correction to compat_sys_kexec_load

correction to compat_sys_kexec_load

[Sharyathi Nagesh]

[-- Attachment #1: Type: text/plain, Size: 972 bytes --]

Hi
   While testing with kexec tool, I observed some problems. When 
application (kexec) is 32 bit and kernel is 64 bit I observed that 
loading crash kernel works without any issues but unloading crash kernel 
fails.
--------------------------------------------------------
running strace over 'kexec -u -p'
show the problem to be with sys_kexec_load() call

sys_kexec_load(0, 0, 0, 0x1, 0)         = -1 EINVAL (Invalid argument)
write(2, "kexec_load (0 segments) failed: "..., 49
kexec_load (0 segments) failed: Invalid argument
) = 4
--------------------------------------------------------

This is patch to fix the problem, I think kernel code had a typo where in:
if((flags & KEXEC_ARCH_MASK) == KEXEC_ARCH) was used instead of
if((flags & KEXEC_ARCH_MASK) != KEXEC_ARCH)

This patch takes care of that, I have tested the patch it worked fine 
for me. Please review the patch and let me know of your views. This 
patch is based on linux-2.6.26-rc3.

Thanks
Yeehaw



[-- Attachment #2: compat_sys_kexec_correction.patch --]
[-- Type: text/x-patch, Size: 596 bytes --]

Index: linux-2.6.25.3-2/kernel/kexec.c
===================================================================
--- linux-2.6.25.3-2.orig/kernel/kexec.c	2008-05-23 14:57:08.000000000 +0530
+++ linux-2.6.25.3-2/kernel/kexec.c	2008-05-23 15:00:41.000000000 +0530
@@ -1068,7 +1068,8 @@
 	/* Don't allow clients that don't understand the native
 	 * architecture to do anything.
 	 */
-	if ((flags & KEXEC_ARCH_MASK) == KEXEC_ARCH_DEFAULT)
+	if (((flags & KEXEC_ARCH_MASK) != KEXEC_ARCH) &&
+		((flags & KEXEC_ARCH_MASK) != KEXEC_ARCH_DEFAULT))
 		return -EINVAL;
 
 	if (nr_segments > KEXEC_SEGMENT_MAX)

Re: correction to compat_sys_kexec_load

[Eric W. Biederman]

Sharyathi Nagesh <sharyath@in.ibm.com> writes:

> Hi
>   While testing with kexec tool, I observed some problems. When application
> (kexec) is 32 bit and kernel is 64 bit I observed that loading crash kernel
> works without any issues but unloading crash kernel fails.
> --------------------------------------------------------
> running strace over 'kexec -u -p'
> show the problem to be with sys_kexec_load() call
>
> sys_kexec_load(0, 0, 0, 0x1, 0)         = -1 EINVAL (Invalid argument)
> write(2, "kexec_load (0 segments) failed: "..., 49
> kexec_load (0 segments) failed: Invalid argument
> ) = 4

Yes.  This is a bug.  Although not in the kernel implementation.

> --------------------------------------------------------
>
> This is patch to fix the problem, I think kernel code had a typo where in:
> if((flags & KEXEC_ARCH_MASK) == KEXEC_ARCH) was used instead of
> if((flags & KEXEC_ARCH_MASK) != KEXEC_ARCH)

Nope.  We do the latter check after we have fixed up the arguments
and call sys_kexec_load.  The check really is meant to filter out
KEXEC_ARCH_DEFAULT.


> This patch takes care of that, I have tested the patch it worked fine for
> me. Please review the patch and let me know of your views. This patch is based
> on linux-2.6.26-rc3.

That patch as it exists is actively bad.  It removes the check for a really
nasty gotcha if someone passes in KEXEC_ARCH_DEFAULT in 32bit mode.  Code
expecting a 32bit handoff and getting a 64bit handoff will explode in fun
ways.  You happened to test the one corner case where this does not matter.

What we need to do is fix /sbin/kexec to pass in the correct
architecture of the kernel for unload as it does for load.

Eric

Re: correction to compat_sys_kexec_load

[Bernhard Walle]

[-- Attachment #1: Type: text/plain, Size: 609 bytes --]

* ebiederm@xmission.com (Eric W. Biederman) [2008-05-23 13:14]:
>
> What we need to do is fix /sbin/kexec to pass in the correct
> architecture of the kernel for unload as it does for load.

How should it know that it unloads a 32 bit kernel on a 64 bit system?
It doesn't have access to the kernel any more once it has been loaded.



	Bernhard

PS: My mail server complains that  maneesh@in.ltcfwd.linux.ibm.com is
invalid because "Recipient address rejected: Domain not found". Maybe
some of the other IBM guys in Cc can take a look at this ... I just
removed that address from Cc for now.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: correction to compat_sys_kexec_load

[Eric W. Biederman]

Bernhard Walle <bwalle@suse.de> writes:

> * ebiederm@xmission.com (Eric W. Biederman) [2008-05-23 13:14]:
>>
>> What we need to do is fix /sbin/kexec to pass in the correct
>> architecture of the kernel for unload as it does for load.
>
> How should it know that it unloads a 32 bit kernel on a 64 bit system?
> It doesn't have access to the kernel any more once it has been loaded.

The architecture parameter is the architecture of the running kernel
that implements sys_kexec_load.

Because it is a pain for testing and in general impossible we don't
change cpu modes during a kexec.  So a 32bit caller of sys_kexec_load
will need to passing in different code if it is running on a 32bit or
a 64bit kernel.

The trampoline code in /sbin/kexec does change modes on x86 when
appropriate.

Caring if you know the architecture in the unload case is a bit
silly. As there is no real justification for it.  At this
point getting user space fixed so that it works on older kernels
seems important. 

Eric

Re: correction to compat_sys_kexec_load

[Bernhard Walle]

* ebiederm@xmission.com (Eric W. Biederman) [2008-05-23 17:36]:
>
> As there is no real justification for it.  At this
> point getting user space fixed so that it works on older kernels
> seems important. 

http://article.gmane.org/gmane.linux.kernel.kexec/1534 should fix this.

Re: correction to compat_sys_kexec_load

[Sharyathi Nagesh]

Bernhard/Eric
   Thanks for helping with this. kexec patch has been submitted up stream
Thanks
Yeehaw
Bernhard Walle wrote:
> * ebiederm@xmission.com (Eric W. Biederman) [2008-05-23 17:36]:
>> As there is no real justification for it.  At this
>> point getting user space fixed so that it works on older kernels
>> seems important. 
> 
> http://article.gmane.org/gmane.linux.kernel.kexec/1534 should fix this.
> 
>