include/net/codel_impl.h

0001 #ifndef __NET_SCHED_CODEL_IMPL_H
0002 #define __NET_SCHED_CODEL_IMPL_H
0003
0004 /*
0005  * Codel - The Controlled-Delay Active Queue Management algorithm
0006  *
0007  *  Copyright (C) 2011-2012 Kathleen Nichols <nichols@pollere.com>
0008  *  Copyright (C) 2011-2012 Van Jacobson <van@pollere.net>
0009  *  Copyright (C) 2012 Michael D. Taht <dave.taht@bufferbloat.net>
0010  *  Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
0011  *
0012  * Redistribution and use in source and binary forms, with or without
0013  * modification, are permitted provided that the following conditions
0014  * are met:
0015  * 1. Redistributions of source code must retain the above copyright
0016  *    notice, this list of conditions, and the following disclaimer,
0017  *    without modification.
0018  * 2. Redistributions in binary form must reproduce the above copyright
0019  *    notice, this list of conditions and the following disclaimer in the
0020  *    documentation and/or other materials provided with the distribution.
0021  * 3. The names of the authors may not be used to endorse or promote products
0022  *    derived from this software without specific prior written permission.
0023  *
0024  * Alternatively, provided that this notice is retained in full, this
0025  * software may be distributed under the terms of the GNU General
0026  * Public License ("GPL") version 2, in which case the provisions of the
0027  * GPL apply INSTEAD OF those given above.
0028  *
0029  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
0030  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
0031  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
0032  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
0033  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
0034  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
0035  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
0036  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
0037  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
0038  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
0039  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
0040  * DAMAGE.
0041  *
0042  */
0043
0044 /* Controlling Queue Delay (CoDel) algorithm
0045  * =========================================
0046  * Source : Kathleen Nichols and Van Jacobson
0047  * http://queue.acm.org/detail.cfm?id=2209336
0048  *
0049  * Implemented on linux by Dave Taht and Eric Dumazet
0050  */
0051
0052 #include <net/inet_ecn.h>
0053
0054 static void codel_params_init(struct codel_params *params)
0055 {
0056     params->interval = MS2TIME(100);
0057     params->target = MS2TIME(5);
0058     params->ce_threshold = CODEL_DISABLED_THRESHOLD;
0059     params->ce_threshold_mask = 0;
0060     params->ce_threshold_selector = 0;
0061     params->ecn = false;
0062 }
0063
0064 static void codel_vars_init(struct codel_vars *vars)
0065 {
0066     memset(vars, 0, sizeof(*vars));
0067 }
0068
0069 static void codel_stats_init(struct codel_stats *stats)
0070 {
0071     stats->maxpacket = 0;
0072 }
0073
0074 /*
0075  * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
0076  * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
0077  *
0078  * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
0079  */
0080 static void codel_Newton_step(struct codel_vars *vars)
0081 {
0082     u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
0083     u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
0084     u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
0085
0086     val >>= 2; /* avoid overflow in following multiply */
0087     val = (val * invsqrt) >> (32 - 2 + 1);
0088
0089     vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
0090 }
0091
0092 /*
0093  * CoDel control_law is t + interval/sqrt(count)
0094  * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
0095  * both sqrt() and divide operation.
0096  */
0097 static codel_time_t codel_control_law(codel_time_t t,
0098                       codel_time_t interval,
0099                       u32 rec_inv_sqrt)
0100 {
0101     return t + reciprocal_scale(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
0102 }
0103
0104 static bool codel_should_drop(const struct sk_buff *skb,
0105                   void *ctx,
0106                   struct codel_vars *vars,
0107                   struct codel_params *params,
0108                   struct codel_stats *stats,
0109                   codel_skb_len_t skb_len_func,
0110                   codel_skb_time_t skb_time_func,
0111                   u32 *backlog,
0112                   codel_time_t now)
0113 {
0114     bool ok_to_drop;
0115     u32 skb_len;
0116
0117     if (!skb) {
0118         vars->first_above_time = 0;
0119         return false;
0120     }
0121
0122     skb_len = skb_len_func(skb);
0123     vars->ldelay = now - skb_time_func(skb);
0124
0125     if (unlikely(skb_len > stats->maxpacket))
0126         stats->maxpacket = skb_len;
0127
0128     if (codel_time_before(vars->ldelay, params->target) ||
0129         *backlog <= params->mtu) {
0130         /* went below - stay below for at least interval */
0131         vars->first_above_time = 0;
0132         return false;
0133     }
0134     ok_to_drop = false;
0135     if (vars->first_above_time == 0) {
0136         /* just went above from below. If we stay above
0137          * for at least interval we'll say it's ok to drop
0138          */
0139         vars->first_above_time = now + params->interval;
0140     } else if (codel_time_after(now, vars->first_above_time)) {
0141         ok_to_drop = true;
0142     }
0143     return ok_to_drop;
0144 }
0145
0146 static struct sk_buff *codel_dequeue(void *ctx,
0147                      u32 *backlog,
0148                      struct codel_params *params,
0149                      struct codel_vars *vars,
0150                      struct codel_stats *stats,
0151                      codel_skb_len_t skb_len_func,
0152                      codel_skb_time_t skb_time_func,
0153                      codel_skb_drop_t drop_func,
0154                      codel_skb_dequeue_t dequeue_func)
0155 {
0156     struct sk_buff *skb = dequeue_func(vars, ctx);
0157     codel_time_t now;
0158     bool drop;
0159
0160     if (!skb) {
0161         vars->dropping = false;
0162         return skb;
0163     }
0164     now = codel_get_time();
0165     drop = codel_should_drop(skb, ctx, vars, params, stats,
0166                  skb_len_func, skb_time_func, backlog, now);
0167     if (vars->dropping) {
0168         if (!drop) {
0169             /* sojourn time below target - leave dropping state */
0170             vars->dropping = false;
0171         } else if (codel_time_after_eq(now, vars->drop_next)) {
0172             /* It's time for the next drop. Drop the current
0173              * packet and dequeue the next. The dequeue might
0174              * take us out of dropping state.
