On Thu, Mar 10, 2022 at 17:05, Tobias Waldekranz <tobias@xxxxxxxxxxxxxx> wrote:
> On Thu, Mar 10, 2022 at 12:35, Vladimir Oltean <olteanv@xxxxxxxxx> wrote:
>> On Thu, Mar 10, 2022 at 09:54:34AM +0100, Tobias Waldekranz wrote:
>>> >> + if (!dsa_port_can_configure_learning(dp) || dp->learning) {
>>> >> + switch (state->state) {
>>> >> + case BR_STATE_DISABLED:
>>> >> + case BR_STATE_BLOCKING:
>>> >> + case BR_STATE_LISTENING:
>>> >> + /* Ideally we would only fast age entries
>>> >> + * belonging to VLANs controlled by this
>>> >> + * MST.
>>> >> + */
>>> >> + dsa_port_fast_age(dp);
>>> >
>>> > Does mv88e6xxx support this? If it does, you might just as well
>>> > introduce another variant of ds->ops->port_fast_age() for an msti.
>>>
>>> You can limit ATU operations to a particular FID. So the way I see it we
>>> could either have:
>>>
>>> int (*port_vlan_fast_age)(struct dsa_switch *ds, int port, u16 vid)
>>>
>>> + Maybe more generic. You could imagine there being a way to trigger
>>> this operation from userspace for example.
>>> - We would have to keep the VLAN<->MSTI mapping in the DSA layer in
>>> order to be able to do the fan-out in dsa_port_set_mst_state.
>>>
>>> or:
>>>
>>> int (*port_msti_fast_age)(struct dsa_switch *ds, int port, u16 msti)
>>>
>>> + Let's the mapping be an internal affair in the driver.
>>> - Perhaps, less generically useful.
>>>
>>> Which one do you prefer? Or is there a hidden third option? :)
>>
>> Yes, I was thinking of "port_msti_fast_age". I don't see a cheap way of
>> keeping VLAN to MSTI associations in the DSA layer. Only if we could
>> retrieve this mapping from the bridge layer - maybe with something
>> analogous to br_vlan_get_info(), but br_mst_get_info(), and this gets
>> passed a VLAN_N_VID sized bitmap, which the bridge populates with ones
>> and zeroes.
>
> That can easily be done. Given that, should we go for port_vlan_fast_age
> instead? port_msti_fast_age feels like an awkward interface, since I
> don't think there is any hardware out there that can actually perform
> that operation without internally fanning it out over all affected VIDs
> (or FIDs in the case of mv88e6xxx).
>
>> The reason why I asked for this is because I'm not sure of the
>> implications of flushing the entire FDB of the port for a single MSTP
>> state change. It would trigger temporary useless flooding in other MSTIs
>> at the very least. There isn't any backwards compatibility concern to
>> speak of, so we can at least try from the beginning to limit the
>> flushing to the required VLANs.
>
> Aside from the performance implications of flows being temporarily
> flooded I don't think there are any.
>
> I suppose if you've disabled flooding of unknown unicast on that port,
> you would loose the flow until you see some return traffic (or when one
> side gives up and ARPs). While somewhat esoteric, it would be nice to
> handle this case if the hardware supports it.
>
>> What I didn't think about, and will be a problem, is
>> dsa_port_notify_bridge_fdb_flush() - we don't know the vid to flush.
>> The easy way out here would be to export dsa_port_notify_bridge_fdb_flush(),
>> add a "vid" argument to it, and let drivers call it. Thoughts?
>
> To me, this seems to be another argument in favor of
> port_vlan_fast_age. That way you would know the VIDs being flushed at
> the DSA layer, and driver writers needn't concern themselves with having
> to remember to generate the proper notifications back to the bridge.
>
>> Alternatively, if you think that cross-flushing FDBs of multiple MSTIs
>> isn't a real problem, I suppose we could keep the "port_fast_age" method.
>
> What about falling back to it if the driver doesn't support per-VLAN
> flushing? Flushing all entries will work in most cases, at the cost of
> some temporary flooding. Seems more useful than refusing the offload
> completely.
Actually now that I think about it, maybe it is more reasonable to risk
having stale entries in the VLANs where the topology changed, rather
than nuking flows in unrelated VLANs.
>>> > And since it is new code, you could require that drivers _do_ support
>>> > configuring learning before they could support MSTP. After all, we don't
>>> > want to keep legacy mechanisms in place forever.
>>>
>>> By "configuring learning", do you mean this new fast-age-per-vid/msti,
>>> or being able to enable/disable learning per port? If it's the latter,
>>> I'm not sure I understand how those two are related.
>>
>> The code from dsa_port_set_state() which you've copied:
>>
>> if (!dsa_port_can_configure_learning(dp) ||
>> (do_fast_age && dp->learning)) {
>>
>> has this explanation:
>>
>> 1. DSA keeps standalone ports in the FORWARDING state.
>> 2. DSA also disables address learning on standalone ports, where this is
>> possible (dsa_port_can_configure_learning(dp) == true).
>> 3. When a port joins a bridge, it leaves its FORWARDING state from
>> standalone mode and inherits the bridge port's BLOCKING state
>> 4. dsa_port_set_state() treats a port transition from FORWARDING to
>> BLOCKING as a transition requiring an FDB flush
>> 5. due to (2), the FDB flush at stage (4) is in fact not needed, because
>> the FDB of that port should already be empty. Flushing the FDB may be
>> a costly operation for some drivers, so it is avoided if possible.
>>
>> So this is why the "dsa_port_can_configure_learning()" check is there -
>> for compatibility with drivers that can't configure learning => they
>> keep learning enabled also in standalone mode => they need an FDB flush
>> when a standalone port joins a bridge.
>>
>> What I'm saying is: for drivers that offload MSTP, let's force them to
>> get the basics right first (have configurable learning), rather than go
>> forward forever with a backwards compatibility mode.
>
> Makes sense, I'll just move it up to the initial capability check.
