1 #ifndef __NET_SCHED_RED_H
2 #define __NET_SCHED_RED_H
3
4 #include <linux/types.h>
5 #include <linux/bug.h>
6 #include <net/pkt_sched.h>
7 #include <net/inet_ecn.h>
8 #include <net/dsfield.h>
9 #include <linux/reciprocal_div.h>
10
11 /* Random Early Detection (RED) algorithm.
12 =======================================
13
14 Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
15 for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
16
17 This file codes a "divisionless" version of RED algorithm
18 as written down in Fig.17 of the paper.
19
20 Short description.
21 ------------------
22
23 When a new packet arrives we calculate the average queue length:
24
25 avg = (1-W)*avg + W*current_queue_len,
26
27 W is the filter time constant (chosen as 2^(-Wlog)), it controls
28 the inertia of the algorithm. To allow larger bursts, W should be
29 decreased.
30
31 if (avg > th_max) -> packet marked (dropped).
32 if (avg < th_min) -> packet passes.
33 if (th_min < avg < th_max) we calculate probability:
34
35 Pb = max_P * (avg - th_min)/(th_max-th_min)
36
37 and mark (drop) packet with this probability.
38 Pb changes from 0 (at avg==th_min) to max_P (avg==th_max).
39 max_P should be small (not 1), usually 0.01..0.02 is good value.
40
41 max_P is chosen as a number, so that max_P/(th_max-th_min)
42 is a negative power of two in order arithmetics to contain
43 only shifts.
44
45
46 Parameters, settable by user:
47 -----------------------------
48
49 qth_min - bytes (should be < qth_max/2)
50 qth_max - bytes (should be at least 2*qth_min and less limit)
51 Wlog - bits (<32) log(1/W).
52 Plog - bits (<32)
53
54 Plog is related to max_P by formula:
55
56 max_P = (qth_max-qth_min)/2^Plog;
57
58 F.e. if qth_max=128K and qth_min=32K, then Plog=22
59 corresponds to max_P=0.02
60
61 Scell_log
62 Stab
63
64 Lookup table for log((1-W)^(t/t_ave).
65
66
67 NOTES:
68
69 Upper bound on W.
70 -----------------
71
72 If you want to allow bursts of L packets of size S,
73 you should choose W:
74
75 L + 1 - th_min/S < (1-(1-W)^L)/W
76
77 th_min/S = 32 th_min/S = 4
78
79 log(W) L
80 -1 33
81 -2 35
82 -3 39
83 -4 46
84 -5 57
85 -6 75
86 -7 101
87 -8 135
88 -9 190
89 etc.
90 */
91
92 /*
93 * Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM
94 * (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001
95 *
96 * Every 500 ms:
97 * if (avg > target and max_p <= 0.5)
98 * increase max_p : max_p += alpha;
99 * else if (avg < target and max_p >= 0.01)
100 * decrease max_p : max_p *= beta;
101 *
102 * target :[qth_min + 0.4*(qth_min - qth_max),
103 * qth_min + 0.6*(qth_min - qth_max)].
104 * alpha : min(0.01, max_p / 4)
105 * beta : 0.9
106 * max_P is a Q0.32 fixed point number (with 32 bits mantissa)
107 * max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ]
108 */
109 #define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100))
110
111 #define MAX_P_MIN (1 * RED_ONE_PERCENT)
112 #define MAX_P_MAX (50 * RED_ONE_PERCENT)
113 #define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4)
114
115 #define RED_STAB_SIZE 256
116 #define RED_STAB_MASK (RED_STAB_SIZE - 1)
117
118 struct red_stats {
119 u32 prob_drop; /* Early probability drops */
120 u32 prob_mark; /* Early probability marks */
121 u32 forced_drop; /* Forced drops, qavg > max_thresh */
122 u32 forced_mark; /* Forced marks, qavg > max_thresh */
123 u32 pdrop; /* Drops due to queue limits */
124 u32 other; /* Drops due to drop() calls */
125 };
126
127 struct red_parms {
128 /* Parameters */
129 u32 qth_min; /* Min avg length threshold: Wlog scaled */
130 u32 qth_max; /* Max avg length threshold: Wlog scaled */
131 u32 Scell_max;
132 u32 max_P; /* probability, [0 .. 1.0] 32 scaled */
133 /* reciprocal_value(max_P / qth_delta) */
134 struct reciprocal_value max_P_reciprocal;
135 u32 qth_delta; /* max_th - min_th */
136 u32 target_min; /* min_th + 0.4*(max_th - min_th) */
137 u32 target_max; /* min_th + 0.6*(max_th - min_th) */
138 u8 Scell_log;
139 u8 Wlog; /* log(W) */
140 u8 Plog; /* random number bits */
141 u8 Stab[RED_STAB_SIZE];
142 };
143
