Loading...
1/* Copyright (C) 2013 Cisco Systems, Inc, 2013.
2 *
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of the GNU General Public License
5 * as published by the Free Software Foundation; either version 2
6 * of the License.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
12 *
13 * Author: Vijay Subramanian <vijaynsu@cisco.com>
14 * Author: Mythili Prabhu <mysuryan@cisco.com>
15 *
16 * ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
17 * University of Oslo, Norway.
18 *
19 * References:
20 * IETF draft submission: http://tools.ietf.org/html/draft-pan-aqm-pie-00
21 * IEEE Conference on High Performance Switching and Routing 2013 :
22 * "PIE: A * Lightweight Control Scheme to Address the Bufferbloat Problem"
23 */
24
25#include <linux/module.h>
26#include <linux/slab.h>
27#include <linux/types.h>
28#include <linux/kernel.h>
29#include <linux/errno.h>
30#include <linux/skbuff.h>
31#include <net/pkt_sched.h>
32#include <net/inet_ecn.h>
33
34#define QUEUE_THRESHOLD 10000
35#define DQCOUNT_INVALID -1
36#define MAX_PROB 0xffffffff
37#define PIE_SCALE 8
38
39/* parameters used */
40struct pie_params {
41 psched_time_t target; /* user specified target delay in pschedtime */
42 u32 tupdate; /* timer frequency (in jiffies) */
43 u32 limit; /* number of packets that can be enqueued */
44 u32 alpha; /* alpha and beta are between 0 and 32 */
45 u32 beta; /* and are used for shift relative to 1 */
46 bool ecn; /* true if ecn is enabled */
47 bool bytemode; /* to scale drop early prob based on pkt size */
48};
49
50/* variables used */
51struct pie_vars {
52 u32 prob; /* probability but scaled by u32 limit. */
53 psched_time_t burst_time;
54 psched_time_t qdelay;
55 psched_time_t qdelay_old;
56 u64 dq_count; /* measured in bytes */
57 psched_time_t dq_tstamp; /* drain rate */
58 u32 avg_dq_rate; /* bytes per pschedtime tick,scaled */
59 u32 qlen_old; /* in bytes */
60};
61
62/* statistics gathering */
63struct pie_stats {
64 u32 packets_in; /* total number of packets enqueued */
65 u32 dropped; /* packets dropped due to pie_action */
66 u32 overlimit; /* dropped due to lack of space in queue */
67 u32 maxq; /* maximum queue size */
68 u32 ecn_mark; /* packets marked with ECN */
69};
70
71/* private data for the Qdisc */
72struct pie_sched_data {
73 struct pie_params params;
74 struct pie_vars vars;
75 struct pie_stats stats;
76 struct timer_list adapt_timer;
77};
78
79static void pie_params_init(struct pie_params *params)
80{
81 params->alpha = 2;
82 params->beta = 20;
83 params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC); /* 30 ms */
84 params->limit = 1000; /* default of 1000 packets */
85 params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC); /* 20 ms */
86 params->ecn = false;
87 params->bytemode = false;
88}
89
90static void pie_vars_init(struct pie_vars *vars)
91{
92 vars->dq_count = DQCOUNT_INVALID;
93 vars->avg_dq_rate = 0;
94 /* default of 100 ms in pschedtime */
95 vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
96}
97
98static bool drop_early(struct Qdisc *sch, u32 packet_size)
99{
100 struct pie_sched_data *q = qdisc_priv(sch);
101 u32 rnd;
102 u32 local_prob = q->vars.prob;
103 u32 mtu = psched_mtu(qdisc_dev(sch));
104
105 /* If there is still burst allowance left skip random early drop */
106 if (q->vars.burst_time > 0)
107 return false;
108
109 /* If current delay is less than half of target, and
110 * if drop prob is low already, disable early_drop
111 */
112 if ((q->vars.qdelay < q->params.target / 2)
113 && (q->vars.prob < MAX_PROB / 5))
114 return false;
115
116 /* If we have fewer than 2 mtu-sized packets, disable drop_early,
117 * similar to min_th in RED
118 */
119 if (sch->qstats.backlog < 2 * mtu)
120 return false;
121
122 /* If bytemode is turned on, use packet size to compute new
123 * probablity. Smaller packets will have lower drop prob in this case
124 */
125 if (q->params.bytemode && packet_size <= mtu)
126 local_prob = (local_prob / mtu) * packet_size;
127 else
128 local_prob = q->vars.prob;
129
130 rnd = prandom_u32();
131 if (rnd < local_prob)
132 return true;
133
134 return false;
135}
136
137static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
138{
139 struct pie_sched_data *q = qdisc_priv(sch);
140 bool enqueue = false;
141
142 if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
143 q->stats.overlimit++;
144 goto out;
145 }
146
147 if (!drop_early(sch, skb->len)) {
148 enqueue = true;
149 } else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
150 INET_ECN_set_ce(skb)) {
151 /* If packet is ecn capable, mark it if drop probability
152 * is lower than 10%, else drop it.
