Support both OpenSSL 1.1.0 and 1.0.2 #16130

In OpenSSL 1.1.0, X509_STORE_CTX is opaque and thus cannot be stack-allocated. This works in OpenSSL 1.1.0 and 1.0.2. Adapted from PR

This is cherry-picked from PR nodejs#8491 and then tidied up. The original had an unnecessarily large diff and messed up some public/private bits.

The exact sizes are not particularly important (the original value was missing all the objects hanging off anyway), only that V8 garbage collector be aware that there is some memory usage beyond the sockets themselves.

These are OpenSSL-internal APIs that are no longer accessible in 1.1.0 and weren't necessary. OpenSSL will push its own errors and, if it doesn't, the calling code would handle it anyway.

This is cherry-picked from PR nodejs#8491 and tidied up. This change does *not* account for the larger ticket key in OpenSSL 1.1.0. That will be done in a follow-up commit as the 48-byte ticket key is part of Node's public API.

Add a regression test for openssl/openssl#4384.

Parts of this were cherry-picked from PR nodejs#8491. Note that this only works with OpenSSL 1.0.2 or 1.1.0g or later. 1.1.0g is, as of writing, not yet released, but the fix is on the branch. See openssl/openssl#4384.

The callbacks are all no-ops in OpenSSL 1.1.0. This isn't necessary (the functions still exist for compatibility), but silences some warnings and avoids allocating a few unused mutexes.

In OpenSSL 1.1.0, EVP_CIPHER_CTX must be heap-allocated. Once we're heap-allocating them, there's no need in a separate initialised_ bit. The presence of ctx_ is sufficient.

Likewise, 1.1.0 requires EVP_MD_CTX be heap-allocated.

1.1.0 requires EVP_MD_CTX be heap-allocated. In doing so, move the Init and Update hooks to shared code because they are the same between Verify and Sign.

1.1.0 requries HMAC_CTX be heap-allocated.

In OpenSSL 1.1.0, EVP_dss1() is removed. These hash names were exposed in Node's public API, so add compatibility hooks for them.

This align the documentation with reality. This API never did what Node claims it did. The SSL_CIPHER_get_version function just isn't useful. In OpenSSL 1.0.2, it always returned the string "TLSv1/SSLv3" for anything but SSLv2 ciphers, which Node does not support. Note how test-tls-multi-pfx.js claims that ECDHE-ECDSA-AES256-GCM-SHA384 was added in TLSv1/SSLv3 which is not true. That cipher is new as of TLS 1.2. The OpenSSL 1.0.2 implementation is: char *SSL_CIPHER_get_version(const SSL_CIPHER *c) { int i; if (c == NULL) return ("(NONE)"); i = (int)(c->id >> 24L); if (i == 3) return ("TLSv1/SSLv3"); else if (i == 2) return ("SSLv2"); else return ("unknown"); } In OpenSSL 1.1.0, SSL_CIPHER_get_version changed to actually behave as Node documented it, but this changes the semantics of the function and breaks tests. The cipher's minimum protocol version is not a useful notion to return to the caller here, so just hardcode the string at "TLSv1/SSLv3" and document it as legacy.

Some errors in the two versions are different. The test-tls-no-sslv3 one because OpenSSL 1.1.x finally does version negotiation properly. 1.0.x's logic was somewhat weird and resulted in very inconsistent errors for SSLv3 in particular. Also the function codes are capitalized differently, but function codes leak implementation details, so don't assert on them to begin with.

"sha" in OpenSSL refers to SHA-0 which was removed from OpenSSL 1.1.0 and is insecure. Replace it with SHA-256 which is present in both 1.0.2 and 1.1.0. Short of shipping a reimplementation in Node, this is an unavoidable behavior change with 1.1.0.

OpenSSL 1.0.x used a 48-byte ticket key, but OpenSSL 1.1.x made it larger by using a larger HMAC-SHA256 key and using AES-256-CBC to encrypt. However, Node's public API exposes the 48-byte key. Implement the ticket key callback to restore the OpenSSL 1.0.x behavior.

OpenSSL 1.1.0 rejects RSA keys smaller than 1024 bits by default. Fix the tests to use larger ones. This test only cares that the PEM blob be missing a trailing newline. Certificate adapted from test/fixtures/cert.pem.

This test is testing what happens to the server if the client shuts off the connection (so the server sees ECONNRESET), but the way it does it is convoluted. It uses a static RSA key exchange with a tiny (384-bit) RSA key. The server doesn't notice (since it is static RSA, the client acts on the key first), so the client tries to encrypt a premaster and fails: rsa routines:RSA_padding_add_PKCS1_type_2:data too large for key size SSL routines:ssl3_send_client_key_exchange:bad rsa encrypt OpenSSL happens not to send an alert in this case, so we get ECONNRESET with no alert. This is quite fragile and, notably, breaks in OpenSSL 1.1.0 now that small RSA keys are rejected by libssl. Instead, test by just connecting a TCP socket and immediately closing it.

These are both no-ops in OpenSSL 1.1.0.

OpenSSL 1.1.0 disables anonymous ciphers unless building with enable-weak-crypto. Avoid unnecessary dependencies on these ciphers in tests.

This test is testing the workaround for an OpenSSL 1.0.x bug, which was fixed in 1.1.0. With the bug fixed, the test expectations need to change slightly.

This is kind of hairy. OpenSSL 1.0.2 ignored the return value and always treated everything as SSL_TLSEXT_ERR_NOACK (so the comment was wrong and Node was never sending a warning alert). OpenSSL 1.1.0 honors SSL_TLSEXT_ERR_NOACK vs SSL_TLSEXT_ERR_FATAL_ALERT and treats everything unknown as SSL_TLSEXT_ERR_FATAL_ALERT. Since this is a behavior change (tests break too), start by aligning everything on SSL_TLSEXT_ERR_NOACK. If sending no_application_protocol is desirable in the future, this can by changed to SSL_TLSEXT_ERR_FATAL_ALERT with whatever deprecation process is appropriate. However, note that, contrary to https://rt.openssl.org/Ticket/Display.html?id=3463#txn-54498, SSL_TLSEXT_ERR_FATAL_ALERT is *not* useful to a server with no fallback protocol. Even if such mismatches were rejected, such a server must *still* account for the fallback protocol case when the client does not advertise ALPN at all. Thus this may not be worth bothering.

Fixing the rest will be rather involved. I think the cleanest option is to deprecate the method string APIs which are weird to begin with.

The first group of tests makes one more connection and leave the server alive for longer. Otherwise the test is just catching that the server has closed the socket, depending on timing. This does not quite make the test pass yet, however. There are some quirks with how the http2 code handles errors which actually affect 1.0.2 as well.

This doesn't work in OpenSSL 1.1.0. Per discussion on the PR, it is preferable to just deprecate this setting. Deprecate it and skip the test in OpenSSL 1.1.0.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support both OpenSSL 1.1.0 and 1.0.2 #16130

Support both OpenSSL 1.1.0 and 1.0.2 #16130

Commits on Nov 2, 2017