100% CPU usage after closing pod log terminal

[rade]

After closing a pod log terminal, the CPU usage of the probe on the host on which the pod resides jumps by 100% and stays there - looks like it's busy-looping on one core.

I've observed this in several environments, using Chrome and FF.

Container logs don't suffer from this.

Popping out the pod log terminal and then closing it does not trigger the problem.

All the probes I've tested with were 1.7.0. I also checked a cluster with 1.6.7 and the problem does not occur there.

I suspect #3013 may have introduced this.

[rade]

I grabbed a go-routine dump from one of the spinning probes and found

goroutine 23452 [runnable]:
github.com/weaveworks/scope/probe/kubernetes.(*logReadCloser).Read(0xc4216ea8c0, 0xc421091000, 0x400, 0x400, 0x0, 0xc42105a600, 0xc421ffbd10)
        /go/src/github.com/weaveworks/scope/probe/kubernetes/logreadcloser.go:88 +0x365
go.(*struct { io.Reader; io.Writer }).Read(0xc4213415a0, 0xc421091000, 0x400, 0x400, 0x400, 0x0, 0x0)
        <autogenerated>:1 +0x5a
github.com/weaveworks/scope/common/xfer.(*pipe).CopyToWebsocket.func2(0x2f22780, 0xc4213415a0, 0xc421b41260, 0xc421275d50, 0x2f31880, 0xc421ffbd00)
        /go/src/github.com/weaveworks/scope/common/xfer/pipes.go:141 +0x9e
created by github.com/weaveworks/scope/common/xfer.(*pipe).CopyToWebsocket
        /go/src/github.com/weaveworks/scope/common/xfer/pipes.go:138 +0x1a8

Here's a theory: The channels in the select were closed by logReadCloser.Close and hence reading from them always succeeds.