Now that we can let an architecture override the timestamping
function, it becomes desirable to ensure that, should the
architecture code switch its timestamping code to sched_clock
once it has been registered, the sched_clock inherits the
timestamp value as its new epoch.
This ensures that the time stamps are continuous and that there
is no jitter other than that introduced by the lack of quality
of the timestamping clock.
Signed-off-by: Marc Zyngier <marc.zyng...@arm.com>
---
kernel/time/sched_clock.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/kernel/time/sched_clock.c b/kernel/time/sched_clock.c
index 142b07619918..ee1bd449ec81 100644
--- a/kernel/time/sched_clock.c
+++ b/kernel/time/sched_clock.c
@@ -192,6 +192,16 @@ sched_clock_register(u64 (*read)(void), int bits, unsigned
long rate)
new_epoch = read();
cyc = cd.actual_read_sched_clock();
ns = rd.epoch_ns + cyc_to_ns((cyc - rd.epoch_cyc) &
rd.sched_clock_mask, rd.mult, rd.shift);
+
+ /*
+ * If the architecture has a timestamp clock, and this is the
+ * first time we register a new sched_clock, use the timestamp
+ * clock as the epoch.
+ */
+ if (IS_ENABLED(CONFIG_ARCH_HAS_TIMESTAMP_CLOCK) &&
+ unlikely(cd.actual_read_sched_clock == jiffy_sched_clock_read))
+ ns = timestamp_clock();
+
cd.actual_read_sched_clock = read;
rd.read_sched_clock = read;
--
2.20.1
