Re: [PATCH 3/3] posix timers: use SIGQUEUE_CANCELLED when the timer is destroyed

[Linus Torvalds <torvalds@linux-foundation.org>]

On Sun, 18 May 2008, Oleg Nesterov wrote:
> 
> Initially, I did
> 
> 	q->flags |= SIGQUEUE_CANCELLED;
> 	spin_lock_irqsave(lock, flags);
> 	q->flags &= ~SIGQUEUE_PREALLOC;
> 
> to document the fact that SIGQUEUE_CANCELLED can be set lockless, but
> then "optimized" the code, couldn't help myself... Besides, the code
> above looks really confusing without the fat comment.

Oh, and the above is *wrong*.

Why?

Becayse if SIGQUEUE_PREALLOC setting needs the lock, then setting any 
*other* bit in that word will also need the lock!

That's because

	q->flags |= SIGQUEUE_CANCELLED;

writes those other bits too - admittedly with the value they were read 
just before, but if it races with something setting SIGQUEUE_PREALLOC that 
doesn't matter - the newly written version will simply be wrong.

So the rule is that if one bit of a word needs locking, then they *all* 
do.

(On alpha, this is true even for whole bytes or shortwords - because a 
byte/shortword write is actually "read word, update byte/short, write 
word" sequence on older CPU's. So you cannot do atomic byte updates, and 
need to use locks).

			Linus