On Wed, 2012-10-31 at 21:15 +0200, Michael S. Tsirkin wrote:

On Sun, Sep 16, 2012 at 04:07:12PM +0100, Ben Hutchings wrote:

quoted

On Sun, 2012-09-16 at 12:17 +0300, Michael S. Tsirkin wrote:

quoted

I notice that dev_hard_start_xmit might invoke
__skb_linearize e.g. if device does not support NETIF_F_SG.

This in turn onvokes __pskb_pull_tail, and
documentation of __pskb_pull_tail says:
  &sk_buff MUST have reference count of 1.

I am guessing 'reference count' means users in this context, right?
IIUC this is because it modifies skb in a way that
isn't safe if anyone else is looking at the skb.


However, I don't see what guarantees that reference
count is 1 when dev_hard_start_xmit invokes
linearize. In particular it calls dev_queue_xmit_nit
which could queue packets on a network tap.

Could someone help me understand please?

Reference count here means references to struct sk_buff itself.  The
header area and data fragments are allowed to be shared.

dev_queue_xmit_nit() clones the skb for each tap, so the reference count
on the original skb remains 1.

Ben.

Interesting. But don't skb clones share the fragment list?

Yes.

Maybe I misunderstand? If they do it looks like the following race
would be possible:

- skb is cloned and queued e.g. at socket receive queue.
  dataref becomes 2.
- On CPU 1, skb_copy_datagram_iovec is called on clone 1, is reads nr_frags and sees
  value > 1.
- On CPU 2, __skb_linearize is now called on clone 2, it modified the
  skb so nr_frags is now 0, and does put_page for all frags > 1.
- On CPU 1, skb_copy_datagram_iovec will now use the previously read
  nr_frags > 1 and access a fragment page that was already freed.

What did I miss?

__skb_linearize() calls __pskb_pull_tail(), which starts with:

	if (eat > 0 || skb_cloned(skb)) {
		if (pskb_expand_head(skb, 0, eat > 0 ? eat + 128 : 0,
				     GFP_ATOMIC))
			return NULL;
	}

pskb_expand_head() will then create a new unshared head area for the skb
being linearised, and will add a reference to each fragment page
(skb_frag_ref()).  __pskb_pull_tail() unreferences the pages later, as
you say, but this all cancels out.

(pskb_expand_head() will also cancel zero-copy (skb_orphan_frags() ->
skb_copy_ubufs()), which seems like it should be done only after the
head area has been unshared.  But skb_clone() will also do that, so I
don't know how the zero-copy flag would still be set on a cloned skb.)

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.