netdev
[Top] [All Lists]

Re: tx_timeout and timer serialisation

To: andrewm@xxxxxxxxxx (Andrew Morton)
Subject: Re: tx_timeout and timer serialisation
From: kuznet@xxxxxxxxxxxxx
Date: Thu, 25 May 2000 22:31:42 +0400 (MSK DST)
Cc: netdev@xxxxxxxxxxx
In-reply-to: <3923F8CD.AECBDA6D@xxxxxxxxxx> from "Andrew Morton" at May 19, 0 00:06:05 am
Sender: owner-netdev@xxxxxxxxxxx
Hello!

>         for (;;) {
>                 unsigned long flags;
>                 int running;
> 
>                 spin_lock_irqsave(&timerlist_lock, flags);
> 
> ** The timer handler could be running now. 

Of course! It is exactly the situation, when del_timer_sync()
is different of del_timer().


>                                        It can delete the
>    timer and kfree it, or reuse its memory for something else,
>    or turn it into a semantically different timer **

Yes, and in this case you cannot use del_timer_sync() and
have to use generic reference counting scheme.

del_timer_sync() is used by process, which _owns_ this timer
and have exclusive right to destroy it. See?

If timer handler is self-restartable, del_timer_sync() guarantees that
timer is not running after exit from del_timer_sync(),
so that you may destroy it safely.

Another (more complicated) scheme is used by TCP (net/ipv4/tcp_timer.c),
by neighbour cache (core/neighbour.c) etc. In these cases timer
"thread" has equal rights with process threads and we have to use
exact reference counting to wait for last timer user.
del_timer_sync() does nothing useful in this case, but it also does not fail,
because anyone operating on timer holds it with reference count.

Alexey

<Prev in Thread] Current Thread [Next in Thread>