ip_vs_sh.c

  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * IPVS:        Source Hashing scheduling module
  4 *
  5 * Authors:     Wensong Zhang <wensong@gnuchina.org>
  6 *
  7 * Changes:
  8 */
  9
 10/*
 11 * The sh algorithm is to select server by the hash key of source IP
 12 * address. The pseudo code is as follows:
 13 *
 14 *       n <- servernode[src_ip];
 15 *       if (n is dead) OR
 16 *          (n is overloaded) or (n.weight <= 0) then
 17 *                 return NULL;
 18 *
 19 *       return n;
 20 *
 21 * Notes that servernode is a 256-bucket hash table that maps the hash
 22 * index derived from packet source IP address to the current server
 23 * array. If the sh scheduler is used in cache cluster, it is good to
 24 * combine it with cache_bypass feature. When the statically assigned
 25 * server is dead or overloaded, the load balancer can bypass the cache
 26 * server and send requests to the original server directly.
 27 *
 28 * The weight destination attribute can be used to control the
 29 * distribution of connections to the destinations in servernode. The
 30 * greater the weight, the more connections the destination
 31 * will receive.
 32 *
 33 */
 34
 35#define pr_fmt(fmt) "IPVS: " fmt
 36
 37#include <linux/ip.h>
 38#include <linux/slab.h>
 39#include <linux/module.h>
 40#include <linux/kernel.h>
 41#include <linux/skbuff.h>
 42
 43#include <net/ip_vs.h>
 44
 45#include <net/tcp.h>
 46#include <linux/udp.h>
 47#include <linux/sctp.h>
 48
 49
 50/*
 51 *      IPVS SH bucket
 52 */
 53struct ip_vs_sh_bucket {
 54	struct ip_vs_dest __rcu	*dest;	/* real server (cache) */
 55};
 56
 57/*
 58 *     for IPVS SH entry hash table
 59 */
 60#ifndef CONFIG_IP_VS_SH_TAB_BITS
 61#define CONFIG_IP_VS_SH_TAB_BITS        8
 62#endif
 63#define IP_VS_SH_TAB_BITS               CONFIG_IP_VS_SH_TAB_BITS
 64#define IP_VS_SH_TAB_SIZE               (1 << IP_VS_SH_TAB_BITS)
 65#define IP_VS_SH_TAB_MASK               (IP_VS_SH_TAB_SIZE - 1)
 66
 67struct ip_vs_sh_state {
 68	struct rcu_head			rcu_head;
 69	struct ip_vs_sh_bucket		buckets[IP_VS_SH_TAB_SIZE];
 70};
 71
 72/* Helper function to determine if server is unavailable */
 73static inline bool is_unavailable(struct ip_vs_dest *dest)
 74{
 75	return atomic_read(&dest->weight) <= 0 ||
 76	       dest->flags & IP_VS_DEST_F_OVERLOAD;
 77}
 78
 79/*
 80 *	Returns hash value for IPVS SH entry
 81 */
 82static inline unsigned int
 83ip_vs_sh_hashkey(int af, const union nf_inet_addr *addr,
 84		 __be16 port, unsigned int offset)
 85{
 86	__be32 addr_fold = addr->ip;
 87
 88#ifdef CONFIG_IP_VS_IPV6
 89	if (af == AF_INET6)
 90		addr_fold = addr->ip6[0]^addr->ip6[1]^
 91			    addr->ip6[2]^addr->ip6[3];
 92#endif
 93	return (offset + hash_32(ntohs(port) + ntohl(addr_fold),
 94				 IP_VS_SH_TAB_BITS)) &
 95		IP_VS_SH_TAB_MASK;
 96}
 97
 98
 99/*
100 *      Get ip_vs_dest associated with supplied parameters.
101 */
102static inline struct ip_vs_dest *
103ip_vs_sh_get(struct ip_vs_service *svc, struct ip_vs_sh_state *s,
104	     const union nf_inet_addr *addr, __be16 port)
105{
106	unsigned int hash = ip_vs_sh_hashkey(svc->af, addr, port, 0);
107	struct ip_vs_dest *dest = rcu_dereference(s->buckets[hash].dest);
108
109	return (!dest || is_unavailable(dest)) ? NULL : dest;
110}
111
112
113/* As ip_vs_sh_get, but with fallback if selected server is unavailable
114 *
115 * The fallback strategy loops around the table starting from a "random"
116 * point (in fact, it is chosen to be the original hash value to make the
117 * algorithm deterministic) to find a new server.
118 */
119static inline struct ip_vs_dest *
120ip_vs_sh_get_fallback(struct ip_vs_service *svc, struct ip_vs_sh_state *s,
121		      const union nf_inet_addr *addr, __be16 port)
122{
123	unsigned int offset, roffset;
124	unsigned int hash, ihash;
125	struct ip_vs_dest *dest;
126
127	/* first try the dest it's supposed to go to */
128	ihash = ip_vs_sh_hashkey(svc->af, addr, port, 0);
129	dest = rcu_dereference(s->buckets[ihash].dest);
130	if (!dest)
131		return NULL;
132	if (!is_unavailable(dest))
133		return dest;
134
135	IP_VS_DBG_BUF(6, "SH: selected unavailable server %s:%d, reselecting",
136		      IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port));
137
138	/* if the original dest is unavailable, loop around the table
139	 * starting from ihash to find a new dest
140	 */
141	for (offset = 0; offset < IP_VS_SH_TAB_SIZE; offset++) {
142		roffset = (offset + ihash) % IP_VS_SH_TAB_SIZE;
143		hash = ip_vs_sh_hashkey(svc->af, addr, port, roffset);
144		dest = rcu_dereference(s->buckets[hash].dest);
145		if (!dest)
146			break;
147		if (!is_unavailable(dest))
148			return dest;
149		IP_VS_DBG_BUF(6, "SH: selected unavailable "
150			      "server %s:%d (offset %d), reselecting",
151			      IP_VS_DBG_ADDR(dest->af, &dest->addr),
152			      ntohs(dest->port), roffset);
153	}
154
155	return NULL;
156}
157
158/*
159 *      Assign all the hash buckets of the specified table with the service.
160 */
161static int
162ip_vs_sh_reassign(struct ip_vs_sh_state *s, struct ip_vs_service *svc)
163{
164	int i;
165	struct ip_vs_sh_bucket *b;
166	struct list_head *p;
167	struct ip_vs_dest *dest;
168	int d_count;
169	bool empty;
170
171	b = &s->buckets[0];
172	p = &svc->destinations;
173	empty = list_empty(p);
174	d_count = 0;
175	for (i=0; i<IP_VS_SH_TAB_SIZE; i++) {
176		dest = rcu_dereference_protected(b->dest, 1);
177		if (dest)
178			ip_vs_dest_put(dest);
179		if (empty)
180			RCU_INIT_POINTER(b->dest, NULL);
181		else {
182			if (p == &svc->destinations)
183				p = p->next;
184
185			dest = list_entry(p, struct ip_vs_dest, n_list);
186			ip_vs_dest_hold(dest);
187			RCU_INIT_POINTER(b->dest, dest);
188
189			IP_VS_DBG_BUF(6, "assigned i: %d dest: %s weight: %d\n",
190				      i, IP_VS_DBG_ADDR(dest->af, &dest->addr),
191				      atomic_read(&dest->weight));
192
193			/* Don't move to next dest until filling weight */
194			if (++d_count >= atomic_read(&dest->weight)) {
195				p = p->next;
196				d_count = 0;
197			}
198
199		}
200		b++;
201	}
202	return 0;
203}
204
205
206/*
207 *      Flush all the hash buckets of the specified table.
208 */
209static void ip_vs_sh_flush(struct ip_vs_sh_state *s)
210{
211	int i;
212	struct ip_vs_sh_bucket *b;
213	struct ip_vs_dest *dest;
214
215	b = &s->buckets[0];
216	for (i=0; i<IP_VS_SH_TAB_SIZE; i++) {
217		dest = rcu_dereference_protected(b->dest, 1);
218		if (dest) {
219			ip_vs_dest_put(dest);
220			RCU_INIT_POINTER(b->dest, NULL);
221		}
222		b++;
223	}
224}
225
226
227static int ip_vs_sh_init_svc(struct ip_vs_service *svc)
228{
229	struct ip_vs_sh_state *s;
230
231	/* allocate the SH table for this service */
232	s = kzalloc(sizeof(struct ip_vs_sh_state), GFP_KERNEL);
233	if (s == NULL)
234		return -ENOMEM;
235
236	svc->sched_data = s;
237	IP_VS_DBG(6, "SH hash table (memory=%zdbytes) allocated for "
238		  "current service\n",
239		  sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE);
240
241	/* assign the hash buckets with current dests */
242	ip_vs_sh_reassign(s, svc);
243
244	return 0;
245}
246
247
248static void ip_vs_sh_done_svc(struct ip_vs_service *svc)
249{
250	struct ip_vs_sh_state *s = svc->sched_data;
251
252	/* got to clean up hash buckets here */
253	ip_vs_sh_flush(s);
254
255	/* release the table itself */
256	kfree_rcu(s, rcu_head);
257	IP_VS_DBG(6, "SH hash table (memory=%zdbytes) released\n",
258		  sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE);
259}
260
261
262static int ip_vs_sh_dest_changed(struct ip_vs_service *svc,
263				 struct ip_vs_dest *dest)
264{
265	struct ip_vs_sh_state *s = svc->sched_data;
266
267	/* assign the hash buckets with the updated service */
268	ip_vs_sh_reassign(s, svc);
269
270	return 0;
271}
272
273
274/* Helper function to get port number */
275static inline __be16
276ip_vs_sh_get_port(const struct sk_buff *skb, struct ip_vs_iphdr *iph)
277{
278	__be16 _ports[2], *ports;
279
280	/* At this point we know that we have a valid packet of some kind.
