Re: [RFC] RT-patch update to remove the global pi_lock

[Ingo Molnar <mingo@elte.hu>]

* Steven Rostedt <rostedt@goodmis.org> wrote:

> Here it is a little ambiguous.  The process use to own the lock, but 
> someone stole it.  When grabbing a lock, I always grab the process 
> lock first before grabbing the lock's lock, but this isn't necessary.  
> So if you already have the two locks (mutex and process) as is the 
> case above, you don't need to unlock them and regrab them (although 
> this doesn't hurt, except for performance), because any race would 
> have been with the grabbing of these two locks in the first place.
> 
> Now, here's why it's safe.  The process that use to own the mutex and 
> had it stolen now is out of the chain.  The pi_lock and the wait_lock 
> here are not in any order.  So no one who is grabbing the process' 
> pi_lock should have owned the wait_lock and vice versa.
> 
> So here's an updated patch without this lock switching:

your patch works great here, on 3 separate systems: a 1-way, a 2/4-way 
and an 8-way.

the 1-way system performed so well running the SMP kernel that i first 
thought i booted the UP kernel by accident :-)

on the 8-way box, "hackbench 10" got _3.7_ times faster (!!!).

i have booted the 8-way box without your patch once more because i didnt 
believe the results initially and thought they were some benchmarking 
fluke. But no, it wasnt a fluke. The kernel profiles are nicely flat 
now.

i've released 2.6.13-rc7-rt2 with your patch included. This is certainly 
a major milestone for PREEMPT_RT, it is now a first-class scalability 
citizen on SMP too. Great work Steven!

	Ingo
-
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/