8234805: (dc) Remove JNI upcall from DatagramChannel.receive implementation

[Daniel Fuchs]

Hi Alan,

This looks like a nice improvement. A couple of comments:

DatagramChannelImpl.java:

   94     // Native buffer for socket address used by receive0
   95     private final NativeSocketAddress socketAddress;

It might be useful to add a comment to stress out that
*all* access to socketAddress must be performed while holding
the readLock.


  609             int n = receive(dst, connected);
  610             while (IOStatus.okayToRetry(n) && isOpen()) {

  642                 int n = receive(dst, connected);
  643                 while (n == IOStatus.UNAVAILABLE && isOpen()) {

I was wondering: why the difference (line 610 vs line 643)?

best regards,

-- daniel

On 26/11/2019 13:22, Alan Bateman wrote:
> This is a patch to replace the JNI upcall in the implementation of 
> DatagramChannel.receive so that the sender SocketAddress is created in 
> Java rather in the native method. This is the issue that was discussed 
> here last year [1]. The replacement marginally improves the single 
> source case, significantly improves the multiple source case.  I've 
> created an issue to add tests to the micro benchmark suite as this an 
> area that lacks benchmarks (Martin Thompson did send benchmarks in the 
> 2018 thread but we couldn't include them without the OCA).
> 
> The webrev with the changes are here. NativeSocketAddres will be a good 
> candidate to replace with the new foreign-memory access API if/when it 
> is promoted from incubating module. The test in the webrev is to add 
> variety to the source address scenarios.
> 
> http://cr.openjdk.java.net/~alanb/8234805/webrev/
> 
> -Alan
> 
> [1] https://mail.openjdk.java.net/pipermail/nio-dev/2018-August/005383.html
> [2] https://bugs.openjdk.java.net/browse/JDK-8234812