On Mon, Nov 02, 2026 at 12:12:36PM -0600, Tushar Dave wrote: > This patch series introduces a mechanism to specify Guest Physical Addresses > (GPAs) for PCI BARs, allowing explicit placement of guest MMIO BAR addresses > to match host physical addresses for assigned devices.
My understanding is that all this does it pass a property to the firmware asking it to set the bar to a given value? There are two standard ways to do this: - EA capability in pci config space - fixed BAR _DSM in ACPI If none of these are appropriate, pls document why. > On some platforms, P2P DMA between devices in the same IOMMU group is required > for the desired data path and performance. The PCI fabric ACS is configured > to permit direct P2P without routing traffic through the host bridge. > > To support this multi-device IOMMU group P2P scenario in virtualization, the > VM may need to preserve the host PCI BAR addresses in the guest. > > A new fixed-bar-N property on a PCI endpoint device specifies the required > guest address for BAR N (N = 0 through 5): > > -device some-device,fixed-bar-0=<addr>[,fixed-bar-1=<addr>]... > > Once any fixed-bar-N property is set on a device, every memory BAR on that > device must have a fixed address. Under a PCIe root port, fixed BAR placement > is all-or-nothing within that root port's subtree. If one endpoint below the > root port has a fixed BAR, every other endpoint with a memory BAR below that > same root port must also have all of its memory BARs fixed. > > QEMU validates the configuration when the machine is done initializing. Each > address must be aligned to its BAR size, fall within the appropriate PCIe MMIO > aperture, and not overlap any other fixed BAR. QEMU aborts on any validation > failure. QEMU performs neither PCI enumeration nor BAR allocation; the > validated configuration is exported to firmware through the "etc/fixed-bars" > fw_cfg blob, which carries each PCI function's vendor/device identification, > firmware device path, and, for fixed devices, the configured address, size, > and type of each BAR. The blob format is documented in > docs/specs/pci-fixed-bars.rst. > > The fixed-bar-N property and its validation are generic and apply to any PCI > device. A machine type opts in by setting the new pci_fixed_bar_supported > MachineClass flag. Machine types that leave it unset reject fixed-bar-N. This > series enables it on the virt machine only, by setting the flag and wiring up > the fw_cfg export. Other machine types would need equivalent wiring to expose > the same mechanism to firmware. > > On the firmware side, the existing IncompatiblePciDeviceSupportDxe driver is > extended to consume this blob. During device discovery, its CheckDevice() > implementation returns ACPI address descriptors with _MIF|_MAF set for the > fixed BARs found in the blob. PciBusDxe's resource allocator already supports > fixed BARs, including sizing and pinning the bridge windows above them. It now > treats a BAR as fixed when the descriptor carries _MIF|_MAF, and programs the > fixed address into the BAR register. > > Fixed BAR placement requires the guest to preserve the PCI resource > configuration established by firmware. This is requested through the PCI boot > configuration preservation mechanism using ACPI _DSM function 5. Fixed BAR > configuration is rejected when firmware is not loaded or ACPI is disabled. > > Example 1: VFIO passthrough device with fixed BARs and the PCIe MMIO > aperture adjusted to accommodate the fixed addresses. > > -machine virt,...,highmem-mmio-base=4T,highmem-mmio-size=4T > -device pcie-root-port,id=pcie.port1,bus=pcie.1,chassis=1,io-reserve=0 > -device x3130-upstream,id=upstream1,bus=pcie.port1 > -device xio3130-downstream,id=downstream1_1,bus=upstream1,chassis=1,slot=2 > -device vfio-pci-nohotplug,host=0018:06:00.0,bus=downstream1_1,id=dev0,\ > > fixed-bar-0=0x48000000000,fixed-bar-2=0x50000000000,fixed-bar-4=0x58000000000 > > Example 2: VFIO and emulated devices under the same root port with fixed BARs > and the PCIe MMIO aperture adjusted using highmem-mmio-base. > > -machine