Motivation:
Highly retained and released objects have contention on their ref
count. Currently, the ref count is updated using compareAndSet
with care to make sure the count doesn't overflow, double free, or
revive the object.
Profiling has shown that a non trivial (~1%) of CPU time on gRPC
latency benchmarks is from the ref count updating.
Modification:
Rather than pessimistically assuming the ref count will be invalid,
optimistically update it assuming it will be. If the update was
wrong, then use the slow path to revert the change and throw an
execption. Most of the time, the ref counts are correct.
This changes from using compareAndSet to getAndAdd, which emits a
different CPU instruction on x86 (CMPXCHG to XADD). Because the
CPU knows it will modifiy the memory, it can avoid contention.
On a highly contended machine, this can be about 2x faster.
There is a downside to the new approach. The ref counters can
temporarily enter invalid states if over retained or over released.
The code does handle these overflow and underflow scenarios, but it
is possible that another concurrent access may push the failure to
a different location. For example:
Time 1 Thread 1: obj.retain(INT_MAX - 1)
Time 2 Thread 1: obj.retain(2)
Time 2 Thread 2: obj.retain(1)
Previously Thread 2 would always succeed and Thread 1 would always
fail on the second access. Now, thread 2 could fail while thread 1
is rolling back its change.
====
There are a few reasons why I think this is okay:
1. Buggy code is going to have bugs. An exception _is_ going to be
thrown. This just causes the other threads to notice the state
is messed up and stop early.
2. If high retention counts are a use case, then ref count should
be a long rather than an int.
3. The critical section is greatly reduced compared to the previous
version, so the likelihood of this happening is lower
4. On error, the code always rollsback the change atomically, so
there is no possibility of corruption.
Result:
Faster refcounting
```
BEFORE:
Benchmark (delay) Mode Cnt Score Error Units
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1 sample 2901361 804.579 ± 1.835 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10 sample 3038729 785.376 ± 16.471 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 100 sample 2899401 817.392 ± 6.668 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1000 sample 3650566 2077.700 ± 0.600 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10000 sample 3005467 19949.334 ± 4.243 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1 sample 456091 48.610 ± 1.162 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10 sample 732051 62.599 ± 0.815 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 100 sample 778925 228.629 ± 1.205 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1000 sample 633682 2002.987 ± 2.856 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10000 sample 506442 19735.345 ± 12.312 ns/op
AFTER:
Benchmark (delay) Mode Cnt Score Error Units
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1 sample 3761980 383.436 ± 1.315 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10 sample 3667304 474.429 ± 1.101 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 100 sample 3039374 479.267 ± 0.435 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1000 sample 3709210 2044.603 ± 0.989 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10000 sample 3011591 19904.227 ± 18.025 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1 sample 494975 52.269 ± 8.345 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10 sample 771094 62.290 ± 0.795 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 100 sample 763230 235.044 ± 1.552 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1000 sample 634037 2006.578 ± 3.574 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10000 sample 506284 19742.605 ± 13.729 ns/op
```
Motivation:
3c8c7fc7e9 introduced some changes to the ResourceLeakDetector that introduced a regression and so would always log that paranoid leak detection should be enabled even it was already.
Modifications:
Correctly not clear the recorded stacktraces when we process the reference queue so we can log these.
Result:
ResourceLeakDetector works again as expected.
Motivation:
NativeLibraryLoader uses ClassLoader#getResource method that can return nulls when the resource cannot be found. The returned url variable should be checked for nullity and fail in a more usable manner than a NullPointerException
Modifications:
Fail with a FileNotFoundException
Result:
Fixes [#7222].
Motiviation:
The ResourceLeakDetector helps to detect and troubleshoot resource leaks and is often used even in production enviroments with a low level. Because of this its import that we try to keep the overhead as low as overhead. Most of the times no leak is detected (as all is correctly handled) so we should keep the overhead for this case as low as possible.
Modifications:
- Only call getStackTrace() if a leak is reported as it is a very expensive native call. Also handle the filtering and creating of the String in a lazy fashion
- Remove the need to mantain a Queue to store the last access records
- Add benchmark
Result:
Huge decrease of performance overhead.
