Re: [PATCHv6 5/5] gnttab: use per-VCPU maptrack free lists

[Malcolm Crossley <malcolm.crossley@citrix.com>]

On 24/04/15 10:50, Jan Beulich wrote:
>>>> On 24.04.15 at 11:09, <malcolm.crossley@citrix.com> wrote:
>> On 23/04/15 17:11, Jan Beulich wrote:
>>>>>> On 22.04.15 at 18:00, <david.vrabel@citrix.com> wrote:
>>>> --- a/xen/include/xen/grant_table.h
>>>> +++ b/xen/include/xen/grant_table.h
>>>> @@ -60,6 +60,8 @@ struct grant_mapping {
>>>>      u32      ref;           /* grant ref */
>>>>      u16      flags;         /* 0-4: GNTMAP_* ; 5-15: unused */
>>>>      domid_t  domid;         /* granting domain */
>>>> +    u32      vcpu;          /* vcpu which created the grant mapping */
>>>> +    u16      pad[2];
>>>>  };
>>>
>>> What is this pad[] good for?
>>
>> The pad is to keep the struct power of 2 sized because this allows the
>> compiler to optimise these macro's to right and left shifts:
>>
>> #define MAPTRACK_PER_PAGE (PAGE_SIZE / sizeof(struct grant_mapping))
>> #define maptrack_entry(t, e) \
>>     ((t)->maptrack[(e)/MAPTRACK_PER_PAGE][(e)%MAPTRACK_PER_PAGE])
> 
> Okay, then why u16[2] instead of u32? And please add a brief
> comment explaining the reason.
> 
> Apart from that I wonder whether fitting vcpu in the 10 unused
> flags bits (not 11, as the comment on the field suggests) would be
> an option. That would require limiting vCPU count to 4k, which I
> don't think would really be a problem for anyone.

I'd like to not have the overhead of bit operations on a hot path.

We're going from 512 grant entries per page to 256 grant entries per
page. This is taking the Xen memory overhead from 0.2% to 0.4% for grant
mapped pages.

Is the extra 0.2% memory overhead that concerning?

Malcolm
> 
> Jan
>