3. Networking Driver

3.1. Introduction

This chapter is intended to provide an introduction to the procedure for writing RTEMS network device drivers. The example code is taken from the ‘Generic 68360’ network device driver. The source code for this driver is located in the c/src/lib/libbsp/m68k/gen68360/network directory in the RTEMS source code distribution. Having a copy of this driver at hand when reading the following notes will help significantly.

3.2. Learn about the network device

Before starting to write the network driver become completely familiar with the programmer’s view of the device. The following points list some of the details of the device that must be understood before a driver can be written.

  • Does the device use DMA to transfer packets to and from memory or does the processor have to copy packets to and from memory on the device?
  • If the device uses DMA, is it capable of forming a single outgoing packet from multiple fragments scattered in separate memory buffers?
  • If the device uses DMA, is it capable of chaining multiple outgoing packets, or does each outgoing packet require intervention by the driver?
  • Does the device automatically pad short frames to the minimum 64 bytes or does the driver have to supply the padding?
  • Does the device automatically retry a transmission on detection of a collision?
  • If the device uses DMA, is it capable of buffering multiple packets to memory, or does the receiver have to be restarted after the arrival of each packet?
  • How are packets that are too short, too long, or received with CRC errors handled? Does the device automatically continue reception or does the driver have to intervene?
  • How is the device Ethernet address set? How is the device programmed to accept or reject broadcast and multicast packets?
  • What interrupts does the device generate? Does it generate an interrupt for each incoming packet, or only for packets received without error? Does it generate an interrupt for each packet transmitted, or only when the transmit queue is empty? What happens when a transmit error is detected?

In addition, some controllers have specific questions regarding board specific configuration. For example, the SONIC Ethernet controller has a very configurable data bus interface. It can even be configured for sixteen and thirty-two bit data buses. This type of information should be obtained from the board vendor.

3.3. Understand the network scheduling conventions

When writing code for the driver transmit and receive tasks, take care to follow the network scheduling conventions. All tasks which are associated with networking share various data structures and resources. To ensure the consistency of these structures the tasks execute only when they hold the network semaphore (rtems_bsdnet_semaphore). The transmit and receive tasks must abide by this protocol. Be very careful to avoid ‘deadly embraces’ with the other network tasks. A number of routines are provided to make it easier for the network driver code to conform to the network task scheduling conventions.

  • void rtems_bsdnet_semaphore_release(void) This function releases the network semaphore. The network driver tasks must call this function immediately before making any blocking RTEMS request.
  • void rtems_bsdnet_semaphore_obtain(void) This function obtains the network semaphore. If a network driver task has released the network semaphore to allow other network-related tasks to run while the task blocks, then this function must be called to reobtain the semaphore immediately after the return from the blocking RTEMS request.
  • rtems_bsdnet_event_receive(rtems_event_set, rtems_option, rtems_interval, rtems_event_set *) The network driver task should call this function when it wishes to wait for an event. This function releases the network semaphore, calls rtems_event_receive to wait for the specified event or events and reobtains the semaphore. The value returned is the value returned by the rtems_event_receive.

3.4. Network Driver Makefile

Network drivers are considered part of the BSD network package and as such are to be compiled with the appropriate flags. This can be accomplished by adding -D__INSIDE_RTEMS_BSD_TCPIP_STACK__ to the command line. If the driver is inside the RTEMS source tree or is built using the RTEMS application Makefiles, then adding the following line accomplishes this:

DEFINES += -D__INSIDE_RTEMS_BSD_TCPIP_STACK__

This is equivalent to the following list of definitions. Early versions of the RTEMS BSD network stack required that all of these be defined.

-D_COMPILING_BSD_KERNEL_ -DKERNEL -DINET -DNFS \
  -DDIAGNOSTIC -DBOOTP_COMPAT

Defining these macros tells the network header files that the driver is to be compiled with extended visibility into the network stack. This is in sharp contrast to applications that simply use the network stack. Applications do not require this level of visibility and should stick to the portable application level API.

As a direct result of being logically internal to the network stack, network drivers use the BSD memory allocation routines This means, for example, that malloc takes three arguments. See the SONIC device driver (c/src/lib/libchip/network/sonic.c) for an example of this. Because of this, network drivers should not include <stdlib.h>. Doing so will result in conflicting definitions of malloc().

Application level code including network servers such as the FTP daemon are not part of the BSD kernel network code and should not be compiled with the BSD network flags. They should include <stdlib.h> and not define the network stack visibility macros.

3.5. Write the Driver Attach Function

The driver attach function is responsible for configuring the driver and making the connection between the network stack and the driver.

