With the introduction of explicit offloading API in switchdev in commit
2f5dc00f7a ("net: bridge: switchdev: let drivers inform which bridge
ports are offloaded"), we started having Ethernet switch drivers calling
directly into a function exported by net/bridge/br_switchdev.c, which is
a function exported by the bridge driver.
This means that drivers that did not have an explicit dependency on the
bridge before, like cpsw and am65-cpsw, now do - otherwise it is not
possible to call a symbol exported by a driver that can be built as
module unless you are a module too.
There was an attempt to solve the dependency issue in the form of commit
b0e8181762 ("net: build all switchdev drivers as modules when the
bridge is a module"). Grygorii Strashko, however, says about it:
| In my opinion, the problem is a bit bigger here than just fixing the
| build :(
|
| In case, of ^cpsw the switchdev mode is kinda optional and in many
| cases (especially for testing purposes, NFS) the multi-mac mode is
| still preferable mode.
|
| There were no such tight dependency between switchdev drivers and
| bridge core before and switchdev serviced as independent, notification
| based layer between them, so ^cpsw still can be "Y" and bridge can be
| "M". Now for mostly every kernel build configuration the CONFIG_BRIDGE
| will need to be set as "Y", or we will have to update drivers to
| support build with BRIDGE=n and maintain separate builds for
| networking vs non-networking testing. But is this enough? Wouldn't
| it cause 'chain reaction' required to add more and more "Y" options
| (like CONFIG_VLAN_8021Q)?
|
| PS. Just to be sure we on the same page - ARM builds will be forced
| (with this patch) to have CONFIG_TI_CPSW_SWITCHDEV=m and so all our
| automation testing will just fail with omap2plus_defconfig.
In the light of this, it would be desirable for some configurations to
avoid dependencies between switchdev drivers and the bridge, and have
the switchdev mode as completely optional within the driver.
Arnd Bergmann also tried to write a patch which better expressed the
build time dependency for Ethernet switch drivers where the switchdev
support is optional, like cpsw/am65-cpsw, and this made the drivers
follow the bridge (compile as module if the bridge is a module) only if
the optional switchdev support in the driver was enabled in the first
place:
https://patchwork.kernel.org/project/netdevbpf/patch/20210802144813.1152762-1-arnd@kernel.org/
but this still did not solve the fact that cpsw and am65-cpsw now must
be built as modules when the bridge is a module - it just expressed
correctly that optional dependency. But the new behavior is an apparent
regression from Grygorii's perspective.
So to support the use case where the Ethernet driver is built-in,
NET_SWITCHDEV (a bool option) is enabled, and the bridge is a module, we
need a framework that can handle the possible absence of the bridge from
the running system, i.e. runtime bloatware as opposed to build-time
bloatware.
Luckily we already have this framework, since switchdev has been using
it extensively. Events from the bridge side are transmitted to the
driver side using notifier chains - this was originally done so that
unrelated drivers could snoop for events emitted by the bridge towards
ports that are implemented by other drivers (think of a switch driver
with LAG offload that listens for switchdev events on a bonding/team
interface that it offloads).
There are also events which are transmitted from the driver side to the
bridge side, which again are modeled using notifiers.
SWITCHDEV_FDB_ADD_TO_BRIDGE is an example of this, and deals with
notifying the bridge that a MAC address has been dynamically learned.
So there is a precedent we can use for modeling the new framework.
The difference compared to SWITCHDEV_FDB_ADD_TO_BRIDGE is that the work
that the bridge needs to do when a port becomes offloaded is blocking in
its nature: replay VLANs, MDBs etc. The calling context is indeed
blocking (we are under rtnl_mutex), but the existing switchdev
notification chain that the bridge is subscribed to is only the atomic
one. So we need to subscribe the bridge to the blocking switchdev
notification chain too.
This patch:
- keeps the driver-side perception of the switchdev_bridge_port_{,un}offload
unchanged
- moves the implementation of switchdev_bridge_port_{,un}offload from
the bridge module into the switchdev module.
- makes everybody that is subscribed to the switchdev blocking notifier
chain "hear" offload & unoffload events
- makes the bridge driver subscribe and handle those events
- moves the bridge driver's handling of those events into 2 new
functions called br_switchdev_port_{,un}offload. These functions
contain in fact the core of the logic that was previously in
switchdev_bridge_port_{,un}offload, just that now we go through an
extra indirection layer to reach them.
Unlike all the other switchdev notification structures, the structure
used to carry the bridge port information, struct
switchdev_notifier_brport_info, does not contain a "bool handled".
This is because in the current usage pattern, we always know that a
switchdev bridge port offloading event will be handled by the bridge,
because the switchdev_bridge_port_offload() call was initiated by a
NETDEV_CHANGEUPPER event in the first place, where info->upper_dev is a
bridge. So if the bridge wasn't loaded, then the CHANGEUPPER event
couldn't have happened.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Tested-by: Grygorii Strashko <grygorii.strashko@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
957e2235e5