irqueue.c

linux 内核源代码

/*********************************************************************
 *
 * Filename:      irqueue.c
 * Version:       0.3
 * Description:   General queue implementation
 * Status:        Experimental.
 * Author:        Dag Brattli <[e-mail]>
 * Created at:    Tue Jun  9 13:29:31 1998
 * Modified at:   Sun Dec 12 13:48:22 1999
 * Modified by:   Dag Brattli <[e-mail]>
 * Modified at:   Thu Jan  4 14:29:10 CET 2001
 * Modified by:   Marc Zyngier <[e-mail]>
 *
 *     Copyright (C) 1998-1999, Aage Kvalnes <[e-mail]>
 *     Copyright (C) 1998, Dag Brattli,
 *     All Rights Reserved.
 *
 *     This code is taken from the Vortex Operating System written by Aage
 *     Kvalnes. Aage has agreed that this code can use the GPL licence,
 *     although he does not use that licence in his own code.
 *
 *     This copyright does however _not_ include the ELF hash() function
 *     which I currently don't know which licence or copyright it
 *     has. Please inform me if you know.
 *
 *     This program is free software; you can redistribute it and/or
 *     modify it under the terms of the GNU General Public License as
 *     published by the Free Software Foundation; either version 2 of
 *     the License, or (at your option) any later version.
 *
 *     Neither Dag Brattli nor University of Tromsø admit liability nor
 *     provide warranty for any of this software. This material is
 *     provided "AS-IS" and at no charge.
 *
 ********************************************************************/

/*
 * NOTE :
 * There are various problems with this package :
 *	o the hash function for ints is pathetic (but could be changed)
 *	o locking is sometime suspicious (especially during enumeration)
 *	o most users have only a few elements (== overhead)
 *	o most users never use seach, so don't benefit from hashing
 * Problem already fixed :
 *	o not 64 bit compliant (most users do hashv = (int) self)
 *	o hashbin_remove() is broken => use hashbin_remove_this()
 * I think most users would be better served by a simple linked list
 * (like include/linux/list.h) with a global spinlock per list.
 * Jean II
 */