virt,...,highmem-mmio-base=64T,highmem-mmio-size=64T > -device pxb-pcie,id=pcie.1,bus_nr=1,bus=pcie.0 > -device pcie-root-port,id=pcie.port1,bus=pcie.1,chassis=1,io-reserve=0 > -device x3130-upstream,id=upstream1,bus=pcie.port1 > -device xio3130-downstream,id=downstream1_1,bus=upstream1,chassis=1,slot=1 > -device > vfio-pci-nohotplug,host=0018:06:00.0,bus=downstream1_1,rombar=0,id=dev0,iommufd=iommufd0,\ > > fixed-bar-0=0x6b4000000000,fixed-bar-2=0x6b8000000000,fixed-bar-4=0x6c0000000000 > -device xio3130-downstream,id=downstream1_2,bus=upstream1,chassis=1,slot=2 > -device > vfio-pci,host=0012:03:00.1,bus=downstream1_2,id=nic1,iommufd=iommufd0,fixed-bar-0=0x6c8000000000 > -device xio3130-downstream,id=downstream1_3,bus=upstream1,chassis=1,slot=3 > -device > e1000,netdev=net0,bus=downstream1_3,id=nic_emu,fixed-bar-0=0x20000000 > > Limitations: > > * I/O BARs and the expansion ROM BAR are not covered by fixed-bar-N. > * SR-IOV VF BARs are not covered. > > Testing: > > The series was tested on the AArch64 virt machine with a range of PCIe > configurations, including emulated and VFIO passthrough devices, PCIe > switches, single and multiple fixed-BAR endpoints behind a shared root port, > devices attached directly to the host bridge, and endpoints behind nested > bridges. All rejection paths were also exercised, with the corresponding > error messages verified. These included unsupported machine types, > hot-plugged devices, non-existent BARs, I/O BARs, unaligned, overlapping, or > out-of-aperture BARs, incomplete BAR configurations, highmem-mmio=off, > missing firmware, and highmem-mmio-base values that cause address-space > overflow. > > Changes since v1: > > * Add a pci_fixed_bar_supported MachineClass flag; fixed-bar-N is rejected > on machine types that do not set it. virt sets it. > * Keep the fixed BAR claims in FixedBarsInfo instead of a static array, and > report validation failures through Error ** to the machine code instead of > calling exit(1). > * Allow a 64-bit non-prefetchable BAR in either aperture when the device is > attached to the host bridge; behind a PCI bridge it must lie below 4 GiB. > A 64-bit prefetchable BAR must still lie in the 64-bit aperture. > * Treat the high PCIe MMIO aperture as absent with highmem-mmio=off, and > reject fixed BAR configuration unless firmware is loaded and ACPI is > enabled. > * Check highmem-mmio-base + highmem-mmio-size against the PA space without > overflow, and document highmem-mmio-base in docs/system/arm/virt.rst. > * Document the etc/fixed-bars format in docs/specs/pci-fixed-bars.rst, and > note that the firmware device path is carried for future matching. > * Move pci-fixed-bar.h to include/hw/pci/. > > v1: > https://lore.kernel.org/qemu-arm/[email protected]/ > > A git branch with this series applied is available at: > https://github.com/tdavenvidia/upstream-qemu/tree/Upstream-fixed-bar-Oct-29-V2 > > The related EDK2 change is available at: > https://github.com/tianocore/edk2/pull/13183 > > Tushar Dave (5): > hw/pci: introduce PCI fixed BAR properties > hw/pci: add validation for fixed BAR configuration > hw/pci: add fixed BAR fw_cfg blob export > hw/arm/virt: export fixed BAR metadata via fw_cfg > hw/arm/virt: add highmem-mmio-base property > > docs/specs/index.rst | 1 + > docs/specs/pci-fixed-bars.rst | 75 +++++++ > docs/system/arm/virt.rst | 4 + > hw/arm/virt.c | 82 ++++++- > hw/pci/meson.build | 2 + > hw/pci/pci-fixed-bar-blob.c | 308 +++++++++++++++++++++++++++ > hw/pci/pci-fixed-bar-validate.c | 365 ++++++++++++++++++++++++++++++++ > hw/pci/pci-fixed-bar-validate.h | 22 ++ > hw/pci/pci.c | 71 +++++++ > include/hw/core/boards.h | 3 + > include/hw/pci/pci-fixed-bar.h | 63 ++++++ > include/hw/pci/pci_device.h | 7 + > 12 files changed, 1002 insertions(+), 1 deletion(-) > create mode 100644 docs/specs/pci-fixed-bars.rst > create mode 100644 hw/pci/pci-fixed-bar-blob.c > create mode 100644 hw/pci/pci-fixed-bar-validate.c > create mode 100644 hw/pci/pci-fixed-bar-validate.h > create mode 100644 include/hw/pci/pci-fixed-bar.h > > -- > 2.34.1
