Linux 2.6.33-rc2
[pohmelfs.git] / include / net / red.h
blob995108e54d9f0adde1344ff76daf821567ae9b83
1 #ifndef __NET_SCHED_RED_H
2 #define __NET_SCHED_RED_H
4 #include <linux/types.h>
5 #include <net/pkt_sched.h>
6 #include <net/inet_ecn.h>
7 #include <net/dsfield.h>
9 /* Random Early Detection (RED) algorithm.
10 =======================================
12 Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
13 for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
15 This file codes a "divisionless" version of RED algorithm
16 as written down in Fig.17 of the paper.
18 Short description.
19 ------------------
21 When a new packet arrives we calculate the average queue length:
23 avg = (1-W)*avg + W*current_queue_len,
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.
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:
33 Pb = max_P * (avg - th_min)/(th_max-th_min)
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.
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.
44 Parameters, settable by user:
45 -----------------------------
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)
52 Plog is related to max_P by formula:
54 max_P = (qth_max-qth_min)/2^Plog;
56 F.e. if qth_max=128K and qth_min=32K, then Plog=22
57 corresponds to max_P=0.02
59 Scell_log
60 Stab
62 Lookup table for log((1-W)^(t/t_ave).
65 NOTES:
67 Upper bound on W.
68 -----------------
70 If you want to allow bursts of L packets of size S,
71 you should choose W:
73 L + 1 - th_min/S < (1-(1-W)^L)/W
75 th_min/S = 32 th_min/S = 4
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.
90 #define RED_STAB_SIZE 256
91 #define RED_STAB_MASK (RED_STAB_SIZE - 1)
93 struct red_stats {
94 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 */
100 u32 backlog;
103 struct red_parms {
104 /* Parameters */
105 u32 qth_min; /* Min avg length threshold: A scaled */
106 u32 qth_max; /* Max avg length threshold: A scaled */
107 u32 Scell_max;
108 u32 Rmask; /* Cached random mask, see red_rmask */
109 u8 Scell_log;
110 u8 Wlog; /* log(W) */
111 u8 Plog; /* random number bits */
112 u8 Stab[RED_STAB_SIZE];
114 /* Variables */
115 int qcount; /* Number of packets since last random
116 number generation */
117 u32 qR; /* Cached random number */
119 unsigned long qavg; /* Average queue length: A scaled */
120 psched_time_t qidlestart; /* Start of current idle period */
123 static inline u32 red_rmask(u8 Plog)
125 return Plog < 32 ? ((1 << Plog) - 1) : ~0UL;
128 static inline void red_set_parms(struct red_parms *p,
129 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
130 u8 Scell_log, u8 *stab)
132 /* Reset average queue length, the value is strictly bound
133 * to the parameters below, reseting hurts a bit but leaving
134 * it might result in an unreasonable qavg for a while. --TGR
136 p->qavg = 0;
138 p->qcount = -1;
139 p->qth_min = qth_min << Wlog;
140 p->qth_max = qth_max << Wlog;
141 p->Wlog = Wlog;
142 p->Plog = Plog;
143 p->Rmask = red_rmask(Plog);
144 p->Scell_log = Scell_log;
145 p->Scell_max = (255 << Scell_log);
147 memcpy(p->Stab, stab, sizeof(p->Stab));
150 static inline int red_is_idling(struct red_parms *p)
152 return p->qidlestart != PSCHED_PASTPERFECT;
155 static inline void red_start_of_idle_period(struct red_parms *p)
157 p->qidlestart = psched_get_time();
160 static inline void red_end_of_idle_period(struct red_parms *p)
162 p->qidlestart = PSCHED_PASTPERFECT;
165 static inline void red_restart(struct red_parms *p)
167 red_end_of_idle_period(p);
168 p->qavg = 0;
169 p->qcount = -1;
172 static inline unsigned long red_calc_qavg_from_idle_time(struct red_parms *p)
174 psched_time_t now;
175 long us_idle;
176 int shift;
178 now = psched_get_time();
179 us_idle = psched_tdiff_bounded(now, p->qidlestart, p->Scell_max);
182 * The problem: ideally, average length queue recalcultion should
183 * be done over constant clock intervals. This is too expensive, so
184 * that the calculation is driven by outgoing packets.
