Re: [PATCH] trace_branch: use per-cpu counters for correct/incorrect stats
From: Steven Rostedt <rostedt@goodmis.org>
Date: 2026-06-29 10:46:54
Also in:
lkml
On Mon, 29 Jun 2026 16:58:38 +0700 Martin Weiss [off-list ref] wrote:
quoted hunk ↗ jump to hunk
Replace per-task counters with per-cpu increments to avoid race conditions in the branch profiler fast path. Fixes FIXME about atomicity. Signed-off-by: Martin Weiss <redacted> --- kernel/trace/trace_branch.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)diff --git a/kernel/trace/trace_branch.c b/kernel/trace/trace_branch.c index d8e97ad798f0..960bcb3d7dbf 100644 --- a/kernel/trace/trace_branch.c +++ b/kernel/trace/trace_branch.c@@ -213,11 +213,11 @@ void ftrace_likely_update(struct ftrace_likely_data *f, int val, */ trace_likely_condition(f, val, expect); - /* FIXME: Make this atomic! */ + /* use per-cpu counters to avoid contention */ if (val == expect) - f->data.correct++; + this_cpu_inc(f->data.correct); else - f->data.incorrect++; + this_cpu_inc(f->data.incorrect);
I don't think this does what you think it does. Did you even test this? I'm guessing it would blow up in some fantastic ways. NAK. -- Steve
user_access_restore(flags); }