1 /* iptables match extension to limit the number of packets per second
2 * seperately for each hashbucket (sourceip/sourceport/dstip/dstport)
4 * (C) 2003-2004 by Harald Welte <laforge@netfilter.org>
6 * $Id: ipt_hashlimit.c 3244 2004-10-20 16:24:29Z laforge@netfilter.org $
8 * Development of this code was funded by Astaro AG, http://www.astaro.com/
10 * based on ipt_limit.c by:
11 * Jérôme de Vivie <devivie@info.enserb.u-bordeaux.fr>
12 * Hervé Eychenne <eychenne@info.enserb.u-bordeaux.fr>
13 * Rusty Russell <rusty@rustcorp.com.au>
15 * The general idea is to create a hash table for every dstip and have a
16 * seperate limit counter per tuple. This way you can do something like 'limit
17 * the number of syn packets for each of my internal addresses.
19 * Ideally this would just be implemented as a general 'hash' match, which would
20 * allow us to attach any iptables target to it's hash buckets. But this is
21 * not possible in the current iptables architecture. As always, pkttables for
24 #include <linux/module.h>
25 #include <linux/skbuff.h>
26 #include <linux/spinlock.h>
27 #include <linux/random.h>
28 #include <linux/jhash.h>
29 #include <linux/slab.h>
30 #include <linux/vmalloc.h>
31 #include <linux/proc_fs.h>
32 #include <linux/seq_file.h>
33 #include <linux/list.h>
35 #include <linux/netfilter_ipv4/ip_tables.h>
36 #include <linux/netfilter_ipv4/ipt_hashlimit.h>
38 /* FIXME: this is just for IP_NF_ASSERRT */
39 #include <linux/netfilter_ipv4/ip_conntrack.h>
40 #include <linux/mutex.h>
42 MODULE_LICENSE("GPL");
43 MODULE_AUTHOR("Harald Welte <laforge@netfilter.org>");
44 MODULE_DESCRIPTION("iptables match for limiting per hash-bucket");
46 /* need to declare this at the top */
47 static struct proc_dir_entry *hashlimit_procdir;
48 static struct file_operations dl_file_ops;
55 /* ports have to be consecutive !!! */
61 /* static / read-only parts in the beginning */
62 struct hlist_node node;
63 struct dsthash_dst dst;
65 /* modified structure members in the end */
66 unsigned long expires; /* precalculated expiry time */
68 unsigned long prev; /* last modification */
70 u_int32_t credit_cap, cost;
74 struct ipt_hashlimit_htable {
75 struct hlist_node node; /* global list of all htables */
78 struct hashlimit_cfg cfg; /* config */
81 spinlock_t lock; /* lock for list_head */
82 u_int32_t rnd; /* random seed for hash */
83 struct timer_list timer; /* timer for gc */
84 atomic_t count; /* number entries in table */
87 struct proc_dir_entry *pde;
89 struct hlist_head hash[0]; /* hashtable itself */
92 static DEFINE_SPINLOCK(hashlimit_lock); /* protects htables list */
93 static DEFINE_MUTEX(hlimit_mutex); /* additional checkentry protection */
94 static HLIST_HEAD(hashlimit_htables);
95 static kmem_cache_t *hashlimit_cachep __read_mostly;
97 static inline int dst_cmp(const struct dsthash_ent *ent, struct dsthash_dst *b)
99 return (ent->dst.dst_ip == b->dst_ip
100 && ent->dst.dst_port == b->dst_port
101 && ent->dst.src_port == b->src_port
102 && ent->dst.src_ip == b->src_ip);
105 static inline u_int32_t
106 hash_dst(const struct ipt_hashlimit_htable *ht, const struct dsthash_dst *dst)
108 return (jhash_3words(dst->dst_ip, (dst->dst_port<<16 | dst->src_port),
109 dst->src_ip, ht->rnd) % ht->cfg.size);
112 static inline struct dsthash_ent *
