Re: Should I add BPF kfuncs for userspace apps? And how?

From: Akihiko Odaki
Date: Thu Dec 14 2023 - 00:56:49 EST


On 2023/12/13 19:22, Benjamin Tissoires wrote:
On Tue, Dec 12, 2023 at 1:41 PM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:

On 2023/12/12 19:39, Benjamin Tissoires wrote:
Hi,

On Tue, Dec 12, 2023 at 9:11 AM Akihiko Odaki <akihiko.odaki@xxxxxxxxxx> wrote:

Hi,

Hi,

Thanks for reply.


It is said eBPF is a safe way to extend kernels and that is very
attarctive, but we need to use kfuncs to add new usage of eBPF and
kfuncs are said as unstable as EXPORT_SYMBOL_GPL. So now I'd like to ask
some questions:

1) Which should I choose, BPF kfuncs or ioctl, when adding a new feature
for userspace apps?
2) How should I use BPF kfuncs from userspace apps if I add them?

Here, a "userspace app" means something not like a system-wide daemon
like systemd (particularly, I have QEMU in mind). I'll describe the
context more below:

I'm probably not the best person in the world to answer your
questions, Alexei and others from the BPF core group are, but given
that you pointed at a thread I was involved in, I feel I can give you
a few pointers.

But first and foremost, I encourage you to schedule an agenda item in
the BPF office hour[4]. Being able to talk with the core people
directly was tremendously helpful to me to understand their point.

I prefer emails because I'm not very fluent when speaking in English and
may have a difficultly to listen to other people, but I may try it in
future.




---

I'm working on a new feature that aids virtio-net implementations using
tuntap virtual network device. You can see [1] for details, but
basically it's to extend BPF_PROG_TYPE_SOCKET_FILTER to report four more
bytes.

However, with long discussions we have confirmed extending
BPF_PROG_TYPE_SOCKET_FILTER is not going to happen, and adding kfuncs is
the way forward. So I decided how to add kfuncs to the kernel and how to
use it. There are rich documentations for the kernel side, but I found
little about the userspace. The best I could find is a systemd change
proposal that is based on WIP kernel changes[2].

Yes, as Alexei already replied, BPF is not adding new stable APIs,
only kfuncs. The reason being that once it's marked as stable, you
can't really remove it, even if you think it's badly designed and
useless.

Kfuncs, OTOH are "unstable" by default meaning that the constraints
around it are more relaxed.

However, "unstable" doesn't mean "unusable". It just means that the
kernel might or might not have the function when you load your program
in userspace. So you have to take that fact into account from day one,
both from the kernel side and the userspace side. The kernel docs have
a nice paragraph explaining that situation and makes the distinction
between relatively unused kfuncs, and well known established ones.

Regarding the systemd discussion you are mentioning ([2]), this is
something that I have on my plate for a long time. I think I even
mentioned it to Alexei at Kernel Recipes this year, and he frowned his
eyebrows when I mentioned it. And looking at the systemd code and the
benefits over a plain ioctl, it is clearer that in that case, a plain
ioctl is better, mostly because we already know the API and the
semantic.

A kfunc would be interesting in cases where you are not sure about the
overall design, and so you can give a shot at various API solutions
without having to keep your bad v1 design forever.


So now I'm wondering how I should use BPF kfuncs from userspace apps if
I add them. In the systemd discussion, it is told that Linus said it's
fine to use BPF kfuncs in a private infrastructure big companies own, or
in systemd as those users know well about the system[3]. Indeed, those
users should be able to make more assumptions on the kernel than
"normal" userspace applications can.

Returning to my proposal, I'm proposing a new feature to be used by QEMU
or other VMM applications. QEMU is more like a normal userspace
application, and usually does not make much assumptions on the kernel it
runs on. For example, it's generally safe to run a Debian container
including QEMU installed with apt on Fedora. BPF kfuncs may work even in
such a situation thanks to CO-RE, but it sounds like *accidentally*
creating UAPIs.

Considering all above, how can I integrate BPF kfuncs to the application?

FWIW, I'm not sure you can rely on BPF calls from a container. There
is a high chance the syscall gets disabled by the runtime.

Right. Container runtimes will not pass CAP_BPF by default, but that
restriction can be lifted and I think that's a valid scenario.



If BPF kfuncs are like EXPORT_SYMBOL_GPL, the natural way to handle them
is to think of BPF programs as some sort of kernel modules and
incorporate logic that behaves like modprobe. More concretely, I can put
eBPF binaries to a directory like:
/usr/local/share/qemu/ebpf/$KERNEL_RELEASE

I would advise against that (one program per kernel release). Simply
because your kfunc may or may not have been backported to kernel
release v6.X.Y+1 while it was not there when v6.X.Y was out. So
relying on the kernel number is just going to be a headache.

As I understand it, the way forward is to rely on the kernel, libbpf
and CO-RE: if the function is not available, the program will simply
not load, and you'll know that this version of the code is not
available (or has changed API).

So what I would do if some kfunc API is becoming deprecated, is
embedding both code paths in the same BPF unit, but marking them as
not loaded by libppf. Then I can load the compilation unit, try v2 of
the API, and if it's not available, try v1, and if not, then mention
that I can not rely on BPF. Of course, this can also be done with
separate compilation units.

Doesn't it mean that the kernel is free to break old versions of QEMU
including BPF programs? That's something I'd like to avoid.

Couple of points here:
- when you say "the kernel", it feels like you are talking about an
external actor tampering with your code. But if you submit a kernel
patch with a specific use case and get yourself involved in the
community, why would anybody change your kfunc API without you knowing
it?

You are right in the practical aspect. I can pay efforts to keep kfunc APIs alive and I'm also sure other developers would also try not to break them for good.

Nevertheless I'm being careful to evaluate APIs from both of the kernel and userspace (QEMU) viewpoints. If I fail to keep kfuncs stable because I die in an accident, for example, it's a poor excuse for other QEMU developers that I intended to keep them stable with my personal effort.

- the whole warning about "unstable" policy means that the user space
component should not take for granted the capability. So if the kfunc
changes/disappears for good reasons (because it was marked as well
used and deprecated for quite some time), qemu should not *break*, it
should not provide the functionality, or have a secondary plan.

But even if you are encountering such issues, in case of a change in
the ABI of your kfunc, it should be easy enough to backport the bpf
changes to your old QEMUs and ask users to upgrade the user space if
they upgrade their kernel.

AFAIU, it is as unstable as you want it to be. It's just that we are
not in the "we don't break user space" contract, because we are
talking about adding a kernel functionality from userspace, which
requires knowing the kernel intrinsics.

I must admit I'm still not convinced the proposed BPF program functionality needs to know internals of the kernel.

The eBPF program QEMU carries is just to calculate hashes from packets. It doesn't need to know the details of how the kernel handles packets. It only needs to have an access to the packet content.

It is exactly what BPF_PROG_TYPE_SOCKET_FILTER does, but it lacks a mechanism to report hash values so I need to extend it or invent a new method. Extending BPF_PROG_TYPE_SOCKET_FILTER is not a way forward since CO-RE is superior to the context rewrite it relies on. But apparently adopting kfuncs and CO-RE also means to lose the "we don't break user space" contract although I have no intention to expose kernel internals to the eBPF program.





Then, QEMU can uname() and get the path to the binary. It will give an
error if it can't find the binary for the current kernel so that it
won't create accidental UAPIs.

The obvious downside of this is that it complicates packaging a lot; it
requires packaging QEMU eBPF binaries each time a new kernel comes up.
This complexity is centrally managed by modprobe for kernel modules, but
apparently each application needs to take care of it for BPF programs.

For my primary use case: HID-BPF, I put kfuncs in kernel v6.3 and
given that I haven't touch this part of the API, the same compilation
unit compiled in the v6.3 era still works on a v6.7-rcx, so no, IMO
it's not complex and doesn't require to follow the kernel releases
(which is the whole point of HID-BPF FWIW).

I also expect BPF kfuncs will work well for long if I introduce its
usage to QEMU in practice. That said, the interface stability is about
when something unexpected happens. What if the interface QEMU relies on
is deemed sub-optimal? Without following kernel releases, QEMU may
accidentally lose the feature relying on eBPF.

In the same way, anybody can tamper with your ioctl or syscall without
QEMU knowing it.

I suppose it's only the case when the ioctl is marked as experimental, and it's also possible to mark it stable when the design is consolidated. The latter part matters; once I and other kernel/QEMU developers become confident, I'd like to drop the experimental notes from the kernel and QEMU and expect kfuncs to keep working.

And what you need to follow is not the kernel *releases*, but the
changes in the kfuncs you are interested in.

That works as long as a version of kfunc is stable across kernel releases. I have to ensure that a breaking change will always require a different symbol for that purpose.





In conclusion, I see too much complexity to use BPF in a userspace
application, which we didn't have to care for
BPF_PROG_TYPE_SOCKET_FILTER. Isn't there a better way? Or shouldn't I
use BPF in my case in the first place?

Given that I'm not a network person, I'm not sure about your use case,
but I would make my decision based on:
- do I know exactly what I want to achieve and I'm confident that I'll
write the proper kernel API from day one? (if not then kfuncs is
appealing because it's less workload in the long run, but userspace
needs to be slightly smarter)

Personally I'm confident that the initial UAPI design will not do a bad
thing at least. However, there is a high chance that the design needs to
be extended to accommodate new features.

