-
Notifications
You must be signed in to change notification settings - Fork 2k
Guide: Writing a device driver in RIOT
This document describes the requirement, design objectives, and some non-function details when writing device drivers in/for RIOT. The term device driver in this context includes all 'CPU-external' devices connected to the CPU typically via peripherals like SPI, I2C, UART, GPIO, and similar. CPU peripherals itself are in RIOT not considered to be device drivers, but peripheral or low-level drivers. Typical devices are network devices like radios, Ethernet adapters, sensors, and actuators.
Device drivers should be as easy to use as possible. This implies an 'intialize->ready' paradigm, meaning, that device drivers should be ready to use and in operation right after they have been initialized. On top, devices should work with physical values where ever possible. So e.g. sensors should return already converted values in some physical unit instead of RAW data, so that users can work directly with data from different devices directly without having to care for device specific conversion.
Additionally towards ease of use, all device drivers in RIOT should provide a similar 'look and feel'. They should behave similar concerning things like their state after initialization, like their used data representation and so on.
Secondly, all device drivers should be optimized for minimal RAM/ROM usage, as RIOT targets (very) constrained devices. This implies, that instead of exposing all thinkable functionality, the drivers should focus exporting and implementing a device's core functionality, thus covering ~95% of the use cases. Furthermore great care should be put into
Third, it should always be possible, to handle more than a single device of one kind. Drivers and their interfaces are thus designed to keep their state information in a parameterized location instead of driver defined global variables.
Fourth, RIOT defines high-level interfaces for certain groups of devices (i.e. netdev for network devices, SAUL for sensors and actuators), which enables users to work with a wide variety of devices without having to know anything about the actual device that is mapped.
Fifth, during initialization we make sure, that we can communicate with a device. After initialization, we assume that the communication with the device works as expected, thus there is no need to check the communication status in each function of the driver.
Sixth, device drivers SHOULD be implemented independent of any CPU/board code. To achieve this, the driver implementations should be strictly be based on platform independent interfaces as the peripheral drivers, xtimer, etc.
Document what your driver does! Most device come with a very large number of features, while the corresponding device driver only supports a sub-set of them. This should be clearly stated in the device driver's documentation, so that anyone wanting to use the driver can find out the supported features without having to scan through the code.
Each device MUST supply a data structure, holding the devices state and configuration, using the naming scheme of DEVNAME_t (e.g. dht_t, or at86rf2xx_t). In context of RIOT, we call this data structure the device descriptor.
This device descriptor MUST contain all the state data of a device. By this, we are not limited on the number of instances of the driver we can run concurrently. The descriptor is hereby used for identifying the device we want to interact with, and SHOULD always be the first parameter for all device driver related functions.
Typical things found in these descriptors are e.g. used peripherals (e.g. SPI or I2C bus used, interfacing GPIO pins), data buffers (e.g. RX/TX buffers where needed), or state machine information.
On top of the device descriptor, each device driver MUST also define a data structure holding the needed configuration data. The naming scheme for this type is DEVNAME_params_t. In contrary to the device descriptor, this data structure should only contain static information, that is needed for the device initialization as it is preferably allocated in ROM.
A simple I2C temperature sensors's device descriptor could look like this:
typedef struct {
tmpabc_params_t p; /**< device configuation parameter like I2C bus and bus addr */
int scale; /**< some custom scaling factor for converting the results */
} tmpabc_t;
/* with params being */
typedef struct {
i2c_t bus; /**< I2C bus the device is connected to */
uint8_t addr; /**< the device's address on the bus */
} tmpabc_params_t;NOTE: In many cases it makes sense, to copy the xxx_params data into the device descriptor during initialization. In some cases, it is however better to just link the params data via pointer and only copy selected data. This way, configuration data that is only used once can be read directly from ROM, while often used fields (e.g. used peripherals) are stored directly in the device descriptor and one saves hereby one de-referencing step when accessing them.
Each device driver in RIOT MUST supply a default configuration file, named DEVNAME_params.h. This file should be located in the RIOT/drivers/.... The idea is, that this file can be overridden by an application or a board, by simply putting a file with the same name in the application's or the board's include folders, while RIOT's build system takes care of prefferring those files instead of the default params file.
A default parameter header file for the example temperature sensor above would look like this (tmpabc_params.h):
...
#include "board.h" /* THIS INCLUDE IS MANDATORY */
#include "tmpabc.h"
...
/**
* @brief Default configuration parameters for TMPABC sensors
* @{
*/
#ifndef TMPABC_PARAM_I2C
#define TMPABC_PARAM_I2C (I2C_DEV(0)
#endif
#ifndef TMPABC_PARAM_ADDR
#define TMPABC_PARAM_ADDR (0xab)
#endif
#ifndef TMPABC_PARAMS
#define TMPABC_PARAMS { .i2c = TMPABC_PARAM_I2C \
.addr = TMPABC_PARAM_ADDR }
#endif
/** @} */
/**
* @brief Allocation of TMPABC configuration
*/
static const tmpabc_params_t tmpabc_params[] = {
TMPABC_PARAMS
}
...Now to influence the default configuration parameters, we have these options:
- We can override one or more of the parameter from the makesystem, e.g.
CFLAGS="-DTMPABC_PARAM_ADDR=0x23" make all- We can override selected parameters from the board configuration (
board.h):
...
/**
* @brief TMPABC sensor configuration
* @{
*/
#define TMPABC_PARAM_I2C (I2C_DEV(1))
#define TMPABC_PARAM_ADDR (0x17)
/** @} */
...- We can define more than a single device in the board configuration (
board.h):
...
