Re: [PATCH] Rename STAT_HAS_TIME32 to KERNEL_STAT64_HAS_TIME32

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

On Fri, Nov 4, 2022, at 02:54, YunQiang Su wrote:
> The macro name STAT_HAS_TIME32 is not so clear.
>
> This macro is used for some arch like MIPSn64.
> The kernel_stat/kernel_stat64 of it has a 32bit unsigned time value.
> Thus that has y2106 problem.
> So we should statx to solve this problem.

I was surprised that you identified this as a y2106 problem
rather than a y2038 problem, so I took a closer look at how this
is handled in the kernel. I found a few interesting bits, in
no particular order:

- mips-n64 does not have a 'stat64' interface, the kernel
  only provides the 'stat', which has the same layout as
  the n32 and o32 'stat64'. From the kernel perspective,
  using KERNEL_STAT64_HAS_TIME32 seems like a misnomer,
  but it's possible that glibc just names things differently
  here.

- parisc64 has the same problem by only exposing a 32-bit
  timestamp in stat, but unlike mips, it has both 'stat'
  and 'stat64' interfaces, with both using the same layout
  as the corresponding parisc32 syscalls. All other 64-bit
  architectures have a 'stat' interface with 64-bit
  timestamps, on alpha and sparc64 there is also a stat64
  interface because the 'stat' version is limited in
  other ways.

- There is a mix of signed and unsigned types in the
  asm/stat.h headers, and you correctly identified mips64
  as using an unsigned type (hence the y2106 limit),
  however mips32 is one of the ones using a signed field,
  and the kernel does not behave any different and
  just relies on gcc's -fno-strict-overflow to truncate
  a 64-bit time64_t into a 32-bit signed or unsigned
  type either way.

- When I worked on the original time64 support for the
  kernel, I'm sure we had discussed truncating the values
  to the correct range on the stat interfaces, but this
  has evidently never happened. We should probably add
  such truncation, but we would have to decide whether to
  truncate this to range of the declared type or more
  sensibly to the range of the traditional time_t type.

- Note that common file systems like xfs and ext3
  intentionally store filesystem timestamps as 'signed'
  for compatibility with 32-bit user space, so there
  is precedent for interpreting the values as signed
  on 64-bit architectures when the range is limited.

        Arnd