Ahhhh! Great spot, yes this was a dumb race condition.

Returning the error from vips_threadpool_free() feels like two ideas are being conflated. How about adding vips_threadpool_wait() to wait for all threads to finish, and calling that from vips_threadpool_free()? Then the end of vips_threadpool_run() can wait for all threads to finish, get any error, and free the threadpool. You wouldn't need to handle errors in the free function.

Good idea. Addressed in the latest revision.

Do you need the wait param? It's fine to call vips_threadpool_free() many times, I think.

I wasn't sure about that one, let me try to remove it. It was inspired by g_thread_pool_free().

... you're right, it doesn't matter that vips_threadpool_wait() is now called twice. It makes the diff also easier to read now. :)

Looks great!