perf_counter: unify and fix delayed counter wakeup

While going over the wakeup code I noticed delayed wakeups only work
for hardware counters but basically all software counters rely on
them.

This patch unifies and generalizes the delayed wakeup to fix this
issue.

Since we're dealing with NMI context bits here, use a cmpxchg() based
single link list implementation to track counters that have pending
wakeups.

[ This should really be generic code for delayed wakeups, but since we
  cannot use cmpxchg()/xchg() in generic code, I've let it live in the
  perf_counter code. -- Eric Dumazet could use it to aggregate the
  network wakeups. ]

Furthermore, the x86 method of using TIF flags was flawed in that its
quite possible to end up setting the bit on the idle task, loosing the
wakeup.

The powerpc method uses per-cpu storage and does appear to be
sufficient.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Paul Mackerras <paulus@samba.org>
Orig-LKML-Reference: <20090330171023.153932974@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

2 files changed, 123 insertions, 8 deletions

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 3b862a7988c..f70ff80e79d 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1197,8 +1197,12 @@ static void free_counter_rcu(struct rcu_head *head)
 	kfree(counter);
 }
 
+static void perf_pending_sync(struct perf_counter *counter);
+
 static void free_counter(struct perf_counter *counter)
 {
+	perf_pending_sync(counter);
+
 	if (counter->destroy)
 		counter->destroy(counter);
 
@@ -1529,6 +1533,118 @@ static const struct file_operations perf_fops = {
 };
 
 /*
+ * Perf counter wakeup
+ *
+ * If there's data, ensure we set the poll() state and publish everything
+ * to user-space before waking everybody up.
+ */
+
+void perf_counter_wakeup(struct perf_counter *counter)
+{
+	struct perf_mmap_data *data;
+
+	rcu_read_lock();
+	data = rcu_dereference(counter->data);
+	if (data) {
+		(void)atomic_xchg(&data->wakeup, POLL_IN);
+		__perf_counter_update_userpage(counter, data);
+	}
+	rcu_read_unlock();
+
+	wake_up_all(&counter->waitq);
+}
+
+/*
+ * Pending wakeups
+ *
+ * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
+ *
+ * The NMI bit means we cannot possibly take locks. Therefore, maintain a
+ * single linked list and use cmpxchg() to add entries lockless.
+ */
+
+#define PENDING_TAIL ((struct perf_wakeup_entry *)-1UL)
+
+static DEFINE_PER_CPU(struct perf_wakeup_entry *, perf_wakeup_head) = {
+	PENDING_TAIL,
+};
+
+static void perf_pending_queue(struct perf_counter *counter)
+{
+	struct perf_wakeup_entry **head;
+	struct perf_wakeup_entry *prev, *next;
+
+	if (cmpxchg(&counter->wakeup.next, NULL, PENDING_TAIL) != NULL)
+		return;
+
+	head = &get_cpu_var(perf_wakeup_head);
+
+	do {
+		prev = counter->wakeup.next = *head;
+		next = &counter->wakeup;
+	} while (cmpxchg(head, prev, next) != prev);
+
+	set_perf_counter_pending();
+
+	put_cpu_var(perf_wakeup_head);
+}
+
+static int __perf_pending_run(void)
+{
+	struct perf_wakeup_entry *list;
+	int nr = 0;
+
+	list = xchg(&__get_cpu_var(perf_wakeup_head), PENDING_TAIL);
+	while (list != PENDING_TAIL) {
+		struct perf_counter *counter = container_of(list,
+				struct perf_counter, wakeup);
+
+		list = list->next;
+
+		counter->wakeup.next = NULL;
+		/*
+		 * Ensure we observe the unqueue before we issue the wakeup,
+		 * so that we won't be waiting forever.
+		 * -- see perf_not_pending().
+		 */
+		smp_wmb();
+
+		perf_counter_wakeup(counter);
+		nr++;
+	}
+
+	return nr;
+}
+
+static inline int perf_not_pending(struct perf_counter *counter)
+{
+	/*
+	 * If we flush on whatever cpu we run, there is a chance we don't
+	 * need to wait.
+	 */
+	get_cpu();
+	__perf_pending_run();
+	put_cpu();
+
+	/*
+	 * Ensure we see the proper queue state before going to sleep
+	 * so that we do not miss the wakeup. -- see perf_pending_handle()
+	 */
+	smp_rmb();
+	return counter->wakeup.next == NULL;
+}
+
+static void perf_pending_sync(struct perf_counter *counter)
+{
+	wait_event(counter->waitq, perf_not_pending(counter));
+}
+
+void perf_counter_do_pending(void)
+{
+	__perf_pending_run();
+}
+
+/*
  * Output
  */
 
@@ -1611,13 +1727,10 @@ static void perf_output_copy(struct perf_output_handle *handle,
 static void perf_output_end(struct perf_output_handle *handle, int nmi)
 {
 	if (handle->wakeup) {
-		(void)atomic_xchg(&handle->data->wakeup, POLL_IN);
-		__perf_counter_update_userpage(handle->counter, handle->data);
-		if (nmi) {
-			handle->counter->wakeup_pending = 1;
-			set_perf_counter_pending();
-		} else
-			wake_up(&handle->counter->waitq);
+		if (nmi)
+			perf_pending_queue(handle->counter);
+		else
+			perf_counter_wakeup(handle->counter);
 	}
 	rcu_read_unlock();
 }
@@ -2211,7 +2324,6 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event,
 
 	counter->cpu			= cpu;
 	counter->hw_event		= *hw_event;
-	counter->wakeup_pending		= 0;
 	counter->group_leader		= group_leader;
 	counter->hw_ops			= NULL;
 	counter->ctx			= ctx;
diff --git a/kernel/timer.c b/kernel/timer.c
index b4555568b4e..672ca25fbc4 100644
--- a/kernel/timer.c
+++ b/kernel/timer.c
@@ -37,6 +37,7 @@
 #include <linux/delay.h>
 #include <linux/tick.h>
 #include <linux/kallsyms.h>
+#include <linux/perf_counter.h>
 
 #include <asm/uaccess.h>
 #include <asm/unistd.h>
@@ -1167,6 +1168,8 @@ static void run_timer_softirq(struct softirq_action *h)
 {
 	struct tvec_base *base = __get_cpu_var(tvec_bases);
 
+	perf_counter_do_pending();
+
 	hrtimer_run_pending();
 
 	if (time_after_eq(jiffies, base->timer_jiffies))