Re: [PATCH] selftest: mm: Test if hugepage does not get leaked during __bio_release_pages()

[David Hildenbrand]

On 24.05.24 08:43, Ritesh Harjani (IBM) wrote:
> David Hildenbrand <david@xxxxxxxxxx> writes:
>
> dropping stable@xxxxxxxxxxxxxxx
>
> > On 24.05.24 04:57, Andrew Morton wrote:
> > > On Thu, 23 May 2024 22:40:25 +0200 David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >
> > > > > You have stable@xxxxxxxxxxxxxxx in the mail headers, so I assume you're
> > > > > proposing this for backporting.  When doing this, please include
> > > > >
> > > > > Cc: <stable@xxxxxxxxxxxxxxx>
> > > > >
> > > > > in the changelog footers and also include a Fixes: target.  I'm
> > > > > assuming the suitable Fixes: target for this patch is 38b43539d64b?
> > > >
> > > > This adds a new selfest to make sure what was fixed (and backported to
> > > > stable) remains fixed.
> > >
> > > Sure.  But we should provide -stable maintainers guidance for "how far
> > > back to go".  There isn't much point in backporting this into kernels
> > > where it's known to fail!
> >
> > I'm probably missing something important.
> >
> > 1) It's a test that does not fall into the common stable kernels
> > categories (see Documentation/process/stable-kernel-rules.rst).
> >
> > 2) If it fails in a kernel *it achieved its goal* of highlighting that
> > something serious is broken.
> >
> > > I'm still thinking that we want this in kernels which have 38b43539d64b?
> >
> > To hide that the other kernels are seriously broken and miss that fix?
> >
> > Really (1) this shouldn't be backported. I'm not even sure it should be
> > a selftest (sounds more like a reproducer that we usually attach to
> > commits, but that's too late). And if people care about backporting it,
> > (2) you really want this test to succeed everywhere. Especially also to
> > find kernels *without* 38b43539d64b
>
>
> Sorry about the noise and cc'd to stable. I believe we don't need to
> backport this test. The idea of adding a selftests was "also" to catch any
> future bugs like this.

Yes, for that purpose it's fine, but it has quite the "specific 
reproducer taste". Having it as part of something that is prepared to 
run against arbitrary kernels (which selftests frequently are not) to 
detect known problems feels better.

> I am unaware of any functional test suite where we could add such tests
> like how filesystems have fstests. Hence the ideas was to add this in
> selftests.

LTP has quite some MM testcases in testcases/kernel/mem/. But it often 
has a different focus (CVE or advanced feature/syscall tests). Now that 
most things are CVEs ... it might not be a bad fit ... :)

> So this begs the question which I also asked few people at LSFMM,
> Does mm has any mmfvt (mm functional verification tests)? Should we have
> something like this? Was it tried in past?

I think LTP is mostly the only "bigger" thing we have that is prepared 
to run against arbitrary kernel versions.

--
Cheers,

David / dhildenb