xref: /dpdk/app/test-crypto-perf/cperf_test_pmd_cyclecount.c (revision 0899a87ce7c71576d445cd956036f6f7e555f01c)
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 		{
78 			rte_cryptodev_sym_session_clear(ctx->dev_id, ctx->sess);
79 			rte_cryptodev_sym_session_free(ctx->sess);
80 		}
81 	}
82 
83 	rte_mempool_free(ctx->pool);
84 
85 	rte_free(ctx->ops);
86 
87 	rte_free(ctx->ops_processed);
88 
89 	rte_free(ctx);
90 }
91 
92 void *
93 cperf_pmd_cyclecount_test_constructor(struct rte_mempool *sess_mp,
94 		struct rte_mempool *sess_priv_mp,
95 		uint8_t dev_id, uint16_t qp_id,
96 		const struct cperf_options *options,
97 		const struct cperf_test_vector *test_vector,
98 		const struct cperf_op_fns *op_fns)
99 {
100 	struct cperf_pmd_cyclecount_ctx *ctx = NULL;
101 
102 	/* preallocate buffers for crypto ops as they can get quite big */
103 	size_t alloc_sz = sizeof(struct rte_crypto_op *) *
104 			options->nb_descriptors;
105 
106 	ctx = rte_malloc(NULL, sizeof(struct cperf_pmd_cyclecount_ctx), 0);
107 	if (ctx == NULL)
108 		goto err;
109 
110 	ctx->dev_id = dev_id;
111 	ctx->qp_id = qp_id;
112 
113 	ctx->populate_ops = op_fns->populate_ops;
114 	ctx->options = options;
115 	ctx->test_vector = test_vector;
116 
117 	/* IV goes at the end of the crypto operation */
118 	uint16_t iv_offset = sizeof(struct rte_crypto_op) +
119 			sizeof(struct rte_crypto_sym_op);
120 
121 	ctx->sess = op_fns->sess_create(sess_mp, sess_priv_mp, dev_id, options,
122 			test_vector, iv_offset);
123 	if (ctx->sess == NULL)
124 		goto err;
125 
126 	if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
127 			&ctx->src_buf_offset, &ctx->dst_buf_offset,
128 			&ctx->pool) < 0)
129 		goto err;
130 
131 	ctx->ops = rte_malloc("ops", alloc_sz, 0);
132 	if (!ctx->ops)
133 		goto err;
134 
135 	ctx->ops_processed = rte_malloc("ops_processed", alloc_sz, 0);
136 	if (!ctx->ops_processed)
137 		goto err;
138 
139 	return ctx;
140 
141 err:
142 	cperf_pmd_cyclecount_test_free(ctx);
143 
144 	return NULL;
145 }
146 
147 /* benchmark alloc-build-free of ops */
148 static inline int
149 pmd_cyclecount_bench_ops(struct pmd_cyclecount_state *state, uint32_t cur_op,
150 		uint16_t test_burst_size)
151 {
152 	uint32_t iter_ops_left = state->opts->total_ops - cur_op;
153 	uint32_t iter_ops_needed =
154 			RTE_MIN(state->opts->nb_descriptors, iter_ops_left);
155 	uint32_t cur_iter_op;
156 	uint32_t imix_idx = 0;
157 
158 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
159 			cur_iter_op += test_burst_size) {
160 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
161 				test_burst_size);
162 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
163 
164 		/* Allocate objects containing crypto operations and mbufs */
165 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
166 					burst_size) != 0) {
167 			RTE_LOG(ERR, USER1,
168 					"Failed to allocate more crypto operations "
169 					"from the crypto operation pool.\n"
170 					"Consider increasing the pool size "
171 					"with --pool-sz\n");
172 				return -1;
173 		}
174 
175 		/* Setup crypto op, attach mbuf etc */
176 		(state->ctx->populate_ops)(ops,
177 				state->ctx->src_buf_offset,
178 				state->ctx->dst_buf_offset,
179 				burst_size,
180 				state->ctx->sess, state->opts,
181 				state->ctx->test_vector, iv_offset,
182 				&imix_idx, NULL);
183 
184 #ifdef CPERF_LINEARIZATION_ENABLE
185 		/* Check if source mbufs require coalescing */
186 		if (state->linearize) {
187 			uint8_t i;
188 			for (i = 0; i < burst_size; i++) {
189 				struct rte_mbuf *src = ops[i]->sym->m_src;
190 				rte_pktmbuf_linearize(src);
191 			}
192 		}
193 #endif /* CPERF_LINEARIZATION_ENABLE */
