1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright (c) 2021 NVIDIA Corporation & Affiliates
7 #include <rte_malloc.h>
9 #include <rte_bus_pci.h>
10 #include <rte_byteorder.h>
13 #include <gpudev_driver.h>
16 #include <cudaTypedefs.h>
20 #define CUDA_DRIVER_MIN_VERSION 11040
21 #define CUDA_API_MIN_VERSION 3020
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);
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;
52 static unsigned int cuda_api_version;
53 static int cuda_driver_version;
56 /* NVIDIA GPU vendor */
57 #define NVIDIA_GPU_VENDOR_ID (0x10de)
59 /* NVIDIA GPU device IDs */
60 #define NVIDIA_GPU_A100_40GB_DEVICE_ID (0x20f1)
61 #define NVIDIA_GPU_A100_80GB_DEVICE_ID (0x20b5)
62 #define NVIDIA_GPU_A100_80GB_DPU_DEVICE_ID (0x20b8)
64 #define NVIDIA_GPU_A30_24GB_DEVICE_ID (0x20b7)
65 #define NVIDIA_GPU_A10_24GB_DEVICE_ID (0x2236)
67 #define NVIDIA_GPU_V100_32GB_SXM_DEVICE_ID (0x1db5)
68 #define NVIDIA_GPU_V100_32GB_PCIE_DEVICE_ID (0x1db6)
69 #define NVIDIA_GPU_V100_16GB_DEVICE_ID (0x1db4)
71 #define NVIDIA_GPU_T4_16GB_DEVICE_ID (0x1eb8)
73 #define CUDA_MAX_ALLOCATION_NUM 512
75 #define GPU_PAGE_SHIFT 16
76 #define GPU_PAGE_SIZE (1UL << GPU_PAGE_SHIFT)
78 RTE_LOG_REGISTER_DEFAULT(cuda_logtype, NOTICE);
80 /* NVIDIA GPU address map */
81 static const struct rte_pci_id pci_id_cuda_map[] = {
83 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
84 NVIDIA_GPU_A100_40GB_DEVICE_ID)
87 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
88 NVIDIA_GPU_A100_80GB_DEVICE_ID)
91 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
92 NVIDIA_GPU_A100_80GB_DPU_DEVICE_ID)
95 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
96 NVIDIA_GPU_A30_24GB_DEVICE_ID)
99 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
100 NVIDIA_GPU_A10_24GB_DEVICE_ID)
103 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
104 NVIDIA_GPU_V100_32GB_SXM_DEVICE_ID)
107 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
108 NVIDIA_GPU_V100_32GB_PCIE_DEVICE_ID)
111 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
112 NVIDIA_GPU_V100_16GB_DEVICE_ID)
115 RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
116 NVIDIA_GPU_T4_16GB_DEVICE_ID)
123 /* Device private info */
125 char gpu_name[RTE_DEV_NAME_MAX_LEN];
128 int gdr_write_ordering;
132 /* Type of memory allocated by CUDA driver */
136 GPU_REGISTERED /* Not used yet */
139 /* key associated to a memory address */
140 typedef uintptr_t cuda_ptr_key;
142 /* Single entry of the memory list */
145 CUdeviceptr ptr_orig_d;
154 struct mem_entry *prev;
155 struct mem_entry *next;
158 static struct mem_entry *mem_alloc_list_head;
159 static struct mem_entry *mem_alloc_list_tail;
160 static uint32_t mem_alloc_list_last_elem;
162 /* Load the CUDA symbols */
167 char cuda_path[1024];
169 if (getenv("CUDA_PATH_L") == NULL)
170 snprintf(cuda_path, 1024, "%s", "libcuda.so");
172 snprintf(cuda_path, 1024, "%s/%s", getenv("CUDA_PATH_L"), "libcuda.so");
174 cudalib = dlopen(cuda_path, RTLD_LAZY);
175 if (cudalib == NULL) {
176 rte_cuda_log(ERR, "Failed to find CUDA library in %s (CUDA_PATH_L=%s)",
177 cuda_path, getenv("CUDA_PATH_L"));
185 cuda_sym_func_loader(void)
190 sym_cuInit = dlsym(cudalib, "cuInit");
191 if (sym_cuInit == NULL) {
192 rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuInit");
196 sym_cuDriverGetVersion = dlsym(cudalib, "cuDriverGetVersion");
197 if (sym_cuDriverGetVersion == NULL) {
