Re: [PATCH] powerpc/eeh: Validate arch in eeh_add_device_early()

On Wed, 2016-01-13 at 10:08 -0200, Guilherme G. Piccoli wrote:

On 01/13/2016 08:38 AM, Michael Ellerman wrote:

quoted

But eeh_enabled() is still false? That seems like it's liable to cause breakage
elsewhere.

Yes, eeh_enabled() is false as expected. Notice that eeh_enabled() is
telling if EEH is enabled or not, and since it's not (because there's no
PCI adapters on machine yet!), makes sense it returns false.

At the end of runs of eeh_add_device_early(), the devices are probed and
EEH is enabled, so the function will reflect this. Notice that the
problem addressed by this patch is the use of eeh_enabled() in hotplug
operations only, which in my opinion is not correct. I checked every
other use of eeh_enable() in the code, and they all seems appropriate,
so I expect no errors at all.

OK.

quoted

Shouldn't the PCI hotplug code instead be taught to initialise EEH correctly
when the first device is added?

Well, for sure there are other ways to achieve this patch's goal, like
refactor PCI hotplug code in lots of places. But notice although the
proposed solution is simple, it solves the issue because
eeh_add_device_early() is ultimately the function used by PCI hotplug
mechanism (no matter if pseries/cell/etc or the type of hotplug) to
initialize EEH by probing the devices at the moment they are added to
the system. In my opinion, this is exactly the location we want to
change code to address this issue.

What do you think? Thanks very much for the review.

I'm still not sure. I'm certainly happy for the fix to be simple, it will need
to be backported after all.

But for example what happens if the user boots with eeh=off on the command
line, and then hotplugs a device. It looks like because you're not using
eeh_enabled() you will incorrectly initialise EEH anyway?

cheers