Suggestion needed to port the fix to JDK17 and JDK11

[Shruthi .]

Hi Alan,

Reordering preClose() in AIX resolves the customer issue. We have validated the fix, and the customer has confirmed it as well.

Thanks
Shruthi

________________________________
From: Alan Bateman <alan.bateman at oracle.com>
Sent: Thursday, February 13, 2025 10:20 PM
To: Shruthi . <Shruthi.Shruthi1 at ibm.com>; net-dev at openjdk.org <net-dev at openjdk.org>
Cc: Syed Moinudeen <smoinud1 at in.ibm.com>
Subject: [EXTERNAL] Re: Suggestion needed to port the fix to JDK17 and JDK11

On 12/02/2025 06: 13, Shruthi . wrote: We discussed the issue with AIX OS team to get their input as this issue is happening only in AIX platform. Update from AIX OS dev team: A thread gets stuck in dup2() while duplicating a file descriptor

On 12/02/2025 06:13, Shruthi . wrote:
We discussed the issue with AIX OS team to get their input as this issue is happening only in AIX platform. Update from AIX OS dev team: A thread gets stuck in dup2() while duplicating a file descriptor that is already being used by two other threads (one for reading, one for writing). The writer thread does not write or close the FIFO, causing the reader to wait indefinitely, which blocks dup2(). Also he is saying that we cannot change OS side as it may break or change its existing behavior which other existing applications may be depending on.

Can we apply the previously proposed fix at java level(reordering the preClose() method) while adding an AIX-specific check to ensure it doesn’t affect other platforms? Or do you have any alternative suggestions?

            if (!tryClose()) {
                long th = thread;
                if (th != 0) {
                    if (!AIX.isAIX)
                        nd.preClose(fd);
                    NativeThread.signal(th);
                    if (AIX.isAIX)
                        nd.preClose(fd);
                }
            }

Are you sure this really fixes the issue on AIX? It means a reader or writer may be signalled before the syscall, is the signal missed or are you saying that it queues up on AIX? If it missed then the close will block in dup2 as before.

-Alan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/net-dev/attachments/20250214/e10d7ba7/attachment-0001.htm>