perf: Optimize the perf_output() path by removing IRQ-disables
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Tue, 18 May 2010 08:50:41 +0000 (10:50 +0200)
committerIngo Molnar <mingo@elte.hu>
Tue, 18 May 2010 16:35:48 +0000 (18:35 +0200)
Since we can now assume there is only a single writer
to each buffer, we can remove per-cpu lock thingy and
use a simply nest-count to the same effect.

This removes the need to disable IRQs.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
include/linux/perf_event.h
kernel/perf_event.c

index 0b521fc8f5b0166599d48d2b952eb6ea30b6fd5e..f1f853a9d5eb0cbd0fbca519b9a4904c94b27e07 100644 (file)
@@ -597,12 +597,12 @@ struct perf_mmap_data {
        atomic_t                        events;         /* event_id limit       */
 
        atomic_long_t                   head;           /* write position    */
-       atomic_long_t                   done_head;      /* completed head    */
 
-       atomic_t                        lock;           /* concurrent writes */
        atomic_t                        wakeup;         /* needs a wakeup    */
        atomic_t                        lost;           /* nr records lost   */
 
+       atomic_t                        nest;           /* nested writers    */
+
        long                            watermark;      /* wakeup watermark  */
 
        struct perf_event_mmap_page     *user_page;
@@ -807,7 +807,6 @@ struct perf_output_handle {
        unsigned long                   offset;
        int                             nmi;
        int                             sample;
-       int                             locked;
 };
 
 #ifdef CONFIG_PERF_EVENTS
index ff5d430d45a74d09898fe1fed399105baea83079..8cf737da3ec4926c23e0a8f58cd137a94c482950 100644 (file)
@@ -2519,8 +2519,6 @@ 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);
@@ -2906,82 +2904,56 @@ 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();
+       atomic_inc(&data->nest);
 }
 
-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 = atomic_long_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 (!atomic_dec_and_test(&data->nest))
+               return;
 
        /*
-        * 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 != atomic_long_read(&data->head))) {
+               atomic_inc(&data->nest);
                goto again;
        }
 
        if (atomic_xchg(&data->wakeup, 0))
                perf_output_wakeup(handle);
-out:
-       put_cpu();
+
+       preempt_enable();
 }
 
 void perf_output_copy(struct perf_output_handle *handle,
@@ -3063,7 +3035,7 @@ int perf_output_begin(struct perf_output_handle *handle,
        if (have_lost)
                size += sizeof(lost_event);
 
-       perf_output_lock(handle);
+       perf_output_get_handle(handle);
 
        do {
                /*
@@ -3083,7 +3055,7 @@ int perf_output_begin(struct perf_output_handle *handle,
        handle->head    = head;
 
        if (head - tail > data->watermark)
-               atomic_set(&data->wakeup, 1);
+               atomic_inc(&data->wakeup);
 
        if (have_lost) {
                lost_event.header.type = PERF_RECORD_LOST;
@@ -3099,7 +3071,7 @@ int perf_output_begin(struct perf_output_handle *handle,
 
 fail:
        atomic_inc(&data->lost);
-       perf_output_unlock(handle);
+       perf_output_put_handle(handle);
 out:
        rcu_read_unlock();
 
@@ -3117,11 +3089,11 @@ void perf_output_end(struct perf_output_handle *handle)
                int events = atomic_inc_return(&data->events);
                if (events >= wakeup_events) {
                        atomic_sub(wakeup_events, &data->events);
-                       atomic_set(&data->wakeup, 1);
+                       atomic_inc(&data->wakeup);
                }
        }
 
-       perf_output_unlock(handle);
+       perf_output_put_handle(handle);
        rcu_read_unlock();
 }
 
@@ -3457,22 +3429,13 @@ 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);
@@ -3483,7 +3446,6 @@ 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)