carp tests: test manually switch between backup and master

[kp]

Summary

There's been at least one issue where we failed to correctly enter
NET_EPOCH that was triggered in this scenario.
Add a test case for it to make it easier to detect issues like this in
the future.

Sponsored by: Rubicon Communications, LLC ("Netgate")

[asomers]

tests/sys/netinet/carp.sh

284

What is the purpose of the sleep?

[kp]

tests/sys/netinet/carp.sh

284

Carp sends packets on a timeout, and this gives it a bit of time for that timer to actually fire and send packets.
We probably don't need it, I mostly have it because Zhenlei's example in https://reviews.freebsd.org/D39113 had it.

[asomers]

tests/sys/netinet/carp.sh

284

I don't think you should include it, unless there's something that you verify after the sleep. Like, poll until you can tell that the timer expired and sent packets.

[zlei]

tests/sys/netinet/carp.sh

279

Well it is a little interesting here.
Without this ifconfig ${epair}a up the virtual host state keep stuck with INIT .

Is it a known bug or feature ?

284

The test script in https://reviews.freebsd.org/D39113 is focusing on testing forced state change, due to the design (and current implementation) of CARP protocol, on ifconfig state master packets (carp / arp / nd) are sent out immediately, so actually no sleep required.

[zlei]

due to the design (and current implementation) of CARP protocol, on ifconfig state master packets (carp / arp / nd) are sent out immediately, so actually no sleep required.

Before I checked carefully the design of CARP protocol, I thought there's a little gap between carp packets and arp / nd packets. I thought the competitor should not send arp / nd packets until it actually become MASTER.

That is by intuition, but is not true (for CARP protocol).

[kp]

Remove sleep

[zlei]

The switch_body() part looks good to me.