xref: /dragonfly/sys/net/altq/altq_rmclass.c (revision b7367ef6)
1 /*	$KAME: altq_rmclass.c,v 1.18 2003/11/06 06:32:53 kjc Exp $	*/
2 /*	$DragonFly: src/sys/net/altq/altq_rmclass.c,v 1.8 2006/12/22 23:44:55 swildner 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 #ident "@(#)rm_class.c  1.48     97/12/05 SMI"
41 
42 #include "opt_altq.h"
43 #include "opt_inet.h"
44 #include "opt_inet6.h"
45 
46 #ifdef ALTQ_CBQ	/* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
47 
48 #include <sys/param.h>
49 #include <sys/malloc.h>
50 #include <sys/mbuf.h>
51 #include <sys/socket.h>
52 #include <sys/systm.h>
53 #include <sys/callout.h>
54 #include <sys/errno.h>
55 #include <sys/time.h>
56 #include <sys/thread.h>
57 
58 #include <net/if.h>
59 
60 #include <net/altq/altq.h>
61 #include <net/altq/altq_rmclass.h>
62 #include <net/altq/altq_rmclass_debug.h>
63 #include <net/altq/altq_red.h>
64 #include <net/altq/altq_rio.h>
65 
66 #include <sys/thread2.h>
67 
68 #ifdef CBQ_TRACE
69 static struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
70 static struct cbqtrace *cbqtrace_ptr = NULL;
71 static int cbqtrace_count;
72 #endif
73 
74 /*
75  * Local Macros
76  */
77 
78 #define	reset_cutoff(ifd)	{ ifd->cutoff_ = RM_MAXDEPTH; }
79 
80 /*
81  * Local routines.
82  */
83 
84 static int	rmc_satisfied(struct rm_class *, struct timeval *);
85 static void	rmc_wrr_set_weights(struct rm_ifdat *);
86 static void	rmc_depth_compute(struct rm_class *);
87 static void	rmc_depth_recompute(rm_class_t *);
88 
89 static struct mbuf *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
90 static struct mbuf *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
91 
92 static int	_rmc_addq(rm_class_t *, struct mbuf *);
93 static void	_rmc_dropq(rm_class_t *);
94 static struct mbuf *_rmc_getq(rm_class_t *);
95 static struct mbuf *_rmc_pollq(rm_class_t *);
96 
97 static int	rmc_under_limit(struct rm_class *, struct timeval *);
98 static void	rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
99 static void	rmc_drop_action(struct rm_class *);
100 static void	rmc_restart(void *);
101 static void	rmc_root_overlimit(struct rm_class *, struct rm_class *);
102 
103 #define	BORROW_OFFTIME
104 /*
105  * BORROW_OFFTIME (experimental):
106  * borrow the offtime of the class borrowing from.
107  * the reason is that when its own offtime is set, the class is unable
108  * to borrow much, especially when cutoff is taking effect.
109  * but when the borrowed class is overloaded (advidle is close to minidle),
110  * use the borrowing class's offtime to avoid overload.
111  */
112 #define	ADJUST_CUTOFF
113 /*
114  * ADJUST_CUTOFF (experimental):
115  * if no underlimit class is found due to cutoff, increase cutoff and
116  * retry the scheduling loop.
117  * also, don't invoke delay_actions while cutoff is taking effect,
118  * since a sleeping class won't have a chance to be scheduled in the
119  * next loop.
120  *
121  * now heuristics for setting the top-level variable (cutoff_) becomes:
122  *	1. if a packet arrives for a not-overlimit class, set cutoff
123  *	   to the depth of the class.
124  *	2. if cutoff is i, and a packet arrives for an overlimit class
125  *	   with an underlimit ancestor at a lower level than i (say j),
126  *	   then set cutoff to j.
127  *	3. at scheduling a packet, if there is no underlimit class
128  *	   due to the current cutoff level, increase cutoff by 1 and
129  *	   then try to schedule again.
130  */
131 
132 /*
133  * rm_class_t *
134  * rmc_newclass(...) - Create a new resource management class at priority
135  * 'pri' on the interface given by 'ifd'.
136  *
137  * nsecPerByte  is the data rate of the interface in nanoseconds/byte.
138  *              E.g., 800 for a 10Mb/s ethernet.  If the class gets less
139  *              than 100% of the bandwidth, this number should be the
140  *              'effective' rate for the class.  Let f be the
141  *              bandwidth fraction allocated to this class, and let
142  *              nsPerByte be the data rate of the output link in
143  *              nanoseconds/byte.  Then nsecPerByte is set to
144  *              nsPerByte / f.  E.g., 1600 (= 800 / .5)
145  *              for a class that gets 50% of an ethernet's bandwidth.
146  *
147  * action       the routine to call when the class is over limit.
148  *
149  * maxq         max allowable queue size for class (in packets).
150  *
151  * parent       parent class pointer.
152  *
153  * borrow       class to borrow from (should be either 'parent' or null).
154  *
155  * maxidle      max value allowed for class 'idle' time estimate (this
156  *              parameter determines how large an initial burst of packets
157  *              can be before overlimit action is invoked.
158  *
159  * offtime      how long 'delay' action will delay when class goes over
160  *              limit (this parameter determines the steady-state burst
161  *              size when a class is running over its limit).
162  *
163  * Maxidle and offtime have to be computed from the following:  If the
164  * average packet size is s, the bandwidth fraction allocated to this
165  * class is f, we want to allow b packet bursts, and the gain of the
166  * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
167  *
168  *   ptime = s * nsPerByte * (1 - f) / f
169  *   maxidle = ptime * (1 - g^b) / g^b
170  *   minidle = -ptime * (1 / (f - 1))
171  *   offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
172  *
173  * Operationally, it's convenient to specify maxidle & offtime in units
174  * independent of the link bandwidth so the maxidle & offtime passed to
175  * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
176  * (The constant factor is a scale factor needed to make the parameters
177  * integers.  This scaling also means that the 'unscaled' values of
178  * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
179  * not nanoseconds.)  Also note that the 'idle' filter computation keeps
180  * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
181  * maxidle also must be scaled upward by this value.  Thus, the passed
182  * values for maxidle and offtime can be computed as follows:
183  *
184  * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
185  * offtime = offtime * 8 / (1000 * nsecPerByte)
186  *
187  * When USE_HRTIME is employed, then maxidle and offtime become:
188  * 	maxidle = maxilde * (8.0 / nsecPerByte);
189  * 	offtime = offtime * (8.0 / nsecPerByte);
190  */
191 struct rm_class *
192 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
193 	     void (*action)(rm_class_t *, rm_class_t *), int maxq,
194 	     struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
195 	     int minidle, u_int offtime, int pktsize, int flags)
196 {
197 	struct rm_class *cl;
198 	struct rm_class *peer;
199 
200 	if (pri >= RM_MAXPRIO)
201 		return (NULL);
202 #ifndef ALTQ_RED
203 	if (flags & RMCF_RED) {
204 #ifdef ALTQ_DEBUG
205 		kprintf("rmc_newclass: RED not configured for CBQ!\n");
206 #endif
207 		return (NULL);
208 	}
209 #endif
210 #ifndef ALTQ_RIO
211 	if (flags & RMCF_RIO) {
212 #ifdef ALTQ_DEBUG
213 		kprintf("rmc_newclass: RIO not configured for CBQ!\n");
214 #endif
215 		return (NULL);
216 	}
217 #endif
218 
219 	cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
220 	callout_init(&cl->callout_);
221 	cl->q_ = kmalloc(sizeof(*cl->q_), M_ALTQ, M_WAITOK | M_ZERO);
222 
223 	/*
224 	 * Class initialization.
