drivers: serial: stm32: Data loss & possible hardfault

[HelligeChris]

Describe the bug
The async_user_callback can be triggered from two places one from the DMA transfer complete interrupt generated by the DMA and from an k_work queue timeout.
This timeout is not in a ISR context, which means the DMA transfer complete event can interrupt the timeout, resulting in data not in sequence and a DMA buffer offset and length that can be changed while the timeout is using these variables.

This can be seen on the callstack, note the timeout is mid copy (memcpy) when the interrupt triggeres;

To Reproduce
We've used this callback handler to debug the issue simply be setting a breakpoint at LOG_ERR("UART FAULT"), the variables bad_message_last_buffer_ptr and expected_new_offset needs to be stored outside the callback.
From what I can see this issue is a race condition, for example if the timeout triggers when the DMA buffer is almost filled (missing 1 byte or something) and the last bytes are received, the DMA transfer complete interrupt will trigger and interrupt the timeout.

static void dma_tx_done_callback(const struct device *dev, struct uart_event *evt, void *user_data) {
        auto *data  = static_cast<UART::dma_callback_data *>(user_data);
        int32_t err = 0;

        switch (evt->type) {
            case UART_RX_RDY: {
                IO::Message message;

                uint8_t *dataLocation = &evt->data.rx.buf[evt->data.rx.offset];

                if (data->bad_message_last_buffer_ptr != evt->data.rx.buf) {
                    data->expected_new_offset = 0U;
                }
                data->bad_message_last_buffer_ptr = evt->data.rx.buf;

                if (data->expected_new_offset != evt->data.rx.offset) {
                    LOG_ERR("UART Fault");
                } else {
                    data->expected_new_offset += evt->data.rx.len;

                    memcpy(message.data.begin(), dataLocation, evt->data.rx.len);
                    message.length = evt->data.rx.len;

                    err = k_msgq_put(&data->recv_buffer, &message, K_NO_WAIT);
                    if (err != 0) {
                        data->uart->rxError = err;
                    }
                }

                err = k_poll_signal_raise(&data->recv_signal, 1U);
                if (err != 0) {
                    data->uart->rxError = err;
                }

            } break;

// Rest of the callback...

Expected behavior
The timeout cannot be interrupted by DMA transfer complete.

Impact
For now we've had to patch the uart_stm32.c driver with an irq_lock/irq_unlock in and start and end of the uart_stm32_async_rx_timeout function;

static void uart_stm32_async_rx_timeout(struct k_work *work)
{
	// Added to stop the dma transfer complete interrupt for interrupting while this is running,
	// Since they share data 
	unsigned int key = irq_lock();

	struct k_work_delayable *dwork = k_work_delayable_from_work(work);
	struct uart_dma_stream *rx_stream =
		CONTAINER_OF(dwork, struct uart_dma_stream, timeout_work);
	struct uart_stm32_data *data = CONTAINER_OF(rx_stream, struct uart_stm32_data, dma_rx);
	const struct device *dev = data->uart_dev;

	LOG_DBG("rx timeout");

	if (data->dma_rx.counter == data->dma_rx.buffer_length) {
		uart_stm32_async_rx_disable(dev);
	} else {
		uart_stm32_dma_rx_flush(dev);
	}
 
	irq_unlock(key);
}

However, without this the driver is completely broken i.e. can lead to hard faults (we've seen the event->data.rx.offset be int max, which when used with a memcpy causes a hardfault) or simply just lost data.

Environment (please complete the following information):

• OS: Linux
• Toolchain GCC 12.3.1
• Zephyr 3.6.0

Additional context
This was tested on the STM32H743

[erwango]

@djiatsaf-st ^^

[djiatsaf-st]

Hi @HelligeChris ,
Thank you for your contributions.

I looked on my side, I have some points to suggest hoping that it can be useful to you.

Already I checked if there is a possibility that a hardfault occurs when the offset reaches the max in combination with the memcpy function. I implemented a test similar to the situation and then I checked the status of the CFSR register (Configurable Fault Status Register), its value is 0x00, which means that there is no any type of hardfault. You can also use the fault analyzer in STM32CubeIDE, the result will be the same.

However rather than using the memcpy() function, you can implement your own, and there you will get results.

I used this function instead of memcpy() for example:

void copyArray(uint8_t *src, uint8_t *dest, uint8_t offset, uint8_t size) { 

            for (int i = offset; i < size; i++) { 

                          *(dest + i) = *(src + i);  

                  } 

    } 

(You can modify it to suit your needs if necessary)

As for the data loss and the race condition, I could not fully reproduce your situation, I would need a little more information on the configurations and the use case you want to implement that pushed you to raise this problem.

Also, I tested your patch, I added your changes by blocking interrupts, then I did a test to see if I would receive data in a good order as you want, and for the moment I did not have the operation you expect, everything was rather "normal" in the communication as usual.

So from my analysis I think that the Zephyr API on asynchronous serial communication is done like this (it seems that it is not directly a FIFO buffer) and it would therefore be necessary to set up a mechanism or a function that processes the data from the buffer well.

In zephyr there is a special mechanism implemented to continuously receive data in case there is an overflow (event->data.rx.offset be int max), that could be a possible solution to the problem too. There is a request made by the Zephyr API driver (uart event type = UART_RX_BUF_REQUEST) to have a second rx buffer.

Here is the link :

https://docs.zephyrproject.org/latest/doxygen/html/group__uart__async.html#ggaf0c9513cbafa36d86b4c83f2b6a03dcda0d250f372702526f1bce6d4dfe166abe

we will have a configuration a bit like this:

case UART_RX_BUF_REQUEST: 
              uart_rx_buf_rsp(dev, rx_second_buffer,sizeof(rx_second_buffer)); 
      break; 

The global array rx_second_buffer will contain the data history for the defined size.

