1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Flow Queue PIE discipline
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>
12 #include <linux/jhash.h>
13 #include <linux/sizes.h>
14 #include <linux/vmalloc.h>
15 #include <net/pkt_cls.h>
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.
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
49 struct list_head flowchain;
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;
59 struct list_head old_flows;
60 struct list_head new_flows;
61 struct pie_params p_params;
70 struct pie_stats stats;
71 struct timer_list adapt_timer;
74 static unsigned int fq_pie_hash(const struct fq_pie_sched_data *q,
77 return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
80 static unsigned int fq_pie_classify(struct sk_buff *skb, struct Qdisc *sch,
83 struct fq_pie_sched_data *q = qdisc_priv(sch);
84 struct tcf_proto *filter;
85 struct tcf_result res;
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);
93 filter = rcu_dereference_bh(q->filter_list);
95 return fq_pie_hash(q, skb) + 1;
97 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
98 result = tcf_classify(skb, filter, &res, false);
100 #ifdef CONFIG_NET_CLS_ACT
105 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
111 if (TC_H_MIN(res.classid) <= q->flows_cnt)
112 return TC_H_MIN(res.classid);
117 /* add skb to flow queue (tail add) */
118 static inline void flow_queue_add(struct fq_pie_flow *flow,
124 flow->tail->next = skb;
129 static int fq_pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
130 struct sk_buff **to_free)
132 struct fq_pie_sched_data *q = qdisc_priv(sch);
133 struct fq_pie_flow *sel_flow;
135 u8 memory_limited = false;
140 /* Classifies packet into corresponding flow */
141 idx = fq_pie_classify(skb, sch, &ret);
143 if (ret & __NET_XMIT_BYPASS)
144 qdisc_qstats_drop(sch);
145 __qdisc_drop(skb, to_free);
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;
155 /* Checks if the qdisc is full */
156 if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
157 q->stats.overlimit++;
159 } else if (unlikely(memory_limited)) {
163 if (!pie_drop_early(sch, &q->p_params, &sel_flow->vars,
164 sel_flow->backlog, skb->len)) {
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.
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);
180 pkt_len = qdisc_pkt_len(skb);
181 q->stats.packets_in++;
182 q->memory_usage += skb->truesize;
183 sch->qstats.backlog += pkt_len;
185 flow_queue_add(sel_flow, skb);
186 if (list_empty(&sel_flow->flowchain)) {
187 list_add_tail(&sel_flow->flowchain, &q->new_flows);
189 sel_flow->deficit = q->quantum;
191 sel_flow->backlog = 0;
194 sel_flow->backlog += pkt_len;
195 return NET_XMIT_SUCCESS;
199 sel_flow->vars.accu_prob = 0;
200 __qdisc_drop(skb, to_free);
201 qdisc_qstats_drop(sch);
205 static struct netlink_range_validation fq_pie_q_range = {
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},
226 static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
228 struct sk_buff *skb = flow->head;
230 flow->head = skb->next;
235 static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
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;
244 head = &q->new_flows;
245 if (list_empty(head)) {
246 head = &q->old_flows;
247 if (list_empty(head))
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);
260 skb = dequeue_head(flow);
261 pkt_len = qdisc_pkt_len(skb);
262 sch->qstats.backlog -= pkt_len;
264 qdisc_bstats_update(sch, 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);
