Fix race conditions for attribute creation

[GlenWalker]

• Attributes created by __onfirstaccess__ not seen by concurrent calls
• Attributes set based on importspec entries not seen by concurrent calls

[RonnyPfannschmidt]

I grasp the issue and recommend a slightly different approach

Instead of taking our on first access, it ough to be replaced with a threading. Event in a short locked block, the same should be done for the values that get taken and replaced

Then other threads may just wait for those to resolve and a Rlock is not needed

As only a single place is using the lock which is global, its OK to make it a private global variable

[GlenWalker]

Will have a go at alternate implementation

[GlenWalker]

Ok, had a first attempt at suggested implementation:

• Code is more complex, especially if still trying to support Python builds without thread support
• test_recursive_import triggers a deadlock by waiting on the event for recmodule.some (submod accesses recmodule.__dict__ which then iterates recmodule.__all__ and calls __makeattr for each attribute, which means __makeattr is called again for some)

[RonnyPfannschmidt]

then back to the initial solution it is, sorry for that one,

[GlenWalker]

then back to the initial solution it is, sorry for that one,

It was definitely worth a try. My first attempt at this problem tried to use __setattr__ to track which attributes, if any, had been set during __onfirstaccess__, but the code for that also ended up more complicated than I was happy with.

Is there anything else you'd like changed, or is the current state of this PR something that you'd be happy to merge?

[GlenWalker]

I haven't added anything to the changelog for this - its missing an entry for 2.0, but I haven't had time to figure out what the changes were. I'm guessing we'd make this 2.1.0 or 2.0.1, apart from fixing the race conditions all other behaviour should be the same as 2.0.0

[RonnyPfannschmidt]

thanks!