One server I am testing responds to an SSLv3 ClientHello with TLSv1.2. If tls_sockets is being used, then testssl.sh responds with "#FIXME: downgraded. still missing a test case here." This PR fixes that, and in general checks the responses in run_protocols() more closely.
If tls_sockets is being used and the connection fails even though the server supports an earlier version of SSL/TLS, then it flags an error. If tls_sockets returns 2, then it verifies that $DETECTED_TLS_VERSION is equal to the highest version number supported by the server (that is also less than the version number in the ClientHello).
In addition, in order to test servers' support for version negotiation, it adds a new test that sends a TLSv1.4 ClientHello and verifies that the server responds with the highest version number that it supports. (This test only runs if both $using_sockets and $EXPERIMENTAL are true and server actually supports some version of SSL/TLS other than SSLv2.)
Changed to only include the signature algorithms extension for TLSv1.2, since RFC 5246 says:
Note: this extension is not meaningful for TLS versions prior to 1.2.
Clients MUST NOT offer it if they are offering prior versions.
However, even if clients do offer it, the rules specified in [TLSEXT]
require servers to ignore extensions they do not understand.
Inclusion of the extension for TLS 1.1 didn't seem to cause any harm, but it seems better to follow the RFC and not include it for TLSv1.0 or TLSv1.1.
RFC 7685 notes that there is at least one TLS implementation that hangs if the client sends a ClientHello with a TLSCiphertext.length between 256 and 511 bytes, and so the padding extension was defined in order to get around this bug. (OpenSSL s_client includes this extension when the -bugs option is used.) So, I changed socksend_tls_clienthello() to include the padding extension if the CLientHello would have a length between 256 and 511 bytes, making the padding extension just large enough to make the ClientHello 512 bytes.
I also fixed a typo (a missing "0x") in the check for whether any ECC ciphers are included in the Client Hello.
In doing some work on cipher_pref_check() I noticed that it was failing on SSLv2 since the call to "$OPENSSL s_client" includes SNI. I've also noticed in my testing that "$OPENSSL s_client" will not connect to an SSLv2-only server unless the "-ssl2" flag is included. So, I carefully checked each call to "$OPENSSL s_client" in the program (other than in run_allciphers and run_cipher_per_proto, since those functions are already addresses in PR #341) to see whether they would inappropriate fail with an SSLv2-only (or SSLv3-only) server.
As a general rule, if the call doesn't currently include the protocol, then I added "-ssl2" if $OPTIMAL_PROTO is "-ssl2", indicating that the server only supports SSLv2, and I removed any $SNI if a protocol is specified if a protocol is specified and it is either SSLv2 or SSLv3.
I tested it on an SSLv2-only server, and the results are much better. I also tested it on a collection of other servers, none of which support SSLv2, and the results are the same as with the current code.
The only thing I haven't been able to test is how the revised code works when the "--starttls" option is used. I don't believe the changes I made would cause anything to break in that case, but I also don't think code will work any better in that case, if the server only supports SSLv2. Of course, since no server should support SSLv2 (let alone only SSLv2), it shouldn't really be an issue.
One thing that I did not change, but that I do not understand; why does determine_optimal_proto() try the protocols in the order "-tls1_2 -tls1 -ssl3 -tls1_1 -ssl2" rather than "-tls1_2 -tls1_1 -tls1 -ssl3 -ssl2"? Doesn't the current ordering imply that TLS v1.0 and SSLv3 are better than TLS v1.1?
Versions of OpenSSL prior to 1.1.0 ignore the options "-tls1_1" and "-tls1_2". So, a call of the form "$OPENSSL ciphers -tls1_2 -V 'ALL:COMPLEMENTOFALL:@STRENGTH' would list all supported ciphers (including SSLv2 ciphers), not just ciphers appropriate for TLS1.2.
This changes the code to use "-tls1" instead of "-tls1_1" or "-tls1_2" if a version of OpenSSL other than 1.1.0 is being used.
I changed the code to use the global $HAS_SSL2 rather than $sslv2_locally_supported.
I don't think there's a need to use $HAS_SSL3 in run_allciphers(), since the call to "$OPENSSL s_client" for non-SSLv2 ciphers does not specify a protocol. It's also not needed in run_cipher_per_proto(), since there is already a call to locally_supported() before anything further is done with a protocol.
This PR adds the signature algorithms, heartbeat, session ticket, and next protocol extensions to the client hello message created by socksend_tls_clienthello() for TLS 1.0 and above. It also adds the supported elliptic curves and ec points format extensions if the client hello message includes any ECC cipher suites.
I tested this version against several servers with $EXPERIMENTAL set to true and get the same results as with the current code with $EXPERIMENTAL set to false.
This PR addresses two problems related to SSLv2 in run_allciphers() and run_cipher_per_proto().
In run_cipher_per_proto(), the call to "$OPENSSL s_client" is changed to that $SNI is not included if $proto is -sslv2 or -sslv3. As noted in a comment within run_prototest_openssl(), "newer openssl throw an error if SNI is supplied with SSLv2" and "SSLv3 doesn't have SNI (openssl doesn't complain though -- yet)."
run_allciphers() will sometimes incorrectly report that a server supports an SSLv2 cipher, even if the server does not support SSLv2 at all. The problem occurs if there is a supported SSLv3 cipher suite that has the same OpenSSL name as an SSLv2 cipher suite (e.g., RC4-MD5). Since the call to "$OPENSSL s_client" only uses the OpenSSL name, but the results report both the name, hexcode, and RFC cipher suite name, both the SSLv2 and SSLv3 cipher suites are reported as being supported (e.g., 0x04=RC4-MD5=TLS_RSA_WITH_RC4_128_MD5 and x010080=RC4-MD5=SSL_CK_RC4_128_WITH_MD5). This PR fixes the problem by testing SSLv2 cipher suites separately from non-SSLv2 cipher suites.
The last line of neat_list currently uses $HEXC as the parameter to show_rfc_style(), but it should use $hexcode. At the moment using $HEXC instead of $hexcode makes no difference, since hexcode="$1" and in all calls to neat_list() the first parameter is $HEXC. However, this bug could create problems in the future since neat_list() will misbehave if the value of the first parameter (hexcode) isn't the same as $HEXC.
This PR makes basically the same changes to run_cipher_per_proto() as I previously made to run_allciphers(). The main difference is that in this function, round 0 consists of a single call to "$OPENSSL s_client" with "-cipher" including all of the locally supported ciphers. The reason for the difference is that in run_allciphers() its saves time to assume the server supports at least one cipher suite. In the case of run_cipher_per_proto(), however, it is likely that the server will not support some protocols at all, so its usually faster to start with a single call to "$OPENSSL s_client" that tests whether the server supports the protocol at all.
There is a tab on the line for SSL_CK_RC2_128_CBC_WITH_MD5. When testssl.sh is called with "-E" and "--show-each," this causes the string "not a/v" to be printed two characters to the right of the same string on every other line (at least on Linux systems). This PR just deletes the tab character.
The run_allciphers() function currently works by calling "$OPENSSL s_client" once for each cipher suite supported by $OPENSSL. In the case of "OpenSSL 1.0.2-chacha (1.0.2e-dev)" that means 195 calls to "$OPENSSL s_client" even though servers tend to only support a small fraction of these cipher suites.
This PR produces the same output as the current run_allciphers() with fewer calls to "$OPENSSL s_client", which results in the function running faster (usually much faster). The basic idea behind the revised function is to test cipher suites in blocks. If $OPENSSL supports 195 cipher suites, then it group these cipher suites into 4 blocks of 64 (with the final block being smaller). It makes one call to "$OPENSSL s_client" with cipher suites 1-64, and if it fails, then it knows that none of these 64 cipher suites are supported by the server and it doesn't need to perform any more tests on these 64 cipher suites. If it succeeds, then it breaks the 64 cipher suites into 4 blocks of 16 and calls "$OPENSSL s_client" with each of those blocks. The blocks of 16 that are successful are broken into blocks of 4, and for each of the successful blocks of 4 the individual cipher suites are tested.
For testssl.sh and www.google.com the number of calls to "$OPENSSL s_client" is reduced from 195 to 88. For github.com the number of calls is reduced to 56!
I haven't made any changes to run_cipher_per_proto yet, but if this PR is accepted I can make the same changes in that function.
Thanks,
David
David Cooper
Dirk Wetter
Niko78
Thomas Patzke
Florian Baumann