xref: /dragonfly/sys/net/altq/altq_hfsc.c (revision 6fb88001)
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 1.5 2005/11/22 00:24:35 dillon 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)
143 {
144 	struct ifnet *ifp;
145 	int error;
146 
147 	if ((ifp = ifunit(a->ifname)) == NULL || a->altq_disc == NULL)
148 		return (EINVAL);
149 	crit_enter();
150 	error = altq_attach(&ifp->if_snd, ALTQT_HFSC, a->altq_disc,
151 	    hfsc_enqueue, hfsc_dequeue, hfsc_request, NULL, NULL);
152 	crit_exit();
153 	return (error);
154 }
155 
156 int
157 hfsc_add_altq(struct pf_altq *a)
158 {
159 	struct hfsc_if *hif;
160 	struct ifnet *ifp;
161 
162 	if ((ifp = ifunit(a->ifname)) == NULL)
163 		return (EINVAL);
164 	if (!ifq_is_ready(&ifp->if_snd))
165 		return (ENODEV);
166 
167 	hif = malloc(sizeof(struct hfsc_if), M_ALTQ, M_WAITOK | M_ZERO);
168 
169 	hif->hif_eligible = ellist_alloc();
170 	hif->hif_ifq = &ifp->if_snd;
171 
172 	/* keep the state in pf_altq */
173 	a->altq_disc = hif;
174 
175 	return (0);
176 }
177 
178 int
179 hfsc_remove_altq(struct pf_altq *a)
180 {
181 	struct hfsc_if *hif;
182 
183 	if ((hif = a->altq_disc) == NULL)
184 		return (EINVAL);
185 	a->altq_disc = NULL;
186 
187 	hfsc_clear_interface(hif);
188 	hfsc_class_destroy(hif->hif_rootclass);
189 
190 	ellist_destroy(hif->hif_eligible);
191 
192 	free(hif, M_ALTQ);
193 
194 	return (0);
195 }
196 
197 int
198 hfsc_add_queue(struct pf_altq *a)
199 {
200 	struct hfsc_if *hif;
201 	struct hfsc_class *cl, *parent;
202 	struct hfsc_opts *opts;
203 	struct service_curve rtsc, lssc, ulsc;
204 
205 	if ((hif = a->altq_disc) == NULL)
206 		return (EINVAL);
207 
208 	opts = &a->pq_u.hfsc_opts;
209 
210 	if (a->parent_qid == HFSC_NULLCLASS_HANDLE && hif->hif_rootclass == NULL)
211 		parent = NULL;
212 	else if ((parent = clh_to_clp(hif, a->parent_qid)) == NULL)
213 		return (EINVAL);
214 
215 	if (a->qid == 0)
216 		return (EINVAL);
217 
218 	if (clh_to_clp(hif, a->qid) != NULL)
219 		return (EBUSY);
220 
221 	rtsc.m1 = opts->rtsc_m1;
222 	rtsc.d  = opts->rtsc_d;
223 	rtsc.m2 = opts->rtsc_m2;
224 	lssc.m1 = opts->lssc_m1;
225 	lssc.d  = opts->lssc_d;
226 	lssc.m2 = opts->lssc_m2;
227 	ulsc.m1 = opts->ulsc_m1;
228 	ulsc.d  = opts->ulsc_d;
229 	ulsc.m2 = opts->ulsc_m2;
230 
231 	cl = hfsc_class_create(hif, &rtsc, &lssc, &ulsc, parent, a->qlimit,
232 			       opts->flags, a->qid);
233 	if (cl == NULL)
234 		return (ENOMEM);
235 
236 	return (0);
237 }
238 
239 int
240 hfsc_remove_queue(struct pf_altq *a)
241 {
242 	struct hfsc_if *hif;
243 	struct hfsc_class *cl;
244 
245 	if ((hif = a->altq_disc) == NULL)
246 		return (EINVAL);
247 
248 	if ((cl = clh_to_clp(hif, a->qid)) == NULL)
249 		return (EINVAL);
250 
251 	return (hfsc_class_destroy(cl));
252 }
253 
254 int
255 hfsc_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
256 {
257 	struct hfsc_if *hif;
258 	struct hfsc_class *cl;
259 	struct hfsc_classstats stats;
260 	int error = 0;
261 
262 	if ((hif = altq_lookup(a->ifname, ALTQT_HFSC)) == NULL)
263 		return (EBADF);
264 
265 	if ((cl = clh_to_clp(hif, a->qid)) == NULL)
266 		return (EINVAL);
267 
268 	if (*nbytes < sizeof(stats))
269 		return (EINVAL);
270 
271 	get_class_stats(&stats, cl);
272 
273 	if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
274 		return (error);
275 	*nbytes = sizeof(stats);
276 	return (0);
277 }
278 
279 /*
280  * bring the interface back to the initial state by discarding
281  * all the filters and classes except the root class.
