Waiter removed twice when waking suspended agent?

[binji]

@AndrewScheidecker pointed out what seems to be a bug in the spec here:
WebAssembly/design#1019 (comment)

Incidentally, I believe I found an erroneous assertion in the SAB spec. In Atomics.wait, step 18 performs RemoveWaiter after Suspend completes. RemoveWaiter asserts that the agent being removed is in the WaiterList. However, if Suspend returned because a call to Atomics.wake called WakeWaiter, the agent will have been removed from the WaiterList already by step 11 of Atomics.wake.

I took a look and this seems right to me. Are we missing something?

cc: @lars-t-hansen @syg

[syg]

@lars-t-hansen wrote that part but I concur with @AndrewScheidecker's reading here that there's double removal in the case of non-spurious wakeup in Atomics.wake. I'd imagine the fix here is to perform RemoveWaiter from Atomics.wake only in the case of spurious wakeup, and in the intentional wakeup case, assert that W is not in WL.

[syg]

@lars-t-hansen wrote that part but I concur with @AndrewScheidecker's reading here that there's double removal in the case of non-spurious wakeup in Atomics.wake. I'd imagine the fix here is to perform RemoveWaiter from Atomics.wake only in the case of spurious wakeup, and in the intentional wakeup case, assert that W is not in WL.

[lars-t-hansen]

@syg, I assume by "spurious wakeup" you mean "the agent was explicitly woken, it did not time out". If so, I agree with your suggestions. (If you actually mean "spurious wakeup" then we have bigger problems, because there shouldn't be any spurious wakeups.)

[lars-t-hansen]

@syg, I assume by "spurious wakeup" you mean "the agent was explicitly woken, it did not time out". If so, I agree with your suggestions. (If you actually mean "spurious wakeup" then we have bigger problems, because there shouldn't be any spurious wakeups.)

[lars-t-hansen]

Opened a PR here: #865

[lars-t-hansen]

Opened a PR here: #865

[syg]

@lars-t-hansen Err, I'm not sure what I meant as I typed that while leaving work yesterday -- but timeout is the right thing here.

[syg]

@lars-t-hansen Err, I'm not sure what I meant as I typed that while leaving work yesterday -- but timeout is the right thing here.