]>
Commit | Line | Data |
---|---|---|
a7834745 TG |
1 | #ifndef __NET_SCHED_RED_H |
2 | #define __NET_SCHED_RED_H | |
3 | ||
a7834745 TG |
4 | #include <linux/types.h> |
5 | #include <net/pkt_sched.h> | |
6 | #include <net/inet_ecn.h> | |
7 | #include <net/dsfield.h> | |
8af2a218 | 8 | #include <linux/reciprocal_div.h> |
a7834745 TG |
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 | ||
8af2a218 ED |
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 | ||
a7834745 TG |
114 | #define RED_STAB_SIZE 256 |
115 | #define RED_STAB_MASK (RED_STAB_SIZE - 1) | |
116 | ||
fd2c3ef7 | 117 | struct red_stats { |
a7834745 TG |
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 */ | |
a7834745 TG |
124 | }; |
125 | ||
fd2c3ef7 | 126 | struct red_parms { |
a7834745 | 127 | /* Parameters */ |
8af2a218 ED |
128 | u32 qth_min; /* Min avg length threshold: Wlog scaled */ |
129 | u32 qth_max; /* Max avg length threshold: Wlog scaled */ | |
a7834745 | 130 | u32 Scell_max; |
8af2a218 ED |
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) */ | |
a7834745 TG |
136 | u8 Scell_log; |
137 | u8 Wlog; /* log(W) */ | |
138 | u8 Plog; /* random number bits */ | |
139 | u8 Stab[RED_STAB_SIZE]; | |
140 | ||
141 | /* Variables */ | |
142 | int qcount; /* Number of packets since last random | |
143 | number generation */ | |
144 | u32 qR; /* Cached random number */ | |
145 | ||
8af2a218 | 146 | unsigned long qavg; /* Average queue length: Wlog scaled */ |
ea6a5d3b | 147 | ktime_t qidlestart; /* Start of current idle period */ |
a7834745 TG |
148 | }; |
149 | ||
8af2a218 | 150 | static inline u32 red_maxp(u8 Plog) |
a7834745 | 151 | { |
8af2a218 | 152 | return Plog < 32 ? (~0U >> Plog) : ~0U; |
a7834745 TG |
153 | } |
154 | ||
8af2a218 | 155 | |
a7834745 TG |
156 | static inline void red_set_parms(struct red_parms *p, |
157 | u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, | |
158 | u8 Scell_log, u8 *stab) | |
159 | { | |
8af2a218 ED |
160 | int delta = qth_max - qth_min; |
161 | ||
a7834745 TG |
162 | /* Reset average queue length, the value is strictly bound |
163 | * to the parameters below, reseting hurts a bit but leaving | |
164 | * it might result in an unreasonable qavg for a while. --TGR | |
165 | */ | |
166 | p->qavg = 0; | |
167 | ||
168 | p->qcount = -1; | |
169 | p->qth_min = qth_min << Wlog; | |
170 | p->qth_max = qth_max << Wlog; | |
171 | p->Wlog = Wlog; | |
172 | p->Plog = Plog; | |
8af2a218 ED |
173 | if (delta < 0) |
174 | delta = 1; | |
175 | p->qth_delta = delta; | |
176 | p->max_P = red_maxp(Plog); | |
177 | p->max_P *= delta; /* max_P = (qth_max-qth_min)/2^Plog */ | |
178 | ||
179 | p->max_P_reciprocal = reciprocal_value(p->max_P / delta); | |
180 | ||
181 | /* RED Adaptative target : | |
182 | * [min_th + 0.4*(min_th - max_th), | |
183 | * min_th + 0.6*(min_th - max_th)]. | |
184 | */ | |
185 | delta /= 5; | |
186 | p->target_min = qth_min + 2*delta; | |
187 | p->target_max = qth_min + 3*delta; | |
188 | ||
a7834745 TG |
189 | p->Scell_log = Scell_log; |
190 | p->Scell_max = (255 << Scell_log); | |
191 | ||
192 | memcpy(p->Stab, stab, sizeof(p->Stab)); | |
193 | } | |
194 | ||
8af2a218 | 195 | static inline int red_is_idling(const struct red_parms *p) |
