rcu: Permit call_rcu() from CPU_DYING notifiers
[linux-2.6.git] / kernel / padata.c
index 82958e0..89fe3d1 100644 (file)
@@ -1,6 +1,8 @@
 /*
  * padata.c - generic interface to process data streams in parallel
  *
+ * See Documentation/padata.txt for an api documentation.
+ *
  * Copyright (C) 2008, 2009 secunet Security Networks AG
  * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
  *
@@ -18,7 +20,7 @@
  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  */
 
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/cpumask.h>
 #include <linux/err.h>
 #include <linux/cpu.h>
 #include <linux/mutex.h>
 #include <linux/sched.h>
 #include <linux/slab.h>
+#include <linux/sysfs.h>
 #include <linux/rcupdate.h>
 
-#define MAX_SEQ_NR INT_MAX - NR_CPUS
 #define MAX_OBJ_NUM 1000
 
 static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
 {
        int cpu, target_cpu;
 
-       target_cpu = cpumask_first(pd->cpumask);
+       target_cpu = cpumask_first(pd->cpumask.pcpu);
        for (cpu = 0; cpu < cpu_index; cpu++)
-               target_cpu = cpumask_next(target_cpu, pd->cpumask);
+               target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
 
        return target_cpu;
 }
 
-static int padata_cpu_hash(struct padata_priv *padata)
+static int padata_cpu_hash(struct parallel_data *pd)
 {
        int cpu_index;
-       struct parallel_data *pd;
-
-       pd =  padata->pd;
 
        /*
         * Hash the sequence numbers to the cpus by taking
         * seq_nr mod. number of cpus in use.
         */
-       cpu_index =  padata->seq_nr % cpumask_weight(pd->cpumask);
+
+       spin_lock(&pd->seq_lock);
+       cpu_index =  pd->seq_nr % cpumask_weight(pd->cpumask.pcpu);
+       pd->seq_nr++;
+       spin_unlock(&pd->seq_lock);
 
        return padata_index_to_cpu(pd, cpu_index);
 }
 
-static void padata_parallel_worker(struct work_struct *work)
+static void padata_parallel_worker(struct work_struct *parallel_work)
 {
-       struct padata_queue *queue;
+       struct padata_parallel_queue *pqueue;
        struct parallel_data *pd;
        struct padata_instance *pinst;
        LIST_HEAD(local_list);
 
        local_bh_disable();
-       queue = container_of(work, struct padata_queue, pwork);
-       pd = queue->pd;
+       pqueue = container_of(parallel_work,
+                             struct padata_parallel_queue, work);
+       pd = pqueue->pd;
        pinst = pd->pinst;
 
-       spin_lock(&queue->parallel.lock);
-       list_replace_init(&queue->parallel.list, &local_list);
-       spin_unlock(&queue->parallel.lock);
+       spin_lock(&pqueue->parallel.lock);
+       list_replace_init(&pqueue->parallel.list, &local_list);
+       spin_unlock(&pqueue->parallel.lock);
 
        while (!list_empty(&local_list)) {
                struct padata_priv *padata;
@@ -88,13 +92,13 @@ static void padata_parallel_worker(struct work_struct *work)
        local_bh_enable();
 }
 
-/*
+/**
  * padata_do_parallel - padata parallelization function
  *
  * @pinst: padata instance
  * @padata: object to be parallelized
  * @cb_cpu: cpu the serialization callback function will run on,
- *          must be in the cpumask of padata.
+ *          must be in the serial cpumask of padata(i.e. cpumask.cbcpu).
  *
  * The parallelization callback function will run with BHs off.
  * Note: Every object which is parallelized by padata_do_parallel
@@ -104,15 +108,18 @@ int padata_do_parallel(struct padata_instance *pinst,
                       struct padata_priv *padata, int cb_cpu)
 {
        int target_cpu, err;
-       struct padata_queue *queue;
+       struct padata_parallel_queue *queue;
        struct parallel_data *pd;
 
        rcu_read_lock_bh();
 
        pd = rcu_dereference(pinst->pd);
 
-       err = 0;
-       if (!(pinst->flags & PADATA_INIT))
+       err = -EINVAL;
+       if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
+               goto out;
+
+       if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
                goto out;
 
        err =  -EBUSY;
@@ -122,28 +129,19 @@ int padata_do_parallel(struct padata_instance *pinst,
        if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
                goto out;
 
-       err = -EINVAL;
-       if (!cpumask_test_cpu(cb_cpu, pd->cpumask))
-               goto out;
-
-       err = -EINPROGRESS;
+       err = 0;
        atomic_inc(&pd->refcnt);
        padata->pd = pd;
        padata->cb_cpu = cb_cpu;
 
-       if (unlikely(atomic_read(&pd->seq_nr) == pd->max_seq_nr))
-               atomic_set(&pd->seq_nr, -1);
-
-       padata->seq_nr = atomic_inc_return(&pd->seq_nr);
-
-       target_cpu = padata_cpu_hash(padata);
-       queue = per_cpu_ptr(pd->queue, target_cpu);
+       target_cpu = padata_cpu_hash(pd);
+       queue = per_cpu_ptr(pd->pqueue, target_cpu);
 
        spin_lock(&queue->parallel.lock);
        list_add_tail(&padata->list, &queue->parallel.list);
        spin_unlock(&queue->parallel.lock);
 
-       queue_work_on(target_cpu, pinst->wq, &queue->pwork);
+       queue_work_on(target_cpu, pinst->wq, &queue->work);
 
 out:
        rcu_read_unlock_bh();
@@ -152,86 +150,62 @@ out:
 }
 EXPORT_SYMBOL(padata_do_parallel);
 
+/*
+ * padata_get_next - Get the next object that needs serialization.
+ *
+ * Return values are:
+ *
+ * A pointer to the control struct of the next object that needs
+ * serialization, if present in one of the percpu reorder queues.
+ *
+ * NULL, if all percpu reorder queues are empty.
+ *
+ * -EINPROGRESS, if the next object that needs serialization will
+ *  be parallel processed by another cpu and is not yet present in
+ *  the cpu's reorder queue.
+ *
+ * -ENODATA, if this cpu has to do the parallel processing for
+ *  the next object.
+ */
 static struct padata_priv *padata_get_next(struct parallel_data *pd)
 {
-       int cpu, num_cpus, empty, calc_seq_nr;
-       int seq_nr, next_nr, overrun, next_overrun;
-       struct padata_queue *queue, *next_queue;
+       int cpu, num_cpus;
+       unsigned int next_nr, next_index;
+       struct padata_parallel_queue *queue, *next_queue;
        struct padata_priv *padata;
        struct padata_list *reorder;
 
-       empty = 0;
-       next_nr = -1;
-       next_overrun = 0;
-       next_queue = NULL;
-
-       num_cpus = cpumask_weight(pd->cpumask);
-
-       for_each_cpu(cpu, pd->cpumask) {
-               queue = per_cpu_ptr(pd->queue, cpu);
-               reorder = &queue->reorder;
-
-               /*
-                * Calculate the seq_nr of the object that should be
-                * next in this queue.
-                */
-               overrun = 0;
-               calc_seq_nr = (atomic_read(&queue->num_obj) * num_cpus)
-                              + queue->cpu_index;
-
-               if (unlikely(calc_seq_nr > pd->max_seq_nr)) {
-                       calc_seq_nr = calc_seq_nr - pd->max_seq_nr - 1;
-                       overrun = 1;
-               }
-
-               if (!list_empty(&reorder->list)) {
-                       padata = list_entry(reorder->list.next,
-                                           struct padata_priv, list);
-
-                       seq_nr  = padata->seq_nr;
-                       BUG_ON(calc_seq_nr != seq_nr);
-               } else {
-                       seq_nr = calc_seq_nr;
-                       empty++;
-               }
+       num_cpus = cpumask_weight(pd->cpumask.pcpu);
 
-               if (next_nr < 0 || seq_nr < next_nr
-                   || (next_overrun && !overrun)) {
-                       next_nr = seq_nr;
-                       next_overrun = overrun;
-                       next_queue = queue;
-               }
-       }
+       /*
+        * Calculate the percpu reorder queue and the sequence
+        * number of the next object.
+        */
+       next_nr = pd->processed;
+       next_index = next_nr % num_cpus;
+       cpu = padata_index_to_cpu(pd, next_index);
+       next_queue = per_cpu_ptr(pd->pqueue, cpu);
 
        padata = NULL;
 
-       if (empty == num_cpus)
-               goto out;
-
        reorder = &next_queue->reorder;
 
        if (!list_empty(&reorder->list)) {
                padata = list_entry(reorder->list.next,
                                    struct padata_priv, list);
 
-               if (unlikely(next_overrun)) {
-                       for_each_cpu(cpu, pd->cpumask) {
-                               queue = per_cpu_ptr(pd->queue, cpu);
-                               atomic_set(&queue->num_obj, 0);
-                       }
-               }
-
                spin_lock(&reorder->lock);
                list_del_init(&padata->list);
                atomic_dec(&pd->reorder_objects);
                spin_unlock(&reorder->lock);
 
-               atomic_inc(&next_queue->num_obj);
+               pd->processed++;
 
                goto out;
        }
 
-       if (next_nr % num_cpus == next_queue->cpu_index) {
+       queue = per_cpu_ptr(pd->pqueue, smp_processor_id());
+       if (queue->cpu_index == next_queue->cpu_index) {
                padata = ERR_PTR(-ENODATA);
                goto out;
        }
@@ -243,56 +217,93 @@ out:
 
 static void padata_reorder(struct parallel_data *pd)
 {
+       int cb_cpu;
        struct padata_priv *padata;
-       struct padata_queue *queue;
+       struct padata_serial_queue *squeue;
        struct padata_instance *pinst = pd->pinst;
 
-try_again:
+       /*
+        * We need to ensure that only one cpu can work on dequeueing of
+        * the reorder queue the time. Calculating in which percpu reorder
+        * queue the next object will arrive takes some time. A spinlock
+        * would be highly contended. Also it is not clear in which order
+        * the objects arrive to the reorder queues. So a cpu could wait to
+        * get the lock just to notice that there is nothing to do at the
+        * moment. Therefore we use a trylock and let the holder of the lock
+        * care for all the objects enqueued during the holdtime of the lock.
+        */
        if (!spin_trylock_bh(&pd->lock))
-               goto out;
+               return;
 
        while (1) {
                padata = padata_get_next(pd);
 
