On Fri, Oct 30, 2020 at 4:02 PM Xie He [off-list ref] wrote:

On Fri, Oct 30, 2020 at 9:35 AM Willem de Bruijn
[off-list ref] wrote:

quoted

In general we try to avoid changing counter behavior like that, as
existing users
may depend on current behavior, e.g., in dashboards or automated monitoring.

I don't know how realistic that is in this specific case, no strong
objections. Use
good judgment.

Originally this function only increases stats.rx_dropped only when
there's a memory squeeze. I don't know the specification for the
meaning of stats.rx_dropped, but as I understand it indicates a frame
is dropped. This is why I wanted to increase it whenever we drop a
frame.

Jakub recently made stats behavior less ambiguous, in commit
0db0c34cfbc9 ("net: tighten the definition of interface statistics").

That said, it's not entirely clear whether rx_dropped would be allowed
to include rx_errors.

My hunch is that it shouldn't. A quick scan of devices did quickly
show at least one example where it does: macvlan. But I expect that to
be an outlier.

Originally this function drops a frame silently if the PVC virtual
device that corresponds to the DLCI number and the protocol type
doesn't exist. I think we may at least need some way to note this.
Originally this function drops a frame with a kernel info message
printed if the protocol type is not supported. I think this is a bad
way because if the other end continuously sends us a lot of frames
with unsupported protocol types, our kernel message log will be
overwhelmed.

I don't know how important it is to keep backwards compatibility. I
usually don't consider this too much. But I can drop this change if we
really want to keep the counter behavior unchanged. I think changing
it is better if we don't consider backwards compatibility.

Please do always consider backward compatibility. In this case, I
don't think that the behavioral change is needed for the core of the
patch (changing control flow).