Re: KCSAN: data-race in __alloc_file / __alloc_file

[Marco Elver]

On Fri, 8 Nov 2019 at 19:40, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Fri, Nov 8, 2019 at 10:16 AM Marco Elver <elver@xxxxxxxxxx> wrote:
> >
> > KCSAN does not use volatile to distinguish accesses. Right now
> > READ_ONCE, WRITE_ONCE, atomic bitops, atomic_t (+ some arch specific
> > primitives) are treated as marked atomic operations.
>
> Ok, so we'd have to do this in terms of ATOMIC_WRITE().
>
> One alternative might be KCSAN enhancement, where you notice the
> following pattern:
>
>  - a field is initialized before the data structure gets exposed (I
> presume KCSAN already must understand about this issue -
> initializations are different and not atomic)
>
>  - while the field is live, there are operations that write the same
> (let's call it "idempotent") value to the field under certain
> circumstances
>
>  - at release time, after all the reference counts are gone, the field
> is read for whether that situation happened. I'm assuming KCSAN
> already understands about this case too?

It's not explicitly aware of initialization or release. We rely on
compiler instrumentation for all memory accesses; KCSAN then sets up
"watchpoints" for sampled memory accesses, delaying execution, and
checking if a concurrent access is observed.

We already have an option (currently disabled on syzbot) where KCSAN
infers a data race not because another instrumented accesses happened
concurrently, but because the data value changed during a watchpoint's
lifetime (e.g. due to uninstrumented write, device write etc.).

This same approach could be used to ignore "idempotent writes" where
we would otherwise report a data race; i.e. if there was a concurrent
write, but the data value did not change, do not report the race. I'm
happy to add this feature if this should always be ignored.

> So it would only be the "idempotent writes" thing that would be
> something KCSAN would have to realize do not involve a race - because
> it simply doesn't matter if two writes of the same value race against
> each other.
>
> But I guess we could also just do
>
>    #define WRITE_IDEMPOTENT(x,y) WRITE_ONCE(x,y)
>
> and use that in the kernel to annotate these things. And if we have
> that kind of annotation, we could then possibly change it to
>
>   #define WRITE_IDEMPOTENT(x,y) \
>        if READ_ONCE(x)!=y WRITE_ONCE(x,y)
>
> if we have numbers that that actually helps (that macro written to be
> intentionally invalid C - it obviously needs statement protection and
> protection against evaluating the arguments multiple times etc).
>
>                 Linus

Thanks,
-- Marco