Re: PROBLEM: Remapping hugepages mappings causes kernel to return EINVAL

[Mike Kravetz]

On 10/23/2017 10:52 AM, C.Wehrmeyer wrote:
> On 2017-10-23 18:57, Michal Hocko wrote:
>> On Mon 23-10-17 18:46:59, C.Wehrmeyer wrote:
>>> On 23-10-17 18:13, Michal Hocko wrote:
>>>> On Mon 23-10-17 16:00:13, C.Wehrmeyer wrote:
>>>>> And just to be very sure I've added:
>>>>>
>>>>> if (madvise(buf1,ALLOC_SIZE_1,MADV_HUGEPAGE)) {
>>>>>           errno_tmp = errno;
>>>>>           fprintf(stderr,"madvise: %u\n",errno_tmp);
>>>>>           goto out;
>>>>> }
>>>>>
>>>>> /*Make sure the mapping is actually used*/
>>>>> memset(buf1,'!',ALLOC_SIZE_1);
>>>>
>>>> Is the buffer aligned to 2MB?
>>>
>>> When I omit MAP_HUGETLB for the flags that mmap receives - no.
>>>
>>> #define ALLOC_SIZE_1 (2 * 1024 * 1024)
>>> [...]
>>> buf1 = mmap (
>>>          NULL,
>>>          ALLOC_SIZE_1,
>>>          prot, /*PROT_READ | PROT_WRITE*/
>>>          flags /*MAP_PRIVATE | MAP_ANONYMOUS*/,
>>>          -1,
>>>          0
>>> );
>>>
>>> In such a case buf1 usually contains addresses which are aligned to 4 KiBs,
>>> such as 0x7f07d76e9000. 2-MiB-aligned addresses, such as 0x7f89f5e00000, are
>>> only produced with MAP_HUGETLB - which, if I understood the documentation
>>> correctly, is not the point of THPs as they are supposed to be transparent.
>>
>> yes. You can use posix_memalign
> 
> Useless. We don't use the memory allocation structures of malloc/free, and yet that's exactly what this function requires us to do. The reason why we use mmap and mremap is to get rid of userspace-crap in the first place.
> 
>> or you can mmap a larger block and
>> munmap the initial unaligned part.
> 
> And how is that supposed to be transparent? When I hear "transparent" I think of a mechanism which I can put under a system so that it benefits from it, while the system does not notice or at least does not need to be aware of it. The system also does not need to be changed for it.
> 
> This approach is even more un-transparent than providing a flag to mmap in order to make hugepages work correctly.

Well at least this has a built in fall back mechanism.  When using hugetlb(fs)
pages, you would need to handle the case where mremap fails due to lack of
configured huge pages.

I assume your allocator will be for somewhat general application usage.  Yet,
for the most reliability the user/admin will need to know at boot time how
many huge pages will be needed and set that up.

-- 
Mike Kravetz