xref: /dpdk/drivers/gpu/cuda/cuda.c (revision 759b5e653580498933e425596b0263084d59bb24)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright (c) 2021 NVIDIA Corporation & Affiliates
3  */
4 
5 #include <dlfcn.h>
6 
7 #include <rte_common.h>
8 #include <rte_log.h>
9 #include <rte_malloc.h>
10 #include <rte_errno.h>
11 #include <rte_pci.h>
12 #include <rte_bus_pci.h>
13 #include <rte_byteorder.h>
14 #include <rte_dev.h>
15 
16 #include <gpudev_driver.h>
17 #include <cuda.h>
18 #include <cudaTypedefs.h>
19 
20 #define CUDA_DRIVER_MIN_VERSION 11040
21 #define CUDA_API_MIN_VERSION 3020
22 
23 /* CUDA Driver functions loaded with dlsym() */
24 static CUresult CUDAAPI (*sym_cuInit)(unsigned int flags);
25 static CUresult CUDAAPI (*sym_cuDriverGetVersion)(int *driverVersion);
26 static CUresult CUDAAPI (*sym_cuGetProcAddress)(const char *symbol,
27 		void **pfn, int cudaVersion, uint64_t flags);
28 
29 /* CUDA Driver functions loaded with cuGetProcAddress for versioning */
30 static PFN_cuGetErrorString pfn_cuGetErrorString;
31 static PFN_cuGetErrorName pfn_cuGetErrorName;
32 static PFN_cuPointerSetAttribute pfn_cuPointerSetAttribute;
33 static PFN_cuDeviceGetAttribute pfn_cuDeviceGetAttribute;
34 static PFN_cuDeviceGetByPCIBusId pfn_cuDeviceGetByPCIBusId;
35 static PFN_cuDevicePrimaryCtxRetain pfn_cuDevicePrimaryCtxRetain;
36 static PFN_cuDevicePrimaryCtxRelease pfn_cuDevicePrimaryCtxRelease;
37 static PFN_cuDeviceTotalMem pfn_cuDeviceTotalMem;
38 static PFN_cuDeviceGetName pfn_cuDeviceGetName;
39 static PFN_cuCtxGetApiVersion pfn_cuCtxGetApiVersion;
40 static PFN_cuCtxSetCurrent pfn_cuCtxSetCurrent;
41 static PFN_cuCtxGetCurrent pfn_cuCtxGetCurrent;
42 static PFN_cuCtxGetDevice pfn_cuCtxGetDevice;
43 static PFN_cuCtxGetExecAffinity pfn_cuCtxGetExecAffinity;
44 static PFN_cuMemAlloc pfn_cuMemAlloc;
45 static PFN_cuMemFree pfn_cuMemFree;
46 static PFN_cuMemHostRegister pfn_cuMemHostRegister;
47 static PFN_cuMemHostUnregister pfn_cuMemHostUnregister;
48 static PFN_cuMemHostGetDevicePointer pfn_cuMemHostGetDevicePointer;
49 static PFN_cuFlushGPUDirectRDMAWrites pfn_cuFlushGPUDirectRDMAWrites;
50 
51 static void *cudalib;
52 static unsigned int cuda_api_version;
53 static int cuda_driver_version;
54 
55 /* NVIDIA GPU vendor */
56 #define NVIDIA_GPU_VENDOR_ID (0x10de)
57 
58 /* NVIDIA GPU device IDs */
59 #define NVIDIA_GPU_A100_40GB_DEVICE_ID (0x20f1)
60 #define NVIDIA_GPU_A100_80GB_DEVICE_ID (0x20b5)
61 #define NVIDIA_GPU_A100_80GB_DPU_DEVICE_ID (0x20b8)
62 
63 #define NVIDIA_GPU_A30_24GB_DEVICE_ID (0x20b7)
64 #define NVIDIA_GPU_A10_24GB_DEVICE_ID (0x2236)
65 
66 #define NVIDIA_GPU_V100_32GB_DEVICE_ID (0x1db6)
67 #define NVIDIA_GPU_V100_16GB_DEVICE_ID (0x1db4)
68 
69 #define NVIDIA_GPU_T4_16GB_DEVICE_ID (0x1eb8)
70 
71 #define CUDA_MAX_ALLOCATION_NUM 512
72 
73 #define GPU_PAGE_SHIFT 16
74 #define GPU_PAGE_SIZE (1UL << GPU_PAGE_SHIFT)
75 
76 static RTE_LOG_REGISTER_DEFAULT(cuda_logtype, NOTICE);
77 
78 /* Helper macro for logging */
79 #define rte_cuda_log(level, fmt, ...) \
80 	rte_log(RTE_LOG_ ## level, cuda_logtype, fmt "\n", ##__VA_ARGS__)
81 
82 #define rte_cuda_debug(fmt, ...) \
83 	rte_cuda_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
84 		##__VA_ARGS__)
85 
86 /* NVIDIA GPU address map */
87 static const struct rte_pci_id pci_id_cuda_map[] = {
88 	{
89 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
90 				NVIDIA_GPU_A100_40GB_DEVICE_ID)
91 	},
92 	{
93 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
94 				NVIDIA_GPU_A100_80GB_DEVICE_ID)
95 	},
96 	{
97 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
98 				NVIDIA_GPU_A100_80GB_DPU_DEVICE_ID)
99 	},
100 	{
101 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
102 				NVIDIA_GPU_A30_24GB_DEVICE_ID)
103 	},
104 	{
105 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
106 				NVIDIA_GPU_A10_24GB_DEVICE_ID)
107 	},
108 	{
109 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
110 				NVIDIA_GPU_V100_32GB_DEVICE_ID)
111 	},
112 	{
113 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
114 				NVIDIA_GPU_V100_16GB_DEVICE_ID)
115 	},
116 	{
117 		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
118 				NVIDIA_GPU_T4_16GB_DEVICE_ID)
119 	},
120 	{
121 		.device_id = 0
122 	}
123 };
124 
125 /* Device private info */
126 struct cuda_info {
127 	char gpu_name[RTE_DEV_NAME_MAX_LEN];
128 	CUdevice cu_dev;
129 	int gdr_supported;
130 	int gdr_write_ordering;
131 	int gdr_flush_type;
132 };
133 
134 /* Type of memory allocated by CUDA driver */
135 enum mem_type {
136 	GPU_MEM = 0,
137 	CPU_REGISTERED,
138 	GPU_REGISTERED /* Not used yet */
139 };
140 
