On Thu, 13 May 2021, Julien Grall wrote:
> On Thu, 13 May 2021, 23:32 Stefano Stabellini, <sstabellini@kernel.org> wrote:
>       On Wed, 12 May 2021, Julien Grall wrote:
>       > Hi Stefano,
>       >
>       > On 12/05/2021 23:00, Stefano Stabellini wrote:
>       > > On Sun, 25 Apr 2021, Julien Grall wrote:
>       > > > From: Julien Grall <jgrall@amazon.com>
>       > > >
>       > > > Only the first 2GB of the virtual address space is shared between all
>       > > > the page-tables on Arm32.
>       > > >
>       > > > There is a long outstanding TODO in xen_pt_update() stating that the
>       > > > function is can only work with shared mapping. Nobody has ever called
>       > >             ^ remove
>       > >
>       > > > the function with private mapping, however as we add more callers
>       > > > there is a risk to mess things up.
>       > > >
>       > > > Introduce a new define to mark the ened of the shared mappings and use
>       > >                                       ^end
>       > >
>       > > > it in xen_pt_update() to verify if the address is correct.
>       > > >
>       > > > Note that on Arm64, all the mappings are shared. Some compiler may
>       > > > complain about an always true check, so the new define is not introduced
>       > > > for arm64 and the code is protected with an #ifdef.
>       > >   On arm64 we could maybe define SHARED_VIRT_END to an arbitrarely large
>       > > value, such as:
>       > >
>       > > #define SHARED_VIRT_END (1UL<<48)
>       > >
>       > > or:
>       > >
>       > > #define SHARED_VIRT_END DIRECTMAP_VIRT_END
>       > >
>       > > ?
>       >
>       > I thought about it but I didn't want to define to a random value... I felt not
>       > define it was better.
> 
>       Yeah, I see your point: any restrictions in addressing (e.g. 48bits)
>       are physical address restrictions. Here we are talking about virtual
>       address restriction, and I don't think there are actually any
>       restrictions there?  We could validly map something at
>       0xffff_ffff_ffff_ffff. So even (1<<48) which makes sense at the physical
>       level, doesn't make sense in terms of virtual addresses.
> 
> 
> The limit for the virtual address is 2^64.
> 
> 
> 
>       > > > Signed-off-by: Julien Grall <jgrall@amazon.com>
>       > > >
>       > > > ---
>       > > >      Changes in v2:
>       > > >          - New patch
>       > > > ---
>       > > >   xen/arch/arm/mm.c            | 11 +++++++++--
>       > > >   xen/include/asm-arm/config.h |  4 ++++
>       > > >   2 files changed, 13 insertions(+), 2 deletions(-)
>       > > >
>       > > > diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c
>       > > > index 8fac24d80086..5c17cafff847 100644
>       > > > --- a/xen/arch/arm/mm.c
>       > > > +++ b/xen/arch/arm/mm.c
>       > > > @@ -1275,11 +1275,18 @@ static int xen_pt_update(unsigned long virt,
>       > > >        * For arm32, page-tables are different on each CPUs. Yet, they
>       > > > share
>       > > >        * some common mappings. It is assumed that only common mappings
>       > > >        * will be modified with this function.
>       > > > -     *
>       > > > -     * XXX: Add a check.
>       > > >        */
>       > > >       const mfn_t root = virt_to_mfn(THIS_CPU_PGTABLE);
>       > > >   +#ifdef SHARED_VIRT_END
>       > > > +    if ( virt > SHARED_VIRT_END ||
>       > > > +         (SHARED_VIRT_END - virt) < nr_mfns )
>       > >
>       > > The following would be sufficient, right?
>       > >
>       > >      if ( virt + nr_mfns > SHARED_VIRT_END )
>       >
>       > This would not protect against an overflow. So I think it is best if we keep
>       > my version.
> 
>       But there can be no overflow with the way SHARED_VIRT_END is defined.
> 
>       Even if SHARED_VIRT_END was defined at (1<<48) there can be no overflow.
>       Only if we defined SHARED_VIRT_END as 0xffff_ffff_ffff_ffff we would
>       have an overflow, but you wrote above that your preference is not to do
>       that.
> 
> 
> You can have an overflow regardless the value of SHARED_VIRT_END.
> 
> Imagine virt = 2^64 - 1 and nr_mfs = 1. The addition would result to 0.
> 
> As a consequence the check would pass when it should not.

Yes you are right, I don't know how I missed it!


> One can argue that the caller will always provide sane values. However given the simplicity of the check, this is not worth the trouble if
> a caller is buggy.
> 
> Now, the problem with SHARED_VIRT_END equals to 2^64 - 1 is not the overflow but the compiler that may throw an error/warning for always
> true check. Hence the reason of not defining SHARED_VIRT_END on arm64.

OK, all checks out.

Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>