113 __dsthash_find(const struct ipt_hashlimit_htable *ht, struct dsthash_dst *dst)
115 struct dsthash_ent *ent;
116 struct hlist_node *pos;
117 u_int32_t hash = hash_dst(ht, dst);
119 if (!hlist_empty(&ht->hash[hash]))
120 hlist_for_each_entry(ent, pos, &ht->hash[hash], node) {
121 if (dst_cmp(ent, dst)) {
129 /* allocate dsthash_ent, initialize dst, put in htable and lock it */
130 static struct dsthash_ent *
131 __dsthash_alloc_init(struct ipt_hashlimit_htable *ht, struct dsthash_dst *dst)
133 struct dsthash_ent *ent;
135 /* initialize hash with random val at the time we allocate
136 * the first hashtable entry */
138 get_random_bytes(&ht->rnd, 4);
141 atomic_read(&ht->count) >= ht->cfg.max) {
142 /* FIXME: do something. question is what.. */
145 "ipt_hashlimit: max count of %u reached\n",
150 ent = kmem_cache_alloc(hashlimit_cachep, GFP_ATOMIC);
154 "ipt_hashlimit: can't allocate dsthash_ent\n");
158 atomic_inc(&ht->count);
160 ent->dst.dst_ip = dst->dst_ip;
161 ent->dst.dst_port = dst->dst_port;
162 ent->dst.src_ip = dst->src_ip;
163 ent->dst.src_port = dst->src_port;
165 hlist_add_head(&ent->node, &ht->hash[hash_dst(ht, dst)]);
171 __dsthash_free(struct ipt_hashlimit_htable *ht, struct dsthash_ent *ent)
173 hlist_del(&ent->node);
174 kmem_cache_free(hashlimit_cachep, ent);
175 atomic_dec(&ht->count);
177 static void htable_gc(unsigned long htlong);
179 static int htable_create(struct ipt_hashlimit_info *minfo)
183 struct ipt_hashlimit_htable *hinfo;
186 size = minfo->cfg.size;
188 size = (((num_physpages << PAGE_SHIFT) / 16384)
189 / sizeof(struct list_head));
190 if (num_physpages > (1024 * 1024 * 1024 / PAGE_SIZE))
195 /* FIXME: don't use vmalloc() here or anywhere else -HW */
196 hinfo = vmalloc(sizeof(struct ipt_hashlimit_htable)
197 + (sizeof(struct list_head) * size));
199 printk(KERN_ERR "ipt_hashlimit: Unable to create hashtable\n");
202 minfo->hinfo = hinfo;
204 /* copy match config into hashtable config */
205 memcpy(&hinfo->cfg, &minfo->cfg, sizeof(hinfo->cfg));
206 hinfo->cfg.size = size;
208 hinfo->cfg.max = 8 * hinfo->cfg.size;
209 else if (hinfo->cfg.max < hinfo->cfg.size)
210 hinfo->cfg.max = hinfo->cfg.size;
212 for (i = 0; i < hinfo->cfg.size; i++)
213 INIT_HLIST_HEAD(&hinfo->hash[i]);
215 atomic_set(&hinfo->count, 0);
216 atomic_set(&hinfo->use, 1);
218 spin_lock_init(&hinfo->lock);
219 hinfo->pde = create_proc_entry(minfo->name, 0, hashlimit_procdir);
224 hinfo->pde->proc_fops = &dl_file_ops;
225 hinfo->pde->data = hinfo;
227 init_timer(&hinfo->timer);
228 hinfo->timer.expires = jiffies + msecs_to_jiffies(hinfo->cfg.gc_interval);
229 hinfo->timer.data = (unsigned long )hinfo;
230 hinfo->timer.function = htable_gc;
231 add_timer(&hinfo->timer);
233 spin_lock_bh(&hashlimit_lock);
234 hlist_add_head(&hinfo->node, &hashlimit_htables);
235 spin_unlock_bh(&hashlimit_lock);
240 static int select_all(struct ipt_hashlimit_htable *ht, struct dsthash_ent *he)
245 static int select_gc(struct ipt_hashlimit_htable *ht, struct dsthash_ent *he)
247 return (jiffies >= he->expires);
250 static void htable_selective_cleanup(struct ipt_hashlimit_htable *ht,
251 int (*select)(struct ipt_hashlimit_htable *ht,
252 struct dsthash_ent *he))
256 IP_NF_ASSERT(ht->cfg.size && ht->cfg.max);
258 /* lock hash table and iterate over it */
