runtime/cgo: longjmp in CGO code causes a crash on Windows

[trink]

System Info

Go: 1.4.1
OS: Windows 7 64bit
Compiler gcc (tdm64-1) 4.7.1

This problem is not new to 1.4 but changes to the libraries being linked can cause the problem to appear/disappear.

Steps to reproduce

 git clone https://github.com/mozilla-services/heka
 cd heka
 git checkout --track origin/sandbox_refactor
 build.bat
 ctest -V -R sandbox

Output

2/3 Test #24: sandbox ..........................***Failed    0.84 sec
2015/02/02 07:15:08 payload_type: txt
payload_name:
payload: Hello World!
2015/02/02 07:15:08 payload_type: txt
payload_name:
payload: Hello World!
2015/02/02 07:15:08 payload_type: txt
payload_name:
payload: Hello World!
Exception 0xc0000005 0x1 0x3d1a0c 0x778e5bce
PC=0x778e5bce
signal arrived during cgo execution

github.com/mozilla-services/heka/sandbox/lua._Cfunc_process_message(0x2e6cc0, 0xc000000000)
        github.com/mozilla-services/heka/sandbox/lua/_obj/_cgo_gotypes.go:96 +0x4a
github.com/mozilla-services/heka/sandbox/lua.(*LuaSandbox).ProcessMessage(0xc08200c780, 0xc08200c500, 0x27)
        D:/work/heka/build/heka/src/github.com/mozilla-services/heka/sandbox/lua/lua_sandbox.go:618 +0x64
github.com/mozilla-services/heka/sandbox/lua_test.TestAPIErrors(0xc082058240)
        D:/work/heka/build/heka/src/github.com/mozilla-services/heka/sandbox/lua/lua_sandbox_test.go:330 +0x4e0
testing.tRunner(0xc082058240, 0x8ae9d8)
        d:/go/x64/src/testing/testing.go:447 +0xc6
created by testing.RunTests
        d:/go/x64/src/testing/testing.go:555 +0xa92

The failure occurs on line https://github.com/mozilla-services/heka/blob/sandbox_refactor/sandbox/lua/lua_sandbox_interface.c#L170 where luaL_error long jumps.

Observation: Adding a return statement to the line changes the behavior but it did not prove to be a viable workaround.

[ianlancetaylor]

Can you write a standalone test case?

I can imagine several ways that longjmp will fail horribly in a Go program, but without knowing what Lua is doing here I don't know whether it can be expected to work or not.

[trink]

Sadly, I haven't had any luck. The unit test I provided is now reproducing the problem consistently but even this code was running successfully with a slightly different verison of the lua.dll (some patches were appiled, but there were no modifications to the error handling routines)

[alexbrainman]

@trink please try https://go-review.googlesource.com/#/c/5711/ see if it helps your problem. I don't have all mingw tools installed to try it myself.

Alex

[trink]

Still seeing the same issue.

go version devel +ad3a99c Tue Feb 24 15:00:34 2015 +1100 windows/amd64

2/3 Test #24: sandbox ..........................***Failed    1.24 sec
2015/02/27 12:58:05 payload_type: txt
payload_name:
payload: Hello World!
2015/02/27 12:58:05 payload_type: txt
payload_name:
payload: Hello World!
2015/02/27 12:58:05 payload_type: txt
payload_name:
payload: Hello World!
Exception 0xc0000005 0x1 0x49eb2c 0x77a35bce
PC=0x77a35bce
signal arrived during cgo execution

github.com/mozilla-services/heka/sandbox/lua._Cfunc_process_message(0x3f6ce0, 0xc000000000)
        github.com/mozilla-services/heka/sandbox/lua/_obj/_cgo_gotypes.go:133 +0x43
github.com/mozilla-services/heka/sandbox/lua.(*LuaSandbox).ProcessMessage(0xc08200caf0, 0xc08200c870, 0x27)
        d:/work/heka/build/heka/src/github.com/mozilla-services/heka/sandbox/lua/lua_sandbox.go:618 +0x64
github.com/mozilla-services/heka/sandbox/lua_test.TestAPIErrors(0xc082056510)
        d:/work/heka/build/heka/src/github.com/mozilla-services/heka/sandbox/lua/lua_sandbox_test.go:330 +0x4e3
testing.tRunner(0xc082056510, 0x904218)
        d:/go/x64/src/testing/testing.go:448 +0xb4
created by testing.RunTests
        d:/go/x64/src/testing/testing.go:556 +0xa53

[alexbrainman]

Fair enough. Thanks for trying.

Alex

[rsc]

Does it also crash on say Linux or OS X, or just Windows?
If it's just Windows we should be able to do whatever we do to defend against this on those systems.

[rafrombrc]

Nope, not an issue on OSX or Linux, only Windows.

[alexbrainman]

@rafrombrc have you tried current tip? @minux made big changes to windows cgo recently.

Alex

[rafrombrc]

Last try was using code from here. I'll try tip and see i that makes a difference.

[alexbrainman]

Please try current tip. Thank you.

Alex

[minux]

If the external linking change fixes the problem then it's great.
However, if it's not, there are possibilities that the usage of
longjmp is just incompatible with cgo.

Without looking at the source code, I can't be sure whether the
longjmp usage is inherently incompatible with cgo.

On the plus side, as Lua's longjmp based error handling is portable,
if the problem doesn't exist on Linux, then I think the longjmp usage
should be fine and the external linking work should fix the windows
issue.

Please try again with the current tip. Thanks.
(the external linking code went in on Mar 23.)

[rafrombrc]

Okay, I've tried the current tip (go version devel +399b3e3 Fri Apr 10 22:51:46 2015 +0000 windows/amd64). Unfortunately, I can't tell whether or not it resolved the original issue, because now I'm hitting a compiler error, it's panicking while trying to compile the executable. I've verified that it's only happening on Windows; on Linux everything compiles and the tests all pass.