Re: [PATCH RFC v7 00/23] DEPT(Dependency Tracker)

[Byungchul Park]

Thomas wrote:
> On Tue, Jan 17 2023 at 10:18, Boqun Feng wrote:
> > On Mon, Jan 16, 2023 at 10:00:52AM -0800, Linus Torvalds wrote:
> > > I also recall this giving a fair amount of false positives, are they all fixed?
> >
> > From the following part in the cover letter, I guess the answer is no?
> >	...
> >       6. Multiple reports are allowed.
> >       7. Deduplication control on multiple reports.
> >       8. Withstand false positives thanks to 6.
> >	...
> >
> > seems to me that the logic is since DEPT allows multiple reports so that
> > false positives are fitlerable by users?
> 
> I really do not know what's so valuable about multiple reports. They
> produce a flood of information which needs to be filtered, while a
> single report ensures that the first detected issue is dumped, which
> increases the probability that it can be recorded and acted upon.

Assuming the following 2 assumptions, you are right.

Assumption 1. There will be too many reports with the multi-report
	      support, like all the combination of dependencies between
	      e.g. in-irq and irq-enabled-context.

Assumption 2. The detection is matured enough so that it barely happens
	      to fix false onces to see true one which is not a big deal.

However, DEPT doesn't generate all the combination of irq things as
Lockdep does so we only see a few multi-reports even with the support,
and I admit DEPT hasn't matured enough yet because fine classification
is required anyway to suppress false alarms. That's why I introduced
multi-report support at least for now. IMHO, it'd be still useful even
if it's gonna report a few true ones at once w/o false ones some day.

> Filtering out false positives is just the wrong approach. Decoding
> dependency issues from any tracker is complex enough given the nature of
> the problem, so adding the burden of filtering out issues from a stream
> of dumps is not helpful at all. It's just a marketing gag.
> 
> > *	Instead of introducing a brand new detector/dependency tracker,
> >	could we first improve the lockdep's dependency tracker? I think
> >	Byungchul also agrees that DEPT and lockdep should share the
> >	same dependency tracker and the benefit of improving the
> >	existing one is that we can always use the self test to catch
> >	any regression. Thoughts?
> 
> Ack. If the internal implementation of lockdep has shortcomings, then we
> can expand and/or replace it instead of having yet another
> infrastructure which is not even remotely as mature.

Ultimately, yes. We should expand or replace it instead of having
another ultimately.

	Byungchul
> 
> Thanks,
> 
>         tglx