blob: 2dc0a849515a7b54e496f5eccb224e6847d97cbd [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>
Eric Dumazet4b549a22012-05-11 09:30:50 +000026#include <net/codel.h>
Michal Kaziord068ca22016-04-22 14:15:59 +020027#include <net/codel_impl.h>
28#include <net/codel_qdisc.h>
Eric Dumazet4b549a22012-05-11 09:30:50 +000029
30/* Fair Queue CoDel.
31 *
32 * Principles :
33 * Packets are classified (internal classifier or external) on flows.
34 * This is a Stochastic model (as we use a hash, several flows
35 * might be hashed on same slot)
36 * Each flow has a CoDel managed queue.
37 * Flows are linked onto two (Round Robin) lists,
38 * so that new flows have priority on old ones.
39 *
40 * For a given flow, packets are not reordered (CoDel uses a FIFO)
41 * head drops only.
42 * ECN capability is on by default.
43 * Low memory footprint (64 bytes per flow)
44 */
45
46struct fq_codel_flow {
47 struct sk_buff *head;
48 struct sk_buff *tail;
49 struct list_head flowchain;
50 int deficit;
51 u32 dropped; /* number of drops (or ECN marks) on this flow */
52 struct codel_vars cvars;
53}; /* please try to keep this structure <= 64 bytes */
54
55struct fq_codel_sched_data {
John Fastabend25d8c0d2014-09-12 20:05:27 -070056 struct tcf_proto __rcu *filter_list; /* optional external classifier */
Eric Dumazet4b549a22012-05-11 09:30:50 +000057 struct fq_codel_flow *flows; /* Flows table [flows_cnt] */
58 u32 *backlogs; /* backlog table [flows_cnt] */
59 u32 flows_cnt; /* number of flows */
60 u32 perturbation; /* hash perturbation */
61 u32 quantum; /* psched_mtu(qdisc_dev(sch)); */
Eric Dumazet9d185622016-05-01 16:47:26 -070062 u32 drop_batch_size;
Eric Dumazet95b58432016-05-06 08:55:12 -070063 u32 memory_limit;
Eric Dumazet4b549a22012-05-11 09:30:50 +000064 struct codel_params cparams;
65 struct codel_stats cstats;
Eric Dumazet95b58432016-05-06 08:55:12 -070066 u32 memory_usage;
67 u32 drop_overmemory;
Eric Dumazet4b549a22012-05-11 09:30:50 +000068 u32 drop_overlimit;
69 u32 new_flow_count;
70
71 struct list_head new_flows; /* list of new flows */
72 struct list_head old_flows; /* list of old flows */
73};
74
75static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
Tom Herbert342db222015-05-01 11:30:13 -070076 struct sk_buff *skb)
Eric Dumazet4b549a22012-05-11 09:30:50 +000077{
Tom Herbert342db222015-05-01 11:30:13 -070078 u32 hash = skb_get_hash_perturb(skb, q->perturbation);
Daniel Borkmann8fc54f62014-08-23 20:58:54 +020079
80 return reciprocal_scale(hash, q->flows_cnt);
Eric Dumazet4b549a22012-05-11 09:30:50 +000081}
82
83static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
84 int *qerr)
85{
86 struct fq_codel_sched_data *q = qdisc_priv(sch);
John Fastabend25d8c0d2014-09-12 20:05:27 -070087 struct tcf_proto *filter;
Eric Dumazet4b549a22012-05-11 09:30:50 +000088 struct tcf_result res;
89 int result;
90
91 if (TC_H_MAJ(skb->priority) == sch->handle &&
92 TC_H_MIN(skb->priority) > 0 &&
93 TC_H_MIN(skb->priority) <= q->flows_cnt)
94 return TC_H_MIN(skb->priority);
95
Valdis.Kletnieks@vt.edu69204cf2014-12-09 16:15:50 -050096 filter = rcu_dereference_bh(q->filter_list);
John Fastabend25d8c0d2014-09-12 20:05:27 -070097 if (!filter)
Eric Dumazet4b549a22012-05-11 09:30:50 +000098 return fq_codel_hash(q, skb) + 1;
99
100 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
Daniel Borkmann3b3ae882015-08-26 23:00:06 +0200101 result = tc_classify(skb, filter, &res, false);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000102 if (result >= 0) {
103#ifdef CONFIG_NET_CLS_ACT
104 switch (result) {
105 case TC_ACT_STOLEN:
106 case TC_ACT_QUEUED:
107 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
108 case TC_ACT_SHOT:
109 return 0;
110 }
111#endif
112 if (TC_H_MIN(res.classid) <= q->flows_cnt)
113 return TC_H_MIN(res.classid);
114 }
115 return 0;
116}
117
118/* helper functions : might be changed when/if skb use a standard list_head */
119
120/* remove one skb from head of slot queue */
