Fix DHT tests failing because of repeated addresses

[Kubuxu]

Due to SO_REUSEPORT it is possible for a localhost:0
address to repeat. This causes failure in DHT tests
where we spun up a lot of nodes inside test.

As for a birthday paradox it is enough to use 140
ports to get 20% chance for collision which was causing
failure in our case.

The fix is to disable REUSE_PORT routine for the
tests and leave it running for sharness tests where
not that many addresses are used at the same time.

Resolves #2745 #2809 #188

[Kubuxu]

[whyrusleeping]

Awesome job catching this, yet another SO_REUSEPORT issue... yay

[jbenet]

• Great find!
• We should continue to have SO_REUSEPORT being stress tested because we may find issues with it (maybe it behaves differently or poorly). If not in these tests, maybe in others? Or would there be another way of signaling ports in the tests here?

[jbenet]

@chriscool was there a reason to do the make after a cd? pls confirm this change is perfectly fine with you.

[chriscool]

Yeah, it's fine. One could also use make -C test/sharness/ TEST_EXPENSIVE=1. It's a matter of preference whether one wants the shell or make to do the cd and to set the environment variable.

[jbenet]

Hold up, this disabled SO_REUSEPORT in ALL the go tests. this is pretty bad and flying blind. if there's any more trouble with SO_REUSEPORT we'll find it with stress testing all the socket use, which the go-ipfs suite does pretty well. Obviously the "accidentally reuse port" problem really sucks... is there another way of going about this? Maybe in a more granular way? (in the failing tests, or in tests which allocate tons of ports). Of course, if we can do a different way of allocating ports even better.

[Kubuxu]

Some smarter way of allocating ports would be the best solution, unfortunately it is hard to do without race of conditions (I don't know if there is a way at all to do it 100% correctly).

The go-reuseport is very deep in the stack so disabling it only locally would require some refactors, that is why I decided to disable it globally.

Also any test that would stress go-reuseport would have quite a high change of triggering the collision.

I am all for re-enabling the go-reuseport in tests but we need to work on smart port selection first (0.0.0.0/0 is just not working with SO_REUSEPORT set).