RFR: 8312166: (dc) DatagramChannel's socket adaptor does not release carrier thread when blocking in receive

[Jaikiran Pai]

On Mon, 17 Jul 2023 10:10:20 GMT, Alan Bateman <alanb at openjdk.org> wrote:

> DatagramChannel is "virtual thread friendly"; when configured blocking, the receive/read methods release the carrier when there is no datagram to receive.
> 
> DatagramChannel's socket adaptor (used by DatagramSocket) is not currently virtual thread friendly. The original changes proposed to address this in JDK 19 interacted with the carrier thread's buffer cache so we decided to not include them at the time. The change proposed here are to address this issue, mostly by moving the implementation of the socket adaptor's send/receive methods into the DatagramChannel implementation so that a temporary direct buffer is not held when parked.
> 
> The only observable change should be to silly/broken code that changes a DatagramPacket's buffer (with setData or setLength) while a receive is in progress. Long standing, and undocumented, behavior would for setXXX to block while a receive is in progress, new behavior will be only access the DatagramPacket's buffer when a datagram is received. I can't think of anything that could rely on behavior like this.

On a general note and more out of curiosity, I see that the `DatagramChannelImpl` dishonours the `blocking` mode set by the caller and forces non-blocking when virtual thread is involved and yet at the same time fulfills the contract of `receive()` and `send()` and even `isBlocking()`.

Is there a reason we couldn't do the same for platform threads and dishonour the blocking mode and always use non-blocking in these `send()` and `receive()` implementations?

-------------

PR Comment: https://git.openjdk.org/jdk/pull/14901#issuecomment-1640270637