Root Cause Analysis
The KCSAN report is triggered by an unmarked read of struct perf_event::state in perf_event_release_kernel() racing with a marked write (WRITE_ONCE()) to event->state in perf_event_set_state() during context-switch scheduling (event_sched_in()).
Locking and Execution Flow
-
Writer path (CPU 1, event_sched_in):
During a context switch (finish_task_switch() -> perf_event_task_sched_in() -> __perf_event_task_sched_in() -> perf_event_context_sched_in()), the scheduler acquires the raw spinlock ctx->lock via perf_ctx_lock(cpuctx, ctx) (it cannot acquire ctx->mutex in atomic/IRQ-disabled context).
In event_sched_in() (kernel/events/core.c:2849), it calls perf_event_set_state(event, PERF_EVENT_STATE_ACTIVE) (kernel/events/core.c:790), which executes:
WRITE_ONCE(event->state, state);
transitioning event->state from PERF_EVENT_STATE_INACTIVE (0) to PERF_EVENT_STATE_ACTIVE (1).
-
Reader path (CPU 0, perf_event_release_kernel):
When closing the perf event file descriptor (close_range() -> perf_release() -> perf_event_release_kernel()), perf_event_release_kernel() acquires ctx->mutex via perf_event_ctx_lock(event) (kernel/events/core.c:5976), without holding ctx->lock. At kernel/events/core.c:5990, it performs an unmarked read of event->state:
if (event->state > PERF_EVENT_STATE_REVOKED) {
perf_remove_from_context(event, DETACH_GROUP|DETACH_DEAD);
} else {
event->state = PERF_EVENT_STATE_DEAD;
}
Why the Race Occurs and State Invariants
- Synchronization mismatch:
event->state is mutated under ctx->lock during scheduling operations (event_sched_in(), event_sched_out(), merge_sched_in()), while perf_event_release_kernel() checks event->state > PERF_EVENT_STATE_REVOKED holding only ctx->mutex.
- Invariant preserved: Transitions into
PERF_EVENT_STATE_REVOKED (-4) or PERF_EVENT_STATE_DEAD (-5) only occur while holding ctx->mutex (pmu_detach_event() -> __pmu_detach_event() -> perf_event_exit_event(..., DETACH_REVOKE) and perf_event_release_kernel()). Once REVOKED, the event is detached from ctx and can no longer be visited by event_sched_in(). Conversely, when event->state > PERF_EVENT_STATE_REVOKED, concurrent ctx->lock-only transitions can only move event->state among PERF_EVENT_STATE_INACTIVE (0), PERF_EVENT_STATE_ACTIVE (1), and PERF_EVENT_STATE_ERROR (-2), all of which are strictly greater than PERF_EVENT_STATE_REVOKED (-4).
- As a result,
perf_event_release_kernel() always takes the perf_remove_from_context(event, DETACH_GROUP|DETACH_DEAD) branch, which then acquires ctx->lock (directly or via event_function_call()) and properly serializes with event_sched_in().
Fix Direction
Use READ_ONCE(event->state) at kernel/events/core.c:5990 (consistent with other lockless/ctx->mutex-only readers of event->state in kernel/events/core.c such as perf_event_read() and perf_event_stop()), preventing KCSAN warnings and potential compiler load-tearing/reloading.