Solaris can't set socket options about jnr-unixsocket

[headius]

It seems like we may be using a bad value for one of the parameters to setsockopt within this code:

jnr-unixsocket/src/main/java/jnr/unixsocket/Common.java

Lines 147 to 149 in fe79d02

	if (0 != Native.setsockopt(fd, SocketLevel.SOL_SOCKET, optname, optvalue)) {
	throw new IOException(Native.getLastErrorString());
	}

The UnixSocketOption value for SO_RCVTIMEO used for the setSoTimeout is mapped to a value from jnr-constants at this line of code:

jnr-unixsocket/src/main/java/jnr/unixsocket/Common.java

Line 157 in fe79d02

wMap.put(UnixSocketOptions.SO_RCVTIMEO, jnr.constants.platform.SocketOption.SO_RCVTIMEO);

So at this point I would suspect that the value defined there is not correct for your flavor of Solaris.

I will need some help investigating this! The values in jnr-constants (linked above) are generated from the actual C constants on a given platform. If you look here you can see the values we are currently using for all Solaris flavors. Can you look into the C socket headers on your system and see if these values match?

If they don't we may finally be forced to split these constants by hardware platform as well as by OS.

The only workaround at this point would be for you to reflectively dig out the file descriptor and call setsockopt yourself with the proper values, either by binding setsockopt on your own (using jnr-ffi, the native binding library behind jnr-unixsocket) or by using reflection to dig into the function already bound by jnr-unixsocket here.

Alternatively, if you are able to give me shell access to a matching environment, I can do the investigation myself.

[wolfchimneyrock]

the constants are the same on my sys/socket.h as in your jnr-constants package:

 #define SO_RCVTIMEO 0x1006      /* receive timeout */

[wolfchimneyrock]

the workaround I have done that works is to switch to non-blocking on the client channel, it is a little more complicated to use but it allows the timeout to be used in the select() call.

Unfortunately we can't give out shell access. Though I would be happy to do an attempt at setsockopt() using reflection

[wolfchimneyrock]

When I run the following:

        UnixSocketChannel channel = UnixSocketChannel.create();

        try {
            Class<?> nativeClass = Class.forName("jnr.unixsocket.Native");
            Method setsockopt = nativeClass.getDeclaredMethod("setsockopt", int.class, SocketLevel.class, SocketOption.class, int.class);
            setsockopt.setAccessible(true);
            Integer result = (Integer)setsockopt.invoke(null, channel.getFD(), SocketLevel.SOL_SOCKET, SocketOption.SO_RCVTIMEO, 10);
            System.out.println("setsockopt result = " + result);
        } catch (Exception e) {
            e.printStackTrace();
        }

on linux/amd64: setsockopt result = 0
on AIX/power9: setsockopt result = 0
on Solaris 11.3/sparc: setsockopt result = -1

Can you walk me through how I could make custom values of the SocketLevel and SocketOption ?

[headius]

Perhaps the mismatch is in SocketLevel?

https://github.com/jnr/jnr-constants/blob/master/src/main/java/jnr/constants/platform/solaris/SocketLevel.java

It seems that the setsockopt function is being called, but we are getting this error. Another thing you could investigate would be to add code to your println to check the errno value. I suspect it will be EINVAL, though.

System.out.println("errno: " + LastError.getLastError(Runtime.getSystemRuntime()));

(untested code but you can see the equivalent here:

jnr-unixsocket/src/main/java/jnr/unixsocket/Native.java

Line 123 in aa2c1e4

return strerror(LastError.getLastError(Runtime.getSystemRuntime()));

[wolfchimneyrock]

errno = 99 => ENOPROTOOPT

[wolfchimneyrock]

If I do the same thing but set SO_RCVBUF it works fine

[wolfchimneyrock]

I notice the man page for setsockopt on Solaris 11.3 doesn't mention SO_RCVTIMEO:
https://docs.oracle.com/cd/E86824_01/html/E54774/setsockopt-3socket.html#scrolltoc

whereas the man page for setsockopt on Solaris 11.4 does mention it:
https://docs.oracle.com/cd/E88353_01/html/E37843/setsockopt-3c.html#scrolltoc

[headius]

So then it is possible this socket option is simply not supported? Maybe there is a different option that has a similar effect?

[wolfchimneyrock]

yeah I think thats the case - can't set a timeout for blocking socket read. Like I said earlier, the work around is to make the channel non-blocking and use a Selector.

[headius]

Okay, that is a non-trivial workaround but I'm glad you have something to go forward with. If you are able to determine that there is some other way we can adjust this timeout on earlier versions of solaris, please open a new issue. I will close this issue since there's not much we can do at the moment.