/**
* @brief Configure the on-board temperature sensors
* @{
*/
#define TMPABC_PARAMS { \
.i2c = I2C_DEV(1), \
.addr = 0x31 \
}, \
{ \
.i2c = I2C_DEV(1), \
.addr = 0x21 \
}
/** @} */
...- And finally, we can simply override the
tmpabc_params.hfile as described above.
In general, the initialization functions should to the following:
- initialize the device descriptor
- initialize non-shared peripherals they use, e.g. GPIO pins
- test for device connectivity, e.g. does a SPI/I2C slave react
- reset the device to a well defined state, e.g. use external reset lines or do a software rest
- do the actual device initialization
For testing a device's connectivity, it is recommended to read certain configuration data with a defined value from the device. Some devices off WHO_AM_I or DEVICE_ID registers for this purpose. Writing and reading back some data to the device is another valid option for testing it's responsiveness.
For more detailed information on how the signature of the init functions should look like, please refer below to the specific requirements for network devices and sensors.
As stated above, we check communication a device only once during initialization and expect the communication to be valid from there on. To prevent subsequent miss-use by passing NULL pointers and similar to the subsequent functions, the recommended way is to check parameter using assert, e.g.:
int16_t tmpabc_read(const tmpabc_t *dev)
{
assert(dev);
....
return value;
}Here in this case we expect the return value to always be a valid temperature value - if anything else that could go wrong is caught by the assert. This saves a lot of conditional code, which would be only concerned with figuring out which return value might be the right one...
Whenever status/error return values are however used in a driver, they should be named, meaning that the driver MUST define an enum assigning names to the actual used value, e.g.
enum {
TMPABC_OK = 0, /**< all went as expected */
TMPABC_NOI2C = -1, /**< error using the I2C bus */
TMPABC_NODEV = -2 /**< no device with the configured address found on the bus */
};- MUST: the supported feature set and any custom behavior is clearly documented
- MUST: device descriptor is defined:
devab_t - MUST: device parameter struct is defined:
devab_params_t - MUST: a default parameter configuration file is present, e.g.
RIOT/drivers/devab/include/devab_params.h - MUST: all error and status return codes are named, e.g.
DEVAB_OK, DEVAB_NOSPI, DEVAB_OVERFLOW, ... - MUST: use
const devab_t *devwhen the device descriptor can be access read-only
All sensor drivers SHOULD implement the SAUL interface. It is however recommended, that the drivers are written in a way, that the drivers to not solely export the SAUL interface, but map the SAUL interface upon a driver specific one.
For example, a temperature driver provides the following function (tmpabc.c):
int16_t tmpabc_read(tmpabc_t *dev);which then can easily be mapped to SAUL (tmpabc_saul.c):
int saul_read(saul_t *dev, phydat_t *data)
{
memset(data, 0, sizeof(phydat_t));
data->x = tmpabc_read((tmpabc_t *)dev);
data->unit = UNIT_TEMP_C;
data->scale = -2;
return 1;
}This ensure the versatility of the device driver, having in mind that one might want to use the driver without SAUL or maybe in a context without RIOT.
Sensor device drivers are expected to implement a single initialization function, DEVNAME_init, taking the device descriptor and the device's parameter struct as argument:
int tmpabc_init(tmpabc_t *dev, const tmpabc_params_t *params);After this function is called, the device MUST be running and usable.
All sensors in RIOT MUST return their reading results as real physical values. When working with sensor data, these are the values of interest, and the overhead of the conversion is normally neglectable.
All values SHOULD be returned using integer types, with int16_t being the preferred type where applicable.
In many situations, the physical values can not be mapped directly to integer values. For example, we do not want to map temperature values to integers directly while using their fraction. The recommended way to solve this by scaling the result value using decimal fixed point arithmetics, in other words just return centi-degree instead of degree (e.g. 2372c°C instead of 23.72°C).
- MUST: mandatory device parameters are configurable through this file, e.g. sampling rate, resolution, sensitivity
- MUST: an init function in the style of
int devab_init(devab_t *dev, const devab_params_t *params);exists - MUST: after initialization, the device must be operational
- MUST: all error and return values are named, e.g.
DEVAB_OK, DEVAB_NODEV, ... - MUST: all 'read' functions return values in their physical representation, e.g.
centi-degree, Pa, lux, mili-G - MUST: all 'read' functions return integer values, prefarably
int16_t - SHOULD: if multiple dimensions are read, they SHOULD be combined into a data structure
- SHOULD: the driver implements the SAUL interface
- SHOULD: the driver exports functions for putting it to sleep and waking the device
The initialization process MUST be split into 2 steps: first initialize the device descriptor and if applicable the used peripherals, and secondly do the actual device initialization. The reason for this is, that before a device is actually activated and can start to process data, the network stack for the device needs to initialized. By supplying a second init function, that does the actual initialization, we can hand the control over when this is done to the actual network stacks.
The initialization functions SHOULD follow this naming scheme:
void netabc_setup(netabc_t *dev, const netabc_params_t *params);
int netabs_init(netabc_t *dev);
Device driver for network device SHOULD implement the netdev interface. It is up to the implementer, if the device driver also offers a device specific interface which is then mapped to the netdev interface, or if the device driver can purely interfaced using netdev. While the second option is recommended for efficiency reasons, this is not mandatory.
- MUST: a setup function in the style of
int devab_setup(devab_t *dev, const devab_params_t *params);exists - MUST: an init function in the style of int devnet_init(devnet_t *dev) exists
- SHOULD: the driver implements 'netdev' [if applicable]
Add some information about how to handle multiple threads, when to use mutexes, and how to deal with interrupts? And especially patterns for being nice from other threads and power consumption point of view...