[Hackathon] Linux session notes

[Hackathon] Linux session notes

[David Vrabel]

[Thanks to Malcolm Crossley for taking notes.]

= Overview =

David Vrabel outlined different kernel modes, their status and
suggested focus:

  PV dom0:  production ready - fixes and new hardware support (e.g. EFI)
  PV domU:  production ready - fixes
  PVHVM:    production ready - fixes and new features
  PVH:      experimental     } the future.
  PVH dom0: not yet          }

David would like PV to be deprecated -- no new PV only features;
features must work for PVH or PVHVM as well.

Jan Beulich said that SuSe planned to jump from classic kernel
straight to PVH.

Konrad Rzeszutek Wilk noted that PV MMU ops are planned to be removed
some time (suggested 5 years) after PVH support is complete.

= Broken Things =

Konrad noted that PAT is broken which makes graphics slow.

512 GiB PV guest limit. The fix (4-level p2m) is present in Xen (but
not tools for save/restore?) but PVHVM or PVH is recommended instead
of extending PV.

In guest kexec when PV drivers are used does not work -- grants and
event channels are not torn down which prevents them being used in the
exec'd guest.

= Plans =

David:

- Fixing m2p override.  Mapping the same MFN (by two grant refs or the
  same) two or more times means get_user_pages() cannot find the right
  page() since m2p override is many-to-one.

- Fixing page-to-gref (needed for grant copy in netback to-guest)
  which is currently netback only (this breaks domU providing an iSCSI
  target to be used via blkfront/blkback by another domU in the same
  host.  Probably by adding a generic struct page extension mechanism.

- Testing (mostly dom0) in XenServer's test system.

Konrad:

- Performance regression testing.
- Microcode loading at run time.
- pciback: more useful SBR fallback if FLR isn't present.

Malcolm:

- PV IOMMU

Daniel Kiper:

- EFI for dom0 (see series on list)

Wei Liu:

- VNUMA for PV and PVH (picking up Elena's work).

Re: [Hackathon] Linux session notes

[Jan Beulich]

>>> On 02.06.14 at 17:36, <david.vrabel@citrix.com> wrote:
> Jan Beulich said that SuSe planned to jump from classic kernel
> straight to PVH.

I didn't say we plan to, I said we might.

> 512 GiB PV guest limit. The fix (4-level p2m) is present in Xen (but
> not tools for save/restore?) but PVHVM or PVH is recommended instead
> of extending PV.

This, btw, also places a limit on kexec/kdump of Dom0 (since that iirc
also involves reading the P2M table and its [multi-level] MFNs). But
hopefully people won't run Dom0 with 512Gb or more anyway...

Jan

Re: [Hackathon] Linux session notes

[Stefano Stabellini]

On Mon, 2 Jun 2014, David Vrabel wrote:
> - Fixing m2p override.  Mapping the same MFN (by two grant refs or the
>   same) two or more times means get_user_pages() cannot find the right
>   page() since m2p override is many-to-one.

Did anybody proposed any ways to fix this issue?
We have the same problem on ARM in p2m.c and I cannot see any way to
solve the issue without slowing down the mfn to pfn conversion too much.

Re: [Hackathon] Linux session notes

[David Vrabel]

On 04/06/14 12:11, Stefano Stabellini wrote:
> On Mon, 2 Jun 2014, David Vrabel wrote:
>> - Fixing m2p override.  Mapping the same MFN (by two grant refs or the
>>   same) two or more times means get_user_pages() cannot find the right
>>   page() since m2p override is many-to-one.
> 
> Did anybody proposed any ways to fix this issue?
> We have the same problem on ARM in p2m.c and I cannot see any way to
> solve the issue without slowing down the mfn to pfn conversion too much.

I was planning to add an array of pages to the vma and use this in
get_user_pages() (instead of the page table walk).

David

Re: [Hackathon] Linux session notes

[Stefano Stabellini]

On Wed, 4 Jun 2014, David Vrabel wrote:
> On 04/06/14 12:11, Stefano Stabellini wrote:
> > On Mon, 2 Jun 2014, David Vrabel wrote:
> >> - Fixing m2p override.  Mapping the same MFN (by two grant refs or the
> >>   same) two or more times means get_user_pages() cannot find the right
> >>   page() since m2p override is many-to-one.
> > 
> > Did anybody proposed any ways to fix this issue?
> > We have the same problem on ARM in p2m.c and I cannot see any way to
> > solve the issue without slowing down the mfn to pfn conversion too much.
> 
> I was planning to add an array of pages to the vma and use this in
> get_user_pages() (instead of the page table walk).

That is a bit too concise.
Does that mean you are not going to fix the mfn to pfn conversion
problem and instead change the common code implementation of
get_user_pages?

Re: [Hackathon] Linux session notes

[David Vrabel]

On 04/06/14 12:37, Stefano Stabellini wrote:
> On Wed, 4 Jun 2014, David Vrabel wrote:
>> On 04/06/14 12:11, Stefano Stabellini wrote:
>>> On Mon, 2 Jun 2014, David Vrabel wrote:
>>>> - Fixing m2p override.  Mapping the same MFN (by two grant refs or the
>>>>   same) two or more times means get_user_pages() cannot find the right
>>>>   page() since m2p override is many-to-one.
>>>
>>> Did anybody proposed any ways to fix this issue?
>>> We have the same problem on ARM in p2m.c and I cannot see any way to
>>> solve the issue without slowing down the mfn to pfn conversion too much.
>>
>> I was planning to add an array of pages to the vma and use this in
>> get_user_pages() (instead of the page table walk).
> 
> That is a bit too concise.
> Does that mean you are not going to fix the mfn to pfn conversion
> problem and instead change the common code implementation of
> get_user_pages?

Yes.  It's impossible to fix mfn_to_pfn() for foreign MFNs.

David

Re: [Hackathon] Linux session notes

[Ian Campbell]

On Wed, 2014-06-04 at 12:42 +0100, David Vrabel wrote:
> On 04/06/14 12:37, Stefano Stabellini wrote:
> > On Wed, 4 Jun 2014, David Vrabel wrote:
> >> On 04/06/14 12:11, Stefano Stabellini wrote:
> >>> On Mon, 2 Jun 2014, David Vrabel wrote:
> >>>> - Fixing m2p override.  Mapping the same MFN (by two grant refs or the
> >>>>   same) two or more times means get_user_pages() cannot find the right
> >>>>   page() since m2p override is many-to-one.
> >>>
> >>> Did anybody proposed any ways to fix this issue?
> >>> We have the same problem on ARM in p2m.c and I cannot see any way to
> >>> solve the issue without slowing down the mfn to pfn conversion too much.
> >>
> >> I was planning to add an array of pages to the vma and use this in
> >> get_user_pages() (instead of the page table walk).

FWIW the classic-XenoLinux patches did something similar (if not
identical) to this.
 
> > That is a bit too concise.
> > Does that mean you are not going to fix the mfn to pfn conversion
> > problem and instead change the common code implementation of
> > get_user_pages?
> 
> Yes.  It's impossible to fix mfn_to_pfn() for foreign MFNs.
> 
> David
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel