Re: [PATCH 1/1] Fix int1 recursion when no perf_bp_event is registeredy

[Jeff Merkey]

This could be a problem down the road for KDB and KGDB as well since
they both directly  fiddle with DR7 directly and by-pass the perf
layer for late breakpoints.   At present they make certain to have a
perf bp pevent registered.

On 12/10/15, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> On Thu, 10 Dec 2015, Andy Lutomirski wrote:
>> On Thu, Dec 10, 2015 at 12:49 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx>
>> wrote:
>> > On Thu, 10 Dec 2015, Andy Lutomirski wrote:
>> >> On brief inspection, this smells like a microcode bug.  Can you send
>> >> /proc/cpuinfo output?
>> >>
>> >> If this is the issue, I'm not sure we want to be in the business of
>> >> working around localized microcode bugs and, if we do, then I think we
>> >> should explicitly detect the bug and log about it.
>> >
>> > I think we should handle such stuff gracefully. Yes, we should log it
>> > and we also should check what the contents of the debug registers are.
>> >
>> > If dr7 has a break point enabled w/o perf having one installed then we
>> > know that someone did a horrible hackery ....
>>
>> I mis-read this.  I don't think this is the microcode bug.
>>
>> Do we know what the actual problem is? Jeff, how did you trigger this?
>>
>> If it's lazy DR switching (which I haven't looked at the details of),
>> then it seems that this could just be triggered by some unfortunate
>> combination of perf config and context switching).  But, if so, then I
>> think that the proposed fix is wrong -- shouldn't we fix dr7 rather
>> than fudging RF to work around this instance of the problem?  After
>> all, if we're hitting this condition in a tight userspace loop, we're
>> going to destroy performance unless we fix dr7.
>
> Right, if dr7 contains a valid breakpoint and perf does not have one
> installed then either that lazy DR stuff is flaky or someone else
> fiddled with dr7. In both cases we want to fix dr7 and yell about it.
>
> If we neither have a breakpoint nor dr7 has one enabled then its
> something spurious and we certainly want to set RF so the machine can
> make progress.
>
> Thanks,
>
> 	tglx
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/