Re: aer_inject vs. apei/einj

[Jean Delvare <jdelvare@suse.de>]

On Wed, 17 Feb 2016 11:03:31 -0600, Bjorn Helgaas wrote:
> [+cc Huang, author of both aer_inject and apei/einj]
> 
> On Wed, Feb 17, 2016 at 7:33 AM, Jean Delvare <jdelvare@suse.de> wrote:
> > Hi all,
> >
> > I am looking for some guidance regarding AER testing. I see that we
> > have two different drivers for error injection in the kernel:
> > aer_inject and apei/einj. The user-space aer-inject tool seems to only
> > care about the former.
> >
> > How does one know which driver should be used on a given system? I
> > suppose that only one of them will work on a given system?
> >
> > My impression is that aer_inject is for "native" AER handling while
> > apei/einj is for ACPI-driven AER. Is it correct? If not I would
> > appreciate some pointers explaining when aer_inject should be used and
> > when apei/einj should be used.
> 
> My understanding is that:
> 
>   - aer_inject does not actually write to any hardware registers
> itself (though I do see it writes to some masks).  It works by
> replacing the PCI config accessors with new ones that make it look
> like the AER registers have errors logged.
> 
>   - apei/einj runs ACPI methods that apparently seed errors.  These
> might use hardware support for seeding errors, which would of course
> be platform-dependent.
> 
> So aer_inject should work on any system at all.  I think apei/einj

My problem is precisely that aer_inject doesn't work on any system I
tested. Either the device doesn't support AER, or its root port doesn't
support AER, or (further I've been) the "error device" of the root port
doesn't exist. I am not too familiar with PCIe but apparently PCIe
devices can have "sub-devices" which do not show in "lspci" but show up
in /sys/bus/pci_express/devices. I have yet to see an aer sub-device
there on any of my systems.

> will only work if the platform supplies an EINJ table, and even when
> it does, I suspect different platforms probably have different
> injection capabilities.
> 
> Huang probably can give a much better response.

Huang, pleeeease? :)

-- 
Jean Delvare
SUSE L3 Support