xref: /spdk/examples/accel/perf/accel_perf.c (revision cfb65ba61180ae80899756239e08f4656866beb3)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include "spdk/stdinc.h"
35 #include "spdk/thread.h"
36 #include "spdk/env.h"
37 #include "spdk/event.h"
38 #include "spdk/log.h"
39 #include "spdk/string.h"
40 #include "spdk/accel_engine.h"
41 
42 static uint64_t	g_tsc_rate;
43 static uint64_t g_tsc_us_rate;
44 static uint64_t g_tsc_end;
45 static int g_xfer_size_bytes = 4096;
46 static int g_queue_depth = 32;
47 static int g_time_in_sec = 5;
48 static bool g_verify = false;
49 static const char *g_workload_type = NULL;
50 static enum accel_capability g_workload_selection;
51 static struct worker_thread *g_workers = NULL;
52 static int g_num_workers = 0;
53 static pthread_mutex_t g_workers_lock = PTHREAD_MUTEX_INITIALIZER;
54 
55 struct worker_thread {
56 	struct spdk_io_channel		*ch;
57 	uint64_t			xfer_completed;
58 	uint64_t			xfer_failed;
59 	uint64_t			current_queue_depth;
60 	struct spdk_mempool		*data_pool;
61 	struct spdk_mempool		*task_pool;
62 	struct worker_thread		*next;
63 	unsigned			core;
64 	struct spdk_thread		*thread;
65 	bool				is_draining;
66 	struct spdk_poller		*is_draining_poller;
67 	struct spdk_poller		*stop_poller;
68 };
69 
70 struct ap_task {
71 	void			*src;
72 	void			*dst;
73 	struct worker_thread	*worker;
74 };
75 
76 inline static struct ap_task *
77 __ap_task_from_accel_task(struct spdk_accel_task *at)
78 {
79 	return (struct ap_task *)((uintptr_t)at - sizeof(struct ap_task));
80 }
81 
82 inline static struct spdk_accel_task *
83 __accel_task_from_ap_task(struct ap_task *ap)
84 {
85 	return (struct spdk_accel_task *)((uintptr_t)ap + sizeof(struct ap_task));
86 }
87 
88 static void
89 dump_user_config(struct spdk_app_opts *opts)
90 {
91 	printf("SPDK Configuration:\n");
92 	printf("Core mask:      %s\n\n", opts->reactor_mask);
93 	printf("Accel Perf Configuration:\n");
94 	printf("Workload Type:  %s\n", g_workload_type);
95 	printf("Transfer size:  %u bytes\n", g_xfer_size_bytes);
96 	printf("Queue depth:    %u\n", g_queue_depth);
97 	printf("Run time:       %u seconds\n", g_time_in_sec);
98 	printf("Verify:         %s\n\n", g_verify ? "Yes" : "No");
99 }
100 
101 static void
102 usage(void)
103 {
104 	printf("accel_perf options:\n");
105 	printf("\t[-h help message]\n");
106 	printf("\t[-q queue depth]\n");
107 	printf("\t[-n number of channels]\n");
108 	printf("\t[-o transfer size in bytes]\n");
109 	printf("\t[-t time in seconds]\n");
110 	printf("\t[-w workload type must be one of these: copy, fill\n");
111 	printf("\t[-y verify result if this switch is on]\n");
112 }
113 
114 static int
115 parse_args(int argc, char *argv)
116 {
117 	switch (argc) {
118 	case 'o':
119 		g_xfer_size_bytes = spdk_strtol(optarg, 10);
120 		break;
121 	case 'q':
122 		g_queue_depth = spdk_strtol(optarg, 10);
123 		break;
124 	case 't':
125 		g_time_in_sec = spdk_strtol(optarg, 10);
126 		break;
127 	case 'y':
128 		g_verify = true;
129 		break;
130 	case 'w':
131 		g_workload_type = optarg;
132 		if (!strcmp(g_workload_type, "copy")) {
133 			g_workload_selection = ACCEL_COPY;
134 		} else if (!strcmp(g_workload_type, "fill")) {
135 			g_workload_selection = ACCEL_FILL;
136 		}
137 		break;
138 	default:
139 		usage();
140 		return 1;
141 	}
142 	return 0;
143 }
144 
145 static void
146 unregister_worker(void *arg1)
147 {
148 	struct worker_thread *worker = arg1;
149 
150 	spdk_mempool_free(worker->data_pool);
151 	spdk_mempool_free(worker->task_pool);
152 	spdk_put_io_channel(worker->ch);
153 	pthread_mutex_lock(&g_workers_lock);
154 	assert(g_num_workers >= 1);
155 	if (--g_num_workers == 0) {
156 		pthread_mutex_unlock(&g_workers_lock);
157 		spdk_app_stop(0);
158 	}
159 	pthread_mutex_unlock(&g_workers_lock);
160 }
161 
162 static void accel_done(void *ref, int status);
163 
164 static void
165 _submit_single(void *arg1, void *arg2)
166 {
167 	struct worker_thread *worker = arg1;
168 	struct ap_task *task = arg2;
169 
170 	assert(worker);
171 
172 	if (g_verify) {
173 		memset(task->src, 0x5a, g_xfer_size_bytes);
174 		memset(task->dst, 0xa5, g_xfer_size_bytes);
175 	}
176 	task->worker = worker;
177 	task->worker->current_queue_depth++;
178 	if (!strcmp(g_workload_type, "copy")) {
179 		spdk_accel_submit_copy(__accel_task_from_ap_task(task),
180 				       worker->ch, task->dst,
181 				       task->src, g_xfer_size_bytes, accel_done);
182 	} else if (!strcmp(g_workload_type, "fill")) {
183 		/* For fill use the first byte of the task->dst buffer */
184 		spdk_accel_submit_fill(__accel_task_from_ap_task(task),
185 				       worker->ch, task->dst, *(uint8_t *)task->src,
186 				       g_xfer_size_bytes, accel_done);
187 	} else {
188 		assert(false);
189 	}
190 }
191 
192 static void
193 _accel_done(void *arg1)
194 {
195 	struct ap_task *task = arg1;
196 	struct worker_thread *worker = task->worker;
197 
198 	assert(worker);
199 	assert(worker->current_queue_depth > 0);
200 
201 	if (g_verify) {
202 		if (memcmp(task->src, task->dst, g_xfer_size_bytes)) {
203 			SPDK_NOTICELOG("Data miscompare\n");
204 			worker->xfer_failed++;
205 			/* TODO: cleanup */
206 			exit(-1);
207 		}
208 	}
209 	worker->xfer_completed++;
210 	worker->current_queue_depth--;
211 
212 	if (!worker->is_draining) {
213 		_submit_single(worker, task);
214 	} else {
215 		spdk_mempool_put(worker->data_pool, task->src);
216 		spdk_mempool_put(worker->data_pool, task->dst);
217 		spdk_mempool_put(worker->task_pool, task);
218 	}
219 }
220 
221 static int
222 dump_result(void)
223 {
224 	uint64_t total_completed = 0;
225 	uint64_t total_failed = 0;
226 	uint64_t total_xfer_per_sec, total_bw_in_MiBps;
227 	struct worker_thread *worker = g_workers;
228 
229 	printf("\nCore           Transfers     Bandwidth     Failed\n");
230 	printf("-------------------------------------------------\n");
231 	while (worker != NULL) {
232 
233 		uint64_t xfer_per_sec = worker->xfer_completed / g_time_in_sec;
234 		uint64_t bw_in_MiBps = (worker->xfer_completed * g_xfer_size_bytes) /
235 				       (g_time_in_sec * 1024 * 1024);
236 
237 		total_completed += worker->xfer_completed;
238 		total_failed += worker->xfer_failed;
239 
240 		if (xfer_per_sec) {
241 			printf("%10d%12" PRIu64 "/s%8" PRIu64 " MiB/s%11" PRIu64 "\n",
242 			       worker->core, xfer_per_sec,
243 			       bw_in_MiBps, worker->xfer_failed);
244 		}
245 
246 		worker = worker->next;
247 	}
248 
249 	total_xfer_per_sec = total_completed / g_time_in_sec;
250 	total_bw_in_MiBps = (total_completed * g_xfer_size_bytes) /
251 			    (g_time_in_sec * 1024 * 1024);
252 
253 	printf("=================================================\n");
254 	printf("Total:%16" PRIu64 "/s%8" PRIu64 " MiB/s%11" PRIu64 "\n\n",
255 	       total_xfer_per_sec, total_bw_in_MiBps, total_failed);
256 
257 	return total_failed ? 1 : 0;
258 }
259 
260 static int
261 _check_draining(void *arg)
262 {
263 	struct worker_thread *worker = arg;
264 
265 	assert(worker);
266 
267 	if (worker->current_queue_depth == 0) {
268 		spdk_poller_unregister(&worker->is_draining_poller);
269 		unregister_worker(worker);
270 	}
271 
272 	return -1;
273 }
274 
275 static int
276 _worker_stop(void *arg)
277 {
278 	struct worker_thread *worker = arg;
279 
280 	assert(worker);
281 
282 	spdk_poller_unregister(&worker->stop_poller);
283 
284 	/* now let the worker drain and check it's outstanding IO with a poller */
285 	worker->is_draining = true;
286 	worker->is_draining_poller = SPDK_POLLER_REGISTER(_check_draining, worker, 0);
287 
288 	return 0;
289 }
290 
291 static void
292 _init_thread_done(void *ctx)
293 {
294 }
295 
296 static void
297 _init_thread(void *arg1)
298 {
299 	struct worker_thread *worker;
300 	char buf_pool_name[30], task_pool_name[30];
301 	struct ap_task *task;
302 	int i;
303 
304 	worker = calloc(1, sizeof(*worker));
305 	if (worker == NULL) {
306 		fprintf(stderr, "Unable to allocate worker\n");
307 		return;
308 	}