0175              * If not, schedule the next drop.
0176              * A large backlog might result in drop rates so high
0177              * that the next drop should happen now,
0178              * hence the while loop.
0179              */
0180             while (vars->dropping &&
0181                    codel_time_after_eq(now, vars->drop_next)) {
0182                 vars->count++; /* dont care of possible wrap
0183                         * since there is no more divide
0184                         */
0185                 codel_Newton_step(vars);
0186                 if (params->ecn && INET_ECN_set_ce(skb)) {
0187                     stats->ecn_mark++;
0188                     vars->drop_next =
0189                         codel_control_law(vars->drop_next,
0190                                   params->interval,
0191                                   vars->rec_inv_sqrt);
0192                     goto end;
0193                 }
0194                 stats->drop_len += skb_len_func(skb);
0195                 drop_func(skb, ctx);
0196                 stats->drop_count++;
0197                 skb = dequeue_func(vars, ctx);
0198                 if (!codel_should_drop(skb, ctx,
0199                                vars, params, stats,
0200                                skb_len_func,
0201                                skb_time_func,
0202                                backlog, now)) {
0203                     /* leave dropping state */
0204                     vars->dropping = false;
0205                 } else {
0206                     /* and schedule the next drop */
0207                     vars->drop_next =
0208                         codel_control_law(vars->drop_next,
0209                                   params->interval,
0210                                   vars->rec_inv_sqrt);
0211                 }
0212             }
0213         }
0214     } else if (drop) {
0215         u32 delta;
0216
0217         if (params->ecn && INET_ECN_set_ce(skb)) {
0218             stats->ecn_mark++;
0219         } else {
0220             stats->drop_len += skb_len_func(skb);
0221             drop_func(skb, ctx);
0222             stats->drop_count++;
0223
0224             skb = dequeue_func(vars, ctx);
0225             drop = codel_should_drop(skb, ctx, vars, params,
0226                          stats, skb_len_func,
0227                          skb_time_func, backlog, now);
0228         }
0229         vars->dropping = true;
0230         /* if min went above target close to when we last went below it
0231          * assume that the drop rate that controlled the queue on the
0232          * last cycle is a good starting point to control it now.
0233          */
0234         delta = vars->count - vars->lastcount;
0235         if (delta > 1 &&
0236             codel_time_before(now - vars->drop_next,
0237                       16 * params->interval)) {
0238             vars->count = delta;
0239             /* we dont care if rec_inv_sqrt approximation
0240              * is not very precise :
0241              * Next Newton steps will correct it quadratically.
0242              */
0243             codel_Newton_step(vars);
0244         } else {
0245             vars->count = 1;
0246             vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
0247         }
0248         vars->lastcount = vars->count;
0249         vars->drop_next = codel_control_law(now, params->interval,
0250                             vars->rec_inv_sqrt);
0251     }
0252 end:
0253     if (skb && codel_time_after(vars->ldelay, params->ce_threshold)) {
0254         bool set_ce = true;
0255
0256         if (params->ce_threshold_mask) {
0257             int dsfield = skb_get_dsfield(skb);
0258
0259             set_ce = (dsfield >= 0 &&
0260                   (((u8)dsfield & params->ce_threshold_mask) ==
0261                    params->ce_threshold_selector));
0262         }
0263         if (set_ce && INET_ECN_set_ce(skb))
0264             stats->ce_mark++;
0265     }
0266     return skb;
0267 }
0268
0269 #endif