xref: /freebsd/sys/net/altq/altq_red.c (revision 4d846d26)
1 /*-
2  * Copyright (C) 1997-2003
3  *	Sony Computer Science Laboratories Inc.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY SONY CSL AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL SONY CSL OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  */
27 /*-
28  * Copyright (c) 1990-1994 Regents of the University of California.
29  * All rights reserved.
30  *
31  * Redistribution and use in source and binary forms, with or without
32  * modification, are permitted provided that the following conditions
33  * are met:
34  * 1. Redistributions of source code must retain the above copyright
35  *    notice, this list of conditions and the following disclaimer.
36  * 2. Redistributions in binary form must reproduce the above copyright
37  *    notice, this list of conditions and the following disclaimer in the
38  *    documentation and/or other materials provided with the distribution.
39  * 3. All advertising materials mentioning features or use of this software
40  *    must display the following acknowledgement:
41  *	This product includes software developed by the Computer Systems
42  *	Engineering Group at Lawrence Berkeley Laboratory.
43  * 4. Neither the name of the University nor of the Laboratory may be used
44  *    to endorse or promote products derived from this software without
45  *    specific prior written permission.
46  *
47  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
48  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
49  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
50  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
51  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
52  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
53  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
54  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
55  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
56  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
57  * SUCH DAMAGE.
58  *
59  * $KAME: altq_red.c,v 1.18 2003/09/05 22:40:36 itojun Exp $
60  * $FreeBSD$
61  */
62 
63 #include "opt_altq.h"
64 #include "opt_inet.h"
65 #include "opt_inet6.h"
66 #ifdef ALTQ_RED	/* red is enabled by ALTQ_RED option in opt_altq.h */
67 
68 #include <sys/param.h>
69 #include <sys/malloc.h>
70 #include <sys/mbuf.h>
71 #include <sys/socket.h>
72 #include <sys/systm.h>
73 #include <sys/errno.h>
74 #if 1 /* ALTQ3_COMPAT */
75 #include <sys/sockio.h>
76 #include <sys/proc.h>
77 #include <sys/kernel.h>
78 #ifdef ALTQ_FLOWVALVE
79 #include <sys/queue.h>
80 #include <sys/time.h>
81 #endif
82 #endif /* ALTQ3_COMPAT */
83 
84 #include <net/if.h>
85 #include <net/if_var.h>
86 
87 #include <netinet/in.h>
88 #include <netinet/in_systm.h>
89 #include <netinet/ip.h>
90 #ifdef INET6
91 #include <netinet/ip6.h>
92 #endif
93 
94 #include <netpfil/pf/pf.h>
95 #include <netpfil/pf/pf_altq.h>
96 #include <netpfil/pf/pf_mtag.h>
97 #include <net/altq/altq.h>
98 #include <net/altq/altq_red.h>
99 
100 /*
101  * ALTQ/RED (Random Early Detection) implementation using 32-bit
102  * fixed-point calculation.
103  *
104  * written by kjc using the ns code as a reference.
105  * you can learn more about red and ns from Sally's home page at
106  * http://www-nrg.ee.lbl.gov/floyd/
107  *
108  * most of the red parameter values are fixed in this implementation
109  * to prevent fixed-point overflow/underflow.
110  * if you change the parameters, watch out for overflow/underflow!
111  *
112  * the parameters used are recommended values by Sally.
113  * the corresponding ns config looks:
114  *	q_weight=0.00195
115  *	minthresh=5 maxthresh=15 queue-size=60
116  *	linterm=30
117  *	dropmech=drop-tail
118  *	bytes=false (can't be handled by 32-bit fixed-point)
119  *	doubleq=false dqthresh=false
120  *	wait=true
121  */
122 /*
123  * alternative red parameters for a slow link.
124  *
125  * assume the queue length becomes from zero to L and keeps L, it takes
126  * N packets for q_avg to reach 63% of L.
127  * when q_weight is 0.002, N is about 500 packets.
128  * for a slow link like dial-up, 500 packets takes more than 1 minute!
129  * when q_weight is 0.008, N is about 127 packets.
130  * when q_weight is 0.016, N is about 63 packets.
131  * bursts of 50 packets are allowed for 0.002, bursts of 25 packets
132  * are allowed for 0.016.
133  * see Sally's paper for more details.
