xref: /netbsd-src/sys/altq/altq_rmclass.c (revision aaf4ece63a859a04e37cf3a7229b5fab0157cc06)
1 /*	$NetBSD: altq_rmclass.c,v 1.11 2005/12/11 12:16:03 christos Exp $	*/
2 /*	$KAME: altq_rmclass.c,v 1.9 2000/12/14 08:12:46 thorpej Exp $	*/
3 
4 /*
5  * Copyright (c) 1991-1997 Regents of the University of California.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by the Network Research
19  *      Group at Lawrence Berkeley Laboratory.
20  * 4. Neither the name of the University nor of the Laboratory may be used
21  *    to endorse or promote products derived from this software without
22  *    specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  * LBL code modified by speer@eng.sun.com, May 1977.
37  * For questions and/or comments, please send mail to cbq@ee.lbl.gov
38  */
39 
40 #include <sys/cdefs.h>
41 __KERNEL_RCSID(0, "$NetBSD: altq_rmclass.c,v 1.11 2005/12/11 12:16:03 christos Exp $");
42 
43 #ident "@(#)rm_class.c  1.48     97/12/05 SMI"
44 
45 #if defined(__FreeBSD__) || defined(__NetBSD__)
46 #include "opt_altq.h"
47 #if (__FreeBSD__ != 2)
48 #include "opt_inet.h"
49 #ifdef __FreeBSD__
50 #include "opt_inet6.h"
51 #endif
52 #endif
53 #endif /* __FreeBSD__ || __NetBSD__ */
54 #ifdef ALTQ_CBQ	/* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
55 
56 #include <sys/param.h>
57 #include <sys/malloc.h>
58 #include <sys/mbuf.h>
59 #include <sys/socket.h>
60 #include <sys/systm.h>
61 #include <sys/errno.h>
62 #include <sys/time.h>
63 #include <sys/kernel.h>
64 
65 #include <net/if.h>
66 #include <netinet/in.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 
70 #include <altq/altq.h>
71 #include <altq/altq_rmclass.h>
72 #include <altq/altq_rmclass_debug.h>
73 #include <altq/altq_red.h>
74 #include <altq/altq_rio.h>
75 
76 #ifdef CBQ_TRACE
77 struct cbqtrace		cbqtrace_buffer[NCBQTRACE+1];
78 struct cbqtrace		*cbqtrace_ptr = NULL;
79 int			cbqtrace_count;
80 #endif
81 
82 /*
83  * Local Macros
84  */
85 
86 #define	reset_cutoff(ifd)	{ ifd->cutoff_ = RM_MAXDEPTH; }
87 
88 /*
89  * Local routines.
90  */
91 
92 static int	rmc_satisfied __P((struct rm_class *, struct timeval *));
93 static void	rmc_wrr_set_weights __P((struct rm_ifdat *));
94 static void	rmc_depth_compute __P((struct rm_class *));
95 static void	rmc_depth_recompute __P((rm_class_t *));
96 
97 static mbuf_t	*_rmc_wrr_dequeue_next __P((struct rm_ifdat *, int));
98 static mbuf_t	*_rmc_prr_dequeue_next __P((struct rm_ifdat *, int));
99 
100 static int	_rmc_addq __P((rm_class_t *, mbuf_t *));
101 static void	_rmc_dropq __P((rm_class_t *));
102 static mbuf_t	*_rmc_getq __P((rm_class_t *));
103 static mbuf_t	*_rmc_pollq __P((rm_class_t *));
104 
105 static int	rmc_under_limit __P((struct rm_class *, struct timeval *));
106 static void	rmc_tl_satisfied __P((struct rm_ifdat *, struct timeval *));
107 static void	rmc_drop_action __P((struct rm_class *));
108 static void	rmc_restart __P((struct rm_class *));
109 static void	rmc_root_overlimit __P((struct rm_class *, struct rm_class *));
110 
111 /*
112  * BORROW_OFFTIME (experimental):
113  * borrow the offtime of the class borrowing from.
114  * the reason is that when its own offtime is set, the class is unable
115  * to borrow much, especially when cutoff is taking effect.
116  * but when the borrowed class is overloaded (advidle is close to minidle),
117  * use the borrowing class's offtime to avoid overload.
118  */
119 /*
120  * ADJUST_CUTOFF (experimental):
121  * if no underlimit class is found due to cutoff, increase cutoff and
122  * retry the scheduling loop.
123  * also, don't invoke delay_actions while cutoff is taking effect,
124  * since a sleeping class won't have a chance to be scheduled in the
125  * next loop.
126  *
127  * now heuristics for setting the top-level variable (cutoff_) becomes:
128  *	1. if a packet arrives for a not-overlimit class, set cutoff
129  *	   to the depth of the class.
130  *	2. if cutoff is i, and a packet arrives for an overlimit class
131  *	   with an underlimit ancestor at a lower level than i (say j),
132  *	   then set cutoff to j.
133  *	3. at scheduling a packet, if there is no underlimit class
134  *	   due to the current cutoff level, increase cutoff by 1 and
135  *	   then try to schedule again.
136  */
137 
138 /*
139  * rm_class_t *
140  * rmc_newclass(...) - Create a new resource management class at priority
141  * 'pri' on the interface given by 'ifd'.
142  *
143  * nsecPerByte  is the data rate of the interface in nanoseconds/byte.
144  *              E.g., 800 for a 10Mb/s ethernet.  If the class gets less
145  *              than 100% of the bandwidth, this number should be the
146  *              'effective' rate for the class.  Let f be the
147  *              bandwidth fraction allocated to this class, and let
148  *              nsPerByte be the data rate of the output link in
149  *              nanoseconds/byte.  Then nsecPerByte is set to
150  *              nsPerByte / f.  E.g., 1600 (= 800 / .5)
151  *              for a class that gets 50% of an ethernet's bandwidth.
152  *
153  * action       the routine to call when the class is over limit.
154  *
155  * maxq         max allowable queue size for class (in packets).
156  *
157  * parent       parent class pointer.
158  *
159  * borrow       class to borrow from (should be either 'parent' or null).
160  *
161  * maxidle      max value allowed for class 'idle' time estimate (this
162  *              parameter determines how large an initial burst of packets
163  *              can be before overlimit action is invoked.
164  *
165  * offtime      how long 'delay' action will delay when class goes over
166  *              limit (this parameter determines the steady-state burst
167  *              size when a class is running over its limit).
168  *
169  * Maxidle and offtime have to be computed from the following:  If the
170  * average packet size is s, the bandwidth fraction allocated to this
171  * class is f, we want to allow b packet bursts, and the gain of the
172  * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
173  *
174  *   ptime = s * nsPerByte * (1 - f) / f
175  *   maxidle = ptime * (1 - g^b) / g^b
176  *   minidle = -ptime * (1 / (f - 1))
177  *   offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
178  *
179  * Operationally, it's convenient to specify maxidle & offtime in units
180  * independent of the link bandwidth so the maxidle & offtime passed to
181  * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
182  * (The constant factor is a scale factor needed to make the parameters
183  * integers.  This scaling also means that the 'unscaled' values of
184  * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
185  * not nanoseconds.)  Also note that the 'idle' filter computation keeps
186  * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
187  * maxidle also must be scaled upward by this value.  Thus, the passed
188  * values for maxidle and offtime can be computed as follows:
189  *
190  * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
191  * offtime = offtime * 8 / (1000 * nsecPerByte)
192  *
193  * When USE_HRTIME is employed, then maxidle and offtime become:
194  * 	maxidle = maxilde * (8.0 / nsecPerByte);
195  * 	offtime = offtime * (8.0 / nsecPerByte);
196  */
197 
198 struct rm_class *
199 rmc_newclass(pri, ifd, nsecPerByte, action, maxq, parent, borrow,
200 	     maxidle, minidle, offtime, pktsize, flags)
201 	int		pri;
202 	struct rm_ifdat	*ifd;
203 	u_int		nsecPerByte;
204 	void		(*action)(rm_class_t *, rm_class_t *);
205 	int		maxq;
206 	struct rm_class	*parent;
207 	struct rm_class	*borrow;
208 	u_int		maxidle;
209 	int		minidle;
210 	u_int		offtime;
211 	int		pktsize;
212 	int		flags;
213 {
214 	struct rm_class *cl;
215 	struct rm_class *peer;
216 	int s;
217 
218 	if (pri >= RM_MAXPRIO)
219 		return (NULL);
220 #ifndef ALTQ_RED
221 	if (flags & RMCF_RED) {
222 		printf("rmc_newclass: RED not configured for CBQ!\n");
223 		return (NULL);
224 	}
225 #endif
226 #ifndef ALTQ_RIO
227 	if (flags & RMCF_RIO) {
228 		printf("rmc_newclass: RIO not configured for CBQ!\n");
229 		return (NULL);
230 	}
231 #endif
232 
233 	MALLOC(cl, struct rm_class *, sizeof(struct rm_class),
234 	       M_DEVBUF, M_WAITOK);
235 	if (cl == NULL)
236 		return (NULL);
237 	(void)memset(cl, 0, sizeof(struct rm_class));
238 	CALLOUT_INIT(&cl->callout_);
239 	MALLOC(cl->q_, class_queue_t *, sizeof(class_queue_t),
240 	       M_DEVBUF, M_WAITOK);
241 	if (cl->q_ == NULL) {
242 		FREE(cl, M_DEVBUF);
243 		return (NULL);
244 	}
245 	(void)memset(cl->q_, 0, sizeof(class_queue_t));
246 
247 	/*
248 	 * Class initialization.
249 	 */
250 	cl->children_ = NULL;
251 	cl->parent_ = parent;
252 	cl->borrow_ = borrow;
253 	cl->leaf_ = 1;
254 	cl->ifdat_ = ifd;
255 	cl->pri_ = pri;
256 	cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
257 	cl->depth_ = 0;
258 	cl->qthresh_ = 0;
259 	cl->ns_per_byte_ = nsecPerByte;
260 
261 	qlimit(cl->q_) = maxq;
262 	qtype(cl->q_) = Q_DROPHEAD;
263 	qlen(cl->q_) = 0;
264 	cl->flags_ = flags;
265 
266 #if 1 /* minidle is also scaled in ALTQ */
267 	cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
268 	if (cl->minidle_ > 0)
269 		cl->minidle_ = 0;
270 #else
271 	cl->minidle_ = minidle;
272 #endif
273 	cl->maxidle_ = (maxidle * nsecPerByte) / 8;
274 	if (cl->maxidle_ == 0)
275 		cl->maxidle_ = 1;
276 #if 1 /* offtime is also scaled in ALTQ */
277 	cl->avgidle_ = cl->maxidle_;
278 	cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
279 	if (cl->offtime_ == 0)
280 		cl->offtime_ = 1;
281 #else
282 	cl->avgidle_ = 0;
283 	cl->offtime_ = (offtime * nsecPerByte) / 8;
284 #endif
285 	cl->overlimit = action;
286 
287 #ifdef ALTQ_RED
288 	if (flags & (RMCF_RED|RMCF_RIO)) {
289 		int red_flags, red_pkttime;
290 
291 		red_flags = 0;
292 		if (flags & RMCF_ECN)
293 			red_flags |= REDF_ECN;
294 		if (flags & RMCF_FLOWVALVE)
295 			red_flags |= REDF_FLOWVALVE;
296 #ifdef ALTQ_RIO
297 		if (flags & RMCF_CLEARDSCP)
298 			red_flags |= RIOF_CLEARDSCP;
299 #endif
300 		red_pkttime = nsecPerByte * pktsize  / 1000;
301 
302 		if (flags & RMCF_RED) {
303 			cl->red_ = red_alloc(0, 0, 0, 0,
304 					     red_flags, red_pkttime);
305 			if (cl->red_ != NULL)
306 				qtype(cl->q_) = Q_RED;
307 		}
308 #ifdef ALTQ_RIO
309 		else {
310 			cl->red_ = (red_t *)rio_alloc(0, NULL,
311 						      red_flags, red_pkttime);
312 			if (cl->red_ != NULL)
313 				qtype(cl->q_) = Q_RIO;
314 		}
315 #endif
316 	}
317 #endif /* ALTQ_RED */
318 
319 	/*
320 	 * put the class into the class tree
321 	 */
322 	s = splnet();
323 	if ((peer = ifd->active_[pri]) != NULL) {
324 		/* find the last class at this pri */
325 		cl->peer_ = peer;
326 		while (peer->peer_ != ifd->active_[pri])
327 			peer = peer->peer_;
328 		peer->peer_ = cl;
329 	} else {
330 		ifd->active_[pri] = cl;
331 		cl->peer_ = cl;
332 	}
333 
334 	if (cl->parent_) {
335 		cl->next_ = parent->children_;
336 		parent->children_ = cl;
337 		parent->leaf_ = 0;
338 	}
339 
340 	/*
341 	 * Compute the depth of this class and it's ancestors in the class
342 	 * hierarchy.
343 	 */
344 	rmc_depth_compute(cl);
345 
346 	/*
347 	 * If CBQ's WRR is enabled, then initailize the class WRR state.
348 	 */
349 	if (ifd->wrr_) {
350 		ifd->num_[pri]++;
351 		ifd->alloc_[pri] += cl->allotment_;
352 		rmc_wrr_set_weights(ifd);
353 	}
354 	splx(s);
355 	return (cl);
356 }
357 
358 int
359 rmc_modclass(cl, nsecPerByte, maxq, maxidle, minidle, offtime, pktsize)
360 	struct rm_class *cl;
361 	u_int		nsecPerByte;
362 	int		maxq;
363 	u_int		maxidle;
364 	int		minidle;
365 	u_int		offtime;
366 	int		pktsize;
367 {
368 	struct rm_ifdat	*ifd;
369 	u_int old_allotment;
370 	int s;
371 
372 	ifd = cl->ifdat_;
373 	old_allotment = cl->allotment_;
374 
375 	s = splnet();
376 	cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
377 	cl->qthresh_ = 0;
378 	cl->ns_per_byte_ = nsecPerByte;
379 
380 	qlimit(cl->q_) = maxq;
381 
382 #if 1 /* minidle is also scaled in ALTQ */
383 	cl->minidle_ = (minidle * nsecPerByte) / 8;
384 	if (cl->minidle_ > 0)
385 		cl->minidle_ = 0;
386 #else
387 	cl->minidle_ = minidle;
388 #endif
389 	cl->maxidle_ = (maxidle * nsecPerByte) / 8;
390 	if (cl->maxidle_ == 0)
391 		cl->maxidle_ = 1;
392 #if 1 /* offtime is also scaled in ALTQ */
393 	cl->avgidle_ = cl->maxidle_;
394 	cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
395 	if (cl->offtime_ == 0)
396 		cl->offtime_ = 1;
397 #else
398 	cl->avgidle_ = 0;
399 	cl->offtime_ = (offtime * nsecPerByte) / 8;
400 #endif
401 
402 	/*
403 	 * If CBQ's WRR is enabled, then initailize the class WRR state.
404 	 */
405 	if (ifd->wrr_) {
406 		ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
407 		rmc_wrr_set_weights(ifd);
408 	}
409 	splx(s);
410 	return (0);
411 }
412 
413 /*
414  * static void
415  * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
416  * 	the appropriate run robin weights for the CBQ weighted round robin
417  *	algorithm.
418  *
419  *	Returns: NONE
420  */
421 
422 static void
423 rmc_wrr_set_weights(ifd)
424 	struct rm_ifdat *ifd;
425 {
426 	int		i;
427 	struct rm_class	*cl, *clh;
428 
429 	for (i = 0; i < RM_MAXPRIO; i++) {
430 		/*
431 		 * This is inverted from that of the simulator to
432 		 * maintain precision.
433 		 */
434 		if (ifd->num_[i] == 0)
435 			ifd->M_[i] = 0;
436 		else
437 			ifd->M_[i] = ifd->alloc_[i] /
438 				(ifd->num_[i] * ifd->maxpkt_);
439 		/*
440 		 * Compute the weigthed allotment for each class.
441 		 * This takes the expensive div instruction out
442 		 * of the main loop for the wrr scheduling path.
443 		 * These only get recomputed when a class comes or
444 		 * goes.
445 		 */
446 		if (ifd->active_[i] != NULL) {
447 			clh = cl = ifd->active_[i];
448 			do {
449 				/* safe-guard for slow link or alloc_ == 0 */
450 				if (ifd->M_[i] == 0)
451 					cl->w_allotment_ = 0;
452 				else
453 					cl->w_allotment_ = cl->allotment_ /
454 						ifd->M_[i];
455 				cl = cl->peer_;
456 			} while ((cl != NULL) && (cl != clh));
457 		}
458 	}
459 }
460 
461 int
462 rmc_get_weight(ifd, pri)
463 	struct rm_ifdat *ifd;
464 	int pri;
465 {
466 	if ((pri >= 0) && (pri < RM_MAXPRIO))
467 		return (ifd->M_[pri]);
468 	else
469 		return (0);
470 }
471 
472 /*
473  * static void
474  * rmc_depth_compute(struct rm_class *cl) - This function computes the
475  * 	appropriate depth of class 'cl' and its ancestors.
476  *
477  *	Returns:	NONE
478  */
479 
480 static void
481 rmc_depth_compute(cl)
482 	struct rm_class *cl;
483 {
484 	rm_class_t *t = cl, *p;
485 
486 	/*
487 	 * Recompute the depth for the branch of the tree.
488 	 */
489 	while (t != NULL) {
490 		p = t->parent_;
491 		if (p && (t->depth_ >= p->depth_)) {
492 			p->depth_ = t->depth_ + 1;
493 			t = p;
494 		} else
495 			t = NULL;
496 	}
497 }
498 
499 /*
500  * static void
501  * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
502  *	the depth of the tree after a class has been deleted.
503  *
504  *	Returns: 	NONE
505  */
506 
507 static void
508 rmc_depth_recompute(rm_class_t *cl)
509 {
510 #if 1 /* ALTQ */
511 	rm_class_t	*p, *t;
512 
513 	p = cl;
514 	while (p != NULL) {
515 		if ((t = p->children_) == NULL) {
516 			p->depth_ = 0;
517 		} else {
518 			int cdepth = 0;
519 
520 			while (t != NULL) {
521 				if (t->depth_ > cdepth)
522 					cdepth = t->depth_;
523 				t = t->next_;
524 			}
525 
526 			if (p->depth_ == cdepth + 1)
527 				/* no change to this parent */
528 				return;
529 
530 			p->depth_ = cdepth + 1;
531 		}
532 
533 		p = p->parent_;
534 	}
535 #else
536 	rm_class_t	*t;
537 
538 	if (cl->depth_ >= 1) {
539 		if (cl->children_ == NULL) {
540 			cl->depth_ = 0;
541 		} else if ((t = cl->children_) != NULL) {
542 			while (t != NULL) {
543 				if (t->children_ != NULL)
544 					rmc_depth_recompute(t);
545 				t = t->next_;
546 			}
547 		} else
548 			rmc_depth_compute(cl);
549 	}
550 #endif
551 }
552 
553 /*
554  * void
555  * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
556  *	function deletes a class from the link-sharing stucture and frees
557  *	all resources associated with the class.
558  *
559  *	Returns: NONE
560  */
561 
562 void
563 rmc_delete_class(ifd, cl)
564 	struct rm_ifdat *ifd;
565 	struct rm_class *cl;
566 {
567 	struct rm_class	*p, *head, *previous;
568 	int		s;
569 
570 	ASSERT(cl->children_ == NULL);
571 
572 	if (cl->sleeping_)
573 		CALLOUT_STOP(&cl->callout_);
574 
575 	s = splnet();
576 	/*
577 	 * Free packets in the packet queue.
578 	 * XXX - this may not be a desired behavior.  Packets should be
579 	 * 		re-queued.
580 	 */
581 	rmc_dropall(cl);
582 
583 	/*
584 	 * If the class has a parent, then remove the class from the
585 	 * class from the parent's children chain.
586 	 */
587 	if (cl->parent_ != NULL) {
588 		head = cl->parent_->children_;
589 		p = previous = head;
590 		if (head->next_ == NULL) {
591 			ASSERT(head == cl);
592 			cl->parent_->children_ = NULL;
593 			cl->parent_->leaf_ = 1;
594 		} else while (p != NULL) {
595 			if (p == cl) {
596 				if (cl == head)
597 					cl->parent_->children_ = cl->next_;
598 				else
599 					previous->next_ = cl->next_;
600 				cl->next_ = NULL;
601 				p = NULL;
602 			} else {
603 				previous = p;
604 				p = p->next_;
605 			}
606 		}
607 	}
608 
609 	/*
610 	 * Delete class from class priority peer list.
611 	 */
612 	if ((p = ifd->active_[cl->pri_]) != NULL) {
613 		/*
614 		 * If there is more than one member of this priority
615 		 * level, then look for class(cl) in the priority level.
616 		 */
617 		if (p != p->peer_) {
618 			while (p->peer_ != cl)
619 				p = p->peer_;
620 			p->peer_ = cl->peer_;
621 
622 			if (ifd->active_[cl->pri_] == cl)
623 				ifd->active_[cl->pri_] = cl->peer_;
624 		} else {
625 			ASSERT(p == cl);
626 			ifd->active_[cl->pri_] = NULL;
627 		}
628 	}
629 
630 	/*
631 	 * Recompute the WRR weights.
632 	 */
633 	if (ifd->wrr_) {
634 		ifd->alloc_[cl->pri_] -= cl->allotment_;
635 		ifd->num_[cl->pri_]--;
636 		rmc_wrr_set_weights(ifd);
637 	}
638 
639 	/*
640 	 * Re-compute the depth of the tree.
641 	 */
642 #if 1 /* ALTQ */
643 	rmc_depth_recompute(cl->parent_);
644 #else
645 	rmc_depth_recompute(ifd->root_);
646 #endif
647 
648 	splx(s);
649 
650 	/*
651 	 * Free the class structure.
652 	 */
653 	if (cl->red_ != NULL) {
654 #ifdef ALTQ_RIO
655 		if (q_is_rio(cl->q_))
656 			rio_destroy((rio_t *)cl->red_);
657 #endif
658 #ifdef ALTQ_RED
659 		if (q_is_red(cl->q_))
660 			red_destroy(cl->red_);
661 #endif
662 	}
663 	FREE(cl->q_, M_DEVBUF);
664 	FREE(cl, M_DEVBUF);
665 }
666 
667 
668 /*
669  * void
670  * rmc_init(...) - Initialize the resource management data structures
671  *	associated with the output portion of interface 'ifp'.  'ifd' is
672  *	where the structures will be built (for backwards compatibility, the
673  *	structures aren't kept in the ifnet struct).  'nsecPerByte'
674  *	gives the link speed (inverse of bandwidth) in nanoseconds/byte.
675  *	'restart' is the driver-specific routine that the generic 'delay
676  *	until under limit' action will call to restart output.  `maxq'
677  *	is the queue size of the 'link' & 'default' classes.  'maxqueued'
678  *	is the maximum number of packets that the resource management
679  *	code will allow to be queued 'downstream' (this is typically 1).
680  *
681  *	Returns:	NONE
682  */
683 
684 void
685 rmc_init(ifq, ifd, nsecPerByte, restart, maxq, maxqueued, maxidle,
686 	 minidle, offtime, flags)
687 	struct ifaltq	*ifq;
688 	struct rm_ifdat *ifd;
689 	u_int	nsecPerByte;
690 	void	(*restart)(struct ifaltq *);
691 	int	maxq, maxqueued;
692 	u_int	maxidle;
693 	int	minidle;
694 	u_int	offtime;
695 	int	flags;
696 {
697 	int		i, mtu;
698 
699 	/*
700 	 * Initialize the CBQ traciing/debug facility.
701 	 */
702 	CBQTRACEINIT();
703 
704 	(void)memset(ifd, 0, sizeof (*ifd));
705 	mtu = ifq->altq_ifp->if_mtu;
706 	ifd->ifq_ = ifq;
707 	ifd->restart = restart;
708 	ifd->maxqueued_ = maxqueued;
709 	ifd->ns_per_byte_ = nsecPerByte;
710 	ifd->maxpkt_ = mtu;
711 	ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
712 	ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
713 #if 1
714 	ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
715 	if (mtu * nsecPerByte > 10 * 1000000)
716 		ifd->maxiftime_ /= 4;
717 #endif
718 
719 	reset_cutoff(ifd);
720 	CBQTRACE(rmc_init, "INIT", ifd->cutoff_);
721 
722 	/*
723 	 * Initialize the CBQ's WRR state.
724 	 */
725 	for (i = 0; i < RM_MAXPRIO; i++) {
726 		ifd->alloc_[i] = 0;
727 		ifd->M_[i] = 0;
728 		ifd->num_[i] = 0;
729 		ifd->na_[i] = 0;
730 		ifd->active_[i] = NULL;
731 	}
732 
733 	/*
734 	 * Initialize current packet state.
735 	 */
736 	ifd->qi_ = 0;
737 	ifd->qo_ = 0;
738 	for (i = 0; i < RM_MAXQUEUED; i++) {
739 		ifd->class_[i] = NULL;
740 		ifd->curlen_[i] = 0;
741 		ifd->borrowed_[i] = NULL;
742 	}
743 
744 	/*
745 	 * Create the root class of the link-sharing structure.
746 	 */
747 	if ((ifd->root_ = rmc_newclass(0, ifd,
748 				       nsecPerByte,
749 				       rmc_root_overlimit, maxq, 0, 0,
750 				       maxidle, minidle, offtime,
751 				       0, 0)) == NULL) {
752 		printf("rmc_init: root class not allocated\n");
753 		return ;
754 	}
755 	ifd->root_->depth_ = 0;
756 }
757 
758 /*
759  * void
760  * rmc_queue_packet(struct rm_class *cl, mbuf_t *m) - Add packet given by
761  *	mbuf 'm' to queue for resource class 'cl'.  This routine is called
762  *	by a driver's if_output routine.  This routine must be called with
763  *	output packet completion interrupts locked out (to avoid racing with
764  *	rmc_dequeue_next).
765  *
766  *	Returns:	0 on successful queueing
767  *			-1 when packet drop occurs
768  */
769 int
770 rmc_queue_packet(cl, m)
771 	struct rm_class *cl;
772 	mbuf_t *m;
773 {
774 	struct timeval	now;
775 	struct rm_ifdat *ifd = cl->ifdat_;
776 	int		cpri = cl->pri_;
777 	int		is_empty = qempty(cl->q_);
778 
779 	RM_GETTIME(now);
780 	if (ifd->cutoff_ > 0) {
781 		if (TV_LT(&cl->undertime_, &now)) {
782 			if (ifd->cutoff_ > cl->depth_)
783 				ifd->cutoff_ = cl->depth_;
784 			CBQTRACE(rmc_queue_packet, "ffoc", cl->depth_);
785 		}
786 #if 1 /* ALTQ */
787 		else {
788 			/*
789 			 * the class is overlimit. if the class has
790 			 * underlimit ancestors, set cutoff to the lowest
791 			 * depth among them.
792 			 */
793 			struct rm_class *borrow = cl->borrow_;
794 
795 			while (borrow != NULL &&
796 			       borrow->depth_ < ifd->cutoff_) {
797 				if (TV_LT(&borrow->undertime_, &now)) {
798 					ifd->cutoff_ = borrow->depth_;
799 					CBQTRACE(rmc_queue_packet, "ffob", ifd->cutoff_);
800 					break;
801 				}
802 				borrow = borrow->borrow_;
803 			}
804 		}
805 #else /* !ALTQ */
806 		else if ((ifd->cutoff_ > 1) && cl->borrow_) {
807 			if (TV_LT(&cl->borrow_->undertime_, &now)) {
808 				ifd->cutoff_ = cl->borrow_->depth_;
809 				CBQTRACE(rmc_queue_packet, "ffob",
810 					 cl->borrow_->depth_);
811 			}
812 		}
813 #endif /* !ALTQ */
814 	}
815 
816 	if (_rmc_addq(cl, m) < 0)
817 		/* failed */
818 		return (-1);
819 
820 	if (is_empty) {
821 		CBQTRACE(rmc_queue_packet, "ytpe", cl->stats_.handle);
822 		ifd->na_[cpri]++;
823 	}
824 
825 	if (qlen(cl->q_) > qlimit(cl->q_)) {
826 		/* note: qlimit can be set to 0 or 1 */
827 		rmc_drop_action(cl);
828 		return (-1);
829 	}
830 	return (0);
831 }
832 
833 /*
834  * void
835  * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
836  *	classes to see if there are satified.
837  */
838 
839 static void
840 rmc_tl_satisfied(ifd, now)
841 	struct rm_ifdat *ifd;
842 	struct timeval *now;
843 {
844 	int	i;
845 	rm_class_t	*p, *bp;
846 
847 	for (i = RM_MAXPRIO - 1; i >= 0; i--) {
848 		if ((bp = ifd->active_[i]) != NULL) {
849 			p = bp;
850 			do {
851 				if (!rmc_satisfied(p, now)) {
852 					ifd->cutoff_ = p->depth_;
853 					return;
854 				}
855 				p = p->peer_;
856 			} while (p != bp);
857 		}
858 	}
859 
860 	reset_cutoff(ifd);
861 }
862 
863 /*
864  * rmc_satisfied - Return 1 of the class is satisfied.  O, otherwise.
865  */
866 
867 static int
868 rmc_satisfied(cl, now)
869 	struct rm_class *cl;
870 	struct timeval *now;
871 {
872 	rm_class_t 	*p;
873 
874 	if (cl == NULL)
875 		return (1);
876 	if (TV_LT(now, &cl->undertime_))
877 		return (1);
878 	if (cl->depth_ == 0) {
879 		if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
880 			return (0);
881 		else
882 			return (1);
883 	}
884 	if (cl->children_ != NULL) {
885 		p = cl->children_;
886 		while (p != NULL) {
887 			if (!rmc_satisfied(p, now))
888 				return (0);
889 			p = p->next_;
890 		}
891 	}
892 
893 	return (1);
894 }
895 
896 /*
897  * Return 1 if class 'cl' is under limit or can borrow from a parent,
898  * 0 if overlimit.  As a side-effect, this routine will invoke the
899  * class overlimit action if the class if overlimit.
900  */
901 
902 static int
903 rmc_under_limit(cl, now)
904 	struct rm_class *cl;
905 	struct timeval *now;
906 {
907 	rm_class_t	*p = cl;
908 	rm_class_t	*top;
909 	struct rm_ifdat	*ifd = cl->ifdat_;
910 
911 	ifd->borrowed_[ifd->qi_] = NULL;
912 	/*
913 	 * If cl is the root class, then always return that it is
914 	 * underlimit.  Otherwise, check to see if the class is underlimit.
915 	 */
916 	if (cl->parent_ == NULL)
917 		return (1);
918 
919 	if (cl->sleeping_) {
920 		if (TV_LT(now, &cl->undertime_))
921 			return (0);
922 
923 		CALLOUT_STOP(&cl->callout_);
924 		cl->sleeping_ = 0;
925 		cl->undertime_.tv_sec = 0;
926 		return (1);
927 	}
928 
929 	top = NULL;
930 	while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
931 		if (((cl = cl->borrow_) == NULL) ||
932 		    (cl->depth_ > ifd->cutoff_)) {
933 #ifdef ADJUST_CUTOFF
934 			if (cl != NULL)
935 				/* cutoff is taking effect, just
936 				   return false without calling
937 				   the delay action. */
938 				return (0);
939 #endif
940 #ifdef BORROW_OFFTIME
941 			/*
942 			 * check if the class can borrow offtime too.
943 			 * borrow offtime from the top of the borrow
944 			 * chain if the top class is not overloaded.
