Re: [Xen-devel] [PATCH v2 COLOPre 04/13] tools/libxc: export xc_bitops.h

[Ian Campbell <ian.campbell@xxxxxxxxxx>]

On Thu, 2015-06-11 at 11:45 +0100, Andrew Cooper wrote:
> On 11/06/15 09:41, Ian Campbell wrote:
> > On Thu, 2015-06-11 at 10:07 +0800, Yang Hongyang wrote:
> >> On 06/10/2015 11:20 PM, Ian Campbell wrote:
> >>> On Mon, 2015-06-08 at 11:43 +0800, Yang Hongyang wrote:
> >>>> When we are under COLO, we will send dirty page bitmap info from
> >>>> secondary to primary at every checkpoint.
> >>> ... and this is a _libxl_ operation? Is that the right layer here?
> >> For the first question, Yes, this is done in the suspend callback on
> >> restore side. We do this in libxl because currently we only added a
> >> back channel on libxl side. There're no back channel in libxc.
> >>
> >> By considering this more, if we do this in libxc part, the code will be
> >> less complex: we can drop the 4th & 9th patch of this series and also
> >> get rid of the get_dirty_pfn() callback. instead we will add a patch to
> >> add back channel in libxc.
> > That sounds better to me, but lets see what Andrew thinks.
> >
> >> For the second question, I'm not sure, what's Andrew's opinion? which
> >> is the right layer to do this operation, libxl or libxc?
> 
> There are a number of bits of information which would be useful going in
> "the backchannel".
> 
> Some are definitely more appropriate at the libxc level, but others are
> more appropriate at the libxl.
> 
> If you recall from the hackathon, there was an Alibaba usecase where
> they wanted a positive success/fail from the receiving side that the VM
> has started up successfully before choosing between cleaning up or
> continuing the VM on the sending side.  This would have to be a libxl
> level backchannel.

FWIW this particular case is currently an xl level backchannel, but I
think your general point stands.

> Whatever happens, backchannel wise, it should be a sensibly
> type/length/chunk'd stream.  (I think there is a spec or two floating
> around somewhere which might be a good start ;p)  There should probably
> be a bit of active negotiation at the start of the backchannel to a)
> confirm you have the correct backchannel and b) the backchannel is
> actually functioning.
> 
> The data on "the backchannel" is always going to be in reply to an
> action taking place in the primary channel, but there are complications
> in that the libxc bit is inherently a blocking model.  In terms of
> coordination, I am leaning towards the view of it being easier and
> cleaner for each level to maintain its own backchannel communication. 
> The libxc bits can expect to read some records out of the backchannel at
> each checkpoint and take appropriate actions before starting the next
> checkpoint.
> 
> Thoughts?
> 
> ~Andrew
> 



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel