On Thu, 21 Jun 2007, Ingo Molnar wrote:
quoted text
> 
> btw., back then we also tried a spin_is_locked() based inner loop but it 
> didnt help the ->tree_lock lockups either. In any case i very much agree 
> that the 'nicer' looping should be added again - the patch below does 
> that. (build and boot tested)

Ok, I'm definitely not going to apply it right now, though.

quoted text
> and the reason that this didnt help the ->tree_lock lockup is likely the 
> same why wait_task_inactive() broke _independently_ of the 'niceness' of 
> the spin-lock operation: there were too few instructions between 
> releasing the lock and re-acquiring it again can cause permanent 
> starvation of another CPU. No amount of logic on the spinning side can 
> overcome this, if acquire/release critical sections are following each 
> other too fast.

Exactly.

The only way to handle that case is to make sure that the person who 
*gets* the spinlock will slow down. The person who doesn't get it can't do 
anything at all about the fact that he's locked out.

A way to do that (as already mentioned) is to have a "this lock is 
contended" flag, and have the person who gets the lock do something about 
it (where the "something" might actually be as simple as saying "When I 
release a lock that somebody marked as having lots of contention, I will 
clear the contention flag, and then just delay myself").

Side note: that trivial approach only really helps for a *single* thread 
that gets it very much (like the example in wait_task_inactive). For true 
contention with multiple different CPU's that can *all* have the bad 
behaviour, you do actually need real queueing.

			Linus
-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/