Before the patch:
Benchmark (recordTimes) Mode Cnt Score Error Units
ResourceLeakDetectorRecordBenchmark.record 8 thrpt 20 4358.367 ± 116.419 ops/s
ResourceLeakDetectorRecordBenchmark.record 16 thrpt 20 2306.027 ± 55.044 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 8 thrpt 20 4220.979 ± 114.046 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 16 thrpt 20 2250.734 ± 55.352 ops/s
With this patch:
Benchmark (recordTimes) Mode Cnt Score Error Units
ResourceLeakDetectorRecordBenchmark.record 8 thrpt 20 71398.957 ± 2695.925 ops/s
ResourceLeakDetectorRecordBenchmark.record 16 thrpt 20 38643.963 ± 1446.694 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 8 thrpt 20 71677.882 ± 2923.622 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 16 thrpt 20 38660.176 ± 1467.732 ops/s
Motivation:
RLD allocates an ArrayDeque in anticipation of recording access
points. If the leak detection level is less than ADVANCED though,
the dequeue is never used. Since SIMPLE is the default level,
there is a minor perf win to not preemptively allocate it.
This showed up in garbage profiling when creation a high number of
buffers.
Modifications:
Only allocate the dequeue if it will be used.
Result:
Less garbage created.
Motivation:
NativeLibraryLoader may only log a debug statement if the library is successfully loaded from java.library.path, but will log failure statements the if load for java.library.path fails which can mislead users to believe the load actually failed when it may have succeeded.
Modifications:
- Always load a debug statement when a library was successfully loaded
Result:
NativeLibraryLoader log statements more clear.
Motivation:
When Log4j2Logger is used with PatternLayout (%F:%L)%c.%M, the log message incorrect shows:
(Log4J2Logger.java:73)io.netty.util.internal.PlatformDependent0.debug ....
Modification:
Extend AbstractLogger
Result:
Fixes [#7186].
Motivation:
We should only try to load the native artifacts if the architecture we are currently running on is the same as the one the native libraries were compiled for.
Modifications:
Include architecture in native lib name and append the current arch when trying to load these. This will fail then if its not the same as the arch of the compiled arch.
Result:
Fixes [#7150].
Motivation:
When the user want to have the direct memory explicitly managed by the GC (just as java.nio does) it is useful to be able to construct an UnpooledByteBufAllocator that allows this without the chances to see any memory leak.
Modifications:
Allow to explicitly disable the usage of reflection to construct direct ByteBufs and so be sure these will be collected by GC.
Result:
More flexible way to use the UnpooledByteBufAllocator.
Motivation:
The documentation for field updates says:
> Note that the guarantees of the {@code compareAndSet}
> method in this class are weaker than in other atomic classes.
> Because this class cannot ensure that all uses of the field
> are appropriate for purposes of atomic access, it can
> guarantee atomicity only with respect to other invocations of
> {@code compareAndSet} and {@code set} on the same updater.
This implies that volatiles shouldn't use normal assignment; the
updater should set them.
Modifications:
Use setter for field updaters that make use of compareAndSet.
Result:
Concurrency compliant code
Motivation:
PD and PD0 Both try to find and use Unsafe. If unavailable, they
try to log why and continue on. However, it is not always east to
enable this logging. Chaining exceptions together is much easier
to reach, and the original exception is relevant when Unsafe is
needed.
Modifications:
* Make PD log why PD0 could not be loaded with a trace level log
* Make PD0 remember why Unsafe wasn't available
* Expose unavailability cause through PD for higher level use.
* Make Epoll and KQueue include the reason when failing
Result:
Easier debugging in hard to reconfigure environments
Motivation:
Continuing to make netty happy when compiling through errorprone.
Modification:
Mostly comments, some minor switch statement changes.
Result:
No more compiler errors!
Motivation:
On restricted systems (e.g. grsecurity), it might not be possible to write a .so on disk and load it afterwards. On those system Netty should check java.library.path for libraries to load.
Modifications:
Changed NativeLibraryLoader.java to first try to load libs from java.library.path before exporting the .so to disk.
Result:
Libraries load fine on restricted systems.
