Motivation:
Each call to SSL_write may introduce about ~100 bytes of overhead. The OpenSslEngine (based upon OpenSSL) is not able to do gathering writes so this means each wrap operation will incur the ~100 byte overhead. This commit attempts to increase goodput by aggregating the plaintext in chunks of <a href="https://tools.ietf.org/html/rfc5246#section-6.2">2^14</a>. If many small chunks are written this can increase goodput, decrease the amount of calls to SSL_write, and decrease overall encryption operations.
Modifications:
- Introduce SslHandlerCoalescingBufferQueue in SslHandler which will aggregate up to 2^14 chunks of plaintext by default
- Introduce SslHandler#setWrapDataSize to control how much data should be aggregated for each write. Aggregation can be disabled by setting this value to <= 0.
Result:
Better goodput when using SslHandler and the OpenSslEngine.
Motivation:
The JDK SSLEngine documentation says that a call to wrap/unwrap "will attempt to consume one complete SSL/TLS network packet" [1]. This limitation can result in thrashing in the pipeline to decode and encode data that may be spread amongst multiple SSL/TLS network packets.
ReferenceCountedOpenSslEngine also does not correct account for the overhead introduced by each individual SSL_write call if there are multiple ByteBuffers passed to the wrap() method.
Modifications:
- OpenSslEngine and SslHandler supports a mode to not comply with the limitation to only deal with a single SSL/TLS network packet per call
- ReferenceCountedOpenSslEngine correctly accounts for the overhead of each call to SSL_write
- SslHandler shouldn't cache maxPacketBufferSize as aggressively because this value may change before/after the handshake.
Result:
OpenSslEngine and SslHanadler can handle multiple SSL/TLS network packet per call.
[1] https://docs.oracle.com/javase/7/docs/api/javax/net/ssl/SSLEngine.html
Motivation:
SslHandlerTest#testCompositeBufSizeEstimationGuaranteesSynchronousWrite has been observed to fail on CI servers. Knowing how many bytes were seen by the client would be helpful.
Modifications:
- Add bytesSeen to the exception if the client closes early.
Result:
More debug info available.
Motivation:
PR #6811 introduced a public utility methods to decode hex dump and its parts, but they are not visible from netty-common.
Modifications:
1. Move the `decodeHexByte`, `decodeHexDump` and `decodeHexNibble` methods into `StringUtils`.
2. Apply these methods where applicable.
3. Remove similar methods from other locations (e.g. `HpackHex` test class).
Result:
Less code duplication.
Motivation:
For historical reasons OpenSSL's internal naming convention for CHACHA20 based cipher suites does not include the HMAC algorithm in the cipher name. This will prevent the CHACHA20 cipher suites from being used if the RFC cipher names are specified.
Modifications:
- Add a special case for CHACHA20 cipher name conversions in CipherSuiteConverter
- Update OPENSSL_CIPHERSUITE_PATTERN to accommodate the new naming scheme for CHACHA20 cipher suites
Result:
CipherSuiteConverter now works with CHACHA20 cipher suites.
Motivation
RFC 6066 (https://tools.ietf.org/html/rfc6066#page-6) says that the hostname in the SNI extension is ASCII encoded but Netty decodes it using UTF-8.
Modifications
Use ASCII instead of UTF-8
Result
Fixes#6717
Motivation:
We not correctly handle LE buffers when try to read the packet length out of the buffer and just assume it always is a BE buffer.
Modifications:
Correctly account for the endianess of the buffer when reading the packet lenght.
Result:
Fixes [#6709].
Motivation:
SSL_write requires a fixed amount of bytes for overhead related to the encryption process for each call. OpenSslEngine#wrap(..) will attempt to encrypt multiple input buffers until MAX_PLAINTEXT_LENGTH are consumed, but the size estimation provided by calculateOutNetBufSize may not leave enough room for each call to SSL_write. If SSL_write is not able to completely write results to the destination buffer it will keep state and attempt to write it later. Netty doesn't account for SSL_write keeping state and assumes all writes will complete synchronously (by attempting to allocate enough space to account for the overhead) and feeds the same data to SSL_write again later which results in corrupted data being generated.
Modifications:
- OpenSslEngine#wrap should only produce a single TLS packet according to the SSLEngine API specificaiton [1].
[1] https://docs.oracle.com/javase/8/docs/api/javax/net/ssl/SSLEngine.html#wrap-java.nio.ByteBuffer:A-int-int-java.nio.ByteBuffer-
- OpenSslEngine#wrap should only consider a single buffer when determining if there is enough space to write, because only a single buffer will ever be consumed.
Result:
OpenSslEngine#wrap will no longer produce corrupted data due to incorrect accounting of space required in the destination buffers.
Motivation:
When adding SNIMatcher support we missed to use static delegating methods and so may try to load classes that not exists in Java7. Which will lead to errors.
Modifications:
- Correctly only try to load classes when running on java8+
- Ensure Java8+ related tests only run when using java8+
Result:
Fixes [#6700]
Motivation:
Java8 adds support for SNIMatcher to reject SNI when the hostname not matches what is expected. We not supported doing this when using SslProvider.OPENSSL*.
Modifications:
- Add support for SNIMatcher when using SslProvider.OPENSSL*
- Add unit tests
Result:
SNIMatcher now support with our own SSLEngine as well.
Motivation:
We need to ensure we only try to to test with the SslProviders that are supported when running the SslHandlerTest.testCompositeBufSizeEstimationGuaranteesSynchronousWrite test.
Modifications:
Skip SslProvider.OPENSSL* if not supported.
Result:
No more test-failures if openssl is not installed on the system.
Motivation:
Java8 adds support for SNIMatcher to reject SNI when the hostname not matches what is expected. We not supported doing this when using SslProvider.OPENSSL*.
Modifications:
- Add support for SNIMatcher when using SslProvider.OPENSSL*
- Add unit tests
Result:
SNIMatcher now support with our own SSLEngine as well.
https://github.com/netty/netty-tcnative/pull/215
Motivation
OCSP stapling (formally known as TLS Certificate Status Request extension) is alternative approach for checking the revocation status of X.509 Certificates. Servers can preemptively fetch the OCSP response from the CA's responder, cache it for some period of time, and pass it along during (a.k.a. staple) the TLS handshake. The client no longer has to reach out on its own to the CA to check the validity of a cetitficate. Some of the key benefits are:
1) Speed. The client doesn't have to crosscheck the certificate.
2) Efficiency. The Internet is no longer DDoS'ing the CA's OCSP responder servers.
3) Safety. Less operational dependence on the CA. Certificate owners can sustain short CA outages.
4) Privacy. The CA can lo longer track the users of a certificate.
https://en.wikipedia.org/wiki/OCSP_staplinghttps://letsencrypt.org/2016/10/24/squarespace-ocsp-impl.html
Modifications
https://www.openssl.org/docs/man1.0.2/ssl/SSL_set_tlsext_status_type.html
Result
High-level API to enable OCSP stapling
Motivation:
`io.netty.handler.logging.LoggingHandler` does not log when these
events happen.
Modifiations:
Add overrides with logging to these methods.
Result:
Logging now happens for these two events.
Motivation:
In OpenSslCertificateException we tried to validate the supplied error code but did not correctly account for all different valid error codes and so threw an IllegalArgumentException.
Modifications:
- Fix validation by updating to latest netty-tcnative and use CertificateVerifier.isValid
- Add unit tests
Result:
Validation of error code works as expected.
Motivation:
1419f5b601 added support for conscrypt but the CI started to fail when running tests with java7 as conscrypt is compiled with java8. This was partly fixed in c4832cd9d9 but we also need to ensure we not try to even load the classes.
