On 5/18/15, 8:48 PM, David Miller wrote:

From: John Fastabend <redacted>
Date: Mon, 18 May 2015 17:21:29 -0700

quoted

So how about having an error strategy sysctl field that we can set
at provisioning time. I think this would align to Roopa's option (b).
This way we can default to "transparent" mode and the users where
this wont work can set the error mode. This way user land software
stacks that work today should continue to work in both modes.

Alert: This is not a switch provisioning issue.

You can frame it like that all day, and continue to talk about
low power cpus or other things which are completely and utterly
irrelevant.

Stop looking at how some specific piece of hardware is configured,
and think about what actually is asking the kernel to do stuff.

That's because the real issue is _semantics_ and what a Linux machine
is expected to do when you insert a route and valid reasons why a
route insertion can fail.

That is the _only_ issue.

And that has to do with what semantics _applcations_ making these
routing change requests expect.

There is nothing else that matters.

And since it is an issue of what semantics those application want and
are able to handle, that is where the request of changed behavior
belongs.

If we added your suggested sysctl, we'd have to name it
"sysctl_break_all_my_apps_please" because that is exactly what it
would be doing. :-)

understood. This seems to lean towards option c) where app explicitly 
requests offload with RTNH_F_OFFLOAD for every route.

from where I see, with the limitations on these boxes,
this requires every app, every `ip route` cmd running on the box to 
explicitly specify offload when running on this hardware. In which case 
having a way to specify a global system policy seemed appropriate. Hence 
the sysctl suggestion.