Re: kmap_atomic and preemption

From: Peter Zijlstra
Date: Thu May 05 2016 - 05:37:21 EST

On Wed, May 04, 2016 at 08:17:55PM +0100, Russell King - ARM Linux wrote:
> On Wed, May 04, 2016 at 03:47:29PM +0200, Peter Zijlstra wrote:
> > Traditionally kmap_atomic() disables preemption; and the reason is that
> > the returned pointer must stay valid. This had a side effect in that it
> > also disabled pagefaults.
> A lowmem page should never change its page_address(), so that much is
> safe.


> I think the question is whether there is any driver code which
> assumes that preemption is unconditionally disabled between a
> kmap_atomic() has been called.

right, this and consistency. Having the function disable preemption
sometimes is just plain weird.

> That wouldn't be an unreasonable assumption given the name of the
> function, so I'd suggest caution with making kmap_atomic() have these
> kinds of differing behaviours depending on whether we're asking to
> kmap a high or lowmem page.

So for -rt I did a preemptible kmap_atomic() for x86 (and maybe someone
did other archs too, I cannot remember), now -rt is funny wrt locking
anyway, but I cannot remember anything breaking because of this, so
there is some hope it will actually work.

> If we are going to allow this, I think it at least needs to be well
> documented.