134  */
135 /* normal red parameters */
136 #define	W_WEIGHT	512	/* inverse of weight of EWMA (511/512) */
137 				/* q_weight = 0.00195 */
138 
139 /* red parameters for a slow link */
140 #define	W_WEIGHT_1	128	/* inverse of weight of EWMA (127/128) */
141 				/* q_weight = 0.0078125 */
142 
143 /* red parameters for a very slow link (e.g., dialup) */
144 #define	W_WEIGHT_2	64	/* inverse of weight of EWMA (63/64) */
145 				/* q_weight = 0.015625 */
146 
147 /* fixed-point uses 12-bit decimal places */
148 #define	FP_SHIFT	12	/* fixed-point shift */
149 
150 /* red parameters for drop probability */
151 #define	INV_P_MAX	10	/* inverse of max drop probability */
152 #define	TH_MIN		5	/* min threshold */
153 #define	TH_MAX		15	/* max threshold */
154 
155 #define	RED_LIMIT	60	/* default max queue length */
156 #define	RED_STATS		/* collect statistics */
157 
158 /*
159  * our default policy for forced-drop is drop-tail.
160  * (in altq-1.1.2 or earlier, the default was random-drop.
161  * but it makes more sense to punish the cause of the surge.)
162  * to switch to the random-drop policy, define "RED_RANDOM_DROP".
163  */
164 
165 /* default red parameter values */
166 static int default_th_min = TH_MIN;
167 static int default_th_max = TH_MAX;
168 static int default_inv_pmax = INV_P_MAX;
169 
170 /*
171  * red support routines
172  */
173 red_t *
174 red_alloc(int weight, int inv_pmax, int th_min, int th_max, int flags,
175    int pkttime)
176 {
177 	red_t	*rp;
178 	int	 w, i;
179 	int	 npkts_per_sec;
180 
181 	rp = malloc(sizeof(red_t), M_DEVBUF, M_NOWAIT | M_ZERO);
182 	if (rp == NULL)
183 		return (NULL);
184 
185 	if (weight == 0)
186 		rp->red_weight = W_WEIGHT;
187 	else
188 		rp->red_weight = weight;
189 
190 	/* allocate weight table */
191 	rp->red_wtab = wtab_alloc(rp->red_weight);
192 	if (rp->red_wtab == NULL) {
193 		free(rp, M_DEVBUF);
194 		return (NULL);
195 	}
196 
197 	rp->red_avg = 0;
198 	rp->red_idle = 1;
199 
200 	if (inv_pmax == 0)
201 		rp->red_inv_pmax = default_inv_pmax;
202 	else
203 		rp->red_inv_pmax = inv_pmax;
204 	if (th_min == 0)
205 		rp->red_thmin = default_th_min;
206 	else
207 		rp->red_thmin = th_min;
208 	if (th_max == 0)
209 		rp->red_thmax = default_th_max;
210 	else
211 		rp->red_thmax = th_max;
212 
213 	rp->red_flags = flags;
214 
215 	if (pkttime == 0)
216 		/* default packet time: 1000 bytes / 10Mbps * 8 * 1000000 */
217 		rp->red_pkttime = 800;
218 	else
219 		rp->red_pkttime = pkttime;
220 
221 	if (weight == 0) {
222 		/* when the link is very slow, adjust red parameters */
223 		npkts_per_sec = 1000000 / rp->red_pkttime;
224 		if (npkts_per_sec < 50) {
225 			/* up to about 400Kbps */
226 			rp->red_weight = W_WEIGHT_2;
227 		} else if (npkts_per_sec < 300) {
228 			/* up to about 2.4Mbps */
229 			rp->red_weight = W_WEIGHT_1;
230 		}
231 	}
232 
233 	/* calculate wshift.  weight must be power of 2 */
234 	w = rp->red_weight;
235 	for (i = 0; w > 1; i++)
236 		w = w >> 1;
237 	rp->red_wshift = i;
238 	w = 1 << rp->red_wshift;
239 	if (w != rp->red_weight) {
240 		printf("invalid weight value %d for red! use %d\n",
241 		       rp->red_weight, w);
242 		rp->red_weight = w;
243 	}
244 
245 	/*
246 	 * thmin_s and thmax_s are scaled versions of th_min and th_max
247 	 * to be compared with avg.
248 	 */
249 	rp->red_thmin_s = rp->red_thmin << (rp->red_wshift + FP_SHIFT);
250 	rp->red_thmax_s = rp->red_thmax << (rp->red_wshift + FP_SHIFT);
251 
252 	/*
253 	 * precompute probability denominator
254 	 *  probd = (2 * (TH_MAX-TH_MIN) / pmax) in fixed-point
255 	 */
256 	rp->red_probd = (2 * (rp->red_thmax - rp->red_thmin)
257 			 * rp->red_inv_pmax) << FP_SHIFT;
258 
259 	microtime(&rp->red_last);
260 	return (rp);
261 }
262 
263 void
264 red_destroy(red_t *rp)
265 {
266 	wtab_destroy(rp->red_wtab);
267 	free(rp, M_DEVBUF);
268 }
269 
270 void
271 red_getstats(red_t *rp, struct redstats *sp)
272 {
273 	sp->q_avg		= rp->red_avg >> rp->red_wshift;
274 	sp->xmit_cnt		= rp->red_stats.xmit_cnt;
275 	sp->drop_cnt		= rp->red_stats.drop_cnt;
276 	sp->drop_forced		= rp->red_stats.drop_forced;
277 	sp->drop_unforced	= rp->red_stats.drop_unforced;
278 	sp->marked_packets	= rp->red_stats.marked_packets;
279 }
280 
281 int
282 red_addq(red_t *rp, class_queue_t *q, struct mbuf *m,
283     struct altq_pktattr *pktattr)
284 {
285 	int avg, droptype;
286 	int n;
287 
288 	avg = rp->red_avg;
289 
290 	/*
291 	 * if we were idle, we pretend that n packets arrived during
292 	 * the idle period.
293 	 */
294 	if (rp->red_idle) {
295 		struct timeval now;
296 		int t;
297 
298 		rp->red_idle = 0;
299 		microtime(&now);
300 		t = (now.tv_sec - rp->red_last.tv_sec);
301 		if (t > 60) {
302 			/*
303 			 * being idle for more than 1 minute, set avg to zero.
304 			 * this prevents t from overflow.
305 			 */
306 			avg = 0;
307 		} else {
308 			t = t * 1000000 + (now.tv_usec - rp->red_last.tv_usec);
309 			n = t / rp->red_pkttime - 1;
310 
311 			/* the following line does (avg = (1 - Wq)^n * avg) */
312 			if (n > 0)
313 				avg = (avg >> FP_SHIFT) *
314 				    pow_w(rp->red_wtab, n);
315 		}
316 	}
317 
318 	/* run estimator. (note: avg is scaled by WEIGHT in fixed-point) */
319 	avg += (qlen(q) << FP_SHIFT) - (avg >> rp->red_wshift);
320 	rp->red_avg = avg;		/* save the new value */
321 
322 	/*
323 	 * red_count keeps a tally of arriving traffic that has not
324 	 * been dropped.
325 	 */
326 	rp->red_count++;
327 
328 	/* see if we drop early */
329 	droptype = DTYPE_NODROP;
330 	if (avg >= rp->red_thmin_s && qlen(q) > 1) {
331 		if (avg >= rp->red_thmax_s) {
332 			/* avg >= th_max: forced drop */
333 			droptype = DTYPE_FORCED;
334 		} else if (rp->red_old == 0) {
335 			/* first exceeds th_min */
336 			rp->red_count = 1;
337 			rp->red_old = 1;
338 		} else if (drop_early((avg - rp->red_thmin_s) >> rp->red_wshift,
339 				      rp->red_probd, rp->red_count)) {
340 			/* mark or drop by red */
341 			if ((rp->red_flags & REDF_ECN) &&
342 			    mark_ecn(m, pktattr, rp->red_flags)) {
343 				/* successfully marked.  do not drop. */
344 				rp->red_count = 0;
345 #ifdef RED_STATS
346 				rp->red_stats.marked_packets++;
347 #endif
348 			} else {
349 				/* unforced drop by red */
350 				droptype = DTYPE_EARLY;
351 			}
352 		}
353 	} else {
354 		/* avg < th_min */
355 		rp->red_old = 0;
356 	}
357 
358 	/*
359 	 * if the queue length hits the hard limit, it's a forced drop.