/*
 * Notes on the concurrent access to hashbin and other SMP issues
 * -------------------------------------------------------------
 *	Hashbins are very often in the IrDA stack a global repository of
 * information, and therefore used in a very asynchronous manner following
 * various events (driver calls, timers, user calls...).
 *	Therefore, very often it is highly important to consider the
 * management of concurrent access to the hashbin and how to guarantee the
 * consistency of the operations on it.
 *
 *	First, we need to define the objective of locking :
 *		1) Protect user data (content pointed by the hashbin)
 *		2) Protect hashbin structure itself (linked list in each bin)
 *
 *			     OLD LOCKING
 *			     -----------
 *
 *	The previous locking strategy, either HB_LOCAL or HB_GLOBAL were
 * both inadequate in *both* aspect.
 *		o HB_GLOBAL was using a spinlock for each bin (local locking).
 *		o HB_LOCAL was disabling irq on *all* CPUs, so use a single
 *		  global semaphore.
 *	The problems were :
 *		A) Global irq disabling is no longer supported by the kernel
 *		B) No protection for the hashbin struct global data
 *			o hashbin_delete()
 *			o hb_current
 *		C) No protection for user data in some cases
 *
 *	A) HB_LOCAL use global irq disabling, so doesn't work on kernel
 * 2.5.X. Even when it is supported (kernel 2.4.X and earlier), its
 * performance is not satisfactory on SMP setups. Most hashbins were
 * HB_LOCAL, so (A) definitely need fixing.
 *	B) HB_LOCAL could be modified to fix (B). However, because HB_GLOBAL
 * lock only the individual bins, it will never be able to lock the
 * global data, so can't do (B).
 *	C) Some functions return pointer to data that is still in the
 * hashbin :
 *		o hashbin_find()
 *		o hashbin_get_first()
 *		o hashbin_get_next()
 *	As the data is still in the hashbin, it may be changed or free'd
 * while the caller is examinimg the data. In those case, locking can't
 * be done within the hashbin, but must include use of the data within
 * the caller.
 *	The caller can easily do this with HB_LOCAL (just disable irqs).
 * However, this is impossible with HB_GLOBAL because the caller has no
 * way to know the proper bin, so don't know which spinlock to use.
 *
 *	Quick summary : can no longer use HB_LOCAL, and HB_GLOBAL is
 * fundamentally broken and will never work.
 *
 *			     NEW LOCKING
 *			     -----------
 *
 *	To fix those problems, I've introduce a few changes in the
 * hashbin locking :
 *		1) New HB_LOCK scheme
 *		2) hashbin->hb_spinlock
 *		3) New hashbin usage policy
 *
 * HB_LOCK :
 * -------
 *	HB_LOCK is a locking scheme intermediate between the old HB_LOCAL
 * and HB_GLOBAL. It uses a single spinlock to protect the whole content
 * of the hashbin. As it is a single spinlock, it can protect the global
 * data of the hashbin and not only the bins themselves.
 *	HB_LOCK can only protect some of the hashbin calls, so it only lock
 * call that can be made 100% safe and leave other call unprotected.
 *	HB_LOCK in theory is slower than HB_GLOBAL, but as the hashbin
 * content is always small contention is not high, so it doesn't matter
 * much. HB_LOCK is probably faster than HB_LOCAL.
 *
 * hashbin->hb_spinlock :
 * --------------------
 *	The spinlock that HB_LOCK uses is available for caller, so that
 * the caller can protect unprotected calls (see below).
 *	If the caller want to do entirely its own locking (HB_NOLOCK), he
 * can do so and may use safely this spinlock.
 *	Locking is done like this :
 *		spin_lock_irqsave(&hashbin->hb_spinlock, flags);
 *	Releasing the lock :
 *		spin_unlock_irqrestore(&hashbin->hb_spinlock, flags);
 *
 * Safe & Protected calls :
 * ----------------------
 *	The following calls are safe or protected via HB_LOCK :
 *		o hashbin_new()		-> safe
 *		o hashbin_delete()
 *		o hashbin_insert()
 *		o hashbin_remove_first()
 *		o hashbin_remove()
 *		o hashbin_remove_this()
 *		o HASHBIN_GET_SIZE()	-> atomic
 *
 *	The following calls only protect the hashbin itself :
 *		o hashbin_lock_find()
 *		o hashbin_find_next()
 *
 * Unprotected calls :
 * -----------------
 *	The following calls need to be protected by the caller :
 *		o hashbin_find()
 *		o hashbin_get_first()
 *		o hashbin_get_next()
 *
 * Locking Policy :
 * --------------
 *	If the hashbin is used only in a single thread of execution
 * (explicitly or implicitely), you can use HB_NOLOCK
 *	If the calling module already provide concurrent access protection,
 * you may use HB_NOLOCK.
 *
 *	In all other cases, you need to use HB_LOCK and lock the hashbin
 * every time before calling one of the unprotected calls. You also must
 * use the pointer returned by the unprotected call within the locked
 * region.
 *
 * Extra care for enumeration :
 * --------------------------
 *	hashbin_get_first() and hashbin_get_next() use the hashbin to
 * store the current position, in hb_current.
 *	As long as the hashbin remains locked, this is safe. If you unlock
 * the hashbin, the current position may change if anybody else modify
 * or enumerate the hashbin.
 *	Summary : do the full enumeration while locked.
 *
 *	Alternatively, you may use hashbin_find_next(). But, this will
 * be slower, is more complex to use and doesn't protect the hashbin
 * content. So, care is needed here as well.
 *
 * Other issues :
 * ------------
 *	I believe that we are overdoing it by using spin_lock_irqsave()
 * and we should use only spin_lock_bh() or similar. But, I don't have
 * the balls to try it out.
 *	Don't believe that because hashbin are now (somewhat) SMP safe
 * that the rest of the code is. Higher layers tend to be safest,
 * but LAP and LMP would need some serious dedicated love.
 *
 * Jean II
 */
#include <linux/module.h>

#include <net/irda/irda.h>
#include <net/irda/irqueue.h>

/************************ QUEUE SUBROUTINES ************************/

/*
 * Hashbin
 */
#define GET_HASHBIN(x) ( x & HASHBIN_MASK )

/*
 * Function hash (name)
 *
 *    This function hash the input string 'name' using the ELF hash
 *    function for strings.
 */
static __u32 hash( const char* name)
{
	__u32 h = 0;
	__u32 g;

	while(*name) {
		h = (h<<4) + *name++;
		if ((g = (h & 0xf0000000)))
			h ^=g>>24;
		h &=~g;
	}
	return h;
}

/*
 * Function enqueue_first (queue, proc)
 *
 *    Insert item first in queue.
 *
 */
static void enqueue_first(irda_queue_t **queue, irda_queue_t* element)
{

	IRDA_DEBUG( 4, "%s()\n", __FUNCTION__);

	/*
	 * Check if queue is empty.
	 */
	if ( *queue == NULL ) {
		/*
		 * Queue is empty.  Insert one element into the queue.
		 */
		element->q_next = element->q_prev = *queue = element;

	} else {
		/*
		 * Queue is not empty.  Insert element into front of queue.
		 */
		element->q_next          = (*queue);
		(*queue)->q_prev->q_next = element;
		element->q_prev          = (*queue)->q_prev;
		(*queue)->q_prev         = element;
		(*queue)                 = element;
	}
}


/*
 * Function dequeue (queue)
 *
 *    Remove first entry in queue
 *
 */
static irda_queue_t *dequeue_first(irda_queue_t **queue)
{
	irda_queue_t *ret;

	IRDA_DEBUG( 4, "dequeue_first()\n");

