Check for overflow of outlen in blake2{s,b}_final

Fix warnings with -Wcast-qual

Fix blake2{sp,bp}_final's return value

12 files changed, 61 insertions, 52 deletions

diff --git a/ref/blake2-impl.h b/ref/blake2-impl.h
index 83d2dcb..5ac7a43 100644
--- a/ref/blake2-impl.h
+++ b/ref/blake2-impl.h
@@ -19,9 +19,11 @@
 static inline uint32_t load32( const void *src )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  return *( uint32_t * )( src );
+  uint32_t w;
+  memcpy(&w, src, sizeof w);
+  return w;
 #else
-  const uint8_t *p = ( uint8_t * )src;
+  const uint8_t *p = ( const uint8_t * )src;
   uint32_t w = *p++;
   w |= ( uint32_t )( *p++ ) <<  8;
   w |= ( uint32_t )( *p++ ) << 16;
@@ -33,9 +35,11 @@ static inline uint32_t load32( const void *src )
 static inline uint64_t load64( const void *src )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  return *( uint64_t * )( src );
+  uint64_t w;
+  memcpy(&w, src, sizeof w);
+  return w;
 #else
-  const uint8_t *p = ( uint8_t * )src;
+  const uint8_t *p = ( const uint8_t * )src;
   uint64_t w = *p++;
   w |= ( uint64_t )( *p++ ) <<  8;
   w |= ( uint64_t )( *p++ ) << 16;
@@ -51,7 +55,7 @@ static inline uint64_t load64( const void *src )
 static inline void store32( void *dst, uint32_t w )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  *( uint32_t * )( dst ) = w;
+  memcpy(dst, &w, sizeof w);
 #else
   uint8_t *p = ( uint8_t * )dst;
   *p++ = ( uint8_t )w; w >>= 8;
@@ -64,7 +68,7 @@ static inline void store32( void *dst, uint32_t w )
 static inline void store64( void *dst, uint64_t w )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  *( uint64_t * )( dst ) = w;
+  memcpy(dst, &w, sizeof w);
 #else
   uint8_t *p = ( uint8_t * )dst;
   *p++ = ( uint8_t )w; w >>= 8;
@@ -125,7 +129,6 @@ static inline uint64_t rotr64( const uint64_t w, const unsigned c )
 static inline void secure_zero_memory( void *v, size_t n )
 {
   volatile uint8_t *p = ( volatile uint8_t * )v;
-
   while( n-- ) *p++ = 0;
 }
 
diff --git a/ref/blake2b-ref.c b/ref/blake2b-ref.c
index a840bb8..fe27f85 100644
--- a/ref/blake2b-ref.c
+++ b/ref/blake2b-ref.c
@@ -149,7 +149,7 @@ static inline int blake2b_init0( blake2b_state *S )
 int blake2b_init_param( blake2b_state *S, const blake2b_param *P )
 {
   blake2b_init0( S );
-  uint8_t *p = ( uint8_t * )( P );
+  const uint8_t *p = ( const uint8_t * )( P );
 
   /* IV XOR ParamBlock */
   for( size_t i = 0; i < 8; ++i )
@@ -310,7 +310,10 @@ int blake2b_update( blake2b_state *S, const uint8_t *in, uint64_t inlen )
 /* Is this correct? */
 int blake2b_final( blake2b_state *S, uint8_t *out, uint8_t outlen )
 {
-  uint8_t buffer[BLAKE2B_OUTBYTES];
+  uint8_t buffer[BLAKE2B_OUTBYTES] = {0};
+
+  if( outlen > BLAKE2B_OUTBYTES )
+    return -1;
 
   if( S->buflen > BLAKE2B_BLOCKBYTES )
   {
@@ -353,7 +356,7 @@ int blake2b( uint8_t *out, const void *in, const void *key, const uint8_t outlen
     if( blake2b_init( S, outlen ) < 0 ) return -1;
   }
 
-  blake2b_update( S, ( uint8_t * )in, inlen );
+  blake2b_update( S, ( const uint8_t * )in, inlen );
   blake2b_final( S, out, outlen );
   return 0;
 }
diff --git a/ref/blake2bp-ref.c b/ref/blake2bp-ref.c
index b50f21d..16542ef 100644
--- a/ref/blake2bp-ref.c
+++ b/ref/blake2bp-ref.c
@@ -179,8 +179,8 @@ int blake2bp_final( blake2bp_state *S, uint8_t *out, const uint8_t outlen )
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2b_update( S->R, hash[i], BLAKE2B_OUTBYTES );
 
-  blake2b_final( S->R, out, outlen );
-  return 0;
+  
+  return blake2b_final( S->R, out, outlen );
 }
 
 int blake2bp( uint8_t *out, const void *in, const void *key, uint8_t outlen, uint64_t inlen, uint8_t keylen )
@@ -252,8 +252,7 @@ int blake2bp( uint8_t *out, const void *in, const void *key, uint8_t outlen, uin
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2b_update( FS, hash[i], BLAKE2B_OUTBYTES );
 
-  blake2b_final( FS, out, outlen );
-  return 0;
+  return blake2b_final( FS, out, outlen );;
 }
 
 #if defined(BLAKE2BP_SELFTEST)
diff --git a/ref/blake2s-ref.c b/ref/blake2s-ref.c
index 2be229e..e9489fe 100644
--- a/ref/blake2s-ref.c
+++ b/ref/blake2s-ref.c
@@ -142,7 +142,7 @@ static inline int blake2s_init0( blake2s_state *S )
 int blake2s_init_param( blake2s_state *S, const blake2s_param *P )
 {
   blake2s_init0( S );
-  uint32_t *p = ( uint32_t * )( P );
+  const uint32_t *p = ( const uint32_t * )( P );
 
   /* IV XOR ParamBlock */
   for( size_t i = 0; i < 8; ++i )
@@ -299,7 +299,10 @@ int blake2s_update( blake2s_state *S, const uint8_t *in, uint64_t inlen )
 
 int blake2s_final( blake2s_state *S, uint8_t *out, uint8_t outlen )
 {
-  uint8_t buffer[BLAKE2S_OUTBYTES];
+  uint8_t buffer[BLAKE2S_OUTBYTES] = {0};
+
+  if( outlen > BLAKE2S_OUTBYTES )
+    return -1;
 
   if( S->buflen > BLAKE2S_BLOCKBYTES )
   {
@@ -316,7 +319,7 @@ int blake2s_final( blake2s_state *S, uint8_t *out, uint8_t outlen )
 
   for( int i = 0; i < 8; ++i ) /* Output full hash to temp buffer */
     store32( buffer + sizeof( S->h[i] ) * i, S->h[i] );
-
+    
   memcpy( out, buffer, outlen );
   return 0;
 }
@@ -341,7 +344,7 @@ int blake2s( uint8_t *out, const void *in, const void *key, const uint8_t outlen
     if( blake2s_init( S, outlen ) < 0 ) return -1;
   }
 
-  blake2s_update( S, ( uint8_t * )in, inlen );
+  blake2s_update( S, ( const uint8_t * )in, inlen );
   blake2s_final( S, out, outlen );
   return 0;
 }
diff --git a/ref/blake2sp-ref.c b/ref/blake2sp-ref.c
index 230187e..0e30e49 100644
--- a/ref/blake2sp-ref.c
+++ b/ref/blake2sp-ref.c
@@ -177,8 +177,7 @@ int blake2sp_final( blake2sp_state *S, uint8_t *out, const uint8_t outlen )
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2s_update( S->R, hash[i], BLAKE2S_OUTBYTES );
 
-  blake2s_final( S->R, out, outlen );
-  return 0;
+  return blake2s_final( S->R, out, outlen );
 }
 
 
@@ -251,8 +250,7 @@ int blake2sp( uint8_t *out, const void *in, const void *key, uint8_t outlen, uin
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2s_update( FS, hash[i], BLAKE2S_OUTBYTES );
 
-  blake2s_final( FS, out, outlen );
-  return 0;
+  return blake2s_final( FS, out, outlen );
 }
 
 
diff --git a/sse/blake2-impl.h b/sse/blake2-impl.h
index c988a94..16219db 100644
--- a/sse/blake2-impl.h
+++ b/sse/blake2-impl.h
@@ -19,9 +19,11 @@
 static inline uint32_t load32( const void *src )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  return *( uint32_t * )( src );
+  uint32_t w;
+  memcpy(&w, src, sizeof w);
+  return w;
 #else
-  const uint8_t *p = ( uint8_t * )src;
+  const uint8_t *p = ( const uint8_t * )src;
   uint32_t w = *p++;
   w |= ( uint32_t )( *p++ ) <<  8;
   w |= ( uint32_t )( *p++ ) << 16;
@@ -33,9 +35,11 @@ static inline uint32_t load32( const void *src )
 static inline uint64_t load64( const void *src )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  return *( uint64_t * )( src );
+  uint64_t w;
+  memcpy(&w, src, sizeof w);
+  return w;
 #else
-  const uint8_t *p = ( uint8_t * )src;
+  const uint8_t *p = ( const uint8_t * )src;
   uint64_t w = *p++;
   w |= ( uint64_t )( *p++ ) <<  8;
   w |= ( uint64_t )( *p++ ) << 16;
@@ -51,7 +55,7 @@ static inline uint64_t load64( const void *src )
 static inline void store32( void *dst, uint32_t w )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  *( uint32_t * )( dst ) = w;
+  memcpy(dst, &w, sizeof w);
 #else
   uint8_t *p = ( uint8_t * )dst;
   *p++ = ( uint8_t )w; w >>= 8;
@@ -64,7 +68,7 @@ static inline void store32( void *dst, uint32_t w )
 static inline void store64( void *dst, uint64_t w )
 {
 #if defined(NATIVE_LITTLE_ENDIAN)
-  *( uint64_t * )( dst ) = w;
+  memcpy(dst, &w, sizeof w);
 #else
   uint8_t *p = ( uint8_t * )dst;
   *p++ = ( uint8_t )w; w >>= 8;
@@ -125,7 +129,6 @@ static inline uint64_t rotr64( const uint64_t w, const unsigned c )
 static inline void secure_zero_memory( void *v, size_t n )
 {
   volatile uint8_t *p = ( volatile uint8_t * )v;
-
   while( n-- ) *p++ = 0;
 }
 
diff --git a/sse/blake2b-round.h b/sse/blake2b-round.h
index 5dfc1fb..b0dc4bc 100644
--- a/sse/blake2b-round.h
+++ b/sse/blake2b-round.h
@@ -14,10 +14,10 @@
 #ifndef __BLAKE2B_ROUND_H__
 #define __BLAKE2B_ROUND_H__
 
-#define LOAD(p)  _mm_load_si128( (__m128i *)(p) )
+#define LOAD(p)  _mm_load_si128( (const __m128i *)(p) )
 #define STORE(p,r) _mm_store_si128((__m128i *)(p), r)
 
-#define LOADU(p)  _mm_loadu_si128( (__m128i *)(p) )
+#define LOADU(p)  _mm_loadu_si128( (const __m128i *)(p) )
 #define STOREU(p,r) _mm_storeu_si128((__m128i *)(p), r)
 
 #define TOF(reg) _mm_castsi128_ps((reg))
diff --git a/sse/blake2b.c b/sse/blake2b.c
index 1d5ad24..1052264 100644
--- a/sse/blake2b.c
+++ b/sse/blake2b.c
@@ -175,11 +175,10 @@ static inline int blake2b_init0( blake2b_state *S )
 /* init xors IV with input parameter block */
 int blake2b_init_param( blake2b_state *S, const blake2b_param *P )
 {
-  uint8_t *p, *h, *v;
   //blake2b_init0( S );
-  v = ( uint8_t * )( blake2b_IV );
-  h = ( uint8_t * )( S->h );
-  p = ( uint8_t * )( P );
+  const uint8_t * v = ( const uint8_t * )( blake2b_IV );
+  const uint8_t * p = ( const uint8_t * )( P );
+  uint8_t * h = ( uint8_t * )( S->h );
   /* IV XOR ParamBlock */
   memset( S, 0, sizeof( blake2b_state ) );
 
@@ -349,6 +348,9 @@ int blake2b_update( blake2b_state *S, const uint8_t *in, uint64_t inlen )
 
 int blake2b_final( blake2b_state *S, uint8_t *out, uint8_t outlen )
 {
+  if( outlen > BLAKE2B_OUTBYTES )
+    return -1;
+
   if( S->buflen > BLAKE2B_BLOCKBYTES )
   {
     blake2b_increment_counter( S, BLAKE2B_BLOCKBYTES );
@@ -386,7 +388,7 @@ int blake2b( uint8_t *out, const void *in, const void *key, const uint8_t outlen
     if( blake2b_init( S, outlen ) < 0 ) return -1;
   }
 
-  blake2b_update( S, ( uint8_t * )in, inlen );
+  blake2b_update( S, ( const uint8_t * )in, inlen );
   blake2b_final( S, out, outlen );
   return 0;
 }
diff --git a/sse/blake2bp.c b/sse/blake2bp.c
index 644f402..d93784c 100644
--- a/sse/blake2bp.c
+++ b/sse/blake2bp.c
@@ -181,8 +181,7 @@ int blake2bp_final( blake2bp_state *S, uint8_t *out, const uint8_t outlen )
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2b_update( S->R, hash[i], BLAKE2B_OUTBYTES );
 
-  blake2b_final( S->R, out, outlen );
-  return 0;
+  return blake2b_final( S->R, out, outlen );
 }
 
 int blake2bp( uint8_t *out, const void *in, const void *key, uint8_t outlen, uint64_t inlen, uint8_t keylen )
@@ -254,8 +253,7 @@ int blake2bp( uint8_t *out, const void *in, const void *key, uint8_t outlen, uin
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2b_update( FS, hash[i], BLAKE2B_OUTBYTES );
 
-  blake2b_final( FS, out, outlen );
-  return 0;
+  return blake2b_final( FS, out, outlen );
 }
 
 
diff --git a/sse/blake2s-round.h b/sse/blake2s-round.h
index 75008b3..804a3e0 100644
--- a/sse/blake2s-round.h
+++ b/sse/blake2s-round.h
@@ -14,10 +14,10 @@
 #ifndef __BLAKE2S_ROUND_H__
 #define __BLAKE2S_ROUND_H__
 
-#define LOAD(p)  _mm_load_si128( (__m128i *)(p) )
+#define LOAD(p)  _mm_load_si128( (const __m128i *)(p) )
 #define STORE(p,r) _mm_store_si128((__m128i *)(p), r)
 
-#define LOADU(p)  _mm_loadu_si128( (__m128i *)(p) )
+#define LOADU(p)  _mm_loadu_si128( (const __m128i *)(p) )
 #define STOREU(p,r) _mm_storeu_si128((__m128i *)(p), r)
 
 #define TOF(reg) _mm_castsi128_ps((reg))
diff --git a/sse/blake2s.c b/sse/blake2s.c
index 9ec2df1..909745c 100644
--- a/sse/blake2s.c
+++ b/sse/blake2s.c
@@ -164,11 +164,10 @@ static inline int blake2s_init0( blake2s_state *S )
 /* init2 xors IV with input parameter block */
 int blake2s_init_param( blake2s_state *S, const blake2s_param *P )
 {
-  uint8_t *p, *h, *v;
   //blake2s_init0( S );
-  v = ( uint8_t * )( blake2s_IV );
-  h = ( uint8_t * )( S->h );
-  p = ( uint8_t * )( P );
+  const uint8_t * v = ( const uint8_t * )( blake2s_IV );
+  const uint8_t * p = ( const uint8_t * )( P );
+  uint8_t * h = ( uint8_t * )( S->h );
   /* IV XOR ParamBlock */
   memset( S, 0, sizeof( blake2s_state ) );
 
@@ -327,7 +326,10 @@ int blake2s_update( blake2s_state *S, const uint8_t *in, uint64_t inlen )
 /* Is this correct? */
 int blake2s_final( blake2s_state *S, uint8_t *out, uint8_t outlen )
 {
-  uint8_t buffer[BLAKE2S_OUTBYTES];
+  uint8_t buffer[BLAKE2S_OUTBYTES] = {0};
+
+  if( outlen > BLAKE2S_OUTBYTES )
+    return -1;
 
   if( S->buflen > BLAKE2S_BLOCKBYTES )
   {
@@ -370,7 +372,7 @@ int blake2s( uint8_t *out, const void *in, const void *key, const uint8_t outlen
     if( blake2s_init( S, outlen ) < 0 ) return -1;
   }
 
-  blake2s_update( S, ( uint8_t * )in, inlen );
+  blake2s_update( S, ( const uint8_t * )in, inlen );
   blake2s_final( S, out, outlen );
   return 0;
 }
diff --git a/sse/blake2sp.c b/sse/blake2sp.c
index a417576..954670f 100644
--- a/sse/blake2sp.c
+++ b/sse/blake2sp.c
@@ -177,8 +177,7 @@ int blake2sp_final( blake2sp_state *S, uint8_t *out, const uint8_t outlen )
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2s_update( S->R, hash[i], BLAKE2S_OUTBYTES );
 
-  blake2s_final( S->R, out, outlen );
-  return 0;
+  return blake2s_final( S->R, out, outlen );
 }
 
 
@@ -251,8 +250,7 @@ int blake2sp( uint8_t *out, const void *in, const void *key, uint8_t outlen, uin
   for( size_t i = 0; i < PARALLELISM_DEGREE; ++i )
     blake2s_update( FS, hash[i], BLAKE2S_OUTBYTES );
 
-  blake2s_final( FS, out, outlen );
-  return 0;
+  return blake2s_final( FS, out, outlen );
 }
 
 #if defined(BLAKE2SP_SELFTEST)