On 5/28/20 2:09 PM, Shuah Khan wrote:
On 5/28/20 1:18 PM, Alexei Starovoitov wrote:
On Thu, May 28, 2020 at 12:59:06PM -0600, Shuah Khan wrote:
On 5/28/20 12:15 PM, Alexei Starovoitov wrote:
On Thu, May 28, 2020 at 11:07:09AM -0600, Shuah Khan wrote:
On 5/28/20 10:14 AM, Alexei Starovoitov wrote:
On Thu, May 28, 2020 at 12:56:31PM +0200, Greg KH wrote:
On Thu, May 28, 2020 at 10:05:57AM +0200, Jiri Benc wrote:
On Wed, 27 May 2020 15:23:13 -0700, Alexei Starovoitov wrote:
I prefer to keep selftests/bpf install broken.
This forced marriage between kselftests and selftests/bpf
never worked well. I think it's a time to free them up from
each other.
Alexei, it would be great if you could cooperate with other people
instead of pushing your own way. The selftests infrastructure
was put
to the kernel to have one place for testing. Inventing yet
another way
to add tests does not help anyone. You don't own the kernel. We're
community, we should cooperate.
I agree, we rely on the infrastructure of the kselftests
framework so
that testing systems do not have to create "custom" frameworks to
handle
all of the individual variants that could easily crop up here.
Let's keep it easy for people to run and use these tests, to not
do so
is to ensure that they are not used, which is the exact opposite
goal of
creating tests.
Greg,
It is easy for people (bpf developers) to run and use the tests.
Every developer runs them before submitting patches.
New tests is a hard requirement for any new features.
Maintainers run them for every push.
What I was and will push back hard is when other people (not bpf
developers)
come back with an excuse that some CI system has a hard time
running these
tests. It's the problem of weak CI. That CI needs to be fixed. Not
the tests.
The example of this is that we already have github/libbpf CI that
runs
selftests/bpf just fine. Anyone who wants to do another CI are
welcome to copy
paste what already works instead of burdening people (bpf
developers) who run
and use existing tests. I frankly have no sympathy to folks who
put their own
interest of their CI development in front of bpf community of
developers.
The main job of CI is to help developers and maintainers.
Where helping means to not impose new dumb rules on developers
because CI
framework is dumb. Fix CI instead.
Here is what CI users are requesting:
- ability to install bpf test with other selftests using kselftest
install. The common framework is in place and with minor changes
to bpf test Makefile, we can make this happen. Others and myself
are willing to work on this, so we can get bpf test coverage in
test rings.
so you're saying that bpf maintainers and all bpf developers now
would need to incorporate new 'make install' step to their workflow
because some unknown CI system that is not even functional decided
to do 'make install' ?
That's exactly my point about selfish CI developers who put their
needs in front of bpf community of developers.
There is no need change bpf maintainer and developer workflow. You
don't have to use install option. Kselftest framework doesn't
require a specific workflow and you can:
1. Build and run your tests from bpf directory if you choose to
2. Install to run on different target.
Adding install install option requires a change to bpf Makefile
only to copy test that are built to install directory.
make kselftest-install from the main kernel Makefile in conjunction
with selftests Makefile and lib.mk will handle all of that.
Sounds like there is a misunderstanding that bpf maintainer/developer
workflow will have to change to support install. That is not the case.
The reason kselftest exists on the first place is to have common
framework to take care of build/run/install as a common layer so
individual test writers don't have to worry about these details
and write tests instead.
I don't think you understand the 'make install' implications.
Not doing 'make install' means that all the Makefile changes that
Yauheni is proposing will immediately bit rot.
People are constantly adding new tests and changing makefile.
'make install' _will_ be broken instantly unless _humans_ incorporate
it in their patch development process and maintainer incorporate
that step into their workflow as well.
I don't think so. I think if you want to work with us on this, we can
find a way. bpf isn't such a unique test that adding install will break
it.
Ohh, but don't worry about this broken 'make install' is not an answer.
It's broken now and I don't want to see patches that move it
from one broken state into another broken state and at the same time
add complexity to it.
Well! What you are saying "I don;t want to collaborate to find a
solution to the problem".
That's very different from 'make install' doing 'cp -r' of the whole dir.
In such case the chances of it going stale and broken are much lower.
- be able to build and run existing tests without breaking the test
build when new tests are added that have hard dependency on new
versions of tools (llvm etc.). This isn't such a novel idea. We
don't break kernel builds every single release and even when we
require newer compiler releases. Plan the new tests with the
intent
to not break existing users and add new tests at the same time.
We use min rev and not bleeding edge as the requirement for kernel
build.
'existing users'?
I said existing tests not users. When you add new bpf tests, existing
tests should continue to build and run without dependency on new revs
of llvm.
Who said that existing test stop building with new llvm ?
Please check your facts.
I am basing this on a previous discussion on this topic. bpf test(s)
that build stop building and the solution always has been upgrade
to new tool chain. If this changed now and there is no hard tie to
bpf test building and llvm release, great.
If bpf test doesn't build and/or installed, it won't run
on test rings that qualify stable/next/main releases.
That's the case today and I prefer to keep it this way.
Why is that? Are you making a decision for the rest of the kernel
with this approach? If bpf doesn't get tested in stable test rings,
this isn't bpf problem alone. This is a kernel release issue.
stable folks ignored our recommendations on how selftests/bpf should
be run, so please don't come up with your ways of doing it and
complain that something doesn't work.
If you want to talk about history, bpf test was added in Oct 2016 and
by then kselftest was in use with its run_tests and install support.
So it is misleading to suggest that the framework came after the bpf
test. bpf test was added to existing kselftest framework used by all
other tests. Expecting the existing framework to change and adapt to
a new test isn't reasonable.
When a new test is added, it has to work within the framework. Framework
can be improved and changed, however not the way you are attempting to
do in this thread. You can do this like others in the community to do
by making changes and improvements.
In any case, kselftest framework lets you override INSTALL_RULE and
define your own. Same is true for other rules: RUN_TESTS, EMIT_TESTS,
CLEAN, INSTALL
bpf does this already for CLEAN.
# override lib.mk's default rules
OVERRIDE_TARGETS := 1
override define CLEAN
If generic install rule doesn't work for you do override so bpf so
users can install it.
thanks,
-- Shuah
