@@ -294,6 +294,8 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
uint64_t va, void *drm_priv,
struct kgd_mem **mem, uint64_t *size,
uint64_t *mmap_offset);
+int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
+ struct dma_buf **dmabuf);
int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
struct tile_config *config);
void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev);
@@ -641,6 +641,21 @@ kfd_mem_attach_userptr(struct amdgpu_device *adev, struct kgd_mem *mem,
return 0;
}
+static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
+{
+ if (!mem->dmabuf) {
+ struct dma_buf *ret = amdgpu_gem_prime_export(
+ &mem->bo->tbo.base,
+ mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
+ DRM_RDWR : 0);
+ if (IS_ERR(ret))
+ return PTR_ERR(ret);
+ mem->dmabuf = ret;
+ }
+
+ return 0;
+}
+
static int
kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
struct amdgpu_bo **bo)
@@ -648,16 +663,9 @@ kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
struct drm_gem_object *gobj;
int ret;
- if (!mem->dmabuf) {
- mem->dmabuf = amdgpu_gem_prime_export(&mem->bo->tbo.base,
- mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
- DRM_RDWR : 0);
- if (IS_ERR(mem->dmabuf)) {
- ret = PTR_ERR(mem->dmabuf);
- mem->dmabuf = NULL;
- return ret;
- }
- }
+ ret = kfd_mem_export_dmabuf(mem);
+ if (ret)
+ return ret;
gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
if (IS_ERR(gobj))
@@ -2065,6 +2073,23 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
return ret;
}
+int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
+ struct dma_buf **dma_buf)
+{
+ int ret;
+
+ mutex_lock(&mem->lock);
+ ret = kfd_mem_export_dmabuf(mem);
+ if (ret)
+ goto out;
+
+ get_dma_buf(mem->dmabuf);
+ *dma_buf = mem->dmabuf;
+out:
+ mutex_unlock(&mem->lock);
+ return ret;
+}
+
/* Evict a userptr BO by stopping the queues if necessary
*
* Runs in MMU notifier, may be in RECLAIM_FS context. This means it
@@ -1787,6 +1787,58 @@ static int kfd_ioctl_import_dmabuf(struct file *filep,
return r;
}
+static int kfd_ioctl_export_dmabuf(struct file *filep,
+ struct kfd_process *p, void *data)
+{
+ struct kfd_ioctl_export_dmabuf_args *args = data;
+ struct kfd_process_device *pdd;
+ struct dma_buf *dmabuf;
+ struct kfd_dev *dev;
+ void *mem;
+ int ret = 0;
+
+ dev = kfd_device_by_id(GET_GPU_ID(args->handle));
+ if (!dev)
+ return -EINVAL;
+
+ mutex_lock(&p->mutex);
+
+ pdd = kfd_get_process_device_data(dev, p);
+ if (!pdd) {
+ ret = -EINVAL;
+ goto err_unlock;
+ }
+
+ mem = kfd_process_device_translate_handle(pdd,
+ GET_IDR_HANDLE(args->handle));
+ if (!mem) {
+ ret = -EINVAL;
+ goto err_unlock;
+ }
+
+ ret = amdgpu_amdkfd_gpuvm_export_dmabuf(mem, &dmabuf);
+ mutex_unlock(&p->mutex);
+ if (ret)
+ goto err_out;
+
+ ret = dma_buf_fd(dmabuf, args->flags);
+ if (ret < 0) {
+ dma_buf_put(dmabuf);
+ goto err_out;
+ }
+ /* dma_buf_fd assigns the reference count to the fd, no need to
+ * put the reference here.
+ */
+ args->dmabuf_fd = ret;
+
+ return 0;
+
+err_unlock:
+ mutex_unlock(&p->mutex);
+err_out:
+ return ret;
+}
+
/* Handle requests for watching SMI events */
static int kfd_ioctl_smi_events(struct file *filep,
struct kfd_process *p, void *data)
@@ -1959,6 +2011,9 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = {
AMDKFD_IOCTL_DEF(AMDKFD_IOC_SET_XNACK_MODE,
kfd_ioctl_set_xnack_mode, 0),
+
+ AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF,
+ kfd_ioctl_export_dmabuf, 0),
};
#define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls)
@@ -32,9 +32,10 @@
* - 1.4 - Indicate new SRAM EDC bit in device properties
* - 1.5 - Add SVM API
* - 1.6 - Query clear flags in SVM get_attr API
+ * - 1.7 - Add DMA buf export ioctl
*/
#define KFD_IOCTL_MAJOR_VERSION 1
-#define KFD_IOCTL_MINOR_VERSION 6
+#define KFD_IOCTL_MINOR_VERSION 7
struct kfd_ioctl_get_version_args {
__u32 major_version; /* from KFD */
@@ -450,6 +451,12 @@ struct kfd_ioctl_import_dmabuf_args {
__u32 dmabuf_fd; /* to KFD */
};
+struct kfd_ioctl_export_dmabuf_args {
+ __u64 handle; /* to KFD */
+ __u32 flags; /* to KFD */
+ __u32 dmabuf_fd; /* from KFD */
+};
+
/*
* KFD SMI(System Management Interface) events
*/
@@ -742,7 +749,10 @@ struct kfd_ioctl_set_xnack_mode_args {
#define AMDKFD_IOC_SET_XNACK_MODE \
AMDKFD_IOWR(0x21, struct kfd_ioctl_set_xnack_mode_args)
+#define AMDKFD_IOC_EXPORT_DMABUF \
+ AMDKFD_IOWR(0x22, struct kfd_ioctl_export_dmabuf_args)
+
#define AMDKFD_COMMAND_START 0x01
-#define AMDKFD_COMMAND_END 0x22
+#define AMDKFD_COMMAND_END 0x23
#endif
Exports a DMA buf fd of a given KFD buffer handle. This is intended for the new upstreamable RDMA solution coming to UCX and libfabric. The corresponding user mode change (Thunk API and kfdtest) is here: https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/commits/fxkamd/dmabuf Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com> --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 45 +++++++++++---- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 55 +++++++++++++++++++ include/uapi/linux/kfd_ioctl.h | 14 ++++- 4 files changed, 104 insertions(+), 12 deletions(-)