Merge tag 'perf-core-for-mingo-5.1-20190321' of git://git.kernel.org/pub/scm/linux...
[sfrench/cifs-2.6.git] / include / net / codel_impl.h
1 #ifndef __NET_SCHED_CODEL_IMPL_H
2 #define __NET_SCHED_CODEL_IMPL_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>
10  *  Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
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 /* Controlling Queue Delay (CoDel) algorithm
45  * =========================================
46  * Source : Kathleen Nichols and Van Jacobson
47  * http://queue.acm.org/detail.cfm?id=2209336
48  *
49  * Implemented on linux by Dave Taht and Eric Dumazet
50  */
51
52 static void codel_params_init(struct codel_params *params)
53 {
54         params->interval = MS2TIME(100);
55         params->target = MS2TIME(5);
56         params->ce_threshold = CODEL_DISABLED_THRESHOLD;
57         params->ecn = false;
58 }
59
60 static void codel_vars_init(struct codel_vars *vars)
61 {
62         memset(vars, 0, sizeof(*vars));
63 }
64
65 static void codel_stats_init(struct codel_stats *stats)
66 {
67         stats->maxpacket = 0;
68 }
69
70 /*
71  * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
72  * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
73  *
74  * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
75  */
76 static void codel_Newton_step(struct codel_vars *vars)
77 {
78         u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
79         u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
80         u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
81
82         val >>= 2; /* avoid overflow in following multiply */
83         val = (val * invsqrt) >> (32 - 2 + 1);
84
85         vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
86 }
87
88 /*
89  * CoDel control_law is t + interval/sqrt(count)
90  * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
91  * both sqrt() and divide operation.
92  */
93 static codel_time_t codel_control_law(codel_time_t t,
94                                       codel_time_t interval,
95                                       u32 rec_inv_sqrt)
96 {
97         return t + reciprocal_scale(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
98 }
99
100 static bool codel_should_drop(const struct sk_buff *skb,
101                               void *ctx,
102                               struct codel_vars *vars,
103                               struct codel_params *params,
104                               struct codel_stats *stats,
105                               codel_skb_len_t skb_len_func,
106                               codel_skb_time_t skb_time_func,
107                               u32 *backlog,
108                               codel_time_t now)
109 {
110         bool ok_to_drop;
111         u32 skb_len;
112
113         if (!skb) {
114                 vars->first_above_time = 0;
115                 return false;
116         }
117
118         skb_len = skb_len_func(skb);
119         vars->ldelay = now - skb_time_func(skb);
120
121         if (unlikely(skb_len > stats->maxpacket))
122                 stats->maxpacket = skb_len;
123
124         if (codel_time_before(vars->ldelay, params->target) ||
125             *backlog <= params->mtu) {
126                 /* went below - stay below for at least interval */
127                 vars->first_above_time = 0;
128                 return false;
129         }
130         ok_to_drop = false;
131         if (vars->first_above_time == 0) {
132                 /* just went above from below. If we stay above
133                  * for at least interval we'll say it's ok to drop
134                  */
135                 vars->first_above_time = now + params->interval;
136         } else if (codel_time_after(now, vars->first_above_time)) {
137                 ok_to_drop = true;
138         }
139         return ok_to_drop;
140 }
141
142 static struct sk_buff *codel_dequeue(void *ctx,
143                                      u32 *backlog,
144                                      struct codel_params *params,
145                                      struct codel_vars *vars,
146                                      struct codel_stats *stats,
147                                      codel_skb_len_t skb_len_func,
148                                      codel_skb_time_t skb_time_func,
149                                      codel_skb_drop_t drop_func,
150                                      codel_skb_dequeue_t dequeue_func)
151 {
152         struct sk_buff *skb = dequeue_func(vars, ctx);
153         codel_time_t now;
154         bool drop;
155
156         if (!skb) {
157                 vars->dropping = false;
158                 return skb;
159         }
160         now = codel_get_time();
161         drop = codel_should_drop(skb, ctx, vars, params, stats,
162                                  skb_len_func, skb_time_func, backlog, now);
163         if (vars->dropping) {
164                 if (!drop) {
165                         /* sojourn time below target - leave dropping state */
166                         vars->dropping = false;
167                 } else if (codel_time_after_eq(now, vars->drop_next)) {
168                         /* It's time for the next drop. Drop the current
169                          * packet and dequeue the next. The dequeue might
170                          * take us out of dropping state.
171                          * If not, schedule the next drop.
172                          * A large backlog might result in drop rates so high
173                          * that the next drop should happen now,
174                          * hence the while loop.
175                          */
176                         while (vars->dropping &&
177                                codel_time_after_eq(now, vars->drop_next)) {
178                                 vars->count++; /* dont care of possible wrap
179                                                 * since there is no more divide
180                                                 */
181                                 codel_Newton_step(vars);
182                                 if (params->ecn && INET_ECN_set_ce(skb)) {
183                                         stats->ecn_mark++;
184                                         vars->drop_next =
185                                                 codel_control_law(vars->drop_next,
186                                                                   params->interval,
187                                                                   vars->rec_inv_sqrt);
188                                         goto end;
189                                 }
190                                 stats->drop_len += skb_len_func(skb);
191                                 drop_func(skb, ctx);
192                                 stats->drop_count++;
193                                 skb = dequeue_func(vars, ctx);
194                                 if (!codel_should_drop(skb, ctx,
195                                                        vars, params, stats,
196                                                        skb_len_func,
197                                                        skb_time_func,
198                                                        backlog, now)) {
199                                         /* leave dropping state */
200                                         vars->dropping = false;
201                                 } else {
202                                         /* and schedule the next drop */
203                                         vars->drop_next =
204                                                 codel_control_law(vars->drop_next,
205                                                                   params->interval,
206                                                                   vars->rec_inv_sqrt);
207                                 }
208                         }
209                 }
210         } else if (drop) {
211                 u32 delta;
212
213                 if (params->ecn && INET_ECN_set_ce(skb)) {
214                         stats->ecn_mark++;
215                 } else {
216                         stats->drop_len += skb_len_func(skb);
217                         drop_func(skb, ctx);
218                         stats->drop_count++;
219
220                         skb = dequeue_func(vars, ctx);
221                         drop = codel_should_drop(skb, ctx, vars, params,
222                                                  stats, skb_len_func,
223                                                  skb_time_func, backlog, now);
224                 }
225                 vars->dropping = true;
226                 /* if min went above target close to when we last went below it
227                  * assume that the drop rate that controlled the queue on the
228                  * last cycle is a good starting point to control it now.
229                  */
230                 delta = vars->count - vars->lastcount;
231                 if (delta > 1 &&
232                     codel_time_before(now - vars->drop_next,
233                                       16 * params->interval)) {
234                         vars->count = delta;
235                         /* we dont care if rec_inv_sqrt approximation
236                          * is not very precise :
237                          * Next Newton steps will correct it quadratically.
238                          */
239                         codel_Newton_step(vars);
240                 } else {
241                         vars->count = 1;
242                         vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
243                 }
244                 vars->lastcount = vars->count;
245                 vars->drop_next = codel_control_law(now, params->interval,
246                                                     vars->rec_inv_sqrt);
247         }
248 end:
249         if (skb && codel_time_after(vars->ldelay, params->ce_threshold) &&
250             INET_ECN_set_ce(skb))
251                 stats->ce_mark++;
252         return skb;
253 }
254
255 #endif