2 * net/sched/cls_u32.c Ugly (or Universal) 32bit key Packet Classifier.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
11 * The filters are packed to hash tables of key nodes
12 * with a set of 32bit key/mask pairs at every node.
13 * Nodes reference next level hash tables etc.
15 * This scheme is the best universal classifier I managed to
16 * invent; it is not super-fast, but it is not slow (provided you
17 * program it correctly), and general enough. And its relative
18 * speed grows as the number of rules becomes larger.
20 * It seems that it represents the best middle point between
21 * speed and manageability both by human and by machine.
23 * It is especially useful for link sharing combined with QoS;
24 * pure RSVP doesn't need such a general approach and can use
25 * much simpler (and faster) schemes, sort of cls_rsvp.c.
27 * JHS: We should remove the CONFIG_NET_CLS_IND from here
28 * eventually when the meta match extension is made available
30 * nfmark match added by Catalin(ux aka Dino) BOIE <catab at umbrella.ro>
33 #include <linux/module.h>
34 #include <linux/slab.h>
35 #include <linux/types.h>
36 #include <linux/kernel.h>
37 #include <linux/string.h>
38 #include <linux/errno.h>
39 #include <linux/percpu.h>
40 #include <linux/rtnetlink.h>
41 #include <linux/skbuff.h>
42 #include <linux/bitmap.h>
43 #include <net/netlink.h>
44 #include <net/act_api.h>
45 #include <net/pkt_cls.h>
48 struct tc_u_knode __rcu *next;
50 struct tc_u_hnode __rcu *ht_up;
52 #ifdef CONFIG_NET_CLS_IND
56 struct tcf_result res;
57 struct tc_u_hnode __rcu *ht_down;
58 #ifdef CONFIG_CLS_U32_PERF
59 struct tc_u32_pcnt __percpu *pf;
61 #ifdef CONFIG_CLS_U32_MARK
64 u32 __percpu *pcpu_success;
68 /* The 'sel' field MUST be the last field in structure to allow for
69 * tc_u32_keys allocated at end of structure.
71 struct tc_u32_sel sel;
75 struct tc_u_hnode __rcu *next;
78 struct tc_u_common *tp_c;
81 struct tc_u_knode __rcu *ht[1];
86 struct tc_u_hnode __rcu *hlist;
93 static inline unsigned int u32_hash_fold(__be32 key,
94 const struct tc_u32_sel *sel,
97 unsigned int h = ntohl(key & sel->hmask) >> fshift;
102 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp, struct tcf_result *res)
105 struct tc_u_knode *knode;
107 } stack[TC_U32_MAXDEPTH];
109 struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
110 unsigned int off = skb_network_offset(skb);
111 struct tc_u_knode *n;
115 #ifdef CONFIG_CLS_U32_PERF
121 n = rcu_dereference_bh(ht->ht[sel]);
125 struct tc_u32_key *key = n->sel.keys;
127 #ifdef CONFIG_CLS_U32_PERF
128 __this_cpu_inc(n->pf->rcnt);
132 #ifdef CONFIG_CLS_U32_MARK
133 if ((skb->mark & n->mask) != n->val) {
134 n = rcu_dereference_bh(n->next);
137 __this_cpu_inc(*n->pcpu_success);
141 for (i = n->sel.nkeys; i > 0; i--, key++) {
142 int toff = off + key->off + (off2 & key->offmask);
145 if (skb_headroom(skb) + toff > INT_MAX)
148 data = skb_header_pointer(skb, toff, 4, &hdata);
151 if ((*data ^ key->val) & key->mask) {
152 n = rcu_dereference_bh(n->next);
155 #ifdef CONFIG_CLS_U32_PERF
156 __this_cpu_inc(n->pf->kcnts[j]);
161 ht = rcu_dereference_bh(n->ht_down);
164 if (n->sel.flags & TC_U32_TERMINAL) {
167 #ifdef CONFIG_NET_CLS_IND
168 if (!tcf_match_indev(skb, n->ifindex)) {
169 n = rcu_dereference_bh(n->next);
173 #ifdef CONFIG_CLS_U32_PERF
174 __this_cpu_inc(n->pf->rhit);
176 r = tcf_exts_exec(skb, &n->exts, res);
178 n = rcu_dereference_bh(n->next);
184 n = rcu_dereference_bh(n->next);
189 if (sdepth >= TC_U32_MAXDEPTH)
191 stack[sdepth].knode = n;
192 stack[sdepth].off = off;
195 ht = rcu_dereference_bh(n->ht_down);
200 data = skb_header_pointer(skb, off + n->sel.hoff, 4,
204 sel = ht->divisor & u32_hash_fold(*data, &n->sel,
207 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
210 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
211 off2 = n->sel.off + 3;
212 if (n->sel.flags & TC_U32_VAROFFSET) {
215 data = skb_header_pointer(skb,
220 off2 += ntohs(n->sel.offmask & *data) >>
225 if (n->sel.flags & TC_U32_EAT) {
236 n = stack[sdepth].knode;
237 ht = rcu_dereference_bh(n->ht_up);