144 struct red_vars {
145 /* Variables */
146 int qcount; /* Number of packets since last random
147 number generation */
148 u32 qR; /* Cached random number */
149
150 unsigned long qavg; /* Average queue length: Wlog scaled */
151 ktime_t qidlestart; /* Start of current idle period */
152 };
153
red_maxp(u8 Plog)154 static inline u32 red_maxp(u8 Plog)
155 {
156 return Plog < 32 ? (~0U >> Plog) : ~0U;
157 }
158
red_set_vars(struct red_vars * v)159 static inline void red_set_vars(struct red_vars *v)
160 {
161 /* Reset average queue length, the value is strictly bound
162 * to the parameters below, reseting hurts a bit but leaving
163 * it might result in an unreasonable qavg for a while. --TGR
164 */
165 v->qavg = 0;
166
167 v->qcount = -1;
168 }
169
red_check_params(u32 qth_min,u32 qth_max,u8 Wlog,u8 Scell_log,u8 * stab)170 static inline bool red_check_params(u32 qth_min, u32 qth_max, u8 Wlog,
171 u8 Scell_log, u8 *stab)
172 {
173 if (fls(qth_min) + Wlog >= 32)
174 return false;
175 if (fls(qth_max) + Wlog >= 32)
176 return false;
177 if (Scell_log >= 32)
178 return false;
179 if (qth_max < qth_min)
180 return false;
181 if (stab) {
182 int i;
183
184 for (i = 0; i < RED_STAB_SIZE; i++)
185 if (stab[i] >= 32)
186 return false;
187 }
188 return true;
189 }
190
red_set_parms(struct red_parms * p,u32 qth_min,u32 qth_max,u8 Wlog,u8 Plog,u8 Scell_log,u8 * stab,u32 max_P)191 static inline void red_set_parms(struct red_parms *p,
192 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
193 u8 Scell_log, u8 *stab, u32 max_P)
194 {
195 int delta = qth_max - qth_min;
196 u32 max_p_delta;
197
198 p->qth_min = qth_min << Wlog;
199 p->qth_max = qth_max << Wlog;
200 p->Wlog = Wlog;
201 p->Plog = Plog;
202 if (delta <= 0)
203 delta = 1;
204 p->qth_delta = delta;
205 if (!max_P) {
206 max_P = red_maxp(Plog);
207 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */
208 }
209 p->max_P = max_P;
210 max_p_delta = max_P / delta;
211 max_p_delta = max(max_p_delta, 1U);
212 p->max_P_reciprocal = reciprocal_value(max_p_delta);
213
214 /* RED Adaptative target :
215 * [min_th + 0.4*(min_th - max_th),
216 * min_th + 0.6*(min_th - max_th)].
217 */
218 delta /= 5;
219 p->target_min = qth_min + 2*delta;
220 p->target_max = qth_min + 3*delta;
221
222 p->Scell_log = Scell_log;
223 p->Scell_max = (255 << Scell_log);
224
225 if (stab)
226 memcpy(p->Stab, stab, sizeof(p->Stab));
227 }
228
red_is_idling(const struct red_vars * v)229 static inline int red_is_idling(const struct red_vars *v)
230 {
231 return v->qidlestart.tv64 != 0;
232 }
233
red_start_of_idle_period(struct red_vars * v)234 static inline void red_start_of_idle_period(struct red_vars *v)
235 {
236 v->qidlestart = ktime_get();
237 }
238
red_end_of_idle_period(struct red_vars * v)239 static inline void red_end_of_idle_period(struct red_vars *v)
240 {
241 v->qidlestart.tv64 = 0;
242 }
243
red_restart(struct red_vars * v)244 static inline void red_restart(struct red_vars *v)
245 {
246 red_end_of_idle_period(v);
247 v->qavg = 0;
248 v->qcount = -1;
249 }
250
red_calc_qavg_from_idle_time(const struct red_parms * p,const struct red_vars * v)251 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p,
252 const struct red_vars *v)
253 {
254 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart);
255 long us_idle = min_t(s64, delta, p->Scell_max);
256 int shift;
257
258 /*
259 * The problem: ideally, average length queue recalcultion should
260 * be done over constant clock intervals. This is too expensive, so
261 * that the calculation is driven by outgoing packets.
262 * When the queue is idle we have to model this clock by hand.
263 *
264 * SF+VJ proposed to "generate":
265 *
266 * m = idletime / (average_pkt_size / bandwidth)
267 *
268 * dummy packets as a burst after idle time, i.e.
269 *
270 * v->qavg *= (1-W)^m
271 *
272 * This is an apparently overcomplicated solution (f.e. we have to
273 * precompute a table to make this calculation in reasonable time)
274 * I believe that a simpler model may be used here,
275 * but it is field for experiments.
