Re: [PATCH net-next V2 00/13] net/mlx5e: Add netdev support for data direct
From: Stanislav Fomichev
Date: Fri Nov 20 2026 - 12:15:11 EST
On 11/20, Dragos Tatulea wrote:
> On 20.11.26 17:38, Mina Almasry wrote:
> > On Fri, Nov 20, 2026 at 6:39 AM Dragos Tatulea <dtatulea@xxxxxxxxxx> wrote:
> >>
> >> Hi all,
> >>
> >> On 10.11.26 19:38, Tariq Toukan wrote:
> >> [...]
> >>>
> >>> .../ethernet/mellanox/mlx5/core/data_direct.c | 9 +-
> >>> drivers/net/ethernet/mellanox/mlx5/core/en.h | 14 +
> >>> .../ethernet/mellanox/mlx5/core/en/params.c | 18 +-
> >>> .../ethernet/mellanox/mlx5/core/en/params.h | 2 +
> >>> .../mellanox/mlx5/core/en/reporter_rx.c | 1 +
> >>> .../mellanox/mlx5/core/en/reporter_tx.c | 1 +
> >>> .../ethernet/mellanox/mlx5/core/en_common.c | 38 ++-
> >>> .../ethernet/mellanox/mlx5/core/en_ethtool.c | 67 +++++
> >>> .../net/ethernet/mellanox/mlx5/core/en_main.c | 253 +++++++++++++++++-
> >>> .../net/ethernet/mellanox/mlx5/core/en_rx.c | 33 +--
> >>> .../net/ethernet/mellanox/mlx5/core/en_tx.c | 4 +-
> >>> .../net/ethernet/mellanox/mlx5/core/main.c | 1 +
> >>> include/linux/mlx5/driver.h | 1 +
> >>> include/net/netdev_queues.h | 1 +
> >>> include/net/netmem.h | 24 ++
> >>> net/core/devmem.c | 85 ++++++
> >>> net/core/devmem.h | 6 +
> >>> net/core/netdev_queues.c | 51 ++++
> >> While working on V3 I realized that keeping the core part in
> >> this series is not the best approach. Initially it made sense
> >> because the change was small.
> >>
> >> I think I should re-spin it in a separate series with netdevsim
> >> based selftests for the new APIs. I would base it on Mina's devmem
> >> leased queues unregister [1].
> >>
> >> Any reason not to?
> >>
> >
> > I honestly wanted to comment that you should base it on the leased
> > queues series from me but didn't want to be annoying, especially
> > because I made you wait for the previous series.
> >
> I realized that quite a few changes were needed after your series
> so I decided to base it on it.
>
> > My series creates binding->bind_dev, and binding->map_dev, and core
> > will (should) handle things correctly across both devices. This seems
> > perfect for you, because you have a bind device and a separate map
> > device and their lifetime is different.
> >
> In my series I am actually skipping leased rx queues in
> net_devmem_unbind_dma_dev() because we need some kind of notification
> mechanism + netkit implementation. I think this should be added later
> if there is a usecase. For now mlx5 simply doesn't support the combo.
>
> One note on net_devmem_unbind_dmabuf(). Posting it here as I don't
> find the right place to mention in your series:
>
> @@ -234,8 +234,17 @@ void net_devmem_unbind_dmabuf(struct net_devmem_dmabuf_binding *binding)
> busy = true;
> }
>
> - if (binding->direction == DMA_FROM_DEVICE && !busy)
> + if (binding->direction == DMA_FROM_DEVICE && !busy) {
> + struct net_device *map_dev = NULL;
> + netdevice_tracker dev_tracker;
> +
> + /* Leased queue: attachment readers hold map_dev's lock. */
> + if (READ_ONCE(binding->map_dev) != binding->bind_dev)
> + map_dev = net_devmem_lock_dev(&binding->map_dev,
> + &dev_tracker);
> net_devmem_dmabuf_binding_unmap(binding);
> + net_devmem_unlock_dev(map_dev, &dev_tracker);
> + }
> WRITE_ONCE(binding->bind_dev, NULL);
> }
>
> map_dev devmem lock is taken to prevent use after free in
> netdev_has_dmabuf_binding() when called in parallel with
> net_devmem_unbind_dmabuf(). As this can happen in my series I added it
> there.
>
> Waiting for feedback from Stan/Jakub/Paolo for a few more hours and
> will post the series tomorrow.
More and more workloads are moving into ns/netkit setup on our side,
so it would be nice to have netkit support as well. We can start with
what you have so you're not blocked for too long, but we do want the
netkit portion as well :-)