xref: /dpdk/app/test-crypto-perf/cperf_test_pmd_cyclecount.c (revision bdce2564dbf78e1fecc0db438b562ae19f0c057c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2017 Intel Corporation
3  */
4 
5 #include <stdbool.h>
6 #include <stdlib.h>
7 
8 #include <rte_crypto.h>
9 #include <rte_cryptodev.h>
10 #include <rte_cycles.h>
11 #include <rte_malloc.h>
12 
13 #include "cperf_ops.h"
14 #include "cperf_test_pmd_cyclecount.h"
15 #include "cperf_test_common.h"
16 
17 #define PRETTY_HDR_FMT "%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n"
18 #define PRETTY_LINE_FMT "%12u%12u%12u%12u%12u%12u%12u%12.0f%12.0f%12.0f\n"
19 #define CSV_HDR_FMT "%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n"
20 #define CSV_LINE_FMT "%10u,%10u,%u,%u,%u,%u,%u,%.3f,%.3f,%.3f\n"
21 
22 struct cperf_pmd_cyclecount_ctx {
23 	uint8_t dev_id;
24 	uint16_t qp_id;
25 	uint8_t lcore_id;
26 
27 	struct rte_mempool *pool;
28 	struct rte_crypto_op **ops;
29 	struct rte_crypto_op **ops_processed;
30 
31 	struct rte_cryptodev_sym_session *sess;
32 
33 	cperf_populate_ops_t populate_ops;
34 
35 	uint32_t src_buf_offset;
36 	uint32_t dst_buf_offset;
37 
38 	const struct cperf_options *options;
39 	const struct cperf_test_vector *test_vector;
40 };
41 
42 struct pmd_cyclecount_state {
43 	struct cperf_pmd_cyclecount_ctx *ctx;
44 	const struct cperf_options *opts;
45 	uint32_t lcore;
46 	uint64_t delay;
47 	int linearize;
48 	uint32_t ops_enqd;
49 	uint32_t ops_deqd;
50 	uint32_t ops_enq_retries;
51 	uint32_t ops_deq_retries;
52 	double cycles_per_build;
53 	double cycles_per_enq;
54 	double cycles_per_deq;
55 };
56 
57 static const uint16_t iv_offset =
58 		sizeof(struct rte_crypto_op) + sizeof(struct rte_crypto_sym_op);
59 
60 static void
61 cperf_pmd_cyclecount_test_free(struct cperf_pmd_cyclecount_ctx *ctx)
62 {
63 	if (!ctx)
64 		return;
65 
66 	if (ctx->sess) {
67 #ifdef RTE_LIB_SECURITY
68 		if (ctx->options->op_type == CPERF_PDCP ||
69 				ctx->options->op_type == CPERF_DOCSIS) {
70 			struct rte_security_ctx *sec_ctx =
71 				(struct rte_security_ctx *)
72 				rte_cryptodev_get_sec_ctx(ctx->dev_id);
73 			rte_security_session_destroy(sec_ctx,
74 				(struct rte_security_session *)ctx->sess);
75 		} else
76 #endif
77 			rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
78 	}
79 
80 	rte_mempool_free(ctx->pool);
81 
82 	rte_free(ctx->ops);
83 
84 	rte_free(ctx->ops_processed);
85 
86 	rte_free(ctx);
87 }
88 
89 void *
90 cperf_pmd_cyclecount_test_constructor(struct rte_mempool *sess_mp,
91 		struct rte_mempool *sess_priv_mp,
92 		uint8_t dev_id, uint16_t qp_id,
93 		const struct cperf_options *options,
94 		const struct cperf_test_vector *test_vector,
95 		const struct cperf_op_fns *op_fns)
96 {
97 	struct cperf_pmd_cyclecount_ctx *ctx = NULL;
98 
99 	/* preallocate buffers for crypto ops as they can get quite big */
100 	size_t alloc_sz = sizeof(struct rte_crypto_op *) *
101 			options->nb_descriptors;
102 
103 	ctx = rte_malloc(NULL, sizeof(struct cperf_pmd_cyclecount_ctx), 0);
104 	if (ctx == NULL)
105 		goto err;
106 
107 	ctx->dev_id = dev_id;
108 	ctx->qp_id = qp_id;
109 
110 	ctx->populate_ops = op_fns->populate_ops;
111 	ctx->options = options;
112 	ctx->test_vector = test_vector;
113 
114 	/* IV goes at the end of the crypto operation */
115 	uint16_t iv_offset = sizeof(struct rte_crypto_op) +
116 			sizeof(struct rte_crypto_sym_op);
117 
118 	ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
119 			test_vector, iv_offset);
120 	if (ctx->sess == NULL)
121 		goto err;
122 
123 	if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
124 			&ctx->src_buf_offset, &ctx->dst_buf_offset,
125 			&ctx->pool) < 0)
126 		goto err;
127 
128 	ctx->ops = rte_malloc("ops", alloc_sz, 0);
129 	if (!ctx->ops)
130 		goto err;
131 