238 off = stack[sdepth].off;
245 net_warn_ratelimited("cls_u32: dead loop\n");
249 static struct tc_u_hnode *
250 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
252 struct tc_u_hnode *ht;
254 for (ht = rtnl_dereference(tp_c->hlist);
256 ht = rtnl_dereference(ht->next))
257 if (ht->handle == handle)
263 static struct tc_u_knode *
264 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
267 struct tc_u_knode *n = NULL;
269 sel = TC_U32_HASH(handle);
270 if (sel > ht->divisor)
273 for (n = rtnl_dereference(ht->ht[sel]);
275 n = rtnl_dereference(n->next))
276 if (n->handle == handle)
283 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
285 struct tc_u_hnode *ht;
286 struct tc_u_common *tp_c = tp->data;
288 if (TC_U32_HTID(handle) == TC_U32_ROOT)
289 ht = rtnl_dereference(tp->root);
291 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
296 if (TC_U32_KEY(handle) == 0)
297 return (unsigned long)ht;
299 return (unsigned long)u32_lookup_key(ht, handle);
302 static void u32_put(struct tcf_proto *tp, unsigned long f)
306 static u32 gen_new_htid(struct tc_u_common *tp_c)
310 /* hgenerator only used inside rtnl lock it is safe to increment
311 * without read _copy_ update semantics
314 if (++tp_c->hgenerator == 0x7FF)
315 tp_c->hgenerator = 1;
316 } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
318 return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
321 static int u32_init(struct tcf_proto *tp)
323 struct tc_u_hnode *root_ht;
324 struct tc_u_common *tp_c;
326 tp_c = tp->q->u32_node;
328 root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
332 root_ht->divisor = 0;
334 root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
335 root_ht->prio = tp->prio;
338 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
344 tp->q->u32_node = tp_c;
348 RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
349 rcu_assign_pointer(tp_c->hlist, root_ht);
350 root_ht->tp_c = tp_c;
352 rcu_assign_pointer(tp->root, root_ht);
357 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n)
359 tcf_unbind_filter(tp, &n->res);
360 tcf_exts_destroy(tp, &n->exts);
362 n->ht_down->refcnt--;
363 #ifdef CONFIG_CLS_U32_PERF
370 static void u32_delete_key_rcu(struct rcu_head *rcu)
372 struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
374 u32_destroy_key(key->tp, key);
377 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
379 struct tc_u_knode __rcu **kp;
380 struct tc_u_knode *pkp;
381 struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
384 kp = &ht->ht[TC_U32_HASH(key->handle)];
385 for (pkp = rtnl_dereference(*kp); pkp;
386 kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
388 RCU_INIT_POINTER(*kp, key->next);
390 call_rcu(&key->rcu, u32_delete_key_rcu);
399 static void u32_clear_hnode(struct tc_u_hnode *ht)
401 struct tc_u_knode *n;
404 for (h = 0; h <= ht->divisor; h++) {
405 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
406 RCU_INIT_POINTER(ht->ht[h],
407 rtnl_dereference(n->next));
408 call_rcu(&n->rcu, u32_delete_key_rcu);
413 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
415 struct tc_u_common *tp_c = tp->data;
416 struct tc_u_hnode __rcu **hn;
417 struct tc_u_hnode *phn;
424 for (phn = rtnl_dereference(*hn);
426 hn = &phn->next, phn = rtnl_dereference(*hn)) {
428 RCU_INIT_POINTER(*hn, ht->next);
437 static void u32_destroy(struct tcf_proto *tp)
439 struct tc_u_common *tp_c = tp->data;
440 struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
442 WARN_ON(root_ht == NULL);
444 if (root_ht && --root_ht->refcnt == 0)
445 u32_destroy_hnode(tp, root_ht);
447 if (--tp_c->refcnt == 0) {
448 struct tc_u_hnode *ht;
450 tp->q->u32_node = NULL;
452 for (ht = rtnl_dereference(tp_c->hlist);
454 ht = rtnl_dereference(ht->next)) {
459 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
460 RCU_INIT_POINTER(tp_c->hlist, ht->next);
470 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
472 struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
473 struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
478 if (TC_U32_KEY(ht->handle))