121static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
122{
123 struct sk_buff *skb = flow->head;
124
125 flow->head = skb->next;
126 skb->next = NULL;
127 return skb;
128}
129
130/* add skb to flow queue (tail add) */
131static inline void flow_queue_add(struct fq_codel_flow *flow,
132 struct sk_buff *skb)
133{
134 if (flow->head == NULL)
135 flow->head = skb;
136 else
137 flow->tail->next = skb;
138 flow->tail = skb;
139 skb->next = NULL;
140}
141
Eric Dumazet9d185622016-05-01 16:47:26 -0700142static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000143{
144 struct fq_codel_sched_data *q = qdisc_priv(sch);
145 struct sk_buff *skb;
146 unsigned int maxbacklog = 0, idx = 0, i, len;
147 struct fq_codel_flow *flow;
Eric Dumazet9d185622016-05-01 16:47:26 -0700148 unsigned int threshold;
Eric Dumazet95b58432016-05-06 08:55:12 -0700149 unsigned int mem = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000150
Eric Dumazet9d185622016-05-01 16:47:26 -0700151 /* Queue is full! Find the fat flow and drop packet(s) from it.
Eric Dumazet4b549a22012-05-11 09:30:50 +0000152 * This might sound expensive, but with 1024 flows, we scan
153 * 4KB of memory, and we dont need to handle a complex tree
154 * in fast path (packet queue/enqueue) with many cache misses.
Eric Dumazet9d185622016-05-01 16:47:26 -0700155 * In stress mode, we'll try to drop 64 packets from the flow,
156 * amortizing this linear lookup to one cache line per drop.
Eric Dumazet4b549a22012-05-11 09:30:50 +0000157 */
158 for (i = 0; i < q->flows_cnt; i++) {
159 if (q->backlogs[i] > maxbacklog) {
160 maxbacklog = q->backlogs[i];
161 idx = i;
162 }
163 }
Eric Dumazet9d185622016-05-01 16:47:26 -0700164
165 /* Our goal is to drop half of this fat flow backlog */
166 threshold = maxbacklog >> 1;
167
Eric Dumazet4b549a22012-05-11 09:30:50 +0000168 flow = &q->flows[idx];
Eric Dumazet9d185622016-05-01 16:47:26 -0700169 len = 0;
170 i = 0;
171 do {
172 skb = dequeue_head(flow);
173 len += qdisc_pkt_len(skb);
Eric Dumazet95b58432016-05-06 08:55:12 -0700174 mem += skb->truesize;
Eric Dumazet9d185622016-05-01 16:47:26 -0700175 kfree_skb(skb);
176 } while (++i < max_packets && len < threshold);
177
178 flow->dropped += i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000179 q->backlogs[idx] -= len;
Eric Dumazet95b58432016-05-06 08:55:12 -0700180 q->memory_usage -= mem;
Eric Dumazet9d185622016-05-01 16:47:26 -0700181 sch->qstats.drops += i;
182 sch->qstats.backlog -= len;
183 sch->q.qlen -= i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000184 return idx;
185}
186
187static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
188{
189 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet9d185622016-05-01 16:47:26 -0700190 unsigned int idx, prev_backlog, prev_qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000191 struct fq_codel_flow *flow;
192 int uninitialized_var(ret);
Eric Dumazet80e509d2016-06-04 12:55:13 -0700193 unsigned int pkt_len;
Eric Dumazet95b58432016-05-06 08:55:12 -0700194 bool memory_limited;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000195
196 idx = fq_codel_classify(skb, sch, &ret);
197 if (idx == 0) {
198 if (ret & __NET_XMIT_BYPASS)
John Fastabend25331d62014-09-28 11:53:29 -0700199 qdisc_qstats_drop(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000200 kfree_skb(skb);
201 return ret;
202 }
203 idx--;
204
205 codel_set_enqueue_time(skb);
206 flow = &q->flows[idx];
207 flow_queue_add(flow, skb);
208 q->backlogs[idx] += qdisc_pkt_len(skb);
John Fastabend25331d62014-09-28 11:53:29 -0700209 qdisc_qstats_backlog_inc(sch, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000210
211 if (list_empty(&flow->flowchain)) {
212 list_add_tail(&flow->flowchain, &q->new_flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000213 q->new_flow_count++;
214 flow->deficit = q->quantum;
215 flow->dropped = 0;
216 }
Eric Dumazet95b58432016-05-06 08:55:12 -0700217 q->memory_usage += skb->truesize;