	/*
	 * Set return value
	 */
	ret =  *queue;

	if ( *queue == NULL ) {
		/*
		 * Queue was empty.
		 */
	} else if ( (*queue)->q_next == *queue ) {
		/*
		 *  Queue only contained a single element. It will now be
		 *  empty.
		 */
		*queue = NULL;
	} else {
		/*
		 * Queue contained several element.  Remove the first one.
		 */
		(*queue)->q_prev->q_next = (*queue)->q_next;
		(*queue)->q_next->q_prev = (*queue)->q_prev;
		*queue = (*queue)->q_next;
	}

	/*
	 * Return the removed entry (or NULL of queue was empty).
	 */
	return ret;
}

/*
 * Function dequeue_general (queue, element)
 *
 *
 */
static irda_queue_t *dequeue_general(irda_queue_t **queue, irda_queue_t* element)
{
	irda_queue_t *ret;

	IRDA_DEBUG( 4, "dequeue_general()\n");

	/*
	 * Set return value
	 */
	ret =  *queue;

	if ( *queue == NULL ) {
		/*
		 * Queue was empty.
		 */
	} else if ( (*queue)->q_next == *queue ) {
		/*
		 *  Queue only contained a single element. It will now be
		 *  empty.
		 */
		*queue = NULL;

	} else {
		/*
		 *  Remove specific element.
		 */
		element->q_prev->q_next = element->q_next;
		element->q_next->q_prev = element->q_prev;
		if ( (*queue) == element)
			(*queue) = element->q_next;
	}

	/*
	 * Return the removed entry (or NULL of queue was empty).
	 */
	return ret;
}

/************************ HASHBIN MANAGEMENT ************************/

/*
 * Function hashbin_create ( type, name )
 *
 *    Create hashbin!
 *
 */
hashbin_t *hashbin_new(int type)
{
	hashbin_t* hashbin;

	/*
	 * Allocate new hashbin
	 */
	hashbin = kzalloc(sizeof(*hashbin), GFP_ATOMIC);
	if (!hashbin)
		return NULL;

	/*
	 * Initialize structure
	 */
	hashbin->hb_type = type;
	hashbin->magic = HB_MAGIC;
	//hashbin->hb_current = NULL;

	/* Make sure all spinlock's are unlocked */
	if ( hashbin->hb_type & HB_LOCK ) {
		spin_lock_init(&hashbin->hb_spinlock);
	}

	return hashbin;
}
EXPORT_SYMBOL(hashbin_new);


/*
 * Function hashbin_delete (hashbin, free_func)
 *
 *    Destroy hashbin, the free_func can be a user supplied special routine
 *    for deallocating this structure if it's complex. If not the user can
 *    just supply kfree, which should take care of the job.
 */
#ifdef CONFIG_LOCKDEP
static int hashbin_lock_depth = 0;
#endif
int hashbin_delete( hashbin_t* hashbin, FREE_FUNC free_func)
{
	irda_queue_t* queue;
	unsigned long flags = 0;
	int i;

	IRDA_ASSERT(hashbin != NULL, return -1;);
	IRDA_ASSERT(hashbin->magic == HB_MAGIC, return -1;);

	/* Synchronize */
	if ( hashbin->hb_type & HB_LOCK ) {
		spin_lock_irqsave_nested(&hashbin->hb_spinlock, flags,
					 hashbin_lock_depth++);
	}

	/*
	 *  Free the entries in the hashbin, TODO: use hashbin_clear when
	 *  it has been shown to work
	 */
	for (i = 0; i < HASHBIN_SIZE; i ++ ) {
		queue = dequeue_first((irda_queue_t**) &hashbin->hb_queue[i]);
		while (queue ) {
			if (free_func)
				(*free_func)(queue);
			queue = dequeue_first(
				(irda_queue_t**) &hashbin->hb_queue[i]);
		}
	}

	/* Cleanup local data */
	hashbin->hb_current = NULL;
	hashbin->magic = ~HB_MAGIC;

	/* Release lock */
	if ( hashbin->hb_type & HB_LOCK) {
		spin_unlock_irqrestore(&hashbin->hb_spinlock, flags);
#ifdef CONFIG_LOCKDEP
		hashbin_lock_depth--;
#endif
	}

	/*
	 *  Free the hashbin structure
	 */
	kfree(hashbin);

	return 0;
}
EXPORT_SYMBOL(hashbin_delete);

/********************* HASHBIN LIST OPERATIONS *********************/

/*
 * Function hashbin_insert (hashbin, entry, name)
 *
 *    Insert an entry into the hashbin
 *
 */
void hashbin_insert(hashbin_t* hashbin, irda_queue_t* entry, long hashv,
		    const char* name)
{
	unsigned long flags = 0;
	int bin;

	IRDA_DEBUG( 4, "%s()\n", __FUNCTION__);

	IRDA_ASSERT( hashbin != NULL, return;);
	IRDA_ASSERT( hashbin->magic == HB_MAGIC, return;);

	/*
	 * Locate hashbin
	 */