xref: /linux/net/sched/sch_fq_codel.c (revision 67f8bc848ee31831336bd478e57d2f993551902e)
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 
42 struct 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 
50 struct 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 
70 static 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 
76 static 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_qdisc(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 */
116 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
117 {
118 	struct sk_buff *skb = flow->head;
119 
120 	WRITE_ONCE(flow->head, skb->next);
121 	skb_mark_not_on_list(skb);
122 	return skb;
123 }
124 
125 /* add skb to flow queue (tail add) */
126 static inline void flow_queue_add(struct fq_codel_flow *flow,
127 				  struct sk_buff *skb)
128 {
129 	if (flow->head == NULL)
130 		WRITE_ONCE(flow->head, skb);
131 	else
132 		flow->tail->next = skb;
133 	flow->tail = skb;
134 	skb->next = NULL;
135 }
136 
137 static 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 		tcf_set_qdisc_drop_reason(skb, QDISC_DROP_OVERLIMIT);
172 		__qdisc_drop(skb, to_free);
173 	} while (++i < max_packets && len < threshold);
174 
175 	/* Tell codel to increase its signal strength also */
176 	WRITE_ONCE(flow->cvars.count, flow->cvars.count + i);
177 	WRITE_ONCE(q->backlogs[idx], q->backlogs[idx] - len);
178 	q->memory_usage -= mem;
179 	__qdisc_qstats_drop(sch, i);
180 	qstats_backlog_sub(sch, len);
181 	WRITE_ONCE(sch->q.qlen, sch->q.qlen - i);
182 	return idx;
183 }
184 
185 static 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 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 	WRITE_ONCE(q->backlogs[idx], 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 		WRITE_ONCE(flow->deficit, q->quantum);
214 	}
215 	get_codel_cb(skb)->mem_usage = is_skb_wmem(skb) ? 0 : skb->truesize;
216 	q->memory_usage += get_codel_cb(skb)->mem_usage;
217 	memory_limited = q->memory_usage > q->memory_limit;
218 	qdisc_qlen_inc(sch);
219 	if (sch->q.qlen <= sch->limit && !memory_limited)
220 		return NET_XMIT_SUCCESS;
221 
222 	prev_backlog = sch->qstats.backlog;
223 	prev_qlen = sch->q.qlen;
224 
225 	/* save this packet length as it might be dropped by fq_codel_drop() */
226 	pkt_len = qdisc_pkt_len(skb);
227 	/* fq_codel_drop() is quite expensive, as it performs a linear search
228 	 * in q->backlogs[] to find a fat flow.
229 	 * So instead of dropping a single packet, drop half of its backlog
230 	 * with a 64 packets limit to not add a too big cpu spike here.
231 	 */
232 	ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
233 
234 	prev_qlen -= sch->q.qlen;
235 	prev_backlog -= sch->qstats.backlog;
236 	q->drop_overlimit += prev_qlen;
237 	if (memory_limited)
238 		q->drop_overmemory += prev_qlen;
239 
240 	/* As we dropped packet(s), better let upper stack know this.
241 	 * If we dropped a packet for this flow, return NET_XMIT_CN,
242 	 * but in this case, our parents wont increase their backlogs.
243 	 */
244 	if (ret == idx) {
245 		qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
246 					  prev_backlog - pkt_len);
247 		return NET_XMIT_CN;
248 	}
249 	qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
250 	return NET_XMIT_SUCCESS;
251 }
252 
253 /* This is the specific function called from codel_dequeue()
254  * to dequeue a packet from queue. Note: backlog is handled in
255  * codel, we dont need to reduce it here.
256  */
257 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
258 {
259 	struct Qdisc *sch = ctx;
260 	struct fq_codel_sched_data *q = qdisc_priv(sch);
261 	struct fq_codel_flow *flow;
262 	struct sk_buff *skb = NULL;
263 
264 	flow = container_of(vars, struct fq_codel_flow, cvars);
265 	if (flow->head) {
266 		skb = dequeue_head(flow);
267 		WRITE_ONCE(q->backlogs[flow - q->flows],
268 			   q->backlogs[flow - q->flows] - qdisc_pkt_len(skb));
269 		q->memory_usage -= get_codel_cb(skb)->mem_usage;
270 		qdisc_qlen_dec(sch);
271 		qdisc_qstats_backlog_dec(sch, skb);
272 	}
273 	return skb;
274 }
275 
276 static void drop_func(struct sk_buff *skb, void *ctx)
277 {
278 	struct Qdisc *sch = ctx;
279 
280 	qdisc_dequeue_drop(sch, skb, QDISC_DROP_CONGESTED);
281 	qdisc_qstats_drop(sch);
282 }
283 
284 static struct sk_buff *__fq_codel_dequeue(struct Qdisc *sch)
285 {
286 	struct fq_codel_sched_data *q = qdisc_priv(sch);
287 	struct sk_buff *skb;
288 	struct fq_codel_flow *flow;
289 	struct list_head *head;
290 
291 begin:
292 	head = &q->new_flows;
293 	if (list_empty(head)) {
294 		head = &q->old_flows;
295 		if (list_empty(head))
296 			return NULL;
297 	}
298 	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
299 
300 	if (flow->deficit <= 0) {
301 		WRITE_ONCE(flow->deficit, flow->deficit + q->quantum);
302 		list_move_tail(&flow->flowchain, &q->old_flows);
303 		goto begin;
304 	}
305 
306 	skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
307 			    &flow->cvars, &q->cstats, qdisc_pkt_len,
308 			    codel_get_enqueue_time, drop_func, dequeue_func);
309 
310 	if (!skb) {
311 		/* force a pass through old_flows to prevent starvation */
312 		if ((head == &q->new_flows) && !list_empty(&q->old_flows))
313 			list_move_tail(&flow->flowchain, &q->old_flows);
314 		else
315 			list_del_init(&flow->flowchain);
316 		goto begin;
317 	}
318 	qdisc_bstats_update(sch, skb);
319 	WRITE_ONCE(flow->deficit, flow->deficit - qdisc_pkt_len(skb));
320 
321 	return skb;
322 }
323 
324 static void fq_codel_dequeue_drop(struct Qdisc *sch)
325 {
326 	struct fq_codel_sched_data *q = qdisc_priv(sch);
327 
328 	if (q->cstats.drop_count) {
329 		qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
330 					  q->cstats.drop_len);
331 		q->cstats.drop_count = 0;
332 		q->cstats.drop_len = 0;
333 	}
334 }
335 
336 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
337 {
338 	struct sk_buff *skb;
339 
340 	skb =  __fq_codel_dequeue(sch);
341 
342 	fq_codel_dequeue_drop(sch);
343 
344 	return skb;
345 }
346 
347 static struct sk_buff *fq_codel_peek(struct Qdisc *sch)
348 {
349 	struct sk_buff *skb = skb_peek(&sch->gso_skb);
350 
351 	if (!skb) {
352 		skb = __fq_codel_dequeue(sch);
353 
354 		if (skb) {
355 			__skb_queue_head(&sch->gso_skb, skb);
356 			/* it's still part of the queue */
357 			qdisc_qstats_backlog_inc(sch, skb);
358 			sch->q.qlen++;
359 		}
360 
361 		fq_codel_dequeue_drop(sch);
362 	}
363 
364 	return skb;
365 }
366 
367 static void fq_codel_flow_purge(struct fq_codel_flow *flow)
368 {
369 	rtnl_kfree_skbs(flow->head, flow->tail);
370 	WRITE_ONCE(flow->head, NULL);
371 }
372 
373 static void fq_codel_reset(struct Qdisc *sch)
374 {
375 	struct fq_codel_sched_data *q = qdisc_priv(sch);
376 	int i;
377 
378 	INIT_LIST_HEAD(&q->new_flows);
379 	INIT_LIST_HEAD(&q->old_flows);
380 	for (i = 0; i < q->flows_cnt; i++) {
381 		struct fq_codel_flow *flow = q->flows + i;
382 
383 		fq_codel_flow_purge(flow);
384 		INIT_LIST_HEAD(&flow->flowchain);
385 		codel_vars_init(&flow->cvars);
386 	}
387 	memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
388 	q->memory_usage = 0;
389 }
390 
391 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
392 	[TCA_FQ_CODEL_TARGET]	= { .type = NLA_U32 },
393 	[TCA_FQ_CODEL_LIMIT]	= { .type = NLA_U32 },
394 	[TCA_FQ_CODEL_INTERVAL]	= { .type = NLA_U32 },
395 	[TCA_FQ_CODEL_ECN]	= { .type = NLA_U32 },
396 	[TCA_FQ_CODEL_FLOWS]	= { .type = NLA_U32 },
397 	[TCA_FQ_CODEL_QUANTUM]	= { .type = NLA_U32 },
