Re: [PATCH v8 net-next 0/2] mv88e6xxx: Add MAB offload support

[netdev]

On 2022-11-15 15:56, Vladimir Oltean wrote:
> On Tue, Nov 15, 2022 at 02:25:13PM +0100, netdev@xxxxxxxxxxxxxxxxxxxx 
> wrote:
> > On 2022-11-15 13:22, Vladimir Oltean wrote:
> > > Do you have a timeline for when the regression was introduced?
> > > Commit 35da1dfd9484 reverts cleanly, so I suppose giving it a go with
> > > that reverted might be worth a shot. Otherwise, a bisect from a known
> > > working version only takes a couple of hours, and shouldn't require
> > > other changes to the setup.
> >
> > Wow! Reverting 35da1dfd9484 and the problem has disappeared. :-)
>
> See? That wasn't very painful, was it.

Indeed it was not, when you get a good tip. Thanks alot! :-)

> Now, why doesn't that commit work for you? that's the real question.
> I'm going to say there's a big assumption made there. The old code used
> to poll up to 16 times with sleeps of up to 2 ms in between.
> The new code polls until at least 50 ms have elapsed.
> I can imagine the thought process being something like "hmm, 16*2=32ms,
> let's round that up to 50 just to be sure". But the effective timeout
> was not really increased. Rather said, in the old code there was never
> really an effective timeout, since the polling code could have been
> preempted many times, and these preemptions would not be accounted
> against the msleep(2) calls. Whereas the new code really tracks
> something approximating 50 ms now.
>
> Could you please add the reverted patch back to your git tree, and see
> by how much do you need to increase the timeout for your system to get
> reliable results?

Yes, so you want me to simply increase the 50ms on line 58 in smi.c...

I have now tried to increase it even to 10000ms == 10s and it didn't 
help,
so something else must be needed...