diff mbox series

arm64: alternatives: Move length validation in alternative_{insn,endif}

Message ID 20210414000803.662534-1-nathan@kernel.org
State Accepted
Commit 22315a2296f4c251fa92aec45fbbae37e9301b6c
Headers show
Series arm64: alternatives: Move length validation in alternative_{insn,endif} | expand

Commit Message

Nathan Chancellor April 14, 2021, 12:08 a.m. UTC 
After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is
set atomically"), LLVM's integrated assembler fails to build entry.S:

<instantiation>:5:7: error: expected assembly-time absolute expression
 .org . - (664b-663b) + (662b-661b)
      ^
<instantiation>:6:7: error: expected assembly-time absolute expression
 .org . - (662b-661b) + (664b-663b)
      ^

The root cause is LLVM's assembler has a one-pass design, meaning it
cannot figure out these instruction lengths when the .org directive is
outside of the subsection that they are in, which was changed by the
.arch_extension directive added in the above commit.

Apply the same fix from commit 966a0acce2fc ("arm64/alternatives: move
length validation inside the subsection") to the alternative_endif
macro, shuffling the .org directives so that the length validation
happen will always happen in the same subsections. alternative_insn has
not shown any issue yet but it appears that it could have the same issue
in the future so just preemptively change it.

Cc: stable@vger.kernel.org
Fixes: f7b93d42945c ("arm64/alternatives: use subsections for replacement sequences")
Link: https://github.com/ClangBuiltLinux/linux/issues/1347
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
---

Apologies if my explanation or terminology is off, I am only now getting
more familiar with assembly.

 arch/arm64/include/asm/alternative-macros.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)


base-commit: 738fa58ee1328481d1d7889e7c430b3401c571b9

Comments

Sami Tolvanen April 14, 2021, 5:46 p.m. UTC | #1 
Hi Nathan,

On Tue, Apr 13, 2021 at 5:09 PM Nathan Chancellor <nathan@kernel.org> wrote:
>

> After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> set atomically"), LLVM's integrated assembler fails to build entry.S:

>

> <instantiation>:5:7: error: expected assembly-time absolute expression

>  .org . - (664b-663b) + (662b-661b)

>       ^

> <instantiation>:6:7: error: expected assembly-time absolute expression

>  .org . - (662b-661b) + (664b-663b)

>       ^

>

> The root cause is LLVM's assembler has a one-pass design, meaning it

> cannot figure out these instruction lengths when the .org directive is

> outside of the subsection that they are in, which was changed by the

> .arch_extension directive added in the above commit.

>

> Apply the same fix from commit 966a0acce2fc ("arm64/alternatives: move

> length validation inside the subsection") to the alternative_endif

> macro, shuffling the .org directives so that the length validation

> happen will always happen in the same subsections. alternative_insn has

> not shown any issue yet but it appears that it could have the same issue

> in the future so just preemptively change it.

>

> Cc: stable@vger.kernel.org

> Fixes: f7b93d42945c ("arm64/alternatives: use subsections for replacement sequences")

> Link: https://github.com/ClangBuiltLinux/linux/issues/1347

> Signed-off-by: Nathan Chancellor <nathan@kernel.org>

> ---

>

> Apologies if my explanation or terminology is off, I am only now getting

> more familiar with assembly.

>

>  arch/arm64/include/asm/alternative-macros.h | 8 ++++----

>  1 file changed, 4 insertions(+), 4 deletions(-)

>

> diff --git a/arch/arm64/include/asm/alternative-macros.h b/arch/arm64/include/asm/alternative-macros.h

> index 5df500dcc627..8a078fc662ac 100644

> --- a/arch/arm64/include/asm/alternative-macros.h

> +++ b/arch/arm64/include/asm/alternative-macros.h

> @@ -97,9 +97,9 @@

>         .popsection

>         .subsection 1

>  663:   \insn2

> -664:   .previous

> -       .org    . - (664b-663b) + (662b-661b)

> +664:   .org    . - (664b-663b) + (662b-661b)

>         .org    . - (662b-661b) + (664b-663b)

> +       .previous

>         .endif

>  .endm

>

> @@ -169,11 +169,11 @@

>   */

>  .macro alternative_endif

>  664:

> +       .org    . - (664b-663b) + (662b-661b)

> +       .org    . - (662b-661b) + (664b-663b)

>         .if .Lasm_alt_mode==0

>         .previous

>         .endif

> -       .org    . - (664b-663b) + (662b-661b)

> -       .org    . - (662b-661b) + (664b-663b)

>  .endm

>

>  /*


Thank you for fixing these!

The patch looks correct to me, next-20210413 builds with LLVM_IAS=1
after I applied it, and defconfig built with both Clang and gcc boots
normally. Please feel free to add:

Reviewed-by: Sami Tolvanen <samitolvanen@google.com>

Tested-by: Sami Tolvanen <samitolvanen@google.com>


Sami
Nick Desaulniers April 14, 2021, 7:22 p.m. UTC | #2 
On Tue, Apr 13, 2021 at 5:09 PM Nathan Chancellor <nathan@kernel.org> wrote:
>

> After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> set atomically"), LLVM's integrated assembler fails to build entry.S:

>

> <instantiation>:5:7: error: expected assembly-time absolute expression

>  .org . - (664b-663b) + (662b-661b)

>       ^

> <instantiation>:6:7: error: expected assembly-time absolute expression

>  .org . - (662b-661b) + (664b-663b)

>       ^

>

> The root cause is LLVM's assembler has a one-pass design, meaning it

> cannot figure out these instruction lengths when the .org directive is

> outside of the subsection that they are in, which was changed by the

> .arch_extension directive added in the above commit.

>

> Apply the same fix from commit 966a0acce2fc ("arm64/alternatives: move

> length validation inside the subsection") to the alternative_endif

> macro, shuffling the .org directives so that the length validation

> happen will always happen in the same subsections. alternative_insn has

> not shown any issue yet but it appears that it could have the same issue

> in the future so just preemptively change it.


Thanks Nathan.
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>

Tested-by: Nick Desaulniers <ndesaulniers@google.com>


I did some additional disassembly comparison.  In case we ever need it
again, I'll copy it below for posterity.

$ ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu make LLVM=1 LLVM_IAS=1
-j72 O=/tmp/a defconfig all
$ b4 am https://lore.kernel.org/lkml/20210414000803.662534-1-nathan@kernel.org/
-o - | git am -3
$ ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu make LLVM=1 LLVM_IAS=1
-j72 O=/tmp/b defconfig all
$ for f in $(find /tmp/a/arch/arm64 -name \*.o); do llvm-objdump -dr
$f > $f.txt; done
$ for f in $(find /tmp/b/arch/arm64 -name \*.o); do llvm-objdump -dr
$f > $f.txt; done
$ for f in $(find /tmp/a/arch/arm64 -name \*.o); do diff -u $f.txt
$(echo $f.txt|sed 's/a/b/'); done | less

For no difference.  You can check more sections than .text by changing
`-d` to `-D` for llvm-objdump, though you're going to get a lot of
noise related to changes in .strtab and relocations referring to debug
info (.debug_str).  But if I drop your patch, rebuild, and recompare,
I see the same differences.

>

> Cc: stable@vger.kernel.org

> Fixes: f7b93d42945c ("arm64/alternatives: use subsections for replacement sequences")

> Link: https://github.com/ClangBuiltLinux/linux/issues/1347

> Signed-off-by: Nathan Chancellor <nathan@kernel.org>

> ---

>

> Apologies if my explanation or terminology is off, I am only now getting

> more familiar with assembly.

>

>  arch/arm64/include/asm/alternative-macros.h | 8 ++++----

>  1 file changed, 4 insertions(+), 4 deletions(-)

>

> diff --git a/arch/arm64/include/asm/alternative-macros.h b/arch/arm64/include/asm/alternative-macros.h

> index 5df500dcc627..8a078fc662ac 100644

> --- a/arch/arm64/include/asm/alternative-macros.h

> +++ b/arch/arm64/include/asm/alternative-macros.h

> @@ -97,9 +97,9 @@

>         .popsection

>         .subsection 1

>  663:   \insn2

> -664:   .previous

> -       .org    . - (664b-663b) + (662b-661b)

> +664:   .org    . - (664b-663b) + (662b-661b)

>         .org    . - (662b-661b) + (664b-663b)

> +       .previous

>         .endif

>  .endm

>

> @@ -169,11 +169,11 @@

>   */

>  .macro alternative_endif

>  664:

> +       .org    . - (664b-663b) + (662b-661b)

> +       .org    . - (662b-661b) + (664b-663b)

>         .if .Lasm_alt_mode==0

>         .previous

>         .endif

> -       .org    . - (664b-663b) + (662b-661b)

> -       .org    . - (662b-661b) + (664b-663b)

