Correctly pass the new object assigned by StringValue() to
ossl_ssl_write_internal_safe().
This is a follow-up to commit 0d8c17aa85 (Reduce
OpenSSL::Buffering#do_write overhead, 2024-12-21).
3ff096196a
Commit aa7f03e18f broke test_minmax_version and test_fallback_scsv
on systems using OpenSSL 1.1.1 with a system-wide configuration file
that specifies MinProtocol=TLSv1.2.
20250228T003003Z.fail.html.gz
20250228T003003Z.fail.html.gz
These test cases were already broken before the commit, but they were
being skipped because check_supported_protocol_versions failed to detect
TLS 1.1 support. To avoid affected by the configuration file, explicitly
reset SSLContext#min_version when TLS 1.1 or older is required.
The test cases are also broken with OpenSSL 3.0 or later, but this is
not currently visible because it still fails to detect TLS 1.1 support.
This is caused by the default SSLContext#security_level value, as
OpenSSL 3.0 changed TLS 1.1 to be disabled at level 1.
6d0ea81b5e
As reported in <https://github.com/ruby/ruby/pull/12823>,
check_supported_protocol_versions is unstable and occasionally fails
with Errno::ECONNABORTED during SSLSocket#connect on Windows.
When the server-side SSLContext specifies an unsupported SSL/TLS
protocol version, start_server accepts a TCP connection but closes it
without reading ClientHello, as SSLSocket#accept immediately raises an
exception. With Winsock, this can cause the client-side
SSLSocket#connect to raise Errno::ECONNABORTED.
While the simplest fix is to add rescue Errno::ECONNABORTED, this method
can be simplified. Instead, let's set up a server that accepts all
protocol versions and test client connections with different settings.
aa7f03e18f
Starting with LibreSSL 3.5, OpenSSL::X509::ExtensionFactory refuses to
create SAN extensions that are not valid according to RFC 6125. While
this behavior makes sense, we need such invalid extensions to test our
own validation routine. Let's construct SAN extensions manually instead.
b420d6d739
There are a few SSL discrepencies in AWS-LC when compared to OpenSSL.
1. AWS-LC has slightly different error messages (in all-caps).
2. AWS-LC has no support for DHE ciphersuites.
3. There are no concepts of SSL security levels within AWS-LC.
4. Similar to LibreSSL, there is no support for OPENSSL_CONF.
a60d050342
OpenSSL::SSL::SSLContext#cert_store= uses SSL_CTX_set_cert_store(). The
store is used for verifying peer certificates and for building
certificate chains to be sent to the peer if there is no chain
explicitly provided by SSLContext#extra_chain_cert=.
Do not specify it in the common test helper start_server, as most
callers do not require either function. Instead, update individual test
cases that use client certificates to explicitly specify it in ctx_proc.
A more direct test case is added to verify the latter function.
9daecee615
assert_handshake_error is useful for checking handshake failures
triggered by the peer, as the underlying socket may be closed
prematurely, leading to different exceptions depending on the platform
and timing.
However, when the local end aborts a handshake, the only possible
exception is OpenSSL::SSL::SSLError. Use stricter assertions in such
cases.
637ba65818
The list of NPN protocols is validated in SSLContext#setup.
The assert_handshake_error is misleading. The client is unable to start
a handshake at all because the server is not running.
e8db6ffd9e
Use start_server instead of start_server_version.
start_server_version is a wrapper around start_server that forces the
server to a specific protocol version using the now-deprecated method
SSLSocket#ssl_version=, but it does more than that. The slightly
different method signature and default values are confusing. Let's
use start_server directly.
22ed31d77e
Make these methods simple wrappers around
SSL_CTX_set_{min,max}_proto_version().
When we introduced these methods in commit 18603949d3 [1], which went
to v2.1.0, we added a private method to SSLContext that set both the
minimum and maximum protocol versions at the same time. This was to
allow emulating the behavior using SSL options on older OpenSSL versions
that lack SSL_CTX_set_{min,max}_proto_version(). Since we no longer
support OpenSSL 1.0.2, the related code has already been removed.
In OpenSSL 1.1.1 or later, setting the minimum or maximum version to 0
is not equivalent to leaving it unset. Similar to SSL options, which we
avoid overwriting as of commit 00bec0d905 and commit 77c3db2d65 [2],
a system-wide configuration file may define a default protocol version
bounds. Setting the minimum version should not unset the maximum
version, and vice versa.
[1] https://github.com/ruby/openssl/pull/142
[2] https://github.com/ruby/openssl/pull/7675766386321
Commit 3bbf5178a9 made blocking methods on SSLSocket follow the
IO#timeout= value. The commit changed io_wait_readable() to potentially
raise an exception without unlocking the String.
The String is currently locked for the entire duration of a #sysread
method call. This does not seem to be necessary, as SSL_read() does not
require that the same buffer is specified when retrying. Locking the
String during each SSL_read() call should be sufficient.
8f791d73f5
Drop support for OpenSSL 1.1.0. OpenSSL 1.1.0 was a non-LTS release and
it has reached upstream EOL in 2019-12 along with OpenSSL 1.0.2.
Distributions that shipped with OpenSSL 1.1.0 include:
- Debian 9 (EOL 2022-06)
- Ubuntu 18.04 LTS (EOL 2023-04)
ba83abe920
This reverts the change made to this test case in commit a0e98d48c9
(Enhance TLS 1.3 support on LibreSSL 3.2/3.3, 2020-12-03).
Part of the test case was skipped on LibreSSL because LibreSSL 3.2.2
introduced a stricter check during creation of the extension. The check
was then relaxed in LibreSSL 3.4.0.
187b176ecd
Apply SSL options set in DEFAULT_PARAMS without clearing existing
options.
It currently clears options in order to avoid setting one of the
options included in OpenSSL::SSL::OP_ALL unless explicitly specified,
namely OpenSSL::SSL::OP_DONT_INSERT_EMPTY_FRAGMENTS. Now that
OpenSSL::SSL::OP_ALL has been removed from SSLContext#initialize, it is
no longer necessary.
77c3db2d65
Respect the SSL options set by default by SSL_CTX() and by the
system-wide OpenSSL configuration file.
OpenSSL::SSL::SSLContext#initialize currently adds OpenSSL::SSL::OP_ALL
on top of the default SSL options. Let's stop doing it.
OpenSSL::SSL::OP_ALL is a set of options that changes OpenSSL's behavior
to workaround various TLS implementation bugs. Using it is considered
usually safe, but is not completely harmless.
00bec0d905
When compiled with OpenSSL <= 1.1.1, OpenSSL::SSL::SSLContext#setup
does not raise an exception on an error return from
SSL_CTX_load_verify_locations(), but instead only prints a verbose-mode
warning. This is not helpful since it very likely indicates an actual
error, such as the specified file not being readable.
Also, OpenSSL's error queue is not correctly cleared:
$ ruby -w -ropenssl -e'OpenSSL.debug=true; ctx=OpenSSL::SSL::SSLContext.new; ctx.ca_file="bad-path"; ctx.setup; pp OpenSSL.errors'
-e:1: warning: can't set verify locations
["error:02001002:system library:fopen:No such file or directory",
"error:2006D080:BIO routines:BIO_new_file:no such file",
"error:0B084002:x509 certificate routines:X509_load_cert_crl_file: system lib"]
The behavior is currently different when compiled with OpenSSL >= 3.0:
SSLError is raised if SSL_CTX_load_verify_file() or
SSL_CTX_load_verify_dir() fails.
This inconsistency was unintentionally introduced by commit 5375a55ffc
("ssl: use SSL_CTX_load_verify_{file,dir}() if available", 2020-02-22).
However, raising SSLError seems more appropriate in this situation.
Let's adjust the OpenSSL <= 1.1.1 code so that it behaves the same way
as the OpenSSL >= 3.0 code currently does.
Fixes: https://github.com/ruby/openssl/issues/6497eb10f7b75
This reverts commit fc4629d246.
The test case "test_connect_certificate_verify_failed_exception_message"
does want to check the reason behind a certificate verification failure
to be included in the exception message.
c309745eb8
A tentative measures fo https://github.com/ruby/openssl/issues/606.
With OpenSSL 3.1.0, the error message at connection using "self-signed
certificate" seems to return `SSL_R_TLSV1_ALERT_UNKNOWN_CA` instead of
`SSL_R_CERTIFICATE_VERIFY_FAILED`.
fc4629d246
As noted in commit a2ed156cc9 ("test/test_ssl: do not run NPN tests
for LibreSSL >= 2.6.1", 2017-08-13), NPN is known not to work properly
on LibreSSL.
Disable NPN support on LibreSSL, whether OPENSSL_NO_NEXTPROTONEG is
defined or not.
NPN is less relevant today anyway. Let's also silence test suite when
it's not available.
289f6e0e1f
- This callback is invoked when TLS key material is generated or
received, in order to allow applications to store this keying material
for debugging purposes.
- It is invoked with an `SSLSocket` and a string containing the key
material in the format used by NSS for its SSLKEYLOGFILE debugging
output.
- This commit adds the Ruby binding `keylog_cb` and the related tests
- It is only compatible with OpenSSL >= 1.1.1. Even if LibreSSL implements
`SSL_CTX_set_keylog_callback()` from v3.4.2, it does nothing (see
648d39f0f0)
3b63232cf1
RFC 6066 states how some wildcard SAN entries MAY be handled, but
it does not say they MUST be handled. LibreSSL 3.5.0 only handles
suffix wildcard SANs, not prefix wildcard SANs, or interior
wildcard SANs, so return early from the wildcard SAN tests on
LibreSSL 3.5.0.
Fixes#471717d7009d6
test_connect_certificate_verify_failed_exception_message occasionally
fails. Is it possible that OpenSSL sees a different clock from Ruby by
more than 10 seconds?
http://ci.rvm.jp/logfiles/brlog.trunk-random0.20211111-072828
Let's give more time after the certificate expiration date to see if
this fixes the flakiness. We had similar occasional failures in
test_x509store.rb before, which disappeared after ruby/ruby commit
7930a352a5 and ruby/openssl commit fb2fcbb13734.
OpenSSL::SSL::SSLSocket allowed #read and #write to be called before an
SSL/TLS handshake is completed. They passed unencrypted data to the
underlying socket.
This behavior is very odd to have in this library. A verbose mode
warning "SSL session is not started yet" was emitted whenever this
happened. It also didn't behave well with OpenSSL::Buffering. Let's
just get rid of it.
Fixes: https://github.com/ruby/openssl/issues/9bf780748b3
SSLSocket#connect eventually calls `GetOpenFile` in order to get the
underlying file descriptor for the IO object passed in on
initialization. `GetOpenFile` assumes that the Ruby object passed in is
a T_FILE object and just casts it to a T_FILE without any checks. If
you pass an object that *isn't* a T_FILE to that function, the program
will segv.
Since we assume the IO object is a file in the `connect` method, this
commit adds a `CheckType` in the initialize method to ensure that the IO
object is actually a T_FILE. If the object *isn't* a T_FILE, this class
will segv on `connect`, so I think this is a backwards compatible
change.
919fa44ec2
A certificate can only have one SubjectAltName extension. OpenSSL 3.0
performs a stricter validation and certificates containing multiple SANs
will be rejected.
558cfbe5f5
OpenSSL 3.0 slightly changed the error message for a certificate
verification failure when an untrusted self-signed certificate is found
in the chain.
b5a0a19850
Disabling ECC support of OpenSSL is impractical nowadays.
We still try to have the C extension compile on no-ec builds (as well
as no-dh or no-engine, etc.) as long as we can, but keeping test cases
for such an extreme scenario is not worth the effort.
2cd01d4676
Provide a wrapper of SSL_set0_tmp_dh_pkey()/SSL_CTX_set_tmp_dh(), which
sets the DH parameters used for ephemeral DH key exchange.
SSLContext#tmp_dh_callback= already exists for this purpose, as a
wrapper around SSL_CTX_set_tmp_dh_callback(), but it is considered
obsolete and the OpenSSL API is deprecated for future removal. There is
no practical use case where an application needs to use different DH
parameters nowadays. This was originally introduced to support export
grade ciphers.
RDoc for #tmp_dh_callback= is updated to recommend the new #tmp_dh=.
Note that current versions of OpenSSL support automatic ECDHE curve
selection which is enabled by default. SSLContext#tmp_dh= should only be
necessary if you must allow ancient clients which don't support ECDHE.
aa43da4f04