276 */
277
278 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
279
280 if (shift)
281 return v->qavg >> shift;
282 else {
283 /* Approximate initial part of exponent with linear function:
284 *
285 * (1-W)^m ~= 1-mW + ...
286 *
287 * Seems, it is the best solution to
288 * problem of too coarse exponent tabulation.
289 */
290 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log;
291
292 if (us_idle < (v->qavg >> 1))
293 return v->qavg - us_idle;
294 else
295 return v->qavg >> 1;
296 }
297 }
298
red_calc_qavg_no_idle_time(const struct red_parms * p,const struct red_vars * v,unsigned int backlog)299 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p,
300 const struct red_vars *v,
301 unsigned int backlog)
302 {
303 /*
304 * NOTE: v->qavg is fixed point number with point at Wlog.
305 * The formula below is equvalent to floating point
306 * version:
307 *
308 * qavg = qavg*(1-W) + backlog*W;
309 *
310 * --ANK (980924)
311 */
312 return v->qavg + (backlog - (v->qavg >> p->Wlog));
313 }
314
red_calc_qavg(const struct red_parms * p,const struct red_vars * v,unsigned int backlog)315 static inline unsigned long red_calc_qavg(const struct red_parms *p,
316 const struct red_vars *v,
317 unsigned int backlog)
318 {
319 if (!red_is_idling(v))
320 return red_calc_qavg_no_idle_time(p, v, backlog);
321 else
322 return red_calc_qavg_from_idle_time(p, v);
323 }
324
325
red_random(const struct red_parms * p)326 static inline u32 red_random(const struct red_parms *p)
327 {
328 return reciprocal_divide(prandom_u32(), p->max_P_reciprocal);
329 }
330
red_mark_probability(const struct red_parms * p,const struct red_vars * v,unsigned long qavg)331 static inline int red_mark_probability(const struct red_parms *p,
332 const struct red_vars *v,
333 unsigned long qavg)
334 {
335 /* The formula used below causes questions.
336
337 OK. qR is random number in the interval
338 (0..1/max_P)*(qth_max-qth_min)
339 i.e. 0..(2^Plog). If we used floating point
340 arithmetics, it would be: (2^Plog)*rnd_num,
341 where rnd_num is less 1.
342
343 Taking into account, that qavg have fixed
344 point at Wlog, two lines
345 below have the following floating point equivalent:
346
347 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
348
349 Any questions? --ANK (980924)
350 */
351 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR);
352 }
353
354 enum {
355 RED_BELOW_MIN_THRESH,
356 RED_BETWEEN_TRESH,
357 RED_ABOVE_MAX_TRESH,
358 };
359
red_cmp_thresh(const struct red_parms * p,unsigned long qavg)360 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg)
361 {
362 if (qavg < p->qth_min)
363 return RED_BELOW_MIN_THRESH;
364 else if (qavg >= p->qth_max)
365 return RED_ABOVE_MAX_TRESH;
366 else
367 return RED_BETWEEN_TRESH;
368 }
369
370 enum {
371 RED_DONT_MARK,
372 RED_PROB_MARK,
373 RED_HARD_MARK,
374 };
375
red_action(const struct red_parms * p,struct red_vars * v,unsigned long qavg)376 static inline int red_action(const struct red_parms *p,
377 struct red_vars *v,
378 unsigned long qavg)
379 {
380 switch (red_cmp_thresh(p, qavg)) {
381 case RED_BELOW_MIN_THRESH:
382 v->qcount = -1;
383 return RED_DONT_MARK;
384
385 case RED_BETWEEN_TRESH:
386 if (++v->qcount) {
387 if (red_mark_probability(p, v, qavg)) {
388 v->qcount = 0;
389 v->qR = red_random(p);
390 return RED_PROB_MARK;
391 }
392 } else
393 v->qR = red_random(p);
394
395 return RED_DONT_MARK;
396
397 case RED_ABOVE_MAX_TRESH:
398 v->qcount = -1;
399 return RED_HARD_MARK;
400 }
401
402 BUG();
403 return RED_DONT_MARK;
404 }
405
red_adaptative_algo(struct red_parms * p,struct red_vars * v)406 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v)
407 {
408 unsigned long qavg;
409 u32 max_p_delta;
410
411 qavg = v->qavg;
412 if (red_is_idling(v))
413 qavg = red_calc_qavg_from_idle_time(p, v);
414
415 /* v->qavg is fixed point number with point at Wlog */
416 qavg >>= p->Wlog;
417
418 if (qavg > p->target_max && p->max_P <= MAX_P_MAX)
419 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */
420 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN)
421 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */
422
423 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta);
424 max_p_delta = max(max_p_delta, 1U);
425 p->max_P_reciprocal = reciprocal_value(max_p_delta);
426 }
427 #endif
428