V4L/DVB (13946): proc_fops: convert av7110
[safe/jmp/linux-2.6] / kernel / perf_event.c
index 0aafe85..2ae7409 100644 (file)
@@ -36,7 +36,7 @@
 /*
  * Each CPU has a list of per CPU events:
  */
-DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
+static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
 
 int perf_max_events __read_mostly = 1;
 static int perf_reserved_percpu __read_mostly;
@@ -203,14 +203,14 @@ perf_lock_task_context(struct task_struct *task, unsigned long *flags)
                 * if so.  If we locked the right context, then it
                 * can't get swapped on us any more.
                 */
-               spin_lock_irqsave(&ctx->lock, *flags);
+               raw_spin_lock_irqsave(&ctx->lock, *flags);
                if (ctx != rcu_dereference(task->perf_event_ctxp)) {
-                       spin_unlock_irqrestore(&ctx->lock, *flags);
+                       raw_spin_unlock_irqrestore(&ctx->lock, *flags);
                        goto retry;
                }
 
                if (!atomic_inc_not_zero(&ctx->refcount)) {
-                       spin_unlock_irqrestore(&ctx->lock, *flags);
+                       raw_spin_unlock_irqrestore(&ctx->lock, *flags);
                        ctx = NULL;
                }
        }
@@ -231,7 +231,7 @@ static struct perf_event_context *perf_pin_task_context(struct task_struct *task
        ctx = perf_lock_task_context(task, &flags);
        if (ctx) {
                ++ctx->pin_count;
-               spin_unlock_irqrestore(&ctx->lock, flags);
+               raw_spin_unlock_irqrestore(&ctx->lock, flags);
        }
        return ctx;
 }
@@ -240,9 +240,9 @@ static void perf_unpin_context(struct perf_event_context *ctx)
 {
        unsigned long flags;
 
-       spin_lock_irqsave(&ctx->lock, flags);
+       raw_spin_lock_irqsave(&ctx->lock, flags);
        --ctx->pin_count;
-       spin_unlock_irqrestore(&ctx->lock, flags);
+       raw_spin_unlock_irqrestore(&ctx->lock, flags);
        put_ctx(ctx);
 }
 
@@ -338,7 +338,16 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
                event->group_leader->nr_siblings--;
 
        update_event_times(event);
-       event->state = PERF_EVENT_STATE_OFF;
+
+       /*
+        * If event was in error state, then keep it
+        * that way, otherwise bogus counts will be
+        * returned on read(). The only way to get out
+        * of error state is by explicit re-enabling
+        * of the event
+        */
+       if (event->state > PERF_EVENT_STATE_OFF)
+               event->state = PERF_EVENT_STATE_OFF;
 
        /*
         * If this was a group event with sibling events then
@@ -418,7 +427,7 @@ static void __perf_event_remove_from_context(void *info)
        if (ctx->task && cpuctx->task_ctx != ctx)
                return;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        /*
         * Protect the list operation against NMI by disabling the
         * events on a global level.
@@ -440,7 +449,7 @@ static void __perf_event_remove_from_context(void *info)
        }
 
        perf_enable();
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 
@@ -467,7 +476,7 @@ static void perf_event_remove_from_context(struct perf_event *event)
        if (!task) {
                /*
                 * Per cpu events are removed via an smp call and
-                * the removal is always sucessful.
+                * the removal is always successful.
                 */
                smp_call_function_single(event->cpu,
                                         __perf_event_remove_from_context,
@@ -479,12 +488,12 @@ retry:
        task_oncpu_function_call(task, __perf_event_remove_from_context,
                                 event);
 
-       spin_lock_irq(&ctx->lock);
+       raw_spin_lock_irq(&ctx->lock);
        /*
         * If the context is active we need to retry the smp call.
         */
        if (ctx->nr_active && !list_empty(&event->group_entry)) {
-               spin_unlock_irq(&ctx->lock);
+               raw_spin_unlock_irq(&ctx->lock);
                goto retry;
        }
 
@@ -495,7 +504,7 @@ retry:
         */
        if (!list_empty(&event->group_entry))
                list_del_event(event, ctx);
-       spin_unlock_irq(&ctx->lock);
+       raw_spin_unlock_irq(&ctx->lock);
 }
 
 /*
@@ -526,7 +535,7 @@ static void __perf_event_disable(void *info)
        if (ctx->task && cpuctx->task_ctx != ctx)
                return;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
 
        /*
         * If the event is on, turn it off.
@@ -542,7 +551,7 @@ static void __perf_event_disable(void *info)
                event->state = PERF_EVENT_STATE_OFF;
        }
 
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 /*
@@ -558,7 +567,7 @@ static void __perf_event_disable(void *info)
  * is the current context on this CPU and preemption is disabled,
  * hence we can't get into perf_event_task_sched_out for this context.
  */
-static void perf_event_disable(struct perf_event *event)
+void perf_event_disable(struct perf_event *event)
 {
        struct perf_event_context *ctx = event->ctx;
        struct task_struct *task = ctx->task;
@@ -575,12 +584,12 @@ static void perf_event_disable(struct perf_event *event)
  retry:
        task_oncpu_function_call(task, __perf_event_disable, event);
 
-       spin_lock_irq(&ctx->lock);
+       raw_spin_lock_irq(&ctx->lock);
        /*
         * If the event is still active, we need to retry the cross-call.
         */
        if (event->state == PERF_EVENT_STATE_ACTIVE) {
-               spin_unlock_irq(&ctx->lock);
+               raw_spin_unlock_irq(&ctx->lock);
                goto retry;
        }
 
@@ -593,7 +602,7 @@ static void perf_event_disable(struct perf_event *event)
                event->state = PERF_EVENT_STATE_OFF;
        }
 
-       spin_unlock_irq(&ctx->lock);
+       raw_spin_unlock_irq(&ctx->lock);
 }
 
 static int
@@ -761,7 +770,7 @@ static void __perf_install_in_context(void *info)
                cpuctx->task_ctx = ctx;
        }
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        ctx->is_active = 1;
        update_context_time(ctx);
 
@@ -773,6 +782,9 @@ static void __perf_install_in_context(void *info)
 
        add_event_to_ctx(event, ctx);
 
+       if (event->cpu != -1 && event->cpu != smp_processor_id())
+               goto unlock;
+
        /*
         * Don't put the event on if it is disabled or if
         * it is in a group and the group isn't on.
@@ -811,7 +823,7 @@ static void __perf_install_in_context(void *info)
  unlock:
        perf_enable();
 
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 /*
@@ -836,7 +848,7 @@ perf_install_in_context(struct perf_event_context *ctx,
        if (!task) {
                /*
                 * Per cpu events are installed via an smp call and
-                * the install is always sucessful.
+                * the install is always successful.
                 */
                smp_call_function_single(cpu, __perf_install_in_context,
                                         event, 1);
@@ -847,12 +859,12 @@ retry:
        task_oncpu_function_call(task, __perf_install_in_context,
                                 event);
 
-       spin_lock_irq(&ctx->lock);
+       raw_spin_lock_irq(&ctx->lock);
        /*
         * we need to retry the smp call.
         */
        if (ctx->is_active && list_empty(&event->group_entry)) {
-               spin_unlock_irq(&ctx->lock);
+               raw_spin_unlock_irq(&ctx->lock);
                goto retry;
        }
 
@@ -863,7 +875,7 @@ retry:
         */
        if (list_empty(&event->group_entry))
                add_event_to_ctx(event, ctx);
-       spin_unlock_irq(&ctx->lock);
+       raw_spin_unlock_irq(&ctx->lock);
 }
 
 /*
@@ -908,7 +920,7 @@ static void __perf_event_enable(void *info)
                cpuctx->task_ctx = ctx;
        }
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        ctx->is_active = 1;
        update_context_time(ctx);
 
@@ -916,6 +928,9 @@ static void __perf_event_enable(void *info)
                goto unlock;
        __perf_event_mark_enabled(event, ctx);
 
+       if (event->cpu != -1 && event->cpu != smp_processor_id())
+               goto unlock;
+
        /*
         * If the event is in a group and isn't the group leader,
         * then don't put it on unless the group is on.
@@ -950,7 +965,7 @@ static void __perf_event_enable(void *info)
        }
 
  unlock:
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 /*
@@ -962,7 +977,7 @@ static void __perf_event_enable(void *info)
  * perf_event_for_each_child or perf_event_for_each as described
  * for perf_event_disable.
  */
-static void perf_event_enable(struct perf_event *event)
+void perf_event_enable(struct perf_event *event)
 {
        struct perf_event_context *ctx = event->ctx;
        struct task_struct *task = ctx->task;
@@ -976,7 +991,7 @@ static void perf_event_enable(struct perf_event *event)
                return;
        }
 
-       spin_lock_irq(&ctx->lock);
+       raw_spin_lock_irq(&ctx->lock);
        if (event->state >= PERF_EVENT_STATE_INACTIVE)
                goto out;
 
@@ -991,10 +1006,10 @@ static void perf_event_enable(struct perf_event *event)
                event->state = PERF_EVENT_STATE_OFF;
 
  retry:
-       spin_unlock_irq(&ctx->lock);
+       raw_spin_unlock_irq(&ctx->lock);
        task_oncpu_function_call(task, __perf_event_enable, event);
 
-       spin_lock_irq(&ctx->lock);
+       raw_spin_lock_irq(&ctx->lock);
 
        /*
         * If the context is active and the event is still off,
@@ -1011,7 +1026,7 @@ static void perf_event_enable(struct perf_event *event)
                __perf_event_mark_enabled(event, ctx);
 
  out:
-       spin_unlock_irq(&ctx->lock);
+       raw_spin_unlock_irq(&ctx->lock);
 }
 
 static int perf_event_refresh(struct perf_event *event, int refresh)
@@ -1033,7 +1048,7 @@ void __perf_event_sched_out(struct perf_event_context *ctx,
 {
        struct perf_event *event;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        ctx->is_active = 0;
        if (likely(!ctx->nr_events))
                goto out;
@@ -1046,7 +1061,7 @@ void __perf_event_sched_out(struct perf_event_context *ctx,
        }
        perf_enable();
  out:
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 /*
@@ -1184,8 +1199,8 @@ void perf_event_task_sched_out(struct task_struct *task,
                 * order we take the locks because no other cpu could
                 * be trying to lock both of these tasks.
                 */
-               spin_lock(&ctx->lock);
-               spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
+               raw_spin_lock(&ctx->lock);
+               raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
                if (context_equiv(ctx, next_ctx)) {
                        /*
                         * XXX do we need a memory barrier of sorts
@@ -1199,8 +1214,8 @@ void perf_event_task_sched_out(struct task_struct *task,
 
                        perf_event_sync_stat(ctx, next_ctx);
                }
-               spin_unlock(&next_ctx->lock);
-               spin_unlock(&ctx->lock);
+               raw_spin_unlock(&next_ctx->lock);
+               raw_spin_unlock(&ctx->lock);
        }
        rcu_read_unlock();
 
@@ -1242,7 +1257,7 @@ __perf_event_sched_in(struct perf_event_context *ctx,
        struct perf_event *event;
        int can_add_hw = 1;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        ctx->is_active = 1;
        if (likely(!ctx->nr_events))
                goto out;
@@ -1297,7 +1312,7 @@ __perf_event_sched_in(struct perf_event_context *ctx,
        }
        perf_enable();
  out:
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 /*
@@ -1361,11 +1376,14 @@ static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
        struct hw_perf_event *hwc;
        u64 interrupts, freq;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
                if (event->state != PERF_EVENT_STATE_ACTIVE)
                        continue;
 
