Re: [PATCH] selftests/mm: unpoison pages in memory-failure teardown

["David Hildenbrand (Arm)" <david@kernel.org>]

On 7/28/26 16:22, Muhammad Usama Anjum wrote:
> On 27/07/2026 1:31 pm, David Hildenbrand (Arm) wrote:
>> On 7/27/26 11:54, Muhammad Usama Anjum wrote:
>>> The memory-failure tests call cleanup() only after all result checks.
>>> A failed ASSERT_* invokes fixture teardown and aborts the test, so it
>>> skips cleanup() and leaves the injected page hardware-poisoned.
>>>
>>> Invoke cleanup() from FIXTURE_TEARDOWN() instead. Guard it with
>>> self->triggered so tests that exit before injection do not try to
>>> unpoison a page when no injection was attempted. This runs the existing
>>> HWPoison and HardwareCorrupted checks on both normal and assertion-failure
>>> paths.
>>>
>>> Fixes: ff4ef2fbd101 ("selftests/mm: add memory failure anonymous page test")
>>> Signed-off-by: Muhammad Usama Anjum <usama.anjum@arm.com>
>>> ---
>>>  tools/testing/selftests/mm/memory-failure.c | 26 +++++++++------------
>>>  1 file changed, 11 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/tools/testing/selftests/mm/memory-failure.c b/tools/testing/selftests/mm/memory-failure.c
>>> index 5d00aab31f9b5..eaa8b9bd401a1 100644
>>> --- a/tools/testing/selftests/mm/memory-failure.c
>>> +++ b/tools/testing/selftests/mm/memory-failure.c
>>> @@ -122,13 +122,6 @@ static void teardown_sighandler(void)
>>>  	sigaction(SIGBUS, &sa, NULL);
>>>  }
>>>  
>>> -FIXTURE_TEARDOWN(memory_failure)
>>> -{
>>> -	close(self->kpageflags_fd);
>>> -	close(self->pagemap_fd);
>>> -	teardown_sighandler();
>>> -}
>>> -
>>>  static void prepare(struct __test_metadata *_metadata, FIXTURE_DATA(memory_failure) * self,
>>>  		    void *vaddr)
>>>  {
>>> @@ -200,8 +193,7 @@ static void check(struct __test_metadata *_metadata, FIXTURE_DATA(memory_failure
>>>  	ASSERT_EQ(pfn_flags & KPF_HWPOISON, KPF_HWPOISON);
>>>  }
>>>  
>>> -static void cleanup(struct __test_metadata *_metadata, FIXTURE_DATA(memory_failure) * self,
>>> -		    void *vaddr)
>>> +static void cleanup(struct __test_metadata *_metadata, FIXTURE_DATA(memory_failure) * self)
>>>  {
>>>  	unsigned long size;
>>>  	uint64_t pfn_flags;
>>> @@ -217,6 +209,16 @@ static void cleanup(struct __test_metadata *_metadata, FIXTURE_DATA(memory_failu
>>>  	ASSERT_EQ(size, self->corrupted_size);
>>>  }
>>>  
>>> +FIXTURE_TEARDOWN(memory_failure)
>>> +{
>>> +	if (self->triggered)
>>> +		cleanup(_metadata, self);
>>
>> If the variant->inject(self, addr) fails, self->triggered would already have
>> been set.
>>
>> Wouldn't it be cleaner to have a new self->poisoned that we set only after
>> ->inject succeeded?
> Setting self->poisoned after variant->inject() returns would not cover the
> MADV_HARD signal path.
> 
> The call path of madvise(MADV_HWPOISON) is:
> 
> madvise() -> madvise_do_behavior() -> madvise_inject_error(MF_ACTION_REQUIRED) ->
> memory_failure() -> hwpoison_user_mappings() -> kill_procs(BUS_MCEERR_AR) ->
> kill_proc() -> force_sig_mceerr()
> 
> SIGBUS is therefore queued while the madvise system call is executing and
> delivered on return to userspace, before execution can continue after
> variant->inject().
> 
> The test's sigbus_action() then calls siglongjmp(), so an assignment placed
> after variant->inject() would be bypassed. The flag would remain false and
> the test would attempt the injection again after returning to sigsetjmp().
> 
> Also, memory_failure() sets PG_hwpoison before several recovery paths that
> can return an error, so a failed injection return does not prove that the
> page was not poisoned.
> 
> Maybe renaming triggered to injection_attempted is more precise here.

Yes, that would make it clearer. + adding a comment that we want to cleanup even
when injection was attempted, but failed.

Thanks!

-- 
Cheers,

David