When building software for the masses, it is sometimes not possible to
decide for all users which SSL backend is appropriate.
Git for Windows, for example, uses cURL to perform clones, fetches and
pushes via HTTPS, and some users strongly prefer OpenSSL, while other
users really need to use Secure Channel because it offers
enterprise-ready tools to manage credentials via Windows' Credential
Store.
The current Git for Windows versions use the ugly work-around of
building libcurl once with OpenSSL support and once with Secure Channel
support, and switching out the binaries in the installer depending on
the user's choice.
Needless to say, this is a super ugly workaround that actually only
works in some cases: Git for Windows also comes in a portable form, and
in a form intended for third-party applications requiring Git
functionality, in which cases this "swap out libcurl-4.dll" simply is
not an option.
Therefore, the Git for Windows project has a vested interest in teaching
cURL to make the SSL backend a *runtime* option.
This patch makes that possible.
By running ./configure with multiple --with-<backend> options, cURL will
be built with multiple backends.
For the moment, the backend can be configured using the environment
variable CURL_SSL_BACKEND (valid values are e.g. "openssl" and
"schannel").
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
So far, all of the SSL backends' private data has been declared as
part of the ssl_connect_data struct, in one big #if .. #elif .. #endif
block.
This can only work as long as the SSL backend is a compile-time option,
something we want to change in the next commits.
Therefore, let's encapsulate the exact data needed by each SSL backend
into a private struct, and let's avoid bleeding any SSL backend-specific
information into urldata.h. This is also necessary to allow multiple SSL
backends to be compiled in at the same time, as e.g. OpenSSL's and
CyaSSL's headers cannot be included in the same .c file.
To avoid too many malloc() calls, we simply append the private structs
to the connectdata struct in allocate_conn().
This requires us to take extra care of alignment issues: struct fields
often need to be aligned on certain boundaries e.g. 32-bit values need to
be stored at addresses that divide evenly by 4 (= 32 bit / 8
bit-per-byte).
We do that by assuming that no SSL backend's private data contains any
fields that need to be aligned on boundaries larger than `long long`
(typically 64-bit) would need. Under this assumption, we simply add a
dummy field of type `long long` to the `struct connectdata` struct. This
field will never be accessed but acts as a placeholder for the four
instances of ssl_backend_data instead. the size of each ssl_backend_data
struct is stored in the SSL backend-specific metadata, to allow
allocate_conn() to know how much extra space to allocate, and how to
initialize the ssl[sockindex]->backend and proxy_ssl[sockindex]->backend
pointers.
This would appear to be a little complicated at first, but is really
necessary to encapsulate the private data of each SSL backend correctly.
And we need to encapsulate thusly if we ever want to allow selecting
CyaSSL and OpenSSL at runtime, as their headers cannot be included within
the same .c file (there are just too many conflicting definitions and
declarations for that).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
At the moment, cURL's SSL backend needs to be configured at build time.
As such, it is totally okay for them to hard-code their backend-specific
data in the ssl_connect_data struct.
In preparation for making the SSL backend a runtime option, let's make
the access of said private data a bit more abstract so that it can be
adjusted later in an easy manner.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
In the ongoing endeavor to abstract out all SSL backend-specific
functionality, this is the next step: Instead of hard-coding how the
different SSL backends access their internal data in getinfo.c, let's
implement backend-specific functions to do that task.
This will also allow for switching SSL backends as a runtime option.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
These functions are all available via the Curl_ssl struct now, no need
to declare them separately anymore.
As the global declarations are removed, the corresponding function
definitions are marked as file-local. The only two exceptions here are
Curl_mbedtls_shutdown() and Curl_polarssl_shutdown(): only the
declarations were removed, there are no function definitions to mark
file-local.
Please note that Curl_nss_force_init() is *still* declared globally, as
the only SSL backend-specific function, because it was introduced
specifically for the use case where cURL was compiled with
`--without-ssl --with-nss`. For details, see f3b77e561 (http_ntlm: add
support for NSS, 2010-06-27).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The _shutdown() function calls the _session_free() function; While this
is not a problem now (because schannel.h declares both functions), a
patch looming in the immediate future with make all of these functions
file-local.
