runtime: bad P in runqempty called from checkRunqsNoP

[mknyszek]

In trying to reproduce #47302 I wrote a little program. I failed to reproduce it, but instead I found a new failure.

The program is:

package main

import (
	"runtime"
	"time"
)

var sink []byte
var sink2 []byte

func main() {
	go func() {
		for {
			sink = make([]byte, 1024)
		}
	}()
	go func() {
		for {
			sink2 = make([]byte, 1024)
		}
	}()
	for i := 0; i < 5; i++ {
		time.Sleep(100 * time.Millisecond)
		runtime.GOMAXPROCS(100)
		time.Sleep(100 * time.Millisecond)
		runtime.GOMAXPROCS(4)
		time.Sleep(100 * time.Millisecond)
		runtime.GOMAXPROCS(6)
	}
}

The failure is:

fatal error: unexpected signal during runtime execution
[signal SIGSEGV: segmentation violation code=0x1 addr=0x5f0 pc=0x434824]

runtime stack:
runtime.throw({0x4696a7, 0x100000081})
	/usr/local/google/home/mknyszek/toolchain/src/runtime/panic.go:1198 +0x71
runtime.sigpanic()
	/usr/local/google/home/mknyszek/toolchain/src/runtime/signal_unix.go:719 +0x396
runtime.runqempty(...)
	/usr/local/google/home/mknyszek/toolchain/src/runtime/proc.go:5922
runtime.checkRunqsNoP({0xc000026000, 0x180, 0x4}, {0xc000018058, 0x7ffcc3668c48, 0x41e725})
	/usr/local/google/home/mknyszek/toolchain/src/runtime/proc.go:3091 +0x84
runtime.findrunnable()
	/usr/local/google/home/mknyszek/toolchain/src/runtime/proc.go:2891 +0x425
runtime.schedule()
	/usr/local/google/home/mknyszek/toolchain/src/runtime/proc.go:3367 +0x239
runtime.park_m(0xc000001380)
	/usr/local/google/home/mknyszek/toolchain/src/runtime/proc.go:3516 +0x14d
runtime.mcall()
	/usr/local/google/home/mknyszek/toolchain/src/runtime/asm_amd64.s:307 +0x43

goroutine 1 [sleep]:
time.Sleep(0x5f5e100)
	/usr/local/google/home/mknyszek/toolchain/src/runtime/time.go:193 +0x12e
main.main()
	/usr/local/google/home/mknyszek/tmp/gmp.go:25 +0x5b

goroutine 5 [runnable]:
main.main.func1()
	/usr/local/google/home/mknyszek/tmp/gmp.go:14 +0x28
created by main.main
	/usr/local/google/home/mknyszek/tmp/gmp.go:12 +0x26

goroutine 6 [runnable]:
main.main.func2()
	/usr/local/google/home/mknyszek/tmp/gmp.go:19 +0x28
created by main.main
	/usr/local/google/home/mknyszek/tmp/gmp.go:17 +0x34

I can reproduce this at tip, and going back about a month. It takes about 90 minutes (sometimes more, sometimes less) to reproduce when running under https://pkg.go.dev/golang.org/x/tools/cmd/stress. I don't think this is a release blocker because we've never seen it out in the wild, and this program is very unrealistic. It may also reproduce with older releases, I'm not sure.

[mknyszek]

CC @prattmic

[gopherbot]

Change https://golang.org/cl/336449 mentions this issue: runtime: move mem profile sampling into m-acquired section

[gopherbot]

Change https://golang.org/cl/359796 mentions this issue: runtime: add always-preempt maymorestack hook