lkml.org 
[lkml]   [2009]   [Sep]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] fix error handling in load_module()
On Mon, 7 Sep 2009 19:45:58 +0530
Kamalesh Babulal <kamalesh@linux.vnet.ibm.com> wrote:

> Hi Rusty,
>
> During our testing following call trace was seen. The testcase was
> to compile the kernel based on the distro config and try to insert all the
> modules compiled.
>
> #!/bin/sh
>
> for module in `modprobe -l | tr '\n' ' '`
> do
> insert_module=`basename $module .ko`
> modprobe -v $insert_module
> done
>
> freq_table sputrace hvcserver axonram pmi ipv6 fuse ehea ib
> Sep 7 15:46:04 mjs22lp5 kernel: mveth ibmvscsic scsi_transport_srp scsi_tgt
> Sep 7 15:46:04 mjs22lp5 kernel: NIP: c0000000000ebba0 LR: c0000000000ee79c CTR: 0000000000000000
> Sep 7 15:46:04 mjs22lp5 kernel: REGS: c00000002c90b8e0 TRAP: 0700 Tainted: P D (2.6.31-rc8)
> Sep 7 15:46:04 mjs22lp5 kernel: MSR: 8000000000029032 <EE,ME,CE,IR,DR> CR: 24222488 XER: 00000008
> Sep 7 15:46:04 mjs22lp5 kernel: TASK = c00000002ff40000[9062] 'modprobe' THREAD: c00000002c908000 CPU: 0
> Sep 7 15:46:04 mjs22lp5 kernel: GPR00: 0000000000000010 c00000002c90bb60 c000000001421e68 0000000000000000
> Sep 7 15:46:04 mjs22lp5 kernel: GPR04: c000000000691a5c c00000000009f5c4 0000000000000000 c0000000167f6630
> Sep 7 15:46:04 mjs22lp5 kernel: GPR08: c0000000167f72a4 000000000000031f c000000000bb9580 000000000000031e
> Sep 7 15:46:04 mjs22lp5 kernel: GPR12: 800000000631b800 c0000000015a2600 0000000000000000 0000000000000000
> Sep 7 15:46:04 mjs22lp5 kernel: GPR16: 0000000000000033 d00000000fb1f6d0 d00000000fb1fe50 000000000000000e
> Sep 7 15:46:04 mjs22lp5 kernel: GPR20: d00000000fb1efb8 d00000000fb62260 d00000000fb00000 8000000000000000
> Sep 7 15:46:04 mjs22lp5 kernel: GPR24: 0000000000000004 d00000000fb1f190 0000000000000035 fffffffffffffff4
> Sep 7 15:46:04 mjs22lp5 kernel: GPR28: 0000000000000000 000000000000031e c00000000137def8 c00000002c90bb60
> Sep 7 15:46:04 mjs22lp5 kernel: NIP [c0000000000ebba0] .percpu_modfree+0xe8/0x210
> Sep 7 15:46:04 mjs22lp5 kernel: LR [c0000000000ee79c] .load_module+0x14f8/0x1650
> Sep 7 15:46:04 mjs22lp5 kernel: Call Trace:
> Sep 7 15:46:04 mjs22lp5 kernel: [c00000002c90bb60] [c00000002c90bc00] 0xc00000002c90bc00 (unreliable)
> Sep 7 15:46:04 mjs22lp5 kernel: [c00000002c90bc00] [c0000000000ee79c] .load_module+0x14f8/0x1650
> Sep 7 15:46:04 mjs22lp5 kernel: [c00000002c90bd90] [c0000000000ee988] .SyS_init_module+0x94/0x2ac
> Sep 7 15:46:04 mjs22lp5 kernel: [c00000002c90be30] [c0000000000084dc] syscall_exit+0x0/0x40
> Sep 7 15:46:04 mjs22lp5 kernel: Instruction dump:
> Sep 7 15:46:05 mjs22lp5 kernel: 48000038 e8080006 793d0020 39080004 78090020 2f800000 409c000c 7c0000d0
> Sep 7 15:46:05 mjs22lp5 kernel: 78090020 7d4a4a14 393d0001 4200ffb0 <0fe00000> 48000000 38a30001 7f83e378
> Sep 7 15:46:05 mjs22lp5 kernel: ---[ end trace 3c8bbdf1034c7f0d ]---
>
> Once the percpu_modalloc fails, percpu_modfree(mod->refptr) is called on a NULL pointer.
> We try calling it on a NULL pointer. The following patch fixes the problem by introducing
> a check for mod->refptr before calling percpu_modfree.

Where did it crash and why did it crash? That trace is pretty unclear.

> diff --git a/kernel/module.c b/kernel/module.c
> index 2d53718..7f89258 100644
> --- a/kernel/module.c
> +++ b/kernel/module.c
> @@ -2379,7 +2379,8 @@ static noinline struct module *load_module(void __user *umod,
> module_unload_free(mod);
> #if defined(CONFIG_MODULE_UNLOAD) && defined(CONFIG_SMP)
> free_init:
> - percpu_modfree(mod->refptr);
> + if (mod->refptr)
> + percpu_modfree(mod->refptr);
> #endif
> module_free(mod, mod->module_init);
> free_core:

My reverse engineering of the secret, undocumented percpu_modfree()
indicates that its mad inventor intended that percpu_modfree(NULL) be a
valid thing to do.

It calls free_percpu(), all implementations of which appear to secretly
support free_percpu(NULL).

So why did your machine crash?

This:

void free_percpu(void *ptr)
{
void *addr = __pcpu_ptr_to_addr(ptr);
struct pcpu_chunk *chunk;
unsigned long flags;
int off;

if (!ptr)
return;

is dangerous. The implementation of __pcpu_ptr_to_addr() can be
overridden by asm/percpu.h and there's no reason why the compiler won't
choose to pass a NULL into __pcpu_ptr_to_addr().

But there doesn't appear to be any overriding of __pcpu_ptr_to_addr()
in 2.6.31 and the default __pcpu_ptr_to_addr() looks like it will
handle a NULL pointer OK.

So again, why did your machine crash?



From: Andrew Morton <akpm@linux-foundation.org>

__pcpu_ptr_to_addr() can be overridden by the architecture and might not
behave well if passed a NULL pointer. So avoid calling it until we have
verified that its arg is not NULL.

Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
---

mm/percpu.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)

diff -puN mm/percpu.c~percpu-avoid-calling-__pcpu_ptr_to_addrnull mm/percpu.c
--- a/mm/percpu.c~percpu-avoid-calling-__pcpu_ptr_to_addrnull
+++ a/mm/percpu.c
@@ -957,7 +957,7 @@ static void pcpu_reclaim(struct work_str
*/
void free_percpu(void *ptr)
{
- void *addr = __pcpu_ptr_to_addr(ptr);
+ void *addr;
struct pcpu_chunk *chunk;
unsigned long flags;
int off;
@@ -965,6 +965,8 @@ void free_percpu(void *ptr)
if (!ptr)
return;

+ addr = __pcpu_ptr_to_addr(ptr);
+
spin_lock_irqsave(&pcpu_lock, flags);

chunk = pcpu_chunk_addr_search(addr);
_


\
 
 \ /
  Last update: 2009-09-10 23:17    [W:0.062 / U:0.372 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site