1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright (c) 2021 NVIDIA Corporation & Affiliates
3 */
4
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <string.h>
8 #include <stdint.h>
9 #include <inttypes.h>
10 #include <stdarg.h>
11 #include <errno.h>
12 #include <getopt.h>
13
14 #include <rte_common.h>
15 #include <rte_malloc.h>
16 #include <rte_memory.h>
17 #include <rte_eal.h>
18 #include <rte_ether.h>
19 #include <rte_ethdev.h>
20 #include <rte_mempool.h>
21 #include <rte_mbuf.h>
22
23 #include <rte_gpudev.h>
24
25 enum app_args {
26 ARG_HELP,
27 ARG_MEMPOOL
28 };
29
30 static void
usage(const char * prog_name)31 usage(const char *prog_name)
32 {
33 printf("%s [EAL options] --\n",
34 prog_name);
35 }
36
37 static void
args_parse(int argc,char ** argv)38 args_parse(int argc, char **argv)
39 {
40 char **argvopt;
41 int opt;
42 int opt_idx;
43
44 static struct option lgopts[] = {
45 { "help", 0, 0, ARG_HELP},
46 /* End of options */
47 { 0, 0, 0, 0 }
48 };
49
50 argvopt = argv;
51 while ((opt = getopt_long(argc, argvopt, "",
52 lgopts, &opt_idx)) != EOF) {
53 switch (opt) {
54 case ARG_HELP:
55 usage(argv[0]);
56 break;
57 default:
58 usage(argv[0]);
59 rte_exit(EXIT_FAILURE, "Invalid option: %s\n", argv[optind]);
60 break;
61 }
62 }
63 }
64
65 static int
alloc_gpu_memory(uint16_t gpu_id)66 alloc_gpu_memory(uint16_t gpu_id)
67 {
68 void *ptr_1 = NULL;
69 void *ptr_2 = NULL;
70 size_t buf_bytes = 1024;
71 unsigned int align = 4096;
72 int ret;
73
74 printf("\n=======> TEST: Allocate GPU memory\n\n");
75
76 /* Alloc memory on GPU 0 without any specific alignment */
77 ptr_1 = rte_gpu_mem_alloc(gpu_id, buf_bytes, 0);
78 if (ptr_1 == NULL) {
79 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n");
80 goto error;
81 }
82 printf("GPU memory allocated at 0x%p size is %zd bytes\n",
83 ptr_1, buf_bytes);
84
85 /* Alloc memory on GPU 0 with 4kB alignment */
86 ptr_2 = rte_gpu_mem_alloc(gpu_id, buf_bytes, align);
87 if (ptr_2 == NULL) {
88 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n");
89 goto error;
90 }
91 printf("GPU memory allocated at 0x%p size is %zd bytes\n",
92 ptr_2, buf_bytes);
93
94 if (((uintptr_t)ptr_2) % align) {
95 fprintf(stderr, "Memory address 0x%p is not aligned to %u\n", ptr_2, align);
96 goto error;
97 }
98
99 ret = rte_gpu_mem_free(gpu_id, (uint8_t *)(ptr_1)+0x700);
100 if (ret < 0) {
101 printf("GPU memory 0x%p NOT freed: GPU driver didn't find this memory address internally.\n",
102 (uint8_t *)(ptr_1)+0x700);
103 } else {
104 fprintf(stderr, "ERROR: rte_gpu_mem_free freed GPU memory 0x%p\n",
105 (uint8_t *)(ptr_1)+0x700);
106 goto error;
107 }
108
109 ret = rte_gpu_mem_free(gpu_id, ptr_2);
110 if (ret < 0) {
111 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret);
112 goto error;
113 }
114 printf("GPU memory 0x%p freed\n", ptr_2);
115
116 ret = rte_gpu_mem_free(gpu_id, ptr_1);
117 if (ret < 0) {
118 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret);
119 goto error;