945 			 */
946 			if (cl != NULL) {
947 				/* cutoff is taking effect, use this class as top. */
948 				top = cl;
949 				CBQTRACE(rmc_under_limit, "ffou", ifd->cutoff_);
950 			}
951 			if (top != NULL && top->avgidle_ == top->minidle_)
952 				top = NULL;
953 			p->overtime_ = *now;
954 			(p->overlimit)(p, top);
955 #else
956 			p->overtime_ = *now;
957 			(p->overlimit)(p, NULL);
958 #endif
959 			return (0);
960 		}
961 		top = cl;
962 	}
963 
964 	if (cl != p)
965 		ifd->borrowed_[ifd->qi_] = cl;
966 	return (1);
967 }
968 
969 /*
970  * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
971  *	Packet-by-packet round robin.
972  *
973  * The heart of the weigthed round-robin scheduler, which decides which
974  * class next gets to send a packet.  Highest priority first, then
975  * weighted round-robin within priorites.
976  *
977  * Each able-to-send class gets to send until its byte allocation is
978  * exhausted.  Thus, the active pointer is only changed after a class has
979  * exhausted its allocation.
980  *
981  * If the scheduler finds no class that is underlimit or able to borrow,
982  * then the first class found that had a nonzero queue and is allowed to
983  * borrow gets to send.
984  */
985 
986 static mbuf_t *
987 _rmc_wrr_dequeue_next(ifd, op)
988 	struct rm_ifdat *ifd;
989 	int op;
990 {
991 	struct rm_class	*cl = NULL, *first = NULL;
992 	u_int		deficit;
993 	int		cpri;
994 	mbuf_t		*m;
995 	struct timeval	now;
996 
997 	RM_GETTIME(now);
998 
999 	/*
1000 	 * if the driver polls the top of the queue and then removes
1001 	 * the polled packet, we must return the same packet.
1002 	 */
1003 	if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1004 		cl = ifd->pollcache_;
1005 		cpri = cl->pri_;
1006 		if (ifd->efficient_) {
1007 			/* check if this class is overlimit */
1008 			if (cl->undertime_.tv_sec != 0 &&
1009 			    rmc_under_limit(cl, &now) == 0)
1010 				first = cl;
1011 		}
1012 		ifd->pollcache_ = NULL;
1013 		goto _wrr_out;
1014 	}
1015 	else {
1016 		/* mode == ALTDQ_POLL || pollcache == NULL */
1017 		ifd->pollcache_ = NULL;
1018 		ifd->borrowed_[ifd->qi_] = NULL;
1019 	}
1020 #ifdef ADJUST_CUTOFF
1021  _again:
1022 #endif
1023 	for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1024 		if (ifd->na_[cpri] == 0)
1025 			continue;
1026 		deficit = 0;
1027 		/*
1028 		 * Loop through twice for a priority level, if some class
1029 		 * was unable to send a packet the first round because
1030 		 * of the weighted round-robin mechanism.
1031 		 * During the second loop at this level, deficit==2.
1032 		 * (This second loop is not needed if for every class,
1033 		 * "M[cl->pri_])" times "cl->allotment" is greater than
1034 		 * the byte size for the largest packet in the class.)
1035 		 */
1036  _wrr_loop:
1037 		cl = ifd->active_[cpri];
1038 		ASSERT(cl != NULL);
1039 		do {
1040 			if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
1041 				cl->bytes_alloc_ += cl->w_allotment_;
1042 			if (!qempty(cl->q_)) {
1043 				if ((cl->undertime_.tv_sec == 0) ||
1044 				    rmc_under_limit(cl, &now)) {
1045 					if (cl->bytes_alloc_ > 0 || deficit > 1)
1046 						goto _wrr_out;
1047 
1048 					/* underlimit but no alloc */
1049 					deficit = 1;
1050 #if 1
1051 					ifd->borrowed_[ifd->qi_] = NULL;
1052 #endif
1053 				}
1054 				else if (first == NULL && cl->borrow_ != NULL)
1055 					first = cl; /* borrowing candidate */
1056 			}
1057 
1058 			cl->bytes_alloc_ = 0;
1059 			cl = cl->peer_;
1060 		} while (cl != ifd->active_[cpri]);
1061 
1062 		if (deficit == 1) {
1063 			/* first loop found an underlimit class with deficit */
1064 			/* Loop on same priority level, with new deficit.  */
1065 			deficit = 2;
1066 			goto _wrr_loop;
1067 		}
1068 	}
1069 
1070 #ifdef ADJUST_CUTOFF
1071 	/*
1072 	 * no underlimit class found.  if cutoff is taking effect,
1073 	 * increase cutoff and try again.
1074 	 */
1075 	if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1076 		ifd->cutoff_++;
1077 		CBQTRACE(_rmc_wrr_dequeue_next, "ojda", ifd->cutoff_);
1078 		goto _again;
1079 	}
1080 #endif /* ADJUST_CUTOFF */
1081 	/*
1082 	 * If LINK_EFFICIENCY is turned on, then the first overlimit
1083 	 * class we encounter will send a packet if all the classes
1084 	 * of the link-sharing structure are overlimit.
1085 	 */
1086 	reset_cutoff(ifd);
1087 	CBQTRACE(_rmc_wrr_dequeue_next, "otsr", ifd->cutoff_);
1088 
1089 	if (!ifd->efficient_ || first == NULL)
1090 		return (NULL);
1091 
1092 	cl = first;
1093 	cpri = cl->pri_;
1094 #if 0	/* too time-consuming for nothing */
1095 	if (cl->sleeping_)
1096 		CALLOUT_STOP(&cl->callout_);
1097 	cl->sleeping_ = 0;
1098 	cl->undertime_.tv_sec = 0;
1099 #endif
1100 	ifd->borrowed_[ifd->qi_] = cl->borrow_;
1101 	ifd->cutoff_ = cl->borrow_->depth_;
1102 
1103 	/*
1104 	 * Deque the packet and do the book keeping...
1105 	 */
1106  _wrr_out:
1107 	if (op == ALTDQ_REMOVE) {
1108 		m = _rmc_getq(cl);
1109 		if (m == NULL)
1110 			panic("_rmc_wrr_dequeue_next");
1111 		if (qempty(cl->q_))
1112 			ifd->na_[cpri]--;
1113 
1114 		/*
1115 		 * Update class statistics and link data.
1116 		 */
1117 		if (cl->bytes_alloc_ > 0)
1118 			cl->bytes_alloc_ -= m_pktlen(m);
1119 
1120 		if ((cl->bytes_alloc_ <= 0) || first == cl)
1121 			ifd->active_[cl->pri_] = cl->peer_;
1122 		else
1123 			ifd->active_[cl->pri_] = cl;
1124 
1125 		ifd->class_[ifd->qi_] = cl;
1126 		ifd->curlen_[ifd->qi_] = m_pktlen(m);
1127 		ifd->now_[ifd->qi_] = now;
1128 		ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1129 		ifd->queued_++;
1130 	} else {
1131 		/* mode == ALTDQ_PPOLL */
1132 		m = _rmc_pollq(cl);
1133 		ifd->pollcache_ = cl;
1134 	}
1135 	return (m);
1136 }
1137 
1138 /*
1139  * Dequeue & return next packet from the highest priority class that
1140  * has a packet to send & has enough allocation to send it.  This
1141  * routine is called by a driver whenever it needs a new packet to
1142  * output.
1143  */
1144 static mbuf_t *
1145 _rmc_prr_dequeue_next(ifd, op)
1146 	struct rm_ifdat *ifd;
1147 	int op;
1148 {
1149 	mbuf_t		*m;
1150 	int		cpri;
1151 	struct rm_class	*cl, *first = NULL;
1152 	struct timeval	now;
1153 
1154 	RM_GETTIME(now);
1155 
1156 	/*
1157 	 * if the driver polls the top of the queue and then removes
1158 	 * the polled packet, we must return the same packet.
