Re: [PATCH v5 08/15] mm: Add unsafe_follow_pfn

[Daniel Vetter]

On Mon, Nov 2, 2020 at 4:52 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Mon, Nov 02, 2020 at 02:23:58PM +0100, Daniel Vetter wrote:
> > On Mon, Nov 2, 2020 at 2:01 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> > >
> > > On Mon, Nov 02, 2020 at 01:56:10PM +0100, Daniel Vetter wrote:
> > > > On Mon, Nov 2, 2020 at 8:29 AM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Oct 30, 2020 at 11:08:08AM +0100, Daniel Vetter wrote:
> > > > > > Also mark up follow_pfn as EXPORT_SYMBOL_GPL. The only safe way to use
> > > > > > that by drivers/modules is together with an mmu_notifier, and that's
> > > > > > all _GPL stuff.
> > > > >
> > > > > I also think it also needs to be renamed to explicitly break any existing
> > > > > users out of tree or int the submission queue.
> > > >
> > > > Ok I looked at the mmu notifier locking again and noticed that
> > > > mm->subscriptions has its own spinlock. Since there usually shouldn't
> > > > be a huge pile of these I think it's feasible to check for the mmu
> > > > notifier in follow_pfn. And that would stuff this gap for good. I'll
> > > > throw that on top as a final patch and see what people think.
> > >
> > > Probably the simplest is to just check mm_has_notifiers() when in
> > > lockdep or something very simple like that
> >
> > lockdep feels wrong, was locking more at CONFIG_DEBUG_VM. And since
> > generally you only have 1 mmu notifier (especially for kvm) I think we
> > can also pay the 2nd cacheline miss and actually check the right mmu
> > notifier is registered.
>
> Need to hold the lock to check that and there are two ways to register
> notifiers these days, so it feels to expensive to me.

Uh I mixed stuff up all along, struct mmu_notifier *subcription that
all the mmu notifier users use has the ->mm pointer we want right
there. That's good enough I think.

Now I'm kinda lost in kvm code trying to wire it through, but it's
looking ok-ish thus far :-)
-Daniel

> CH's 'export symbol only for kvm' really does seem the most robust way
> to handle this though.
>
> Jason



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