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;
69 struct pie_stats stats;
70 struct timer_list adapt_timer;
73 static unsigned int fq_pie_hash(const struct fq_pie_sched_data *q,
76 return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
79 static unsigned int fq_pie_classify(struct sk_buff *skb, struct Qdisc *sch,
82 struct fq_pie_sched_data *q = qdisc_priv(sch);
83 struct tcf_proto *filter;
84 struct tcf_result res;
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);
92 filter = rcu_dereference_bh(q->filter_list);
94 return fq_pie_hash(q, skb) + 1;
96 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
97 result = tcf_classify(skb, filter, &res, false);
99 #ifdef CONFIG_NET_CLS_ACT
104 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
110 if (TC_H_MIN(res.classid) <= q->flows_cnt)
111 return TC_H_MIN(res.classid);
116 /* add skb to flow queue (tail add) */
117 static inline void flow_queue_add(struct fq_pie_flow *flow,
123 flow->tail->next = skb;
128 static int fq_pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
129 struct sk_buff **to_free)
131 struct fq_pie_sched_data *q = qdisc_priv(sch);
132 struct fq_pie_flow *sel_flow;
133 int uninitialized_var(ret);
134 u8 memory_limited = false;
139 /* Classifies packet into corresponding flow */
140 idx = fq_pie_classify(skb, sch, &ret);
141 sel_flow = &q->flows[idx];
143 /* Checks whether adding a new packet would exceed memory limit */
144 get_pie_cb(skb)->mem_usage = skb->truesize;
145 memory_limited = q->memory_usage > q->memory_limit + skb->truesize;
147 /* Checks if the qdisc is full */
148 if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
149 q->stats.overlimit++;
151 } else if (unlikely(memory_limited)) {
155 if (!pie_drop_early(sch, &q->p_params, &sel_flow->vars,
156 sel_flow->backlog, skb->len)) {
158 } else if (q->p_params.ecn &&
159 sel_flow->vars.prob <= (MAX_PROB / 100) * q->ecn_prob &&
160 INET_ECN_set_ce(skb)) {
161 /* If packet is ecn capable, mark it if drop probability
162 * is lower than the parameter ecn_prob, else drop it.
168 /* Set enqueue time only when dq_rate_estimator is disabled. */
169 if (!q->p_params.dq_rate_estimator)
170 pie_set_enqueue_time(skb);
172 pkt_len = qdisc_pkt_len(skb);
173 q->stats.packets_in++;
174 q->memory_usage += skb->truesize;
175 sch->qstats.backlog += pkt_len;
177 flow_queue_add(sel_flow, skb);
178 if (list_empty(&sel_flow->flowchain)) {
179 list_add_tail(&sel_flow->flowchain, &q->new_flows);
181 sel_flow->deficit = q->quantum;
183 sel_flow->backlog = 0;
186 sel_flow->backlog += pkt_len;
187 return NET_XMIT_SUCCESS;
191 sel_flow->vars.accu_prob = 0;
192 sel_flow->vars.accu_prob_overflows = 0;
193 __qdisc_drop(skb, to_free);
194 qdisc_qstats_drop(sch);
198 static const struct nla_policy fq_pie_policy[TCA_FQ_PIE_MAX + 1] = {
199 [TCA_FQ_PIE_LIMIT] = {.type = NLA_U32},
200 [TCA_FQ_PIE_FLOWS] = {.type = NLA_U32},
201 [TCA_FQ_PIE_TARGET] = {.type = NLA_U32},
202 [TCA_FQ_PIE_TUPDATE] = {.type = NLA_U32},
203 [TCA_FQ_PIE_ALPHA] = {.type = NLA_U32},
204 [TCA_FQ_PIE_BETA] = {.type = NLA_U32},
205 [TCA_FQ_PIE_QUANTUM] = {.type = NLA_U32},
206 [TCA_FQ_PIE_MEMORY_LIMIT] = {.type = NLA_U32},
207 [TCA_FQ_PIE_ECN_PROB] = {.type = NLA_U32},
208 [TCA_FQ_PIE_ECN] = {.type = NLA_U32},
209 [TCA_FQ_PIE_BYTEMODE] = {.type = NLA_U32},
210 [TCA_FQ_PIE_DQ_RATE_ESTIMATOR] = {.type = NLA_U32},
213 static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
215 struct sk_buff *skb = flow->head;
217 flow->head = skb->next;
222 static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
224 struct fq_pie_sched_data *q = qdisc_priv(sch);
225 struct sk_buff *skb = NULL;
226 struct fq_pie_flow *flow;
227 struct list_head *head;
231 head = &q->new_flows;
232 if (list_empty(head)) {
233 head = &q->old_flows;
234 if (list_empty(head))
