From 69e386c2792b211bffcb3ed70bab91af6347a40e Mon Sep 17 00:00:00 2001
From: JR Lanteigne <jrlanteigne@gmail.com>
Date: Sat, 15 Aug 2026 20:56:48 -0300
Subject: [PATCH] runtime: recompute timer channel send delay at send time

The timer struct documents the delay argument as nanotime() - t.when.
time.sendTime uses it to reconstruct the delivered value as
Now().Add(-delta). unlockAndRun computes delta from the now cached by
the caller. The heap update, the sendLock acquisition, other timers
due in the same timers.run pass, and preemption of the timer-running M
all happen after that clock read. Their duration lands in the
delivered value as one-sided, load-dependent error. The common case is
a few hundred nanoseconds. Under scheduler or interrupt pressure the
error reaches several microseconds.

Consumers that bin ticker values against a wall-clock grid see the
outliers as phantom missed ticks. A value that crosses a period
boundary makes floor arithmetic count two elapsed periods where one
elapsed.

Recompute delta immediately before the send. This matches its
documented definition at the moment of use, and the reconstruction
then recovers the fire time to within two adjacent clock reads.
Measured on linux/amd64 (i5-8500, idle), the mean error of 1s ticker
values against the scheduled grid drops from ~600ns with one-sided
tails beyond 1.4us to 0ns +/- 0.5us with no observed tails. The cost
is one nanotime call per channel-timer firing. Synctest bubbles keep
their fake-clock delta.

Fixes #80893
---
 src/runtime/time.go | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/src/runtime/time.go b/src/runtime/time.go
index daf20cf4331197..477b342d2f58d7 100644
--- a/src/runtime/time.go
+++ b/src/runtime/time.go
@@ -1202,6 +1202,9 @@ func (t *timer) unlockAndRun(now int64, bubble *synctestBubble) {
 		}
 	}
 
+	if t.isChan && bubble == nil {
+		delay += nanotime() - now
+	}
 	f(arg, seq, delay)
 
 	if t.isChan {