runtime: don't adjust timer pp field in timerWaiting status
Before this CL, the following sequence was possible:
* GC scavenger starts and sets up scavenge.timer
* GC calls readyForScavenger, but sysmon is sleeping
* program calls runtime.GOMAXPROCS to shrink number of processors
* procresize destroys a P, the one that scavenge.timer is on
* (*pp).destroy calls moveTimers, which gets to the scavenger timer
* scavenger timer is timerWaiting, and moveTimers clears t.pp
* sysmon wakes up and calls wakeScavenger
* wakeScavengers calls stopTimer on scavenger.timer, still timerWaiting
* stopTimer calls deltimer which loads t.pp, which is still nil
* stopTimer tries to increment deletedTimers on nil t.pp, and crashes
The point of vulnerability is the time that t.pp is set to nil by
moveTimers and the time that t.pp is set to non-nil by moveTimers,
which is a few instructions at most. So it's not likely and in
particular is quite unlikely on x86. But with a more relaxed memory
model the area of vulnerability can be somewhat larger. This appears
to tbe the cause of two builder failures in a few months on linux-mips.
This CL fixes the problem by making moveTimers change the status from
timerWaiting to timerMoving while t.pp is clear. That will cause
deltimer to wait until the status is back to timerWaiting, at which
point t.pp has been set again.
Fixes #43712
Change-Id: I66838319ecfbf15be66c1fac88d9bd40e2295852
Reviewed-on: https://go-review.googlesource.com/c/go/+/284775
Trust: Ian Lance Taylor <iant@golang.org> Reviewed-by: Michael Knyszek <mknyszek@google.com> Reviewed-by: Michael Pratt <mpratt@google.com>