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 31 usage(const char *prog_name) 32 { 33 printf("%s [EAL options] --\n", 34 prog_name); 35 } 36 37 static void 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 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 int ret; 72 73 printf("\n=======> TEST: Allocate GPU memory\n\n"); 74 75 /* Alloc memory on GPU 0 */ 76 ptr_1 = rte_gpu_mem_alloc(gpu_id, buf_bytes); 77 if (ptr_1 == NULL) { 78 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n"); 79 goto error; 80 } 81 printf("GPU memory allocated at 0x%p size is %zd bytes\n", 82 ptr_1, buf_bytes); 83 84 ptr_2 = rte_gpu_mem_alloc(gpu_id, buf_bytes); 85 if (ptr_2 == NULL) { 86 fprintf(stderr, "rte_gpu_mem_alloc GPU memory returned error\n"); 87 goto error; 88 } 89 printf("GPU memory allocated at 0x%p size is %zd bytes\n", 90 ptr_2, buf_bytes); 91 92 ret = rte_gpu_mem_free(gpu_id, (uint8_t *)(ptr_1)+0x700); 93 if (ret < 0) { 94 printf("GPU memory 0x%p NOT freed: GPU driver didn't find this memory address internally.\n", 95 (uint8_t *)(ptr_1)+0x700); 96 } else { 97 fprintf(stderr, "ERROR: rte_gpu_mem_free freed GPU memory 0x%p\n", 98 (uint8_t *)(ptr_1)+0x700); 99 goto error; 100 } 101 102 ret = rte_gpu_mem_free(gpu_id, ptr_2); 103 if (ret < 0) { 104 fprintf(stderr, "rte_gpu_mem_free returned error %d\n", ret); 105 goto error; 106 } 107 printf("GPU memory 0x%p freed\n", ptr_2); 108 109 ret = rte_gpu_mem_free(gpu_id, ptr_1); 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_1); 115 116 printf("\n=======> TEST: PASSED\n"); 117 return 0; 118 119 error: 120 printf("\n=======> TEST: FAILED\n"); 121 return -1; 122 } 123 124 static int 125 register_cpu_memory(uint16_t gpu_id) 126 { 127 void *ptr = NULL; 128 size_t buf_bytes = 1024; 129 int ret; 130 131 printf("\n=======> TEST: Register CPU memory\n\n"); 132 133 /* Alloc memory on CPU visible from GPU 0 */ 134 ptr = rte_zmalloc(NULL, buf_bytes, 0); 135 if (ptr == NULL) { 136 fprintf(stderr, "Failed to allocate CPU memory.\n"); 137 goto error; 138 } 139 140 ret = rte_gpu_mem_register(gpu_id, buf_bytes, ptr); 141 if (ret < 0) { 142 fprintf(stderr, "rte_gpu_mem_register CPU memory returned error %d\n", ret); 143 goto error; 144 } 145 printf("CPU memory registered at 0x%p %zdB\n", ptr, buf_bytes); 146 147 ret = rte_gpu_mem_unregister(gpu_id, (uint8_t *)(ptr)+0x700); 148 if (ret < 0) { 149 printf("CPU memory 0x%p NOT unregistered: GPU driver didn't find this memory address internally\n", 150 (uint8_t *)(ptr)+0x700); 151 } else { 152 fprintf(stderr, "ERROR: rte_gpu_mem_unregister unregistered GPU memory 0x%p\n", 153 (uint8_t *)(ptr)+0x700); 154 goto error; 155 } 156 157 ret = rte_gpu_mem_unregister(gpu_id, ptr); 158 if (ret < 0) { 159 fprintf(stderr, "rte_gpu_mem_unregister returned error %d\n", ret); 160 goto error; 161 } 162 printf("CPU memory 0x%p unregistered\n", ptr); 163 164 printf("\n=======> TEST: PASSED\n"); 165 return 0; 166 167 error: 168 