patch-2.4.23 linux-2.4.23/net/ipv4/ipvs/ip_vs_lblc.c

• Lines: 625
• Date: 2003-11-28 10:26:21.000000000 -0800
• Orig file: linux-2.4.22/net/ipv4/ipvs/ip_vs_lblc.c
• Orig date: 1969-12-31 16:00:00.000000000 -0800

diff -urN linux-2.4.22/net/ipv4/ipvs/ip_vs_lblc.c linux-2.4.23/net/ipv4/ipvs/ip_vs_lblc.c
@@ -0,0 +1,624 @@
+/*
+ * IPVS:        Locality-Based Least-Connection scheduling module
+ *
+ * Version:     $Id: ip_vs_lblc.c,v 1.9 2002/03/25 12:44:35 wensong Exp $
+ *
+ * Authors:     Wensong Zhang <wensong@gnuchina.org>
+ *
+ *              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.
+ *
+ * Changes:
+ *     Martin Hamilton         :    fixed the terrible locking bugs
+ *                                   *lock(tbl->lock) ==> *lock(&tbl->lock)
+ *     Wensong Zhang           :    fixed the uninitilized tbl->lock bug
+ *     Wensong Zhang           :    added doing full expiration check to
+ *                                   collect stale entries of 24+ hours when
+ *                                   no partial expire check in a half hour
+ *     Julian Anastasov        :    replaced del_timer call with del_timer_sync
+ *                                   to avoid the possible race between timer
+ *                                   handler and del_timer thread in SMP
+ *
+ */
+
+/*
+ * The lblc algorithm is as follows (pseudo code):
+ *
+ *       if cachenode[dest_ip] is null then
+ *               n, cachenode[dest_ip] <- {weighted least-conn node};
+ *       else
+ *               n <- cachenode[dest_ip];
+ *               if (n is dead) OR
+ *                  (n.conns>n.weight AND
+ *                   there is a node m with m.conns<m.weight/2) then
+ *                 n, cachenode[dest_ip] <- {weighted least-conn node};
+ *
+ *       return n;
+ *
+ * Thanks must go to Wenzhuo Zhang for talking WCCP to me and pushing
+ * me to write this module.
+ */
+
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/errno.h>
+
+/* for systcl */
+#include <linux/fs.h>
+#include <linux/sysctl.h>
+
+#include <net/ip_vs.h>
+
+
+/*
+ *    It is for garbage collection of stale IPVS lblc entries,
+ *    when the table is full.
+ */
+#define CHECK_EXPIRE_INTERVAL   (60*HZ)
+#define ENTRY_TIMEOUT           (6*60*HZ)
+
+/*
+ *    It is for full expiration check.
+ *    When there is no partial expiration check (garbage collection)
+ *    in a half hour, do a full expiration check to collect stale
+ *    entries that haven't been touched for a day.
