Re: [PATCH] net: af_packet: don't call tpacket_destruct_skb() until the skb is sent out

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

Le samedi 11 septembre 2010 à 00:47 +0800, Changli Gao a écrit :
> On Fri, Sep 10, 2010 at 10:26 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:


> > Are you sure sock_wfree(skb) is still needed ?
> 
> sock_wfree(skb) is also used to wake up the users who sleep on
> poll(2). If sock_wfree(skb) is moved into skb->destructor(), and
> called before skb is sent out, pollers will be waked up without
> POLLOUT, and since the later skb_shinfo(skb)->destructor() doesn't
> wake up the pollers, POLLOUT events will be lost, and the poller will
> be blocked forever.
> 

Then implement poll() to use the number of available slots.
(not use the default poll() that relies on generic sk / inet queues and
counters)

Really, sock_wfree() cannot be used at all, or we also must disable
early orphaning of these skbs.

Goal is to replace skb->destructor use in af_packet by
shinfo->destructor, not mix the two.


> >
> >
> >> +     kfree(arg);
> >
> > this new kmalloc()/kfree() for each sent packet wont please the guys
> > using af_packet/mmap interface...
> 
> Embed these two pointers into skb_shared_info? It may slow the others.

we have some room because of SKB_PAD alignment,