Kernel KVM virtualization development
 help / color / mirror / Atom feed
From: bugzilla-daemon@kernel.org
To: kvm@vger.kernel.org
Subject: [Bug 222317] New: pci: 8086:4c43 host bridge enumerated as class 0x0604 breaks Intel IGD passthrough
Date: Sun, 22 Nov 2026 19:51:11 +0000	[thread overview]
Message-ID: <bug-222317-28872@https.bugzilla.kernel.org/> (raw)

https://bugzilla.kernel.org/show_bug.cgi?id=222317

            Bug ID: 222317
           Summary: pci: 8086:4c43 host bridge enumerated as class 0x0604
                    breaks Intel IGD passthrough
           Product: Drivers
           Version: 2.5
          Hardware: Intel
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P3
         Component: PCI
          Assignee: drivers_pci@kernel-bugs.osdl.org
          Reporter: 2039475150@qq.com
                CC: kvm@vger.kernel.org
        Regression: No

Filed under Drivers/PCI because the underlying defect is the PCI class code
enumeration of the CPU host bridge; the visible failure happens in vfio-pci.
CC'd VFIO/KVM people for the driver side.

Passing the integrated GPU (Intel UHD Graphics P750, 8086:4c90) to a QEMU/KVM
guest fails on this machine. vfio_pci_igd_cfg_init() returns -EINVAL during
vfio_pci_open_device(), so QEMU never gets the device:

    vfio-pci 0000:00:02.0: resetting
    vfio-pci 0000:00:02.0: reset done
    vfio-pci 0000:00:02.0: Failed to setup Intel IGD regions
    kvm: -device vfio-pci,host=0000:00:02.0,...: vfio 0000:00:02.0:
         error getting device from group 0: Invalid argument

Root cause
----------

drivers/vfio/pci/vfio_pci_igd.c, vfio_pci_igd_cfg_init():

    host_bridge = pci_get_domain_bus_and_slot(0, 0, PCI_DEVFN(0, 0));
    if (!host_bridge)
            return -ENODEV;

    if (host_bridge->vendor != PCI_VENDOR_ID_INTEL ||
        host_bridge->class != (PCI_CLASS_BRIDGE_HOST << 8)) {
            pci_dev_put(host_bridge);
            return -EINVAL;
    }

This firmware enumerates the CPU host bridge 00:00.0 (Intel 8086:4c43,
Rocket Lake-S Host Bridge/DRAM Controller) with class code 0x060400
(PCI-to-PCI bridge) instead of the architecturally correct 0x060000
(Host Bridge):

    $ setpci -s 00:00.0 0x09.b 0x0a.b 0x0b.b
    00 04 06          # class 0x060400

    $ setpci -s 00:00.0 HEADER_TYPE.b
    01                # type 1 header, i.e. looks like a PCI bridge

    $ setpci -s 00:00.0 0x18.l
    ffffffff          # primary/secondary/subordinate bus numbers all 0xff

lspci -t shows the PCI core actually walking that bogus secondary bus:

    -[0000:00]-+-00.0-[ff]--
               +-01.0-[02]----00.0
               +-02.0
               ...

The other half of the IGD check passes: the OpRegion is valid.

    $ setpci -s 00:02.0 fc.l
    953ce018          # OpRegion base in PCI config space 0xFC

    # read at that physical address:
    000000  49 6e 74 65 6c 47 72 61 70 68 69 63 73 4d 65 6d  IntelGraphicsMem
    000010  08 00 00 00 00 00 01 02 ...                      size = 8 KB

So vfio_pci_igd_opregion_init() succeeds and only
vfio_pci_igd_cfg_init() fails. The IGD itself is fine; the failure is purely
the host-bridge class check.

Why this is worth fixing in the kernel
--------------------------------------

The class code is a read-only PCI register and 0x0604 here is a firmware
defect. A user cannot fix it. The only workaround I found is patching the
cmpl immediate inside vfio-pci.ko after every kernel upgrade:

    bf3: 81 78 44 00 00 06 00   cmpl $0x60000,0x44(%rax)   # stock
         81 78 44 00 04 06 00   cmpl $0x60400,0x44(%rax)   # patched

With that two-byte change the check passes and the host-bridge, LPC and
OpRegion regions all register correctly; QEMU then reports a real Intel IGD
instead of a plain PCI device. This is a poor long-term solution: it breaks on
every rebuild and requires stripping the module signature.

Suggested fixes
---------------

1. Preferred: a DMI-scoped quirk in drivers/pci/quirks.c that restores
   dev->class = PCI_CLASS_BRIDGE_HOST << 8 for 8086:4c43 on the affected
   board(s) only. This keeps the change off the other machines that report
   8086:4c43 with the correct 0x0600, and it fixes the value for every
   consumer instead of only vfio.

       static void quirk_intel_host_bridge_class(struct pci_dev *dev)
       {
               dev->class = PCI_CLASS_BRIDGE_HOST << 8;
       }
       DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, 0x4c43,
                               quirk_intel_host_bridge_class);

   guarded by dmi_check_system() for Gigabyte/W480 VISION W.

2. Alternative: an explicit escape hatch in vfio-pci (module parameter or
   sysfs attribute) to relax the host-bridge class check, so affected users
   do not have to binary-patch the module.

I understand neither option "fixes the firmware", and that the class
check\nexists for a reason. Option 1 seems the least invasive way to make this
hardware usable.

Environment
-----------

    CPU:      Intel Xeon W-1390P (Rocket Lake)
    Board:    Gigabyte W480 VISION W, BIOS F25a (2026-06-11)
    iGPU:     0000:00:02.0  Intel UHD Graphics P750 [8086:4c90] rev 04
    Host Br:  0000:00:00.0  Intel [8086:4c43] rev 01, class 0x060400
    IOMMU:    group 0, iGPU alone
    Kernel:   7.0.14-28-pve (Proxmox VE 9.2, based on Ubuntu-7.0.0 + stable)
    cmdline:  BOOT_IMAGE=/boot/vmlinuz-7.0.14-28-pve root=/dev/mapper/pve-root
ro quiet
    CONFIG_VFIO_PCI=m

Note: this board also hit a separate PCI bug involving
CONFIG_PCI_DYNAMIC_OF_NODES
(CVE-2026-98504, fixed in 7.3-rc5 / 6.18.55). That is unrelated to this report;
CONFIG_PCI_DYNAMIC_OF_NODES is not even enabled in this build.

Note 2: the failing sequence above was observed on an earlier kernel on this
machine; on the current 7.0.14-28-pve the module is locally patched, so dmesg
no longer shows it. drivers/vfio/pci/vfio_pci_igd.c is unchanged in v7.3.3.

-- 
You are receiving this mail because:
You are on the CC list for the bug.

                 reply	other threads:[~2026-11-22 19:51 UTC|newest]

Thread overview: [no followups] expand[flat|nested]  mbox.gz  Atom feed

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=bug-222317-28872@https.bugzilla.kernel.org/ \
    --to=bugzilla-daemon@kernel.org \
    --cc=kvm@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox