Re: RFC: Automatically making a PCI device assignable in the config file

[Gordan Bobic <gordan@bobich.net>]

 On Wed, 10 Jul 2013 16:29:33 +0100, George Dunlap 
 <george.dunlap@eu.citrix.com> wrote:
> On 10/07/13 16:12, Gordan Bobic wrote:
>> On Wed, 10 Jul 2013 15:45:49 +0100, George Dunlap 
>> <george.dunlap@eu.citrix.com> wrote:
>>> On 10/07/13 14:55, Ian Jackson wrote:
>>>> George Dunlap writes ("Re: [Xen-devel] RFC: Automatically making a 
>>>> PCI device assignable in the config file"):
>>>>> Auto-seizing is fairly dangerous; you could easily accidentally 
>>>>> yank out
>>>>> the ethernet card, or even the disk that dom0 is using.  I really 
>>>>> think
>>>>> it should have to be enabled on a device-by-device basis.
>>>> I don't think this makes any sense.
>>>>
>>>> In practice you are saying that in order to avoid mistakes, the 
>>>> local
>>>> admin must provide the BDF of the device to be passed through in 
>>>> two
>>>> separate places.
>>>
>>> That's not what I had in mind; what I had in mind was something 
>>> like this:
>>>
>>> pci = [ '08:04.1,seize=1' ]
>>>
>>> Or alternately:
>>>
>>> xl pci-attach h0 08:04.1,seize=1
>>>
>>> One could also imagine having something in xl.conf like the 
>>> following:
>>>
>>> pci.autoseize = [ '08:04.1','01:00.0' ]
>>>
>>> In this case you wouldn't be simply copy and pasting, because you'd
>>> probably have different domains handling each device.  But in any
>>> case, this was just exploring the alternatives -- I don't think 
>>> that's
>>> the best thing to do.
>>>
>>>> But that doesn't actually help.  If this is all properly 
>>>> documented
>>>> and so forth (ie, if the admin has a smooth experience and doesn't
>>>> trip over having dailed to "seize" the device), they will just
>>>> cut-and-paste the same value into both places in the config.
>>>>
>>>> They will also mutter under their breath to ask why this is
>>>> necessary...
>>>
>>> If someone can accidentally type "xl pci-attach 08:04.0" instead of
>>> "xl pci-attach 08.04.1" and suddenly completely lose their network
>>> connectivity (or yank their hard drive out), then I think they'll
>>> appreciate it.
>>>
>>> In any case, at the moment it's much worse: you have to either
>>> scriptwise run "xl pci-assignable-add" a device, or add an 
>>> exclusion
>>> on the Linux commandline in grub.  So far I'm the only person I 
>>> know
>>> who has complained about it. :-)
>>
>> This would be an awesome feature, but success can be mixed. My
>> experience is that Nvidia drivers (as one example, I'm sure there
>> are others) don't handle device detaching very gracefully. ACS
>> might help but I have no means of testing
>> that at the moment. So ultimately, you still have to either
>> exclude at least some devices from grub (if
>> xen-pciback is built in) or in modprobe.d (attach device to
>> pciback in driver pre-install). I'm not sure what can be
>> done about this considering we are talking about 3rd
>> party binary drivers. :(
>
> Yes, unfortunately we'll always have to have the fallback for devices
> that don't play well with being re-assigned.  But I my very limited
> experience with network cards gives me some hope that at least for
> many devices for which driver domains might be in use, the
> auto-reassign thing will be useful in a number of cases.

 What about having a feature that upon parsing the config file modifies
 the grub.conf or modprobe.d/xen-pciback.conf (depending on whether
 xen-pciback is module or built in) and adds the required entries
 for the device(s)?

 Gordan