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

[Muchun Song]

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?
> 
> 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));

Hi Oscar,

After more thinkging, I think here should be:

  struct mem_section *ms = __pfn_to_section(ALIGN_DOWN(page_to_pfn(head), memory_block_size_bytes()));

Why?

[                  hotplugged memory                  ]
[        section        ][...][        section        ]
[ vmemmap ][              usable memory               ]
  ^   |      |                                      |
  +---+      |                                      |
    ^        |                                      |
    +--------+                                      |
        ^                                           |
        +-------------------------------------------+

The page_to_pfn(head) can falls onto the non-1st section, actually, we desire 
1st section which ->section_mem_map is the start vmemmap of the vmemmap.
If we align the page_to_pfn(head) with the start pfn of the hotplugged memory,
then we can simplify the code further.

  unsigned long size = memory_block_size_bytes();
  unsigned long pfn = ALIGN_DOWN(page_to_pfn(head), size);

  if (pfn_valid(pfn) && PageVmemmapSelfHosted(pfn_to_page(pfn)))
          return;

Hotplugged memory block never has non-present sections, while boot memory block
can have one or more. So pfn_valid() is used to filter out the first section if
it is non-present.

Hopefully I am not wrong.

Thanks.

> +	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;
>  
>  	vmemmap_pages = hugetlb_optimize_vmemmap_pages(h);
>  	if (!vmemmap_pages)
> @@ -199,10 +207,10 @@ static struct ctl_table hugetlb_vmemmap_sysctls[] = {
>  static __init int hugetlb_vmemmap_sysctls_init(void)
>  {
>  	/*
> -	 * If "memory_hotplug.memmap_on_memory" is enabled or "struct page"
> -	 * crosses page boundaries, the vmemmap pages cannot be optimized.
> +	 * If "struct page" crosses page boundaries, the vmemmap pages cannot
> +	 * be optimized.
>  	 */
> -	if (!mhp_memmap_on_memory() && is_power_of_2(sizeof(struct page)))
> +	if (is_power_of_2(sizeof(struct page)))
>  		register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
>  
>  	return 0;
> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index 1213d0c67a53..863966c2c6f1 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -45,8 +45,6 @@
>  #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
>  static int memmap_on_memory_set(const char *val, const struct kernel_param *kp)
>  {
> -	if (hugetlb_optimize_vmemmap_enabled())
> -		return 0;
>  	return param_set_bool(val, kp);
>  }
>  
> @@ -1032,6 +1030,7 @@ int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
>  {
>  	unsigned long end_pfn = pfn + nr_pages;
>  	int ret;
> +	int i;
>  
>  	ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
>  	if (ret)
> @@ -1039,6 +1038,12 @@ int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
>  
>  	move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
>  
> +	/*
> +	 * Let us flag self-hosted memmap
> +	 */
> +	for (i = 0; i < nr_pages; i++)
> +		SetPageVmemmap_self_hosted(pfn_to_page(pfn + i));
> +
>  	/*
>  	 * It might be that the vmemmap_pages fully span sections. If that is
>  	 * the case, mark those sections online here as otherwise they will be
> 
> 
> -- 
> Oscar Salvador
> SUSE Labs
>