xref: /dragonfly/sys/net/altq/altq_hfsc.c (revision 9ddb8543)
1 /*	$KAME: altq_hfsc.c,v 1.25 2004/04/17 10:54:48 kjc Exp $	*/
2 /*	$DragonFly: src/sys/net/altq/altq_hfsc.c,v 1.9 2008/05/14 11:59:23 sephe Exp $ */
3 
4 /*
5  * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
6  *
7  * Permission to use, copy, modify, and distribute this software and
8  * its documentation is hereby granted (including for commercial or
9  * for-profit use), provided that both the copyright notice and this
10  * permission notice appear in all copies of the software, derivative
11  * works, or modified versions, and any portions thereof.
12  *
13  * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
14  * WHICH MAY HAVE SERIOUS CONSEQUENCES.  CARNEGIE MELLON PROVIDES THIS
15  * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
16  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
18  * DISCLAIMED.  IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
21  * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
22  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
23  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
25  * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
26  * DAMAGE.
27  *
28  * Carnegie Mellon encourages (but does not require) users of this
29  * software to return any improvements or extensions that they make,
30  * and to grant Carnegie Mellon the rights to redistribute these
31  * changes without encumbrance.
32  */
33 /*
34  * H-FSC is described in Proceedings of SIGCOMM'97,
35  * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
36  * Real-Time and Priority Service"
37  * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
38  *
39  * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
40  * when a class has an upperlimit, the fit-time is computed from the
41  * upperlimit service curve.  the link-sharing scheduler does not schedule
42  * a class whose fit-time exceeds the current time.
43  */
44 
45 #include "opt_altq.h"
46 #include "opt_inet.h"
47 #include "opt_inet6.h"
48 
49 #ifdef ALTQ_HFSC  /* hfsc is enabled by ALTQ_HFSC option in opt_altq.h */
50 
51 #include <sys/param.h>
52 #include <sys/malloc.h>
53 #include <sys/mbuf.h>
54 #include <sys/socket.h>
55 #include <sys/systm.h>
56 #include <sys/errno.h>
57 #include <sys/queue.h>
58 #include <sys/thread.h>
59 
60 #include <net/if.h>
61 #include <net/ifq_var.h>
62 #include <netinet/in.h>
63 
64 #include <net/pf/pfvar.h>
65 #include <net/altq/altq.h>
66 #include <net/altq/altq_hfsc.h>
67 
68 #include <sys/thread2.h>
69 
70 /*
71  * function prototypes
72  */
73 static int	hfsc_clear_interface(struct hfsc_if *);
74 static int	hfsc_request(struct ifaltq *, int, void *);
75 static void	hfsc_purge(struct hfsc_if *);
76 static struct hfsc_class *hfsc_class_create(struct hfsc_if *,
77 					    struct service_curve *,
78 					    struct service_curve *,
79 					    struct service_curve *,
80 					    struct hfsc_class *, int, int, int);
81 static int	hfsc_class_destroy(struct hfsc_class *);
82 static struct hfsc_class *hfsc_nextclass(struct hfsc_class *);
83 static int	hfsc_enqueue(struct ifaltq *, struct mbuf *,
84 			     struct altq_pktattr *);
85 static struct mbuf *hfsc_dequeue(struct ifaltq *, struct mbuf *, int);
86 
87 static int	hfsc_addq(struct hfsc_class *, struct mbuf *);
88 static struct mbuf *hfsc_getq(struct hfsc_class *);
89 static struct mbuf *hfsc_pollq(struct hfsc_class *);
90 static void	hfsc_purgeq(struct hfsc_class *);
91 
92 static void	update_cfmin(struct hfsc_class *);
93 static void	set_active(struct hfsc_class *, int);
94 static void	set_passive(struct hfsc_class *);
95 
96 static void	init_ed(struct hfsc_class *, int);
97 static void	update_ed(struct hfsc_class *, int);
98 static void	update_d(struct hfsc_class *, int);
99 static void	init_vf(struct hfsc_class *, int);
100 static void	update_vf(struct hfsc_class *, int, uint64_t);
101 static ellist_t *ellist_alloc(void);
102 static void	ellist_destroy(ellist_t *);
103 static void	ellist_insert(struct hfsc_class *);
104 static void	ellist_remove(struct hfsc_class *);
105 static void	ellist_update(struct hfsc_class *);
106 struct hfsc_class *ellist_get_mindl(ellist_t *, uint64_t);
107 static actlist_t *actlist_alloc(void);
108 static void	actlist_destroy(actlist_t *);
109 static void	actlist_insert(struct hfsc_class *);
110 static void	actlist_remove(struct hfsc_class *);
111 static void	actlist_update(struct hfsc_class *);
112 
113 static struct hfsc_class *actlist_firstfit(struct hfsc_class *, uint64_t);
114 
115 static __inline uint64_t	seg_x2y(uint64_t, uint64_t);
116 static __inline uint64_t	seg_y2x(uint64_t, uint64_t);
117 static __inline uint64_t	m2sm(u_int);
118 static __inline uint64_t	m2ism(u_int);
119 static __inline uint64_t	d2dx(u_int);
120 static u_int			sm2m(uint64_t);
121 static u_int			dx2d(uint64_t);
122 
123 static void	sc2isc(struct service_curve *, struct internal_sc *);
124 static void	rtsc_init(struct runtime_sc *, struct internal_sc *,
125 			  uint64_t, uint64_t);
126 static uint64_t	rtsc_y2x(struct runtime_sc *, uint64_t);
127 static uint64_t	rtsc_x2y(struct runtime_sc *, uint64_t);
128 static void	rtsc_min(struct runtime_sc *, struct internal_sc *,
129 			 uint64_t, uint64_t);
130 
131 static void	get_class_stats(struct hfsc_classstats *, struct hfsc_class *);
132 static struct hfsc_class *clh_to_clp(struct hfsc_if *, uint32_t);
133 
134 /*
135  * macros
136  */
137 #define	is_a_parent_class(cl)	((cl)->cl_children != NULL)
138 
139 #define	HT_INFINITY	0xffffffffffffffffLL	/* infinite time value */
140 
141 int
142 hfsc_pfattach(struct pf_altq *a, struct ifaltq *ifq)
143 {
144 	return altq_attach(ifq, ALTQT_HFSC, a->altq_disc,
145 	    hfsc_enqueue, hfsc_dequeue, hfsc_request, NULL, NULL);
146 }
147 
148 int
149 hfsc_add_altq(struct pf_altq *a)
150 {
151 	struct hfsc_if *hif;
152 	struct ifnet *ifp;
153 
154 	if ((ifp = ifunit(a->ifname)) == NULL)
155 		return (EINVAL);
156 	if (!ifq_is_ready(&ifp->if_snd))
157 		return (ENODEV);
158 
159 	hif = kmalloc(sizeof(struct hfsc_if), M_ALTQ, M_WAITOK | M_ZERO);
160 
161 	hif->hif_eligible = ellist_alloc();
162 	hif->hif_ifq = &ifp->if_snd;
163 	ifq_purge(&ifp->if_snd);
164 
165 	/* keep the state in pf_altq */
166 	a->altq_disc = hif;
167 
168 	return (0);
169 }
170 
171 int
172 hfsc_remove_altq(struct pf_altq *a)
173 {
174 	struct hfsc_if *hif;
175 
176 	if ((hif = a->altq_disc) == NULL)
177 		return (EINVAL);
178 	a->altq_disc = NULL;
179 
180 	hfsc_clear_interface(hif);
181 	hfsc_class_destroy(hif->hif_rootclass);
182 
183 	ellist_destroy(hif->hif_eligible);
184 
185 	kfree(hif, M_ALTQ);
186 
187 	return (0);
188 }
189 
190 static int
191 hfsc_add_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
192 {
193 	struct hfsc_class *cl, *parent;
194 	struct hfsc_opts *opts;
195 	struct service_curve rtsc, lssc, ulsc;
196 
197 	KKASSERT(a->qid != 0);
198 
199 	opts = &a->pq_u.hfsc_opts;
200 
201 	if (a->parent_qid == HFSC_NULLCLASS_HANDLE && hif->hif_rootclass == NULL)
202 		parent = NULL;
203 	else if ((parent = clh_to_clp(hif, a->parent_qid)) == NULL)
204 		return (EINVAL);
205 
206 	if (clh_to_clp(hif, a->qid) != NULL)
207 		return (EBUSY);
208 
209 	rtsc.m1 = opts->rtsc_m1;
210 	rtsc.d  = opts->rtsc_d;
211 	rtsc.m2 = opts->rtsc_m2;
212 	lssc.m1 = opts->lssc_m1;
213 	lssc.d  = opts->lssc_d;
214 	lssc.m2 = opts->lssc_m2;
215 	ulsc.m1 = opts->ulsc_m1;
216 	ulsc.d  = opts->ulsc_d;
217 	ulsc.m2 = opts->ulsc_m2;
218 
219 	cl = hfsc_class_create(hif, &rtsc, &lssc, &ulsc, parent, a->qlimit,
220 			       opts->flags, a->qid);
221 	if (cl == NULL)
222 		return (ENOMEM);
223 
224 	return (0);
225 }
226 
227 int
228 hfsc_add_queue(struct pf_altq *a)
229 {
230 	struct hfsc_if *hif;
231 	struct ifaltq *ifq;
232 	int error;
233 
234 	if (a->qid == 0)
235 		return (EINVAL);
236 
237 	/* XXX not MP safe */
238 	if ((hif = a->altq_disc) == NULL)
239 		return (EINVAL);
240 	ifq = hif->hif_ifq;
241 
242 	ALTQ_LOCK(ifq);
243 	error = hfsc_add_queue_locked(a, hif);
244 	ALTQ_UNLOCK(ifq);
245 
246 	return error;
247 }
248 
249 static int
250 hfsc_remove_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
251 {
252 	struct hfsc_class *cl;
253 
254 	if ((cl = clh_to_clp(hif, a->qid)) == NULL)
255 		return (EINVAL);
256 
257 	return (hfsc_class_destroy(cl));
258 }
259 
260 int
261 hfsc_remove_queue(struct pf_altq *a)
262 {
263 	struct hfsc_if *hif;
264 	struct ifaltq *ifq;
265 	int error;
266 
267 	/* XXX not MP safe */
268 	if ((hif = a->altq_disc) == NULL)
269 		return (EINVAL);
270 	ifq = hif->hif_ifq;
271 
272 	ALTQ_LOCK(ifq);
273 	error = hfsc_remove_queue_locked(a, hif);
274 	ALTQ_UNLOCK(ifq);
275 
276 	return error;
277 }
278 
279 int
280 hfsc_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
281 {
282 	struct hfsc_if *hif;
283 	struct hfsc_class *cl;
284 	struct hfsc_classstats stats;
285 	struct ifaltq *ifq;
286 	int error = 0;
287 
288 	if (*nbytes < sizeof(stats))
289 		return (EINVAL);
290 
291 	/* XXX not MP safe */
292 	if ((hif = altq_lookup(a->ifname, ALTQT_HFSC)) == NULL)
293 		return (EBADF);
294 	ifq = hif->hif_ifq;
295 
296 	ALTQ_LOCK(ifq);
297 
298 	if ((cl = clh_to_clp(hif, a->qid)) == NULL) {
299 		ALTQ_UNLOCK(ifq);
300 		return (EINVAL);
301 	}
302 
303 	get_class_stats(&stats, cl);
304 
305 	ALTQ_UNLOCK(ifq);
306 
307 	if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
308 		return (error);
309 	*nbytes = sizeof(stats);
310 	return (0);
311 }
312 
313 /*
314  * bring the interface back to the initial state by discarding
315  * all the filters and classes except the root class.
316  */
317 static int
318 hfsc_clear_interface(struct hfsc_if *hif)
319 {
320 	struct hfsc_class *cl;
321 
322 	if (hif->hif_rootclass == NULL)
323 		return (0);
324 
325 
326 	/* clear out the classes */
327 	while ((cl = hif->hif_rootclass->cl_children) != NULL) {
328 		/*
329 		 * remove the first leaf class found in the hierarchy
330 		 * then start over
331 		 */
332 		for (; cl != NULL; cl = hfsc_nextclass(cl)) {
333 			if (!is_a_parent_class(cl)) {
334 				hfsc_class_destroy(cl);
335 				break;
336 			}
337 		}
338 	}
339 
340 	return (0);
341 }
342 
343 static int
344 hfsc_request(struct ifaltq *ifq, int req, void *arg)
345 {
346 	struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
347 
348 	crit_enter();
349 	switch (req) {
350 	case ALTRQ_PURGE:
351 		hfsc_purge(hif);
352 		break;
353 	}
354 	crit_exit();
355 	return (0);
356 }
357 
358 /* discard all the queued packets on the interface */
359 static void
360 hfsc_purge(struct hfsc_if *hif)
361 {
362 	struct hfsc_class *cl;
363 
364 	for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl)) {
365 		if (!qempty(cl->cl_q))
366 			hfsc_purgeq(cl);
367 	}
368 	if (ifq_is_enabled(hif->hif_ifq))
369 		hif->hif_ifq->ifq_len = 0;
370 }
371 
372 struct hfsc_class *
373 hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
374 		  struct service_curve *fsc, struct service_curve *usc,
375 		  struct hfsc_class *parent, int qlimit, int flags, int qid)
376 {
377 	struct hfsc_class *cl, *p;
378 	int i;
379 
380 	if (hif->hif_classes >= HFSC_MAX_CLASSES)
381 		return (NULL);
382 
383 #ifndef ALTQ_RED
384 	if (flags & HFCF_RED) {
385 #ifdef ALTQ_DEBUG
386 		kprintf("hfsc_class_create: RED not configured for HFSC!\n");
387 #endif
388 		return (NULL);
389 	}
390 #endif
391 
392 	cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
393 	cl->cl_q = kmalloc(sizeof(*cl->cl_q), M_ALTQ, M_WAITOK | M_ZERO);
394 	cl->cl_actc = actlist_alloc();
395 
396 	if (qlimit == 0)
397 		qlimit = 50;  /* use default */
398 	qlimit(cl->cl_q) = qlimit;
399 	qtype(cl->cl_q) = Q_DROPTAIL;
400 	qlen(cl->cl_q) = 0;
401 	cl->cl_flags = flags;
402 #ifdef ALTQ_RED
403 	if (flags & (HFCF_RED|HFCF_RIO)) {
404 		int red_flags, red_pkttime;
405 		u_int m2;
406 
407 		m2 = 0;
408 		if (rsc != NULL && rsc->m2 > m2)
409 			m2 = rsc->m2;
410 		if (fsc != NULL && fsc->m2 > m2)
411 			m2 = fsc->m2;
412 		if (usc != NULL && usc->m2 > m2)
413 			m2 = usc->m2;
414 
415 		red_flags = 0;
416 		if (flags & HFCF_ECN)
417 			red_flags |= REDF_ECN;
418 #ifdef ALTQ_RIO
419 		if (flags & HFCF_CLEARDSCP)
420 			red_flags |= RIOF_CLEARDSCP;
421 #endif
422 		if (m2 < 8)
423 			red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
424 		else
425 			red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
426 				* 1000 * 1000 * 1000 / (m2 / 8);
427 		if (flags & HFCF_RED) {
428 			cl->cl_red = red_alloc(0, 0,
429 			    qlimit(cl->cl_q) * 10/100,
430 			    qlimit(cl->cl_q) * 30/100,
431 			    red_flags, red_pkttime);
432 			if (cl->cl_red != NULL)
433 				qtype(cl->cl_q) = Q_RED;
434 		}
435 #ifdef ALTQ_RIO
436 		else {
437 			cl->cl_red = (red_t *)rio_alloc(0, NULL,
438 			    red_flags, red_pkttime);
439 			if (cl->cl_red != NULL)
440 				qtype(cl->cl_q) = Q_RIO;
441 		}
442 #endif
443 	}
444 #endif /* ALTQ_RED */
445 
446 	if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
447 		cl->cl_rsc = kmalloc(sizeof(*cl->cl_rsc), M_ALTQ, M_WAITOK);
448 		sc2isc(rsc, cl->cl_rsc);
449 		rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
450 		rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
451 	}
452 	if (fsc != NULL && (fsc->m1 != 0 || fsc->m2 != 0)) {
453 		cl->cl_fsc = kmalloc(sizeof(*cl->cl_fsc), M_ALTQ, M_WAITOK);
454 		if (cl->cl_fsc == NULL)
455 			goto err_ret;
456 		sc2isc(fsc, cl->cl_fsc);
457 		rtsc_init(&cl->cl_virtual, cl->cl_fsc, 0, 0);
458 	}
459 	if (usc != NULL && (usc->m1 != 0 || usc->m2 != 0)) {
460 		cl->cl_usc = kmalloc(sizeof(*cl->cl_usc), M_ALTQ, M_WAITOK);
461 		if (cl->cl_usc == NULL)
462 			goto err_ret;
463 		sc2isc(usc, cl->cl_usc);
464 		rtsc_init(&cl->cl_ulimit, cl->cl_usc, 0, 0);
465 	}
466 
467 	cl->cl_id = hif->hif_classid++;
468 	cl->cl_handle = qid;
469 	cl->cl_hif = hif;
470 	cl->cl_parent = parent;
471 
472 	crit_enter();
473 	hif->hif_classes++;
474 
475 	/*
476 	 * find a free slot in the class table.  if the slot matching
477 	 * the lower bits of qid is free, use this slot.  otherwise,
478 	 * use the first free slot.
479 	 */
480 	i = qid % HFSC_MAX_CLASSES;
481 	if (hif->hif_class_tbl[i] == NULL)
482 		hif->hif_class_tbl[i] = cl;
483 	else {
484 		for (i = 0; i < HFSC_MAX_CLASSES; i++) {
485 			if (hif->hif_class_tbl[i] == NULL) {
486 				hif->hif_class_tbl[i] = cl;
487 				break;
488 			}
489 		}
490 		if (i == HFSC_MAX_CLASSES) {
491 			crit_exit();
492 			goto err_ret;
493 		}
494 	}
495 
496 	if (flags & HFCF_DEFAULTCLASS)
497 		hif->hif_defaultclass = cl;
498 
499 	if (parent == NULL) {
500 		/* this is root class */
501 		hif->hif_rootclass = cl;
502 	} else if (parent->cl_children == NULL) {
503 		/* add this class to the children list of the parent */
504 		parent->cl_children = cl;
505 	} else {
506 		p = parent->cl_children;
507 		while (p->cl_siblings != NULL)
508 			p = p->cl_siblings;
509 		p->cl_siblings = cl;
510 	}
511 	crit_exit();
512 
513 	return (cl);
514 
515  err_ret:
516 	if (cl->cl_actc != NULL)
517 		actlist_destroy(cl->cl_actc);
518 	if (cl->cl_red != NULL) {
519 #ifdef ALTQ_RIO
520 		if (q_is_rio(cl->cl_q))
521 			rio_destroy((rio_t *)cl->cl_red);
522 #endif
523 #ifdef ALTQ_RED
524 		if (q_is_red(cl->cl_q))
525 			red_destroy(cl->cl_red);
526 #endif
527 	}
528 	if (cl->cl_fsc != NULL)
529 		kfree(cl->cl_fsc, M_ALTQ);
530 	if (cl->cl_rsc != NULL)
531 		kfree(cl->cl_rsc, M_ALTQ);
532 	if (cl->cl_usc != NULL)
533 		kfree(cl->cl_usc, M_ALTQ);
534 	if (cl->cl_q != NULL)
535 		kfree(cl->cl_q, M_ALTQ);
536 	kfree(cl, M_ALTQ);
537 	return (NULL);
538 }
539 
540 static int
541 hfsc_class_destroy(struct hfsc_class *cl)
542 {
543 	int i;
544 
545 	if (cl == NULL)
546 		return (0);
547 
548 	if (is_a_parent_class(cl))
549 		return (EBUSY);
550 
551 	crit_enter();
552 
553 	if (!qempty(cl->cl_q))
554 		hfsc_purgeq(cl);
555 
556 	if (cl->cl_parent == NULL) {
557 		/* this is root class */
558 	} else {
559 		struct hfsc_class *p = cl->cl_parent->cl_children;
560 
561 		if (p == cl) {
562 			cl->cl_parent->cl_children = cl->cl_siblings;
563 		} else {
564 			do {
565 				if (p->cl_siblings == cl) {
566 					p->cl_siblings = cl->cl_siblings;
567 					break;
568 				}
569 			} while ((p = p->cl_siblings) != NULL);
570 		}
571 		KKASSERT(p != NULL);
572 	}
573 
574 	for (i = 0; i < HFSC_MAX_CLASSES; i++) {
575 		if (cl->cl_hif->hif_class_tbl[i] == cl) {
576 			cl->cl_hif->hif_class_tbl[i] = NULL;
577 			break;
578 		}
579 	}
580 
581 	cl->cl_hif->hif_classes--;
582 	crit_exit();
583 
584 	actlist_destroy(cl->cl_actc);
585 
586 	if (cl->cl_red != NULL) {
587 #ifdef ALTQ_RIO
588 		if (q_is_rio(cl->cl_q))
589 			rio_destroy((rio_t *)cl->cl_red);
590 #endif
591 #ifdef ALTQ_RED
592 		if (q_is_red(cl->cl_q))
593 			red_destroy(cl->cl_red);
594 #endif
595 	}
596 
597 	if (cl == cl->cl_hif->hif_rootclass)
598 		cl->cl_hif->hif_rootclass = NULL;
599 	if (cl == cl->cl_hif->hif_defaultclass)
600 		cl->cl_hif->hif_defaultclass = NULL;
601 
602 	if (cl->cl_usc != NULL)
603 		kfree(cl->cl_usc, M_ALTQ);
604 	if (cl->cl_fsc != NULL)
605 		kfree(cl->cl_fsc, M_ALTQ);
606 	if (cl->cl_rsc != NULL)
607 		kfree(cl->cl_rsc, M_ALTQ);
608 	kfree(cl->cl_q, M_ALTQ);
609 	kfree(cl, M_ALTQ);
610 
611 	return (0);
612 }
613 
614 /*
615  * hfsc_nextclass returns the next class in the tree.
