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

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Mon, Nov 12, 2018 at 07:30:50AM +0100, Ingo Molnar wrote:
> 
> * Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> 
> > On Mon, Nov 12, 2018 at 06:10:33AM +0100, Ingo Molnar wrote:
> > > 
> > > * Waiman Long <longman@redhat.com> wrote:
> > > 
> > > > On 11/10/2018 09:10 AM, Peter Zijlstra wrote:
> > > > > On Fri, Nov 09, 2018 at 09:04:12AM +0100, Ingo Molnar wrote:
> > > > >> 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.)
> > > > >
> > > > > Something like this is needed for objtool LTO support as well. I just
> > > > > dread the build time 'regressions' this will introduce :/
> > > > >
> > > > > The final link pass is already by far the most expensive part (as
> > > > > measured in wall-time) of building a kernel, adding more work there
> > > > > would really suck :/
> > > > 
> > > > I think the idea is to make objtool have the capability to do that. It
> > > > doesn't mean we need to turn it on by default in every build.
> > > 
> > > Yeah.
> > > 
> > > Also note that much of the objtool legwork would be on a per file basis 
> > > which is reasonably parallelized already. On x86 it's also already done 
> > > for every ORC build i.e. every distro build and the incremental overhead 
> > > from also extracting locking dependencies should be reasonably small.
> > > 
> > > The final search of the global graph would be serialized but still 
> > > reasonably fast as these are all 'class' level dependencies which are 
> > > much less numerous than runtime dependencies.
> > > 
> > > I.e. I think we are talking about tens of thousands of dependencies, not 
> > > tens of millions.
> > > 
> > > At least in theory. ;-)
> > 
> > Generating a unified call graph sounds very expensive (and very far
> > beyond what objtool can do today).
> 
> Well, objtool already goes through the instruction stream and recognizes 
> function calls - so it can in effect generate a stream of "function x 
> called by function y" data, correct?

Yeah, though it would be quite simple to get the same data with a simple
awk script at link time.

> >  Also, what about function pointers?
> 
> So maybe it's possible to enumerate all potential values for function 
> pointers with a reasonably simple compiler plugin and work from there?

I think this would be somewhere between very difficult and impossible to
do properly.  I can't even imagine how this would be implemented in a
compiler plugin.  But I'd love to be proven wrong on that.

> One complication would be function pointers encoded as opaque data 
> types...
> 
> > BTW there's another kernel static analysis tool which attempts to 
> > create such a call graph already: smatch.
> 
> It's not included in the kernel tree though and I'd expect tight coupling 
> (or at least lock-step improvements) between tooling and lockdep here.

Fair enough.  Smatch's call tree isn't perfect anyway, but I don't think
perfect is attainable.

-- 
Josh