blob: 22fa13cf5d8b89916cb004438ed5132adbad521a [file] [log] [blame]
Eric Dumazet4b549a22012-05-11 09:30:50 +00001/*
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 *
Eric Dumazet80ba92f2015-05-08 15:05:12 -07009 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
Eric Dumazet4b549a22012-05-11 09:30:50 +000010 */
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>
Jiri Pirkocf1facd2017-02-09 14:38:56 +010026#include <net/pkt_cls.h>
Eric Dumazet4b549a22012-05-11 09:30:50 +000027#include <net/codel.h>
Michal Kaziord068ca22016-04-22 14:15:59 +020028#include <net/codel_impl.h>
29#include <net/codel_qdisc.h>
Eric Dumazet4b549a22012-05-11 09:30:50 +000030
31/* Fair Queue CoDel.
32 *
33 * Principles :
34 * Packets are classified (internal classifier or external) on flows.
35 * This is a Stochastic model (as we use a hash, several flows
36 * might be hashed on same slot)
37 * Each flow has a CoDel managed queue.
38 * Flows are linked onto two (Round Robin) lists,
39 * so that new flows have priority on old ones.
40 *
41 * For a given flow, packets are not reordered (CoDel uses a FIFO)
42 * head drops only.
43 * ECN capability is on by default.
44 * Low memory footprint (64 bytes per flow)
45 */
46
47struct fq_codel_flow {
48 struct sk_buff *head;
49 struct sk_buff *tail;
50 struct list_head flowchain;
51 int deficit;
52 u32 dropped; /* number of drops (or ECN marks) on this flow */
53 struct codel_vars cvars;
54}; /* please try to keep this structure <= 64 bytes */
55
56struct fq_codel_sched_data {
John Fastabend25d8c0d2014-09-12 20:05:27 -070057 struct tcf_proto __rcu *filter_list; /* optional external classifier */
Jiri Pirko6529eab2017-05-17 11:07:55 +020058 struct tcf_block *block;
Eric Dumazet4b549a22012-05-11 09:30:50 +000059 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
60 u32 *backlogs; /* backlog table [flows_cnt] */
61 u32 flows_cnt; /* number of flows */
Eric Dumazet4b549a22012-05-11 09:30:50 +000062 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
Eric Dumazet9d185622016-05-01 16:47:26 -070063 u32 drop_batch_size;
Eric Dumazet95b58432016-05-06 08:55:12 -070064 u32 memory_limit;
Eric Dumazet4b549a22012-05-11 09:30:50 +000065 struct codel_params cparams;
66 struct codel_stats cstats;
Eric Dumazet95b58432016-05-06 08:55:12 -070067 u32 memory_usage;
68 u32 drop_overmemory;
Eric Dumazet4b549a22012-05-11 09:30:50 +000069 u32 drop_overlimit;
70 u32 new_flow_count;
71
72 struct list_head new_flows; /* list of new flows */
73 struct list_head old_flows; /* list of old flows */
74};
75
76static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
Tom Herbert342db222015-05-01 11:30:13 -070077 struct sk_buff *skb)
Eric Dumazet4b549a22012-05-11 09:30:50 +000078{
Andrew Collins264b87f2017-01-18 14:04:28 -070079 return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
Eric Dumazet4b549a22012-05-11 09:30:50 +000080}
81
82static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
83 int *qerr)
84{
85 struct fq_codel_sched_data *q = qdisc_priv(sch);
John Fastabend25d8c0d2014-09-12 20:05:27 -070086 struct tcf_proto *filter;
Eric Dumazet4b549a22012-05-11 09:30:50 +000087 struct tcf_result res;
88 int result;
89
90 if (TC_H_MAJ(skb->priority) == sch->handle &&
91 TC_H_MIN(skb->priority) > 0 &&
92 TC_H_MIN(skb->priority) <= q->flows_cnt)
93 return TC_H_MIN(skb->priority);
94
Valdis.Kletnieks@vt.edu69204cf2014-12-09 16:15:50 -050095 filter = rcu_dereference_bh(q->filter_list);
John Fastabend25d8c0d2014-09-12 20:05:27 -070096 if (!filter)
Eric Dumazet4b549a22012-05-11 09:30:50 +000097 return fq_codel_hash(q, skb) + 1;
98
99 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