616  *   usage:
617  *	for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl))
618  *		do_something;
619  */
620 static struct hfsc_class *
621 hfsc_nextclass(struct hfsc_class *cl)
622 {
623 	if (cl->cl_children != NULL) {
624 		cl = cl->cl_children;
625 	} else if (cl->cl_siblings != NULL) {
626 		cl = cl->cl_siblings;
627 	} else {
628 		while ((cl = cl->cl_parent) != NULL) {
629 			if (cl->cl_siblings != NULL) {
630 				cl = cl->cl_siblings;
631 				break;
632 			}
633 		}
634 	}
635 
636 	return (cl);
637 }
638 
639 /*
640  * hfsc_enqueue is an enqueue function to be registered to
641  * (*altq_enqueue) in struct ifaltq.
642  */
643 static int
644 hfsc_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
645 {
646 	struct hfsc_if	*hif = (struct hfsc_if *)ifq->altq_disc;
647 	struct hfsc_class *cl;
648 	int len;
649 
650 	/* grab class set by classifier */
651 	if ((m->m_flags & M_PKTHDR) == 0) {
652 		/* should not happen */
653 		if_printf(ifq->altq_ifp, "altq: packet does not have pkthdr\n");
654 		m_freem(m);
655 		return (ENOBUFS);
656 	}
657 	crit_enter();
658 	if (m->m_pkthdr.fw_flags & ALTQ_MBUF_TAGGED)
659 		cl = clh_to_clp(hif, m->m_pkthdr.altq_qid);
660 	else
661 		cl = NULL;
662 	if (cl == NULL || is_a_parent_class(cl)) {
663 		cl = hif->hif_defaultclass;
664 		if (cl == NULL) {
665 			m_freem(m);
666 			crit_exit();
667 			return (ENOBUFS);
668 		}
669 	}
670 	cl->cl_pktattr = NULL;
671 	len = m_pktlen(m);
672 	if (hfsc_addq(cl, m) != 0) {
673 		/* drop occurred.  mbuf was freed in hfsc_addq. */
674 		PKTCNTR_ADD(&cl->cl_stats.drop_cnt, len);
675 		crit_exit();
676 		return (ENOBUFS);
677 	}
678 	ifq->ifq_len++;
679 	cl->cl_hif->hif_packets++;
680 
681 	/* successfully queued. */
682 	if (qlen(cl->cl_q) == 1)
683 		set_active(cl, m_pktlen(m));
684 	crit_exit();
685 	return (0);
686 }
687 
688 /*
689  * hfsc_dequeue is a dequeue function to be registered to
690  * (*altq_dequeue) in struct ifaltq.
691  *
692  * note: ALTDQ_POLL returns the next packet without removing the packet
693  *	from the queue.  ALTDQ_REMOVE is a normal dequeue operation.
694  *	ALTDQ_REMOVE must return the same packet if called immediately
695  *	after ALTDQ_POLL.
696  */
697 static struct mbuf *
698 hfsc_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
699 {
700 	struct hfsc_if	*hif = (struct hfsc_if *)ifq->altq_disc;
701 	struct hfsc_class *cl;
702 	struct mbuf *m;
703 	int len, next_len;
704 	int realtime = 0;
705 	uint64_t cur_time;
706 
707 	if (hif->hif_packets == 0) {
708 		/* no packet in the tree */
709 		return (NULL);
710 	}
711 
712 	crit_enter();
713 	cur_time = read_machclk();
714 
715 	if (op == ALTDQ_REMOVE && hif->hif_pollcache != NULL) {
716 		cl = hif->hif_pollcache;
717 		hif->hif_pollcache = NULL;
718 		/* check if the class was scheduled by real-time criteria */
719 		if (cl->cl_rsc != NULL)
720 			realtime = (cl->cl_e <= cur_time);
721 	} else {
722 		/*
723 		 * if there are eligible classes, use real-time criteria.
724 		 * find the class with the minimum deadline among
725 		 * the eligible classes.
726 		 */
727 		if ((cl = ellist_get_mindl(hif->hif_eligible, cur_time)) != NULL) {
728 			realtime = 1;
729 		} else {
730 #ifdef ALTQ_DEBUG
731 			int fits = 0;
732 #endif
733 			/*
734 			 * use link-sharing criteria
735 			 * get the class with the minimum vt in the hierarchy
736 			 */
737 			cl = hif->hif_rootclass;
738 			while (is_a_parent_class(cl)) {
739 
740 				cl = actlist_firstfit(cl, cur_time);
741 				if (cl == NULL) {
742 #ifdef ALTQ_DEBUG
743 					if (fits > 0)
744 						kprintf("%d fit but none found\n",fits);
745 #endif
746 					m = NULL;
747 					goto done;
748 				}
749 				/*
750 				 * update parent's cl_cvtmin.
751 				 * don't update if the new vt is smaller.