141 /* key associated to a memory address */
142 typedef uintptr_t cuda_ptr_key;
143 
144 /* Single entry of the memory list */
145 struct mem_entry {
146 	CUdeviceptr ptr_d;
147 	CUdeviceptr ptr_orig_d;
148 	void *ptr_h;
149 	size_t size;
150 	size_t size_orig;
151 	struct rte_gpu *dev;
152 	CUcontext ctx;
153 	cuda_ptr_key pkey;
154 	enum mem_type mtype;
155 	struct mem_entry *prev;
156 	struct mem_entry *next;
157 };
158 
159 static struct mem_entry *mem_alloc_list_head;
160 static struct mem_entry *mem_alloc_list_tail;
161 static uint32_t mem_alloc_list_last_elem;
162 
163 /* Load the CUDA symbols */
164 
165 static int
166 cuda_loader(void)
167 {
168 	char cuda_path[1024];
169 
170 	if (getenv("CUDA_PATH_L") == NULL)
171 		snprintf(cuda_path, 1024, "%s", "libcuda.so");
172 	else
173 		snprintf(cuda_path, 1024, "%s%s", getenv("CUDA_PATH_L"), "libcuda.so");
174 
175 	cudalib = dlopen(cuda_path, RTLD_LAZY);
176 	if (cudalib == NULL) {
177 		rte_cuda_log(ERR, "Failed to find CUDA library in %s (CUDA_PATH_L=%s)",
178 				cuda_path, getenv("CUDA_PATH_L"));
179 		return -1;
180 	}
181 
182 	return 0;
183 }
184 
185 static int
186 cuda_sym_func_loader(void)
187 {
188 	if (cudalib == NULL)
189 		return -1;
190 
191 	sym_cuInit = dlsym(cudalib, "cuInit");
192 	if (sym_cuInit == NULL) {
193 		rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuInit");
194 		return -1;
195 	}
196 
197 	sym_cuDriverGetVersion = dlsym(cudalib, "cuDriverGetVersion");
198 	if (sym_cuDriverGetVersion == NULL) {
199 		rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuDriverGetVersion");
200 		return -1;
201 	}
202 
203 	sym_cuGetProcAddress = dlsym(cudalib, "cuGetProcAddress");
204 	if (sym_cuGetProcAddress == NULL) {
205 		rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuGetProcAddress");
206 		return -1;
207 	}
208 
209 	return 0;
210 }
211 
212 static int
213 cuda_pfn_func_loader(void)
214 {
215 	CUresult res;
216 
217 	res = sym_cuGetProcAddress("cuGetErrorString",
218 			(void **) (&pfn_cuGetErrorString), cuda_driver_version, 0);
219 	if (res != 0) {
220 		rte_cuda_log(ERR, "Retrieve pfn_cuGetErrorString failed with %d", res);
221 		return -1;
222 	}
223 
224 	res = sym_cuGetProcAddress("cuGetErrorName",
225 			(void **)(&pfn_cuGetErrorName), cuda_driver_version, 0);
226 	if (res != 0) {
227 		rte_cuda_log(ERR, "Retrieve pfn_cuGetErrorName failed with %d", res);
228 		return -1;
229 	}
230 
231 	res = sym_cuGetProcAddress("cuPointerSetAttribute",
232 			(void **)(&pfn_cuPointerSetAttribute), cuda_driver_version, 0);
233 	if (res != 0) {
234 		rte_cuda_log(ERR, "Retrieve pfn_cuPointerSetAttribute failed with %d", res);
235 		return -1;
236 	}
237 
238 	res = sym_cuGetProcAddress("cuDeviceGetAttribute",
239 			(void **)(&pfn_cuDeviceGetAttribute), cuda_driver_version, 0);
240 	if (res != 0) {
241 		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetAttribute failed with %d", res);
242 		return -1;
243 	}
244 
245 	res = sym_cuGetProcAddress("cuDeviceGetByPCIBusId",
246 			(void **)(&pfn_cuDeviceGetByPCIBusId), cuda_driver_version, 0);
247 	if (res != 0) {
248 		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetByPCIBusId failed with %d", res);
249 		return -1;
250 	}
251 
252 	res = sym_cuGetProcAddress("cuDeviceGetName",
253 			(void **)(&pfn_cuDeviceGetName), cuda_driver_version, 0);
254 	if (res != 0) {
255 		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetName failed with %d", res);
256 		return -1;
257 	}
258 
259 	res = sym_cuGetProcAddress("cuDevicePrimaryCtxRetain",
260 			(void **)(&pfn_cuDevicePrimaryCtxRetain), cuda_driver_version, 0);
261 	if (res != 0) {
262 		rte_cuda_log(ERR, "Retrieve pfn_cuDevicePrimaryCtxRetain failed with %d", res);
263 		return -1;
264 	}
265 
266 	res = sym_cuGetProcAddress("cuDevicePrimaryCtxRelease",
267 			(void **)(&pfn_cuDevicePrimaryCtxRelease), cuda_driver_version, 0);
268 	if (res != 0) {
269 		rte_cuda_log(ERR, "Retrieve pfn_cuDevicePrimaryCtxRelease failed with %d", res);
270 		return -1;
271 	}
272 
273 	res = sym_cuGetProcAddress("cuDeviceTotalMem",
274 			(void **)(&pfn_cuDeviceTotalMem), cuda_driver_version, 0);
275 	if (res != 0) {
276 		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceTotalMem failed with %d", res);
277 		return -1;
278 	}
279 
280 	res = sym_cuGetProcAddress("cuCtxGetApiVersion",
281 			(void **)(&pfn_cuCtxGetApiVersion), cuda_driver_version, 0);
282 	if (res != 0) {
283 		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetApiVersion failed with %d", res);
284 		return -1;
285 	}
286 
287 	res = sym_cuGetProcAddress("cuCtxGetDevice",
288 			(void **)(&pfn_cuCtxGetDevice), cuda_driver_version, 0);
