Re: [PATCH rdma-rc v2] IB/IPoIB: Fix queue count for non-enhanced IPoIB over netlink

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Jan 26, 2023 at 08:02:00PM +0100, Dragos Tatulea wrote:
> On 01/26, Leon Romanovsky wrote:
> > On Thu, Jan 26, 2023 at 02:40:56PM -0400, Jason Gunthorpe wrote:
> > > On Thu, Jan 26, 2023 at 08:35:18PM +0200, Leon Romanovsky wrote:
> > > > On Tue, Jan 24, 2023 at 08:24:18PM +0200, Leon Romanovsky wrote:
> > > > > From: Dragos Tatulea <dtatulea@nvidia.com>
> > > > > 
> > > > > Make sure that non-enhanced IPoIB queues are configured with only
> > > > > 1 tx and rx queues over netlink. This behavior is consistent with the
> > > > > sysfs child_create configuration.
> > > > > 
> > > > > The cited commit opened up the possibility for child PKEY interface
> > > > > to have multiple tx/rx queues. It is the driver's responsibility to
> > > > > re-adjust the queue count accordingly. This patch does exactly that:
> > > > > non-enhanced IPoIB supports only 1 tx and 1 rx queue.
> > > > > 
> > > > > Fixes: dbc94a0fb817 ("IB/IPoIB: Fix queue count inconsistency for PKEY child interfaces")
> > > > > Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
> > > > > Signed-off-by: Leon Romanovsky <leon@nvidia.coma
> > > > > ---
> > > > > Changelog:
> > > > > v2:
> > > > >  * Changed implementation
> > > > > v1: https://lore.kernel.org/all/752143b0eef72a966662ce94526b1ceb5ba4bbb3.1674234106.git.leon@kernel.org
> > > > >  * Fixed typo in warning print.
> > > > > v0: https://lore.kernel.org/all/4a7ecec08ee30ad8004019818fadf1e58057e945.1674137153.git.leon@kernel.org>
> > > > > ---
> > > > >  drivers/infiniband/ulp/ipoib/ipoib_main.c | 8 ++++++++
> > > > >  1 file changed, 8 insertions(+)
> > > > 
> > > > 
> > > > Dragos pointed to me that I sent commit with "old" commit message.
> > > > The right one is below and I'll fix it locally once will apply it.
> > > > 
> > > > Jason, are you happy with the patch?
> > > 
> > > Why not use min?
> > 
> > It doesn't give anything as we are in legacy IPoIB path and it will be
> > min with 1 anyway.
> > 
> To add to Leon's comment:
> 
> It is making it explicit that IPoIB is using only one queue. Similar to
> how ipoib_alloc_netdev() calls alloc_netdev_mq() with 1 tx and 1 rx queue
> for legacy IPoIB when the parent interface is created and also when
> child interfaces are created over sysfs.

Ok then, Leon please apply it and fix the commit message

Jason