Re: [PATCH] mm: madvise: make MADV_NOHUGEPAGE a no-op if !THP

[Yang Shi <yang@os.amperecomputing.com>]

On 5/14/25 8:29 AM, Lorenzo Stoakes wrote:
> On Wed, May 14, 2025 at 08:25:05AM -0700, Yang Shi wrote:
>>
>> On 5/14/25 8:07 AM, Ignacio Moreno Gonzalez wrote:
>>> On 5/14/2025 3:52 PM, Lorenzo Stoakes wrote:
>>>
>>>> I have pinged s390 people on there, but I don't think this is going to make this
>>>> cycle given we will probably need to coordinate with them to fix up this enum
>>>> name (which seems the best solution to me!...)
>>> They answered that it would be ok for them to do a quick fix over the mm tree:
>>>
>>> https://lore.kernel.org/linux-mm/6f8f3780-902b-49d4-a766-ea2e1a8f85ea@linux.ibm.com/
>>>
>>>
>>>>> +#include <uapi/asm/mman.h>
>>>> Also this should be linux/mman.h I believe, sorry for not catching first time round...!
>>>>
>>> Including linux/mman.h leads to a compilation error:
>>>
>>> ../include/linux/huge_mm.h:601:23: error: ‘MADV_NOHUGEPAGE’
>>>
>>> Including uapi/linux/mman.h works, but I am not sure if that is correct.
>> Is this build on x86? I actually tried this on my arm64 build before I
>> suggested uapi/asm/mman.h. But I saw a lot of compilation errors, like:
>>
>> In file included from ./include/linux/memcontrol.h:22,
>>                   from ./include/linux/swap.h:9,
>>                   from ./include/linux/userfaultfd_k.h:18,
>>                   from ./include/linux/hugetlb.h:16,
>>                   from ./arch/arm64/include/asm/mman.h:10,
>>                   from ./include/uapi/linux/mman.h:5,
>>                   from ./include/linux/huge_mm.h:512,
>>                   from ./include/linux/mm.h:1238,
>>                   from ./include/linux/memblock.h:12,
>>                   from ./arch/arm64/include/asm/acpi.h:14,
>>                   from ./include/acpi/acpi_io.h:7,
>>                   from ./include/linux/acpi.h:39,
>>                   from ./include/acpi/apei.h:9,
>>                   from ./include/acpi/ghes.h:5,
>>                   from ./include/linux/arm_sdei.h:8,
>>                   from arch/arm64/kernel/asm-offsets.c:10:
>> ./include/linux/vmstat.h: In function ‘__zone_stat_mod_folio’:
>> ./include/linux/vmstat.h:414:31: error: implicit declaration of function
>> ‘folio_zone’; did you mean ‘folio_zonenum’?
>> [-Wimplicit-function-declaration]
>>    414 |         __mod_zone_page_state(folio_zone(folio), item, nr);
>>        |                               ^~~~~~~~~~
>>        |                               folio_zonenum
>> ./include/linux/vmstat.h:414:31: error: passing argument 1 of
>> ‘__mod_zone_page_state’ makes pointer from integer without a cast
>> [-Wint-conversion]
>>    414 |         __mod_zone_page_state(folio_zone(folio), item, nr);
>>        |                               ^~~~~~~~~~~~~~~~~
>>        |                               |
>>        |                               int
>> ./include/linux/vmstat.h:273:28: note: expected ‘struct zone *’ but argument
>> is of type ‘int’
>>    273 | void __mod_zone_page_state(struct zone *, enum zone_stat_item item,
>> long);
>>        |                            ^~~~~~~~~~~~~
>> ./include/linux/vmstat.h: In function ‘__zone_stat_add_folio’:
>> ./include/linux/vmstat.h:420:56: error: implicit declaration of function
>> ‘folio_nr_pages’; did you mean ‘folio_page’?
>> [-Wimplicit-function-declaration]
>>    420 |         __mod_zone_page_state(folio_zone(folio), item,
>> folio_nr_pages(folio));
>>        | ^~~~~~~~~~~~~~
>>
>>
>> The build used default Fedora kernel config with THP disabled and v6.15-rc6
>> kernel.
> Yeah I suspect this is because of circular dependencies as I mentioned in reply
> to Ignacio :) Sorry I missed you'd suggested it Yang, you were right to do so :)

No problem :)

>
> You can see mm.h -> huge_mm.h there so it does seem to be same thing.

Yeah, it caused some circular dependencies.

>
> C headers are a source of many sighs.
>
>>>