blob: 6883a89715625fc7874780ef962e51dc7773a35b [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
WANG Congc0afd9c2015-07-14 11:21:58 -0700187static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch)
188{
189 unsigned int prev_backlog;
190
191 prev_backlog = sch->qstats.backlog;
Eric Dumazet9d185622016-05-01 16:47:26 -0700192 fq_codel_drop(sch, 1U);
WANG Congc0afd9c2015-07-14 11:21:58 -0700193 return prev_backlog - sch->qstats.backlog;
194}
195
Eric Dumazet4b549a22012-05-11 09:30:50 +0000196static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
197{
198 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet9d185622016-05-01 16:47:26 -0700199 unsigned int idx, prev_backlog, prev_qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000200 struct fq_codel_flow *flow;
201 int uninitialized_var(ret);
Eric Dumazet95b58432016-05-06 08:55:12 -0700202 bool memory_limited;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000203
204 idx = fq_codel_classify(skb, sch, &ret);
205 if (idx == 0) {
206 if (ret & __NET_XMIT_BYPASS)
John Fastabend25331d62014-09-28 11:53:29 -0700207 qdisc_qstats_drop(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000208 kfree_skb(skb);
209 return ret;
210 }
211 idx--;
212
213 codel_set_enqueue_time(skb);
214 flow = &q->flows[idx];
215 flow_queue_add(flow, skb);
216 q->backlogs[idx] += qdisc_pkt_len(skb);
John Fastabend25331d62014-09-28 11:53:29 -0700217 qdisc_qstats_backlog_inc(sch, skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000218
219 if (list_empty(&flow->flowchain)) {
220 list_add_tail(&flow->flowchain, &q->new_flows);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000221 q->new_flow_count++;
222 flow->deficit = q->quantum;
223 flow->dropped = 0;
224 }
Eric Dumazet95b58432016-05-06 08:55:12 -0700225 q->memory_usage += skb->truesize;
226 memory_limited = q->memory_usage > q->memory_limit;
227 if (++sch->q.qlen <= sch->limit && !memory_limited)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000228 return NET_XMIT_SUCCESS;
229
WANG Cong2ccccf52016-02-25 14:55:01 -0800230 prev_backlog = sch->qstats.backlog;
Eric Dumazet9d185622016-05-01 16:47:26 -0700231 prev_qlen = sch->q.qlen;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000232
Eric Dumazet9d185622016-05-01 16:47:26 -0700233 /* fq_codel_drop() is quite expensive, as it performs a linear search
234 * in q->backlogs[] to find a fat flow.
235 * So instead of dropping a single packet, drop half of its backlog
236 * with a 64 packets limit to not add a too big cpu spike here.
237 */
238 ret = fq_codel_drop(sch, q->drop_batch_size);
239
240 q->drop_overlimit += prev_qlen - sch->q.qlen;
Eric Dumazet95b58432016-05-06 08:55:12 -0700241 if (memory_limited)
242 q->drop_overmemory += prev_qlen - sch->q.qlen;
Eric Dumazet9d185622016-05-01 16:47:26 -0700243 /* As we dropped packet(s), better let upper stack know this */
244 qdisc_tree_reduce_backlog(sch, prev_qlen - sch->q.qlen,
245 prev_backlog - sch->qstats.backlog);
246
247 return ret == idx ? NET_XMIT_CN : NET_XMIT_SUCCESS;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000248}
249
250/* This is the specific function called from codel_dequeue()
251 * to dequeue a packet from queue. Note: backlog is handled in
252 * codel, we dont need to reduce it here.
253 */
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200254static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000255{
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200256 struct Qdisc *sch = ctx;
Eric Dumazet865ec552012-05-16 04:39:09 +0000257 struct fq_codel_sched_data *q = qdisc_priv(sch);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000258 struct fq_codel_flow *flow;
259 struct sk_buff *skb = NULL;
260
261 flow = container_of(vars, struct fq_codel_flow, cvars);
262 if (flow->head) {
263 skb = dequeue_head(flow);
Eric Dumazet865ec552012-05-16 04:39:09 +0000264 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
Eric Dumazet77f57762016-05-15 18:16:38 -0700265 q->memory_usage -= skb->truesize;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000266 sch->q.qlen--;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200267 sch->qstats.backlog -= qdisc_pkt_len(skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000268 }
269 return skb;
270}
271
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200272static void drop_func(struct sk_buff *skb, void *ctx)
273{
274 struct Qdisc *sch = ctx;
275
276 qdisc_drop(skb, sch);
277}
278
Eric Dumazet4b549a22012-05-11 09:30:50 +0000279static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
280{
281 struct fq_codel_sched_data *q = qdisc_priv(sch);
282 struct sk_buff *skb;
283 struct fq_codel_flow *flow;
284 struct list_head *head;
285 u32 prev_drop_count, prev_ecn_mark;
WANG Cong2ccccf52016-02-25 14:55:01 -0800286 unsigned int prev_backlog;
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
303 prev_drop_count = q->cstats.drop_count;
304 prev_ecn_mark = q->cstats.ecn_mark;
WANG Cong2ccccf52016-02-25 14:55:01 -0800305 prev_backlog = sch->qstats.backlog;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000306
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200307 skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
308 &flow->cvars, &q->cstats, qdisc_pkt_len,
309 codel_get_enqueue_time, drop_func, dequeue_func);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000310
311 flow->dropped += q->cstats.drop_count - prev_drop_count;
312 flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
313
314 if (!skb) {
315 /* force a pass through old_flows to prevent starvation */
316 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
317 list_move_tail(&flow->flowchain, &q->old_flows);
318 else
319 list_del_init(&flow->flowchain);
320 goto begin;
321 }
322 qdisc_bstats_update(sch, skb);
323 flow->deficit -= qdisc_pkt_len(skb);
WANG Cong2ccccf52016-02-25 14:55:01 -0800324 /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000325 * or HTB crashes. Defer it for next round.
326 */
327 if (q->cstats.drop_count && sch->q.qlen) {
WANG Cong2ccccf52016-02-25 14:55:01 -0800328 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
329 q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000330 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800331 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000332 }
333 return skb;
334}
335
336static void fq_codel_reset(struct Qdisc *sch)
337{
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700338 struct fq_codel_sched_data *q = qdisc_priv(sch);
339 int i;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000340
Eric Dumazet3d0e0af2015-07-31 17:53:39 -0700341 INIT_LIST_HEAD(&q->new_flows);
342 INIT_LIST_HEAD(&q->old_flows);
343 for (i = 0; i < q->flows_cnt; i++) {
344 struct fq_codel_flow *flow = q->flows + i;
345
346 while (flow->head) {
347 struct sk_buff *skb = dequeue_head(flow);
348
349 qdisc_qstats_backlog_dec(sch, skb);
350 kfree_skb(skb);
351 }
352
353 INIT_LIST_HEAD(&flow->flowchain);
354 codel_vars_init(&flow->cvars);
355 }
356 memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
357 sch->q.qlen = 0;
Eric Dumazet77f57762016-05-15 18:16:38 -0700358 q->memory_usage = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000359}
360
361static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
362 [TCA_FQ_CODEL_TARGET] = { .type = NLA_U32 },
363 [TCA_FQ_CODEL_LIMIT] = { .type = NLA_U32 },
364 [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
365 [TCA_FQ_CODEL_ECN] = { .type = NLA_U32 },
366 [TCA_FQ_CODEL_FLOWS] = { .type = NLA_U32 },
367 [TCA_FQ_CODEL_QUANTUM] = { .type = NLA_U32 },
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700368 [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
Eric Dumazet9d185622016-05-01 16:47:26 -0700369 [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
Eric Dumazet95b58432016-05-06 08:55:12 -0700370 [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
Eric Dumazet4b549a22012-05-11 09:30:50 +0000371};
372
373static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
374{