198 rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuDriverGetVersion");
202 sym_cuGetProcAddress = dlsym(cudalib, "cuGetProcAddress");
203 if (sym_cuGetProcAddress == NULL) {
204 rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuGetProcAddress");
212 cuda_pfn_func_loader(void)
216 res = sym_cuGetProcAddress("cuGetErrorString",
217 (void **) (&pfn_cuGetErrorString), cuda_driver_version, 0);
219 rte_cuda_log(ERR, "Retrieve pfn_cuGetErrorString failed with %d", res);
223 res = sym_cuGetProcAddress("cuGetErrorName",
224 (void **)(&pfn_cuGetErrorName), cuda_driver_version, 0);
226 rte_cuda_log(ERR, "Retrieve pfn_cuGetErrorName failed with %d", res);
230 res = sym_cuGetProcAddress("cuPointerSetAttribute",
231 (void **)(&pfn_cuPointerSetAttribute), cuda_driver_version, 0);
233 rte_cuda_log(ERR, "Retrieve pfn_cuPointerSetAttribute failed with %d", res);
237 res = sym_cuGetProcAddress("cuDeviceGetAttribute",
238 (void **)(&pfn_cuDeviceGetAttribute), cuda_driver_version, 0);
240 rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetAttribute failed with %d", res);
244 res = sym_cuGetProcAddress("cuDeviceGetByPCIBusId",
245 (void **)(&pfn_cuDeviceGetByPCIBusId), cuda_driver_version, 0);
247 rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetByPCIBusId failed with %d", res);
251 res = sym_cuGetProcAddress("cuDeviceGetName",
252 (void **)(&pfn_cuDeviceGetName), cuda_driver_version, 0);
254 rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetName failed with %d", res);
258 res = sym_cuGetProcAddress("cuDevicePrimaryCtxRetain",
259 (void **)(&pfn_cuDevicePrimaryCtxRetain), cuda_driver_version, 0);
261 rte_cuda_log(ERR, "Retrieve pfn_cuDevicePrimaryCtxRetain failed with %d", res);
265 res = sym_cuGetProcAddress("cuDevicePrimaryCtxRelease",
266 (void **)(&pfn_cuDevicePrimaryCtxRelease), cuda_driver_version, 0);
268 rte_cuda_log(ERR, "Retrieve pfn_cuDevicePrimaryCtxRelease failed with %d", res);
272 res = sym_cuGetProcAddress("cuDeviceTotalMem",
273 (void **)(&pfn_cuDeviceTotalMem), cuda_driver_version, 0);
275 rte_cuda_log(ERR, "Retrieve pfn_cuDeviceTotalMem failed with %d", res);
279 res = sym_cuGetProcAddress("cuCtxGetApiVersion",
280 (void **)(&pfn_cuCtxGetApiVersion), cuda_driver_version, 0);
282 rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetApiVersion failed with %d", res);
286 res = sym_cuGetProcAddress("cuCtxGetDevice",
287 (void **)(&pfn_cuCtxGetDevice), cuda_driver_version, 0);
289 rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetDevice failed with %d", res);
293 res = sym_cuGetProcAddress("cuCtxSetCurrent",
294 (void **)(&pfn_cuCtxSetCurrent), cuda_driver_version, 0);
296 rte_cuda_log(ERR, "Retrieve pfn_cuCtxSetCurrent failed with %d", res);
300 res = sym_cuGetProcAddress("cuCtxGetCurrent",
301 (void **)(&pfn_cuCtxGetCurrent), cuda_driver_version, 0);
303 rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetCurrent failed with %d", res);
307 res = sym_cuGetProcAddress("cuCtxGetExecAffinity",
308 (void **)(&pfn_cuCtxGetExecAffinity), cuda_driver_version, 0);
310 rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetExecAffinity failed with %d", res);
314 res = sym_cuGetProcAddress("cuMemAlloc",
315 (void **)(&pfn_cuMemAlloc), cuda_driver_version, 0);
317 rte_cuda_log(ERR, "Retrieve pfn_cuMemAlloc failed with %d", res);
321 res = sym_cuGetProcAddress("cuMemFree",
322 (void **)(&pfn_cuMemFree), cuda_driver_version, 0);
324 rte_cuda_log(ERR, "Retrieve pfn_cuMemFree failed with %d", res);
328 res = sym_cuGetProcAddress("cuMemHostRegister",
329 (void **)(&pfn_cuMemHostRegister), cuda_driver_version, 0);
331 rte_cuda_log(ERR, "Retrieve pfn_cuMemHostRegister failed with %d", res);
335 res = sym_cuGetProcAddress("cuMemHostUnregister",
336 (void **)(&pfn_cuMemHostUnregister), cuda_driver_version, 0);
338 rte_cuda_log(ERR, "Retrieve pfn_cuMemHostUnregister failed with %d", res);
342 res = sym_cuGetProcAddress("cuMemHostGetDevicePointer",
343 (void **)(&pfn_cuMemHostGetDevicePointer), cuda_driver_version, 0);
345 rte_cuda_log(ERR, "Retrieve pfn_cuMemHostGetDevicePointer failed with %d", res);
349 res = sym_cuGetProcAddress("cuFlushGPUDirectRDMAWrites",
350 (void **)(&pfn_cuFlushGPUDirectRDMAWrites), cuda_driver_version, 0);
352 rte_cuda_log(ERR, "Retrieve cuFlushGPUDirectRDMAWrites failed with %d", res);
359 /* Generate a key from a memory pointer */
361 get_hash_from_ptr(void *ptr)
363 return (uintptr_t)ptr;
367 mem_list_count_item(void)
369 return mem_alloc_list_last_elem;
372 /* Initiate list of memory allocations if not done yet */
373 static struct mem_entry *
374 mem_list_add_item(void)
376 /* Initiate list of memory allocations if not done yet */
377 if (mem_alloc_list_head == NULL) {
378 mem_alloc_list_head = rte_zmalloc(NULL,
379 sizeof(struct mem_entry),
380 RTE_CACHE_LINE_SIZE);
381 if (mem_alloc_list_head == NULL) {
382 rte_cuda_log(ERR, "Failed to allocate memory for memory list");
386 mem_alloc_list_head->next = NULL;
387 mem_alloc_list_head->prev = NULL;
388 mem_alloc_list_tail = mem_alloc_list_head;
390 struct mem_entry *mem_alloc_list_cur = rte_zmalloc(NULL,
391 sizeof(struct mem_entry),
392 RTE_CACHE_LINE_SIZE);
394 if (mem_alloc_list_cur == NULL) {
395 rte_cuda_log(ERR, "Failed to allocate memory for memory list");
399 mem_alloc_list_tail->next = mem_alloc_list_cur;
400 mem_alloc_list_cur->prev = mem_alloc_list_tail;
401 mem_alloc_list_tail = mem_alloc_list_tail->next;
402 mem_alloc_list_tail->next = NULL;
405 mem_alloc_list_last_elem++;
407 return mem_alloc_list_tail;
410 static struct mem_entry *
411 mem_list_find_item(cuda_ptr_key pk)
413 struct mem_entry *mem_alloc_list_cur = NULL;
415 if (mem_alloc_list_head == NULL) {
416 rte_cuda_log(ERR, "Memory list doesn't exist");
420 if (mem_list_count_item() == 0) {
421 rte_cuda_log(ERR, "No items in memory list");
425 mem_alloc_list_cur = mem_alloc_list_head;
427 while (mem_alloc_list_cur != NULL) {
428 if (mem_alloc_list_cur->pkey == pk)
429 return mem_alloc_list_cur;
430 mem_alloc_list_cur = mem_alloc_list_cur->next;
433 return mem_alloc_list_cur;
437 mem_list_del_item(cuda_ptr_key pk)
439 struct mem_entry *mem_alloc_list_cur = NULL;
441 mem_alloc_list_cur = mem_list_find_item(pk);
442 if (mem_alloc_list_cur == NULL)
445 /* if key is in head */
446 if (mem_alloc_list_cur->prev == NULL) {
447 mem_alloc_list_head = mem_alloc_list_cur->next;
448 if (mem_alloc_list_head != NULL)
449 mem_alloc_list_head->prev = NULL;
451 mem_alloc_list_cur->prev->next = mem_alloc_list_cur->next;
452 if (mem_alloc_list_cur->next != NULL)
453 mem_alloc_list_cur->next->prev = mem_alloc_list_cur->prev;
456 rte_free(mem_alloc_list_cur);
458 mem_alloc_list_last_elem--;
464 cuda_dev_info_get(struct rte_gpu *dev, struct rte_gpu_info *info)
468 struct rte_gpu_info parent_info;
469 CUexecAffinityParam affinityPrm;
470 const char *err_string;
471 struct cuda_info *private;
472 CUcontext current_ctx;
480 /* Child initialization time probably called by rte_gpu_add_child() */
481 if (dev->mpshared->info.parent != RTE_GPU_ID_NONE &&
482 dev->mpshared->dev_private == NULL) {
483 /* Store current ctx */
484 res = pfn_cuCtxGetCurrent(¤t_ctx);
486 pfn_cuGetErrorString(res, &(err_string));
487 rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
493 /* Set child ctx as current ctx */
494 input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
495 res = pfn_cuCtxSetCurrent(input_ctx);
497 pfn_cuGetErrorString(res, &(err_string));
498 rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
509 res = pfn_cuCtxGetExecAffinity(&affinityPrm,
510 CU_EXEC_AFFINITY_TYPE_SM_COUNT);
512 pfn_cuGetErrorString(res, &(err_string));
513 rte_cuda_log(ERR, "cuCtxGetExecAffinity failed with %s",
516 dev->mpshared->info.processor_count =
517 (uint32_t)affinityPrm.param.smCount.val;
519 ret = rte_gpu_info_get(dev->mpshared->info.parent, &parent_info);
524 dev->mpshared->info.total_memory = parent_info.total_memory;
526 dev->mpshared->info.page_size = parent_info.page_size;
529 * GPU Device private info
531 dev->mpshared->dev_private = rte_zmalloc(NULL,
532 sizeof(struct cuda_info),
533 RTE_CACHE_LINE_SIZE);
534 if (dev->mpshared->dev_private == NULL) {
535 rte_cuda_log(ERR, "Failed to allocate memory for GPU process private");
540 private = (struct cuda_info *)dev->mpshared->dev_private;
542 res = pfn_cuCtxGetDevice(&(private->cu_dev));
544 pfn_cuGetErrorString(res, &(err_string));
545 rte_cuda_log(ERR, "cuCtxGetDevice failed with %s",
551 res = pfn_cuDeviceGetName(private->gpu_name,
552 RTE_DEV_NAME_MAX_LEN, private->cu_dev);
554 pfn_cuGetErrorString(res, &(err_string));
555 rte_cuda_log(ERR, "cuDeviceGetName failed with %s",
561 /* Restore original ctx as current ctx */
562 res = pfn_cuCtxSetCurrent(current_ctx);
564 pfn_cuGetErrorString(res, &(err_string));
565 rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
572 *info = dev->mpshared->info;
582 cuda_mem_alloc(struct rte_gpu *dev, size_t size, unsigned int align, void **ptr)
585 const char *err_string;
586 CUcontext current_ctx;
588 unsigned int flag = 1;
593 /* Store current ctx */
594 res = pfn_cuCtxGetCurrent(¤t_ctx);
596 pfn_cuGetErrorString(res, &(err_string));
597 rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
603 /* Set child ctx as current ctx */
604 input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
605 res = pfn_cuCtxSetCurrent(input_ctx);
607 pfn_cuGetErrorString(res, &(err_string));
608 rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
614 /* Get next memory list item */
615 mem_alloc_list_tail = mem_list_add_item();
616 if (mem_alloc_list_tail == NULL) {
621 /* Allocate memory */
622 mem_alloc_list_tail->size = size;
623 mem_alloc_list_tail->size_orig = size + align;
625 res = pfn_cuMemAlloc(&(mem_alloc_list_tail->ptr_orig_d),
626 mem_alloc_list_tail->size_orig);
628 pfn_cuGetErrorString(res, &(err_string));
629 rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
635 /* Align memory address */
636 mem_alloc_list_tail->ptr_d = mem_alloc_list_tail->ptr_orig_d;
637 if (align && ((uintptr_t)mem_alloc_list_tail->ptr_d) % align)
638 mem_alloc_list_tail->ptr_d += (align -
639 (((uintptr_t)mem_alloc_list_tail->ptr_d) % align));
