At the moment it is currently possible to sneak a core into nohz_full
that lies between nr_possible and NR_CPUS - but you won't "see" it
because cpumask_pr_args() implicitly hides anything above nr_cpu_ids.
This becomes a problem when the nohz_full CPU set doesn't contain at
least one other valid nohz CPU - in which case we end up with the
tick_nohz_full_running set and no tick core specified, which trips an
endless sequence of WARN() and renders the machine unusable.
I inadvertently opened the door for this when fixing an overly
restrictive nohz_full conditional in the below Fixes: commit - and then
courtesy of my optimistic ACPI reporting nr_possible of 64 (the default
Kconfig for NR_CPUS) and the not-so helpful implict filtering done by
cpumask_pr_args, I unfortunately did not spot it during my testing.
So here, I don't rely on what was printed anymore, but code exactly what
our restrictions should be in order to be aligned with rcu_nocbs - which
was the original goal. Since the checks lie in "__init" code it is largely
free for us to do this anyway.
Building with NOHZ_FULL and NR_CPUS=128 on an otherwise defconfig, and
booting with "rcu_nocbs=8-127 nohz_full=96-127" on the same 16 core T5500
Dell machine now results in the following (only relevant lines shown):
smpboot: Allowing 64 CPUs, 48 hotplug CPUs
setup_percpu: NR_CPUS:128 nr_cpumask_bits:128 nr_cpu_ids:64 nr_node_ids:2
housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' contains nonexistent CPUs.
housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' has no valid CPUs.
rcu: RCU restricting CPUs from NR_CPUS=128 to nr_cpu_ids=64.
rcu: Note: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.
rcu: Offload RCU callbacks from CPUs: 8-63.
One can see both new housekeeping checks are triggered in the above.
The same invalid boot arg combination would have previously resulted in
an infinitely scrolling mix of WARN from all cores per tick on this box.
Fixes: 915a2bc3c6b7 ("sched/isolation: Reconcile rcu_nocbs= and nohz_full=")
Cc: Paul E. McKenney <[email protected]>
Cc: Frederic Weisbecker <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Signed-off-by: Paul Gortmaker <[email protected]>
---
kernel/sched/isolation.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
index 7f06eaf12818..01abc8400d6c 100644
--- a/kernel/sched/isolation.c
+++ b/kernel/sched/isolation.c
@@ -89,6 +89,18 @@ static int __init housekeeping_setup(char *str, enum hk_flags flags)
return 0;
}
+ if (!cpumask_subset(non_housekeeping_mask, cpu_possible_mask)) {
+ pr_info("housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' contains nonexistent CPUs.\n");
+ cpumask_and(non_housekeeping_mask, cpu_possible_mask,
+ non_housekeeping_mask);
+ }
+
+ if (cpumask_empty(non_housekeeping_mask)) {
+ pr_info("housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' has no valid CPUs.\n");
+ free_bootmem_cpumask_var(non_housekeeping_mask);
+ return 0;
+ }
+
alloc_bootmem_cpumask_var(&tmp);
if (!housekeeping_flags) {
alloc_bootmem_cpumask_var(&housekeeping_mask);
--
2.17.1
On Mon, Dec 06, 2021 at 09:59:49AM -0500, Paul Gortmaker wrote:
> At the moment it is currently possible to sneak a core into nohz_full
> that lies between nr_possible and NR_CPUS - but you won't "see" it
> because cpumask_pr_args() implicitly hides anything above nr_cpu_ids.
>
> This becomes a problem when the nohz_full CPU set doesn't contain at
> least one other valid nohz CPU - in which case we end up with the
> tick_nohz_full_running set and no tick core specified, which trips an
> endless sequence of WARN() and renders the machine unusable.
>
> I inadvertently opened the door for this when fixing an overly
> restrictive nohz_full conditional in the below Fixes: commit - and then
> courtesy of my optimistic ACPI reporting nr_possible of 64 (the default
> Kconfig for NR_CPUS) and the not-so helpful implict filtering done by
> cpumask_pr_args, I unfortunately did not spot it during my testing.
>
> So here, I don't rely on what was printed anymore, but code exactly what
> our restrictions should be in order to be aligned with rcu_nocbs - which
> was the original goal. Since the checks lie in "__init" code it is largely
> free for us to do this anyway.
>
> Building with NOHZ_FULL and NR_CPUS=128 on an otherwise defconfig, and
> booting with "rcu_nocbs=8-127 nohz_full=96-127" on the same 16 core T5500
> Dell machine now results in the following (only relevant lines shown):
>
> smpboot: Allowing 64 CPUs, 48 hotplug CPUs
> setup_percpu: NR_CPUS:128 nr_cpumask_bits:128 nr_cpu_ids:64 nr_node_ids:2
> housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' contains nonexistent CPUs.
> housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' has no valid CPUs.
> rcu: RCU restricting CPUs from NR_CPUS=128 to nr_cpu_ids=64.
> rcu: Note: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.
> rcu: Offload RCU callbacks from CPUs: 8-63.
>
> One can see both new housekeeping checks are triggered in the above.
> The same invalid boot arg combination would have previously resulted in
> an infinitely scrolling mix of WARN from all cores per tick on this box.
>
> Fixes: 915a2bc3c6b7 ("sched/isolation: Reconcile rcu_nocbs= and nohz_full=")
> Cc: Paul E. McKenney <[email protected]>
> Cc: Frederic Weisbecker <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Signed-off-by: Paul Gortmaker <[email protected]>
> ---
> kernel/sched/isolation.c | 12 ++++++++++++
> 1 file changed, 12 insertions(+)
>
> diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
> index 7f06eaf12818..01abc8400d6c 100644
> --- a/kernel/sched/isolation.c
> +++ b/kernel/sched/isolation.c
> @@ -89,6 +89,18 @@ static int __init housekeeping_setup(char *str, enum hk_flags flags)
> return 0;
> }
>
> + if (!cpumask_subset(non_housekeeping_mask, cpu_possible_mask)) {
> + pr_info("housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' contains nonexistent CPUs.\n");
> + cpumask_and(non_housekeeping_mask, cpu_possible_mask,
> + non_housekeeping_mask);
> + }
> +
> + if (cpumask_empty(non_housekeeping_mask)) {
> + pr_info("housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' has no valid CPUs.\n");
> + free_bootmem_cpumask_var(non_housekeeping_mask);
> + return 0;
If Frederic applies his rcu_nocbs work to nohz_full, it may some day be
valid to specify an empty nohz_full CPU mask. Of course, it might well
be that warning in the meantime is a good thing, but I figured that I
should call attention to the possibility.
Thanx, Paul
> + }
> +
> alloc_bootmem_cpumask_var(&tmp);
> if (!housekeeping_flags) {
> alloc_bootmem_cpumask_var(&housekeeping_mask);
> --
> 2.17.1
>
[Re: [PATCH 1/2] sched/isolation: really align nohz_full with rcu_nocbs] On 06/12/2021 (Mon 13:33) Paul E. McKenney wrote:
> On Mon, Dec 06, 2021 at 09:59:49AM -0500, Paul Gortmaker wrote:
> > At the moment it is currently possible to sneak a core into nohz_full
> > that lies between nr_possible and NR_CPUS - but you won't "see" it
> > because cpumask_pr_args() implicitly hides anything above nr_cpu_ids.
> >
> > This becomes a problem when the nohz_full CPU set doesn't contain at
> > least one other valid nohz CPU - in which case we end up with the
> > tick_nohz_full_running set and no tick core specified, which trips an
> > endless sequence of WARN() and renders the machine unusable.
