Shmulik Regev fixed a flaw in the multi interface that occurred when doing

HTTP CONNECT over a proxy

diff --git a/CHANGES b/CHANGES
index 570a75d8abe8c18748a503271cbcaa15f39155ca..9759d6512552d99849d8cc01c568b2437a6eb97b 100644 (file)
--- a/CHANGES
+++ b/CHANGES
@@ -7,6 +7,32 @@
                                   Changelog
 
 Daniel S (14 June 2007)
+- Shmulik Regev:
+
+  I've encountered (and hopefully fixed) a problem involving proxy CONNECT
+  requests and easy handles state management. The problem isn't simple to
+  reproduce since it depends on socket state. It only manifests itself when
+  working with non-blocking sockets.
+
+  Here is the scenario:
+
+  1. in multi_runsingle the easy handle is in the CURLM_STATE_WAITCONNECT and
+  calls Curl_protocol_connect
+
+  2. in Curl_proxyCONNECT, line 1247, if the socket isn't ready the function
+  returns and conn->bits.tunnel_connecting is TRUE
+
+  3. when the call to Curl_protocol_connect returns the protocol_connect flag
+  is false and the easy state is changed to CURLM_STATE_PROTOCONNECT which
+  isn't correct if a proxy is used.  Rather CURLM_STATE_WAITPROXYCONNECT
+  should be used.
+
+  I discovered this while performing an HTTPS request through a proxy (squid)
+  on my local network. The problem caused openssl to fail as it read the proxy
+  response to the CONNECT call ('HTTP/1.0 Established') rather than the SSL
+  handshake (the exact openssl error was 'wrong ssl version' but this isn't
+  very important)
+
 - Dave Vasilevsky filed bug report #1736875
   (http://curl.haxx.se/bug/view.cgi?id=1736875) almost simultanouesly as Dan
   Fandrich mentioned a related build problem on the libcurl mailing list:

diff --git a/RELEASE-NOTES b/RELEASE-NOTES
index 38e5272dbb106b61a8c4cae14b30e6b697be49aa..2039645ad57a109ccc8155da8b08042f0f84d740 100644 (file)
--- a/RELEASE-NOTES
+++ b/RELEASE-NOTES
@@ -57,6 +57,7 @@ This release includes the following bugfixes:
  o out-of-boundary write in Curl_select()
  o -s/--silent can now be used to toggle off the silence again
  o builds fine on 64bit HP-UX
+ o multi interface HTTP CONNECT glitch
 
 This release includes the following known bugs:
 
@@ -83,6 +84,6 @@ advice from friends like these:
  Frank Hempel, Michael Wallner, Jeff Pohlmeyer, Tobias Rundström,
  Anders Gustafsson, James Bursa, Kristian Gunstone, Feng Tu,
  Andre Guibert de Bruet, Rob Crittenden, Rich Rauenzahn, Tom Regner,
- Dave Vasilevsky
+ Dave Vasilevsky, Shmulik Regev
 
         Thanks! (and sorry if I forgot to mention someone)

diff --git a/lib/multi.c b/lib/multi.c
index bbcf6319e788bd04b74c651c562035fad075698b..5e91a5e7cefc9a4b04b0d20cbb0938e1db7e017c 100644 (file)
--- a/lib/multi.c
+++ b/lib/multi.c
@@ -1007,8 +1007,15 @@ static CURLMcode multi_runsingle(struct Curl_multi *multi,
         if(!protocol_connect) {
           /* We have a TCP connection, but 'protocol_connect' may be false
              and then we continue to 'STATE_PROTOCONNECT'. If protocol
-             connect is TRUE, we move on to STATE_DO. */
-          multistate(easy, CURLM_STATE_PROTOCONNECT);
+             connect is TRUE, we move on to STATE_DO.
+             BUT if we are using a proxy we must change to WAITPROXYCONNECT
+             */
+#ifndef CURL_DISABLE_HTTP
+          if (easy->easy_conn->bits.tunnel_connecting)
+            multistate(easy, CURLM_STATE_WAITPROXYCONNECT);
+          else
+#endif
+            multistate(easy, CURLM_STATE_PROTOCONNECT);
         }
         else {
           /* after the connect has completed, go WAITDO */