merged utp fix from RC_0_16

ChangeLog

@@ -10,6 +10,7 @@
 	* fix uTP edge case where udp socket buffer fills up
 	* fix nagle implementation in uTP
 
+	* fix potential packet allocation alignment issue in utp
 	* make 'close_redudnant_connections' cover more cases
 	* set_piece_deadline() also unfilters the piece (if its priority is 0)
 	* add work-around for bug in windows vista and earlier in GetOverlappedResult

include/libtorrent/string_util.hpp

@@ -54,6 +54,10 @@ namespace libtorrent
 	// in strict ansi mode
 	char* allocate_string_copy(char const* str);
 
+	// returns p + x such that the pointer is 8 bytes aligned
+	// x cannot be greater than 7
+	void* align_pointer(void* p);
+
 }
 
 #endif

src/string_util.cpp

@@ -130,5 +130,18 @@ namespace libtorrent
 		return tmp;
 	}
 
+	// 8-byte align pointer
+	void* align_pointer(void* p)
+	{
+		int offset = uintptr_t(p) & 0x7;
+		// if we're already aligned, don't do anything
+		if (offset == 0) return p;
+		
+		// offset is how far passed the last aligned address
+		// we are. We need to go forward to the next aligned
+		// one. Since aligned addresses are 8 bytes apart, add
+		// 8 - offset.
+		return static_cast<char*>(p) + (8 - offset);
+	}
 }
 

src/utp_stream.cpp

@@ -1682,9 +1682,11 @@ bool utp_socket_impl::send_pkt(int flags)
 			stack_alloced = true;
 #endif
 			TORRENT_ASSERT(force);
-			// TODO: 3 this alloca() statement won't necessarily produce
-			// correctly aligned memory. do something about that
-			p = (packet*)TORRENT_ALLOCA(char, sizeof(packet) + packet_size);
+			// this alloca() statement won't necessarily produce
+			// correctly aligned memory. That's why we ask for 7 more bytes
+			// and adjust our pointer to be aligned later
+			p = (packet*)TORRENT_ALLOCA(char, sizeof(packet) + packet_size + 7);
+			p = (packet*)align_pointer(p);
 			UTP_LOGV("%8p: allocating %d bytes on the stack\n", this, packet_size);
 			p->allocated = packet_size;
 		}