1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
|
risc-v: 0.956
user-level: 0.949
vnc: 0.937
permissions: 0.936
ppc: 0.934
i386: 0.931
performance: 0.928
arm: 0.926
debug: 0.923
KVM: 0.921
peripherals: 0.917
network: 0.917
register: 0.916
graphic: 0.914
semantic: 0.914
architecture: 0.914
virtual: 0.914
TCG: 0.913
hypervisor: 0.912
x86: 0.911
socket: 0.909
device: 0.908
assembly: 0.908
VMM: 0.905
kernel: 0.897
boot: 0.890
PID: 0.888
mistranslation: 0.871
files: 0.860
Go programs crash on qemu-sh4 due to issues with atomics
After #1738545 [1] was fixed, Go applications work fine on qemu-arm but still crash on qemu-sh4. From the backtrace, it looks like an issue with the atomics in qemu-sh4:
(sid-sh4-sbuild)root@epyc:/# cat hello.go
package main
import "fmt"
func main() {
fmt.Println("hello world")
}
(sid-sh4-sbuild)root@epyc:/# gccgo-9 hello.go -o hello
(sid-sh4-sbuild)root@epyc:/# ./hello
panic: ( runtime runtime.errorString) (0x7f74527c,0x80a038)
fatal error: panic on system stack
panic: ( runtime runtime.errorString) (0x7f74527c,0x80a038)
fatal error: panic on system stack
runtime stack:
runtime..z2finternal..z2fatomic.Load64
../../../src/libgo/go/runtime/internal/atomic/atomic.c:37
runtime_mstart
../../../src/libgo/runtime/proc.c:596
goroutine 1 [running]:
goroutine running on other thread; stack unavailable
runtime stack:
runtime..z2finternal..z2fatomic.Load64
../../../src/libgo/go/runtime/internal/atomic/atomic.c:37
runtime_mstart
../../../src/libgo/runtime/proc.c:596
(sid-sh4-sbuild)root@epyc:/#
The same sample Go program runs fine on my SH7785LCR SH4 evaluation board:
root@tirpitz:~> uname -a
Linux tirpitz 3.16.7-ckt7 #8 PREEMPT Fri Oct 21 18:47:41 CEST 2016 sh4a GNU/Linux
root@tirpitz:~> cat hello.go
package main
import "fmt"
func main() {
fmt.Println("hello world")
}
root@tirpitz:~> gccgo-9 hello.go -o hello
root@tirpitz:~> ./hello
hello world
root@tirpitz:~>
Please note: In order to be able to reproduce this, one also needs to revert commit 61dedf2af7 [2], otherwise the Go application crashes differently:
(sid-sh4-sbuild)root@epyc:/# ./hello
Unhandled trap: 0x180
pc=0x7e5f7f9e sr=0x00000000 pr=0x7ee3d582 fpscr=0x00080004
spc=0x00000000 ssr=0x00000000 gbr=0x7e590480 vbr=0x00000000
sgr=0x00000000 dbr=0x00000000 delayed_pc=0x7e5f7f60 fpul=0x00034f3b
r0=0x008007d4 r1=0x00000000 r2=0xfffe0b2a r3=0x00000002
r4=0x008006e4 r5=0x00872000 r6=0x00200000 r7=0x00000000
r8=0x7f7bca7c r9=0x7fffebd4 r10=0x00800480 r11=0x7f7bc0f0
r12=0x7f7a3fa4 r13=0x008004c0 r14=0x7f7b2238 r15=0x7fffebd0
r16=0x00000000 r17=0x00000000 r18=0x00000000 r19=0x00000000
r20=0x00000000 r21=0x00000000 r22=0x00000000 r23=0x00000000
(sid-sh4-sbuild)root@epyc:/#
> [1] https://bugs.launchpad.net/bugs/1738545
> [2] https://bugs.launchpad.net/bugs/1796520
The immediate cause of the crash here is that the runtime invokes the Load64() function on a pointer that isn't 8-aligned, which triggers a hand-coded check-and-panic in the libgo code. I haven't yet tracked down where that pointer came from.
The non-8-aligned pointer is the runtime.work.empty field. The compilation that I have of this binary has put the 'runtime.work' struct at 0x6bfadc, which is only 4-aligned, and this won't work as the lfstack fields it starts with are supposed to be 8-aligned. So it looks to me like the compiler has miscompiled the binary somehow, and QEMU's actual execution of it is OK.
I don't know if this is a general bug in the sh4 gccgo support (in which case we must be succeeding on the real hardware by accident, probably by finishing fast enough that the gc never kicks in), or if QEMU is mis-executing the compiler somehow and a build done on the real hardware puts the work struct at an 8-aligned address.
I just did an objdump -x of the /usr/lib/sh4-linux-gnu/libgo.so.14, which will be the shipped version from the Debian package, and in the section header it has:
24 .bss 000191f8 00fe74ec 00fe74ec 00fd74ec 2**2
ALLOC
and in the symbol table it has:
00ff98f4 l O .bss 00000104 runtime.work
So the compiler has put the 'runtime.work' struct at a non-multiple-of-8 offset into the bss, and it's given the BSS alignment requirements that are only 4-aligned, not 8-aligned. That means it's random luck whether the struct gets 8-aligned or not.
This looks to me like it's a bug in the sh4 gccgo -- https://go101.org/article/memory-layout.html says that the first word in a struct or variable is supposed to be guaranteed to be 8-aligned, so the compiler needs to align things more strictly than it is currently doing.
Thanks. I will report this to Go/gcc upstream.
I'm going to close this bug, since it's a problem with the gccgo codegen, not a QEMU bug. (I'm interested in any response you get from the go/gcc upstream folks, though.)
|