Re: [RFC PATCH] Xen PCI back - do slot and bus reset (v0).

[Ruediger Otte <ruediger.otte@gmx.net>]

Hi,

this patch actually solves the problem I experienced when trying to
passthrough a Radeon 7750 to a windows guest with a custom build of
Xen 4.4.0 on Debian Wheezy.

While I was always getting dom0 lockups at the first reboot of the
guest, I'm now able to do several reboots without sending the host to
standby in between. I haven't done extensive testing yet, but gpu
passthrough seems to just work now.

However I'm now seeing a different issue at dom0 boot when the
devices are assigned to xen-pciback (in kernel, no module). After the
message "xen_pciback: backend is passthrough" the host hangs for 2-3
seconds, then reboots. Strangely sometimes the host just boots ok,
but then again I get three or more failed boots in a row before it
finally works.

Before I applied your patch I have already followed every hint and
tried every available workaround with no success so far, so I would
be glad if this code could be further improved. If there's the need I
can of course provide debug output and configuration details from my
setup.

Kind regards,
Ruediger Otte

> Hey,
> 
> While I was trying to narrow down the state of GPU passthrough
> (still not finished) and figuring what needs to be done I realized
> that Xen PCIback did not reset my GPU properly (when I crashed the
> Windows guest by mistake). It does an FLR reset or Power one - if
> the device supports it. But it seems that some of these GPUs
> are liars and actually don't do the power part properly.
> 
> One way to fix this is by doing a slot (aka device) and bus reset.
> Of course to do that - you need to make sure that all of the
> functions of a device are under the ownership of xen-pciback.
> Otherwise you might accidently reset your sound card while it is
> being used.
> 
> These RFC patches cleanup pci back a bit and also make it possible
> for Xen pciback to do the whole gamma of 'reset' for PCI devices:
> FLR, power management, AER, slot and bus reset if neccessary.
> 
> Thanks go to Gordan Bobic for educating me on how to "reprogram"
> and GFX460 in a Quardro 4000M and also reporting oddities when
> a PCI device was reset but it looked like it was not reset.
> 
> 
>  drivers/xen/xen-pciback/pci_stub.c | 142
> +++++++++++++++++++++++++++++++------
> drivers/xen/xen-pciback/xenbus.c   |   5 +- 2 files changed, 124
> insertions(+), 23 deletions(-)
> 
> 
> Konrad Rzeszutek Wilk (5):
>       xen-pciback: Cleanup up pcistub_put_pci_dev
>       xen-pciback: First reset, then free.
>       xen-pciback: Document when we FLR an PCI device.
>       xen/pciback: Move the FLR code to a function.
>       xen/pciback: PCI reset slot or bus