]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
sched/psi: Optimize psi_group_change() cpu_clock() usage
authorPeter Zijlstra <peterz@infradead.org>
Fri, 23 May 2025 15:28:00 +0000 (17:28 +0200)
committerPeter Zijlstra <peterz@infradead.org>
Wed, 9 Jul 2025 11:40:21 +0000 (13:40 +0200)
Dietmar reported that commit 3840cbe24cf0 ("sched: psi: fix bogus
pressure spikes from aggregation race") caused a regression for him on
a high context switch rate benchmark (schbench) due to the now
repeating cpu_clock() calls.

In particular the problem is that get_recent_times() will extrapolate
the current state to 'now'. But if an update uses a timestamp from
before the start of the update, it is possible to get two reads
with inconsistent results. It is effectively back-dating an update.

(note that this all hard-relies on the clock being synchronized across
CPUs -- if this is not the case, all bets are off).

Combine this problem with the fact that there are per-group-per-cpu
seqcounts, the commit in question pushed the clock read into the group
iteration, causing tree-depth cpu_clock() calls. On architectures
where cpu_clock() has appreciable overhead, this hurts.

Instead move to a per-cpu seqcount, which allows us to have a single
clock read for all group updates, increasing internal consistency and
lowering update overhead. This comes at the cost of a longer update
side (proportional to the tree depth) which can cause the read side to
retry more often.

Fixes: 3840cbe24cf0 ("sched: psi: fix bogus pressure spikes from aggregation race")
Reported-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com>,
Link: https://lkml.kernel.org/20250522084844.GC31726@noisy.programming.kicks-ass.net
include/linux/psi_types.h
kernel/sched/psi.c

index f1fd3a8044e0eca4fcea4487243bd6b0e438aaf4..dd10c22299ab82639605dfe0cbfae13a25f76318 100644 (file)
@@ -84,11 +84,9 @@ enum psi_aggregators {
 struct psi_group_cpu {
        /* 1st cacheline updated by the scheduler */
 
-       /* Aggregator needs to know of concurrent changes */
-       seqcount_t seq ____cacheline_aligned_in_smp;
-
        /* States of the tasks belonging to this group */
-       unsigned int tasks[NR_PSI_TASK_COUNTS];
+       unsigned int tasks[NR_PSI_TASK_COUNTS]
+                       ____cacheline_aligned_in_smp;
 
        /* Aggregate pressure state derived from the tasks */
        u32 state_mask;
index 5837cd8e7b97e055af0257cb2676174d259a47d2..2024c1d36402d4dde9b7f0a3b73ca8cf1a1681e0 100644 (file)
@@ -176,6 +176,28 @@ struct psi_group psi_system = {
        .pcpu = &system_group_pcpu,
 };
 
+static DEFINE_PER_CPU(seqcount_t, psi_seq);
+
+static inline void psi_write_begin(int cpu)
+{
+       write_seqcount_begin(per_cpu_ptr(&psi_seq, cpu));
+}
+
+static inline void psi_write_end(int cpu)
+{
+       write_seqcount_end(per_cpu_ptr(&psi_seq, cpu));
+}
+
+static inline u32 psi_read_begin(int cpu)
+{
+       return read_seqcount_begin(per_cpu_ptr(&psi_seq, cpu));
+}
+
+static inline bool psi_read_retry(int cpu, u32 seq)
+{
+       return read_seqcount_retry(per_cpu_ptr(&psi_seq, cpu), seq);
+}
+
 static void psi_avgs_work(struct work_struct *work);
 
 static void poll_timer_fn(struct timer_list *t);
@@ -186,7 +208,7 @@ static void group_init(struct psi_group *group)
 
        group->enabled = true;
        for_each_possible_cpu(cpu)
-               seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq);
+               seqcount_init(per_cpu_ptr(&psi_seq, cpu));
        group->avg_last_update = sched_clock();
        group->avg_next_update = group->avg_last_update + psi_period;
        mutex_init(&group->avgs_lock);
@@ -266,14 +288,14 @@ static void get_recent_times(struct psi_group *group, int cpu,
 
        /* Snapshot a coherent view of the CPU state */
        do {
-               seq = read_seqcount_begin(&groupc->seq);
+               seq = psi_read_begin(cpu);
                now = cpu_clock(cpu);
                memcpy(times, groupc->times, sizeof(groupc->times));
                state_mask = groupc->state_mask;
                state_start = groupc->state_start;
                if (cpu == current_cpu)
                        memcpy(tasks, groupc->tasks, sizeof(groupc->tasks));
-       } while (read_seqcount_retry(&groupc->seq, seq));
+       } while (psi_read_retry(cpu, seq));
 
        /* Calculate state time deltas against the previous snapshot */
        for (s = 0; s < NR_PSI_STATES; s++) {
@@ -772,30 +794,20 @@ static void record_times(struct psi_group_cpu *groupc, u64 now)
                groupc->times[PSI_NONIDLE] += delta;
 }
 
+#define for_each_group(iter, group) \
+       for (typeof(group) iter = group; iter; iter = iter->parent)
+
 static void psi_group_change(struct psi_group *group, int cpu,
                             unsigned int clear, unsigned int set,
-                            bool wake_clock)
+                            u64 now, bool wake_clock)
 {
        struct psi_group_cpu *groupc;
        unsigned int t, m;
        u32 state_mask;
-       u64 now;
 
        lockdep_assert_rq_held(cpu_rq(cpu));
        groupc = per_cpu_ptr(group->pcpu, cpu);
 
-       /*
-        * First we update the task counts according to the state
-        * change requested through the @clear and @set bits.
-        *
-        * Then if the cgroup PSI stats accounting enabled, we
-        * assess the aggregate resource states this CPU's tasks
-        * have been in since the last change, and account any
-        * SOME and FULL time these may have resulted in.
-        */
-       write_seqcount_begin(&groupc->seq);
-       now = cpu_clock(cpu);
-
        /*
         * Start with TSK_ONCPU, which doesn't have a corresponding
         * task count - it's just a boolean flag directly encoded in
@@ -847,7 +859,6 @@ static void psi_group_change(struct psi_group *group, int cpu,
 
                groupc->state_mask = state_mask;
 
-               write_seqcount_end(&groupc->seq);
                return;
        }
 
@@ -868,8 +879,6 @@ static void psi_group_change(struct psi_group *group, int cpu,
 
        groupc->state_mask = state_mask;
 
-       write_seqcount_end(&groupc->seq);
-
        if (state_mask & group->rtpoll_states)
                psi_schedule_rtpoll_work(group, 1, false);
 
@@ -904,24 +913,29 @@ static void psi_flags_change(struct task_struct *task, int clear, int set)
 void psi_task_change(struct task_struct *task, int clear, int set)
 {
        int cpu = task_cpu(task);
-       struct psi_group *group;
+       u64 now;
 
        if (!task->pid)
                return;
 
        psi_flags_change(task, clear, set);
 
-       group = task_psi_group(task);
-       do {
-               psi_group_change(group, cpu, clear, set, true);
-       } while ((group = group->parent));
+       psi_write_begin(cpu);
+       now = cpu_clock(cpu);
+       for_each_group(group, task_psi_group(task))
+               psi_group_change(group, cpu, clear, set, now, true);
+       psi_write_end(cpu);
 }
 
 void psi_task_switch(struct task_struct *prev, struct task_struct *next,
                     bool sleep)
 {
-       struct psi_group *group, *common = NULL;
+       struct psi_group *common = NULL;
        int cpu = task_cpu(prev);
+       u64 now;
+
+       psi_write_begin(cpu);
+       now = cpu_clock(cpu);
 
