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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_choke.c CHOKE scheduler
4 *
5 * Copyright (c) 2011 Stephen Hemminger <shemminger@vyatta.com>
6 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
7 */
8
9#include <linux/module.h>
10#include <linux/types.h>
11#include <linux/kernel.h>
12#include <linux/skbuff.h>
13#include <linux/vmalloc.h>
14#include <net/pkt_sched.h>
15#include <net/pkt_cls.h>
16#include <net/inet_ecn.h>
17#include <net/red.h>
18#include <net/flow_dissector.h>
19
20/*
21 CHOKe stateless AQM for fair bandwidth allocation
22 =================================================
23
24 CHOKe (CHOose and Keep for responsive flows, CHOose and Kill for
25 unresponsive flows) is a variant of RED that penalizes misbehaving flows but
26 maintains no flow state. The difference from RED is an additional step
27 during the enqueuing process. If average queue size is over the
28 low threshold (qmin), a packet is chosen at random from the queue.
29 If both the new and chosen packet are from the same flow, both
30 are dropped. Unlike RED, CHOKe is not really a "classful" qdisc because it
31 needs to access packets in queue randomly. It has a minimal class
32 interface to allow overriding the builtin flow classifier with
33 filters.
34
35 Source:
36 R. Pan, B. Prabhakar, and K. Psounis, "CHOKe, A Stateless
37 Active Queue Management Scheme for Approximating Fair Bandwidth Allocation",
38 IEEE INFOCOM, 2000.
39
40 A. Tang, J. Wang, S. Low, "Understanding CHOKe: Throughput and Spatial
41 Characteristics", IEEE/ACM Transactions on Networking, 2004
42
43 */
44
45/* Upper bound on size of sk_buff table (packets) */
46#define CHOKE_MAX_QUEUE (128*1024 - 1)
47
48struct choke_sched_data {
49/* Parameters */
50 u32 limit;
51 unsigned char flags;
52
53 struct red_parms parms;
54
55/* Variables */
56 struct red_vars vars;
57 struct {
58 u32 prob_drop; /* Early probability drops */
59 u32 prob_mark; /* Early probability marks */
60 u32 forced_drop; /* Forced drops, qavg > max_thresh */
61 u32 forced_mark; /* Forced marks, qavg > max_thresh */
62 u32 pdrop; /* Drops due to queue limits */
63 u32 matched; /* Drops to flow match */
64 } stats;
65
66 unsigned int head;
67 unsigned int tail;
68
69 unsigned int tab_mask; /* size - 1 */
70
71 struct sk_buff **tab;
72};
73
74/* number of elements in queue including holes */
75static unsigned int choke_len(const struct choke_sched_data *q)
76{
77 return (q->tail - q->head) & q->tab_mask;
78}
79
80/* Is ECN parameter configured */
81static int use_ecn(const struct choke_sched_data *q)
82{
83 return q->flags & TC_RED_ECN;
84}
85
86/* Should packets over max just be dropped (versus marked) */
87static int use_harddrop(const struct choke_sched_data *q)
88{
89 return q->flags & TC_RED_HARDDROP;
90}
91
92/* Move head pointer forward to skip over holes */
93static void choke_zap_head_holes(struct choke_sched_data *q)
94{
95 do {
96 q->head = (q->head + 1) & q->tab_mask;
97 if (q->head == q->tail)
98 break;
99 } while (q->tab[q->head] == NULL);
100}
101
102/* Move tail pointer backwards to reuse holes */
103static void choke_zap_tail_holes(struct choke_sched_data *q)
104{
105 do {
106 q->tail = (q->tail - 1) & q->tab_mask;
107 if (q->head == q->tail)
108 break;
109 } while (q->tab[q->tail] == NULL);
110}
111
112/* Drop packet from queue array by creating a "hole" */
113static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx,
114 struct sk_buff **to_free)
115{
116 struct choke_sched_data *q = qdisc_priv(sch);
117 struct sk_buff *skb = q->tab[idx];
118
119 q->tab[idx] = NULL;
120
121 if (idx == q->head)
122 choke_zap_head_holes(q);
123 if (idx == q->tail)
124 choke_zap_tail_holes(q);
125
126 qdisc_qstats_backlog_dec(sch, skb);
127 qdisc_tree_reduce_backlog(sch, 1, qdisc_pkt_len(skb));
128 qdisc_drop(skb, sch, to_free);
129 --sch->q.qlen;
130}
131
132struct choke_skb_cb {
133 u8 keys_valid;
134 struct flow_keys_digest keys;
135};
136
137static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
138{
139 qdisc_cb_private_validate(skb, sizeof(struct choke_skb_cb));
140 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
141}
142
143/*
144 * Compare flow of two packets
145 * Returns true only if source and destination address and port match.