So let's just move the _session_free() function's definition before it
is called.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The entire idea of introducing the Curl_ssl struct to describe SSL
backends is to prepare for choosing the SSL backend at runtime.
To that end, convert all the #ifdef have_curlssl_* style conditionals
to use bit flags instead.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The SHA-256 checksumming is also an SSL backend-specific function.
Let's include it in the struct declaring the functionality of SSL
backends.
In contrast to MD5, there is no fall-back code. To indicate this, the
respective entries are NULL for those backends that offer no support for
SHA-256 checksumming.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The MD5 summing is also an SSL backend-specific function. So let's
include it, offering the previous fall-back code as a separate function
now: Curl_none_md5sum(). To allow for that, the signature had to be
changed so that an error could be returned from the implementation
(Curl_none_md5sum() can run out of memory).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
This is the first step to unify the SSL backend handling. Now all the
SSL backend-specific functionality is accessed via a global instance of
the Curl_ssl struct.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The idea of introducing the Curl_ssl struct was to unify how the SSL
backends are declared and called. To this end, we now provide an
instance of the Curl_ssl struct for each and every SSL backend.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
- Track when the cached encrypted data contains only a partial record
that can't be decrypted without more data (SEC_E_INCOMPLETE_MESSAGE).
- Change Curl_schannel_data_pending to return false in such a case.
Other SSL libraries have pending data functions that behave similarly.
Ref: https://github.com/curl/curl/pull/1387
Closes https://github.com/curl/curl/pull/1392
When UNICODE is not defined, the Curl_convert_UTF8_to_tchar macro maps
directly to its argument. As it is declared as a pointer to const and
InitializeSecurityContext expects a pointer to non-const, both MSVC and MinGW
issue a warning about implicitly casting away the const. Fix this by declaring
the variables as pointers to non-const.
Closes https://github.com/curl/curl/pull/1394
This commit introduces the CURL_SSLVERSION_MAX_* constants as well as
the --tls-max option of the curl tool.
Closes https://github.com/curl/curl/pull/1166
- Remove the SNI disabled when host verification disabled message
since that is incorrect.
- Show a message for legacy versions of Windows <= XP that connections
may fail since those versions of WinSSL lack SNI, algorithms, etc.
Bug: https://github.com/curl/curl/pull/1240
Fixes a few issues in manual wildcard cert name validation in
schannel support code for Win32 CE:
- when comparing the wildcard name to the hostname, the wildcard
character was removed from the cert name and the hostname
was checked to see if it ended with the modified cert name.
This allowed cert names like *.com to match the connection
hostname. This violates recommendations from RFC 6125.
- when the wildcard name in the certificate is longer than the
connection hostname, a buffer overread of the connection
hostname buffer would occur during the comparison of the
certificate name and the connection hostname.
* HTTPS proxies:
An HTTPS proxy receives all transactions over an SSL/TLS connection.
Once a secure connection with the proxy is established, the user agent
uses the proxy as usual, including sending CONNECT requests to instruct
the proxy to establish a [usually secure] TCP tunnel with an origin
server. HTTPS proxies protect nearly all aspects of user-proxy
communications as opposed to HTTP proxies that receive all requests
(including CONNECT requests) in vulnerable clear text.
With HTTPS proxies, it is possible to have two concurrent _nested_
SSL/TLS sessions: the "outer" one between the user agent and the proxy
and the "inner" one between the user agent and the origin server
(through the proxy). This change adds supports for such nested sessions
as well.
