Hello,
On Wed, Jan 3, 2018 at 9:51 PM, Willy Tarreau <w...@1wt.eu> wrote: > On Wed, Jan 03, 2018 at 09:31:47PM +0100, Willy Tarreau wrote: >> Oh I think you've just put your finger on it. I remember taking care >> of handling 0-sized frames, and facing certain difficulties with them >> (eg: sometimes returning size 0 just means nothing was done). I sounds >> very likely that we can still have a bug around this. It would also >> explain why your patch could get rid of it. >> >> I'll have a look at the code in case I have an idea. > > Could you please try the attached patch? I'm pretty sure it is *very* > related to your observations. In fact till now we would not update the > parser's state on an empty data frame, which explains why you had to > move the stuff around. Yes, I can confirm this fixes the problem for me, and in a good looking way too: all streams are correctly handled within that single connection (no sudden connection teardown and subsequent new connection for follow-up streams). Now we won't even bother the ML subscribers with a 50+ messages mega-thread debugging an obscure H2 issue :) regards, lukas