Re: [PATCH v2] isolcpus: affine kernel threads to specified cpumask

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Mar 26, 2020 at 05:20:05PM +0100, Frederic Weisbecker wrote:
> On Wed, Mar 25, 2020 at 08:47:36AM -0300, Marcelo Tosatti wrote:
> > 
> > Hi Frederic,
> > 
> > On Wed, Mar 25, 2020 at 01:30:00AM +0100, Frederic Weisbecker wrote:
> > > On Tue, Mar 24, 2020 at 12:20:16PM -0300, Marcelo Tosatti wrote:
> > > > 
> > > > This is a kernel enhancement to configure the cpu affinity of kernel
> > > > threads via kernel boot option isolcpus=no_kthreads,<isolcpus_params>,<cpulist>
> > > > 
> > > > When this option is specified, the cpumask is immediately applied upon
> > > > thread launch. This does not affect kernel threads that specify cpu
> > > > and node.
> > > > 
> > > > This allows CPU isolation (that is not allowing certain threads
> > > > to execute on certain CPUs) without using the isolcpus=domain parameter,
> > > > making it possible to enable load balancing on such CPUs
> > > > during runtime (see
> > > > 
> > > > Note-1: this is based off on Wind River's patch at
> > > > https://github.com/starlingx-staging/stx-integ/blob/master/kernel/kernel-std/centos/patches/affine-compute-kernel-threads.patch
> > > > 
> > > > Difference being that this patch is limited to modifying
> > > > kernel thread cpumask: Behaviour of other threads can
> > > > be controlled via cgroups or sched_setaffinity.
> > > > 
> > > > Note-2: MontaVista's patch was based off Christoph Lameter's patch at
> > > > https://lwn.net/Articles/565932/ with the only difference being
> > > > the kernel parameter changed from kthread to kthread_cpus.
> > > > 
> > > > Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
> > > 
> > > I'm wondering, why do you need such a boot shift at all when you
> > > can actually affine kthreads on runtime?
> > 
> > New, unbound kernel threads inherit the cpumask of kthreadd.
> > 
> > Therefore there is a race between kernel thread creation 
> > and affine.
> > 
> > If you know of a solution to that problem, that can be used instead.
> 
> Well, you could first set the affinity of kthreadd and only then the affinity
> of the others. But I can still imagine some tiny races with fork().
> 
> > > 
> > > >  };
> > > >  
> > > >  #ifdef CONFIG_CPU_ISOLATION
> > > > diff --git a/kernel/kthread.c b/kernel/kthread.c
> > > > index b262f47046ca..be9c8d53a986 100644
> > > > --- a/kernel/kthread.c
> > > > +++ b/kernel/kthread.c
> > > > @@ -347,7 +347,7 @@ struct task_struct *__kthread_create_on_node(int (*threadfn)(void *data),
> > > >  		 * The kernel thread should not inherit these properties.
> > > >  		 */
> > > >  		sched_setscheduler_nocheck(task, SCHED_NORMAL, &param);
> > > > -		set_cpus_allowed_ptr(task, cpu_all_mask);
> > > > +		set_cpus_allowed_ptr(task, cpu_kthread_mask);
> > > 
> > > I'm wondering, why are we using cpu_all_mask and not cpu_possible_mask here?
> > > If we used the latter, you wouldn't need to create cpu_kthread_mask and
> > > you could directly rely on housekeeping_cpumask(HK_FLAG_KTHREAD).
> > 
> > I suppose that either work: CPUs can only be online from
> > cpu_possible_mask (and is contained in cpu_possible_mask).
> > 
> > Nice cleanup, thanks.
> 
> But may I suggest you to do:
> 
> -         set_cpus_allowed_ptr(task, cpu_all_mask);
> +         set_cpus_allowed_ptr(task, cpu_possible_mask);
> 
> as a first step in its own patch in the series. I just want to make sure that change
> isn't missed by reviewers or bisections, in case someone catches something we
> overlooked.
> 
> > 
> > > 
> > > > diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
> > > > index 008d6ac2342b..e9d48729efd4 100644
> > > > --- a/kernel/sched/isolation.c
> > > > +++ b/kernel/sched/isolation.c
> > > > @@ -169,6 +169,12 @@ static int __init housekeeping_isolcpus_setup(char *str)
> > > >  			continue;
> > > >  		}
> > > >  
> > > > +		if (!strncmp(str, "no_kthreads,", 12)) {
> > > > +			str += 12;
> > > > +			flags |= HK_FLAG_NO_KTHREADS;
> > > 
> > > You will certainly want HK_FLAG_WQ as well since workqueue has its own
> > > way to deal with unbound affinity.
> > 
> > Yep. HK_FLAG_WQ is simply a convenience so that the user does not have
> > to configure this separately: OK.
> 
> Also, and that's a larger debate, are you interested in isolating kthreads
> only or any kind of kernel unbound work that could be affine outside
> a given CPU?

Any kind of kernel work.

> In case of all the unbound work, I may suggest an all-in-one "unbound"
> flag that would do:
> 
>     HK_FLAG_KTHREAD | HK_FLAG_WQ | HK_FLAG_TIMER | HK_FLAG_RCU | HK_FLAG_MISC
>         | HK_FLAG_SCHED
> 
> Otherwise we can stick with HK_FLAG_KTHREAD, but I'd be curious about your usecase.
> 
> Thanks.

BTW HK_FLAG_SCHED is not settable at the moment.

Any reason why nohz_full= is not setting it ?

Thanks