Re: [PATCH] selftests/mm: add THP sysfs interface test

[Breno Leitao <leitao@debian.org>]

On Mon, Mar 16, 2026 at 03:44:14PM +0100, David Hildenbrand (Arm) wrote:
> On 3/16/26 14:47, Breno Leitao wrote:
> > On Mon, Mar 16, 2026 at 12:55:13PM +0000, Lorenzo Stoakes (Oracle) wrote:
> >> On Mon, Mar 09, 2026 at 05:00:34AM -0700, Breno Leitao wrote:
> >>> Add a shell-based selftest that exercises the full set of THP sysfs
> >>> knobs: enabled (global and per-size anon), defrag, use_zero_page,
> >>> hpage_pmd_size, shmem_enabled (global and per-size), shrink_underused,
> >>> khugepaged/ tunables, and per-size stats files.
> >>>
> >>> Each writable knob is tested for valid writes, invalid-input rejection,
> >>> idempotent writes, and mode transitions where applicable. All original
> >>> values are saved before testing and restored afterwards.
> >>>
> >>> The test uses the kselftest KTAP framework (ktap_helpers.sh) for
> >>> structured TAP 13 output, making results parseable by the kselftest
> >>> harness. The test plan is printed at the end since the number of test
> >>> points is dynamic (depends on available hugepage sizes and sysfs files).
> >>>
> >>> This is particularly useful for validating the refactoring of
> >>> enabled_store() and anon_enabled_store() to use sysfs_match_string()
> >>> and the new change_enabled()/change_anon_orders() helpers.
> >>>
> >>> Signed-off-by: Breno Leitao <leitao@debian.org>
> >>
> >> The test is broken locally for me, returning error code 127.
> >>
> >> I do appreciate the effort here, so I'm sorry to push back negatively, but I
> >> feel a bash script here is pretty janky, and frankly if any of these interfaces
> >> were as broken as this it'd be a major failure that would surely get picked up
> >> far sooner elsewhere.
> >>
> >> So while I think this might be useful as a local test for your sysfs interface
> >> changes, I don't think this is really suited to the mm selftests.
> > 
> > That is totally fine. This test is what I have been using to test the
> > changes, and I decide to share it in case someone find it useful.
> > 
> > Let's drop it.
> 
> Out of interest, to we know why the test is failing for Lorenzo?

I really don't know, but, it sounds like ktap was not found?

Then the first early-exit path hit:
ktap_skip_all "..."   # undefined → returns 127 exit "$KSFT_SKIP"
# expands to: exit "" → exits with last $? = 127

> I agree that the test is a bit excessive, in particular when it comes to
> invalid/idempotent values etc. I could see some value for testing
> whether setting the modes keeps working, but also then I wonder if that
> is really something we'll be changing frequently (and that breaks easily).

yea, I make it very excessive, because there were some intrinsics in
those sysfs that I was gettingit wrong when doing the intial conversion.

So, the test is something that I trust now, and I found it useful when
finding regressiosn.

Is is something that will chagne frequently? probably not!

That said, would you like to have a simplified/different version of this
test?