259 spin_lock_bh(&ht->lock);
260 for (i = 0; i < ht->cfg.size; i++) {
261 struct dsthash_ent *dh;
262 struct hlist_node *pos, *n;
263 hlist_for_each_entry_safe(dh, pos, n, &ht->hash[i], node) {
264 if ((*select)(ht, dh))
265 __dsthash_free(ht, dh);
268 spin_unlock_bh(&ht->lock);
271 /* hash table garbage collector, run by timer */
272 static void htable_gc(unsigned long htlong)
274 struct ipt_hashlimit_htable *ht = (struct ipt_hashlimit_htable *)htlong;
276 htable_selective_cleanup(ht, select_gc);
278 /* re-add the timer accordingly */
279 ht->timer.expires = jiffies + msecs_to_jiffies(ht->cfg.gc_interval);
280 add_timer(&ht->timer);
283 static void htable_destroy(struct ipt_hashlimit_htable *hinfo)
285 /* remove timer, if it is pending */
286 if (timer_pending(&hinfo->timer))
287 del_timer(&hinfo->timer);
289 /* remove proc entry */
290 remove_proc_entry(hinfo->pde->name, hashlimit_procdir);
292 htable_selective_cleanup(hinfo, select_all);
296 static struct ipt_hashlimit_htable *htable_find_get(char *name)
298 struct ipt_hashlimit_htable *hinfo;
299 struct hlist_node *pos;
301 spin_lock_bh(&hashlimit_lock);
302 hlist_for_each_entry(hinfo, pos, &hashlimit_htables, node) {
303 if (!strcmp(name, hinfo->pde->name)) {
304 atomic_inc(&hinfo->use);
305 spin_unlock_bh(&hashlimit_lock);
309 spin_unlock_bh(&hashlimit_lock);
314 static void htable_put(struct ipt_hashlimit_htable *hinfo)
316 if (atomic_dec_and_test(&hinfo->use)) {
317 spin_lock_bh(&hashlimit_lock);
318 hlist_del(&hinfo->node);
319 spin_unlock_bh(&hashlimit_lock);
320 htable_destroy(hinfo);
325 /* The algorithm used is the Simple Token Bucket Filter (TBF)
326 * see net/sched/sch_tbf.c in the linux source tree
329 /* Rusty: This is my (non-mathematically-inclined) understanding of
330 this algorithm. The `average rate' in jiffies becomes your initial
331 amount of credit `credit' and the most credit you can ever have
332 `credit_cap'. The `peak rate' becomes the cost of passing the
335 `prev' tracks the last packet hit: you gain one credit per jiffy.
336 If you get credit balance more than this, the extra credit is
337 discarded. Every time the match passes, you lose `cost' credits;
338 if you don't have that many, the test fails.
340 See Alexey's formal explanation in net/sched/sch_tbf.c.
342 To get the maximum range, we multiply by this factor (ie. you get N
343 credits per jiffy). We want to allow a rate as low as 1 per day
344 (slowest userspace tool allows), which means
345 CREDITS_PER_JIFFY*HZ*60*60*24 < 2^32 ie.
347 #define MAX_CPJ (0xFFFFFFFF / (HZ*60*60*24))
349 /* Repeated shift and or gives us all 1s, final shift and add 1 gives
350 * us the power of 2 below the theoretical max, so GCC simply does a
352 #define _POW2_BELOW2(x) ((x)|((x)>>1))
353 #define _POW2_BELOW4(x) (_POW2_BELOW2(x)|_POW2_BELOW2((x)>>2))
354 #define _POW2_BELOW8(x) (_POW2_BELOW4(x)|_POW2_BELOW4((x)>>4))
355 #define _POW2_BELOW16(x) (_POW2_BELOW8(x)|_POW2_BELOW8((x)>>8))
356 #define _POW2_BELOW32(x) (_POW2_BELOW16(x)|_POW2_BELOW16((x)>>16))
357 #define POW2_BELOW32(x) ((_POW2_BELOW32(x)>>1) + 1)
359 #define CREDITS_PER_JIFFY POW2_BELOW32(MAX_CPJ)
361 /* Precision saver. */
362 static inline u_int32_t
363 user2credits(u_int32_t user)
365 /* If multiplying would overflow... */
366 if (user > 0xFFFFFFFF / (HZ*CREDITS_PER_JIFFY))
