Commit 1dde0ea9 authored by Felix Kuehling's avatar Felix Kuehling Committed by Alex Deucher
Browse files

drm/amdkfd: Add DMABuf import functionality 



This is used for interoperability between ROCm compute and graphics
APIs. It allows importing graphics driver BOs into the ROCm SVM
address space for zero-copy GPU access.

The API is split into two steps (query and import) to allow user mode
to manage the virtual address space allocation for the imported buffer.

Acked-by: default avatarAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: default avatarFelix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: default avatarAlex Deucher <alexander.deucher@amd.com>
parent 3704d56e

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c

+57 −0
Original line number Diff line number Diff line
@@ -26,6 +26,7 @@ 
#include "amdgpu.h"

#include "amdgpu_gfx.h"

#include <linux/module.h>

#include <linux/dma-buf.h>



const struct kgd2kfd_calls *kgd2kfd;
@@ -433,6 +434,62 @@ void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info) 
	cu_info->lds_size = acu_info.lds_size;

}



int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,

				  struct kgd_dev **dma_buf_kgd,

				  uint64_t *bo_size, void *metadata_buffer,

				  size_t buffer_size, uint32_t *metadata_size,

				  uint32_t *flags)

{

	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;

	struct dma_buf *dma_buf;

	struct drm_gem_object *obj;

	struct amdgpu_bo *bo;

	uint64_t metadata_flags;

	int r = -EINVAL;



	dma_buf = dma_buf_get(dma_buf_fd);

	if (IS_ERR(dma_buf))

		return PTR_ERR(dma_buf);



	if (dma_buf->ops != &amdgpu_dmabuf_ops)

		/* Can't handle non-graphics buffers */

		goto out_put;



	obj = dma_buf->priv;

	if (obj->dev->driver != adev->ddev->driver)

		/* Can't handle buffers from different drivers */

		goto out_put;



	adev = obj->dev->dev_private;

	bo = gem_to_amdgpu_bo(obj);

	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |

				    AMDGPU_GEM_DOMAIN_GTT)))

		/* Only VRAM and GTT BOs are supported */

		goto out_put;



	r = 0;

	if (dma_buf_kgd)

		*dma_buf_kgd = (struct kgd_dev *)adev;

	if (bo_size)

		*bo_size = amdgpu_bo_size(bo);

	if (metadata_size)

		*metadata_size = bo->metadata_size;

	if (metadata_buffer)

		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,

					   metadata_size, &metadata_flags);

	if (flags) {

		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?

			ALLOC_MEM_FLAGS_VRAM : ALLOC_MEM_FLAGS_GTT;



		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)

			*flags |= ALLOC_MEM_FLAGS_PUBLIC;

	}



out_put:

	dma_buf_put(dma_buf);

	return r;

}



uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd)

{

	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h

+11 −0
Original line number Diff line number Diff line
@@ -149,6 +149,11 @@ uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd); 


uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd);

void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info);

int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,

				  struct kgd_dev **dmabuf_kgd,

				  uint64_t *bo_size, void *metadata_buffer,

				  size_t buffer_size, uint32_t *metadata_size,

				  uint32_t *flags);

uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd);

uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd);
@@ -200,6 +205,12 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *process_info, 
int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd,

					      struct kfd_vm_fault_info *info);



int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,

				      struct dma_buf *dmabuf,

				      uint64_t va, void *vm,

				      struct kgd_mem **mem, uint64_t *size,

				      uint64_t *mmap_offset);



void amdgpu_amdkfd_gpuvm_init_mem_limits(void);

void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo);

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c

+55 −0
Original line number Diff line number Diff line
@@ -25,6 +25,7 @@ 
#include <linux/list.h>

#include <linux/pagemap.h>

#include <linux/sched/mm.h>

#include <linux/dma-buf.h>

#include <drm/drmP.h>

#include "amdgpu_object.h"

#include "amdgpu_vm.h"
@@ -1664,6 +1665,60 @@ int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct kgd_dev *kgd, 
	return 0;

}



int amdgpu_amdkfd_gpuvm_import_dmabuf(struct kgd_dev *kgd,

				      struct dma_buf *dma_buf,

				      uint64_t va, void *vm,

				      struct kgd_mem **mem, uint64_t *size,

				      uint64_t *mmap_offset)

{

	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;

	struct drm_gem_object *obj;

	struct amdgpu_bo *bo;

	struct amdgpu_vm *avm = (struct amdgpu_vm *)vm;



	if (dma_buf->ops != &amdgpu_dmabuf_ops)

		/* Can't handle non-graphics buffers */

		return -EINVAL;



	obj = dma_buf->priv;

	if (obj->dev->dev_private != adev)

		/* Can't handle buffers from other devices */

		return -EINVAL;



	bo = gem_to_amdgpu_bo(obj);

	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |

				    AMDGPU_GEM_DOMAIN_GTT)))

		/* Only VRAM and GTT BOs are supported */

		return -EINVAL;



	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);

	if (!*mem)

		return -ENOMEM;



	if (size)

		*size = amdgpu_bo_size(bo);



	if (mmap_offset)

		*mmap_offset = amdgpu_bo_mmap_offset(bo);



	INIT_LIST_HEAD(&(*mem)->bo_va_list);

	mutex_init(&(*mem)->lock);

	(*mem)->mapping_flags =

		AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE |

		AMDGPU_VM_PAGE_EXECUTABLE | AMDGPU_VM_MTYPE_NC;



	(*mem)->bo = amdgpu_bo_ref(bo);

	(*mem)->va = va;

	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?

		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;

	(*mem)->mapped_to_gpu_memory = 0;

	(*mem)->process_info = avm->process_info;

	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);

	amdgpu_sync_create(&(*mem)->sync);



	return 0;

}



/* Evict a userptr BO by stopping the queues if necessary

 *

 * Runs in MMU notifier, may be in RECLAIM_FS context. This means it

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h

+2 −0
Original line number Diff line number Diff line
@@ -54,6 +54,8 @@ void *amdgpu_gem_prime_vmap(struct drm_gem_object *obj); 
void amdgpu_gem_prime_vunmap(struct drm_gem_object *obj, void *vaddr);

int amdgpu_gem_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);



extern const struct dma_buf_ops amdgpu_dmabuf_ops;



/*

 * GEM objects.

 */

drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c

+1 −3
Original line number Diff line number Diff line
@@ -39,8 +39,6 @@ 
#include <drm/amdgpu_drm.h>

#include <linux/dma-buf.h>



static const struct dma_buf_ops amdgpu_dmabuf_ops;



/**

 * amdgpu_gem_prime_get_sg_table - &drm_driver.gem_prime_get_sg_table

 * implementation
@@ -332,7 +330,7 @@ static int amdgpu_gem_begin_cpu_access(struct dma_buf *dma_buf, 
	return ret;

}



static const struct dma_buf_ops amdgpu_dmabuf_ops = {

const struct dma_buf_ops amdgpu_dmabuf_ops = {

	.attach = amdgpu_gem_map_attach,

	.detach = amdgpu_gem_map_detach,

	.map_dma_buf = drm_gem_map_dma_buf,

CRA Git | Maintained and supported by SUSTech CRA and CCSE