>  .endm

>

>  /*

>

> base-commit: 738fa58ee1328481d1d7889e7c430b3401c571b9

> --

> 2.31.1.272.g89b43f80a5

>



-- 
Thanks,
~Nick Desaulniers
Catalin Marinas April 15, 2021, 9:17 a.m. UTC | #3 
Hi Nathan,

On Tue, Apr 13, 2021 at 05:08:04PM -0700, Nathan Chancellor wrote:
> After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> set atomically"), LLVM's integrated assembler fails to build entry.S:

> 

> <instantiation>:5:7: error: expected assembly-time absolute expression

>  .org . - (664b-663b) + (662b-661b)

>       ^

> <instantiation>:6:7: error: expected assembly-time absolute expression

>  .org . - (662b-661b) + (664b-663b)

>       ^


I tried the latest Linus' tree and linux-next (defconfig) with this
commit in and I can't get your build error. I used both clang-10 from
Debian stable and clang-11 from Debian sid. So, which clang version did
you use or which kernel config options?

-- 
Catalin
Nathan Chancellor April 15, 2021, 1:25 p.m. UTC | #4 
On Thu, Apr 15, 2021 at 10:17:43AM +0100, Catalin Marinas wrote:
> Hi Nathan,

> 

> On Tue, Apr 13, 2021 at 05:08:04PM -0700, Nathan Chancellor wrote:

> > After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> > set atomically"), LLVM's integrated assembler fails to build entry.S:

> > 

> > <instantiation>:5:7: error: expected assembly-time absolute expression

> >  .org . - (664b-663b) + (662b-661b)

> >       ^

> > <instantiation>:6:7: error: expected assembly-time absolute expression

> >  .org . - (662b-661b) + (664b-663b)

> >       ^

> 

> I tried the latest Linus' tree and linux-next (defconfig) with this

> commit in and I can't get your build error. I used both clang-10 from

> Debian stable and clang-11 from Debian sid. So, which clang version did

> you use or which kernel config options?

> 

> -- 

> Catalin

> 


Hi Catalin,

Interesting, this reproduces for me with LLVM 12 or newer with just
defconfig.

$ make -j"$(nproc)" ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- LLVM=1 LLVM_IAS=1 defconfig arch/arm64/kernel/entry.o

https://github.com/ClangBuiltLinux/continuous-integration2/runs/2350258778?check_suite_focus=true

Cheers,
Nathan
Catalin Marinas April 15, 2021, 2:02 p.m. UTC | #5 
On Thu, Apr 15, 2021 at 06:25:57AM -0700, Nathan Chancellor wrote:
> On Thu, Apr 15, 2021 at 10:17:43AM +0100, Catalin Marinas wrote:

> > On Tue, Apr 13, 2021 at 05:08:04PM -0700, Nathan Chancellor wrote:

> > > After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> > > set atomically"), LLVM's integrated assembler fails to build entry.S:

> > > 

> > > <instantiation>:5:7: error: expected assembly-time absolute expression

> > >  .org . - (664b-663b) + (662b-661b)

> > >       ^

> > > <instantiation>:6:7: error: expected assembly-time absolute expression

> > >  .org . - (662b-661b) + (664b-663b)

> > >       ^

> > 

> > I tried the latest Linus' tree and linux-next (defconfig) with this

> > commit in and I can't get your build error. I used both clang-10 from

> > Debian stable and clang-11 from Debian sid. So, which clang version did

> > you use or which kernel config options?

> 

> Interesting, this reproduces for me with LLVM 12 or newer with just

> defconfig.


It fails for me as well with clang-12. Do you happen to know why it
works fine with previous clang versions?

-- 
Catalin
Sami Tolvanen April 15, 2021, 3:50 p.m. UTC | #6 
On Thu, Apr 15, 2021 at 7:02 AM Catalin Marinas <catalin.marinas@arm.com> wrote:
>

> On Thu, Apr 15, 2021 at 06:25:57AM -0700, Nathan Chancellor wrote:

> > On Thu, Apr 15, 2021 at 10:17:43AM +0100, Catalin Marinas wrote:

> > > On Tue, Apr 13, 2021 at 05:08:04PM -0700, Nathan Chancellor wrote:

> > > > After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> > > > set atomically"), LLVM's integrated assembler fails to build entry.S:

> > > >

> > > > <instantiation>:5:7: error: expected assembly-time absolute expression

> > > >  .org . - (664b-663b) + (662b-661b)

> > > >       ^

> > > > <instantiation>:6:7: error: expected assembly-time absolute expression

> > > >  .org . - (662b-661b) + (664b-663b)

> > > >       ^

> > >

> > > I tried the latest Linus' tree and linux-next (defconfig) with this

> > > commit in and I can't get your build error. I used both clang-10 from

> > > Debian stable and clang-11 from Debian sid. So, which clang version did

> > > you use or which kernel config options?

