This application note contains explanation with examples for 2 distinct topics:

• Data reception with UART and DMA when application does not know size of bytes to receive in advance
• Data transmission with UART and DMA to avoid CPU stalling and use CPU for other purposes

Github supports ToC by default. It is available in the top-left corner of this document.

• DMA: Direct Memory Access controller in STM32
• UART: Universal Asynchronous Receiver Transmitter
• USART: Universal Synchronous Asynchronous Receiver Transmitter
• TX: Transmit
• RX: Receive
• HT: Half-Transfer Complete DMA event/flag
• TC: Transfer Complete DMA event/flag
• RTO: Receiver Timeout UART event/flag
• IRQ: Interrupt

STM32 has peripherals such as USART, UART or LPUART. Difference between them is not relevant for this purpose since concept can be applied to all of them. In few words, USART supports synchronous operation on top of asynchronous (UART) and LPUART supports Low-Power operation in STOP mode. When synchronous mode or low-power mode is not used, USART, UART and LPUART can be consideted identical. For complete set of details, check product's reference manual and datasheet.

For the sake of this application note, we will only use term UART.

UART in STM32 allows configurion using different transmit (TX) and receive (RX) modes:

• Polling mode (no DMA, no IRQ)
  • P: Application is polling for status bits to check if any character has been transmitted/received and read it fast enough in order to not-miss any byte
  • P: Easy to implement, simply few code lines
  • C: Can easily miss received data in complex application if CPU cannot read registers quickly enough
  • C: Works only for low baudrates, 9600 or lower
• Interrupt mode (no DMA)
  • P: UART triggers interrupt and CPU jumps to service routine to handle each received byte separately
  • P: Commonly used approach in embedded applications
  • P: Works well with common baudrates, 115200, up to ~921600 bauds
  • C: Interrupt service routine is executed for every received character
  • C: May decrease system performance if interrupts are triggered for every character for high-speed baudrates
• DMA mode
  • DMA is used to transfer data from USART RX data register to user memory on hardware level. No application interaction is needed at this point except processing received data by application once necessary
  • P: Transfer from USART peripheral to memory is done on hardware level without CPU interaction
  • P: Can work very easily with operating systems
  • P: Optimized for highest baudrates > 1Mbps and low-power applications
  • P: In case of big bursts of data, increasing data buffer size can improve functionality
  • C: Number of bytes to transfer must be known in advance by DMA hardware
  • C: If communication fails, DMA may not notify application about all bytes transferred

This article focuses only on DMA mode for RX operation and explain how to handle unknown data length

Every STM32 has at least one (1) UART IP and at least one (1) DMA controller available in its DNA. This is all we need for successful data transmission. Application uses default features to implement very efficient transmit system using DMA.

While implementation happens to be pretty straight-forward for TX (set pointer to data, define its length and go) operation, this may not be the case for receive. Implementing DMA receive, application should understand number of received bytes to process by DMA before its considered done. However, UART protocol does not offer such information (it could work with higher-level protocol, but that's way another story that we don't touch here. We assume we have to implement very reliable low-level communication protocol).

STM32s have capability in UART to detect when RX line has not been active for period of time. This is achieved using 2 methods:

• IDLE LINE event: Triggered when RX line has been in idle state (normally high state) for 1 frame time, after last received byte. Frame time is based on baudrate. Higher baudrate means lower frame time for single byte.
• RTO (Receiver Timeout) event: Triggered when line has been in idle state for programmable time. It is fully configured by firmware.

Both events can trigger an interrupt which is an essential feature to allow effective receive operation

Not all STM32 have IDLE LINE or RTO features available. When not available, examples concerning these features may not be used.

An example: To transmit 1 byte at 115200 bauds, it takes approximately (for easier estimation) ~100us; for 3 bytes it would be ~300us in total. IDLE line event triggers an interrupt when line has been in idle state for 1 frame time (in this case 100us), after third byte has been received.

This is a real experiment demo using STM32F4 and IDLE LINE event. After IDLE event is triggered, data are echoed back (loopback mode):

• Application receives 3 bytes, takes approx ~300us at 115200 bauds
• RX goes to high state (yellow rectangle) and UART RX detects it has been idle for at least 1 frame time (approx 100us)
  • Width of yellow rectangle represents 1 frame time
• IDLE line interrupt is triggered at green arrow
• Application echoes data back from interrupt context

DMA in STM32 can be configured in normal or circular mode. For each mode, DMA requires number of elements to transfer before its events (half-transfer complete, transfer complete) are triggered.

• Normal mode: DMA starts with data transfer, once it transfers all elements, it stops and sets enable bit to 0.
  • Application is using this mode when transmitting data
• Circular mode: DMA starts with transfer, once it transfers all elements (as written in corresponding length register), it starts from beginning of memory and transfers more
  • Applicaton is using this mode when receiving data

While transfer is active, 2 (among others) interrupts may be triggered:

• Half-Transfer complete HT: Triggers when DMA transfers half count of elements
• Transfer-Complete TC: Triggers when DMA transfers all elements

When DMA operates in circular mode, these interrupts are triggered periodically

Number of elements to transfer by DMA hardware must be written to relevant DMA register before start of transfer

Now it is time to understand which features to use to receive data with UART and DMA to offload CPU. As for the sake of this example, we use memory buffer array of 20 bytes. DMA will transfer data received from UART to this buffer.

Listed are steps to begin. Initial assumption is that UART has been initialized prior reaching this step, same for basic DMA setup, the rest:

• Application writes 20 to relevant DMA register for data length
• Application writes memory & peripheral addresses to relevant DMA registers
• Application sets DMA direction to peripheral-to-memory mode
• Application puts DMA to circular mode. This is to assure DMA does not stop transferring data after it reaches end of memory. Instead, it will roll over and continue with transferring possible more data from UART to memory
• Application enables DMA & UART in reception mode. Receive can not start & DMA will wait UART to receive first character and transmit it to array. This is done for every received byte
• Application is notified by DMA HT event (or interrupt) after first 10 have been transferred from UART to memory
• Application is notified by DMA TC event (or interrupt) after 20 bytes are transferred from UART to memory
• Application is notified by UART IDLE line (or RTO) in case of IDLE line or timeout detected on RX line
• Application needs to reach on all of these events for most efficient receive

This configuration is important as we do not know length in advance. Application needs to assume it may be endless number of bytes received, therefore DMA must be operational endlessly.

We have used 20 bytes long array for demonstration purposes. In real app this size may need to be increased. It all depends on UART baudrate (higher speed, more data may be received in fixed window) and how fast application can process the received data (either using interrupt notification, RTOS, or polling mode)

Everything gets simplier when application transmits data, length of data is known in advance and memory to transmit is ready. For the sake of this example, we use memory for Helloworld message. In C language it would be:

const char
hello_world_arr[] = "HelloWorld";

• Application writes number of bytes to transmit to relevant DMA register, that would be strlen(hello_world_arr) or 10
• Application writes memory & peripheral addresses to relevant DMA registers
• Application sets DMA direction to memory-to-peripheral mode
• Application sets DMA to normal mode. This effectively disables DMA once all the bytes are successfully transferred
• Application enables DMA & UART in transmitter mode. Transmit starts immediately when UART requests first byte via DMA to be shifted to UART TX register
• Application is notified by TC event (or interrupt) after all bytes have been transmitted from memory to UART via DMA
• DMA is stopped and application may prepare next transfer immediately

Please note that TC event is triggered before last UART byte has been fully transmitted over UART. That's because TC event is part of DMA and not part of UART. It is triggered when DMA transfers all the bytes from point A to point B. That is, point A for DMA is memory, point B is UART data register. Now it is up to UART to clock out byte to GPIO pin

This section describes 4 possible cases and one additional which explains why HT/TC events are necessary by application

Abbrevations used for the image:

• R: Read pointer, used by application to read data from memory. Later also used as old_ptr
• W: Write pointer, used by DMA to write next byte to. Increased every time DMA writes new byte. Later also used as new_ptr
• HT: Half-Transfer Complete event triggered by DMA
• TC: Transfer-Complete event - triggered by DMA
• I: IDLE line event - triggered by USART

DMA configuration:

• Circular mode
• 20 bytes data length
  • Consequently HT event gets triggered at 10 bytes being transmitted
  • Consequently TC event gets triggered at 20 bytes being transmitted

Possible cases during real-life execution:

• Case A: DMA transfers 10 bytes. Application gets notification with HT event and may process received data
• Case B: DMA transfers next 10 bytes. Application gets notification thanks to TC event. Processing now starts from last known position until the end of memory
  • DMA is in circular mode, thus it will continue right from beginning of the buffer, on top of the picture
• Case C: DMA transfers 10 bytes, but not aligned with HT nor TC events
  • Application gets notified with HT event when first 6 bytes are transfered. Processing may start from last known read location
  • Application receives IDLE line event after next 4 bytes are successfully transfered to memory
• Case D: DMA transfers 10 bytes in overflow mode and but not aligned with HT nor TC events
  • Application gets notification by TC event when first 4 bytes are transfered. Processing may start from last known read location
  • Application gets notification by IDLE event after next 6 bytes are transfered. Processing may start from beginning of buffer
• Case E: Example what may happen when application relies only on IDLE event
  • If application receives 30 bytes in burst, 10 bytes get overwritten by DMA as application did not process it quickly enough
  • Application gets IDLE line event once there is steady RX line for 1 byte timeframe
  • Red part of data represents first 10 received bytes from burst which were overwritten by last 10 bytes in burst
  • Option to avoid such scenario is to poll for DMA changes quicker than burst of 20 bytes take; or by using TC and HT events

