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, NULL, 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 q->memory_usage = 0;
351}
352
353static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
354 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
355 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
356 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
357 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
358 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
359 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
360 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
361 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
362 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
363 [TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR] = { .type = NLA_U8 },
364 [TCA_FQ_CODEL_CE_THRESHOLD_MASK] = { .type = NLA_U8 },
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 u32 quantum = 0;
373 int err;
374
375 err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
376 fq_codel_policy, NULL);
377 if (err < 0)
378 return err;
379 if (tb[TCA_FQ_CODEL_FLOWS]) {
380 if (q->flows)
381 return -EINVAL;
382 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
383 if (!q->flows_cnt ||
384 q->flows_cnt > 65536)
385 return -EINVAL;
386 }
387 if (tb[TCA_FQ_CODEL_QUANTUM]) {
388 quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
389 if (quantum > FQ_CODEL_QUANTUM_MAX) {
390 NL_SET_ERR_MSG(extack, "Invalid quantum");
391 return -EINVAL;
392 }
393 }
394 sch_tree_lock(sch);
395
396 if (tb[TCA_FQ_CODEL_TARGET]) {
397 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
398
399 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
400 }
401
402 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
403 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
404
405 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
406 }
407
408 if (tb[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR])
409 q->cparams.ce_threshold_selector = nla_get_u8(tb[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR]);
410 if (tb[TCA_FQ_CODEL_CE_THRESHOLD_MASK])
411 q->cparams.ce_threshold_mask = nla_get_u8(tb[TCA_FQ_CODEL_CE_THRESHOLD_MASK]);
412
413 if (tb[TCA_FQ_CODEL_INTERVAL]) {
414 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
415
416 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
417 }
418
419 if (tb[TCA_FQ_CODEL_LIMIT])
420 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
421
422 if (tb[TCA_FQ_CODEL_ECN])
423 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
424
425 if (quantum)
426 q->quantum = quantum;
427
428 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
429 q->drop_batch_size = max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
430
431 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
432 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
433
434 while (sch->q.qlen > sch->limit ||
435 q->memory_usage > q->memory_limit) {
436 struct sk_buff *skb = fq_codel_dequeue(sch);
437
438 q->cstats.drop_len += qdisc_pkt_len(skb);
439 rtnl_kfree_skbs(skb, skb);
440 q->cstats.drop_count++;
441 }
442 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
443 q->cstats.drop_count = 0;
444 q->cstats.drop_len = 0;
445
446 sch_tree_unlock(sch);
447 return 0;
448}
449
450static void fq_codel_destroy(struct Qdisc *sch)
451{
452 struct fq_codel_sched_data *q = qdisc_priv(sch);
453
454 tcf_block_put(q->block);
455 kvfree(q->backlogs);
456 kvfree(q->flows);
457}
458
459static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
460 struct netlink_ext_ack *extack)
461{
462 struct fq_codel_sched_data *q = qdisc_priv(sch);
463 int i;
464 int err;
465
466 sch->limit = 10*1024;
467 q->flows_cnt = 1024;
468 q->memory_limit = 32 << 20; /* 32 MBytes */
469 q->drop_batch_size = 64;
470 q->quantum = psched_mtu(qdisc_dev(sch));
471 INIT_LIST_HEAD(&q->new_flows);
472 INIT_LIST_HEAD(&q->old_flows);
473 codel_params_init(&q->cparams);
474 codel_stats_init(&q->cstats);
475 q->cparams.ecn = true;
476 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
477
478 if (opt) {
479 err = fq_codel_change(sch, opt, extack);
480 if (err)
481 goto init_failure;
482 }
483
484 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
485 if (err)
486 goto init_failure;
487
488 if (!q->flows) {
489 q->flows = kvcalloc(q->flows_cnt,
490 sizeof(struct fq_codel_flow),
491 GFP_KERNEL);
492 if (!q->flows) {
493 err = -ENOMEM;
494 goto init_failure;
495 }
496 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
497 if (!q->backlogs) {
498 err = -ENOMEM;
