xref: /openbmc/linux/net/sched/sch_fq_pie.c (revision be2b81b5)
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 const struct nla_policy fq_pie_policy[TCA_FQ_PIE_MAX + 1] = {
205 	[TCA_FQ_PIE_LIMIT]		= {.type = NLA_U32},
206 	[TCA_FQ_PIE_FLOWS]		= {.type = NLA_U32},
207 	[TCA_FQ_PIE_TARGET]		= {.type = NLA_U32},
208 	[TCA_FQ_PIE_TUPDATE]		= {.type = NLA_U32},
209 	[TCA_FQ_PIE_ALPHA]		= {.type = NLA_U32},
210 	[TCA_FQ_PIE_BETA]		= {.type = NLA_U32},
211 	[TCA_FQ_PIE_QUANTUM]		= {.type = NLA_U32},
212 	[TCA_FQ_PIE_MEMORY_LIMIT]	= {.type = NLA_U32},
213 	[TCA_FQ_PIE_ECN_PROB]		= {.type = NLA_U32},
214 	[TCA_FQ_PIE_ECN]		= {.type = NLA_U32},
215 	[TCA_FQ_PIE_BYTEMODE]		= {.type = NLA_U32},
216 	[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]	= {.type = NLA_U32},
217 };
218 
219 static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
220 {
221 	struct sk_buff *skb = flow->head;
222 
223 	flow->head = skb->next;
224 	skb->next = NULL;
225 	return skb;
226 }
227 
228 static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
229 {
230 	struct fq_pie_sched_data *q = qdisc_priv(sch);
231 	struct sk_buff *skb = NULL;
232 	struct fq_pie_flow *flow;
233 	struct list_head *head;
234 	u32 pkt_len;
235 
236 begin:
237 	head = &q->new_flows;
238 	if (list_empty(head)) {
239 		head = &q->old_flows;
240 		if (list_empty(head))
241 			return NULL;
242 	}
243 
244 	flow = list_first_entry(head, struct fq_pie_flow, flowchain);
245 	/* Flow has exhausted all its credits */
246 	if (flow->deficit <= 0) {
247 		flow->deficit += q->quantum;
248 		list_move_tail(&flow->flowchain, &q->old_flows);
249 		goto begin;
250 	}
251 
252 	if (flow->head) {
253 		skb = dequeue_head(flow);
254 		pkt_len = qdisc_pkt_len(skb);
255 		sch->qstats.backlog -= pkt_len;
256 		sch->q.qlen--;
257 		qdisc_bstats_update(sch, skb);
258 	}
259 
260 	if (!skb) {
261 		/* force a pass through old_flows to prevent starvation */
262 		if (head == &q->new_flows && !list_empty(&q->old_flows))
263 			list_move_tail(&flow->flowchain, &q->old_flows);
264 		else
265 			list_del_init(&flow->flowchain);
266 		goto begin;
267 	}
268 
269 	flow->qlen--;
270 	flow->deficit -= pkt_len;
271 	flow->backlog -= pkt_len;
272 	q->memory_usage -= get_pie_cb(skb)->mem_usage;
273 	pie_process_dequeue(skb, &q->p_params, &flow->vars, flow->backlog);
274 	return skb;
275 }
276 
277 static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
278 			 struct netlink_ext_ack *extack)
279 {
280 	struct fq_pie_sched_data *q = qdisc_priv(sch);
281 	struct nlattr *tb[TCA_FQ_PIE_MAX + 1];
282 	unsigned int len_dropped = 0;
283 	unsigned int num_dropped = 0;
284 	int err;
285 
286 	err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, opt, fq_pie_policy, extack);
287 	if (err < 0)
288 		return err;
289 
290 	sch_tree_lock(sch);
291 	if (tb[TCA_FQ_PIE_LIMIT]) {
292 		u32 limit = nla_get_u32(tb[TCA_FQ_PIE_LIMIT]);
293 
294 		q->p_params.limit = limit;
295 		sch->limit = limit;
296 	}
297 	if (tb[TCA_FQ_PIE_FLOWS]) {
298 		if (q->flows) {
299 			NL_SET_ERR_MSG_MOD(extack,
300 					   "Number of flows cannot be changed");
301 			goto flow_error;
302 		}
303 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_PIE_FLOWS]);
304 		if (!q->flows_cnt || q->flows_cnt > 65536) {
305 			NL_SET_ERR_MSG_MOD(extack,
306 					   "Number of flows must range in [1..65536]");
307 			goto flow_error;
308 		}
309 	}
310 
311 	/* convert from microseconds to pschedtime */
312 	if (tb[TCA_FQ_PIE_TARGET]) {
313 		/* target is in us */
314 		u32 target = nla_get_u32(tb[TCA_FQ_PIE_TARGET]);
315 
316 		/* convert to pschedtime */
317 		q->p_params.target =