Modifications:
Only try to load conscrypt classes when on java8+-
Result:
CI not fails anymore.
Motivation:
Conscrypt is a Java Security provider that wraps OpenSSL (specifically BoringSSL). It's a possible alternative to Netty-tcnative that we should explore. So this commit is just to enable us to further investigate its use.
Modifications:
Modifying the SslContext creation path to support the Conscrypt provider.
Result:
Netty will support OpenSSL with conscrypt.
Motivation:
We should limit the size of the allocated outbound buffer to MAX_ENCRYPTED_PACKET_LENGTH to ensure we not cause an OOME when the user tries to encrypt a very big buffer.
Modifications:
Limit the size of the allocated outbound buffer to MAX_ENCRYPTED_PACKET_LENGTH
Result:
Fixes [#6564]
Motivation:
ChunkedWriteHandler queues written messages and actually writes them
when flush is called. In its doFlush method, it needs to flush after
each chunk is written to preserve memory. However, non-chunked messages
(those that aren't of type ChunkedInput) are treated in the same way,
which means that flush is called after each message is written.
Modifications:
Moved the call to flush() inside the if block that tests if the message
is an instance of ChunkedInput. To ensure flush is called at least once,
the existing boolean flushed is checked at the end of doFlush. This
check was previously in ChunkedWriteHandler.flush(), but wasn't checked in
other invocations of doFlush, e.g. in channelInactive.
Result:
When this handler is present in a pipeline, writing a series
of non-chunked messages will be flushed as the developer intended.
Motivation:
Some pipelines require support for both SSL and non-SSL messaging.
Modifications:
Add utility decoder to support both SSL and non-SSL handlers based on the initial message.
Result:
Less boilerplate code to write for developers.
Motivation:
SslContext and SslContextBuilder do not support a way to specify the desired TLS protocols. This currently requires that the user extracts the SSLEngine once a context is built and manually call SSLEngine#setEnabledProtocols(String[]). Something this critical should be supported at the SslContext level.
Modifications:
- SslContextBuilder should accept a list of protocols to configure for each SslEngine
Result:
SslContext consistently sets the supported TLS/SSL protocols.
Motivaiton:
It is possible that if the OpenSSL library supports the interfaces required to use the KeyManagerFactory, but we fail to get the io.netty.handler.ssl.openssl.useKeyManagerFactory system property (or this property is set to false) that SSLEngineTest based unit tests which use a KeyManagerFactory will fail.
Modifications:
- We should check if the OpenSSL library supports the KeyManagerFactory interfaces and if the system property allows them to be used in OpenSslEngineTests
Result:
Unit tests which use OpenSSL and KeyManagerFactory will be skipped instead of failing.
Motivation:
When we do a wrap operation we calculate the maximum size of the destination buffer ahead of time, and return a BUFFER_OVERFLOW exception if the destination buffer is not big enough. However if there is a CompositeByteBuf the wrap operation may consist of multiple ByteBuffers and each incurs its own overhead during the encryption. We currently don't account for the overhead required for encryption if there are multiple ByteBuffers and we assume the overhead will only apply once to the entire input size. If there is not enough room to write an entire encrypted packed into the BIO SSL_write will return -1 despite having actually written content to the BIO. We then attempt to retry the write with a bigger buffer, but because SSL_write is stateful the remaining bytes from the previous operation are put into the BIO. This results in sending the second half of the encrypted data being sent to the peer which is not of proper format and the peer will be confused and ultimately not get the expected data (which may result in a fatal error). In this case because SSL_write returns -1 we have no way to know how many bytes were actually consumed and so the best we can do is ensure that we always allocate a destination buffer with enough space so we are guaranteed to complete the write operation synchronously.
