sched/core: Add rq->lock wrappers
authorPeter Zijlstra <peterz@infradead.org>
Tue, 4 Oct 2016 14:04:35 +0000 (16:04 +0200)
committerIngo Molnar <mingo@kernel.org>
Thu, 16 Mar 2017 08:46:22 +0000 (09:46 +0100)
The missing update_rq_clock() check can work with partial rq->lock
wrappery, since a missing wrapper can cause the warning to not be
emitted when it should have, but cannot cause the warning to trigger
when it should not have.

The duplicate update_rq_clock() check however can cause false warnings
to trigger. Therefore add more comprehensive rq->lock wrappery.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/sched/core.c
kernel/sched/fair.c
kernel/sched/sched.h

index 1bd15d0d03075eb4a5fd2535f74347d177cb768d..c5a514b1668db8ea7cef3112d141d775fc6aca2b 100644 (file)
@@ -85,21 +85,6 @@ int sysctl_sched_rt_runtime = 950000;
 /* CPUs with isolated domains */
 cpumask_var_t cpu_isolated_map;
 
-/*
- * this_rq_lock - lock this runqueue and disable interrupts.
- */
-static struct rq *this_rq_lock(void)
-       __acquires(rq->lock)
-{
-       struct rq *rq;
-
-       local_irq_disable();
-       rq = this_rq();
-       raw_spin_lock(&rq->lock);
-
-       return rq;
-}
-
 /*
  * __task_rq_lock - lock the rq @p resides on.
  */
@@ -264,13 +249,14 @@ static void hrtick_clear(struct rq *rq)
 static enum hrtimer_restart hrtick(struct hrtimer *timer)
 {
        struct rq *rq = container_of(timer, struct rq, hrtick_timer);
+       struct rq_flags rf;
 
        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 
-       raw_spin_lock(&rq->lock);
+       rq_lock(rq, &rf);
        update_rq_clock(rq);
        rq->curr->sched_class->task_tick(rq, rq->curr, 1);
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, &rf);
 
        return HRTIMER_NORESTART;
 }
@@ -290,11 +276,12 @@ static void __hrtick_restart(struct rq *rq)
 static void __hrtick_start(void *arg)
 {
        struct rq *rq = arg;
+       struct rq_flags rf;
 
-       raw_spin_lock(&rq->lock);
+       rq_lock(rq, &rf);
        __hrtick_restart(rq);
        rq->hrtick_csd_pending = 0;
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, &rf);
 }
 
 /*
@@ -949,18 +936,19 @@ void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
  *
  * Returns (locked) new rq. Old rq's lock is released.
  */
-static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
+static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
+                                  struct task_struct *p, int new_cpu)
 {
        lockdep_assert_held(&rq->lock);
 
        p->on_rq = TASK_ON_RQ_MIGRATING;
        dequeue_task(rq, p, 0);
        set_task_cpu(p, new_cpu);
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, rf);
 
        rq = cpu_rq(new_cpu);
 
-       raw_spin_lock(&rq->lock);
+       rq_lock(rq, rf);
        BUG_ON(task_cpu(p) != new_cpu);
        enqueue_task(rq, p, 0);
        p->on_rq = TASK_ON_RQ_QUEUED;
@@ -983,7 +971,8 @@ struct migration_arg {
  * So we race with normal scheduler movements, but that's OK, as long
  * as the task is no longer on this CPU.
  */
-static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
+static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
+                                struct task_struct *p, int dest_cpu)
 {
        if (unlikely(!cpu_active(dest_cpu)))
                return rq;
@@ -992,7 +981,7 @@ static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_
        if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
                return rq;
 
-       rq = move_queued_task(rq, p, dest_cpu);
+       rq = move_queued_task(rq, rf, p, dest_cpu);
 
        return rq;
 }
@@ -1007,6 +996,7 @@ static int migration_cpu_stop(void *data)
        struct migration_arg *arg = data;
        struct task_struct *p = arg->task;
        struct rq *rq = this_rq();
+       struct rq_flags rf;
 
        /*
         * The original target CPU might have gone down and we might
@@ -1021,7 +1011,7 @@ static int migration_cpu_stop(void *data)
        sched_ttwu_pending();
 
        raw_spin_lock(&p->pi_lock);
-       raw_spin_lock(&rq->lock);
+       rq_lock(rq, &rf);
        /*
         * If task_rq(p) != rq, it cannot be migrated here, because we're
         * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
@@ -1029,11 +1019,11 @@ static int migration_cpu_stop(void *data)
         */
        if (task_rq(p) == rq) {
                if (task_on_rq_queued(p))
-                       rq = __migrate_task(rq, p, arg->dest_cpu);
+                       rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
                else
                        p->wake_cpu = arg->dest_cpu;
        }
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, &rf);
        raw_spin_unlock(&p->pi_lock);
 
        local_irq_enable();
@@ -1153,9 +1143,7 @@ static int __set_cpus_allowed_ptr(struct task_struct *p,
                 * OK, since we're going to drop the lock immediately
                 * afterwards anyway.
                 */
-               rq_unpin_lock(rq, &rf);
-               rq = move_queued_task(rq, p, dest_cpu);
-               rq_repin_lock(rq, &rf);
+               rq = move_queued_task(rq, &rf, p, dest_cpu);
        }
 out:
        task_rq_unlock(rq, p, &rf);
@@ -1220,16 +1208,24 @@ static void __migrate_swap_task(struct task_struct *p, int cpu)
 {
        if (task_on_rq_queued(p)) {
                struct rq *src_rq, *dst_rq;
+               struct rq_flags srf, drf;
 
                src_rq = task_rq(p);
                dst_rq = cpu_rq(cpu);
 
+               rq_pin_lock(src_rq, &srf);
+               rq_pin_lock(dst_rq, &drf);
+
                p->on_rq = TASK_ON_RQ_MIGRATING;
                deactivate_task(src_rq, p, 0);
                set_task_cpu(p, cpu);
                activate_task(dst_rq, p, 0);
                p->on_rq = TASK_ON_RQ_QUEUED;
                check_preempt_curr(dst_rq, p, 0);
+
+               rq_unpin_lock(dst_rq, &drf);
+               rq_unpin_lock(src_rq, &srf);
+
        } else {
                /*
                 * Task isn't running anymore; make it appear like we migrated
@@ -1729,14 +1725,12 @@ void sched_ttwu_pending(void)
        struct rq *rq = this_rq();
        struct llist_node *llist = llist_del_all(&rq->wake_list);
        struct task_struct *p;
-       unsigned long flags;
        struct rq_flags rf;
 
        if (!llist)
                return;
 
-       raw_spin_lock_irqsave(&rq->lock, flags);
-       rq_pin_lock(rq, &rf);
+       rq_lock_irqsave(rq, &rf);
 
        while (llist) {
                int wake_flags = 0;
@@ -1750,8 +1744,7 @@ void sched_ttwu_pending(void)
                ttwu_do_activate(rq, p, wake_flags, &rf);
        }
 
-       rq_unpin_lock(rq, &rf);
-       raw_spin_unlock_irqrestore(&rq->lock, flags);
+       rq_unlock_irqrestore(rq, &rf);
 }
 
 void scheduler_ipi(void)
@@ -1809,7 +1802,7 @@ static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
 void wake_up_if_idle(int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
-       unsigned long flags;
+       struct rq_flags rf;
 
        rcu_read_lock();
 
@@ -1819,11 +1812,11 @@ void wake_up_if_idle(int cpu)
        if (set_nr_if_polling(rq->idle)) {
                trace_sched_wake_idle_without_ipi(cpu);
        } else {
-               raw_spin_lock_irqsave(&rq->lock, flags);
+               rq_lock_irqsave(rq, &rf);
                if (is_idle_task(rq->curr))
                        smp_send_reschedule(cpu);
                /* Else CPU is not idle, do nothing here: */
-               raw_spin_unlock_irqrestore(&rq->lock, flags);
+               rq_unlock_irqrestore(rq, &rf);
        }
 
 out:
@@ -1849,11 +1842,9 @@ static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
        }
 #endif
 
