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

[Matthew Wilcox <willy@infradead.org>]

On Wed, May 20, 2020 at 09:39:49PM -0700, Dan Williams wrote:
> 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?
> >
> > > > +
> > > > +     /* 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.

Completed, no, but I think it does enforce that they're visible to other
CPUs before this write is visible to other CPUs.

I'll quote relevant bits from the document ...

 (2) Data dependency barriers.

     A data dependency barrier is a weaker form of read barrier.  In the case
     where two loads are performed such that the second depends on the result
     of the first (eg: the first load retrieves the address to which the second
     load will be directed), a data dependency barrier would be required to
     make sure that the target of the second load is updated after the address
     obtained by the first load is accessed.

[...]
SMP BARRIER PAIRING
-------------------
[...]
        CPU 1                 CPU 2
        ===============       ===============================
        a = 1;
        <write barrier>
        WRITE_ONCE(b, &a);    x = READ_ONCE(b);
                              <data dependency barrier>
                              y = *x;


> > >
> > > > +     /*
> > > > +      * 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.

I don't think I explained myself well enough.

When we initialise devmem_inode, does devmem_inode->i_mapping->host point
to somewhere other than devmem_inode?

I appreciate in this function, inode->i_mapping->host will point to inode.
But we're now changing i_mapping to be devmem_inode's i_mapping.  Why
do we need to change devmem_inode's i_mapping->host pointer?