net: core: allow fast GRO for skbs with Ethernet header in head

Commit 78d3fd0b7d ("gro: Only use skb_gro_header for completely
non-linear packets") back in May'09 (v2.6.31-rc1) has changed the
original condition '!skb_headlen(skb)' to
'skb->mac_header == skb->tail' in gro_reset_offset() saying: "Since
the drivers that need this optimisation all provide completely
non-linear packets" (note that this condition has become the current
'skb_mac_header(skb) == skb_tail_pointer(skb)' later with commmit
ced14f6804 ("net: Correct comparisons and calculations using
skb->tail and skb-transport_header") without any functional changes).

For now, we have the following rough statistics for v5.4-rc7:
1) napi_gro_frags: 14
2) napi_gro_receive with skb->head containing (most of) payload: 83
3) napi_gro_receive with skb->head containing all the headers: 20
4) napi_gro_receive with skb->head containing only Ethernet header: 2

With the current condition, fast GRO with the usage of
NAPI_GRO_CB(skb)->frag0 is available only in the [1] case.
Packets pushed by [2] and [3] go through the 'slow' path, but
it's not a problem for them as they already contain all the needed
headers in skb->head, so pskb_may_pull() only moves skb->data.

The layout of skbs in the fourth [4] case at the moment of
dev_gro_receive() is identical to skbs that have come through [1],
as napi_frags_skb() pulls Ethernet header to skb->head. The only
difference is that the mentioned condition is always false for them,
because skb_put() and friends irreversibly alter the tail pointer.
They also go through the 'slow' path, but now every single
pskb_may_pull() in every single .gro_receive() will call the *really*
slow __pskb_pull_tail() to pull headers to head. This significantly
decreases the overall performance for no visible reasons.

The only two users of method [4] is:
* drivers/staging/qlge
* drivers/net/wireless/iwlwifi (all three variants: dvm, mvm, mvm-mq)

Note that in case with wireless drivers we can't use [1]
(napi_gro_frags()) at least for now and mac80211 stack always
performs pushes and pulls anyways, so performance hit is inavoidable.

At the moment of v2.6.31 the mentioned change was necessary (that's
why I don't add the "Fixes:" tag), but it became obsolete since
skb_gro_mac_header() has gone in commit a50e233c50 ("net-gro:
restore frag0 optimization"), so we can simply revert the condition
in gro_reset_offset() to allow skbs from [4] go through the 'fast'
path just like in case [1].

This was tested on a 600 MHz MIPS CPU and a custom driver and this
patch gave boosts up to 40 Mbps to method [4] in both directions
comparing to net-next, which made overall performance relatively
close to [1] (without it, [4] is the slowest).

v2:
- Add more references and explanations to commit message
- Fix some typos ibid
- No functional changes

Signed-off-by: Alexander Lobakin <alobakin@dlink.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/dev.c

@@ -5611,8 +5611,7 @@ static void skb_gro_reset_offset(struct sk_buff *skb)
 	NAPI_GRO_CB(skb)->frag0 = NULL;
 	NAPI_GRO_CB(skb)->frag0_len = 0;
 
-	if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
-	    pinfo->nr_frags &&
+	if (!skb_headlen(skb) && pinfo->nr_frags &&
 	    !PageHighMem(skb_frag_page(frag0))) {
 		NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
 		NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,