499 goto alloc_failure;
500 }
501 for (i = 0; i < q->flows_cnt; i++) {
502 struct fq_codel_flow *flow = q->flows + i;
503
504 INIT_LIST_HEAD(&flow->flowchain);
505 codel_vars_init(&flow->cvars);
506 }
507 }
508 if (sch->limit >= 1)
509 sch->flags |= TCQ_F_CAN_BYPASS;
510 else
511 sch->flags &= ~TCQ_F_CAN_BYPASS;
512 return 0;
513
514alloc_failure:
515 kvfree(q->flows);
516 q->flows = NULL;
517init_failure:
518 q->flows_cnt = 0;
519 return err;
520}
521
522static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
523{
524 struct fq_codel_sched_data *q = qdisc_priv(sch);
525 struct nlattr *opts;
526
527 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
528 if (opts == NULL)
529 goto nla_put_failure;
530
531 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
532 codel_time_to_us(q->cparams.target)) ||
533 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
534 sch->limit) ||
535 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
536 codel_time_to_us(q->cparams.interval)) ||
537 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
538 q->cparams.ecn) ||
539 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
540 q->quantum) ||
541 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
542 q->drop_batch_size) ||
543 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
544 q->memory_limit) ||
545 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
546 q->flows_cnt))
547 goto nla_put_failure;
548
549 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD) {
550 if (nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
551 codel_time_to_us(q->cparams.ce_threshold)))
552 goto nla_put_failure;
553 if (nla_put_u8(skb, TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR, q->cparams.ce_threshold_selector))
554 goto nla_put_failure;
555 if (nla_put_u8(skb, TCA_FQ_CODEL_CE_THRESHOLD_MASK, q->cparams.ce_threshold_mask))
556 goto nla_put_failure;
557 }
558
559 return nla_nest_end(skb, opts);
560
561nla_put_failure:
562 return -1;
563}
564
565static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
566{
567 struct fq_codel_sched_data *q = qdisc_priv(sch);
568 struct tc_fq_codel_xstats st = {
569 .type = TCA_FQ_CODEL_XSTATS_QDISC,
570 };
571 struct list_head *pos;
572
573 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
574 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
575 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
576 st.qdisc_stats.new_flow_count = q->new_flow_count;
577 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
578 st.qdisc_stats.memory_usage = q->memory_usage;
579 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
580
581 sch_tree_lock(sch);
582 list_for_each(pos, &q->new_flows)
583 st.qdisc_stats.new_flows_len++;
584
585 list_for_each(pos, &q->old_flows)
586 st.qdisc_stats.old_flows_len++;
587 sch_tree_unlock(sch);
588
589 return gnet_stats_copy_app(d, &st, sizeof(st));
590}
591
592static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
593{
594 return NULL;
595}
596
597static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
598{
599 return 0;
600}
601
602static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
603 u32 classid)
604{
605 return 0;
606}
607
608static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
609{
610}
611
612static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
613 struct netlink_ext_ack *extack)
614{
615 struct fq_codel_sched_data *q = qdisc_priv(sch);
616
617 if (cl)
618 return NULL;
619 return q->block;
620}
621
622static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
623 struct sk_buff *skb, struct tcmsg *tcm)
624{
625 tcm->tcm_handle |= TC_H_MIN(cl);
626 return 0;
627}
628
629static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
630 struct gnet_dump *d)
631{
632 struct fq_codel_sched_data *q = qdisc_priv(sch);
633 u32 idx = cl - 1;
634 struct gnet_stats_queue qs = { 0 };
635 struct tc_fq_codel_xstats xstats;
636
637 if (idx < q->flows_cnt) {
638 const struct fq_codel_flow *flow = &q->flows[idx];
639 const struct sk_buff *skb;
640
641 memset(&xstats, 0, sizeof(xstats));
642 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
643 xstats.class_stats.deficit = flow->deficit;
644 xstats.class_stats.ldelay =
645 codel_time_to_us(flow->cvars.ldelay);
646 xstats.class_stats.count = flow->cvars.count;
647 xstats.class_stats.lastcount = flow->cvars.lastcount;
648 xstats.class_stats.dropping = flow->cvars.dropping;
649 if (flow->cvars.dropping) {
650 codel_tdiff_t delta = flow->cvars.drop_next -
651 codel_get_time();
652
653 xstats.class_stats.drop_next = (delta >= 0) ?
