Re: [PATCH] x86: mark native_set_p4d() as __always_inline

[Kirill A. Shutemov]

On Tue, Jun 05, 2018 at 11:35:15AM +, Arnd Bergmann wrote:
> When CONFIG_OPTIMIZE_INLINING is enabled, the function native_set_p4d()
> may not be fully inlined into the caller, resulting in a false-positive
> warning about an access to the __pgtable_l5_enabled variable from a
> non-__init function, despite the original caller being an __init function:
> 
> WARNING: vmlinux.o(.text.unlikely+0x1429): Section mismatch in reference from 
> the function native_set_p4d() to the variable .init.data:__pgtable_l5_enabled
> WARNING: vmlinux.o(.text.unlikely+0x1429): Section mismatch in reference from 
> the function native_p4d_clear() to the variable 
> .init.data:__pgtable_l5_enabled
> The function native_set_p4d() references
> the variable __initdata __pgtable_l5_enabled.
> This is often because native_set_p4d lacks a __initdata
> annotation or the annotation of __pgtable_l5_enabled is wrong.
> 
> Marking the native_set_p4d function and its caller native_p4d_clear()
> avoids this problem.
> 
> I did not bisect the original cause, but I assume this is related to
> the recent rework that turned pgtable_l5_enabled() into an inline
> function, which in turn caused the compiler to make different inlining
> decisions.
> 
> Fixes: ad3fe525b950 ("x86/mm: Unify pgtable_l5_enabled usage in early boot 
> code")
> Signed-off-by: Arnd Bergmann 

Thanks for fixing this.

Acked-by: Kirill A. Shutemov 

Do you have a clue which native_set_p4d() call causes the issue?

-- 
 Kirill A. Shutemov

[PATCH] x86: mark native_set_p4d() as __always_inline

[Arnd Bergmann]

When CONFIG_OPTIMIZE_INLINING is enabled, the function native_set_p4d()
may not be fully inlined into the caller, resulting in a false-positive
warning about an access to the __pgtable_l5_enabled variable from a
non-__init function, despite the original caller being an __init function:

WARNING: vmlinux.o(.text.unlikely+0x1429): Section mismatch in reference from 
the function native_set_p4d() to the variable .init.data:__pgtable_l5_enabled
WARNING: vmlinux.o(.text.unlikely+0x1429): Section mismatch in reference from 
the function native_p4d_clear() to the variable .init.data:__pgtable_l5_enabled
The function native_set_p4d() references
the variable __initdata __pgtable_l5_enabled.
This is often because native_set_p4d lacks a __initdata
annotation or the annotation of __pgtable_l5_enabled is wrong.

Marking the native_set_p4d function and its caller native_p4d_clear()
avoids this problem.

I did not bisect the original cause, but I assume this is related to
the recent rework that turned pgtable_l5_enabled() into an inline
function, which in turn caused the compiler to make different inlining
decisions.

Fixes: ad3fe525b950 ("x86/mm: Unify pgtable_l5_enabled usage in early boot 
code")
Signed-off-by: Arnd Bergmann 
---
 arch/x86/include/asm/pgtable_64.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/include/asm/pgtable_64.h 
b/arch/x86/include/asm/pgtable_64.h
index 3c5385f9a88f..0fdcd21dadbd 100644
--- a/arch/x86/include/asm/pgtable_64.h
+++ b/arch/x86/include/asm/pgtable_64.h
@@ -216,7 +216,7 @@ static inline pgd_t pti_set_user_pgd(pgd_t *pgdp, pgd_t pgd)
 }
 #endif
 
-static inline void native_set_p4d(p4d_t *p4dp, p4d_t p4d)
+static __always_inline void native_set_p4d(p4d_t *p4dp, p4d_t p4d)
 {
pgd_t pgd;
 
@@ -230,7 +230,7 @@ static inline void native_set_p4d(p4d_t *p4dp, p4d_t p4d)
*p4dp = native_make_p4d(native_pgd_val(pgd));
 }
 
-static inline void native_p4d_clear(p4d_t *p4d)
+static __always_inline void native_p4d_clear(p4d_t *p4d)
 {
native_set_p4d(p4d, native_make_p4d(0));
 }
-- 
2.9.0