Re: [PATCH 3/4 v2] cgroup: separate rstat locks for subsystems

[JP Kobryn <inwardvessel@gmail.com>]

On 2/27/25 1:55 PM, inwardvessel wrote:
> From: JP Kobryn <inwardvessel@gmail.com>
> 
> Let the existing locks be dedicated to the base stats and rename them as
> such. Also add new rstat locks for each enabled subsystem. When handling
> cgroup subsystem states, distinguish between formal subsystems (memory,
> io, etc) and the base stats subsystem state (represented by
> cgroup::self) to decide on which locks to take. This change is made to
> prevent contention between subsystems when updating/flushing stats.
> 
> Signed-off-by: JP Kobryn <inwardvessel@gmail.com>
> ---
>   kernel/cgroup/rstat.c | 93 +++++++++++++++++++++++++++++++++----------
>   1 file changed, 72 insertions(+), 21 deletions(-)
> 
> diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
> index 88908ef9212d..b3eaefc1fd07 100644
> --- a/kernel/cgroup/rstat.c
> +++ b/kernel/cgroup/rstat.c
> @@ -9,8 +9,12 @@
>   
>   #include <trace/events/cgroup.h>
>   
> -static DEFINE_SPINLOCK(cgroup_rstat_lock);
> -static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock);
> +static DEFINE_SPINLOCK(cgroup_rstat_base_lock);
> +static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_base_cpu_lock);
> +
> +static spinlock_t cgroup_rstat_subsys_lock[CGROUP_SUBSYS_COUNT];
> +static DEFINE_PER_CPU(raw_spinlock_t,
> +		cgroup_rstat_subsys_cpu_lock[CGROUP_SUBSYS_COUNT]);
>   
>   static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu);
>   
> @@ -20,8 +24,13 @@ static struct cgroup_rstat_cpu *cgroup_rstat_cpu(
>   	return per_cpu_ptr(css->rstat_cpu, cpu);
>   }
>   
> +static inline bool is_base_css(struct cgroup_subsys_state *css)
> +{
> +	return css->ss == NULL;
> +}
> +
>   /*
> - * Helper functions for rstat per CPU lock (cgroup_rstat_cpu_lock).
> + * Helper functions for rstat per CPU locks.
>    *
>    * This makes it easier to diagnose locking issues and contention in
>    * production environments. The parameter @fast_path determine the
> @@ -36,12 +45,12 @@ unsigned long _cgroup_rstat_cpu_lock(raw_spinlock_t *cpu_lock, int cpu,
>   	bool contended;
>   
>   	/*
> -	 * The _irqsave() is needed because cgroup_rstat_lock is
> -	 * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
> -	 * this lock with the _irq() suffix only disables interrupts on
> -	 * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
> -	 * interrupts on both configurations. The _irqsave() ensures
> -	 * that interrupts are always disabled and later restored.
> +	 * The _irqsave() is needed because the locks used for flushing are
> +	 * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring this lock
> +	 * with the _irq() suffix only disables interrupts on a non-PREEMPT_RT
> +	 * kernel. The raw_spinlock_t below disables interrupts on both
> +	 * configurations. The _irqsave() ensures that interrupts are always
> +	 * disabled and later restored.
>   	 */
>   	contended = !raw_spin_trylock_irqsave(cpu_lock, flags);
>   	if (contended) {
> @@ -87,7 +96,7 @@ __bpf_kfunc void cgroup_rstat_updated(
>   		struct cgroup_subsys_state *css, int cpu)
>   {
>   	struct cgroup *cgrp = css->cgroup;
> -	raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
> +	raw_spinlock_t *cpu_lock;
>   	unsigned long flags;
>   
>   	/*
> @@ -101,6 +110,12 @@ __bpf_kfunc void cgroup_rstat_updated(
>   	if (data_race(cgroup_rstat_cpu(css, cpu)->updated_next))
>   		return;
>   
> +	if (is_base_css(css))
> +		cpu_lock = per_cpu_ptr(&cgroup_rstat_base_cpu_lock, cpu);
> +	else
> +		cpu_lock = per_cpu_ptr(cgroup_rstat_subsys_cpu_lock, cpu) +
> +			css->ss->id;
> +
>   	flags = _cgroup_rstat_cpu_lock(cpu_lock, cpu, cgrp, true);
>   
>   	/* put @css and all ancestors on the corresponding updated lists */
> @@ -208,11 +223,17 @@ static struct cgroup_subsys_state *cgroup_rstat_updated_list(
>   		struct cgroup_subsys_state *root, int cpu)
>   {
>   	struct cgroup *cgrp = root->cgroup;
> -	raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
>   	struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(root, cpu);
>   	struct cgroup_subsys_state *head = NULL, *parent, *child;
> +	raw_spinlock_t *cpu_lock;
>   	unsigned long flags;
>   
> +	if (is_base_css(root))
> +		cpu_lock = per_cpu_ptr(&cgroup_rstat_base_cpu_lock, cpu);
> +	else
> +		cpu_lock = per_cpu_ptr(cgroup_rstat_subsys_cpu_lock, cpu) +
> +			root->ss->id;
> +
>   	flags = _cgroup_rstat_cpu_lock(cpu_lock, cpu, cgrp, false);
>   
>   	/* Return NULL if this subtree is not on-list */
> @@ -315,7 +336,7 @@ static void cgroup_rstat_flush_locked(struct cgroup_subsys_state *css,
>   	struct cgroup *cgrp = css->cgroup;
>   	int cpu;
>   
> -	lockdep_assert_held(&cgroup_rstat_lock);
> +	lockdep_assert_held(&lock);

