linux进程调度机制剖析（基于3.16-rc4）

1 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);

 1 struct sched_entity {
 2     struct load_weight    load;        /* for load-balancing */
 3     struct rb_node        run_node;
 4     struct list_head    group_node;
 5     unsigned int        on_rq;
 6 
 7     u64            exec_start;
 8     u64            sum_exec_runtime;
 9     u64            vruntime;
10     u64            prev_sum_exec_runtime;
11 
12     u64            nr_migrations;
13 
14 #ifdef CONFIG_SCHEDSTATS
15     struct sched_statistics statistics;
16 #endif
17 
18 #ifdef CONFIG_FAIR_GROUP_SCHED
19     int            depth;
20     struct sched_entity    *parent;
21     /* rq on which this entity is (to be) queued: */
22     struct cfs_rq        *cfs_rq;
23     /* rq "owned" by this entity/group: */
24     struct cfs_rq        *my_q;
25 #endif
26 
27 #ifdef CONFIG_SMP
28     /* Per-entity load-tracking */
29     struct sched_avg    avg;
30 #endif
31 };

 1 struct sched_rt_entity {
 2     struct list_head run_list;
 3     unsigned long timeout;
 4     unsigned long watchdog_stamp;
 5     unsigned int time_slice;
 6 
 7     struct sched_rt_entity *back;
 8 #ifdef CONFIG_RT_GROUP_SCHED
 9     struct sched_rt_entity    *parent;
10     /* rq on which this entity is (to be) queued: */
11     struct rt_rq        *rt_rq;
12     /* rq "owned" by this entity/group: */
13     struct rt_rq        *my_q;
14 #endif
15 };

 1 struct sched_class {
 2     const struct sched_class *next;
 3 
 4     void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
 5     void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
 6     void (*yield_task) (struct rq *rq);
 7     bool (*yield_to_task) (struct rq *rq, struct task_struct *p, bool preempt);
 8 
 9     void (*check_preempt_curr) (struct rq *rq, struct task_struct *p, int flags);
10 
11     /*
12      * It is the responsibility of the pick_next_task() method that will
13      * return the next task to call put_prev_task() on the @prev task or
14      * something equivalent.
15      *
16      * May return RETRY_TASK when it finds a higher prio class has runnable
17      * tasks.
18      */
19     struct task_struct * (*pick_next_task) (struct rq *rq,
20                         struct task_struct *prev);
21     void (*put_prev_task) (struct rq *rq, struct task_struct *p);
22 
23 #ifdef CONFIG_SMP
24     int  (*select_task_rq)(struct task_struct *p, int task_cpu, int sd_flag, int flags);
25     void (*migrate_task_rq)(struct task_struct *p, int next_cpu);
26 
27     void (*post_schedule) (struct rq *this_rq);
28     void (*task_waking) (struct task_struct *task);
29     void (*task_woken) (struct rq *this_rq, struct task_struct *task);
30 
31     void (*set_cpus_allowed)(struct task_struct *p,
32                  const struct cpumask *newmask);
33 
34     void (*rq_online)(struct rq *rq);
35     void (*rq_offline)(struct rq *rq);
36 #endif
37 
38     void (*set_curr_task) (struct rq *rq);
39     void (*task_tick) (struct rq *rq, struct task_struct *p, int queued);
40     void (*task_fork) (struct task_struct *p);
41     void (*task_dead) (struct task_struct *p);
42 
43     void (*switched_from) (struct rq *this_rq, struct task_struct *task);
44     void (*switched_to) (struct rq *this_rq, struct task_struct *task);
45     void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
46                  int oldprio);
47 
48     unsigned int (*get_rr_interval) (struct rq *rq,
49                      struct task_struct *task);
50 
51 #ifdef CONFIG_FAIR_GROUP_SCHED
52     void (*task_move_group) (struct task_struct *p, int on_rq);
53 #endif
54 };

1 const struct sched_class fair_sched_class;

 1 const struct sched_class fair_sched_class = {
 2     .next            = &idle_sched_class,
 3     .enqueue_task        = enqueue_task_fair,
 4     .dequeue_task        = dequeue_task_fair,
 5     .yield_task        = yield_task_fair,
 6     .yield_to_task        = yield_to_task_fair,
 7 
 8     .check_preempt_curr    = check_preempt_wakeup,
 9 
10     .pick_next_task        = pick_next_task_fair,
11     .put_prev_task        = put_prev_task_fair,
12 
13 #ifdef CONFIG_SMP
14     .select_task_rq        = select_task_rq_fair,
15     .migrate_task_rq    = migrate_task_rq_fair,
16 
17     .rq_online        = rq_online_fair,
18     .rq_offline        = rq_offline_fair,
19 
20     .task_waking        = task_waking_fair,
21 #endif
22 
23     .set_curr_task          = set_curr_task_fair,
24     .task_tick        = task_tick_fair,
25     .task_fork        = task_fork_fair,
26 
27     .prio_changed        = prio_changed_fair,
28     .switched_from        = switched_from_fair,
29     .switched_to        = switched_to_fair,
30 
31     .get_rr_interval    = get_rr_interval_fair,
32 
33 #ifdef CONFIG_FAIR_GROUP_SCHED
34     .task_move_group    = task_move_group_fair,
35 #endif
36 };

 1 struct task_struct {
 2     volatile long state;    /* -1 unrunnable, 0 runnable, >0 stopped */
 3     .....
 4     int on_rq;
 5 
 6     int prio, static_prio, normal_prio;
 7     unsigned int rt_priority;
 8     const struct sched_class *sched_class;
 9     struct sched_entity se;
10     struct sched_rt_entity rt;
11 #ifdef CONFIG_CGROUP_SCHED
12     struct task_group *sched_task_group;
13 #endif
14     struct sched_dl_entity dl;
15     .....
16     .....
17     unsigned int policy;
18     .....
19     .....
20     struct sched_info sched_info;
21     .....
22     .....
23 };

 1 void scheduler_tick(void)
 2 {
 3     int cpu = smp_processor_id();
 4     struct rq *rq = cpu_rq(cpu);
 5     struct task_struct *curr = rq->curr;
 6 
 7     sched_clock_tick();
 8 
 9     raw_spin_lock(&rq->lock);
10     update_rq_clock(rq);
11     curr->sched_class->task_tick(rq, curr, 0);
12     update_cpu_load_active(rq);
13     raw_spin_unlock(&rq->lock);
14 
15     perf_event_task_tick();
16 
17 #ifdef CONFIG_SMP
18     rq->idle_balance = idle_cpu(cpu);
19     trigger_load_balance(rq);
20 #endif
21     rq_last_tick_reset(rq);
22 }

 1 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
 2 {
 3     struct cfs_rq *cfs_rq;
 4     struct sched_entity *se = &curr->se;
 5 
 6     for_each_sched_entity(se) {
 7         cfs_rq = cfs_rq_of(se);
 8         entity_tick(cfs_rq, se, queued);
 9     }
10 
11     if (numabalancing_enabled)
12         task_tick_numa(rq, curr);
13 
14     update_rq_runnable_avg(rq, 1);
15 }

 1 static void
 2 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
 3 {
 4     /*
 5      * Update run-time statistics of the ‘current‘.
 6      */
 7     update_curr(cfs_rq);
 8 
 9     /*
10      * Ensure that runnable average is periodically updated.
11      */
12     update_entity_load_avg(curr, 1);
13     update_cfs_rq_blocked_load(cfs_rq, 1);
14     update_cfs_shares(cfs_rq);
15 
16 #ifdef CONFIG_SCHED_HRTICK
17     /*
18      * queued ticks are scheduled to match the slice, so don‘t bother
19      * validating it and just reschedule.
20      */
21     if (queued) {
22         resched_task(rq_of(cfs_rq)->curr);
23         return;
24     }
25     /*
26      * don‘t let the period tick interfere with the hrtick preemption
27      */
28     if (!sched_feat(DOUBLE_TICK) &&
29             hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
30         return;
31 #endif
32 
33     if (cfs_rq->nr_running > 1)
34         check_preempt_tick(cfs_rq, curr);
35 }

 1 static void update_curr(struct cfs_rq *cfs_rq)
 2 {
 3     struct sched_entity *curr = cfs_rq->curr;
 4     u64 now = rq_clock_task(rq_of(cfs_rq));
 5     u64 delta_exec;
 6 
 7     if (unlikely(!curr))
 8         return;
 9 
10     delta_exec = now - curr->exec_start;
11     if (unlikely((s64)delta_exec <= 0))
12         return;
13 
14     curr->exec_start = now;
15 
16     schedstat_set(curr->statistics.exec_max,
17               max(delta_exec, curr->statistics.exec_max));
18 
19     curr->sum_exec_runtime += delta_exec;
20     schedstat_add(cfs_rq, exec_clock, delta_exec);
21 
22     curr->vruntime += calc_delta_fair(delta_exec, curr);
23     update_min_vruntime(cfs_rq);
24 
25     if (entity_is_task(curr)) {
26         struct task_struct *curtask = task_of(curr);
27 
28         trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
29         cpuacct_charge(curtask, delta_exec);
30         account_group_exec_runtime(curtask, delta_exec);
31     }
32 
33     account_cfs_rq_runtime(cfs_rq, delta_exec);
34 } 

 1 static void update_min_vruntime(struct cfs_rq *cfs_rq)
 2 {
 3     u64 vruntime = cfs_rq->min_vruntime;
 4 
 5     if (cfs_rq->curr)
 6         vruntime = cfs_rq->curr->vruntime;
 7 
 8     if (cfs_rq->rb_leftmost) {
 9         struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
10                            struct sched_entity,
11                            run_node);
12 
13         if (!cfs_rq->curr)
14             vruntime = se->vruntime;
15         else
16             vruntime = min_vruntime(vruntime, se->vruntime);
17     }
18 
19     /* ensure we never gain time by being placed backwards. */
20     cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
21 #ifndef CONFIG_64BIT
22     smp_wmb();
23     cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
24 #endif
25 } 

1 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
2 {
3     if (unlikely(se->load.weight != NICE_0_LOAD))
4         delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
5 
6     return delta;
7 } 

 1 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
 2 {
 3     u64 fact = scale_load_down(weight);
 4     int shift = WMULT_SHIFT;
 5 
 6     __update_inv_weight(lw);
 7 
 8     if (unlikely(fact >> 32)) {
 9         while (fact >> 32) {
10             fact >>= 1;
11             shift--;
12         }
13     }
14 
15     /* hint to use a 32x32->64 mul */
16     fact = (u64)(u32)fact * lw->inv_weight;
17 
18     while (fact >> 32) {
19         fact >>= 1;
20         shift--;
21     }
22 
23     return mul_u64_u32_shr(delta_exec, fact, shift);
24 }

 1 static int effective_prio(struct task_struct *p)
 2 {
 3     p->normal_prio = normal_prio(p);
 4     /*
 5      * If we are RT tasks or we were boosted to RT priority,
 6      * keep the priority unchanged. Otherwise, update priority
 7      * to the normal priority:
 8      */
 9     if (!rt_prio(p->prio))
10         return p->normal_prio;
11     return p->prio;
12 }

 1 static inline int normal_prio(struct task_struct *p)
 2 {
 3     int prio;
 4 
 5     if (task_has_dl_policy(p))
 6         prio = MAX_DL_PRIO-1;
 7     else if (task_has_rt_policy(p))
 8         prio = MAX_RT_PRIO-1 - p->rt_priority;
 9     else
10         prio = __normal_prio(p);
11     return prio;
12 }

1 struct load_weight {
2     unsigned long weight;
3     u32 inv_weight;
4 };

 1 static const int prio_to_weight[40] = {
 2  /* -20 */     88761,     71755,     56483,     46273,     36291,
 3  /* -15 */     29154,     23254,     18705,     14949,     11916,
 4  /* -10 */      9548,      7620,      6100,      4904,      3906,
 5  /*  -5 */      3121,      2501,      1991,      1586,      1277,
 6  /*   0 */      1024,       820,       655,       526,       423,
 7  /*   5 */       335,       272,       215,       172,       137,
 8  /*  10 */       110,        87,        70,        56,        45,
 9  /*  15 */        36,        29,        23,        18,        15,
10 };

 1 static const u32 prio_to_wmult[40] = {
 2  /* -20 */     48388,     59856,     76040,     92818,    118348,
 3  /* -15 */    147320,    184698,    229616,    287308,    360437,
 4  /* -10 */    449829,    563644,    704093,    875809,   1099582,
 5  /*  -5 */   1376151,   1717300,   2157191,   2708050,   3363326,
 6  /*   0 */   4194304,   5237765,   6557202,   8165337,  10153587,
 7  /*   5 */  12820798,  15790321,  19976592,  24970740,  31350126,
 8  /*  10 */  39045157,  49367440,  61356676,  76695844,  95443717,
 9  /*  15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
10 };

 1 static void set_load_weight(struct task_struct *p)
 2 {
 3     int prio = p->static_prio - MAX_RT_PRIO;
 4     struct load_weight *load = &p->se.load;
 5 
 6     /*
 7      * SCHED_IDLE tasks get minimal weight:
 8      */
 9     if (p->policy == SCHED_IDLE) {
10         load->weight = scale_load(WEIGHT_IDLEPRIO);
11         load->inv_weight = WMULT_IDLEPRIO;
12         return;
13     }
14 
15     load->weight = scale_load(prio_to_weight[prio]);
16     load->inv_weight = prio_to_wmult[prio];
17 } 

1 static inline void update_load_set(struct load_weight *lw, unsigned long w)
2 {
3     lw->weight = w;
4     lw->inv_weight = 0;
5 }

1 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
2 {
3     lw->weight += inc;
4     lw->inv_weight = 0;
5 }

1 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
2 {
3     lw->weight -= dec;
4     lw->inv_weight = 0;
5 }

 1 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
 2 {
 3     u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
 4 
 5     for_each_sched_entity(se) {
 6         struct load_weight *load;
 7         struct load_weight lw;
 8 
 9         cfs_rq = cfs_rq_of(se);
10         load = &cfs_rq->load;
11 
12         if (unlikely(!se->on_rq)) {
13             lw = cfs_rq->load;
14 
15             update_load_add(&lw, se->load.weight);
16             load = &lw;
17         }
18         slice = __calc_delta(slice, se->load.weight, load);
19     }
20     return slice;
21 }

1 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
2 {
3     return calc_delta_fair(sched_slice(cfs_rq, se), se);
4 }

 1 static struct sched_entity *
 2 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
 3 {
 4     struct sched_entity *left = __pick_first_entity(cfs_rq);
 5     struct sched_entity *se;
 6 
 7     /*
 8      * If curr is set we have to see if its left of the leftmost entity
 9      * still in the tree, provided there was anything in the tree at all.
10      */
11     if (!left || (curr && entity_before(curr, left)))
12         left = curr;
13 
14     se = left; /* ideally we run the leftmost entity */
15 
16     /*
17      * Avoid running the skip buddy, if running something else can
18      * be done without getting too unfair.
19      */
20     if (cfs_rq->skip == se) {
21         struct sched_entity *second;
22 
23         if (se == curr) {
24             second = __pick_first_entity(cfs_rq);
25         } else {
26             second = __pick_next_entity(se);
27             if (!second || (curr && entity_before(curr, second)))
28                 second = curr;
29         }
30 
31         if (second && wakeup_preempt_entity(second, left) < 1)
32             se = second;
33     }
34 
35     /*
36      * Prefer last buddy, try to return the CPU to a preempted task.
37      */
38     if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
39         se = cfs_rq->last;
40 
41     /*
42      * Someone really wants this to run. If it‘s not unfair, run it.
43      */
44     if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
45         se = cfs_rq->next;
46 
47     clear_buddies(cfs_rq, se);
48 
49     return se;
50 }

