Re: [PATCH v3] /dev/mem: Revoke mappings when a driver claims the region

[Dan Williams <dan.j.williams@intel.com>]

On Wed, May 20, 2020 at 9:37 PM Dan Williams <dan.j.williams@intel.com> wrote:
>
> On Wed, May 20, 2020 at 7:26 PM Matthew Wilcox <willy@infradead.org> wrote:
> >
> > On Wed, May 20, 2020 at 06:35:25PM -0700, Dan Williams wrote:
> > > +static struct inode *devmem_inode;
> > > +
> > > +#ifdef CONFIG_IO_STRICT_DEVMEM
> > > +void revoke_devmem(struct resource *res)
> > > +{
> > > +     struct inode *inode = READ_ONCE(devmem_inode);
> > > +
> > > +     /*
> > > +      * Check that the initialization has completed. Losing the race
> > > +      * is ok because it means drivers are claiming resources before
> > > +      * the fs_initcall level of init and prevent /dev/mem from
> > > +      * establishing mappings.
> > > +      */
> > > +     smp_rmb();
> > > +     if (!inode)
> > > +             return;
> >
> > But we don't need the smp_rmb() here, right?  READ_ONCE and WRITE_ONCE
> > are a DATA DEPENDENCY barrier (in Documentation/memory-barriers.txt parlance)
> > so the smp_rmb() is superfluous ...
>
> Is it? I did not grok that from Documentation/memory-barriers.txt.
> READ_ONCE and WRITE_ONCE are certainly ordered with respect to each
> other in the same function, but I thought they still depend on
> barriers for smp ordering?
>
> >
> > > +     /*
> > > +      * Use a unified address space to have a single point to manage
> > > +      * revocations when drivers want to take over a /dev/mem mapped
> > > +      * range.
> > > +      */
> > > +     inode->i_mapping = devmem_inode->i_mapping;
> > > +     inode->i_mapping->host = devmem_inode;
> >
> > umm ... devmem_inode->i_mapping->host doesn't already point to devmem_inode?
>
> Not if inode is coming from:
>
>      mknod ./newmem c 1 1
>
> ...that's the problem that a unified inode solves. You can mknod all
> you want, but mapping and mapping->host will point to a common
> instance.
>
> >
> > > +
> > > +     /* publish /dev/mem initialized */
> > > +     smp_wmb();
> > > +     WRITE_ONCE(devmem_inode, inode);
> >
> > As above, unnecessary barrier, I think.
>
> Well, if you're not sure, how sure should I be?

I'm pretty sure they are needed, because I need the prior writes to
initialize the inode to be fenced before the final write to publish
the inode. I don't think WRITE_ONCE() enforces that prior writes have
completed.