diff --git a/tmk_core/common/uart.c b/tmk_core/common/uart.c
index f2e4bc4f3..412fcf8e1 100644
--- a/tmk_core/common/uart.c
+++ b/tmk_core/common/uart.c
@@ -31,9 +31,41 @@
 
 #include "uart.h"
 
+#if defined(__AVR_ATmega168__) || defined(__AVR_ATmega168P__) || defined(__AVR_ATmega328P__)
+#   define UDRn         UDR0
+#   define UBRRn        UBRR0
+#   define UCSRnA       UCSR0A
+#   define UCSRnB       UCSR0B
+#   define UCSRnC       UCSR0C
+#   define U2Xn         U2X0
+#   define RXENn        RXEN0
+#   define TXENn        TXEN0
+#   define RXCIEn       RXCIE0
+#   define UCSZn1       UCSZ01
+#   define UCSZn0       UCSZ00
+#   define UDRIEn       UDRIE0
+#   define UDRE_vect    USART_UDRE_vect
+#   define RX_vect      USART_RX_vect
+#elif defined(__AVR_ATmega32U4__) || defined(__AVR_ATmega32U2__) || defined(__AVR_AT90USB1286__) || defined(__AVR_AT90USB1287__)
+#   define UDRn         UDR1
+#   define UBRRn        UBRR1
+#   define UCSRnA       UCSR1A
+#   define UCSRnB       UCSR1B
+#   define UCSRnC       UCSR1C
+#   define U2Xn         U2X1
+#   define RXENn        RXEN1
+#   define TXENn        TXEN1
+#   define RXCIEn       RXCIE1
+#   define UCSZn1       UCSZ11
+#   define UCSZn0       UCSZ10
+#   define UDRIEn       UDRIE1
+#   define UDRE_vect    USART1_UDRE_vect
+#   define RX_vect      USART1_RX_vect
+#endif
+
 // These buffers may be any size from 2 to 256 bytes.
 #define RX_BUFFER_SIZE 64
-#define TX_BUFFER_SIZE 40
+#define TX_BUFFER_SIZE 256
 
 static volatile uint8_t tx_buffer[TX_BUFFER_SIZE];
 static volatile uint8_t tx_buffer_head;
@@ -45,10 +77,10 @@ static volatile uint8_t rx_buffer_tail;
 // Initialize the UART
 void uart_init(uint32_t baud) {
     cli();
-    UBRR0          = (F_CPU / 4 / baud - 1) / 2;
-    UCSR0A         = (1 << U2X0);
-    UCSR0B         = (1 << RXEN0) | (1 << TXEN0) | (1 << RXCIE0);
-    UCSR0C         = (1 << UCSZ01) | (1 << UCSZ00);
+    UBRRn          = (F_CPU / 4 / baud - 1) / 2;
+    UCSRnA         = (1 << U2Xn);
+    UCSRnB         = (1 << RXENn) | (1 << TXENn) | (1 << RXCIEn);
+    UCSRnC         = (1 << UCSZn1) | (1 << UCSZn0);
     tx_buffer_head = tx_buffer_tail = 0;
     rx_buffer_head = rx_buffer_tail = 0;
     sei();
@@ -60,12 +92,14 @@ void uart_putchar(uint8_t c) {
 
     i = tx_buffer_head + 1;
     if (i >= TX_BUFFER_SIZE) i = 0;
+	// return immediately to avoid deadlock when interrupt is disabled(called from ISR)
+	if (tx_buffer_tail == i && (SREG & (1<<SREG_I)) == 0) return;
     while (tx_buffer_tail == i)
         ;  // wait until space in buffer
     // cli();
     tx_buffer[i]   = c;
     tx_buffer_head = i;
-    UCSR0B         = (1 << RXEN0) | (1 << TXEN0) | (1 << RXCIE0) | (1 << UDRIE0);
+    UCSRB         = (1 << RXENn) | (1 << TXENn) | (1 << RXCIEn) | (1 << UDRIEn);
     // sei();
 }
 
@@ -95,12 +129,12 @@ uint8_t uart_available(void) {
 }
 
 // Transmit Interrupt
-ISR(USART_UDRE_vect) {
+ISR(UDRE_vect) {
     uint8_t i;
 
     if (tx_buffer_head == tx_buffer_tail) {
         // buffer is empty, disable transmit interrupt
-        UCSR0B = (1 << RXEN0) | (1 << TXEN0) | (1 << RXCIE0);
+        UCSRnB = (1 << RXENn) | (1 << TXENn) | (1 << RXCIEn);
     } else {
         i = tx_buffer_tail + 1;
         if (i >= TX_BUFFER_SIZE) i = 0;
@@ -110,10 +144,10 @@ ISR(USART_UDRE_vect) {
 }
 
 // Receive Interrupt
-ISR(USART_RX_vect) {
+ISR(RX_vect) {
     uint8_t c, i;
 
-    c = UDR0;
+    c = UDRn;
     i = rx_buffer_head + 1;
     if (i >= RX_BUFFER_SIZE) i = 0;
     if (i != rx_buffer_tail) {