Re: xen-pciback.hide syntax

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Tue, Jul 31, 2012 at 05:40:14PM +0200, Sander Eikelenboom wrote:
> Hello Konrad,
> 
> Tuesday, July 31, 2012, 5:25:58 PM, you wrote:
> 
> > On Mon, Jul 30, 2012 at 09:47:41PM +0200, Sander Eikelenboom wrote:
> >> Monday, July 30, 2012, 9:00:06 PM, you wrote:
> >> 
> >> > On Sat, May 19, 2012 at 10:46:15AM +0200, Sander Eikelenboom wrote:
> >> >> Hi Konrad,
> >> >> 
> >> >> The syntax for specifying the devices for pciback to hide is "bus:device.function".
> >> >> While thinking about cooking up a patch to be able to use a "*" wildcard for the function, i was wondering if not hiding all functions of a device is feasible at all.
> >> >> 
> >> >> For what I understand of PCI, function 0 is always required, so if I only hide function 0, i can't use the other functions in dom0, since those functions would require a function 0, which is hidden.
> >> >> 
> >> >> So would it be more logical to drop/ignore the function from the BDF, and always hide all functions from a device ?
> >> 
> >> > That might run afoul of the SR-IOV virtual devices. They (when loaded) provide a fake
> >> > bus:device:function, where the device is port (so if the SR-IOV card has two
> >> > jacks, you get 00 and 01), and the function is for the amount of VFs it can make.
> >> > On the Intel SR-IOV NIC with 'igbvf.max_vfs=7' I end up with 14 PCI devices, where
> >> > the function bear no resemblence to each other (and can be passed in different guests).
> >> 
> >> > The PCI restriction I know of is if the device is behind a bridge. The issue here
> >> > is that .. well, you could pass in a different function to a different guest, but
> >> > one guest's hardware device could listen on the other guests' function. It would
> >> > require tweaking the driver to dump the contents of some registers and some deep
> >> > hacking, but that is the security issue with that.
> >> 
> >> Hmm that would mean there are three possibilities:
> >> 1) Accept a Wildcard syntax like "bus:device.*", which would mean hide all functions of device.
> 
> > Which in this context actually makes sense. You probably don't want to use the VF's in
> > your host.
> 
> In my use cases i always hide all functions, and since my usb controllers have 7 functions, that leads to quite some long lines.
> 
> >> 2) Accept not providing the function as a wildcard "bus:device", would mean hide all functions of device.
> 
> > <nods>.
> >> 
> >> 3) Do nothing, the gained overview on grub lines isn't worth the effort :-)
> 
> > Heh!
> 
> > I think I like 2).
> 
> I think that would be the most simple and straightforward to implement, the only thing is that the .cfg files seem to use the "bus:device.*" scheme ...
> Don't know if there are any other related cmd options for the kernel that use a certain syntax that could be preferred ?
> 

So Jan implemented this and it is in v3.7.