[PATCH] pi-futex: scheduler support for pi
authorIngo Molnar <mingo@elte.hu>
Tue, 27 Jun 2006 09:54:51 +0000 (02:54 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 28 Jun 2006 00:32:46 +0000 (17:32 -0700)
Add framework to boost/unboost the priority of RT tasks.

This consists of:

 - caching the 'normal' priority in ->normal_prio
 - providing a functions to set/get the priority of the task
 - make sched_setscheduler() aware of boosting

The effective_prio() cleanups also fix a priority-calculation bug pointed out
by Andrey Gelman, in set_user_nice().

has_rt_policy() fix: Peter Williams <pwil3058@bigpond.net.au>

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Andrey Gelman <agelman@012.net.il>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
include/linux/init_task.h
include/linux/sched.h
kernel/sched.c

index e127ef7..678c1a9 100644 (file)
@@ -87,6 +87,7 @@ extern struct group_info init_groups;
        .lock_depth     = -1,                                           \
        .prio           = MAX_PRIO-20,                                  \
        .static_prio    = MAX_PRIO-20,                                  \
+       .normal_prio    = MAX_PRIO-20,                                  \
        .policy         = SCHED_NORMAL,                                 \
        .cpus_allowed   = CPU_MASK_ALL,                                 \
        .mm             = NULL,                                         \
@@ -122,6 +123,7 @@ extern struct group_info init_groups;
        .journal_info   = NULL,                                         \
        .cpu_timers     = INIT_CPU_TIMERS(tsk.cpu_timers),              \
        .fs_excl        = ATOMIC_INIT(0),                               \
+       .pi_lock        = SPIN_LOCK_UNLOCKED,                           \
 }
 
 
index 0bc81a1..6f16764 100644 (file)
@@ -495,8 +495,11 @@ struct signal_struct {
 
 #define MAX_PRIO               (MAX_RT_PRIO + 40)
 
-#define rt_task(p)             (unlikely((p)->prio < MAX_RT_PRIO))
+#define rt_prio(prio)          unlikely((prio) < MAX_RT_PRIO)
+#define rt_task(p)             rt_prio((p)->prio)
 #define batch_task(p)          (unlikely((p)->policy == SCHED_BATCH))
+#define has_rt_policy(p) \
+       unlikely((p)->policy != SCHED_NORMAL && (p)->policy != SCHED_BATCH)
 
 /*
  * Some day this will be a full-fledged user tracking system..
@@ -725,7 +728,7 @@ struct task_struct {
 #endif
 #endif
        int load_weight;        /* for niceness load balancing purposes */
-       int prio, static_prio;
+       int prio, static_prio, normal_prio;
        struct list_head run_list;
        prio_array_t *array;
 
@@ -852,6 +855,9 @@ struct task_struct {
 /* Protection of (de-)allocation: mm, files, fs, tty, keyrings */
        spinlock_t alloc_lock;
 
+       /* Protection of the PI data structures: */
+       spinlock_t pi_lock;
+
 #ifdef CONFIG_DEBUG_MUTEXES
        /* mutex deadlock detection */
        struct mutex_waiter *blocked_on;
@@ -1018,6 +1024,17 @@ static inline void idle_task_exit(void) {}
 #endif
 
 extern void sched_idle_next(void);
+
+#ifdef CONFIG_RT_MUTEXES
+extern int rt_mutex_getprio(task_t *p);
+extern void rt_mutex_setprio(task_t *p, int prio);
+#else
+static inline int rt_mutex_getprio(task_t *p)
+{
+       return p->normal_prio;
+}
+#endif
+
 extern void set_user_nice(task_t *p, long nice);
 extern int task_prio(const task_t *p);
 extern int task_nice(const task_t *p);
index 15abf08..08431f0 100644 (file)
@@ -355,6 +355,25 @@ static inline void finish_lock_switch(runqueue_t *rq, task_t *prev)
 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
 
 /*
+ * __task_rq_lock - lock the runqueue a given task resides on.
+ * Must be called interrupts disabled.
+ */
+static inline runqueue_t *__task_rq_lock(task_t *p)
+       __acquires(rq->lock)
+{
+       struct runqueue *rq;
+
+repeat_lock_task:
+       rq = task_rq(p);
+       spin_lock(&rq->lock);
+       if (unlikely(rq != task_rq(p))) {
+               spin_unlock(&rq->lock);
+               goto repeat_lock_task;
+       }
+       return rq;
+}
+
+/*
  * task_rq_lock - lock the runqueue a given task resides on and disable
  * interrupts.  Note the ordering: we can safely lookup the task_rq without
  * explicitly disabling preemption.
@@ -375,6 +394,12 @@ repeat_lock_task:
        return rq;
 }
 
+static inline void __task_rq_unlock(runqueue_t *rq)
+       __releases(rq->lock)
+{
+       spin_unlock(&rq->lock);
+}
+
 static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
        __releases(rq->lock)
 {
@@ -638,7 +663,7 @@ static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
 }
 
 /*
- * effective_prio - return the priority that is based on the static
+ * __normal_prio - return the priority that is based on the static
  * priority but is modified by bonuses/penalties.
  *
  * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
@@ -651,13 +676,11 @@ static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
  *
  * Both properties are important to certain workloads.
  */
-static int effective_prio(task_t *p)
+
+static inline int __normal_prio(task_t *p)
 {
        int bonus, prio;
 
-       if (rt_task(p))
-               return p->prio;
-
        bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
 
        prio = p->static_prio - bonus;
@@ -692,7 +715,7 @@ static int effective_prio(task_t *p)
 
 static void set_load_weight(task_t *p)
 {
-       if (rt_task(p)) {
+       if (has_rt_policy(p)) {
 #ifdef CONFIG_SMP
                if (p == task_rq(p)->migration_thread)
                        /*
@@ -731,6 +754,44 @@ static inline void dec_nr_running(task_t *p, runqueue_t *rq)
 }
 
 /*
+ * Calculate the expected normal priority: i.e. priority
+ * without taking RT-inheritance into account. Might be
+ * boosted by interactivity modifiers. Changes upon fork,
+ * setprio syscalls, and whenever the interactivity
+ * estimator recalculates.
+ */
+static inline int normal_prio(task_t *p)
+{
+       int prio;
+
+       if (has_rt_policy(p))
+               prio = MAX_RT_PRIO-1 - p->rt_priority;
+       else
+               prio = __normal_prio(p);
+       return prio;
+}
+
+/*
+ * Calculate the current priority, i.e. the priority
+ * taken into account by the scheduler. This value might
+ * be boosted by RT tasks, or might be boosted by
+ * interactivity modifiers. Will be RT if the task got
+ * RT-boosted. If not then it returns p->normal_prio.
+ */
+static int effective_prio(task_t *p)
+{
+       p->normal_prio = normal_prio(p);
+       /*
+        * If we are RT tasks or we were boosted to RT priority,
+        * keep the priority unchanged. Otherwise, update priority
+        * to the normal priority:
+        */
+       if (!rt_prio(p->prio))
+               return p->normal_prio;
+       return p->prio;
+}
+
+/*
  * __activate_task - move a task to the runqueue.
  */
 static void __activate_task(task_t *p, runqueue_t *rq)
@@ -752,6 +813,10 @@ static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
        inc_nr_running(p, rq);
 }
 
+/*
+ * Recalculate p->normal_prio and p->prio after having slept,
+ * updating the sleep-average too:
+ */
 static int recalc_task_prio(task_t *p, unsigned long long now)
 {
        /* Caller must always ensure 'now >= p->timestamp' */
@@ -1448,6 +1513,12 @@ void fastcall sched_fork(task_t *p, int clone_flags)
         * event cannot wake it up and insert it on the runqueue either.
         */
        p->state = TASK_RUNNING;
+
+       /*
+        * Make sure we do not leak PI boosting priority to the child:
+        */
+       p->prio = current->normal_prio;
+
        INIT_LIST_HEAD(&p->run_list);
        p->array = NULL;
 #ifdef CONFIG_SCHEDSTATS
@@ -1527,6 +1598,7 @@ void fastcall wake_up_new_task(task_t *p, unsigned long clone_flags)
                                __activate_task(p, rq);
                        else {
                                p->prio = current->prio;
+                               p->normal_prio = current->normal_prio;
                                list_add_tail(&p->run_list, &current->run_list);
                                p->array = current->array;
                                p->array->nr_active++;
@@ -3668,12 +3740,65 @@ long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
 
 EXPORT_SYMBOL(sleep_on_timeout);
 
+#ifdef CONFIG_RT_MUTEXES
+
+/*
+ * rt_mutex_setprio - set the current priority of a task
+ * @p: task
+ * @prio: prio value (kernel-internal form)
+ *
+ * This function changes the 'effective' priority of a task. It does
+ * not touch ->normal_prio like __setscheduler().
+ *
+ * Used by the rt_mutex code to implement priority inheritance logic.
+ */
+void rt_mutex_setprio(task_t *p, int prio)
+{
+       unsigned long flags;
+       prio_array_t *array;
+       runqueue_t *rq;
+       int oldprio;
+
+       BUG_ON(prio < 0 || prio > MAX_PRIO);
+
+       rq = task_rq_lock(p, &flags);
+
+       oldprio = p->prio;
+       array = p->array;
+       if (array)
+               dequeue_task(p, array);
+       p->prio = prio;
+
+       if (array) {
+               /*
+                * If changing to an RT priority then queue it
+                * in the active array!
+                */
+               if (rt_task(p))
+                       array = rq->active;
+               enqueue_task(p, array);
+               /*
+                * Reschedule if we are currently running on this runqueue and
+                * our priority decreased, or if we are not currently running on
+                * this runqueue and our priority is higher than the current's
+                */
+               if (task_running(rq, p)) {
+                       if (p->prio > oldprio)
+                               resched_task(rq->curr);
+               } else if (TASK_PREEMPTS_CURR(p, rq))
+                       resched_task(rq->curr);
+       }
+       task_rq_unlock(rq, &flags);
+}
+
+#endif
+
 void set_user_nice(task_t *p, long nice)
 {
        unsigned long flags;
        prio_array_t *array;
        runqueue_t *rq;
-       int old_prio, new_prio, delta;
+       int old_prio, delta;
 
