xref: /dpdk/app/test-crypto-perf/cperf_test_pmd_cyclecount.c (revision 76a48d8ac0928d1a0afbd9f7abfe1d2d9f6d9aaf)
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 	void *sess;
32 	uint8_t sess_owner;
33 
34 	cperf_populate_ops_t populate_ops;
35 
36 	uint32_t src_buf_offset;
37 	uint32_t dst_buf_offset;
38 
39 	const struct cperf_options *options;
40 	const struct cperf_test_vector *test_vector;
41 };
42 
43 struct pmd_cyclecount_state {
44 	struct cperf_pmd_cyclecount_ctx *ctx;
45 	const struct cperf_options *opts;
46 	uint32_t lcore;
47 	uint64_t delay;
48 	int linearize;
49 	uint32_t ops_enqd;
50 	uint32_t ops_deqd;
51 	uint32_t ops_enq_retries;
52 	uint32_t ops_deq_retries;
53 	double cycles_per_build;
54 	double cycles_per_enq;
55 	double cycles_per_deq;
56 };
57 
58 static const uint16_t iv_offset =
59 		sizeof(struct rte_crypto_op) + sizeof(struct rte_crypto_sym_op);
60 
61 static void
62 cperf_pmd_cyclecount_test_free(struct cperf_pmd_cyclecount_ctx *ctx)
63 {
64 	if (!ctx)
65 		return;
66 
67 	if (ctx->sess != NULL && ctx->sess_owner) {
68 #ifdef RTE_LIB_SECURITY
69 		if (ctx->options->op_type == CPERF_PDCP ||
70 				ctx->options->op_type == CPERF_DOCSIS) {
71 			void *sec_ctx = rte_cryptodev_get_sec_ctx(ctx->dev_id);
72 
73 			rte_security_session_destroy(sec_ctx, (void *)ctx->sess);
74 		} else
75 #endif
76 			rte_cryptodev_sym_session_free(ctx->dev_id, ctx->sess);
77 	}
78 
79 	rte_mempool_free(ctx->pool);
80 
81 	rte_free(ctx->ops);
82 
83 	rte_free(ctx->ops_processed);
84 
85 	rte_free(ctx);
86 }
87 
88 void *
89 cperf_pmd_cyclecount_test_constructor(struct rte_mempool *sess_mp,
90 		uint8_t dev_id, uint16_t qp_id,
91 		const struct cperf_options *options,
92 		const struct cperf_test_vector *test_vector,
93 		const struct cperf_op_fns *op_fns,
94 		void **sess)
95 {
96 	struct cperf_pmd_cyclecount_ctx *ctx = NULL;
97 
98 	/* preallocate buffers for crypto ops as they can get quite big */
99 	size_t alloc_sz = sizeof(struct rte_crypto_op *) *
100 			options->nb_descriptors;
101 
102 	ctx = rte_malloc(NULL, sizeof(struct cperf_pmd_cyclecount_ctx), 0);
103 	if (ctx == NULL)
104 		goto err;
105 
106 	ctx->dev_id = dev_id;
107 	ctx->qp_id = qp_id;
108 
109 	ctx->populate_ops = op_fns->populate_ops;
110 	ctx->options = options;
111 	ctx->test_vector = test_vector;
112 
113 	/* IV goes at the end of the crypto operation */
114 	uint16_t iv_offset = sizeof(struct rte_crypto_op) +
115 			sizeof(struct rte_crypto_sym_op);
116 
117 	if (*sess != NULL) {
118 		ctx->sess = *sess;
119 		ctx->sess_owner = false;
120 	} else {
121 		ctx->sess = op_fns->sess_create(sess_mp, dev_id, options, test_vector,
122 			iv_offset);
123 		if (ctx->sess == NULL)
124 			goto err;
125 		*sess = ctx->sess;
126 		ctx->sess_owner = true;
127 	}
128 
129 	if (cperf_alloc_common_memory(options, test_vector, dev_id, qp_id, 0,
130 			&ctx->src_buf_offset, &ctx->dst_buf_offset,
131 			&ctx->pool) < 0)
132 		goto err;
133 
134 	ctx->ops = rte_malloc("ops", alloc_sz, 0);
135 	if (!ctx->ops)
136 		goto err;
137 
138 	ctx->ops_processed = rte_malloc("ops_processed", alloc_sz, 0);
139 	if (!ctx->ops_processed)
140 		goto err;
141 
142 	return ctx;
143 
144 err:
145 	cperf_pmd_cyclecount_test_free(ctx);
146 
147 	return NULL;
148 }
149 
150 /* benchmark alloc-build-free of ops */
151 static inline int
152 pmd_cyclecount_bench_ops(struct pmd_cyclecount_state *state, uint32_t cur_op,
153 		uint16_t test_burst_size)
154 {
155 	uint32_t iter_ops_left = state->opts->total_ops - cur_op;
156 	uint32_t iter_ops_needed =
