Commit | Line | Data |
---|---|---|
76e3cc12 ED |
1 | #ifndef __NET_SCHED_CODEL_H |
2 | #define __NET_SCHED_CODEL_H | |
3 | ||
4 | /* | |
5 | * Codel - The Controlled-Delay Active Queue Management algorithm | |
6 | * | |
7 | * Copyright (C) 2011-2012 Kathleen Nichols <nichols@pollere.com> | |
8 | * Copyright (C) 2011-2012 Van Jacobson <van@pollere.net> | |
9 | * Copyright (C) 2012 Michael D. Taht <dave.taht@bufferbloat.net> | |
80ba92fa | 10 | * Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com> |
76e3cc12 ED |
11 | * |
12 | * Redistribution and use in source and binary forms, with or without | |
13 | * modification, are permitted provided that the following conditions | |
14 | * are met: | |
15 | * 1. Redistributions of source code must retain the above copyright | |
16 | * notice, this list of conditions, and the following disclaimer, | |
17 | * without modification. | |
18 | * 2. Redistributions in binary form must reproduce the above copyright | |
19 | * notice, this list of conditions and the following disclaimer in the | |
20 | * documentation and/or other materials provided with the distribution. | |
21 | * 3. The names of the authors may not be used to endorse or promote products | |
22 | * derived from this software without specific prior written permission. | |
23 | * | |
24 | * Alternatively, provided that this notice is retained in full, this | |
25 | * software may be distributed under the terms of the GNU General | |
26 | * Public License ("GPL") version 2, in which case the provisions of the | |
27 | * GPL apply INSTEAD OF those given above. | |
28 | * | |
29 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
30 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
31 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
32 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
33 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
34 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
35 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
36 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
37 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
38 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
39 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH | |
40 | * DAMAGE. | |
41 | * | |
42 | */ | |
43 | ||
44 | #include <linux/types.h> | |
45 | #include <linux/ktime.h> | |
46 | #include <linux/skbuff.h> | |
47 | #include <net/pkt_sched.h> | |
48 | #include <net/inet_ecn.h> | |
49 | ||
50 | /* Controlling Queue Delay (CoDel) algorithm | |
51 | * ========================================= | |
52 | * Source : Kathleen Nichols and Van Jacobson | |
53 | * http://queue.acm.org/detail.cfm?id=2209336 | |
54 | * | |
55 | * Implemented on linux by Dave Taht and Eric Dumazet | |
56 | */ | |
57 | ||
58 | ||
59 | /* CoDel uses a 1024 nsec clock, encoded in u32 | |
60 | * This gives a range of 2199 seconds, because of signed compares | |
61 | */ | |
62 | typedef u32 codel_time_t; | |
63 | typedef s32 codel_tdiff_t; | |
64 | #define CODEL_SHIFT 10 | |
65 | #define MS2TIME(a) ((a * NSEC_PER_MSEC) >> CODEL_SHIFT) | |
66 | ||
67 | static inline codel_time_t codel_get_time(void) | |
68 | { | |
d2de875c | 69 | u64 ns = ktime_get_ns(); |
76e3cc12 ED |
70 | |
71 | return ns >> CODEL_SHIFT; | |
72 | } | |
73 | ||
1ba3aab3 JDB |
74 | /* Dealing with timer wrapping, according to RFC 1982, as desc in wikipedia: |
75 | * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution | |
76 | * codel_time_after(a,b) returns true if the time a is after time b. | |
77 | */ | |
78 | #define codel_time_after(a, b) \ | |
79 | (typecheck(codel_time_t, a) && \ | |
80 | typecheck(codel_time_t, b) && \ | |
81 | ((s32)((a) - (b)) > 0)) | |
82 | #define codel_time_before(a, b) codel_time_after(b, a) | |
83 | ||
84 | #define codel_time_after_eq(a, b) \ | |
85 | (typecheck(codel_time_t, a) && \ | |
86 | typecheck(codel_time_t, b) && \ | |
87 | ((s32)((a) - (b)) >= 0)) | |
88 | #define codel_time_before_eq(a, b) codel_time_after_eq(b, a) | |
76e3cc12 ED |
89 | |
90 | /* Qdiscs using codel plugin must use codel_skb_cb in their own cb[] */ | |
91 | struct codel_skb_cb { | |
92 | codel_time_t enqueue_time; | |
93 | }; | |
94 | ||
95 | static struct codel_skb_cb *get_codel_cb(const struct sk_buff *skb) | |
96 | { | |
97 | qdisc_cb_private_validate(skb, sizeof(struct codel_skb_cb)); | |
98 | return (struct codel_skb_cb *)qdisc_skb_cb(skb)->data; | |
99 | } | |
100 | ||
101 | static codel_time_t codel_get_enqueue_time(const struct sk_buff *skb) | |
102 | { | |
103 | return get_codel_cb(skb)->enqueue_time; | |
104 | } | |
105 | ||
106 | static void codel_set_enqueue_time(struct sk_buff *skb) | |
107 | { | |
108 | get_codel_cb(skb)->enqueue_time = codel_get_time(); | |
109 | } | |
110 | ||
111 | static inline u32 codel_time_to_us(codel_time_t val) | |
112 | { | |
113 | u64 valns = ((u64)val << CODEL_SHIFT); | |
114 | ||
115 | do_div(valns, NSEC_PER_USEC); | |
116 | return (u32)valns; | |
117 | } | |
118 | ||
119 | /** | |
120 | * struct codel_params - contains codel parameters | |
121 | * @target: target queue size (in time units) | |
80ba92fa | 122 | * @ce_threshold: threshold for marking packets with ECN CE |
76e3cc12 | 123 | * @interval: width of moving time window |
a5d28090 | 124 | * @mtu: device mtu, or minimal queue backlog in bytes. |
76e3cc12 ED |
125 | * @ecn: is Explicit Congestion Notification enabled |
126 | */ | |
127 | struct codel_params { | |
128 | codel_time_t target; | |
80ba92fa | 129 | codel_time_t ce_threshold; |
76e3cc12 | 130 | codel_time_t interval; |
a5d28090 | 131 | u32 mtu; |
76e3cc12 ED |
132 | bool ecn; |
133 | }; | |
134 | ||
135 | /** | |
136 | * struct codel_vars - contains codel variables | |
137 | * @count: how many drops we've done since the last time we | |
138 | * entered dropping state | |
139 | * @lastcount: count at entry to dropping state | |
140 | * @dropping: set to true if in dropping state | |
536edd67 | 141 | * @rec_inv_sqrt: reciprocal value of sqrt(count) >> 1 |
76e3cc12 ED |
142 | * @first_above_time: when we went (or will go) continuously above target |
143 | * for interval | |
144 | * @drop_next: time to drop next packet, or when we dropped last | |
145 | * @ldelay: sojourn time of last dequeued packet | |
146 | */ | |
147 | struct codel_vars { | |
148 | u32 count; | |
149 | u32 lastcount; | |
6ff272c9 ED |
150 | bool dropping; |
151 | u16 rec_inv_sqrt; | |
76e3cc12 ED |
152 | codel_time_t first_above_time; |
153 | codel_time_t drop_next; | |
154 | codel_time_t ldelay; | |
155 | }; | |
156 | ||
6ff272c9 ED |
157 | #define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */ |
158 | /* needed shift to get a Q0.32 number from rec_inv_sqrt */ | |
159 | #define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS) | |
160 | ||
76e3cc12 ED |
161 | /** |
162 | * struct codel_stats - contains codel shared variables and stats | |
163 | * @maxpacket: largest packet we've seen so far | |
164 | * @drop_count: temp count of dropped packets in dequeue() | |
2ccccf5f | 165 | * @drop_len: bytes of dropped packets in dequeue() |
76e3cc12 | 166 | * ecn_mark: number of packets we ECN marked instead of dropping |
80ba92fa | 167 | * ce_mark: number of packets CE marked because sojourn time was above ce_threshold |
76e3cc12 ED |
168 | */ |
169 | struct codel_stats { | |
170 | u32 maxpacket; | |
171 | u32 drop_count; | |
2ccccf5f | 172 | u32 drop_len; |
76e3cc12 | 173 | u32 ecn_mark; |
80ba92fa | 174 | u32 ce_mark; |
76e3cc12 ED |
175 | }; |
176 | ||
80ba92fa ED |
177 | #define CODEL_DISABLED_THRESHOLD INT_MAX |
178 | ||
a5d28090 ED |
