perf_counter: PERF_SAMPLE_ID and inherited counters

Anton noted that for inherited counters the counter-id as provided by
PERF_SAMPLE_ID isn't mappable to the id found through PERF_RECORD_ID
because each inherited counter gets its own id.

His suggestion was to always return the parent counter id, since that
is the primary counter id as exposed. However, these inherited
counters have a unique identifier so that events like
PERF_EVENT_PERIOD and PERF_EVENT_THROTTLE can be specific about which
counter gets modified, which is important when trying to normalize the
sample streams.

This patch removes PERF_EVENT_PERIOD in favour of PERF_SAMPLE_PERIOD,
which is more useful anyway, since changing periods became a lot more
common than initially thought -- rendering PERF_EVENT_PERIOD the less
useful solution (also, PERF_SAMPLE_PERIOD reports the more accurate
value, since it reports the value used to trigger the overflow,
whereas PERF_EVENT_PERIOD simply reports the requested period changed,
which might only take effect on the next cycle).

This still leaves us PERF_EVENT_THROTTLE to consider, but since that
_should_ be a rare occurrence, and linking it to a primary id is the
most useful bit to diagnose the problem, we introduce a
PERF_SAMPLE_STREAM_ID, for those few cases where the full
reconstruction is important.

[Does change the ABI a little, but I see no other way out]

Suggested-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1248095846.15751.8781.camel@twins>

1 files changed, 31 insertions, 61 deletions

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index e1d6a3aa1333..7530588fa5c5 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -155,6 +155,20 @@ static void unclone_ctx(struct perf_counter_context *ctx)
 }
 
 /*
+ * If we inherit counters we want to return the parent counter id
+ * to userspace.
+ */
+static u64 primary_counter_id(struct perf_counter *counter)
+{
+        u64 id = counter->id;
+
+        if (counter->parent)
+                id = counter->parent->id;
+
+        return id;
+}
+
+/*
  * Get the perf_counter_context for a task and lock it.
  * This has to cope with with the fact that until it is locked,
  * the context could get moved to another task.
@@ -1296,7 +1310,6 @@ static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
 #define MAX_INTERRUPTS (~0ULL)
 
 static void perf_log_throttle(struct perf_counter *counter, int enable);
-static void perf_log_period(struct perf_counter *counter, u64 period);
 
 static void perf_adjust_period(struct perf_counter *counter, u64 events)
 {
@@ -1315,8 +1328,6 @@ static void perf_adjust_period(struct perf_counter *counter, u64 events)
         if (!sample_period)
                 sample_period = 1;
 
-        perf_log_period(counter, sample_period);
-
         hwc->sample_period = sample_period;
 }
 
@@ -1705,7 +1716,7 @@ perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
                 values[n++] = counter->total_time_running +
                         atomic64_read(&counter->child_total_time_running);
         if (counter->attr.read_format & PERF_FORMAT_ID)
-                values[n++] = counter->id;
+                values[n++] = primary_counter_id(counter);
         mutex_unlock(&counter->child_mutex);
 
         if (count < n * sizeof(u64))
@@ -1812,8 +1823,6 @@ static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
 
                 counter->attr.sample_freq = value;
         } else {
-                perf_log_period(counter, value);
-
                 counter->attr.sample_period = value;
                 counter->hw.sample_period = value;
         }
@@ -2662,6 +2671,9 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
         if (sample_type & PERF_SAMPLE_ID)
                 header.size += sizeof(u64);
 
+        if (sample_type & PERF_SAMPLE_STREAM_ID)
+                header.size += sizeof(u64);
+
         if (sample_type & PERF_SAMPLE_CPU) {
                 header.size += sizeof(cpu_entry);
 
@@ -2705,7 +2717,13 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
         if (sample_type & PERF_SAMPLE_ADDR)
                 perf_output_put(&handle, data->addr);
 
-        if (sample_type & PERF_SAMPLE_ID)
+        if (sample_type & PERF_SAMPLE_ID) {
+                u64 id = primary_counter_id(counter);
+
+                perf_output_put(&handle, id);
+        }
+
+        if (sample_type & PERF_SAMPLE_STREAM_ID)
                 perf_output_put(&handle, counter->id);
 
         if (sample_type & PERF_SAMPLE_CPU)
@@ -2728,7 +2746,7 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
                         if (sub != counter)
                                 sub->pmu->read(sub);
 
-                        group_entry.id = sub->id;
+                        group_entry.id = primary_counter_id(sub);
                         group_entry.counter = atomic64_read(&sub->count);
 
                         perf_output_put(&handle, group_entry);
@@ -2788,15 +2806,8 @@ perf_counter_read_event(struct perf_counter *counter,
         }
 
         if (counter->attr.read_format & PERF_FORMAT_ID) {
-                u64 id;
-
                 event.header.size += sizeof(u64);
-                if (counter->parent)
-                        id = counter->parent->id;
-                else
-                        id = counter->id;
-
-                event.format[i++] = id;
+                event.format[i++] = primary_counter_id(counter);
         }
 
         ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
@@ -3191,49 +3202,6 @@ void __perf_counter_mmap(struct vm_area_struct *vma)
 }
 
 /*
- * Log sample_period changes so that analyzing tools can re-normalize the
- * event flow.
- */
-
-struct freq_event {
-        struct perf_event_header        header;
-        u64                             time;
-        u64                             id;
-        u64                             period;
-};
-
-static void perf_log_period(struct perf_counter *counter, u64 period)
-{
-        struct perf_output_handle handle;
-        struct freq_event event;
-        int ret;
-
-        if (counter->hw.sample_period == period)
-                return;
-
-        if (counter->attr.sample_type & PERF_SAMPLE_PERIOD)
-                return;
-
-        event = (struct freq_event) {
-                .header = {
-                        .type = PERF_EVENT_PERIOD,
-                        .misc = 0,
-                        .size = sizeof(event),
-                },
-                .time = sched_clock(),
-                .id = counter->id,
-                .period = period,
-        };
-
-        ret = perf_output_begin(&handle, counter, sizeof(event), 1, 0);
-        if (ret)
-                return;
-
-        perf_output_put(&handle, event);
-        perf_output_end(&handle);
-}
-
-/*
  * IRQ throttle logging
  */
 
@@ -3246,14 +3214,16 @@ static void perf_log_throttle(struct perf_counter *counter, int enable)
                 struct perf_event_header        header;
                 u64                             time;
                 u64                             id;
+                u64                             stream_id;
         } throttle_event = {
                 .header = {
                         .type = PERF_EVENT_THROTTLE + 1,
                         .misc = 0,
                         .size = sizeof(throttle_event),
                 },
-                .time   = sched_clock(),
-                .id     = counter->id,
+                .time           = sched_clock(),
+                .id             = primary_counter_id(counter),
+                .stream_id      = counter->id,
         };
 
         ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);