194 		rte_mempool_put_bulk(state->ctx->pool, (void **)ops,
195 				burst_size);
196 	}
197 
198 	return 0;
199 }
200 
201 /* allocate and build ops (no free) */
202 static int
203 pmd_cyclecount_build_ops(struct pmd_cyclecount_state *state,
204 		uint32_t iter_ops_needed, uint16_t test_burst_size)
205 {
206 	uint32_t cur_iter_op;
207 	uint32_t imix_idx = 0;
208 
209 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
210 			cur_iter_op += test_burst_size) {
211 		uint32_t burst_size = RTE_MIN(
212 				iter_ops_needed - cur_iter_op, test_burst_size);
213 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
214 
215 		/* Allocate objects containing crypto operations and mbufs */
216 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
217 					burst_size) != 0) {
218 			RTE_LOG(ERR, USER1,
219 					"Failed to allocate more crypto operations "
220 					"from the crypto operation pool.\n"
221 					"Consider increasing the pool size "
222 					"with --pool-sz\n");
223 				return -1;
224 		}
225 
226 		/* Setup crypto op, attach mbuf etc */
227 		(state->ctx->populate_ops)(ops,
228 				state->ctx->src_buf_offset,
229 				state->ctx->dst_buf_offset,
230 				burst_size,
231 				state->ctx->sess, state->opts,
232 				state->ctx->test_vector, iv_offset,
233 				&imix_idx, NULL);
234 	}
235 	return 0;
236 }
237 
238 /* benchmark enqueue, returns number of ops enqueued */
239 static uint32_t
240 pmd_cyclecount_bench_enq(struct pmd_cyclecount_state *state,
241 		uint32_t iter_ops_needed, uint16_t test_burst_size)
242 {
243 	/* Enqueue full descriptor ring of ops on crypto device */
244 	uint32_t cur_iter_op = 0;
245 	while (cur_iter_op < iter_ops_needed) {
246 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
247 				test_burst_size);
248 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
249 		uint32_t burst_enqd;
250 
251 		burst_enqd = rte_cryptodev_enqueue_burst(state->ctx->dev_id,
252 				state->ctx->qp_id, ops, burst_size);
253 
254 		/* if we couldn't enqueue anything, the queue is full */
255 		if (!burst_enqd) {
256 			/* don't try to dequeue anything we didn't enqueue */
257 			return cur_iter_op;
258 		}
259 
260 		if (burst_enqd < burst_size)
261 			state->ops_enq_retries++;
262 		state->ops_enqd += burst_enqd;
263 		cur_iter_op += burst_enqd;
264 	}
265 	return iter_ops_needed;
266 }
267 
268 /* benchmark dequeue */
269 static void
270 pmd_cyclecount_bench_deq(struct pmd_cyclecount_state *state,
271 		uint32_t iter_ops_needed, uint16_t test_burst_size)
272 {
273 	/* Dequeue full descriptor ring of ops on crypto device */
274 	uint32_t cur_iter_op = 0;
275 	while (cur_iter_op < iter_ops_needed) {
276 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
277 				test_burst_size);
278 		struct rte_crypto_op **ops_processed =
279 				&state->ctx->ops[cur_iter_op];
280 		uint32_t burst_deqd;
281 
282 		burst_deqd = rte_cryptodev_dequeue_burst(state->ctx->dev_id,
283 				state->ctx->qp_id, ops_processed, burst_size);
284 
285 		if (burst_deqd < burst_size)
286 			state->ops_deq_retries++;
287 		state->ops_deqd += burst_deqd;
288 		cur_iter_op += burst_deqd;
289 	}
290 }
291 
292 /* run benchmark per burst size */
293 static inline int
294 pmd_cyclecount_bench_burst_sz(
295 		struct pmd_cyclecount_state *state, uint16_t test_burst_size)
296 {
297 	uint64_t tsc_start;
298 	uint64_t tsc_end;
299 	uint64_t tsc_op;
300 	uint64_t tsc_enq;
301 	uint64_t tsc_deq;
302 	uint32_t cur_op;
303 
304 	/* reset all counters */
305 	tsc_enq = 0;
306 	tsc_deq = 0;
307 	state->ops_enqd = 0;
308 	state->ops_enq_retries = 0;
309 	state->ops_deqd = 0;
310 	state->ops_deq_retries = 0;
311 
312 	/*
313 	 * Benchmark crypto op alloc-build-free separately.
