Re: [dpdk-dev] [PATCH] bus/pci: set intr_handle type for secondary processes

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 18-Oct-18 5:41 PM, Alejandro Lucero wrote:
> 
> 
> On Thu, Oct 18, 2018 at 5:26 PM Burakov, Anatoly 
> <anatoly.burakov@intel.com <mailto:anatoly.burakov@intel.com>> wrote:
> 
>     On 27-Sep-18 1:30 PM, Alejandro Lucero wrote:
>      > Invoking rte_pci_read/write_config functions requires device with
>      > a intr_handle type for using VFIO or UIO driver related functions.
>      >
>      > Secondary processes rely on primary processes for device
>     initialization
>      > so they do not usually require using these functions. However,
>     some PMDs,
>      > like NFP PMD, require using these functions even for secondary
>     processes.
>      >
>      > Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com
>     <mailto:alejandro.lucero@netronome.com>>
>      > ---
> 
>     Hi Alejandro,
> 
>     I’m curious of consequences of setting intr handle to a valid value
>     when
>     we don’t have an interrupt thread. Something may try to use it
>     (although
>     I couldn’t find any such usage).
> 
> 
> The point is secondary processes do not deal with interrupts so I assume 
> setting the type does not change anything but it allows to use PCI 
> read/write functions by secondary processes.
> 
>     PCI config read really uses intr handle type to discover userspace
>     driver type – this seems ever so slightly wrong, and looks like
>     something that should be part of rte_device somewhere, independent of
>     interrupt types. Do we have any other alternative to do the same thing
>     (i.e. know what userspace driver is used for a particular PCI device)?
> 
> 
> I agree current way not being specially good.
> 
> Your comment has reminded me there is another way: just using the kdrv 
> field from the rte_pci_device struct. I have code using that field for 
> doing a different thing in the NFP PMD depending on the driver in use, 
> UIO or VFIO. So I think a better patch would be just to modify those pci 
> functions for using kdrv field instead.
> 
> Adding Ferruh in the thread for commenting on this potential change.
> 

I definitely think the way you describe would be a better way to fix 
this (i.e. use kdrv in PCI config functions rather than intr handle type).

-- 
Thanks,
Anatoly