Re: [Xen-devel] [PATCH v5 3/8] arch, arm: let map_mmio_regions() take pfn as parameters

[Arianna Avanzini <avanzini.arianna@xxxxxxxxx>]

Hello,

thank you for your comments, and sorry for the huge delay in replying.

On 04/09/2014 03:54 PM, Ian Campbell wrote:
> On Mon, 2014-04-07 at 01:31 +0200, Arianna Avanzini wrote:
>> -        res = map_mmio_regions(d, addr & PAGE_MASK,
>> -                               PAGE_ALIGN(addr + size) - 1,
>> -                               addr & PAGE_MASK);
>> +        res = map_mmio_regions(d,
>> +                               paddr_to_pfn(addr & PAGE_MASK),
>> +                               paddr_to_pfn_aligned(addr + size - 1),
> 
> 
> With
> +#define paddr_to_pfn_aligned(paddr)    paddr_to_pfn(PAGE_ALIGN(paddr))
> 
> There is a subtle difference here, which is that the "- 1" is now inside
> the align. Does this always have the same result? I'm not sure.
> 
> If addr+size == 0x1000 (page aligned) then:
> 
>         PAGE_ALIGN(0x10000)-1 = 0x10000-1 = 0xffff
> 
> But
> 
>         paddr_to_pfn_aligned(0x10000 - 1) =
>         paddr_to_pfn(PAGE_ALIGN(0xffff)) = paddr_to_pfn(0x10000) = 0x10
> 
> The new map_mmio_regions uses pfn_to_paddr which is:
> #define pfn_to_paddr(pfn) ((paddr_t)(pfn) << PAGE_SHIFT)
> 
> So with the old code the end address would be 0xffff, while with the new
> code it is 0x10<<12 = 0x10000.
> 
> I suspect the implementation of apply_to_p2m_changes is such that it
> doesn't actually change anything. Can you confirm that this was your
> intention?
> 

It wasn't my intention to change the size of the address range to be mapped,
thank you for pointing that out. As far as I can understand, the changes in this
patch would let apply_p2m_changes() use one extra address for the mapping; sorry
for that.

While preparing the patch I saw that, e.g., with addr + size = 0x10000,

end_gfn = paddr_to_pfn(PAGE_ALIGN(0x10000) - 1) =
          paddr_to_pfn(PAGE_ALIGN(0x10000)) - 1 = 0xf

and

pfn_to_paddr(end_gfn) = 0xf000

which seemed to me to be wrong, as the previous end address was, as you wrote,
0xffff.

Instead, if

end_gfn = paddr_to_pfn(PAGE_ALIGN(0x10000 - 1)) = 0x10

then

pfn_to_paddr(end_gfn) = 0x10000

which I thought lets the needed address range be mapped; however I didn't see
what you pointed out, that it also lets one extra address be used for the 
mapping.

> Is the end argument tio map_mmio_regions now intended to be inclusive or
> exclusive? 

As far as I understand, apply_p2m_changes(), which is called by the ARM version
of map_mmio_regions(), seems to take it as exclusive, as the mapping is
performed while (addr < end_gpaddr).
In the x86 version of map_mmio_regions() it is instead intended to be inclusive;
sorry for this mismatch, I'll try to make the behavior of the two versions
consistent.

> This sort of issue can be avoided by using a count instead of
> an end in the interface.
> 

Thank you very much for the hint.


> Ian.
> 



-- 
/*
 * Arianna Avanzini
 * avanzini.arianna@xxxxxxxxx
 * 73628@xxxxxxxxxxxxxxxxxxx
 */

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel