missing unplug of SCSI devices in HVM guest

missing unplug of SCSI devices in HVM guest

[Olaf Hering]

[-- Attachment #1.1: Type: text/plain, Size: 932 bytes --]

Does anyone remember why the the vbd frontend drivers also claim the
SCSI disks, but the vbd backend in qemu has no unplug support for SCSI?

The current situation for qemu-xen and qemu-xen-traditional is that both
will create an emulated LSI controller with disk=[vdev=sda]. The
xenlinux and pvops frontend drivers will claim the disk, but the
sym53c8xx will see and claim it as well. As a result each disk is
visible twice. One has to blacklist the sym53c8xx driver to avoid that.

What should be done to fix this?
#1 new unplug protocol for SCSI, but old guests dont know about it
#2 reuse IDE flag to also unplug SCSI. This would cover pvops and guests
   where xenlinux based xen-platform-pci.ko is loaded before
   sym53c8xx.ko. It would break guests with frontend drivers that do not
   claim SCSI disks, the SCSI disk would disappear (if such frontends
   really exist).
#3 do not provide SCSI if guest has PV drivers

Olaf

[-- Attachment #1.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 181 bytes --]

[-- Attachment #2: Type: text/plain, Size: 127 bytes --]

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

Re: missing unplug of SCSI devices in HVM guest

[George Dunlap]

On Wed, Aug 24, 2016 at 10:24 AM, Olaf Hering <olaf@aepfle.de> wrote:
> Does anyone remember why the the vbd frontend drivers also claim the
> SCSI disks, but the vbd backend in qemu has no unplug support for SCSI?
>
> The current situation for qemu-xen and qemu-xen-traditional is that both
> will create an emulated LSI controller with disk=[vdev=sda]. The
> xenlinux and pvops frontend drivers will claim the disk, but the
> sym53c8xx will see and claim it as well. As a result each disk is
> visible twice. One has to blacklist the sym53c8xx driver to avoid that.
>
> What should be done to fix this?
> #1 new unplug protocol for SCSI, but old guests dont know about it
> #2 reuse IDE flag to also unplug SCSI. This would cover pvops and guests
>    where xenlinux based xen-platform-pci.ko is loaded before
>    sym53c8xx.ko. It would break guests with frontend drivers that do not
>    claim SCSI disks, the SCSI disk would disappear (if such frontends
>    really exist).
> #3 do not provide SCSI if guest has PV drivers

I think #3 was has been done in practice, but obviously not enforced
by the toolstack -- i.e., "Doctor, it hurts when I do this <makes
motion>."  "Then don't do that."

The problem with enforcing #3 is that there's no real way for the
toolstack to know if the guest will have PV drivers before booting.

There's also #4: Do not provide a PV backend for SCSI disks.  Not sure
that's actually possible, as libxl has historically used the PV
backend as the canonical place for listing disks associated with a
domain (although that may have changed since XSA-whatever which
resulted in libxl having its own local copy of the backend
information).

Paul, do you have any thoughts?

 -George

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

Re: missing unplug of SCSI devices in HVM guest

[Paul Durrant]

> -----Original Message-----
> From: dunlapg@gmail.com [mailto:dunlapg@gmail.com] On Behalf Of
> George Dunlap
> Sent: 06 September 2016 17:42
> To: Olaf Hering <olaf@aepfle.de>; Paul Durrant <Paul.Durrant@citrix.com>
> Cc: xen-devel@lists.xen.org
> Subject: Re: [Xen-devel] missing unplug of SCSI devices in HVM guest
> 
> On Wed, Aug 24, 2016 at 10:24 AM, Olaf Hering <olaf@aepfle.de> wrote:
> > Does anyone remember why the the vbd frontend drivers also claim the
> > SCSI disks, but the vbd backend in qemu has no unplug support for SCSI?
> >
> > The current situation for qemu-xen and qemu-xen-traditional is that
> > both will create an emulated LSI controller with disk=[vdev=sda]. The
> > xenlinux and pvops frontend drivers will claim the disk, but the
> > sym53c8xx will see and claim it as well. As a result each disk is
> > visible twice. One has to blacklist the sym53c8xx driver to avoid that.
> >
> > What should be done to fix this?
> > #1 new unplug protocol for SCSI, but old guests dont know about it
> > #2 reuse IDE flag to also unplug SCSI. This would cover pvops and guests
> >    where xenlinux based xen-platform-pci.ko is loaded before
> >    sym53c8xx.ko. It would break guests with frontend drivers that do not
> >    claim SCSI disks, the SCSI disk would disappear (if such frontends
> >    really exist).
> > #3 do not provide SCSI if guest has PV drivers
> 
> I think #3 was has been done in practice, but obviously not enforced by the
> toolstack -- i.e., "Doctor, it hurts when I do this <makes
> motion>."  "Then don't do that."
> 
> The problem with enforcing #3 is that there's no real way for the toolstack to
> know if the guest will have PV drivers before booting.
> 
> There's also #4: Do not provide a PV backend for SCSI disks.  Not sure that's
> actually possible, as libxl has historically used the PV backend as the canonical
> place for listing disks associated with a domain (although that may have
> changed since XSA-whatever which resulted in libxl having its own local copy
> of the backend information).
> 
> Paul, do you have any thoughts?
> 

I agree that #3 is not practical.

I would have thought option #1 is the most logical and desirable in the long run, but #2 could perhaps be used (by means of a configuration option to qemu) in the meantime. In practice I doubt there is anything out there that would use emulated SCSI as well as PV storage.

  Paul

>  -George
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: missing unplug of SCSI devices in HVM guest

[Olaf Hering]

Am 7. September 2016 12:38:09 MESZ, schrieb Paul Durrant <Paul.Durrant@citrix.com>:

>I would have thought option #1 is the most logical and desirable in the
>long run, but #2 could perhaps be used (by means of a configuration
>option to qemu) in the meantime. In practice I doubt there is anything
>out there that would use emulated SCSI as well as PV storage.

The qemu changes I sent earlier this month implement #2. And my testing shows there will be no loss.

Emulated SCSI has the benefit that loading kernel+initrd is faster.

Olaf 


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

Re: missing unplug of SCSI devices in HVM guest

[Paul Durrant]

> -----Original Message-----
> From: Olaf Hering [mailto:olaf@aepfle.de]
> Sent: 07 September 2016 11:49
> To: Paul Durrant <Paul.Durrant@citrix.com>; George Dunlap
> <dunlapg@umich.edu>
> Cc: xen-devel@lists.xen.org
> Subject: RE: [Xen-devel] missing unplug of SCSI devices in HVM guest
> 
> Am 7. September 2016 12:38:09 MESZ, schrieb Paul Durrant
> <Paul.Durrant@citrix.com>:
> 
> >I would have thought option #1 is the most logical and desirable in the
> >long run, but #2 could perhaps be used (by means of a configuration
> >option to qemu) in the meantime. In practice I doubt there is anything
> >out there that would use emulated SCSI as well as PV storage.
> 
> The qemu changes I sent earlier this month implement #2. And my testing
> shows there will be no loss.
> 

Ok. I guess #1 could be implemented later if it proves necessary in some circumstances.

  Paul

> Emulated SCSI has the benefit that loading kernel+initrd is faster.
> 

With seabios I'd certainly believe it. The IDE emulation is PIO only (unless you hack the config ... which I tend to do).

  Paul

> Olaf

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