[lijinf2]

close #3282

Manual backport of #3308 to 12.9.x.

The callback registry shared by cuLogsRegisterCallback and cuDeviceRegisterAsyncNotification was not thread-safe. This adds locking so concurrent register/unregister works correctly.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[lijinf2]

/ok to test 24eeaf5

[lijinf2]

Hi Mike, could you review this 12.9.x backport of #3308 (global callback map lock)?

[mdboom]

Did you generate this from cybind? Generated files should always be handled that way to avoid things getting out of sync. It doesn't look like it updated the hash at the top of the file.

[lijinf2]

Did you generate this from cybind? Generated files should always be handled that way to avoid things getting out of sync. It doesn't look like it updated the hash at the top of the file.

@mdboom It's not generated, for two reasons: (1) 12.9.x uses a Tempita template (runtime.pyx.in), and I'm not sure how cybind can still produce that; (2) regenerating changes far more lines than this PR does, which would make it hard to review.

I updated the hash with grep -v 'CYTHON-BINDINGS-GENERATED' runtime.pyx.in | sha256sum and committed it. But please let me know if you'd still prefer going through cybind. If so, could you point me to how to generate the Tempita (.pyx.in) files with it?

[lijinf2]

/ok to test 86320d7