[PATCH v3 00/30] Driver for Hyper-v virtual compute device

From: Iouri Tarassov
Date: Tue Mar 01 2022 - 14:46:35 EST


The driver code in the new patch series is split to more simple patches
for easy review. The definitions are added to the header files as
they become to be used.

The patches introduces a driver for the Hyper-V virtual compute hardware
devices. This driver powers the Windows Subsystem for Linux (WSL) and will
soon power the Windows Subsystem for Android (WSA).

Per earlier feedback, the driver is now located under the Hyper-V node as it
is not a classic Linux GPU (KMS/DRM) driver and really only make sense
when running under a Windows host under Hyper-V. Although we refer to this
driver as VGPU for shorthand, in reality this is a generic virtualization
infrastructure for various class of compute accelerators, the most popular
and ubiquitous being the GPU. We support virtualization of non-GPU devices
through this infrastructure. These devices are exposed to user-space and
used by various API and framework, such as CUDA, OpenCL, OpenVINO,
OneAPI, DX12, etc...

One critical piece of feedback, provided by the community in our earlier
submission, was the lack of an open source user-space to go along with our
kernel compute driver. At the time we only had CUDA and DX12 user-space APIs,
both of which being closed source. This is a divisive issue in the community
and is heavily debated (https://lwn.net/Articles/821817/).

We took this feedback to heart and we spent the last year working on a way
to address this key piece of feedback. Working closely with our partner Intel,
we're happy to say that we now have a fully open source user-space for the
OpenCL, OpenVINO and OneAPI compute family of APIs on Intel GPU platforms.
This is supported by this open source project
(https://github.com/intel/compute-runtime).

To make it easy for our partners to build compute drivers, which are usable
in both Windows and WSL environment, we provide a library, that provides a
stable interface to our compute device abstraction. This was originally part
of the libdxcore closed source API, but have now spawn that off into it's
own open source project (https://github.com/microsoft/libdxg).

Between the Intel compute runtime project and libdxg, we now have a fully
open source implementation of our virtualized compute stack inside of WSL.
We will continue to support both open source user-space API against our
compute abstraction as well as closed source one (CUDA, DX12), leaving
it to the API owners and partners to decide what makes the most sense for them.

A lot of efforts went into addressing community feedback in this revised
set of patches and we hope this is getting closer to what the community
would like to see.

We're looking forward additional feedback.

Driver overview
-------------------------------------------------------
dxgkrnl is a driver for Hyper-V virtual compute devices, such as VGPU
devices, which are projected to a Linux virtual machine (VM) by a Windows
host. dxgkrnl works in context of WDDM (Windows Display Driver Model)
for GPU or MCDM (Microsoft Compute Driver Model) for non-GPU devices.
WDDM/MCDM consists of the following components:
- Graphics or Compute applications
- A graphics or compute user mode API (for example OpenGL, Vulkan, OpenCL,
OpenVINO, OneAPI, CUDA, DX12, ...)
- User Mode Driver (UMD), written by a hardware vendor
- optional libdxg library helping UMD portability across Windows and Linux
- dxgkrnl Linux kernel driver (this driver)
- Kernel mode port driver on the Windows host (dxgkrnl.sys / dxgmms2.sys)
- Kernel mode miniport driver (KMD) on the Windows host, written by a
hardware vendor running on the Windows host and interfacing with the
hardware device.

dxgkrnl exposes a subset the WDDM/MCDM D3DKMT interface to user space. See
https://docs.microsoft.com/en-us/windows-hardware/drivers/ddi/d3dkmthk/
This interface provides user space with an abstracted view and control
of compute devices in a portable way across Windows and WSL. It is used for
devices such as GPU or AI/ML processors. This interface is mapped to custom
IOCTLs on Linux (see d3dkmthk.h). The libdxg library translates the D3DKMT
interface calls to the driver IOCTLs.

A more detailed overview of this architecture is available here:
https://devblogs.microsoft.com/directx/directx-heart-linux/

Virtual compute devices are paravirtualized, meaning that the actual
communication with the corresponding device hardware happens on the host.
The version of dxgkrnl inside of the Linux kernel coordinates with the
version of dxgkrnl running on Windows to provide a consistent and
portable abstraction for the device that the various APIs and UMD can
rely on across Windows and Linux.

Dxgkrnl creates the /dev/dxg device, which can be used to enumerate virtual
compute devices (also called adapters). UMD or an API runtime open the
device and send ioctls to create various device objects (dxgdevice,
dxgcontext, dxgallocation, etc., defined in dxgkrnl.h) and to submit work
to the device. The WDDM objects are represented in user mode as opaque
handles (struct d3dkmthandle). Dxgkrnl creates a dxgprocess object
for each process, which opens the /dev/dxg device. This object has a handle
table, which is used to translate d3dkmt handles to kernel objects. Handle
table definitions are in hmgr.h. There is also a global handle table for
objects, which do not belong to a particular process.

