binfmt_misc.c: avoid potential kernel stack overflow
Subsystems (2)
(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?
fs/binfmt_misc.c +2 -2
@@ -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;
Quoted 45 lines · Pavel Emelyanov
Quoted 7 lines
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 theload_misc_binary->open_exec->load_misc_binarychain and stack overflow.Quoted 8 lines
There's a similar problem with
binfmt_script, and there's ash_bangmemner onlinux_binprmstructure to handle this, but simply raising this inbinfmt_miscmay break some setups when the interpreter of some misc binaries is a script.So the proposal is to turn
sh_banginto a bit, add a new one (themisc_bang) and raise it inload_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.
Quoted 6 lines · 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_bangflag even too early.
Quoted 7 lines
Thus when we launch a script the
load_misc_binarysets this bang, then returns error, since the binary is actually a script, then theload_script_binarysuccessfully loads the script, then it loads the misc binary again, which exits with the-ENOEXECerror due to bang set.This patch helped my box, what about yours?
It works. Thank you.
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;
Signature · 3 lines
Regards, Kirill A. Shutemov + Belarus, Minsk + ALT Linux Team, http://www.altlinux.com/
Quoted 10 lines · Kirill A. Shutemov
I have noticed yet another problem: more than one bit of
sh_bangcan be used on alpha:
fs/exec.c
1189 return retval;
1190
1191 /* Remember if the application is TASO. */
1192bprm->sh_bang=eh->ah.entry<0x100000000UL;
1193
1194bprm->file= file;
Sorry. Please ignore this part of the message. %)
Signature · 3 lines
Regards, Kirill A. Shutemov + Belarus, Minsk + ALT Linux Team, http://www.altlinux.com/
Quoted 28 lines · On Mon, 18 Aug 2008 18:09:23 +0400 Pavel Emelyanov <[email protected]> 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_bangflag even too early.Thus when we launch a script the
load_misc_binarysets this bang, then returns error, since the binary is actually a script, then theload_script_binarysuccessfully loads the script, then it loads the misc binary again, which exits with the-ENOEXECerror due to bang set.This patch helped my box, what about yours?
fs/binfmt_misc.c +0
-2@@ -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);
fs/binfmt_misc.c static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs) · line 204
Quoted 8 lines
fs/binfmt_misc.c +2
-0@@ -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 <[email protected]>
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.
---
fs/binfmt_misc.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
fs/binfmt_misc.c renamed from fs/binfmt_misc.c~binfmt_miscc-avoid-potential-kernel-stack-overflow +2 -2
@@ -120,8 +120,6 @@ static int load_misc_binary(struct linuxif (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 linuxif (retval <0 ) goto _error; + bprm->misc_bang =1 ; + retval =search_binary_handler (bprm, regs);if (retval <0 ) goto _error;
_
Quoted 42 lines · Andrew Morton
Quoted 28 lines · On Mon, 18 Aug 2008 18:09:23 +0400 Pavel Emelyanov <[email protected]> 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_bangflag even too early.Thus when we launch a script the
load_misc_binarysets this bang, then returns error, since the binary is actually a script, then theload_script_binarysuccessfully loads the script, then it loads the misc binary again, which exits with the-ENOEXECerror due to bang set.This patch helped my box, what about yours?
fs/binfmt_misc.c +0
-2@@ -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);fs/binfmt_misc.c static int load_misc_binary(struct linux_binprm *bprm, struct pt_regs *regs) · line 204
Quoted 8 lines
fs/binfmt_misc.c +2
-0@@ -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>
Quoted 1 line
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)?
Quoted 2 lines
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.
Quoted 1 line
Thanks.
Thanks,
Pavel