2 * net/sched/cls_tcindex.c Packet classifier for skb->tc_index
4 * Written 1998,1999 by Werner Almesberger, EPFL ICA
7 #include <linux/module.h>
8 #include <linux/types.h>
9 #include <linux/kernel.h>
10 #include <linux/skbuff.h>
11 #include <linux/errno.h>
12 #include <linux/slab.h>
13 #include <net/act_api.h>
14 #include <net/netlink.h>
15 #include <net/pkt_cls.h>
18 * Passing parameters to the root seems to be done more awkwardly than really
19 * necessary. At least, u32 doesn't seem to use such dirty hacks. To be
23 #define PERFECT_HASH_THRESHOLD 64 /* use perfect hash if not bigger */
24 #define DEFAULT_HASH_SIZE 64 /* optimized for diffserv */
27 struct tcindex_filter_result {
29 struct tcf_result res;
33 struct tcindex_filter {
35 struct tcindex_filter_result result;
36 struct tcindex_filter __rcu *next;
42 struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
43 struct tcindex_filter __rcu **h; /* imperfect hash; */
45 u16 mask; /* AND key with mask */
46 u32 shift; /* shift ANDed key to the right */
47 u32 hash; /* hash table size; 0 if undefined */
48 u32 alloc_hash; /* allocated size */
49 u32 fall_through; /* 0: only classify if explicit match */
54 tcindex_filter_is_set(struct tcindex_filter_result *r)
56 return tcf_exts_is_predicative(&r->exts) || r->res.classid;
59 static struct tcindex_filter_result *
60 tcindex_lookup(struct tcindex_data *p, u16 key)
63 struct tcindex_filter_result *f = p->perfect + key;
65 return tcindex_filter_is_set(f) ? f : NULL;
67 struct tcindex_filter __rcu **fp;
68 struct tcindex_filter *f;
70 fp = &p->h[key % p->hash];
71 for (f = rcu_dereference_bh_rtnl(*fp);
73 fp = &f->next, f = rcu_dereference_bh_rtnl(*fp))
82 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
83 struct tcf_result *res)
85 struct tcindex_data *p = rcu_dereference_bh(tp->root);
86 struct tcindex_filter_result *f;
87 int key = (skb->tc_index & p->mask) >> p->shift;
89 pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
92 f = tcindex_lookup(p, key);
96 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
98 pr_debug("alg 0x%x\n", res->classid);
102 pr_debug("map 0x%x\n", res->classid);
104 return tcf_exts_exec(skb, &f->exts, res);
108 static unsigned long tcindex_get(struct tcf_proto *tp, u32 handle)
110 struct tcindex_data *p = rtnl_dereference(tp->root);
111 struct tcindex_filter_result *r;
113 pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
114 if (p->perfect && handle >= p->alloc_hash)
116 r = tcindex_lookup(p, handle);
117 return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
120 static int tcindex_init(struct tcf_proto *tp)
122 struct tcindex_data *p;
124 pr_debug("tcindex_init(tp %p)\n", tp);
125 p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
130 p->hash = DEFAULT_HASH_SIZE;
133 rcu_assign_pointer(tp->root, p);
137 static void tcindex_destroy_rexts(struct rcu_head *head)
139 struct tcindex_filter_result *r;
141 r = container_of(head, struct tcindex_filter_result, rcu);
142 tcf_exts_destroy(&r->exts);
145 static void tcindex_destroy_fexts(struct rcu_head *head)
147 struct tcindex_filter *f = container_of(head, struct tcindex_filter, rcu);
149 tcf_exts_destroy(&f->result.exts);
153 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
155 struct tcindex_data *p = rtnl_dereference(tp->root);
156 struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
157 struct tcindex_filter __rcu **walk;
158 struct tcindex_filter *f = NULL;
160 pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p\n", tp, arg, p);
167 for (i = 0; i < p->hash; i++) {
169 for (f = rtnl_dereference(*walk); f;
170 walk = &f->next, f = rtnl_dereference(*walk)) {
178 rcu_assign_pointer(*walk, rtnl_dereference(f->next));
180 tcf_unbind_filter(tp, &r->res);
181 /* all classifiers are required to call tcf_exts_destroy() after rcu
182 * grace period, since converted-to-rcu actions are relying on that
183 * in cleanup() callback
186 call_rcu(&f->rcu, tcindex_destroy_fexts);
188 call_rcu(&r->rcu, tcindex_destroy_rexts);
192 static int tcindex_destroy_element(struct tcf_proto *tp,
194 struct tcf_walker *walker)
196 return tcindex_delete(tp, arg);
199 static void __tcindex_destroy(struct rcu_head *head)
201 struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
209 valid_perfect_hash(struct tcindex_data *p)
211 return p->hash > (p->mask >> p->shift);
214 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
215 [TCA_TCINDEX_HASH] = { .type = NLA_U32 },
216 [TCA_TCINDEX_MASK] = { .type = NLA_U16 },
217 [TCA_TCINDEX_SHIFT] = { .type = NLA_U32 },
218 [TCA_TCINDEX_FALL_THROUGH] = { .type = NLA_U32 },
219 [TCA_TCINDEX_CLASSID] = { .type = NLA_U32 },
222 static void tcindex_filter_result_init(struct tcindex_filter_result *r)
224 memset(r, 0, sizeof(*r));
225 tcf_exts_init(&r->exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
228 static void __tcindex_partial_destroy(struct rcu_head *head)
230 struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
237 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
238 u32 handle, struct tcindex_data *p,
239 struct tcindex_filter_result *r, struct nlattr **tb,
240 struct nlattr *est, bool ovr)
243 struct tcindex_filter_result new_filter_result, *old_r = r;
244 struct tcindex_filter_result cr;
245 struct tcindex_data *cp, *oldp;
246 struct tcindex_filter *f = NULL; /* make gcc behave */
249 tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
250 err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
255 /* tcindex_data attributes must look atomic to classifier/lookup so
256 * allocate new tcindex data and RCU assign it onto root. Keeping
257 * perfect hash and hash pointers from old data.
259 cp = kzalloc(sizeof(*cp), GFP_KERNEL);
264 cp->shift = p->shift;
266 cp->alloc_hash = p->alloc_hash;
267 cp->fall_through = p->fall_through;
273 cp->perfect = kmemdup(p->perfect,
274 sizeof(*r) * cp->hash, GFP_KERNEL);
277 for (i = 0; i < cp->hash; i++)
278 tcf_exts_init(&cp->perfect[i].exts,
279 TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
284 tcindex_filter_result_init(&new_filter_result);
285 tcindex_filter_result_init(&cr);
289 if (tb[TCA_TCINDEX_HASH])
290 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
292 if (tb[TCA_TCINDEX_MASK])
293 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
295 if (tb[TCA_TCINDEX_SHIFT])
296 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
300 /* Hash already allocated, make sure that we still meet the
301 * requirements for the allocated hash.
304 if (!valid_perfect_hash(cp) ||
305 cp->hash > cp->alloc_hash)
307 } else if (cp->h && cp->hash != cp->alloc_hash) {
312 if (tb[TCA_TCINDEX_FALL_THROUGH])
313 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
316 /* Hash not specified, use perfect hash if the upper limit
317 * of the hashing index is below the threshold.
319 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
320 cp->hash = (cp->mask >> cp->shift) + 1;
322 cp->hash = DEFAULT_HASH_SIZE;
325 if (!cp->perfect && !cp->h)
326 cp->alloc_hash = cp->hash;
328 /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
329 * but then, we'd fail handles that may become valid after some future
330 * mask change. While this is extremely unlikely to ever matter,
331 * the check below is safer (and also more backwards-compatible).
