Revision 4, 2004-03-30
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.

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:

* [Includes] Get rid of "version.h". Replace <linux/i2c-proc.h> with
  <linux/i2c-sensor.h>. Includes typically look like that:
  #include <linux/module.h>
  #include <linux/init.h>
  #include <linux/slab.h>
  #include <linux/i2c.h>
  #include <linux/i2c-sensor.h>
  #include <linux/i2c-vid.h>	/* if you need VRM support */
  #include <asm/io.h>		/* if you have I/O operations */
  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.

* [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 void lm75_update_client(struct i2c_client *client);

* [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 (lm78 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/i2c/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/i2c/sysfs-interface file.

* [Attach] For I2C drivers, the attach function should make sure
  that the adapter's class has I2C_CLASS_HWMON, using the
  following construct:
  if (!(adapter->class & I2C_CLASS_HWMON))
          return 0;
  ISA-only drivers of course don't need this.

* [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
  (typically the driver name, e.g. "lm75").
  In i2c-only drivers, drop the i2c_is_isa_adapter check, it's
  useless. Same for isa-only drivers, as the test would always be
  true. Only hybrid drivers (which are quite rare) still need it.
  The errorN labels are reduced to the number needed. If that number
  is 2 (i2c-only drivers), it is advised that the labels are named
  exit and exit_free. For i2c+isa drivers, labels should be named
  ERROR0, ERROR1 and ERROR2. Don't forget to properly set err before
  jumping to error labels. By the way, labels should be left-aligned.
  Use memset to fill the client and data area with 0x00.
  Use i2c_set_clientdata to set the client data (as opposed to
  a direct access to client->data).
  Use strlcpy instead of strcpy 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.
  Drop client->id.

* [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 is), and initialization should be
  limited to the strictly necessary steps.

* [Detach] Get rid of data, 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.

* [Update] Don't access client->data directly, use
  i2c_get_clientdata(client) instead.

* [Interface] Init function should not print anything. Make sure
  there is a MODULE_LICENSE() line, at the bottom of the file
  (after MODULE_AUTHOR() and MODULE_DESCRIPTION(), in this order).

Coding policy:

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

* [Debug/log] Get rid of #ifdef DEBUG/#endif constructs whenever you
  can. Calls to printk/pr_debug for debugging purposes are replaced
  by calls to dev_dbg. Here is an example on 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, initial
  values) should be aligned. This greatly improves readability.
  Same goes for variables declarations. Alignments are achieved by the
  means of tabs, not spaces. Remember that tabs are set to 8 in the
  Linux kernel code.

* [Structure definition] The name field should be standardized. All
  lowercase and as simple as the driver name itself (e.g. "lm75").

* [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.