Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Fair Queue CoDel discipline
4 *
5 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
6 */
7
8#include <linux/module.h>
9#include <linux/types.h>
10#include <linux/kernel.h>
11#include <linux/jiffies.h>
12#include <linux/string.h>
13#include <linux/in.h>
14#include <linux/errno.h>
15#include <linux/init.h>
16#include <linux/skbuff.h>
17#include <linux/slab.h>
18#include <linux/vmalloc.h>
19#include <net/netlink.h>
20#include <net/pkt_sched.h>
21#include <net/pkt_cls.h>
22#include <net/codel.h>
23#include <net/codel_impl.h>
24#include <net/codel_qdisc.h>
25
26/* Fair Queue CoDel.
27 *
28 * Principles :
29 * Packets are classified (internal classifier or external) on flows.
30 * This is a Stochastic model (as we use a hash, several flows
31 * might be hashed on same slot)
32 * Each flow has a CoDel managed queue.
33 * Flows are linked onto two (Round Robin) lists,
34 * so that new flows have priority on old ones.
35 *
36 * For a given flow, packets are not reordered (CoDel uses a FIFO)
37 * head drops only.
38 * ECN capability is on by default.
39 * Low memory footprint (64 bytes per flow)
40 */
41
42struct fq_codel_flow {
43 struct sk_buff *head;
44 struct sk_buff *tail;
45 struct list_head flowchain;
46 int deficit;
47 struct codel_vars cvars;
48}; /* please try to keep this structure <= 64 bytes */
49
50struct fq_codel_sched_data {
51 struct tcf_proto __rcu *filter_list; /* optional external classifier */
52 struct tcf_block *block;
53 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
54 u32 *backlogs; /* backlog table [flows_cnt] */
55 u32 flows_cnt; /* number of flows */
56 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
57 u32 drop_batch_size;
58 u32 memory_limit;
59 struct codel_params cparams;
60 struct codel_stats cstats;
61 u32 memory_usage;
62 u32 drop_overmemory;
63 u32 drop_overlimit;
64 u32 new_flow_count;
65
66 struct list_head new_flows; /* list of new flows */
67 struct list_head old_flows; /* list of old flows */
68};
69
70static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
71 struct sk_buff *skb)
72{
73 return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
74}
75
76static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
77 int *qerr)
78{
79 struct fq_codel_sched_data *q = qdisc_priv(sch);
80 struct tcf_proto *filter;
81 struct tcf_result res;
82 int result;
83
84 if (TC_H_MAJ(skb->priority) == sch->handle &&
85 TC_H_MIN(skb->priority) > 0 &&
86 TC_H_MIN(skb->priority) <= q->flows_cnt)
87 return TC_H_MIN(skb->priority);
88
89 filter = rcu_dereference_bh(q->filter_list);
90 if (!filter)
91 return fq_codel_hash(q, skb) + 1;
92
93 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
94 result = tcf_classify(skb, filter, &res, false);
95 if (result >= 0) {
96#ifdef CONFIG_NET_CLS_ACT
97 switch (result) {
98 case TC_ACT_STOLEN:
99 case TC_ACT_QUEUED:
100 case TC_ACT_TRAP:
101 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
102 fallthrough;
103 case TC_ACT_SHOT:
104 return 0;
105 }
106#endif
107 if (TC_H_MIN(res.classid) <= q->flows_cnt)
108 return TC_H_MIN(res.classid);
109 }
110 return 0;
111}
112
113/* helper functions : might be changed when/if skb use a standard list_head */
114
115/* remove one skb from head of slot queue */
116static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
117{
118 struct sk_buff *skb = flow->head;
119
120 flow->head = skb->next;
121 skb_mark_not_on_list(skb);
122 return skb;
123}
124
125/* add skb to flow queue (tail add) */
126static inline void flow_queue_add(struct fq_codel_flow *flow,
127 struct sk_buff *skb)
128{
129 if (flow->head == NULL)
130 flow->head = skb;
131 else
132 flow->tail->next = skb;
133 flow->tail = skb;
134 skb->next = NULL;
135}
136
137static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
138 struct sk_buff **to_free)
139{
140 struct fq_codel_sched_data *q = qdisc_priv(sch);
141 struct sk_buff *skb;
142 unsigned int maxbacklog = 0, idx = 0, i, len;
143 struct fq_codel_flow *flow;
144 unsigned int threshold;
145 unsigned int mem = 0;
146
147 /* Queue is full! Find the fat flow and drop packet(s) from it.
148 * This might sound expensive, but with 1024 flows, we scan
149 * 4KB of memory, and we dont need to handle a complex tree
150 * in fast path (packet queue/enqueue) with many cache misses.
151 * In stress mode, we'll try to drop 64 packets from the flow,
152 * amortizing this linear lookup to one cache line per drop.
153 */
154 for (i = 0; i < q->flows_cnt; i++) {
155 if (q->backlogs[i] > maxbacklog) {
156 maxbacklog = q->backlogs[i];
157 idx = i;
158 }
159 }
160
161 /* Our goal is to drop half of this fat flow backlog */
162 threshold = maxbacklog >> 1;
163
164 flow = &q->flows[idx];
165 len = 0;
166 i = 0;
167 do {
168 skb = dequeue_head(flow);
169 len += qdisc_pkt_len(skb);
170 mem += get_codel_cb(skb)->mem_usage;
171 __qdisc_drop(skb, to_free);
172 } while (++i < max_packets && len < threshold);
173
174 /* Tell codel to increase its signal strength also */
175 flow->cvars.count += i;
176 q->backlogs[idx] -= len;
177 q->memory_usage -= mem;
178 sch->qstats.drops += i;
179 sch->qstats.backlog -= len;
180 sch->q.qlen -= i;
181 return idx;
182}
183
184static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
185 struct sk_buff **to_free)
186{
187 struct fq_codel_sched_data *q = qdisc_priv(sch);
188 unsigned int idx, prev_backlog, prev_qlen;
189 struct fq_codel_flow *flow;
190 int ret;
191 unsigned int pkt_len;
192 bool memory_limited;
193
194 idx = fq_codel_classify(skb, sch, &ret);
195 if (idx == 0) {
196 if (ret & __NET_XMIT_BYPASS)
197 qdisc_qstats_drop(sch);
198 __qdisc_drop(skb, to_free);
199 return ret;
200 }
201 idx--;
202
203 codel_set_enqueue_time(skb);
204 flow = &q->flows[idx];
205 flow_queue_add(flow, skb);
206 q->backlogs[idx] += qdisc_pkt_len(skb);
207 qdisc_qstats_backlog_inc(sch, skb);
208
209 if (list_empty(&flow->flowchain)) {
210 list_add_tail(&flow->flowchain, &q->new_flows);
211 q->new_flow_count++;
212 flow->deficit = q->quantum;
213 }
214 get_codel_cb(skb)->mem_usage = skb->truesize;
215 q->memory_usage += get_codel_cb(skb)->mem_usage;
216 memory_limited = q->memory_usage > q->memory_limit;
217 if (++sch->q.qlen <= sch->limit && !memory_limited)
218 return NET_XMIT_SUCCESS;
219
220 prev_backlog = sch->qstats.backlog;
221 prev_qlen = sch->q.qlen;
222
223 /* save this packet length as it might be dropped by fq_codel_drop() */
224 pkt_len = qdisc_pkt_len(skb);
225 /* fq_codel_drop() is quite expensive, as it performs a linear search
226 * in q->backlogs[] to find a fat flow.
227 * So instead of dropping a single packet, drop half of its backlog
228 * with a 64 packets limit to not add a too big cpu spike here.
229 */
230 ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
231
232 prev_qlen -= sch->q.qlen;
233 prev_backlog -= sch->qstats.backlog;
234 q->drop_overlimit += prev_qlen;
235 if (memory_limited)
236 q->drop_overmemory += prev_qlen;
237
238 /* As we dropped packet(s), better let upper stack know this.
239 * If we dropped a packet for this flow, return NET_XMIT_CN,
240 * but in this case, our parents wont increase their backlogs.
241 */
242 if (ret == idx) {
243 qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
244 prev_backlog - pkt_len);
245 return NET_XMIT_CN;
246 }
247 qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
248 return NET_XMIT_SUCCESS;
249}
250
251/* This is the specific function called from codel_dequeue()
252 * to dequeue a packet from queue. Note: backlog is handled in
253 * codel, we dont need to reduce it here.
254 */
255static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
256{
257 struct Qdisc *sch = ctx;
258 struct fq_codel_sched_data *q = qdisc_priv(sch);
259 struct fq_codel_flow *flow;
260 struct sk_buff *skb = NULL;
261
262 flow = container_of(vars, struct fq_codel_flow, cvars);
263 if (flow->head) {
264 skb = dequeue_head(flow);
265 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
266 q->memory_usage -= get_codel_cb(skb)->mem_usage;
267 sch->q.qlen--;
268 sch->qstats.backlog -= qdisc_pkt_len(skb);
269 }
270 return skb;
271}
272
273static void drop_func(struct sk_buff *skb, void *ctx)
274{
275 struct Qdisc *sch = ctx;
276
277 kfree_skb(skb);
278 qdisc_qstats_drop(sch);
279}
280
281static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
282{
283 struct fq_codel_sched_data *q = qdisc_priv(sch);
284 struct sk_buff *skb;
285 struct fq_codel_flow *flow;
286 struct list_head *head;
287
288begin:
289 head = &q->new_flows;
290 if (list_empty(head)) {
291 head = &q->old_flows;
292 if (list_empty(head))
293 return NULL;
294 }
295 flow = list_first_entry(head, struct fq_codel_flow, flowchain);
296
297 if (flow->deficit <= 0) {
298 flow->deficit += q->quantum;
299 list_move_tail(&flow->flowchain, &q->old_flows);
300 goto begin;
301 }
302
303 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
304 &flow->cvars, &q->cstats, qdisc_pkt_len,
305 codel_get_enqueue_time, drop_func, dequeue_func);
306
307 if (!skb) {
308 /* force a pass through old_flows to prevent starvation */
309 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
310 list_move_tail(&flow->flowchain, &q->old_flows);
311 else
312 list_del_init(&flow->flowchain);
313 goto begin;
314 }
315 qdisc_bstats_update(sch, skb);
316 flow->deficit -= qdisc_pkt_len(skb);
317 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
318 * or HTB crashes. Defer it for next round.
319 */
320 if (q->cstats.drop_count && sch->q.qlen) {
321 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
322 q->cstats.drop_len);
323 q->cstats.drop_count = 0;
324 q->cstats.drop_len = 0;
325 }
326 return skb;
327}
328
329static void fq_codel_flow_purge(struct fq_codel_flow *flow)
330{
331 rtnl_kfree_skbs(flow->head, flow->tail);
332 flow->head = NULL;
333}
334
335static void fq_codel_reset(struct Qdisc *sch)
336{
337 struct fq_codel_sched_data *q = qdisc_priv(sch);
338 int i;
339
340 INIT_LIST_HEAD(&q->new_flows);
341 INIT_LIST_HEAD(&q->old_flows);
342 for (i = 0; i < q->flows_cnt; i++) {
343 struct fq_codel_flow *flow = q->flows + i;
344
345 fq_codel_flow_purge(flow);
346 INIT_LIST_HEAD(&flow->flowchain);
347 codel_vars_init(&flow->cvars);
348 }
349 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
350 sch->q.qlen = 0;
351 sch->qstats.backlog = 0;
352 q->memory_usage = 0;
353}
354
355static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
356 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
357 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
358 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
359 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
360 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
361 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
362 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
363 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
364 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
365};
366
367static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
368 struct netlink_ext_ack *extack)
369{
370 struct fq_codel_sched_data *q = qdisc_priv(sch);
371 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
372 int err;
373
374 if (!opt)
375 return -EINVAL;
376
377 err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
378 fq_codel_policy, NULL);
379 if (err < 0)
380 return err;
381 if (tb[TCA_FQ_CODEL_FLOWS]) {
382 if (q->flows)
383 return -EINVAL;
384 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
385 if (!q->flows_cnt ||
386 q->flows_cnt > 65536)
387 return -EINVAL;
388 }
389 sch_tree_lock(sch);
390
391 if (tb[TCA_FQ_CODEL_TARGET]) {
392 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
393
394 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
395 }
396
397 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
398 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
399
400 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
401 }
402
403 if (tb[TCA_FQ_CODEL_INTERVAL]) {
404 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
405
406 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
407 }
408
409 if (tb[TCA_FQ_CODEL_LIMIT])
410 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
411
412 if (tb[TCA_FQ_CODEL_ECN])
413 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
414
415 if (tb[TCA_FQ_CODEL_QUANTUM])
416 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
417
418 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
419 q->drop_batch_size = max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
420
421 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
422 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
423
424 while (sch->q.qlen > sch->limit ||
425 q->memory_usage > q->memory_limit) {
426 struct sk_buff *skb = fq_codel_dequeue(sch);
427
428 q->cstats.drop_len += qdisc_pkt_len(skb);
429 rtnl_kfree_skbs(skb, skb);
430 q->cstats.drop_count++;
431 }
432 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
433 q->cstats.drop_count = 0;
434 q->cstats.drop_len = 0;
435
436 sch_tree_unlock(sch);
437 return 0;
438}
439
440static void fq_codel_destroy(struct Qdisc *sch)
441{
442 struct fq_codel_sched_data *q = qdisc_priv(sch);
443
444 tcf_block_put(q->block);
445 kvfree(q->backlogs);
446 kvfree(q->flows);
447}
448
449static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
450 struct netlink_ext_ack *extack)
451{
452 struct fq_codel_sched_data *q = qdisc_priv(sch);
453 int i;
454 int err;
455
456 sch->limit = 10*1024;
457 q->flows_cnt = 1024;
458 q->memory_limit = 32 << 20; /* 32 MBytes */
459 q->drop_batch_size = 64;
460 q->quantum = psched_mtu(qdisc_dev(sch));
461 INIT_LIST_HEAD(&q->new_flows);
462 INIT_LIST_HEAD(&q->old_flows);
463 codel_params_init(&q->cparams);
464 codel_stats_init(&q->cstats);
465 q->cparams.ecn = true;
466 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
467
468 if (opt) {
469 err = fq_codel_change(sch, opt, extack);
470 if (err)
471 goto init_failure;
472 }
473
474 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
475 if (err)
476 goto init_failure;
477
478 if (!q->flows) {
479 q->flows = kvcalloc(q->flows_cnt,
480 sizeof(struct fq_codel_flow),
481 GFP_KERNEL);
482 if (!q->flows) {
483 err = -ENOMEM;
484 goto init_failure;
485 }
486 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
487 if (!q->backlogs) {
488 err = -ENOMEM;
489 goto alloc_failure;
490 }
491 for (i = 0; i < q->flows_cnt; i++) {
492 struct fq_codel_flow *flow = q->flows + i;
493
494 INIT_LIST_HEAD(&flow->flowchain);
495 codel_vars_init(&flow->cvars);
496 }
497 }
498 if (sch->limit >= 1)
499 sch->flags |= TCQ_F_CAN_BYPASS;
500 else
501 sch->flags &= ~TCQ_F_CAN_BYPASS;
502 return 0;
503
504alloc_failure:
505 kvfree(q->flows);
506 q->flows = NULL;
507init_failure:
508 q->flows_cnt = 0;
509 return err;
510}
511
512static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
513{
514 struct fq_codel_sched_data *q = qdisc_priv(sch);
515 struct nlattr *opts;
516
517 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
518 if (opts == NULL)
519 goto nla_put_failure;
520
521 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
522 codel_time_to_us(q->cparams.target)) ||
523 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
524 sch->limit) ||
525 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
526 codel_time_to_us(q->cparams.interval)) ||
527 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
528 q->cparams.ecn) ||
529 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
530 q->quantum) ||
531 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
532 q->drop_batch_size) ||
533 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
534 q->memory_limit) ||
535 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
536 q->flows_cnt))
537 goto nla_put_failure;
538
539 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
540 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
541 codel_time_to_us(q->cparams.ce_threshold)))
542 goto nla_put_failure;
543
544 return nla_nest_end(skb, opts);
545
546nla_put_failure:
547 return -1;
548}
549
550static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
551{
552 struct fq_codel_sched_data *q = qdisc_priv(sch);
553 struct tc_fq_codel_xstats st = {
554 .type = TCA_FQ_CODEL_XSTATS_QDISC,
555 };
556 struct list_head *pos;
557
558 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
559 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
560 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
561 st.qdisc_stats.new_flow_count = q->new_flow_count;
562 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
563 st.qdisc_stats.memory_usage = q->memory_usage;
564 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
565
566 sch_tree_lock(sch);
567 list_for_each(pos, &q->new_flows)
568 st.qdisc_stats.new_flows_len++;
569
570 list_for_each(pos, &q->old_flows)
571 st.qdisc_stats.old_flows_len++;
572 sch_tree_unlock(sch);
573
574 return gnet_stats_copy_app(d, &st, sizeof(st));
575}
576
577static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
578{
579 return NULL;
580}
581
582static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
583{
584 return 0;
585}
586
587static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
588 u32 classid)
589{
590 return 0;
591}
592
593static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
594{
595}
596
597static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
598 struct netlink_ext_ack *extack)
599{
600 struct fq_codel_sched_data *q = qdisc_priv(sch);
601
602 if (cl)
603 return NULL;
604 return q->block;
605}
606
607static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
608 struct sk_buff *skb, struct tcmsg *tcm)
609{
610 tcm->tcm_handle |= TC_H_MIN(cl);
611 return 0;
612}
613
614static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
615 struct gnet_dump *d)
616{
617 struct fq_codel_sched_data *q = qdisc_priv(sch);
618 u32 idx = cl - 1;
619 struct gnet_stats_queue qs = { 0 };
620 struct tc_fq_codel_xstats xstats;
621
622 if (idx < q->flows_cnt) {
623 const struct fq_codel_flow *flow = &q->flows[idx];
624 const struct sk_buff *skb;
625
626 memset(&xstats, 0, sizeof(xstats));
627 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
628 xstats.class_stats.deficit = flow->deficit;
629 xstats.class_stats.ldelay =
630 codel_time_to_us(flow->cvars.ldelay);
631 xstats.class_stats.count = flow->cvars.count;
632 xstats.class_stats.lastcount = flow->cvars.lastcount;
633 xstats.class_stats.dropping = flow->cvars.dropping;
634 if (flow->cvars.dropping) {
635 codel_tdiff_t delta = flow->cvars.drop_next -
636 codel_get_time();
637
638 xstats.class_stats.drop_next = (delta >= 0) ?
639 codel_time_to_us(delta) :
640 -codel_time_to_us(-delta);
641 }
642 if (flow->head) {
643 sch_tree_lock(sch);
644 skb = flow->head;
645 while (skb) {
646 qs.qlen++;
647 skb = skb->next;
648 }
649 sch_tree_unlock(sch);
650 }
651 qs.backlog = q->backlogs[idx];
652 qs.drops = 0;
653 }
654 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
655 return -1;
656 if (idx < q->flows_cnt)
657 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
658 return 0;
659}
660
661static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
662{
663 struct fq_codel_sched_data *q = qdisc_priv(sch);
664 unsigned int i;
665
666 if (arg->stop)
667 return;
668
669 for (i = 0; i < q->flows_cnt; i++) {
670 if (list_empty(&q->flows[i].flowchain) ||
671 arg->count < arg->skip) {
672 arg->count++;
673 continue;
674 }
675 if (arg->fn(sch, i + 1, arg) < 0) {
676 arg->stop = 1;
677 break;
678 }
679 arg->count++;
680 }
681}
682
683static const struct Qdisc_class_ops fq_codel_class_ops = {
684 .leaf = fq_codel_leaf,
685 .find = fq_codel_find,
686 .tcf_block = fq_codel_tcf_block,
687 .bind_tcf = fq_codel_bind,
688 .unbind_tcf = fq_codel_unbind,
689 .dump = fq_codel_dump_class,
690 .dump_stats = fq_codel_dump_class_stats,
691 .walk = fq_codel_walk,
692};
693
694static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
695 .cl_ops = &fq_codel_class_ops,
696 .id = "fq_codel",
697 .priv_size = sizeof(struct fq_codel_sched_data),
698 .enqueue = fq_codel_enqueue,
699 .dequeue = fq_codel_dequeue,
700 .peek = qdisc_peek_dequeued,
701 .init = fq_codel_init,
702 .reset = fq_codel_reset,
703 .destroy = fq_codel_destroy,
704 .change = fq_codel_change,
705 .dump = fq_codel_dump,
706 .dump_stats = fq_codel_dump_stats,
707 .owner = THIS_MODULE,
708};
709
710static int __init fq_codel_module_init(void)
711{
712 return register_qdisc(&fq_codel_qdisc_ops);
713}
714
715static void __exit fq_codel_module_exit(void)
716{
717 unregister_qdisc(&fq_codel_qdisc_ops);
718}
719
720module_init(fq_codel_module_init)
721module_exit(fq_codel_module_exit)
722MODULE_AUTHOR("Eric Dumazet");
723MODULE_LICENSE("GPL");
724MODULE_DESCRIPTION("Fair Queue CoDel discipline");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Fair Queue CoDel discipline
4 *
5 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
6 */
7
8#include <linux/module.h>
9#include <linux/types.h>
10#include <linux/kernel.h>
11#include <linux/jiffies.h>
12#include <linux/string.h>
13#include <linux/in.h>
14#include <linux/errno.h>
15#include <linux/init.h>
16#include <linux/skbuff.h>
17#include <linux/jhash.h>
18#include <linux/slab.h>
19#include <linux/vmalloc.h>
20#include <net/netlink.h>
21#include <net/pkt_sched.h>
22#include <net/pkt_cls.h>
23#include <net/codel.h>
24#include <net/codel_impl.h>
25#include <net/codel_qdisc.h>
26
27/* Fair Queue CoDel.
28 *
29 * Principles :
30 * Packets are classified (internal classifier or external) on flows.
31 * This is a Stochastic model (as we use a hash, several flows
32 * might be hashed on same slot)
33 * Each flow has a CoDel managed queue.
34 * Flows are linked onto two (Round Robin) lists,
35 * so that new flows have priority on old ones.
36 *
37 * For a given flow, packets are not reordered (CoDel uses a FIFO)
38 * head drops only.
39 * ECN capability is on by default.
40 * Low memory footprint (64 bytes per flow)
41 */
42
43struct fq_codel_flow {
44 struct sk_buff *head;
45 struct sk_buff *tail;
46 struct list_head flowchain;
47 int deficit;
48 struct codel_vars cvars;
49}; /* please try to keep this structure <= 64 bytes */
50
51struct fq_codel_sched_data {
52 struct tcf_proto __rcu *filter_list; /* optional external classifier */
53 struct tcf_block *block;
54 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
55 u32 *backlogs; /* backlog table [flows_cnt] */
56 u32 flows_cnt; /* number of flows */
57 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
58 u32 drop_batch_size;
59 u32 memory_limit;
60 struct codel_params cparams;
61 struct codel_stats cstats;
62 u32 memory_usage;
63 u32 drop_overmemory;
64 u32 drop_overlimit;
65 u32 new_flow_count;
66
67 struct list_head new_flows; /* list of new flows */
68 struct list_head old_flows; /* list of old flows */
69};
70
71static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
72 struct sk_buff *skb)
73{
74 return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
75}
76
77static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
78 int *qerr)
79{
80 struct fq_codel_sched_data *q = qdisc_priv(sch);
81 struct tcf_proto *filter;
82 struct tcf_result res;
83 int result;
84
85 if (TC_H_MAJ(skb->priority) == sch->handle &&
86 TC_H_MIN(skb->priority) > 0 &&
87 TC_H_MIN(skb->priority) <= q->flows_cnt)
88 return TC_H_MIN(skb->priority);
89
90 filter = rcu_dereference_bh(q->filter_list);
91 if (!filter)
92 return fq_codel_hash(q, skb) + 1;
93
94 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
95 result = tcf_classify(skb, filter, &res, false);
96 if (result >= 0) {
97#ifdef CONFIG_NET_CLS_ACT
98 switch (result) {
99 case TC_ACT_STOLEN:
100 case TC_ACT_QUEUED:
101 case TC_ACT_TRAP:
102 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
103 /* fall through */
104 case TC_ACT_SHOT:
105 return 0;
106 }
107#endif
108 if (TC_H_MIN(res.classid) <= q->flows_cnt)
109 return TC_H_MIN(res.classid);
110 }
111 return 0;
112}
113
114/* helper functions : might be changed when/if skb use a standard list_head */
115
116/* remove one skb from head of slot queue */
117static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
118{
119 struct sk_buff *skb = flow->head;
120
121 flow->head = skb->next;
122 skb_mark_not_on_list(skb);
123 return skb;
124}
125
126/* add skb to flow queue (tail add) */
127static inline void flow_queue_add(struct fq_codel_flow *flow,
128 struct sk_buff *skb)
129{
130 if (flow->head == NULL)
131 flow->head = skb;
132 else
133 flow->tail->next = skb;
134 flow->tail = skb;
135 skb->next = NULL;
136}
137
138static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
139 struct sk_buff **to_free)
140{
141 struct fq_codel_sched_data *q = qdisc_priv(sch);
142 struct sk_buff *skb;
143 unsigned int maxbacklog = 0, idx = 0, i, len;
144 struct fq_codel_flow *flow;
145 unsigned int threshold;
146 unsigned int mem = 0;
147
148 /* Queue is full! Find the fat flow and drop packet(s) from it.
149 * This might sound expensive, but with 1024 flows, we scan
150 * 4KB of memory, and we dont need to handle a complex tree
151 * in fast path (packet queue/enqueue) with many cache misses.
152 * In stress mode, we'll try to drop 64 packets from the flow,
153 * amortizing this linear lookup to one cache line per drop.
