Re: KCSAN: data-race in __alloc_file / __alloc_file

["Paul E. McKenney" <paulmck@kernel.org>]

On Mon, Nov 11, 2019 at 04:10:20PM +0100, Marco Elver wrote:
> On Mon, 11 Nov 2019 at 15:31, Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > On Mon, Nov 11, 2019 at 03:17:51PM +0100, Marco Elver wrote:
> > > On Sun, 10 Nov 2019 at 21:44, Paul E. McKenney <paulmck@kernel.org> wrote:
> > > >
> > > > On Sun, Nov 10, 2019 at 11:20:53AM -0800, Linus Torvalds wrote:
> > > > > On Sun, Nov 10, 2019 at 11:12 AM Linus Torvalds
> > > > > <torvalds@linux-foundation.org> wrote:
> > > > > >
> > > > > > And this is where WRITE_IDEMPOTENT would make a possible difference.
> > > > > > In particular, if we make the optimization to do the "read and only
> > > > > > write if changed"
> > > > >
> > > > > It might be useful for checking too. IOW, something like KCSAN could
> > > > > actually check that if a field has an idempotent write to it, all
> > > > > writes always have the same value.
> > > > >
> > > > > Again, there's the issue with lifetime.
> > > > >
> > > > > Part of that is "initialization is different". Those writes would not
> > > > > be marked idempotent, of course, and they'd write another value.
> > > > >
> > > > > There's also the issue of lifetime at the _end_ of the use, of course.
> > > > > There _are_ interesting data races at the end of the lifetime, both
> > > > > reads and writes.
> > > > >
> > > > > In particular, if it's a sticky flag, in order for there to not be any
> > > > > races, all the writes have to happen with a refcount held, and the
> > > > > final read has to happen after the final refcount is dropped (and the
> > > > > refcounts have to have atomicity and ordering, of course). I'm not
> > > > > sure how easy something like that is model in KSAN. Maybe it already
> > > > > does things like that for all the other refcount stuff we do.
> > > > >
> > > > > But the lifetime can be problematic for other reasons too - in this
> > > > > particular case we have a union for that sticky flag (which is used
> > > > > under the refcount), and then when the final refcount is released we
> > > > > read that value (thus no data race) but because of the union we will
> > > > > now start using that field with *different* data. It becomes that RCU
> > > > > list head instead.
> > > > >
> > > > > That kind of "it used to be a sticky flag, but now the lifetime of the
> > > > > flag is over, and it's something entirely different" might be a
> > > > > nightmare for something like KCSAN. It sounds complicated to check
> > > > > for, but I have no idea what KCSAN really considers complicated or
> > > > > not.
> > > >
> > > > But will "one size fits all" be practical and useful?
> > > >
> > > > For my code, I would be happy to accept a significant "false positive"
> > > > rate to get even a probabilistic warning of other-task accesses to some
> > > > of RCU's fields.  Even if the accesses were perfect from a functional
> > > > viewpoint, they could be problematic from a performance and scalability
> > > > viewpoint.  And for something like RCU, real bugs, even those that are
> > > > very improbable, need to be fixed.
> > > >
> > > > But other code (and thus other developers and maintainers) are going to
> > > > have different needs.  For all I know, some might have good reasons to
> > > > exclude their code from KCSAN analysis entirely.
> > > >
> > > > Would it make sense for KCSAN to have per-file/subsystem/whatever flags
> > > > specifying the depth of the analysis?
> > >
> > > Just to answer this: we already have this, and disable certain files
> > > already. So it's an option if required. Just need maintainers to add
> > > KCSAN_SANITIZE := n, or KCSAN_SANITIZE_file.o := n to Makefiles, and
> > > KCSAN will simply ignore those.
> > >
> > > FWIW we now also have a config option to "ignore repeated writes with
> > > the same value". It may be a little overaggressive/imprecise in
> > > filtering data races, but anything else like the super precise
> > > analysis involving tracking lifetimes and values (and whatever else
> > > the rules would require) is simply too complex. So, the current
> > > solution will avoid reporting cases like the original report here
> > > (__alloc_file), but at the cost of maybe being a little imprecise.
> > > It's probably a reasonable trade-off, given that we have too many data
> > > races to deal with on syzbot anyway.
> >
> > Nice!
> >
> > Is this added repeated-writes analysis something that can be disabled?
> > I would prefer that the analysis of RCU complain in this case as a
> > probabilistic cache-locality warning.  If it can be disabled, please
> > let me know if there is anything that I need to do to make this happen.
> 
> It's hidden behind a Kconfig config option, and actually disabled by
> default. We can't enable/disable this on a per-file basis.
> 
> Right now, we'll just enable it on the public syzbot instance, which
> will use the most conservative config.  Of course you can still run
> your own fuzzer/stress test of choice with KCSAN and the option
> disabled. Is that enough?
> 
> Otherwise I could also just say if the symbolized top stack frame
> contains "rcu_", don't ignore -- which would be a little hacky and
> imprecise though. What do you prefer?

Tough question.  ;-)

We could try the "rcu_" trick, and if it doesn't cause problems for
others, why not?

						Thanx, Paul