132 	ctx->ops_processed = rte_malloc("ops_processed", alloc_sz, 0);
133 	if (!ctx->ops_processed)
134 		goto err;
135 
136 	return ctx;
137 
138 err:
139 	cperf_pmd_cyclecount_test_free(ctx);
140 
141 	return NULL;
142 }
143 
144 /* benchmark alloc-build-free of ops */
145 static inline int
146 pmd_cyclecount_bench_ops(struct pmd_cyclecount_state *state, uint32_t cur_op,
147 		uint16_t test_burst_size)
148 {
149 	uint32_t iter_ops_left = state->opts->total_ops - cur_op;
150 	uint32_t iter_ops_needed =
151 			RTE_MIN(state->opts->nb_descriptors, iter_ops_left);
152 	uint32_t cur_iter_op;
153 	uint32_t imix_idx = 0;
154 
155 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
156 			cur_iter_op += test_burst_size) {
157 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
158 				test_burst_size);
159 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
160 
161 		/* Allocate objects containing crypto operations and mbufs */
162 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
163 					burst_size) != 0) {
164 			RTE_LOG(ERR, USER1,
165 					"Failed to allocate more crypto operations "
166 					"from the crypto operation pool.\n"
167 					"Consider increasing the pool size "
168 					"with --pool-sz\n");
169 				return -1;
170 		}
171 
172 		/* Setup crypto op, attach mbuf etc */
173 		(state->ctx->populate_ops)(ops,
174 				state->ctx->src_buf_offset,
175 				state->ctx->dst_buf_offset,
176 				burst_size,
177 				state->ctx->sess, state->opts,
178 				state->ctx->test_vector, iv_offset,
179 				&imix_idx, NULL);
180 
181 #ifdef CPERF_LINEARIZATION_ENABLE
182 		/* Check if source mbufs require coalescing */
183 		if (state->linearize) {
184 			uint8_t i;
185 			for (i = 0; i < burst_size; i++) {
186 				struct rte_mbuf *src = ops[i]->sym->m_src;
187 				rte_pktmbuf_linearize(src);
188 			}
189 		}
190 #endif /* CPERF_LINEARIZATION_ENABLE */
191 		rte_mempool_put_bulk(state->ctx->pool, (void **)ops,
192 				burst_size);
193 	}
194 
195 	return 0;
196 }
197 
198 /* allocate and build ops (no free) */
199 static int
200 pmd_cyclecount_build_ops(struct pmd_cyclecount_state *state,
201 		uint32_t iter_ops_needed, uint16_t test_burst_size)
202 {
203 	uint32_t cur_iter_op;
204 	uint32_t imix_idx = 0;
205 
206 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
207 			cur_iter_op += test_burst_size) {
208 		uint32_t burst_size = RTE_MIN(
209 				iter_ops_needed - cur_iter_op, test_burst_size);
210 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
211 
212 		/* Allocate objects containing crypto operations and mbufs */
213 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
214 					burst_size) != 0) {
215 			RTE_LOG(ERR, USER1,
216 					"Failed to allocate more crypto operations "
217 					"from the crypto operation pool.\n"
218 					"Consider increasing the pool size "
219 					"with --pool-sz\n");
220 				return -1;
221 		}
222 
223 		/* Setup crypto op, attach mbuf etc */
224 		(state->ctx->populate_ops)(ops,
225 				state->ctx->src_buf_offset,
226 				state->ctx->dst_buf_offset,
227 				burst_size,
228 				state->ctx->sess, state->opts,
229 				state->ctx->test_vector, iv_offset,
230 				&imix_idx, NULL);
231 	}
232 	return 0;
233 }
234 
235 /* benchmark enqueue, returns number of ops enqueued */
236 static uint32_t
237 pmd_cyclecount_bench_enq(struct pmd_cyclecount_state *state,
238 		uint32_t iter_ops_needed, uint16_t test_burst_size)
239 {
240 	/* Enqueue full descriptor ring of ops on crypto device */
241 	uint32_t cur_iter_op = 0;
242 	while (cur_iter_op < iter_ops_needed) {
243 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
244 				test_burst_size);
245 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
246 		uint32_t burst_enqd;
247 
248 		burst_enqd = rte_cryptodev_enqueue_burst(state->ctx->dev_id,
249 				state->ctx->qp_id, ops, burst_size);