Modifications:
- SslHandler#allocateNetBuf should take into account how many ByteBuffers will be wrapped and apply the encryption overhead for each
- Include the TLS header length in the overhead computation
Result:
Fixes https://github.com/netty/netty/issues/6481
Motivation:
325cc84a2e introduced new tests which uses classes only provided by Java8+. We need to ensure we only try to load classes needed for these when we run the tests on Java8+ so we still can run the testsuite with Java7.
Modifications:
Add extra class which only gets loaded when Java8+ is used and move code there.
Result:
No more class-loader issue when running tests with Java7.
Motivation:
We not support all SSLParameters settings so we should better throw if a user try to use them.
Modifications:
- Check for unsupported parameters
- Add unit test
Result:
Less surprising behavior.
Motivation:
d8e6fbb9c3 attempted to account for the JDK not throwing the expected SSLHandshakeException by allowing a SSLException to also pass the test. However in some situations the SSLException will not be the top level exception and the Throwable must be unwrapped to see if the root cause is an SSLException.
Modifications:
- Unwrap exceptions thrown by the JDK's SSLEngine to check for SSLException.
Result:
SSLEngineTest (and derived classes) are more reliable.
Motivation:
OpenSSL doesn't automatically verify hostnames and requires extract method calls to enable this feature [1]. We should allow this to be configured.
Modifications:
- SSLParamaters#getEndpointIdentificationAlgorithm() should be respected and configured via tcnative interfaces.
Result:
OpenSslEngine respects hostname verification.
[1] https://wiki.openssl.org/index.php/Hostname_validation
Motivation:
We have our own ThreadLocalRandom implementation to support older JDKs . That said we should prefer the JDK provided when running on JDK >= 7
Modification:
Using ThreadLocalRandom implementation of the JDK when possible.
Result:
Make use of JDK implementations when possible.
Motivation:
Commit cd3bf3df58 made netty observe the latest version of netty-tcnative which changed the way how static fields are computed for various SSL.* values. This lead to have SSL_OP_NO_SSLv2 become 0 when using boringssl as boringssl not supports SSLv2 at all. In the logic of ReferenceCountedOpenSslEngine.getEnabledProtocols() we not expect to have a zero value and so our logic fails.
Modifications:
Check we actual support the protocol before return it as enabled.
Result:
SSLEngineTest.testEnablingAnAlreadyDisabledSslProtocol passes again with boringssl
Motivation:
JdkOpenSslEngineInteroptTest.mySetupMutualAuthServerIsValidClientException(...) delegated to the wrong super method.
Modifications:
Fix delegate
Result:
Correct test-code.
Motivation:
Some version of openssl dont support the needed APIs to use a KeyManagerFactory. In this case we should skip the tests.
Modifications:
- Use assumeTrue(...) to skip tests that need a KeyManagerFactory and its not supported.
Result:
Tests pass on all openssl versions we support.
Motivation:
tcnative was moved into an internal package.
Modifications:
Update package for tcnative imports.
Result:
Use correct package names for tcnative.
Motivation:
OpenSslEngineTest has unused imports and SSLEngineTest uses a fixed port which was used for debugging.
Modifications:
- Remove unused imports
- Use ephemeral port
Result:
Cleaner test code.
Modifications:
tcnative made some fixes and API changes related to setVerify. We should absorb these changes in Netty.
Modifications:
- Use tcnatives updated APIs
- Add unit tests to demonstrate correct behavior
Result:
Updated to latest tcnative code and more unit tests to verify expected behavior.
Motivation:
We should remove the restriction to only allow to call unwrap with a ByteBuffer[] whose cumulative length exceeds MAX_ENCRYPTED_PACKET_LENGTH.
Modifications:
Remove guard.
Result:
Fixes [#6335].
Motivation:
CipherSuiteConverter may throw a NPE if a cipher suite from OpenSSL does not match the precomputed regular expression for OpenSSL ciphers. This method shouldn't throw and instead just return null.
