Re: Bug in computing data_len in tcp_sendmsg?

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

Le mercredi 30 novembre 2011 à 20:06 -0800, Tom Herbert a écrit :
> On Wed, Nov 30, 2011 at 7:42 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > Le mercredi 30 novembre 2011 à 17:48 -0800, Tom Herbert a écrit :
> >> I believe that skb->data_len might no be computed correctly in
> >> tcp_sendmsg.  Specifically, when skb_add_data_nocache (or
> >> skb_add_data) is called skb->data_len is not updated (skb_put only
> >> updates skb->len).  This results in the datalen in the head skbuf
> >> being zero so any subsequent uses of the value lead to incorrect
> >> results.  For instance, skb_headlen returns the length of the head
> >> skbu data and not just that of the headers.  If I'm reading this
> >> correctly, it's a pretty fundamental bug.
> >>
> >> I don't have a fix for this yet.
> >>
> >> Tom
> >
> > On which tree do you see a problem ?
> >
> > For example net-next seems fine to me :
> >
> >
> 
> On net-next.  Look at skb_add_data_nocache.  This doesn't touch
> data_len, and really can't since skb_put wants skb linearized.
> 
> Actually, I think I am misinterpreting the meaning of data_len, looks
> like it is intended to return the number of bytes in the page buffer
> area (comment on the function would be have been nice ;-) ).  It would
> seem that e1000e driver might be misinterpreting it also:
> 
>         segs = skb_shinfo(skb)->gso_segs ? : 1;
>         /* multiply data chunks by size of headers */
>         bytecount = ((segs - 1) * skb_headlen(skb)) + skb->len;
> 
> So maybe the bug is  that e1000e is misusing the function (and it
> looks like some other drivers would be too)?.

Or the "bug" was to assume that skb was headless.
It was true until recently.

We recently added commit f07d960df33c5aef
(tcp: avoid frag allocation for small frames)

to avoid page allocation for small frames.

So now, skb can contain in head part of tcp data.