Re: [PATCH 1/1] Documentation: hyperv: Add overview of PCI pass-thru device support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 2/14/2024 3:22 PM, mhkelley58@xxxxxxxxx wrote:
> From: Michael Kelley <mhklinux@xxxxxxxxxxx>
> 
> Add documentation topic for PCI pass-thru devices in Linux guests
> on Hyper-V and for the associated PCI controller driver (pci-hyperv.c).
> 
> Signed-off-by: Michael Kelley <mhklinux@xxxxxxxxxxx>
> ---
>  Documentation/virt/hyperv/index.rst |   1 +
>  Documentation/virt/hyperv/vpci.rst  | 316 ++++++++++++++++++++++++++++
>  2 files changed, 317 insertions(+)
>  create mode 100644 Documentation/virt/hyperv/vpci.rst
> 
> diff --git a/Documentation/virt/hyperv/index.rst b/Documentation/virt/hyperv/index.rst
> index 4a7a1b738bbe..de447e11b4a5 100644
> --- a/Documentation/virt/hyperv/index.rst
> +++ b/Documentation/virt/hyperv/index.rst
> @@ -10,3 +10,4 @@ Hyper-V Enlightenments
>     overview
>     vmbus
>     clocks
> +   vpci
> diff --git a/Documentation/virt/hyperv/vpci.rst b/Documentation/virt/hyperv/vpci.rst
> new file mode 100644
> index 000000000000..dbca50f31923
> --- /dev/null
> +++ b/Documentation/virt/hyperv/vpci.rst
> @@ -0,0 +1,316 @@
> +.. SPDX-License-Identifier: GPL-2.0
> +
> +PCI pass-thru devices
> +=========================
> +In a Hyper-V guest VM, PCI pass-thru devices (also called
> +virtual PCI devices, or vPCI devices) are physical PCI devices
> +that are mapped directly into the VM's physical address space.
> +Guest device drivers can interact directly with the hardware
> +without intermediation by the host hypervisor.  This approach
> +provides higher bandwidth access to the device with lower
> +latency, compared with devices that are virtualized by the
> +hypervisor.  The device should appear to the guest just as it
> +would when running on bare metal, so no changes are required
> +to the Linux device drivers for the device.
> +
> +Hyper-V terminology for vPCI devices is "Discrete Device
> +Assignment" (DDA).  Public documentation for Hyper-V DDA is
> +available here: `DDA`_
> +
> +.. _DDA: https://learn.microsoft.com/en-us/windows-server/virtualization/hyper-v/plan/plan-for-deploying-devices-using-discrete-device-assignment
> +
> +DDA is typically used for storage controllers, such as NVMe,
> +and for GPUs.  A similar mechanism for NICs is called SR-IOV
> +and produces the same benefits by allowing a guest device
> +driver to interact directly with the hardware.  See Hyper-V
> +public documentation here: `SR-IOV`_
> +
> +.. _SR-IOV: https://learn.microsoft.com/en-us/windows-hardware/drivers/network/overview-of-single-root-i-o-virtualization--sr-iov-
> +
> +This discussion of vPCI devices includes DDA and SR-IOV
> +devices.
> +
> +Device Presentation
> +-------------------
> +Hyper-V provides full PCI functionality for a vPCI device when
> +it is operating, so the Linux device driver for the device can
> +be used unchanged, provided it uses the correct Linux kernel
> +APIs for accessing PCI config space and for other integration
> +with Linux.  But the initial detection of the PCI device and
> +its integration with the Linux PCI subsystem must use Hyper-V
> +specific mechanisms.  Consequently, vPCI devices on Hyper-V
> +have a dual identity.  They are initially presented to Linux
> +guests as VMBus devices via the standard VMBus "offer"
> +mechanism, so they have a VMBus identity and appear under
> +/sys/bus/vmbus/devices.  The VMBus vPCI driver in Linux at
> +drivers/pci/controller/pci-hyperv.c handles a newly introduced
> +vPCI device by fabricating a PCI bus topology and creating all
> +the normal PCI device data structures in Linux that would
> +exist if the PCI device were discovered via ACPI on a bare-
> +metal system.  Once those data structures are set up, the
> +device also has a normal PCI identity in Linux, and the normal
> +Linux device driver for the vPCI device can function as if it
> +were running in Linux on bare-metal.  Because vPCI devices are
> +presented dynamically through the VMBus offer mechanism, they
> +do not appear in the Linux guest's ACPI tables.  vPCI devices
> +may be added to a VM or removed from a VM at any time during
> +the life of the VM, and not just during initial boot.
> +
> +With this approach, the vPCI device is a VMBus device and a
> +PCI device at the same time.  In response to the VMBus offer
> +message, the hv_pci_probe() function runs and establishes a
> +VMBus connection to the vPCI VSP on the Hyper-V host.  That
> +connection has a single VMBus channel.  The channel is used to
> +exchange messages with the vPCI VSP for the purpose of setting
> +up and configuring the vPCI device in Linux.  Once the device
> +is fully configured in Linux as a PCI device, the VMBus
> +channel is used only if Linux changes the vCPU to be
> +interrupted in the guest, or 


> ..............................if the vPCI device is removed by
> +the VM while the VM is running.

This seems to conflict with the statement called out below. Did you
mean to say "if the vPCI device is removed *from* the VM..."?


> The ongoing operation of the
> +device happens directly between the Linux device driver for
> +the device and the hardware, with VMBus and the VMBus channel
> +playing no role.
> +
> +PCI Device Setup
> +----------------
> +PCI device setup follows a sequence that Hyper-V originally
> +created for Windows guests, and that can be ill-suited for
> +Linux guests due to differences in the overall structure of
> +the Linux PCI subsystem compared with Windows.  Nonetheless,
> +with a bit of hackery in the Hyper-V virtual PCI driver for
> +Linux, the virtual PCI device is setup in Linux so that
> +generic Linux PCI subsystem code and the Linux driver for the
> +device "just work".
> +
> +Each vPCI device is set up in Linux to be in its own PCI
> +domain with a host bridge.  The PCI domainID is derived from
> +bytes 4 and 5 of the instance GUID assigned to the VMBus vPCI
> +device.  The Hyper-V host does not guarantee that these bytes
> +are unique, so hv_pci_probe() has an algorithm to resolve
> +collisions.  The collision resolution is intended to be stable
> +across reboots of the same VM so that the PCI domainIDs don't
> +change, as the domainID appears in the user space
> +configuration of some devices.
> +
> +hv_pci_probe() allocates a guest MMIO range to be used as PCI
> +config space for the device.  This MMIO range is communicated
> +to the Hyper-V host over the VMBus channel as part of telling
> +the host that the device is ready to enter d0.  See
> +hv_pci_enter_d0().  When the guest subsequently accesses this
> +MMIO range, the Hyper-V host intercepts the accesses and maps
> +them to the physical device PCI config space.
> +
> +hv_pci_probe() also gets BAR information for the device from
> +the Hyper-V host, and uses this information to allocate MMIO
> +space for the BARs.  That MMIO space is then setup to be
> +associated with the host bridge so that it works when generic
> +PCI subsystem code in Linux processes the BARs.
> +
> +Finally, hv_pci_probe() creates the root PCI bus.  At this
> +point the Hyper-V virtual PCI driver hackery is done, and the
> +normal Linux PCI machinery for scanning the root bus works to
> +detect the device, to perform driver matching, and to
> +initialize the driver and device.
> +
> +PCI Device Removal
> +------------------
> +A Hyper-V host may initiate removal of a vPCI device from a
> +guest VM at any time during the life of the VM.  The removal
> +is instigated by an admin action taken on the Hyper-V host and


> +is not under the control of the guest OS.

See conflict here.

