We only really have two options:
(1) Reject conflicting filters when scheduling events. At event open
time we have ot check whether an entire group is actually
schedulable.
(2) Merge filters when scheduling events, and then filter out
branches which don't match a particular event's filters when reading
the branches.
> > There are a number of cases of that shape given the set of configurable
> > filters. In theory it's possible to retain those in some cases, but I
> > don't think that the complexity is justified.
> >
> > Similarly, whenever kernel branches are recorded it's necessary to drop
> > the stale branches whenever branch recording is paused, as there's
> > necessarily a blackout period and hence a gap in the records.
>
> If we save BRBE record when a process is switched out and then restore
> the record when a process is switched in, should we can keep a decent
> branch record for performance profiling?
I cannot parse this question. What are you trying to suggest here?
> I understand it might be many things happen in the middle of a task
> switching or migration, but it is fine for not recording branches during
> the blackout period. The missed branch records are not very helpful for
> forming a flow for a profiled program itself, especially, if we
> consider we will optimize userspace program in many cases.
Just to be clear, you're talking about userspace specifically, right?
There are users that want a contiguous set of branches, which is what
hardware tries to guarantee, and that's what LBR tries to gaurantee
today, so I don't think that we can say gaps are always fine.
If we want a "give me as many arbitrary samples branches as possible,
with arbitrary potential gaps" option, I'm not necessarily opposed to
that, but I do not think that should be the default behaviour.
> > Do you think that you have a case where losing branches across rotation
> > *really* matters?
>
> I agreed that event rotation case might be rare and complex. Please see
> a comment below.
>
> > > For a context-switch case, we need to save and re-inject branch record.
> > > BRBE record sticks to a process context, no matter what events have been
> > > enabled.
> >
> > I had originally wanted to keep per-event records around, but it doesn't
> > work in all cases. One reason events get discarded at context-switch
> > time is that CPU-bound events can sample branches, and would
> > mis-attribute stale userspace branches to the wrong context when
> > switching tasks. There are explicit comments about this in
> > amd_pmu_brs_sched_task() and intel_pmu_lbr_sched_task().
> >
> > Given we discard records when reprogramming events, we *could* try to
> > preserve events in some cases, but I suspect that as with the rotation
> > case this'll be a lot of complexity for little gain. Note that as we
> > discard events when enabling the PMU, we'd throw some task-bound records
> > away anyway, and practically the gain would be limited to cpu-bound
> > records.
> >
> > Do you have a reason why you think we *must* keep events around?
>
> Here I am really concerned are cases when a process is preempted or
> migrated. The driver doesn't save and restore branch records for these
> cases, it just invalidates all records when a task is scheduled in.
>
> As a result, if an event overflow is close to context switch, it is
> likely to capture incomplete branch records. For a userspace-only
> tracing, it is risk to capture empty branch record after preemption
> and migrations.
Yes, and I was initially concerned about that, but is that really a
problem? So long as the event period doesn't *always* coincide with
preemption you'll get records from other samples anyway.
While I agree this is a theoretical concern, I don't think its of
practical importance. This discarding happens on x86 *today* with LBR,
and (AFAICT) there hav been no complaints. Note that the LBR logic to
save/restore records for task contexts is only done for user callstack
recording, which BRBE does not support.
Regardless, there's still the problematic interaction with event
rotation; you cannot save/restore the records safely here if events (and
the relevant filters) can change between the save and restore.
Discarding here is significantly simpler, and correct.
Mark.
