1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef __NET_SCHED_RED_H
3 #define __NET_SCHED_RED_H
4
5 #include <linux/types.h>
6 #include <linux/bug.h>
7 #include <net/pkt_sched.h>
8 #include <net/inet_ecn.h>
9 #include <net/dsfield.h>
10 #include <linux/reciprocal_div.h>
11
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
93 /*
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
116 #define RED_STAB_SIZE 256
117 #define RED_STAB_MASK (RED_STAB_SIZE - 1)
118
119 struct red_stats {
120 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 };
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
154 static inline u32 red_maxp(u8 Plog)
155 {
156 return Plog < 32 ? (~0U >> Plog) : ~0U;
157 }
158
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
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
191 static inline int red_get_flags(unsigned char qopt_flags,
192 unsigned char historic_mask,
193 struct nlattr *flags_attr,
194 unsigned char supported_mask,
195 struct nla_bitfield32 *p_flags,
196 unsigned char *p_userbits,
197 struct netlink_ext_ack *extack)
198 {
199 struct nla_bitfield32 flags;
200
201 if (qopt_flags && flags_attr) {
202 NL_SET_ERR_MSG_MOD(extack, "flags should be passed either through qopt, or through a dedicated attribute");
203 return -EINVAL;
204 }
205
206 if (flags_attr) {
207 flags = nla_get_bitfield32(flags_attr);
208 } else {
209 flags.selector = historic_mask;
210 flags.value = qopt_flags & historic_mask;
211 }
212
213 *p_flags = flags;
214 *p_userbits = qopt_flags & ~historic_mask;
215 return 0;
216 }
217
218 static inline int red_validate_flags(unsigned char flags,
219 struct netlink_ext_ack *extack)
220 {
221 if ((flags & TC_RED_NODROP) && !(flags & TC_RED_ECN)) {
222 NL_SET_ERR_MSG_MOD(extack, "nodrop mode is only meaningful with ECN");
223 return -EINVAL;
224 }
225
226 return 0;
227 }
228
229 static inline void red_set_parms(struct red_parms *p,
230 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog,
231 u8 Scell_log, u8 *stab, u32 max_P)
232 {
233 int delta = qth_max - qth_min;
234 u32 max_p_delta;
235
236 WRITE_ONCE(p->qth_min, qth_min << Wlog);
237 WRITE_ONCE(p->qth_max, qth_max << Wlog);
238 WRITE_ONCE(p->Wlog, Wlog);
239 WRITE_ONCE(p->Plog, Plog);
240 if (delta <= 0)
241 delta = 1;
242 p->qth_delta = delta;
243 if (!max_P) {
244 max_P = red_maxp(Plog);
245 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */
246 }
247 WRITE_ONCE(p->max_P, max_P);
248 max_p_delta = max_P / delta;
249 max_p_delta = max(max_p_delta, 1U);
250 p->max_P_reciprocal = reciprocal_value(max_p_delta);
251
252 /* RED Adaptative target :
253 * [min_th + 0.4*(min_th - max_th),
254 * min_th + 0.6*(min_th - max_th)].
255 */
256 delta /= 5;
257 p->target_min = qth_min + 2*delta;
258 p->target_max = qth_min + 3*delta;
259
260 WRITE_ONCE(p->Scell_log, Scell_log);
261 p->Scell_max = (255 << Scell_log);
262
263 if (stab)
264 memcpy(p->Stab, stab, sizeof(p->Stab));
265 }
266
267 static inline int red_is_idling(const struct red_vars *v)
268 {
269 return v->qidlestart != 0;
270 }
271
272 static inline void red_start_of_idle_period(struct red_vars *v)
273 {
274 v->qidlestart = ktime_get();
275 }
276
277 static inline void red_end_of_idle_period(struct red_vars *v)
278 {
279 v->qidlestart = 0;
280 }
281
282 static inline void red_restart(struct red_vars *v)
283 {
284 red_end_of_idle_period(v);
285 v->qavg = 0;
286 v->qcount = -1;
287 }
288
289 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p,
290 const struct red_vars *v)
291 {
292 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart);
293 long us_idle = min_t(s64, delta, p->Scell_max);
294 int shift;
295
296 /*
297 * The problem: ideally, average length queue recalculation should
298 * be done over constant clock intervals. This is too expensive, so
299 * that the calculation is driven by outgoing packets.
300 * When the queue is idle we have to model this clock by hand.
301 *
302 * SF+VJ proposed to "generate":
303 *
304 * m = idletime / (average_pkt_size / bandwidth)
305 *
306 * dummy packets as a burst after idle time, i.e.
307 *
308 * v->qavg *= (1-W)^m
309 *
310 * This is an apparently overcomplicated solution (f.e. we have to
311 * precompute a table to make this calculation in reasonable time)
312 * I believe that a simpler model may be used here,
313 * but it is field for experiments.
314 */
315
316 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK];
317
318 if (shift)
319 return v->qavg >> shift;
320 else {
321 /* Approximate initial part of exponent with linear function:
322 *
323 * (1-W)^m ~= 1-mW + ...
324 *
325 * Seems, it is the best solution to
326 * problem of too coarse exponent tabulation.
327 */
328 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log;
329
330 if (us_idle < (v->qavg >> 1))
331 return v->qavg - us_idle;
332 else
333 return v->qavg >> 1;
334 }
335 }
336
337 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p,
338 const struct red_vars *v,
339 unsigned int backlog)
340 {
341 /*
342 * NOTE: v->qavg is fixed point number with point at Wlog.
343 * The formula below is equvalent to floating point
344 * version:
345 *
346 * qavg = qavg*(1-W) + backlog*W;
347 *
348 * --ANK (980924)
349 */
350 return v->qavg + (backlog - (v->qavg >> p->Wlog));
351 }
352
353 static inline unsigned long red_calc_qavg(const struct red_parms *p,
354 const struct red_vars *v,
355 unsigned int backlog)
356 {
357 if (!red_is_idling(v))
358 return red_calc_qavg_no_idle_time(p, v, backlog);
359 else
360 return red_calc_qavg_from_idle_time(p, v);
361 }
362
363
364 static inline u32 red_random(const struct red_parms *p)
365 {
366 return reciprocal_divide(get_random_u32(), p->max_P_reciprocal);
367 }
368
369 static inline int red_mark_probability(const struct red_parms *p,
370 const struct red_vars *v,
371 unsigned long qavg)
372 {
373 /* The formula used below causes questions.
374
375 OK. qR is random number in the interval
376 (0..1/max_P)*(qth_max-qth_min)
377 i.e. 0..(2^Plog). If we used floating point
378 arithmetics, it would be: (2^Plog)*rnd_num,
379 where rnd_num is less 1.
380
381 Taking into account, that qavg have fixed
382 point at Wlog, two lines
383 below have the following floating point equivalent:
384
385 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount
386
387 Any questions? --ANK (980924)
388 */
389 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR);
390 }
391
392 enum {
393 RED_BELOW_MIN_THRESH,
394 RED_BETWEEN_TRESH,
395 RED_ABOVE_MAX_TRESH,
396 };
397
398 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg)
399 {
400 if (qavg < p->qth_min)
401 return RED_BELOW_MIN_THRESH;
402 else if (qavg >= p->qth_max)
403 return RED_ABOVE_MAX_TRESH;
404 else
405 return RED_BETWEEN_TRESH;
406 }
407
408 enum {
409 RED_DONT_MARK,
410 RED_PROB_MARK,
411 RED_HARD_MARK,
412 };
413
414 static inline int red_action(const struct red_parms *p,
415 struct red_vars *v,
416 unsigned long qavg)
417 {
418 switch (red_cmp_thresh(p, qavg)) {
419 case RED_BELOW_MIN_THRESH:
420 v->qcount = -1;
421 return RED_DONT_MARK;
422
423 case RED_BETWEEN_TRESH:
424 if (++v->qcount) {
425 if (red_mark_probability(p, v, qavg)) {
426 v->qcount = 0;
427 v->qR = red_random(p);
428 return RED_PROB_MARK;
429 }
430 } else
431 v->qR = red_random(p);
432
433 return RED_DONT_MARK;
434
435 case RED_ABOVE_MAX_TRESH:
436 v->qcount = -1;
437 return RED_HARD_MARK;
438 }
439
440 BUG();
441 return RED_DONT_MARK;
442 }
443
444 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v)
445 {
446 unsigned long qavg;
447 u32 max_p_delta;
448
449 qavg = v->qavg;
450 if (red_is_idling(v))
451 qavg = red_calc_qavg_from_idle_time(p, v);
452
453 /* v->qavg is fixed point number with point at Wlog */
454 qavg >>= p->Wlog;
455
456 if (qavg > p->target_max && p->max_P <= MAX_P_MAX)
457 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */
458 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN)
459 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */
460
461 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta);
462 max_p_delta = max(max_p_delta, 1U);
463 p->max_P_reciprocal = reciprocal_value(max_p_delta);
464 }
465 #endif
466
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.