Jiri Pirko87d83092017-05-17 11:07:54 +0200100 result = tcf_classify(skb, filter, &res, false);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000101 if (result >= 0) {
102#ifdef CONFIG_NET_CLS_ACT
103 switch (result) {
104 case TC_ACT_STOLEN:
105 case TC_ACT_QUEUED:
Jiri Pirkoe25ea212017-06-06 14:12:02 +0200106 case TC_ACT_TRAP:
Eric Dumazet4b549a22012-05-11 09:30:50 +0000107 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
Gustavo A. R. Silvaf3ae6082017-10-19 16:28:24 -0500108 /* fall through */
Eric Dumazet4b549a22012-05-11 09:30:50 +0000109 case TC_ACT_SHOT:
110 return 0;
111 }
112#endif
113 if (TC_H_MIN(res.classid) <= q->flows_cnt)
114 return TC_H_MIN(res.classid);
115 }
116 return 0;
117}
118
119/* helper functions : might be changed when/if skb use a standard list_head */
120
121/* remove one skb from head of slot queue */
122static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
123{
124 struct sk_buff *skb = flow->head;
125
126 flow->head = skb->next;
127 skb->next = NULL;
128 return skb;
129}
130
131/* add skb to flow queue (tail add) */
132static inline void flow_queue_add(struct fq_codel_flow *flow,
133 struct sk_buff *skb)
134{
135 if (flow->head == NULL)
136 flow->head = skb;
137 else
138 flow->tail->next = skb;
139 flow->tail = skb;
140 skb->next = NULL;
141}
142
Eric Dumazet520ac302016-06-21 23:16:49 -0700143static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
144 struct sk_buff **to_free)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000145{
146 struct fq_codel_sched_data *q = qdisc_priv(sch);
147 struct sk_buff *skb;
148 unsigned int maxbacklog = 0, idx = 0, i, len;
149 struct fq_codel_flow *flow;
Eric Dumazet9d185622016-05-01 16:47:26 -0700150 unsigned int threshold;
Eric Dumazet95b58432016-05-06 08:55:12 -0700151 unsigned int mem = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000152
Eric Dumazet9d185622016-05-01 16:47:26 -0700153 /* Queue is full! Find the fat flow and drop packet(s) from it.
Eric Dumazet4b549a22012-05-11 09:30:50 +0000154 * This might sound expensive, but with 1024 flows, we scan
155 * 4KB of memory, and we dont need to handle a complex tree
156 * in fast path (packet queue/enqueue) with many cache misses.
Eric Dumazet9d185622016-05-01 16:47:26 -0700157 * In stress mode, we'll try to drop 64 packets from the flow,
158 * amortizing this linear lookup to one cache line per drop.
Eric Dumazet4b549a22012-05-11 09:30:50 +0000159 */
160 for (i = 0; i < q->flows_cnt; i++) {
161 if (q->backlogs[i] > maxbacklog) {
162 maxbacklog = q->backlogs[i];
163 idx = i;
164 }
165 }
Eric Dumazet9d185622016-05-01 16:47:26 -0700166
167 /* Our goal is to drop half of this fat flow backlog */
168 threshold = maxbacklog >> 1;
169
Eric Dumazet4b549a22012-05-11 09:30:50 +0000170 flow = &q->flows[idx];
Eric Dumazet9d185622016-05-01 16:47:26 -0700171 len = 0;
172 i = 0;
173 do {
174 skb = dequeue_head(flow);
175 len += qdisc_pkt_len(skb);
Eric Dumazet008830b2016-06-21 23:16:50 -0700176 mem += get_codel_cb(skb)->mem_usage;
Eric Dumazet520ac302016-06-21 23:16:49 -0700177 __qdisc_drop(skb, to_free);
Eric Dumazet9d185622016-05-01 16:47:26 -0700178 } while (++i < max_packets && len < threshold);
179
180 flow->dropped += i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000181 q->backlogs[idx] -= len;
Eric Dumazet95b58432016-05-06 08:55:12 -0700182 q->memory_usage -= mem;
Eric Dumazet9d185622016-05-01 16:47:26 -0700183 sch->qstats.drops += i;
184 sch->qstats.backlog -= len;
185 sch->q.qlen -= i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000186 return idx;
187}
188
Eric Dumazet520ac302016-06-21 23:16:49 -0700189static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
