Re: [PATCH] PCI/ACPI: Do not fiddle with ExtTag and RO in program_hpx_type2

[Bjorn Helgaas <helgaas@kernel.org>]

On Wed, Jan 07, 2026 at 05:22:05PM -0600, Bjorn Helgaas wrote:
> On Fri, Dec 05, 2025 at 03:28:29PM +0100, Håkon Bugge wrote:
> > After commit 60db3a4d8cc9 ("PCI: Enable PCIe Extended Tags if
> > supported"), the kernel controls enablement of extended tags
> > (ExtTag). Similar, after commit a99b646afa8a ("PCI: Disable PCIe
> > Relaxed Ordering if unsupported"), the kernel controls the relaxed
> > ordering bit (RO), in the sense that the kernel keeps the bit set (if
> > already set) unless the RC does not support it.
> > 
> > On some platforms, when program_hpx_type2() is called and the _HPX
> > object's Type2 records are, let's say, dubious, we may end up
> > resetting ExtTag and RO, although they were explicit set or kept set
> > by the OSPM [1].
> 
> This text about Type 2 records in ACPI r6.6, sec 6.2.10.3, seems a
> little ambiguous to me:
> 
>   A PCI Express-aware OS that has assumed ownership of native hot plug
>   (via _OSC) but does not support or does not have ownership of the
>   AER register set must use the data values returned by the _HPX
>   object’s Type 2 record to program the AER registers of a hot-added
>   PCI Express device. However, since the Type 2 record also includes
>   register bits that have functions other than AER, OSPM must ignore
>   values contained within this setting record that are not applicable.
> 
> If I squint, I can read that as meaning that Type 2 is really there
> just for AER, and the OS:
> 
>   - should only use a Type 2 record when it owns PCIe native hotplug
>     (native_pcie_hotplug) but does not own AER (!native_aer),
> 
>   - should only program AER registers, and
> 
>   - should *ignore* bits unrelated to AER.
> 
> Most of the registers in Type 2 are in the AER Capability.  Device
> Control is in the PCIe Capability, but if _OSC has granted AER
> ownership to the OS, that includes the Error Reporting Enable bits in
> Device Control (there's a PCI Firmware spec ECN to this effect:
> https://members.pcisig.com/wg/Firmware/document/20514).
> 
> Type 2 does include Link Control, which is in the PCIe Capability and
> doesn't seem related to AER, so maybe I'm on the wrong track.  But if
> Type 2 was intended to handle things *other* than AER, I would think
> the PCIe Capability Slot Control and Root Control would have been
> included.
> 
> So *maybe* program_hpx_type2() should mask out everything from
> pci_exp_devctl_or except PCI_EXP_DEVCTL_CERE, PCI_EXP_DEVCTL_NFERE,
> PCI_EXP_DEVCTL_FERE, and PCI_EXP_DEVCTL_URRE?  I have no idea what we
> would do with Link Control though.
> 
> I wish I knew the history of this, but I don't.

I poked around in the old specs and found that _HPX was added in ACPI
r3.0, where the Type 2 record only contained AER registers
(Uncorrectable Error Mask, Uncorrectable Error Severity, and
Correctable Error Mask), and the description said this:

  OSPM will only evaluate _HPX with Setting Record – Type 2 if OSPM is
  not controlling the PCI Express Advanced Error Reporting capability.

ACPI r4.0 changed _HPX to the current r6.6 description and added all
the registers included in r6.6 (without changing the record revision).

So I think we should do what I proposed above (only do
program_hpx_type2() if the OS owns PCIe hotplug but not AER, and only
update the CERE/NFERE/FERE/URRE bits in Device Control).  I would
probably log the Link Control values if they're set, but otherwise
ignore them.

> > The Advanced Configuration and Power Interface (ACPI) Specification
> > version 6.6 has a provision that gives the OSPM the ability to
> > control these bits any way. In a note in section 6.2.9, it is stated:
> > 
> > "OSPM may override the settings provided by the _HPX object's Type2
> > record (PCI Express Settings) or Type3 record (PCI Express Descriptor
> > Settings) when OSPM has assumed native control of the corresponding
> > feature."
> > 
> > So, in order to preserve the increased performance of ExtTag and RO on
> > platforms that support any of these, we make sure program_hpx_type2()
> > doesn't reset them.
> > 
> > [1] Operating System-directed configuration and Power Management
> > 
> > Fixes: 60db3a4d8cc9 ("PCI: Enable PCIe Extended Tags if supported")
> > Fixes: a99b646afa8a ("PCI: Disable PCIe Relaxed Ordering if unsupported")
> > Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
> > ---
> >  drivers/pci/pci-acpi.c | 15 +++++++++------
> >  1 file changed, 9 insertions(+), 6 deletions(-)
> > 
> > diff --git a/drivers/pci/pci-acpi.c b/drivers/pci/pci-acpi.c
> > index 9369377725fa0..6a2ae1b821732 100644
> > --- a/drivers/pci/pci-acpi.c
> > +++ b/drivers/pci/pci-acpi.c
> > @@ -324,15 +324,18 @@ static void program_hpx_type2(struct pci_dev *dev, struct hpx_type2 *hpx)
> >  		return;
> >  	}
> >  
> > -	/*
> > -	 * Don't allow _HPX to change MPS or MRRS settings.  We manage
> > -	 * those to make sure they're consistent with the rest of the
> > -	 * platform.
> > +	/* Don't allow _HPX to change MPS, MRRS, ExtTag, or RO
> > +	 * settings.  We manage those to make sure they're consistent
> > +	 * with the rest of the platform.
> >  	 */
> >  	hpx->pci_exp_devctl_and |= PCI_EXP_DEVCTL_PAYLOAD |
> > -				    PCI_EXP_DEVCTL_READRQ;
> > +				   PCI_EXP_DEVCTL_READRQ  |
> > +				   PCI_EXP_DEVCTL_EXT_TAG |
> > +				   PCI_EXP_DEVCTL_RELAX_EN;
> >  	hpx->pci_exp_devctl_or &= ~(PCI_EXP_DEVCTL_PAYLOAD |
> > -				    PCI_EXP_DEVCTL_READRQ);
> > +				    PCI_EXP_DEVCTL_READRQ  |
> > +				    PCI_EXP_DEVCTL_EXT_TAG |
> > +				    PCI_EXP_DEVCTL_RELAX_EN);
> >  
> >  	/* Initialize Device Control Register */
> >  	pcie_capability_clear_and_set_word(dev, PCI_EXP_DEVCTL,
> > -- 
> > 2.43.5
> >