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/module.h>
14#include <linux/sizes.h>
15#include <linux/vmalloc.h>
16#include <net/pkt_cls.h>
17#include <net/pie.h>
18
19/* Flow Queue PIE
20 *
21 * Principles:
22 * - Packets are classified on flows.
23 * - This is a Stochastic model (as we use a hash, several flows might
24 * be hashed to the same slot)
25 * - Each flow has a PIE managed queue.
26 * - Flows are linked onto two (Round Robin) lists,
27 * so that new flows have priority on old ones.
28 * - For a given flow, packets are not reordered.
29 * - Drops during enqueue only.
30 * - ECN capability is off by default.
31 * - ECN threshold (if ECN is enabled) is at 10% by default.
32 * - Uses timestamps to calculate queue delay by default.
33 */
34
35/**
36 * struct fq_pie_flow - contains data for each flow
37 * @vars: pie vars associated with the flow
38 * @deficit: number of remaining byte credits
39 * @backlog: size of data in the flow
40 * @qlen: number of packets in the flow
41 * @flowchain: flowchain for the flow
42 * @head: first packet in the flow
43 * @tail: last packet in the flow
44 */
45struct fq_pie_flow {
46 struct pie_vars vars;
47 s32 deficit;
48 u32 backlog;
49 u32 qlen;
50 struct list_head flowchain;
51 struct sk_buff *head;
52 struct sk_buff *tail;
53};
54
55struct fq_pie_sched_data {
56 struct tcf_proto __rcu *filter_list; /* optional external classifier */
57 struct tcf_block *block;
58 struct fq_pie_flow *flows;
59 struct Qdisc *sch;
60 struct list_head old_flows;
61 struct list_head new_flows;
62 struct pie_params p_params;
63 u32 ecn_prob;
64 u32 flows_cnt;
65 u32 flows_cursor;
66 u32 quantum;
67 u32 memory_limit;
68 u32 new_flow_count;
69 u32 memory_usage;
70 u32 overmemory;
71 struct pie_stats stats;
72 struct timer_list adapt_timer;
73};
74
75static unsigned int fq_pie_hash(const struct fq_pie_sched_data *q,
76 struct sk_buff *skb)
77{
78 return reciprocal_scale(val: skb_get_hash(skb), ep_ro: q->flows_cnt);
79}
80
81static unsigned int fq_pie_classify(struct sk_buff *skb, struct Qdisc *sch,
82 int *qerr)
83{
84 struct fq_pie_sched_data *q = qdisc_priv(sch);
85 struct tcf_proto *filter;
86 struct tcf_result res;
87 int result;
88
89 if (TC_H_MAJ(skb->priority) == sch->handle &&
90 TC_H_MIN(skb->priority) > 0 &&
91 TC_H_MIN(skb->priority) <= q->flows_cnt)
92 return TC_H_MIN(skb->priority);
93
94 filter = rcu_dereference_bh(q->filter_list);
95 if (!filter)
96 return fq_pie_hash(q, skb) + 1;
97
98 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
99 result = tcf_classify(skb, NULL, tp: filter, res: &res, compat_mode: false);
100 if (result >= 0) {
101#ifdef CONFIG_NET_CLS_ACT
102 switch (result) {
103 case TC_ACT_STOLEN:
104 case TC_ACT_QUEUED:
105 case TC_ACT_TRAP:
106 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
107 fallthrough;
108 case TC_ACT_SHOT:
109 return 0;
110 }
111#endif
112 if (TC_H_MIN(res.classid) <= q->flows_cnt)
113 return TC_H_MIN(res.classid);
114 }
115 return 0;
116}
117
118/* add skb to flow queue (tail add) */
119static inline void flow_queue_add(struct fq_pie_flow *flow,
120 struct sk_buff *skb)
121{
122 if (!flow->head)
123 flow->head = skb;
124 else
125 flow->tail->next = skb;
126 flow->tail = skb;
127 skb->next = NULL;
128}
129
130static int fq_pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
131 struct sk_buff **to_free)
132{
133 struct fq_pie_sched_data *q = qdisc_priv(sch);
134 struct fq_pie_flow *sel_flow;
135 int ret;
136 u8 memory_limited = false;
137 u8 enqueue = false;
138 u32 pkt_len;
139 u32 idx;
140
141 /* Classifies packet into corresponding flow */
142 idx = fq_pie_classify(skb, sch, qerr: &ret);
143 if (idx == 0) {
144 if (ret & __NET_XMIT_BYPASS)
145 qdisc_qstats_drop(sch);
146 __qdisc_drop(skb, to_free);
147 return ret;
148 }
149 idx--;
150
151 sel_flow = &q->flows[idx];
152 /* Checks whether adding a new packet would exceed memory limit */
153 get_pie_cb(skb)->mem_usage = skb->truesize;
154 memory_limited = q->memory_usage > q->memory_limit + skb->truesize;
155
156 /* Checks if the qdisc is full */
157 if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
158 q->stats.overlimit++;
159 goto out;
160 } else if (unlikely(memory_limited)) {
161 q->overmemory++;
162 }
163
164 if (!pie_drop_early(sch, params: &q->p_params, vars: &sel_flow->vars,
165 backlog: sel_flow->backlog, packet_size: skb->len)) {
166 enqueue = true;
167 } else if (q->p_params.ecn &&
168 sel_flow->vars.prob <= (MAX_PROB / 100) * q->ecn_prob &&
169 INET_ECN_set_ce(skb)) {
170 /* If packet is ecn capable, mark it if drop probability
171 * is lower than the parameter ecn_prob, else drop it.
