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 <net/flow_keys.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 tcf_proto *filter_list;
62 struct {
63 u32 prob_drop; /* Early probability drops */
64 u32 prob_mark; /* Early probability marks */
65 u32 forced_drop; /* Forced drops, qavg > max_thresh */
66 u32 forced_mark; /* Forced marks, qavg > max_thresh */
67 u32 pdrop; /* Drops due to queue limits */
68 u32 other; /* Drops due to drop() calls */
69 u32 matched; /* Drops to flow match */
70 } stats;
71
72 unsigned int head;
73 unsigned int tail;
74
75 unsigned int tab_mask; /* size - 1 */
76
77 struct sk_buff **tab;
78};
79
80/* deliver a random number between 0 and N - 1 */
81static u32 random_N(unsigned int N)
82{
83 return reciprocal_divide(random32(), N);
84}
85
86/* number of elements in queue including holes */
87static unsigned int choke_len(const struct choke_sched_data *q)
88{
89 return (q->tail - q->head) & q->tab_mask;
90}
91
92/* Is ECN parameter configured */
93static int use_ecn(const struct choke_sched_data *q)
94{
95 return q->flags & TC_RED_ECN;
96}
97
98/* Should packets over max just be dropped (versus marked) */
99static int use_harddrop(const struct choke_sched_data *q)
100{
101 return q->flags & TC_RED_HARDDROP;
102}
103
104/* Move head pointer forward to skip over holes */
105static void choke_zap_head_holes(struct choke_sched_data *q)
106{
107 do {
108 q->head = (q->head + 1) & q->tab_mask;
109 if (q->head == q->tail)
110 break;
111 } while (q->tab[q->head] == NULL);
112}
113
114/* Move tail pointer backwards to reuse holes */
115static void choke_zap_tail_holes(struct choke_sched_data *q)
116{
117 do {
118 q->tail = (q->tail - 1) & q->tab_mask;
119 if (q->head == q->tail)
120 break;
121 } while (q->tab[q->tail] == NULL);
122}
123
124/* Drop packet from queue array by creating a "hole" */
125static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx)
126{
127 struct choke_sched_data *q = qdisc_priv(sch);
128 struct sk_buff *skb = q->tab[idx];
129
130 q->tab[idx] = NULL;
131
132 if (idx == q->head)
133 choke_zap_head_holes(q);
134 if (idx == q->tail)
135 choke_zap_tail_holes(q);
136
137 sch->qstats.backlog -= qdisc_pkt_len(skb);
138 qdisc_drop(skb, sch);
139 qdisc_tree_decrease_qlen(sch, 1);
140 --sch->q.qlen;
141}
142
143struct choke_skb_cb {
144 u16 classid;
145 u8 keys_valid;
146 struct flow_keys keys;
147};
148
149static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
150{
151 qdisc_cb_private_validate(skb, sizeof(struct choke_skb_cb));
152 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
153}
154
155static inline void choke_set_classid(struct sk_buff *skb, u16 classid)
156{
157 choke_skb_cb(skb)->classid = classid;
158}
159
160static u16 choke_get_classid(const struct sk_buff *skb)
161{
162 return choke_skb_cb(skb)->classid;
163}
164
165/*
166 * Compare flow of two packets
167 * Returns true only if source and destination address and port match.
