Re: [PATCH 10/11] tracing/perf: Fix lock events recursions in the fast path

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Feb 05, 2010 at 02:01:55PM +0100, Peter Zijlstra wrote:
> On Fri, 2010-02-05 at 13:12 +0100, Peter Zijlstra wrote:
> > On Fri, 2010-02-05 at 13:10 +0100, Peter Zijlstra wrote:
> > > On Fri, 2010-02-05 at 11:49 +0100, Ingo Molnar wrote:
> > > > 
> > > > > That said, I'm not at all happy about removing lockdep annotations to make 
> > > > > the tracer faster, that's really counter productive.
> > > > 
> > > > Are there no dynamic techniques that could be used here?
> > > > 
> > > > Lockdep obviously wants maximum instrumentation coverage - performance be 
> > > > damned.
> > > > 
> > > > Lock profiling/tracing/visualization wants the minimum subset of events it is 
> > > > interested in - everything else is unnecessary overhead.
> > > 
> > > Well, they could start by moving the tracepoint inside the lockdep
> > > recursion check.
> > 
> > IIRC the reason its now outside is that you'd loose tracepoint on
> > lockdep_off() usage, but having the tracer folks help on removing any
> > such usage is of course a good thing.
> > 
> > The usage thereof in nmi_enter() doesn't seem like a problem, since
> > you're not supposed to be using locks from nmi context anyway, more so,
> > I'd not be adverse to putting BUG_ON(in_nmi()) in every lockdep hook.
> 
> Another nasty side effect is that it (lockdep recursion) isn't IRQ aware
> in that we don't do any tracking for IRQ's that hit while we're doing
> lockdep. We can fix that using a recursion context like we did for perf,
> that would actually improve lockdep itself too.


Yep, I agree with you. With the lockdep recursion check fixed to be
subtle enough for that + the lock events under lockdep recursion checks,
it fixes the situation while keeping the lockdep coverage in perf
tracing path for other cases.

I will start by adressing this.

That said, I think this is good for a first step, but we can't continue
to force the lock events -> lockdep dependency in the long term. We
can't have a serious lock profiling if we are doomed to suffer the
slowness due to lockdep checks at the same time.

Sure we can continue to support having both, but I think we should also
think about a solution to handle lock events without it in the future.
That will require some minimal lockdep functionalities (keeping the
lockdep map, and class hashes).

Thanks.