Re: [patch] Real-Time Preemption, -RT-2.6.12-rc1-V0.7.41-07

[Steven Rostedt <rostedt@goodmis.org>]

On Wed, 2005-03-30 at 08:50 +0200, Ingo Molnar wrote:
> * Steven Rostedt <rostedt@goodmis.org> wrote:
> 
> > OK, I'm declaring defeat here. I've been fighting race conditions all 
> > day, and it's now 1 in the morning where I live. It looks like this 
> > implementation has no other choice but to have the waking up "pending 
> > owner" take the wait_list lock once again.  How heavy of a overhead is 
> > that really?
> 
> as i mentioned it before, taking a lock is not a big issue at all. Since 
> you have to touch the lock data structure anyway (and all of it fits 
> into a single cacheline), it doesnt really matter whether it's atomic 
> flag setting/clearing, or raw spinlock based.c0267ad4
> 

OK, I've implemented adding the write_lock, but I still have a nasty
race condition, and I finally figured out why!

The damn BKL!  Here's the situation I'm having.

Process A grabs BKL then tries to grab another semaphore (say sem X)
But process B has sem X so process A sleeps.
Since semaphores don't have the restriction of saving the BKL, the BKL
   gets released from process A.
Process C comes along and grabs the BKL.
Finally B releases sem X and process A becomes the new "pending owner" 
  and wakes up.
When A schedules in it tries to grab the BKL but blocks. Now it is at
  the point where A doesn't actually own either the BKL or sem X. 
When C releases the BKL and gives it to A, A is now the pending owner
  of both the BKL and sem X.

When this occurs, all hell breaks loose. 

I believe I can solve this by making the BKL a special case and not
implement the pending owner at all for it.

> later on, once things are stable and well-understood, we can still 
> attempt to micro-optimize it.
> 

Heck, I'll make it bloat city till I get it working, and then tone it
down a little :-)  And maybe later we can have a better solution for the
BKL.

-- Steve