gh-144281: Fix crash on memoryview slice assignment with shared memory

[VanshAgarwal24036]

Fix a possible interpreter crash when assigning to a memoryview slice backed by shared memory by validating the underlying buffer before writing. A regression test and NEWS entry are included.

• Issue: SIGBUS (Invalid Write) in memory_ass_sub via multiprocessing.shared_memory during buffer assignment #144281

[duaneg]

It is not clear this is a necessary or sufficient fix without a valid specific trigger.

[VanshAgarwal24036]

I’ve updated the test that should also make the need for the view->buf NULL check clearer.

[duaneg]

The issue is moot since this has been closed, but I'm afraid this still wouldn't have been useful. Your fix is unrelated to the test; it passes with or without it. It is good practice to always run your test scripts against affected versions, to confirm it reproduces the bug. Without that you don't know the test works.

Also note the bug in the OP is a SIGBUS, a fatal signal. It is completely expected that invalid operations like this raise exceptions, but it shouldn't be possible to trigger a fatal signal via the public API^{-ish. Terms and conditions apply. Offer void when ctypes or mmap or gc or various other low-level power tools.}

[da-woods]

I don't think this can be right. Partly because if it is you'd need to guard every use of view->buf and not just that one. And you'd probably have to do that in a thread-safe way. But also because memoryview owns view so nothing else should even have access to modify view and make it "no longer valid".

[picnixz]

If this is a legitimate issue, I believe this should be fixed as part of #143324. I will take care of that once I have time as well. More generally, I would prefer that you avoid addressing crash issues because many of your PRs are eventually closed and the issue is usually non-trivial so it is better to have a more seasoned contributor for those issues.