[PATCH 1/2] drm/amdgpu: Generalize KFD dmabuf import

[PATCH 1/2] drm/amdgpu: Generalize KFD dmabuf import

[Felix Kuehling]

Use proper amdgpu_gem_prime_import function to handle all kinds of
imports. Remember the dmabuf reference to enable proper multi-GPU
attachment to multiple VMs without erroneously re-exporting the
underlying BO multiple times.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
---
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 45 ++++++++++++-------
 1 file changed, 28 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 90b985436878..d53d19b9d6dc 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -2000,30 +2000,34 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
 	struct amdgpu_bo *bo;
 	int ret;
 
-	if (dma_buf->ops != &amdgpu_dmabuf_ops)
-		/* Can't handle non-graphics buffers */
-		return -EINVAL;
-
-	obj = dma_buf->priv;
-	if (drm_to_adev(obj->dev) != adev)
-		/* Can't handle buffers from other devices */
-		return -EINVAL;
+	obj = amdgpu_gem_prime_import(adev_to_drm(adev), dma_buf);
+	if (IS_ERR(obj))
+		return PTR_ERR(obj);
+	if (obj->import_attach) {
+		/* Import takes an extra reference if it creates a dynamic
+		 * dmabuf attachment. Release that reference to avoid
+		 * leaking it. FIXME: I think this is a bug.
+		 */
+		dma_buf_put(dma_buf);
+	}
 
 	bo = gem_to_amdgpu_bo(obj);
 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
-				    AMDGPU_GEM_DOMAIN_GTT)))
+				    AMDGPU_GEM_DOMAIN_GTT))) {
 		/* Only VRAM and GTT BOs are supported */
-		return -EINVAL;
+		ret = -EINVAL;
+		goto err_put_obj;
+	}
 
 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
-	if (!*mem)
-		return -ENOMEM;
+	if (!*mem) {
+		ret = -ENOMEM;
+		goto err_put_obj;
+	}
 
 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
-	if (ret) {
-		kfree(mem);
-		return ret;
-	}
+	if (ret)
+		goto err_free_mem;
 
 	if (size)
 		*size = amdgpu_bo_size(bo);
@@ -2040,7 +2044,8 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
 
-	drm_gem_object_get(&bo->tbo.base);
+	get_dma_buf(dma_buf);
+	(*mem)->dmabuf = dma_buf;
 	(*mem)->bo = bo;
 	(*mem)->va = va;
 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
@@ -2052,6 +2057,12 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
 	(*mem)->is_imported = true;
 
 	return 0;
+
+err_free_mem:
+	kfree(mem);
+err_put_obj:
+	drm_gem_object_put(obj);
+	return ret;
 }
 
 /* Evict a userptr BO by stopping the queues if necessary
-- 
2.32.0

[PATCH 2/2] drm/amdkfd: Implement DMA buf fd export for RDMA

[Felix Kuehling]

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(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index fcbc8a9c9e06..840de82460a3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -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);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index d53d19b9d6dc..9f57e5091fa8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -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
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 4bfc0c8ab764..ddbc28951ac1 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -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)
diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index af96af174dc4..3b80af5f5bdf 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -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
-- 
2.32.0