289 	if (res != 0) {
290 		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetDevice failed with %d", res);
291 		return -1;
292 	}
293 
294 	res = sym_cuGetProcAddress("cuCtxSetCurrent",
295 			(void **)(&pfn_cuCtxSetCurrent), cuda_driver_version, 0);
296 	if (res != 0) {
297 		rte_cuda_log(ERR, "Retrieve pfn_cuCtxSetCurrent failed with %d", res);
298 		return -1;
299 	}
300 
301 	res = sym_cuGetProcAddress("cuCtxGetCurrent",
302 			(void **)(&pfn_cuCtxGetCurrent), cuda_driver_version, 0);
303 	if (res != 0) {
304 		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetCurrent failed with %d", res);
305 		return -1;
306 	}
307 
308 	res = sym_cuGetProcAddress("cuCtxGetExecAffinity",
309 			(void **)(&pfn_cuCtxGetExecAffinity), cuda_driver_version, 0);
310 	if (res != 0) {
311 		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetExecAffinity failed with %d", res);
312 		return -1;
313 	}
314 
315 	res = sym_cuGetProcAddress("cuMemAlloc",
316 			(void **)(&pfn_cuMemAlloc), cuda_driver_version, 0);
317 	if (res != 0) {
318 		rte_cuda_log(ERR, "Retrieve pfn_cuMemAlloc failed with %d", res);
319 		return -1;
320 	}
321 
322 	res = sym_cuGetProcAddress("cuMemFree",
323 			(void **)(&pfn_cuMemFree), cuda_driver_version, 0);
324 	if (res != 0) {
325 		rte_cuda_log(ERR, "Retrieve pfn_cuMemFree failed with %d", res);
326 		return -1;
327 	}
328 
329 	res = sym_cuGetProcAddress("cuMemHostRegister",
330 			(void **)(&pfn_cuMemHostRegister), cuda_driver_version, 0);
331 	if (res != 0) {
332 		rte_cuda_log(ERR, "Retrieve pfn_cuMemHostRegister failed with %d", res);
333 		return -1;
334 	}
335 
336 	res = sym_cuGetProcAddress("cuMemHostUnregister",
337 			(void **)(&pfn_cuMemHostUnregister), cuda_driver_version, 0);
338 	if (res != 0) {
339 		rte_cuda_log(ERR, "Retrieve pfn_cuMemHostUnregister failed with %d", res);
340 		return -1;
341 	}
342 
343 	res = sym_cuGetProcAddress("cuMemHostGetDevicePointer",
344 			(void **)(&pfn_cuMemHostGetDevicePointer), cuda_driver_version, 0);
345 	if (res != 0) {
346 		rte_cuda_log(ERR, "Retrieve pfn_cuMemHostGetDevicePointer failed with %d", res);
347 		return -1;
348 	}
349 
350 	res = sym_cuGetProcAddress("cuFlushGPUDirectRDMAWrites",
351 			(void **)(&pfn_cuFlushGPUDirectRDMAWrites), cuda_driver_version, 0);
352 	if (res != 0) {
353 		rte_cuda_log(ERR, "Retrieve cuFlushGPUDirectRDMAWrites failed with %d", res);
354 		return -1;
355 	}
356 
357 	return 0;
358 }
359 
360 /* Generate a key from a memory pointer */
361 static cuda_ptr_key
362 get_hash_from_ptr(void *ptr)
363 {
364 	return (uintptr_t)ptr;
365 }
366 
367 static uint32_t
368 mem_list_count_item(void)
369 {
370 	return mem_alloc_list_last_elem;
371 }
372 
373 /* Initiate list of memory allocations if not done yet */
374 static struct mem_entry *
375 mem_list_add_item(void)
376 {
377 	/* Initiate list of memory allocations if not done yet */
378 	if (mem_alloc_list_head == NULL) {
379 		mem_alloc_list_head = rte_zmalloc(NULL,
380 				sizeof(struct mem_entry),
381 				RTE_CACHE_LINE_SIZE);
382 		if (mem_alloc_list_head == NULL) {
383 			rte_cuda_log(ERR, "Failed to allocate memory for memory list");
384 			return NULL;
385 		}
386 
387 		mem_alloc_list_head->next = NULL;
388 		mem_alloc_list_head->prev = NULL;
389 		mem_alloc_list_tail = mem_alloc_list_head;
390 	} else {
391 		struct mem_entry *mem_alloc_list_cur = rte_zmalloc(NULL,
392 				sizeof(struct mem_entry),
393 				RTE_CACHE_LINE_SIZE);
394 
395 		if (mem_alloc_list_cur == NULL) {
396 			rte_cuda_log(ERR, "Failed to allocate memory for memory list");
397 			return NULL;
398 		}
399 
400 		mem_alloc_list_tail->next = mem_alloc_list_cur;
401 		mem_alloc_list_cur->prev = mem_alloc_list_tail;
402 		mem_alloc_list_tail = mem_alloc_list_tail->next;
403 		mem_alloc_list_tail->next = NULL;
404 	}
405 
406 	mem_alloc_list_last_elem++;
407 
408 	return mem_alloc_list_tail;
409 }
410 
411 static struct mem_entry *
412 mem_list_find_item(cuda_ptr_key pk)
413 {
414 	struct mem_entry *mem_alloc_list_cur = NULL;
415 
416 	if (mem_alloc_list_head == NULL) {
417 		rte_cuda_log(ERR, "Memory list doesn't exist");
418 		return NULL;
419 	}
420 
421 	if (mem_list_count_item() == 0) {
422 		rte_cuda_log(ERR, "No items in memory list");
423 		return NULL;
424 	}
425 
426 	mem_alloc_list_cur = mem_alloc_list_head;
427 
428 	while (mem_alloc_list_cur != NULL) {
429 		if (mem_alloc_list_cur->pkey == pk)
430 			return mem_alloc_list_cur;
431 		mem_alloc_list_cur = mem_alloc_list_cur->next;
432 	}
433 
434 	return mem_alloc_list_cur;
435 }
436 
437 static int
438 mem_list_del_item(cuda_ptr_key pk)
439 {
