Re: [PATCH v4 18/49] mlock: Convert mlock to vma iterator

[Ryan Roberts]

On 11/07/2023 16:27, Liam R. Howlett wrote:
> * Ryan Roberts <ryan.roberts@xxxxxxx> [230711 10:09]:
>> On 20/01/2023 16:26, Liam R. Howlett wrote:
>>> From: "Liam R. Howlett" <Liam.Howlett@xxxxxxxxxx>
>>>
>>> Use the vma iterator so that the iterator can be invalidated or updated
>>> to avoid each caller doing so.
>>
>> Hi,
> 
> 
> Hello!
> 
>>
>> I've bisected 2 mm selftest regressions back to this patch, so hoping someone can help debug and fix? The failures are reproducible on x86_64 and arm64.
> 
> Thanks!  That is a big help.  Where did you start your bisection?  I
> assume 6.4?

Yes, I'm working to get all the mm selftests running (and ideally passing!) on
arm64. I working on v6.4 and it was broken there. I went arbitrarily back to
v5.10 and it was working there, so bisected between them.

> 
>>
>>
>> mlock-random-test:
>>
>> $ ./run_kselftest.sh -t mm:mlock-random-test
>> TAP version 13
>> 1..1
>> # selftests: mm: mlock-random-test
>> mlock() failure at |0xaaaaaaab52d0(131072)| mlock:|0xaaaaaaacc65d(26551)|
>> not ok 1 selftests: mm: mlock-random-test # exit=255
>>
>> This mallocs a buffer then loops 100 times, trying to mlock random parts of it. After this patch, the test fails after a variable number of iterations; mlock() returns ENOMEM. If I explicitly munlock at the end of each loop, it works.
>>
>>
>> mlock2-tests:
>>
>> $ ./run_kselftest.sh -t mm:mlock2-tests
>> TAP version 13
>> 1..1
>> # selftests: mm: mlock2-tests
>> munlock(): Cannot allocate memory
>> munlock(): Cannot allocate memory
>> not ok 1 selftests: mm: mlock2-tests # exit=2
>>
>> Here, a 3 page buffer is mlock2()ed, then the middle page is munlocked. Finally the whole 3 page range is munlocked, and after this patch it fails with ENOMEM. If I modify the test to split the final munlock into 2, one for the first page and one for the last, the test passes.
>>
>>
>> Immediately prior to this patch (2286a6914c77 "mm: change mprotect_fixup to vma iterator"), both tests pass.
>>
>> From a quick scan of the man page, I don't think it explicitly says that its ok to call mlock/munlock on already locked/unlocked pages, but it's certainly a change of behavior and the tests notice, so I'm guessing this wasn't intentional?
>>
>> I'm not familiar with this code so it's not obvious to me exactly what the problem is, but I'm hoping someone can help debug?
> 
> I think I see the issue and I'm working on a fix. I appreciate the
> analysis and report, it really helps narrow things down.

You're welcome!

> 
> Regards,
> Liam