Re: Regression associated with commit c8628155ece3 - "tcp: reduce out_of_order memory use"

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

On Fri, 2012-08-24 at 11:58 -0500, Larry Finger wrote:
> On 08/24/2012 11:23 AM, Eric Dumazet wrote:
> > On Fri, 2012-08-24 at 18:18 +0200, Eric Dumazet wrote:
> >> On Fri, 2012-08-24 at 10:58 -0500, Larry Finger wrote:
> >>> On 08/24/2012 10:19 AM, David Miller wrote:
> >>>>
> >>>> This looks like full-on data corruption to me.
> >>>
> >>> I agree. The question is why does it happen with r8712u, and only after the
> >>> commit in the subject. Drivers for other devices that I have are OK. Thus far, I
> >>> have tested b43, rtl8187, ath9k_htc, and rtl8192cu. To my knowledge, there are
> >>> no reports posted for this bug with any other device.
> >>
> >> bugs can sit unnoticed, and one change somewhere can uncover them.
> >>
> >> Really this driver must have a bug, if not half a dozen of bugs.
> >>
> >> For example this sequence of code is a clear bug :
> >>
> >> sub_skb = dev_alloc_skb(nSubframe_Length + 12);
> >> skb_reserve(sub_skb, 12);
> >>
> >>
> >> Also the free_recv_skb_queue looks really suspect to me
> >>
> >> What the hell is doing recv_tasklet() I really wonder.
> >>
> >> This code, combined with the skb_clone() in recvbuf2recvframe()
> >> can clearly reuse an skb passed to upper stacks.
> >>
> >>
> >> queueing one skb in free_recv_skb_queue should be done
> >> only if no clone of this skb exist somewhere.
> >>
> >> Please someone fix this buggy driver.
> >>
> >
> > Try the following patch for a start
> >
> > diff --git a/drivers/staging/rtl8712/rtl8712_recv.c b/drivers/staging/rtl8712/rtl8712_recv.c
> > index 8e82ce2..88e3ca6 100644
> > --- a/drivers/staging/rtl8712/rtl8712_recv.c
> > +++ b/drivers/staging/rtl8712/rtl8712_recv.c
> > @@ -1127,6 +1127,9 @@ static void recv_tasklet(void *priv)
> >   		recvbuf2recvframe(padapter, pskb);
> >   		skb_reset_tail_pointer(pskb);
> >   		pskb->len = 0;
> > -		skb_queue_tail(&precvpriv->free_recv_skb_queue, pskb);
> > +		if (!skb_cloned(pskb))
> > +			skb_queue_tail(&precvpriv->free_recv_skb_queue, pskb);
> > +		else
> > +			consume_skb(pskb);
> >   	}
> >   }
> 
> This one did not help. There is no doubt it is needed for the case where memory 
> is tight, an allocation fails, and the driver clones the skb. In the present 
> case, debug statements have shown that the skb_clone() call was not made.
> 
> In the long term, this driver will be replaced with one that uses mac80211, but 
> in the short term, I am trying to fix it.
> 
> As I said earlier, my skb skills are minimal. Could you explain what is wrong 
> with the following sequence?
> 
>   sub_skb = dev_alloc_skb(nSubframe_Length + 12);
>   skb_reserve(sub_skb, 12);

dev_alloc_skb() can return NULL

-> crash

skb_clone() can also return NULL 

-> crash