Re: [PATCH v2 2/2] mm: memory_hotplug: introduce SECTION_CANNOT_OPTIMIZE_VMEMMAP

From: David Hildenbrand
Date: Fri Jun 17 2022 - 05:25:35 EST


On 17.06.22 11:10, Muchun Song wrote:
> On Fri, Jun 17, 2022 at 09:39:27AM +0200, David Hildenbrand wrote:
>> On 17.06.22 09:28, Muchun Song wrote:
>>> On Fri, Jun 17, 2022 at 07:46:53AM +0200, Oscar Salvador wrote:
>>>> On Thu, Jun 16, 2022 at 09:30:33AM +0200, David Hildenbrand wrote:
>>>>> IIRC, that was used to skip these patches on the offlining path before
>>>>> we provided the ranges to offline_pages().
>>>>
>>>> Yeah, it was designed for that purpose back then.
>>>>
>>>>> I'd not mess with PG_reserved, and give them a clearer name, to not
>>>>> confuse them with other, ordinary, vmemmap pages that are not
>>>>> self-hosted (maybe in the future we might want to flag all vmemmap pages
>>>>> with a new type?).
>>>>
>>>> Not sure whether a new type is really needed, or to put it another way, I
>>>> cannot see the benefit.
>>>>
>>>>>
>>>>> I'd just try reusing the flag PG_owner_priv_1. And eventually, flag all
>>>>> (v)memmap pages with a type PG_memmap. However, the latter would be
>>>>> optional and might not be strictly required
>>>>>
>>>>>
>>>>> So what think could make sense is
>>>>>
>>>>> /* vmemmap pages that are self-hosted and cannot be optimized/freed. */
>>>>> PG_vmemmap_self_hosted = PG_owner_priv_1,
>>>>
>>>> Sure, I just lightly tested the below, and seems to work, but not sure
>>>> whether that is what you are referring to.
>>>> @Munchun: thoughts?
>>>>
>>>
>>> I think it works and fits my requirement.
>>>
>>>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>>>> index e66f7aa3191d..a4556afd7bda 100644
>>>> --- a/include/linux/page-flags.h
>>>> +++ b/include/linux/page-flags.h
>>>> @@ -193,6 +193,11 @@ enum pageflags {
>>>>
>>>> /* Only valid for buddy pages. Used to track pages that are reported */
>>>> PG_reported = PG_uptodate,
>>>> +
>>>> +#ifdef CONFIG_MEMORY_HOTPLUG
>>>> + /* For self-hosted memmap pages */
>>>> + PG_vmemmap_self_hosted = PG_owner_priv_1,
>>>> +#endif
>>>> };
>>>>
>>>> #define PAGEFLAGS_MASK ((1UL << NR_PAGEFLAGS) - 1)
>>>> @@ -628,6 +633,10 @@ PAGEFLAG_FALSE(SkipKASanPoison, skip_kasan_poison)
>>>> */
>>>> __PAGEFLAG(Reported, reported, PF_NO_COMPOUND)
>>>>
>>>> +#ifdef CONFIG_MEMORY_HOTPLUG
>>>> +PAGEFLAG(Vmemmap_self_hosted, vmemmap_self_hosted, PF_ANY)
>>>> +#endif
>>>> +
>>>> /*
>>>> * On an anonymous page mapped into a user virtual memory area,
>>>> * page->mapping points to its anon_vma, not to a struct address_space;
>>>> diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
>>>> index 1089ea8a9c98..e2de7ed27e9e 100644
>>>> --- a/mm/hugetlb_vmemmap.c
>>>> +++ b/mm/hugetlb_vmemmap.c
>>>> @@ -101,6 +101,14 @@ void hugetlb_vmemmap_free(struct hstate *h, struct page *head)
>>>> {
>>>> unsigned long vmemmap_addr = (unsigned long)head;
>>>> unsigned long vmemmap_end, vmemmap_reuse, vmemmap_pages;
>>>> + struct mem_section *ms = __pfn_to_section(page_to_pfn(head));
>>>> + struct page *memmap;
>>>> +
>>>> + memmap = sparse_decode_mem_map(ms->section_mem_map,
>>>> + pfn_to_section_nr(page_to_pfn(head)));
>>>> +
>>>> + if (PageVmemmap_self_hosted(memmap))
>>>> + return;
>>>
>>> I think here needs a loop if it is a 1GB page (spans multiple sections).
>>> Right? Here is an implementation based on another approach. But I think
>>> your implementation is more simpler and efficient. Would you mind me
>>> squash your diff into my patch and with your "Co-developed-by"?
>>
>> Due to hugtlb alignment requirements, and the vmemmap pages being at the
>> start of the hotplugged memory region, I think that cannot currently
>> happen. Checking the first vmemmap page might be good enough for now,
>> and probably for the future.
>>
>
> If the memory block size is 128MB, then a 1GB huge page spans 8 blocks.
> Is it possible that some blocks of them are vmemmap-hosted?

No, don't think so. If you think about it, a huge/gigantic page can only
start in a memmap-on-memory region but never end in on (or overlap one)
-- because the reserved memmap part of the memory block always precedes
actually usable data.

So even with variable-size memory blocks and weird address alignment,
checking the first memmap of a huge page for vmemmp-on-memory should be
sufficient.

Unless I am missing something.

--
Thanks,

David / dhildenb