When TE: is inserted in the request, we must add a "Connection: TE" as
well to be HTTP 1.1 compliant. If a custom Connection: header is passed
in, we must use that and only append TE to it. Test case 1125 verifies
TE: + custom Connection:.
Since this struct member is used in the code to determine what and how
to decode automatically and since it is now also used for compressed
Transfer-Encodings, I renamed it to the more suitable 'auto_decoding'
Transfer-Encoding differs from Content-Encoding in a few subtle ways,
but primarily it concerns the transfer only and not the content so when
discovered to be compressed we know we have to uncompress it. There will
only arrive compressed transfers in a response after we have requested
them with the appropriate TE: header.
Test case 1122 and 1123 verify.
When checking if an existing RTSP connection is alive or not, the
checkconnection function might be called with a SessionHandle pointer
being NULL and then referenced causing a crash. This happened only using
the multi interface.
Reported by: Tinus van den Berg
Bug: http://curl.haxx.se/bug/view.cgi?id=3280739
In case a client certificate is used, invalidate SSL session cache
at the end of a session. This forces NSS to ask for a new client
certificate when connecting second time to the same host.
Bug: https://bugzilla.redhat.com/689031
* Rename the object object directory from 'objs' to 'BCC_obj'. I feel
it should be named properly. Ref. Makefile.Watcom where it's called
'WC_Win32.obj'.
* Turn off these warnings to keep the build totally silent (with CBuilder-6
that is).
-w-inl 8026 Functions X are not expanded inline.
-w-pia 8060 Possibly incorrect assignment
-w-pin 8061 Initialization is only partially bracketed
I'm sure the warnings could be fixed the "proper" way or with some added
"#pragma" statements. But that just clutters the sources IMHO.
* $(MKDIR) and $(RMDIR) have been replaced with the shell-commands 'md'
and 'rd'. When having MingW/Msys programs 'mkdir.exe' and 'rmdir.exe' in
$PATH, this confuses Borland's make and the result (the cleaning etc.) would
not be as expected.
* Added a ".path.int = $(OBJDIR)" to tell make where the $(PREPROCESSED)
files are. Why we need the preprocess step in the fist place is beyond me
(Yang?). But I'll leave that for now.
Stop the abuse of CURLE_FAILED_INIT as return code for things not being
init related by introducing two new return codes:
CURLE_NOT_BUILT_IN and CURLE_UNKNOWN_OPTION
CURLE_NOT_BUILT_IN replaces return code 4 that has been obsoleted for
several years. It is used for returning error when something is
attempted to be used but the feature/option was not enabled or
explictitly disabled at build-time. Getting this error mostly means that
libcurl needs to be rebuilt.
CURLE_FAILED_INIT is now saved and used strictly for init
failures. Getting this problem means something went seriously wrong,
like a resource shortage or similar.
CURLE_UNKNOWN_OPTION is the option formerly known as
CURLE_UNKNOWN_TELNET_OPTION (and the old name is still present,
separately defined to be removed in a very distant future). This error
code is meant to be used to return when an option is given to libcurl
that isn't known. This problem would mostly indicate a problem in the
program that uses libcurl.
In my attempts to reduce #ifdefs in code, the SOCKS functions are now
macros when libcurl is built without proxy support and therefore the FTP
code could avoid some #ifs.
The new http_proxy.* files now host HTTP proxy specific code (500+ lines
moved out from http.c), and as a consequence there is a macro introduced
for the Curl_proxyCONNECT() function so that code can use it without
actually supporting proxy (or HTTP) in builds.
1 - make sure to #define macros for cookie functions in the cookie
header when cookies are disabled to avoid having to use #ifdefs in code
using those functions.
2 - move cookie-specific code to cookie.c and use the functio
conditionally as mentioned in (1).
net result: 6 #if lines removed, and 9 lines of code less
Within multi_socket when conn is used as a shorthand, data could be
changed and multi_runsingle could modify the connectdata struct to deal
with. This bug has not been included in a public release.
Using 'conn' like that turned out to be ugly. This change is a partial
revert of commit f1c6cd42f4.
Reported by: Miroslav Spousta
Bug: http://curl.haxx.se/bug/view.cgi?id=3265485
When asked to bind the local end of a connection when doing a request,
the code will now disqualify other existing connections from re-use even
if they are connected to the correct remote host.
