GNU Linux-libre 4.19.207-gnu1
[releases.git] / net / sched / cls_tcindex.c
1 /*
2  * net/sched/cls_tcindex.c      Packet classifier for skb->tc_index
3  *
4  * Written 1998,1999 by Werner Almesberger, EPFL ICA
5  */
6
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>
16 #include <net/sch_generic.h>
17
18 /*
19  * Passing parameters to the root seems to be done more awkwardly than really
20  * necessary. At least, u32 doesn't seem to use such dirty hacks. To be
21  * verified. FIXME.
22  */
23
24 #define PERFECT_HASH_THRESHOLD  64      /* use perfect hash if not bigger */
25 #define DEFAULT_HASH_SIZE       64      /* optimized for diffserv */
26
27
28 struct tcindex_filter_result {
29         struct tcf_exts         exts;
30         struct tcf_result       res;
31         struct rcu_work         rwork;
32 };
33
34 struct tcindex_filter {
35         u16 key;
36         struct tcindex_filter_result result;
37         struct tcindex_filter __rcu *next;
38         struct rcu_work rwork;
39 };
40
41
42 struct tcindex_data {
43         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
44         struct tcindex_filter __rcu **h; /* imperfect hash; */
45         struct tcf_proto *tp;
46         u16 mask;               /* AND key with mask */
47         u32 shift;              /* shift ANDed key to the right */
48         u32 hash;               /* hash table size; 0 if undefined */
49         u32 alloc_hash;         /* allocated size */
50         u32 fall_through;       /* 0: only classify if explicit match */
51         struct rcu_work rwork;
52 };
53
54 static inline int tcindex_filter_is_set(struct tcindex_filter_result *r)
55 {
56         return tcf_exts_has_actions(&r->exts) || r->res.classid;
57 }
58
59 static struct tcindex_filter_result *tcindex_lookup(struct tcindex_data *p,
60                                                     u16 key)
61 {
62         if (p->perfect) {
63                 struct tcindex_filter_result *f = p->perfect + key;
64
65                 return tcindex_filter_is_set(f) ? f : NULL;
66         } else if (p->h) {
67                 struct tcindex_filter __rcu **fp;
68                 struct tcindex_filter *f;
69
70                 fp = &p->h[key % p->hash];
71                 for (f = rcu_dereference_bh_rtnl(*fp);
72                      f;
73                      fp = &f->next, f = rcu_dereference_bh_rtnl(*fp))
74                         if (f->key == key)
75                                 return &f->result;
76         }
77
78         return NULL;
79 }
80
81
82 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
83                             struct tcf_result *res)
84 {
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;
88
89         pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
90                  skb, tp, res, p);
91
92         f = tcindex_lookup(p, key);
93         if (!f) {
94                 struct Qdisc *q = tcf_block_q(tp->chain->block);
95
96                 if (!p->fall_through)
97                         return -1;
98                 res->classid = TC_H_MAKE(TC_H_MAJ(q->handle), key);
99                 res->class = 0;
100                 pr_debug("alg 0x%x\n", res->classid);
101                 return 0;
102         }
103         *res = f->res;
104         pr_debug("map 0x%x\n", res->classid);
105
106         return tcf_exts_exec(skb, &f->exts, res);
107 }
108
109
110 static void *tcindex_get(struct tcf_proto *tp, u32 handle)
111 {
112         struct tcindex_data *p = rtnl_dereference(tp->root);
113         struct tcindex_filter_result *r;
114
115         pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
116         if (p->perfect && handle >= p->alloc_hash)
117                 return NULL;
118         r = tcindex_lookup(p, handle);
119         return r && tcindex_filter_is_set(r) ? r : NULL;
120 }
121
122 static int tcindex_init(struct tcf_proto *tp)
123 {
124         struct tcindex_data *p;
125
126         pr_debug("tcindex_init(tp %p)\n", tp);
127         p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
128         if (!p)
129                 return -ENOMEM;
130
131         p->mask = 0xffff;
132         p->hash = DEFAULT_HASH_SIZE;
133         p->fall_through = 1;
134
135         rcu_assign_pointer(tp->root, p);
136         return 0;
137 }
138
139 static void __tcindex_destroy_rexts(struct tcindex_filter_result *r)
140 {
141         tcf_exts_destroy(&r->exts);
142         tcf_exts_put_net(&r->exts);
143 }
144
145 static void tcindex_destroy_rexts_work(struct work_struct *work)
146 {
147         struct tcindex_filter_result *r;
148
149         r = container_of(to_rcu_work(work),
