perf_counter: inheritable sample counters
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Tue, 5 May 2009 15:50:26 +0000 (17:50 +0200)
committerIngo Molnar <mingo@elte.hu>
Tue, 5 May 2009 18:18:33 +0000 (20:18 +0200)
Redirect the output to the parent counter and put in some sanity checks.

[ Impact: new perfcounter feature - inherited sampling counters ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <20090505155437.331556171@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/perf_counter.c

index c881afef997b94c1c2b9fcb05c6c3a3e2cfb8996..60e55f0b48f4b94b4ed4d3a8bb9f26935903f161 100644 (file)
@@ -738,10 +738,18 @@ static void perf_counter_enable(struct perf_counter *counter)
        spin_unlock_irq(&ctx->lock);
 }
 
-static void perf_counter_refresh(struct perf_counter *counter, int refresh)
+static int perf_counter_refresh(struct perf_counter *counter, int refresh)
 {
+       /*
+        * not supported on inherited counters
+        */
+       if (counter->hw_event.inherit)
+               return -EINVAL;
+
        atomic_add(refresh, &counter->event_limit);
        perf_counter_enable(counter);
+
+       return 0;
 }
 
 /*
@@ -1307,7 +1315,7 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
                perf_counter_disable_family(counter);
                break;
        case PERF_COUNTER_IOC_REFRESH:
-               perf_counter_refresh(counter, arg);
+               err = perf_counter_refresh(counter, arg);
                break;
        case PERF_COUNTER_IOC_RESET:
                perf_counter_reset(counter);
@@ -1814,6 +1822,12 @@ static int perf_output_begin(struct perf_output_handle *handle,
        struct perf_mmap_data *data;
        unsigned int offset, head;
 
+       /*
+        * For inherited counters we send all the output towards the parent.
+        */
+       if (counter->parent)
+               counter = counter->parent;
+
        rcu_read_lock();
        data = rcu_dereference(counter->data);
        if (!data)
@@ -1995,6 +2009,9 @@ static void perf_counter_output(struct perf_counter *counter,
        if (record_type & PERF_RECORD_ADDR)
                perf_output_put(&handle, addr);
 
+       /*
+        * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
+        */
        if (record_type & PERF_RECORD_GROUP) {
                struct perf_counter *leader, *sub;
                u64 nr = counter->nr_siblings;
@@ -2281,6 +2298,11 @@ int perf_counter_overflow(struct perf_counter *counter,
        int events = atomic_read(&counter->event_limit);
        int ret = 0;
 
+       /*
+        * XXX event_limit might not quite work as expected on inherited
+        * counters
+        */
+
        counter->pending_kill = POLL_IN;
        if (events && atomic_dec_and_test(&counter->event_limit)) {
                ret = 1;
@@ -2801,6 +2823,12 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event,
 
        pmu = NULL;
 
+       /*
+        * we currently do not support PERF_RECORD_GROUP on inherited counters
+        */
+       if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
+               goto done;
+
        if (perf_event_raw(hw_event)) {
                pmu = hw_perf_counter_init(counter);
                goto done;