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:
There are numerous usages of internalNioBuffer which hard code 0 for the index when the intention was to use the readerIndex().
Modifications:
- Remove hard coded 0 for the index and use readerIndex()
Result:
We are less susceptible to using the wrong index, and don't make assumptions about the ByteBufAllocator.
Motivation:
ReferenceCountedOpenSslEngine#wrap must have a direct buffer for a destination to interact with JNI. If the user doesn't supply a direct buffer we internally allocate one to write the results of wrap into. After this operation completes we copy the contents of the direct buffer into the heap buffer and use internalNioBuffer to get the content. However we pass in the end index but the internalNioBuffer expects a length.
Modifications:
- pass the length instead of end index to internalNioBuffer
Result:
ReferenceCountedOpenSslEngine#wrap will copy the correct amount of data into the destination buffer when heap buffers are wrapped.
Motivation:
SslContextBuilder sill state the KeyManagerFactory and TrustManagerFactory are only supported when SslProvider.JDK is used. This is not correct anymore.
Modifications:
Fix javadocs.
Result:
Correct javadocs.
Motivation:
SslContext#newHandler currently creates underlying SSLEngine without
enabling HTTPS endpointIdentificationAlgorithm. This behavior in
unsecured when used on the client side.
We can’t harden the behavior for now, as it would break existing
behavior, for example tests using self signed certificates.
Proper hardening will happen in a future major version when we can
break behavior.
Modifications:
Add javadoc warnings with code snippets.
Result:
Existing unsafe behavior and workaround documented.
Motivation:
Normally if a decoder produces an exception its wrapped with DecodingException. This is not the cause for NotSslRecordException in SslHandler and SniHandler.
Modifications:
Just throw the NotSslRecordException exception for decode(...) and so ensure its correctly wrapped in a DecodingException before its passed through the pipeline.
Result:
Consist behavior.
Motivation:
To ensure that all bytes queued in OpenSSL/tcnative internal buffers we invoke SSL_shutdown again to stimulate OpenSSL to write any pending bytes. If this call fails we may call SSL_free and the associated shutdown method to free resources. At this time we may attempt to use the networkBIO which has already been freed and get a NPE.
Modifications:
- Don't call bioLengthByteBuffer(networkBIO) if we have called shutdown() in ReferenceCountedOpenSslEngine
Result:
Fixes https://github.com/netty/netty/issues/6466
Motivation:
Realization of `AbstractTrafficShapingHandler.userDefinedWritabilityIndex()` has references to subclasses.
In addition, one of the subclasses overriding it, but the other does not.
Modifications:
Add overriding to the second subclass. Remove references to subclasses from parent class.
Result:
More consistent and clean code (OOP-stylish).
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:
As netty-tcnative can be build against different native libraries and versions we should log the used one.
Modifications:
Log the used native library after netty-tcnative was loaded.
Result:
Easier to understand what native SSL library was used.
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:
ThreadLocalInsecureRandom still referenced ThreadLocalRandom directly, but shouldn't.
Modifications:
ThreadLocalInsecureRandom should reference PlatformDependent#threadLocalRandom() in comments
Result:
Less usage of internal.ThreadLocalRandom.
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:
If an event occurs which generates non-application data (shutdown, handshake failure, alert generation, etc...) and the non-application buffer in the ByteBuffer BIO is full (or sufficiently small) we may not propagate all data to our peer before tearing down the socket.
Modifications:
- when wrap() detects the outbound is closed, but there is more data pending in the non-application buffers, we must also check if OpenSSL will generate more data from calling SSL_shutdown again
- when wrap() detects a handshakeExcpetion during failure we should check if OpenSSL has any pending data (in addition to the non-application buff) before throwing the handshake exception
Result:
OpenSslEngine more reliably transmits data to the peer before closing the socket.
Motivation:
tcnative was moved into an internal package.
Modifications:
Update package for tcnative imports.
Result:
Use correct package names for tcnative.
Motivation:
If the OpenSslEngine has bytes pending in the non-application buffer and also generates wrapped data during the handshake then the handshake data will be missed. This will lead to a handshake stall and eventually timeout. This can occur if the non-application buffer becomes full due to a large certificate/hello message.
Modification:
- ReferenceCountedOpenSslEngine should not assume if no data is flushed from the non-application buffer that no data will be produced by the handshake.
Result:
New unit tests with larger certificate chains don't fail.
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:
tcnative has updated how constants are defined and removed some constants which are either obsolete or now set directly in tcnative.
Modifications:
- update to compile against tcnative changes.
Result:
Netty compiles with tcnative options changes.
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:
There were some warnings for the code in the ssl package.
Modifications:
- Remove not needed else blocks
- Use correctly base class for static usage
- Replace String.length() == 0 with String.isEmpty()
- Remove unused code
Result:
Less warnings and cleaner code.
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:
The SSLEngine wrap and unwrap methods can be called in a way that has no side effects, but this could involve costly validation and allocation. The SslHandler should avoid calling into these methods if possible.
Modifications:
- wrapNonAppData should provide additional status which can be used by wrap to breakout early if possible
Result:
SslHandler invokes the SSLEngine less.
Motivation:
Previous versions of netty-tcnative used the org.apache.tomcat namespace which could lead to problems when a user tried to use tomcat and netty in the same app.
Modifications:
Use netty-tcnative which now uses a different namespace and adjust code to some API changes.
Result:
Its now possible to use netty-tcnative even when running together with tomcat.
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:
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].
Motivation:
SslHandler has multiple methods which have better replacements now or are obsolete. We should mark these as `@Deprecated`.
Modifications:
Mark methods as deprecated.
Result:
API cleanup preparation.
