Was seeing spurious SSL connection aborts using libcurl and
OpenSSL. I tracked it down to uncleared error state on the
OpenSSL error stack - patch attached deals with that.
Rough idea of problem:
Code that uses libcurl calls some library that uses OpenSSL but
don't clear the OpenSSL error stack after an error.
ssluse.c calls SSL_read which eventually gets an EWOULDBLOCK from
the OS. Returns -1 to indicate an error
ssluse.c calls SSL_get_error. First thing, SSL_get_error calls
ERR_get_error to check the OpenSSL error stack, finds an old
error and returns SSL_ERROR_SSL instead of SSL_ERROR_WANT_READ or
SSL_ERROR_WANT_WRITE.
ssluse.c returns an error and aborts the connection
Solution:
Clear the openssl error stack before calling SSL_* operation if
we're going to call SSL_get_error afterwards.
Notes:
This is much more likely to happen with multi because it's easier
to intersperse other calls to the OpenSSL library in the same
thread.
Enable OpenLDAP support for cygwin builds. This support was disabled back
in 2008 due to incompatibilities between OpenSSL and OpenLDAP headers.
cygwin's OpenSSL 0.9.8l and OpenLDAP 2.3.43 versions on cygwin 1.5.25
allow building an OpenLDAP enabled libcurl supporting back to Windows 95.
Remove non-functional CURL_LDAP_HYBRID code and references.
Jason McDonald posted bug report #3006786 when he found that the
SFTP code didn't timeout properly in several places in the code
even if a timeout was set properly.
Based on his suggested patch, I wrote a different implementation
that I think addressed the issue better and also uses the connect
timeout for the initial part of the SSH/SFTP done during the
"protocol connect" phase.
(http://curl.haxx.se/bug/view.cgi?id=3006786)
Igor Novoseltsev reported a problem with the multi socket API and
using timeouts and timers. It boiled down to a problem with
libcurl's use of GetTickCount() interally to figure out the
current time, while Igor's own application code used another
function call.
It made his app call the socket API timeout function a bit
_before_ libcurl would consider the timeout to trigger, and that
could easily lead to timeouts or stalls in the app. It seems
GetTickCount() in general often has no better resolution than
16ms and switching to the alternative function
QueryPerformanceCounter has its share of problems:
http://www.virtualdub.org/blog/pivot/entry.php?id=106
We address this problem by simply having libcurl treat timers
that already has occured or will occur within 40ms subject for
treatment. I'm confident that there are other implementations and
operating systems with similarly in accurate timer functions so
it makes sense to have applied generically and I don't believe we
sacrifice much by adding a 40ms inaccuracy on these timeouts.
makes the LDAP code much cleaner, nicer and in general being a
better libcurl citizen. If a new enough OpenLDAP version is
detect, the new and shiny lib/openldap.c code is then used
instead of the old cruft
Code by Howard, minor cleanups by Daniel.
Eric Mertens posted bug #3003705: when we made TFTP use the
correct timeout option when sent to the server (fixed May 18th
2010) it became obvious that libcurl used invalid timeout values
(300 by default while the RFC allows nothing above 255). While of
course it is obvious that as TFTP has worked thus far without
being able to set timeout at all, just removing the setting
wouldn't make any difference in behavior. I decided to still keep
it (but fix the problem) as it now actually allows for easier
(future) customization of the timeout.
(http://curl.haxx.se/bug/view.cgi?id=3003705)
In a normal expression, doing [unsigned short] + 1 will not wrap
at 16 bits so the comparisons and outputs were done wrong. I
added a macro do make sure it gets done right.
Douglas Kilpatrick filed bug report #3004787 about it:
http://curl.haxx.se/bug/view.cgi?id=3004787
By undefing a bunch of E* defines that VC10 has started to define
but that we redefine internally to their WSA* alternatives when
building for Windows.
Eric Mertens posted bug report #3003005 pointing out that the
libcurl TFTP code was not sending the timeout option properly to
the server, and suggested a fix.
(http://curl.haxx.se/bug/view.cgi?id=3003005)
John-Mark Bell filed bug #3000052 that identified a problem (with
an associated patch) with the OpenSSL handshake state machine
when the multi interface is used:
Performing an https request using a curl multi handle and using
select or epoll to wait for events results in a hang. It appears
that the cause is the fix for bug #2958179, which makes
ossl_connect_common unconditionally return from the step 2 loop
when fetching from a multi handle.
When ossl_connect_step2 has completed, it updates
connssl->connecting_state to ssl_connect_3. ossl_connect_common
will then return to the caller, as a multi handle is in
use. Eventually, the client code will call curl_multi_fdset to
obtain an updated fdset to select or epoll on. For https
requests, curl_multi_fdset will cause https_getsock to be called.
https_getsock will only return a socket handle if the
connecting_state is ssl_connect_2_reading or
ssl_connect_2_writing. Therefore, the client will never obtain a
valid fdset, and thus not drive the multi handle, resulting in a
hang.
(http://curl.haxx.se/bug/view.cgi?id=3000052)
Sebastian V reported bug #3000056 identifying a problem with
redirect following. It showed that when curl followed redirects
it didn't properly ignore the response body of the 30X response
if that response was using compressed Content-Encoding!
(http://curl.haxx.se/bug/view.cgi?id=3000056)
Dirk Manske reported a regression. When connecting with the multi
interface, there were situations where libcurl wouldn't store
connect time correctly as it used to (and is documented to) do.
Using his fine sample program we could repeat it, and I wrote up
test case 573 using that code. The problem does not easily show
itself using the local test suite though.
The fix, also as suggested by Dirk, is a bit on the ugly side as
it adds yet another call to Curl_verboseconnect() and setting the
TIMER_CONNECT time. That situation is subject for some closer
inspection in the future.
Prefixing the FTP quote commands with an asterisk really only
worked for the postquote actions. This is now fixed and test case
227 has been extended to verify.
Matt Wixson found and fixed a bug in the SCP/SFTP area where the
code treated a 0 return code from libssh2 to be the same as
EAGAIN while in reality it isn't. The problem caused a hang in
SFTP transfers from a MessageWay server.
Ben Greear brought a patch that from now on allows all protocols
to specify name and user within the URL, in the same manner HTTP
and FTP have been allowed to in the past - although far from all
of the libcurl supported protocols actually have that feature in
their URL definition spec.
Bob Richmond: There's an annoying situation where libcurl will
read new HTTP response data from a socket, then check if it's a
timeout if one is set. If the last packet received constitutes
the end of the response body, libcurl still treats it as a
timeout condition and reports a message like:
"Operation timed out after 3000 milliseconds with 876 out of 876
bytes received"
It should only a timeout if the timer lapsed and we DIDN'T
receive the end of the response body yet.
Christopher Conroy fixed a problem with RTSP and GET_PARAMETER
reported to us by Massimo Callegari. There's a new test case 572
that verifies this now.