153 */
154 q->stats.ecn_mark++;
155 enqueue = true;
156 }
157
158 /* we can enqueue the packet */
159 if (enqueue) {
160 q->stats.packets_in++;
161 if (qdisc_qlen(sch) > q->stats.maxq)
162 q->stats.maxq = qdisc_qlen(sch);
163
164 return qdisc_enqueue_tail(skb, sch);
165 }
166
167out:
168 q->stats.dropped++;
169 return qdisc_drop(skb, sch);
170}
171
172static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
173 [TCA_PIE_TARGET] = {.type = NLA_U32},
174 [TCA_PIE_LIMIT] = {.type = NLA_U32},
175 [TCA_PIE_TUPDATE] = {.type = NLA_U32},
176 [TCA_PIE_ALPHA] = {.type = NLA_U32},
177 [TCA_PIE_BETA] = {.type = NLA_U32},
178 [TCA_PIE_ECN] = {.type = NLA_U32},
179 [TCA_PIE_BYTEMODE] = {.type = NLA_U32},
180};
181
182static int pie_change(struct Qdisc *sch, struct nlattr *opt)
183{
184 struct pie_sched_data *q = qdisc_priv(sch);
185 struct nlattr *tb[TCA_PIE_MAX + 1];
186 unsigned int qlen, dropped = 0;
187 int err;
188
189 if (!opt)
190 return -EINVAL;
191
192 err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy);
193 if (err < 0)
194 return err;
195
196 sch_tree_lock(sch);
197
198 /* convert from microseconds to pschedtime */
199 if (tb[TCA_PIE_TARGET]) {
200 /* target is in us */
201 u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
202
203 /* convert to pschedtime */
204 q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
205 }
206
207 /* tupdate is in jiffies */
208 if (tb[TCA_PIE_TUPDATE])
209 q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
210
211 if (tb[TCA_PIE_LIMIT]) {
212 u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
213
214 q->params.limit = limit;
215 sch->limit = limit;
216 }
217
218 if (tb[TCA_PIE_ALPHA])
219 q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
220
221 if (tb[TCA_PIE_BETA])
222 q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
223
224 if (tb[TCA_PIE_ECN])
225 q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
226
227 if (tb[TCA_PIE_BYTEMODE])
228 q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
229
230 /* Drop excess packets if new limit is lower */
231 qlen = sch->q.qlen;
232 while (sch->q.qlen > sch->limit) {
233 struct sk_buff *skb = __skb_dequeue(&sch->q);
234
235 dropped += qdisc_pkt_len(skb);
236 qdisc_qstats_backlog_dec(sch, skb);
237 qdisc_drop(skb, sch);
238 }
239 qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
240
241 sch_tree_unlock(sch);
242 return 0;
243}
244
245static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
246{
247
248 struct pie_sched_data *q = qdisc_priv(sch);
249 int qlen = sch->qstats.backlog; /* current queue size in bytes */
250
251 /* If current queue is about 10 packets or more and dq_count is unset
252 * we have enough packets to calculate the drain rate. Save
253 * current time as dq_tstamp and start measurement cycle.
254 */
255 if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
256 q->vars.dq_tstamp = psched_get_time();
257 q->vars.dq_count = 0;
258 }
259
260 /* Calculate the average drain rate from this value. If queue length
261 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
262 * the dq_count to -1 as we don't have enough packets to calculate the
263 * drain rate anymore The following if block is entered only when we
264 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
265 * and we calculate the drain rate for the threshold here. dq_count is
266 * in bytes, time difference in psched_time, hence rate is in
267 * bytes/psched_time.
