Re: [PATCH v3 1/9] s390: vfio_ap: link the vfio_ap devices to the vfio_ap bus subsystem

[Cornelia Huck]

On Thu, 14 Feb 2019 13:30:59 -0500
Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:

> On 2/14/19 12:36 PM, Pierre Morel wrote:
> > On 14/02/2019 17:57, Cornelia Huck wrote:  
> >> On Thu, 14 Feb 2019 16:47:30 +0100 Pierre Morel
> >> <pmorel@xxxxxxxxxxxxx> wrote:
> >>  
> >>> On 14/02/2019 15:54, Cornelia Huck wrote:  
> >>>> On Thu, 14 Feb 2019 14:51:01 +0100 Pierre Morel
> >>>> <pmorel@xxxxxxxxxxxxx> wrote:  

> >>>>> -ÂÂÂ matrix_dev->device.type = &vfio_ap_dev_type; 
> >>>>> dev_set_name(&matrix_dev->device, "%s", VFIO_AP_DEV_NAME); 
> >>>>> matrix_dev->device.parent = root_device; +
> >>>>> matrix_dev->device.bus = &matrix_bus; matrix_dev->device.release = 
> >>>>> vfio_ap_matrix_dev_release; -
> >>>>> matrix_dev->device.driver = &vfio_ap_drv.driver; +
> >>>>> matrix_dev->vfio_ap_drv = &vfio_ap_drv;  
> >>>>
> >>>> Can't you get that structure through matrix_dev->device.driver
> >>>> instead when you need it in the function below?  
> >>>
> >>> Not anymore. We have two different drivers and devices matrix_drv
> >>> <-> matrix_dev and vfio_ap_drv <-> ap_devices
> >>>
> >>> The driver behind the matrix_dev->dev->driver is matrix_drv what is
> >>> needed here is vfio_ap_drv.  
> >>
> >> Wait, we had tacked a driver for ap devices unto a matrix device,
> >> which is not on the ap bus?  
> 
> It's really a bit more complicated than that. Without going into a
> lengthy description of the history of AP passthrough support, suffice it
> to say that we needed a device to serve as the parent of each mediated
> device used to configure a matrix of AP adapter IDs and domain indexes
> identifying the devices to which a guest would be granted access. The
> AP devices themselves are attached to the AP bus, but the matrix device
> is an artificial (virtual?) device whose sole purpose in life is to
> serve as an anchor for the mediated devices whose sysfs interfaces are
> created and managed by the vfio_ap device driver. The matrix device
> itself is created by the vfio_ap device driver - when it is initialized 
> - for that purpose. In hindsight, maybe there was a better way to
> implement this, but neither this patch nor this discussion belongs in
> this series. It distracts from discussion of interrupt support which is
> the sole purpose of the patch series.

The we-need-a-parent part is fine; but whatever we're doing with that
driver just looks wrong, so that even the new bus that basically does
nothing looks better...

> 
> > 
> > ...yes -(
> >   
> >> Maybe that's what trips libudev? >
> >> (And reading further in the current code, it seems we clear that 
> >> structure _after_ the matrix device had been setup, so how can that 
> >> even work? Where am I confused?)  
> > 
> > On device_register there were no bus, so the core just do not look for a 
> > driver and this field was nor tested nor overwritten.

Hm... so has the callback in driver_for_each_device() in
vfio_ap_verify_queue_reserved() ever been invoked at all? It seems this
patch fixes more than just libudev issues...

> >   
> >>  
> >>>  
> >>>>  
> >>>>>
> >>>>> ret = device_register(&matrix_dev->device); if (ret) goto
> >>>>> matrix_reg_err;
> >>>>>
> >>>>> +ÂÂÂ ret = driver_register(&matrix_driver.drv); +ÂÂÂ if (ret) 
> >>>>> +ÂÂÂÂÂÂÂ goto
> >>>>> matrix_drv_err; +  
> >>>>
> >>>> As you already have several structures that can be registered
> >>>> exactly once (the root device, the bus, the driver, ...), you can
> >>>> already be sure that there's only one device on the bus, can't
> >>>> you?  
> >>>
> >>> hum, no I don't think so, no device can register before this module
> >>> is loaded, but what does prevent a device to register later from
> >>> another module?  
> >>
> >> Not unless you export the interface, I guess.
> >>  
> > 
> > :) definitively right
> > thanks, this will simplify the code in the next version.
> > I will take the patch away from this series to get the way to stable as 
> > Christian requested.

Yeah, makes sense.