On Wed, 07 Jun 2017 05:58:50 PDT (-0700), peterz@xxxxxxxxxxxxx wrote:
> On Wed, Jun 07, 2017 at 02:36:27PM +0200, Peter Zijlstra wrote:
>> Which (pending the sub confusion) will generate the entire set of:
>>
>> atomic_add, atomic_add_return{_relaxed,_acquire,_release,} atomic_fetch_add{_relaxed,_acquire,_release,}
>> atomic_sub, atomic_sub_return{_relaxed,_acquire,_release,} atomic_fetch_sub{_relaxed,_acquire,_release,}
>>
>> atomic_and, atomic_fetch_and{_relaxed,_acquire,_release,}
>> atomic_or, atomic_fetch_or{_relaxed,_acquire,_release,}
>> atomic_xor, atomic_fetch_xor{_relaxed,_acquire,_release,}
>>
>
> Another approach would be to override __atomic_op_{acquire,release} and
> use things like:
>
> "FENCE r,rw" -- (load) ACQUIRE
> "FENCE rw,w" -- (store) RELEASE
>
> And then you only need to provide _relaxed atomics.
>
> Also, and I didn't check for that, you need to provide:
>
> smp_load_acquire(), smp_store_release(), atomic_read_acquire(),
> atomic_store_release().
OK, thanks for looking so deeply into this. Sorry it was such a mess, I
thought I included a note somewhere that this all needed to be redone -- I just
wanted to get a v2 out first as that split all the drivers out. I've went
ahead and completely rewrote atomic.h using your suggestions in a slightly
modified way. It includes
* _relaxed, _acquire, and _release versions of everything via a bunch of macros.
* What I believe to be correct aqrl bits on every op.
* 64-bit and 32-bit atomics (as opposed to just copying everything)
I didn't implement try_cmpxchg yet. I'm going to go ahead and sort through our
memory barriers, look at the few remaining CR comments from our v2, and then
submit a v3 patch set.
I'm only replying to this message, but I believe I'll have taken into account
all your comments for the v3.
Thanks, again, for your time!
