Re: [PATCH] sched: Fix Core-wide rq->lock for uninitialized CPUs

From: Tao Zhou
Date: Thu Aug 19 2021 - 11:50:16 EST


On Thu, Aug 19, 2021 at 01:09:17PM +0200, Peter Zijlstra wrote:

> On Wed, Aug 18, 2021 at 01:17:34AM +0200, Eugene Syromiatnikov wrote:
> > On Tue, Aug 17, 2021 at 05:52:43PM +0200, Peter Zijlstra wrote:
> > > Urgh... lemme guess, your HP BIOS is funny and reports more possible
> > > CPUs than you actually have resulting in cpu_possible_mask !=
> > > cpu_online_mask. Alternatively, you booted with nr_cpus= or something
> > > daft like that.
> >
> > Yep, it seems to be the case:
> >
> > # cat /sys/devices/system/cpu/possible
> > 0-7
> > # cat /sys/devices/system/cpu/online
> > 0-3
> >
>
> I think the below should work... can you please verify?
>
> ---
> Subject: sched: Fix Core-wide rq->lock for uninitialized CPUs
>
> Eugene tripped over the case where rq_lock(), as called in a
> for_each_possible_cpu() loop came apart because rq->core hadn't been
> setup yet.
>
> This is a somewhat unusual, but valid case.
>
> Rework things such that rq->core is initialized to point at itself. IOW
> initialize each CPU as a single threaded Core. CPU online will then join
> the new CPU (thread) to an existing Core where needed.
>
> For completeness sake, have CPU offline fully undo the state so as to
> not presume the topology will match the next time it comes online.
>
> Fixes: 9edeaea1bc45 ("sched: Core-wide rq->lock")
> Reported-by: Eugene Syromiatnikov <esyr@xxxxxxxxxx>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> ---
> kernel/sched/core.c | 143 +++++++++++++++++++++++++++++++++++++++++----------
> kernel/sched/sched.h | 2 +-
> 2 files changed, 118 insertions(+), 27 deletions(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 0f14e09a4f99..21d633971fcf 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -237,9 +237,30 @@ static DEFINE_MUTEX(sched_core_mutex);
> static atomic_t sched_core_count;
> static struct cpumask sched_core_mask;
>
> +static void sched_core_lock(int cpu, unsigned long *flags)
> +{
> + const struct cpumask *smt_mask = cpu_smt_mask(cpu);
> + int t, i = 0;
> +
> + local_irq_save(*flags);
> + for_each_cpu(t, smt_mask)
> + raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
> +}
> +
> +static void sched_core_unlock(int cpu, unsigned long *flags)
> +{
> + const struct cpumask *smt_mask = cpu_smt_mask(cpu);
> + int t;
> +
> + for_each_cpu(t, smt_mask)
> + raw_spin_unlock(&cpu_rq(t)->__lock);
> + local_irq_restore(*flags);
> +}
> +
> static void __sched_core_flip(bool enabled)
> {
> - int cpu, t, i;
> + unsigned long flags;
> + int cpu, t;
>
> cpus_read_lock();
>
> @@ -250,19 +271,12 @@ static void __sched_core_flip(bool enabled)
> for_each_cpu(cpu, &sched_core_mask) {
> const struct cpumask *smt_mask = cpu_smt_mask(cpu);
>
> - i = 0;
> - local_irq_disable();
> - for_each_cpu(t, smt_mask) {
> - /* supports up to SMT8 */
> - raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
> - }
> + sched_core_lock(cpu, &flags);
>
> for_each_cpu(t, smt_mask)
> cpu_rq(t)->core_enabled = enabled;
>
> - for_each_cpu(t, smt_mask)
> - raw_spin_unlock(&cpu_rq(t)->__lock);
> - local_irq_enable();
> + sched_core_unlock(cpu, &flags);
>
> cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
> }
> @@ -5979,35 +5993,109 @@ void queue_core_balance(struct rq *rq)
> queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
> }
>
> -static inline void sched_core_cpu_starting(unsigned int cpu)
> +static void sched_core_cpu_starting(unsigned int cpu)
> {
> const struct cpumask *smt_mask = cpu_smt_mask(cpu);
> - struct rq *rq, *core_rq = NULL;
> - int i;
> + struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
> + unsigned long flags;
> + int t;
>
> - core_rq = cpu_rq(cpu)->core;
> + sched_core_lock(cpu, &flags);
>
> - if (!core_rq) {
> - for_each_cpu(i, smt_mask) {
> - rq = cpu_rq(i);
> - if (rq->core && rq->core == rq)
> - core_rq = rq;
> + WARN_ON_ONCE(rq->core != rq);
> +
> + /* if we're the first, we'll be our own leader */
> + if (cpumask_weight(smt_mask) == 1)
> + goto unlock;
> +
> + /* find the leader */
> + for_each_cpu(t, smt_mask) {
> + if (t == cpu)
> + continue;
> + rq = cpu_rq(t);
> + if (rq->core == rq) {
> + core_rq = rq;
> + break;
> }
> + }
>
> - if (!core_rq)
> - core_rq = cpu_rq(cpu);
> + if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
> + goto unlock;
>
> - for_each_cpu(i, smt_mask) {
> - rq = cpu_rq(i);
> + /* install and validate core_rq */
> + for_each_cpu(t, smt_mask) {
> + rq = cpu_rq(t);
>
> - WARN_ON_ONCE(rq->core && rq->core != core_rq);
> + if (t == cpu)
> rq->core = core_rq;
> - }
> +
> + WARN_ON_ONCE(rq->core != core_rq);
> }

