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