168 * false for special cases
169 */
170static bool choke_match_flow(struct sk_buff *skb1,
171 struct sk_buff *skb2)
172{
173 if (skb1->protocol != skb2->protocol)
174 return false;
175
176 if (!choke_skb_cb(skb1)->keys_valid) {
177 choke_skb_cb(skb1)->keys_valid = 1;
178 skb_flow_dissect(skb1, &choke_skb_cb(skb1)->keys);
179 }
180
181 if (!choke_skb_cb(skb2)->keys_valid) {
182 choke_skb_cb(skb2)->keys_valid = 1;
183 skb_flow_dissect(skb2, &choke_skb_cb(skb2)->keys);
184 }
185
186 return !memcmp(&choke_skb_cb(skb1)->keys,
187 &choke_skb_cb(skb2)->keys,
188 sizeof(struct flow_keys));
189}
190
191/*
192 * Classify flow using either:
193 * 1. pre-existing classification result in skb
194 * 2. fast internal classification
195 * 3. use TC filter based classification
196 */
197static bool choke_classify(struct sk_buff *skb,
198 struct Qdisc *sch, int *qerr)
199
200{
201 struct choke_sched_data *q = qdisc_priv(sch);
202 struct tcf_result res;
203 int result;
204
205 result = tc_classify(skb, q->filter_list, &res);
206 if (result >= 0) {
207#ifdef CONFIG_NET_CLS_ACT
208 switch (result) {
209 case TC_ACT_STOLEN:
210 case TC_ACT_QUEUED:
211 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
212 case TC_ACT_SHOT:
213 return false;
214 }
215#endif
216 choke_set_classid(skb, TC_H_MIN(res.classid));
217 return true;
218 }
219
220 return false;
221}
222
223/*
224 * Select a packet at random from queue
225 * HACK: since queue can have holes from previous deletion; retry several
226 * times to find a random skb but then just give up and return the head
227 * Will return NULL if queue is empty (q->head == q->tail)
228 */
229static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
230 unsigned int *pidx)
231{
232 struct sk_buff *skb;
233 int retrys = 3;
234
235 do {
236 *pidx = (q->head + random_N(choke_len(q))) & q->tab_mask;
237 skb = q->tab[*pidx];
238 if (skb)
239 return skb;
240 } while (--retrys > 0);
241
242 return q->tab[*pidx = q->head];
243}
244
245/*
246 * Compare new packet with random packet in queue
247 * returns true if matched and sets *pidx
248 */
249static bool choke_match_random(const struct choke_sched_data *q,
250 struct sk_buff *nskb,
251 unsigned int *pidx)
252{
253 struct sk_buff *oskb;
254
255 if (q->head == q->tail)
256 return false;
257
258 oskb = choke_peek_random(q, pidx);
259 if (q->filter_list)
260 return choke_get_classid(nskb) == choke_get_classid(oskb);
261
262 return choke_match_flow(oskb, nskb);
263}
264
265static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch)
266{
267 struct choke_sched_data *q = qdisc_priv(sch);
268 const struct red_parms *p = &q->parms;
269 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
270
271 if (q->filter_list) {
272 /* If using external classifiers, get result and record it. */
273 if (!choke_classify(skb, sch, &ret))
274 goto other_drop; /* Packet was eaten by filter */
275 }
276
277 choke_skb_cb(skb)->keys_valid = 0;
278 /* Compute average queue usage (see RED) */
279 q->vars.qavg = red_calc_qavg(p, &q->vars, sch->q.qlen);
280 if (red_is_idling(&q->vars))
281 red_end_of_idle_period(&q->vars);
282
283 /* Is queue small? */
284 if (q->vars.qavg <= p->qth_min)
285 q->vars.qcount = -1;
286 else {
287 unsigned int idx;
288
289 /* Draw a packet at random from queue and compare flow */
290 if (choke_match_random(q, skb, &idx)) {
291 q->stats.matched++;
292 choke_drop_by_idx(sch, idx);
293 goto congestion_drop;
294 }
295
296 /* Queue is large, always mark/drop */
297 if (q->vars.qavg > p->qth_max) {
298 q->vars.qcount = -1;
299
300 sch->qstats.overlimits++;
301 if (use_harddrop(q) || !use_ecn(q) ||
302 !INET_ECN_set_ce(skb)) {
303 q->stats.forced_drop++;
304 goto congestion_drop;
305 }
306
307 q->stats.forced_mark++;
308 } else if (++q->vars.qcount) {
309 if (red_mark_probability(p, &q->vars, q->vars.qavg)) {
310 q->vars.qcount = 0;
311 q->vars.qR = red_random(p);
312
313 sch->qstats.overlimits++;
314 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
315 q->stats.prob_drop++;
316 goto congestion_drop;
317 }
318
319 q->stats.prob_mark++;
320 }
321 } else
322 q->vars.qR = red_random(p);
323 }
324
325 /* Admit new packet */
326 if (sch->q.qlen < q->limit) {
327 q->tab[q->tail] = skb;
328 q->tail = (q->tail + 1) & q->tab_mask;
329 ++sch->q.qlen;
330 sch->qstats.backlog += qdisc_pkt_len(skb);
331 return NET_XMIT_SUCCESS;
332 }
333
334 q->stats.pdrop++;
335 return qdisc_drop(skb, sch);
336
337congestion_drop:
338 qdisc_drop(skb, sch);
339 return NET_XMIT_CN;
340
341other_drop:
342 if (ret & __NET_XMIT_BYPASS)
343 sch->qstats.drops++;
344 kfree_skb(skb);
345 return ret;
346}
347
348static struct sk_buff *choke_dequeue(struct Qdisc *sch)
349{
350 struct choke_sched_data *q = qdisc_priv(sch);
351 struct sk_buff *skb;
352
353 if (q->head == q->tail) {
354 if (!red_is_idling(&q->vars))
355 red_start_of_idle_period(&q->vars);
356 return NULL;
357 }
358
359 skb = q->tab[q->head];
360 q->tab[q->head] = NULL;
361 choke_zap_head_holes(q);
362 --sch->q.qlen;
363 sch->qstats.backlog -= qdisc_pkt_len(skb);
364 qdisc_bstats_update(sch, skb);
365
366 return skb;
367}
368
369static unsigned int choke_drop(struct Qdisc *sch)
370{
371 struct choke_sched_data *q = qdisc_priv(sch);
372 unsigned int len;
373
374 len = qdisc_queue_drop(sch);
375 if (len > 0)
376 q->stats.other++;
377 else {
378 if (!red_is_idling(&q->vars))
379 red_start_of_idle_period(&q->vars);
380 }
381
382 return len;
383}
384
385static void choke_reset(struct Qdisc *sch)
386{
387 struct choke_sched_data *q = qdisc_priv(sch);
388
389 red_restart(&q->vars);
390}
391
392static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
393 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
394 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
395 [TCA_CHOKE_MAX_P] = { .type = NLA_U32 },
396};
397
398
399static void choke_free(void *addr)
400{
401 if (addr) {
402 if (is_vmalloc_addr(addr))
403 vfree(addr);
404 else
405 kfree(addr);
406 }
407}
408
409static int choke_change(struct Qdisc *sch, struct nlattr *opt)
410{
411 struct choke_sched_data *q = qdisc_priv(sch);
412 struct nlattr *tb[TCA_CHOKE_MAX + 1];
413 const struct tc_red_qopt *ctl;
414 int err;
415 struct sk_buff **old = NULL;
416 unsigned int mask;
417 u32 max_P;
418
419 if (opt == NULL)
420 return -EINVAL;
421
422 err = nla_parse_nested(tb, TCA_CHOKE_MAX, opt, choke_policy);
423 if (err < 0)
424 return err;
425
426 if (tb[TCA_CHOKE_PARMS] == NULL ||
427 tb[TCA_CHOKE_STAB] == NULL)
428 return -EINVAL;
429
430 max_P = tb[TCA_CHOKE_MAX_P] ? nla_get_u32(tb[TCA_CHOKE_MAX_P]) : 0;
431
432 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
433
434 if (ctl->limit > CHOKE_MAX_QUEUE)
435 return -EINVAL;
436
437 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
438 if (mask != q->tab_mask) {
439 struct sk_buff **ntab;
440
441 ntab = kcalloc(mask + 1, sizeof(struct sk_buff *), GFP_KERNEL);
442 if (!ntab)
443 ntab = vzalloc((mask + 1) * sizeof(struct sk_buff *));
