Re: [PATCH 1/2] hugetlb: fix update_and_free_page contig page struct assumption

[Zi Yan <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 3016 bytes --]

On 18 Feb 2021, at 12:51, Mike Kravetz wrote:

> On 2/18/21 9:40 AM, Zi Yan wrote:
>> On 18 Feb 2021, at 12:32, Jason Gunthorpe wrote:
>>
>>> On Thu, Feb 18, 2021 at 12:27:58PM -0500, Zi Yan wrote:
>>>> On 18 Feb 2021, at 12:25, Jason Gunthorpe wrote:
>>>>
>>>>> On Thu, Feb 18, 2021 at 02:45:54PM +0000, Matthew Wilcox wrote:
>>>>>> On Wed, Feb 17, 2021 at 11:02:52AM -0800, Andrew Morton wrote:
>>>>>>> On Wed, 17 Feb 2021 10:49:25 -0800 Mike Kravetz <mike.kravetz@oracle.com> wrote:
>>>>>>>> page structs are not guaranteed to be contiguous for gigantic pages.  The
>>>>>>>
>>>>>>> June 2014.  That's a long lurk time for a bug.  I wonder if some later
>>>>>>> commit revealed it.
>>>>>>
>>>>>> I would suggest that gigantic pages have not seen much use.  Certainly
>>>>>> performance with Intel CPUs on benchmarks that I've been involved with
>>>>>> showed lower performance with 1GB pages than with 2MB pages until quite
>>>>>> recently.
>>>>>
>>>>> I suggested in another thread that maybe it is time to consider
>>>>> dropping this "feature"
>>>>
>>>> You mean dropping gigantic page support in hugetlb?
>>>
>>> No, I mean dropping support for arches that want to do:
>>>
>>>    tail_page != head_page + tail_page_nr
>>>
>>> because they can't allocate the required page array either virtually
>>> or physically contiguously.
>>>
>>> It seems like quite a burden on the core mm for a very niche, and
>>> maybe even non-existant, case.
>>>
>>> It was originally done for PPC, can these PPC systems use VMEMMAP now?
>>>
>>>>> The cost to fix GUP to be compatible with this will hurt normal
>>>>> GUP performance - and again, that nobody has hit this bug in GUP
>>>>> further suggests the feature isn't used..
>>>>
>>>> A easy fix might be to make gigantic hugetlb page depends on
>>>> CONFIG_SPARSEMEM_VMEMMAP, which guarantee all struct pages are contiguous.
>>>
>>> Yes, exactly.
>>
>> I actually have a question on CONFIG_SPARSEMEM_VMEMMAP. Can we assume
>> PFN_A - PFN_B == struct_page_A - struct_page_B, meaning all struct pages
>> are ordered based on physical addresses? I just wonder for two PFN ranges,
>> e.g., [0 - 128MB], [128MB - 256MB], if it is possible to first online
>> [128MB - 256MB] then [0 - 128MB] and the struct pages of [128MB - 256MB]
>> are in front of [0 - 128MB] in the vmemmap due to online ordering.
>
> I have not looked at the code which does the onlining and vmemmap setup.
> But, these definitions make me believe it is true:
>
> #elif defined(CONFIG_SPARSEMEM_VMEMMAP)
>
> /* memmap is virtually contiguous.  */
> #define __pfn_to_page(pfn)      (vmemmap + (pfn))
> #define __page_to_pfn(page)     (unsigned long)((page) - vmemmap)

Makes sense. Thank you for checking.

I guess making gigantic page depends on CONFIG_SPARSEMEM_VMEMMAP might
be a good way of simplifying code and avoiding future bugs unless
there is an arch really needs gigantic page and cannot have VMEMMAP.

—
Best Regards,
Yan Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]