patch-2.4.19 linux-2.4.19/include/net/iw_handler.h

Next file: linux-2.4.19/include/net/ndisc.h
Previous file: linux-2.4.19/include/net/irda/irlap_event.h
Back to the patch index
Back to the overall index

diff -urN linux-2.4.18/include/net/iw_handler.h linux-2.4.19/include/net/iw_handler.h
@@ -0,0 +1,374 @@
+/*
+ * This file define the new driver API for Wireless Extensions
+ *
+ * Version :	2	6.12.01
+ *
+ * Authors :	Jean Tourrilhes - HPL - <jt@hpl.hp.com>
+ * Copyright (c) 2001 Jean Tourrilhes, All Rights Reserved.
+ */
+
+#ifndef _IW_HANDLER_H
+#define _IW_HANDLER_H
+
+/************************** DOCUMENTATION **************************/
+/*
+ * Initial driver API (1996 -> onward) :
+ * -----------------------------------
+ * The initial API just sends the IOCTL request received from user space
+ * to the driver (via the driver ioctl handler). The driver has to
+ * handle all the rest...
+ *
+ * The initial API also defines a specific handler in struct net_device
+ * to handle wireless statistics.
+ *
+ * The initial APIs served us well and has proven a reasonably good design.
+ * However, there is a few shortcommings :
+ *	o No events, everything is a request to the driver.
+ *	o Large ioctl function in driver with gigantic switch statement
+ *	  (i.e. spaghetti code).
+ *	o Driver has to mess up with copy_to/from_user, and in many cases
+ *	  does it unproperly. Common mistakes are :
+ *		* buffer overflows (no checks or off by one checks)
+ *		* call copy_to/from_user with irq disabled
+ *	o The user space interface is tied to ioctl because of the use
+ *	  copy_to/from_user.
+ *
+ * New driver API (2001 -> onward) :
+ * -------------------------------
+ * The new driver API is just a bunch of standard functions (handlers),
+ * each handling a specific Wireless Extension. The driver just export
+ * the list of handler it supports, and those will be called apropriately.
+ *
+ * I tried to keep the main advantage of the previous API (simplicity,
+ * efficiency and light weight), and also I provide a good dose of backward
+ * compatibility (most structures are the same, driver can use both API
+ * simultaneously, ...).
+ * Hopefully, I've also addressed the shortcomming of the initial API.
+ *
+ * The advantage of the new API are :
+ *	o Handling of Extensions in driver broken in small contained functions
+ *	o Tighter checks of ioctl before calling the driver
+ *	o Flexible commit strategy (at least, the start of it)
+ *	o Backward compatibility (can be mixed with old API)
+ *	o Driver doesn't have to worry about memory and user-space issues
+ * The last point is important for the following reasons :
+ *	o You are now able to call the new driver API from any API you
+ *		want (including from within other parts of the kernel).
+ *	o Common mistakes are avoided (buffer overflow, user space copy
+ *		with irq disabled and so on).
+ *
+ * The Drawback of the new API are :
+ *	o bloat (especially kernel)
+ *	o need to migrate existing drivers to new API
+ * My initial testing shows that the new API adds around 3kB to the kernel
+ * and save between 0 and 5kB from a typical driver.
+ * Also, as all structures and data types are unchanged, the migration is
+ * quite straightforward (but tedious).
+ *
+ * ---
+ *
+ * The new driver API is defined below in this file. User space should
+ * not be aware of what's happening down there...
+ *
+ * A new kernel wrapper is in charge of validating the IOCTLs and calling
+ * the appropriate driver handler. This is implemented in :
+ *	# net/core/wireless.c
+ *
+ * The driver export the list of handlers in :
+ *	# include/linux/netdevice.h (one place)
+ *
+ * The new driver API is available for WIRELESS_EXT >= 13.
+ * Good luck with migration to the new API ;-)
+ */
+
+/* ---------------------- THE IMPLEMENTATION ---------------------- */
+/*
+ * Some of the choice I've made are pretty controversials. Defining an
+ * API is very much weighting compromises. This goes into some of the
+ * details and the thinking behind the implementation.
+ *
+ * Implementation goals :
+ * --------------------
+ * The implementation goals were as follow :
+ *	o Obvious : you should not need a PhD to understand what's happening,
+ *		the benefit is easier maintainance.
+ *	o Flexible : it should accomodate a wide variety of driver
+ *		implementations and be as flexible as the old API.
+ *	o Lean : it should be efficient memory wise to minimise the impact
+ *		on kernel footprint.
+ *	o Transparent to user space : the large number of user space
+ *		applications that use Wireless Extensions should not need
+ *		any modifications.
+ *
+ * Array of functions versus Struct of functions
+ * ---------------------------------------------
+ * 1) Having an array of functions allow the kernel code to access the
+ * handler in a single lookup, which is much more efficient (think hash
+ * table here).
+ * 2) The only drawback is that driver writer may put their handler in
+ * the wrong slot. This is trivial to test (I set the frequency, the
+ * bitrate changes). Once the handler is in the proper slot, it will be
+ * there forever, because the array is only extended at the end.
+ * 3) Backward/forward compatibility : adding new handler just require
+ * extending the array, so you can put newer driver in older kernel
+ * without having to patch the kernel code (and vice versa).
+ *
+ * All handler are of the same generic type
+ * ----------------------------------------
+ * That's a feature !!!
+ * 1) Having a generic handler allow to have generic code, which is more
+ * efficient. If each of the handler was individually typed I would need
+ * to add a big switch in the kernel (== more bloat). This solution is
+ * more scalable, adding new Wireless Extensions doesn't add new code.
+ * 2) You can use the same handler in different slots of the array. For
+ * hardware, it may be more efficient or logical to handle multiple
+ * Wireless Extensions with a single function, and the API allow you to
+ * do that. (An example would be a single record on the card to control
+ * both bitrate and frequency, the handler would read the old record,
+ * modify it according to info->cmd and rewrite it).
+ *
+ * Functions prototype uses union iwreq_data
+ * -----------------------------------------
+ * Some would have prefered functions defined this way :
+ *	static int mydriver_ioctl_setrate(struct net_device *dev, 
+ *					  long rate, int auto)
+ * 1) The kernel code doesn't "validate" the content of iwreq_data, and
+ * can't do it (different hardware may have different notion of what a
+ * valid frequency is), so we don't pretend that we do it.
+ * 2) The above form is not extendable. If I want to add a flag (for
+ * example to distinguish setting max rate and basic rate), I would
+ * break the prototype. Using iwreq_data is more flexible.
+ * 3) Also, the above form is not generic (see above).
+ * 4) I don't expect driver developper using the wrong field of the
+ * union (Doh !), so static typechecking doesn't add much value.
+ * 5) Lastly, you can skip the union by doing :
+ *	static int mydriver_ioctl_setrate(struct net_device *dev,
+ *					  struct iw_request_info *info,
+ *					  struct iw_param *rrq,
+ *					  char *extra)
+ * And then adding the handler in the array like this :
+ *        (iw_handler) mydriver_ioctl_setrate,             // SIOCSIWRATE
+ *
+ * Using functions and not a registry
+ * ----------------------------------
+ * Another implementation option would have been for every instance to
+ * define a registry (a struct containing all the Wireless Extensions)
+ * and only have a function to commit the registry to the hardware.
+ * 1) This approach can be emulated by the current code, but not
+ * vice versa.
+ * 2) Some drivers don't keep any configuration in the driver, for them
+ * adding such a registry would be a significant bloat.
+ * 3) The code to translate from Wireless Extension to native format is
+ * needed anyway, so it would not reduce significantely the amount of code.
+ * 4) The current approach only selectively translate Wireless Extensions
+ * to native format and only selectively set, whereas the registry approach
+ * would require to translate all WE and set all parameters for any single
+ * change.
+ * 5) For many Wireless Extensions, the GET operation return the current
+ * dynamic value, not the value that was set.
