Add support for Unix domain sockets

[astei]

Motivation:

The io_uring transport does not currently support Unix domain sockets. The epoll transport does, though. There should be parity between the two.

Modifications:

• Add IOUringDomainSocketChannel, IOUringDomainSocketChannelConfig, and IOUringServerDomainSocketChannel classes.
• Add tests for io_uring domain sockets, including sanity checks with the known working epoll transport.

I've marked thie PR as a draft for a few reasons:

• Transferring file descriptors isn't implemented yet. Appears to be a Linux kernel restriction for security reasons. A future version of Linux may allow for this.
• IOUringDomainSocketShutdownOutputByPeerTest fails for some reason.
• IOUringSocketConnectionAttemptTest fails with Invalid argument for a reason that is not clear to me. Fixed, reason was IPv6-mapped IPv4 addresses

Result:

The io_uring transport supports Unix domain sockets.

[normanmaurer]

I will have a look soon...

[normanmaurer]

@astei can you rebase as well ?

[astei]

@normanmaurer rebased

[normanmaurer]

@astei seems like all pass ?

[astei]

@astei seems like all pass ?

They did not pass on my system (Fedora 33 Workstation, Linux 5.10.x kernel). It’s very strange the CI thinks it’s good. I don’t have a ready explanation for why.

[astei]

@normanmaurer when I am back home I will rerun the tests again.

[normanmaurer]

👍

…

Am 02.02.2021 um 21:20 schrieb Andrew Steinborn ***@***.***>:  @normanmaurer when I am back home I will rerun the tests again. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[astei]

@normanmaurer I have re-run the tests, and the same two tests are failing: https://gist.github.com/astei/e3fc19a705c56a086b83ef8b042ef070

[normanmaurer]

Do the same tests also fail when you try to run the epoll tests on netty ?

…

Am 03.02.2021 um 06:10 schrieb Andrew Steinborn ***@***.***>:  @normanmaurer I have re-run the tests, and the same two tests are failing: https://gist.github.com/astei/e3fc19a705c56a086b83ef8b042ef070 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[astei]

Do the same tests also fail when you try to run the epoll tests on netty ?

@normanmaurer all the epoll transport tests pass, which is very strange.

[LifeIsStrange]

Unrelated to this PR:
Openjdk 16 release next month and support natively unix domain socket
https://openjdk.java.net/jeps/380
https://inside.java/2021/02/03/jep380-unix-domain-sockets-channels/

[normanmaurer]

@astei the connect error can be fixed by:

diff --git a/src/main/java/io/netty/incubator/channel/uring/AbstractIOUringChannel.java b/src/main/java/io/netty/incubator/channel/uring/AbstractIOUringChannel.java
index ebd5228..876cd8a 100644
--- a/src/main/java/io/netty/incubator/channel/uring/AbstractIOUringChannel.java
+++ b/src/main/java/io/netty/incubator/channel/uring/AbstractIOUringChannel.java
@@ -676,7 +676,7 @@ abstract class AbstractIOUringChannel extends AbstractChannel implements UnixCha
                 remoteAddressMemory = Buffer.allocateDirectWithNativeOrder(Native.SIZEOF_SOCKADDR_STORAGE);
                 long remoteAddressMemoryAddress = Buffer.memoryAddress(remoteAddressMemory);

-                int addrLen = Sockaddr.write(remoteAddressMemoryAddress, remoteAddress);
+                int addrLen = Sockaddr.write(socket.isIpv6(), remoteAddressMemoryAddress, remoteAddress);

                 final IOUringSubmissionQueue ioUringSubmissionQueue = submissionQueue();
                 ioUringSubmissionQueue.addConnect(socket.intValue(), remoteAddressMemoryAddress, addrLen, (short) 0);
diff --git a/src/main/java/io/netty/incubator/channel/uring/MsgHdrMemory.java b/src/main/java/io/netty/incubator/channel/uring/MsgHdrMemory.java
index 8fb3fa5..dacdb25 100644
--- a/src/main/java/io/netty/incubator/channel/uring/MsgHdrMemory.java
+++ b/src/main/java/io/netty/incubator/channel/uring/MsgHdrMemory.java
@@ -40,7 +40,7 @@ final class MsgHdrMemory {
             addressLength = socket.isIpv6() ? Native.SIZEOF_SOCKADDR_IN6 : Native.SIZEOF_SOCKADDR_IN;
             PlatformDependent.setMemory(sockAddress, Native.SIZEOF_SOCKADDR_STORAGE, (byte) 0);
         } else {
-            addressLength = Sockaddr.write(sockAddress, address);
+            addressLength = Sockaddr.write(socket.isIpv6(), sockAddress, address);
         }
         Iov.write(iovAddress, bufferAddress, length);
         MsgHdr.write(memory, sockAddress, addressLength, iovAddress, 1);
diff --git a/src/main/java/io/netty/incubator/channel/uring/Sockaddr.java b/src/main/java/io/netty/incubator/channel/uring/Sockaddr.java
index e0782a6..734389f 100644
--- a/src/main/java/io/netty/incubator/channel/uring/Sockaddr.java
+++ b/src/main/java/io/netty/incubator/channel/uring/Sockaddr.java
@@ -37,12 +37,11 @@ final class Sockaddr {

     private Sockaddr() { }

-    static int write(long memory, SocketAddress address) {
+    static int write(boolean ipv6, long memory, SocketAddress address) {
         if (address instanceof DomainSocketAddress) {
             return Sockaddr.writeDomain(memory, ((DomainSocketAddress) address).path());
         } else if (address instanceof InetSocketAddress) {
             InetSocketAddress inetSocketAddress = (InetSocketAddress) address;
-            boolean ipv6 = ((InetSocketAddress) address).getAddress() instanceof Inet6Address;
             if (ipv6) {
                 return Sockaddr.writeIPv6(memory, inetSocketAddress.getAddress(), inetSocketAddress.getPort());
             } else {

This is due the fact of ipv6 mapped ipv4 addresses

[normanmaurer]

If you like I can also try to check the other error. But most likely not before next week. Let me know @astei

[victorstewart]

saw this on jen's twitter. fyi you won't be able to transfer file descriptors because sendmsg cmsghdrs are only allowed on inet(6) + stream... and coming in 5.12 inet(6) + dgram.

there's a whole bunch of security issues that were solved with a blunt hammer.

[astei]

Well, that fixes 2/3 issues. I'm not a huge fan of adding back socket.isIpv6() since it looks a little ugly (and quite out of place with a Unix domain socket).

[normanmaurer]

@astei sorry for the slow turn around ... I will try to push this over the finish line next week.

[Kamillaova]

bump

[franz1981]

@astei do you think to rebase? If not I can pick this for the next weeks

[astei]

@franz1981 I can try to get this moving again at some point, but it's not been high on my list of priorities. Since submitting this PR I've gone through a few job changes and my current job is pretty far removed from Netty.

[Kamillaova]

bump