test_data += segment_sz; 124 segments_nb--; 125 126 while (segments_nb) { 127 struct rte_mbuf *m; 128 129 m = rte_pktmbuf_alloc(mempool); 130 if (m == NULL) 131 goto error; 132 133 rte_pktmbuf_chain(mbuf, m); 134 135 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, segment_sz); 136 if (mbuf_data == NULL) 137 goto error; 138 139 memcpy(mbuf_data, test_data, segment_sz); 140 test_data += segment_sz; 141 segments_nb--; 142 } 143 144 if (last_sz) { 145 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, last_sz); 146 if (mbuf_data == NULL) 147 goto error; 148 149 memcpy(mbuf_data, test_data, last_sz); 150 } 151 152 if (options->op_type != CPERF_CIPHER_ONLY) { 153 mbuf_data = (uint8_t *)rte_pktmbuf_append(mbuf, 154 options->digest_sz); 155 if (mbuf_data == NULL) 156 goto error; 157 } 158 159 if (options->op_type == CPERF_AEAD) { 160 uint8_t *aead = (uint8_t *)rte_pktmbuf_prepend(mbuf, 161 RTE_ALIGN_CEIL(options->aead_aad_sz, 16)); 162 163 if (aead == NULL) 164 goto error; 165 166 memcpy(aead, test_vector->aad.data, test_vector->aad.length); 167 } 168 169 return mbuf; 170 error: 171 if (mbuf != NULL) 172 rte_pktmbuf_free(mbuf); 173 174 return NULL; 175 } 176 177 void * 178 cperf_throughput_test_constructor(uint8_t dev_id, uint16_t qp_id, 179 const struct cperf_options *options, 180 const struct cperf_test_vector *test_vector, 181 const struct cperf_op_fns *op_fns) 182 { 183 struct cperf_throughput_ctx *ctx = NULL; 184 unsigned int mbuf_idx = 0; 185 char pool_name[32] = ""; 186 187 ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0); 188 if (ctx == NULL) 189 goto err; 190 191 ctx->dev_id = dev_id; 192 ctx->qp_id = qp_id; 193 194 ctx->populate_ops = op_fns->populate_ops; 195 ctx->options = options; 196 ctx->test_vector = test_vector; 197 198 /* IV goes at the end of the cryptop operation */ 199 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 200 sizeof(struct rte_crypto_sym_op); 201 202 ctx->sess = op_fns->sess_create(dev_id, options, test_vector, iv_offset); 203 if (ctx->sess == NULL) 204 goto err; 205 206 snprintf(pool_name, sizeof(pool_name), "cperf_pool_in_cdev_%d", 207 dev_id); 208 209 ctx->pkt_mbuf_pool_in = rte_pktmbuf_pool_create(pool_name, 210 options->pool_sz * options->segments_nb, 0, 0, 211 RTE_PKTMBUF_HEADROOM + 212 RTE_CACHE_LINE_ROUNDUP( 213 (options->max_buffer_size / options->segments_nb) + 214 (options->max_buffer_size % options->segments_nb) + 215 options->digest_sz), 216 rte_socket_id()); 217 218 if (ctx->pkt_mbuf_pool_in == NULL) 219 goto err; 220 221 /* Generate mbufs_in with plaintext populated for test */ 222 ctx->mbufs_in = rte_malloc(NULL, 223 (sizeof(struct rte_mbuf *) * ctx->options->pool_sz), 0); 224 225 for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) { 226 ctx->mbufs_in[mbuf_idx] = cperf_mbuf_create( 227 ctx->pkt_mbuf_pool_in, options->segments_nb, 228 options, test_vector); 229 if (ctx->mbufs_in[mbuf_idx] == NULL) 230 goto err; 231 } 232 233 if (options->out_of_place == 1) { 234 235 snprintf(pool_name, sizeof(pool_name), "cperf_pool_out_cdev_%d", 236 dev_id); 237 238 ctx->pkt_mbuf_pool_out = rte_pktmbuf_pool_create( 239 pool_name, options->pool_sz, 0, 0, 240 RTE_PKTMBUF_HEADROOM + 241 RTE_CACHE_LINE_ROUNDUP( 242 options->max_buffer_size + 243 options->digest_sz), 244 rte_socket_id()); 245 246 if (ctx->pkt_mbuf_pool_out == NULL) 247 goto err; 248 } 249 250 ctx->mbufs_out = rte_malloc(NULL, 251 (sizeof(struct rte_mbuf *) * 252 ctx->options->pool_sz), 0); 253 254 for (mbuf_idx = 0; mbuf_idx < options->pool_sz; mbuf_idx++) { 255 if (options->out_of_place == 1) { 256 ctx->mbufs_out[mbuf_idx] = cperf_mbuf_create( 257 ctx->pkt_mbuf_pool_out, 1, 258 options, test_vector); 259 if (ctx->mbufs_out[mbuf_idx] == NULL) 260 goto err; 261 } else { 262 ctx->mbufs_out[mbuf_idx] = NULL; 