PyErr_CheckSignals should be callable without holding the GIL #133465
Labels
interpreter-core
(Objects, Python, Grammar, and Parser dirs)
topic-C-API
type-feature
A feature request or enhancement
Comments
|
Please see https://research.owlfolio.org/pubs/2025-pyext-ctrlc-talk/ for a "decompressed" rationale for this change, with analysis of the status quo, performance measurements, and discussion. |
|
Discussion in #133466 wound up requesting that design issues be forwarded to the C-API working group; that thread is linked above. I have also posted https://discuss.python.org/t/ergonomics-of-signal-checks-with-detached-thread-state/94355 soliciting feedback from extension developers. |
zackw
added a commit
to MillionConcepts/cpython-patches
that referenced
this issue
Jun 10, 2025
Add new C-API functions `PyErr_CheckSignalsDetached` and `PyErr_AreSignalsPending`. `PyErr_CheckSignalsDetached` can *only* be called by threads that *don’t* have an attached thread state. It does the same thing as `PyErr_CheckSignals`, except that it guarantees it will only reattach the supplied thread state if necessary in order to run signal handlers. (Also, it never runs the cycle collector.) `PyErr_AreSignalsPending` can be called with or without an attached thread state. It reports to its caller whether signals are pending, but never runs any handlers itself. Rationale: Compiled-code modules that implement time-consuming operations that don’t require manipulating Python objects, are supposed to call PyErr_CheckSignals frequently throughout each such operation, so that if the user interrupts the operation with control-C, it is canceled promptly. In the normal case where no signals are pending, PyErr_CheckSignals is cheap (two atomic memory operations); however, callers must have an attached thread state, and compiled-code modules that implement time-consuming operations are also supposed to detach their thread state during each such operation. The overhead of re-attaching a thread state in order to call PyErr_CheckSignals, and then releasing it again, sufficiently often for reasonable user responsiveness, can be substantial, particularly in traditional (non-free-threaded) builds. These new functions permit compiled-code modules to avoid that extra overhead.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Feature or enhancement
Proposal:
Compiled-code modules that implement time-consuming operations that don’t require manipulating Python objects, are supposed to call
PyErr_CheckSignalsfrequently throughout each such operation, so that if the user interrupts the operation with control-C, it is cancelled promptly.In the normal case where no signals are pending,
PyErr_CheckSignalsis cheap; however, callers must hold the GIL, and compiled-code modules that implement time-consuming operations are also supposed to release the GIL during each such operation. The overhead of reclaiming the GIL in order to callPyErr_CheckSignals,and then releasing it again, sufficiently often for reasonable user responsiveness, can be substantial.
If my understanding of the thread-state rules is correct, PyErr_CheckSignals only needs the GIL if it has work to do. Checking whether there is a pending signal, or a pending request to run the cycle collector, requires only a couple of atomic loads. Therefore, I propose that we should restructure
PyErr_CheckSignalsand its close relatives (_PyErr_CheckSignalsand_PyErr_CheckSignalsTstate) so that all the “do we have anything to do” checks are done in a batch before anything that needs the GIL. If any of them are true, then the function should acquire the GIL itself, repeat the check (because another thread could have stolen the event while we were waiting for the GIL), and then actually do the work, thus enabling callers to not hold the GIL.I have already prepared and tested an implementation of this change and will submit a PR shortly.
Has this already been discussed elsewhere?
This is a minor feature, which does not need previous discussion elsewhere
Links to previous discussion of this feature:
No response
Linked PRs
The text was updated successfully, but these errors were encountered: