fix 'hf iclass sim'

[proxmark3-svn] / client / cmdhfmfhard.c

diff --git a/client/cmdhfmfhard.c b/client/cmdhfmfhard.c
index b5eabb82b0bd51ab3dfda578897f88b4f5844001..4a379c7129e14a4c948cf2d7a0ef2f3b898d9b6c 100644 (file)
--- a/client/cmdhfmfhard.c
+++ b/client/cmdhfmfhard.c
@@ -25,6 +25,7 @@
 #include <locale.h>
 #include <math.h>
 #include "proxmark3.h"
 #include <locale.h>
 #include <math.h>
 #include "proxmark3.h"

+#include "comms.h"

 #include "cmdmain.h"
 #include "ui.h"
 #include "util.h"
 #include "cmdmain.h"
 #include "ui.h"
 #include "util.h"


@@ -32,7 +33,9 @@
 #include "crapto1/crapto1.h"
 #include "parity.h"
 #include "hardnested/hardnested_bruteforce.h"
 #include "crapto1/crapto1.h"
 #include "parity.h"
 #include "hardnested/hardnested_bruteforce.h"

+#include "hardnested/hardnested_bf_core.h"

 #include "hardnested/hardnested_bitarray_core.h"
 #include "hardnested/hardnested_bitarray_core.h"

+#include "zlib.h"

 
 #define NUM_CHECK_BITFLIPS_THREADS             (num_CPUs())
 #define NUM_REDUCTION_WORKING_THREADS  (num_CPUs())
 
 #define NUM_CHECK_BITFLIPS_THREADS             (num_CPUs())
 #define NUM_REDUCTION_WORKING_THREADS  (num_CPUs())


@@ -40,7 +43,7 @@
 #define IGNORE_BITFLIP_THRESHOLD               0.99    // ignore bitflip arrays which have nearly only valid states
 
 #define STATE_FILES_DIRECTORY                  "hardnested/tables/"
 #define IGNORE_BITFLIP_THRESHOLD               0.99    // ignore bitflip arrays which have nearly only valid states
 
 #define STATE_FILES_DIRECTORY                  "hardnested/tables/"

-#define STATE_FILE_TEMPLATE                            "bitflip_%d_%03" PRIx16 "_states.bin"
+#define STATE_FILE_TEMPLATE                            "bitflip_%d_%03" PRIx16 "_states.bin.z"

 
 #define DEBUG_KEY_ELIMINATION
 // #define DEBUG_REDUCTION
 
 #define DEBUG_KEY_ELIMINATION
 // #define DEBUG_REDUCTION


@@ -70,25 +73,32 @@ static float brute_force_per_second;
 
 
 static void get_SIMD_instruction_set(char* instruction_set) {
 
 
 static void get_SIMD_instruction_set(char* instruction_set) {

-       #if !defined(__APPLE__) || (defined(__APPLE__) && (__clang_major__ > 8))
-               #if (__GNUC__ >= 5) && (__GNUC__ > 5 || __GNUC_MINOR__ > 2)
-       if (__builtin_cpu_supports("avx512f")) strcpy(instruction_set, "AVX512F");
-       else if (__builtin_cpu_supports("avx2")) strcpy(instruction_set, "AVX2");
-               #else 
-       if (__builtin_cpu_supports("avx2")) strcpy(instruction_set, "AVX2");
-               #endif
-       else if (__builtin_cpu_supports("avx")) strcpy(instruction_set, "AVX");
-       else if (__builtin_cpu_supports("sse2")) strcpy(instruction_set, "SSE2");
-       else if (__builtin_cpu_supports("mmx")) strcpy(instruction_set, "MMX");
-       else 
-       #endif
-               strcpy(instruction_set, "unsupported");
+       switch(GetSIMDInstrAuto()) {
+               case SIMD_AVX512:
+                       strcpy(instruction_set, "AVX512F");
+                       break;
+               case SIMD_AVX2:
+                       strcpy(instruction_set, "AVX2");
+                       break;
+               case SIMD_AVX:
+                       strcpy(instruction_set, "AVX");
+                       break;
+               case SIMD_SSE2:
+                       strcpy(instruction_set, "SSE2");
+                       break;
+               case SIMD_MMX:
+                       strcpy(instruction_set, "MMX");
+                       break;
+               default:
+                       strcpy(instruction_set, "no");
+                       break;
+       }       

 }
 
 
 static void print_progress_header(void) {
        char progress_text[80];
 }
 
 
 static void print_progress_header(void) {
        char progress_text[80];

-       char instr_set[12] = "";
+       char instr_set[12] = {0};

        get_SIMD_instruction_set(instr_set);
        sprintf(progress_text, "Start using %d threads and %s SIMD core", num_CPUs(), instr_set);
        PrintAndLog("\n\n");
        get_SIMD_instruction_set(instr_set);
        sprintf(progress_text, "Start using %d threads and %s SIMD core", num_CPUs(), instr_set);
        PrintAndLog("\n\n");


@@ -141,12 +151,6 @@ static inline void set_bit24(uint32_t *bitarray, uint32_t index)
 }
 
 
 }
 
 

-static inline void clear_bit24(uint32_t *bitarray, uint32_t index)
-{
-       bitarray[index>>5] &= ~(0x80000000>>(index&0x0000001f));
-}
-
-

 static inline uint32_t test_bit24(uint32_t *bitarray, uint32_t index)
 {
        return  bitarray[index>>5] & (0x80000000>>(index&0x0000001f));
 static inline uint32_t test_bit24(uint32_t *bitarray, uint32_t index)
 {
        return  bitarray[index>>5] & (0x80000000>>(index&0x0000001f));


@@ -187,40 +191,6 @@ static inline uint32_t next_state(uint32_t *bitarray, uint32_t state)
 }
 
 
 }
 
 

-static inline uint32_t next_not_state(uint32_t *bitarray, uint32_t state)
-{
-       if (++state == 1<<24) return 1<<24;
-       uint32_t index = state >> 5;
-       uint_fast8_t bit = state & 0x1f;
-       uint32_t line = bitarray[index] << bit;
-       while (bit <= 0x1f) {
-               if ((line & 0x80000000) == 0) return state;
-               state++;
-               bit++;
-               line <<= 1;
-       }
-       index++;
-       while (bitarray[index] == 0xffffffff && state < 1<<24) {
-               index++;
-               state += 0x20;
-       }
-       if (state >= 1<<24) return 1<<24;
-#if defined __GNUC__
-       return state + __builtin_clz(~bitarray[index]);
-#else
-       bit = 0x00;
-       line = bitarray[index];
-       while (bit <= 0x1f) {
-               if ((line & 0x80000000) == 0) return state;
-               state++;
-               bit++;
-               line <<= 1;
-       }
-       return 1<<24;
-#endif
-}
-
-

 
 
 #define BITFLIP_2ND_BYTE                               0x0200
 
 
 #define BITFLIP_2ND_BYTE                               0x0200


@@ -240,14 +210,50 @@ static int compare_count_bitflip_bitarrays(const void *b1, const void *b2)
 }
 
 
 }
 
 

+static voidpf inflate_malloc(voidpf opaque, uInt items, uInt size)
+{
+       return malloc(items*size);
+}
+
+
+static void inflate_free(voidpf opaque, voidpf address)
+{
+       free(address);
+}
+
+#define OUTPUT_BUFFER_LEN 80
+#define INPUT_BUFFER_LEN 80
+
+//----------------------------------------------------------------------------
+// Initialize decompression of the respective (HF or LF) FPGA stream 
+//----------------------------------------------------------------------------
+static void init_inflate(z_streamp compressed_stream, uint8_t *input_buffer, uint32_t insize, uint8_t *output_buffer, uint32_t outsize)
+{
+
+       // initialize z_stream structure for inflate:
+       compressed_stream->next_in = input_buffer;
+       compressed_stream->avail_in = insize;
+       compressed_stream->next_out = output_buffer;
+       compressed_stream->avail_out = outsize;
+       compressed_stream->zalloc = &inflate_malloc;
+       compressed_stream->zfree = &inflate_free;
+
+       inflateInit2(compressed_stream, 0);
+       
+}
+
+

 static void init_bitflip_bitarrays(void)
 {
 #if defined (DEBUG_REDUCTION)
        uint8_t line = 0;
 #endif 
 
 static void init_bitflip_bitarrays(void)
 {
 #if defined (DEBUG_REDUCTION)
        uint8_t line = 0;
 #endif 
 

+
+       z_stream compressed_stream;
+       

        char state_files_path[strlen(get_my_executable_directory()) + strlen(STATE_FILES_DIRECTORY) + strlen(STATE_FILE_TEMPLATE) + 1];
        char state_files_path[strlen(get_my_executable_directory()) + strlen(STATE_FILES_DIRECTORY) + strlen(STATE_FILE_TEMPLATE) + 1];

-       char state_file_name[strlen(STATE_FILE_TEMPLATE)];
+       char state_file_name[strlen(STATE_FILE_TEMPLATE)+1];

        
        for (odd_even_t odd_even = EVEN_STATE; odd_even <= ODD_STATE; odd_even++) {
                num_effective_bitflips[odd_even] = 0;
        
        for (odd_even_t odd_even = EVEN_STATE; odd_even <= ODD_STATE; odd_even++) {
                num_effective_bitflips[odd_even] = 0;


@@ -262,22 +268,30 @@ static void init_bitflip_bitarrays(void)
                        if (statesfile == NULL) {
                                continue;
                        } else {
                        if (statesfile == NULL) {
                                continue;
                        } else {

-                               uint32_t *bitset = (uint32_t *)malloc_bitarray(sizeof(uint32_t) * (1<<19));
-                               if (bitset == NULL) {
-                                       printf("Out of memory error in init_bitflip_statelists(). Aborting...\n");
-                                       fclose(statesfile);
-                                       exit(4);
-                               }
-                               size_t bytesread = fread(bitset, 1, sizeof(uint32_t) * (1<<19), statesfile);
-                               if (bytesread != sizeof(uint32_t) * (1<<19)) {
-                                       printf("File read error with %s. Aborting...", state_file_name);
+                               fseek(statesfile, 0, SEEK_END);
+                               uint32_t filesize = (uint32_t)ftell(statesfile);
+                               rewind(statesfile);
+                               uint8_t input_buffer[filesize];
+                               size_t bytesread = fread(input_buffer, 1, filesize, statesfile);
+                               if (bytesread != filesize) {
+                                       printf("File read error with %s. Aborting...\n", state_file_name);

                                        fclose(statesfile);
                                        fclose(statesfile);

-                                       free_bitarray(bitset);

                                        exit(5);
                                }
                                fclose(statesfile);
                                        exit(5);
                                }
                                fclose(statesfile);

-                               uint32_t count = count_states(bitset);
+                               uint32_t count = 0;
+                               init_inflate(&compressed_stream, input_buffer, filesize, (uint8_t *)&count, sizeof(count));
+                               inflate(&compressed_stream, Z_SYNC_FLUSH);

                                if ((float)count/(1<<24) < IGNORE_BITFLIP_THRESHOLD) {
                                if ((float)count/(1<<24) < IGNORE_BITFLIP_THRESHOLD) {

+                                       uint32_t *bitset = (uint32_t *)malloc_bitarray(sizeof(uint32_t) * (1<<19));
+                                       if (bitset == NULL) {
+                                               printf("Out of memory error in init_bitflip_statelists(). Aborting...\n");
+                                               inflateEnd(&compressed_stream);
+                                               exit(4);
+                                       }
+                                       compressed_stream.next_out = (uint8_t *)bitset;
+                                       compressed_stream.avail_out = sizeof(uint32_t) * (1<<19);
+                                       inflate(&compressed_stream, Z_SYNC_FLUSH);

                                        effective_bitflip[odd_even][num_effective_bitflips[odd_even]++] = bitflip;
                                        bitflip_bitarrays[odd_even][bitflip] = bitset;
                                        count_bitflip_bitarrays[odd_even][bitflip] = count;
                                        effective_bitflip[odd_even][num_effective_bitflips[odd_even]++] = bitflip;
                                        bitflip_bitarrays[odd_even][bitflip] = bitset;
                                        count_bitflip_bitarrays[odd_even][bitflip] = count;


@@ -289,9 +303,8 @@ static void init_bitflip_bitarrays(void)
                                                line = 0;
                                        }
 #endif
                                                line = 0;
                                        }
 #endif

-                               } else {
-                                       free_bitarray(bitset);

                                }
                                }

+                               inflateEnd(&compressed_stream);

                        }
                }
                effective_bitflip[odd_even][num_effective_bitflips[odd_even]] = 0x400;  // EndOfList marker
                        }
                }
                effective_bitflip[odd_even][num_effective_bitflips[odd_even]] = 0x400;  // EndOfList marker


@@ -1145,7 +1158,13 @@ static bool timeout(void)
 }
 
 
 }
 
 

-static void *check_for_BitFlipProperties_thread(void *args)
+static void 
+#ifdef __has_attribute
+#if __has_attribute(force_align_arg_pointer)
+__attribute__((force_align_arg_pointer)) 
+#endif
+#endif
+*check_for_BitFlipProperties_thread(void *args)

 {
        uint8_t first_byte = ((uint8_t *)args)[0];
        uint8_t last_byte = ((uint8_t *)args)[1];
 {
        uint8_t first_byte = ((uint8_t *)args)[0];
        uint8_t last_byte = ((uint8_t *)args)[1];


@@ -1893,7 +1912,13 @@ static void init_book_of_work(void)
 }
 
 
 }
 
 

-static void *generate_candidates_worker_thread(void *args)
+static void 
+#ifdef __has_attribute
+#if __has_attribute(force_align_arg_pointer)
+__attribute__((force_align_arg_pointer)) 
+#endif
+#endif
+*generate_candidates_worker_thread(void *args)

 {
        uint16_t *sum_args = (uint16_t *)args;
        uint16_t sum_a0 = sums[sum_args[0]];
 {
        uint16_t *sum_args = (uint16_t *)args;
        uint16_t sum_a0 = sums[sum_args[0]];


@@ -2481,6 +2506,10 @@ static void set_test_state(uint8_t byte)
 int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBlockNo, uint8_t trgKeyType, uint8_t *trgkey, bool nonce_file_read, bool nonce_file_write, bool slow, int tests) 
 {
        char progress_text[80];
 int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBlockNo, uint8_t trgKeyType, uint8_t *trgkey, bool nonce_file_read, bool nonce_file_write, bool slow, int tests) 
 {
        char progress_text[80];

+       
+       char instr_set[12] = {0};
+       get_SIMD_instruction_set(instr_set);
+       PrintAndLog("Using %s SIMD core.", instr_set);

 
        srand((unsigned) time(NULL));
        brute_force_per_second = brute_force_benchmark();
 
        srand((unsigned) time(NULL));
        brute_force_per_second = brute_force_benchmark();


@@ -2549,6 +2578,7 @@ int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBloc
                                best_first_bytes[0] = best_first_byte_smallest_bitarray;
                                pre_XOR_nonces();
                                prepare_bf_test_nonces(nonces, best_first_bytes[0]);
                                best_first_bytes[0] = best_first_byte_smallest_bitarray;
                                pre_XOR_nonces();
                                prepare_bf_test_nonces(nonces, best_first_bytes[0]);

+                               hardnested_print_progress(num_acquired_nonces, "Starting brute force...", expected_brute_force1, 0);

                                key_found = brute_force();
                                free(candidates->states[ODD_STATE]);
                                free(candidates->states[EVEN_STATE]);
                                key_found = brute_force();
                                free(candidates->states[ODD_STATE]);
                                free(candidates->states[EVEN_STATE]);


@@ -2568,6 +2598,7 @@ int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBloc
                                        // printf("Estimated remaining states: %" PRIu64 " (2^%1.1f)\n", nonces[best_first_bytes[0]].sum_a8_guess[j].num_states, log(nonces[best_first_bytes[0]].sum_a8_guess[j].num_states)/log(2.0));
                                        generate_candidates(first_byte_Sum, nonces[best_first_bytes[0]].sum_a8_guess[j].sum_a8_idx);
                                        // printf("Time for generating key candidates list: %1.0f sec (%1.1f sec CPU)\n", difftime(time(NULL), start_time), (float)(msclock() - start_clock)/1000.0);
                                        // printf("Estimated remaining states: %" PRIu64 " (2^%1.1f)\n", nonces[best_first_bytes[0]].sum_a8_guess[j].num_states, log(nonces[best_first_bytes[0]].sum_a8_guess[j].num_states)/log(2.0));
                                        generate_candidates(first_byte_Sum, nonces[best_first_bytes[0]].sum_a8_guess[j].sum_a8_idx);
                                        // printf("Time for generating key candidates list: %1.0f sec (%1.1f sec CPU)\n", difftime(time(NULL), start_time), (float)(msclock() - start_clock)/1000.0);

+                                       hardnested_print_progress(num_acquired_nonces, "Starting brute force...", expected_brute_force, 0);

                                        key_found = brute_force();
                                        free_statelist_cache();
                                        free_candidates_memory(candidates);
                                        key_found = brute_force();
                                        free_statelist_cache();
                                        free_candidates_memory(candidates);


@@ -2608,6 +2639,12 @@ int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBloc
 
                if (nonce_file_read) {          // use pre-acquired data from file nonces.bin
                        if (read_nonce_file() != 0) {
 
                if (nonce_file_read) {          // use pre-acquired data from file nonces.bin
                        if (read_nonce_file() != 0) {

+                               free_bitflip_bitarrays();
+                               free_nonces_memory();
+                               free_bitarray(all_bitflips_bitarray[ODD_STATE]);
+                               free_bitarray(all_bitflips_bitarray[EVEN_STATE]);
+                               free_sum_bitarrays();
+                               free_part_sum_bitarrays();

                                return 3;
                        }
                        hardnested_stage = CHECK_1ST_BYTES | CHECK_2ND_BYTES;
                                return 3;
                        }
                        hardnested_stage = CHECK_1ST_BYTES | CHECK_2ND_BYTES;


@@ -2617,6 +2654,12 @@ int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBloc
                } else {                                        // acquire nonces.
                        uint16_t is_OK = acquire_nonces(blockNo, keyType, key, trgBlockNo, trgKeyType, nonce_file_write, slow);
                        if (is_OK != 0) {
                } else {                                        // acquire nonces.
                        uint16_t is_OK = acquire_nonces(blockNo, keyType, key, trgBlockNo, trgKeyType, nonce_file_write, slow);
                        if (is_OK != 0) {

+                               free_bitflip_bitarrays();
+                               free_nonces_memory();
+                               free_bitarray(all_bitflips_bitarray[ODD_STATE]);
+                               free_bitarray(all_bitflips_bitarray[EVEN_STATE]);
+                               free_sum_bitarrays();
+                               free_part_sum_bitarrays();

                                return is_OK;
                        }
                }
                                return is_OK;
                        }
                }


@@ -2646,10 +2689,11 @@ int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBloc
                        for (statelist_t *sl = candidates; sl != NULL; sl = sl->next) {
                                maximum_states += (uint64_t)sl->len[ODD_STATE] * sl->len[EVEN_STATE];
                        }
                        for (statelist_t *sl = candidates; sl != NULL; sl = sl->next) {
                                maximum_states += (uint64_t)sl->len[ODD_STATE] * sl->len[EVEN_STATE];
                        }

-                       printf("Number of remaining possible keys: %" PRIu64 " (2^%1.1f)\n", maximum_states, log(maximum_states)/log(2.0));
+                       // printf("Number of remaining possible keys: %" PRIu64 " (2^%1.1f)\n", maximum_states, log(maximum_states)/log(2.0));

                        best_first_bytes[0] = best_first_byte_smallest_bitarray;
                        pre_XOR_nonces();
                        prepare_bf_test_nonces(nonces, best_first_bytes[0]);
                        best_first_bytes[0] = best_first_byte_smallest_bitarray;
                        pre_XOR_nonces();
                        prepare_bf_test_nonces(nonces, best_first_bytes[0]);

+                       hardnested_print_progress(num_acquired_nonces, "Starting brute force...", expected_brute_force1, 0);

                        key_found = brute_force();
                        free(candidates->states[ODD_STATE]);
                        free(candidates->states[EVEN_STATE]);
                        key_found = brute_force();
                        free(candidates->states[ODD_STATE]);
                        free(candidates->states[EVEN_STATE]);


@@ -2669,6 +2713,7 @@ int mfnestedhard(uint8_t blockNo, uint8_t keyType, uint8_t *key, uint8_t trgBloc
                                // printf("Estimated remaining states: %" PRIu64 " (2^%1.1f)\n", nonces[best_first_bytes[0]].sum_a8_guess[j].num_states, log(nonces[best_first_bytes[0]].sum_a8_guess[j].num_states)/log(2.0));
                                generate_candidates(first_byte_Sum, nonces[best_first_bytes[0]].sum_a8_guess[j].sum_a8_idx);
                                // printf("Time for generating key candidates list: %1.0f sec (%1.1f sec CPU)\n", difftime(time(NULL), start_time), (float)(msclock() - start_clock)/1000.0);
                                // printf("Estimated remaining states: %" PRIu64 " (2^%1.1f)\n", nonces[best_first_bytes[0]].sum_a8_guess[j].num_states, log(nonces[best_first_bytes[0]].sum_a8_guess[j].num_states)/log(2.0));
                                generate_candidates(first_byte_Sum, nonces[best_first_bytes[0]].sum_a8_guess[j].sum_a8_idx);
                                // printf("Time for generating key candidates list: %1.0f sec (%1.1f sec CPU)\n", difftime(time(NULL), start_time), (float)(msclock() - start_clock)/1000.0);

+                               hardnested_print_progress(num_acquired_nonces, "Starting brute force...", expected_brute_force, 0);

                                key_found = brute_force();
                                free_statelist_cache();
                                free_candidates_memory(candidates);
                                key_found = brute_force();
                                free_statelist_cache();
                                free_candidates_memory(candidates);