On Fri, 30 Oct 2020 13:36:05 -0700
John Fastabend [off-list ref] wrote:

Jesper Dangaard Brouer wrote:

quoted

The use-case for dropping the MTU check when TC-BPF does redirect to
ingress, is described by Eyal Birger in email[0]. The summary is the
ability to increase packet size (e.g. with IPv6 headers for NAT64) and
ingress redirect packet and let normal netstack fragment packet as needed.

[0] https://lore.kernel.org/netdev/CAHsH6Gug-hsLGHQ6N0wtixdOa85LDZ3HNRHVd0opR=19Qo4W4Q@mail.gmail.com/ (local)

V4:
 - Keep net_device "up" (IFF_UP) check.
 - Adjustment to handle bpf_redirect_peer() helper

Signed-off-by: Jesper Dangaard Brouer <redacted>
---
 include/linux/netdevice.h |   31 +++++++++++++++++++++++++++++--
 net/core/dev.c            |   19 ++-----------------
 net/core/filter.c         |   14 +++++++++++---
 3 files changed, 42 insertions(+), 22 deletions(-)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 964b494b0e8d..bd02ddab8dfe 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h

@@ -3891,11 +3891,38 @@ int dev_forward_skb(struct net_device *dev, struct sk_buff *skb);
 bool is_skb_forwardable(const struct net_device *dev,
 			const struct sk_buff *skb);
 
+static __always_inline bool __is_skb_forwardable(const struct net_device *dev,
+						 const struct sk_buff *skb,
+						 const bool check_mtu)  

It looks like if check_mtu=false then this is just an interface up check.
Can we leave is_skb_forwardable logic alone and just change the spots where
this is called with false to something with a name that describes the check,
such as is_dev_up(dev). I think it will make this change smaller and the
code easier to read. Did I miss something?

People should realized that this is constructed such, the compiler will
compile-time remove the actual argument (the const bool check_mtu).
And this propagates also to ____dev_forward_skb() where the call places
are also inlined.

Yes, this (check_mtu=false) is basically an interface up check, but the
only place it is used directly is in the ndo_get_peer_dev() case, and
reading the code I find it more readable that is says
__is_skb_forwardable because this is used as part of a forwarding step,
and is_dev_up() doesn't convey the intent in this use-case.

quoted

+{
+	const u32 vlan_hdr_len = 4; /* VLAN_HLEN */
+	unsigned int len;
+
+	if (!(dev->flags & IFF_UP))
+		return false;
+
+	if (!check_mtu)
+		return true;
+
+	len = dev->mtu + dev->hard_header_len + vlan_hdr_len;
+	if (skb->len <= len)
+		return true;
+
+	/* if TSO is enabled, we don't care about the length as the packet
+	 * could be forwarded without being segmented before
+	 */
+	if (skb_is_gso(skb))
+		return true;
+
+	return false;
+}
+
 static __always_inline int ____dev_forward_skb(struct net_device *dev,
-					       struct sk_buff *skb)
+					       struct sk_buff *skb,
+					       const bool check_mtu)
 {  

I guess you will get some duplication here if you have a dev_forward_skb()
and a dev_forward_skb_nocheck() or something. Take it or leave it. I know
I've added my share of bool swivel bits like this, but better to avoid
it if possible IMO.

As I wrote the bool will actually get compile-time removed, so I don't
see that as problematic.  And I avoided replicating the code in more
places.

The problematic part (which you didn't comment) on is this:

On Fri, 30 Oct 2020 17:51:07 +0100
Jesper Dangaard Brouer [off-list ref] wrote:

I'm replicating two lines from dev_forward_skb(), but I couldn't find a
way to avoid this, without causing larger code changes (and slower code).

Other than style aspects it looks correct to me.

quoted

 	if (skb_orphan_frags(skb, GFP_ATOMIC) ||
-	    unlikely(!is_skb_forwardable(dev, skb))) {
+	    unlikely(!__is_skb_forwardable(dev, skb, check_mtu))) {
 		atomic_long_inc(&dev->rx_dropped);
 		kfree_skb(skb);
 		return NET_RX_DROP;

diff --git a/net/core/dev.c b/net/core/dev.c
index 9499a414d67e..445ccf92c149 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -2188,28 +2188,13 @@ static inline void net_timestamp_set(struct sk_buff *skb)
   

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer