On Mon, 28 Apr 2025 14:00:07 -0400
Mathieu Desnoyers [off-list ref] wrote:

quoted

quoted

It is enough to guard with RCU ? See syscall_regfunc() from
tracepoint.c where we do:

                  read_lock(&tasklist_lock);
                  for_each_process_thread(p, t) {
                          set_task_syscall_work(t, SYSCALL_TRACEPOINT);
                  }
                  read_unlock(&tasklist_lock);

To prevent concurrent fork from adding threads while we
iterate, thus opening the possibility of missing a clear
due to a concurrent fork + set bit.  

A set_bit only would happen if the callback was live and accepting new
callback requests. It's a bug for a tracer to call unwind_deferred_cancel()
and then call unwind_deferred_request() (which would set the bit). We could
possibly set the tracer's unwind descriptor id to -1, and do an
WARN_ON_ONCE() in unwind_deferred_request() if the tracer's id is negative.

The loop is called under the callback_mutex, where no new tracer could
register and be assigned that bit.  

Ah, that's the piece I missed. The callback_mutex prevents reallocation
of the ID by unwind_deferred_init while iterating on the tasks.

One more comment: if we change the linked list for an array (or make the 
linked list an RCU list), can we remove the callback_mutex from
unwind_deferred_task_work by turning it into an RCU read-side ?

Then we just need to wait for a grace period before returning from
unwind_deferred_cancel, which then allows the caller to reclaim "work".

Taking the callback_mutex in unwind_deferred_task_work will end up being
the single thing that does a lot of cache line bouncing across CPUs when
hit heavily by tracers.

I'm not against this, but again, that's an optimization. I want to keep the
initial code simple. And then add the more complex optimizations when this
is stable.

-- Steve