Re: [PATCH 1/3] x86/quirks: Scan all busses for early PCI quirks

[Guilherme G. Piccoli]

On 23/10/2018 14:03, Bjorn Helgaas wrote:
> On Mon, Oct 22, 2018 at 05:35:06PM -0300, Guilherme G. Piccoli wrote:
>> On 18/10/2018 19:15, Bjorn Helgaas wrote:
>>> On Thu, Oct 18, 2018 at 03:37:19PM -0300, Guilherme G. Piccoli wrote:
>>> [...] 
>> I understand your point, but I think this is inherently an architecture
>> problem. No matter what solution we decide for, it'll need to be applied
>> in early boot time, like before the PCI layer gets initialized.
> 
> This is the part I want to know more about.  Apparently there's some
> event X between early_quirks() and the PCI device enumeration, and we
> must disable MSIs before X:
> 
>   setup_arch()
>       early_quirks()                     # arch/x86/kernel/early-quirks.c
>       early_pci_clear_msi()
>   ...
>   X
>   ...
>   pci_scan_root_bus_bridge()
>     ...
>     DECLARE_PCI_FIXUP_EARLY              # drivers/pci/quirks.c
> 
> I want to know specifically what X is.  If we don't know what X is and
> all we know is "we have to disable MSIs earlier than PCI init", then
> we're likely to break things again in the future by changing the order
> of disabling MSIs and whatever X is.
> 
> Bjorn
> 

Hi Bjorn (and all CCed), I'm sorry to necro-bump a thread >2 years
later, but recent discussions led to a better understanding of this 'X'
point, thanks to Thomas!

For those that deleted this thread from their email clients, it's
available in [0] - the summary is that we faced an IRQ storm really
early in boot, due to a bogus PCIe device MSI behavior, when booting a
kdump kernel. This led the machine to get stuck in the boot and we
couldn't kdump. The solution hereby proposed is to clear MSI interrupts
early in x86, if a parameter is provided. I don't have the reproducer
anymore and it was pretty hard to reproduce in virtual environments.

So, about the 'X' Bjorn, in another recent thread about IRQ storms [1],
Thomas clarified that and after a brief discussion, it seems there's no
better way to prevent the MSI storm other than clearing the MSI
capability early in boot. As discussed both here and in thread [1], this
is indeed a per-architecture issue (powerpc is not subject to that, due
to a better FW reset mechanism), so I think we still could benefit in
having this idea implemented upstream, at least in x86 (we could expand
to other architectures if desired, in the future).

As a "test" data point, this was implemented in Ubuntu (same 3 patches
present in this series) for ~2 years and we haven't received bug reports
- I'm saying that because I understand your concerns about expanding the
early PCI quirks scope.

Let me know your thoughts. I'd suggest all to read thread [1], which
addresses a similar issue but in a different "moment" of the system boot
and provides some more insight on why the early MSI clearing seems to
make sense.

Thanks,


Guilherme


[0]
https://lore.kernel.org/linux-pci/20181018183721.27467-1-gpiccoli@xxxxxxxxxxxxx

[1] https://lore.kernel.org/lkml/87y2js3ghv.fsf@xxxxxxxxxxxxxxxxxxxxxxx