blob: bbd5f8753600624f2eba9a91d2efa21333f4221f [file] [log] [blame]
Thomas Gleixner2874c5f2019-05-27 08:55:01 +02001// SPDX-License-Identifier: GPL-2.0-or-later
Eric Dumazet4b549a22012-05-11 09:30:50 +00002/*
3 * Fair Queue CoDel discipline
4 *
Eric Dumazet80ba92f2015-05-08 15:05:12 -07005 * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
Eric Dumazet4b549a22012-05-11 09:30:50 +00006 */
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>
Eric Dumazet4b549a22012-05-11 09:30:50 +000017#include <linux/slab.h>
18#include <linux/vmalloc.h>
19#include <net/netlink.h>
20#include <net/pkt_sched.h>
Jiri Pirkocf1facd2017-02-09 14:38:56 +010021#include <net/pkt_cls.h>
Eric Dumazet4b549a22012-05-11 09:30:50 +000022#include <net/codel.h>
Michal Kaziord068ca22016-04-22 14:15:59 +020023#include <net/codel_impl.h>
24#include <net/codel_qdisc.h>
Eric Dumazet4b549a22012-05-11 09:30:50 +000025
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;
Eric Dumazet4b549a22012-05-11 09:30:50 +000047 struct codel_vars cvars;
48}; /* please try to keep this structure <= 64 bytes */
49
50struct fq_codel_sched_data {
John Fastabend25d8c0d2014-09-12 20:05:27 -070051 struct tcf_proto __rcu *filter_list; /* optional external classifier */
Jiri Pirko6529eab2017-05-17 11:07:55 +020052 struct tcf_block *block;
Eric Dumazet4b549a22012-05-11 09:30:50 +000053 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
54 u32 *backlogs; /* backlog table [flows_cnt] */
55 u32 flows_cnt; /* number of flows */
Eric Dumazet4b549a22012-05-11 09:30:50 +000056 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
Eric Dumazet9d185622016-05-01 16:47:26 -070057 u32 drop_batch_size;
Eric Dumazet95b58432016-05-06 08:55:12 -070058 u32 memory_limit;
Eric Dumazet4b549a22012-05-11 09:30:50 +000059 struct codel_params cparams;
60 struct codel_stats cstats;
Eric Dumazet95b58432016-05-06 08:55:12 -070061 u32 memory_usage;
62 u32 drop_overmemory;
Eric Dumazet4b549a22012-05-11 09:30:50 +000063 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,
Tom Herbert342db222015-05-01 11:30:13 -070071 struct sk_buff *skb)
Eric Dumazet4b549a22012-05-11 09:30:50 +000072{
Andrew Collins264b87f2017-01-18 14:04:28 -070073 return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
Eric Dumazet4b549a22012-05-11 09:30:50 +000074}
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);
John Fastabend25d8c0d2014-09-12 20:05:27 -070080 struct tcf_proto *filter;
Eric Dumazet4b549a22012-05-11 09:30:50 +000081 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
Valdis.Kletnieks@vt.edu69204cf2014-12-09 16:15:50 -050089 filter = rcu_dereference_bh(q->filter_list);
John Fastabend25d8c0d2014-09-12 20:05:27 -070090 if (!filter)
Eric Dumazet4b549a22012-05-11 09:30:50 +000091 return fq_codel_hash(q, skb) + 1;
92
93 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
Jiri Pirko87d83092017-05-17 11:07:54 +020094 result = tcf_classify(skb, filter, &res, false);
Eric Dumazet4b549a22012-05-11 09:30:50 +000095 if (result >= 0) {
96#ifdef CONFIG_NET_CLS_ACT
97 switch (result) {
98 case TC_ACT_STOLEN:
99 case TC_ACT_QUEUED:
Jiri Pirkoe25ea212017-06-06 14:12:02 +0200100 case TC_ACT_TRAP:
Eric Dumazet4b549a22012-05-11 09:30:50 +0000101 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
Gustavo A. R. Silva964201d2020-07-07 12:21:38 -0500102 fallthrough;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000103 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;
David S. Millera8305bf2018-07-29 20:42:53 -0700121 skb_mark_not_on_list(skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000122 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
Eric Dumazet520ac302016-06-21 23:16:49 -0700137static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
138 struct sk_buff **to_free)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000139{
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;
Eric Dumazet9d185622016-05-01 16:47:26 -0700144 unsigned int threshold;
Eric Dumazet95b58432016-05-06 08:55:12 -0700145 unsigned int mem = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000146
Eric Dumazet9d185622016-05-01 16:47:26 -0700147 /* Queue is full! Find the fat flow and drop packet(s) from it.
Eric Dumazet4b549a22012-05-11 09:30:50 +0000148 * 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.
Eric Dumazet9d185622016-05-01 16:47:26 -0700151 * In stress mode, we'll try to drop 64 packets from the flow,
152 * amortizing this linear lookup to one cache line per drop.
Eric Dumazet4b549a22012-05-11 09:30:50 +0000153 */
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 }
Eric Dumazet9d185622016-05-01 16:47:26 -0700160
161 /* Our goal is to drop half of this fat flow backlog */
162 threshold = maxbacklog >> 1;
163
Eric Dumazet4b549a22012-05-11 09:30:50 +0000164 flow = &q->flows[idx];
Eric Dumazet9d185622016-05-01 16:47:26 -0700165 len = 0;
166 i = 0;
167 do {
168 skb = dequeue_head(flow);
169 len += qdisc_pkt_len(skb);
Eric Dumazet008830b2016-06-21 23:16:50 -0700170 mem += get_codel_cb(skb)->mem_usage;
Eric Dumazet520ac302016-06-21 23:16:49 -0700171 __qdisc_drop(skb, to_free);
Eric Dumazet9d185622016-05-01 16:47:26 -0700172 } while (++i < max_packets && len < threshold);
173
Dave Tahtae697f32019-08-03 16:37:28 -0700174 /* Tell codel to increase its signal strength also */
175 flow->cvars.count += i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000176 q->backlogs[idx] -= len;
Eric Dumazet95b58432016-05-06 08:55:12 -0700177 q->memory_usage -= mem;
Eric Dumazet9d185622016-05-01 16:47:26 -0700178 sch->qstats.drops += i;
179 sch->qstats.backlog -= len;
180 sch->q.qlen -= i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000181 return idx;
182}
183
Eric Dumazet520ac302016-06-21 23:16:49 -0700184static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
185 struct sk_buff **to_free)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000186{
187 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet9d185622016-05-01 16:47:26 -0700188 unsigned int idx, prev_backlog, prev_qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000189 struct fq_codel_flow *flow;
Kees Cook3f649ab2020-06-03 13:09:38 -0700190 int ret;
Eric Dumazet80e509d2016-06-04 12:55:13 -0700191 unsigned int pkt_len;
Eric Dumazet95b58432016-05-06 08:55:12 -0700192 bool memory_limited;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000193
194 idx = fq_codel_classify(skb, sch, &ret);
195 if (idx == 0) {
196 if (ret & __NET_XMIT_BYPASS)
John Fastabend25331d62014-09-28 11:53:29 -0700197 qdisc_qstats_drop(sch);
Eric Dumazet520ac302016-06-21 23:16:49 -0700198 __qdisc_drop(skb, to_free);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000199 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);
John Fastabend25331d62014-09-28 11:53:29 -0700207 qdisc_qstats_backlog_inc(sch, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000208
209 if (list_empty(&flow->flowchain)) {
210 list_add_tail(&flow->flowchain, &q->new_flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000211 q->new_flow_count++;
212 flow->deficit = q->quantum;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000213 }
Eric Dumazet008830b2016-06-21 23:16:50 -0700214 get_codel_cb(skb)->mem_usage = skb->truesize;
215 q->memory_usage += get_codel_cb(skb)->mem_usage;
Eric Dumazet95b58432016-05-06 08:55:12 -0700216 memory_limited = q->memory_usage > q->memory_limit;
217 if (++sch->q.qlen <= sch->limit && !memory_limited)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000218 return NET_XMIT_SUCCESS;
219
WANG Cong2ccccf52016-02-25 14:55:01 -0800220 prev_backlog = sch->qstats.backlog;
Eric Dumazet9d185622016-05-01 16:47:26 -0700221 prev_qlen = sch->q.qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000222
Eric Dumazet80e509d2016-06-04 12:55:13 -0700223 /* save this packet length as it might be dropped by fq_codel_drop() */
224 pkt_len = qdisc_pkt_len(skb);
Eric Dumazet9d185622016-05-01 16:47:26 -0700225 /* 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 */
Eric Dumazet520ac302016-06-21 23:16:49 -0700230 ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
Eric Dumazet9d185622016-05-01 16:47:26 -0700231
Eric Dumazet80e509d2016-06-04 12:55:13 -0700232 prev_qlen -= sch->q.qlen;
233 prev_backlog -= sch->qstats.backlog;
234 q->drop_overlimit += prev_qlen;
Eric Dumazet95b58432016-05-06 08:55:12 -0700235 if (memory_limited)
Eric Dumazet80e509d2016-06-04 12:55:13 -0700236 q->drop_overmemory += prev_qlen;
Eric Dumazet9d185622016-05-01 16:47:26 -0700237
Eric Dumazet80e509d2016-06-04 12:55:13 -0700238 /* 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;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000249}
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 */
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200255static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000256{
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200257 struct Qdisc *sch = ctx;
Eric Dumazet865ec552012-05-16 04:39:09 +0000258 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000259 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);
