Re: [PATCH 09/15] sh: rework sync_file_range ABI

["Arnd Bergmann" <arnd@arndb.de>]

On Fri, Jun 21, 2024, at 10:44, John Paul Adrian Glaubitz wrote:
> On Thu, 2024-06-20 at 18:23 +0200, Arnd Bergmann wrote:
>> From: Arnd Bergmann <arnd@arndb.de>
>> 
>> The unusual function calling conventions on superh ended up causing
>                                               ^^^^^^
>                                        It's spelled SuperH

Fixed now.

>> diff --git a/arch/sh/kernel/sys_sh32.c b/arch/sh/kernel/sys_sh32.c
>> index 9dca568509a5..d5a4f7c697d8 100644
>> --- a/arch/sh/kernel/sys_sh32.c
>> +++ b/arch/sh/kernel/sys_sh32.c
>> @@ -59,3 +59,14 @@ asmlinkage int sys_fadvise64_64_wrapper(int fd, u32 offset0, u32 offset1,
>>  				 (u64)len0 << 32 | len1, advice);
>>  #endif
>>  }
>> +
>> +/*
>> + * swap the arguments the way that libc wants it instead of
>
> I think "swap the arguments to the order that libc wants them" would
> be easier to understand here.

Done

>> diff --git a/arch/sh/kernel/syscalls/syscall.tbl b/arch/sh/kernel/syscalls/syscall.tbl
>> index bbf83a2db986..c55fd7696d40 100644
>> --- a/arch/sh/kernel/syscalls/syscall.tbl
>> +++ b/arch/sh/kernel/syscalls/syscall.tbl
>> @@ -321,7 +321,7 @@
>>  311	common	set_robust_list			sys_set_robust_list
>>  312	common	get_robust_list			sys_get_robust_list
>>  313	common	splice				sys_splice
>> -314	common	sync_file_range			sys_sync_file_range
>> +314	common	sync_file_range			sys_sh_sync_file_range6
>                                                                  ^^^^^^ 
> Why the suffix 6 here?

In a later part of my cleanup, I'm consolidating all the
copies of this function (arm64, mips, parisc, powerpc,
s390, sh, sparc, x86) and picked the name
sys_sync_file_range6() for common implementation.

I end up with four entry points here, so the naming is a bit
confusing:

- sys_sync_file_range() is only used on 64-bit architectures,
  on x32 and on mips-n32. This uses four arguments, including
  two 64-bit wide ones.

- sys_sync_file_range2() continues to be used on arm, powerpc,
  xtensa and now on sh, hexagon and csky. I change the
  implementation to take six 32-bit arguments, but the ABI
  remains the same as before, with the flags before offset.

- sys_sync_file_range6() is used for most other 32-bit ABIs:
  arc, m68k, microblaze, nios2, openrisc, parisc, s390, sh, sparc
  and x86. This also has six 32-bit arguments but in the
  default order (fd, offset, nbytes, flags).

- sys_sync_file_range7() is exclusive to mips-o32, this one
  has an unused argument and is otherwise the same as
  sys_sync_file_range6().

My plan is to then have some infrastructure to ensure
userspace tools (libc, strace, qemu, rust, ...) use the
same calling conventions as the kernel. I'm doing the
same thing for all other syscalls that have architecture
specific calling conventions, so far I'm using

fadvise64_64_7
fanotify_mark6
truncate3
truncate4
ftruncate3
ftruncate4
fallocate6
pread5
pread6
pwrite5
pwrite6
preadv5
preadv6
pwritev5
pwritev6
sync_file_range6
fadvise64_64_2
fadvise64_64_6
fadvise64_5
fadvise64_6
readahead4
readahead5

The last number here is usually the number of 32-bit
arguments, except for fadvise64_64_2 that uses the
same argument reordering trick as sync_file_range2.

I'm not too happy with the naming but couldn't come up with
anything clearer either, so let me know if you have any
ideas there.

>>  315	common	tee				sys_tee
>>  316	common	vmsplice			sys_vmsplice
>>  317	common	move_pages			sys_move_pages
>> @@ -395,6 +395,7 @@
>>  385	common	pkey_alloc			sys_pkey_alloc
>>  386	common	pkey_free			sys_pkey_free
>>  387	common	rseq				sys_rseq
>> +388	common	sync_file_range2		sys_sync_file_range2
>>  # room for arch specific syscalls
>>  393	common	semget				sys_semget
>>  394	common	semctl				sys_semctl
>
> I wonder how you discovered this bug. Did you look up the calling 
> convention on SuperH
> and compare the argument order for the sys_sync_file_range system call 
> documented there
> with the order in the kernel?

I had to categorize all architectures based on their calling
conventions to see if 64-bit arguments need aligned pairs or
not, so I wrote a set of simple C files that I compiled for
all architectures to see in which cases they insert unused
arguments or swap the order of the upper and lower halves.

SuperH, parisc and s390 are each slightly different from all the
others here, so I ended up reading the ELF psABI docs and/or
the compiler sources to be sure.
I also a lot of git history.

> Did you also check what order libc uses? I would expect libc on SuperH 
> misordering the
> arguments as well unless I am missing something. Or do we know that the 
> code is actually
> currently broken?

Yes, I checked glibc, musl and uclibc-ng for all the cases in
which the ABI made no sense, as well as to check that my analysis
of the kernel sources matches the expectations of the libc.

     Arnd