diff mbox

[net,1/2] sfc: handle nonlinear SKBs in efx_filter_rfs()

Message ID 57471DF5.8010901@solarflare.com
State Superseded, archived
Delegated to: David Miller
Headers show

Commit Message

Edward Cree May 26, 2016, 4:01 p.m. UTC 
Previously efx_filter_rfs() assumed that the headers it needed (802.1Q, IP)
 would be present in the linear data area of the SKB.
When running with debugging I found that this is not always the case and
 that in fact the data may all be paged.
So now use skb_header_pointer() to extract the data.

Also replace EFX_BUG_ON_PARANOID checks for insufficient data with checks
 that return -EINVAL, as this case is possible if the received packet was
 too short.

Signed-off-by: Edward Cree <ecree@solarflare.com>
---
 drivers/net/ethernet/sfc/rx.c | 39 +++++++++++++++++++++++----------------
 1 file changed, 23 insertions(+), 16 deletions(-)

Comments

Eric Dumazet May 26, 2016, 4:56 p.m. UTC | #1 
On Thu, 2016-05-26 at 17:01 +0100, Edward Cree wrote:
> Previously efx_filter_rfs() assumed that the headers it needed (802.1Q, IP)
>  would be present in the linear data area of the SKB.
> When running with debugging I found that this is not always the case and
>  that in fact the data may all be paged.
> So now use skb_header_pointer() to extract the data.
> 
> Also replace EFX_BUG_ON_PARANOID checks for insufficient data with checks
>  that return -EINVAL, as this case is possible if the received packet was
>  too short.
> 
> Signed-off-by: Edward Cree <ecree@solarflare.com>
> ---
>  drivers/net/ethernet/sfc/rx.c | 39 +++++++++++++++++++++++----------------
>  1 file changed, 23 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
> index 8956995..52790f0 100644
> --- a/drivers/net/ethernet/sfc/rx.c
> +++ b/drivers/net/ethernet/sfc/rx.c
> @@ -842,25 +842,32 @@ int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
>  	struct efx_nic *efx = netdev_priv(net_dev);
>  	struct efx_channel *channel;
>  	struct efx_filter_spec spec;
> +	/* 60 octets is the maximum length of an IPv4 header (all IPv6 headers
> +	 * are 40 octets), and we pull 4 more to get the port numbers
> +	 */
> +	#define EFX_RFS_HEADER_LENGTH	(sizeof(struct vlan_hdr) + 60 + 4)

Lot of magic here. Yet another flow dissection.

Seems to be a good place to use net/core/flow_dissector.c helpers.

I truly believe that every time a driver has a private flow dissector,
it always have at least one bug.
Edward Cree May 26, 2016, 5:11 p.m. UTC | #2 
On 26/05/16 17:56, Eric Dumazet wrote:
> Lot of magic here. Yet another flow dissection.
>
> Seems to be a good place to use net/core/flow_dissector.c helpers.
Fair point, but that doesn't really feel like 'net' material.
I'll look into flow_dissector and see if I can get something ready for when net-next opens back up.  But in the meantime I think this fix is still needed.
David Miller May 26, 2016, 7:47 p.m. UTC | #3 
From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Thu, 26 May 2016 09:56:36 -0700

> I truly believe that every time a driver has a private flow dissector,
> it always have at least one bug.

+1
diff mbox

Patch

diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
index 8956995..52790f0 100644
--- a/drivers/net/ethernet/sfc/rx.c
+++ b/drivers/net/ethernet/sfc/rx.c
@@ -842,25 +842,32 @@  int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 	struct efx_nic *efx = netdev_priv(net_dev);
 	struct efx_channel *channel;
 	struct efx_filter_spec spec;
+	/* 60 octets is the maximum length of an IPv4 header (all IPv6 headers
+	 * are 40 octets), and we pull 4 more to get the port numbers
+	 */
+	#define EFX_RFS_HEADER_LENGTH	(sizeof(struct vlan_hdr) + 60 + 4)
+	unsigned char header[EFX_RFS_HEADER_LENGTH];
+	int headlen = min_t(int, EFX_RFS_HEADER_LENGTH, skb->len);
+	#undef EFX_RFS_HEADER_LENGTH
+	void *hptr;
 	const __be16 *ports;
 	__be16 ether_type;
 	int nhoff;
 	int rc;
 
-	/* The core RPS/RFS code has already parsed and validated
-	 * VLAN, IP and transport headers.  We assume they are in the
-	 * header area.
-	 */
+	hptr = skb_header_pointer(skb, 0, headlen, header);
+	if (!hptr)
+		return -EINVAL;
 
 	if (skb->protocol == htons(ETH_P_8021Q)) {
-		const struct vlan_hdr *vh =
-			(const struct vlan_hdr *)skb->data;
+		const struct vlan_hdr *vh = hptr;
 
 		/* We can't filter on the IP 5-tuple and the vlan
 		 * together, so just strip the vlan header and filter
 		 * on the IP part.
 		 */
-		EFX_BUG_ON_PARANOID(skb_headlen(skb) < sizeof(*vh));
+		if (headlen < sizeof(*vh))
+			return -EINVAL;
 		ether_type = vh->h_vlan_encapsulated_proto;
 		nhoff = sizeof(struct vlan_hdr);
 	} else {
@@ -881,23 +888,23 @@  int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 	spec.ether_type = ether_type;
 
 	if (ether_type == htons(ETH_P_IP)) {
-		const struct iphdr *ip =
-			(const struct iphdr *)(skb->data + nhoff);
+		const struct iphdr *ip = hptr + nhoff;
 
-		EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
+		if (headlen < nhoff + sizeof(*ip))
+			return -EINVAL;
 		if (ip_is_fragment(ip))
 			return -EPROTONOSUPPORT;
 		spec.ip_proto = ip->protocol;
 		spec.rem_host[0] = ip->saddr;
 		spec.loc_host[0] = ip->daddr;
-		EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
-		ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
+		if (headlen < nhoff + 4 * ip->ihl + 4)
+			return -EINVAL;
+		ports = (const __be16 *)(hptr + nhoff + 4 * ip->ihl);
 	} else {
-		const struct ipv6hdr *ip6 =
-			(const struct ipv6hdr *)(skb->data + nhoff);
+		const struct ipv6hdr *ip6 = (hptr + nhoff);
 
-		EFX_BUG_ON_PARANOID(skb_headlen(skb) <
-				    nhoff + sizeof(*ip6) + 4);
+		if (headlen < nhoff + sizeof(*ip6) + 4)
+			return -EINVAL;
 		spec.ip_proto = ip6->nexthdr;
 		memcpy(spec.rem_host, &ip6->saddr, sizeof(ip6->saddr));
 		memcpy(spec.loc_host, &ip6->daddr, sizeof(ip6->daddr));