[PATCH RFC] net: bridge: mcast: don't clear L2 host_joined on port group deletion

[PATCH RFC] net: bridge: mcast: don't clear L2 host_joined on port group deletion

[Cedric Jehasse via B4 Relay]

From: Cedric Jehasse <cedric.jehasse@luminex.be>

For a static L2 multicast group that has both a host entry and a port
entry, deleting the port entry also removes the host entry, and the
whole group then disappears from "bridge mdb show".

To reproduce:
  bridge mdb add dev br0 port br0  grp 01:02:03:04:05:06 permanent
  bridge mdb add dev br0 port swp1 grp 01:02:03:04:05:06 permanent
  bridge mdb del dev br0 port swp1 grp 01:02:03:04:05:06 permanent
  bridge mdb show # the "port br0" host entry is gone, too

br_multicast_del_pg() processes every non-(*,G) entry through the S,G
path, which removes the port group from br->sg_port_tbl and then calls
br_multicast_sg_del_exclude_ports(). L2 entries are stored in
sg_port_tbl as well, so they take this path too.

When the last port is removed in br_multicast_sg_del_exclude_ports it
sets "sgmp->host_joined = false", clearing the host membership directly
and bypassing br_multicast_host_leave(). With host_joined now false and
no ports left, br_multicast_del_pg() arms the group timer and
br_multicast_group_expired() tears down the whole mdb entry -- even
though the host membership was explicitly and permanently configured
from user space.

Keep removing L2 port groups from sg_port_tbl, but skip the S,G
EXCLUDE-mode handling for them. The host membership of an L2 group is
managed solely via br_multicast_host_join() / br_multicast_host_leave().

Signed-off-by: Cedric Jehasse <cedric.jehasse@luminex.be>
---
For a static L2 multicast group that has both a host entry and a port
entry, deleting the port entry also removes the host entry, and the
whole group then disappears from "bridge mdb show".

To reproduce:
  bridge mdb add dev br0 port br0  grp 01:02:03:04:05:06 permanent
  bridge mdb add dev br0 port swp1 grp 01:02:03:04:05:06 permanent
  bridge mdb del dev br0 port swp1 grp 01:02:03:04:05:06 permanent
  bridge mdb show # the "port br0" host entry is gone, too

br_multicast_del_pg() processes every non-(*,G) entry through the S,G
path, which removes the port group from br->sg_port_tbl and then calls
br_multicast_sg_del_exclude_ports(). L2 entries are stored in
sg_port_tbl as well, so they take this path too.

When the last port is removed in br_multicast_sg_del_exclude_ports it
sets "sgmp->host_joined = false", clearing the host membership directly
and bypassing br_multicast_host_leave(). With host_joined now false and
no ports left, br_multicast_del_pg() arms the group timer and
br_multicast_group_expired() tears down the whole mdb entry -- even
though the host membership was explicitly and permanently configured
from user space.

Keep removing L2 port groups from sg_port_tbl, but skip the S,G
EXCLUDE-mode handling for them. The host membership of an L2 group is
managed solely via br_multicast_host_join() / br_multicast_host_leave().

This fixes the issue, but i'd like a second opinion on if this is the
correct way to fix it.
---
 net/bridge/br_multicast.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/net/bridge/br_multicast.c b/net/bridge/br_multicast.c
index 881d866d687a..d718a6d1ddb1 100644
--- a/net/bridge/br_multicast.c
+++ b/net/bridge/br_multicast.c
@@ -816,7 +816,13 @@ void br_multicast_del_pg(struct net_bridge_mdb_entry *mp,
 	if (!br_multicast_is_star_g(&mp->addr)) {
 		rhashtable_remove_fast(&br->sg_port_tbl, &pg->rhnode,
 				       br_sg_port_rht_params);
-		br_multicast_sg_del_exclude_ports(mp);
+		/* L2 entries share sg_port_tbl with S,G entries but have no
+		 * *,G/S,G EXCLUDE-mode semantics; their host membership is
+		 * managed explicitly via br_multicast_host_join()/leave() and
+		 * must not be cleared here when the last port group is removed.
+		 */
+		if (!br_group_is_l2(&mp->addr))
+			br_multicast_sg_del_exclude_ports(mp);
 	} else {
 		br_multicast_star_g_handle_mode(pg, MCAST_INCLUDE);
 	}

---
base-commit: 022bdd9c0d036863c4bacd1688b73c6be3001cee
change-id: 20260609-mdb_l2_host_joined_fix-fb2de21580c7

Best regards,
-- 
Cedric Jehasse <cedric.jehasse@luminex.be>

Re: [PATCH RFC] net: bridge: mcast: don't clear L2 host_joined on port group deletion

[Ido Schimmel]

On Wed, Jun 10, 2026 at 10:31:23AM +0200, Cedric Jehasse via B4 Relay wrote:
> From: Cedric Jehasse <cedric.jehasse@luminex.be>
> 
> For a static L2 multicast group that has both a host entry and a port
> entry, deleting the port entry also removes the host entry, and the
> whole group then disappears from "bridge mdb show".
> 
> To reproduce:
>   bridge mdb add dev br0 port br0  grp 01:02:03:04:05:06 permanent
>   bridge mdb add dev br0 port swp1 grp 01:02:03:04:05:06 permanent
>   bridge mdb del dev br0 port swp1 grp 01:02:03:04:05:06 permanent
>   bridge mdb show # the "port br0" host entry is gone, too

Please show the output in the commit message and also show that this
differs from regular (*, G) entries where the host entry is not removed
following the deletion of the port entry.

> 
> br_multicast_del_pg() processes every non-(*,G) entry through the S,G
> path, which removes the port group from br->sg_port_tbl and then calls
> br_multicast_sg_del_exclude_ports(). L2 entries are stored in
> sg_port_tbl as well, so they take this path too.
> 
> When the last port is removed in br_multicast_sg_del_exclude_ports it
> sets "sgmp->host_joined = false", clearing the host membership directly
> and bypassing br_multicast_host_leave(). With host_joined now false and
> no ports left, br_multicast_del_pg() arms the group timer and
> br_multicast_group_expired() tears down the whole mdb entry -- even
> though the host membership was explicitly and permanently configured
> from user space.
> 
> Keep removing L2 port groups from sg_port_tbl, but skip the S,G
> EXCLUDE-mode handling for them. The host membership of an L2 group is
> managed solely via br_multicast_host_join() / br_multicast_host_leave().
> 
> Signed-off-by: Cedric Jehasse <cedric.jehasse@luminex.be>

The patch seems OK to me, but please add a test case in bridge_mdb.sh.

I checked the code and AFAICT this never worked, so target at net-next
without a fixes tag: Support for L2 multicast groups was added in
955062b03fa62, but at this point the mode handling already existed in
br_multicast_del_pg().