157 			RTE_MIN(state->opts->nb_descriptors, iter_ops_left);
158 	uint32_t cur_iter_op;
159 	uint32_t imix_idx = 0;
160 
161 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
162 			cur_iter_op += test_burst_size) {
163 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
164 				test_burst_size);
165 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
166 
167 		/* Allocate objects containing crypto operations and mbufs */
168 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
169 					burst_size) != 0) {
170 			RTE_LOG(ERR, USER1,
171 					"Failed to allocate more crypto operations "
172 					"from the crypto operation pool.\n"
173 					"Consider increasing the pool size "
174 					"with --pool-sz\n");
175 				return -1;
176 		}
177 
178 		/* Setup crypto op, attach mbuf etc */
179 		(state->ctx->populate_ops)(ops,
180 				state->ctx->src_buf_offset,
181 				state->ctx->dst_buf_offset,
182 				burst_size,
183 				state->ctx->sess, state->opts,
184 				state->ctx->test_vector, iv_offset,
185 				&imix_idx, NULL);
186 
187 #ifdef CPERF_LINEARIZATION_ENABLE
188 		/* Check if source mbufs require coalescing */
189 		if (state->linearize) {
190 			uint8_t i;
191 			for (i = 0; i < burst_size; i++) {
192 				struct rte_mbuf *src = ops[i]->sym->m_src;
193 				rte_pktmbuf_linearize(src);
194 			}
195 		}
196 #endif /* CPERF_LINEARIZATION_ENABLE */
197 		rte_mempool_put_bulk(state->ctx->pool, (void **)ops,
198 				burst_size);
199 	}
200 
201 	return 0;
202 }
203 
204 /* allocate and build ops (no free) */
205 static int
206 pmd_cyclecount_build_ops(struct pmd_cyclecount_state *state,
207 		uint32_t iter_ops_needed, uint16_t test_burst_size)
208 {
209 	uint32_t cur_iter_op;
210 	uint32_t imix_idx = 0;
211 
212 	for (cur_iter_op = 0; cur_iter_op < iter_ops_needed;
213 			cur_iter_op += test_burst_size) {
214 		uint32_t burst_size = RTE_MIN(
215 				iter_ops_needed - cur_iter_op, test_burst_size);
216 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
217 
218 		/* Allocate objects containing crypto operations and mbufs */
219 		if (rte_mempool_get_bulk(state->ctx->pool, (void **)ops,
220 					burst_size) != 0) {
221 			RTE_LOG(ERR, USER1,
222 					"Failed to allocate more crypto operations "
223 					"from the crypto operation pool.\n"
224 					"Consider increasing the pool size "
225 					"with --pool-sz\n");
226 				return -1;
227 		}
228 
229 		/* Setup crypto op, attach mbuf etc */
230 		(state->ctx->populate_ops)(ops,
231 				state->ctx->src_buf_offset,
232 				state->ctx->dst_buf_offset,
233 				burst_size,
234 				state->ctx->sess, state->opts,
235 				state->ctx->test_vector, iv_offset,
236 				&imix_idx, NULL);
237 	}
238 	return 0;
239 }
240 
241 /* benchmark enqueue, returns number of ops enqueued */
242 static uint32_t
243 pmd_cyclecount_bench_enq(struct pmd_cyclecount_state *state,
244 		uint32_t iter_ops_needed, uint16_t test_burst_size)
245 {
246 	/* Enqueue full descriptor ring of ops on crypto device */
247 	uint32_t cur_iter_op = 0;
248 	while (cur_iter_op < iter_ops_needed) {
249 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
250 				test_burst_size);
251 		struct rte_crypto_op **ops = &state->ctx->ops[cur_iter_op];
252 		uint32_t burst_enqd;
253 
254 		burst_enqd = rte_cryptodev_enqueue_burst(state->ctx->dev_id,
255 				state->ctx->qp_id, ops, burst_size);
256 
257 		/* if we couldn't enqueue anything, the queue is full */
258 		if (!burst_enqd) {
259 			/* don't try to dequeue anything we didn't enqueue */
260 			return cur_iter_op;
261 		}
262 
263 		if (burst_enqd < burst_size)
