John Clayton's weird explorations in the wonders of the windows tcp/ip

stack ;-)

diff --git a/docs/KNOWN_BUGS b/docs/KNOWN_BUGS
index 3e054f1bce94e2e33443ae015daec5ae18533ff6..31e39fe3f1debfffad330a38bbdb1e3c70d395d5 100644 (file)
--- a/docs/KNOWN_BUGS
+++ b/docs/KNOWN_BUGS
@@ -14,3 +14,63 @@ may have been fixed since this was written!
 * libcurl does not deal nicely with files larger than 2GB
 
 * GOPHER transfers seem broken
+
+------------------------------------------------------------------------------
+
+Q: My program blows up when I run lots of curl_easy_perform() calls on a
+single thread
+Q: My program dies when a single thread re-enters the win32 select() call
+via curl_easy_perform()
+Q: --- add your own flavour here ---
+
+Single Threaded Re-Entracy
+--------------------------
+
+There is a glitch / trick to using cURL on Win32 related to re-entrancy.
+This experience was gained on verion 7.9.4 using Windows NT SP3 in a banking
+environment (just in case you wanted to know).
+
+If you have already called curl_easy_perform(), and *somehow* you cause your
+single thread of execution to make another call to curl_easy_perform() - the
+windows socket() call used to create a new socket for the second connection
+can return with 10044 / 10043 error codes.
+
+The WSA errors we experienced are:
+WSAEPROTONOSUPPORT 
+(10043) 
+Protocol not supported. 
+The requested protocol has not been configured into the system, or no
+implementation for it exists. For example, a socket call requests a
+SOCK_DGRAM socket, but specifies a stream protocol. 
+
+WSAESOCKTNOSUPPORT 
+(10044) 
+Socket type not supported. 
+The support for the specified socket type does not exist in this address
+family. For example, the optional type SOCK_RAW might be selected in a
+socket call, and the implementation does not support SOCK_RAW sockets at
+all. 
+
+We have experienced this by creating a timer that ticks every 20ms, and on
+the tick making a curl_easy_perform() call.  The call usually completed in
+about 300ms.  And we expected (before this test) that the timer would NOT be
+fired during a call to curl_easy_perform(), howvever, while the first
+curl_easy_perform() is running a tick *is* fired by the windows API somehow,
+and we then call curl_easy_perform() again - thus single threaded
+re-entrancy is achieved.
+
+Notes:
+* We made sure that a new CURL structure was being used for each
+curl_easy_perform() request, and that the curl_global_init() had been called
+beforehand.  
+* I'm happy to answer any questions about this problem to try to track it
+down.
+* Once the socket() call started failing, there is no hope - it never works
+again.
+* Slowing the timer down to give each request enough time to complete solves
+this problem completely.
+
+If anyone has the source code to the WinNT implementation of socket() and
+can figure out WHY this can occur, more tracing can be performed.
+
+        John Clayton <John.Clayton at barclayscapital.com>