150                          struct tcindex_filter_result,
151                          rwork);
152         rtnl_lock();
153         __tcindex_destroy_rexts(r);
154         rtnl_unlock();
155 }
156
157 static void __tcindex_destroy_fexts(struct tcindex_filter *f)
158 {
159         tcf_exts_destroy(&f->result.exts);
160         tcf_exts_put_net(&f->result.exts);
161         kfree(f);
162 }
163
164 static void tcindex_destroy_fexts_work(struct work_struct *work)
165 {
166         struct tcindex_filter *f = container_of(to_rcu_work(work),
167                                                 struct tcindex_filter,
168                                                 rwork);
169
170         rtnl_lock();
171         __tcindex_destroy_fexts(f);
172         rtnl_unlock();
173 }
174
175 static int tcindex_delete(struct tcf_proto *tp, void *arg, bool *last,
176                           struct netlink_ext_ack *extack)
177 {
178         struct tcindex_data *p = rtnl_dereference(tp->root);
179         struct tcindex_filter_result *r = arg;
180         struct tcindex_filter __rcu **walk;
181         struct tcindex_filter *f = NULL;
182
183         pr_debug("tcindex_delete(tp %p,arg %p),p %p\n", tp, arg, p);
184         if (p->perfect) {
185                 if (!r->res.class)
186                         return -ENOENT;
187         } else {
188                 int i;
189
190                 for (i = 0; i < p->hash; i++) {
191                         walk = p->h + i;
192                         for (f = rtnl_dereference(*walk); f;
193                              walk = &f->next, f = rtnl_dereference(*walk)) {
194                                 if (&f->result == r)
195                                         goto found;
196                         }
197                 }
198                 return -ENOENT;
199
200 found:
201                 rcu_assign_pointer(*walk, rtnl_dereference(f->next));
202         }
203         tcf_unbind_filter(tp, &r->res);
204         /* all classifiers are required to call tcf_exts_destroy() after rcu
205          * grace period, since converted-to-rcu actions are relying on that
206          * in cleanup() callback
207          */
208         if (f) {
209                 if (tcf_exts_get_net(&f->result.exts))
210                         tcf_queue_work(&f->rwork, tcindex_destroy_fexts_work);
211                 else
212                         __tcindex_destroy_fexts(f);
213         } else {
214                 if (tcf_exts_get_net(&r->exts))
215                         tcf_queue_work(&r->rwork, tcindex_destroy_rexts_work);
216                 else
217                         __tcindex_destroy_rexts(r);
218         }
219
220         *last = false;
221         return 0;
222 }
223
224 static void tcindex_destroy_work(struct work_struct *work)
225 {
226         struct tcindex_data *p = container_of(to_rcu_work(work),
227                                               struct tcindex_data,
228                                               rwork);
229
230         kfree(p->perfect);
231         kfree(p->h);
232         kfree(p);
233 }
234
235 static inline int
236 valid_perfect_hash(struct tcindex_data *p)
237 {
238         return  p->hash > (p->mask >> p->shift);
239 }
240
241 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
242         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
243         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
244         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
245         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
246         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
247 };
248
249 static int tcindex_filter_result_init(struct tcindex_filter_result *r)
250 {
251         memset(r, 0, sizeof(*r));
252         return tcf_exts_init(&r->exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
253 }
254
255 static void tcindex_partial_destroy_work(struct work_struct *work)
256 {
257         struct tcindex_data *p = container_of(to_rcu_work(work),
258                                               struct tcindex_data,
259                                               rwork);
260
261         kfree(p->perfect);
262         kfree(p);
263 }
264
265 static void tcindex_free_perfect_hash(struct tcindex_data *cp)
266 {
267         int i;
268
269         for (i = 0; i < cp->hash; i++)
270                 tcf_exts_destroy(&cp->perfect[i].exts);
271         kfree(cp->perfect);
272 }
273
274 static int tcindex_alloc_perfect_hash(struct net *net, struct tcindex_data *cp)
275 {
276         int i, err = 0;
277
278         cp->perfect = kcalloc(cp->hash, sizeof(struct tcindex_filter_result),
