| Message ID | 20220118102756.1259149-1-ardb@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | ARM: Thumb2: align ALT_UP() sections sufficiently | expand |
On Tue, Jan 18, 2022 at 11:27:56AM +0100, Ard Biesheuvel wrote: > When building for Thumb2, the .alt.smp.init sections that are emitted by > the ALT_UP() patching code may not be 32-bit aligned, even though the > fixup_smp_on_up() routine expects that. This results in alignment faults > at module load time, which need to be fixed up by the fault handler. > > So let's align those sections explicitly, and avoid this from occurring. Are you seeing a problem that this patch fixes? This really should not matter. .alt.smp.init contents are always a whole number of 32-bit words. These are gathered by the linker into the .init.smpalt section, so the contents should always be a whole number of 32-bit words. This follows the .init.tagtable section, which is also a 32-bit word aligned structure built by the linker... which follows the .init.arch.info section and .init.proc.info sections which all have 32-bit alignment requirements.
On Tue, 18 Jan 2022 at 12:21, Russell King (Oracle) <linux@armlinux.org.uk> wrote: > > On Tue, Jan 18, 2022 at 11:27:56AM +0100, Ard Biesheuvel wrote: > > When building for Thumb2, the .alt.smp.init sections that are emitted by > > the ALT_UP() patching code may not be 32-bit aligned, even though the > > fixup_smp_on_up() routine expects that. This results in alignment faults > > at module load time, which need to be fixed up by the fault handler. > > > > So let's align those sections explicitly, and avoid this from occurring. > > Are you seeing a problem that this patch fixes? > > This really should not matter. .alt.smp.init contents are always a whole > number of 32-bit words. These are gathered by the linker into the > .init.smpalt section, so the contents should always be a whole number > of 32-bit words. > > This follows the .init.tagtable section, which is also a 32-bit word > aligned structure built by the linker... which follows the > .init.arch.info section and .init.proc.info sections which all have > 32-bit alignment requirements. > This only affects modules, not the core kernel. The .alt.smp.init section in a module is visible to the module loader, which means the module loader will make no attempt to position it at a 32-bit aligned address if the ELF alignment is only 16 bits, which appears to be the default in my Thumb2 build [gcc version 10.3.1 20211117 (Debian 10.3.0-13)] I only spotted this because do_fixup_smp_on_up() was shown as the most recent in-kernel fixup location in /proc/cpu/alignment.
On Tue, Jan 18, 2022 at 12:32:55PM +0100, Ard Biesheuvel wrote: > On Tue, 18 Jan 2022 at 12:21, Russell King (Oracle) > <linux@armlinux.org.uk> wrote: > > > > On Tue, Jan 18, 2022 at 11:27:56AM +0100, Ard Biesheuvel wrote: > > > When building for Thumb2, the .alt.smp.init sections that are emitted by > > > the ALT_UP() patching code may not be 32-bit aligned, even though the > > > fixup_smp_on_up() routine expects that. This results in alignment faults > > > at module load time, which need to be fixed up by the fault handler. > > > > > > So let's align those sections explicitly, and avoid this from occurring. > > > > Are you seeing a problem that this patch fixes? > > > > This really should not matter. .alt.smp.init contents are always a whole > > number of 32-bit words. These are gathered by the linker into the > > .init.smpalt section, so the contents should always be a whole number > > of 32-bit words. > > > > This follows the .init.tagtable section, which is also a 32-bit word > > aligned structure built by the linker... which follows the > > .init.arch.info section and .init.proc.info sections which all have > > 32-bit alignment requirements. > > > > This only affects modules, not the core kernel. The .alt.smp.init > section in a module is visible to the module loader, which means the > module loader will make no attempt to position it at a 32-bit aligned > address if the ELF alignment is only 16 bits, which appears to be the > default in my Thumb2 build [gcc version 10.3.1 20211117 (Debian > 10.3.0-13)] > > I only spotted this because do_fixup_smp_on_up() was shown as the most > recent in-kernel fixup location in /proc/cpu/alignment. Ok, thanks for the explanation.
diff --git a/arch/arm/include/asm/assembler.h b/arch/arm/include/asm/assembler.h index 5a530e25ea1a..51ae4674ca49 100644 --- a/arch/arm/include/asm/assembler.h +++ b/arch/arm/include/asm/assembler.h @@ -261,6 +261,7 @@ THUMB( fpreg .req r7 ) */ #define ALT_UP(instr...) \ .pushsection ".alt.smp.init", "a" ;\ + .align 2 ;\ .long 9998b - . ;\ 9997: instr ;\ .if . - 9997b == 2 ;\ @@ -272,6 +273,7 @@ THUMB( fpreg .req r7 ) .popsection #define ALT_UP_B(label) \ .pushsection ".alt.smp.init", "a" ;\ + .align 2 ;\ .long 9998b - . ;\ W(b) . + (label - 9998b) ;\ .popsection diff --git a/arch/arm/include/asm/processor.h b/arch/arm/include/asm/processor.h index 6af68edfa53a..bdc35c0e8dfb 100644 --- a/arch/arm/include/asm/processor.h +++ b/arch/arm/include/asm/processor.h @@ -96,6 +96,7 @@ unsigned long __get_wchan(struct task_struct *p); #define __ALT_SMP_ASM(smp, up) \ "9998: " smp "\n" \ " .pushsection \".alt.smp.init\", \"a\"\n" \ + " .align 2\n" \ " .long 9998b - .\n" \ " " up "\n" \ " .popsection\n"
When building for Thumb2, the .alt.smp.init sections that are emitted by the ALT_UP() patching code may not be 32-bit aligned, even though the fixup_smp_on_up() routine expects that. This results in alignment faults at module load time, which need to be fixed up by the fault handler. So let's align those sections explicitly, and avoid this from occurring. Cc: <stable@vger.kernel.org> Signed-off-by: Ard Biesheuvel <ardb@kernel.org> --- arch/arm/include/asm/assembler.h | 2 ++ arch/arm/include/asm/processor.h | 1 + 2 files changed, 3 insertions(+)