blob: 932f0d79d60cbbab60db73baddc8b650935b3bf6 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001/* SPDX-License-Identifier: GPL-2.0 */
Thomas Grafa7834742005-11-05 21:14:03 +01002#ifndef __NET_SCHED_RED_H
3#define __NET_SCHED_RED_H
4
Thomas Grafa7834742005-11-05 21:14:03 +01005#include <linux/types.h>
Paul Gortmaker187f1882011-11-23 20:12:59 -05006#include <linux/bug.h>
Thomas Grafa7834742005-11-05 21:14:03 +01007#include <net/pkt_sched.h>
8#include <net/inet_ecn.h>
9#include <net/dsfield.h>
Eric Dumazet8af2a212011-12-08 06:06:03 +000010#include <linux/reciprocal_div.h>
Thomas Grafa7834742005-11-05 21:14:03 +010011
12/* Random Early Detection (RED) algorithm.
13 =======================================
14
15 Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways
16 for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking.
17
18 This file codes a "divisionless" version of RED algorithm
19 as written down in Fig.17 of the paper.
20
21 Short description.
22 ------------------
23
24 When a new packet arrives we calculate the average queue length:
25
26 avg = (1-W)*avg + W*current_queue_len,
27
28 W is the filter time constant (chosen as 2^(-Wlog)), it controls
29 the inertia of the algorithm. To allow larger bursts, W should be
30 decreased.
31
32 if (avg > th_max) -> packet marked (dropped).
33 if (avg < th_min) -> packet passes.
34 if (th_min < avg < th_max) we calculate probability:
35
36 Pb = max_P * (avg - th_min)/(th_max-th_min)
37
38 and mark (drop) packet with this probability.
39 Pb changes from 0 (at avg==th_min) to max_P (avg==th_max).
40 max_P should be small (not 1), usually 0.01..0.02 is good value.
41
42 max_P is chosen as a number, so that max_P/(th_max-th_min)
43 is a negative power of two in order arithmetics to contain
44 only shifts.
45
46
47 Parameters, settable by user:
48 -----------------------------
49
50 qth_min - bytes (should be < qth_max/2)
51 qth_max - bytes (should be at least 2*qth_min and less limit)
52 Wlog - bits (<32) log(1/W).
53 Plog - bits (<32)
54
55 Plog is related to max_P by formula:
56
57 max_P = (qth_max-qth_min)/2^Plog;
58
59 F.e. if qth_max=128K and qth_min=32K, then Plog=22
60 corresponds to max_P=0.02
61
62 Scell_log
63 Stab
64
65 Lookup table for log((1-W)^(t/t_ave).
66
67
68 NOTES:
69
70 Upper bound on W.
71 -----------------
72
73 If you want to allow bursts of L packets of size S,
74 you should choose W:
75
76 L + 1 - th_min/S < (1-(1-W)^L)/W
77
78 th_min/S = 32 th_min/S = 4
79
80 log(W) L
81 -1 33
82 -2 35
83 -3 39
84 -4 46
85 -5 57
86 -6 75
87 -7 101
88 -8 135
89 -9 190
90 etc.
91 */
92
Eric Dumazet8af2a212011-12-08 06:06:03 +000093/*
94 * Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM
95 * (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001
96 *
97 * Every 500 ms:
98 * if (avg > target and max_p <= 0.5)
99 * increase max_p : max_p += alpha;
100 * else if (avg < target and max_p >= 0.01)
101 * decrease max_p : max_p *= beta;
102 *
103 * target :[qth_min + 0.4*(qth_min - qth_max),
104 * qth_min + 0.6*(qth_min - qth_max)].
105 * alpha : min(0.01, max_p / 4)
106 * beta : 0.9
107 * max_P is a Q0.32 fixed point number (with 32 bits mantissa)
108 * max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ]
109 */
110#define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100))
111
112#define MAX_P_MIN (1 * RED_ONE_PERCENT)
113#define MAX_P_MAX (50 * RED_ONE_PERCENT)
114#define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4)
115
Thomas Grafa7834742005-11-05 21:14:03 +0100116#define RED_STAB_SIZE 256
117#define RED_STAB_MASK (RED_STAB_SIZE - 1)
118
Eric Dumazetfd2c3ef2009-11-03 03:26:03 +0000119struct red_stats {
Thomas Grafa7834742005-11-05 21:14:03 +0100120 u32 prob_drop; /* Early probability drops */
121 u32 prob_mark; /* Early probability marks */
122 u32 forced_drop; /* Forced drops, qavg > max_thresh */
123 u32 forced_mark; /* Forced marks, qavg > max_thresh */
124 u32 pdrop; /* Drops due to queue limits */
125 u32 other; /* Drops due to drop() calls */
Thomas Grafa7834742005-11-05 21:14:03 +0100126};
127
Eric Dumazetfd2c3ef2009-11-03 03:26:03 +0000128struct red_parms {
Thomas Grafa7834742005-11-05 21:14:03 +0100129 /* Parameters */
Eric Dumazet8af2a212011-12-08 06:06:03 +0000130 u32 qth_min; /* Min avg length threshold: Wlog scaled */
131 u32 qth_max; /* Max avg length threshold: Wlog scaled */
Thomas Grafa7834742005-11-05 21:14:03 +0100132 u32 Scell_max;
Eric Dumazet8af2a212011-12-08 06:06:03 +0000133 u32 max_P; /* probability, [0 .. 1.0] 32 scaled */
Hannes Frederic Sowa809fa972014-01-22 02:29:41 +0100134 /* reciprocal_value(max_P / qth_delta) */
135 struct reciprocal_value max_P_reciprocal;
Eric Dumazet8af2a212011-12-08 06:06:03 +0000136 u32 qth_delta; /* max_th - min_th */
137 u32 target_min; /* min_th + 0.4*(max_th - min_th) */
138 u32 target_max; /* min_th + 0.6*(max_th - min_th) */
Thomas Grafa7834742005-11-05 21:14:03 +0100139 u8 Scell_log;
140 u8 Wlog; /* log(W) */
141 u8 Plog; /* random number bits */
142 u8 Stab[RED_STAB_SIZE];
Eric Dumazeteeca6682012-01-05 02:25:16 +0000143};
Thomas Grafa7834742005-11-05 21:14:03 +0100144
Eric Dumazeteeca6682012-01-05 02:25:16 +0000145struct red_vars {
Thomas Grafa7834742005-11-05 21:14:03 +0100146 /* Variables */
147 int qcount; /* Number of packets since last random
148 number generation */
149 u32 qR; /* Cached random number */
150
Eric Dumazet8af2a212011-12-08 06:06:03 +0000151 unsigned long qavg; /* Average queue length: Wlog scaled */
Eric Dumazetea6a5d32011-11-30 12:10:53 +0000152 ktime_t qidlestart; /* Start of current idle period */
Thomas Grafa7834742005-11-05 21:14:03 +0100153};
154
Eric Dumazet8af2a212011-12-08 06:06:03 +0000155static inline u32 red_maxp(u8 Plog)
Thomas Grafa7834742005-11-05 21:14:03 +0100156{
Eric Dumazet8af2a212011-12-08 06:06:03 +0000157 return Plog < 32 ? (~0U >> Plog) : ~0U;
Thomas Grafa7834742005-11-05 21:14:03 +0100158}
159
Eric Dumazeteeca6682012-01-05 02:25:16 +0000160static inline void red_set_vars(struct red_vars *v)
161{
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 v->qavg = 0;
167
168 v->qcount = -1;
169}
Eric Dumazet8af2a212011-12-08 06:06:03 +0000170
Randy Dunlap7a209692020-12-24 22:23:44 -0800171static inline bool red_check_params(u32 qth_min, u32 qth_max, u8 Wlog, u8 Scell_log)
Nogah Frankel8afa10c2017-12-04 13:31:11 +0200172{
173 if (fls(qth_min) + Wlog > 32)
174 return false;
175 if (fls(qth_max) + Wlog > 32)
176 return false;
Randy Dunlap7a209692020-12-24 22:23:44 -0800177 if (Scell_log >= 32)
178 return false;
Nogah Frankel8afa10c2017-12-04 13:31:11 +0200179 if (qth_max < qth_min)
180 return false;
181 return true;
182}
183
Petr Machata14bc1752020-03-13 01:10:56 +0200184static inline int red_get_flags(unsigned char qopt_flags,
185 unsigned char historic_mask,
186 struct nlattr *flags_attr,
187 unsigned char supported_mask,
188 struct nla_bitfield32 *p_flags,
189 unsigned char *p_userbits,
190 struct netlink_ext_ack *extack)
191{
192 struct nla_bitfield32 flags;
193
194 if (qopt_flags && flags_attr) {
195 NL_SET_ERR_MSG_MOD(extack, "flags should be passed either through qopt, or through a dedicated attribute");
196 return -EINVAL;
197 }
198
199 if (flags_attr) {
200 flags = nla_get_bitfield32(flags_attr);
201 } else {
202 flags.selector = historic_mask;
203 flags.value = qopt_flags & historic_mask;
204 }
205
206 *p_flags = flags;
207 *p_userbits = qopt_flags & ~historic_mask;
208 return 0;
209}
210
211static inline int red_validate_flags(unsigned char flags,
212 struct netlink_ext_ack *extack)
213{
Petr Machata0a7fad22020-03-13 01:10:57 +0200214 if ((flags & TC_RED_NODROP) && !(flags & TC_RED_ECN)) {
215 NL_SET_ERR_MSG_MOD(extack, "nodrop mode is only meaningful with ECN");
216 return -EINVAL;
217 }
218
Petr Machata14bc1752020-03-13 01:10:56 +0200219 return 0;
220}
221
Thomas Grafa7834742005-11-05 21:14:03 +0100222static inline void red_set_parms(struct red_parms *p,
223 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
Eric Dumazeta73ed262011-12-09 02:46:45 +0000224 u8 Scell_log, u8 *stab, u32 max_P)
Thomas Grafa7834742005-11-05 21:14:03 +0100225{
Eric Dumazet8af2a212011-12-08 06:06:03 +0000226 int delta = qth_max - qth_min;
Eric Dumazeta73ed262011-12-09 02:46:45 +0000227 u32 max_p_delta;
Eric Dumazet8af2a212011-12-08 06:06:03 +0000228
Thomas Grafa7834742005-11-05 21:14:03 +0100229 p->qth_min = qth_min << Wlog;
230 p->qth_max = qth_max << Wlog;
231 p->Wlog = Wlog;
232 p->Plog = Plog;
Nogah Frankel5c472202017-12-04 13:31:10 +0200233 if (delta <= 0)
Eric Dumazet8af2a212011-12-08 06:06:03 +0000234 delta = 1;
235 p->qth_delta = delta;
Eric Dumazeta73ed262011-12-09 02:46:45 +0000236 if (!max_P) {
237 max_P = red_maxp(Plog);
238 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */
239 }
240 p->max_P = max_P;
241 max_p_delta = max_P / delta;
242 max_p_delta = max(max_p_delta, 1U);
243 p->max_P_reciprocal = reciprocal_value(max_p_delta);
Eric Dumazet8af2a212011-12-08 06:06:03 +0000244
245 /* RED Adaptative target :
246 * [min_th + 0.4*(min_th - max_th),
247 * min_th + 0.6*(min_th - max_th)].
248 */
249 delta /= 5;
250 p->target_min = qth_min + 2*delta;
251 p->target_max = qth_min + 3*delta;
252
Thomas Grafa7834742005-11-05 21:14:03 +0100253 p->Scell_log = Scell_log;
254 p->Scell_max = (255 << Scell_log);
255
Eric Dumazetddecf0f2012-01-06 06:31:44 +0000256 if (stab)
257 memcpy(p->Stab, stab, sizeof(p->Stab));
Thomas Grafa7834742005-11-05 21:14:03 +0100258}
259
Eric Dumazeteeca6682012-01-05 02:25:16 +0000260static inline int red_is_idling(const struct red_vars *v)
Thomas Grafa7834742005-11-05 21:14:03 +0100261{
Thomas Gleixner2456e852016-12-25 11:38:40 +0100262 return v->qidlestart != 0;
Thomas Grafa7834742005-11-05 21:14:03 +0100263}
264
Eric Dumazeteeca6682012-01-05 02:25:16 +0000265static inline void red_start_of_idle_period(struct red_vars *v)
Thomas Grafa7834742005-11-05 21:14:03 +0100266{
Eric Dumazeteeca6682012-01-05 02:25:16 +0000267 v->qidlestart = ktime_get();
Thomas Grafa7834742005-11-05 21:14:03 +0100268}
269
Eric Dumazeteeca6682012-01-05 02:25:16 +0000270static inline void red_end_of_idle_period(struct red_vars *v)
Thomas Grafa7834742005-11-05 21:14:03 +0100271{
Thomas Gleixner2456e852016-12-25 11:38:40 +0100272 v->qidlestart = 0;
Thomas Grafa7834742005-11-05 21:14:03 +0100273}
274
Eric Dumazeteeca6682012-01-05 02:25:16 +0000275static inline void red_restart(struct red_vars *v)
Thomas Grafa7834742005-11-05 21:14:03 +0100276{
Eric Dumazeteeca6682012-01-05 02:25:16 +0000277 red_end_of_idle_period(v);
278 v->qavg = 0;
279 v->qcount = -1;
Thomas Grafa7834742005-11-05 21:14:03 +0100280}
281
Eric Dumazeteeca6682012-01-05 02:25:16 +0000282static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p,
283 const struct red_vars *v)
Thomas Grafa7834742005-11-05 21:14:03 +0100284{
Eric Dumazeteeca6682012-01-05 02:25:16 +0000285 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart);
Eric Dumazetea6a5d32011-11-30 12:10:53 +0000286 long us_idle = min_t(s64, delta, p->Scell_max);
Thomas Grafa7834742005-11-05 21:14:03 +0100287 int shift;
288
Thomas Grafa7834742005-11-05 21:14:03 +0100289 /*
290 * The problem: ideally, average length queue recalcultion should
291 * be done over constant clock intervals. This is too expensive, so
292 * that the calculation is driven by outgoing packets.
