On 4/20/21 12:07 AM, Arnd Bergmann wrote:

On Tue, Apr 20, 2021 at 5:10 AM Matthew Wilcox [off-list ref] wrote:

quoted

On Tue, Apr 20, 2021 at 02:48:17AM +0000, Vineet Gupta wrote:

quoted

quoted

32-bit architectures which expect 8-byte alignment for 8-byte integers
and need 64-bit DMA addresses (arc, arm, mips, ppc) had their struct
page inadvertently expanded in 2019.

FWIW, ARC doesn't require 8 byte alignment for 8 byte integers. This is
only needed for 8-byte atomics due to the requirements of LLOCKD/SCOND
instructions.

Ah, like x86?  OK, great, I'll drop your arch from the list of
affected.  Thanks!

I mistakenly assumed that i386 and m68k were the only supported
architectures with 32-bit alignment on u64. I checked it now and found

$ for i in /home/arnd/cross/x86_64/gcc-10.1.0-nolibc/*/bin/*-gcc ; do
echo `echo 'int a = __alignof__(long long);' | $i -xc - -Wall -S -o- |
grep -A1 a: | tail -n 1 | cut -f 3 -d\   `
${i#/home/arnd/cross/x86_64/gcc-10.1.0-nolibc/*/bin/} ; done
8 aarch64-linux-gcc
8 alpha-linux-gcc
4 arc-linux-gcc
8 arm-linux-gnueabi-gcc
8 c6x-elf-gcc
4 csky-linux-gcc
4 h8300-linux-gcc
8 hppa-linux-gcc
8 hppa64-linux-gcc
8 i386-linux-gcc
8 ia64-linux-gcc
2 m68k-linux-gcc
4 microblaze-linux-gcc
8 mips-linux-gcc
8 mips64-linux-gcc
8 nds32le-linux-gcc
4 nios2-linux-gcc
4 or1k-linux-gcc
8 powerpc-linux-gcc
8 powerpc64-linux-gcc
8 riscv32-linux-gcc
8 riscv64-linux-gcc
8 s390-linux-gcc
4 sh2-linux-gcc
4 sh4-linux-gcc
8 sparc-linux-gcc
8 sparc64-linux-gcc
8 x86_64-linux-gcc
8 xtensa-linux-gcc

which means that half the 32-bit architectures do this. This may
cause more problems when arc and/or microblaze want to support
64-bit kernels and compat mode in the future on their latest hardware,
as that means duplicating the x86 specific hacks we have for compat.

What is alignof(u64) on 64-bit arc?

$ echo 'int a = __alignof__(long long);' | arc64-linux-gnu-gcc -xc - 
-Wall -S -o - | grep -A1 a: | tail -n 1 | cut -f 3
8

Yeah ARCv2 alignment of 4 for 64-bit data was a bit of surprise finding 
for me as well. When 64-bit load/stores were initially targeted by the 
internal Metaware compiler (llvm based) they decided to keep alignment 
to 4 still (granted hardware allowed this) and then gcc guys decided to 
follow the same ABI. I only found this by accident :-)

Can you point me to some specifics on the compat issue. For better of 
worse, arc64 does''t have a compat 32-bit mode, so everything is 
64-on-64 or 32-on-32 (ARC32 flavor of ARCv3)

Thx,
-Vineet