a7834745 | 196 | { |
ea6a5d3b | 197 | return p->qidlestart.tv64 != 0; |
a7834745 TG |
198 | } |
199 | ||
200 | static inline void red_start_of_idle_period(struct red_parms *p) | |
201 | { | |
ea6a5d3b | 202 | p->qidlestart = ktime_get(); |
a7834745 TG |
203 | } |
204 | ||
205 | static inline void red_end_of_idle_period(struct red_parms *p) | |
206 | { | |
ea6a5d3b | 207 | p->qidlestart.tv64 = 0; |
a7834745 TG |
208 | } |
209 | ||
210 | static inline void red_restart(struct red_parms *p) | |
211 | { | |
212 | red_end_of_idle_period(p); | |
213 | p->qavg = 0; | |
214 | p->qcount = -1; | |
215 | } | |
216 | ||
8af2a218 | 217 | static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p) |
a7834745 | 218 | { |
ea6a5d3b ED |
219 | s64 delta = ktime_us_delta(ktime_get(), p->qidlestart); |
220 | long us_idle = min_t(s64, delta, p->Scell_max); | |
a7834745 TG |
221 | int shift; |
222 | ||
a7834745 TG |
223 | /* |
224 | * The problem: ideally, average length queue recalcultion should | |
225 | * be done over constant clock intervals. This is too expensive, so | |
226 | * that the calculation is driven by outgoing packets. | |
227 | * When the queue is idle we have to model this clock by hand. | |
228 | * | |
229 | * SF+VJ proposed to "generate": | |
230 | * | |
231 | * m = idletime / (average_pkt_size / bandwidth) | |
232 | * | |
233 | * dummy packets as a burst after idle time, i.e. | |
234 | * | |
235 | * p->qavg *= (1-W)^m | |
236 | * | |
237 | * This is an apparently overcomplicated solution (f.e. we have to | |
238 | * precompute a table to make this calculation in reasonable time) | |
239 | * I believe that a simpler model may be used here, | |
240 | * but it is field for experiments. | |
241 | */ | |
242 | ||
243 | shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; | |
244 | ||
245 | if (shift) | |
246 | return p->qavg >> shift; | |
247 | else { | |
248 | /* Approximate initial part of exponent with linear function: | |
249 | * | |
250 | * (1-W)^m ~= 1-mW + ... | |
251 | * | |
252 | * Seems, it is the best solution to | |
253 | * problem of too coarse exponent tabulation. | |
254 | */ | |
c4c0ce5c | 255 | us_idle = (p->qavg * (u64)us_idle) >> p->Scell_log; |
a7834745 TG |
256 | |
257 | if (us_idle < (p->qavg >> 1)) | |
258 | return p->qavg - us_idle; | |
259 | else | |
260 | return p->qavg >> 1; | |
261 | } | |
262 | } | |
263 | ||
8af2a218 | 264 | static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p, |
a7834745 TG |
265 | unsigned int backlog) |
266 | { | |
267 | /* | |
268 | * NOTE: p->qavg is fixed point number with point at Wlog. | |
269 | * The formula below is equvalent to floating point | |
270 | * version: | |
271 | * | |
272 | * qavg = qavg*(1-W) + backlog*W; | |
273 | * | |
274 | * --ANK (980924) | |
275 | */ | |
276 | return p->qavg + (backlog - (p->qavg >> p->Wlog)); | |
277 | } | |
278 | ||
8af2a218 | 279 | static inline unsigned long red_calc_qavg(const struct red_parms *p, |
a7834745 TG |
280 | unsigned int backlog) |
281 | { | |
282 | if (!red_is_idling(p)) | |
283 | return red_calc_qavg_no_idle_time(p, backlog); | |
284 | else | |
285 | return red_calc_qavg_from_idle_time(p); | |
286 | } | |
287 | ||