268 */
269 if (q->vars.dq_count != DQCOUNT_INVALID) {
270 q->vars.dq_count += skb->len;
271
272 if (q->vars.dq_count >= QUEUE_THRESHOLD) {
273 psched_time_t now = psched_get_time();
274 u32 dtime = now - q->vars.dq_tstamp;
275 u32 count = q->vars.dq_count << PIE_SCALE;
276
277 if (dtime == 0)
278 return;
279
280 count = count / dtime;
281
282 if (q->vars.avg_dq_rate == 0)
283 q->vars.avg_dq_rate = count;
284 else
285 q->vars.avg_dq_rate =
286 (q->vars.avg_dq_rate -
287 (q->vars.avg_dq_rate >> 3)) + (count >> 3);
288
289 /* If the queue has receded below the threshold, we hold
290 * on to the last drain rate calculated, else we reset
291 * dq_count to 0 to re-enter the if block when the next
292 * packet is dequeued
293 */
294 if (qlen < QUEUE_THRESHOLD)
295 q->vars.dq_count = DQCOUNT_INVALID;
296 else {
297 q->vars.dq_count = 0;
298 q->vars.dq_tstamp = psched_get_time();
299 }
300
301 if (q->vars.burst_time > 0) {
302 if (q->vars.burst_time > dtime)
303 q->vars.burst_time -= dtime;
304 else
305 q->vars.burst_time = 0;
306 }
307 }
308 }
309}
310
311static void calculate_probability(struct Qdisc *sch)
312{
313 struct pie_sched_data *q = qdisc_priv(sch);
314 u32 qlen = sch->qstats.backlog; /* queue size in bytes */
315 psched_time_t qdelay = 0; /* in pschedtime */
316 psched_time_t qdelay_old = q->vars.qdelay; /* in pschedtime */
317 s32 delta = 0; /* determines the change in probability */
318 u32 oldprob;
319 u32 alpha, beta;
320 bool update_prob = true;
321
322 q->vars.qdelay_old = q->vars.qdelay;
323
324 if (q->vars.avg_dq_rate > 0)
325 qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
326 else
327 qdelay = 0;
328
329 /* If qdelay is zero and qlen is not, it means qlen is very small, less
330 * than dequeue_rate, so we do not update probabilty in this round
331 */
332 if (qdelay == 0 && qlen != 0)
333 update_prob = false;
334
335 /* In the algorithm, alpha and beta are between 0 and 2 with typical
336 * value for alpha as 0.125. In this implementation, we use values 0-32
337 * passed from user space to represent this. Also, alpha and beta have
338 * unit of HZ and need to be scaled before they can used to update
339 * probability. alpha/beta are updated locally below by 1) scaling them
340 * appropriately 2) scaling down by 16 to come to 0-2 range.
341 * Please see paper for details.
342 *
343 * We scale alpha and beta differently depending on whether we are in
344 * light, medium or high dropping mode.
345 */
346 if (q->vars.prob < MAX_PROB / 100) {
347 alpha =
348 (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
349 beta =
350 (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
351 } else if (q->vars.prob < MAX_PROB / 10) {
352 alpha =
353 (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
354 beta =
355 (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
356 } else {
357 alpha =
358 (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
359 beta =
360 (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
361 }
362
363 /* alpha and beta should be between 0 and 32, in multiples of 1/16 */
364 delta += alpha * ((qdelay - q->params.target));
365 delta += beta * ((qdelay - qdelay_old));
366
367 oldprob = q->vars.prob;
368
369 /* to ensure we increase probability in steps of no more than 2% */
370 if (delta > (s32) (MAX_PROB / (100 / 2)) &&
371 q->vars.prob >= MAX_PROB / 10)
372 delta = (MAX_PROB / 100) * 2;
373
374 /* Non-linear drop:
375 * Tune drop probability to increase quickly for high delays(>= 250ms)
376 * 250ms is derived through experiments and provides error protection
377 */
378
379 if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
380 delta += MAX_PROB / (100 / 2);
381
382 q->vars.prob += delta;
383
384 if (delta > 0) {
385 /* prevent overflow */
386 if (q->vars.prob < oldprob) {
387 q->vars.prob = MAX_PROB;
388 /* Prevent normalization error. If probability is at
389 * maximum value already, we normalize it here, and
390 * skip the check to do a non-linear drop in the next
391 * section.
392 */
393 update_prob = false;
394 }
395 } else {
396 /* prevent underflow */
397 if (q->vars.prob > oldprob)
398 q->vars.prob = 0;
399 }
400
401 /* Non-linear drop in probability: Reduce drop probability quickly if
402 * delay is 0 for 2 consecutive Tupdate periods.
