Re: [PATCH v6] taskstats: fix data-race

[Christian Brauner]

On Thu, Oct 24, 2019 at 03:13:48PM +0200, Dmitry Vyukov wrote:
> On Thu, Oct 24, 2019 at 3:05 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote:
> >
> > On Thu, Oct 24, 2019 at 01:51:20PM +0200, Dmitry Vyukov wrote:
> > > On Thu, Oct 24, 2019 at 1:32 PM Andrea Parri <parri.andrea@xxxxxxxxx> wrote:
> > > >
> > > > > How these later loads can be completely independent of the pointer
> > > > > value? They need to obtain the pointer value from somewhere. And this
> > > > > can only be done by loaded it. And if a thread loads a pointer and
> > > > > then dereferences that pointer, that's a data/address dependency and
> > > > > we assume this is now covered by READ_ONCE.
> > > >
> > > > The "dependency" I was considering here is a dependency _between the
> > > > load of sig->stats in taskstats_tgid_alloc() and the (program-order)
> > > > later loads of *(sig->stats) in taskstats_exit().  Roughly speaking,
> > > > such a dependency should correspond to a dependency chain at the asm
> > > > or registers level from the first load to the later loads; e.g., in:
> > > >
> > > >   Thread [register r0 contains the address of sig->stats]
> > > >
> > > >   A: LOAD r1,[r0]       // LOAD_ACQUIRE sig->stats
> > > >      ...
> > > >   B: LOAD r2,[r0]       // LOAD *(sig->stats)
> > > >   C: LOAD r3,[r2]
> > > >
> > > > there would be no such dependency from A to C.  Compare, e.g., with:
> > > >
> > > >   Thread [register r0 contains the address of sig->stats]
> > > >
> > > >   A: LOAD r1,[r0]       // LOAD_ACQUIRE sig->stats
> > > >      ...
> > > >   C: LOAD r3,[r1]       // LOAD *(sig->stats)
> > > >
> > > > AFAICT, there's no guarantee that the compilers will generate such a
> > > > dependency from the code under discussion.
> > >
> > > Fixing this by making A ACQUIRE looks like somewhat weird code pattern
> > > to me (though correct). B is what loads the address used to read
> > > indirect data, so B ought to be ACQUIRE (or LOAD-DEPENDS which we get
> > > from READ_ONCE).
> > >
> > > What you are suggesting is:
> > >
> > > addr = ptr.load(memory_order_acquire);
> > > if (addr) {
> > >   addr = ptr.load(memory_order_relaxed);
> > >   data = *addr;
> > > }
> > >
> > > whereas the canonical/non-convoluted form of this pattern is:
> > >
> > > addr = ptr.load(memory_order_consume);
> > > if (addr)
> > >   data = *addr;
> >
> > No, I'd rather be suggesting:
> >
> >   addr = ptr.load(memory_order_acquire);
> >   if (addr)
> >     data = *addr;
> >
> > since I'd not expect any form of encouragement to rely on "consume" or
> > on "READ_ONCE() + true-address-dependency" from myself.  ;-)
> 
> But why? I think kernel contains lots of such cases and it seems to be
> officially documented by the LKMM:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/memory-model/Documentation/explanation.txt
> address dependencies and ppo

You mean this section:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/memory-model/Documentation/explanation.txt#n955
and specifically:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/memory-model/Documentation/explanation.txt#n982
?