lkml.org 
[lkml]   [2021]   [Nov]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 4/6] rcu/nocb: Create nocb kthreads on all CPUs as long as the "rcu_nocb=" is passed
On Wed, Nov 17, 2021 at 10:57:53PM +0100, Frederic Weisbecker wrote:
> On Wed, Nov 17, 2021 at 11:27:10AM -0800, Paul E. McKenney wrote:
> > On Wed, Nov 17, 2021 at 04:56:35PM +0100, Frederic Weisbecker wrote:
> > > In order to be able to (de-)offload any CPU using cpuset in the future,
> > > create a NOCB kthread for all possible CPUs. For now this is done only
> > > as long as the "rcu_nocb=" kernel parameter is passed to avoid
> > > the unnecessary overhead for most users.
> >
> > The "nohz_full=" kernel parameter would also cause these kthreads to
> > be created, correct? (Yeah, a nit, but...)
> >
> > And some fallout of my forgetfulness below. :-/
>
> Ah right, that too!
> > >
> > > @@ -1268,7 +1265,7 @@ static void __init rcu_spawn_nocb_kthreads(void)
> > > int cpu;
> > >
> > > if (rcu_nocb_is_setup) {
> > > - for_each_online_cpu(cpu)
> > > + for_each_possible_cpu(cpu)
> >
> > Gah... I had forgotten. :-/
> >
> > Some firmware lies about the OS instance's age. Other firmware lies
> > about the number of CPUs, sometimes claiming large multiples of the
> > actual number of CPUs.
> >
> > So this needs to stay "for_each_online_cpu". Otherwise, Paul Gortmaker
> > will once again be afflicted with hundreds of unnecessary rcuo kthreads.
> >
> > The later calls to rcutree_prepare_cpu() from rcutree_prepare_cpu()
> > will take care of any CPUs that first come online later.
>
> Ok that makes sense.
>
> >
> > > rcu_spawn_cpu_nocb_kthread(cpu);
> > > }
> > > }
> > > @@ -1303,7 +1300,7 @@ static void __init rcu_organize_nocb_kthreads(void)
> > > * Should the corresponding CPU come online in the future, then
> > > * we will spawn the needed set of rcu_nocb_kthread() kthreads.
> > > */
> > > - for_each_cpu(cpu, rcu_nocb_mask) {
> > > + for_each_possible_cpu(cpu) {
> >
> > This needs to change, but to for_each_online_cpu() instead of
> > for_each_possible_cpu(). That handles the case where the boot CPU is
> > not initially offloaded, but where the sysadm later needs to offload it.
>
> I'm less sure about that one. This is called early from rcu_init_nohz(). Only
> the boot CPU should be online at that point. We really need to integrate all
> possible CPUs within the nocb groups.

Ah, yes, this is creating the data-structure linkages, not the kthreads.
Your for_each_possible_cpu() is quite correct, good!

Thanx, Paul

> Thanks.
>
> >
> > > rdp = per_cpu_ptr(&rcu_data, cpu);
> > > if (rdp->cpu >= nl) {
> > > /* New GP kthread, set up for CBs & next GP. */
> > > @@ -1327,7 +1324,8 @@ static void __init rcu_organize_nocb_kthreads(void)
> > > pr_cont(" %d", cpu);
> > > }
> > > rdp->nocb_gp_rdp = rdp_gp;
> > > - list_add_tail(&rdp->nocb_entry_rdp, &rdp_gp->nocb_head_rdp);
> > > + if (cpumask_test_cpu(cpu, rcu_nocb_mask))
> > > + list_add_tail(&rdp->nocb_entry_rdp, &rdp_gp->nocb_head_rdp);
> > > }
> > > if (gotnocbs && dump_tree)
> > > pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
> > > --
> > > 2.25.1
> > >

\
 
 \ /
  Last update: 2021-11-17 23:31    [W:0.037 / U:0.076 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site