On Fri, Oct 09, 2020 at 04:24:45PM +0200, Daniel Vetter wrote:
> On Fri, Oct 9, 2020 at 2:31 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> >
> > On Fri, Oct 09, 2020 at 09:59:31AM +0200, Daniel Vetter wrote:
> >
> > > +struct address_space *iomem_get_mapping(void)
> > > +{
> > > + return iomem_inode->i_mapping;
> >
> > This should pair an acquire with the release below
> >
> > > + /*
> > > + * Publish /dev/mem initialized.
> > > + * Pairs with smp_load_acquire() in revoke_iomem().
> > > + */
> > > + smp_store_release(&iomem_inode, inode);
> >
> > However, this seems abnormal, initcalls rarely do this kind of stuff
> > with global data..
> >
> > The kernel crashes if this fs_initcall is raced with
> > iomem_get_mapping() due to the unconditional dereference, so I think
> > it can be safely switched to a simple assignment.
>
> Ah yes I checked this all, but forgot to correctly annotate the
> iomem_get_mapping access. For reference, see b34e7e298d7a ("/dev/mem:
> Add missing memory barriers for devmem_inode").
Oh yikes, so revoke_iomem can run concurrently during early boot,
tricky.
> The reasons for the annotations is that iomem requests can happen
> fairly early, way before fs_initcalls happen. That means revoke_iomem
> needs to check for that and bail out if we race - nothing bad can
> happen since userspace isn't running at this point anyway. And
> apparently it needs to be a full acquire fence since we don't just
> write a value, but need a barrier for the struct stuff.
Yes, if that is what is happening it release/acquire is needed.
Jason