375 struct fq_codel_sched_data *q = qdisc_priv(sch);
376 struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
377 int err;
378
379 if (!opt)
380 return -EINVAL;
381
382 err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
383 if (err < 0)
384 return err;
385 if (tb[TCA_FQ_CODEL_FLOWS]) {
386 if (q->flows)
387 return -EINVAL;
388 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
389 if (!q->flows_cnt ||
390 q->flows_cnt > 65536)
391 return -EINVAL;
392 }
393 sch_tree_lock(sch);
394
395 if (tb[TCA_FQ_CODEL_TARGET]) {
396 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
397
398 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
399 }
400
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700401 if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
402 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
403
404 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
405 }
406
Eric Dumazet4b549a22012-05-11 09:30:50 +0000407 if (tb[TCA_FQ_CODEL_INTERVAL]) {
408 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
409
410 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
411 }
412
413 if (tb[TCA_FQ_CODEL_LIMIT])
414 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
415
416 if (tb[TCA_FQ_CODEL_ECN])
417 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
418
419 if (tb[TCA_FQ_CODEL_QUANTUM])
420 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
421
Eric Dumazet9d185622016-05-01 16:47:26 -0700422 if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
423 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
424
Eric Dumazet95b58432016-05-06 08:55:12 -0700425 if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
426 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
427
428 while (sch->q.qlen > sch->limit ||
429 q->memory_usage > q->memory_limit) {
Eric Dumazet4b549a22012-05-11 09:30:50 +0000430 struct sk_buff *skb = fq_codel_dequeue(sch);
431
WANG Cong2ccccf52016-02-25 14:55:01 -0800432 q->cstats.drop_len += qdisc_pkt_len(skb);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000433 kfree_skb(skb);
434 q->cstats.drop_count++;
435 }
WANG Cong2ccccf52016-02-25 14:55:01 -0800436 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000437 q->cstats.drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800438 q->cstats.drop_len = 0;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000439
440 sch_tree_unlock(sch);
441 return 0;
442}
443
444static void *fq_codel_zalloc(size_t sz)
445{
446 void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
447
448 if (!ptr)
449 ptr = vzalloc(sz);
450 return ptr;
451}
452
453static void fq_codel_free(void *addr)
454{
WANG Cong4cb28972014-06-02 15:55:22 -0700455 kvfree(addr);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000456}
457
458static void fq_codel_destroy(struct Qdisc *sch)
459{
460 struct fq_codel_sched_data *q = qdisc_priv(sch);
461
462 tcf_destroy_chain(&q->filter_list);
463 fq_codel_free(q->backlogs);
464 fq_codel_free(q->flows);
465}
466
467static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
468{
469 struct fq_codel_sched_data *q = qdisc_priv(sch);
470 int i;
471
472 sch->limit = 10*1024;
473 q->flows_cnt = 1024;
Eric Dumazet95b58432016-05-06 08:55:12 -0700474 q->memory_limit = 32 << 20; /* 32 MBytes */
Eric Dumazet9d185622016-05-01 16:47:26 -0700475 q->drop_batch_size = 64;
Eric Dumazet4b549a22012-05-11 09:30:50 +0000476 q->quantum = psched_mtu(qdisc_dev(sch));
Aruna-Hewapathirane63862b52014-01-11 07:15:59 -0500477 q->perturbation = prandom_u32();
Eric Dumazet4b549a22012-05-11 09:30:50 +0000478 INIT_LIST_HEAD(&q->new_flows);
479 INIT_LIST_HEAD(&q->old_flows);
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200480 codel_params_init(&q->cparams);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000481 codel_stats_init(&q->cstats);
482 q->cparams.ecn = true;
Michal Kazior79bdc4c2016-04-22 14:15:58 +0200483 q->cparams.mtu = psched_mtu(qdisc_dev(sch));
Eric Dumazet4b549a22012-05-11 09:30:50 +0000484
485 if (opt) {
486 int err = fq_codel_change(sch, opt);
487 if (err)
488 return err;
489 }
490
491 if (!q->flows) {
492 q->flows = fq_codel_zalloc(q->flows_cnt *
493 sizeof(struct fq_codel_flow));
494 if (!q->flows)
495 return -ENOMEM;
496 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
497 if (!q->backlogs) {
498 fq_codel_free(q->flows);
499 return -ENOMEM;
500 }
501 for (i = 0; i < q->flows_cnt; i++) {
502 struct fq_codel_flow *flow = q->flows + i;
503
504 INIT_LIST_HEAD(&flow->flowchain);
Eric Dumazetb3791352012-09-01 03:19:57 +0000505 codel_vars_init(&flow->cvars);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000506 }
507 }
508 if (sch->limit >= 1)
509 sch->flags |= TCQ_F_CAN_BYPASS;
510 else
511 sch->flags &= ~TCQ_F_CAN_BYPASS;
512 return 0;
513}
514