printf("\n=======> TEST: FAILED\n"); 169 return -1; 170 } 171 172 static int 173 create_update_comm_flag(uint16_t gpu_id) 174 { 175 struct rte_gpu_comm_flag devflag; 176 int ret = 0; 177 uint32_t set_val; 178 uint32_t get_val; 179 180 printf("\n=======> TEST: Communication flag\n\n"); 181 182 ret = rte_gpu_comm_create_flag(gpu_id, &devflag, RTE_GPU_COMM_FLAG_CPU); 183 if (ret < 0) { 184 fprintf(stderr, "rte_gpu_comm_create_flag returned error %d\n", ret); 185 goto error; 186 } 187 188 set_val = 25; 189 ret = rte_gpu_comm_set_flag(&devflag, set_val); 190 if (ret < 0) { 191 fprintf(stderr, "rte_gpu_comm_set_flag returned error %d\n", ret); 192 goto error; 193 } 194 195 ret = rte_gpu_comm_get_flag_value(&devflag, &get_val); 196 if (ret < 0) { 197 fprintf(stderr, "rte_gpu_comm_get_flag_value returned error %d\n", ret); 198 goto error; 199 } 200 201 printf("Communication flag value at 0x%p was set to %d and current value is %d\n", 202 devflag.ptr, set_val, get_val); 203 204 set_val = 38; 205 ret = rte_gpu_comm_set_flag(&devflag, set_val); 206 if (ret < 0) { 207 fprintf(stderr, "rte_gpu_comm_set_flag returned error %d\n", ret); 208 goto error; 209 } 210 211 ret = rte_gpu_comm_get_flag_value(&devflag, &get_val); 212 if (ret < 0) { 213 fprintf(stderr, "rte_gpu_comm_get_flag_value returned error %d\n", ret); 214 goto error; 215 } 216 217 printf("Communication flag value at 0x%p was set to %d and current value is %d\n", 218 devflag.ptr, set_val, get_val); 219 220 ret = rte_gpu_comm_destroy_flag(&devflag); 221 if (ret < 0) { 222 fprintf(stderr, "rte_gpu_comm_destroy_flags returned error %d\n", ret); 223 goto error; 224 } 225 226 printf("\n=======> TEST: PASSED\n"); 227 return 0; 228 229 error: 230 printf("\n=======> TEST: FAILED\n"); 231 return -1; 232 } 233 234 static int 235 simulate_gpu_task(struct rte_gpu_comm_list *comm_list_item, int num_pkts) 236 { 237 int idx; 238 239 if (comm_list_item == NULL) 240 return -1; 241 242 for (idx = 0; idx < num_pkts; idx++) { 243 /** 244 * consume(comm_list_item->pkt_list[idx].addr); 245 */ 246 } 247 comm_list_item->status = RTE_GPU_COMM_LIST_DONE; 248 249 return 0; 250 } 251 252 static int 253 create_update_comm_list(uint16_t gpu_id) 254 { 255 int ret = 0; 256 int i = 0; 257 struct rte_gpu_comm_list *comm_list; 258 uint32_t num_comm_items = 1024; 259 struct rte_mbuf *mbufs[10]; 260 261 printf("\n=======> TEST: Communication list\n\n"); 262 263 comm_list = rte_gpu_comm_create_list(gpu_id, num_comm_items); 264 if (comm_list == NULL) { 265 fprintf(stderr, "rte_gpu_comm_create_list returned error %d\n", ret); 266 goto error; 267 } 268 269 /** 270 * Simulate DPDK receive functions like rte_eth_rx_burst() 271 */ 272 for (i = 0; i < 10; i++) { 273 mbufs[i] = rte_zmalloc(NULL, sizeof(struct rte_mbuf), 0); 274 if (mbufs[i] == NULL) { 275 fprintf(stderr, "Failed to allocate fake mbufs in CPU memory.