> +
> +A guest VM is notified of the removal by an unsolicited
> +"Eject" message sent from the host to the guest over the VMBus
> +channel associated with the vPCI device.  Upon receipt of such
> +a message, the Hyper-V virtual PCI driver in Linux
> +asynchronously invokes Linux kernel PCI subsystem calls to
> +shutdown and remove the device.  When those calls are
> +complete, an "Ejection Complete" message is sent back to
> +Hyper-V over the VMBus channel indicating that the device has
> +been removed.  At this point, Hyper-V sends a VMBus rescind
> +message to the Linux guest, which the VMBus driver in Linux
> +processes by removing the VMBus identity for the device.  Once
> +that processing is complete, all vestiges of the device having
> +been present are gone from the Linux kernel.  The rescind
> +message also indicates to the guest that Hyper-V has stopped
> +providing support for the vPCI device in the guest.  If the
> +guest were to attempt to access that device's MMIO space, it
> +would be an invalid reference. Hypercalls affecting the device
> +return errors, and any further messages sent in the VMBus
> +channel are ignored.
> +
> +After sending the Eject message, Hyper-V allows the guest VM
> +60 seconds to cleanly shutdown the device and respond with
> +Ejection Complete before sending the VMBus rescind
> +message.  If for any reason the Eject steps don't complete
> +within the allowed 60 seconds, the Hyper-V host forcibly
> +performs the rescind steps, which will likely result in
> +cascading errors in the guest because the device is now no
> +longer present from the guest standpoint and accessing the
> +device MMIO space will fail.
> +
> +Because ejection is asynchronous and can happen at any point
> +during the guest VM lifecycle, proper synchronization in the
> +Hyper-V virtual PCI driver is very tricky.  Ejection has been
> +observed even before a newly offered vPCI device has been
> +fully setup.  The Hyper-V virtual PCI driver has been updated
> +several times over the years to fix race conditions when
> +ejections happen at inopportune times. Care must be taken when
> +modifying this code to prevent re-introducing such problems.
> +See comments in the code.
> +
> +Interrupt Assignment
> +--------------------
> +The Hyper-V virtual PCI driver supports vPCI devices using
> +MSI, multi-MSI, or MSI-X.  Assigning the guest vCPU that will> +receive the interrupt for a particular MSI or MSI-X message is
> +complex because of the way the Linux setup of IRQs maps onto
> +the Hyper-V interfaces.  For the single-MSI and MSI-X cases,
> +Linux calls hv_compse_msi_msg() twice, with the first call
> +containing a dummy vCPU and the second call containing the
> +real vCPU.  Furthermore, hv_irq_unmask() is finally called
> +(on x86) or the GICD registers are set (on arm64) to specify
> +the real vCPU again.  Each of these three calls interact
> +with Hyper-V, which must decide which physical CPU should
> +receive the interrupt before it is forwarded to the guest VM.
> +Unfortunately, the Hyper-V decision-making process is a bit
> +limited, and can result in concentrating the physical
> +interrupts on a single CPU, causing a performance bottleneck.
> +See details about how this is resolved in the extensive
> +comment above the function hv_compose_msi_req_get_cpu().
> +
> +The Hyper-V virtual PCI driver implements the
> +irq_chip.irq_compose_msi_msg function as hv_compose_msi_msg().
> +Unfortunately, on Hyper-V the implementation requires sending
> +a VMBus message to the Hyper-V host and awaiting an interrupt
> +indicating receipt of a reply message.  Since
> +irq_chip.irq_compose_msi_msg can be called with IRQ locks
> +held, it doesn't work to do the normal sleep until awakened by
> +the interrupt. Instead hv_compose_msi_msg() must send the
> +VMBus message, and then poll for the completion message. As
> +further complexity, the vPCI device could be ejected/rescinded
> +while the polling is in progress, so this scenario must be
> +detected as well.  See comments in the code regarding this
> +very tricky area.
> +
> +Most of the code in the Hyper-V virtual PCI driver (pci-
> +hyperv.c) applies to Hyper-V and Linux guests running on x86
> +and on arm64 architectures.  But there are differences in how
> +interrupt assignments are managed.  On x86, the Hyper-V
> +virtual PCI driver in the guest must make a hypercall to tell
> +Hyper-V which guest vCPU should be interrupted by each
> +MSI/MSI-X interrupt, and the x86 interrupt vector number that
> +the x86_vector IRQ domain has picked for the interrupt.  This
> +hypercall is made by hv_arch_irq_unmask().  On arm64, the
> +Hyper-V virtual PCI driver manages the allocation of an SPI
> +for each MSI/MSI-X interrupt.  The Hyper-V virtual PCI driver
> +stores the allocated SPI in the architectural GICD registers,
> +which Hyper-V emulates, so no hypercall is necessary as with
> +x86.  Hyper-V does not support using LPIs for vPCI devices in
> +arm64 guest VMs because it does not emulate a GICv3 ITS.
> +
> +The Hyper-V virtual PCI driver in Linux supports vPCI devices
> +whose drivers create managed or unmanaged Linux IRQs.  If the
> +smp_affinity for an unmanaged IRQ is updated via the /proc/irq
> +interface, the Hyper-V virtual PCI driver is called to tell
> +the Hyper-V host to change the interrupt targeting and
