workqueue: don't migrate pending works from the dead CPU
authorOleg Nesterov <oleg@tv-sign.ru>
Wed, 9 May 2007 09:34:09 +0000 (02:34 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 9 May 2007 19:30:52 +0000 (12:30 -0700)
Currently CPU_DEAD uses kthread_stop() to stop cwq->thread and then
transfers cwq->worklist to another CPU.  However, it is very unlikely that
worker_thread() will notice kthread_should_stop() before flushing
cwq->worklist.  It is only possible if worker_thread() was preempted after
run_workqueue(cwq), a new work_struct was added, and CPU_DEAD happened
before cwq->thread has a chance to run.

This means that take_over_work() mostly adds unneeded complications.  Note
also that kthread_stop() is not good per se, wake_up_process() may confuse
work->func() if it sleeps waiting for some event.

Remove take_over_work() and migrate_sequence complications.  CPU_DEAD sets
the cwq->should_stop flag (introduced by this patch) and waits for
cwq->thread to flush cwq->worklist and exit.  Because the dead CPU is not
on cpu_online_map, no more works can be added to that cwq.

cpu_populated_map was introduced to optimize for_each_possible_cpu(), it is
not strictly needed, and it is more a documentation in fact.

Saves 418 bytes.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Cc: "Pallipadi, Venkatesh" <venkatesh.pallipadi@intel.com>
Cc: Gautham shenoy <ego@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kernel/workqueue.c

index 398c34f..a981add 100644 (file)
@@ -43,10 +43,11 @@ struct cpu_workqueue_struct {
 
        struct list_head worklist;
        wait_queue_head_t more_work;
+       struct work_struct *current_work;
 
        struct workqueue_struct *wq;
        struct task_struct *thread;
-       struct work_struct *current_work;
+       int should_stop;
 
        int run_depth;          /* Detect run_workqueue() recursion depth */
 } ____cacheline_aligned;
@@ -64,11 +65,12 @@ struct workqueue_struct {
 
 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
    threads to each one as cpus come/go. */
-static long migrate_sequence __read_mostly;
 static DEFINE_MUTEX(workqueue_mutex);
 static LIST_HEAD(workqueues);
 
-static int singlethread_cpu;
+static int singlethread_cpu __read_mostly;
+/* optimization, we could use cpu_possible_map */
+static cpumask_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)
@@ -344,10 +346,28 @@ static void run_workqueue(struct cpu_workqueue_struct *cwq)
        spin_unlock_irqrestore(&cwq->lock, flags);
 }
 
+/*
+ * 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;
-       DECLARE_WAITQUEUE(wait, current);
+       DEFINE_WAIT(wait);
        struct k_sigaction sa;
        sigset_t blocked;
 
@@ -373,23 +393,21 @@ static int worker_thread(void *__cwq)
        siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
        do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
 
-       set_current_state(TASK_INTERRUPTIBLE);
-       while (!kthread_should_stop()) {
+       for (;;) {
                if (cwq->wq->freezeable)
                        try_to_freeze();
 
-               add_wait_queue(&cwq->more_work, &wait);
-               if (list_empty(&cwq->worklist))
+               prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
+               if (!cwq->should_stop && list_empty(&cwq->worklist))
                        schedule();
-               else
-                       __set_current_state(TASK_RUNNING);
-               remove_wait_queue(&cwq->more_work, &wait);
+               finish_wait(&cwq->more_work, &wait);
+
+               if (cwq_should_stop(cwq))
+                       break;
 
-               if (!list_empty(&cwq->worklist))
-                       run_workqueue(cwq);
-               set_current_state(TASK_INTERRUPTIBLE);
+               run_workqueue(cwq);
        }
-       __set_current_state(TASK_RUNNING);
+
        return 0;
 }
 
@@ -454,20 +472,13 @@ static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
  */
 void fastcall flush_workqueue(struct workqueue_struct *wq)
 {
-       if (is_single_threaded(wq)) {
-               /* Always use first cpu's area. */
+       if (is_single_threaded(wq))
                flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
-       } else {
-               long sequence;
+       else {
                int cpu;
-again:
-               sequence = migrate_sequence;
 
-               for_each_possible_cpu(cpu)
+               for_each_cpu_mask(cpu, cpu_populated_map)
                        flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
-
-               if (unlikely(sequence != migrate_sequence))
-                       goto again;
        }
 }
 EXPORT_SYMBOL_GPL(flush_workqueue);
@@ -485,11 +496,8 @@ static void wait_on_work(struct cpu_workqueue_struct *cwq,
        }
        spin_unlock_irq(&cwq->lock);
 
