Re: [PATCH v2] mm/memblock: Correct totalram_pages accounting with KMSAN

[Wei Yang <richard.weiyang@gmail.com>]

On Thu, Sep 25, 2025 at 05:50:53PM +0300, Mike Rapoport wrote:
>On Thu, Sep 25, 2025 at 05:37:59AM -0700, SeongJae Park wrote:
>> Hello,
>> 
>> On Wed, 24 Sep 2025 12:03:01 +0200 Alexander Potapenko <glider@google.com> wrote:
>> 
>> > When KMSAN is enabled, `kmsan_memblock_free_pages()` can hold back pages
>> > for metadata instead of returning them to the early allocator. The callers,
>> > however, would unconditionally increment `totalram_pages`, assuming the
>> > pages were always freed. This resulted in an incorrect calculation of the
>> > total available RAM, causing the kernel to believe it had more memory than
>> > it actually did.
>> > 
>> > This patch refactors `memblock_free_pages()` to return the number of pages
>> > it successfully frees. If KMSAN stashes the pages, the function now
>> > returns 0; otherwise, it returns the number of pages in the block.
>> > 
>> > The callers in `memblock.c` have been updated to use this return value,
>> > ensuring that `totalram_pages` is incremented only by the number of pages
>> > actually returned to the allocator. This corrects the total RAM accounting
>> > when KMSAN is active.
>> > 
>> > Cc: Aleksandr Nogikh <nogikh@google.com>
>> > Fixes: 3c2065098260 ("init: kmsan: call KMSAN initialization routines")
>> > Signed-off-by: Alexander Potapenko <glider@google.com>
>> > Reviewed-by: David Hildenbrand <david@redhat.com>
>> [...]
>> > --- a/mm/mm_init.c
>> > +++ b/mm/mm_init.c
>> > @@ -2548,24 +2548,25 @@ void *__init alloc_large_system_hash(const char *tablename,
>> >  	return table;
>> >  }
>> >  
>> > -void __init memblock_free_pages(struct page *page, unsigned long pfn,
>> > -							unsigned int order)
>> > +unsigned long __init memblock_free_pages(struct page *page, unsigned long pfn,
>> > +					 unsigned int order)
>> >  {
>> >  	if (IS_ENABLED(CONFIG_DEFERRED_STRUCT_PAGE_INIT)) {
>> >  		int nid = early_pfn_to_nid(pfn);
>> >  
>> >  		if (!early_page_initialised(pfn, nid))
>> > -			return;
>> > +			return 0;
>> >  	}
>> 
>> I found this patch on mm-new tree is making my test machine (QEMU) reports much
>> less MemTotal even though KMSAN is disabled.  And modifying the above part to
>> be considered as free success (returning '1UL << order') fixed my issue.
>> Because the commit message says the purpose of this change is only for
>> KMSAN-stashed memory, maybe the above behavior change is not really intended?
>> 
>> I'm not familiar with this code so I'm unsure if the workaround is the right
>> fix.  But since I have no time to look this in deep for now, reporting first
>
>With DEFERRED_STRUCT_PAGE_INIT we count totalram_pages in
>memblock_free_all() but actually free them in deferred_init_memmap() and
>deferred_grow_zone().
>
>So returning '1UL << order' is a correct workaround, but the proper fix
>should update totalram_pages in the deferred path IMHO.
>

Maybe I did something similar at [1].

But this hit a problem for shmem, since shmem_fill_super() use
totalram_pages(). And before DEFERRED_STRUCT_PAGE_INIT finish, the size is too
small, so it can't boot up.

Per my understanding, shmem_fill_super() could be invoked after
memblock_discard(), so it is not proper to refactor to get ram size from
memblock.

Could we adjust shmem_default_max_blocks/shmem_default_max_inodes use memblock
at boot stage and use totalram_pages() after system is fully up? Or any other
suggestions?

[1]: http://lkml.kernel.org/r/20240726003612.5578-1-richard.weiyang@gmail.com

>-- 
>Sincerely yours,
>Mike.

-- 
Wei Yang
Help you, Help me