225 	 */
226 	cl->children_ = NULL;
227 	cl->parent_ = parent;
228 	cl->borrow_ = borrow;
229 	cl->leaf_ = 1;
230 	cl->ifdat_ = ifd;
231 	cl->pri_ = pri;
232 	cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
233 	cl->depth_ = 0;
234 	cl->qthresh_ = 0;
235 	cl->ns_per_byte_ = nsecPerByte;
236 
237 	qlimit(cl->q_) = maxq;
238 	qtype(cl->q_) = Q_DROPHEAD;
239 	qlen(cl->q_) = 0;
240 	cl->flags_ = flags;
241 
242 #if 1 /* minidle is also scaled in ALTQ */
243 	cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
244 	if (cl->minidle_ > 0)
245 		cl->minidle_ = 0;
246 #else
247 	cl->minidle_ = minidle;
248 #endif
249 	cl->maxidle_ = (maxidle * nsecPerByte) / 8;
250 	if (cl->maxidle_ == 0)
251 		cl->maxidle_ = 1;
252 #if 1 /* offtime is also scaled in ALTQ */
253 	cl->avgidle_ = cl->maxidle_;
254 	cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
255 	if (cl->offtime_ == 0)
256 		cl->offtime_ = 1;
257 #else
258 	cl->avgidle_ = 0;
259 	cl->offtime_ = (offtime * nsecPerByte) / 8;
260 #endif
261 	cl->overlimit = action;
262 
263 #ifdef ALTQ_RED
264 	if (flags & (RMCF_RED|RMCF_RIO)) {
265 		int red_flags, red_pkttime;
266 
267 		red_flags = 0;
268 		if (flags & RMCF_ECN)
269 			red_flags |= REDF_ECN;
270 #ifdef ALTQ_RIO
271 		if (flags & RMCF_CLEARDSCP)
272 			red_flags |= RIOF_CLEARDSCP;
273 #endif
274 		red_pkttime = nsecPerByte * pktsize  / 1000;
275 
276 		if (flags & RMCF_RED) {
277 			cl->red_ = red_alloc(0, 0,
278 			    qlimit(cl->q_) * 10/100,
279 			    qlimit(cl->q_) * 30/100,
280 			    red_flags, red_pkttime);
281 			if (cl->red_ != NULL)
282 				qtype(cl->q_) = Q_RED;
283 		}
284 #ifdef ALTQ_RIO
285 		else {
286 			cl->red_ = (red_t *)rio_alloc(0, NULL,
287 						      red_flags, red_pkttime);
288 			if (cl->red_ != NULL)
289 				qtype(cl->q_) = Q_RIO;
290 		}
291 #endif
292 	}
293 #endif /* ALTQ_RED */
294 
295 	/*
296 	 * put the class into the class tree
297 	 */
298 	crit_enter();
299 	if ((peer = ifd->active_[pri]) != NULL) {
300 		/* find the last class at this pri */
301 		cl->peer_ = peer;
302 		while (peer->peer_ != ifd->active_[pri])
303 			peer = peer->peer_;
304 		peer->peer_ = cl;
305 	} else {
306 		ifd->active_[pri] = cl;
307 		cl->peer_ = cl;
308 	}
309 
310 	if (cl->parent_) {
311 		cl->next_ = parent->children_;
312 		parent->children_ = cl;
313 		parent->leaf_ = 0;
314 	}
315 
316 	/*
317 	 * Compute the depth of this class and its ancestors in the class
318 	 * hierarchy.
319 	 */
320 	rmc_depth_compute(cl);
321 
322 	/*
323 	 * If CBQ's WRR is enabled, then initialize the class WRR state.
324 	 */
325 	if (ifd->wrr_) {
326 		ifd->num_[pri]++;
327 		ifd->alloc_[pri] += cl->allotment_;
328 		rmc_wrr_set_weights(ifd);
329 	}
330 	crit_exit();
331 	return (cl);
332 }
333 
334 int
335 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
336 	     int minidle, u_int offtime, int pktsize)
337 {
338 	struct rm_ifdat *ifd;
339 	u_int old_allotment;
340 
341 	ifd = cl->ifdat_;
342 	old_allotment = cl->allotment_;
343 
344 	crit_enter();
345 	cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
346 	cl->qthresh_ = 0;
347 	cl->ns_per_byte_ = nsecPerByte;
348 
349 	qlimit(cl->q_) = maxq;
350 
351 #if 1 /* minidle is also scaled in ALTQ */
352 	cl->minidle_ = (minidle * nsecPerByte) / 8;
353 	if (cl->minidle_ > 0)
354 		cl->minidle_ = 0;
355 #else
356 	cl->minidle_ = minidle;
357 #endif
358 	cl->maxidle_ = (maxidle * nsecPerByte) / 8;
359 	if (cl->maxidle_ == 0)
360 		cl->maxidle_ = 1;
361 #if 1 /* offtime is also scaled in ALTQ */
362 	cl->avgidle_ = cl->maxidle_;
363 	cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
364 	if (cl->offtime_ == 0)
365 		cl->offtime_ = 1;
366 #else
367 	cl->avgidle_ = 0;
368 	cl->offtime_ = (offtime * nsecPerByte) / 8;
369 #endif
370 
371 	/*
372 	 * If CBQ's WRR is enabled, then initialize the class WRR state.
373 	 */
374 	if (ifd->wrr_) {
375 		ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
376 		rmc_wrr_set_weights(ifd);
377 	}
378 	crit_exit();
379 	return (0);
380 }
381 
382 /*
383  * static void
384  * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
385  *	the appropriate run robin weights for the CBQ weighted round robin
386  *	algorithm.
387  *
388  *	Returns: NONE
389  */
390 
391 static void
392 rmc_wrr_set_weights(struct rm_ifdat *ifd)
393 {
394 	int i;
395 	struct rm_class *cl, *clh;
396 
397 	for (i = 0; i < RM_MAXPRIO; i++) {
398 		/*
399 		 * This is inverted from that of the simulator to
400 		 * maintain precision.
401 		 */
402 		if (ifd->num_[i] == 0)
403 			ifd->M_[i] = 0;
404 		else
405 			ifd->M_[i] = ifd->alloc_[i] /
406 				(ifd->num_[i] * ifd->maxpkt_);
407 		/*
408 		 * Compute the weighted allotment for each class.
409 		 * This takes the expensive div instruction out
410 		 * of the main loop for the wrr scheduling path.
411 		 * These only get recomputed when a class comes or
412 		 * goes.
413 		 */
414 		if (ifd->active_[i] != NULL) {
415 			clh = cl = ifd->active_[i];
416 			do {
417 				/* safe-guard for slow link or alloc_ == 0 */
418 				if (ifd->M_[i] == 0)
419 					cl->w_allotment_ = 0;
420 				else
421 					cl->w_allotment_ = cl->allotment_ /
422 						ifd->M_[i];
423 				cl = cl->peer_;
424 			} while ((cl != NULL) && (cl != clh));
425 		}
426 	}
427 }
428 
429 int
430 rmc_get_weight(struct rm_ifdat *ifd, int pri)
431 {
432 	if ((pri >= 0) && (pri < RM_MAXPRIO))
433 		return (ifd->M_[pri]);
434 	else
435 		return (0);
436 }
437 
438 /*
439  * static void
440  * rmc_depth_compute(struct rm_class *cl) - This function computes the
441  *	appropriate depth of class 'cl' and its ancestors.
442  *
443  *	Returns:	NONE
444  */
445 
446 static void
447 rmc_depth_compute(struct rm_class *cl)
448 {
449 	rm_class_t *t = cl, *p;
450 
451 	/*
452 	 * Recompute the depth for the branch of the tree.
453 	 */
454 	while (t != NULL) {
455 		p = t->parent_;
456 		if (p && (t->depth_ >= p->depth_)) {
457 			p->depth_ = t->depth_ + 1;
458 			t = p;
459 		} else
460 			t = NULL;
461 	}
462 }
463 
464 /*
465  * static void
466  * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
467  *	the depth of the tree after a class has been deleted.
468  *
469  *	Returns:	NONE
470  */
471 
472 static void
473 rmc_depth_recompute(rm_class_t *cl)
474 {
475 #if 1 /* ALTQ */
476 	rm_class_t *p, *t;
477 
478 	p = cl;
479 	while (p != NULL) {
480 		if ((t = p->children_) == NULL) {
481 			p->depth_ = 0;
482 		} else {
483 			int cdepth = 0;
484 
485 			while (t != NULL) {
486 				if (t->depth_ > cdepth)
487 					cdepth = t->depth_;
488 				t = t->next_;
489 			}
490 
491 			if (p->depth_ == cdepth + 1)
492 				/* no change to this parent */
493 				return;
494 
495 			p->depth_ = cdepth + 1;
496 		}
497 
498 		p = p->parent_;
499 	}
500 #else
501 	rm_class_t	*t;
502 
503 	if (cl->depth_ >= 1) {
504 		if (cl->children_ == NULL) {
505 			cl->depth_ = 0;
506 		} else if ((t = cl->children_) != NULL) {
507 			while (t != NULL) {
508 				if (t->children_ != NULL)
509 					rmc_depth_recompute(t);
510 				t = t->next_;
511 			}
512 		} else
513 			rmc_depth_compute(cl);
514 	}
515 #endif
516 }
517 
518 /*
519  * void
520  * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
521  *	function deletes a class from the link-sharing structure and frees
522  *	all resources associated with the class.
523  *
524  *	Returns: NONE
525  */
526 
527 void
528 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
529 {
530 	struct rm_class *p, *head, *previous;
531 
532 	KKASSERT(cl->children_ == NULL);
533 
534 	if (cl->sleeping_)
535 		callout_stop(&cl->callout_);
536 
537 	crit_enter();
538 	/*
539 	 * Free packets in the packet queue.
540 	 * XXX - this may not be a desired behavior.  Packets should be
541 	 *		re-queued.
542 	 */
543 	rmc_dropall(cl);
544 
545 	/*
546 	 * If the class has a parent, then remove the class from the
547 	 * class from the parent's children chain.
548 	 */
549 	if (cl->parent_ != NULL) {
550 		head = cl->parent_->children_;
551 		p = previous = head;
552 		if (head->next_ == NULL) {
553 			KKASSERT(head == cl);
554 			cl->parent_->children_ = NULL;
555 			cl->parent_->leaf_ = 1;
556 		} else while (p != NULL) {
557 			if (p == cl) {
558 				if (cl == head)
559 					cl->parent_->children_ = cl->next_;
560 				else
561 					previous->next_ = cl->next_;
562 				cl->next_ = NULL;
563 				p = NULL;
564 			} else {
565 				previous = p;
566 				p = p->next_;
567 			}
568 		}
569 	}
570 
571 	/*
572 	 * Delete class from class priority peer list.
573 	 */
574 	if ((p = ifd->active_[cl->pri_]) != NULL) {
575 		/*
576 		 * If there is more than one member of this priority
577 		 * level, then look for class(cl) in the priority level.
578 		 */
579 		if (p != p->peer_) {
580 			while (p->peer_ != cl)
581 				p = p->peer_;
582 			p->peer_ = cl->peer_;
583 
584 			if (ifd->active_[cl->pri_] == cl)
585 				ifd->active_[cl->pri_] = cl->peer_;
586 		} else {
587 			KKASSERT(p == cl);
588 			ifd->active_[cl->pri_] = NULL;
589 		}
590 	}
591 
592 	/*
593 	 * Recompute the WRR weights.
594 	 */
595 	if (ifd->wrr_) {
596 		ifd->alloc_[cl->pri_] -= cl->allotment_;
597 		ifd->num_[cl->pri_]--;
598 		rmc_wrr_set_weights(ifd);
599 	}
600 
601 	/*
602 	 * Re-compute the depth of the tree.
603 	 */
604 #if 1 /* ALTQ */
605 	rmc_depth_recompute(cl->parent_);
606 #else
607 	rmc_depth_recompute(ifd->root_);
608 #endif
609 
610 	crit_exit();
611 
612 	/*
613 	 * Free the class structure.
614 	 */
615 	if (cl->red_ != NULL) {
616 #ifdef ALTQ_RIO
617 		if (q_is_rio(cl->q_))
618 			rio_destroy((rio_t *)cl->red_);
619 #endif
620 #ifdef ALTQ_RED
621 		if (q_is_red(cl->q_))
622 			red_destroy(cl->red_);
623 #endif
624 	}
625 	kfree(cl->q_, M_ALTQ);
626 	kfree(cl, M_ALTQ);
627 }
628 
629 /*
630  * void
631  * rmc_init(...) - Initialize the resource management data structures
632  *	associated with the output portion of interface 'ifp'.  'ifd' is
633  *	where the structures will be built (for backwards compatibility, the
634  *	structures aren't kept in the ifnet struct).  'nsecPerByte'
635  *	gives the link speed (inverse of bandwidth) in nanoseconds/byte.
