xref: /linux/include/net/red.h (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
1 #ifndef __NET_SCHED_RED_H
2 #define __NET_SCHED_RED_H
3 
4 #include <linux/types.h>
5 #include <net/pkt_sched.h>
6 #include <net/inet_ecn.h>
7 #include <net/dsfield.h>
8 #include <linux/reciprocal_div.h>
9 
10 /*	Random Early Detection (RED) algorithm.
11 	=======================================
12 
13 	Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
14 	for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
15 
16 	This file codes a "divisionless" version of RED algorithm
17 	as written down in Fig.17 of the paper.
18 
19 	Short description.
20 	------------------
21 
22 	When a new packet arrives we calculate the average queue length:
23 
24 	avg = (1-W)*avg + W*current_queue_len,
25 
26 	W is the filter time constant (chosen as 2^(-Wlog)), it controls
27 	the inertia of the algorithm. To allow larger bursts, W should be
28 	decreased.
29 
30 	if (avg > th_max) -> packet marked (dropped).
31 	if (avg < th_min) -> packet passes.
32 	if (th_min < avg < th_max) we calculate probability:
33 
34 	Pb = max_P * (avg - th_min)/(th_max-th_min)
35 
36 	and mark (drop) packet with this probability.
37 	Pb changes from 0 (at avg==th_min) to max_P (avg==th_max).
38 	max_P should be small (not 1), usually 0.01..0.02 is good value.
39 
40 	max_P is chosen as a number, so that max_P/(th_max-th_min)
41 	is a negative power of two in order arithmetics to contain
42 	only shifts.
43 
44 
45 	Parameters, settable by user:
46 	-----------------------------
47 
48 	qth_min		- bytes (should be < qth_max/2)
49 	qth_max		- bytes (should be at least 2*qth_min and less limit)
50 	Wlog	       	- bits (<32) log(1/W).
51 	Plog	       	- bits (<32)
52 
53 	Plog is related to max_P by formula:
54 
55 	max_P = (qth_max-qth_min)/2^Plog;
56 
57 	F.e. if qth_max=128K and qth_min=32K, then Plog=22
58 	corresponds to max_P=0.02
59 
60 	Scell_log
61 	Stab
62 
63 	Lookup table for log((1-W)^(t/t_ave).
64 
65 
66 	NOTES:
67 
68 	Upper bound on W.
69 	-----------------
70 
71 	If you want to allow bursts of L packets of size S,
72 	you should choose W:
73 
74 	L + 1 - th_min/S < (1-(1-W)^L)/W
75 
76 	th_min/S = 32         th_min/S = 4
77 
78 	log(W)	L
79 	-1	33
80 	-2	35
81 	-3	39
82 	-4	46
83 	-5	57
84 	-6	75
85 	-7	101
86 	-8	135
87 	-9	190
88 	etc.
89  */
90 
91 /*
92  * Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM
93  * (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001
94  *
95  * Every 500 ms:
96  *  if (avg > target and max_p <= 0.5)
97  *   increase max_p : max_p += alpha;
98  *  else if (avg < target and max_p >= 0.01)
99  *   decrease max_p : max_p *= beta;
100  *
101  * target :[qth_min + 0.4*(qth_min - qth_max),
102  *          qth_min + 0.6*(qth_min - qth_max)].
