Fix native module loading in multi-context (--experimental-workers)

[rpetrich]

Fixes native module loading from multiple contexts by saving the module metadata provided by native modules when they call node_module_register. Also fixes a race condition when multiple contexts happen to load a module at the same time.

Checklist

• make -j4 test (UNIX), or vcbuild test (Windows) passes
• tests and/or benchmarks are included
• commit message follows commit guidelines

[rpetrich]

Not sure if further work should be done to cause dlopen followed by require of a native module to fail, or if addons/dlopen-ping-pong should be updated to match the new behaviour.

Also, it may be worthwhile to introduce new metadata in the node_module struct denoting whether or not a module is capable of being loaded in a alternate thread. Many native modules expect to call back into node from uv_default_loop, and this fails in a very confusing way when a native module is called from a worker.

[bnoordhuis]

See #19731 (review). This PR has the same issue.

[rpetrich]

Thanks, @bnoordhuis. Do you have any opinions on whether the well-known symbol is the recommended approach for multi-context aware modules? Unless all modules that use well-known are expected to be multi-context aware, it suffers from the same problem as you've described in #19731.

Should I split out the race condition fix into a separate PR? Also, perhaps it makes sense to raise an error if a multi-context unaware module is loaded in an alternate context? (though that would be a breaking change—as it is now, some of these modules can be used from a single worker, provided they don't reference the uv_default_loop in an unsafe way)

[bnoordhuis]

Do you have any opinions on whether the well-known symbol is the recommended approach for multi-context aware modules?

I think it could be. It's new enough that we can simply say it should be multi-thread ready.

perhaps it makes sense to raise an error if a multi-context unaware module is loaded in an alternate context?

No strong opinion except that it's a breaking change. That's something to avoid unless there's a very clear benefit/cost ratio.