blob: 3319f16b3beb899727c7a434e75fb1010d7ee139 [file] [log] [blame]
Thomas Grafa7834742005-11-05 21:14:03 +01001#ifndef __NET_SCHED_RED_H
2#define __NET_SCHED_RED_H
3
Thomas Grafa7834742005-11-05 21:14:03 +01004#include <linux/types.h>
5#include <net/pkt_sched.h>
6#include <net/inet_ecn.h>
7#include <net/dsfield.h>
8
9/* Random Early Detection (RED) algorithm.
10 =======================================
11
12 Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
13 for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
14
15 This file codes a "divisionless" version of RED algorithm
16 as written down in Fig.17 of the paper.
17
18 Short description.
19 ------------------
20
21 When a new packet arrives we calculate the average queue length:
22
23 avg = (1-W)*avg + W*current_queue_len,
24
25 W is the filter time constant (chosen as 2^(-Wlog)), it controls
26 the inertia of the algorithm. To allow larger bursts, W should be
27 decreased.
28
29 if (avg > th_max) -> packet marked (dropped).
30 if (avg < th_min) -> packet passes.
31 if (th_min < avg < th_max) we calculate probability:
32
33 Pb = max_P * (avg - th_min)/(th_max-th_min)
34
35 and mark (drop) packet with this probability.
36 Pb changes from 0 (at avg==th_min) to max_P (avg==th_max).
37 max_P should be small (not 1), usually 0.01..0.02 is good value.
38
39 max_P is chosen as a number, so that max_P/(th_max-th_min)
40 is a negative power of two in order arithmetics to contain
41 only shifts.
42
43
44 Parameters, settable by user:
45 -----------------------------
46
47 qth_min - bytes (should be < qth_max/2)
48 qth_max - bytes (should be at least 2*qth_min and less limit)
49 Wlog - bits (<32) log(1/W).
50 Plog - bits (<32)
51
52 Plog is related to max_P by formula:
53
54 max_P = (qth_max-qth_min)/2^Plog;
55
56 F.e. if qth_max=128K and qth_min=32K, then Plog=22
57 corresponds to max_P=0.02
58
59 Scell_log
60 Stab
61
62 Lookup table for log((1-W)^(t/t_ave).
63
64
65 NOTES:
66
67 Upper bound on W.
68 -----------------
69
70 If you want to allow bursts of L packets of size S,
71 you should choose W:
72
73 L + 1 - th_min/S < (1-(1-W)^L)/W
74
75 th_min/S = 32 th_min/S = 4
76
77 log(W) L
78 -1 33
79 -2 35
80 -3 39
81 -4 46
82 -5 57
83 -6 75
84 -7 101
85 -8 135
86 -9 190
87 etc.
88 */
89
90#define RED_STAB_SIZE 256
91#define RED_STAB_MASK (RED_STAB_SIZE - 1)
92
Eric Dumazetfd2c3ef2009-11-03 03:26:03 +000093struct red_stats {
Thomas Grafa7834742005-11-05 21:14:03 +010094 u32 prob_drop; /* Early probability drops */
95 u32 prob_mark; /* Early probability marks */
96 u32 forced_drop; /* Forced drops, qavg > max_thresh */
97 u32 forced_mark; /* Forced marks, qavg > max_thresh */
98 u32 pdrop; /* Drops due to queue limits */
99 u32 other; /* Drops due to drop() calls */
Thomas Grafa7834742005-11-05 21:14:03 +0100100};
101
Eric Dumazetfd2c3ef2009-11-03 03:26:03 +0000102struct red_parms {
Thomas Grafa7834742005-11-05 21:14:03 +0100103 /* Parameters */
104 u32 qth_min; /* Min avg length threshold: A scaled */
105 u32 qth_max; /* Max avg length threshold: A scaled */
106 u32 Scell_max;
107 u32 Rmask; /* Cached random mask, see red_rmask */
108 u8 Scell_log;
109 u8 Wlog; /* log(W) */
110 u8 Plog; /* random number bits */
111 u8 Stab[RED_STAB_SIZE];
112
113 /* Variables */
114 int qcount; /* Number of packets since last random
115 number generation */
116 u32 qR; /* Cached random number */
117
118 unsigned long qavg; /* Average queue length: A scaled */
119 psched_time_t qidlestart; /* Start of current idle period */
120};
121
122static inline u32 red_rmask(u8 Plog)
123{
124 return Plog < 32 ? ((1 << Plog) - 1) : ~0UL;
125}
126
127static inline void red_set_parms(struct red_parms *p,
128 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
129 u8 Scell_log, u8 *stab)
130{
131 /* Reset average queue length, the value is strictly bound
132 * to the parameters below, reseting hurts a bit but leaving
133 * it might result in an unreasonable qavg for a while. --TGR
134 */
135 p->qavg = 0;
136
137 p->qcount = -1;
138 p->qth_min = qth_min << Wlog;
139 p->qth_max = qth_max << Wlog;
140 p->Wlog = Wlog;
141 p->Plog = Plog;
142 p->Rmask = red_rmask(Plog);
143 p->Scell_log = Scell_log;
144 p->Scell_max = (255 << Scell_log);
145
146 memcpy(p->Stab, stab, sizeof(p->Stab));
147}
148
149static inline int red_is_idling(struct red_parms *p)
150{
Patrick McHardya0849802007-03-23 11:28:30 -0700151 return p->qidlestart != PSCHED_PASTPERFECT;
Thomas Grafa7834742005-11-05 21:14:03 +0100152}
153
154static inline void red_start_of_idle_period(struct red_parms *p)
155{
Patrick McHardy3bebcda2007-03-23 11:29:25 -0700156 p->qidlestart = psched_get_time();
Thomas Grafa7834742005-11-05 21:14:03 +0100157}
158
159static inline void red_end_of_idle_period(struct red_parms *p)
160{
Patrick McHardya0849802007-03-23 11:28:30 -0700161 p->qidlestart = PSCHED_PASTPERFECT;
Thomas Grafa7834742005-11-05 21:14:03 +0100162}
163
164static inline void red_restart(struct red_parms *p)
165{
166 red_end_of_idle_period(p);
167 p->qavg = 0;
168 p->qcount = -1;
169}
170
171static inline unsigned long red_calc_qavg_from_idle_time(struct red_parms *p)
172{
173 psched_time_t now;
174 long us_idle;
175 int shift;
176
Patrick McHardy3bebcda2007-03-23 11:29:25 -0700177 now = psched_get_time();
Patrick McHardy03cc45c2007-03-23 11:29:11 -0700178 us_idle = psched_tdiff_bounded(now, p->qidlestart, p->Scell_max);
Thomas Grafa7834742005-11-05 21:14:03 +0100179
180 /*
181 * The problem: ideally, average length queue recalcultion should
182 * be done over constant clock intervals. This is too expensive, so
183 * that the calculation is driven by outgoing packets.