360 	 */
361 	if (droptype == DTYPE_NODROP && qlen(q) >= qlimit(q))
362 		droptype = DTYPE_FORCED;
363 
364 #ifdef RED_RANDOM_DROP
365 	/* if successful or forced drop, enqueue this packet. */
366 	if (droptype != DTYPE_EARLY)
367 		_addq(q, m);
368 #else
369 	/* if successful, enqueue this packet. */
370 	if (droptype == DTYPE_NODROP)
371 		_addq(q, m);
372 #endif
373 	if (droptype != DTYPE_NODROP) {
374 		if (droptype == DTYPE_EARLY) {
375 			/* drop the incoming packet */
376 #ifdef RED_STATS
377 			rp->red_stats.drop_unforced++;
378 #endif
379 		} else {
380 			/* forced drop, select a victim packet in the queue. */
381 #ifdef RED_RANDOM_DROP
382 			m = _getq_random(q);
383 #endif
384 #ifdef RED_STATS
385 			rp->red_stats.drop_forced++;
386 #endif
387 		}
388 #ifdef RED_STATS
389 		PKTCNTR_ADD(&rp->red_stats.drop_cnt, m_pktlen(m));
390 #endif
391 		rp->red_count = 0;
392 		m_freem(m);
393 		return (-1);
394 	}
395 	/* successfully queued */
396 #ifdef RED_STATS
397 	PKTCNTR_ADD(&rp->red_stats.xmit_cnt, m_pktlen(m));
398 #endif
399 	return (0);
400 }
401 
402 /*
403  * early-drop probability is calculated as follows:
404  *   prob = p_max * (avg - th_min) / (th_max - th_min)
405  *   prob_a = prob / (2 - count*prob)
406  *	    = (avg-th_min) / (2*(th_max-th_min)*inv_p_max - count*(avg-th_min))
407  * here prob_a increases as successive undrop count increases.
408  * (prob_a starts from prob/2, becomes prob when (count == (1 / prob)),
409  * becomes 1 when (count >= (2 / prob))).
410  */
411 int
412 drop_early(int fp_len, int fp_probd, int count)
413 {
414 	int	d;		/* denominator of drop-probability */
415 
416 	d = fp_probd - count * fp_len;
417 	if (d <= 0)
418 		/* count exceeds the hard limit: drop or mark */
419 		return (1);
420 
421 	/*
422 	 * now the range of d is [1..600] in fixed-point. (when
423 	 * th_max-th_min=10 and p_max=1/30)
424 	 * drop probability = (avg - TH_MIN) / d
425 	 */
426 
427 	if ((arc4random() % d) < fp_len) {
428 		/* drop or mark */
429 		return (1);
430 	}
431 	/* no drop/mark */
432 	return (0);
433 }
434 
435 /*
436  * try to mark CE bit to the packet.
437  *    returns 1 if successfully marked, 0 otherwise.
438  */
439 int
440 mark_ecn(struct mbuf *m, struct altq_pktattr *pktattr, int flags)
441 {
442 	struct mbuf	*m0;
443 	struct pf_mtag	*at;
444 	void		*hdr;
445 
446 	at = pf_find_mtag(m);
447 	if (at != NULL) {
448 		hdr = at->hdr;
449 	} else
450 		return (0);
451 
452 	/* verify that pattr_hdr is within the mbuf data */
453 	for (m0 = m; m0 != NULL; m0 = m0->m_next)
454 		if (((caddr_t)hdr >= m0->m_data) &&
455 		    ((caddr_t)hdr < m0->m_data + m0->m_len))
456 			break;
457 	if (m0 == NULL) {
458 		/* ick, tag info is stale */
459 		return (0);
460 	}
461 
462 	switch (((struct ip *)hdr)->ip_v) {
463 	case IPVERSION:
464 		if (flags & REDF_ECN4) {
465 			struct ip *ip = hdr;
466 			u_int8_t otos;
467 			int sum;
468 
469 			if (ip->ip_v != 4)
470 				return (0);	/* version mismatch! */
471 
472 			if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT)
473 				return (0);	/* not-ECT */
474 			if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE)
475 				return (1);	/* already marked */
476 
477 			/*
478 			 * ecn-capable but not marked,
479 			 * mark CE and update checksum
480 			 */
481 			otos = ip->ip_tos;
482 			ip->ip_tos |= IPTOS_ECN_CE;
483 			/*
