ArthurChiao's Blog

Differentiate three types of eBPF redirections (2022)

Published at 2022-07-25 | Last Update 2022-07-25


There are three types of eBPF redirection fashions in Linux kernel that may confuse developers often:

  1. bpf_redirect_peer()
  2. bpf_redirect_neighbor()
  3. bpf_redirect()

This post helps to clarify them by digging into the code in history order, and also discusses usages & related problems in real world.

1 The foundation: bpf_redirect(), 2015

This BPF helper was introduced via this patch in 2015,

bpf: add bpf_redirect() helper

Existing bpf_clone_redirect() helper clones skb before redirecting it to RX or
TX of destination netdev.  Introduce bpf_redirect() helper that does that without cloning.

1.1 The documentation


long bpf_redirect(ifindex, flags) can be used to redirect the given packet to the given net device identified with index ifindex. This helper is somewhat similar to bpf_clone_redirect(), except that the packet is not cloned, which provides increased performance (25% pps increase compares with clone_redirect() according to the commit message).

Return value: TC/XDP verdict.

Comparison with bpf_clone_redirect()

  bpf_clone_redirect() bpf_redirect()
Effeciency Low (involve skb clone) High
Where the redirection happens Inside function call After function call (this func only returns a verdict, the real redirection happens in skb_do_redirect())
Can be used out of eBPF program Yes No
May change underlying skb buffer Yes (need more re-validations) No
Redirect across netns No No

1.2 Kernel implementations/changes

Now let’s look at what changes were made against the Linux kernel to support this function.

1. Add TC action type TC_ACT_REDIRECT

diff --git a/include/uapi/linux/pkt_cls.h b/include/uapi/linux/pkt_cls.h
@@ -87,6 +87,7 @@ enum {
 #define TC_ACT_STOLEN          4
 #define TC_ACT_QUEUED          5
 #define TC_ACT_REPEAT          6
+#define TC_ACT_REDIRECT                7

2. Add new BPF helper & syscall

+static u64 bpf_redirect(u64 ifindex, u64 flags, u64 r3, u64 r4, u64 r5)
+       struct redirect_info *ri = this_cpu_ptr(&redirect_info);
+       ri->ifindex = ifindex;
+       ri->flags = flags;
+       return TC_ACT_REDIRECT;

We can see that this helper only sets ifindex and flags and then returns a TC_ACT_REDIRECT to the caller, that’s why we say that the real redirection happens after bpf_redirect() finishes.

3. Process redirect logic in TC BPF

When BPF program (with bpf_redirect() in the program) was attached to the TC ingress hook, bpf_redirect() will be executed in the tc_classify() method:

@@ -3670,6 +3670,14 @@ static inline struct sk_buff *handle_ing(struct sk_buff *skb,
        switch(tc_classify()) { // <-- bpf_redirect() executes in the tc_classify() method
        case TC_ACT_QUEUED:
                return NULL;
+       case TC_ACT_REDIRECT:
+               /* skb_mac_header check was done by cls/act_bpf, so
+                * we can safely push the L2 header back before
+                * redirecting to another netdev
+                */
+               __skb_push(skb, skb->mac_len);
+               skb_do_redirect(skb);
+               return NULL;
+struct redirect_info {
+       u32 ifindex;
+       u32 flags;
+static DEFINE_PER_CPU(struct redirect_info, redirect_info);
+int skb_do_redirect(struct sk_buff *skb)
+       struct redirect_info *ri = this_cpu_ptr(&redirect_info);
+       struct net_device *dev;
+       dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
+       ri->ifindex = 0;
+       if (BPF_IS_REDIRECT_INGRESS(ri->flags))
+               return dev_forward_skb(dev, skb);
+       skb->dev = dev;
+       return dev_queue_xmit(skb);

If it returns TC_ACT_REDIRECT, then the skb_do_redirect() will perform the real redirection.

1.3 Call stack

pkt -> NIC -> TC ingress -> handle_ing()
                             |-verdict = tc_classify()     // exec BPF code
                             |            |-bpf_redirect() // return verdict
                             |-switch (verdict) {
                               case TC_ACK_REDIRECT:
                                   skb_do_redirect()       // to the target net device
                                     |-if ingress:
                                     |   dev_forward_skb()

We can see from the last few lines that bpf_redirect() supports both ingress and egress redirection.

2 Egress optimization: bpf_redirect_neighbor(), 2020

Five years after bpf_redirect() emerged in the Linux kernel, an egress side optimization was introduced for it, in patch:

bpf: Add redirect_neigh helper as redirect drop-in

Add a redirect_neigh() helper as redirect() drop-in replacement
for the xmit side. Main idea for the helper is to be very similar
in semantics to the latter just that the skb gets injected into
the neighboring subsystem in order to let the stack do the work
it knows best anyway to populate the L2 addresses of the packet
and then hand over to dev_queue_xmit() as redirect() does.