368 return (user / IPT_HASHLIMIT_SCALE) * HZ * CREDITS_PER_JIFFY;
370 return (user * HZ * CREDITS_PER_JIFFY) / IPT_HASHLIMIT_SCALE;
373 static inline void rateinfo_recalc(struct dsthash_ent *dh, unsigned long now)
375 dh->rateinfo.credit += (now - xchg(&dh->rateinfo.prev, now))
377 if (dh->rateinfo.credit > dh->rateinfo.credit_cap)
378 dh->rateinfo.credit = dh->rateinfo.credit_cap;
382 hashlimit_match(const struct sk_buff *skb,
383 const struct net_device *in,
384 const struct net_device *out,
385 const struct xt_match *match,
386 const void *matchinfo,
388 unsigned int protoff,
391 struct ipt_hashlimit_info *r =
392 ((struct ipt_hashlimit_info *)matchinfo)->u.master;
393 struct ipt_hashlimit_htable *hinfo = r->hinfo;
394 unsigned long now = jiffies;
395 struct dsthash_ent *dh;
396 struct dsthash_dst dst;
398 /* build 'dst' according to hinfo->cfg and current packet */
399 memset(&dst, 0, sizeof(dst));
400 if (hinfo->cfg.mode & IPT_HASHLIMIT_HASH_DIP)
401 dst.dst_ip = skb->nh.iph->daddr;
402 if (hinfo->cfg.mode & IPT_HASHLIMIT_HASH_SIP)
403 dst.src_ip = skb->nh.iph->saddr;
404 if (hinfo->cfg.mode & IPT_HASHLIMIT_HASH_DPT
405 ||hinfo->cfg.mode & IPT_HASHLIMIT_HASH_SPT) {
406 u_int16_t _ports[2], *ports;
408 switch (skb->nh.iph->protocol) {
413 ports = skb_header_pointer(skb, skb->nh.iph->ihl*4,
414 sizeof(_ports), &_ports);
417 _ports[0] = _ports[1] = 0;
422 /* We've been asked to examine this packet, and we
423 can't. Hence, no choice but to drop. */
427 if (hinfo->cfg.mode & IPT_HASHLIMIT_HASH_SPT)
428 dst.src_port = ports[0];
429 if (hinfo->cfg.mode & IPT_HASHLIMIT_HASH_DPT)
430 dst.dst_port = ports[1];
433 spin_lock_bh(&hinfo->lock);
434 dh = __dsthash_find(hinfo, &dst);
436 dh = __dsthash_alloc_init(hinfo, &dst);
439 /* enomem... don't match == DROP */
441 printk(KERN_ERR "%s: ENOMEM\n", __FUNCTION__);
442 spin_unlock_bh(&hinfo->lock);
446 dh->expires = jiffies + msecs_to_jiffies(hinfo->cfg.expire);
448 dh->rateinfo.prev = jiffies;
449 dh->rateinfo.credit = user2credits(hinfo->cfg.avg *
451 dh->rateinfo.credit_cap = user2credits(hinfo->cfg.avg *
453 dh->rateinfo.cost = user2credits(hinfo->cfg.avg);
455 spin_unlock_bh(&hinfo->lock);
459 /* update expiration timeout */
460 dh->expires = now + msecs_to_jiffies(hinfo->cfg.expire);
462 rateinfo_recalc(dh, now);
463 if (dh->rateinfo.credit >= dh->rateinfo.cost) {
464 /* We're underlimit. */
465 dh->rateinfo.credit -= dh->rateinfo.cost;
466 spin_unlock_bh(&hinfo->lock);
470 spin_unlock_bh(&hinfo->lock);
472 /* default case: we're overlimit, thus don't match */
477 hashlimit_checkentry(const char *tablename,
479 const struct xt_match *match,
481 unsigned int matchsize,
482 unsigned int hook_mask)
484 struct ipt_hashlimit_info *r = matchinfo;
486 /* Check for overflow. */
487 if (r->cfg.burst == 0
488 || user2credits(r->cfg.avg * r->cfg.burst) <
489 user2credits(r->cfg.avg)) {
490 printk(KERN_ERR "ipt_hashlimit: Overflow, try lower: %u/%u\n",
491 r->cfg.avg, r->cfg.burst);
496 || r->cfg.mode > (IPT_HASHLIMIT_HASH_DPT
497 |IPT_HASHLIMIT_HASH_DIP
498 |IPT_HASHLIMIT_HASH_SIP
499 |IPT_HASHLIMIT_HASH_SPT))
502 if (!r->cfg.gc_interval)
508 /* This is the best we've got: We cannot release and re-grab lock,
509 * since checkentry() is called before ip_tables.c grabs ipt_mutex.
