Re: [Qemu-devel] [PATCH v9 10/12] vfio: Add function to get device_api string from vfio_device_info.flags

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 21 Oct 2016 08:30:53 +0530
Kirti Wankhede <kwankhede@nvidia.com> wrote:

> On 10/21/2016 2:52 AM, Alex Williamson wrote:
> > On Fri, 21 Oct 2016 02:44:37 +0530
> > Kirti Wankhede <kwankhede@nvidia.com> wrote:
> >   
> ...
> 
> >>>>>>  
> >>>>>> +extern const char *vfio_device_api_string(u32 flags);
> >>>>>> +
> >>>>>>  struct pci_dev;
> >>>>>>  #ifdef CONFIG_EEH
> >>>>>>  extern void vfio_spapr_pci_eeh_open(struct pci_dev *pdev);      
> >>>>>
> >>>>> Couldn't this simply be a #define in the uapi header?
> >>>>>
> >>>>> #define VFIO_DEVICE_PCI_API_STRING "vfio-pci"
> >>>>>
> >>>>> I don't really see why we need a lookup function.
> >>>>>       
> >>>>
> >>>> String is tightly coupled with the FLAG, right?
> >>>> Instead user need to take care of making sure to return proper string,
> >>>> and don't mis-match the string, I think having function is easier.    
> >>>
> >>> That's exactly why I proposed putting the #define string in the uapi,
> >>> by that I mean the vfio uapi header.  That keeps the tight coupling to
> >>> the flag, they're both defined in the same place, plus it gives
> >>> userspace a reference so they're not just inventing a string to compare
> >>> against.  IOW, the vendor driver simply does an sprintf of
> >>> VFIO_DEVICE_PCI_API_STRING and userspace (ie. libvirt) can do a strcmp
> >>> with VFIO_DEVICE_PCI_API_STRING from the same header and everybody
> >>> arrives at the same result.
> >>>     
> >>>> Vendor driver should decide the type of device they want to expose and
> >>>> set the flag, using this function vendor driver would return string
> >>>> which is based on flag they set.    
> >>>
> >>> Being a function adds no intrinsic value and being in a uapi header does
> >>> add value to userspace.  Thanks,
> >>>     
> >>
> >> Ok. The strings should be in uapi, but having function (like below) to
> >> return proper string based on flag would be good to have for vendor driver.
> >>
> >>  +const char *vfio_device_api_string(u32 flags)
> >>  +{
> >>  +	if (flags & VFIO_DEVICE_FLAGS_PCI)
> >>  +		return VFIO_DEVICE_API_PCI_STRING;
> >>  +
> >>  +	if (flags & VFIO_DEVICE_FLAGS_PLATFORM)
> >>  +		return VFIO_DEVICE_API_PLATFORM_STRING;
> >>  +
> >>  +	if (flags & VFIO_DEVICE_FLAGS_AMBA)
> >>  +		return VFIO_DEVICE_API_AMBA_STRING;
> >>  +
> >>  +	return "";
> >>  +}
> >>  +EXPORT_SYMBOL(vfio_device_api_string);  
> > 
> > I disagree, it's pointless maintenance overhead.  It's yet another
> > function that we need to care about for kABI and it offers almost no
> > value.  Thanks,
> >   
> 
> If any vendor driver sets VFIO_DEVICE_FLAGS_PLATFORM flag but sets
> VFIO_DEVICE_API_PCI_STRING, we don't have a way to verify this in kernel
> driver. Is that acceptable?

a) The function doesn't solve that problem, as seen in the mtty sample
driver there's no guarantee that the vendor driver passes
device_info.flags vs the #define for the flag itself.  So we're already
expecting them to get it right in two separate places.

b) It's not going to take them very long to figure this out if they
care about userspace tools that make use of this field to determine the
device API.  This is a very obvious and simple bug.

c) We can check and correct how open source vendor drivers work.

Thanks,
Alex