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

[Jan Beulich <jbeulich@suse.com>]

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@xen.org 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.

>>> @@ -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.

>>> --- 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.

Jan