440 	struct mem_entry *mem_alloc_list_cur = NULL;
441 
442 	mem_alloc_list_cur = mem_list_find_item(pk);
443 	if (mem_alloc_list_cur == NULL)
444 		return -EINVAL;
445 
446 	/* if key is in head */
447 	if (mem_alloc_list_cur->prev == NULL) {
448 		mem_alloc_list_head = mem_alloc_list_cur->next;
449 		if (mem_alloc_list_head != NULL)
450 			mem_alloc_list_head->prev = NULL;
451 	} else {
452 		mem_alloc_list_cur->prev->next = mem_alloc_list_cur->next;
453 		if (mem_alloc_list_cur->next != NULL)
454 			mem_alloc_list_cur->next->prev = mem_alloc_list_cur->prev;
455 	}
456 
457 	rte_free(mem_alloc_list_cur);
458 
459 	mem_alloc_list_last_elem--;
460 
461 	return 0;
462 }
463 
464 static int
465 cuda_dev_info_get(struct rte_gpu *dev, struct rte_gpu_info *info)
466 {
467 	int ret = 0;
468 	CUresult res;
469 	struct rte_gpu_info parent_info;
470 	CUexecAffinityParam affinityPrm;
471 	const char *err_string;
472 	struct cuda_info *private;
473 	CUcontext current_ctx;
474 	CUcontext input_ctx;
475 
476 	if (dev == NULL) {
477 		rte_errno = ENODEV;
478 		return -rte_errno;
479 	}
480 
481 	/* Child initialization time probably called by rte_gpu_add_child() */
482 	if (dev->mpshared->info.parent != RTE_GPU_ID_NONE &&
483 			dev->mpshared->dev_private == NULL) {
484 		/* Store current ctx */
485 		res = pfn_cuCtxGetCurrent(&current_ctx);
486 		if (res != 0) {
487 			pfn_cuGetErrorString(res, &(err_string));
488 			rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
489 					err_string);
490 			rte_errno = EPERM;
491 			return -rte_errno;
492 		}
493 
494 		/* Set child ctx as current ctx */
495 		input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
496 		res = pfn_cuCtxSetCurrent(input_ctx);
497 		if (res != 0) {
498 			pfn_cuGetErrorString(res, &(err_string));
499 			rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
500 					err_string);
501 			rte_errno = EPERM;
502 			return -rte_errno;
503 		}
504 
505 		/*
506 		 * Ctx capacity info
507 		 */
508 
509 		/* MPS compatible */
510 		res = pfn_cuCtxGetExecAffinity(&affinityPrm,
511 				CU_EXEC_AFFINITY_TYPE_SM_COUNT);
512 		if (res != 0) {
513 			pfn_cuGetErrorString(res, &(err_string));
514 			rte_cuda_log(ERR, "cuCtxGetExecAffinity failed with %s",
515 					err_string);
516 		}
517 		dev->mpshared->info.processor_count =
518 				(uint32_t)affinityPrm.param.smCount.val;
519 
520 		ret = rte_gpu_info_get(dev->mpshared->info.parent, &parent_info);
521 		if (ret) {
522 			rte_errno = ENODEV;
523 			return -rte_errno;
524 		}
525 		dev->mpshared->info.total_memory = parent_info.total_memory;
526 
527 		/*
528 		 * GPU Device private info
529 		 */
530 		dev->mpshared->dev_private = rte_zmalloc(NULL,
531 				sizeof(struct cuda_info),
532 				RTE_CACHE_LINE_SIZE);
533 		if (dev->mpshared->dev_private == NULL) {
534 			rte_cuda_log(ERR, "Failed to allocate memory for GPU process private");
535 			rte_errno = EPERM;
536 			return -rte_errno;
537 		}
538 
539 		private = (struct cuda_info *)dev->mpshared->dev_private;
540 
541 		res = pfn_cuCtxGetDevice(&(private->cu_dev));
542 		if (res != 0) {
543 			pfn_cuGetErrorString(res, &(err_string));
544 			rte_cuda_log(ERR, "cuCtxGetDevice failed with %s",
545 					err_string);
546 			rte_errno = EPERM;
547 			return -rte_errno;
548 		}
549 
550 		res = pfn_cuDeviceGetName(private->gpu_name,
551 				RTE_DEV_NAME_MAX_LEN, private->cu_dev);
552 		if (res != 0) {
553 			pfn_cuGetErrorString(res, &(err_string));
554 			rte_cuda_log(ERR, "cuDeviceGetName failed with %s",
555 					err_string);
556 			rte_errno = EPERM;
557 			return -rte_errno;
558 		}
559 
560 		/* Restore original ctx as current ctx */
561 		res = pfn_cuCtxSetCurrent(current_ctx);
562 		if (res != 0) {
563 			pfn_cuGetErrorString(res, &(err_string));
564 			rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
565 					err_string);
566 			rte_errno = EPERM;
567 			return -rte_errno;
568 		}
569 	}
570 
571 	*info = dev->mpshared->info;
572 
573 	return 0;
574 }
575 
576 /*
577  * GPU Memory
578  */
579 
580 static int
581 cuda_mem_alloc(struct rte_gpu *dev, size_t size, unsigned int align, void **ptr)
582 {
583 	CUresult res;
584 	const char *err_string;
585 	CUcontext current_ctx;
586 	CUcontext input_ctx;
587 	unsigned int flag = 1;
588 
589 	if (dev == NULL)
590 		return -ENODEV;
591 
592 	/* Store current ctx */
593 	res = pfn_cuCtxGetCurrent(&current_ctx);
594 	if (res != 0) {
595 		pfn_cuGetErrorString(res, &(err_string));
596 		rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
597 				err_string);
598 		rte_errno = EPERM;
599 		return -rte_errno;
600 	}
601 
602 	/* Set child ctx as current ctx */