479 return u32_delete_key(tp, (struct tc_u_knode *)ht);
484 if (ht->refcnt == 1) {
486 u32_destroy_hnode(tp, ht);
494 #define NR_U32_NODE (1<<12)
495 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
497 struct tc_u_knode *n;
499 unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
502 return handle | 0xFFF;
504 for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
506 n = rtnl_dereference(n->next))
507 set_bit(TC_U32_NODE(n->handle), bitmap);
509 i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
510 if (i >= NR_U32_NODE)
511 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
514 return handle | (i >= NR_U32_NODE ? 0xFFF : i);
517 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
518 [TCA_U32_CLASSID] = { .type = NLA_U32 },
519 [TCA_U32_HASH] = { .type = NLA_U32 },
520 [TCA_U32_LINK] = { .type = NLA_U32 },
521 [TCA_U32_DIVISOR] = { .type = NLA_U32 },
522 [TCA_U32_SEL] = { .len = sizeof(struct tc_u32_sel) },
523 [TCA_U32_INDEV] = { .type = NLA_STRING, .len = IFNAMSIZ },
524 [TCA_U32_MARK] = { .len = sizeof(struct tc_u32_mark) },
527 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
528 unsigned long base, struct tc_u_hnode *ht,
529 struct tc_u_knode *n, struct nlattr **tb,
530 struct nlattr *est, bool ovr)
535 tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
536 err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
541 if (tb[TCA_U32_LINK]) {
542 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
543 struct tc_u_hnode *ht_down = NULL, *ht_old;
545 if (TC_U32_KEY(handle))
549 ht_down = u32_lookup_ht(ht->tp_c, handle);
556 ht_old = rtnl_dereference(n->ht_down);
557 rcu_assign_pointer(n->ht_down, ht_down);
562 if (tb[TCA_U32_CLASSID]) {
563 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
564 tcf_bind_filter(tp, &n->res, base);
567 #ifdef CONFIG_NET_CLS_IND
568 if (tb[TCA_U32_INDEV]) {
570 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
576 tcf_exts_change(tp, &n->exts, &e);
580 tcf_exts_destroy(tp, &e);
584 static int u32_change(struct net *net, struct sk_buff *in_skb,
585 struct tcf_proto *tp, unsigned long base, u32 handle,
587 unsigned long *arg, bool ovr)
589 struct tc_u_common *tp_c = tp->data;
590 struct tc_u_hnode *ht;
591 struct tc_u_knode *n;
592 struct tc_u32_sel *s;
593 struct nlattr *opt = tca[TCA_OPTIONS];
594 struct nlattr *tb[TCA_U32_MAX + 1];
597 #ifdef CONFIG_CLS_U32_PERF
602 return handle ? -EINVAL : 0;
604 err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
608 n = (struct tc_u_knode *)*arg;
610 if (TC_U32_KEY(n->handle) == 0)
613 return u32_set_parms(net, tp, base,
614 rtnl_dereference(n->ht_up), n, tb,
618 if (tb[TCA_U32_DIVISOR]) {
619 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
621 if (--divisor > 0x100)
623 if (TC_U32_KEY(handle))
626 handle = gen_new_htid(tp->data);
630 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
635 ht->divisor = divisor;
638 RCU_INIT_POINTER(ht->next, tp_c->hlist);
639 rcu_assign_pointer(tp_c->hlist, ht);
640 *arg = (unsigned long)ht;
644 if (tb[TCA_U32_HASH]) {
645 htid = nla_get_u32(tb[TCA_U32_HASH]);
646 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
647 ht = rtnl_dereference(tp->root);
650 ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
655 ht = rtnl_dereference(tp->root);
659 if (ht->divisor < TC_U32_HASH(htid))
663 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
665 handle = htid | TC_U32_NODE(handle);
667 handle = gen_new_kid(ht, htid);
669 if (tb[TCA_U32_SEL] == NULL)
672 s = nla_data(tb[TCA_U32_SEL]);
674 n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
678 #ifdef CONFIG_CLS_U32_PERF
679 size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
680 n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
687 memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
688 RCU_INIT_POINTER(n->ht_up, ht);
690 n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
691 tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
694 #ifdef CONFIG_CLS_U32_MARK
695 n->pcpu_success = alloc_percpu(u32);
697 if (tb[TCA_U32_MARK]) {
698 struct tc_u32_mark *mark;
