On Thu, Jun 11, 2015 at 08:33:41AM -0700, Paul E. McKenney wrote:
> On Thu, Jun 11, 2015 at 02:46:47PM +0200, Peter Zijlstra wrote:
> > Introduce raw_write_seqcount_barrier(), a new construct that can be
> > used to provide write barrier semantics in seqcount read loops instead
> > of the usual consistency guarantee.
> >
> > Cc: Al Viro <v...@zeniv.linux.org.uk>
> > Cc: Linus Torvalds <torva...@linux-foundation.org>
> > Cc: Paul McKenney <paul...@linux.vnet.ibm.com>
> > Suggested-by: Oleg Nesterov <o...@redhat.com>
> > Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org>
> > ---
> > include/linux/seqlock.h | 42 ++++++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 42 insertions(+)
> >
> > --- a/include/linux/seqlock.h
> > +++ b/include/linux/seqlock.h
> > @@ -233,6 +233,48 @@ static inline void raw_write_seqcount_en
> > s->sequence++;
> > }
> >
> > +/**
> > + * raw_write_seqcount_barrier - do a seq write barrier
> > + * @s: pointer to seqcount_t
> > + *
> > + * This can be used to provide an ordering guarantee instead of the
> > + * usual consistency guarantee. It is one wmb cheaper, because we can
> > + * collapse the two back-to-back wmb()s.
> > + *
> > + * seqcount_t seq;
> > + * bool X = true, Y = false;
> > + *
> > + * void read(void)
> > + * {
> > + * bool x, y;
> > + *
> > + * do {
> > + * int s = read_seqcount_begin(&seq);
> > + *
> > + * x = X; y = Y;
> > + *
> > + * } while (read_seqcount_retry(&seq, s));
> > + *
> > + * BUG_ON(!x && !y);
> > + * }
> > + *
> > + * void write(void)
> > + * {
> > + * Y = true;
> > + *
> > + * write_seqcount_begin(seq);
> > + * write_seqcount_end(seq);
> > + *
> > + * X = false;
> > + * }
>
> So when using this, write() would instead look like this?
>
> void write(void)
> {
> Y = true;
> raw_write_seqcount_barrier(seq);
> X = false;
> }
>
> I suggest calling this out explicitly. Agreed, it should be obvious,
> but some poor sot is going to be reading this at 3AM local time after
> a couple days of no sleep, in which case obvious might not be so obvious.
>
> I also would suggest READ_ONCE() and WRITE_ONCE() to keep the compiler
> trickiness down to a dull roar. Understood, it is hard to make anything
> bad happen in this case, but small changes could result in badness.
>
> > + */
> > +static inline void raw_write_seqcount_barrier(seqcount_t *s)
> > +{
> > + s->sequence++;
> > + smp_wmb();
> > + s->sequence++;
> > +}
> > +
> > /*
> > * raw_write_seqcount_latch - redirect readers to even/odd copy
> > * @s: pointer to seqcount_t
>
> Looks good otherwise.
>
> Reviewed-by: Paul E. McKenney <paul...@linux.vnet.ibm.com>
Color me slow and stupid. Maybe due to reviewing a patch too early in
the morning, who knows?
There is nothing above that prevents the compiler and the CPU from
reordering the assignments to X and Y with the increment of s->sequence++.
One fix would be as follows:
static inline void raw_write_seqcount_barrier(seqcount_t *s)
{
smp_wmb();
s->sequence++;
smp_wmb();
s->sequence++;
smp_wmb();
}
Of course, this assumes that the accesses surrounding the call to
raw_write_seqcount_barrier() are writes. If they can be a reads,
the two added smp_wmb() calls need to be full barriers.
Thanx, Paul
--
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/
