xref: /linux/include/net/codel_impl.h (revision 7f4a5ec6258fd7c92633ec4b0493fc51166d9398)
1 #ifndef __NET_SCHED_CODEL_IMPL_H
2 #define __NET_SCHED_CODEL_IMPL_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) 2012,2015 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 /* Controlling Queue Delay (CoDel) algorithm
45  * =========================================
46  * Source : Kathleen Nichols and Van Jacobson
47  * http://queue.acm.org/detail.cfm?id=2209336
48  *
49  * Implemented on linux by Dave Taht and Eric Dumazet
50  */
51 
52 #include <net/inet_ecn.h>
53 
54 static void codel_params_init(struct codel_params *params)
55 {
56 	params->interval = MS2TIME(100);
57 	params->target = MS2TIME(5);
58 	params->ce_threshold = CODEL_DISABLED_THRESHOLD;
59 	params->ce_threshold_mask = 0;
60 	params->ce_threshold_selector = 0;
61 	params->ecn = false;
62 }
63 
64 static void codel_vars_init(struct codel_vars *vars)
65 {
66 	memset(vars, 0, sizeof(*vars));
67 }
68 
69 static void codel_stats_init(struct codel_stats *stats)
70 {
71 	stats->maxpacket = 0;
72 }
73 
74 /*
75  * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
76  * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
77  *
78  * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
79  */
80 static void codel_Newton_step(struct codel_vars *vars)
81 {
82 	u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
83 	u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
84 	u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
85 
86 	val >>= 2; /* avoid overflow in following multiply */
87 	val = (val * invsqrt) >> (32 - 2 + 1);
88 
89 	vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
90 }
91 
92 /*
93  * CoDel control_law is t + interval/sqrt(count)
94  * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
95  * both sqrt() and divide operation.
96  *
97  * Clamp the increment to at least 1 tick: a very small interval (or a
98  * large count) can truncate it to zero, stalling the dropping loop.
99  */
100 static codel_time_t codel_control_law(codel_time_t t,
101 				      codel_time_t interval,
102 				      u32 rec_inv_sqrt)
103 {
104 	return t + max_t(u32, 1,
105 			 reciprocal_scale(interval,
106 					  rec_inv_sqrt << REC_INV_SQRT_SHIFT));
107 }
108 
109 static bool codel_should_drop(const struct sk_buff *skb,
110 			      void *ctx,
111 			      struct codel_vars *vars,
112 			      struct codel_params *params,
113 			      struct codel_stats *stats,
114 			      codel_skb_len_t skb_len_func,
115 			      codel_skb_time_t skb_time_func,
116 			      u32 *backlog,
117 			      codel_time_t now)
118 {
119 	bool ok_to_drop;
120 	u32 skb_len;
121 
122 	if (!skb) {
123 		vars->first_above_time = 0;
124 		return false;
125 	}
126 
127 	skb_len = skb_len_func(skb);
128 	WRITE_ONCE(vars->ldelay, now - skb_time_func(skb));
129 
130 	if (unlikely(skb_len > stats->maxpacket))
131 		WRITE_ONCE(stats->maxpacket, skb_len);
132 
133 	if (codel_time_before(vars->ldelay, params->target) ||
134 	    *backlog <= params->mtu) {
135 		/* went below - stay below for at least interval */
136 		vars->first_above_time = 0;
137 		return false;
138 	}
139 	ok_to_drop = false;
140 	if (vars->first_above_time == 0) {
141 		/* just went above from below. If we stay above
142 		 * for at least interval we'll say it's ok to drop
143 		 */
144 		vars->first_above_time = now + params->interval;
145 	} else if (codel_time_after(now, vars->first_above_time)) {
146 		ok_to_drop = true;
147 	}
148 	return ok_to_drop;
149 }
150 
151 static struct sk_buff *codel_dequeue(void *ctx,
152 				     u32 *backlog,
153 				     struct codel_params *params,
154 				     struct codel_vars *vars,
155 				     struct codel_stats *stats,
156 				     codel_skb_len_t skb_len_func,
157 				     codel_skb_time_t skb_time_func,
158 				     codel_skb_drop_t drop_func,
159 				     codel_skb_dequeue_t dequeue_func)
160 {
161 	struct sk_buff *skb = dequeue_func(vars, ctx);
162 	unsigned int drops = 0;
163 	codel_time_t now;
164 	bool drop;
165 
166 	if (!skb) {
167 		vars->first_above_time = 0;
168 		WRITE_ONCE(vars->dropping, false);
169 		return skb;
170 	}
171 	now = codel_get_time();
172 	drop = codel_should_drop(skb, ctx, vars, params, stats,
173 				 skb_len_func, skb_time_func, backlog, now);
174 	if (vars->dropping) {
175 		if (!drop) {
176 			/* sojourn time below target - leave dropping state */
177 			WRITE_ONCE(vars->dropping, false);
178 		} else if (codel_time_after_eq(now, vars->drop_next)) {
179 			/* It's time for the next drop. Drop the current
180 			 * packet and dequeue the next. The dequeue might
181 			 * take us out of dropping state.
