Re: [-mm PATCH] RCU: debug sleep check

[Dipankar Sarma <dipankar@in.ibm.com>]

On Sun, Sep 24, 2006 at 02:44:40PM -0700, Andrew Morton wrote:
> On Mon, 25 Sep 2006 03:05:08 +0530
> Dipankar Sarma <dipankar@in.ibm.com> wrote:
> 
> > On Sun, Sep 24, 2006 at 11:56:46AM -0700, Andrew Morton wrote:
> > > 
> > > Does this actually change anything?  rcu_read_lock is preempt_disable(), and
> > > might_sleep() already triggers if called inside preempt_disable().
> > 
> > It makes a difference if CONFIG_PREEMPT=n. AFAICS, preempt_disable()
> > is a nop then and rcu needs its own check for sleeping while
> > in read-side critical section.
> 
> Right.  I think enough people run with CONFIG_PREEMPT=y to make this
> change rather unnecessary.

I would mostly agree except that for the rcupreempt (-rt) implementation
we don't use preempt_disable/enable in rcu read-side critical section.
So, I have to add a rcu_read_in_atomic() API anyway and there is
no harm in adding the same for rcuclassic (current default)
so that sleeping-while-rcu-atomic check happens irrespective
of CONFIG_PREEMPT. It is included only if CONFIG_DEBUG_SPINLOCK_SLEEP=y.

> 
> And if there are developers out there who are testing their code without an
> effective CONFIG_DEBUG_SPINLOCK_SLEEP, then *that* is what we need to fix,
> no?

Yes. With this patch, I am relying on people to test their
code with CONFIG_DEBUG_SPINLOCK_SLEEP so that we can catch
bad rcu users.

Thanks
Dipankar