Skip to contentSkip to conversation

net: bridge: fix client roaming from DSA user port

RFC PATCH In review DENG Qingfang 4 participants 7 messages Apr 19, 2020 – Dec 18, 2020

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 net: bridge: fix client roaming from DSA user port

When a client roams from a DSA user port to a soft-bridged port (such as WiFi interface), the left-over MAC entry in the switch HW is not deleted, causing inconsistency between Linux fdb and the switch MAC table. As a result, the client cannot talk to other hosts which are on that DSA user port until the MAC entry expires.

Solve this by notifying switchdev fdb to delete the leftover entry when an entry is updated. Remove the added_by_user check in DSA

---
I tried this on mt7530 and mv88e6xxx, but only mt7530 works.
In previous discussion[1], Andrew Lunn said "try playing with auto learning for the CPU port" but it didn't work on mv88e6xxx either

I think commit 7e26bf45e4cb ("net: bridge: allow SW learn to take over HW fdb entries") already tried to fix this issue..

[1] https://lore.kernel.org/netdev/[email protected]/

 net/bridge/br_fdb.c | 3 +++
 net/dsa/slave.c     | 2 --
 2 files changed, 3 insertions(+), 2 deletions(-)
net/bridge/br_fdb.c +3 -0
index 4877a0db16c6..46003e78f2ac 100644
@@ -579,6 +579,9 @@ void br_fdb_update(struct net_bridge *br, struct net_bridge_port *source,
 			/* fastpath: update of existing entry */
 			if (unlikely(source != fdb->dst &&
 				     !test_bit(BR_FDB_STICKY, &fdb->flags))) {
+				/* Remove the entry in HW */
+				br_switchdev_fdb_notify(fdb, RTM_DELNEIGH);
+
 				fdb->dst = source;
 				fdb_modified = true;
 				/* Take over HW learned entry */
net/dsa/slave.c +0 -2
index e94eb1aac602..6133a1be1a74 100644
@@ -1932,8 +1932,6 @@ static void dsa_slave_switchdev_event_work(struct work_struct *work)
 
 	case SWITCHDEV_FDB_DEL_TO_DEVICE:
 		fdb_info = &switchdev_work->fdb_info;
-		if (!fdb_info->added_by_user)
-			break;
 
 		err = dsa_port_fdb_del(dp, fdb_info->addr, fdb_info->vid);
 		if (err) {
Signature · 1 line
2.26.0
maintainer
#1 lore raw
Quoted 6 lines · DENG Qingfang

When a client roams from a DSA user port to a soft-bridged port (such as WiFi interface), the left-over MAC entry in the switch HW is not deleted, causing inconsistency between Linux fdb and the switch MAC table. As a result, the client cannot talk to other hosts which are on that DSA user port until the MAC entry expires.

Quoted 8 lines

Solve this by notifying switchdev fdb to delete the leftover entry when an entry is updated. Remove the added_by_user check in DSA

---
I tried this on mt7530 and mv88e6xxx, but only mt7530 works.
In previous discussion[1], Andrew Lunn said "try playing with auto learning for the CPU port" but it didn't work on mv88e6xxx either

Hi Deng

We should probably first define how we expect moving MAC to work. Then we can make any core fixes, and driver fixes.

For DSA, we have assumed that the software bridge and the hardware bridge are independent, each performs its own learning. Only static entries are kept in sync.

How should this separate learning work for a MAC address which moves?

Andrew

#2 lore raw
Quoted 19 lines · Andrew Lunn
Quoted 6 lines · DENG Qingfang

When a client roams from a DSA user port to a soft-bridged port (such as WiFi interface), the left-over MAC entry in the switch HW is not deleted, causing inconsistency between Linux fdb and the switch MAC table. As a result, the client cannot talk to other hosts which are on that DSA user port until the MAC entry expires.

Quoted 8 lines

Solve this by notifying switchdev fdb to delete the leftover entry when an entry is updated. Remove the added_by_user check in DSA

---
I tried this on mt7530 and mv88e6xxx, but only mt7530 works.
In previous discussion[1], Andrew Lunn said "try playing with auto learning for the CPU port" but it didn't work on mv88e6xxx either

Hi Deng

Quoted 8 lines

We should probably first define how we expect moving MAC to work. Then we can make any core fixes, and driver fixes.

For DSA, we have assumed that the software bridge and the hardware bridge are independent, each performs its own learning. Only static entries are kept in sync.

How should this separate learning work for a MAC address which moves?

When a client moves from a hardware port (e.g. sw0p1) to a software port (wlan0) or another hardware port that belongs to a different switch (sw1p1), that MAC entry
in sw0's MAC table should be deleted, or replaced with the CPU port as destination,
by DSA. Otherwise the client is unable to talk to other hosts on sw0 because sw0 still thinks the client is on sw0p1.

Discussion in OpenWrt GitHub:
https://github.com/openwrt/openwrt/pull/2798#issuecomment-602221971

Quoted 2 lines

Andrew

maintainer
#3 lore raw
Quoted 7 lines

When a client moves from a hardware port (e.g. sw0p1) to a software port (wlan0) or another hardware port that belongs to a different switch (sw1p1), that MAC entry
in sw0's MAC table should be deleted, or replaced with the CPU port as destination,
by DSA. Otherwise the client is unable to talk to other hosts on sw0 because sw0 still thinks the client is on sw0p1.

The MAC address needs to move, no argument there. But what are the mechanisms which cause this. Is learning sufficient, or does DSA need to take an active role?

Forget about DSA for the moment. How does this work for two normal bridges? Is the flow of unicast packets sufficient? Does it requires a broadcast packet from the device after it moves?

Andrew

#4 lore raw

Hi!

Quoted 4 lines · Andrew Lunn

The MAC address needs to move, no argument there. But what are the mechanisms which cause this. Is learning sufficient, or does DSA need to take an active role?

cpu port learning will break switch operation if for whatever reason we want to disable bridge offloading (e.g. ebtables?). In this case a packet received from cpu port need to be sent back through cpu port to another switch port, and the switch will learn from this packet incorrectly.

If we want cpu port learning to kick in, we need to make sure that:
1. When bridge offload is enabled, the switch takes care of packet flooding on switch ports itself, instead of flooding with software bridge.
2. Software bridge shouldn't forward any packet between ports on the same switch.
3. cpu port learning should only be enabled when bridge offloading is used.

Otherwise we need to manually sync fdb between software bridge and switch, specifically we need to take over fdb management on cpu port and tell the switch what devices are on the software bridge port side.

I haven't read kernel bridge code thoroughly and have no idea which one is better/easier.

Some switches (e.g. mt753x) have an option to forward packets with unknown destination port to cpu port only, instead of flooding. For this type of switch, the solution proposed in this patch is fine, because removing fdb entries has the same effect as telling switch that a device is on cpu port. If there's a switch without this feature, (which I have no idea if it exists) there will be issues on packet flooding behavior.

Quoted 4 lines

Forget about DSA for the moment. How does this work for two normal bridges? Is the flow of unicast packets sufficient? Does it requires a broadcast packet from the device after it moves?

It doesn't have to be a broadcast packet but it needs a packet to go through both bridges.

Say we have bridge A and bridge B, port A1 and B1 are connected together and a device is on port A2 first:
Bridge A knows that this device is on port A2 and will forward traffic through A1 to B1 if needed. Bridge B sees these packets and knows device is on port B1.
When the device move from A2 to B2, B updates its fdb and if a packet reaches A, A will update its fdb too.

The problem addressed in this patch is that switch doesn't update its fdb when a device moves from switch to software bridge, because cpu port learning is disabled.

Signature · 2 lines
Regards,
Chuanhong Guo
maintainer
#5 lore raw
Quoted 14 lines · Chuanhong Guo

Hi!

Quoted 4 lines · Andrew Lunn

The MAC address needs to move, no argument there. But what are the mechanisms which cause this. Is learning sufficient, or does DSA need to take an active role?

cpu port learning will break switch operation if for whatever reason we want to disable bridge offloading (e.g. ebtables?). In this case a packet received from cpu port need to be sent back through cpu port to another switch port, and the switch will learn from this packet incorrectly.

Quoted 4 lines

If we want cpu port learning to kick in, we need to make sure that:
1. When bridge offload is enabled, the switch takes care of packet flooding on switch ports itself, instead of flooding with software bridge.

Hi Chuanhong

This is what the skb->offload_fwd_mark is all about. If this is set to 1, it means the switch has done all the forwarding needed for ports in that switch. Most of the tag drivers set this unconditionally true.

Quoted 2 lines

2. Software bridge shouldn't forward any packet between ports on the same switch.

If skb->offload_fwd_mark is true, it won't.

Quoted 2 lines

3. cpu port learning should only be enabled when bridge offloading is used.

So it should be safe for most switch drivers. And the ones which don't set offload_fwd_mark are probably relying of software bridging, or are broken and replicating frames.

Quoted 10 lines

It doesn't have to be a broadcast packet but it needs a packet to go through both bridges.

Say we have bridge A and bridge B, port A1 and B1 are connected together and a device is on port A2 first:
Bridge A knows that this device is on port A2 and will forward traffic through A1 to B1 if needed. Bridge B sees these packets and knows device is on port B1.
When the device move from A2 to B2, B updates its fdb and if a packet reaches A, A will update its fdb too.

The issue here is 'if a packet reaches A'. B might have no reason to send a unicast packet to A, if none of the destinations the device is talking to is reached via A. Which is why i think a broadcast/multicast packet is more likely to be involved.

Andrew

#6 lore raw

Hi Andrew and Chuanhong,

What is the status of this patch?

I am running into the same problem on my wrt3200acm and wrt1900acs, running on linux kernel, dsa driver and ubuntu.

I tried the patch below, without any luck. Although I have conceived two other work-arounds, I am curious if there is a real fix for this issue.

drivers/net/dsa/mv88e6xxx/chip.c    2020-12-18 15:01:58.010404773 +0100 renamed from drivers/net/dsa/mv88e6xxx/chip.c    2020-10-03 11:21:32.000000000 +0200 +0 -3
diff -Naur a/drivers/net/dsa/mv88e6xxx/chip.c b/drivers/net/dsa/mv88e6xxx/chip.c
@@ -2476,3 +2475,0 @@
-    /* Disable learning for CPU port */
-    if (dsa_is_cpu_port(ds, port))
-        reg = 0;

Regards,

Eric Woudstra