mmc: s3c6410: enable ADMA feature in 6410 sdhci controller
[safe/jmp/linux-2.6] / kernel / workqueue.c
index fb56fed..77dabbf 100644 (file)
@@ -9,11 +9,11 @@
  * 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>
  *
- * Made to use alloc_percpu by Christoph Lameter <clameter@sgi.com>.
+ * Made to use alloc_percpu by Christoph Lameter.
  */
 
 #include <linux/module.h>
@@ -32,6 +32,9 @@
 #include <linux/freezer.h>
 #include <linux/kallsyms.h>
 #include <linux/debug_locks.h>
+#include <linux/lockdep.h>
+#define CREATE_TRACE_POINTS
+#include <trace/events/workqueue.h>
 
 /*
  * The per-CPU workqueue (if single thread, we always use the first
@@ -47,9 +50,6 @@ struct cpu_workqueue_struct {
 
        struct workqueue_struct *wq;
        struct task_struct *thread;
-       int should_stop;
-
-       int run_depth;          /* Detect run_workqueue() recursion depth */
 } ____cacheline_aligned;
 
 /*
@@ -62,34 +62,153 @@ 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
+};
+
+#ifdef CONFIG_DEBUG_OBJECTS_WORK
+
+static struct debug_obj_descr work_debug_descr;
+
+/*
+ * fixup_init is called when:
+ * - an active object is initialized
+ */
+static int work_fixup_init(void *addr, enum debug_obj_state state)
+{
+       struct work_struct *work = addr;
+
+       switch (state) {
+       case ODEBUG_STATE_ACTIVE:
+               cancel_work_sync(work);
+               debug_object_init(work, &work_debug_descr);
+               return 1;
+       default:
+               return 0;
+       }
+}
+
+/*
+ * fixup_activate is called when:
+ * - an active object is activated
+ * - an unknown object is activated (might be a statically initialized object)
+ */
+static int work_fixup_activate(void *addr, enum debug_obj_state state)
+{
+       struct work_struct *work = addr;
+
+       switch (state) {
+
+       case ODEBUG_STATE_NOTAVAILABLE:
+               /*
+                * This is not really a fixup. The work struct was
+                * statically initialized. We just make sure that it
+                * is tracked in the object tracker.
+                */
+               if (test_bit(WORK_STRUCT_STATIC, work_data_bits(work))) {
+                       debug_object_init(work, &work_debug_descr);
+                       debug_object_activate(work, &work_debug_descr);
+                       return 0;
+               }
+               WARN_ON_ONCE(1);
+               return 0;
+
+       case ODEBUG_STATE_ACTIVE:
+               WARN_ON(1);
+
+       default:
+               return 0;
+       }
+}
+
+/*
+ * fixup_free is called when:
+ * - an active object is freed
+ */
+static int work_fixup_free(void *addr, enum debug_obj_state state)
+{
+       struct work_struct *work = addr;
+
+       switch (state) {
+       case ODEBUG_STATE_ACTIVE:
+               cancel_work_sync(work);
+               debug_object_free(work, &work_debug_descr);
+               return 1;
+       default:
+               return 0;
+       }
+}
+
+static struct debug_obj_descr work_debug_descr = {
+       .name           = "work_struct",
+       .fixup_init     = work_fixup_init,
+       .fixup_activate = work_fixup_activate,
+       .fixup_free     = work_fixup_free,
 };
 
-/* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
-   threads to each one as cpus come/go. */
-static DEFINE_MUTEX(workqueue_mutex);
+static inline void debug_work_activate(struct work_struct *work)
+{
+       debug_object_activate(work, &work_debug_descr);
+}
+
+static inline void debug_work_deactivate(struct work_struct *work)
+{
+       debug_object_deactivate(work, &work_debug_descr);
+}
+
+void __init_work(struct work_struct *work, int onstack)
+{
+       if (onstack)
+               debug_object_init_on_stack(work, &work_debug_descr);
+       else
+               debug_object_init(work, &work_debug_descr);
+}
+EXPORT_SYMBOL_GPL(__init_work);
+
+void destroy_work_on_stack(struct work_struct *work)
+{
+       debug_object_free(work, &work_debug_descr);
+}
+EXPORT_SYMBOL_GPL(destroy_work_on_stack);
+
+#else
+static inline void debug_work_activate(struct work_struct *work) { }
+static inline void debug_work_deactivate(struct work_struct *work) { }
+#endif
+
+/* Serializes the accesses to the list of workqueues. */
+static DEFINE_SPINLOCK(workqueue_lock);
 static LIST_HEAD(workqueues);
 
 static int singlethread_cpu __read_mostly;
-static cpumask_t cpu_singlethread_map __read_mostly;
-/* optimization, we could use cpu_possible_map */
-static cpumask_t cpu_populated_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
+ * which comes in between can't use for_each_online_cpu(). We could
+ * use cpu_possible_map, the cpumask below is more a documentation
+ * than optimization.
+ */
+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);
 }
@@ -110,6 +229,16 @@ static inline void set_wq_data(struct work_struct *work,
        atomic_long_set(&work->data, new);
 }
 
+/*
+ * Clear WORK_STRUCT_PENDING and the workqueue on which it was queued.
+ */
+static inline void clear_wq_data(struct work_struct *work)
+{
+       unsigned long flags = *work_data_bits(work) &
+                               (1UL << WORK_STRUCT_STATIC);
+       atomic_long_set(&work->data, flags);
+}
+
 static inline
 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
 {
@@ -117,29 +246,28 @@ struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
 }
 
 static void insert_work(struct cpu_workqueue_struct *cwq,
-                               struct work_struct *work, int tail)
+                       struct work_struct *work, struct list_head *head)
 {
+       trace_workqueue_insertion(cwq->thread, work);
+
        set_wq_data(work, cwq);
        /*
         * Ensure that we get the right work->data if we see the
         * result of list_add() below, see try_to_grab_pending().
         */
        smp_wmb();
-       if (tail)
-               list_add_tail(&work->entry, &cwq->worklist);
-       else
-               list_add(&work->entry, &cwq->worklist);
+       list_add_tail(&work->entry, head);
        wake_up(&cwq->more_work);
 }
 
-/* Preempt must be disabled. */
 static void __queue_work(struct cpu_workqueue_struct *cwq,
                         struct work_struct *work)
 {
        unsigned long flags;
 
+       debug_work_activate(work);
        spin_lock_irqsave(&cwq->lock, flags);
-       insert_work(cwq, work, 1);
+       insert_work(cwq, work, &cwq->worklist);
        spin_unlock_irqrestore(&cwq->lock, flags);
 }
 
@@ -150,24 +278,46 @@ static void __queue_work(struct cpu_workqueue_struct *cwq,
  *
  * Returns 0 if @work was already on a queue, non-zero otherwise.
  *
- * We queue the work to the CPU it was submitted, but there is no
- * guarantee that it will be processed by that CPU.
+ * We queue the work to the CPU on which it was submitted, but if the CPU dies
+ * it can be processed by another CPU.
  */
-int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
+int queue_work(struct workqueue_struct *wq, struct work_struct *work)
+{
+       int ret;
+
+       ret = queue_work_on(get_cpu(), wq, work);
+       put_cpu();
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(queue_work);
+
+/**
+ * queue_work_on - queue work on specific cpu
+ * @cpu: CPU number to execute work on
+ * @wq: workqueue to use
+ * @work: work to queue
+ *
+ * Returns 0 if @work was already on a queue, non-zero otherwise.
+ *
+ * We queue the work to a specific CPU, the caller must ensure it
+ * can't go away.
+ */
+int
+queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
 {
        int ret = 0;
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
                BUG_ON(!list_empty(&work->entry));
-               __queue_work(wq_per_cpu(wq, get_cpu()), work);
-               put_cpu();
+               __queue_work(wq_per_cpu(wq, cpu), work);
                ret = 1;
        }
        return ret;
 }
-EXPORT_SYMBOL_GPL(queue_work);
+EXPORT_SYMBOL_GPL(queue_work_on);
 
-void delayed_work_timer_fn(unsigned long __data)
+static void delayed_work_timer_fn(unsigned long __data)
 {
        struct delayed_work *dwork = (struct delayed_work *)__data;
        struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
@@ -184,10 +334,9 @@ void delayed_work_timer_fn(unsigned long __data)
  *
  * Returns 0 if @work was already on a queue, non-zero otherwise.
  */
-int fastcall queue_delayed_work(struct workqueue_struct *wq,
+int queue_delayed_work(struct workqueue_struct *wq,
                        struct delayed_work *dwork, unsigned long delay)
 {
-       timer_stats_timer_set_start_info(&dwork->timer);
        if (delay == 0)
                return queue_work(wq, &dwork->work);
 
@@ -215,6 +364,8 @@ int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
                BUG_ON(timer_pending(timer));
                BUG_ON(!list_empty(&work->entry));
 
+               timer_stats_timer_set_start_info(&dwork->timer);
+
                /* This stores cwq for the moment, for the timer_fn */
                set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
                timer->expires = jiffies + delay;
@@ -234,31 +385,40 @@ EXPORT_SYMBOL_GPL(queue_delayed_work_on);
 static void run_workqueue(struct cpu_workqueue_struct *cwq)
 {
        spin_lock_irq(&cwq->lock);
-       cwq->run_depth++;
-       if (cwq->run_depth > 3) {
-               /* morton gets to eat his hat */
-               printk("%s: recursion depth exceeded: %d\n",
-                       __FUNCTION__, cwq->run_depth);
-               dump_stack();
-       }
        while (!list_empty(&cwq->worklist)) {
                struct work_struct *work = list_entry(cwq->worklist.next,
                                                struct work_struct, entry);
                work_func_t f = work->func;
-
+#ifdef CONFIG_LOCKDEP
+               /*
+                * It is permissible to free the struct work_struct
+                * from inside the function that is called from it,
+                * this we need to take into account for lockdep too.
+                * To avoid bogus "held lock freed" warnings as well
+                * as problems when looking into work->lockdep_map,
+                * make a copy and use that here.
+                */
+               struct lockdep_map lockdep_map = work->lockdep_map;
+#endif
+               trace_workqueue_execution(cwq->thread, work);
+               debug_work_deactivate(work);
                cwq->current_work = work;
                list_del_init(cwq->worklist.next);
                spin_unlock_irq(&cwq->lock);
 
                BUG_ON(get_wq_data(work) != cwq);
                work_clear_pending(work);
+               lock_map_acquire(&cwq->wq->lockdep_map);
+               lock_map_acquire(&lockdep_map);
                f(work);
+               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: "
                                        "%s/0x%08x/%d\n",
                                        current->comm, preempt_count(),
-                                       current->pid);
+                                       task_pid_nr(current));
                        printk(KERN_ERR "    last function: ");
                        print_symbol("%s\n", (unsigned long)f);
                        debug_show_held_locks(current);
@@ -268,48 +428,28 @@ static void run_workqueue(struct cpu_workqueue_struct *cwq)
                spin_lock_irq(&cwq->lock);
                cwq->current_work = NULL;
        }
-       cwq->run_depth--;
        spin_unlock_irq(&cwq->lock);
 }
 
-/*
- * NOTE: the caller must not touch *cwq if this func returns true
- */
-static int cwq_should_stop(struct cpu_workqueue_struct *cwq)
-{
-       int should_stop = cwq->should_stop;
-
-       if (unlikely(should_stop)) {
-               spin_lock_irq(&cwq->lock);
-               should_stop = cwq->should_stop && list_empty(&cwq->worklist);
-               if (should_stop)
-                       cwq->thread = NULL;
-               spin_unlock_irq(&cwq->lock);
-       }
-
-       return should_stop;
-}
-
 static int worker_thread(void *__cwq)
 {
        struct cpu_workqueue_struct *cwq = __cwq;
        DEFINE_WAIT(wait);
 
-       if (!cwq->wq->freezeable)
-               current->flags |= PF_NOFREEZE;
-
-       set_user_nice(current, -5);
+       if (cwq->wq->freezeable)
+               set_freezable();
 
        for (;;) {
                prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
-               if (!freezing(current) && !cwq->should_stop
-                   && list_empty(&cwq->worklist))
+               if (!freezing(current) &&
+                   !kthread_should_stop() &&
+                   list_empty(&cwq->worklist))
                        schedule();
                finish_wait(&cwq->more_work, &wait);
 
                try_to_freeze();
 
-               if (cwq_should_stop(cwq))
+               if (kthread_should_stop())
                        break;
 
                run_workqueue(cwq);
@@ -330,38 +470,43 @@ static void wq_barrier_func(struct work_struct *work)
 }
 
