testing: loses log messages made in a goroutine after test has completed

[ianlancetaylor]

The testing package silently drops t.Log messages that are made by a goroutine after the test has completed. For example, the following test passes and logs nothing.

This should be fixed by extending the code added in 1.12 that panics if a goroutine logs after a test has completed to not silently discard the log message. The silent discarding is occurring in the current code because every test runs in a parent context, and the output accumulated by the parent context is discarded.

package main

import (
	"testing"
	"time"
)

func TestLateLog(t *testing.T) {
	c := make(chan bool)
	go func() {
		<-c
		time.Sleep(time.Millisecond)
		t.Log("log after test")
	}()
	close(c)
}

[rhnvrm]

Hey,

I tried investigating and wrote a test and modifying various places in testing.go like the log and logDepth function in my fork with debug statements (fmt.Fprintf(os.Stdout, "%s", string(s))) but the log never ran.

I might be naively making a mistake here but I guess after the TestLateLog function exits t.output or t.parent output is never changed and flush output never has anything to print.

func TestLateLogWrapper(t *testing.T) {
	testenv.MustHaveGoBuild(t)
	tempDir, err := ioutil.TempDir("", "issue-30389")
	if err != nil {
		t.Fatalf("Failed to create temp directory: %v", err)
	}
	defer os.RemoveAll(tempDir)
	code := `
package main

import (
	"testing"
	"time"
	"fmt"
)

func TestLateLog(t *testing.T) {
	c := make(chan bool)
	go func(t *testing.T) {
		<-c
		time.Sleep(10*time.Millisecond)
		fmt.Printf("-> %v\n", t)
		t.Log("log after test")
	}(t)
	close(c)
}
`
	destFile := filepath.Join(tempDir, "issue30389_test.go")
	if err := ioutil.WriteFile(destFile, []byte(code), 0644); err != nil {
		t.Fatalf("Failed to save test code: %v", err)
	}
	got, err := exec.Command(testenv.GoToolPath(t), "test", destFile).CombinedOutput()
	if err == nil {
		t.Fatal("The command unexpectedly passed")
	}
	want := []byte(`xxx`)
	if !bytes.HasPrefix(got, want) {
		t.Errorf("Did not print previous output\nGot:\n%s\n\nWant prefix:\n%s", got, want)
	}
}

Another thing I had a doubt about was, if we are to add a panic before the following to notify users that this information will be lost (or even for debugging if we reach there for this specific case)
parent.output = append(parent.output, parent.decorate(s, depth+1)...)
is there a way to detect that it was a test without a tRunner