go: testing: panic in Init if flag.Parse has already been called

In case anyone else lands here and is looking for a simple answer:

With Go 1.13 you can’t do this anymore:

// foo_test.go
func init() {
	if !flag.Parsed() {
		flag.Parse()
	}
}

You must do this instead:

// foo_test.go
func TestMain(m *testing.M) {
	flag.Parse()
	os.Exit(m.Run())
}

What we (or at least I) intended with #21051 / CL 173722 was:

• Anyone using go test shouldn’t observe any difference or need to explicitly call testing.Init.
• People using testing outside of go test (notably, testing.Benchmark within a package main) would need to use testing.Init if they want to use test flags.

So @vearutop this means that you shouldn’t need to use testing.Init in your example code. In any case, it certainly shouldn’t fail with such a mysterious error.

As @bcmills alluded to, the problem is that code which calls flag.Parse during initialization won’t get the test flags. The initialization of the user’s test package (“ptest”/“pxtest” in the go tool internal parlance) happens before the package main test runner initialization (“testmain”) so there’s nothing we can do in the generated testmain code to fix this.

The workaround that makes sense to me, which is what I think @bcmills is suggesting, is that we can insert a bit of code into the ptest/pxtest packages before we build them:

var _ = func() bool {
	testing.Init()
	return true
}()

(Putting this in a variable declaration expression means that it happens before flag.Parse is called in a variable initialization as well as an init function. Calling flag.Parse as part of variable declaration is ill-advised, of course, but it does work today.)

I think this is pretty straightforward. I’ll send a CL soon.

If I understand correctly, the init functions within a package run in the order in which the source files were presented to the compiler, right?

Yeah, and same for variable declarations, modulo initialization dependencies. I’ll ensure that the synthesized file is passed to the compiler first.

The design of that pin feature is horrible… the pinned issue appears in the main issues page and it looks like a “notification” of some sort, with a prominent “X” in the upper-right area that people will press thinking they’re just removing a notification… instead, they unpin the issue and usually they don’t even notice they performed a public, global action.

Re-pinning this issue, for the same reason @bcmills originally did.

FYI @davecheney and everyone else who participated in the little pinning/unpinning skirmish.

I know this is closed and I’m late to the party, but I want to express how sad this makes me. Go has gone this long (until 1.13) telling people to put flag.Parse() in the init() and then all the sudden the behavior is silently broken. Do we expect all those users to find this thread? We will never know how many users we pissed off or alienated with this one.

I would go as far as to say that this is very close to being against the principle of “no breaking language changes until go 2”. The fact that turning this failure into a panic is “too disruptive” tells you exactly how bad this is. instead of those users being greeted with an error, we just confuse them with broken behavior and maybe they’ll end up finding this thread some day. This is not the “go” I remember.

Just because we feel like it should be new tribal knowledge that flag.Parse() shouldn’t be called in init does not mean we can just break the user experience.

Just look above this comment at the long list of people whose evenings have been ruined by this.

From a high level of programming obviousness, calling flag.Parse() in init makes perfect sense. If the problem is that other people may have added flags or something in a downstream package, make it so flag.Parse() can only be called in package main or something.

It’s hard to understand why code would call flag.Init in an init function that is not in the main package. That will ignore flags defined in other packages that are initialized later. I think I would want to see a clear explanation as to why we need to cater to that case at all, since it seems broken by design.

I’m afraid that I have to disagree with your premise. We never told people to call flag.Parse in init. We always told people to call it in main. Calling flag.Parse in init doesn’t make sense, as it means that all flags added by subsequent init functions would be unknown. That is true even in the main package.

I talked with @rsc and @jayconrod today about what to do about this. Given the fact that CL 176098 didn’t completely resolve the breakage, and the complexity of both that and CL 184923, we’re inclined to roll back CL 176098 and take the position that it’s just not appropriate for init functions to call flag.Parse.

(I think that’s roughly in line with @mvdan’s opinion that the fixes are “too magical […] and in the end not worth doing”.)

Alternatively env vars can be used instead of flags, though I’m not sure what is violated when flag.* is called outside main package.

To be clear, we are talking only about flag.Parse here, not flag.*. The flag package is designed so that multiple packages can define flags, without those packages being aware of each other. It’s essential that flag.Parse be called only after all flags are defined, meaning only after all packages are initialized. The normal practice is that the main package calls flag.Parse, because the main package is the only package that is sure to run after all packages are initialized.

Your program may well be designed so that it is OK to call flag.Parse in an init function of some non-main package. But most programs do not and cannot work that way. So I am questioning how much we should twist the testing package to support that use case. Of course if the change to the testing package is simple, then that is fine. But I would not like to see a lot of complexity added to support a very unusual use case.

Hello, I know I drop late to this, we also stumbled into this issue and while we kind of worked around it, I feel that what we do now to address this is a bit more convoluted than it used/needs to be. So the essence of the problem is the following question:

• what if multiple packages of the application need for the init()-ialization part parameters that are to be extracted from the command line via Parse?. We initially implemented that with a special singleton object that was created in a configuration package that was imported by every other package needing to access some command line parameters. The configuration package did the flag.Parse in its init() function and everything worked just fine, until this change. Now we had to re-implement the initialization to be explicitly called via a chain of events that follow back to an initial call in the application’s main function that actually triggers the flag.Parse() at a later time just to make it compatible with the new rules. So what I’m saying is that this change renders unusable any package init()-ialization that would be dependent on cli parameters. I don’t think this should be the case, this seems an artificial limitation to me. I do understand the reasoning for this change, but maybe something else can be made to address this scenario as well? Maybe change in the flag.Parse() implementation to allow multiple interleaved calls to flag.Var and flag.Parse without generating errors?