641 /* GPUDirect RDMA attribute required */
642 res = pfn_cuPointerSetAttribute(&flag,
643 CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
644 mem_alloc_list_tail->ptr_d);
646 rte_cuda_log(ERR, "Could not set SYNC MEMOP attribute for "
647 "GPU memory at %"PRIu32", err %d",
648 (uint32_t)mem_alloc_list_tail->ptr_d, res);
653 mem_alloc_list_tail->pkey = get_hash_from_ptr((void *)mem_alloc_list_tail->ptr_d);
654 mem_alloc_list_tail->ptr_h = NULL;
655 mem_alloc_list_tail->dev = dev;
656 mem_alloc_list_tail->ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
657 mem_alloc_list_tail->mtype = GPU_MEM;
659 /* Restore original ctx as current ctx */
660 res = pfn_cuCtxSetCurrent(current_ctx);
662 pfn_cuGetErrorString(res, &(err_string));
663 rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
669 *ptr = (void *)mem_alloc_list_tail->ptr_d;
675 cuda_mem_register(struct rte_gpu *dev, size_t size, void *ptr)
678 const char *err_string;
679 CUcontext current_ctx;
681 unsigned int flag = 1;
687 /* Store current ctx */
688 res = pfn_cuCtxGetCurrent(¤t_ctx);
690 pfn_cuGetErrorString(res, &(err_string));
691 rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
697 /* Set child ctx as current ctx */
698 input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
699 res = pfn_cuCtxSetCurrent(input_ctx);
701 pfn_cuGetErrorString(res, &(err_string));
702 rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
708 /* Get next memory list item */
709 mem_alloc_list_tail = mem_list_add_item();
710 if (mem_alloc_list_tail == NULL) {
715 /* Allocate memory */
716 mem_alloc_list_tail->size = size;
717 mem_alloc_list_tail->ptr_h = ptr;
719 res = pfn_cuMemHostRegister(mem_alloc_list_tail->ptr_h,
720 mem_alloc_list_tail->size,
721 CU_MEMHOSTREGISTER_PORTABLE |
722 CU_MEMHOSTREGISTER_DEVICEMAP);
724 pfn_cuGetErrorString(res, &(err_string));
725 rte_cuda_log(ERR, "cuMemHostRegister failed with %s ptr %p size %zd",
727 mem_alloc_list_tail->ptr_h,
728 mem_alloc_list_tail->size);
733 res = pfn_cuDeviceGetAttribute(&(use_ptr_h),
734 CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM,
735 ((struct cuda_info *)(dev->mpshared->dev_private))->cu_dev);
737 pfn_cuGetErrorString(res, &(err_string));
738 rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
744 if (use_ptr_h == 0) {
745 res = pfn_cuMemHostGetDevicePointer(&(mem_alloc_list_tail->ptr_d),
746 mem_alloc_list_tail->ptr_h, 0);
748 pfn_cuGetErrorString(res, &(err_string));
749 rte_cuda_log(ERR, "cuMemHostGetDevicePointer failed with %s",
755 if ((uintptr_t)mem_alloc_list_tail->ptr_d !=
756 (uintptr_t)mem_alloc_list_tail->ptr_h) {
757 rte_cuda_log(ERR, "Host input pointer is different wrt GPU registered pointer");
762 mem_alloc_list_tail->ptr_d = (CUdeviceptr)mem_alloc_list_tail->ptr_h;
765 /* GPUDirect RDMA attribute required */
766 res = pfn_cuPointerSetAttribute(&flag,
767 CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
768 mem_alloc_list_tail->ptr_d);
770 rte_cuda_log(ERR, "Could not set SYNC MEMOP attribute for GPU memory at %"PRIu32
771 ", err %d", (uint32_t)mem_alloc_list_tail->ptr_d, res);
776 mem_alloc_list_tail->pkey = get_hash_from_ptr((void *)mem_alloc_list_tail->ptr_h);
777 mem_alloc_list_tail->size = size;
778 mem_alloc_list_tail->dev = dev;
779 mem_alloc_list_tail->ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
780 mem_alloc_list_tail->mtype = CPU_REGISTERED;