218 memory_limited = q->memory_usage > q->memory_limit;
219 if (++sch->q.qlen <= sch->limit && !memory_limited)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000220 return NET_XMIT_SUCCESS;
221
WANG Cong2ccccf52016-02-25 14:55:01 -0800222 prev_backlog = sch->qstats.backlog;
Eric Dumazet9d185622016-05-01 16:47:26 -0700223 prev_qlen = sch->q.qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000224
Eric Dumazet80e509d2016-06-04 12:55:13 -0700225 /* save this packet length as it might be dropped by fq_codel_drop() */
226 pkt_len = qdisc_pkt_len(skb);
Eric Dumazet9d185622016-05-01 16:47:26 -0700227 /* 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);
233
Eric Dumazet80e509d2016-06-04 12:55:13 -0700234 prev_qlen -= sch->q.qlen;
235 prev_backlog -= sch->qstats.backlog;
236 q->drop_overlimit += prev_qlen;
Eric Dumazet95b58432016-05-06 08:55:12 -0700237 if (memory_limited)
Eric Dumazet80e509d2016-06-04 12:55:13 -0700238 q->drop_overmemory += prev_qlen;
Eric Dumazet9d185622016-05-01 16:47:26 -0700239
Eric Dumazet80e509d2016-06-04 12:55:13 -0700240 /* 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;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000251}
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 */
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200257static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000258{
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200259 struct Qdisc *sch = ctx;
Eric Dumazet865ec552012-05-16 04:39:09 +0000260 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000261 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);
Eric Dumazet865ec552012-05-16 04:39:09 +0000267 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
Eric Dumazet77f57762016-05-15 18:16:38 -0700268 q->memory_usage -= skb->truesize;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000269 sch->q.qlen--;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200270 sch->qstats.backlog -= qdisc_pkt_len(skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000271 }
272 return skb;
273}
274
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200275static void drop_func(struct sk_buff *skb, void *ctx)
276{
277 struct Qdisc *sch = ctx;
278
279 qdisc_drop(skb, sch);
280}
281
Eric Dumazet4b549a22012-05-11 09:30:50 +0000282static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
283{
284 struct fq_codel_sched_data *q = qdisc_priv(sch);
285 struct sk_buff *skb;
286 struct fq_codel_flow *flow;
287 struct list_head *head;
288 u32 prev_drop_count, prev_ecn_mark;
WANG Cong2ccccf52016-02-25 14:55:01 -0800289 unsigned int prev_backlog;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000290
291begin:
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 flow->deficit += q->quantum;
302 list_move_tail(&flow->flowchain, &q->old_flows);
303 goto begin;
304 }
305
306 prev_drop_count = q->cstats.drop_count;
307 prev_ecn_mark = q->cstats.ecn_mark;
WANG Cong2ccccf52016-02-25 14:55:01 -0800308 prev_backlog = sch->qstats.backlog;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000309
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200310 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
311 &flow->cvars, &q->cstats, qdisc_pkt_len,
312 codel_get_enqueue_time, drop_func, dequeue_func);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000313
314 flow->dropped += q->cstats.drop_count - prev_drop_count;
315 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
316
317 if (!skb) {
318 /* force a pass through old_flows to prevent starvation */
319 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
320 list_move_tail(&flow->flowchain, &q->old_flows);
321 else
322 list_del_init(&flow->flowchain);
323 goto begin;
324 }
325 qdisc_bstats_update(sch, skb);
326 flow->deficit -= qdisc_pkt_len(skb);
WANG Cong2ccccf52016-02-25 14:55:01 -0800327 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000328 * or HTB crashes. Defer it for next round.
