Re: [PATCH] binfmt_misc.c: avoid potential kernel stack overflow

Re: [PATCH] binfmt_misc.c: avoid potential kernel stack overflow

[Pavel Emelyanov]

(Put lkml in Cc. The original message is beyond)

Oops! My fault. The problem is that in case of modularized binfmt,
the appropriate binary handler gets registered _before_ the script
one and sets the misc_bang flag even too early.

Thus when we launch a script the load_misc_binary sets this bang,
then returns error, since the binary is actually a script, then the
load_script_binary successfully loads the script, then it loads the
misc binary again, which exits with the -ENOEXEC error due to bang 
set.

This patch helped my box, what about yours?

diff --git a/fs/binfmt_misc.c b/fs/binfmt_misc.c
index 7562053..8d7e88e 100644
--- a/fs/binfmt_misc.c
+++ b/fs/binfmt_misc.c
@@ -120,8 +120,6 @@ static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs)
 	if (bprm->misc_bang)
 		goto _ret;
 
-	bprm->misc_bang = 1;
-
 	/* to keep locking time low, we copy the interpreter string */
 	read_lock(&entries_lock);
 	fmt = check_file(bprm);
@@ -199,6 +197,8 @@ static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs)
 	if (retval < 0)
 		goto _error;
 
+	bprm->misc_bang = 1;
+
 	retval = search_binary_handler (bprm, regs);
 	if (retval < 0)
 		goto _error;

> On Tue, Apr 29, 2008 at 12:59:24AM -0700, Pavel Emelyanov wrote:
>> This can be triggered with root help only, but...
>>
>> Register the ":text:E::txt::/root/cat.txt:' rule in binfmt_misc (by root) and
>> try launching the cat.txt file (by anyone) :) The result is - the endless
>> recursion in the load_misc_binary -> open_exec -> load_misc_binary chain and
>> stack overflow.
>>
>> There's a similar problem with binfmt_script, and there's a sh_bang memner on
>> linux_binprm structure to handle this, but simply raising this in binfmt_misc
>> may break some setups when the interpreter of some misc binaries is a script.
>>
>> So the proposal is to turn sh_bang into a bit, add a new one (the misc_bang)
>> and raise it in load_misc_binary.  After this, even if we set up the misc ->
>> script -> misc loop for binfmts one of them will step on its own bang and
>> exit.
> 
> This patch causes problem in some cases, if the kernel compiled with 
> CONFIG_BINFMT_MISC=m:
> 
> $ pwd
> /tmp/chroot
> $ cat test0.c 
> #include <stdio.h>
> 
> int main(void)
> {
>         printf("test\n");
>         return 0;
> }
> $ gcc test0.c -o test0 -static
> $ sudo sh -c 'echo ":test:M::123::/test0:" > /proc/sys/fs/binfmt_misc/register'
> $ cat test1
> 123
> $ cat test2
> #!/test1
> $ sudo chroot /tmp/chroot /test2
> chroot: cannot run command `/test2': No such file or directory
> $ sudo strace chroot /tmp/chroot /test2
> ...
> execve("/test2", ["/test2"], [/* 54 vars */]) = -1 ENOEXEC (Exec format error)
> ...
> 
> This test works fine if I revert the patch or compile with CONFIG_BINFMT_MISC=y.
> 

Re: [PATCH] binfmt_misc.c: avoid potential kernel stack overflow

[Kirill A. Shutemov]

[-- Attachment #1: Type: text/plain, Size: 1192 bytes --]

On Mon, Aug 18, 2008 at 06:09:23PM +0400, Pavel Emelyanov wrote:
> (Put lkml in Cc. The original message is beyond)
> 
> Oops! My fault. The problem is that in case of modularized binfmt,
> the appropriate binary handler gets registered _before_ the script
> one and sets the misc_bang flag even too early.
> 
> Thus when we launch a script the load_misc_binary sets this bang,
> then returns error, since the binary is actually a script, then the
> load_script_binary successfully loads the script, then it loads the
> misc binary again, which exits with the -ENOEXEC error due to bang 
> set.
> 
> This patch helped my box, what about yours?

It works. Thank you.

Reported-and-tested-by: Kirill A. Shutemov <kirill@shutemov.name>

I have noticed yet another problem: more than one bit of sh_bang can be
used on alpha:

fs/exec.c
1189                         return retval;
1190 
1191                 /* Remember if the application is TASO.  */
1192                 bprm->sh_bang = eh->ah.entry < 0x100000000UL;
1193 
1194                 bprm->file = file;


-- 
Regards,  Kirill A. Shutemov
 + Belarus, Minsk
 + ALT Linux Team, http://www.altlinux.com/

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 197 bytes --]

Re: [PATCH] binfmt_misc.c: avoid potential kernel stack overflow

[Kirill A. Shutemov]

[-- Attachment #1: Type: text/plain, Size: 579 bytes --]

On Mon, Aug 18, 2008 at 05:44:19PM +0300, Kirill A. Shutemov wrote:
> I have noticed yet another problem: more than one bit of sh_bang can be
> used on alpha:
> 
> fs/exec.c
> 1189                         return retval;
> 1190 
> 1191                 /* Remember if the application is TASO.  */
> 1192                 bprm->sh_bang = eh->ah.entry < 0x100000000UL;
> 1193 
> 1194                 bprm->file = file;

Sorry. Please ignore this part of the message. %)

-- 
Regards,  Kirill A. Shutemov
 + Belarus, Minsk
 + ALT Linux Team, http://www.altlinux.com/

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 197 bytes --]

Re: [PATCH] binfmt_misc.c: avoid potential kernel stack overflow

[Andrew Morton]

On Mon, 18 Aug 2008 18:09:23 +0400
Pavel Emelyanov <xemul@openvz.org> wrote:

> (Put lkml in Cc. The original message is beyond)
> 
> Oops! My fault. The problem is that in case of modularized binfmt,
> the appropriate binary handler gets registered _before_ the script
> one and sets the misc_bang flag even too early.
> 
> Thus when we launch a script the load_misc_binary sets this bang,
> then returns error, since the binary is actually a script, then the
> load_script_binary successfully loads the script, then it loads the
> misc binary again, which exits with the -ENOEXEC error due to bang 
> set.
> 
> This patch helped my box, what about yours?
> 
> diff --git a/fs/binfmt_misc.c b/fs/binfmt_misc.c
> index 7562053..8d7e88e 100644
> --- a/fs/binfmt_misc.c
> +++ b/fs/binfmt_misc.c
> @@ -120,8 +120,6 @@ static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs)
>  	if (bprm->misc_bang)
>  		goto _ret;
>  
> -	bprm->misc_bang = 1;
> -
>  	/* to keep locking time low, we copy the interpreter string */
>  	read_lock(&entries_lock);
>  	fmt = check_file(bprm);
> @@ -199,6 +197,8 @@ static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs)
>  	if (retval < 0)
>  		goto _error;
>  
> +	bprm->misc_bang = 1;
> +
>  	retval = search_binary_handler (bprm, regs);
>  	if (retval < 0)
>  		goto _error;

<scrabble, hunt>

I put together the below description.  It has no signed-off-by: (yet).
Has this been sufficiently well tested and checked to be in a merge-ready
state?

Thanks.



From: Pavel Emelyanov <xemul@openvz.org>

Fix a regression introduced by 3a2e7f47d71e1df86acc1dda6826890b6546a4e1
("binfmt_misc.c: avoid potential kernel stack overflow").

In the case of modularized binfmt, the appropriate binary handler gets
registered _before_ the script one and sets the misc_bang flag even too
early.

Thus when we launch a script the load_misc_binary sets this bang, then
returns error, since the binary is actually a script, then the
load_script_binary successfully loads the script, then it loads the misc
binary again, which exits with the -ENOEXEC error due to bang set.

Reported-and-tested-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: <stable@kernel.org>		[2.6.26.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
---

 fs/binfmt_misc.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff -puN fs/binfmt_misc.c~binfmt_miscc-avoid-potential-kernel-stack-overflow fs/binfmt_misc.c
--- a/fs/binfmt_misc.c~binfmt_miscc-avoid-potential-kernel-stack-overflow
+++ a/fs/binfmt_misc.c
@@ -120,8 +120,6 @@ static int load_misc_binary(struct linux
 	if (bprm->misc_bang)
 		goto _ret;
 
-	bprm->misc_bang = 1;
-
 	/* to keep locking time low, we copy the interpreter string */
 	read_lock(&entries_lock);
 	fmt = check_file(bprm);
@@ -199,6 +197,8 @@ static int load_misc_binary(struct linux
 	if (retval < 0)
 		goto _error;
 
+	bprm->misc_bang = 1;
+
 	retval = search_binary_handler (bprm, regs);
 	if (retval < 0)
 		goto _error;
_

Re: [PATCH] binfmt_misc.c: avoid potential kernel stack overflow

[Pavel Emelyanov]

Andrew Morton wrote:
> On Mon, 18 Aug 2008 18:09:23 +0400
> Pavel Emelyanov <xemul@openvz.org> wrote:
> 
>> (Put lkml in Cc. The original message is beyond)
>>
>> Oops! My fault. The problem is that in case of modularized binfmt,
>> the appropriate binary handler gets registered _before_ the script
>> one and sets the misc_bang flag even too early.
>>
>> Thus when we launch a script the load_misc_binary sets this bang,
>> then returns error, since the binary is actually a script, then the
>> load_script_binary successfully loads the script, then it loads the
>> misc binary again, which exits with the -ENOEXEC error due to bang 
>> set.
>>
>> This patch helped my box, what about yours?
>>
>> diff --git a/fs/binfmt_misc.c b/fs/binfmt_misc.c
>> index 7562053..8d7e88e 100644
>> --- a/fs/binfmt_misc.c
>> +++ b/fs/binfmt_misc.c
>> @@ -120,8 +120,6 @@ static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs)
>>  	if (bprm->misc_bang)
>>  		goto _ret;
>>  
>> -	bprm->misc_bang = 1;
>> -
>>  	/* to keep locking time low, we copy the interpreter string */
>>  	read_lock(&entries_lock);
>>  	fmt = check_file(bprm);
>> @@ -199,6 +197,8 @@ static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs)
>>  	if (retval < 0)
>>  		goto _error;
>>  
>> +	bprm->misc_bang = 1;
>> +
>>  	retval = search_binary_handler (bprm, regs);
>>  	if (retval < 0)
>>  		goto _error;
> 
> <scrabble, hunt>
> 
> I put together the below description.  It has no signed-off-by: (yet).

Well, sorry for that, I just wanted to get the Kirill's approval of the
fix, while testing other things myself. I sent the properly formatted
patch later. So can you, please, pick the comment and/or subject from
that one (which is a bit less messy, I think)?

> Has this been sufficiently well tested and checked to be in a merge-ready
> state?

I have checked different combinations, so I believe it has.

> Thanks.

Thanks,
Pavel