Handle SIGBREAK for Windows

[yanyongyu]

On Windows, subprocess.CTRL_BREAK_EVENT may be sent to subprocess using os.kill or process.send_signal as SIGBREAK signal. This signal can be catched by subprocess to do a graceful shutdown instead of terminating the subprocess directly.
And, this signal is also catched in hypercorn.

[Kludex]

Is this an alternative to #1584 ?

[yanyongyu]

Is this an alternative to #1584 ?

This pr may be related to #1584 but not an alternative. The reloader should do something else on Windows including send CTRL_BREAK_EVENT(or maybe CTRL_C_EVENT) to worker process and preventing SIGBREAK(or CTRL_C_EVENT) being handled in main process when reloading.

In our application, we use subprocess to run uvicorn now and SIGBREAK is sent to subprocess on Windows when shutting down (as process.terminate will call winapi to force exit). uvicorn uses multiprocessing to spawn workers, i'm not sure if the signal handling is the same. maybe @StarHeartHunt we need a further testing on uvicorn reloader.

Another problem is that CTRL_C_EVENT will be sent to all processes in the process group on Windows. We cannot use it as we have multiple subprocess workers running. We finally change to using subprocess.CREATE_NEW_PROCESS_GROUP and signal.CTRL_BREAK_EVENT.😓

Reference:

1. note in python docs: https://docs.python.org/3/library/subprocess.html#subprocess.Popen.send_signal
2. our code: https://github.com/nonebot/nb-cli/blob/0221588fdc132891fbe7ccd8a8b15487380e722c/nb_cli/handlers/process.py#L56-L104
3. windows docs: https://learn.microsoft.com/en-us/windows/console/generateconsolectrlevent, https://learn.microsoft.com/en-us/windows/win32/procthread/process-creation-flags

[StarHeartHunt]

As mentioned above, The Ctrl + Break signal is supposed to be handled by uvicorn as a kind of exit signal too, for external subprocess calls with subprocess.CREATE_NEW_PROCESS_GROUP to gracefully shutdown the server process, This is what this PR meant for.

And #1584 tries fixing the graceful exit in reloader when uvicorn it self is the root process, it seems in this case we can use Ctrl+C signal to achieve this goal (Maybe the internal code is different in multiprocessing?). The solution has been proved to be valid on my Windows 10 laptop.

TL;DR: Based on current investigation and discussions in our team @nonebot, To implement a graceful exit:

1. subprocess with a subprocess.CREATE_NEW_PROCESS_GROUP flag or multiprocessing
2. The main process can only use CTRL_BREAK_EVENT as exit signal in subprocess situations, but multiprocessing could use CTRL_C_EVENT, and in both situations, the main process should implement a shield context to avoid being terminated by itself.
3. A signal (CTRL_BREAK_EVENT / CTRL_C_EVENT) is sent to the subprocess.

[Kludex]

Can you also show me the reference on hypercorn that you mentioned in the description, please?

[yanyongyu]

Can you also show me the reference on hypercorn that you mentioned in the description, please?

That's here: https://github.com/pgjones/hypercorn/blob/8ae17ca68204d9718389fb3649ca0ed6ba851906/src/hypercorn/run.py#L61-L63

By the way, hypercorn uses a multiprocessing spawn Event instead of signals to notify the worker to exit.

[Kludex]

Did this caused #1972 ?

[yanyongyu]

Our team member test the 0.22.0 version but can not repreduce the issue on windows 10. The detailed info is shown in the #1584 (comment)