[Intel-gfx] [patch RFC 00/15] mm/highmem: Provide a preemptible variant of kmap_atomic & friends

[Daniel Vetter]

On Sun, Sep 20, 2020 at 08:23:26AM +0200, Thomas Gleixner wrote:
> On Sat, Sep 19 2020 at 12:37, Daniel Vetter wrote:
> > On Sat, Sep 19, 2020 at 12:35 PM Daniel Vetter <daniel at ffwll.ch> wrote:
> >> I think it should be the case, but I want to double check: Will
> >> copy_*_user be allowed within a kmap_temporary section? This would
> >> allow us to ditch an absolute pile of slowpaths.
> >
> > (coffee just kicked in) copy_*_user is ofc allowed, but if you hit a
> > page fault you get a short read/write. This looks like it would remove
> > the need to handle these in a slowpath, since page faults can now be
> > served in this new kmap_temporary sections. But this sounds too good
> > to be true, so I'm wondering what I'm missing.
> 
> In principle we could allow pagefaults, but not with the currently
> proposed interface which can be called from any context. Obviously if
> called from atomic context it can't handle user page faults.
 
Yeah that's clear, but does the implemention need to disable pagefaults
unconditionally?

> In theory we could make a variant which does not disable pagefaults, but
> that's what kmap() already provides.

Currently we have a bunch of code which roughly does

	kmap_atomic();
	copy_*_user();
	kunmap_atomic();

	if (short_copy_user) {
		kmap();
		copy_*_user(remaining_stuff);
		kunmap();
	}

And the only reason is that kmap is a notch slower, hence the fastpath. If
we get a kmap which is fast and allows pagefaults (only in contexts that
allow pagefaults already ofc) then we can ditch a pile of kmap users.

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch