1 #ifndef __NET_SCHED_CODEL_H
2 #define __NET_SCHED_CODEL_H
5 * Codel - The Controlled-Delay Active Queue Management algorithm
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 Eric Dumazet <edumazet@google.com>
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
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.
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.
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
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 #include <linux/reciprocal_div.h>
51 /* Controlling Queue Delay (CoDel) algorithm
52 * =========================================
53 * Source : Kathleen Nichols and Van Jacobson
54 * http://queue.acm.org/detail.cfm?id=2209336
56 * Implemented on linux by Dave Taht and Eric Dumazet
60 /* CoDel uses a 1024 nsec clock, encoded in u32
61 * This gives a range of 2199 seconds, because of signed compares
63 typedef u32 codel_time_t
;
64 typedef s32 codel_tdiff_t
;
65 #define CODEL_SHIFT 10
66 #define MS2TIME(a) ((a * NSEC_PER_MSEC) >> CODEL_SHIFT)
68 static inline codel_time_t
codel_get_time(void)
70 u64 ns
= ktime_to_ns(ktime_get());
72 return ns
>> CODEL_SHIFT
;
75 /* Dealing with timer wrapping, according to RFC 1982, as desc in wikipedia:
76 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
77 * codel_time_after(a,b) returns true if the time a is after time b.
79 #define codel_time_after(a, b) \
80 (typecheck(codel_time_t, a) && \
81 typecheck(codel_time_t, b) && \
82 ((s32)((a) - (b)) > 0))
83 #define codel_time_before(a, b) codel_time_after(b, a)
85 #define codel_time_after_eq(a, b) \
86 (typecheck(codel_time_t, a) && \
87 typecheck(codel_time_t, b) && \
88 ((s32)((a) - (b)) >= 0))
89 #define codel_time_before_eq(a, b) codel_time_after_eq(b, a)
91 /* Qdiscs using codel plugin must use codel_skb_cb in their own cb[] */
93 codel_time_t enqueue_time
;
96 static struct codel_skb_cb
*get_codel_cb(const struct sk_buff
*skb
)
98 qdisc_cb_private_validate(skb
, sizeof(struct codel_skb_cb
));
99 return (struct codel_skb_cb
*)qdisc_skb_cb(skb
)->data
;
102 static codel_time_t
codel_get_enqueue_time(const struct sk_buff
*skb
)
104 return get_codel_cb(skb
)->enqueue_time
;
107 static void codel_set_enqueue_time(struct sk_buff
*skb
)
109 get_codel_cb(skb
)->enqueue_time
= codel_get_time();
112 static inline u32
codel_time_to_us(codel_time_t val
)
114 u64 valns
= ((u64
)val
<< CODEL_SHIFT
);
116 do_div(valns
, NSEC_PER_USEC
);
121 * struct codel_params - contains codel parameters
122 * @target: target queue size (in time units)
123 * @interval: width of moving time window
124 * @ecn: is Explicit Congestion Notification enabled
126 struct codel_params
{
128 codel_time_t interval
;
133 * struct codel_vars - contains codel variables
134 * @count: how many drops we've done since the last time we
135 * entered dropping state
136 * @lastcount: count at entry to dropping state
137 * @dropping: set to true if in dropping state
138 * @rec_inv_sqrt: reciprocal value of sqrt(count) >> 1
139 * @first_above_time: when we went (or will go) continuously above target
141 * @drop_next: time to drop next packet, or when we dropped last
142 * @ldelay: sojourn time of last dequeued packet
149 codel_time_t first_above_time
;
150 codel_time_t drop_next
;
154 #define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */
155 /* needed shift to get a Q0.32 number from rec_inv_sqrt */
156 #define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS)
159 * struct codel_stats - contains codel shared variables and stats
160 * @maxpacket: largest packet we've seen so far
161 * @drop_count: temp count of dropped packets in dequeue()
162 * ecn_mark: number of packets we ECN marked instead of dropping
170 static void codel_params_init(struct codel_params
*params
)
172 params
->interval
= MS2TIME(100);
173 params
->target
= MS2TIME(5);
177 static void codel_vars_init(struct codel_vars
*vars
)
179 memset(vars
, 0, sizeof(*vars
));
182 static void codel_stats_init(struct codel_stats
*stats
)
184 stats
->maxpacket
= 256;
188 * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
189 * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
191 * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
193 static void codel_Newton_step(struct codel_vars
*vars
)
195 u32 invsqrt
= ((u32
)vars
->rec_inv_sqrt
) << REC_INV_SQRT_SHIFT
;
196 u32 invsqrt2
= ((u64
)invsqrt
* invsqrt
) >> 32;
197 u64 val
= (3LL << 32) - ((u64
)vars
->count
* invsqrt2
);
199 val
>>= 2; /* avoid overflow in following multiply */
200 val
= (val
* invsqrt
) >> (32 - 2 + 1);
202 vars
->rec_inv_sqrt
= val
>> REC_INV_SQRT_SHIFT
;
206 * CoDel control_law is t + interval/sqrt(count)
207 * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
208 * both sqrt() and divide operation.