318 			PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
319 	}
320 
321 	/* tupdate is in jiffies */
322 	if (tb[TCA_FQ_PIE_TUPDATE])
323 		q->p_params.tupdate =
324 			usecs_to_jiffies(nla_get_u32(tb[TCA_FQ_PIE_TUPDATE]));
325 
326 	if (tb[TCA_FQ_PIE_ALPHA])
327 		q->p_params.alpha = nla_get_u32(tb[TCA_FQ_PIE_ALPHA]);
328 
329 	if (tb[TCA_FQ_PIE_BETA])
330 		q->p_params.beta = nla_get_u32(tb[TCA_FQ_PIE_BETA]);
331 
332 	if (tb[TCA_FQ_PIE_QUANTUM])
333 		q->quantum = nla_get_u32(tb[TCA_FQ_PIE_QUANTUM]);
334 
335 	if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
336 		q->memory_limit = nla_get_u32(tb[TCA_FQ_PIE_MEMORY_LIMIT]);
337 
338 	if (tb[TCA_FQ_PIE_ECN_PROB])
339 		q->ecn_prob = nla_get_u32(tb[TCA_FQ_PIE_ECN_PROB]);
340 
341 	if (tb[TCA_FQ_PIE_ECN])
342 		q->p_params.ecn = nla_get_u32(tb[TCA_FQ_PIE_ECN]);
343 
344 	if (tb[TCA_FQ_PIE_BYTEMODE])
345 		q->p_params.bytemode = nla_get_u32(tb[TCA_FQ_PIE_BYTEMODE]);
346 
347 	if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
348 		q->p_params.dq_rate_estimator =
349 			nla_get_u32(tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
350 
351 	/* Drop excess packets if new limit is lower */
352 	while (sch->q.qlen > sch->limit) {
353 		struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
354 
355 		len_dropped += qdisc_pkt_len(skb);
356 		num_dropped += 1;
357 		rtnl_kfree_skbs(skb, skb);
358 	}
359 	qdisc_tree_reduce_backlog(sch, num_dropped, len_dropped);
360 
361 	sch_tree_unlock(sch);
362 	return 0;
363 
364 flow_error:
365 	sch_tree_unlock(sch);
366 	return -EINVAL;
367 }
368 
369 static void fq_pie_timer(struct timer_list *t)
370 {
371 	struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
372 	struct Qdisc *sch = q->sch;
373 	spinlock_t *root_lock; /* to lock qdisc for probability calculations */
374 	u32 idx;
375 
376 	root_lock = qdisc_lock(qdisc_root_sleeping(sch));
377 	spin_lock(root_lock);
378 
379 	for (idx = 0; idx < q->flows_cnt; idx++)
380 		pie_calculate_probability(&q->p_params, &q->flows[idx].vars,
381 					  q->flows[idx].backlog);
382 
383 	/* reset the timer to fire after 'tupdate' jiffies. */
384 	if (q->p_params.tupdate)
385 		mod_timer(&q->adapt_timer, jiffies + q->p_params.tupdate);
386 
387 	spin_unlock(root_lock);
388 }
389 
390 static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
391 		       struct netlink_ext_ack *extack)
392 {
393 	struct fq_pie_sched_data *q = qdisc_priv(sch);
394 	int err;
395 	u32 idx;
396 
397 	pie_params_init(&q->p_params);
398 	sch->limit = 10 * 1024;
399 	q->p_params.limit = sch->limit;
400 	q->quantum = psched_mtu(qdisc_dev(sch));
401 	q->sch = sch;
402 	q->ecn_prob = 10;
403 	q->flows_cnt = 1024;
404 	q->memory_limit = SZ_32M;
405 
406 	INIT_LIST_HEAD(&q->new_flows);
407 	INIT_LIST_HEAD(&q->old_flows);
408 	timer_setup(&q->adapt_timer, fq_pie_timer, 0);
409 
410 	if (opt) {
411 		err = fq_pie_change(sch, opt, extack);
412 
413 		if (err)
414 			return err;
415 	}
416 
417 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
418 	if (err)
419 		goto init_failure;
420 
421 	q->flows = kvcalloc(q->flows_cnt, sizeof(struct fq_pie_flow),
422 			    GFP_KERNEL);
423 	if (!q->flows) {
424 		err = -ENOMEM;
425 		goto init_failure;
426 	}
427 	for (idx = 0; idx < q->flows_cnt; idx++) {
428 		struct fq_pie_flow *flow = q->flows + idx;
429 
430 		INIT_LIST_HEAD(&flow->flowchain);
431 		pie_vars_init(&flow->vars);
432 	}
433 
434 	mod_timer(&q->adapt_timer, jiffies + HZ / 2);
435 
436 	return 0;
437 
438 init_failure:
439 	q->flows_cnt = 0;
440 
441 	return err;
442 }
443 
444 static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
445 {
446 	struct fq_pie_sched_data *q = qdisc_priv(sch);
447 	struct nlattr *opts;
448 
449 	opts = nla_nest_start(skb, TCA_OPTIONS);
450 	if (!opts)
451 		return -EMSGSIZE;
452 
453 	/* convert target from pschedtime to us */
454 	if (nla_put_u32(skb, TCA_FQ_PIE_LIMIT, sch->limit) ||
455 	    nla_put_u32(skb, TCA_FQ_PIE_FLOWS, q->flows_cnt) ||
456 	    nla_put_u32(skb, TCA_FQ_PIE_TARGET,
457 			((u32)PSCHED_TICKS2NS(q->p_params.target)) /
458 			NSEC_PER_USEC) ||
459 	    nla_put_u32(skb, TCA_FQ_PIE_TUPDATE,
460 			jiffies_to_usecs(q->p_params.tupdate)) ||
461 	    nla_put_u32(skb, TCA_FQ_PIE_ALPHA, q->p_params.alpha) ||
462 	    nla_put_u32(skb, TCA_FQ_PIE_BETA, q->p_params.beta) ||
463 	    nla_put_u32(skb, TCA_FQ_PIE_QUANTUM, q->quantum) ||
464 	    nla_put_u32(skb, TCA_FQ_PIE_MEMORY_LIMIT, q->memory_limit) ||
465 	    nla_put_u32(skb, TCA_FQ_PIE_ECN_PROB, q->ecn_prob) ||
466 	    nla_put_u32(skb, TCA_FQ_PIE_ECN, q->p_params.ecn) ||
467 	    nla_put_u32(skb, TCA_FQ_PIE_BYTEMODE, q->p_params.bytemode) ||
468 	    nla_put_u32(skb, TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
469 			q->p_params.dq_rate_estimator))
470 		goto nla_put_failure;
471 
472 	return nla_nest_end(skb, opts);
473 
474 nla_put_failure:
475 	nla_nest_cancel(skb, opts);
476 	return -EMSGSIZE;
477 }
478 
479 static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
480 {
481 	struct fq_pie_sched_data *q = qdisc_priv(sch);
482 	struct tc_fq_pie_xstats st = {
483 		.packets_in	= q->stats.packets_in,
484 		.overlimit	= q->stats.overlimit,
485 		.overmemory	= q->overmemory,
486 		.dropped	= q->stats.dropped,
487 		.ecn_mark	= q->stats.ecn_mark,
488 		.new_flow_count = q->new_flow_count,
489 		.memory_usage   = q->memory_usage,
490 	};
491 	struct list_head *pos;
492 
493 	sch_tree_lock(sch);
494 	list_for_each(pos, &q->new_flows)
495 		st.new_flows_len++;
496 
497 	list_for_each(pos, &q->old_flows)
498 		st.old_flows_len++;
499 	sch_tree_unlock(sch);
500 
501 	return gnet_stats_copy_app(d, &st, sizeof(st));
502 }
503 
504 static void fq_pie_reset(struct Qdisc *sch)
505 {
506 	struct fq_pie_sched_data *q = qdisc_priv(sch);
507 	u32 idx;
508 
509 	INIT_LIST_HEAD(&q->new_flows);
510 	INIT_LIST_HEAD(&q->old_flows);
511 	for (idx = 0; idx < q->flows_cnt; idx++) {
512 		struct fq_pie_flow *flow = q->flows + idx;
513 
514 		/* Removes all packets from flow */
515 		rtnl_kfree_skbs(flow->head, flow->tail);
516 		flow->head = NULL;
517 
518 		INIT_LIST_HEAD(&flow->flowchain);
519 		pie_vars_init(&flow->vars);
520 	}
521 }
522 
523 static void fq_pie_destroy(struct Qdisc *sch)
524 {
525 	struct fq_pie_sched_data *q = qdisc_priv(sch);
526 
527 	tcf_block_put(q->block);
528 	q->p_params.tupdate = 0;
529 	del_timer_sync(&q->adapt_timer);
530 	kvfree(q->flows);
531 }
532 
533 static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
534 	.id		= "fq_pie",
535 	.priv_size	= sizeof(struct fq_pie_sched_data),
536 	.enqueue	= fq_pie_qdisc_enqueue,
537 	.dequeue	= fq_pie_qdisc_dequeue,
538 	.peek		= qdisc_peek_dequeued,
539 	.init		= fq_pie_init,
540 	.destroy	= fq_pie_destroy,
541 	.reset		= fq_pie_reset,
542 	.change		= fq_pie_change,
543 	.dump		= fq_pie_dump,
544 	.dump_stats	= fq_pie_dump_stats,
545 	.owner		= THIS_MODULE,
546 };
547 
548 static int __init fq_pie_module_init(void)
549 {
550 	return register_qdisc(&fq_pie_qdisc_ops);
551 }
552 
553 static void __exit fq_pie_module_exit(void)
554 {
555 	unregister_qdisc(&fq_pie_qdisc_ops);
556 }
557 
558 module_init(fq_pie_module_init);
559 module_exit(fq_pie_module_exit);
560 
561 MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
562 MODULE_AUTHOR("Mohit P. Tahiliani");
563 MODULE_LICENSE("GPL");
564