On Tue, 6 Aug 2024 at 09:32, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > So I think smp_wmb() and smp_rmb() would be sufficient. I also find it
> > clearer in this case.
>
> It's not the question of sufficiency; it's whether anything cheaper can be
> had.
I'd avoid smp_rmb(), since that can be very expensive, but
"smp_load_acquire()" is about as cheap as it comes. As is
"smp_store_release()" and "smp_wmb()".
So if it's mostly about just ordering with max_fds, I'd suggest just
doing the store of the bigger max_fds value with a smp_store_release()
(which guarantees that any previous stores of the bitmap pointers etc
have been done first and are visible), and then before accessing the
arrays, do the load of the max_fds thing with a "smp_load_acquire()",
and it should all be good.
But the array pointers themselves also need the same
smp_store_release() when updating, just to guarantee that the newly
copied contents are ordered. Which implies having to read those too
with smp_load_acquire(), regardless of the max_fds value read.
But on x86, smp_store_release/smp_load_acquire is completely free. And
on other architectures it's generally one of the cheapest possible
ordering constraints. So I doubt it's a performance issue even outside
of x86, and avoiding a double indirection is probably a win.
Linus