+ */
+#define COUNT_FOR_FULL_EXPIRATION   30
+int sysctl_ip_vs_lblc_expiration = 24*60*60*HZ;
+
+
+/*
+ *     for IPVS lblc entry hash table
+ */
+#ifndef CONFIG_IP_VS_LBLC_TAB_BITS
+#define CONFIG_IP_VS_LBLC_TAB_BITS      10
+#endif
+#define IP_VS_LBLC_TAB_BITS     CONFIG_IP_VS_LBLC_TAB_BITS
+#define IP_VS_LBLC_TAB_SIZE     (1 << IP_VS_LBLC_TAB_BITS)
+#define IP_VS_LBLC_TAB_MASK     (IP_VS_LBLC_TAB_SIZE - 1)
+
+
+/*
+ *      IPVS lblc entry represents an association between destination
+ *      IP address and its destination server
+ */
+struct ip_vs_lblc_entry {
+	struct list_head        list;
+	__u32                   addr;           /* destination IP address */
+	struct ip_vs_dest       *dest;          /* real server (cache) */
+	unsigned long           lastuse;        /* last used time */
+};
+
+
+/*
+ *      IPVS lblc hash table
+ */
+struct ip_vs_lblc_table {
+	rwlock_t	        lock;           /* lock for this table */
+	struct list_head        bucket[IP_VS_LBLC_TAB_SIZE];  /* hash bucket */
+	atomic_t                entries;        /* number of entries */
+	int                     max_size;       /* maximum size of entries */
+	struct timer_list       periodic_timer; /* collect stale entries */
+	int                     rover;          /* rover for expire check */
+	int                     counter;        /* counter for no expire */
+};
+
+
+/*
+ *      IPVS LBLC sysctl table
+ */
+struct ip_vs_lblc_sysctl_table {
+	struct ctl_table_header *sysctl_header;
+	ctl_table vs_vars[2];
+	ctl_table vs_dir[2];
+	ctl_table ipv4_dir[2];
+	ctl_table root_dir[2];
+};
+
+
+static struct ip_vs_lblc_sysctl_table lblc_sysctl_table = {
+	NULL,
+	{{NET_IPV4_VS_LBLC_EXPIRE, "lblc_expiration",
+	  &sysctl_ip_vs_lblc_expiration,
+	  sizeof(int), 0644, NULL, &proc_dointvec_jiffies},
+	 {0}},
+	{{NET_IPV4_VS, "vs", NULL, 0, 0555, lblc_sysctl_table.vs_vars},
+	 {0}},
+	{{NET_IPV4, "ipv4", NULL, 0, 0555, lblc_sysctl_table.vs_dir},
+	 {0}},
+	{{CTL_NET, "net", NULL, 0, 0555, lblc_sysctl_table.ipv4_dir},
+	 {0}}
+};
+
+
+/*
+ *      new/free a ip_vs_lblc_entry, which is a mapping of a destionation
+ *      IP address to a server.
+ */
+static inline struct ip_vs_lblc_entry *
+ip_vs_lblc_new(__u32 daddr, struct ip_vs_dest *dest)
+{
+	struct ip_vs_lblc_entry *en;
+
+	en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
+	if (en == NULL) {
+		IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
+		return NULL;
+	}
+
+	INIT_LIST_HEAD(&en->list);
+	en->addr = daddr;
+
+	atomic_inc(&dest->refcnt);
+	en->dest = dest;
+
+	return en;
+}
+
+
+static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
+{
+	list_del(&en->list);
+	/*
+	 * We don't kfree dest because it is refered either by its service
+	 * or the trash dest list.
+	 */
+	atomic_dec(&en->dest->refcnt);
+	kfree(en);
+}
+
+
+/*
+ *	Returns hash value for IPVS LBLC entry
+ */
+static inline unsigned ip_vs_lblc_hashkey(__u32 addr)
+{
+	return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
+}
+
+
+/*
+ *	Hash an entry in the ip_vs_lblc_table.
+ *	returns bool success.
+ */
+static int
+ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
+{
+	unsigned hash;
+
+	if (!list_empty(&en->list)) {
+		IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
+			  "called from %p\n", __builtin_return_address(0));
+		return 0;
+	}
+
+	/*
+	 *	Hash by destination IP address
+	 */
+	hash = ip_vs_lblc_hashkey(en->addr);
+
+	write_lock(&tbl->lock);
+	list_add(&en->list, &tbl->bucket[hash]);
+	atomic_inc(&tbl->entries);
+	write_unlock(&tbl->lock);
+
+	return 1;
+}
+
+
+#if 0000
+/*
+ *	Unhash ip_vs_lblc_entry from ip_vs_lblc_table.
+ *	returns bool success.
