Re: [Xen-devel] [PATCH] libbxl: add support for pvscsi, iteration 1

[Ian Campbell <Ian.Campbell@xxxxxxxxxx>]

On Fri, 2014-05-02 at 16:36 +0200, Olaf Hering wrote:
> On Wed, Apr 30, Olaf Hering wrote:
> 
> > Add support for vscsi= in domU.cfg, add new xl commands scsi-attach,
> > scsi-list, scsi-detach. The syntax follows what xend understood.
> > 
> > pvscsi provides a dom0 SCSI device as-is to a domU. The backend and
> > frontend drivers can be found in xen-linux, or its forward-ported
> > variants. This patch was tested with an openSUSE dom0 and a SLES12
> > guest. Any openSUSE/SLE11/SLE12 dom0/domU combination will work.
> 
> One question regarding device index handling in xenstore:
> 
> If devices get removed and added at runtime, they will appear as
> /local/domain/0/backend/<devtype>/<domid>/<index>, like
> /local/domain/0/backend/vif/3/0.
> 
> What is supposed to happen with index if, in this example,
> network-attach adds another device, network-detach removes it again,
> later network-attach adds yet another device? The first attach will most
> likely use index==1. network-detach will remove the device with index 1.
> Is the last network-attach supposed to use index 2, or is it allowed to
> reuse index 1? Right now I'm not seeing an index counter for
> /local/domain/0/backend/vif/3, so I assume it will pick 1.

The function to look at is libxl__device_nextid, which I think is buggy
in the scenario you describe since it only looks at the number of
entries, not which ones are used.

> While working on scsi-attach/scsi-detach I was wondering if the new code
> should maintain some counter so that new vscsi hosts get a new number
> which was not in use before. Same for vscsi devs. It looks like xend
> maintained an internal global counter for (at least) all devs. Surely
> such counter was easy to maintain for xend as it controls everything.
> 
> If no such counter is required I have to assume that by the time the
> toolstack reuses index 1 everything in backend/frontend has released all
> references to index 1. No races should happen.
> 
> Is that assumption correct for all kind of backend devices?

I'm not sure.

Perhaps libxl__device_nextid should use a xenstore transaction to read ,
increment and write a suitable counter, perhaps under /libxl?

Ian.


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