314 	 */
315 	tsc_start = rte_rdtsc_precise();
316 
317 	for (cur_op = 0; cur_op < state->opts->total_ops;
318 			cur_op += state->opts->nb_descriptors) {
319 		if (unlikely(pmd_cyclecount_bench_ops(
320 				state, cur_op, test_burst_size)))
321 			return -1;
322 	}
323 
324 	tsc_end = rte_rdtsc_precise();
325 	tsc_op = tsc_end - tsc_start;
326 
327 
328 	/*
329 	 * Hardware acceleration cyclecount benchmarking loop.
330 	 *
331 	 * We're benchmarking raw enq/deq performance by filling up the device
332 	 * queue, so we never get any failed enqs unless the driver won't accept
333 	 * the exact number of descriptors we requested, or the driver won't
334 	 * wrap around the end of the TX ring. However, since we're only
335 	 * dequeuing once we've filled up the queue, we have to benchmark it
336 	 * piecemeal and then average out the results.
337 	 */
338 	cur_op = 0;
339 	while (cur_op < state->opts->total_ops) {
340 		uint32_t iter_ops_left = state->opts->total_ops - cur_op;
341 		uint32_t iter_ops_needed = RTE_MIN(
342 				state->opts->nb_descriptors, iter_ops_left);
343 		uint32_t iter_ops_allocd = iter_ops_needed;
344 
345 		/* allocate and build ops */
346 		if (unlikely(pmd_cyclecount_build_ops(state, iter_ops_needed,
347 				test_burst_size)))
348 			return -1;
349 
350 		tsc_start = rte_rdtsc_precise();
351 
352 		/* fill up TX ring */
353 		iter_ops_needed = pmd_cyclecount_bench_enq(state,
354 				iter_ops_needed, test_burst_size);
355 
356 		tsc_end = rte_rdtsc_precise();
357 
358 		tsc_enq += tsc_end - tsc_start;
359 
360 		/* allow for HW to catch up */
361 		if (state->delay)
362 			rte_delay_us_block(state->delay);
363 
364 		tsc_start = rte_rdtsc_precise();
365 
366 		/* drain RX ring */
367 		pmd_cyclecount_bench_deq(state, iter_ops_needed,
368 				test_burst_size);
369 
370 		tsc_end = rte_rdtsc_precise();
371 
372 		tsc_deq += tsc_end - tsc_start;
373 
374 		cur_op += iter_ops_needed;
375 
376 		/*
377 		 * we may not have processed all ops that we allocated, so
378 		 * free everything we've allocated.