172 */
173 q->stats.ecn_mark++;
174 enqueue = true;
175 }
176 if (enqueue) {
177 /* Set enqueue time only when dq_rate_estimator is disabled. */
178 if (!q->p_params.dq_rate_estimator)
179 pie_set_enqueue_time(skb);
180
181 pkt_len = qdisc_pkt_len(skb);
182 q->stats.packets_in++;
183 q->memory_usage += skb->truesize;
184 sch->qstats.backlog += pkt_len;
185 sch->q.qlen++;
186 flow_queue_add(flow: sel_flow, skb);
187 if (list_empty(head: &sel_flow->flowchain)) {
188 list_add_tail(new: &sel_flow->flowchain, head: &q->new_flows);
189 q->new_flow_count++;
190 sel_flow->deficit = q->quantum;
191 sel_flow->qlen = 0;
192 sel_flow->backlog = 0;
193 }
194 sel_flow->qlen++;
195 sel_flow->backlog += pkt_len;
196 return NET_XMIT_SUCCESS;
197 }
198out:
199 q->stats.dropped++;
200 sel_flow->vars.accu_prob = 0;
201 __qdisc_drop(skb, to_free);
202 qdisc_qstats_drop(sch);
203 return NET_XMIT_CN;
204}
205
206static const struct netlink_range_validation fq_pie_q_range = {
207 .min = 1,
208 .max = 1 << 20,
209};
210
211static const struct nla_policy fq_pie_policy[TCA_FQ_PIE_MAX + 1] = {
212 [TCA_FQ_PIE_LIMIT] = {.type = NLA_U32},
213 [TCA_FQ_PIE_FLOWS] = {.type = NLA_U32},
214 [TCA_FQ_PIE_TARGET] = {.type = NLA_U32},
215 [TCA_FQ_PIE_TUPDATE] = {.type = NLA_U32},
216 [TCA_FQ_PIE_ALPHA] = {.type = NLA_U32},
217 [TCA_FQ_PIE_BETA] = {.type = NLA_U32},
218 [TCA_FQ_PIE_QUANTUM] =
219 NLA_POLICY_FULL_RANGE(NLA_U32, &fq_pie_q_range),
220 [TCA_FQ_PIE_MEMORY_LIMIT] = {.type = NLA_U32},
221 [TCA_FQ_PIE_ECN_PROB] = {.type = NLA_U32},
222 [TCA_FQ_PIE_ECN] = {.type = NLA_U32},
223 [TCA_FQ_PIE_BYTEMODE] = {.type = NLA_U32},
224 [TCA_FQ_PIE_DQ_RATE_ESTIMATOR] = {.type = NLA_U32},
225};
226
227static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
228{
229 struct sk_buff *skb = flow->head;
230
231 flow->head = skb->next;
232 skb->next = NULL;
233 return skb;
234}
235
236static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
237{
238 struct fq_pie_sched_data *q = qdisc_priv(sch);
239 struct sk_buff *skb = NULL;
240 struct fq_pie_flow *flow;
241 struct list_head *head;
242 u32 pkt_len;
243
244begin:
245 head = &q->new_flows;
246 if (list_empty(head)) {
247 head = &q->old_flows;
248 if (list_empty(head))
249 return NULL;
250 }
251
252 flow = list_first_entry(head, struct fq_pie_flow, flowchain);
253 /* Flow has exhausted all its credits */
254 if (flow->deficit <= 0) {
255 flow->deficit += q->quantum;
256 list_move_tail(list: &flow->flowchain, head: &q->old_flows);
257 goto begin;
258 }
259
260 if (flow->head) {
261 skb = dequeue_head(flow);
262 pkt_len = qdisc_pkt_len(skb);
263 sch->qstats.backlog -= pkt_len;
264 sch->q.qlen--;
265 qdisc_bstats_update(sch, skb);
266 }
267
268 if (!skb) {
269 /* force a pass through old_flows to prevent starvation */
270 if (head == &q->new_flows && !list_empty(head: &q->old_flows))
271 list_move_tail(list: &flow->flowchain, head: &q->old_flows);
272 else
273 list_del_init(entry: &flow->flowchain);
274 goto begin;
275 }
276
