Technical Explanation of Root Cause
1. Participating Entities and Locations
2. Mechanism and Lifecycle Violation
When a perf_event with attr.sigtrap set triggers an overflow in __perf_event_overflow(), it schedules SIGTRAP signal delivery using task work:
if (!event->pending_work &&
!task_work_add(current, &event->pending_task, notify_mode)) {
event->pending_work = pending_id;
local_inc(&event->ctx->nr_no_switch_fast);
WARN_ON_ONCE(!atomic_long_inc_not_zero(&event->refcount));
event->pending_addr = 0;
if (valid_sample && (data->sample_flags & PERF_SAMPLE_ADDR))
event->pending_addr = data->addr;
...
Here, event->pending_work functions as an exclusivity/ownership guard: while non-zero, subsequent overflows are prevented from overwriting event->pending_addr or adding another event->pending_task to the task's work list.
On return to user mode (e.g., via exit_to_user_mode_loop() -> task_work_run()), the task runs queued task works. In task_work_run() (kernel/task_work.c:231-236), the callback entry head = &event->pending_task is already detached from task->task_works prior to calling work->func(work) (perf_pending_task()).
In perf_pending_task():
static void perf_pending_task(struct callback_head *head)
{
struct perf_event *event = container_of(head, struct perf_event, pending_task);
int rctx;
rctx = perf_swevent_get_recursion_context();
if (event->pending_work) {
event->pending_work = 0;
perf_sigtrap(event);
local_dec(&event->ctx->nr_no_switch_fast);
}
put_event(event);
if (rctx >= 0)
perf_swevent_put_recursion_context(rctx);
}
perf_pending_task() executes in process/task context with hardware interrupts enabled. Crucially:
event->pending_work is reset to 0 before calling perf_sigtrap(event) and reading event->pending_addr.
- An interrupt (such as an hrtimer interrupt driving
perf_swevent_hrtimer) can arrive on the current CPU after event->pending_work = 0 is executed and while perf_sigtrap() is reading event->pending_addr.
- In interrupt context,
__perf_event_overflow() evaluates !event->pending_work as true (since pending_work was prematurely reset to 0).
- Because
event->pending_task was already detached from the task's work list by task_work_run(), task_work_add(current, &event->pending_task, notify_mode) succeeds, and __perf_event_overflow() resets event->pending_addr = 0 (and event->pending_addr = data->addr).
- This write to
event->pending_addr races directly with the interrupted task reading event->pending_addr at kernel/events/core.c:7676.