On 2018-12-23 5:23 a.m., Andrew Lunn wrote:
On Sat, Dec 22, 2018 at 08:29:08PM -0600, Robert Hancock wrote:
I have a device using a KSZ9897 switch
(CONFIG_MICROCHIP_KSZ_SPI_DRIVER=y) using a 4.19.9 kernel, where I am
trying to enable a bridge on some of the ports using systemd-networkd.
Hi Robert
I don't think many of use use such a setup, so we probably have not
noticed this.
However, it seems to be getting an error when it tries to configure the
ports to be part of the bridge. systemd-networkd complains with:
lan1: Set link
lan1: Could not join netdev: Operation not supported
lan1: Failed
and the kernel complains:
lan: bridge flag offload is not supported 4(lan1)
That message is coming from br_switchdev_set_port_flag in
net/bridge/br_switchdev.c. Adding some more output to that statement
tells me:
lan: bridge flag offload is not supported, flags 18656 mask 64 support
0, 4(lan1)
which appears to mean something is trying to enable BR_BCAST_FLOOD,
BR_MCAST_FLOOD, BR_PROMISC, BR_FLOOD, BR_LEARNING on the port, but it is
failing to enable BR_FLOOD because brport_flags_support is 0. I am just
using the default bridge settings in systemd-networkd, so that is
nothing that I am specifying explicitly.
I think this is a systemd problem. It should first query what flags
are supported. See
dc0ecabd6231 ("net: switchdev: Add support for querying supported bridge flags by
hardware")
And then only try to turn on flags which are supported. If it tries
to turn on flags which are not supported, an error is the correct
thing to do.
I may be missing some context about how the switchdev/DSA code works,
but are those flags actually unsupported, or just unsupported in
hardware offload? Userspace doesn't know or care about the hardware
offload capability of the switch, it's just trying to set standard flags
on the bridge. As far as I can tell, BR_FLOOD is enabled by default (as
a bridge without that set isn't very generally useful), but systemd
trying to explicitly set it to enabled causes the entire userspace
netlink operation to fail. That doesn't seem right.
As I mentioned in my follow-up message, it seems like the function
producing that message shouldn't be failing the operation in this case,
just bailing out in the same way as if the port doesn't support
retrieving switchdev attributes at all.
