The import lock is lacking tests that it functions properly.

Coming up with a way of checking it was unexpectedly hard. I finally managed to write a script which quite reliably fails if you patch Python/import.c to not take the import lock. The script relies on garbage collection trickery to release the GIL at arbitrary moments.
It should probably be converted to an unit test.

test_threaded_import is designed to check the import lock (and it does
at least to some degree - I changed runpy's handling of the import
lock because my original approach broke that test).

(Haven't looked at this patch in detail as yet)

> test_threaded_import is designed to check the import lock (and it does
> at least to some degree - I changed runpy's handling of the import
> lock because my original approach broke that test).

Apparently, Lib/test/test_threaded_import only works if run directly. If
run through regrtest, it succeeds even with the import lock disabled
(probably because random.py is already included by regrtest).

Ok, here is a patch for test_threaded_import that makes it work for regrtest too. I've removed all global variables and converted it to unittest.

I've committed the fix in r82885 (3.2) and r82886 (3.1). The reliance on random still looks a bit quirky to me, but at least the test now does what it should do, and has a cleaned up coding style.

I'm leaving this issue open, for the other test might be interesting to integrate as well.

Here is another patch that also tests that calls to path hooks and meta_path entries are serialized. Again, they pass in normal conditions and fail when the import lock is disabled.

Ping?

With the new setUp and tearDown methods, the threadedimp2 patch doesn't apply cleanly any more.

Looks good and passes for me - fixed patch attached.

Was I meant to still be looking at mtimport or threadimp, or have both of those been applied?

> Was I meant to still be looking at mtimport or threadimp, or have both 
> of those been applied?

threadimp has already been committed.
As for mtimport, I'll try to refactor it into a proper unit test.

The new patch was committed in r84258, thanks Nick.