On Sat, Jan 9, 2016 at 9:45 AM, Tony Luck <tony.luck(a)gmail.com> wrote:
On Fri, Jan 8, 2016 at 5:52 PM, Andy Lutomirski
> Also, I think it would be nicer if the machine check code would invoke
> the handler regardless of which handler (or class) is selected. Then
> the handlers that don't want to handle #MC can just reject them.
The machine check code is currently a two pass process.
First we scan all the machine check banks (on all processors
at the moment because machine checks are broadcast). We
assess the severity of all errors found.
Then we take action. Panic if the most severe error was fatal,
recover if not.
This patch series tweaks the severity calculation. In-kernel
errors at IPs with a EXTABLE_CLASS_FAULT handler are
now ranked as recoverable. All other kernel errors remain
I don't think it is right to unconditionally execute the fix code in the
severity assessment phase.
I would argue that unconditionally calling the handler would be
cleaner. The handler would return 0 or false to indicate that it
refuses to fix the exception.
This is similar to the logic that, for regular user memory access, we
shouldn't fix up faults other than #PF. Given that we're adding
flexible handler callbacks, lets push all the "is this an acceptable
fault to fix up" down into the callback. Does that make sense?
Perhaps later we can revisit the two pass process?
Oh, I see. Is it the case that the MC code can't cleanly handle the
case where the error was nominally recoverable but the kernel doesn't
know how to recover from it due to the lack of a handler that's okay
with it, because the handler's refusal to handle the fault wouldn't be
known until too late?