603 	input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
604 	res = pfn_cuCtxSetCurrent(input_ctx);
605 	if (res != 0) {
606 		pfn_cuGetErrorString(res, &(err_string));
607 		rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
608 				err_string);
609 		rte_errno = EPERM;
610 		return -rte_errno;
611 	}
612 
613 	/* Get next memory list item */
614 	mem_alloc_list_tail = mem_list_add_item();
615 	if (mem_alloc_list_tail == NULL) {
616 		rte_errno = EPERM;
617 		return -rte_errno;
618 	}
619 
620 	/* Allocate memory */
621 	mem_alloc_list_tail->size = size;
622 	mem_alloc_list_tail->size_orig = size + align;
623 
624 	res = pfn_cuMemAlloc(&(mem_alloc_list_tail->ptr_orig_d),
625 			mem_alloc_list_tail->size_orig);
626 	if (res != 0) {
627 		pfn_cuGetErrorString(res, &(err_string));
628 		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
629 				err_string);
630 		rte_errno = EPERM;
631 		return -rte_errno;
632 	}
633 
634 	/* Align memory address */
635 	mem_alloc_list_tail->ptr_d = mem_alloc_list_tail->ptr_orig_d;
636 	if (align && ((uintptr_t)mem_alloc_list_tail->ptr_d) % align)
637 		mem_alloc_list_tail->ptr_d += (align -
638 				(((uintptr_t)mem_alloc_list_tail->ptr_d) % align));
639 
640 	/* GPUDirect RDMA attribute required */
641 	res = pfn_cuPointerSetAttribute(&flag,
642 			CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
643 			mem_alloc_list_tail->ptr_d);
644 	if (res != 0) {
645 		rte_cuda_log(ERR, "Could not set SYNC MEMOP attribute for "
646 				"GPU memory at  %"PRIu32", err %d",
647 				(uint32_t)mem_alloc_list_tail->ptr_d, res);
648 		rte_errno = EPERM;
649 		return -rte_errno;
650 	}
651 
652 	mem_alloc_list_tail->pkey = get_hash_from_ptr((void *)mem_alloc_list_tail->ptr_d);
653 	mem_alloc_list_tail->ptr_h = NULL;
654 	mem_alloc_list_tail->dev = dev;
655 	mem_alloc_list_tail->ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
656 	mem_alloc_list_tail->mtype = GPU_MEM;
657 
658 	/* Restore original ctx as current ctx */
659 	res = pfn_cuCtxSetCurrent(current_ctx);
660 	if (res != 0) {
661 		pfn_cuGetErrorString(res, &(err_string));
662 		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
663 				err_string);
664 		rte_errno = EPERM;
665 		return -rte_errno;
666 	}
667 
668 	*ptr = (void *)mem_alloc_list_tail->ptr_d;
669 
670 	return 0;
671 }
672 
673 static int
674 cuda_mem_register(struct rte_gpu *dev, size_t size, void *ptr)
675 {
676 	CUresult res;
677 	const char *err_string;
678 	CUcontext current_ctx;
679 	CUcontext input_ctx;
680 	unsigned int flag = 1;
681 	int use_ptr_h = 0;
682 
683 	if (dev == NULL)
684 		return -ENODEV;
685 
686 	/* Store current ctx */
687 	res = pfn_cuCtxGetCurrent(&current_ctx);
688 	if (res != 0) {
689 		pfn_cuGetErrorString(res, &(err_string));
690 		rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
691 				err_string);
692 		rte_errno = EPERM;
693 		return -rte_errno;
694 	}
695 
696 	/* Set child ctx as current ctx */
697 	input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
698 	res = pfn_cuCtxSetCurrent(input_ctx);
699 	if (res != 0) {
700 		pfn_cuGetErrorString(res, &(err_string));
701 		rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
702 				err_string);
703 		rte_errno = EPERM;
704 		return -rte_errno;
705 	}
706 
707 	/* Get next memory list item */
708 	mem_alloc_list_tail = mem_list_add_item();
709 	if (mem_alloc_list_tail == NULL) {
710 		rte_errno = EPERM;
711 		return -rte_errno;
712 	}
713 
714 	/* Allocate memory */
715 	mem_alloc_list_tail->size = size;
716 	mem_alloc_list_tail->ptr_h = ptr;
717 
718 	res = pfn_cuMemHostRegister(mem_alloc_list_tail->ptr_h,
719 			mem_alloc_list_tail->size,
720 			CU_MEMHOSTREGISTER_PORTABLE |
721 			CU_MEMHOSTREGISTER_DEVICEMAP);
722 	if (res != 0) {
723 		pfn_cuGetErrorString(res, &(err_string));
724 		rte_cuda_log(ERR, "cuMemHostRegister failed with %s ptr %p size %zd",
725 				err_string,
726 				mem_alloc_list_tail->ptr_h,
727 				mem_alloc_list_tail->size);
728 		rte_errno = EPERM;
729 		return -rte_errno;
730 	}
731 
732 	res = pfn_cuDeviceGetAttribute(&(use_ptr_h),
733 			CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM,
734 			((struct cuda_info *)(dev->mpshared->dev_private))->cu_dev);
735 	if (res != 0) {
736 		pfn_cuGetErrorString(res, &(err_string));
737 		rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
738 				err_string);
739 		rte_errno = EPERM;
740 		return -rte_errno;
741 	}
742 
743 	if (use_ptr_h == 0) {
744 		res = pfn_cuMemHostGetDevicePointer(&(mem_alloc_list_tail->ptr_d),
745 				mem_alloc_list_tail->ptr_h, 0);
746 		if (res != 0) {
747 			pfn_cuGetErrorString(res, &(err_string));