103  * alpha : min(0.01, max_p / 4)
104  * beta : 0.9
105  * max_P is a Q0.32 fixed point number (with 32 bits mantissa)
106  * max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ]
107  */
108 #define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100))
109 
110 #define MAX_P_MIN (1 * RED_ONE_PERCENT)
111 #define MAX_P_MAX (50 * RED_ONE_PERCENT)
112 #define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4)
113 
114 #define RED_STAB_SIZE	256
115 #define RED_STAB_MASK	(RED_STAB_SIZE - 1)
116 
117 struct red_stats {
118 	u32		prob_drop;	/* Early probability drops */
119 	u32		prob_mark;	/* Early probability marks */
120 	u32		forced_drop;	/* Forced drops, qavg > max_thresh */
121 	u32		forced_mark;	/* Forced marks, qavg > max_thresh */
122 	u32		pdrop;          /* Drops due to queue limits */
123 	u32		other;          /* Drops due to drop() calls */
124 };
125 
126 struct red_parms {
127 	/* Parameters */
128 	u32		qth_min;	/* Min avg length threshold: Wlog scaled */
129 	u32		qth_max;	/* Max avg length threshold: Wlog scaled */
130 	u32		Scell_max;
131 	u32		max_P;		/* probability, [0 .. 1.0] 32 scaled */
132 	u32		max_P_reciprocal; /* reciprocal_value(max_P / qth_delta) */
133 	u32		qth_delta;	/* max_th - min_th */
134 	u32		target_min;	/* min_th + 0.4*(max_th - min_th) */
135 	u32		target_max;	/* min_th + 0.6*(max_th - min_th) */
136 	u8		Scell_log;
137 	u8		Wlog;		/* log(W)		*/
138 	u8		Plog;		/* random number bits	*/
139 	u8		Stab[RED_STAB_SIZE];
140 };
141 
142 struct red_vars {
143 	/* Variables */
144 	int		qcount;		/* Number of packets since last random
145 					   number generation */
146 	u32		qR;		/* Cached random number */
147 
148 	unsigned long	qavg;		/* Average queue length: Wlog scaled */
149 	ktime_t		qidlestart;	/* Start of current idle period */
150 };
151 
152 static inline u32 red_maxp(u8 Plog)
153 {
154 	return Plog < 32 ? (~0U >> Plog) : ~0U;
155 }
156 
157 static inline void red_set_vars(struct red_vars *v)
158 {
159 	/* Reset average queue length, the value is strictly bound
160 	 * to the parameters below, reseting hurts a bit but leaving
161 	 * it might result in an unreasonable qavg for a while. --TGR
162 	 */
163 	v->qavg		= 0;
164 
165 	v->qcount	= -1;
166 }
167 
168 static inline void red_set_parms(struct red_parms *p,
169 				 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
170 				 u8 Scell_log, u8 *stab, u32 max_P)
171 {
172 	int delta = qth_max - qth_min;
173 	u32 max_p_delta;
174 
175 	p->qth_min	= qth_min << Wlog;
176 	p->qth_max	= qth_max << Wlog;
177 	p->Wlog		= Wlog;
178 	p->Plog		= Plog;
179 	if (delta < 0)
180 		delta = 1;
181 	p->qth_delta	= delta;
182 	if (!max_P) {
183 		max_P = red_maxp(Plog);
184 		max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */
185 	}
186 	p->max_P = max_P;
187 	max_p_delta = max_P / delta;
188 	max_p_delta = max(max_p_delta, 1U);
189 	p->max_P_reciprocal  = reciprocal_value(max_p_delta);
190 
191 	/* RED Adaptative target :
192 	 * [min_th + 0.4*(min_th - max_th),
193 	 *  min_th + 0.6*(min_th - max_th)].
194 	 */
195 	delta /= 5;
196 	p->target_min = qth_min + 2*delta;
197 	p->target_max = qth_min + 3*delta;
198 
199 	p->Scell_log	= Scell_log;
200 	p->Scell_max	= (255 << Scell_log);
201 
202 	if (stab)
203 		memcpy(p->Stab, stab, sizeof(p->Stab));
204 }
205 
206 static inline int red_is_idling(const struct red_vars *v)
207 {
208 	return v->qidlestart.tv64 != 0;
209 }
210 
211 static inline void red_start_of_idle_period(struct red_vars *v)
212 {
213 	v->qidlestart = ktime_get();
214 }
215 
216 static inline void red_end_of_idle_period(struct red_vars *v)
217 {
218 	v->qidlestart.tv64 = 0;
219 }
220 
221 static inline void red_restart(struct red_vars *v)
222 {
223 	red_end_of_idle_period(v);
224 	v->qavg = 0;
225 	v->qcount = -1;
226 }
227 
228 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p,
229 							 const struct red_vars *v)
230 {
231 	s64 delta = ktime_us_delta(ktime_get(), v->qidlestart);
232 	long us_idle = min_t(s64, delta, p->Scell_max);
233 	int  shift;
234 
235 	/*
236 	 * The problem: ideally, average length queue recalcultion should
237 	 * be done over constant clock intervals. This is too expensive, so
238 	 * that the calculation is driven by outgoing packets.
239 	 * When the queue is idle we have to model this clock by hand.
240 	 *
241 	 * SF+VJ proposed to "generate":
242 	 *
243 	 *	m = idletime / (average_pkt_size / bandwidth)
244 	 *
245 	 * dummy packets as a burst after idle time, i.e.
246 	 *
247 	 * 	p->qavg *= (1-W)^m
248 	 *
249 	 * This is an apparently overcomplicated solution (f.e. we have to
250 	 * precompute a table to make this calculation in reasonable time)
251 	 * I believe that a simpler model may be used here,
252 	 * but it is field for experiments.
253 	 */
254 
255 	shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
256 
257 	if (shift)
258 		return v->qavg >> shift;
259 	else {
260 		/* Approximate initial part of exponent with linear function:
261 		 *
262 		 * 	(1-W)^m ~= 1-mW + ...
