Hello Linus, On Tue, Feb 02, 2016 at 12:19:04AM -0800, Linus Torvalds wrote: > On Tue, Feb 2, 2016 at 12:07 AM, Linus Torvalds > <torva...@linux-foundation.org> wrote: > > > > So we *absolutely* should say that *OF COURSE* these things work: > > > > - CPU A: > > > > .. initialize data structure -> smp_wmb() -> WRITE_ONCE(ptr); > > > > - CPU B: > > > > smp_load_acquire(ptr) - we can rely on things behind "ptr" being > > initialized > > That's a bad example, btw. I shouldn't have made it be a "pointer", > because then we get the whole address dependency chain ordering > anyway. > > So instead of "ptr", read "state flag". It might just be an "int" that > says "data has been initialized". > > So > > .. initialize memory .. > smp_wmb(); > WRITE_ONCE(&is_initialized, 1); > > should pair with > > if (smp_load_acquire(&is_initialized)) > ... we can read and write the data, knowing it has been initialized .. > > exactly because "smp_wmb()" (cheap write barrier) might be cheaper > than "smp_store_release()" (expensive full barrier) and thus > preferred. >
Just to be clear, what Will, Paul and I are discussing here is about
local transitivity, which refers to something like this following
example:
(a, b and is_initialized are all initially zero)
P0:
WRITE_ONCE(a, 1);
smp_store_release(&is_initialized, 1);
P1:
r1 = smp_load_acquire(&is_initialized);
smp_store_release(&b, 1);
P2:
r2 = smp_load_acquire(&b);
r3 = READ_ONCE(a);
, in which case, r1 == 1 && r2 == 1 && r3 == 0 can not happen because
RELEASE+ACQUIRE pairs guarantee local transitivity.
More on local transitvity:
http://article.gmane.org/gmane.linux.kernel.virtualization/26856
And what I'm asking here is something like this following example:
(a, b and is_initialized are all initially zero)
P0:
WRITE_ONCE(a, 1);
smp_store_release(&is_initialized, 1);
P1:
if (r1 = READ_ONCE(is_initialized))
smp_store_release(&b, 1);
P2:
if (r2 = READ_ONCE(b)) {
smp_rmb();
r3 = READ_ONCE(a);
}
, in which case, can r1 == 1 && r2 == 1 && r3 == 0 happen?
Please note this example is about two questions on local transitivity:
1. Could "READ_ONCE(); if" extend a locally transitive chain.
2. Could "READ_ONCE(); if; smp_rmb()" at least be a locally
transitive chain termination?
> So mixing ordering metaphors actually does make sense, and should be
> entirely well-defined.
>
I think Paul does agree that smp_{r,w}mb() with applicative memory
operations around could pair with smp_store_release() or
smp_load_acquire().
Hope I didn't misunderstand any of you or make you misunderstood with
each other..
Regards,
Boqun
> There's likely less reason to do it the other way (ie
> "smp_store_release()" on one side pairing with "LOAD_ONCE() +
> smp_rmb()" on the other) since there likely isn't the same kind of
> performance reason for that pairing. But even if we would never
> necessarily want to do it, I think our memory ordering rules would be
> *much* better for strongly stating that it has to work, than being
> timid and trying to make the rules weak.
>
> Memory ordering is confusing enough as it is. We should not make
> people worry more than they already have to. Strong rules are good.
>
> Linus
signature.asc
Description: PGP signature
