HTTP 1.1 connection leaks

[bitly]

HTTP 1.1 connection leaks

[ploxiln]

excellent minimal fix (if it works)

[jayridge]

danger, danger!!! this is a good opt for low volume or simpleq mget
clients. however, it will increase latency, cpu and sockets in timed wait (
typically 2 * the MSL ).

[ploxiln]

ok the situation is becoming clearer: the bug with libevent-1.4 is that if the connection is keep-alive and the client closes it, it doesn't close the fd and clean up the resources related to the connection. (I was under the impression that HTTP/1.1 / keep-alive was generally buggy and we didn't really use it)

[ploxiln]

references for libevent-1.4 keep-alive fix strategies:
http://www.mail-archive.com/[email protected]/msg01721.html
http://www.mail-archive.com/[email protected]/msg01722.html

[ploxiln]

btw, my latest fix, in libevent-1.4.14b:

diff -urp libevent-1.4.14b-stable/http.c libevent-1.4.14b-stable-fix/http.c
--- libevent-1.4.14b-stable/http.c  2010-06-20 09:06:04.000000000 -0400
+++ libevent-1.4.14b-stable-fix/http.c  2012-05-09 14:51:37.000000000 -0400
@@ -2231,6 +2231,7 @@ evhttp_handle_request(struct evhttp_requ
    if (req->uri == NULL) {
        event_debug(("%s: bad request", __func__));
        if (req->evcon->state == EVCON_DISCONNECTED) {
+           req->userdone = 1;
            evhttp_connection_fail(req->evcon, EVCON_HTTP_EOF);
        } else {
            event_debug(("%s: sending error", __func__));

[mreiferson]

Closing (we're attempting to push libevent changes upstream)