Bug report

Bug description:

As @gaogaotiantian pointed out in #132451 (comment) PDB's new remote attaching feature can enter an uninterruptible infinite loop if you type while True: pass at a PDB prompt. We're able to interrupt the script itself, when you 'cont'inue past a breakpoint, but aren't able to interrupt a statement evaluated interactively at a PDB prompt.

I have a few different ideas for how this could be fixed, though they all have different tradeoffs.

Options:

1. the client can send a SIGINT signal to the remote.

• • Pro: it can interrupt IO, as long as PyErr_CheckSignals() is being called by the main thread
• • Con: it will not work on Windows
• • Con: it will kill the process if the SIGINT handler has been unset (and PDB does mess with the signal handler - I'm not sure this is just a theoretical problem)

2. we can raise a KeyboardInterrupt from an injected script

• • Pro: It will work on Windows
• • Con: We need to change sys.remote_exec() to allow KeyboardInterrupt to escape
• • Pro: A looping injected script can be killed now, though!
• • Con: It cannot interrupt IO
• • Con: The KeyboardInterrupt will get raised even in places where signals were blocked, potentially leading to a crash where something that should never get an exception got one

3. we can make PDB stop if set_trace() is called while it's evaluating a user-supplied statement

• • Pro: it will work on Windows
• • Con: It cannot interrupt IO
• • Con: it will lead to recursion of PDB's trace function if the script is interrupted, and I don't know how well PDB will handle that

4. the PDB server can start a thread that listens to a socket and calls signal.raise_signal(signal.SIGINT) every time it gets a message on the socket

• • Pro: it can interrupt IO in the same cases as 1 can
• • Pro: it works in "true remote" scenarios where the client and server are on different hosts (which isn't currently a feature, but would be a nice future feature)
• • Pro: it works on Windows
• • Con: more complexity, more failure modes

I'm currently leaning towards (4) - it's the most complete fix and works on the most platforms and the most scenarios (interrupting both CPU-bound and IO-bound stuff). I plan to prototype this soon.

CPython versions tested on:

CPython main branch

Operating systems tested on:

No response

Linked PRs

• gh-132975: Improve Remote PDB interrupt handling #133223