279                               GFP_KERNEL | __GFP_NOWARN);
280         if (!cp->perfect)
281                 return -ENOMEM;
282
283         for (i = 0; i < cp->hash; i++) {
284                 err = tcf_exts_init(&cp->perfect[i].exts,
285                                     TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
286                 if (err < 0)
287                         goto errout;
288 #ifdef CONFIG_NET_CLS_ACT
289                 cp->perfect[i].exts.net = net;
290 #endif
291         }
292
293         return 0;
294
295 errout:
296         tcindex_free_perfect_hash(cp);
297         return err;
298 }
299
300 static int
301 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
302                   u32 handle, struct tcindex_data *p,
303                   struct tcindex_filter_result *r, struct nlattr **tb,
304                   struct nlattr *est, bool ovr, struct netlink_ext_ack *extack)
305 {
306         struct tcindex_filter_result new_filter_result, *old_r = r;
307         struct tcindex_data *cp = NULL, *oldp;
308         struct tcindex_filter *f = NULL; /* make gcc behave */
309         struct tcf_result cr = {};
310         int err, balloc = 0;
311         struct tcf_exts e;
312
313         err = tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
314         if (err < 0)
315                 return err;
316         err = tcf_exts_validate(net, tp, tb, est, &e, ovr, extack);
317         if (err < 0)
318                 goto errout;
319
320         err = -ENOMEM;
321         /* tcindex_data attributes must look atomic to classifier/lookup so
322          * allocate new tcindex data and RCU assign it onto root. Keeping
323          * perfect hash and hash pointers from old data.
324          */
325         cp = kzalloc(sizeof(*cp), GFP_KERNEL);
326         if (!cp)
327                 goto errout;
328
329         cp->mask = p->mask;
330         cp->shift = p->shift;
331         cp->hash = p->hash;
332         cp->alloc_hash = p->alloc_hash;
333         cp->fall_through = p->fall_through;
334         cp->tp = tp;
335
336         if (tb[TCA_TCINDEX_HASH])
337                 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
338
339         if (tb[TCA_TCINDEX_MASK])
340                 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
341
342         if (tb[TCA_TCINDEX_SHIFT]) {
343                 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
344                 if (cp->shift > 16) {
345                         err = -EINVAL;
346                         goto errout;
347                 }
348         }
349         if (!cp->hash) {
350                 /* Hash not specified, use perfect hash if the upper limit
351                  * of the hashing index is below the threshold.
352                  */
353                 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
354                         cp->hash = (cp->mask >> cp->shift) + 1;
355                 else
356                         cp->hash = DEFAULT_HASH_SIZE;
357         }
358
359         if (p->perfect) {
360                 int i;
361
362                 if (tcindex_alloc_perfect_hash(net, cp) < 0)
363                         goto errout;
364                 cp->alloc_hash = cp->hash;
365                 for (i = 0; i < min(cp->hash, p->hash); i++)
366                         cp->perfect[i].res = p->perfect[i].res;
367                 balloc = 1;
368         }
369         cp->h = p->h;
370
371         err = tcindex_filter_result_init(&new_filter_result);
372         if (err < 0)
373                 goto errout_alloc;
374         if (old_r)
375                 cr = r->res;
376
377         err = -EBUSY;
378
379         /* Hash already allocated, make sure that we still meet the
380          * requirements for the allocated hash.
381          */
382         if (cp->perfect) {
383                 if (!valid_perfect_hash(cp) ||
384                     cp->hash > cp->alloc_hash)
385                         goto errout_alloc;
386         } else if (cp->h && cp->hash != cp->alloc_hash) {
387                 goto errout_alloc;
388         }
389
390         err = -EINVAL;
391         if (tb[TCA_TCINDEX_FALL_THROUGH])
392                 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
393
394         if (!cp->perfect && !cp->h)
395                 cp->alloc_hash = cp->hash;
396
397         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
398          * but then, we'd fail handles that may become valid after some future
399          * mask change. While this is extremely unlikely to ever matter,
400          * the check below is safer (and also more backwards-compatible).
