Re: [PATCH net-next] net: take care of cloned skbs in tcp_try_coalesce()

[Alexander Duyck <alexander.h.duyck@intel.com>]

On 05/02/2012 01:13 AM, Eric Dumazet wrote:
> From: Eric Dumazet <edumazet@google.com>
>
> Before stealing fragments or skb head, we must make sure skb is not
> cloned.
>
> If skb is cloned, we must take references on pages instead.
>
> Bug happened using tcpdump (if not using mmap())
>
> Reported-by: Alexander Duyck <alexander.h.duyck@intel.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> ---
>  net/ipv4/tcp_input.c |   17 ++++++++++++++---
>  1 file changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> index 96a631d..7686d7f 100644
> --- a/net/ipv4/tcp_input.c
> +++ b/net/ipv4/tcp_input.c
> @@ -4467,7 +4467,7 @@ static bool tcp_try_coalesce(struct sock *sk,
>  			     struct sk_buff *from,
>  			     bool *fragstolen)
>  {
> -	int delta, len = from->len;
> +	int i, delta, len = from->len;
>  
>  	*fragstolen = false;
>  	if (tcp_hdr(from)->fin)
> @@ -4497,7 +4497,13 @@ copyfrags:
>  		       skb_shinfo(from)->frags,
>  		       skb_shinfo(from)->nr_frags * sizeof(skb_frag_t));
>  		skb_shinfo(to)->nr_frags += skb_shinfo(from)->nr_frags;
> -		skb_shinfo(from)->nr_frags = 0;
> +
> +		if (skb_cloned(from))
> +			for (i = 0; i < skb_shinfo(from)->nr_frags; i++)
> +				skb_frag_ref(from, i);
> +		else
> +			skb_shinfo(from)->nr_frags = 0;
> +
>  		to->truesize += delta;
>  		atomic_add(delta, &sk->sk_rmem_alloc);
>  		sk_mem_charge(sk, delta);
I am fairly certain the bug I saw is only masked over by this change. 
The underlying problem is that we shouldn't be messing with nr_frags on
the from or the to if either one is clone.  You now have a check in
place for the from, but what about the to?  This function should
probably be calling a pskb_expand_head on the to skb in order to
guarantee that the skb->head isn't shared.  Otherwise this is going to
cause other issues for any functions that are sharing these skbs that
just walk through frags without checking skb->len or skb->data_len first. 

> @@ -4515,7 +4521,12 @@ copyfrags:
>  		offset = from->data - (unsigned char *)page_address(page);
>  		skb_fill_page_desc(to, skb_shinfo(to)->nr_frags,
>  				   page, offset, skb_headlen(from));
> -		*fragstolen = true;
> +
> +		if (skb_cloned(from))
> +			get_page(page);
> +		else
> +			*fragstolen = true;
> +
>  		delta = len; /* we dont know real truesize... */
>  		goto copyfrags;
>  	}
>
>
I don't see where we are now addressing the put_page call to release the
page afterwards.  By calling get_page you are incrementing the page
count by one, but where are you decrementing dataref in the shared
info?  Without that we are looking at a memory leak because __kfree_skb
will decrement the dataref but it will never reach 0 so it will never
call put_page on the head frag.

Thanks,

Alex