On Tue, 11 Aug 2026 14:00:05 +0800
Michael Wu <[email protected]> wrote:
> > What does the above mean? Are you loading two modules at the same time?
> Two modules (A and B) are loaded simultaneously on different CPUs. On the
> arm64,
> when CPU0's trace_module_notify [pri=1] and CPU1's trace_module_notify
> [pri=0]
> simultaneously perform operations on call_A, because they are in different
> cache lines,
> CPU1 may observe WRITE_ONCE(head->next, &f->link) in step (4) before
> f->link.next=next in step (2).
> At this time, CPU1 reads an uninitialized f->link.next and performs an
> operation that causes to crash.
This is still way too verbose. Is this AI written? If so, AI is *not* your
friend.
>
> > What does "pri=X notifier" mean? What function calls are these coming from?
> >
> `pri=X notifier` represents `trace_events.c:trace_module_notify [pri=1]` and
> `trace.c:trace_module_notify [pri=0]`, respectively.
Why are the priorities of the notifiers important here?
I honestly didn't know one was allowed to load two modules at the same time
and thought that it the module logic would prevent that. But if that's not
the case, then yeah, we need protection.
> CPU0 (loads module A) CPU1 (loads module B)
> =============================== ===============================
> load_module(A) load_module(B)
> blocking_notifier_call_chain_robust
> blocking_notifier_call_chain_robust
> notifier_call_chain notifier_call_chain
> nb = trace_events.c: nb = trace.c:
> trace_module_notify [pri=1] trace_module_notify [pri=0]
> mutex_lock(&event_mutex) trace_event_update_all()
> trace_module_add_events(A)
> down_write(&trace_event_sem)
> __register_event(call_A)
> __add_event_to_tracers(call_A)
> event_define_fields(call_A)
> for each f:
> f = kmem_cache_alloc()
> list_add(&f->link,
> &class->fields)
> f->link.next=next; (2)
> WRITE_ONCE(head->next,
> &f->link); (4)
> update_event_fields(call_A)
>
> mutex_unlock(&event_mutex)
> list_for_each_entry(field,
> &class->fields,
> link)
> field =
> class->fields->next
> = &f->link
> = f
> (offset 0)
>
> up_write(&trace_event_sem)
Basically this can be summed up to being:
CPU0 (loads module A) CPU1 (loads module B)
=============================== ===============================
load_module(A) load_module(B)
notifier_call_chain notifier_call_chain
trace_module_notify trace_module_notify
mutex_lock(&event_mutex) trace_event_update_all()
trace_module_add_events(A)
down_write(&trace_event_sem)
__register_event(call_A)
__add_event_to_tracers(call_A)
event_define_fields(call_A)
for each f:
list_for_each_entry(field,
list_add(&f->link,
&class->fields, link)
&class->fields) field =
class->fields->next;
Where you can see that one is being read while the other is being written
to. You do not need to go into details of the cache visibility here because
this is an obvious race condition. All that information just distracts from
the real issue that is being fixed.
Less is more when it comes to describing a bug.
I'll rewrite you change log and take the patch.
Thanks,
-- Steve