636  *	'restart' is the driver-specific routine that the generic 'delay
637  *	until under limit' action will call to restart output.  `maxq'
638  *	is the queue size of the 'link' & 'default' classes.  'maxqueued'
639  *	is the maximum number of packets that the resource management
640  *	code will allow to be queued 'downstream' (this is typically 1).
641  *
642  *	Returns:	NONE
643  */
644 
645 void
646 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
647          void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
648 	 int minidle, u_int offtime, int flags)
649 {
650 	int i, mtu;
651 
652 	/*
653 	 * Initialize the CBQ tracing/debug facility.
654 	 */
655 	CBQTRACEINIT();
656 
657 	bzero(ifd, sizeof (*ifd));
658 	mtu = ifq->altq_ifp->if_mtu;
659 	ifd->ifq_ = ifq;
660 	ifd->restart = restart;
661 	ifd->maxqueued_ = maxqueued;
662 	ifd->ns_per_byte_ = nsecPerByte;
663 	ifd->maxpkt_ = mtu;
664 	ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
665 	ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
666 #if 1
667 	ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
668 	if (mtu * nsecPerByte > 10 * 1000000)
669 		ifd->maxiftime_ /= 4;
670 #endif
671 
672 	reset_cutoff(ifd);
673 	CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
674 
675 	/*
676 	 * Initialize the CBQ's WRR state.
677 	 */
678 	for (i = 0; i < RM_MAXPRIO; i++) {
679 		ifd->alloc_[i] = 0;
680 		ifd->M_[i] = 0;
681 		ifd->num_[i] = 0;
682 		ifd->na_[i] = 0;
683 		ifd->active_[i] = NULL;
684 	}
685 
686 	/*
687 	 * Initialize current packet state.
688 	 */
689 	ifd->qi_ = 0;
690 	ifd->qo_ = 0;
691 	for (i = 0; i < RM_MAXQUEUED; i++) {
692 		ifd->class_[i] = NULL;
693 		ifd->curlen_[i] = 0;
694 		ifd->borrowed_[i] = NULL;
695 	}
696 
697 	/*
698 	 * Create the root class of the link-sharing structure.
699 	 */
700 	ifd->root_ = rmc_newclass(0, ifd, nsecPerByte, rmc_root_overlimit,
701 				  maxq, 0, 0, maxidle, minidle, offtime, 0, 0);
702 	if (ifd->root_ == NULL) {
703 		kprintf("rmc_init: root class not allocated\n");
704 		return ;
705 	}
706 	ifd->root_->depth_ = 0;
707 }
708 
709 /*
710  * void
711  * rmc_queue_packet(struct rm_class *cl, struct mbuf *m) - Add packet given by
712  *	mbuf 'm' to queue for resource class 'cl'.  This routine is called
713  *	by a driver's if_output routine.  This routine must be called with
714  *	output packet completion interrupts locked out (to avoid racing with
715  *	rmc_dequeue_next).
716  *
717  *	Returns:	0 on successful queueing
718  *			-1 when packet drop occurs
719  */
720 int
721 rmc_queue_packet(struct rm_class *cl, struct mbuf *m)
722 {
723 	struct timeval now;
724 	struct rm_ifdat *ifd = cl->ifdat_;
725 	int cpri = cl->pri_;
726 	int is_empty = qempty(cl->q_);
727 
728 	RM_GETTIME(now);
729 	if (ifd->cutoff_ > 0) {
730 		if (TV_LT(&cl->undertime_, &now)) {
731 			if (ifd->cutoff_ > cl->depth_)
732 				ifd->cutoff_ = cl->depth_;
733 			CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
734 		}
735 #if 1 /* ALTQ */
736 		else {
737 			/*
738 			 * the class is overlimit. if the class has
739 			 * underlimit ancestors, set cutoff to the lowest
740 			 * depth among them.
741 			 */
742 			struct rm_class *borrow = cl->borrow_;
743 
744 			while (borrow != NULL &&
745 			       borrow->depth_ < ifd->cutoff_) {
746 				if (TV_LT(&borrow->undertime_, &now)) {
747 					ifd->cutoff_ = borrow->depth_;
748 					CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
749 					break;
750 				}
751 				borrow = borrow->borrow_;
752 			}
753 		}
754 #else /* !ALTQ */
755 		else if ((ifd->cutoff_ > 1) && cl->borrow_) {
756 			if (TV_LT(&cl->borrow_->undertime_, &now)) {
757 				ifd->cutoff_ = cl->borrow_->depth_;
758 				CBQTRACE(rmc_queue_packet, 'ffob',
759 					 cl->borrow_->depth_);
760 			}
761 		}
762 #endif /* !ALTQ */
763 	}
764 
765 	if (_rmc_addq(cl, m) < 0)
766 		/* failed */
767 		return (-1);
768 
769 	if (is_empty) {
770 		CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
771 		ifd->na_[cpri]++;
772 	}
773 
774 	if (qlen(cl->q_) > qlimit(cl->q_)) {
775 		/* note: qlimit can be set to 0 or 1 */
776 		rmc_drop_action(cl);
777 		return (-1);
778 	}
779 	return (0);
780 }
781 
782 /*
783  * void
784  * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
785  *	classes to see if there are satified.
786  */
787 
788 static void
789 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
790 {
791 	int i;
792 	rm_class_t *p, *bp;
793 
794 	for (i = RM_MAXPRIO - 1; i >= 0; i--) {
795 		if ((bp = ifd->active_[i]) != NULL) {
796 			p = bp;
797 			do {
798 				if (!rmc_satisfied(p, now)) {
799 					ifd->cutoff_ = p->depth_;
800 					return;
801 				}
802 				p = p->peer_;
803 			} while (p != bp);
804 		}
805 	}
806 
807 	reset_cutoff(ifd);
808 }
809 
810 /*
811  * rmc_satisfied - Return 1 of the class is satisfied.  O, otherwise.
812  */
813 
814 static int
815 rmc_satisfied(struct rm_class *cl, struct timeval *now)
816 {
817 	rm_class_t *p;
818 
819 	if (cl == NULL)
820 		return (1);
821 	if (TV_LT(now, &cl->undertime_))
822 		return (1);
823 	if (cl->depth_ == 0) {
824 		if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
825 			return (0);
826 		else
827 			return (1);
828 	}
829 	if (cl->children_ != NULL) {
830 		p = cl->children_;
831 		while (p != NULL) {
832 			if (!rmc_satisfied(p, now))
833 				return (0);
834 			p = p->next_;
835 		}
836 	}
837 
838 	return (1);
839 }
840 
841 /*
842  * Return 1 if class 'cl' is under limit or can borrow from a parent,
843  * 0 if overlimit.  As a side-effect, this routine will invoke the
844  * class overlimit action if the class if overlimit.
