xref: /openbmc/linux/net/sched/sch_fq_pie.c (revision c4c3c32d)
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 quantum;
65 	u32 memory_limit;
66 	u32 new_flow_count;
67 	u32 memory_usage;
68 	u32 overmemory;
69 	struct pie_stats stats;
70 	struct timer_list adapt_timer;
71 };
72 
73 static unsigned int fq_pie_hash(const struct fq_pie_sched_data *q,
74 				struct sk_buff *skb)
75 {
76 	return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
77 }
78 
79 static unsigned int fq_pie_classify(struct sk_buff *skb, struct Qdisc *sch,
80 				    int *qerr)
81 {
82 	struct fq_pie_sched_data *q = qdisc_priv(sch);
83 	struct tcf_proto *filter;
84 	struct tcf_result res;
85 	int result;
86 
87 	if (TC_H_MAJ(skb->priority) == sch->handle &&
88 	    TC_H_MIN(skb->priority) > 0 &&
89 	    TC_H_MIN(skb->priority) <= q->flows_cnt)
90 		return TC_H_MIN(skb->priority);
91 
92 	filter = rcu_dereference_bh(q->filter_list);
93 	if (!filter)
94 		return fq_pie_hash(q, skb) + 1;
95 
96 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
97 	result = tcf_classify(skb, NULL, filter, &res, false);
98 	if (result >= 0) {
99 #ifdef CONFIG_NET_CLS_ACT
100 		switch (result) {
101 		case TC_ACT_STOLEN:
102 		case TC_ACT_QUEUED:
103 		case TC_ACT_TRAP:
104 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
105 			fallthrough;
106 		case TC_ACT_SHOT:
107 			return 0;
108 		}
109 #endif
110 		if (TC_H_MIN(res.classid) <= q->flows_cnt)
111 			return TC_H_MIN(res.classid);
112 	}
113 	return 0;
114 }
115 
116 /* add skb to flow queue (tail add) */
117 static inline void flow_queue_add(struct fq_pie_flow *flow,
118 				  struct sk_buff *skb)
119 {
120 	if (!flow->head)
121 		flow->head = skb;
122 	else
123 		flow->tail->next = skb;
124 	flow->tail = skb;
125 	skb->next = NULL;
126 }
127 
128 static int fq_pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
129 				struct sk_buff **to_free)
130 {
131 	struct fq_pie_sched_data *q = qdisc_priv(sch);
132 	struct fq_pie_flow *sel_flow;
133 	int ret;
134 	u8 memory_limited = false;
135 	u8 enqueue = false;
136 	u32 pkt_len;
137 	u32 idx;
138 
139 	/* Classifies packet into corresponding flow */
140 	idx = fq_pie_classify(skb, sch, &ret);
141 	if (idx == 0) {
142 		if (ret & __NET_XMIT_BYPASS)
143 			qdisc_qstats_drop(sch);
144 		__qdisc_drop(skb, to_free);
145 		return ret;
146 	}
147 	idx--;
148 
149 	sel_flow = &q->flows[idx];
150 	/* Checks whether adding a new packet would exceed memory limit */
151 	get_pie_cb(skb)->mem_usage = skb->truesize;
152 	memory_limited = q->memory_usage > q->memory_limit + skb->truesize;
153 
154 	/* Checks if the qdisc is full */
155 	if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
156 		q->stats.overlimit++;
157 		goto out;
158 	} else if (unlikely(memory_limited)) {
159 		q->overmemory++;
160 	}
161 
162 	if (!pie_drop_early(sch, &q->p_params, &sel_flow->vars,
163 			    sel_flow->backlog, skb->len)) {
164 		enqueue = true;
165 	} else if (q->p_params.ecn &&
166 		   sel_flow->vars.prob <= (MAX_PROB / 100) * q->ecn_prob &&
167 		   INET_ECN_set_ce(skb)) {
168 		/* If packet is ecn capable, mark it if drop probability
169 		 * is lower than the parameter ecn_prob, else drop it.
170 		 */
171 		q->stats.ecn_mark++;
172 		enqueue = true;
173 	}
174 	if (enqueue) {
175 		/* Set enqueue time only when dq_rate_estimator is disabled. */
176 		if (!q->p_params.dq_rate_estimator)
177 			pie_set_enqueue_time(skb);
178 
179 		pkt_len = qdisc_pkt_len(skb);
180 		q->stats.packets_in++;
181 		q->memory_usage += skb->truesize;
182 		sch->qstats.backlog += pkt_len;
183 		sch->q.qlen++;
184 		flow_queue_add(sel_flow, skb);
185 		if (list_empty(&sel_flow->flowchain)) {
186 			list_add_tail(&sel_flow->flowchain, &q->new_flows);
187 			q->new_flow_count++;
188 			sel_flow->deficit = q->quantum;
189 			sel_flow->qlen = 0;
190 			sel_flow->backlog = 0;
191 		}
192 		sel_flow->qlen++;
193 		sel_flow->backlog += pkt_len;
194 		return NET_XMIT_SUCCESS;
195 	}
196 out:
197 	q->stats.dropped++;
198 	sel_flow->vars.accu_prob = 0;
199 	__qdisc_drop(skb, to_free);
200 	qdisc_qstats_drop(sch);
201 	return NET_XMIT_CN;
202 }
203 
204 static struct netlink_range_validation fq_pie_q_range = {
205 	.min = 1,
206 	.max = 1 << 20,
207 };
208 
209 static const struct nla_policy fq_pie_policy[TCA_FQ_PIE_MAX + 1] = {
210 	[TCA_FQ_PIE_LIMIT]		= {.type = NLA_U32},
211 	[TCA_FQ_PIE_FLOWS]		= {.type = NLA_U32},
212 	[TCA_FQ_PIE_TARGET]		= {.type = NLA_U32},
213 	[TCA_FQ_PIE_TUPDATE]		= {.type = NLA_U32},
214 	[TCA_FQ_PIE_ALPHA]		= {.type = NLA_U32},
215 	[TCA_FQ_PIE_BETA]		= {.type = NLA_U32},
216 	[TCA_FQ_PIE_QUANTUM]		=
217 			NLA_POLICY_FULL_RANGE(NLA_U32, &fq_pie_q_range),
218 	[TCA_FQ_PIE_MEMORY_LIMIT]	= {.type = NLA_U32},
219 	[TCA_FQ_PIE_ECN_PROB]		= {.type = NLA_U32},
220 	[TCA_FQ_PIE_ECN]		= {.type = NLA_U32},
221 	[TCA_FQ_PIE_BYTEMODE]		= {.type = NLA_U32},
222 	[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]	= {.type = NLA_U32},
223 };
224 
225 static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
226 {
227 	struct sk_buff *skb = flow->head;
228 
229 	flow->head = skb->next;
230 	skb->next = NULL;
231 	return skb;
232 }
233 
234 static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
235 {
236 	struct fq_pie_sched_data *q = qdisc_priv(sch);
237 	struct sk_buff *skb = NULL;
238 	struct fq_pie_flow *flow;
239 	struct list_head *head;
240 	u32 pkt_len;
241 
242 begin:
243 	head = &q->new_flows;
244 	if (list_empty(head)) {
245 		head = &q->old_flows;
246 		if (list_empty(head))
247 			return NULL;
248 	}
249 
250 	flow = list_first_entry(head, struct fq_pie_flow, flowchain);
251 	/* Flow has exhausted all its credits */
252 	if (flow->deficit <= 0) {
253 		flow->deficit += q->quantum;
254 		list_move_tail(&flow->flowchain, &q->old_flows);
255 		goto begin;
256 	}
257 
258 	if (flow->head) {
259 		skb = dequeue_head(flow);
260 		pkt_len = qdisc_pkt_len(skb);
261 		sch->qstats.backlog -= pkt_len;
262 		sch->q.qlen--;
263 		qdisc_bstats_update(sch, skb);
264 	}
265 
266 	if (!skb) {
267 		/* force a pass through old_flows to prevent starvation */
268 		if (head == &q->new_flows && !list_empty(&q->old_flows))
