Code Review Request : 7011591 JDWP socket transport should restart interrupted system calls (EINTR)

[Dmitry Samersoff]

David,

To be on safe side with connect() (on Linux, not sure for Solaris)
you need to call connect() again and if it returns EINPROGRESS, run 
polling loop. After poll() it's recommended to call getsockopt() and 
read the SO_ERROR to determine whether connect was successful or not.

-Dmitry


On 2011-08-04 16:02, David Buck wrote:
> Hi!
>
> I would like to request that my fix for 7011591 be reviewed for push
> into JDK8 (and JDK7u if possible).
>
> CR: 7011591 JDWP socket transport should restart interrupted system
> calls (EINTR)
>
> Weblink: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=2205030
> (for some reason, the public URL for the parent bug is not working. (I
> believe that SubCR 2205030 will have all needed information.)
>
> Webrev: http://cr.openjdk.java.net/~dbuck/7011591/webrev.00/
>
> Description:
> Syscalls that can be interrupted by signal handling should automatically
> retry if they fail with EINTR. Unfortunately, the JDWP socket transport
> implementation does not do this. Since HotSpot does not use signals for
> thread suspension, HotSpot customers have not run into any problem (as
> far as I know). But because JRockit R27 and earlier releases use signals
> to suspend threads, all threads are constantly being interrupted thus
> JRockit customers are much more likely (almost guaranteed under some
> circumstances) to have system calls interrupted by signal handling.
>
> I have added a loop to automatically retry send(), recv(), sendto(),
> recvfrom(), accept(), and poll() calls if the call fails with EINTR.
> Note that I do not believe that the UDP versions of the calls
> (recvfrom() and sendto()) are actualy ever used in the JDK, but I fixed
> them in case they are used in the future.
>
> The implementation for connect() was more complicated because Solaris
> does not let us just call connect() again after EINTR. Once a call to
> connect() fails with EINTR on Solaris, the connection setup continues
> asynchronously and we need to use select() or poll() to determine when
> the socket can be used.
>
> The patch for jdk7u2 is identical to this (jdk8) fix, if possible, I
> would like to receive pre-approval to push to jdk7u-dev repository as well.
>
> This fix has already been done for 1.4.2/5/6 as part of the JRockit
> Simplification Project. This needs to be forward ported into 7 and 8 to
> avoid a regression.
>
> Best Regards,
> -Buck


-- 
Dmitry Samersoff
Java Hotspot development team, SPB04
* There will come soft rains ...