[PATCH v2 0/3] vdso: Keep the CLOCK_AUX base at full precision

[Zhan Xusheng]

v1: https://lore.kernel.org/all/20260831125557.1490398-1-zhanxusheng@xxxxxxxxxx

Changes in v2, all from Thomas' review of 1/3:

 - Drop the nsec_per_sec local and spell the divisor out at each site.
 - Store the remainder straight into the basetime, as the coarse clocks
   already do.  It works at every site including CLOCK_MONOTONIC, whose
   remainder feeds CLOCK_BOOTTIME, because the copy can read it back from
   the CLOCK_MONOTONIC basetime.  That also makes the copy take sec and
   nsec from the same place instead of one from the basetime and one from
   a local left over from the division.  It is not a codegen win: gcc 13
   emits the same 419 instructions either way, with two movs scheduled a
   slot later.
 - Keep "vdso_ts->sec = sec;" on its own line under the copy comment.
 - Keep the asm() barrier rather than OPTIMIZER_HIDE_VAR(), matching
   __iter_div_u64_rem().
 - 2/3 drops the same local and stores into the basetime too.
 - Picked up the Reviewed-by on 3/3, which is unchanged.

The problem 2/3 fixes: the CLOCK_AUX basetime is shifted down to
nanoseconds and back up, which drops the fractional nanoseconds of
xtime_nsec, so the vDSO floors the base and the cycle delta separately
where ktime_get_aux() floors their sum.  The vDSO reading ends up 0 or
1 ns below the syscall for the same clock, and clock_getres() advertises
1 ns for these clocks.  Thomas confirmed this was unintentional.

Checked under QEMU with an auxiliary clock enabled through
/sys/kernel/time/aux_clocks/0/aux_clock_enable, comparing CLOCK_AUX via
the vDSO against the raw syscall at offset 0, at +5.123456789 s, and with
offs_aux driven negative.  The two agree within read latency in all three,
and over 100000 interleaved pairs each the vDSO reading is never ahead of
a syscall reading taken after it.

The 1 ns bias itself is not measurable that way: a pair of reads costs
several hundred nanoseconds, so it cannot be separated from read latency.
The argument for it is the algebra in 2/3, plus a sweep over
(xtime_nsec, delta) which puts the 1 ns case at 60% for shift 24.

Zhan Xusheng (3):
  vdso/math64: Add and use __iter_div64_u64_rem()
  vdso/vsyscall: Keep the CLOCK_AUX base scaled
  vdso/gettimeofday: Assert that the clock id fits the dispatch mask

 include/vdso/math64.h   | 21 +++++++++++++++++++++
 kernel/time/vsyscall.c  | 28 ++++++++++++----------------
 lib/vdso/gettimeofday.c |  2 ++
 3 files changed, 35 insertions(+), 16 deletions(-)


base-commit: cee9395acd8043be0644b25c34bfa86623f2b935
-- 
2.43.0