1 #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 */
62 typedef u32 codel_time_t;
63 typedef s32 codel_tdiff_t;
64 #define CODEL_SHIFT 10
65 #define MS2TIME(a) ((a * NSEC_PER_MSEC) >> CODEL_SHIFT)
66
codel_get_time(void)67 static inline codel_time_t codel_get_time(void)
68 {
69 u64 ns = ktime_get_ns();
70
71 return ns >> CODEL_SHIFT;
72 }
73
74 /* 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)
89
90 /* Qdiscs using codel plugin must use codel_skb_cb in their own cb[] */
91 struct codel_skb_cb {
92 codel_time_t enqueue_time;
93 };
94
get_codel_cb(const struct sk_buff * skb)95 static 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
codel_get_enqueue_time(const struct sk_buff * skb)101 static codel_time_t codel_get_enqueue_time(const struct sk_buff *skb)
102 {
103 return get_codel_cb(skb)->enqueue_time;
104 }
105
codel_set_enqueue_time(struct sk_buff * skb)106 static void codel_set_enqueue_time(struct sk_buff *skb)
107 {
108 get_codel_cb(skb)->enqueue_time = codel_get_time();
109 }
110
codel_time_to_us(codel_time_t val)111 static 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 * @mtu: device mtu, or minimal queue backlog in bytes.
124 * @ecn: is Explicit Congestion Notification enabled
125 */
126 struct codel_params {
127 codel_time_t target;
128 codel_time_t interval;
129 u32 mtu;
130 bool ecn;
131 };
132
133 /**
134 * struct codel_vars - contains codel variables
135 * @count: how many drops we've done since the last time we
136 * entered dropping state
137 * @lastcount: count at entry to dropping state
138 * @dropping: set to true if in dropping state
139 * @rec_inv_sqrt: reciprocal value of sqrt(count) >> 1
140 * @first_above_time: when we went (or will go) continuously above target
141 * for interval
142 * @drop_next: time to drop next packet, or when we dropped last
143 * @ldelay: sojourn time of last dequeued packet
144 */
145 struct codel_vars {
146 u32 count;
147 u32 lastcount;
148 bool dropping;
149 u16 rec_inv_sqrt;
150 codel_time_t first_above_time;
151 codel_time_t drop_next;
152 codel_time_t ldelay;
153 };
154
155 #define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */
156 /* needed shift to get a Q0.32 number from rec_inv_sqrt */
157 #define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS)
158
159 /**
160 * struct codel_stats - contains codel shared variables and stats
161 * @maxpacket: largest packet we've seen so far
162 * @drop_count: temp count of dropped packets in dequeue()
163 * ecn_mark: number of packets we ECN marked instead of dropping
164 */
165 struct codel_stats {
166 u32 maxpacket;
167 u32 drop_count;
168 u32 ecn_mark;
169 };
170
codel_params_init(struct codel_params * params,const struct Qdisc * sch)171 static void codel_params_init(struct codel_params *params,
172 const struct Qdisc *sch)
173 {
174 params->interval = MS2TIME(100);
175 params->target = MS2TIME(5);
176 params->mtu = psched_mtu(qdisc_dev(sch));
177 params->ecn = false;
178 }
179
codel_vars_init(struct codel_vars * vars)180 static void codel_vars_init(struct codel_vars *vars)
181 {
182 memset(vars, 0, sizeof(*vars));
183 }
184
codel_stats_init(struct codel_stats * stats)185 static void codel_stats_init(struct codel_stats *stats)
186 {
187 stats->maxpacket = 0;
188 }
189
190 /*
191 * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
192 * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
193 *
194 * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
195 */
codel_Newton_step(struct codel_vars * vars)196 static void codel_Newton_step(struct codel_vars *vars)
197 {
198 u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
199 u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
200 u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
201
202 val >>= 2; /* avoid overflow in following multiply */
203 val = (val * invsqrt) >> (32 - 2 + 1);
204
205 vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
206 }
207
208 /*
209 * CoDel control_law is t + interval/sqrt(count)
210 * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
211 * both sqrt() and divide operation.
