Re: [Xen-devel] [PATCH 04/17] xen: Convert virt_to_mfn() and mfn_to_virt() to use typesafe MFN

[Julien Grall <julien@xxxxxxx>]

Hi,

On 26/03/2020 09:09, Jan Beulich wrote:
> On 25.03.2020 19:21, Julien Grall wrote:
> > On 25/03/2020 15:27, Jan Beulich wrote:
> > > On 22.03.2020 17:14, julien@xxxxxxx wrote:
> > > > @@ -785,21 +781,21 @@ bool is_iomem_page(mfn_t mfn)
> > > >        return (page_get_owner(page) == dom_io);
> > > >    }
> > > >    -static int update_xen_mappings(unsigned long mfn, unsigned int cacheattr)
> > > > +static int update_xen_mappings(mfn_t mfn, unsigned int cacheattr)
> > > >    {
> > > >        int err = 0;
> > > > -    bool alias = mfn >= PFN_DOWN(xen_phys_start) &&
> > > > -         mfn < PFN_UP(xen_phys_start + xen_virt_end - XEN_VIRT_START);
> > > > +    bool alias = mfn_x(mfn) >= PFN_DOWN(xen_phys_start) &&
> > > > +         mfn_x(mfn) < PFN_UP(xen_phys_start + xen_virt_end - XEN_VIRT_START);
> > > >        unsigned long xen_va =
> > > > -        XEN_VIRT_START + ((mfn - PFN_DOWN(xen_phys_start)) << PAGE_SHIFT);
> > > > +        XEN_VIRT_START + mfn_to_maddr(mfn_add(mfn, -PFN_DOWN(xen_phys_start)));
> > > Depending on the types involved (e.g. in PFN_DOWN()) this may
> > > or may not be safe, so I consider such a transformation at
> > > least fragile. I think we either want to gain mfn_sub() or
> > > keep this as a "real" subtraction.
> > I want to avoid mfn_x() as much as possible when everything can
> > be done using typesafe operation. But i am not sure how
> > mfn_sub() would solve the problem. Do you mind providing more
> > information?
> Consider PFN_DOWN() potentially returning "unsigned int". The
> negation of an unsigned int is still an unsigned int, and hence
> e.g. -1U (which might result here) is really 0xFFFFFFFF rather
> than -1L / -1UL as intended. Whereas with mfn_sub() the
> conversion to unsigned long of the (positive) value to subtract
> would occur as part of evaluating function arguments, and the
> resulting subtraction would then be correct.
I will have a look to introduce mfn_sub().

> > > > @@ -584,21 +584,21 @@ static unsigned long init_node_heap(int node, unsigned 
> > > > long mfn,
> > > >            needed = 0;
> > > >        }
> > > >        else if ( *use_tail && nr >= needed &&
> > > > -              arch_mfn_in_directmap(mfn + nr) &&
> > > > +              arch_mfn_in_directmap(mfn_x(mfn_add(mfn, nr))) &&
> > > >                  (!xenheap_bits ||
> > > > -               !((mfn + nr - 1) >> (xenheap_bits - PAGE_SHIFT))) )
> > > > +               !((mfn_x(mfn) + nr - 1) >> (xenheap_bits - PAGE_SHIFT))) )
> > > May I suggest consistency here: This one uses +, while ...
> > >
> > > >        {
> > > > -        _heap[node] = mfn_to_virt(mfn + nr - needed);
> > > > -        avail[node] = mfn_to_virt(mfn + nr - 1) +
> > > > +        _heap[node] = mfn_to_virt(mfn_add(mfn, nr - needed));
> > > > +        avail[node] = mfn_to_virt(mfn_add(mfn, nr - 1)) +
> > > >                          PAGE_SIZE - sizeof(**avail) * NR_ZONES;
> > > >        }
> > > >        else if ( nr >= needed &&
> > > > -              arch_mfn_in_directmap(mfn + needed) &&
> > > > +              arch_mfn_in_directmap(mfn_x(mfn_add(mfn, needed))) &&
> > > ... this one uses mfn_add() despite the mfn_x() around it, and ...
> > So the reason I used mfn_x(mfn_add(mfn, needed)) here is I plan
> > to convert arch_mfn_in_directmap() to use typesafe soon. In the
> > two others cases...
> >
> > > >                  (!xenheap_bits ||
> > > > -               !((mfn + needed - 1) >> (xenheap_bits - PAGE_SHIFT))) )
> > > > +               !((mfn_x(mfn) + needed - 1) >> (xenheap_bits - PAGE_SHIFT))) )
> > > ... here you use + again. My personal preference would be to avoid
> > > constructs like mfn_x(mfn_add()).
> > ... I am still unsure how to avoid mfn_x(). Do you have any ideas?
> I don't see how it can be avoided right now. But I also don't see
> why - for consistency, as said - you couldn't use mfn_x() also in
> the middle case. You could then still convert to mfn_add() with
> that future change of yours.
I could have as I could also have converted arch_mfn_in_directmap() to 
use typesafe MFN. Anything around typesafe is a can of worms and this is 
the fine line I found.
Anyway, I could not be bother to bikeshed... So I going to switch the 
other one to mfn_x(...) + needed.
> > > > --- a/xen/include/asm-x86/mm.h
> > > > +++ b/xen/include/asm-x86/mm.h
> > > > @@ -667,7 +667,7 @@ static inline bool arch_mfn_in_directmap(unsigned long mfn)
> > > >    {
> > > >        unsigned long eva = min(DIRECTMAP_VIRT_END, HYPERVISOR_VIRT_END);
> > > >    -    return mfn <= (virt_to_mfn(eva - 1) + 1);
> > > > +    return mfn <= mfn_x(mfn_add(virt_to_mfn(eva - 1),  1));
> > > Even if you wanted to stick to using mfn_add() here, there's one
> > > blank too many after the comma.
> > I will remove the extra blank. Regarding the construction, I have
> > been wondering for a couple of years now whether we should
> > introduce mfn_{lt, gt}. What do you think?
> I too have been wondering, and wouldn't mind their introduction
> (plus mfn_le / mfn_ge perhaps). But it'll truly help you here
> anyway only once the function parameter is also mfn_t.
This is a longer term plan. So I am going to leave it like that for now 
until I manage to find time.
Cheers,

--
Julien Grall