184 * When the queue is idle we have to model this clock by hand.
185 *
186 * SF+VJ proposed to "generate":
187 *
188 * m = idletime / (average_pkt_size / bandwidth)
189 *
190 * dummy packets as a burst after idle time, i.e.
191 *
192 * p->qavg *= (1-W)^m
193 *
194 * This is an apparently overcomplicated solution (f.e. we have to
195 * precompute a table to make this calculation in reasonable time)
196 * I believe that a simpler model may be used here,
197 * but it is field for experiments.
198 */
199
200 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
201
202 if (shift)
203 return p->qavg >> shift;
204 else {
205 /* Approximate initial part of exponent with linear function:
206 *
207 * (1-W)^m ~= 1-mW + ...
208 *
209 * Seems, it is the best solution to
210 * problem of too coarse exponent tabulation.
211 */
Ilpo Järvinenc4c0ce52006-08-04 16:36:18 -0700212 us_idle = (p->qavg * (u64)us_idle) >> p->Scell_log;
Thomas Grafa7834742005-11-05 21:14:03 +0100213
214 if (us_idle < (p->qavg >> 1))
215 return p->qavg - us_idle;
216 else
217 return p->qavg >> 1;
218 }
219}
220
221static inline unsigned long red_calc_qavg_no_idle_time(struct red_parms *p,
222 unsigned int backlog)
223{
224 /*
225 * NOTE: p->qavg is fixed point number with point at Wlog.
226 * The formula below is equvalent to floating point
227 * version:
228 *
229 * qavg = qavg*(1-W) + backlog*W;
230 *
231 * --ANK (980924)
232 */
233 return p->qavg + (backlog - (p->qavg >> p->Wlog));
234}
235
236static inline unsigned long red_calc_qavg(struct red_parms *p,
237 unsigned int backlog)
238{
239 if (!red_is_idling(p))
240 return red_calc_qavg_no_idle_time(p, backlog);
241 else
242 return red_calc_qavg_from_idle_time(p);
243}
244
245static inline u32 red_random(struct red_parms *p)
246{
247 return net_random() & p->Rmask;
248}
249
250static inline int red_mark_probability(struct red_parms *p, unsigned long qavg)
251{
252 /* The formula used below causes questions.
253
254 OK. qR is random number in the interval 0..Rmask
255 i.e. 0..(2^Plog). If we used floating point
256 arithmetics, it would be: (2^Plog)*rnd_num,
257 where rnd_num is less 1.
258
259 Taking into account, that qavg have fixed
260 point at Wlog, and Plog is related to max_P by
261 max_P = (qth_max-qth_min)/2^Plog; two lines
262 below have the following floating point equivalent:
263
264 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
265
266 Any questions? --ANK (980924)
267 */
268 return !(((qavg - p->qth_min) >> p->Wlog) * p->qcount < p->qR);
269}
270
271enum {
272 RED_BELOW_MIN_THRESH,
273 RED_BETWEEN_TRESH,
274 RED_ABOVE_MAX_TRESH,
275};
276
277static inline int red_cmp_thresh(struct red_parms *p, unsigned long qavg)
278{
279 if (qavg < p->qth_min)
280 return RED_BELOW_MIN_THRESH;
281 else if (qavg >= p->qth_max)
282 return RED_ABOVE_MAX_TRESH;
283 else
284 return RED_BETWEEN_TRESH;
285}
286
287enum {
288 RED_DONT_MARK,
289 RED_PROB_MARK,
290 RED_HARD_MARK,
291};
292
293static inline int red_action(struct red_parms *p, unsigned long qavg)
294{
295 switch (red_cmp_thresh(p, qavg)) {
296 case RED_BELOW_MIN_THRESH:
297 p->qcount = -1;
298 return RED_DONT_MARK;
299
300 case RED_BETWEEN_TRESH:
301 if (++p->qcount) {
302 if (red_mark_probability(p, qavg)) {
303 p->qcount = 0;
304 p->qR = red_random(p);
305 return RED_PROB_MARK;
306 }
307 } else
308 p->qR = red_random(p);
309
310 return RED_DONT_MARK;
311
312 case RED_ABOVE_MAX_TRESH:
313 p->qcount = -1;
314 return RED_HARD_MARK;
315 }
316
317 BUG();
318 return RED_DONT_MARK;
319}
320
321#endif