Re: [PATCH v3] mm/memory hotplug/unplug: Optimize zone contiguous check when changing pfn range

["David Hildenbrand (Arm)" <david@kernel.org>]

>   With the last memblock region fits in Node 1 Zone Normal. 
> 
>   Then I punch a hole in this region with 2M(subsection) size with following
>   change, to mimic there is a hole in memory range:
> 
>     @@ -1372,5 +1372,8 @@ __init void e820__memblock_setup(void)
>             /* Throw away partial pages: */
>             memblock_trim_memory(PAGE_SIZE);
>      
>     +       memblock_remove(0x140000000, 0x200000);
>     +
>             memblock_dump_all();
>      }
>     
>   Then the memblock dump shows:
> 
>      MEMBLOCK configuration:
>       memory size = 0x000000017fd7dc00 reserved size = 0x0000000005a97 9c2
>       memory.cnt  = 0x4
>       memory[0x0]     [0x0000000000001000-0x000000000009efff], 0x000000000009e000 bytes on node 0 flags: 0x0
>       memory[0x1]     [0x0000000000100000-0x00000000bffdefff], 0x00000000bfedf000 bytes on node 0 flags: 0x0
>   +-  memory[0x2]     [0x0000000100000000-0x000000013fffffff], 0x0000000040000000 bytes on node 1 flags: 0x0
>   +-  memory[0x3]     [0x0000000140200000-0x00000001bfffffff], 0x000000007fe00000 bytes on node 1 flags: 0x0
> 
>   We can see the original one memblock region is divided into two, with a hole
>   of 2M in the middle.

Yes, that makes sense.

> 
>   Not sure this is a reasonable mimic of memory hole. Also I tried to
>   punch a larger hole, e.g. 10M, still see the behavioral change.
> 
> The /proc/zoneinfo result:
> 
>     w/o patch
> 
>     Node 1, zone   Normal
>       pages free     469271
>             boost    0
>             min      8567
>             low      10708
>             high     12849
>             promo    14990
>             spanned  786432
>             present  785920
>             contigu  0           <--- zone is non-contiguous
>             managed  766024
>             cma      0
>     
>     with patch
> 
>     Node 1, zone   Normal
>       pages free     121098
>             boost    0
>             min      8665
>             low      10831
>             high     12997
>             promo    15163
>             spanned  786432
>             present  785920
>             contigu  1           <--- zone is contiguous
>             managed  773041
>             cma      0
> 
>   This shows we treat Node 1 Zone Normal as non-contiguous before, but treat
>   it a contiguous zone after this patch.
> 
> Reason:
> 
>   set_zone_contiguous()
>       __pageblock_pfn_to_page()
>           pfn_to_online_page()
> 	      pfn_section_valid() <--- check subsection
> 
>   When SPARSEMEM_VMEMMEP is set, pfn_section_valid() checks subsection bit to
>   decide if it is valid. For a hole, the corresponding bit is not set. So it
>   is non-contiguous before the patch.
> 
>   After this patch, the memory map in this hole also contributes to
>   pages_with_online_memmap, so it is treated as contiguous.

That means that mm init code actually initialized a memmap, so there is
a memmap there that is properly initialized?

So init_unavailable_range()->for_each_valid_pfn() processed these
sub-section holes I guess.

subsection_map_init() takes care of initializing the subsections. That
happens before memmap_init() in free_area_init().

Is there a problem in for_each_valid_pfn()?


And I think there is in first_valid_pfn:

	if (valid_section(ms) &&
	    (early_section(ms) || pfn_section_first_valid(ms, &pfn))) {
		rcu_read_unlock_sched();
		return pfn;
	}

The PFN is valid, but we actually care about whether it will be online.
So likely, we should skip over sub-sections here also for early sections
(even though the memmap exist, nobody should be looking at it, just like
for an offline memory section).

> 
> Some question:
> 
>   I suspect with !SPARSEMEM_VMEMMEP, we always treat Zone Normal as
>   contiguous, because we don't set subsection. So it looks the behavior is
>   different from SPARSEMEM_VMEMMEP. But I didn't manage to build kernel with
>   !SPARSEMEM_VMEMMEP to verify.
> 
>   I see the discussion on defining zone->contiguous as safe to use
>   pfn_to_page() for the whole zone. For this purpose, current change looks
>   good to me. Since we do allocate and init memory map for holes.

Right.

> 
>   But pageblock_pfn_to_page() is used for compaction and other. A pfn with
>   memory map but no actual memory seems not guarantee to be a usable page. So
>   the correct usage of pageblock_pfn_to_page() is after
>   pageblock_pfn_to_page() return a page, we should validate each page in the
>   range before using? I am a little lost here.

These non-existent pages (holes) are no different than allocated
un-movable memory. So compaction code must deal with them. Just like
smaller memory holes that don't cover a full memory section.

-- 
Cheers,

David