> Am 19.05.2025 um 16:46 schrieb Yann Ylavic <ylavic....@gmail.com>:
>
> On Mon, May 19, 2025 at 4:05 PM Ruediger Pluem <rpl...@apache.org> wrote:
>>
>>
>>
>> On 5/19/25 1:53 PM, jean-frederic clere wrote:
>>> Hi,
>>>
>>> While investigating a customer case I note a regression with forward proxy.
>>> While testing I have:
>>> 2.4.57 finished in 3.05s, 3278.62 req/s, 743.70KB/s
>>> 2.4.x finished in 15.38s, 650.17 req/s, 147.88KB/s
>>>
>>> The regression goes way with:
>>> +++
>>> jfclere@fedora:~/httpd-2.4.x$ svn diff .
>>> Index: modules/proxy/proxy_util.c
>>> ===================================================================
>>> --- modules/proxy/proxy_util.c (revision 1925472)
>>> +++ modules/proxy/proxy_util.c (working copy)
>>> @@ -1660,7 +1660,6 @@
>>> conn = connection_make(p, worker);
>>> }
>>> else if (conn->close
>>> - || conn->forward
>>> || (conn->connection
>>> && conn->connection->keepalive == AP_CONN_CLOSE)
>>> || worker->s->disablereuse) {
>>> +++
>>> But that doesn't look quite right. any hints?
>>
>> Agreed that is does not look right. What configuration do you use for the
>> test?
>> What is the test scenario?
>
> Hm, r1912459 added this test to disable reuse for CONNECT connections
> to ProxyRemote, because I thought that CONNECT requests were not
> reusable. However this kinds of defeats the code in
> ap_proxy_determine_connection() allowing to reuse a ProxyRemote
> connection [1] if its forward_info did not change (same target
> host/port and same auth), and I think that's what your use case was
> reaching before r1912459.
> It's probably safe for successive requests to the same host/port/auth
> to be pipelined in the same CONNECT payload, saving new connections
> and TLS handshakes, so possibly Jean-Frederic's patch is correct
> (though not complete, we need to close the connection if the
> forward_info does not match in [1])?
>
> [1]
> https://github.com/apache/httpd/blob/2.4.63/modules/proxy/proxy_util.c#L3372
Hmm, what kind of CONNECT connections are reusable? The common use case is a
TLS tunnel, I think. It would not work to reuse that.
- Stefan
