Re: Xen blktap driver for Ceph RBD : Anybody wants to test ? :p

From: Wei Liu <wei.liu2@citrix.com>
To: George Dunlap <george.dunlap@eu.citrix.com>
Cc: James Harper <james.harper@bendigoit.com.au>,
	Wei Liu <Wei.Liu2@citrix.com>,
	Stefano Stabellini <stefano.stabellini@eu.citrix.com>,
	Sylvain Munaut <s.munaut@whatever-company.com>,
	"xen-devel@lists.xen.org" <xen-devel@lists.xen.org>,
	"ceph-devel@vger.kernel.org" <ceph-devel@vger.kernel.org>,
	Roger Pau Monne <roger.pau@citrix.com>
Subject: Re: Xen blktap driver for Ceph RBD : Anybody wants to test ? :p
Date: Mon, 5 Aug 2013 16:18:14 +0100	[thread overview]
Message-ID: <20130805151814.GG27858__27537.2225013145$1375716003$gmane$org@zion.uk.xensource.com> (raw)
In-Reply-To: <51FFB0FF.7070607@eu.citrix.com>

On Mon, Aug 05, 2013 at 03:04:47PM +0100, George Dunlap wrote:
> On 05/08/13 14:55, Sylvain Munaut wrote:
> >Hi George,
> >
> >
> >>Yes; qemu knows how to be a Xen PV block back-end.
> >Very interesting. Is there documentation about this somewhere ?
> >I had a look some time ago and it was really not very clear.
> >
> >Things like what Xen version support this. And with which features (
> >indirect descriptors, persistent grants, discard, flush, ...) and/or
> >which limitation.
> 
> I don't think this is documented anywhere; you'll need to ask the
> experts.  Stefano? Roger? Wei?
> 

These are Linux features not Xen ones AFAICT. In theory they are not
bound to specific Xen versions.

For the network part I don't think new features depend on any specific
hypercall. However for block Roger and Stefano seem to introduce
new hypercalls for certain features (I might be wrong though).

Wei.

> >
> >
> >>One of the reasons for stopping work on blktap3 (AIUI) was that it
> >>should in theory have performance characteristics similar to blktap3,
> >And did anyone check the theory currently ? :)
> 
> I say "in theory" because they are using the same basic
> architecture: a normal process running in dom0, with no special
> kernel support.  If there were a performance difference, it would be
> something that should (in theory) be able to be optimized.
> 
> I don't think we have comparisons between qdisk (which is what we
> call qemu-as-pv-backend in Xen) and blktap3 (and since blktap3
> wasn't finished they wouldn't mean much anyway); but I think qdisk
> compares reasonably with blkback.
> 
>  -George