 1 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
 2 {
 3     /*
 4      * If still on the runqueue then deactivate_task()
 5      * was not called and update_curr() has to be done:
 6      */
 7     if (prev->on_rq)
 8         update_curr(cfs_rq);
 9 
10     /* throttle cfs_rqs exceeding runtime */
11     check_cfs_rq_runtime(cfs_rq);
12 
13     check_spread(cfs_rq, prev);
14     if (prev->on_rq) {
15         update_stats_wait_start(cfs_rq, prev);
16         /* Put ‘current‘ back into the tree. */
17         __enqueue_entity(cfs_rq, prev);
18         /* in !on_rq case, update occurred at dequeue */
19         update_entity_load_avg(prev, 1);
20     }
21     cfs_rq->curr = NULL;
22 } 

 1 static void
 2 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 3 {
 4     /* ‘current‘ is not kept within the tree. */
 5     if (se->on_rq) {
 6         /*
 7          * Any task has to be enqueued before it get to execute on
 8          * a CPU. So account for the time it spent waiting on the
 9          * runqueue.
10          */
11         update_stats_wait_end(cfs_rq, se);
12         __dequeue_entity(cfs_rq, se);
13     }
14 
15     update_stats_curr_start(cfs_rq, se);
16     cfs_rq->curr = se;
17 #ifdef CONFIG_SCHEDSTATS
18     /*
19      * Track our maximum slice length, if the CPU‘s load is at
20      * least twice that of our own weight (i.e. dont track it
21      * when there are only lesser-weight tasks around):
22      */
23     if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
24         se->statistics.slice_max = max(se->statistics.slice_max,
25             se->sum_exec_runtime - se->prev_sum_exec_runtime);
26     }
27 #endif
28     se->prev_sum_exec_runtime = se->sum_exec_runtime;
29 }

 1 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 2 {
 3     struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
 4     struct rb_node *parent = NULL;
 5     struct sched_entity *entry;
 6     int leftmost = 1;
 7 
 8     /*
 9      * Find the right place in the rbtree:
10      */
11     while (*link) {
12         parent = *link;
13         entry = rb_entry(parent, struct sched_entity, run_node);
14         /*
15          * We dont care about collisions. Nodes with
16          * the same key stay together.
17          */
18         if (entity_before(se, entry)) {
19             link = &parent->rb_left;
20         } else {
21             link = &parent->rb_right;
22             leftmost = 0;
23         }
24     }
25 
26     /*
27      * Maintain a cache of leftmost tree entries (it is frequently
28      * used):
29      */
30     if (leftmost)
31         cfs_rq->rb_leftmost = &se->run_node;
32 
33     rb_link_node(&se->run_node, parent, link);
34     rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
35 }

 1 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 2 {
 3     if (cfs_rq->rb_leftmost == &se->run_node) {
 4         struct rb_node *next_node;
 5 
 6         next_node = rb_next(&se->run_node);
 7         cfs_rq->rb_leftmost = next_node;
 8     }
 9 
10     rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
11 }

 1 static int
 2 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
 3 {
 4     unsigned long flags;
 5     int cpu, success = 0;
 6 
 7     /*
 8      * If we are going to wake up a thread waiting for CONDITION we
 9      * need to ensure that CONDITION=1 done by the caller can not be
10      * reordered with p->state check below. This pairs with mb() in
11      * set_current_state() the waiting thread does.
12      */
13     smp_mb__before_spinlock();
14     raw_spin_lock_irqsave(&p->pi_lock, flags);
15     if (!(p->state & state))
16         goto out;
17 
18     success = 1; /* we‘re going to change ->state */
19     cpu = task_cpu(p);
20 
21     if (p->on_rq && ttwu_remote(p, wake_flags))
22         goto stat;
23 
24 #ifdef CONFIG_SMP
25     /*
26      * If the owning (remote) cpu is still in the middle of schedule() with
27      * this task as prev, wait until its done referencing the task.
28      */
29     while (p->on_cpu)
30         cpu_relax();
31     /*
32      * Pairs with the smp_wmb() in finish_lock_switch().
33      */
34     smp_rmb();
35 
36     p->sched_contributes_to_load = !!task_contributes_to_load(p);
37     p->state = TASK_WAKING;
38 
39     if (p->sched_class->task_waking)
40         p->sched_class->task_waking(p);
41 
42     cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
43     if (task_cpu(p) != cpu) {
44         wake_flags |= WF_MIGRATED;
45         set_task_cpu(p, cpu);
46     }
47 #endif /* CONFIG_SMP */
48 
49     ttwu_queue(p, cpu);
50 stat:
51     ttwu_stat(p, cpu, wake_flags);
52 out:
53     raw_spin_unlock_irqrestore(&p->pi_lock, flags);
54 
55     return success;
56 }