282  */
283 static int
284 hfsc_clear_interface(struct hfsc_if *hif)
285 {
286 	struct hfsc_class *cl;
287 
288 	if (hif->hif_rootclass == NULL)
289 		return (0);
290 
291 
292 	/* clear out the classes */
293 	while ((cl = hif->hif_rootclass->cl_children) != NULL) {
294 		/*
295 		 * remove the first leaf class found in the hierarchy
296 		 * then start over
297 		 */
298 		for (; cl != NULL; cl = hfsc_nextclass(cl)) {
299 			if (!is_a_parent_class(cl)) {
300 				hfsc_class_destroy(cl);
301 				break;
302 			}
303 		}
304 	}
305 
306 	return (0);
307 }
308 
309 static int
310 hfsc_request(struct ifaltq *ifq, int req, void *arg)
311 {
312 	struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
313 
314 	crit_enter();
315 	switch (req) {
316 	case ALTRQ_PURGE:
317 		hfsc_purge(hif);
318 		break;
319 	}
320 	crit_exit();
321 	return (0);
322 }
323 
324 /* discard all the queued packets on the interface */
325 static void
326 hfsc_purge(struct hfsc_if *hif)
327 {
328 	struct hfsc_class *cl;
329 
330 	for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl)) {
331 		if (!qempty(cl->cl_q))
332 			hfsc_purgeq(cl);
333 	}
334 	if (ifq_is_enabled(hif->hif_ifq))
335 		hif->hif_ifq->ifq_len = 0;
336 }
337 
338 struct hfsc_class *
339 hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
340 		  struct service_curve *fsc, struct service_curve *usc,
341 		  struct hfsc_class *parent, int qlimit, int flags, int qid)
342 {
343 	struct hfsc_class *cl, *p;
344 	int i;
345 
346 	if (hif->hif_classes >= HFSC_MAX_CLASSES)
347 		return (NULL);
348 
349 #ifndef ALTQ_RED
350 	if (flags & HFCF_RED) {
351 #ifdef ALTQ_DEBUG
352 		printf("hfsc_class_create: RED not configured for HFSC!\n");
353 #endif
354 		return (NULL);
355 	}
356 #endif
357 
358 	cl = malloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
359 	cl->cl_q = malloc(sizeof(*cl->cl_q), M_ALTQ, M_WAITOK | M_ZERO);
360 	cl->cl_actc = actlist_alloc();
361 
362 	if (qlimit == 0)
363 		qlimit = 50;  /* use default */
364 	qlimit(cl->cl_q) = qlimit;
365 	qtype(cl->cl_q) = Q_DROPTAIL;
366 	qlen(cl->cl_q) = 0;
367 	cl->cl_flags = flags;
368 #ifdef ALTQ_RED
369 	if (flags & (HFCF_RED|HFCF_RIO)) {
370 		int red_flags, red_pkttime;
371 		u_int m2;
372 
373 		m2 = 0;
374 		if (rsc != NULL && rsc->m2 > m2)
375 			m2 = rsc->m2;
376 		if (fsc != NULL && fsc->m2 > m2)
377 			m2 = fsc->m2;
378 		if (usc != NULL && usc->m2 > m2)
379 			m2 = usc->m2;
380 
381 		red_flags = 0;
382 		if (flags & HFCF_ECN)
383 			red_flags |= REDF_ECN;
384 #ifdef ALTQ_RIO
385 		if (flags & HFCF_CLEARDSCP)
386 			red_flags |= RIOF_CLEARDSCP;
387 #endif
388 		if (m2 < 8)
389 			red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
390 		else
391 			red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
392 				* 1000 * 1000 * 1000 / (m2 / 8);
393 		if (flags & HFCF_RED) {
394 			cl->cl_red = red_alloc(0, 0,
395 			    qlimit(cl->cl_q) * 10/100,
396 			    qlimit(cl->cl_q) * 30/100,
397 			    red_flags, red_pkttime);
398 			if (cl->cl_red != NULL)
399 				qtype(cl->cl_q) = Q_RED;
400 		}
401 #ifdef ALTQ_RIO
402 		else {
403 			cl->cl_red = (red_t *)rio_alloc(0, NULL,
404 			    red_flags, red_pkttime);
405 			if (cl->cl_red != NULL)
406 				qtype(cl->cl_q) = Q_RIO;
407 		}
408 #endif
409 	}
410 #endif /* ALTQ_RED */
411 
412 	if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
413 		cl->cl_rsc = malloc(sizeof(*cl->cl_rsc), M_ALTQ, M_WAITOK);
414 		sc2isc(rsc, cl->cl_rsc);
415 		rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
416 		rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
417 	}
418 	if (fsc != NULL && (fsc->m1 != 0 || fsc->m2 != 0)) {
419 		cl->cl_fsc = malloc(sizeof(*cl->cl_fsc), M_ALTQ, M_WAITOK);
420 		if (cl->cl_fsc == NULL)
421 			goto err_ret;
422 		sc2isc(fsc, cl->cl_fsc);
423 		rtsc_init(&cl->cl_virtual, cl->cl_fsc, 0, 0);
424 	}
425 	if (usc != NULL && (usc->m1 != 0 || usc->m2 != 0)) {
426 		cl->cl_usc = malloc(sizeof(*cl->cl_usc), M_ALTQ, M_WAITOK);
427 		if (cl->cl_usc == NULL)
428 			goto err_ret;
429 		sc2isc(usc, cl->cl_usc);
430 		rtsc_init(&cl->cl_ulimit, cl->cl_usc, 0, 0);
431 	}
432 
433 	cl->cl_id = hif->hif_classid++;
434 	cl->cl_handle = qid;
435 	cl->cl_hif = hif;
436 	cl->cl_parent = parent;
437 
438 	crit_enter();
439 	hif->hif_classes++;
440 
441 	/*
442 	 * find a free slot in the class table.  if the slot matching
443 	 * the lower bits of qid is free, use this slot.  otherwise,
444 	 * use the first free slot.
445 	 */
446 	i = qid % HFSC_MAX_CLASSES;
447 	if (hif->hif_class_tbl[i] == NULL)
448 		hif->hif_class_tbl[i] = cl;
449 	else {
450 		for (i = 0; i < HFSC_MAX_CLASSES; i++) {
451 			if (hif->hif_class_tbl[i] == NULL) {
452 				hif->hif_class_tbl[i] = cl;
453 				break;
454 			}
455 		}
456 		if (i == HFSC_MAX_CLASSES) {
457 			crit_exit();
458 			goto err_ret;
459 		}
460 	}
461 
462 	if (flags & HFCF_DEFAULTCLASS)
463 		hif->hif_defaultclass = cl;
464 
465 	if (parent == NULL) {
466 		/* this is root class */
467 		hif->hif_rootclass = cl;
468 	} else if (parent->cl_children == NULL) {
469 		/* add this class to the children list of the parent */
470 		parent->cl_children = cl;
471 	} else {
472 		p = parent->cl_children;
473 		while (p->cl_siblings != NULL)
474 			p = p->cl_siblings;
475 		p->cl_siblings = cl;
476 	}
477 	crit_exit();
478 
479 	return (cl);
480 
481  err_ret:
482 	if (cl->cl_actc != NULL)
483 		actlist_destroy(cl->cl_actc);
484 	if (cl->cl_red != NULL) {
485 #ifdef ALTQ_RIO
486 		if (q_is_rio(cl->cl_q))
487 			rio_destroy((rio_t *)cl->cl_red);
488 #endif
489 #ifdef ALTQ_RED
490 		if (q_is_red(cl->cl_q))
491 			red_destroy(cl->cl_red);
492 #endif
493 	}
494 	if (cl->cl_fsc != NULL)
495 		free(cl->cl_fsc, M_ALTQ);
496 	if (cl->cl_rsc != NULL)
497 		free(cl->cl_rsc, M_ALTQ);
498 	if (cl->cl_usc != NULL)
499 		free(cl->cl_usc, M_ALTQ);
500 	if (cl->cl_q != NULL)
501 		free(cl->cl_q, M_ALTQ);
502 	free(cl, M_ALTQ);
503 	return (NULL);
504 }
505 
506 static int
507 hfsc_class_destroy(struct hfsc_class *cl)
508 {
509 	int i;
510 
511 	if (cl == NULL)
512 		return (0);
513 
514 	if (is_a_parent_class(cl))
515 		return (EBUSY);
516 
517 	crit_enter();
518 
519 	if (!qempty(cl->cl_q))
520 		hfsc_purgeq(cl);
521 
522 	if (cl->cl_parent == NULL) {
523 		/* this is root class */
524 	} else {
525 		struct hfsc_class *p = cl->cl_parent->cl_children;
526 
527 		if (p == cl) {
528 			cl->cl_parent->cl_children = cl->cl_siblings;
529 		} else {
530 			do {
531 				if (p->cl_siblings == cl) {
532 					p->cl_siblings = cl->cl_siblings;
533 					break;
534 				}
535 			} while ((p = p->cl_siblings) != NULL);
536 		}
537 		KKASSERT(p != NULL);
538 	}
539 
540 	for (i = 0; i < HFSC_MAX_CLASSES; i++) {
541 		if (cl->cl_hif->hif_class_tbl[i] == cl) {
542 			cl->cl_hif->hif_class_tbl[i] = NULL;
543 			break;
544 		}
545 	}
546 
547 	cl->cl_hif->hif_classes--;
548 	crit_exit();
549 
550 	actlist_destroy(cl->cl_actc);
551 
552 	if (cl->cl_red != NULL) {
553 #ifdef ALTQ_RIO
554 		if (q_is_rio(cl->cl_q))
555 			rio_destroy((rio_t *)cl->cl_red);
556 #endif
557 #ifdef ALTQ_RED
558 		if (q_is_red(cl->cl_q))
559 			red_destroy(cl->cl_red);
560 #endif
561 	}
562 
563 	if (cl == cl->cl_hif->hif_rootclass)
564 		cl->cl_hif->hif_rootclass = NULL;
565 	if (cl == cl->cl_hif->hif_defaultclass)
566 		cl->cl_hif->hif_defaultclass = NULL;
567 
568 	if (cl->cl_usc != NULL)
569 		free(cl->cl_usc, M_ALTQ);
570 	if (cl->cl_fsc != NULL)
571 		free(cl->cl_fsc, M_ALTQ);
572 	if (cl->cl_rsc != NULL)
573 		free(cl->cl_rsc, M_ALTQ);
574 	free(cl->cl_q, M_ALTQ);
575 	free(cl, M_ALTQ);
576 
577 	return (0);
578 }
579 
580 /*
581  * hfsc_nextclass returns the next class in the tree.
582  *   usage:
583  *	for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl))
584  *		do_something;
585  */
586 static struct hfsc_class *
587 hfsc_nextclass(struct hfsc_class *cl)
588 {
589 	if (cl->cl_children != NULL) {
590 		cl = cl->cl_children;
591 	} else if (cl->cl_siblings != NULL) {
592 		cl = cl->cl_siblings;
593 	} else {
594 		while ((cl = cl->cl_parent) != NULL) {
595 			if (cl->cl_siblings != NULL) {
596 				cl = cl->cl_siblings;
597 				break;
598 			}
599 		}
600 	}
601 
602 	return (cl);
603 }
604 
605 /*
606  * hfsc_enqueue is an enqueue function to be registered to
607  * (*altq_enqueue) in struct ifaltq.
608  */
609 static int
610 hfsc_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
611 {
612 	struct hfsc_if	*hif = (struct hfsc_if *)ifq->altq_disc;
613 	struct hfsc_class *cl;
614 	int len;
615 
616 	/* grab class set by classifier */
617 	if ((m->m_flags & M_PKTHDR) == 0) {
618 		/* should not happen */
619 		if_printf(ifq->altq_ifp, "altq: packet does not have pkthdr\n");
620 		m_freem(m);
621 		return (ENOBUFS);
622 	}
623 	crit_enter();
624 	if (m->m_pkthdr.fw_flags & ALTQ_MBUF_TAGGED)
625 		cl = clh_to_clp(hif, m->m_pkthdr.altq_qid);
626 	else
627 		cl = NULL;
628 	if (cl == NULL || is_a_parent_class(cl)) {
629 		cl = hif->hif_defaultclass;
630 		if (cl == NULL) {
631 			m_freem(m);
632 			crit_exit();
633 			return (ENOBUFS);
634 		}
635 	}
636 	cl->cl_pktattr = NULL;
637 	len = m_pktlen(m);
638 	if (hfsc_addq(cl, m) != 0) {
639 		/* drop occurred.  mbuf was freed in hfsc_addq. */
640 		PKTCNTR_ADD(&cl->cl_stats.drop_cnt, len);
641 		crit_exit();
642 		return (ENOBUFS);
643 	}
644 	ifq->ifq_len++;
645 	cl->cl_hif->hif_packets++;
646 
647 	/* successfully queued. */
648 	if (qlen(cl->cl_q) == 1)
649 		set_active(cl, m_pktlen(m));
650 	crit_exit();
651 	return (0);
652 }
653 
654 /*
655  * hfsc_dequeue is a dequeue function to be registered to
656  * (*altq_dequeue) in struct ifaltq.
657  *
658  * note: ALTDQ_POLL returns the next packet without removing the packet
659  *	from the queue.  ALTDQ_REMOVE is a normal dequeue operation.
660  *	ALTDQ_REMOVE must return the same packet if called immediately
661  *	after ALTDQ_POLL.
662  */
663 static struct mbuf *
664 hfsc_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
665 {
666 	struct hfsc_if	*hif = (struct hfsc_if *)ifq->altq_disc;
667 	struct hfsc_class *cl;
668 	struct mbuf *m;
669 	int len, next_len;
670 	int realtime = 0;
671 	uint64_t cur_time;
672 
673 	if (hif->hif_packets == 0) {
674 		/* no packet in the tree */
675 		return (NULL);
676 	}
677 
678 	crit_enter();
679 	cur_time = read_machclk();
680 
681 	if (op == ALTDQ_REMOVE && hif->hif_pollcache != NULL) {
682 		cl = hif->hif_pollcache;
683 		hif->hif_pollcache = NULL;
684 		/* check if the class was scheduled by real-time criteria */
685 		if (cl->cl_rsc != NULL)
686 			realtime = (cl->cl_e <= cur_time);
687 	} else {
688 		/*
689 		 * if there are eligible classes, use real-time criteria.
690 		 * find the class with the minimum deadline among
691 		 * the eligible classes.