309 
310 	worker->core = spdk_env_get_current_core();
311 	worker->thread = spdk_get_thread();
312 	worker->next = g_workers;
313 	worker->ch = spdk_accel_engine_get_io_channel();
314 	snprintf(buf_pool_name, sizeof(buf_pool_name), "buf_pool_%d", g_num_workers);
315 	snprintf(task_pool_name, sizeof(task_pool_name), "task_pool_%d", g_num_workers);
316 	worker->data_pool = spdk_mempool_create(buf_pool_name,
317 						g_queue_depth * 2, /* src + dst */
318 						g_xfer_size_bytes,
319 						SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
320 						SPDK_ENV_SOCKET_ID_ANY);
321 	worker->task_pool = spdk_mempool_create(task_pool_name,
322 						g_queue_depth,
323 						spdk_accel_task_size() + sizeof(struct ap_task),
324 						SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
325 						SPDK_ENV_SOCKET_ID_ANY);
326 	if (!worker->data_pool || !worker->task_pool) {
327 		fprintf(stderr, "Could not allocate buffer pool.\n");
328 		spdk_mempool_free(worker->data_pool);
329 		spdk_mempool_free(worker->task_pool);
330 		free(worker);
331 		return;
332 	}
333 
334 	/* Register a poller that will stop the worker at time elapsed */
335 	worker->stop_poller = SPDK_POLLER_REGISTER(_worker_stop, worker,
336 			      g_time_in_sec * 1000000ULL);
337 
338 	g_workers = worker;
339 	pthread_mutex_lock(&g_workers_lock);
340 	g_num_workers++;
341 	pthread_mutex_unlock(&g_workers_lock);
342 
343 	for (i = 0; i < g_queue_depth; i++) {
344 		task = spdk_mempool_get(worker->task_pool);
345 		if (!task) {
346 			fprintf(stderr, "Unable to get accel_task\n");
347 			return;
348 		}
349 		task->src = spdk_mempool_get(worker->data_pool);
350 		task->dst = spdk_mempool_get(worker->data_pool);
351 		_submit_single(worker, task);
352 	}
353 }
354 
355 static void
356 accel_done(void *ref, int status)
357 {
358 	struct ap_task *task = __ap_task_from_accel_task(ref);
359 	struct worker_thread *worker = task->worker;
360 
361 	assert(worker);
362 
363 	spdk_thread_send_msg(worker->thread, _accel_done, task);
364 }
365 
366 static void
367 accel_perf_start(void *arg1)
368 {
369 	uint64_t capabilites;
370 	struct spdk_io_channel *accel_ch;
371 
372 	accel_ch = spdk_accel_engine_get_io_channel();
373 	capabilites = spdk_accel_get_capabilities(accel_ch);
374 	spdk_put_io_channel(accel_ch);
375 
376 	if ((capabilites & g_workload_selection) != g_workload_selection) {
377 		SPDK_ERRLOG("Selected workload is not supported by the current engine\n");
378 		SPDK_NOTICELOG("Software engine is selected by default, enable a HW engine via RPC\n\n");
379 		spdk_app_stop(-1);
380 		return;
381 	}
382 
383 	g_tsc_rate = spdk_get_ticks_hz();
384 	g_tsc_us_rate = g_tsc_rate / (1000 * 1000);
385 	g_tsc_end = spdk_get_ticks() + g_time_in_sec * g_tsc_rate;
386 
387 	printf("Running for %d seconds...\n", g_time_in_sec);
388 	fflush(stdout);
389 
390 	spdk_for_each_thread(_init_thread, NULL, _init_thread_done);
391 }
392 
393 int
394 main(int argc, char **argv)
395 {
396 	struct spdk_app_opts opts = {};
397 	struct worker_thread *worker, *tmp;
398 	int rc = 0;
399 
400 	pthread_mutex_init(&g_workers_lock, NULL);
401 	spdk_app_opts_init(&opts);
402 	opts.reactor_mask = "0x1";
403 	if ((rc = spdk_app_parse_args(argc, argv, &opts, "o:q:t:yw:", NULL, parse_args,
404 				      usage)) != SPDK_APP_PARSE_ARGS_SUCCESS) {
405 		rc = -1;
406 		goto cleanup;
407 	}
408 
409 	if (g_workload_type == NULL ||
410 	    (strcmp(g_workload_type, "copy") &&
411 	     strcmp(g_workload_type, "fill"))) {
412 		usage();
413 		rc = -1;
414 		goto cleanup;
415 	}
416 
417 	dump_user_config(&opts);
418 	rc = spdk_app_start(&opts, accel_perf_start, NULL);
419 	if (rc) {
420 		SPDK_ERRLOG("ERROR starting application\n");
421 	} else {
422 		dump_result();
423 	}
424 
425 	pthread_mutex_destroy(&g_workers_lock);
426 
427 	worker = g_workers;
428 	while (worker) {
429 		tmp = worker->next;
430 		free(worker);
431 		worker = tmp;
432 	}
433 cleanup:
434 	spdk_app_fini();
435 	return rc;
436 }
437