Loading...
1/*
2 * net/sched/sch_choke.c CHOKE scheduler
3 *
4 * Copyright (c) 2011 Stephen Hemminger <shemminger@vyatta.com>
5 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * version 2 as published by the Free Software Foundation.
10 *
11 */
12
13#include <linux/module.h>
14#include <linux/types.h>
15#include <linux/kernel.h>
16#include <linux/skbuff.h>
17#include <linux/reciprocal_div.h>
18#include <linux/vmalloc.h>
19#include <net/pkt_sched.h>
20#include <net/inet_ecn.h>
21#include <net/red.h>
22#include <linux/ip.h>
23#include <net/ip.h>
24#include <linux/ipv6.h>
25#include <net/ipv6.h>
26
27/*
28 CHOKe stateless AQM for fair bandwidth allocation
29 =================================================
30
31 CHOKe (CHOose and Keep for responsive flows, CHOose and Kill for
32 unresponsive flows) is a variant of RED that penalizes misbehaving flows but
33 maintains no flow state. The difference from RED is an additional step
34 during the enqueuing process. If average queue size is over the
35 low threshold (qmin), a packet is chosen at random from the queue.
36 If both the new and chosen packet are from the same flow, both
37 are dropped. Unlike RED, CHOKe is not really a "classful" qdisc because it
38 needs to access packets in queue randomly. It has a minimal class
39 interface to allow overriding the builtin flow classifier with
40 filters.
41
42 Source:
43 R. Pan, B. Prabhakar, and K. Psounis, "CHOKe, A Stateless
44 Active Queue Management Scheme for Approximating Fair Bandwidth Allocation",
45 IEEE INFOCOM, 2000.
46
47 A. Tang, J. Wang, S. Low, "Understanding CHOKe: Throughput and Spatial
48 Characteristics", IEEE/ACM Transactions on Networking, 2004
49
50 */
51
52/* Upper bound on size of sk_buff table (packets) */
53#define CHOKE_MAX_QUEUE (128*1024 - 1)
54
55struct choke_sched_data {
56/* Parameters */
57 u32 limit;
58 unsigned char flags;
59
60 struct red_parms parms;
61
62/* Variables */
63 struct tcf_proto *filter_list;
64 struct {
65 u32 prob_drop; /* Early probability drops */
66 u32 prob_mark; /* Early probability marks */
67 u32 forced_drop; /* Forced drops, qavg > max_thresh */
68 u32 forced_mark; /* Forced marks, qavg > max_thresh */
69 u32 pdrop; /* Drops due to queue limits */
70 u32 other; /* Drops due to drop() calls */
71 u32 matched; /* Drops to flow match */
72 } stats;
73
74 unsigned int head;
75 unsigned int tail;
76
77 unsigned int tab_mask; /* size - 1 */
78
79 struct sk_buff **tab;
80};
81
82/* deliver a random number between 0 and N - 1 */
83static u32 random_N(unsigned int N)
84{
85 return reciprocal_divide(random32(), N);
86}
87
88/* number of elements in queue including holes */
89static unsigned int choke_len(const struct choke_sched_data *q)
90{
91 return (q->tail - q->head) & q->tab_mask;
92}
93
94/* Is ECN parameter configured */
95static int use_ecn(const struct choke_sched_data *q)
96{
97 return q->flags & TC_RED_ECN;
98}
99
100/* Should packets over max just be dropped (versus marked) */
101static int use_harddrop(const struct choke_sched_data *q)
102{
103 return q->flags & TC_RED_HARDDROP;
104}
105
106/* Move head pointer forward to skip over holes */
107static void choke_zap_head_holes(struct choke_sched_data *q)
108{
109 do {
110 q->head = (q->head + 1) & q->tab_mask;
111 if (q->head == q->tail)
112 break;
113 } while (q->tab[q->head] == NULL);
114}
115
116/* Move tail pointer backwards to reuse holes */
117static void choke_zap_tail_holes(struct choke_sched_data *q)
118{
119 do {
120 q->tail = (q->tail - 1) & q->tab_mask;
121 if (q->head == q->tail)
122 break;
123 } while (q->tab[q->tail] == NULL);
124}
125
126/* Drop packet from queue array by creating a "hole" */
127static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx)
128{
129 struct choke_sched_data *q = qdisc_priv(sch);
130 struct sk_buff *skb = q->tab[idx];
131
132 q->tab[idx] = NULL;
133
134 if (idx == q->head)
135 choke_zap_head_holes(q);
136 if (idx == q->tail)
137 choke_zap_tail_holes(q);
138
139 sch->qstats.backlog -= qdisc_pkt_len(skb);
140 qdisc_drop(skb, sch);
141 qdisc_tree_decrease_qlen(sch, 1);
142 --sch->q.qlen;
143}
144
145/*
146 * Compare flow of two packets
147 * Returns true only if source and destination address and port match.
