Re: [RFC PATCH 00/12] locking/lockdep: Add a new class of terminal locks

From: Ingo Molnar
Date: Mon Nov 12 2018 - 00:15:50 EST



* Waiman Long <longman@xxxxxxxxxx> wrote:

> > Could you please measure a locking intense workload instead, such as:
> >
> > $ perf stat --null --sync --repeat 10 perf bench sched messaging
> >
> > and profile which locks used there could be marked terminal, and measure
> > the before/after performance impact?
>
> I will run the test. It will probably be done after the LPC next week.

Thanks!

> >> Below were selected output lines from the lockdep_stats files of the
> >> patched and unpatched kernels after bootup and running parallel kernel
> >> builds.
> >>
> >> Item Unpatched kernel Patched kernel % Change
> >> ---- ---------------- -------------- --------
> >> direct dependencies 9732 8994 -7.6%
> >> dependency chains 18776 17033 -9.3%
> >> dependency chain hlocks 76044 68419 -10.0%
> >> stack-trace entries 110403 104341 -5.5%
> > That's pretty impressive!
> >
> >> There were some reductions in the size of the lockdep tables. They were
> >> not significant, but it is still a good start to rein in the number of
> >> entries in those tables to make it harder to overflow them.
> > Agreed.
> >
> > BTW., if you are interested in more radical approaches to optimize
> > lockdep, we could also add a static checker via objtool driven call graph
> > analysis, and mark those locks terminal that we can prove are terminal.
> >
> > This would require the unified call graph of the kernel image and of all
> > modules to be examined in a final pass, but that's within the principal
> > scope of objtool. (This 'final pass' could also be done during bootup, at
> > least in initial versions.)
> >
> > Note that beyond marking it 'terminal' such a static analysis pass would
> > also allow the detection of obvious locking bugs at the build (or boot)
> > stage already - plus it would allow the disabling of lockdep for
> > self-contained locks that don't interact with anything else.
> >
> > I.e. the static analysis pass would 'augment' lockdep and leave only
> > those locks active for runtime lockdep tracking whose dependencies it
> > cannot prove to be correct yet.
>
> It is a pretty interesting idea to use objtool to scan for locks. The
> list of locks that I marked as terminal in this patch was found by
> looking at /proc/lockdep for those that only have backward dependencies,
> but no forward dependency. I focused on those with a large number of BDs
> and check the code to see if they could marked as terminal. This is a
> rather labor intensive process and is subject to error.

Yeah.

> [...] It would be nice if it can be done by an automated tool. So I am
> going to look into that, but it won't be part of this initial patchset,
> though.

Of course!

> I sent this patchset out to see if anyone has any objection to it. It
> seems you don't have any objection to that. So I am going to move ahead
> to do more testing and performance analysis.

The one worry I have is that this interim solution removes the benefit of
a proper static analysis method.

But if you promise to make a serious effort on the static analysis
tooling as well (which should have awesome performance results and
automate the manual markup), then I have no fundamental objections to the
interim approach either.

If static analysis works as well as I expect it to then in principle we
might even be able to have lockdep enabled in production kernels: it
would only add overhead to locks that are overly complex - which would
create incentives to improve those dependencies.

Thanks,

Ingo