This will also affect which connections that can be used for pipelining,
so that only connections that aren't bound or bound to the same
device/port you're asking for will be considered.
The RTSP-specific function for checking for "dead" connection is better
located in rtsp.c. The code using this is now written without #ifdefs as
the function call is instead turned into a macro (in rtsp.h) when RTSP
is disabled.
Move ipv6-functional-probe into a single function that is used from all
places that need to know.
Make the probe function store the result in a static variable so that
subsequent invokes just returns the previous result and won't have to
probe again.
Curl_posttransfer is called too soon to add the final new line.
Moved the new line logic to pgrsDone as there is no more call to
update the progress status after this call.
Reported by: Dmitri Shubin <sbn_at_tbricks.com>
http://curl.haxx.se/mail/lib-2010-12/0162.html
When libcurl sends a HTTP request on a re-used connection and detects it
being closed (ie no data at all was read from it), it is important to
rewind if any data in the request was sent using the read callback or
was read from file, as otherwise the retried request will be broken.
Reported by: Chris Smowton
Bug: http://curl.haxx.se/bug/view.cgi?id=3195205
When NSS-powered libcurl connected to a SSL server with
CURLOPT_SSL_VERIFYPEER equal to zero, NSS remembered that the peer
certificate was accepted by libcurl and did not ask the second time when
connecting to the same server with CURLOPT_SSL_VERIFYPEER equal to one.
This patch turns off the SSL session cache for the particular SSL socket
if peer verification is disabled. In order to avoid any performance
impact, the peer verification is completely skipped in that case, which
makes it even faster than before.
Bug: https://bugzilla.redhat.com/678580
The PROT_* set of internal defines for the protocols is no longer
used. We now use the same bits internally as we have defined in the
public header using the CURLPROTO_ prefix. This is for simplicity and
because the PROT_* prefix was already used duplicated internally for a
set of KRB4 values.
The PROTOPT_* defines were moved up to just below the struct definition
within which they are used.
The protocol handler struct got a 'flags' field for special information
and characteristics of the given protocol.
This now enables us to move away central protocol information such as
CLOSEACTION and DUALCHANNEL from single defines in a central place, out
to each protocol's definition. It also made us stop abusing the protocol
field for other info than the protocol, and we could start cleaning up
other protocol-specific things by adding flags bits to set in the
handler struct.
The "protocol" field connectdata struct was removed as well and the code
now refers directly to the conn->handler->protocol field instead. To
make things work properly, the code now always store a conn->given
pointer that points out the original handler struct so that the code can
learn details from the original protocol even if conn->handler is
modified along the way - for example when switching to go over a HTTP
proxy.
The non-blocking connect improvement for IMAP showed that we didn't
properly define the Curl_ssl_connect_nonblocking function for non-SSL
builds.
Reported by: Tor Arntsen
Only download and convert the certdata to the ca-bundle.crt if Mozilla
changed the data
The Perl LWP module (which in a bit of a circular reference is used by
mk-ca-bundle.pl) is now indirectly using this script. I made this small
tweak to make it easier to automatically maintain the generated
ca-bundle.crt file in version control.
Some protocols have to call the underlying functions without regard to
what exact state the socket signals. For example even if the socket says
"readable", the send function might need to be called while uploading,
or vice versa. This is the case for libssh2 based protocols: SCP and
SFTP and we now introduce a define to set those protocols and we make
the multi interface code aware of this concept.
This is another fix to make test 582 run properly.
As a new state recently was added to the IMAP state machine it has to be
in the array of names as well as otherwise libcurl crashes when a debug
version runs...
For uploads we want to use the _sending_ function even when the socket
turns out readable as the underlying libssh2 sftp send function will
deal with both accordingly. This is what the cselect_bits magic is for.
Fixes test 582.
Make GSS authentication work when a curl handle is reused for multiple
authenticated requests, by always setting negdata->state in
output_auth_headers().
Signed-off-by: Marcus Sundberg <marcus.sundberg@aptilo.com>
When using the multi interface and a handle using SFTP was removed very
early on, we would get a segfault due to the code assumed data was there
that hadn't yet been setup.
