diff --git a/make/mcu/STM32F1.mk b/make/mcu/STM32F1.mk
index a6c6a69067..9bdd33cc63 100644
--- a/make/mcu/STM32F1.mk
+++ b/make/mcu/STM32F1.mk
@@ -67,7 +67,7 @@ MCU_COMMON_SRC = \
             drivers/dma.c \
             drivers/inverter.c \
             drivers/light_ws2811strip_stdperiph.c \
-            drivers/serial_uart_init.c \
+            drivers/serial_uart_stdperiph.c \
             drivers/serial_uart_stm32f10x.c \
             drivers/system_stm32f10x.c \
             drivers/timer_stm32f10x.c
diff --git a/make/mcu/STM32F3.mk b/make/mcu/STM32F3.mk
index b1641a6ef0..e498012a53 100644
--- a/make/mcu/STM32F3.mk
+++ b/make/mcu/STM32F3.mk
@@ -86,7 +86,7 @@ MCU_COMMON_SRC = \
             drivers/transponder_ir_io_stdperiph.c \
             drivers/pwm_output_dshot.c \
             drivers/pwm_output_dshot_shared.c \
-            drivers/serial_uart_init.c \
+            drivers/serial_uart_stdperiph.c \
             drivers/serial_uart_stm32f30x.c \
             drivers/system_stm32f30x.c \
             drivers/timer_stm32f30x.c
diff --git a/make/mcu/STM32F4.mk b/make/mcu/STM32F4.mk
index 5dcf4b765e..ea7b0645cc 100644
--- a/make/mcu/STM32F4.mk
+++ b/make/mcu/STM32F4.mk
@@ -180,7 +180,7 @@ MCU_COMMON_SRC = \
             drivers/transponder_ir_io_stdperiph.c \
             drivers/pwm_output_dshot.c \
             drivers/pwm_output_dshot_shared.c \
-            drivers/serial_uart_init.c \
+            drivers/serial_uart_stdperiph.c \
             drivers/serial_uart_stm32f4xx.c \
             drivers/system_stm32f4xx.c \
             drivers/timer_stm32f4xx.c \
diff --git a/make/mcu/STM32F7.mk b/make/mcu/STM32F7.mk
index 925d6ce975..c8f00135b6 100644
--- a/make/mcu/STM32F7.mk
+++ b/make/mcu/STM32F7.mk
@@ -179,13 +179,12 @@ MCU_COMMON_SRC = \
             drivers/timer_hal.c \
             drivers/timer_stm32f7xx.c \
             drivers/system_stm32f7xx.c \
-            drivers/serial_uart_stm32f7xx.c \
-            drivers/serial_uart_hal.c
+            drivers/serial_uart_hal.c \
+            drivers/serial_uart_stm32f7xx.c
 
 MCU_EXCLUDES = \
             drivers/bus_i2c.c \
-            drivers/timer.c \
-            drivers/serial_uart.c
+            drivers/timer.c
 
 MSC_SRC = \
             drivers/usb_msc_f7xx.c \
diff --git a/make/mcu/STM32H7.mk b/make/mcu/STM32H7.mk
index a3ccf14aa0..50e267dd8f 100644
--- a/make/mcu/STM32H7.mk
+++ b/make/mcu/STM32H7.mk
@@ -235,8 +235,8 @@ MCU_COMMON_SRC = \
             drivers/system_stm32h7xx.c \
             drivers/timer_hal.c \
             drivers/timer_stm32h7xx.c \
-            drivers/serial_uart_stm32h7xx.c \
             drivers/serial_uart_hal.c \
+            drivers/serial_uart_stm32h7xx.c \
             drivers/bus_quadspi_hal.c \
             drivers/bus_spi_hal.c \
             drivers/dma_stm32h7xx.c \
@@ -253,8 +253,7 @@ MCU_COMMON_SRC = \
 
 MCU_EXCLUDES = \
             drivers/bus_i2c.c \
-            drivers/timer.c \
-            drivers/serial_uart.c
+            drivers/timer.c
 
 #MSC_SRC = \
 #            drivers/usb_msc_h7xx.c \
diff --git a/src/main/drivers/serial_uart.c b/src/main/drivers/serial_uart.c
index e82e1ca77a..742a2f16b5 100644
--- a/src/main/drivers/serial_uart.c
+++ b/src/main/drivers/serial_uart.c
@@ -33,19 +33,42 @@
 #ifdef USE_UART
 
 #include "build/build_config.h"
-#include "build/atomic.h"
 
 #include "common/utils.h"
 
 #include "drivers/dma.h"
-#include "drivers/inverter.h"
-#include "drivers/nvic.h"
 #include "drivers/rcc.h"
 
 #include "drivers/serial.h"
 #include "drivers/serial_uart.h"
 #include "drivers/serial_uart_impl.h"
 
+serialPort_t *uartOpen(UARTDevice_e device, serialReceiveCallbackPtr rxCallback, void *rxCallbackData, uint32_t baudRate, portMode_e mode, portOptions_e options)
+{   
+    uartPort_t *s = serialUART(device, baudRate, mode, options);
+    
+    if (!s)
+        return (serialPort_t *)s;
+    
+#ifdef USE_DMA
+    s->txDMAEmpty = true;
+#endif
+    
+    // common serial initialisation code should move to serialPort::init()
+    s->port.rxBufferHead = s->port.rxBufferTail = 0;
+    s->port.txBufferHead = s->port.txBufferTail = 0;
+    // callback works for IRQ-based RX ONLY
+    s->port.rxCallback = rxCallback;
+    s->port.rxCallbackData = rxCallbackData;
+    s->port.mode = mode;
+    s->port.baudRate = baudRate;
+    s->port.options = options;
+    
+    uartReconfigure(s);
+    
+    return (serialPort_t *)s;
+}
+
 static void uartSetBaudRate(serialPort_t *instance, uint32_t baudRate)
 {
     uartPort_t *uartPort = (uartPort_t *)instance;
@@ -60,66 +83,26 @@ static void uartSetMode(serialPort_t *instance, portMode_e mode)
     uartReconfigure(uartPort);
 }
 
-void uartTryStartTxDMA(uartPort_t *s)
-{
-    // uartTryStartTxDMA must be protected, since it is called from
-    // uartWrite and handleUsartTxDma (an ISR).
-
-    ATOMIC_BLOCK(NVIC_PRIO_SERIALUART_TXDMA) {
-        if (IS_DMA_ENABLED(s->txDMAResource)) {
-            // DMA is already in progress
-            return;
-        }
-
-        // For F4 (and F1), there are cases that NDTR (CNDTR for F1) is non-zero upon TC interrupt.
-        // We couldn't find out the root cause, so mask the case here.
-        // F3 is not confirmed to be vulnerable, but not excluded as a safety.
-
-        if (xDMA_GetCurrDataCounter(s->txDMAResource)) {
-            // Possible premature TC case.
-            goto reenable;
-        }
-
-        if (s->port.txBufferHead == s->port.txBufferTail) {
-            // No more data to transmit.
-            s->txDMAEmpty = true;
-            return;
-        }
-
-        // Start a new transaction.
-
-#ifdef STM32F4
-        xDMA_MemoryTargetConfig(s->txDMAResource, (uint32_t)&s->port.txBuffer[s->port.txBufferTail], DMA_Memory_0);
-#else
-        DMAx_SetMemoryAddress(s->txDMAResource, (uint32_t)&s->port.txBuffer[s->port.txBufferTail]);
-#endif
-
-        if (s->port.txBufferHead > s->port.txBufferTail) {
-            xDMA_SetCurrDataCounter(s->txDMAResource, s->port.txBufferHead - s->port.txBufferTail);
-            s->port.txBufferTail = s->port.txBufferHead;
-        } else {
-            xDMA_SetCurrDataCounter(s->txDMAResource, s->port.txBufferSize - s->port.txBufferTail);
-            s->port.txBufferTail = 0;
-        }
-        s->txDMAEmpty = false;
-
-    reenable:
-        xDMA_Cmd(s->txDMAResource, ENABLE);
-    }
-}
-
 static uint32_t uartTotalRxBytesWaiting(const serialPort_t *instance)
 {
     const uartPort_t *s = (const uartPort_t*)instance;
 
+#ifdef USE_DMA
     if (s->rxDMAResource) {
+        // XXX Could be consolidated
+#ifdef USE_HAL_DRIVER
+        uint32_t rxDMAHead = __HAL_DMA_GET_COUNTER(s->Handle.hdmarx);
+#else
         uint32_t rxDMAHead = xDMA_GetCurrDataCounter(s->rxDMAResource);
+#endif
+
         if (rxDMAHead >= s->rxDMAPos) {
             return rxDMAHead - s->rxDMAPos;
         } else {
             return s->port.rxBufferSize + rxDMAHead - s->rxDMAPos;
         }
     }
+#endif
 
     if (s->port.rxBufferHead >= s->port.rxBufferTail) {
         return s->port.rxBufferHead - s->port.rxBufferTail;
@@ -140,12 +123,17 @@ static uint32_t uartTotalTxBytesFree(const serialPort_t *instance)
         bytesUsed = s->port.txBufferSize + s->port.txBufferHead - s->port.txBufferTail;
     }
 
+#ifdef USE_DMA
     if (s->txDMAResource) {
         /*
          * When we queue up a DMA request, we advance the Tx buffer tail before the transfer finishes, so we must add
          * the remaining size of that in-progress transfer here instead:
          */
+#ifdef USE_HAL_DRIVER
+        bytesUsed += __HAL_DMA_GET_COUNTER(s->Handle.hdmatx);
+#else
         bytesUsed += xDMA_GetCurrDataCounter(s->txDMAResource);
+#endif
 
