Re: kernel thread exit race

[Russell King <rmk@arm.linux.org.uk>]

On Wed, Aug 07, 2002 at 03:30:17PM +0400, Nikita Danilov wrote:
> So, complete() is not-arch dependent because spinlocks are "good" in all
> architectures? complete() ends with spin_unlock_irqrestore() so it
> cannot be any better than spinlocks, until there is some hidden magic.

It works like this:

	cpu1					cpu2
kill thread (on cpu2)
					complete_and_exit()
					- takes spinlock
wait_for_completion()
- spins on completion spinlock
					- increments x->done
					- wakes up anyone waiting on
					  the completion
					- releases spinlock
- checks x->done
- decrements x->done
- releases spinlock

OR:

	cpu1					cpu2
kill thread (on cpu2)
wait_for_completion()
- takes spinlock
					complete_and_exit()
					- spins on spinlock
- checks x->done
- adds to waitqueue
- releases spinlock
					- increments x->done
- sleeps
					- wakes up anyone waiting on
					  the completion
- wakes up
- spins on spinlock
					- releases spinlock
- decrements x->done
- releases spinlock

OR:

	cpu1					cpu2
kill thread (on cpu2)
wait_for_completion()
- takes spinlock
- checks x->done
- adds to waitqueue
- releases spinlock
- sleeps
					complete_and_exit()
					- takes spinlock
					- increments x->done
					- wakes up anyone waiting on
					  the completion
- wakes up
- spins on spinlock
					- releases spinlock
- decrements x->done
- releases spinlock

As you can see, wait_for_completion() will never return until complete()
has released the spinlock.

-- 
Russell King (rmk@arm.linux.org.uk)                The developer of ARM Linux
             http://www.arm.linux.org.uk/personal/aboutme.html