blob: aeee28081245c9215f10badd611f58ba0124fcd0 [file] [log] [blame]
Eric Dumazet76e3cc12012-05-10 07:51:25 +00001#ifndef __NET_SCHED_CODEL_H
2#define __NET_SCHED_CODEL_H
3
4/*
5 * Codel - The Controlled-Delay Active Queue Management algorithm
6 *
7 * Copyright (C) 2011-2012 Kathleen Nichols <nichols@pollere.com>
8 * Copyright (C) 2011-2012 Van Jacobson <van@pollere.net>
9 * Copyright (C) 2012 Michael D. Taht <dave.taht@bufferbloat.net>
10 * Copyright (C) 2012 Eric Dumazet <edumazet@google.com>
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions, and the following disclaimer,
17 * without modification.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. The names of the authors may not be used to endorse or promote products
22 * derived from this software without specific prior written permission.
23 *
24 * Alternatively, provided that this notice is retained in full, this
25 * software may be distributed under the terms of the GNU General
26 * Public License ("GPL") version 2, in which case the provisions of the
27 * GPL apply INSTEAD OF those given above.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
40 * DAMAGE.
41 *
42 */
43
44#include <linux/types.h>
45#include <linux/ktime.h>
46#include <linux/skbuff.h>
47#include <net/pkt_sched.h>
48#include <net/inet_ecn.h>
49
50/* Controlling Queue Delay (CoDel) algorithm
51 * =========================================
52 * Source : Kathleen Nichols and Van Jacobson
53 * http://queue.acm.org/detail.cfm?id=2209336
54 *
55 * Implemented on linux by Dave Taht and Eric Dumazet
56 */
57
58
59/* CoDel uses a 1024 nsec clock, encoded in u32
60 * This gives a range of 2199 seconds, because of signed compares
61 */
62typedef u32 codel_time_t;
63typedef s32 codel_tdiff_t;
64#define CODEL_SHIFT 10
65#define MS2TIME(a) ((a * NSEC_PER_MSEC) >> CODEL_SHIFT)
66
67static inline codel_time_t codel_get_time(void)
68{
Eric Dumazetd2de8752014-08-22 18:32:09 -070069 u64 ns = ktime_get_ns();
Eric Dumazet76e3cc12012-05-10 07:51:25 +000070
71 return ns >> CODEL_SHIFT;
72}
73
Jesper Dangaard Brouer1ba3aab2013-10-31 22:10:55 +010074/* Dealing with timer wrapping, according to RFC 1982, as desc in wikipedia:
75 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
76 * codel_time_after(a,b) returns true if the time a is after time b.
77 */
78#define codel_time_after(a, b) \
79 (typecheck(codel_time_t, a) && \
80 typecheck(codel_time_t, b) && \
81 ((s32)((a) - (b)) > 0))
82#define codel_time_before(a, b) codel_time_after(b, a)
83
84#define codel_time_after_eq(a, b) \
85 (typecheck(codel_time_t, a) && \
86 typecheck(codel_time_t, b) && \
87 ((s32)((a) - (b)) >= 0))
88#define codel_time_before_eq(a, b) codel_time_after_eq(b, a)
Eric Dumazet76e3cc12012-05-10 07:51:25 +000089
90/* Qdiscs using codel plugin must use codel_skb_cb in their own cb[] */
91struct codel_skb_cb {
92 codel_time_t enqueue_time;
93};
94
95static struct codel_skb_cb *get_codel_cb(const struct sk_buff *skb)
96{
97 qdisc_cb_private_validate(skb, sizeof(struct codel_skb_cb));
98 return (struct codel_skb_cb *)qdisc_skb_cb(skb)->data;
99}
100
101static codel_time_t codel_get_enqueue_time(const struct sk_buff *skb)
102{
103 return get_codel_cb(skb)->enqueue_time;
104}
105
106static void codel_set_enqueue_time(struct sk_buff *skb)
107{
108 get_codel_cb(skb)->enqueue_time = codel_get_time();
109}
110
111static inline u32 codel_time_to_us(codel_time_t val)
112{
113 u64 valns = ((u64)val << CODEL_SHIFT);
114
115 do_div(valns, NSEC_PER_USEC);
116 return (u32)valns;
117}
118
119/**
120 * struct codel_params - contains codel parameters
121 * @target: target queue size (in time units)
122 * @interval: width of moving time window
123 * @ecn: is Explicit Congestion Notification enabled
124 */
125struct codel_params {
126 codel_time_t target;
127 codel_time_t interval;
128 bool ecn;
129};
130
131/**
132 * struct codel_vars - contains codel variables
133 * @count: how many drops we've done since the last time we
134 * entered dropping state
135 * @lastcount: count at entry to dropping state
136 * @dropping: set to true if in dropping state
Eric Dumazet536edd62012-05-12 03:32:13 +0000137 * @rec_inv_sqrt: reciprocal value of sqrt(count) >> 1
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000138 * @first_above_time: when we went (or will go) continuously above target
139 * for interval
140 * @drop_next: time to drop next packet, or when we dropped last
141 * @ldelay: sojourn time of last dequeued packet
142 */
143struct codel_vars {
144 u32 count;
145 u32 lastcount;
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000146 bool dropping;
147 u16 rec_inv_sqrt;
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000148 codel_time_t first_above_time;
149 codel_time_t drop_next;
150 codel_time_t ldelay;
151};
152
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000153#define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */
154/* needed shift to get a Q0.32 number from rec_inv_sqrt */
155#define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS)
156