264 			state->ops_enq_retries++;
265 		state->ops_enqd += burst_enqd;
266 		cur_iter_op += burst_enqd;
267 	}
268 	return iter_ops_needed;
269 }
270 
271 /* benchmark dequeue */
272 static void
273 pmd_cyclecount_bench_deq(struct pmd_cyclecount_state *state,
274 		uint32_t iter_ops_needed, uint16_t test_burst_size)
275 {
276 	/* Dequeue full descriptor ring of ops on crypto device */
277 	uint32_t cur_iter_op = 0;
278 	while (cur_iter_op < iter_ops_needed) {
279 		uint32_t burst_size = RTE_MIN(iter_ops_needed - cur_iter_op,
280 				test_burst_size);
281 		struct rte_crypto_op **ops_processed =
282 				&state->ctx->ops[cur_iter_op];
283 		uint32_t burst_deqd;
284 
285 		burst_deqd = rte_cryptodev_dequeue_burst(state->ctx->dev_id,
286 				state->ctx->qp_id, ops_processed, burst_size);
287 
288 		if (burst_deqd < burst_size)
289 			state->ops_deq_retries++;
290 		state->ops_deqd += burst_deqd;
291 		cur_iter_op += burst_deqd;
292 	}
293 }
294 
295 /* run benchmark per burst size */
296 static inline int
297 pmd_cyclecount_bench_burst_sz(
298 		struct pmd_cyclecount_state *state, uint16_t test_burst_size)
299 {
300 	uint64_t tsc_start;
301 	uint64_t tsc_end;
302 	uint64_t tsc_op;
303 	uint64_t tsc_enq;
304 	uint64_t tsc_deq;
305 	uint32_t cur_op;
306 
307 	/* reset all counters */
308 	tsc_enq = 0;
309 	tsc_deq = 0;
310 	state->ops_enqd = 0;
311 	state->ops_enq_retries = 0;
312 	state->ops_deqd = 0;
313 	state->ops_deq_retries = 0;
314 
315 	/*
316 	 * Benchmark crypto op alloc-build-free separately.
317 	 */
318 	tsc_start = rte_rdtsc_precise();
319 
320 	for (cur_op = 0; cur_op < state->opts->total_ops;
321 			cur_op += state->opts->nb_descriptors) {
322 		if (unlikely(pmd_cyclecount_bench_ops(
323 				state, cur_op, test_burst_size)))
324 			return -1;
325 	}
326 
327 	tsc_end = rte_rdtsc_precise();
328 	tsc_op = tsc_end - tsc_start;
329 
330 
331 	/*
332 	 * Hardware acceleration cyclecount benchmarking loop.
333 	 *
334 	 * We're benchmarking raw enq/deq performance by filling up the device
335 	 * queue, so we never get any failed enqs unless the driver won't accept
336 	 * the exact number of descriptors we requested, or the driver won't
337 	 * wrap around the end of the TX ring. However, since we're only
338 	 * dequeuing once we've filled up the queue, we have to benchmark it
339 	 * piecemeal and then average out the results.
340 	 */
341 	cur_op = 0;
342 	while (cur_op < state->opts->total_ops) {
343 		uint32_t iter_ops_left = state->opts->total_ops - cur_op;
344 		uint32_t iter_ops_needed = RTE_MIN(
345 				state->opts->nb_descriptors, iter_ops_left);
346 		uint32_t iter_ops_allocd = iter_ops_needed;
347 
348 		/* allocate and build ops */
349 		if (unlikely(pmd_cyclecount_build_ops(state, iter_ops_needed,
350 				test_burst_size)))
351 			return -1;
352 
353 		tsc_start = rte_rdtsc_precise();
354 
355 		/* fill up TX ring */
356 		iter_ops_needed = pmd_cyclecount_bench_enq(state,
357 				iter_ops_needed, test_burst_size);
358 
359 		tsc_end = rte_rdtsc_precise();
360 
361 		tsc_enq += tsc_end - tsc_start;
362 
363 		/* allow for HW to catch up */
364 		if (state->delay)
365 			rte_delay_us_block(state->delay);
366 
367 		tsc_start = rte_rdtsc_precise();
368 
369 		/* drain RX ring */
370 		pmd_cyclecount_bench_deq(state, iter_ops_needed,
371 				test_burst_size);
372 
373 		tsc_end = rte_rdtsc_precise();
374 
375 		tsc_deq += tsc_end - tsc_start;
376 
377 		cur_op += iter_ops_needed;
378 
379 		/*
380 		 * we may not have processed all ops that we allocated, so
381 		 * free everything we've allocated.
