perf_counter: Collapse inherit on read()

Currently the counter value returned by read() is the value of
the parent counter, to which child counters are only fed back
on child exit.

Thus read() can return rather erratic (and meaningless) numbers
depending on the state of the child processes.

Change this by always iterating the full child hierarchy on
read() and sum all counters.

Suggested-by: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 9509310419545e03d015f851169436921c2fbeb3..48471d75ae0169abd5d92a1cf1b1caae3de392fb 100644 (file)
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1688,6 +1688,18 @@ static int perf_release(struct inode *inode, struct file *file)
        return 0;
 }
 
+static u64 perf_counter_read_tree(struct perf_counter *counter)
+{
+       struct perf_counter *child;
+       u64 total = 0;
+
+       total += perf_counter_read(counter);
+       list_for_each_entry(child, &counter->child_list, child_list)
+               total += perf_counter_read(child);
+
+       return total;
+}
+
 /*
  * Read the performance counter - simple non blocking version for now
  */
@@ -1707,7 +1719,7 @@ perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
 
        WARN_ON_ONCE(counter->ctx->parent_ctx);
        mutex_lock(&counter->child_mutex);
-       values[0] = perf_counter_read(counter);
+       values[0] = perf_counter_read_tree(counter);
        n = 1;
        if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
                values[n++] = counter->total_time_enabled +