179 | static void codel_params_init(struct codel_params *params, |
180 | const struct Qdisc *sch) | |
76e3cc12 ED |
181 | { |
182 | params->interval = MS2TIME(100); | |
183 | params->target = MS2TIME(5); | |
a5d28090 | 184 | params->mtu = psched_mtu(qdisc_dev(sch)); |
80ba92fa | 185 | params->ce_threshold = CODEL_DISABLED_THRESHOLD; |
76e3cc12 ED |
186 | params->ecn = false; |
187 | } | |
188 | ||
189 | static void codel_vars_init(struct codel_vars *vars) | |
190 | { | |
536edd67 | 191 | memset(vars, 0, sizeof(*vars)); |
76e3cc12 ED |
192 | } |
193 | ||
194 | static void codel_stats_init(struct codel_stats *stats) | |
195 | { | |
a5d28090 | 196 | stats->maxpacket = 0; |
76e3cc12 ED |
197 | } |
198 | ||
536edd67 ED |
199 | /* |
200 | * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots | |
201 | * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2) | |
202 | * | |
6ff272c9 | 203 | * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32 |
76e3cc12 | 204 | */ |
536edd67 | 205 | static void codel_Newton_step(struct codel_vars *vars) |
76e3cc12 | 206 | { |
6ff272c9 ED |
207 | u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT; |
208 | u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32; | |
209 | u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2); | |
76e3cc12 | 210 | |
6ff272c9 ED |
211 | val >>= 2; /* avoid overflow in following multiply */ |
212 | val = (val * invsqrt) >> (32 - 2 + 1); | |
76e3cc12 | 213 | |
6ff272c9 | 214 | vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT; |
76e3cc12 ED |
215 | } |
216 | ||
536edd67 ED |
217 | /* |
218 | * CoDel control_law is t + interval/sqrt(count) | |
219 | * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid | |
220 | * both sqrt() and divide operation. | |
221 | */ | |
76e3cc12 ED |
222 | static codel_time_t codel_control_law(codel_time_t t, |
223 | codel_time_t interval, | |
536edd67 | 224 | u32 rec_inv_sqrt) |
76e3cc12 | 225 | { |
89770b0a | 226 | return t + reciprocal_scale(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT); |
76e3cc12 ED |
227 | } |
228 | ||
536edd67 | 229 | static bool codel_should_drop(const struct sk_buff *skb, |
865ec552 | 230 | struct Qdisc *sch, |
76e3cc12 ED |
231 | struct codel_vars *vars, |
232 | struct codel_params *params, | |
233 | struct codel_stats *stats, | |
234 | codel_time_t now) | |
235 | { | |
236 | bool ok_to_drop; | |
237 | ||
238 | if (!skb) { | |
239 | vars->first_above_time = 0; | |
240 | return false; | |
241 | } | |
242 | ||
243 | vars->ldelay = now - codel_get_enqueue_time(skb); | |
865ec552 | 244 | sch->qstats.backlog -= qdisc_pkt_len(skb); |
76e3cc12 ED |
245 | |
246 | if (unlikely(qdisc_pkt_len(skb) > stats->maxpacket)) | |
247 | stats->maxpacket = qdisc_pkt_len(skb); | |
248 | ||
249 | if (codel_time_before(vars->ldelay, params->target) || | |
a5d28090 | 250 | sch->qstats.backlog <= params->mtu) { |
76e3cc12 ED |
251 | /* went below - stay below for at least interval */ |
252 | vars->first_above_time = 0; | |
253 | return false; | |
254 | } | |
255 | ok_to_drop = false; | |
256 | if (vars->first_above_time == 0) { | |
257 | /* just went above from below. If we stay above | |
258 | * for at least interval we'll say it's ok to drop | |
259 | */ | |
260 | vars->first_above_time = now + params->interval; | |
261 | } else if (codel_time_after(now, vars->first_above_time)) { | |
262 | ok_to_drop = true; | |
263 | } | |
264 | return ok_to_drop; | |
265 | } | |
266 | ||
267 | typedef struct sk_buff * (*codel_skb_dequeue_t)(struct codel_vars *vars, | |
268 | struct Qdisc *sch); | |
269 | ||
270 | static struct sk_buff *codel_dequeue(struct Qdisc *sch, | |
271 | struct codel_params *params, | |
272 | struct codel_vars *vars, | |
273 | struct codel_stats *stats, | |
865ec552 | 274 | codel_skb_dequeue_t dequeue_func) |