692 		 */
693 		if ((cl = ellist_get_mindl(hif->hif_eligible, cur_time)) != NULL) {
694 			realtime = 1;
695 		} else {
696 #ifdef ALTQ_DEBUG
697 			int fits = 0;
698 #endif
699 			/*
700 			 * use link-sharing criteria
701 			 * get the class with the minimum vt in the hierarchy
702 			 */
703 			cl = hif->hif_rootclass;
704 			while (is_a_parent_class(cl)) {
705 
706 				cl = actlist_firstfit(cl, cur_time);
707 				if (cl == NULL) {
708 #ifdef ALTQ_DEBUG
709 					if (fits > 0)
710 						printf("%d fit but none found\n",fits);
711 #endif
712 					m = NULL;
713 					goto done;
714 				}
715 				/*
716 				 * update parent's cl_cvtmin.
717 				 * don't update if the new vt is smaller.
718 				 */
719 				if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
720 					cl->cl_parent->cl_cvtmin = cl->cl_vt;
721 #ifdef ALTQ_DEBUG
722 				fits++;
723 #endif
724 			}
725 		}
726 
727 		if (op == ALTDQ_POLL) {
728 			hif->hif_pollcache = cl;
729 			m = hfsc_pollq(cl);
730 			goto done;
731 		}
732 	}
733 
734 	m = hfsc_getq(cl);
735 	if (m == NULL)
736 		panic("hfsc_dequeue:");
737 	len = m_pktlen(m);
738 	cl->cl_hif->hif_packets--;
739 	ifq->ifq_len--;
740 	PKTCNTR_ADD(&cl->cl_stats.xmit_cnt, len);
741 
742 	update_vf(cl, len, cur_time);
743 	if (realtime)
744 		cl->cl_cumul += len;
745 
746 	if (!qempty(cl->cl_q)) {
747 		if (cl->cl_rsc != NULL) {
748 			/* update ed */
749 			next_len = m_pktlen(qhead(cl->cl_q));
750 
751 			if (realtime)
752 				update_ed(cl, next_len);
753 			else
754 				update_d(cl, next_len);
755 		}
756 	} else {
757 		/* the class becomes passive */
758 		set_passive(cl);
759 	}
760 done:
761 	crit_exit();
762 	KKASSERT(mpolled == NULL || m == mpolled);
763 	return (m);
764 }
765 
766 static int
767 hfsc_addq(struct hfsc_class *cl, struct mbuf *m)
768 {
769 
770 #ifdef ALTQ_RIO
771 	if (q_is_rio(cl->cl_q))
772 		return rio_addq((rio_t *)cl->cl_red, cl->cl_q,
773 				m, cl->cl_pktattr);
774 #endif
775 #ifdef ALTQ_RED
776 	if (q_is_red(cl->cl_q))
777 		return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
778 #endif
779 	if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
780 		m_freem(m);
781 		return (-1);
782 	}
783 
784 	if (cl->cl_flags & HFCF_CLEARDSCP)
785 		write_dsfield(m, cl->cl_pktattr, 0);
786 
787 	_addq(cl->cl_q, m);
788 
789 	return (0);
790 }
791 
792 static struct mbuf *
793 hfsc_getq(struct hfsc_class *cl)
794 {
795 #ifdef ALTQ_RIO
796 	if (q_is_rio(cl->cl_q))
797 		return rio_getq((rio_t *)cl->cl_red, cl->cl_q);
798 #endif
799 #ifdef ALTQ_RED
800 	if (q_is_red(cl->cl_q))
801 		return red_getq(cl->cl_red, cl->cl_q);
802 #endif
803 	return _getq(cl->cl_q);
804 }
805 
806 static struct mbuf *
807 hfsc_pollq(struct hfsc_class *cl)
808 {
809 	return qhead(cl->cl_q);
810 }
811 
812 static void
813 hfsc_purgeq(struct hfsc_class *cl)
814 {
815 	struct mbuf *m;
816 
817 	if (qempty(cl->cl_q))
818 		return;
819 
820 	while ((m = _getq(cl->cl_q)) != NULL) {
821 		PKTCNTR_ADD(&cl->cl_stats.drop_cnt, m_pktlen(m));
822 		m_freem(m);
823 		cl->cl_hif->hif_packets--;
824 		cl->cl_hif->hif_ifq->ifq_len--;
825 	}
826 	KKASSERT(qlen(cl->cl_q) == 0);
827 
828 	update_vf(cl, 0, 0);	/* remove cl from the actlist */
829 	set_passive(cl);
830 }
831 
832 static void
833 set_active(struct hfsc_class *cl, int len)
834 {
835 	if (cl->cl_rsc != NULL)
836 		init_ed(cl, len);
837 	if (cl->cl_fsc != NULL)
838 		init_vf(cl, len);
839 
840 	cl->cl_stats.period++;
841 }
842 
843 static void
844 set_passive(struct hfsc_class *cl)
845 {
846 	if (cl->cl_rsc != NULL)
847 		ellist_remove(cl);
848 
849 	/*
850 	 * actlist is now handled in update_vf() so that update_vf(cl, 0, 0)
851 	 * needs to be called explicitly to remove a class from actlist
852 	 */
853 }
854 
855 static void
856 init_ed(struct hfsc_class *cl, int next_len)
857 {
858 	uint64_t cur_time;
859 
860 	cur_time = read_machclk();
861 
862 	/* update the deadline curve */
863 	rtsc_min(&cl->cl_deadline, cl->cl_rsc, cur_time, cl->cl_cumul);
864 
865 	/*
866 	 * update the eligible curve.
867 	 * for concave, it is equal to the deadline curve.
868 	 * for convex, it is a linear curve with slope m2.
869 	 */
870 	cl->cl_eligible = cl->cl_deadline;
871 	if (cl->cl_rsc->sm1 <= cl->cl_rsc->sm2) {
872 		cl->cl_eligible.dx = 0;
873 		cl->cl_eligible.dy = 0;
874 	}
875 
876 	/* compute e and d */
877 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
878 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
879 
880 	ellist_insert(cl);
881 }
882 
883 static void
884 update_ed(struct hfsc_class *cl, int next_len)
885 {
886 	cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
887 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
888 
889 	ellist_update(cl);
890 }
891 
892 static void
893 update_d(struct hfsc_class *cl, int next_len)
894 {
895 	cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
896 }
897 
898 static void
899 init_vf(struct hfsc_class *cl, int len)
900 {
901 	struct hfsc_class *max_cl, *p;
902 	uint64_t vt, f, cur_time;
903 	int go_active;
904 
905 	cur_time = 0;
906 	go_active = 1;
907 	for ( ; cl->cl_parent != NULL; cl = cl->cl_parent) {
908 		if (go_active && cl->cl_nactive++ == 0)
909 			go_active = 1;
910 		else
911 			go_active = 0;
912 
913 		if (go_active) {
914 			max_cl = actlist_last(cl->cl_parent->cl_actc);
915 			if (max_cl != NULL) {
916 				/*
917 				 * set vt to the average of the min and max
918 				 * classes.  if the parent's period didn't
919 				 * change, don't decrease vt of the class.