444 if (!ntab)
445 return -ENOMEM;
446
447 sch_tree_lock(sch);
448 old = q->tab;
449 if (old) {
450 unsigned int oqlen = sch->q.qlen, tail = 0;
451
452 while (q->head != q->tail) {
453 struct sk_buff *skb = q->tab[q->head];
454
455 q->head = (q->head + 1) & q->tab_mask;
456 if (!skb)
457 continue;
458 if (tail < mask) {
459 ntab[tail++] = skb;
460 continue;
461 }
462 sch->qstats.backlog -= qdisc_pkt_len(skb);
463 --sch->q.qlen;
464 qdisc_drop(skb, sch);
465 }
466 qdisc_tree_decrease_qlen(sch, oqlen - sch->q.qlen);
467 q->head = 0;
468 q->tail = tail;
469 }
470
471 q->tab_mask = mask;
472 q->tab = ntab;
473 } else
474 sch_tree_lock(sch);
475
476 q->flags = ctl->flags;
477 q->limit = ctl->limit;
478
479 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
480 ctl->Plog, ctl->Scell_log,
481 nla_data(tb[TCA_CHOKE_STAB]),
482 max_P);
483 red_set_vars(&q->vars);
484
485 if (q->head == q->tail)
486 red_end_of_idle_period(&q->vars);
487
488 sch_tree_unlock(sch);
489 choke_free(old);
490 return 0;
491}
492
493static int choke_init(struct Qdisc *sch, struct nlattr *opt)
494{
495 return choke_change(sch, opt);
496}
497
498static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
499{
500 struct choke_sched_data *q = qdisc_priv(sch);
501 struct nlattr *opts = NULL;
502 struct tc_red_qopt opt = {
503 .limit = q->limit,
504 .flags = q->flags,
505 .qth_min = q->parms.qth_min >> q->parms.Wlog,
506 .qth_max = q->parms.qth_max >> q->parms.Wlog,
507 .Wlog = q->parms.Wlog,
508 .Plog = q->parms.Plog,
509 .Scell_log = q->parms.Scell_log,
510 };
511
512 opts = nla_nest_start(skb, TCA_OPTIONS);
513 if (opts == NULL)
514 goto nla_put_failure;
515
516 if (nla_put(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt) ||
517 nla_put_u32(skb, TCA_CHOKE_MAX_P, q->parms.max_P))
518 goto nla_put_failure;
519 return nla_nest_end(skb, opts);
520
521nla_put_failure:
522 nla_nest_cancel(skb, opts);
523 return -EMSGSIZE;
524}
525
526static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
527{
528 struct choke_sched_data *q = qdisc_priv(sch);
529 struct tc_choke_xstats st = {
530 .early = q->stats.prob_drop + q->stats.forced_drop,
531 .marked = q->stats.prob_mark + q->stats.forced_mark,
532 .pdrop = q->stats.pdrop,
533 .other = q->stats.other,
534 .matched = q->stats.matched,
535 };
536
537 return gnet_stats_copy_app(d, &st, sizeof(st));
538}
539
540static void choke_destroy(struct Qdisc *sch)
541{
542 struct choke_sched_data *q = qdisc_priv(sch);
543
544 tcf_destroy_chain(&q->filter_list);
545 choke_free(q->tab);
546}
547
548static struct Qdisc *choke_leaf(struct Qdisc *sch, unsigned long arg)
549{
550 return NULL;
551}
552
553static unsigned long choke_get(struct Qdisc *sch, u32 classid)
554{
555 return 0;
556}
557
558static void choke_put(struct Qdisc *q, unsigned long cl)
559{
560}
561
562static unsigned long choke_bind(struct Qdisc *sch, unsigned long parent,
563 u32 classid)
564{
565 return 0;
566}
567
568static struct tcf_proto **choke_find_tcf(struct Qdisc *sch, unsigned long cl)
569{
570 struct choke_sched_data *q = qdisc_priv(sch);
571
572 if (cl)
573 return NULL;
574 return &q->filter_list;
575}
576
577static int choke_dump_class(struct Qdisc *sch, unsigned long cl,
578 struct sk_buff *skb, struct tcmsg *tcm)
579{
580 tcm->tcm_handle |= TC_H_MIN(cl);
581 return 0;
582}
583
584static void choke_walk(struct Qdisc *sch, struct qdisc_walker *arg)
585{
586 if (!arg->stop) {
587 if (arg->fn(sch, 1, arg) < 0) {
588 arg->stop = 1;
589 return;
590 }
591 arg->count++;
592 }
593}
594
595static const struct Qdisc_class_ops choke_class_ops = {
596 .leaf = choke_leaf,
597 .get = choke_get,
598 .put = choke_put,
599 .tcf_chain = choke_find_tcf,
600 .bind_tcf = choke_bind,
601 .unbind_tcf = choke_put,
602 .dump = choke_dump_class,
603 .walk = choke_walk,
604};
605
606static struct sk_buff *choke_peek_head(struct Qdisc *sch)
607{
608 struct choke_sched_data *q = qdisc_priv(sch);
609
610 return (q->head != q->tail) ? q->tab[q->head] : NULL;
611}
612
613static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
614 .id = "choke",
615 .priv_size = sizeof(struct choke_sched_data),
616
617 .enqueue = choke_enqueue,
618 .dequeue = choke_dequeue,
619 .peek = choke_peek_head,
620 .drop = choke_drop,
621 .init = choke_init,
622 .destroy = choke_destroy,
623 .reset = choke_reset,
624 .change = choke_change,
625 .dump = choke_dump,
626 .dump_stats = choke_dump_stats,
627 .owner = THIS_MODULE,
628};
629
630static int __init choke_module_init(void)
631{
632 return register_qdisc(&choke_qdisc_ops);
633}
634
635static void __exit choke_module_exit(void)
636{
637 unregister_qdisc(&choke_qdisc_ops);
638}
639
640module_init(choke_module_init)
641module_exit(choke_module_exit)
642
643MODULE_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 other; /* Drops due to drop() calls */
64 u32 matched; /* Drops to flow match */
65 } stats;
66
67 unsigned int head;
68 unsigned int tail;
69
70 unsigned int tab_mask; /* size - 1 */
71
72 struct sk_buff **tab;
73};
74
75/* number of elements in queue including holes */
76static unsigned int choke_len(const struct choke_sched_data *q)
77{
78 return (q->tail - q->head) & q->tab_mask;
79}
80
81/* Is ECN parameter configured */
82static int use_ecn(const struct choke_sched_data *q)
83{
84 return q->flags & TC_RED_ECN;
85}
86
87/* Should packets over max just be dropped (versus marked) */
88static int use_harddrop(const struct choke_sched_data *q)
89{
90 return q->flags & TC_RED_HARDDROP;
91}
92
93/* Move head pointer forward to skip over holes */
94static void choke_zap_head_holes(struct choke_sched_data *q)
95{
96 do {
97 q->head = (q->head + 1) & q->tab_mask;
98 if (q->head == q->tail)
99 break;
100 } while (q->tab[q->head] == NULL);
101}
102
103/* Move tail pointer backwards to reuse holes */
104static void choke_zap_tail_holes(struct choke_sched_data *q)
105{
106 do {
107 q->tail = (q->tail - 1) & q->tab_mask;
108 if (q->head == q->tail)
109 break;
110 } while (q->tab[q->tail] == NULL);
111}
112
113/* Drop packet from queue array by creating a "hole" */
114static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx,
115 struct sk_buff **to_free)
116{
117 struct choke_sched_data *q = qdisc_priv(sch);
118 struct sk_buff *skb = q->tab[idx];
119
120 q->tab[idx] = NULL;
121
122 if (idx == q->head)
123 choke_zap_head_holes(q);
124 if (idx == q->tail)
125 choke_zap_tail_holes(q);
126
127 qdisc_qstats_backlog_dec(sch, skb);
128 qdisc_tree_reduce_backlog(sch, 1, qdisc_pkt_len(skb));
129 qdisc_drop(skb, sch, to_free);
130 --sch->q.qlen;
131}
132
133struct choke_skb_cb {
134 u8 keys_valid;
135 struct flow_keys_digest keys;
136};
137
138static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
139{
140 qdisc_cb_private_validate(skb, sizeof(struct choke_skb_cb));
141 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
142}
143
144/*
145 * Compare flow of two packets
146 * Returns true only if source and destination address and port match.
