Root Cause Analysis
The KCSAN data race occurs on the flags field (unsigned long flags) of struct trace_event_file (include/linux/trace_events.h:680) between a non-atomic write in filter_disable() and a lockless read in __trace_trigger_soft_disabled().
As documented in struct trace_event_file (include/linux/trace_events.h:672-678):
"The bits must be set atomically to prevent races from other writers. Reads of flags do not need to be in sync as they occur in critical sections."
While other subsystems (kernel/trace/trace_events.c, kernel/trace/trace_events_trigger.c) use set_bit() / clear_bit() to update file->flags, filter_disable() (kernel/trace/trace_events_filter.c:1321-1329) and event_set_filtered_flag() (kernel/trace/trace_events_filter.c:2288-2296) perform plain non-atomic read-modify-write operations:
static void filter_disable(struct trace_event_file *file)
{
unsigned long old_flags = file->flags;
file->flags &= ~EVENT_FILE_FL_FILTERED;
if (old_flags != file->flags)
trace_buffered_event_disable();
}
static inline void event_set_filtered_flag(struct trace_event_file *file)
{
unsigned long old_flags = file->flags;
file->flags |= EVENT_FILE_FL_FILTERED;
if (old_flags != file->flags)
trace_buffered_event_enable();
}
In addition to racing with lockless tracepoint readers, these non-atomic RMW sequences can race with event_enable_trigger() (kernel/trace/trace_events_trigger.c:1670-1680), which calls set_bit() / clear_bit() on EVENT_FILE_FL_SOFT_DISABLED_BIT in tracepoint context without holding event_mutex. That can cause lost updates to EVENT_FILE_FL_SOFT_DISABLED_BIT or cause old_flags != file->flags to evaluate true spuriously and miscount trace_buffered_event_ref.
2. Unmarked lockless reads in tracepoint paths
On the reader side, tracepoint handlers execute without holding event_mutex and read file->flags with plain loads rather than READ_ONCE():
3. Triggering Sequence
- CPU 0 (writer): Task 5433 writes a filter string to a tracefs subsystem
filter file (subsystem_filter_write() -> apply_subsystem_event_filter(), acquiring event_mutex). create_system_filter() -> process_system_preds() (kernel/trace/trace_events_filter.c:2329-2331) iterates over events in the subsystem (dir); when process_preds() returns an error for an event, it invokes filter_disable(file), which writes file->flags &= ~EVENT_FILE_FL_FILTERED at kernel/trace/trace_events_filter.c:1325.
- CPU 1 (reader): Concurrently, task 5440 hits the
sched_switch tracepoint (__schedule() -> trace_event_raw_event_sched_switch() -> trace_trigger_soft_disabled() -> __trace_trigger_soft_disabled()), reading file->flags at kernel/trace/trace_events_trigger.c:217 without event_mutex or READ_ONCE().
Fix Direction
- Replace the non-atomic RMW in
filter_disable() and event_set_filtered_flag() (kernel/trace/trace_events_filter.c) with test_and_clear_bit(EVENT_FILE_FL_FILTERED_BIT, &file->flags) and test_and_set_bit(EVENT_FILE_FL_FILTERED_BIT, &file->flags).
- Use
READ_ONCE(file->flags) (or data_race() / test_bit()) in lockless readers such as trace_trigger_soft_disabled() (include/linux/trace_events.h) and __trace_trigger_soft_disabled() (kernel/trace/trace_events_trigger.c), as well as __event_trigger_test_discard() and trace_event_buffer_lock_reserve().