+               if (event->cpu != -1 && event->cpu != smp_processor_id())
+                       continue;
+
                hwc = &event->hw;
 
                interrupts = hwc->interrupts;
@@ -1416,7 +1434,7 @@ static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
                        perf_enable();
                }
        }
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 /*
@@ -1429,7 +1447,7 @@ static void rotate_ctx(struct perf_event_context *ctx)
        if (!ctx->nr_events)
                return;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        /*
         * Rotate the first entry last (works just fine for group events too):
         */
@@ -1440,7 +1458,7 @@ static void rotate_ctx(struct perf_event_context *ctx)
        }
        perf_enable();
 
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 }
 
 void perf_event_task_tick(struct task_struct *curr, int cpu)
@@ -1489,7 +1507,7 @@ static void perf_event_enable_on_exec(struct task_struct *task)
 
        __perf_event_task_sched_out(ctx);
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
 
        list_for_each_entry(event, &ctx->group_list, group_entry) {
                if (!event->attr.enable_on_exec)
@@ -1507,7 +1525,7 @@ static void perf_event_enable_on_exec(struct task_struct *task)
        if (enabled)
                unclone_ctx(ctx);
 
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 
        perf_event_task_sched_in(task, smp_processor_id());
  out:
@@ -1533,10 +1551,10 @@ static void __perf_event_read(void *info)
        if (ctx->task && cpuctx->task_ctx != ctx)
                return;
 
-       spin_lock(&ctx->lock);
+       raw_spin_lock(&ctx->lock);
        update_context_time(ctx);
        update_event_times(event);
-       spin_unlock(&ctx->lock);
+       raw_spin_unlock(&ctx->lock);
 
        event->pmu->read(event);
 }
@@ -1554,10 +1572,10 @@ static u64 perf_event_read(struct perf_event *event)
                struct perf_event_context *ctx = event->ctx;
                unsigned long flags;
 
-               spin_lock_irqsave(&ctx->lock, flags);
+               raw_spin_lock_irqsave(&ctx->lock, flags);
                update_context_time(ctx);
                update_event_times(event);
-               spin_unlock_irqrestore(&ctx->lock, flags);
+               raw_spin_unlock_irqrestore(&ctx->lock, flags);
        }
 
        return atomic64_read(&event->count);
@@ -1570,8 +1588,7 @@ static void
 __perf_event_init_context(struct perf_event_context *ctx,
                            struct task_struct *task)
 {
-       memset(ctx, 0, sizeof(*ctx));
-       spin_lock_init(&ctx->lock);
+       raw_spin_lock_init(&ctx->lock);
        mutex_init(&ctx->mutex);
        INIT_LIST_HEAD(&ctx->group_list);
        INIT_LIST_HEAD(&ctx->event_list);
@@ -1587,15 +1604,12 @@ static struct perf_event_context *find_get_context(pid_t pid, int cpu)
        unsigned long flags;
        int err;
 
-       /*
-        * If cpu is not a wildcard then this is a percpu event:
-        */
-       if (cpu != -1) {
+       if (pid == -1 && cpu != -1) {
                /* Must be root to operate on a CPU event: */
                if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
                        return ERR_PTR(-EACCES);
 
-               if (cpu < 0 || cpu > num_possible_cpus())
+               if (cpu < 0 || cpu >= nr_cpumask_bits)
                        return ERR_PTR(-EINVAL);
 
                /*
@@ -1603,7 +1617,7 @@ static struct perf_event_context *find_get_context(pid_t pid, int cpu)
                 * offline CPU and activate it when the CPU comes up, but
                 * that's for later.
                 */
-               if (!cpu_isset(cpu, cpu_online_map))
+               if (!cpu_online(cpu))
                        return ERR_PTR(-ENODEV);
 
                cpuctx = &per_cpu(perf_cpu_context, cpu);
@@ -1641,11 +1655,11 @@ static struct perf_event_context *find_get_context(pid_t pid, int cpu)
        ctx = perf_lock_task_context(task, &flags);
        if (ctx) {
                unclone_ctx(ctx);
-               spin_unlock_irqrestore(&ctx->lock, flags);
+               raw_spin_unlock_irqrestore(&ctx->lock, flags);
        }
 
        if (!ctx) {
-               ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
+               ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
                err = -ENOMEM;
                if (!ctx)
                        goto errout;
@@ -1831,7 +1845,7 @@ static int perf_event_read_group(struct perf_event *event,
 
                size = n * sizeof(u64);
 
-               if (copy_to_user(buf + size, values, size)) {
+               if (copy_to_user(buf + ret, values, size)) {
                        ret = -EFAULT;
                        goto unlock;
                }
@@ -1979,7 +1993,7 @@ static int perf_event_period(struct perf_event *event, u64 __user *arg)
        if (!value)
                return -EINVAL;
 
-       spin_lock_irq(&ctx->lock);
+       raw_spin_lock_irq(&ctx->lock);
        if (event->attr.freq) {
                if (value > sysctl_perf_event_sample_rate) {
                        ret = -EINVAL;
@@ -1992,7 +2006,7 @@ static int perf_event_period(struct perf_event *event, u64 __user *arg)
                event->hw.sample_period = value;
        }
 unlock:
-       spin_unlock_irq(&ctx->lock);
+       raw_spin_unlock_irq(&ctx->lock);
 
        return ret;
 }
@@ -2201,6 +2215,7 @@ static void perf_mmap_data_free(struct perf_mmap_data *data)
        perf_mmap_free_page((unsigned long)data->user_page);
        for (i = 0; i < data->nr_pages; i++)
                perf_mmap_free_page((unsigned long)data->data_pages[i]);
+       kfree(data);
 }
 
 #else
@@ -2241,6 +2256,7 @@ static void perf_mmap_data_free_work(struct work_struct *work)
                perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 
        vfree(base);
+       kfree(data);
 }
 
 static void perf_mmap_data_free(struct perf_mmap_data *data)
@@ -2346,7 +2362,6 @@ static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
 
        data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
        perf_mmap_data_free(data);
-       kfree(data);
 }
 
 static void perf_mmap_data_release(struct perf_event *event)
@@ -3244,8 +3259,6 @@ static void perf_event_task_output(struct perf_event *event,
        task_event->event_id.tid = perf_event_tid(event, task);
        task_event->event_id.ptid = perf_event_tid(event, current);
 
-       task_event->event_id.time = perf_clock();
-
        perf_output_put(&handle, task_event->event_id);
 
        perf_output_end(&handle);
@@ -3253,6 +3266,12 @@ static void perf_event_task_output(struct perf_event *event,
 
 static int perf_event_task_match(struct perf_event *event)
 {
+       if (event->state < PERF_EVENT_STATE_INACTIVE)
+               return 0;
+
+       if (event->cpu != -1 && event->cpu != smp_processor_id())
+               return 0;
+
        if (event->attr.comm || event->attr.mmap || event->attr.task)
                return 1;
 
@@ -3278,12 +3297,11 @@ static void perf_event_task_event(struct perf_task_event *task_event)
        rcu_read_lock();
        cpuctx = &get_cpu_var(perf_cpu_context);
        perf_event_task_ctx(&cpuctx->ctx, task_event);
-       put_cpu_var(perf_cpu_context);
-
        if (!ctx)
-               ctx = rcu_dereference(task_event->task->perf_event_ctxp);
+               ctx = rcu_dereference(current->perf_event_ctxp);
        if (ctx)
                perf_event_task_ctx(ctx, task_event);
+       put_cpu_var(perf_cpu_context);
        rcu_read_unlock();
 }
 
@@ -3311,6 +3329,7 @@ static void perf_event_task(struct task_struct *task,
                        /* .ppid */
                        /* .tid  */
                        /* .ptid */
+                       .time = perf_clock(),
                },
        };
 
@@ -3360,6 +3379,12 @@ static void perf_event_comm_output(struct perf_event *event,
 
 static int perf_event_comm_match(struct perf_event *event)
 {
+       if (event->state < PERF_EVENT_STATE_INACTIVE)
+               return 0;
+
+       if (event->cpu != -1 && event->cpu != smp_processor_id())
+               return 0;
+
        if (event->attr.comm)
                return 1;
 
@@ -3396,15 +3421,10 @@ static void perf_event_comm_event(struct perf_comm_event *comm_event)
        rcu_read_lock();
        cpuctx = &get_cpu_var(perf_cpu_context);
        perf_event_comm_ctx(&cpuctx->ctx, comm_event);
-       put_cpu_var(perf_cpu_context);
-
-       /*
-        * doesn't really matter which of the child contexts the
-        * events ends up in.
-        */
        ctx = rcu_dereference(current->perf_event_ctxp);
        if (ctx)
                perf_event_comm_ctx(ctx, comm_event);
+       put_cpu_var(perf_cpu_context);
        rcu_read_unlock();
 }
 
@@ -3479,6 +3499,12 @@ static void perf_event_mmap_output(struct perf_event *event,
 static int perf_event_mmap_match(struct perf_event *event,
                                   struct perf_mmap_event *mmap_event)
 {
+       if (event->state < PERF_EVENT_STATE_INACTIVE)
+               return 0;
+
+       if (event->cpu != -1 && event->cpu != smp_processor_id())
+               return 0;
+
        if (event->attr.mmap)
                return 1;
 
@@ -3552,15 +3578,10 @@ got_name:
        rcu_read_lock();
        cpuctx = &get_cpu_var(perf_cpu_context);
        perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
-       put_cpu_var(perf_cpu_context);
-
-       /*
-        * doesn't really matter which of the child contexts the
-        * events ends up in.
-        */
        ctx = rcu_dereference(current->perf_event_ctxp);
        if (ctx)
                perf_event_mmap_ctx(ctx, mmap_event);
+       put_cpu_var(perf_cpu_context);
        rcu_read_unlock();
 
        kfree(buf);
@@ -3831,27 +3852,40 @@ static int perf_swevent_is_counting(struct perf_event *event)
 static int perf_tp_event_match(struct perf_event *event,
                                struct perf_sample_data *data);
 
+static int perf_exclude_event(struct perf_event *event,
+                             struct pt_regs *regs)
+{
+       if (regs) {
+               if (event->attr.exclude_user && user_mode(regs))
+                       return 1;
+
+               if (event->attr.exclude_kernel && !user_mode(regs))
+                       return 1;
+       }
+
+       return 0;
+}
+
 static int perf_swevent_match(struct perf_event *event,
                                enum perf_type_id type,
                                u32 event_id,
                                struct perf_sample_data *data,
                                struct pt_regs *regs)
 {
+       if (event->cpu != -1 && event->cpu != smp_processor_id())
+               return 0;
+
        if (!perf_swevent_is_counting(event))
                return 0;
 
        if (event->attr.type != type)
                return 0;
+
        if (event->attr.config != event_id)
                return 0;
 
-       if (regs) {
-               if (event->attr.exclude_user && user_mode(regs))
-                       return 0;
-
-               if (event->attr.exclude_kernel && !user_mode(regs))
-                       return 0;
-       }
+       if (perf_exclude_event(event, regs))
+               return 0;
 
        if (event->attr.type == PERF_TYPE_TRACEPOINT &&
            !perf_tp_event_match(event, data))
@@ -3904,7 +3938,7 @@ void perf_swevent_put_recursion_context(int rctx)
 {
        struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
        barrier();
-       cpuctx->recursion[rctx]++;
+       cpuctx->recursion[rctx]--;
        put_cpu_var(perf_cpu_context);
 }
 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
@@ -3991,6 +4025,8 @@ static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
        event->pmu->read(event);
 
        data.addr = 0;
+       data.raw = NULL;
+       data.period = event->hw.last_period;
        regs = get_irq_regs();
        /*
         * In case we exclude kernel IPs or are somehow not in interrupt
@@ -4059,8 +4095,7 @@ static void cpu_clock_perf_event_update(struct perf_event *event)
        u64 now;
 
        now = cpu_clock(cpu);
-       prev = atomic64_read(&event->hw.prev_count);
-       atomic64_set(&event->hw.prev_count, now);
+       prev = atomic64_xchg(&event->hw.prev_count, now);
        atomic64_add(now - prev, &event->count);
 }
 
@@ -4265,15 +4300,8 @@ static void bp_perf_event_destroy(struct perf_event *event)
 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
 {
        int err;
-       /*
-        * The breakpoint is already filled if we haven't created the counter
-        * through perf syscall
-        * FIXME: manage to get trigerred to NULL if it comes from syscalls
-        */
-       if (!bp->callback)
-               err = register_perf_hw_breakpoint(bp);
-       else
-               err = __register_perf_hw_breakpoint(bp);
+
+       err = register_perf_hw_breakpoint(bp);
        if (err)
                return ERR_PTR(err);
 
@@ -4282,15 +4310,18 @@ static const struct pmu *bp_perf_event_init(struct perf_event *bp)
        return &perf_ops_bp;
 }
 
-void perf_bp_event(struct perf_event *bp, void *regs)
+void perf_bp_event(struct perf_event *bp, void *data)
 {
-       /* TODO */
+       struct perf_sample_data sample;
+       struct pt_regs *regs = data;
+
+       sample.raw = NULL;
+       sample.addr = bp->attr.bp_addr;
+
+       if (!perf_exclude_event(bp, regs))
+               perf_swevent_add(bp, 1, 1, &sample, regs);
 }
 #else
-static void bp_perf_event_destroy(struct perf_event *event)
-{
-}
-
 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
 {
        return NULL;
@@ -4367,7 +4398,7 @@ perf_event_alloc(struct perf_event_attr *attr,
                   struct perf_event_context *ctx,
                   struct perf_event *group_leader,
                   struct perf_event *parent_event,
-                  perf_callback_t callback,
+                  perf_overflow_handler_t overflow_handler,
                   gfp_t gfpflags)
 {
        const struct pmu *pmu;
@@ -4410,10 +4441,10 @@ perf_event_alloc(struct perf_event_attr *attr,
 
        event->state            = PERF_EVENT_STATE_INACTIVE;
 
-       if (!callback && parent_event)
-               callback = parent_event->callback;
+       if (!overflow_handler && parent_event)
+               overflow_handler = parent_event->overflow_handler;
        
-       event->callback = callback;
+       event->overflow_handler = overflow_handler;
 
        if (attr->disabled)
                event->state = PERF_EVENT_STATE_OFF;
@@ -4548,7 +4579,7 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,
        if (attr->type >= PERF_TYPE_MAX)
                return -EINVAL;
 
-       if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
+       if (attr->__reserved_1)
                return -EINVAL;
 
        if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
@@ -4701,7 +4732,7 @@ SYSCALL_DEFINE5(perf_event_open,
        if (IS_ERR(event))
                goto err_put_context;
 
-       err = anon_inode_getfd("[perf_event]", &perf_fops, event, 0);
+       err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
        if (err < 0)
                goto err_free_put_context;
 
@@ -4753,7 +4784,8 @@ err_put_context:
  */
 struct perf_event *
 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
-                                pid_t pid, perf_callback_t callback)
+                                pid_t pid,
+                                perf_overflow_handler_t overflow_handler)
 {
        struct perf_event *event;
        struct perf_event_context *ctx;
@@ -4764,14 +4796,17 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
         */
 
        ctx = find_get_context(pid, cpu);
-       if (IS_ERR(ctx))
-               return NULL;
+       if (IS_ERR(ctx)) {
+               err = PTR_ERR(ctx);
+               goto err_exit;
+       }
 
        event = perf_event_alloc(attr, cpu, ctx, NULL,
-                                    NULL, callback, GFP_KERNEL);
-       err = PTR_ERR(event);
-       if (IS_ERR(event))
+                                NULL, overflow_handler, GFP_KERNEL);
+       if (IS_ERR(event)) {
+               err = PTR_ERR(event);
                goto err_put_context;
+       }
 
        event->filp = NULL;
        WARN_ON_ONCE(ctx->parent_ctx);
@@ -4788,11 +4823,10 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
 
        return event;
 
