xref: /dpdk/app/test-crypto-perf/cperf_test_pmd_cyclecount.c (revision 3cc6ecfdfe85d2577fef30e1791bb7534e3d60b3)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2017 Intel Corporation
3  */
4 
5 #include <stdbool.h>
6 
7 #include <rte_crypto.h>
8 #include <rte_cryptodev.h>
9 #include <rte_cycles.h>
10 #include <rte_malloc.h>
11 
12 #include "cperf_ops.h"
13 #include "cperf_test_pmd_cyclecount.h"
14 #include "cperf_test_common.h"
15 
16 #define PRETTY_HDR_FMT "%12s%12s%12s%12s%12s%12s%12s%12s%12s%12s\n\n"
17 #define PRETTY_LINE_FMT "%12u%12u%12u%12u%12u%12u%12u%12.0f%12.0f%12.0f\n"
18 #define CSV_HDR_FMT "%s,%s,%s,%s,%s,%s,%s,%s,%s,%s\n"
19 #define CSV_LINE_FMT "%10u;%10u;%u;%u;%u;%u;%u;%.3f;%.3f;%.3f\n"
20 
21 struct cperf_pmd_cyclecount_ctx {
22 	uint8_t dev_id;
23 	uint16_t qp_id;
24 	uint8_t lcore_id;
25 
26 	struct rte_mempool *pool;
27 	struct rte_crypto_op **ops;
28 	struct rte_crypto_op **ops_processed;
29 
30 	struct rte_cryptodev_sym_session *sess;
31 
32 	cperf_populate_ops_t populate_ops;
33 
34 	uint32_t src_buf_offset;
35 	uint32_t dst_buf_offset;
36 
37 	const struct cperf_options *options;
38 	const struct cperf_test_vector *test_vector;
39 };
40 
41 struct pmd_cyclecount_state {
42 	struct cperf_pmd_cyclecount_ctx *ctx;
43 	const struct cperf_options *opts;
44 	uint32_t lcore;
45 	uint64_t delay;
46 	int linearize;
47 	uint32_t ops_enqd;
48 	uint32_t ops_deqd;
49 	uint32_t ops_enq_retries;
50 	uint32_t ops_deq_retries;
51 	double cycles_per_build;
52 	double cycles_per_enq;
53 	double cycles_per_deq;
54 };
55 
56 static const uint16_t iv_offset =
57 		sizeof(struct rte_crypto_op) + sizeof(struct rte_crypto_sym_op);
58 
59 static void
60 cperf_pmd_cyclecount_test_free(struct cperf_pmd_cyclecount_ctx *ctx)
61 {
62 	if (!ctx)
63 		return;
64 
65 	if (ctx->sess) {
66 #ifdef RTE_LIBRTE_SECURITY
67 		if (ctx->options->op_type == CPERF_PDCP ||
68 				ctx->options->op_type == CPERF_DOCSIS) {
69 			struct rte_security_ctx *sec_ctx =
70 				(struct rte_security_ctx *)
71 				rte_cryptodev_get_sec_ctx(ctx->dev_id);
72 			rte_security_session_destroy(sec_ctx,
73 				(struct rte_security_session *)ctx->sess);
74 		} else
75 #endif
76 		{
77 			rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
78 			rte_cryptodev_sym_session_free(ctx->sess);
79 		}
80 	}
81 
82 	if (ctx->pool)
83 		rte_mempool_free(ctx->pool);
84 
85 	if (ctx->ops)
86 		rte_free(ctx->ops);
87 
88 	if (ctx->ops_processed)
89 		rte_free(ctx->ops_processed);
90 
91 	rte_free(ctx);
92 }
93 
94 void *
95 cperf_pmd_cyclecount_test_constructor(struct rte_mempool *sess_mp,
96 		struct rte_mempool *sess_priv_mp,
97 		uint8_t dev_id, uint16_t qp_id,
98 		const struct cperf_options *options,
99 		const struct cperf_test_vector *test_vector,
100 		const struct cperf_op_fns *op_fns)
101 {
102 	struct cperf_pmd_cyclecount_ctx *ctx = NULL;
103 
104 	/* preallocate buffers for crypto ops as they can get quite big */
105 	size_t alloc_sz = sizeof(struct rte_crypto_op *) *
106 			options->nb_descriptors;
107 
108 	ctx = rte_malloc(NULL, sizeof(struct cperf_pmd_cyclecount_ctx), 0);
109 	if (ctx == NULL)
110 		goto err;
111 
112 	ctx->dev_id = dev_id;
113 	ctx->qp_id = qp_id;
114 
115 	ctx->populate_ops = op_fns->populate_ops;
116 	ctx->options = options;
117 	ctx->test_vector = test_vector;
118 
119 	/* IV goes at the end of the crypto operation */
120 	uint16_t iv_offset = sizeof(struct rte_crypto_op) +
121 			sizeof(struct rte_crypto_sym_op);
122 
123 	ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
124 			test_vector, iv_offset);
125 	if (ctx->sess == NULL)
126 		goto err;
127 
128 	if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
129 			&ctx->src_buf_offset, &ctx->dst_buf_offset,
130 			&ctx->pool) < 0)