920 				 */
921 				vt = max_cl->cl_vt;
922 				if (cl->cl_parent->cl_cvtmin != 0)
923 					vt = (cl->cl_parent->cl_cvtmin + vt)/2;
924 
925 				if (cl->cl_parent->cl_vtperiod !=
926 				    cl->cl_parentperiod || vt > cl->cl_vt)
927 					cl->cl_vt = vt;
928 			} else {
929 				/*
930 				 * first child for a new parent backlog period.
931 				 * add parent's cvtmax to vtoff of children
932 				 * to make a new vt (vtoff + vt) larger than
933 				 * the vt in the last period for all children.
934 				 */
935 				vt = cl->cl_parent->cl_cvtmax;
936 				for (p = cl->cl_parent->cl_children; p != NULL;
937 				     p = p->cl_siblings)
938 					p->cl_vtoff += vt;
939 				cl->cl_vt = 0;
940 				cl->cl_parent->cl_cvtmax = 0;
941 				cl->cl_parent->cl_cvtmin = 0;
942 			}
943 			cl->cl_initvt = cl->cl_vt;
944 
945 			/* update the virtual curve */
946 			vt = cl->cl_vt + cl->cl_vtoff;
947 			rtsc_min(&cl->cl_virtual, cl->cl_fsc, vt, cl->cl_total);
948 			if (cl->cl_virtual.x == vt) {
949 				cl->cl_virtual.x -= cl->cl_vtoff;
950 				cl->cl_vtoff = 0;
951 			}
952 			cl->cl_vtadj = 0;
953 
954 			cl->cl_vtperiod++;  /* increment vt period */
955 			cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
956 			if (cl->cl_parent->cl_nactive == 0)
957 				cl->cl_parentperiod++;
958 			cl->cl_f = 0;
959 
960 			actlist_insert(cl);
961 
962 			if (cl->cl_usc != NULL) {
963 				/* class has upper limit curve */
964 				if (cur_time == 0)
965 					cur_time = read_machclk();
966 
967 				/* update the ulimit curve */
968 				rtsc_min(&cl->cl_ulimit, cl->cl_usc, cur_time,
969 				    cl->cl_total);
970 				/* compute myf */
971 				cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
972 				    cl->cl_total);
973 				cl->cl_myfadj = 0;
974 			}
975 		}
976 
977 		if (cl->cl_myf > cl->cl_cfmin)
978 			f = cl->cl_myf;
979 		else
980 			f = cl->cl_cfmin;
981 		if (f != cl->cl_f) {
982 			cl->cl_f = f;
983 			update_cfmin(cl->cl_parent);
984 		}
985 	}
986 }
987 
988 static void
989 update_vf(struct hfsc_class *cl, int len, uint64_t cur_time)
990 {
991 	uint64_t f, myf_bound, delta;
992 	int go_passive;
993 
994 	go_passive = qempty(cl->cl_q);
995 
996 	for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
997 		cl->cl_total += len;
998 
999 		if (cl->cl_fsc == NULL || cl->cl_nactive == 0)
1000 			continue;
1001 
1002 		if (go_passive && --cl->cl_nactive == 0)
1003 			go_passive = 1;
1004 		else
1005 			go_passive = 0;
1006 
1007 		if (go_passive) {
1008 			/* no more active child, going passive */
1009 
1010 			/* update cvtmax of the parent class */
1011 			if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
1012 				cl->cl_parent->cl_cvtmax = cl->cl_vt;
1013 
1014 			/* remove this class from the vt list */
1015 			actlist_remove(cl);
1016 
1017 			update_cfmin(cl->cl_parent);
1018 
1019 			continue;
1020 		}
1021 
1022 		/*
1023 		 * update vt and f
1024 		 */
1025 		cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
1026 		    - cl->cl_vtoff + cl->cl_vtadj;
1027 
1028 		/*
1029 		 * if vt of the class is smaller than cvtmin,
1030 		 * the class was skipped in the past due to non-fit.
1031 		 * if so, we need to adjust vtadj.
1032 		 */
1033 		if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
1034 			cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
1035 			cl->cl_vt = cl->cl_parent->cl_cvtmin;
1036 		}
1037 
1038 		/* update the vt list */
1039 		actlist_update(cl);
1040 
1041 		if (cl->cl_usc != NULL) {
1042 			cl->cl_myf = cl->cl_myfadj
1043 			    + rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
1044 
1045 			/*
1046 			 * if myf lags behind by more than one clock tick
1047 			 * from the current time, adjust myfadj to prevent
1048 			 * a rate-limited class from going greedy.
1049 			 * in a steady state under rate-limiting, myf
1050 			 * fluctuates within one clock tick.
1051 			 */
1052 			myf_bound = cur_time - machclk_per_tick;
1053 			if (cl->cl_myf < myf_bound) {
1054 				delta = cur_time - cl->cl_myf;
1055 				cl->cl_myfadj += delta;
1056 				cl->cl_myf += delta;
1057 			}
1058 		}
1059 
1060 		/* cl_f is max(cl_myf, cl_cfmin) */
1061 		if (cl->cl_myf > cl->cl_cfmin)
1062 			f = cl->cl_myf;
1063 		else
1064 			f = cl->cl_cfmin;
1065 		if (f != cl->cl_f) {
1066 			cl->cl_f = f;
1067 			update_cfmin(cl->cl_parent);
1068 		}
1069 	}
1070 }
1071 
1072 static void
1073 update_cfmin(struct hfsc_class *cl)
1074 {
1075 	struct hfsc_class *p;
1076 	uint64_t cfmin;
1077 
1078 	if (TAILQ_EMPTY(cl->cl_actc)) {
1079 		cl->cl_cfmin = 0;
1080 		return;
1081 	}
1082 	cfmin = HT_INFINITY;
1083 	TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1084 		if (p->cl_f == 0) {
1085 			cl->cl_cfmin = 0;
1086 			return;
1087 		}
1088 		if (p->cl_f < cfmin)
1089 			cfmin = p->cl_f;
1090 	}
1091 	cl->cl_cfmin = cfmin;
1092 }
1093 
1094 /*
1095  * TAILQ based ellist and actlist implementation
1096  * (ion wanted to make a calendar queue based implementation)
1097  */
1098 /*
1099  * eligible list holds backlogged classes being sorted by their eligible times.