I need to remove the ampersand since the variable is already a pointer.

>   
>   	for_each_possible_cpu(cpu) {
>   		struct cgroup_subsys_state *pos;
> @@ -356,12 +377,18 @@ static void cgroup_rstat_flush_locked(struct cgroup_subsys_state *css,
>   __bpf_kfunc void cgroup_rstat_flush(struct cgroup_subsys_state *css)
>   {
>   	struct cgroup *cgrp = css->cgroup;
> +	spinlock_t *lock;
> +
> +	if (is_base_css(css))
> +		lock = &cgroup_rstat_base_lock;
> +	else
> +		lock = &cgroup_rstat_subsys_lock[css->ss->id];
>   
>   	might_sleep();
>   
> -	__cgroup_rstat_lock(&cgroup_rstat_lock, cgrp, -1);
> -	cgroup_rstat_flush_locked(css, &cgroup_rstat_lock);
> -	__cgroup_rstat_unlock(&cgroup_rstat_lock, cgrp, -1);
> +	__cgroup_rstat_lock(lock, cgrp, -1);
> +	cgroup_rstat_flush_locked(css, lock);
> +	__cgroup_rstat_unlock(lock, cgrp, -1);
>   }
>   
>   /**
> @@ -376,10 +403,16 @@ __bpf_kfunc void cgroup_rstat_flush(struct cgroup_subsys_state *css)
>   void cgroup_rstat_flush_hold(struct cgroup_subsys_state *css)
>   {
>   	struct cgroup *cgrp = css->cgroup;
> +	spinlock_t *lock;
> +
> +	if (is_base_css(css))
> +		lock = &cgroup_rstat_base_lock;
> +	else
> +		lock = &cgroup_rstat_subsys_lock[css->ss->id];
>   
>   	might_sleep();
> -	__cgroup_rstat_lock(&cgroup_rstat_lock, cgrp, -1);
> -	cgroup_rstat_flush_locked(css, &cgroup_rstat_lock);
> +	__cgroup_rstat_lock(lock, cgrp, -1);
> +	cgroup_rstat_flush_locked(css, lock);
>   }
>   
>   /**
> @@ -389,7 +422,14 @@ void cgroup_rstat_flush_hold(struct cgroup_subsys_state *css)
>   void cgroup_rstat_flush_release(struct cgroup_subsys_state *css)
>   {
>   	struct cgroup *cgrp = css->cgroup;
> -	__cgroup_rstat_unlock(&cgroup_rstat_lock, cgrp, -1);
> +	spinlock_t *lock;
> +
> +	if (is_base_css(css))
> +		lock = &cgroup_rstat_base_lock;
> +	else
> +		lock = &cgroup_rstat_subsys_lock[css->ss->id];
> +
> +	__cgroup_rstat_unlock(lock, cgrp, -1);
>   }
>   
>   int cgroup_rstat_init(struct cgroup_subsys_state *css)
> @@ -435,10 +475,21 @@ void cgroup_rstat_exit(struct cgroup_subsys_state *css)
>   
>   void __init cgroup_rstat_boot(void)
>   {
> -	int cpu;
> +	struct cgroup_subsys *ss;
> +	int cpu, ssid;
>   
> -	for_each_possible_cpu(cpu)
> -		raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu));
> +	for_each_subsys(ss, ssid) {
> +		spin_lock_init(&cgroup_rstat_subsys_lock[ssid]);
> +	}
> +
> +	for_each_possible_cpu(cpu) {
> +		raw_spin_lock_init(per_cpu_ptr(&cgroup_rstat_base_cpu_lock, cpu));
> +
> +		for_each_subsys(ss, ssid) {
> +			raw_spin_lock_init(
> +					per_cpu_ptr(cgroup_rstat_subsys_cpu_lock, cpu) + ssid);
> +		}
> +	}
>   }
>   
>   /*