Re: [RFC 1/1] s390/cio: Remove uevent-suppress from css driver

[Cornelia Huck <cohuck@redhat.com>]

On Wed, 9 Dec 2020 13:52:03 +0100
Halil Pasic <pasic@linux.ibm.com> wrote:

> On Tue, 8 Dec 2020 18:30:54 +0100
> Cornelia Huck <cohuck@redhat.com> wrote:
> 
> > > 
> > > But, the more i look at this patch and discuss on this, i think this is 
> > > not complete.
> > > i.e as you know, the main reason for this RFC was the the below thread.
> > > https://marc.info/?l=linux-s390&m=158591045732735&w=2
> > > We are still not solving the problem that was mentioned in that RFD.
> > > 
> > > There are couple of things which we needs to consider here. With this 
> > > patch, the uevents
> > > are generated before doing the initialization or before finding the 
> > > ccw-device
> > > connected. Which means, the udev-rules have to manage with a 
> > > non-initialized setup
> > > compared to the previous version (Version without this patch). As you 
> > > mentioned, the
> > > current user-space programs which works with this uevent, especially in 
> > > case of vfio-ccw
> > > will have a problem.    
> > 
> > IIUC, we'll get the "normal" ADD uevent when the subchannel device is
> > registered (i.e. made visible). For the vfio-ccw case, we want the
> > driverctl rule to match in this case, so that the driver override can
> > be set before the subchannel device ends up being bound to the I/O
> > subchannel driver. So I think that removing the suppression is giving
> > us exactly what we want? Modulo any errors in the initialization
> > sequence we might currently have in the css bus code, of course.
> >  
> 
> I believe, I'm the originator of these concerns, yet I find my
> concern hard to recognize in the comment of Vineeth, so let me
> please try to explain this in a different way.
> 
> AFAIK the uevent handling is asynchronous with regards to matching and
> probing, in a sense that there is no synchronization mechanism that
> ensures, the userspace has had the ADD event handled (e.g.
> driver_override set_ before the kernel proceeds with matching and
> probing of the device. Am I wrong about this?
> 
> If I'm, with the suppression gone we end up with race, where userspace
> may or may not set driver_override in time.
> 
> The man page of driverctl
> (https://manpages.debian.org/testing/driverctl/driverctl.8.en.html)
> claims that: "driverctl integrates with udev to support overriding driver
> selection for both cold- and hotplugged devices from the moment of discovery, ..."
> and "The driver overrides created by driverctl are persistent across
> system reboots by default."
> 
> Writing to the driver_override sysfs attribute does not auto-rebind. So
> if we can't ensure being in time to set driver_override for the
> subchannel before the io_subchannel driver binds, then the userspace
> should handle this situation (by unbind and bind) to ensure the
> effectiveness of 'driver override'. I couldn't find that code in
> driverctl, and I assume if we had that, driver override would work
> without this patch. Conny, does that sound about right?
> 
> My argument is purely speculative. I didn't try this out, but trying
> stuff out is of limited value with races anyway. Vineeth did you try?
> If not, I could check this out myself some time later.

Whenever we delegate policy decisions like that to userspace, we'll end
up with uncertainty depending on timing. I don't think that there's any
way around it. (FWIW, driver_override for pci behaves just the same,
unless I misread the code.)

What removing the uevent suppression does give us is at least a chance
to influence the driver that is being bound and not wait until we have
a fully setup ccw device as a child as well.

>  
> > I'm not sure how many rules actually care about events for the
> > subchannel device; the ccw device seems like the more helpful device to
> > watch out for.  
> 
> I tend to agree, but the problem with vfio-ccw is that (currently) we
> don't have a ccw device in the host, because we pass-through the
> subchannel. When we interrogate the subchannel, we do learn if there
> is a device and if, what is its devno. If I were to run a system with
> vfio-ccw passthrough, I would want to passthrough the subchannel that
> talks to the DASD (identified by devno) I need passed through to my
> guest.

I think that can be solved by simply adding the devno as a variable to
the uevent (valid if it's an I/O subchannel; we don't register the
subchannel in the first place if dnv is not set.)