293 * When the queue is idle we have to model this clock by hand.
294 *
295 * SF+VJ proposed to "generate":
296 *
297 * m = idletime / (average_pkt_size / bandwidth)
298 *
299 * dummy packets as a burst after idle time, i.e.
300 *
David Ward4362aaf2012-04-16 03:17:22 +0000301 * v->qavg *= (1-W)^m
Thomas Grafa7834742005-11-05 21:14:03 +0100302 *
303 * This is an apparently overcomplicated solution (f.e. we have to
304 * precompute a table to make this calculation in reasonable time)
305 * I believe that a simpler model may be used here,
306 * but it is field for experiments.
307 */
308
309 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
310
311 if (shift)
Eric Dumazeteeca6682012-01-05 02:25:16 +0000312 return v->qavg >> shift;
Thomas Grafa7834742005-11-05 21:14:03 +0100313 else {
314 /* Approximate initial part of exponent with linear function:
315 *
316 * (1-W)^m ~= 1-mW + ...
317 *
318 * Seems, it is the best solution to
319 * problem of too coarse exponent tabulation.
320 */
Eric Dumazeteeca6682012-01-05 02:25:16 +0000321 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log;
Thomas Grafa7834742005-11-05 21:14:03 +0100322
Eric Dumazeteeca6682012-01-05 02:25:16 +0000323 if (us_idle < (v->qavg >> 1))
324 return v->qavg - us_idle;
Thomas Grafa7834742005-11-05 21:14:03 +0100325 else
Eric Dumazeteeca6682012-01-05 02:25:16 +0000326 return v->qavg >> 1;
Thomas Grafa7834742005-11-05 21:14:03 +0100327 }
328}
329
Eric Dumazet8af2a212011-12-08 06:06:03 +0000330static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p,
Eric Dumazeteeca6682012-01-05 02:25:16 +0000331 const struct red_vars *v,
Thomas Grafa7834742005-11-05 21:14:03 +0100332 unsigned int backlog)
333{
334 /*
David Ward4362aaf2012-04-16 03:17:22 +0000335 * NOTE: v->qavg is fixed point number with point at Wlog.
Thomas Grafa7834742005-11-05 21:14:03 +0100336 * The formula below is equvalent to floating point
337 * version:
338 *
339 * qavg = qavg*(1-W) + backlog*W;
340 *
341 * --ANK (980924)
342 */
Eric Dumazeteeca6682012-01-05 02:25:16 +0000343 return v->qavg + (backlog - (v->qavg >> p->Wlog));
Thomas Grafa7834742005-11-05 21:14:03 +0100344}
345
Eric Dumazet8af2a212011-12-08 06:06:03 +0000346static inline unsigned long red_calc_qavg(const struct red_parms *p,
Eric Dumazeteeca6682012-01-05 02:25:16 +0000347 const struct red_vars *v,
Thomas Grafa7834742005-11-05 21:14:03 +0100348 unsigned int backlog)
349{
Eric Dumazeteeca6682012-01-05 02:25:16 +0000350 if (!red_is_idling(v))
351 return red_calc_qavg_no_idle_time(p, v, backlog);
Thomas Grafa7834742005-11-05 21:14:03 +0100352 else
Eric Dumazeteeca6682012-01-05 02:25:16 +0000353 return red_calc_qavg_from_idle_time(p, v);
Thomas Grafa7834742005-11-05 21:14:03 +0100354}
355
Eric Dumazet8af2a212011-12-08 06:06:03 +0000356
357static inline u32 red_random(const struct red_parms *p)
Thomas Grafa7834742005-11-05 21:14:03 +0100358{
Aruna-Hewapathirane63862b52014-01-11 07:15:59 -0500359 return reciprocal_divide(prandom_u32(), p->max_P_reciprocal);
Thomas Grafa7834742005-11-05 21:14:03 +0100360}
361
Eric Dumazeteeca6682012-01-05 02:25:16 +0000362static inline int red_mark_probability(const struct red_parms *p,
363 const struct red_vars *v,
364 unsigned long qavg)
Thomas Grafa7834742005-11-05 21:14:03 +0100365{
366 /* The formula used below causes questions.
