netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Jesper Dangaard Brouer <brouer@redhat.com>
To: Daniel Borkmann <daniel@iogearbox.net>
Cc: bpf@vger.kernel.org, netdev@vger.kernel.org,
	Daniel Borkmann <borkmann@iogearbox.net>,
	Alexei Starovoitov <alexei.starovoitov@gmail.com>,
	maze@google.com, lmb@cloudflare.com, shaun@tigera.io,
	Lorenzo Bianconi <lorenzo@kernel.org>,
	marek@cloudflare.com, John Fastabend <john.fastabend@gmail.com>,
	Jakub Kicinski <kuba@kernel.org>,
	eyal.birger@gmail.com, colrack@gmail.com, brouer@redhat.com
Subject: Re: [PATCH bpf-next V8 5/8] bpf: drop MTU check when doing TC-BPF redirect to ingress
Date: Thu, 17 Dec 2020 15:46:55 +0100	[thread overview]
Message-ID: <20201217154655.42e89d08@carbon> (raw)
In-Reply-To: <af28e4e7-8089-b252-3927-a962b98ad7b8@iogearbox.net>

On Thu, 3 Dec 2020 00:43:36 +0100
Daniel Borkmann <daniel@iogearbox.net> wrote:

> On 11/27/20 7:06 PM, Jesper Dangaard Brouer wrote:
> > 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/
> > 
> > V4:
> >   - Keep net_device "up" (IFF_UP) check.
> >   - Adjustment to handle bpf_redirect_peer() helper
> > 
> > Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > ---
> >   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 7ce648a564f7..4a854e09e918 100644
> > --- a/include/linux/netdevice.h
> > +++ b/include/linux/netdevice.h
> > @@ -3917,11 +3917,38 @@ 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)
> > +{
> > +	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)
> >   {
> >   	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 60d325bda0d7..6ceb6412ee97 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -2189,28 +2189,13 @@ static inline void net_timestamp_set(struct sk_buff *skb)
> >   
> >   bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
> >   {
> > -	unsigned int len;
> > -
> > -	if (!(dev->flags & IFF_UP))
> > -		return false;
> > -
> > -	len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
> > -	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;
> > +	return __is_skb_forwardable(dev, skb, true);
> >   }
> >   EXPORT_SYMBOL_GPL(is_skb_forwardable);  
> 
> Only user of is_skb_forwardable() that is left after this patch is bridge, maybe
> the whole thing should be moved into the header?

Well, yes, maybe... I just felt it belongs in another patchset.


> >   int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
> >   {
> > -	int ret = ____dev_forward_skb(dev, skb);
> > +	int ret = ____dev_forward_skb(dev, skb, true);
> >   
> >   	if (likely(!ret)) {
> >   		skb->protocol = eth_type_trans(skb, dev);
> > diff --git a/net/core/filter.c b/net/core/filter.c
> > index d6125cfc49c3..4673afe59533 100644
> > --- a/net/core/filter.c
> > +++ b/net/core/filter.c
> > @@ -2083,13 +2083,21 @@ static const struct bpf_func_proto bpf_csum_level_proto = {
> >   
> >   static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
> >   {
> > -	return dev_forward_skb(dev, skb);
> > +	int ret = ____dev_forward_skb(dev, skb, false);
> > +
> > +	if (likely(!ret)) {
> > +		skb->protocol = eth_type_trans(skb, dev);
> > +		skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
> > +		ret = netif_rx(skb);  
> 
> Why netif_rx() and not netif_rx_internal() as in dev_forward_skb() originally?
> One extra call otherwise.

This is because the function below calls netif_rx(), which is just
outside patch-diff-window.  Thus, it looked wrong/strange to call
netif_rx_internal(), but sure I can use netif_rx_internal() instead.

> 
> > +	}
> > +
> > +	return ret;
> >   }
> >   
> >   static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
> >   				      struct sk_buff *skb)
> >   {
> > -	int ret = ____dev_forward_skb(dev, skb);
> > +	int ret = ____dev_forward_skb(dev, skb, false);
> >   
> >   	if (likely(!ret)) {
> >   		skb->dev = dev;
> > @@ -2480,7 +2488,7 @@ int skb_do_redirect(struct sk_buff *skb)
> >   			goto out_drop;
> >   		dev = ops->ndo_get_peer_dev(dev);
> >   		if (unlikely(!dev ||
> > -			     !is_skb_forwardable(dev, skb) ||
> > +			     !__is_skb_forwardable(dev, skb, false) ||  
> 
> If we only use __is_skb_forwardable() with false directly here, maybe then
> lets just have the !(dev->flags & IFF_UP) test here instead..

Sure, let do that.

> >   			     net_eq(net, dev_net(dev))))
> >   			goto out_drop;
> >   		skb->dev = dev;
> > 


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


  reply	other threads:[~2020-12-17 14:48 UTC|newest]

Thread overview: 16+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-11-27 18:06 [PATCH bpf-next V8 0/8] bpf: New approach for BPF MTU handling Jesper Dangaard Brouer
2020-11-27 18:06 ` [PATCH bpf-next V8 1/8] bpf: Remove MTU check in __bpf_skb_max_len Jesper Dangaard Brouer
2020-11-27 18:06 ` [PATCH bpf-next V8 2/8] bpf: fix bpf_fib_lookup helper MTU check for SKB ctx Jesper Dangaard Brouer
2020-11-27 18:06 ` [PATCH bpf-next V8 3/8] bpf: bpf_fib_lookup return MTU value as output when looked up Jesper Dangaard Brouer
2020-12-02 22:23   ` Daniel Borkmann
2020-11-27 18:06 ` [PATCH bpf-next V8 4/8] bpf: add BPF-helper for MTU checking Jesper Dangaard Brouer
2020-12-02 23:23   ` Daniel Borkmann
2020-12-14 13:51     ` Jesper Dangaard Brouer
2020-11-27 18:06 ` [PATCH bpf-next V8 5/8] bpf: drop MTU check when doing TC-BPF redirect to ingress Jesper Dangaard Brouer
2020-12-02 23:43   ` Daniel Borkmann
2020-12-17 14:46     ` Jesper Dangaard Brouer [this message]
2020-12-17 16:10       ` Jesper Dangaard Brouer
2020-11-27 18:06 ` [PATCH bpf-next V8 6/8] bpf: make it possible to identify BPF redirected SKBs Jesper Dangaard Brouer
2020-12-03  0:06   ` Daniel Borkmann
2020-11-27 18:06 ` [PATCH bpf-next V8 7/8] selftests/bpf: use bpf_check_mtu in selftest test_cls_redirect Jesper Dangaard Brouer
2020-11-27 18:06 ` [PATCH bpf-next V8 8/8] bpf/selftests: tests using bpf_check_mtu BPF-helper Jesper Dangaard Brouer

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20201217154655.42e89d08@carbon \
    --to=brouer@redhat.com \
    --cc=alexei.starovoitov@gmail.com \
    --cc=borkmann@iogearbox.net \
    --cc=bpf@vger.kernel.org \
    --cc=colrack@gmail.com \
    --cc=daniel@iogearbox.net \
    --cc=eyal.birger@gmail.com \
    --cc=john.fastabend@gmail.com \
    --cc=kuba@kernel.org \
    --cc=lmb@cloudflare.com \
    --cc=lorenzo@kernel.org \
    --cc=marek@cloudflare.com \
    --cc=maze@google.com \
    --cc=netdev@vger.kernel.org \
    --cc=shaun@tigera.io \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).