[PATCH v4 6/7] sched/fair: skip busy cores in SIS search
From: Abel Wu
Date: Sun Jun 19 2022 - 08:05:55 EST
Try to improve efficiency of SIS domain search by filtering out busy
cores, and as a result the more overloaded the system is, the less
cpus will be scanned.
The filter is supposed to contain unoccupied cpus of the LLC. And we
propagate these cpus to the filter one at a time at core granule.
This can help spreading load to different cores given that the search
depth is limited. The chosen cpu to be propagated is guaranteed to be
unoccupied at that time.
When idle cpu exists, the last one is preferred in order not to
conflict with periodic load balancing during which the first idle
cpu (if any) is chosen to be fed with tasks.
Signed-off-by: Abel Wu <wuyun.abel@xxxxxxxxxxxxx>
---
include/linux/sched/topology.h | 20 ++++++++
kernel/sched/fair.c | 90 +++++++++++++++++++++++++++++++---
kernel/sched/topology.c | 12 ++++-
3 files changed, 115 insertions(+), 7 deletions(-)
diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h
index 3e99ac98d766..b93edf587d84 100644
--- a/include/linux/sched/topology.h
+++ b/include/linux/sched/topology.h
@@ -103,6 +103,10 @@ struct sched_group;
* load balancing on each SMT domain inside the LLC, the state will be
* re-evaluated and switch from sd_is_busy to sd_has_icpus if idle cpus
* exist.
+ *
+ * For SMT domains, the state is updated during load balancing at SMT
+ * level. Upper levels are ignored due to the long intervals that make
+ * information out-of-date quickly.
*/
enum sd_state {
sd_has_icores,
@@ -113,7 +117,18 @@ enum sd_state {
struct sched_domain_shared {
atomic_t ref;
atomic_t nr_busy_cpus;
+
+ int updating;
int state; /* see enum sd_state */
+
+ /*
+ * Record unoccupied cpus for SIS domain search.
+ *
+ * NOTE: this field is variable length. (Allocated dynamically
+ * by attaching extra space to the end of the structure,
+ * depending on how many CPUs the kernel has booted up with)
+ */
+ unsigned long idle_cpus[];
};
struct sched_domain {
@@ -199,6 +214,11 @@ static inline struct cpumask *sched_domain_span(struct sched_domain *sd)
return to_cpumask(sd->span);
}
+static inline struct cpumask *sched_domain_icpus(struct sched_domain_shared *sds)
+{
+ return to_cpumask(sds->idle_cpus);
+}
+
extern void partition_sched_domains_locked(int ndoms_new,
cpumask_var_t doms_new[],
struct sched_domain_attr *dattr_new);
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 2ca37fdc6c4d..d55fdcedf2c0 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6241,6 +6241,28 @@ static inline bool test_idle_cpus(int cpu)
return sd_get_state(cpu) != sd_is_busy;
}
+static void sd_update_icpus(int core, int icpu)
+{
+ struct sched_domain_shared *sds;
+ struct cpumask *icpus;
+
+ sds = rcu_dereference(per_cpu(sd_llc_shared, core));
+ if (!sds)
+ return;
+
+ icpus = sched_domain_icpus(sds);
+
+ /*
+ * XXX: The update is racy between different cores.
+ * The non-atomic ops here is a tradeoff of accuracy
+ * for easing the cache traffic.
+ */
+ if (icpu == -1)
+ cpumask_andnot(icpus, icpus, cpu_smt_mask(core));
+ else if (!cpumask_test_cpu(icpu, icpus))
+ __cpumask_set_cpu(icpu, icpus);
+}
+
/*
* Scans the local SMT mask to see if the entire core is idle, and records this
* information in sd_llc_shared->has_idle_cores.
