There is a small race between when the cycle count is read from
the hardware and when the epoch stabilizes. Consider this
scenario:

 CPU0                           CPU1
 ----                           ----
 cyc = read_sched_clock()
 cyc_to_sched_clock()
                                 update_sched_clock()
                                  ...
                                  cd.epoch_cyc = cyc;
  epoch_cyc = cd.epoch_cyc;
  ...
  epoch_ns + cyc_to_ns((cyc - epoch_cyc)

The cyc on cpu0 was read before the epoch changed. But we
calculate the nanoseconds based on the new epoch by subtracting
the new epoch from the old cycle count. Since epoch is most likely
larger than the old cycle count we calculate a large number that
will be converted to nanoseconds and added to epoch_ns, causing
time to jump forward too much.

Fix this problem by reading the hardware after the epoch has
stabilized.

Cc: Russell King <linux@arm.linux.org.uk>
Signed-off-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: John Stultz <john.stultz@linaro.org>