329 */
330 if (q->cstats.drop_count && sch->q.qlen) {
WANG Cong2ccccf52016-02-25 14:55:01 -0800331 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
332 q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000333 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800334 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000335 }
336 return skb;
337}
338
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700339static void fq_codel_flow_purge(struct fq_codel_flow *flow)
340{
341 rtnl_kfree_skbs(flow->head, flow->tail);
342 flow->head = NULL;
343}
344
Eric Dumazet4b549a22012-05-11 09:30:50 +0000345static void fq_codel_reset(struct Qdisc *sch)
346{
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700347 struct fq_codel_sched_data *q = qdisc_priv(sch);
348 int i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000349
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700350 INIT_LIST_HEAD(&q->new_flows);
351 INIT_LIST_HEAD(&q->old_flows);
352 for (i = 0; i < q->flows_cnt; i++) {
353 struct fq_codel_flow *flow = q->flows + i;
354
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700355 fq_codel_flow_purge(flow);
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700356 INIT_LIST_HEAD(&flow->flowchain);
357 codel_vars_init(&flow->cvars);
358 }
359 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
360 sch->q.qlen = 0;
Eric Dumazetece5d4c2016-06-13 20:21:54 -0700361 sch->qstats.backlog = 0;
Eric Dumazet77f57762016-05-15 18:16:38 -0700362 q->memory_usage = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000363}
364
365static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
366 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
367 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
368 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
369 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
370 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
371 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700372 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
Eric Dumazet9d185622016-05-01 16:47:26 -0700373 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
Eric Dumazet95b58432016-05-06 08:55:12 -0700374 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
Eric Dumazet4b549a22012-05-11 09:30:50 +0000375};
376
377static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
378{
379 struct fq_codel_sched_data *q = qdisc_priv(sch);
380 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
381 int err;
382
383 if (!opt)
384 return -EINVAL;
385
386 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
387 if (err < 0)
388 return err;
389 if (tb[TCA_FQ_CODEL_FLOWS]) {
390 if (q->flows)
391 return -EINVAL;
392 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
393 if (!q->flows_cnt ||
394 q->flows_cnt > 65536)
395 return -EINVAL;
396 }
397 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
423 if (tb[TCA_FQ_CODEL_QUANTUM])
424 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
425
Eric Dumazet9d185622016-05-01 16:47:26 -0700426 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
427 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
428
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
448static void *fq_codel_zalloc(size_t sz)
449{
450 void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
451
452 if (!ptr)
453 ptr = vzalloc(sz);
454 return ptr;
455}
456
457static void fq_codel_free(void *addr)
458{
WANG Cong4cb28972014-06-02 15:55:22 -0700459 kvfree(addr);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000460}
461
462static void fq_codel_destroy(struct Qdisc *sch)
463{
464 struct fq_codel_sched_data *q = qdisc_priv(sch);
465
466 tcf_destroy_chain(&q->filter_list);
467 fq_codel_free(q->backlogs);
468 fq_codel_free(q->flows);
469}
470
471static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
472{
473 struct fq_codel_sched_data *q = qdisc_priv(sch);
474 int i;
475
476 sch->limit = 10*1024;
477 q->flows_cnt = 1024;
Eric Dumazet95b58432016-05-06 08:55:12 -0700478 q->memory_limit = 32 << 20; /* 32 MBytes */
Eric Dumazet9d185622016-05-01 16:47:26 -0700479 q->drop_batch_size = 64;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000480 q->quantum = psched_mtu(qdisc_dev(sch));
