diff --git a/firmware/hackrf_usb/m0_state.h b/firmware/hackrf_usb/m0_state.h
index 2ce24980..ef28afaf 100644
--- a/firmware/hackrf_usb/m0_state.h
+++ b/firmware/hackrf_usb/m0_state.h
@@ -37,7 +37,8 @@ struct m0_state {
 enum m0_mode {
 	M0_MODE_IDLE = 0,
 	M0_MODE_RX = 1,
-	M0_MODE_TX = 2,
+	M0_MODE_TX_START = 2,
+	M0_MODE_TX_RUN = 3,
 };
 
 /* Address of m0_state is set in ldscripts. If you change the name of this
diff --git a/firmware/hackrf_usb/sgpio_m0.s b/firmware/hackrf_usb/sgpio_m0.s
index b2189c9c..6f4613bb 100644
--- a/firmware/hackrf_usb/sgpio_m0.s
+++ b/firmware/hackrf_usb/sgpio_m0.s
@@ -68,8 +68,8 @@ There are four key code paths, with the following worst-case timings:
 
 RX, normal:     147 cycles
 RX, overrun:    76 cycles
-TX, normal:     133 cycles
-TX, underrun:   140 cycles
+TX, normal:     138 cycles
+TX, underrun:   142 cycles
 
 Design
 ======
@@ -118,7 +118,8 @@ registers and fixed memory addresses.
 // Operating modes.
 .equ MODE_IDLE,                            0
 .equ MODE_RX,                              1
-.equ MODE_TX,                              2
+.equ MODE_TX_START,                        2
+.equ MODE_TX_RUN,                          3
 
 // Our slice chain is set up as follows (ascending data age; arrows are reversed for flow):
 //     L  -> F  -> K  -> C -> J  -> E  -> I  -> A
@@ -179,7 +180,7 @@ main:
 
 idle:
 	// Wait for RX or TX mode to be set.
-	mode .req r0
+	mode .req r3
 	ldr mode, [state, #MODE]                        // mode = state.mode                    // 2
 	cmp mode, #MODE_IDLE                            // if mode == IDLE:                     // 1
 	beq idle                                        //      goto idle                       // 1 thru, 3 taken
@@ -230,7 +231,7 @@ loop:
 	add buf_ptr, buf_base                           // buf_ptr += buf_base                  // 1
 
 	// Load mode.
-	mode .req r0
+	mode .req r3
 	ldr mode, [state, #MODE]                        // mode = state.mode                    // 2
 
 	// Branch according to mode setting.
@@ -255,6 +256,15 @@ direction_tx:
 	sub buf_margin, #32                             // buf_margin -= 32                     // 1
 	bmi tx_zeros                                    // if buf_margin < 0: goto tx_zeros     // 1 thru, 3 taken
 
+	// At this point we know there is TX data available.
+	// If still in TX start mode, switch to TX run.
+	cmp mode, #MODE_TX_START                        // if mode != TX_START:                 // 1
+	bne tx_write                                    //      goto tx_write                   // 1 thru, 3 taken
+	mov mode, #MODE_TX_RUN                          // mode = TX_RUN                        // 1
+	str mode, [state, #MODE]                        // state.mode = mode                    // 2
+
+tx_write:
+
 	// Write data to SGPIO.
 	ldm buf_ptr!, {r0-r3}                           // r0-r3 = buf_ptr[0:16]; buf_ptr += 16 // 5
 	str r0, [sgpio_data, #SLICE0]                   // SGPIO_REG_SS[SLICE0] = r0            // 8
@@ -282,6 +292,10 @@ tx_zeros:
 	str zero, [sgpio_data, #SLICE6]                 // SGPIO_REG_SS[SLICE6] = zero          // 8
 	str zero, [sgpio_data, #SLICE7]                 // SGPIO_REG_SS[SLICE7] = zero          // 8
 
+	// If in TX start mode, don't count this as a shortfall.
+	cmp mode, #MODE_TX_START                        // if mode == TX_START:                 // 1
+	beq loop                                        //      goto loop                       // 1 thru, 3 taken
+
 shortfall:
 
 	// Get current shortfall length from high register.
diff --git a/firmware/hackrf_usb/usb_api_transceiver.c b/firmware/hackrf_usb/usb_api_transceiver.c
index f0f9b137..03d99747 100644
--- a/firmware/hackrf_usb/usb_api_transceiver.c
+++ b/firmware/hackrf_usb/usb_api_transceiver.c
@@ -296,7 +296,7 @@ void transceiver_startup(const transceiver_mode_t mode) {
 		led_off(LED2);
 		led_on(LED3);
 		rf_path_set_direction(&rf_path, RF_PATH_DIRECTION_TX);
-		m0_state.mode = M0_MODE_TX;
+		m0_state.mode = M0_MODE_TX_START;
 		m0_state.shortfall_limit = _tx_underrun_limit;
 		break;
 	default:
diff --git a/host/hackrf-tools/src/hackrf_debug.c b/host/hackrf-tools/src/hackrf_debug.c
index 38e832bd..3d107e06 100644
--- a/host/hackrf-tools/src/hackrf_debug.c
+++ b/host/hackrf-tools/src/hackrf_debug.c
@@ -378,7 +378,7 @@ int write_register(hackrf_device* device, uint8_t part,
 }
 
 static void print_state(hackrf_m0_state *state) {
-	const char *mode_names[] = {"IDLE", "RX", "TX"};
+	const char *mode_names[] = {"IDLE", "RX", "TX_START", "TX_RUN"};
 	const uint32_t num_modes = sizeof(mode_names) / sizeof(mode_names[0]);
 	printf("M0 state:\n");
 	printf("Mode: %u (%s)\n",