Bug: http://curl.haxx.se/mail/lib-2011-03/0066.html
Reported by: Saqib Ali
Both SFTP and SCP are protocols that need to shut down stuff properly
when the connection is about to get torned down. The primary effect of
not doing this shows up as memory leaks (when using SCP or SFTP with the
multi interface).
This is one of the problems detected by test 582.
As we know how much to send, we can and should stop once we've sent that
much data as it avoids having to rely on other mechanisms to detect the
end.
This is one of the problems detected by test 582.
Reported by: Henry Ludemann <misc@hl.id.au>
When using the multi_socket API to do SFTP upload, it is important that
we set a quick expire when leaving the SSH_SFTP_UPLOAD_INIT state as
there's nothing happening on the socket so there's no read or write to
wait for, but the next libssh2 API function needs to be called to get
the ball rolling.
This is one of the problems detected by test 582.
Reported by: Henry Ludemann <misc@hl.id.au>
All C and H files now (should) feature the proper project curl source
code header, which includes basic info, a copyright statement and some
basic disclaimers.
CyaSSL (available from git@github.com:cyassl/cyassl.git) has been
added to the SSL abstraction layer.
To test:
1) git CyaSSL sources
2) autoreconf -i
3) ./configure --disable-static
4) make
5) sudo make install
6) autoreconf -i
7) git curl sources (and this patch)
8) ./configure --disable-shared --with-cyassl --without-ssl --enable-debug
9) make
10) normal testing
Please send questions or comments to todd@yassl.com .
libssh2_knownhost_readfile() returns a negative value on error or
otherwise number of parsed known hosts - this was previously not
documented correctly in the libssh2 man page for the function.
Bug: http://curl.haxx.se/mail/lib-2011-02/0327.html
Reported by: murat
Removed the "netrc_debug" keyword replaced with --netrc-file additions.
Removed the debug code from Curl_parsenetrc as it is superseeded by
--netrc-file.
After a request times out, the connection wasn't properly closed and
prevented to get re-used, so subsequent transfers could still mistakenly
get to use the previously aborted connection.
When failing to connect the protocol during the CURLM_STATE_PROTOCONNECT
state, Curl_done() has to be called with the premature flag set TRUE as
for the pingpong protocols this can be important.
When Curl_done() is called with premature == TRUE, it needs to call
Curl_disconnect() with its 'dead_connection' argument set to TRUE as
well so that any protocol handler's disconnect function won't attempt to
use the (control) connection for anything.
This problem caused the pingpong protocols to fail to disconnect when
STARTTLS failed.
Reported by: Alona Rossen
Bug: http://curl.haxx.se/mail/lib-2011-02/0195.html
Introducing a few CURL_SOCKOPT* defines for conveniance. The new
CURL_SOCKOPT_ALREADY_CONNECTED signals to libcurl that the socket is to
be treated as already connected and thus it will skip the connect()
call.
It turns out some systems rely on the gmtime or gmtime_r to be defined
already in the system headers and thus my "precaution" redefining of
them only caused trouble. They are now removed.
On second thought, I think CURLE_TLSAUTH_FAILED should be eliminated. It
was only being raised when an internal error occurred while allocating
or setting the GnuTLS SRP client credentials struct. For TLS
authentication failures, the general CURLE_SSL_CONNECT_ERROR seems
appropriate; its error string already includes "passwords" as a possible
cause. Having a separate TLS auth error code might also cause people to
think that a TLS auth failure means the wrong username or password was
entered, when it could also be a sign of a man-in-the-middle attack.
When the callback returns an error, this function must make sure to return
CURLE_ABORTED_BY_CALLBACK properly and not CURLE_OK as before to allow the
callback to properly abort the operation.
The main has not been updated from some time and is out of sync with
the code. The code is now tested by several test cases so no need for
a seperate code path.
Instead of polluting many places with #ifdefs, we create a single place
for this function, and also check return code properly so that a NULL
pointer returned won't cause problems.
The official Mozilla page at http://www.mozilla.org/projects/security/certs/
points out a new place as the "proper" place to get Mozilla's CA certs from
so this script is now updated to use that instead.
Reported by: Daniel Mentz
The official Mozilla page at
http://www.mozilla.org/projects/security/certs/ points out a new place
as the "proper" place to get Mozilla's CA certs from so this script is
now updated to use that instead.
Reported by: Daniel Mentz
The code in the toofast state needs to first recalculate the values
before it uses them again since it may have been a while since it last
did it when it reaches this point.
