xref: /linux/net/sched/sch_fq_pie.c (revision 5e2cb28dd7e182dfa641550dfa225913509ad45d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Flow Queue PIE discipline
3  *
4  * Copyright (C) 2019 Mohit P. Tahiliani <tahiliani@nitk.edu.in>
5  * Copyright (C) 2019 Sachin D. Patil <sdp.sachin@gmail.com>
6  * Copyright (C) 2019 V. Saicharan <vsaicharan1998@gmail.com>
7  * Copyright (C) 2019 Mohit Bhasi <mohitbhasi1998@gmail.com>
8  * Copyright (C) 2019 Leslie Monis <lesliemonis@gmail.com>
9  * Copyright (C) 2019 Gautam Ramakrishnan <gautamramk@gmail.com>
10  */
11 
12 #include <linux/jhash.h>
13 #include <linux/sizes.h>
14 #include <linux/vmalloc.h>
15 #include <net/pkt_cls.h>
16 #include <net/pie.h>
17 
18 /* Flow Queue PIE
19  *
20  * Principles:
21  *   - Packets are classified on flows.
22  *   - This is a Stochastic model (as we use a hash, several flows might
23  *                                 be hashed to the same slot)
24  *   - Each flow has a PIE managed queue.
25  *   - Flows are linked onto two (Round Robin) lists,
26  *     so that new flows have priority on old ones.
27  *   - For a given flow, packets are not reordered.
28  *   - Drops during enqueue only.
29  *   - ECN capability is off by default.
30  *   - ECN threshold (if ECN is enabled) is at 10% by default.
31  *   - Uses timestamps to calculate queue delay by default.
32  */
33 
34 /**
35  * struct fq_pie_flow - contains data for each flow
36  * @vars:	pie vars associated with the flow
37  * @deficit:	number of remaining byte credits
38  * @backlog:	size of data in the flow
39  * @qlen:	number of packets in the flow
40  * @flowchain:	flowchain for the flow
41  * @head:	first packet in the flow
42  * @tail:	last packet in the flow
43  */
44 struct fq_pie_flow {
45 	struct pie_vars vars;
46 	s32 deficit;
47 	u32 backlog;
48 	u32 qlen;
49 	struct list_head flowchain;
50 	struct sk_buff *head;
51 	struct sk_buff *tail;
52 };
53 
54 struct fq_pie_sched_data {
55 	struct tcf_proto __rcu *filter_list; /* optional external classifier */
56 	struct tcf_block *block;
57 	struct fq_pie_flow *flows;
58 	struct Qdisc *sch;
59 	struct list_head old_flows;
60 	struct list_head new_flows;
61 	struct pie_params p_params;
62 	u32 ecn_prob;
63 	u32 flows_cnt;
64 	u32 flows_cursor;
65 	u32 quantum;
66 	u32 memory_limit;
67 	u32 new_flow_count;
68 	u32 memory_usage;
69 	u32 overmemory;
70 	struct pie_stats stats;
71 	struct timer_list adapt_timer;
72 };
73 
74 static unsigned int fq_pie_hash(const struct fq_pie_sched_data *q,
75 				struct sk_buff *skb)
76 {
77 	return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
78 }
79 
80 static unsigned int fq_pie_classify(struct sk_buff *skb, struct Qdisc *sch,
