IPv6 QoS fix

[rdica]

Short description of changes

When IPv6 is enabled, this provides a fix to properly typecast tos so setsockopt can properly add DSCP field value.
Also adds error checking for setsockopt() calls and ensures correct QoS operation for dual-protocol sockets.

CHANGELOG: Client+Server: Correct QoS setting for IPv6

Context: Fixes an issue?

Fixes: #3623

Does this change need documentation? What needs to be documented and how?
No, just a bugfix

Status of this Pull Request

What is missing until this pull request can be merged?

Needs testing on all but Linux OSes.

Checklist

• I've verified that this Pull Request follows the general code principles
• I tested my code and it does what I want
• My code follows the style guide
• I waited some time after this Pull Request was opened and all GitHub checks completed without errors.
• I've filled all the content above

[softins]

I've been doing some tests of this on Linux (RPi), and it needs a little more investigation.

I verified that the setsockopt() previously returned an error, and that with this fix it succeeds.

I also tested it with different QoS values for server and client, running on the same machine connecting via localhost (::1) on IPv6, while monitoring the traffic with Wireshark. The correct QoS (DSCP+ECN) was sent from each end.

However, if the client initiated the connection to 127.0.0.1 instead, to force an IPv4 connection, both ends sent a QoS value of 0.

If I started the client without -6, and connected to 127.0.0.1 or to localhost, then the traffic from the client to the server carried the expected QoS value, but that from the server (still running with -6, so a dual-protocol socket) back to the client sent QoS of 0.

So I need to check how to get a dual-stack socket to send the correct QoS value when it has an IPv4 connection. Maybe it will also accept the setsockopt ( UdpSocket, IPPROTO_IP, IP_TOS, ...) call?

[softins]

So I need to check how to get a dual-stack socket to send the correct QoS value when it has an IPv4 connection. Maybe it will also accept the setsockopt ( UdpSocket, IPPROTO_IP, IP_TOS, ...) call?

Yes, it works. I'll update the PR directly.

[softins]

I have done some builds on Windows. Here are my findings:

1. IPV6_TCLASS is not supported on Windows and returns an error from setsockopt()
2. IP_TOS is supported on Windows, but appears to do nothing. Whatever value I give it, the Differentiated Services Field in the IPv4 header is always zero.

Looking at https://learn.microsoft.com/en-us/previous-versions/windows/desktop/qos/sending-qos-enabled-data it would appear that setting QoS in Windows uses a completely different API from that on Linux.

For 3.12.0, I would suggest that we just specify QoS is not supported on Windows. We may want to reject -Q or emit a warning that it is ignored.

It would be worth checking on Mac too, but hopefully that should be ok, as Mac is Unix-based.

[pljones]

I vaguely remember a long and detailed explanation of what was needed to explain to Windows (10 at the time, IIRC) that QoS/TOS should be respected. This was either on the original PR or in the issue or discussion it was related to. It's probably worth digging that up and reviewing it.

[softins]

It would be worth checking on Mac too, but hopefully that should be ok, as Mac is Unix-based.

Sadly, it doesn't work completely on macOS. Using IP_TOS on an AF_INET socket works to set TOS for IPv4, and using IPV6_TCLASS works to set TOS for IPv6 traffic. But macOS doesn't accept the IP_TOS call on an AF_INET6 socket in the way that Linux does, so there is no way on macOS or iOS, when IPv6 is enabled, to set the default TOS for IPv4 traffic.

ChatGPT confirmed this is the case, and suggested the best way round it would be to create a drop-in replacement for sendto() called sendto_with_tos(), that uses sendmsg() to set the TOS for each outgoing packet. It said this would be compatible with Linux too. Although not with Windows.

[softins]

Well so far, I have #if-ed out the calls that will error on Windows and Apple. With this change, the following are true:

1. On Linux and Android, QoS will be correctly sent for all traffic whether or not the -6 option was given, including traffic sent to an IPv4 address via an IPv6 socket (-6 given).
2. On macOS and iOS, QoS will be correctly sent for all (IPv4) traffic if -6 was not given. If -6 was given to enable IPv6 support, QoS will be correctly sent for traffic to an IPv6 address, but is not sent (field is always zero) with traffic to an IPv4 address.
3. On Windows, QoS is never sent - the field is always zero.

[pljones]

Gemini says for Windows (though I didn't check versions):
...

qWAVE (Quality Windows Audio/Video Experience)
This is the modern, recommended API for Windows applications (especially multimedia) to handle QoS. It handles both the marking of packets and the "bandwidth admission" (checking if the network can handle the stream).
Header: qos2.h
Library: Qwave.dll
Functionality: It uses "Traffic Types" rather than raw hex values:

Traffic Type	Description	DSCP Value (Approx)
QOSTrafficTypeBestEffort	Standard traffic	0
QOSTrafficTypeBackground	Low priority	8
QOSTrafficTypeVideo	Video streaming	34 (AF41)
QOSTrafficTypeVoice	VoIP/Audio	46 (EF)
QOSTrafficTypeControl	Network control	48

---

Sounds like it's very version specific but originated around Windows Vista/7 and was fully supported in Windows 10 onwards. (Not Windows Server, though.)

It did also reference the "hack" I vaguely remembered:

The Registry "Fix": To make Windows respect the user-defined TOS value via Winsock, you must often set a registry key:
Path: HKEY_LOCAL_MACHINE\System\CurrentControlSet\Services\Tcpip\Parameters
Value: DisableUserTOSSetting (DWORD) set to 0.

And then "Socket Option: IPPROTO_IP level, option IP_TOS", so long as Group Policy allows it.

I think I went with group policy:

"Any traffic from 'MyApp.exe' should be marked with DSCP 46."

I'd imagine the NSIS installer has a Group Policy hook? Maybe it could be done there.

[pljones]

And it found the original discussion for me: https://github.com/orgs/jamulussoftware/discussions/1072
And that has the PR linked: #1310

The original discussion makes a good point: the application itself should know the right value -- it shouldn't be the user picking numbers at random from the command line.

[softins]

Interesting stuff. So what to do with this PR, and what to defer to another? This one has already improved a bit on @rdica's original, without much radical change, and now fully works on Linux. It seems to me that this is worth merging now for 3.12.0, and then we raise two more separate PRs for after 3.12.0:

1. Add use of sendmsg() when required for correct sending of TOS for IPv4 over an AF_INET6 socket on macOS and iOS.
2. Implementation of TOS in the right way for Windows.

Comments?

[softins]

The original discussion makes a good point: the application itself should know the right value -- it shouldn't be the user picking numbers at random from the command line.

Yes, agreed. Although if -Q is not specified, the default number in Jamulus is 0x80, which is DSCP of CS4 and ECN of 0, giving improved priority over untagged traffic.

It would be worth understanding which other values are useful for Jamulus, and providing useful names for the user to specify them. Again, post-3.12