185 * When the queue is idle we have to model this clock by hand.
187 * SF+VJ proposed to "generate":
189 * m = idletime / (average_pkt_size / bandwidth)
191 * dummy packets as a burst after idle time, i.e.
193 * p->qavg *= (1-W)^m
195 * This is an apparently overcomplicated solution (f.e. we have to
196 * precompute a table to make this calculation in reasonable time)
197 * I believe that a simpler model may be used here,
198 * but it is field for experiments.
201 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
203 if (shift)
204 return p->qavg >> shift;
205 else {
206 /* Approximate initial part of exponent with linear function:
208 * (1-W)^m ~= 1-mW + ...
210 * Seems, it is the best solution to
211 * problem of too coarse exponent tabulation.
213 us_idle = (p->qavg * (u64)us_idle) >> p->Scell_log;
215 if (us_idle < (p->qavg >> 1))
216 return p->qavg - us_idle;
217 else
218 return p->qavg >> 1;
222 static inline unsigned long red_calc_qavg_no_idle_time(struct red_parms *p,
223 unsigned int backlog)
226 * NOTE: p->qavg is fixed point number with point at Wlog.
227 * The formula below is equvalent to floating point
228 * version:
230 * qavg = qavg*(1-W) + backlog*W;
232 * --ANK (980924)
234 return p->qavg + (backlog - (p->qavg >> p->Wlog));
237 static inline unsigned long red_calc_qavg(struct red_parms *p,
238 unsigned int backlog)
240 if (!red_is_idling(p))
241 return red_calc_qavg_no_idle_time(p, backlog);
242 else
243 return red_calc_qavg_from_idle_time(p);
246 static inline u32 red_random(struct red_parms *p)
248 return net_random() & p->Rmask;
251 static inline int red_mark_probability(struct red_parms *p, unsigned long qavg)
253 /* The formula used below causes questions.
255 OK. qR is random number in the interval 0..Rmask
256 i.e. 0..(2^Plog). If we used floating point
257 arithmetics, it would be: (2^Plog)*rnd_num,
258 where rnd_num is less 1.
260 Taking into account, that qavg have fixed
261 point at Wlog, and Plog is related to max_P by
262 max_P = (qth_max-qth_min)/2^Plog; two lines
263 below have the following floating point equivalent:
265 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
267 Any questions? --ANK (980924)
269 return !(((qavg - p->qth_min) >> p->Wlog) * p->qcount < p->qR);
272 enum {
273 RED_BELOW_MIN_THRESH,
274 RED_BETWEEN_TRESH,
275 RED_ABOVE_MAX_TRESH,
278 static inline int red_cmp_thresh(struct red_parms *p, unsigned long qavg)
280 if (qavg < p->qth_min)
281 return RED_BELOW_MIN_THRESH;
282 else if (qavg >= p->qth_max)
283 return RED_ABOVE_MAX_TRESH;
284 else
285 return RED_BETWEEN_TRESH;
288 enum {
289 RED_DONT_MARK,
290 RED_PROB_MARK,
291 RED_HARD_MARK,
294 static inline int red_action(struct red_parms *p, unsigned long qavg)
296 switch (red_cmp_thresh(p, qavg)) {
297 case RED_BELOW_MIN_THRESH:
298 p->qcount = -1;
299 return RED_DONT_MARK;
301 case RED_BETWEEN_TRESH:
302 if (++p->qcount) {
303 if (red_mark_probability(p, qavg)) {
304 p->qcount = 0;
305 p->qR = red_random(p);
306 return RED_PROB_MARK;
308 } else
309 p->qR = red_random(p);
311 return RED_DONT_MARK;
313 case RED_ABOVE_MAX_TRESH:
314 p->qcount = -1;
315 return RED_HARD_MARK;
318 BUG();
319 return RED_DONT_MARK;
322 #endif