333 if (cp->perfect || valid_perfect_hash(cp))
334 if (handle >= cp->alloc_hash)
339 if (!cp->perfect && !cp->h) {
340 if (valid_perfect_hash(cp)) {
343 cp->perfect = kcalloc(cp->hash, sizeof(*r), GFP_KERNEL);
346 for (i = 0; i < cp->hash; i++)
347 tcf_exts_init(&cp->perfect[i].exts,
352 struct tcindex_filter __rcu **hash;
354 hash = kcalloc(cp->hash,
355 sizeof(struct tcindex_filter *),
367 r = cp->perfect + handle;
369 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
371 if (r == &new_filter_result) {
372 f = kzalloc(sizeof(*f), GFP_KERNEL);
376 tcindex_filter_result_init(&f->result);
380 if (tb[TCA_TCINDEX_CLASSID]) {
381 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
382 tcf_bind_filter(tp, &cr.res, base);
386 tcf_exts_change(tp, &r->exts, &e);
388 tcf_exts_change(tp, &cr.exts, &e);
390 if (old_r && old_r != r)
391 tcindex_filter_result_init(old_r);
395 rcu_assign_pointer(tp->root, cp);
397 if (r == &new_filter_result) {
398 struct tcindex_filter *nfp;
399 struct tcindex_filter __rcu **fp;
401 tcf_exts_change(tp, &f->result.exts, &r->exts);
403 fp = cp->h + (handle % cp->hash);
404 for (nfp = rtnl_dereference(*fp);
406 fp = &nfp->next, nfp = rtnl_dereference(*fp))
409 rcu_assign_pointer(*fp, f);
413 call_rcu(&oldp->rcu, __tcindex_partial_destroy);
419 else if (balloc == 2)
423 tcf_exts_destroy(&e);
428 tcindex_change(struct net *net, struct sk_buff *in_skb,
429 struct tcf_proto *tp, unsigned long base, u32 handle,
430 struct nlattr **tca, unsigned long *arg, bool ovr)
432 struct nlattr *opt = tca[TCA_OPTIONS];
433 struct nlattr *tb[TCA_TCINDEX_MAX + 1];
434 struct tcindex_data *p = rtnl_dereference(tp->root);
435 struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
438 pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
439 "p %p,r %p,*arg 0x%lx\n",
440 tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
445 err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
449 return tcindex_set_parms(net, tp, base, handle, p, r, tb,
453 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
455 struct tcindex_data *p = rtnl_dereference(tp->root);
456 struct tcindex_filter *f, *next;
459 pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
461 for (i = 0; i < p->hash; i++) {
462 if (!p->perfect[i].res.class)
464 if (walker->count >= walker->skip) {
466 (unsigned long) (p->perfect+i), walker)
477 for (i = 0; i < p->hash; i++) {
478 for (f = rtnl_dereference(p->h[i]); f; f = next) {
479 next = rtnl_dereference(f->next);
480 if (walker->count >= walker->skip) {
481 if (walker->fn(tp, (unsigned long) &f->result,
492 static bool tcindex_destroy(struct tcf_proto *tp, bool force)
494 struct tcindex_data *p = rtnl_dereference(tp->root);
495 struct tcf_walker walker;
500 pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
503 walker.fn = tcindex_destroy_element;
504 tcindex_walk(tp, &walker);
506 RCU_INIT_POINTER(tp->root, NULL);
507 call_rcu(&p->rcu, __tcindex_destroy);
512 static int tcindex_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
513 struct sk_buff *skb, struct tcmsg *t)
515 struct tcindex_data *p = rtnl_dereference(tp->root);
516 struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
519 pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p\n",
520 tp, fh, skb, t, p, r);
521 pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
523 nest = nla_nest_start(skb, TCA_OPTIONS);
525 goto nla_put_failure;
528 t->tcm_handle = ~0; /* whatever ... */
529 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
530 nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
531 nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
532 nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
533 goto nla_put_failure;
534 nla_nest_end(skb, nest);
537 t->tcm_handle = r - p->perfect;
539 struct tcindex_filter *f;
540 struct tcindex_filter __rcu **fp;
544 for (i = 0; !t->tcm_handle && i < p->hash; i++) {
546 for (f = rtnl_dereference(*fp);
548 fp = &f->next, f = rtnl_dereference(*fp)) {
550 t->tcm_handle = f->key;
554 pr_debug("handle = %d\n", t->tcm_handle);
556 nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
557 goto nla_put_failure;
559 if (tcf_exts_dump(skb, &r->exts) < 0)
560 goto nla_put_failure;
561 nla_nest_end(skb, nest);
563 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
564 goto nla_put_failure;
570 nla_nest_cancel(skb, nest);
574 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
576 .classify = tcindex_classify,
577 .init = tcindex_init,
578 .destroy = tcindex_destroy,
580 .change = tcindex_change,
581 .delete = tcindex_delete,
582 .walk = tcindex_walk,
583 .dump = tcindex_dump,
584 .owner = THIS_MODULE,
587 static int __init init_tcindex(void)
589 return register_tcf_proto_ops(&cls_tcindex_ops);
592 static void __exit exit_tcindex(void)
594 unregister_tcf_proto_ops(&cls_tcindex_ops);
597 module_init(init_tcindex)
598 module_exit(exit_tcindex)
599 MODULE_LICENSE("GPL");