        if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
                return;
@@ -3688,7 +3813,7 @@ void set_user_nice(task_t *p, long nice)
         * it wont have any effect on scheduling until the task is
         * not SCHED_NORMAL/SCHED_BATCH:
         */
-       if (rt_task(p)) {
+       if (has_rt_policy(p)) {
                p->static_prio = NICE_TO_PRIO(nice);
                goto out_unlock;
        }
@@ -3698,12 +3823,11 @@ void set_user_nice(task_t *p, long nice)
                dec_raw_weighted_load(rq, p);
        }
 
-       old_prio = p->prio;
-       new_prio = NICE_TO_PRIO(nice);
-       delta = new_prio - old_prio;
        p->static_prio = NICE_TO_PRIO(nice);
        set_load_weight(p);
-       p->prio += delta;
+       old_prio = p->prio;
+       p->prio = effective_prio(p);
+       delta = p->prio - old_prio;
 
        if (array) {
                enqueue_task(p, array);
@@ -3718,7 +3842,6 @@ void set_user_nice(task_t *p, long nice)
 out_unlock:
        task_rq_unlock(rq, &flags);
 }
-
 EXPORT_SYMBOL(set_user_nice);
 
 /*
@@ -3833,16 +3956,14 @@ static void __setscheduler(struct task_struct *p, int policy, int prio)
        BUG_ON(p->array);
        p->policy = policy;
        p->rt_priority = prio;
-       if (policy != SCHED_NORMAL && policy != SCHED_BATCH) {
-               p->prio = MAX_RT_PRIO-1 - p->rt_priority;
-       } else {
-               p->prio = p->static_prio;
-               /*
-                * SCHED_BATCH tasks are treated as perpetual CPU hogs:
-                */
-               if (policy == SCHED_BATCH)
-                       p->sleep_avg = 0;
-       }
+       p->normal_prio = normal_prio(p);
+       /* we are holding p->pi_lock already */
+       p->prio = rt_mutex_getprio(p);
+       /*
+        * SCHED_BATCH tasks are treated as perpetual CPU hogs:
+        */
+       if (policy == SCHED_BATCH)
+               p->sleep_avg = 0;
        set_load_weight(p);
 }
 
@@ -3912,14 +4033,20 @@ recheck:
        if (retval)
                return retval;
        /*
+        * make sure no PI-waiters arrive (or leave) while we are
+        * changing the priority of the task:
+        */
+       spin_lock_irqsave(&p->pi_lock, flags);
+       /*
         * To be able to change p->policy safely, the apropriate
         * runqueue lock must be held.
         */
-       rq = task_rq_lock(p, &flags);
+       rq = __task_rq_lock(p);
        /* recheck policy now with rq lock held */
        if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
                policy = oldpolicy = -1;
-               task_rq_unlock(rq, &flags);
+               __task_rq_unlock(rq);
+               spin_unlock_irqrestore(&p->pi_lock, flags);
                goto recheck;
        }
        array = p->array;
@@ -3940,7 +4067,9 @@ recheck:
                } else if (TASK_PREEMPTS_CURR(p, rq))
                        resched_task(rq->curr);
        }
-       task_rq_unlock(rq, &flags);
+       __task_rq_unlock(rq);
+       spin_unlock_irqrestore(&p->pi_lock, flags);
+
        return 0;
 }
 EXPORT_SYMBOL_GPL(sched_setscheduler);
@@ -4575,7 +4704,7 @@ void __devinit init_idle(task_t *idle, int cpu)
        idle->timestamp = sched_clock();
        idle->sleep_avg = 0;
        idle->array = NULL;
-       idle->prio = MAX_PRIO;
+       idle->prio = idle->normal_prio = MAX_PRIO;
        idle->state = TASK_RUNNING;
        idle->cpus_allowed = cpumask_of_cpu(cpu);
        set_task_cpu(idle, cpu);
@@ -6582,7 +6711,8 @@ void normalize_rt_tasks(void)
                if (!rt_task(p))
                        continue;
 
-               rq = task_rq_lock(p, &flags);
+               spin_lock_irqsave(&p->pi_lock, flags);
+               rq = __task_rq_lock(p);
 
                array = p->array;
                if (array)
@@ -6593,7 +6723,8 @@ void normalize_rt_tasks(void)
                        resched_task(rq->curr);
                }
 
-               task_rq_unlock(rq, &flags);
+               __task_rq_unlock(rq);
+               spin_unlock_irqrestore(&p->pi_lock, flags);
        }
        read_unlock_irq(&tasklist_lock);
 }