This will be used by file_do() and Curl_readwrite() as a unified method
of checking to see if a remote document meets the supplied
CURLOPT_TIMEVAL and CURLOPT_TIMECONDITION.
Signed-off-by: Dave Reisner <d@falconindy.com>
When this callback is called due to the destruction of the ares handle,
the connection pointer passed in as an argument may no longer pointing
to valid data and this function doesn't need to do anything with it
anyway so we make sure it doesn't.
Bug: http://curl.haxx.se/mail/lib-2011-01/0333.html
Reported by: Vsevolod Novikov
The HTTP parser allocated memory on each received Location: header
without properly freeing old data. Starting now, the code only considers
the first Location: header and will blissfully ignore subsequent ones.
Bug: http://curl.haxx.se/bug/view.cgi?id=3165129
Reported by: Martin Lemke
... and update the curl.1 and curl_easy_setopt.3 man pages such that
they do not suggest to use an OpenSSL utility if curl is not built
against OpenSSL.
Bug: https://bugzilla.redhat.com/669702
The idea that the protocol and socktype is part of name resolving in the
libc functions is nuts. We keep the name resolver functions assume
TCP/STREAM and we make sure that when we want to connect to a UDP
service we use the correct UDP/DGRAM set instead. This bug was because
the ->protocol field was not always set correctly.
This bug was only affecting ipv6-disabled non-cares non-threaded builds.
Bug: http://curl.haxx.se/bug/view.cgi?id=3154436
Reported by: "dperham"
When configure --enable-debug has been used, all files in lib/ are now
built twice and a separate static library crafted for unit-testing will
be linked. The unit tests in the tests/unit subdir will use that
library.
Since some systems don't have PATH_MAX and it isn't that clever to
assume a fixed maximum path length, the code now allocates buffer space
instead of using stack.
Reported by: Samuel Thibault
Bug: http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=608521
Sending "pwd" as a QUOTE command only sent the reply to the
DEBUGFUNCTION. Now it also sends an FTP-like header to the header
callback to allow similar operations as with FTP, and apps can re-use
the same parser.
When built IPv6-enabled, we could do Curl_done() with one of the two
resolves having returned already, so when ares_cancel() is called the
resolve callback ends up doing funny things (sometimes resulting in a
segfault) since it would try to actually store the previous resolve even
though we're shutting down the resolve.
This bug was introduced in commit 8ab137b2bc so it hasn't been
included in any public release.
Bug: http://curl.haxx.se/bug/view.cgi?id=3145445
Reported by: Pedro Larroy
Providing multiple dots in a series in the domain field (domain=..com) could
trick the cookie engine to wrongly accept the cookie believing it to be
fine. Since the tailmatching would then match all .com sites, the cookie would
then be sent to all of them.
The code now requires at least one letter between each dot for them to be
counted. Edited test case 61 to verify this.
When using the multi interface and connecting to a host name that
resolves to multiple IP addresses, there was no logic that made it
continue to the next IP if connecting to the first address times
out. This is now corrected.
The info about pipe status and expire cleared are clearly debug-related
and not anything mere mortals will or should care about so they are now
ifdef'ed DEBUGBUILD
Similar to what is done already for RCPT TO, the code now checks for and
adds angle brackets (<>) around the email address that is provided for
CURLOPT_MAIL_RCPT unless the app has done so itself.
Make sure that Curl_cache_addr() errors are propagated to callers of
loadhostpairs().
(this loadhostpairs function caused a scan-build warning due to the
'dns' variable getting assigned but never used)
Doing curlx_strtoofft() on the size just to figure out the end of it
causes a compiler warning since the result wasn't used, but is also a
bit of a waste.
Since the original `conn' pointer was used after the `connectdata' it
points to has been closed/cleaned up by Curl_reconnect_request it caused
a crash. We must make sure to use the newly created connection instead!
URL: http://curl.haxx.se/mail/lib-2010-12/0202.html
Make the c-ares resolver code ask for both IPv4 and IPv6 addresses when
IPv6 is enabled.
This is a workaround for the missing ares_getaddrinfo() and is a lot
easier to implement.
Note that as long as c-ares returns IPv4 addresses when IPv6 addresses
were requested but missing, this will cause a host's IPv4 addresses to
occur twice in the DNS cache.