238 flow = list_first_entry(head, struct fq_pie_flow, flowchain);
239 /* Flow has exhausted all its credits */
240 if (flow->deficit <= 0) {
241 flow->deficit += q->quantum;
242 list_move_tail(&flow->flowchain, &q->old_flows);
247 skb = dequeue_head(flow);
248 pkt_len = qdisc_pkt_len(skb);
249 sch->qstats.backlog -= pkt_len;
251 qdisc_bstats_update(sch, skb);
255 /* force a pass through old_flows to prevent starvation */
256 if (head == &q->new_flows && !list_empty(&q->old_flows))
257 list_move_tail(&flow->flowchain, &q->old_flows);
259 list_del_init(&flow->flowchain);
264 flow->deficit -= pkt_len;
265 flow->backlog -= pkt_len;
266 q->memory_usage -= get_pie_cb(skb)->mem_usage;
267 pie_process_dequeue(skb, &q->p_params, &flow->vars, flow->backlog);
271 static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
272 struct netlink_ext_ack *extack)
274 struct fq_pie_sched_data *q = qdisc_priv(sch);
275 struct nlattr *tb[TCA_FQ_PIE_MAX + 1];
276 unsigned int len_dropped = 0;
277 unsigned int num_dropped = 0;
283 err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, opt, fq_pie_policy, extack);
288 if (tb[TCA_FQ_PIE_LIMIT]) {
289 u32 limit = nla_get_u32(tb[TCA_FQ_PIE_LIMIT]);
291 q->p_params.limit = limit;
294 if (tb[TCA_FQ_PIE_FLOWS]) {
296 NL_SET_ERR_MSG_MOD(extack,
297 "Number of flows cannot be changed");
300 q->flows_cnt = nla_get_u32(tb[TCA_FQ_PIE_FLOWS]);
301 if (!q->flows_cnt || q->flows_cnt > 65536) {
302 NL_SET_ERR_MSG_MOD(extack,
303 "Number of flows must be < 65536");
308 /* convert from microseconds to pschedtime */
309 if (tb[TCA_FQ_PIE_TARGET]) {
310 /* target is in us */
311 u32 target = nla_get_u32(tb[TCA_FQ_PIE_TARGET]);
313 /* convert to pschedtime */
315 PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
318 /* tupdate is in jiffies */
319 if (tb[TCA_FQ_PIE_TUPDATE])
320 q->p_params.tupdate =
321 usecs_to_jiffies(nla_get_u32(tb[TCA_FQ_PIE_TUPDATE]));
323 if (tb[TCA_FQ_PIE_ALPHA])
324 q->p_params.alpha = nla_get_u32(tb[TCA_FQ_PIE_ALPHA]);
326 if (tb[TCA_FQ_PIE_BETA])
327 q->p_params.beta = nla_get_u32(tb[TCA_FQ_PIE_BETA]);
329 if (tb[TCA_FQ_PIE_QUANTUM])
330 q->quantum = nla_get_u32(tb[TCA_FQ_PIE_QUANTUM]);
332 if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
333 q->memory_limit = nla_get_u32(tb[TCA_FQ_PIE_MEMORY_LIMIT]);
335 if (tb[TCA_FQ_PIE_ECN_PROB])
336 q->ecn_prob = nla_get_u32(tb[TCA_FQ_PIE_ECN_PROB]);
338 if (tb[TCA_FQ_PIE_ECN])
339 q->p_params.ecn = nla_get_u32(tb[TCA_FQ_PIE_ECN]);
341 if (tb[TCA_FQ_PIE_BYTEMODE])
342 q->p_params.bytemode = nla_get_u32(tb[TCA_FQ_PIE_BYTEMODE]);
344 if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
345 q->p_params.dq_rate_estimator =
346 nla_get_u32(tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
348 /* Drop excess packets if new limit is lower */
349 while (sch->q.qlen > sch->limit) {
350 struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
352 len_dropped += qdisc_pkt_len(skb);
354 rtnl_kfree_skbs(skb, skb);
356 qdisc_tree_reduce_backlog(sch, num_dropped, len_dropped);
358 sch_tree_unlock(sch);
362 sch_tree_unlock(sch);
366 static void fq_pie_timer(struct timer_list *t)
368 struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
369 struct Qdisc *sch = q->sch;
370 spinlock_t *root_lock; /* to lock qdisc for probability calculations */
373 root_lock = qdisc_lock(qdisc_root_sleeping(sch));
374 spin_lock(root_lock);
376 for (idx = 0; idx < q->flows_cnt; idx++)
377 pie_calculate_probability(&q->p_params, &q->flows[idx].vars,
378 q->flows[idx].backlog);
380 /* reset the timer to fire after 'tupdate' jiffies. */
381 if (q->p_params.tupdate)
382 mod_timer(&q->adapt_timer, jiffies + q->p_params.tupdate);
384 spin_unlock(root_lock);
387 static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
388 struct netlink_ext_ack *extack)
390 struct fq_pie_sched_data *q = qdisc_priv(sch);
394 pie_params_init(&q->p_params);
395 sch->limit = 10 * 1024;
396 q->p_params.limit = sch->limit;
397 q->quantum = psched_mtu(qdisc_dev(sch));
401 q->memory_limit = SZ_32M;
403 INIT_LIST_HEAD(&q->new_flows);
404 INIT_LIST_HEAD(&q->old_flows);
407 err = fq_pie_change(sch, opt, extack);
413 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
417 q->flows = kvcalloc(q->flows_cnt, sizeof(struct fq_pie_flow),
423 for (idx = 0; idx < q->flows_cnt; idx++) {
424 struct fq_pie_flow *flow = q->flows + idx;
426 INIT_LIST_HEAD(&flow->flowchain);
427 pie_vars_init(&flow->vars);
430 timer_setup(&q->adapt_timer, fq_pie_timer, 0);
431 mod_timer(&q->adapt_timer, jiffies + HZ / 2);
441 static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
443 struct fq_pie_sched_data *q = qdisc_priv(sch);
446 opts = nla_nest_start(skb, TCA_OPTIONS);
450 /* convert target from pschedtime to us */
451 if (nla_put_u32(skb, TCA_FQ_PIE_LIMIT, sch->limit) ||
452 nla_put_u32(skb, TCA_FQ_PIE_FLOWS, q->flows_cnt) ||
453 nla_put_u32(skb, TCA_FQ_PIE_TARGET,
454 ((u32)PSCHED_TICKS2NS(q->p_params.target)) /
456 nla_put_u32(skb, TCA_FQ_PIE_TUPDATE,
457 jiffies_to_usecs(q->p_params.tupdate)) ||
458 nla_put_u32(skb, TCA_FQ_PIE_ALPHA, q->p_params.alpha) ||
459 nla_put_u32(skb, TCA_FQ_PIE_BETA, q->p_params.beta) ||
460 nla_put_u32(skb, TCA_FQ_PIE_QUANTUM, q->quantum) ||
461 nla_put_u32(skb, TCA_FQ_PIE_MEMORY_LIMIT, q->memory_limit) ||
462 nla_put_u32(skb, TCA_FQ_PIE_ECN_PROB, q->ecn_prob) ||
463 nla_put_u32(skb, TCA_FQ_PIE_ECN, q->p_params.ecn) ||
464 nla_put_u32(skb, TCA_FQ_PIE_BYTEMODE, q->p_params.bytemode) ||
465 nla_put_u32(skb, TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
466 q->p_params.dq_rate_estimator))
467 goto nla_put_failure;
469 return nla_nest_end(skb, opts);
472 nla_nest_cancel(skb, opts);
476 static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
478 struct fq_pie_sched_data *q = qdisc_priv(sch);
479 struct tc_fq_pie_xstats st = {
480 .packets_in = q->stats.packets_in,
481 .overlimit = q->stats.overlimit,
482 .overmemory = q->overmemory,
483 .dropped = q->stats.dropped,
484 .ecn_mark = q->stats.ecn_mark,
485 .new_flow_count = q->new_flow_count,
486 .memory_usage = q->memory_usage,
488 struct list_head *pos;
491 list_for_each(pos, &q->new_flows)
494 list_for_each(pos, &q->old_flows)
496 sch_tree_unlock(sch);
498 return gnet_stats_copy_app(d, &st, sizeof(st));
501 static void fq_pie_reset(struct Qdisc *sch)
503 struct fq_pie_sched_data *q = qdisc_priv(sch);
506 INIT_LIST_HEAD(&q->new_flows);
507 INIT_LIST_HEAD(&q->old_flows);
508 for (idx = 0; idx < q->flows_cnt; idx++) {
509 struct fq_pie_flow *flow = q->flows + idx;
511 /* Removes all packets from flow */
512 rtnl_kfree_skbs(flow->head, flow->tail);
515 INIT_LIST_HEAD(&flow->flowchain);
516 pie_vars_init(&flow->vars);
520 sch->qstats.backlog = 0;
523 static void fq_pie_destroy(struct Qdisc *sch)
525 struct fq_pie_sched_data *q = qdisc_priv(sch);
527 tcf_block_put(q->block);
528 del_timer_sync(&q->adapt_timer);
532 static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
534 .priv_size = sizeof(struct fq_pie_sched_data),
535 .enqueue = fq_pie_qdisc_enqueue,
536 .dequeue = fq_pie_qdisc_dequeue,
537 .peek = qdisc_peek_dequeued,
539 .destroy = fq_pie_destroy,
540 .reset = fq_pie_reset,
541 .change = fq_pie_change,
543 .dump_stats = fq_pie_dump_stats,
544 .owner = THIS_MODULE,
547 static int __init fq_pie_module_init(void)
549 return register_qdisc(&fq_pie_qdisc_ops);
552 static void __exit fq_pie_module_exit(void)
554 unregister_qdisc(&fq_pie_qdisc_ops);
557 module_init(fq_pie_module_init);
558 module_exit(fq_pie_module_exit);
560 MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
561 MODULE_AUTHOR("Mohit P. Tahiliani");
562 MODULE_LICENSE("GPL");