1159 	 */
1160 	if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1161 		cl = ifd->pollcache_;
1162 		cpri = cl->pri_;
1163 		ifd->pollcache_ = NULL;
1164 		goto _prr_out;
1165 	} else {
1166 		/* mode == ALTDQ_POLL || pollcache == NULL */
1167 		ifd->pollcache_ = NULL;
1168 		ifd->borrowed_[ifd->qi_] = NULL;
1169 	}
1170 #ifdef ADJUST_CUTOFF
1171  _again:
1172 #endif
1173 	for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1174 		if (ifd->na_[cpri] == 0)
1175 			continue;
1176 		cl = ifd->active_[cpri];
1177 		ASSERT(cl != NULL);
1178 		do {
1179 			if (!qempty(cl->q_)) {
1180 				if ((cl->undertime_.tv_sec == 0) ||
1181 				    rmc_under_limit(cl, &now))
1182 					goto _prr_out;
1183 				if (first == NULL && cl->borrow_ != NULL)
1184 					first = cl;
1185 			}
1186 			cl = cl->peer_;
1187 		} while (cl != ifd->active_[cpri]);
1188 	}
1189 
1190 #ifdef ADJUST_CUTOFF
1191 	/*
1192 	 * no underlimit class found.  if cutoff is taking effect, increase
1193 	 * cutoff and try again.
1194 	 */
1195 	if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1196 		ifd->cutoff_++;
1197 		goto _again;
1198 	}
1199 #endif /* ADJUST_CUTOFF */
1200 	/*
1201 	 * If LINK_EFFICIENCY is turned on, then the first overlimit
1202 	 * class we encounter will send a packet if all the classes
1203 	 * of the link-sharing structure are overlimit.
1204 	 */
1205 	reset_cutoff(ifd);
1206 	if (!ifd->efficient_ || first == NULL)
1207 		return (NULL);
1208 
1209 	cl = first;
1210 	cpri = cl->pri_;
1211 #if 0	/* too time-consuming for nothing */
1212 	if (cl->sleeping_)
1213 		CALLOUT_STOP(&cl->callout_);
1214 	cl->sleeping_ = 0;
1215 	cl->undertime_.tv_sec = 0;
1216 #endif
1217 	ifd->borrowed_[ifd->qi_] = cl->borrow_;
1218 	ifd->cutoff_ = cl->borrow_->depth_;
1219 
1220 	/*
1221 	 * Deque the packet and do the book keeping...
1222 	 */
1223  _prr_out:
1224 	if (op == ALTDQ_REMOVE) {
1225 		m = _rmc_getq(cl);
1226 		if (m == NULL)
1227 			panic("_rmc_prr_dequeue_next");
1228 		if (qempty(cl->q_))
1229 			ifd->na_[cpri]--;
1230 
1231 		ifd->active_[cpri] = cl->peer_;
1232 
1233 		ifd->class_[ifd->qi_] = cl;
1234 		ifd->curlen_[ifd->qi_] = m_pktlen(m);
1235 		ifd->now_[ifd->qi_] = now;
1236 		ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1237 		ifd->queued_++;
1238 	} else {
1239 		/* mode == ALTDQ_POLL */
1240 		m = _rmc_pollq(cl);
1241 		ifd->pollcache_ = cl;
1242 	}
1243 	return (m);
1244 }
1245 
1246 /*
1247  * mbuf_t *
1248  * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1249  *	is invoked by the packet driver to get the next packet to be
1250  *	dequeued and output on the link.  If WRR is enabled, then the
1251  *	WRR dequeue next routine will determine the next packet to sent.
1252  *	Otherwise, packet-by-packet round robin is invoked.
1253  *
1254  *	Returns:	NULL, if a packet is not available or if all
1255  *			classes are overlimit.
1256  *
1257  *			Otherwise, Pointer to the next packet.
1258  */
1259 
1260 mbuf_t *
1261 rmc_dequeue_next(ifd, mode)
1262 	struct rm_ifdat *ifd;
1263 	int mode;
1264 {
1265 	if (ifd->queued_ >= ifd->maxqueued_)
1266 		return (NULL);
1267 	else if (ifd->wrr_)
1268 		return (_rmc_wrr_dequeue_next(ifd, mode));
1269 	else
1270 		return (_rmc_prr_dequeue_next(ifd, mode));
1271 }
1272 
1273 /*
1274  * Update the utilization estimate for the packet that just completed.
1275  * The packet's class & the parent(s) of that class all get their
1276  * estimators updated.  This routine is called by the driver's output-
1277  * packet-completion interrupt service routine.
1278  */
1279 
1280 /*
1281  * a macro to approximate "divide by 1000" that gives 0.000999,
1282  * if a value has enough effective digits.
1283  * (on pentium, mul takes 9 cycles but div takes 46!)
1284  */
1285 #define	NSEC_TO_USEC(t)	(((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1286 void
1287 rmc_update_class_util(ifd)
1288 	struct rm_ifdat *ifd;
1289 {
1290 	int		idle, avgidle, pktlen;
1291 	int 		pkt_time, tidle;
1292 	rm_class_t	*cl, *borrowed;
1293 	rm_class_t	*borrows;
1294 	struct timeval	*nowp;
1295 
1296 	/*
1297 	 * Get the most recent completed class.
1298 	 */
1299 	if ((cl = ifd->class_[ifd->qo_]) == NULL)
1300 		return;
1301 
1302 	pktlen = ifd->curlen_[ifd->qo_];
1303 	borrowed = ifd->borrowed_[ifd->qo_];
1304 	borrows = borrowed;
1305 
1306 	PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1307 
1308 	/*
1309 	 * Run estimator on class and it's ancesstors.
1310 	 */
1311 	/*
1312 	 * rm_update_class_util is designed to be called when the
1313 	 * transfer is completed from a xmit complete interrupt,
1314 	 * but most drivers don't implement an upcall for that.
1315 	 * so, just use estimated completion time.
1316 	 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1317 	 */
1318 	nowp = &ifd->now_[ifd->qo_];
1319 	/* get pkt_time (for link) in usec */
1320 #if 1  /* use approximation */
1321 	pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1322 	pkt_time = NSEC_TO_USEC(pkt_time);
1323 #else
1324 	pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1325 #endif
1326 #if 1 /* ALTQ4PPP */
1327 	if (TV_LT(nowp, &ifd->ifnow_)) {
1328 		int iftime;
1329 
1330 		/*
1331 		 * make sure the estimated completion time does not go
1332 		 * too far.  it can happen when the link layer supports
1333 		 * data compression or the interface speed is set to
1334 		 * a much lower value.
1335 		 */
1336 		TV_DELTA(&ifd->ifnow_, nowp, iftime);
1337 		if (iftime+pkt_time < ifd->maxiftime_) {
1338 			TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1339 		} else {
1340 			TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1341 		}
1342 	} else {
1343 		TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1344 	}
1345 #else
1346 	if (TV_LT(nowp, &ifd->ifnow_)) {
1347 		TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1348 	} else {
1349 		TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1350 	}
1351 #endif
1352 
1353 	while (cl != NULL) {
1354 		TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1355 		if (idle >= 2000000)
1356 			/*
1357 			 * this class is idle enough, reset avgidle.
1358 			 * (TV_DELTA returns 2000000 us when delta is large.)
1359 			 */
1360 			cl->avgidle_ = cl->maxidle_;
1361 
1362 		/* get pkt_time (for class) in usec */
1363 #if 1  /* use approximation */
1364 		pkt_time = pktlen * cl->ns_per_byte_;
1365 		pkt_time = NSEC_TO_USEC(pkt_time);
1366 #else
1367 		pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1368 #endif
1369 		idle -= pkt_time;
1370 
1371 		avgidle = cl->avgidle_;
1372 		avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1373 		cl->avgidle_ = avgidle;
1374 
1375 		/* Are we overlimit ? */
1376 		if (avgidle <= 0) {
1377 			CBQTRACE(rmc_update_class_util, "milo", cl->stats_.handle);
1378 #if 1 /* ALTQ */
1379 			/*
1380 			 * need some lower bound for avgidle, otherwise
1381 			 * a borrowing class gets unbounded penalty.