+ *
+ * This header is <net/iw_handler.h>
+ * ---------------------------------
+ * 1) This header is kernel space only and should not be exported to
+ * user space. Headers in "include/linux/" are exported, headers in
+ * "include/net/" are not.
+ *
+ * Mixed 32/64 bit issues
+ * ----------------------
+ * The Wireless Extensions are designed to be 64 bit clean, by using only
+ * datatypes with explicit storage size.
+ * There are some issues related to kernel and user space using different
+ * memory model, and in particular 64bit kernel with 32bit user space.
+ * The problem is related to struct iw_point, that contains a pointer
+ * that *may* need to be translated.
+ * This is quite messy. The new API doesn't solve this problem (it can't),
+ * but is a step in the right direction :
+ * 1) Meta data about each ioctl is easily available, so we know what type
+ * of translation is needed.
+ * 2) The move of data between kernel and user space is only done in a single
+ * place in the kernel, so adding specific hooks in there is possible.
+ * 3) In the long term, it allows to move away from using ioctl as the
+ * user space API.
+ *
+ * So many comments and so few code
+ * --------------------------------
+ * That's a feature. Comments won't bloat the resulting kernel binary.
+ */
+
+/***************************** INCLUDES *****************************/
+
+#include <linux/wireless.h>		/* IOCTL user space API */
+
+/***************************** VERSION *****************************/
+/*
+ * This constant is used to know which version of the driver API is
+ * available. Hopefully, this will be pretty stable and no changes
+ * will be needed...
+ * I just plan to increment with each new version.
+ */
+#define IW_HANDLER_VERSION	2
+
+/**************************** CONSTANTS ****************************/
+
+/* Special error message for the driver to indicate that we
+ * should do a commit after return from the iw_handler */
+#define EIWCOMMIT	EINPROGRESS
+
+/* Flags available in struct iw_request_info */
+#define IW_REQUEST_FLAG_NONE	0x0000	/* No flag so far */
+
+/* Type of headers we know about (basically union iwreq_data) */
+#define IW_HEADER_TYPE_NULL	0	/* Not available */
+#define IW_HEADER_TYPE_CHAR	2	/* char [IFNAMSIZ] */
+#define IW_HEADER_TYPE_UINT	4	/* __u32 */
+#define IW_HEADER_TYPE_FREQ	5	/* struct iw_freq */
+#define IW_HEADER_TYPE_POINT	6	/* struct iw_point */
+#define IW_HEADER_TYPE_PARAM	7	/* struct iw_param */
+#define IW_HEADER_TYPE_ADDR	8	/* struct sockaddr */
+
+/* Handling flags */
+/* Most are not implemented. I just use them as a reminder of some
+ * cool features we might need one day ;-) */
+#define IW_DESCR_FLAG_NONE	0x0000	/* Obvious */
+/* Wrapper level flags */
+#define IW_DESCR_FLAG_DUMP	0x0001	/* Not part of the dump command */
+#define IW_DESCR_FLAG_EVENT	0x0002	/* Generate an event on SET */
+#define IW_DESCR_FLAG_RESTRICT	0x0004	/* GET request is ROOT only */
+/* Driver level flags */
+#define IW_DESCR_FLAG_WAIT	0x0100	/* Wait for driver event */
+
+/****************************** TYPES ******************************/
+
+/* ----------------------- WIRELESS HANDLER ----------------------- */
+/*
+ * A wireless handler is just a standard function, that looks like the
+ * ioctl handler.
+ * We also define there how a handler list look like... As the Wireless
+ * Extension space is quite dense, we use a simple array, which is faster
+ * (that's the perfect hash table ;-).
+ */
+
+/*
+ * Meta data about the request passed to the iw_handler.
+ * Most handlers can safely ignore what's in there.
+ * The 'cmd' field might come handy if you want to use the same handler
+ * for multiple command...
+ * This struct is also my long term insurance. I can add new fields here
+ * without breaking the prototype of iw_handler...
+ */
+struct iw_request_info
+{
+	__u16		cmd;		/* Wireless Extension command */
+	__u16		flags;		/* More to come ;-) */
+};
+
+/*
+ * This is how a function handling a Wireless Extension should look
+ * like (both get and set, standard and private).
+ */
+typedef int (*iw_handler)(struct net_device *dev, struct iw_request_info *info,
+			  union iwreq_data *wrqu, char *extra);
+
+/*
+ * This define all the handler that the driver export.
+ * As you need only one per driver type, please use a static const
+ * shared by all driver instances... Same for the members...
+ * This will be linked from net_device in <linux/netdevice.h>
+ */
+struct iw_handler_def
+{
+	/* Number of handlers defined (more precisely, index of the
+	 * last defined handler + 1) */
+	__u16			num_standard;
+	__u16			num_private;
+	/* Number of private arg description */
+	__u16			num_private_args;
+
+	/* Array of handlers for standard ioctls
+	 * We will call dev->wireless_handlers->standard[ioctl - SIOCSIWNAME]
+	 */
+	iw_handler *		standard;
+
+	/* Array of handlers for private ioctls
+	 * Will call dev->wireless_handlers->private[ioctl - SIOCIWFIRSTPRIV]
+	 */
+	iw_handler *		private;
+
+	/* Arguments of private handler. This one is just a list, so you
+	 * can put it in any order you want and should not leave holes...
+	 * We will automatically export that to user space... */
+	struct iw_priv_args *	private_args;
+
+	/* In the long term, get_wireless_stats will move from
+	 * 'struct net_device' to here, to minimise bloat. */
+};
+
+/* ----------------------- WIRELESS EVENTS ----------------------- */
+/*
+ * Currently we don't support events, so let's just plan for the
+ * future...
+ */
+
+/*
+ * A Wireless Event.
+ */
+// How do we define short header ? We don't want a flag on length.
+// Probably a flag on event ? Highest bit to zero...
+struct iw_event
+{
+	__u16		length;			/* Lenght of this stuff */
+	__u16		event;			/* Wireless IOCTL */
+	union	iwreq_data	header;		/* IOCTL fixed payload */
+	char		extra[0];		/* Optional IOCTL data */
+};
+
+/* ---------------------- IOCTL DESCRIPTION ---------------------- */
+/*
+ * One of the main goal of the new interface is to deal entirely with
+ * user space/kernel space memory move.
+ * For that, we need to know :
+ *	o if iwreq is a pointer or contain the full data
+ *	o what is the size of the data to copy
+ *
+ * For private IOCTLs, we use the same rules as used by iwpriv and
+ * defined in struct iw_priv_args.
+ *
+ * For standard IOCTLs, things are quite different and we need to
+ * use the stuctures below. Actually, this struct is also more
+ * efficient, but that's another story...
+ */
+
+/*
+ * Describe how a standard IOCTL looks like.
+ */
+struct iw_ioctl_description
+{
+	__u8	header_type;		/* NULL, iw_point or other */
+	__u8	token_type;		/* Future */
+	__u16	token_size;		/* Granularity of payload */
+	__u16	min_tokens;		/* Min acceptable token number */
+	__u16	max_tokens;		/* Max acceptable token number */
+	__u32	flags;			/* Special handling of the request */
+};
+
+/* Need to think of short header translation table. Later. */
+
+/**************************** PROTOTYPES ****************************/
+/*
+ * Functions part of the Wireless Extensions (defined in net/core/wireless.c).
+ * Those may be called only within the kernel.
+ */
+
+/* First : function strictly used inside the kernel */
+
+/* Handle /proc/net/wireless, called in net/code/dev.c */
+extern int dev_get_wireless_info(char * buffer, char **start, off_t offset,
+				 int length);
+
+/* Handle IOCTLs, called in net/code/dev.c */
+extern int wireless_process_ioctl(struct ifreq *ifr, unsigned int cmd);
+
+/* Second : functions that may be called by driver modules */
+/* None yet */
+
+#endif	/* _LINUX_WIRELESS_H */

FUNET's LINUX-ADM group, linux-adm@nic.funet.fi
TCL-scripts by Sam Shen (who was at: slshen@lbl.gov)