On Thu, Nov 22, 2018 at 04:05:24PM +0100, Oleg Nesterov wrote:
> On 11/22, Andrea Parri wrote:
> >
> > > See 142b18ddc81439acda4bc4231b291e99fe67d507 ("uprobes: Fix handle_swbp()
> > > vs unregister() + register() race") and the comment above this rmb().
> >
> > Mmh..., at first glance, this suggests me that the above set_bit() and
> > test_bit() to/from uprobe->flags are among these memory accesses. But
> > this doesn't make sense to me: these accesses do not "alternate" (i.e.,
> > they both appear after the corresponding barrier..); instead I'd expect
> > something like (on top of the above diff):
> >
> > diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
> > index 2d29977522017..a75b9a08dee54 100644
> > --- a/kernel/events/uprobes.c
> > +++ b/kernel/events/uprobes.c
> > @@ -2178,10 +2178,18 @@ static void handle_swbp(struct pt_regs *regs)
> > * After we hit the bp, _unregister + _register can install the
> > * new and not-yet-analyzed uprobe at the same address, restart.
> > */
> > - smp_rmb(); /* pairs with the smp_wmb() in prepare_uprobe() */
> > if (unlikely(!test_bit(UPROBE_COPY_INSN, &uprobe->flags)))
> > goto out;
> >
> > + /*
> > + * Pairs with the smp_wmb() in prepare_uprobe().
> > + *
> > + * Guarantees that if we see the UPROBE_COPY_INSN bit set, then
> > + * we must (can) also see the stores to &uprobe->arch performed
> > + * by prepare_uprobe() (say).
> > + */
> > + smp_rmb();
>
> OOPS, you are right! Thanks.
Thank you for the clarification; I'll send a patch with the fix shortly.
Andrea
>
> Oleg.
>