515static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
516{
517 struct fq_codel_sched_data *q = qdisc_priv(sch);
518 struct nlattr *opts;
519
520 opts = nla_nest_start(skb, TCA_OPTIONS);
521 if (opts == NULL)
522 goto nla_put_failure;
523
524 if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
525 codel_time_to_us(q->cparams.target)) ||
526 nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
527 sch->limit) ||
528 nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
529 codel_time_to_us(q->cparams.interval)) ||
530 nla_put_u32(skb, TCA_FQ_CODEL_ECN,
531 q->cparams.ecn) ||
532 nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
533 q->quantum) ||
Eric Dumazet9d185622016-05-01 16:47:26 -0700534 nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
535 q->drop_batch_size) ||
Eric Dumazet95b58432016-05-06 08:55:12 -0700536 nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
537 q->memory_limit) ||
Eric Dumazet4b549a22012-05-11 09:30:50 +0000538 nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
539 q->flows_cnt))
540 goto nla_put_failure;
541
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700542 if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
543 nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
544 codel_time_to_us(q->cparams.ce_threshold)))
545 goto nla_put_failure;
546
Yang Yingliangd59b7d82014-03-12 10:20:32 +0800547 return nla_nest_end(skb, opts);
Eric Dumazet4b549a22012-05-11 09:30:50 +0000548
549nla_put_failure:
550 return -1;
551}
552
553static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
554{
555 struct fq_codel_sched_data *q = qdisc_priv(sch);
556 struct tc_fq_codel_xstats st = {
557 .type = TCA_FQ_CODEL_XSTATS_QDISC,
Eric Dumazet4b549a22012-05-11 09:30:50 +0000558 };
559 struct list_head *pos;
560
Sasha Levin669d67b2012-05-14 11:57:06 +0000561 st.qdisc_stats.maxpacket = q->cstats.maxpacket;
562 st.qdisc_stats.drop_overlimit = q->drop_overlimit;
563 st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
564 st.qdisc_stats.new_flow_count = q->new_flow_count;
Eric Dumazet80ba92f2015-05-08 15:05:12 -0700565 st.qdisc_stats.ce_mark = q->cstats.ce_mark;
Eric Dumazet95b58432016-05-06 08:55:12 -0700566 st.qdisc_stats.memory_usage = q->memory_usage;
567 st.qdisc_stats.drop_overmemory = q->drop_overmemory;
Sasha Levin669d67b2012-05-14 11:57:06 +0000568
Eric Dumazet4b549a22012-05-11 09:30:50 +0000569 list_for_each(pos, &q->new_flows)
570 st.qdisc_stats.new_flows_len++;
571
572 list_for_each(pos, &q->old_flows)
573 st.qdisc_stats.old_flows_len++;
574
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
583static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
584{
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
596static void fq_codel_put(struct Qdisc *q, unsigned long cl)
597{
598}
599
John Fastabend25d8c0d2014-09-12 20:05:27 -0700600static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
601 unsigned long cl)
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;
607 return &q->filter_list;
608}
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];
627 const struct sk_buff *skb = flow->head;
628
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 }
645 while (skb) {
646 qs.qlen++;
647 skb = skb->next;
648 }
649 qs.backlog = q->backlogs[idx];
650 qs.drops = flow->dropped;
651 }
John Fastabendb0ab6f92014-09-28 11:54:24 -0700652 if (gnet_stats_copy_queue(d, NULL, &qs, 0) < 0)
Eric Dumazet4b549a22012-05-11 09:30:50 +0000653 return -1;
654 if (idx < q->flows_cnt)
655 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
656 return 0;
657}
658
659static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
660{
661 struct fq_codel_sched_data *q = qdisc_priv(sch);
662 unsigned int i;
663
664 if (arg->stop)
665 return;
666
667 for (i = 0; i < q->flows_cnt; i++) {
668 if (list_empty(&q->flows[i].flowchain) ||
669 arg->count < arg->skip) {
670 arg->count++;
671 continue;
672 }
673 if (arg->fn(sch, i + 1, arg) < 0) {
674 arg->stop = 1;
675 break;
676 }
677 arg->count++;
678 }
679}
680
681static const struct Qdisc_class_ops fq_codel_class_ops = {
682 .leaf = fq_codel_leaf,
683 .get = fq_codel_get,
684 .put = fq_codel_put,
685 .tcf_chain = fq_codel_find_tcf,
686 .bind_tcf = fq_codel_bind,
687 .unbind_tcf = fq_codel_put,
688 .dump = fq_codel_dump_class,
689 .dump_stats = fq_codel_dump_class_stats,
690 .walk = fq_codel_walk,
691};
692
693static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
694 .cl_ops = &fq_codel_class_ops,
695 .id = "fq_codel",
696 .priv_size = sizeof(struct fq_codel_sched_data),
697 .enqueue = fq_codel_enqueue,
698 .dequeue = fq_codel_dequeue,
699 .peek = qdisc_peek_dequeued,
WANG Congc0afd9c2015-07-14 11:21:58 -0700700 .drop = fq_codel_qdisc_drop,
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");