Re: regarding use of various cmpxchg* API

From: Peter Zijlstra
Date: Fri Jun 06 2014 - 03:29:36 EST

On Thu, Jun 05, 2014 at 06:44:37PM -0400, Pranith Kumar wrote:
> Hi Peter, Ingo,
> I see that in the drivers there are the following types of cmpxchg API:
> __cmpxchg64()

That shouldn't really be used, and its only used in intel-iommu.c.

> atomic_cmpxchg()
> atomic64_cmpxchg()
> atomic_long_cmpxchg()

They're part of the atomic*_t family and should be used on their
respective types, no point in changing that.

> local_cmpxchg()(in kernel/events/ring_buffer.c)

This is part of the local_t API, also no real point to change that.

> this_cpu_cmpxchg()

This is part of the this_cpu*() family, and there's no way you can
generate the same code using the below to.

> cmpxchg()
> cmpxchg_local()

Which leaves these, which are used in all other cases.

> Since cmpxchg() internally handles the width, do you think it makes
> sense to replace the above uses of cmpxchg with the document API
> (cmpxchg, atomic_cmpxchg, cmpxchg_local)?
> I am willing to do this and wanted to know if it something you think
> is worth pursuing.

Don't think that's useful. If you really want to go do something, try
the annotation I suggested to get the parisc/sparc32 things correct
again. Add the __atomic sparse address space and the store()/load()

Attachment: pgpcmauKDU9bG.pgp
Description: PGP signature