1 files changed, 443 insertions, 310 deletions

diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index a4fa381..ff86c55 100644
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -283,14 +283,15 @@ ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
 static void
 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 {
-	struct perf_event *group_leader = event->group_leader;
+	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
+	event->attach_state |= PERF_ATTACH_CONTEXT;
 
 	/*
-	 * Depending on whether it is a standalone or sibling event,
-	 * add it straight to the context's event list, or to the group
-	 * leader's sibling list:
+	 * If we're a stand alone event or group leader, we go to the context
+	 * list, group events are kept attached to the group so that
+	 * perf_group_detach can, at all times, locate all siblings.
 	 */
-	if (group_leader == event) {
+	if (event->group_leader == event) {
 		struct list_head *list;
 
 		if (is_software_event(event))
@@ -298,13 +299,6 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 
 		list = ctx_group_list(event, ctx);
 		list_add_tail(&event->group_entry, list);
-	} else {
-		if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
-		    !is_software_event(event))
-			group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
-
-		list_add_tail(&event->group_entry, &group_leader->sibling_list);
-		group_leader->nr_siblings++;
 	}
 
 	list_add_rcu(&event->event_entry, &ctx->event_list);
@@ -313,6 +307,24 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 		ctx->nr_stat++;
 }
 
+static void perf_group_attach(struct perf_event *event)
+{
+	struct perf_event *group_leader = event->group_leader;
+
+	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP);
+	event->attach_state |= PERF_ATTACH_GROUP;
+
+	if (group_leader == event)
+		return;
+
+	if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
+			!is_software_event(event))
+		group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
+
+	list_add_tail(&event->group_entry, &group_leader->sibling_list);
+	group_leader->nr_siblings++;
+}
+
 /*
  * Remove a event from the lists for its context.
  * Must be called with ctx->mutex and ctx->lock held.
@@ -320,17 +332,22 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 static void
 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
 {
-	if (list_empty(&event->group_entry))
+	/*
+	 * We can have double detach due to exit/hot-unplug + close.
+	 */
+	if (!(event->attach_state & PERF_ATTACH_CONTEXT))
 		return;
+
+	event->attach_state &= ~PERF_ATTACH_CONTEXT;
+
 	ctx->nr_events--;
 	if (event->attr.inherit_stat)
 		ctx->nr_stat--;
 
-	list_del_init(&event->group_entry);
 	list_del_rcu(&event->event_entry);
 
-	if (event->group_leader != event)
-		event->group_leader->nr_siblings--;
+	if (event->group_leader == event)
+		list_del_init(&event->group_entry);
 
 	update_group_times(event);
 
@@ -345,21 +362,39 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
 		event->state = PERF_EVENT_STATE_OFF;
 }
 
-static void
-perf_destroy_group(struct perf_event *event, struct perf_event_context *ctx)
+static void perf_group_detach(struct perf_event *event)
 {
 	struct perf_event *sibling, *tmp;
+	struct list_head *list = NULL;
+
+	/*
+	 * We can have double detach due to exit/hot-unplug + close.
+	 */
+	if (!(event->attach_state & PERF_ATTACH_GROUP))
+		return;
+
+	event->attach_state &= ~PERF_ATTACH_GROUP;
+
+	/*
+	 * If this is a sibling, remove it from its group.
+	 */
+	if (event->group_leader != event) {
+		list_del_init(&event->group_entry);
+		event->group_leader->nr_siblings--;
+		return;
+	}
+
+	if (!list_empty(&event->group_entry))
+		list = &event->group_entry;
 
 	/*
 	 * If this was a group event with sibling events then
 	 * upgrade the siblings to singleton events by adding them
-	 * to the context list directly:
+	 * to whatever list we are on.
 	 */
 	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
-		struct list_head *list;
-
-		list = ctx_group_list(event, ctx);
-		list_move_tail(&sibling->group_entry, list);
+		if (list)
+			list_move_tail(&sibling->group_entry, list);
 		sibling->group_leader = sibling;
 
 		/* Inherit group flags from the previous leader */
@@ -652,8 +687,11 @@ group_sched_in(struct perf_event *group_event,
 	if (txn)
 		pmu->start_txn(pmu);
 
-	if (event_sched_in(group_event, cpuctx, ctx))
+	if (event_sched_in(group_event, cpuctx, ctx)) {
+		if (txn)
+			pmu->cancel_txn(pmu);
 		return -EAGAIN;
+	}
 
 	/*
 	 * Schedule in siblings as one group (if any):
@@ -675,9 +713,6 @@ group_sched_in(struct perf_event *group_event,
 	}
 
 group_error:
-	if (txn)
-		pmu->cancel_txn(pmu);
-
 	/*
 	 * Groups can be scheduled in as one unit only, so undo any
 	 * partial group before returning:
@@ -689,6 +724,9 @@ group_error:
 	}
 	event_sched_out(group_event, cpuctx, ctx);
 
+	if (txn)
+		pmu->cancel_txn(pmu);
+
 	return -EAGAIN;
 }
 
@@ -727,6 +765,7 @@ static void add_event_to_ctx(struct perf_event *event,
 			       struct perf_event_context *ctx)
 {
 	list_add_event(event, ctx);
+	perf_group_attach(event);
 	event->tstamp_enabled = ctx->time;
 	event->tstamp_running = ctx->time;
 	event->tstamp_stopped = ctx->time;
@@ -1468,6 +1507,9 @@ do {					\
 		divisor = nsec * frequency;
 	}
 