484 			 * update checksum (from RFC1624)
485 			 *	   HC' = ~(~HC + ~m + m')
486 			 */
487 			sum = ~ntohs(ip->ip_sum) & 0xffff;
488 			sum += (~otos & 0xffff) + ip->ip_tos;
489 			sum = (sum >> 16) + (sum & 0xffff);
490 			sum += (sum >> 16);  /* add carry */
491 			ip->ip_sum = htons(~sum & 0xffff);
492 			return (1);
493 		}
494 		break;
495 #ifdef INET6
496 	case (IPV6_VERSION >> 4):
497 		if (flags & REDF_ECN6) {
498 			struct ip6_hdr *ip6 = hdr;
499 			u_int32_t flowlabel;
500 
501 			flowlabel = ntohl(ip6->ip6_flow);
502 			if ((flowlabel >> 28) != 6)
503 				return (0);	/* version mismatch! */
504 			if ((flowlabel & (IPTOS_ECN_MASK << 20)) ==
505 			    (IPTOS_ECN_NOTECT << 20))
506 				return (0);	/* not-ECT */
507 			if ((flowlabel & (IPTOS_ECN_MASK << 20)) ==
508 			    (IPTOS_ECN_CE << 20))
509 				return (1);	/* already marked */
510 			/*
511 			 * ecn-capable but not marked,  mark CE
512 			 */
513 			flowlabel |= (IPTOS_ECN_CE << 20);
514 			ip6->ip6_flow = htonl(flowlabel);
515 			return (1);
516 		}
517 		break;
518 #endif  /* INET6 */
519 	}
520 
521 	/* not marked */
522 	return (0);
523 }
524 
525 struct mbuf *
526 red_getq(red_t *rp, class_queue_t *q)
527 {
528 	struct mbuf *m;
529 
530 	if ((m = _getq(q)) == NULL) {
531 		if (rp->red_idle == 0) {
532 			rp->red_idle = 1;
533 			microtime(&rp->red_last);
534 		}
535 		return NULL;
536 	}
537 
538 	rp->red_idle = 0;
539 	return (m);
540 }
541 
542 /*
543  * helper routine to calibrate avg during idle.
544  * pow_w(wtab, n) returns (1 - Wq)^n in fixed-point
545  * here Wq = 1/weight and the code assumes Wq is close to zero.
546  *
547  * w_tab[n] holds ((1 - Wq)^(2^n)) in fixed-point.
548  */
549 static struct wtab *wtab_list = NULL;	/* pointer to wtab list */
550 
551 struct wtab *
552 wtab_alloc(int weight)
553 {
554 	struct wtab	*w;
555 	int		 i;
556 
557 	for (w = wtab_list; w != NULL; w = w->w_next)
558 		if (w->w_weight == weight) {
559 			w->w_refcount++;
560 			return (w);
561 		}
562 
563 	w = malloc(sizeof(struct wtab), M_DEVBUF, M_NOWAIT | M_ZERO);
564 	if (w == NULL)
565 		return (NULL);
566 	w->w_weight = weight;
567 	w->w_refcount = 1;
568 	w->w_next = wtab_list;
569 	wtab_list = w;
570 
571 	/* initialize the weight table */
572 	w->w_tab[0] = ((weight - 1) << FP_SHIFT) / weight;
573 	for (i = 1; i < 32; i++) {
574 		w->w_tab[i] = (w->w_tab[i-1] * w->w_tab[i-1]) >> FP_SHIFT;
575 		if (w->w_tab[i] == 0 && w->w_param_max == 0)
576 			w->w_param_max = 1 << i;
577 	}
578 
579 	return (w);
580 }
581 
582 int
583 wtab_destroy(struct wtab *w)
584 {
585 	struct wtab	*prev;
586 
587 	if (--w->w_refcount > 0)
588 		return (0);
589 
590 	if (wtab_list == w)
591 		wtab_list = w->w_next;
592 	else for (prev = wtab_list; prev->w_next != NULL; prev = prev->w_next)
593 		if (prev->w_next == w) {
594 			prev->w_next = w->w_next;
595 			break;
596 		}
597 
598 	free(w, M_DEVBUF);
599 	return (0);
600 }
601 
602 int32_t
603 pow_w(struct wtab *w, int n)
604 {
605 	int	i, bit;
606 	int32_t	val;
607 
608 	if (n >= w->w_param_max)
609 		return (0);
610 
611 	val = 1 << FP_SHIFT;
612 	if (n <= 0)
613 		return (val);
614 
615 	bit = 1;
616 	i = 0;
617 	while (n) {
618 		if (n & bit) {
619 			val = (val * w->w_tab[i]) >> FP_SHIFT;
620 			n &= ~bit;
621 		}
622 		i++;
623 		bit <<=  1;
624 	}
625 	return (val);
626 }
627 
628 #endif /* ALTQ_RED */
629