Re: [Xen-devel] [PATCH v3] xen/grant-table: Avoid m2p_override during mapping

[Zoltan Kiss <zoltan.kiss@citrix.com>]

On 21/01/14 14:22, Stefano Stabellini wrote:
> On Tue, 21 Jan 2014, David Vrabel wrote:
>> On 21/01/14 12:26, Stefano Stabellini wrote:
>>> On Mon, 20 Jan 2014, Zoltan Kiss wrote:
>>>
>>>> -		ret = m2p_add_override(mfn, pages[i], kmap_ops ?
>>>> -				       &kmap_ops[i] : NULL);
>>>> +		if (m2p_override)
>>>> +			ret = m2p_add_override(mfn, pages[i], kmap_ops ?
>>>> +					       &kmap_ops[i] : NULL);
>>>> +		else {
>>>> +			unsigned long pfn = page_to_pfn(pages[i]);
>>>> +			WARN_ON(PagePrivate(pages[i]));
>>>> +			SetPagePrivate(pages[i]);
>>>> +			set_page_private(pages[i], mfn);
>>>> +			pages[i]->index = pfn_to_mfn(pfn);
>>>> +			if (unlikely(!set_phys_to_machine(pfn, FOREIGN_FRAME(mfn))))
>>>> +				return -ENOMEM;
>>>
>>> What happens if the page is PageHighMem?
>>>
>>> This looks like a subset of m2p_add_override, but it is missing some
>>> relevant bits, like the PageHighMem check, or the p2m(m2p(mfn)) == mfn
>>> check.  Maybe we can find a way to avoid duplicating the code.
>>> We could split m2p_add_override in two functions or add yet another
>>> parameter to it.
>>
>> The PageHighMem() check isn't relevant as we're not mapping anything
>> here.  Also, a page for a kernel grant mapping only cannot be highmem.
>>
>> The check for a local mfn and the additional set_phys_to_machine() is
>> only necessary if something tries an mfn_to_pfn() on the local mfn.  We
>> can only omit adding an m2p override if we know thing will try
>> mfn_to_pfn(), therefore the check and set_phys_to_machine() is unnecessary.
>
> OK, you convinced me that the two checks are superfluous for this case.
>
> Can we still avoid the code duplication by removing the corresponding
> code from m2p_add_override and m2p_remove_override and doing the
> set_page_private thing uniquely in grant-table.c?
Yes, I moved these parts out from the m2p* funcions to the gntmap 
functions. One change is that now we pass pfn/mfn to m2p* functions as 
they are changing right before the call. Also, to avoid racing I clear 
the page->private value before calling m2p_remove_override. I'll send in 
the patch soon.

Zoli