1382 			 */
1383 			if (avgidle < cl->minidle_)
1384 				avgidle = cl->avgidle_ = cl->minidle_;
1385 #endif
1386 			/* set next idle to make avgidle 0 */
1387 			tidle = pkt_time +
1388 				(((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1389 			TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1390 			++cl->stats_.over;
1391 		} else {
1392 			cl->avgidle_ =
1393 			    (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1394 			cl->undertime_.tv_sec = 0;
1395 			if (cl->sleeping_) {
1396 				CALLOUT_STOP(&cl->callout_);
1397 				cl->sleeping_ = 0;
1398 			}
1399 		}
1400 
1401 		if (borrows != NULL) {
1402 			if (borrows != cl)
1403 				++cl->stats_.borrows;
1404 			else
1405 				borrows = NULL;
1406 		}
1407 		cl->last_ = ifd->ifnow_;
1408 		cl->last_pkttime_ = pkt_time;
1409 
1410 #if 1
1411 		if (cl->parent_ == NULL) {
1412 			/* take stats of root class */
1413 			PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1414 		}
1415 #endif
1416 
1417 		cl = cl->parent_;
1418 	}
1419 
1420 	/*
1421 	 * Check to see if cutoff needs to set to a new level.
1422 	 */
1423 	cl = ifd->class_[ifd->qo_];
1424 	if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1425 #if 1 /* ALTQ */
1426 		if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1427 			rmc_tl_satisfied(ifd, nowp);
1428 			CBQTRACE(rmc_update_class_util, "broe", ifd->cutoff_);
1429 		} else {
1430 			ifd->cutoff_ = borrowed->depth_;
1431 			CBQTRACE(rmc_update_class_util, "ffob", borrowed->depth_);
1432 		}
1433 #else /* !ALTQ */
1434 		if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1435 			reset_cutoff(ifd);
1436 #ifdef notdef
1437 			rmc_tl_satisfied(ifd, &now);
1438 #endif
1439 			CBQTRACE(rmc_update_class_util, "broe", ifd->cutoff_);
1440 		} else {
1441 			ifd->cutoff_ = borrowed->depth_;
1442 			CBQTRACE(rmc_update_class_util, "ffob", borrowed->depth_);
1443 		}
1444 #endif /* !ALTQ */
1445 	}
1446 
1447 	/*
1448 	 * Release class slot
1449 	 */
1450 	ifd->borrowed_[ifd->qo_] = NULL;
1451 	ifd->class_[ifd->qo_] = NULL;
1452 	ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1453 	ifd->queued_--;
1454 }
1455 
1456 /*
1457  * void
1458  * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1459  *	over-limit action routines.  These get invoked by rmc_under_limit()
1460  *	if a class with packets to send if over its bandwidth limit & can't
1461  *	borrow from a parent class.
1462  *
1463  *	Returns: NONE
1464  */
1465 
1466 static void
1467 rmc_drop_action(cl)
1468 	struct rm_class *cl;
1469 {
1470 	struct rm_ifdat	*ifd = cl->ifdat_;
1471 
1472 	ASSERT(qlen(cl->q_) > 0);
1473 	_rmc_dropq(cl);
1474 	if (qempty(cl->q_))
1475 		ifd->na_[cl->pri_]--;
1476 }
1477 
1478 void rmc_dropall(cl)
1479     struct rm_class *cl;
1480 {
1481 	struct rm_ifdat *ifd = cl->ifdat_;
1482 
1483 	if (!qempty(cl->q_)) {
1484 		_flushq(cl->q_);
1485 
1486 		ifd->na_[cl->pri_]--;
1487 	}
1488 }
1489 
1490 #if (__FreeBSD_version > 300000)
1491 /* hzto() is removed from FreeBSD-3.0 */
1492 static int hzto __P((struct timeval *));
1493 
1494 static int
1495 hzto(tv)
1496 	struct timeval *tv;
1497 {
1498 	struct timeval t2;
1499 
1500 	getmicrotime(&t2);
1501 	t2.tv_sec = tv->tv_sec - t2.tv_sec;
1502 	t2.tv_usec = tv->tv_usec - t2.tv_usec;
1503 	return (tvtohz(&t2));
1504 }
1505 #endif /* __FreeBSD_version > 300000 */
1506 
1507 /*
1508  * void
1509  * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1510  *	delay action routine.  It is invoked via rmc_under_limit when the
1511  *	packet is discoverd to be overlimit.
1512  *
1513  *	If the delay action is result of borrow class being overlimit, then
1514  *	delay for the offtime of the borrowing class that is overlimit.
1515  *
1516  *	Returns: NONE
1517  */
1518 
1519 void
1520 rmc_delay_action(cl, borrow)
1521 	struct rm_class *cl, *borrow;
1522 {
1523 	int	nowdelay, t, extradelay;
1524 
1525 	cl->stats_.overactions++;
1526 	TV_DELTA(&cl->undertime_, &cl->overtime_, nowdelay);
1527 #ifndef BORROW_OFFTIME
1528 	nowdelay += cl->offtime_;
1529 #endif
1530 
1531 	if (!cl->sleeping_) {
1532 		CBQTRACE(rmc_delay_action, "yled", cl->stats_.handle);
1533 #ifdef BORROW_OFFTIME
1534 		if (borrow != NULL)
1535 			extradelay = borrow->offtime_;
1536 		else
1537 #endif
1538 			extradelay = cl->offtime_;
1539 
1540 #ifdef ALTQ
1541 		/*
1542 		 * XXX recalculate suspend time:
1543 		 * current undertime is (tidle + pkt_time) calculated
1544 		 * from the last transmission.
1545 		 *	tidle: time required to bring avgidle back to 0
1546 		 *	pkt_time: target waiting time for this class
1547 		 * we need to replace pkt_time by offtime
1548 		 */
1549 		extradelay -= cl->last_pkttime_;
1550 #endif
1551 		if (extradelay > 0) {
1552 			TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1553 			nowdelay += extradelay;
1554 		}
1555 
1556 		cl->sleeping_ = 1;
1557 		cl->stats_.delays++;
1558 
1559 		/*
1560 		 * Since packets are phased randomly with respect to the
1561 		 * clock, 1 tick (the next clock tick) can be an arbitrarily
1562 		 * short time so we have to wait for at least two ticks.
1563 		 * NOTE:  If there's no other traffic, we need the timer as
1564 		 * a 'backstop' to restart this class.
1565 		 */
1566 		if (nowdelay > tick * 2) {
1567 #ifdef __FreeBSD__
1568 			/* FreeBSD rounds up the tick */
1569 			t = hzto(&cl->undertime_);
1570 #else
1571 			/* other BSDs round down the tick */
1572 			t = hzto(&cl->undertime_) + 1;
1573 #endif
1574 		} else
1575 			t = 2;
1576 		CALLOUT_RESET(&cl->callout_, t,
1577 			      (timeout_t *)rmc_restart, (caddr_t)cl);
1578 	}
1579 }
1580 
1581 /*
1582  * void
1583  * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1584  *	called by the system timer code & is responsible checking if the
1585  *	class is still sleeping (it might have been restarted as a side
1586  *	effect of the queue scan on a packet arrival) and, if so, restarting
1587  *	output for the class.  Inspecting the class state & restarting output
1588  *	require locking the class structure.  In general the driver is
1589  *	responsible for locking but this is the only routine that is not
1590  *	called directly or indirectly from the interface driver so it has
1591  *	know about system locking conventions.  Under bsd, locking is done
1592  *	by raising IPL to splnet so that's what's implemented here.  On a
1593  *	different system this would probably need to be changed.
1594  *
1595  *	Returns:	NONE
1596  */
1597 
1598 static void
1599 rmc_restart(cl)
1600 	struct rm_class *cl;
1601 {
1602 	struct rm_ifdat *ifd = cl->ifdat_;
1603 	int s;
1604 
1605 	s = splnet();
1606 	if (cl->sleeping_) {
1607 		cl->sleeping_ = 0;
1608 		cl->undertime_.tv_sec = 0;
1609 
1610 		if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1611 			CBQTRACE(rmc_restart, "trts", cl->stats_.handle);
1612 			(ifd->restart)(ifd->ifq_);
1613 		}
1614 	}
1615 	splx(s);
1616 }
1617 
1618 /*
1619  * void
1620  * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1621  *	handling routine for the root class of the link sharing structure.
1622  *
1623  *	Returns: NONE
1624  */
1625 
1626 static void
1627 rmc_root_overlimit(cl, borrow)
1628 	struct rm_class *cl, *borrow;
1629 {
1630     panic("rmc_root_overlimit");
1631 }
1632 
1633 /*
1634  * Packet Queue handling routines.  Eventually, this is to localize the
1635  *	effects on the code whether queues are red queues or droptail
1636  *	queues.