-       if (unlikely(running)) {
-               mutex_unlock(&workqueue_mutex);
+       if (unlikely(running))
                wait_for_completion(&barr.done);
-               mutex_lock(&workqueue_mutex);
-       }
 }
 
 /**
@@ -510,155 +518,31 @@ void flush_work(struct workqueue_struct *wq, struct work_struct *work)
 {
        struct cpu_workqueue_struct *cwq;
 
-       mutex_lock(&workqueue_mutex);
        cwq = get_wq_data(work);
        /* Was it ever queued ? */
        if (!cwq)
-               goto out;
+               return;
 
        /*
-        * This work can't be re-queued, and the lock above protects us
-        * from take_over_work(), no need to re-check that get_wq_data()
-        * is still the same when we take cwq->lock.
+        * This work can't be re-queued, no need to re-check that
+        * get_wq_data() is still the same when we take cwq->lock.
         */
        spin_lock_irq(&cwq->lock);
        list_del_init(&work->entry);
        work_release(work);
        spin_unlock_irq(&cwq->lock);
 
-       if (is_single_threaded(wq)) {
-               /* Always use first cpu's area. */
+       if (is_single_threaded(wq))
                wait_on_work(per_cpu_ptr(wq->cpu_wq, singlethread_cpu), work);
-       else {
+       else {
                int cpu;
 
-               for_each_online_cpu(cpu)
+               for_each_cpu_mask(cpu, cpu_populated_map)
                        wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
        }
-out:
-       mutex_unlock(&workqueue_mutex);
 }
 EXPORT_SYMBOL_GPL(flush_work);
 
-static void init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
-{
-       struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
-
-       cwq->wq = wq;
-       spin_lock_init(&cwq->lock);
-       INIT_LIST_HEAD(&cwq->worklist);
-       init_waitqueue_head(&cwq->more_work);
-}
-
-static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
-                                                       int cpu)
-{
-       struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
-       struct task_struct *p;
-
-       if (is_single_threaded(wq))
-               p = kthread_create(worker_thread, cwq, "%s", wq->name);
-       else
-               p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
-       if (IS_ERR(p))
-               return NULL;
-       cwq->thread = p;
-       return p;
-}
-
-struct workqueue_struct *__create_workqueue(const char *name,
-                                           int singlethread, int freezeable)
-{
-       int cpu, destroy = 0;
-       struct workqueue_struct *wq;
-       struct task_struct *p;
-
-       wq = kzalloc(sizeof(*wq), GFP_KERNEL);
-       if (!wq)
-               return NULL;
-
-       wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
-       if (!wq->cpu_wq) {
-               kfree(wq);
-               return NULL;
-       }
-
-       wq->name = name;
-       wq->freezeable = freezeable;
-
-       mutex_lock(&workqueue_mutex);
-       if (singlethread) {
-               INIT_LIST_HEAD(&wq->list);
-               init_cpu_workqueue(wq, singlethread_cpu);
-               p = create_workqueue_thread(wq, singlethread_cpu);
-               if (!p)
-                       destroy = 1;
-               else
-                       wake_up_process(p);
-       } else {
-               list_add(&wq->list, &workqueues);
-               for_each_possible_cpu(cpu) {
-                       init_cpu_workqueue(wq, cpu);
-                       if (!cpu_online(cpu))
-                               continue;
-
-                       p = create_workqueue_thread(wq, cpu);
-                       if (p) {
-                               kthread_bind(p, cpu);
-                               wake_up_process(p);
-                       } else
-                               destroy = 1;
-               }
-       }
-       mutex_unlock(&workqueue_mutex);
-
-       /*
-        * Was there any error during startup? If yes then clean up:
-        */
-       if (destroy) {
-               destroy_workqueue(wq);
-               wq = NULL;
-       }
-       return wq;
-}
-EXPORT_SYMBOL_GPL(__create_workqueue);
-
-static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
-{
-       struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
-
-       if (cwq->thread) {
-               kthread_stop(cwq->thread);
-               cwq->thread = NULL;
-       }
-}
-
-/**
- * destroy_workqueue - safely terminate a workqueue
- * @wq: target workqueue
- *
- * Safely destroy a workqueue. All work currently pending will be done first.
- */
-void destroy_workqueue(struct workqueue_struct *wq)
-{
-       int cpu;
-
-       flush_workqueue(wq);
-
-       /* We don't need the distraction of CPUs appearing and vanishing. */
-       mutex_lock(&workqueue_mutex);
-       if (is_single_threaded(wq))
-               cleanup_workqueue_thread(wq, singlethread_cpu);
-       else {
-               for_each_online_cpu(cpu)
-                       cleanup_workqueue_thread(wq, cpu);
-               list_del(&wq->list);
-       }
-       mutex_unlock(&workqueue_mutex);
-       free_percpu(wq->cpu_wq);
-       kfree(wq);
-}
-EXPORT_SYMBOL_GPL(destroy_workqueue);
 