1100  * there is one eligible list per interface.
1101  */
1102 
1103 static ellist_t *
1104 ellist_alloc(void)
1105 {
1106 	ellist_t *head;
1107 
1108 	head = malloc(sizeof(ellist_t *), M_ALTQ, M_WAITOK);
1109 	TAILQ_INIT(head);
1110 	return (head);
1111 }
1112 
1113 static void
1114 ellist_destroy(ellist_t *head)
1115 {
1116 	free(head, M_ALTQ);
1117 }
1118 
1119 static void
1120 ellist_insert(struct hfsc_class *cl)
1121 {
1122 	struct hfsc_if *hif = cl->cl_hif;
1123 	struct hfsc_class *p;
1124 
1125 	/* check the last entry first */
1126 	if ((p = TAILQ_LAST(hif->hif_eligible, _eligible)) == NULL ||
1127 	    p->cl_e <= cl->cl_e) {
1128 		TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1129 		return;
1130 	}
1131 
1132 	TAILQ_FOREACH(p, hif->hif_eligible, cl_ellist) {
1133 		if (cl->cl_e < p->cl_e) {
1134 			TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1135 			return;
1136 		}
1137 	}
1138 	KKASSERT(0); /* should not reach here */
1139 }
1140 
1141 static void
1142 ellist_remove(struct hfsc_class *cl)
1143 {
1144 	struct hfsc_if *hif = cl->cl_hif;
1145 
1146 	TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1147 }
1148 
1149 static void
1150 ellist_update(struct hfsc_class *cl)
1151 {
1152 	struct hfsc_if *hif = cl->cl_hif;
1153 	struct hfsc_class *p, *last;
1154 
1155 	/*
1156 	 * the eligible time of a class increases monotonically.
1157 	 * if the next entry has a larger eligible time, nothing to do.
1158 	 */
1159 	p = TAILQ_NEXT(cl, cl_ellist);
1160 	if (p == NULL || cl->cl_e <= p->cl_e)
1161 		return;
1162 
1163 	/* check the last entry */
1164 	last = TAILQ_LAST(hif->hif_eligible, _eligible);
1165 	KKASSERT(last != NULL);
1166 	if (last->cl_e <= cl->cl_e) {
1167 		TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1168 		TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1169 		return;
1170 	}
1171 
1172 	/*
1173 	 * the new position must be between the next entry
1174 	 * and the last entry
1175 	 */
1176 	while ((p = TAILQ_NEXT(p, cl_ellist)) != NULL) {
1177 		if (cl->cl_e < p->cl_e) {
1178 			TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1179 			TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1180 			return;
1181 		}
1182 	}
1183 	KKASSERT(0); /* should not reach here */
1184 }
1185 
1186 /* find the class with the minimum deadline among the eligible classes */
1187 struct hfsc_class *
1188 ellist_get_mindl(ellist_t *head, uint64_t cur_time)
1189 {
1190 	struct hfsc_class *p, *cl = NULL;
1191 
1192 	TAILQ_FOREACH(p, head, cl_ellist) {
1193 		if (p->cl_e > cur_time)
1194 			break;
1195 		if (cl == NULL || p->cl_d < cl->cl_d)
1196 			cl = p;
1197 	}
1198 	return (cl);
1199 }
1200 
1201 /*
1202  * active children list holds backlogged child classes being sorted
1203  * by their virtual time.
1204  * each intermediate class has one active children list.
1205  */
1206 static actlist_t *
1207 actlist_alloc(void)
1208 {
1209 	actlist_t *head;
1210 
1211 	head = malloc(sizeof(*head), M_ALTQ, M_WAITOK);
1212 	TAILQ_INIT(head);
1213 	return (head);
1214 }
1215 
1216 static void
1217 actlist_destroy(actlist_t *head)
1218 {
1219 	free(head, M_ALTQ);
1220 }
1221 static void
1222 actlist_insert(struct hfsc_class *cl)
1223 {
1224 	struct hfsc_class *p;
1225 
1226 	/* check the last entry first */
1227 	if ((p = TAILQ_LAST(cl->cl_parent->cl_actc, _active)) == NULL
1228 	    || p->cl_vt <= cl->cl_vt) {
1229 		TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1230 		return;
1231 	}
1232 
1233 	TAILQ_FOREACH(p, cl->cl_parent->cl_actc, cl_actlist) {
1234 		if (cl->cl_vt < p->cl_vt) {
1235 			TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1236 			return;
1237 		}
1238 	}
1239 	KKASSERT(0); /* should not reach here */
1240 }
1241 
1242 static void
1243 actlist_remove(struct hfsc_class *cl)
1244 {
1245 	TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1246 }
1247 
1248 static void
1249 actlist_update(struct hfsc_class *cl)
1250 {
1251 	struct hfsc_class *p, *last;
1252 
1253 	/*
1254 	 * the virtual time of a class increases monotonically during its
1255 	 * backlogged period.
1256 	 * if the next entry has a larger virtual time, nothing to do.
1257 	 */
1258 	p = TAILQ_NEXT(cl, cl_actlist);
1259 	if (p == NULL || cl->cl_vt < p->cl_vt)
1260 		return;
1261 
1262 	/* check the last entry */
1263 	last = TAILQ_LAST(cl->cl_parent->cl_actc, _active);
1264 	KKASSERT(last != NULL);
1265 	if (last->cl_vt <= cl->cl_vt) {
1266 		TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1267 		TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1268 		return;
1269 	}
1270 
1271 	/*
1272 	 * the new position must be between the next entry
1273 	 * and the last entry
1274 	 */
1275 	while ((p = TAILQ_NEXT(p, cl_actlist)) != NULL) {
1276 		if (cl->cl_vt < p->cl_vt) {
1277 			TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1278 			TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1279 			return;
1280 		}
1281 	}
1282 	KKASSERT(0); /* should not reach here */
1283 }
1284 
1285 static struct hfsc_class *
1286 actlist_firstfit(struct hfsc_class *cl, uint64_t cur_time)
1287 {
1288 	struct hfsc_class *p;
1289 
1290 	TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1291 		if (p->cl_f <= cur_time)
1292 			return (p);
1293 	}
1294 	return (NULL);
1295 }
1296 
1297 /*
1298  * service curve support functions
1299  *
1300  *  external service curve parameters
1301  *	m: bits/sec
1302  *	d: msec
1303  *  internal service curve parameters
1304  *	sm: (bytes/tsc_interval) << SM_SHIFT
1305  *	ism: (tsc_count/byte) << ISM_SHIFT
1306  *	dx: tsc_count
1307  *
1308  * SM_SHIFT and ISM_SHIFT are scaled in order to keep effective digits.
