Formalize linetable parsing and propagate out-of-bounds errors

[noahbkim]

We've noticed a transient issue where py-spy will abort due to unmet assumptions about the contents of co_linetable (for example we've seen [231], which doesn't even satisfy the internal specification). To address this, I've patched the co_linetable parsing to propagate out-of-bounds errors so that the samples are simply ignored (and a warning is raised).

This change is not as minimal as it could be, for example it reorders a couple lines in python_interpreters.rs for clarity. If that's a problem, I'd be happy to go back and pare down such changes.

Taking a glance, I would guess this is the root cause for #735 and #569, although I have no explanation for why the co_linetable's appear invalid.

[noahbkim]

Updated the PR to include a fix for an off-by-one introduced by #635, trivially reproducible with the following:

# a.py
import b

# b.py
import time
time.sleep(10)

# Command
cargo run -F unwind --release -- record --native --rate 1 -- python3.12 a.py

[korniltsev]

This change is not as minimal as it could be, for example it reorders a couple lines in python_interpreters.rs for clarity. If that's a problem, I'd be happy to go back and pare down such changes.

Let's change only what's needed for the fix, so that it's easier to cherry-pick & revert & reason about

[Hnasar]

@benfred can approve this workflow so the build can run?

Otherwise this LGTM, and we've manually built it and confirmed it's fixed the issues we were seeing.

[benfred]

@benfred can approve this workflow so the build can run?

I believe that this branch needs to be updated to the latest master first (both to get some CI fixes in - and to let me approve the workflow run) - and I don't seem to have permissions on this branch to update it myself =(

[korniltsev]

I suggest we also fix int overflows either in the same PR or in a followup, so that debug builds of pyspy will not panic

[noahbkim]

Updated! @benfred thanks for taking a look.
@korniltsev can you point to the overflows to fix?

[korniltsev]

@korniltsev can you point to the overflows to fix?

Basically every time some uncontrolled user-provided number (from the linetable) is added, subtracted, or shifted, etc.
You can search for "checked" in this diff https://github.com/grafana/py-spy/pull/1/files
Fuzzing is a great and easy way to find some of those overflows.

Btw, I am not a maintainer here, this is just a suggestion for improvements. Maybe even better to do as a separate change. Feel free to ignore.

Thank you for working on this!

[benfred]

Updated the PR to include a fix for an off-by-one introduced by #635, trivially reproducible with the following:

I believe this is fixed by #751 - can you verify if this still happens with the latest code?

Also, There was also an alternative fix for the same linetable problem in #736 - I've merged that one since it passed CI and I want to get a release out.

I like the unittests that you've added here though, any chance that you can merge in the other PR here so we can get those tests included?