Re: [PATCH v4 7/7] selftests: prctl: introduce tests for disabling THPs except for madvise

[Usama Arif]

On 14/08/2025 16:02, Lorenzo Stoakes wrote:
> On Thu, Aug 14, 2025 at 02:08:57PM +0100, Mark Brown wrote:
>> On Thu, Aug 14, 2025 at 02:59:13PM +0200, David Hildenbrand wrote:
>>> On 14.08.25 14:09, Mark Brown wrote:
>>
>>>> Perhaps this is something that needs considering in the ABI, so
>>>> userspace can reasonably figure out if it failed to configure whatever
>>>> is being configured due to a missing feature (in which case it should
>>>> fall back to not using that feature somehow) or due to it messing
>>>> something else up?  We might be happy with the tests being version
>>>> specific but general userspace should be able to be a bit more robust.
>>
>>> Yeah, the whole prctl() ship has sailed, unfortunately :(
>>
>> Perhaps a second call or sysfs file or something that returns the
>> supported mask?  You'd still have a boostrapping issue with existing
>> versions but at least at any newer stuff would be helped.
> 
> Ack yeah I do wish we had better APIs for expressing what was
> available/not. Will put this sort of thing on the TODO...
> 
> Overall I don't want to hold this up unnecesarily, and I bow to the
> consensus if others feel we ought not to _assume_ same kernel at least best
> effort.
> 
> Usama - It's ok to leave it as is in this case since obviously only tip
> kernel will have this feature.

ah ok, so will keep it at skipping if prctl doesnt work in fixture as is
in the current v4 version.

I only have the below diff and its equivalent for patch 7 as a difference over
this version. Will wait until tomorrow morning incase there are more comments
and hopefully send out a last revision!

Thanks!



diff --git a/tools/testing/selftests/mm/prctl_thp_disable.c b/tools/testing/selftests/mm/prctl_thp_disable.c
index 8845e9f414560..e9e519c85224c 100644
--- a/tools/testing/selftests/mm/prctl_thp_disable.c
+++ b/tools/testing/selftests/mm/prctl_thp_disable.c
@@ -18,6 +18,7 @@
 
 enum thp_collapse_type {
        THP_COLLAPSE_NONE,
+       THP_COLLAPSE_MADV_NOHUGEPAGE,
        THP_COLLAPSE_MADV_HUGEPAGE,     /* MADV_HUGEPAGE before access */
        THP_COLLAPSE_MADV_COLLAPSE,     /* MADV_COLLAPSE after access */
 };
@@ -49,6 +50,8 @@ static int test_mmap_thp(enum thp_collapse_type madvise_buf, size_t pmdsize)
 
        if (madvise_buf == THP_COLLAPSE_MADV_HUGEPAGE)
                madvise(mem, pmdsize, MADV_HUGEPAGE);
+       else if (madvise_buf == THP_COLLAPSE_MADV_NOHUGEPAGE)
+               madvise(mem, pmdsize, MADV_NOHUGEPAGE);
 
        /* Ensure memory is allocated */
        memset(mem, 1, pmdsize);
@@ -73,6 +76,8 @@ static void prctl_thp_disable_completely_test(struct __test_metadata *const _met
        /* tests after prctl overrides global policy */
        ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_NONE, pmdsize), 0);
 
+       ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_MADV_NOHUGEPAGE, pmdsize), 0);
+
        ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_MADV_HUGEPAGE, pmdsize), 0);
 
        ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_MADV_COLLAPSE, pmdsize), 0);
@@ -84,6 +89,8 @@ static void prctl_thp_disable_completely_test(struct __test_metadata *const _met
        ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_NONE, pmdsize),
                  thp_policy == THP_ALWAYS ? 1 : 0);
 
+       ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_MADV_NOHUGEPAGE, pmdsize), 0);
+
        ASSERT_EQ(test_mmap_thp(THP_COLLAPSE_MADV_HUGEPAGE, pmdsize),
                  thp_policy == THP_NEVER ? 0 : 1);

> 
> Cheers, Lorenzo