Currently we use perf_event_context::task_ctx_data to save
and restore the LBR status when the task is scheduled out
and in.
We don't allocate it for child contexts, which results in
shorter task's LBR stack, because we don't save the history
from previous run and start over every time we schedule the
task in.
I made a test to generate samples with LBR call stack
and got higher numbers on bigger chain depths:
before: after:
LBR call chain: nr: 1 60561 498127
LBR call chain: nr: 2 0 0
LBR call chain: nr: 3 107030 2172
LBR call chain: nr: 4 466685 62758
LBR call chain: nr: 5 2307319 878046
LBR call chain: nr: 6 48713 495218
LBR call chain: nr: 7 1040 4551
LBR call chain: nr: 8 481 172
LBR call chain: nr: 9 878 120
LBR call chain: nr: 10 2377 6698
LBR call chain: nr: 11 28830 151487
LBR call chain: nr: 12 29347 339867
LBR call chain: nr: 13 4 22
LBR call chain: nr: 14 3 53
Cc: Andi Kleen <a...@linux.intel.com>
Signed-off-by: Jiri Olsa <jo...@kernel.org>
---
kernel/events/core.c | 14 ++++++++++++++
1 file changed, 14 insertions(+)
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 4df5b695bf0d..55fb648a32b0 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -10703,6 +10703,19 @@ inherit_event(struct perf_event *parent_event,
if (IS_ERR(child_event))
return child_event;
+
+ if ((child_event->attach_state & PERF_ATTACH_TASK_DATA) &&
+ !child_ctx->task_ctx_data) {
+ struct pmu *pmu = child_event->pmu;
+
+ child_ctx->task_ctx_data = kzalloc(pmu->task_ctx_size,
+ GFP_KERNEL);
+ if (!child_ctx->task_ctx_data) {
+ free_event(child_event);
+ return NULL;
+ }
+ }
+
/*
* is_orphaned_event() and list_add_tail(&parent_event->child_list)
* must be under the same lock in order to serialize against
@@ -10713,6 +10726,7 @@ inherit_event(struct perf_event *parent_event,
if (is_orphaned_event(parent_event) ||
!atomic_long_inc_not_zero(&parent_event->refcount)) {
mutex_unlock(&parent_event->child_mutex);
+ /* task_ctx_data is freed with child_ctx */
free_event(child_event);
return NULL;
}
--
2.13.6