398 	[TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
399 	[TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
400 	[TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
401 	[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR] = { .type = NLA_U8 },
402 	[TCA_FQ_CODEL_CE_THRESHOLD_MASK] = { .type = NLA_U8 },
403 };
404 
405 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
406 			   struct netlink_ext_ack *extack)
407 {
408 	unsigned int dropped_pkts = 0, dropped_bytes = 0;
409 	struct fq_codel_sched_data *q = qdisc_priv(sch);
410 	struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
411 	u32 quantum = 0;
412 	int err;
413 
414 	err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
415 					  fq_codel_policy, NULL);
416 	if (err < 0)
417 		return err;
418 	if (tb[TCA_FQ_CODEL_FLOWS]) {
419 		if (q->flows)
420 			return -EINVAL;
421 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
422 		if (!q->flows_cnt ||
423 		    q->flows_cnt > 65536)
424 			return -EINVAL;
425 	}
426 	if (tb[TCA_FQ_CODEL_QUANTUM]) {
427 		quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
428 		if (quantum > FQ_CODEL_QUANTUM_MAX) {
429 			NL_SET_ERR_MSG(extack, "Invalid quantum");
430 			return -EINVAL;
431 		}
432 	}
433 	sch_tree_lock(sch);
434 
435 	if (tb[TCA_FQ_CODEL_TARGET]) {
436 		u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
437 
438 		WRITE_ONCE(q->cparams.target,
439 			   (target * NSEC_PER_USEC) >> CODEL_SHIFT);
440 	}
441 
442 	if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
443 		u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
444 
445 		WRITE_ONCE(q->cparams.ce_threshold,
446 			   (val * NSEC_PER_USEC) >> CODEL_SHIFT);
447 	}
448 
449 	if (tb[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR])
450 		WRITE_ONCE(q->cparams.ce_threshold_selector,
451 			   nla_get_u8(tb[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR]));
452 	if (tb[TCA_FQ_CODEL_CE_THRESHOLD_MASK])
453 		WRITE_ONCE(q->cparams.ce_threshold_mask,
454 			   nla_get_u8(tb[TCA_FQ_CODEL_CE_THRESHOLD_MASK]));
455 
456 	if (tb[TCA_FQ_CODEL_INTERVAL]) {
457 		u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
458 
459 		WRITE_ONCE(q->cparams.interval,
460 			   (interval * NSEC_PER_USEC) >> CODEL_SHIFT);
461 	}
462 
463 	if (tb[TCA_FQ_CODEL_LIMIT])
464 		WRITE_ONCE(sch->limit,
465 			   nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]));
466 
467 	if (tb[TCA_FQ_CODEL_ECN])
468 		WRITE_ONCE(q->cparams.ecn,
469 			   !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]));
470 
471 	if (quantum)
472 		WRITE_ONCE(q->quantum, quantum);
473 
474 	if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
475 		WRITE_ONCE(q->drop_batch_size,
476 			   max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])));
477 
478 	if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
479 		WRITE_ONCE(q->memory_limit,
480 			   min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT])));
481 
482 	while (sch->q.qlen > sch->limit ||
483 	       q->memory_usage > q->memory_limit) {
484 		struct sk_buff *skb = qdisc_dequeue_internal(sch, false);
485 
486 		if (!skb)
487 			break;
488 
489 		dropped_pkts++;
490 		dropped_bytes += qdisc_pkt_len(skb);
491 		rtnl_kfree_skbs(skb, skb);
492 	}
493 	qdisc_tree_reduce_backlog(sch, dropped_pkts, dropped_bytes);
494 
495 	sch_tree_unlock(sch);
496 	return 0;
497 }
498 
499 static void fq_codel_destroy(struct Qdisc *sch)
500 {
501 	struct fq_codel_sched_data *q = qdisc_priv(sch);
502 
503 	tcf_block_put(q->block);
504 	kvfree(q->backlogs);
505 	kvfree(q->flows);
506 }
507 
508 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
509 			 struct netlink_ext_ack *extack)
510 {
511 	struct fq_codel_sched_data *q = qdisc_priv(sch);
512 	u32 mtu;
513 	int i;
514 	int err;
515 
516 	sch->limit = 10*1024;
517 	q->flows_cnt = 1024;
518 	q->memory_limit = 32 << 20; /* 32 MBytes */
519 	q->drop_batch_size = 64;
520 	mtu = clamp_t(u32, psched_mtu(qdisc_dev(sch)), 256, FQ_CODEL_QUANTUM_MAX);
521 	q->quantum = mtu;
522 	INIT_LIST_HEAD(&q->new_flows);
523 	INIT_LIST_HEAD(&q->old_flows);
524 	codel_params_init(&q->cparams);
525 	codel_stats_init(&q->cstats);
526 	q->cparams.ecn = true;
527 	q->cparams.mtu = mtu;
528 
529 	if (opt) {
530 		err = fq_codel_change(sch, opt, extack);
531 		if (err)
532 			goto init_failure;
533 	}
534 
535 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
536 	if (err)
537 		goto init_failure;
538 
539 	if (!q->flows) {
540 		q->flows = kvzalloc_objs(struct fq_codel_flow, q->flows_cnt);
541 		if (!q->flows) {
542 			err = -ENOMEM;
543 			goto init_failure;
544 		}
545 		q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
546 		if (!q->backlogs) {
547 			err = -ENOMEM;
548 			goto alloc_failure;
549 		}
550 		for (i = 0; i < q->flows_cnt; i++) {
551 			struct fq_codel_flow *flow = q->flows + i;
552 
553 			INIT_LIST_HEAD(&flow->flowchain);
554 			codel_vars_init(&flow->cvars);
555 		}
556 	}
557 	if (sch->limit >= 1)
558 		sch->flags |= TCQ_F_CAN_BYPASS;
559 	else
560 		sch->flags &= ~TCQ_F_CAN_BYPASS;
561 
562 	sch->flags |= TCQ_F_DEQUEUE_DROPS;
563 
564 	return 0;
565 
566 alloc_failure:
567 	kvfree(q->flows);
568 	q->flows = NULL;
569 init_failure:
570 	q->flows_cnt = 0;
571 	return err;
572 }
573 
574 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
575 {
576 	struct fq_codel_sched_data *q = qdisc_priv(sch);
577 	codel_time_t ce_threshold;
578 	struct nlattr *opts;
579 
580 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
581 	if (opts == NULL)
582 		goto nla_put_failure;
583 
584 	if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
585 			codel_time_to_us(READ_ONCE(q->cparams.target))) ||
586 	    nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
587 			READ_ONCE(sch->limit)) ||
588 	    nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
589 			codel_time_to_us(READ_ONCE(q->cparams.interval))) ||
590 	    nla_put_u32(skb, TCA_FQ_CODEL_ECN,
591 			READ_ONCE(q->cparams.ecn)) ||
592 	    nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
593 			READ_ONCE(q->quantum)) ||
594 	    nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
595 			READ_ONCE(q->drop_batch_size)) ||
596 	    nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
597 			READ_ONCE(q->memory_limit)) ||
598 	    nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
599 			READ_ONCE(q->flows_cnt)))
600 		goto nla_put_failure;
601 
602 	ce_threshold = READ_ONCE(q->cparams.ce_threshold);
603 	if (ce_threshold != CODEL_DISABLED_THRESHOLD) {
604 		if (nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
605 				codel_time_to_us(ce_threshold)))
606 			goto nla_put_failure;
607 		if (nla_put_u8(skb, TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR,
608 			       READ_ONCE(q->cparams.ce_threshold_selector)))
609 			goto nla_put_failure;
610 		if (nla_put_u8(skb, TCA_FQ_CODEL_CE_THRESHOLD_MASK,
611 			       READ_ONCE(q->cparams.ce_threshold_mask)))
612 			goto nla_put_failure;
613 	}
614 
615 	return nla_nest_end(skb, opts);
616 
617 nla_put_failure:
618 	return -1;
619 }
620 
621 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
622 {
623 	struct fq_codel_sched_data *q = qdisc_priv(sch);
624 	struct tc_fq_codel_xstats st = {
625 		.type				= TCA_FQ_CODEL_XSTATS_QDISC,
626 	};
627 	struct list_head *pos;
628 
629 	sch_tree_lock(sch);
630 
631 	st.qdisc_stats.maxpacket = q->cstats.maxpacket;
632 	st.qdisc_stats.drop_overlimit = q->drop_overlimit;
633 	st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
634 	st.qdisc_stats.new_flow_count = q->new_flow_count;
635 	st.qdisc_stats.ce_mark = q->cstats.ce_mark;
636 	st.qdisc_stats.memory_usage  = q->memory_usage;
637 	st.qdisc_stats.drop_overmemory = q->drop_overmemory;
638 
639 	list_for_each(pos, &q->new_flows)
640 		st.qdisc_stats.new_flows_len++;
641 
642 	list_for_each(pos, &q->old_flows)
643 		st.qdisc_stats.old_flows_len++;
644 	sch_tree_unlock(sch);
645 
646 	return gnet_stats_copy_app(d, &st, sizeof(st));
647 }
648 
649 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
650 {
651 	return NULL;
652 }
653 
654 static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
655 {
656 	return 0;
657 }
658 
659 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
660 			      u32 classid)
661 {
662 	return 0;
663 }
664 
665 static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
666 {
667 }
668 
669 static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
670 					    struct netlink_ext_ack *extack)
671 {
672 	struct fq_codel_sched_data *q = qdisc_priv(sch);
673 
674 	if (cl)
675 		return NULL;
676 	return q->block;
677 }
678 
679 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
680 			  struct sk_buff *skb, struct tcmsg *tcm)
681 {
682 	tcm->tcm_handle |= TC_H_MIN(cl);
683 	return 0;
684 }
685 
686 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
687 				     struct gnet_dump *d)
688 {
689 	struct fq_codel_sched_data *q = qdisc_priv(sch);
690 	u32 idx = cl - 1;
691 	struct gnet_stats_queue qs = { 0 };
692 	struct tc_fq_codel_xstats xstats;
693 
694 	if (idx < q->flows_cnt) {
695 		const struct fq_codel_flow *flow = &q->flows[idx];
696 		const struct sk_buff *skb;
697 
698 		memset(&xstats, 0, sizeof(xstats));
699 		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
700 		xstats.class_stats.deficit = READ_ONCE(flow->deficit);
701 		xstats.class_stats.ldelay =
702 			codel_time_to_us(READ_ONCE(flow->cvars.ldelay));
703 		xstats.class_stats.count = READ_ONCE(flow->cvars.count);
704 		xstats.class_stats.lastcount = READ_ONCE(flow->cvars.lastcount);
705 		xstats.class_stats.dropping = READ_ONCE(flow->cvars.dropping);
706 		if (xstats.class_stats.dropping) {
707 			codel_tdiff_t delta = READ_ONCE(flow->cvars.drop_next) -
708 					      codel_get_time();
709 
710 			xstats.class_stats.drop_next = (delta >= 0) ?
711 				codel_time_to_us(delta) :
712 				-codel_time_to_us(-delta);
713 		}
714 		if (READ_ONCE(flow->head)) {
715 			sch_tree_lock(sch);
716 			skb = flow->head;
717 			while (skb) {
718 				qs.qlen++;
719 				skb = skb->next;
720 			}
721 			sch_tree_unlock(sch);
722 		}
723 		qs.backlog = READ_ONCE(q->backlogs[idx]);
724 		qs.drops = 0;
725 	}
726 	if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
727 		return -1;
728 	if (idx < q->flows_cnt)
729 		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
730 	return 0;
731 }
732 
733 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
734 {
735 	struct fq_codel_sched_data *q = qdisc_priv(sch);
736 	unsigned int i;
737 
738 	if (arg->stop)
739 		return;
740 
741 	for (i = 0; i < q->flows_cnt; i++) {
742 		if (list_empty(&q->flows[i].flowchain)) {
743 			arg->count++;
744 			continue;
745 		}
746 		if (!tc_qdisc_stats_dump(sch, i + 1, arg))
747 			break;
748 	}
749 }
750 
751 static const struct Qdisc_class_ops fq_codel_class_ops = {
752 	.leaf		=	fq_codel_leaf,
753 	.find		=	fq_codel_find,
754 	.tcf_block	=	fq_codel_tcf_block,
755 	.bind_tcf	=	fq_codel_bind,
756 	.unbind_tcf	=	fq_codel_unbind,
757 	.dump		=	fq_codel_dump_class,
758 	.dump_stats	=	fq_codel_dump_class_stats,
759 	.walk		=	fq_codel_walk,
760 };
761 
762 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
763 	.cl_ops		=	&fq_codel_class_ops,
764 	.id		=	"fq_codel",
765 	.priv_size	=	sizeof(struct fq_codel_sched_data),
766 	.enqueue	=	fq_codel_enqueue,
767 	.dequeue	=	fq_codel_dequeue,
768 	.peek		=	fq_codel_peek,
769 	.init		=	fq_codel_init,
770 	.reset		=	fq_codel_reset,
771 	.destroy	=	fq_codel_destroy,
772 	.change		=	fq_codel_change,
773 	.dump		=	fq_codel_dump,
774 	.dump_stats =	fq_codel_dump_stats,
775 	.owner		=	THIS_MODULE,
776 };
777 MODULE_ALIAS_NET_SCH("fq_codel");
778 
779 static int __init fq_codel_module_init(void)
780 {
781 	return register_qdisc(&fq_codel_qdisc_ops);
782 }
783 
784 static void __exit fq_codel_module_exit(void)
785 {
786 	unregister_qdisc(&fq_codel_qdisc_ops);
787 }
788 
789 module_init(fq_codel_module_init)
790 module_exit(fq_codel_module_exit)
791 MODULE_AUTHOR("Eric Dumazet");
792 MODULE_LICENSE("GPL");
793 MODULE_DESCRIPTION("Fair Queue CoDel discipline");
794