752 				 */
753 				if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
754 					cl->cl_parent->cl_cvtmin = cl->cl_vt;
755 #ifdef ALTQ_DEBUG
756 				fits++;
757 #endif
758 			}
759 		}
760 
761 		if (op == ALTDQ_POLL) {
762 			hif->hif_pollcache = cl;
763 			m = hfsc_pollq(cl);
764 			goto done;
765 		}
766 	}
767 
768 	m = hfsc_getq(cl);
769 	if (m == NULL)
770 		panic("hfsc_dequeue:");
771 	len = m_pktlen(m);
772 	cl->cl_hif->hif_packets--;
773 	ifq->ifq_len--;
774 	PKTCNTR_ADD(&cl->cl_stats.xmit_cnt, len);
775 
776 	update_vf(cl, len, cur_time);
777 	if (realtime)
778 		cl->cl_cumul += len;
779 
780 	if (!qempty(cl->cl_q)) {
781 		if (cl->cl_rsc != NULL) {
782 			/* update ed */
783 			next_len = m_pktlen(qhead(cl->cl_q));
784 
785 			if (realtime)
786 				update_ed(cl, next_len);
787 			else
788 				update_d(cl, next_len);
789 		}
790 	} else {
791 		/* the class becomes passive */
792 		set_passive(cl);
793 	}
794 done:
795 	crit_exit();
796 	KKASSERT(mpolled == NULL || m == mpolled);
797 	return (m);
798 }
799 
800 static int
801 hfsc_addq(struct hfsc_class *cl, struct mbuf *m)
802 {
803 
804 #ifdef ALTQ_RIO
805 	if (q_is_rio(cl->cl_q))
806 		return rio_addq((rio_t *)cl->cl_red, cl->cl_q,
807 				m, cl->cl_pktattr);
808 #endif
809 #ifdef ALTQ_RED
810 	if (q_is_red(cl->cl_q))
811 		return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
812 #endif
813 	if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
814 		m_freem(m);
815 		return (-1);
816 	}
817 
818 	if (cl->cl_flags & HFCF_CLEARDSCP)
819 		write_dsfield(m, cl->cl_pktattr, 0);
820 
821 	_addq(cl->cl_q, m);
822 
823 	return (0);
824 }
825 
826 static struct mbuf *
827 hfsc_getq(struct hfsc_class *cl)
828 {
829 #ifdef ALTQ_RIO
830 	if (q_is_rio(cl->cl_q))
831 		return rio_getq((rio_t *)cl->cl_red, cl->cl_q);
832 #endif
833 #ifdef ALTQ_RED
834 	if (q_is_red(cl->cl_q))
835 		return red_getq(cl->cl_red, cl->cl_q);
836 #endif
837 	return _getq(cl->cl_q);
838 }
839 
840 static struct mbuf *
841 hfsc_pollq(struct hfsc_class *cl)
842 {
843 	return qhead(cl->cl_q);
844 }
845 
846 static void
847 hfsc_purgeq(struct hfsc_class *cl)
848 {
849 	struct mbuf *m;
850 
851 	if (qempty(cl->cl_q))
852 		return;
853 
854 	while ((m = _getq(cl->cl_q)) != NULL) {
855 		PKTCNTR_ADD(&cl->cl_stats.drop_cnt, m_pktlen(m));
856 		m_freem(m);
857 		cl->cl_hif->hif_packets--;
858 		cl->cl_hif->hif_ifq->ifq_len--;
859 	}
860 	KKASSERT(qlen(cl->cl_q) == 0);
861 
862 	update_vf(cl, 0, 0);	/* remove cl from the actlist */
863 	set_passive(cl);
864 }
865 
866 static void
867 set_active(struct hfsc_class *cl, int len)
868 {
869 	if (cl->cl_rsc != NULL)
870 		init_ed(cl, len);
871 	if (cl->cl_fsc != NULL)
872 		init_vf(cl, len);
873 
874 	cl->cl_stats.period++;
875 }
876 
877 static void
878 set_passive(struct hfsc_class *cl)
879 {
880 	if (cl->cl_rsc != NULL)
881 		ellist_remove(cl);
882 
883 	/*
884 	 * actlist is now handled in update_vf() so that update_vf(cl, 0, 0)
885 	 * needs to be called explicitly to remove a class from actlist
886 	 */
887 }
888 
889 static void
890 init_ed(struct hfsc_class *cl, int next_len)
891 {
892 	uint64_t cur_time;
893 
894 	cur_time = read_machclk();
895 
896 	/* update the deadline curve */
897 	rtsc_min(&cl->cl_deadline, cl->cl_rsc, cur_time, cl->cl_cumul);
898 
899 	/*
900 	 * update the eligible curve.
901 	 * for concave, it is equal to the deadline curve.
902 	 * for convex, it is a linear curve with slope m2.
903 	 */
904 	cl->cl_eligible = cl->cl_deadline;
905 	if (cl->cl_rsc->sm1 <= cl->cl_rsc->sm2) {
906 		cl->cl_eligible.dx = 0;
907 		cl->cl_eligible.dy = 0;
908 	}
909 
910 	/* compute e and d */
911 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
912 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
913 
914 	ellist_insert(cl);
915 }
916 
917 static void
918 update_ed(struct hfsc_class *cl, int next_len)
919 {
920 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
921 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
922 
923 	ellist_update(cl);
924 }
925 
926 static void
927 update_d(struct hfsc_class *cl, int next_len)
928 {
929 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
930 }
931 
932 static void
933 init_vf(struct hfsc_class *cl, int len)
934 {
935 	struct hfsc_class *max_cl, *p;
936 	uint64_t vt, f, cur_time;
937 	int go_active;
938 
939 	cur_time = 0;
940 	go_active = 1;
941 	for ( ; cl->cl_parent != NULL; cl = cl->cl_parent) {
942 		if (go_active && cl->cl_nactive++ == 0)
943 			go_active = 1;
944 		else
945 			go_active = 0;
946 
947 		if (go_active) {
948 			max_cl = actlist_last(cl->cl_parent->cl_actc);
949 			if (max_cl != NULL) {
950 				/*
951 				 * set vt to the average of the min and max
952 				 * classes.  if the parent's period didn't
953 				 * change, don't decrease vt of the class.
954 				 */
955 				vt = max_cl->cl_vt;
956 				if (cl->cl_parent->cl_cvtmin != 0)
957 					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
958 
959 				if (cl->cl_parent->cl_vtperiod !=
960 				    cl->cl_parentperiod || vt > cl->cl_vt)
961 					cl->cl_vt = vt;
962 			} else {
963 				/*
964 				 * first child for a new parent backlog period.
965 				 * add parent's cvtmax to vtoff of children
966 				 * to make a new vt (vtoff + vt) larger than
967 				 * the vt in the last period for all children.
968 				 */
969 				vt = cl->cl_parent->cl_cvtmax;
970 				for (p = cl->cl_parent->cl_children; p != NULL;
971 				     p = p->cl_siblings)
972 					p->cl_vtoff += vt;
973 				cl->cl_vt = 0;
974 				cl->cl_parent->cl_cvtmax = 0;
975 				cl->cl_parent->cl_cvtmin = 0;
976 			}
977 			cl->cl_initvt = cl->cl_vt;
978 
979 			/* update the virtual curve */
980 			vt = cl->cl_vt + cl->cl_vtoff;
981 			rtsc_min(&cl->cl_virtual, cl->cl_fsc, vt, cl->cl_total);
982 			if (cl->cl_virtual.x == vt) {
983 				cl->cl_virtual.x -= cl->cl_vtoff;
984 				cl->cl_vtoff = 0;
985 			}
986 			cl->cl_vtadj = 0;
987 
988 			cl->cl_vtperiod++;  /* increment vt period */
989 			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
990 			if (cl->cl_parent->cl_nactive == 0)
991 				cl->cl_parentperiod++;
992 			cl->cl_f = 0;
993 
994 			actlist_insert(cl);
995 
996 			if (cl->cl_usc != NULL) {
997 				/* class has upper limit curve */
998 				if (cur_time == 0)
999 					cur_time = read_machclk();
1000 
1001 				/* update the ulimit curve */
1002 				rtsc_min(&cl->cl_ulimit, cl->cl_usc, cur_time,
1003 				    cl->cl_total);
1004 				/* compute myf */
1005 				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
1006 				    cl->cl_total);
1007 				cl->cl_myfadj = 0;
1008 			}
1009 		}
1010 
1011 		if (cl->cl_myf > cl->cl_cfmin)
1012 			f = cl->cl_myf;
1013 		else
1014 			f = cl->cl_cfmin;
1015 		if (f != cl->cl_f) {
1016 			cl->cl_f = f;
1017 			update_cfmin(cl->cl_parent);
1018 		}
1019 	}
1020 }
1021 
1022 static void
1023 update_vf(struct hfsc_class *cl, int len, uint64_t cur_time)
1024 {
1025 	uint64_t f, myf_bound, delta;
1026 	int go_passive;
1027 
1028 	go_passive = qempty(cl->cl_q);
1029 
1030 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
1031 		cl->cl_total += len;
1032 
1033 		if (cl->cl_fsc == NULL || cl->cl_nactive == 0)
1034 			continue;
1035 
1036 		if (go_passive && --cl->cl_nactive == 0)
1037 			go_passive = 1;
1038 		else
1039 			go_passive = 0;
1040 
1041 		if (go_passive) {
1042 			/* no more active child, going passive */
1043 
1044 			/* update cvtmax of the parent class */
1045 			if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
1046 				cl->cl_parent->cl_cvtmax = cl->cl_vt;
1047 
1048 			/* remove this class from the vt list */
1049 			actlist_remove(cl);
1050 
1051 			update_cfmin(cl->cl_parent);
1052 
1053 			continue;
1054 		}
1055 
1056 		/*
1057 		 * update vt and f
1058 		 */
1059 		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
1060 		    - cl->cl_vtoff + cl->cl_vtadj;
1061 
1062 		/*
1063 		 * if vt of the class is smaller than cvtmin,
1064 		 * the class was skipped in the past due to non-fit.