146 * false for special cases
147 */
148static bool choke_match_flow(struct sk_buff *skb1,
149 struct sk_buff *skb2)
150{
151 struct flow_keys temp;
152
153 if (skb1->protocol != skb2->protocol)
154 return false;
155
156 if (!choke_skb_cb(skb1)->keys_valid) {
157 choke_skb_cb(skb1)->keys_valid = 1;
158 skb_flow_dissect_flow_keys(skb1, &temp, 0);
159 make_flow_keys_digest(&choke_skb_cb(skb1)->keys, &temp);
160 }
161
162 if (!choke_skb_cb(skb2)->keys_valid) {
163 choke_skb_cb(skb2)->keys_valid = 1;
164 skb_flow_dissect_flow_keys(skb2, &temp, 0);
165 make_flow_keys_digest(&choke_skb_cb(skb2)->keys, &temp);
166 }
167
168 return !memcmp(&choke_skb_cb(skb1)->keys,
169 &choke_skb_cb(skb2)->keys,
170 sizeof(choke_skb_cb(skb1)->keys));
171}
172
173/*
174 * Select a packet at random from queue
175 * HACK: since queue can have holes from previous deletion; retry several
176 * times to find a random skb but then just give up and return the head
177 * Will return NULL if queue is empty (q->head == q->tail)
178 */
179static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
180 unsigned int *pidx)
181{
182 struct sk_buff *skb;
183 int retrys = 3;
184
185 do {
186 *pidx = (q->head + get_random_u32_below(choke_len(q))) & q->tab_mask;
187 skb = q->tab[*pidx];
188 if (skb)
189 return skb;
190 } while (--retrys > 0);
191
192 return q->tab[*pidx = q->head];
193}
194
195/*
196 * Compare new packet with random packet in queue
197 * returns true if matched and sets *pidx
198 */
199static bool choke_match_random(const struct choke_sched_data *q,
200 struct sk_buff *nskb,
201 unsigned int *pidx)
202{
203 struct sk_buff *oskb;
204
205 if (q->head == q->tail)
206 return false;
207
208 oskb = choke_peek_random(q, pidx);
209 return choke_match_flow(oskb, nskb);
210}
211
212static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch,
213 struct sk_buff **to_free)
214{
215 struct choke_sched_data *q = qdisc_priv(sch);
216 const struct red_parms *p = &q->parms;
217
218 choke_skb_cb(skb)->keys_valid = 0;
219 /* Compute average queue usage (see RED) */
220 q->vars.qavg = red_calc_qavg(p, &q->vars, sch->q.qlen);
221 if (red_is_idling(&q->vars))
222 red_end_of_idle_period(&q->vars);
223
224 /* Is queue small? */
225 if (q->vars.qavg <= p->qth_min)
226 q->vars.qcount = -1;
227 else {
228 unsigned int idx;
229
230 /* Draw a packet at random from queue and compare flow */
231 if (choke_match_random(q, skb, &idx)) {
232 q->stats.matched++;
233 choke_drop_by_idx(sch, idx, to_free);
234 goto congestion_drop;
235 }
236
237 /* Queue is large, always mark/drop */
238 if (q->vars.qavg > p->qth_max) {
239 q->vars.qcount = -1;
240
241 qdisc_qstats_overlimit(sch);
242 if (use_harddrop(q) || !use_ecn(q) ||
243 !INET_ECN_set_ce(skb)) {
244 q->stats.forced_drop++;
245 goto congestion_drop;
246 }
247
248 q->stats.forced_mark++;
249 } else if (++q->vars.qcount) {
250 if (red_mark_probability(p, &q->vars, q->vars.qavg)) {
251 q->vars.qcount = 0;
252 q->vars.qR = red_random(p);
253
254 qdisc_qstats_overlimit(sch);
255 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
256 q->stats.prob_drop++;
257 goto congestion_drop;
258 }
259
260 q->stats.prob_mark++;
261 }
262 } else
263 q->vars.qR = red_random(p);
264 }
265
266 /* Admit new packet */
267 if (sch->q.qlen < q->limit) {
268 q->tab[q->tail] = skb;
269 q->tail = (q->tail + 1) & q->tab_mask;
270 ++sch->q.qlen;
271 qdisc_qstats_backlog_inc(sch, skb);
272 return NET_XMIT_SUCCESS;
273 }
274
275 q->stats.pdrop++;
276 return qdisc_drop(skb, sch, to_free);
277
278congestion_drop:
279 qdisc_drop(skb, sch, to_free);
280 return NET_XMIT_CN;
281}
282
283static struct sk_buff *choke_dequeue(struct Qdisc *sch)
284{
285 struct choke_sched_data *q = qdisc_priv(sch);
286 struct sk_buff *skb;
287
288 if (q->head == q->tail) {
289 if (!red_is_idling(&q->vars))
290 red_start_of_idle_period(&q->vars);
291 return NULL;
292 }
293
294 skb = q->tab[q->head];
295 q->tab[q->head] = NULL;
296 choke_zap_head_holes(q);
297 --sch->q.qlen;
298 qdisc_qstats_backlog_dec(sch, skb);
299 qdisc_bstats_update(sch, skb);
300
301 return skb;
302}
303
304static void choke_reset(struct Qdisc *sch)
305{
306 struct choke_sched_data *q = qdisc_priv(sch);
307
308 while (q->head != q->tail) {
309 struct sk_buff *skb = q->tab[q->head];
310
311 q->head = (q->head + 1) & q->tab_mask;
312 if (!skb)
313 continue;
314 rtnl_qdisc_drop(skb, sch);
315 }
316
317 if (q->tab)
318 memset(q->tab, 0, (q->tab_mask + 1) * sizeof(struct sk_buff *));
319 q->head = q->tail = 0;
320 red_restart(&q->vars);
321}
322
323static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
324 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
325 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
326 [TCA_CHOKE_MAX_P] = { .type = NLA_U32 },
327};
328
329
330static void choke_free(void *addr)
331{
332 kvfree(addr);
333}
334
335static int choke_change(struct Qdisc *sch, struct nlattr *opt,
336 struct netlink_ext_ack *extack)
337{
338 struct choke_sched_data *q = qdisc_priv(sch);
339 struct nlattr *tb[TCA_CHOKE_MAX + 1];
340 const struct tc_red_qopt *ctl;
341 int err;
342 struct sk_buff **old = NULL;
343 unsigned int mask;
344 u32 max_P;
345 u8 *stab;
346
347 if (opt == NULL)
348 return -EINVAL;
349
350 err = nla_parse_nested_deprecated(tb, TCA_CHOKE_MAX, opt,
351 choke_policy, NULL);
352 if (err < 0)
353 return err;
354
355 if (tb[TCA_CHOKE_PARMS] == NULL ||
356 tb[TCA_CHOKE_STAB] == NULL)
357 return -EINVAL;
358
359 max_P = tb[TCA_CHOKE_MAX_P] ? nla_get_u32(tb[TCA_CHOKE_MAX_P]) : 0;
360
361 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
362 stab = nla_data(tb[TCA_CHOKE_STAB]);
363 if (!red_check_params(ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Scell_log, stab))
364 return -EINVAL;
365
366 if (ctl->limit > CHOKE_MAX_QUEUE)
367 return -EINVAL;
368
369 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
370 if (mask != q->tab_mask) {
371 struct sk_buff **ntab;
372
373 ntab = kvcalloc(mask + 1, sizeof(struct sk_buff *), GFP_KERNEL);
374 if (!ntab)
375 return -ENOMEM;
376
377 sch_tree_lock(sch);
378 old = q->tab;
379 if (old) {
380 unsigned int oqlen = sch->q.qlen, tail = 0;
381 unsigned dropped = 0;
382
383 while (q->head != q->tail) {
384 struct sk_buff *skb = q->tab[q->head];
385
386 q->head = (q->head + 1) & q->tab_mask;
387 if (!skb)
388 continue;
