Re: [PATCH 2/2] sched: Move the sleeping while atomic checks early in cond_resched()

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Jul 10, 2009 at 04:59:55PM +0200, Peter Zijlstra wrote:
> On Fri, 2009-07-10 at 16:49 +0200, Frederic Weisbecker wrote:
> 
> > index 0cb0d8d..e357dc7 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -2279,11 +2279,13 @@ extern int _cond_resched(void);
> >  #ifdef CONFIG_PREEMPT_BKL
> >  static inline int cond_resched(void)
> >  {
> > +	might_sleep();
> >  	return 0;
> >  }
> >  #else
> >  static inline int cond_resched(void)
> >  {
> > +	might_sleep();
> >  	return _cond_resched();
> >  }
> >  #endif
> 
> # define might_resched() _cond_resched()


Argh, indeed.
I thought might_sleep() only wrapped __might_sleep(__FILE__, __LINE__)

 
> # define might_sleep() \
> 	do { __might_sleep(__FILE__, __LINE__); might_resched(); } while (0)
> 
> 
> Doesn't seem to make it any better that, but yeah, moving that
> __might_sleep() did occur to me earlier today when I touched that code.


Ok.

Another idea: if cond_resched() was a macro and __might_sleep() was
called inside, the given __FILE__ __LINE__ would be much more useful.

Only the backtraces would be useful in the current state, __FILE__
and __LINE__ point to sched.h, which is not exactly what is needed,
right?

Thanks,
Frederic.


> > diff --git a/kernel/sched.c b/kernel/sched.c
> > index 87ecac1..c22804b 100644
> > --- a/kernel/sched.c
> > +++ b/kernel/sched.c
> > @@ -6605,9 +6605,6 @@ SYSCALL_DEFINE0(sched_yield)
> >  
> >  static void __cond_resched(void)
> >  {
> > -#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
> > -	__might_sleep(__FILE__, __LINE__);
> > -#endif
> >  	/*
> >  	 * The BKS might be reacquired before we have dropped
> >  	 * PREEMPT_ACTIVE, which could trigger a second
> 
> 
>