120 }
121 printf("GPU memory 0x%p freed\n", ptr_1);
122
123 printf("\n=======> TEST: PASSED\n");
124 return 0;
125
126 error:
127
128 rte_gpu_mem_free(gpu_id, ptr_1);
129 rte_gpu_mem_free(gpu_id, ptr_2);
130
131 printf("\n=======> TEST: FAILED\n");
132 return -1;
133 }
134
135 static int
register_cpu_memory(uint16_t gpu_id)136 register_cpu_memory(uint16_t gpu_id)
137 {
138 void *ptr = NULL;
139 size_t buf_bytes = 1024;
140 int ret;
141
142 printf("\n=======> TEST: Register CPU memory\n\n");
143
144 /* Alloc memory on CPU visible from GPU 0 */
145 ptr = rte_zmalloc(NULL, buf_bytes, 0);
146 if (ptr == NULL) {
147 fprintf(stderr, "Failed to allocate CPU memory.\n");
148 goto error;
149 }
150
151 ret = rte_gpu_mem_register(gpu_id, buf_bytes, ptr);
152 if (ret < 0) {
153 fprintf(stderr, "rte_gpu_mem_register CPU memory returned error %d\n", ret);
154 goto error;
155 }
156 printf("CPU memory registered at 0x%p %zdB\n", ptr, buf_bytes);
157
158 ret = rte_gpu_mem_unregister(gpu_id, (uint8_t *)(ptr)+0x700);
159 if (ret < 0) {
160 printf("CPU memory 0x%p NOT unregistered: GPU driver didn't find this memory address internally\n",
161 (uint8_t *)(ptr)+0x700);
162 } else {
163 fprintf(stderr, "ERROR: rte_gpu_mem_unregister unregistered GPU memory 0x%p\n",
164 (uint8_t *)(ptr)+0x700);
165 goto error;
166 }
167
168 ret = rte_gpu_mem_unregister(gpu_id, ptr);
169 if (ret < 0) {
170 fprintf(stderr, "rte_gpu_mem_unregister returned error %d\n", ret);
171 goto error;
172 }
173 printf("CPU memory 0x%p unregistered\n", ptr);
174
175 rte_free(ptr);
176
177 printf("\n=======> TEST: PASSED\n");
178 return 0;
179
180 error:
181
182 rte_gpu_mem_unregister(gpu_id, ptr);
183 rte_free(ptr);
184 printf("\n=======> TEST: FAILED\n");
185 return -1;
186 }
187
188 static int
gpu_mem_cpu_map(uint16_t gpu_id)189 gpu_mem_cpu_map(uint16_t gpu_id)
190 {
191 void *ptr_gpu = NULL;
192 void *ptr_cpu = NULL;
193 size_t buf_bytes = 1024;
194 unsigned int align = 4096;
195 int ret;
196
197 printf("\n=======> TEST: Map GPU memory for CPU visibility\n\n");
198
199 /* Alloc memory on GPU 0 with 4kB alignment */
200 ptr_gpu = rte_gpu_mem_alloc(gpu_id, buf_bytes, align);
201 if (ptr_gpu == NULL) {
202 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n");
203 goto error;
204 }
205 printf("GPU memory allocated at 0x%p size is %zd bytes\n",
206 ptr_gpu, buf_bytes);
207
208 ptr_cpu = rte_gpu_mem_cpu_map(gpu_id, buf_bytes, ptr_gpu);
209 if (ptr_cpu == NULL) {
210 fprintf(stderr, "rte_gpu_mem_cpu_map returned error\n");
211 goto error;
212 }
213 printf("GPU memory CPU mapped at 0x%p\n", ptr_cpu);
214
215 ((uint8_t *)ptr_cpu)[0] = 0x4;
216 ((uint8_t *)ptr_cpu)[1] = 0x5;
217 ((uint8_t *)ptr_cpu)[2] = 0x6;
218
219 printf("GPU memory first 3 bytes set from CPU: %x %x %x\n",
220 ((uint8_t *)ptr_cpu)[0],
221 ((uint8_t *)ptr_cpu)[1],
222 ((uint8_t *)ptr_cpu)[2]);
223
224 ret = rte_gpu_mem_cpu_unmap(gpu_id, ptr_gpu);
225 if (ret < 0) {