1309  * we should be able to handle 100K-1Gbps linkspeed with 200Hz-1GHz CPU
1310  * speed.  SM_SHIFT and ISM_SHIFT are selected to have at least 3 effective
1311  * digits in decimal using the following table.
1312  *
1313  *  bits/sec    100Kbps     1Mbps     10Mbps     100Mbps    1Gbps
1314  *  ----------+-------------------------------------------------------
1315  *  bytes/nsec  12.5e-6    125e-6     1250e-6    12500e-6   125000e-6
1316  *  sm(500MHz)  25.0e-6    250e-6     2500e-6    25000e-6   250000e-6
1317  *  sm(200MHz)  62.5e-6    625e-6     6250e-6    62500e-6   625000e-6
1318  *
1319  *  nsec/byte   80000      8000       800        80         8
1320  *  ism(500MHz) 40000      4000       400        40         4
1321  *  ism(200MHz) 16000      1600       160        16         1.6
1322  */
1323 #define	SM_SHIFT	24
1324 #define	ISM_SHIFT	10
1325 
1326 #define	SM_MASK		((1LL << SM_SHIFT) - 1)
1327 #define	ISM_MASK	((1LL << ISM_SHIFT) - 1)
1328 
1329 static __inline uint64_t
1330 seg_x2y(uint64_t x, uint64_t sm)
1331 {
1332 	uint64_t y;
1333 
1334 	/*
1335 	 * compute
1336 	 *	y = x * sm >> SM_SHIFT
1337 	 * but divide it for the upper and lower bits to avoid overflow
1338 	 */
1339 	y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
1340 	return (y);
1341 }
1342 
1343 static __inline uint64_t
1344 seg_y2x(uint64_t y, uint64_t ism)
1345 {
1346 	uint64_t x;
1347 
1348 	if (y == 0)
1349 		x = 0;
1350 	else if (ism == HT_INFINITY)
1351 		x = HT_INFINITY;
1352 	else
1353 		x = (y >> ISM_SHIFT) * ism + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
1354 
1355 	return (x);
1356 }
1357 
1358 static __inline uint64_t
1359 m2sm(u_int m)
1360 {
1361 	uint64_t sm;
1362 
1363 	sm = ((uint64_t)m << SM_SHIFT) / 8 / machclk_freq;
1364 	return (sm);
1365 }
1366 
1367 static __inline uint64_t
1368 m2ism(u_int m)
1369 {
1370 	uint64_t ism;
1371 
1372 	if (m == 0)
1373 		ism = HT_INFINITY;
1374 	else
1375 		ism = ((uint64_t)machclk_freq << ISM_SHIFT) * 8 / m;
1376 	return (ism);
1377 }
1378 
1379 static __inline uint64_t
1380 d2dx(u_int d)
1381 {
1382 	uint64_t dx;
1383 
1384 	dx = ((uint64_t)d * machclk_freq) / 1000;
1385 	return (dx);
1386 }
1387 
1388 static u_int
1389 sm2m(uint64_t sm)
1390 {
1391 	uint64_t m;
1392 
1393 	m = (sm * 8 * machclk_freq) >> SM_SHIFT;
1394 	return ((u_int)m);
1395 }
1396 
1397 static u_int
1398 dx2d(uint64_t dx)
1399 {
1400 	uint64_t d;
1401 
1402 	d = dx * 1000 / machclk_freq;
1403 	return ((u_int)d);
1404 }
1405 
1406 static void
1407 sc2isc(struct service_curve *sc, struct internal_sc *isc)
1408 {
1409 	isc->sm1 = m2sm(sc->m1);
1410 	isc->ism1 = m2ism(sc->m1);
1411 	isc->dx = d2dx(sc->d);
1412 	isc->dy = seg_x2y(isc->dx, isc->sm1);
1413 	isc->sm2 = m2sm(sc->m2);
1414 	isc->ism2 = m2ism(sc->m2);
1415 }
1416 
1417 /*
1418  * initialize the runtime service curve with the given internal
1419  * service curve starting at (x, y).
1420  */
1421 static void
1422 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1423 {
1424 	rtsc->x = x;
1425 	rtsc->y = y;
1426 	rtsc->sm1 = isc->sm1;
1427 	rtsc->ism1 = isc->ism1;
1428 	rtsc->dx = isc->dx;
1429 	rtsc->dy = isc->dy;
1430 	rtsc->sm2 = isc->sm2;
1431 	rtsc->ism2 = isc->ism2;
1432 }
1433 
1434 /*
1435  * calculate the y-projection of the runtime service curve by the
1436  * given x-projection value
1437  */
1438 static uint64_t
1439 rtsc_y2x(struct runtime_sc *rtsc, uint64_t y)
1440 {
1441 	uint64_t x;
1442 
1443 	if (y < rtsc->y) {
1444 		x = rtsc->x;
1445 	} else if (y <= rtsc->y + rtsc->dy) {
1446 		/* x belongs to the 1st segment */
1447 		if (rtsc->dy == 0)
1448 			x = rtsc->x + rtsc->dx;
1449 		else
1450 			x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
1451 	} else {
1452 		/* x belongs to the 2nd segment */
1453 		x = rtsc->x + rtsc->dx
1454 		    + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
1455 	}
1456 	return (x);
1457 }
1458 
1459 static uint64_t
1460 rtsc_x2y(struct runtime_sc *rtsc, uint64_t x)
1461 {
1462 	uint64_t y;
1463 
1464 	if (x <= rtsc->x) {
1465 		y = rtsc->y;
1466 	} else if (x <= rtsc->x + rtsc->dx) {
1467 		/* y belongs to the 1st segment */
1468 		y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
1469 	} else
1470 		/* y belongs to the 2nd segment */
1471 		y = rtsc->y + rtsc->dy
1472 		    + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
1473 	return (y);
1474 }
1475 
1476 /*
1477  * update the runtime service curve by taking the minimum of the current
1478  * runtime service curve and the service curve starting at (x, y).
