1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2016-2017 Intel Corporation
5 #include <rte_malloc.h>
6 #include <rte_cycles.h>
7 #include <rte_crypto.h>
8 #include <rte_cryptodev.h>
10 #include "cperf_test_latency.h"
11 #include "cperf_ops.h"
12 #include "cperf_test_common.h"
14 struct cperf_op_result {
17 enum rte_crypto_op_status status;
20 struct cperf_latency_ctx {
25 struct rte_mempool *pool;
27 struct rte_cryptodev_sym_session *sess;
29 cperf_populate_ops_t populate_ops;
31 uint32_t src_buf_offset;
32 uint32_t dst_buf_offset;
34 const struct cperf_options *options;
35 const struct cperf_test_vector *test_vector;
36 struct cperf_op_result *res;
40 struct cperf_op_result *result;
43 #define max(a, b) (a > b ? (uint64_t)a : (uint64_t)b)
44 #define min(a, b) (a < b ? (uint64_t)a : (uint64_t)b)
47 cperf_latency_test_free(struct cperf_latency_ctx *ctx)
51 rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
52 rte_cryptodev_sym_session_free(ctx->sess);
56 rte_mempool_free(ctx->pool);
64 cperf_latency_test_constructor(struct rte_mempool *sess_mp,
65 struct rte_mempool *sess_priv_mp,
66 uint8_t dev_id, uint16_t qp_id,
67 const struct cperf_options *options,
68 const struct cperf_test_vector *test_vector,
69 const struct cperf_op_fns *op_fns)
71 struct cperf_latency_ctx *ctx = NULL;
72 size_t extra_op_priv_size = sizeof(struct priv_op_data);
74 ctx = rte_malloc(NULL, sizeof(struct cperf_latency_ctx), 0);
81 ctx->populate_ops = op_fns->populate_ops;
82 ctx->options = options;
83 ctx->test_vector = test_vector;
85 /* IV goes at the end of the crypto operation */
86 uint16_t iv_offset = sizeof(struct rte_crypto_op) +
87 sizeof(struct rte_crypto_sym_op) +
88 sizeof(struct cperf_op_result *);
90 ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
91 test_vector, iv_offset);
92 if (ctx->sess == NULL)
95 if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id,
97 &ctx->src_buf_offset, &ctx->dst_buf_offset,
101 ctx->res = rte_malloc(NULL, sizeof(struct cperf_op_result) *
102 ctx->options->total_ops, 0);
104 if (ctx->res == NULL)
109 cperf_latency_test_free(ctx);
115 store_timestamp(struct rte_crypto_op *op, uint64_t timestamp)
117 struct priv_op_data *priv_data;
119 priv_data = (struct priv_op_data *) (op->sym + 1);
120 priv_data->result->status = op->status;
121 priv_data->result->tsc_end = timestamp;
125 cperf_latency_test_runner(void *arg)
127 struct cperf_latency_ctx *ctx = arg;
128 uint16_t test_burst_size;
129 uint8_t burst_size_idx = 0;
130 uint32_t imix_idx = 0;
132 static int only_once;
137 struct rte_crypto_op *ops[ctx->options->max_burst_size];
138 struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
140 struct priv_op_data *priv_data;
142 uint32_t lcore = rte_lcore_id();
144 #ifdef CPERF_LINEARIZATION_ENABLE
145 struct rte_cryptodev_info dev_info;
148 /* Check if source mbufs require coalescing */
149 if (ctx->options->segment_sz < ctx->options->max_buffer_size) {
150 rte_cryptodev_info_get(ctx->dev_id, &dev_info);
151 if ((dev_info.feature_flags &
152 RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) == 0)
155 #endif /* CPERF_LINEARIZATION_ENABLE */
157 ctx->lcore_id = lcore;
159 /* Warm up the host CPU before starting the test */
160 for (i = 0; i < ctx->options->total_ops; i++)
161 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
163 /* Get first size from range or list */
164 if (ctx->options->inc_burst_size != 0)
165 test_burst_size = ctx->options->min_burst_size;
167 test_burst_size = ctx->options->burst_size_list[0];
169 uint16_t iv_offset = sizeof(struct rte_crypto_op) +
170 sizeof(struct rte_crypto_sym_op) +
171 sizeof(struct cperf_op_result *);
173 while (test_burst_size <= ctx->options->max_burst_size) {
174 uint64_t ops_enqd = 0, ops_deqd = 0;
177 uint64_t tsc_val, tsc_end, tsc_start;
178 uint64_t tsc_max = 0, tsc_min = ~0UL, tsc_tot = 0, tsc_idx = 0;
179 uint64_t enqd_max = 0, enqd_min = ~0UL, enqd_tot = 0;
180 uint64_t deqd_max = 0, deqd_min = ~0UL, deqd_tot = 0;
182 while (enqd_tot < ctx->options->total_ops) {
184 uint16_t burst_size = ((enqd_tot + test_burst_size)
185 <= ctx->options->total_ops) ?
187 ctx->options->total_ops -
190 /* Allocate objects containing crypto operations and mbufs */
191 if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
194 "Failed to allocate more crypto operations "
195 "from the crypto operation pool.\n"
196 "Consider increasing the pool size "
201 /* Setup crypto op, attach mbuf etc */
202 (ctx->populate_ops)(ops, ctx->src_buf_offset,
204 burst_size, ctx->sess, ctx->options,
205 ctx->test_vector, iv_offset,
208 tsc_start = rte_rdtsc_precise();
210 #ifdef CPERF_LINEARIZATION_ENABLE
212 /* PMD doesn't support scatter-gather and source buffer
214 * We need to linearize it before enqueuing.
