Hello netdev, Let me start by describing an issue that I have found in the bridge/switchdev replay logic, which will lead to a larger question about how switchdev ports are synchronized with a bridge. Anyway, on the bug! The steps to reproduce the issue are very simple on my board, but I've never seen it before, so it could be that you need a pretty fast (by embedded standards) SMP CPU to trigger the race. Hardware: - Marvell CN9130 SoC - Marvell 88E6393X Switch Steps: 1. Create a bridge with multicast snooping enabled 2. Attach an switch port to the bridge 3. Remove the bridge If (2) is done immediately after (1), then my switch's hardware MDB will be left with two orphan entries after removing the bridge (3). On a system running net-next with the switchdev tracepoint series[1] applied, this is what it looks like (with comm, pid, timestamp etc. trimmed from the trace output to make the lines a bit narrower, with an event number added instead): root@infix-06-0b-00:~$ echo 1 >/sys/kernel/debug/tracing/events/switchdev/enable root@infix-06-0b-00:~$ cat /sys/kernel/debug/tracing/trace_pipe | grep HOST_MDB & [1] 2602 root@infix-06-0b-00:~$ ip link add dev br0 up type bridge mcast_snooping 1 && ip link set dev x3 up master br0 01: switchdev_defer: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x4 orig br0) vid 0 addr 33:33:ff:87:e4:3f 02: switchdev_defer: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x4 orig br0) vid 0 addr 33:33:00:00:00:6a 03: switchdev_call_replay: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x0 orig br0) vid 0 addr 33:33:00:00:00:6a -> 0 04: switchdev_call_replay: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x0 orig br0) vid 0 addr 33:33:ff:87:e4:3f -> 0 05: switchdev_call_blocking: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x4 orig br0) vid 0 addr 33:33:ff:87:e4:3f -> 0 06: switchdev_call_blocking: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x4 orig br0) vid 0 addr 33:33:00:00:00:6a -> 0 07: switchdev_defer: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x4 orig br0) vid 0 addr 33:33:00:00:00:fb 08: switchdev_call_blocking: dev x3 PORT_OBJ_ADD HOST_MDB(flags 0x4 orig br0) vid 0 addr 33:33:00:00:00:fb -> 0 root@infix-06-0b-00:~$ ip link del dev br0 09: switchdev_call_replay: dev x3 PORT_OBJ_DEL HOST_MDB(flags 0x0 orig br0) vid 0 addr 33:33:00:00:00:fb -> 0 10: switchdev_call_replay: dev x3 PORT_OBJ_DEL HOST_MDB(flags 0x0 orig br0) vid 0 addr 33:33:00:00:00:6a -> 0 11: switchdev_call_replay: dev x3 PORT_OBJ_DEL HOST_MDB(flags 0x0 orig br0) vid 0 addr 33:33:ff:87:e4:3f -> 0 root@infix-06-0b-00:~$ mvls atu ADDRESS FID STATE Q F 0 1 2 3 4 5 6 7 8 9 a DEV:0 Marvell 88E6393X 33:33:00:00:00:6a 1 static - - 0 . . . . . . . . . . 33:33:ff:87:e4:3f 1 static - - 0 . . . . . . . . . . ff:ff:ff:ff:ff:ff 1 static - - 0 1 2 3 4 5 6 7 8 9 a root@infix-06-0b-00:~$ Event 01 and 02 are generated when the bridge is parsing an MLD report generated by the kernel to register membership in the interface's link local group address and the All-Snoopers group. As we are in softirq context, the events are placed in the deferred queue. More or less concurrently, the port is joining the bridge, and the DSA layer is sending a SWITCHDEV_BRPORT_OFFLOADED to request a replay of all relevant events that happened before it joined. Since we now find the two host entries in the bridge's MDB, we replay the events (03 and 04). Once the replay is done, at some later point in time, we start to process the deferred events and send 05 and 06 (originally 01 and 02) to the driver again. At this point, the DSA layer will have a recorded refcount of 2 for both of the groups in question, whereas the bridge only holds a single reference count of each membership. Therefore, when we tear down the bridge, the corresponding UNOFFLOADED event will trigger another replay cycle, which will send a single delete event for each group, leaving the DSA layer's refcount at 1. This is confirmed by the mvls output, showing the two groups are still loaded in the ATU. The end (of the bug report) How can we make sure that this does not happen? Broadly, it seems to me like we need the bridge to signal the start and end of a replay sequence (via the deferred queue), which the drivers can then use to determine when it should start/stop accepting deferred messages. I think this also means that we have to guarantee that no new events of a particular class can be generated while we are scanning the database of those objects and generating replay events. Some complicating factors: - There is no single stream of events to synchronize with. - Some calls are deferred for later dispatch - Some are run synchronously in a blocking context - Some are run synchronously in an atomic context - The rhashtable which backs the FDB is designed to be lock-free, making it hard to ensure atomicity between a replay cycle and new addresses being learned in the hotpath - If we take this approach, how can we make sure that most of the driver-side implementation can be shared in switchdev.c, and doesn't have to be reinvented by each driver? I really hope that someone can tell my why this problem is much easier than this! [1]: https://lore.kernel.org/netdev/20240123153707.550795-1-tobias@xxxxxxxxxxxxxx/