654 codel_time_to_us(delta) :
655 -codel_time_to_us(-delta);
656 }
657 if (flow->head) {
658 sch_tree_lock(sch);
659 skb = flow->head;
660 while (skb) {
661 qs.qlen++;
662 skb = skb->next;
663 }
664 sch_tree_unlock(sch);
665 }
666 qs.backlog = q->backlogs[idx];
667 qs.drops = 0;
668 }
669 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
670 return -1;
671 if (idx < q->flows_cnt)
672 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
673 return 0;
674}
675
676static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
677{
678 struct fq_codel_sched_data *q = qdisc_priv(sch);
679 unsigned int i;
680
681 if (arg->stop)
682 return;
683
684 for (i = 0; i < q->flows_cnt; i++) {
685 if (list_empty(&q->flows[i].flowchain)) {
686 arg->count++;
687 continue;
688 }
689 if (!tc_qdisc_stats_dump(sch, i + 1, arg))
690 break;
691 }
692}
693
694static const struct Qdisc_class_ops fq_codel_class_ops = {
695 .leaf = fq_codel_leaf,
696 .find = fq_codel_find,
697 .tcf_block = fq_codel_tcf_block,
698 .bind_tcf = fq_codel_bind,
699 .unbind_tcf = fq_codel_unbind,
700 .dump = fq_codel_dump_class,
701 .dump_stats = fq_codel_dump_class_stats,
702 .walk = fq_codel_walk,
703};
704
705static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
706 .cl_ops = &fq_codel_class_ops,
707 .id = "fq_codel",
708 .priv_size = sizeof(struct fq_codel_sched_data),
709 .enqueue = fq_codel_enqueue,
710 .dequeue = fq_codel_dequeue,
711 .peek = qdisc_peek_dequeued,
712 .init = fq_codel_init,
713 .reset = fq_codel_reset,
714 .destroy = fq_codel_destroy,
715 .change = fq_codel_change,
716 .dump = fq_codel_dump,
717 .dump_stats = fq_codel_dump_stats,
718 .owner = THIS_MODULE,
719};
720
721static int __init fq_codel_module_init(void)
722{
723 return register_qdisc(&fq_codel_qdisc_ops);
724}
725
726static void __exit fq_codel_module_exit(void)
727{
728 unregister_qdisc(&fq_codel_qdisc_ops);
729}
730
731module_init(fq_codel_module_init)
732module_exit(fq_codel_module_exit)
733MODULE_AUTHOR("Eric Dumazet");
734MODULE_LICENSE("GPL");
735MODULE_DESCRIPTION("Fair Queue CoDel discipline");
1/*
2 * Fair Queue CoDel discipline
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
10 */
11
12#include <linux/module.h>
13#include <linux/types.h>
14#include <linux/kernel.h>
15#include <linux/jiffies.h>
16#include <linux/string.h>
17#include <linux/in.h>
18#include <linux/errno.h>
19#include <linux/init.h>
20#include <linux/skbuff.h>
21#include <linux/jhash.h>
22#include <linux/slab.h>
23#include <linux/vmalloc.h>
24#include <net/netlink.h>
25#include <net/pkt_sched.h>
26#include <net/codel.h>
27#include <net/codel_impl.h>
28#include <net/codel_qdisc.h>
29
30/* Fair Queue CoDel.
31 *
32 * Principles :
33 * Packets are classified (internal classifier or external) on flows.
34 * This is a Stochastic model (as we use a hash, several flows
35 * might be hashed on same slot)
36 * Each flow has a CoDel managed queue.
37 * Flows are linked onto two (Round Robin) lists,
38 * so that new flows have priority on old ones.
39 *
40 * For a given flow, packets are not reordered (CoDel uses a FIFO)
41 * head drops only.
42 * ECN capability is on by default.