Motivation:
When compiling this code and running it through errorprone[1], this message appears:
```
StringUtil.java:493: error: [FallThrough] Switch case may fall through; add a `// fall through` comment if it was deliberate
case LINE_FEED:
^
(see http://errorprone.info/bugpattern/FallThrough)
```
By adding that comment, it silences the error and also makes clear the intention of that statement.
[1]http://errorprone.info/index
Modification:
Add simple comment.
Result:
Errorprone is happier with the code.
Motivation:
The `AsciiString#toString` method calculate string value and cache it into field. If an `AsciiString` created from the `String` value, we can avoid rebuilding strings if we cache them immediately when creating `AsciiString`. It would be useful for constants strings, which already stored in the JVMs string table, or in cases where an unavoidable `#toString `method call is assumed.
Modifications:
- Add new static method `AsciiString#cache(String)` which save string value into cache field.
- Apply a "benign" data race in the `#hashCode` and `#toString` methods.
Result:
Less memory usage in some `AsciiString` use cases.
Motivation:
At the moment we try to load the library using multiple names which includes names using - but also _ . We should just use _ all the time.
Modifications:
Replace - with _
Result:
Fixes [#7069]
Motivation:
In some cases of using an `InternalThreadLocalMap#stringBuilder`, the `StringBuilder`s size can often exceed the exist limit (1024 bytes). This can lead to permanent memory reallocation.
Modifications:
Add custom properties for the initial capacity and maximum size (after which the `StringBuilder`s capacity will be reduced to the initial capacity).
Result:
An `InternalThreadLocalMap#stringBuilder`s initial and max size is configurable. Fixes [#7092].
Motivation:
NativeLibraryLoader has some code-duplication that can be removed.
Modifications:
Remove duplicated code and just use provided methods of PlatformDependent.
Result:
Less code duplication, fixes [#3756].
Motivation:
Now that the NativeLibraryLoader implicitly detects the shaded package prefix we no longer need the io.netty.packagePrefix system property.
Modifications:
- Remove io.netty.packagePrefix processing from NativeLibraryLoader
Result:
Code is cleaner.
Motivation:
Shading requires renaming binary components (.so, .dll; for tcnative,
epoll, etc). But the rename then requires setting the
io.netty.packagePrefix system property on the command line or runtime,
which is either a burden or not feasible.
If you don't rename the binary components everything appears to
work, until a dependency on a second version of the binary component is
added. At that point, only one version of the binary will be loaded...
which is what shading is supposed to prevent. So for valid shading, the
binaries must be renamed.
Modifications:
Automatically detect the package prefix by comparing the actual class
name to the non-shaded expected class name. The expected class name must
be obfuscated to prevent shading utilities from changing it.
Result:
When shading and using binary components, runtime configuration is no
longer necessary.
Pre-existing shading users that were not renaming the binary components
will break, because the packagePrefix previously defaulted to "". Since
these pre-existing users had broken configurations that only _appeared_
to work, this breakage is considered a Good Thing. Users may workaround
this breakage temporarily by setting -Dio.netty.packagePrefix= to
restore packagePrefix to "".
Fixes#6963
Motivation:
ResourceLeakDetector records at most MAX_RECORDS+1 records
Modifications:
Make room before add to lastRecords
Result:
ResourceLeakDetector will record at most MAX_RECORDS records
Motivation:
JCTools 2.0.2 provides an unbounded MPSC linked queue. Before we shaded JCTools we had our own unbounded MPSC linked queue and used it in various places but gave this up because there was no public equivalent available in JCTools at the time.
Modifications:
- Use JCTool's MPSC linked queue when no upper bound is specified
Result:
Fixes https://github.com/netty/netty/issues/5951
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:
A `StringUtil#escapeCsv` creates new `StringBuilder` on each value even if the same string is returned in the end.
Modifications:
Create new `StringBuilder` only if it really needed. Otherwise, return the original string (or just trimmed substring).
Result:
Less GC load. Up to 4x faster work for not changed strings.
Motivation:
AppendableCharSequence depends upon IndexOutOfBoundsException to trigger a resize operation under the assumption that the resize operation will be rare if the initial size guess is good. However if the initial size guess is not good then the performance will be more unpredictable and likely suffer.