81 				    int *qerr)
82 {
83 	struct fq_pie_sched_data *q = qdisc_priv(sch);
84 	struct tcf_proto *filter;
85 	struct tcf_result res;
86 	int result;
87 
88 	if (TC_H_MAJ(skb->priority) == sch->handle &&
89 	    TC_H_MIN(skb->priority) > 0 &&
90 	    TC_H_MIN(skb->priority) <= q->flows_cnt)
91 		return TC_H_MIN(skb->priority);
92 
93 	filter = rcu_dereference_bh(q->filter_list);
94 	if (!filter)
95 		return fq_pie_hash(q, skb) + 1;
96 
97 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
98 	result = tcf_classify(skb, NULL, filter, &res, false);
99 	if (result >= 0) {
100 #ifdef CONFIG_NET_CLS_ACT
101 		switch (result) {
102 		case TC_ACT_STOLEN:
103 		case TC_ACT_QUEUED:
104 		case TC_ACT_TRAP:
105 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
106 			fallthrough;
107 		case TC_ACT_SHOT:
108 			return 0;
109 		}
110 #endif
111 		if (TC_H_MIN(res.classid) <= q->flows_cnt)
112 			return TC_H_MIN(res.classid);
113 	}
114 	return 0;
115 }
116 
117 /* add skb to flow queue (tail add) */
118 static inline void flow_queue_add(struct fq_pie_flow *flow,
119 				  struct sk_buff *skb)
120 {
121 	if (!flow->head)
122 		flow->head = skb;
123 	else
124 		flow->tail->next = skb;
125 	flow->tail = skb;
126 	skb->next = NULL;
127 }
128 
129 static int fq_pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
130 				struct sk_buff **to_free)
131 {
132 	struct fq_pie_sched_data *q = qdisc_priv(sch);
133 	struct fq_pie_flow *sel_flow;
134 	int ret;
135 	u8 memory_limited = false;
136 	u8 enqueue = false;
137 	u32 pkt_len;
138 	u32 idx;
139 
140 	/* Classifies packet into corresponding flow */
141 	idx = fq_pie_classify(skb, sch, &ret);
142 	if (idx == 0) {
143 		if (ret & __NET_XMIT_BYPASS)
144 			qdisc_qstats_drop(sch);
145 		__qdisc_drop(skb, to_free);
146 		return ret;
147 	}
148 	idx--;
149 
150 	sel_flow = &q->flows[idx];
151 	/* Checks whether adding a new packet would exceed memory limit */
152 	get_pie_cb(skb)->mem_usage = skb->truesize;
153 	memory_limited = q->memory_usage > q->memory_limit + skb->truesize;
154 
155 	/* Checks if the qdisc is full */
156 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
157 		q->stats.overlimit++;
158 		goto out;
159 	} else if (unlikely(memory_limited)) {
160 		q->overmemory++;
161 	}
162 
163 	if (!pie_drop_early(sch, &q->p_params, &sel_flow->vars,
164 			    sel_flow->backlog, skb->len)) {
165 		enqueue = true;
166 	} else if (q->p_params.ecn &&
167 		   sel_flow->vars.prob <= (MAX_PROB / 100) * q->ecn_prob &&
168 		   INET_ECN_set_ce(skb)) {
169 		/* If packet is ecn capable, mark it if drop probability
170 		 * is lower than the parameter ecn_prob, else drop it.
171 		 */
172 		q->stats.ecn_mark++;
173 		enqueue = true;
174 	}
175 	if (enqueue) {
176 		/* Set enqueue time only when dq_rate_estimator is disabled. */
177 		if (!q->p_params.dq_rate_estimator)
178 			pie_set_enqueue_time(skb);
179 
180 		pkt_len = qdisc_pkt_len(skb);
181 		q->stats.packets_in++;
182 		q->memory_usage += skb->truesize;
183 		sch->qstats.backlog += pkt_len;
184 		sch->q.qlen++;
185 		flow_queue_add(sel_flow, skb);
186 		if (list_empty(&sel_flow->flowchain)) {
187 			list_add_tail(&sel_flow->flowchain, &q->new_flows);
188 			q->new_flow_count++;
189 			sel_flow->deficit = q->quantum;
190 			sel_flow->qlen = 0;
191 			sel_flow->backlog = 0;
192 		}
193 		sel_flow->qlen++;
194 		sel_flow->backlog += pkt_len;
195 		return NET_XMIT_SUCCESS;
196 	}
197 out:
198 	q->stats.dropped++;
199 	sel_flow->vars.accu_prob = 0;
200 	__qdisc_drop(skb, to_free);
201 	qdisc_qstats_drop(sch);
202 	return NET_XMIT_CN;
203 }
204 
205 static const struct netlink_range_validation fq_pie_q_range = {
206 	.min = 1,
207 	.max = 1 << 20,
208 };
209 
210 static const struct nla_policy fq_pie_policy[TCA_FQ_PIE_MAX + 1] = {
211 	[TCA_FQ_PIE_LIMIT]		= {.type = NLA_U32},
212 	[TCA_FQ_PIE_FLOWS]		= {.type = NLA_U32},
213 	[TCA_FQ_PIE_TARGET]		= {.type = NLA_U32},
214 	[TCA_FQ_PIE_TUPDATE]		= {.type = NLA_U32},
215 	[TCA_FQ_PIE_ALPHA]		= {.type = NLA_U32},
216 	[TCA_FQ_PIE_BETA]		= {.type = NLA_U32},
217 	[TCA_FQ_PIE_QUANTUM]		=
218 			NLA_POLICY_FULL_RANGE(NLA_U32, &fq_pie_q_range),
219 	[TCA_FQ_PIE_MEMORY_LIMIT]	= {.type = NLA_U32},
220 	[TCA_FQ_PIE_ECN_PROB]		= {.type = NLA_U32},
221 	[TCA_FQ_PIE_ECN]		= {.type = NLA_U32},
222 	[TCA_FQ_PIE_BYTEMODE]		= {.type = NLA_U32},
223 	[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]	= {.type = NLA_U32},
224 };
225 
226 static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
227 {
228 	struct sk_buff *skb = flow->head;
229 
230 	flow->head = skb->next;
231 	skb->next = NULL;
232 	return skb;
233 }
234 
235 static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
236 {
237 	struct fq_pie_sched_data *q = qdisc_priv(sch);
238 	struct sk_buff *skb = NULL;
239 	struct fq_pie_flow *flow;
240 	struct list_head *head;
241 	u32 pkt_len;
242 
243 begin:
244 	head = &q->new_flows;
245 	if (list_empty(head)) {
246 		head = &q->old_flows;
247 		if (list_empty(head))
248 			return NULL;
249 	}
250 
251 	flow = list_first_entry(head, struct fq_pie_flow, flowchain);
252 	/* Flow has exhausted all its credits */
253 	if (flow->deficit <= 0) {
254 		flow->deficit += q->quantum;
255 		list_move_tail(&flow->flowchain, &q->old_flows);
256 		goto begin;
257 	}
258 
259 	if (flow->head) {
260 		skb = dequeue_head(flow);
261 		pkt_len = qdisc_pkt_len(skb);
262 		sch->qstats.backlog -= pkt_len;
263 		sch->q.qlen--;
264 		qdisc_bstats_update(sch, skb);
265 	}
266 
267 	if (!skb) {
268 		/* force a pass through old_flows to prevent starvation */
269 		if (head == &q->new_flows && !list_empty(&q->old_flows))
270 			list_move_tail(&flow->flowchain, &q->old_flows);
271 		else
272 			list_del_init(&flow->flowchain);
273 		goto begin;
274 	}
275 
276 	flow->qlen--;
277 	flow->deficit -= pkt_len;
278 	flow->backlog -= pkt_len;
279 	q->memory_usage -= get_pie_cb(skb)->mem_usage;
280 	pie_process_dequeue(skb, &q->p_params, &flow->vars, flow->backlog);
281 	return skb;
282 }
283 
284 static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
285 			 struct netlink_ext_ack *extack)
286 {
287 	struct fq_pie_sched_data *q = qdisc_priv(sch);
288 	struct nlattr *tb[TCA_FQ_PIE_MAX + 1];
289 	unsigned int len_dropped = 0;
290 	unsigned int num_dropped = 0;
291 	int err;
292 
293 	err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, opt, fq_pie_policy, extack);
294 	if (err < 0)
295 		return err;
296 
297 	sch_tree_lock(sch);
298 	if (tb[TCA_FQ_PIE_LIMIT]) {
299 		u32 limit = nla_get_u32(tb[TCA_FQ_PIE_LIMIT]);
300 
301 		q->p_params.limit = limit;
302 		sch->limit = limit;
303 	}
304 	if (tb[TCA_FQ_PIE_FLOWS]) {
305 		if (q->flows) {
306 			NL_SET_ERR_MSG_MOD(extack,
307 					   "Number of flows cannot be changed");
308 			goto flow_error;
309 		}
310 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_PIE_FLOWS]);
311 		if (!q->flows_cnt || q->flows_cnt > 65536) {
312 			NL_SET_ERR_MSG_MOD(extack,
313 					   "Number of flows must range in [1..65536]");
314 			goto flow_error;
315 		}
316 	}
317 
318 	/* convert from microseconds to pschedtime */
319 	if (tb[TCA_FQ_PIE_TARGET]) {
320 		/* target is in us */
321 		u32 target = nla_get_u32(tb[TCA_FQ_PIE_TARGET]);
322 
323 		/* convert to pschedtime */
324 		q->p_params.target =
325 			PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
326 	}
327 
328 	/* tupdate is in jiffies */
329 	if (tb[TCA_FQ_PIE_TUPDATE])
330 		q->p_params.tupdate =
331 			usecs_to_jiffies(nla_get_u32(tb[TCA_FQ_PIE_TUPDATE]));
332 
333 	if (tb[TCA_FQ_PIE_ALPHA])
334 		q->p_params.alpha = nla_get_u32(tb[TCA_FQ_PIE_ALPHA]);
335 
336 	if (tb[TCA_FQ_PIE_BETA])
337 		q->p_params.beta = nla_get_u32(tb[TCA_FQ_PIE_BETA]);
338 
339 	if (tb[TCA_FQ_PIE_QUANTUM])
340 		q->quantum = nla_get_u32(tb[TCA_FQ_PIE_QUANTUM]);
341 
342 	if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
343 		q->memory_limit = nla_get_u32(tb[TCA_FQ_PIE_MEMORY_LIMIT]);
344 
345 	if (tb[TCA_FQ_PIE_ECN_PROB])
346 		q->ecn_prob = nla_get_u32(tb[TCA_FQ_PIE_ECN_PROB]);
347 
348 	if (tb[TCA_FQ_PIE_ECN])
349 		q->p_params.ecn = nla_get_u32(tb[TCA_FQ_PIE_ECN]);
350 
351 	if (tb[TCA_FQ_PIE_BYTEMODE])
352 		q->p_params.bytemode = nla_get_u32(tb[TCA_FQ_PIE_BYTEMODE]);
353 
354 	if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
355 		q->p_params.dq_rate_estimator =
356 			nla_get_u32(tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
357 
358 	/* Drop excess packets if new limit is lower */
359 	while (sch->q.qlen > sch->limit) {
360 		struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
361 
362 		len_dropped += qdisc_pkt_len(skb);
363 		num_dropped += 1;
364 		rtnl_kfree_skbs(skb, skb);
365 	}
366 	qdisc_tree_reduce_backlog(sch, num_dropped, len_dropped);
367 
368 	sch_tree_unlock(sch);
369 	return 0;
370 
371 flow_error:
372 	sch_tree_unlock(sch);
373 	return -EINVAL;
374 }
375 
376 static void fq_pie_timer(struct timer_list *t)
377 {
378 	struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
379 	unsigned long next, tupdate;
380 	struct Qdisc *sch = q->sch;
381 	spinlock_t *root_lock; /* to lock qdisc for probability calculations */
382 	int max_cnt, i;
383 
384 	rcu_read_lock();
385 	root_lock = qdisc_lock(qdisc_root_sleeping(sch));
386 	spin_lock(root_lock);
387 
388 	/* Limit this expensive loop to 2048 flows per round. */
389 	max_cnt = min_t(int, q->flows_cnt - q->flows_cursor, 2048);
390 	for (i = 0; i < max_cnt; i++) {
391 		pie_calculate_probability(&q->p_params,
392 					  &q->flows[q->flows_cursor].vars,
393 					  q->flows[q->flows_cursor].backlog);
394 		q->flows_cursor++;
395 	}
396 
397 	tupdate = q->p_params.tupdate;
398 	next = 0;
399 	if (q->flows_cursor >= q->flows_cnt) {
400 		q->flows_cursor = 0;
401 		next = tupdate;
402 	}
403 	if (tupdate)
404 		mod_timer(&q->adapt_timer, jiffies + next);
405 	spin_unlock(root_lock);
406 	rcu_read_unlock();
407 }
408 
409 static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
410 		       struct netlink_ext_ack *extack)
411 {
412 	struct fq_pie_sched_data *q = qdisc_priv(sch);
413 	int err;
414 	u32 idx;
415 
416 	pie_params_init(&q->p_params);
417 	sch->limit = 10 * 1024;
418 	q->p_params.limit = sch->limit;
419 	q->quantum = psched_mtu(qdisc_dev(sch));
420 	q->sch = sch;
421 	q->ecn_prob = 10;
422 	q->flows_cnt = 1024;
423 	q->memory_limit = SZ_32M;
424 
425 	INIT_LIST_HEAD(&q->new_flows);
426 	INIT_LIST_HEAD(&q->old_flows);
427 	timer_setup(&q->adapt_timer, fq_pie_timer, 0);
428 
429 	if (opt) {
430 		err = fq_pie_change(sch, opt, extack);
431 
432 		if (err)
433 			return err;
434 	}
435 
436 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
437 	if (err)
438 		goto init_failure;
439 
440 	q->flows = kvcalloc(q->flows_cnt, sizeof(struct fq_pie_flow),
441 			    GFP_KERNEL);
442 	if (!q->flows) {
443 		err = -ENOMEM;
444 		goto init_failure;
445 	}
446 	for (idx = 0; idx < q->flows_cnt; idx++) {
447 		struct fq_pie_flow *flow = q->flows + idx;
448 
449 		INIT_LIST_HEAD(&flow->flowchain);
450 		pie_vars_init(&flow->vars);
451 	}
452 
453 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
454 
455 	return 0;
456 
457 init_failure:
458 	q->flows_cnt = 0;
459 
460 	return err;
461 }
462 
463 static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
464 {
465 	struct fq_pie_sched_data *q = qdisc_priv(sch);
466 	struct nlattr *opts;
467 
468 	opts = nla_nest_start(skb, TCA_OPTIONS);
469 	if (!opts)
470 		return -EMSGSIZE;
471 
472 	/* convert target from pschedtime to us */
473 	if (nla_put_u32(skb, TCA_FQ_PIE_LIMIT, sch->limit) ||
474 	    nla_put_u32(skb, TCA_FQ_PIE_FLOWS, q->flows_cnt) ||
475 	    nla_put_u32(skb, TCA_FQ_PIE_TARGET,
476 			((u32)PSCHED_TICKS2NS(q->p_params.target)) /
477 			NSEC_PER_USEC) ||
478 	    nla_put_u32(skb, TCA_FQ_PIE_TUPDATE,
479 			jiffies_to_usecs(q->p_params.tupdate)) ||
480 	    nla_put_u32(skb, TCA_FQ_PIE_ALPHA, q->p_params.alpha) ||
481 	    nla_put_u32(skb, TCA_FQ_PIE_BETA, q->p_params.beta) ||
482 	    nla_put_u32(skb, TCA_FQ_PIE_QUANTUM, q->quantum) ||
483 	    nla_put_u32(skb, TCA_FQ_PIE_MEMORY_LIMIT, q->memory_limit) ||
484 	    nla_put_u32(skb, TCA_FQ_PIE_ECN_PROB, q->ecn_prob) ||
485 	    nla_put_u32(skb, TCA_FQ_PIE_ECN, q->p_params.ecn) ||
486 	    nla_put_u32(skb, TCA_FQ_PIE_BYTEMODE, q->p_params.bytemode) ||
487 	    nla_put_u32(skb, TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
488 			q->p_params.dq_rate_estimator))
489 		goto nla_put_failure;
490 
491 	return nla_nest_end(skb, opts);
492 
493 nla_put_failure:
494 	nla_nest_cancel(skb, opts);
495 	return -EMSGSIZE;
496 }
497 
498 static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
499 {
500 	struct fq_pie_sched_data *q = qdisc_priv(sch);
501 	struct tc_fq_pie_xstats st = {
502 		.packets_in	= q->stats.packets_in,
503 		.overlimit	= q->stats.overlimit,
504 		.overmemory	= q->overmemory,
505 		.dropped	= q->stats.dropped,
506 		.ecn_mark	= q->stats.ecn_mark,
507 		.new_flow_count = q->new_flow_count,
508 		.memory_usage   = q->memory_usage,
509 	};
510 	struct list_head *pos;
511 
512 	sch_tree_lock(sch);
513 	list_for_each(pos, &q->new_flows)
514 		st.new_flows_len++;
515 
516 	list_for_each(pos, &q->old_flows)
517 		st.old_flows_len++;
518 	sch_tree_unlock(sch);
519 
520 	return gnet_stats_copy_app(d, &st, sizeof(st));
521 }
522 
523 static void fq_pie_reset(struct Qdisc *sch)
524 {
525 	struct fq_pie_sched_data *q = qdisc_priv(sch);
526 	u32 idx;
527 
528 	INIT_LIST_HEAD(&q->new_flows);
529 	INIT_LIST_HEAD(&q->old_flows);
530 	for (idx = 0; idx < q->flows_cnt; idx++) {
531 		struct fq_pie_flow *flow = q->flows + idx;
532 
533 		/* Removes all packets from flow */
534 		rtnl_kfree_skbs(flow->head, flow->tail);
535 		flow->head = NULL;
536 
537 		INIT_LIST_HEAD(&flow->flowchain);
538 		pie_vars_init(&flow->vars);
539 	}
540 }
541 
542 static void fq_pie_destroy(struct Qdisc *sch)
543 {
544 	struct fq_pie_sched_data *q = qdisc_priv(sch);
545 
546 	tcf_block_put(q->block);
547 	q->p_params.tupdate = 0;
548 	del_timer_sync(&q->adapt_timer);
549 	kvfree(q->flows);
550 }
551 
552 static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
553 	.id		= "fq_pie",
554 	.priv_size	= sizeof(struct fq_pie_sched_data),
555 	.enqueue	= fq_pie_qdisc_enqueue,
556 	.dequeue	= fq_pie_qdisc_dequeue,
557 	.peek		= qdisc_peek_dequeued,
558 	.init		= fq_pie_init,
559 	.destroy	= fq_pie_destroy,
560 	.reset		= fq_pie_reset,
561 	.change		= fq_pie_change,
562 	.dump		= fq_pie_dump,
563 	.dump_stats	= fq_pie_dump_stats,
564 	.owner		= THIS_MODULE,
565 };
566 
567 static int __init fq_pie_module_init(void)
568 {
569 	return register_qdisc(&fq_pie_qdisc_ops);
570 }
571 
572 static void __exit fq_pie_module_exit(void)
573 {
574 	unregister_qdisc(&fq_pie_qdisc_ops);
575 }
576 
577 module_init(fq_pie_module_init);
578 module_exit(fq_pie_module_exit);
579 
580 MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
581 MODULE_AUTHOR("Mohit P. Tahiliani");
582 MODULE_LICENSE("GPL");
583