On Mon, Oct 06, 2014 at 11:19:15AM +0200, Peter Zijlstra wrote:
> On Mon, Oct 06, 2014 at 02:25:09AM +0200, Oleg Nesterov wrote:
> > Yes, and the comments ;)
> >
> > I showed this patch only to complete the discussion, I am not going to
> > send it now.
>
> Fair enough :-)
>
> > But thanks for the review!
> >
> > > > +static void kthread_kill(struct task_struct *k, struct kthread
> > > > *kthread)
> > > > +{
> > > > + smp_mb__before_atomic();
> > >
> > > test_bit isn't actually an atomic op so this barrier is 'wrong'. If you
> > > need an MB there smp_mb() it is.
> >
> > Hmm. I specially checked Documentation/memory-barriers.txt,
> >
> > (*) smp_mb__before_atomic();
> > (*) smp_mb__after_atomic();
> >
> > These are for use with atomic (such as add, subtract, increment and
> > decrement) functions that don't return a value, especially when used
> > for
> > reference counting. These functions do not imply memory barriers.
> >
> > These are also used for atomic bitop functions that do not return a
> > value (such as set_bit and clear_bit).
> > ^^^^^^^^^^^^^^^^^^^^^
> >
> > Either you or memory-barriers.txt should be fixed ;)
>
> Its in there, just not explicitly. All those functions listed are
> read-modify-write ops, test_bit() is not, its just a read. But yes I
> suppose we could make that more explicit.
>
> Also test_bit() obviously does return a value, so it doesn't fall in the
> {set,clear}_bit() class.
>
> Does the change below clarify things?
>
> > > > + if (test_bit(KTHREAD_WANTS_SIGNAL, &kthread->flags)) {
> > > > + unsigned long flags;
> > > > + bool kill = true;
> > > > +
> > > > + if (lock_task_sighand(k, &flags)) {
> > >
> > > Since we do the double test thing here, with the set side also done
> > > under the lock, so we really need a barrier above?
> >
> > Yes, otherwise set_kthread_wants_signal() can miss a signal. And note
> > that the 2nd check is only needed to ensure that we can not race
> > with set_kthread_wants_signal(false).
> >
> > BUT!!! I have to admit that I simply do not know if there is any arch
> >
> > set_bit(&word, X);
> > test_bit(&word, Y);
> >
> > which actually needs mb() in between, the word is the same. Probably
> > not.
>
> DEC Alpha? Wasn't it the problem there that dependencies didn't actually
> work as expected?
This looks to me to be an issue of cache coherence rather than
dependency ordering, so I would expect that DEC Alpha would respect
the ordering.
Thanx, Paul
> Added Paul to Cc.
>
> ---
> Documentation/memory-barriers.txt | 9 +++------
> 1 file changed, 3 insertions(+), 6 deletions(-)
>
> diff --git a/Documentation/memory-barriers.txt
> b/Documentation/memory-barriers.txt
> index 22a969cdd476..0d97c99ad957 100644
> --- a/Documentation/memory-barriers.txt
> +++ b/Documentation/memory-barriers.txt
> @@ -1594,12 +1594,9 @@ CPU from reordering them.
> (*) smp_mb__before_atomic();
> (*) smp_mb__after_atomic();
>
> - These are for use with atomic (such as add, subtract, increment and
> - decrement) functions that don't return a value, especially when used for
> - reference counting. These functions do not imply memory barriers.
> -
> - These are also used for atomic bitop functions that do not return a
> - value (such as set_bit and clear_bit).
> + These are for use with atomic/bitop (r-m-w) functions that don't return
> + a value (eg. atomic_{add,sub,inc,dec}(), {set,clear}_bit()). These
> + functions do not imply memory barriers.
>
> As an example, consider a piece of code that marks an object as being
> dead
> and then decrements the object's reference count:
>
--
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/