700 mark = nla_data(tb[TCA_U32_MARK]);
702 n->mask = mark->mask;
706 err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
708 struct tc_u_knode __rcu **ins;
709 struct tc_u_knode *pins;
711 ins = &ht->ht[TC_U32_HASH(handle)];
712 for (pins = rtnl_dereference(*ins); pins;
713 ins = &pins->next, pins = rtnl_dereference(*ins))
714 if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
717 RCU_INIT_POINTER(n->next, pins);
718 rcu_assign_pointer(*ins, n);
720 *arg = (unsigned long)n;
723 #ifdef CONFIG_CLS_U32_PERF
730 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
732 struct tc_u_common *tp_c = tp->data;
733 struct tc_u_hnode *ht;
734 struct tc_u_knode *n;
740 for (ht = rtnl_dereference(tp_c->hlist);
742 ht = rtnl_dereference(ht->next)) {
743 if (ht->prio != tp->prio)
745 if (arg->count >= arg->skip) {
746 if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
752 for (h = 0; h <= ht->divisor; h++) {
753 for (n = rtnl_dereference(ht->ht[h]);
755 n = rtnl_dereference(n->next)) {
756 if (arg->count < arg->skip) {
760 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
770 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
771 struct sk_buff *skb, struct tcmsg *t)
773 struct tc_u_knode *n = (struct tc_u_knode *)fh;
774 struct tc_u_hnode *ht_up, *ht_down;
780 t->tcm_handle = n->handle;
782 nest = nla_nest_start(skb, TCA_OPTIONS);
784 goto nla_put_failure;
786 if (TC_U32_KEY(n->handle) == 0) {
787 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
788 u32 divisor = ht->divisor + 1;
790 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
791 goto nla_put_failure;
793 #ifdef CONFIG_CLS_U32_PERF
794 struct tc_u32_pcnt *gpf;
798 if (nla_put(skb, TCA_U32_SEL,
799 sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
801 goto nla_put_failure;
803 ht_up = rtnl_dereference(n->ht_up);
805 u32 htid = n->handle & 0xFFFFF000;
806 if (nla_put_u32(skb, TCA_U32_HASH, htid))
807 goto nla_put_failure;
809 if (n->res.classid &&
810 nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
811 goto nla_put_failure;
813 ht_down = rtnl_dereference(n->ht_down);
815 nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
816 goto nla_put_failure;
818 #ifdef CONFIG_CLS_U32_MARK
819 if ((n->val || n->mask)) {
820 struct tc_u32_mark mark = {.val = n->val,
825 for_each_possible_cpu(cpum) {
826 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
831 if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
832 goto nla_put_failure;
836 if (tcf_exts_dump(skb, &n->exts) < 0)
837 goto nla_put_failure;
839 #ifdef CONFIG_NET_CLS_IND
841 struct net_device *dev;
842 dev = __dev_get_by_index(net, n->ifindex);
843 if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
844 goto nla_put_failure;
847 #ifdef CONFIG_CLS_U32_PERF
848 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
849 n->sel.nkeys * sizeof(u64),
852 goto nla_put_failure;
854 for_each_possible_cpu(cpu) {
856 struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
858 gpf->rcnt += pf->rcnt;
859 gpf->rhit += pf->rhit;
860 for (i = 0; i < n->sel.nkeys; i++)
861 gpf->kcnts[i] += pf->kcnts[i];
864 if (nla_put(skb, TCA_U32_PCNT,
865 sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(u64),
868 goto nla_put_failure;
874 nla_nest_end(skb, nest);
876 if (TC_U32_KEY(n->handle))
877 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
878 goto nla_put_failure;
882 nla_nest_cancel(skb, nest);
886 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
888 .classify = u32_classify,
890 .destroy = u32_destroy,
893 .change = u32_change,
894 .delete = u32_delete,
897 .owner = THIS_MODULE,
900 static int __init init_u32(void)
902 pr_info("u32 classifier\n");
903 #ifdef CONFIG_CLS_U32_PERF
904 pr_info(" Performance counters on\n");
906 #ifdef CONFIG_NET_CLS_IND
907 pr_info(" input device check on\n");
909 #ifdef CONFIG_NET_CLS_ACT
910 pr_info(" Actions configured\n");
912 return register_tcf_proto_ops(&cls_u32_ops);
915 static void __exit exit_u32(void)
917 unregister_tcf_proto_ops(&cls_u32_ops);
920 module_init(init_u32)
921 module_exit(exit_u32)
922 MODULE_LICENSE("GPL");