+ */
+static int ip_vs_lblc_unhash(struct ip_vs_lblc_table *tbl,
+			     struct ip_vs_lblc_entry *en)
+{
+	if (list_empty(&en->list)) {
+		IP_VS_ERR("ip_vs_lblc_unhash(): request for not hashed entry, "
+			  "called from %p\n", __builtin_return_address(0));
+		return 0;
+	}
+
+	/*
+	 * Remove it from the table
+	 */
+	write_lock(&tbl->lock);
+	list_del(&en->list);
+	INIT_LIST_HEAD(&en->list);
+	write_unlock(&tbl->lock);
+
+	return 1;
+}
+#endif
+
+
+/*
+ *  Get ip_vs_lblc_entry associated with supplied parameters.
+ */
+static inline struct ip_vs_lblc_entry *
+ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __u32 addr)
+{
+	unsigned hash;
+	struct ip_vs_lblc_entry *en;
+	struct list_head *l,*e;
+
+	hash = ip_vs_lblc_hashkey(addr);
+	l = &tbl->bucket[hash];
+
+	read_lock(&tbl->lock);
+
+	for (e=l->next; e!=l; e=e->next) {
+		en = list_entry(e, struct ip_vs_lblc_entry, list);
+		if (en->addr == addr) {
+			/* HIT */
+			read_unlock(&tbl->lock);
+			return en;
+		}
+	}
+
+	read_unlock(&tbl->lock);
+
+	return NULL;
+}
+
+
+/*
+ *      Flush all the entries of the specified table.
+ */
+static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
+{
+	int i;
+	struct list_head *l;
+	struct ip_vs_lblc_entry *en;
+
+	for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
+		write_lock(&tbl->lock);
+		for (l=&tbl->bucket[i]; l->next!=l; ) {
+			en = list_entry(l->next,
+					struct ip_vs_lblc_entry, list);
+			ip_vs_lblc_free(en);
+			atomic_dec(&tbl->entries);
+		}
+		write_unlock(&tbl->lock);
+	}
+}
+
+
+static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
+{
+	unsigned long now = jiffies;
+	int i, j;
+	struct list_head *l, *e;
+	struct ip_vs_lblc_entry *en;
+
+	for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
+		j = (j + 1) & IP_VS_LBLC_TAB_MASK;
+		e = l = &tbl->bucket[j];
+		write_lock(&tbl->lock);
+		while (e->next != l) {
+			en = list_entry(e->next,
+					struct ip_vs_lblc_entry, list);
+			if ((now - en->lastuse) <
+			    sysctl_ip_vs_lblc_expiration) {
+				e = e->next;
+				continue;
+			}
+			ip_vs_lblc_free(en);
+			atomic_dec(&tbl->entries);
+		}
+		write_unlock(&tbl->lock);
+	}
+	tbl->rover = j;
+}
+
+
+/*
+ *      Periodical timer handler for IPVS lblc table
+ *      It is used to collect stale entries when the number of entries
+ *      exceeds the maximum size of the table.
+ *
+ *      Fixme: we probably need more complicated algorithm to collect
+ *             entries that have not been used for a long time even
+ *             if the number of entries doesn't exceed the maximum size
+ *             of the table.
+ *      The full expiration check is for this purpose now.
+ */
+static void ip_vs_lblc_check_expire(unsigned long data)
+{
+	struct ip_vs_lblc_table *tbl;
+	unsigned long now = jiffies;
+	int goal;
+	int i, j;
+	struct list_head *l, *e;
+	struct ip_vs_lblc_entry *en;
+
+	tbl = (struct ip_vs_lblc_table *)data;
+
+	if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
+		/* do full expiration check */
+		ip_vs_lblc_full_check(tbl);
+		tbl->counter = 1;
+		goto out;
+	}
+
+	if (atomic_read(&tbl->entries) <= tbl->max_size) {
+		tbl->counter++;
+		goto out;
+	}
+
+	goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
+	if (goal > tbl->max_size/2)
+		goal = tbl->max_size/2;
+
+	for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
+		j = (j + 1) & IP_VS_LBLC_TAB_MASK;
+		e = l = &tbl->bucket[j];
+		write_lock(&tbl->lock);
+		while (e->next != l) {
+			en = list_entry(e->next,
+					struct ip_vs_lblc_entry, list);
+			if ((now - en->lastuse) < ENTRY_TIMEOUT) {
+				e = e->next;
+				continue;
+			}
+			ip_vs_lblc_free(en);
+			atomic_dec(&tbl->entries);
+			goal--;
+		}
+		write_unlock(&tbl->lock);
+		if (goal <= 0)
+			break;
+	}
+	tbl->rover = j;
+
+  out:
+	mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
+}
+
+
+static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
+{
+	int i;
+	struct ip_vs_lblc_table *tbl;
+
+	/*
+	 *    Allocate the ip_vs_lblc_table for this service
+	 */
+	tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
+	if (tbl == NULL) {
+		IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
+		return -ENOMEM;
+	}
+	svc->sched_data = tbl;
+	IP_VS_DBG(6, "LBLC hash table (memory=%dbytes) allocated for "
+		  "current service\n",
+		  sizeof(struct ip_vs_lblc_table));
+
+	/*
+	 *    Initialize the hash buckets
+	 */
+	for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
+		INIT_LIST_HEAD(&tbl->bucket[i]);
+	}
+	tbl->lock = RW_LOCK_UNLOCKED;
+	tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
+	tbl->rover = 0;
+	tbl->counter = 1;
+
+	/*
+	 *    Hook periodic timer for garbage collection
+	 */
+	init_timer(&tbl->periodic_timer);
+	tbl->periodic_timer.data = (unsigned long)tbl;
+	tbl->periodic_timer.function = ip_vs_lblc_check_expire;
+	tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
+	add_timer(&tbl->periodic_timer);
+
+	return 0;
+}
+
+
+static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
+{
+	struct ip_vs_lblc_table *tbl = svc->sched_data;
+
+	/* remove periodic timer */
+	del_timer_sync(&tbl->periodic_timer);
+
+	/* got to clean up table entries here */
+	ip_vs_lblc_flush(tbl);
+
+	/* release the table itself */
+	kfree(svc->sched_data);
+	IP_VS_DBG(6, "LBLC hash table (memory=%dbytes) released\n",
+		  sizeof(struct ip_vs_lblc_table));
+
+	return 0;
+}
+
+
+static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
+{
+	return 0;
+}
+
+
+static inline struct ip_vs_dest *
+__ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
+{
+	register struct list_head *l, *e;
+	struct ip_vs_dest *dest, *least;
+	int loh, doh;
+
+	/*
+	 * We think the overhead of processing active connections is fifty
+	 * times higher than that of inactive connections in average. (This
+	 * fifty times might not be accurate, we will change it later.) We
+	 * use the following formula to estimate the overhead:
+	 *                dest->activeconns*50 + dest->inactconns
+	 * and the load:
+	 *                (dest overhead) / dest->weight
+	 *
+	 * Remember -- no floats in kernel mode!!!
+	 * The comparison of h1*w2 > h2*w1 is equivalent to that of
+	 *                h1/w1 > h2/w2
+	 * if every weight is larger than zero.
+	 *
+	 * The server with weight=0 is quiesced and will not receive any
+	 * new connection.
+	 */
+
+	l = &svc->destinations;
+	for (e=l->next; e!=l; e=e->next) {
+		least = list_entry(e, struct ip_vs_dest, n_list);
+		if (atomic_read(&least->weight) > 0) {
+			loh = atomic_read(&least->activeconns) * 50
+				+ atomic_read(&least->inactconns);
+			goto nextstage;
+		}
+	}
+	return NULL;
+
+	/*
+	 *    Find the destination with the least load.
