Re: [PATCH 08/11] rcu: don't use negative ->rcu_read_lock_nesting

[Lai Jiangshan <laijs@linux.alibaba.com>]

On 2019/11/18 5:53 上午, Paul E. McKenney wrote:
> On Sat, Nov 16, 2019 at 09:04:56PM +0800, Lai Jiangshan wrote:
>> On 2019/11/1 8:33 下午, Paul E. McKenney wrote:
>>> On Thu, Oct 31, 2019 at 10:08:03AM +0000, Lai Jiangshan wrote:
>>>> Negative ->rcu_read_lock_nesting was introduced to prevent
>>>> scheduler deadlock which was just prevented by deferred qs.
>>>> So negative ->rcu_read_lock_nesting is useless now and
>>>> rcu_read_unlock() can be simplified.
>>>>
>>>> And negative ->rcu_read_lock_nesting is bug-prone,
>>>> it is good to kill it.
>>>>
>>>> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
>>>> ---
>>>>    kernel/rcu/tree_exp.h    | 30 ++----------------------------
>>>>    kernel/rcu/tree_plugin.h | 21 +++++----------------
>>>>    2 files changed, 7 insertions(+), 44 deletions(-)
>>>>
>>>> diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
>>>> index c0d06bce35ea..9dcbd2734620 100644
>>>> --- a/kernel/rcu/tree_exp.h
>>>> +++ b/kernel/rcu/tree_exp.h
>>>> @@ -621,11 +621,11 @@ static void rcu_exp_handler(void *unused)
>>>>    	 * report the quiescent state, otherwise defer.
>>>>    	 */
>>>>    	if (!t->rcu_read_lock_nesting) {
>>>> +		rdp->exp_deferred_qs = true;
>>>>    		if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
>>>>    		    rcu_dynticks_curr_cpu_in_eqs()) {
>>>> -			rcu_report_exp_rdp(rdp);
>>>> +			rcu_preempt_deferred_qs(t);
>>>>    		} else {
>>>> -			rdp->exp_deferred_qs = true;
>>>>    			set_tsk_need_resched(t);
>>>>    			set_preempt_need_resched();
>>>>    		}
>>>> @@ -646,32 +646,6 @@ static void rcu_exp_handler(void *unused)
>>>>    		WRITE_ONCE(t->rcu_read_unlock_special.b.exp_hint, true);
>>>>    		return;
>>>>    	}
>>>> -
>>>> -	/*
>>>> -	 * The final and least likely case is where the interrupted
>>>> -	 * code was just about to or just finished exiting the RCU-preempt
>>>> -	 * read-side critical section, and no, we can't tell which.
>>>> -	 * So either way, set ->deferred_qs to flag later code that
>>>> -	 * a quiescent state is required.
>>>> -	 *
>>>> -	 * If the CPU is fully enabled (or if some buggy RCU-preempt
>>>> -	 * read-side critical section is being used from idle), just
>>>> -	 * invoke rcu_preempt_deferred_qs() to immediately report the
>>>> -	 * quiescent state.  We cannot use rcu_read_unlock_special()
>>>> -	 * because we are in an interrupt handler, which will cause that
>>>> -	 * function to take an early exit without doing anything.
>>>> -	 *
>>>> -	 * Otherwise, force a context switch after the CPU enables everything.
>>>> -	 */
>>>> -	rdp->exp_deferred_qs = true;
>>>> -	if (rcu_preempt_need_deferred_qs(t) &&
>>>> -	    (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
>>>> -	    WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs()))) {
>>>> -		rcu_preempt_deferred_qs(t);
>>>> -	} else {
>>>> -		set_tsk_need_resched(t);
>>>> -		set_preempt_need_resched();
>>>> -	}
>>>>    }
>>>>    /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
>>>> index dbded2b8c792..c62631c79463 100644
>>>> --- a/kernel/rcu/tree_plugin.h
>>>> +++ b/kernel/rcu/tree_plugin.h
>>>> @@ -344,8 +344,6 @@ static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
>>>>    }
>>>>    /* Bias and limit values for ->rcu_read_lock_nesting. */
>>>> -#define RCU_NEST_BIAS INT_MAX
>>>> -#define RCU_NEST_NMAX (-INT_MAX / 2)
>>>>    #define RCU_NEST_PMAX (INT_MAX / 2)
>>>>    /*
>>>> @@ -373,21 +371,15 @@ void __rcu_read_unlock(void)
>>>>    {
>>>>    	struct task_struct *t = current;
>>>> -	if (t->rcu_read_lock_nesting != 1) {
>>>> -		--t->rcu_read_lock_nesting;
>>>> -	} else {
>>>> +	if (--t->rcu_read_lock_nesting == 0) {
>>>>    		barrier();  /* critical section before exit code. */
>>>> -		t->rcu_read_lock_nesting = -RCU_NEST_BIAS;
>>>> -		barrier();  /* assign before ->rcu_read_unlock_special load */
>>>
>>> But if we take an interrupt here, and the interrupt handler contains
>>> an RCU read-side critical section, don't we end up in the same hole
>>> that resulted in this article when the corresponding rcu_read_unlock()
>>> executes?  https://lwn.net/Articles/453002/
>>
>> Hello, Paul
>>
>> I'm replying the email of V1, which is relying on deferred_qs changes
>> in [PATCH 07/11] (V1).
>> ([PATCH 04/11](V1) relies on it too as you pointed out)
>>
>> I hope I can answer the question wrt https://lwn.net/Articles/453002/
>> maybe partially.
>>
>> With the help of deferred_qs mechanism and the special.b.deferred_qs
>> bit, I HOPED rcu_read_unlock_special() can find if itself is
>> risking in scheduler locks via special.b.deferred_qs bit.
>>
>> --t->rcu_read_lock_nesting;
>> //outmost rcu c.s, rcu_read_lock_nesting is 0. but special is not zero
>> INTERRUPT
>>   // the fallowing code will normally be in_interrupt()
>>   // or NOT in_interrupt() when wakeup_softirqd() in invoke_softirq()
>>   // or NOT in_interrupt() when preempt_shedule_irq()
>>   // or other cases I missed.
>>   scheduler_lock()
>>   rcu_read_lock()
>>   rcu_read_unlock()
>>    // special has been set but with no special.b.deferred_qs
>>    rcu_read_unlock_special()
>>     raise_softirq_irqoff()
>>      wake_up() when !in_interrupt() // dead lock
>>
>> preempt_shedule_irq() is guaranteed to clear rcu_read_unlock_special
>> when rcu_read_lock_nesting = 0 before calling into scheduler locks.
>>
>> But, at least, what caused my hope to be failed was the case
>> wakeup_softirqd() in invoke_softirq() (which was once protected by
>> softirq in about 2 years between ec433f0c5152 and facd8b80c67a).
>> I don't think it is hard to fix it if we keep using
>> special.b.deferred_qs as this V1 series.
> 
> It is quite possible that special.b.deferred_qs might be useful
> for debugging.  But it should now be possible to take care of the
> nohz_full issue for expedited grace periods, which might in turn allow
> rcu_read_unlock_special() to avoid acquiring scheduler locks.
> 
> This could avoid the need for negative ->rcu_read_lock_nesting,
> in turn allowing your simplified _rcu_read_unlock().
> 
> Would you like to do the expedited grace-period modifications, or
> would you rather that I do so?
> 

Hello, Paul

To be honest, I didn't known there was special issue about
nohz_full with expedited grace periods until several days before
you told me. I just thought that it is requested to be expedited
so that we need to wake up something to handle it ASAP.

IOW, I'm not in a position to do the expedited grace-period
modifications before I learnt enough about it. I would be very
obliged that you do so. I believe it will be a better solution
than this one or the one in V2 relying on preempt_count.

Thanks
Lai