Re: [PATCH net] net: dsa: felix: fix NULL pointer dereference in felix_port_enable()

From: Aristo Chen

Date: Fri Nov 20 2026 - 01:31:44 EST


Hi Vladimir,

Thanks for the quick review and the suggestion!

On Thu, Nov 19, 2026 at 7:59 PM Vladimir Oltean <vladimir.oltean@xxxxxxx> wrote:
>
> Hi Aristo,
>
> On Thu, Nov 19, 2026 at 10:07:10AM +0000, Aristo Chen wrote:
> > DSA registers the user ports in dsa_tree_setup_ports() before it sets
> > conduit->dsa_ptr in dsa_tree_setup_conduit().
>
> I wish I had seen this when writing commit 11fd667dac31 ("net: dsa:
> setup master before ports"), because I later had to revert it in commit
> 762c2998c9625f642f0d23da7d3f7e4f90665fdf due to some unforeseen MTU
> handling complications. Had I realized the conduit->dsa_ptr
> implications, I would have tried harder to solve the MTU stuff.
>
> > On teardown, it clears
> > conduit->dsa_ptr in dsa_tree_teardown_conduit() before
> > dsa_tree_teardown_ports() removes the user ports. RTNL is not held
> > across either pair of steps, so user space can open a user port while
> > conduit->dsa_ptr is NULL.
>
> Correct, even though conduit->dsa_ptr itself *is* assigned under
> rtnl_lock() since commit c146f9bc195a ("net: dsa: hold rtnl_mutex when
> calling dsa_master_{setup,teardown}").
>
> >
> > In NPI mode, felix_port_enable() calls felix_cpu_port_for_conduit().
> > For a physical conduit, that function reads the CPU port index through
> > conduit->dsa_ptr without checking for NULL. This caused an oops on an
> > LS1028A RDB when systemd-networkd tried to open a user port while the
> > switch was being probed again at runtime:
>
> Yeah, I think this isn't the only problem. conduit->dsa_ptr is accessed
> from the data path as well, so IMO patching place after place is just
> playing the whack-a-mole game.
>
> I realize reapplying commit 11fd667dac31 ("net: dsa: setup master before
> ports") for stable kernels introduces a high regression risk, but what
> about this change instead (not even compile-tested)?
>
> diff --git a/net/dsa/user.c b/net/dsa/user.c
> index 041f9060c8ef..c68ed0389404 100644
> --- a/net/dsa/user.c
> +++ b/net/dsa/user.c
> @@ -403,6 +403,9 @@ static int dsa_user_open(struct net_device *dev)
> struct dsa_port *dp = dsa_user_to_port(dev);
> int err;
>
> + if (!conduit->dsa_ptr)
> + return -ENETDOWN;
> +
> err = dev_open(conduit, NULL);
> if (err < 0) {
> netdev_err(dev, "failed to open conduit %s\n", conduit->name);
>
> I know what you're thinking - if we refuse bringing the interface up
> altogether, won't that regress user space and be even worse than risking
> race conditions with dsa_conduit_setup() assigning dev->dsa_ptr fast enough?

I tested your change on the LS1028A RDB, on a 7.0-based kernel without
the felix patch, with systemd 259. It does prevent the crash. The
conduit was already up in all 10 runtime probes, and I saw no oops. In
10 unbinds, 28 opens during teardown failed with -ENETDOWN, and the
other 2 found the port already gone.

But networkd did not bring the refused ports back. In all 17 runtime
probes where only networkd tried to bring up the ports, swp0-swp2
stayed down until I ran "networkctl reconfigure". In the first seven,
they were still "off (configuring)" after two minutes. Only swp3 came
up on its own.

In four debug runs, networkd logged two -ENETDOWN "retrying" messages
per affected port. Before it retries, networkd checks that the conduit
is up, but not that DSA has set it up. In link_up_or_down_handler(), an
-ENETDOWN reply to the activation request makes link_up_dsa_slave()
queue an "up" request. If that request gets -ENETDOWN too, its handler
tries to queue another one while the failed request is still in the
queue, and request_new() drops the new one as a duplicate. That would
explain the two messages and why the ports stayed down. The handler and
the duplicate check look the same in v262.2, but I only tested 259.

So the core change can leave a port down if an early open used to
succeed and user space does not retry after DSA sets up the conduit. In
these tests, networkd did not retry after setup. For drivers whose early
opens work today, that would be a new regression.

I agree that the packet path still needs a proper lifetime fix. The
felix patch only fixes the crash I reproduced. For a physical conduit,
it checks the user port's assigned CPU port against the NPI port
without reading dsa_ptr, and it leaves the LAG check unchanged. I also
checked the open path and the other drivers' .port_enable callbacks,
and did not find another place that dereferences conduit->dsa_ptr
without checking it. On the same board, all ports came up in 117
runtime probes with the felix patch, and there was no oops in 10
teardown runs.

Would you take the felix patch for net, to fix this crash without
making early opens fail in the core? The setup and teardown order could
then be fixed in net-next.

>
> And the reason why I think this alternative change has a chance is
> because of DSA history, where bringing a user port up had as prerequisite
> to bring the conduit up first. I removed that dependency in commit
> 9d5ef190e561 ("net: dsa: automatically bring up DSA master when opening
> user port"), but user space still has code to deal with this for older
> kernels. See:
> https://github.com/systemd/systemd/blob/v262.2/src/network/networkd-setlink.c#L1081-L1082
> and I'm hoping other network managers than systemd-networkd have similar
> code as well.
>
> So I'm just proposing to play dumb for a while in stable kernels, return
> the -ENETDOWN that used to be established practice, and buy some time to
> do a more proper reordering of the setup/teardown in net-next, which
> gives us the guarantee that we need that conduit->dsa_ptr is a valid
> pointer for the entire registered lifetime of the user ports.

Thanks,
Aristo