147 * false for special cases
148 */
149static bool choke_match_flow(struct sk_buff *skb1,
150 struct sk_buff *skb2)
151{
152 struct flow_keys temp;
153
154 if (skb1->protocol != skb2->protocol)
155 return false;
156
157 if (!choke_skb_cb(skb1)->keys_valid) {
158 choke_skb_cb(skb1)->keys_valid = 1;
159 skb_flow_dissect_flow_keys(skb1, &temp, 0);
160 make_flow_keys_digest(&choke_skb_cb(skb1)->keys, &temp);
161 }
162
163 if (!choke_skb_cb(skb2)->keys_valid) {
164 choke_skb_cb(skb2)->keys_valid = 1;
165 skb_flow_dissect_flow_keys(skb2, &temp, 0);
166 make_flow_keys_digest(&choke_skb_cb(skb2)->keys, &temp);
167 }
168
169 return !memcmp(&choke_skb_cb(skb1)->keys,
170 &choke_skb_cb(skb2)->keys,
171 sizeof(choke_skb_cb(skb1)->keys));
172}
173
174/*
175 * Select a packet at random from queue
176 * HACK: since queue can have holes from previous deletion; retry several
177 * times to find a random skb but then just give up and return the head
178 * Will return NULL if queue is empty (q->head == q->tail)
179 */
180static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
181 unsigned int *pidx)
182{
183 struct sk_buff *skb;
184 int retrys = 3;
185
186 do {
187 *pidx = (q->head + prandom_u32_max(choke_len(q))) & q->tab_mask;
188 skb = q->tab[*pidx];
189 if (skb)
190 return skb;
191 } while (--retrys > 0);
192
193 return q->tab[*pidx = q->head];
194}
195
196/*
197 * Compare new packet with random packet in queue
198 * returns true if matched and sets *pidx
199 */
200static bool choke_match_random(const struct choke_sched_data *q,
201 struct sk_buff *nskb,
202 unsigned int *pidx)
203{
204 struct sk_buff *oskb;
205
206 if (q->head == q->tail)
207 return false;
208
209 oskb = choke_peek_random(q, pidx);
210 return choke_match_flow(oskb, nskb);
211}
212
213static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch,
214 struct sk_buff **to_free)
215{
216 struct choke_sched_data *q = qdisc_priv(sch);
217 const struct red_parms *p = &q->parms;
218
219 choke_skb_cb(skb)->keys_valid = 0;
220 /* Compute average queue usage (see RED) */
221 q->vars.qavg = red_calc_qavg(p, &q->vars, sch->q.qlen);
222 if (red_is_idling(&q->vars))
223 red_end_of_idle_period(&q->vars);
224
225 /* Is queue small? */
226 if (q->vars.qavg <= p->qth_min)
227 q->vars.qcount = -1;
228 else {
229 unsigned int idx;
230
231 /* Draw a packet at random from queue and compare flow */
232 if (choke_match_random(q, skb, &idx)) {
233 q->stats.matched++;
234 choke_drop_by_idx(sch, idx, to_free);
235 goto congestion_drop;
236 }
237
238 /* Queue is large, always mark/drop */
239 if (q->vars.qavg > p->qth_max) {
240 q->vars.qcount = -1;
241
242 qdisc_qstats_overlimit(sch);
243 if (use_harddrop(q) || !use_ecn(q) ||
244 !INET_ECN_set_ce(skb)) {
245 q->stats.forced_drop++;
246 goto congestion_drop;
247 }
248
249 q->stats.forced_mark++;
250 } else if (++q->vars.qcount) {
251 if (red_mark_probability(p, &q->vars, q->vars.qavg)) {
252 q->vars.qcount = 0;
253 q->vars.qR = red_random(p);
254
255 qdisc_qstats_overlimit(sch);
256 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
257 q->stats.prob_drop++;
258 goto congestion_drop;
259 }
260
261 q->stats.prob_mark++;
262 }