 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
-                                       struct wq_barrier *barr, int tail)
+                       struct wq_barrier *barr, struct list_head *head)
 {
-       INIT_WORK(&barr->work, wq_barrier_func);
+       /*
+        * debugobject calls are safe here even with cwq->lock locked
+        * as we know for sure that this will not trigger any of the
+        * checks and call back into the fixup functions where we
+        * might deadlock.
+        */
+       INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
        __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
 
        init_completion(&barr->done);
 
-       insert_work(cwq, &barr->work, tail);
+       debug_work_activate(&barr->work);
+       insert_work(cwq, &barr->work, head);
 }
 
-static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
+static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 {
-       if (cwq->thread == current) {
-               /*
-                * Probably keventd trying to flush its own queue. So simply run
-                * it by hand rather than deadlocking.
-                */
-               run_workqueue(cwq);
-       } else {
-               struct wq_barrier barr;
-               int active = 0;
+       int active = 0;
+       struct wq_barrier barr;
 
-               spin_lock_irq(&cwq->lock);
-               if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
-                       insert_wq_barrier(cwq, &barr, 1);
-                       active = 1;
-               }
-               spin_unlock_irq(&cwq->lock);
+       WARN_ON(cwq->thread == current);
+
+       spin_lock_irq(&cwq->lock);
+       if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
+               insert_wq_barrier(cwq, &barr, &cwq->worklist);
+               active = 1;
+       }
+       spin_unlock_irq(&cwq->lock);
 
-               if (active)
-                       wait_for_completion(&barr.done);
+       if (active) {
+               wait_for_completion(&barr.done);
+               destroy_work_on_stack(&barr.work);
        }
+
+       return active;
 }
 
 /**
@@ -377,28 +522,82 @@ static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
  * This function used to run the workqueues itself.  Now we just wait for the
  * helper threads to do it.
  */
-void fastcall flush_workqueue(struct workqueue_struct *wq)
+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();
-       for_each_cpu_mask(cpu, *cpu_map)
+       lock_map_acquire(&wq->lockdep_map);
+       lock_map_release(&wq->lockdep_map);
+       for_each_cpu(cpu, cpu_map)
                flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
 }
 EXPORT_SYMBOL_GPL(flush_workqueue);
 
+/**
+ * flush_work - block until a work_struct's callback has terminated
+ * @work: the work which is to be flushed
+ *
+ * Returns false if @work has already terminated.
+ *
+ * It is expected that, prior to calling flush_work(), the caller has
+ * arranged for the work to not be requeued, otherwise it doesn't make
+ * sense to use this function.
+ */
+int flush_work(struct work_struct *work)
+{
+       struct cpu_workqueue_struct *cwq;
+       struct list_head *prev;
+       struct wq_barrier barr;
+
+       might_sleep();
+       cwq = get_wq_data(work);
+       if (!cwq)
+               return 0;
+
+       lock_map_acquire(&cwq->wq->lockdep_map);
+       lock_map_release(&cwq->wq->lockdep_map);
+
+       prev = NULL;
+       spin_lock_irq(&cwq->lock);
+       if (!list_empty(&work->entry)) {
+               /*
+                * See the comment near try_to_grab_pending()->smp_rmb().
+                * If it was re-queued under us we are not going to wait.
+                */
+               smp_rmb();
+               if (unlikely(cwq != get_wq_data(work)))
+                       goto out;
+               prev = &work->entry;
+       } else {
+               if (cwq->current_work != work)
+                       goto out;
+               prev = &cwq->worklist;
+       }
+       insert_wq_barrier(cwq, &barr, prev->next);
+out:
+       spin_unlock_irq(&cwq->lock);
+       if (!prev)
+               return 0;
+
+       wait_for_completion(&barr.done);
+       destroy_work_on_stack(&barr.work);
+       return 1;
+}
+EXPORT_SYMBOL_GPL(flush_work);
+
 /*
- * Upon a successful return, the caller "owns" WORK_STRUCT_PENDING bit,
+ * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
  * so this work can't be re-armed in any way.
  */
 static int try_to_grab_pending(struct work_struct *work)
 {
        struct cpu_workqueue_struct *cwq;
-       int ret = 0;
+       int ret = -1;
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
-               return 1;
+               return 0;
 