382 		 */
383 		rte_mempool_put_bulk(state->ctx->pool,
384 				(void **)state->ctx->ops, iter_ops_allocd);
385 	}
386 
387 	state->cycles_per_build = (double)tsc_op / state->opts->total_ops;
388 	state->cycles_per_enq = (double)tsc_enq / state->ops_enqd;
389 	state->cycles_per_deq = (double)tsc_deq / state->ops_deqd;
390 
391 	return 0;
392 }
393 
394 int
395 cperf_pmd_cyclecount_test_runner(void *test_ctx)
396 {
397 	struct pmd_cyclecount_state state = {0};
398 	const struct cperf_options *opts;
399 	uint16_t test_burst_size;
400 	uint8_t burst_size_idx = 0;
401 
402 	state.ctx = test_ctx;
403 	opts = state.ctx->options;
404 	state.opts = opts;
405 	state.lcore = rte_lcore_id();
406 	state.linearize = 0;
407 
408 	static RTE_ATOMIC(uint16_t) display_once;
409 	static bool warmup = true;
410 
411 	/*
412 	 * We need a small delay to allow for hardware to process all the crypto
413 	 * operations. We can't automatically figure out what the delay should
414 	 * be, so we leave it up to the user (by default it's 0).
415 	 */
416 	state.delay = 1000 * opts->pmdcc_delay;
417 
418 #ifdef CPERF_LINEARIZATION_ENABLE
419 	struct rte_cryptodev_info dev_info;
420 
421 	/* Check if source mbufs require coalescing */
422 	if (opts->segments_sz < ctx->options->max_buffer_size) {
423 		rte_cryptodev_info_get(state.ctx->dev_id, &dev_info);
424 		if ((dev_info.feature_flags &
425 				    RTE_CRYPTODEV_FF_MBUF_SCATTER_GATHER) ==
426 				0) {
427 			state.linearize = 1;
428 		}
429 	}
430 #endif /* CPERF_LINEARIZATION_ENABLE */
431 
432 	state.ctx->lcore_id = state.lcore;
433 
434 	/* Get first size from range or list */
435 	if (opts->inc_burst_size != 0)
436 		test_burst_size = opts->min_burst_size;
437 	else
438 		test_burst_size = opts->burst_size_list[0];
439 
440 	while (test_burst_size <= opts->max_burst_size) {
441 		/* do a benchmark run */
442 		if (pmd_cyclecount_bench_burst_sz(&state, test_burst_size))
443 			return -1;
444 
445 		/*
446 		 * First run is always a warm up run.
447 		 */
448 		if (warmup) {
449 			warmup = false;
450 			continue;
451 		}
452 
453 		uint16_t exp = 0;
454 		if (!opts->csv) {
455 			if (rte_atomic_compare_exchange_strong_explicit(&display_once, &exp, 1,
456 					rte_memory_order_relaxed, rte_memory_order_relaxed))
457 				printf(PRETTY_HDR_FMT, "lcore id", "Buf Size",
458 						"Burst Size", "Enqueued",
459 						"Dequeued", "Enq Retries",
460 						"Deq Retries", "Cycles/Op",
461 						"Cycles/Enq", "Cycles/Deq");
462 
463 			printf(PRETTY_LINE_FMT, state.ctx->lcore_id,
464 					opts->test_buffer_size, test_burst_size,
465 					state.ops_enqd, state.ops_deqd,
466 					state.ops_enq_retries,
467 					state.ops_deq_retries,
468 					state.cycles_per_build,
469 					state.cycles_per_enq,
470 					state.cycles_per_deq);
471 		} else {
472 			if (rte_atomic_compare_exchange_strong_explicit(&display_once, &exp, 1,
473 					rte_memory_order_relaxed, rte_memory_order_relaxed))
474 				printf(CSV_HDR_FMT, "# lcore id", "Buf Size",
475 						"Burst Size", "Enqueued",
476 						"Dequeued", "Enq Retries",
477 						"Deq Retries", "Cycles/Op",
478 						"Cycles/Enq", "Cycles/Deq");
479 
480 			printf(CSV_LINE_FMT, state.ctx->lcore_id,
481 					opts->test_buffer_size, test_burst_size,
482 					state.ops_enqd, state.ops_deqd,
483 					state.ops_enq_retries,
484 					state.ops_deq_retries,
485 					state.cycles_per_build,
486 					state.cycles_per_enq,
487 					state.cycles_per_deq);
488 		}
489 
490 		/* Get next size from range or list */
491 		if (opts->inc_burst_size != 0)
492 			test_burst_size += opts->inc_burst_size;
493 		else {
494 			if (++burst_size_idx == opts->burst_size_count)
495 				break;
496 			test_burst_size = opts->burst_size_list[burst_size_idx];
497 		}
498 	}
499 
500 	return 0;
501 }
502 
503 void
504 cperf_pmd_cyclecount_test_destructor(void *arg)
505 {
506 	struct cperf_pmd_cyclecount_ctx *ctx = arg;
507 
508 	if (ctx == NULL)
509 		return;
510 
511 	cperf_pmd_cyclecount_test_free(ctx);
512 }
513