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 */
25 struct 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
pie_drop_early(struct Qdisc * sch,struct pie_params * params,struct pie_vars * vars,u32 backlog,u32 packet_size)33 bool 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 }
83 EXPORT_SYMBOL_GPL(pie_drop_early);
84
pie_qdisc_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)85 static 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
121 out:
122 q->stats.dropped++;
123 q->vars.accu_prob = 0;
124 return qdisc_drop(skb, sch, to_free);
125 }
126
127 static 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
pie_change(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)138 static 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 WRITE_ONCE(q->params.target,
160 PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC));
161 }
162
163 /* tupdate is in jiffies */
164 if (tb[TCA_PIE_TUPDATE])
165 WRITE_ONCE(q->params.tupdate,
166 usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE])));
167
168 if (tb[TCA_PIE_LIMIT]) {
169 u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
170
171 WRITE_ONCE(q->params.limit, limit);
172 WRITE_ONCE(sch->limit, limit);
173 }
174
175 if (tb[TCA_PIE_ALPHA])
176 WRITE_ONCE(q->params.alpha, nla_get_u32(tb[TCA_PIE_ALPHA]));
177
178 if (tb[TCA_PIE_BETA])
179 WRITE_ONCE(q->params.beta, nla_get_u32(tb[TCA_PIE_BETA]));
180
181 if (tb[TCA_PIE_ECN])
182 WRITE_ONCE(q->params.ecn, nla_get_u32(tb[TCA_PIE_ECN]));
183
184 if (tb[TCA_PIE_BYTEMODE])
185 WRITE_ONCE(q->params.bytemode,
186 nla_get_u32(tb[TCA_PIE_BYTEMODE]));
187
188 if (tb[TCA_PIE_DQ_RATE_ESTIMATOR])
189 WRITE_ONCE(q->params.dq_rate_estimator,
190 nla_get_u32(tb[TCA_PIE_DQ_RATE_ESTIMATOR]));
191
192 /* Drop excess packets if new limit is lower */
193 qlen = sch->q.qlen;
194 while (sch->q.qlen > sch->limit) {
195 struct sk_buff *skb = __qdisc_dequeue_head(&sch->q);
196
197 dropped += qdisc_pkt_len(skb);
198 qdisc_qstats_backlog_dec(sch, skb);
199 rtnl_qdisc_drop(skb, sch);
200 }
201 qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped);
202
203 sch_tree_unlock(sch);
204 return 0;
205 }
206
pie_process_dequeue(struct sk_buff * skb,struct pie_params * params,struct pie_vars * vars,u32 backlog)207 void pie_process_dequeue(struct sk_buff *skb, struct pie_params *params,
208 struct pie_vars *vars, u32 backlog)
209 {
210 psched_time_t now = psched_get_time();
211 u32 dtime = 0;
212
213 /* If dq_rate_estimator is disabled, calculate qdelay using the
214 * packet timestamp.
215 */
216 if (!params->dq_rate_estimator) {
217 vars->qdelay = now - pie_get_enqueue_time(skb);
218
219 if (vars->dq_tstamp != DTIME_INVALID)
220 dtime = now - vars->dq_tstamp;
221
222 vars->dq_tstamp = now;
223
224 if (backlog == 0)
225 vars->qdelay = 0;
226
227 if (dtime == 0)
228 return;
229
230 goto burst_allowance_reduction;
231 }
232
233 /* If current queue is about 10 packets or more and dq_count is unset
234 * we have enough packets to calculate the drain rate. Save
235 * current time as dq_tstamp and start measurement cycle.
236 */
237 if (backlog >= QUEUE_THRESHOLD && vars->dq_count == DQCOUNT_INVALID) {
238 vars->dq_tstamp = psched_get_time();
239 vars->dq_count = 0;
240 }
241
242 /* Calculate the average drain rate from this value. If queue length
243 * has receded to a small value viz., <= QUEUE_THRESHOLD bytes, reset
244 * the dq_count to -1 as we don't have enough packets to calculate the
245 * drain rate anymore. The following if block is entered only when we
246 * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
247 * and we calculate the drain rate for the threshold here. dq_count is
248 * in bytes, time difference in psched_time, hence rate is in
249 * bytes/psched_time.