250 
251 		/* if we couldn't enqueue anything, the queue is full */
252 		if (!burst_enqd) {
253 			/* don't try to dequeue anything we didn't enqueue */
254 			return cur_iter_op;
255 		}
256 
257 		if (burst_enqd < burst_size)
258 			state->ops_enq_retries++;
259 		state->ops_enqd += burst_enqd;
260 		cur_iter_op += burst_enqd;
261 	}
262 	return iter_ops_needed;
263 }
264 
265 /* benchmark dequeue */
266 static void
267 pmd_cyclecount_bench_deq(struct pmd_cyclecount_state *state,
268 		uint32_t iter_ops_needed, uint16_t test_burst_size)
269 {
270 	/* Dequeue full descriptor ring of ops on crypto device */
271 	uint32_t cur_iter_op = 0;
272 	while (cur_iter_op < iter_ops_needed) {
273 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
274 				test_burst_size);
275 		struct rte_crypto_op **ops_processed =
276 				&state->ctx->ops[cur_iter_op];
277 		uint32_t burst_deqd;
278 
279 		burst_deqd = rte_cryptodev_dequeue_burst(state->ctx->dev_id,
280 				state->ctx->qp_id, ops_processed, burst_size);
281 
282 		if (burst_deqd < burst_size)
283 			state->ops_deq_retries++;
284 		state->ops_deqd += burst_deqd;
285 		cur_iter_op += burst_deqd;
286 	}
287 }
288 
289 /* run benchmark per burst size */
290 static inline int
291 pmd_cyclecount_bench_burst_sz(
292 		struct pmd_cyclecount_state *state, uint16_t test_burst_size)
293 {
294 	uint64_t tsc_start;
295 	uint64_t tsc_end;
296 	uint64_t tsc_op;
297 	uint64_t tsc_enq;
298 	uint64_t tsc_deq;
299 	uint32_t cur_op;
300 
301 	/* reset all counters */
302 	tsc_enq = 0;
303 	tsc_deq = 0;
304 	state->ops_enqd = 0;
305 	state->ops_enq_retries = 0;
306 	state->ops_deqd = 0;
307 	state->ops_deq_retries = 0;
308 
309 	/*
310 	 * Benchmark crypto op alloc-build-free separately.
311 	 */
312 	tsc_start = rte_rdtsc_precise();
313 
314 	for (cur_op = 0; cur_op < state->opts->total_ops;
315 			cur_op += state->opts->nb_descriptors) {
316 		if (unlikely(pmd_cyclecount_bench_ops(
317 				state, cur_op, test_burst_size)))
318 			return -1;
319 	}
320 
321 	tsc_end = rte_rdtsc_precise();
322 	tsc_op = tsc_end - tsc_start;
323 
324 
325 	/*
326 	 * Hardware acceleration cyclecount benchmarking loop.
327 	 *
328 	 * We're benchmarking raw enq/deq performance by filling up the device
329 	 * queue, so we never get any failed enqs unless the driver won't accept
330 	 * the exact number of descriptors we requested, or the driver won't
331 	 * wrap around the end of the TX ring. However, since we're only
332 	 * dequeuing once we've filled up the queue, we have to benchmark it
333 	 * piecemeal and then average out the results.
334 	 */
335 	cur_op = 0;
336 	while (cur_op < state->opts->total_ops) {
337 		uint32_t iter_ops_left = state->opts->total_ops - cur_op;
338 		uint32_t iter_ops_needed = RTE_MIN(
339 				state->opts->nb_descriptors, iter_ops_left);
340 		uint32_t iter_ops_allocd = iter_ops_needed;
341 
342 		/* allocate and build ops */
343 		if (unlikely(pmd_cyclecount_build_ops(state, iter_ops_needed,
344 				test_burst_size)))
345 			return -1;
346 
347 		tsc_start = rte_rdtsc_precise();
348 
349 		/* fill up TX ring */
350 		iter_ops_needed = pmd_cyclecount_bench_enq(state,
351 				iter_ops_needed, test_burst_size);
352 
353 		tsc_end = rte_rdtsc_precise();
354 
355 		tsc_enq += tsc_end - tsc_start;
356 
357 		/* allow for HW to catch up */
358 		if (state->delay)
359 			rte_delay_us_block(state->delay);
360 
361 		tsc_start = rte_rdtsc_precise();
362 
363 		/* drain RX ring */
364 		pmd_cyclecount_bench_deq(state, iter_ops_needed,
365 				test_burst_size);
366 
367 		tsc_end = rte_rdtsc_precise();
368 
369 		tsc_deq += tsc_end - tsc_start;
370 
371 		cur_op += iter_ops_needed;
372 
373 		/*
374 		 * we may not have processed all ops that we allocated, so
375 		 * free everything we've allocated.
