Re: [patch 00/15] Generic Mutex Subsystem

[Oleg Nesterov <oleg@tv-sign.ru>]

Balbir Singh wrote:
>
> On Wed, Dec 21, 2005 at 07:42:14PM +0300, Oleg Nesterov wrote:
> >
> > Note that subsequent up() will not call wakeup(): ->count == 0,
> > it just increment it. That is why we are waking the next waiter
> > in advance. When it gets cpu, it will decrement ->count by 1,
> > because ->sleepers == 0. If up() (++->count) was already called,
> > it takes semaphore. If not - goes to sleep again.
> >
> > Or my understanding is completely broken?
>
> [ ... long snip ... ]
>
> The question now remains as to why we have the atomic_add_negative()? Why do
> we change the count in __down(), when down() has already decremented it
> for us?

... and why __down() always sets ->sleepers = 0 on return.

I don't have an answer, only a wild guess.

Note that if P1 releases this semaphore before pre-woken P2 actually
gets cpu what happens is:

	P1->up() just increments ->count, no wake_up() (fastpath)

	P2 takes the semaphore without schedule.

So *may be* it was designed this way as some form of optimization,
in this scenario P2 has some chances to run with sem held earlier.

Oleg.