131 		goto err;
132 
133 	ctx->ops = rte_malloc("ops", alloc_sz, 0);
134 	if (!ctx->ops)
135 		goto err;
136 
137 	ctx->ops_processed = rte_malloc("ops_processed", alloc_sz, 0);
138 	if (!ctx->ops_processed)
139 		goto err;
140 
141 	return ctx;
142 
143 err:
144 	cperf_pmd_cyclecount_test_free(ctx);
145 
146 	return NULL;
147 }
148 
149 /* benchmark alloc-build-free of ops */
150 static inline int
151 pmd_cyclecount_bench_ops(struct pmd_cyclecount_state *state, uint32_t cur_op,
152 		uint16_t test_burst_size)
153 {
154 	uint32_t iter_ops_left = state->opts->total_ops - cur_op;
155 	uint32_t iter_ops_needed =
156 			RTE_MIN(state->opts->nb_descriptors, iter_ops_left);
157 	uint32_t cur_iter_op;
158 	uint32_t imix_idx = 0;
159 
160 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
161 			cur_iter_op += test_burst_size) {
162 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
163 				test_burst_size);
164 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
165 
166 		/* Allocate objects containing crypto operations and mbufs */
167 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
168 					burst_size) != 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 		(state->ctx->populate_ops)(ops,
179 				state->ctx->src_buf_offset,
180 				state->ctx->dst_buf_offset,
181 				burst_size,
182 				state->ctx->sess, state->opts,
183 				state->ctx->test_vector, iv_offset,
184 				&imix_idx);
185 
186 #ifdef CPERF_LINEARIZATION_ENABLE
187 		/* Check if source mbufs require coalescing */
188 		if (state->linearize) {
189 			uint8_t i;
190 			for (i = 0; i < burst_size; i++) {
191 				struct rte_mbuf *src = ops[i]->sym->m_src;
192 				rte_pktmbuf_linearize(src);
193 			}
194 		}
195 #endif /* CPERF_LINEARIZATION_ENABLE */
196 		rte_mempool_put_bulk(state->ctx->pool, (void **)ops,
197 				burst_size);
198 	}
199 
200 	return 0;
201 }
202 
203 /* allocate and build ops (no free) */
204 static int
205 pmd_cyclecount_build_ops(struct pmd_cyclecount_state *state,
206 		uint32_t iter_ops_needed, uint16_t test_burst_size)
207 {
208 	uint32_t cur_iter_op;
209 	uint32_t imix_idx = 0;
210 
211 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
212 			cur_iter_op += test_burst_size) {
213 		uint32_t burst_size = RTE_MIN(
214 				iter_ops_needed - cur_iter_op, test_burst_size);
215 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
216 
217 		/* Allocate objects containing crypto operations and mbufs */
218 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
219 					burst_size) != 0) {
220 			RTE_LOG(ERR, USER1,
221 					"Failed to allocate more crypto operations "
222 					"from the crypto operation pool.\n"
223 					"Consider increasing the pool size "
224 					"with --pool-sz\n");
225 				return -1;
226 		}
227 
228 		/* Setup crypto op, attach mbuf etc */
229 		(state->ctx->populate_ops)(ops,
230 				state->ctx->src_buf_offset,
231 				state->ctx->dst_buf_offset,
232 				burst_size,
233 				state->ctx->sess, state->opts,
234 				state->ctx->test_vector, iv_offset,
235 				&imix_idx);
236 	}
237 	return 0;
238 }
239 
240 /* benchmark enqueue, returns number of ops enqueued */
241 static uint32_t
242 pmd_cyclecount_bench_enq(struct pmd_cyclecount_state *state,
243 		uint32_t iter_ops_needed, uint16_t test_burst_size)
244 {
245 	/* Enqueue full descriptor ring of ops on crypto device */
246 	uint32_t cur_iter_op = 0;
247 	while (cur_iter_op < iter_ops_needed) {
248 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
249 				test_burst_size);
250 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
251 		uint32_t burst_enqd;
252 
253 		burst_enqd = rte_cryptodev_enqueue_burst(state->ctx->dev_id,
254 				state->ctx->qp_id, ops, burst_size);
255 
256 		/* if we couldn't enqueue anything, the queue is full */