510 * We also cannot grab the hashtable spinlock, since htable_create will
511 * call vmalloc, and that can sleep. And we cannot just re-search
512 * the list of htable's in htable_create(), since then we would
513 * create duplicate proc files. -HW */
514 mutex_lock(&hlimit_mutex);
515 r->hinfo = htable_find_get(r->name);
516 if (!r->hinfo && (htable_create(r) != 0)) {
517 mutex_unlock(&hlimit_mutex);
520 mutex_unlock(&hlimit_mutex);
522 /* Ugly hack: For SMP, we only want to use one set */
529 hashlimit_destroy(const struct xt_match *match, void *matchinfo,
530 unsigned int matchsize)
532 struct ipt_hashlimit_info *r = matchinfo;
534 htable_put(r->hinfo);
537 static struct ipt_match ipt_hashlimit = {
539 .match = hashlimit_match,
540 .matchsize = sizeof(struct ipt_hashlimit_info),
541 .checkentry = hashlimit_checkentry,
542 .destroy = hashlimit_destroy,
548 static void *dl_seq_start(struct seq_file *s, loff_t *pos)
550 struct proc_dir_entry *pde = s->private;
551 struct ipt_hashlimit_htable *htable = pde->data;
552 unsigned int *bucket;
554 spin_lock_bh(&htable->lock);
555 if (*pos >= htable->cfg.size)
558 bucket = kmalloc(sizeof(unsigned int), GFP_ATOMIC);
560 return ERR_PTR(-ENOMEM);
566 static void *dl_seq_next(struct seq_file *s, void *v, loff_t *pos)
568 struct proc_dir_entry *pde = s->private;
569 struct ipt_hashlimit_htable *htable = pde->data;
570 unsigned int *bucket = (unsigned int *)v;
573 if (*pos >= htable->cfg.size) {
580 static void dl_seq_stop(struct seq_file *s, void *v)
582 struct proc_dir_entry *pde = s->private;
583 struct ipt_hashlimit_htable *htable = pde->data;
584 unsigned int *bucket = (unsigned int *)v;
588 spin_unlock_bh(&htable->lock);
591 static inline int dl_seq_real_show(struct dsthash_ent *ent, struct seq_file *s)
593 /* recalculate to show accurate numbers */
594 rateinfo_recalc(ent, jiffies);
596 return seq_printf(s, "%ld %u.%u.%u.%u:%u->%u.%u.%u.%u:%u %u %u %u\n",
597 (long)(ent->expires - jiffies)/HZ,
598 NIPQUAD(ent->dst.src_ip), ntohs(ent->dst.src_port),
599 NIPQUAD(ent->dst.dst_ip), ntohs(ent->dst.dst_port),
600 ent->rateinfo.credit, ent->rateinfo.credit_cap,
604 static int dl_seq_show(struct seq_file *s, void *v)
606 struct proc_dir_entry *pde = s->private;
607 struct ipt_hashlimit_htable *htable = pde->data;
608 unsigned int *bucket = (unsigned int *)v;
609 struct dsthash_ent *ent;
610 struct hlist_node *pos;
612 if (!hlist_empty(&htable->hash[*bucket]))
613 hlist_for_each_entry(ent, pos, &htable->hash[*bucket], node) {
614 if (dl_seq_real_show(ent, s)) {
615 /* buffer was filled and unable to print that tuple */
623 static struct seq_operations dl_seq_ops = {
624 .start = dl_seq_start,
630 static int dl_proc_open(struct inode *inode, struct file *file)
632 int ret = seq_open(file, &dl_seq_ops);
635 struct seq_file *sf = file->private_data;
636 sf->private = PDE(inode);
641 static struct file_operations dl_file_ops = {
642 .owner = THIS_MODULE,
643 .open = dl_proc_open,
646 .release = seq_release
649 static int init_or_fini(int fini)
656 if (ipt_register_match(&ipt_hashlimit)) {
658 goto cleanup_nothing;
661 hashlimit_cachep = kmem_cache_create("ipt_hashlimit",
662 sizeof(struct dsthash_ent), 0,
664 if (!hashlimit_cachep) {
665 printk(KERN_ERR "Unable to create ipt_hashlimit slab cache\n");
667 goto cleanup_unreg_match;
670 hashlimit_procdir = proc_mkdir("ipt_hashlimit", proc_net);
671 if (!hashlimit_procdir) {
672 printk(KERN_ERR "Unable to create proc dir entry\n");
674 goto cleanup_free_slab;
680 remove_proc_entry("ipt_hashlimit", proc_net);
682 kmem_cache_destroy(hashlimit_cachep);
684 ipt_unregister_match(&ipt_hashlimit);
690 static int __init ipt_hashlimit_init(void)
692 return init_or_fini(0);
695 static void __exit ipt_hashlimit_fini(void)
700 module_init(ipt_hashlimit_init);
701 module_exit(ipt_hashlimit_fini);