Peter, Holger, thanks for your reports. At the moment I have absolutely no idea about what can happen here.
If you could give 1.8.5 a try, as a few bugs were fixed after the commit referenced in this thread, that would be great. The commit related to the bogus buffer wrapping in bi_putblk() could possibly strike on H2 when receiving POSTs. The likeliness is low but it's worth being certain we're not chasing a fantom! Holger, if you could send me your core and unstripped haproxy executable in private, that would be really cool. Just be aware that it can reveal some sensitive data (client addresses etc) so I can understand if you don't want to. I'll go back recheck the H2 error path because if it takes two hours before it dies, it's most likely there. Thanks, Willy