+	if (!divisor)
+		return dividend;
+
 	return div64_u64(dividend, divisor);
 }
 
@@ -1490,7 +1532,7 @@ static int perf_event_start(struct perf_event *event)
 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
 {
 	struct hw_perf_event *hwc = &event->hw;
-	u64 period, sample_period;
+	s64 period, sample_period;
 	s64 delta;
 
 	period = perf_calculate_period(event, nsec, count);
@@ -1841,6 +1883,7 @@ static void free_event_rcu(struct rcu_head *head)
 }
 
 static void perf_pending_sync(struct perf_event *event);
+static void perf_mmap_data_put(struct perf_mmap_data *data);
 
 static void free_event(struct perf_event *event)
 {
@@ -1856,9 +1899,9 @@ static void free_event(struct perf_event *event)
 			atomic_dec(&nr_task_events);
 	}
 
-	if (event->output) {
-		fput(event->output->filp);
-		event->output = NULL;
+	if (event->data) {
+		perf_mmap_data_put(event->data);
+		event->data = NULL;
 	}
 
 	if (event->destroy)
@@ -1893,8 +1936,8 @@ int perf_event_release_kernel(struct perf_event *event)
 	 */
 	mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
 	raw_spin_lock_irq(&ctx->lock);
+	perf_group_detach(event);
 	list_del_event(event, ctx);
-	perf_destroy_group(event, ctx);
 	raw_spin_unlock_irq(&ctx->lock);
 	mutex_unlock(&ctx->mutex);
 
@@ -2175,7 +2218,27 @@ unlock:
 	return ret;
 }
 
-static int perf_event_set_output(struct perf_event *event, int output_fd);
+static const struct file_operations perf_fops;
+
+static struct perf_event *perf_fget_light(int fd, int *fput_needed)
+{
+	struct file *file;
+
+	file = fget_light(fd, fput_needed);
+	if (!file)
+		return ERR_PTR(-EBADF);
+
+	if (file->f_op != &perf_fops) {
+		fput_light(file, *fput_needed);
+		*fput_needed = 0;
+		return ERR_PTR(-EBADF);
+	}
+
+	return file->private_data;
+}
+
+static int perf_event_set_output(struct perf_event *event,
+				 struct perf_event *output_event);
 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
 
 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
@@ -2202,7 +2265,23 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 		return perf_event_period(event, (u64 __user *)arg);
 
 	case PERF_EVENT_IOC_SET_OUTPUT:
-		return perf_event_set_output(event, arg);
+	{
+		struct perf_event *output_event = NULL;
+		int fput_needed = 0;
+		int ret;
+
+		if (arg != -1) {
+			output_event = perf_fget_light(arg, &fput_needed);
+			if (IS_ERR(output_event))
+				return PTR_ERR(output_event);
+		}
+
+		ret = perf_event_set_output(event, output_event);
+		if (output_event)
+			fput_light(output_event->filp, fput_needed);
+
+		return ret;
+	}
 
 	case PERF_EVENT_IOC_SET_FILTER:
 		return perf_event_set_filter(event, (void __user *)arg);
@@ -2297,11 +2376,6 @@ unlock:
 	rcu_read_unlock();
 }
 
-static unsigned long perf_data_size(struct perf_mmap_data *data)
-{
-	return data->nr_pages << (PAGE_SHIFT + data->data_order);
-}
-
 #ifndef CONFIG_PERF_USE_VMALLOC
 
 /*
@@ -2320,6 +2394,19 @@ perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
 	return virt_to_page(data->data_pages[pgoff - 1]);
 }
 
+static void *perf_mmap_alloc_page(int cpu)
+{
+	struct page *page;
+	int node;
+
+	node = (cpu == -1) ? cpu : cpu_to_node(cpu);
+	page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
+	if (!page)
+		return NULL;
+
+	return page_address(page);
+}
+
 static struct perf_mmap_data *
 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
 {
@@ -2327,8 +2414,6 @@ perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
 	unsigned long size;
 	int i;
 
-	WARN_ON(atomic_read(&event->mmap_count));
-
 	size = sizeof(struct perf_mmap_data);
 	size += nr_pages * sizeof(void *);
 
@@ -2336,17 +2421,16 @@ perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
 	if (!data)
 		goto fail;
 
-	data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
+	data->user_page = perf_mmap_alloc_page(event->cpu);
 	if (!data->user_page)
 		goto fail_user_page;
 
 	for (i = 0; i < nr_pages; i++) {
-		data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
+		data->data_pages[i] = perf_mmap_alloc_page(event->cpu);
 		if (!data->data_pages[i])
 			goto fail_data_pages;
 	}
 
-	data->data_order = 0;
 	data->nr_pages = nr_pages;
 
 	return data;
@@ -2382,6 +2466,11 @@ static void perf_mmap_data_free(struct perf_mmap_data *data)
 	kfree(data);
 }
 
+static inline int page_order(struct perf_mmap_data *data)
+{
+	return 0;
+}
+
 #else
 
 /*
@@ -2390,10 +2479,15 @@ static void perf_mmap_data_free(struct perf_mmap_data *data)
  * Required for architectures that have d-cache aliasing issues.
  */
 
+static inline int page_order(struct perf_mmap_data *data)
+{
+	return data->page_order;
+}
+
 static struct page *
 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
 {
-	if (pgoff > (1UL << data->data_order))
+	if (pgoff > (1UL << page_order(data)))
 		return NULL;
 
 	return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
@@ -2413,7 +2507,7 @@ static void perf_mmap_data_free_work(struct work_struct *work)
 	int i, nr;
 
 	data = container_of(work, struct perf_mmap_data, work);
-	nr = 1 << data->data_order;
+	nr = 1 << page_order(data);
 
 	base = data->user_page;
 	for (i = 0; i < nr + 1; i++)
@@ -2435,8 +2529,6 @@ perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
 	unsigned long size;
 	void *all_buf;
 
-	WARN_ON(atomic_read(&event->mmap_count));
-
 	size = sizeof(struct perf_mmap_data);
 	size += sizeof(void *);
 
@@ -2452,7 +2544,7 @@ perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
 
 	data->user_page = all_buf;
 	data->data_pages[0] = all_buf + PAGE_SIZE;
-	data->data_order = ilog2(nr_pages);
+	data->page_order = ilog2(nr_pages);
 	data->nr_pages = 1;
 
 	return data;
@@ -2466,6 +2558,11 @@ fail:
 
 #endif
 
+static unsigned long perf_data_size(struct perf_mmap_data *data)
+{
+	return data->nr_pages << (PAGE_SHIFT + page_order(data));
+}
+
 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
 	struct perf_event *event = vma->vm_file->private_data;
@@ -2506,8 +2603,6 @@ perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
 {
 	long max_size = perf_data_size(data);
 
-	atomic_set(&data->lock, -1);
-
 	if (event->attr.watermark) {
 		data->watermark = min_t(long, max_size,
 					event->attr.wakeup_watermark);
@@ -2516,7 +2611,7 @@ perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
 	if (!data->watermark)
 		data->watermark = max_size / 2;
 
-
+	atomic_set(&data->refcount, 1);
 	rcu_assign_pointer(event->data, data);
 }
 
@@ -2528,13 +2623,26 @@ static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
 	perf_mmap_data_free(data);
 }
 
-static void perf_mmap_data_release(struct perf_event *event)
+static struct perf_mmap_data *perf_mmap_data_get(struct perf_event *event)
 {
-	struct perf_mmap_data *data = event->data;
+	struct perf_mmap_data *data;
 
-	WARN_ON(atomic_read(&event->mmap_count));
+	rcu_read_lock();
+	data = rcu_dereference(event->data);
+	if (data) {
+		if (!atomic_inc_not_zero(&data->refcount))
+			data = NULL;
+	}
+	rcu_read_unlock();
+
+	return data;
+}
+
+static void perf_mmap_data_put(struct perf_mmap_data *data)
+{
+	if (!atomic_dec_and_test(&data->refcount))
+		return;
 
-	rcu_assign_pointer(event->data, NULL);
 	call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
 }
 
@@ -2549,15 +2657,18 @@ static void perf_mmap_close(struct vm_area_struct *vma)
 {
 	struct perf_event *event = vma->vm_file->private_data;
 
-	WARN_ON_ONCE(event->ctx->parent_ctx);
 	if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
 		unsigned long size = perf_data_size(event->data);
-		struct user_struct *user = current_user();
+		struct user_struct *user = event->mmap_user;
+		struct perf_mmap_data *data = event->data;
 
 		atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
-		vma->vm_mm->locked_vm -= event->data->nr_locked;
-		perf_mmap_data_release(event);
+		vma->vm_mm->locked_vm -= event->mmap_locked;
+		rcu_assign_pointer(event->data, NULL);
 		mutex_unlock(&event->mmap_mutex);
+
+		perf_mmap_data_put(data);
+		free_uid(user);
 	}
 }
 
@@ -2580,6 +2691,14 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 	long user_extra, extra;
 	int ret = 0;
 
+	/*
+	 * Don't allow mmap() of inherited per-task counters. This would
+	 * create a performance issue due to all children writing to the
+	 * same buffer.
+	 */
+	if (event->cpu == -1 && event->attr.inherit)
+		return -EINVAL;
+
 	if (!(vma->vm_flags & VM_SHARED))
 		return -EINVAL;
 
@@ -2601,13 +2720,10 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 
 	WARN_ON_ONCE(event->ctx->parent_ctx);
 	mutex_lock(&event->mmap_mutex);
-	if (event->output) {
-		ret = -EINVAL;
-		goto unlock;
-	}
-
-	if (atomic_inc_not_zero(&event->mmap_count)) {
-		if (nr_pages != event->data->nr_pages)
+	if (event->data) {
+		if (event->data->nr_pages == nr_pages)
+			atomic_inc(&event->data->refcount);
+		else
 			ret = -EINVAL;
 		goto unlock;
 	}
@@ -2639,21 +2755,23 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 	WARN_ON(event->data);
 
 	data = perf_mmap_data_alloc(event, nr_pages);
-	ret = -ENOMEM;
-	if (!data)
+	if (!data) {
+		ret = -ENOMEM;
 		goto unlock;
+	}
 
-	ret = 0;
 	perf_mmap_data_init(event, data);
-
-	atomic_set(&event->mmap_count, 1);
-	atomic_long_add(user_extra, &user->locked_vm);
-	vma->vm_mm->locked_vm += extra;
-	event->data->nr_locked = extra;
 	if (vma->vm_flags & VM_WRITE)
 		event->data->writable = 1;
 
+	atomic_long_add(user_extra, &user->locked_vm);
+	event->mmap_locked = extra;
+	event->mmap_user = get_current_user();
+	vma->vm_mm->locked_vm += event->mmap_locked;
+
 unlock:
+	if (!ret)
+		atomic_inc(&event->mmap_count);
 	mutex_unlock(&event->mmap_mutex);
 
 	vma->vm_flags |= VM_RESERVED;
@@ -2885,127 +3003,87 @@ static void perf_output_wakeup(struct perf_output_handle *handle)
 }
 
 /*
- * Curious locking construct.
- *
  * We need to ensure a later event_id doesn't publish a head when a former
- * event_id isn't done writing. However since we need to deal with NMIs we
+ * event isn't done writing. However since we need to deal with NMIs we
  * cannot fully serialize things.
  *
- * What we do is serialize between CPUs so we only have to deal with NMI
- * nesting on a single CPU.
- *
  * We only publish the head (and generate a wakeup) when the outer-most
- * event_id completes.
+ * event completes.
  */
-static void perf_output_lock(struct perf_output_handle *handle)
+static void perf_output_get_handle(struct perf_output_handle *handle)
 {
 	struct perf_mmap_data *data = handle->data;
-	int cur, cpu = get_cpu();
-
-	handle->locked = 0;
 
-	for (;;) {
-		cur = atomic_cmpxchg(&data->lock, -1, cpu);
-		if (cur == -1) {
-			handle->locked = 1;
-			break;
-		}
-		if (cur == cpu)
-			break;
-
-		cpu_relax();
-	}
+	preempt_disable();
+	local_inc(&data->nest);
+	handle->wakeup = local_read(&data->wakeup);
 }
 
-static void perf_output_unlock(struct perf_output_handle *handle)
+static void perf_output_put_handle(struct perf_output_handle *handle)
 {
 	struct perf_mmap_data *data = handle->data;
 	unsigned long head;
-	int cpu;
-
-	data->done_head = data->head;
-
-	if (!handle->locked)
-		goto out;
 
 again:
-	/*
-	 * The xchg implies a full barrier that ensures all writes are done
-	 * before we publish the new head, matched by a rmb() in userspace when
-	 * reading this position.
-	 */
-	while ((head = atomic_long_xchg(&data->done_head, 0)))
-		data->user_page->data_head = head;
+	head = local_read(&data->head);
 
 	/*
-	 * NMI can happen here, which means we can miss a done_head update.
+	 * IRQ/NMI can happen here, which means we can miss a head update.
 	 */
 
-	cpu = atomic_xchg(&data->lock, -1);
-	WARN_ON_ONCE(cpu != smp_processor_id());
+	if (!local_dec_and_test(&data->nest))
+		goto out;
 
 	/*
-	 * Therefore we have to validate we did not indeed do so.
+	 * Publish the known good head. Rely on the full barrier implied
+	 * by atomic_dec_and_test() order the data->head read and this
+	 * write.
 	 */
-	if (unlikely(atomic_long_read(&data->done_head))) {
-		/*
-		 * Since we had it locked, we can lock it again.
-		 */
-		while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
-			cpu_relax();
+	data->user_page->data_head = head;
 
+	/*
+	 * Now check if we missed an update, rely on the (compiler)
+	 * barrier in atomic_dec_and_test() to re-read data->head.
+	 */
+	if (unlikely(head != local_read(&data->head))) {
+		local_inc(&data->nest);
 		goto again;
 	}
 
-	if (atomic_xchg(&data->wakeup, 0))
+	if (handle->wakeup != local_read(&data->wakeup))
 		perf_output_wakeup(handle);
-out:
-	put_cpu();
+
+ out:
+	preempt_enable();
 }
 
-void perf_output_copy(struct perf_output_handle *handle,
+__always_inline void perf_output_copy(struct perf_output_handle *handle,
 		      const void *buf, unsigned int len)
 {
-	unsigned int pages_mask;
-	unsigned long offset;
-	unsigned int size;
-	void **pages;
-
-	offset		= handle->offset;
-	pages_mask	= handle->data->nr_pages - 1;
-	pages		= handle->data->data_pages;
-
 	do {
-		unsigned long page_offset;
-		unsigned long page_size;
-		int nr;
+		unsigned long size = min_t(unsigned long, handle->size, len);
 
-		nr	    = (offset >> PAGE_SHIFT) & pages_mask;
-		page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
-		page_offset = offset & (page_size - 1);
-		size	    = min_t(unsigned int, page_size - page_offset, len);
+		memcpy(handle->addr, buf, size);
 
-		memcpy(pages[nr] + page_offset, buf, size);
+		len -= size;
+		handle->addr += size;
+		buf += size;
+		handle->size -= size;
+		if (!handle->size) {
+			struct perf_mmap_data *data = handle->data;
 
-		len	    -= size;
-		buf	    += size;
-		offset	    += size;
+			handle->page++;
+			handle->page &= data->nr_pages - 1;
+			handle->addr = data->data_pages[handle->page];
+			handle->size = PAGE_SIZE << page_order(data);
+		}
 	} while (len);
-
-	handle->offset = offset;
-
-	/*
-	 * Check we didn't copy past our reservation window, taking the
-	 * possible unsigned int wrap into account.
-	 */
-	WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
 }
 
 int perf_output_begin(struct perf_output_handle *handle,
 		      struct perf_event *event, unsigned int size,
 		      int nmi, int sample)
 {
-	struct perf_event *output_event;
 	struct perf_mmap_data *data;
 	unsigned long tail, offset, head;
 	int have_lost;
@@ -3022,10 +3100,6 @@ int perf_output_begin(struct perf_output_handle *handle,
 	if (event->parent)
 		event = event->parent;
 
-	output_event = rcu_dereference(event->output);
-	if (output_event)
-		event = output_event;
-
 	data = rcu_dereference(event->data);
 	if (!data)
 		goto out;
@@ -3036,13 +3110,13 @@ int perf_output_begin(struct perf_output_handle *handle,
 	handle->sample	= sample;
 
 	if (!data->nr_pages)
-		goto fail;
+		goto out;
 
-	have_lost = atomic_read(&data->lost);
+	have_lost = local_read(&data->lost);
 	if (have_lost)
 		size += sizeof(lost_event);
 
-	perf_output_lock(handle);
+	perf_output_get_handle(handle);
 
 	do {
 		/*
@@ -3052,24 +3126,28 @@ int perf_output_begin(struct perf_output_handle *handle,
 		 */
 		tail = ACCESS_ONCE(data->user_page->data_tail);
 		smp_rmb();
-		offset = head = atomic_long_read(&data->head);
+		offset = head = local_read(&data->head);
 		head += size;
 		if (unlikely(!perf_output_space(data, tail, offset, head)))
 			goto fail;
-	} while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
+	} while (local_cmpxchg(&data->head, offset, head) != offset);
 
-	handle->offset	= offset;
-	handle->head	= head;
+	if (head - local_read(&data->wakeup) > data->watermark)
+		local_add(data->watermark, &data->wakeup);
 
-	if (head - tail > data->watermark)
-		atomic_set(&data->wakeup, 1);
+	handle->page = offset >> (PAGE_SHIFT + page_order(data));
+	handle->page &= data->nr_pages - 1;
+	handle->size = offset & ((PAGE_SIZE << page_order(data)) - 1);
+	handle->addr = data->data_pages[handle->page];
+	handle->addr += handle->size;
+	handle->size = (PAGE_SIZE << page_order(data)) - handle->size;
 
 	if (have_lost) {
 		lost_event.header.type = PERF_RECORD_LOST;
 		lost_event.header.misc = 0;
 		lost_event.header.size = sizeof(lost_event);
 		lost_event.id          = event->id;
-		lost_event.lost        = atomic_xchg(&data->lost, 0);
+		lost_event.lost        = local_xchg(&data->lost, 0);
 
 		perf_output_put(handle, lost_event);
 	}
@@ -3077,8 +3155,8 @@ int perf_output_begin(struct perf_output_handle *handle,
 	return 0;
 
 fail:
-	atomic_inc(&data->lost);
-	perf_output_unlock(handle);
+	local_inc(&data->lost);
+	perf_output_put_handle(handle);
 out:
 	rcu_read_unlock();
 
@@ -3093,14 +3171,14 @@ void perf_output_end(struct perf_output_handle *handle)
 	int wakeup_events = event->attr.wakeup_events;
 
 	if (handle->sample && wakeup_events) {
-		int events = atomic_inc_return(&data->events);
+		int events = local_inc_return(&data->events);
 		if (events >= wakeup_events) {
-			atomic_sub(wakeup_events, &data->events);
-			atomic_set(&data->wakeup, 1);
+			local_sub(wakeup_events, &data->events);
+			local_inc(&data->wakeup);
 		}
 	}
 
-	perf_output_unlock(handle);
+	perf_output_put_handle(handle);
 	rcu_read_unlock();
 }
 
@@ -3436,22 +3514,13 @@ static void perf_event_task_output(struct perf_event *event,
 {
 	struct perf_output_handle handle;
 	struct task_struct *task = task_event->task;
-	unsigned long flags;
 	int size, ret;
 
-	/*
-	 * If this CPU attempts to acquire an rq lock held by a CPU spinning
-	 * in perf_output_lock() from interrupt context, it's game over.
-	 */
-	local_irq_save(flags);
-
 	size  = task_event->event_id.header.size;
 	ret = perf_output_begin(&handle, event, size, 0, 0);
 
-	if (ret) {
-		local_irq_restore(flags);
+	if (ret)
 		return;
-	}
 
 	task_event->event_id.pid = perf_event_pid(event, task);
 	task_event->event_id.ppid = perf_event_pid(event, current);
@@ -3462,7 +3531,6 @@ static void perf_event_task_output(struct perf_event *event,
 	perf_output_put(&handle, task_event->event_id);
 
 	perf_output_end(&handle);
-	local_irq_restore(flags);
 }
 
 static int perf_event_task_match(struct perf_event *event)
@@ -3990,13 +4058,6 @@ static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
 	}
 }
 
-static void perf_swevent_unthrottle(struct perf_event *event)
-{
-	/*
-	 * Nothing to do, we already reset hwc->interrupts.
-	 */
-}
-
 static void perf_swevent_add(struct perf_event *event, u64 nr,
 			       int nmi, struct perf_sample_data *data,
 			       struct pt_regs *regs)
@@ -4020,9 +4081,6 @@ static void perf_swevent_add(struct perf_event *event, u64 nr,
 	perf_swevent_overflow(event, 0, nmi, data, regs);
 }
 
-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)
 {
@@ -4052,10 +4110,6 @@ static int perf_swevent_match(struct perf_event *event,
 	if (perf_exclude_event(event, regs))
 		return 0;
 
