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_throughput.h"
11 #include "cperf_ops.h"
12 #include "cperf_test_common.h"
14 struct cperf_throughput_ctx {
19 struct rte_mempool *pool;
21 struct rte_cryptodev_sym_session *sess;
23 cperf_populate_ops_t populate_ops;
25 uint32_t src_buf_offset;
26 uint32_t dst_buf_offset;
28 const struct cperf_options *options;
29 const struct cperf_test_vector *test_vector;
33 cperf_throughput_test_free(struct cperf_throughput_ctx *ctx)
37 rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
38 rte_cryptodev_sym_session_free(ctx->sess);
42 rte_mempool_free(ctx->pool);
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)
55 struct cperf_throughput_ctx *ctx = NULL;
57 ctx = rte_malloc(NULL, sizeof(struct cperf_throughput_ctx), 0);
64 ctx->populate_ops = op_fns->populate_ops;
65 ctx->options = options;
66 ctx->test_vector = test_vector;
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);
72 ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, test_vector,
74 if (ctx->sess == NULL)
77 if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
78 &ctx->src_buf_offset, &ctx->dst_buf_offset,
84 cperf_throughput_test_free(ctx);
90 cperf_throughput_test_runner(void *test_ctx)
92 struct cperf_throughput_ctx *ctx = test_ctx;
93 uint16_t test_burst_size;
94 uint8_t burst_size_idx = 0;
98 struct rte_crypto_op *ops[ctx->options->max_burst_size];
99 struct rte_crypto_op *ops_processed[ctx->options->max_burst_size];
102 uint32_t lcore = rte_lcore_id();
104 #ifdef CPERF_LINEARIZATION_ENABLE
105 struct rte_cryptodev_info dev_info;
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)
115 #endif /* CPERF_LINEARIZATION_ENABLE */
117 ctx->lcore_id = lcore;
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);
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;
127 test_burst_size = ctx->options->burst_size_list[0];
129 uint16_t iv_offset = sizeof(struct rte_crypto_op) +
130 sizeof(struct rte_crypto_sym_op);
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;
136 uint64_t tsc_start, tsc_end, tsc_duration;
138 uint16_t ops_unused = 0;
140 tsc_start = rte_rdtsc_precise();
142 while (ops_enqd_total < ctx->options->total_ops) {
144 uint16_t burst_size = ((ops_enqd_total + test_burst_size)
145 <= ctx->options->total_ops) ?
147 ctx->options->total_ops -
150 uint16_t ops_needed = burst_size - ops_unused;
152 /* Allocate objects containing crypto operations and mbufs */
153 if (rte_mempool_get_bulk(ctx->pool, (void **)ops,
156 "Failed to allocate more crypto operations "
157 "from the crypto operation pool.\n"
158 "Consider increasing the pool size "
163 /* Setup crypto op, attach mbuf etc */
164 (ctx->populate_ops)(ops, ctx->src_buf_offset,
166 ops_needed, ctx->sess,
167 ctx->options, ctx->test_vector,
171 * When ops_needed is smaller than ops_enqd, the
172 * unused ops need to be moved to the front for
175 if (unlikely(ops_enqd > ops_needed)) {
176 size_t nb_b_to_mov = ops_unused * sizeof(
177 struct rte_crypto_op *);
179 memmove(&ops[ops_needed], &ops[ops_enqd],
183 #ifdef CPERF_LINEARIZATION_ENABLE
185 /* PMD doesn't support scatter-gather and source buffer
187 * We need to linearize it before enqueuing.
189 for (i = 0; i < burst_size; i++)
190 rte_pktmbuf_linearize(ops[i]->sym->m_src);
192 #endif /* CPERF_LINEARIZATION_ENABLE */
194 /* Enqueue burst of ops on crypto device */
195 ops_enqd = rte_cryptodev_enqueue_burst(ctx->dev_id, ctx->qp_id,
197 if (ops_enqd < burst_size)
201 * Calculate number of ops not enqueued (mainly for hw
202 * accelerators whose ingress queue can fill up).
204 ops_unused = burst_size - ops_enqd;
205 ops_enqd_total += ops_enqd;
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);
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);
217 ops_deqd_total += ops_deqd;
220 * Count dequeue polls which didn't return any
221 * processed operations. This statistic is mainly
222 * relevant to hw accelerators.
229 /* Dequeue any operations still in the crypto device */
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);
236 ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
237 ops_processed, test_burst_size);
241 rte_mempool_put_bulk(ctx->pool,
242 (void **)ops_processed, ops_deqd);
243 ops_deqd_total += ops_deqd;
247 tsc_end = rte_rdtsc_precise();
248 tsc_duration = (tsc_end - tsc_start);
250 /* Calculate average operations processed per second */
251 double ops_per_second = ((double)ctx->options->total_ops /
252 tsc_duration) * rte_get_tsc_hz();
254 /* Calculate average throughput (Gbps) in bits per second */
255 double throughput_gbps = ((ops_per_second *
256 ctx->options->test_buffer_size * 8) / 1000000000);
258 /* Calculate average cycles per packet */
259 double cycles_per_packet = ((double)tsc_duration /
260 ctx->options->total_ops);
262 if (!ctx->options->csv) {
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",
271 printf("%12u%12u%12u%12"PRIu64"%12"PRIu64"%12"PRIu64
272 "%12"PRIu64"%12.4f%12.4f%12.2f\n",
274 ctx->options->test_buffer_size,
280 ops_per_second/1000000,
285 printf("#lcore id,Buffer Size(B),"
286 "Burst Size,Enqueued,Dequeued,Failed Enq,"
287 "Failed Deq,Ops(Millions),Throughput(Gbps),"
291 printf("%u;%u;%u;%"PRIu64";%"PRIu64";%"PRIu64";%"PRIu64";"
294 ctx->options->test_buffer_size,
300 ops_per_second/1000000,
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;
309 if (++burst_size_idx == ctx->options->burst_size_count)
311 test_burst_size = ctx->options->burst_size_list[burst_size_idx];
321 cperf_throughput_test_destructor(void *arg)
323 struct cperf_throughput_ctx *ctx = arg;
328 cperf_throughput_test_free(ctx);