1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2016-2017 Intel Corporation 3 */ 4 5 #include <stdlib.h> 6 7 #include <rte_malloc.h> 8 #include <rte_cycles.h> 9 #include <rte_crypto.h> 10 #include <rte_cryptodev.h> 11 12 #include "cperf_test_throughput.h" 13 #include "cperf_ops.h" 14 #include "cperf_test_common.h" 15 16 struct cperf_throughput_ctx { 17 uint8_t dev_id; 18 uint16_t qp_id; 19 uint8_t lcore_id; 20 21 struct rte_mempool *pool; 22 23 void *sess; 24 25 cperf_populate_ops_t populate_ops; 26 27 uint32_t src_buf_offset; 28 uint32_t dst_buf_offset; 29 30 const struct cperf_options *options; 31 const struct cperf_test_vector *test_vector; 32 }; 33 34 static void 35 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx) 36 { 37 if (!ctx) 38 return; 39 if (ctx->sess) { 40 if (ctx->options->op_type == CPERF_ASYM_MODEX) 41 rte_cryptodev_asym_session_free(ctx->dev_id, 42 (void *)ctx->sess); 43 #ifdef RTE_LIB_SECURITY 44 else if (ctx->options->op_type == CPERF_PDCP || 45 ctx->options->op_type == CPERF_DOCSIS || 46 ctx->options->op_type == CPERF_IPSEC) { 47 void *sec_ctx = rte_cryptodev_get_sec_ctx(ctx->dev_id); 48 49 rte_security_session_destroy(sec_ctx, (void *)ctx->sess); 50 } 51 #endif 52 else 53 rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess); 54 } 55 rte_mempool_free(ctx->pool); 56 57 rte_free(ctx); 58 } 59 60 void * 61 cperf_throughput_test_constructor(struct rte_mempool *sess_mp, 62 uint8_t dev_id, uint16_t qp_id, 63 const struct cperf_options *options, 64 const struct cperf_test_vector *test_vector, 65 const struct cperf_op_fns *op_fns) 66 { 67 struct cperf_throughput_ctx *ctx = NULL; 68 69 ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0); 70 if (ctx == NULL) 71 goto err; 72 73 ctx->dev_id = dev_id; 74 ctx->qp_id = qp_id; 75 76 ctx->populate_ops = op_fns->populate_ops; 77 ctx->options = options; 78 ctx->test_vector = test_vector; 79 80 /* IV goes at the end of the crypto operation */ 81 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 82 sizeof(struct rte_crypto_sym_op); 83 84 ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, 85 test_vector, iv_offset); 86 if (ctx->sess == NULL) 87 goto err; 88 89 if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0, 90 &ctx->src_buf_offset, &ctx->dst_buf_offset, 91 &ctx->pool) < 0) 92 goto err; 93 94 return ctx; 95 err: 96 cperf_throughput_test_free(ctx); 97 98 return NULL; 99 } 100 101 int 102 cperf_throughput_test_runner(void *test_ctx) 103 { 104 struct cperf_throughput_ctx *ctx = test_ctx; 105 uint16_t test_burst_size; 106 uint8_t burst_size_idx = 0; 107 uint32_t imix_idx = 0; 108 109 static uint16_t display_once; 110 111 struct rte_crypto_op *ops[ctx->options->max_burst_size]; 112 struct rte_crypto_op *ops_processed[ctx->options->max_burst_size]; 113 uint64_t i; 114 115 uint32_t lcore = rte_lcore_id(); 116 117 #ifdef CPERF_LINEARIZATION_ENABLE 118 struct rte_cryptodev_info dev_info; 119 int linearize = 0; 120 121 /* Check if source mbufs require coalescing */ 122 if ((ctx->options->op_type != CPERF_ASYM_MODEX) && 123 (ctx->options->segment_sz < ctx->options->max_buffer_size)) { 124 rte_cryptodev_info_get(ctx->dev_id, &dev_info); 125 if ((dev_info.feature_flags & 126 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0) 127 linearize = 1; 128 } 129 #endif /* CPERF_LINEARIZATION_ENABLE */ 130 131 ctx->lcore_id = lcore; 132 133 /* Warm up the host CPU before starting the test */ 134 for (i = 0; i < ctx->options->total_ops; i++) 135 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 136 137 /* Get first size from range or list */ 138 if (ctx->options->inc_burst_size != 0) 139 test_burst_size = ctx->options->min_burst_size; 140 else 141 test_burst_size = ctx->options->burst_size_list[0]; 142 143 uint16_t iv_offset = sizeof(struct rte_crypto_op) + 144 sizeof(struct rte_crypto_sym_op); 145 146 while (test_burst_size <= ctx->options->max_burst_size) { 147 uint64_t ops_enqd = 0, ops_enqd_total = 0, ops_enqd_failed = 0; 148 uint64_t ops_deqd = 0, ops_deqd_total = 0, ops_deqd_failed = 0; 149 150 uint64_t tsc_start, tsc_end, tsc_duration; 151 152 uint16_t ops_unused = 0; 153 154 tsc_start = rte_rdtsc_precise(); 155 156 while (ops_enqd_total < ctx->options->total_ops) { 157 158 uint16_t burst_size = ((ops_enqd_total + test_burst_size) 159 <= ctx->options->total_ops) ? 160 test_burst_size : 161 ctx->options->total_ops - 162 ops_enqd_total; 163 164 uint16_t ops_needed = burst_size - ops_unused; 165 166 /* Allocate objects containing crypto operations and mbufs */ 167 if (rte_mempool_get_bulk(ctx->pool, (void **)ops, 168 ops_needed) != 0) { 169 RTE_LOG(ERR, USER1, 170 "Failed to allocate more crypto operations " 171 "from the crypto operation pool.\n" 172 "Consider increasing the pool size " 173 "with --pool-sz\n"); 174 return -1; 175 } 176 177 /* Setup crypto op, attach mbuf etc */ 178 (ctx->populate_ops)(ops, ctx->src_buf_offset, 179 ctx->dst_buf_offset, 180 ops_needed, ctx->sess, 181 ctx->options, ctx->test_vector, 182 iv_offset, &imix_idx, &tsc_start); 183 184 /** 185 * When ops_needed is smaller than ops_enqd, the 186 * unused ops need to be moved to the front for 187 * next round use. 188 */ 189 if (unlikely(ops_enqd > ops_needed)) { 190 size_t nb_b_to_mov = ops_unused * sizeof( 191 struct rte_crypto_op *); 192 193 memmove(&ops[ops_needed], &ops[ops_enqd], 194 nb_b_to_mov); 195 } 196 197 #ifdef CPERF_LINEARIZATION_ENABLE 198 if (linearize) { 199 /* PMD doesn't support scatter-gather and source buffer 200 * is segmented. 201 * We need to linearize it before enqueuing. 202 */ 203 for (i = 0; i < burst_size; i++) 204 rte_pktmbuf_linearize( 205 ops[i]->sym->m_src); 206 } 207 #endif /* CPERF_LINEARIZATION_ENABLE */ 208 209 /* Enqueue burst of ops on crypto device */ 210 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, 211 ops, burst_size); 212 if (ops_enqd < burst_size) 213 ops_enqd_failed++; 214 215 /** 216 * Calculate number of ops not enqueued (mainly for hw 217 * accelerators whose ingress queue can fill up). 218 */ 219 ops_unused = burst_size - ops_enqd; 220 ops_enqd_total += ops_enqd; 221 222 223 /* Dequeue processed burst of ops from crypto device */ 224 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 225 ops_processed, test_burst_size); 226 227 if (likely(ops_deqd)) { 228 /* Free crypto ops so they can be reused. */ 229 rte_mempool_put_bulk(ctx->pool, 230 (void **)ops_processed, ops_deqd); 231 232 ops_deqd_total += ops_deqd; 233 } else { 234 /** 235 * Count dequeue polls which didn't return any 236 * processed operations. This statistic is mainly 237 * relevant to hw accelerators. 