Yoh, Peter!

Read not more than ten times but stay here long enough until I realize
something that I am not sure. You are not only just fix the problem,
you change the rq's core selection method(manner.). Original source now
set the core to one rq, here you distribute the core_rq to not just only
one rq.

For example(SMT3): rq1 --> rq0, rq0 --> rq2, rq2 --> rq2(/* whoopsie */).

And the core can change when deactivate.
Not sure I'm getting all the things right, but this is more performance
efficient because the lock contention is avoided when pick task from my
imagination.
Again, I'm not know much about everything, just put some words.

> +
> +unlock:
> + sched_core_unlock(cpu, &flags);
> }
> +
> +static void sched_core_cpu_deactivate(unsigned int cpu)
> +{
> + const struct cpumask *smt_mask = cpu_smt_mask(cpu);
> + struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
> + unsigned long flags;
> + int t;
> +
> + sched_core_lock(cpu, &flags);
> +
> + /* if we're the last man standing, nothing to do */
> + if (cpumask_weight(smt_mask) == 1) {
> + WARN_ON_ONCE(rq->core != rq);
> + goto unlock;
> + }
> +
> + /* if we're not the leader, nothing to do */
> + if (rq->core != rq)
> + goto unlock;
> +
> + /* find a new leader */
> + for_each_cpu(t, smt_mask) {
> + if (t == cpu)
> + continue;
> + core_rq = cpu_rq(t);
> + break;
> + }
> +
> + if (WARN_ON_ONCE(!core_rq)) /* impossible */
> + goto unlock;
> +
> + /* copy the shared state to the new leader */
> + core_rq->core_task_seq = rq->core_task_seq;
> + core_rq->core_pick_seq = rq->core_pick_seq;
> + core_rq->core_cookie = rq->core_cookie;
> + core_rq->core_forceidle = rq->core_forceidle;
> + core_rq->core_forceidle_seq = rq->core_forceidle_seq;
> +
> + /* install new leader */
> + for_each_cpu(t, smt_mask) {
> + rq = cpu_rq(t);
> + rq->core = core_rq;
> + }
> +
> +unlock:
> + sched_core_unlock(cpu, &flags);
> +}
> +
> +static inline void sched_core_cpu_dying(unsigned int cpu)
> +{
> + struct rq *rq = cpu_rq(cpu);
> +
> + if (rq->core != rq)
> + rq->core = rq;
> +}
> +
> #else /* !CONFIG_SCHED_CORE */
>
> static inline void sched_core_cpu_starting(unsigned int cpu) {}
> +static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
> +static inline void sched_core_cpu_dying(unsigned int cpu) {}
>
> static struct task_struct *
> pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
> @@ -9009,6 +9097,8 @@ int sched_cpu_deactivate(unsigned int cpu)
> */
> if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
> static_branch_dec_cpuslocked(&sched_smt_present);
> +
> + sched_core_cpu_deactivate(cpu);
> #endif
>
> if (!sched_smp_initialized)
> @@ -9113,6 +9203,7 @@ int sched_cpu_dying(unsigned int cpu)
> calc_load_migrate(rq);
> update_max_interval();
> hrtick_clear(rq);
> + sched_core_cpu_dying(cpu);
> return 0;
> }
> #endif
> @@ -9324,7 +9415,7 @@ void __init sched_init(void)
> atomic_set(&rq->nr_iowait, 0);
>
> #ifdef CONFIG_SCHED_CORE
> - rq->core = NULL;
> + rq->core = rq;
> rq->core_pick = NULL;
> rq->core_enabled = 0;
> rq->core_tree = RB_ROOT;
> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> index a9a660c6e08a..e6347c88c467 100644
> --- a/kernel/sched/sched.h
> +++ b/kernel/sched/sched.h
> @@ -1103,7 +1103,7 @@ struct rq {
> unsigned int core_sched_seq;
> struct rb_root core_tree;
>
> - /* shared state */
> + /* shared state -- careful with sched_core_cpu_deactivate() */
> unsigned int core_task_seq;
> unsigned int core_pick_seq;
> unsigned long core_cookie;


Thanks,
Tao