379 		 */
380 		rte_mempool_put_bulk(state->ctx->pool,
381 				(void **)state->ctx->ops, iter_ops_allocd);
382 	}
383 
384 	state->cycles_per_build = (double)tsc_op / state->opts->total_ops;
385 	state->cycles_per_enq = (double)tsc_enq / state->ops_enqd;
386 	state->cycles_per_deq = (double)tsc_deq / state->ops_deqd;
387 
388 	return 0;
389 }
390 
391 int
392 cperf_pmd_cyclecount_test_runner(void *test_ctx)
393 {
394 	struct pmd_cyclecount_state state = {0};
395 	const struct cperf_options *opts;
396 	uint16_t test_burst_size;
397 	uint8_t burst_size_idx = 0;
398 
399 	state.ctx = test_ctx;
400 	opts = state.ctx->options;
401 	state.opts = opts;
402 	state.lcore = rte_lcore_id();
403 	state.linearize = 0;
404 
405 	static uint16_t display_once;
406 	static bool warmup = true;
407 
408 	/*
409 	 * We need a small delay to allow for hardware to process all the crypto
410 	 * operations. We can't automatically figure out what the delay should
411 	 * be, so we leave it up to the user (by default it's 0).
412 	 */
413 	state.delay = 1000 * opts->pmdcc_delay;
414 
415 #ifdef CPERF_LINEARIZATION_ENABLE
416 	struct rte_cryptodev_info dev_info;
417 
418 	/* Check if source mbufs require coalescing */
419 	if (opts->segments_sz < ctx->options->max_buffer_size) {
420 		rte_cryptodev_info_get(state.ctx->dev_id, &dev_info);
421 		if ((dev_info.feature_flags &
422 				    RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) ==
423 				0) {
424 			state.linearize = 1;
425 		}
426 	}
427 #endif /* CPERF_LINEARIZATION_ENABLE */
428 
429 	state.ctx->lcore_id = state.lcore;
430 
431 	/* Get first size from range or list */
432 	if (opts->inc_burst_size != 0)
433 		test_burst_size = opts->min_burst_size;
434 	else
435 		test_burst_size = opts->burst_size_list[0];
436 
437 	while (test_burst_size <= opts->max_burst_size) {
438 		/* do a benchmark run */
439 		if (pmd_cyclecount_bench_burst_sz(&state, test_burst_size))
440 			return -1;
441 
442 		/*
443 		 * First run is always a warm up run.
444 		 */
445 		if (warmup) {
446 			warmup = false;
447 			continue;
448 		}
449 
450 		uint16_t exp = 0;
451 		if (!opts->csv) {
452 			if (__atomic_compare_exchange_n(&display_once, &exp, 1, 0,
453 					__ATOMIC_RELAXED, __ATOMIC_RELAXED))
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 (__atomic_compare_exchange_n(&display_once, &exp, 1, 0,
470 					__ATOMIC_RELAXED, __ATOMIC_RELAXED))
471 				printf(CSV_HDR_FMT, "# lcore id", "Buf Size",
472 						"Burst Size", "Enqueued",
473 						"Dequeued", "Enq Retries",
474 						"Deq Retries", "Cycles/Op",
475 						"Cycles/Enq", "Cycles/Deq");
476 
477 			printf(CSV_LINE_FMT, state.ctx->lcore_id,
478 					opts->test_buffer_size, test_burst_size,
479 					state.ops_enqd, state.ops_deqd,
480 					state.ops_enq_retries,
481 					state.ops_deq_retries,
482 					state.cycles_per_build,
483 					state.cycles_per_enq,
484 					state.cycles_per_deq);
485 		}
486 
487 		/* Get next size from range or list */
488 		if (opts->inc_burst_size != 0)
489 			test_burst_size += opts->inc_burst_size;
490 		else {
491 			if (++burst_size_idx == opts->burst_size_count)
492 				break;
493 			test_burst_size = opts->burst_size_list[burst_size_idx];
494 		}
495 	}
496 
497 	return 0;
498 }
499 
500 void
501 cperf_pmd_cyclecount_test_destructor(void *arg)
502 {
503 	struct cperf_pmd_cyclecount_ctx *ctx = arg;
504 
505 	if (ctx == NULL)
506 		return;
507 
508 	cperf_pmd_cyclecount_test_free(ctx);
509 }
510