Re: [PATCH/RFC net] net: dsa: lantiq_gswip: honor dsa_db passed to port_fdb_{add,del}

[Daniel Golle <daniel@makrotopia.org>]

On Sun, Aug 10, 2025 at 07:32:29PM +0300, Vladimir Oltean wrote:
> On Sun, Aug 10, 2025 at 03:48:30PM +0100, Daniel Golle wrote:
> > [   66.300000] gswip 1e108000.switch: port 3 failed to add 6a:94:c2:xx:xx:xx vid 1 to fdb: -22
> > [   66.300000] gswip 1e108000.switch: port 3 failed to add 1a:f8:a8:xx:xx:xx vid 0 to fdb: -22
> > [   66.320000] gswip 1e108000.switch: port 3 failed to add 1a:f8:a8:xx:xx:xx vid 1 to fdb: -22
> > [   66.320000] gswip 1e108000.switch: port 3 failed to delete 6a:94:c2:xx:xx:xx vid 1 from fdb: -2
> > 
> > So the problem is apparently that at the point of calling br_add_if() the
> > port obviously isn't (yet) a member of the bridge and hence
> > dsa_port_bridge_dev_get() would still return NULL at this point, which
> > then causes gswip_port_fdb() to return -EINVAL.
> 
> Nope, this theory is false because the user port _is_ a member of the
> bridge when it processes the SWITCHDEV_FDB_ADD_TO_DEVICE events.
> 
> There are 2 cases for handling these events. One is handling past events
> which were missed and are re-generated during FDB replay:
> 
> [   65.510000] [<807ed128>] dsa_user_fdb_event+0x110/0x1c8
> [   65.510000] [<807f89b8>] __switchdev_handle_fdb_event_to_device+0x138/0x228
> [   65.510000] [<807f8ad8>] switchdev_handle_fdb_event_to_device+0x30/0x48
> [   65.510000] [<807ec328>] dsa_user_switchdev_event+0x90/0xb0
> [   65.510000] [<807c43c0>] br_switchdev_fdb_replay+0xd0/0x138
> [   65.510000] [<807c4de8>] br_switchdev_port_offload+0x240/0x39c
> [   65.510000] [<80799b6c>] br_switchdev_blocking_event+0x80/0xec
> [   65.510000] [<80065e20>] raw_notifier_call_chain+0x48/0x88
> [   65.510000] [<807f83e0>] switchdev_bridge_port_offload+0x5c/0xd0
> [   65.510000] [<807e4c90>] dsa_port_bridge_join+0x170/0x410
> [   65.510000] [<807ed5fc>] dsa_user_changeupper.part.0+0x40/0x180
> [   65.510000] [<807f0ac0>] dsa_user_netdevice_event+0x5b4/0xc34
> [   65.510000] [<80065e20>] raw_notifier_call_chain+0x48/0x88
> [   65.510000] [<805edeec>] __netdev_upper_dev_link+0x1bc/0x450
> [   65.510000] [<805ee1dc>] netdev_master_upper_dev_link+0x2c/0x38
> [   65.510000] [<807a055c>] br_add_if+0x494/0x890
> 
> If you look at the order of operations, you'll see that:
> 
> int dsa_port_bridge_join(struct dsa_port *dp, struct net_device *br,
> 			 struct netlink_ext_ack *extack)
> {
> 	...
> 	err = dsa_port_bridge_create(dp, br, extack); // this sets dp->bridge
> 	if (err)
> 		return err;
> 
> 	brport_dev = dsa_port_to_bridge_port(dp);
> 
> 	info.bridge = *dp->bridge;
> 	err = dsa_broadcast(DSA_NOTIFIER_BRIDGE_JOIN, &info); // this calls ds->ops->port_bridge_join()
> 	if (err)
> 		goto out_rollback;
> 
> 	/* Drivers which support bridge TX forwarding should set this */
> 	dp->bridge->tx_fwd_offload = info.tx_fwd_offload;
> 
> 	err = switchdev_bridge_port_offload(brport_dev, dev, dp,
> 					    &dsa_user_switchdev_notifier,
> 					    &dsa_user_switchdev_blocking_notifier,
> 					    dp->bridge->tx_fwd_offload, extack); // this calls br_switchdev_fdb_replay()
> 	if (err)
> 		goto out_rollback_unbridge;
> 	...
> }
> 
> by the time br_switchdev_fdb_replay() is called, dp->bridge correctly
> reflects the bridge that is generating the events.
> 
> The problem is not a race condition, the problem is that the driver does
> not correctly handle host FDB entries.
> 
> The truly revealing step is to uncomment this:
> 
> 	netdev_dbg(dev, "%s FDB entry towards %s, addr %pM vid %d%s\n",
> 		   event == SWITCHDEV_FDB_ADD_TO_DEVICE ? "Adding" : "Deleting",
> 		   orig_dev->name, fdb_info->addr, fdb_info->vid,
> 		   host_addr ? " as host address" : "");
> 
> and see it will print "as host address", meaning dsa_port_bridge_host_fdb_add()
> will be called.

Thank you for explaining the details, it makes perfect sense to me now.

> 
> At the DSA cross-chip notifier layer, this generates a DSA_NOTIFIER_HOST_FDB_ADD
> event rather than the port-level DSA_NOTIFIER_FDB_ADD. The major difference in
> handling is that HOST_FDB_ADD events are matched by the *CPU* port, see
> dsa_port_host_address_match().
> 
> The CPU port is not part of the bridge that generated the host FDB entry,
> only the user port it services is.
> 
> The problem originates, roughly speaking, since commit 10fae4ac89ce
> ("net: dsa: include bridge addresses which are local in the host fdb
> list"), which first appeared in v5.14. We rolled out a new feature using
> existing API, and didn't notice the gswip driver wouldn't tolerate
> ds->ops->port_fdb_add() getting called on the CPU port.
> 
> Anyway, your intuition for fixing this properly was somewhat correct.
> Even if gswip does not implement FDB isolation, it is correct to look at
> the dsa_db :: bridge member to see which bridge originated the host FDB
> entry. That was its exact purpose, as documented in section "Address
> databases" of Documentation/networking/dsa/dsa.rst. Even if cpu_dp->bridge
> is NULL (as expected), drivers have to know on behalf of which user port
> (member of a bridge) the CPU port is filtering an entry towards the host.
> This is because CPU and DSA ports are servicing multiple address databases.
> 
> The only problem is that the API you're making use for fixing this was
> introduced in commit c26933639b54 ("net: dsa: request drivers to perform
> FDB isolation"), first appeared in v5.18. Thus, you can't fix an issue
> in linux-5.15.y using this method, unless the API change is backported.
> 
> Alternatively, for stable kernels you could suppress the error and
> return 0 in the gswip driver, with the appropriate comment that the API
> doesn't communicate the bridge ID for host FDB entries, and thus, the
> driver just won't filter them. Then, you could develop the solution
> further for net-next, keeping in mind how things are supposed to work.

As it would be nice to have the proper fix backported at least all the way
down to linux-6.1.y, do you think it would be ok to have that solution
I proposed (and picked from the GPL-2.0 licensed vendor driver) applied to
the 'net' tree (with a more appropriate Fixes: tag and commit description,
obviously) and either just not fix it for linux-5.15.y, or only there
replace the 'return -EINVAL;' with a 'dev_warn(...); return 0;'?

In fact, commit c26933639b54 ("net: dsa: request drivers to perform FDB
isolation") also touches drivers/net/dsa/lantiq_gswip.c and does add
struct dsa_db as parameter for the .port_fdb_{add,del} ops. Would it be
ok to hence target that commit in the Fixes: tag?