[RFC PATCH 4/4] linux: remove time64 redirects for __TIMESIZE=64 architectures

Adhemerval Zanella Netto adhemerval.zanella@linaro.org
Wed Nov 11 16:46:05 GMT 2026



On 11/11/26 12:49, Arnd Bergmann wrote:
> On Wed, Nov 11, 2026, at 14:07, Arnd Bergmann wrote:
>> On Wed, Nov 11, 2026, at 13:09, Adhemerval Zanella Netto wrote:
>>>
>>> The symbol is already set on all four of them, because their configure
>>> fragments raise the minimum kernel to 5.1 or higher (arc 5.1.0, or1k
>>> 5.4.0, riscv32 5.4.0, loongarch32 6.19.0).  So the kernel-features.h hunk
>>> does not change anything for them.
>>
>> Ok
> 
> One more questions about the minimum kernel version: are the versions
> going to be the constant for the foreseeable future, or is there a
> process for changing them at some point, possibly to 5.4 or higher
> for 32-bit targets?
> 
> I see that linux-3.2 was four years old when it became the minimum
> supported version back in 2016. Now linux-3.2 is 15 years old, and
> even linux-5.1 at 7.5 years feels like an eternity ago. The oldest
> LTS kernel is currently linux-5.10 from 2020. The CIP SLTS-4.19
> still has two years left but has a very limited scope.
> 
> From what I understand, keeping this long backwards compatibility
> is useful to run new containers on older enterprise or desktop
> distro kernels that may still have extended support (RHEL8,
> Ubuntu 18.04). It feels much less relevant for 32-bit Arm and
> other embedded targets.
We removed the kernel runtime startup check (b46d2506567 - 2.36) and the
LD_ASSUME_KERNEL check during process startup, so the minimum kernel version
is not enforced anymore.

The --enable-kernel just define which __ASSUME_* macros are set in
kernel-features.h and the ENOSYS fallbacks. It also sets the NT_GNU_ABI_TAG
note, although I think it is just used by ld.so and file to prints the expected
kernel version.

And the minimal requires kernel has been raised before (7aab07e4e57, 6dad2c06889,
d0f5b3f851c, 5b4ecd3f956, 139ace95756).

Rising could be interesting to remove some fallbacks, for instance:

* Up to 4.4/4.11: the fallbacks for renameat2 (3.15), execveat (3.19),
  mlock2 (4.4) and statx (4.11). Also the arch socketcall gates: 4.3 on i386,
  s390 and m68k, and 4.4 on sparc32. Each is one or two files.

* 5.1 (the largest cleanup): __ASSUME_TIME64_SYSCALLS becomes unconditional
  on 32-bit ABIs.  It also drops the ipc() multiplexer path on i386, s390,
  m68k and similar arches.

* 5.3 to 6.6: clone3 (5.3), which removes the clone3_supported probe in
  clone-internal.c:76. Then faccessat2 (5.8), close_range (5.9),
  FUTEX_LOCK_PI2 (5.14) and fchmodat2 (6.6).

It does come with some costs though:

* Containers on old hosts, and this is the main reason the floor is still low.
  RHEL 7 is 3.10, RHEL 8 is 4.18 (maintained until 2029), Ubuntu 20.04 is 5.4,
  and Amazon Linux 2 is 4.14 or 5.10. Any floor above 4.18 drops RHEL 8 hosts.

* No clear error on older kernels, so users will need to strace or rely on
  some old debug tracing.

* Also, the ENOSYS fallback is used by some environments to advertise some
  missing features. For instance, seccomp sandboxes, valgrind, gVisor and
  older qemu-user.


More information about the Libc-alpha mailing list