From: Tom Lane <tgl@sss.pgh.pa.us>
To: exclusion@gmail.com
cc: pgsql-bugs@lists.postgresql.org
Subject: Re: BUG #16685: The ecpg thread/descriptor test fails sometimes on
 Windows
In-reply-to: <367782.1603554604@sss.pgh.pa.us>
References: <16685-d6cd241872c101d3@postgresql.org>
 <367782.1603554604@sss.pgh.pa.us>
Comments: In-reply-to Tom Lane <tgl@sss.pgh.pa.us>
	message dated "Sat, 24 Oct 2020 11:50:04 -0400"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <380143.1603558463.1@sss.pgh.pa.us>
Date: Sat, 24 Oct 2020 12:54:23 -0400
Message-ID: <380144.1603558463@sss.pgh.pa.us>
Precedence: bulk

I wrote:
> I'll bet that the correct fix is

>  static void
>  set_descriptors(struct descriptor *value)
>  {
> +	pthread_once(&descriptor_once, descriptor_key_init);
> 	pthread_setspecific(descriptor_key, value);
>  }

No, looking closer, there are only two calls of set_descriptors()
and each of them is preceded by a call to get_descriptors().
So it should not be possible to get here without descriptor_key
having been initialized.

I'm forced to the conclusion that there's something wrong with
ECPG's emulation of pthread_once ...

... and now that I look at it, it seems just as obvious what
is wrong there:

void
win32_pthread_once(volatile pthread_once_t *once, void (*fn) (void))
{
	if (!*once)
	{
		pthread_mutex_lock(&win32_pthread_once_lock);
		if (!*once)
		{
			*once = true;
			fn();
		}
		pthread_mutex_unlock(&win32_pthread_once_lock);
	}
}

We should not set *once until AFTER we execute fn().
Otherwise, other threads passing through pthread_once()
will mistakenly fall through, expecting the initialization
to be done already.

(So in this view, adding a sleep just before fn() would
make the failure more reproducible.)

			regards, tom lane