runtime: qemu-arm fails to emulate go user programs

[arges]

qemu-arm fails to emulate go user programs due to how qemu handles certain signals.

The original issue/diff that seems to solve the issue is here:
https://codereview.appspot.com/124900043/diff/60001/src/pkg/runtime/os_linux.c

The qemu code is here:
http://git.qemu.org/?p=qemu.git;a=blob;f=linux-user/signal.c;h=1141054be2170128d6f7a340b41484b49a255936;hb=HEAD#l82

Test Case:

Build this on an non-arm platform such as x86.

test.go:

package main

import "fmt"

func main() {
 fmt.Println("Hi")
}

1. GOARCH=arm go build test.go
2. qemu-arm test

This will fail with the following error:

qemu-arm test
fatal error: rt_sigaction failure

runtime stack:
runtime.throw(0x106d68, 0x14)
    /usr/lib/go/src/runtime/panic.go:527 +0x78
runtime.setsig(0x40, 0x6f8b0, 0x1)
    /usr/lib/go/src/runtime/os1_linux.go:297 +0x148
runtime.initsig()
    /usr/lib/go/src/runtime/signal1_unix.go:67 +0x180
runtime.mstart1()
    /usr/lib/go/src/runtime/proc1.go:717 +0xd4
runtime.mstart()
    /usr/lib/go/src/runtime/proc1.go:691 +0x84

goroutine 1 [runnable]:
runtime.main()
    /usr/lib/go/src/runtime/proc.go:28
runtime.goexit()
    /usr/lib/go/src/runtime/asm_arm.s:1036 +0x4

Testing with the aforementioned patch solves the issue, but I'm unsure its the best approach to solving this. I can re-propose that patch if there isn't a better solution, but this would be a very useful feature for using qemu user space emulation with go.

--chris

[rsc]

/cc @minux who wrote CL 124900043.

I'm not sure how/why that code got lost/reverted. It seems fine to put in again.

[gopherbot]

CL https://golang.org/cl/16853 mentions this issue.

[davecheney]

Could this be backported to 1.5.2 ?

[minux]

Probably not critical enough, we never supported running binaries under qemu user emulation before.

[rsc]

@davecheney, if you can make a case for why it's important, it's certainly simple enough that I am OK with putting it in 1.5.2.

[davecheney]

Thanks @rsc. Canonical have gone through a lot of internal bureaucratic pain to get up to Go 1.5, if this patch could make it into 1.5.2 that would be much appreciated.

[streamnsight]

👍 for the patch
/cc @minux
Can someone tell which go version had the original patch? or when 1.5.2 is scheduled to be released?
I'm having the same problem trying to compile code in qemu

thanks

[rsc]

No released Go version has had this patch, as far as I understand. Go 1.5.2 should be out by the end of the calendar year, probably sooner.

[streamnsight]

@rsc thanks
I thought I understood the patch was there but was 'reversed' at some point. the very first link mentioned by @arges seems to be it, but there is no clear indication of what the version was at the time.

[mwhudson]

AIUI the fix was committed to the C version of the runtime on the dev.power64, but I don't think it ever got to master. Perhaps it got lost as part of the transition of the runtime to pure go? Doesn't really matter now anyway.

[gopherbot]

CL https://golang.org/cl/16986 mentions this issue.

[arges]

@streamnsight FWIW I originally reproduced this issue in go1.5.1

[luxas]

I compiled a binary on ARM (etcd) with latest go1.5.2.
However, now this issue came up instead (on amd64):

# On host machine
$ qemu-arm-static etcd
Segmentation fault (core dumped)

# Running in docker
$ etcd
/build/buildd/qemu-2.0.0+dfsg/tcg/optimize.c:362: tcg fatal error
Aborted (core dumped)

$ etcd
/build/buildd/qemu-2.0.0+dfsg/tcg/tcg.c:1693: tcg fatal error
Aborted (core dumped)

I'm using https://github.com/moul/docker-binfmt-register for running ARM docker images on amd64
I'll try to reproduce what @arges did

[davecheney]

@luxas this is probably a different issue. Please open a new issue, but note that qemu is not reliable when running go programs. qemu's emulation is too glibc specific.

[luxas]

Yep, I'm fine with creating a new issue. Just said it here first, so you guys know about it.

[minux]

Yeah, I think we lost the ability to run Go binaries with qemu-arm long time ago. tcg fatal error also seems to suggest that it's a qemu bug. (the arm port used to be developed with qemu-arm, but then as the port is able to run on real hardwares, we no longer test on qemu-arm.) The signal patch is originally intended for qemu-ppc64, btw. I doubt that even qemu-ppc64 is able to run Go binaries, other than the most trivial one. As beefy arm boards are widely available, I don't think we should invest into qemu usermode emulation any more. It's slower than the more powerful arm boards, and it doesn't support multithreaded programs well.