Re: [PATCH 0/3] selftests/mm: add memory failure selftests

[Miaohe Lin <linmiaohe@huawei.com>]

On 2026/1/12 19:33, Miaohe Lin wrote:
> On 2026/1/12 17:40, David Hildenbrand (Red Hat) wrote:
>> On 1/12/26 10:19, Miaohe Lin wrote:
>>> On 2026/1/9 21:45, David Hildenbrand (Red Hat) wrote:
>>>> On 1/7/26 10:37, Miaohe Lin wrote:
>>>>> Introduce selftests to validate the functionality of memory failure.
>>>>> These tests help ensure that memory failure handling for anonymous
>>>>> pages, pagecaches pages works correctly, including proper SIGBUS
>>>>> delivery to user processes, page isolation, and recovery paths.
>>>>>
>>>>> Currently madvise syscall is used to inject memory failures. And only
>>>>> anonymous pages and pagecaches are tested. More test scenarios, e.g.
>>>>> hugetlb, shmem, thp, will be added. Also more memory failure injecting
>>>>> methods will be supported, e.g. APEI Error INJection, if required.
>>>>
>>>
>>> Thanks for test and report. :)
>>>
>>>> 0day reports that these tests fail:
>>>>
>>>> # # ------------------------
>>>> # # running ./memory-failure
>>>> # # ------------------------
>>>> # # TAP version 13
>>>> # # 1..6
>>>> # # # Starting 6 tests from 2 test cases.
>>>> # # #  RUN           memory_failure.madv_hard.anon ...
>>>> # # #            OK  memory_failure.madv_hard.anon
>>>> # # ok 1 memory_failure.madv_hard.anon
>>>> # # #  RUN           memory_failure.madv_hard.clean_pagecache ...
>>>> # # # memory-failure.c:166:clean_pagecache:Expected setjmp (1) == 0 (0)
>>>> # # # clean_pagecache: Test terminated by assertion
>>>> # # #          FAIL  memory_failure.madv_hard.clean_pagecache
>>>> # # not ok 2 memory_failure.madv_hard.clean_pagecache
>>>> # # #  RUN           memory_failure.madv_hard.dirty_pagecache ...
>>>> # # # memory-failure.c:207:dirty_pagecache:Expected unpoison_memory(self->pfn) (-16) == 0 (0)
>>>> # # # dirty_pagecache: Test terminated by assertion
>>>> # # #          FAIL  memory_failure.madv_hard.dirty_pagecache
>>>> # # not ok 3 memory_failure.madv_hard.dirty_pagecache
>>>> # # #  RUN           memory_failure.madv_soft.anon ...
>>>> # # #            OK  memory_failure.madv_soft.anon
>>>> # # ok 4 memory_failure.madv_soft.anon
>>>> # # #  RUN           memory_failure.madv_soft.clean_pagecache ...
>>>> # # # memory-failure.c:282:clean_pagecache:Expected variant->inject(self, addr) (-1) == 0 (0)
>>>> # # # clean_pagecache: Test terminated by assertion
>>>> # # #          FAIL  memory_failure.madv_soft.clean_pagecache
>>>> # # not ok 5 memory_failure.madv_soft.clean_pagecache
>>>> # # #  RUN           memory_failure.madv_soft.dirty_pagecache ...
>>>> # # # memory-failure.c:319:dirty_pagecache:Expected variant->inject(self, addr) (-1) == 0 (0)
>>>> # # # dirty_pagecache: Test terminated by assertion
>>>> # # #          FAIL  memory_failure.madv_soft.dirty_pagecache
>>>> # # not ok 6 memory_failure.madv_soft.dirty_pagecache
>>>> # # # FAILED: 2 / 6 tests passed.
>>>> # # # Totals: pass:2 fail:4 xfail:0 xpass:0 skip:0 error:0
>>>> # # [FAIL]
>>>> # not ok 71 memory-failure # exit=1
>>>>
>>>>
>>>> Can the test maybe not deal with running in certain environments (config options etc)?
>>>
>>> To run the test, I think there should be:
>>>    1.CONFIG_MEMORY_FAILURE and CONFIG_HWPOISON_INJECT should be enabled.
>>>    2.Root privilege is required.
>>>    3.For dirty/clean pagecache testcases, the test file "./clean-page-cache-test-file" and
>>>      "./dirty-page-cache-test-file" are assumed to be created on non-memory file systems
>>>      such as xfs, ext4, etc.
>>>
>>> Does your test environment break any of the above rules?
>>
>> It is 0day environment, so very likely yes. I suspect 1).

Hi David,

After taking a more close look, I think CONFIG_MEMORY_FAILURE and CONFIG_HWPOISON_INJECT should have been
enabled in 0day environment or testcase memory_failure.madv_hard.anon should fail. memory_failure.madv_hard.anon
will inject memory failure and expects seeing a SIGBUG signal.

>>
>>> Am I expected to add some code to
>>> guard against this?
>>
>> Yes, at least some.
>>
>> Checking for root privileges is not required. The tests are commonly run from non-memory file systems, but, in theory, could be run from nfs etc.
>>
>> If you require special file systems, take a look at gup_longterm.o where we test for some fileystsem types.

And I think the cause of failures of testcases memory_failure.madv_hard.clean_pagecache and memory_failure.madv_hard.dirty_pagecache
is they running on memory filesystems. The error pages are kept in page cache in that case while memory_failure.madv_hard.clean_pagecache
expects to see the error page truncated.

But I have no idea why memory_failure.madv_soft.dirty_pagecache and memory_failure.madv_soft.clean_pagecache return -1(-EPERM?) when try
to inject memory error through madvise syscall. It could be really helpful if more information can be provided.

Thanks!
.