Re: [RFC PATCH 2/3] arch, arm32: add the XEN_DOMCTL_memory_mapping hypercall

[Julien Grall <julien.grall@linaro.org>]

Hello Arianna,

On 07/03/14 08:56, Arianna Avanzini wrote:
>>> Beware:  I found that map_mmio_regions was off by one page unless I changed
>>> "gfn + nr_mfns - 1" to "gfn + nr_mfns".
>>> This may have been fixed in more recent upstream code.
>>
>> That makes me think, map_mmio_regions is buggy if the range is not correctly
>> aligned. For instance:
>>
>> start_gaddr = 0x1001 and end_gaddr = 0x2001, we only map 0x1000-0x2000 to the
>> domain.
>>
>> The problem is the same with p2m_populate_ram (even if I doubt that RAM range is
>> not page-aligned).
>>
>> I see 2 solutions to fix it:
>>      - Change map_mmio_regions prototype to use frame is instead address.
>>          - Aligned the input addresses.
>>
>> IHMO, the former seems to be better, at least it avoid developer to discover
>> that some bits was mapped under his feet.
>>
>
> Sorry if I intrude here, thank you for pointing that out. As of now I have tried
> to simply pass aligned addresses to map_mmio_regions() when it is invoked by the
> memory_mapping hypercall.

FYI, the end address should be PAGE-aligned minus 1. Otherwise you will 
map a spurious page to the guest.

> Sorry for the additional question, do you think that it would be appropriate to
> add to the series a patch implementing one of the above-described solutions (the
> one believed to be better)?

I'd like to have some input from Ian before. I think for now it's fine 
to stay with the current implementation.

Regards,

-- 
Julien Grall