Driver attach functions take a pointer to an rtems_bsdnet_ifconfig structure as their only argument. and set the driver parameters based on the values in this structure. If an entry in the configuration structure is zero the attach function chooses an appropriate default value for that parameter.

The driver should then set up several fields in the ifnet structure in the device-dependent data structure supplied and maintained by the driver:

ifp->if_softc
Pointer to the device-dependent data. The first entry in the device-dependent data structure must be an arpcom structure.
ifp->if_name
The name of the device. The network stack uses this string and the device number for device name lookups. The device name should be obtained from the name entry in the configuration structure.
ifp->if_unit
The device number. The network stack uses this number and the device name for device name lookups. For example, if ifp->if_name is scc and ifp->if_unit is 1, the full device name would be scc1. The unit number should be obtained from the ‘name’ entry in the configuration structure.
ifp->if_mtu
The maximum transmission unit for the device. For Ethernet devices this value should almost always be 1500.
ifp->if_flags
The device flags. Ethernet devices should set the flags to IFF_BROADCAST|IFF_SIMPLEX, indicating that the device can broadcast packets to multiple destinations and does not receive and transmit at the same time.
ifp->if_snd.ifq_maxlen
The maximum length of the queue of packets waiting to be sent to the driver. This is normally set to ifqmaxlen.
ifp->if_init
The address of the driver initialization function.
ifp->if_start
The address of the driver start function.
ifp->if_ioctl
The address of the driver ioctl function.
ifp->if_output
The address of the output function. Ethernet devices should set this to ether_output.

RTEMS provides a function to parse the driver name in the configuration structure into a device name and unit number.

int rtems_bsdnet_parse_driver_name (
    const struct rtems_bsdnet_ifconfig *config,
    char **namep
);

The function takes two arguments; a pointer to the configuration structure and a pointer to a pointer to a character. The function parses the configuration name entry, allocates memory for the driver name, places the driver name in this memory, sets the second argument to point to the name and returns the unit number. On error, a message is printed and -1 is returned.

Once the attach function has set up the above entries it must link the driver data structure onto the list of devices by calling if_attach. Ethernet devices should then call ether_ifattach. Both functions take a pointer to the device’s ifnet structure as their only argument.

The attach function should return a non-zero value to indicate that the driver has been successfully configured and attached.

3.6. Write the Driver Start Function.

This function is called each time the network stack wants to start the transmitter. This occures whenever the network stack adds a packet to a device’s send queue and the IFF_OACTIVE bit in the device’s if_flags is not set.

For many devices this function need only set the IFF_OACTIVE bit in the if_flags and send an event to the transmit task indicating that a packet is in the driver transmit queue.

3.7. Write the Driver Initialization Function.

This function should initialize the device, attach to interrupt handler, and start the driver transmit and receive tasks. The function

rtems_id
rtems_bsdnet_newproc (char *name,
    int stacksize,
    void(*entry)(void *),
    void *arg);

should be used to start the driver tasks.

Note that the network stack may call the driver initialization function more than once. Make sure multiple versions of the receive and transmit tasks are not accidentally started.

3.8. Write the Driver Transmit Task

This task is reponsible for removing packets from the driver send queue and sending them to the device. The task should block waiting for an event from the driver start function indicating that packets are waiting to be transmitted. When the transmit task has drained the driver send queue the task should clear the IFF_OACTIVE bit in if_flags and block until another outgoing packet is queued.

3.9. Write the Driver Receive Task

This task should block until a packet arrives from the device. If the device is an Ethernet interface the function ether_input should be called to forward the packet to the network stack. The arguments to ether_input are a pointer to the interface data structure, a pointer to the ethernet header and a pointer to an mbuf containing the packet itself.

3.10. Write the Driver Interrupt Handler

A typical interrupt handler will do nothing more than the hardware manipulation required to acknowledge the interrupt and send an RTEMS event to wake up the driver receive or transmit task waiting for the event. Network interface interrupt handlers must not make any calls to other network routines.

3.11. Write the Driver IOCTL Function

This function handles ioctl requests directed at the device. The ioctl commands which must be handled are:

SIOCGIFADDR

SIOCSIFADDR
If the device is an Ethernet interface these commands should be passed on to ether_ioctl.
SIOCSIFFLAGS

This command should be used to start or stop the device, depending on the state of the interface IFF_UP and IFF_RUNNING bits in if_flags:

IFF_RUNNING
Stop the device.
IFF_UP
Start the device.
IFF_UP|IFF_RUNNING
Stop then start the device.
0
Do nothing.

3.12. Write the Driver Statistic-Printing Function

This function should print the values of any statistic/diagnostic counters the network driver may use. The driver ioctl function should call the statistic-printing function when the ioctl command is SIO_RTEMS_SHOW_STATS.