Re: DSA: Suspicious RCU usage (via rtnl_bridge_getlink)

Tue, Sep 20, 2016 at 04:32:53PM CEST, vivien.didelot@savoirfairelinux.com wrote:

Hi Andrew, Russell,

Andrew Lunn [off-list ref] writes:

quoted

On Tue, Sep 20, 2016 at 11:26:12AM +0100, Russell King - ARM Linux wrote:

quoted

Issuing "bridge vlan show" on clearfog provokes a "suspicious RCU usage"
warning from the kernel (see below).

As it's illegal to schedule while holding the RCU read lock, there's the
possibility for this happening much earlier in the call sequence -
mv88e6xxx_port_vlan_dump() takes a mutex, and if that mutex were already
held, we'd schedule at that point.  The RCU read lock was taken by
rtnl_bridge_getlink().

It looks horrible to fix - mvmdio.c as well as DSA locking are involved.

I would say this needs fixing higher up, in the bridge code. DSA has
to be able to sleep, since the switch can be on any arbitrary bus,
MDIO, SPI, etc. This will affect pure switchdev devices as well, since
they often need to send a request to the switch and wait for a reply.

It looks similar to when a switchdev object/attribute is added/deleted
without the SWITCHDEV_F_DEFER flag, used in the bridge code to defer
switchdev operations until switchdev_deferred_process() is called.

This is usually used to process switchdev ops outside the bridge lock.

Jiri, can switchdev_port_vlan_fill not using SWITCHDEV_F_DEFER be the
reason for this suspicious RCU usage when issuing "bridge vlan show"?

If it is called from atomic context, it should be deferred.

Thanks,

       Vivien