Re: Question: Clearing error bits in the root port post enumeration

[Bjorn Helgaas <helgaas@kernel.org>]

On Tue, Nov 07, 2023 at 08:44:53AM +0530, Vidya Sagar wrote:
> On 11/3/2023 11:50 PM, Bjorn Helgaas wrote:
> > On Tue, Oct 31, 2023 at 12:26:31PM +0000, Vidya Sagar wrote:
> > > Hi folks,
> > > 
> > > I would like to know your comments on the following scenario where
> > > we are observing the root port logging errors because of the
> > > enumeration flow being followed.
> > > 
> > > DUT information:
> > > - Has a root port and an endpoint connected to it
> > > - Uses ECAM mechanism to access the configuration space
> > > - Booted through ACPI flow
> > > - Has a Firmware-First approach for handling the errors
> > > - System is configured to treat Unsupported Requests as
> > >    AdvisoryNon-Fatal errors
> > > 
> > > As we all know, when a configuration read request comes in for a
> > > device number that is not implemented, a UR would be returned as per
> > > the PCIe spec.
> > > 
> > > As part of the enumeration flow on DUT, when the kernel reads offset
> > > 0x0 of B:D:F=0:0:0, the root port responds with its valid Vendor-ID
> > > and Device-ID values.  But, when B:D:F=0:1:0 is probed, since there
> > > is no device present there, the root port responds with an
> > > Unsupported Request and simultaneously logs the same in the Device
> > > Status register (i.e. bit-3).  Because of it, there is a UR logged
> > > in the Device Status register of the RP by the time enumeration is
> > > complete.
> > > 
> > > In the case of AER capability natively owned by the kernel, the AER
> > > driver's init call would clear all such pending bits.
> > > 
> > > Since we are going with the Firmware-First approach, and the system
> > > is configured to treat Unsupported Requests as AdvisoryNon-Fatal
> > > errors, only a correctable error interrupt can be raised to the
> > > Firmware which takes care of clearing the corresponding status
> > > registers.  The firmware can't know about the UnsupReq bit being set
> > > as the interrupt it received is for a correctable error hence it
> > > clears only bits related to correctable error.
> > > 
> > > All these events leave a freshly booted system with the following
> > > bits set.
> > > 
> > > Secondary status: 66MHz- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- <MAbort+ <SERR- <PERR-          (MAbort)
> > > DevSta: CorrErr- NonFatalErr- FatalErr- UnsupReq+ AuxPwr- TransPend-                                                              (UnsupReq)
> > > UESta:  DLP- SDES- TLP- FCP- CmpltTO- CmpltAbrt- UnxCmplt- RxOF- MalfTLP- ECRC- UnsupReq+ ACSViol-   (UnsupReq)
> > > 
> > > Since the reason for UR is well understood at this point, I would
> > > like to weigh in on the idea of clearing the aforementioned bits in
> > > the root port once the enumeration is done particularly to cater to
> > > the configurations where Firmware-First approach is in place.
> > > Please let me know your comments on this approach.
> > 
> > I think Secondary status (PCI_SEC_STATUS) is always owned by the OS
> > and is not affected by _OSC negotiation, right?  Linux does basically
> > nothing with that today, but I think it *could* clear the "Received
> > Master Abort" bit.
>
> Yes. PCI_SEC_STATUS is always owned by the OS and _OSC negotiation doesn't
> really affect that.
> 
> > I'm not very familiar with Advisory Non-Fatal errors.  I'm curious
> > about the UESta situation: why can't firmware know about UnsupReq
> > being set?  I assume PCI_ERR_COR_ADV_NFAT is the Correctable Error
> > Status bit the firmware *does* see and clear.
>
> Yes, PCI_ERR_COR_ADV_NFAT is indeed cleared by the firmware.
> > 
> > But isn't the whole point of Advisory Non-Fatal errors that an error
> > that is logged as an Uncorrectable Error and that normally would be
> > signaled with ERR_NONFATAL is signaled with ERR_COR instead?  So
> > doesn't PCI_ERR_COR_ADV_NFAT being set imply that some
> > PCI_ERR_UNCOR_STATUS must be set as well?  If so, I would think
> > firmware *could* figure that out and clear the PCI_ERR_UNCOR_STATUS
> > bit.
>
> So, are you suggesting that let the firmware only clear the
> PCI_ERR_UNCOR_STATUS also?

In this firmware-first scenario, I'm assuming the platform retained
ownership of the AER capability, so I would think firmware certainly
should be allowed to clear PCI_ERR_UNCOR_STATUS.

> if so, then, I can even make the firmware clear the PCI_SEC_STATUS
> also thereby leaving the firmware responsible for clearing all the
> error bits. Does that sound ok?

It doesn't sound quite right to me for firmware to clear
PCI_SEC_STATUS because it doesn't own that register.  I suspect we
would probably see the "Received Master Abort" bit set after
enumeration even on Conventional PCI systems, so I doubt this is
anything specific to PCIe or AER, and maybe Linux should clear it
after enumerating devices below the bridge.

Bjorn