sched: fix buddie group latency
[safe/jmp/linux-2.6] / kernel / workqueue.c
index 4a26a13..2f44583 100644 (file)
@@ -9,7 +9,7 @@
  * Derived from the taskqueue/keventd code by:
  *
  *   David Woodhouse <dwmw2@infradead.org>
- *   Andrew Morton <andrewm@uow.edu.au>
+ *   Andrew Morton
  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
  *   Theodore Ts'o <tytso@mit.edu>
  *
@@ -62,6 +62,7 @@ struct workqueue_struct {
        const char *name;
        int singlethread;
        int freezeable;         /* Freeze threads during suspend */
+       int rt;
 #ifdef CONFIG_LOCKDEP
        struct lockdep_map lockdep_map;
 #endif
@@ -72,7 +73,7 @@ static DEFINE_SPINLOCK(workqueue_lock);
 static LIST_HEAD(workqueues);
 
 static int singlethread_cpu __read_mostly;
-static cpumask_t cpu_singlethread_map __read_mostly;
+static const struct cpumask *cpu_singlethread_map __read_mostly;
 /*
  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
@@ -80,24 +81,24 @@ static cpumask_t cpu_singlethread_map __read_mostly;
  * use cpu_possible_map, the cpumask below is more a documentation
  * than optimization.
  */
-static cpumask_t cpu_populated_map __read_mostly;
+static cpumask_var_t cpu_populated_map __read_mostly;
 
 /* If it's single threaded, it isn't in the list of workqueues. */
-static inline int is_single_threaded(struct workqueue_struct *wq)
+static inline int is_wq_single_threaded(struct workqueue_struct *wq)
 {
        return wq->singlethread;
 }
 
-static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
+static const struct cpumask *wq_cpu_map(struct workqueue_struct *wq)
 {
-       return is_single_threaded(wq)
-               ? &cpu_singlethread_map : &cpu_populated_map;
+       return is_wq_single_threaded(wq)
+               ? cpu_singlethread_map : cpu_populated_map;
 }
 
 static
 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
 {
-       if (unlikely(is_single_threaded(wq)))
+       if (unlikely(is_wq_single_threaded(wq)))
                cpu = singlethread_cpu;
        return per_cpu_ptr(wq->cpu_wq, cpu);
 }
@@ -290,11 +291,11 @@ static void run_workqueue(struct cpu_workqueue_struct *cwq)
 
                BUG_ON(get_wq_data(work) != cwq);
                work_clear_pending(work);
-               lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
-               lock_acquire(&lockdep_map, 0, 0, 0, 2, _THIS_IP_);
+               lock_map_acquire(&cwq->wq->lockdep_map);
+               lock_map_acquire(&lockdep_map);
                f(work);
-               lock_release(&lockdep_map, 1, _THIS_IP_);
-               lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
+               lock_map_release(&lockdep_map);
+               lock_map_release(&cwq->wq->lockdep_map);
 
                if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
                        printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
@@ -409,12 +410,12 @@ static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
  */
 void flush_workqueue(struct workqueue_struct *wq)
 {
-       const cpumask_t *cpu_map = wq_cpu_map(wq);
+       const struct cpumask *cpu_map = wq_cpu_map(wq);
        int cpu;
 
        might_sleep();
-       lock_acquire(&wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
-       lock_release(&wq->lockdep_map, 1, _THIS_IP_);
+       lock_map_acquire(&wq->lockdep_map);
+       lock_map_release(&wq->lockdep_map);
        for_each_cpu_mask_nr(cpu, *cpu_map)
                flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 }
@@ -441,8 +442,8 @@ int flush_work(struct work_struct *work)
        if (!cwq)
                return 0;
 
-       lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
-       lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
+       lock_map_acquire(&cwq->wq->lockdep_map);
+       lock_map_release(&cwq->wq->lockdep_map);
 
        prev = NULL;
        spin_lock_irq(&cwq->lock);
@@ -531,13 +532,13 @@ static void wait_on_work(struct work_struct *work)
 {
        struct cpu_workqueue_struct *cwq;
        struct workqueue_struct *wq;
-       const cpumask_t *cpu_map;
+       const struct cpumask *cpu_map;
        int cpu;
 
        might_sleep();
 
-       lock_acquire(&work->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
-       lock_release(&work->lockdep_map, 1, _THIS_IP_);
+       lock_map_acquire(&work->lockdep_map);
+       lock_map_release(&work->lockdep_map);
 
        cwq = get_wq_data(work);
        if (!cwq)
@@ -766,8 +767,9 @@ init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 
 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 {
+       struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
        struct workqueue_struct *wq = cwq->wq;
-       const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
+       const char *fmt = is_wq_single_threaded(wq) ? "%s" : "%s/%d";
        struct task_struct *p;
 
        p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
@@ -781,7 +783,8 @@ static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
         */
        if (IS_ERR(p))
                return PTR_ERR(p);
-
+       if (cwq->wq->rt)
+               sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
        cwq->thread = p;
 
        return 0;
@@ -801,6 +804,7 @@ static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
 struct workqueue_struct *__create_workqueue_key(const char *name,
                                                int singlethread,
                                                int freezeable,
+                                               int rt,
                                                struct lock_class_key *key,
                                                const char *lock_name)
 {
@@ -822,6 +826,7 @@ struct workqueue_struct *__create_workqueue_key(const char *name,
        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
        wq->singlethread = singlethread;
        wq->freezeable = freezeable;
+       wq->rt = rt;
        INIT_LIST_HEAD(&wq->list);
 
        if (singlethread) {
@@ -872,8 +877,8 @@ static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
        if (cwq->thread == NULL)
                return;
 
-       lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
-       lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
+       lock_map_acquire(&cwq->wq->lockdep_map);
+       lock_map_release(&cwq->wq->lockdep_map);
 
        flush_cpu_workqueue(cwq);
        /*
@@ -898,7 +903,7 @@ static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
  */
 void destroy_workqueue(struct workqueue_struct *wq)
 {
-       const cpumask_t *cpu_map = wq_cpu_map(wq);
+       const struct cpumask *cpu_map = wq_cpu_map(wq);
        int cpu;
 
        cpu_maps_update_begin();
@@ -928,7 +933,7 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 
        switch (action) {
        case CPU_UP_PREPARE:
-               cpu_set(cpu, cpu_populated_map);
+               cpumask_set_cpu(cpu, cpu_populated_map);
        }
 undo:
        list_for_each_entry(wq, &workqueues, list) {
@@ -959,17 +964,64 @@ undo:
        switch (action) {
        case CPU_UP_CANCELED:
        case CPU_POST_DEAD:
-               cpu_clear(cpu, cpu_populated_map);
+               cpumask_clear_cpu(cpu, cpu_populated_map);
        }
 
        return ret;
 }
 
+#ifdef CONFIG_SMP
+struct work_for_cpu {
+       struct work_struct work;
+       long (*fn)(void *);
+       void *arg;
+       long ret;
+};
+
+static void do_work_for_cpu(struct work_struct *w)
+{
+       struct work_for_cpu *wfc = container_of(w, struct work_for_cpu, work);
+
+       wfc->ret = wfc->fn(wfc->arg);
+}
+
+/**
+ * work_on_cpu - run a function in user context on a particular cpu
+ * @cpu: the cpu to run on
+ * @fn: the function to run
+ * @arg: the function arg
+ *
+ * This will return -EINVAL in the cpu is not online, or the return value
+ * of @fn otherwise.
+ */
+long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
+{
+       struct work_for_cpu wfc;
+
+       INIT_WORK(&wfc.work, do_work_for_cpu);
+       wfc.fn = fn;
+       wfc.arg = arg;
+       get_online_cpus();
+       if (unlikely(!cpu_online(cpu)))
+               wfc.ret = -EINVAL;
+       else {
+               schedule_work_on(cpu, &wfc.work);
+               flush_work(&wfc.work);
+       }
+       put_online_cpus();
+
+       return wfc.ret;
+}
+EXPORT_SYMBOL_GPL(work_on_cpu);
+#endif /* CONFIG_SMP */
+
 void __init init_workqueues(void)
 {
-       cpu_populated_map = cpu_online_map;
-       singlethread_cpu = first_cpu(cpu_possible_map);
-       cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
+       alloc_cpumask_var(&cpu_populated_map, GFP_KERNEL);
+
+       cpumask_copy(cpu_populated_map, cpu_online_mask);
+       singlethread_cpu = cpumask_first(cpu_possible_mask);
+       cpu_singlethread_map = cpumask_of(singlethread_cpu);
        hotcpu_notifier(workqueue_cpu_callback, 0);
        keventd_wq = create_workqueue("events");
        BUG_ON(!keventd_wq);