On Thu, 14 Apr 2022 11:18:03 +0200
Kurt Kanzenbach <kurt@xxxxxxxxxxxxx> wrote:
I finally ran this series through my tests, and it has some issues.
> Introduce fast/NMI safe accessor to clock tai for tracing. The Linux kernel
> tracing infrastructure has support for using different clocks to generate
> timestamps for trace events. Especially in TSN networks it's useful to have TAI
> as trace clock, because the application scheduling is done in accordance to the
> network time, which is based on TAI. With a tai trace_clock in place, it becomes
> very convenient to correlate network activity with Linux kernel application
> traces.
>
> Use the same implementation as ktime_get_boot_fast_ns() does by reading the
> monotonic time and adding the TAI offset. The same limitations as for the fast
> boot implementation apply. The TAI offset may change at run time e.g., by
> setting the time or using adjtimex() with an offset. However, these kind of
> offset changes are rare events. Nevertheless, the user has to be aware and deal
> with it in post processing.
>
> An alternative approach would be to use the same implementation as
> ktime_get_real_fast_ns() does. However, this requires to add an additional u64
> member to the tk_read_base struct. This struct together with a seqcount is
> designed to fit into a single cache line on 64 bit architectures. Adding a new
> member would violate this constraint.
>
> Signed-off-by: Kurt Kanzenbach <kurt@xxxxxxxxxxxxx>
> ---
> Documentation/core-api/timekeeping.rst | 1 +
> include/linux/timekeeping.h | 1 +
> kernel/time/timekeeping.c | 17 +++++++++++++++++
> 3 files changed, 19 insertions(+)
>
> diff --git a/Documentation/core-api/timekeeping.rst b/Documentation/core-api/timekeeping.rst
> index 729e24864fe7..22ec68f24421 100644
> --- a/Documentation/core-api/timekeeping.rst
> +++ b/Documentation/core-api/timekeeping.rst
> @@ -132,6 +132,7 @@ Some additional variants exist for more specialized cases:
> .. c:function:: u64 ktime_get_mono_fast_ns( void )
> u64 ktime_get_raw_fast_ns( void )
> u64 ktime_get_boot_fast_ns( void )
> + u64 ktime_get_tai_fast_ns( void )
> u64 ktime_get_real_fast_ns( void )
>
> These variants are safe to call from any context, including from
> diff --git a/include/linux/timekeeping.h b/include/linux/timekeeping.h
> index 78a98bdff76d..fe1e467ba046 100644
> --- a/include/linux/timekeeping.h
> +++ b/include/linux/timekeeping.h
> @@ -177,6 +177,7 @@ static inline u64 ktime_get_raw_ns(void)
> extern u64 ktime_get_mono_fast_ns(void);
> extern u64 ktime_get_raw_fast_ns(void);
> extern u64 ktime_get_boot_fast_ns(void);
> +extern u64 ktime_get_tai_fast_ns(void);
> extern u64 ktime_get_real_fast_ns(void);
>
> /*
> diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c
> index dcdcb85121e4..2c22023fbf5f 100644
> --- a/kernel/time/timekeeping.c
> +++ b/kernel/time/timekeeping.c
> @@ -532,6 +532,23 @@ u64 notrace ktime_get_boot_fast_ns(void)
> }
> EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
>
> +/**
> + * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
> + *
> + * The same limitations as described for ktime_get_boot_fast_ns() apply. The
> + * mono time and the TAI offset are not read atomically which may yield wrong
> + * readouts. However, an update of the TAI offset is an rare event e.g., caused
> + * by settime or adjtimex with an offset. The user of this function has to deal
> + * with the possibility of wrong timestamps in post processing.
> + */
> +u64 notrace ktime_get_tai_fast_ns(void)
> +{
> + struct timekeeper *tk = &tk_core.timekeeper;
> +
> + return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
As you are using this for tracing, can you open code the
ktime_get_mono_fast_ns(), otherwise we need to mark that function as
notrace. Not to mention, this is a fast path and using the noinline of
__ktime_get_fast_ns() should be less overhead.
That said, I hit this too:
less-5071 [000] d.h2. 498087876.351330: do_raw_spin_trylock <-_raw_spin_lock
less-5071 [000] d.h4. 498087876.351334: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
less-5071 [000] d.h5. 498087876.351334: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
less-5071 [000] d.h3. 498087876.351334: rcu_read_lock_sched_held <-lock_acquired
less-5071 [000] d.h5. 498087876.351337: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
kworker/u8:1-45 [003] d.h7. 1651009380.982749: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
kworker/u8:1-45 [003] d.h7. 1651009380.982749: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
kworker/u8:1-45 [003] d.h5. 1651009380.982749: rcu_read_lock_held_common <-rcu_read_lock_sched_held
kworker/u8:1-45 [003] d.h7. 498087876.375905: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
kworker/u8:1-45 [003] d.h7. 498087876.375905: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
kworker/u8:1-45 [003] d.h5. 498087876.375905: update_cfs_group <-task_tick_fair
kworker/u8:1-45 [003] d.h7. 498087876.375909: ktime_get_mono_fast_ns <-ktime_get_tai_fast_ns
The clock seems to be toggling between 1651009380 and 498087876 causing the
ftrace ring buffer to shutdown (it doesn't allow for time to go backwards).
This is running on a 32 bit x86.
-- Steve
> +}
> +EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
> +
> static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
> {
> struct tk_read_base *tkr;