257 		if (!burst_enqd) {
258 			/* don't try to dequeue anything we didn't enqueue */
259 			return cur_iter_op;
260 		}
261 
262 		if (burst_enqd < burst_size)
263 			state->ops_enq_retries++;
264 		state->ops_enqd += burst_enqd;
265 		cur_iter_op += burst_enqd;
266 	}
267 	return iter_ops_needed;
268 }
269 
270 /* benchmark dequeue */
271 static void
272 pmd_cyclecount_bench_deq(struct pmd_cyclecount_state *state,
273 		uint32_t iter_ops_needed, uint16_t test_burst_size)
274 {
275 	/* Dequeue full descriptor ring of ops on crypto device */
276 	uint32_t cur_iter_op = 0;
277 	while (cur_iter_op < iter_ops_needed) {
278 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
279 				test_burst_size);
280 		struct rte_crypto_op **ops_processed =
281 				&state->ctx->ops[cur_iter_op];
282 		uint32_t burst_deqd;
283 
284 		burst_deqd = rte_cryptodev_dequeue_burst(state->ctx->dev_id,
285 				state->ctx->qp_id, ops_processed, burst_size);
286 
287 		if (burst_deqd < burst_size)
288 			state->ops_deq_retries++;
289 		state->ops_deqd += burst_deqd;
290 		cur_iter_op += burst_deqd;
291 	}
292 }
293 
294 /* run benchmark per burst size */
295 static inline int
296 pmd_cyclecount_bench_burst_sz(
297 		struct pmd_cyclecount_state *state, uint16_t test_burst_size)
298 {
299 	uint64_t tsc_start;
300 	uint64_t tsc_end;
301 	uint64_t tsc_op;
302 	uint64_t tsc_enq;
303 	uint64_t tsc_deq;
304 	uint32_t cur_op;
305 
306 	/* reset all counters */
307 	tsc_enq = 0;
308 	tsc_deq = 0;
309 	state->ops_enqd = 0;
310 	state->ops_enq_retries = 0;
311 	state->ops_deqd = 0;
312 	state->ops_deq_retries = 0;
313 
314 	/*
315 	 * Benchmark crypto op alloc-build-free separately.
316 	 */
317 	tsc_start = rte_rdtsc_precise();
318 
319 	for (cur_op = 0; cur_op < state->opts->total_ops;
320 			cur_op += state->opts->nb_descriptors) {
321 		if (unlikely(pmd_cyclecount_bench_ops(
322 				state, cur_op, test_burst_size)))
323 			return -1;
324 	}
325 
326 	tsc_end = rte_rdtsc_precise();
327 	tsc_op = tsc_end - tsc_start;
328 
329 
330 	/*
331 	 * Hardware acceleration cyclecount benchmarking loop.
332 	 *
333 	 * We're benchmarking raw enq/deq performance by filling up the device
334 	 * queue, so we never get any failed enqs unless the driver won't accept
335 	 * the exact number of descriptors we requested, or the driver won't
336 	 * wrap around the end of the TX ring. However, since we're only
337 	 * dequeueing once we've filled up the queue, we have to benchmark it
338 	 * piecemeal and then average out the results.
339 	 */
340 	cur_op = 0;
341 	while (cur_op < state->opts->total_ops) {
342 		uint32_t iter_ops_left = state->opts->total_ops - cur_op;
343 		uint32_t iter_ops_needed = RTE_MIN(
344 				state->opts->nb_descriptors, iter_ops_left);
345 		uint32_t iter_ops_allocd = iter_ops_needed;
346 
347 		/* allocate and build ops */
348 		if (unlikely(pmd_cyclecount_build_ops(state, iter_ops_needed,
349 				test_burst_size)))
350 			return -1;
351 
352 		tsc_start = rte_rdtsc_precise();
353 
354 		/* fill up TX ring */
355 		iter_ops_needed = pmd_cyclecount_bench_enq(state,
356 				iter_ops_needed, test_burst_size);
357 
358 		tsc_end = rte_rdtsc_precise();
359 
360 		tsc_enq += tsc_end - tsc_start;
361 
362 		/* allow for HW to catch up */
363 		if (state->delay)
364 			rte_delay_us_block(state->delay);
365 
366 		tsc_start = rte_rdtsc_precise();
367 
368 		/* drain RX ring */
369 		pmd_cyclecount_bench_deq(state, iter_ops_needed,
370 				test_burst_size);
371 
372 		tsc_end = rte_rdtsc_precise();
373 
374 		tsc_deq += tsc_end - tsc_start;
375 
376 		cur_op += iter_ops_needed;
377 
378 		/*
379 		 * we may not have processed all ops that we allocated, so
380 		 * free everything we've allocated.
