[PATCH 2/2] unwind: relax the pass condition of strace-k.test test case
Masatake YAMATO
yamato at redhat.com
Wed Jun 11 18:44:54 UTC 2014
strace-k.test is failed on i386 GNU/Linux. I inspect the reason and it
seems that libunwind does not have enough ability to unwinding the
stack when pc is at [vdso]. However, I think stacktrace feature is
still useful even on i386 GNU/Linux. So this patch relaxes the pass
condition of strace-k.test test case.
Background:
It seems that call convention at vdso on i386 linux is different
from explanations on widely available texts.
Generally $ebp + 4 points return address for the current frame.
libunwind may use this convention for unwinding.
However, this convention is not applicable at vdso on i386 linux.
Gdb may know the convention used in the vdso.
I run a program extended from stack-fcall.c under gdb.
Call sequence of extended stack-fcall.c is
main -> f0 -> f1 -> f2 -> getpid -> sysenter@[vdso].
Following gdb session is stopped just after sysenter@[vdso].
With where command of gdb, the call sequence is printed.
However, with unwinding manually, in other words tracking
$ebp, the addresses of f2 and getpid are not found.
f1, f0 and main can be unwined well.
---- GDB SESSION ----
(gdb) stepi
0xb7fff424 in __kernel_vsyscall ()
(gdb) disass
Dump of assembler code for function __kernel_vsyscall:
0xb7fff414 <+0>: push %ecx
0xb7fff415 <+1>: push %edx
0xb7fff416 <+2>: push %ebp
0xb7fff417 <+3>: mov %esp,%ebp
0xb7fff419 <+5>: sysenter
0xb7fff41b <+7>: nop
0xb7fff41c <+8>: nop
0xb7fff41d <+9>: nop
0xb7fff41e <+10>: nop
0xb7fff41f <+11>: nop
0xb7fff420 <+12>: nop
0xb7fff421 <+13>: nop
0xb7fff422 <+14>: int $0x80
=> 0xb7fff424 <+16>: pop %ebp
0xb7fff425 <+17>: pop %edx
0xb7fff426 <+18>: pop %ecx
0xb7fff427 <+19>: ret
End of assembler dump.
(gdb) print/x *(void **)($ebp + 4)
$57 = 0x8048457
(gdb) where
#0 0xb7fff424 in __kernel_vsyscall ()
#1 0x47636fde in getpid () from /lib/libc.so.6
#2 0x0804843d in f2 (i=1) at stack-fcall.c:9
#3 0x08048457 in f1 (i=1) at stack-fcall.c:15
#4 0x08048471 in f0 (i=1) at stack-fcall.c:21
#5 0x0804848e in main (argc=1, argv=0xbffff534) at stack-fcall.c:27
---- GDB SESSION ----
How the test case is relaxed:
Though two frames on the stack are lost, I think stacktrace feature
is useful. So I decide to relax the pass condition of test case.
* tests/stack-fcall.c (f2): New function.
(f1): Call f2. Added one more function call.
* tests/strace-k.test: check the output of test
target with matching two patterns. The test is passed
when the output matches one of them.
One is "getpid f2 f1 f0 main". This is strict pattern we
used before. "f1 f0 main" is the new relaxed pattern.
Reference:
https://www.kernel.org/pub/linux/kernel/people/akpm/patches/2.5/2.5.67/2.5.67-mm3/broken-out/vsyscall-unwinding.patch
This patch may explains the call convention in vdso of i386.
Signed-off-by: Masatake YAMATO <yamato at redhat.com>
---
tests/stack-fcall.c | 8 +++++++-
tests/strace-k.test | 7 +++++--
2 files changed, 12 insertions(+), 3 deletions(-)
diff --git a/tests/stack-fcall.c b/tests/stack-fcall.c
index 1c66fef..b49014a 100644
--- a/tests/stack-fcall.c
+++ b/tests/stack-fcall.c
@@ -3,12 +3,18 @@
/* Use "volatile" to avoid compiler optimization. */
-int f1(int i)
+int f2(int i)
{
static pid_t (* volatile g)(void) = getpid;
return g() + i;
}
+int f1(int i)
+{
+ static int (* volatile g)(int) = f2;
+ return g(i) + i;
+}
+
int f0(volatile int i)
{
static int (* volatile g)(int) = f1;
diff --git a/tests/strace-k.test b/tests/strace-k.test
index 3845c8c..eae0a71 100755
--- a/tests/strace-k.test
+++ b/tests/strace-k.test
@@ -23,11 +23,14 @@ $STRACE $args > $LOG 2>&1 || {
fail_ "$STRACE $args failed"
}
-expected='getpid f1 f0 main '
+expected_x86_64='getpid f2 f1 f0 main '
+expected_i386='f1 f0 main '
result=$(sed -n '1,/(main+0x[a-f0-9]\+) .*/ s/^.*(\([^+]\+\)+0x[a-f0-9]\+) .*/\1/p' $LOG |
tr '\n' ' ')
-test "$result" = "$expected" || {
+test "$result" = "$expected_x86_64" ||
+test "$result" = "$expected_i386" ||
+{
cat $LOG
echo "expected: \"$expected\""
echo "result: \"$result\""
--
1.9.0
More information about the Strace-devel
mailing list