263 		 *
264 		 * Seems, it is the best solution to
265 		 * problem of too coarse exponent tabulation.
266 		 */
267 		us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log;
268 
269 		if (us_idle < (v->qavg >> 1))
270 			return v->qavg - us_idle;
271 		else
272 			return v->qavg >> 1;
273 	}
274 }
275 
276 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p,
277 						       const struct red_vars *v,
278 						       unsigned int backlog)
279 {
280 	/*
281 	 * NOTE: p->qavg is fixed point number with point at Wlog.
282 	 * The formula below is equvalent to floating point
283 	 * version:
284 	 *
285 	 * 	qavg = qavg*(1-W) + backlog*W;
286 	 *
287 	 * --ANK (980924)
288 	 */
289 	return v->qavg + (backlog - (v->qavg >> p->Wlog));
290 }
291 
292 static inline unsigned long red_calc_qavg(const struct red_parms *p,
293 					  const struct red_vars *v,
294 					  unsigned int backlog)
295 {
296 	if (!red_is_idling(v))
297 		return red_calc_qavg_no_idle_time(p, v, backlog);
298 	else
299 		return red_calc_qavg_from_idle_time(p, v);
300 }
301 
302 
303 static inline u32 red_random(const struct red_parms *p)
304 {
305 	return reciprocal_divide(net_random(), p->max_P_reciprocal);
306 }
307 
308 static inline int red_mark_probability(const struct red_parms *p,
309 				       const struct red_vars *v,
310 				       unsigned long qavg)
311 {
312 	/* The formula used below causes questions.
313 
314 	   OK. qR is random number in the interval
315 		(0..1/max_P)*(qth_max-qth_min)
316 	   i.e. 0..(2^Plog). If we used floating point
317 	   arithmetics, it would be: (2^Plog)*rnd_num,
318 	   where rnd_num is less 1.
319 
320 	   Taking into account, that qavg have fixed
321 	   point at Wlog, two lines
322 	   below have the following floating point equivalent:
323 
324 	   max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
325 
326 	   Any questions? --ANK (980924)
327 	 */
328 	return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR);
329 }
330 
331 enum {
332 	RED_BELOW_MIN_THRESH,
333 	RED_BETWEEN_TRESH,
334 	RED_ABOVE_MAX_TRESH,
335 };
336 
337 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg)
338 {
339 	if (qavg < p->qth_min)
340 		return RED_BELOW_MIN_THRESH;
341 	else if (qavg >= p->qth_max)
342 		return RED_ABOVE_MAX_TRESH;
343 	else
344 		return RED_BETWEEN_TRESH;
345 }
346 
347 enum {
348 	RED_DONT_MARK,
349 	RED_PROB_MARK,
350 	RED_HARD_MARK,
351 };
352 
353 static inline int red_action(const struct red_parms *p,
354 			     struct red_vars *v,
355 			     unsigned long qavg)
356 {
357 	switch (red_cmp_thresh(p, qavg)) {
358 		case RED_BELOW_MIN_THRESH:
359 			v->qcount = -1;
360 			return RED_DONT_MARK;
361 
362 		case RED_BETWEEN_TRESH:
363 			if (++v->qcount) {
364 				if (red_mark_probability(p, v, qavg)) {
365 					v->qcount = 0;
366 					v->qR = red_random(p);
367 					return RED_PROB_MARK;
368 				}
369 			} else
370 				v->qR = red_random(p);
371 
372 			return RED_DONT_MARK;
373 
374 		case RED_ABOVE_MAX_TRESH:
375 			v->qcount = -1;
376 			return RED_HARD_MARK;
377 	}
378 
379 	BUG();
380 	return RED_DONT_MARK;
381 }
382 
383 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v)
384 {
385 	unsigned long qavg;
386 	u32 max_p_delta;
387 
388 	qavg = v->qavg;
389 	if (red_is_idling(v))
390 		qavg = red_calc_qavg_from_idle_time(p, v);
391 
392 	/* p->qavg is fixed point number with point at Wlog */
393 	qavg >>= p->Wlog;
394 
395 	if (qavg > p->target_max && p->max_P <= MAX_P_MAX)
396 		p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */
397 	else if (qavg < p->target_min && p->max_P >= MAX_P_MIN)
398 		p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */
399 
400 	max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta);
401 	max_p_delta = max(max_p_delta, 1U);
402 	p->max_P_reciprocal = reciprocal_value(max_p_delta);
403 }
404 #endif
405