Re: [patch] mm: rewrite vmap layer

[Nick Piggin <npiggin@suse.de>]

On Wed, Aug 20, 2008 at 09:03:59AM -0500, Christoph Lameter wrote:
> Nick Piggin wrote:
> 
> >> Or run purge_vma_area_lazy from keventd?
> >  
> > Right. But that's only needed if we want to vmap from irq context too
> > (otherwise we can just do the purge check at vmap time).
> > 
> > Is there any good reason to be able to vmap or vunmap from interrupt
> > time, though?
> 
> It would be good to have vunmap work in an interrupt context like other free
> operations. One may hold spinlocks while freeing structure.

I don't know if just-in-case is a strong argument to make the locks
interrupt safe and logic to handle deferred flushing. I'd be happy
to add it if there are some specific cases though.
 

> vmap from interrupt context would be useful f.e. for general fallback in the
> page allocator to virtually mapped memory if no linear physical memory is
> available (virtualizable compound pages). Without a vmap that can be run in an
> interrupt context we cannot support GFP_ATOMIC allocs there.

Indeed that would be a good use for it if this general fallback mechanism
were to be merged.