Re: PCIe AER generates no interrupts on host (ZynqMP)

[Bjorn Helgaas <helgaas@kernel.org>]

On Fri, Jan 07, 2022 at 10:31:06PM +0100, Pali Rohár wrote:
> On Friday 07 January 2022 14:34:15 Bjorn Helgaas wrote:
> > On Fri, Jan 07, 2022 at 11:04:58AM +0100, Pali Rohár wrote:
> > > Hello! You asked me in another email for comments to this email, so I'm
> > > replying directly to this email...
> > > 
> > > On Tuesday 04 January 2022 10:02:18 Stefan Roese wrote:
> > > > Hi,
> > > > 
> > > > I'm trying to get the Kernel PCIe AER infrastructure to work on my
> > > > ZynqMP based system. E.g. handle the events (correctable, uncorrectable
> > > > etc). In my current tests, no AER interrupt is generated though. I'm
> > > > currently using the "surprise down error status" in the uncorrectable
> > > > error status register of the connected PCIe switch (PLX / Broadcom
> > > > PEX8718). Here the bit is correctly logged in the PEX switch
> > > > uncorrectable error status register but no interrupt is generated
> > > > to the root-port / system. And hence no AER message(s) reported.
> > 
> > I think the error should also be logged in the Root Port AER
> > Capability.  And of course the interrupt enable bits in the Root Error
> > Command register would have to be set.
> > 
> > > > Does any one of you have some ideas on what might be missing? Why are
> > > > these events not reported to the PCIe rootport driver via IRQ? Might
> > > > this be a problem of the missing MSI-X support of the ZynqMP? The AER
> > > > interrupt is connected as legacy IRQ:
> > > > 
> > > > cat /proc/interrupts | grep -i aer
> > > >  58:          0          0          0          0  nwl_pcie:legacy   0 Level
> > > > PCIe PME, aerdrv
> > 
> > I guess this means whatever INTx the Root Port is using is connected
> > to IRQ 58?  Can you tell whether that INTx works if a device below the
> > Root Port uses it?  Or whether it is asserted for PMEs?
> > 
> > > Error events (correctable, non-fatal and fatal) are reported by PCIe
> > > devices to the Root Complex via PCIe error messages (Message code of TLP
> > > is set to Error Message) and not via interrupts. Root Port is then
> > > responsible to "convert" these PCIe error messages to MSI(X) interrupt
> > > and report it to the system. According to PCIe spec, AER is supported
> > > only via MSI(X) interrupts, not legacy INTx.
> > 
> > Where does it say that?  PCIe r5.0, sec 6.2.4.1.2 and 6.2.6, both
> > mention INTx, and the diagram in 6.2.6 even shows possible
> > platform-specific System Error signaling.
> 
> Kernel AER driver is not available when MSI is not supported:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/pci/pcie/aer.c?h=v5.15#n112
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/pci/pcie/portdrv_core.c?h=v5.15#n224
> Originally this was my primary indication that AER is MSI(X)-only.

I think that is broken.  Looks like it was added by 3e77a3f7895e
("PCI: Disable AER with pci=nomsi"), which says that with "pci=nomsi",
we see lost interrupts and lockdep inversions.

I don't know any more of the history behind that, but I suspect that
turning off AER in that case just covered up some other Linux issue.

> And my understanding is that AER Root Error Status Register (7.8.4.10)
> specifies Advanced Error Interrupt Message Number which indicates which
> MSI(X) interrupt is used. And there is no information about INTx if you
> enable particular reporting category via AER Root Error Command Register.
> That is why I was in impression that AER interrupts are MSI-only.
> 
> But now I'm looking at 6.2.4.1.2 section and seems that AER can really
> use INTx. So I was wrong here.
> 
> But why then kernel AER driver has check that AER is available only when
> MSI is enabled? And not available when MSI is disabled?

Apart from the issue behind 3e77a3f7895e, I think this is just because
the intersection of platforms that lack MSI and people with enough
interest in AER is small.

The interrupt configuration in portdrv is nasty.  Although it looks
like pcie_init_service_irqs() might now be smart enough to use the
legacy INTx when MSI is not available.

Bjorn