Root Cause Analysis
The KCSAN data race is on the unsigned long flags field of struct trace_event_file (include/linux/trace_events.h:680).
-
Write side (kernel/trace/trace_events_trigger.c):
When an enable_event / disable_event trigger attached to an event (here sched:sched_switch on CPU 0) fires, __trace_trigger_soft_disabled() -> event_triggers_call() -> data_ops_trigger() invokes event_enable_trigger() (kernel/trace/trace_events_trigger.c:1670-1680). event_enable_trigger() updates the target struct trace_event_file (enable_data->file) via atomic bit operations:
if (enable_data->enable)
clear_bit(EVENT_FILE_FL_SOFT_DISABLED_BIT, &enable_data->file->flags);
else
set_bit(EVENT_FILE_FL_SOFT_DISABLED_BIT, &enable_data->file->flags);
-
Read side (kernel/trace/trace.h / kernel/trace/trace.c):
Concurrently on CPU 1, the target trace event (sched:sched_process_fork) executes trace_event_buffer_commit() (kernel/trace/trace.c:2020-2042), which calls the inline helper __event_trigger_test_discard() (kernel/trace/trace.h:1726-1757). At line 1732, __event_trigger_test_discard() performs a plain, unmarked read of file->flags:
unsigned long eflags = file->flags;
It then re-reads file->flags plainly at lines 1737, 1742, 1745, and 1749 (checking EVENT_FILE_FL_SOFT_DISABLED, EVENT_FILE_FL_FILTERED, and EVENT_FILE_FL_PID_FILTER) rather than reusing a single READ_ONCE(file->flags) snapshot.
-
Invariant and Trigger Mechanism:
As documented in include/linux/trace_events.h:672-678, trace_event_file::flags is intentionally read locklessly in tracepoint fast paths while writers update individual bits atomically (set_bit() / clear_bit()). Because __event_trigger_test_discard() (as well as other lockless readers such as trace_trigger_soft_disabled(), __trace_trigger_soft_disabled(), and trace_event_buffer_lock_reserve()) accesses file->flags using plain C reads instead of READ_ONCE(), KCSAN reports a data race against the marked atomic write in event_enable_trigger().
Additionally, if EVENT_FILE_FL_SOFT_DISABLED changes between trace_event_buffer_lock_reserve() (kernel/trace/trace.c:1882) and trace_event_buffer_commit(), both commit (__buffer_unlock_commit() at kernel/trace/trace.h:1687) and discard (__trace_event_discard_commit() at kernel/trace/trace.h:1702) distinguish temporary per-CPU trace_buffered_event buffers from regular ring-buffer entries via this_cpu_read(trace_buffered_event) == event, keeping ring-buffer state and preemption counts balanced.
Fix Direction
Use READ_ONCE(file->flags) when sampling file->flags in __event_trigger_test_discard() (and reuse the sampled eflags variable for the subsequent flag checks in kernel/trace/trace.h:1732-1750, as well as in other lockless readers of trace_event_file::flags such as trace_trigger_soft_disabled(), __trace_trigger_soft_disabled(), and trace_event_buffer_lock_reserve()).