        /*
         * The queueing is in progress, or it is already queued. Try to
@@ -418,6 +617,7 @@ static int try_to_grab_pending(struct work_struct *work)
                 */
                smp_rmb();
                if (cwq == get_wq_data(work)) {
+                       debug_work_deactivate(work);
                        list_del_init(&work->entry);
                        ret = 1;
                }
@@ -435,24 +635,29 @@ static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
 
        spin_lock_irq(&cwq->lock);
        if (unlikely(cwq->current_work == work)) {
-               insert_wq_barrier(cwq, &barr, 0);
+               insert_wq_barrier(cwq, &barr, cwq->worklist.next);
                running = 1;
        }
        spin_unlock_irq(&cwq->lock);
 
-       if (unlikely(running))
+       if (unlikely(running)) {
                wait_for_completion(&barr.done);
+               destroy_work_on_stack(&barr.work);
+       }
 }
 
 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_map_acquire(&work->lockdep_map);
+       lock_map_release(&work->lockdep_map);
+
        cwq = get_wq_data(work);
        if (!cwq)
                return;
@@ -460,14 +665,32 @@ static void wait_on_work(struct work_struct *work)
        wq = cwq->wq;
        cpu_map = wq_cpu_map(wq);
 
-       for_each_cpu_mask(cpu, *cpu_map)
+       for_each_cpu(cpu, cpu_map)
                wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
 }
 
+static int __cancel_work_timer(struct work_struct *work,
+                               struct timer_list* timer)
+{
+       int ret;
+
+       do {
+               ret = (timer && likely(del_timer(timer)));
+               if (!ret)
+                       ret = try_to_grab_pending(work);
+               wait_on_work(work);
+       } while (unlikely(ret < 0));
+
+       clear_wq_data(work);
+       return ret;
+}
+
 /**
  * cancel_work_sync - block until a work_struct's callback has terminated
  * @work: the work which is to be flushed
  *
+ * Returns true if @work was pending.
+ *
  * cancel_work_sync() will cancel the work if it is queued. If the work's
  * callback appears to be running, cancel_work_sync() will block until it
  * has completed.
@@ -483,31 +706,26 @@ static void wait_on_work(struct work_struct *work)
  * The caller must ensure that workqueue_struct on which this work was last
  * queued can't be destroyed before this function returns.
  */
-void cancel_work_sync(struct work_struct *work)
+int cancel_work_sync(struct work_struct *work)
 {
-       while (!try_to_grab_pending(work))
-               cpu_relax();
-       wait_on_work(work);
-       work_clear_pending(work);
+       return __cancel_work_timer(work, NULL);
 }
 EXPORT_SYMBOL_GPL(cancel_work_sync);
 
 /**
- * cancel_rearming_delayed_work - reliably kill off a delayed work.
+ * cancel_delayed_work_sync - reliably kill off a delayed work.
  * @dwork: the delayed work struct
  *
+ * Returns true if @dwork was pending.
+ *
  * It is possible to use this function if @dwork rearms itself via queue_work()
  * or queue_delayed_work(). See also the comment for cancel_work_sync().
  */
-void cancel_rearming_delayed_work(struct delayed_work *dwork)
+int cancel_delayed_work_sync(struct delayed_work *dwork)
 {
-       while (!del_timer(&dwork->timer) &&
-              !try_to_grab_pending(&dwork->work))
-               cpu_relax();
-       wait_on_work(&dwork->work);
-       work_clear_pending(&dwork->work);
+       return __cancel_work_timer(&dwork->work, &dwork->timer);
 }
-EXPORT_SYMBOL(cancel_rearming_delayed_work);
+EXPORT_SYMBOL(cancel_delayed_work_sync);
 
 static struct workqueue_struct *keventd_wq __read_mostly;
 
@@ -515,14 +733,32 @@ static struct workqueue_struct *keventd_wq __read_mostly;
  * schedule_work - put work task in global workqueue
  * @work: job to be done
  *
- * This puts a job in the kernel-global workqueue.
+ * Returns zero if @work was already on the kernel-global workqueue and
+ * non-zero otherwise.
+ *
+ * This puts a job in the kernel-global workqueue if it was not already
+ * queued and leaves it in the same position on the kernel-global
+ * workqueue otherwise.
  */
-int fastcall schedule_work(struct work_struct *work)
+int schedule_work(struct work_struct *work)
 {
        return queue_work(keventd_wq, work);
 }
 EXPORT_SYMBOL(schedule_work);
 
