Re: [PATCH 1/2] rcu: Delete a redundant check in rcu_check_gp_kthread_starvation()

[Paul E. McKenney]

On Tue, Jul 11, 2023 at 11:20:07AM +0800, Leizhen (ThunderTown) wrote:
> 
> 
> On 2023/7/11 3:03, Paul E. McKenney wrote:
> > On Wed, Jul 05, 2023 at 03:30:19PM +0800, Zhen Lei wrote:
> >> The above condition "if (gpk)" already ensures that gp_kthread is created,
> >> so the local variable 'cpu' cannot be negative here.
> >>
> >> Signed-off-by: Zhen Lei <thunder.leizhen@xxxxxxxxxx>
> >> ---
> >>  kernel/rcu/tree_stall.h | 12 +++++-------
> >>  1 file changed, 5 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
> >> index b10b8349bb2a48b..dcfaa3d5db2cbc7 100644
> >> --- a/kernel/rcu/tree_stall.h
> >> +++ b/kernel/rcu/tree_stall.h
> >> @@ -537,13 +537,11 @@ static void rcu_check_gp_kthread_starvation(void)
> >>  			pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
> >>  			pr_err("RCU grace-period kthread stack dump:\n");
> >>  			sched_show_task(gpk);
> >> -			if (cpu >= 0) {
> > 
> > I am not quite this trusting of the relation between the relationship
> > between the existence of the grace-period khread and its CPU number
> > being in range.  Let's please start with something like this:
> > 
> > 			if (!WARN_ON_ONCE(cpu < 0)) {
> > 
> > Please note that this is not just me.  See for example the use of the
> > cpumask_check() function, albeit the opposite concern.
> 
> git grep -wn "\->cpu" kernel/ include/
> kernel/kthread.c:583:   to_kthread(p)->cpu = cpu;				//kthread_create_on_cpu()
> kernel/sched/sched.h:2024:      WRITE_ONCE(task_thread_info(p)->cpu, cpu);	//__set_task_cpu()
> include/linux/sched.h:2250:     return READ_ONCE(task_thread_info(p)->cpu);	//task_cpu()
> 
> git grep -wn "\.cpu" kernel/ include/						//There is no task related, the search result is omitted.
> 
> Therefore, there is only one path "set_task_cpu()-->__set_task_cpu()" that can dynamically
> change the value of task_cpu(p). In fact, this guarantee has been made in set_task_cpu().
> set_task_cpu
> 	WARN_ON_ONCE(!cpu_online(new_cpu));
> 	__set_task_cpu(p, new_cpu);
> 
> In addition, task_struct has member 'on_rq'. Therefore, when a task leaves the scheduling
> queue, setting the member 'cpu' to an invalid value will be thankless.

Thank you for digging into this!  Given that, as you say, we can dispense
with the check.

> Sorry, these two patches was posted too quickly, and I'm still regretting that I should have
> attached this to the commit description these days.

Please do resend the patches with this explanation in the commit log.
And please don't worry about making the English pretty, as I can always
wordsmith.

							Thanx, Paul

> >> -				if (cpu_is_offline(cpu)) {
> >> -					pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
> >> -				} else  {
> >> -					pr_err("Stack dump where RCU GP kthread last ran:\n");
> >> -					dump_cpu_task(cpu);
> >> -				}
> >> +			if (cpu_is_offline(cpu)) {
> >> +				pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
> >> +			} else  {
> >> +				pr_err("Stack dump where RCU GP kthread last ran:\n");
> >> +				dump_cpu_task(cpu);
> >>  			}
> >>  			wake_up_process(gpk);
> >>  		}
> >> -- 
> >> 2.25.1
> >>
> > .
> > 
> 
> -- 
> Regards,
>   Zhen Lei