dns: add dns_rr_dup and test_dns_rr_dup

[maximilianfridrich]

No description provided.

[cspiel1]

Can confirm that two requests compete to have the same rr out of the DNS cache in their req->addrl. But still digging into the details what happens in the main thread when the DNS responses are handled nearly simultaneously.

[alfredh]

Would it be possible to move this to the regular tests ?

The regular tests include more variations like OOM, threading etc..

[maximilianfridrich]

Yes, thanks for the feedback. We will keep working on this. We're pretty sure the shared pointers to the rr objects lead to the DNS problems we're experiencing with parallel outgoing SIP requests. We're trying to pinpoint it exactly and come up with a proper solution.

[cspiel1]

The test does not check what currently is expected when dnsc API is used in request.c.

ASSERT_TRUE(parallel.rrv[0] != parallel.rrv[1]);

It checks that rr is not shared.

[cspiel1]

Currently we observe a failed DNS request if two SIP INVITE are tried to be sent directly one after another.

I am still not sure that there is a bug in client.c or request.c. Because the req->addrl is filled and consumed during one call of addr_handler(), making it possible to call two query handlers in a row that report the same rr.

Anyway, we have a failed SIP requests with addr_handler() returning EDESTADDRREQ with the parcall module. But only if empty AAAA responses are involved. Currently we are trying to understand, reproduce and fix what we observed at one of our customers.

[cspiel1]

Note, that q->rrlv[i] use rr->le_priv to link the elements, while req->addrl uses re->le. Thus it looks correct.

[cspiel1]

I think that I found the bug.
The empty AAAA response is not put into the DNS cache. Now suppose that two SIP requests are sent close enough before the first empty AAAA arrives. Both A queries can be found in the cash, but can't be processed immediately because of these check

	/* wait for other (A/AAAA) query to complete */
	if (req->dnsq || req->dnsq2)
		return;

The continuity of filling and consuming req->addlr is interrupted in this case. Only one of the requests will succeed.

I'll try to write a unit test that reproduces this.

[alfredh]

This method is called Happy Eyeballs and can be used for e.g. SIP and HTTP, please take a look here:

https://datatracker.ietf.org/doc/html/rfc6555

Both DNS A+AAAA query is sent at the same time and the first one to respond will be used.
It may be used for a dualstack-host that has both IPv4+IPv6 address.

[cspiel1]

As a first step, should we add AAAA tests? #1515

Then we could add a test that runs A+AAAA query in parallel and reproduces the issue.

[cspiel1]

This looks good. The windows test run was already fixed. A rebase should solve this.