269 			list_move_tail(&flow->flowchain, &q->old_flows);
270 		else
271 			list_del_init(&flow->flowchain);
272 		goto begin;
273 	}
274 
275 	flow->qlen--;
276 	flow->deficit -= pkt_len;
277 	flow->backlog -= pkt_len;
278 	q->memory_usage -= get_pie_cb(skb)->mem_usage;
279 	pie_process_dequeue(skb, &q->p_params, &flow->vars, flow->backlog);
280 	return skb;
281 }
282 
283 static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
284 			 struct netlink_ext_ack *extack)
285 {
286 	struct fq_pie_sched_data *q = qdisc_priv(sch);
287 	struct nlattr *tb[TCA_FQ_PIE_MAX + 1];
288 	unsigned int len_dropped = 0;
289 	unsigned int num_dropped = 0;
290 	int err;
291 
292 	err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, opt, fq_pie_policy, extack);
293 	if (err < 0)
294 		return err;
295 
296 	sch_tree_lock(sch);
297 	if (tb[TCA_FQ_PIE_LIMIT]) {
298 		u32 limit = nla_get_u32(tb[TCA_FQ_PIE_LIMIT]);
299 
300 		q->p_params.limit = limit;
301 		sch->limit = limit;
302 	}
303 	if (tb[TCA_FQ_PIE_FLOWS]) {
304 		if (q->flows) {
305 			NL_SET_ERR_MSG_MOD(extack,
306 					   "Number of flows cannot be changed");
307 			goto flow_error;
308 		}
309 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_PIE_FLOWS]);
310 		if (!q->flows_cnt || q->flows_cnt > 65536) {
311 			NL_SET_ERR_MSG_MOD(extack,
312 					   "Number of flows must range in [1..65536]");
313 			goto flow_error;
314 		}
315 	}
316 
317 	/* convert from microseconds to pschedtime */
318 	if (tb[TCA_FQ_PIE_TARGET]) {
319 		/* target is in us */
320 		u32 target = nla_get_u32(tb[TCA_FQ_PIE_TARGET]);
321 
322 		/* convert to pschedtime */
323 		q->p_params.target =
324 			PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
325 	}
326 
327 	/* tupdate is in jiffies */
328 	if (tb[TCA_FQ_PIE_TUPDATE])
329 		q->p_params.tupdate =
330 			usecs_to_jiffies(nla_get_u32(tb[TCA_FQ_PIE_TUPDATE]));
331 
332 	if (tb[TCA_FQ_PIE_ALPHA])
333 		q->p_params.alpha = nla_get_u32(tb[TCA_FQ_PIE_ALPHA]);
334 
335 	if (tb[TCA_FQ_PIE_BETA])
336 		q->p_params.beta = nla_get_u32(tb[TCA_FQ_PIE_BETA]);
337 
338 	if (tb[TCA_FQ_PIE_QUANTUM])
339 		q->quantum = nla_get_u32(tb[TCA_FQ_PIE_QUANTUM]);
340 
341 	if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
342 		q->memory_limit = nla_get_u32(tb[TCA_FQ_PIE_MEMORY_LIMIT]);
343 
344 	if (tb[TCA_FQ_PIE_ECN_PROB])
345 		q->ecn_prob = nla_get_u32(tb[TCA_FQ_PIE_ECN_PROB]);
346 
347 	if (tb[TCA_FQ_PIE_ECN])
348 		q->p_params.ecn = nla_get_u32(tb[TCA_FQ_PIE_ECN]);
349 
350 	if (tb[TCA_FQ_PIE_BYTEMODE])
351 		q->p_params.bytemode = nla_get_u32(tb[TCA_FQ_PIE_BYTEMODE]);
352 
353 	if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
354 		q->p_params.dq_rate_estimator =
355 			nla_get_u32(tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
356 
357 	/* Drop excess packets if new limit is lower */
358 	while (sch->q.qlen > sch->limit) {
359 		struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
360 
361 		len_dropped += qdisc_pkt_len(skb);
362 		num_dropped += 1;
363 		rtnl_kfree_skbs(skb, skb);
364 	}
365 	qdisc_tree_reduce_backlog(sch, num_dropped, len_dropped);
366 
367 	sch_tree_unlock(sch);
368 	return 0;
369 
370 flow_error:
371 	sch_tree_unlock(sch);
372 	return -EINVAL;
373 }
374 
375 static void fq_pie_timer(struct timer_list *t)
376 {
377 	struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
378 	struct Qdisc *sch = q->sch;
379 	spinlock_t *root_lock; /* to lock qdisc for probability calculations */
380 	u32 idx;
381 
382 	rcu_read_lock();
383 	root_lock = qdisc_lock(qdisc_root_sleeping(sch));
384 	spin_lock(root_lock);
385 
386 	for (idx = 0; idx < q->flows_cnt; idx++)
387 		pie_calculate_probability(&q->p_params, &q->flows[idx].vars,
388 					  q->flows[idx].backlog);
389 
390 	/* reset the timer to fire after 'tupdate' jiffies. */
391 	if (q->p_params.tupdate)
392 		mod_timer(&q->adapt_timer, jiffies + q->p_params.tupdate);
393 
394 	spin_unlock(root_lock);
395 	rcu_read_unlock();
396 }
397 
398 static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
399 		       struct netlink_ext_ack *extack)
400 {
401 	struct fq_pie_sched_data *q = qdisc_priv(sch);
402 	int err;
403 	u32 idx;
404 
405 	pie_params_init(&q->p_params);
406 	sch->limit = 10 * 1024;
407 	q->p_params.limit = sch->limit;
408 	q->quantum = psched_mtu(qdisc_dev(sch));
409 	q->sch = sch;
410 	q->ecn_prob = 10;
411 	q->flows_cnt = 1024;
412 	q->memory_limit = SZ_32M;
413 
414 	INIT_LIST_HEAD(&q->new_flows);
415 	INIT_LIST_HEAD(&q->old_flows);
416 	timer_setup(&q->adapt_timer, fq_pie_timer, 0);
417 
418 	if (opt) {
419 		err = fq_pie_change(sch, opt, extack);
420 
421 		if (err)
422 			return err;
423 	}
424 
425 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
426 	if (err)
427 		goto init_failure;
428 
429 	q->flows = kvcalloc(q->flows_cnt, sizeof(struct fq_pie_flow),
430 			    GFP_KERNEL);
431 	if (!q->flows) {
432 		err = -ENOMEM;
433 		goto init_failure;
434 	}
435 	for (idx = 0; idx < q->flows_cnt; idx++) {
436 		struct fq_pie_flow *flow = q->flows + idx;
437 
438 		INIT_LIST_HEAD(&flow->flowchain);
439 		pie_vars_init(&flow->vars);
440 	}
441 
442 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
443 
444 	return 0;
445 
446 init_failure:
447 	q->flows_cnt = 0;
448 
449 	return err;
450 }
451 
452 static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
453 {
454 	struct fq_pie_sched_data *q = qdisc_priv(sch);
455 	struct nlattr *opts;
456 
457 	opts = nla_nest_start(skb, TCA_OPTIONS);
458 	if (!opts)
459 		return -EMSGSIZE;
460 
461 	/* convert target from pschedtime to us */
462 	if (nla_put_u32(skb, TCA_FQ_PIE_LIMIT, sch->limit) ||
463 	    nla_put_u32(skb, TCA_FQ_PIE_FLOWS, q->flows_cnt) ||
464 	    nla_put_u32(skb, TCA_FQ_PIE_TARGET,
465 			((u32)PSCHED_TICKS2NS(q->p_params.target)) /
466 			NSEC_PER_USEC) ||
467 	    nla_put_u32(skb, TCA_FQ_PIE_TUPDATE,
468 			jiffies_to_usecs(q->p_params.tupdate)) ||
469 	    nla_put_u32(skb, TCA_FQ_PIE_ALPHA, q->p_params.alpha) ||
470 	    nla_put_u32(skb, TCA_FQ_PIE_BETA, q->p_params.beta) ||
471 	    nla_put_u32(skb, TCA_FQ_PIE_QUANTUM, q->quantum) ||
472 	    nla_put_u32(skb, TCA_FQ_PIE_MEMORY_LIMIT, q->memory_limit) ||
473 	    nla_put_u32(skb, TCA_FQ_PIE_ECN_PROB, q->ecn_prob) ||
474 	    nla_put_u32(skb, TCA_FQ_PIE_ECN, q->p_params.ecn) ||
475 	    nla_put_u32(skb, TCA_FQ_PIE_BYTEMODE, q->p_params.bytemode) ||
476 	    nla_put_u32(skb, TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
477 			q->p_params.dq_rate_estimator))
478 		goto nla_put_failure;
479 
480 	return nla_nest_end(skb, opts);
481 
482 nla_put_failure:
483 	nla_nest_cancel(skb, opts);
484 	return -EMSGSIZE;
485 }
486 
487 static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
488 {
489 	struct fq_pie_sched_data *q = qdisc_priv(sch);
490 	struct tc_fq_pie_xstats st = {
491 		.packets_in	= q->stats.packets_in,
492 		.overlimit	= q->stats.overlimit,
493 		.overmemory	= q->overmemory,
494 		.dropped	= q->stats.dropped,
495 		.ecn_mark	= q->stats.ecn_mark,
496 		.new_flow_count = q->new_flow_count,
497 		.memory_usage   = q->memory_usage,
498 	};
499 	struct list_head *pos;
500 
501 	sch_tree_lock(sch);
502 	list_for_each(pos, &q->new_flows)
503 		st.new_flows_len++;
504 
505 	list_for_each(pos, &q->old_flows)
506 		st.old_flows_len++;
507 	sch_tree_unlock(sch);
508 
509 	return gnet_stats_copy_app(d, &st, sizeof(st));
510 }
511 
512 static void fq_pie_reset(struct Qdisc *sch)
513 {
514 	struct fq_pie_sched_data *q = qdisc_priv(sch);
515 	u32 idx;
516 
517 	INIT_LIST_HEAD(&q->new_flows);
518 	INIT_LIST_HEAD(&q->old_flows);
519 	for (idx = 0; idx < q->flows_cnt; idx++) {
520 		struct fq_pie_flow *flow = q->flows + idx;
521 
522 		/* Removes all packets from flow */
523 		rtnl_kfree_skbs(flow->head, flow->tail);
524 		flow->head = NULL;
525 
526 		INIT_LIST_HEAD(&flow->flowchain);
527 		pie_vars_init(&flow->vars);
528 	}
529 }
530 
531 static void fq_pie_destroy(struct Qdisc *sch)
532 {
533 	struct fq_pie_sched_data *q = qdisc_priv(sch);
534 
535 	tcf_block_put(q->block);
536 	q->p_params.tupdate = 0;
537 	del_timer_sync(&q->adapt_timer);
538 	kvfree(q->flows);
539 }
540 
541 static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
542 	.id		= "fq_pie",
543 	.priv_size	= sizeof(struct fq_pie_sched_data),
544 	.enqueue	= fq_pie_qdisc_enqueue,
545 	.dequeue	= fq_pie_qdisc_dequeue,
546 	.peek		= qdisc_peek_dequeued,
547 	.init		= fq_pie_init,
548 	.destroy	= fq_pie_destroy,
549 	.reset		= fq_pie_reset,
550 	.change		= fq_pie_change,
551 	.dump		= fq_pie_dump,
552 	.dump_stats	= fq_pie_dump_stats,
553 	.owner		= THIS_MODULE,
554 };
555 
556 static int __init fq_pie_module_init(void)
557 {
558 	return register_qdisc(&fq_pie_qdisc_ops);
559 }
560 
561 static void __exit fq_pie_module_exit(void)
562 {
563 	unregister_qdisc(&fq_pie_qdisc_ops);
564 }
565 
566 module_init(fq_pie_module_init);
567 module_exit(fq_pie_module_exit);
568 
569 MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
570 MODULE_AUTHOR("Mohit P. Tahiliani");
571 MODULE_LICENSE("GPL");
572