On Sat, Dec 13, 2014 at 04:14:58PM -0800, Linus Torvalds wrote:
> On Sat, Dec 13, 2014 at 3:47 PM, Al Viro <v...@zeniv.linux.org.uk> wrote:
> >
> > static inline void mnt_dec_writers(struct mount *mnt)
> > {
> > #ifdef CONFIG_SMP
> > this_cpu_dec(mnt->mnt_pcp->mnt_writers);
> > #else
> > mnt->mnt_writers--;
> > #endif
> > }
> > It's load/modify/store, without any kind of atomicity; get preempted in the
> > middle of that sequence by another caller of mnt_dec_writers() and obvious
> > bad
> > things will happen...
>
> Ugh, yes ok, the UP case needs it for the actual counter itself. Ugh.
> What an ugly mess. I'd rather have the preemption disable where it is
> actually *needed*, in that function itself for the UP case (or just
> make it "atomic_t", which would likely be better still.
So does SMP - this_cpu_dec() relies on preemption being disabled. On x86
we might get away with that, what with having it compiled into decl %gs:const,
but on generic it turns into
*raw_cpu_ptr(&pcp) -= 1;
and compiler has every right to turn it into
p = raw_cpu_ptr(&pcp);
(*p)--;
again, with no locking. Lose the timeslice in the middle of that and you
are risking to get a different CPU when you are scheduled again, with
another process doing this_cpu_dec() on your old CPU. Have fun - two
non-atomic decrements of the same variable by different CPUs in parallel...
We really need preemtion disabled there, UP or no UP.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
