Re: [PATCH v3 5/5] powerpc/mm: Fix growth direction for hugepages mmaps with slice

[Christophe LEROY <christophe.leroy@c-s.fr>]

Le 24/01/2018 à 11:08, Aneesh Kumar K.V a écrit :
> 
> 
> On 01/24/2018 03:33 PM, Christophe LEROY wrote:
>>
>>
>> Le 24/01/2018 à 10:51, Aneesh Kumar K.V a écrit :
>>>
>>>
>>> On 01/24/2018 03:09 PM, Christophe LEROY wrote:
>>>>
>>>>
>>>> Le 24/01/2018 à 10:35, Aneesh Kumar K.V a écrit :
>>>>>
>>>
>>>>>> Did you try with HUGETLB_MORECORE_HEAPBASE=0x11000000 on PPC64 as 
>>>>>> I suggested in my last email on this subject (22/01/2018 9:22) ?
>>>>>
>>>>>
>>>>> yes. The test ran fine for me
>>>>
>>>> You tried with 0x30000000, it works as well on PPC32.
>>>>
>>>> I'd really like you to try with 0x11000000 which is in the same 
>>>> slice as the 10020000-10030000 range.
>>>>
>>>>
>>>
>>> Now that explains is better. But then the requested HEAPBASE was not 
>>> free and hence topdown search got an address in the below range.
>>>
>>> 7efffd000000-7f0000000000 rw-p 00000000 00:0d 1082770 /anon_hugepage 
>>> (deleted)
>>>
>>>
>>> The new range allocated is such that there is no scope for expansion 
>>> of heap if we do a topdown search. But why should that require us to 
>>> change from topdown/bottomup search?
>>>
>>>
>>> 10000000-10010000 r-xp 00000000 fc:00 9044312 /home/kvaneesh/a.out
>>> 10010000-10020000 r--p 00000000 fc:00 9044312 /home/kvaneesh/a.out
>>> 10020000-10030000 rw-p 00010000 fc:00 9044312 /home/kvaneesh/a.out
>>> 7efffd000000-7f0000000000 rw-p 00000000 00:0d 1082770 /anon_hugepage 
>>> (deleted)
>>> 7ffff2d40000-7ffff7d60000 rw-p 00000000 00:00 0
>>> 7ffff7d60000-7ffff7f10000 r-xp 00000000 fc:00 9250090 
>>> /lib/powerpc64le-linux-gnu/libc-2.23.so
>>> 7ffff7f10000-7ffff7f20000 r--p 001a0000 fc:00 9250090 
>>> /lib/powerpc64le-linux-gnu/libc-2.23.so
>>> 7ffff7f20000-7ffff7f30000 rw-p 001b0000 fc:00 9250090 
>>> /lib/powerpc64le-linux-gnu/libc-2.23.so
>>> 7ffff7f40000-7ffff7f60000 r-xp 00000000 fc:00 10754812 
>>> /usr/lib/libhugetlbfs.so.0
>>> 7ffff7f60000-7ffff7f70000 r--p 00010000 fc:00 10754812 
>>> /usr/lib/libhugetlbfs.so.0
>>> 7ffff7f70000-7ffff7f80000 rw-p 00020000 fc:00 10754812 
>>> /usr/lib/libhugetlbfs.so.0
>>> 7ffff7f80000-7ffff7fa0000 r-xp 00000000 00:00 0 [vdso]
>>> 7ffff7fa0000-7ffff7fe0000 r-xp 00000000 fc:00 9250107 
>>> /lib/powerpc64le-linux-gnu/ld-2.23.so
>>> 7ffff7fe0000-7ffff7ff0000 r--p 00030000 fc:00 9250107 
>>> /lib/powerpc64le-linux-gnu/ld-2.23.so
>>> 7ffff7ff0000-7ffff8000000 rw-p 00040000 fc:00 9250107 
>>> /lib/powerpc64le-linux-gnu/ld-2.23.so
>>> 7ffffffd0000-800000000000 rw-p 00000000 00:00 0 [stack]
>>>
>>>
>>> For the specific test, one should pass the HEAPBASE value such that 
>>> it can be expanded if required isn't it ?
>>
>> For the test, yes, it is dumb to pass an unusable HEAPBASE, but what 
>> happens in real life:
>> * PPC32: No HEAPBASE, hugetlbfs defines a HEAPBASE at sbrk(0) + 
>> PAGE_SIZE = 0x10800000 ==> This is in the same slice as already 
>> allocated ==> the kernel does as if mmap() had been called with no 
>> hint address and allocates something unusable instead.
>> * PPC64: No HEAPBASE, hugetlbfs seems to define a HEAPBASE at 
>> 100000000000, which doesn't conflict with an already allocated mapping 
>> ==> it works.
>>
>> Now, when we take the generic case, ie when slice is not activated, 
>> when you call mmap() without a hint address, it allocates a suitable 
>> address because it does bottom-up. Why do differently with slices ?
>>
> 
> IIUC that is largely arch dependent, PPC64 always did topdown search. 
> Even for regular non hugetlb mmap it did topdown search. If you set 
> legacy mmap we selected bottom up approach. You can check 
> arch_pick_mmap_layout() for more details. Now x86 is slightly different.
> For the default search if we can't find a mapping address it will try a 
> bottomup search. Having said that if you think libhugetlbfs made 
> assumptions with respect to 8xx and you don't want to break it make
> 8xx unmapped area search bottomup.
> 

Or would there be a way to make libhugetlbfs aware of the slices 
constraints and make it choose a suitable hint address at first try ?

Christophe