43 * Low memory footprint (64 bytes per flow)
44 */
45
46struct fq_codel_flow {
47 struct sk_buff *head;
48 struct sk_buff *tail;
49 struct list_head flowchain;
50 int deficit;
51 u32 dropped; /* number of drops (or ECN marks) on this flow */
52 struct codel_vars cvars;
53}; /* please try to keep this structure <= 64 bytes */
54
55struct fq_codel_sched_data {
56 struct tcf_proto __rcu *filter_list; /* optional external classifier */
57 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
58 u32 *backlogs; /* backlog table [flows_cnt] */
59 u32 flows_cnt; /* number of flows */
60 u32 perturbation; /* hash perturbation */
61 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
62 u32 drop_batch_size;
63 u32 memory_limit;
64 struct codel_params cparams;
65 struct codel_stats cstats;
66 u32 memory_usage;
67 u32 drop_overmemory;
68 u32 drop_overlimit;
69 u32 new_flow_count;
70
71 struct list_head new_flows; /* list of new flows */
72 struct list_head old_flows; /* list of old flows */
73};
74
75static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
76 struct sk_buff *skb)
77{
78 u32 hash = skb_get_hash_perturb(skb, q->perturbation);
79
80 return reciprocal_scale(hash, q->flows_cnt);
81}
82
83static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
84 int *qerr)
85{
86 struct fq_codel_sched_data *q = qdisc_priv(sch);
87 struct tcf_proto *filter;
88 struct tcf_result res;
89 int result;
90
91 if (TC_H_MAJ(skb->priority) == sch->handle &&
92 TC_H_MIN(skb->priority) > 0 &&
93 TC_H_MIN(skb->priority) <= q->flows_cnt)
94 return TC_H_MIN(skb->priority);
95
96 filter = rcu_dereference_bh(q->filter_list);
97 if (!filter)
98 return fq_codel_hash(q, skb) + 1;
99
100 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
101 result = tc_classify(skb, filter, &res, false);
102 if (result >= 0) {
103#ifdef CONFIG_NET_CLS_ACT
104 switch (result) {
105 case TC_ACT_STOLEN:
106 case TC_ACT_QUEUED:
107 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
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/* helper functions : might be changed when/if skb use a standard list_head */
119
120/* remove one skb from head of slot queue */
121static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
122{
123 struct sk_buff *skb = flow->head;
124
125 flow->head = skb->next;
126 skb->next = NULL;
127 return skb;
128}
129
130/* add skb to flow queue (tail add) */
131static inline void flow_queue_add(struct fq_codel_flow *flow,
132 struct sk_buff *skb)
133{
134 if (flow->head == NULL)
135 flow->head = skb;
136 else
137 flow->tail->next = skb;
138 flow->tail = skb;
139 skb->next = NULL;
140}
141
142static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
143 struct sk_buff **to_free)
144{
145 struct fq_codel_sched_data *q = qdisc_priv(sch);
146 struct sk_buff *skb;
147 unsigned int maxbacklog = 0, idx = 0, i, len;
148 struct fq_codel_flow *flow;
149 unsigned int threshold;
150 unsigned int mem = 0;
151
152 /* Queue is full! Find the fat flow and drop packet(s) from it.
153 * This might sound expensive, but with 1024 flows, we scan
154 * 4KB of memory, and we dont need to handle a complex tree
155 * in fast path (packet queue/enqueue) with many cache misses.
156 * In stress mode, we'll try to drop 64 packets from the flow,
157 * amortizing this linear lookup to one cache line per drop.