845  */
846 
847 static int
848 rmc_under_limit(struct rm_class *cl, struct timeval *now)
849 {
850 	rm_class_t *p = cl;
851 	rm_class_t *top;
852 	struct rm_ifdat *ifd = cl->ifdat_;
853 
854 	ifd->borrowed_[ifd->qi_] = NULL;
855 	/*
856 	 * If cl is the root class, then always return that it is
857 	 * underlimit.  Otherwise, check to see if the class is underlimit.
858 	 */
859 	if (cl->parent_ == NULL)
860 		return (1);
861 
862 	if (cl->sleeping_) {
863 		if (TV_LT(now, &cl->undertime_))
864 			return (0);
865 
866 		callout_stop(&cl->callout_);
867 		cl->sleeping_ = 0;
868 		cl->undertime_.tv_sec = 0;
869 		return (1);
870 	}
871 
872 	top = NULL;
873 	while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
874 		if (((cl = cl->borrow_) == NULL) ||
875 		    (cl->depth_ > ifd->cutoff_)) {
876 #ifdef ADJUST_CUTOFF
877 			if (cl != NULL)
878 				/* cutoff is taking effect, just
879 				   return false without calling
880 				   the delay action. */
881 				return (0);
882 #endif
883 #ifdef BORROW_OFFTIME
884 			/*
885 			 * check if the class can borrow offtime too.
886 			 * borrow offtime from the top of the borrow
887 			 * chain if the top class is not overloaded.
888 			 */
889 			if (cl != NULL) {
890 				/* cutoff is taking effect, use this class as top. */
891 				top = cl;
892 				CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
893 			}
894 			if (top != NULL && top->avgidle_ == top->minidle_)
895 				top = NULL;
896 			p->overtime_ = *now;
897 			(p->overlimit)(p, top);
898 #else
899 			p->overtime_ = *now;
900 			(p->overlimit)(p, NULL);
901 #endif
902 			return (0);
903 		}
904 		top = cl;
905 	}
906 
907 	if (cl != p)
908 		ifd->borrowed_[ifd->qi_] = cl;
909 	return (1);
910 }
911 
912 /*
913  * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
914  *	Packet-by-packet round robin.
915  *
916  * The heart of the weighted round-robin scheduler, which decides which
917  * class next gets to send a packet.  Highest priority first, then
918  * weighted round-robin within priorites.
919  *
920  * Each able-to-send class gets to send until its byte allocation is
921  * exhausted.  Thus, the active pointer is only changed after a class has
922  * exhausted its allocation.
923  *
924  * If the scheduler finds no class that is underlimit or able to borrow,
925  * then the first class found that had a nonzero queue and is allowed to
926  * borrow gets to send.
927  */
928 
929 static struct mbuf *
930 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
931 {
932 	struct rm_class *cl = NULL, *first = NULL;
933 	u_int deficit;
934 	int cpri;
935 	struct mbuf *m;
936 	struct timeval now;
937 
938 	RM_GETTIME(now);
939 
940 	/*
941 	 * if the driver polls the top of the queue and then removes
942 	 * the polled packet, we must return the same packet.
943 	 */
944 	if (op == ALTDQ_REMOVE && ifd->pollcache_) {
945 		cl = ifd->pollcache_;
946 		cpri = cl->pri_;
947 		if (ifd->efficient_) {
948 			/* check if this class is overlimit */
949 			if (cl->undertime_.tv_sec != 0 &&
950 			    rmc_under_limit(cl, &now) == 0)
951 				first = cl;
952 		}
953 		ifd->pollcache_ = NULL;
954 		goto _wrr_out;
955 	}
956 	else {
957 		/* mode == ALTDQ_POLL || pollcache == NULL */
958 		ifd->pollcache_ = NULL;
959 		ifd->borrowed_[ifd->qi_] = NULL;
960 	}
961 #ifdef ADJUST_CUTOFF
962  _again:
963 #endif
964 	for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
965 		if (ifd->na_[cpri] == 0)
966 			continue;
967 		deficit = 0;
968 		/*
969 		 * Loop through twice for a priority level, if some class
970 		 * was unable to send a packet the first round because
971 		 * of the weighted round-robin mechanism.
972 		 * During the second loop at this level, deficit==2.
973 		 * (This second loop is not needed if for every class,
974 		 * "M[cl->pri_])" times "cl->allotment" is greater than
975 		 * the byte size for the largest packet in the class.)
976 		 */
977  _wrr_loop:
978 		cl = ifd->active_[cpri];
979 		KKASSERT(cl != NULL);
980 		do {
981 			if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
982 				cl->bytes_alloc_ += cl->w_allotment_;
983 			if (!qempty(cl->q_)) {
984 				if ((cl->undertime_.tv_sec == 0) ||
985 				    rmc_under_limit(cl, &now)) {
986 					if (cl->bytes_alloc_ > 0 || deficit > 1)
987 						goto _wrr_out;
988 
989 					/* underlimit but no alloc */
990 					deficit = 1;
991 #if 1
992 					ifd->borrowed_[ifd->qi_] = NULL;
993 #endif
994 				}
995 				else if (first == NULL && cl->borrow_ != NULL)
996 					first = cl; /* borrowing candidate */
997 			}
998 
999 			cl->bytes_alloc_ = 0;
1000 			cl = cl->peer_;
1001 		} while (cl != ifd->active_[cpri]);
1002 
1003 		if (deficit == 1) {
1004 			/* first loop found an underlimit class with deficit */
1005 			/* Loop on same priority level, with new deficit.  */
1006 			deficit = 2;
1007 			goto _wrr_loop;
1008 		}
1009 	}
1010 
1011 #ifdef ADJUST_CUTOFF
1012 	/*
1013 	 * no underlimit class found.  if cutoff is taking effect,
1014 	 * increase cutoff and try again.
1015 	 */
1016 	if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1017 		ifd->cutoff_++;
1018 		CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1019 		goto _again;
1020 	}
1021 #endif /* ADJUST_CUTOFF */
1022 	/*
1023 	 * If LINK_EFFICIENCY is turned on, then the first overlimit
1024 	 * class we encounter will send a packet if all the classes
1025 	 * of the link-sharing structure are overlimit.
1026 	 */
1027 	reset_cutoff(ifd);
1028 	CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1029 
1030 	if (!ifd->efficient_ || first == NULL)
1031 		return (NULL);
1032 
1033 	cl = first;
1034 	cpri = cl->pri_;
1035 #if 0	/* too time-consuming for nothing */
1036 	if (cl->sleeping_)
1037 		callout_stop(&cl->callout_);
1038 	cl->sleeping_ = 0;
1039 	cl->undertime_.tv_sec = 0;
1040 #endif
1041 	ifd->borrowed_[ifd->qi_] = cl->borrow_;
1042 	ifd->cutoff_ = cl->borrow_->depth_;
1043 
1044 	/*
1045 	 * Deque the packet and do the book keeping...
