/* Task command name length: */
#define TASK_COMM_LEN 16
+enum task_event {
+ PUT_PREV_TASK = 0,
+ PICK_NEXT_TASK = 1,
+ TASK_WAKE = 2,
+ TASK_MIGRATE = 3,
+ TASK_UPDATE = 4,
+ IRQ_UPDATE = 5,
+};
+
extern cpumask_var_t cpu_isolated_map;
extern void scheduler_tick(void);
#endif
};
+#ifdef CONFIG_SCHED_WALT
+#define RAVG_HIST_SIZE_MAX 5
+
+/* ravg represents frequency scaled cpu-demand of tasks */
+struct ravg {
+ /*
+ * 'mark_start' marks the beginning of an event (task waking up, task
+ * starting to execute, task being preempted) within a window
+ *
+ * 'sum' represents how runnable a task has been within current
+ * window. It incorporates both running time and wait time and is
+ * frequency scaled.
+ *
+ * 'sum_history' keeps track of history of 'sum' seen over previous
+ * RAVG_HIST_SIZE windows. Windows where task was entirely sleeping are
+ * ignored.
+ *
+ * 'demand' represents maximum sum seen over previous
+ * sysctl_sched_ravg_hist_size windows. 'demand' could drive frequency
+ * demand for tasks.
+ *
+ * 'curr_window' represents task's contribution to cpu busy time
+ * statistics (rq->curr_runnable_sum) in current window
+ *
+ * 'prev_window' represents task's contribution to cpu busy time
+ * statistics (rq->prev_runnable_sum) in previous window
+ */
+ u64 mark_start;
+ u32 sum, demand;
+ u32 sum_history[RAVG_HIST_SIZE_MAX];
+ u32 curr_window, prev_window;
+ u16 active_windows;
+};
+#endif
+
struct sched_rt_entity {
struct list_head run_list;
unsigned long timeout;
const struct sched_class *sched_class;
struct sched_entity se;
struct sched_rt_entity rt;
+#ifdef CONFIG_SCHED_WALT
+ struct ravg ravg;
+ /*
+ * 'init_load_pct' represents the initial task load assigned to children
+ * of this task
+ */
+ u32 init_load_pct;
+ u64 last_sleep_ts;
+#endif
+
#ifdef CONFIG_CGROUP_SCHED
struct task_group *sched_task_group;
#endif
extern unsigned int sysctl_sched_cstate_aware;
extern unsigned int sysctl_sched_wakeup_granularity;
extern unsigned int sysctl_sched_child_runs_first;
+#ifdef CONFIG_SCHED_WALT
+extern unsigned int sysctl_sched_use_walt_cpu_util;
+extern unsigned int sysctl_sched_use_walt_task_util;
+extern unsigned int sysctl_sched_walt_init_task_load_pct;
+extern unsigned int sysctl_sched_walt_cpu_high_irqload;
+#endif
enum sched_tunable_scaling {
SCHED_TUNABLESCALING_NONE,
__entry->util)
);
+#ifdef CONFIG_SCHED_WALT
+extern unsigned int sysctl_sched_use_walt_cpu_util;
+extern unsigned int sysctl_sched_use_walt_task_util;
+extern unsigned int walt_ravg_window;
+extern bool walt_disabled;
+#endif
+
+/*
+ * Tracepoint for accounting cpu root cfs_rq
+ */
+TRACE_EVENT(sched_load_avg_cpu,
+
+ TP_PROTO(int cpu, struct cfs_rq *cfs_rq),
+
+ TP_ARGS(cpu, cfs_rq),
+
+ TP_STRUCT__entry(
+ __field( int, cpu )
+ __field( unsigned long, load_avg )
+ __field( unsigned long, util_avg )
+ __field( unsigned long, util_avg_pelt )
+ __field( unsigned long, util_avg_walt )
+ ),
+
+ TP_fast_assign(
+ __entry->cpu = cpu;
+ __entry->load_avg = cfs_rq->avg.load_avg;
+ __entry->util_avg = cfs_rq->avg.util_avg;
+ __entry->util_avg_pelt = cfs_rq->avg.util_avg;
+ __entry->util_avg_walt = 0;
+#ifdef CONFIG_SCHED_WALT
+ __entry->util_avg_walt =
+ cpu_rq(cpu)->prev_runnable_sum << SCHED_CAPACITY_SHIFT;
+ do_div(__entry->util_avg_walt, walt_ravg_window);
+ if (!walt_disabled && sysctl_sched_use_walt_cpu_util)
+ __entry->util_avg = __entry->util_avg_walt;
+#endif
+ ),
+
+ TP_printk("cpu=%d load_avg=%lu util_avg=%lu "
+ "util_avg_pelt=%lu util_avg_walt=%lu",
+ __entry->cpu, __entry->load_avg, __entry->util_avg,
+ __entry->util_avg_pelt, __entry->util_avg_walt)
+);
+
+
/*
* Tracepoint for sched_entity load tracking:
*/
__field( pid_t, pid )
__field( unsigned long, load )
__field( unsigned long, util )
+ __field( unsigned long, util_pelt )
+ __field( unsigned long, util_walt )
),
TP_fast_assign(
__entry->pid = p ? p->pid : -1;
__entry->load = se->avg.load_avg;
__entry->util = se->avg.util_avg;
+ __entry->util_pelt = __entry->util;
+ __entry->util_walt = 0;
+#ifdef CONFIG_SCHED_WALT
+ if (!se->my_q) {
+ struct task_struct *p = container_of(se, struct task_struct, se);
+ __entry->util_walt = p->ravg.demand;
+ do_div(__entry->util_walt, walt_ravg_window >> SCHED_CAPACITY_SHIFT);
+ if (!walt_disabled && sysctl_sched_use_walt_task_util)
+ __entry->util = __entry->util_walt;
+ }
+#endif
),
- TP_printk("cpu=%d path=%s comm=%s pid=%d load=%lu util=%lu",
+ TP_printk("cpu=%d path=%s comm=%s pid=%d load=%lu util=%lu util_pelt=%lu util_walt=%lu",
__entry->cpu, __get_str(path), __entry->comm,
- __entry->pid, __entry->load, __entry->util)
+ __entry->pid, __entry->load, __entry->util,
+ __entry->util_pelt, __entry->util_walt)
);
/*
__entry->target)
);
+#ifdef CONFIG_SCHED_WALT
+struct rq;
+
+TRACE_EVENT(walt_update_task_ravg,
+
+ TP_PROTO(struct task_struct *p, struct rq *rq, int evt,
+ u64 wallclock, u64 irqtime),
+
+ TP_ARGS(p, rq, evt, wallclock, irqtime),
+
+ TP_STRUCT__entry(
+ __array( char, comm, TASK_COMM_LEN )
+ __field( pid_t, pid )
+ __field( pid_t, cur_pid )
+ __field( u64, wallclock )
+ __field( u64, mark_start )
+ __field( u64, delta_m )
+ __field( u64, win_start )
+ __field( u64, delta )
+ __field( u64, irqtime )
+ __array( char, evt, 16 )
+ __field(unsigned int, demand )
+ __field(unsigned int, sum )
+ __field( int, cpu )
+ __field( u64, cs )
+ __field( u64, ps )
+ __field( u32, curr_window )
+ __field( u32, prev_window )
+ __field( u64, nt_cs )
+ __field( u64, nt_ps )
+ __field( u32, active_windows )
+ ),
+
+ TP_fast_assign(
+ static const char* walt_event_names[] =
+ {
+ "PUT_PREV_TASK",
+ "PICK_NEXT_TASK",
+ "TASK_WAKE",
+ "TASK_MIGRATE",
+ "TASK_UPDATE",
+ "IRQ_UPDATE"
+ };
+ __entry->wallclock = wallclock;
+ __entry->win_start = rq->window_start;
+ __entry->delta = (wallclock - rq->window_start);
+ strcpy(__entry->evt, walt_event_names[evt]);
+ __entry->cpu = rq->cpu;
+ __entry->cur_pid = rq->curr->pid;
+ memcpy(__entry->comm, p->comm, TASK_COMM_LEN);
+ __entry->pid = p->pid;
+ __entry->mark_start = p->ravg.mark_start;
+ __entry->delta_m = (wallclock - p->ravg.mark_start);
+ __entry->demand = p->ravg.demand;
+ __entry->sum = p->ravg.sum;
+ __entry->irqtime = irqtime;
+ __entry->cs = rq->curr_runnable_sum;
+ __entry->ps = rq->prev_runnable_sum;
+ __entry->curr_window = p->ravg.curr_window;
+ __entry->prev_window = p->ravg.prev_window;
+ __entry->nt_cs = rq->nt_curr_runnable_sum;
+ __entry->nt_ps = rq->nt_prev_runnable_sum;
