| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In OpenSSL < 1.1.0, the session_remove_cb callback is called inside the
global lock for CRYPTO_LOCK_SSL_CTX which is shared across the entire
process, not just for the specific SSL_CTX object. It is possible that
the callback releases GVL while the lock for CRYPTO_LOCK_SSL_CTX is
held, causing another thread calling an OpenSSL function that tries to
acquire the same lock stuck forever.
Add a note about the possible deadlock to the docs for
SSLContext#session_remove_cb=, and skip the relevant test cases unless
the OSSL_TEST_ALL environment variable is set to 1.
A deadlock due to this issue is observed:
http://ci.rvm.jp/results/trunk-test@frontier/104428
|
| |
|
|\
| |
| | |
ossl.c: make legacy locking callbacks reentrant
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Although it's not documented explicitly that the locking callbacks must
provide reentrant mutexes, it seems to be required.
Specifically, the session_remove_cb callback function of an SSL_CTX is
called in a critical section for CRYPTO_LOCK_SSL_CTX, which is shared
across the library. This leads, if the callback function calls another
OpenSSL function that will attempt to lock CRYPTO_LOCK_SSL_CTX, to
deadlock. SSL_CTX_free() is one example of such a function.
http://ci.rvm.jp/results/trunk@P895/64001
|
| |
| |
| |
| |
| | |
In preparation for making the mutexes reentrant. It is common to the
non-dynamic and the dynamic locking callbacks.
|
|\ \
| | |
| | | |
ssl: prevent SSLSocket#sysread* from leaking uninitialized data
|
| | |
| | |
| | |
| | |
| | |
| | | |
Set the length of the buffer string to 0 first, and adjust to the size
successfully read by the SSL_read() call later. This is needed because
the buffer string may be provided by the caller.
|
| |/
| |
| |
| |
| | |
The sleep was to ensure that the SSLSocket#read_nonblock will get
close_notify alert. A simple IO.select will suffice.
|
| |
| |
| |
| |
| |
| |
| | |
Skip test_openssl_engine_cipher_rc4 which will fail without RC4 support.
It may be disabled by 'no-rc4' configure option of the OpenSSL library.
Reference: https://github.com/ruby/openssl/issues/154
|
|/
|
|
|
|
|
| |
Use ignore_stderr option of assert_separately instead of $stderr.reopen
which may not work if the OpenSSL library uses a different stdio.
Reference: https://github.com/ruby/openssl/issues/154
|
|
|
|
|
|
|
| |
- Upgrade to latest Ubuntu LTS.
- Remove unnecessary packages.
- Update OpenSSL, LibreSSL, and Ruby versions. Notably, LibreSSL 2.6 is
added. Accordingly, .travis.yml is also updated to use that.
|
|
|
|
|
|
| |
Similar to the previous one, LibreSSL 2.6.1 has relevant functions such
as SSL_CTX_set_next_proto_select_cb(), but they are broken and do
nothing.
|
|
|
|
|
| |
LibreSSL 2.6.1 has SSL_CTX_set_tmp_ecdh_callback() function, but it does
not work.
|
|
|
|
|
|
|
| |
The test cases added by commit 8ed81ff4b0a8 ("test/test_pair: fix
test_write_nonblock{,_no_exceptions}", 2017-09-04) can consume much
memory and time if the OpenSSL supports SSL/TLS compression. Disable it
explicitly.
|
|
|
|
| |
Suppress "using default DH parameters" message.
|
|
|
|
|
|
|
|
|
|
| |
* include/ruby/ruby.h (ST2FIX): fix unnormalized Fixnum value bug
on mingw/mswin. [ruby-core:82687] [Bug #13877]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@59765 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
[ky: add ST2FIX() definition to ext/openssl/ruby_missing.h, and adapt
the test case to the 2.0 branch.]
Sync-with-trunk: r59765
|
|
|
|
|
|
|
|
|
| |
When the previous SSLSocket#write_nonblock call does not finish writing
the complete contents, SSL_shutdown() which is called through
SSLSocket#close will not send a close_notify alert.
As of commit e3a305063675 ssl_pair no longer uses the sync_close
feature. Do not expect that SSL_read() would get ECONNRESET.
|
|\
| |
| | |
test/test_ssl: explicitly accept TLS 1.1 in corresponding test
|
|/
|
|
|
|
| |
OpenSSL in Debian sid has recently disabled TLS < 1.2 by default, so in
order to test that TLS 1.1 works, we need to explicitly make our test
client accept it.
|
|\
| |
| | |
ssl: remove useless call to rb_thread_wait_fd()
|
| |
| |
| |
| |
| |
| | |
That there is no immediately readable data in the SSL instance does not
imply it has to read more bytes from the underlying socket. Just call
SSL_read() and check the return value.
|
|\ \
| |/
|/| |
Fix test failures with TLS 1.3-capable OpenSSL
|
| |
| |
| |
| | |
Fix test cases failing with TLS 1.3-enabled OpenSSL master.
|
| |
| |
| |
| |
| | |
Use TLS 1.2 explicitly where needed, since TLS 1.3 will remove session
ID based session resumption.
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The very patch that added this test case made the dfree function not
send close_notify alert when an SSLSocket is being GCed.
Anyway, the new OSSL_GC_STRESS option added by 6ee4b285036e ("test: run
test cases under GC.stress if OSSL_GC_STRESS is specified", 2016-12-04)
will cover this kind of issues.
|
| |
| |
| |
| |
| |
| | |
LibreSSL 2.6.1 removed DSA support from its SSL code. Also, TLS 1.3 will
not support DSA certificates. Use an RSA certificate as the client
certificate in the tests, too.
|
| |
| |
| |
| |
| |
| | |
Add methods that check whether the running OpenSSL is an OpenSSL or a
LibreSSL, and optionally check whether the version is newer or equal to
the given version number.
|
| |
| |
| |
| |
| |
| | |
start_server can hang if the given block exits before closing sockets
that the block opens. While this is a carelessness of the caller, we
can do a better job.
|
| |
| |
| |
| |
| | |
Close the socket by server_loop rather than by server_proc. This reduces
code duplication.
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
An assumption in OpenSSL::TestSSL#test_get_ephemeral_key is that the
ephemeral key type is always EVP_PKEY_EC when negotiated with an ECDHE
cipher suite. This is not true if X25519 is chosen.
The test is passing because we happen to fix the group to P-256 in
start_server, but let's make it explicit.
|
| |
| |
| |
| |
| |
| |
| | |
The block passed to start_server is invoked with two arguments, the
running thread object for the server and the automatically-selected port
number. The first argument is completely useless and actually is not
used anywhere.
|
| |
| |
| |
| |
| | |
Add a method that returns whether the OpenSSL supports TLS 1.2 or not.
This will be useful for test cases that are specific to TLS ~1.2.
|
| |
| |
| |
| |
| | |
Use EnvUtil.suppress_warning instead. We have started to use it already,
and the name 'suppress_warning' expresses what it does more clearly.
|
| |
| |
| |
| |
| |
| | |
Fix wrong nesting in test/utils.rb. Remove unnecessary requires. Wrap
the code with 'if defined?(OpenSSL::TestUtils) ~ end' and avoid class
definition with modifier if.
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
* topic/test-memory-leak:
Enable OSSL_MDEBUG on CI builds
Add OpenSSL.print_mem_leaks
test: prepare test PKey instances on demand
test: let OpenSSL::TestCase include OpenSSL::TestUtils
Don't define main() when built with --enable-debug
(cherry picked from commit 5c586acc387834ab4e09260937dc21064fc59de4)
Note that fix for new test cases that use the old constants removed by
this is squashed in.
|
|/
|
|
|
|
|
| |
The authentication tag can be set after starting the decryption, if
only it is before Cipher#final is called.
Fixes: https://github.com/ruby/openssl/issues/74
|
|
|
|
|
|
|
| |
As noted in the SSL_CTX_sess_set_remove_cb(3) manpage, SSL_CTX_free()
will call the callback function for each session in the internal session
store. We can't call the callback Proc since it may do a new object
allocation which is prohibited during GC.
|
| |
|
|
|
|
|
|
|
| |
OpenSSL <= 1.0.0 did not support TLS 1.1/1.2, and thus we must still
check the existence of the symbols. This fixes the previous commit,
3e5a009966bd ("ssl: remove unsupported TLS versions from
SSLContext::METHODS", 2017-08-08).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Check for all version-specific SSL methods. We do check for existence of
TLSv1_1_method() and TLSv1_2_method(), but not for TLSv1_method(). This
fixes compile error when OpenSSL is configured with no-tls1-method.
Also check the OPENSSL_NO_TLS{1,1_1,1_2} macros for whether OpenSSL
supports the corresponding versions or not. This prevents :TLSv1 from
being in SSLContext::METHODS when OpenSSL is compiled with no-tls1.
In particular, Debian sid has disabled TLS 1.0/1.1 support recently.
The changes in ext/openssl are partial backport of 4eb4b3297a92 ("Remove
support for OpenSSL 0.9.8 and 1.0.0", 2016-11-30).
|
|\
| |
| | |
Add msys2 library dependency tag in gem metadata
|
|/
|
|
|
|
|
|
|
|
|
|
|
|
| |
RubyInstaller2 supports metadata tags for installation of dependent
MSYS2/MINGW libraries. The openssl gem requires the mingw-openssl
package to be installed on the system, which the gem installer takes
care about, when this tag is set.
The feature is documented here:
https://github.com/oneclick/rubyinstaller2/wiki/For-gem-developers#msys2-library-dependency
This fixes issues like
https://github.com/oneclick/rubyinstaller2/issues/54 and
https://github.com/oneclick/rubyinstaller2/issues/53 .
|
|\
| |
| |
| |
| |
| |
| | |
* ky/pem-passwd-cb-get-rid-of-minlen:
ossl_pem_passwd_cb: handle nil from the block explicitly
ossl_pem_passwd_cb: do not check for taintedness
ossl_pem_passwd_cb: relax passphrase length constraint
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
There is code that returns nil in the passphrase block on purpose (to
prevent OpenSSL from prompting on stdin):
OpenSSL::PKey.read(File.read("file.pem")) { nil }
This is working just by chance because the TypeError from StringValue()
is silently ignored. Let's short circuit in that case and save raising
a needless exception, as this pattern has become too common.
|
| |
| |
| |
| | |
It is perfectly permissible to take passwords from an untrusted source.
|
|/
|
|
|
|
| |
The minimum passphrase length of 4 bytes is only a limitation of
PEM_def_callback() which isn't relevant here. Commit f38501249f33
introduced this bug.
|
| |
|
|
|
|
|
| |
Emulate the behavior of 'gem install --conservative'. This would prevent
overwriting the existing Rake installation.
|