On Thu, May 02, 2013 at 11:19:12PM +0300, Julian Anastasov wrote:
>
> Hello,
>
> On Thu, 2 May 2013, Paul E. McKenney wrote:
>
> > > Only the new cond_resched_rcu() macro provides
> > > PREEMPT_ACTIVE flag to skip the rcu_preempt_sleep_check()
> > > call. The old macros provide locked=0 as you noticed. Does it
> > > answer your question or I'm missing something?
> >
> > PREEMPT_ACTIVE's value is usually 0x10000000. Did it change
> > since 3.9? If not, rcu_sleep_check(preempt_offset & PREEMPT_ACTIVE)
> > is the same as rcu_sleep_check(0).
>
> Yes, the different platforms use different bit,
> that is why I mentioned about my failed attempt at
> changing hardirq.h. PREEMPT_ACTIVE is always != 0.
>
> But I don't understand what do you mean by
> 'preempt_offset & PREEMPT_ACTIVE' being always 0.
> It is always 0 for cond_resched(), cond_resched_lock()
> and cond_resched_softirq(), not for cond_resched_rcu():
>
> (PREEMPT_ACTIVE | PREEMPT_RCU_OFFSET) & PREEMPT_ACTIVE
> should give PREEMPT_ACTIVE, not 0. We have 2 cases in
> rcu_sleep_check() for the if:
>
> 1. !(PREEMPT_ACTIVE) => FALSE for cond_resched_rcu
> 2. !(0) => TRUE for other cond_resched_* macros
>
> On x86 the code is:
>
> __might_sleep:
> pushl %ebp #
> testl $268435456, %ecx #, preempt_offset
> ...
> jne .L240 #,
> // rcu_lock_map checked when PREEMPT_ACTIVE is missing
> .L240:
> // rcu_bh_lock_map checked
OK, apologies -- I was looking at the calls to __might_sleep() in
mainline, and missed the one that you added. Revisiting that, a
question:
> +#ifdef CONFIG_PREEMPT_RCU
> +#define PREEMPT_RCU_OFFSET 1
Does this really want to be "1" instead of PREEMPT_OFFSET?
> +#else
> +#define PREEMPT_RCU_OFFSET PREEMPT_CHECK_OFFSET
> +#endif
> +
> +extern int __cond_resched_rcu(void);
> +
> +#define cond_resched_rcu() ({ \
> + __might_sleep(__FILE__, __LINE__, PREEMPT_ACTIVE | \
> + PREEMPT_RCU_OFFSET); \
> + __cond_resched_rcu(); \
> +})
> +
For the rest, I clearly need to revisit when more alert, because right
now I am not seeing the connection to preemptible RCU's rcu_read_lock()
implementation.
Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe lvs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