76e3cc12 ED |
275 | { |
276 | struct sk_buff *skb = dequeue_func(vars, sch); | |
277 | codel_time_t now; | |
278 | bool drop; | |
279 | ||
280 | if (!skb) { | |
281 | vars->dropping = false; | |
282 | return skb; | |
283 | } | |
284 | now = codel_get_time(); | |
865ec552 | 285 | drop = codel_should_drop(skb, sch, vars, params, stats, now); |
76e3cc12 ED |
286 | if (vars->dropping) { |
287 | if (!drop) { | |
288 | /* sojourn time below target - leave dropping state */ | |
289 | vars->dropping = false; | |
290 | } else if (codel_time_after_eq(now, vars->drop_next)) { | |
291 | /* It's time for the next drop. Drop the current | |
292 | * packet and dequeue the next. The dequeue might | |
293 | * take us out of dropping state. | |
294 | * If not, schedule the next drop. | |
295 | * A large backlog might result in drop rates so high | |
296 | * that the next drop should happen now, | |
297 | * hence the while loop. | |
298 | */ | |
299 | while (vars->dropping && | |
300 | codel_time_after_eq(now, vars->drop_next)) { | |
536edd67 ED |
301 | vars->count++; /* dont care of possible wrap |
302 | * since there is no more divide | |
303 | */ | |
304 | codel_Newton_step(vars); | |
76e3cc12 ED |
305 | if (params->ecn && INET_ECN_set_ce(skb)) { |
306 | stats->ecn_mark++; | |
307 | vars->drop_next = | |
308 | codel_control_law(vars->drop_next, | |
309 | params->interval, | |
536edd67 | 310 | vars->rec_inv_sqrt); |
76e3cc12 ED |
311 | goto end; |
312 | } | |
2ccccf5f | 313 | stats->drop_len += qdisc_pkt_len(skb); |
76e3cc12 ED |
314 | qdisc_drop(skb, sch); |
315 | stats->drop_count++; | |
316 | skb = dequeue_func(vars, sch); | |
865ec552 | 317 | if (!codel_should_drop(skb, sch, |
76e3cc12 ED |
318 | vars, params, stats, now)) { |
319 | /* leave dropping state */ | |
320 | vars->dropping = false; | |
321 | } else { | |
322 | /* and schedule the next drop */ | |
323 | vars->drop_next = | |
324 | codel_control_law(vars->drop_next, | |
325 | params->interval, | |
536edd67 | 326 | vars->rec_inv_sqrt); |
76e3cc12 ED |
327 | } |
328 | } | |
329 | } | |
330 | } else if (drop) { | |
2359a476 ED |
331 | u32 delta; |
332 | ||
76e3cc12 ED |
333 | if (params->ecn && INET_ECN_set_ce(skb)) { |
334 | stats->ecn_mark++; | |
335 | } else { | |
2ccccf5f | 336 | stats->drop_len += qdisc_pkt_len(skb); |
76e3cc12 ED |
337 | qdisc_drop(skb, sch); |
338 | stats->drop_count++; | |
339 | ||
340 | skb = dequeue_func(vars, sch); | |
865ec552 | 341 | drop = codel_should_drop(skb, sch, vars, params, |
76e3cc12 ED |
342 | stats, now); |
343 | } | |
344 | vars->dropping = true; | |
345 | /* if min went above target close to when we last went below it | |
346 | * assume that the drop rate that controlled the queue on the | |
347 | * last cycle is a good starting point to control it now. | |
348 | */ | |
2359a476 ED |
349 | delta = vars->count - vars->lastcount; |
350 | if (delta > 1 && | |
351 | codel_time_before(now - vars->drop_next, | |
76e3cc12 | 352 | 16 * params->interval)) { |
2359a476 | 353 | vars->count = delta; |
536edd67 ED |
354 | /* we dont care if rec_inv_sqrt approximation |
355 | * is not very precise : | |
356 | * Next Newton steps will correct it quadratically. | |
357 | */ | |
358 | codel_Newton_step(vars); | |
76e3cc12 ED |
359 | } else { |
360 | vars->count = 1; | |
6ff272c9 | 361 | vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT; |
76e3cc12 ED |
362 | } |
363 | vars->lastcount = vars->count; | |
364 | vars->drop_next = codel_control_law(now, params->interval, | |
536edd67 | 365 | vars->rec_inv_sqrt); |
76e3cc12 ED |
366 | } |
367 | end: | |
80ba92fa ED |
368 | if (skb && codel_time_after(vars->ldelay, params->ce_threshold) && |
369 | INET_ECN_set_ce(skb)) | |
370 | stats->ce_mark++; | |
76e3cc12 ED |
371 | return skb; |
372 | } | |
373 | #endif |