+ __entry->active_windows = p->ravg.active_windows;
+ ),
+
+ TP_printk("wallclock=%llu window_start=%llu delta=%llu event=%s cpu=%d cur_pid=%d pid=%d comm=%s"
+ " mark_start=%llu delta=%llu demand=%u sum=%u irqtime=%llu"
+ " curr_runnable_sum=%llu prev_runnable_sum=%llu cur_window=%u"
+ " prev_window=%u nt_curr_runnable_sum=%llu nt_prev_runnable_sum=%llu active_windows=%u",
+ __entry->wallclock, __entry->win_start, __entry->delta,
+ __entry->evt, __entry->cpu, __entry->cur_pid,
+ __entry->pid, __entry->comm, __entry->mark_start,
+ __entry->delta_m, __entry->demand,
+ __entry->sum, __entry->irqtime,
+ __entry->cs, __entry->ps,
+ __entry->curr_window, __entry->prev_window,
+ __entry->nt_cs, __entry->nt_ps,
+ __entry->active_windows
+ )
+);
+
+TRACE_EVENT(walt_update_history,
+
+ TP_PROTO(struct rq *rq, struct task_struct *p, u32 runtime, int samples,
+ int evt),
+
+ TP_ARGS(rq, p, runtime, samples, evt),
+
+ TP_STRUCT__entry(
+ __array( char, comm, TASK_COMM_LEN )
+ __field( pid_t, pid )
+ __field(unsigned int, runtime )
+ __field( int, samples )
+ __field( int, evt )
+ __field( u64, demand )
+ __field(unsigned int, walt_avg )
+ __field(unsigned int, pelt_avg )
+ __array( u32, hist, RAVG_HIST_SIZE_MAX)
+ __field( int, cpu )
+ ),
+
+ TP_fast_assign(
+ memcpy(__entry->comm, p->comm, TASK_COMM_LEN);
+ __entry->pid = p->pid;
+ __entry->runtime = runtime;
+ __entry->samples = samples;
+ __entry->evt = evt;
+ __entry->demand = p->ravg.demand;
+ __entry->walt_avg = (__entry->demand << 10) / walt_ravg_window,
+ __entry->pelt_avg = p->se.avg.util_avg;
+ memcpy(__entry->hist, p->ravg.sum_history,
+ RAVG_HIST_SIZE_MAX * sizeof(u32));
+ __entry->cpu = rq->cpu;
+ ),
+
+ TP_printk("pid=%d comm=%s runtime=%u samples=%d event=%d demand=%llu ravg_window=%u"
+ " walt=%u pelt=%u hist0=%u hist1=%u hist2=%u hist3=%u hist4=%u cpu=%d",
+ __entry->pid, __entry->comm,
+ __entry->runtime, __entry->samples, __entry->evt,
+ __entry->demand,
+ walt_ravg_window,
+ __entry->walt_avg,
+ __entry->pelt_avg,
+ __entry->hist[0], __entry->hist[1],
+ __entry->hist[2], __entry->hist[3],
+ __entry->hist[4], __entry->cpu)
+);
+
+TRACE_EVENT(walt_migration_update_sum,
+
+ TP_PROTO(struct rq *rq, struct task_struct *p),
+
+ TP_ARGS(rq, p),
+
+ TP_STRUCT__entry(
+ __field(int, cpu )
+ __field(int, pid )
+ __field( u64, cs )
+ __field( u64, ps )
+ __field( s64, nt_cs )
+ __field( s64, nt_ps )
+ ),
+
+ TP_fast_assign(
+ __entry->cpu = cpu_of(rq);
+ __entry->cs = rq->curr_runnable_sum;
+ __entry->ps = rq->prev_runnable_sum;
+ __entry->nt_cs = (s64)rq->nt_curr_runnable_sum;
+ __entry->nt_ps = (s64)rq->nt_prev_runnable_sum;
+ __entry->pid = p->pid;
+ ),
+
+ TP_printk("cpu=%d curr_runnable_sum=%llu prev_runnable_sum=%llu nt_curr_runnable_sum=%lld nt_prev_runnable_sum=%lld pid=%d",
+ __entry->cpu, __entry->cs, __entry->ps,
+ __entry->nt_cs, __entry->nt_ps, __entry->pid)
+);
+#endif /* CONFIG_SCHED_WALT */
#endif /* CONFIG_SMP */
#endif /* _TRACE_SCHED_H */
If in doubt, say N here.
+config SCHED_WALT
+ bool "Support window based load tracking"
+ depends on SMP
+ help
+ This feature will allow the scheduler to maintain a tunable window
+ based set of metrics for tasks and runqueues. These metrics can be
+ used to guide task placement as well as task frequency requirements
+ for cpufreq governors.
+
config BSD_PROCESS_ACCT
bool "BSD Process Accounting"
depends on MULTIUSER
obj-y += wait.o wait_bit.o swait.o completion.o idle.o
obj-$(CONFIG_SMP) += cpupri.o cpudeadline.o topology.o stop_task.o
obj-$(CONFIG_GENERIC_ARCH_TOPOLOGY) += energy.o
+obj-$(CONFIG_SCHED_WALT) += walt.o
obj-$(CONFIG_SCHED_AUTOGROUP) += autogroup.o
obj-$(CONFIG_SCHEDSTATS) += stats.o
obj-$(CONFIG_SCHED_DEBUG) += debug.o
#define CREATE_TRACE_POINTS
#include <trace/events/sched.h>
+#include "walt.h"
DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
p->sched_class->migrate_task_rq(p);
p->se.nr_migrations++;
perf_event_task_migrate(p);
+
+ walt_fixup_busy_time(p, new_cpu);
}
__set_task_cpu(p, new_cpu);
*
*/
+#ifdef CONFIG_SMP
+#ifdef CONFIG_SCHED_WALT
+/* utility function to update walt signals at wakeup */
+static inline void walt_try_to_wake_up(struct task_struct *p)
+{
+ struct rq *rq = cpu_rq(task_cpu(p));
+ struct rq_flags rf;
+ u64 wallclock;
+
+ rq_lock_irqsave(rq, &rf);
+ wallclock = walt_ktime_clock();
+ walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
+ walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
+ rq_unlock_irqrestore(rq, &rf);
+}
+#else
+#define walt_try_to_wake_up(a) {}
+#endif
+#endif
+
/**
* try_to_wake_up - wake up a thread
* @p: the thread to be awakened
*/
smp_cond_load_acquire(&p->on_cpu, !VAL);
+ walt_try_to_wake_up(p);
+
p->sched_contributes_to_load = !!task_contributes_to_load(p);
p->state = TASK_WAKING;
trace_sched_waking(p);
if (!task_on_rq_queued(p)) {
+ u64 wallclock = walt_ktime_clock();
+
+ walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
+ walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
+
if (p->in_iowait) {
delayacct_blkio_end();
atomic_dec(&rq->nr_iowait);
p->se.prev_sum_exec_runtime = 0;
p->se.nr_migrations = 0;
p->se.vruntime = 0;
+#ifdef CONFIG_SCHED_WALT
+ p->last_sleep_ts = 0;
+#endif
+
INIT_LIST_HEAD(&p->se.group_node);
+ walt_init_new_task_load(p);
#ifdef CONFIG_FAIR_GROUP_SCHED
p->se.cfs_rq = NULL;
struct rq *rq;
raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
+
+ walt_init_new_task_load(p);
+
p->state = TASK_RUNNING;
#ifdef CONFIG_SMP
/*
post_init_entity_util_avg(&p->se);
activate_task(rq, p, ENQUEUE_NOCLOCK);
+ walt_mark_task_starting(p);
+
p->on_rq = TASK_ON_RQ_QUEUED;
trace_sched_wakeup_new(p);
check_preempt_curr(rq, p, WF_FORK);
rq_lock(rq, &rf);
+ walt_set_window_start(rq, &rf);
+ walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
+ walt_ktime_clock(), 0);
update_rq_clock(rq);
curr->sched_class->task_tick(rq, curr, 0);
cpu_load_update_active(rq);
struct rq_flags rf;
struct rq *rq;
int cpu;
+ u64 wallclock;
cpu = smp_processor_id();
rq = cpu_rq(cpu);
}
next = pick_next_task(rq, prev, &rf);
+ wallclock = walt_ktime_clock();
+ walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
+ walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
clear_tsk_need_resched(prev);
clear_preempt_need_resched();
if (likely(prev != next)) {
+#ifdef CONFIG_SCHED_WALT
+ if (!prev->on_rq)
+ prev->last_sleep_ts = wallclock;
+#endif
rq->nr_switches++;
rq->curr = next;
/*
sched_ttwu_pending();
rq_lock_irqsave(rq, &rf);
+
+ walt_migrate_sync_cpu(cpu);
+
if (rq->rd) {
BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
set_rq_offline(rq);
rq->idle_stamp = 0;
rq->avg_idle = 2*sysctl_sched_migration_cost;
rq->max_idle_balance_cost = sysctl_sched_migration_cost;
+#ifdef CONFIG_SCHED_WALT
+ rq->cur_irqload = 0;
+ rq->avg_irqload = 0;
+ rq->irqload_ts = 0;
+#endif
INIT_LIST_HEAD(&rq->cfs_tasks);
#include <linux/context_tracking.h>
#include <linux/sched/cputime.h>
#include "sched.h"
+#include "walt.h"
#ifdef CONFIG_IRQ_TIME_ACCOUNTING
struct irqtime *irqtime = this_cpu_ptr(&cpu_irqtime);
s64 delta;
int cpu;
+#ifdef CONFIG_SCHED_WALT
+ u64 wallclock;
+ bool account = true;
+#endif
if (!sched_clock_irqtime)
return;
cpu = smp_processor_id();
+#ifdef CONFIG_SCHED_WALT
+ wallclock = sched_clock_cpu(cpu);
+#endif
delta = sched_clock_cpu(cpu) - irqtime->irq_start_time;
irqtime->irq_start_time += delta;
irqtime_account_delta(irqtime, delta, CPUTIME_IRQ);
else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
irqtime_account_delta(irqtime, delta, CPUTIME_SOFTIRQ);
+#ifdef CONFIG_SCHED_WALT
+ else
+ account = false;
+
+ if (account)
+ walt_account_irqtime(cpu, curr, delta, wallclock);
+#endif
}
EXPORT_SYMBOL_GPL(irqtime_account_irq);
#include <linux/slab.h>
#include <uapi/linux/sched/types.h>
+#include "walt.h"
+
struct dl_bandwidth def_dl_bandwidth;
static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
WARN_ON(!dl_prio(prio));
dl_rq->dl_nr_running++;
add_nr_running(rq_of_dl_rq(dl_rq), 1);
+ walt_inc_cumulative_runnable_avg(rq_of_dl_rq(dl_rq), dl_task_of(dl_se));
inc_dl_deadline(dl_rq, deadline);
inc_dl_migration(dl_se, dl_rq);
WARN_ON(!dl_rq->dl_nr_running);
dl_rq->dl_nr_running--;
sub_nr_running(rq_of_dl_rq(dl_rq), 1);
+ walt_dec_cumulative_runnable_avg(rq_of_dl_rq(dl_rq), dl_task_of(dl_se));
dec_dl_deadline(dl_rq, dl_se->deadline);
dec_dl_migration(dl_se, dl_rq);
deactivate_task(rq, next_task, 0);
sub_running_bw(next_task->dl.dl_bw, &rq->dl);
sub_rq_bw(next_task->dl.dl_bw, &rq->dl);
+ next_task->on_rq = TASK_ON_RQ_MIGRATING;
set_task_cpu(next_task, later_rq->cpu);
+ next_task->on_rq = TASK_ON_RQ_QUEUED;
add_rq_bw(next_task->dl.dl_bw, &later_rq->dl);
add_running_bw(next_task->dl.dl_bw, &later_rq->dl);
activate_task(later_rq, next_task, 0);
deactivate_task(src_rq, p, 0);
sub_running_bw(p->dl.dl_bw, &src_rq->dl);
sub_rq_bw(p->dl.dl_bw, &src_rq->dl);
+ p->on_rq = TASK_ON_RQ_MIGRATING;
set_task_cpu(p, this_cpu);
+ p->on_rq = TASK_ON_RQ_QUEUED;
add_rq_bw(p->dl.dl_bw, &this_rq->dl);
add_running_bw(p->dl.dl_bw, &this_rq->dl);
activate_task(this_rq, p, 0);
#include "sched.h"
#include "tune.h"
+#include "walt.h"
/*
* Targeted preemption latency for CPU-bound tasks:
const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
+#ifdef CONFIG_SCHED_WALT
+unsigned int sysctl_sched_use_walt_cpu_util = 1;
+unsigned int sysctl_sched_use_walt_task_util = 1;
+__read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
+ (10 * NSEC_PER_MSEC);
+#endif
+
#ifdef CONFIG_SMP
/*
* For asym packing, by default the lower numbered cpu has higher priority.
static unsigned long weighted_cpuload(struct rq *rq);
static unsigned long source_load(int cpu, int type);
static unsigned long target_load(int cpu, int type);
-static unsigned long capacity_of(int cpu);
/* Cached statistics for all CPUs within a node */
struct numa_stats {
* See cpu_util().
*/
cpufreq_update_util(rq, 0);
+#ifdef CONFIG_SMP
+ trace_sched_load_avg_cpu(cpu_of(rq), cfs_rq);
+#endif
}
}
set_sd_overutilized(sd);
rcu_read_unlock();
}
+
+unsigned long boosted_cpu_util(int cpu);
#else
+
#define update_overutilized_status(rq) do {} while (0)
+#define boosted_cpu_util(cpu) cpu_util_freq(cpu)
+
#endif /* CONFIG_SMP */
/*
if (cfs_rq_throttled(cfs_rq))
break;
cfs_rq->h_nr_running++;
+ walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
flags = ENQUEUE_WAKEUP;
}
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
cfs_rq->h_nr_running++;
+ walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
if (cfs_rq_throttled(cfs_rq))
break;
add_nr_running(rq, 1);
if (!task_new)
update_overutilized_status(rq);
+ walt_inc_cumulative_runnable_avg(rq, p);
}
+
hrtick_update(rq);
}
if (cfs_rq_throttled(cfs_rq))
break;
cfs_rq->h_nr_running--;
+ walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
/* Don't dequeue parent if it has other entities besides us */
if (cfs_rq->load.weight) {
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
cfs_rq->h_nr_running--;
+ walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
if (cfs_rq_throttled(cfs_rq))
break;
*/
schedtune_dequeue_task(p, cpu_of(rq));
- if (!se)
+ if (!se) {
sub_nr_running(rq, 1);
+ walt_dec_cumulative_runnable_avg(rq, p);
+ }
hrtick_update(rq);
}
return max(rq->cpu_load[type-1], total);
}
-static unsigned long capacity_of(int cpu)
-{
- return cpu_rq(cpu)->cpu_capacity;
-}
-
-static unsigned long capacity_orig_of(int cpu)
-{
- return cpu_rq(cpu)->cpu_capacity_orig;
-}
-
static unsigned long cpu_avg_load_per_task(int cpu)
{
struct rq *rq = cpu_rq(cpu);
return 0;
}
-/*
- * cpu_util returns the amount of capacity of a CPU that is used by CFS
- * tasks. The unit of the return value must be the one of capacity so we can
- * compare the utilization with the capacity of the CPU that is available for
- * CFS task (ie cpu_capacity).
- *
- * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
- * recent utilization of currently non-runnable tasks on a CPU. It represents
- * the amount of utilization of a CPU in the range [0..capacity_orig] where
- * capacity_orig is the cpu_capacity available at the highest frequency
- * (arch_scale_freq_capacity()).
- * The utilization of a CPU converges towards a sum equal to or less than the
- * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
- * the running time on this CPU scaled by capacity_curr.
- *
- * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
- * higher than capacity_orig because of unfortunate rounding in
- * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
- * the average stabilizes with the new running time. We need to check that the
- * utilization stays within the range of [0..capacity_orig] and cap it if
- * necessary. Without utilization capping, a group could be seen as overloaded
- * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
- * available capacity. We allow utilization to overshoot capacity_curr (but not
- * capacity_orig) as it useful for predicting the capacity required after task
- * migrations (scheduler-driven DVFS).
- */
-static unsigned long __cpu_util(int cpu, int delta)
-{
- unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
- unsigned long capacity = capacity_orig_of(cpu);
-
- delta += util;
- if (delta < 0)
- return 0;
-
- return (delta >= capacity) ? capacity : delta;
-}
-
-static unsigned long cpu_util(int cpu)
-{
- return __cpu_util(cpu, 0);
-}
-
static void record_wakee(struct task_struct *p)
{
/*
unsigned long
boosted_cpu_util(int cpu)
{
- unsigned long util = cpu_util(cpu);
+ unsigned long util = cpu_util_freq(cpu);
long margin = schedtune_cpu_margin(util, cpu);
trace_sched_boost_cpu(cpu, util, margin);
static inline unsigned long task_util(struct task_struct *p)
{
+#ifdef CONFIG_SCHED_WALT
+ if (!walt_disabled && sysctl_sched_use_walt_task_util) {
+ return (p->ravg.demand / (walt_ravg_window >> SCHED_CAPACITY_SHIFT));
+ }
+#endif
return p->se.avg.util_avg;
}
{
unsigned long util, capacity;
+#ifdef CONFIG_SCHED_WALT
+ /*
+ * WALT does not decay idle tasks in the same manner
+ * as PELT, so it makes little sense to subtract task
+ * utilization from cpu utilization. Instead just use
+ * cpu_util for this case.
+ */
+ if (!walt_disabled && sysctl_sched_use_walt_cpu_util)
+ return cpu_util(cpu);
+#endif
/* Task has no contribution or is new */
if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
return cpu_util(cpu);
if (!cpu_online(i))
continue;
+ if (walt_cpu_high_irqload(i))
+ continue;
+
/*
* p's blocked utilization is still accounted for on prev_cpu
* so prev_cpu will receive a negative bias due to the double
#include <linux/slab.h>
#include <linux/irq_work.h>
+#include "walt.h"
+
int sched_rr_timeslice = RR_TIMESLICE;
int sysctl_sched_rr_timeslice = (MSEC_PER_SEC / HZ) * RR_TIMESLICE;
rt_se->timeout = 0;
enqueue_rt_entity(rt_se, flags);
+ walt_inc_cumulative_runnable_avg(rq, p);
if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
enqueue_pushable_task(rq, p);
update_curr_rt(rq);
dequeue_rt_entity(rt_se, flags);
+ walt_dec_cumulative_runnable_avg(rq, p);
dequeue_pushable_task(rq, p);
}
}
deactivate_task(rq, next_task, 0);
+ next_task->on_rq = TASK_ON_RQ_MIGRATING;
set_task_cpu(next_task, lowest_rq->cpu);
+ next_task->on_rq = TASK_ON_RQ_QUEUED;
activate_task(lowest_rq, next_task, 0);
ret = 1;
resched = true;
deactivate_task(src_rq, p, 0);
+ p->on_rq = TASK_ON_RQ_MIGRATING;
set_task_cpu(p, this_cpu);
+ p->on_rq = TASK_ON_RQ_QUEUED;
activate_task(this_rq, p, 0);
/*
* We continue with the search, just in
struct list_head leaf_cfs_rq_list;
struct task_group *tg; /* group that "owns" this runqueue */
+#ifdef CONFIG_SCHED_WALT
+ u64 cumulative_runnable_avg;
+#endif
+
#ifdef CONFIG_CFS_BANDWIDTH
int runtime_enabled;
u64 runtime_expires;
u64 max_idle_balance_cost;
#endif
+#ifdef CONFIG_SCHED_WALT
+ u64 cumulative_runnable_avg;
+ u64 window_start;
+ u64 curr_runnable_sum;
+ u64 prev_runnable_sum;
+ u64 nt_curr_runnable_sum;
+ u64 nt_prev_runnable_sum;
+ u64 cur_irqload;
+ u64 avg_irqload;
+ u64 irqload_ts;
+ u64 cum_window_demand;
+#endif /* CONFIG_SCHED_WALT */
+
+
#ifdef CONFIG_IRQ_TIME_ACCOUNTING
u64 prev_irq_time;
#endif
}
#endif
+#ifdef CONFIG_SMP
+static inline unsigned long capacity_of(int cpu)
+{
+ return cpu_rq(cpu)->cpu_capacity;
+}
+
+static inline unsigned long capacity_orig_of(int cpu)
+{
+ return cpu_rq(cpu)->cpu_capacity_orig;
+}
+
+extern unsigned int sysctl_sched_use_walt_cpu_util;
+extern unsigned int walt_ravg_window;
+extern bool walt_disabled;
+
+/*
+ * cpu_util returns the amount of capacity of a CPU that is used by CFS
+ * tasks. The unit of the return value must be the one of capacity so we can
+ * compare the utilization with the capacity of the CPU that is available for
+ * CFS task (ie cpu_capacity).
+ *
+ * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
+ * recent utilization of currently non-runnable tasks on a CPU. It represents
+ * the amount of utilization of a CPU in the range [0..capacity_orig] where
+ * capacity_orig is the cpu_capacity available at the highest frequency
+ * (arch_scale_freq_capacity()).
+ * The utilization of a CPU converges towards a sum equal to or less than the
+ * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
+ * the running time on this CPU scaled by capacity_curr.
+ *
+ * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
+ * higher than capacity_orig because of unfortunate rounding in
+ * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
+ * the average stabilizes with the new running time. We need to check that the
+ * utilization stays within the range of [0..capacity_orig] and cap it if
+ * necessary. Without utilization capping, a group could be seen as overloaded
+ * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
+ * available capacity. We allow utilization to overshoot capacity_curr (but not
+ * capacity_orig) as it useful for predicting the capacity required after task
+ * migrations (scheduler-driven DVFS).
+ */
+static inline unsigned long __cpu_util(int cpu, int delta)
+{
+ unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
+ unsigned long capacity = capacity_orig_of(cpu);
+
+#ifdef CONFIG_SCHED_WALT
+ if (!walt_disabled && sysctl_sched_use_walt_cpu_util) {
+ util = cpu_rq(cpu)->cumulative_runnable_avg << SCHED_CAPACITY_SHIFT;
+ util = div_u64(util, walt_ravg_window);
+ }
+#endif
+ delta += util;
+ if (delta < 0)
+ return 0;
+
+ return (delta >= capacity) ? capacity : delta;
+}
+
+static inline unsigned long cpu_util(int cpu)
+{
+ return __cpu_util(cpu, 0);
+}
+
+static inline unsigned long cpu_util_freq(int cpu)
+{
+ unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
+ unsigned long capacity = capacity_orig_of(cpu);
+
+#ifdef CONFIG_SCHED_WALT
+ if (!walt_disabled && sysctl_sched_use_walt_cpu_util) {
+ util = cpu_rq(cpu)->prev_runnable_sum << SCHED_CAPACITY_SHIFT;
+ do_div(util, walt_ravg_window);
+ }
+#endif
+ return (util >= capacity) ? capacity : util;
+}
+
+#endif
+
static inline void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
{
rq->rt_avg += rt_delta * arch_scale_freq_capacity(NULL, cpu_of(rq));
static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
#endif /* CONFIG_CPU_FREQ */
+#ifdef CONFIG_SCHED_WALT
+
+static inline bool
+walt_task_in_cum_window_demand(struct rq *rq, struct task_struct *p)
+{
+ return cpu_of(rq) == task_cpu(p) &&
+ (p->on_rq || p->last_sleep_ts >= rq->window_start);
+}
+
+#endif /* CONFIG_SCHED_WALT */
+
#ifdef arch_scale_freq_capacity
#ifndef arch_scale_freq_invariant
#define arch_scale_freq_invariant() (true)
// SPDX-License-Identifier: GPL-2.0
#include "sched.h"
+#include "walt.h"
/*
* stop-task scheduling class.
enqueue_task_stop(struct rq *rq, struct task_struct *p, int flags)
{
add_nr_running(rq, 1);
+ walt_inc_cumulative_runnable_avg(rq, p);
}
static void
dequeue_task_stop(struct rq *rq, struct task_struct *p, int flags)
{
sub_nr_running(rq, 1);
+ walt_dec_cumulative_runnable_avg(rq, p);
}
static void yield_task_stop(struct rq *rq)
--- /dev/null
+/*
+ * Copyright (c) 2016, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ *
+ * Window Assisted Load Tracking (WALT) implementation credits:
+ * Srivatsa Vaddagiri, Steve Muckle, Syed Rameez Mustafa, Joonwoo Park,
+ * Pavan Kumar Kondeti, Olav Haugan
+ *
+ * 2016-03-06: Integration with EAS/refactoring by Vikram Mulukutla
+ * and Todd Kjos
+ */
+
+#include <linux/acpi.h>
+#include <linux/syscore_ops.h>
+#include <trace/events/sched.h>
+#include "sched.h"
+#include "walt.h"
+
+#define WINDOW_STATS_RECENT 0
+#define WINDOW_STATS_MAX 1
+#define WINDOW_STATS_MAX_RECENT_AVG 2
+#define WINDOW_STATS_AVG 3
+#define WINDOW_STATS_INVALID_POLICY 4
+
+#define EXITING_TASK_MARKER 0xdeaddead
+
+static __read_mostly unsigned int walt_ravg_hist_size = 5;
+static __read_mostly unsigned int walt_window_stats_policy =
+ WINDOW_STATS_MAX_RECENT_AVG;
+static __read_mostly unsigned int walt_account_wait_time = 1;
+static __read_mostly unsigned int walt_freq_account_wait_time = 0;
+static __read_mostly unsigned int walt_io_is_busy = 0;
+
+unsigned int sysctl_sched_walt_init_task_load_pct = 15;
+
+/* true -> use PELT based load stats, false -> use window-based load stats */
+bool __read_mostly walt_disabled = false;
+
+/*
+ * Window size (in ns). Adjust for the tick size so that the window
+ * rollover occurs just before the tick boundary.
+ */
+__read_mostly unsigned int walt_ravg_window =
+ (20000000 / TICK_NSEC) * TICK_NSEC;
+#define MIN_SCHED_RAVG_WINDOW ((10000000 / TICK_NSEC) * TICK_NSEC)
+#define MAX_SCHED_RAVG_WINDOW ((1000000000 / TICK_NSEC) * TICK_NSEC)
+
+static unsigned int sync_cpu;
+static ktime_t ktime_last;
+static __read_mostly bool walt_ktime_suspended;
+
+static unsigned int task_load(struct task_struct *p)
+{
+ return p->ravg.demand;
+}
+
+static inline void fixup_cum_window_demand(struct rq *rq, s64 delta)
+{
+ rq->cum_window_demand += delta;
+ if (unlikely((s64)rq->cum_window_demand < 0))
+ rq->cum_window_demand = 0;
+}
+
+void
+walt_inc_cumulative_runnable_avg(struct rq *rq,
+ struct task_struct *p)
+{
+ rq->cumulative_runnable_avg += p->ravg.demand;
+
+ /*
+ * Add a task's contribution to the cumulative window demand when
+ *
+ * (1) task is enqueued with on_rq = 1 i.e migration,
+ * prio/cgroup/class change.
+ * (2) task is waking for the first time in this window.
+ */
+ if (p->on_rq || (p->last_sleep_ts < rq->window_start))
+ fixup_cum_window_demand(rq, p->ravg.demand);
+}
+
+void
+walt_dec_cumulative_runnable_avg(struct rq *rq,
+ struct task_struct *p)
+{
+ rq->cumulative_runnable_avg -= p->ravg.demand;
+ BUG_ON((s64)rq->cumulative_runnable_avg < 0);
+
+ /*
+ * on_rq will be 1 for sleeping tasks. So check if the task
+ * is migrating or dequeuing in RUNNING state to change the
+ * prio/cgroup/class.
+ */
+ if (task_on_rq_migrating(p) || p->state == TASK_RUNNING)
+ fixup_cum_window_demand(rq, -(s64)p->ravg.demand);
+}
+
+static void
+fixup_cumulative_runnable_avg(struct rq *rq,
+ struct task_struct *p, u64 new_task_load)
+{
+ s64 task_load_delta = (s64)new_task_load - task_load(p);
+
+ rq->cumulative_runnable_avg += task_load_delta;
+ if ((s64)rq->cumulative_runnable_avg < 0)
+ panic("cra less than zero: tld: %lld, task_load(p) = %u\n",
+ task_load_delta, task_load(p));
+
+ fixup_cum_window_demand(rq, task_load_delta);
+}
+
+u64 walt_ktime_clock(void)
+{
+ if (unlikely(walt_ktime_suspended))
+ return ktime_to_ns(ktime_last);
+ return ktime_get_ns();
+}
+
+static void walt_resume(void)
+{
+ walt_ktime_suspended = false;
+}
+
+static int walt_suspend(void)
+{
+ ktime_last = ktime_get();
+ walt_ktime_suspended = true;
+ return 0;
+}
+
+static struct syscore_ops walt_syscore_ops = {
+ .resume = walt_resume,
+ .suspend = walt_suspend
+};
+
+static int __init walt_init_ops(void)
+{
+ register_syscore_ops(&walt_syscore_ops);
+ return 0;
+}
+late_initcall(walt_init_ops);
+
+void walt_inc_cfs_cumulative_runnable_avg(struct cfs_rq *cfs_rq,
+ struct task_struct *p)
+{
+ cfs_rq->cumulative_runnable_avg += p->ravg.demand;
+}
+
+void walt_dec_cfs_cumulative_runnable_avg(struct cfs_rq *cfs_rq,
+ struct task_struct *p)
+{
+ cfs_rq->cumulative_runnable_avg -= p->ravg.demand;
+}
+
+static int exiting_task(struct task_struct *p)
+{
+ if (p->flags & PF_EXITING) {
+ if (p->ravg.sum_history[0] != EXITING_TASK_MARKER) {
+ p->ravg.sum_history[0] = EXITING_TASK_MARKER;
+ }
+ return 1;
+ }
+ return 0;
+}
+
+static int __init set_walt_ravg_window(char *str)
+{
+ unsigned int adj_window;
+ bool no_walt = walt_disabled;
+
+ get_option(&str, &walt_ravg_window);
+
+ /* Adjust for CONFIG_HZ */
+ adj_window = (walt_ravg_window / TICK_NSEC) * TICK_NSEC;
+
+ /* Warn if we're a bit too far away from the expected window size */
+ WARN(adj_window < walt_ravg_window - NSEC_PER_MSEC,
+ "tick-adjusted window size %u, original was %u\n", adj_window,
+ walt_ravg_window);
+
+ walt_ravg_window = adj_window;
+
+ walt_disabled = walt_disabled ||
+ (walt_ravg_window < MIN_SCHED_RAVG_WINDOW ||
+ walt_ravg_window > MAX_SCHED_RAVG_WINDOW);
+
+ WARN(!no_walt && walt_disabled,
+ "invalid window size, disabling WALT\n");
+
+ return 0;
+}
+
+early_param("walt_ravg_window", set_walt_ravg_window);
+
+static void
+update_window_start(struct rq *rq, u64 wallclock)
+{
+ s64 delta;
+ int nr_windows;
+
+ delta = wallclock - rq->window_start;
+ /* If the MPM global timer is cleared, set delta as 0 to avoid kernel BUG happening */
+ if (delta < 0) {
+ delta = 0;
+ WARN_ONCE(1, "WALT wallclock appears to have gone backwards or reset\n");
+ }
+
+ if (delta < walt_ravg_window)
+ return;
+
+ nr_windows = div64_u64(delta, walt_ravg_window);
+ rq->window_start += (u64)nr_windows * (u64)walt_ravg_window;
+
+ rq->cum_window_demand = rq->cumulative_runnable_avg;
+}
+
+extern unsigned long capacity_curr_of(int cpu);
+/*
+ * Translate absolute delta time accounted on a CPU
+ * to a scale where 1024 is the capacity of the most
+ * capable CPU running at FMAX
+ */
+static u64 scale_exec_time(u64 delta, struct rq *rq)
+{
+ unsigned long capcurr = capacity_curr_of(cpu_of(rq));
+
+ return (delta * capcurr) >> SCHED_CAPACITY_SHIFT;
+}
+
+static int cpu_is_waiting_on_io(struct rq *rq)
+{
+ if (!walt_io_is_busy)
+ return 0;
+
+ return atomic_read(&rq->nr_iowait);
+}
+
+void walt_account_irqtime(int cpu, struct task_struct *curr,
+ u64 delta, u64 wallclock)
+{
+ struct rq *rq = cpu_rq(cpu);
+ unsigned long flags, nr_windows;
+ u64 cur_jiffies_ts;
+
+ raw_spin_lock_irqsave(&rq->lock, flags);
+
+ /*
+ * cputime (wallclock) uses sched_clock so use the same here for
+ * consistency.
+ */
+ delta += sched_clock() - wallclock;
+ cur_jiffies_ts = get_jiffies_64();
+
+ if (is_idle_task(curr))
+ walt_update_task_ravg(curr, rq, IRQ_UPDATE, walt_ktime_clock(),
+ delta);
+
+ nr_windows = cur_jiffies_ts - rq->irqload_ts;
+
+ if (nr_windows) {
+ if (nr_windows < 10) {
+ /* Decay CPU's irqload by 3/4 for each window. */
+ rq->avg_irqload *= (3 * nr_windows);
+ rq->avg_irqload = div64_u64(rq->avg_irqload,
+ 4 * nr_windows);
+ } else {
+ rq->avg_irqload = 0;
+ }
+ rq->avg_irqload += rq->cur_irqload;
+ rq->cur_irqload = 0;
+ }
+
+ rq->cur_irqload += delta;
+ rq->irqload_ts = cur_jiffies_ts;
+ raw_spin_unlock_irqrestore(&rq->lock, flags);
+}
+
+
+#define WALT_HIGH_IRQ_TIMEOUT 3
+
+u64 walt_irqload(int cpu) {
+ struct rq *rq = cpu_rq(cpu);
+ s64 delta;
+ delta = get_jiffies_64() - rq->irqload_ts;
+
+ /*
+ * Current context can be preempted by irq and rq->irqload_ts can be
+ * updated by irq context so that delta can be negative.
+ * But this is okay and we can safely return as this means there
+ * was recent irq occurrence.
+ */
+
+ if (delta < WALT_HIGH_IRQ_TIMEOUT)
+ return rq->avg_irqload;
+ else
+ return 0;
+}
+
+int walt_cpu_high_irqload(int cpu) {
+ return walt_irqload(cpu) >= sysctl_sched_walt_cpu_high_irqload;
+}
+
+static int account_busy_for_cpu_time(struct rq *rq, struct task_struct *p,
+ u64 irqtime, int event)
+{
+ if (is_idle_task(p)) {
+ /* TASK_WAKE && TASK_MIGRATE is not possible on idle task! */
+ if (event == PICK_NEXT_TASK)
+ return 0;
+
+ /* PUT_PREV_TASK, TASK_UPDATE && IRQ_UPDATE are left */
+ return irqtime || cpu_is_waiting_on_io(rq);
+ }
+
+ if (event == TASK_WAKE)
+ return 0;
+
+ if (event == PUT_PREV_TASK || event == IRQ_UPDATE ||
+ event == TASK_UPDATE)
+ return 1;
+
+ /* Only TASK_MIGRATE && PICK_NEXT_TASK left */
+ return walt_freq_account_wait_time;
+}
+
+/*
+ * Account cpu activity in its busy time counters (rq->curr/prev_runnable_sum)
+ */
+static void update_cpu_busy_time(struct task_struct *p, struct rq *rq,
+ int event, u64 wallclock, u64 irqtime)
+{
+ int new_window, nr_full_windows = 0;
+ int p_is_curr_task = (p == rq->curr);
+ u64 mark_start = p->ravg.mark_start;
+ u64 window_start = rq->window_start;
+ u32 window_size = walt_ravg_window;
+ u64 delta;
+
+ new_window = mark_start < window_start;
+ if (new_window) {
+ nr_full_windows = div64_u64((window_start - mark_start),
+ window_size);
+ if (p->ravg.active_windows < USHRT_MAX)
+ p->ravg.active_windows++;
+ }
+
+ /* Handle per-task window rollover. We don't care about the idle
+ * task or exiting tasks. */
+ if (new_window && !is_idle_task(p) && !exiting_task(p)) {
+ u32 curr_window = 0;
+
+ if (!nr_full_windows)
+ curr_window = p->ravg.curr_window;
+
+ p->ravg.prev_window = curr_window;
+ p->ravg.curr_window = 0;
+ }
+
+ if (!account_busy_for_cpu_time(rq, p, irqtime, event)) {
+ /* account_busy_for_cpu_time() = 0, so no update to the
+ * task's current window needs to be made. This could be
+ * for example
+ *
+ * - a wakeup event on a task within the current
+ * window (!new_window below, no action required),
+ * - switching to a new task from idle (PICK_NEXT_TASK)
+ * in a new window where irqtime is 0 and we aren't
+ * waiting on IO */
+
+ if (!new_window)
+ return;
+
+ /* A new window has started. The RQ demand must be rolled
+ * over if p is the current task. */
+ if (p_is_curr_task) {
+ u64 prev_sum = 0;
+
+ /* p is either idle task or an exiting task */
+ if (!nr_full_windows) {
+ prev_sum = rq->curr_runnable_sum;
+ }
+
+ rq->prev_runnable_sum = prev_sum;
+ rq->curr_runnable_sum = 0;
+ }
+
+ return;
+ }
+
+ if (!new_window) {
+ /* account_busy_for_cpu_time() = 1 so busy time needs
+ * to be accounted to the current window. No rollover
+ * since we didn't start a new window. An example of this is
+ * when a task starts execution and then sleeps within the
+ * same window. */
+
+ if (!irqtime || !is_idle_task(p) || cpu_is_waiting_on_io(rq))
+ delta = wallclock - mark_start;
+ else
+ delta = irqtime;
+ delta = scale_exec_time(delta, rq);
+ rq->curr_runnable_sum += delta;
+ if (!is_idle_task(p) && !exiting_task(p))
+ p->ravg.curr_window += delta;
+
+ return;
+ }
+
+ if (!p_is_curr_task) {
+ /* account_busy_for_cpu_time() = 1 so busy time needs
+ * to be accounted to the current window. A new window
+ * has also started, but p is not the current task, so the
+ * window is not rolled over - just split up and account
+ * as necessary into curr and prev. The window is only
+ * rolled over when a new window is processed for the current
+ * task.
+ *
+ * Irqtime can't be accounted by a task that isn't the
+ * currently running task. */
+
+ if (!nr_full_windows) {
+ /* A full window hasn't elapsed, account partial
+ * contribution to previous completed window. */
+ delta = scale_exec_time(window_start - mark_start, rq);
+ if (!exiting_task(p))
+ p->ravg.prev_window += delta;
+ } else {
+ /* Since at least one full window has elapsed,
+ * the contribution to the previous window is the
+ * full window (window_size). */
+ delta = scale_exec_time(window_size, rq);
+ if (!exiting_task(p))
+ p->ravg.prev_window = delta;
+ }
+ rq->prev_runnable_sum += delta;
+
+ /* Account piece of busy time in the current window. */
+ delta = scale_exec_time(wallclock - window_start, rq);
+ rq->curr_runnable_sum += delta;
+ if (!exiting_task(p))
+ p->ravg.curr_window = delta;
+
+ return;
+ }
+
+ if (!irqtime || !is_idle_task(p) || cpu_is_waiting_on_io(rq)) {
+ /* account_busy_for_cpu_time() = 1 so busy time needs
+ * to be accounted to the current window. A new window
+ * has started and p is the current task so rollover is
+ * needed. If any of these three above conditions are true
+ * then this busy time can't be accounted as irqtime.
+ *
+ * Busy time for the idle task or exiting tasks need not
+ * be accounted.
+ *
+ * An example of this would be a task that starts execution
+ * and then sleeps once a new window has begun. */
+
+ if (!nr_full_windows) {
+ /* A full window hasn't elapsed, account partial
+ * contribution to previous completed window. */
+ delta = scale_exec_time(window_start - mark_start, rq);
+ if (!is_idle_task(p) && !exiting_task(p))
+ p->ravg.prev_window += delta;
+
+ delta += rq->curr_runnable_sum;
+ } else {
+ /* Since at least one full window has elapsed,
+ * the contribution to the previous window is the
+ * full window (window_size). */
+ delta = scale_exec_time(window_size, rq);
+ if (!is_idle_task(p) && !exiting_task(p))
+ p->ravg.prev_window = delta;
+
+ }
+ /*
+ * Rollover for normal runnable sum is done here by overwriting
+ * the values in prev_runnable_sum and curr_runnable_sum.
+ * Rollover for new task runnable sum has completed by previous
+ * if-else statement.
+ */
+ rq->prev_runnable_sum = delta;
+
+ /* Account piece of busy time in the current window. */
+ delta = scale_exec_time(wallclock - window_start, rq);
+ rq->curr_runnable_sum = delta;
+ if (!is_idle_task(p) && !exiting_task(p))
+ p->ravg.curr_window = delta;
+
+ return;
+ }
+
+ if (irqtime) {
+ /* account_busy_for_cpu_time() = 1 so busy time needs
+ * to be accounted to the current window. A new window
+ * has started and p is the current task so rollover is
+ * needed. The current task must be the idle task because
+ * irqtime is not accounted for any other task.
+ *
+ * Irqtime will be accounted each time we process IRQ activity
+ * after a period of idleness, so we know the IRQ busy time
+ * started at wallclock - irqtime. */
+
+ BUG_ON(!is_idle_task(p));
+ mark_start = wallclock - irqtime;
+
+ /* Roll window over. If IRQ busy time was just in the current
+ * window then that is all that need be accounted. */
+ rq->prev_runnable_sum = rq->curr_runnable_sum;
+ if (mark_start > window_start) {
+ rq->curr_runnable_sum = scale_exec_time(irqtime, rq);
+ return;
+ }
+
+ /* The IRQ busy time spanned multiple windows. Process the
+ * busy time preceding the current window start first. */
+ delta = window_start - mark_start;
+ if (delta > window_size)
+ delta = window_size;
+ delta = scale_exec_time(delta, rq);
+ rq->prev_runnable_sum += delta;
+
+ /* Process the remaining IRQ busy time in the current window. */
+ delta = wallclock - window_start;
+ rq->curr_runnable_sum = scale_exec_time(delta, rq);
+
+ return;
+ }
+
+ BUG();
+}
+
+static int account_busy_for_task_demand(struct task_struct *p, int event)
+{
+ /* No need to bother updating task demand for exiting tasks
+ * or the idle task. */
+ if (exiting_task(p) || is_idle_task(p))
+ return 0;
+
+ /* When a task is waking up it is completing a segment of non-busy
+ * time. Likewise, if wait time is not treated as busy time, then
+ * when a task begins to run or is migrated, it is not running and
+ * is completing a segment of non-busy time. */
+ if (event == TASK_WAKE || (!walt_account_wait_time &&
+ (event == PICK_NEXT_TASK || event == TASK_MIGRATE)))
+ return 0;
+
+ return 1;
+}
+
+/*
+ * Called when new window is starting for a task, to record cpu usage over
+ * recently concluded window(s). Normally 'samples' should be 1. It can be > 1
+ * when, say, a real-time task runs without preemption for several windows at a
+ * stretch.
+ */
+static void update_history(struct rq *rq, struct task_struct *p,
+ u32 runtime, int samples, int event)
+{
+ u32 *hist = &p->ravg.sum_history[0];
+ int ridx, widx;
+ u32 max = 0, avg, demand;
+ u64 sum = 0;
+
+ /* Ignore windows where task had no activity */
+ if (!runtime || is_idle_task(p) || exiting_task(p) || !samples)
+ goto done;
+
+ /* Push new 'runtime' value onto stack */
+ widx = walt_ravg_hist_size - 1;
+ ridx = widx - samples;
+ for (; ridx >= 0; --widx, --ridx) {
+ hist[widx] = hist[ridx];
+ sum += hist[widx];
+ if (hist[widx] > max)
+ max = hist[widx];
+ }
+
+ for (widx = 0; widx < samples && widx < walt_ravg_hist_size; widx++) {
+ hist[widx] = runtime;
+ sum += hist[widx];
+ if (hist[widx] > max)
+ max = hist[widx];
+ }
+
+ p->ravg.sum = 0;
+
+ if (walt_window_stats_policy == WINDOW_STATS_RECENT) {
+ demand = runtime;
+ } else if (walt_window_stats_policy == WINDOW_STATS_MAX) {
+ demand = max;
+ } else {
+ avg = div64_u64(sum, walt_ravg_hist_size);
+ if (walt_window_stats_policy == WINDOW_STATS_AVG)
+ demand = avg;
+ else
+ demand = max(avg, runtime);
+ }
+
+ /*
+ * A throttled deadline sched class task gets dequeued without
+ * changing p->on_rq. Since the dequeue decrements hmp stats
+ * avoid decrementing it here again.
+ *
+ * When window is rolled over, the cumulative window demand
+ * is reset to the cumulative runnable average (contribution from
+ * the tasks on the runqueue). If the current task is dequeued
+ * already, it's demand is not included in the cumulative runnable
+ * average. So add the task demand separately to cumulative window
+ * demand.
+ */
+ if (!task_has_dl_policy(p) || !p->dl.dl_throttled) {
+ if (task_on_rq_queued(p))
+ fixup_cumulative_runnable_avg(rq, p, demand);
+ else if (rq->curr == p)
+ fixup_cum_window_demand(rq, demand);
+ }
+
+ p->ravg.demand = demand;
+
+done:
+ trace_walt_update_history(rq, p, runtime, samples, event);
+ return;
+}
+
+static void add_to_task_demand(struct rq *rq, struct task_struct *p,
+ u64 delta)
+{
+ delta = scale_exec_time(delta, rq);
+ p->ravg.sum += delta;
+ if (unlikely(p->ravg.sum > walt_ravg_window))
+ p->ravg.sum = walt_ravg_window;
+}
+
+/*
+ * Account cpu demand of task and/or update task's cpu demand history
+ *
+ * ms = p->ravg.mark_start;
+ * wc = wallclock
+ * ws = rq->window_start
+ *
+ * Three possibilities:
+ *
+ * a) Task event is contained within one window.
+ * window_start < mark_start < wallclock
+ *
+ * ws ms wc
+ * | | |
+ * V V V
+ * |---------------|
+ *
+ * In this case, p->ravg.sum is updated *iff* event is appropriate
+ * (ex: event == PUT_PREV_TASK)
+ *
+ * b) Task event spans two windows.
+ * mark_start < window_start < wallclock
+ *
+ * ms ws wc
+ * | | |
+ * V V V
+ * -----|-------------------
+ *
+ * In this case, p->ravg.sum is updated with (ws - ms) *iff* event
+ * is appropriate, then a new window sample is recorded followed
+ * by p->ravg.sum being set to (wc - ws) *iff* event is appropriate.
+ *
+ * c) Task event spans more than two windows.
+ *
+ * ms ws_tmp ws wc
+ * | | | |
+ * V V V V
+ * ---|-------|-------|-------|-------|------
+ * | |
+ * |<------ nr_full_windows ------>|
+ *
+ * In this case, p->ravg.sum is updated with (ws_tmp - ms) first *iff*
+ * event is appropriate, window sample of p->ravg.sum is recorded,
+ * 'nr_full_window' samples of window_size is also recorded *iff*
+ * event is appropriate and finally p->ravg.sum is set to (wc - ws)
+ * *iff* event is appropriate.
+ *
+ * IMPORTANT : Leave p->ravg.mark_start unchanged, as update_cpu_busy_time()
+ * depends on it!