212 */
codel_control_law(codel_time_t t,codel_time_t interval,u32 rec_inv_sqrt)213 static codel_time_t codel_control_law(codel_time_t t,
214 codel_time_t interval,
215 u32 rec_inv_sqrt)
216 {
217 return t + reciprocal_scale(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
218 }
219
codel_should_drop(const struct sk_buff * skb,struct Qdisc * sch,struct codel_vars * vars,struct codel_params * params,struct codel_stats * stats,codel_time_t now)220 static bool codel_should_drop(const struct sk_buff *skb,
221 struct Qdisc *sch,
222 struct codel_vars *vars,
223 struct codel_params *params,
224 struct codel_stats *stats,
225 codel_time_t now)
226 {
227 bool ok_to_drop;
228
229 if (!skb) {
230 vars->first_above_time = 0;
231 return false;
232 }
233
234 vars->ldelay = now - codel_get_enqueue_time(skb);
235 sch->qstats.backlog -= qdisc_pkt_len(skb);
236
237 if (unlikely(qdisc_pkt_len(skb) > stats->maxpacket))
238 stats->maxpacket = qdisc_pkt_len(skb);
239
240 if (codel_time_before(vars->ldelay, params->target) ||
241 sch->qstats.backlog <= params->mtu) {
242 /* went below - stay below for at least interval */
243 vars->first_above_time = 0;
244 return false;
245 }
246 ok_to_drop = false;
247 if (vars->first_above_time == 0) {
248 /* just went above from below. If we stay above
249 * for at least interval we'll say it's ok to drop
250 */
251 vars->first_above_time = now + params->interval;
252 } else if (codel_time_after(now, vars->first_above_time)) {
253 ok_to_drop = true;
254 }
255 return ok_to_drop;
256 }
257
258 typedef struct sk_buff * (*codel_skb_dequeue_t)(struct codel_vars *vars,
259 struct Qdisc *sch);
260
261 static struct sk_buff *codel_dequeue(struct Qdisc *sch,
262 struct codel_params *params,
263 struct codel_vars *vars,
264 struct codel_stats *stats,
265 codel_skb_dequeue_t dequeue_func)
266 {
267 struct sk_buff *skb = dequeue_func(vars, sch);
268 codel_time_t now;
269 bool drop;
270
271 if (!skb) {
272 vars->dropping = false;
273 return skb;
274 }
275 now = codel_get_time();
276 drop = codel_should_drop(skb, sch, vars, params, stats, now);
277 if (vars->dropping) {
278 if (!drop) {
279 /* sojourn time below target - leave dropping state */
280 vars->dropping = false;
281 } else if (codel_time_after_eq(now, vars->drop_next)) {
282 /* It's time for the next drop. Drop the current
283 * packet and dequeue the next. The dequeue might
284 * take us out of dropping state.
285 * If not, schedule the next drop.
286 * A large backlog might result in drop rates so high
287 * that the next drop should happen now,
288 * hence the while loop.
289 */
290 while (vars->dropping &&
291 codel_time_after_eq(now, vars->drop_next)) {
292 vars->count++; /* dont care of possible wrap
293 * since there is no more divide
294 */
295 codel_Newton_step(vars);
296 if (params->ecn && INET_ECN_set_ce(skb)) {
297 stats->ecn_mark++;
298 vars->drop_next =
299 codel_control_law(vars->drop_next,
300 params->interval,
301 vars->rec_inv_sqrt);
302 goto end;
303 }
304 qdisc_drop(skb, sch);
305 stats->drop_count++;
306 skb = dequeue_func(vars, sch);
307 if (!codel_should_drop(skb, sch,
308 vars, params, stats, now)) {
309 /* leave dropping state */
310 vars->dropping = false;
311 } else {
312 /* and schedule the next drop */
313 vars->drop_next =
314 codel_control_law(vars->drop_next,
315 params->interval,
316 vars->rec_inv_sqrt);
317 }
318 }
319 }
320 } else if (drop) {
321 u32 delta;
322
323 if (params->ecn && INET_ECN_set_ce(skb)) {
324 stats->ecn_mark++;
325 } else {
326 qdisc_drop(skb, sch);
327 stats->drop_count++;
328
329 skb = dequeue_func(vars, sch);
330 drop = codel_should_drop(skb, sch, vars, params,
331 stats, now);
332 }
333 vars->dropping = true;
334 /* if min went above target close to when we last went below it
335 * assume that the drop rate that controlled the queue on the
336 * last cycle is a good starting point to control it now.
337 */
338 delta = vars->count - vars->lastcount;
339 if (delta > 1 &&
340 codel_time_before(now - vars->drop_next,
341 16 * params->interval)) {
342 vars->count = delta;
343 /* we dont care if rec_inv_sqrt approximation
344 * is not very precise :
345 * Next Newton steps will correct it quadratically.
346 */
347 codel_Newton_step(vars);
348 } else {
349 vars->count = 1;
350 vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
351 }
352 vars->lastcount = vars->count;
353 vars->drop_next = codel_control_law(now, params->interval,
354 vars->rec_inv_sqrt);
355 }
356 end:
357 return skb;
358 }
359 #endif
360