[Tarantool-patches] [PATCH 1/1] fiber_cond: remove rlist_shift usages

[Serge Petrenko]

02.03.2021 10:37, Cyrill Gorcunov пишет:
> On Mon, Mar 01, 2021 at 11:23:13PM +0100, Vladislav Shpilevoy wrote:
>> @@ -94,7 +94,7 @@ fiber_cond_broadcast(struct fiber_cond *e)
>>   {
>>   	while (! rlist_empty(&e->waiters)) {
>>   		struct fiber *f;
>> -		f = rlist_shift_entry(&e->waiters, struct fiber, state);
>> +		f = rlist_first_entry(&e->waiters, struct fiber, state);
>>   		fiber_wakeup(f);
>>   	}
>>   }
> The fiber_wakeup ignores
>
> 	if (f->flags & (FIBER_IS_READY | FIBER_IS_DEAD))
> 		return;
>
> can't we hit the situation where fiber_cond_broadcast called with
> dead fiber so that it won't be deleted from the list with new code?

Good point, I never thought of this.

Once a fiber dies, it's removed from any list that could wake it up:

         /* reset pending wakeups */
         rlist_del(&fiber->state);

And FIBER_IS_READY is set only from fiber_wakeup() to guard from
concurrent wake-ups. So the patch must be fine indeed.

Vlad, LGTM.

>
> Actually looking into fiber_loop code I see
>
> static void
> fiber_loop(MAYBE_UNUSED void *data)
> {
> 	...
> 	fiber->flags |= FIBER_IS_DEAD;
> 	while (! rlist_empty(&fiber->wake)) {
> 		       struct fiber *f;
> 		       f = rlist_shift_entry(&fiber->wake, struct fiber,
> 					     state);
> 		       assert(f != fiber);
> 		       fiber_wakeup(f);
> 	}
>
> so it should be safe with your patch, but just to make sure I didn't
> miss something obvious.

-- 
Serge Petrenko