Re: [take21 1/4] kevent: Core files.

[Eric Dumazet <dada1@cosmosbay.com>]

Evgeniy Polyakov a e'crit :
> On Sat, Oct 28, 2006 at 12:28:12PM +0200, Eric Dumazet (dada1@cosmosbay.com) wrote:
>>
>> I really dont understand how you manage to queue multiple kevents in the 
>> 'overflow list'. You just queue one kevent at most. What am I missing ?
> 
> There is no overflow list - it is a pointer to the first kevent in the
> ready queue, which was not put into ring buffer. It is an optimisation, 
> which allows to not search for that position each time new event should 
> be placed into the buffer, when it starts to have an empty slot.

This overflow list (you may call it differently, but still it IS a list), is 
not complete. I feel you add it just to make me happy, but I am not (yet :) )

For example, you make no test at kevent_finish_user_complete() time.

Obviously, you can have a dangling pointer, and crash your box in certain 
conditions.

static void kevent_finish_user_complete(struct kevent *k, int deq)
{
	struct kevent_user *u = k->user;
	unsigned long flags;

	if (deq)
		kevent_dequeue(k);

	spin_lock_irqsave(&u->ready_lock, flags);
	if (k->flags & KEVENT_READY) {
+               if (u->overflow_event == k) {
+		/* MUST do something to change u->overflow_kevent */
+		}
		list_del(&k->ready_entry);
		k->flags &= ~KEVENT_READY;
		u->ready_num--;
	}
	spin_unlock_irqrestore(&u->ready_lock, flags);

	kevent_user_put(u);
	call_rcu(&k->rcu_head, kevent_free_rcu);
}

Eric