diff --git a/attribute.go b/attribute.go
index b2ac945..61112ae 100644
--- a/attribute.go
+++ b/attribute.go
@@ -107,26 +107,27 @@ func MarshalAttributes(attrs []Attribute) ([]byte, error) {
 // It is recommend to use the AttributeDecoder type where possible instead of calling
 // UnmarshalAttributes and using package nlenc functions directly.
 func UnmarshalAttributes(b []byte) ([]Attribute, error) {
-	var attrs []Attribute
-	var i int
-	for {
-		if i > len(b) || len(b[i:]) == 0 {
-			break
-		}
 
-		var a Attribute
-		if err := (&a).unmarshal(b[i:]); err != nil {
-			return nil, err
-		}
+	ad, err := NewAttributeDecoder(b)
+	if err != nil {
+		return nil, err
+	}
 
-		if a.Length == 0 {
-			i += nlaHeaderLen
-			continue
-		}
+	// Return a nil slice when there are no attributes to decode.
+	if ad.Len() == 0 {
+		return nil, nil
+	}
 
-		i += nlaAlign(int(a.Length))
+	attrs := make([]Attribute, 0, ad.Len())
+
+	for ad.Next() {
+		if ad.attr().Length != 0 {
+			attrs = append(attrs, ad.attr())
+		}
+	}
 
-		attrs = append(attrs, a)
+	if err := ad.Err(); err != nil {
+		return nil, err
 	}
 
 	return attrs, nil
@@ -148,10 +149,14 @@ type AttributeDecoder struct {
 	// If not set, the native byte order will be used.
 	ByteOrder binary.ByteOrder
 
-	// The attributes being worked on, and the iterator index into the slice of
-	// attributes.
-	attrs []Attribute
-	i     int
+	// The current attribute being worked on.
+	a Attribute
+
+	// The slice of input bytes and its iterator index.
+	b []byte
+	i int
+
+	length int
 
 	// Any error encountered while decoding attributes.
 	err error
@@ -160,17 +165,21 @@ type AttributeDecoder struct {
 // NewAttributeDecoder creates an AttributeDecoder that unpacks Attributes
 // from b and prepares the decoder for iteration.
 func NewAttributeDecoder(b []byte) (*AttributeDecoder, error) {
-	attrs, err := UnmarshalAttributes(b)
-	if err != nil {
-		return nil, err
-	}
 
-	return &AttributeDecoder{
+	ad := &AttributeDecoder{
 		// By default, use native byte order.
 		ByteOrder: nlenc.NativeEndian(),
 
-		attrs: attrs,
-	}, nil
+		b: b,
+	}
+
+	var err error
+	ad.length, err = ad.available()
+	if err != nil {
+		return nil, err
+	}
+
+	return ad, nil
 }
 
 // Next advances the decoder to the next netlink attribute.  It returns false
@@ -181,34 +190,93 @@ func (ad *AttributeDecoder) Next() bool {
 		return false
 	}
 
-	ad.i++
+	// Exit if array pointer is at or beyond the end of the slice.
+	if ad.i >= len(ad.b) {
+		return false
+	}
+
+	if err := ad.a.unmarshal(ad.b[ad.i:]); err != nil {
+		ad.err = err
+		return false
+	}
+
+	// Advance the pointer by at least one header's length.
+	if int(ad.a.Length) < nlaHeaderLen {
+		ad.i += nlaHeaderLen
+	} else {
+		ad.i += nlaAlign(int(ad.a.Length))
+	}
 
-	// More attributes?
-	return len(ad.attrs) >= ad.i
+	return true
 }
 
 // Type returns the Attribute.Type field of the current netlink attribute
 // pointed to by the decoder.
 //
 // Type masks off the high bits of the netlink attribute type which may contain
-// the Nested and NetByteOrder flags. If you need direct access to these flags,
-// consider using UnmarshalAttributes instead.
+// the Nested and NetByteOrder flags. These can be obtained by calling TypeFlags.
 func (ad *AttributeDecoder) Type() uint16 {
 	// Mask off any flags stored in the high bits.
-	return ad.attr().Type & attrTypeMask
+	return ad.a.Type & attrTypeMask
+}
+
+// TypeFlags returns the two high bits of the Attribute.Type field of the current
+// netlink attribute pointed to by the decoder.
+//
+// These bits of the netlink attribute type are used for the Nested and NetByteOrder
+// flags, available as the Nested and NetByteOrder constants in this package.
+func (ad *AttributeDecoder) TypeFlags() uint16 {
+	return ad.a.Type & ^attrTypeMask
 }
 
 // Len returns the number of netlink attributes pointed to by the decoder.
-func (ad *AttributeDecoder) Len() int { return len(ad.attrs) }
+func (ad *AttributeDecoder) Len() int { return ad.length }
+
+// count scans the input slice to count the number of netlink attributes
+// that could be decoded by Next().
+func (ad *AttributeDecoder) available() (int, error) {
+
+	var i, count int
+	for {
+
+		// No more data to read.
+		if i >= len(ad.b) {
+			break
+		}
+
+		// Make sure there's at least a header's worth
+		// of data to read on each iteration.
+		if len(ad.b[i:]) < nlaHeaderLen {
+			return 0, errInvalidAttribute
+		}
+
+		// Extract the length of the attribute.
+		l := int(nlenc.Uint16(ad.b[i : i+2]))
+
+		// Ignore zero-length attributes.
+		if l != 0 {
+			count++
+		}
+
+		// Advance by at least a header's worth of bytes.
+		if l < nlaHeaderLen {
+			l = nlaHeaderLen
+		}
+
+		i += nlaAlign(l)
+	}
+
+	return count, nil
+}
 
 // attr returns the current Attribute pointed to by the decoder.
 func (ad *AttributeDecoder) attr() Attribute {
-	return ad.attrs[ad.i-1]
+	return ad.a
 }
 
 // data returns the Data field of the current Attribute pointed to by the decoder.
 func (ad *AttributeDecoder) data() []byte {
-	return ad.attr().Data
+	return ad.a.Data
 }
 
 // Err returns the first error encountered by the decoder.
diff --git a/attribute_test.go b/attribute_test.go
index eeecab2..01f9a35 100644
--- a/attribute_test.go
+++ b/attribute_test.go
@@ -754,6 +754,22 @@ func TestAttributeDecoderOK(t *testing.T) {
 				})
 			},
 		},
+		{
+			name: "typeflags",
+			attrs: []Attribute{{
+				Type: 0xffff,
+			}},
+			fn: func(ad *AttributeDecoder) {
+
+				if diff := cmp.Diff(ad.Type(), uint16(0x3fff)); diff != "" {
+					panicf("unexpected Type (-want +got):\n%s", diff)
+				}
+
+				if diff := cmp.Diff(ad.TypeFlags(), uint16(0xc000)); diff != "" {
+					panicf("unexpected TypeFlags (-want +got):\n%s", diff)
+				}
+			},
+		},
 	}
 
 	for _, tt := range tests {