Re: pv_ops & gntdev?

[Jeremy Fitzhardinge <jeremy@goop.org>]

Gerd Hoffmann wrote:
>> Would the unmap operation on the vma be enough to do that?   Hm, I guess
>> vm_operations_struct needs something that gets called before the zap
>> rather than after (->close).
>>     
>
> mmu motifier will fit the bill and additionally work fine on partial unmaps.
>   

OK, it'll be interesting to see how that turns out.

>>> The (unclean) exit case might be tricky though.  Could be the kernel
>>> tries to unpin before zapping all mappings, so we don't trap into xen
>>> all the time.
>>>       
>> Yes, the unpin needs to make sure it clears out all the grant refs,
>> because you just can't unpin otherwise.
>>
>> The alternative is to hook into the fault handler so that a
>> fault-n-emulate operation on the pagetable which fails because of the
>> grant reference can be fixed up at that point.  But that might be even
>> more intrusive.
>>     
>
> Or just allow to flag "skip the unpin please for this mm".  It's an
> optimization after all, it isn't required, right?
>   

If its pinned you have to unpin; you can't free pagetable pages while 
they're pinned.  If you don't pin the pagetable then context switches 
get very slow (it would need to revalidate the whole pagetable every 
context switch).

We could, however, defer the unpin and unpin late, after all the vmas 
have been unmapped.  But that's a bit ugly (it would require a separate 
hook in exit_mmap).

    J