[REGRESSION] PCI: Dynamic OF node creation hangs on invalid bridge configuration

[Angel J]

Hello,

This regression was introduced in v6.17-rc1 and remains in the dynamic
PCI OF node code. Linux 6.18 fails very early during boot on my Dell XPS
8940 when CONFIG_PCI_DYNAMIC_OF_NODES=y. The display goes black before
any useful console or pstore output appears, and the machine remains
hung.

Disabling CONFIG_PCI_DYNAMIC_OF_NODES makes the same kernel boot normally.
Linux v6.19-rc5 also boots because the RP1 driver stopped selecting that
option, not because the dynamic PCI OF node code changed.

Hardware:

  System: Dell XPS 8940, board 0K3CM7, BIOS 2.27.1 (2025-04-03)
  CPU: Intel Core i7-11700 (Rocket Lake)
  Triggering device: 0000:00:00.0, Intel 8086:4c43
  Firmware boot: UEFI, ACPI-based x86_64

Test results:

  Linux 6.12.103: good
  Linux v6.17-rc1: bad
  Linux 6.18.44, CONFIG_PCI_DYNAMIC_OF_NODES=y: bad
  Linux 6.18.44, CONFIG_PCI_DYNAMIC_OF_NODES=n: good
  Linux 6.18.44, CONFIG_PCI_DYNAMIC_OF_NODES=y, skip invalid bridge: good
  Linux v6.19-rc4: bad
  Linux v6.19-rc5: good
  Linux 4621c338d33f: bad
  Linux e55feea3a03a: good

The first-parent boundary between the last two tests is:

  e55feea3a03a ("Merge tag 'soc-fixes-6.19' ...")

That merge contains ce26f588c831 ("misc: rp1: drop overlay support"), which
removes:

  select PCI_DYNAMIC_OF_NODES

from MISC_RP1. The generated configurations differed only by:

  -CONFIG_PCI_DYNAMIC_OF_NODES=y
  +# CONFIG_PCI_DYNAMIC_OF_NODES is not set

CONFIG_MISC_RP1=m in both. The known-bad 4621c338d33f and Linux 6.18.44
both boot after disabling PCI_DYNAMIC_OF_NODES (and MISC_RP1, whose Kconfig
select otherwise forces it on).

There is no RP1 device in this machine, and its driver does not probe.
MISC_RP1 affects this x86 system only because its Kconfig select enables
PCI_DYNAMIC_OF_NODES globally.

I narrowed the hang to dynamic OF node creation for 0000:00:00.0
(8086:4c43), the Intel Rocket Lake-S integrated memory controller. On
this machine it reports PCI class 0604 and header type 1, but it is bound
to icl_uncore and is not a usable forwarding bridge:

  Bus: primary=ff, secondary=ff, subordinate=ff
  I/O and memory bridge windows: invalid/all ones
  kernel: bridge configuration invalid ([bus ff-ff]), reconfiguring

A minimal dynamic node for this device boots. Adding only
device_type="pci" makes it hang. Full PCI dynamic OF node generation also
boots when only 0000:00:00.0 is skipped.

I also tested a generic guard in of_pci_make_dev_node(). It reads
PCI_PRIMARY_BUS and skips node creation if the primary bus does not match
pdev->bus->number, the secondary bus is not greater than the primary bus,
or the secondary bus is greater than the subordinate bus. Linux 6.18.44
boots with CONFIG_PCI_DYNAMIC_OF_NODES=y and logs:

  pci 0000:00:00.0: skipping dynamic OF node for invalid bridge

The regression was introduced by 49d63971f963 in v6.17-rc1. Commit
ce26f588c831 made v6.19-rc5 boot by removing MISC_RP1's select of
PCI_DYNAMIC_OF_NODES. That avoids the failing path in the tested
configuration but does not change of_pci_make_dev_node().

I do not have a log from the failed boots. No output appeared with
earlyprintk=efi,keep, keep_bootcon, ignore_loglevel, loglevel=8,
initcall_debug, and EFI pstore enabled. Blacklisting both GPU drivers and
disabling the Intel IOMMU did not make the bad kernel boot.

Should of_pci_make_dev_node() reject devices with invalid bridge bus
configuration, or should 8086:4c43 use a device-specific quirk? I can
prepare and test a patch for either approach. For Linux 6.18 stable,
would that fix be preferable to backporting ce26f588c831 and its
dependencies?

#regzbot introduced: 49d63971f963
#regzbot title: PCI_DYNAMIC_OF_NODES early boot hang on ACPI x86

Thanks,
Angel J

Attachment: signature.asc
Description: OpenPGP digital signature