Re: shutil.rmtree() fails when used in Fedora (rpm) "mock" environment

On 2024-10-24 at 20:54:53 +0100,
MRAB via Python-list <python-list@python.org> wrote:

On 2024-10-24 20:21, Left Right wrote:

The stack is created on line 760 with os.lstat and entries are appended
on lines 677 (os.rmdir), 679 (os.close) and 689 (os.lstat).
>
'func' is popped off the stack on line 651 and check in the following lines.
>
I can't see anywhere else where something else is put onto the stack or
an entry is replaced.

 
But the _rmtree_safe_fd() compares func to a *dynamically* resolved
reference: os.lstat. If the reference to os changed (or os object was
modified to have new reference at lstat) between the time os.lstat was
added to the stack and the time of comparison, then comparison
would've failed.  To illustrate my idea:
 
os.lstat = lambda x: x # thread 1
stack.append((os.lstat, ...)) # thread 1
os.lstat = lambda x: x # thread 2
func, *_ = stack.pop() # thread 1
assert func is os.lstat # thread 1 (failure!)
 
The only question is: is it possible to modify os.lstat like that, and
if so, how?
 
Other alternatives include a malfunctioning "is" operator,
malfunctioning module cache... all those are a lot less likely.

What is the probability of replacing os.lstat, os.close or os.rmdir from
another thread at just the right time?

That is never the right question in a multi-threaded system.  The answer
is always that is doesn't matter, the odds will beat you in the end.  Or
sometimes right in the middle of a CPU instruction; does anyone remember
the MC680XX series?

Yes, as a matter of fact, I did used to make my living designing,
building, delivering, and maintaining such systems.