Re: [PATCH net-next v4 06/10] netdev: add netdevice notifier type to trigger a reprogramming of offloads

On 10.01.2016 18:16, Tom Herbert wrote:

On Sat, Jan 9, 2016 at 2:52 PM, Hannes Frederic Sowa <
hannes@stressinduktion.org> wrote:

quoted

On 09.01.2016 23:27, Tom Herbert wrote:

quoted

On Sat, Jan 9, 2016 at 9:30 AM, Hannes Frederic Sowa
[off-list ref] wrote:

quoted

On 09.01.2016 18:25, Tom Herbert wrote:

quoted

On Sat, Jan 9, 2016 at 7:07 AM, Hannes Frederic Sowa
[off-list ref] wrote:

quoted

Signed-off-by: Hannes Frederic Sowa <redacted>
---
include/linux/netdevice.h | 1 +
1 file changed, 1 insertion(+)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 8d8e5ca951b493..9750e46760695d 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h

@@ -2183,6 +2183,7 @@ struct netdev_lag_lower_state_info {

#define NETDEV_BONDING_INFO 0x0019
#define NETDEV_PRECHANGEUPPER 0x001A
#define NETDEV_CHANGELOWERSTATE 0x001B
+#define NETDEV_REFRESH_OFFLOADS 0x001C

Per previous discussion we don't want to generalize this current
offload interface. Can we just NETDEV_UP as the notifier?

The problem with only using NETDEV_UP/REGISTER is that some drivers need
to
reconfigure their offloads during operation while keeping the netdevice
in
UP state. One example is ixgbe. This was my first idea also.

It's configuration. Just get it once and save the port number(s) in
the driver. Configure the device only when
IXGBE_FLAG_VXLAN_OFFLOAD_CAPABLE is set.

I don't see much value that each driver has to hold a database of ports to
offload, the kernel knows much better and can easily be queried via this
mechanism.

This is not only about ixgbe, but also about benet, which f.e. needs the
list of offloads available again during resume. Of course, drivers can do
their own stateful handling of ports, but this seems to much more generate
problems. I rather would like to see more logic moved into the core and

less

quoted

code in the drivers.

All the drivers that support VXLAN already save the ports in a stateful
fashion. Some support a single port, some and array, some a list. The only
question is whether they bother to save the information when the offload
feature is disabled. ixgbe does not for instance, but it looks like fm10k
does since it doesn't check any feature flags. Fixing those that don't save
the information all the times should be straightforward.

Yes, would probably be straight forward to add another list plus extra 
reference counting and maybe a lock, but I don't see why we need to 
duplicate the policy in the drivers? I personally think ixgbe does ok in 
this regard.

Anyway, this series merely tries to decouple 10gigbit cards from vxlan 
and/or geneve. It will be no problem to deal with this later on.

Bye,
Hannes