250 */
251 if (vars->dq_count != DQCOUNT_INVALID) {
252 vars->dq_count += skb->len;
253
254 if (vars->dq_count >= QUEUE_THRESHOLD) {
255 u32 count = vars->dq_count << PIE_SCALE;
256
257 dtime = now - vars->dq_tstamp;
258
259 if (dtime == 0)
260 return;
261
262 count = count / dtime;
263
264 if (vars->avg_dq_rate == 0)
265 vars->avg_dq_rate = count;
266 else
267 vars->avg_dq_rate =
268 (vars->avg_dq_rate -
269 (vars->avg_dq_rate >> 3)) + (count >> 3);
270
271 /* If the queue has receded below the threshold, we hold
272 * on to the last drain rate calculated, else we reset
273 * dq_count to 0 to re-enter the if block when the next
274 * packet is dequeued
275 */
276 if (backlog < QUEUE_THRESHOLD) {
277 vars->dq_count = DQCOUNT_INVALID;
278 } else {
279 vars->dq_count = 0;
280 vars->dq_tstamp = psched_get_time();
281 }
282
283 goto burst_allowance_reduction;
284 }
285 }
286
287 return;
288
289 burst_allowance_reduction:
290 if (vars->burst_time > 0) {
291 if (vars->burst_time > dtime)
292 vars->burst_time -= dtime;
293 else
294 vars->burst_time = 0;
295 }
296 }
297 EXPORT_SYMBOL_GPL(pie_process_dequeue);
298
pie_calculate_probability(struct pie_params * params,struct pie_vars * vars,u32 backlog)299 void pie_calculate_probability(struct pie_params *params, struct pie_vars *vars,
300 u32 backlog)
301 {
302 psched_time_t qdelay = 0; /* in pschedtime */
303 psched_time_t qdelay_old = 0; /* in pschedtime */
304 s64 delta = 0; /* determines the change in probability */
305 u64 oldprob;
306 u64 alpha, beta;
307 u32 power;
308 bool update_prob = true;
309
310 if (params->dq_rate_estimator) {
311 qdelay_old = vars->qdelay;
312 vars->qdelay_old = vars->qdelay;
313
314 if (vars->avg_dq_rate > 0)
315 qdelay = (backlog << PIE_SCALE) / vars->avg_dq_rate;
316 else
317 qdelay = 0;
318 } else {
319 qdelay = vars->qdelay;
320 qdelay_old = vars->qdelay_old;
321 }
322
323 /* If qdelay is zero and backlog is not, it means backlog is very small,
324 * so we do not update probability in this round.
325 */
326 if (qdelay == 0 && backlog != 0)
327 update_prob = false;
328
329 /* In the algorithm, alpha and beta are between 0 and 2 with typical
330 * value for alpha as 0.125. In this implementation, we use values 0-32
331 * passed from user space to represent this. Also, alpha and beta have
332 * unit of HZ and need to be scaled before they can used to update
333 * probability. alpha/beta are updated locally below by scaling down
334 * by 16 to come to 0-2 range.
335 */
336 alpha = ((u64)params->alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
337 beta = ((u64)params->beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
338
339 /* We scale alpha and beta differently depending on how heavy the
340 * congestion is. Please see RFC 8033 for details.
341 */
342 if (vars->prob < MAX_PROB / 10) {
343 alpha >>= 1;
344 beta >>= 1;
345
346 power = 100;
347 while (vars->prob < div_u64(MAX_PROB, power) &&
348 power <= 1000000) {
349 alpha >>= 2;
350 beta >>= 2;
351 power *= 10;
352 }
353 }
354
355 /* alpha and beta should be between 0 and 32, in multiples of 1/16 */
356 delta += alpha * (qdelay - params->target);
357 delta += beta * (qdelay - qdelay_old);
358
359 oldprob = vars->prob;
360
361 /* to ensure we increase probability in steps of no more than 2% */
362 if (delta > (s64)(MAX_PROB / (100 / 2)) &&
363 vars->prob >= MAX_PROB / 10)
364 delta = (MAX_PROB / 100) * 2;
365
366 /* Non-linear drop:
367 * Tune drop probability to increase quickly for high delays(>= 250ms)
368 * 250ms is derived through experiments and provides error protection
369 */
370
371 if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
372 delta += MAX_PROB / (100 / 2);
373
374 vars->prob += delta;
375
376 if (delta > 0) {
377 /* prevent overflow */
378 if (vars->prob < oldprob) {
379 vars->prob = MAX_PROB;
380 /* Prevent normalization error. If probability is at
381 * maximum value already, we normalize it here, and
382 * skip the check to do a non-linear drop in the next
383 * section.
384 */
385 update_prob = false;
386 }
387 } else {
388 /* prevent underflow */
389 if (vars->prob > oldprob)
390 vars->prob = 0;
391 }
392
393 /* Non-linear drop in probability: Reduce drop probability quickly if
394 * delay is 0 for 2 consecutive Tupdate periods.
395 */
396
397 if (qdelay == 0 && qdelay_old == 0 && update_prob)
398 /* Reduce drop probability to 98.4% */
399 vars->prob -= vars->prob / 64;
400
401 vars->qdelay = qdelay;
402 vars->backlog_old = backlog;
403
404 /* We restart the measurement cycle if the following conditions are met
405 * 1. If the delay has been low for 2 consecutive Tupdate periods
406 * 2. Calculated drop probability is zero
407 * 3. If average dq_rate_estimator is enabled, we have at least one
408 * estimate for the avg_dq_rate ie., is a non-zero value
409 */
410 if ((vars->qdelay < params->target / 2) &&
411 (vars->qdelay_old < params->target / 2) &&
412 vars->prob == 0 &&
413 (!params->dq_rate_estimator || vars->avg_dq_rate > 0)) {
414 pie_vars_init(vars);
415 }
416
417 if (!params->dq_rate_estimator)
418 vars->qdelay_old = qdelay;
419 }
420 EXPORT_SYMBOL_GPL(pie_calculate_probability);
421
pie_timer(struct timer_list * t)422 static void pie_timer(struct timer_list *t)
423 {
424 struct pie_sched_data *q = from_timer(q, t, adapt_timer);
425 struct Qdisc *sch = q->sch;
426 spinlock_t *root_lock;
427
428 rcu_read_lock();
429 root_lock = qdisc_lock(qdisc_root_sleeping(sch));
430 spin_lock(root_lock);
431 pie_calculate_probability(&q->params, &q->vars, sch->qstats.backlog);
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 rcu_read_unlock();
438 }
439
pie_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)440 static int pie_init(struct Qdisc *sch, struct nlattr *opt,
441 struct netlink_ext_ack *extack)
442 {
443 struct pie_sched_data *q = qdisc_priv(sch);
444
445 pie_params_init(&q->params);
446 pie_vars_init(&q->vars);
447 sch->limit = q->params.limit;
448
449 q->sch = sch;
450 timer_setup(&q->adapt_timer, pie_timer, 0);
451
452 if (opt) {
453 int err = pie_change(sch, opt, extack);
454
455 if (err)
456 return err;
457 }
458
459 mod_timer(&q->adapt_timer, jiffies + HZ / 2);
460 return 0;
461 }
462
pie_dump(struct Qdisc * sch,struct sk_buff * skb)463 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
464 {
465 struct pie_sched_data *q = qdisc_priv(sch);
466 struct nlattr *opts;
467
468 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
469 if (!opts)
470 goto nla_put_failure;
471
472 /* convert target from pschedtime to us */
473 if (nla_put_u32(skb, TCA_PIE_TARGET,
474 ((u32)PSCHED_TICKS2NS(READ_ONCE(q->params.target))) /
475 NSEC_PER_USEC) ||
476 nla_put_u32(skb, TCA_PIE_LIMIT, READ_ONCE(sch->limit)) ||
477 nla_put_u32(skb, TCA_PIE_TUPDATE,
478 jiffies_to_usecs(READ_ONCE(q->params.tupdate))) ||
479 nla_put_u32(skb, TCA_PIE_ALPHA, READ_ONCE(q->params.alpha)) ||
480 nla_put_u32(skb, TCA_PIE_BETA, READ_ONCE(q->params.beta)) ||
481 nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
482 nla_put_u32(skb, TCA_PIE_BYTEMODE,
483 READ_ONCE(q->params.bytemode)) ||
484 nla_put_u32(skb, TCA_PIE_DQ_RATE_ESTIMATOR,
485 READ_ONCE(q->params.dq_rate_estimator)))
486 goto nla_put_failure;
487
488 return nla_nest_end(skb, opts);
489
490 nla_put_failure:
491 nla_nest_cancel(skb, opts);
492 return -1;
493 }
494
pie_dump_stats(struct Qdisc * sch,struct gnet_dump * d)495 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
496 {
497 struct pie_sched_data *q = qdisc_priv(sch);
498 struct tc_pie_xstats st = {
499 .prob = q->vars.prob << BITS_PER_BYTE,
500 .delay = ((u32)PSCHED_TICKS2NS(q->vars.qdelay)) /
501 NSEC_PER_USEC,
502 .packets_in = q->stats.packets_in,
503 .overlimit = q->stats.overlimit,
504 .maxq = q->stats.maxq,
505 .dropped = q->stats.dropped,
506 .ecn_mark = q->stats.ecn_mark,
507 };
508
509 /* avg_dq_rate is only valid if dq_rate_estimator is enabled */
510 st.dq_rate_estimating = q->params.dq_rate_estimator;
511
512 /* unscale and return dq_rate in bytes per sec */
513 if (q->params.dq_rate_estimator)
514 st.avg_dq_rate = q->vars.avg_dq_rate *
515 (PSCHED_TICKS_PER_SEC) >> PIE_SCALE;
516
517 return gnet_stats_copy_app(d, &st, sizeof(st));
518 }
519
pie_qdisc_dequeue(struct Qdisc * sch)520 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
521 {
522 struct pie_sched_data *q = qdisc_priv(sch);
523 struct sk_buff *skb = qdisc_dequeue_head(sch);
524
525 if (!skb)
526 return NULL;
527
528 pie_process_dequeue(skb, &q->params, &q->vars, sch->qstats.backlog);
529 return skb;
530 }
531
pie_reset(struct Qdisc * sch)532 static void pie_reset(struct Qdisc *sch)
533 {
534 struct pie_sched_data *q = qdisc_priv(sch);
535
536 qdisc_reset_queue(sch);
537 pie_vars_init(&q->vars);
538 }
539
pie_destroy(struct Qdisc * sch)540 static void pie_destroy(struct Qdisc *sch)
541 {
542 struct pie_sched_data *q = qdisc_priv(sch);
543
544 q->params.tupdate = 0;
545 del_timer_sync(&q->adapt_timer);
546 }
547
548 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
549 .id = "pie",
550 .priv_size = sizeof(struct pie_sched_data),
551 .enqueue = pie_qdisc_enqueue,
552 .dequeue = pie_qdisc_dequeue,
553 .peek = qdisc_peek_dequeued,
554 .init = pie_init,
555 .destroy = pie_destroy,
556 .reset = pie_reset,
557 .change = pie_change,
558 .dump = pie_dump,
559 .dump_stats = pie_dump_stats,
560 .owner = THIS_MODULE,
561 };
562 MODULE_ALIAS_NET_SCH("pie");
563
pie_module_init(void)564 static int __init pie_module_init(void)
565 {
566 return register_qdisc(&pie_qdisc_ops);
567 }
568
pie_module_exit(void)569 static void __exit pie_module_exit(void)
570 {
571 unregister_qdisc(&pie_qdisc_ops);
572 }
573
574 module_init(pie_module_init);
575 module_exit(pie_module_exit);
576
577 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
578 MODULE_AUTHOR("Vijay Subramanian");
579 MODULE_AUTHOR("Mythili Prabhu");
580 MODULE_LICENSE("GPL");
581