1479  */
1480 static void
1481 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1482 {
1483 	uint64_t y1, y2, dx, dy;
1484 
1485 	if (isc->sm1 <= isc->sm2) {
1486 		/* service curve is convex */
1487 		y1 = rtsc_x2y(rtsc, x);
1488 		if (y1 < y)
1489 			/* the current rtsc is smaller */
1490 			return;
1491 		rtsc->x = x;
1492 		rtsc->y = y;
1493 		return;
1494 	}
1495 
1496 	/*
1497 	 * service curve is concave
1498 	 * compute the two y values of the current rtsc
1499 	 *	y1: at x
1500 	 *	y2: at (x + dx)
1501 	 */
1502 	y1 = rtsc_x2y(rtsc, x);
1503 	if (y1 <= y) {
1504 		/* rtsc is below isc, no change to rtsc */
1505 		return;
1506 	}
1507 
1508 	y2 = rtsc_x2y(rtsc, x + isc->dx);
1509 	if (y2 >= y + isc->dy) {
1510 		/* rtsc is above isc, replace rtsc by isc */
1511 		rtsc->x = x;
1512 		rtsc->y = y;
1513 		rtsc->dx = isc->dx;
1514 		rtsc->dy = isc->dy;
1515 		return;
1516 	}
1517 
1518 	/*
1519 	 * the two curves intersect
1520 	 * compute the offsets (dx, dy) using the reverse
1521 	 * function of seg_x2y()
1522 	 *	seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
1523 	 */
1524 	dx = ((y1 - y) << SM_SHIFT) / (isc->sm1 - isc->sm2);
1525 	/*
1526 	 * check if (x, y1) belongs to the 1st segment of rtsc.
1527 	 * if so, add the offset.
1528 	 */
1529 	if (rtsc->x + rtsc->dx > x)
1530 		dx += rtsc->x + rtsc->dx - x;
1531 	dy = seg_x2y(dx, isc->sm1);
1532 
1533 	rtsc->x = x;
1534 	rtsc->y = y;
1535 	rtsc->dx = dx;
1536 	rtsc->dy = dy;
1537 }
1538 
1539 static void
1540 get_class_stats(struct hfsc_classstats *sp, struct hfsc_class *cl)
1541 {
1542 	sp->class_id = cl->cl_id;
1543 	sp->class_handle = cl->cl_handle;
1544 
1545 	if (cl->cl_rsc != NULL) {
1546 		sp->rsc.m1 = sm2m(cl->cl_rsc->sm1);
1547 		sp->rsc.d = dx2d(cl->cl_rsc->dx);
1548 		sp->rsc.m2 = sm2m(cl->cl_rsc->sm2);
1549 	} else {
1550 		sp->rsc.m1 = 0;
1551 		sp->rsc.d = 0;
1552 		sp->rsc.m2 = 0;
1553 	}
1554 	if (cl->cl_fsc != NULL) {
1555 		sp->fsc.m1 = sm2m(cl->cl_fsc->sm1);
1556 		sp->fsc.d = dx2d(cl->cl_fsc->dx);
1557 		sp->fsc.m2 = sm2m(cl->cl_fsc->sm2);
1558 	} else {
1559 		sp->fsc.m1 = 0;
1560 		sp->fsc.d = 0;
1561 		sp->fsc.m2 = 0;
1562 	}
1563 	if (cl->cl_usc != NULL) {
1564 		sp->usc.m1 = sm2m(cl->cl_usc->sm1);
1565 		sp->usc.d = dx2d(cl->cl_usc->dx);
1566 		sp->usc.m2 = sm2m(cl->cl_usc->sm2);
1567 	} else {
1568 		sp->usc.m1 = 0;
1569 		sp->usc.d = 0;
1570 		sp->usc.m2 = 0;
1571 	}
1572 
1573 	sp->total = cl->cl_total;
1574 	sp->cumul = cl->cl_cumul;
1575 
1576 	sp->d = cl->cl_d;
1577 	sp->e = cl->cl_e;
1578 	sp->vt = cl->cl_vt;
1579 	sp->f = cl->cl_f;
1580 
1581 	sp->initvt = cl->cl_initvt;
1582 	sp->vtperiod = cl->cl_vtperiod;
1583 	sp->parentperiod = cl->cl_parentperiod;
1584 	sp->nactive = cl->cl_nactive;
1585 	sp->vtoff = cl->cl_vtoff;
1586 	sp->cvtmax = cl->cl_cvtmax;
1587 	sp->myf = cl->cl_myf;
1588 	sp->cfmin = cl->cl_cfmin;
1589 	sp->cvtmin = cl->cl_cvtmin;
1590 	sp->myfadj = cl->cl_myfadj;
1591 	sp->vtadj = cl->cl_vtadj;
1592 
1593 	sp->cur_time = read_machclk();
1594 	sp->machclk_freq = machclk_freq;
1595 
1596 	sp->qlength = qlen(cl->cl_q);
1597 	sp->qlimit = qlimit(cl->cl_q);
1598 	sp->xmit_cnt = cl->cl_stats.xmit_cnt;
1599 	sp->drop_cnt = cl->cl_stats.drop_cnt;
1600 	sp->period = cl->cl_stats.period;
1601 
1602 	sp->qtype = qtype(cl->cl_q);
1603 #ifdef ALTQ_RED
1604 	if (q_is_red(cl->cl_q))
1605 		red_getstats(cl->cl_red, &sp->red[0]);
1606 #endif
1607 #ifdef ALTQ_RIO
1608 	if (q_is_rio(cl->cl_q))
1609 		rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
1610 #endif
1611 }
1612 
1613 /* convert a class handle to the corresponding class pointer */
1614 static struct hfsc_class *
1615 clh_to_clp(struct hfsc_if *hif, uint32_t chandle)
1616 {
1617 	int i;
1618 	struct hfsc_class *cl;
1619 
1620 	if (chandle == 0)
1621 		return (NULL);
1622 	/*
1623 	 * first, try optimistically the slot matching the lower bits of
1624 	 * the handle.  if it fails, do the linear table search.
1625 	 */
1626 	i = chandle % HFSC_MAX_CLASSES;
1627 	if ((cl = hif->hif_class_tbl[i]) != NULL && cl->cl_handle == chandle)
1628 		return (cl);
1629 	for (i = 0; i < HFSC_MAX_CLASSES; i++)
1630 		if ((cl = hif->hif_class_tbl[i]) != NULL &&
1631 		    cl->cl_handle == chandle)
1632 			return (cl);
1633 	return (NULL);
1634 }
1635 
1636 #endif /* ALTQ_HFSC */
1637