A secure connection with a proxy requires its own set of the usual SSL
options (their actual descriptions differ and need polishing, see TODO):
--proxy-cacert FILE CA certificate to verify peer against
--proxy-capath DIR CA directory to verify peer against
--proxy-cert CERT[:PASSWD] Client certificate file and password
--proxy-cert-type TYPE Certificate file type (DER/PEM/ENG)
--proxy-ciphers LIST SSL ciphers to use
--proxy-crlfile FILE Get a CRL list in PEM format from the file
--proxy-insecure Allow connections to proxies with bad certs
--proxy-key KEY Private key file name
--proxy-key-type TYPE Private key file type (DER/PEM/ENG)
--proxy-pass PASS Pass phrase for the private key
--proxy-ssl-allow-beast Allow security flaw to improve interop
--proxy-sslv2 Use SSLv2
--proxy-sslv3 Use SSLv3
--proxy-tlsv1 Use TLSv1
--proxy-tlsuser USER TLS username
--proxy-tlspassword STRING TLS password
--proxy-tlsauthtype STRING TLS authentication type (default SRP)
All --proxy-foo options are independent from their --foo counterparts,
except --proxy-crlfile which defaults to --crlfile and --proxy-capath
which defaults to --capath.
Curl now also supports %{proxy_ssl_verify_result} --write-out variable,
similar to the existing %{ssl_verify_result} variable.
Supported backends: OpenSSL, GnuTLS, and NSS.
* A SOCKS proxy + HTTP/HTTPS proxy combination:
If both --socks* and --proxy options are given, Curl first connects to
the SOCKS proxy and then connects (through SOCKS) to the HTTP or HTTPS
proxy.
TODO: Update documentation for the new APIs and --proxy-* options.
Look for "Added in 7.XXX" marks.
- Fix GnuTLS code for CURL_SSLVERSION_TLSv1_2 that broke when the
TLS 1.3 support was added in 6ad3add.
- Homogenize across code for all backends the error message when TLS 1.3
is not available to "<backend>: TLS 1.3 is not yet supported".
- Return an error when a user-specified ssl version is unrecognized.
---
Prior to this change our code for some of the backends used the
'default' label in the switch statement (ie ver unrecognized) for
ssl.version and treated it the same as CURL_SSLVERSION_DEFAULT.
Bug: https://curl.haxx.se/mail/lib-2016-11/0048.html
Reported-by: Kamil Dudka
Curl_select_ready() was the former API that was replaced with
Curl_select_check() a while back and the former arg setup was provided
with a define (in order to leave existing code unmodified).
Now we instead offer SOCKET_READABLE and SOCKET_WRITABLE for the most
common shortcuts where only one socket is checked. They're also more
visibly macros.
- Disable ALPN on Wine.
- Don't pass input secbuffer when ALPN is disabled.
When ALPN support was added a change was made to pass an input secbuffer
to initialize the context. When ALPN is enabled the buffer contains the
ALPN information, and when it's disabled the buffer is empty. In either
case this input buffer caused problems with Wine and connections would
not complete.
Bug: https://github.com/curl/curl/issues/983
Reported-by: Christian Fillion
Prior to this change we called Curl_ssl_getsessionid and
Curl_ssl_addsessionid regardless of whether session ID reusing was
enabled. According to comments that is in case session ID reuse was
disabled but then later enabled.
The old way was not intuitive and probably not something users expected.
When a user disables session ID caching I'd guess they don't expect the
session ID to be cached anyway in case the caching is later enabled.
Calling QueryContextAttributes with SECPKG_ATTR_APPLICATION_PROTOCOL
fails on Windows < 8.1 so we need to disable ALPN on these OS versions.
Inspiration provide by: Daniel Seither
Closes#848Fixes#840
Sessionid cache management is inseparable from managing individual
session lifetimes. E.g. for reference-counted sessions (like those in
SChannel and OpenSSL engines) every session addition and removal
should be accompanied with refcount increment and decrement
respectively. Failing to do so synchronously leads to a race condition
that causes symptoms like use-after-free and memory corruption.
This commit:
- makes existing session cache locking explicit, thus allowing
individual engines to manage lock's scope.
- fixes OpenSSL and SChannel engines by putting refcount management
inside this lock's scope in relevant places.