263 } else
264 q->vars.qR = red_random(p);
265 }
266
267 /* Admit new packet */
268 if (sch->q.qlen < q->limit) {
269 q->tab[q->tail] = skb;
270 q->tail = (q->tail + 1) & q->tab_mask;
271 ++sch->q.qlen;
272 qdisc_qstats_backlog_inc(sch, skb);
273 return NET_XMIT_SUCCESS;
274 }
275
276 q->stats.pdrop++;
277 return qdisc_drop(skb, sch, to_free);
278
279congestion_drop:
280 qdisc_drop(skb, sch, to_free);
281 return NET_XMIT_CN;
282}
283
284static struct sk_buff *choke_dequeue(struct Qdisc *sch)
285{
286 struct choke_sched_data *q = qdisc_priv(sch);
287 struct sk_buff *skb;
288
289 if (q->head == q->tail) {
290 if (!red_is_idling(&q->vars))
291 red_start_of_idle_period(&q->vars);
292 return NULL;
293 }
294
295 skb = q->tab[q->head];
296 q->tab[q->head] = NULL;
297 choke_zap_head_holes(q);
298 --sch->q.qlen;
299 qdisc_qstats_backlog_dec(sch, skb);
300 qdisc_bstats_update(sch, skb);
301
302 return skb;
303}
304
305static void choke_reset(struct Qdisc *sch)
306{
307 struct choke_sched_data *q = qdisc_priv(sch);
308
309 while (q->head != q->tail) {
310 struct sk_buff *skb = q->tab[q->head];
311
312 q->head = (q->head + 1) & q->tab_mask;
313 if (!skb)
314 continue;
315 rtnl_qdisc_drop(skb, sch);
316 }
317
318 sch->q.qlen = 0;
319 sch->qstats.backlog = 0;
320 if (q->tab)
321 memset(q->tab, 0, (q->tab_mask + 1) * sizeof(struct sk_buff *));
322 q->head = q->tail = 0;
323 red_restart(&q->vars);
324}
325
326static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
327 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
328 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
329 [TCA_CHOKE_MAX_P] = { .type = NLA_U32 },
330};
331
332
333static void choke_free(void *addr)
334{
335 kvfree(addr);
336}
337
338static int choke_change(struct Qdisc *sch, struct nlattr *opt,
339 struct netlink_ext_ack *extack)
340{
341 struct choke_sched_data *q = qdisc_priv(sch);
342 struct nlattr *tb[TCA_CHOKE_MAX + 1];
343 const struct tc_red_qopt *ctl;
344 int err;
345 struct sk_buff **old = NULL;
346 unsigned int mask;
347 u32 max_P;
348
349 if (opt == NULL)
350 return -EINVAL;
351
352 err = nla_parse_nested_deprecated(tb, TCA_CHOKE_MAX, opt,
353 choke_policy, NULL);
354 if (err < 0)
355 return err;
356
357 if (tb[TCA_CHOKE_PARMS] == NULL ||
358 tb[TCA_CHOKE_STAB] == NULL)
359 return -EINVAL;
360
361 max_P = tb[TCA_CHOKE_MAX_P] ? nla_get_u32(tb[TCA_CHOKE_MAX_P]) : 0;
362
363 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
364
365 if (!red_check_params(ctl->qth_min, ctl->qth_max, ctl->Wlog))
366 return -EINVAL;
367
368 if (ctl->limit > CHOKE_MAX_QUEUE)
369 return -EINVAL;
370
371 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
372 if (mask != q->tab_mask) {
373 struct sk_buff **ntab;
374
375 ntab = kvcalloc(mask + 1, sizeof(struct sk_buff *), GFP_KERNEL);
376 if (!ntab)
377 return -ENOMEM;
378
379 sch_tree_lock(sch);
380 old = q->tab;
381 if (old) {
382 unsigned int oqlen = sch->q.qlen, tail = 0;
383 unsigned dropped = 0;
384
385 while (q->head != q->tail) {
386 struct sk_buff *skb = q->tab[q->head];
387
388 q->head = (q->head + 1) & q->tab_mask;
389 if (!skb)
390 continue;
391 if (tail < mask) {
392 ntab[tail++] = skb;