226 fprintf(stderr, "rte_gpu_mem_cpu_unmap returned error %d\n", ret);
227 goto error;
228 }
229 printf("GPU memory CPU unmapped, 0x%p not valid anymore\n", ptr_cpu);
230
231 ret = rte_gpu_mem_free(gpu_id, ptr_gpu);
232 if (ret < 0) {
233 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret);
234 goto error;
235 }
236 printf("GPU memory 0x%p freed\n", ptr_gpu);
237
238 printf("\n=======> TEST: PASSED\n");
239 return 0;
240
241 error:
242
243 rte_gpu_mem_cpu_unmap(gpu_id, ptr_gpu);
244 rte_gpu_mem_free(gpu_id, ptr_gpu);
245
246 printf("\n=======> TEST: FAILED\n");
247 return -1;
248 }
249
250 static int
create_update_comm_flag(uint16_t gpu_id)251 create_update_comm_flag(uint16_t gpu_id)
252 {
253 struct rte_gpu_comm_flag devflag;
254 int ret = 0;
255 uint32_t set_val;
256 uint32_t get_val;
257
258 printf("\n=======> TEST: Communication flag\n\n");
259
260 ret = rte_gpu_comm_create_flag(gpu_id, &devflag, RTE_GPU_COMM_FLAG_CPU);
261 if (ret < 0) {
262 fprintf(stderr, "rte_gpu_comm_create_flag returned error %d\n", ret);
263 goto error;
264 }
265
266 set_val = 25;
267 ret = rte_gpu_comm_set_flag(&devflag, set_val);
268 if (ret < 0) {
269 fprintf(stderr, "rte_gpu_comm_set_flag returned error %d\n", ret);
270 goto error;
271 }
272
273 ret = rte_gpu_comm_get_flag_value(&devflag, &get_val);
274 if (ret < 0) {
275 fprintf(stderr, "rte_gpu_comm_get_flag_value returned error %d\n", ret);
276 goto error;
277 }
278
279 printf("Communication flag value at 0x%p was set to %d and current value is %d\n",
280 devflag.ptr, set_val, get_val);
281
282 set_val = 38;
283 ret = rte_gpu_comm_set_flag(&devflag, set_val);
284 if (ret < 0) {
285 fprintf(stderr, "rte_gpu_comm_set_flag returned error %d\n", ret);
286 goto error;
287 }
288
289 ret = rte_gpu_comm_get_flag_value(&devflag, &get_val);
290 if (ret < 0) {
291 fprintf(stderr, "rte_gpu_comm_get_flag_value returned error %d\n", ret);
292 goto error;
293 }
294
295 printf("Communication flag value at 0x%p was set to %d and current value is %d\n",
296 devflag.ptr, set_val, get_val);
297
298 ret = rte_gpu_comm_destroy_flag(&devflag);
299 if (ret < 0) {
300 fprintf(stderr, "rte_gpu_comm_destroy_flags returned error %d\n", ret);
301 goto error;
302 }
303
304 printf("\n=======> TEST: PASSED\n");
305 return 0;
306
307 error:
308
309 rte_gpu_comm_destroy_flag(&devflag);
310 printf("\n=======> TEST: FAILED\n");
311 return -1;
312 }
313
314 static int
simulate_gpu_task(struct rte_gpu_comm_list * comm_list_item,int num_pkts)315 simulate_gpu_task(struct rte_gpu_comm_list *comm_list_item, int num_pkts)
316 {
317 int idx;
318
319 if (comm_list_item == NULL)
320 return -1;
321
322 for (idx = 0; idx < num_pkts; idx++) {
323 /**
324 * consume(comm_list_item->pkt_list[idx].addr);
325 */
326 }
327 /*
328 * A real GPU workload function can't directly call rte_gpu_comm_set_status
329 * because it's a CPU-only function.
330 * A real GPU workload should implement the content
331 * of rte_gpu_comm_set_status() in GPU specific code.
332 */
333 rte_gpu_comm_set_status(comm_list_item, RTE_GPU_COMM_LIST_DONE);
334
335 return 0;
336 }
337
338 static int
create_update_comm_list(uint16_t gpu_id)339 create_update_comm_list(uint16_t gpu_id)
340 {
341 int ret = 0;
342 int i = 0;
343 struct rte_gpu_comm_list *comm_list = NULL;
344 uint32_t num_comm_items = 1024;
345 struct rte_mbuf *mbufs[10];
346
347 printf("\n=======> TEST: Communication list\n\n");
348
349 comm_list = rte_gpu_comm_create_list(gpu_id, num_comm_items);
350 if (comm_list == NULL) {
351 fprintf(stderr, "rte_gpu_comm_create_list returned error %d\n", ret);
352 goto error;
353 }
354
355 /**
356 * Simulate DPDK receive functions like rte_eth_rx_burst()
357 */
358 for (i = 0; i < 10; i++) {
359 mbufs[i] = rte_zmalloc(NULL, sizeof(struct rte_mbuf), 0);
360 if (mbufs[i] == NULL) {
361 fprintf(stderr, "Failed to allocate fake mbufs in CPU memory.\n");
362 goto error;
363 }
364
365 memset(mbufs[i], 0, sizeof(struct rte_mbuf));
366 }
367
368 /**
369 * Populate just the first item of the list
370 */
371 ret = rte_gpu_comm_populate_list_pkts(&(comm_list[0]), mbufs, 10);
372 if (ret < 0) {
373 fprintf(stderr, "rte_gpu_comm_populate_list_pkts returned error %d\n", ret);
374 goto error;
375 }
376
377 ret = rte_gpu_comm_cleanup_list(&(comm_list[0]));
378 if (ret == 0) {
379 fprintf(stderr, "rte_gpu_comm_cleanup_list erroneously cleaned the list even if packets have not been consumed yet\n");
380 goto error;
381 }
382 printf("Communication list not cleaned because packets have not been consumed yet.\n");
383
384 /**
385 * Simulate a GPU tasks going through the packet list to consume
386 * mbufs packets and release them
387 */
388 printf("Consuming packets...\n");
389 simulate_gpu_task(&(comm_list[0]), 10);
390
391 /**
392 * Packets have been consumed, now the communication item
393 * and the related mbufs can be all released
394 */
395 ret = rte_gpu_comm_cleanup_list(&(comm_list[0]));
396 if (ret < 0) {
397 fprintf(stderr, "rte_gpu_comm_cleanup_list returned error %d\n", ret);
398 goto error;
399 }
400
401 printf("Communication list cleaned because packets have been consumed now.\n");
402
403 ret = rte_gpu_comm_destroy_list(comm_list, num_comm_items);
404 if (ret < 0) {
405 fprintf(stderr, "rte_gpu_comm_destroy_list returned error %d\n", ret);
406 goto error;
407 }
408
409 for (i = 0; i < 10; i++)
410 rte_free(mbufs[i]);
411
412 printf("\n=======> TEST: PASSED\n");
413 return 0;
414
415 error:
416
417 rte_gpu_comm_destroy_list(comm_list, num_comm_items);
418 for (i = 0; i < 10; i++)
419 rte_free(mbufs[i]);
420 printf("\n=======> TEST: FAILED\n");
421 return -1;
422 }
423
424 int
main(int argc,char ** argv)425 main(int argc, char **argv)
426 {
427 int ret;
428 int nb_gpus = 0;
429 int16_t gpu_id = 0;
430 struct rte_gpu_info ginfo;
431
432 /* Init EAL. */
433 ret = rte_eal_init(argc, argv);
434 if (ret < 0)
435 rte_exit(EXIT_FAILURE, "EAL init failed\n");
436 argc -= ret;
437 argv += ret;
438 if (argc > 1)
439 args_parse(argc, argv);
440 argc -= ret;
441 argv += ret;
442
443 nb_gpus = rte_gpu_count_avail();
444 printf("\n\nDPDK found %d GPUs:\n", nb_gpus);
445 RTE_GPU_FOREACH(gpu_id)
446 {
447 if (rte_gpu_info_get(gpu_id, &ginfo))
448 rte_exit(EXIT_FAILURE, "rte_gpu_info_get error - bye\n");
449
450 printf("\tGPU ID %d\n\t\tparent ID %d GPU Bus ID %s NUMA node %d Tot memory %.02f MB, Tot processors %d\n",
451 ginfo.dev_id,
452 ginfo.parent,
453 ginfo.name,
454 ginfo.numa_node,
455 (((float)ginfo.total_memory)/(float)1024)/(float)1024,
456 ginfo.processor_count
457 );
458 }
459 printf("\n\n");
460
461 if (nb_gpus == 0) {
462 fprintf(stderr, "Need at least one GPU on the system to run the example\n");
463 return EXIT_FAILURE;
464 }
465
466 gpu_id = 0;
467
468 /**
469 * Memory tests
470 */
471 alloc_gpu_memory(gpu_id);
472 register_cpu_memory(gpu_id);
473 gpu_mem_cpu_map(gpu_id);
474
475 /**
476 * Communication items test
477 */
478 create_update_comm_flag(gpu_id);
479 create_update_comm_list(gpu_id);
480
481 /* clean up the EAL */
482 rte_eal_cleanup();
483
484 return EXIT_SUCCESS;
485 }
486