Re: [PATCH] bcachefs: six locks: Fix missing barrier on wait->lock_acquired

[Kent Overstreet]

On Sat, Aug 12, 2023 at 01:58:58PM -0700, Boqun Feng wrote:
> Given the whole percpu counters for readers thing is similar to
> percpu_rw_semaphore, I took a look at percpu_rwsem and wonder there is
> a path to combine that with SIX lock. And that makes me realize another
> fix may be needed here, considering the following case:
> 
> 	Task A					Task B
> 	======					======
> 	__six_lock_wakeup():
> 	  task = w->task;
> 	  ...
> 	  smp_store_release(&w->locked_acquire, true);
> 	  					six_lock_slowpath():
> 						  while (1) {
> 						    if (smp_load_acquire(->lock_acquired))
> 						      break;
> 						  }
> 
> 						six_unlock();
> 						do_exit(); // Task B ends its life :(
> 
> 	  wake_up_process(task); // @task is a dangling task pointer!!!
> 
> Looks like get_task_struct() and put_task_struct() are needed here:
> similar to percpu_rwsem_wake_function().

yeah, looks like you're right. Shame about adding more atomics here,
maybe we can convert task->ref to a percpu refcount at some point.