For an example program illustrating why a user might reasonably expect (and attempt to make use of) this behavior, see https://play.golang.org/p/mDkMG67ehAI.

Thanks for filing this.

I also found this behavior puzzling and apparently I got it wrong too as I was expecting no values to be sent after t.Stop() returned.

If, as I understand now, a value might be sent after the call to Stop has returned, is there any way to correctly re-use a timer?

If, as I understand now, a value might be sent after the call to Stop has returned, is there any way to correctly re-use a timer?

If you know that no other goroutine is receiving on the C channel, then you can receive unconditionally after the call to Stop returns false, and it won't actually matter whether the value was delivered before or after the call to Stop returned.

(The select-with-default after Stop is where the bug occurs, and none of the examples in your post do that — they don't need to, because in your examples the reads on t.C are all on the same goroutine as the calls to t.Stop.)

In the current implementation, this can probably be done by changing runOneTimer to move the Cas away from timerRunning until after the invocation of f.

Added to proposal process to resolve whether to do this, after mention in #38945.

This made its first appearance in the proposal minutes last week.
You can imagine ways to break Go programs with this, such as by

t := time.NewTimer(1*time.Second)
time.Sleep(2*time.Second)
t.Stop()
<-t.C

This would succeed today, but if we clear the channel during Stop it would start blocking.
We would have to define that in the case where Stop pulls the element back out of the channel buffer, it returns true, meaning it "stopped" the timer (false means the timer already expired; expired would now mean "expired and was observed by a channel receive on t.C").

On the other hand, there is technically a race here, and it's not guaranteed that this snippet never blocks, especially on a loaded system. The correct snippet, even today, is:

t := time.NewTimer(1*time.Second)
time.Sleep(2*time.Second)
if !t.Stop() {
    <-t.C
}

That example would be unaffected, because Stop would pull the buffered element back out of the channel and then return true, signaling to the caller that there is nothing to receive.

We could make the change at the start of a release cycle and be ready to roll it back if real (as opposed to hypothetical) problems arose.

Thoughts?

While I like the behavior of pulling the value back out of the channel and having Stop return true, I don't think it is actually backward-compatible, depending on how the len builtin for channels interacts with the memory model.

For example, this program today is not reported as a race and the receive is guaranteed to never block, but would deadlock if Stop clears the channel:

package main

import (
        "fmt"
        "time"
)

func main() {
        t := time.NewTimer(1 * time.Second)
        for len(t.C) == 0 {
                time.Sleep(time.Millisecond)
        }
        t.Stop()
        <-t.C
        fmt.Println("ok")
}

That is why I had restricted my original proposal to guarantee only that no value is sent after Stop returns.

Sorry, I'm not sure I understand what it means to guarantee that no value is sent on the channel after Stop returns. As far as I can see, that is already guaranteed--maybe not in the documentation, but certainly in the implementation. What is the actual change that you are suggesting?

@ianlancetaylor, the program given in https://play.golang.org/p/Wm1x8DmYoQo should run indefinitely without reporting any delayed times. Instead, it reports a nonzero rate of delayed timer sends.

On my Xeon workstation:

example.com$ gotip version
go version devel +b2ce3931 Tue May 26 22:28:45 2020 +0000 linux/amd64

example.com$ gotip run main.go
Timer stopped at 2020-05-27 22:16:46.951252699 -0400 EDT m=+2.886904949 subsequently received time
                 2020-05-27 22:16:46.951252454 -0400 EDT m=+2.886904697.
Timer stopped at 2020-05-27 22:17:00.794561424 -0400 EDT m=+16.730213661 subsequently received time
                 2020-05-27 22:17:00.794559263 -0400 EDT m=+16.730211512.
Timer stopped at 2020-05-27 22:17:00.774413697 -0400 EDT m=+16.710065937 subsequently received time
                 2020-05-27 22:17:00.77440917 -0400 EDT m=+16.710061421.
Timer stopped at 2020-05-27 22:17:03.679525664 -0400 EDT m=+19.615177943 subsequently received time
                 2020-05-27 22:17:03.679522832 -0400 EDT m=+19.615178663.
Timer stopped at 2020-05-27 22:17:06.024962313 -0400 EDT m=+21.960614549 subsequently received time
                 2020-05-27 22:17:06.024960025 -0400 EDT m=+21.960612270.
Timer stopped at 2020-05-27 22:17:06.692670731 -0400 EDT m=+22.628322974 subsequently received time
                 2020-05-27 22:17:06.692669848 -0400 EDT m=+22.628322092.
Timer stopped at 2020-05-27 22:17:22.942118175 -0400 EDT m=+38.877770415 subsequently received time
                 2020-05-27 22:17:22.942117468 -0400 EDT m=+38.877769713.
Timer stopped at 2020-05-27 22:17:22.901857123 -0400 EDT m=+38.837509358 subsequently received time
                 2020-05-27 22:17:22.901855958 -0400 EDT m=+38.837508205.
Timer stopped at 2020-05-27 22:17:27.342696328 -0400 EDT m=+43.278348567 subsequently received time
                 2020-05-27 22:17:27.342693224 -0400 EDT m=+43.278345463.
Timer stopped at 2020-05-27 22:17:28.993568136 -0400 EDT m=+44.929220391 subsequently received time
                 2020-05-27 22:17:28.993571601 -0400 EDT m=+44.929223839.
Received 10 delayed times after 128517449 iterations.

In all cases, the failure mode is the same: at some point after Stop returns, the channel receives a very-slightly-earlier timestamp. That is: the return from Stop is not synchronized with the actual send to the channel.

Instead, I propose that the send to the channel, if it occurs at all, ought to happen before the return from Stop.

Thanks for the example. If the code is written per the documentation of time.Timer.Stop, it works correctly (https://play.golang.org/p/KlC__ujuE5M).

And now I see how this happens. It's because in runOneTimer we set the status to timerNoStatus before we call the function that sends a value on the channel, which permits deltimer and therefore time.Timer.Stop to return false before we send the value. The return of false does reliably mean that a value will be sent on the channel, but the value may not have been sent when Stop returns.

I don't see an easy fix in the current implementation. We can't leave the timer in timerRunning state while we unlock pp.timersLock in order to run the function.

Would it make sense to change the timer to the timerModifying state while running the callback (which should be short), and then to timerNoStatus after the callback completes?

(Or would that potentially induce starvation for a time.Ticker with a very short period?)

The same send-after-Stop problem appears to also applies to time.Ticker, which I suppose is not a surprise because they share the same underlying timer implementation.

(Code in https://play.golang.org/p/Y_Hz4xkYr07, but it doesn't reproduce there due to the Playground's faketime patch.)

It's actually a bit more complicated than I thought. If we leave the timer status as timerRunning (or timerModifying), then if the timer function does anything that touches the timer, it can hang. For example, the scavenger uses a timer with a function of wakeScavenger that expects to stop the timer. That would hang waiting for the timer to leave timerRunning status. The netpoll code also manipulates the timer in the timer function, though I think it wouldn't cause any problems here.

Let's try to separate the discussion of semantics from implementation.

The docs say:

To ensure the channel is empty after a call to Stop, check the return value
and drain the channel. For example, assuming the program has not received
from t.C already:

    if !t.Stop() {
    	<-t.C
    }

The proposed change in semantics is to make it that no receive can ever happen after t.Stop returns. That would mean that under the assumption - "assuming the program has not received from t.C already" - t.Stop would never return false anymore. So the above code would no longer be required (and if left alone would just never execute the if body).

This would avoid the problem of people not knowing to write that fixup code or not understanding the subtlety involved. Certainly we've all confused ourselves enough just thinking through this over the years.

That's the proposed semantic change.

There are also questions about how to implement that - it's not just the easy "pull the item back out of the buffer" that I suggested before - but certainly it is possible. Let's figure out if we agree about the semantic change.

(Another side-effect would probably be that len(t.C) would always return 0, meaning the channel would appear to be unbuffered. But we've never promised buffering in the docs. The buffering is only there because I was trying to avoid blocking the timer thread. It's a hack.)

I agree that if we could change the semantics to ensure that no receive is possible after Stop, that would be preferable to both the current behavior and the originally-proposed change to only disallow a send-after-Stop.

However, I don't see any way to do that without breaking the example in https://play.golang.org/p/r77N1PfXuu5.

That program explicitly relies on the buffering behaviors of t.C: specifically, that len(t.C) becomes nonzero when the timer fires, and that the buffered value remains receivable after t.Stop(). Neither of those behaviors is strictly guaranteed by the documentation, but nor does anything in the documentation explicitly reserve them as subject to change — and they have been observable since CL 4063043 over nine(!) years ago.

Disallowing a receive after Stop must necessarily break one of those two behaviors. If we marked the buffer as non-empty but then retracted the buffered value in t.Stop(), then the program would block forever on the subsequent receive. If we instead made the channel unbuffered — or, equivalently, hooked the runtime to avoid sending to it until some goroutine is ready to receive — then the program would loop forever waiting for len(t.C) to become nonzero.

So the only way I see to make that change in semantics would be to declare that this example program is already incorrect, because it is relying on undocumented behavior regardless of the stability of that behavior. I could be wrong, but given how long this behavior has been in place I suspect that changing it would break real programs.

Do you know of any programs that call len on a timer channel?

I do not know of any specifically, but given the 9-year interval and Hyrum's law I would be surprised if they did not exist.

If we are willing to assume that such programs do not exist (or declare that they are invalid if they do exist, because the buffering behavior was never documented), then preventing the receive seems ok.

Is this actually the same issue as this old one? #11513

@rogpeppe, yes it certainly subsumes that one. The example I gave there in #11513 (comment) would definitely change but after seeing real uses, I think more code will be fixed than would break. It can be an early in cycle change.

Also, in that comment I said it's not the library's job to "throw away values already sent on the channel". The fix I'm suggesting above makes the channel (indistinguishable from) unbuffered instead, so there are no values "already sent".

Based on the discussion above, this sounds like a likely accept.

One possible implementation would be:

1. Find a bit in the channel struct to mark timer channels.
2. Stop inlining chanlen/chancap.
3. Make chanlen/chancap return 0 for timer channels even though the representation has buffer size 1.
4. In Stop/Reset, clear the buffer if there's something in it.

There are other more involved implementations (see for example #8898), but this simple change seems like the shortest path to the semantics we've been discussing.

No change in consensus, so accepting.

Change https://go.dev/cl/568341 mentions this issue: time: avoid stale receives after Timer/Ticker Stop/Reset return

Change https://go.dev/cl/568375 mentions this issue: cmd/compile: compile cap(ch) as call to runtime.chancap

Looks like this may happen for Go 1.23, finally.

In the discussion above #37196 (comment) I said the channel would "appear to be unbuffered", but I don't think I realized at the time (or I forgot) that that's literally the entire semantic change: timer channels will now be synchronous aka unbuffered. All the simplification follows from that. (The implementation is more subtle, including a hidden buffer as an optimization, but the semantic effect is simply that timer channels will now behave as if unbuffered.)

For a graceful rollout, this change and the change for #8898 can be disabled by GODEBUG=asynctimerchan=1.

Just found this thread while trying to track down a test flake - the test was trying to drain the channel with a <-timer.C but also provided a default: flag which caused the old event to fail to be read. I am glad that the behavior here is changing to make it more difficult to do the subtly wrong thing.

With tip, the finalizer of a timer will not get executed. Is it intended?

package main

import "time"
import "runtime"

func main() {
	timer := time.NewTimer(time.Second)
	runtime.SetFinalizer(timer, func(c *time.Timer){
		println("done")
	})
	timer.Stop()
	runtime.GC()
	time.Sleep(time.Second)
}

$ gotv 1.22. run main.go
[Run]: $HOME/.cache/gotv/tag_go1.22.1/bin/go run main.go
done
$ gotv :tip run main.go
[Run]: $HOME/.cache/gotv/bra_master/bin/go run main.go
$ 

BTW, since Go 1.23, does the Stop method of a Timer still need to be called to make the Timer collectable as earlier as possible?

BTW, since Go 1.23, does the Stop method of a Timer still need to be called to make the Timer collectable as earlier as possible?

https://go-review.googlesource.com/c/go/+/512355#:~:text=This%20CL%20fixes,be%20GC%27ed%20immediately.

@go101 The finalizer change sounds like a bug to me. Please open a new issue for that. Thanks.

@ianlancetaylor I'm not sure whether or not it is a bug, because the comment in the PR mentions a Timer object might be not allocated on heap, Anyway, I created the issue.

Thanks for creating the issue. Seems to me it is a bug either in the implementation or in the documentation.