FROMLIST: psi: track changed states
authorSuren Baghdasaryan <surenb@google.com>
Wed, 6 Mar 2019 18:25:50 +0000 (10:25 -0800)
committerlingsen1 <lingsen1@lenovo.com>
Sun, 7 Feb 2021 09:37:07 +0000 (17:37 +0800)
Introduce changed_states parameter into collect_percpu_times to track
the states changed since the last update.

Signed-off-by: Suren Baghdasaryan <surenb@google.com>
(not upstream yet, latest version published at: https://lore.kernel.org/patchwork/patch/1052420/)

Bug: 127712811
Bug: 129157727
Test: lmkd in PSI mode
Change-Id: Idb2f7d73013bff16bb101b62a2609917a5353bf9
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
kernel/sched/psi.c

index ace5ed97b186228bc2bfea362ac146954f85468d..1b99eeffaa257f388ac2ffcdcb211fda38c8765d 100644 (file)
@@ -210,7 +210,8 @@ static bool test_state(unsigned int *tasks, enum psi_states state)
        }
 }
 
-static void get_recent_times(struct psi_group *group, int cpu, u32 *times)
+static void get_recent_times(struct psi_group *group, int cpu, u32 *times,
+                            u32 *pchanged_states)
 {
        struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
        u64 now, state_start;
@@ -218,6 +219,8 @@ static void get_recent_times(struct psi_group *group, int cpu, u32 *times)
        unsigned int seq;
        u32 state_mask;
 
+       *pchanged_states = 0;
+
        /* Snapshot a coherent view of the CPU state */
        do {
                seq = read_seqcount_begin(&groupc->seq);
@@ -246,6 +249,8 @@ static void get_recent_times(struct psi_group *group, int cpu, u32 *times)
                groupc->times_prev[s] = times[s];
 
                times[s] = delta;
+               if (delta)
+                       *pchanged_states |= (1 << s);
        }
 }
 
@@ -269,10 +274,11 @@ static void calc_avgs(unsigned long avg[3], int missed_periods,
        avg[2] = calc_load(avg[2], EXP_300s, pct);
 }
 
-static bool collect_percpu_times(struct psi_group *group)
+static void collect_percpu_times(struct psi_group *group, u32 *pchanged_states)
 {
        u64 deltas[NR_PSI_STATES - 1] = { 0, };
        unsigned long nonidle_total = 0;
+       u32 changed_states = 0;
        int cpu;
        int s;
 
@@ -287,8 +293,11 @@ static bool collect_percpu_times(struct psi_group *group)
        for_each_possible_cpu(cpu) {
                u32 times[NR_PSI_STATES];
                u32 nonidle;
+               u32 cpu_changed_states;
 
-               get_recent_times(group, cpu, times);
+               get_recent_times(group, cpu, times,
+                               &cpu_changed_states);
+               changed_states |= cpu_changed_states;
 
                nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]);
                nonidle_total += nonidle;
@@ -313,7 +322,8 @@ static bool collect_percpu_times(struct psi_group *group)
        for (s = 0; s < NR_PSI_STATES - 1; s++)
                group->total[s] += div_u64(deltas[s], max(nonidle_total, 1UL));
 
-       return nonidle_total;
+       if (pchanged_states)
+               *pchanged_states = changed_states;
 }
 
 static u64 update_averages(struct psi_group *group, u64 now)
@@ -373,6 +383,7 @@ static void psi_avgs_work(struct work_struct *work)
 {
        struct delayed_work *dwork;
        struct psi_group *group;
+       u32 changed_states;
        bool nonidle;
        u64 now;
 
@@ -383,7 +394,8 @@ static void psi_avgs_work(struct work_struct *work)
 
        now = sched_clock();
 
-       nonidle = collect_percpu_times(group);
+       collect_percpu_times(group, &changed_states);
+       nonidle = changed_states & (1 << PSI_NONIDLE);
        /*
         * If there is task activity, periodically fold the per-cpu
         * times and feed samples into the running averages. If things
@@ -719,7 +731,7 @@ int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res)
        /* Update averages before reporting them */
        mutex_lock(&group->avgs_lock);
        now = sched_clock();
-       collect_percpu_times(group);
+       collect_percpu_times(group, NULL);
        if (now >= group->avg_next_update)
                group->avg_next_update = update_averages(group, now);
        mutex_unlock(&group->avgs_lock);