Investigate io_uring submission queue's enqueueSqe performance

[franz1981]

#14962 has introduced a safe method (using NIO ByteBuffer) to interact with io_uring queues, but as usual, this doesn't come for free: it would be good to verify that performance is not impacted and/or mitigate it, in case.

The completion's queue array decoding seems to read way less fields, but should require the same analysis.

This work has been motivated by https://github.com/openjdk/loom/blob/c25b25646e23bccd04265ee5aaa907d0945f323d/src/java.base/linux/classes/jdk/internal/ffi/generated/iouring/io_uring_sqe.java and https://github.com/openjdk/loom/blob/c25b25646e23bccd04265ee5aaa907d0945f323d/src/java.base/linux/classes/sun/nio/ch/iouring/IOUring.java#L570-L609 which:

• uses sliced MemorySegments
• zeoed

The last one is something I'm happy we don't do, but fixed offset accesses on trusted memory segment (i.e. with base offset, memory address and endiannes constant) could give some performance advantage over NIO ByteBuffer,

In #15486 we've evaluated that VarHandles byte buffer view to be very performant, but we lack to verify how it fear compared to MemorySegment in this scenario (indexed accesses) - and if it matters, in the grand scheme.

[franz1981]

@dreamlike-ocean With all the io_uring features enabled IDK how much we expect to interact with the submission queue, so I would like to know your opinion and if you have any data/hints to share on this 🙏

[franz1981]

@AlanBateman and @mcimadamore too: this issue is just another way of validating the benefits of using sliced MemorySegments over VarHandles byteBufferView vs NIO ByteBuffer which I remember:

1. there are tons of improvements on Hotspot
2. existing JMH benchmark in the JDK repo (which I have to search for, yet)

[normanmaurer]

I mean yeah it would be nice if we can improve that said I am not overly concerned atm.

[franz1981]

With all the features of io_uring enabled is probably a drop in the ocean as I won't expect that much sqe submission activity, too. But I have never measured it to be sure of it

[franz1981]

The other nice part is if we can get these classes to be autogenerated as the loom branch @normanmaurer

[dreamlike-ocean]

@dreamlike-ocean With all the io_uring features enabled IDK how much we expect to interact with the submission queue, so I would like to know your opinion and if you have any data/hints to share on this 🙏

We don’t have hard data to prove it, but in scenarios running on Linux 6.1 where most connections are in auto-read mode, once a TCP connection is established, only write operations generally need to interact with the SQ. Reading/writing SQEs themselves is not really a performance-sensitive point.

https://inside.java/2025/06/12/ffm-vs-unsafe/
Moreover, the random access performance of MemorySegment is somewhat lower compared to Unsafe, so some small tricks may be needed to reach the desired performance level.