 static struct workqueue_struct *keventd_wq;
 
@@ -822,85 +706,193 @@ int current_is_keventd(void)
 
 }
 
-/* Take the work from this (downed) CPU. */
-static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
+static struct cpu_workqueue_struct *
+init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
 {
        struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
-       struct list_head list;
-       struct work_struct *work;
 
-       spin_lock_irq(&cwq->lock);
-       list_replace_init(&cwq->worklist, &list);
-       migrate_sequence++;
-
-       while (!list_empty(&list)) {
-               printk("Taking work for %s\n", wq->name);
-               work = list_entry(list.next,struct work_struct,entry);
-               list_del(&work->entry);
-               __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
-       }
-       spin_unlock_irq(&cwq->lock);
+       cwq->wq = wq;
+       spin_lock_init(&cwq->lock);
+       INIT_LIST_HEAD(&cwq->worklist);
+       init_waitqueue_head(&cwq->more_work);
+
+       return cwq;
 }
 
-/* We're holding the cpucontrol mutex here */
-static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
-                                 unsigned long action,
-                                 void *hcpu)
+static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
+{
+       struct workqueue_struct *wq = cwq->wq;
+       const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
+       struct task_struct *p;
+
+       p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
+       /*
+        * Nobody can add the work_struct to this cwq,
+        *      if (caller is __create_workqueue)
+        *              nobody should see this wq
+        *      else // caller is CPU_UP_PREPARE
+        *              cpu is not on cpu_online_map
+        * so we can abort safely.
+        */
+       if (IS_ERR(p))
+               return PTR_ERR(p);
+
+       cwq->thread = p;
+       cwq->should_stop = 0;
+       if (!is_single_threaded(wq))
+               kthread_bind(p, cpu);
+
+       if (is_single_threaded(wq) || cpu_online(cpu))
+               wake_up_process(p);
+
+       return 0;
+}
+
+struct workqueue_struct *__create_workqueue(const char *name,
+                                           int singlethread, int freezeable)
 {
-       unsigned int hotcpu = (unsigned long)hcpu;
        struct workqueue_struct *wq;
+       struct cpu_workqueue_struct *cwq;
+       int err = 0, cpu;
 
-       switch (action) {
-       case CPU_UP_PREPARE:
+       wq = kzalloc(sizeof(*wq), GFP_KERNEL);
+       if (!wq)
+               return NULL;
+
+       wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
+       if (!wq->cpu_wq) {
+               kfree(wq);
+               return NULL;
+       }
+
+       wq->name = name;
+       wq->freezeable = freezeable;
+
+       if (singlethread) {
+               INIT_LIST_HEAD(&wq->list);
+               cwq = init_cpu_workqueue(wq, singlethread_cpu);
+               err = create_workqueue_thread(cwq, singlethread_cpu);
+       } else {
                mutex_lock(&workqueue_mutex);
-               /* Create a new workqueue thread for it. */
-               list_for_each_entry(wq, &workqueues, list) {
-                       if (!create_workqueue_thread(wq, hotcpu)) {
-                               printk("workqueue for %i failed\n", hotcpu);
-                               return NOTIFY_BAD;
-                       }
+               list_add(&wq->list, &workqueues);
+
+               for_each_possible_cpu(cpu) {
+                       cwq = init_cpu_workqueue(wq, cpu);
+                       if (err || !cpu_online(cpu))
+                               continue;
+                       err = create_workqueue_thread(cwq, cpu);
                }
-               break;
+               mutex_unlock(&workqueue_mutex);
+       }
+
+       if (err) {
+               destroy_workqueue(wq);
+               wq = NULL;
+       }
+       return wq;
+}
+EXPORT_SYMBOL_GPL(__create_workqueue);
 
