Re: [PATCH v2] mm/migrate: Fix do_pages_stat in 32-bit mode

[Zi Yan <ziy@nvidia.com>]

On 1 Jul 2025, at 12:58, Christoph Berg wrote:

> Re: David Hildenbrand
>> Subject should start with "mm/migrate:"
>> Likely we want a
>> Fixes:
>> and then this is probably "Reported-by:" paired with a "Closes:" link
>> to any such report.
>
> I included these now, except for "Closes:" which I have to idea what
> to put in.

Fixes should be:

Fixes: 5b1b561ba73c ("mm: simplify compat_sys_move_pages")

Closes could be a link to the bug report.

>
>> But I'm wondering how long this has already been like that. :)
>
> The now-offending "pages += chunk_nr" line is from 2010, but I think
> the bug is rather from 5b1b561ba73c8ab9c98e5dfd14dc7ee47efb6530 (2021)
> which reshuffled the array-vs-32-bit handling.
>
>> Something a bit more elegant might be:
>
> Thanks, I used your patch draft with some minor changes.
>
>>  static int get_compat_pages_array(const void __user *chunk_pages[],
>>                                   const void __user * __user *pages,
>> +                                 unsigned long chunk_offs,
>
> I replaced chunk_offs with "chunk_offset" since "offs" looked too much
> like plural (list of offsets) to me.
>
>>                 if (in_compat_syscall()) {
>>                         if (get_compat_pages_array(chunk_pages, pages,
>> -                                                  chunk_nr))
>> +                                                  chunk_offs, chunk_nr))
>>                                 break;
>>                 } else {
>>                         if (copy_from_user(chunk_pages, pages,
>
> The else branch here needs tweaking as well:
>
>                 } else {
> -                       if (copy_from_user(chunk_pages, pages,
> +                       if (copy_from_user(chunk_pages, pages + chunk_offset,
>                                       chunk_nr * sizeof(*chunk_pages)))
>
>
>> @@ -2440,11 +2442,11 @@ static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
>>                 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
>> -               if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
>> +               if (copy_to_user(status + chunk_offs, chunk_status,
>> +                                chunk_nr * sizeof(*status)))
>
> This seems to work, but honestly I am wondering, if copy_from_user
> needs a special 32-bit case, doesn't copy_to_user need special casing
> as well?
>
>> (untested, of course)
>
> The attached patch makes PG18's new numa test pass on amd64 kernels
> both in amd64 and i386 userlands.
>
> (In the meantime, PG git head got a workaround that limits the chunk
> size to the same 16 as used in do_pages_stat; I tested with the
> version before that.)
>
> Christoph
>
>
> From fdbcbc88825bc2e857dfeeebc91d62864e0774dd Mon Sep 17 00:00:00 2001
> From: Christoph Berg <myon@debian.org>
> Date: Tue, 24 Jun 2025 16:44:27 +0200
> Subject: [PATCH v2] mm/migrate: Fix do_pages_stat in 32-bit mode
>
> For arrays with more than 16 entries, the old code would incorrectly
> advance the pages pointer by 16 words instead of 16 compat_uptr_t.
> Fix by doing the pointer arithmetic inside get_compat_pages_array where
> pages32 is already a correctly-typed pointer.
>
> Discovered while working on PostgreSQL 18's new NUMA introspection code.
>
> Signed-off-by: Christoph Berg <myon@debian.org>
> Reported-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
> Reported-by: Tomas Vondra <tomas@vondra.me>
> Suggested-by: David Hildenbrand <david@redhat.com>
> Fixes: 5b1b561ba73c8ab9c98e5dfd14dc7ee47efb6530
> ---
>  mm/migrate.c | 14 ++++++++------
>  1 file changed, 8 insertions(+), 6 deletions(-)
>
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 8cf0f9c9599d..2c88f3b33833 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -2399,6 +2399,7 @@ static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
>
>  static int get_compat_pages_array(const void __user *chunk_pages[],
>  				  const void __user * __user *pages,
> +				  unsigned long chunk_offset,
>  				  unsigned long chunk_nr)
>  {
>  	compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
> @@ -2406,7 +2407,7 @@ static int get_compat_pages_array(const void __user *chunk_pages[],
>  	int i;
>
>  	for (i = 0; i < chunk_nr; i++) {
> -		if (get_user(p, pages32 + i))
> +		if (get_user(p, pages32 + chunk_offset + i))
>  			return -EFAULT;
>  		chunk_pages[i] = compat_ptr(p);
>  	}
> @@ -2425,27 +2426,28 @@ static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
>  #define DO_PAGES_STAT_CHUNK_NR 16UL
>  	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
>  	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
> +	unsigned long chunk_offset = 0;
>
>  	while (nr_pages) {
>  		unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
>
>  		if (in_compat_syscall()) {
>  			if (get_compat_pages_array(chunk_pages, pages,
> -						   chunk_nr))
> +						   chunk_offset, chunk_nr))
>  				break;
>  		} else {
> -			if (copy_from_user(chunk_pages, pages,
> +			if (copy_from_user(chunk_pages, pages + chunk_offset,
>  				      chunk_nr * sizeof(*chunk_pages)))
>  				break;
>  		}
>
>  		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
>
> -		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
> +		if (copy_to_user(status + chunk_offset, chunk_status,
> +				 chunk_nr * sizeof(*status)))
>  			break;
>
> -		pages += chunk_nr;
> -		status += chunk_nr;
> +		chunk_offset += chunk_nr;
>  		nr_pages -= chunk_nr;
>  	}
>  	return nr_pages ? -EFAULT : 0;
> -- 
> 2.47.2


Best Regards,
Yan, Zi