Re: [PATCH] mm/hugetlb: change ENOSPC to ENOMEM in alloc_hugetlb_folio

[Dafna Hirschfeld <dafna.hirschfeld@intel.com>]

On 02.12.2024 11:49, Muchun Song wrote:
>
>
>> On Dec 1, 2024, at 09:03, Dafna Hirschfeld <dafna.hirschfeld@intel.com> wrote:
>>
>> The error ENOSPC is translated in vmf_error to VM_FAULT_SIGBUS which is
>> further translated in EFAULT in i.e. pin/get_user_pages.
>> But when running out of pages/hugepages we expect to see ENOMEM and
>> not EFAULT.
>
>Hi Dafna,
>
>Refers to Documentation/mm/hugetlbfs_reserv.rst. I saw:
>
>    If no huge page exists at page fault time, the task is sent
>    a **SIGBUS** and often dies an unhappy death.
>
>Seems SIGBUS is expected since it is introduced.

Hi,
I see that mlock do return ENOMEM when out of hugepages.
It has the function "__mlock_posix_error_return" that converts EFAULT returned from
"get_user_pages" to ENOMEM.
So callers to "get_user_pages" can override EFAULT with ENOMEM but then it obscure
cases where EFUALT should really be the return code.

thanks,
Dafna

>
>Thanks.