Re: [PATCH] mm/sparse: set SECTION_NID_SHIFT to 6

[David Hildenbrand]

> Sounds nice to me, so here's a patch. Could you review this?

Hi,

sorry for the late reply, I was on vacation. Please send it as a proper 
stand-alone patch next time, such that it

1. won't get silently ignored by reviewers/maintainers within a thread
2. Can easily get picked up/tested

Some minor comments below.

> Thanks,
> Naoya Horiguchi
> ---
>  From a146c9f12ae8985c8985a5861330f7528cd14fe8 Mon Sep 17 00:00:00 2001
> From: Naoya Horiguchi <naoya.horiguchi@xxxxxxx>
> Date: Mon, 28 Jun 2021 15:50:37 +0900
> Subject: [PATCH] mm/sparse: set SECTION_NID_SHIFT to 6
>
> Hagio-san reported that crash utility can see bit 4 in section_mem_map
> (SECTION_TAINT_ZONE_DEVICE) to be set, even if we do not use any
> ZONE_DEVICE ilke pmem or HMM.  This problem could break crash-related

s/ilke/like/

> toolsets and/or other memory analysis tools.

I'd rephrase this to "Having SECTION_TAINT_ZONE_DEVICE set for wrong 
sections forces pfn_to_online_page() through the slow path, but doesn't 
actually break the kernel. However, it can break crash-related toolsets."

However, I am not sure why it actually breaks crash? crash would have to 
implement the same slow-path check and would have to double-check the 
sub-section present map. Then, it should just work like 
pfn_to_online_page() and not have a real issue. What am I missing?

> The root cause is that SECTION_NID_SHIFT is incorrectly set to 3,
> while we use lower 5 bits for SECTION_* flags.  So bit 3 and 4 can be
> overlapped by sub-field for early NID, and bit 4 is unexpectedly set
> on (for example) NUMA node id is 2 or 3.
>
> To fix it, set SECTION_NID_SHIFT to 6 which is the minimum number of
> available bits of section flag field.
>
> [1]: https://github.com/crash-utility/crash/commit/0b5435e10161345cf713ed447a155a611a1b408b

[1] is never referenced

> Fixes: 1f90a3477df3 ("mm: teach pfn_to_online_page() about ZONE_DEVICE section collisions")
> Cc: stable@xxxxxxxxxxxxxxx # v5.12+

^ I am not really convinced that this is a stable fix. It forces 
something through the slow path, but the kernel itself is not broken, no?

> Reported-by: Kazuhito Hagio <k-hagio-ab@xxxxxxx>
> Suggested-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> Signed-off-by: Naoya Horiguchi <naoya.horiguchi@xxxxxxx>
> ---
>   include/linux/mmzone.h | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index fcb535560028..d6aa2a196aeb 100644
> --- a/include/linux/mmzone.h
> +++ b/include/linux/mmzone.h
> @@ -1357,6 +1357,7 @@ extern size_t mem_section_usage_size(void);
>    *      worst combination is powerpc with 256k pages,
>    *      which results in PFN_SECTION_SHIFT equal 6.
>    * To sum it up, at least 6 bits are available.
> + * SECTION_NID_SHIFT is set to 6 based on this fact.

I'd drop that comment or rephrase to ("once this changes, don't forget 
to adjust SECTION_NID_SHIFT")

>    */
>   #define SECTION_MARKED_PRESENT		(1UL<<0)
>   #define SECTION_HAS_MEM_MAP		(1UL<<1)
> @@ -1365,7 +1366,7 @@ extern size_t mem_section_usage_size(void);
>   #define SECTION_TAINT_ZONE_DEVICE	(1UL<<4)
>   #define SECTION_MAP_LAST_BIT		(1UL<<5)
>   #define SECTION_MAP_MASK		(~(SECTION_MAP_LAST_BIT-1))
> -#define SECTION_NID_SHIFT		3
> +#define SECTION_NID_SHIFT		6
>   
>   static inline struct page *__section_mem_map_addr(struct mem_section *section)
>   {

Change itself looks correct to me.

Acked-by: David Hildenbrand <david@xxxxxxxxxx>

--
Thanks,

David / dhildenb