Thread safety of PyCapsule

[eendebakpt]

Bug report

Bug description:

The PyCapsule object is not thread-safe. For example the read and write operations to the underlying pointer do not use atomics.

cpython/Objects/capsule.c

Line 111 in e094544

return capsule->pointer;

This can be fixed with atomic read and write operations. (same for the name field and some others)

Another potential issue is that some operations on a capsule are not atomic. For example a thread can attempt to retrieve a pointer using PyCapsule_GetPointer while a concurrent thread is updating the name and pointer using PyCapsule_SetName and PyCapsule_SetPointer. That could be addressed with locks, but I suspect that is too heavy for the capsule object.

I am leaning towards adding atomics for the read and write operations and documenting that in a PyCapsule does not guarantee its methods are atomic.

@colesbury @seberg

CPython versions tested on:

CPython main branch

Operating systems tested on:

Windows

[colesbury]

It seems fine to me to require that callers to the capsule API are required to do their own synchronization.

[eendebakpt]

It seems fine to me to require that callers to the capsule API are required to do their own synchronization.

To be clear: does that mean we also do not add atomic reads/writes to the pointer in PyCapsule_GetPointer?

[colesbury]

Yes