263 } 264 } 265 266 snprintf(pool_name, sizeof(pool_name), "cperf_op_pool_cdev_%d", 267 dev_id); 268 269 uint16_t priv_size = test_vector->cipher_iv.length + 270 test_vector->auth_iv.length; 271 272 ctx->crypto_op_pool = rte_crypto_op_pool_create(pool_name, 273 RTE_CRYPTO_OP_TYPE_SYMMETRIC, options->pool_sz, 274 512, priv_size, rte_socket_id()); 275 if (ctx->crypto_op_pool == NULL) 276 goto err; 277 278 return ctx; 279 err: 280 cperf_throughput_test_free(ctx, mbuf_idx); 281 282 return NULL; 283 } 284 285 int 286 cperf_throughput_test_runner(void *test_ctx) 287 { 288 struct cperf_throughput_ctx *ctx = test_ctx; 289 uint16_t test_burst_size; 290 uint8_t burst_size_idx = 0; 291 292 static int only_once; 293 294 struct rte_crypto_op *ops[ctx->options->max_burst_size]; 295 struct rte_crypto_op *ops_processed[ctx->options->max_burst_size]; 296 uint64_t i; 297 298 uint32_t lcore = rte_lcore_id(); 299 300 #ifdef CPERF_LINEARIZATION_ENABLE 301 struct rte_cryptodev_info dev_info; 302 int linearize = 0; 303 304 /* Check if source mbufs require coalescing */ 305 if (ctx->options->segments_nb > 1) { 306 rte_cryptodev_info_get(ctx->dev_id, &dev_info); 307 if ((dev_info.feature_flags & 308 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0) 309 linearize = 1; 310 } 311 #endif /* CPERF_LINEARIZATION_ENABLE */ 312 313 ctx->lcore_id = lcore; 314 315 /* Warm up the host CPU before starting the test */ 316 for (i = 0; i < ctx->options->total_ops; i++) 317 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 318 319 /* Get first size from range or list */ 320 if (ctx->options->inc_burst_size != 0) 321 test_burst_size = ctx->options->min_burst_size; 322 else 323 test_burst_size = ctx->options->burst_size_list[0]; 324 325 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 326 sizeof(struct rte_crypto_sym_op); 327 328 while (test_burst_size <= ctx->options->max_burst_size) { 329 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0; 330 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0; 331 332 uint64_t m_idx = 0, tsc_start, tsc_end, tsc_duration; 333 334 uint16_t ops_unused = 0; 335 336 tsc_start = rte_rdtsc_precise(); 337 338 while (ops_enqd_total < ctx->options->total_ops) { 339 340 uint16_t burst_size = ((ops_enqd_total + test_burst_size) 341 <= ctx->options->total_ops) ? 342 test_burst_size : 343 ctx->options->total_ops - 344 ops_enqd_total; 345 346 uint16_t ops_needed = burst_size - ops_unused; 347 348 /* Allocate crypto ops from pool */ 349 if (ops_needed != rte_crypto_op_bulk_alloc( 350 ctx->crypto_op_pool, 351 RTE_CRYPTO_OP_TYPE_SYMMETRIC, 352 ops, ops_needed)) 353 return -1; 354 355 /* Setup crypto op, attach mbuf etc */ 356 (ctx->populate_ops)(ops, &ctx->mbufs_in[m_idx], 357 &ctx->mbufs_out[m_idx], 358 ops_needed, ctx->sess, ctx->options, 359 ctx->test_vector, iv_offset); 360 361 /** 362 * When ops_needed is smaller than ops_enqd, the 363 * unused ops need to be moved to the front for 364 * next round use. 365 */ 366 if (unlikely(ops_enqd > ops_needed)) { 367 size_t nb_b_to_mov = ops_unused * sizeof( 368 struct rte_crypto_op *); 369 370 memmove(&ops[ops_needed], &ops[ops_enqd], 371 nb_b_to_mov); 372 } 373 374 #ifdef CPERF_LINEARIZATION_ENABLE 375 if (linearize) { 376 /* PMD doesn't support scatter-gather and source buffer 377 * is segmented. 378 * We need to linearize it before enqueuing. 379 */ 380 for (i = 0; i < burst_size; i++) 381 rte_pktmbuf_linearize(ops[i]->sym->m_src); 382 } 383 #endif /* CPERF_LINEARIZATION_ENABLE */ 384 385 /* Enqueue burst of ops on crypto device */ 386 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, 387 ops, burst_size); 388 if (ops_enqd < burst_size) 389 ops_enqd_failed++; 390 391 /** 392 * Calculate number of ops not enqueued (mainly for hw 393 * accelerators whose ingress queue can fill up). 394 */ 395 ops_unused = burst_size - ops_enqd; 396 ops_enqd_total += ops_enqd; 397 398 399 /* Dequeue processed burst of ops from crypto device */ 400 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 401 ops_processed, test_burst_size); 402 403 if (likely(ops_deqd)) { 404 /* free crypto ops so they can be reused. We don't free 405 * the mbufs here as we don't want to reuse them as 406 * the crypto operation will change the data and cause 407 * failures. 408 */ 409 rte_mempool_put_bulk(ctx->crypto_op_pool, 410 (void **)ops_processed, ops_deqd); 411 412 ops_deqd_total += ops_deqd; 413 } else { 414 /** 415 * Count dequeue polls which didn't return any 416 * processed operations. This statistic is mainly 417 * relevant to hw accelerators. 418 */ 419 ops_deqd_failed++; 420 } 421 422 m_idx += ops_needed; 423 m_idx = m_idx + test_burst_size > ctx->options->pool_sz ? 424 0 : m_idx; 425 } 426 427 /* Dequeue any operations still in the crypto device */ 428 429 while (ops_deqd_total < ctx->options->total_ops) { 430 /* Sending 0 length burst to flush sw crypto device */ 431 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 432 433 /* dequeue burst */ 434 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 435 ops_processed, test_burst_size); 436 if (ops_deqd == 0) 437 ops_deqd_failed++; 438 else { 439 rte_mempool_put_bulk(ctx->crypto_op_pool, 440 (void **)ops_processed, ops_deqd); 441 442 ops_deqd_total += ops_deqd; 443 } 444 } 445 446 tsc_end = rte_rdtsc_precise(); 447 tsc_duration = (tsc_end - tsc_start); 448 449 /* Calculate average operations processed per second */ 450 double ops_per_second = ((double)ctx->options->total_ops / 451 tsc_duration) * rte_get_tsc_hz(); 452 453 /* Calculate average throughput (Gbps) in bits per second */ 454 double throughput_gbps = ((ops_per_second * 455 ctx->options->test_buffer_size * 8) / 1000000000); 456 457 /* Calculate average cycles per packet */ 458 double cycles_per_packet = ((double)tsc_duration / 459 ctx->options->total_ops); 460 461 if (!ctx->options->csv) { 462 if (!only_once) 463 printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n", 464 "lcore id", "Buf Size", "Burst Size", 465 "Enqueued", "Dequeued", "Failed Enq", 466 "Failed Deq", "MOps", "Gbps", 467 "Cycles/Buf"); 468 only_once = 1; 469 470 printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64 471 "%12"PRIu64"%12.4f%12.4f%12.2f\n", 472 ctx->lcore_id, 473 ctx->options->test_buffer_size, 474 test_burst_size, 475 ops_enqd_total, 476 ops_deqd_total, 477 ops_enqd_failed, 478 ops_deqd_failed, 479 ops_per_second/1000000, 480 throughput_gbps, 481 cycles_per_packet); 482 } else { 483 if (!only_once) 484 printf("# lcore id, Buffer Size(B)," 485 "Burst Size,Enqueued,Dequeued,Failed Enq," 486 "Failed Deq,Ops(Millions),Throughput(Gbps)," 487 "Cycles/Buf\n\n"); 488 only_once = 1; 489 490 printf("%10u;%10u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";" 491 "%.f3;%.f3;%.f3\n", 492 ctx->lcore_id, 493 ctx->options->test_buffer_size, 494 test_burst_size, 495 ops_enqd_total, 496 ops_deqd_total, 497 ops_enqd_failed, 498 ops_deqd_failed, 499 ops_per_second/1000000, 500 throughput_gbps, 501 cycles_per_packet); 502 } 503 504 /* Get next size from range or list */ 505 if (ctx->options->inc_burst_size != 0) 506 test_burst_size += ctx->options->inc_burst_size; 507 else { 508 if (++burst_size_idx == ctx->options->burst_size_count) 509 break; 510 test_burst_size = ctx->options->burst_size_list[burst_size_idx]; 511 } 512 513 } 514 515 return 0; 516 } 517 518 519 void 520 cperf_throughput_test_destructor(void *arg) 521 { 522 struct cperf_throughput_ctx *ctx = arg; 523 524 if (ctx == NULL) 525 return; 526 527 cperf_throughput_test_free(ctx, ctx->options->pool_sz); 528 } 529