238 */ 239 ops_deqd_failed++; 240 } 241 242 } 243 244 /* Dequeue any operations still in the crypto device */ 245 246 while (ops_deqd_total < ctx->options->total_ops) { 247 /* Sending 0 length burst to flush sw crypto device */ 248 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0); 249 250 /* dequeue burst */ 251 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id, 252 ops_processed, test_burst_size); 253 if (ops_deqd == 0) 254 ops_deqd_failed++; 255 else { 256 rte_mempool_put_bulk(ctx->pool, 257 (void **)ops_processed, ops_deqd); 258 ops_deqd_total += ops_deqd; 259 } 260 } 261 262 tsc_end = rte_rdtsc_precise(); 263 tsc_duration = (tsc_end - tsc_start); 264 265 /* Calculate average operations processed per second */ 266 double ops_per_second = ((double)ctx->options->total_ops / 267 tsc_duration) * rte_get_tsc_hz(); 268 269 /* Calculate average throughput (Gbps) in bits per second */ 270 double throughput_gbps = ((ops_per_second * 271 ctx->options->test_buffer_size * 8) / 1000000000); 272 273 /* Calculate average cycles per packet */ 274 double cycles_per_packet = ((double)tsc_duration / 275 ctx->options->total_ops); 276 277 uint16_t exp = 0; 278 if (!ctx->options->csv) { 279 if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0, 280 __ATOMIC_RELAXED, __ATOMIC_RELAXED)) 281 printf("%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n", 282 "lcore id", "Buf Size", "Burst Size", 283 "Enqueued", "Dequeued", "Failed Enq", 284 "Failed Deq", "MOps", "Gbps", 285 "Cycles/Buf"); 286 287 printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64 288 "%12"PRIu64"%12.4f%12.4f%12.2f\n", 289 ctx->lcore_id, 290 ctx->options->test_buffer_size, 291 test_burst_size, 292 ops_enqd_total, 293 ops_deqd_total, 294 ops_enqd_failed, 295 ops_deqd_failed, 296 ops_per_second/1000000, 297 throughput_gbps, 298 cycles_per_packet); 299 } else { 300 if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0, 301 __ATOMIC_RELAXED, __ATOMIC_RELAXED)) 302 printf("#lcore id,Buffer Size(B)," 303 "Burst Size,Enqueued,Dequeued,Failed Enq," 304 "Failed Deq,Ops(Millions),Throughput(Gbps)," 305 "Cycles/Buf\n\n"); 306 307 printf("%u,%u,%u,%"PRIu64",%"PRIu64",%"PRIu64",%"PRIu64"," 308 "%.3f,%.3f,%.3f\n", 309 ctx->lcore_id, 310 ctx->options->test_buffer_size, 311 test_burst_size, 312 ops_enqd_total, 313 ops_deqd_total, 314 ops_enqd_failed, 315 ops_deqd_failed, 316 ops_per_second/1000000, 317 throughput_gbps, 318 cycles_per_packet); 319 } 320 321 /* Get next size from range or list */ 322 if (ctx->options->inc_burst_size != 0) 323 test_burst_size += ctx->options->inc_burst_size; 324 else { 325 if (++burst_size_idx == ctx->options->burst_size_count) 326 break; 327 test_burst_size = ctx->options->burst_size_list[burst_size_idx]; 328 } 329 330 } 331 332 return 0; 333 } 334 335 336 void 337 cperf_throughput_test_destructor(void *arg) 338 { 339 struct cperf_throughput_ctx *ctx = arg; 340 341 if (ctx == NULL) 342 return; 343 344 cperf_throughput_test_free(ctx); 345 } 346