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000157/**
158 * struct codel_stats - contains codel shared variables and stats
159 * @maxpacket: largest packet we've seen so far
160 * @drop_count: temp count of dropped packets in dequeue()
161 * ecn_mark: number of packets we ECN marked instead of dropping
162 */
163struct codel_stats {
164 u32 maxpacket;
165 u32 drop_count;
166 u32 ecn_mark;
167};
168
169static void codel_params_init(struct codel_params *params)
170{
171 params->interval = MS2TIME(100);
172 params->target = MS2TIME(5);
173 params->ecn = false;
174}
175
176static void codel_vars_init(struct codel_vars *vars)
177{
Eric Dumazet536edd62012-05-12 03:32:13 +0000178 memset(vars, 0, sizeof(*vars));
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000179}
180
181static void codel_stats_init(struct codel_stats *stats)
182{
183 stats->maxpacket = 256;
184}
185
Eric Dumazet536edd62012-05-12 03:32:13 +0000186/*
187 * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
188 * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
189 *
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000190 * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000191 */
Eric Dumazet536edd62012-05-12 03:32:13 +0000192static void codel_Newton_step(struct codel_vars *vars)
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000193{
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000194 u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
195 u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
196 u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000197
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000198 val >>= 2; /* avoid overflow in following multiply */
199 val = (val * invsqrt) >> (32 - 2 + 1);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000200
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000201 vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000202}
203
Eric Dumazet536edd62012-05-12 03:32:13 +0000204/*
205 * CoDel control_law is t + interval/sqrt(count)
206 * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
207 * both sqrt() and divide operation.
208 */
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000209static codel_time_t codel_control_law(codel_time_t t,
210 codel_time_t interval,
Eric Dumazet536edd62012-05-12 03:32:13 +0000211 u32 rec_inv_sqrt)
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000212{
Daniel Borkmann89770b02014-01-22 02:29:40 +0100213 return t + reciprocal_scale(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000214}
215
Eric Dumazet536edd62012-05-12 03:32:13 +0000216static bool codel_should_drop(const struct sk_buff *skb,
Eric Dumazet865ec552012-05-16 04:39:09 +0000217 struct Qdisc *sch,
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000218 struct codel_vars *vars,
219 struct codel_params *params,
220 struct codel_stats *stats,
221 codel_time_t now)
222{
223 bool ok_to_drop;
224
225 if (!skb) {
226 vars->first_above_time = 0;
227 return false;
228 }
229
230 vars->ldelay = now - codel_get_enqueue_time(skb);
Eric Dumazet865ec552012-05-16 04:39:09 +0000231 sch->qstats.backlog -= qdisc_pkt_len(skb);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000232
233 if (unlikely(qdisc_pkt_len(skb) > stats->maxpacket))
234 stats->maxpacket = qdisc_pkt_len(skb);
235
236 if (codel_time_before(vars->ldelay, params->target) ||
Eric Dumazet865ec552012-05-16 04:39:09 +0000237 sch->qstats.backlog <= stats->maxpacket) {
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000238 /* went below - stay below for at least interval */
239 vars->first_above_time = 0;
240 return false;
241 }
242 ok_to_drop = false;
243 if (vars->first_above_time == 0) {
244 /* just went above from below. If we stay above
245 * for at least interval we'll say it's ok to drop
246 */
247 vars->first_above_time = now + params->interval;
248 } else if (codel_time_after(now, vars->first_above_time)) {
249 ok_to_drop = true;
250 }
251 return ok_to_drop;
252}
253
254typedef struct sk_buff * (*codel_skb_dequeue_t)(struct codel_vars *vars,
255 struct Qdisc *sch);
256
257static struct sk_buff *codel_dequeue(struct Qdisc *sch,
258 struct codel_params *params,
259 struct codel_vars *vars,
260 struct codel_stats *stats,
Eric Dumazet865ec552012-05-16 04:39:09 +0000261 codel_skb_dequeue_t dequeue_func)
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000262{
263 struct sk_buff *skb = dequeue_func(vars, sch);
264 codel_time_t now;
265 bool drop;
266
267 if (!skb) {
268 vars->dropping = false;
269 return skb;
270 }
271 now = codel_get_time();
Eric Dumazet865ec552012-05-16 04:39:09 +0000272 drop = codel_should_drop(skb, sch, vars, params, stats, now);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000273 if (vars->dropping) {
274 if (!drop) {
275 /* sojourn time below target - leave dropping state */
276 vars->dropping = false;
277 } else if (codel_time_after_eq(now, vars->drop_next)) {
278 /* It's time for the next drop. Drop the current
279 * packet and dequeue the next. The dequeue might
280 * take us out of dropping state.