- adds these explicit locking calls to other engines that use
sessionid cache to accommodate for this change. Note, however,
that it is unknown whether any of these engines could also have
this race.
Bug: https://github.com/curl/curl/issues/815Fixes#815Closes#847
Only protocols that actually have a protocol registered for ALPN and NPN
should try to get that negotiated in the TLS handshake. That is only
HTTPS (well, http/1.1 and http/2) right now. Previously ALPN and NPN
would wrongly be used in all handshakes if libcurl was built with it
enabled.
Reported-by: Jay Satiro
Fixes#789
New tool option --ssl-no-revoke.
New value CURLSSLOPT_NO_REVOKE for CURLOPT_SSL_OPTIONS.
Currently this option applies only to WinSSL where we have automatic
certificate revocation checking by default. According to the
ssl-compared chart there are other backends that have automatic checking
(NSS, wolfSSL and DarwinSSL) so we could possibly accommodate them at
some later point.
Bug: https://github.com/bagder/curl/issues/264
Reported-by: zenden2k <zenden2k@gmail.com>
This commit is several drafts squashed together. The changes from each
draft are noted below. If any changes are similar and possibly
contradictory the change in the latest draft takes precedence.
Bug: https://github.com/bagder/curl/issues/244
Reported-by: Chris Araman
%%
%% Draft 1
%%
- return 0 if len == 0. that will have to be documented.
- continue on and process the caches regardless of raw recv
- if decrypted data will be returned then set the error code to CURLE_OK
and return its count
- if decrypted data will not be returned and the connection has closed
(eg nread == 0) then return 0 and CURLE_OK
- if decrypted data will not be returned and the connection *hasn't*
closed then set the error code to CURLE_AGAIN --only if an error code
isn't already set-- and return -1
- narrow the Win2k workaround to only Win2k
%%
%% Draft 2
%%
- Trying out a change in flow to handle corner cases.
%%
%% Draft 3
%%
- Back out the lazier decryption change made in draft2.
%%
%% Draft 4
%%
- Some formatting and branching changes
- Decrypt all encrypted cached data when len == 0
- Save connection closed state
- Change special Win2k check to use connection closed state
%%
%% Draft 5
%%
- Default to CURLE_AGAIN in cleanup if an error code wasn't set and the
connection isn't closed.
%%
%% Draft 6
%%
- Save the last error only if it is an unrecoverable error.
Prior to this I saved the last error state in all cases; unfortunately
the logic to cover that in all cases would lead to some muddle and I'm
concerned that could then lead to a bug in the future so I've replaced
it by only recording an unrecoverable error and that state will persist.
- Do not recurse on renegotiation.
Instead we'll continue on to process any trailing encrypted data
received during the renegotiation only.
- Move the err checks in cleanup after the check for decrypted data.
In either case decrypted data is always returned but I think it's easier
to understand when those err checks come after the decrypted data check.
%%
%% Draft 7
%%
- Regardless of len value go directly to cleanup if there is an
unrecoverable error or a close_notify was already received. Prior to
this change we only acknowledged those two states if len != 0.
- Fix a bug in connection closed behavior: Set the error state in the
cleanup, because we don't know for sure it's an error until that time.
- (Related to above) In the case the connection is closed go "greedy"
with the decryption to make sure all remaining encrypted data has been
decrypted even if it is not needed at that time by the caller. This is
necessary because we can only tell if the connection closed gracefully
(close_notify) once all encrypted data has been decrypted.
- Do not renegotiate when an unrecoverable error is pending.
%%
%% Draft 8
%%
- Don't show 'server closed the connection' info message twice.
- Show an info message if server closed abruptly (missing close_notify).
Some servers will request a client certificate, but not require one.
This change allows libcurl to connect to such servers when using
schannel as its ssl/tls backend. When a server requests a client
certificate, libcurl will now continue the handshake without one,
rather than terminating the handshake. The server can then decide
if that is acceptable or not. Prior to this change, libcurl would
terminate the handshake, reporting a SEC_I_INCOMPLETE_CREDENTIALS
error.
