2009-06-25 10:11:53

by Heiko Carstens

[permalink] [raw]
Subject: [PATCH] gcov: fix __ctors_start alignment

From: Heiko Carstens <[email protected]>

The ctors section for each object file is eight byte aligned (on 64 bit).
However the __ctors_start symbol starts at an arbitrary address dependent
on the size of the previous sections.
Therefore the linker may add some zeroes after __ctors_start to make sure
the ctors contents are properly aligned. However the extra zeroes at the
beginning aren't expected by the code. When walking the functions pointers
contained in there and extra zeroes are added this may result in random
jumps.
So make sure that the __ctors_start symbol is always aligned as well.

Fixes this crash on an allyesconfig on s390:

[ 0.582482] Kernel BUG at 0000000000000012 [verbose debug info unavailable]
[ 0.582489] illegal operation: 0001 [#1] SMP DEBUG_PAGEALLOC
[ 0.582496] Modules linked in:
[ 0.582501] CPU: 0 Tainted: G W 2.6.31-rc1-dirty #273
[ 0.582506] Process swapper (pid: 1, task: 000000003f218000, ksp: 000000003f2238e8)
[ 0.582510] Krnl PSW : 0704200180000000 0000000000000012 (0x12)
[ 0.582518] R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:0 CC:2 PM:0 EA:3
[ 0.582524] Krnl GPRS: 0000000000036727 0000000000000010 0000000000000001 0000000000000001
[ 0.582529] 00000000001dfefa 0000000000000000 0000000000000000 0000000000000040
[ 0.582534] 0000000001fff0f0 0000000001790628 0000000002296048 0000000002296048
[ 0.582540] 00000000020c438e 0000000001786000 0000000002014a66 000000003f223e60
[ 0.582553] Krnl Code:>0000000000000012: 0000 unknown
[ 0.582559] 0000000000000014: 0000 unknown
[ 0.582564] 0000000000000016: 0000 unknown
[ 0.582570] 0000000000000018: 0000 unknown
[ 0.582575] 000000000000001a: 0000 unknown
[ 0.582580] 000000000000001c: 0000 unknown
[ 0.582585] 000000000000001e: 0000 unknown
[ 0.582591] 0000000000000020: 0000 unknown
[ 0.582596] Call Trace:
[ 0.582599] ([<0000000002014a46>] kernel_init+0x622/0x7a0)
[ 0.582607] [<0000000000113e22>] kernel_thread_starter+0x6/0xc
[ 0.582615] [<0000000000113e1c>] kernel_thread_starter+0x0/0xc
[ 0.582621] INFO: lockdep is turned off.
[ 0.582624] Last Breaking-Event-Address:
[ 0.582627] [<0000000002014a64>] kernel_init+0x640/0x7a0

Cc: Peter Oberparleiter <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Martin Schwidefsky <[email protected]>
Signed-off-by: Heiko Carstens <[email protected]>
---
include/asm-generic/vmlinux.lds.h | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

Index: linux-2.6/include/asm-generic/vmlinux.lds.h
===================================================================
--- linux-2.6.orig/include/asm-generic/vmlinux.lds.h
+++ linux-2.6/include/asm-generic/vmlinux.lds.h
@@ -441,7 +441,8 @@
}

#ifdef CONFIG_CONSTRUCTORS
-#define KERNEL_CTORS() VMLINUX_SYMBOL(__ctors_start) = .; \
+#define KERNEL_CTORS() . = ALIGN(8); \
+ VMLINUX_SYMBOL(__ctors_start) = .; \
*(.ctors) \
VMLINUX_SYMBOL(__ctors_end) = .;
#else


2009-06-25 13:57:04

by Peter Oberparleiter

[permalink] [raw]
Subject: Re: [PATCH] gcov: fix __ctors_start alignment

Heiko Carstens wrote:
> From: Heiko Carstens <[email protected]>
>
> The ctors section for each object file is eight byte aligned (on 64 bit).
> However the __ctors_start symbol starts at an arbitrary address dependent
> on the size of the previous sections.
> Therefore the linker may add some zeroes after __ctors_start to make sure
> the ctors contents are properly aligned. However the extra zeroes at the
> beginning aren't expected by the code. When walking the functions pointers
> contained in there and extra zeroes are added this may result in random
> jumps.
> So make sure that the __ctors_start symbol is always aligned as well.

Ack. And good find! :) Also with this patch applied, the x86_64 boot
problem reported by Ingo Molnar no longer occurs.