748 			rte_cuda_log(ERR, "cuMemHostGetDevicePointer failed with %s",
749 					err_string);
750 			rte_errno = EPERM;
751 			return -rte_errno;
752 		}
753 
754 		if ((uintptr_t)mem_alloc_list_tail->ptr_d !=
755 				(uintptr_t)mem_alloc_list_tail->ptr_h) {
756 			rte_cuda_log(ERR, "Host input pointer is different wrt GPU registered pointer");
757 			rte_errno = ENOTSUP;
758 			return -rte_errno;
759 		}
760 	} else {
761 		mem_alloc_list_tail->ptr_d = (CUdeviceptr)mem_alloc_list_tail->ptr_h;
762 	}
763 
764 	/* GPUDirect RDMA attribute required */
765 	res = pfn_cuPointerSetAttribute(&flag,
766 			CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
767 			mem_alloc_list_tail->ptr_d);
768 	if (res != 0) {
769 		rte_cuda_log(ERR, "Could not set SYNC MEMOP attribute for GPU memory at %"PRIu32
770 				", err %d", (uint32_t)mem_alloc_list_tail->ptr_d, res);
771 		rte_errno = EPERM;
772 		return -rte_errno;
773 	}
774 
775 	mem_alloc_list_tail->pkey = get_hash_from_ptr((void *)mem_alloc_list_tail->ptr_h);
776 	mem_alloc_list_tail->size = size;
777 	mem_alloc_list_tail->dev = dev;
778 	mem_alloc_list_tail->ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
779 	mem_alloc_list_tail->mtype = CPU_REGISTERED;
780 	mem_alloc_list_tail->ptr_orig_d = mem_alloc_list_tail->ptr_d;
781 
782 	/* Restore original ctx as current ctx */
783 	res = pfn_cuCtxSetCurrent(current_ctx);
784 	if (res != 0) {
785 		pfn_cuGetErrorString(res, &(err_string));
786 		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
787 				err_string);
788 		rte_errno = EPERM;
789 		return -rte_errno;
790 	}
791 
792 	return 0;
793 }
794 
795 static int
796 cuda_mem_free(struct rte_gpu *dev, void *ptr)
797 {
798 	CUresult res;
799 	struct mem_entry *mem_item;
800 	const char *err_string;
801 	cuda_ptr_key hk;
802 
803 	if (dev == NULL)
804 		return -ENODEV;
805 
806 	hk = get_hash_from_ptr((void *)ptr);
807 
808 	mem_item = mem_list_find_item(hk);
809 	if (mem_item == NULL) {
810 		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
811 		rte_errno = EPERM;
812 		return -rte_errno;
813 	}
814 
815 	if (mem_item->mtype == GPU_MEM) {
816 		res = pfn_cuMemFree(mem_item->ptr_orig_d);
817 		if (res != 0) {
818 			pfn_cuGetErrorString(res, &(err_string));
819 			rte_cuda_log(ERR, "cuMemFree current failed with %s",
820 					err_string);
821 			rte_errno = EPERM;
822 			return -rte_errno;
823 		}
824 
825 		return mem_list_del_item(hk);
826 	}
827 
828 	rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
829 
830 	return -EPERM;
831 }
832 
833 static int
834 cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
835 {
836 	CUresult res;
837 	struct mem_entry *mem_item;
838 	const char *err_string;
839 	cuda_ptr_key hk;
840 
841 	if (dev == NULL)
842 		return -ENODEV;
843 
844 	hk = get_hash_from_ptr((void *)ptr);
845 
846 	mem_item = mem_list_find_item(hk);
847 	if (mem_item == NULL) {
848 		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
849 		rte_errno = EPERM;
850 		return -rte_errno;
851 	}
852 
853 	if (mem_item->mtype == CPU_REGISTERED) {
854 		res = pfn_cuMemHostUnregister(ptr);
855 		if (res != 0) {
856 			pfn_cuGetErrorString(res, &(err_string));
857 			rte_cuda_log(ERR, "cuMemHostUnregister current failed with %s",
858 					err_string);
859 			rte_errno = EPERM;
860 			return -rte_errno;
861 		}
862 
863 		return mem_list_del_item(hk);
864 	}
865 
866 	rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
867 
868 	rte_errno = EPERM;
869 	return -rte_errno;
870 }
871 
872 static int
873 cuda_dev_close(struct rte_gpu *dev)
874 {
875 	if (dev == NULL)
876 		return -EINVAL;
877 
878 	rte_free(dev->mpshared->dev_private);
879 
880 	return 0;
881 }
882 
883 static int
884 cuda_wmb(struct rte_gpu *dev)
885 {
886 	CUresult res;
887 	const char *err_string;
888 	CUcontext current_ctx;
889 	CUcontext input_ctx;
890 	struct cuda_info *private;
891 
892 	if (dev == NULL) {
893 		rte_errno = ENODEV;
894 		return -rte_errno;
895 	}
896 
897 	private = (struct cuda_info *)dev->mpshared->dev_private;
898 
899 	if (private->gdr_write_ordering != CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE) {
900 		/*
901 		 * No need to explicitly force the write ordering because
902 		 * the device natively supports it
903 		 */
904 		return 0;
905 	}
906 
907 	if (private->gdr_flush_type != CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST) {
908 		/*
909 		 * Can't flush GDR writes with cuFlushGPUDirectRDMAWrites CUDA function.
910 		 * Application needs to use alternative methods.
911 		 */
912 		rte_cuda_log(WARNING, "Can't flush GDR writes with cuFlushGPUDirectRDMAWrites CUDA function."
