4 files changed, 30 insertions, 29 deletions

diff --git a/src/crypto/mac/ghash/builtin/ghash_builtin.c b/src/crypto/mac/ghash/builtin/ghash_builtin.c
index 341408f..c518663 100644
--- a/src/crypto/mac/ghash/builtin/ghash_builtin.c
+++ b/src/crypto/mac/ghash/builtin/ghash_builtin.c
@@ -61,8 +61,10 @@ static inline void mulH_a(fastd_block128_t *x, const fastd_mac_state_t *cstate)
 }
 
 
-static fastd_mac_state_t* ghash_init(fastd_context_t *ctx UNUSED, const uint8_t *key) {
-	fastd_mac_state_t *state = malloc(sizeof(fastd_mac_state_t));
+static fastd_mac_state_t* ghash_init(const uint8_t *key) {
+	fastd_mac_state_t *state;
+	if (posix_memalign((void**)&state, 16, sizeof(fastd_mac_state_t)))
+		abort();
 
 	fastd_block128_t Hbase[4];
 	fastd_block128_t Rbase[4];
@@ -105,7 +107,7 @@ static fastd_mac_state_t* ghash_init(fastd_context_t *ctx UNUSED, const uint8_t
 	return state;
 }
 
-static bool ghash_hash(fastd_context_t *ctx UNUSED, const fastd_mac_state_t *state, fastd_block128_t *out, const fastd_block128_t *in, size_t n_blocks) {
+static bool ghash_hash(const fastd_mac_state_t *state, fastd_block128_t *out, const fastd_block128_t *in, size_t n_blocks) {
 	memset(out, 0, sizeof(fastd_block128_t));
 
 	size_t i;
@@ -117,7 +119,7 @@ static bool ghash_hash(fastd_context_t *ctx UNUSED, const fastd_mac_state_t *sta
 	return true;
 }
 
-static void ghash_free(fastd_context_t *ctx UNUSED, fastd_mac_state_t *state) {
+static void ghash_free(fastd_mac_state_t *state) {
 	if (state) {
 		secure_memzero(state, sizeof(*state));
 		free(state);
diff --git a/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.c b/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.c
index 8f1edb0..419d406 100644
--- a/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.c
+++ b/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.c
@@ -34,17 +34,10 @@ static bool ghash_available(void) {
 	return ((fastd_cpuid()&REQ) == REQ);
 }
 
-static void ghash_free(fastd_context_t *ctx UNUSED, fastd_mac_state_t *state) {
-	if (state) {
-		secure_memzero(state, sizeof(*state));
-		free(state);
-	}
-}
-
 const fastd_mac_t fastd_mac_ghash_pclmulqdq = {
 	.available = ghash_available,
 
 	.init = fastd_ghash_pclmulqdq_init,
 	.hash = fastd_ghash_pclmulqdq_hash,
-	.free = ghash_free,
+	.free = fastd_ghash_pclmulqdq_free,
 };
diff --git a/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.h b/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.h
index ccb1ecf..81313a3 100644
--- a/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.h
+++ b/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq.h
@@ -27,10 +27,6 @@
 #include "../../../../crypto.h"
 
 
-struct fastd_mac_state {
-	fastd_block128_t H;
-};
-
-
-fastd_mac_state_t* fastd_ghash_pclmulqdq_init(fastd_context_t *ctx, const uint8_t *key);
-bool fastd_ghash_pclmulqdq_hash(fastd_context_t *ctx, const fastd_mac_state_t *state, fastd_block128_t *out, const fastd_block128_t *in, size_t n_blocks);
+fastd_mac_state_t* fastd_ghash_pclmulqdq_init(const uint8_t *key);
+bool fastd_ghash_pclmulqdq_hash(const fastd_mac_state_t *state, fastd_block128_t *out, const fastd_block128_t *in, size_t n_blocks);
+void fastd_ghash_pclmulqdq_free(fastd_mac_state_t *state);
diff --git a/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq_impl.c b/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq_impl.c
index 134d4ed..201ec9a 100644
--- a/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq_impl.c
+++ b/src/crypto/mac/ghash/pclmulqdq/ghash_pclmulqdq_impl.c
@@ -35,6 +35,11 @@ typedef union vecblock {
 	fastd_block128_t b;
 } vecblock_t;
 
+struct fastd_mac_state {
+	vecblock_t H;
+};
+
+
 static inline __m128i shl(__m128i v, int a) {
 	__m128i tmpl = _mm_slli_epi64(v, a);
 	__m128i tmpr = _mm_srli_epi64(v, 64-a);
@@ -58,18 +63,24 @@ static inline __m128i byteswap(__m128i v) {
 }
 
 
-fastd_mac_state_t* fastd_ghash_pclmulqdq_init(fastd_context_t *ctx UNUSED, const uint8_t *key) {
-	fastd_mac_state_t *state = malloc(sizeof(fastd_mac_state_t));
-
-	vecblock_t h;
-	memcpy(&h, key, sizeof(__m128i));
+fastd_mac_state_t* fastd_ghash_pclmulqdq_init(const uint8_t *key) {
+	fastd_mac_state_t *state;
+	if (posix_memalign((void**)&state, 16, sizeof(fastd_mac_state_t)))
+		abort();
 
-	h.v = byteswap(h.v);
-	state->H = h.b;
+	memcpy(&state->H, key, sizeof(__m128i));
+	state->H.v = byteswap(state->H.v);
 
 	return state;
 }
 
+void fastd_ghash_pclmulqdq_free(fastd_mac_state_t *state) {
+	if (state) {
+		secure_memzero(state, sizeof(*state));
+		free(state);
+	}
+}
+
 static __m128i gmul(__m128i v, __m128i h) {
 	/* multiply */
 	__m128i z0, z1, z2, tmp;
@@ -123,15 +134,14 @@ static __m128i gmul(__m128i v, __m128i h) {
 }
 
 
-bool fastd_ghash_pclmulqdq_hash(fastd_context_t *ctx UNUSED, const fastd_mac_state_t *state, fastd_block128_t *out, const fastd_block128_t *in, size_t n_blocks) {
-	vecblock_t h = {.b = state->H};
+bool fastd_ghash_pclmulqdq_hash(const fastd_mac_state_t *state, fastd_block128_t *out, const fastd_block128_t *in, size_t n_blocks) {
 	vecblock_t v = {.v = _mm_setzero_si128()};
 
 	size_t i;
 	for (i = 0; i < n_blocks; i++) {
 		__m128i b = ((vecblock_t)in[i]).v;
 		v.v = _mm_xor_si128(v.v, byteswap(b));
-		v.v = gmul(v.v, h.v);
+		v.v = gmul(v.v, state->H.v);
 	}
 
 	v.v = byteswap(v.v);