[PATCH] malloc: Use correct C11 atomics for fastbin

[Wilco Dijkstra]

Hi Florian,

>> 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.

That's because tcache has hardly any effect on programs with a high rate
of (de)allocations. It's so small that you usually end up using the fastbin
code anyway. All the singlethreaded optimizations in fastbins are absolutely
essential - that's the code that shows up in profiles.

If we want to make tcache actually work, it will have to support far more
allocations, particularly for smaller sizes.

Cheers,
Wilco