148 * false for special cases
149 */
150static bool choke_match_flow(struct sk_buff *skb1,
151 struct sk_buff *skb2)
152{
153 int off1, off2, poff;
154 const u32 *ports1, *ports2;
155 u8 ip_proto;
156 __u32 hash1;
157
158 if (skb1->protocol != skb2->protocol)
159 return false;
160
161 /* Use hash value as quick check
162 * Assumes that __skb_get_rxhash makes IP header and ports linear
163 */
164 hash1 = skb_get_rxhash(skb1);
165 if (!hash1 || hash1 != skb_get_rxhash(skb2))
166 return false;
167
168 /* Probably match, but be sure to avoid hash collisions */
169 off1 = skb_network_offset(skb1);
170 off2 = skb_network_offset(skb2);
171
172 switch (skb1->protocol) {
173 case __constant_htons(ETH_P_IP): {
174 const struct iphdr *ip1, *ip2;
175
176 ip1 = (const struct iphdr *) (skb1->data + off1);
177 ip2 = (const struct iphdr *) (skb2->data + off2);
178
179 ip_proto = ip1->protocol;
180 if (ip_proto != ip2->protocol ||
181 ip1->saddr != ip2->saddr || ip1->daddr != ip2->daddr)
182 return false;
183
184 if (ip_is_fragment(ip1) | ip_is_fragment(ip2))
185 ip_proto = 0;
186 off1 += ip1->ihl * 4;
187 off2 += ip2->ihl * 4;
188 break;
189 }
190
191 case __constant_htons(ETH_P_IPV6): {
192 const struct ipv6hdr *ip1, *ip2;
193
194 ip1 = (const struct ipv6hdr *) (skb1->data + off1);
195 ip2 = (const struct ipv6hdr *) (skb2->data + off2);
196
197 ip_proto = ip1->nexthdr;
198 if (ip_proto != ip2->nexthdr ||
199 ipv6_addr_cmp(&ip1->saddr, &ip2->saddr) ||
200 ipv6_addr_cmp(&ip1->daddr, &ip2->daddr))
201 return false;
202 off1 += 40;
203 off2 += 40;
204 }
205
206 default: /* Maybe compare MAC header here? */
207 return false;
208 }
209
210 poff = proto_ports_offset(ip_proto);
211 if (poff < 0)
212 return true;
213
214 off1 += poff;
215 off2 += poff;
216
217 ports1 = (__force u32 *)(skb1->data + off1);
218 ports2 = (__force u32 *)(skb2->data + off2);
219 return *ports1 == *ports2;
220}
221
222struct choke_skb_cb {
223 u16 classid;
224};
225
226static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
227{
228 BUILD_BUG_ON(sizeof(skb->cb) <
229 sizeof(struct qdisc_skb_cb) + sizeof(struct choke_skb_cb));
230 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
231}
232
233static inline void choke_set_classid(struct sk_buff *skb, u16 classid)
234{
235 choke_skb_cb(skb)->classid = classid;
236}
237
238static u16 choke_get_classid(const struct sk_buff *skb)
239{
240 return choke_skb_cb(skb)->classid;
241}
242
243/*
244 * Classify flow using either:
245 * 1. pre-existing classification result in skb
246 * 2. fast internal classification
247 * 3. use TC filter based classification
248 */
249static bool choke_classify(struct sk_buff *skb,
250 struct Qdisc *sch, int *qerr)
251
252{
253 struct choke_sched_data *q = qdisc_priv(sch);
254 struct tcf_result res;
255 int result;
256
257 result = tc_classify(skb, q->filter_list, &res);
258 if (result >= 0) {
259#ifdef CONFIG_NET_CLS_ACT
260 switch (result) {
261 case TC_ACT_STOLEN:
262 case TC_ACT_QUEUED:
263 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
264 case TC_ACT_SHOT:
265 return false;
266 }
267#endif
268 choke_set_classid(skb, TC_H_MIN(res.classid));
269 return true;
270 }
271
272 return false;
273}
274
275/*
276 * Select a packet at random from queue
277 * HACK: since queue can have holes from previous deletion; retry several
278 * times to find a random skb but then just give up and return the head
279 * Will return NULL if queue is empty (q->head == q->tail)
280 */
281static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
282 unsigned int *pidx)
283{
284 struct sk_buff *skb;
285 int retrys = 3;
286
287 do {
