Go 1.9rc1 is due out imminently, which means the time to report issues with Go 1.8 was about 7-9 months ago. I suggest you try out Go 1.9 and report any bugs you find before it's too late for Go 1.9.

Go 1.8 documented:

https://golang.org/doc/go1.8#net_http

The Context returned by Request.Context is canceled if the underlying net.Conn closes. For instance, if the user closes their browser in the middle of a slow request, the Handler can now detect that the user is gone. This complements the existing CloseNotifier support. This functionality requires that the underlying net.Conn implements recently clarified interface documentation.

And https://golang.org/doc/go1.8#net

The Conn documentation has been updated to clarify expectations of an interface implementation. Updates in the net/http packages depend on implementations obeying the documentation.

Updating: implementations of the Conn interface should verify they implement the documented semantics. The golang.org/x/net/nettest package will exercise a Conn and validate it behaves properly.

But your net.Conn implementation looks fine, so I think this is actually a bug in net/http, but it's too late to fix for Go 1.9, so I'l target this at Go 1.10.

Using the modified program https://play.golang.org/p/eRBx-bv4Rr with more logging against Go 1.7 vs Go 1.8:

bradfitz@gdev:~$ GOROOT=$HOME/go1.7 go run hijack.go
2017/07/23 16:39:47 Fetching...
2017/07/23 16:39:47 Conn.Read(4096) ...
2017/07/23 16:39:47 Conn.Read = 96, <nil>
2017/07/23 16:39:47 in Handler.
2017/07/23 16:39:47 Got: &{200 Connected 200 HTTP/1.0 1 0 map[] 0xc4200f4100 -1 [] true false map[] 0xc4200d40f0 <nil>}, <nil>
HTTP/1.0 200 Connected
Connection: close

And Go 1.8:

bradfitz@gdev:~$ GOROOT=$HOME/go1.8 go run hijack.go
2017/07/23 16:39:51 Fetching...
2017/07/23 16:39:51 SetReadDeadline(off)
2017/07/23 16:39:51 Conn.Read(4096) ...
2017/07/23 16:39:51 Conn.Read = 96, <nil>
2017/07/23 16:39:51 SetReadDeadline(off)
2017/07/23 16:39:51 in Handler.
2017/07/23 16:39:51 SetReadDeadline(-352054h39m51.732834171s)
2017/07/23 16:39:51 Conn.Read(1) ...
2017/07/23 16:39:56 Conn.Read = 0, read tcp 127.0.0.1:42428->127.0.0.1:37735: i/o timeout
2017/07/23 16:39:56 SetReadDeadline(off)
2017/07/23 16:39:56 Got: &{200 Connected 200 HTTP/1.0 1 0 map[] 0xc4200184c0 -1 [] true false map[] 0xc42000a800 <nil>}, <nil>
HTTP/1.0 200 Connected
Connection: close

random data

What happens in Go 1.8 is that the Hijack call calls SetReadDeadline(aLongTimeAgo) to interrupt any currently-blocked net.Conn.Read call or any soon-to-be-called net.Conn.Read 1 byte read.

But this net.Conn implementation in its Read method unconditionally calls SetReadDeadline itself, extending it from aLongTimeAgo (1 second after Unix time began) to 5 seconds from current time.

As a workaround, you could make your net.Conn type implement SetReadDeadline itself and track whether it's in aLongTimeAgo mode (I'd just check for any deadline before current time) and if so, don't do your own extensions.

I'll have to think about how to fix this in Go. Too bad we don't have context support in io.Reader. The the Hijack could've just canceled the context of the background 1 byte read, not affecting other deadlines.

/cc @tombergan

Btw, the workaround would look like:

// timeoutableConn
type timeoutableConn struct {
	net.Conn
	hadReadDeadlineInPast int32 // atomic
}

func (c *timeoutableConn) Read(b []byte) (n int, err error) {
	if atomic.LoadInt32(&c.hadReadDeadlineInPast) != 1 {
		c.Conn.SetReadDeadline(time.Now().Add(5 * time.Second))
	}
	return c.Conn.Read(b)
}

func (c *timeoutableConn) SetReadDeadline(t time.Time) error {
	inPast := int32(0)
	if t.Before(time.Now()) {
		inPast = 1
	}
	atomic.StoreInt32(&c.hadReadDeadlineInPast, inPast)
	return c.Conn.SetReadDeadline(t)
}

I've verified that successfully works around this.

@bradfitz, thanks. It also works for me. I appreciate you suggesting the workaround.

Change https://golang.org/cl/62610 mentions this issue: net/http: avoid locking by using goroutine

@tombergan Could you take a look at the CL?

Not convinced this is a bug in net/http. Hijack is not the only code being broken by this Conn implementation. Conn clients also expect that:

conn.SetReadDeadline(time.Now().Add(30*time.Second))
readHeader(conn)

will not read for more than 30 seconds. But this Conn implementation keeps bumping the deadline forward on every read, so as long as the header is coming in 1 byte every 4.9s, the readHeader could go on forever.

If a Conn.Read is going to change the timeout, presumably to keep Read from blocking for too long, it should still use the min of whatever timeout it wants to impose and any previously-set deadline.

I would be in favor of simply closing this bug.

Moving to Go 1.11 in any event but feel free to close, @tombergan or @bradfitz, if you agree with the rationale here.

Agreed, will close. We've gone two releases so far with the Conn requirements spelled out in docs and net/http strictly depending on, and we've had minimum fallout, and most custom Conns have simply fixed their implementations given the new docs.