Re: [PATCH bpf-next 3/3] selftests/bpf: allow compiling BPF objects without BTF

On Mon, Mar 22, 2021 at 10:54 AM Alexei Starovoitov
[off-list ref] wrote:

On Mon, Mar 22, 2021 at 09:56:19AM -0700, Andrii Nakryiko wrote:

quoted

On Sun, Mar 21, 2021 at 6:07 PM Alexei Starovoitov
[off-list ref] wrote:

quoted

On Sat, Mar 20, 2021 at 10:00:57AM -0700, Andrii Nakryiko wrote:

quoted

On Fri, Mar 19, 2021 at 7:22 PM Alexei Starovoitov
[off-list ref] wrote:

quoted

On Fri, Mar 19, 2021 at 01:59:09PM -0700, Andrii Nakryiko wrote:

quoted

Add ability to skip BTF generation for some BPF object files. This is done
through using a convention of .nobtf.c file name suffix.

Also add third statically linked file to static_linked selftest. This file has
no BTF, causing resulting object file to have only some of DATASEC BTF types.
It also is using (from BPF code) global variables. This tests both libbpf's
static linking logic and bpftool's skeleton generation logic.

I don't like the long term direction of patch 1 and 3.
BTF is mandatory for the most bpf kernel features added in the last couple years.
Making user space do quirks for object files without BTF is not something
we should support or maintain. If there is no BTF the linker and skeleton
generation shouldn't crash, of course, but they should reject such object.

I don't think tools need to enforce any policies like that. They are
tools and should be unassuming about the way they are going to be used
to the extent possible.

Right and bpftool/skeleton was used with BTF since day one.
Without BTF the skeleton core ideas are lost. The skeleton api
gives no benefit. So what's the point of adding support for skeleton without BTF?
Is there a user that would benefit? If so, what will they gain from
such BTF-less skeleton?

The only part of skeleton API that's not available is convenient
user-space access to global variables. If you don't use global
variables you don't use BTF at all with skeleton. So all features but
one work without BTF just fine: compile-time maps and progs (and
links) references, embedding object file in .skel.h, and even
automatic memory-mapping of .data/.rodata/.bss (just unknown struct
layout).

Compile-time maps and progs and separately object file embedding in C
header are useful in their own rights, even individually. There is no
single "core idea" of the BPF skeleton in my mind. What is it for you?

So given none of the fixes are horrible hacks and won't incur
additional maintenance costs, what's the problem with accepting them?

Because they double the maintenance cost now and double the support forever.
We never needed to worry about skeleton without BTF and now it would be
a thing ? So all tests realistically need to be doubled: with and without BTF.

2x? Realistically?.. No, I wouldn't say so. :) Extra test or a few
might be warranted, but doubling the amount of testing is a huge
exaggeration.

Even more so for static linking. If one .o has BTF and another doesn't
what linker suppose to do? Keep it, but the linked BTF will sort of cover
both .o-s, but line info in some funcs will be missing.
All these weird combinations would need to be tested.

BPF static linker already supports that mode, btw. And yes, it
shouldn't crash the kernel. And you don't need a skeleton or static
linker to do that to the kernel, so I don't know how that is a new
mode of operation.

The sensible thing to do would be to reject skel generation without BTF
and reject linking without BTF. The user most likely forgot -g during
compilation of bpf prog. The bpftool should give the helpful message
in such case. Whether it's generating skel or linking. Silently proceeding
and generating half-featured skeleton is not what user wanted.

Sure, a warning message makes sense. Outright disabling this - not so
much. I still can't get why I can't get BPF skeleton and static
linking without BTF, if I really want to. Both are useful without BTF.

So I don't know, it's the third different argument I'm addressing
without any conclusion on the previous two. It, sadly, feels rather
like fighting subjective personal preferences, rather than a
discussion.