389 if (tail < mask) {
390 ntab[tail++] = skb;
391 continue;
392 }
393 dropped += qdisc_pkt_len(skb);
394 qdisc_qstats_backlog_dec(sch, skb);
395 --sch->q.qlen;
396 rtnl_qdisc_drop(skb, sch);
397 }
398 qdisc_tree_reduce_backlog(sch, oqlen - sch->q.qlen, dropped);
399 q->head = 0;
400 q->tail = tail;
401 }
402
403 q->tab_mask = mask;
404 q->tab = ntab;
405 } else
406 sch_tree_lock(sch);
407
408 q->flags = ctl->flags;
409 q->limit = ctl->limit;
410
411 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
412 ctl->Plog, ctl->Scell_log,
413 stab,
414 max_P);
415 red_set_vars(&q->vars);
416
417 if (q->head == q->tail)
418 red_end_of_idle_period(&q->vars);
419
420 sch_tree_unlock(sch);
421 choke_free(old);
422 return 0;
423}
424
425static int choke_init(struct Qdisc *sch, struct nlattr *opt,
426 struct netlink_ext_ack *extack)
427{
428 return choke_change(sch, opt, extack);
429}
430
431static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
432{
433 struct choke_sched_data *q = qdisc_priv(sch);
434 struct nlattr *opts = NULL;
435 struct tc_red_qopt opt = {
436 .limit = q->limit,
437 .flags = q->flags,
438 .qth_min = q->parms.qth_min >> q->parms.Wlog,
439 .qth_max = q->parms.qth_max >> q->parms.Wlog,
440 .Wlog = q->parms.Wlog,
441 .Plog = q->parms.Plog,
442 .Scell_log = q->parms.Scell_log,
443 };
444
445 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
446 if (opts == NULL)
447 goto nla_put_failure;
448
449 if (nla_put(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt) ||
450 nla_put_u32(skb, TCA_CHOKE_MAX_P, q->parms.max_P))
451 goto nla_put_failure;
452 return nla_nest_end(skb, opts);
453
454nla_put_failure:
455 nla_nest_cancel(skb, opts);
456 return -EMSGSIZE;
457}
458
459static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
460{
461 struct choke_sched_data *q = qdisc_priv(sch);
462 struct tc_choke_xstats st = {
463 .early = q->stats.prob_drop + q->stats.forced_drop,
464 .marked = q->stats.prob_mark + q->stats.forced_mark,
465 .pdrop = q->stats.pdrop,
466 .matched = q->stats.matched,
467 };
468
469 return gnet_stats_copy_app(d, &st, sizeof(st));
470}
471
472static void choke_destroy(struct Qdisc *sch)
473{
474 struct choke_sched_data *q = qdisc_priv(sch);
475
476 choke_free(q->tab);
477}
478
479static struct sk_buff *choke_peek_head(struct Qdisc *sch)
480{
481 struct choke_sched_data *q = qdisc_priv(sch);
482
483 return (q->head != q->tail) ? q->tab[q->head] : NULL;
484}
485
486static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
487 .id = "choke",
488 .priv_size = sizeof(struct choke_sched_data),
489
490 .enqueue = choke_enqueue,
491 .dequeue = choke_dequeue,
492 .peek = choke_peek_head,
493 .init = choke_init,
494 .destroy = choke_destroy,
495 .reset = choke_reset,
496 .change = choke_change,
497 .dump = choke_dump,
498 .dump_stats = choke_dump_stats,
499 .owner = THIS_MODULE,
500};
501
502static int __init choke_module_init(void)
503{
504 return register_qdisc(&choke_qdisc_ops);
505}
506
507static void __exit choke_module_exit(void)
508{
509 unregister_qdisc(&choke_qdisc_ops);
510}
511
512module_init(choke_module_init)
513module_exit(choke_module_exit)
514
515MODULE_LICENSE("GPL");