154 */
155 for (i = 0; i < q->flows_cnt; i++) {
156 if (q->backlogs[i] > maxbacklog) {
157 maxbacklog = q->backlogs[i];
158 idx = i;
159 }
160 }
161
162 /* Our goal is to drop half of this fat flow backlog */
163 threshold = maxbacklog >> 1;
164
165 flow = &q->flows[idx];
166 len = 0;
167 i = 0;
168 do {
169 skb = dequeue_head(flow);
170 len += qdisc_pkt_len(skb);
171 mem += get_codel_cb(skb)->mem_usage;
172 __qdisc_drop(skb, to_free);
173 } while (++i < max_packets && len < threshold);
174
175 /* Tell codel to increase its signal strength also */
176 flow->cvars.count += i;
177 q->backlogs[idx] -= len;
178 q->memory_usage -= mem;
179 sch->qstats.drops += i;
180 sch->qstats.backlog -= len;
181 sch->q.qlen -= i;
182 return idx;
183}
184
185static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
186 struct sk_buff **to_free)
187{
188 struct fq_codel_sched_data *q = qdisc_priv(sch);
189 unsigned int idx, prev_backlog, prev_qlen;
190 struct fq_codel_flow *flow;
191 int uninitialized_var(ret);
192 unsigned int pkt_len;
193 bool memory_limited;
194
195 idx = fq_codel_classify(skb, sch, &ret);
196 if (idx == 0) {
197 if (ret & __NET_XMIT_BYPASS)
198 qdisc_qstats_drop(sch);
199 __qdisc_drop(skb, to_free);
200 return ret;
201 }
202 idx--;
203
204 codel_set_enqueue_time(skb);
205 flow = &q->flows[idx];
206 flow_queue_add(flow, skb);
207 q->backlogs[idx] += qdisc_pkt_len(skb);
208 qdisc_qstats_backlog_inc(sch, skb);
209
210 if (list_empty(&flow->flowchain)) {
211 list_add_tail(&flow->flowchain, &q->new_flows);
212 q->new_flow_count++;
213 flow->deficit = q->quantum;
214 }
215 get_codel_cb(skb)->mem_usage = skb->truesize;
216 q->memory_usage += get_codel_cb(skb)->mem_usage;
217 memory_limited = q->memory_usage > q->memory_limit;
218 if (++sch->q.qlen <= sch->limit && !memory_limited)
219 return NET_XMIT_SUCCESS;
220
221 prev_backlog = sch->qstats.backlog;
222 prev_qlen = sch->q.qlen;
223
224 /* save this packet length as it might be dropped by fq_codel_drop() */
225 pkt_len = qdisc_pkt_len(skb);
226 /* fq_codel_drop() is quite expensive, as it performs a linear search
227 * in q->backlogs[] to find a fat flow.
228 * So instead of dropping a single packet, drop half of its backlog
229 * with a 64 packets limit to not add a too big cpu spike here.
230 */
231 ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
232
233 prev_qlen -= sch->q.qlen;
234 prev_backlog -= sch->qstats.backlog;
235 q->drop_overlimit += prev_qlen;
236 if (memory_limited)
237 q->drop_overmemory += prev_qlen;
238
239 /* As we dropped packet(s), better let upper stack know this.
240 * If we dropped a packet for this flow, return NET_XMIT_CN,
241 * but in this case, our parents wont increase their backlogs.
242 */
243 if (ret == idx) {
244 qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
245 prev_backlog - pkt_len);
246 return NET_XMIT_CN;
247 }
248 qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
249 return NET_XMIT_SUCCESS;
250}
251
252/* This is the specific function called from codel_dequeue()
253 * to dequeue a packet from queue. Note: backlog is handled in
254 * codel, we dont need to reduce it here.
255 */
256static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
257{
258 struct Qdisc *sch = ctx;
259 struct fq_codel_sched_data *q = qdisc_priv(sch);
260 struct fq_codel_flow *flow;
261 struct sk_buff *skb = NULL;
262
263 flow = container_of(vars, struct fq_codel_flow, cvars);
264 if (flow->head) {
265 skb = dequeue_head(flow);
266 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
267 q->memory_usage -= get_codel_cb(skb)->mem_usage;
268 sch->q.qlen--;
269 sch->qstats.backlog -= qdisc_pkt_len(skb);
270 }
271 return skb;
272}
273
274static void drop_func(struct sk_buff *skb, void *ctx)
275{
276 struct Qdisc *sch = ctx;
277
278 kfree_skb(skb);
279 qdisc_qstats_drop(sch);
280}
281
282static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
283{
284 struct fq_codel_sched_data *q = qdisc_priv(sch);
285 struct sk_buff *skb;
286 struct fq_codel_flow *flow;
287 struct list_head *head;
288
289begin:
290 head = &q->new_flows;
291 if (list_empty(head)) {
292 head = &q->old_flows;
293 if (list_empty(head))
294 return NULL;
295 }
296 flow = list_first_entry(head, struct fq_codel_flow, flowchain);
297
298 if (flow->deficit <= 0) {
299 flow->deficit += q->quantum;
300 list_move_tail(&flow->flowchain, &q->old_flows);
301 goto begin;
302 }
303
304 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
305 &flow->cvars, &q->cstats, qdisc_pkt_len,
306 codel_get_enqueue_time, drop_func, dequeue_func);
307
308 if (!skb) {
309 /* force a pass through old_flows to prevent starvation */
310 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
311 list_move_tail(&flow->flowchain, &q->old_flows);
312 else
313 list_del_init(&flow->flowchain);
314 goto begin;
315 }
316 qdisc_bstats_update(sch, skb);
317 flow->deficit -= qdisc_pkt_len(skb);
318 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
319 * or HTB crashes. Defer it for next round.
320 */
321 if (q->cstats.drop_count && sch->q.qlen) {
322 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
323 q->cstats.drop_len);
324 q->cstats.drop_count = 0;
325 q->cstats.drop_len = 0;
326 }
327 return skb;
328}
329
330static void fq_codel_flow_purge(struct fq_codel_flow *flow)
331{
332 rtnl_kfree_skbs(flow->head, flow->tail);
333 flow->head = NULL;
334}
335
336static void fq_codel_reset(struct Qdisc *sch)
337{
338 struct fq_codel_sched_data *q = qdisc_priv(sch);
339 int i;
340
341 INIT_LIST_HEAD(&q->new_flows);
342 INIT_LIST_HEAD(&q->old_flows);
343 for (i = 0; i < q->flows_cnt; i++) {
344 struct fq_codel_flow *flow = q->flows + i;
345
346 fq_codel_flow_purge(flow);
347 INIT_LIST_HEAD(&flow->flowchain);
348 codel_vars_init(&flow->cvars);
349 }
350 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
351 sch->q.qlen = 0;
352 sch->qstats.backlog = 0;
353 q->memory_usage = 0;
354}
355
356static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
357 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
358 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
359 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
360 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
361 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
362 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
363 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
364 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
365 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
366};
367
368static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
369 struct netlink_ext_ack *extack)
370{
371 struct fq_codel_sched_data *q = qdisc_priv(sch);
372 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
373 int err;
374
375 if (!opt)
376 return -EINVAL;
377
378 err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
379 fq_codel_policy, NULL);
380 if (err < 0)
381 return err;
382 if (tb[TCA_FQ_CODEL_FLOWS]) {
383 if (q->flows)
384 return -EINVAL;
385 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
386 if (!q->flows_cnt ||
387 q->flows_cnt > 65536)
388 return -EINVAL;
389 }
390 sch_tree_lock(sch);
391
392 if (tb[TCA_FQ_CODEL_TARGET]) {
393 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
394
395 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
396 }
397
398 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
399 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
400
401 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
402 }
403
404 if (tb[TCA_FQ_CODEL_INTERVAL]) {
405 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
406
407 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
408 }
409
410 if (tb[TCA_FQ_CODEL_LIMIT])
411 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
412
413 if (tb[TCA_FQ_CODEL_ECN])
414 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
415
416 if (tb[TCA_FQ_CODEL_QUANTUM])
417 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
418
419 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
420 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
421
422 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
423 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
424
425 while (sch->q.qlen > sch->limit ||
426 q->memory_usage > q->memory_limit) {
427 struct sk_buff *skb = fq_codel_dequeue(sch);
428
429 q->cstats.drop_len += qdisc_pkt_len(skb);
430 rtnl_kfree_skbs(skb, skb);
431 q->cstats.drop_count++;
432 }
433 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
434 q->cstats.drop_count = 0;
435 q->cstats.drop_len = 0;
436
437 sch_tree_unlock(sch);
438 return 0;
439}
440
441static void fq_codel_destroy(struct Qdisc *sch)
442{
443 struct fq_codel_sched_data *q = qdisc_priv(sch);
444
445 tcf_block_put(q->block);
446 kvfree(q->backlogs);
447 kvfree(q->flows);
448}
449
450static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
451 struct netlink_ext_ack *extack)
452{
453 struct fq_codel_sched_data *q = qdisc_priv(sch);
454 int i;
455 int err;
456
457 sch->limit = 10*1024;
458 q->flows_cnt = 1024;
459 q->memory_limit = 32 << 20; /* 32 MBytes */
460 q->drop_batch_size = 64;
461 q->quantum = psched_mtu(qdisc_dev(sch));
462 INIT_LIST_HEAD(&q->new_flows);
463 INIT_LIST_HEAD(&q->old_flows);
464 codel_params_init(&q->cparams);
465 codel_stats_init(&q->cstats);
466 q->cparams.ecn = true;
467 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
468
469 if (opt) {
470 err = fq_codel_change(sch, opt, extack);
471 if (err)
472 goto init_failure;
473 }
474
475 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
476 if (err)
477 goto init_failure;
478
479 if (!q->flows) {
480 q->flows = kvcalloc(q->flows_cnt,
481 sizeof(struct fq_codel_flow),
482 GFP_KERNEL);
483 if (!q->flows) {
484 err = -ENOMEM;
485 goto init_failure;
486 }
487 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
488 if (!q->backlogs) {
489 err = -ENOMEM;
490 goto alloc_failure;
491 }
492 for (i = 0; i < q->flows_cnt; i++) {
493 struct fq_codel_flow *flow = q->flows + i;
494
495 INIT_LIST_HEAD(&flow->flowchain);
496 codel_vars_init(&flow->cvars);
497 }
498 }
499 if (sch->limit >= 1)
500 sch->flags |= TCQ_F_CAN_BYPASS;
501 else
502 sch->flags &= ~TCQ_F_CAN_BYPASS;
503 return 0;
504
505alloc_failure:
506 kvfree(q->flows);
507 q->flows = NULL;
508init_failure:
509 q->flows_cnt = 0;
510 return err;
511}
512
513static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
514{
515 struct fq_codel_sched_data *q = qdisc_priv(sch);
516 struct nlattr *opts;
517
518 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
519 if (opts == NULL)
520 goto nla_put_failure;
521
522 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
523 codel_time_to_us(q->cparams.target)) ||
524 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
525 sch->limit) ||
526 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
527 codel_time_to_us(q->cparams.interval)) ||
528 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
529 q->cparams.ecn) ||
530 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
531 q->quantum) ||
532 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
533 q->drop_batch_size) ||
534 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
535 q->memory_limit) ||
536 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
537 q->flows_cnt))
538 goto nla_put_failure;
539
540 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
541 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
542 codel_time_to_us(q->cparams.ce_threshold)))
543 goto nla_put_failure;
544
545 return nla_nest_end(skb, opts);
546
547nla_put_failure:
548 return -1;
549}
550
551static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
552{
553 struct fq_codel_sched_data *q = qdisc_priv(sch);
554 struct tc_fq_codel_xstats st = {
555 .type = TCA_FQ_CODEL_XSTATS_QDISC,
556 };
557 struct list_head *pos;
558
559 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
560 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
561 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
562 st.qdisc_stats.new_flow_count = q->new_flow_count;
563 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
564 st.qdisc_stats.memory_usage = q->memory_usage;
565 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
566
567 sch_tree_lock(sch);
568 list_for_each(pos, &q->new_flows)
569 st.qdisc_stats.new_flows_len++;
570
571 list_for_each(pos, &q->old_flows)
572 st.qdisc_stats.old_flows_len++;
573 sch_tree_unlock(sch);
574
575 return gnet_stats_copy_app(d, &st, sizeof(st));
576}
577
578static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
579{
580 return NULL;
581}
582
583static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
584{
585 return 0;
586}
587
588static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
589 u32 classid)
590{
591 return 0;
592}
593
594static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
595{
596}
597
598static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
599 struct netlink_ext_ack *extack)
600{
601 struct fq_codel_sched_data *q = qdisc_priv(sch);
602
603 if (cl)
604 return NULL;
605 return q->block;
606}
607
608static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
609 struct sk_buff *skb, struct tcmsg *tcm)
610{
611 tcm->tcm_handle |= TC_H_MIN(cl);
612 return 0;
613}
614
615static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
616 struct gnet_dump *d)
617{
618 struct fq_codel_sched_data *q = qdisc_priv(sch);
619 u32 idx = cl - 1;
620 struct gnet_stats_queue qs = { 0 };
621 struct tc_fq_codel_xstats xstats;
622
623 if (idx < q->flows_cnt) {
624 const struct fq_codel_flow *flow = &q->flows[idx];
625 const struct sk_buff *skb;
626
627 memset(&xstats, 0, sizeof(xstats));
628 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
629 xstats.class_stats.deficit = flow->deficit;
630 xstats.class_stats.ldelay =
631 codel_time_to_us(flow->cvars.ldelay);
632 xstats.class_stats.count = flow->cvars.count;
633 xstats.class_stats.lastcount = flow->cvars.lastcount;
634 xstats.class_stats.dropping = flow->cvars.dropping;
635 if (flow->cvars.dropping) {
636 codel_tdiff_t delta = flow->cvars.drop_next -
637 codel_get_time();
638
639 xstats.class_stats.drop_next = (delta >= 0) ?
640 codel_time_to_us(delta) :
641 -codel_time_to_us(-delta);
642 }
643 if (flow->head) {
644 sch_tree_lock(sch);
645 skb = flow->head;
646 while (skb) {
647 qs.qlen++;
648 skb = skb->next;
649 }
650 sch_tree_unlock(sch);
651 }
652 qs.backlog = q->backlogs[idx];
653 qs.drops = 0;
654 }
655 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
656 return -1;
657 if (idx < q->flows_cnt)
658 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
659 return 0;
660}
661
662static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
663{
664 struct fq_codel_sched_data *q = qdisc_priv(sch);
665 unsigned int i;
666
667 if (arg->stop)
668 return;
669
670 for (i = 0; i < q->flows_cnt; i++) {
671 if (list_empty(&q->flows[i].flowchain) ||
672 arg->count < arg->skip) {
673 arg->count++;
674 continue;
675 }
676 if (arg->fn(sch, i + 1, arg) < 0) {
677 arg->stop = 1;
678 break;
679 }
680 arg->count++;
681 }
682}
683
684static const struct Qdisc_class_ops fq_codel_class_ops = {
685 .leaf = fq_codel_leaf,
686 .find = fq_codel_find,
687 .tcf_block = fq_codel_tcf_block,
688 .bind_tcf = fq_codel_bind,
689 .unbind_tcf = fq_codel_unbind,
690 .dump = fq_codel_dump_class,
691 .dump_stats = fq_codel_dump_class_stats,
692 .walk = fq_codel_walk,
693};
694
695static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
696 .cl_ops = &fq_codel_class_ops,
697 .id = "fq_codel",
698 .priv_size = sizeof(struct fq_codel_sched_data),
699 .enqueue = fq_codel_enqueue,
700 .dequeue = fq_codel_dequeue,
701 .peek = qdisc_peek_dequeued,
702 .init = fq_codel_init,
703 .reset = fq_codel_reset,
704 .destroy = fq_codel_destroy,
705 .change = fq_codel_change,
706 .dump = fq_codel_dump,
707 .dump_stats = fq_codel_dump_stats,
708 .owner = THIS_MODULE,
709};
710
711static int __init fq_codel_module_init(void)
712{
713 return register_qdisc(&fq_codel_qdisc_ops);
714}
715
716static void __exit fq_codel_module_exit(void)
717{
718 unregister_qdisc(&fq_codel_qdisc_ops);
719}
720
721module_init(fq_codel_module_init)
722module_exit(fq_codel_module_exit)
723MODULE_AUTHOR("Eric Dumazet");
724MODULE_LICENSE("GPL");