-       case CPU_ONLINE:
-               /* Kick off worker threads. */
-               list_for_each_entry(wq, &workqueues, list) {
-                       struct cpu_workqueue_struct *cwq;
+static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
+{
+       struct wq_barrier barr;
+       int alive = 0;
 
-                       cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
-                       kthread_bind(cwq->thread, hotcpu);
-                       wake_up_process(cwq->thread);
-               }
+       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);
+
+       if (alive) {
+               wait_for_completion(&barr.done);
+
+               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);
+       }
+}
+
+/**
+ * destroy_workqueue - safely terminate a workqueue
+ * @wq: target workqueue
+ *
+ * Safely destroy a workqueue. All work currently pending will be done first.
+ */
+void destroy_workqueue(struct workqueue_struct *wq)
+{
+       struct cpu_workqueue_struct *cwq;
+
+       if (is_single_threaded(wq)) {
+               cwq = per_cpu_ptr(wq->cpu_wq, singlethread_cpu);
+               cleanup_workqueue_thread(cwq, singlethread_cpu);
+       } else {
+               int cpu;
+
+               mutex_lock(&workqueue_mutex);
+               list_del(&wq->list);
                mutex_unlock(&workqueue_mutex);
-               break;
 
-       case CPU_UP_CANCELED:
-               list_for_each_entry(wq, &workqueues, list) {
-                       if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
-                               continue;
-                       /* Unbind so it can run. */
-                       kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
-                                    any_online_cpu(cpu_online_map));
-                       cleanup_workqueue_thread(wq, hotcpu);
+               for_each_cpu_mask(cpu, cpu_populated_map) {
+                       cwq = per_cpu_ptr(wq->cpu_wq, cpu);
+                       cleanup_workqueue_thread(cwq, cpu);
                }
-               mutex_unlock(&workqueue_mutex);
-               break;
+       }
 
-       case CPU_DOWN_PREPARE:
+       free_percpu(wq->cpu_wq);
+       kfree(wq);
+}
+EXPORT_SYMBOL_GPL(destroy_workqueue);
+
+static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
+                                               unsigned long action,
+                                               void *hcpu)
+{
+       unsigned int cpu = (unsigned long)hcpu;
+       struct cpu_workqueue_struct *cwq;
+       struct workqueue_struct *wq;
+
+       switch (action) {
+       case CPU_LOCK_ACQUIRE:
                mutex_lock(&workqueue_mutex);
-               break;
+               return NOTIFY_OK;
 
-       case CPU_DOWN_FAILED:
+       case CPU_LOCK_RELEASE:
                mutex_unlock(&workqueue_mutex);
-               break;
+               return NOTIFY_OK;
 
-       case CPU_DEAD:
-               list_for_each_entry(wq, &workqueues, list)
-                       cleanup_workqueue_thread(wq, hotcpu);
-               list_for_each_entry(wq, &workqueues, list)
-                       take_over_work(wq, hotcpu);
-               mutex_unlock(&workqueue_mutex);
-               break;
+       case CPU_UP_PREPARE:
+               cpu_set(cpu, cpu_populated_map);
+       }
+
+       list_for_each_entry(wq, &workqueues, list) {
+               cwq = per_cpu_ptr(wq->cpu_wq, cpu);
+
+               switch (action) {
+               case CPU_UP_PREPARE:
+                       if (!create_workqueue_thread(cwq, cpu))
+                               break;
+                       printk(KERN_ERR "workqueue for %i failed\n", cpu);
+                       return NOTIFY_BAD;
+
+               case CPU_ONLINE:
+                       wake_up_process(cwq->thread);
+                       break;
+
+               case CPU_UP_CANCELED:
+                       if (cwq->thread)
+                               wake_up_process(cwq->thread);
+               case CPU_DEAD:
+                       cleanup_workqueue_thread(cwq, cpu);
+                       break;
+               }
        }
 
        return NOTIFY_OK;
@@ -908,9 +900,9 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 
 void init_workqueues(void)
 {
+       cpu_populated_map = cpu_online_map;
        singlethread_cpu = first_cpu(cpu_possible_map);
        hotcpu_notifier(workqueue_cpu_callback, 0);
        keventd_wq = create_workqueue("events");
        BUG_ON(!keventd_wq);
 }
-