Re: [dpdk-dev] [PATCH v1 1/1] bus/pci: probe PCI devices in whitelisted order

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

On Wed, Sep 25, 2019 at 06:41:36AM +0000, Slava Ovsiienko wrote:
> > -----Original Message-----
> > From: dev <dev-bounces@dpdk.org> On Behalf Of vattunuru@marvell.com
> > Sent: Monday, September 23, 2019 14:57
> > To: dev@dpdk.org
> > Cc: gaetan.rivet@6wind.com; ferruh.yigit@intel.com;
> > anatoly.burakov@intel.com; Thomas Monjalon <thomas@monjalon.net>;
> > jerinj@marvell.com; Vamsi Attunuru <vattunuru@marvell.com>
> > Subject: [dpdk-dev] [PATCH v1 1/1] bus/pci: probe PCI devices in whitelisted
> > order
> > 
> > From: Vamsi Attunuru <vattunuru@marvell.com>
> > 
> > Current pci bus driver scans pci devices in the order that it read from sysfs.
> > Accordingly all or whitelisted devices are getting probed.
> > 
> > Patch modifies the probing order of whitelisted pci devices in a sequence the
> > devices are whitelisted(using EAL flags).
> 
> Thanks, it would be nice to have opportunity to control probing order,
> it might be useful for bonded devices and representors either.
> 
> Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
> 
> > 
> > It ensures the eth devices that application uses are probed in device
> > whitelisted sequence, in turn it facilitates the packet forwarding applications
> > to work without any packet loss or performance drop when the underneath
> > network ports have different bandwidths. By altering the whitelist order
> > applications like testpmd, l2fwd can forward the ingress traffic to egress port
> > that has of equivalent bandwidth.
> > 
> > Signed-off-by: Vamsi Attunuru <vattunuru@marvell.com>

Hello Vamsi, Viacheslav,

This is a nice patch. I agree that port dependency could be better
handled. The port-mapping part however should be managed at the app
level.

Vamsi, you gave the example of l2fwd and testpmd, being able to
properly configure forwarding directions implicitly. I think the better
approach here is to add these configurations items within the apps
directly. Configuring the mapping at the port level is not precise
enough. The proper control is about cores, port and queues, not only ports.
This patch only solves a limited part of this issue with testpmd.

I wrote a command to do this, that collided with some stream
rework from Intel at the time (3, 4 years back?), so I did not take the
time to force it through. If there is a need we could discuss about
adding this back. I had needed it to write a PMD, that could be useful
to others.

As you say Viacheslav, there are use-cases that will rely on
fine-grained probe order. However, this patch solves this issue only
regarding PCI devices, depending on other PCI devices. We have in EAL
an improper hack about it, forcing the vdev probe last, because
usually ports depending on others are virtual ones. As this patch shows,
the hack is not sufficient, and as the hack shows, this patch does not
cover everything.

A solution, would be an EAL parameter (I propose --no-dev), that
disable probing for all buses. Applications and devices requiring a
fine-grained probe order, are then free to start in this mode (and maybe
force it through EAL conf), then hotplug ports as they see fit.

This will keep the existing behavior stable for current apps, while allowing
flexibility for the more advanced ones.

-- 
Gaëtan Rivet
6WIND