Serial port communication failure

1. DMA Failure causes serial communication failure

1.1 background

        Android The application layer passes through write Methods write serial port and MCU signal communication , If not received MCU Of ack Then rewrite the previously written information , Until received MCU Of ack;

        bug Phenomenon is ：Android The application layer has not received ack And constantly rewrite ,2 Minutes later write Method stuck .

1.2 analysis

         After source code analysis , Think serial Layer of write Function may cause sleep , One of the buffer spaces is 4096 byte , The upper level passes through log Statistical discovery of the sent data , When data is sent accumulatively 4102 Bytes will get stuck .

         Add log Reproduce and analyze , give the result as follows ：

          according to log You know , stay write When the method gets stuck, the last batch log Show uart_write The return value of the method is 0, If the upper application is blocking access , It will call schedule Method hibernate the current process, that is, the upper intuitive feeling is stuck .

         Reference resources log Analyze the source code to know ,serial The ring buffer space of layer is full, resulting in bug Appearance ,serial Layer data to hardware TX_FIFO By DMA To carry , So preliminary estimates DMA There is a problem , towards SoC After consulting the manufacturer, the other party indicated , The SoC Of DMA There are some flaws .

         Continue to add log Find out , exactly DMA There is a failure of not handling data , give the result as follows ：

2.  close DMA After that, the serial communication fails

2.1  background

         Since last shutdown DMA in the future , In a specific scenario, the serial port communication with high frequency fails , The basic process is MCU to SoC Send a message , Then a frame of message will be lost , Which leads to Android Application upper layer verification failed , Judge communication failure .

         For various reasons MCU The end shields the retransmission mechanism for this situation , So we need to SoC Look for reasons and solutions .

         One noteworthy phenomenon is ： The application layer finds , When the data read at one time is 32 Byte time , The next frame of data will be lost .

2.2 analysis

        SoC Serial port hardware RX_FIFO Space, too 32 byte , The threshold set in the driver code is 16 byte , namely RX_FIFO In more than 16 Byte will trigger interrupt , remind CPU Transport data to tty layer , Combined with the phenomenon found in the application layer , The preliminary conclusion is RX_FIFO overflow .

        SoC The serial port controller of has overflow flag bit , Add log Print , When communication failure is finally confirmed RX_FIFO It does overflow .

         The basic judgment is UART Hardware interrupt priority is too low, resulting in delayed interrupt response , Which leads to RX_FIFO overflow .

2.3  terms of settlement

2.3.1 Adjust the threshold

         Adjust the threshold of triggering serial port interrupt , There is some improvement after adjustment , But the frequency of overflow is still high , So this method pass fall .

2.3.2  Modify interrupt priority

         Modifying interrupt priority has a wide range of influence , And SoC The manufacturer does not recommend modifying the interrupt priority , It is suggested to modify the interruption of kinship , So this method pass fall .

2.3.3 Modify interrupt affinity .

Mode one ：

//1.  Check the number corresponding to the interrupt 
cat /proc/interrupts 
  ······
  59:        205          0       GIC  21e8000.serial
  
//2.  Change the interrupt affinity of serial port from CPU0 Change to CPU1  
echo 2 > /proc/irq/59/smp_affi nity
// Decimal system 2 Corresponding binary 0010, If you want to change to CPU2, You need to write 0100 Decimal system 4, however IMX6D Only 2 individual CPU The core , So set 4 It's invalid ( Don't take effect )

Mode two ：

         Call in driver code irq_set_affinity() Bind the serial port interrupt to a CPU At the heart of it .

retval = request_irq(sport->port.irq, imx_int, 0,
                                     dev_name(port->dev), sport);
irq_set_affinity(sport->port.irq, cpumask_of(2));

         By adjusting the interruption of kinship, it can only have a certain improvement , You can't 100% Solve this problem .

2.3.4 Add flow control

         Flow control is divided into software flow control and hardware flow control , Personally, I prefer hardware flow control .