Re: libxl / xen-pciback interaction

[Jan Beulich <jbeulich@xxxxxxxx>]

On 16.03.2021 16:03, Jürgen Groß wrote:
> On 16.03.21 15:27, Jan Beulich wrote:
>> All,
>>
>> while trying to test a pciback fix for how SR-IOV VFs get placed in the
>> guest PCI topology I noticed that my test VM would only ever see the 1st
>> out of 3 VFs that I passed to it. As it turns out, the driver watches
>> its own (backend) node, and upon first receiving notification it
>> evaluates state found in xenstore to set up the backend device.
>> Subsequently it switches the device to Initialised. After this switching,
>> not further instances of the watch triggering would do anything.
>>
>> In all instances I observed the watch event getting processed when the
>> "num_devs" node still reported 1. Trying to deal with this in libxl, by
>> delaying the writing of the "num_devs" node, led to a fatal error
>> ("num_devs" not being available to read) in the driver, causing the
>> device to move to Closing state. Therefore I decided that the issue has
>> to be addressed in the driver, resulting in a patch (reproduced below)
>> that I'm not overly happy with. I think the present libxl behavior is
>> wrong - it shouldn't trigger driver initialization without having fully
>> populated the information the driver is supposed to consume for its
>> device initialization. The only solution that I can think of, however,
>> doesn't look very appealing either: Instead of putting all pieces of the
>> data for one device in a transaction, make a single transaction cover
>> all devices collectively.
> 
> Any reason why you don't like this solution?

It would be quite a bit of code churn afaict (i.e. not an undertaking I
would like to start for code I'm not really familiar with), and error
cleanup might also become quite ugly.

> Its not as if there would
> be a large problem to be expected with using a single transaction for
> all PCI devices passed through (assuming you didn't mean to pack really
> all devices of the guest into that single transaction).

I meant all PCI devices; I can't even see a remote reason why all
devices regardless of type would want packing in a single transaction.

Jan