376 		 */
377 		rte_mempool_put_bulk(state->ctx->pool,
378 				(void **)state->ctx->ops, iter_ops_allocd);
379 	}
380 
381 	state->cycles_per_build = (double)tsc_op / state->opts->total_ops;
382 	state->cycles_per_enq = (double)tsc_enq / state->ops_enqd;
383 	state->cycles_per_deq = (double)tsc_deq / state->ops_deqd;
384 
385 	return 0;
386 }
387 
388 int
389 cperf_pmd_cyclecount_test_runner(void *test_ctx)
390 {
391 	struct pmd_cyclecount_state state = {0};
392 	const struct cperf_options *opts;
393 	uint16_t test_burst_size;
394 	uint8_t burst_size_idx = 0;
395 
396 	state.ctx = test_ctx;
397 	opts = state.ctx->options;
398 	state.opts = opts;
399 	state.lcore = rte_lcore_id();
400 	state.linearize = 0;
401 
402 	static uint16_t display_once;
403 	static bool warmup = true;
404 
405 	/*
406 	 * We need a small delay to allow for hardware to process all the crypto
407 	 * operations. We can't automatically figure out what the delay should
408 	 * be, so we leave it up to the user (by default it's 0).
409 	 */
410 	state.delay = 1000 * opts->pmdcc_delay;
411 
412 #ifdef CPERF_LINEARIZATION_ENABLE
413 	struct rte_cryptodev_info dev_info;
414 
415 	/* Check if source mbufs require coalescing */
416 	if (opts->segments_sz < ctx->options->max_buffer_size) {
417 		rte_cryptodev_info_get(state.ctx->dev_id, &dev_info);
418 		if ((dev_info.feature_flags &
419 				    RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) ==
420 				0) {
421 			state.linearize = 1;
422 		}
423 	}
424 #endif /* CPERF_LINEARIZATION_ENABLE */
425 
426 	state.ctx->lcore_id = state.lcore;
427 
428 	/* Get first size from range or list */
429 	if (opts->inc_burst_size != 0)
430 		test_burst_size = opts->min_burst_size;
431 	else
432 		test_burst_size = opts->burst_size_list[0];
433 
434 	while (test_burst_size <= opts->max_burst_size) {
435 		/* do a benchmark run */
436 		if (pmd_cyclecount_bench_burst_sz(&state, test_burst_size))
437 			return -1;
438 
439 		/*
440 		 * First run is always a warm up run.
441 		 */
442 		if (warmup) {
443 			warmup = false;
444 			continue;
445 		}
446 
447 		uint16_t exp = 0;
448 		if (!opts->csv) {
449 			if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0,
450 					__ATOMIC_RELAXED, __ATOMIC_RELAXED))
451 				printf(PRETTY_HDR_FMT, "lcore id", "Buf Size",
452 						"Burst Size", "Enqueued",
453 						"Dequeued", "Enq Retries",
454 						"Deq Retries", "Cycles/Op",
455 						"Cycles/Enq", "Cycles/Deq");
456 
457 			printf(PRETTY_LINE_FMT, state.ctx->lcore_id,
458 					opts->test_buffer_size, test_burst_size,
459 					state.ops_enqd, state.ops_deqd,
460 					state.ops_enq_retries,
461 					state.ops_deq_retries,
462 					state.cycles_per_build,
463 					state.cycles_per_enq,
464 					state.cycles_per_deq);
465 		} else {
466 			if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0,
467 					__ATOMIC_RELAXED, __ATOMIC_RELAXED))
468 				printf(CSV_HDR_FMT, "# lcore id", "Buf Size",
469 						"Burst Size", "Enqueued",
470 						"Dequeued", "Enq Retries",
471 						"Deq Retries", "Cycles/Op",
472 						"Cycles/Enq", "Cycles/Deq");
473 
474 			printf(CSV_LINE_FMT, state.ctx->lcore_id,
475 					opts->test_buffer_size, test_burst_size,
476 					state.ops_enqd, state.ops_deqd,
477 					state.ops_enq_retries,
478 					state.ops_deq_retries,
479 					state.cycles_per_build,
480 					state.cycles_per_enq,
481 					state.cycles_per_deq);
482 		}
483 
484 		/* Get next size from range or list */
485 		if (opts->inc_burst_size != 0)
486 			test_burst_size += opts->inc_burst_size;
487 		else {
488 			if (++burst_size_idx == opts->burst_size_count)
489 				break;
490 			test_burst_size = opts->burst_size_list[burst_size_idx];
491 		}
492 	}
493 
494 	return 0;
495 }
496 
497 void
498 cperf_pmd_cyclecount_test_destructor(void *arg)
499 {
500 	struct cperf_pmd_cyclecount_ctx *ctx = arg;
501 
502 	if (ctx == NULL)
503 		return;
504 
505 	cperf_pmd_cyclecount_test_free(ctx);
506 }
507