blob: bb0cd6d3d2c2749d54e26368fb2558beedea85c9 [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;
Davide Caratti3aa26052021-07-28 20:08:00 +020094 result = tcf_classify(skb, NULL, 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];
Eric Dumazetc7c5e6f2021-09-03 15:03:43 -0700372 u32 quantum = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000373 int err;
374
375 if (!opt)
376 return -EINVAL;
377
Johannes Berg8cb08172019-04-26 14:07:28 +0200378 err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
379 fq_codel_policy, NULL);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000380 if (err < 0)
381 return err;
382 if (tb[TCA_FQ_CODEL_FLOWS]) {
383 if (q->flows)
384 return -EINVAL;
385 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
386 if (!q->flows_cnt ||
387 q->flows_cnt > 65536)
388 return -EINVAL;
389 }
Eric Dumazetc7c5e6f2021-09-03 15:03:43 -0700390 if (tb[TCA_FQ_CODEL_QUANTUM]) {
391 quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
392 if (quantum > FQ_CODEL_QUANTUM_MAX) {
393 NL_SET_ERR_MSG(extack, "Invalid quantum");
394 return -EINVAL;
395 }
396 }
Eric Dumazet4b549a22012-05-11 09:30:50 +0000397 sch_tree_lock(sch);
398
399 if (tb[TCA_FQ_CODEL_TARGET]) {
400 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
401
402 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
403 }
404
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700405 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
406 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
407
408 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
409 }
410
Eric Dumazet4b549a22012-05-11 09:30:50 +0000411 if (tb[TCA_FQ_CODEL_INTERVAL]) {
412 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
413
414 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
415 }
416
417 if (tb[TCA_FQ_CODEL_LIMIT])
418 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
419
420 if (tb[TCA_FQ_CODEL_ECN])
421 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
422
Eric Dumazetc7c5e6f2021-09-03 15:03:43 -0700423 if (quantum)
424 q->quantum = quantum;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000425
Eric Dumazet9d185622016-05-01 16:47:26 -0700426 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
Eric Dumazet14695212020-04-25 12:40:25 -0700427 q->drop_batch_size = max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
Eric Dumazet9d185622016-05-01 16:47:26 -0700428
Eric Dumazet95b58432016-05-06 08:55:12 -0700429 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
430 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
431
432 while (sch->q.qlen > sch->limit ||
433 q->memory_usage > q->memory_limit) {
Eric Dumazet4b549a22012-05-11 09:30:50 +0000434 struct sk_buff *skb = fq_codel_dequeue(sch);
435
WANG Cong2ccccf52016-02-25 14:55:01 -0800436 q->cstats.drop_len += qdisc_pkt_len(skb);
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700437 rtnl_kfree_skbs(skb, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000438 q->cstats.drop_count++;
439 }
WANG Cong2ccccf52016-02-25 14:55:01 -0800440 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000441 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800442 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000443
444 sch_tree_unlock(sch);
445 return 0;
446}
447
Eric Dumazet4b549a22012-05-11 09:30:50 +0000448static void fq_codel_destroy(struct Qdisc *sch)
449{
450 struct fq_codel_sched_data *q = qdisc_priv(sch);
451
Jiri Pirko6529eab2017-05-17 11:07:55 +0200452 tcf_block_put(q->block);
Michal Hocko752ade62017-05-08 15:57:27 -0700453 kvfree(q->backlogs);
454 kvfree(q->flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000455}
456
Alexander Aringe63d7df2017-12-20 12:35:13 -0500457static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
458 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000459{
460 struct fq_codel_sched_data *q = qdisc_priv(sch);
461 int i;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200462 int err;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000463
464 sch->limit = 10*1024;
465 q->flows_cnt = 1024;
Eric Dumazet95b58432016-05-06 08:55:12 -0700466 q->memory_limit = 32 << 20; /* 32 MBytes */
Eric Dumazet9d185622016-05-01 16:47:26 -0700467 q->drop_batch_size = 64;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000468 q->quantum = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000469 INIT_LIST_HEAD(&q->new_flows);
470 INIT_LIST_HEAD(&q->old_flows);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200471 codel_params_init(&q->cparams);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000472 codel_stats_init(&q->cstats);
473 q->cparams.ecn = true;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200474 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000475
476 if (opt) {
Jacob Keller83fe6b82018-07-10 14:22:27 -0700477 err = fq_codel_change(sch, opt, extack);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000478 if (err)
Jacob Keller83fe6b82018-07-10 14:22:27 -0700479 goto init_failure;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000480 }
481
Alexander Aring8d1a77f2017-12-20 12:35:19 -0500482 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
Jiri Pirko6529eab2017-05-17 11:07:55 +0200483 if (err)
Jacob Keller83fe6b82018-07-10 14:22:27 -0700484 goto init_failure;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200485
Eric Dumazet4b549a22012-05-11 09:30:50 +0000486 if (!q->flows) {
Kees Cook778e1cd2018-06-12 14:04:48 -0700487 q->flows = kvcalloc(q->flows_cnt,
488 sizeof(struct fq_codel_flow),
489 GFP_KERNEL);
Jacob Keller83fe6b82018-07-10 14:22:27 -0700490 if (!q->flows) {
491 err = -ENOMEM;
492 goto init_failure;
493 }
Kees Cook778e1cd2018-06-12 14:04:48 -0700494 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
Jacob Keller83fe6b82018-07-10 14:22:27 -0700495 if (!q->backlogs) {
496 err = -ENOMEM;
497 goto alloc_failure;
498 }
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;
Jacob Keller83fe6b82018-07-10 14:22:27 -0700511
512alloc_failure:
513 kvfree(q->flows);
514 q->flows = NULL;
515init_failure:
516 q->flows_cnt = 0;
517 return err;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000518}
519
520static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
521{
522 struct fq_codel_sched_data *q = qdisc_priv(sch);
523 struct nlattr *opts;
524
Michal Kubecekae0be8d2019-04-26 11:13:06 +0200525 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000526 if (opts == NULL)
527 goto nla_put_failure;
528
529 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
530 codel_time_to_us(q->cparams.target)) ||
531 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
532 sch->limit) ||
533 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
534 codel_time_to_us(q->cparams.interval)) ||
535 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
536 q->cparams.ecn) ||
537 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
538 q->quantum) ||
Eric Dumazet9d185622016-05-01 16:47:26 -0700539 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
540 q->drop_batch_size) ||
Eric Dumazet95b58432016-05-06 08:55:12 -0700541 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
542 q->memory_limit) ||
Eric Dumazet4b549a22012-05-11 09:30:50 +0000543 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
544 q->flows_cnt))
545 goto nla_put_failure;
546
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700547 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
548 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
549 codel_time_to_us(q->cparams.ce_threshold)))
550 goto nla_put_failure;
551
Yang Yingliangd59b7d82014-03-12 10:20:32 +0800552 return nla_nest_end(skb, opts);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000553
554nla_put_failure:
555 return -1;
556}
557
558static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
559{
560 struct fq_codel_sched_data *q = qdisc_priv(sch);
561 struct tc_fq_codel_xstats st = {
562 .type = TCA_FQ_CODEL_XSTATS_QDISC,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000563 };
564 struct list_head *pos;
565
Sasha Levin669d67b2012-05-14 11:57:06 +0000566 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
567 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
568 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
569 st.qdisc_stats.new_flow_count = q->new_flow_count;
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700570 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
Eric Dumazet95b58432016-05-06 08:55:12 -0700571 st.qdisc_stats.memory_usage = q->memory_usage;
572 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
Sasha Levin669d67b2012-05-14 11:57:06 +0000573
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700574 sch_tree_lock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000575 list_for_each(pos, &q->new_flows)
576 st.qdisc_stats.new_flows_len++;
577
578 list_for_each(pos, &q->old_flows)
579 st.qdisc_stats.old_flows_len++;
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700580 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000581
582 return gnet_stats_copy_app(d, &st, sizeof(st));
583}
584
585static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
586{
587 return NULL;
588}
589
WANG Cong143976c2017-08-24 16:51:29 -0700590static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000591{
592 return 0;
593}
594
595static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
596 u32 classid)
597{
Eric Dumazet4b549a22012-05-11 09:30:50 +0000598 return 0;
599}
600
WANG Cong143976c2017-08-24 16:51:29 -0700601static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000602{
603}
604
Alexander Aringcbaacc42017-12-20 12:35:16 -0500605static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
606 struct netlink_ext_ack *extack)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000607{
608 struct fq_codel_sched_data *q = qdisc_priv(sch);
609
610 if (cl)
611 return NULL;
Jiri Pirko6529eab2017-05-17 11:07:55 +0200612 return q->block;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000613}
614
615static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
616 struct sk_buff *skb, struct tcmsg *tcm)
617{
618 tcm->tcm_handle |= TC_H_MIN(cl);
619 return 0;
620}
621
622static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
623 struct gnet_dump *d)
624{
625 struct fq_codel_sched_data *q = qdisc_priv(sch);
626 u32 idx = cl - 1;
627 struct gnet_stats_queue qs = { 0 };
628 struct tc_fq_codel_xstats xstats;
629
630 if (idx < q->flows_cnt) {
631 const struct fq_codel_flow *flow = &q->flows[idx];
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700632 const struct sk_buff *skb;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000633
634 memset(&xstats, 0, sizeof(xstats));
635 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
636 xstats.class_stats.deficit = flow->deficit;
637 xstats.class_stats.ldelay =
638 codel_time_to_us(flow->cvars.ldelay);
639 xstats.class_stats.count = flow->cvars.count;
640 xstats.class_stats.lastcount = flow->cvars.lastcount;
641 xstats.class_stats.dropping = flow->cvars.dropping;
642 if (flow->cvars.dropping) {
643 codel_tdiff_t delta = flow->cvars.drop_next -
644 codel_get_time();
645
646 xstats.class_stats.drop_next = (delta >= 0) ?
647 codel_time_to_us(delta) :
648 -codel_time_to_us(-delta);
649 }
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700650 if (flow->head) {
651 sch_tree_lock(sch);
652 skb = flow->head;
653 while (skb) {
654 qs.qlen++;
655 skb = skb->next;
656 }
657 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000658 }
659 qs.backlog = q->backlogs[idx];
Dave Taht77ddaff2019-08-03 16:37:29 -0700660 qs.drops = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000661 }
Eric Dumazetaafddbf2016-06-06 09:12:39 -0700662 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000663 return -1;
664 if (idx < q->flows_cnt)
665 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
666 return 0;
667}
668
669static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
670{
671 struct fq_codel_sched_data *q = qdisc_priv(sch);
672 unsigned int i;
673
674 if (arg->stop)
675 return;
676
677 for (i = 0; i < q->flows_cnt; i++) {
678 if (list_empty(&q->flows[i].flowchain) ||
679 arg->count < arg->skip) {
680 arg->count++;
681 continue;
682 }
683 if (arg->fn(sch, i + 1, arg) < 0) {
684 arg->stop = 1;
685 break;
686 }
687 arg->count++;
688 }
689}
690
691static const struct Qdisc_class_ops fq_codel_class_ops = {
692 .leaf = fq_codel_leaf,
WANG Cong143976c2017-08-24 16:51:29 -0700693 .find = fq_codel_find,
Jiri Pirko6529eab2017-05-17 11:07:55 +0200694 .tcf_block = fq_codel_tcf_block,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000695 .bind_tcf = fq_codel_bind,
WANG Cong143976c2017-08-24 16:51:29 -0700696 .unbind_tcf = fq_codel_unbind,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000697 .dump = fq_codel_dump_class,
698 .dump_stats = fq_codel_dump_class_stats,
699 .walk = fq_codel_walk,
700};
701
702static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
703 .cl_ops = &fq_codel_class_ops,
704 .id = "fq_codel",
705 .priv_size = sizeof(struct fq_codel_sched_data),
706 .enqueue = fq_codel_enqueue,
707 .dequeue = fq_codel_dequeue,
708 .peek = qdisc_peek_dequeued,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000709 .init = fq_codel_init,
710 .reset = fq_codel_reset,
711 .destroy = fq_codel_destroy,
712 .change = fq_codel_change,
713 .dump = fq_codel_dump,
714 .dump_stats = fq_codel_dump_stats,
715 .owner = THIS_MODULE,
716};
717
718static int __init fq_codel_module_init(void)
719{
720 return register_qdisc(&fq_codel_qdisc_ops);
721}
722
723static void __exit fq_codel_module_exit(void)
724{
725 unregister_qdisc(&fq_codel_qdisc_ops);
726}
727
728module_init(fq_codel_module_init)
729module_exit(fq_codel_module_exit)
730MODULE_AUTHOR("Eric Dumazet");
731MODULE_LICENSE("GPL");
Rob Gill67c20de2020-06-20 02:08:25 +0000732MODULE_DESCRIPTION("Fair Queue CoDel discipline");