GODRIVER-3009 Fix concurrent panic in struct codec.

[qingyang-hu]

GODRIVER-3009

Summary

Remove the unnecessary nil cache, so nil encoder/decoder will not be stored.

Background & Motivation

The original panic was:

interface conversion: interface is nil, not bsoncodec.ValueEncoder

from bson/bsoncodec/codec_cache.go:45

The root cause is at https://github.com/mongodb/mongo-go-driver/blob/master/bson/bsoncodec/struct_codec.go#L481

[mongodb-drivers-pr-bot]

API Change Report

No changes found!

[prestonvasquez]

We need to add the following test case to TestRegistryBuilder/Register/Lookup:

				{
					"nil",
					nil,
					nil,
					ErrNilType,
					false,
				},

We might also add a test that captures the repro case in bson package:

func TestConcurrentEncoding(t *testing.T) {
	t.Parallel()

	wg := sync.WaitGroup{}
	wg.Add(10_000)
	for i := 0; i < 10_000; i++ {
		go func() {
			defer wg.Done()
			Marshal(struct{ LastError error }{})
		}()
	}

	wg.Wait()
}

The latter might be caught by the race detector, but it couldn't hurt to have it around. What are your thoughts?

[prestonvasquez]

The LookupEncoder and LookupDecoder return values should match when valueType is nil. I think it makes sense to return ErrNilType here rather than ErrNoEncoder

Suggested change

	return nil, ErrNoEncoder{Type: valueType}
	return nil, ErrNilType

[qingyang-hu]

I agree. However, it will be a breaking change. How do you think about filing a v2.0 ticket instead?

[matthewdale]

The error message for ErrNilType is

"cannot perform a decoder lookup on <nil>"

We'll need to update the error message to make sense for either case.

[matthewdale]

Looks good 👍