Re: [PATCH net-next v4 2/3] virtio_net: add optional flow dissection in virtio_net_hdr_to_skb

On Thu, Jun 10, 2021 at 10:11 PM Jason Wang [off-list ref] wrote:

在 2021/6/10 下午10:04, Willem de Bruijn 写道:

quoted

On Thu, Jun 10, 2021 at 1:25 AM Jason Wang [off-list ref] wrote:

quoted

在 2021/6/10 下午12:19, Alexei Starovoitov 写道:

quoted

On Wed, Jun 9, 2021 at 9:13 PM Jason Wang [off-list ref] wrote:

quoted

So I wonder why not simply use helpers to access the vnet header like
how tcp-bpf access the tcp header?

Short answer - speed.
tcp-bpf accesses all uapi and non-uapi structs directly.

Ok, this makes sense. But instead of coupling device specific stuffs
like vnet header and neediness into general flow_keys as a context.

It would be better to introduce a vnet header context which contains

1) vnet header
2) flow keys
3) other contexts like endian and virtio-net features

So we preserve the performance and decouple the virtio-net stuffs from
general structures like flow_keys or __sk_buff.

You are advocating for a separate BPF program that takes a vnet hdr
and flow_keys as context and is run separately after flow dissection?

Yes.

quoted

I don't understand the benefit of splitting the program in two in this manner.

It decouples a device specific attributes from the general structures
like flow keys. We have xen-netfront, netvsc and a lot of drivers that
works for the emulated devices. We could not add all those metadatas as
the context of flow keys.

What are device-specific attributes here? What kind of metadata?

The only metadata that can be passed with tuntap, pf_packet et al is
virtio_net_hdr.

I quite don't understand where xen-netfront et al come in.

That's why I suggest to use something more
generic like XDP from the start. Yes, GSO stuffs was disabled by
virtio-net on XDP but it's not something that can not be fixed. If the
GSO and s/g support can not be done in short time

An alternative interface does not address that we already have this
interface and it is already causing problems.

then a virtio-net
specific BPF program still looks much better than coupling virtio-net
metadata into flow keys or other general data structures.

quoted

Your previous comment mentions two vnet_hdr definitions that can get
out of sync. Do you mean v1 of this patch, that adds the individual
fields to bpf_flow_dissector?

No, I meant this part of the patch:


+static int check_virtio_net_hdr_access(struct bpf_verifier_env *env,
int off,
+                       int size)
+{
+    if (size < 0 || off < 0 ||
+        (u64)off + size > sizeof(struct virtio_net_hdr)) {
+        verbose(env, "invalid access to virtio_net_hdr off=%d size=%d\n",
+            off, size);
+        return -EACCES;
+    }
+    return 0;
+}
+


It prevents the program from accessing e.g num_buffers.

I see, thanks. See my response to your following point.

quoted

That is no longer the case: the latest
version directly access the real struct. As Alexei points out, doing
this does not set virtio_net_hdr in stone in the ABI. That is a valid
worry. But so this patch series will not restrict how that struct may
develop over time. A version field allows a BPF program to parse the
different variants of the struct -- in the same manner as other
protocol headers.

The format of the virtio-net header depends on the virtio features, any
reason for another version? The correct way is to provide features in
the context, in this case you don't event need the endian hint.

That might work. It clearly works for virtio. Not sure how to apply it
to pf_packet or tuntap callers of virtio_net_hdr_to_skb.

quoted

If you prefer, we can add that field from the start.
I don't see a benefit to an extra layer of indirection in the form of
helper functions.

I do see downsides to splitting the program. The goal is to ensure
consistency between vnet_hdr and packet payload. A program split
limits to checking vnet_hdr against what the flow_keys struct has
extracted. That is a great reduction over full packet access.

Full packet access could be still done in bpf flow dissector.

quoted

For
instance, does the packet contain IP options? No idea.

I don't understand here. You can figure out this in flow dissector, and
you can extend the flow keys to carry out this information if necessary.

This I disagree with. flow_keys are a distillation/simplification of
the packet contents. It is unlikely to capture every feature of every
packet. We end up having to extend it for every new case we're
interested in. That is ugly and a lot of busywork. And for what
purpose? The virtio_net_hdr field prefaces the protocol headers in the
same buffer in something like tpacket. Processing the metadata
together with the data is straightforward. I don't see what isolation
or layering that breaks.

And if you want to have more capability, XDP which is designed for early
packet filtering is the right way to go which have even more functions
that a simple bpf flow dissector.

quoted

If stable ABI is not a concern and there are no different struct
definitions that can go out of sync, does that address your main
concerns?

I think not. Assuming we provide sufficient contexts (e.g the virtio
features), problem still: 1) coupling virtio-net with flow_keys

A flow dissection program is allowed to read both contents of struct
virtio_net_hdr and packet contents. virtio_net_hdr is not made part of
struct bpf_flow_keys. The pointer there is just a way to give access
to multiple data sources through the single bpf program context.

2) can't work for XDP.

This future work (AF_?)XDP based alternative to
pf_packet/tuntap/virtio does not exist yet, so it's hard to fully
prepare for. But any replacement interface will observe the same
issue: when specifying offloads like GSO/csum, that metadata may not
agree with actual packet contents. We have to have a way to validate
that. I could imagine that this XDP program attached to the AF_XDP
interface would do the validation itself? Is that what you mean?