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 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 --sch->q.qlen;
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}
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 = nla_get_u32_default(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 WRITE_ONCE(q->flags, ctl->flags);
409 WRITE_ONCE(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 u8 Wlog = READ_ONCE(q->parms.Wlog);
435 struct nlattr *opts = NULL;
436 struct tc_red_qopt opt = {
437 .limit = READ_ONCE(q->limit),
438 .flags = READ_ONCE(q->flags),
439 .qth_min = READ_ONCE(q->parms.qth_min) >> Wlog,
440 .qth_max = READ_ONCE(q->parms.qth_max) >> Wlog,
441 .Wlog = Wlog,
442 .Plog = READ_ONCE(q->parms.Plog),
443 .Scell_log = READ_ONCE(q->parms.Scell_log),
444 };
445
446 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
447 if (opts == NULL)
448 goto nla_put_failure;
449
450 if (nla_put(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt) ||
451 nla_put_u32(skb, TCA_CHOKE_MAX_P, READ_ONCE(q->parms.max_P)))
452 goto nla_put_failure;
453 return nla_nest_end(skb, opts);
454
455nla_put_failure:
456 nla_nest_cancel(skb, opts);
457 return -EMSGSIZE;
458}
459
460static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
461{
462 struct choke_sched_data *q = qdisc_priv(sch);
463 struct tc_choke_xstats st = {
464 .early = q->stats.prob_drop + q->stats.forced_drop,
465 .marked = q->stats.prob_mark + q->stats.forced_mark,
466 .pdrop = q->stats.pdrop,
467 .matched = q->stats.matched,
468 };
469
470 return gnet_stats_copy_app(d, &st, sizeof(st));
471}
472
473static void choke_destroy(struct Qdisc *sch)
474{
475 struct choke_sched_data *q = qdisc_priv(sch);
476
477 choke_free(q->tab);
478}
479
480static struct sk_buff *choke_peek_head(struct Qdisc *sch)
481{
482 struct choke_sched_data *q = qdisc_priv(sch);
483
484 return (q->head != q->tail) ? q->tab[q->head] : NULL;
485}
486
487static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
488 .id = "choke",
489 .priv_size = sizeof(struct choke_sched_data),
490
491 .enqueue = choke_enqueue,
492 .dequeue = choke_dequeue,
493 .peek = choke_peek_head,
494 .init = choke_init,
495 .destroy = choke_destroy,
496 .reset = choke_reset,
497 .change = choke_change,
498 .dump = choke_dump,
499 .dump_stats = choke_dump_stats,
500 .owner = THIS_MODULE,
501};
502MODULE_ALIAS_NET_SCH("choke");
503
504static int __init choke_module_init(void)
505{
506 return register_qdisc(&choke_qdisc_ops);
507}
508
509static void __exit choke_module_exit(void)
510{
511 unregister_qdisc(&choke_qdisc_ops);
512}
513
514module_init(choke_module_init)
515module_exit(choke_module_exit)
516
517MODULE_LICENSE("GPL");
518MODULE_DESCRIPTION("Choose and keep responsive flows scheduler");