For Go 1.12, I plan to implement non-cooperative preemption directly in the runtime, which may move much of this logic into the runtime itself.

That being the case, personally, I'd rather wait for 1.12's cycle for this. I think it's unlikely that we are going to finish implementing the debugger side of call injection before we go is deep in 1.12's cycle, even as it is now.

cc @derekparker

That being the case, personally, I'd rather wait for 1.12's cycle for this.

Fair enough, but if we do end up making more movement on the implementation of call injection on our end reliably, I'd rather get this feature into users hands sooner rather than later, this is a particularly exciting one.

I'd rather get this feature into users hands sooner rather than later

Note that debugger function call injection is supported by the runtime now. It's just that the debugger has to let the whole process resume execution during the injected call. You probably have a better sense of the impact that has on the user experience than I do.

As a data point, GDB by default continues the entire process during things like call injection and stepping (though set scheduler-locking can turn that off on Linux at least).

Change https://golang.org/cl/134779 mentions this issue: runtime: support disabling goroutine scheduling by class

I wanted to ping for an update on this issue as @aarzilli @heschik @hyangah and I spoke about it today during our meeting.

It seems like some of the foundation is there to prevent the runtime from scheduling goroutines that aren't already in the running state. However if I understand correctly were still a bit short on this feature. We would still need cooperation from the runtime to "unschedule" any running goroutine aside from the one Delve wants to be allowed to execute correct? As I understand it, even if we inject schedEnableUser(false) into every thread, that thread will still be _Grunning meaning it will still execute, correct? I assume the above CL prevents any non-running goroutine from being scheduled, but does not deschedule any running goroutine.

If that is the case, then what is the proposed path forward in order to finish this feature? From the original description @aclements suggested injecting a call on each thread which would stop the goroutine. From the standpoint of Delve that makes sense as it would allow us to selectively inject this call onto any thread running a goroutine besides the one we wish to continue execution.

We would really like to get this feature completed as I think that having other side effects during a user-injected function call is less than ideal. Is there any bandwidth to get this done during the 1.13 cycle?

We have asynchronous preemption now, which makes things a bit simpler, but there are still several edge cases to consider.

• We really want direct Go scheduler cooperation here. Bypassing cooperation and just continuing a single OS thread would work for a little while, but would fall apart when that thread enters the scheduler, or tries to allocate if another thread is holding a runtime lock.
• The previous point likely means we want to do this by preempting all goroutines and putting them in some stopped state. However, asynchronous preemption is asynchronous (as advertised!), so doing this naively would allow a little execution skew on resume, as we try to stop all goroutines.
• We need to be careful about coordinating with other scheduling events. e.g., if the resumed goroutines spawns another goroutine, that goroutine shouldn't start running. A STW should continue to work, but also not resume frozen goroutines when continued.
• If the goroutine that the debugger wishes to resume isn't currently running, we need to transition it to running.

A scheme similar to what @aclements originally proposed is a good start, but there are still a lot of holes here:

1. Inject a debugger preempt request on all threads. On success, this puts the running goroutine in _Grunnable (or _Gdebugstop?) and then returns to the debugger.
   1. This fails at unsafe points, and the debugger needs to continue execution until at a safe point.
   2. In user code, the debugger could look at _PCDATA_UnsafePoint and continue / single step until the next safe point.
   3. In runtime code, single stepping may be too slow, but setting a breakpoint on function return may be good enough?
      1. I am vaguely concerned we cause deadlocks here especially if one of the threads in the runtime is trying to do something like STW.
2. Once all goroutines are preempted, tell the runtime to freeze all execution except requested goroutines.
   1. How to do this is a bit unclear. We can't inject a call anymore because all threads will be in the runtime.
3. Tell the runtime to run specified goroutines, which get _Grunning|Gdebugrun.
4. Continue execution on all threads. The runtime will only run _Grunning|Gdebugrun threads.

I think a lot of this would be simplified if we could allow more skew. i.e., set a global debugPreempt and then periodically send sigPreempt to all threads until all goroutines stop. But this is much more asynchronous and would allow more code to run on goroutines that you want to freeze. I don't know if that would be an acceptable user experience in a debugger?

   2. In user code, the debugger could look at `_PCDATA_UnsafePoint` and continue / single step until the next safe point.

At the moment delve doesn't read any of the pcdata/funcdata structures of the runtime. This is nice because we don't have to worry about keeping up with changes made to them. If we decide that delve needs to start reading those we'll also have to figure out a way to at least detect that the format has changed and we need to update delve.

I think a lot of this would be simplified if we could allow more skew. i.e., set a global debugPreempt and then periodically send sigPreempt to all threads until all goroutines stop. But this is much more asynchronous and would allow more code to run on goroutines that you want to freeze. I don't know if that would be an acceptable user experience in a debugger?

Probably not. Much of the motivation behind this feature is to be used in conjunction with either call injection or next/step/stepout, if we allow other goroutines to execute a lot of code before actually stopping it wouldn't be much different from the status quo.

However, if we allow delve to read _PCDATA_UnsafePoint we could do something like the debugPreempt with minimal skew:

• set debugPreempt = true, also "somehow" communicate to the runtime which goroutine(s) we want to keep running
• for every thread running a goroutine that needs to be stopped:
  • if it is running non-runtime code
    • if it's already at a safe point do nothing, otherwise:
    • set a breakpoint on every safepoint of the current function
    • set a breakpoint on the return address (I'm assuming the return address is always a safe point, if it isn't then a breakpoint on every safe point of the caller function)
  • if it is running runtime code:
    • set a breakpoint on the first non-runtime function in the stack
• resume all threads, sending a SIGURG to threads already at a safepoint
• every time one of the breakpoints set above is hit resume the thread converting SIGTRAP into SIGURG
• eventually every goroutine but the ones we want to run will be stopped

This should work fine with other STW in progress (according to my limited understanding of GC).
But what about windows? The asyncpreempt mechanism used on windows is incompatible with debuggers, I'm guessing there's no solution here.

   2. In user code, the debugger could look at `_PCDATA_UnsafePoint` and continue / single step until the next safe point.

At the moment delve doesn't read any of the pcdata/funcdata structures of the runtime. This is nice because we don't have to worry about keeping up with changes made to them. If we decide that delve needs to start reading those we'll also have to figure out a way to at least detect that the format has changed and we need to update delve.

FWIW, I recently added a test to detect these failures in x/debug, which does some pcdata/funcdata parsing. If we revamped x/debug to provide more public API, then perhaps delve could use that.

But what about windows? The asyncpreempt mechanism used on windows is incompatible with debuggers, I'm guessing there's no solution here.

I am not familiar with debugging on Windows. Am I correct that this is because we use SuspendThread, which is what I assume debuggers use as well?

If so, then in theory I think delve itself could implement preempM. Though that of course makes it very sensitive to runtime internals.

I am not familiar with debugging on Windows. Am I correct that this is because we use SuspendThread, which is what I assume debuggers use as well?

Yes, see: #36494

Also, I forgot to mention one thing. Because one possible use of this is in combination with call injection I think we should plan for it to play nice with it, but call injection invoves spawning a new goroutine to execute the call, which means that that goroutine should run, should this apply to all new goroutines, or no? weird.