281	 * Because ICMP packets are only guaranteed to have the first 8
282	 * bytes, let's just grab the ports.  Fortunately they're in the
283	 * same position for all three of the protocols we care about.
284	 */
285	switch (iph->protocol) {
286	case IPPROTO_TCP:
287	case IPPROTO_UDP:
288	case IPPROTO_SCTP:
289		ports = skb_header_pointer(skb, iph->len, sizeof(_ports),
290					   &_ports);
291		if (unlikely(!ports))
292			return 0;
293
294		if (likely(!ip_vs_iph_inverse(iph)))
295			return ports[0];
296		else
297			return ports[1];
298	default:
299		return 0;
300	}
301}
302
303
304/*
305 *      Source Hashing scheduling
306 */
307static struct ip_vs_dest *
308ip_vs_sh_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
309		  struct ip_vs_iphdr *iph)
310{
311	struct ip_vs_dest *dest;
312	struct ip_vs_sh_state *s;
313	__be16 port = 0;
314	const union nf_inet_addr *hash_addr;
315
316	hash_addr = ip_vs_iph_inverse(iph) ? &iph->daddr : &iph->saddr;
317
318	IP_VS_DBG(6, "ip_vs_sh_schedule(): Scheduling...\n");
319
320	if (svc->flags & IP_VS_SVC_F_SCHED_SH_PORT)
321		port = ip_vs_sh_get_port(skb, iph);
322
323	s = (struct ip_vs_sh_state *) svc->sched_data;
324
325	if (svc->flags & IP_VS_SVC_F_SCHED_SH_FALLBACK)
326		dest = ip_vs_sh_get_fallback(svc, s, hash_addr, port);
327	else
328		dest = ip_vs_sh_get(svc, s, hash_addr, port);
329
330	if (!dest) {
331		ip_vs_scheduler_err(svc, "no destination available");
332		return NULL;
333	}
334
335	IP_VS_DBG_BUF(6, "SH: source IP address %s --> server %s:%d\n",
336		      IP_VS_DBG_ADDR(svc->af, hash_addr),
337		      IP_VS_DBG_ADDR(dest->af, &dest->addr),
338		      ntohs(dest->port));
339
340	return dest;
341}
342
343
344/*
345 *      IPVS SH Scheduler structure
346 */
347static struct ip_vs_scheduler ip_vs_sh_scheduler =
348{
349	.name =			"sh",
350	.refcnt =		ATOMIC_INIT(0),
351	.module =		THIS_MODULE,
352	.n_list	 =		LIST_HEAD_INIT(ip_vs_sh_scheduler.n_list),
353	.init_service =		ip_vs_sh_init_svc,
354	.done_service =		ip_vs_sh_done_svc,
355	.add_dest =		ip_vs_sh_dest_changed,
356	.del_dest =		ip_vs_sh_dest_changed,
357	.upd_dest =		ip_vs_sh_dest_changed,
358	.schedule =		ip_vs_sh_schedule,
359};
360
361
362static int __init ip_vs_sh_init(void)
363{
364	return register_ip_vs_scheduler(&ip_vs_sh_scheduler);
365}
366
367
368static void __exit ip_vs_sh_cleanup(void)
369{
370	unregister_ip_vs_scheduler(&ip_vs_sh_scheduler);
371	synchronize_rcu();
372}
373
374
375module_init(ip_vs_sh_init);
376module_exit(ip_vs_sh_cleanup);
377MODULE_LICENSE("GPL");
378MODULE_DESCRIPTION("ipvs source hashing scheduler");