1046 	 */
1047  _wrr_out:
1048 	if (op == ALTDQ_REMOVE) {
1049 		m = _rmc_getq(cl);
1050 		if (m == NULL)
1051 			panic("_rmc_wrr_dequeue_next");
1052 		if (qempty(cl->q_))
1053 			ifd->na_[cpri]--;
1054 
1055 		/*
1056 		 * Update class statistics and link data.
1057 		 */
1058 		if (cl->bytes_alloc_ > 0)
1059 			cl->bytes_alloc_ -= m_pktlen(m);
1060 
1061 		if ((cl->bytes_alloc_ <= 0) || first == cl)
1062 			ifd->active_[cl->pri_] = cl->peer_;
1063 		else
1064 			ifd->active_[cl->pri_] = cl;
1065 
1066 		ifd->class_[ifd->qi_] = cl;
1067 		ifd->curlen_[ifd->qi_] = m_pktlen(m);
1068 		ifd->now_[ifd->qi_] = now;
1069 		ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1070 		ifd->queued_++;
1071 	} else {
1072 		/* mode == ALTDQ_PPOLL */
1073 		m = _rmc_pollq(cl);
1074 		ifd->pollcache_ = cl;
1075 	}
1076 	return (m);
1077 }
1078 
1079 /*
1080  * Dequeue & return next packet from the highest priority class that
1081  * has a packet to send & has enough allocation to send it.  This
1082  * routine is called by a driver whenever it needs a new packet to
1083  * output.
1084  */
1085 static struct mbuf *
1086 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1087 {
1088 	struct mbuf *m;
1089 	int cpri;
1090 	struct rm_class *cl, *first = NULL;
1091 	struct timeval now;
1092 
1093 	RM_GETTIME(now);
1094 
1095 	/*
1096 	 * if the driver polls the top of the queue and then removes
1097 	 * the polled packet, we must return the same packet.
1098 	 */
1099 	if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1100 		cl = ifd->pollcache_;
1101 		cpri = cl->pri_;
1102 		ifd->pollcache_ = NULL;
1103 		goto _prr_out;
1104 	} else {
1105 		/* mode == ALTDQ_POLL || pollcache == NULL */
1106 		ifd->pollcache_ = NULL;
1107 		ifd->borrowed_[ifd->qi_] = NULL;
1108 	}
1109 #ifdef ADJUST_CUTOFF
1110  _again:
1111 #endif
1112 	for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1113 		if (ifd->na_[cpri] == 0)
1114 			continue;
1115 		cl = ifd->active_[cpri];
1116 		KKASSERT(cl != NULL);
1117 		do {
1118 			if (!qempty(cl->q_)) {
1119 				if ((cl->undertime_.tv_sec == 0) ||
1120 				    rmc_under_limit(cl, &now))
1121 					goto _prr_out;
1122 				if (first == NULL && cl->borrow_ != NULL)
1123 					first = cl;
1124 			}
1125 			cl = cl->peer_;
1126 		} while (cl != ifd->active_[cpri]);
1127 	}
1128 
1129 #ifdef ADJUST_CUTOFF
1130 	/*
1131 	 * no underlimit class found.  if cutoff is taking effect, increase
1132 	 * cutoff and try again.
1133 	 */
1134 	if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1135 		ifd->cutoff_++;
1136 		goto _again;
1137 	}
1138 #endif /* ADJUST_CUTOFF */
1139 	/*
1140 	 * If LINK_EFFICIENCY is turned on, then the first overlimit
1141 	 * class we encounter will send a packet if all the classes
1142 	 * of the link-sharing structure are overlimit.
1143 	 */
1144 	reset_cutoff(ifd);
1145 	if (!ifd->efficient_ || first == NULL)
1146 		return (NULL);
1147 
1148 	cl = first;
1149 	cpri = cl->pri_;
1150 #if 0	/* too time-consuming for nothing */
1151 	if (cl->sleeping_)
1152 		callout_stop(&cl->callout_);
1153 	cl->sleeping_ = 0;
1154 	cl->undertime_.tv_sec = 0;
1155 #endif
1156 	ifd->borrowed_[ifd->qi_] = cl->borrow_;
1157 	ifd->cutoff_ = cl->borrow_->depth_;
1158 
1159 	/*
1160 	 * Deque the packet and do the book keeping...
1161 	 */
1162  _prr_out:
1163 	if (op == ALTDQ_REMOVE) {
1164 		m = _rmc_getq(cl);
1165 		if (m == NULL)
1166 			panic("_rmc_prr_dequeue_next");
1167 		if (qempty(cl->q_))
1168 			ifd->na_[cpri]--;
1169 
1170 		ifd->active_[cpri] = cl->peer_;
1171 
1172 		ifd->class_[ifd->qi_] = cl;
1173 		ifd->curlen_[ifd->qi_] = m_pktlen(m);
1174 		ifd->now_[ifd->qi_] = now;
1175 		ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1176 		ifd->queued_++;
1177 	} else {
1178 		/* mode == ALTDQ_POLL */
1179 		m = _rmc_pollq(cl);
1180 		ifd->pollcache_ = cl;
1181 	}
1182 	return (m);
1183 }
1184 
1185 /*
1186  * struct mbuf *
1187  * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1188  *	is invoked by the packet driver to get the next packet to be
1189  *	dequeued and output on the link.  If WRR is enabled, then the
1190  *	WRR dequeue next routine will determine the next packet to sent.
1191  *	Otherwise, packet-by-packet round robin is invoked.
1192  *
1193  *	Returns:	NULL, if a packet is not available or if all
1194  *			classes are overlimit.
1195  *
1196  *			Otherwise, Pointer to the next packet.
1197  */
1198 
1199 struct mbuf *
1200 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1201 {
1202 	if (ifd->queued_ >= ifd->maxqueued_)
1203 		return (NULL);
1204 	else if (ifd->wrr_)
1205 		return (_rmc_wrr_dequeue_next(ifd, mode));
1206 	else
1207 		return (_rmc_prr_dequeue_next(ifd, mode));
1208 }
1209 
1210 /*
1211  * Update the utilization estimate for the packet that just completed.
1212  * The packet's class & the parent(s) of that class all get their
1213  * estimators updated.  This routine is called by the driver's output-
1214  * packet-completion interrupt service routine.
1215  */
1216 
1217 /*
1218  * a macro to approximate "divide by 1000" that gives 0.000999,
1219  * if a value has enough effective digits.
1220  * (on pentium, mul takes 9 cycles but div takes 46!)
1221  */
1222 #define	NSEC_TO_USEC(t)	(((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1223 void
1224 rmc_update_class_util(struct rm_ifdat *ifd)
1225 {
1226 	int idle, avgidle, pktlen;
1227 	int pkt_time, tidle;
1228 	rm_class_t *cl, *borrowed;
1229 	rm_class_t *borrows;
1230 	struct timeval *nowp;
1231 
1232 	/*
1233 	 * Get the most recent completed class.
1234 	 */
1235 	if ((cl = ifd->class_[ifd->qo_]) == NULL)
1236 		return;
1237 
1238 	pktlen = ifd->curlen_[ifd->qo_];
1239 	borrowed = ifd->borrowed_[ifd->qo_];
1240 	borrows = borrowed;
1241 
1242 	PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1243 
1244 	/*
1245 	 * Run estimator on class and its ancestors.
1246 	 */
1247 	/*
1248 	 * rm_update_class_util is designed to be called when the
1249 	 * transfer is completed from a xmit complete interrupt,
1250 	 * but most drivers don't implement an upcall for that.
1251 	 * so, just use estimated completion time.
1252 	 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1253 	 */
1254 	nowp = &ifd->now_[ifd->qo_];
1255 	/* get pkt_time (for link) in usec */
1256 #if 1  /* use approximation */
1257 	pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1258 	pkt_time = NSEC_TO_USEC(pkt_time);
1259 #else
1260 	pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1261 #endif
1262 #if 1 /* ALTQ4PPP */
1263 	if (TV_LT(nowp, &ifd->ifnow_)) {
1264 		int iftime;
1265 
1266 		/*
1267 		 * make sure the estimated completion time does not go
1268 		 * too far.  it can happen when the link layer supports
1269 		 * data compression or the interface speed is set to
1270 		 * a much lower value.
1271 		 */
1272 		TV_DELTA(&ifd->ifnow_, nowp, iftime);
1273 		if (iftime+pkt_time < ifd->maxiftime_) {
1274 			TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1275 		} else {
1276 			TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1277 		}
1278 	} else {
1279 		TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1280 	}
1281 #else
1282 	if (TV_LT(nowp, &ifd->ifnow_)) {
1283 		TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1284 	} else {
1285 		TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1286 	}
1287 #endif
1288 
1289 	while (cl != NULL) {
1290 		TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1291 		if (idle >= 2000000)
1292 			/*
1293 			 * this class is idle enough, reset avgidle.
1294 			 * (TV_DELTA returns 2000000 us when delta is large.)
1295 			 */
1296 			cl->avgidle_ = cl->maxidle_;
1297 
1298 		/* get pkt_time (for class) in usec */
1299 #if 1  /* use approximation */
1300 		pkt_time = pktlen * cl->ns_per_byte_;
1301 		pkt_time = NSEC_TO_USEC(pkt_time);
1302 #else
1303 		pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1304 #endif
1305 		idle -= pkt_time;
1306 
1307 		avgidle = cl->avgidle_;
1308 		avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1309 		cl->avgidle_ = avgidle;
1310 
1311 		/* Are we overlimit ? */
1312 		if (avgidle <= 0) {
1313 			CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1314 #if 1 /* ALTQ */
1315 			/*
1316 			 * need some lower bound for avgidle, otherwise
1317 			 * a borrowing class gets unbounded penalty.
1318 			 */
1319 			if (avgidle < cl->minidle_)
1320 				avgidle = cl->avgidle_ = cl->minidle_;
1321 #endif
1322 			/* set next idle to make avgidle 0 */
1323 			tidle = pkt_time +
1324 				(((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1325 			TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1326 			++cl->stats_.over;
1327 		} else {
1328 			cl->avgidle_ =
1329 			    (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1330 			cl->undertime_.tv_sec = 0;
1331 			if (cl->sleeping_) {
1332 				callout_stop(&cl->callout_);
1333 				cl->sleeping_ = 0;
1334 			}
1335 		}
1336 
1337 		if (borrows != NULL) {
1338 			if (borrows != cl)
1339 				++cl->stats_.borrows;
1340 			else
1341 				borrows = NULL;
1342 		}
1343 		cl->last_ = ifd->ifnow_;
1344 		cl->last_pkttime_ = pkt_time;
1345 
1346 #if 1
1347 		if (cl->parent_ == NULL) {
1348 			/* take stats of root class */
1349 			PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1350 		}
1351 #endif
1352 
1353 		cl = cl->parent_;
1354 	}
1355 
1356 	/*
1357 	 * Check to see if cutoff needs to set to a new level.
1358 	 */
1359 	cl = ifd->class_[ifd->qo_];
1360 	if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1361 #if 1 /* ALTQ */
1362 		if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1363 			rmc_tl_satisfied(ifd, nowp);
1364 			CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1365 		} else {
1366 			ifd->cutoff_ = borrowed->depth_;
1367 			CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1368 		}
1369 #else /* !ALTQ */
1370 		if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1371 			reset_cutoff(ifd);
1372 #ifdef notdef
1373 			rmc_tl_satisfied(ifd, &now);
1374 #endif
1375 			CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1376 		} else {
1377 			ifd->cutoff_ = borrowed->depth_;
1378 			CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1379 		}
1380 #endif /* !ALTQ */
1381 	}
1382 
1383 	/*
1384 	 * Release class slot
1385 	 */
1386 	ifd->borrowed_[ifd->qo_] = NULL;
1387 	ifd->class_[ifd->qo_] = NULL;
1388 	ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1389 	ifd->queued_--;
1390 }
1391 
1392 /*
1393  * void
1394  * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1395  *	over-limit action routines.  These get invoked by rmc_under_limit()
1396  *	if a class with packets to send if over its bandwidth limit & can't
1397  *	borrow from a parent class.
1398  *
1399  *	Returns: NONE
1400  */
1401 
1402 static void
1403 rmc_drop_action(struct rm_class *cl)
1404 {
1405 	struct rm_ifdat *ifd = cl->ifdat_;
1406 
1407 	KKASSERT(qlen(cl->q_) > 0);
1408 	_rmc_dropq(cl);
1409 	if (qempty(cl->q_))
1410 		ifd->na_[cl->pri_]--;
1411 }
1412 
1413 void
1414 rmc_dropall(struct rm_class *cl)
1415 {
1416 	struct rm_ifdat *ifd = cl->ifdat_;
1417 
1418 	if (!qempty(cl->q_)) {
1419 		_flushq(cl->q_);
1420 
1421 		ifd->na_[cl->pri_]--;
1422 	}
1423 }
1424 
1425 /*
1426  * void
1427  * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1428  *	delay action routine.  It is invoked via rmc_under_limit when the
1429  *	packet is discoverd to be overlimit.
1430  *
1431  *	If the delay action is result of borrow class being overlimit, then
1432  *	delay for the offtime of the borrowing class that is overlimit.
1433  *
1434  *	Returns: NONE
1435  */
1436 
1437 void
1438 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1439 {
1440 	int delay, t, extradelay;
1441 
1442 	cl->stats_.overactions++;
1443 	TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1444 #ifndef BORROW_OFFTIME
1445 	delay += cl->offtime_;
1446 #endif
1447 
1448 	if (!cl->sleeping_) {
1449 		CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1450 #ifdef BORROW_OFFTIME
1451 		if (borrow != NULL)
1452 			extradelay = borrow->offtime_;
1453 		else
1454 #endif
1455 			extradelay = cl->offtime_;
1456 
1457 #ifdef ALTQ
1458 		/*
1459 		 * XXX recalculate suspend time:
1460 		 * current undertime is (tidle + pkt_time) calculated
1461 		 * from the last transmission.
