Re: [PATCH bpf-next 3/5] libbpf: add low level TC-BPF API

Andrii Nakryiko [off-list ref] writes:

On Wed, Apr 14, 2021 at 3:51 PM Toke Høiland-Jørgensen [off-list ref] wrote:

quoted

Andrii Nakryiko [off-list ref] writes:

quoted

On Wed, Apr 14, 2021 at 3:58 AM Toke Høiland-Jørgensen [off-list ref] wrote:

quoted

Andrii Nakryiko [off-list ref] writes:

quoted

On Tue, Apr 6, 2021 at 3:06 AM Toke Høiland-Jørgensen [off-list ref] wrote:

quoted

Andrii Nakryiko [off-list ref] writes:

quoted

On Sat, Apr 3, 2021 at 10:47 AM Alexei Starovoitov
[off-list ref] wrote:

quoted

On Sat, Apr 03, 2021 at 12:38:06AM +0530, Kumar Kartikeya Dwivedi wrote:

quoted

On Sat, Apr 03, 2021 at 12:02:14AM IST, Alexei Starovoitov wrote:

quoted

On Fri, Apr 2, 2021 at 8:27 AM Kumar Kartikeya Dwivedi [off-list ref] wrote:

quoted

[...]

All of these things are messy because of tc legacy. bpf tried to follow tc style
with cls and act distinction and it didn't quite work. cls with
direct-action is the only
thing that became mainstream while tc style attach wasn't really addressed.
There were several incidents where tc had tens of thousands of progs attached
because of this attach/query/index weirdness described above.
I think the only way to address this properly is to introduce bpf_link style of
attaching to tc. Such bpf_link would support ingress/egress only.
direction-action will be implied. There won't be any index and query
will be obvious.

Note that we already have bpf_link support working (without support for pinning
ofcourse) in a limited way. The ifindex, protocol, parent_id, priority, handle,
chain_index tuple uniquely identifies a filter, so we stash this in the bpf_link
and are able to operate on the exact filter during release.

Except they're not unique. The library can stash them, but something else
doing detach via iproute2 or their own netlink calls will detach the prog.
This other app can attach to the same spot a different prog and now
bpf_link__destroy will be detaching somebody else prog.

quoted

quoted

So I would like to propose to take this patch set a step further from
what Daniel said:
int bpf_tc_attach(prog_fd, ifindex, {INGRESS,EGRESS}):
and make this proposed api to return FD.
To detach from tc ingress/egress just close(fd).

You mean adding an fd-based TC API to the kernel?

yes.

I'm totally for bpf_link-based TC attachment.

But I think *also* having "legacy" netlink-based APIs will allow
applications to handle older kernels in a much nicer way without extra
dependency on iproute2. We have a similar situation with kprobe, where
currently libbpf only supports "modern" fd-based attachment, but users
periodically ask questions and struggle to figure out issues on older
kernels that don't support new APIs.

+1; I am OK with adding a new bpf_link-based way to attach TC programs,
but we still need to support the netlink API in libbpf.

quoted

So I think we'd have to support legacy TC APIs, but I agree with
Alexei and Daniel that we should keep it to the simplest and most
straightforward API of supporting direction-action attachments and
setting up qdisc transparently (if I'm getting all the terminology
right, after reading Quentin's blog post). That coincidentally should
probably match how bpf_link-based TC API will look like, so all that
can be abstracted behind a single bpf_link__attach_tc() API as well,
right? That's the plan for dealing with kprobe right now, btw. Libbpf
will detect the best available API and transparently fall back (maybe
with some warning for awareness, due to inherent downsides of legacy
APIs: no auto-cleanup being the most prominent one).

Yup, SGTM: Expose both in the low-level API (in bpf.c), and make the
high-level API auto-detect. That way users can also still use the
netlink attach function if they don't want the fd-based auto-close
behaviour of bpf_link.

So I thought a bit more about this, and it feels like the right move
would be to expose only higher-level TC BPF API behind bpf_link. It
will keep the API complexity and amount of APIs that libbpf will have
to support to the minimum, and will keep the API itself simple:
direct-attach with the minimum amount of input arguments. By not
exposing low-level APIs we also table the whole bpf_tc_cls_attach_id
design discussion, as we now can keep as much info as needed inside
bpf_link_tc (which will embed bpf_link internally as well) to support
detachment and possibly some additional querying, if needed.

But then there would be no way for the caller to explicitly select a
mechanism? I.e., if I write a BPF program using this mechanism targeting
a 5.12 kernel, I'll get netlink attachment, which can stick around when
I do bpf_link__disconnect(). But then if the kernel gets upgraded to
support bpf_link for TC programs I'll suddenly transparently get
bpf_link and the attachments will go away unless I pin them. This
seems... less than ideal?

That's what we are doing with bpf_program__attach_kprobe(), though.
And so far I've only seen people (privately) saying how good it would
be to have bpf_link-based TC APIs, doesn't seem like anyone with a
realistic use case prefers the current APIs. So I suspect it's not
going to be a problem in practice. But at least I'd start there and
see how people are using it and if they need anything else.

*sigh* - I really wish you would stop arbitrarily declaring your own use
cases "realistic" and mine (implied) "unrealistic". Makes it really hard
to have a productive discussion...

Well (sigh?..), this wasn't my intention, sorry you read it this way.
But we had similar discussions when I was adding bpf_link-based XDP
attach APIs.

Great, thank you! And yeah, we did discuss exactly this before, which is
where my mental sigh came from - I feel like we already covered this
ground and that I'm just being dismissed with "that is not a real use
case". But OK, I'll give it another shot, see below.

And guess what, now I see that samples/bpf/whatever_xdp is switched to
bpf_link-based XDP, because that makes everything simpler and more
reliable. What I also know is that in production we ran into multiple
issues with anything that doesn't auto-detach on process exit/crash
(unless pinned explicitly, of course). And that people that are trying
to use TC right now are saying how having bpf_link-based TC APIs would
make everything *simpler* and *safer*. So I don't know... I understand
it might be convenient in some cases to not care about a lifetime of
BPF programs you are attaching, but then there are usually explicit
and intentional ways to achieve at least similar behavior with safety
by default.

So I guess call me unconvinced (yet? still?). Give it another shot,
though.

I'm not arguing against adding bpf_link support, and I'm even fine with
making it the default. As you say, there are plenty of use cases where
the bpf_link semantics make sense, and the XDP programs in samples all
fall in this category. So sure, let's add this support and make this
convenient to use.

But there are also use cases where the BPF program lifetime absolutely
shouldn't follow that of the userspace application. This includes both
applications that don't have a long-running daemon at all (like a
firewall that just loads a ruleset at boot; xdp-filter is such an
application in the BPF world, but I'm sure there are others). And
daemons that use BPF as a data path and want the packets to keep flowing
even when they restart, like Cilium as Daniel mentioned.

So the latter category of applications need their BPF programs to be
permanently attached to the interface. And sure, this can sorta be done
by pinning the bpf_link; but not really, because then:

- You incur a new dependency on bpffs, so you have to make sure that is
  mounted and that you can get at the particular fs instance you're
  using; the latter is especially painful if you switch namespaces.

- Your BPF program lifetime is no longer tied to the interface, so you
  have to deal with garbage collecting your pinned files somehow. This
  is especially painful if you don't have a daemon.

Together, these issues make bpf_link a much less compelling proposition,
to the point where it's no longer the better API for these use cases,
IMO. And I know that because I had to work around just these issues with
bpf_link for xdp-tools.

But I'm not even asking for the netlink API to be the default, I'm
fine with bpf_link being the default and encouraged API. I'm just asking
for a way to make it *possible* to select which attach mode I want.
Either by a flag to bpf_program__attach_tc(), or by exposing the
low-level bpf_tc_cls_*() netlink functions, like we do for XDP.

quoted

quoted

quoted

If we expose the low-level API I can elect to just use this if I know I
want netlink behaviour, but if bpf_program__attach_tc() is the only API
available it would at least need a flag to enforce one mode or the other
(I can see someone wanting to enforce kernel bpf_link semantics as well,
so a flag for either mode seems reasonable?).

Sophisticated enough users can also do feature detection to know if
it's going to work or not.

Sure, but that won't help if there's no API to pick the attach mode they
want.

I'm not intending to allow legacy kprobe APIs to be "chosen", for
instance. Because I'm convinced it's a bad API that no one should use
if they can use an FD-based one.

I'd tend to agree with you for the tracing APIs, actually. But a
BPF-based data plane is different, as I tried to explain above.

It might be a different case for TC, who knows. I'd just start with
safer APIs and then evaluate whether there is a real demand for less
safe ones. It's just some minor refactoring and exposing more APIs,
when/if we need them.

There you go again with the "real demand" argument. How can I read this
in any other way than that you don't consider my use case "real" (as you
just assured me above was not the case)? What do you consider "real
demand"?

-Toke