158 */
159 for (i = 0; i < q->flows_cnt; i++) {
160 if (q->backlogs[i] > maxbacklog) {
161 maxbacklog = q->backlogs[i];
162 idx = i;
163 }
164 }
165
166 /* Our goal is to drop half of this fat flow backlog */
167 threshold = maxbacklog >> 1;
168
169 flow = &q->flows[idx];
170 len = 0;
171 i = 0;
172 do {
173 skb = dequeue_head(flow);
174 len += qdisc_pkt_len(skb);
175 mem += get_codel_cb(skb)->mem_usage;
176 __qdisc_drop(skb, to_free);
177 } while (++i < max_packets && len < threshold);
178
179 flow->dropped += i;
180 q->backlogs[idx] -= len;
181 q->memory_usage -= mem;
182 sch->qstats.drops += i;
183 sch->qstats.backlog -= len;
184 sch->q.qlen -= i;
185 return idx;
186}
187
188static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
189 struct sk_buff **to_free)
190{
191 struct fq_codel_sched_data *q = qdisc_priv(sch);
192 unsigned int idx, prev_backlog, prev_qlen;
193 struct fq_codel_flow *flow;
194 int uninitialized_var(ret);
195 unsigned int pkt_len;
196 bool memory_limited;
197
198 idx = fq_codel_classify(skb, sch, &ret);
199 if (idx == 0) {
200 if (ret & __NET_XMIT_BYPASS)
201 qdisc_qstats_drop(sch);
202 __qdisc_drop(skb, to_free);
203 return ret;
204 }
205 idx--;
206
207 codel_set_enqueue_time(skb);
208 flow = &q->flows[idx];
209 flow_queue_add(flow, skb);
210 q->backlogs[idx] += qdisc_pkt_len(skb);
211 qdisc_qstats_backlog_inc(sch, skb);
212
213 if (list_empty(&flow->flowchain)) {
214 list_add_tail(&flow->flowchain, &q->new_flows);
215 q->new_flow_count++;
216 flow->deficit = q->quantum;
217 flow->dropped = 0;
218 }
219 get_codel_cb(skb)->mem_usage = skb->truesize;
220 q->memory_usage += get_codel_cb(skb)->mem_usage;
221 memory_limited = q->memory_usage > q->memory_limit;
222 if (++sch->q.qlen <= sch->limit && !memory_limited)
223 return NET_XMIT_SUCCESS;
224
225 prev_backlog = sch->qstats.backlog;
226 prev_qlen = sch->q.qlen;
227
228 /* save this packet length as it might be dropped by fq_codel_drop() */
229 pkt_len = qdisc_pkt_len(skb);
230 /* fq_codel_drop() is quite expensive, as it performs a linear search
231 * in q->backlogs[] to find a fat flow.
232 * So instead of dropping a single packet, drop half of its backlog
233 * with a 64 packets limit to not add a too big cpu spike here.
234 */
235 ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
236
237 prev_qlen -= sch->q.qlen;
238 prev_backlog -= sch->qstats.backlog;
239 q->drop_overlimit += prev_qlen;
240 if (memory_limited)
241 q->drop_overmemory += prev_qlen;
242
243 /* As we dropped packet(s), better let upper stack know this.
244 * If we dropped a packet for this flow, return NET_XMIT_CN,
245 * but in this case, our parents wont increase their backlogs.
246 */
247 if (ret == idx) {
248 qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
249 prev_backlog - pkt_len);
250 return NET_XMIT_CN;
251 }
252 qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
253 return NET_XMIT_SUCCESS;
254}
255
256/* This is the specific function called from codel_dequeue()
257 * to dequeue a packet from queue. Note: backlog is handled in
258 * codel, we dont need to reduce it here.
259 */
260static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
261{
262 struct Qdisc *sch = ctx;
263 struct fq_codel_sched_data *q = qdisc_priv(sch);
264 struct fq_codel_flow *flow;
265 struct sk_buff *skb = NULL;
266
267 flow = container_of(vars, struct fq_codel_flow, cvars);
268 if (flow->head) {
269 skb = dequeue_head(flow);
270 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
271 q->memory_usage -= get_codel_cb(skb)->mem_usage;
272 sch->q.qlen--;
273 sch->qstats.backlog -= qdisc_pkt_len(skb);
274 }
275 return skb;
276}
277
278static void drop_func(struct sk_buff *skb, void *ctx)
279{
280 struct Qdisc *sch = ctx;
281
282 kfree_skb(skb);
283 qdisc_qstats_drop(sch);
284}
285
286static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
287{
288 struct fq_codel_sched_data *q = qdisc_priv(sch);
289 struct sk_buff *skb;
290 struct fq_codel_flow *flow;
291 struct list_head *head;
292 u32 prev_drop_count, prev_ecn_mark;
293 unsigned int prev_backlog;
294
295begin:
296 head = &q->new_flows;
297 if (list_empty(head)) {
298 head = &q->old_flows;
299 if (list_empty(head))
300 return NULL;
301 }
302 flow = list_first_entry(head, struct fq_codel_flow, flowchain);
303
304 if (flow->deficit <= 0) {
305 flow->deficit += q->quantum;
306 list_move_tail(&flow->flowchain, &q->old_flows);
307 goto begin;
308 }
309
310 prev_drop_count = q->cstats.drop_count;
311 prev_ecn_mark = q->cstats.ecn_mark;
312 prev_backlog = sch->qstats.backlog;
313
314 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
315 &flow->cvars, &q->cstats, qdisc_pkt_len,
316 codel_get_enqueue_time, drop_func, dequeue_func);
317
318 flow->dropped += q->cstats.drop_count - prev_drop_count;
319 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
320
321 if (!skb) {
322 /* force a pass through old_flows to prevent starvation */
323 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
324 list_move_tail(&flow->flowchain, &q->old_flows);
325 else
326 list_del_init(&flow->flowchain);
327 goto begin;
328 }
329 qdisc_bstats_update(sch, skb);
330 flow->deficit -= qdisc_pkt_len(skb);
331 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
332 * or HTB crashes. Defer it for next round.
333 */
334 if (q->cstats.drop_count && sch->q.qlen) {
335 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
336 q->cstats.drop_len);
337 q->cstats.drop_count = 0;
338 q->cstats.drop_len = 0;
339 }
340 return skb;
341}
342
343static void fq_codel_flow_purge(struct fq_codel_flow *flow)
344{
345 rtnl_kfree_skbs(flow->head, flow->tail);
346 flow->head = NULL;
347}
348
349static void fq_codel_reset(struct Qdisc *sch)
350{
351 struct fq_codel_sched_data *q = qdisc_priv(sch);
352 int i;
353
354 INIT_LIST_HEAD(&q->new_flows);
355 INIT_LIST_HEAD(&q->old_flows);
356 for (i = 0; i < q->flows_cnt; i++) {
357 struct fq_codel_flow *flow = q->flows + i;
358
359 fq_codel_flow_purge(flow);
360 INIT_LIST_HEAD(&flow->flowchain);
361 codel_vars_init(&flow->cvars);
362 }
363 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
364 sch->q.qlen = 0;
365 sch->qstats.backlog = 0;
366 q->memory_usage = 0;
367}
368
369static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
370 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
371 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
372 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
373 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
374 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
375 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
376 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
377 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
378 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
379};
380
381static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
382{
383 struct fq_codel_sched_data *q = qdisc_priv(sch);
384 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
385 int err;
386
387 if (!opt)
388 return -EINVAL;
389
390 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
391 if (err < 0)
392 return err;
393 if (tb[TCA_FQ_CODEL_FLOWS]) {
394 if (q->flows)
395 return -EINVAL;
396 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
397 if (!q->flows_cnt ||
398 q->flows_cnt > 65536)
399 return -EINVAL;
400 }
401 sch_tree_lock(sch);
402
403 if (tb[TCA_FQ_CODEL_TARGET]) {
404 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
405
406 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
407 }
408
409 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
410 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
411
412 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
413 }
414
415 if (tb[TCA_FQ_CODEL_INTERVAL]) {
416 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
417
418 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
419 }
420
421 if (tb[TCA_FQ_CODEL_LIMIT])
422 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
423
424 if (tb[TCA_FQ_CODEL_ECN])
425 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
426
427 if (tb[TCA_FQ_CODEL_QUANTUM])
428 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
429
430 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
431 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
432
433 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
434 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
435
436 while (sch->q.qlen > sch->limit ||
437 q->memory_usage > q->memory_limit) {
438 struct sk_buff *skb = fq_codel_dequeue(sch);
439
440 q->cstats.drop_len += qdisc_pkt_len(skb);
441 rtnl_kfree_skbs(skb, skb);
442 q->cstats.drop_count++;
443 }
444 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
445 q->cstats.drop_count = 0;
446 q->cstats.drop_len = 0;
447
448 sch_tree_unlock(sch);
449 return 0;
450}
451
452static void *fq_codel_zalloc(size_t sz)
453{
454 void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
455
456 if (!ptr)
457 ptr = vzalloc(sz);
458 return ptr;
459}
460
461static void fq_codel_free(void *addr)
462{
463 kvfree(addr);
