Vladimir Oltean
2021-Feb-09 22:51 UTC
[Bridge] [PATCH v2 net-next 04/11] net: bridge: offload initial and final port flags through switchdev
On Wed, Feb 10, 2021 at 12:01:24AM +0200, Ido Schimmel wrote:> On Tue, Feb 09, 2021 at 10:20:45PM +0200, Vladimir Oltean wrote: > > On Tue, Feb 09, 2021 at 08:51:00PM +0200, Ido Schimmel wrote: > > > On Tue, Feb 09, 2021 at 05:19:29PM +0200, Vladimir Oltean wrote: > > > > So switchdev drivers operating in standalone mode should disable address > > > > learning. As a matter of practicality, we can reduce code duplication in > > > > drivers by having the bridge notify through switchdev of the initial and > > > > final brport flags. Then, drivers can simply start up hardcoded for no > > > > address learning (similar to how they already start up hardcoded for no > > > > forwarding), then they only need to listen for > > > > SWITCHDEV_ATTR_ID_PORT_BRIDGE_FLAGS and their job is basically done, no > > > > need for special cases when the port joins or leaves the bridge etc. > > > > > > How are you handling the case where a port leaves a LAG that is linked > > > to a bridge? In this case the port becomes a standalone port, but will > > > not get this notification. > > > > Apparently the answer to that question is "I delete the code that makes > > this use case work", how smart of me. Thanks. > > Not sure how you expect to interpret this.Next patch (05/11) deletes that explicit notification from dsa_port_bridge_leave, function which is called from dsa_port_lag_leave too, apparently with good reason.> > Unless you have any idea how I could move the logic into the bridge, I > > guess I'm stuck with DSA and all the other switchdev drivers having this > > forest of corner cases to deal with. At least I can add a comment so I'm > > not tempted to delete it next time. > > There are too many moving pieces with stacked devices. It is not only > LAG/bridge. In L3 you have VRFs, SVIs, macvlans etc. It might be better > to gracefully / explicitly not handle a case rather than pretending to > handle it correctly with complex / buggy code. > > For example, you should refuse to be enslaved to a LAG that already has > upper devices such as a bridge. You are probably not handling this > correctly / at all. This is easy. Just a call to > netdev_has_any_upper_dev().Correct, good point, in particular this means that joining a bridged LAG will not get me any notifications of that LAG's CHANGEUPPER because that was consumed a long time ago. An equally valid approach seems to be to check for netdev_master_upper_dev_get_rcu in dsa_port_lag_join, and call dsa_port_bridge_join on the upper if that is present.> The reverse, during unlinking, would be to refuse unlinking if the upper > has uppers of its own. netdev_upper_dev_unlink() needs to learn to > return an error and callers such as team/bond need to learn to handle > it, but it seems patchable.Again, this was treated prior to my deletion in this series and not by erroring out, I just really didn't think it through. So you're saying that if we impose that all switchdev drivers restrict the house of cards to be constructed from the bottom up, and destructed from the top down, then the notification of bridge port flags can stay in the bridge layer?
Ido Schimmel
2021-Feb-10 10:59 UTC
[Bridge] [PATCH v2 net-next 04/11] net: bridge: offload initial and final port flags through switchdev
On Wed, Feb 10, 2021 at 12:51:53AM +0200, Vladimir Oltean wrote:> On Wed, Feb 10, 2021 at 12:01:24AM +0200, Ido Schimmel wrote: > > On Tue, Feb 09, 2021 at 10:20:45PM +0200, Vladimir Oltean wrote: > > > On Tue, Feb 09, 2021 at 08:51:00PM +0200, Ido Schimmel wrote: > > > > On Tue, Feb 09, 2021 at 05:19:29PM +0200, Vladimir Oltean wrote: > > > > > So switchdev drivers operating in standalone mode should disable address > > > > > learning. As a matter of practicality, we can reduce code duplication in > > > > > drivers by having the bridge notify through switchdev of the initial and > > > > > final brport flags. Then, drivers can simply start up hardcoded for no > > > > > address learning (similar to how they already start up hardcoded for no > > > > > forwarding), then they only need to listen for > > > > > SWITCHDEV_ATTR_ID_PORT_BRIDGE_FLAGS and their job is basically done, no > > > > > need for special cases when the port joins or leaves the bridge etc. > > > > > > > > How are you handling the case where a port leaves a LAG that is linked > > > > to a bridge? In this case the port becomes a standalone port, but will > > > > not get this notification. > > > > > > Apparently the answer to that question is "I delete the code that makes > > > this use case work", how smart of me. Thanks. > > > > Not sure how you expect to interpret this. > > Next patch (05/11) deletes that explicit notification from dsa_port_bridge_leave, > function which is called from dsa_port_lag_leave too, apparently with good reason. > > > > Unless you have any idea how I could move the logic into the bridge, I > > > guess I'm stuck with DSA and all the other switchdev drivers having this > > > forest of corner cases to deal with. At least I can add a comment so I'm > > > not tempted to delete it next time. > > > > There are too many moving pieces with stacked devices. It is not only > > LAG/bridge. In L3 you have VRFs, SVIs, macvlans etc. It might be better > > to gracefully / explicitly not handle a case rather than pretending to > > handle it correctly with complex / buggy code. > > > > For example, you should refuse to be enslaved to a LAG that already has > > upper devices such as a bridge. You are probably not handling this > > correctly / at all. This is easy. Just a call to > > netdev_has_any_upper_dev(). > > Correct, good point, in particular this means that joining a bridged LAG > will not get me any notifications of that LAG's CHANGEUPPER because that > was consumed a long time ago. An equally valid approach seems to be to > check for netdev_master_upper_dev_get_rcu in dsa_port_lag_join, and call > dsa_port_bridge_join on the upper if that is present.The bridge might already have a state you are not familiar with (e.g., FDB entry pointing to the LAG), so best to just forbid this. I think it's fair to impose such limitations (assuming they are properly communicated to user space) given it results in a much less buggy/complex code to maintain.> > > The reverse, during unlinking, would be to refuse unlinking if the upper > > has uppers of its own. netdev_upper_dev_unlink() needs to learn to > > return an error and callers such as team/bond need to learn to handle > > it, but it seems patchable. > > Again, this was treated prior to my deletion in this series and not by > erroring out, I just really didn't think it through. > > So you're saying that if we impose that all switchdev drivers restrict > the house of cards to be constructed from the bottom up, and destructed > from the top down, then the notification of bridge port flags can stay > in the bridge layer?I actually don't think it's a good idea to have this in the bridge in any case. I understand that it makes sense for some devices where learning, flooding, etc are port attributes, but in other devices these can be {port,vlan} attributes and then you need to take care of them when a vlan is added / deleted and not only when a port is removed from the bridge. So for such devices this really won't save anything. I would thus leave it to the lower levels to decide.