        if (next->pid) {
                psi_flags_change(next, 0, TSK_ONCPU);
@@ -930,16 +944,15 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
                 * ancestors with @prev, those will already have @prev's
                 * TSK_ONCPU bit set, and we can stop the iteration there.
                 */
-               group = task_psi_group(next);
-               do {
-                       if (per_cpu_ptr(group->pcpu, cpu)->state_mask &
-                           PSI_ONCPU) {
+               for_each_group(group, task_psi_group(next)) {
+                       struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
+
+                       if (groupc->state_mask & PSI_ONCPU) {
                                common = group;
                                break;
                        }
-
-                       psi_group_change(group, cpu, 0, TSK_ONCPU, true);
-               } while ((group = group->parent));
+                       psi_group_change(group, cpu, 0, TSK_ONCPU, now, true);
+               }
        }
 
        if (prev->pid) {
@@ -972,12 +985,11 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
 
                psi_flags_change(prev, clear, set);
 
-               group = task_psi_group(prev);
-               do {
+               for_each_group(group, task_psi_group(prev)) {
                        if (group == common)
                                break;
-                       psi_group_change(group, cpu, clear, set, wake_clock);
-               } while ((group = group->parent));
+                       psi_group_change(group, cpu, clear, set, now, wake_clock);
+               }
 
                /*
                 * TSK_ONCPU is handled up to the common ancestor. If there are
@@ -987,20 +999,21 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
                 */
                if ((prev->psi_flags ^ next->psi_flags) & ~TSK_ONCPU) {
                        clear &= ~TSK_ONCPU;
-                       for (; group; group = group->parent)
-                               psi_group_change(group, cpu, clear, set, wake_clock);
+                       for_each_group(group, common)
+                               psi_group_change(group, cpu, clear, set, now, wake_clock);
                }
        }
+       psi_write_end(cpu);
 }
 
 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
 void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_struct *prev)
 {
        int cpu = task_cpu(curr);
-       struct psi_group *group;
        struct psi_group_cpu *groupc;
        s64 delta;
        u64 irq;
+       u64 now;
 
        if (static_branch_likely(&psi_disabled) || !irqtime_enabled())
                return;
@@ -1009,8 +1022,7 @@ void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_st
                return;
 
        lockdep_assert_rq_held(rq);
-       group = task_psi_group(curr);
-       if (prev && task_psi_group(prev) == group)
+       if (prev && task_psi_group(prev) == task_psi_group(curr))
                return;
 
        irq = irq_time_read(cpu);
@@ -1019,25 +1031,22 @@ void psi_account_irqtime(struct rq *rq, struct task_struct *curr, struct task_st
                return;
        rq->psi_irq_time = irq;
 
-       do {
-               u64 now;
+       psi_write_begin(cpu);
+       now = cpu_clock(cpu);
 
+       for_each_group(group, task_psi_group(curr)) {
                if (!group->enabled)
                        continue;
 
                groupc = per_cpu_ptr(group->pcpu, cpu);
 
-               write_seqcount_begin(&groupc->seq);
-               now = cpu_clock(cpu);
-
                record_times(groupc, now);
                groupc->times[PSI_IRQ_FULL] += delta;
 
-               write_seqcount_end(&groupc->seq);
-
                if (group->rtpoll_states & (1 << PSI_IRQ_FULL))
                        psi_schedule_rtpoll_work(group, 1, false);
-       } while ((group = group->parent));
+       }
+       psi_write_end(cpu);
 }
 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
 
@@ -1225,12 +1234,14 @@ void psi_cgroup_restart(struct psi_group *group)
                return;
 
        for_each_possible_cpu(cpu) {
-               struct rq *rq = cpu_rq(cpu);
-               struct rq_flags rf;
+               u64 now;
 
-               rq_lock_irq(rq, &rf);
-               psi_group_change(group, cpu, 0, 0, true);
-               rq_unlock_irq(rq, &rf);
+               guard(rq_lock_irq)(cpu_rq(cpu));
+
+               psi_write_begin(cpu);
+               now = cpu_clock(cpu);
+               psi_group_change(group, cpu, 0, 0, now, true);
+               psi_write_end(cpu);
        }
 }
 #endif /* CONFIG_CGROUPS */