1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright (c) 2021 NVIDIA Corporation & Affiliates
14 #include <rte_common.h>
15 #include <rte_malloc.h>
16 #include <rte_memory.h>
18 #include <rte_ether.h>
19 #include <rte_ethdev.h>
20 #include <rte_mempool.h>
23 #include <rte_gpudev.h>
31 usage(const char *prog_name)
33 printf("%s [EAL options] --\n",
38 args_parse(int argc, char **argv)
44 static struct option lgopts[] = {
45 { "help", 0, 0, ARG_HELP},
51 while ((opt = getopt_long(argc, argvopt, "",
52 lgopts, &opt_idx)) != EOF) {
59 rte_exit(EXIT_FAILURE, "Invalid option: %s\n", argv[optind]);
66 alloc_gpu_memory(uint16_t gpu_id)
70 size_t buf_bytes = 1024;
71 unsigned int align = 4096;
74 printf("\n=======> TEST: Allocate GPU memory\n\n");
76 /* Alloc memory on GPU 0 without any specific alignment */
77 ptr_1 = rte_gpu_mem_alloc(gpu_id, buf_bytes, 0);
79 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n");
82 printf("GPU memory allocated at 0x%p size is %zd bytes\n",
85 /* Alloc memory on GPU 0 with 4kB alignment */
86 ptr_2 = rte_gpu_mem_alloc(gpu_id, buf_bytes, align);
88 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n");
91 printf("GPU memory allocated at 0x%p size is %zd bytes\n",
94 if (((uintptr_t)ptr_2) % align) {
95 fprintf(stderr, "Memory address 0x%p is not aligned to %u\n", ptr_2, align);
99 ret = rte_gpu_mem_free(gpu_id, (uint8_t *)(ptr_1)+0x700);
101 printf("GPU memory 0x%p NOT freed: GPU driver didn't find this memory address internally.\n",
102 (uint8_t *)(ptr_1)+0x700);
104 fprintf(stderr, "ERROR: rte_gpu_mem_free freed GPU memory 0x%p\n",
105 (uint8_t *)(ptr_1)+0x700);
109 ret = rte_gpu_mem_free(gpu_id, ptr_2);
111 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret);
114 printf("GPU memory 0x%p freed\n", ptr_2);
116 ret = rte_gpu_mem_free(gpu_id, ptr_1);
118 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret);
121 printf("GPU memory 0x%p freed\n", ptr_1);
123 printf("\n=======> TEST: PASSED\n");
128 rte_gpu_mem_free(gpu_id, ptr_1);
129 rte_gpu_mem_free(gpu_id, ptr_2);
131 printf("\n=======> TEST: FAILED\n");
136 register_cpu_memory(uint16_t gpu_id)
139 size_t buf_bytes = 1024;
142 printf("\n=======> TEST: Register CPU memory\n\n");
144 /* Alloc memory on CPU visible from GPU 0 */
145 ptr = rte_zmalloc(NULL, buf_bytes, 0);
147 fprintf(stderr, "Failed to allocate CPU memory.\n");
151 ret = rte_gpu_mem_register(gpu_id, buf_bytes, ptr);
153 fprintf(stderr, "rte_gpu_mem_register CPU memory returned error %d\n", ret);
156 printf("CPU memory registered at 0x%p %zdB\n", ptr, buf_bytes);
158 ret = rte_gpu_mem_unregister(gpu_id, (uint8_t *)(ptr)+0x700);
160 printf("CPU memory 0x%p NOT unregistered: GPU driver didn't find this memory address internally\n",
161 (uint8_t *)(ptr)+0x700);
163 fprintf(stderr, "ERROR: rte_gpu_mem_unregister unregistered GPU memory 0x%p\n",
164 (uint8_t *)(ptr)+0x700);
168 ret = rte_gpu_mem_unregister(gpu_id, ptr);
170 fprintf(stderr, "rte_gpu_mem_unregister returned error %d\n", ret);
173 printf("CPU memory 0x%p unregistered\n", ptr);
177 printf("\n=======> TEST: PASSED\n");
182 rte_gpu_mem_unregister(gpu_id, ptr);
184 printf("\n=======> TEST: FAILED\n");
189 gpu_mem_cpu_map(uint16_t gpu_id)
191 void *ptr_gpu = NULL;
192 void *ptr_cpu = NULL;
193 size_t buf_bytes = 1024;
194 unsigned int align = 4096;
197 printf("\n=======> TEST: Map GPU memory for CPU visibility\n\n");
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");
205 printf("GPU memory allocated at 0x%p size is %zd bytes\n",
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");
213 printf("GPU memory CPU mapped at 0x%p\n", ptr_cpu);
215 ((uint8_t *)ptr_cpu)[0] = 0x4;
216 ((uint8_t *)ptr_cpu)[1] = 0x5;
217 ((uint8_t *)ptr_cpu)[2] = 0x6;
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]);
224 ret = rte_gpu_mem_cpu_unmap(gpu_id, ptr_gpu);
226 fprintf(stderr, "rte_gpu_mem_cpu_unmap returned error %d\n", ret);
229 printf("GPU memory CPU unmapped, 0x%p not valid anymore\n", ptr_cpu);
231 ret = rte_gpu_mem_free(gpu_id, ptr_gpu);
233 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret);
236 printf("GPU memory 0x%p freed\n", ptr_gpu);
238 printf("\n=======> TEST: PASSED\n");