Not trying to offend you or anything, but designs can change for
multiple reasons. Floppy disks were a good design at the time, and it
took decades to remove support for it in the kernel. In the same way,
removing an architecture from the kernel is hard, because even if you
can not run a new kernel on those architectures, "we do not break
userspace".

It's a totally valid point. I expect the underlying networking technique (RSS) will matter for the next decade, but I'm not sure after that.


The whole BPF approach is to say that users of BPF are not plain
random users, and they have to know a little bit of the kernel, and
they know that once the kfunc is here, it doesn't mean it'll stay here
forever.


- are all of my use cases covered by using BPF? (what happens if I run
QEMU in a container?) -> BPF might or might not be a solution

Yes. Containers can be used to 1) have a different userspace or 2)
isolate things for security.

Regarding 2), QEMU and libvirt has sandbox mechanisms so we can rely on
them instead of containers so we can just pass capabilities to the
container. At least, we can always have a setuid helper outside
container, and pass around file descriptors it generates.

So 1) is the only problem that matters.


But the nice thing about using BPF kfuncs is that it allows you to
have a testing (not-)UAPI kernel interface. You can then implement the
userspace changes and see how it behaves. And then, once you got the
right design, you can decide to promote it to a proper syscall or
ioctl if you want.

I expect it's possible to have testing ioctls. Quickly searching online,
there are experimental ioctls[1][2]. I also know DRM has a relaxed
policy for closed-source userspace[3].

Sure, but handling a change in the API in those cases is tough in the
kernel. You probably need to bump versions, return different values
depending on how many parameters you are given, and you are never sure
the caller is using the right parameters. BPF simplifies this by
actually checking the types of the caller, and if there is a
discrepancy, it'll notify userspace that it is doing something bad.

That's the whole reason I'm reluctant to have an ioctl while I intend to use it like a UAPI. BPF looks nice; it's safer, allows to extend the functionality when the virtio spec QEMU implements gets updated and to reuse the existing code for BPF_PROG_TYPE_SOCKET_FILTER.



So I'm seeing the distinction of UAPI/kfunc even less definitive; UAPIs
can also be broken if the subsystem maintainers agree and there is no
real user. I also think it's natural to say a kfunc will be stable as
long as there is a user, but it contradicts with the current situation.

Please read more carefully the kernel docs [4] (just quoting here the
beginning):

"""
Like any other change to the kernel, maintainers will not change or
remove a kfunc without having a reasonable justification. Whether or
not they'll choose to change a kfunc will ultimately depend on a
variety of factors, such as how widely used the kfunc is, how long the
kfunc has been in the kernel, whether an alternative kfunc exists,
what the norm is in terms of stability for the subsystem in question,
and of course what the technical cost is of continuing to support the
kfunc.
"""

kfunc is expressed as EXPORT_SYMBOL_GPL in the documentation, and Linus
expects kfunc is for users like big companies or systemd, which closely
follow the kernel, according to the systemd discussion I cited in the
last email.

Please re-read the doc[4], it's not a 1-to-1 matching to EXPORT_SYMBOL_GPL.
And being the one who reported Linus' words in that systemd thread,
Linus was not concerned about "potential situations that may or may
not happen", because he expected the people who use kfunc to do the
right thing. Because they are not average programmers. And QEMU
developers would definitely fit in that category IMO.

And the whole "you can consider kfunc similar to EXPORT_SYMBOL_GPL" is
just a warning for user space that the kfunc will never be kept only
for stability reasons. So when you want to use a kfunc, you need to be
aware of it and not segfault if it's not there (which can not happen
TBH unless you don't check that your program was correctly loaded).


According to the discussion above, it may be better off abandoning BPF
and implementing all in kernel, with ioctl as I have a (hopefully) sound
idea of UAPI design. But I'll also continue considering the BPF option;
BPF is still attractive due to its extensibility and safety.

We can not tell you to choose one solution over the other. The choice
is yours. I personally find BPF more appealing because it allows the
user space application to define its own kernel API for its own needs
while relying on just a few defined kfuncs.

But again, sometimes it doesn't work, like the systemd thread you
linked, it's too big overhead for little gain compared to an ioctl in
that particular case.

IMO the biggest issue for you is not the stability of the API, but the
container capabilities. Because allowing CAP_BPF allows for a whole
lot of nasty things to happen :)

I'll keep thinking about what QEMU developers can do for kfuncs, lacking UAPI stability, and if it makes sense to require CAP_BPF.

Regards,
Akihiko Odaki