Hacktoberfest 2026:维护者为十月标记出来的 issue,仍然开放、适合新手。 浏览 Hacktoberfest issue

render: on macOS the context function's advertised 127.0.0.1:PORT can reach another process's listener (TestListenerRoundTrip flakes)

未关闭
#424 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 3 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
go, macos
领域
cli, networking

调研方向

Start in cmd/crossplane/render/contextfn/listener.go at Start, where the listener is bound to ":0" and the target is advertised as 127.0.0.1:PORT, and read TestListenerRoundTrip in listener_test.go. Run the small two-listener program from the issue on macOS to see the shadowing, and check the lsof output the issue describes. Done means the advertised loopback port always reaches this process on macOS and Linux, and go test -count=100 on TestListenerRoundTrip passes without flakes.

由索引模型根据 Issue 内容生成。

描述

bug

Usernames and home paths in captured output are redacted to me / /Users/me.

What happened?

TestListenerRoundTrip (cmd/crossplane/render/contextfn) fails intermittently on macOS (found while working on unrelated render PRs) because its RPC reaches a server that isn't the context function:

  • rpc error: code = Unimplemented desc = unknown service apiextensions.fn.proto.v1.FunctionRunnerService
  • rpc error: code = Unavailable desc = connection error: desc = "error reading server preface: EOF"
  • ... error reading server preface: http2: failed reading the frame payload: http2: frame too large, note that the frame header looked like an HTTP/1.1 header
  • ... error reading server preface: read tcp 127.0.0.1:57131->127.0.0.1:57130: read: connection reset by peer

The cause is in the code under test. contextfn.Start listens on ":0" (a dual-stack wildcard, [::]:PORT) and advertises 127.0.0.1:PORT. On macOS (as on other BSD-derived kernels) the wildcard doesn't reserve 127.0.0.1:PORT: another process can listen there too, before or after us, and connections to 127.0.0.1:PORT go to that more specific listener. Linux refuses the second bind, so it isn't affected. Which error you get depends on what that listener speaks: gRPC gives unknown service, HTTP/1.1 gives frame too large, a raw TCP forwarder gives EOF or a reset.

Expected: a client dialing the advertised target always reaches the context function.

A small program shows it deterministically. Listening like Start, then binding 127.0.0.1 on the same port and dialing the advertised target, reaches the second listener on macOS; Linux rejects the second bind. The other bind order behaves the same on macOS.

Program and output
// Listens like contextfn.Start (":0"), then lets a second listener bind
// 127.0.0.1 on the same port, then dials the advertised target
// 127.0.0.1:PORT and reports which listener accepted the connection.
package main

import (
	"fmt"
	"net"
	"time"
)

func try(first string) {
	a, err := net.Listen("tcp", first)
	if err != nil {
		panic(err)
	}
	defer a.Close()
	port := a.Addr().(*net.TCPAddr).Port

	b, err := net.Listen("tcp", fmt.Sprintf("127.0.0.1:%d", port))
	if err != nil {
		fmt.Printf("listen %-12q -> %s; then listen 127.0.0.1:%d -> %v\n", first, a.Addr(), port, err)
		return
	}
	defer b.Close()

	got := make(chan string, 2)
	go func() { if _, err := a.Accept(); err == nil { got <- "the first listener" } }()
	go func() { if _, err := b.Accept(); err == nil { got <- "the SECOND listener" } }()
	c, err := net.Dial("tcp", fmt.Sprintf("127.0.0.1:%d", port))
	if err != nil {
		panic(err)
	}
	defer c.Close()
	select {
	case w := <-got:
		fmt.Printf("listen %-12q -> %s; then listen 127.0.0.1:%d -> ok; dial 127.0.0.1:%d reached %s\n", first, a.Addr(), port, port, w)
	case <-time.After(2 * time.Second):
		fmt.Println("no accept")
	}
}

func main() {
	try(":0")
	try("127.0.0.1:0")
}

macOS:

$ go run .
listen ":0"         -> [::]:52543; then listen 127.0.0.1:52543 -> ok; dial 127.0.0.1:52543 reached the SECOND listener
listen "127.0.0.1:0" -> 127.0.0.1:52545; then listen 127.0.0.1:52545 -> listen tcp 127.0.0.1:52545: bind: address already in use

Linux (arm64, in a container):

listen ":0"         -> [::]:43583; then listen 127.0.0.1:43583 -> listen tcp 127.0.0.1:43583: bind: address already in use
listen "127.0.0.1:0" -> 127.0.0.1:44079; then listen 127.0.0.1:44079 -> listen tcp 127.0.0.1:44079: bind: address already in use

Long-lived 127.0.0.1 listeners in the ephemeral range are common on a developer Mac (Rancher Desktop's limactl and steve, Chrome, IDEs, ssh forwards). Instrumenting the test to run lsof -nP -iTCP:<port> on failure caught them: the test binary (probe-con…) held *:PORT while steve or limactl held 127.0.0.1:PORT.

Two captured failures
--- FAIL: TestListenerRoundTrip (0.06s)
    listener_test.go:60: RunFunction: rpc error: code = Unavailable desc = connection error: desc = "error reading server preface: http2: failed reading the frame payload: http2: frame too large, note that the frame header looked like an HTTP/1.1 header"
        pid=42477 lsof:
        COMMAND     PID             USER   FD   TYPE             DEVICE SIZE/OFF NODE NAME
        steve      9356 me    7u  IPv4 0x1762ec606a975469      0t0  TCP 127.0.0.1:49574 (LISTEN)
        probe-con 42477 me    4u  IPv6 0xc9c60a2a9dce0c78      0t0  TCP *:49574 (LISTEN)
FAIL
--- FAIL: TestListenerRoundTrip (2.06s)
    listener_test.go:60: RunFunction: rpc error: code = Unavailable desc = connection error: desc = "error reading server preface: EOF"
        pid=42486 lsof:
        COMMAND     PID             USER   FD   TYPE             DEVICE SIZE/OFF NODE NAME
        limactl    4651 me    9u  IPv4 0x4e3222da67fe7a05      0t0  TCP 127.0.0.1:49340 (LISTEN)
        probe-con 42486 me    4u  IPv6 0xc5d00989bf2fe556      0t0  TCP *:49340 (LISTEN)
FAIL

Failure rate varies with the host's listeners: about 10% of runs on one day (go test -count=50: 5–8 failures), 0 on a quieter one (5 × -count=100 of the test binary, 3 × go test -count=100). For comparable numbers I ran 12 test processes at once, each -test.run '^TestListenerRoundTrip$' -test.count=200 (2,400 runs per round):

Build Rounds Failed runs
origin/main (":0") 12 (28,800 runs) 19: per round 2, 1, 2, 0, 1, 1, 2, 1, 1, 3, 1, 4
Proposed fix below 8 (19,200 runs) 0: every round 0

In 10 of the 19 baseline failures lsof showed another process's 127.0.0.1 listener on the port (limactl 4, steve 2, Chrome 2, a desktop backup agent 2). In the other 9 it had gone by the time lsof ran: 4 frame too large (so an HTTP/1.1 server answered), 1 reset, and 4 DeadlineExceeded … while waiting for connections to become ready. None of these occurred with the fix.

Not just tests. Start's target is what render xr/render op give the render engine, rewritten to host.docker.internal by RewriteAddressesForDocker. With Rancher Desktop's default mapping (Lima's 192.168.5.2), Lima forwards to the host's 127.0.0.1:PORT, so the engine's connection is shadowed the same way. The engine container itself uses host.docker.internal:host-gateway (engine_docker.go), which Rancher resolves to 172.17.0.1; I didn't investigate that path.

Container check

A host program opened a ":0" listener, optionally a second 127.0.0.1:PORT listener, then ran docker run --rm alpine nc host.docker.internal PORT:

shadow=false port=51120 accepted by: contextfn-style listener [::]:51120 (remote 127.0.0.1:51122)
container output: nc exit 0
shadow=true port=51170 accepted by: SHADOW listener 127.0.0.1:51170 (remote 127.0.0.1:51171)
container output: nc exit 0
How can we reproduce it?

On macOS, repeat or parallelize until a run fails (the program above reproduces the underlying behaviour deterministically):

go test -count=100 ./cmd/crossplane/render/contextfn/ -run '^TestListenerRoundTrip$'

Possible fix. Keep the all-interfaces listener: the engine container must reach the context function (e80ffa7: "we have to listen on 0.0.0.0 (not localhost/127.0.0.1) to ensure we're reachable from Docker"). Also claim 127.0.0.1:PORT so nothing else can:

  • Listen on ":0" (port P), then on 127.0.0.1:P. If that succeeds (macOS/BSD), serve gRPC on both: grpc.Server.Serve can be called per listener, and GracefulStop closes them all.
  • If it fails with EADDRINUSE, either our wildcard already owns loopback (Linux) or another process does (macOS). Dial 127.0.0.1:P and check our listener accepted it: keep the single listener in the first case, try another port in the second.

The patch measured above (against 29316fe) passes go vet and golangci-lint run ./cmd/crossplane/render/contextfn/... and the package's tests on macOS, and go test -count=500 on Linux in a golang:1.26 container.

Patch
--- a/cmd/crossplane/render/contextfn/listener.go
+++ b/cmd/crossplane/render/contextfn/listener.go
@@ -22,6 +22,7 @@ import (
 	"fmt"
 	"net"
 	"sync"
+	"syscall"
 	"time"
 
 	fnv1 "github.com/crossplane/function-sdk-go/proto/v1"
@@ -63,8 +64,7 @@ func Start(ctx context.Context, log logging.Logger, contextData map[string]any)
 		return nil, errors.Wrap(err, "cannot create capture context function input")
 	}
 
-	var lc net.ListenConfig
-	lis, err := lc.Listen(ctx, "tcp", ":0")
+	lis, err := listen(ctx)
 	if err != nil {
 		return nil, errors.Wrap(err, "cannot create listener for context function")
 	}
@@ -73,7 +73,7 @@ func Start(ctx context.Context, log logging.Logger, contextData map[string]any)
 	fn := newServer(contextData)
 	fnv1.RegisterFunctionRunnerServiceServer(srv, fn)
 
-	addr := lis.Addr().(*net.TCPAddr) //nolint:forcetypeassert // We specified "tcp" above.
+	addr := lis[0].Addr().(*net.TCPAddr) //nolint:forcetypeassert // listen only returns TCP listeners.
 
 	h := &Handle{
 		// Report the target as 127.0.0.1:PORT since the render machinery knows
@@ -85,15 +85,81 @@ func Start(ctx context.Context, log logging.Logger, contextData map[string]any)
 		captureInput: &runtime.RawExtension{Raw: ci},
 	}
 
-	go func() {
-		if err := srv.Serve(lis); err != nil {
-			log.Debug("Context function gRPC server stopped", "error", err)
-		}
-	}()
+	for _, l := range lis {
+		go func() {
+			if err := srv.Serve(l); err != nil {
+				log.Debug("Context function gRPC server stopped", "error", err)
+			}
+		}()
+	}
 
 	return h, nil
 }
 
+// listen returns listeners for a single port P that is reachable on all
+// interfaces (so the render engine container can reach it via
+// host.docker.internal) and whose 127.0.0.1:P address - the target we
+// advertise - is guaranteed to reach this process.
+//
+// On BSD-derived kernels such as macOS's, a dual-stack wildcard listener on
+// [::]:P does not stop another process from listening on 127.0.0.1:P (before or
+// after us), and connections to 127.0.0.1:P go to that more specific listener.
+// We therefore also claim 127.0.0.1:P ourselves. On Linux that bind fails
+// because our wildcard listener already owns 127.0.0.1:P; we confirm that by
+// dialing ourselves. If another process owns 127.0.0.1:P we try another port.
+func listen(ctx context.Context) ([]net.Listener, error) {
+	var lc net.ListenConfig
+	for range 10 {
+		all, err := lc.Listen(ctx, "tcp", ":0")
+		if err != nil {
+			return nil, err
+		}
+		port := all.Addr().(*net.TCPAddr).Port //nolint:forcetypeassert // We specified "tcp" above.
+
+		lo, err := lc.Listen(ctx, "tcp4", fmt.Sprintf("127.0.0.1:%d", port))
+		if err == nil {
+			return []net.Listener{all, lo}, nil
+		}
+		if !errors.Is(err, syscall.EADDRINUSE) {
+			_ = all.Close()
+			return nil, err
+		}
+		if servesLoopback(ctx, all, port) {
+			return []net.Listener{all}, nil
+		}
+		_ = all.Close()
+	}
+	return nil, errors.New("cannot find a port whose loopback address is not held by another process")
+}
+
+// servesLoopback reports whether a connection to 127.0.0.1:port is accepted by
+// lis. It must be called before lis is served.
+func servesLoopback(ctx context.Context, lis net.Listener, port int) bool {
+	ctx, cancel := context.WithTimeout(ctx, time.Second)
+	defer cancel()
+
+	var d net.Dialer
+	c, err := d.DialContext(ctx, "tcp4", fmt.Sprintf("127.0.0.1:%d", port))
+	if err != nil {
+		return false
+	}
+	defer c.Close() //nolint:errcheck // Nothing useful to do.
+
+	tl := lis.(*net.TCPListener) //nolint:forcetypeassert // listen only creates TCP listeners.
+	_ = tl.SetDeadline(time.Now().Add(time.Second))
+	defer tl.SetDeadline(time.Time{}) //nolint:errcheck // Nothing useful to do.
+
+	a, err := tl.Accept()
+	if err != nil {
+		return false
+	}
+	defer a.Close() //nolint:errcheck // Nothing useful to do.
+
+	ra := a.RemoteAddr().(*net.TCPAddr) //nolint:forcetypeassert // Accepted from a TCP listener.
+	la := c.LocalAddr().(*net.TCPAddr)  //nolint:forcetypeassert // Dialed over TCP.
+	return ra.Port == la.Port && ra.IP.Equal(la.IP)
+}
+
 // Stop gracefully stops the function server, which closes its listener. Safe to
 // call multiple times.
 func (h *Handle) Stop() {

Listening on 127.0.0.1:0 instead also stops the flake (0 failures across 19,200 runs), but reverts e80ffa7: on Linux host.docker.internal:host-gateway resolves to the bridge IP, so the engine container couldn't reach a loopback-only listener.

A deterministic regression test could assert the invariant directly: after Start, binding 127.0.0.1:<port> must fail with EADDRINUSE. On macOS that bind succeeds against origin/main; on Linux it fails before and after the fix.

What environment did it happen in?
  • Crossplane CLI version: built from main @ 29316fea54f2ede9d2c039d9c54f0c29cbad4b65 (go version go1.26.7 darwin/arm64)
  • Platform (e.g., linux/amd64): darwin/arm64, Docker Engine 29.5.3 (Rancher Desktop)
  • Crossplane version (if applicable): n/a (unit test)
主要语言
Go
星标
20
派生
34
平均合并
5 天 4 小时
30 天内合并 PR
21

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

crossplane/cli 的其他 Issue

查看 crossplane/cli 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。