> +everything works properly.  However, on x86 if the x86_vector
> +IRQ domain needs to reassign an interrupt vector due to
> +running out of vectors on a CPU, there's no path to inform the
> +Hyper-V host of the change, and things break.  Fortunately,
> +guest VMs operate in a constrained device environment where
> +using all the vectors on a CPU doesn't happen. Since such a
> +problem is only a theoretical concern rather than a practical
> +concern, it has been left unaddressed.
> +
> +DMA
> +---
> +By default, Hyper-V pins all guest VM memory in the host
> +when the VM is created, and programs the physical IOMMU to
> +allow the VM to have DMA access to all its memory.  Hence
> +it is safe to assign PCI devices to the VM, and allow the
> +guest operating system to program the DMA transfers.  The
> +physical IOMMU prevents a malicious guest from initiating
> +DMA to memory belonging to the host or to other VMs on the
> +host. From the Linux guest standpoint, such DMA transfers
> +are in "direct" mode since Hyper-V does not provide a virtual
> +IOMMU in the guest.
> +
> +Hyper-V assumes that physical PCI devices always perform
> +cache-coherent DMA.  When running on x86, this behavior is
> +required by the architecture.  When running on arm64, the
> +architecture allows for both cache-coherent and
> +non-cache-coherent devices, with the behavior of each device
> +specified in the ACPI DSDT.  But when a PCI device is assigned
> +to a guest VM, that device does not appear in the DSDT, so the
> +Hyper-V VMBus driver propagates cache-coherency information
> +from the VMBus node in the ACPI DSDT to all VMBus devices,
> +including vPCI devices (since they have a dual identity as a VMBus
> +device and as a PCI device).  See vmbus_dma_configure().
> +Current Hyper-V versions always indicate that the VMBus is
> +cache coherent, so vPCI devices on arm64 always get marked as
> +cache coherent and the CPU does not perform any sync
> +operations as part of dma_map/unmap_*() calls.
> +
> +vPCI protocol versions
> +----------------------
> +As previously described, during vPCI device setup and teardown
> +messages are passed over a VMBus channel between the Hyper-V
> +host and the Hyper-v vPCI driver in the Linux guest.  Some
> +messages have been revised in newer versions of Hyper-V, so
> +the guest and host must agree on the vPCI protocol version to
> +be used.  The version is negotiated when communication over
> +the VMBus channel is first established.  See
> +hv_pci_protocol_negotiation(). Newer versions of the protocol
> +extend support to VMs with more than 64 vCPUs, and provide
> +additional information about the vPCI device, such as the
> +guest virtual NUMA node to which it is most closely affined in
> +the underlying hardware.
> +
> +Guest NUMA node affinity
> +------------------------
> +When the vPCI protocol version provides it, the guest NUMA
> +node affinity of the vPCI device is stored as part of the Linux
> +device information for subsequent use by the Linux driver. See
> +hv_pci_assign_numa_node().  If the negotiated protocol version
> +does not support the host providing NUMA affinity information,
> +the Linux guest defaults the device NUMA node to 0.  But even
> +when the negotiated protocol version includes NUMA affinity
> +information, the ability of the host to provide such
> +information depends on certain host configuration options.  If
> +the guest receives NUMA node value "0", it could mean NUMA
> +node 0, or it could mean "no information is available".
> +Unfortunately it is not possible to distinguish the two cases
> +from the guest side.
> +
> +PCI config space access in a CoCo VM
> +------------------------------------
> +Linux PCI device drivers access PCI config space using a
> +standard set of functions provided by the Linux PCI subsystem.
> +In Hyper-V guests these standard functions map to functions
> +hv_pcifront_read_config() and hv_pcifront_write_config()
> +in the Hyper-V virtual PCI driver.  In normal VMs,
> +these hv_pcifront_*() functions directly access the PCI config
> +space, and the accesses trap to Hyper-V to be handled.
> +But in CoCo VMs, memory encryption prevents Hyper-V
> +from reading the guest instruction stream to emulate the
> +access, so the hv_pcifront_*() functions must invoke
> +hypercalls with explicit arguments describing the access to be
> +made.
> +
> +Config Block back-channel
> +-------------------------
> +The Hyper-V host and Hyper-V virtual PCI driver in Linux
> +together implement a non-standard back-channel communication
> +path between the host and guest.  The back-channel path uses
> +messages sent over the VMBus channel associated with the vPCI
> +device.  The functions hyperv_read_cfg_blk() and
> +hyperv_write_cfg_blk() are the primary interfaces provided to
> +other parts of the Linux kernel.  As of this writing, these
> +interfaces are used only by the Mellanox mlx5 driver to pass
> +diagnostic data to a Hyper-V host running in the Azure public
> +cloud.  The functions hyperv_read_cfg_blk() and
> +hyperv_write_cfg_blk() are implemented in a separate module
> +(pci-hyperv-intf.c, under CONFIG_PCI_HYPERV_INTERFACE) that
> +effectively stubs them out when running in non-Hyper-V
> +environments.

Otherwise, FWIW

Reviewed-by: Easwar Hariharan <eahariha@xxxxxxxxxxxxxxxxxxx>





[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux