You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
hashcat/OpenCL/m26100-pure.cl

429 lines
11 KiB

/**
* Author......: See docs/credits.txt
* License.....: MIT
*/
#define NEW_SIMD_CODE
#ifdef KERNEL_STATIC
#include M2S(INCLUDE_PATH/inc_vendor.h)
#include M2S(INCLUDE_PATH/inc_types.h)
#include M2S(INCLUDE_PATH/inc_platform.cl)
#include M2S(INCLUDE_PATH/inc_common.cl)
#include M2S(INCLUDE_PATH/inc_simd.cl)
#include M2S(INCLUDE_PATH/inc_hash_sha1.cl)
#include M2S(INCLUDE_PATH/inc_hash_sha256.cl)
#include M2S(INCLUDE_PATH/inc_cipher_aes.cl)
#endif
#define COMPARE_S M2S(INCLUDE_PATH/inc_comp_single.cl)
#define COMPARE_M M2S(INCLUDE_PATH/inc_comp_multi.cl)
typedef struct mozilla_aes_tmp
{
u32 ipad[8];
u32 opad[8];
u32 dgst[8];
u32 out[8];
} mozilla_aes_tmp_t;
typedef struct mozilla_aes
{
u32 iv_buf[4];
u32 ct_buf[4];
} mozilla_aes_t;
DECLSPEC void hmac_sha256_run_V (PRIVATE_AS u32x *w0, PRIVATE_AS u32x *w1, PRIVATE_AS u32x *w2, PRIVATE_AS u32x *w3, PRIVATE_AS u32x *ipad, PRIVATE_AS u32x *opad, PRIVATE_AS u32x *digest)
{
digest[0] = ipad[0];
digest[1] = ipad[1];
digest[2] = ipad[2];
digest[3] = ipad[3];
digest[4] = ipad[4];
digest[5] = ipad[5];
digest[6] = ipad[6];
digest[7] = ipad[7];
sha256_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0];
w0[1] = digest[1];
w0[2] = digest[2];
w0[3] = digest[3];
w1[0] = digest[4];
w1[1] = digest[5];
w1[2] = digest[6];
w1[3] = digest[7];
w2[0] = 0x80000000;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = (64 + 32) * 8;
digest[0] = opad[0];
digest[1] = opad[1];
digest[2] = opad[2];
digest[3] = opad[3];
digest[4] = opad[4];
digest[5] = opad[5];
digest[6] = opad[6];
digest[7] = opad[7];
sha256_transform_vector (w0, w1, w2, w3, digest);
}
KERNEL_FQ void m26100_init (KERN_ATTR_TMPS_ESALT (mozilla_aes_tmp_t, mozilla_aes_t))
{
/**
* base
*/
const u64 gid = get_global_id (0);
if (gid >= GID_CNT) return;
sha1_ctx_t ctx;
sha1_init (&ctx);
// there's some data in salt_buf[8] and onwards so we need to move this to local variable
u32 gs[16] = { 0 };
gs[0] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[0];
gs[1] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[1];
gs[2] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[2];
gs[3] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[3];
gs[4] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[4];
sha1_update_swap (&ctx, gs, 20);
sha1_update_global_swap (&ctx, pws[gid].i, pws[gid].pw_len);
sha1_final (&ctx);
u32 ek[16] = { 0 };
ek[0] = ctx.h[0];
ek[1] = ctx.h[1];
ek[2] = ctx.h[2];
ek[3] = ctx.h[3];
ek[4] = ctx.h[4];
sha256_hmac_ctx_t sha256_hmac_ctx;
sha256_hmac_init (&sha256_hmac_ctx, ek, 20);
tmps[gid].ipad[0] = sha256_hmac_ctx.ipad.h[0];
tmps[gid].ipad[1] = sha256_hmac_ctx.ipad.h[1];
tmps[gid].ipad[2] = sha256_hmac_ctx.ipad.h[2];
tmps[gid].ipad[3] = sha256_hmac_ctx.ipad.h[3];
tmps[gid].ipad[4] = sha256_hmac_ctx.ipad.h[4];
tmps[gid].ipad[5] = sha256_hmac_ctx.ipad.h[5];
tmps[gid].ipad[6] = sha256_hmac_ctx.ipad.h[6];
tmps[gid].ipad[7] = sha256_hmac_ctx.ipad.h[7];
tmps[gid].opad[0] = sha256_hmac_ctx.opad.h[0];
tmps[gid].opad[1] = sha256_hmac_ctx.opad.h[1];
tmps[gid].opad[2] = sha256_hmac_ctx.opad.h[2];
tmps[gid].opad[3] = sha256_hmac_ctx.opad.h[3];
tmps[gid].opad[4] = sha256_hmac_ctx.opad.h[4];
tmps[gid].opad[5] = sha256_hmac_ctx.opad.h[5];
tmps[gid].opad[6] = sha256_hmac_ctx.opad.h[6];
tmps[gid].opad[7] = sha256_hmac_ctx.opad.h[7];
// accessing esalt from _init can lead to false negatives because,
// but we know the global salt is unique for each entry so its fine
u32 es[16] = { 0 };
es[0] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[ 8];
es[1] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[ 9];
es[2] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[10];
es[3] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[11];
es[4] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[12];
es[5] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[13];
es[6] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[14];
es[7] = salt_bufs[DIGESTS_OFFSET_HOST].salt_buf[15];
sha256_hmac_update_swap (&sha256_hmac_ctx, es, 32);
for (u32 i = 0, j = 1; i < 8; i += 8, j += 1)
{
sha256_hmac_ctx_t sha256_hmac_ctx2 = sha256_hmac_ctx;
u32 w0[4];
u32 w1[4];
u32 w2[4];
u32 w3[4];
w0[0] = j;
w0[1] = 0;
w0[2] = 0;
w0[3] = 0;
w1[0] = 0;
w1[1] = 0;
w1[2] = 0;
w1[3] = 0;
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
sha256_hmac_update_64 (&sha256_hmac_ctx2, w0, w1, w2, w3, 4);
sha256_hmac_final (&sha256_hmac_ctx2);
tmps[gid].dgst[i + 0] = sha256_hmac_ctx2.opad.h[0];
tmps[gid].dgst[i + 1] = sha256_hmac_ctx2.opad.h[1];
tmps[gid].dgst[i + 2] = sha256_hmac_ctx2.opad.h[2];
tmps[gid].dgst[i + 3] = sha256_hmac_ctx2.opad.h[3];
tmps[gid].dgst[i + 4] = sha256_hmac_ctx2.opad.h[4];
tmps[gid].dgst[i + 5] = sha256_hmac_ctx2.opad.h[5];
tmps[gid].dgst[i + 6] = sha256_hmac_ctx2.opad.h[6];
tmps[gid].dgst[i + 7] = sha256_hmac_ctx2.opad.h[7];
tmps[gid].out[i + 0] = tmps[gid].dgst[i + 0];
tmps[gid].out[i + 1] = tmps[gid].dgst[i + 1];
tmps[gid].out[i + 2] = tmps[gid].dgst[i + 2];
tmps[gid].out[i + 3] = tmps[gid].dgst[i + 3];
tmps[gid].out[i + 4] = tmps[gid].dgst[i + 4];
tmps[gid].out[i + 5] = tmps[gid].dgst[i + 5];
tmps[gid].out[i + 6] = tmps[gid].dgst[i + 6];
tmps[gid].out[i + 7] = tmps[gid].dgst[i + 7];
}
}
KERNEL_FQ void m26100_loop (KERN_ATTR_TMPS_ESALT (mozilla_aes_tmp_t, mozilla_aes_t))
{
const u64 gid = get_global_id (0);
if ((gid * VECT_SIZE) >= GID_CNT) return;
u32x ipad[8];
u32x opad[8];
ipad[0] = packv (tmps, ipad, gid, 0);
ipad[1] = packv (tmps, ipad, gid, 1);
ipad[2] = packv (tmps, ipad, gid, 2);
ipad[3] = packv (tmps, ipad, gid, 3);
ipad[4] = packv (tmps, ipad, gid, 4);
ipad[5] = packv (tmps, ipad, gid, 5);
ipad[6] = packv (tmps, ipad, gid, 6);
ipad[7] = packv (tmps, ipad, gid, 7);
opad[0] = packv (tmps, opad, gid, 0);
opad[1] = packv (tmps, opad, gid, 1);
opad[2] = packv (tmps, opad, gid, 2);
opad[3] = packv (tmps, opad, gid, 3);
opad[4] = packv (tmps, opad, gid, 4);
opad[5] = packv (tmps, opad, gid, 5);
opad[6] = packv (tmps, opad, gid, 6);
opad[7] = packv (tmps, opad, gid, 7);
for (u32 i = 0; i < 8; i += 8)
{
u32x dgst[8];
u32x out[8];
dgst[0] = packv (tmps, dgst, gid, i + 0);
dgst[1] = packv (tmps, dgst, gid, i + 1);
dgst[2] = packv (tmps, dgst, gid, i + 2);
dgst[3] = packv (tmps, dgst, gid, i + 3);
dgst[4] = packv (tmps, dgst, gid, i + 4);
dgst[5] = packv (tmps, dgst, gid, i + 5);
dgst[6] = packv (tmps, dgst, gid, i + 6);
dgst[7] = packv (tmps, dgst, gid, i + 7);
out[0] = packv (tmps, out, gid, i + 0);
out[1] = packv (tmps, out, gid, i + 1);
out[2] = packv (tmps, out, gid, i + 2);
out[3] = packv (tmps, out, gid, i + 3);
out[4] = packv (tmps, out, gid, i + 4);
out[5] = packv (tmps, out, gid, i + 5);
out[6] = packv (tmps, out, gid, i + 6);
out[7] = packv (tmps, out, gid, i + 7);
for (u32 j = 0; j < LOOP_CNT; j++)
{
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = dgst[0];
w0[1] = dgst[1];
w0[2] = dgst[2];
w0[3] = dgst[3];
w1[0] = dgst[4];
w1[1] = dgst[5];
w1[2] = dgst[6];
w1[3] = dgst[7];
w2[0] = 0x80000000;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = (64 + 32) * 8;
hmac_sha256_run_V (w0, w1, w2, w3, ipad, opad, dgst);
out[0] ^= dgst[0];
out[1] ^= dgst[1];
out[2] ^= dgst[2];
out[3] ^= dgst[3];
out[4] ^= dgst[4];
out[5] ^= dgst[5];
out[6] ^= dgst[6];
out[7] ^= dgst[7];
}
unpackv (tmps, dgst, gid, i + 0, dgst[0]);
unpackv (tmps, dgst, gid, i + 1, dgst[1]);
unpackv (tmps, dgst, gid, i + 2, dgst[2]);
unpackv (tmps, dgst, gid, i + 3, dgst[3]);
unpackv (tmps, dgst, gid, i + 4, dgst[4]);
unpackv (tmps, dgst, gid, i + 5, dgst[5]);
unpackv (tmps, dgst, gid, i + 6, dgst[6]);
unpackv (tmps, dgst, gid, i + 7, dgst[7]);
unpackv (tmps, out, gid, i + 0, out[0]);
unpackv (tmps, out, gid, i + 1, out[1]);
unpackv (tmps, out, gid, i + 2, out[2]);
unpackv (tmps, out, gid, i + 3, out[3]);
unpackv (tmps, out, gid, i + 4, out[4]);
unpackv (tmps, out, gid, i + 5, out[5]);
unpackv (tmps, out, gid, i + 6, out[6]);
unpackv (tmps, out, gid, i + 7, out[7]);
}
}
KERNEL_FQ void m26100_comp (KERN_ATTR_TMPS_ESALT (mozilla_aes_tmp_t, mozilla_aes_t))
{
const u64 gid = get_global_id (0);
const u64 lid = get_local_id (0);
const u64 lsz = get_local_size (0);
/**
* aes shared
*/
#ifdef REAL_SHM
LOCAL_VK u32 s_td0[256];
LOCAL_VK u32 s_td1[256];
LOCAL_VK u32 s_td2[256];
LOCAL_VK u32 s_td3[256];
LOCAL_VK u32 s_td4[256];
LOCAL_VK u32 s_te0[256];
LOCAL_VK u32 s_te1[256];
LOCAL_VK u32 s_te2[256];
LOCAL_VK u32 s_te3[256];
LOCAL_VK u32 s_te4[256];
for (u32 i = lid; i < 256; i += lsz)
{
s_td0[i] = td0[i];
s_td1[i] = td1[i];
s_td2[i] = td2[i];
s_td3[i] = td3[i];
s_td4[i] = td4[i];
s_te0[i] = te0[i];
s_te1[i] = te1[i];
s_te2[i] = te2[i];
s_te3[i] = te3[i];
s_te4[i] = te4[i];
}
SYNC_THREADS ();
#else
CONSTANT_AS u32a *s_td0 = td0;
CONSTANT_AS u32a *s_td1 = td1;
CONSTANT_AS u32a *s_td2 = td2;
CONSTANT_AS u32a *s_td3 = td3;
CONSTANT_AS u32a *s_td4 = td4;
CONSTANT_AS u32a *s_te0 = te0;
CONSTANT_AS u32a *s_te1 = te1;
CONSTANT_AS u32a *s_te2 = te2;
CONSTANT_AS u32a *s_te3 = te3;
CONSTANT_AS u32a *s_te4 = te4;
#endif
if (gid >= GID_CNT) return;
u32 ukey[8];
ukey[0] = tmps[gid].out[0];
ukey[1] = tmps[gid].out[1];
ukey[2] = tmps[gid].out[2];
ukey[3] = tmps[gid].out[3];
ukey[4] = tmps[gid].out[4];
ukey[5] = tmps[gid].out[5];
ukey[6] = tmps[gid].out[6];
ukey[7] = tmps[gid].out[7];
u32 ks[60];
AES256_set_decrypt_key (ks, ukey, s_te0, s_te1, s_te2, s_te3, s_td0, s_td1, s_td2, s_td3);
// first check the padding
u32 iv_buf[4];
iv_buf[0] = esalt_bufs[DIGESTS_OFFSET_HOST].iv_buf[0];
iv_buf[1] = esalt_bufs[DIGESTS_OFFSET_HOST].iv_buf[1];
iv_buf[2] = esalt_bufs[DIGESTS_OFFSET_HOST].iv_buf[2];
iv_buf[3] = esalt_bufs[DIGESTS_OFFSET_HOST].iv_buf[3];
u32 ct_buf[4];
ct_buf[0] = esalt_bufs[DIGESTS_OFFSET_HOST].ct_buf[0];
ct_buf[1] = esalt_bufs[DIGESTS_OFFSET_HOST].ct_buf[1];
ct_buf[2] = esalt_bufs[DIGESTS_OFFSET_HOST].ct_buf[2];
ct_buf[3] = esalt_bufs[DIGESTS_OFFSET_HOST].ct_buf[3];
u32 pt_buf[4];
aes256_decrypt (ks, ct_buf, pt_buf, s_td0, s_td1, s_td2, s_td3, s_td4);
pt_buf[0] ^= iv_buf[0];
pt_buf[1] ^= iv_buf[1];
pt_buf[2] ^= iv_buf[2];
pt_buf[3] ^= iv_buf[3];
// password-check\x02\x02
if (pt_buf[0] != 0x73736170) return;
if (pt_buf[1] != 0x64726f77) return;
if (pt_buf[2] != 0x6568632d) return;
if (pt_buf[3] != 0x02026b63) return;
const u32 r0 = ct_buf[0];
const u32 r1 = ct_buf[1];
const u32 r2 = ct_buf[2];
const u32 r3 = ct_buf[3];
#define il_pos 0
#ifdef KERNEL_STATIC
#include COMPARE_M
#endif
}