Skip to contentSkip to conversation

timekeeping: Update multiplier when NTP frequency is set directly

RFC PATCH In review Miroslav Lichvar 2 participants 3 messages May 23, 2018 – May 24, 2018 Timekeeping, clocksource core, NTP, alarmtimer

lore
Share & tools
lore

Miroslav Lichvar

author reviewer
#0 lore raw

RFC PATCH 1/1 timekeeping: Update multiplier when NTP frequency is set directly

When the NTP frequency is set directly from userspace using the ADJ_FREQUENCY or ADJ_TICK timex mode, immediately update the timekeeper's multiplier instead of waiting for the next tick.

This removes a hidden non-deterministic delay in setting of the frequency and allows an extremely tight control of the system clock with update rates close to or even exceeding the kernel HZ.

Cc Thomas Gleixner <[email protected]>
Cc John Stultz <[email protected]>
Cc Richard Cochran <[email protected]>
Cc Prarit Bhargava <[email protected]>
---
 kernel/time/timekeeping.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/time/timekeeping.c +4 -4
index 49cbceef5deb..6922dae7317c 100644
@@ -2047,10 +2047,6 @@ void update_wall_time(void)
 				   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 #endif
 
-	/* Check if there's really nothing to do */
-	if (offset < real_tk->cycle_interval)
-		goto out;
-
 	/* Do some additional sanity checking */
 	timekeeping_check_update(tk, offset);
 
@@ -2332,6 +2328,10 @@ int do_adjtimex(struct timex *txc)
 	write_seqcount_end(&tk_core.seq);
 	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
 
+	/* Update the multiplier immediately if frequency was set directly */
+	if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
+		update_wall_time();
+
 	if (tai != orig_tai)
 		clock_was_set();
 
Signature · 1 line
2.14.3
#1 lore raw
Quoted 7 lines · Miroslav Lichvar

When the NTP frequency is set directly from userspace using the ADJ_FREQUENCY or ADJ_TICK timex mode, immediately update the timekeeper's multiplier instead of waiting for the next tick.

This removes a hidden non-deterministic delay in setting of the frequency and allows an extremely tight control of the system clock with update rates close to or even exceeding the kernel HZ.

Thanks for the patch!

I feel like we tried this years back, but had to revert it. But its been awhile. Am I confusing things?

kernel/time/timekeeping.c void update_wall_time(void) · line 2053

Quoted 12 lines
kernel/time/timekeeping.c +0 -4
index 49cbceef5deb..6922dae7317c 100644
@@ -2047,10 +2047,6 @@ void update_wall_time(void)
                                   tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
 #endif
 
-       /* Check if there's really nothing to do */
-       if (offset < real_tk->cycle_interval)
-               goto out;
-

Apologies again, as I don't have a lot of context here these days, but
this could mean we end up doing unnecessary work on every
update_wall_time, no?

Would a "force" flag be better to pass to update_wall_time() to only avoid the short-cut in the non-adjtimex case?

thanks
-john

author reviewer
#2 lore raw
Quoted 7 lines · John Stultz
Quoted 3 lines · Miroslav Lichvar

This removes a hidden non-deterministic delay in setting of the frequency and allows an extremely tight control of the system clock with update rates close to or even exceeding the kernel HZ.

I feel like we tried this years back, but had to revert it. But its been awhile. Am I confusing things?

IIRC we only talked about doing this. Before the recent changes in timekeeping, namely c2cda2a5 (Don't align NTP frequency adjustments to ticks), it wouldn't make much sense. There would still be a hidden delay, it would just be negative (from the applications point of view).

Quoted 8 lines

kernel/time/timekeeping.c void update_wall_time(void) · line 2053

Quoted 4 lines
kernel/time/timekeeping.c +0 -4
-       /* Check if there's really nothing to do */
-       if (offset < real_tk->cycle_interval)
-               goto out;
-

Apologies again, as I don't have a lot of context here these days, but
this could mean we end up doing unnecessary work on every
update_wall_time, no?

I'm not sure. If I understand it correctly, update_wall_time() is normally not called more than once per tick. Only when an update is late and it happens to include the next tick, the subsequent call might be unnecessary, right?

Quoted 2 lines

Would a "force" flag be better to pass to update_wall_time() to only avoid the short-cut in the non-adjtimex case?

Yes, that makes sense.

Thanks,

Signature · 1 line
Miroslav Lichvar