-	if (event->attr.type == PERF_TYPE_TRACEPOINT &&
-	    !perf_tp_event_match(event, data))
-		return 0;
-
 	return 1;
 }
 
@@ -4066,19 +4120,46 @@ static inline u64 swevent_hash(u64 type, u32 event_id)
 	return hash_64(val, SWEVENT_HLIST_BITS);
 }
 
-static struct hlist_head *
-find_swevent_head(struct perf_cpu_context *ctx, u64 type, u32 event_id)
+static inline struct hlist_head *
+__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
 {
-	u64 hash;
-	struct swevent_hlist *hlist;
+	u64 hash = swevent_hash(type, event_id);
 
-	hash = swevent_hash(type, event_id);
+	return &hlist->heads[hash];
+}
+
+/* For the read side: events when they trigger */
+static inline struct hlist_head *
+find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
+{
+	struct swevent_hlist *hlist;
 
 	hlist = rcu_dereference(ctx->swevent_hlist);
 	if (!hlist)
 		return NULL;
 
-	return &hlist->heads[hash];
+	return __find_swevent_head(hlist, type, event_id);
+}
+
+/* For the event head insertion and removal in the hlist */
+static inline struct hlist_head *
+find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
+{
+	struct swevent_hlist *hlist;
+	u32 event_id = event->attr.config;
+	u64 type = event->attr.type;
+
+	/*
+	 * Event scheduling is always serialized against hlist allocation
+	 * and release. Which makes the protected version suitable here.
+	 * The context lock guarantees that.
+	 */
+	hlist = rcu_dereference_protected(ctx->swevent_hlist,
+					  lockdep_is_held(&event->ctx->lock));
+	if (!hlist)
+		return NULL;
+
+	return __find_swevent_head(hlist, type, event_id);
 }
 
 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
@@ -4095,7 +4176,7 @@ static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
 
 	rcu_read_lock();
 
-	head = find_swevent_head(cpuctx, type, event_id);
+	head = find_swevent_head_rcu(cpuctx, type, event_id);
 
 	if (!head)
 		goto end;
@@ -4110,7 +4191,7 @@ end:
 
 int perf_swevent_get_recursion_context(void)
 {
-	struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
+	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
 	int rctx;
 
 	if (in_nmi())
@@ -4122,10 +4203,8 @@ int perf_swevent_get_recursion_context(void)
 	else
 		rctx = 0;
 
-	if (cpuctx->recursion[rctx]) {
-		put_cpu_var(perf_cpu_context);
+	if (cpuctx->recursion[rctx])
 		return -1;
-	}
 
 	cpuctx->recursion[rctx]++;
 	barrier();
@@ -4139,7 +4218,6 @@ void perf_swevent_put_recursion_context(int rctx)
 	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
 	barrier();
 	cpuctx->recursion[rctx]--;
-	put_cpu_var(perf_cpu_context);
 }
 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
 
@@ -4150,6 +4228,7 @@ void __perf_sw_event(u32 event_id, u64 nr, int nmi,
 	struct perf_sample_data data;
 	int rctx;
 
+	preempt_disable_notrace();
 	rctx = perf_swevent_get_recursion_context();
 	if (rctx < 0)
 		return;
@@ -4159,6 +4238,7 @@ void __perf_sw_event(u32 event_id, u64 nr, int nmi,
 	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
 
 	perf_swevent_put_recursion_context(rctx);
+	preempt_enable_notrace();
 }
 
 static void perf_swevent_read(struct perf_event *event)
@@ -4178,7 +4258,7 @@ static int perf_swevent_enable(struct perf_event *event)
 		perf_swevent_set_period(event);
 	}
 
-	head = find_swevent_head(cpuctx, event->attr.type, event->attr.config);
+	head = find_swevent_head(cpuctx, event);
 	if (WARN_ON_ONCE(!head))
 		return -EINVAL;
 
@@ -4192,11 +4272,22 @@ static void perf_swevent_disable(struct perf_event *event)
 	hlist_del_rcu(&event->hlist_entry);
 }
 
+static void perf_swevent_void(struct perf_event *event)
+{
+}
+
+static int perf_swevent_int(struct perf_event *event)
+{
+	return 0;
+}
+
 static const struct pmu perf_ops_generic = {
 	.enable		= perf_swevent_enable,
 	.disable	= perf_swevent_disable,
+	.start		= perf_swevent_int,
+	.stop		= perf_swevent_void,
 	.read		= perf_swevent_read,
-	.unthrottle	= perf_swevent_unthrottle,
+	.unthrottle	= perf_swevent_void, /* hwc->interrupts already reset */
 };
 
 /*
@@ -4366,6 +4457,14 @@ static const struct pmu perf_ops_task_clock = {
 	.read		= task_clock_perf_event_read,
 };
 
+/* Deref the hlist from the update side */
+static inline struct swevent_hlist *
+swevent_hlist_deref(struct perf_cpu_context *cpuctx)
+{
+	return rcu_dereference_protected(cpuctx->swevent_hlist,
+					 lockdep_is_held(&cpuctx->hlist_mutex));
+}
+
 static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
 {
 	struct swevent_hlist *hlist;
@@ -4376,12 +4475,11 @@ static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
 
 static void swevent_hlist_release(struct perf_cpu_context *cpuctx)
 {
-	struct swevent_hlist *hlist;
+	struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx);
 
-	if (!cpuctx->swevent_hlist)
+	if (!hlist)
 		return;
 
-	hlist = cpuctx->swevent_hlist;
 	rcu_assign_pointer(cpuctx->swevent_hlist, NULL);
 	call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
 }
@@ -4418,7 +4516,7 @@ static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
 
 	mutex_lock(&cpuctx->hlist_mutex);
 
-	if (!cpuctx->swevent_hlist && cpu_online(cpu)) {
+	if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) {
 		struct swevent_hlist *hlist;
 
 		hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
@@ -4467,10 +4565,48 @@ static int swevent_hlist_get(struct perf_event *event)
 
 #ifdef CONFIG_EVENT_TRACING
 
-void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
-		   int entry_size, struct pt_regs *regs)
+static const struct pmu perf_ops_tracepoint = {
+	.enable		= perf_trace_enable,
+	.disable	= perf_trace_disable,
+	.start		= perf_swevent_int,
+	.stop		= perf_swevent_void,
+	.read		= perf_swevent_read,
+	.unthrottle	= perf_swevent_void,
+};
+
+static int perf_tp_filter_match(struct perf_event *event,
+				struct perf_sample_data *data)
+{
+	void *record = data->raw->data;
+
+	if (likely(!event->filter) || filter_match_preds(event->filter, record))
+		return 1;
+	return 0;
+}
+
+static int perf_tp_event_match(struct perf_event *event,
+				struct perf_sample_data *data,
+				struct pt_regs *regs)
+{
+	/*
+	 * All tracepoints are from kernel-space.
+	 */
+	if (event->attr.exclude_kernel)
+		return 0;
+
+	if (!perf_tp_filter_match(event, data))
+		return 0;
+
+	return 1;
+}
+
+void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
+		   struct pt_regs *regs, struct hlist_head *head)
 {
 	struct perf_sample_data data;
+	struct perf_event *event;
+	struct hlist_node *node;
+
 	struct perf_raw_record raw = {
 		.size = entry_size,
 		.data = record,
@@ -4479,26 +4615,18 @@ void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
 	perf_sample_data_init(&data, addr);
 	data.raw = &raw;
 
-	/* Trace events already protected against recursion */
-	do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
-			 &data, regs);
+	rcu_read_lock();
+	hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
+		if (perf_tp_event_match(event, &data, regs))
+			perf_swevent_add(event, count, 1, &data, regs);
+	}
+	rcu_read_unlock();
 }
 EXPORT_SYMBOL_GPL(perf_tp_event);
 
-static int perf_tp_event_match(struct perf_event *event,
-				struct perf_sample_data *data)
-{
-	void *record = data->raw->data;
-
-	if (likely(!event->filter) || filter_match_preds(event->filter, record))
-		return 1;
-	return 0;
-}
-
 static void tp_perf_event_destroy(struct perf_event *event)
 {
-	perf_trace_disable(event->attr.config);
-	swevent_hlist_put(event);
+	perf_trace_destroy(event);
 }
 
 static const struct pmu *tp_perf_event_init(struct perf_event *event)
@@ -4514,17 +4642,13 @@ static const struct pmu *tp_perf_event_init(struct perf_event *event)
 			!capable(CAP_SYS_ADMIN))
 		return ERR_PTR(-EPERM);
 
-	if (perf_trace_enable(event->attr.config))
+	err = perf_trace_init(event);
+	if (err)
 		return NULL;
 
 	event->destroy = tp_perf_event_destroy;
-	err = swevent_hlist_get(event);
-	if (err) {
-		perf_trace_disable(event->attr.config);
-		return ERR_PTR(err);
-	}
 
-	return &perf_ops_generic;
+	return &perf_ops_tracepoint;
 }
 
 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
@@ -4552,12 +4676,6 @@ static void perf_event_free_filter(struct perf_event *event)
 
 #else
 
-static int perf_tp_event_match(struct perf_event *event,
-				struct perf_sample_data *data)
-{
-	return 1;
-}
-
 static const struct pmu *tp_perf_event_init(struct perf_event *event)
 {
 	return NULL;
@@ -4886,54 +5004,53 @@ err_size:
 	goto out;
 }
 
-static int perf_event_set_output(struct perf_event *event, int output_fd)
+static int
+perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 {
-	struct perf_event *output_event = NULL;
-	struct file *output_file = NULL;
-	struct perf_event *old_output;
-	int fput_needed = 0;
+	struct perf_mmap_data *data = NULL, *old_data = NULL;
 	int ret = -EINVAL;
 
-	if (!output_fd)
+	if (!output_event)
 		goto set;
 
-	output_file = fget_light(output_fd, &fput_needed);
-	if (!output_file)
-		return -EBADF;
-
-	if (output_file->f_op != &perf_fops)
+	/* don't allow circular references */
+	if (event == output_event)
 		goto out;
 
-	output_event = output_file->private_data;
-
-	/* Don't chain output fds */
-	if (output_event->output)
+	/*
+	 * Don't allow cross-cpu buffers
+	 */
+	if (output_event->cpu != event->cpu)
 		goto out;
 
-	/* Don't set an output fd when we already have an output channel */
-	if (event->data)
+	/*
+	 * If its not a per-cpu buffer, it must be the same task.
+	 */
+	if (output_event->cpu == -1 && output_event->ctx != event->ctx)
 		goto out;
 