Eric Dumazet865ec552012-05-16 04:39:09 +0000265 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
Eric Dumazet008830b2016-06-21 23:16:50 -0700266 q->memory_usage -= get_codel_cb(skb)->mem_usage;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000267 sch->q.qlen--;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200268 sch->qstats.backlog -= qdisc_pkt_len(skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000269 }
270 return skb;
271}
272
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200273static void drop_func(struct sk_buff *skb, void *ctx)
274{
275 struct Qdisc *sch = ctx;
276
Eric Dumazet520ac302016-06-21 23:16:49 -0700277 kfree_skb(skb);
278 qdisc_qstats_drop(sch);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200279}
280
Eric Dumazet4b549a22012-05-11 09:30:50 +0000281static 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;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000287
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
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200303 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);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000306
Eric Dumazet4b549a22012-05-11 09:30:50 +0000307 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);
WANG Cong2ccccf52016-02-25 14:55:01 -0800317 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000318 * or HTB crashes. Defer it for next round.
319 */
320 if (q->cstats.drop_count && sch->q.qlen) {
WANG Cong2ccccf52016-02-25 14:55:01 -0800321 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
322 q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000323 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800324 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000325 }
326 return skb;
327}
328
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700329static 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
Eric Dumazet4b549a22012-05-11 09:30:50 +0000335static void fq_codel_reset(struct Qdisc *sch)
336{
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700337 struct fq_codel_sched_data *q = qdisc_priv(sch);
338 int i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000339
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700340 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
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700345 fq_codel_flow_purge(flow);
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700346 INIT_LIST_HEAD(&flow->flowchain);
347 codel_vars_init(&flow->cvars);
348 }
349 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
350 sch->q.qlen = 0;
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700351 sch->qstats.backlog = 0;
Eric Dumazet77f57762016-05-15 18:16:38 -0700352 q->memory_usage = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000353}
354
355static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
356 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
357 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
358 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
359 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
360 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
361 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700362 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
Eric Dumazet9d185622016-05-01 16:47:26 -0700363 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
Eric Dumazet95b58432016-05-06 08:55:12 -0700364 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
Eric Dumazet4b549a22012-05-11 09:30:50 +0000365};
366
Alexander Aring20307212017-12-20 12:35:14 -0500367static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
368 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000369{
370 struct fq_codel_sched_data *q = qdisc_priv(sch);
371 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
372 int err;
373
374 if (!opt)
375 return -EINVAL;
376
Johannes Berg8cb08172019-04-26 14:07:28 +0200377 err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
378 fq_codel_policy, NULL);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000379 if (err < 0)
380 return err;
381 if (tb[TCA_FQ_CODEL_FLOWS]) {
382 if (q->flows)
383 return -EINVAL;
384 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
385 if (!q->flows_cnt ||
386 q->flows_cnt > 65536)
387 return -EINVAL;
388 }
389 sch_tree_lock(sch);
390
391 if (tb[TCA_FQ_CODEL_TARGET]) {
392 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
393
394 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
395 }
396
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700397 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
398 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
399
400 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
401 }
402
Eric Dumazet4b549a22012-05-11 09:30:50 +0000403 if (tb[TCA_FQ_CODEL_INTERVAL]) {
404 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
405
406 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
407 }
408
409 if (tb[TCA_FQ_CODEL_LIMIT])
410 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
411
412 if (tb[TCA_FQ_CODEL_ECN])
413 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
414
415 if (tb[TCA_FQ_CODEL_QUANTUM])
416 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
417
Eric Dumazet9d185622016-05-01 16:47:26 -0700418 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
Eric Dumazet14695212020-04-25 12:40:25 -0700419 q->drop_batch_size = max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
Eric Dumazet9d185622016-05-01 16:47:26 -0700420
Eric Dumazet95b58432016-05-06 08:55:12 -0700421 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
422 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
423
424 while (sch->q.qlen > sch->limit ||
425 q->memory_usage > q->memory_limit) {
Eric Dumazet4b549a22012-05-11 09:30:50 +0000426 struct sk_buff *skb = fq_codel_dequeue(sch);
427
WANG Cong2ccccf52016-02-25 14:55:01 -0800428 q->cstats.drop_len += qdisc_pkt_len(skb);
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700429 rtnl_kfree_skbs(skb, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000430 q->cstats.drop_count++;
431 }
WANG Cong2ccccf52016-02-25 14:55:01 -0800432 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000433 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800434 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000435
436 sch_tree_unlock(sch);
437 return 0;
438}
439
Eric Dumazet4b549a22012-05-11 09:30:50 +0000440static void fq_codel_destroy(struct Qdisc *sch)
441{
442 struct fq_codel_sched_data *q = qdisc_priv(sch);
443
Jiri Pirko6529eab2017-05-17 11:07:55 +0200444 tcf_block_put(q->block);
Michal Hocko752ade62017-05-08 15:57:27 -0700445 kvfree(q->backlogs);
446 kvfree(q->flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000447}
448
Alexander Aringe63d7df2017-12-20 12:35:13 -0500449static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
450 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000451{
452 struct fq_codel_sched_data *q = qdisc_priv(sch);
453 int i;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200454 int err;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000455
456 sch->limit = 10*1024;
457 q->flows_cnt = 1024;
Eric Dumazet95b58432016-05-06 08:55:12 -0700458 q->memory_limit = 32 << 20; /* 32 MBytes */
Eric Dumazet9d185622016-05-01 16:47:26 -0700459 q->drop_batch_size = 64;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000460 q->quantum = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000461 INIT_LIST_HEAD(&q->new_flows);
462 INIT_LIST_HEAD(&q->old_flows);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200463 codel_params_init(&q->cparams);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000464 codel_stats_init(&q->cstats);
465 q->cparams.ecn = true;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200466 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000467
468 if (opt) {
Jacob Keller83fe6b82018-07-10 14:22:27 -0700469 err = fq_codel_change(sch, opt, extack);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000470 if (err)
Jacob Keller83fe6b82018-07-10 14:22:27 -0700471 goto init_failure;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000472 }
473
Alexander Aring8d1a77f2017-12-20 12:35:19 -0500474 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
Jiri Pirko6529eab2017-05-17 11:07:55 +0200475 if (err)
Jacob Keller83fe6b82018-07-10 14:22:27 -0700476 goto init_failure;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200477
Eric Dumazet4b549a22012-05-11 09:30:50 +0000478 if (!q->flows) {
Kees Cook778e1cd2018-06-12 14:04:48 -0700479 q->flows = kvcalloc(q->flows_cnt,
480 sizeof(struct fq_codel_flow),
481 GFP_KERNEL);
Jacob Keller83fe6b82018-07-10 14:22:27 -0700482 if (!q->flows) {
483 err = -ENOMEM;
484 goto init_failure;
485 }
Kees Cook778e1cd2018-06-12 14:04:48 -0700486 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
Jacob Keller83fe6b82018-07-10 14:22:27 -0700487 if (!q->backlogs) {
488 err = -ENOMEM;
489 goto alloc_failure;
490 }
Eric Dumazet4b549a22012-05-11 09:30:50 +0000491 for (i = 0; i < q->flows_cnt; i++) {
492 struct fq_codel_flow *flow = q->flows + i;
493
494 INIT_LIST_HEAD(&flow->flowchain);
Eric Dumazetb3791352012-09-01 03:19:57 +0000495 codel_vars_init(&flow->cvars);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000496 }
497 }
498 if (sch->limit >= 1)
499 sch->flags |= TCQ_F_CAN_BYPASS;
500 else
501 sch->flags &= ~TCQ_F_CAN_BYPASS;
502 return 0;
Jacob Keller83fe6b82018-07-10 14:22:27 -0700503
504alloc_failure:
505 kvfree(q->flows);
506 q->flows = NULL;
507init_failure:
508 q->flows_cnt = 0;
509 return err;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000510}
511
512static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
513{
514 struct fq_codel_sched_data *q = qdisc_priv(sch);
515 struct nlattr *opts;
516
Michal Kubecekae0be8d2019-04-26 11:13:06 +0200517 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000518 if (opts == NULL)
519 goto nla_put_failure;
520
521 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
522 codel_time_to_us(q->cparams.target)) ||
523 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
524 sch->limit) ||
525 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
526 codel_time_to_us(q->cparams.interval)) ||
527 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
528 q->cparams.ecn) ||
529 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
530 q->quantum) ||
Eric Dumazet9d185622016-05-01 16:47:26 -0700531 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
532 q->drop_batch_size) ||
Eric Dumazet95b58432016-05-06 08:55:12 -0700533 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
534 q->memory_limit) ||
Eric Dumazet4b549a22012-05-11 09:30:50 +0000535 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
536 q->flows_cnt))
537 goto nla_put_failure;
538
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700539 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
540 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
541 codel_time_to_us(q->cparams.ce_threshold)))
542 goto nla_put_failure;
543
Yang Yingliangd59b7d82014-03-12 10:20:32 +0800544 return nla_nest_end(skb, opts);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000545
546nla_put_failure:
547 return -1;
548}
549
550static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
551{
552 struct fq_codel_sched_data *q = qdisc_priv(sch);
553 struct tc_fq_codel_xstats st = {
554 .type = TCA_FQ_CODEL_XSTATS_QDISC,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000555 };
556 struct list_head *pos;
557
Sasha Levin669d67b2012-05-14 11:57:06 +0000558 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
559 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
560 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
561 st.qdisc_stats.new_flow_count = q->new_flow_count;
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700562 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
Eric Dumazet95b58432016-05-06 08:55:12 -0700563 st.qdisc_stats.memory_usage = q->memory_usage;
564 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
Sasha Levin669d67b2012-05-14 11:57:06 +0000565
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700566 sch_tree_lock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000567 list_for_each(pos, &q->new_flows)
568 st.qdisc_stats.new_flows_len++;
569
570 list_for_each(pos, &q->old_flows)
571 st.qdisc_stats.old_flows_len++;
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700572 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000573
574 return gnet_stats_copy_app(d, &st, sizeof(st));
575}
576
577static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
578{
579 return NULL;
580}
581
WANG Cong143976c2017-08-24 16:51:29 -0700582static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000583{
584 return 0;
585}
586
587static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
588 u32 classid)
589{
Eric Dumazet4b549a22012-05-11 09:30:50 +0000590 return 0;
591}
592
WANG Cong143976c2017-08-24 16:51:29 -0700593static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000594{
595}
596
Alexander Aringcbaacc42017-12-20 12:35:16 -0500597static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
598 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000599{
600 struct fq_codel_sched_data *q = qdisc_priv(sch);
601
602 if (cl)
603 return NULL;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200604 return q->block;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000605}
606
607static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
608 struct sk_buff *skb, struct tcmsg *tcm)
609{
610 tcm->tcm_handle |= TC_H_MIN(cl);
611 return 0;
612}
613
614static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
615 struct gnet_dump *d)
616{
617 struct fq_codel_sched_data *q = qdisc_priv(sch);
618 u32 idx = cl - 1;
619 struct gnet_stats_queue qs = { 0 };
620 struct tc_fq_codel_xstats xstats;
621
622 if (idx < q->flows_cnt) {
623 const struct fq_codel_flow *flow = &q->flows[idx];
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700624 const struct sk_buff *skb;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000625
626 memset(&xstats, 0, sizeof(xstats));
627 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
628 xstats.class_stats.deficit = flow->deficit;
629 xstats.class_stats.ldelay =
630 codel_time_to_us(flow->cvars.ldelay);
631 xstats.class_stats.count = flow->cvars.count;
632 xstats.class_stats.lastcount = flow->cvars.lastcount;
633 xstats.class_stats.dropping = flow->cvars.dropping;
634 if (flow->cvars.dropping) {
635 codel_tdiff_t delta = flow->cvars.drop_next -
636 codel_get_time();
637
638 xstats.class_stats.drop_next = (delta >= 0) ?
639 codel_time_to_us(delta) :
640 -codel_time_to_us(-delta);
641 }
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700642 if (flow->head) {
643 sch_tree_lock(sch);
644 skb = flow->head;
645 while (skb) {
646 qs.qlen++;
647 skb = skb->next;
648 }
649 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000650 }
651 qs.backlog = q->backlogs[idx];
Dave Taht77ddaff2019-08-03 16:37:29 -0700652 qs.drops = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000653 }
Eric Dumazetaafddbf2016-06-06 09:12:39 -0700654 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000655 return -1;
656 if (idx < q->flows_cnt)
657 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
658 return 0;
659}
660
661static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
662{
663 struct fq_codel_sched_data *q = qdisc_priv(sch);
664 unsigned int i;
665
666 if (arg->stop)
667 return;
668
669 for (i = 0; i < q->flows_cnt; i++) {
670 if (list_empty(&q->flows[i].flowchain) ||
671 arg->count < arg->skip) {
672 arg->count++;
673 continue;
674 }
675 if (arg->fn(sch, i + 1, arg) < 0) {
676 arg->stop = 1;
677 break;
678 }
679 arg->count++;
680 }
681}
682
683static const struct Qdisc_class_ops fq_codel_class_ops = {
684 .leaf = fq_codel_leaf,
WANG Cong143976c2017-08-24 16:51:29 -0700685 .find = fq_codel_find,
Jiri Pirko6529eab2017-05-17 11:07:55 +0200686 .tcf_block = fq_codel_tcf_block,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000687 .bind_tcf = fq_codel_bind,
WANG Cong143976c2017-08-24 16:51:29 -0700688 .unbind_tcf = fq_codel_unbind,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000689 .dump = fq_codel_dump_class,
690 .dump_stats = fq_codel_dump_class_stats,
691 .walk = fq_codel_walk,
692};
693
694static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
695 .cl_ops = &fq_codel_class_ops,
696 .id = "fq_codel",
697 .priv_size = sizeof(struct fq_codel_sched_data),
698 .enqueue = fq_codel_enqueue,
699 .dequeue = fq_codel_dequeue,
700 .peek = qdisc_peek_dequeued,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000701 .init = fq_codel_init,
702 .reset = fq_codel_reset,
703 .destroy = fq_codel_destroy,
704 .change = fq_codel_change,
705 .dump = fq_codel_dump,
706 .dump_stats = fq_codel_dump_stats,
707 .owner = THIS_MODULE,
708};
709
710static int __init fq_codel_module_init(void)
711{
712 return register_qdisc(&fq_codel_qdisc_ops);
713}
714
715static void __exit fq_codel_module_exit(void)
716{
717 unregister_qdisc(&fq_codel_qdisc_ops);
718}
719
720module_init(fq_codel_module_init)
721module_exit(fq_codel_module_exit)
722MODULE_AUTHOR("Eric Dumazet");
723MODULE_LICENSE("GPL");
Rob Gill67c20de2020-06-20 02:08:25 +0000724MODULE_DESCRIPTION("Fair Queue CoDel discipline");