Aruna-Hewapathirane63862b52014-01-11 07:15:59 -0500481 q->perturbation = prandom_u32();
Eric Dumazet4b549a22012-05-11 09:30:50 +0000482 INIT_LIST_HEAD(&q->new_flows);
483 INIT_LIST_HEAD(&q->old_flows);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200484 codel_params_init(&q->cparams);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000485 codel_stats_init(&q->cstats);
486 q->cparams.ecn = true;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200487 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000488
489 if (opt) {
490 int err = fq_codel_change(sch, opt);
491 if (err)
492 return err;
493 }
494
495 if (!q->flows) {
496 q->flows = fq_codel_zalloc(q->flows_cnt *
497 sizeof(struct fq_codel_flow));
498 if (!q->flows)
499 return -ENOMEM;
500 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
501 if (!q->backlogs) {
502 fq_codel_free(q->flows);
503 return -ENOMEM;
504 }
505 for (i = 0; i < q->flows_cnt; i++) {
506 struct fq_codel_flow *flow = q->flows + i;
507
508 INIT_LIST_HEAD(&flow->flowchain);
Eric Dumazetb3791352012-09-01 03:19:57 +0000509 codel_vars_init(&flow->cvars);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000510 }
511 }
512 if (sch->limit >= 1)
513 sch->flags |= TCQ_F_CAN_BYPASS;
514 else
515 sch->flags &= ~TCQ_F_CAN_BYPASS;
516 return 0;
517}
518
519static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
520{
521 struct fq_codel_sched_data *q = qdisc_priv(sch);
522 struct nlattr *opts;
523
524 opts = nla_nest_start(skb, TCA_OPTIONS);
525 if (opts == NULL)
526 goto nla_put_failure;
527
528 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
529 codel_time_to_us(q->cparams.target)) ||
530 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
531 sch->limit) ||
532 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
533 codel_time_to_us(q->cparams.interval)) ||
534 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
535 q->cparams.ecn) ||
536 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
537 q->quantum) ||
Eric Dumazet9d185622016-05-01 16:47:26 -0700538 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
539 q->drop_batch_size) ||
Eric Dumazet95b58432016-05-06 08:55:12 -0700540 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
541 q->memory_limit) ||
Eric Dumazet4b549a22012-05-11 09:30:50 +0000542 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
543 q->flows_cnt))
544 goto nla_put_failure;
545
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700546 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
547 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
548 codel_time_to_us(q->cparams.ce_threshold)))
549 goto nla_put_failure;
550
Yang Yingliangd59b7d82014-03-12 10:20:32 +0800551 return nla_nest_end(skb, opts);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000552
553nla_put_failure:
554 return -1;
555}
556
557static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
558{
559 struct fq_codel_sched_data *q = qdisc_priv(sch);
560 struct tc_fq_codel_xstats st = {
561 .type = TCA_FQ_CODEL_XSTATS_QDISC,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000562 };
563 struct list_head *pos;
564
Sasha Levin669d67b2012-05-14 11:57:06 +0000565 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
566 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
567 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
568 st.qdisc_stats.new_flow_count = q->new_flow_count;
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700569 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
Eric Dumazet95b58432016-05-06 08:55:12 -0700570 st.qdisc_stats.memory_usage = q->memory_usage;
571 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
Sasha Levin669d67b2012-05-14 11:57:06 +0000572
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700573 sch_tree_lock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000574 list_for_each(pos, &q->new_flows)
575 st.qdisc_stats.new_flows_len++;
576
577 list_for_each(pos, &q->old_flows)
578 st.qdisc_stats.old_flows_len++;
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700579 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000580
581 return gnet_stats_copy_app(d, &st, sizeof(st));
582}
583
584static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
585{
586 return NULL;
587}
588
589static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
590{
591 return 0;
592}
593
594static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
595 u32 classid)
596{
597 /* we cannot bypass queue discipline anymore */
598 sch->flags &= ~TCQ_F_CAN_BYPASS;
599 return 0;
600}
601
602static void fq_codel_put(struct Qdisc *q, unsigned long cl)
603{
604}
605
John Fastabend25d8c0d2014-09-12 20:05:27 -0700606static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
607 unsigned long cl)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000608{
609 struct fq_codel_sched_data *q = qdisc_priv(sch);
610
611 if (cl)
612 return NULL;
613 return &q->filter_list;
614}
615
616static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
617 struct sk_buff *skb, struct tcmsg *tcm)
618{
619 tcm->tcm_handle |= TC_H_MIN(cl);
620 return 0;
621}
622
623static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
624 struct gnet_dump *d)
625{
626 struct fq_codel_sched_data *q = qdisc_priv(sch);
627 u32 idx = cl - 1;
628 struct gnet_stats_queue qs = { 0 };
629 struct tc_fq_codel_xstats xstats;
630
631 if (idx < q->flows_cnt) {
632 const struct fq_codel_flow *flow = &q->flows[idx];
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700633 const struct sk_buff *skb;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000634
635 memset(&xstats, 0, sizeof(xstats));
636 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
637 xstats.class_stats.deficit = flow->deficit;
638 xstats.class_stats.ldelay =
639 codel_time_to_us(flow->cvars.ldelay);
640 xstats.class_stats.count = flow->cvars.count;
641 xstats.class_stats.lastcount = flow->cvars.lastcount;
642 xstats.class_stats.dropping = flow->cvars.dropping;
643 if (flow->cvars.dropping) {
644 codel_tdiff_t delta = flow->cvars.drop_next -
645 codel_get_time();
646
647 xstats.class_stats.drop_next = (delta >= 0) ?
648 codel_time_to_us(delta) :
649 -codel_time_to_us(-delta);
650 }
Eric Dumazetedb09eb2016-06-06 09:37:16 -0700651 if (flow->head) {
652 sch_tree_lock(sch);
653 skb = flow->head;
654 while (skb) {
655 qs.qlen++;
656 skb = skb->next;
657 }
658 sch_tree_unlock(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000659 }
660 qs.backlog = q->backlogs[idx];
661 qs.drops = flow->dropped;
662 }
Eric Dumazetaafddbf2016-06-06 09:12:39 -0700663 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000664 return -1;
665 if (idx < q->flows_cnt)
666 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
667 return 0;
668}
669
670static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
671{
672 struct fq_codel_sched_data *q = qdisc_priv(sch);
673 unsigned int i;
674
675 if (arg->stop)
676 return;
677
678 for (i = 0; i < q->flows_cnt; i++) {
679 if (list_empty(&q->flows[i].flowchain) ||
680 arg->count < arg->skip) {
681 arg->count++;
682 continue;
683 }
684 if (arg->fn(sch, i + 1, arg) < 0) {
685 arg->stop = 1;
686 break;
687 }
688 arg->count++;
689 }
690}
691
692static const struct Qdisc_class_ops fq_codel_class_ops = {
693 .leaf = fq_codel_leaf,
694 .get = fq_codel_get,
695 .put = fq_codel_put,
696 .tcf_chain = fq_codel_find_tcf,
697 .bind_tcf = fq_codel_bind,
698 .unbind_tcf = fq_codel_put,
699 .dump = fq_codel_dump_class,
700 .dump_stats = fq_codel_dump_class_stats,
701 .walk = fq_codel_walk,
702};
703
704static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
705 .cl_ops = &fq_codel_class_ops,
706 .id = "fq_codel",
707 .priv_size = sizeof(struct fq_codel_sched_data),
708 .enqueue = fq_codel_enqueue,
709 .dequeue = fq_codel_dequeue,
710 .peek = qdisc_peek_dequeued,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000711 .init = fq_codel_init,
712 .reset = fq_codel_reset,
713 .destroy = fq_codel_destroy,
714 .change = fq_codel_change,
715 .dump = fq_codel_dump,
716 .dump_stats = fq_codel_dump_stats,
717 .owner = THIS_MODULE,
718};
719
720static int __init fq_codel_module_init(void)
721{
722 return register_qdisc(&fq_codel_qdisc_ops);
723}
724
725static void __exit fq_codel_module_exit(void)
726{
727 unregister_qdisc(&fq_codel_qdisc_ops);
728}
729
730module_init(fq_codel_module_init)
731module_exit(fq_codel_module_exit)
732MODULE_AUTHOR("Eric Dumazet");
733MODULE_LICENSE("GPL");