mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
synced 2026-08-28 01:43:47 -04:00
509ca545d4
We have several use cases where a pod injects traffic into the datapath of another so that the traffic appears to have originated from that pod. One such use case is a synthetic flow generator which injects synthetic traffic into a pod's datapath to enable dynamic probing and debugging. Another is a transparent proxy where connections originating from one pod are redirected towards another which proxies that connection. The new connection is bound to the IP of the original pod using IP_TRANSPARENT and its traffic is injected into that pod's datapath and handled as if it had originated there. This can be used for mTLS, etc. We use bpf_redirect(BPF_F_INGRESS) to direct traffic leaving the proxy, flow generator, etc. towards the target pod, ensuring that eBPF programs that are meant to intercept traffic leaving that pod are executed. However, this doesn't work with netkit. With netkit, an ingress redirection from proxy to workload skips eBPF programs that are meant to intercept traffic leaving the pod, since they reside on the netkit peer device. One workaround is to attach the same program to both the netkit peer device and the TCX ingress hook for the netkit pair's primary interface, but a) This seems hacky and we need to be careful not to run the same program twice for the same skb in cases where we want to pass that traffic to the host stack. b) We're trying to keep the proxy redirection / traffic injection systems as modular and separated from Cilium as possible, the system that manages netkit setup and core eBPF programming. It would be handy if instead we could redirect traffic directly from one netkit peer device to another. This patch proposes an extension to bpf_redirect_peer to allow us to do just that. With this patch, the BPF_F_EGRESS flag tells bpf_redirect_peer to emit the skb in the egress direction of the target interface's peer device While the main use case is netkit, I suppose you could also use this mode with veth as well if, e.g., there were some eBPF programs attached to that side of the veth pair that needed to intercept traffic. +---------------------------------------------------------------------+ | +-------------------------+ 6. bpf_redirect_neigh(eth0) | | | pod (10.244.0.10) | ------------------------ | | | | | | | | | +--------+ | | +---------+ | | | | 1. packet -->| | | | | | | | | | leaves ^ | netkit |<===========|======| netkit | | | | | | | peer |=======(eBPF)=====>| primary | | | | | | | | | | | | | | | | | +--------+ | | +---------+ | | | | | | | 2. bpf_redirect v | | +-----------|-------------+ |___________________ +-------| | | | | eth0 | | | 5. bpf_redirect_peer(BPF_F_EGRESS) | +-------| | |________________________ | | | +-------------------------+ | | | | | proxy (10.244.0.11) | | | | | | IP_TRANSPARENT | | | | | | +--------+ | | +---------+ | | | | 3. packet <--| | | | | |<-- | | | enters | netkit |<===========|======| netkit | | | | [proxy] | peer |=======(eBPF)=====>| primary | | | | 4. packet -->| | | | | | | | leaves +--------+ | +---------+ | | | sip=10.244.0.10 | | | +-------------------------+ | +---------------------------------------------------------------------+ Using the proxy use case as an example, in step 5 we would redirect traffic leaving the proxy towards the pod's peer device using bpf_redirect_peer(BPF_F_EGRESS). As a bonus, since the skb doesn't have to go through the backlog queue it can take full advantage of netkit's performance benefits. I set up a test where outgoing iperf3 traffic is injected into the datapath of another pod using either bpf_redirect_peer(BPF_F_EGRESS) or bpf_redirect(BPF_F_INGRESS). I used Cilium's eBPF host routing mode which skips the host stack and uses BPF redirect helpers to do all the routing. (net.ipv4.tcp_congestion_control=cubic,mtu=1500,100GiB link,Cilium eBPF host routing mode) BASELINE [bpf_redirect(BPF_F_INGRESS)] 1. [iperf pod] ==bpf_redirect([pod b], BPF_F_INGRESS)==> [pod b] 2. [pod b] ==bpf_redirect_neigh([eth0])==> eth0 3. eth0 ==over network==> [host b] [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-60.00 sec 231 GBytes 33.0 Gbits/sec 12060 sender [ 5] 0.00-60.00 sec 230 GBytes 33.0 Gbits/sec receiver TEST [bpf_redirect_peer(BPF_F_EGRESS)] 1. [iperf pod] ==bpf_redirect_peer([pod b], BPF_F_EGRESS)==> [pod b] 2. [pod b] ==bpf_redirect_neigh([eth0])==> eth0 3. eth0 ==over network==> [host b] [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-60.00 sec 272 GBytes 38.9 Gbits/sec 0 sender [ 5] 0.00-60.00 sec 272 GBytes 38.9 Gbits/sec receiver In this test, using bpf_redirect_peer(BPF_F_EGRESS) for the hop from [iperf pod] to [pod b] led to ~18% more throughput compared to bpf_redirect(BPF_F_INGRESS). Signed-off-by: Jordan Rife <jordan@jrife.io> Acked-by: Daniel Borkmann <daniel@iogearbox.net> Acked-by: Paul Chaignon <paul.chaignon@gmail.com> Reviewed-by: Jiayuan Chen <jiayuan.chen@linux.dev> Link: https://lore.kernel.org/r/20260618182035.43811-2-jordan@jrife.io Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Why we want a copy of kernel headers in tools?
==============================================
There used to be no copies, with tools/ code using kernel headers
directly. From time to time tools/perf/ broke due to legitimate kernel
hacking. At some point Linus complained about such direct usage. Then we
adopted the current model.
The way these headers are used in perf are not restricted to just
including them to compile something.
There are sometimes used in scripts that convert defines into string
tables, etc, so some change may break one of these scripts, or new MSRs
may use some different #define pattern, etc.
E.g.:
$ ls -1 tools/perf/trace/beauty/*.sh | head -5
tools/perf/trace/beauty/arch_errno_names.sh
tools/perf/trace/beauty/drm_ioctl.sh
tools/perf/trace/beauty/fadvise.sh
tools/perf/trace/beauty/fsconfig.sh
tools/perf/trace/beauty/fsmount.sh
$
$ tools/perf/trace/beauty/fadvise.sh
static const char *fadvise_advices[] = {
[0] = "NORMAL",
[1] = "RANDOM",
[2] = "SEQUENTIAL",
[3] = "WILLNEED",
[4] = "DONTNEED",
[5] = "NOREUSE",
};
$
The tools/perf/check-headers.sh script, part of the tools/ build
process, points out changes in the original files.
So its important not to touch the copies in tools/ when doing changes in
the original kernel headers, that will be done later, when
check-headers.sh inform about the change to the perf tools hackers.
Another explanation from Ingo Molnar:
It's better than all the alternatives we tried so far:
- Symbolic links and direct #includes: this was the original approach but
was pushed back on from the kernel side, when tooling modified the
headers and broke them accidentally for kernel builds.
- Duplicate self-defined ABI headers like glibc: double the maintenance
burden, double the chance for mistakes, plus there's no tech-driven
notification mechanism to look at new kernel side changes.
What we are doing now is a third option:
- A software-enforced copy-on-write mechanism of kernel headers to
tooling, driven by non-fatal warnings on the tooling side build when
kernel headers get modified:
Warning: Kernel ABI header differences:
diff -u tools/include/uapi/drm/i915_drm.h include/uapi/drm/i915_drm.h
diff -u tools/include/uapi/linux/fs.h include/uapi/linux/fs.h
diff -u tools/include/uapi/linux/kvm.h include/uapi/linux/kvm.h
...
The tooling policy is to always pick up the kernel side headers as-is,
and integate them into the tooling build. The warnings above serve as a
notification to tooling maintainers that there's changes on the kernel
side.
We've been using this for many years now, and it might seem hacky, but
works surprisingly well.