+/*
+ * schedule_work_on - put work task on a specific cpu
+ * @cpu: cpu to put the work task on
+ * @work: job to be done
+ *
+ * This puts a job on a specific cpu
+ */
+int schedule_work_on(int cpu, struct work_struct *work)
+{
+       return queue_work_on(cpu, keventd_wq, work);
+}
+EXPORT_SYMBOL(schedule_work_on);
+
 /**
  * schedule_delayed_work - put work task in global workqueue after delay
  * @dwork: job to be done
@@ -531,15 +767,32 @@ EXPORT_SYMBOL(schedule_work);
  * After waiting for a given time this puts a job in the kernel-global
  * workqueue.
  */
-int fastcall schedule_delayed_work(struct delayed_work *dwork,
+int schedule_delayed_work(struct delayed_work *dwork,
                                        unsigned long delay)
 {
-       timer_stats_timer_set_start_info(&dwork->timer);
        return queue_delayed_work(keventd_wq, dwork, delay);
 }
 EXPORT_SYMBOL(schedule_delayed_work);
 
 /**
+ * flush_delayed_work - block until a dwork_struct's callback has terminated
+ * @dwork: the delayed work which is to be flushed
+ *
+ * Any timeout is cancelled, and any pending work is run immediately.
+ */
+void flush_delayed_work(struct delayed_work *dwork)
+{
+       if (del_timer_sync(&dwork->timer)) {
+               struct cpu_workqueue_struct *cwq;
+               cwq = wq_per_cpu(get_wq_data(&dwork->work)->wq, get_cpu());
+               __queue_work(cwq, &dwork->work);
+               put_cpu();
+       }
+       flush_work(&dwork->work);
+}
+EXPORT_SYMBOL(flush_delayed_work);
+
+/**
  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
  * @cpu: cpu to use
  * @dwork: job to be done
@@ -562,33 +815,70 @@ EXPORT_SYMBOL(schedule_delayed_work_on);
  * Returns zero on success.
  * Returns -ve errno on failure.
  *
- * Appears to be racy against CPU hotplug.
- *
  * schedule_on_each_cpu() is very slow.
  */
 int schedule_on_each_cpu(work_func_t func)
 {
        int cpu;
+       int orig = -1;
        struct work_struct *works;
 
        works = alloc_percpu(struct work_struct);
        if (!works)
                return -ENOMEM;
 
-       preempt_disable();              /* CPU hotplug */
+       get_online_cpus();
+
+       /*
+        * When running in keventd don't schedule a work item on
+        * itself.  Can just call directly because the work queue is
+        * already bound.  This also is faster.
+        */
+       if (current_is_keventd())
+               orig = raw_smp_processor_id();
+
        for_each_online_cpu(cpu) {
                struct work_struct *work = per_cpu_ptr(works, cpu);
 
                INIT_WORK(work, func);
-               set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
-               __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
+               if (cpu != orig)
+                       schedule_work_on(cpu, work);
        }
-       preempt_enable();
-       flush_workqueue(keventd_wq);
+       if (orig >= 0)
+               func(per_cpu_ptr(works, orig));
+
+       for_each_online_cpu(cpu)
+               flush_work(per_cpu_ptr(works, cpu));
+
+       put_online_cpus();
        free_percpu(works);
        return 0;
 }
 
+/**
+ * flush_scheduled_work - ensure that any scheduled work has run to completion.
+ *
+ * Forces execution of the kernel-global workqueue and blocks until its
+ * completion.
+ *
+ * Think twice before calling this function!  It's very easy to get into
+ * trouble if you don't take great care.  Either of the following situations
+ * will lead to deadlock:
+ *
+ *     One of the work items currently on the workqueue needs to acquire
+ *     a lock held by your code or its caller.
+ *
+ *     Your code is running in the context of a work routine.
+ *
+ * They will be detected by lockdep when they occur, but the first might not
+ * occur very often.  It depends on what work items are on the workqueue and
+ * what locks they need, which you have no control over.
+ *
+ * In most situations flushing the entire workqueue is overkill; you merely
+ * need to know that a particular work item isn't queued and isn't running.
+ * In such cases you should use cancel_delayed_work_sync() or
+ * cancel_work_sync() instead.
+ */
 void flush_scheduled_work(void)
 {
        flush_workqueue(keventd_wq);
@@ -629,7 +919,7 @@ int keventd_up(void)
 int current_is_keventd(void)
 {
        struct cpu_workqueue_struct *cwq;
-       int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
+       int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
        int ret = 0;
 