URL: http://curl.haxx.se/mail/lib-2010-12/0041.html
The SSL_SERVER_VERIFY_LATER bit in the ssl_ctx_new() call allows the
code to verify the peer certificate explicitly after the handshake and
then the "data->set.ssl.verifypeer" option works.
The public axTLS header (at least as of 1.2.7) redefines the memory
functions. We #undef those again immediately after the public header to
limit the damage. This should be fixed in axTLS.
Failed HTTPS tests: 301, 306, 311, 312, 313, 560
311, 312 need more detailed error reporting from axTLS.
313 relates to CRL, which hasn't been implemented yet.
Added axTLS to autotool files and glue code to misc other files.
axtls.h maps SSL API functions, but may change.
axtls.c is just a stub file and will definitely change.
The function that checks if pipelining is possible now requires the HTTP
bit to be set so that it doesn't mistakenly tries to do it for other
protocols.
Bug: http://curl.haxx.se/mail/lib-2010-12/0152.html
Reported by: Dmitri Shubin
The generic timeout code must not check easy handles that are already
completed. Going to completed (again) within there risked decreasing the
number of alive handles again and thus it could go negative.
This regression bug was added in 7.21.2 in commit ca10e28f06
ossl_connect_common() now checks whether or not 'struct
connectdata->state' is equal 'ssl_connection_complete' and if so, will
return CURLE_OK with 'done' set to 'TRUE'. This check prevents
ossl_connect_common() from creating a new ssl connection on an existing
ssl session which causes openssl to fail when it tries to parse an
encrypted TLS packet since the cipher data was effectively thrown away
when the new ssl connection was created.
Bug: http://curl.haxx.se/mail/lib-2010-11/0169.html
It helps to prevent a hangup with some FTP servers in case idle session
timeout has exceeded. But it may be useful also for other protocols
that send any quit message on disconnect. Currently used by FTP, POP3,
IMAP and SMTP.
When looping in this function and checking for the timeout being
expired, it was not updating the reference time when calculating the
timediff since previous round which made it think each subsequent loop
to have taken longer than it actually did.
I also modified the function to use the generic Curl_timeleft() function
instead of the custom logic.
Bug: http://curl.haxx.se/bug/view.cgi?id=3112579
Ensure that spurious results from system's getaddrinfo() ares not propagated
by Curl_getaddrinfo_ex() into the library.
Also ensure that the ai_addrlen member of Curl_getaddrinfo_ex()'s output linked
list of Curl_addrinfo structures has appropriate family-specific address size.
On Windows, translate WSAGetLastError() to errno values as GNU
TLS does it internally, too. This is necessary because send() and
recv() on Windows don't set errno when they fail but GNU TLS
expects a proper errno value.
Bug: http://curl.haxx.se/bug/view.cgi?id=3110991
When no timeout is set, we call the socket_ready function with a timeout
value of 0 during handshake, which makes it loop too much/fast in this
function. It also made this function return CURLE_OPERATION_TIMEDOUT
wrongly on a slow handshake.
However, the particular bug report that highlighted this problem is not
solved by this fix, as this fix only makes the more proper error get
reported instead.
Bug: http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=594150
Reported by: Johannes Ernst
While changing Curl_sec_read_msg to accept an enum protection_level
instead of an int, I went ahead and fixed the usage of the associated
fields.
Some code was assuming that prot_clear == 0. Fixed those to use the
proper value. Added assertions prior to any code that would set the
protection level.
This is the advised way of checking for errors in the GSS-API RFC.
Also added some '\n' to the error message so that they are not mixed
with other outputs.
The IP version choice was previously only in the UserDefined struct
within the SessionHandle, but since we sometimes alter that option
during a request we need to have it on a per-connection basis.
I also moved more "init conn" code into the allocate_conn() function
which is designed for that purpose more or less.
CURLOPT_RESOLVE is a new option that sends along a curl_slist with
name:port:address sets that will populate the DNS cache with entries so
that request can be "fooled" to use another host than what otherwise
would've been used. Previously we've encouraged the use of Host: for
that when dealing with HTTP, but this new feature has the added bonus
that it allows the name from the URL to be used for TLS SNI and server
certificate name checks as well.
This is a first change. Surely more will follow to make it decent.