381 		 */
382 		rte_mempool_put_bulk(state->ctx->pool,
383 				(void **)state->ctx->ops, iter_ops_allocd);
384 	}
385 
386 	state->cycles_per_build = (double)tsc_op / state->opts->total_ops;
387 	state->cycles_per_enq = (double)tsc_enq / state->ops_enqd;
388 	state->cycles_per_deq = (double)tsc_deq / state->ops_deqd;
389 
390 	return 0;
391 }
392 
393 int
394 cperf_pmd_cyclecount_test_runner(void *test_ctx)
395 {
396 	struct pmd_cyclecount_state state = {0};
397 	const struct cperf_options *opts;
398 	uint16_t test_burst_size;
399 	uint8_t burst_size_idx = 0;
400 
401 	state.ctx = test_ctx;
402 	opts = state.ctx->options;
403 	state.opts = opts;
404 	state.lcore = rte_lcore_id();
405 	state.linearize = 0;
406 
407 	static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
408 	static bool warmup = true;
409 
410 	/*
411 	 * We need a small delay to allow for hardware to process all the crypto
412 	 * operations. We can't automatically figure out what the delay should
413 	 * be, so we leave it up to the user (by default it's 0).
414 	 */
415 	state.delay = 1000 * opts->pmdcc_delay;
416 
417 #ifdef CPERF_LINEARIZATION_ENABLE
418 	struct rte_cryptodev_info dev_info;
419 
420 	/* Check if source mbufs require coalescing */
421 	if (opts->segments_sz < ctx->options->max_buffer_size) {
422 		rte_cryptodev_info_get(state.ctx->dev_id, &dev_info);
423 		if ((dev_info.feature_flags &
424 				    RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) ==
425 				0) {
426 			state.linearize = 1;
427 		}
428 	}
429 #endif /* CPERF_LINEARIZATION_ENABLE */
430 
431 	state.ctx->lcore_id = state.lcore;
432 
433 	/* Get first size from range or list */
434 	if (opts->inc_burst_size != 0)
435 		test_burst_size = opts->min_burst_size;
436 	else
437 		test_burst_size = opts->burst_size_list[0];
438 
439 	while (test_burst_size <= opts->max_burst_size) {
440 		/* do a benchmark run */
441 		if (pmd_cyclecount_bench_burst_sz(&state, test_burst_size))
442 			return -1;
443 
444 		/*
445 		 * First run is always a warm up run.
446 		 */
447 		if (warmup) {
448 			warmup = false;
449 			continue;
450 		}
451 
452 		if (!opts->csv) {
453 			if (rte_atomic16_test_and_set(&display_once))
454 				printf(PRETTY_HDR_FMT, "lcore id", "Buf Size",
455 						"Burst Size", "Enqueued",
456 						"Dequeued", "Enq Retries",
457 						"Deq Retries", "Cycles/Op",
458 						"Cycles/Enq", "Cycles/Deq");
459 
460 			printf(PRETTY_LINE_FMT, state.ctx->lcore_id,
461 					opts->test_buffer_size, test_burst_size,
462 					state.ops_enqd, state.ops_deqd,
463 					state.ops_enq_retries,
464 					state.ops_deq_retries,
465 					state.cycles_per_build,
466 					state.cycles_per_enq,
467 					state.cycles_per_deq);
468 		} else {
469 			if (rte_atomic16_test_and_set(&display_once))
470 				printf(CSV_HDR_FMT, "# lcore id", "Buf Size",
471 						"Burst Size", "Enqueued",
472 						"Dequeued", "Enq Retries",
473 						"Deq Retries", "Cycles/Op",
474 						"Cycles/Enq", "Cycles/Deq");
475 
476 			printf(CSV_LINE_FMT, state.ctx->lcore_id,
477 					opts->test_buffer_size, test_burst_size,
478 					state.ops_enqd, state.ops_deqd,
479 					state.ops_enq_retries,
480 					state.ops_deq_retries,
481 					state.cycles_per_build,
482 					state.cycles_per_enq,
483 					state.cycles_per_deq);
484 		}
485 
486 		/* Get next size from range or list */
487 		if (opts->inc_burst_size != 0)
488 			test_burst_size += opts->inc_burst_size;
489 		else {
490 			if (++burst_size_idx == opts->burst_size_count)
491 				break;
492 			test_burst_size = opts->burst_size_list[burst_size_idx];
493 		}
494 	}
495 
496 	return 0;
497 }
498 
499 void
500 cperf_pmd_cyclecount_test_destructor(void *arg)
501 {
502 	struct cperf_pmd_cyclecount_ctx *ctx = arg;
503 
504 	if (ctx == NULL)
505 		return;
506 
507 	cperf_pmd_cyclecount_test_free(ctx);
508 }
509