Modifications:
- Check the position in AppendableCharSequence#append to determine if a resize is necessary
Result:
More predictable performance in AppendableCharSequence#append.
Motivation:
We need to use FQCN to prevent classloader issues for classes that are > Java6. This is a cleanup of ed5fcbb773.
Modifications:
Just remove the imports and use FQCN.
Result:
No classloader issues with java6
Motivation:
Docker's `--tmpfs` flag mounts the temp volume with `noexec` by default,
resulting in an UnsatisfiedLinkError. While this is good security
practice, it is a surprising failure from a seemingly innocuous flag.
Modifications:
Add a best-effort attempt in `NativeLibraryLoader` to detect when temp
files beng loaded cannot be executed even when execution permissions
are set, often because the `noexec` flag is set on the volume.
Requires numerous additional exclusions to the Animal Sniffer config
for Java7 POSIX permissions manipulation.
Result:
Fixes [#6678].
Motivation:
For our native libraries in netty we support shading, to have this work on runtime the user needs to set a system property. This code should shared.
Modifications:
Move logic to NativeLbiraryLoader and so share for all native libs.
Result:
Less code duplication and also will work for netty-tcnative out of the box once it support shading
Motivation:
`FormattingTuple.getArgArray()` is never used.
In the `MessageFormatter` it is possible to make
some improvements, e.g. replace `StringBuffer`
with `StringBuilder`, avoid redundant allocations, etc.
Modifications:
- Remove `argArray` field from the `FormattingTuple`.
- In `MessageFormatter`:
- replace `StringBuffer` with `StringBuilder`,
- replace `HashMap` with `HashSet` and make it lazy initialized.
- avoid redundant allocations (`substring()`, etc.)
- use appropriate StringBuilder's methods for the some `Number` values.
- Porting unit tests from `slf4j`.
Result:
Less GC load on logging with internal `MessageFormatter`.
Motivation:
IPv4/6 validation methods use allocations, which can be avoided.
IPv4 parse method use StringTokenizer.
Modifications:
Rewriting IPv4/6 validation methods to avoid allocations.
Rewriting IPv4 parse method without use StringTokenizer.
Result:
IPv4/6 validation and IPv4 parsing faster up to 2-10x.
Motivation:
We should only try to load jdk.internal.misc.Unsafe if we run on Java9+ to eliminate noise in the log.
Modifications:
- Move javaVersion() and related methods to PlatformDependent0 to be able to use these in the static initializer without creating a cycle.
- Only try to load jdk.internal.misc.Unsafe when running in Java9+
Result:
Less noise in the log when running pre java9.
Motivation:
A previous change allocated a new thread local string builder if it
was getting too large. This is a good change, these string builders
can accidentally get too large and then never shrunk and that is sort
of a memory leak. However, the change allocates an entirely new string
builder which is more allocations than necessary. Instead, we can trim
the string builder if its too large, this only allocates an extra
backing array instead of a whole new object.
Modifications:
If the string builder is above a threshold, we trim the string builder
and then ensure its capacity is reasonable to we do not allocate too
much as we start using the string builder.
Result:
The thread local string builder do not serve as a memory yet we do not
allocate too many new objects.
Motivation:
`NetUtil`'s methods `isValidIpV6Address` and `getIPv6ByName` incorrectly validate some IPv6 addresses.
Modifications:
- `getIPv6ByName`: add checks for single colon at the start or end.
- `isValidIpV6Address`: fix checks for the count of colons and use `endOffset` instead of `ipAddress.length()` for the cases with the brackets or '%'.
Result:
More correct implementation of `NetUtil#isValidIpV6Address` and `NetUtil#getIPv6ByName`.
Motivation:
AtomicIntegerFieldUpdater#get is unnecessary, I think use simple volatile read is cleaner
Modifications:
Replace code STATE_UPDATER.get(this) to state in SingleThreadEventExecutor
Result:
Cleaner code
Motivation:
InternalThreadLocalMap#stringBuilder: ensure memory overhead
Modification:
If the capacity of StringBuilder is greater than 65536 then release it on the next time you get StringBuilder and re-create a StringBuilder.
Result:
Possible less memory usage.
