Hi Cyril, On Sun, Sep 07, 2014 at 06:30:40PM +0200, Cyril Bonté wrote: > Hi all, > > Le 06/09/2014 08:28, Willy Tarreau a écrit : > >To be honnest, I think it is necessary, and I hope it will be enough. We > >had this issue in 1.5-dev as well a long time ago and the problem was the > >same, the beginning of the message was not moved. Now 1.5's buffer > >semantics > >have changed to make all this much simpler. There is one thing to test to > >confirm it's enough : make a farm with two untested servers, one which does > >not respond and a valid one so that there's a retry and redispatch. Haproxy > >will first pick server1, prepend the header, and try to send the request > >there. After the failure it will redispatch to server2 and rewind the > >buffer to change the header. On server2 you should see a valid request if > >the fix is OK. That's how I managed to fix it in 1.5-dev. The config below > >should be enough to reproduce it : > > > > listen test :10002 > > balance roundrobin > > mode http > > option http-server-close > > http-send-name-header Host > > retries 1 > > option redispatch > > server bad 127.0.0.1:12345 > > server good 127.0.0.1:80 > > > >I also remember that rewinding POST requests is harder than rewinding > >GET because the body was already partially parsed. I remember something > >in 1.4 which we did to take care of that, either we refrain from parsing > >the body before connecting when there's the send-name-header, or we could > >correctly rewind. So sending a POST request through the config above will > >tell you if that's OK :-) > > I was optimistic during my tests unless I tried sending POST requests > with a body longer than the buffer size. In certain conditions, haproxy > segfaults. For example, this is the case with the "bad" server, no need > to redispatch. > > I need some time to remember how buffers work in 1.4, in order to fix > the segfault.
Argh, crap! I fear that we're opening a new can of worms :-( At least we have working code in 1.5 so we know there's a fix, but the two versions are so much different that we cannot port 1.5 changes back to 1.4. It's possible however that we'll have to add a WAIT_CONN flag to prevent the body from being parsed/forwarded until the connection is opened. Willy
