The Virtual Dmabuf or Virtio based Dmabuf (Vdmabuf) driver can be used to "transfer" a page-backed dmabuf created in the Guest to the Host without making any copies. This is mostly accomplished by recreating the dmabuf on the Host using the PFNs and other meta-data shared by the guest. A use-case where this driver would be a good fit is a multi-GPU system (perhaps one discrete and one integrated) where one of the GPUs does not have access to the display/connectors/outputs. This could be an embedded system design decision or a restriction made at the firmware/BIOS level or perhaps the device is setup in UPT (Universal Passthrough) mode. When such a GPU is passthrough'd to a Guest OS, this driver can help in transferring the scanout buffer(s) (rendered using the native rendering stack) to the Host for the purpose of displaying them. Or, quite simply, this driver can be used to transfer a dmabuf created by an application running on the Guest to another application running on the Host.
The userspace component running in the Guest that transfers the dmabuf is referred to as the producer or exporter and its counterpart running in the Host is referred to as importer or consumer. For instance, a Wayland compositor would potentially be a producer and Qemu UI would be a consumer. It is the producer's responsibility to not reuse or destroy the shared buffer while it is still being used by the consumer. The consumer would send a release cmd indicating that it is done after which the shared buffer can be safely used again by the producer. One way the producer can prevent accidental re-use of the shared buffer is to lock the buffer when it exports it and unlock it after it gets a release cmd. As an example, the GBM API provides a simple way to lock and unlock a surface's buffers.
For each dmabuf that is to be shared with the Host, a 128-bit unique ID is generated that identifies this buffer across the whole system. This ID is a combination of the Qemu process ID, a counter and a randomizer. We could potentially use UUID API but we currently use the above mentioned combination to identify the source of the buffer at any given time for potential bookkeeping.
A typical cycle starts with the producer or exporter calling the export IOCTL to export a dmabuf; a new unique ID is generated for this buffer and it gets registered with the Host. The Host then alerts the consumer or importer by raising an event and shares the ID. In response, the consumer calls the import IOCTL using the ID and gets a newly created dmabuf fd in return. After it is done using the dmabuf, the consumer finally calls the release IOCTL and the Guest is notified which in turn notifies the producer letting it know that the buffer is now safe to reuse.
v2: - Added a notifier mechanism for getting the kvm pointer. - Added start and stop routines in the Vhost backend. - Augmented the cover letter and made some minor improvements.
v3: - Refactored the code to make it similar to vsock
Vivek Kasireddy (3): kvm: Add a notifier for create and destroy VM events virtio: Introduce Vdmabuf driver vhost: Add Vdmabuf backend
drivers/vhost/Kconfig | 9 + drivers/vhost/Makefile | 3 + drivers/vhost/vdmabuf.c | 1446 +++++++++++++++++++++++++++ drivers/virtio/Kconfig | 8 + drivers/virtio/Makefile | 1 + drivers/virtio/virtio_vdmabuf.c | 1090 ++++++++++++++++++++ include/linux/kvm_host.h | 5 + include/linux/virtio_vdmabuf.h | 271 +++++ include/uapi/linux/vhost.h | 3 + include/uapi/linux/virtio_ids.h | 1 + include/uapi/linux/virtio_vdmabuf.h | 99 ++ virt/kvm/kvm_main.c | 20 +- 12 files changed, 2954 insertions(+), 2 deletions(-) create mode 100644 drivers/vhost/vdmabuf.c create mode 100644 drivers/virtio/virtio_vdmabuf.c create mode 100644 include/linux/virtio_vdmabuf.h create mode 100644 include/uapi/linux/virtio_vdmabuf.h