Re: [PATCH v3 4/7] xen/pciback: Implement PCI reset slot or bus with 'do_flr' SysFS attribute

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

On Wed, Jul 09, 2014 at 03:26:05PM +0100, David Vrabel wrote:
> On 09/07/14 15:12, Konrad Rzeszutek Wilk wrote:
> > On Wed, Jul 09, 2014 at 01:32:10PM +0100, David Vrabel wrote:
> >> On 08/07/14 19:46, Konrad Rzeszutek Wilk wrote:
> >>> On Tue, Jul 08, 2014 at 07:02:51PM +0100, David Vrabel wrote:
> >>>> On 08/07/14 19:58, konrad@kernel.org wrote:
> >>>>> --- a/Documentation/ABI/testing/sysfs-driver-pciback
> >>>>> +++ b/Documentation/ABI/testing/sysfs-driver-pciback
> >>>>> @@ -82,3 +82,14 @@ Description:
> >>>>>                  device is shared, enabled, or on a level interrupt line.
> >>>>>                  Writing a string of DDDD:BB:DD.F will toggle the state.
> >>>>>                  This is Domain:Bus:Device.Function where domain is optional.
> >>>>> +
> >>>>> +What:           /sys/bus/pci/drivers/pciback/do_flr
> >>>>> +Date:           July 2014
> >>>>> +KernelVersion:  3.16
> >>>>> +Contact:        xen-devel@lists.xenproject.org
> >>>>> +Description:
> >>>>> +                An option to slot or bus reset an PCI device owned by
> >>>>> +                Xen PCI backend. Writing a string of DDDD:BB:DD.F will cause
> >>>>> +                the driver to perform an slot or bus reset if the device
> >>>>> +                supports. It also checks to make sure that all of the devices
> >>>>> +                under the bridge are owned by Xen PCI backend.
> >>>>
> >>>> Not sure I like this new interface.  I solved this by adding a new reset
> >>>> file that looked like the regular one the pci would have if it supported
> >>>> FLR.  I'm fairly sure I posted a series for this.  Was there a reason
> >>>> you didn't do this?
> >>>
> >>> It did not work.
> >>>
> >>> During bootup kobject would complain about a secondary 'reset' SysFS
> >>> on the PCI device.
> >>
> >> I think this because of pciback registering a driver too early, before
> >> the device is fully initialized.  You can see in the trace that it is
> >> the common pci code that is trying to add the "reset" file so it must be
> >> doing this /after/ pciback's probe has been called.
> >>
> >> I would consider:
> >>
> >> 1. Removing the "hide" module parameter -- it doesn't work if pciback is
> >> a module anyway.
> > 
> > I find it incredibly useful and so do a lot of other people.
> 
> PCI passthrough must work well without hide and without pciback being
> built-in (and it does with the "reset" change).
> 
> What are you using "hide" for?

For hiding the AHCI driver from the likes of multipath and lvm.
I want the device driver domain to set those up and don't want the
initial domain have to create this and then have to tear it down.

Ditto for the bttv - it ends up loading tons of drivers and
just sorting out the dependency is tedious. If I hide it away
I can easily pass it in the backend.
> 
> >> 2. Making pciback initialize like a regular driver module (no
> >> fs_initcall() shenanigans).
> > 
> > The point is to take the PCI device before the drivers touch it.
> > 
> > We want it to be in a pristine state so that the device driver
> > domains can use it.
> 
> But hide only ensures this the first time the device is assigned.  Using
> a function reset ensures this all the time.

The hide also saves the registers:

 514         pci_save_state(dev);
 515         dev_data->pci_saved_state = pci_store_saved_state(dev);

before they are used by device drivers. The function reset (or bus
reset) won't ensure that those registers will always be the same
from the initial bootup state. As in: device drivers -> pciback ->
save states -> FLR -> give to guest. There is still the taint of the
device driver modifying the registers.

> 
> >> 3. Require userspace to sort out binding the device to pciback (e.g.,
> >> libxl already does the unbind if requested).
> > 
> > How would you do the bus/slot reset? Or are you thinking that at
> > that point the 'reset' functionality would be over-written to point
> > to Xen pciback and it would do the job?
> 
> I'm not sure I understand your question.  libxl already does the
> function reset (by writing to "reset").

Right.  And it also does 'do_flr'.
> 
> >> 4. Finally, I would consider generic driver core functionality for
> >> prioritizing drivers so they get probed first.
> > 
> > Not sure I understand why you want the drivers to use the device
> > first? The point is that we can 'hide' them from the generic
> > drivers and present them to the backend domains.
> 
> The pciback driver would be prioritized, so it would be probed first.

Which would require it to be some early_initcall variant and use
the override (which I think is going in 3.17?) for 'struct device'?
(Alex implemented it). I think that is what you are saying? However
that looks to be orthogonal to what this patch tries to do?

> 
> > Regardless of these - I am curious to why you don't like do_flr
> > as it is even implemented in the the toolstack (but buggy) and
> > it does a good job of allowing us to do slot/bus reset?
> 
> Because there is already a documented interface for resetting devices
> (the "reset" file), we don't want a second interface.

Which just does the FLR. This is doing bus/slot reset and doing an
override of the 'reset' in SysFS does not work without some clever
coding.

The 'do_flr' seems to be already baked in libxl - why not use it?
> 
> David