Re: Regression, bisected: reference leak with IPSec since ~2.6.31

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 21 septembre 2010 à 11:55 +0200, Eric Dumazet a écrit :
> Le mardi 21 septembre 2010 à 09:38 +0000, Jarek Poplawski a écrit :
> > On Tue, Sep 21, 2010 at 11:21:00AM +0200, Eric Dumazet wrote:
> 
> > I hope you're right with this.
> > 
> > > 
> > > I liked the :
> > > 
> > > <check everything without changing state>
> > > if something wrong
> > > 	goto slow_path
> > > else
> > > 	<OK, lets do destructive things>
> > > 
> > 
> > But it's an optimization of the "unlikely" case btw. ;-)
> > 
> 
> This is a bug fix, in a complex function.
> 
> in -next branch, we can try to be smart, adding more code in slow_path
> to revert what was done in the hope fast path would be taken.
> And pray we dont add another bug :)
> 

My initial patch was somehing like :

<loop to check everything is ok>
  compute truesizes = sum (frag->truesize)
  not modifying frag->sk/frag->destructor

if problem -> goto slow_path
else {
	skb->truesize -= truesizes;
       send(skb)

	for each frag
		set frag->sk = original_sk
                set frag->destructor = sock_wfree
		send(frag)

Problem was if we got an error while sending skb or one frag :
It was leaking some truesize accounting.
(an extra loop was needed to set frag->sk/frag->destructor for remaining
frags before freeing them)

So I added the extra loop to :

for each frag {
	set frag->sk = skb->sk
	set frag->destructor = sock_wfree
	skb->truesize -= frag->truesize
}

It removed the need to compute sum(frag->truesize) in a temp variable
(truesizes)