-       raw_spin_lock(&rq->lock);
-       rq_pin_lock(rq, &rf);
+       rq_lock(rq, &rf);
        ttwu_do_activate(rq, p, wake_flags, &rf);
-       rq_unpin_lock(rq, &rf);
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, &rf);
 }
 
 /*
@@ -2100,11 +2091,9 @@ static void try_to_wake_up_local(struct task_struct *p, struct rq_flags *rf)
                 * disabled avoiding further scheduler activity on it and we've
                 * not yet picked a replacement task.
                 */
-               rq_unpin_lock(rq, rf);
-               raw_spin_unlock(&rq->lock);
+               rq_unlock(rq, rf);
                raw_spin_lock(&p->pi_lock);
-               raw_spin_lock(&rq->lock);
-               rq_repin_lock(rq, rf);
+               rq_relock(rq, rf);
        }
 
        if (!(p->state & TASK_NORMAL))
@@ -2778,9 +2767,9 @@ static void __balance_callback(struct rq *rq)
 {
        struct callback_head *head, *next;
        void (*func)(struct rq *rq);
-       unsigned long flags;
+       struct rq_flags rf;
 
-       raw_spin_lock_irqsave(&rq->lock, flags);
+       rq_lock_irqsave(rq, &rf);
        head = rq->balance_callback;
        rq->balance_callback = NULL;
        while (head) {
@@ -2791,7 +2780,7 @@ static void __balance_callback(struct rq *rq)
 
                func(rq);
        }
-       raw_spin_unlock_irqrestore(&rq->lock, flags);
+       rq_unlock_irqrestore(rq, &rf);
 }
 
 static inline void balance_callback(struct rq *rq)
@@ -3096,15 +3085,18 @@ void scheduler_tick(void)
        int cpu = smp_processor_id();
        struct rq *rq = cpu_rq(cpu);
        struct task_struct *curr = rq->curr;
+       struct rq_flags rf;
 
        sched_clock_tick();
 
-       raw_spin_lock(&rq->lock);
+       rq_lock(rq, &rf);
+
        update_rq_clock(rq);
        curr->sched_class->task_tick(rq, curr, 0);
        cpu_load_update_active(rq);
        calc_global_load_tick(rq);
-       raw_spin_unlock(&rq->lock);
+
+       rq_unlock(rq, &rf);
 
        perf_event_task_tick();
 
@@ -3389,8 +3381,7 @@ static void __sched notrace __schedule(bool preempt)
         * done by the caller to avoid the race with signal_wake_up().
         */
        smp_mb__before_spinlock();
-       raw_spin_lock(&rq->lock);
-       rq_pin_lock(rq, &rf);
+       rq_lock(rq, &rf);
 
        /* Promote REQ to ACT */
        rq->clock_update_flags <<= 1;
@@ -3442,8 +3433,7 @@ static void __sched notrace __schedule(bool preempt)
                rq = context_switch(rq, prev, next, &rf);
        } else {
                rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
-               rq_unpin_lock(rq, &rf);
-               raw_spin_unlock_irq(&rq->lock);
+               rq_unlock_irq(rq, &rf);
        }
 
        balance_callback(rq);
@@ -4926,7 +4916,12 @@ SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
  */
 SYSCALL_DEFINE0(sched_yield)
 {
-       struct rq *rq = this_rq_lock();
+       struct rq_flags rf;
+       struct rq *rq;
+
+       local_irq_disable();
+       rq = this_rq();
+       rq_lock(rq, &rf);
 
        schedstat_inc(rq->yld_count);
        current->sched_class->yield_task(rq);
@@ -4935,9 +4930,8 @@ SYSCALL_DEFINE0(sched_yield)
         * Since we are going to call schedule() anyway, there's
         * no need to preempt or enable interrupts:
         */
-       __release(rq->lock);
-       spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
-       do_raw_spin_unlock(&rq->lock);
+       preempt_disable();
+       rq_unlock(rq, &rf);
        sched_preempt_enable_no_resched();
 
        schedule();
@@ -5582,11 +5576,11 @@ static struct task_struct fake_task = {
  * there's no concurrency possible, we hold the required locks anyway
  * because of lock validation efforts.
  */
-static void migrate_tasks(struct rq *dead_rq)
+static void migrate_tasks(struct rq *dead_rq, struct rq_flags *rf)
 {
        struct rq *rq = dead_rq;
        struct task_struct *next, *stop = rq->stop;
-       struct rq_flags rf;
+       struct rq_flags orf = *rf;
        int dest_cpu;
 
        /*
@@ -5605,9 +5599,7 @@ static void migrate_tasks(struct rq *dead_rq)
         * class method both need to have an up-to-date
         * value of rq->clock[_task]
         */
-       rq_pin_lock(rq, &rf);
        update_rq_clock(rq);
-       rq_unpin_lock(rq, &rf);
 
        for (;;) {
                /*
@@ -5620,8 +5612,7 @@ static void migrate_tasks(struct rq *dead_rq)
                /*
                 * pick_next_task() assumes pinned rq->lock:
                 */
-               rq_repin_lock(rq, &rf);
-               next = pick_next_task(rq, &fake_task, &rf);
+               next = pick_next_task(rq, &fake_task, rf);
                BUG_ON(!next);
                next->sched_class->put_prev_task(rq, next);
 
@@ -5634,10 +5625,9 @@ static void migrate_tasks(struct rq *dead_rq)
                 * because !cpu_active at this point, which means load-balance
                 * will not interfere. Also, stop-machine.
                 */
-               rq_unpin_lock(rq, &rf);
-               raw_spin_unlock(&rq->lock);
+               rq_unlock(rq, rf);
                raw_spin_lock(&next->pi_lock);
-               raw_spin_lock(&rq->lock);
+               rq_relock(rq, rf);
 
                /*
                 * Since we're inside stop-machine, _nothing_ should have
@@ -5651,12 +5641,12 @@ static void migrate_tasks(struct rq *dead_rq)
 
                /* Find suitable destination for @next, with force if needed. */
                dest_cpu = select_fallback_rq(dead_rq->cpu, next);
-
-               rq = __migrate_task(rq, next, dest_cpu);
+               rq = __migrate_task(rq, rf, next, dest_cpu);
                if (rq != dead_rq) {
-                       raw_spin_unlock(&rq->lock);
+                       rq_unlock(rq, rf);
                        rq = dead_rq;
-                       raw_spin_lock(&rq->lock);
+                       *rf = orf;
+                       rq_relock(rq, rf);
                }
                raw_spin_unlock(&next->pi_lock);
        }
@@ -5769,7 +5759,7 @@ static int cpuset_cpu_inactive(unsigned int cpu)
 int sched_cpu_activate(unsigned int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
-       unsigned long flags;
+       struct rq_flags rf;
 
        set_cpu_active(cpu, true);
 
@@ -5787,12 +5777,12 @@ int sched_cpu_activate(unsigned int cpu)
         * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
         *    domains.
         */
-       raw_spin_lock_irqsave(&rq->lock, flags);
+       rq_lock_irqsave(rq, &rf);
        if (rq->rd) {
                BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
                set_rq_online(rq);
        }
-       raw_spin_unlock_irqrestore(&rq->lock, flags);
+       rq_unlock_irqrestore(rq, &rf);
 
        update_max_interval();
 
@@ -5850,18 +5840,20 @@ int sched_cpu_starting(unsigned int cpu)
 int sched_cpu_dying(unsigned int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
-       unsigned long flags;
+       struct rq_flags rf;
 
        /* Handle pending wakeups and then migrate everything off */
        sched_ttwu_pending();
-       raw_spin_lock_irqsave(&rq->lock, flags);
+
+       rq_lock_irqsave(rq, &rf);
        if (rq->rd) {
                BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
                set_rq_offline(rq);
        }
-       migrate_tasks(rq);
+       migrate_tasks(rq, &rf);
        BUG_ON(rq->nr_running != 1);
-       raw_spin_unlock_irqrestore(&rq->lock, flags);
+       rq_unlock_irqrestore(rq, &rf);
+
        calc_load_migrate(rq);
        update_max_interval();
        nohz_balance_exit_idle(cpu);
@@ -7011,14 +7003,15 @@ static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
        for_each_online_cpu(i) {
                struct cfs_rq *cfs_rq = tg->cfs_rq[i];
                struct rq *rq = cfs_rq->rq;
+               struct rq_flags rf;
 
-               raw_spin_lock_irq(&rq->lock);
+               rq_lock_irq(rq, &rf);
                cfs_rq->runtime_enabled = runtime_enabled;
                cfs_rq->runtime_remaining = 0;
 
                if (cfs_rq->throttled)
                        unthrottle_cfs_rq(cfs_rq);
-               raw_spin_unlock_irq(&rq->lock);
+               rq_unlock_irq(rq, &rf);
        }
        if (runtime_was_enabled && !runtime_enabled)
                cfs_bandwidth_usage_dec();
index dea138964b9107b3e22542a8b80f5cf1d43c1dee..72b081b9a24967d750e76c112022746e2c2feb34 100644 (file)
@@ -4271,8 +4271,9 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
        list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
                                throttled_list) {
                struct rq *rq = rq_of(cfs_rq);
+               struct rq_flags rf;
 
-               raw_spin_lock(&rq->lock);
+               rq_lock(rq, &rf);
                if (!cfs_rq_throttled(cfs_rq))
                        goto next;
 
@@ -4289,7 +4290,7 @@ static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
                        unthrottle_cfs_rq(cfs_rq);
 
 next:
-               raw_spin_unlock(&rq->lock);
+               rq_unlock(rq, &rf);
 
                if (!remaining)
                        break;
@@ -5097,15 +5098,16 @@ void cpu_load_update_nohz_stop(void)
        unsigned long curr_jiffies = READ_ONCE(jiffies);
        struct rq *this_rq = this_rq();
        unsigned long load;
+       struct rq_flags rf;
 
        if (curr_jiffies == this_rq->last_load_update_tick)
                return;
 
        load = weighted_cpuload(cpu_of(this_rq));
-       raw_spin_lock(&this_rq->lock);
+       rq_lock(this_rq, &rf);
        update_rq_clock(this_rq);
        cpu_load_update_nohz(this_rq, curr_jiffies, load);
-       raw_spin_unlock(&this_rq->lock);
+       rq_unlock(this_rq, &rf);
 }
 #else /* !CONFIG_NO_HZ_COMMON */
 static inline void cpu_load_update_nohz(struct rq *this_rq,
@@ -6913,9 +6915,11 @@ static void attach_task(struct rq *rq, struct task_struct *p)
  */
 static void attach_one_task(struct rq *rq, struct task_struct *p)
 {
-       raw_spin_lock(&rq->lock);
+       struct rq_flags rf;
+
+       rq_lock(rq, &rf);
        attach_task(rq, p);
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, &rf);
 }
 
 /*
@@ -6926,8 +6930,9 @@ static void attach_tasks(struct lb_env *env)
 {
        struct list_head *tasks = &env->tasks;
        struct task_struct *p;
+       struct rq_flags rf;
 
-       raw_spin_lock(&env->dst_rq->lock);
+       rq_lock(env->dst_rq, &rf);
 
        while (!list_empty(tasks)) {
                p = list_first_entry(tasks, struct task_struct, se.group_node);
@@ -6936,7 +6941,7 @@ static void attach_tasks(struct lb_env *env)
                attach_task(env->dst_rq, p);
        }
 
-       raw_spin_unlock(&env->dst_rq->lock);
+       rq_unlock(env->dst_rq, &rf);
 }
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
@@ -6944,9 +6949,9 @@ static void update_blocked_averages(int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
        struct cfs_rq *cfs_rq;
-       unsigned long flags;
+       struct rq_flags rf;
 
-       raw_spin_lock_irqsave(&rq->lock, flags);
+       rq_lock_irqsave(rq, &rf);
        update_rq_clock(rq);
 
        /*
@@ -6965,7 +6970,7 @@ static void update_blocked_averages(int cpu)
                if (cfs_rq->tg->se[cpu])
                        update_load_avg(cfs_rq->tg->se[cpu], 0);
        }
-       raw_spin_unlock_irqrestore(&rq->lock, flags);
+       rq_unlock_irqrestore(rq, &rf);
 }
 
 /*
@@ -7019,12 +7024,12 @@ static inline void update_blocked_averages(int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
        struct cfs_rq *cfs_rq = &rq->cfs;
-       unsigned long flags;
+       struct rq_flags rf;
 
-       raw_spin_lock_irqsave(&rq->lock, flags);
+       rq_lock_irqsave(rq, &rf);
        update_rq_clock(rq);
        update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
-       raw_spin_unlock_irqrestore(&rq->lock, flags);
+       rq_unlock_irqrestore(rq, &rf);
 }
 
 static unsigned long task_h_load(struct task_struct *p)
@@ -8042,7 +8047,7 @@ static int load_balance(int this_cpu, struct rq *this_rq,
        struct sched_domain *sd_parent = sd->parent;
        struct sched_group *group;
        struct rq *busiest;
-       unsigned long flags;
+       struct rq_flags rf;
        struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
 
        struct lb_env env = {
@@ -8105,7 +8110,7 @@ redo:
                env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
 
 more_balance:
-               raw_spin_lock_irqsave(&busiest->lock, flags);
+               rq_lock_irqsave(busiest, &rf);
                update_rq_clock(busiest);
 
                /*
@@ -8122,14 +8127,14 @@ more_balance:
                 * See task_rq_lock() family for the details.
                 */
 
-               raw_spin_unlock(&busiest->lock);
+               rq_unlock(busiest, &rf);
 
                if (cur_ld_moved) {
                        attach_tasks(&env);
                        ld_moved += cur_ld_moved;
                }
 
-               local_irq_restore(flags);
+               local_irq_restore(rf.flags);
 
                if (env.flags & LBF_NEED_BREAK) {
                        env.flags &= ~LBF_NEED_BREAK;
@@ -8207,6 +8212,8 @@ more_balance:
                        sd->nr_balance_failed++;
 
                if (need_active_balance(&env)) {
+                       unsigned long flags;
+
                        raw_spin_lock_irqsave(&busiest->lock, flags);
 
                        /* don't kick the active_load_balance_cpu_stop,
@@ -8444,8 +8451,9 @@ static int active_load_balance_cpu_stop(void *data)
        struct rq *target_rq = cpu_rq(target_cpu);
        struct sched_domain *sd;
        struct task_struct *p = NULL;
+       struct rq_flags rf;
 
-       raw_spin_lock_irq(&busiest_rq->lock);
+       rq_lock_irq(busiest_rq, &rf);
 
        /* make sure the requested cpu hasn't gone down in the meantime */
        if (unlikely(busiest_cpu != smp_processor_id() ||
@@ -8496,7 +8504,7 @@ static int active_load_balance_cpu_stop(void *data)
        rcu_read_unlock();
 out_unlock:
        busiest_rq->active_balance = 0;
-       raw_spin_unlock(&busiest_rq->lock);
+       rq_unlock(busiest_rq, &rf);
 
        if (p)
                attach_one_task(target_rq, p);
@@ -8794,10 +8802,13 @@ static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
                 * do the balance.
                 */
                if (time_after_eq(jiffies, rq->next_balance)) {
-                       raw_spin_lock_irq(&rq->lock);
+                       struct rq_flags rf;
+
+                       rq_lock_irq(rq, &rf);
                        update_rq_clock(rq);
                        cpu_load_update_idle(rq);
-                       raw_spin_unlock_irq(&rq->lock);
+                       rq_unlock_irq(rq, &rf);
+
                        rebalance_domains(rq, CPU_IDLE);
                }
 
@@ -8988,8 +8999,9 @@ static void task_fork_fair(struct task_struct *p)
        struct cfs_rq *cfs_rq;
        struct sched_entity *se = &p->se, *curr;
        struct rq *rq = this_rq();
+       struct rq_flags rf;
 
-       raw_spin_lock(&rq->lock);
+       rq_lock(rq, &rf);
        update_rq_clock(rq);
 
        cfs_rq = task_cfs_rq(current);
@@ -9010,7 +9022,7 @@ static void task_fork_fair(struct task_struct *p)
        }
 
        se->vruntime -= cfs_rq->min_vruntime;
-       raw_spin_unlock(&rq->lock);
+       rq_unlock(rq, &rf);
 }
 
 /*
@@ -9372,7 +9384,6 @@ static DEFINE_MUTEX(shares_mutex);
 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
 {
        int i;
-       unsigned long flags;
 
        /*
         * We can't change the weight of the root cgroup.
@@ -9389,19 +9400,17 @@ int sched_group_set_shares(struct task_group *tg, unsigned long shares)
        tg->shares = shares;
        for_each_possible_cpu(i) {
                struct rq *rq = cpu_rq(i);
-               struct sched_entity *se;
+               struct sched_entity *se = tg->se[i];
+               struct rq_flags rf;
 
-               se = tg->se[i];
                /* Propagate contribution to hierarchy */
-               raw_spin_lock_irqsave(&rq->lock, flags);
-
-               /* Possible calls to update_curr() need rq clock */
+               rq_lock_irqsave(rq, &rf);
                update_rq_clock(rq);
                for_each_sched_entity(se) {
                        update_load_avg(se, UPDATE_TG);
                        update_cfs_shares(se);
                }
-               raw_spin_unlock_irqrestore(&rq->lock, flags);
+               rq_unlock_irqrestore(rq, &rf);
        }
 
 done:
index 5cbf92214ad89287d111ab8300e5b55923d83ffe..7d4f69329634f2cc8c01fbf9d2cd3960d79807ff 100644 (file)
@@ -1624,6 +1624,7 @@ static inline void sched_avg_update(struct rq *rq) { }
 
 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
        __acquires(rq->lock);
+
 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
        __acquires(p->pi_lock)
        __acquires(rq->lock);
@@ -1645,6 +1646,62 @@ task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
        raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
 }
 
+static inline void
+rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
+       __acquires(rq->lock)
+{
+       raw_spin_lock_irqsave(&rq->lock, rf->flags);
+       rq_pin_lock(rq, rf);
+}
+
+static inline void
+rq_lock_irq(struct rq *rq, struct rq_flags *rf)
+       __acquires(rq->lock)
+{
+       raw_spin_lock_irq(&rq->lock);
+       rq_pin_lock(rq, rf);
+}
+
+static inline void
+rq_lock(struct rq *rq, struct rq_flags *rf)
+       __acquires(rq->lock)
+{
+       raw_spin_lock(&rq->lock);
+       rq_pin_lock(rq, rf);
+}
+
+static inline void
+rq_relock(struct rq *rq, struct rq_flags *rf)
+       __acquires(rq->lock)
+{
+       raw_spin_lock(&rq->lock);
+       rq_repin_lock(rq, rf);
+}
+
+static inline void
+rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
+       __releases(rq->lock)
+{
+       rq_unpin_lock(rq, rf);
+       raw_spin_unlock_irqrestore(&rq->lock, rf->flags);
+}
+
+static inline void
+rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
+       __releases(rq->lock)
+{
+       rq_unpin_lock(rq, rf);
+       raw_spin_unlock_irq(&rq->lock);
+}
+
+static inline void
+rq_unlock(struct rq *rq, struct rq_flags *rf)
+       __releases(rq->lock)
+{
+       rq_unpin_lock(rq, rf);
+       raw_spin_unlock(&rq->lock);
+}
+
 #ifdef CONFIG_SMP
 #ifdef CONFIG_PREEMPT