Re: [PATCH v1 4/6] perf: Allow using AUX data in perf samples

[Alexander Shishkin]

Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:

> On Tue, Jun 19, 2018 at 01:47:25PM +0300, Alexander Shishkin wrote:
>> Right, the SW stuff may then race with event_function_call() stuff. Hmm.
>> For the HW stuff, I'm hoping that some kind of a sleight of hand may
>> suffice. Let me think some more.
>
> I currently don't see how the SW driven snapshot can ever work, see my
> comment on the last patch.

Yes, this should be solved by grouping, similarly PEBS->PT.

>> > Why can't you just snapshot the current location and let the thing
>> > 'run' ?
>> 
>> Because the buffer will overwrite itself and the location will be useless.
>
> Not if it's large enough ;-)
>
>> We don't write the AUX data out in this 'mode' at all, only the samples,
>> which allows for much less data in the resulting perf.data, less work for
>> the consumer, less IO bandwidth etc, and as a bonus, no AUX-related
>> interrupts.
>> 
>> But actually, even to snapshot the location we need to stop the event.
>
> Maybe new methods that should only be called from NMI context? 

Right, I went with that and added a ->snapshot_aux() method that can't
change any states or do anything that would affect the operation of
potentially preempted in-IRQ callbacks. So, if NMI hits in the middle of
->stop() or whatnot, we should be fine. Some context: in AUX overwrite
mode (aka "snapshot mode") we don't get PT related NMIs, the only NMIs
are from the HW events for which we're writing samples. We use the cpu
local ->handle_nmi (a kind of misnomer) to tell us if the NMI hit
between ->start() and ->stop() and we can safely take the sample.

The other problem is sampling SW events, that would require a ctx->lock
to prevent racing with event_function_call()s from other cpus, resulting
in somewhat cringy "if (!in_nmi()) raw_spin_lock(...)", but I don't have
better idea as to how to handle that.

The whole thing is squashed into one patch below for convenience.