implement GC keyword for CNI spec 1.1

[henry118]

This PR implements the GC keyword in libcni for CNI spec 1.1

Fixes: #974

[coveralls]

coverage: 71.605% (-1.1%) from 72.685% when pulling e86c9ca on henry118:gc into 743a44d on containernetworking:main.

[s1061123]

This requires func.Check must not be nil (otherwise nil pointer reference happen). Should we add validate all func.{Add,Del,Check,GC} in initiliazation (or support 'nil' to mention unsupported)

[henry118]

This patch added the nil reference check here.

[squeed]

Chatted with @dcbw for a bit, and we both agreed we want to see just one more thing: can you somehow expand the debug plugin so we can determine the containerID and ifname that is being passed to the DEL?

After that we can merge this. Thanks!

[s1061123]

I'm wondering that plugin's GC might be causes several race conditions:

For example:

1. plugin A's GC is invoked by container runtime and plugin A starts the GC call
2. During pluing A's GC call, (unfortunately) container runtime invokes plugin A's GC again (oh god...)

In addition, we also have another example:

1. plugin A's GC is invoked by container runtime and plugin A starts the GC call
2. During pluing A's GC call, (unfortunately) container runtime invokes plugin A's Add

To simplify plugin code, both should be blocked (i.e. prevent to invoke), I suppose and it should be implemented common code, in libcni.

What do you think about that?

[henry118]

I'm wondering that plugin's GC might be causes several race conditions:

For example:

1. plugin A's GC is invoked by container runtime and plugin A starts the GC call

2. During pluing A's GC call, (unfortunately) container runtime invokes plugin A's GC again (oh god...)

In addition, we also have another example:

1. plugin A's GC is invoked by container runtime and plugin A starts the GC call

2. During pluing A's GC call, (unfortunately) container runtime invokes plugin A's Add

To simplify plugin code, both should be blocked (i.e. prevent to invoke), I suppose and it should be implemented common code, in libcni.

What do you think about that?

The GC spec explicitly called out those race conditions. The proposed mechanism is letting the runtimes to serialize the operations.