243 rte_gpu_mem_cpu_unmap(gpu_id, ptr_gpu);
244 rte_gpu_mem_free(gpu_id, ptr_gpu);
246 printf("\n=======> TEST: FAILED\n");
251 create_update_comm_flag(uint16_t gpu_id)
253 struct rte_gpu_comm_flag devflag;
258 printf("\n=======> TEST: Communication flag\n\n");
260 ret = rte_gpu_comm_create_flag(gpu_id, &devflag, RTE_GPU_COMM_FLAG_CPU);
262 fprintf(stderr, "rte_gpu_comm_create_flag returned error %d\n", ret);
267 ret = rte_gpu_comm_set_flag(&devflag, set_val);
269 fprintf(stderr, "rte_gpu_comm_set_flag returned error %d\n", ret);
273 ret = rte_gpu_comm_get_flag_value(&devflag, &get_val);
275 fprintf(stderr, "rte_gpu_comm_get_flag_value returned error %d\n", ret);
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);
283 ret = rte_gpu_comm_set_flag(&devflag, set_val);
285 fprintf(stderr, "rte_gpu_comm_set_flag returned error %d\n", ret);
289 ret = rte_gpu_comm_get_flag_value(&devflag, &get_val);
291 fprintf(stderr, "rte_gpu_comm_get_flag_value returned error %d\n", ret);
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);
298 ret = rte_gpu_comm_destroy_flag(&devflag);
300 fprintf(stderr, "rte_gpu_comm_destroy_flags returned error %d\n", ret);
304 printf("\n=======> TEST: PASSED\n");
309 rte_gpu_comm_destroy_flag(&devflag);
310 printf("\n=======> TEST: FAILED\n");
315 simulate_gpu_task(struct rte_gpu_comm_list *comm_list_item, int num_pkts)
319 if (comm_list_item == NULL)
322 for (idx = 0; idx < num_pkts; idx++) {
324 * consume(comm_list_item->pkt_list[idx].addr);
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.
333 rte_gpu_comm_set_status(comm_list_item, RTE_GPU_COMM_LIST_DONE);
339 create_update_comm_list(uint16_t gpu_id)
343 struct rte_gpu_comm_list *comm_list = NULL;
344 uint32_t num_comm_items = 1024;
345 struct rte_mbuf *mbufs[10];
347 printf("\n=======> TEST: Communication list\n\n");
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);
356 * Simulate DPDK receive functions like rte_eth_rx_burst()
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");
365 memset(mbufs[i], 0, sizeof(struct rte_mbuf));
369 * Populate just the first item of the list
371 ret = rte_gpu_comm_populate_list_pkts(&(comm_list[0]), mbufs, 10);
373 fprintf(stderr, "rte_gpu_comm_populate_list_pkts returned error %d\n", ret);
377 ret = rte_gpu_comm_cleanup_list(&(comm_list[0]));
379 fprintf(stderr, "rte_gpu_comm_cleanup_list erroneously cleaned the list even if packets have not been consumed yet\n");
382 printf("Communication list not cleaned because packets have not been consumed yet.\n");
385 * Simulate a GPU tasks going through the packet list to consume
386 * mbufs packets and release them
388 printf("Consuming packets...\n");
389 simulate_gpu_task(&(comm_list[0]), 10);
392 * Packets have been consumed, now the communication item
393 * and the related mbufs can be all released
395 ret = rte_gpu_comm_cleanup_list(&(comm_list[0]));
397 fprintf(stderr, "rte_gpu_comm_cleanup_list returned error %d\n", ret);
401 printf("Communication list cleaned because packets have been consumed now.\n");
403 ret = rte_gpu_comm_destroy_list(comm_list, num_comm_items);
405 fprintf(stderr, "rte_gpu_comm_destroy_list returned error %d\n", ret);
409 for (i = 0; i < 10; i++)
412 printf("\n=======> TEST: PASSED\n");
417 rte_gpu_comm_destroy_list(comm_list, num_comm_items);
418 for (i = 0; i < 10; i++)
420 printf("\n=======> TEST: FAILED\n");
425 main(int argc, char **argv)
430 struct rte_gpu_info ginfo;
433 ret = rte_eal_init(argc, argv);
435 rte_exit(EXIT_FAILURE, "EAL init failed\n");
439 args_parse(argc, argv);
443 nb_gpus = rte_gpu_count_avail();
444 printf("\n\nDPDK found %d GPUs:\n", nb_gpus);
445 RTE_GPU_FOREACH(gpu_id)
447 if (rte_gpu_info_get(gpu_id, &ginfo))
448 rte_exit(EXIT_FAILURE, "rte_gpu_info_get error - bye\n");
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",
455 (((float)ginfo.total_memory)/(float)1024)/(float)1024,
456 ginfo.processor_count
462 fprintf(stderr, "Need at least one GPU on the system to run the example\n");
471 alloc_gpu_memory(gpu_id);
472 register_cpu_memory(gpu_id);
473 gpu_mem_cpu_map(gpu_id);
476 * Communication items test
478 create_update_comm_flag(gpu_id);
479 create_update_comm_list(gpu_id);
481 /* clean up the EAL */