[Intel-gfx] [PATCH 00/34] put_user_pages(): miscellaneous call sites

[Matthew Wilcox]

On Fri, Aug 02, 2019 at 02:41:46PM +0200, Jan Kara wrote:
> On Fri 02-08-19 11:12:44, Michal Hocko wrote:
> > On Thu 01-08-19 19:19:31, john.hubbard at gmail.com wrote:
> > [...]
> > > 2) Convert all of the call sites for get_user_pages*(), to
> > > invoke put_user_page*(), instead of put_page(). This involves dozens of
> > > call sites, and will take some time.
> > 
> > How do we make sure this is the case and it will remain the case in the
> > future? There must be some automagic to enforce/check that. It is simply
> > not manageable to do it every now and then because then 3) will simply
> > be never safe.
> > 
> > Have you considered coccinele or some other scripted way to do the
> > transition? I have no idea how to deal with future changes that would
> > break the balance though.
> 
> Yeah, that's why I've been suggesting at LSF/MM that we may need to create
> a gup wrapper - say vaddr_pin_pages() - and track which sites dropping
> references got converted by using this wrapper instead of gup. The
> counterpart would then be more logically named as unpin_page() or whatever
> instead of put_user_page().  Sure this is not completely foolproof (you can
> create new callsite using vaddr_pin_pages() and then just drop refs using
> put_page()) but I suppose it would be a high enough barrier for missed
> conversions... Thoughts?

I think the API we really need is get_user_bvec() / put_user_bvec(),
and I know Christoph has been putting some work into that.  That avoids
doing refcount operations on hundreds of pages if the page in question is
a huge page.  Once people are switched over to that, they won't be tempted
to manually call put_page() on the individual constituent pages of a bvec.