On 10/23/20 10:41 AM, Peter Zijlstra wrote:
On Fri, Oct 23, 2020 at 09:49:53AM +0100, Suzuki Poulose wrote:
On 10/23/20 8:39 AM, Peter Zijlstra wrote:
So then I don't understand the !->owner issue, that only happens when
the task dies, which cannot be concurrent with event creation. Are you
Part of the patch from Sai, fixes this by avoiding the dereferencing
after event creation (by caching it). But the kernel events needs
fixing.
I'm fundamentally failing here. Creating a link to the sink is strictly
event-creation time. Why would you ever need it again later? Later you
already have the sink setup.
Sorry for the lack of clarity here, and you are not alone unless you
have drowned in the CoreSight topologies ;-)
Typically current generation of systems have the following topology :
CPU0
etm0 \
\ ________
/ \
CPU1 / \
etm1 \
\
/------- sink0
CPU2 /
etm2 \ /
\ ________ /
/
CPU3 /
etm3
i.e, Multiple ETMs share a sink. [for the sake of simplicity, I have
used one sink. Even though there could be potential sinks (of different
types), none of them are private to the ETMs. So, in a nutshell, a sink
can be reached by multiple ETMs. ]
Now, for a session :
perf record -e cs_etm/sinkid=sink0/u workload
We create an event per CPU (say eventN, which are scheduled based on the
threads that could execute on the CPU. At this point we have finalized
the sink0, and have allocated necessary buffer for the sink0.
Now, when the threads are scheduled on the CPUs, we start the
appropriate events for the CPUs.
e.g,
CPU0 sched -> workload:0 - > etm0->event0_start -> Turns all
the components upto sink0, starting the trace collection in the buffer.
Now, if another CPU, CPU1 starts tracing event1 for workload:1 thread,
it will eventually try to turn ON the sink0.Since sink0 is already
active tracing event0, we could allow this to go through and collect
the trace in the *same hardware buffer* (which can be demuxed from the
single AUX record using the TraceID in the packets). Please note that
we do double buffering and hardware buffer is copied only when the sink0
is stopped (see below).
But, if the event scheduled on CPU1 doesn't belong to the above session,
but belongs to different perf session
(say, perf record -e cs_etm/sinkid=sink0/u benchmark),
we can't allow this to succeed and mix the trace data in to that of
workload and thus fail the operation.
In a nutshell, since the sinks are shared, we start the sink on the
first event and keeps sharing the sink buffer with any event that
belongs to the same session (using refcounts). The sink is only released
for other sessions, when there are no more events in the session tracing
on any of the ETMs.
I know this is fundamentally a topology issue, but that is not something
we can fix. But the situation is changing and we are starting to see
systems with per-CPU sinks.
Hope this helps.
Suzuki
