remote pdb can not attach when breakpoint() and exit

[yihong0618]

Bug report

Bug description:

reproduce

on terminal

./python.exe
breakpoint()
# ctrl + D

another terminal

sudo ./python.exe -m pdb -p $(pid)

more can check png below

debug:

Pdb._last_pdb_instance still when do_quit in repl so remote pdb can not exec in it

fix:

diff --git a/Lib/pdb.py b/Lib/pdb.py
index f695a39332..3e32c50f96 100644
--- a/Lib/pdb.py
+++ b/Lib/pdb.py
@@ -2008,6 +2008,7 @@ def do_quit(self, arg):
                     return
 
         self._user_requested_quit = True
+        Pdb._last_pdb_instance = None
         self.set_quit()
         return 1
 

CPython versions tested on:

CPython main branch

Operating systems tested on:

macOS

Linked PRs

• gh-139688: make _last_pdb_instance to None when pdb do_quit for remote debug #139690

[gaogaotiantian]

There’s a trade off here I believe. So using breakpoint directly in REPL is not super common. Keeping the pdb instance has the advantage that it can keep all the aliases and the displays. Comparing to the capability to attach to the process remotely, I think making it consistent is slightly better.

I think the major argument here is that attaching to a repl session after a breakpoint is too rare to be useful. If there’s no downside, I’m more than happy to make it work. But to me it seems like that’s not the case?

[yihong0618]

There’s a trade off here I believe. So using breakpoint directly in REPL is not super common. Keeping the pdb instance has the advantage that it can keep all the aliases and the displays. Comparing to the capability to attach to the process remotely, I think making it consistent is slightly better.

I think the major argument here is that attaching to a repl session after a breakpoint is too rare to be useful. If there’s no downside, I’m more than happy to make it work. But to me it seems like that’s not the case?

the problem here is that

    def do_quit(self, arg):
        """q(uit) | exit

        Quit from the debugger. The program being executed is aborted.
        """
        # Show prompt to kill process when in 'inline' mode and if pdb was not
        # started from an interactive console. The attribute sys.ps1 is only
        # defined if the interpreter is in interactive mode.
        if self.mode == 'inline' and not hasattr(sys, 'ps1'):
            while True:
                try:
                    reply = input('Quitting pdb will kill the process. Quit anyway? [y/n] ')
                    reply = reply.lower().strip()
                except EOFError:
                    reply = 'y'
                    self.message('')
                if reply == 'y' or reply == '':
                    sys.exit(1)
                elif reply.lower() == 'n':
                    return

        self._user_requested_quit = True
        # gh-139688 set _last_pdb_instance to None for remote debugging.
        Pdb._last_pdb_instance = None
        self.set_quit()
        return 1

seems only repl and rare script we can goes into this logic

        self._user_requested_quit = True
        # gh-139688 set _last_pdb_instance to None for remote debugging.
        Pdb._last_pdb_instance = None
        self.set_quit()
        return 1

I checked the test_pdb and not found others way to goes into this and has one Pdb._last_pdb_instance, if the check is for repl or others
I am thinking this is right.

[yihong0618]

hi @gaogaotiantian

I made a simpe script without repl

import pdb
import time
import sys

sys.ps1 = ">>> "

def auto_yes_input(prompt=''):
    if 'Quit anyway' in prompt or 'y/n' in prompt:
        print(f"{prompt}y")
        return 'y'
    else:
        print(f"{prompt}q")  
        return 'q'

import builtins
original_input = builtins.input
builtins.input = auto_yes_input


try:
    breakpoint()  
except SystemExit:
    pass
except Exception as e:
    pass

builtins.input = original_input


try:
    for i in range(100):
        print("sleep here")
        time.sleep(2)
except KeyboardInterrupt:
    print("quit")

➜  cpython git:(main) ✗ sudo ./python.exe -m pdb -p 45490
*** Another PDB instance is already attached.

with the patch can fix

update

this script can simpe

import time


try:
    breakpoint()  
except SystemExit:
    pass
except Exception as e:
    pass


try:
    for i in range(100):
        print("sleep here")
        time.sleep(2)
except KeyboardInterrupt:
    print("quit")

[gaogaotiantian]

Maybe I did not make myself clear.

With or without your new code, this is going to be a rare case, where breakpoint() is used with REPL.

We cache the last used pdb instance for a reason - to keep instance setings like display and alias. This is not super critical, but it's still a benefit.

Now the question is - if it's worth it to trade this benefit to the capability to be able to attach to it. I'm leaning towards no, just because it's so rare in real life for users to remote attach to an interactive shell after they use breakpoint() in it. The "bug" this fixes is too theoretical.

Like I said, it's okay to fix a theoretical bug, but only when it does not compromise other use cases. In this specific case, it deprives the benefit I mentioned above, when the user uses breakpoint() multiple times in their REPL session (which I think is slightly more common to remote attach to the REPL session after breakpoint()).

[yihong0618]

Maybe I did not make myself clear.

With or without your new code, this is going to be a rare case, where breakpoint() is used with REPL.

We cache the last used pdb instance for a reason - to keep instance setings like display and alias. This is not super critical, but it's still a benefit.

Now the question is - if it's worth it to trade this benefit to the capability to be able to attach to it. I'm leaning towards no, just because it's so rare in real life for users to remote attach to an interactive shell after they use breakpoint() in it. The "bug" this fixes is too theoretical.

Like I said, it's okay to fix a theoretical bug, but only when it does not compromise other use cases. In this specific case, it deprives the benefit I mentioned above, when the user uses breakpoint() multiple times in their REPL session (which I think is slightly more common to remote attach to the REPL session after breakpoint()).

yes agree with that, but after some digging my question is its not only in REPL which maybe not rare.
and I made a script in the last comment, if this is still a rare, maybe this issue can be closed