913 				"Application needs to use alternative methods.");
914 
915 		rte_errno = ENOTSUP;
916 		return -rte_errno;
917 	}
918 
919 	/* Store current ctx */
920 	res = pfn_cuCtxGetCurrent(&current_ctx);
921 	if (res != 0) {
922 		pfn_cuGetErrorString(res, &(err_string));
923 		rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
924 				err_string);
925 		rte_errno = EPERM;
926 		return -rte_errno;
927 	}
928 
929 	/* Set child ctx as current ctx */
930 	input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
931 	res = pfn_cuCtxSetCurrent(input_ctx);
932 	if (res != 0) {
933 		pfn_cuGetErrorString(res, &(err_string));
934 		rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
935 				err_string);
936 		rte_errno = EPERM;
937 		return -rte_errno;
938 	}
939 
940 	res = pfn_cuFlushGPUDirectRDMAWrites(CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TARGET_CURRENT_CTX,
941 			CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TO_ALL_DEVICES);
942 	if (res != 0) {
943 		pfn_cuGetErrorString(res, &(err_string));
944 		rte_cuda_log(ERR, "cuFlushGPUDirectRDMAWrites current failed with %s",
945 				err_string);
946 		rte_errno = EPERM;
947 		return -rte_errno;
948 	}
949 
950 	/* Restore original ctx as current ctx */
951 	res = pfn_cuCtxSetCurrent(current_ctx);
952 	if (res != 0) {
953 		pfn_cuGetErrorString(res, &(err_string));
954 		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
955 				err_string);
956 		rte_errno = EPERM;
957 		return -rte_errno;
958 	}
959 
960 	return 0;
961 }
962 
963 static int
964 cuda_gpu_probe(__rte_unused struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
965 {
966 	struct rte_gpu *dev = NULL;
967 	CUresult res;
968 	CUdevice cu_dev_id;
969 	CUcontext pctx;
970 	char dev_name[RTE_DEV_NAME_MAX_LEN];
971 	const char *err_string;
972 	int processor_count = 0;
973 	struct cuda_info *private;
974 
975 	if (pci_dev == NULL) {
976 		rte_cuda_log(ERR, "NULL PCI device");
977 		rte_errno = ENODEV;
978 		return -rte_errno;
979 	}
980 
981 	rte_pci_device_name(&pci_dev->addr, dev_name, sizeof(dev_name));
982 
983 	/* Allocate memory to be used privately by drivers */
984 	dev = rte_gpu_allocate(pci_dev->device.name);
985 	if (dev == NULL) {
986 		rte_errno = ENODEV;
987 		return -rte_errno;
988 	}
989 
990 	/* Initialize values only for the first CUDA driver call */
991 	if (dev->mpshared->info.dev_id == 0) {
992 		mem_alloc_list_head = NULL;
993 		mem_alloc_list_tail = NULL;
994 		mem_alloc_list_last_elem = 0;
995 
996 		/* Load libcuda.so library */
997 		if (cuda_loader()) {
998 			rte_cuda_log(ERR, "CUDA Driver library not found");
999 			rte_errno = ENOTSUP;
1000 			return -rte_errno;
1001 		}
1002 
1003 		/* Load initial CUDA functions */
1004 		if (cuda_sym_func_loader()) {
1005 			rte_cuda_log(ERR, "CUDA functions not found in library");
1006 			rte_errno = ENOTSUP;
1007 			return -rte_errno;
1008 		}
1009 
1010 		/*
1011 		 * Required to initialize the CUDA Driver.
1012 		 * Multiple calls of cuInit() will return immediately
1013 		 * without making any relevant change
1014 		 */
1015 		sym_cuInit(0);
1016 
1017 		res = sym_cuDriverGetVersion(&cuda_driver_version);
1018 		if (res != 0) {
1019 			rte_cuda_log(ERR, "cuDriverGetVersion failed with %d", res);
1020 			rte_errno = ENOTSUP;
1021 			return -rte_errno;
1022 		}
1023 
1024 		if (cuda_driver_version < CUDA_DRIVER_MIN_VERSION) {
1025 			rte_cuda_log(ERR, "CUDA Driver version found is %d. "
1026 					"Minimum requirement is %d",
1027 					cuda_driver_version,
1028 					CUDA_DRIVER_MIN_VERSION);
1029 			rte_errno = ENOTSUP;
1030 			return -rte_errno;
1031 		}
1032 
1033 		if (cuda_pfn_func_loader()) {
1034 			rte_cuda_log(ERR, "CUDA PFN functions not found in library");
1035 			rte_errno = ENOTSUP;
1036 			return -rte_errno;
1037 		}
1038 	}
1039 
1040 	/* Fill HW specific part of device structure */
1041 	dev->device = &pci_dev->device;
1042 	dev->mpshared->info.numa_node = pci_dev->device.numa_node;
1043 
1044 	/* Get NVIDIA GPU Device descriptor */
1045 	res = pfn_cuDeviceGetByPCIBusId(&cu_dev_id, dev->device->name);
1046 	if (res != 0) {
1047 		pfn_cuGetErrorString(res, &(err_string));
1048 		rte_cuda_log(ERR, "cuDeviceGetByPCIBusId name %s failed with %d: %s",
1049 				dev->device->name, res, err_string);
1050 		rte_errno = EPERM;
1051 		return -rte_errno;
1052 	}
1053 
1054 	res = pfn_cuDevicePrimaryCtxRetain(&pctx, cu_dev_id);
1055 	if (res != 0) {
1056 		pfn_cuGetErrorString(res, &(err_string));
1057 		rte_cuda_log(ERR, "cuDevicePrimaryCtxRetain name %s failed with %d: %s",
1058 				dev->device->name, res, err_string);
1059 		rte_errno = EPERM;
1060 		return -rte_errno;
1061 	}
1062 
1063 	res = pfn_cuCtxGetApiVersion(pctx, &cuda_api_version);
1064 	if (res != 0) {
1065 		rte_cuda_log(ERR, "cuCtxGetApiVersion failed with %d", res);
1066 		rte_errno = ENOTSUP;
1067 		return -rte_errno;
1068 	}
1069 
1070 	if (cuda_api_version < CUDA_API_MIN_VERSION) {
1071 		rte_cuda_log(ERR, "CUDA API version found is %d Minimum requirement is %d",