393 continue;
394 }
395 dropped += qdisc_pkt_len(skb);
396 qdisc_qstats_backlog_dec(sch, skb);
397 --sch->q.qlen;
398 rtnl_qdisc_drop(skb, sch);
399 }
400 qdisc_tree_reduce_backlog(sch, oqlen - sch->q.qlen, dropped);
401 q->head = 0;
402 q->tail = tail;
403 }
404
405 q->tab_mask = mask;
406 q->tab = ntab;
407 } else
408 sch_tree_lock(sch);
409
410 q->flags = ctl->flags;
411 q->limit = ctl->limit;
412
413 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
414 ctl->Plog, ctl->Scell_log,
415 nla_data(tb[TCA_CHOKE_STAB]),
416 max_P);
417 red_set_vars(&q->vars);
418
419 if (q->head == q->tail)
420 red_end_of_idle_period(&q->vars);
421
422 sch_tree_unlock(sch);
423 choke_free(old);
424 return 0;
425}
426
427static int choke_init(struct Qdisc *sch, struct nlattr *opt,
428 struct netlink_ext_ack *extack)
429{
430 return choke_change(sch, opt, extack);
431}
432
433static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
434{
435 struct choke_sched_data *q = qdisc_priv(sch);
436 struct nlattr *opts = NULL;
437 struct tc_red_qopt opt = {
438 .limit = q->limit,
439 .flags = q->flags,
440 .qth_min = q->parms.qth_min >> q->parms.Wlog,
441 .qth_max = q->parms.qth_max >> q->parms.Wlog,
442 .Wlog = q->parms.Wlog,
443 .Plog = q->parms.Plog,
444 .Scell_log = q->parms.Scell_log,
445 };
446
447 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
448 if (opts == NULL)
449 goto nla_put_failure;
450
451 if (nla_put(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt) ||
452 nla_put_u32(skb, TCA_CHOKE_MAX_P, q->parms.max_P))
453 goto nla_put_failure;
454 return nla_nest_end(skb, opts);
455
456nla_put_failure:
457 nla_nest_cancel(skb, opts);
458 return -EMSGSIZE;
459}
460
461static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
462{
463 struct choke_sched_data *q = qdisc_priv(sch);
464 struct tc_choke_xstats st = {
465 .early = q->stats.prob_drop + q->stats.forced_drop,
466 .marked = q->stats.prob_mark + q->stats.forced_mark,
467 .pdrop = q->stats.pdrop,
468 .other = q->stats.other,
469 .matched = q->stats.matched,
470 };
471
472 return gnet_stats_copy_app(d, &st, sizeof(st));
473}
474
475static void choke_destroy(struct Qdisc *sch)
476{
477 struct choke_sched_data *q = qdisc_priv(sch);
478
479 choke_free(q->tab);
480}
481
482static struct sk_buff *choke_peek_head(struct Qdisc *sch)
483{
484 struct choke_sched_data *q = qdisc_priv(sch);
485
486 return (q->head != q->tail) ? q->tab[q->head] : NULL;
487}
488
489static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
490 .id = "choke",
491 .priv_size = sizeof(struct choke_sched_data),
492
493 .enqueue = choke_enqueue,
494 .dequeue = choke_dequeue,
495 .peek = choke_peek_head,
496 .init = choke_init,
497 .destroy = choke_destroy,
498 .reset = choke_reset,
499 .change = choke_change,
500 .dump = choke_dump,
501 .dump_stats = choke_dump_stats,
502 .owner = THIS_MODULE,
503};
504
505static int __init choke_module_init(void)
506{
507 return register_qdisc(&choke_qdisc_ops);
508}
509
510static void __exit choke_module_exit(void)
511{
512 unregister_qdisc(&choke_qdisc_ops);
513}
514
515module_init(choke_module_init)
516module_exit(choke_module_exit)
517
518MODULE_LICENSE("GPL");