288 *pidx = (q->head + random_N(choke_len(q))) & q->tab_mask;
289 skb = q->tab[*pidx];
290 if (skb)
291 return skb;
292 } while (--retrys > 0);
293
294 return q->tab[*pidx = q->head];
295}
296
297/*
298 * Compare new packet with random packet in queue
299 * returns true if matched and sets *pidx
300 */
301static bool choke_match_random(const struct choke_sched_data *q,
302 struct sk_buff *nskb,
303 unsigned int *pidx)
304{
305 struct sk_buff *oskb;
306
307 if (q->head == q->tail)
308 return false;
309
310 oskb = choke_peek_random(q, pidx);
311 if (q->filter_list)
312 return choke_get_classid(nskb) == choke_get_classid(oskb);
313
314 return choke_match_flow(oskb, nskb);
315}
316
317static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch)
318{
319 struct choke_sched_data *q = qdisc_priv(sch);
320 struct red_parms *p = &q->parms;
321 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
322
323 if (q->filter_list) {
324 /* If using external classifiers, get result and record it. */
325 if (!choke_classify(skb, sch, &ret))
326 goto other_drop; /* Packet was eaten by filter */
327 }
328
329 /* Compute average queue usage (see RED) */
330 p->qavg = red_calc_qavg(p, sch->q.qlen);
331 if (red_is_idling(p))
332 red_end_of_idle_period(p);
333
334 /* Is queue small? */
335 if (p->qavg <= p->qth_min)
336 p->qcount = -1;
337 else {
338 unsigned int idx;
339
340 /* Draw a packet at random from queue and compare flow */
341 if (choke_match_random(q, skb, &idx)) {
342 q->stats.matched++;
343 choke_drop_by_idx(sch, idx);
344 goto congestion_drop;
345 }
346
347 /* Queue is large, always mark/drop */
348 if (p->qavg > p->qth_max) {
349 p->qcount = -1;
350
351 sch->qstats.overlimits++;
352 if (use_harddrop(q) || !use_ecn(q) ||
353 !INET_ECN_set_ce(skb)) {
354 q->stats.forced_drop++;
355 goto congestion_drop;
356 }
357
358 q->stats.forced_mark++;
359 } else if (++p->qcount) {
360 if (red_mark_probability(p, p->qavg)) {
361 p->qcount = 0;
362 p->qR = red_random(p);
363
364 sch->qstats.overlimits++;
365 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
366 q->stats.prob_drop++;
367 goto congestion_drop;
368 }
369
370 q->stats.prob_mark++;
371 }
372 } else
373 p->qR = red_random(p);
374 }
375
376 /* Admit new packet */
377 if (sch->q.qlen < q->limit) {
378 q->tab[q->tail] = skb;
379 q->tail = (q->tail + 1) & q->tab_mask;
380 ++sch->q.qlen;
381 sch->qstats.backlog += qdisc_pkt_len(skb);
382 return NET_XMIT_SUCCESS;
383 }
384
385 q->stats.pdrop++;
386 sch->qstats.drops++;
387 kfree_skb(skb);
388 return NET_XMIT_DROP;
389
390 congestion_drop:
391 qdisc_drop(skb, sch);
392 return NET_XMIT_CN;
393
394 other_drop:
395 if (ret & __NET_XMIT_BYPASS)
396 sch->qstats.drops++;
397 kfree_skb(skb);
398 return ret;
399}
400
401static struct sk_buff *choke_dequeue(struct Qdisc *sch)
402{
403 struct choke_sched_data *q = qdisc_priv(sch);
404 struct sk_buff *skb;
405
406 if (q->head == q->tail) {
407 if (!red_is_idling(&q->parms))
408 red_start_of_idle_period(&q->parms);
409 return NULL;
410 }
411
412 skb = q->tab[q->head];
413 q->tab[q->head] = NULL;
414 choke_zap_head_holes(q);
415 --sch->q.qlen;
416 sch->qstats.backlog -= qdisc_pkt_len(skb);
417 qdisc_bstats_update(sch, skb);
418
419 return skb;
420}
421
422static unsigned int choke_drop(struct Qdisc *sch)
423{
424 struct choke_sched_data *q = qdisc_priv(sch);
425 unsigned int len;
426
427 len = qdisc_queue_drop(sch);
428 if (len > 0)
429 q->stats.other++;
430 else {
431 if (!red_is_idling(&q->parms))
432 red_start_of_idle_period(&q->parms);
433 }
434
435 return len;
436}
437
438static void choke_reset(struct Qdisc *sch)
439{
440 struct choke_sched_data *q = qdisc_priv(sch);
441
442 red_restart(&q->parms);
443}
444
445static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
446 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
447 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
448};
449
450
451static void choke_free(void *addr)
452{
453 if (addr) {
454 if (is_vmalloc_addr(addr))
455 vfree(addr);
456 else
457 kfree(addr);
458 }
459}
460
461static int choke_change(struct Qdisc *sch, struct nlattr *opt)
462{
463 struct choke_sched_data *q = qdisc_priv(sch);
464 struct nlattr *tb[TCA_CHOKE_MAX + 1];
465 const struct tc_red_qopt *ctl;
466 int err;
467 struct sk_buff **old = NULL;
468 unsigned int mask;
469
470 if (opt == NULL)
471 return -EINVAL;
472
473 err = nla_parse_nested(tb, TCA_CHOKE_MAX, opt, choke_policy);
474 if (err < 0)
475 return err;
476
477 if (tb[TCA_CHOKE_PARMS] == NULL ||
478 tb[TCA_CHOKE_STAB] == NULL)
479 return -EINVAL;
480
481 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
482
483 if (ctl->limit > CHOKE_MAX_QUEUE)
484 return -EINVAL;
485
486 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
487 if (mask != q->tab_mask) {
488 struct sk_buff **ntab;
489
490 ntab = kcalloc(mask + 1, sizeof(struct sk_buff *), GFP_KERNEL);
491 if (!ntab)
492 ntab = vzalloc((mask + 1) * sizeof(struct sk_buff *));
493 if (!ntab)
494 return -ENOMEM;
495
496 sch_tree_lock(sch);
497 old = q->tab;
498 if (old) {
499 unsigned int oqlen = sch->q.qlen, tail = 0;
500
501 while (q->head != q->tail) {
502 struct sk_buff *skb = q->tab[q->head];
503
504 q->head = (q->head + 1) & q->tab_mask;
505 if (!skb)
506 continue;
507 if (tail < mask) {
508 ntab[tail++] = skb;
509 continue;
510 }
511 sch->qstats.backlog -= qdisc_pkt_len(skb);
512 --sch->q.qlen;
513 qdisc_drop(skb, sch);
514 }
515 qdisc_tree_decrease_qlen(sch, oqlen - sch->q.qlen);
516 q->head = 0;
517 q->tail = tail;
518 }
519
520 q->tab_mask = mask;
521 q->tab = ntab;
522 } else
523 sch_tree_lock(sch);
524
525 q->flags = ctl->flags;
526 q->limit = ctl->limit;
527
528 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
529 ctl->Plog, ctl->Scell_log,
530 nla_data(tb[TCA_CHOKE_STAB]));
531
532 if (q->head == q->tail)
533 red_end_of_idle_period(&q->parms);
534
535 sch_tree_unlock(sch);
536 choke_free(old);
537 return 0;
538}
539
540static int choke_init(struct Qdisc *sch, struct nlattr *opt)
541{
542 return choke_change(sch, opt);
543}
544
545static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
546{
547 struct choke_sched_data *q = qdisc_priv(sch);
548 struct nlattr *opts = NULL;
549 struct tc_red_qopt opt = {
550 .limit = q->limit,
551 .flags = q->flags,
552 .qth_min = q->parms.qth_min >> q->parms.Wlog,
553 .qth_max = q->parms.qth_max >> q->parms.Wlog,
554 .Wlog = q->parms.Wlog,
555 .Plog = q->parms.Plog,
556 .Scell_log = q->parms.Scell_log,
557 };
558
559 opts = nla_nest_start(skb, TCA_OPTIONS);
560 if (opts == NULL)
561 goto nla_put_failure;
562
563 NLA_PUT(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt);
564 return nla_nest_end(skb, opts);
565
566nla_put_failure:
567 nla_nest_cancel(skb, opts);
568 return -EMSGSIZE;
569}
570
571static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
572{
573 struct choke_sched_data *q = qdisc_priv(sch);
574 struct tc_choke_xstats st = {
575 .early = q->stats.prob_drop + q->stats.forced_drop,
576 .marked = q->stats.prob_mark + q->stats.forced_mark,
577 .pdrop = q->stats.pdrop,
578 .other = q->stats.other,
579 .matched = q->stats.matched,
580 };
581
582 return gnet_stats_copy_app(d, &st, sizeof(st));
583}
584
585static void choke_destroy(struct Qdisc *sch)
586{
587 struct choke_sched_data *q = qdisc_priv(sch);
588
589 tcf_destroy_chain(&q->filter_list);
590 choke_free(q->tab);
591}
592
593static struct Qdisc *choke_leaf(struct Qdisc *sch, unsigned long arg)
594{
595 return NULL;
596}
597
598static unsigned long choke_get(struct Qdisc *sch, u32 classid)
599{
600 return 0;
601}
602
603static void choke_put(struct Qdisc *q, unsigned long cl)
604{
605}
606
607static unsigned long choke_bind(struct Qdisc *sch, unsigned long parent,
608 u32 classid)
609{
610 return 0;
611}
612
613static struct tcf_proto **choke_find_tcf(struct Qdisc *sch, unsigned long cl)
614{
615 struct choke_sched_data *q = qdisc_priv(sch);
616
617 if (cl)
618 return NULL;
619 return &q->filter_list;
620}
621
622static int choke_dump_class(struct Qdisc *sch, unsigned long cl,
623 struct sk_buff *skb, struct tcmsg *tcm)
624{
625 tcm->tcm_handle |= TC_H_MIN(cl);
626 return 0;
627}
628
629static void choke_walk(struct Qdisc *sch, struct qdisc_walker *arg)
630{
631 if (!arg->stop) {
632 if (arg->fn(sch, 1, arg) < 0) {
633 arg->stop = 1;
634 return;
635 }
636 arg->count++;
637 }
638}
639
640static const struct Qdisc_class_ops choke_class_ops = {
641 .leaf = choke_leaf,
642 .get = choke_get,
643 .put = choke_put,
644 .tcf_chain = choke_find_tcf,
645 .bind_tcf = choke_bind,
646 .unbind_tcf = choke_put,
647 .dump = choke_dump_class,
648 .walk = choke_walk,
649};
650
651static struct sk_buff *choke_peek_head(struct Qdisc *sch)
652{
653 struct choke_sched_data *q = qdisc_priv(sch);
654
655 return (q->head != q->tail) ? q->tab[q->head] : NULL;
656}
657
658static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
659 .id = "choke",
660 .priv_size = sizeof(struct choke_sched_data),
661
662 .enqueue = choke_enqueue,
663 .dequeue = choke_dequeue,
664 .peek = choke_peek_head,
665 .drop = choke_drop,
666 .init = choke_init,
667 .destroy = choke_destroy,
668 .reset = choke_reset,
669 .change = choke_change,
670 .dump = choke_dump,
671 .dump_stats = choke_dump_stats,
672 .owner = THIS_MODULE,
673};
674
675static int __init choke_module_init(void)
676{
677 return register_qdisc(&choke_qdisc_ops);
678}
679
680static void __exit choke_module_exit(void)
681{
682 unregister_qdisc(&choke_qdisc_ops);
683}
684
685module_init(choke_module_init)
686module_exit(choke_module_exit)
687
688MODULE_LICENSE("GPL");
1/*
2 * net/sched/sch_choke.c CHOKE scheduler
3 *
4 * Copyright (c) 2011 Stephen Hemminger <shemminger@vyatta.com>
5 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * version 2 as published by the Free Software Foundation.
10 *
11 */
12
13#include <linux/module.h>
14#include <linux/types.h>
15#include <linux/kernel.h>
16#include <linux/skbuff.h>
17#include <linux/vmalloc.h>
18#include <net/pkt_sched.h>
19#include <net/pkt_cls.h>
20#include <net/inet_ecn.h>
21#include <net/red.h>
22#include <net/flow_dissector.h>
23
24/*
25 CHOKe stateless AQM for fair bandwidth allocation
26 =================================================
27
28 CHOKe (CHOose and Keep for responsive flows, CHOose and Kill for
29 unresponsive flows) is a variant of RED that penalizes misbehaving flows but
30 maintains no flow state. The difference from RED is an additional step
31 during the enqueuing process. If average queue size is over the
32 low threshold (qmin), a packet is chosen at random from the queue.
33 If both the new and chosen packet are from the same flow, both
34 are dropped. Unlike RED, CHOKe is not really a "classful" qdisc because it
35 needs to access packets in queue randomly. It has a minimal class
36 interface to allow overriding the builtin flow classifier with
37 filters.