401          */
402         if (cp->perfect || valid_perfect_hash(cp))
403                 if (handle >= cp->alloc_hash)
404                         goto errout_alloc;
405
406
407         err = -ENOMEM;
408         if (!cp->perfect && !cp->h) {
409                 if (valid_perfect_hash(cp)) {
410                         if (tcindex_alloc_perfect_hash(net, cp) < 0)
411                                 goto errout_alloc;
412                         balloc = 1;
413                 } else {
414                         struct tcindex_filter __rcu **hash;
415
416                         hash = kcalloc(cp->hash,
417                                        sizeof(struct tcindex_filter *),
418                                        GFP_KERNEL);
419
420                         if (!hash)
421                                 goto errout_alloc;
422
423                         cp->h = hash;
424                         balloc = 2;
425                 }
426         }
427
428         if (cp->perfect)
429                 r = cp->perfect + handle;
430         else
431                 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
432
433         if (r == &new_filter_result) {
434                 f = kzalloc(sizeof(*f), GFP_KERNEL);
435                 if (!f)
436                         goto errout_alloc;
437                 f->key = handle;
438                 f->next = NULL;
439                 err = tcindex_filter_result_init(&f->result);
440                 if (err < 0) {
441                         kfree(f);
442                         goto errout_alloc;
443                 }
444         }
445
446         if (tb[TCA_TCINDEX_CLASSID]) {
447                 cr.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
448                 tcf_bind_filter(tp, &cr, base);
449         }
450
451         if (old_r && old_r != r) {
452                 err = tcindex_filter_result_init(old_r);
453                 if (err < 0) {
454                         kfree(f);
455                         goto errout_alloc;
456                 }
457         }
458
459         oldp = p;
460         r->res = cr;
461         tcf_exts_change(&r->exts, &e);
462
463         rcu_assign_pointer(tp->root, cp);
464
465         if (r == &new_filter_result) {
466                 struct tcindex_filter *nfp;
467                 struct tcindex_filter __rcu **fp;
468
469                 f->result.res = r->res;
470                 tcf_exts_change(&f->result.exts, &r->exts);
471
472                 fp = cp->h + (handle % cp->hash);
473                 for (nfp = rtnl_dereference(*fp);
474                      nfp;
475                      fp = &nfp->next, nfp = rtnl_dereference(*fp))
476                                 ; /* nothing */
477
478                 rcu_assign_pointer(*fp, f);
479         } else {
480                 tcf_exts_destroy(&new_filter_result.exts);
481         }
482
483         if (oldp)
484                 tcf_queue_work(&oldp->rwork, tcindex_partial_destroy_work);
485         return 0;
486
487 errout_alloc:
488         if (balloc == 1)
489                 tcindex_free_perfect_hash(cp);
490         else if (balloc == 2)
491                 kfree(cp->h);
492         tcf_exts_destroy(&new_filter_result.exts);
493 errout:
494         kfree(cp);
495         tcf_exts_destroy(&e);
496         return err;
497 }
498
499 static int
500 tcindex_change(struct net *net, struct sk_buff *in_skb,
501                struct tcf_proto *tp, unsigned long base, u32 handle,
502                struct nlattr **tca, void **arg, bool ovr,
503                struct netlink_ext_ack *extack)
504 {
505         struct nlattr *opt = tca[TCA_OPTIONS];
506         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
507         struct tcindex_data *p = rtnl_dereference(tp->root);
508         struct tcindex_filter_result *r = *arg;
509         int err;
510
511         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
512             "p %p,r %p,*arg %p\n",
513             tp, handle, tca, arg, opt, p, r, arg ? *arg : NULL);
514
515         if (!opt)
516                 return 0;
517
518         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy, NULL);
519         if (err < 0)
520                 return err;
521
522         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
523                                  tca[TCA_RATE], ovr, extack);
524 }
525
526 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
527 {
528         struct tcindex_data *p = rtnl_dereference(tp->root);
529         struct tcindex_filter *f, *next;
530         int i;
531
532         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
533         if (p->perfect) {
534                 for (i = 0; i < p->hash; i++) {
535                         if (!p->perfect[i].res.class)
536                                 continue;
537                         if (walker->count >= walker->skip) {
538                                 if (walker->fn(tp, p->perfect + i, walker) < 0) {
539                                         walker->stop = 1;
540                                         return;
541                                 }
542                         }
543                         walker->count++;
544                 }
545         }
546         if (!p->h)
547                 return;
548         for (i = 0; i < p->hash; i++) {
549                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
550                         next = rtnl_dereference(f->next);
551                         if (walker->count >= walker->skip) {
552                                 if (walker->fn(tp, &f->result, walker) < 0) {
553                                         walker->stop = 1;
554                                         return;
555                                 }
556                         }
557                         walker->count++;
558                 }
559         }
560 }
561
562 static void tcindex_destroy(struct tcf_proto *tp,
563                             struct netlink_ext_ack *extack)
564 {
565         struct tcindex_data *p = rtnl_dereference(tp->root);
566         int i;
567
568         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
569
570         if (p->perfect) {
571                 for (i = 0; i < p->hash; i++) {
572                         struct tcindex_filter_result *r = p->perfect + i;
573
574                         tcf_unbind_filter(tp, &r->res);
575                         if (tcf_exts_get_net(&r->exts))
576                                 tcf_queue_work(&r->rwork,
577                                                tcindex_destroy_rexts_work);
578                         else
579                                 __tcindex_destroy_rexts(r);
580                 }
581         }
582
583         for (i = 0; p->h && i < p->hash; i++) {
584                 struct tcindex_filter *f, *next;
585                 bool last;
586
587                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
588                         next = rtnl_dereference(f->next);
589                         tcindex_delete(tp, &f->result, &last, NULL);
590                 }
591         }
592
593         tcf_queue_work(&p->rwork, tcindex_destroy_work);
594 }
595
596
597 static int tcindex_dump(struct net *net, struct tcf_proto *tp, void *fh,
598                         struct sk_buff *skb, struct tcmsg *t)
599 {
600         struct tcindex_data *p = rtnl_dereference(tp->root);
601         struct tcindex_filter_result *r = fh;
602         struct nlattr *nest;
603
604         pr_debug("tcindex_dump(tp %p,fh %p,skb %p,t %p),p %p,r %p\n",
605                  tp, fh, skb, t, p, r);
606         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
607
608         nest = nla_nest_start(skb, TCA_OPTIONS);
609         if (nest == NULL)
610                 goto nla_put_failure;
611
612         if (!fh) {
613                 t->tcm_handle = ~0; /* whatever ... */
614                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
615                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
616                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
617                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
618                         goto nla_put_failure;
619                 nla_nest_end(skb, nest);
620         } else {
621                 if (p->perfect) {
622                         t->tcm_handle = r - p->perfect;
623                 } else {
624                         struct tcindex_filter *f;
625                         struct tcindex_filter __rcu **fp;
626                         int i;
627
628                         t->tcm_handle = 0;
629                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
630                                 fp = &p->h[i];
631                                 for (f = rtnl_dereference(*fp);
632                                      !t->tcm_handle && f;
633                                      fp = &f->next, f = rtnl_dereference(*fp)) {
634                                         if (&f->result == r)
635                                                 t->tcm_handle = f->key;
636                                 }
637                         }
638                 }
639                 pr_debug("handle = %d\n", t->tcm_handle);
640                 if (r->res.class &&
641                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
642                         goto nla_put_failure;
643
644                 if (tcf_exts_dump(skb, &r->exts) < 0)
645                         goto nla_put_failure;
646                 nla_nest_end(skb, nest);
647
648                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
649                         goto nla_put_failure;
650         }
651
652         return skb->len;
653
654 nla_put_failure:
655         nla_nest_cancel(skb, nest);
656         return -1;
657 }
658
659 static void tcindex_bind_class(void *fh, u32 classid, unsigned long cl,
660                                void *q, unsigned long base)
661 {
662         struct tcindex_filter_result *r = fh;
663
664         if (r && r->res.classid == classid) {
665                 if (cl)
666                         __tcf_bind_filter(q, &r->res, base);
667                 else
668                         __tcf_unbind_filter(q, &r->res);
669         }
670 }
671
672 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
673         .kind           =       "tcindex",
674         .classify       =       tcindex_classify,
675         .init           =       tcindex_init,
676         .destroy        =       tcindex_destroy,
677         .get            =       tcindex_get,
678         .change         =       tcindex_change,
679         .delete         =       tcindex_delete,
680         .walk           =       tcindex_walk,
681         .dump           =       tcindex_dump,
682         .bind_class     =       tcindex_bind_class,
683         .owner          =       THIS_MODULE,
684 };
685
686 static int __init init_tcindex(void)
687 {
688         return register_tcf_proto_ops(&cls_tcindex_ops);
689 }
690
691 static void __exit exit_tcindex(void)
692 {
693         unregister_tcf_proto_ops(&cls_tcindex_ops);
694 }
695
696 module_init(init_tcindex)
697 module_exit(exit_tcindex)
698 MODULE_LICENSE("GPL");