Re: [PATCH v2 09/12] x86/altp2m: add remaining support routines.

[Razvan Cojocaru <rcojocaru@bitdefender.com>]

On 06/25/2015 07:48 PM, Ed White wrote:
> On 06/25/2015 06:40 AM, Razvan Cojocaru wrote:
>> On 06/25/2015 03:44 PM, Lengyel, Tamas wrote:
>>> On Wed, Jun 24, 2015 at 2:06 PM, Ed White <edmund.h.white@intel.com
>>> <mailto:edmund.h.white@intel.com>> wrote:
>>>     On 06/24/2015 09:15 AM, Lengyel, Tamas wrote:
>>>     >> +bool_t p2m_set_altp2m_mem_access(struct domain *d, uint16_t idx,
>>>     >> +                                 unsigned long pfn, xenmem_access_t
>>>     >> access)
>>>     >> +{
>>>     >>
>>>     >
>>>     > This function IMHO should be merged with p2m_set_mem_access and should be
>>>     > triggerable with the same memop (XENMEM_access_op) hypercall instead of
>>>     > introducing a new hvmop one.
>>>
>>>     I think we should vote on this. My view is that it makes
>>>     XENMEM_access_op
>>>     too complicated to use.
>>>
>>> The two functions are not very long and share enough code that it would
>>> justify merging. The only big change added is the copy from host->alt
>>> when the entry doesn't exists in alt, and that itself is pretty self
>>> contained. Let's see if we can get a third opinion on it..
>>
>> At first sight (I admit I'm rather late in the game and haven't had a
>> chance to follow the series closely from the beginning), the two
>> functions do seem to be mergeable (or at least the common code factored
>> out in static helper functions).
>>
>> Also, if Ed's concern is that the libxc API would look unnatural if
>> xc_set_mem_access() is used for both purposes, as far as I can tell the
>> only difference could be a non-zero last altp2m parameter, so I agree
>> with you that the less functions doing almost the same thing the better
>> (I have been guilty of this in the past too, for example with my
>> xc_enable_introspection() function ;) ).
>>
>> So I'd say, yes, if possible merge them.
> 
> So here are my reasons why I don't think we should merge the hypercalls,
> in more detail:
> 
> Although the two hypercalls are similar, they are not identical. For one
> thing, the existing hypercall can only be used cross-domain whereas the
> altp2m one can be used cross-domain or intra-domain. Also, the existing
> hypercall can be used to modify a range of pages and the new one can only
> modify a single page, and that is intentional.
> 
> As I see it, the implementation in hvm.c would become a lot less clean,
> and every direct user of the existing hypercall would have to change for
> no good reason.

Thank you for the explanation. While it could be argued that a non-zero
altp2m parameter passed to a merged xc_set_mem_access() could be the
xc_set_altp2m_mem_access() selector, and that the function can then
return EINVAL for parameters that don't fit the semantics of the
selected behaviour, I also don't have a strong aversion to those
functions not being merged. So I'll defer this to Tamas.

> Razvan's suggestion to merge the functions that implement the p2m changes
> I'm more ambivalent about. Personally, I prefer not to have code that
> contains lots of conditional logic, which would be the result, but I
> don't feel that strongly about it.

Well, not necessarily merge the functions, but at least have as much
common code as possible factored out in helper static functions that
both of them call.


Thanks,
Razvan