When a bridge FDB entry roams between ports, br_fdb_update() notifies
switchdev of the deletion and the addition without a lock. With two
CPUs learning one address on different ports, DSA counts the host
address on its CPU port more or fewer times than the bridge has it,
which ends in "failed to delete ... from fdb: -2" or an address left on
the CPU port.
The earlier attempt took hash_lock in br_fdb_update(); Nikolay rejected
it because it slows down learning and roaming. This series leaves
br_fdb_update() and the rest of the bridge's learning path unchanged
and fixes DSA instead: each user port records the host addresses it
holds and, on every event, checks where the bridge has the entry now.
Patch 1 is the only bridge change: an RCU lookup that DSA calls from
its workqueue. Patch 2 replaces the counting.
The counting has been exposed to the race since commit 90dc8fd36078
("net: bridge: notify switchdev of disappearance of old FDB entry upon
migration"). The change redesigns how DSA tracks host addresses, so it
goes to net-next without a Fixes: tag.
Tested in QEMU with dsa_loop given FDB operations and assisted
learning, PROVE_LOCKING, PROVE_RCU and DEBUG_ATOMIC_SLEEP enabled, one
source address sent from two veth ports on two CPUs for 5 s, 5 boots
per kernel: net-next logs the -2 error and leaves the address on the
CPU port in 4 of 5 boots, this series in none, and it removes the
address when the user ports leave the bridge. A "bridge fdb replace" of
the address onto a user port leaves it on the CPU port in 5 of 5 boots
on net-next and removes it in 5 of 5 with this series. The kernel log
of every boot has no lockdep, RCU or atomic sleep reports.
Changes since the hash_lock patch:
- Leave br_fdb_update() unchanged (Nikolay).
- Sashiko's finding on that patch, a deletion that races the locked
roam, has no counterpart here: DSA looks the entry up when it handles
each event and treats a deleted entry as gone.
- Of the three findings it lists as pre-existing, the MAB and
BR_FDB_LOCAL ones are in the bridge, which this series does not
change; the peer-site one, a "bridge fdb replace" onto a DSA user port
that notifies no deletion, is handled in patch 2.
Link: https://lore.kernel.org/netdev/20261004124254.3525496-1-julius@bairaktaris.de/ (local)
I used Claude (Anthropic) to help write these patches.
Julius Bairaktaris (2):
net: bridge: switchdev: add br_switchdev_fdb_find_port()
net: dsa: fix host fdb refcount when a bridge fdb entry roams
include/linux/if_bridge.h | 13 +++
net/bridge/br_switchdev.c | 63 +++++++++--
net/dsa/port.c | 36 +++++--
net/dsa/port.h | 6 ++
net/dsa/user.c | 216 +++++++++++++++++++++++++++++++++++---
net/dsa/user.h | 8 ++
6 files changed, 309 insertions(+), 33 deletions(-)
base-commit: fa5196392691051963726787944db872420dfa75
--
2.53.0