Re: [patch] scheduler fixes, 2.5.32-BK

[Linus Torvalds]

On Fri, 30 Aug 2002, Ingo Molnar wrote:
> 
> it touches the waitqueue spinlock - and the __down() path [ie. the process
> that gets woken up, which has the semaphore on the stack] takes the
> spinlock after waking up. Ie. there's guaranteed synchronization, the
> semaphore will not be 'unused' before the __down() path takes the spinlock
> - ie. after the __up() path releases the spinlock. What am i missing?

So why couldn't this happen? This is what used to happen before, I don't 
see that consolidating the spinlock had any impact at all.

	CPU #0						CPU #1

	down()						up()

		lock decl (negative)
		__down()				lock incl
			spinlock()			__up()
			atomic_add_negative()
				success - break
			spinunlock();
		}					wake_up()
	return - semaphore is now invalid		spin_lock()

							BOOM!

The fact is, that as long as down() and up() avoid taking the spinlock 
_before_ they touch "count", they aren't synchronized. 

And we definitely do _not_ want to take the spinlock before we touch 
count, since that would make the fast path a lot slower.

What?

		Linus