781 mem_alloc_list_tail->ptr_orig_d = mem_alloc_list_tail->ptr_d;
783 /* Restore original ctx as current ctx */
784 res = pfn_cuCtxSetCurrent(current_ctx);
786 pfn_cuGetErrorString(res, &(err_string));
787 rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
797 cuda_mem_cpu_map(struct rte_gpu *dev, __rte_unused size_t size, void *ptr_in, void **ptr_out)
799 struct mem_entry *mem_item;
805 hk = get_hash_from_ptr((void *)ptr_in);
807 mem_item = mem_list_find_item(hk);
808 if (mem_item == NULL) {
809 rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory.", ptr_in);
814 if (mem_item->mtype != GPU_MEM) {
815 rte_cuda_log(ERR, "Memory address 0x%p is not GPU memory type.", ptr_in);
820 if (mem_item->size != size)
821 rte_cuda_log(WARNING,
822 "Can't expose memory area with size (%zd) different from original size (%zd).",
823 size, mem_item->size);
825 if (gdrcopy_pin(&gdrc_h, &(mem_item->mh), (uint64_t)mem_item->ptr_d,
826 mem_item->size, &(mem_item->ptr_h))) {
827 rte_cuda_log(ERR, "Error exposing GPU memory address 0x%p.", ptr_in);
832 *ptr_out = mem_item->ptr_h;
838 cuda_mem_free(struct rte_gpu *dev, void *ptr)
841 struct mem_entry *mem_item;
842 const char *err_string;
848 hk = get_hash_from_ptr((void *)ptr);
850 mem_item = mem_list_find_item(hk);
851 if (mem_item == NULL) {
852 rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
857 if (mem_item->mtype == GPU_MEM) {
858 res = pfn_cuMemFree(mem_item->ptr_orig_d);
860 pfn_cuGetErrorString(res, &(err_string));
861 rte_cuda_log(ERR, "cuMemFree current failed with %s",
867 return mem_list_del_item(hk);
870 rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
876 cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
879 struct mem_entry *mem_item;
880 const char *err_string;
886 hk = get_hash_from_ptr((void *)ptr);
888 mem_item = mem_list_find_item(hk);
889 if (mem_item == NULL) {
890 rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
895 if (mem_item->mtype == CPU_REGISTERED) {
896 res = pfn_cuMemHostUnregister(ptr);
898 pfn_cuGetErrorString(res, &(err_string));
899 rte_cuda_log(ERR, "cuMemHostUnregister current failed with %s",
905 return mem_list_del_item(hk);
908 rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
915 cuda_mem_cpu_unmap(struct rte_gpu *dev, void *ptr_in)
917 struct mem_entry *mem_item;
923 hk = get_hash_from_ptr((void *)ptr_in);
925 mem_item = mem_list_find_item(hk);
926 if (mem_item == NULL) {
927 rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory.", ptr_in);
932 if (gdrcopy_unpin(gdrc_h, mem_item->mh, (void *)mem_item->ptr_d,
934 rte_cuda_log(ERR, "Error unexposing GPU memory address 0x%p.", ptr_in);
943 cuda_dev_close(struct rte_gpu *dev)
948 rte_free(dev->mpshared->dev_private);
954 cuda_wmb(struct rte_gpu *dev)
957 const char *err_string;
958 CUcontext current_ctx;
960 struct cuda_info *private;
967 private = (struct cuda_info *)dev->mpshared->dev_private;
969 if (private->gdr_write_ordering != CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE) {
971 * No need to explicitly force the write ordering because
972 * the device natively supports it
977 if (private->gdr_flush_type != CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST) {
979 * Can't flush GDR writes with cuFlushGPUDirectRDMAWrites CUDA function.
980 * Application needs to use alternative methods.
982 rte_cuda_log(WARNING, "Can't flush GDR writes with cuFlushGPUDirectRDMAWrites CUDA function."