Modifications:
- if cacheFromOpenSsl(..) fails the conversion toJava should return null
Result:
Fixes https://github.com/netty/netty/issues/6336.
Motivation:
Currently Netty utilizes BIO_new_bio_pair so we can control all FD lifetime and event notification but delegates to OpenSSL for encryption/decryption. The current implementation sets up a pair of BIO buffers to read/write encrypted/plaintext data. This approach requires copying of data from Java ByteBuffers to native memory BIO buffers, and also requires both BIO buffers to be sufficiently large to hold application data. If direct ByteBuffers are used we can avoid coyping to/from the intermediate BIO buffer and just read/write directly from the direct ByteBuffer memory. We still need an internal buffer because OpenSSL may generate write data as a result of read calls (e.g. handshake, alerts, renegotiation, etc..), but this buffer doesn't have to be be large enough to hold application data.
Modifications:
- Take advantage of the new ByteBuffer based BIO provided by netty-tcnative instead of using BIO_read and BIO_write.
Result:
Less copying and lower memory footprint requirement per TLS connection.
Motivation:
We used various mocking frameworks. We should only use one...
Modifications:
Make usage of mocking framework consistent by only using Mockito.
Result:
Less dependencies and more consistent mocking usage.
Motiviation:
We should ensure we not need any extra wrap / unwrap calls during handshake once the handshake was signaled as finished
Modifications:
More strict testing
Result:
Better testing of handshake behaviour
Motivation:
Tests were added in 91f050d2ef to run with different protocols / ciphers. These may fail currently when openssl was compiled without support for the protocol / ciphers.
Modifications:
- Refactor tests to easier understand for which protocol / cipher it failed
- Not fail the test if the protocol is not supported with the used openssl version.
Result:
More robust testing.
Motivation:
We failed to properly test if a protocol is supported on an OpenSSL installation and just always returned all protocols.
Modifications:
- Detect which protocols are supported on a platform.
- Skip protocols in tests when not supported. This fixes a build error on some platforms introduced by [#6276].
Result:
Correctly return only the supported protocols
Motivation:
We used ca 2k as maximum overhead for encrypted packets which is a lot more then what is needed in reality by OpenSSL. This could lead to the need of more memory.
Modification:
- Use a lower overhead of 86 bytes as defined by the spec and openssl itself
- Fix unit test to use the correct session to calculate needed buffer size
Result:
Less memory usage.
Motivation:
We should test that we correctly return BUFFER_UNDERFLOW if the src buffer not contains enough data to unwrap it.
Modification:
Add unit test to verify behaviour.
Result:
Better test coverrage of SSLEngine implementations.
Motivation:
SslHandler closed the channel as soon as it was able to write out the close_notify message. This may not be what the user want as it may make sense to only close it after the actual response to the close_notify was received in order to guarantee a clean-shutdown of the connection in all cases.
Beside this closeNotifyFlushTimeoutMillis is volatile so may change between two reads. We need to cache it in a local variable to ensure it not change int between. Beside this we also need to check if the flush promise was complete the schedule timeout as this may happened but we were not able to cancel the timeout yet. Otherwise we will produce an missleading log message.
Modifications:
- Add new setter / getter to SslHandler which allows to specify the behavior (old behavior is preserved as default)
- Added unit test.
- Cache volatile closeNotifyTimeoutMillis.
- Correctly check if flush promise was complete before we try to forcibly close the Channel and log a warning.
- Add missing javadocs.
Result:
More clean shutdown of connection possible when using SSL and fix racy way of schedule close_notify flush timeouts and javadocs.
Motivation:
PR [#6238] added guards to be able to call wrap(...) / unwrap(...) after the engine was shutdown. Unfortunally one case was missed which is when closeOutbound() was called and produced some data while closeInbound() was not called yet.
Modifications:
Correctly guard against SSLException when closeOutbound() was called, produced data and someone calls wrap(...) after it.
Result:
No more SSLException. Fixes [#6260].