216 for (i = 0; i < burst_size; i++)
217 rte_pktmbuf_linearize(ops[i]->sym->m_src);
219 #endif /* CPERF_LINEARIZATION_ENABLE */
221 /* Enqueue burst of ops on crypto device */
222 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
225 /* Dequeue processed burst of ops from crypto device */
226 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
227 ops_processed, test_burst_size);
229 tsc_end = rte_rdtsc_precise();
231 /* Free memory for not enqueued operations */
232 if (ops_enqd != burst_size)
233 rte_mempool_put_bulk(ctx->pool,
234 (void **)&ops[ops_enqd],
235 burst_size - ops_enqd);
237 for (i = 0; i < ops_enqd; i++) {
238 ctx->res[tsc_idx].tsc_start = tsc_start;
240 * Private data structure starts after the end of the
241 * rte_crypto_sym_op structure.
243 priv_data = (struct priv_op_data *) (ops[i]->sym + 1);
244 priv_data->result = (void *)&ctx->res[tsc_idx];
248 if (likely(ops_deqd)) {
249 /* Free crypto ops so they can be reused. */
250 for (i = 0; i < ops_deqd; i++)
251 store_timestamp(ops_processed[i], tsc_end);
253 rte_mempool_put_bulk(ctx->pool,
254 (void **)ops_processed, ops_deqd);
256 deqd_tot += ops_deqd;
257 deqd_max = max(ops_deqd, deqd_max);
258 deqd_min = min(ops_deqd, deqd_min);
261 enqd_tot += ops_enqd;
262 enqd_max = max(ops_enqd, enqd_max);
263 enqd_min = min(ops_enqd, enqd_min);
268 /* Dequeue any operations still in the crypto device */
269 while (deqd_tot < ctx->options->total_ops) {
270 /* Sending 0 length burst to flush sw crypto device */
271 rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id, NULL, 0);
274 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
275 ops_processed, test_burst_size);
277 tsc_end = rte_rdtsc_precise();
280 for (i = 0; i < ops_deqd; i++)
281 store_timestamp(ops_processed[i], tsc_end);
283 rte_mempool_put_bulk(ctx->pool,
284 (void **)ops_processed, ops_deqd);
286 deqd_tot += ops_deqd;
287 deqd_max = max(ops_deqd, deqd_max);
288 deqd_min = min(ops_deqd, deqd_min);
292 for (i = 0; i < tsc_idx; i++) {
293 tsc_val = ctx->res[i].tsc_end - ctx->res[i].tsc_start;
294 tsc_max = max(tsc_val, tsc_max);
295 tsc_min = min(tsc_val, tsc_min);
299 double time_tot, time_avg, time_max, time_min;
301 const uint64_t tunit = 1000000; /* us */
302 const uint64_t tsc_hz = rte_get_tsc_hz();
304 uint64_t enqd_avg = enqd_tot / b_idx;
305 uint64_t deqd_avg = deqd_tot / b_idx;
306 uint64_t tsc_avg = tsc_tot / tsc_idx;
308 time_tot = tunit*(double)(tsc_tot) / tsc_hz;
309 time_avg = tunit*(double)(tsc_avg) / tsc_hz;
310 time_max = tunit*(double)(tsc_max) / tsc_hz;
311 time_min = tunit*(double)(tsc_min) / tsc_hz;
313 if (ctx->options->csv) {
315 printf("\n# lcore, Buffer Size, Burst Size, Pakt Seq #, "
316 "Packet Size, cycles, time (us)");
318 for (i = 0; i < ctx->options->total_ops; i++) {
320 printf("\n%u;%u;%u;%"PRIu64";%"PRIu64";%.3f",
321 ctx->lcore_id, ctx->options->test_buffer_size,
322 test_burst_size, i + 1,
323 ctx->res[i].tsc_end - ctx->res[i].tsc_start,
324 tunit * (double) (ctx->res[i].tsc_end
325 - ctx->res[i].tsc_start)
331 printf("\n# Device %d on lcore %u\n", ctx->dev_id,
333 printf("\n# total operations: %u", ctx->options->total_ops);
334 printf("\n# Buffer size: %u", ctx->options->test_buffer_size);
335 printf("\n# Burst size: %u", test_burst_size);
336 printf("\n# Number of bursts: %"PRIu64,
340 printf("\n# \t Total\t Average\t "
341 "Maximum\t Minimum");
342 printf("\n# enqueued\t%12"PRIu64"\t%10"PRIu64"\t"
343 "%10"PRIu64"\t%10"PRIu64, enqd_tot,
344 enqd_avg, enqd_max, enqd_min);
345 printf("\n# dequeued\t%12"PRIu64"\t%10"PRIu64"\t"
346 "%10"PRIu64"\t%10"PRIu64, deqd_tot,
347 deqd_avg, deqd_max, deqd_min);
348 printf("\n# cycles\t%12"PRIu64"\t%10"PRIu64"\t"
349 "%10"PRIu64"\t%10"PRIu64, tsc_tot,
350 tsc_avg, tsc_max, tsc_min);
351 printf("\n# time [us]\t%12.0f\t%10.3f\t%10.3f\t%10.3f",
352 time_tot, time_avg, time_max, time_min);
357 /* Get next size from range or list */
358 if (ctx->options->inc_burst_size != 0)
359 test_burst_size += ctx->options->inc_burst_size;
361 if (++burst_size_idx == ctx->options->burst_size_count)
364 ctx->options->burst_size_list[burst_size_idx];
372 cperf_latency_test_destructor(void *arg)
374 struct cperf_latency_ctx *ctx = arg;
379 cperf_latency_test_free(ctx);