190 struct sk_buff **to_free)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000191{
192 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet9d185622016-05-01 16:47:26 -0700193 unsigned int idx, prev_backlog, prev_qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000194 struct fq_codel_flow *flow;
195 int uninitialized_var(ret);
Eric Dumazet80e509d2016-06-04 12:55:13 -0700196 unsigned int pkt_len;
Eric Dumazet95b58432016-05-06 08:55:12 -0700197 bool memory_limited;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000198
199 idx = fq_codel_classify(skb, sch, &ret);
200 if (idx == 0) {
201 if (ret & __NET_XMIT_BYPASS)
John Fastabend25331d62014-09-28 11:53:29 -0700202 qdisc_qstats_drop(sch);
Eric Dumazet520ac302016-06-21 23:16:49 -0700203 __qdisc_drop(skb, to_free);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000204 return ret;
205 }
206 idx--;
207
208 codel_set_enqueue_time(skb);
209 flow = &q->flows[idx];
210 flow_queue_add(flow, skb);
211 q->backlogs[idx] += qdisc_pkt_len(skb);
John Fastabend25331d62014-09-28 11:53:29 -0700212 qdisc_qstats_backlog_inc(sch, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000213
214 if (list_empty(&flow->flowchain)) {
215 list_add_tail(&flow->flowchain, &q->new_flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000216 q->new_flow_count++;
217 flow->deficit = q->quantum;
218 flow->dropped = 0;
219 }
Eric Dumazet008830b2016-06-21 23:16:50 -0700220 get_codel_cb(skb)->mem_usage = skb->truesize;
221 q->memory_usage += get_codel_cb(skb)->mem_usage;
Eric Dumazet95b58432016-05-06 08:55:12 -0700222 memory_limited = q->memory_usage > q->memory_limit;
223 if (++sch->q.qlen <= sch->limit && !memory_limited)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000224 return NET_XMIT_SUCCESS;
225
WANG Cong2ccccf52016-02-25 14:55:01 -0800226 prev_backlog = sch->qstats.backlog;
Eric Dumazet9d185622016-05-01 16:47:26 -0700227 prev_qlen = sch->q.qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000228
Eric Dumazet80e509d2016-06-04 12:55:13 -0700229 /* save this packet length as it might be dropped by fq_codel_drop() */
230 pkt_len = qdisc_pkt_len(skb);
Eric Dumazet9d185622016-05-01 16:47:26 -0700231 /* fq_codel_drop() is quite expensive, as it performs a linear search
232 * in q->backlogs[] to find a fat flow.
233 * So instead of dropping a single packet, drop half of its backlog
234 * with a 64 packets limit to not add a too big cpu spike here.
235 */
Eric Dumazet520ac302016-06-21 23:16:49 -0700236 ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
Eric Dumazet9d185622016-05-01 16:47:26 -0700237
Eric Dumazet80e509d2016-06-04 12:55:13 -0700238 prev_qlen -= sch->q.qlen;
239 prev_backlog -= sch->qstats.backlog;
240 q->drop_overlimit += prev_qlen;
Eric Dumazet95b58432016-05-06 08:55:12 -0700241 if (memory_limited)
Eric Dumazet80e509d2016-06-04 12:55:13 -0700242 q->drop_overmemory += prev_qlen;
Eric Dumazet9d185622016-05-01 16:47:26 -0700243
Eric Dumazet80e509d2016-06-04 12:55:13 -0700244 /* As we dropped packet(s), better let upper stack know this.
245 * If we dropped a packet for this flow, return NET_XMIT_CN,
246 * but in this case, our parents wont increase their backlogs.
247 */
248 if (ret == idx) {
249 qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
250 prev_backlog - pkt_len);
251 return NET_XMIT_CN;
252 }
253 qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
254 return NET_XMIT_SUCCESS;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000255}
256
257/* This is the specific function called from codel_dequeue()
258 * to dequeue a packet from queue. Note: backlog is handled in
259 * codel, we dont need to reduce it here.
260 */
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200261static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000262{
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200263 struct Qdisc *sch = ctx;
Eric Dumazet865ec552012-05-16 04:39:09 +0000264 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000265 struct fq_codel_flow *flow;
266 struct sk_buff *skb = NULL;
267
268 flow = container_of(vars, struct fq_codel_flow, cvars);
269 if (flow->head) {
270 skb = dequeue_head(flow);
Eric Dumazet865ec552012-05-16 04:39:09 +0000271 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
Eric Dumazet008830b2016-06-21 23:16:50 -0700272 q->memory_usage -= get_codel_cb(skb)->mem_usage;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000273 sch->q.qlen--;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200274 sch->qstats.backlog -= qdisc_pkt_len(skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000275 }
276 return skb;
277}
278
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200279static void drop_func(struct sk_buff *skb, void *ctx)
280{
281 struct Qdisc *sch = ctx;
282
Eric Dumazet520ac302016-06-21 23:16:49 -0700283 kfree_skb(skb);
284 qdisc_qstats_drop(sch);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200285}
286
Eric Dumazet4b549a22012-05-11 09:30:50 +0000287static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
288{
289 struct fq_codel_sched_data *q = qdisc_priv(sch);
290 struct sk_buff *skb;
291 struct fq_codel_flow *flow;
292 struct list_head *head;
293 u32 prev_drop_count, prev_ecn_mark;
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
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200313 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
314 &flow->cvars, &q->cstats, qdisc_pkt_len,
315 codel_get_enqueue_time, drop_func, dequeue_func);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000316
317 flow->dropped += q->cstats.drop_count - prev_drop_count;
318 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
319
320 if (!skb) {
321 /* force a pass through old_flows to prevent starvation */
322 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
323 list_move_tail(&flow->flowchain, &q->old_flows);
324 else
325 list_del_init(&flow->flowchain);
326 goto begin;
327 }
328 qdisc_bstats_update(sch, skb);
329 flow->deficit -= qdisc_pkt_len(skb);
WANG Cong2ccccf52016-02-25 14:55:01 -0800330 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000331 * or HTB crashes. Defer it for next round.
332 */
333 if (q->cstats.drop_count && sch->q.qlen) {
WANG Cong2ccccf52016-02-25 14:55:01 -0800334 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
335 q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000336 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800337 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000338 }
339 return skb;
340}
341
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700342static void fq_codel_flow_purge(struct fq_codel_flow *flow)
343{
344 rtnl_kfree_skbs(flow->head, flow->tail);
345 flow->head = NULL;
346}
347
Eric Dumazet4b549a22012-05-11 09:30:50 +0000348static void fq_codel_reset(struct Qdisc *sch)
349{
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700350 struct fq_codel_sched_data *q = qdisc_priv(sch);
351 int i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000352
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700353 INIT_LIST_HEAD(&q->new_flows);
354 INIT_LIST_HEAD(&q->old_flows);
355 for (i = 0; i < q->flows_cnt; i++) {
356 struct fq_codel_flow *flow = q->flows + i;
357
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700358 fq_codel_flow_purge(flow);
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700359 INIT_LIST_HEAD(&flow->flowchain);
360 codel_vars_init(&flow->cvars);
361 }
362 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
363 sch->q.qlen = 0;
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700364 sch->qstats.backlog = 0;
Eric Dumazet77f57762016-05-15 18:16:38 -0700365 q->memory_usage = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000366}
367
368static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
369 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
370 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
371 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
372 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
373 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
374 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700375 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
Eric Dumazet9d185622016-05-01 16:47:26 -0700376 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
Eric Dumazet95b58432016-05-06 08:55:12 -0700377 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
Eric Dumazet4b549a22012-05-11 09:30:50 +0000378};
379
Alexander Aring20307212017-12-20 12:35:14 -0500380static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
381 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000382{
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
Johannes Bergfceb6432017-04-12 14:34:07 +0200390 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy,
391 NULL);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000392 if (err < 0)
393 return err;
394 if (tb[TCA_FQ_CODEL_FLOWS]) {
395 if (q->flows)
396 return -EINVAL;
397 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
398 if (!q->flows_cnt ||
399 q->flows_cnt > 65536)
400 return -EINVAL;
401 }
402 sch_tree_lock(sch);
403
404 if (tb[TCA_FQ_CODEL_TARGET]) {
405 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
406
407 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
408 }
409
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700410 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
411 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
412
413 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
414 }
415
Eric Dumazet4b549a22012-05-11 09:30:50 +0000416 if (tb[TCA_FQ_CODEL_INTERVAL]) {
417 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
418
419 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
420 }
421
422 if (tb[TCA_FQ_CODEL_LIMIT])
423 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
424
425 if (tb[TCA_FQ_CODEL_ECN])
426 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
427
428 if (tb[TCA_FQ_CODEL_QUANTUM])
429 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
430
Eric Dumazet9d185622016-05-01 16:47:26 -0700431 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
432 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
433
Eric Dumazet95b58432016-05-06 08:55:12 -0700434 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
435 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
436
437 while (sch->q.qlen > sch->limit ||
438 q->memory_usage > q->memory_limit) {
Eric Dumazet4b549a22012-05-11 09:30:50 +0000439 struct sk_buff *skb = fq_codel_dequeue(sch);
440
WANG Cong2ccccf52016-02-25 14:55:01 -0800441 q->cstats.drop_len += qdisc_pkt_len(skb);
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700442 rtnl_kfree_skbs(skb, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000443 q->cstats.drop_count++;
444 }
WANG Cong2ccccf52016-02-25 14:55:01 -0800445 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000446 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800447 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000448
449 sch_tree_unlock(sch);
450 return 0;
451}
452
Eric Dumazet4b549a22012-05-11 09:30:50 +0000453static void fq_codel_destroy(struct Qdisc *sch)
454{
455 struct fq_codel_sched_data *q = qdisc_priv(sch);
456
Jiri Pirko6529eab2017-05-17 11:07:55 +0200457 tcf_block_put(q->block);
Michal Hocko752ade62017-05-08 15:57:27 -0700458 kvfree(q->backlogs);
459 kvfree(q->flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000460}
461
Alexander Aringe63d7df2017-12-20 12:35:13 -0500462static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
463 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000464{
465 struct fq_codel_sched_data *q = qdisc_priv(sch);
466 int i;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200467 int err;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000468
469 sch->limit = 10*1024;
470 q->flows_cnt = 1024;
Eric Dumazet95b58432016-05-06 08:55:12 -0700471 q->memory_limit = 32 << 20; /* 32 MBytes */
Eric Dumazet9d185622016-05-01 16:47:26 -0700472 q->drop_batch_size = 64;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000473 q->quantum = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000474 INIT_LIST_HEAD(&q->new_flows);
475 INIT_LIST_HEAD(&q->old_flows);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200476 codel_params_init(&q->cparams);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000477 codel_stats_init(&q->cstats);
478 q->cparams.ecn = true;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200479 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000480
481 if (opt) {
Alexander Aring793d81d2017-12-20 12:35:15 -0500482 int err = fq_codel_change(sch, opt, extack);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000483 if (err)
484 return err;
485 }
486
Alexander Aring8d1a77f2017-12-20 12:35:19 -0500487 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
Jiri Pirko6529eab2017-05-17 11:07:55 +0200488 if (err)
489 return err;
490
Eric Dumazet4b549a22012-05-11 09:30:50 +0000491 if (!q->flows) {
Michal Hocko752ade62017-05-08 15:57:27 -0700492 q->flows = kvzalloc(q->flows_cnt *
493 sizeof(struct fq_codel_flow), GFP_KERNEL);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000494 if (!q->flows)
495 return -ENOMEM;
Michal Hocko752ade62017-05-08 15:57:27 -0700496 q->backlogs = kvzalloc(q->flows_cnt * sizeof(u32), GFP_KERNEL);
Nikolay Aleksandrov30c31d72017-08-30 12:49:02 +0300497 if (!q->backlogs)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000498 return -ENOMEM;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000499 for (i = 0; i < q->flows_cnt; i++) {
500 struct fq_codel_flow *flow = q->flows + i;
501
502 INIT_LIST_HEAD(&flow->flowchain);
Eric Dumazetb3791352012-09-01 03:19:57 +0000503 codel_vars_init(&flow->cvars);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000504 }
505 }
506 if (sch->limit >= 1)
507 sch->flags |= TCQ_F_CAN_BYPASS;
508 else
509 sch->flags &= ~TCQ_F_CAN_BYPASS;
510 return 0;
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(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) ||
Eric Dumazet9d185622016-05-01 16:47:26 -0700532 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
533 q->drop_batch_size) ||
Eric Dumazet95b58432016-05-06 08:55:12 -0700534 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
535 q->memory_limit) ||
Eric Dumazet4b549a22012-05-11 09:30:50 +0000536 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
537 q->flows_cnt))
538 goto nla_put_failure;
539
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700540 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
Yang Yingliangd59b7d82014-03-12 10:20:32 +0800545 return nla_nest_end(skb, opts);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000546
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,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000556 };
557 struct list_head *pos;
558
Sasha Levin669d67b2012-05-14 11:57:06 +0000559 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;
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700563 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
Eric Dumazet95b58432016-05-06 08:55:12 -0700564 st.qdisc_stats.memory_usage = q->memory_usage;
565 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
Sasha Levin669d67b2012-05-14 11:57:06 +0000566
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700567 sch_tree_lock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000568 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++;
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700573 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000574
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
WANG Cong143976c2017-08-24 16:51:29 -0700583static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000584{
585 return 0;
586}
587
588static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
589 u32 classid)
590{
591 /* we cannot bypass queue discipline anymore */
592 sch->flags &= ~TCQ_F_CAN_BYPASS;
593 return 0;
594}
595
WANG Cong143976c2017-08-24 16:51:29 -0700596static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000597{
598}
599
Alexander Aringcbaacc42017-12-20 12:35:16 -0500600static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
601 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000602{
603 struct fq_codel_sched_data *q = qdisc_priv(sch);
604
605 if (cl)
606 return NULL;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200607 return q->block;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000608}
609
610static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
611 struct sk_buff *skb, struct tcmsg *tcm)
612{
613 tcm->tcm_handle |= TC_H_MIN(cl);
614 return 0;
615}
616
617static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
618 struct gnet_dump *d)
619{
620 struct fq_codel_sched_data *q = qdisc_priv(sch);
621 u32 idx = cl - 1;
622 struct gnet_stats_queue qs = { 0 };
623 struct tc_fq_codel_xstats xstats;
624
625 if (idx < q->flows_cnt) {
626 const struct fq_codel_flow *flow = &q->flows[idx];
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700627 const struct sk_buff *skb;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000628
629 memset(&xstats, 0, sizeof(xstats));
630 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
631 xstats.class_stats.deficit = flow->deficit;
632 xstats.class_stats.ldelay =
633 codel_time_to_us(flow->cvars.ldelay);
634 xstats.class_stats.count = flow->cvars.count;
635 xstats.class_stats.lastcount = flow->cvars.lastcount;
636 xstats.class_stats.dropping = flow->cvars.dropping;
637 if (flow->cvars.dropping) {
638 codel_tdiff_t delta = flow->cvars.drop_next -
639 codel_get_time();
640
641 xstats.class_stats.drop_next = (delta >= 0) ?
642 codel_time_to_us(delta) :
643 -codel_time_to_us(-delta);
644 }
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700645 if (flow->head) {
646 sch_tree_lock(sch);
647 skb = flow->head;
648 while (skb) {
649 qs.qlen++;
650 skb = skb->next;
651 }
652 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000653 }
654 qs.backlog = q->backlogs[idx];
655 qs.drops = flow->dropped;
656 }
Eric Dumazetaafddbf2016-06-06 09:12:39 -0700657 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000658 return -1;
659 if (idx < q->flows_cnt)
660 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
661 return 0;
662}
663
664static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
665{
666 struct fq_codel_sched_data *q = qdisc_priv(sch);
667 unsigned int i;
668
669 if (arg->stop)
670 return;
671
672 for (i = 0; i < q->flows_cnt; i++) {
673 if (list_empty(&q->flows[i].flowchain) ||
674 arg->count < arg->skip) {
675 arg->count++;
676 continue;
677 }
678 if (arg->fn(sch, i + 1, arg) < 0) {
679 arg->stop = 1;
680 break;
681 }
682 arg->count++;
683 }
684}
685
686static const struct Qdisc_class_ops fq_codel_class_ops = {
687 .leaf = fq_codel_leaf,
WANG Cong143976c2017-08-24 16:51:29 -0700688 .find = fq_codel_find,
Jiri Pirko6529eab2017-05-17 11:07:55 +0200689 .tcf_block = fq_codel_tcf_block,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000690 .bind_tcf = fq_codel_bind,
WANG Cong143976c2017-08-24 16:51:29 -0700691 .unbind_tcf = fq_codel_unbind,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000692 .dump = fq_codel_dump_class,
693 .dump_stats = fq_codel_dump_class_stats,
694 .walk = fq_codel_walk,
695};
696
697static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
698 .cl_ops = &fq_codel_class_ops,
699 .id = "fq_codel",
700 .priv_size = sizeof(struct fq_codel_sched_data),
701 .enqueue = fq_codel_enqueue,
702 .dequeue = fq_codel_dequeue,
703 .peek = qdisc_peek_dequeued,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000704 .init = fq_codel_init,
705 .reset = fq_codel_reset,
706 .destroy = fq_codel_destroy,
707 .change = fq_codel_change,
708 .dump = fq_codel_dump,
709 .dump_stats = fq_codel_dump_stats,
710 .owner = THIS_MODULE,
711};
712
713static int __init fq_codel_module_init(void)
714{
715 return register_qdisc(&fq_codel_qdisc_ops);
716}
717
718static void __exit fq_codel_module_exit(void)
719{
720 unregister_qdisc(&fq_codel_qdisc_ops);
721}
722
723module_init(fq_codel_module_init)
724module_exit(fq_codel_module_exit)
725MODULE_AUTHOR("Eric Dumazet");
726MODULE_LICENSE("GPL");