Re: [tip:perfcounters/core] x86: Add NMI types for kmap_atomic

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2009-06-17 at 17:43 +0900, Tejun Heo wrote:
> Hello,
> 
> Peter Zijlstra wrote:
> > On Tue, 2009-06-16 at 13:38 +0100, Hugh Dickins wrote:
> >> Something else to throw in: what if they were not just atomic,
> >> but also replaced the current sleeping kmaps? i.e. a task context
> >> carries around its own stack of these.
> > 
> > I actually did that once, but it means the task needs to be cpu-affine,
> > because fixmaps have different addresses between cpus. And disabling
> > migration for tasks has subtle side-effects so I dropped that again.
> > 
> > However, I recently considered the possiblity of putting the fixmaps in
> > the new per-cpu address space so that we might use the %gs segment to
> > normalize the fixmap addresses between the cpus.
> > 
> > This would allow full preemptible kmaps (yay for -rt).
> > 
> > However I suspect it might greatly complicate kmaps for the !i386 world.
> 
> Other archs are in the process of conversion so once that is complete,
> there's no reason this should be more difficult but it means that
> kmapped addresses should be accessed differently from regular ones
> which we can't do.  :-(

Right, we'd need percpu_memset, percpu_copy_from, percpu_copy_to, etc.
that all operate on %gs like addrs. might be a tad invasive.