Driver initialization
-------------------------------------------------------
When dxgkrnl is loaded, dxgkrnl registers for virtual PCI device arrival
notifications and VM bus channel device notifications (see dxgmodule.c). When
the first virtual device is started, dxgkrnl creates a misc device (/dev/dxg).
A user mode client can open the /dev/dxg device and send IOCTLs to enumerate
and control virtual compute devices.

Virtual compute device initialization
-------------------------------------------------------
A virtual device is represented by a dxgadapter object. It is created when
the corresponding device arrives on the virtual PCI bus. The device vendor
is PCI_VENDOR_ID_MICROSOFT and the device id is PCI_DEVICE_ID_VIRTUAL_RENDER.
The adapter is started when the corresponding VM bus channel and the global
VM bus channel are initialized. Dynamic arrival/removal of devices is
supported.

Internal objects
-------------------------------------------------------
dxgkrnl creates various internal objects in response to IOCTL calls. The
corresponsing objects are also created on the host. Each object is placed to
a process or global handle table. Object handles (d3dkmthandle) are returned
to user mode. The object handles are also used to reference objects on the host.
Corresponding objects in the guest and the host have the same handle value to
avoid handle translation. The internal objects are:
- dxgadapter
Represents a virtual compute device object. It is created for every device
projected by the host to the VM.
- dxgprocess
The object is created for each Linux process, which opens /dxg/dev. It has the
object handle table, which holds pointers to all internal objects, which are
created by this process.
- dxgcontext
Represents a device execution thread in the packet scheduling mode (as oppose
to the hardware scheduling mode).
- dxghwqueue
Represents a device execution thread in the hardware scheduling mode.
- dxgdevice
A collection of device contexts, allocations, resources, sync objects, etc.
- dxgallocation
Represents a device accessible memory allocation.
- dxgresource
A collection of dxgallocation objects. This object could be shared between
devices and processes.
- dxgsharedresource
Represents a dxgresource object, which is sharable between processes.
- dxgsyncobject
Represents a device synchronization object, which is used to synchronize
execution of DMA buffers in the device execution contexts.
- dxgsharedsyncobject
Represent a device synchronization object, which is sharable between processes.
- dxgpagingqueue
Represents a queue, which is used to manage residency of the device allocation
objects.

Communications with the host
-------------------------------------------------------
Dxgkrnl communicates with the host via Hyper-V VM bus channels. There
is a global channel and a per virtual computed device channel. The VM bus
messages are defined in dxgvmbus.h and the implementation is in dxgvmbus.c.

Most VM bus messages to the host are synchronous. When the host enables the
asynchronous mode, some high frequency VM bus messages are sent asynchronously
to improve performance. When async messages are enabled, all VM bus messages are
sent only via the global channel to maintain the order of messages on the host.

The host could send asynchronous messages to the Linux dxgkrnl driver via
the global VM bus channel. The host messages are handled by
dxgvmbuschannel_receive().

PCI config space of the device is used to exchange information between the host
and the guest during dxgkrnl initialization. Look at dxg_pci_probe_device().

CPU access to device accessible allocations
-------------------------------------------------------
D3DKMT API allows creation of allocations, which are accessible by the device
and the CPU. The global VM bus channels has associated IO space, which is used
to implement CPU access to CPU visible allocations. For each such allocation
the host allocates a portion of the guest IO space and maps it to the
allocation memory (it could be in system memory or in device local memory).
A user mode application calls the LX_DXLOCK2 ioctl to get the allocation
CPU address. Dxgkrnl uses vm_mmap to allocate a user space virtual address
range and maps it to the allocation IO space using io_remap_pfn_range().
This way Linux user mode virtual addresses point to the host system memory
or device local memory.