1065 		 * if so, we need to adjust vtadj.
1066 		 */
1067 		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
1068 			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
1069 			cl->cl_vt = cl->cl_parent->cl_cvtmin;
1070 		}
1071 
1072 		/* update the vt list */
1073 		actlist_update(cl);
1074 
1075 		if (cl->cl_usc != NULL) {
1076 			cl->cl_myf = cl->cl_myfadj
1077 			    + rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
1078 
1079 			/*
1080 			 * if myf lags behind by more than one clock tick
1081 			 * from the current time, adjust myfadj to prevent
1082 			 * a rate-limited class from going greedy.
1083 			 * in a steady state under rate-limiting, myf
1084 			 * fluctuates within one clock tick.
1085 			 */
1086 			myf_bound = cur_time - machclk_per_tick;
1087 			if (cl->cl_myf < myf_bound) {
1088 				delta = cur_time - cl->cl_myf;
1089 				cl->cl_myfadj += delta;
1090 				cl->cl_myf += delta;
1091 			}
1092 		}
1093 
1094 		/* cl_f is max(cl_myf, cl_cfmin) */
1095 		if (cl->cl_myf > cl->cl_cfmin)
1096 			f = cl->cl_myf;
1097 		else
1098 			f = cl->cl_cfmin;
1099 		if (f != cl->cl_f) {
1100 			cl->cl_f = f;
1101 			update_cfmin(cl->cl_parent);
1102 		}
1103 	}
1104 }
1105 
1106 static void
1107 update_cfmin(struct hfsc_class *cl)
1108 {
1109 	struct hfsc_class *p;
1110 	uint64_t cfmin;
1111 
1112 	if (TAILQ_EMPTY(cl->cl_actc)) {
1113 		cl->cl_cfmin = 0;
1114 		return;
1115 	}
1116 	cfmin = HT_INFINITY;
1117 	TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1118 		if (p->cl_f == 0) {
1119 			cl->cl_cfmin = 0;
1120 			return;
1121 		}
1122 		if (p->cl_f < cfmin)
1123 			cfmin = p->cl_f;
1124 	}
1125 	cl->cl_cfmin = cfmin;
1126 }
1127 
1128 /*
1129  * TAILQ based ellist and actlist implementation
1130  * (ion wanted to make a calendar queue based implementation)
1131  */
1132 /*
1133  * eligible list holds backlogged classes being sorted by their eligible times.
1134  * there is one eligible list per interface.
1135  */
1136 
1137 static ellist_t *
1138 ellist_alloc(void)
1139 {
1140 	ellist_t *head;
1141 
1142 	head = kmalloc(sizeof(ellist_t *), M_ALTQ, M_WAITOK);
1143 	TAILQ_INIT(head);
1144 	return (head);
1145 }
1146 
1147 static void
1148 ellist_destroy(ellist_t *head)
1149 {
1150 	kfree(head, M_ALTQ);
1151 }
1152 
1153 static void
1154 ellist_insert(struct hfsc_class *cl)
1155 {
1156 	struct hfsc_if *hif = cl->cl_hif;
1157 	struct hfsc_class *p;
1158 
1159 	/* check the last entry first */
1160 	if ((p = TAILQ_LAST(hif->hif_eligible, _eligible)) == NULL ||
1161 	    p->cl_e <= cl->cl_e) {
1162 		TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1163 		return;
1164 	}
1165 
1166 	TAILQ_FOREACH(p, hif->hif_eligible, cl_ellist) {
1167 		if (cl->cl_e < p->cl_e) {
1168 			TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1169 			return;
1170 		}
1171 	}
1172 	KKASSERT(0); /* should not reach here */
1173 }
1174 
1175 static void
1176 ellist_remove(struct hfsc_class *cl)
1177 {
1178 	struct hfsc_if *hif = cl->cl_hif;
1179 
1180 	TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1181 }
1182 
1183 static void
1184 ellist_update(struct hfsc_class *cl)
1185 {
1186 	struct hfsc_if *hif = cl->cl_hif;
1187 	struct hfsc_class *p, *last;
1188 
1189 	/*
1190 	 * the eligible time of a class increases monotonically.
1191 	 * if the next entry has a larger eligible time, nothing to do.
1192 	 */
1193 	p = TAILQ_NEXT(cl, cl_ellist);
1194 	if (p == NULL || cl->cl_e <= p->cl_e)
1195 		return;
1196 
1197 	/* check the last entry */
1198 	last = TAILQ_LAST(hif->hif_eligible, _eligible);
1199 	KKASSERT(last != NULL);
1200 	if (last->cl_e <= cl->cl_e) {
1201 		TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1202 		TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1203 		return;
1204 	}
1205 
1206 	/*
1207 	 * the new position must be between the next entry
1208 	 * and the last entry
1209 	 */
1210 	while ((p = TAILQ_NEXT(p, cl_ellist)) != NULL) {
1211 		if (cl->cl_e < p->cl_e) {
1212 			TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1213 			TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1214 			return;
1215 		}
1216 	}
1217 	KKASSERT(0); /* should not reach here */
1218 }
1219 
1220 /* find the class with the minimum deadline among the eligible classes */
1221 struct hfsc_class *
1222 ellist_get_mindl(ellist_t *head, uint64_t cur_time)
1223 {
1224 	struct hfsc_class *p, *cl = NULL;
1225 
1226 	TAILQ_FOREACH(p, head, cl_ellist) {
1227 		if (p->cl_e > cur_time)
1228 			break;
1229 		if (cl == NULL || p->cl_d < cl->cl_d)
1230 			cl = p;
1231 	}
1232 	return (cl);
1233 }
1234 
1235 /*
1236  * active children list holds backlogged child classes being sorted
1237  * by their virtual time.
1238  * each intermediate class has one active children list.
1239  */
1240 static actlist_t *
1241 actlist_alloc(void)
1242 {
1243 	actlist_t *head;
1244 
1245 	head = kmalloc(sizeof(*head), M_ALTQ, M_WAITOK);
1246 	TAILQ_INIT(head);
1247 	return (head);
1248 }
1249 
1250 static void
1251 actlist_destroy(actlist_t *head)
1252 {
1253 	kfree(head, M_ALTQ);
1254 }
1255 static void
1256 actlist_insert(struct hfsc_class *cl)
1257 {
1258 	struct hfsc_class *p;
1259 
1260 	/* check the last entry first */
1261 	if ((p = TAILQ_LAST(cl->cl_parent->cl_actc, _active)) == NULL
1262 	    || p->cl_vt <= cl->cl_vt) {
1263 		TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1264 		return;
1265 	}
1266 
1267 	TAILQ_FOREACH(p, cl->cl_parent->cl_actc, cl_actlist) {
1268 		if (cl->cl_vt < p->cl_vt) {
1269 			TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1270 			return;
1271 		}
1272 	}
1273 	KKASSERT(0); /* should not reach here */
1274 }
1275 
1276 static void
1277 actlist_remove(struct hfsc_class *cl)
1278 {
1279 	TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1280 }
1281 
1282 static void
1283 actlist_update(struct hfsc_class *cl)
1284 {
1285 	struct hfsc_class *p, *last;
1286 
1287 	/*
1288 	 * the virtual time of a class increases monotonically during its
1289 	 * backlogged period.
1290 	 * if the next entry has a larger virtual time, nothing to do.
1291 	 */
1292 	p = TAILQ_NEXT(cl, cl_actlist);
1293 	if (p == NULL || cl->cl_vt < p->cl_vt)
1294 		return;
1295 
1296 	/* check the last entry */
1297 	last = TAILQ_LAST(cl->cl_parent->cl_actc, _active);
1298 	KKASSERT(last != NULL);
1299 	if (last->cl_vt <= cl->cl_vt) {
1300 		TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1301 		TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1302 		return;
1303 	}
1304 
1305 	/*
1306 	 * the new position must be between the next entry
1307 	 * and the last entry
1308 	 */
1309 	while ((p = TAILQ_NEXT(p, cl_actlist)) != NULL) {
1310 		if (cl->cl_vt < p->cl_vt) {
1311 			TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1312 			TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1313 			return;
1314 		}
1315 	}
1316 	KKASSERT(0); /* should not reach here */
1317 }
1318 
1319 static struct hfsc_class *
1320 actlist_firstfit(struct hfsc_class *cl, uint64_t cur_time)
1321 {
1322 	struct hfsc_class *p;
1323 
1324 	TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1325 		if (p->cl_f <= cur_time)
1326 			return (p);
1327 	}
1328 	return (NULL);
1329 }
1330 
1331 /*
1332  * service curve support functions
1333  *
1334  *  external service curve parameters
1335  *	m: bits/sec
1336  *	d: msec
1337  *  internal service curve parameters
1338  *	sm: (bytes/tsc_interval) << SM_SHIFT
1339  *	ism: (tsc_count/byte) << ISM_SHIFT
1340  *	dx: tsc_count
1341  *
1342  * SM_SHIFT and ISM_SHIFT are scaled in order to keep effective digits.
1343  * we should be able to handle 100K-1Gbps linkspeed with 200Hz-1GHz CPU
1344  * speed.  SM_SHIFT and ISM_SHIFT are selected to have at least 3 effective
1345  * digits in decimal using the following table.
1346  *
1347  *  bits/sec    100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
1348  *  ----------+-------------------------------------------------------
1349  *  bytes/nsec  12.5e-6    125e-6     1250e-6    12500e-6   125000e-6
1350  *  sm(500MHz)  25.0e-6    250e-6     2500e-6    25000e-6   250000e-6
1351  *  sm(200MHz)  62.5e-6    625e-6     6250e-6    62500e-6   625000e-6
1352  *
1353  *  nsec/byte   80000      8000       800        80         8
1354  *  ism(500MHz) 40000      4000       400        40         4
1355  *  ism(200MHz) 16000      1600       160        16         1.6
1356  */
1357 #define	SM_SHIFT	24
1358 #define	ISM_SHIFT	10
1359 
1360 #define	SM_MASK		((1LL << SM_SHIFT) - 1)
1361 #define	ISM_MASK	((1LL << ISM_SHIFT) - 1)
1362 
1363 static __inline uint64_t
1364 seg_x2y(uint64_t x, uint64_t sm)
1365 {
1366 	uint64_t y;
1367 
1368 	/*
1369 	 * compute
1370 	 *	y = x * sm >> SM_SHIFT
1371 	 * but divide it for the upper and lower bits to avoid overflow
1372 	 */
1373 	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
1374 	return (y);
1375 }
1376 
1377 static __inline uint64_t
1378 seg_y2x(uint64_t y, uint64_t ism)
1379 {
1380 	uint64_t x;
1381 
1382 	if (y == 0)
1383 		x = 0;
1384 	else if (ism == HT_INFINITY)
1385 		x = HT_INFINITY;
1386 	else
1387 		x = (y >> ISM_SHIFT) * ism + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
1388 
1389 	return (x);
1390 }
1391 
1392 static __inline uint64_t
1393 m2sm(u_int m)
1394 {
1395 	uint64_t sm;
1396 
1397 	sm = ((uint64_t)m << SM_SHIFT) / 8 / machclk_freq;
1398 	return (sm);
1399 }
1400 
1401 static __inline uint64_t
1402 m2ism(u_int m)
1403 {
1404 	uint64_t ism;
1405 
1406 	if (m == 0)
1407 		ism = HT_INFINITY;
1408 	else
1409 		ism = ((uint64_t)machclk_freq << ISM_SHIFT) * 8 / m;
1410 	return (ism);
1411 }
1412 
1413 static __inline uint64_t
1414 d2dx(u_int d)
1415 {
1416 	uint64_t dx;
1417 
1418 	dx = ((uint64_t)d * machclk_freq) / 1000;
1419 	return (dx);
1420 }
1421 
1422 static u_int
1423 sm2m(uint64_t sm)
1424 {
1425 	uint64_t m;
1426 
1427 	m = (sm * 8 * machclk_freq) >> SM_SHIFT;
1428 	return ((u_int)m);
1429 }
1430 
1431 static u_int
1432 dx2d(uint64_t dx)
1433 {
1434 	uint64_t d;
1435 
1436 	d = dx * 1000 / machclk_freq;
1437 	return ((u_int)d);
1438 }
1439 
1440 static void
1441 sc2isc(struct service_curve *sc, struct internal_sc *isc)
1442 {
1443 	isc->sm1 = m2sm(sc->m1);
1444 	isc->ism1 = m2ism(sc->m1);
1445 	isc->dx = d2dx(sc->d);
1446 	isc->dy = seg_x2y(isc->dx, isc->sm1);
1447 	isc->sm2 = m2sm(sc->m2);
1448 	isc->ism2 = m2ism(sc->m2);
1449 }
1450 
1451 /*
1452  * initialize the runtime service curve with the given internal
1453  * service curve starting at (x, y).
1454  */
1455 static void
1456 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1457 {
1458 	rtsc->x = x;
1459 	rtsc->y = y;
1460 	rtsc->sm1 = isc->sm1;
1461 	rtsc->ism1 = isc->ism1;
1462 	rtsc->dx = isc->dx;
1463 	rtsc->dy = isc->dy;
1464 	rtsc->sm2 = isc->sm2;
1465 	rtsc->ism2 = isc->ism2;
1466 }
1467 
1468 /*
1469  * calculate the y-projection of the runtime service curve by the
1470  * given x-projection value
1471  */
1472 static uint64_t
1473 rtsc_y2x(struct runtime_sc *rtsc, uint64_t y)
1474 {
1475 	uint64_t x;
1476 
1477 	if (y < rtsc->y) {
1478 		x = rtsc->x;
1479 	} else if (y <= rtsc->y + rtsc->dy) {
1480 		/* x belongs to the 1st segment */
1481 		if (rtsc->dy == 0)
1482 			x = rtsc->x + rtsc->dx;
1483 		else
1484 			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
1485 	} else {
1486 		/* x belongs to the 2nd segment */
1487 		x = rtsc->x + rtsc->dx
1488 		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
1489 	}
1490 	return (x);
1491 }
1492 
1493 static uint64_t
1494 rtsc_x2y(struct runtime_sc *rtsc, uint64_t x)
1495 {
1496 	uint64_t y;
1497 
1498 	if (x <= rtsc->x) {
1499 		y = rtsc->y;
1500 	} else if (x <= rtsc->x + rtsc->dx) {
1501 		/* y belongs to the 1st segment */
1502 		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
1503 	} else
1504 		/* y belongs to the 2nd segment */
1505 		y = rtsc->y + rtsc->dy
1506 		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
1507 	return (y);
1508 }
1509 
1510 /*
1511  * update the runtime service curve by taking the minimum of the current
1512  * runtime service curve and the service curve starting at (x, y).
1513  */
1514 static void
1515 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1516 {
1517 	uint64_t y1, y2, dx, dy;
1518 
1519 	if (isc->sm1 <= isc->sm2) {
1520 		/* service curve is convex */
1521 		y1 = rtsc_x2y(rtsc, x);
1522 		if (y1 < y)
1523 			/* the current rtsc is smaller */
1524 			return;
1525 		rtsc->x = x;
1526 		rtsc->y = y;
1527 		return;
1528 	}
1529 
1530 	/*
1531 	 * service curve is concave
1532 	 * compute the two y values of the current rtsc
1533 	 *	y1: at x
1534 	 *	y2: at (x + dx)
1535 	 */
1536 	y1 = rtsc_x2y(rtsc, x);
1537 	if (y1 <= y) {
1538 		/* rtsc is below isc, no change to rtsc */
1539 		return;
1540 	}
1541 
1542 	y2 = rtsc_x2y(rtsc, x + isc->dx);
1543 	if (y2 >= y + isc->dy) {
1544 		/* rtsc is above isc, replace rtsc by isc */
1545 		rtsc->x = x;
1546 		rtsc->y = y;
1547 		rtsc->dx = isc->dx;
1548 		rtsc->dy = isc->dy;
1549 		return;
1550 	}
1551 
1552 	/*
1553 	 * the two curves intersect
1554 	 * compute the offsets (dx, dy) using the reverse
1555 	 * function of seg_x2y()
1556 	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
1557 	 */
1558 	dx = ((y1 - y) << SM_SHIFT) / (isc->sm1 - isc->sm2);
1559 	/*
1560 	 * check if (x, y1) belongs to the 1st segment of rtsc.
1561 	 * if so, add the offset.
1562 	 */
1563 	if (rtsc->x + rtsc->dx > x)
1564 		dx += rtsc->x + rtsc->dx - x;
1565 	dy = seg_x2y(dx, isc->sm1);
1566 
1567 	rtsc->x = x;
1568 	rtsc->y = y;
1569 	rtsc->dx = dx;
1570 	rtsc->dy = dy;
1571 }
1572 
1573 static void
1574 get_class_stats(struct hfsc_classstats *sp, struct hfsc_class *cl)
1575 {
1576 	sp->class_id = cl->cl_id;
1577 	sp->class_handle = cl->cl_handle;
1578 
1579 	if (cl->cl_rsc != NULL) {
1580 		sp->rsc.m1 = sm2m(cl->cl_rsc->sm1);
1581 		sp->rsc.d = dx2d(cl->cl_rsc->dx);
1582 		sp->rsc.m2 = sm2m(cl->cl_rsc->sm2);
1583 	} else {
1584 		sp->rsc.m1 = 0;
1585 		sp->rsc.d = 0;
1586 		sp->rsc.m2 = 0;
1587 	}
1588 	if (cl->cl_fsc != NULL) {
1589 		sp->fsc.m1 = sm2m(cl->cl_fsc->sm1);
1590 		sp->fsc.d = dx2d(cl->cl_fsc->dx);
1591 		sp->fsc.m2 = sm2m(cl->cl_fsc->sm2);
1592 	} else {
1593 		sp->fsc.m1 = 0;
1594 		sp->fsc.d = 0;
1595 		sp->fsc.m2 = 0;
1596 	}
1597 	if (cl->cl_usc != NULL) {
1598 		sp->usc.m1 = sm2m(cl->cl_usc->sm1);
1599 		sp->usc.d = dx2d(cl->cl_usc->dx);
1600 		sp->usc.m2 = sm2m(cl->cl_usc->sm2);
1601 	} else {
1602 		sp->usc.m1 = 0;
1603 		sp->usc.d = 0;
1604 		sp->usc.m2 = 0;
1605 	}
1606 
1607 	sp->total = cl->cl_total;
1608 	sp->cumul = cl->cl_cumul;
1609 
1610 	sp->d = cl->cl_d;
1611 	sp->e = cl->cl_e;
1612 	sp->vt = cl->cl_vt;
1613 	sp->f = cl->cl_f;
1614 
1615 	sp->initvt = cl->cl_initvt;
1616 	sp->vtperiod = cl->cl_vtperiod;
1617 	sp->parentperiod = cl->cl_parentperiod;
1618 	sp->nactive = cl->cl_nactive;
1619 	sp->vtoff = cl->cl_vtoff;
1620 	sp->cvtmax = cl->cl_cvtmax;
1621 	sp->myf = cl->cl_myf;
1622 	sp->cfmin = cl->cl_cfmin;
1623 	sp->cvtmin = cl->cl_cvtmin;
1624 	sp->myfadj = cl->cl_myfadj;
1625 	sp->vtadj = cl->cl_vtadj;
1626 
1627 	sp->cur_time = read_machclk();
1628 	sp->machclk_freq = machclk_freq;
1629 
1630 	sp->qlength = qlen(cl->cl_q);
1631 	sp->qlimit = qlimit(cl->cl_q);
1632 	sp->xmit_cnt = cl->cl_stats.xmit_cnt;
1633 	sp->drop_cnt = cl->cl_stats.drop_cnt;
1634 	sp->period = cl->cl_stats.period;
1635 
1636 	sp->qtype = qtype(cl->cl_q);
1637 #ifdef ALTQ_RED
1638 	if (q_is_red(cl->cl_q))
1639 		red_getstats(cl->cl_red, &sp->red[0]);
1640 #endif
1641 #ifdef ALTQ_RIO
1642 	if (q_is_rio(cl->cl_q))
1643 		rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
1644 #endif
1645 }
1646 
1647 /* convert a class handle to the corresponding class pointer */
1648 static struct hfsc_class *
1649 clh_to_clp(struct hfsc_if *hif, uint32_t chandle)
1650 {
1651 	int i;
1652 	struct hfsc_class *cl;
1653 
1654 	if (chandle == 0)
1655 		return (NULL);
1656 	/*
1657 	 * first, try optimistically the slot matching the lower bits of
1658 	 * the handle.  if it fails, do the linear table search.
1659 	 */
1660 	i = chandle % HFSC_MAX_CLASSES;
1661 	if ((cl = hif->hif_class_tbl[i]) != NULL && cl->cl_handle == chandle)
1662 		return (cl);
1663 	for (i = 0; i < HFSC_MAX_CLASSES; i++)
1664 		if ((cl = hif->hif_class_tbl[i]) != NULL &&
1665 		    cl->cl_handle == chandle)
1666 			return (cl);
1667 	return (NULL);
1668 }
1669 
1670 #endif /* ALTQ_HFSC */
1671