Example code to read data from memory and process it, for cases A-D

/**
 * \brief           Check for new data received with DMA
 *
 * User must select context to call this function from:
 * - Only interrupts (DMA HT, DMA TC, UART IDLE) with same preemption priority level
 * - Only thread context (outside interrupts)
 *
 * If called from both context-es, exclusive access protection must be implemented
 * This mode is not advised as it usually means architecture design problems
 *
 * When IDLE interrupt is not present, application must rely only on thread context,
 * by manually calling function as quickly as possible, to make sure
 * data are read from raw buffer and processed.
 *
 * Not doing reads fast enough may cause DMA to overflow unread received bytes,
 * hence application will lost useful data.
 *
 * Solutions to this are:
 * - Improve architecture design to achieve faster reads
 * - Increase raw buffer size and allow DMA to write more data before this function is called
 */
void
usart_rx_check(void) {
    /*
     * Set old position variable as static.
     *
     * Linker should (with default C configuration) set this variable to `0`.
     * It is used to keep latest read start position,
     * transforming this function to not being reentrant or thread-safe
     */
    static size_t old_pos;
    size_t pos;

    /* Calculate current position in buffer and check for new data available */
    pos = ARRAY_LEN(usart_rx_dma_buffer) - LL_DMA_GetDataLength(DMA1, LL_DMA_CHANNEL_5);
    if (pos != old_pos) {                       /* Check change in received data */
        if (pos > old_pos) {                    /* Current position is over previous one */
            /*
             * Processing is done in "linear" mode.
             *
             * Application processing is fast with single data block,
             * length is simply calculated by subtracting pointers
             *
             * [   0   ]
             * [   1   ] <- old_pos |------------------------------------|
             * [   2   ]            |                                    |
             * [   3   ]            | Single block (len = pos - old_pos) |
             * [   4   ]            |                                    |
             * [   5   ]            |------------------------------------|
             * [   6   ] <- pos
             * [   7   ]
             * [ N - 1 ]
             */
            usart_process_data(&usart_rx_dma_buffer[old_pos], pos - old_pos);
        } else {
            /*
             * Processing is done in "overflow" mode..
             *
             * Application must process data twice,
             * since there are 2 linear memory blocks to handle
             *
             * [   0   ]            |---------------------------------|
             * [   1   ]            | Second block (len = pos)        |
             * [   2   ]            |---------------------------------|
             * [   3   ] <- pos
             * [   4   ] <- old_pos |---------------------------------|
             * [   5   ]            |                                 |
             * [   6   ]            | First block (len = N - old_pos) |
             * [   7   ]            |                                 |
             * [ N - 1 ]            |---------------------------------|
             */
            usart_process_data(&usart_rx_dma_buffer[old_pos], ARRAY_LEN(usart_rx_dma_buffer) - old_pos);
            if (pos > 0) {
                usart_process_data(&usart_rx_dma_buffer[0], pos);
            }
        }
        old_pos = pos;                          /* Save current position as old for next transfers */
    }
}

Thanks to Cortex-M NVIC's (Nested Vectored Interrupt Controller) flexibility, user can configure priority level for each of the NVIC interrupt lines; it has full control over execution profile for each of the interrupt lines separately.

There are 2 priority types in Cortex-M:

• Preemption priority: Interrupt with higher logical priority level can preempt already running lower priority interrupt
• Subpriority: Interrupt with higher subpriority (but same preemption priority) will execute first when 2 (or more) interrupt lines become active at the same time; such interrupt will also never stop currently executed interrupt (if any) by the CPU.

STM32s have different interrupt lines (interrupt service routines later too) for DMA and UART, one for each peripheral and its priority could be software configurable.

Function that gets called to process received data must keep position of last read value, hence processing function is not thread-safe or reentrant and requires special attention.

Application must assure, DMA and UART interrupts utilize same preemption priority level. This is the only configuration to guarantee processing function never gets preempted by itself (DMA interrupt to preempty UART, or opposite), otherwise last-known read position may get corrupted and application will operate with wrong data.

Examples can be used as reference code to implement your own DMA TX and RX functionality.

There are 2 sets of examples:

• Examples for RX only
  • Available in projects folder with usart_rx_ prefix
  • DMA is used to receive data, polling is used to echo data back
• Examples for RX & TX
  • DMA is used to receive data and to transmit data back
  • It uses ring buffer to copy data from DMA buffer to application before it is sent back

Common for all examples:

• Developed in STM32CubeIDE for easier evaluation on STM32 boards
• Fully developed using LL drivers for various STM32 families
• UART common configuration: 115200 bauds, 1 stop bit, no-parity
• DMA RX common configuration: Circular mode, TC and HT events enabled
• DMA TX common configuration: Normal mode, TC event enabled
• All RX examples implement loop-back functionality. Every character received by UART and transfered by DMA is sent back to same UART

• It is possible to run H743 (single-core) examples on dual-core STM32H7 Nucleo boards, NUCLEO-H745 or NUCLEO-H755. Special care needs to be taken as dual-core H7 Nucleo boards use DCDC for MCU power hence application must check clock configuration in main file and uncomment code to enable SMPS.

Examples demonstrate different use cases for RX only or RX&TX combined.

Demos part of this repository are all based on Low-Level (LL) drivers to maximize user understanding - how to convert theory into practice. Some STM32Cube firmware packages include same example using HAL drivers too. Some of them are (with link to example; list is not exhausted) listed below. All examples are identified as UART_ReceptionToIdle_CircularDMA - you can search for it in your local Cube firmware repository.

• STM32U5 UART_ReceptionToIdle_CircularDMA
• STM32L5 UART_ReceptionToIdle_CircularDMA
• STM32G4 UART_ReceptionToIdle_CircularDMA
• STM32F4 UART_ReceptionToIdle_CircularDMA
• STM32G0 UART_ReceptionToIdle_CircularDMA
• STM32L4 UART_ReceptionToIdle_CircularDMA
• STM32WB UART_ReceptionToIdle_CircularDMA
• STM32WL UART_ReceptionToIdle_CircularDMA

• DMA hardware takes care to transfer received data to memory
• Application must constantly poll for new changes in DMA registers and read received data quick enough to make sure DMA will not overwrite data in buffer
• Processing of received data is in thread mode (not in interrupt)
• P: Easy to implement
• P: No interrupts, no consideration of priority and race conditions
• P: Fits for devices without USART IDLE line detection
• C: Application takes care of data periodically
• C: Not possible to put application to low-power mode (sleep mode)

• Same as polling for changes but with dedicated thread in operating system to process data
• P: Easy to implement to RTOS systems, uses single thread without additional RTOS features (no mutexes, semaphores, memory queues)
• P: No interrupts, no consideration of priority and race conditions
• P: Data processing always on-time with maximum delay given by thread delay, thus with known maximum latency between received character and processed time
  • Unless system has higher priority threads
• P: Fits for devices without UART IDLE line detection
• C: Application takes care of data periodically
• C: Uses memory resources dedicated for separate thread for data processing
• C: Not possible to put application to low-power mode (sleep mode)

• Application gets notification by IDLE line detection or DMA TC/HT events
• Application has to process data only when it receives any of the 3 interrupts
• P: Application does not need to poll for new changes
• P: Application receives interrupts on events
• P: Application may enter low-power modes to increase battery life (if operated on battery)
• C: Data are read (processed) in the interrupt. We strive to execute interrupt routine as fast as possible
• C: Long interrupt execution may break other compatibility in the application

Processing of incoming data is from 2 interrupt vectors, hence it is important that they do not preempt each-other. Set both to the same preemption priority!

• Application gets notification by IDLE line detection or DMA TC/HT events
• Application uses separate thread to process the data only when notified in one of interrupts
• P: Processing is not in the interrupt but in separate thread
• P: Interrupt only informs processing thread to process (or to wakeup)
• P: Operating system may put processing thread to blocked state while waiting for event
• C: Memory usage for separate thread message queue (or semaphore)

This is the most preferred way to use and process UART received character

• Application is using DMA in normal mode to transfer data
• Application is always using ringbuffer between high-level write and low-level transmit operation
• DMA TC interrupt is triggered when transfer has finished. Application can then send more data

This is a demo application available in projects folder. Its purpose is to show how can application implement output of debug messages without drastically affect CPU performance. It is using DMA to transfer data (no CPU to wait for UART flags) and can achieve very high or very low data rates

• All debug messages from application are written to intermediate ringbuffer
• Application will try to start & configure DMA after every successfive write to ringbuffer
• If transfer is on-going, next start is configured from DMA TC interrupt

As a result of this demo application for STM32F413-Nucleo board, observations are as following:

• Demo code sends 1581 bytes every second at 115200 bauds, which is approx 142ms.
• With DMA disabled, CPU load was 14%, in-line with time to transmit the data
• With DMA enabled, CPU load was 0%
• DMA can be enabled/disabled with USE_DMA_TX macro configuration in main.c

1. run git clone --recurse-submodules https://github.com/MaJerle/stm32-usart-dma-rx-tx to clone repository including submodules
2. run examples from projects directory using STM32CubeIDE IDE