+               /*
+                * All reorder queues are empty, or the next object that needs
+                * serialization is parallel processed by another cpu and is
+                * still on it's way to the cpu's reorder queue, nothing to
+                * do for now.
+                */
                if (!padata || PTR_ERR(padata) == -EINPROGRESS)
                        break;
 
+               /*
+                * This cpu has to do the parallel processing of the next
+                * object. It's waiting in the cpu's parallelization queue,
+                * so exit immediately.
+                */
                if (PTR_ERR(padata) == -ENODATA) {
+                       del_timer(&pd->timer);
                        spin_unlock_bh(&pd->lock);
-                       goto out;
+                       return;
                }
 
-               queue = per_cpu_ptr(pd->queue, padata->cb_cpu);
+               cb_cpu = padata->cb_cpu;
+               squeue = per_cpu_ptr(pd->squeue, cb_cpu);
 
-               spin_lock(&queue->serial.lock);
-               list_add_tail(&padata->list, &queue->serial.list);
-               spin_unlock(&queue->serial.lock);
+               spin_lock(&squeue->serial.lock);
+               list_add_tail(&padata->list, &squeue->serial.list);
+               spin_unlock(&squeue->serial.lock);
 
-               queue_work_on(padata->cb_cpu, pinst->wq, &queue->swork);
+               queue_work_on(cb_cpu, pinst->wq, &squeue->work);
        }
 
        spin_unlock_bh(&pd->lock);
 
-       if (atomic_read(&pd->reorder_objects))
-               goto try_again;
+       /*
+        * The next object that needs serialization might have arrived to
+        * the reorder queues in the meantime, we will be called again
+        * from the timer function if no one else cares for it.
+        */
+       if (atomic_read(&pd->reorder_objects)
+                       && !(pinst->flags & PADATA_RESET))
+               mod_timer(&pd->timer, jiffies + HZ);
+       else
+               del_timer(&pd->timer);
 
-out:
        return;
 }
 
-static void padata_serial_worker(struct work_struct *work)
+static void padata_reorder_timer(unsigned long arg)
 {
-       struct padata_queue *queue;
+       struct parallel_data *pd = (struct parallel_data *)arg;
+
+       padata_reorder(pd);
+}
+
+static void padata_serial_worker(struct work_struct *serial_work)
+{
+       struct padata_serial_queue *squeue;
        struct parallel_data *pd;
        LIST_HEAD(local_list);
 
        local_bh_disable();
-       queue = container_of(work, struct padata_queue, swork);
-       pd = queue->pd;
+       squeue = container_of(serial_work, struct padata_serial_queue, work);
+       pd = squeue->pd;
 
-       spin_lock(&queue->serial.lock);
-       list_replace_init(&queue->serial.list, &local_list);
-       spin_unlock(&queue->serial.lock);
+       spin_lock(&squeue->serial.lock);
+       list_replace_init(&squeue->serial.list, &local_list);
+       spin_unlock(&squeue->serial.lock);
 
        while (!list_empty(&local_list)) {
                struct padata_priv *padata;
@@ -308,7 +319,7 @@ static void padata_serial_worker(struct work_struct *work)
        local_bh_enable();
 }
 
-/*
+/**
  * padata_do_serial - padata serialization function
  *
  * @padata: object to be serialized.
@@ -319,18 +330,18 @@ static void padata_serial_worker(struct work_struct *work)
 void padata_do_serial(struct padata_priv *padata)
 {
        int cpu;
-       struct padata_queue *queue;
+       struct padata_parallel_queue *pqueue;
        struct parallel_data *pd;
 
        pd = padata->pd;
 
        cpu = get_cpu();
-       queue = per_cpu_ptr(pd->queue, cpu);
+       pqueue = per_cpu_ptr(pd->pqueue, cpu);
 
-       spin_lock(&queue->reorder.lock);
+       spin_lock(&pqueue->reorder.lock);
        atomic_inc(&pd->reorder_objects);
-       list_add_tail(&padata->list, &queue->reorder.list);
-       spin_unlock(&queue->reorder.lock);
+       list_add_tail(&padata->list, &pqueue->reorder.list);
+       spin_unlock(&pqueue->reorder.lock);
 
        put_cpu();
 
@@ -338,52 +349,88 @@ void padata_do_serial(struct padata_priv *padata)
 }
 EXPORT_SYMBOL(padata_do_serial);
 
-static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
-                                            const struct cpumask *cpumask)
+static int padata_setup_cpumasks(struct parallel_data *pd,
+                                const struct cpumask *pcpumask,
+                                const struct cpumask *cbcpumask)
 {
-       int cpu, cpu_index, num_cpus;
-       struct padata_queue *queue;
-       struct parallel_data *pd;
+       if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
+               return -ENOMEM;
 
-       cpu_index = 0;
-
-       pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
-       if (!pd)
-               goto err;
+       cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
+       if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
+               free_cpumask_var(pd->cpumask.cbcpu);
+               return -ENOMEM;
+       }
 
-       pd->queue = alloc_percpu(struct padata_queue);
-       if (!pd->queue)
-               goto err_free_pd;
+       cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
+       return 0;
+}
 
-       if (!alloc_cpumask_var(&pd->cpumask, GFP_KERNEL))
-               goto err_free_queue;
+static void __padata_list_init(struct padata_list *pd_list)
+{
+       INIT_LIST_HEAD(&pd_list->list);
+       spin_lock_init(&pd_list->lock);
+}
 
-       cpumask_and(pd->cpumask, cpumask, cpu_active_mask);
+/* Initialize all percpu queues used by serial workers */
+static void padata_init_squeues(struct parallel_data *pd)
+{
+       int cpu;
+       struct padata_serial_queue *squeue;
 
-       for_each_cpu(cpu, pd->cpumask) {
-               queue = per_cpu_ptr(pd->queue, cpu);
+       for_each_cpu(cpu, pd->cpumask.cbcpu) {
+               squeue = per_cpu_ptr(pd->squeue, cpu);
+               squeue->pd = pd;
+               __padata_list_init(&squeue->serial);
+               INIT_WORK(&squeue->work, padata_serial_worker);
+       }
+}
 
-               queue->pd = pd;
+/* Initialize all percpu queues used by parallel workers */
+static void padata_init_pqueues(struct parallel_data *pd)
+{
+       int cpu_index, cpu;
+       struct padata_parallel_queue *pqueue;
 
-               queue->cpu_index = cpu_index;
+       cpu_index = 0;
+       for_each_cpu(cpu, pd->cpumask.pcpu) {
+               pqueue = per_cpu_ptr(pd->pqueue, cpu);
+               pqueue->pd = pd;
+               pqueue->cpu_index = cpu_index;
                cpu_index++;
 
-               INIT_LIST_HEAD(&queue->reorder.list);
-               INIT_LIST_HEAD(&queue->parallel.list);
-               INIT_LIST_HEAD(&queue->serial.list);
-               spin_lock_init(&queue->reorder.lock);
-               spin_lock_init(&queue->parallel.lock);
-               spin_lock_init(&queue->serial.lock);
-
-               INIT_WORK(&queue->pwork, padata_parallel_worker);
-               INIT_WORK(&queue->swork, padata_serial_worker);
-               atomic_set(&queue->num_obj, 0);
+               __padata_list_init(&pqueue->reorder);
+               __padata_list_init(&pqueue->parallel);
+               INIT_WORK(&pqueue->work, padata_parallel_worker);
+               atomic_set(&pqueue->num_obj, 0);
        }
+}
 
-       num_cpus = cpumask_weight(pd->cpumask);
-       pd->max_seq_nr = (MAX_SEQ_NR / num_cpus) * num_cpus - 1;
+/* Allocate and initialize the internal cpumask dependend resources. */
+static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
+                                            const struct cpumask *pcpumask,
+                                            const struct cpumask *cbcpumask)
+{
+       struct parallel_data *pd;
+
+       pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
+       if (!pd)
+               goto err;
 
-       atomic_set(&pd->seq_nr, -1);
+       pd->pqueue = alloc_percpu(struct padata_parallel_queue);
+       if (!pd->pqueue)
+               goto err_free_pd;
+
+       pd->squeue = alloc_percpu(struct padata_serial_queue);
+       if (!pd->squeue)
+               goto err_free_pqueue;
+       if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
+               goto err_free_squeue;
+
+       padata_init_pqueues(pd);
+       padata_init_squeues(pd);
+       setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
+       pd->seq_nr = 0;
        atomic_set(&pd->reorder_objects, 0);
        atomic_set(&pd->refcnt, 0);
        pd->pinst = pinst;
@@ -391,8 +438,10 @@ static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
 
        return pd;
 
-err_free_queue:
-       free_percpu(pd->queue);
+err_free_squeue:
+       free_percpu(pd->squeue);
+err_free_pqueue:
+       free_percpu(pd->pqueue);
 err_free_pd:
        kfree(pd);
 err:
@@ -401,15 +450,63 @@ err:
 
 static void padata_free_pd(struct parallel_data *pd)
 {
-       free_cpumask_var(pd->cpumask);
-       free_percpu(pd->queue);
+       free_cpumask_var(pd->cpumask.pcpu);
+       free_cpumask_var(pd->cpumask.cbcpu);
+       free_percpu(pd->pqueue);
+       free_percpu(pd->squeue);
        kfree(pd);
 }
 
+/* Flush all objects out of the padata queues. */
+static void padata_flush_queues(struct parallel_data *pd)
+{
+       int cpu;
+       struct padata_parallel_queue *pqueue;
+       struct padata_serial_queue *squeue;
+
+       for_each_cpu(cpu, pd->cpumask.pcpu) {
+               pqueue = per_cpu_ptr(pd->pqueue, cpu);
+               flush_work(&pqueue->work);
+       }
+
+       del_timer_sync(&pd->timer);
+
+       if (atomic_read(&pd->reorder_objects))
+               padata_reorder(pd);
+
+       for_each_cpu(cpu, pd->cpumask.cbcpu) {
+               squeue = per_cpu_ptr(pd->squeue, cpu);
+               flush_work(&squeue->work);
+       }
+
+       BUG_ON(atomic_read(&pd->refcnt) != 0);
+}
+
+static void __padata_start(struct padata_instance *pinst)
+{
+       pinst->flags |= PADATA_INIT;
+}
+
+static void __padata_stop(struct padata_instance *pinst)
+{
+       if (!(pinst->flags & PADATA_INIT))
+               return;
+
+       pinst->flags &= ~PADATA_INIT;
+
+       synchronize_rcu();
+
+       get_online_cpus();
+       padata_flush_queues(pinst->pd);
+       put_online_cpus();
+}
+
+/* Replace the internal control structure with a new one. */
 static void padata_replace(struct padata_instance *pinst,
                           struct parallel_data *pd_new)
 {
        struct parallel_data *pd_old = pinst->pd;
+       int notification_mask = 0;
 
        pinst->flags |= PADATA_RESET;
 
@@ -417,45 +514,162 @@ static void padata_replace(struct padata_instance *pinst,
 
        synchronize_rcu();
 
-       while (atomic_read(&pd_old->refcnt) != 0)
-               yield();
-
-       flush_workqueue(pinst->wq);
+       if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
+               notification_mask |= PADATA_CPU_PARALLEL;
+       if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
+               notification_mask |= PADATA_CPU_SERIAL;
 
+       padata_flush_queues(pd_old);
        padata_free_pd(pd_old);
 
+       if (notification_mask)
+               blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
+                                            notification_mask,
+                                            &pd_new->cpumask);
+
        pinst->flags &= ~PADATA_RESET;
 }
 
-/*
- * padata_set_cpumask - set the cpumask that padata should use
+/**
+ * padata_register_cpumask_notifier - Registers a notifier that will be called
+ *                             if either pcpu or cbcpu or both cpumasks change.
  *
- * @pinst: padata instance
- * @cpumask: the cpumask to use
+ * @pinst: A poineter to padata instance
+ * @nblock: A pointer to notifier block.
  */
-int padata_set_cpumask(struct padata_instance *pinst,
-                       cpumask_var_t cpumask)
+int padata_register_cpumask_notifier(struct padata_instance *pinst,
+                                    struct notifier_block *nblock)
 {
+       return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
+                                               nblock);
+}
+EXPORT_SYMBOL(padata_register_cpumask_notifier);
+
+/**
+ * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
+ *        registered earlier  using padata_register_cpumask_notifier
+ *
+ * @pinst: A pointer to data instance.
+ * @nlock: A pointer to notifier block.
+ */
+int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
+                                      struct notifier_block *nblock)
+{
+       return blocking_notifier_chain_unregister(
+               &pinst->cpumask_change_notifier,
+               nblock);
+}
+EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
+
+
+/* If cpumask contains no active cpu, we mark the instance as invalid. */
+static bool padata_validate_cpumask(struct padata_instance *pinst,
+                                   const struct cpumask *cpumask)
+{
+       if (!cpumask_intersects(cpumask, cpu_online_mask)) {
+               pinst->flags |= PADATA_INVALID;
+               return false;
+       }
+
+       pinst->flags &= ~PADATA_INVALID;
+       return true;
+}
+
+static int __padata_set_cpumasks(struct padata_instance *pinst,
+                                cpumask_var_t pcpumask,
+                                cpumask_var_t cbcpumask)
+{
+       int valid;
        struct parallel_data *pd;
-       int err = 0;
+
+       valid = padata_validate_cpumask(pinst, pcpumask);
+       if (!valid) {
+               __padata_stop(pinst);
+               goto out_replace;
+       }
+
+       valid = padata_validate_cpumask(pinst, cbcpumask);
+       if (!valid)
+               __padata_stop(pinst);
+
+out_replace:
+       pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
+       if (!pd)
+               return -ENOMEM;
+
+       cpumask_copy(pinst->cpumask.pcpu, pcpumask);
+       cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
+
+       padata_replace(pinst, pd);
+
+       if (valid)
+               __padata_start(pinst);
+
+       return 0;
+}
+
+/**
+ * padata_set_cpumasks - Set both parallel and serial cpumasks. The first
+ *                       one is used by parallel workers and the second one
+ *                       by the wokers doing serialization.
+ *
+ * @pinst: padata instance
+ * @pcpumask: the cpumask to use for parallel workers
+ * @cbcpumask: the cpumsak to use for serial workers
+ */
+int padata_set_cpumasks(struct padata_instance *pinst, cpumask_var_t pcpumask,
+                       cpumask_var_t cbcpumask)
+{
+       int err;
 
        mutex_lock(&pinst->lock);
+       get_online_cpus();
 
+       err = __padata_set_cpumasks(pinst, pcpumask, cbcpumask);
+
+       put_online_cpus();
+       mutex_unlock(&pinst->lock);
+
+       return err;
+
+}
+EXPORT_SYMBOL(padata_set_cpumasks);
+
+/**
+ * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
+ *                     equivalent to @cpumask.
+ *
+ * @pinst: padata instance
+ * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
+ *                to parallel and serial cpumasks respectively.
+ * @cpumask: the cpumask to use
+ */
+int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
+                      cpumask_var_t cpumask)
+{
+       struct cpumask *serial_mask, *parallel_mask;
+       int err = -EINVAL;
+
+       mutex_lock(&pinst->lock);
        get_online_cpus();
 
-       pd = padata_alloc_pd(pinst, cpumask);
-       if (!pd) {
-               err = -ENOMEM;
-               goto out;
+       switch (cpumask_type) {
+       case PADATA_CPU_PARALLEL:
+               serial_mask = pinst->cpumask.cbcpu;
+               parallel_mask = cpumask;
+               break;
+       case PADATA_CPU_SERIAL:
+               parallel_mask = pinst->cpumask.pcpu;
+               serial_mask = cpumask;
+               break;
+       default:
+                goto out;
        }
 
-       cpumask_copy(pinst->cpumask, cpumask);
-
-       padata_replace(pinst, pd);
+       err =  __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
 
 out:
        put_online_cpus();
-
        mutex_unlock(&pinst->lock);
 
        return err;
@@ -466,31 +680,49 @@ static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
 {
        struct parallel_data *pd;
 
-       if (cpumask_test_cpu(cpu, cpu_active_mask)) {
-               pd = padata_alloc_pd(pinst, pinst->cpumask);
+       if (cpumask_test_cpu(cpu, cpu_online_mask)) {
+               pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
+                                    pinst->cpumask.cbcpu);
                if (!pd)
                        return -ENOMEM;
 
                padata_replace(pinst, pd);
+
+               if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
+                   padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
+                       __padata_start(pinst);
        }
 
        return 0;
 }
 
-/*
- * padata_add_cpu - add a cpu to the padata cpumask
+ /**
+ * padata_add_cpu - add a cpu to one or both(parallel and serial)
+ *                  padata cpumasks.
  *
  * @pinst: padata instance
  * @cpu: cpu to add
+ * @mask: bitmask of flags specifying to which cpumask @cpu shuld be added.
+ *        The @mask may be any combination of the following flags:
+ *          PADATA_CPU_SERIAL   - serial cpumask
+ *          PADATA_CPU_PARALLEL - parallel cpumask
  */
-int padata_add_cpu(struct padata_instance *pinst, int cpu)
+
+int padata_add_cpu(struct padata_instance *pinst, int cpu, int mask)
 {
        int err;
 
+       if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
+               return -EINVAL;
+
        mutex_lock(&pinst->lock);
 
        get_online_cpus();
-       cpumask_set_cpu(cpu, pinst->cpumask);
+       if (mask & PADATA_CPU_SERIAL)
+               cpumask_set_cpu(cpu, pinst->cpumask.cbcpu);
+       if (mask & PADATA_CPU_PARALLEL)
+               cpumask_set_cpu(cpu, pinst->cpumask.pcpu);
+
        err = __padata_add_cpu(pinst, cpu);
        put_online_cpus();
 
@@ -502,33 +734,54 @@ EXPORT_SYMBOL(padata_add_cpu);
 
 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
 {
-       struct parallel_data *pd;
+       struct parallel_data *pd = NULL;
 
        if (cpumask_test_cpu(cpu, cpu_online_mask)) {
-               pd = padata_alloc_pd(pinst, pinst->cpumask);
+
+               if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
+                   !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
+                       __padata_stop(pinst);
+
+               pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
+                                    pinst->cpumask.cbcpu);
                if (!pd)
                        return -ENOMEM;
 
                padata_replace(pinst, pd);
+
+               cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
+               cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
        }
 
        return 0;
 }
 
-/*
- * padata_remove_cpu - remove a cpu from the padata cpumask
+ /**
+ * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
+ *                     padata cpumasks.
  *
  * @pinst: padata instance
  * @cpu: cpu to remove
+ * @mask: bitmask specifying from which cpumask @cpu should be removed
+ *        The @mask may be any combination of the following flags:
+ *          PADATA_CPU_SERIAL   - serial cpumask
+ *          PADATA_CPU_PARALLEL - parallel cpumask
  */
-int padata_remove_cpu(struct padata_instance *pinst, int cpu)
+int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
 {
        int err;
 
+       if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
+               return -EINVAL;
+
        mutex_lock(&pinst->lock);
 
        get_online_cpus();
-       cpumask_clear_cpu(cpu, pinst->cpumask);
+       if (mask & PADATA_CPU_SERIAL)
+               cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
+       if (mask & PADATA_CPU_PARALLEL)
+               cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
+
        err = __padata_remove_cpu(pinst, cpu);
        put_online_cpus();
 
@@ -538,20 +791,29 @@ int padata_remove_cpu(struct padata_instance *pinst, int cpu)
 }
 EXPORT_SYMBOL(padata_remove_cpu);
 
-/*
+/**
  * padata_start - start the parallel processing
  *
  * @pinst: padata instance to start
  */
-void padata_start(struct padata_instance *pinst)
+int padata_start(struct padata_instance *pinst)
 {
+       int err = 0;
+
        mutex_lock(&pinst->lock);
-       pinst->flags |= PADATA_INIT;
+
+       if (pinst->flags & PADATA_INVALID)
+               err =-EINVAL;
+
+        __padata_start(pinst);
+
        mutex_unlock(&pinst->lock);
+
+       return err;
 }
 EXPORT_SYMBOL(padata_start);
 
-/*
+/**
  * padata_stop - stop the parallel processing
  *
  * @pinst: padata instance to stop
@@ -559,12 +821,20 @@ EXPORT_SYMBOL(padata_start);
 void padata_stop(struct padata_instance *pinst)
 {
        mutex_lock(&pinst->lock);
-       pinst->flags &= ~PADATA_INIT;
+       __padata_stop(pinst);
        mutex_unlock(&pinst->lock);
 }
 EXPORT_SYMBOL(padata_stop);
 
 #ifdef CONFIG_HOTPLUG_CPU
+
+static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
+{
+       return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
+               cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
+}
+
+
 static int padata_cpu_callback(struct notifier_block *nfb,
                               unsigned long action, void *hcpu)
 {
@@ -577,29 +847,29 @@ static int padata_cpu_callback(struct notifier_block *nfb,
        switch (action) {
        case CPU_ONLINE:
        case CPU_ONLINE_FROZEN:
-               if (!cpumask_test_cpu(cpu, pinst->cpumask))
+               if (!pinst_has_cpu(pinst, cpu))
                        break;
                mutex_lock(&pinst->lock);
                err = __padata_add_cpu(pinst, cpu);
                mutex_unlock(&pinst->lock);
                if (err)
-                       return NOTIFY_BAD;
+                       return notifier_from_errno(err);
                break;
 
        case CPU_DOWN_PREPARE:
        case CPU_DOWN_PREPARE_FROZEN:
-               if (!cpumask_test_cpu(cpu, pinst->cpumask))
+               if (!pinst_has_cpu(pinst, cpu))
                        break;
                mutex_lock(&pinst->lock);
                err = __padata_remove_cpu(pinst, cpu);
                mutex_unlock(&pinst->lock);
                if (err)
-                       return NOTIFY_BAD;
+                       return notifier_from_errno(err);
                break;
 
        case CPU_UP_CANCELED:
        case CPU_UP_CANCELED_FROZEN:
-               if (!cpumask_test_cpu(cpu, pinst->cpumask))
+               if (!pinst_has_cpu(pinst, cpu))
                        break;
                mutex_lock(&pinst->lock);
                __padata_remove_cpu(pinst, cpu);
@@ -607,7 +877,7 @@ static int padata_cpu_callback(struct notifier_block *nfb,
 
        case CPU_DOWN_FAILED:
        case CPU_DOWN_FAILED_FROZEN:
-               if (!cpumask_test_cpu(cpu, pinst->cpumask))
+               if (!pinst_has_cpu(pinst, cpu))
                        break;
                mutex_lock(&pinst->lock);
                __padata_add_cpu(pinst, cpu);
@@ -618,36 +888,202 @@ static int padata_cpu_callback(struct notifier_block *nfb,
 }
 #endif
 
+static void __padata_free(struct padata_instance *pinst)
+{
+#ifdef CONFIG_HOTPLUG_CPU
+       unregister_hotcpu_notifier(&pinst->cpu_notifier);
+#endif
+
+       padata_stop(pinst);
+       padata_free_pd(pinst->pd);
+       free_cpumask_var(pinst->cpumask.pcpu);
+       free_cpumask_var(pinst->cpumask.cbcpu);
+       kfree(pinst);
+}
+
+#define kobj2pinst(_kobj)                                      \
+       container_of(_kobj, struct padata_instance, kobj)
+#define attr2pentry(_attr)                                     \
+       container_of(_attr, struct padata_sysfs_entry, attr)
+
+static void padata_sysfs_release(struct kobject *kobj)
+{
+       struct padata_instance *pinst = kobj2pinst(kobj);
+       __padata_free(pinst);
+}
+
+struct padata_sysfs_entry {
+       struct attribute attr;
+       ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
+       ssize_t (*store)(struct padata_instance *, struct attribute *,
+                        const char *, size_t);
+};
+
+static ssize_t show_cpumask(struct padata_instance *pinst,
+                           struct attribute *attr,  char *buf)
+{
+       struct cpumask *cpumask;
+       ssize_t len;
+
+       mutex_lock(&pinst->lock);
+       if (!strcmp(attr->name, "serial_cpumask"))
+               cpumask = pinst->cpumask.cbcpu;
+       else
+               cpumask = pinst->cpumask.pcpu;
+
+       len = bitmap_scnprintf(buf, PAGE_SIZE, cpumask_bits(cpumask),
+                              nr_cpu_ids);
+       if (PAGE_SIZE - len < 2)
+               len = -EINVAL;
+       else
+               len += sprintf(buf + len, "\n");
+
+       mutex_unlock(&pinst->lock);
+       return len;
+}
+
+static ssize_t store_cpumask(struct padata_instance *pinst,
+                            struct attribute *attr,
+                            const char *buf, size_t count)
+{
+       cpumask_var_t new_cpumask;
+       ssize_t ret;
+       int mask_type;
+
+       if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
+               return -ENOMEM;
+
+       ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
+                          nr_cpumask_bits);
+       if (ret < 0)
+               goto out;
+
+       mask_type = !strcmp(attr->name, "serial_cpumask") ?
+               PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
+       ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
+       if (!ret)
+               ret = count;
+
+out:
+       free_cpumask_var(new_cpumask);
+       return ret;
+}
+
+#define PADATA_ATTR_RW(_name, _show_name, _store_name)         \
+       static struct padata_sysfs_entry _name##_attr =         \
+               __ATTR(_name, 0644, _show_name, _store_name)
+#define PADATA_ATTR_RO(_name, _show_name)              \
+       static struct padata_sysfs_entry _name##_attr = \
+               __ATTR(_name, 0400, _show_name, NULL)
+
+PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
+PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
+
 /*
- * padata_alloc - allocate and initialize a padata instance
+ * Padata sysfs provides the following objects:
+ * serial_cpumask   [RW] - cpumask for serial workers
+ * parallel_cpumask [RW] - cpumask for parallel workers
+ */
+static struct attribute *padata_default_attrs[] = {
+       &serial_cpumask_attr.attr,
+       &parallel_cpumask_attr.attr,
+       NULL,
+};
+
+static ssize_t padata_sysfs_show(struct kobject *kobj,
+                                struct attribute *attr, char *buf)
+{
+       struct padata_instance *pinst;
+       struct padata_sysfs_entry *pentry;
+       ssize_t ret = -EIO;
+
+       pinst = kobj2pinst(kobj);
+       pentry = attr2pentry(attr);
+       if (pentry->show)
+               ret = pentry->show(pinst, attr, buf);
+
+       return ret;
+}
+
+static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
+                                 const char *buf, size_t count)
+{
+       struct padata_instance *pinst;
+       struct padata_sysfs_entry *pentry;
+       ssize_t ret = -EIO;
+
+       pinst = kobj2pinst(kobj);
+       pentry = attr2pentry(attr);
+       if (pentry->show)
+               ret = pentry->store(pinst, attr, buf, count);
+
+       return ret;
+}
+
+static const struct sysfs_ops padata_sysfs_ops = {
+       .show = padata_sysfs_show,
+       .store = padata_sysfs_store,
+};
+
+static struct kobj_type padata_attr_type = {
+       .sysfs_ops = &padata_sysfs_ops,
+       .default_attrs = padata_default_attrs,
+       .release = padata_sysfs_release,
+};
+
+/**
+ * padata_alloc_possible - Allocate and initialize padata instance.
+ *                         Use the cpu_possible_mask for serial and
+ *                         parallel workers.
  *
- * @cpumask: cpumask that padata uses for parallelization
  * @wq: workqueue to use for the allocated padata instance
  */
-struct padata_instance *padata_alloc(const struct cpumask *cpumask,
-                                    struct workqueue_struct *wq)
+struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
+{
+       return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
+}
+EXPORT_SYMBOL(padata_alloc_possible);
+
+/**
+ * padata_alloc - allocate and initialize a padata instance and specify
+ *                cpumasks for serial and parallel workers.
+ *
+ * @wq: workqueue to use for the allocated padata instance
+ * @pcpumask: cpumask that will be used for padata parallelization
+ * @cbcpumask: cpumask that will be used for padata serialization
+ */
+struct padata_instance *padata_alloc(struct workqueue_struct *wq,
+                                    const struct cpumask *pcpumask,
+                                    const struct cpumask *cbcpumask)
 {
        struct padata_instance *pinst;
-       struct parallel_data *pd;
+       struct parallel_data *pd = NULL;
 
        pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
        if (!pinst)
                goto err;
 
        get_online_cpus();
-
-       pd = padata_alloc_pd(pinst, cpumask);
-       if (!pd)
+       if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
+               goto err_free_inst;
+       if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
+               free_cpumask_var(pinst->cpumask.pcpu);
                goto err_free_inst;
+       }
+       if (!padata_validate_cpumask(pinst, pcpumask) ||
+           !padata_validate_cpumask(pinst, cbcpumask))
+               goto err_free_masks;
 
-       if (!alloc_cpumask_var(&pinst->cpumask, GFP_KERNEL))
-               goto err_free_pd;
+       pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
+       if (!pd)
+               goto err_free_masks;
 
        rcu_assign_pointer(pinst->pd, pd);
 
        pinst->wq = wq;
 
-       cpumask_copy(pinst->cpumask, cpumask);
+       cpumask_copy(pinst->cpumask.pcpu, pcpumask);
+       cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
 
        pinst->flags = 0;
 
@@ -659,12 +1095,15 @@ struct padata_instance *padata_alloc(const struct cpumask *cpumask,
 
        put_online_cpus();
 
+       BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
+       kobject_init(&pinst->kobj, &padata_attr_type);
        mutex_init(&pinst->lock);
 
        return pinst;
 
-err_free_pd:
-       padata_free_pd(pd);
+err_free_masks:
+       free_cpumask_var(pinst->cpumask.pcpu);
+       free_cpumask_var(pinst->cpumask.cbcpu);
 err_free_inst:
        kfree(pinst);
        put_online_cpus();
@@ -673,25 +1112,13 @@ err:
 }
 EXPORT_SYMBOL(padata_alloc);
 
-/*
+/**
  * padata_free - free a padata instance
  *
- * @ padata_inst: padata instance to free
+ * @padata_inst: padata instance to free
  */
 void padata_free(struct padata_instance *pinst)
 {
-       padata_stop(pinst);
-
-       synchronize_rcu();
-
-       while (atomic_read(&pinst->pd->refcnt) != 0)
-               yield();
-
-#ifdef CONFIG_HOTPLUG_CPU
-       unregister_hotcpu_notifier(&pinst->cpu_notifier);
-#endif
-       padata_free_pd(pinst->pd);
-       free_cpumask_var(pinst->cpumask);
-       kfree(pinst);
+       kobject_put(&pinst->kobj);
 }
 EXPORT_SYMBOL(padata_free);