Re: [Xen-devel] [PATCH] Xen/pciback: Implement PCI slot or bus reset with 'do_flr' SysFS attribute

[Govinda Tatti <govinda.tatti@oracle.com>]

On 11/29/2017 12:05 PM, Pasi Kärkkäinen wrote:
> On Wed, Nov 29, 2017 at 11:25:09AM -0600, Govinda Tatti wrote:
>>>>> Furthermore, contrary to what you claim in
>>>>> your reply to Pasi, I can't see where you try an actual FLR first -
>>>>> you go straight to pci_probe_reset_{slot,bus}(). If you actually
>>>>> tried FLR first, only falling back to the other methods as "emulation",
>>>>> I could certainly agree with the file name chosen.
>>>> Currently, multiple resets are being invoked (independently) in the context
>>>> of "xl attach/detach/shutdown/reboot".
>>>>
>>>> - pci_reset_function_locked (invoked by pcistub_put_pci_dev())
>>>>     This function tries various PCI reset methods including FLR.
>>>> - pcistub_reset_dev (called by toolsstack based on "do_flr" attribute)
>>> While related in a certain way, I can't really see how this addresses
>>> the comment.
>> pcistub_reset_dev() just tries slot or bus reset but not FLR since it is
>> being
>> checked and executed by pci_reset_function_locked() if supported. May be we
>> can
>> add FLR reset code to pcistub_reset_dev() and try FLR first before fall-back
>> to
>> slot/bus reset.
>>
> Hmm.. is the suitable reset method something that can be figured out automatically?
>
> I mean if FLR is tried first, but it isn't supported by the device, is it OK to go ahead and do a slot/bus reset automatically?
Yes, we are moving in the same direction.

> What if there are other devices in the same bus, wouldn't automatic bus reset be a bad thing?
We will perform BUS or SLOT reset only if all device/functions behind 
the bus/slot are owned by pcistub.
Otherwise, we will skipthis reset.

> Or should the reset-method be configured by the user for the VM / PCI device ?
>
> Just thinking out loud here..
>

Cheers
GOVINDA