The quick check in tick_do_update_jiffies64() whether jiffies need to be
updated is not really correct under all circumstances and on all
architectures, especially not on 32bit systems.

The quick check does:

    if (now < READ_ONCE(tick_next_period))
    	return;

and the counterpart in the update is:

    WRITE_ONCE(tick_next_period, next_update_time);

This has two problems:

  1) On weakly ordered architectures there is no guarantee that the stores
     before the WRITE_ONCE() are visible which means that other CPUs can
     operate on a stale jiffies value.

  2) On 32bit the store of tick_next_period which is an u64 is split into
     two 32bit stores. If the first 32bit store advances tick_next_period
     far out and the second 32bit store is delayed (virt, NMI ...) then
     jiffies will become stale until the second 32bit store happens.

Address this by seperating the handling for 32bit and 64bit.

On 64bit problem #1 is addressed by replacing READ_ONCE() / WRITE_ONCE()
with smp_load_acquire() / smp_store_release().

On 32bit problem #2 is addressed by protecting the quick check with the
jiffies sequence counter. The load and stores can be plain because the
sequence count mechanics provides the required barriers already.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Link: https://lore.kernel.org/r/87czzpc02w.fsf@nanos.tec.linutronix.de