iclass.c: speeding up MAC calculation

diff --git a/armsrc/iclass.c b/armsrc/iclass.c
index ee305a26fc6b6ac521d3f633f01ff1bf2439b817..5d7375a7bd8ad982778ad41046925570e9fa33e0 100644 (file)
--- a/armsrc/iclass.c
+++ b/armsrc/iclass.c
@@ -1146,7 +1146,7 @@ int doIClassSimulation(int simulationMode, uint8_t *reader_mac_buf) {
                                memcpy(data_response, ToSend, ToSendMax);
                                modulated_response = data_response;
                                modulated_response_size = ToSendMax;
-                               response_delay = 0;//We need to hurry here...
+                               response_delay = 0; //We need to hurry here... (but maybe not too much... ??)
                                //exitLoop = true;
                        } else {    //Not fullsim, we don't respond
                                // We do not know what to answer, so lets keep quiet

diff --git a/armsrc/optimized_cipher.c b/armsrc/optimized_cipher.c
index ee9d556817772567b5ff3c8b6c991e8c79748ff4..2ac72ec0fa9857a76f18355345e4e99655988d0f 100644 (file)
--- a/armsrc/optimized_cipher.c
+++ b/armsrc/optimized_cipher.c
@@ -60,15 +60,63 @@
   -- MHS 2015
 **/
 
+/**
+
+  The runtime of opt_doTagMAC_2() with the MHS optimized version was 403 microseconds on Proxmark3.
+  This was still to slow for some newer readers which didn't want to wait that long.
+
+  Further optimizations to speedup the MAC calculations:
+  * Optimized opt_Tt logic
+  * Look up table for opt_select
+  * Removing many unnecessary bit maskings (& 0x1)
+  * updating state in place instead of alternating use of a second state structure
+  * remove the necessity to reverse bits of input and output bytes
+ 
+  opt_doTagMAC_2() now completes in 270 microseconds.
+
+  -- piwi 2019
+**/
+
 #include "optimized_cipher.h"
 #include <stddef.h>
 #include <stdbool.h>
 #include <stdint.h>
-
-
-#define opt_T(s) (0x1 & ((s->t >> 15) ^ (s->t >> 14)^ (s->t >> 10)^ (s->t >> 8)^ (s->t >> 5)^ (s->t >> 4)^ (s->t >> 1)^ s->t))
-
-#define opt_B(s) (((s->b >> 6) ^ (s->b >> 5) ^ (s->b >> 4) ^ (s->b)) & 0x1)
+#include "string.h"
+
+static const uint8_t opt_select_LUT[256] = {
+       00, 03, 02, 01, 02, 03, 00, 01, 04, 07, 07, 04, 06, 07, 05, 04,
+       01, 02, 03, 00, 02, 03, 00, 01, 05, 06, 06, 05, 06, 07, 05, 04,
+       06, 05, 04, 07, 04, 05, 06, 07, 06, 05, 05, 06, 04, 05, 07, 06,
+       07, 04, 05, 06, 04, 05, 06, 07, 07, 04, 04, 07, 04, 05, 07, 06,
+       06, 05, 04, 07, 04, 05, 06, 07, 02, 01, 01, 02, 00, 01, 03, 02,
+       03, 00, 01, 02, 00, 01, 02, 03, 07, 04, 04, 07, 04, 05, 07, 06,
+       00, 03, 02, 01, 02, 03, 00, 01, 00, 03, 03, 00, 02, 03, 01, 00,
+       05, 06, 07, 04, 06, 07, 04, 05, 05, 06, 06, 05, 06, 07, 05, 04,
+       02, 01, 00, 03, 00, 01, 02, 03, 06, 05, 05, 06, 04, 05, 07, 06,
+       03, 00, 01, 02, 00, 01, 02, 03, 07, 04, 04, 07, 04, 05, 07, 06,
+       02, 01, 00, 03, 00, 01, 02, 03, 02, 01, 01, 02, 00, 01, 03, 02,
+       03, 00, 01, 02, 00, 01, 02, 03, 03, 00, 00, 03, 00, 01, 03, 02,
+       04, 07, 06, 05, 06, 07, 04, 05, 00, 03, 03, 00, 02, 03, 01, 00,
+       01, 02, 03, 00, 02, 03, 00, 01, 05, 06, 06, 05, 06, 07, 05, 04,
+       04, 07, 06, 05, 06, 07, 04, 05, 04, 07, 07, 04, 06, 07, 05, 04,
+       01, 02, 03, 00, 02, 03, 00, 01, 01, 02, 02, 01, 02, 03, 01, 00
+};
+
+/********************** the table above has been generated with this code: ********
+#include "util.h"
+static void init_opt_select_LUT(void) {
+       for (int r = 0; r < 256; r++) {
+               uint8_t r_ls2 = r << 2;
+               uint8_t r_and_ls2 = r & r_ls2;
+               uint8_t r_or_ls2  = r | r_ls2;
+               uint8_t z0 = (r_and_ls2 >> 5) ^ ((r & ~r_ls2) >> 4) ^ ( r_or_ls2 >> 3);
+               uint8_t z1 = (r_or_ls2 >> 6) ^ ( r_or_ls2 >> 1) ^ (r >> 5) ^ r;
+               uint8_t z2 = ((r & ~r_ls2) >> 4) ^ (r_and_ls2 >> 3) ^ r;
+               opt_select_LUT[r] = (z0 & 4) | (z1 & 2) | (z2 & 1);
+       }
+       print_result("", opt_select_LUT, 256);
+}
+***********************************************************************************/
 
 #define opt__select(x,y,r)  (4 & (((r & (r << 2)) >> 5) ^ ((r & ~(r << 2)) >> 4) ^ ( (r | r << 2) >> 3)))\
        |(2 & (((r | r << 2) >> 6) ^ ( (r | r << 2) >> 1) ^ (r >> 5) ^ r ^ ((x^y) << 1)))\
@@ -78,9 +126,6 @@
  * Some background on the expression above can be found here...
 uint8_t xopt__select(bool x, bool y, uint8_t r)
 {
-       uint8_t r_ls2 = r << 2;
-       uint8_t r_and_ls2 = r & r_ls2;
-       uint8_t r_or_ls2  = r | r_ls2;
 
        //r:      r0 r1 r2 r3 r4 r5 r6 r7
        //r_ls2:  r2 r3 r4 r5 r6 r7  0  0
@@ -100,81 +145,95 @@ uint8_t xopt__select(bool x, bool y, uint8_t r)
 }
 */
 
-void opt_successor(const uint8_t *k, State *s, bool y, State *successor) {
-       uint8_t Tt = 1 & opt_T(s);
-
-       successor->t = (s->t >> 1);
-       successor->t |= (Tt ^ (s->r >> 7 & 0x1) ^ (s->r >> 3 & 0x1)) << 15;
-
-       successor->b = s->b >> 1;
-       successor->b |= (opt_B(s) ^ (s->r & 0x1)) << 7;
-
-       successor->r = (k[opt__select(Tt, y, s->r)] ^ successor->b) + s->l ;
-       successor->l = successor->r + s->r;
+static void opt_successor(const uint8_t *k, State *s, uint8_t y) {
+// #define opt_T(s) (0x1 & ((s->t >> 15) ^ (s->t >> 14) ^ (s->t >> 10) ^ (s->t >> 8) ^ (s->t >> 5) ^ (s->t >> 4)^ (s->t >> 1) ^ s->t))
+       // uint8_t Tt = opt_T(s);
+       uint16_t Tt = s->t & 0xc533;
+       Tt = Tt ^ (Tt >> 1);
+       Tt = Tt ^ (Tt >> 4);
+       Tt = Tt ^ (Tt >> 10);
+       Tt = Tt ^ (Tt >> 8);
+
+       s->t = (s->t >> 1);
+       s->t |= (Tt ^ (s->r >> 7) ^ (s->r >> 3)) << 15;
+
+       uint8_t opt_B = s->b;
+       opt_B ^= s->b >> 6;
+       opt_B ^= s->b >> 5;
+       opt_B ^= s->b >> 4;
+
+       s->b = s->b >> 1;
+       s->b |= (opt_B ^ s->r) << 7;
+
+       uint8_t opt_select = opt_select_LUT[s->r] & 0x04;
+       opt_select |= (opt_select_LUT[s->r] ^ ((Tt ^ y) << 1)) & 0x02;
+       opt_select |= (opt_select_LUT[s->r] ^ Tt) & 0x01;
+
+       uint8_t r = s->r;
+       s->r = (k[opt_select] ^ s->b) + s->l ;
+       s->l = s->r + r;
 }
 
-void opt_suc(const uint8_t *k, State *s, uint8_t *in, uint8_t length, bool add32Zeroes) {
-       State x2;
+static void opt_suc(const uint8_t *k, State *s, uint8_t *in, uint8_t length, bool add32Zeroes) {
        for (int i = 0; i < length; i++) {
                uint8_t head;
-               head = 1 & (in[i] >> 7);
-               opt_successor(k, s, head, &x2);
+               head = in[i];
+               opt_successor(k, s, head);
 
-               head = 1 & (in[i] >> 6);
-               opt_successor(k, &x2, head, s);
+               head >>= 1;
+               opt_successor(k, s, head);
 
-               head = 1 & (in[i] >> 5);
-               opt_successor(k, s, head, &x2);
+               head >>= 1;
+               opt_successor(k, s, head);
 
-               head = 1 & (in[i] >> 4);
-               opt_successor(k, &x2, head, s);
+               head >>= 1;
+               opt_successor(k, s, head);
 
-               head = 1 & (in[i] >> 3);
-               opt_successor(k, s, head, &x2);
+               head >>= 1;
+               opt_successor(k, s, head);
 
-               head = 1 & (in[i] >> 2);
-               opt_successor(k, &x2, head, s);
+               head >>= 1;
+               opt_successor(k, s, head);
 
-               head = 1 & (in[i] >> 1);
-               opt_successor(k, s, head, &x2);
+               head >>= 1;
+               opt_successor(k, s, head);
 
-               head = 1 & in[i];
-               opt_successor(k, &x2, head, s);
+               head >>= 1;
+               opt_successor(k, s, head);
        }
        //For tag MAC, an additional 32 zeroes
        if (add32Zeroes) {
                for(int i = 0; i < 16; i++) {
-                       opt_successor(k, s, 0, &x2);
-                       opt_successor(k, &x2, 0, s);
+                       opt_successor(k, s, 0);
+                       opt_successor(k, s, 0);
                }
        }
 }
 
-void opt_output(const uint8_t *k, State *s,  uint8_t *buffer) {
-       State temp = {0, 0, 0, 0};
+static void opt_output(const uint8_t *k, State *s,  uint8_t *buffer) {
        for (uint8_t times = 0; times < 4; times++) {
                uint8_t bout = 0;
-               bout |= (s->r & 0x4) << 5;
-               opt_successor(k, s, 0, &temp);
-               bout |= (temp.r & 0x4) << 4;
-               opt_successor(k, &temp, 0, s);
-               bout |= (s->r & 0x4) << 3;
-               opt_successor(k, s, 0, &temp);
-               bout |= (temp.r & 0x4) << 2;
-               opt_successor(k, &temp, 0, s);
-               bout |= (s->r & 0x4) << 1;
-               opt_successor(k, s, 0, &temp);
-               bout |= (temp.r & 0x4) ;
-               opt_successor(k, &temp, 0, s);
+               bout |= (s->r & 0x4) >> 2;
+               opt_successor(k, s, 0);
                bout |= (s->r & 0x4) >> 1;
-               opt_successor(k, s, 0, &temp);
-               bout |= (temp.r & 0x4) >> 2;
-               opt_successor(k, &temp, 0, s);
+               opt_successor(k, s, 0);
+               bout |= (s->r & 0x4);
+               opt_successor(k, s, 0);
+               bout |= (s->r & 0x4) << 1;
+               opt_successor(k, s, 0);
+               bout |= (s->r & 0x4) << 2;
+               opt_successor(k, s, 0);
+               bout |= (s->r & 0x4) << 3;
+               opt_successor(k, s, 0);
+               bout |= (s->r & 0x4) << 4;
+               opt_successor(k, s, 0);
+               bout |= (s->r & 0x4) << 5;
+               opt_successor(k, s, 0);
                buffer[times] = bout;
        }
 }
 
-void opt_MAC(uint8_t *k, uint8_t *input, uint8_t *out) {
+static void opt_MAC(uint8_t *k, uint8_t *input, uint8_t *out) {
        State _init  =  {
                ((k[0] ^ 0x4c) + 0xEC) & 0xFF,// l
                ((k[0] ^ 0x4c) + 0x21) & 0xFF,// r
@@ -187,43 +246,22 @@ void opt_MAC(uint8_t *k, uint8_t *input, uint8_t *out) {
        opt_output(k, &_init, out);
 }
 
-uint8_t rev_byte(uint8_t b) {
-       b = (b & 0xF0) >> 4 | (b & 0x0F) << 4;
-       b = (b & 0xCC) >> 2 | (b & 0x33) << 2;
-       b = (b & 0xAA) >> 1 | (b & 0x55) << 1;
-       return b;
-}
-
-void opt_reverse_arraybytecpy(uint8_t *dest, uint8_t *src, size_t len) {
-       for (size_t i = 0; i < len; i++) {
-               dest[i] = rev_byte(src[i]);
-       }
-}
-
 void opt_doReaderMAC(uint8_t *cc_nr_p, uint8_t *div_key_p, uint8_t mac[4]) {
-       static uint8_t cc_nr[12];
-       opt_reverse_arraybytecpy(cc_nr, cc_nr_p, 12);
        uint8_t dest[] = {0, 0, 0, 0, 0, 0, 0, 0};
-       opt_MAC(div_key_p, cc_nr, dest);
-       //The output MAC must also be reversed
-       opt_reverse_arraybytecpy(mac, dest, 4);
+       opt_MAC(div_key_p, cc_nr_p, dest);
+       memcpy(mac, dest, 4);
        return;
 }
 
 void opt_doTagMAC(uint8_t *cc_p, const uint8_t *div_key_p, uint8_t mac[4]) {
-       static uint8_t cc_nr[8+4+4];
-       opt_reverse_arraybytecpy(cc_nr, cc_p, 12);
        State _init = {
                ((div_key_p[0] ^ 0x4c) + 0xEC) & 0xFF,// l
                ((div_key_p[0] ^ 0x4c) + 0x21) & 0xFF,// r
                0x4c, // b
                0xE012 // t
        };
-       opt_suc(div_key_p, &_init,cc_nr, 12, true);
-       uint8_t dest[] = {0, 0, 0, 0};
-       opt_output(div_key_p, &_init, dest);
-       //The output MAC must also be reversed
-       opt_reverse_arraybytecpy(mac, dest, 4);
+       opt_suc(div_key_p, &_init, cc_p, 12, true);
+       opt_output(div_key_p, &_init, mac);
        return;
 }
 
@@ -236,15 +274,13 @@ void opt_doTagMAC(uint8_t *cc_p, const uint8_t *div_key_p, uint8_t mac[4]) {
  * @return the cipher state
  */
 State opt_doTagMAC_1(uint8_t *cc_p, const uint8_t *div_key_p) {
-       static uint8_t cc_nr[8];
-       opt_reverse_arraybytecpy(cc_nr, cc_p, 8);
        State _init = {
                ((div_key_p[0] ^ 0x4c) + 0xEC) & 0xFF,// l
                ((div_key_p[0] ^ 0x4c) + 0x21) & 0xFF,// r
                0x4c, // b
                0xE012 // t
        };
-       opt_suc(div_key_p, &_init, cc_nr, 8, false);
+       opt_suc(div_key_p, &_init, cc_p, 8, false);
        return _init;
 }
 
@@ -258,13 +294,7 @@ State opt_doTagMAC_1(uint8_t *cc_p, const uint8_t *div_key_p) {
  * @param div_key_p - the key to use
  */
 void opt_doTagMAC_2(State _init, uint8_t *nr, uint8_t mac[4], const uint8_t *div_key_p) {
-       static uint8_t _nr[4];
-       opt_reverse_arraybytecpy(_nr, nr, 4);
-       opt_suc(div_key_p, &_init, _nr, 4, true);
-       //opt_suc(div_key_p, &_init,nr, 4, false);
-       uint8_t dest[] = {0, 0, 0, 0};
-       opt_output(div_key_p, &_init, dest);
-       //The output MAC must also be reversed
-       opt_reverse_arraybytecpy(mac, dest, 4);
+       opt_suc(div_key_p, &_init, nr, 4, true);
+       opt_output(div_key_p, &_init, mac);
        return;
 }