@bcmills that sounds like a great resolution to me.

Ideally I think we’d reject flag.Parse during init entirely (by panicking), but I’m sure that’s too disruptive and backwards-incompatible of a change to do now. The issues in this thread brought to my attention how common this is, and https://github.com/golang/go/issues/31859#issuecomment-492563946 makes me wonder how many cases there are in the wild where such code only happens to work as intended due to unspecified package initialization order (see #31636).

Perhaps it would be appropriate to check for flag.Parse during init as a vet check or in some other static analysis tool.

~We’re hitting flag provided but not defined: -test.testlogfile issue, even though we call flag.Parse() in the TestMain() function and we don’t do any flag related calls anywhere else.~

~Do you think we might be hitting the same or a different issue?~

EDIT: It turns out that we, in fact, did call flags.Parse(os.Args[1:]) somewhere in the runtime (really big code base). What a stupid code; triggering a stupid side effect 😃

Now, I wish Go 2 restricted flags.Parse() into Main/TestMain functions only 😄

Adding a note to the next release documentation

There was one in the Go 1.13 release notes (https://golang.org/doc/go1.13#testing):

Testing flags are now registered in the new Init function, which is invoked by the generated main function for the test. As a result, testing flags are now only registered when running a test binary, and packages that call flag.Parse during package initialization may cause tests to fail.

(That’s not to say that we shouldn’t also clarify documentation and implement the vet check, of course.)

We too hit this at work, and I agree that we shouldn’t have been doing it in the first place - a library that declares a flag shouldn’t also call flag.Parse.

I personally would close this issue as fixed, but I’d make it much simpler for developers upgrading to 1.13 to notice what’s wrong and know how to fix it. For example, we could make the testing package panic if flag.Parse is called before the main package’s init; this would be similar to what was done in https://golang.org/cl/177419

Hi – sorry, I’ve been meaning to revisit this issue but haven’t had time recently. I’ll try to put together a CL which implements the approach I mentioned and then we can decide if we want to put it in or do something else for 1.13. I’ll have that in the next 2 days.

@mvdan

wouldn’t the testing package be different when built as part of a test binary versus as part of a non-test binary? Not that that’s a bad thing per se, but it could lead to confusing behavior or unnecessarily increase build times.

Can you be more specific about confusing behavior? I’m not sure how you’d observe the difference.

Build times shouldn’t be an issue; the build cache should still cache both versions.

we wouldn’t have to mess with build tags.

To be clear, my suggested workaround does not involve build tags.

@mvdan I agree: we should panic in testing.Init if flag.Parse has already been called.

I think it’s important that we allow flag.Parse in an init function within a *_test.go file, since (in most cases) that file fills the same role as a main package.

Using https://golang.org/pkg/testing/#hdr-Main is a bit cleaner.

Right, I’m saying that without the _ "testing" import your repro code is buggy due to depending on undefined behavior, and that bugginess is exposed by the init ordering changes unrelated to my testing changes that are on tip.

By adding the _ "testing" import, your repro code no longer depends on undefined behavior, but it was broken by my change.

So:

• You need to fix your code if it depends on testing being initialized first without actually importing it
• We need to fix Go so that any package (not just the package under test) sees that the test flags have been registered when running as a test

Sorry, I meant to say that there are some changes on tip (that will be in Go 1.13). I updated my comment.

@vearutop thanks for bringing it to my attention.

It turns out that this particular repro code relies on undefined behavior: it assumes that the testing package is initialized before the somewhere package. There’s nothing in the spec that says it will be. And indeed, it so happens that on tip there are some initialization changes that already broke this code before any of my testing changes landed; see https://golang.org/cl/161337, https://golang.org/cl/170318, and related discussion over on #31636. If I run the repro with Go at 8515d9cf656dedce3dbcb09ac7dc00f036e454d3, which is this commit where CL 170318 went in and before any of my testing changes, I get a failure:

~/apps/godev/bin/go test
flag provided but not defined: -test.timeout
Usage of /tmp/go-build059416027/b001/aaa.test:
  -stop-on-failure
        Stop processing on first failed scenario.
exit status 2
FAIL    aaa     0.001s

(And if I run with -race, it sometimes fails and sometimes succeeds.)

However, this all kind of doesn’t matter because you could fix your repro code by making somewhere import testing, perhaps just by adding _ "testing" to the imports. Then you’re guaranteed that testing is initialized first, and so it works up until my changes break it. So this is a problem.

To state it more generally, the “testinginit” workaround I did over in https://golang.org/cl/176098 ensures that testing.Init is called before the package under test is initialized. It does not ensure that testing.Init is called before that package’s dependencies are initialized, so if they call flag.Parse during initialization they may not see the testing flags.

I think that a solution is to move the testinginit workaround from the package under test into the testing package. That is, go test would insert this file into the testing package:

package testing
func init() { Init() }

(This would be very simple with a test build tag, but we decided in #21360 not to do that.)

@bcmills thoughts?