983 "Application needs to use alternative methods.");
989 /* Store current ctx */
990 res = pfn_cuCtxGetCurrent(¤t_ctx);
992 pfn_cuGetErrorString(res, &(err_string));
993 rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
999 /* Set child ctx as current ctx */
1000 input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
1001 res = pfn_cuCtxSetCurrent(input_ctx);
1003 pfn_cuGetErrorString(res, &(err_string));
1004 rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
1010 res = pfn_cuFlushGPUDirectRDMAWrites(CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TARGET_CURRENT_CTX,
1011 CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TO_ALL_DEVICES);
1013 pfn_cuGetErrorString(res, &(err_string));
1014 rte_cuda_log(ERR, "cuFlushGPUDirectRDMAWrites current failed with %s",
1020 /* Restore original ctx as current ctx */
1021 res = pfn_cuCtxSetCurrent(current_ctx);
1023 pfn_cuGetErrorString(res, &(err_string));
1024 rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
1034 cuda_gpu_probe(__rte_unused struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
1036 struct rte_gpu *dev = NULL;
1040 char dev_name[RTE_DEV_NAME_MAX_LEN];
1041 const char *err_string;
1042 int processor_count = 0;
1043 struct cuda_info *private;
1045 if (pci_dev == NULL) {
1046 rte_cuda_log(ERR, "NULL PCI device");
1051 rte_pci_device_name(&pci_dev->addr, dev_name, sizeof(dev_name));
1053 /* Allocate memory to be used privately by drivers */
1054 dev = rte_gpu_allocate(pci_dev->device.name);
1060 /* Initialize values only for the first CUDA driver call */
1061 if (dev->mpshared->info.dev_id == 0) {
1062 mem_alloc_list_head = NULL;
1063 mem_alloc_list_tail = NULL;
1064 mem_alloc_list_last_elem = 0;
1066 /* Load libcuda.so library */
1067 if (cuda_loader()) {
1068 rte_cuda_log(ERR, "CUDA Driver library not found");
1069 rte_errno = ENOTSUP;
1073 /* Load initial CUDA functions */
1074 if (cuda_sym_func_loader()) {
1075 rte_cuda_log(ERR, "CUDA functions not found in library");
1076 rte_errno = ENOTSUP;
1081 * Required to initialize the CUDA Driver.
1082 * Multiple calls of cuInit() will return immediately
1083 * without making any relevant change
1087 res = sym_cuDriverGetVersion(&cuda_driver_version);
1089 rte_cuda_log(ERR, "cuDriverGetVersion failed with %d", res);
1090 rte_errno = ENOTSUP;
1094 if (cuda_driver_version < CUDA_DRIVER_MIN_VERSION) {
1095 rte_cuda_log(ERR, "CUDA Driver version found is %d. "
1096 "Minimum requirement is %d",
1097 cuda_driver_version,
1098 CUDA_DRIVER_MIN_VERSION);
1099 rte_errno = ENOTSUP;
1103 if (cuda_pfn_func_loader()) {
1104 rte_cuda_log(ERR, "CUDA PFN functions not found in library");
1105 rte_errno = ENOTSUP;
1112 /* Fill HW specific part of device structure */
1113 dev->device = &pci_dev->device;
1114 dev->mpshared->info.numa_node = pci_dev->device.numa_node;
1116 /* Get NVIDIA GPU Device descriptor */
1117 res = pfn_cuDeviceGetByPCIBusId(&cu_dev_id, dev->device->name);
1119 pfn_cuGetErrorString(res, &(err_string));
1120 rte_cuda_log(ERR, "cuDeviceGetByPCIBusId name %s failed with %d: %s",
1121 dev->device->name, res, err_string);
1126 res = pfn_cuDevicePrimaryCtxRetain(&pctx, cu_dev_id);
1128 pfn_cuGetErrorString(res, &(err_string));
1129 rte_cuda_log(ERR, "cuDevicePrimaryCtxRetain name %s failed with %d: %s",
1130 dev->device->name, res, err_string);
1135 res = pfn_cuCtxGetApiVersion(pctx, &cuda_api_version);
1137 rte_cuda_log(ERR, "cuCtxGetApiVersion failed with %d", res);
1138 rte_errno = ENOTSUP;
1142 if (cuda_api_version < CUDA_API_MIN_VERSION) {
1143 rte_cuda_log(ERR, "CUDA API version found is %d Minimum requirement is %d",
1144 cuda_api_version, CUDA_API_MIN_VERSION);
1145 rte_errno = ENOTSUP;
