Linus Torvalds <torva...@linux-foundation.org> writes:
> On Wed, Nov 2, 2022 at 7:09 PM Michael Ellerman <m...@ellerman.id.au> wrote:
>>
>> - Fix an endian thinko in the asm-generic compat_arg_u64() which led to
>> syscall arguments
>> being swapped for some compat syscalls.
>
> Am I mis-reading this, or did this bug (introduced in this merge
> window by commit 43d5de2b67d7 "asm-generic: compat: Support BE for
> long long args in 32-bit ABIs") break *every* architecture?
No. Just RISC-V and powerpc.
> And people just didn't scream, because 32-bit code has just become so rare?
I had two systems and several VMs that booted happily with the bug
present, so there's some luck involved as to whether your userspace
trips over the bug in a way that matters.
But we did have people scream eventually :/
> Or is it just because those compat macros are effectively not used
> elsewhere, and x86 has its own versions? Looks like possibly mainly
> RISC-V?
Yeah. Although compat_arg_u64() is defined for all arches, it's only
used in places guarded by __ARCH_WANT_COMPAT_FOO macros, and those are
only selected by RISC-V and powerpc.
Full list is:
__ARCH_WANT_COMPAT_FADVISE64_64 riscv
__ARCH_WANT_COMPAT_FALLOCATE riscv, powerpc
__ARCH_WANT_COMPAT_FTRUNCATE64 riscv
__ARCH_WANT_COMPAT_PREAD64 riscv
__ARCH_WANT_COMPAT_PWRITE64 riscv
__ARCH_WANT_COMPAT_READAHEAD riscv
__ARCH_WANT_COMPAT_SYNC_FILE_RANGE riscv
__ARCH_WANT_COMPAT_TRUNCATE64 riscv
> Side note: why is it doing
>
> #ifndef compat_arg_u64
>
> at all? That macro is not actually defined anywhere else, so that
> #ifdef seems to be just confused.
That goes back to the original submission:
59c10c52f573 ("riscv: compat: syscall: Add compat_sys_call_table
implementation")
I guess it was following the example in asm-generic/compat.h where a
bunch of other things are guarded by ifndefs. But agree it's
unnecessarily flexible in this case until we have another definition.
cheers