-err_put_context:
-       if (err < 0)
-               put_ctx(ctx);
-
-       return NULL;
+ err_put_context:
+       put_ctx(ctx);
+ err_exit:
+       return ERR_PTR(err);
 }
 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
 
@@ -4973,7 +5007,7 @@ void perf_event_exit_task(struct task_struct *child)
         * reading child->perf_event_ctxp, we wait until it has
         * incremented the context's refcount before we do put_ctx below.
         */
-       spin_lock(&child_ctx->lock);
+       raw_spin_lock(&child_ctx->lock);
        child->perf_event_ctxp = NULL;
        /*
         * If this context is a clone; unclone it so it can't get
@@ -4982,7 +5016,7 @@ void perf_event_exit_task(struct task_struct *child)
         */
        unclone_ctx(child_ctx);
        update_context_time(child_ctx);
-       spin_unlock_irqrestore(&child_ctx->lock, flags);
+       raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
 
        /*
         * Report the task dead after unscheduling the events so that we
@@ -5065,7 +5099,7 @@ again:
  */
 int perf_event_init_task(struct task_struct *child)
 {
-       struct perf_event_context *child_ctx, *parent_ctx;
+       struct perf_event_context *child_ctx = NULL, *parent_ctx;
        struct perf_event_context *cloned_ctx;
        struct perf_event *event;
        struct task_struct *parent = current;
@@ -5081,20 +5115,6 @@ int perf_event_init_task(struct task_struct *child)
                return 0;
 
        /*
-        * This is executed from the parent task context, so inherit
-        * events that have been marked for cloning.
-        * First allocate and initialize a context for the child.
-        */
-
-       child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
-       if (!child_ctx)
-               return -ENOMEM;
-
-       __perf_event_init_context(child_ctx, child);
-       child->perf_event_ctxp = child_ctx;
-       get_task_struct(child);
-
-       /*
         * If the parent's context is a clone, pin it so it won't get
         * swapped under us.
         */
@@ -5124,6 +5144,26 @@ int perf_event_init_task(struct task_struct *child)
                        continue;
                }
 
+               if (!child->perf_event_ctxp) {
+                       /*
+                        * This is executed from the parent task context, so
+                        * inherit events that have been marked for cloning.
+                        * First allocate and initialize a context for the
+                        * child.
+                        */
+
+                       child_ctx = kzalloc(sizeof(struct perf_event_context),
+                                           GFP_KERNEL);
+                       if (!child_ctx) {
+                               ret = -ENOMEM;
+                               break;
+                       }
+
+                       __perf_event_init_context(child_ctx, child);
+                       child->perf_event_ctxp = child_ctx;
+                       get_task_struct(child);
+               }
+
                ret = inherit_group(event, parent, parent_ctx,
                                             child, child_ctx);
                if (ret) {
@@ -5132,7 +5172,7 @@ int perf_event_init_task(struct task_struct *child)
                }
        }
 
-       if (inherited_all) {
+       if (child_ctx && inherited_all) {
                /*
                 * Mark the child context as a clone of the parent
                 * context, or of whatever the parent is a clone of.
@@ -5266,11 +5306,11 @@ perf_set_reserve_percpu(struct sysdev_class *class,
        perf_reserved_percpu = val;
        for_each_online_cpu(cpu) {
                cpuctx = &per_cpu(perf_cpu_context, cpu);
-               spin_lock_irq(&cpuctx->ctx.lock);
+               raw_spin_lock_irq(&cpuctx->ctx.lock);
                mpt = min(perf_max_events - cpuctx->ctx.nr_events,
                          perf_max_events - perf_reserved_percpu);
                cpuctx->max_pertask = mpt;
-               spin_unlock_irq(&cpuctx->ctx.lock);
+               raw_spin_unlock_irq(&cpuctx->ctx.lock);
        }
        spin_unlock(&perf_resource_lock);