Re: [dpdk-dev] [PATCH v2] eal/devargs: add option to supply PCI dev args

["Gaëtan Rivet" <gaetan.rivet@6wind.com>]

On Wed, Jun 27, 2018 at 02:25:30PM +0530, Pavan Nikhilesh wrote:
> Hi Gaëtan,
> 
> On Wed, Jun 27, 2018 at 10:39:59AM +0200, Gaëtan Rivet wrote:
> > Hi Ferruh, Pavan,
> >
> > sorry for the delay,
> >
> > On Tue, Jun 26, 2018 at 04:40:21PM +0100, Ferruh Yigit wrote:
> > > On 6/26/2018 1:48 PM, Shahaf Shuler wrote:
> > > > Hi Pavan,
> > > >
> > > > Friday, June 15, 2018 7:44 AM, Pavan Nikhilesh:
> > > >> Subject: [dpdk-dev] [PATCH v2] eal/devargs: add option to supply PCI dev
> > > >> args
> > > >>
> > > >> Currently, the only way of supplying device argument to a pci device is to
> > > >> whitelist it i.e. -w 000X:00:0X.0,self_test=1. This is not a very feasible method
> > > >> as whitelisting a device has its own side effects i.e only the whitelisted pci
> > > >> devices are probed.
> > > >>
> > > >> Add a new eal command line option --pci-args to pass device args without the
> > > >> need to whitelist the devices.
> > > >>            --pci-args 000X:00:0X.0,self_test=1
> > > >>
> > > >> Signed-off-by: Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>
> > > >
> > > > Tested-by: Shahaf Shuler <shahafs@mellanox.com>
> > > >
> > > > It seems to work.
> > > > Please see small comments below
> > >
> > > Isn't this conflict with Gaetan's devarg work which has wider scope?
> > >
> >
> > Indeed it does.
> >
> > Pavan, I have submitted a new version of a series adding generic kvargs
> > to several layers (bus, class, driver).
> >
> > It does cover this exact use-case.
> >
> > However, while writing it, I wasn't able to find PCI bus specific
> > parameters, that could showcase the functionality.
> 
> The idea of the patch is to avoid whitelising a device when we want to
> supply kvargs to it, I tried mapping it to devargs rework patchset but couldn't
> do it at a glance. For example, the following patch[1] reads kvargs through
> whitelisting which should be avoided.
> 
> [1]http://patches.dpdk.org/patch/41223/
> 

I see.

Actually, your use-case won't be covered by the devargs rework.

I am still dumbfounded by how this blacklist/whitelist mode stuff is
kept against all odds. But that's not the time to deal with it.

The issue is that the two features "declaring a device" and
"configuring a bus" are currently awkwardly merged. You are piling stuff
on the "declaring a device" part to enhance the "configuring a bus"
feature.

Instead of going this way, I would advise to separate the two features.

If buses could be configured with a generic EAL option
"--blacklist=pci,vdev" for example, then you could provide devargs as
much as you want, the buses themselves would stay properly configured.

This means removing devargs policy, device types and rewriting bus logic
about it.

Regards,
-- 
Gaëtan Rivet
6WIND