x86/pgtable: Always inline p4d_index
PATCH 1/1 x86/pgtable: Always inline p4d_index
When building an allyesconfig build with Clang, the kernel fails to link arch/x86/platform/efi/efi_64.o because of a failed BUILD_BUG_ON:
ld: arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
(.text+0x8e5): undefined reference to `__compiletime_assert_277'
Since there are several BUILD_BUG_ONs in efi_64.c, I isolated it down to the following:
BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));
After some research, it turns out that there is a new config option called NO_AUTO_INLINE, which adds '-fno-inline-functions' to KBUILD_CFLAGS so that the compiler does not auto inline small functions, which causes this BUILD_BUG_ON to fail because p4d_index is no longer an integer constant expression.
According to the help text of the config, functions explicitly marked inline should still be inlined. As it turns out, GCC and Clang both support '-fno-inline-functions' but Clang only inlines functions when they are marked with an always inline annotation[1].
Since it's expected that p4d_index should always be inlined so that its value can be evaluated at build time, mark it as __always_inline.
[1]: https://github.com/llvm-mirror/clang/blob/master/lib/Frontend/CompilerInvocation.cpp#L637-L656
--- arch/x86/include/asm/pgtable.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/include/asm/pgtable.h +1 -1
@@ -918,7 +918,7 @@ static inline int p4d_bad(p4d_t p4d) }#endif /* CONFIG_PGTABLE_LEVELS > 3 */ -static inline unsigned long p4d_index (unsigned long address) +static __always_inline unsigned long p4d_index (unsigned long address) {return (address >>P4D_SHIFT ) & (PTRS_PER_P4D -1 ); }
Signature · 1 line
2.20.0
Quoted 23 lines · Nathan Chancellor
When building an allyesconfig build with Clang, the kernel fails to link
arch/x86/platform/efi/efi_64.obecause of a failedBUILD_BUG_ON:ld:
arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
(.text+0x8e5): undefined reference to `__compiletime_assert_277'Since there are several
BUILD_BUG_ONsinefi_64.c, I isolated it down to the following:BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));After some research, it turns out that there is a new config option called
NO_AUTO_INLINE, which adds '-fno-inline-functions' toKBUILD_CFLAGSso that the compiler does not auto inline small functions, which causes thisBUILD_BUG_ONto fail becausep4d_indexis no longer an integer constant expression.According to the help text of the config, functions explicitly marked inline should still be inlined. As it turns out, GCC and Clang both support '-fno-inline-functions' but Clang only inlines functions when they are marked with an always inline annotation[1].
Quoted 4 lines
Since it's expected that
p4d_indexshould always be inlined so that its value can be evaluated at build time, mark it as__always_inline.[1]: https://github.com/llvm-mirror/clang/blob/master/lib/Frontend/CompilerInvocation.cpp#L637-L656
(resending as text-plain; why does gmail silently switch it back to html...)
Just some top secret github-fu: this link is bound to break over time as the source changes. While on that page, if you hit y, it's hotkeyed to update the URL to pin it to a SHA. That is a little more robust (the branch or even the tree could still disappear, or history gets rewritten (hopefully not on a master branch, but I've seen it happen accidentally (Firefox OS))).
This doesn't look like an invasive change to me, though I'm curious about the usefulness of CONFIG_NO_AUTO_INLINE if the kernel is still compiled at -O2.
Thanks for sending!
Quoted 23 lines
Signed-off-by Nathan Chancellor <[email protected]>--- arch/x86/include/asm/pgtable.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/pgtable.h +1
-2@@ -918,7 +918,7 @@ static inline int p4d_bad(p4d_t p4d) }#endif /* CONFIG_PGTABLE_LEVELS > 3 */ -static inline unsigned long p4d_index (unsigned long address) +static __always_inline unsigned long p4d_index (unsigned long address) {return (address >>P4D_SHIFT ) & (PTRS_PER_P4D -1 ); } --2.20.0
Signature · 2 lines
Thanks, ~Nick Desaulniers
PATCH v2 1/1 x86/pgtable: Always inline p4d_index
In reply to Nathan ChancellorWhen building an allyesconfig build with Clang, the kernel fails to link arch/x86/platform/efi/efi_64.o because of a failed BUILD_BUG_ON:
ld: arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
(.text+0x8e5): undefined reference to `__compiletime_assert_277'
Since there are several BUILD_BUG_ONs in efi_64.c, I isolated it down to the following:
BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));
After some research, it turns out that there is a new config option called NO_AUTO_INLINE, which adds '-fno-inline-functions' to KBUILD_CFLAGS so that the compiler does not auto inline small functions, which causes this BUILD_BUG_ON to fail because p4d_index is no longer an integer constant expression.
According to the help text of the config, functions explicitly marked inline should still be inlined. As it turns out, GCC and Clang both support '-fno-inline-functions' but Clang only inlines functions when they are marked with an always inline annotation[1].
Since it's expected that p4d_index should always be inlined so that its value can be evaluated at build time, mark it as __always_inline.
---
v1 -> v2:
* Add Nick's Reviewed-by tag.
* Update the GitHub URL to use the official LLVM repository and pin to a SHA so that it is always valid, as suggested by Nick.
arch/x86/include/asm/pgtable.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/include/asm/pgtable.h +1 -1
@@ -918,7 +918,7 @@ static inline int p4d_bad(p4d_t p4d) }#endif /* CONFIG_PGTABLE_LEVELS > 3 */ -static inline unsigned long p4d_index (unsigned long address) +static __always_inline unsigned long p4d_index (unsigned long address) {return (address >>P4D_SHIFT ) & (PTRS_PER_P4D -1 ); }
Signature · 1 line
2.20.1
Quoted 22 lines · Nathan Chancellor
When building an allyesconfig build with Clang, the kernel fails to link
arch/x86/platform/efi/efi_64.obecause of a failedBUILD_BUG_ON:ld:
arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
(.text+0x8e5): undefined reference to `__compiletime_assert_277'Since there are several
BUILD_BUG_ONsinefi_64.c, I isolated it down to the following:BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));After some research, it turns out that there is a new config option called
NO_AUTO_INLINE, which adds '-fno-inline-functions' toKBUILD_CFLAGSso that the compiler does not auto inline small functions, which causes thisBUILD_BUG_ONto fail becausep4d_indexis no longer an integer constant expression.According to the help text of the config, functions explicitly marked inline should still be inlined. As it turns out, GCC and Clang both support '-fno-inline-functions' but Clang only inlines functions when they are marked with an always inline annotation[1].
Quoted 8 lines
Since it's expected that
p4d_indexshould always be inlined so that its value can be evaluated at build time, mark it as__always_inline.Signed-off-by Nathan Chancellor <[email protected]>Reviewed-by Nick Desaulniers <[email protected]>
Hi,
I see the same build error when using gcc 4.8.5 (SUSE Linux). This patch fixes that build error.
Thanks.
Quoted 26 lines
---
v1 -> v2:
* Add Nick's Reviewed-by tag.
* Update the GitHub URL to use the official LLVM repository and pin to a SHA so that it is always valid, as suggested by Nick.
arch/x86/include/asm/pgtable.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/pgtable.h +1
-1@@ -918,7 +918,7 @@ static inline int p4d_bad(p4d_t p4d) }#endif /* CONFIG_PGTABLE_LEVELS > 3 */ -static inline unsigned long p4d_index (unsigned long address) +static __always_inline unsigned long p4d_index (unsigned long address) {return (address >>P4D_SHIFT ) & (PTRS_PER_P4D -1 ); }
Signature · 1 line
~Randy
Quoted 9 lines · Nathan Chancellor
When building an allyesconfig build with Clang, the kernel fails to link
arch/x86/platform/efi/efi_64.obecause of a failedBUILD_BUG_ON:ld:
arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
(.text+0x8e5): undefined reference to `__compiletime_assert_277'Since there are several
BUILD_BUG_ONsinefi_64.c, I isolated it down to the following:
Quoted 7 lines
BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));After some research, it turns out that there is a new config option called
NO_AUTO_INLINE, which adds '-fno-inline-functions' toKBUILD_CFLAGSso that the compiler does not auto inline small functions, which causes thisBUILD_BUG_ONto fail becausep4d_indexis no longer an integer constant expression.
There is no tree where this config option exists.
Quoted 7 lines
According to the help text of the config, functions explicitly marked inline should still be inlined. As it turns out, GCC and Clang both support '-fno-inline-functions' but Clang only inlines functions when they are marked with an always inline annotation[1].
Since it's expected that
p4d_indexshould always be inlined so that its value can be evaluated at build time, mark it as__always_inline.
Sorry no, that's just papering over the problem.
The point is that NO_AUTO_INLINE is/was meant to prevent the compiler from automatically inlining functions, which are NOT marked inline in any form in order to expand the traceability.
With GCC this makes sense, because it still inlines functions which are solely marked inline and even if it decides not to inline them it will evaluate them if they expand to a build time constant expression.
Now Clang decided to give -fno-inline-functions a different meaning, i.e. the same as GCC has for -fno-inline, which prevents inlining for everything except functions which are marked __always_inline.
So just "fixing" the build wreckage by duct taping one single instance of inline functions is really a bad idea. The resulting kernel will be bloatware because all regular inlines and we have tons of them will turn into function calls even if the function overhead is larger than the resulting inline code. Not to talk about the performance impact.
Anyway, as this option is found to be nowhere there is no point to apply this patch.
Thanks,
tglx
Quoted 13 lines · Thomas Gleixner
Quoted 11 lines · Nathan Chancellor
When building an allyesconfig build with Clang, the kernel fails to link
arch/x86/platform/efi/efi_64.obecause of a failedBUILD_BUG_ON:ld:
arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
(.text+0x8e5): undefined reference to `__compiletime_assert_277'Since there are several
BUILD_BUG_ONsinefi_64.c, I isolated it down to the following:BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));
Quoted 8 lines
Quoted 5 lines
After some research, it turns out that there is a new config option called
NO_AUTO_INLINE, which adds '-fno-inline-functions' toKBUILD_CFLAGSso that the compiler does not auto inline small functions, which causes thisBUILD_BUG_ONto fail becausep4d_indexis no longer an integer constant expression.There is no tree where this config option exists.
Hmmm, appears that Linus rejected the pull that would have added this option, which sat in -next for a month and a half.
https://lore.kernel.org/lkml/CAHk-=wiLt3rgeyM3BWAd5VJrKcnxxuHybwoQiDGMgyspo6oXDg@mail.gmail.com/
It is no longer in -next or the kbuild tree so I guess it was scrapped.
Quoted 22 lines
Quoted 7 lines
According to the help text of the config, functions explicitly marked inline should still be inlined. As it turns out, GCC and Clang both support '-fno-inline-functions' but Clang only inlines functions when they are marked with an always inline annotation[1].
Since it's expected that
p4d_indexshould always be inlined so that its value can be evaluated at build time, mark it as__always_inline.Sorry no, that's just papering over the problem.
The point is that
NO_AUTO_INLINEis/was meant to prevent the compiler from automatically inlining functions, which are NOT marked inline in any form in order to expand the traceability.With GCC this makes sense, because it still inlines functions which are solely marked inline and even if it decides not to inline them it will evaluate them if they expand to a build time constant expression.
Now Clang decided to give -fno-inline-functions a different meaning, i.e. the same as GCC has for -fno-inline, which prevents inlining for everything except functions which are marked
__always_inline.
Quoted 6 lines
So just "fixing" the build wreckage by duct taping one single instance of inline functions is really a bad idea. The resulting kernel will be bloatware because all regular inlines and we have tons of them will turn into function calls even if the function overhead is larger than the resulting inline code. Not to talk about the performance impact.
Far points. I don't disagree that this is a band aid patch but I was more concerned about the build error than the runtime impact since this was uncovered with an allyesconfig build, which we aren't booting anyways.
Quoted 3 lines
Anyway, as this option is found to be nowhere there is no point to apply this patch.
Indeed. If the config reappears, it should probably 'depends on CC_IS_GCC'.
Quoted 4 lines
Thanks,
tglx
Thank you for the reply and review!
Nathan