Sharing objects between processes
-------------------------------------------------------
Some dxgkrnl objects could be shared between processes. This includes resources
(dxgresource) and synchronization objects (dxgsyncobject).
The WDDM API provides a way to share objects using so called "NT handles".
"NT handle" on Windows is a native Windows process handle (HANDLE). "NT handles"
are implemented as file descriptors (FD) on Linux.
The LX_DXSHAREOBJECTS ioctl is used to get an FD for a shared object.
Before a shared object can be used, it needs to be "opened" to get the "local"
d3dkmthandle handle. The LX_DXOPENRESOURCEFROMNTHANDLE and
LX_DXOPENSYNCOBJECTFROMNTHANDLE2 ioctls are used to open a shared object

Iouri Tarassov (30):
drivers: hv: dxgkrnl: Add virtual compute device VM bus
channel guids
drivers: hv: dxgkrnl: Driver initialization and loading
drivers: hv: dxgkrnl: Add VM bus message support, initialize VM bus
channels.
drivers: hv: dxgkrnl: Creation of dxgadapter object
drivers: hv: dxgkrnl: Opening of /dev/dxg device and dxgprocess
creation
drivers: hv: dxgkrnl: Enumerate and open dxgadapter objects
drivers: hv: dxgkrnl: Creation of dxgdevice objects
drivers: hv: dxgkrnl: Creation of dxgcontext objects
drivers: hv: dxgkrnl: Creation of compute device allocations and
resources
drivers: hv: dxgkrnl: Creation of compute device sync objects
drivers: hv: dxgkrnl: Operations using sync objects
drivers: hv: dxgkrnl: Sharing of dxgresource objects
drivers: hv: dxgkrnl: Sharing of sync objects
drivers: hv: dxgkrnl: Creation of hardware queues. Sync object
operations to hw queue.
drivers: hv: dxgkrnl: Creation of paging queue objects.
drivers: hv: dxgkrnl: Submit execution commands to the compute device
drivers: hv: dxgkrnl: Share objects with the host
drivers: hv: dxgkrnl: Query the dxgdevice state
drivers: hv: dxgkrnl: Map(unmap) CPU address to device allocation
drivers: hv: dxgkrnl: Manage device allocation properties
drivers: hv: dxgkrnl: Flush heap transitions
drivers: hv: dxgkrnl: Query video memory information
drivers: hv: dxgkrnl: The escape ioctl
drivers: hv: dxgkrnl: Ioctl to put device to error state
drivers: hv: dxgkrnl: Ioctls to query statistics and clock calibration
drivers: hv: dxgkrnl: Offer and reclaim allocations
drivers: hv: dxgkrnl: Ioctls to manage scheduling priority
drivers: hv: dxgkrnl: Manage residency of allocations
drivers: hv: dxgkrnl: Manage compute device virtual addresses
drivers: hv: dxgkrnl: Add support to map guest pages by host

MAINTAINERS | 7 +
drivers/hv/Kconfig | 2 +
drivers/hv/Makefile | 1 +
drivers/hv/dxgkrnl/Kconfig | 26 +
drivers/hv/dxgkrnl/Makefile | 5 +
drivers/hv/dxgkrnl/dxgadapter.c | 1375 ++++++++
drivers/hv/dxgkrnl/dxgkrnl.h | 962 ++++++
drivers/hv/dxgkrnl/dxgmodule.c | 942 ++++++
drivers/hv/dxgkrnl/dxgprocess.c | 334 ++
drivers/hv/dxgkrnl/dxgvmbus.c | 3788 ++++++++++++++++++++++
drivers/hv/dxgkrnl/dxgvmbus.h | 868 +++++
drivers/hv/dxgkrnl/hmgr.c | 566 ++++
drivers/hv/dxgkrnl/hmgr.h | 112 +
drivers/hv/dxgkrnl/ioctl.c | 5368 +++++++++++++++++++++++++++++++
drivers/hv/dxgkrnl/misc.c | 43 +
drivers/hv/dxgkrnl/misc.h | 96 +
include/linux/hyperv.h | 16 +
include/uapi/misc/d3dkmthk.h | 1679 ++++++++++
18 files changed, 16190 insertions(+)
create mode 100644 drivers/hv/dxgkrnl/Kconfig
create mode 100644 drivers/hv/dxgkrnl/Makefile
create mode 100644 drivers/hv/dxgkrnl/dxgadapter.c
create mode 100644 drivers/hv/dxgkrnl/dxgkrnl.h
create mode 100644 drivers/hv/dxgkrnl/dxgmodule.c
create mode 100644 drivers/hv/dxgkrnl/dxgprocess.c
create mode 100644 drivers/hv/dxgkrnl/dxgvmbus.c
create mode 100644 drivers/hv/dxgkrnl/dxgvmbus.h
create mode 100644 drivers/hv/dxgkrnl/hmgr.c
create mode 100644 drivers/hv/dxgkrnl/hmgr.h
create mode 100644 drivers/hv/dxgkrnl/ioctl.c
create mode 100644 drivers/hv/dxgkrnl/misc.c
create mode 100644 drivers/hv/dxgkrnl/misc.h
create mode 100644 include/uapi/misc/d3dkmthk.h

--
2.35.1