 1 void wake_up_new_task(struct task_struct *p)
 2 {
 3     unsigned long flags;
 4     struct rq *rq;
 5 
 6     raw_spin_lock_irqsave(&p->pi_lock, flags);
 7 #ifdef CONFIG_SMP
 8     /*
 9      * Fork balancing, do it here and not earlier because:
10      *  - cpus_allowed can change in the fork path
11      *  - any previously selected cpu might disappear through hotplug
12      */
13     set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
14 #endif
15 
16     /* Initialize new task‘s runnable average */
17     init_task_runnable_average(p);
18     rq = __task_rq_lock(p);
19     activate_task(rq, p, 0);
20     p->on_rq = 1;
21     trace_sched_wakeup_new(p, true);
22     check_preempt_curr(rq, p, WF_FORK);
23 #ifdef CONFIG_SMP
24     if (p->sched_class->task_woken)
25         p->sched_class->task_woken(rq, p);
26 #endif
27     task_rq_unlock(rq, p, &flags);
28 }

 1 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
 2 {
 3     struct task_struct *curr = rq->curr;
 4     struct sched_entity *se = &curr->se, *pse = &p->se;
 5     struct cfs_rq *cfs_rq = task_cfs_rq(curr);
 6     int scale = cfs_rq->nr_running >= sched_nr_latency;
 7     int next_buddy_marked = 0;
 8 
 9     if (unlikely(se == pse))
10         return;
11 
12     /*
13      * This is possible from callers such as move_task(), in which we
14      * unconditionally check_prempt_curr() after an enqueue (which may have
15      * lead to a throttle).  This both saves work and prevents false
16      * next-buddy nomination below.
17      */
18     if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
19         return;
20 
21     if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
22         set_next_buddy(pse);
23         next_buddy_marked = 1;
24     }
25 
26     /*
27      * We can come here with TIF_NEED_RESCHED already set from new task
28      * wake up path.
29      *
30      * Note: this also catches the edge-case of curr being in a throttled
31      * group (e.g. via set_curr_task), since update_curr() (in the
32      * enqueue of curr) will have resulted in resched being set.  This
33      * prevents us from potentially nominating it as a false LAST_BUDDY
34      * below.
35      */
36     if (test_tsk_need_resched(curr))
37         return;
38 
39     /* Idle tasks are by definition preempted by non-idle tasks. */
40     if (unlikely(curr->policy == SCHED_IDLE) &&
41         likely(p->policy != SCHED_IDLE))
42         goto preempt;
43 
44     /*
45      *  do not preempt non-idle tasks (their preemption
46      * is driven by the tick):
47      */
48     if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
49         return;
50 
51     find_matching_se(&se, &pse);
52     update_curr(cfs_rq_of(se));
53     BUG_ON(!pse);
54     if (wakeup_preempt_entity(se, pse) == 1) {
55         /*
56          * Bias pick_next to pick the sched entity that is
57          * triggering this preemption.
58          */
59         if (!next_buddy_marked)
60             set_next_buddy(pse);
61         goto preempt;
62     }
63 
64     return;
65 
66 preempt:
67     resched_task(curr);
68     /*
69      * Only set the backward buddy when the current task is still
70      * on the rq. This can happen when a wakeup gets interleaved
71      * with schedule on the ->pre_schedule() or idle_balance()
72      * point, either of which can * drop the rq lock.
73      *
74      * Also, during early boot the idle thread is in the fair class,
75      * for obvious reasons its a bad idea to schedule back to it.
76      */
77     if (unlikely(!se->on_rq || curr == rq->idle))
78         return;
79 
80     if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
81         set_last_buddy(se);
82 }