         /*
          * If the Tx buffer is being written to very quickly, we might have advanced the head into the buffer
@@ -159,6 +147,7 @@ static uint32_t uartTotalTxBytesFree(const serialPort_t *instance)
             return 0;
         }
     }
+#endif
 
     return (s->port.txBufferSize - 1) - bytesUsed;
 }
@@ -166,9 +155,12 @@ static uint32_t uartTotalTxBytesFree(const serialPort_t *instance)
 static bool isUartTransmitBufferEmpty(const serialPort_t *instance)
 {
     const uartPort_t *s = (const uartPort_t *)instance;
+#ifdef USE_DMA
     if (s->txDMAResource) {
         return s->txDMAEmpty;
-    } else {
+    } else
+#endif
+    {
         return s->port.txBufferTail == s->port.txBufferHead;
     }
 }
@@ -178,11 +170,14 @@ static uint8_t uartRead(serialPort_t *instance)
     uint8_t ch;
     uartPort_t *s = (uartPort_t *)instance;
 
+#ifdef USE_DMA
     if (s->rxDMAResource) {
         ch = s->port.rxBuffer[s->port.rxBufferSize - s->rxDMAPos];
         if (--s->rxDMAPos == 0)
             s->rxDMAPos = s->port.rxBufferSize;
-    } else {
+    } else
+#endif
+    {
         ch = s->port.rxBuffer[s->port.rxBufferTail];
         if (s->port.rxBufferTail + 1 >= s->port.rxBufferSize) {
             s->port.rxBufferTail = 0;
@@ -197,17 +192,26 @@ static uint8_t uartRead(serialPort_t *instance)
 static void uartWrite(serialPort_t *instance, uint8_t ch)
 {
     uartPort_t *s = (uartPort_t *)instance;
+
     s->port.txBuffer[s->port.txBufferHead] = ch;
+
     if (s->port.txBufferHead + 1 >= s->port.txBufferSize) {
         s->port.txBufferHead = 0;
     } else {
         s->port.txBufferHead++;
     }
 
+#ifdef USE_DMA
     if (s->txDMAResource) {
         uartTryStartTxDMA(s);
-    } else {
+    } else
+#endif
+    {
+#ifdef USE_HAL_DRIVER
+        __HAL_UART_ENABLE_IT(&s->Handle, UART_IT_TXE);
+#else
         USART_ITConfig(s->USARTx, USART_IT_TXE, ENABLE);
+#endif
     }
 }
 
@@ -228,75 +232,43 @@ const struct serialPortVTable uartVTable[] = {
     }
 };
 
+#define UART_IRQHandler(type, dev)                            \
+    void type ## dev ## _IRQHandler(void)                     \
+    {                                                         \
+        uartPort_t *s = &(uartDevmap[UARTDEV_ ## dev]->port); \
+        uartIrqHandler(s);                                    \
+    }
+
 #ifdef USE_UART1
-// USART1 Rx/Tx IRQ Handler
-void USART1_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_1]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(USART, 1) // USART1 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART2
-// USART2 Rx/Tx IRQ Handler
-void USART2_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_2]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(USART, 2) // USART2 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART3
-// USART3 Rx/Tx IRQ Handler
-void USART3_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_3]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(USART, 3) // USART3 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART4
-// UART4 Rx/Tx IRQ Handler
-void UART4_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_4]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(UART, 4)  // UART4 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART5
-// UART5 Rx/Tx IRQ Handler
-void UART5_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_5]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(UART, 5)  // UART5 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART6
-// USART6 Rx/Tx IRQ Handler
-void USART6_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_6]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(USART, 6) // USART6 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART7
-// UART7 Rx/Tx IRQ Handler
-void UART7_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_7]->port);
-    uartIrqHandler(s);
-}
+UART_IRQHandler(UART, 7)  // UART7 Rx/Tx IRQ Handler
 #endif
 
 #ifdef USE_UART8
-// UART8 Rx/Tx IRQ Handler
-void UART8_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_8]->port);
-    uartIrqHandler(s);
-}
-#endif
+UART_IRQHandler(UART, 8)  // UART8 Rx/Tx IRQ Handler
 #endif
+
+#endif // USE_UART
diff --git a/src/main/drivers/serial_uart_hal.c b/src/main/drivers/serial_uart_hal.c
index a4e7860fc1..f8fb2c274f 100644
--- a/src/main/drivers/serial_uart_hal.c
+++ b/src/main/drivers/serial_uart_hal.c
@@ -68,19 +68,6 @@ static void usartConfigurePinInversion(uartPort_t *uartPort) {
 
 void uartReconfigure(uartPort_t *uartPort)
 {
-    /*RCC_PeriphCLKInitTypeDef RCC_PeriphClkInit;
-    RCC_PeriphClkInit.PeriphClockSelection = RCC_PERIPHCLK_USART1|RCC_PERIPHCLK_USART2|RCC_PERIPHCLK_USART3|
-            RCC_PERIPHCLK_UART4|RCC_PERIPHCLK_UART5|RCC_PERIPHCLK_USART6|RCC_PERIPHCLK_UART7|RCC_PERIPHCLK_UART8;
-    RCC_PeriphClkInit.Usart1ClockSelection = RCC_USART1CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Usart2ClockSelection = RCC_USART2CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Usart3ClockSelection = RCC_USART3CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Uart4ClockSelection = RCC_UART4CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Uart5ClockSelection = RCC_UART5CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Usart6ClockSelection = RCC_USART6CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Uart7ClockSelection = RCC_UART7CLKSOURCE_SYSCLK;
-    RCC_PeriphClkInit.Uart8ClockSelection = RCC_UART8CLKSOURCE_SYSCLK;
-    HAL_RCCEx_PeriphCLKConfig(&RCC_PeriphClkInit);*/
-
     HAL_UART_DeInit(&uartPort->Handle);
     uartPort->Handle.Init.BaudRate = uartPort->port.baudRate;
     // according to the stm32 documentation wordlen has to be 9 for parity bits
@@ -208,47 +195,6 @@ void uartReconfigure(uartPort_t *uartPort)
     return;
 }
 
-serialPort_t *uartOpen(UARTDevice_e device, serialReceiveCallbackPtr callback, void *callbackData, uint32_t baudRate, portMode_e mode, portOptions_e options)
-{
-    uartPort_t *s = serialUART(device, baudRate, mode, options);
-
-    if (!s) {
-        return (serialPort_t *)s;
-    }
-
-#ifdef USE_DMA
-    s->txDMAEmpty = true;
-#endif
-
-    // common serial initialisation code should move to serialPort::init()
-    s->port.rxBufferHead = s->port.rxBufferTail = 0;
-    s->port.txBufferHead = s->port.txBufferTail = 0;
-    // callback works for IRQ-based RX ONLY
-    s->port.rxCallback = callback;
-    s->port.rxCallbackData = callbackData;
-    s->port.mode = mode;
-    s->port.baudRate = baudRate;
-    s->port.options = options;
-
-    uartReconfigure(s);
-
-    return (serialPort_t *)s;
-}
-
-void uartSetBaudRate(serialPort_t *instance, uint32_t baudRate)
-{
-    uartPort_t *uartPort = (uartPort_t *)instance;
-    uartPort->port.baudRate = baudRate;
-    uartReconfigure(uartPort);
-}
-
-void uartSetMode(serialPort_t *instance, portMode_e mode)
-{
-    uartPort_t *uartPort = (uartPort_t *)instance;
-    uartPort->port.mode = mode;
-    uartReconfigure(uartPort);
-}
-
 #ifdef USE_DMA
 void uartTryStartTxDMA(uartPort_t *s)
 {
@@ -286,209 +232,4 @@ void uartTryStartTxDMA(uartPort_t *s)
 }
 #endif
 
-uint32_t uartTotalRxBytesWaiting(const serialPort_t *instance)
-{
-    uartPort_t *s = (uartPort_t*)instance;
-
-#ifdef USE_DMA
-    if (s->rxDMAResource) {
-        uint32_t rxDMAHead = __HAL_DMA_GET_COUNTER(s->Handle.hdmarx);
-
-        if (rxDMAHead >= s->rxDMAPos) {
-            return rxDMAHead - s->rxDMAPos;
-        } else {
-            return s->port.rxBufferSize + rxDMAHead - s->rxDMAPos;
-        }
-    }
-#endif
-
-    if (s->port.rxBufferHead >= s->port.rxBufferTail) {
-        return s->port.rxBufferHead - s->port.rxBufferTail;
-    } else {
-        return s->port.rxBufferSize + s->port.rxBufferHead - s->port.rxBufferTail;
-    }
-}
-
-uint32_t uartTotalTxBytesFree(const serialPort_t *instance)
-{
-    uartPort_t *s = (uartPort_t*)instance;
-
-    uint32_t bytesUsed;
-
-    if (s->port.txBufferHead >= s->port.txBufferTail) {
-        bytesUsed = s->port.txBufferHead - s->port.txBufferTail;
-    } else {
-        bytesUsed = s->port.txBufferSize + s->port.txBufferHead - s->port.txBufferTail;
-    }
-
-#ifdef USE_DMA
-    if (s->txDMAResource) {
-        /*
-         * When we queue up a DMA request, we advance the Tx buffer tail before the transfer finishes, so we must add
-         * the remaining size of that in-progress transfer here instead:
-         */
-        bytesUsed += __HAL_DMA_GET_COUNTER(s->Handle.hdmatx);
-
-        /*
-         * If the Tx buffer is being written to very quickly, we might have advanced the head into the buffer
-         * space occupied by the current DMA transfer. In that case the "bytesUsed" total will actually end up larger
-         * than the total Tx buffer size, because we'll end up transmitting the same buffer region twice. (So we'll be
-         * transmitting a garbage mixture of old and new bytes).
-         *
-         * Be kind to callers and pretend like our buffer can only ever be 100% full.
-         */
-        if (bytesUsed >= s->port.txBufferSize - 1) {
-            return 0;
-        }
-    }
-#endif
-
-    return (s->port.txBufferSize - 1) - bytesUsed;
-}
-
-bool isUartTransmitBufferEmpty(const serialPort_t *instance)
-{
-    const uartPort_t *s = (uartPort_t *)instance;
-#ifdef USE_DMA
-    if (s->txDMAResource)
-        return s->txDMAEmpty;
-    else
-#endif
-        return s->port.txBufferTail == s->port.txBufferHead;
-}
-
-uint8_t uartRead(serialPort_t *instance)
-{
-    uint8_t ch;
-    uartPort_t *s = (uartPort_t *)instance;
-
-#ifdef USE_DMA
-    if (s->rxDMAResource) {
-        ch = s->port.rxBuffer[s->port.rxBufferSize - s->rxDMAPos];
-        if (--s->rxDMAPos == 0)
-            s->rxDMAPos = s->port.rxBufferSize;
-    } else
-#endif
-    {
-        ch = s->port.rxBuffer[s->port.rxBufferTail];
-        if (s->port.rxBufferTail + 1 >= s->port.rxBufferSize) {
-            s->port.rxBufferTail = 0;
-        } else {
-            s->port.rxBufferTail++;
-        }
-    }
-
-    return ch;
-}
-
-void uartWrite(serialPort_t *instance, uint8_t ch)
-{
-    uartPort_t *s = (uartPort_t *)instance;
-
-    s->port.txBuffer[s->port.txBufferHead] = ch;
-
-    if (s->port.txBufferHead + 1 >= s->port.txBufferSize) {
-        s->port.txBufferHead = 0;
-    } else {
-        s->port.txBufferHead++;
-    }
-
-#ifdef USE_DMA
-    if (s->txDMAResource) {
-        uartTryStartTxDMA(s);
-    } else
-#endif
-    {
-        __HAL_UART_ENABLE_IT(&s->Handle, UART_IT_TXE);
-    }
-}
-
-const struct serialPortVTable uartVTable[] = {
-    {
-        .serialWrite = uartWrite,
-        .serialTotalRxWaiting = uartTotalRxBytesWaiting,
-        .serialTotalTxFree = uartTotalTxBytesFree,
-        .serialRead = uartRead,
-        .serialSetBaudRate = uartSetBaudRate,
-        .isSerialTransmitBufferEmpty = isUartTransmitBufferEmpty,
-        .setMode = uartSetMode,
-        .setCtrlLineStateCb = NULL,
-        .setBaudRateCb = NULL,
-        .writeBuf = NULL,
-        .beginWrite = NULL,
-        .endWrite = NULL,
-    }
-};
-
-#ifdef USE_UART1
-// USART1 Rx/Tx IRQ Handler
-void USART1_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_1]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART2
-// USART2 Rx/Tx IRQ Handler
-void USART2_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_2]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART3
-// USART3 Rx/Tx IRQ Handler
-void USART3_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_3]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART4
-// UART4 Rx/Tx IRQ Handler
-void UART4_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_4]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART5
-// UART5 Rx/Tx IRQ Handler
-void UART5_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_5]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART6
-// USART6 Rx/Tx IRQ Handler
-void USART6_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_6]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART7
-// UART7 Rx/Tx IRQ Handler
-void UART7_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_7]->port);
-    uartIrqHandler(s);
-}
-#endif
-
-#ifdef USE_UART8
-// UART8 Rx/Tx IRQ Handler
-void UART8_IRQHandler(void)
-{
-    uartPort_t *s = &(uartDevmap[UARTDEV_8]->port);
-    uartIrqHandler(s);
-}
-#endif
-#endif
+#endif // USE_UART
diff --git a/src/main/drivers/serial_uart_init.c b/src/main/drivers/serial_uart_stdperiph.c
similarity index 59%
rename from src/main/drivers/serial_uart_init.c
rename to src/main/drivers/serial_uart_stdperiph.c
index 816d464923..b74ddcc1e0 100644
--- a/src/main/drivers/serial_uart_init.c
+++ b/src/main/drivers/serial_uart_stdperiph.c
@@ -37,9 +37,11 @@
 #ifdef USE_UART
 
 #include "build/build_config.h"
+#include "build/atomic.h"
 
 #include "common/utils.h"
 #include "drivers/inverter.h"
+#include "drivers/nvic.h"
 #include "drivers/rcc.h"
 
 #include "drivers/serial.h"
@@ -114,137 +116,140 @@ void uartReconfigure(uartPort_t *uartPort)
         USART_HalfDuplexCmd(uartPort->USARTx, DISABLE);
 
     USART_Cmd(uartPort->USARTx, ENABLE);
-}
-
-serialPort_t *uartOpen(UARTDevice_e device, serialReceiveCallbackPtr rxCallback, void *rxCallbackData, uint32_t baudRate, portMode_e mode, portOptions_e options)
-{
-    uartPort_t *s = serialUART(device, baudRate, mode, options);
-
-    if (!s)
-        return (serialPort_t *)s;
-
-    s->txDMAEmpty = true;
-
-    // common serial initialisation code should move to serialPort::init()
-    s->port.rxBufferHead = s->port.rxBufferTail = 0;
-    s->port.txBufferHead = s->port.txBufferTail = 0;
-    // callback works for IRQ-based RX ONLY
-    s->port.rxCallback = rxCallback;
-    s->port.rxCallbackData = rxCallbackData;
-    s->port.mode = mode;
-    s->port.baudRate = baudRate;
-    s->port.options = options;
-
-    uartReconfigure(s);
 
     // Receive DMA or IRQ
     DMA_InitTypeDef DMA_InitStructure;
-    if (mode & MODE_RX) {
-#ifdef STM32F4
-        if (s->rxDMAResource) {
+    if (uartPort->port.mode & MODE_RX) {
+        if (uartPort->rxDMAResource) {
             DMA_StructInit(&DMA_InitStructure);
-            DMA_InitStructure.DMA_PeripheralBaseAddr = s->rxDMAPeripheralBaseAddr;
+            DMA_InitStructure.DMA_Mode = DMA_Mode_Circular;
+            DMA_InitStructure.DMA_PeripheralBaseAddr = uartPort->rxDMAPeripheralBaseAddr;
             DMA_InitStructure.DMA_Priority = DMA_Priority_Medium;
             DMA_InitStructure.DMA_PeripheralInc = DMA_PeripheralInc_Disable;
             DMA_InitStructure.DMA_PeripheralDataSize = DMA_PeripheralDataSize_Byte;
             DMA_InitStructure.DMA_MemoryInc = DMA_MemoryInc_Enable;
             DMA_InitStructure.DMA_MemoryDataSize = DMA_MemoryDataSize_Byte;
+            DMA_InitStructure.DMA_BufferSize = uartPort->port.rxBufferSize;
+#ifdef STM32F4
             DMA_InitStructure.DMA_FIFOMode = DMA_FIFOMode_Disable ;
             DMA_InitStructure.DMA_FIFOThreshold = DMA_FIFOThreshold_1QuarterFull ;
             DMA_InitStructure.DMA_MemoryBurst = DMA_MemoryBurst_Single ;
             DMA_InitStructure.DMA_PeripheralBurst = DMA_PeripheralBurst_Single;
-#else
-        if (s->rxDMAResource) {
-            DMA_StructInit(&DMA_InitStructure);
-            DMA_InitStructure.DMA_PeripheralBaseAddr = s->rxDMAPeripheralBaseAddr;
-            DMA_InitStructure.DMA_Priority = DMA_Priority_Medium;
-            DMA_InitStructure.DMA_M2M = DMA_M2M_Disable;
-            DMA_InitStructure.DMA_PeripheralInc = DMA_PeripheralInc_Disable;
-            DMA_InitStructure.DMA_PeripheralDataSize = DMA_PeripheralDataSize_Byte;
-            DMA_InitStructure.DMA_MemoryInc = DMA_MemoryInc_Enable;
-            DMA_InitStructure.DMA_MemoryDataSize = DMA_MemoryDataSize_Byte;
-#endif
-            DMA_InitStructure.DMA_BufferSize = s->port.rxBufferSize;
 
-#ifdef STM32F4
-            DMA_InitStructure.DMA_Channel = s->rxDMAChannel;
+            DMA_InitStructure.DMA_Channel = uartPort->rxDMAChannel;
             DMA_InitStructure.DMA_DIR = DMA_DIR_PeripheralToMemory;
-            DMA_InitStructure.DMA_Mode = DMA_Mode_Circular;
-            DMA_InitStructure.DMA_Memory0BaseAddr = (uint32_t)s->port.rxBuffer;
-            xDMA_DeInit(s->rxDMAResource);
-            xDMA_Init(s->rxDMAResource, &DMA_InitStructure);
-            xDMA_Cmd(s->rxDMAResource, ENABLE);
-            USART_DMACmd(s->USARTx, USART_DMAReq_Rx, ENABLE);
-            s->rxDMAPos = xDMA_GetCurrDataCounter(s->rxDMAResource);
+            DMA_InitStructure.DMA_Memory0BaseAddr = (uint32_t)uartPort->port.rxBuffer;
 #else
+            DMA_InitStructure.DMA_M2M = DMA_M2M_Disable;
             DMA_InitStructure.DMA_DIR = DMA_DIR_PeripheralSRC;
-            DMA_InitStructure.DMA_Mode = DMA_Mode_Circular;
-            DMA_InitStructure.DMA_MemoryBaseAddr = (uint32_t)s->port.rxBuffer;
-            xDMA_DeInit(s->rxDMAResource);
-            xDMA_Init(s->rxDMAResource, &DMA_InitStructure);
-            xDMA_Cmd(s->rxDMAResource, ENABLE);
-            USART_DMACmd(s->USARTx, USART_DMAReq_Rx, ENABLE);
-            s->rxDMAPos = xDMA_GetCurrDataCounter(s->rxDMAResource);
+            DMA_InitStructure.DMA_MemoryBaseAddr = (uint32_t)uartPort->port.rxBuffer;
 #endif
+
+            xDMA_DeInit(uartPort->rxDMAResource);
+            xDMA_Init(uartPort->rxDMAResource, &DMA_InitStructure);
+            xDMA_Cmd(uartPort->rxDMAResource, ENABLE);
+            USART_DMACmd(uartPort->USARTx, USART_DMAReq_Rx, ENABLE);
+            uartPort->rxDMAPos = xDMA_GetCurrDataCounter(uartPort->rxDMAResource);
         } else {
-            USART_ClearITPendingBit(s->USARTx, USART_IT_RXNE);
-            USART_ITConfig(s->USARTx, USART_IT_RXNE, ENABLE);
-            USART_ITConfig(s->USARTx, USART_IT_IDLE, ENABLE);
+            USART_ClearITPendingBit(uartPort->USARTx, USART_IT_RXNE);
+            USART_ITConfig(uartPort->USARTx, USART_IT_RXNE, ENABLE);
+            USART_ITConfig(uartPort->USARTx, USART_IT_IDLE, ENABLE);
         }
     }
 
     // Transmit DMA or IRQ
-    if (mode & MODE_TX) {
-#ifdef STM32F4
-        if (s->txDMAResource) {
+    if (uartPort->port.mode & MODE_TX) {
+        if (uartPort->txDMAResource) {
             DMA_StructInit(&DMA_InitStructure);
-            DMA_InitStructure.DMA_PeripheralBaseAddr = s->txDMAPeripheralBaseAddr;
+            DMA_InitStructure.DMA_Mode = DMA_Mode_Normal;
+            DMA_InitStructure.DMA_PeripheralBaseAddr = uartPort->txDMAPeripheralBaseAddr;
             DMA_InitStructure.DMA_Priority = DMA_Priority_Medium;
             DMA_InitStructure.DMA_PeripheralInc = DMA_PeripheralInc_Disable;
             DMA_InitStructure.DMA_PeripheralDataSize = DMA_PeripheralDataSize_Byte;
             DMA_InitStructure.DMA_MemoryInc = DMA_MemoryInc_Enable;
             DMA_InitStructure.DMA_MemoryDataSize = DMA_MemoryDataSize_Byte;
+            DMA_InitStructure.DMA_BufferSize = uartPort->port.txBufferSize;
+
+#ifdef STM32F4
             DMA_InitStructure.DMA_FIFOMode = DMA_FIFOMode_Disable ;
             DMA_InitStructure.DMA_FIFOThreshold = DMA_FIFOThreshold_1QuarterFull ;
             DMA_InitStructure.DMA_MemoryBurst = DMA_MemoryBurst_Single ;
             DMA_InitStructure.DMA_PeripheralBurst = DMA_PeripheralBurst_Single;
+
+            DMA_InitStructure.DMA_Channel = uartPort->txDMAChannel;
+            DMA_InitStructure.DMA_DIR = DMA_DIR_MemoryToPeripheral;
 #else
-        if (s->txDMAResource) {
-            DMA_StructInit(&DMA_InitStructure);
-            DMA_InitStructure.DMA_PeripheralBaseAddr = s->txDMAPeripheralBaseAddr;
-            DMA_InitStructure.DMA_Priority = DMA_Priority_Medium;
             DMA_InitStructure.DMA_M2M = DMA_M2M_Disable;
-            DMA_InitStructure.DMA_PeripheralInc = DMA_PeripheralInc_Disable;
-            DMA_InitStructure.DMA_PeripheralDataSize = DMA_PeripheralDataSize_Byte;
-            DMA_InitStructure.DMA_MemoryInc = DMA_MemoryInc_Enable;
-            DMA_InitStructure.DMA_MemoryDataSize = DMA_MemoryDataSize_Byte;
+            DMA_InitStructure.DMA_DIR = DMA_DIR_PeripheralDST;
 #endif
-            DMA_InitStructure.DMA_BufferSize = s->port.txBufferSize;
+
+            xDMA_DeInit(uartPort->txDMAResource);
+            xDMA_Init(uartPort->txDMAResource, &DMA_InitStructure);
 
 #ifdef STM32F4
-            DMA_InitStructure.DMA_Channel = s->txDMAChannel;
-            DMA_InitStructure.DMA_DIR = DMA_DIR_MemoryToPeripheral;
-            DMA_InitStructure.DMA_Mode = DMA_Mode_Normal;
-            xDMA_DeInit(s->txDMAResource);
-            xDMA_Init(s->txDMAResource, &DMA_InitStructure);
-            xDMA_ITConfig(s->txDMAResource, DMA_IT_TC | DMA_IT_FE | DMA_IT_TE | DMA_IT_DME, ENABLE);
-            xDMA_SetCurrDataCounter(s->txDMAResource, 0);
+            xDMA_ITConfig(uartPort->txDMAResource, DMA_IT_TC | DMA_IT_FE | DMA_IT_TE | DMA_IT_DME, ENABLE);
 #else
-            DMA_InitStructure.DMA_DIR = DMA_DIR_PeripheralDST;
-            DMA_InitStructure.DMA_Mode = DMA_Mode_Normal;
-            xDMA_DeInit(s->txDMAResource);
-            xDMA_Init(s->txDMAResource, &DMA_InitStructure);
-            xDMA_ITConfig(s->txDMAResource, DMA_IT_TC, ENABLE);
-            xDMA_SetCurrDataCounter(s->txDMAResource, 0);
+            xDMA_ITConfig(uartPort->txDMAResource, DMA_IT_TC, ENABLE);
 #endif
-            USART_DMACmd(s->USARTx, USART_DMAReq_Tx, ENABLE);
+
+            xDMA_SetCurrDataCounter(uartPort->txDMAResource, 0);
+            USART_DMACmd(uartPort->USARTx, USART_DMAReq_Tx, ENABLE);
         } else {
-            USART_ITConfig(s->USARTx, USART_IT_TXE, ENABLE);
+            USART_ITConfig(uartPort->USARTx, USART_IT_TXE, ENABLE);
         }
     }
 
-    USART_Cmd(s->USARTx, ENABLE);
+    USART_Cmd(uartPort->USARTx, ENABLE);
+}
 
-    return (serialPort_t *)s;
+#ifdef USE_DMA
+void uartTryStartTxDMA(uartPort_t *s)
+{
+    // uartTryStartTxDMA must be protected, since it is called from
+    // uartWrite and handleUsartTxDma (an ISR).
+
+    ATOMIC_BLOCK(NVIC_PRIO_SERIALUART_TXDMA) {
+        if (IS_DMA_ENABLED(s->txDMAResource)) {
+            // DMA is already in progress
+            return;
+        }
+
+        // For F4 (and F1), there are cases that NDTR (CNDTR for F1) is non-zero upon TC interrupt.
+        // We couldn't find out the root cause, so mask the case here.
+        // F3 is not confirmed to be vulnerable, but not excluded as a safety.
+
+        if (xDMA_GetCurrDataCounter(s->txDMAResource)) {
+            // Possible premature TC case.
+            goto reenable;
+        }
+
+        if (s->port.txBufferHead == s->port.txBufferTail) {
+            // No more data to transmit.
+            s->txDMAEmpty = true;
+            return;
+        }
+
+        // Start a new transaction.
+
+#ifdef STM32F4
+        xDMA_MemoryTargetConfig(s->txDMAResource, (uint32_t)&s->port.txBuffer[s->port.txBufferTail], DMA_Memory_0);
+#else
+        DMAx_SetMemoryAddress(s->txDMAResource, (uint32_t)&s->port.txBuffer[s->port.txBufferTail]);
+#endif
+
+        if (s->port.txBufferHead > s->port.txBufferTail) {
+            xDMA_SetCurrDataCounter(s->txDMAResource, s->port.txBufferHead - s->port.txBufferTail);
+            s->port.txBufferTail = s->port.txBufferHead;
+        } else {
+            xDMA_SetCurrDataCounter(s->txDMAResource, s->port.txBufferSize - s->port.txBufferTail);
+            s->port.txBufferTail = 0;
+        }
+        s->txDMAEmpty = false;
+
+    reenable:
+        xDMA_Cmd(s->txDMAResource, ENABLE);
+    }
 }
 #endif
+
+#endif // USE_UART