Re: Malloc improvements

[Florian Weimer <fweimer at redhat dot com>]

On 07/19/2016 02:03 PM, Anton Blanchard wrote:
> Hi Florian.
>
> > x86_64 checks __libc_multiple_threads and avoids atomics if possible.
> > Do you already do this in POWER?
>
> x86 has CISCy instructions that do a local atomic increment/decrement,
> and I suspect they use that in the single threaded case. We do not have
> instructions like that on PowerPC, and it presents an issue for signals.
>
> I think we got resistance when we suggested using load/add/store
> instructions in single threaded mode on PowerPC because it made the
> async signal situation somewhat worse. Of course malloc is not defined
> to be async signal safe, and I bet there are places we fall apart today.
> Perhaps we can revisit this now we have an unlocked per thread cache in
> DJ's branch.

Yes, with the lockless, waitless cache, you get data corruption instead 
of a self-deadlock. :)

I think non-atomic updates are okay.  I doubt that fork will be more 
broken than it is today.  (fork is supposedly async-signal-safe, but our 
implementation is not because it touches the malloc locks.)

Thanks,
Florian