403 */
404
405 if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
406 q->vars.prob = (q->vars.prob * 98) / 100;
407
408 q->vars.qdelay = qdelay;
409 q->vars.qlen_old = qlen;
410
411 /* We restart the measurement cycle if the following conditions are met
412 * 1. If the delay has been low for 2 consecutive Tupdate periods
413 * 2. Calculated drop probability is zero
414 * 3. We have atleast one estimate for the avg_dq_rate ie.,
415 * is a non-zero value
416 */
417 if ((q->vars.qdelay < q->params.target / 2) &&
418 (q->vars.qdelay_old < q->params.target / 2) &&
419 (q->vars.prob == 0) &&
420 (q->vars.avg_dq_rate > 0))
421 pie_vars_init(&q->vars);
422}
423
424static void pie_timer(unsigned long arg)
425{
426 struct Qdisc *sch = (struct Qdisc *)arg;
427 struct pie_sched_data *q = qdisc_priv(sch);
428 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
429
430 spin_lock(root_lock);
431 calculate_probability(sch);
432
433 /* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
434 if (q->params.tupdate)
435 mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
436 spin_unlock(root_lock);
437
438}
439
440static int pie_init(struct Qdisc *sch, struct nlattr *opt)
441{
442 struct pie_sched_data *q = qdisc_priv(sch);
443
444 pie_params_init(&q->params);
445 pie_vars_init(&q->vars);
446 sch->limit = q->params.limit;
447
448 setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);
449
450 if (opt) {
451 int err = pie_change(sch, opt);
452
453 if (err)
454 return err;
455 }
456
457 mod_timer(&q->adapt_timer, jiffies + HZ / 2);
458 return 0;
459}
460
461static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
462{
463 struct pie_sched_data *q = qdisc_priv(sch);
464 struct nlattr *opts;
465
466 opts = nla_nest_start(skb, TCA_OPTIONS);
467 if (opts == NULL)
468 goto nla_put_failure;
469
470 /* convert target from pschedtime to us */
471 if (nla_put_u32(skb, TCA_PIE_TARGET,
472 ((u32) PSCHED_TICKS2NS(q->params.target)) /
473 NSEC_PER_USEC) ||
474 nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
475 nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
476 nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
477 nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
478 nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
479 nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
480 goto nla_put_failure;
481
482 return nla_nest_end(skb, opts);
483
484nla_put_failure:
485 nla_nest_cancel(skb, opts);
486 return -1;
487
488}
489
490static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
491{
492 struct pie_sched_data *q = qdisc_priv(sch);
493 struct tc_pie_xstats st = {
494 .prob = q->vars.prob,
495 .delay = ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
496 NSEC_PER_USEC,
497 /* unscale and return dq_rate in bytes per sec */
498 .avg_dq_rate = q->vars.avg_dq_rate *
499 (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
500 .packets_in = q->stats.packets_in,
501 .overlimit = q->stats.overlimit,
502 .maxq = q->stats.maxq,
503 .dropped = q->stats.dropped,
504 .ecn_mark = q->stats.ecn_mark,
505 };
506
507 return gnet_stats_copy_app(d, &st, sizeof(st));
508}
509
510static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
511{
512 struct sk_buff *skb;
513 skb = __qdisc_dequeue_head(sch, &sch->q);
514
515 if (!skb)
516 return NULL;
517
518 pie_process_dequeue(sch, skb);
519 return skb;
520}
521
522static void pie_reset(struct Qdisc *sch)
523{
524 struct pie_sched_data *q = qdisc_priv(sch);
525 qdisc_reset_queue(sch);
526 pie_vars_init(&q->vars);
527}
528
529static void pie_destroy(struct Qdisc *sch)
530{
531 struct pie_sched_data *q = qdisc_priv(sch);
532 q->params.tupdate = 0;
533 del_timer_sync(&q->adapt_timer);
534}
535
536static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
537 .id = "pie",
538 .priv_size = sizeof(struct pie_sched_data),
539 .enqueue = pie_qdisc_enqueue,
540 .dequeue = pie_qdisc_dequeue,
541 .peek = qdisc_peek_dequeued,
542 .init = pie_init,
543 .destroy = pie_destroy,
544 .reset = pie_reset,
545 .change = pie_change,
546 .dump = pie_dump,
547 .dump_stats = pie_dump_stats,
548 .owner = THIS_MODULE,
549};
550
551static int __init pie_module_init(void)
552{
553 return register_qdisc(&pie_qdisc_ops);
554}
555
556static void __exit pie_module_exit(void)
557{
558 unregister_qdisc(&pie_qdisc_ops);
559}
560
561module_init(pie_module_init);
562module_exit(pie_module_exit);
563
564MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
565MODULE_AUTHOR("Vijay Subramanian");
566MODULE_AUTHOR("Mythili Prabhu");
567MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (C) 2013 Cisco Systems, Inc, 2013.
3 *
4 * Author: Vijay Subramanian <vijaynsu@cisco.com>
5 * Author: Mythili Prabhu <mysuryan@cisco.com>
6 *
7 * ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
8 * University of Oslo, Norway.
9 *
10 * References:
11 * RFC 8033: https://tools.ietf.org/html/rfc8033
12 */
13
14#include <linux/module.h>
15#include <linux/slab.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/errno.h>
19#include <linux/skbuff.h>
20#include <net/pkt_sched.h>
21#include <net/inet_ecn.h>
22#include <net/pie.h>
23
24/* private data for the Qdisc */
25struct pie_sched_data {
26 struct pie_vars vars;
27 struct pie_params params;
28 struct pie_stats stats;
29 struct timer_list adapt_timer;
30 struct Qdisc *sch;
31};
32
33bool pie_drop_early(struct Qdisc *sch, struct pie_params *params,
34 struct pie_vars *vars, u32 backlog, u32 packet_size)
35{
36 u64 rnd;
37 u64 local_prob = vars->prob;
38 u32 mtu = psched_mtu(qdisc_dev(sch));
39
40 /* If there is still burst allowance left skip random early drop */
41 if (vars->burst_time > 0)
42 return false;
43
44 /* If current delay is less than half of target, and
45 * if drop prob is low already, disable early_drop
46 */
47 if ((vars->qdelay < params->target / 2) &&
48 (vars->prob < MAX_PROB / 5))
49 return false;
50
51 /* If we have fewer than 2 mtu-sized packets, disable pie_drop_early,
52 * similar to min_th in RED
53 */
54 if (backlog < 2 * mtu)
55 return false;
56
57 /* If bytemode is turned on, use packet size to compute new
58 * probablity. Smaller packets will have lower drop prob in this case
59 */
60 if (params->bytemode && packet_size <= mtu)
61 local_prob = (u64)packet_size * div_u64(local_prob, mtu);
62 else
63 local_prob = vars->prob;
64
65 if (local_prob == 0)
66 vars->accu_prob = 0;
67 else
68 vars->accu_prob += local_prob;
69
70 if (vars->accu_prob < (MAX_PROB / 100) * 85)
71 return false;
72 if (vars->accu_prob >= (MAX_PROB / 2) * 17)
73 return true;
74
75 get_random_bytes(&rnd, 8);
76 if ((rnd >> BITS_PER_BYTE) < local_prob) {
77 vars->accu_prob = 0;
78 return true;
79 }
80
81 return false;
82}
83EXPORT_SYMBOL_GPL(pie_drop_early);
84
85static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
86 struct sk_buff **to_free)
87{
88 struct pie_sched_data *q = qdisc_priv(sch);
89 bool enqueue = false;
90
91 if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
92 q->stats.overlimit++;
93 goto out;
94 }
95
96 if (!pie_drop_early(sch, &q->params, &q->vars, sch->qstats.backlog,
97 skb->len)) {
98 enqueue = true;
99 } else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
100 INET_ECN_set_ce(skb)) {
101 /* If packet is ecn capable, mark it if drop probability
102 * is lower than 10%, else drop it.
103 */
104 q->stats.ecn_mark++;
105 enqueue = true;
106 }
107
108 /* we can enqueue the packet */
109 if (enqueue) {
110 /* Set enqueue time only when dq_rate_estimator is disabled. */
111 if (!q->params.dq_rate_estimator)
112 pie_set_enqueue_time(skb);
113
114 q->stats.packets_in++;
115 if (qdisc_qlen(sch) > q->stats.maxq)
116 q->stats.maxq = qdisc_qlen(sch);
117
118 return qdisc_enqueue_tail(skb, sch);
119 }
120
121out:
122 q->stats.dropped++;
123 q->vars.accu_prob = 0;
124 return qdisc_drop(skb, sch, to_free);
125}
126
127static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
128 [TCA_PIE_TARGET] = {.type = NLA_U32},
129 [TCA_PIE_LIMIT] = {.type = NLA_U32},
130 [TCA_PIE_TUPDATE] = {.type = NLA_U32},
131 [TCA_PIE_ALPHA] = {.type = NLA_U32},
132 [TCA_PIE_BETA] = {.type = NLA_U32},
133 [TCA_PIE_ECN] = {.type = NLA_U32},
134 [TCA_PIE_BYTEMODE] = {.type = NLA_U32},
135 [TCA_PIE_DQ_RATE_ESTIMATOR] = {.type = NLA_U32},
136};
137
138static int pie_change(struct Qdisc *sch, struct nlattr *opt,
139 struct netlink_ext_ack *extack)
140{
141 struct pie_sched_data *q = qdisc_priv(sch);
142 struct nlattr *tb[TCA_PIE_MAX + 1];
143 unsigned int qlen, dropped = 0;
144 int err;
145
146 err = nla_parse_nested_deprecated(tb, TCA_PIE_MAX, opt, pie_policy,
147 NULL);
148 if (err < 0)
149 return err;
150
151 sch_tree_lock(sch);
152
153 /* convert from microseconds to pschedtime */
154 if (tb[TCA_PIE_TARGET]) {
155 /* target is in us */
156 u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
157
158 /* convert to pschedtime */
159 q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
160 }
161
162 /* tupdate is in jiffies */
163 if (tb[TCA_PIE_TUPDATE])
164 q->params.tupdate =
165 usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
166
167 if (tb[TCA_PIE_LIMIT]) {
168 u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
169
170 q->params.limit = limit;
171 sch->limit = limit;
172 }
173
174 if (tb[TCA_PIE_ALPHA])
175 q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
176
177 if (tb[TCA_PIE_BETA])
178 q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
179
180 if (tb[TCA_PIE_ECN])
181 q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
182
183 if (tb[TCA_PIE_BYTEMODE])
184 q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
185
186 if (tb[TCA_PIE_DQ_RATE_ESTIMATOR])
187 q->params.dq_rate_estimator =
188 nla_get_u32(tb[TCA_PIE_DQ_RATE_ESTIMATOR]);
189
190 /* Drop excess packets if new limit is lower */
191 qlen = sch->q.qlen;
192 while (sch->q.qlen > sch->limit) {
193 struct sk_buff *skb = __qdisc_dequeue_head(&sch->q);
194
195 dropped += qdisc_pkt_len(skb);
196 qdisc_qstats_backlog_dec(sch, skb);
197 rtnl_qdisc_drop(skb, sch);
198 }
199 qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
200
201 sch_tree_unlock(sch);
202 return 0;
203}
204
205void pie_process_dequeue(struct sk_buff *skb, struct pie_params *params,
206 struct pie_vars *vars, u32 backlog)
207{
208 psched_time_t now = psched_get_time();
209 u32 dtime = 0;
210
211 /* If dq_rate_estimator is disabled, calculate qdelay using the
212 * packet timestamp.
213 */
214 if (!params->dq_rate_estimator) {
215 vars->qdelay = now - pie_get_enqueue_time(skb);
216
217 if (vars->dq_tstamp != DTIME_INVALID)
218 dtime = now - vars->dq_tstamp;
219
220 vars->dq_tstamp = now;
221
222 if (backlog == 0)
223 vars->qdelay = 0;
224
225 if (dtime == 0)
226 return;
227
228 goto burst_allowance_reduction;
229 }
230
231 /* If current queue is about 10 packets or more and dq_count is unset
232 * we have enough packets to calculate the drain rate. Save
233 * current time as dq_tstamp and start measurement cycle.
234 */
235 if (backlog >= QUEUE_THRESHOLD && vars->dq_count == DQCOUNT_INVALID) {
236 vars->dq_tstamp = psched_get_time();
237 vars->dq_count = 0;
238 }
239
240 /* Calculate the average drain rate from this value. If queue length
241 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes, reset
242 * the dq_count to -1 as we don't have enough packets to calculate the
243 * drain rate anymore. The following if block is entered only when we
244 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
245 * and we calculate the drain rate for the threshold here. dq_count is
246 * in bytes, time difference in psched_time, hence rate is in
247 * bytes/psched_time.
248 */
249 if (vars->dq_count != DQCOUNT_INVALID) {
250 vars->dq_count += skb->len;
251
252 if (vars->dq_count >= QUEUE_THRESHOLD) {
253 u32 count = vars->dq_count << PIE_SCALE;
254
255 dtime = now - vars->dq_tstamp;
256
257 if (dtime == 0)
258 return;
259
260 count = count / dtime;
261
262 if (vars->avg_dq_rate == 0)
263 vars->avg_dq_rate = count;
264 else
265 vars->avg_dq_rate =
266 (vars->avg_dq_rate -
267 (vars->avg_dq_rate >> 3)) + (count >> 3);
268
269 /* If the queue has receded below the threshold, we hold
270 * on to the last drain rate calculated, else we reset
271 * dq_count to 0 to re-enter the if block when the next
272 * packet is dequeued
273 */
274 if (backlog < QUEUE_THRESHOLD) {
275 vars->dq_count = DQCOUNT_INVALID;
276 } else {
277 vars->dq_count = 0;
278 vars->dq_tstamp = psched_get_time();
279 }
280
281 goto burst_allowance_reduction;
282 }
283 }
284
285 return;
286
287burst_allowance_reduction:
288 if (vars->burst_time > 0) {
289 if (vars->burst_time > dtime)
290 vars->burst_time -= dtime;
291 else
292 vars->burst_time = 0;
293 }
294}
295EXPORT_SYMBOL_GPL(pie_process_dequeue);
296
297void pie_calculate_probability(struct pie_params *params, struct pie_vars *vars,
298 u32 backlog)
299{
300 psched_time_t qdelay = 0; /* in pschedtime */
301 psched_time_t qdelay_old = 0; /* in pschedtime */
302 s64 delta = 0; /* determines the change in probability */
303 u64 oldprob;
304 u64 alpha, beta;
305 u32 power;
306 bool update_prob = true;
307
308 if (params->dq_rate_estimator) {
309 qdelay_old = vars->qdelay;
310 vars->qdelay_old = vars->qdelay;
311
312 if (vars->avg_dq_rate > 0)
313 qdelay = (backlog << PIE_SCALE) / vars->avg_dq_rate;
314 else
315 qdelay = 0;
316 } else {
317 qdelay = vars->qdelay;
318 qdelay_old = vars->qdelay_old;
319 }
320
321 /* If qdelay is zero and backlog is not, it means backlog is very small,
322 * so we do not update probabilty in this round.
323 */
324 if (qdelay == 0 && backlog != 0)
325 update_prob = false;
326
327 /* In the algorithm, alpha and beta are between 0 and 2 with typical
328 * value for alpha as 0.125. In this implementation, we use values 0-32
329 * passed from user space to represent this. Also, alpha and beta have
330 * unit of HZ and need to be scaled before they can used to update
331 * probability. alpha/beta are updated locally below by scaling down
332 * by 16 to come to 0-2 range.
333 */
334 alpha = ((u64)params->alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
335 beta = ((u64)params->beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
336
337 /* We scale alpha and beta differently depending on how heavy the
338 * congestion is. Please see RFC 8033 for details.
339 */
340 if (vars->prob < MAX_PROB / 10) {
341 alpha >>= 1;
342 beta >>= 1;
343
344 power = 100;
345 while (vars->prob < div_u64(MAX_PROB, power) &&
346 power <= 1000000) {
347 alpha >>= 2;
348 beta >>= 2;
349 power *= 10;
350 }
351 }
352
353 /* alpha and beta should be between 0 and 32, in multiples of 1/16 */
354 delta += alpha * (qdelay - params->target);
355 delta += beta * (qdelay - qdelay_old);
356
357 oldprob = vars->prob;
358
359 /* to ensure we increase probability in steps of no more than 2% */
360 if (delta > (s64)(MAX_PROB / (100 / 2)) &&
361 vars->prob >= MAX_PROB / 10)
362 delta = (MAX_PROB / 100) * 2;
363
364 /* Non-linear drop:
365 * Tune drop probability to increase quickly for high delays(>= 250ms)
366 * 250ms is derived through experiments and provides error protection
367 */
368
369 if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
370 delta += MAX_PROB / (100 / 2);
371
372 vars->prob += delta;
373
374 if (delta > 0) {
375 /* prevent overflow */
376 if (vars->prob < oldprob) {
377 vars->prob = MAX_PROB;
378 /* Prevent normalization error. If probability is at
379 * maximum value already, we normalize it here, and
380 * skip the check to do a non-linear drop in the next
381 * section.
382 */
383 update_prob = false;
384 }
385 } else {
386 /* prevent underflow */
387 if (vars->prob > oldprob)
388 vars->prob = 0;
389 }
390
391 /* Non-linear drop in probability: Reduce drop probability quickly if
392 * delay is 0 for 2 consecutive Tupdate periods.
393 */
394
395 if (qdelay == 0 && qdelay_old == 0 && update_prob)
396 /* Reduce drop probability to 98.4% */
397 vars->prob -= vars->prob / 64;
398
399 vars->qdelay = qdelay;
400 vars->backlog_old = backlog;
401
402 /* We restart the measurement cycle if the following conditions are met
403 * 1. If the delay has been low for 2 consecutive Tupdate periods
404 * 2. Calculated drop probability is zero
405 * 3. If average dq_rate_estimator is enabled, we have at least one
406 * estimate for the avg_dq_rate ie., is a non-zero value
407 */
408 if ((vars->qdelay < params->target / 2) &&
409 (vars->qdelay_old < params->target / 2) &&
410 vars->prob == 0 &&
411 (!params->dq_rate_estimator || vars->avg_dq_rate > 0)) {
412 pie_vars_init(vars);
413 }
414
415 if (!params->dq_rate_estimator)
416 vars->qdelay_old = qdelay;
417}
418EXPORT_SYMBOL_GPL(pie_calculate_probability);
419
420static void pie_timer(struct timer_list *t)
421{
422 struct pie_sched_data *q = from_timer(q, t, adapt_timer);
423 struct Qdisc *sch = q->sch;
424 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
425
426 spin_lock(root_lock);
427 pie_calculate_probability(&q->params, &q->vars, sch->qstats.backlog);
428
429 /* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
430 if (q->params.tupdate)
431 mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
432 spin_unlock(root_lock);
433}
434
435static int pie_init(struct Qdisc *sch, struct nlattr *opt,
436 struct netlink_ext_ack *extack)
437{
438 struct pie_sched_data *q = qdisc_priv(sch);
439
440 pie_params_init(&q->params);
441 pie_vars_init(&q->vars);
442 sch->limit = q->params.limit;
443
444 q->sch = sch;
445 timer_setup(&q->adapt_timer, pie_timer, 0);
446
447 if (opt) {
448 int err = pie_change(sch, opt, extack);
449
450 if (err)
451 return err;
452 }
453
454 mod_timer(&q->adapt_timer, jiffies + HZ / 2);
455 return 0;
456}
457
458static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
459{
460 struct pie_sched_data *q = qdisc_priv(sch);
461 struct nlattr *opts;
462
463 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
464 if (!opts)
465 goto nla_put_failure;
466
467 /* convert target from pschedtime to us */
468 if (nla_put_u32(skb, TCA_PIE_TARGET,
469 ((u32)PSCHED_TICKS2NS(q->params.target)) /
470 NSEC_PER_USEC) ||
471 nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
472 nla_put_u32(skb, TCA_PIE_TUPDATE,
473 jiffies_to_usecs(q->params.tupdate)) ||
474 nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
475 nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
476 nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
477 nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode) ||
478 nla_put_u32(skb, TCA_PIE_DQ_RATE_ESTIMATOR,
479 q->params.dq_rate_estimator))
480 goto nla_put_failure;
481
482 return nla_nest_end(skb, opts);
483
484nla_put_failure:
485 nla_nest_cancel(skb, opts);
486 return -1;
487}
488
489static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
490{
491 struct pie_sched_data *q = qdisc_priv(sch);
492 struct tc_pie_xstats st = {
493 .prob = q->vars.prob << BITS_PER_BYTE,
494 .delay = ((u32)PSCHED_TICKS2NS(q->vars.qdelay)) /
495 NSEC_PER_USEC,
496 .packets_in = q->stats.packets_in,
497 .overlimit = q->stats.overlimit,
498 .maxq = q->stats.maxq,
499 .dropped = q->stats.dropped,
500 .ecn_mark = q->stats.ecn_mark,
501 };
502
503 /* avg_dq_rate is only valid if dq_rate_estimator is enabled */
504 st.dq_rate_estimating = q->params.dq_rate_estimator;
505
506 /* unscale and return dq_rate in bytes per sec */
507 if (q->params.dq_rate_estimator)
508 st.avg_dq_rate = q->vars.avg_dq_rate *
509 (PSCHED_TICKS_PER_SEC) >> PIE_SCALE;
510
511 return gnet_stats_copy_app(d, &st, sizeof(st));
512}
513
514static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
515{
516 struct pie_sched_data *q = qdisc_priv(sch);
517 struct sk_buff *skb = qdisc_dequeue_head(sch);
518
519 if (!skb)
520 return NULL;
521
522 pie_process_dequeue(skb, &q->params, &q->vars, sch->qstats.backlog);
523 return skb;
524}
525
526static void pie_reset(struct Qdisc *sch)
527{
528 struct pie_sched_data *q = qdisc_priv(sch);
529
530 qdisc_reset_queue(sch);
531 pie_vars_init(&q->vars);
532}
533
534static void pie_destroy(struct Qdisc *sch)
535{
536 struct pie_sched_data *q = qdisc_priv(sch);
537
538 q->params.tupdate = 0;
539 del_timer_sync(&q->adapt_timer);
540}
541
542static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
543 .id = "pie",
544 .priv_size = sizeof(struct pie_sched_data),
545 .enqueue = pie_qdisc_enqueue,
546 .dequeue = pie_qdisc_dequeue,
547 .peek = qdisc_peek_dequeued,
548 .init = pie_init,
549 .destroy = pie_destroy,
550 .reset = pie_reset,
551 .change = pie_change,
552 .dump = pie_dump,
553 .dump_stats = pie_dump_stats,
554 .owner = THIS_MODULE,
555};
556
557static int __init pie_module_init(void)
558{
559 return register_qdisc(&pie_qdisc_ops);
560}
561
562static void __exit pie_module_exit(void)
563{
564 unregister_qdisc(&pie_qdisc_ops);
565}
566
567module_init(pie_module_init);
568module_exit(pie_module_exit);
569
570MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
571MODULE_AUTHOR("Vijay Subramanian");
572MODULE_AUTHOR("Mythili Prabhu");
573MODULE_LICENSE("GPL");