1637  */
1638 
1639 static int
1640 _rmc_addq(cl, m)
1641 	rm_class_t *cl;
1642 	mbuf_t *m;
1643 {
1644 #ifdef ALTQ_RIO
1645 	if (q_is_rio(cl->q_))
1646 		return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1647 #endif
1648 #ifdef ALTQ_RED
1649 	if (q_is_red(cl->q_))
1650 		return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1651 #endif /* ALTQ_RED */
1652 
1653 	if (cl->flags_ & RMCF_CLEARDSCP)
1654 		write_dsfield(m, cl->pktattr_, 0);
1655 
1656 	_addq(cl->q_, m);
1657 	return (0);
1658 }
1659 
1660 /* note: _rmc_dropq is not called for red */
1661 static void
1662 _rmc_dropq(cl)
1663 	rm_class_t *cl;
1664 {
1665 	mbuf_t  *m;
1666 
1667 	if ((m = _getq(cl->q_)) != NULL)
1668 		m_freem(m);
1669 }
1670 
1671 static mbuf_t *
1672 _rmc_getq(cl)
1673 	rm_class_t *cl;
1674 {
1675 #ifdef ALTQ_RIO
1676 	if (q_is_rio(cl->q_))
1677 		return rio_getq((rio_t *)cl->red_, cl->q_);
1678 #endif
1679 #ifdef ALTQ_RED
1680 	if (q_is_red(cl->q_))
1681 		return red_getq(cl->red_, cl->q_);
1682 #endif
1683 	return _getq(cl->q_);
1684 }
1685 
1686 static mbuf_t *
1687 _rmc_pollq(cl)
1688 	rm_class_t *cl;
1689 {
1690 	return qhead(cl->q_);
1691 }
1692 
1693 #ifdef CBQ_TRACE
1694 
1695 /*
1696  * DDB hook to trace cbq events:
1697  *  the last 1024 events are held in a circular buffer.
1698  *  use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1699  */
1700 void cbqtrace_dump(int);
1701 static char *rmc_funcname(void *);
1702 
1703 static struct rmc_funcs {
1704 	void *func;
1705 	char *name;
1706 } rmc_funcs[] = {
1707 	{ rmc_init, 			"rmc_init" },
1708 	{ rmc_queue_packet, 		"rmc_queue_packet" },
1709 	{ rmc_under_limit, 		"rmc_under_limit" },
1710 	{ rmc_update_class_util, 	"rmc_update_class_util" },
1711 	{ rmc_delay_action, 		"rmc_delay_action" },
1712 	{ rmc_restart, 			"rmc_restart" },
1713 	{ _rmc_wrr_dequeue_next, 	"_rmc_wrr_dequeue_next" },
1714 	{ NULL, 			NULL }
1715 };
1716 
1717 static char *rmc_funcname(func)
1718 	void *func;
1719 {
1720 	struct rmc_funcs *fp;
1721 
1722 	for (fp = rmc_funcs; fp->func != NULL; fp++)
1723 		if (fp->func == func)
1724 			return (fp->name);
1725 	return ("unknown");
1726 }
1727 
1728 void cbqtrace_dump(counter)
1729 	int counter;
1730 {
1731 	int i, *p;
1732 	char *cp;
1733 
1734 	counter = counter % NCBQTRACE;
1735 	p = (int *)&cbqtrace_buffer[counter];
1736 
1737 	for (i=0; i<20; i++) {
1738 		printf("[0x%x] ", *p++);
1739 		printf("%s: ", rmc_funcname((void *)*p++));
1740 		cp = (char *)p++;
1741 		printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1742 		printf("%d\n",*p++);
1743 
1744 		if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1745 			p = (int *)cbqtrace_buffer;
1746 	}
1747 }
1748 #endif /* CBQ_TRACE */
1749 
1750 #endif /* ALTQ_CBQ */
1751 
1752 #if defined(ALTQ_CBQ) || defined(ALTQ_RED) || defined(ALTQ_RIO) || defined(ALTQ_HFSC) || defined(ALTQ_PRIQ)
1753 #if !defined(__GNUC__) || defined(ALTQ_DEBUG)
1754 
1755 void
1756 _addq(q, m)
1757 	class_queue_t *q;
1758 	mbuf_t *m;
1759 {
1760         mbuf_t  *m0;
1761 
1762 	if ((m0 = qtail(q)) != NULL)
1763 		m->m_nextpkt = m0->m_nextpkt;
1764 	else
1765 		m0 = m;
1766 	m0->m_nextpkt = m;
1767 	qtail(q) = m;
1768 	qlen(q)++;
1769 }
1770 
1771 mbuf_t *
1772 _getq(q)
1773 	class_queue_t *q;
1774 {
1775 	mbuf_t  *m, *m0;
1776 
1777 	if ((m = qtail(q)) == NULL)
1778 		return (NULL);
1779 	if ((m0 = m->m_nextpkt) != m)
1780 		m->m_nextpkt = m0->m_nextpkt;
1781 	else {
1782 		ASSERT(qlen(q) == 1);
1783 		qtail(q) = NULL;
1784 	}
1785 	qlen(q)--;
1786 	m0->m_nextpkt = NULL;
1787 	return (m0);
1788 }
1789 
1790 /* drop a packet at the tail of the queue */
1791 mbuf_t *
1792 _getq_tail(q)
1793 	class_queue_t *q;
1794 {
1795 	mbuf_t *m, *m0, *prev;
1796 
1797 	if ((m = m0 = qtail(q)) == NULL)
1798 		return NULL;
1799 	do {
1800 		prev = m0;
1801 		m0 = m0->m_nextpkt;
1802 	} while (m0 != m);
1803 	prev->m_nextpkt = m->m_nextpkt;
1804 	if (prev == m)  {
1805 		ASSERT(qlen(q) == 1);
1806 		qtail(q) = NULL;
1807 	} else
1808 		qtail(q) = prev;
1809 	qlen(q)--;
1810 	m->m_nextpkt = NULL;
1811 	return (m);
1812 }
1813 
1814 /* randomly select a packet in the queue */
1815 mbuf_t *
1816 _getq_random(q)
1817 	class_queue_t *q;
1818 {
1819 	struct mbuf *m;
1820 	int i, n;
1821 
1822 	if ((m = qtail(q)) == NULL)
1823 		return NULL;
1824 	if (m->m_nextpkt == m) {
1825 		ASSERT(qlen(q) == 1);
1826 		qtail(q) = NULL;
1827 	} else {
1828 		struct mbuf *prev = NULL;
1829 
1830 		n = random() % qlen(q) + 1;
1831 		for (i = 0; i < n; i++) {
1832 			prev = m;
1833 			m = m->m_nextpkt;
1834 		}
1835 		prev->m_nextpkt = m->m_nextpkt;
1836 		if (m == qtail(q))
1837 			qtail(q) = prev;
1838 	}
1839 	qlen(q)--;
1840 	m->m_nextpkt = NULL;
1841 	return (m);
1842 }
1843 
1844 void
1845 _removeq(q, m)
1846 	class_queue_t *q;
1847 	mbuf_t *m;
1848 {
1849 	mbuf_t *m0, *prev;
1850 
1851 	m0 = qtail(q);
1852 	do {
1853 		prev = m0;
1854 		m0 = m0->m_nextpkt;
1855 	} while (m0 != m);
1856 	prev->m_nextpkt = m->m_nextpkt;
1857 	if (prev == m)
1858 		qtail(q) = NULL;
1859 	else if (qtail(q) == m)
1860 		qtail(q) = prev;
1861 	qlen(q)--;
1862 }
1863 
1864 void
1865 _flushq(q)
1866 	class_queue_t *q;
1867 {
1868 	mbuf_t *m;
1869 
1870 	while ((m = _getq(q)) != NULL)
1871 		m_freem(m);
1872 	ASSERT(qlen(q) == 0);
1873 }
1874 
1875 #endif /* !__GNUC__ || ALTQ_DEBUG */
1876 #endif /* ALTQ_CBQ || ALTQ_RED || ALTQ_RIO || ALTQ_HFSC || ALTQ_PRIQ */
1877