38
39 Source:
40 R. Pan, B. Prabhakar, and K. Psounis, "CHOKe, A Stateless
41 Active Queue Management Scheme for Approximating Fair Bandwidth Allocation",
42 IEEE INFOCOM, 2000.
43
44 A. Tang, J. Wang, S. Low, "Understanding CHOKe: Throughput and Spatial
45 Characteristics", IEEE/ACM Transactions on Networking, 2004
46
47 */
48
49/* Upper bound on size of sk_buff table (packets) */
50#define CHOKE_MAX_QUEUE (128*1024 - 1)
51
52struct choke_sched_data {
53/* Parameters */
54 u32 limit;
55 unsigned char flags;
56
57 struct red_parms parms;
58
59/* Variables */
60 struct red_vars vars;
61 struct {
62 u32 prob_drop; /* Early probability drops */
63 u32 prob_mark; /* Early probability marks */
64 u32 forced_drop; /* Forced drops, qavg > max_thresh */
65 u32 forced_mark; /* Forced marks, qavg > max_thresh */
66 u32 pdrop; /* Drops due to queue limits */
67 u32 other; /* Drops due to drop() calls */
68 u32 matched; /* Drops to flow match */
69 } stats;
70
71 unsigned int head;
72 unsigned int tail;
73
74 unsigned int tab_mask; /* size - 1 */
75
76 struct sk_buff **tab;
77};
78
79/* number of elements in queue including holes */
80static unsigned int choke_len(const struct choke_sched_data *q)
81{
82 return (q->tail - q->head) & q->tab_mask;
83}
84
85/* Is ECN parameter configured */
86static int use_ecn(const struct choke_sched_data *q)
87{
88 return q->flags & TC_RED_ECN;
89}
90
91/* Should packets over max just be dropped (versus marked) */
92static int use_harddrop(const struct choke_sched_data *q)
93{
94 return q->flags & TC_RED_HARDDROP;
95}
96
97/* Move head pointer forward to skip over holes */
98static void choke_zap_head_holes(struct choke_sched_data *q)
99{
100 do {
101 q->head = (q->head + 1) & q->tab_mask;
102 if (q->head == q->tail)
103 break;
104 } while (q->tab[q->head] == NULL);
105}
106
107/* Move tail pointer backwards to reuse holes */
108static void choke_zap_tail_holes(struct choke_sched_data *q)
109{
110 do {
111 q->tail = (q->tail - 1) & q->tab_mask;
112 if (q->head == q->tail)
113 break;
114 } while (q->tab[q->tail] == NULL);
115}
116
117/* Drop packet from queue array by creating a "hole" */
118static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx,
119 struct sk_buff **to_free)
120{
121 struct choke_sched_data *q = qdisc_priv(sch);
122 struct sk_buff *skb = q->tab[idx];
123
124 q->tab[idx] = NULL;
125
126 if (idx == q->head)
127 choke_zap_head_holes(q);
128 if (idx == q->tail)
129 choke_zap_tail_holes(q);
130
131 qdisc_qstats_backlog_dec(sch, skb);
132 qdisc_tree_reduce_backlog(sch, 1, qdisc_pkt_len(skb));
133 qdisc_drop(skb, sch, to_free);
134 --sch->q.qlen;
135}
136
137struct choke_skb_cb {
138 u16 classid;
139 u8 keys_valid;
140 struct flow_keys_digest keys;
141};
142
143static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
144{
145 qdisc_cb_private_validate(skb, sizeof(struct choke_skb_cb));
146 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
147}
148
149static inline void choke_set_classid(struct sk_buff *skb, u16 classid)
150{
151 choke_skb_cb(skb)->classid = classid;
152}
153
154/*
155 * Compare flow of two packets
156 * Returns true only if source and destination address and port match.
157 * false for special cases
158 */
159static bool choke_match_flow(struct sk_buff *skb1,
160 struct sk_buff *skb2)
161{
162 struct flow_keys temp;
163
164 if (skb1->protocol != skb2->protocol)
165 return false;
166
167 if (!choke_skb_cb(skb1)->keys_valid) {
168 choke_skb_cb(skb1)->keys_valid = 1;
169 skb_flow_dissect_flow_keys(skb1, &temp, 0);
170 make_flow_keys_digest(&choke_skb_cb(skb1)->keys, &temp);
171 }
172
173 if (!choke_skb_cb(skb2)->keys_valid) {
174 choke_skb_cb(skb2)->keys_valid = 1;
175 skb_flow_dissect_flow_keys(skb2, &temp, 0);
176 make_flow_keys_digest(&choke_skb_cb(skb2)->keys, &temp);
177 }
178
179 return !memcmp(&choke_skb_cb(skb1)->keys,
180 &choke_skb_cb(skb2)->keys,
181 sizeof(choke_skb_cb(skb1)->keys));
182}
183
184/*
185 * Select a packet at random from queue
186 * HACK: since queue can have holes from previous deletion; retry several
187 * times to find a random skb but then just give up and return the head
188 * Will return NULL if queue is empty (q->head == q->tail)
189 */
190static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
191 unsigned int *pidx)
192{
193 struct sk_buff *skb;
194 int retrys = 3;
195
196 do {
197 *pidx = (q->head + prandom_u32_max(choke_len(q))) & q->tab_mask;
198 skb = q->tab[*pidx];
199 if (skb)
200 return skb;
201 } while (--retrys > 0);
202
203 return q->tab[*pidx = q->head];
204}
205
206/*
207 * Compare new packet with random packet in queue
208 * returns true if matched and sets *pidx
209 */
210static bool choke_match_random(const struct choke_sched_data *q,
211 struct sk_buff *nskb,
212 unsigned int *pidx)
213{
214 struct sk_buff *oskb;
215
216 if (q->head == q->tail)
217 return false;
218
219 oskb = choke_peek_random(q, pidx);
220 return choke_match_flow(oskb, nskb);
221}
222
223static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch,
224 struct sk_buff **to_free)
225{
226 struct choke_sched_data *q = qdisc_priv(sch);
227 const struct red_parms *p = &q->parms;
228
229 choke_skb_cb(skb)->keys_valid = 0;
230 /* Compute average queue usage (see RED) */
231 q->vars.qavg = red_calc_qavg(p, &q->vars, sch->q.qlen);
232 if (red_is_idling(&q->vars))
233 red_end_of_idle_period(&q->vars);
234
235 /* Is queue small? */
236 if (q->vars.qavg <= p->qth_min)
237 q->vars.qcount = -1;
238 else {
239 unsigned int idx;
240
241 /* Draw a packet at random from queue and compare flow */
242 if (choke_match_random(q, skb, &idx)) {
243 q->stats.matched++;
244 choke_drop_by_idx(sch, idx, to_free);
245 goto congestion_drop;
246 }
247
248 /* Queue is large, always mark/drop */
249 if (q->vars.qavg > p->qth_max) {
250 q->vars.qcount = -1;
251
252 qdisc_qstats_overlimit(sch);
253 if (use_harddrop(q) || !use_ecn(q) ||
254 !INET_ECN_set_ce(skb)) {
255 q->stats.forced_drop++;
256 goto congestion_drop;
257 }
258
259 q->stats.forced_mark++;
260 } else if (++q->vars.qcount) {
261 if (red_mark_probability(p, &q->vars, q->vars.qavg)) {
262 q->vars.qcount = 0;
263 q->vars.qR = red_random(p);
264
265 qdisc_qstats_overlimit(sch);
266 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
267 q->stats.prob_drop++;
268 goto congestion_drop;
269 }
270
271 q->stats.prob_mark++;
272 }
273 } else
274 q->vars.qR = red_random(p);
275 }
276
277 /* Admit new packet */
278 if (sch->q.qlen < q->limit) {
279 q->tab[q->tail] = skb;
280 q->tail = (q->tail + 1) & q->tab_mask;
281 ++sch->q.qlen;
282 qdisc_qstats_backlog_inc(sch, skb);
283 return NET_XMIT_SUCCESS;
284 }
285
286 q->stats.pdrop++;
287 return qdisc_drop(skb, sch, to_free);
288
289congestion_drop:
290 qdisc_drop(skb, sch, to_free);
291 return NET_XMIT_CN;
292}
293
294static struct sk_buff *choke_dequeue(struct Qdisc *sch)
295{
296 struct choke_sched_data *q = qdisc_priv(sch);
297 struct sk_buff *skb;
298
299 if (q->head == q->tail) {
300 if (!red_is_idling(&q->vars))
301 red_start_of_idle_period(&q->vars);
302 return NULL;
303 }
304
305 skb = q->tab[q->head];
306 q->tab[q->head] = NULL;
307 choke_zap_head_holes(q);
308 --sch->q.qlen;
309 qdisc_qstats_backlog_dec(sch, skb);
310 qdisc_bstats_update(sch, skb);
311
312 return skb;
313}
314
315static void choke_reset(struct Qdisc *sch)
316{
317 struct choke_sched_data *q = qdisc_priv(sch);
318
319 while (q->head != q->tail) {
320 struct sk_buff *skb = q->tab[q->head];
321
322 q->head = (q->head + 1) & q->tab_mask;
323 if (!skb)
324 continue;
325 rtnl_qdisc_drop(skb, sch);
326 }
327
328 sch->q.qlen = 0;
329 sch->qstats.backlog = 0;
330 memset(q->tab, 0, (q->tab_mask + 1) * sizeof(struct sk_buff *));
331 q->head = q->tail = 0;
332 red_restart(&q->vars);
333}
334
335static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
336 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
337 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
338 [TCA_CHOKE_MAX_P] = { .type = NLA_U32 },
339};
340
341
342static void choke_free(void *addr)
343{
344 kvfree(addr);
345}
346
347static int choke_change(struct Qdisc *sch, struct nlattr *opt,
348 struct netlink_ext_ack *extack)
349{
350 struct choke_sched_data *q = qdisc_priv(sch);
351 struct nlattr *tb[TCA_CHOKE_MAX + 1];
352 const struct tc_red_qopt *ctl;
353 int err;
354 struct sk_buff **old = NULL;
355 unsigned int mask;
356 u32 max_P;
357
358 if (opt == NULL)
359 return -EINVAL;
360
