Re: [PATCH] malloc: Use correct C11 atomics for fastbin

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

Hi Florian,

>> Fix memory ordering issues in the fastbin implementation: in REMOVE_FB the
>> next pointer is read before any MO synchronization, however in the C11 atomic
>> model this is only correct after a load acquire. Refactor the fastbin code
>> and add a dedicated fastbin_push/pop implementation. The number of acquire
>> or release atomics remains the same, and the new functions are inlined, so
>> performance is unaffected.
>>
>> Passes regress, OK for commit?
>
> Did you actually observe any problems resulting from this?  We
> previously concluded that the dependency chain would mae this valid
> (outside the C11 memory model).

No, I believe it works in most weak memory models (except for Alpha).
However the code always looked weird with the acquire done after several
non-atomic accesses... The key is that we need to do the acquire/release
only once per iteration in the CAS loop, and that avoids regressions.

Cheers,
Wilco