Re: [dpdk-dev] [PATCH 2/2] net/mlx5: avoid unbind step to enable switchdev mode

[Jan Viktorin <viktorin@cesnet.cz>]

On Mon, 15 Mar 2021 19:31:25 +0000
Slava Ovsiienko <viacheslavo@nvidia.com> wrote:

> Hi, Jan

Hello Salva,

I am sorry, but I didn't get the point of your response. For me, there
was 1 PF 0000:05:00.1, SRIOV disabled. So, I followed the guide [1].

1. Enable SRIOV via mlxconfig. Done.

2. Configure max number of VFs. Done.

3. Reset firmware. Done.

In fact, I had to reboot the machine here so I am not convinced the
reset of firmware is the thing. But I can image that there is some
other use case for it. After reboot, I configured 2 VFs: 0000:05:00.4,
0000:05:00.5.

4. Unbind the PF 0000:05:00.1. Done.

5. Fail, no way how to configure switchdev because of previous step.
Thus, I strongly believe that the guide is misleading here.

When I bound the PF 0000:05:00.1 back, I could configure the switchdev
mode and everything was just OK and seemed to be working. So **after**
the bind I configured the switchdev mode properly.

[1] https://doc.dpdk.org/guides/nics/mlx5.html#enable-switchdev-mode

Jan

> 
> 1. Let's suppose we have the one port NIC, PCI BDF 08:00.0, SR-IOV enabled, 0 VFs is configured,
> netdev name is ens1f0 (depends on distro and kernel version). 
> 
> 2. Then, we configure 3 VFs with SR-IOV, we get :
> - 08:00.0 - PF, ens1f0
> - 08:00.1 - VF0, ens1f1
> - 08:00.2 - VF1, ens1f2
> - 08:00.3 - VF2, ens1f3
> 
> 3. Then, we should unbind VFs ONLY, PF should be kept untouched.
> - 08:00.0 - PF, ens1f0
> - 08:00.1 - VF0, -
> - 08:00.2 - VF1, -
> - 08:00.3 - VF2, -
> 
> 4. Then, configure switchdev mode
> - 08:00.0 - PF, ens1f0
> - 08:00.1 - VF0, -
> - 08:00.2 - VF1, -
> - 08:00.3 - VF2, -
> + we get 3 representor netdevs:
> ens1f0_0 (no dedicated PCI device)
> ens1f0_1 (no dedicated PCI device)
> ens1f0_2 (no dedicated PCI device)
> 
> 5. Then, bind VFs back (assume we are no going to map them to VMs)
> - 08:00.0 - PF, ens1f0
> - 08:00.1 - VF0, ens1f1
> - 08:00.2 - VF1, ens1f2
> - 08:00.3 - VF2, ens1f3
> - ens1f0_0 (no dedicated PCI device)
> - ens1f0_1 (no dedicated PCI device)
> - ens1f0_2 (no dedicated PCI device)
> 
> 6. As result we should get 7 netdevs - 1 PF (Uplink representor), 3 VFs, 3 representors (no PCI device, pure netdev)
> 
> With best regards, Slava
> 
> > -----Original Message-----
> > From: Jan Viktorin <viktorin@cesnet.cz>
> > Sent: Monday, March 15, 2021 18:11
> > To: Slava Ovsiienko <viacheslavo@nvidia.com>
> > Cc: dev@dpdk.org; Asaf Penso <asafp@nvidia.com>; Shahaf Shuler
> > <shahafs@nvidia.com>; Matan Azrad <matan@nvidia.com>
> > Subject: Re: [PATCH 2/2] net/mlx5: avoid unbind step to enable switchdev
> > mode
> > 
> > Hello Salva,
> > 
> > On Mon, 15 Mar 2021 15:53:51 +0000
> > Slava Ovsiienko <viacheslavo@nvidia.com> wrote:
> >   
> > > Hi, Jan
> > >
> > > Yes, bullet [4] explicitly requires to unbind VFs, and detach the netdevs  
> > from the mlx5_core driver.  
> > > Otherwise, kernel driver refuses to be configured with switchdev mode in  
> > [5]. So, [4] can't be skipped.  
> > > After setting swithdev mode, VFs can be bound back (if it is needed, and  
> > these ones are not mapped to VMs):
> > 
> > OK, but I believe that it is **not possible** to follow the rule [5].
> > The guide explicitly says in [4] "can be rebind **after** the switchdev
> > mode".
> > Just, if you unbind the device, there is no way how to configure the
> > switchdev mode, this is the contradiction I mentioned in the commit. You
> > cannot configure switchdev mode because the interface is gone and the path
> > /sys/class/net/<net device>/compat/devlink/mode no longer exists.
> > 
> > So, maybe, just the formulation is wrong. So, what is the **exact
> > right** way how to do it? I would change the commit accordingly. Just, let's
> > make it right. Would it work this way?
> > 
> >  # echo -n "<device pci address>" > /sys/bus/pci/drivers/mlx5_core/unbind
> >  # echo -n "<device pci address>" > /sys/bus/pci/drivers/mlx5_core/bind
> >  # echo switchdev > /sys/class/net/<net device>/compat/devlink/mode
> > 
> > It is good to mention that after the rebind, the <net device> can change.
> > 
> > Regards,
> > Jan
> >   
> > >
> > > echo -n "<device pci address>" > > /sys/bus/pci/drivers/mlx5_core/bind
> > >
> > > With best regards,
> > > Slava
> > >  
> > > > -----Original Message-----
> > > > From: Jan Viktorin <iviktorin@fit.vutbr.cz>
> > > > Sent: Monday, March 15, 2021 17:34
> > > > To: dev@dpdk.org
> > > > Cc: Jan Viktorin <viktorin@cesnet.cz>; Asaf Penso
> > > > <asafp@nvidia.com>; Shahaf Shuler <shahafs@nvidia.com>; Slava
> > > > Ovsiienko <viacheslavo@nvidia.com>; Matan Azrad  
> > <matan@nvidia.com>  
> > > > Subject: [PATCH 2/2] net/mlx5: avoid unbind step to enable switchdev
> > > > mode
> > > >
> > > > From: Jan Viktorin <viktorin@cesnet.cz>
> > > >
> > > > The step 4 is a contradiction. It advices to unbind the device from
> > > > the mlx5_core which removes the associated system network interface  
> > (e.g.  
> > > > eth0). In the step 5, the same system network interface (e.g. eth0)
> > > > is required to exist.
> > > >
> > > > Signed-off-by: Jan Viktorin <viktorin@cesnet.cz>
> > > > ---
> > > >  doc/guides/nics/mlx5.rst | 6 +-----
> > > >  1 file changed, 1 insertion(+), 5 deletions(-)
> > > >
> > > > diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst
> > > > index 0a2dc3dee..122d8e0fc 100644
> > > > --- a/doc/guides/nics/mlx5.rst
> > > > +++ b/doc/guides/nics/mlx5.rst
> > > > @@ -1370,11 +1370,7 @@ the DPDK application.
> > > >
> > > >          echo <num of vfs > /sys/class/net/<net  
> > > > device>/device/sriov_numvfs  
> > > >
> > > > -4. Unbind the device (can be rebind after the switchdev mode)::
> > > > -
> > > > -        echo -n "<device pci address>" >
> > > > /sys/bus/pci/drivers/mlx5_core/unbind
> > > > -
> > > > -5. Enable switchdev mode::
> > > > +4. Enable switchdev mode::
> > > >
> > > >          echo switchdev > /sys/class/net/<net  
> > > > device>/compat/devlink/mode  
> > > >
> > > > --
> > > > 2.30.1  
> > >  
>