1072 				cuda_api_version, CUDA_API_MIN_VERSION);
1073 		rte_errno = ENOTSUP;
1074 		return -rte_errno;
1075 	}
1076 
1077 	dev->mpshared->info.context = (uint64_t)pctx;
1078 
1079 	/*
1080 	 * GPU Device generic info
1081 	 */
1082 
1083 	/* Processor count */
1084 	res = pfn_cuDeviceGetAttribute(&(processor_count),
1085 			CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
1086 			cu_dev_id);
1087 	if (res != 0) {
1088 		pfn_cuGetErrorString(res, &(err_string));
1089 		rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
1090 				err_string);
1091 		rte_errno = EPERM;
1092 		return -rte_errno;
1093 	}
1094 	dev->mpshared->info.processor_count = (uint32_t)processor_count;
1095 
1096 	/* Total memory */
1097 	res = pfn_cuDeviceTotalMem(&dev->mpshared->info.total_memory, cu_dev_id);
1098 	if (res != 0) {
1099 		pfn_cuGetErrorString(res, &(err_string));
1100 		rte_cuda_log(ERR, "cuDeviceTotalMem failed with %s",
1101 				err_string);
1102 		rte_errno = EPERM;
1103 		return -rte_errno;
1104 	}
1105 
1106 	/*
1107 	 * GPU Device private info
1108 	 */
1109 	dev->mpshared->dev_private = rte_zmalloc(NULL,
1110 			sizeof(struct cuda_info),
1111 			RTE_CACHE_LINE_SIZE);
1112 	if (dev->mpshared->dev_private == NULL) {
1113 		rte_cuda_log(ERR, "Failed to allocate memory for GPU process private");
1114 		rte_errno = EPERM;
1115 		return -rte_errno;
1116 	}
1117 
1118 	private = (struct cuda_info *)dev->mpshared->dev_private;
1119 	private->cu_dev = cu_dev_id;
1120 	res = pfn_cuDeviceGetName(private->gpu_name,
1121 			RTE_DEV_NAME_MAX_LEN,
1122 			cu_dev_id);
1123 	if (res != 0) {
1124 		pfn_cuGetErrorString(res, &(err_string));
1125 		rte_cuda_log(ERR, "cuDeviceGetName failed with %s",
1126 				err_string);
1127 		rte_errno = EPERM;
1128 		return -rte_errno;
1129 	}
1130 
1131 	res = pfn_cuDeviceGetAttribute(&(private->gdr_supported),
1132 			CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED,
1133 			cu_dev_id);
1134 	if (res != 0) {
1135 		pfn_cuGetErrorString(res, &(err_string));
1136 		rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
1137 				err_string);
1138 		rte_errno = EPERM;
1139 		return -rte_errno;
1140 	}
1141 
1142 	if (private->gdr_supported == 0)
1143 		rte_cuda_log(WARNING, "GPU %s doesn't support GPUDirect RDMA",
1144 				pci_dev->device.name);
1145 
1146 	res = pfn_cuDeviceGetAttribute(&(private->gdr_write_ordering),
1147 			CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING,
1148 			cu_dev_id);
1149 	if (res != 0) {
1150 		pfn_cuGetErrorString(res, &(err_string));
1151 		rte_cuda_log(ERR,
1152 				"cuDeviceGetAttribute failed with %s",
1153 				err_string);
1154 		rte_errno = EPERM;
1155 		return -rte_errno;
1156 	}
1157 
1158 	if (private->gdr_write_ordering == CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE) {
1159 		res = pfn_cuDeviceGetAttribute(&(private->gdr_flush_type),
1160 				CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS,
1161 				cu_dev_id);
1162 		if (res != 0) {
1163 			pfn_cuGetErrorString(res, &(err_string));
1164 			rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
1165 					err_string);
1166 			rte_errno = EPERM;
1167 			return -rte_errno;
1168 		}
1169 
1170 		if (private->gdr_flush_type != CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST)
1171 			rte_cuda_log(ERR, "GPUDirect RDMA flush writes API is not supported");
1172 	}
1173 
1174 	dev->ops.dev_info_get = cuda_dev_info_get;
1175 	dev->ops.dev_close = cuda_dev_close;
1176 	dev->ops.mem_alloc = cuda_mem_alloc;
1177 	dev->ops.mem_free = cuda_mem_free;
1178 	dev->ops.mem_register = cuda_mem_register;
1179 	dev->ops.mem_unregister = cuda_mem_unregister;
1180 	dev->ops.wmb = cuda_wmb;
1181 
1182 	rte_gpu_complete_new(dev);
1183 
1184 	rte_cuda_debug("dev id = %u name = %s",
1185 			dev->mpshared->info.dev_id, private->gpu_name);
1186 
1187 	return 0;
1188 }
1189 
1190 static int
1191 cuda_gpu_remove(struct rte_pci_device *pci_dev)
1192 {
1193 	struct rte_gpu *dev;
1194 	int ret;
1195 	uint8_t gpu_id;
1196 
1197 	if (pci_dev == NULL) {
1198 		rte_errno = ENODEV;
1199 		return -rte_errno;
1200 	}
1201 
1202 	dev = rte_gpu_get_by_name(pci_dev->device.name);
1203 	if (dev == NULL) {
1204 		rte_cuda_log(ERR, "Couldn't find HW dev \"%s\" to uninitialise it",
1205 				pci_dev->device.name);
1206 		rte_errno = ENODEV;
1207 		return -rte_errno;
1208 	}
1209 	gpu_id = dev->mpshared->info.dev_id;
1210 
1211 	/* release dev from library */
1212 	ret = rte_gpu_release(dev);
1213 	if (ret)
1214 		rte_cuda_log(ERR, "Device %i failed to uninit: %i", gpu_id, ret);
1215 
1216 	rte_cuda_debug("Destroyed dev = %u", gpu_id);
1217 
1218 	return 0;
1219 }
1220 
1221 static struct rte_pci_driver rte_cuda_driver = {
1222 	.id_table = pci_id_cuda_map,
1223 	.drv_flags = RTE_PCI_DRV_WC_ACTIVATE,
1224 	.probe = cuda_gpu_probe,
1225 	.remove = cuda_gpu_remove,
1226 };
1227 
1228 RTE_PMD_REGISTER_PCI(gpu_cuda, rte_cuda_driver);
1229 RTE_PMD_REGISTER_PCI_TABLE(gpu_cuda, pci_id_cuda_map);
1230 RTE_PMD_REGISTER_KMOD_DEP(gpu_cuda, "* nvidia & (nv_peer_mem | nvpeer_mem)");
1231