1149 dev->mpshared->info.context = (uint64_t)pctx;
1152 * GPU Device generic info
1155 /* Processor count */
1156 res = pfn_cuDeviceGetAttribute(&(processor_count),
1157 CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
1160 pfn_cuGetErrorString(res, &(err_string));
1161 rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
1166 dev->mpshared->info.processor_count = (uint32_t)processor_count;
1169 res = pfn_cuDeviceTotalMem(&dev->mpshared->info.total_memory, cu_dev_id);
1171 pfn_cuGetErrorString(res, &(err_string));
1172 rte_cuda_log(ERR, "cuDeviceTotalMem failed with %s",
1178 dev->mpshared->info.page_size = (size_t)GPU_PAGE_SIZE;
1181 * GPU Device private info
1183 dev->mpshared->dev_private = rte_zmalloc(NULL,
1184 sizeof(struct cuda_info),
1185 RTE_CACHE_LINE_SIZE);
1186 if (dev->mpshared->dev_private == NULL) {
1187 rte_cuda_log(ERR, "Failed to allocate memory for GPU process private");
1192 private = (struct cuda_info *)dev->mpshared->dev_private;
1193 private->cu_dev = cu_dev_id;
1194 res = pfn_cuDeviceGetName(private->gpu_name,
1195 RTE_DEV_NAME_MAX_LEN,
1198 pfn_cuGetErrorString(res, &(err_string));
1199 rte_cuda_log(ERR, "cuDeviceGetName failed with %s",
1205 res = pfn_cuDeviceGetAttribute(&(private->gdr_supported),
1206 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED,
1209 pfn_cuGetErrorString(res, &(err_string));
1210 rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
1216 if (private->gdr_supported == 0)
1217 rte_cuda_log(WARNING, "GPU %s doesn't support GPUDirect RDMA",
1218 pci_dev->device.name);
1220 res = pfn_cuDeviceGetAttribute(&(private->gdr_write_ordering),
1221 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING,
1224 pfn_cuGetErrorString(res, &(err_string));
1226 "cuDeviceGetAttribute failed with %s",
1232 if (private->gdr_write_ordering == CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE) {
1233 res = pfn_cuDeviceGetAttribute(&(private->gdr_flush_type),
1234 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS,
1237 pfn_cuGetErrorString(res, &(err_string));
1238 rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
1244 if (private->gdr_flush_type != CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST)
1245 rte_cuda_log(ERR, "GPUDirect RDMA flush writes API is not supported");
1248 dev->ops.dev_info_get = cuda_dev_info_get;
1249 dev->ops.dev_close = cuda_dev_close;
1250 dev->ops.mem_alloc = cuda_mem_alloc;
1251 dev->ops.mem_free = cuda_mem_free;
1252 dev->ops.mem_register = cuda_mem_register;
1253 dev->ops.mem_unregister = cuda_mem_unregister;
1254 dev->ops.mem_cpu_map = cuda_mem_cpu_map;
1255 dev->ops.mem_cpu_unmap = cuda_mem_cpu_unmap;
1256 dev->ops.wmb = cuda_wmb;
1258 rte_gpu_complete_new(dev);
1260 rte_cuda_debug("dev id = %u name = %s",
1261 dev->mpshared->info.dev_id, private->gpu_name);
1267 cuda_gpu_remove(struct rte_pci_device *pci_dev)
1269 struct rte_gpu *dev;
1273 if (pci_dev == NULL) {
1278 dev = rte_gpu_get_by_name(pci_dev->device.name);
1280 rte_cuda_log(ERR, "Couldn't find HW dev \"%s\" to uninitialise it",
1281 pci_dev->device.name);
1285 gpu_id = dev->mpshared->info.dev_id;
1287 /* release dev from library */
1288 ret = rte_gpu_release(dev);
1290 rte_cuda_log(ERR, "Device %i failed to uninit: %i", gpu_id, ret);
1292 rte_cuda_debug("Destroyed dev = %u", gpu_id);
1297 static struct rte_pci_driver rte_cuda_driver = {
1298 .id_table = pci_id_cuda_map,
1299 .drv_flags = RTE_PCI_DRV_WC_ACTIVATE,
1300 .probe = cuda_gpu_probe,
1301 .remove = cuda_gpu_remove,
1304 RTE_PMD_REGISTER_PCI(gpu_cuda, rte_cuda_driver);
1305 RTE_PMD_REGISTER_PCI_TABLE(gpu_cuda, pci_id_cuda_map);
1306 RTE_PMD_REGISTER_KMOD_DEP(gpu_cuda, "* nvidia & (nv_peer_mem | nvpeer_mem)");