Re: Buggy interaction of live migration and p2m updates

[Ian Campbell <Ian.Campbell@citrix.com>]

On Fri, 2014-11-21 at 12:20 +0100, Juergen Gross wrote:
> On 11/21/2014 12:15 PM, Ian Campbell wrote:
> > On Fri, 2014-11-21 at 11:07 +0000, Andrew Cooper wrote:
> >> On 21/11/14 10:46, Ian Campbell wrote:
> >>> On Fri, 2014-11-21 at 10:24 +0000, Andrew Cooper wrote:
> >>>> On 21/11/14 09:43, Ian Campbell wrote:
> >>>>> I don't see any (explicit) mention of the pfn_to_mfn_frame_list_list
> >>>>> here, where does that fit in?
> >>>>>
> >>>> It is referenced several times, although not by its exact name.
> >>> Hence no explicit mention.
> >>>
> >>> It's ambiguous when you refer to "higher level frames" (which I presume
> >>> are the reference you are referring to) because some kernels (perhaps
> >>> only historic ones, I've not been keeping up) keep both an N-level tree
> >>> of their own internally and the toolstack visible frame_list_list
> >>> (sometimes partially overlapping at some level). Is every reference to
> >>> "higher level frames" actually intended to be a reference to
> >>> pfn_to_mfn_frame_list_list or not?
> >>
> >> "higher level frames" would be the toolstack-abi-defined first and
> >> second level lists.  The logdirty infrastructure can be used to detect
> >> writes to these frames, and therefore detect structural changes to the p2m.
> >>
> >> I would like to hope that every kernel out there keeps this information
> >> correctly up-to-date and updates it in an appropriate order...
> >>
> >>>
> >>> It seems like sometimes you are talking at times about tracking the
> >>> kernel's internal structure and not just pfn_to_mfn_frame_list_list and
> >>> I'm not sure why that would be.
> >>
> >> I apologise for giving this impression.  It was not intended.
> >
> > Great, I just wanted to be sure we were all on the same page, since
> > scrobbling around in the kernel's internal data structures would clearly
> > be mad...
> >
> >>
> >>> I'm also not sure why
> >>> pfn_to_mfn_frame_list_list is apparently discounted in the linear case,
> >>> AFAIK the guest is still obliged to keep that up to date regardless of
> >>> the scheme it uses internally for accessing the p2m.
> >>
> >> There are two reasons for the virtual linear p2m, the primary one being
> >> to break the hard 512GB limit given the old 3-level table.
> >>
> >> A 64bit PV guest cannot possibly use the pfn_to_mfn_frame_list_list if
> >> it needs to actually exceed 512GB of RAM.  Therefore, to signal the use
> >> the virtual linear method, a PV guest explicitly sets
> >> pfn_to_mfn_frame_list_list to INVALID_MFN, and fills in the brand new
> >> adjacent information.
> >
> > Oh, I hadn't realised this linear p2m stuff involved a guest ABI change.
> > Have I somehow completely missed the xen.git side of these patches? I
> > thought I'd only seen linux.git ones (and hence wasn't looking very
> > closely).
> 
> V1 of the patches suggesting such a change have been posted a week ago:
> 
> http://lists.xen.org/archives/html/xen-devel/2014-11/msg01276.html

Ah, thanks. I had indeed ignored that as "just another iteration of the
linux patches", oops!

> The linear p2m stuff is a prerequisite for this change, not the reason
> for it.

Ian.