> >

> > Interesting, this reproduces for me with LLVM 12 or newer with just

> > defconfig.

>

> It fails for me as well with clang-12. Do you happen to know why it

> works fine with previous clang versions?


It looks like CONFIG_ARM64_AS_HAS_MTE is not set when we use the
integrated assembler with LLVM 11, and the code that breaks later
versions is gated behind CONFIG_ARM64_MTE.

Sami
Catalin Marinas April 15, 2021, 4:57 p.m. UTC | #7 
On Thu, Apr 15, 2021 at 08:50:25AM -0700, Sami Tolvanen wrote:
> On Thu, Apr 15, 2021 at 7:02 AM Catalin Marinas <catalin.marinas@arm.com> wrote:

> >

> > On Thu, Apr 15, 2021 at 06:25:57AM -0700, Nathan Chancellor wrote:

> > > On Thu, Apr 15, 2021 at 10:17:43AM +0100, Catalin Marinas wrote:

> > > > On Tue, Apr 13, 2021 at 05:08:04PM -0700, Nathan Chancellor wrote:

> > > > > After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> > > > > set atomically"), LLVM's integrated assembler fails to build entry.S:

> > > > >

> > > > > <instantiation>:5:7: error: expected assembly-time absolute expression

> > > > >  .org . - (664b-663b) + (662b-661b)

> > > > >       ^

> > > > > <instantiation>:6:7: error: expected assembly-time absolute expression

> > > > >  .org . - (662b-661b) + (664b-663b)

> > > > >       ^

> > > >

> > > > I tried the latest Linus' tree and linux-next (defconfig) with this

> > > > commit in and I can't get your build error. I used both clang-10 from

> > > > Debian stable and clang-11 from Debian sid. So, which clang version did

> > > > you use or which kernel config options?

> > >

> > > Interesting, this reproduces for me with LLVM 12 or newer with just

> > > defconfig.

> >

> > It fails for me as well with clang-12. Do you happen to know why it

> > works fine with previous clang versions?

> 

> It looks like CONFIG_ARM64_AS_HAS_MTE is not set when we use the

> integrated assembler with LLVM 11, and the code that breaks later

> versions is gated behind CONFIG_ARM64_MTE.


That explains it, thanks.

-- 
Catalin
Catalin Marinas April 15, 2021, 5:48 p.m. UTC | #8 
On Tue, 13 Apr 2021 17:08:04 -0700, Nathan Chancellor wrote:
> After commit 2decad92f473 ("arm64: mte: Ensure TIF_MTE_ASYNC_FAULT is

> set atomically"), LLVM's integrated assembler fails to build entry.S:

> 

> <instantiation>:5:7: error: expected assembly-time absolute expression

>  .org . - (664b-663b) + (662b-661b)

>       ^

> <instantiation>:6:7: error: expected assembly-time absolute expression

>  .org . - (662b-661b) + (664b-663b)

>       ^

> 

> [...]


Applied to arm64 (for-next/fixes), thanks!

[1/1] arm64: alternatives: Move length validation in alternative_{insn, endif}
      https://git.kernel.org/arm64/c/22315a2296f4

-- 
Catalin
diff mbox series

Patch

diff --git a/arch/arm64/include/asm/alternative-macros.h b/arch/arm64/include/asm/alternative-macros.h
index 5df500dcc627..8a078fc662ac 100644
--- a/arch/arm64/include/asm/alternative-macros.h
+++ b/arch/arm64/include/asm/alternative-macros.h
@@ -97,9 +97,9 @@ 
 	.popsection
 	.subsection 1
 663:	\insn2
-664:	.previous
-	.org	. - (664b-663b) + (662b-661b)
+664:	.org	. - (664b-663b) + (662b-661b)
 	.org	. - (662b-661b) + (664b-663b)
+	.previous
 	.endif
 .endm
 
@@ -169,11 +169,11 @@ 
  */
 .macro alternative_endif
 664:
+	.org	. - (664b-663b) + (662b-661b)
+	.org	. - (662b-661b) + (664b-663b)
 	.if .Lasm_alt_mode==0
 	.previous
 	.endif
-	.org	. - (664b-663b) + (662b-661b)
-	.org	. - (662b-661b) + (664b-663b)
 .endm
 
 /*