I just noticed that the binary that stack trace is from was compiled with Go 1.14, not Go 1.14.1 like I had thought originally. But we got the same panics with a binary compiled on Linux with Go 1.14.1.

@bradfitz @kardianos @kevinburke

@andybalholm, could you try testing with a go build from head, or patching in CL 220905? I wonder if this is another symptom of the bug that was fixed there.

Building with

go version devel +620208790e Wed Apr 1 19:32:57 2020 +0000 darwin/amd64

I still get the panic:

panic: net/http: internal error: connCount underflow

goroutine 67124 [running]:
net/http.(*Transport).decConnsPerHost(0x108df80, 0x0, 0x0, 0xc00b2800e4, 0x4, 0xc002061900, 0x17, 0x0)
	/usr/local/go/src/net/http/transport.go:1388 +0x590
net/http.(*persistConn).closeLocked(0xc00af806c0, 0xb847e0, 0xc00006cb60)
	/usr/local/go/src/net/http/transport.go:2586 +0xf5
net/http.(*persistConn).close(0xc00af806c0, 0xb847e0, 0xc00006cb60)
	/usr/local/go/src/net/http/transport.go:2576 +0x83
net/http.(*persistConn).closeConnIfStillIdle(0xc00af806c0)
	/usr/local/go/src/net/http/transport.go:1890 +0xcb
created by time.goFunc
	/usr/local/go/src/time/sleep.go:169 +0x44

@bcmills This is an http request issue not http2.

@andybalholm, could you share some more detail about your usage? (Any unusual Transport customizations, proxies, or anything like that?)

If you have a bunch of these panic traces, are there any that don't have that stack sequence (decConnsPerHost called from closeLocked called from close)?

@bcmills Here is my transport:

var dialer = &net.Dialer{
	Timeout:   30 * time.Second,
	KeepAlive: 30 * time.Second,
	DualStack: true,
}

var httpTransport = &http.Transport{
	Proxy:                 http.ProxyFromEnvironment,
	DialContext:           dialer.DialContext,
	TLSHandshakeTimeout:   10 * time.Second,
	ExpectContinueTimeout: 1 * time.Second,
}

The environment doesn't specify a proxy. I call http2.ConfigureTransport(httpTransport) elsewhere in the code. (But not making that call doesn't fix the problem.)

@rsc All the stack traces seem to have that sequence. And close is being called from either net/http.(*persistConn).readLoop.func1 or net/http.(*persistConn).closeConnIfStillIdle.

One thing I realize is that we dont check the return of removeIdleConnLocked during closeConnIfStillIdle before we blindly close the persistConn.

More interesting is that the order of pc.close() and removeIdleConn vary which means we could easily get into this situation.

@andybalholm I am confused. You mention that you set MaxConnsPerHost = 8, but the above does not show you setting it. Where/When did you do that.

@andybalholm, do you set those after the Transport is already in use? I don't believe that's safe in general. (It wouldn't surprise me if it's even a data race.)

What happens if you run your program under the race detector?

Can you run with -race? That would catch this pretty fast if it was done later.

I can't run it with -race on the server where I was getting the panics; it's too heavily loaded. But I tried it on a local server with just me proxied through it, and that was enough:

WARNING: DATA RACE
Read at 0x000001587540 by goroutine 87:
  net/http.(*Transport).onceSetNextProtoDefaults()
      /usr/local/go/src/net/http/transport.go:344 +0x61
  net/http.(*Transport).onceSetNextProtoDefaults-fm()
      /usr/local/go/src/net/http/transport.go:343 +0x41
  sync.(*Once).doSlow()
      /usr/local/go/src/sync/once.go:66 +0x103
  sync.(*Once).Do()
      /usr/local/go/src/sync/once.go:57 +0x68
  net/http.(*Transport).roundTrip()
      /usr/local/go/src/net/http/transport.go:485 +0x8b
  net/http.(*Transport).RoundTrip()
      /usr/local/go/src/net/http/roundtrip.go:17 +0x42
  main.proxyHandler.ServeHTTP()
      /root/src/redwood/proxy.go:328 +0x17fb
  main.(*proxyHandler).ServeHTTP()
      <autogenerated>:1 +0x114
  net/http.serverHandler.ServeHTTP()
      /usr/local/go/src/net/http/server.go:2807 +0xce
  net/http.initALPNRequest.ServeHTTP()
      /usr/local/go/src/net/http/server.go:3381 +0xfc
  net/http.(*initALPNRequest).ServeHTTP()
      <autogenerated>:1 +0xa6
  net/http.Handler.ServeHTTP-fm()
      /usr/local/go/src/net/http/server.go:87 +0x64
  golang.org/x/net/http2.(*serverConn).runHandler()
      /root/go/pkg/mod/golang.org/x/net@v0.0.0-20190213061140-3a22650c66bd/http2/server.go:2095 +0xc0

Previous write at 0x000001587540 by main goroutine:
  golang.org/x/net/http2.configureTransport()
      /root/go/pkg/mod/golang.org/x/net@v0.0.0-20190213061140-3a22650c66bd/http2/transport.go:174 +0x420
  golang.org/x/net/http2.ConfigureTransport()
      /root/go/pkg/mod/golang.org/x/net@v0.0.0-20190213061140-3a22650c66bd/http2/transport.go:136 +0x7ca
  main.main()
      /root/src/redwood/redwood.go:88 +0x7ab

It's been several years since I tried running with -race. I didn't find anything then, and I figured only a heavily-loaded production server would have enough concurrency to expose races if there were any. But now I know that running with -race locally is worth doing.

Reordering the startup steps gets rid of the race.

Running go 1.14.6 or 1.14.7

Since this is a different code path and happening during automatic cleanup, seems like this shouldn't happen...

panic: net/http: internal error: connCount underflow
goroutine 25 [running]:
net/http.(*Transport).decConnsPerHost(0xc0004ba280, 0x0, 0x0, 0x7fae7523e750, 0x5, 0xc0004af9e0, 0x18, 0x0)
/usr/local/go/src/net/http/transport.go:1391 +0x590
net/http.(*persistConn).closeLocked(0xc0000a4000, 0x132adcf00, 0xc000118c50)
/usr/local/go/src/net/http/transport.go:2584 +0xf5
net/http.(*persistConn).close(0xc0000a4000, 0x132adcf00, 0xc000118c50)
/usr/local/go/src/net/http/transport.go:2574 +0x83
net/http.(*persistConn).closeConnIfStillIdle(0xc0000a4000)
/usr/local/go/src/net/http/transport.go:1888 +0xcb
created by time.goFunc
/usr/local/go/src/time/sleep.go:168 +0x44

Should I open a new issue for this?

@mrjrieke please try your code with the race detector, as that proved to be the problem above. If that doesn't reveal a mistake, please file a new issue.