Re: [PATCH] elf: Use C11 like atomics on _dl_mcount

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 01/09/22 11:19, Wilco Dijkstra wrote:
> Hi Adhemerval,
> 
>> Right, I was not trying to be clever here and acquire might be a more
>> strict memory order.
>>
>> And I haven't replied yet to you atomic refactor, but I also think changing
>> step by step is indeed better than changing all the atomic altogether.
> 
> Well that's what I did, I split into about 15-20 patches rather than a single one.

Is the https://patchwork.sourceware.org/project/glibc/list/?series=10727 right? I 
think it has failed CI, could you submit again?

> 
>> Do you think using relaxed on all atomic is suffice here?
> 
> Indeed, adding acquire on these counters will not ensure correct memory ordering.
> 
> Note this particular code looks very unsafe - it is not clear to me how it blocks
> other threads from accessing new elements before they are properly initialized.
> Also the counters that are being updated atomically are read without atomic
> accesses, so may get a cached value or something newer - you don't know.
> 

I thinking we are tracking two different issues, although they are correlated.
My take is to first try to map the current code to a C11-like atomic, so we
can always use compiler builtins where applicable (there are couple of 
architecture that we might avoid using libgcc for performance reasons), and
then check if the atomic usage does make sense.