On Tue, Sep 22, 2015 at 9:10 AM, Tejun Heo <t...@kernel.org> wrote:
>
> That's a pentium pro era errata. Virtually no working machine is
> affected by that anymore and nobody builds kernel with that option.
> In most cases, store_release and load_acquire are cheaper as they're
> more specific. On x86, store_release and load_acquire boil down to
> compiler reordering barriers. You're running in the opposite
> direction.
Well, to be fair, there are lots of machines where acquire/release is
actually quite expensive.
In general, the cheapest barrier there is (apart from the "no barrier
at all" or just "compiler barrier") is "smp_wmb()". If an
architecture gets that one wrong, the architects were f*cking morons.
It should be a fundamentally cheap operation, since writes are
buffered and it should simply be a buffer barrier.
The acquire/release things are generally fairly cheap on modern
architectures. Not free (except on x86), but fairly low-cost. HOWEVER,
they are not at all free on some older architectures, including 32-bit
ARM.
smp_rmb() should generally be about the same cost as an acquire. It
can go either way.
So *if* the algorithm is amenable to smp_wmb()/smp_rmb() kind of
barriers, that's actually quite possibly better than acquire/release.
smp_mb() is expensive pretty much everywhere.
Looking forward, I suspect long-term acquire/release is what hardware
is going to be "reasonably good at", but as things are right now, you
can't necessarily rely on them being fast.
Linus
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