This solves two bigger items:
i) skbs don't need to go up to the stack on the host facing veth ingress side for traffic egressing
  the container to achieve the same for populating L2 which also has the huge advantage that
ii) the skb->sk won't get orphaned in ip_rcv_core() when entering the IP routing layer on the host stack.

Given that skb->sk neither gets orphaned when crossing the netns
as per 9c4c325 ("skbuff: preserve sock reference when scrubbing
the skb.") the helper can then push the skbs directly to the phys
device where FQ scheduler can do its work and TCP stack gets proper
backpressure given we hold on to skb->sk as long as skb is still
residing in queues.

With the helper used in BPF data path to then push the skb to the
phys device, I observed a stable/consistent TCP_STREAM improvement
on veth devices for traffic going container -> host -> host ->
container from ~10Gbps to ~15Gbps for a single stream in my test

2.1 Comparison with bpf_redirect()

  bpf_redirect() bpf_redirect_neighbor()
Supported direction ingress & egress egress only
Fill L2 addr by kernel stack (neighbor subsys) No Yes (fills in e.g. MAC addr based on L3 info from the pkt)
Redirect across netns No No
Others   flags argument is reserved and must be 0; currently only supported for tc BPF program types

Return: TC_ACT_REDIRECT on success or TC_ACT_SHOT on error.

2.2 Kernel implementations/changes

1. Modify skb_do_redirect(), prefer the new one whenever available

 int skb_do_redirect(struct sk_buff *skb)
        struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
        struct net_device *dev;
+       u32 flags = ri->flags;

        dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->tgt_index);
        ri->tgt_index = 0;
@@ -2231,7 +2439,22 @@ int skb_do_redirect(struct sk_buff *skb)
                return -EINVAL;

-       return __bpf_redirect(skb, dev, ri->flags);
+       return flags & BPF_F_NEIGH ?
+              __bpf_redirect_neigh(skb, dev) :
+              __bpf_redirect(skb, dev, flags);

2. Add bpf_redirect_neigh() helper/wrapper and syscall

Omitted, see the call stack below.

2.3 Call stack

  |-__bpf_redirect_neigh(skb, dev) :
          |-rt = ip_route_output_flow()
          |-skb_dst_set(skb, &rt->dst);
          |-bpf_out_neigh_v4(net, skb)
              |-neigh = ip_neigh_for_gw(rt, skb, &is_v6gw);
              |-sock_confirm_neigh(skb, neigh);
              |-neigh_output(neigh, skb, is_v6gw); // xmit with L2 header properly set

Note that although this is an optimization to bpf_redirec(), it still needs to go through the entire kernel stack.

3 Ingress optimization: bpf_redirect_peer(), 2020

Introduced in patch, 2020,

bpf: Add redirect_peer helper

Add an efficient ingress to ingress netns switch that can be used out of tc BPF
programs in order to redirect traffic from host ns ingress into a container
veth device ingress without having to go via CPU backlog queue [0].

For local containers this can also be utilized and path via CPU backlog queue only needs
to be taken once, not twice. On a high level this borrows from ipvlan which does
similar switch in __netif_receive_skb_core() and then iterates via another_round.
This helps to reduce latency for mentioned use cases.

3.1 Comparison with bpf_redirect()

  bpf_redirect() bpf_redirect_peer()
Supported direction ingress & egress ingress only
Redirect across netns No Yes (the netns switch takes place from ingress to ingress without going through the CPU's backlog queue)
Others   flags argument is reserved and must be 0; currently only supported for tc BPF program types; peer device must reside in a different netns

Return: TC_ACT_REDIRECT on success or TC_ACT_SHOT on error.

3.2 Kernel implementations/changes

1. Add new redirection flags

diff --git a/net/core/filter.c b/net/core/filter.c
index 5da44b11e1ec..fab951c6be57 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -2380,8 +2380,9 @@ static int __bpf_redirect_neigh(struct sk_buff *skb, struct net_device *dev)

 /* Internal, non-exposed redirect flags. */
 enum {
-       BPF_F_NEIGH = (1ULL << 1),
+       BPF_F_NEIGH     = (1ULL << 1),
+       BPF_F_PEER      = (1ULL << 2),

2. Add helper/syscall

+BPF_CALL_2(bpf_redirect_peer, u32, ifindex, u64, flags)
+       struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+       if (unlikely(flags))
+               return TC_ACT_SHOT;
+       ri->flags = BPF_F_PEER;
+       ri->tgt_index = ifindex;
+       return TC_ACT_REDIRECT;

3. Allow to re-enter TC ingress processing (for the peer device here)

@@ -5163,7 +5167,12 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
        if (static_branch_unlikely(&ingress_needed_key)) {
-               skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
+               bool another = false;
+               skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
+                                        &another);
+               if (another)
+                       goto another_round;

sch_handle_ingress() changes:

@@ -4974,7 +4974,11 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
                 * redirecting to another netdev
                __skb_push(skb, skb->mac_len);
-               skb_do_redirect(skb);
+               if (skb_do_redirect(skb) == -EAGAIN) {
+                       __skb_pull(skb, skb->mac_len);
+                       *another = true;
+                       break;
+               }

 int skb_do_redirect(struct sk_buff *skb)
        struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+       struct net *net = dev_net(skb->dev);
        struct net_device *dev;
        u32 flags = ri->flags;
-       dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->tgt_index);
+       dev = dev_get_by_index_rcu(net, ri->tgt_index);
        ri->tgt_index = 0;
+       ri->flags = 0;
+       if (flags & BPF_F_PEER) {
+               const struct net_device_ops *ops = dev->netdev_ops;
+               dev = ops->ndo_get_peer_dev(dev);
+               if (unlikely(!dev || !is_skb_forwardable(dev, skb) || net_eq(net, dev_net(dev))))
+                       goto out_drop;
+               skb->dev = dev;
+               return -EAGAIN;
        return flags & BPF_F_NEIGH ?  __bpf_redirect_neigh(skb, dev) : __bpf_redirect(skb, dev, flags);

3.3 Call stack

With a picture from my networking stack post,

Fig. Entering kernel stack: L2 processing steps

  |-// Device driver processing, e.g. update device's rx/tx stats
  |-another_round:  <------------------<-----------+
  |-// Generic XDP processing                      |
  |                                                | with skb->dev changed to the peer device, the next round
  |-// Tap processing if not skipped               | "G-XDP -> TAP -> TC" processings will be for the peer device,
  |                                                | which means we successfully bypassed tons of stuffs
  |-// TC BPF ingress processing if not skipped    | (and entered container's netns from the default/host netns)
  |-sch_handle_ingress(&another)                   | as shown in the above picture
  |-if another:                                    |
  |   goto another_round -------------->-----------+
  |-// Netfilter processing if not skipped

Some explanations:

  1. The first time execution of sch_handle_ingress() is for the current network device (e.g. eth0 of the physical host);
  2. If it returns with another==true, then the execution will go to another_round; then
  3. We come to sch_handle_ingress() for the second time, and this time, we are executing on the TC ingress hook of the device (e.g. the eth0 inside a container) we redirected to.

3.4 Use cases & performance evaluation

Two scenarios in Cilium networking solution:

  1. Physical NIC -> container NIC redirection
  2. "Container A -> container B" redirection in the same host

And in Cilium, this behavior is controlled by a dedicated option --enable-host-legacy-routing=true/false:

  1. With true: turn off peer redirection optimization, still goes through the entire kernel stack as normal;
  2. With false: turn on peer redirection (if kernel supports), dramatic performance gain is expected.

Performance benchmark see Cilium 1.9 Release Notes, and we have double confirmed the benchmark with Cilium 1.10.7 + 5.10 kernel in our clusters.

3.5 Impacts & known issues

Kubernetes: incorrect Pod ingress statistics

kubelet collects each pod’s network stats (e.g. rx_packets, rx_bytes) via cadvisor/netlink, and exposes these metrics via 10250 metrics port.

See kubelet doc for more information.

On a Kubernetes node:

$ curl -H 'Authorization: Bearer eyJh...' -X GET -d '{"num_stats": 1,"containerName": "/kubepods","subContainers": true}' --insecure > stats.json

Get stats of a specific pod:

$ cat stats.json | jq '."/kubepods/burstable/pod42ef7dc5-a27f-4ee5-ac97-54c3ce93bc9b/47585f9593"."stats"[0]."network"' | head -n8 | egrep "(rx_bytes|rx_packets|tx_bytes|tx_packets)"
  "rx_bytes": 34344009,
  "rx_packets": 505446,
  "tx_bytes": 5344339889,
  "tx_packets": 6214124,

The data source of these statistics actually come from sysfs/procfs. Such as, a pod’s rx_bytes is retrieved by cat /sys/class/net/<device>/statistics/rx_bytes where <device> is the pod’s network interface.

Problems arise when using bpf_redirect_peer(), as packets fly directly to the TC ingress point of a Pod from the physical NIC’s TC ingress point, which skips the driver processing steps of the pod's NIC, so rx/tx statistics like the above will not be correctly updated (only few packets will go through the driver). As a result, ingress statistics such as pps/bandwidth will be nearly zeros.