Re: a question about __down() in Linux/arch/i386/kernel/semaphore.c

[Bob Miller <rem@osdl.org>]

On Mon, Aug 05, 2002 at 10:42:37PM +0800, Edward Shao \(ªòªv°ê\) wrote:
> sorry, i found it!
> wake_up_locked(&sem->wait);
> but why do we need to wake up the sleepers again?
> Thank you very much.
> 
> -Edward Shao-
> 
> ----- Original Message -----
> From: "Edward Shao (ªòªv°ê)" <szg90@cs.ccu.edu.tw>
> To: <linux-kernel@vger.kernel.org>
> Sent: Monday, August 05, 2002 10:33 PM
> Subject: a question about __down() in Linux/arch/i386/kernel/semaphore.c
> 
> 
> > Hi,
> >
> > I have a question about __down() in kernel 2.4.18
> > (Linux/arch/i386/kernel/semaphore.c)
> > I found the last line of __down() is
> > wake_up(&sem->wait);
> > but in kernel 2.5.28, i didn't see this line..
> > is this line necessary in kernel 2.4.18?
> > why?
> >
> > Thank you very much.
> >
> > Best Regard!!!
> >
> > -Edward Shao-
> >
> >

The quick answer: so we don't miss waking someone up.  But, seriously,
the semaphore code is very subtle.

This semaphore implementation allows more than one process to be in the
critical section at a time (a.k.a. a counting semaphore).  In order to
support those semantics, more than one wakeup may occur before a process
is pulled off the wake_q and changed to running.  Because the process that
is waiting to run (in the __down() code) is responsible for pulling itself 
off the wait_q, if the 2 __up()s happen before the __down() can finish,
the 2 __up()s will wakeup the same process twice.  So, the __down() code
needs to protect agaist this.

-- 
Bob Miller					Email: rem@osdl.org
Open Source Development Lab			Phone: 503.626.2455 Ext. 17