Revision 7, 2007-04-19
Jean Delvare <khali@linux-fr.org>
Greg KH <greg@kroah.com>

This is a guide on how to convert I2C chip drivers from Linux 2.4 to
Linux 2.6. I have been using existing drivers (lm75, lm78) as examples.
Then I converted a driver myself (lm83) and updated this document.
Note that this guide is strongly oriented towards hardware monitoring
drivers. Many points are still valid for other type of drivers, but
others may be irrelevant.

There are two sets of points below. The first set concerns technical
changes. The second set concerns coding policy. Both are mandatory.

Although reading this guide will help you porting drivers, I suggest
you keep an eye on an already ported driver while porting your own
driver. This will help you a lot understanding what this guide
exactly means. Choose the chip driver that is the more similar to
yours for best results.

Technical changes:

* [Driver type] Any driver that was relying on i2c-isa has to be
  converted to a proper isa, platform or pci driver. This is not
  covered by this guide.

* [Includes] Get rid of "version.h" and <linux/i2c-proc.h>.
  Includes typically look like that:
  #include <linux/module.h>
  #include <linux/init.h>
  #include <linux/slab.h>
  #include <linux/jiffies.h>
  #include <linux/i2c.h>
  #include <linux/hwmon.h>	/* for hardware monitoring drivers */
  #include <linux/hwmon-sysfs.h>
  #include <linux/hwmon-vid.h>	/* if you need VRM support */
  #include <linux/err.h>	/* for class registration */
  Please respect this inclusion order. Some extra headers may be
  required for a given driver (e.g. "lm75.h").

* [Addresses] SENSORS_I2C_END becomes I2C_CLIENT_END, ISA addresses
  are no more handled by the i2c core. Address ranges are no more
  supported either, define each individual address separately.
  SENSORS_INSMOD_<n> becomes I2C_CLIENT_INSMOD_<n>.

* [Client data] Get rid of sysctl_id. Try using standard names for
  register values (for example, temp_os becomes temp_max). You're
  still relatively free here, but you *have* to follow the standard
  names for sysfs files (see the Sysctl section below).

* [Function prototypes] The detect functions loses its flags
  parameter. Sysctl (e.g. lm75_temp) and miscellaneous functions
  are off the list of prototypes. This usually leaves five
  prototypes:
  static int lm75_attach_adapter(struct i2c_adapter *adapter);
  static int lm75_detect(struct i2c_adapter *adapter, int address,
      int kind);
  static void lm75_init_client(struct i2c_client *client);
  static int lm75_detach_client(struct i2c_client *client);
  static struct lm75_data lm75_update_device(struct device *dev);

* [Sysctl] All sysctl stuff is of course gone (defines, ctl_table
  and functions). Instead, you have to define show and set functions for
  each sysfs file. Only define set for writable values. Take a look at an
  existing 2.6 driver for details (it87 for example). Don't forget
  to define the attributes for each file (this is that step that
  links callback functions). Use the file names specified in
  Documentation/hwmon/sysfs-interface for the individual files. Also
  convert the units these files read and write to the specified ones.
  If you need to add a new type of file, please discuss it on the
  sensors mailing list <lm-sensors@lm-sensors.org> by providing a
  patch to the Documentation/hwmon/sysfs-interface file.

* [Attach] The attach function should make sure that the adapter's
  class has I2C_CLASS_HWMON (or whatever class is suitable for your
  driver), using the following construct:
  if (!(adapter->class & I2C_CLASS_HWMON))
          return 0;
  Call i2c_probe() instead of i2c_detect().

* [Detect] As mentioned earlier, the flags parameter is gone.
  The type_name and client_name strings are replaced by a single
  name string, which will be filled with a lowercase, short string.
  The labels used for error paths are reduced to the number needed.
  It is advised that the labels are given descriptive names such as
  exit and exit_free. Don't forget to properly set err before
  jumping to error labels. By the way, labels should be left-aligned.
  Use kzalloc instead of kmalloc.
  Use i2c_set_clientdata to set the client data (as opposed to
  a direct access to client->data).
  Use strlcpy instead of strcpy or snprintf to copy the client name.
  Replace the sysctl directory registration by calls to
  device_create_file. Move the driver initialization before any
  sysfs file creation.
  Register the client with the hwmon class (using hwmon_device_register)
  if applicable.
  Drop client->id.
  Drop any 24RF08 corruption prevention you find, as this is now done
  at the i2c-core level, and doing it twice voids it.
  Don't add I2C_CLIENT_ALLOW_USE to client->flags, it's the default now.

* [Init] Limits must not be set by the driver (can be done later in
  user-space). Chip should not be reset default (although a module
  parameter may be used to force it), and initialization should be
  limited to the strictly necessary steps.

* [Detach] Remove the call to i2c_deregister_entry. Do not log an
  error message if i2c_detach_client fails, as i2c-core will now do
  it for you.
  Unregister from the hwmon class if applicable.

* [Update] The function prototype changed, it is now
  passed a device structure, which you have to convert to a client
  using to_i2c_client(dev). The update function should return a
  pointer to the client data.
  Don't access client->data directly, use i2c_get_clientdata(client)
  instead.
  Use time_after() instead of direct jiffies comparison.

* [Interface] Make sure there is a MODULE_LICENSE() line, at the bottom
  of the file (after MODULE_AUTHOR() and MODULE_DESCRIPTION(), in this
  order).

* [Driver] The flags field of the i2c_driver structure is gone.
  I2C_DF_NOTIFY is now the default behavior.
  The i2c_driver structure has a driver member, which is itself a
  structure, those name member should be initialized to a driver name
  string. i2c_driver itself has no name member anymore.

* [Driver model] Instead of shutdown or reboot notifiers, provide a
  shutdown() method in your driver.

* [Power management] Use the driver model suspend() and resume()
  callbacks instead of the obsolete pm_register() calls.

Coding policy:

* [Copyright] Use (C), not (c), for copyright.

* [Debug/log] Get rid of #ifdef DEBUG/#endif constructs whenever you
  can. Calls to printk for debugging purposes are replaced by calls to
  dev_dbg where possible, else to pr_debug. Here is an example of how
  to call it (taken from lm75_detect):
  dev_dbg(&client->dev, "Starting lm75 update\n");
  Replace other printk calls with the dev_info, dev_err or dev_warn
  function, as appropriate.

* [Constants] Constants defines (registers, conversions) should be
  aligned. This greatly improves readability.
  Alignments are achieved by the means of tabs, not spaces. Remember
  that tabs are set to 8 in the Linux kernel code.

* [Layout] Avoid extra empty lines between comments and what they
  comment. Respect the coding style (see Documentation/CodingStyle),
  in particular when it comes to placing curly braces.

* [Comments] Make sure that no comment refers to a file that isn't
  part of the Linux source tree (typically doc/chips/<chip name>),
  and that remaining comments still match the code. Merging comment
  lines when possible is encouraged.