[PATCH] malloc: Use correct C11 atomics for fastbin

[Florian Weimer]

* Wilco Dijkstra:

> Hi,
>
>>> Wilco Dijkstra via Libc-alpha <libc-alpha@sourceware.org> writes:
>>>> +/* Atomically pop from the fastbin list.  The arena lock must be held to
>>>> +   block other threads removing entries, avoiding the ABA issue.  */
>>>
>>> If the arena lock must be held anyway, why go through the compare and
>>> exchange overhead?  We know we're the only thread accessing it.
>>
>> Other threads are adding entries without the arena lock.
>
> Yes, malloc is blocking but free isn't and accesses the freelist
> concurrently.  It's a really weird design. Splitting the free list
> into a local one and a shared one would be far better - no atomics
> when you have the malloc lock, and if the local free list is empty it
> takes one atomic to copy all shared entries.

The local free list is in the tcache.  I think DJ said that removing the
fastbins still resulted in a performance loss.

The tcache has also the benefit that the chain length is bounded.

Thanks,
Florian