Re: [RFC][PATCH 5/5] arch: Sanitize atomic_t bitwise ops

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Feb 06, 2014 at 08:53:02AM -0800, Linus Torvalds wrote:
> On Thu, Feb 6, 2014 at 5:48 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > Many archs have atomic_{set,clear}_mask() but not all. Remove these
> > and provide a comprehensive set of bitops:
> >
> >   atomic{,64}_{and,or,xor}{,_return}()
> 
> Who uses these, and why?

s390 is stuffed with atomic_{set,clear}_mask usage, no clue whatfor.

There's 2 atomic_set_mask() users in drm/i915, again haven't looked if
it makes sense.

Various archs use atomic_set_mask() for their tlb flush mask, which
arguably should be done using the bitmap functions we have.

> The "_return()" versions of atomic ops are noticeably slower and more
> complex on common architectures (ie x86), and apparently there is no
> use of them since they didn't exist.
> 
> So why add them? Just to encourage people to do bad things?

Fair enough, they're mostly an accident of how I implemented the macro
generation magic. I suppose I can change that and avoid generating the
_return thingies.

I don't particularly care for this patch too much; but I do dislike the
atomic_{set,clear}_mask() things -- mostly because they don't really fit
the normal atomic_t functions and because some archs have then and
others do no.

And atomic_{set,clear}_mask() can be used for setting/clearing multiple
bits as opposed to {set,clear}_bit, which only does a single bit.