182 			 * If not, schedule the next drop.
183 			 * A large backlog might result in drop rates so high
184 			 * that the next drop should happen now,
185 			 * hence the while loop.
186 			 */
187 			while (vars->dropping &&
188 			       codel_time_after_eq(now, vars->drop_next)) {
189 				if (++drops > CODEL_MAX_DROPS_PER_DEQUEUE) {
190 					/* fell far behind the schedule */
191 					WRITE_ONCE(vars->drop_next,
192 						   codel_control_law(now,
193 								     params->interval,
194 								     vars->rec_inv_sqrt));
195 					break;
196 				}
197 				/* dont care of possible wrap
198 				 * since there is no more divide.
199 				 */
200 				WRITE_ONCE(vars->count, vars->count + 1);
201 				codel_Newton_step(vars);
202 				if (params->ecn && INET_ECN_set_ce(skb)) {
203 					WRITE_ONCE(stats->ecn_mark,
204 						   stats->ecn_mark + 1);
205 					WRITE_ONCE(vars->drop_next,
206 						codel_control_law(vars->drop_next,
207 								  params->interval,
208 								  vars->rec_inv_sqrt));
209 					goto end;
210 				}
211 				stats->drop_len += skb_len_func(skb);
212 				drop_func(skb, ctx);
213 				stats->drop_count++;
214 				skb = dequeue_func(vars, ctx);
215 				if (!codel_should_drop(skb, ctx,
216 						       vars, params, stats,
217 						       skb_len_func,
218 						       skb_time_func,
219 						       backlog, now)) {
220 					/* leave dropping state */
221 					WRITE_ONCE(vars->dropping, false);
222 				} else {
223 					/* and schedule the next drop */
224 					WRITE_ONCE(vars->drop_next,
225 						codel_control_law(vars->drop_next,
226 								  params->interval,
227 								  vars->rec_inv_sqrt));
228 				}
229 			}
230 		}
231 	} else if (drop) {
232 		u32 delta;
233 
234 		if (params->ecn && INET_ECN_set_ce(skb)) {
235 			WRITE_ONCE(stats->ecn_mark, stats->ecn_mark + 1);
236 		} else {
237 			stats->drop_len += skb_len_func(skb);
238 			drop_func(skb, ctx);
239 			stats->drop_count++;
240 
241 			skb = dequeue_func(vars, ctx);
242 			drop = codel_should_drop(skb, ctx, vars, params,
243 						 stats, skb_len_func,
244 						 skb_time_func, backlog, now);
245 		}
246 		WRITE_ONCE(vars->dropping, true);
247 		/* if min went above target close to when we last went below it
248 		 * assume that the drop rate that controlled the queue on the
249 		 * last cycle is a good starting point to control it now.
250 		 */
251 		delta = vars->count - vars->lastcount;
252 		if (delta > 1 &&
253 		    codel_time_before(now - vars->drop_next,
254 				      16 * params->interval)) {
255 			WRITE_ONCE(vars->count, delta);
256 			/* we dont care if rec_inv_sqrt approximation
257 			 * is not very precise :
258 			 * Next Newton steps will correct it quadratically.
259 			 */
260 			codel_Newton_step(vars);
261 		} else {
262 			WRITE_ONCE(vars->count, 1);
263 			vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
264 		}
265 		WRITE_ONCE(vars->lastcount, vars->count);
266 		WRITE_ONCE(vars->drop_next,
267 			   codel_control_law(now, params->interval,
268 					     vars->rec_inv_sqrt));
269 	}
270 end:
271 	if (skb && codel_time_after(vars->ldelay, params->ce_threshold)) {
272 		bool set_ce = true;
273 
274 		if (params->ce_threshold_mask) {
275 			int dsfield = skb_get_dsfield(skb);
276 
277 			set_ce = (dsfield >= 0 &&
278 				  (((u8)dsfield & params->ce_threshold_mask) ==
279 				   params->ce_threshold_selector));
280 		}
281 		if (set_ce && INET_ECN_set_ce(skb))
282 			WRITE_ONCE(stats->ce_mark, stats->ce_mark + 1);
283 	}
284 	return skb;
285 }
286 
287 #endif
288