> >
> > I inadvertently opened the door for this when fixing an overly
> > restrictive nohz_full conditional in the below Fixes: commit - and then
> > courtesy of my optimistic ACPI reporting nr_possible of 64 (the default
> > Kconfig for NR_CPUS) and the not-so helpful implict filtering done by
> > cpumask_pr_args, I unfortunately did not spot it during my testing.
> >
> > So here, I don't rely on what was printed anymore, but code exactly what
> > our restrictions should be in order to be aligned with rcu_nocbs - which
> > was the original goal. Since the checks lie in "__init" code it is largely
> > free for us to do this anyway.
> >
> > Building with NOHZ_FULL and NR_CPUS=128 on an otherwise defconfig, and
> > booting with "rcu_nocbs=8-127 nohz_full=96-127" on the same 16 core T5500
> > Dell machine now results in the following (only relevant lines shown):
> >
> > smpboot: Allowing 64 CPUs, 48 hotplug CPUs
> > setup_percpu: NR_CPUS:128 nr_cpumask_bits:128 nr_cpu_ids:64 nr_node_ids:2
> > housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' contains nonexistent CPUs.
> > housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' has no valid CPUs.
> > rcu: RCU restricting CPUs from NR_CPUS=128 to nr_cpu_ids=64.
> > rcu: Note: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.
> > rcu: Offload RCU callbacks from CPUs: 8-63.
> >
> > One can see both new housekeeping checks are triggered in the above.
> > The same invalid boot arg combination would have previously resulted in
> > an infinitely scrolling mix of WARN from all cores per tick on this box.
> >
> > Fixes: 915a2bc3c6b7 ("sched/isolation: Reconcile rcu_nocbs= and nohz_full=")
> > Cc: Paul E. McKenney <[email protected]>
> > Cc: Frederic Weisbecker <[email protected]>
> > Cc: Thomas Gleixner <[email protected]>
> > Cc: Ingo Molnar <[email protected]>
> > Signed-off-by: Paul Gortmaker <[email protected]>
> > ---
> > kernel/sched/isolation.c | 12 ++++++++++++
> > 1 file changed, 12 insertions(+)
> >
> > diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
> > index 7f06eaf12818..01abc8400d6c 100644
> > --- a/kernel/sched/isolation.c
> > +++ b/kernel/sched/isolation.c
> > @@ -89,6 +89,18 @@ static int __init housekeeping_setup(char *str, enum hk_flags flags)
> > return 0;
> > }
> >
> > + if (!cpumask_subset(non_housekeeping_mask, cpu_possible_mask)) {
> > + pr_info("housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' contains nonexistent CPUs.\n");
> > + cpumask_and(non_housekeeping_mask, cpu_possible_mask,
> > + non_housekeeping_mask);
> > + }
> > +
> > + if (cpumask_empty(non_housekeeping_mask)) {
> > + pr_info("housekeeping: kernel parameter 'nohz_full=' or 'isolcpus=' has no valid CPUs.\n");
> > + free_bootmem_cpumask_var(non_housekeeping_mask);
> > + return 0;
>
> If Frederic applies his rcu_nocbs work to nohz_full, it may some day be
> valid to specify an empty nohz_full CPU mask. Of course, it might well
> be that warning in the meantime is a good thing, but I figured that I
> should call attention to the possibility.
It isn't just a good thing ; it is required. Call chain is as this:
nohz_full= / isolcpus=
housekeeping_nohz_full_setup / housekeeping_isolcpus_setup
housekeeping_setup
tick_nohz_full_setup
tick_nohz_full_running = true;
So housekeeping setup is the "last chance" to validate inputs and
avoid calling tick_nohz_full_setup which unconditionally sets the
tick_nohz_full_running (as the crux of this problem).
At least that is as things stand today based on my understanding.
Paul.
--
>
> Thanx, Paul
>
> > + }
> > +
> > alloc_bootmem_cpumask_var(&tmp);
> > if (!housekeeping_flags) {
> > alloc_bootmem_cpumask_var(&housekeeping_mask);
> > --
> > 2.17.1
> >