If the query result has a binary attribute, the binary attribute is
base64 encoded. But all following non binary attributes are also base64
encoded which is wrong.
This is a test (LDAP server is public).
curl
ldap://x500.bund.de:389/o=Bund,c=DE?userCertificate,certificateSerialNumber?sub
?cn=*Woehleke*
If you use a custom Host: name in a request to a SSL server, libcurl
will now use that given name when it verifies the server certificate to
be correct rather than using the host name used in the actual URL.
When given a custom host name in a Host: header, we can use it for
several different purposes other than just cookies, so we rename it and
use it for SSL SNI etc.
Some FTP servers (e.g. Pure-ftpd) end up hanging if we close the data
connection before transferring all the requested data. If we send ABOR
in that case, it prevents the server from hanging.
Bug: https://bugzilla.redhat.com/643656
Reported by: Pasi Karkkainen, Patrick Monnerat
These haven't worked in at least 8 years due to missing source
files, and most active RiscOS developers these days apparently
cross-compile anyway.
Signed-off-by: James Bursa <james@zamez.org>
In libssh2 1.2.8, libssh2_session_handshake() replaces
libssh2_session_startup() to fix the previous portability problem with
the socket type that was too small for win64 and thus easily could cause
crashes and more.
It is a bad idea to use the public prefix used by another library and
now we realize that libssh2 introduces a symbol in the upcoming version
1.2.8 that conflicts with our static function named libssh2_free.
When failing to build form post due to an error, the code now does a
proper failf(). Previously libcurl would report an error like "failed
creating formpost data" when a file wasn't possible to open which was
not easy for users to figure out.
I also lower cased a function name to be named more curl-style and
removed some unnecessary code.
The URL parser got a little stricter as it now considers a ? to be a
host name divider so that the slightly sloppier URLs work too. The
problem that made me do this change was the reported problem with an URL
like: www.example.com?email=name@example.com This form of URL is not
really a legal URL (due to the missing slash after the host name) but is
widely accepted by all major browsers and libcurl also already accepted
it, it was just the '@' letter that triggered the problem now.
The side-effect of this change is that now libcurl no longer accepts the
? letter as part of user-name or password when given in the URL, which
it used to accept (and is tested in test 191). That letter is however
mentioned in RFC3986 to be required to be percent encoded since it is
used as a divider.
Bug: http://curl.haxx.se/bug/view.cgi?id=3090268
In order to avoid for example the pingpong protocols to issue STARTTLS
(or equivalent) even though there's no SSL support built-in.
Reported by: Sune Ahlgren
Bug: http://curl.haxx.se/mail/archive-2010-10/0045.html
As the change in 5f0ae7a062 added a precaution against negative
file sizes that for some reason managed to get returned, this change now
introduces the same check at the second place in the code where the file
size from the libssh2 stat call is used.
This check might not be suitable for a 32 bit curl_off_t, but libssh2.h
assumes long long to work and to be 64 bit so I believe such a small
curl_off_t will be very unlikely to occur in the wild.
Renamed SDK_* to NDK_*; made NDK_* defines overwriteable from
environment; removed now obsolete YACC macro;
moved some curl_config.h defines to IPv6 section since they
are only needed when IPv6 is enabled - this makes libcurl compile
with older NDKs too which were not IPv6-aware.
We forgot to release the buffer passed to gss_init_sec_context.
The previous logic was difficult to read as we were reusing the same
variable (gssbuf) for both input buffer and output buffer. Splitted the
logic in 2 variables to better underline who needs to be released.
Also made the code break at 80 lines.
This fixes a memory leak related to the GSS-API code.
Added a krb5_init and krb5_end functions. Also removed a work-around
the lack of proper initialization of the GSS-API context.
It was pointed out that the special case libcurl did for 416 was
incorrect and wrong. 416 is not really different to other errors so the
response body must be handled like for other errors/http responses.
Reported by: Chris Smowton
Bug: http://curl.haxx.se/bug/view.cgi?id=3076808
It is still not clarified exactly why this happens, but libssh2
sometimes report a negative file size for the remote SFTP file and that
deeply confuses libcurl (or crashes it) so this precaution is added to
avoid badness.
Reported by: Ernest Beinrohr
Bug: http://curl.haxx.se/bug/view.cgi?id=3076430