272 list_del_init(&flow->flowchain);
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);
284 static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
285 struct netlink_ext_ack *extack)
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;
296 err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, opt, fq_pie_policy, extack);
301 if (tb[TCA_FQ_PIE_LIMIT]) {
302 u32 limit = nla_get_u32(tb[TCA_FQ_PIE_LIMIT]);
304 q->p_params.limit = limit;
307 if (tb[TCA_FQ_PIE_FLOWS]) {
309 NL_SET_ERR_MSG_MOD(extack,
310 "Number of flows cannot be changed");
313 q->flows_cnt = nla_get_u32(tb[TCA_FQ_PIE_FLOWS]);
314 if (!q->flows_cnt || q->flows_cnt > 65536) {
315 NL_SET_ERR_MSG_MOD(extack,
316 "Number of flows must range in [1..65536]");
321 /* convert from microseconds to pschedtime */
322 if (tb[TCA_FQ_PIE_TARGET]) {
323 /* target is in us */
324 u32 target = nla_get_u32(tb[TCA_FQ_PIE_TARGET]);
326 /* convert to pschedtime */
328 PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
331 /* tupdate is in jiffies */
332 if (tb[TCA_FQ_PIE_TUPDATE])
333 q->p_params.tupdate =
334 usecs_to_jiffies(nla_get_u32(tb[TCA_FQ_PIE_TUPDATE]));
336 if (tb[TCA_FQ_PIE_ALPHA])
337 q->p_params.alpha = nla_get_u32(tb[TCA_FQ_PIE_ALPHA]);
339 if (tb[TCA_FQ_PIE_BETA])
340 q->p_params.beta = nla_get_u32(tb[TCA_FQ_PIE_BETA]);
342 if (tb[TCA_FQ_PIE_QUANTUM])
343 q->quantum = nla_get_u32(tb[TCA_FQ_PIE_QUANTUM]);
345 if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
346 q->memory_limit = nla_get_u32(tb[TCA_FQ_PIE_MEMORY_LIMIT]);
348 if (tb[TCA_FQ_PIE_ECN_PROB])
349 q->ecn_prob = nla_get_u32(tb[TCA_FQ_PIE_ECN_PROB]);
351 if (tb[TCA_FQ_PIE_ECN])
352 q->p_params.ecn = nla_get_u32(tb[TCA_FQ_PIE_ECN]);
354 if (tb[TCA_FQ_PIE_BYTEMODE])
355 q->p_params.bytemode = nla_get_u32(tb[TCA_FQ_PIE_BYTEMODE]);
357 if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
358 q->p_params.dq_rate_estimator =
359 nla_get_u32(tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
361 /* Drop excess packets if new limit is lower */
362 while (sch->q.qlen > sch->limit) {
363 struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
365 len_dropped += qdisc_pkt_len(skb);
367 rtnl_kfree_skbs(skb, skb);
369 qdisc_tree_reduce_backlog(sch, num_dropped, len_dropped);
371 sch_tree_unlock(sch);
375 sch_tree_unlock(sch);
379 static void fq_pie_timer(struct timer_list *t)
381 struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
382 unsigned long next, tupdate;
383 struct Qdisc *sch = q->sch;
384 spinlock_t *root_lock; /* to lock qdisc for probability calculations */
387 root_lock = qdisc_lock(qdisc_root_sleeping(sch));
388 spin_lock(root_lock);
390 /* Limit this expensive loop to 2048 flows per round. */
391 max_cnt = min_t(int, q->flows_cnt - q->flows_cursor, 2048);
392 for (i = 0; i < max_cnt; i++) {
393 pie_calculate_probability(&q->p_params,
394 &q->flows[q->flows_cursor].vars,
395 q->flows[q->flows_cursor].backlog);
399 tupdate = q->p_params.tupdate;
401 if (q->flows_cursor >= q->flows_cnt) {
406 mod_timer(&q->adapt_timer, jiffies + next);
407 spin_unlock(root_lock);
410 static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
411 struct netlink_ext_ack *extack)
413 struct fq_pie_sched_data *q = qdisc_priv(sch);
417 pie_params_init(&q->p_params);
418 sch->limit = 10 * 1024;
419 q->p_params.limit = sch->limit;
420 q->quantum = psched_mtu(qdisc_dev(sch));
424 q->memory_limit = SZ_32M;
426 INIT_LIST_HEAD(&q->new_flows);
427 INIT_LIST_HEAD(&q->old_flows);
428 timer_setup(&q->adapt_timer, fq_pie_timer, 0);
431 err = fq_pie_change(sch, opt, extack);
437 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
441 q->flows = kvcalloc(q->flows_cnt, sizeof(struct fq_pie_flow),
447 for (idx = 0; idx < q->flows_cnt; idx++) {
448 struct fq_pie_flow *flow = q->flows + idx;
450 INIT_LIST_HEAD(&flow->flowchain);
451 pie_vars_init(&flow->vars);
454 mod_timer(&q->adapt_timer, jiffies + HZ / 2);
464 static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
466 struct fq_pie_sched_data *q = qdisc_priv(sch);
469 opts = nla_nest_start(skb, TCA_OPTIONS);
473 /* convert target from pschedtime to us */
474 if (nla_put_u32(skb, TCA_FQ_PIE_LIMIT, sch->limit) ||
475 nla_put_u32(skb, TCA_FQ_PIE_FLOWS, q->flows_cnt) ||
476 nla_put_u32(skb, TCA_FQ_PIE_TARGET,
477 ((u32)PSCHED_TICKS2NS(q->p_params.target)) /
479 nla_put_u32(skb, TCA_FQ_PIE_TUPDATE,
480 jiffies_to_usecs(q->p_params.tupdate)) ||
481 nla_put_u32(skb, TCA_FQ_PIE_ALPHA, q->p_params.alpha) ||
482 nla_put_u32(skb, TCA_FQ_PIE_BETA, q->p_params.beta) ||
483 nla_put_u32(skb, TCA_FQ_PIE_QUANTUM, q->quantum) ||
484 nla_put_u32(skb, TCA_FQ_PIE_MEMORY_LIMIT, q->memory_limit) ||
485 nla_put_u32(skb, TCA_FQ_PIE_ECN_PROB, q->ecn_prob) ||
486 nla_put_u32(skb, TCA_FQ_PIE_ECN, q->p_params.ecn) ||
487 nla_put_u32(skb, TCA_FQ_PIE_BYTEMODE, q->p_params.bytemode) ||
488 nla_put_u32(skb, TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
489 q->p_params.dq_rate_estimator))
490 goto nla_put_failure;
492 return nla_nest_end(skb, opts);
495 nla_nest_cancel(skb, opts);
499 static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
501 struct fq_pie_sched_data *q = qdisc_priv(sch);
502 struct tc_fq_pie_xstats st = {
503 .packets_in = q->stats.packets_in,
504 .overlimit = q->stats.overlimit,
505 .overmemory = q->overmemory,
506 .dropped = q->stats.dropped,
507 .ecn_mark = q->stats.ecn_mark,
508 .new_flow_count = q->new_flow_count,
509 .memory_usage = q->memory_usage,
511 struct list_head *pos;
514 list_for_each(pos, &q->new_flows)
517 list_for_each(pos, &q->old_flows)
519 sch_tree_unlock(sch);
521 return gnet_stats_copy_app(d, &st, sizeof(st));
524 static void fq_pie_reset(struct Qdisc *sch)
526 struct fq_pie_sched_data *q = qdisc_priv(sch);
529 INIT_LIST_HEAD(&q->new_flows);
530 INIT_LIST_HEAD(&q->old_flows);
531 for (idx = 0; idx < q->flows_cnt; idx++) {
532 struct fq_pie_flow *flow = q->flows + idx;
534 /* Removes all packets from flow */
535 rtnl_kfree_skbs(flow->head, flow->tail);
538 INIT_LIST_HEAD(&flow->flowchain);
539 pie_vars_init(&flow->vars);
543 static void fq_pie_destroy(struct Qdisc *sch)
545 struct fq_pie_sched_data *q = qdisc_priv(sch);
547 tcf_block_put(q->block);
548 q->p_params.tupdate = 0;
549 del_timer_sync(&q->adapt_timer);
553 static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
555 .priv_size = sizeof(struct fq_pie_sched_data),
556 .enqueue = fq_pie_qdisc_enqueue,
557 .dequeue = fq_pie_qdisc_dequeue,
558 .peek = qdisc_peek_dequeued,
560 .destroy = fq_pie_destroy,
561 .reset = fq_pie_reset,
562 .change = fq_pie_change,
564 .dump_stats = fq_pie_dump_stats,
565 .owner = THIS_MODULE,
568 static int __init fq_pie_module_init(void)
570 return register_qdisc(&fq_pie_qdisc_ops);
573 static void __exit fq_pie_module_exit(void)
575 unregister_qdisc(&fq_pie_qdisc_ops);
578 module_init(fq_pie_module_init);
579 module_exit(fq_pie_module_exit);
581 MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
582 MODULE_AUTHOR("Mohit P. Tahiliani");
583 MODULE_LICENSE("GPL");