Re: [PATCH bpf-next 1/4] xdp: Support specifying expected existing program when attaching XDP

Toke Høiland-Jørgensen wrote:

Andrii Nakryiko [off-list ref] writes:

quoted

On Mon, Mar 23, 2020 at 12:23 PM Toke Høiland-Jørgensen [off-list ref] wrote:

quoted

Andrii Nakryiko [off-list ref] writes:

quoted

On Mon, Mar 23, 2020 at 4:24 AM Toke Høiland-Jørgensen [off-list ref] wrote:

quoted

Andrii Nakryiko [off-list ref] writes:

quoted

On Fri, Mar 20, 2020 at 11:31 AM John Fastabend
[off-list ref] wrote:

quoted

Jakub Kicinski wrote:

quoted

On Fri, 20 Mar 2020 09:48:10 +0100 Toke Høiland-Jørgensen wrote:

quoted

Jakub Kicinski [off-list ref] writes:

quoted

On Thu, 19 Mar 2020 14:13:13 +0100 Toke Høiland-Jørgensen wrote:

quoted

From: Toke Høiland-Jørgensen <redacted>

While it is currently possible for userspace to specify that an existing
XDP program should not be replaced when attaching to an interface, there is
no mechanism to safely replace a specific XDP program with another.

This patch adds a new netlink attribute, IFLA_XDP_EXPECTED_FD, which can be
set along with IFLA_XDP_FD. If set, the kernel will check that the program
currently loaded on the interface matches the expected one, and fail the
operation if it does not. This corresponds to a 'cmpxchg' memory operation.

A new companion flag, XDP_FLAGS_EXPECT_FD, is also added to explicitly
request checking of the EXPECTED_FD attribute. This is needed for userspace
to discover whether the kernel supports the new attribute.

Signed-off-by: Toke Høiland-Jørgensen <redacted>

I didn't know we wanted to go ahead with this...

Well, I'm aware of the bpf_link discussion, obviously. Not sure what's
happening with that, though. So since this is a straight-forward
extension of the existing API, that doesn't carry a high implementation
cost, I figured I'd just go ahead with this. Doesn't mean we can't have
something similar in bpf_link as well, of course.

I'm not really in the loop, but from what I overheard - I think the
bpf_link may be targeting something non-networking first.

My preference is to avoid building two different APIs one for XDP and another
for everything else. If we have userlands that already understand links and
pinning support is on the way imo lets use these APIs for networking as well.

I agree here. And yes, I've been working on extending bpf_link into
cgroup and then to XDP. We are still discussing some cgroup-specific
details, but the patch is ready. I'm going to post it as an RFC to get
the discussion started, before we do this for XDP.

Well, my reason for being skeptic about bpf_link and proposing the
netlink-based API is actually exactly this, but in reverse: With
bpf_link we will be in the situation that everything related to a netdev
is configured over netlink *except* XDP.

One can argue that everything related to use of BPF is going to be
uniform and done through BPF syscall? Given variety of possible BPF
hooks/targets, using custom ways to attach for all those many cases is
really bad as well, so having a unifying concept and single entry to
do this is good, no?

Well, it depends on how you view the BPF subsystem's relation to the
rest of the kernel, I suppose. I tend to view it as a subsystem that
provides a bunch of functionality, which you can setup (using "internal"
BPF APIs), and then attach that object to a different subsystem
(networking) using that subsystem's configuration APIs.

Seeing as this really boils down to a matter of taste, though, I'm not
sure we'll find agreement on this :)

Yeah, seems like so. But then again, your view and reality don't seem
to correlate completely. cgroup, a lot of tracing,
flow_dissector/lirc_mode2 attachments all are done through BPF
syscall.

Well, I wasn't talking about any of those subsystems, I was talking
about networking :)

My experience has been that networking in the strict sense of XDP no
longer exists on its own without cgroups, flow dissector, sockops,
sockmap, tracing, etc. All of these pieces are built, patched, loaded,
pinned and otherwise managed and manipulated as BPF objects via libbpf.

Because I have all this infra in place for other items its a bit odd
imo to drop out of BPF apis to then swap a program differently in the
XDP case from how I would swap a program in any other place. I'm
assuming ability to swap links will be enabled at some point.

Granted it just means I have some extra functions on the side to manage
the swap similar to how 'qdisc' would be handled today but still not as
nice an experience in my case as if it was handled natively.

Anyways the netlink API is going to have to call into the BPF infra
on the kernel side for verification, etc so its already not pure
networking.

In particular, networking already has a consistent and fairly
well-designed configuration mechanism (i.e., netlink) that we are
generally trying to move more functionality *towards* not *away from*
(see, e.g., converting ethtool to use netlink).

True. But BPF programs are going to exist and interop with other
programs not exactly in the networking space. Actually library calls
might be used in tracing, cgroups, and XDP side. It gets a bit more
interesting if the "same" object file (with some patching) runs in both
XDP and sockops land for example.

quoted

LINK_CREATE provides an opportunity to finally unify all those
different ways to achieve the same "attach my BPF program to some
target object" semantics.

Well I also happen to think that "attach a BPF program to an object" is
the wrong way to think about XDP. Rather, in my mind the model is
"instruct the netdevice to execute this piece of BPF code".

quoted

quoted

quoted

quoted

Other than that, I don't see any reason why the bpf_link API won't work.
So I guess that if no one else has any problem with BPF insisting on
being a special snowflake, I guess I can live with it as well... *shrugs* :)

Apart from derogatory remark,

Yeah, should have left out the 'snowflake' bit, sorry about that...

quoted

BPF is a bit special here, because it requires every potential BPF
hook (be it cgroups, xdp, perf_event, etc) to be aware of BPF
program(s) and execute them with special macro. So like it or not, it
is special and each driver supporting BPF needs to implement this BPF
wiring.

All that is about internal implementation, though. I'm bothered by the
API discrepancy (i.e., from the user PoV we'll end up with: "netlink is
what you use to configure your netdev except if you want to attach an
XDP program to it").

See my reply to David. Depends on where you define user API. Is it
libbpf API, which is what most users are using? Or kernel API?

Well I'm talking about the kernel<->userspace API, obviously :)

quoted

If everyone is using libbpf, does kernel system (bpf syscall vs
netlink) matter all that much?

This argument works the other way as well, though: If libbpf can
abstract the subsystem differences and provide a consistent interface to
"the BPF world", why does BPF need to impose its own syscall API on the
networking subsystem?

I can make it work either way as a netlink or syscall its not going
to be a blocker. If we go netlink route then the next question is
does libbpf pull in the ability to swap XDP progs via netlink or
is that some other lib?

-Toke