+ */
+static void update_task_demand(struct task_struct *p, struct rq *rq,
+ int event, u64 wallclock)
+{
+ u64 mark_start = p->ravg.mark_start;
+ u64 delta, window_start = rq->window_start;
+ int new_window, nr_full_windows;
+ u32 window_size = walt_ravg_window;
+
+ new_window = mark_start < window_start;
+ if (!account_busy_for_task_demand(p, event)) {
+ if (new_window)
+ /* If the time accounted isn't being accounted as
+ * busy time, and a new window started, only the
+ * previous window need be closed out with the
+ * pre-existing demand. Multiple windows may have
+ * elapsed, but since empty windows are dropped,
+ * it is not necessary to account those. */
+ update_history(rq, p, p->ravg.sum, 1, event);
+ return;
+ }
+
+ if (!new_window) {
+ /* The simple case - busy time contained within the existing
+ * window. */
+ add_to_task_demand(rq, p, wallclock - mark_start);
+ return;
+ }
+
+ /* Busy time spans at least two windows. Temporarily rewind
+ * window_start to first window boundary after mark_start. */
+ delta = window_start - mark_start;
+ nr_full_windows = div64_u64(delta, window_size);
+ window_start -= (u64)nr_full_windows * (u64)window_size;
+
+ /* Process (window_start - mark_start) first */
+ add_to_task_demand(rq, p, window_start - mark_start);
+
+ /* Push new sample(s) into task's demand history */
+ update_history(rq, p, p->ravg.sum, 1, event);
+ if (nr_full_windows)
+ update_history(rq, p, scale_exec_time(window_size, rq),
+ nr_full_windows, event);
+
+ /* Roll window_start back to current to process any remainder
+ * in current window. */
+ window_start += (u64)nr_full_windows * (u64)window_size;
+
+ /* Process (wallclock - window_start) next */
+ mark_start = window_start;
+ add_to_task_demand(rq, p, wallclock - mark_start);
+}
+
+/* Reflect task activity on its demand and cpu's busy time statistics */
+void walt_update_task_ravg(struct task_struct *p, struct rq *rq,
+ int event, u64 wallclock, u64 irqtime)
+{
+ if (walt_disabled || !rq->window_start)
+ return;
+
+ lockdep_assert_held(&rq->lock);
+
+ update_window_start(rq, wallclock);
+
+ if (!p->ravg.mark_start)
+ goto done;
+
+ update_task_demand(p, rq, event, wallclock);
+ update_cpu_busy_time(p, rq, event, wallclock, irqtime);
+
+done:
+ trace_walt_update_task_ravg(p, rq, event, wallclock, irqtime);
+
+ p->ravg.mark_start = wallclock;
+}
+
+static void reset_task_stats(struct task_struct *p)
+{
+ u32 sum = 0;
+
+ if (exiting_task(p))
+ sum = EXITING_TASK_MARKER;
+
+ memset(&p->ravg, 0, sizeof(struct ravg));
+ /* Retain EXITING_TASK marker */
+ p->ravg.sum_history[0] = sum;
+}
+
+void walt_mark_task_starting(struct task_struct *p)
+{
+ u64 wallclock;
+ struct rq *rq = task_rq(p);
+
+ if (!rq->window_start) {
+ reset_task_stats(p);
+ return;
+ }
+
+ wallclock = walt_ktime_clock();
+ p->ravg.mark_start = wallclock;
+}
+
+void walt_set_window_start(struct rq *rq, struct rq_flags *rf)
+{
+ if (likely(rq->window_start))
+ return;
+
+ if (cpu_of(rq) == sync_cpu) {
+ rq->window_start = 1;
+ } else {
+ struct rq *sync_rq = cpu_rq(sync_cpu);
+ rq_unpin_lock(rq, rf);
+ double_lock_balance(rq, sync_rq);
+ rq->window_start = sync_rq->window_start;
+ rq->curr_runnable_sum = rq->prev_runnable_sum = 0;
+ raw_spin_unlock(&sync_rq->lock);
+ rq_repin_lock(rq, rf);
+ }
+
+ rq->curr->ravg.mark_start = rq->window_start;
+}
+
+void walt_migrate_sync_cpu(int cpu)
+{
+ if (cpu == sync_cpu)
+ sync_cpu = smp_processor_id();
+}
+
+void walt_fixup_busy_time(struct task_struct *p, int new_cpu)
+{
+ struct rq *src_rq = task_rq(p);
+ struct rq *dest_rq = cpu_rq(new_cpu);
+ u64 wallclock;
+
+ if (!p->on_rq && p->state != TASK_WAKING)
+ return;
+
+ if (exiting_task(p)) {
+ return;
+ }
+
+ if (p->state == TASK_WAKING)
+ double_rq_lock(src_rq, dest_rq);
+
+ wallclock = walt_ktime_clock();
+
+ walt_update_task_ravg(task_rq(p)->curr, task_rq(p),
+ TASK_UPDATE, wallclock, 0);
+ walt_update_task_ravg(dest_rq->curr, dest_rq,
+ TASK_UPDATE, wallclock, 0);
+
+ walt_update_task_ravg(p, task_rq(p), TASK_MIGRATE, wallclock, 0);
+
+ /*
+ * When a task is migrating during the wakeup, adjust
+ * the task's contribution towards cumulative window
+ * demand.
+ */
+ if (p->state == TASK_WAKING &&
+ p->last_sleep_ts >= src_rq->window_start) {
+ fixup_cum_window_demand(src_rq, -(s64)p->ravg.demand);
+ fixup_cum_window_demand(dest_rq, p->ravg.demand);
+ }
+
+ if (p->ravg.curr_window) {
+ src_rq->curr_runnable_sum -= p->ravg.curr_window;
+ dest_rq->curr_runnable_sum += p->ravg.curr_window;
+ }
+
+ if (p->ravg.prev_window) {
+ src_rq->prev_runnable_sum -= p->ravg.prev_window;
+ dest_rq->prev_runnable_sum += p->ravg.prev_window;
+ }
+
+ if ((s64)src_rq->prev_runnable_sum < 0) {
+ src_rq->prev_runnable_sum = 0;
+ WARN_ON(1);
+ }
+ if ((s64)src_rq->curr_runnable_sum < 0) {
+ src_rq->curr_runnable_sum = 0;
+ WARN_ON(1);
+ }
+
+ trace_walt_migration_update_sum(src_rq, p);
+ trace_walt_migration_update_sum(dest_rq, p);
+
+ if (p->state == TASK_WAKING)
+ double_rq_unlock(src_rq, dest_rq);
+}
+
+void walt_init_new_task_load(struct task_struct *p)
+{
+ int i;
+ u32 init_load_windows =
+ div64_u64((u64)sysctl_sched_walt_init_task_load_pct *
+ (u64)walt_ravg_window, 100);
+ u32 init_load_pct = current->init_load_pct;
+
+ p->init_load_pct = 0;
+ memset(&p->ravg, 0, sizeof(struct ravg));
+
+ if (init_load_pct) {
+ init_load_windows = div64_u64((u64)init_load_pct *
+ (u64)walt_ravg_window, 100);
+ }
+
+ p->ravg.demand = init_load_windows;
+ for (i = 0; i < RAVG_HIST_SIZE_MAX; ++i)
+ p->ravg.sum_history[i] = init_load_windows;
+}
--- /dev/null
+/*
+ * Copyright (c) 2016, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ */
+
+#ifndef __WALT_H
+#define __WALT_H
+
+#ifdef CONFIG_SCHED_WALT
+
+void walt_update_task_ravg(struct task_struct *p, struct rq *rq, int event,
+ u64 wallclock, u64 irqtime);
+void walt_inc_cumulative_runnable_avg(struct rq *rq, struct task_struct *p);
+void walt_dec_cumulative_runnable_avg(struct rq *rq, struct task_struct *p);
+void walt_inc_cfs_cumulative_runnable_avg(struct cfs_rq *rq,
+ struct task_struct *p);
+void walt_dec_cfs_cumulative_runnable_avg(struct cfs_rq *rq,
+ struct task_struct *p);
+void walt_fixup_busy_time(struct task_struct *p, int new_cpu);
+void walt_init_new_task_load(struct task_struct *p);
+void walt_mark_task_starting(struct task_struct *p);
+void walt_set_window_start(struct rq *rq, struct rq_flags *rf);
+void walt_migrate_sync_cpu(int cpu);
+u64 walt_ktime_clock(void);
+void walt_account_irqtime(int cpu, struct task_struct *curr, u64 delta,
+ u64 wallclock);
+
+u64 walt_irqload(int cpu);
+int walt_cpu_high_irqload(int cpu);
+
+#else /* CONFIG_SCHED_WALT */
+
+static inline void walt_update_task_ravg(struct task_struct *p, struct rq *rq,
+ int event, u64 wallclock, u64 irqtime) { }
+static inline void walt_inc_cumulative_runnable_avg(struct rq *rq, struct task_struct *p) { }
+static inline void walt_dec_cumulative_runnable_avg(struct rq *rq, struct task_struct *p) { }
+static inline void walt_inc_cfs_cumulative_runnable_avg(struct cfs_rq *rq,
+ struct task_struct *p) { }
+static inline void walt_dec_cfs_cumulative_runnable_avg(struct cfs_rq *rq,
+ struct task_struct *p) { }
+static inline void walt_fixup_busy_time(struct task_struct *p, int new_cpu) { }
+static inline void walt_init_new_task_load(struct task_struct *p) { }
+static inline void walt_mark_task_starting(struct task_struct *p) { }
+static inline void walt_set_window_start(struct rq *rq, struct rq_flags *rf) { }
+static inline void walt_migrate_sync_cpu(int cpu) { }
+static inline u64 walt_ktime_clock(void) { return 0; }
+
+#define walt_cpu_high_irqload(cpu) false
+
+#endif /* CONFIG_SCHED_WALT */
+
+extern bool walt_disabled;
+
+#endif
.extra1 = &min_sched_granularity_ns,
.extra2 = &max_sched_granularity_ns,
},
+#ifdef CONFIG_SCHED_WALT
+ {
+ .procname = "sched_use_walt_cpu_util",
+ .data = &sysctl_sched_use_walt_cpu_util,
+ .maxlen = sizeof(unsigned int),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ },
+ {
+ .procname = "sched_use_walt_task_util",
+ .data = &sysctl_sched_use_walt_task_util,
+ .maxlen = sizeof(unsigned int),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ },
+ {
+ .procname = "sched_walt_init_task_load_pct",
+ .data = &sysctl_sched_walt_init_task_load_pct,
+ .maxlen = sizeof(unsigned int),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ },
+ {
+ .procname = "sched_walt_cpu_high_irqload",
+ .data = &sysctl_sched_walt_cpu_high_irqload,
+ .maxlen = sizeof(unsigned int),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ },
+#endif
{
.procname = "sched_sync_hint_enable",
.data = &sysctl_sched_sync_hint_enable,