8af2a218 ED |
288 | |
289 | static inline u32 red_random(const struct red_parms *p) | |
a7834745 | 290 | { |
8af2a218 | 291 | return reciprocal_divide(net_random(), p->max_P_reciprocal); |
a7834745 TG |
292 | } |
293 | ||
8af2a218 | 294 | static inline int red_mark_probability(const struct red_parms *p, unsigned long qavg) |
a7834745 TG |
295 | { |
296 | /* The formula used below causes questions. | |
297 | ||
8af2a218 ED |
298 | OK. qR is random number in the interval |
299 | (0..1/max_P)*(qth_max-qth_min) | |
a7834745 TG |
300 | i.e. 0..(2^Plog). If we used floating point |
301 | arithmetics, it would be: (2^Plog)*rnd_num, | |
302 | where rnd_num is less 1. | |
303 | ||
304 | Taking into account, that qavg have fixed | |
8af2a218 | 305 | point at Wlog, two lines |
a7834745 TG |
306 | below have the following floating point equivalent: |
307 | ||
308 | max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount | |
309 | ||
310 | Any questions? --ANK (980924) | |
311 | */ | |
312 | return !(((qavg - p->qth_min) >> p->Wlog) * p->qcount < p->qR); | |
313 | } | |
314 | ||
315 | enum { | |
316 | RED_BELOW_MIN_THRESH, | |
317 | RED_BETWEEN_TRESH, | |
318 | RED_ABOVE_MAX_TRESH, | |
319 | }; | |
320 | ||
321 | static inline int red_cmp_thresh(struct red_parms *p, unsigned long qavg) | |
322 | { | |
323 | if (qavg < p->qth_min) | |
324 | return RED_BELOW_MIN_THRESH; | |
325 | else if (qavg >= p->qth_max) | |
326 | return RED_ABOVE_MAX_TRESH; | |
327 | else | |
328 | return RED_BETWEEN_TRESH; | |
329 | } | |
330 | ||
331 | enum { | |
332 | RED_DONT_MARK, | |
333 | RED_PROB_MARK, | |
334 | RED_HARD_MARK, | |
335 | }; | |
336 | ||
337 | static inline int red_action(struct red_parms *p, unsigned long qavg) | |
338 | { | |
339 | switch (red_cmp_thresh(p, qavg)) { | |
340 | case RED_BELOW_MIN_THRESH: | |
341 | p->qcount = -1; | |
342 | return RED_DONT_MARK; | |
343 | ||
344 | case RED_BETWEEN_TRESH: | |
345 | if (++p->qcount) { | |
346 | if (red_mark_probability(p, qavg)) { | |
347 | p->qcount = 0; | |
348 | p->qR = red_random(p); | |
349 | return RED_PROB_MARK; | |
350 | } | |
351 | } else | |
352 | p->qR = red_random(p); | |
353 | ||
354 | return RED_DONT_MARK; | |
355 | ||
356 | case RED_ABOVE_MAX_TRESH: | |
357 | p->qcount = -1; | |
358 | return RED_HARD_MARK; | |
359 | } | |
360 | ||
361 | BUG(); | |
362 | return RED_DONT_MARK; | |
363 | } | |
364 | ||
8af2a218 ED |
365 | static inline void red_adaptative_algo(struct red_parms *p) |
366 | { | |
367 | unsigned long qavg; | |
368 | u32 max_p_delta; | |
369 | ||
370 | qavg = p->qavg; | |
371 | if (red_is_idling(p)) | |
372 | qavg = red_calc_qavg_from_idle_time(p); | |
373 | ||
374 | /* p->qavg is fixed point number with point at Wlog */ | |
375 | qavg >>= p->Wlog; | |
376 | ||
377 | if (qavg > p->target_max && p->max_P <= MAX_P_MAX) | |
378 | p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */ | |
379 | else if (qavg < p->target_min && p->max_P >= MAX_P_MIN) | |
380 | p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */ | |
381 | ||
382 | max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta); | |
383 | p->max_P_reciprocal = reciprocal_value(max_p_delta); | |
384 | } | |
a7834745 | 385 | #endif |