Messages in this thread | | | Date | Mon, 11 May 2015 10:50:42 -0700 | Subject | Re: [tip:locking/core] locking/pvqspinlock: Replace xchg() by the more descriptive set_mb() | From | Linus Torvalds <> |
| |
On Mon, May 11, 2015 at 7:54 AM, Peter Zijlstra <peterz@infradead.org> wrote: > > Hmm, so I looked at the set_mb() definitions and I figure we want to do > something like the below, right?
I don't think you need to do this for the non-smp cases. The whole thing is about smp memory ordering, so on UP you don't even need the WRITE_ONCE(), much less a barrier.
That said, I do wonder if we should make that "it's only an smp barrier" more explicit. We have non-smp barriers for people who do DMA, and while they should probably never use anything like set_mb() anyway (they tend to want *release* semantics, not a full barrier), from a conceptual standpoint the "set_mb()" function really is closer to the "smp_store_release()/smp_load_acquire()" family of macros.
So I wonder if we should change the name to match.
IOW, if we are really cleaning up smp_mb() and changing most of the lines associated with it (we really have very few users, and there seems to be more lines *defining* smp_mb() than there are lines *using* it in the kernel), maybe we should also just rename it "smp_store_mb()" at the same time.
I dunno. Maybe the churn isn't worth it.
Linus
| |