runtime: fix interactions between synctest, race detector, and timers
When an AfterFunc executes in a synctest bubble, there is a series of
happens-before relationships:
1. The AfterFunc is created.
2. The AfterFunc goroutine executes.
3. The AfterFunc goroutine returns.
4. A subsequent synctest.Wait call returns.
We were failing to correctly establish the happens-before relationship
between the AfterFunc goroutine and the AfterFunc itself being created.
When an AfterFunc executes, the G running the timer temporarily switches
to the timer heap's racectx. It then calls time.goFunc, which starts a
new goroutine to execute the timer. time.goFunc relies on the new goroutine
inheriting the racectx of the G running the timer.
Normal, non-synctest timers, execute with m.curg == nil, which causes
new goroutines to inherit the g0 racectx. We were running synctest
timers with m.curg set (to the G executing synctest.Run), so the new
AfterFunc goroutine was created using m.curg's racectx. This resulted
in us not properly establishing the happens-before relationship between
AfterFunc being called and the AfterFunc goroutine starting.
Fix this by setting m.curg to nil while executing timers.
As one additional fix, when waking a blocked bubble, wake the root
goroutine rather than a goroutine blocked in Wait if there is a
timer that can fire.
Fixes #72750
Change-Id: I2b2d6b0f17f64649409adc93c2603f720494af89
Reviewed-on: https://go-review.googlesource.com/c/go/+/658595
Auto-Submit: Damien Neil <dneil@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Michael Pratt <mpratt@google.com>