        BUG_ON(!keventd_wq);
@@ -657,8 +947,9 @@ 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);
@@ -672,9 +963,11 @@ 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;
-       cwq->should_stop = 0;
+
+       trace_workqueue_creation(cwq->thread, cpu);
 
        return 0;
 }
@@ -690,8 +983,12 @@ static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
        }
 }
 
-struct workqueue_struct *__create_workqueue(const char *name,
-                                           int singlethread, int freezeable)
+struct workqueue_struct *__create_workqueue_key(const char *name,
+                                               int singlethread,
+                                               int freezeable,
+                                               int rt,
+                                               struct lock_class_key *key,
+                                               const char *lock_name)
 {
        struct workqueue_struct *wq;
        struct cpu_workqueue_struct *cwq;
@@ -708,8 +1005,10 @@ struct workqueue_struct *__create_workqueue(const char *name,
        }
 
        wq->name = 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) {
@@ -717,9 +1016,22 @@ struct workqueue_struct *__create_workqueue(const char *name,
                err = create_workqueue_thread(cwq, singlethread_cpu);
                start_workqueue_thread(cwq, -1);
        } else {
-               mutex_lock(&workqueue_mutex);
+               cpu_maps_update_begin();
+               /*
+                * We must place this wq on list even if the code below fails.
+                * cpu_down(cpu) can remove cpu from cpu_populated_map before
+                * destroy_workqueue() takes the lock, in that case we leak
+                * cwq[cpu]->thread.
+                */
+               spin_lock(&workqueue_lock);
                list_add(&wq->list, &workqueues);
-
+               spin_unlock(&workqueue_lock);
+               /*
+                * We must initialize cwqs for each possible cpu even if we
+                * are going to call destroy_workqueue() finally. Otherwise
+                * cpu_up() can hit the uninitialized cwq once we drop the
+                * lock.
+                */
                for_each_possible_cpu(cpu) {
                        cwq = init_cpu_workqueue(wq, cpu);
                        if (err || !cpu_online(cpu))
@@ -727,7 +1039,7 @@ struct workqueue_struct *__create_workqueue(const char *name,
                        err = create_workqueue_thread(cwq, cpu);
                        start_workqueue_thread(cwq, cpu);
                }
-               mutex_unlock(&workqueue_mutex);
+               cpu_maps_update_done();
        }
 
        if (err) {
@@ -736,33 +1048,34 @@ struct workqueue_struct *__create_workqueue(const char *name,
        }
        return wq;
 }
-EXPORT_SYMBOL_GPL(__create_workqueue);
+EXPORT_SYMBOL_GPL(__create_workqueue_key);
 
-static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
+static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
 {
-       struct wq_barrier barr;
-       int alive = 0;
-
-       spin_lock_irq(&cwq->lock);
-       if (cwq->thread != NULL) {
-               insert_wq_barrier(cwq, &barr, 1);
-               cwq->should_stop = 1;
-               alive = 1;
-       }
-       spin_unlock_irq(&cwq->lock);
+       /*
+        * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
+        * cpu_add_remove_lock protects cwq->thread.
+        */
+       if (cwq->thread == NULL)
+               return;
 
-       if (alive) {
-               wait_for_completion(&barr.done);
+       lock_map_acquire(&cwq->wq->lockdep_map);
+       lock_map_release(&cwq->wq->lockdep_map);
 
-               while (unlikely(cwq->thread != NULL))
-                       cpu_relax();
-               /*
-                * Wait until cwq->thread unlocks cwq->lock,
-                * it won't touch *cwq after that.
-                */
-               smp_rmb();
-               spin_unlock_wait(&cwq->lock);
-       }
+       flush_cpu_workqueue(cwq);
+       /*
+        * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
+        * a concurrent flush_workqueue() can insert a barrier after us.
+        * However, in that case run_workqueue() won't return and check
+        * kthread_should_stop() until it flushes all work_struct's.
+        * When ->worklist becomes empty it is safe to exit because no
+        * more work_structs can be queued on this cwq: flush_workqueue
+        * checks list_empty(), and a "normal" queue_work() can't use
+        * a dead CPU.
+        */
+       trace_workqueue_destruction(cwq->thread);
+       kthread_stop(cwq->thread);
+       cwq->thread = NULL;
 }
 
 /**
@@ -773,18 +1086,17 @@ static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
  */
 void destroy_workqueue(struct workqueue_struct *wq)
 {
-       const cpumask_t *cpu_map = wq_cpu_map(wq);
-       struct cpu_workqueue_struct *cwq;
+       const struct cpumask *cpu_map = wq_cpu_map(wq);
        int cpu;
 
-       mutex_lock(&workqueue_mutex);
+       cpu_maps_update_begin();
+       spin_lock(&workqueue_lock);
        list_del(&wq->list);
-       mutex_unlock(&workqueue_mutex);
+       spin_unlock(&workqueue_lock);
 
-       for_each_cpu_mask(cpu, *cpu_map) {
-               cwq = per_cpu_ptr(wq->cpu_wq, cpu);
-               cleanup_workqueue_thread(cwq, cpu);
-       }
+       for_each_cpu(cpu, cpu_map)
+               cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
+       cpu_maps_update_done();
 
        free_percpu(wq->cpu_wq);
        kfree(wq);
@@ -798,22 +1110,15 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
        unsigned int cpu = (unsigned long)hcpu;
        struct cpu_workqueue_struct *cwq;
        struct workqueue_struct *wq;
+       int ret = NOTIFY_OK;
 
        action &= ~CPU_TASKS_FROZEN;
 
        switch (action) {
-       case CPU_LOCK_ACQUIRE:
-               mutex_lock(&workqueue_mutex);
-               return NOTIFY_OK;
-
-       case CPU_LOCK_RELEASE:
-               mutex_unlock(&workqueue_mutex);
-               return NOTIFY_OK;
-
        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) {
                cwq = per_cpu_ptr(wq->cpu_wq, cpu);
 
@@ -821,8 +1126,11 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
                case CPU_UP_PREPARE:
                        if (!create_workqueue_thread(cwq, cpu))
                                break;
-                       printk(KERN_ERR "workqueue for %i failed\n", cpu);
-                       return NOTIFY_BAD;
+                       printk(KERN_ERR "workqueue [%s] for %i failed\n",
+                               wq->name, cpu);
+                       action = CPU_UP_CANCELED;
+                       ret = NOTIFY_BAD;
+                       goto undo;
 
                case CPU_ONLINE:
                        start_workqueue_thread(cwq, cpu);
@@ -830,20 +1138,75 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 
                case CPU_UP_CANCELED:
                        start_workqueue_thread(cwq, -1);
-               case CPU_DEAD:
-                       cleanup_workqueue_thread(cwq, cpu);
+               case CPU_POST_DEAD:
+                       cleanup_workqueue_thread(cwq);
                        break;
                }
        }
 
-       return NOTIFY_OK;
+       switch (action) {
+       case CPU_UP_CANCELED:
+       case CPU_POST_DEAD:
+               cpumask_clear_cpu(cpu, cpu_populated_map);
+       }
+
+       return ret;
+}
+
+#ifdef CONFIG_SMP
+
+struct work_for_cpu {
+       struct completion completion;
+       long (*fn)(void *);
+       void *arg;
+       long ret;
+};
+
+static int do_work_for_cpu(void *_wfc)
+{
+       struct work_for_cpu *wfc = _wfc;
+       wfc->ret = wfc->fn(wfc->arg);
+       complete(&wfc->completion);
+       return 0;
+}
+
+/**
+ * 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 the value @fn returns.
+ * It is up to the caller to ensure that the cpu doesn't go offline.
+ * The caller must not hold any locks which would prevent @fn from completing.
+ */
+long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
+{
+       struct task_struct *sub_thread;
+       struct work_for_cpu wfc = {
+               .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
+               .fn = fn,
+               .arg = arg,
+       };
+
+       sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
+       if (IS_ERR(sub_thread))
+               return PTR_ERR(sub_thread);
+       kthread_bind(sub_thread, cpu);
+       wake_up_process(sub_thread);
+       wait_for_completion(&wfc.completion);
+       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);