367
Eric Dumazet8af2a212011-12-08 06:06:03 +0000368 OK. qR is random number in the interval
369 (0..1/max_P)*(qth_max-qth_min)
Thomas Grafa7834742005-11-05 21:14:03 +0100370 i.e. 0..(2^Plog). If we used floating point
371 arithmetics, it would be: (2^Plog)*rnd_num,
372 where rnd_num is less 1.
373
374 Taking into account, that qavg have fixed
Eric Dumazet8af2a212011-12-08 06:06:03 +0000375 point at Wlog, two lines
Thomas Grafa7834742005-11-05 21:14:03 +0100376 below have the following floating point equivalent:
377
378 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
379
380 Any questions? --ANK (980924)
381 */
Eric Dumazeteeca6682012-01-05 02:25:16 +0000382 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR);
Thomas Grafa7834742005-11-05 21:14:03 +0100383}
384
385enum {
386 RED_BELOW_MIN_THRESH,
387 RED_BETWEEN_TRESH,
388 RED_ABOVE_MAX_TRESH,
389};
390
Eric Dumazeteeca6682012-01-05 02:25:16 +0000391static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg)
Thomas Grafa7834742005-11-05 21:14:03 +0100392{
393 if (qavg < p->qth_min)
394 return RED_BELOW_MIN_THRESH;
395 else if (qavg >= p->qth_max)
396 return RED_ABOVE_MAX_TRESH;
397 else
398 return RED_BETWEEN_TRESH;
399}
400
401enum {
402 RED_DONT_MARK,
403 RED_PROB_MARK,
404 RED_HARD_MARK,
405};
406
Eric Dumazeteeca6682012-01-05 02:25:16 +0000407static inline int red_action(const struct red_parms *p,
408 struct red_vars *v,
409 unsigned long qavg)
Thomas Grafa7834742005-11-05 21:14:03 +0100410{
411 switch (red_cmp_thresh(p, qavg)) {
412 case RED_BELOW_MIN_THRESH:
Eric Dumazeteeca6682012-01-05 02:25:16 +0000413 v->qcount = -1;
Thomas Grafa7834742005-11-05 21:14:03 +0100414 return RED_DONT_MARK;
415
416 case RED_BETWEEN_TRESH:
Eric Dumazeteeca6682012-01-05 02:25:16 +0000417 if (++v->qcount) {
418 if (red_mark_probability(p, v, qavg)) {
419 v->qcount = 0;
420 v->qR = red_random(p);
Thomas Grafa7834742005-11-05 21:14:03 +0100421 return RED_PROB_MARK;
422 }
423 } else
Eric Dumazeteeca6682012-01-05 02:25:16 +0000424 v->qR = red_random(p);
Thomas Grafa7834742005-11-05 21:14:03 +0100425
426 return RED_DONT_MARK;
427
428 case RED_ABOVE_MAX_TRESH:
Eric Dumazeteeca6682012-01-05 02:25:16 +0000429 v->qcount = -1;
Thomas Grafa7834742005-11-05 21:14:03 +0100430 return RED_HARD_MARK;
431 }
432
433 BUG();
434 return RED_DONT_MARK;
435}
436
Eric Dumazeteeca6682012-01-05 02:25:16 +0000437static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v)
Eric Dumazet8af2a212011-12-08 06:06:03 +0000438{
439 unsigned long qavg;
440 u32 max_p_delta;
441
Eric Dumazeteeca6682012-01-05 02:25:16 +0000442 qavg = v->qavg;
443 if (red_is_idling(v))
444 qavg = red_calc_qavg_from_idle_time(p, v);
Eric Dumazet8af2a212011-12-08 06:06:03 +0000445
David Ward4362aaf2012-04-16 03:17:22 +0000446 /* v->qavg is fixed point number with point at Wlog */
Eric Dumazet8af2a212011-12-08 06:06:03 +0000447 qavg >>= p->Wlog;
448
449 if (qavg > p->target_max && p->max_P <= MAX_P_MAX)
450 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */
451 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN)
452 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */
453
454 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta);
Eric Dumazeta73ed262011-12-09 02:46:45 +0000455 max_p_delta = max(max_p_delta, 1U);
Eric Dumazet8af2a212011-12-08 06:06:03 +0000456 p->max_P_reciprocal = reciprocal_value(max_p_delta);
457}
Thomas Grafa7834742005-11-05 21:14:03 +0100458#endif