runtime: netpoll failed on android/386

[eliasnaur]

From the android/386 builder:

https://build.golang.org/log/e11446723639eb0b41a7268c1f19c0de5f8da7cc

runtime: epollwait on fd 6 failed with 22
fatal error: runtime: netpoll failed

runtime stack:
runtime.throw(0xea8e5d96, 0x17)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/runtime/panic.go:617 +0x69
runtime.netpoll(0xeab24a00, 0x1a5c7cf0)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/runtime/netpoll_epoll.go:75 +0x1f9
runtime.findrunnable(0xb9c23300, 0x0)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/runtime/proc.go:2226 +0x8bb
runtime.schedule()
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/runtime/proc.go:2525 +0x207
runtime.park_m(0xb9cc0000)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/runtime/proc.go:2605 +0x84
runtime.mcall(0xea78ab32)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/runtime/asm_386.s:345 +0x49

goroutine 1 [chan receive]:
testing.tRunner.func1(0xb9c98000)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:838 +0x1a2
testing.tRunner(0xb9c98000, 0xb9c3ff1c)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:866 +0xac
testing.runTests(0xb9c0e0a0, 0xeab10180, 0x23, 0x23, 0x0)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:1152 +0x286
testing.(*M).Run(0xb9c52380, 0x0)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:1069 +0x141
main.main()
	_testmain.go:112 +0x14d

goroutine 41 [runnable]:
testing.tRunner.func1(0xb9cd2dc0)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:857 +0x1fa
testing.tRunner(0xb9cd2dc0, 0xea9c93d8)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:866 +0xac
created by testing.(*T).Run
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:913 +0x2d1

goroutine 58 [runnable]:
testing.runTests.func1.1(0xb9c98000)
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:1159
created by testing.runTests.func1
	/private/var/folders/f6/d2bhfqss2716nxm8gkv1fmb80000gn/T/workdir-host-darwin-amd64-eliasnaur-android/go/src/testing/testing.go:1159 +0x84
exitcode=2go_android_exec: adb -e shell rm -rf /data/local/tmp/unix.test-17686
FAIL	cmd/vendor/golang.org/x/sys/unix	0.532s

[odeke-em]

Thank you for this report @eliasnaur!

Kindly paging @ianlancetaylor.

[ianlancetaylor]

That is the only instance I can find in the builder logs. It should be impossible.

[eliasnaur]

Another, again on android/386:

https://build.golang.org/log/11cd9309ba16cce696dc4763a7be1a433a590a7d

[eliasnaur]

Another, this time on android/arm64:

https://build.golang.org/log/ec6550708e7d674f80b19c3a987d0be24eb7dbe4

[eliasnaur]

FWIW, I can reproduce this crash locally by running the cmd/vendor/golang.org/x/sys/unix tests in a loop. I don't know how to debug it though.

[eliasnaur]

It could be a coincidence, but I once got the crash just as I was ctrl-C'ing the test loop that had run for minutes. If so, there might be a racy interaction between epoll_pwait and signals or during process exit.

[eliasnaur]

Edit: deleted confused analysis. runtime.netpollclose only deregisters a filedescriptor, it doesn't close the poll fd itself.

[gopherbot]

Change https://golang.org/cl/163638 mentions this issue: unix: don't overwrite unrelated file descriptors in TestDup

[gopherbot]

Change https://golang.org/cl/163623 mentions this issue: cmd/vendor/golang.org/x/sys: re-vendor