Re: [PATCH 2/2] x86/intel_rdt: Update task closid immediately on CPU in rmdir and unmount

From: Fenghua Yu
Date: Sat Nov 26 2016 - 16:07:15 EST


On Sat, Nov 26, 2016 at 10:08:57AM +0100, Thomas Gleixner wrote:
> On Fri, 25 Nov 2016, Fenghua Yu wrote:
> > On Wed, Nov 23, 2016 at 03:23:50PM +0100, Thomas Gleixner wrote:
> > > +#ifdef CONFIG_SMP
> > > + /*
> > > + * This is safe on x86 w/o barriers as the ordering
> > > + * of writing to task_cpu() and t->on_cpu is
> > > + * reverse to the reading here. The detection is
> > > + * inaccurate as tasks might move or schedule
> > > + * before the smp function call takes place. In
> > > + * such a case the function call is pointless, but
> > > + * there is no other side effect.
> > > + */
> >
> > If process p1 is running on CPU1 before this point,
> >
> > > + if (mask && t->on_cpu)
> > > + cpumask_set_cpu(task_cpu(t), mask);
> >
> > If between CPU1 is set in mask and rdt_update_closid(tmpmask, NULL) is
> > called, p1 is switched to CPU2, and process p2 with its own closid
> > (e.g. 2) is switched to CPU1.
> >
> > Then closid in PQR_ASSOC is set incorrectly as 0 instead of 2 on CPU1.
> > 0 may stay in PQR_ASSOC until next context switch which may take long time
> > in cases of real time or HPC.
> >
> > Don't we need to care this situation? In this situation, the function call
> > is not "pointless" but it's wrong, right?
>
> No.
>
> CPU0 CPU1 CPU2
> T1 (closid 0) T2 (closid 2)
>
> (t1->on_cpu)
> set(1, mask)
> preemption
> T1 ->CPU2
> switch_to(T3) preemption
> switch_to(idle)
> T2 -> CPU1
> switch_to(T2) switch_to(T1)
> intel_rdt_sched_in() intel_rdt_sched_in()
> closid = T2->closid closid = T1->closid
> closid =2 closid = CPU2->closid
>
> rdt_update_closid(mask)
>
> rdt_update_cpu_closid()
> intel_rdt_sched_in()
> closid = T2->closid
> closid = 2
>
> IOW, whatever comes first, sched_switch() or function call will update the
> closid to the correct value.
>
> If CPU2 was in the removed group then this looks the following way:
>
> CPU0 CPU1 CPU2
> T1 (closid 0) T2 (closid 2)
>
> (t1->on_cpu)
> set(1, mask)
> preemption
> T1 ->CPU2
> switch_to(T3) preemption
> switch_to(idle)
> T2 -> CPU1
> switch_to(T2) switch_to(T1)
> intel_rdt_sched_in() intel_rdt_sched_in()
> closid = T2->closid closid = T1->closid (0)
> closid =2 closid = CPU2->closid
> closid = 5
> for_each_cpu(grp->mask)
> CPU2->closid = 0
>
> rdt_update_closid(mask)
>
> rdt_update_cpu_closid() rdt_update_cpu_closid()
> intel_rdt_sched_in( intel_rdt_sched_in()
> closid = T2->closid closid = T1->closid (0)
> closid = 2 closid = CPU2->closid
> closid = 0
>
> But on CPU2 the function call might be pointless as well in the following
> situation:
>
> CPU0 CPU1 CPU2
> T1 (closid 0) T2 (closid 2)
>
> (t1->on_cpu)
> set(1, mask)
> preemption
> T1 ->CPU2
> switch_to(T3) preemption
> switch_to(idle)
>
> for_each_cpu(grp->mask)
> CPU2->closid = 0
> T2 -> CPU1
> switch_to(T2) switch_to(T1)
> intel_rdt_sched_in() intel_rdt_sched_in()
> closid = T2->closid closid = T1->closid (0)
> closid =2 closid = CPU2->closid
> closid = 0
>
> rdt_update_closid(mask)
>
> rdt_update_cpu_closid() rdt_update_cpu_closid()
> intel_rdt_sched_in( intel_rdt_sched_in()
> closid = T2->closid closid = T1->closid (0)
> closid = 2 closid = CPU2->closid
> closid = 0
>
> The whole thing works by ordering:
>
> 1) Update closids of each task in the group and if a task is running on a
> cpu then mark the CPU on which the task is running for the function
> call.
>
> 2) Update closids of each CPU in the group
>
> 3) Or the cpumasks of the tasks and the groups and invoke the function call
> on all of them
>
> If an affected task does a sched_switch after task->closid is updated and
> before the function call is invoked then the function call is pointless.
>
> If a sched switch happens on a CPU after cpu->closid is updated and before
> the function call is invoked then the function call is pointless.
>
> But there is no case where the function call can result in a wrong value.
>
> Thanks,
>
> tglx

Thank you for your clear explanation. You are absolutely correct. I know
I must miss something.

The reworked second patch and the first patch were tested successfully.
I assume you will check them in tip tree and I will not send v2.

Thanks.

-Fenghua