[RFC 09/11] coresight: etm-perf: Disable the path before capturing the trace data

[Mathieu Poirier]

On Fri, Nov 27, 2020 at 10:32:28AM +0000, Suzuki K Poulose wrote:
> On 11/10/20 12:45 PM, Anshuman Khandual wrote:
> > perf handle structure needs to be shared with the TRBE IRQ handler for
> > capturing trace data and restarting the handle. There is a probability
> > of an undefined reference based crash when etm event is being stopped
> > while a TRBE IRQ also getting processed. This happens due the release
> > of perf handle via perf_aux_output_end(). This stops the sinks via the
> > link before releasing the handle, which will ensure that a simultaneous
> > TRBE IRQ could not happen.
> 
> Or in other words :
> 
> We now have :
> 
> 	update_buffer()
> 
> 	perf_aux_output_end(handle)
> 
> 	...
> 	disable_path()
> 
> This is problematic due to various reasons :
> 
> 1) The semantics of update_buffer() is not clear. i.e, whether it
>    should leave the "sink" "stopped" or "disabled" or "active"

I'm a little confused by the above as the modes that apply here are
CS_MODE_DISABLED and CS_MODE_PERF, so I'll go with those.  Let me know if you
meant something else.

So far ->update_buffer() doesn't touch drvdata->mode and as such it is still set
to CS_MODE_PERF when the update has completed. 

> 
> 2) This breaks the recommended trace collection sequence of
>    "flush" and "stop" from source to the sink for trace collection.
>     i.e, we stop the source now. But don't flush the components
>     from source to sink, rather we stop and flush from the sink.
>     And we flush and stop the path after we have collected the
>     trace data at sink, which is pointless.

The above assesment is correct.  Fixing it though has far reaching ramifications
that go far beyond the scope of this patch.   

> 
> 3) For a sink with IRQ handler, if we don't stop the sink with
>    update_buffer(), we could have a situation :
> 
>    update_buffer()
> 
>    perf_aux_outpuf_end(handle) # handle is invalid now
> 
>  -----------------> IRQ    -> irq_handler()
>                                perf_aux_output_end(handle) # Wrong !
> 
> 
>    disable_path()

That's the picture of the issue I had in my head when looking at the code -
I'm glad we came to the same conclusion.

> 
> The sysfs mode is fine, as we defer the trace collection to disable_path().
> 
> The proposed patch is still racy, as we could still hit the problem.
> 
> So, to avoid all of these situations, I think we should defer the the
> update_buffer() to sink_ops->disable(), when we have flushed and stopped
> the all the components upstream and avoid any races with the IRQ
> handler.
> 
> i.e,
> 
> 	source_ops->stop(csdev);
> 
> 	disable_path(handle); // similar to the enable_path
> 
> 
> sink_ops->disable(csdev, handle)
> {
>   /* flush & stop */
> 
>   /* collect trace */
>   perf_aux_output_end(handle, size);
> }

That is one solution.  The advantage here is that it takes care of the
flusing problem you described above.  On the flip side it is moving a lot of
code around, something that is better to do in another set.

Another solution is to disable the TRBE IRQ in ->udpate_buffer().  The ETR does
the same kind of thing with tmc_flush_and_stop().  I don't know how feasible
that is but it would be a simple solution for this set.  Properly flushing the
pipeline could be done later.  I'm fine with either approach.

Thanks,
Mathieu 

> 
> 
> Kind regards
> Suzuki
> 
> 
> 
> > 
> > Signed-off-by: Anshuman Khandual <anshuman.khandual at arm.com>
> > ---
> > This might cause problem with traditional sink devices which can be
> > operated in both sysfs and perf mode. This needs to be addressed
> > correctly. One option would be to move the update_buffer callback
> > into the respective sink devices. e.g, disable().
> > 
> >   drivers/hwtracing/coresight/coresight-etm-perf.c | 2 ++
> >   1 file changed, 2 insertions(+)
> > 
> > diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
> > index 534e205..1a37991 100644
> > --- a/drivers/hwtracing/coresight/coresight-etm-perf.c
> > +++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
> > @@ -429,7 +429,9 @@ static void etm_event_stop(struct perf_event *event, int mode)
> >   		size = sink_ops(sink)->update_buffer(sink, handle,
> >   					      event_data->snk_config);
> > +		coresight_disable_path(path);
> >   		perf_aux_output_end(handle, size);
> > +		return;
> >   	}
> >   	/* Disabling the path make its elements available to other sessions */
> > 
>