DSA registers the user ports in dsa_tree_setup_ports() before it sets conduit->dsa_ptr in dsa_tree_setup_conduit(). 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. 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: [ 166.722423] Unable to handle kernel access to user memory outside uaccess routines at virtual address 0000000000000028 [ 166.816908] Internal error: Oops: 0000000096000004 [#1] SMP [ 166.961766] pc : felix_cpu_port_for_conduit.isra.0+0x90/0xb8 [mscc_felix_dsa_lib] [ 166.969289] lr : felix_port_enable+0x120/0x1a0 [mscc_felix_dsa_lib] The oops left RTNL locked. Later operations that waited for RTNL hung, and the board had to be power cycled. NPI mode uses one CPU port for the whole switch. Changing the conduit of a user port can move the NPI port to another CPU port, and the check in felix_port_enable() keeps user ports that still use the old CPU port down. For a physical conduit, compare dp->cpu_dp->index with the NPI port instead. DSA sets dp->cpu_dp before it registers the user port and clears it only after it removes the user port, so the check keeps working while conduit->dsa_ptr is NULL. As long as the physical conduit is set up, conduit->dsa_ptr and dp->cpu_dp point to the same CPU port, so the check gives the same result as before. LAG conduits keep using felix_cpu_port_for_conduit(), which does not read dsa_ptr for them. Fixes: eca70102cfb1 ("net: dsa: felix: add support for changing DSA master") Signed-off-by: Aristo Chen --- The crash was seen on a 7.0-based distribution kernel. I applied this patch to that kernel and probed the switch again at runtime 117 times on an LS1028A RDB, with systemd-networkd bringing up the user ports. Each time, three user ports were opened before the conduit was set up. All of them came up, and there was no crash. The check also still works: after one port was moved to the other conduit, bringing up a port that still used the old conduit failed with "Multiple conduits are not allowed". I also unbound the switch 10 times while bringing up user ports that were down. 22 of those ports came up after the conduit was torn down and before they were removed, with no crash. On net, this patch is only compile-tested. drivers/net/dsa/ocelot/felix.c | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-) diff --git a/drivers/net/dsa/ocelot/felix.c b/drivers/net/dsa/ocelot/felix.c index c99fd3d23bb3..76aabb8e2887 100644 --- a/drivers/net/dsa/ocelot/felix.c +++ b/drivers/net/dsa/ocelot/felix.c @@ -1239,8 +1239,19 @@ static int felix_port_enable(struct dsa_switch *ds, int port, if (ocelot->npi >= 0) { struct net_device *conduit = dsa_port_to_conduit(dp); + int cpu; - if (felix_cpu_port_for_conduit(ds, conduit) != ocelot->npi) { + /* Do not use conduit->dsa_ptr for a physical conduit. It is + * NULL before DSA sets up the conduit and after it tears it + * down, and opening a user port can reach this check in both + * cases. + */ + if (netif_is_lag_master(conduit)) + cpu = felix_cpu_port_for_conduit(ds, conduit); + else + cpu = dp->cpu_dp->index; + + if (cpu != ocelot->npi) { dev_err(ds->dev, "Multiple conduits are not allowed\n"); return -EINVAL; } -- 2.53.0