Re: fstat seems broken on sparc64

Re: fstat seems broken on sparc64

[Jan Engelhardt]

On Sunday 2010-01-10 22:11, David Miller wrote:
>
>> libc6   2.10.2-4
>
>Perfect, now you just have to audit what changed between the -2 and
>the -4 package to find the bug :-)

I really do not believe you when you say your glibc is not affected.
Maybe some things did not became clear as part of the discussion,
so let me quickly restate:

 - talking about -m64 / libc6-sparc64_2.10.2-*_sparc.deb only

 - syscall is screwed up. strace outputs:

C func  | i386      | x86_64     | sparc32   | sparc64
--------+-----------+------------+-----------+---------------
stat    | stat64    | stat       | stat64    | stat64
lstat   | lstat64   | lstat      | lstat64   | lstat64
fstatat | fstatat64 | newfstatat | fstatat64 | fstatat64
fstat   | fstat64   | fstat      | fstat64   | fstat (*)


 - I now found and fixed the problem, here is the patch for
   you to study (as code speaks louder)

Index: glibc-2.10.1/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
===================================================================
--- glibc-2.10.1.orig/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
+++ glibc-2.10.1/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
@@ -1 +1 @@
-#include "../../fxstat.c"
+#include "../../i386/fxstat.c"

(Hint: Compare with sparc64/lxstat.c)

Re: fstat seems broken on sparc64

[David Miller]

From: Jan Engelhardt <jengelh@medozas.de>
Date: Mon, 11 Jan 2010 00:15:17 +0100 (CET)

> On Sunday 2010-01-10 22:11, David Miller wrote:
>>
>>> libc6   2.10.2-4
>>
>>Perfect, now you just have to audit what changed between the -2 and
>>the -4 package to find the bug :-)
> 
> I really do not believe you when you say your glibc is not affected.

Why not :-)

davem@sunset:~$ dpkg-query -W | grep libc6-sparc64
libc6-sparc64	2.10.2-2
davem@sunset:~$ cat fstat_test.c
#include <sys/stat.h>
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>

int main(void)
{
	struct stat sb;
	int fd = open("foo", O_RDWR | O_CREAT, 0666);

	if (fd < 0)
		abort();
	if (fstat(fd, &sb) < 0)
		abort();
	printf("%lu.%lu\n", sb.st_mtim.tv_sec, sb.st_mtim.tv_nsec);
	if (stat("foo", &sb) < 0)
		abort();
	printf("%lu.%lu\n", sb.st_mtim.tv_sec, sb.st_mtim.tv_nsec);
	if (fstatat(AT_FDCWD, "foo", &sb, 0) < 0)
		abort();
	printf("%lu.%lu\n", sb.st_mtim.tv_sec, sb.st_mtim.tv_nsec);
	return 0;
}
davem@sunset:~$ gcc -m64 -o fstat_test fstat_test.c
davem@sunset:~$ ./fstat_test
1261510364.0
1261510364.0
1261510364.0
davem@sunset:~$ 

But, indeed when I run it under strace64 it goes:

open("foo", O_RDWR|O_CREAT, 0666)       = 3
fstat(3, {st_mode=S_IFREG|0644, st_size=0, ...}) = 0
fstat(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 2), ...}) = 0

I wonder why I get a sane output from the test program even though
it's not using fstat64.

>  - I now found and fixed the problem, here is the patch for
>    you to study (as code speaks louder)
> 
> Index: glibc-2.10.1/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
> ===================================================================
> --- glibc-2.10.1.orig/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
> +++ glibc-2.10.1/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
> @@ -1 +1 @@
> -#include "../../fxstat.c"
> +#include "../../i386/fxstat.c"
> 
> (Hint: Compare with sparc64/lxstat.c

Thanks a lot!

Please submit this to the glibc/eglibc folks for inclusion!

Re: fstat seems broken on sparc64

[Jan Engelhardt]

On Monday 2010-01-11 07:42, David Miller wrote:
>> On Sunday 2010-01-10 22:11, David Miller wrote:
>>>
>>>> libc6   2.10.2-4
>>>
>>>Perfect, now you just have to audit what changed between the -2 and
>>>the -4 package to find the bug :-)
>> 
>> I really do not believe you when you say your glibc is not affected.
>
>Why not :-)
>
>davem@sunset:~$ dpkg-query -W | grep libc6-sparc64
>libc6-sparc64	2.10.2-2
>davem@sunset:~$ cat fstat_test.c
>#include <sys/stat.h>
>#include <fcntl.h>
>#include <stdio.h>
>#include <stdlib.h>
>
>int main(void)
>{
>	struct stat sb;
>	int fd = open("foo", O_RDWR | O_CREAT, 0666);
>
>	if (fd < 0)
>		abort();
>	if (fstat(fd, &sb) < 0)
>		abort();
>	printf("%lu.%lu\n", sb.st_mtim.tv_sec, sb.st_mtim.tv_nsec);
>	if (stat("foo", &sb) < 0)
>		abort();
>	printf("%lu.%lu\n", sb.st_mtim.tv_sec, sb.st_mtim.tv_nsec);
>	if (fstatat(AT_FDCWD, "foo", &sb, 0) < 0)
>		abort();
>	printf("%lu.%lu\n", sb.st_mtim.tv_sec, sb.st_mtim.tv_nsec);
>	return 0;
>}
>davem@sunset:~$ gcc -m64 -o fstat_test fstat_test.c
>davem@sunset:~$ ./fstat_test
>1261510364.0
>1261510364.0
>1261510364.0
>davem@sunset:~$ 
>But, indeed when I run it under strace64 it goes:
>
>open("foo", O_RDWR|O_CREAT, 0666)       = 3
>fstat(3, {st_mode=S_IFREG|0644, st_size=0, ...}) = 0
>fstat(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 2), ...}) = 0
>
>I wonder why I get a sane output from the test program even though
>it's not using fstat64.

I would recommend that you run the testcase on a filesystem that
supports and records subsecond timestamps, and that the file is
created at a moment where the subsecond part is not 0. When in doubt,
use touch(1) to apply some subseconds, but usually, just running it
will create it at a non-zero subsecond with a high probability.

>> ===================================================================
>> --- glibc-2.10.1.orig/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
>> +++ glibc-2.10.1/sysdeps/unix/sysv/linux/sparc/sparc64/fxstat.c
>> @@ -1 +1 @@
>> -#include "../../fxstat.c"
>> +#include "../../i386/fxstat.c"
>> 
>> (Hint: Compare with sparc64/lxstat.c
>
>Thanks a lot!
>
>Please submit this to the glibc/eglibc folks for inclusion!
>
http://sources.redhat.com/bugzilla/show_bug.cgi?id=11155