Re: selftests/bpf doesn't compile

From: Daniel Borkmann
Date: Mon Sep 18 2017 - 09:31:52 EST


On 09/16/2017 12:41 AM, Shuah Khan wrote:
On 09/15/2017 12:48 PM, Daniel Borkmann wrote:
On 09/15/2017 08:23 PM, Daniel Borkmann wrote:
On 09/15/2017 08:07 PM, Alexei Starovoitov wrote:
On Fri, Sep 15, 2017 at 05:58:40PM +0100, Edward Cree wrote:
On 15/09/17 17:02, Alexei Starovoitov wrote:
On Thu, Sep 14, 2017 at 09:33:48AM -0600, Shuah Khan wrote:
Is bpf test intended to be run in kselftest run? The clang dependency might
not be met on majority of the systems. Is this a hard dependency??
It is a hard dependency and clang should be present on majority of the systems.
I think this is the wrong approach. Making kselftest hard-require clang doesn't
mean that the bpf tests will be run more often, it means that the rest of the
kselftests will be run less often. clang is quite big (when I tried to install
it on one of my test servers, I didn't have enough disk space & had to go on a
clear-out of unused packages), and most people aren't interested in the bpf
subsystem specifically; they would rather be able to skip those tests.
I feel that as long as they know they are skipping some tests (so e.g. they
won't consider it a sufficient test of a kselftest refactor), that's fine.
It's not even as though all of the bpf tests require clang; the (smaller) tests
written directly in raw eBPF instructions could still be run on such a system.
So I think we should attempt to run as much as possible but accept that clang
may not be available and have an option to skip some tests in that case.

imo the value of selftests/bpf is twofold:
1. it helps bpf developers avoid regressions
2. as part of continuous integration it helps to catch bpf regressions
that were somehow caused by changes in other parts of the kernel

If a developer didn't bother to satisfy all bpf tests dependencies
(which includes clang) and ran all tests before sending a patch,
I don't want to see such patches. It just wastes maintainers time
to review code and spot bugs that could have been caught by tests.
Collectively we invested years of work into these tests and
developers better take advantage of it by running all.

+1

If a CI server didn't satisfy all bpf test dependencies,
I don't want such CI setup to be running and reporting results,
since it will give false sense of test coverage.
Test failures due to missing dependencies are hard failures.
We cannot skip them.

+1

Btw, on that note, the folks from zero-day bot run the BPF kselftests
for a while now just fine and they do run them together with clang,
so they have the full, proper coverage how it should be. It's not
how it used to be in the early days, you can just go and install
llvm/clang package on all the major distros today and you get the
bpf target by default enabled already.

I'd like generic XDP tests to be added to selftests/bpf which
would mean that the latest iproute2 will become a hard dependency
and bpf developers and CI host owners would need to upgrade
their iproute2.
The tests either pass or fail. Skipping them due to missing
dependencies is the same as fail and in that sense I don't want
to change selftests/bpf/Makefile to make it skip clang.

I fully agree that for the BPF selftests it is very desirable
to not only test the verifier with couple of BPF insn snippets,
but to actually load and run programs that more closely resemble
real world programs. For more complex interactions these snippets
are just limited, think of tail calls, testing perf event output
helper, etc, which would all require to write these tests with
restricted C when we add them (unless we want to make writing
these tests a real pain ;) in which case no-one will bother to
write tests at all for them). Mid to long term I would definitely
like to see more programs in BPF selftests (e.g. moved over from
samples/bpf/) to increase the test coverage.

As I said in my earlier email:

Unless users choose to install clang, bpf will always fail run without
clang. So clang dependency is an issue for bpf test coverage in general.
That is your choice as to whether you want to increase the scope of
regression test coverage for bpf or not.

I fully understand you have weigh that against ease of writing tests.

We can leave things the way they are since:

- You can't force users to install clang and run bpf test. Users might
opt for letting bpf test fail due to unmet dependency.

- You have reasons to continue use clang and you have been using it for
a longtime.

I'm definitely for leaving it as it currently is and having clang as
hard dependency; there will be more BPF selftests over time that will
require to compile BPF progs (through clang's BPF backend) written in
restricted C, so just skipping these tests would give a false sense
of coverage. clang is pretty much needed anyway for writing more complex
programs, thus leaving requirements the way they are is the much better
option.

I will try to see why make ksefltest fails on bpf even with my patch
series that addresses the source directory issue. All other tests build
and run. It is not an issue with bpf specifically, it is something that
has never been tested in this use-case.

Great, thanks!