[HelligeChris]

Hi @djiatsaf-st,

The core issue is the race condition between the timeout (uart_stm32_async_rx_timeout) and actual DMA interrupt (dma_stm32_irq_handler in dma_stm32.c, calls uart_stm32_dma_rx_cb in uart_stm32.c). Since both callbacks makes changes to the variable data->dma_rx.counter, depending on when and if the DMA interrupt interrupts the timeout the data->rx.len can go "negativ" though since the type is size_t it would be 4,294,967,295. This is what we've seen lead to hardfaults.

Already I checked if there is a possibility that a hardfault occurs when the offset reaches the max in combination with the memcpy function. I implemented a test similar to the situation and then I checked the status of the CFSR register (Configurable Fault Status Register), its value is 0x00, which means that there is no any type of hardfault. You can also use the fault analyzer in STM32CubeIDE, the result will be the same.

By max do you mean the size of the buffer or the size_t max size?

As for the data loss and the race condition, I could not fully reproduce your situation, I would need a little more information on the configurations and the use case you want to implement that pushed you to raise this problem.

In our application we have two different controllers with serial communication between them, where one acts as a sensor while the other periodically and asynchronously gets the newest sensor value. The UART rx is enabled like this:

rc = uart_rx_enable(
                this->device,
                this->rx_buffer.at(this->cb_data.rx_buffer_chosen).begin(),
                this->rx_buffer.at(this->cb_data.rx_buffer_chosen).max_size(),
                1000
            );

So, we expect the driver to timeout after 1 ms with no new data received. The problem occurs in the case where the timeout is triggered and the dma rx buffer is almost full, while the timeout callback is running more bytes is received and the dma rx buffer is completely filled triggering the DMA interrupt. Since the DMA interrupt is an ISR the timeout callback is "paused" and the DMA interrupt ISR is called. After the DMA interrupt ISR is finished the timeout callback continues where it left off. This leads to data received out of order/not in sequence.

Also, I tested your patch, I added your changes by blocking interrupts, then I did a test to see if I would receive data in a good order as you want, and for the moment I did not have the operation you expect, everything was rather "normal" in the communication as usual.

Just to be clear, by "normal" do you mean data is received sequentially? Since this is also the operation I expect after applying the patch.

In zephyr there is a special mechanism implemented to continuously receive data in case there is an overflow (event->data.rx.offset be int max), that could be a possible solution to the problem too. There is a request made by the Zephyr API driver (uart event type = UART_RX_BUF_REQUEST) to have a second rx buffer.

I believe we already use this mechanism in our use case, here's our complete callback:

    static void dma_tx_done_callback(const struct device *dev, struct uart_event *evt, void *user_data) {
        auto *data  = static_cast<UART::dma_callback_data *>(user_data);
        int32_t err = 0;

        switch (evt->type) {
            case UART_RX_RDY: {
                IO::Message message;

                uint8_t *dataLocation = &evt->data.rx.buf[evt->data.rx.offset];

                memcpy(message.data.begin(), dataLocation, evt->data.rx.len);
                message.length = evt->data.rx.len;

                err = k_msgq_put(&data->recv_buffer, &message, K_NO_WAIT);
                if (err != 0) {
                    data->uart->rxError = err;
                }

                err = k_poll_signal_raise(&data->recv_signal, 1U);
                if (err != 0) {
                    data->uart->rxError = err;
                }
            } break;
            case UART_TX_ABORTED:
                [[fallthrough]];
            case UART_TX_DONE: {
                k_sem_give(&data->tx_sem);
            } break;
            case UART_RX_BUF_REQUEST: {
                data->uart->next_dma_buffer();
                err = uart_rx_buf_rsp(
                    dev,
                    data->uart->rx_buffer.at(data->rx_buffer_chosen).begin(),
                    data->uart->rx_buffer.at(data->rx_buffer_chosen).max_size()
                );
                if (err == -EACCES) {
                    data->uart->rx_enable(true);
                }
            } break;
            case UART_RX_BUF_RELEASED:
                break;
            case UART_RX_DISABLED: {
                if (data->uart->get_rx_enabled()) {
                    data->uart->rx_enable(true);
                }
            } break;
            default:
                break;
        }
    }

[djiatsaf-st]

Thanks for your answers

By max do you mean the size of the buffer or the size_t max size?

i mean a buffer size

Just to be clear, by "normal" do you mean data is received sequentially? Since this is also the operation I expect after applying the patch.

No, "normal" does not mean that the data will be received sequentially, that's why I put normal in quotes. I meant by normal, since I use uart communication on zephyr, I always had to make some modifications on the buffer data to have the right values.

For the rest, I took note and I understand well the use case and the issue.

I have to do some tests on my side and check if the patch does not impact the current state of the driver.

If I have other questions I will come back to you.

[github-actions]

This issue has been marked as stale because it has been open (more than) 60 days with no activity. Remove the stale label or add a comment saying that you would like to have the label removed otherwise this issue will automatically be closed in 14 days. Note, that you can always re-open a closed issue at any time.

[vestom]

Please do not close this issue, if it has not been fixed in newest Zephyr. This issue is real and causing data loss/mangling when using the STM32 UART in DMA mode. Until fixed we will have to manually implement the fix for each Zephyr update to our production code.

The fix is literally a two line patch of irq_lock()/irq_unlock() as shown above. Also the need for the fix can be realized by inspecting the code and seeing the race between the DMA transfer complete interrupt and the k_work queue timeout. Other async functions in the driver already has a similar irq_lock()/irq_unlock() protection.