 1 static int
 2 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
 3 {
 4     s64 gran, vdiff = curr->vruntime - se->vruntime;
 5 
 6     if (vdiff <= 0)
 7         return -1;
 8 
 9     gran = wakeup_gran(curr, se);
10     if (vdiff > gran)
11         return 1;
12 
13     return 0;
14 }

 1 static void task_waking_fair(struct task_struct *p)
 2 {
 3     struct sched_entity *se = &p->se;
 4     struct cfs_rq *cfs_rq = cfs_rq_of(se);
 5     u64 min_vruntime;
 6 
 7 #ifndef CONFIG_64BIT
 8     u64 min_vruntime_copy;
 9 
10     do {
11         min_vruntime_copy = cfs_rq->min_vruntime_copy;
12         smp_rmb();
13         min_vruntime = cfs_rq->min_vruntime;
14     } while (min_vruntime != min_vruntime_copy);
15 #else
16     min_vruntime = cfs_rq->min_vruntime;
17 #endif
18 
19     se->vruntime -= min_vruntime;
20     record_wakee(p);
21 }

 1 static void task_fork_fair(struct task_struct *p)
 2 {
 3     struct cfs_rq *cfs_rq;
 4     struct sched_entity *se = &p->se, *curr;
 5     int this_cpu = smp_processor_id();
 6     struct rq *rq = this_rq();
 7     unsigned long flags;
 8 
 9     raw_spin_lock_irqsave(&rq->lock, flags);
10 
11     update_rq_clock(rq);
12 
13     cfs_rq = task_cfs_rq(current);
14     curr = cfs_rq->curr;
15 
16     /*
17      * Not only the cpu but also the task_group of the parent might have
18      * been changed after parent->se.parent,cfs_rq were copied to
19      * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
20      * of child point to valid ones.
21      */
22     rcu_read_lock();
23     __set_task_cpu(p, this_cpu);
24     rcu_read_unlock();
25 
26     update_curr(cfs_rq);
27 
28     if (curr)
29         se->vruntime = curr->vruntime;
30     place_entity(cfs_rq, se, 1);
31 
32     if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
33         /*
34          * Upon rescheduling, sched_class::put_prev_task() will place
35          * ‘current‘ within the tree based on its new key value.
36          */
37         swap(curr->vruntime, se->vruntime);
38         resched_task(rq->curr);
39     }
40 
41     se->vruntime -= cfs_rq->min_vruntime;
42 
43     raw_spin_unlock_irqrestore(&rq->lock, flags);
44 }

 1 static void
 2 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
 3 {
 4     u64 vruntime = cfs_rq->min_vruntime;
 5 
 6     /*
 7      * The ‘current‘ period is already promised to the current tasks,
 8      * however the extra weight of the new task will slow them down a
 9      * little, place the new task so that it fits in the slot that
10      * stays open at the end.
11      */
12     if (initial && sched_feat(START_DEBIT))
13         vruntime += sched_vslice(cfs_rq, se);
14 
15     /* sleeps up to a single latency don‘t count. */
16     if (!initial) {
17         unsigned long thresh = sysctl_sched_latency;
18 
19         /*
20          * Halve their sleep time‘s effect, to allow
21          * for a gentler effect of sleepers:
22          */
23         if (sched_feat(GENTLE_FAIR_SLEEPERS))
24             thresh >>= 1;
25 
26         vruntime -= thresh;
27     }
28 
29     /* ensure we never gain time by being placed backwards. */
30     se->vruntime = max_vruntime(se->vruntime, vruntime);
31 }