Re: [PATCH v2] mm: fix KASAN build error due to p*d_populate_kernel()

[Harry Yoo <harry.yoo@oracle.com>]

On Fri, Aug 22, 2025 at 06:02:40PM +0200, Andrey Ryabinin wrote:
> On 8/22/25 3:11 AM, Harry Yoo wrote:
> > On Thu, Aug 21, 2025 at 10:36:12AM -0700, Dave Hansen wrote:
> >> On 8/21/25 04:57, Harry Yoo wrote:
> >>> However, {pgd,p4d}_populate_kernel() is defined as a function regardless
> >>> of the number of page table levels, so the compiler may not optimize
> >>> them away. In this case, the following linker error occurs:
> > 
> > Hi, thanks for taking a look, Dave!
> > 
> > First of all, this is a fix-up patch of a mm-hotfixes patch series that
> > fixes a bug (I should have explained that in the changelog) [1].
> > 
> > [1] https://lore.kernel.org/linux-mm/20250818020206.4517-1-harry.yoo@oracle.com
> > 
> > I think we can continue discussing it and perhaps do that as part of
> > a follow-up series, because the current patch series need to be backported
> > to -stable and your suggestion to improve existing code doesn't require
> > -stable backports.
> > 
> > Does that sound fine?
> > 
> >> This part of the changelog confused me. I think it's focusing on the
> >> wrong thing.
> >>
> >> The code that's triggering this is literally:
> >>
> >>>                         pgd_populate(&init_mm, pgd,
> >>>                                         lm_alias(kasan_early_shadow_p4d));
> >>
> >> It sure _looks_ like it's unconditionally referencing the
> >> 'kasan_early_shadow_p4d' symbol. I think it's wrong to hide that with
> >> macro magic and just assume that the macros won't reference it.
> >>
> >> If a symbol isn't being defined, it shouldn't be referenced in C code.:q
>  
> That's not exactly the case for the kernel. It historically relied on being
> compiled with optimization and compiler being able to eliminate unused references.
> AFAIR BUILD_BUG_ON() works like that, there are also plenty of code like
> 
> if  (IS_ENABLED(CONFIG_SOMETHING))
> 	ptr = &something;
> else
> 	ptr = &something_else; 
> 
> e.g. irq_remaping_prepare();

Agreed. I've seen this pattern in many places.

> > A fair point, and that's what KASAN code has been doing for years.
> > 
> >> The right way to do it is to have an #ifdef in a header that avoids
> >> compiling in the reference to the symbol.
> > 
> > You mean defining some wrapper functions for p*d_populate_kernel() in
> > KASAN with different implementations based on ifdeffery?
> > 
> > Just to clarify, what should be the exact ifdeffery to cover these cases?
> > #if CONFIG_PGTABLE_LEVELS == 4 and 5, or
> > #ifdef __PAGETABLE_P4D_FOLDED and __PAGETABLE_PUD_FOLDED ?
> > 
> 
> I think ifdef should be the same as for symbol, so '#if CONFIG_PGTABLE_LEVELS > 4'
> for *_p4d and '#if CONFIG_PGTABLE_LEVELS > 3' for *_pud

Right.

> > I have no strong opinion on this, let's hear what KASAN folks think.
> > 
> 
> So, I think we have following options:
> 
> 1. Macros as you did.
> 2. Hide references in function under  '#if CONFIG_PGTABLE_LEVELS > x', like Dave suggested.
> 3. It should be enough to just add if in code like
>             if (CONFIG_PGTABLE_LEVELS > 4)
> 		pgd_populate_kernel(addr, pgd,
>                                           lm_alias(kasan_early_shadow_p4d));
> Compiler should be able to optimize it away.
> 
> 4. I guess that the link error is due to enabled CONFIG_DEBUG_VIRTUAL=y
> lm_alias() ends up with __phys_addr_symbol() function call which compiler can't optimize away.
> Technically we can declare __phys_addr_symbol() with __attribute__((pure)), so compiler will
> be able to optimize away this call, because the result should be unused.
> But I'm not sure we really want that, because it's debug function and even if the result is unused
> we might want to still have a check if symbol address is correct.
> 
> 
> I would probably prefer 3rd option, but I don't really have very strong opinion, so either way is fine.

I also prefer 3rd option (but 1st or 2nd is also fine to me)

That's two votes, I'll do 2nd option in the follow-up series unless
Dave or somebody else objects?

-- 
Cheers,
Harry / Hyeonggon