281 * If not, schedule the next drop.
282 * A large backlog might result in drop rates so high
283 * that the next drop should happen now,
284 * hence the while loop.
285 */
286 while (vars->dropping &&
287 codel_time_after_eq(now, vars->drop_next)) {
Eric Dumazet536edd62012-05-12 03:32:13 +0000288 vars->count++; /* dont care of possible wrap
289 * since there is no more divide
290 */
291 codel_Newton_step(vars);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000292 if (params->ecn && INET_ECN_set_ce(skb)) {
293 stats->ecn_mark++;
294 vars->drop_next =
295 codel_control_law(vars->drop_next,
296 params->interval,
Eric Dumazet536edd62012-05-12 03:32:13 +0000297 vars->rec_inv_sqrt);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000298 goto end;
299 }
300 qdisc_drop(skb, sch);
301 stats->drop_count++;
302 skb = dequeue_func(vars, sch);
Eric Dumazet865ec552012-05-16 04:39:09 +0000303 if (!codel_should_drop(skb, sch,
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000304 vars, params, stats, now)) {
305 /* leave dropping state */
306 vars->dropping = false;
307 } else {
308 /* and schedule the next drop */
309 vars->drop_next =
310 codel_control_law(vars->drop_next,
311 params->interval,
Eric Dumazet536edd62012-05-12 03:32:13 +0000312 vars->rec_inv_sqrt);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000313 }
314 }
315 }
316 } else if (drop) {
Eric Dumazet2359a472012-07-29 20:52:21 +0000317 u32 delta;
318
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000319 if (params->ecn && INET_ECN_set_ce(skb)) {
320 stats->ecn_mark++;
321 } else {
322 qdisc_drop(skb, sch);
323 stats->drop_count++;
324
325 skb = dequeue_func(vars, sch);
Eric Dumazet865ec552012-05-16 04:39:09 +0000326 drop = codel_should_drop(skb, sch, vars, params,
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000327 stats, now);
328 }
329 vars->dropping = true;
330 /* if min went above target close to when we last went below it
331 * assume that the drop rate that controlled the queue on the
332 * last cycle is a good starting point to control it now.
333 */
Eric Dumazet2359a472012-07-29 20:52:21 +0000334 delta = vars->count - vars->lastcount;
335 if (delta > 1 &&
336 codel_time_before(now - vars->drop_next,
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000337 16 * params->interval)) {
Eric Dumazet2359a472012-07-29 20:52:21 +0000338 vars->count = delta;
Eric Dumazet536edd62012-05-12 03:32:13 +0000339 /* we dont care if rec_inv_sqrt approximation
340 * is not very precise :
341 * Next Newton steps will correct it quadratically.
342 */
343 codel_Newton_step(vars);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000344 } else {
345 vars->count = 1;
Eric Dumazet6ff272c2012-05-12 21:23:23 +0000346 vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000347 }
348 vars->lastcount = vars->count;
349 vars->drop_next = codel_control_law(now, params->interval,
Eric Dumazet536edd62012-05-12 03:32:13 +0000350 vars->rec_inv_sqrt);
Eric Dumazet76e3cc12012-05-10 07:51:25 +0000351 }
352end:
353 return skb;
354}
355#endif