277 flow->qlen--;
278 flow->deficit -= pkt_len;
279 flow->backlog -= pkt_len;
280 q->memory_usage -= get_pie_cb(skb)->mem_usage;
281 pie_process_dequeue(skb, params: &q->p_params, vars: &flow->vars, backlog: flow->backlog);
282 return skb;
283}
284
285static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
286 struct netlink_ext_ack *extack)
287{
288 struct fq_pie_sched_data *q = qdisc_priv(sch);
289 struct nlattr *tb[TCA_FQ_PIE_MAX + 1];
290 unsigned int len_dropped = 0;
291 unsigned int num_dropped = 0;
292 int err;
293
294 err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, nla: opt, policy: fq_pie_policy, extack);
295 if (err < 0)
296 return err;
297
298 sch_tree_lock(q: sch);
299 if (tb[TCA_FQ_PIE_LIMIT]) {
300 u32 limit = nla_get_u32(nla: tb[TCA_FQ_PIE_LIMIT]);
301
302 q->p_params.limit = limit;
303 sch->limit = limit;
304 }
305 if (tb[TCA_FQ_PIE_FLOWS]) {
306 if (q->flows) {
307 NL_SET_ERR_MSG_MOD(extack,
308 "Number of flows cannot be changed");
309 goto flow_error;
310 }
311 q->flows_cnt = nla_get_u32(nla: tb[TCA_FQ_PIE_FLOWS]);
312 if (!q->flows_cnt || q->flows_cnt > 65536) {
313 NL_SET_ERR_MSG_MOD(extack,
314 "Number of flows must range in [1..65536]");
315 goto flow_error;
316 }
317 }
318
319 /* convert from microseconds to pschedtime */
320 if (tb[TCA_FQ_PIE_TARGET]) {
321 /* target is in us */
322 u32 target = nla_get_u32(nla: tb[TCA_FQ_PIE_TARGET]);
323
324 /* convert to pschedtime */
325 q->p_params.target =
326 PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
327 }
328
329 /* tupdate is in jiffies */
330 if (tb[TCA_FQ_PIE_TUPDATE])
331 q->p_params.tupdate =
332 usecs_to_jiffies(u: nla_get_u32(nla: tb[TCA_FQ_PIE_TUPDATE]));
333
334 if (tb[TCA_FQ_PIE_ALPHA])
335 q->p_params.alpha = nla_get_u32(nla: tb[TCA_FQ_PIE_ALPHA]);
336
337 if (tb[TCA_FQ_PIE_BETA])
338 q->p_params.beta = nla_get_u32(nla: tb[TCA_FQ_PIE_BETA]);
339
340 if (tb[TCA_FQ_PIE_QUANTUM])
341 q->quantum = nla_get_u32(nla: tb[TCA_FQ_PIE_QUANTUM]);
342
343 if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
344 q->memory_limit = nla_get_u32(nla: tb[TCA_FQ_PIE_MEMORY_LIMIT]);
345
346 if (tb[TCA_FQ_PIE_ECN_PROB])
347 q->ecn_prob = nla_get_u32(nla: tb[TCA_FQ_PIE_ECN_PROB]);
348
349 if (tb[TCA_FQ_PIE_ECN])
350 q->p_params.ecn = nla_get_u32(nla: tb[TCA_FQ_PIE_ECN]);
351
352 if (tb[TCA_FQ_PIE_BYTEMODE])
353 q->p_params.bytemode = nla_get_u32(nla: tb[TCA_FQ_PIE_BYTEMODE]);
354
355 if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
356 q->p_params.dq_rate_estimator =
357 nla_get_u32(nla: tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
358
359 /* Drop excess packets if new limit is lower */
360 while (sch->q.qlen > sch->limit) {
361 struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
362
363 len_dropped += qdisc_pkt_len(skb);
364 num_dropped += 1;
365 rtnl_kfree_skbs(head: skb, tail: skb);
366 }
367 qdisc_tree_reduce_backlog(qdisc: sch, n: num_dropped, len: len_dropped);
368
369 sch_tree_unlock(q: sch);
370 return 0;
371
372flow_error:
373 sch_tree_unlock(q: sch);
374 return -EINVAL;
375}
376
377static void fq_pie_timer(struct timer_list *t)
378{
379 struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
380 unsigned long next, tupdate;
381 struct Qdisc *sch = q->sch;
382 spinlock_t *root_lock; /* to lock qdisc for probability calculations */
383 int max_cnt, i;
384
385 rcu_read_lock();
386 root_lock = qdisc_lock(qdisc: qdisc_root_sleeping(qdisc: sch));
387 spin_lock(lock: root_lock);
388
389 /* Limit this expensive loop to 2048 flows per round. */
390 max_cnt = min_t(int, q->flows_cnt - q->flows_cursor, 2048);
391 for (i = 0; i < max_cnt; i++) {
392 pie_calculate_probability(params: &q->p_params,
393 vars: &q->flows[q->flows_cursor].vars,
394 backlog: q->flows[q->flows_cursor].backlog);
395 q->flows_cursor++;
396 }
397
398 tupdate = q->p_params.tupdate;
399 next = 0;
400 if (q->flows_cursor >= q->flows_cnt) {
401 q->flows_cursor = 0;
402 next = tupdate;
403 }
404 if (tupdate)
405 mod_timer(timer: &q->adapt_timer, expires: jiffies + next);
406 spin_unlock(lock: root_lock);
407 rcu_read_unlock();
408}
409
410static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
411 struct netlink_ext_ack *extack)
412{
413 struct fq_pie_sched_data *q = qdisc_priv(sch);
414 int err;
415 u32 idx;
416
417 pie_params_init(params: &q->p_params);
418 sch->limit = 10 * 1024;
419 q->p_params.limit = sch->limit;
420 q->quantum = psched_mtu(dev: qdisc_dev(qdisc: sch));
421 q->sch = sch;
422 q->ecn_prob = 10;
423 q->flows_cnt = 1024;
424 q->memory_limit = SZ_32M;
425
426 INIT_LIST_HEAD(list: &q->new_flows);
427 INIT_LIST_HEAD(list: &q->old_flows);
428 timer_setup(&q->adapt_timer, fq_pie_timer, 0);
429
430 if (opt) {
431 err = fq_pie_change(sch, opt, extack);
432
433 if (err)
434 return err;
435 }
436
437 err = tcf_block_get(p_block: &q->block, p_filter_chain: &q->filter_list, q: sch, extack);
438 if (err)
439 goto init_failure;
440
441 q->flows = kvcalloc(n: q->flows_cnt, size: sizeof(struct fq_pie_flow),
442 GFP_KERNEL);
443 if (!q->flows) {
444 err = -ENOMEM;
445 goto init_failure;
446 }
447 for (idx = 0; idx < q->flows_cnt; idx++) {
448 struct fq_pie_flow *flow = q->flows + idx;
449
450 INIT_LIST_HEAD(list: &flow->flowchain);
451 pie_vars_init(vars: &flow->vars);
452 }
453
454 mod_timer(timer: &q->adapt_timer, expires: jiffies + HZ / 2);
455
456 return 0;
457
458init_failure:
459 q->flows_cnt = 0;
460
461 return err;
462}
463
464static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
465{
466 struct fq_pie_sched_data *q = qdisc_priv(sch);
467 struct nlattr *opts;
468
469 opts = nla_nest_start(skb, attrtype: TCA_OPTIONS);
470 if (!opts)
471 return -EMSGSIZE;
472
473 /* convert target from pschedtime to us */
474 if (nla_put_u32(skb, attrtype: TCA_FQ_PIE_LIMIT, value: sch->limit) ||
475 nla_put_u32(skb, attrtype: TCA_FQ_PIE_FLOWS, value: q->flows_cnt) ||
476 nla_put_u32(skb, attrtype: TCA_FQ_PIE_TARGET,
477 value: ((u32)PSCHED_TICKS2NS(q->p_params.target)) /
478 NSEC_PER_USEC) ||
479 nla_put_u32(skb, attrtype: TCA_FQ_PIE_TUPDATE,
480 value: jiffies_to_usecs(j: q->p_params.tupdate)) ||
481 nla_put_u32(skb, attrtype: TCA_FQ_PIE_ALPHA, value: q->p_params.alpha) ||
482 nla_put_u32(skb, attrtype: TCA_FQ_PIE_BETA, value: q->p_params.beta) ||
483 nla_put_u32(skb, attrtype: TCA_FQ_PIE_QUANTUM, value: q->quantum) ||
484 nla_put_u32(skb, attrtype: TCA_FQ_PIE_MEMORY_LIMIT, value: q->memory_limit) ||
485 nla_put_u32(skb, attrtype: TCA_FQ_PIE_ECN_PROB, value: q->ecn_prob) ||
486 nla_put_u32(skb, attrtype: TCA_FQ_PIE_ECN, value: q->p_params.ecn) ||
487 nla_put_u32(skb, attrtype: TCA_FQ_PIE_BYTEMODE, value: q->p_params.bytemode) ||
488 nla_put_u32(skb, attrtype: TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
489 value: q->p_params.dq_rate_estimator))
490 goto nla_put_failure;
491
492 return nla_nest_end(skb, start: opts);
493
494nla_put_failure:
495 nla_nest_cancel(skb, start: opts);
496 return -EMSGSIZE;
497}
498
499static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
500{
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,
510 };
511 struct list_head *pos;
512
513 sch_tree_lock(q: sch);
514 list_for_each(pos, &q->new_flows)
515 st.new_flows_len++;
516
517 list_for_each(pos, &q->old_flows)
518 st.old_flows_len++;
519 sch_tree_unlock(q: sch);
520
521 return gnet_stats_copy_app(d, st: &st, len: sizeof(st));
522}
523
524static void fq_pie_reset(struct Qdisc *sch)
525{
526 struct fq_pie_sched_data *q = qdisc_priv(sch);
527 u32 idx;
528
529 INIT_LIST_HEAD(list: &q->new_flows);
530 INIT_LIST_HEAD(list: &q->old_flows);
531 for (idx = 0; idx < q->flows_cnt; idx++) {
532 struct fq_pie_flow *flow = q->flows + idx;
533
534 /* Removes all packets from flow */
535 rtnl_kfree_skbs(head: flow->head, tail: flow->tail);
536 flow->head = NULL;
537
538 INIT_LIST_HEAD(list: &flow->flowchain);
539 pie_vars_init(vars: &flow->vars);
540 }
541}
542
543static void fq_pie_destroy(struct Qdisc *sch)
544{
545 struct fq_pie_sched_data *q = qdisc_priv(sch);
546
547 tcf_block_put(block: q->block);
548 q->p_params.tupdate = 0;
549 del_timer_sync(timer: &q->adapt_timer);
550 kvfree(addr: q->flows);
551}
552
553static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
554 .id = "fq_pie",
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,
559 .init = fq_pie_init,
560 .destroy = fq_pie_destroy,
561 .reset = fq_pie_reset,
562 .change = fq_pie_change,
563 .dump = fq_pie_dump,
564 .dump_stats = fq_pie_dump_stats,
565 .owner = THIS_MODULE,
566};
567MODULE_ALIAS_NET_SCH("fq_pie");
568
569static int __init fq_pie_module_init(void)
570{
571 return register_qdisc(qops: &fq_pie_qdisc_ops);
572}
573
574static void __exit fq_pie_module_exit(void)
575{
576 unregister_qdisc(qops: &fq_pie_qdisc_ops);
577}
578
579module_init(fq_pie_module_init);
580module_exit(fq_pie_module_exit);
581
582MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
583MODULE_AUTHOR("Mohit P. Tahiliani");
584MODULE_LICENSE("GPL");
585

source code of linux/net/sched/sch_fq_pie.c