Re: [PATCH net-next RFC 05/12] net: dsa: Add support for learning FDB through notification

On 07/11/2017 06:05 PM, Nikolay Aleksandrov wrote:

On 11/07/17 13:26, Arkadi Sharshevsky wrote:

quoted

On 07/10/2017 11:59 PM, Vivien Didelot wrote:

quoted

Hi Arkadi,

Arkadi Sharshevsky [off-list ref] writes:

quoted

quoted

quoted

+		err = dsa_port_fdb_add(p->dp, fdb_info->addr, fdb_info->vid);
+		if (err) {
+			netdev_dbg(dev, "fdb add failed err=%d\n", err);
+			break;
+		}
+		call_switchdev_notifiers(SWITCHDEV_FDB_OFFLOADED, dev,
+					 &fdb_info->info);
+		break;
+
+	case SWITCHDEV_FDB_DEL_TO_DEVICE:
+		fdb_info = &switchdev_work->fdb_info;
+		err = dsa_port_fdb_del(p->dp, fdb_info->addr, fdb_info->vid);
+		if (err)
+			netdev_dbg(dev, "fdb del failed err=%d\n", err);

OK I must have missed from the off-list discussion why we are not
calling the switchdev notifier here?

We do not agree on it actually, that is why it was moved to the list.
I think that delete should succeed, you should retry until succession.

The deletion is done under spinlock in the bridge so you cannot block,
thus delete cannot fail due to hardware failure. Calling it here doesn't
make sense because the bridge probably already deleted this FDB.

So as we discussed, the problem here is that if dsa_port_fdb_del fails
for some probable reasons (MDIO timeout, weak GPIO lines, etc.), Linux
bridge will delete the entry in software, dumping bridge fdb will show
nothing, but the entry would still be programmed in hardware and the
network can thus be inconsistent, unsupposedly switching frames.

IMHO the correct way for bridge to use the notification chain is to make
SWITCHDEV_FDB_DEL_TO_DEVICE symmetrical to SWITCHDEV_FDB_ADD_TO_DEVICE:
if an entry has been marked as offloaded, bridge must mark the entry as
to-be-deleted and do not delete the software entry until the driver
notifies back the successful deletion.

If that is hardly feasible due to some bridge limitations, we must
explain this in a comment and use something more explosive than a simple
netdev_dbg to warn the user about the broken network setup...


Thanks,

        Vivien

Hi Nikolay,

Vivien raised inconsistency issue with the current switchdev
notification chain in case of FDB del. In case of static FDB delete,
notification will be sent to the driver, followed by deletion of the
software entry without waiting for the hardware delete. In case of
hardware deletion failure the consecutive FDB dump will not show the
deleted entry, yet, the entry will stay in hardware.

The deletion is done under lock thus the hardware deletion is deferred,
and cannot fail due to hardware removal failure. Thus the above proposed
solution by Vivien can lead to confusing situation:

1. User deletes the entry
2. Deletion succeed
3. User dumps FDB and still sees this entry due to hardware failure,
   what should he do? retry to delete until the FDB dump will not show
   the entry?

Would like to hear you opinion about this solution.

IMHO in this case the driver should retry to delete, in case of
several retries the driver should maybe:
1. Trap the traffic to CPU (dint think it possible in case of DSA).
2. Disable the port (its more explosive then netdev_dbg).

Thanks,
Arkadi

Hi,
Looking at the code - it would seem that retrying is the only current option
with the way these switchdev notifications are handled. They cannot fail, meaning
from the bridge POV these ops must always succeed and errors are ignored, so the
driver should do everything possible to stay in sync, and in case all fails
then disabling the port seems like the best option to me, to show that something is
clearly wrong and avoid further issues, but DSA maintainers can comment more
on how to handle failure.

That being said:
This sounds a lot like the switchdev notifications vs callbacks discussions that we've
had in the past. Also what happened with the prepare+commit and all that ? If the hash_lock
is the main problem let's work towards improving that and making the fdb code handle
switchdev similar to the vlan code.

Cheers,
 Nik

Vlans can be only added by user under rtnl so its possible to sleep. On
the other hand FDBs can be learned in atomic context, thus the
notification chain is atomic. One Possible way I thought about doing it
is to maintain bridge internal ordered workqueue for the FDBs learned in
atomic context, furthermore the FDB table will be protected by mutex.
The STATIC entries are added in user context so the user will add
directly to this hash table and could sleep.

Its not very good, but I don't think about another solution here.

Regarding notification or not, drivers should know about FDBs that are
not directly related to them like in case of vxlan (vxlan device is not
upper device of some siwthcdev port) so notification is better approach
in my opinion.

Regarding prepare/commit, it is not used in mlxsw. And we would like
to remove it from DSA (in case of FDBs). The prepare phase cannot
assure success because errors could happen unrelated to reserved space
in the switch in the commit stage..

Thanks,
Arkadi