+	 */
+  nextstage:
+	for (e=e->next; e!=l; e=e->next) {
+		dest = list_entry(e, struct ip_vs_dest, n_list);
+		doh = atomic_read(&dest->activeconns) * 50
+			+ atomic_read(&dest->inactconns);
+		if (loh * atomic_read(&dest->weight) >
+		    doh * atomic_read(&least->weight)) {
+			least = dest;
+			loh = doh;
+		}
+	}
+
+	IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
+		  "activeconns %d refcnt %d weight %d overhead %d\n",
+		  NIPQUAD(least->addr), ntohs(least->port),
+		  atomic_read(&least->activeconns),
+		  atomic_read(&least->refcnt),
+		  atomic_read(&least->weight), loh);
+
+	return least;
+}
+
+
+/*
+ *   If this destination server is overloaded and there is a less loaded
+ *   server, then return true.
+ */
+static inline int
+is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
+{
+	if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
+		register struct list_head *l, *e;
+		struct ip_vs_dest *d;
+
+		l = &svc->destinations;
+		for (e=l->next; e!=l; e=e->next) {
+			d = list_entry(e, struct ip_vs_dest, n_list);
+			if (atomic_read(&d->activeconns)*2
+			    < atomic_read(&d->weight)) {
+				return 1;
+			}
+		}
+	}
+	return 0;
+}
+
+
+/*
+ *    Locality-Based (weighted) Least-Connection scheduling
+ */
+static struct ip_vs_dest *
+ip_vs_lblc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
+{
+	struct ip_vs_dest *dest;
+	struct ip_vs_lblc_table *tbl;
+	struct ip_vs_lblc_entry *en;
+
+	IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
+
+	tbl = (struct ip_vs_lblc_table *)svc->sched_data;
+	en = ip_vs_lblc_get(tbl, iph->daddr);
+	if (en == NULL) {
+		dest = __ip_vs_wlc_schedule(svc, iph);
+		if (dest == NULL) {
+			IP_VS_DBG(1, "no destination available\n");
+			return NULL;
+		}
+		en = ip_vs_lblc_new(iph->daddr, dest);
+		if (en == NULL) {
+			return NULL;
+		}
+		ip_vs_lblc_hash(tbl, en);
+	} else {
+		dest = en->dest;
+		if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
+		    || atomic_read(&dest->weight) <= 0
+		    || is_overloaded(dest, svc)) {
+			dest = __ip_vs_wlc_schedule(svc, iph);
+			if (dest == NULL) {
+				IP_VS_DBG(1, "no destination available\n");
+				return NULL;
+			}
+			atomic_dec(&en->dest->refcnt);
+			atomic_inc(&dest->refcnt);
+			en->dest = dest;
+		}
+	}
+	en->lastuse = jiffies;
+
+	IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
+		  "--> server %u.%u.%u.%u:%d\n",
+		  NIPQUAD(en->addr),
+		  NIPQUAD(dest->addr),
+		  ntohs(dest->port));
+
+	return dest;
+}
+
+
+/*
+ *      IPVS LBLC Scheduler structure
+ */
+static struct ip_vs_scheduler ip_vs_lblc_scheduler =
+{
+	{0},                    /* n_list */
+	"lblc",                 /* name */
+	ATOMIC_INIT(0),         /* refcnt */
+	THIS_MODULE,		/* this module */
+	ip_vs_lblc_init_svc,    /* service initializer */
+	ip_vs_lblc_done_svc,    /* service done */
+	ip_vs_lblc_update_svc,  /* service updater */
+	ip_vs_lblc_schedule,    /* select a server from the destination list */
+};
+
+
+static int __init ip_vs_lblc_init(void)
+{
+	INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
+	lblc_sysctl_table.sysctl_header =
+		register_sysctl_table(lblc_sysctl_table.root_dir, 0);
+	return register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
+}
+
+
+static void __exit ip_vs_lblc_cleanup(void)
+{
+	unregister_sysctl_table(lblc_sysctl_table.sysctl_header);
+	unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
+}
+
+
+module_init(ip_vs_lblc_init);
+module_exit(ip_vs_lblc_cleanup);
+MODULE_LICENSE("GPL");