On ons, mar 23, 2022 at 12:16, Vladimir Oltean <olteanv@xxxxxxxxx> wrote: > On Wed, Mar 23, 2022 at 11:13:51AM +0100, Hans Schultz wrote: >> On tis, mar 22, 2022 at 13:08, Vladimir Oltean <olteanv@xxxxxxxxx> wrote: >> > On Tue, Mar 22, 2022 at 12:01:13PM +0100, Hans Schultz wrote: >> >> On fre, mar 18, 2022 at 15:19, Vladimir Oltean <olteanv@xxxxxxxxx> wrote: >> >> > On Fri, Mar 18, 2022 at 02:10:26PM +0100, Hans Schultz wrote: >> >> >> In the offloaded case there is no difference between static and dynamic >> >> >> flags, which I see as a general issue. (The resulting ATU entry is static >> >> >> in either case.) >> >> > >> >> > It _is_ a problem. We had the same problem with the is_local bit. >> >> > Independently of this series, you can add the dynamic bit to struct >> >> > switchdev_notifier_fdb_info and make drivers reject it. >> >> > >> >> >> These FDB entries are removed when link goes down (soft or hard). The >> >> >> zero DPV entries that the new code introduces age out after 5 minutes, >> >> >> while the locked flagged FDB entries are removed by link down (thus the >> >> >> FDB and the ATU are not in sync in this case). >> >> > >> >> > Ok, so don't let them disappear from hardware, refresh them from the >> >> > driver, since user space and the bridge driver expect that they are >> >> > still there. >> >> >> >> I have now tested with two extra unmanaged switches (each connected to a >> >> seperate port on our managed switch, and when migrating from one port to >> >> another, there is member violations, but as the initial entry ages out, >> >> a new miss violation occurs and the new port adds the locked entry. In >> >> this case I only see one locked entry, either on the initial port or >> >> later on the port the host migrated to (via switch). >> >> >> >> If I refresh the ATU entries indefinitly, then this migration will for >> >> sure not work, and with the member violation suppressed, it will be >> >> silent about it. >> > >> > Manual says that migrations should trigger miss violations if configured >> > adequately, is this not the case? >> > >> >> So I don't think it is a good idea to refresh the ATU entries >> >> indefinitely. >> >> >> >> Another issue I see, is that there is a deadlock or similar issue when >> >> receiving violations and running 'bridge fdb show' (it seemed that >> >> member violations also caused this, but not sure yet...), as the unit >> >> freezes, not to return... >> > >> > Have you enabled lockdep, debug atomic sleep, detect hung tasks, things >> > like that? >> >> I have now determined that it is the rtnl_lock() that causes the >> "deadlock". The doit() in rtnetlink.c is under rtnl_lock() and is what >> takes care of getting the fdb entries when running 'bridge fdb show'. In >> principle there should be no problem with this, but I don't know if some >> interrupt queue is getting jammed as they are blocked from rtnetlink.c? > > Sorry, I forgot to respond yesterday to this. > By any chance do you maybe have an AB/BA lock inversion, where from the > ATU interrupt handler you do mv88e6xxx_reg_lock() -> rtnl_lock(), while > from the port_fdb_dump() handler you do rtnl_lock() -> mv88e6xxx_reg_lock()? If I release the mv88e6xxx_reg_lock() before calling the handler, I need to get it again for the mv88e6xxx_g1_atu_loadpurge() call at least. But maybe the vtu_walk also needs the mv88e6xxx_reg_lock()? I could also just release the mv88e6xxx_reg_lock() before the call_switchdev_notifiers() call and reacquire it immediately after?
