xen pv hugepages

xen pv hugepages

[Elena Ufimtseva]

Hello

As I can see there is no support for hugepages in Xen for PV kernels.
Is this something in the future plans or there is no really need to
have this implemented?

I came across this paper https://www.kernel.org/doc/ols/2011/ols2011-gadre.pdf,
not sure if this was somehow presented to xen community, I can't find
any code related to this.

Thank you!
-- 
Elena

Re: xen pv hugepages

[Konrad Rzeszutek Wilk]

On February 6, 2014 5:17:44 PM EST, Elena Ufimtseva <ufimtseva@gmail.com> wrote:
>Hello
>
>As I can see there is no support for hugepages in Xen for PV kernels.
>Is this something in the future plans or there is no really need to
>have this implemented?

It was implemented. The UEK2 kernel (Oracle's) and SLES (I think) has this implemented. But making it upstream means adding some Xen specific code in the generic code which is a no-go.

If you use PVH you get it for free. With the patches that Mukesh posted to enable some cpuid flags I was able to boot a guest with 2MB and 1GB pages without any trouble.
>
>I came across this paper
>https://www.kernel.org/doc/ols/2011/ols2011-gadre.pdf,
>not sure if this was somehow presented to xen community, I can't find
>any code related to this.
>
>Thank you!

Re: xen pv hugepages

[Elena Ufimtseva]

On Thu, Feb 6, 2014 at 6:59 PM, Konrad Rzeszutek Wilk
<konrad.wilk@oracle.com> wrote:
> On February 6, 2014 5:17:44 PM EST, Elena Ufimtseva <ufimtseva@gmail.com> wrote:
>>Hello
>>
>>As I can see there is no support for hugepages in Xen for PV kernels.
>>Is this something in the future plans or there is no really need to
>>have this implemented?
>
> It was implemented. The UEK2 kernel (Oracle's) and SLES (I think) has this implemented. But making it upstream means adding some Xen specific code in the generic code which is a no-go.
>
> If you use PVH you get it for free. With the patches that Mukesh posted to enable some cpuid flags I was able to boot a guest with 2MB and 1GB pages without any trouble.

Ok, I see. Well, I would think then that the kernel then should not
provide any possibility to use it? Otherwise there are oopses and some
other unpleasant things happening when trying to use huge pages. I am
not sure, maybe if its clearly states that there is no hugetlb support
for pv, then there is no need for this?

>>
>>I came across this paper
>>https://www.kernel.org/doc/ols/2011/ols2011-gadre.pdf,
>>not sure if this was somehow presented to xen community, I can't find
>>any code related to this.
>>
>>Thank you!
>
>



-- 
Elena

Re: xen pv hugepages

[Jan Beulich]

>>> On 07.02.14 at 01:12, Elena Ufimtseva <ufimtseva@gmail.com> wrote:
> Ok, I see. Well, I would think then that the kernel then should not
> provide any possibility to use it? Otherwise there are oopses and some
> other unpleasant things happening when trying to use huge pages. I am
> not sure, maybe if its clearly states that there is no hugetlb support
> for pv, then there is no need for this?

One of the downsides of pv-ops: You can't hide functionality
known to not work under Xen already at kernel configuration
time, and hence need to find ways to suppress their use at
run time if enabled for the purpose of using it when run on
bare hardware.

Btw., if you're after the underlying code in the hypervisor,
look for the uses of "opt_allow_superpage".

Jan

Re: xen pv hugepages

[Konrad Rzeszutek Wilk]

On Thu, Feb 06, 2014 at 07:12:24PM -0500, Elena Ufimtseva wrote:
> On Thu, Feb 6, 2014 at 6:59 PM, Konrad Rzeszutek Wilk
> <konrad.wilk@oracle.com> wrote:
> > On February 6, 2014 5:17:44 PM EST, Elena Ufimtseva <ufimtseva@gmail.com> wrote:
> >>Hello
> >>
> >>As I can see there is no support for hugepages in Xen for PV kernels.
> >>Is this something in the future plans or there is no really need to
> >>have this implemented?
> >
> > It was implemented. The UEK2 kernel (Oracle's) and SLES (I think) has this implemented. But making it upstream means adding some Xen specific code in the generic code which is a no-go.
> >
> > If you use PVH you get it for free. With the patches that Mukesh posted to enable some cpuid flags I was able to boot a guest with 2MB and 1GB pages without any trouble.
> 
> Ok, I see. Well, I would think then that the kernel then should not
> provide any possibility to use it? Otherwise there are oopses and some
> other unpleasant things happening when trying to use huge pages. I am
> not sure, maybe if its clearly states that there is no hugetlb support
> for pv, then there is no need for this?

I was under the impression that the hugetlb option only gets enabled
if the 'cpu_has_pse' bit is set. And for PV that is cleared so it won't
be set.

> 
> >>
> >>I came across this paper
> >>https://www.kernel.org/doc/ols/2011/ols2011-gadre.pdf,
> >>not sure if this was somehow presented to xen community, I can't find
> >>any code related to this.
> >>
> >>Thank you!
> >
> >
> 
> 
> 
> -- 
> Elena

Re: xen pv hugepages

[Jan Beulich]

>>> On 07.02.14 at 00:59, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> On February 6, 2014 5:17:44 PM EST, Elena Ufimtseva <ufimtseva@gmail.com> wrote:
>>As I can see there is no support for hugepages in Xen for PV kernels.
>>Is this something in the future plans or there is no really need to
>>have this implemented?
> 
> It was implemented. The UEK2 kernel (Oracle's) and SLES (I think) has this 
> implemented.

No, neither SLES nor openSUSE ever did (I guess it's tmem you're
mixing this up with) - the hypervisor side implementation just
doesn't look to be ready for other than experimental use.

Jan