Motivation:
In commit fc3c9c9523 I changes the way how we calculate the capacity of the needed ByteBuf for wrap operations that happen during writes when the SslHandler is used. This had the effect that the same capacity for ByteBufs is needed for the JDK implementation of SSLEngine but also for our SSLEngine implementation that uses OpenSSL / BoringSSL / LibreSSL. Unfortunally this had the side-effect that applications that used our SSLEngine implementation now need a lot more memory as bascially the JDK implementation always needs a 16kb buffer for each wrap while we can do a lot better for our SSLEngine implementation.
Modification:
- Resurrect code that calculate a better ByteBuf capacity when using our SSLEngine implementation and so be able to safe a lot of memory
- Add test-case to ensure it works as expected and is not removed again later on.
Result:
Memory footprint of applications that uses our SSLEngine implementation based on OpenSSL / BoringSSL / LibreSSL is back to the same amount of before commit fc3c9c9523.
Motivation:
Currently Netty does not wrap socket connect, bind, or accept
operations in doPrivileged blocks. Nor does it wrap cases where a dns
lookup might happen.
This prevents an application utilizing the SecurityManager from
isolating SocketPermissions to Netty.
Modifications:
I have introduced a class (SocketUtils) that wraps operations
requiring SocketPermissions in doPrivileged blocks.
Result:
A user of Netty can grant SocketPermissions explicitly to the Netty
jar, without granting it to the rest of their application.
Motivation:
For the completion of a handshake we already fire a SslHandshakeCompletionEvent which the user can intercept. We should do the same for the receiving of close_notify.
Modifications:
Add SslCloseCompletionEvent and test-case.
Result:
More consistent API.
Motivation:
https://github.com/netty/netty/pull/6042 only addressed PlatformDependent#getSystemClassLoader but getClassLoader is also called in an optional manner in some common code paths but fails to catch a general enough exception to continue working.
Modifications:
- Calls to getClassLoader which can continue if results fail should catch Throwable
Result:
More resilient code in the presense of restrictive class loaders.
Fixes https://github.com/netty/netty/issues/6246.
Motivation:
The SslHandler.sslCloseFuture() may not be notified when the Channel is closed before a closify_notify is received.
Modifications:
Ensure we try to fail the sslCloseFuture() when the Channel is closed.
Result:
Correctly notify the ssl close future.
Motivation:
The JDK implementation of SSLEngine allows to have unwrap(...) / wrap(...) called even after closeInbound() and closeOutbound() were called. We need to support the same in ReferenceCountedSslEngine.
Modification:
- Allow calling ReferenceCountedSslEngine.unwrap(...) / wrap(...) after the engine was closed
- Modify unit test to ensure correct behaviour.
Result:
Implementation works as expected.
Motivation:
fc3c9c9523 introduced a bug which will have ReferenceCountedSslEngine.unwrap(...) produce an IOOBE when be called with an BŷteBuffer as src that contains multiple SSLRecords and has a position != 0.
Modification:
- Correctly set the limit on the ByteBuffer and so fix the IOOBE.
- Add test-case to verify the fix
Result:
Correctly handle heap buffers as well.
Motivation:
Openssl provider should behave same as JDK provider when mutual authentication is required and a specific set of trusted Certificate Authorities are specified. The SSL handshake should return back to the connected peer the same list of configured Certificate Authorities.
Modifications:
Correctly set the CA list.
Result:
Correct and same behaviour as the JDK implementation.
Motivation:
We need to ensure we not swallow the close_notify that should be send back to the remote peer. See [#6167]
Modifications:
- Only call shutdown() in closeInbound() if there is nothing pending that should be send back to the remote peer.
- Return the correct HandshakeStatus when the close_notify was received.
- Only shutdown() when close_notify was received after closeOutbound() was called.
Result:
close_notify is correctly send back to the remote peer and handled when received.
Motivation
The IdleStateHandler tracks write() idleness on message granularity but does not take into consideration that the client may be just slow and has managed to consume a subset of the message's bytes in the configured period of time.
Modifications
Adding an optional configuration parameter to IdleStateHandler which tells it to observe ChannelOutboundBuffer's state.
Result
Fixes https://github.com/netty/netty/issues/6150
Motivation:
Our ReferenceCountedOpenSslEngine does not support compression so we should explicit disable it.
This is related to #3722.
Modifications:
Set SSL_OP_NO_COMPRESSION option.
Result:
Not use compression.
Motivation:
In later Java8 versions our Atomic*FieldUpdater are slower then the JDK implementations so we should not use ours anymore. Even worse the JDK implementations provide for example an optimized version of addAndGet(...) using intrinsics which makes it a lot faster for this use-case.
Modifications:
- Remove methods that return our own Atomic*FieldUpdaters.
- Use the JDK implementations everywhere.
Result:
Faster code.
Motivation:
We need to ensure we handle the case when BUFFER_OVERFLOW happens during unwrap but the readable bytes are bigger then the expected applicationBufferSize. Otherwise we may produce an IllegalArgumentException as we will try to allocate a buffer with capacity < 0.
Modifications:
- Guard against this case.
- Ensure we not double release buffer on exception when doing unwrap.
Result:
No more exception when running testsuite with java 9.
Motivation:
We need to ensure the tracked object can not be GC'ed before ResourceLeak.close() is called as otherwise we may get false-positives reported by the ResourceLeakDetector. This can happen as the JIT / GC may be able to figure out that we do not need the tracked object anymore and so already enqueue it for collection before we actually get a chance to close the enclosing ResourceLeak.
Modifications:
- Add ResourceLeakTracker and deprecate the old ResourceLeak
- Fix some javadocs to correctly release buffers.
- Add a unit test for ResourceLeakDetector that shows that ResourceLeakTracker has not the problems.
Result:
No more false-positives reported by ResourceLeakDetector when ResourceLeakDetector.track(...) is used.