1462 		 *	tidle: time required to bring avgidle back to 0
1463 		 *	pkt_time: target waiting time for this class
1464 		 * we need to replace pkt_time by offtime
1465 		 */
1466 		extradelay -= cl->last_pkttime_;
1467 #endif
1468 		if (extradelay > 0) {
1469 			TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1470 			delay += extradelay;
1471 		}
1472 
1473 		cl->sleeping_ = 1;
1474 		cl->stats_.delays++;
1475 
1476 		/*
1477 		 * Since packets are phased randomly with respect to the
1478 		 * clock, 1 tick (the next clock tick) can be an arbitrarily
1479 		 * short time so we have to wait for at least two ticks.
1480 		 * NOTE:  If there's no other traffic, we need the timer as
1481 		 * a 'backstop' to restart this class.
1482 		 */
1483 		if (delay > tick * 2)
1484 			t = (delay + tick - 1) / tick;
1485 		else
1486 			t = 2;
1487 		callout_reset(&cl->callout_, t, rmc_restart, cl);
1488 	}
1489 }
1490 
1491 /*
1492  * void
1493  * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1494  *	called by the system timer code & is responsible checking if the
1495  *	class is still sleeping (it might have been restarted as a side
1496  *	effect of the queue scan on a packet arrival) and, if so, restarting
1497  *	output for the class.  Inspecting the class state & restarting output
1498  *	require locking the class structure.  In general the driver is
1499  *	responsible for locking but this is the only routine that is not
1500  *	called directly or indirectly from the interface driver so it has
1501  *	know about system locking conventions.  Under bsd, locking is done
1502  *	by raising IPL to splimp so that's what's implemented here.  On a
1503  *	different system this would probably need to be changed.
1504  *
1505  *	Since this function is called from an independant timeout, we
1506  *	have to set up the lock conditions expected for the ALTQ operation.
1507  *	Note that the restart will probably fall through to an if_start.
1508  *
1509  *	Returns:	NONE
1510  */
1511 
1512 static void
1513 rmc_restart(void *arg)
1514 {
1515 	struct rm_class *cl = arg;
1516 	struct rm_ifdat *ifd = cl->ifdat_;
1517 
1518 	lwkt_serialize_enter(ifd->ifq_->altq_ifp->if_serializer);
1519 	if (cl->sleeping_) {
1520 		cl->sleeping_ = 0;
1521 		cl->undertime_.tv_sec = 0;
1522 
1523 		if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1524 			CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1525 			(ifd->restart)(ifd->ifq_);
1526 		}
1527 	}
1528 	lwkt_serialize_exit(ifd->ifq_->altq_ifp->if_serializer);
1529 }
1530 
1531 /*
1532  * void
1533  * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1534  *	handling routine for the root class of the link sharing structure.
1535  *
1536  *	Returns: NONE
1537  */
1538 
1539 static void
1540 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1541 {
1542         panic("rmc_root_overlimit");
1543 }
1544 
1545 /*
1546  * Packet Queue handling routines.  Eventually, this is to localize the
1547  *	effects on the code whether queues are red queues or droptail
1548  *	queues.
1549  */
1550 
1551 static int
1552 _rmc_addq(rm_class_t *cl, struct mbuf *m)
1553 {
1554 #ifdef ALTQ_RIO
1555 	if (q_is_rio(cl->q_))
1556 		return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1557 #endif
1558 #ifdef ALTQ_RED
1559 	if (q_is_red(cl->q_))
1560 		return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1561 #endif /* ALTQ_RED */
1562 
1563 	if (cl->flags_ & RMCF_CLEARDSCP)
1564 		write_dsfield(m, cl->pktattr_, 0);
1565 
1566 	_addq(cl->q_, m);
1567 	return (0);
1568 }
1569 
1570 /* note: _rmc_dropq is not called for red */
1571 static void
1572 _rmc_dropq(rm_class_t *cl)
1573 {
1574 	struct mbuf *m;
1575 
1576 	if ((m = _getq(cl->q_)) != NULL)
1577 		m_freem(m);
1578 }
1579 
1580 static struct mbuf *
1581 _rmc_getq(rm_class_t *cl)
1582 {
1583 #ifdef ALTQ_RIO
1584 	if (q_is_rio(cl->q_))
1585 		return rio_getq((rio_t *)cl->red_, cl->q_);
1586 #endif
1587 #ifdef ALTQ_RED
1588 	if (q_is_red(cl->q_))
1589 		return red_getq(cl->red_, cl->q_);
1590 #endif
1591 	return _getq(cl->q_);
1592 }
1593 
1594 static struct mbuf *
1595 _rmc_pollq(rm_class_t *cl)
1596 {
1597 	return qhead(cl->q_);
1598 }
1599 
1600 #ifdef CBQ_TRACE
1601 /*
1602  * DDB hook to trace cbq events:
1603  *  the last 1024 events are held in a circular buffer.
1604  *  use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1605  */
1606 void		cbqtrace_dump(int);
1607 static char	*rmc_funcname(void *);
1608 
1609 static struct rmc_funcs {
1610 	void	*func;
1611 	char	*name;
1612 } rmc_funcs[] = {
1613 	rmc_init,		"rmc_init",
1614 	rmc_queue_packet,	"rmc_queue_packet",
1615 	rmc_under_limit,	"rmc_under_limit",
1616 	rmc_update_class_util,	"rmc_update_class_util",
1617 	rmc_delay_action,	"rmc_delay_action",
1618 	rmc_restart,		"rmc_restart",
1619 	_rmc_wrr_dequeue_next,	"_rmc_wrr_dequeue_next",
1620 	NULL,			NULL
1621 };
1622 
1623 static chari *
1624 rmc_funcname(void *func)
1625 {
1626 	struct rmc_funcs *fp;
1627 
1628 	for (fp = rmc_funcs; fp->func != NULL; fp++) {
1629 		if (fp->func == func)
1630 			return (fp->name);
1631 	}
1632 
1633 	return ("unknown");
1634 }
1635 
1636 void
1637 cbqtrace_dump(int counter)
1638 {
1639 	int i, *p;
1640 	char *cp;
1641 
1642 	counter = counter % NCBQTRACE;
1643 	p = (int *)&cbqtrace_buffer[counter];
1644 
1645 	for (i=0; i<20; i++) {
1646 		kprintf("[0x%x] ", *p++);
1647 		kprintf("%s: ", rmc_funcname((void *)*p++));
1648 		cp = (char *)p++;
1649 		kprintf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1650 		kprintf("%d\n",*p++);
1651 
1652 		if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1653 			p = (int *)cbqtrace_buffer;
1654 	}
1655 }
1656 #endif /* CBQ_TRACE */
1657 #endif /* ALTQ_CBQ */
1658