Re: [PATCHv2 2/4] coresight: tmc-etf: Fix NULL ptr dereference in tmc_enable_etf_sink_perf()

[Suzuki Poulose]

On 10/22/20 4:06 PM, Peter Zijlstra wrote:
> On Thu, Oct 22, 2020 at 02:30:21PM +0100, Suzuki Poulose wrote:
> > On 10/22/20 12:32 PM, Peter Zijlstra wrote:
> > > On Thu, Oct 22, 2020 at 04:27:52PM +0530, Sai Prakash Ranjan wrote:
> > >
> > > > Looking at the ETR and other places in the kernel, ETF and the
> > > > ETB are the only places trying to dereference the task(owner)
> > > > in tmc_enable_etf_sink_perf() which is also called from the
> > > > sched_in path as in the call trace.
> > >
> > > > @@ -391,6 +392,10 @@ static void *tmc_alloc_etf_buffer(struct coresight_device *csdev,
> > > >    {
> > > >    	int node;
> > > >    	struct cs_buffers *buf;
> > > > +	struct task_struct *task = READ_ONCE(event->owner);
> > > > +
> > > > +	if (!task || is_kernel_event(event))
> > > > +		return NULL;
> > >
> > >
> > > This is *wrong*... why do you care about who owns the events?
> >
> > This is due to the special case of the CoreSight configuration, where
> > a "sink" (where the trace data is captured) is shared by multiple Trace
> > units. So, we could share the "sink" for multiple trace units if they
> > are tracing the events that belong to the same "perf" session. (The
> > userspace tool could decode the trace data based on the TraceID
> > in the trace packets). Is there a better way to do this ?
>
> I thought we added sink identification through perf_event_attr::config2
> ?

Correct. attr:config2 identifies the "sink" for the collection. But,
that doesn't solve the problem we have here. If two separate perf
sessions use the "same sink", we don't want to mix the
trace data into the same sink for events from different sessions.

Thus, we need a way to check if a new event starting the tracing on
an ETM belongs to the same session as the one already pumping the trace
into the sink.

We use event->owner pid for this check and thats where we encountered
a NULL event->owner. Looking at the code further, we identified that
kernel events could also trigger this issue.

Suzuki