x/mobile: gomobile's JNI crashes under android x86 with ARM translation

[wxiaoguang]

1. What version of Go are you using (go version)?
   go version go1.6.2 darwin/amd64
2. What operating system and processor architecture are you using (go env)?
   Host: x86_64

GOHOSTARCH="amd64"
GOHOSTOS="darwin"
GOOS="darwin"
GOGCCFLAGS="-fPIC -m64 -pthread -fno-caret-diagnostics -Qunused-arguments -fmessage-length=0 -fno-common"
CXX="clang++"

Target:
genymotion android 4.2/4.3(x86) with arm translation 1.1

1. What did you do?

https://play.golang.org/p/zUeejSnLmO

armeabi-v7a under android x86 with arm translation, the CompareAndSwapUint64 crashes like this:

panic: runtime error: invalid memory address or nil pointer dereference
[signal 0xb code=0x1 addr=0x0 pc=0x38273e98]

I only want to build armeabi-v7a JNIs to reduce my APK's size.
The crash only occurs under x86 with arm translation.
The code under armeabi-v7a and x86_64 with arm translation works correctly.

I made a very dirty hack to make the code work:

    1. clean gomobile compiled packages:  rm -rf ~/go/pkg/gomobile/pkg_*
    2. modify /usr/local/go/src/sync/atomic/asm_linux_arm.s
       comment out "__kuser_cmpxchg64"
       force arm arch to 7: MOVW     $7, R0

TEXT setupAndCallCAS64<>(SB),NOSPLIT,$-4-21
    //MOVB  runtime·armArch(SB), R0
    MOVW    $7, R0
    // LDREXD, STREXD only present on ARMv6K or higher
    CMP $6, R0
    MOVW.CS $·armCompareAndSwapUint64(SB), R1
    MOVW.CS R1, armCAS64(SB)
    MOVW.CS R1, R15

    3. re-init gomobile:  gomobile init
    4. re-compile library & app

[hyangah]

Can you include the crash stack trace, and the output of 'gomobile bind -x -v'?

Is the problem present with go1.7 (or tip) as well?

[wxiaoguang]

Attached files are about the bind command and the panic stack trace.

gomobile-bind.txt.zip
gomobile-panic.txt

The crash occurred in the net module, I debugged for a while and found that it was caused by CompareAndSwapUint64.

I am pretty sure that the problem is caused by using golang's linux-kernel atomic CAS64(__kuser_cmpxchg64) under a android x86 with ARM translation.

You can refer to my "dirty solution", I just removed the code for "__kuser_cmpxchg64" in /usr/local/go/src/sync/atomic/asm_linux_arm.s, and force golang to use "armCompareAndSwapUint64", then everything works.

I will try go 1.7 later, but I don't think it works if it still uses the same code for "__kuser_cmpxchg64".

Maybe this problem is not very serious, because seldom people do what I do: unpack the aar and remove everything(x86/x86_64) else besides armeabi-v7a to keep the APK as small as possible :)

[wxiaoguang]

I have some new thoughts about this problem.

Since gomobile only supports armeabi-v7a and above, can we make a special CompareAndSwapUint64 for armeabi-v7a and above? Then this special implementation can call armCompareAndSwapUint64 directly which is efficient enough. I read the source code about __kuser_cmpxchg64 and thought that it's not necessary to use it under an arm-v7a and above CPU.

[hyangah]

cc @crawshaw

@wxiaoguang do you see the same problem when running it on real arm devices?
btw, you can configure -target flag to limit the target architecture. e.g. -target=android/arm

[wxiaoguang]

Real arm devices works OK. This problem only occurs under x86(i386 only) with arm translation.

I have made a cleaner fix, just change the detection order:

/src/sync/atomic/asm_linux_arm.s

TEXT setupAndCallCAS64<>(SB),NOSPLIT,$-4-21
    //1. try to use armCAS64(LDREXD, STREXD)
    MOVB    runtime·armArch(SB), R0
    // LDREXD, STREXD only present on ARMv6K or higher
    CMP $6, R0 // TODO(minux): how to differentiate ARMv6 with ARMv6K?
    MOVW.CS $·armCompareAndSwapUint64(SB), R1
    MOVW.CS R1, armCAS64(SB)
    MOVW.CS R1, R15
    //2. try to use __kuser_cmpxchg64
    MOVW    $0xffff0ffc, R0 // __kuser_helper_version
    MOVW    (R0), R0
    // __kuser_cmpxchg64 only present if helper version >= 5
    CMP     $5, R0
    MOVW.CS $kernelCAS64<>(SB), R1
    MOVW.CS R1, armCAS64(SB)
    MOVW.CS R1, R15 // R15 = hardware PC
    //3. we are out of luck, can only use runtime's emulated 64-bit cas
    MOVW    $·generalCAS64(SB), R1
    MOVW    R1, armCAS64(SB)
    MOVW    R1, R15