-	atomic_long_inc(&output_file->f_count);
-
 set:
 	mutex_lock(&event->mmap_mutex);
-	old_output = event->output;
-	rcu_assign_pointer(event->output, output_event);
-	mutex_unlock(&event->mmap_mutex);
+	/* Can't redirect output if we've got an active mmap() */
+	if (atomic_read(&event->mmap_count))
+		goto unlock;
 
-	if (old_output) {
-		/*
-		 * we need to make sure no existing perf_output_*()
-		 * is still referencing this event.
-		 */
-		synchronize_rcu();
-		fput(old_output->filp);
+	if (output_event) {
+		/* get the buffer we want to redirect to */
+		data = perf_mmap_data_get(output_event);
+		if (!data)
+			goto unlock;
 	}
 
+	old_data = event->data;
+	rcu_assign_pointer(event->data, data);
 	ret = 0;
+unlock:
+	mutex_unlock(&event->mmap_mutex);
+
+	if (old_data)
+		perf_mmap_data_put(old_data);
 out:
-	fput_light(output_file, fput_needed);
 	return ret;
 }
 
@@ -4949,13 +5066,13 @@ SYSCALL_DEFINE5(perf_event_open,
 		struct perf_event_attr __user *, attr_uptr,
 		pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
 {
-	struct perf_event *event, *group_leader;
+	struct perf_event *event, *group_leader = NULL, *output_event = NULL;
 	struct perf_event_attr attr;
 	struct perf_event_context *ctx;
 	struct file *event_file = NULL;
 	struct file *group_file = NULL;
+	int event_fd;
 	int fput_needed = 0;
-	int fput_needed2 = 0;
 	int err;
 
 	/* for future expandability... */
@@ -4976,26 +5093,38 @@ SYSCALL_DEFINE5(perf_event_open,
 			return -EINVAL;
 	}
 
+	event_fd = get_unused_fd_flags(O_RDWR);
+	if (event_fd < 0)
+		return event_fd;
+
 	/*
 	 * Get the target context (task or percpu):
 	 */
 	ctx = find_get_context(pid, cpu);
-	if (IS_ERR(ctx))
-		return PTR_ERR(ctx);
+	if (IS_ERR(ctx)) {
+		err = PTR_ERR(ctx);
+		goto err_fd;
+	}
+
+	if (group_fd != -1) {
+		group_leader = perf_fget_light(group_fd, &fput_needed);
+		if (IS_ERR(group_leader)) {
+			err = PTR_ERR(group_leader);
+			goto err_put_context;
+		}
+		group_file = group_leader->filp;
+		if (flags & PERF_FLAG_FD_OUTPUT)
+			output_event = group_leader;
+		if (flags & PERF_FLAG_FD_NO_GROUP)
+			group_leader = NULL;
+	}
 
 	/*
 	 * Look up the group leader (we will attach this event to it):
 	 */
-	group_leader = NULL;
-	if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
+	if (group_leader) {
 		err = -EINVAL;
-		group_file = fget_light(group_fd, &fput_needed);
-		if (!group_file)
-			goto err_put_context;
-		if (group_file->f_op != &perf_fops)
-			goto err_put_context;
 
-		group_leader = group_file->private_data;
 		/*
 		 * Do not allow a recursive hierarchy (this new sibling
 		 * becoming part of another group-sibling):
@@ -5017,22 +5146,21 @@ SYSCALL_DEFINE5(perf_event_open,
 
 	event = perf_event_alloc(&attr, cpu, ctx, group_leader,
 				     NULL, NULL, GFP_KERNEL);
-	err = PTR_ERR(event);
-	if (IS_ERR(event))
+	if (IS_ERR(event)) {
+		err = PTR_ERR(event);
 		goto err_put_context;
+	}
 
-	err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
-	if (err < 0)
-		goto err_free_put_context;
+	if (output_event) {
+		err = perf_event_set_output(event, output_event);
+		if (err)
+			goto err_free_put_context;
+	}
 
-	event_file = fget_light(err, &fput_needed2);
-	if (!event_file)
+	event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
+	if (IS_ERR(event_file)) {
+		err = PTR_ERR(event_file);
 		goto err_free_put_context;
-
-	if (flags & PERF_FLAG_FD_OUTPUT) {
-		err = perf_event_set_output(event, group_fd);
-		if (err)
-			goto err_fput_free_put_context;
 	}
 
 	event->filp = event_file;
@@ -5048,19 +5176,23 @@ SYSCALL_DEFINE5(perf_event_open,
 	list_add_tail(&event->owner_entry, &current->perf_event_list);
 	mutex_unlock(&current->perf_event_mutex);
 
-err_fput_free_put_context:
-	fput_light(event_file, fput_needed2);
+	/*
+	 * Drop the reference on the group_event after placing the
+	 * new event on the sibling_list. This ensures destruction
+	 * of the group leader will find the pointer to itself in
+	 * perf_group_detach().
+	 */
+	fput_light(group_file, fput_needed);
+	fd_install(event_fd, event_file);
+	return event_fd;
 
 err_free_put_context:
-	if (err < 0)
-		free_event(event);
-
+	free_event(event);
 err_put_context:
-	if (err < 0)
-		put_ctx(ctx);
-
 	fput_light(group_file, fput_needed);
-
+	put_ctx(ctx);
+err_fd:
+	put_unused_fd(event_fd);
 	return err;
 }
 
@@ -5371,6 +5503,7 @@ static void perf_free_event(struct perf_event *event,
 
 	fput(parent->filp);
 
+	perf_group_detach(event);
 	list_del_event(event, ctx);
 	free_event(event);
 }