464}
465
466static void fq_codel_destroy(struct Qdisc *sch)
467{
468 struct fq_codel_sched_data *q = qdisc_priv(sch);
469
470 tcf_destroy_chain(&q->filter_list);
471 fq_codel_free(q->backlogs);
472 fq_codel_free(q->flows);
473}
474
475static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
476{
477 struct fq_codel_sched_data *q = qdisc_priv(sch);
478 int i;
479
480 sch->limit = 10*1024;
481 q->flows_cnt = 1024;
482 q->memory_limit = 32 << 20; /* 32 MBytes */
483 q->drop_batch_size = 64;
484 q->quantum = psched_mtu(qdisc_dev(sch));
485 q->perturbation = prandom_u32();
486 INIT_LIST_HEAD(&q->new_flows);
487 INIT_LIST_HEAD(&q->old_flows);
488 codel_params_init(&q->cparams);
489 codel_stats_init(&q->cstats);
490 q->cparams.ecn = true;
491 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
492
493 if (opt) {
494 int err = fq_codel_change(sch, opt);
495 if (err)
496 return err;
497 }
498
499 if (!q->flows) {
500 q->flows = fq_codel_zalloc(q->flows_cnt *
501 sizeof(struct fq_codel_flow));
502 if (!q->flows)
503 return -ENOMEM;
504 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
505 if (!q->backlogs) {
506 fq_codel_free(q->flows);
507 return -ENOMEM;
508 }
509 for (i = 0; i < q->flows_cnt; i++) {
510 struct fq_codel_flow *flow = q->flows + i;
511
512 INIT_LIST_HEAD(&flow->flowchain);
513 codel_vars_init(&flow->cvars);
514 }
515 }
516 if (sch->limit >= 1)
517 sch->flags |= TCQ_F_CAN_BYPASS;
518 else
519 sch->flags &= ~TCQ_F_CAN_BYPASS;
520 return 0;
521}
522
523static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
524{
525 struct fq_codel_sched_data *q = qdisc_priv(sch);
526 struct nlattr *opts;
527
528 opts = nla_nest_start(skb, TCA_OPTIONS);
529 if (opts == NULL)
530 goto nla_put_failure;
531
532 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
533 codel_time_to_us(q->cparams.target)) ||
534 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
535 sch->limit) ||
536 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
537 codel_time_to_us(q->cparams.interval)) ||
538 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
539 q->cparams.ecn) ||
540 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
541 q->quantum) ||
542 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
543 q->drop_batch_size) ||
544 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
545 q->memory_limit) ||
546 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
547 q->flows_cnt))
548 goto nla_put_failure;
549
550 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
551 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
552 codel_time_to_us(q->cparams.ce_threshold)))
553 goto nla_put_failure;
554
555 return nla_nest_end(skb, opts);
556
557nla_put_failure:
558 return -1;
559}
560
561static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
562{
563 struct fq_codel_sched_data *q = qdisc_priv(sch);
564 struct tc_fq_codel_xstats st = {
565 .type = TCA_FQ_CODEL_XSTATS_QDISC,
566 };
567 struct list_head *pos;
568
569 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
570 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
571 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
572 st.qdisc_stats.new_flow_count = q->new_flow_count;
573 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
574 st.qdisc_stats.memory_usage = q->memory_usage;
575 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
576
577 sch_tree_lock(sch);
578 list_for_each(pos, &q->new_flows)
579 st.qdisc_stats.new_flows_len++;
580
581 list_for_each(pos, &q->old_flows)
582 st.qdisc_stats.old_flows_len++;
583 sch_tree_unlock(sch);
584
585 return gnet_stats_copy_app(d, &st, sizeof(st));
586}
587
588static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
589{
590 return NULL;
591}
592
593static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
594{
595 return 0;
596}
597
598static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
599 u32 classid)
600{
601 /* we cannot bypass queue discipline anymore */
602 sch->flags &= ~TCQ_F_CAN_BYPASS;
603 return 0;
604}
605
606static void fq_codel_put(struct Qdisc *q, unsigned long cl)
607{
608}
609
610static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
611 unsigned long cl)
612{
613 struct fq_codel_sched_data *q = qdisc_priv(sch);
614
615 if (cl)
616 return NULL;
617 return &q->filter_list;
618}
619
620static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
621 struct sk_buff *skb, struct tcmsg *tcm)
622{
623 tcm->tcm_handle |= TC_H_MIN(cl);
624 return 0;
625}
626
627static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
628 struct gnet_dump *d)
629{
630 struct fq_codel_sched_data *q = qdisc_priv(sch);
631 u32 idx = cl - 1;
632 struct gnet_stats_queue qs = { 0 };
633 struct tc_fq_codel_xstats xstats;
634
635 if (idx < q->flows_cnt) {
636 const struct fq_codel_flow *flow = &q->flows[idx];
637 const struct sk_buff *skb;
638
639 memset(&xstats, 0, sizeof(xstats));
640 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
641 xstats.class_stats.deficit = flow->deficit;
642 xstats.class_stats.ldelay =
643 codel_time_to_us(flow->cvars.ldelay);
644 xstats.class_stats.count = flow->cvars.count;
645 xstats.class_stats.lastcount = flow->cvars.lastcount;
646 xstats.class_stats.dropping = flow->cvars.dropping;
647 if (flow->cvars.dropping) {
648 codel_tdiff_t delta = flow->cvars.drop_next -
649 codel_get_time();
650
651 xstats.class_stats.drop_next = (delta >= 0) ?
652 codel_time_to_us(delta) :
653 -codel_time_to_us(-delta);
654 }
655 if (flow->head) {
656 sch_tree_lock(sch);
657 skb = flow->head;
658 while (skb) {
659 qs.qlen++;
660 skb = skb->next;
661 }
662 sch_tree_unlock(sch);
663 }
664 qs.backlog = q->backlogs[idx];
665 qs.drops = flow->dropped;
666 }
667 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
668 return -1;
669 if (idx < q->flows_cnt)
670 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
671 return 0;
672}
673
674static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
675{
676 struct fq_codel_sched_data *q = qdisc_priv(sch);
677 unsigned int i;
678
679 if (arg->stop)
680 return;
681
682 for (i = 0; i < q->flows_cnt; i++) {
683 if (list_empty(&q->flows[i].flowchain) ||
684 arg->count < arg->skip) {
685 arg->count++;
686 continue;
687 }
688 if (arg->fn(sch, i + 1, arg) < 0) {
689 arg->stop = 1;
690 break;
691 }
692 arg->count++;
693 }
694}
695
696static const struct Qdisc_class_ops fq_codel_class_ops = {
697 .leaf = fq_codel_leaf,
698 .get = fq_codel_get,
699 .put = fq_codel_put,
700 .tcf_chain = fq_codel_find_tcf,
701 .bind_tcf = fq_codel_bind,
702 .unbind_tcf = fq_codel_put,
703 .dump = fq_codel_dump_class,
704 .dump_stats = fq_codel_dump_class_stats,
705 .walk = fq_codel_walk,
706};
707
708static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
709 .cl_ops = &fq_codel_class_ops,
710 .id = "fq_codel",
711 .priv_size = sizeof(struct fq_codel_sched_data),
712 .enqueue = fq_codel_enqueue,
713 .dequeue = fq_codel_dequeue,
714 .peek = qdisc_peek_dequeued,
715 .init = fq_codel_init,
716 .reset = fq_codel_reset,
717 .destroy = fq_codel_destroy,
718 .change = fq_codel_change,
719 .dump = fq_codel_dump,
720 .dump_stats = fq_codel_dump_stats,
721 .owner = THIS_MODULE,
722};
723
724static int __init fq_codel_module_init(void)
725{
726 return register_qdisc(&fq_codel_qdisc_ops);
727}
728
729static void __exit fq_codel_module_exit(void)
730{
731 unregister_qdisc(&fq_codel_qdisc_ops);
732}
733
734module_init(fq_codel_module_init)
735module_exit(fq_codel_module_exit)
736MODULE_AUTHOR("Eric Dumazet");
737MODULE_LICENSE("GPL");