Re: set_page_dirty races (was: Re: [patch 2/4] vfs: add set_page_dirty_notag)

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2009-02-17 at 12:55 +0100, Nick Piggin wrote:
> On Tue, Feb 17, 2009 at 12:39:32PM +0100, Peter Zijlstra wrote:
> > On Tue, 2009-02-17 at 12:25 +0100, Nick Piggin wrote:
> > 
> > > Introduce put_user_pages function.
> > > 
> > > In order to have more flexibility to deal with issues surrounding
> > > get_user_pages difficulties[*], introduce put_user_pages function
> > > intended to release pages acquired by get_user_pages. For now, just
> > > do the regular put_page thing. If all callers are converted, it could
> > > be used to help with such races. In the meantime, it will actually
> > > serve as a small extra piece of documentation for the code.
> > > 
> > > [*] eg. get_user_pages caller can bypass page_mkwrite calls into the
> > >     filesystem to notify of page dirty activity if the page gets cleaned
> > >     before the caller calls its final set_page_dirty).
> > 
> > Hmm, if we want to distinguish between .write=1 and .write=0, we would
> > have to pass .write to pup too, right?
> 
> Doh, yeah. I hand edited the patch to put that parameter in, but quilt
> refresh must have outsmarted me!
> 
> If nobody thinks it is insane, I'll resend to Andrew in a new thread.

Right, gup_fast() seems to also respect .write properly, so it would
also be used to balance that.

I guess gup_fast() would need to use trylock_page(), and fall back to
the slow path when we start taking PG_locked on .write.

I suppose we should start converting a few gup users over to pup before
handing the thing to Andrew, to have at least a few examples in-kernel.