Re: binding/unbinding devices to vfio-pci

[Hannes Reinecke <hare@suse.de>]

On 07/02/2013 05:13 PM, Yoder Stuart-B08248 wrote:
>
>
>> -----Original Message-----
>> From: Alex Williamson [mailto:alex.williamson@redhat.com]
>> Sent: Tuesday, July 02, 2013 9:46 AM
>> To: Yoder Stuart-B08248
>> Cc: kvm@vger.kernel.org list; Alexander Graf; Bhushan Bharat-R65777; a.motakis@virtualopensystems.com;
>> virtualization@lists.linux-foundation.org
>> Subject: Re: binding/unbinding devices to vfio-pci
>>
>> On Tue, 2013-07-02 at 14:15 +0000, Yoder Stuart-B08248 wrote:
>>> Alex,
>>>
>>> I'm trying to think through how binding/unbinding of devices will
>>> work with VFIO for platform devices and have a couple of questions
>>> about how vfio-pci works.
>>>
>>> When you bind a device to vfio-pci, e.g.:
>>> # echo 1102 0002 > /sys/bus/pci/drivers/vfio-pci/new_id
>>>
>>> ...I understand that the echo into 'new_id' tells the
>>> vfio pci driver that it now handles the specified PCI ID.
>>>
>>> But now there are 2 drivers that handle that PCI ID,
>>> the original host driver and vfio-pci.   Say that
>>> you hotplug a PCI device that matches that ID.   Which of
>>> the 2 drivers are going to get bound to the device?
>>>
>>> Also, if you unbind a device from vfio-pci and want to
>>> bind it again to the normal host driver you would just
>>> echo the full device info into the 'bind' sysfs file
>>> for the host driver, right?
>>>
>>>      echo 0000:06:0d.0 > /sys/bus/pci/drivers/...
>>
>> Hi Stuart,
>>
>> The driver binding interface is far from perfect.  In your scenario
>> where you've added the ID for one device, then hotplug another device
>> with the same ID, the results are indeterminate.  Both vfio-pci and the
>> host driver, assuming it's still loaded, can claim the device, it's just
>> a matter of which gets probed first.
>>
>> Generally that window should be very short though.  To bind a device,
>> the user should do:
>>
>> 1) echo ssss:bb:dd.f > /sys/bus/pci/devices/ssss:bb:dd.f/driver/unbind
>> 2) echo vvvv dddd > /sys/bus/pci/drivers/vfio-pci/new_id
>> 3) echo ssss:bb:dd.f > /sys/bus/pci/drivers/vfio-pci/bind
>> 4) echo vvvv dddd > /sys/bus/pci/drivers/vfio-pci/remove_id
>>
>> There are actually a number of ways you can do this and the default
>> autoprobe behavior really makes step 3) unnecessary as the driver core
>> will probe any unbound devices as soon as a new_id is added to vfio-pci.
>> That can be changed by:
>>
>> # echo 0 > /sys/bus/pci/drivers_autoprobe
>>
>> But then we have to worry about races from any devices that might have
>> been hotplugged in the interim.
>
> But, even apart from hot-plugged devices, what about the device
> we just unbound?  There are now 2 host drivers that can handle the
> device when the autoprobe happens.  Is it just luck that vfio-pci
> is the one that gets the device?
>
Probably the best way of handling this would be to disallow autobinding 
for vfio in general.
Then the 'normal' driver would be bound to the device, and vfio must be 
enabled via manual binding.
Much like it is today.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)