@@ -6340,6 +6362,10 @@ static inline bool test_idle_cpus(int cpu)
return true;
}
+static inline void sd_update_icpus(int core, int icpu)
+{
+}
+
static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
{
return __select_idle_cpu(core, p);
@@ -6370,7 +6396,8 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
if (!this_sd)
return -1;
- cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
+ cpumask_and(cpus, has_idle_core ? sched_domain_span(sd) :
+ sched_domain_icpus(sd->shared), p->cpus_ptr);
if (sched_feat(SIS_PROP) && !has_idle_core) {
u64 avg_cost, avg_idle, span_avg;
@@ -8342,6 +8369,7 @@ struct sd_lb_stats {
unsigned int prefer_sibling; /* tasks should go to sibling first */
int sd_state;
+ int idle_cpu;
struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
struct sg_lb_stats local_stat; /* Statistics of the local group */
@@ -8362,6 +8390,7 @@ static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
.total_load = 0UL,
.total_capacity = 0UL,
.sd_state = sd_is_busy,
+ .idle_cpu = -1,
.busiest_stat = {
.idle_cpus = UINT_MAX,
.group_type = group_has_spare,
@@ -8702,10 +8731,18 @@ sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs
return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
}
-static inline void sd_classify(struct sd_lb_stats *sds, struct rq *rq)
+static inline void sd_classify(struct sd_lb_stats *sds, struct rq *rq, int cpu)
{
- if (sds->sd_state != sd_has_icpus && unoccupied_rq(rq))
+ if (sds->sd_state != sd_has_icpus && unoccupied_rq(rq)) {
+ /*
+ * Prefer idle cpus than unoccupied ones. This
+ * is achieved by only allowing the idle ones
+ * unconditionally overwrite the preious record
+ * while the occupied ones can't.
+ */
+ sds->idle_cpu = cpu;
sds->sd_state = sd_has_icpus;
+ }
}
/**
@@ -8741,7 +8778,7 @@ static inline void update_sg_lb_stats(struct lb_env *env,
sgs->sum_nr_running += nr_running;
if (update_core)
- sd_classify(sds, rq);
+ sd_classify(sds, rq, i);
if (nr_running > 1)
*sg_status |= SG_OVERLOAD;
@@ -8757,7 +8794,16 @@ static inline void update_sg_lb_stats(struct lb_env *env,
* No need to call idle_cpu() if nr_running is not 0
*/
if (!nr_running && idle_cpu(i)) {
+ /*
+ * Prefer the last idle cpu by overwriting
+ * preious one. The first idle cpu in this
+ * domain (if any) can trigger balancing
+ * and fed with tasks, so we'd better choose
+ * a candidate in an opposite way.
+ */
+ sds->idle_cpu = i;
sgs->idle_cpus++;
+
/* Idle cpu can't have misfit task */
continue;
}
@@ -9273,8 +9319,40 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
static void sd_update_state(struct lb_env *env, struct sd_lb_stats *sds)
{
- if (sds->sd_state == sd_has_icpus && !test_idle_cpus(env->dst_cpu))
- set_idle_cpus(env->dst_cpu, true);
+ struct sched_domain_shared *sd_smt_shared = env->sd->shared;
+ enum sd_state new = sds->sd_state;
+ int this = env->dst_cpu;
+
+ /*
+ * Parallel updating can hardly contribute accuracy to
+ * the filter, besides it can be one of the burdens on
+ * cache traffic.
+ */
+ if (cmpxchg(&sd_smt_shared->updating, 0, 1))
+ return;
+
+ /*
+ * There is at least one unoccupied cpu available, so
+ * propagate it to the filter to avoid false negative
+ * issue which could result in lost tracking of some
+ * idle cpus thus throughupt downgraded.
+ */
+ if (new != sd_is_busy) {
+ if (!test_idle_cpus(this))
+ set_idle_cpus(this, true);
+ } else {
+ /*
+ * Nothing changes so nothing to update or
+ * propagate.
+ */
+ if (sd_smt_shared->state == sd_is_busy)
+ goto out;
+ }
+
+ sd_update_icpus(this, sds->idle_cpu);
+ sd_smt_shared->state = new;
+out:
+ xchg(&sd_smt_shared->updating, 0);
}
/**
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 8739c2a5a54e..d3cd7cf5a136 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -1641,6 +1641,16 @@ sd_init(struct sched_domain_topology_level *tl,
sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
atomic_inc(&sd->shared->ref);
atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
+
+ /*
+ * Initialize SMT domains to be busy, so that we don't
+ * need to propagate idle cpus to LLC domains which are
+ * default to fully busy (no cpus set). This will be
+ * updated in the first load balancing on SMT domains
+ * if necessary.
+ */
+ if (sd->flags & SD_SHARE_CPUCAPACITY)
+ WRITE_ONCE(sd->shared->state, sd_is_busy);
}
sd->private = sdd;
@@ -2106,7 +2116,7 @@ static int __sdt_alloc(const struct cpumask *cpu_map)
*per_cpu_ptr(sdd->sd, j) = sd;
- sds = kzalloc_node(sizeof(struct sched_domain_shared),
+ sds = kzalloc_node(sizeof(struct sched_domain_shared) + cpumask_size(),
GFP_KERNEL, cpu_to_node(j));
if (!sds)
return -ENOMEM;
--
2.31.1