210 static codel_time_t
codel_control_law(codel_time_t t
,
211 codel_time_t interval
,
214 return t
+ reciprocal_divide(interval
, rec_inv_sqrt
<< REC_INV_SQRT_SHIFT
);
218 static bool codel_should_drop(const struct sk_buff
*skb
,
220 struct codel_vars
*vars
,
221 struct codel_params
*params
,
222 struct codel_stats
*stats
,
228 vars
->first_above_time
= 0;
232 vars
->ldelay
= now
- codel_get_enqueue_time(skb
);
233 sch
->qstats
.backlog
-= qdisc_pkt_len(skb
);
235 if (unlikely(qdisc_pkt_len(skb
) > stats
->maxpacket
))
236 stats
->maxpacket
= qdisc_pkt_len(skb
);
238 if (codel_time_before(vars
->ldelay
, params
->target
) ||
239 sch
->qstats
.backlog
<= stats
->maxpacket
) {
240 /* went below - stay below for at least interval */
241 vars
->first_above_time
= 0;
245 if (vars
->first_above_time
== 0) {
246 /* just went above from below. If we stay above
247 * for at least interval we'll say it's ok to drop
249 vars
->first_above_time
= now
+ params
->interval
;
250 } else if (codel_time_after(now
, vars
->first_above_time
)) {
256 typedef struct sk_buff
* (*codel_skb_dequeue_t
)(struct codel_vars
*vars
,
259 static struct sk_buff
*codel_dequeue(struct Qdisc
*sch
,
260 struct codel_params
*params
,
261 struct codel_vars
*vars
,
262 struct codel_stats
*stats
,
263 codel_skb_dequeue_t dequeue_func
)
265 struct sk_buff
*skb
= dequeue_func(vars
, sch
);
270 vars
->dropping
= false;
273 now
= codel_get_time();
274 drop
= codel_should_drop(skb
, sch
, vars
, params
, stats
, now
);
275 if (vars
->dropping
) {
277 /* sojourn time below target - leave dropping state */
278 vars
->dropping
= false;
279 } else if (codel_time_after_eq(now
, vars
->drop_next
)) {
280 /* It's time for the next drop. Drop the current
281 * packet and dequeue the next. The dequeue might
282 * take us out of dropping state.
283 * If not, schedule the next drop.
284 * A large backlog might result in drop rates so high
285 * that the next drop should happen now,
286 * hence the while loop.
288 while (vars
->dropping
&&
289 codel_time_after_eq(now
, vars
->drop_next
)) {
290 vars
->count
++; /* dont care of possible wrap
291 * since there is no more divide
293 codel_Newton_step(vars
);
294 if (params
->ecn
&& INET_ECN_set_ce(skb
)) {
297 codel_control_law(vars
->drop_next
,
302 qdisc_drop(skb
, sch
);
304 skb
= dequeue_func(vars
, sch
);
305 if (!codel_should_drop(skb
, sch
,
306 vars
, params
, stats
, now
)) {
307 /* leave dropping state */
308 vars
->dropping
= false;
310 /* and schedule the next drop */
312 codel_control_law(vars
->drop_next
,
321 if (params
->ecn
&& INET_ECN_set_ce(skb
)) {
324 qdisc_drop(skb
, sch
);
327 skb
= dequeue_func(vars
, sch
);
328 drop
= codel_should_drop(skb
, sch
, vars
, params
,
331 vars
->dropping
= true;
332 /* if min went above target close to when we last went below it
333 * assume that the drop rate that controlled the queue on the
334 * last cycle is a good starting point to control it now.
336 delta
= vars
->count
- vars
->lastcount
;
338 codel_time_before(now
- vars
->drop_next
,
339 16 * params
->interval
)) {
341 /* we dont care if rec_inv_sqrt approximation
342 * is not very precise :
343 * Next Newton steps will correct it quadratically.
345 codel_Newton_step(vars
);
348 vars
->rec_inv_sqrt
= ~0U >> REC_INV_SQRT_SHIFT
;
350 vars
->lastcount
= vars
->count
;
351 vars
->drop_next
= codel_control_law(now
, params
->interval
,