\n"); 276 goto error; 277 } 278 279 memset(mbufs[i], 0, sizeof(struct rte_mbuf)); 280 } 281 282 /** 283 * Populate just the first item of the list 284 */ 285 ret = rte_gpu_comm_populate_list_pkts(&(comm_list[0]), mbufs, 10); 286 if (ret < 0) { 287 fprintf(stderr, "rte_gpu_comm_populate_list_pkts returned error %d\n", ret); 288 goto error; 289 } 290 291 ret = rte_gpu_comm_cleanup_list(&(comm_list[0])); 292 if (ret == 0) { 293 fprintf(stderr, "rte_gpu_comm_cleanup_list erroneously cleaned the list even if packets have not been consumed yet\n"); 294 goto error; 295 } 296 printf("Communication list not cleaned because packets have not been consumed yet.\n"); 297 298 /** 299 * Simulate a GPU tasks going through the packet list to consume 300 * mbufs packets and release them 301 */ 302 printf("Consuming packets...\n"); 303 simulate_gpu_task(&(comm_list[0]), 10); 304 305 /** 306 * Packets have been consumed, now the communication item 307 * and the related mbufs can be all released 308 */ 309 ret = rte_gpu_comm_cleanup_list(&(comm_list[0])); 310 if (ret < 0) { 311 fprintf(stderr, "rte_gpu_comm_cleanup_list returned error %d\n", ret); 312 goto error; 313 } 314 315 printf("Communication list cleaned because packets have been consumed now.\n"); 316 317 ret = rte_gpu_comm_destroy_list(comm_list, num_comm_items); 318 if (ret < 0) { 319 fprintf(stderr, "rte_gpu_comm_destroy_list returned error %d\n", ret); 320 goto error; 321 } 322 323 for (i = 0; i < 10; i++) 324 rte_free(mbufs[i]); 325 326 printf("\n=======> TEST: PASSED\n"); 327 return 0; 328 329 error: 330 printf("\n=======> TEST: FAILED\n"); 331 return -1; 332 } 333 334 int 335 main(int argc, char **argv) 336 { 337 int ret; 338 int nb_gpus = 0; 339 int16_t gpu_id = 0; 340 struct rte_gpu_info ginfo; 341 342 /* Init EAL. */ 343 ret = rte_eal_init(argc, argv); 344 if (ret < 0) 345 rte_exit(EXIT_FAILURE, "EAL init failed\n"); 346 argc -= ret; 347 argv += ret; 348 if (argc > 1) 349 args_parse(argc, argv); 350 argc -= ret; 351 argv += ret; 352 353 nb_gpus = rte_gpu_count_avail(); 354 printf("\n\nDPDK found %d GPUs:\n", nb_gpus); 355 RTE_GPU_FOREACH(gpu_id) 356 { 357 if (rte_gpu_info_get(gpu_id, &ginfo)) 358 rte_exit(EXIT_FAILURE, "rte_gpu_info_get error - bye\n"); 359 360 printf("\tGPU ID %d\n\t\tparent ID %d GPU Bus ID %s NUMA node %d Tot memory %.02f MB, Tot processors %d\n", 361 ginfo.dev_id, 362 ginfo.parent, 363 ginfo.name, 364 ginfo.numa_node, 365 (((float)ginfo.total_memory)/(float)1024)/(float)1024, 366 ginfo.processor_count 367 ); 368 } 369 printf("\n\n"); 370 371 if (nb_gpus == 0) { 372 fprintf(stderr, "Need at least one GPU on the system to run the example\n"); 373 return EXIT_FAILURE; 374 } 375 376 gpu_id = 0; 377 378 /** 379 * Memory tests 380 */ 381 alloc_gpu_memory(gpu_id); 382 register_cpu_memory(gpu_id); 383 384 /** 385 * Communication items test 386 */ 387 create_update_comm_flag(gpu_id); 388 create_update_comm_list(gpu_id); 389 390 /* clean up the EAL */ 391 rte_eal_cleanup(); 392 393 return EXIT_SUCCESS; 394 } 395