361 err = nla_parse_nested(tb, TCA_CHOKE_MAX, opt, choke_policy, NULL);
362 if (err < 0)
363 return err;
364
365 if (tb[TCA_CHOKE_PARMS] == NULL ||
366 tb[TCA_CHOKE_STAB] == NULL)
367 return -EINVAL;
368
369 max_P = tb[TCA_CHOKE_MAX_P] ? nla_get_u32(tb[TCA_CHOKE_MAX_P]) : 0;
370
371 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
372
373 if (!red_check_params(ctl->qth_min, ctl->qth_max, ctl->Wlog))
374 return -EINVAL;
375
376 if (ctl->limit > CHOKE_MAX_QUEUE)
377 return -EINVAL;
378
379 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
380 if (mask != q->tab_mask) {
381 struct sk_buff **ntab;
382
383 ntab = kvmalloc_array((mask + 1), sizeof(struct sk_buff *), GFP_KERNEL | __GFP_ZERO);
384 if (!ntab)
385 return -ENOMEM;
386
387 sch_tree_lock(sch);
388 old = q->tab;
389 if (old) {
390 unsigned int oqlen = sch->q.qlen, tail = 0;
391 unsigned dropped = 0;
392
393 while (q->head != q->tail) {
394 struct sk_buff *skb = q->tab[q->head];
395
396 q->head = (q->head + 1) & q->tab_mask;
397 if (!skb)
398 continue;
399 if (tail < mask) {
400 ntab[tail++] = skb;
401 continue;
402 }
403 dropped += qdisc_pkt_len(skb);
404 qdisc_qstats_backlog_dec(sch, skb);
405 --sch->q.qlen;
406 rtnl_qdisc_drop(skb, sch);
407 }
408 qdisc_tree_reduce_backlog(sch, oqlen - sch->q.qlen, dropped);
409 q->head = 0;
410 q->tail = tail;
411 }
412
413 q->tab_mask = mask;
414 q->tab = ntab;
415 } else
416 sch_tree_lock(sch);
417
418 q->flags = ctl->flags;
419 q->limit = ctl->limit;
420
421 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
422 ctl->Plog, ctl->Scell_log,
423 nla_data(tb[TCA_CHOKE_STAB]),
424 max_P);
425 red_set_vars(&q->vars);
426
427 if (q->head == q->tail)
428 red_end_of_idle_period(&q->vars);
429
430 sch_tree_unlock(sch);
431 choke_free(old);
432 return 0;
433}
434
435static int choke_init(struct Qdisc *sch, struct nlattr *opt,
436 struct netlink_ext_ack *extack)
437{
438 return choke_change(sch, opt, extack);
439}
440
441static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
442{
443 struct choke_sched_data *q = qdisc_priv(sch);
444 struct nlattr *opts = NULL;
445 struct tc_red_qopt opt = {
446 .limit = q->limit,
447 .flags = q->flags,
448 .qth_min = q->parms.qth_min >> q->parms.Wlog,
449 .qth_max = q->parms.qth_max >> q->parms.Wlog,
450 .Wlog = q->parms.Wlog,
451 .Plog = q->parms.Plog,
452 .Scell_log = q->parms.Scell_log,
453 };
454
455 opts = nla_nest_start(skb, TCA_OPTIONS);
456 if (opts == NULL)
457 goto nla_put_failure;
458
459 if (nla_put(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt) ||
460 nla_put_u32(skb, TCA_CHOKE_MAX_P, q->parms.max_P))
461 goto nla_put_failure;
462 return nla_nest_end(skb, opts);
463
464nla_put_failure:
465 nla_nest_cancel(skb, opts);
466 return -EMSGSIZE;
467}
468
469static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
470{
471 struct choke_sched_data *q = qdisc_priv(sch);
472 struct tc_choke_xstats st = {
473 .early = q->stats.prob_drop + q->stats.forced_drop,
474 .marked = q->stats.prob_mark + q->stats.forced_mark,
475 .pdrop = q->stats.pdrop,
476 .other = q->stats.other,
477 .matched = q->stats.matched,
478 };
479
480 return gnet_stats_copy_app(d, &st, sizeof(st));
481}
482
483static void choke_destroy(struct Qdisc *sch)
484{
485 struct choke_sched_data *q = qdisc_priv(sch);
486
487 choke_free(q->tab);
488}
489
490static struct sk_buff *choke_peek_head(struct Qdisc *sch)
491{
492 struct choke_sched_data *q = qdisc_priv(sch);
493
494 return (q->head != q->tail) ? q->tab[q->head] : NULL;
495}
496
497static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
498 .id = "choke",
499 .priv_size = sizeof(struct choke_sched_data),
500
501 .enqueue = choke_enqueue,
502 .dequeue = choke_dequeue,
503 .peek = choke_peek_head,
504 .init = choke_init,
505 .destroy = choke_destroy,
506 .reset = choke_reset,
507 .change = choke_change,
508 .dump = choke_dump,
509 .dump_stats = choke_dump_stats,
510 .owner = THIS_MODULE,
511};
512
513static int __init choke_module_init(void)
514{
515 return register_qdisc(&choke_qdisc_ops);
516}
517
518static void __exit choke_module_exit(void)
519{
520 unregister_qdisc(&choke_qdisc_ops);
521}
522
523module_init(choke_module_init)
524module_exit(choke_module_exit)
525
526MODULE_LICENSE("GPL");