Re: [PATCH 0/7] OvmfPkg/PciHotPlugInitDxe: obey PCI resource reservation hints from QEMU

[Laszlo Ersek <lersek@redhat.com>]

On 09/25/17 21:58, Laszlo Ersek wrote:
> Repo:   https://github.com/lersek/edk2.git
> Branch: pci_resource_reserve
> 
> "OvmfPkg/PciHotPlugInitDxe" implements EFI_PCI_HOT_PLUG_INIT_PROTOCOL
> from the PI spec. Currently it returns fixed resource reservations (512B
> IO space and 2MB non-prefetchable 32-bit MMIO space) for all bridges.
> This is not flexible enough:
> 
> (1) For PCI Express bridges (such as root ports and downstream ports),
> reserving any IO space is wasteful, and limits the number of ports.
> (This is because each such port can only accept one (multi-function)
> device, and because the 512B reservation is rounded up to 4KB, and
> because the IO space under OVMF/Q35 consists of 10 * 4KB.)
> 
> (2) For hot-plugging PCI Express devices with large MMIO BARs into such
> ports, the reservation of only 2MB MMIO (and even that limited to
> non-prefetchable, hence 32-bit, address space) is insufficient.
> 
> QEMU has recently gained the ability to control these reservation sizes
> on the command line. At this point the hints are only supported for the
> "pcie-root-port" device -- which implies the Q35 machine type, given
> that i440fx only supports conventional PCI, not PCI Express --, with the
> following properties:
> 
> - pcie-root-port.bus-reserve=uint32
> - pcie-root-port.io-reserve=size
> - pcie-root-port.mem-reserve=size
> - pcie-root-port.pref32-reserve=size
> - pcie-root-port.pref64-reserve=size
> 
> The hints are exposed to the guest in the conventional config space of
> the device, using a vendor-specific capability (which is documented in
> the QEMU source tree; see all references in the individual patches). It
> is expected that, if any future hotplug controllers in QEMU gain the
> same ability, they will reuse the mechanism identically.
> 
> This series generally implements the parsing of this capability in
> OvmfPkg/PciHotPlugInitDxe, and the translation thereof to the ACPI
> address space descriptor format that the Platform Init spec defines for
> EFI_PCI_HOT_PLUG_INIT_PROTOCOL.GetResourcePadding().
> 
> The first patch is for "MdePkg/IndustryStandard"; it introduces the
> vendor-specific capability header, from the PCI 2.3 spec. The rest of
> the patches is for OvmfPkg.
> 
> Cc: Jordan Justen <jordan.l.justen@intel.com>
> Cc: Liming Gao <liming.gao@intel.com>
> Cc: Marcel Apfelbaum <marcel@redhat.com>
> Cc: Michael D Kinney <michael.d.kinney@intel.com>
> Cc: Ruiyu Ni <ruiyu.ni@intel.com>
> 
> Thanks
> Laszlo
> 
> Laszlo Ersek (7):
>   MdePkg/IndustryStandard/Pci23: add vendor-specific capability header
>   OvmfPkg/IndustryStandard: define PCI Capabilities for QEMU's PCI
>     Bridges
>   OvmfPkg/PciHotPlugInitDxe: clean up protocol usage comment
>   OvmfPkg/PciHotPlugInitDxe: clean up addr. range for non-prefetchable
>     MMIO
>   OvmfPkg/PciHotPlugInitDxe: generalize RESOURCE_PADDING composition
>   OvmfPkg/PciHotPlugInitDxe: add helper functions for setting up
>     paddings
>   OvmfPkg/PciHotPlugInitDxe: translate QEMU's resource reservation hints
> 
>  MdePkg/Include/IndustryStandard/Pci23.h                      |  10 +
>  OvmfPkg/Include/IndustryStandard/QemuPciBridgeCapabilities.h |  60 ++
>  OvmfPkg/PciHotPlugInitDxe/PciHotPlugInit.c                   | 661 ++++++++++++++++++--
>  OvmfPkg/PciHotPlugInitDxe/PciHotPlugInit.inf                 |   6 +-
>  4 files changed, 674 insertions(+), 63 deletions(-)
>  create mode 100644 OvmfPkg/Include/IndustryStandard/QemuPciBridgeCapabilities.h
> 

Thanks All for the feedback, pushed as commit range
9425b34925d0..fe4049471bdf.

Laszlo