1 /* $KAME: altq_hfsc.c,v 1.25 2004/04/17 10:54:48 kjc Exp $ */
4 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
6 * Permission to use, copy, modify, and distribute this software and
7 * its documentation is hereby granted (including for commercial or
8 * for-profit use), provided that both the copyright notice and this
9 * permission notice appear in all copies of the software, derivative
10 * works, or modified versions, and any portions thereof.
12 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
13 * WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON PROVIDES THIS
14 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
15 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
16 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
17 * DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
19 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
20 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
21 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
22 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
24 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
27 * Carnegie Mellon encourages (but does not require) users of this
28 * software to return any improvements or extensions that they make,
29 * and to grant Carnegie Mellon the rights to redistribute these
30 * changes without encumbrance.
33 * H-FSC is described in Proceedings of SIGCOMM'97,
34 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
35 * Real-Time and Priority Service"
36 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
38 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
39 * when a class has an upperlimit, the fit-time is computed from the
40 * upperlimit service curve. the link-sharing scheduler does not schedule
41 * a class whose fit-time exceeds the current time.
46 #include "opt_inet6.h"
48 #ifdef ALTQ_HFSC /* hfsc is enabled by ALTQ_HFSC option in opt_altq.h */
50 #include <sys/param.h>
51 #include <sys/malloc.h>
53 #include <sys/socket.h>
54 #include <sys/systm.h>
55 #include <sys/errno.h>
56 #include <sys/queue.h>
57 #include <sys/thread.h>
60 #include <net/ifq_var.h>
61 #include <netinet/in.h>
63 #include <net/pf/pfvar.h>
64 #include <net/altq/altq.h>
65 #include <net/altq/altq_hfsc.h>
67 #include <sys/thread2.h>
72 static int hfsc_clear_interface(struct hfsc_if *);
73 static int hfsc_request(struct ifaltq *, int, void *);
74 static void hfsc_purge(struct hfsc_if *);
75 static struct hfsc_class *hfsc_class_create(struct hfsc_if *,
76 struct service_curve *,
77 struct service_curve *,
78 struct service_curve *,
79 struct hfsc_class *, int, int, int);
80 static int hfsc_class_destroy(struct hfsc_class *);
81 static struct hfsc_class *hfsc_nextclass(struct hfsc_class *);
82 static int hfsc_enqueue(struct ifaltq *, struct mbuf *,
83 struct altq_pktattr *);
84 static struct mbuf *hfsc_dequeue(struct ifaltq *, struct mbuf *, int);
86 static int hfsc_addq(struct hfsc_class *, struct mbuf *);
87 static struct mbuf *hfsc_getq(struct hfsc_class *);
88 static struct mbuf *hfsc_pollq(struct hfsc_class *);
89 static void hfsc_purgeq(struct hfsc_class *);
91 static void update_cfmin(struct hfsc_class *);
92 static void set_active(struct hfsc_class *, int);
93 static void set_passive(struct hfsc_class *);
95 static void init_ed(struct hfsc_class *, int);
96 static void update_ed(struct hfsc_class *, int);
97 static void update_d(struct hfsc_class *, int);
98 static void init_vf(struct hfsc_class *, int);
99 static void update_vf(struct hfsc_class *, int, uint64_t);
100 static ellist_t *ellist_alloc(void);
101 static void ellist_destroy(ellist_t *);
102 static void ellist_insert(struct hfsc_class *);
103 static void ellist_remove(struct hfsc_class *);
104 static void ellist_update(struct hfsc_class *);
105 struct hfsc_class *ellist_get_mindl(ellist_t *, uint64_t);
106 static actlist_t *actlist_alloc(void);
107 static void actlist_destroy(actlist_t *);
108 static void actlist_insert(struct hfsc_class *);
109 static void actlist_remove(struct hfsc_class *);
110 static void actlist_update(struct hfsc_class *);
112 static struct hfsc_class *actlist_firstfit(struct hfsc_class *, uint64_t);
114 static __inline uint64_t seg_x2y(uint64_t, uint64_t);
115 static __inline uint64_t seg_y2x(uint64_t, uint64_t);
116 static __inline uint64_t m2sm(u_int);
117 static __inline uint64_t m2ism(u_int);
118 static __inline uint64_t d2dx(u_int);
119 static u_int sm2m(uint64_t);
120 static u_int dx2d(uint64_t);
122 static void sc2isc(struct service_curve *, struct internal_sc *);
123 static void rtsc_init(struct runtime_sc *, struct internal_sc *,
125 static uint64_t rtsc_y2x(struct runtime_sc *, uint64_t);
126 static uint64_t rtsc_x2y(struct runtime_sc *, uint64_t);
127 static void rtsc_min(struct runtime_sc *, struct internal_sc *,
130 static void get_class_stats(struct hfsc_classstats *, struct hfsc_class *);
131 static struct hfsc_class *clh_to_clp(struct hfsc_if *, uint32_t);
136 #define is_a_parent_class(cl) ((cl)->cl_children != NULL)
138 #define HT_INFINITY 0xffffffffffffffffLL /* infinite time value */
141 hfsc_pfattach(struct pf_altq *a, struct ifaltq *ifq)
143 return altq_attach(ifq, ALTQT_HFSC, a->altq_disc,
144 hfsc_enqueue, hfsc_dequeue, hfsc_request, NULL, NULL);
148 hfsc_add_altq(struct pf_altq *a)
153 if ((ifp = ifunit(a->ifname)) == NULL)
155 if (!ifq_is_ready(&ifp->if_snd))
158 hif = kmalloc(sizeof(struct hfsc_if), M_ALTQ, M_WAITOK | M_ZERO);
160 hif->hif_eligible = ellist_alloc();
161 hif->hif_ifq = &ifp->if_snd;
162 ifq_purge(&ifp->if_snd);
164 /* keep the state in pf_altq */
171 hfsc_remove_altq(struct pf_altq *a)
175 if ((hif = a->altq_disc) == NULL)
179 hfsc_clear_interface(hif);
180 hfsc_class_destroy(hif->hif_rootclass);
182 ellist_destroy(hif->hif_eligible);
190 hfsc_add_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
192 struct hfsc_class *cl, *parent;
193 struct hfsc_opts *opts;
194 struct service_curve rtsc, lssc, ulsc;
196 KKASSERT(a->qid != 0);
198 opts = &a->pq_u.hfsc_opts;
200 if (a->parent_qid == HFSC_NULLCLASS_HANDLE && hif->hif_rootclass == NULL)
202 else if ((parent = clh_to_clp(hif, a->parent_qid)) == NULL)
205 if (clh_to_clp(hif, a->qid) != NULL)
208 rtsc.m1 = opts->rtsc_m1;
209 rtsc.d = opts->rtsc_d;
210 rtsc.m2 = opts->rtsc_m2;
211 lssc.m1 = opts->lssc_m1;
212 lssc.d = opts->lssc_d;
213 lssc.m2 = opts->lssc_m2;
214 ulsc.m1 = opts->ulsc_m1;
215 ulsc.d = opts->ulsc_d;
216 ulsc.m2 = opts->ulsc_m2;
218 cl = hfsc_class_create(hif, &rtsc, &lssc, &ulsc, parent, a->qlimit,
219 opts->flags, a->qid);
227 hfsc_add_queue(struct pf_altq *a)
236 /* XXX not MP safe */
237 if ((hif = a->altq_disc) == NULL)
242 error = hfsc_add_queue_locked(a, hif);
249 hfsc_remove_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
251 struct hfsc_class *cl;
253 if ((cl = clh_to_clp(hif, a->qid)) == NULL)
256 return (hfsc_class_destroy(cl));
260 hfsc_remove_queue(struct pf_altq *a)
266 /* XXX not MP safe */
267 if ((hif = a->altq_disc) == NULL)
272 error = hfsc_remove_queue_locked(a, hif);
279 hfsc_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
282 struct hfsc_class *cl;
283 struct hfsc_classstats stats;
287 if (*nbytes < sizeof(stats))
290 /* XXX not MP safe */
291 if ((hif = altq_lookup(a->ifname, ALTQT_HFSC)) == NULL)
297 if ((cl = clh_to_clp(hif, a->qid)) == NULL) {
302 get_class_stats(&stats, cl);
306 if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
308 *nbytes = sizeof(stats);
313 * bring the interface back to the initial state by discarding
314 * all the filters and classes except the root class.
317 hfsc_clear_interface(struct hfsc_if *hif)
319 struct hfsc_class *cl;
321 if (hif->hif_rootclass == NULL)
325 /* clear out the classes */
326 while ((cl = hif->hif_rootclass->cl_children) != NULL) {
328 * remove the first leaf class found in the hierarchy
331 for (; cl != NULL; cl = hfsc_nextclass(cl)) {
332 if (!is_a_parent_class(cl)) {
333 hfsc_class_destroy(cl);
343 hfsc_request(struct ifaltq *ifq, int req, void *arg)
345 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
357 /* discard all the queued packets on the interface */
359 hfsc_purge(struct hfsc_if *hif)
361 struct hfsc_class *cl;
363 for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl)) {
364 if (!qempty(cl->cl_q))
367 if (ifq_is_enabled(hif->hif_ifq))
368 hif->hif_ifq->ifq_len = 0;
372 hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
373 struct service_curve *fsc, struct service_curve *usc,
374 struct hfsc_class *parent, int qlimit, int flags, int qid)
376 struct hfsc_class *cl, *p;
379 if (hif->hif_classes >= HFSC_MAX_CLASSES)
383 if (flags & HFCF_RED) {
385 kprintf("hfsc_class_create: RED not configured for HFSC!\n");
391 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
392 cl->cl_q = kmalloc(sizeof(*cl->cl_q), M_ALTQ, M_WAITOK | M_ZERO);
393 cl->cl_actc = actlist_alloc();
396 qlimit = 50; /* use default */
397 qlimit(cl->cl_q) = qlimit;
398 qtype(cl->cl_q) = Q_DROPTAIL;
400 cl->cl_flags = flags;
402 if (flags & (HFCF_RED|HFCF_RIO)) {
403 int red_flags, red_pkttime;
407 if (rsc != NULL && rsc->m2 > m2)
409 if (fsc != NULL && fsc->m2 > m2)
411 if (usc != NULL && usc->m2 > m2)
415 if (flags & HFCF_ECN)
416 red_flags |= REDF_ECN;
418 if (flags & HFCF_CLEARDSCP)
419 red_flags |= RIOF_CLEARDSCP;
422 red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
424 red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
425 * 1000 * 1000 * 1000 / (m2 / 8);
426 if (flags & HFCF_RED) {
427 cl->cl_red = red_alloc(0, 0,
428 qlimit(cl->cl_q) * 10/100,
429 qlimit(cl->cl_q) * 30/100,
430 red_flags, red_pkttime);
431 if (cl->cl_red != NULL)
432 qtype(cl->cl_q) = Q_RED;
436 cl->cl_red = (red_t *)rio_alloc(0, NULL,
437 red_flags, red_pkttime);
438 if (cl->cl_red != NULL)
439 qtype(cl->cl_q) = Q_RIO;
443 #endif /* ALTQ_RED */
445 if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
446 cl->cl_rsc = kmalloc(sizeof(*cl->cl_rsc), M_ALTQ, M_WAITOK);
447 sc2isc(rsc, cl->cl_rsc);
448 rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
449 rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
451 if (fsc != NULL && (fsc->m1 != 0 || fsc->m2 != 0)) {
452 cl->cl_fsc = kmalloc(sizeof(*cl->cl_fsc), M_ALTQ, M_WAITOK);
453 if (cl->cl_fsc == NULL)
455 sc2isc(fsc, cl->cl_fsc);
456 rtsc_init(&cl->cl_virtual, cl->cl_fsc, 0, 0);
458 if (usc != NULL && (usc->m1 != 0 || usc->m2 != 0)) {
459 cl->cl_usc = kmalloc(sizeof(*cl->cl_usc), M_ALTQ, M_WAITOK);
460 if (cl->cl_usc == NULL)
462 sc2isc(usc, cl->cl_usc);
463 rtsc_init(&cl->cl_ulimit, cl->cl_usc, 0, 0);
466 cl->cl_id = hif->hif_classid++;
469 cl->cl_parent = parent;
475 * find a free slot in the class table. if the slot matching
476 * the lower bits of qid is free, use this slot. otherwise,
477 * use the first free slot.
479 i = qid % HFSC_MAX_CLASSES;
480 if (hif->hif_class_tbl[i] == NULL)
481 hif->hif_class_tbl[i] = cl;
483 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
484 if (hif->hif_class_tbl[i] == NULL) {
485 hif->hif_class_tbl[i] = cl;
489 if (i == HFSC_MAX_CLASSES) {
495 if (flags & HFCF_DEFAULTCLASS)
496 hif->hif_defaultclass = cl;
498 if (parent == NULL) {
499 /* this is root class */
500 hif->hif_rootclass = cl;
501 } else if (parent->cl_children == NULL) {
502 /* add this class to the children list of the parent */
503 parent->cl_children = cl;
505 p = parent->cl_children;
506 while (p->cl_siblings != NULL)
515 if (cl->cl_actc != NULL)
516 actlist_destroy(cl->cl_actc);
517 if (cl->cl_red != NULL) {
519 if (q_is_rio(cl->cl_q))
520 rio_destroy((rio_t *)cl->cl_red);
523 if (q_is_red(cl->cl_q))
524 red_destroy(cl->cl_red);
527 if (cl->cl_fsc != NULL)
528 kfree(cl->cl_fsc, M_ALTQ);
529 if (cl->cl_rsc != NULL)
530 kfree(cl->cl_rsc, M_ALTQ);
531 if (cl->cl_usc != NULL)
532 kfree(cl->cl_usc, M_ALTQ);
533 if (cl->cl_q != NULL)
534 kfree(cl->cl_q, M_ALTQ);
540 hfsc_class_destroy(struct hfsc_class *cl)
549 if (is_a_parent_class(cl))
554 if (!qempty(cl->cl_q))
557 if (cl->cl_parent == NULL) {
558 /* this is root class */
560 struct hfsc_class *p = cl->cl_parent->cl_children;
563 cl->cl_parent->cl_children = cl->cl_siblings;
566 if (p->cl_siblings == cl) {
567 p->cl_siblings = cl->cl_siblings;
570 } while ((p = p->cl_siblings) != NULL);
575 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
576 if (hif->hif_class_tbl[i] == cl) {
577 hif->hif_class_tbl[i] = NULL;
585 actlist_destroy(cl->cl_actc);
587 if (cl->cl_red != NULL) {
589 if (q_is_rio(cl->cl_q))
590 rio_destroy((rio_t *)cl->cl_red);
593 if (q_is_red(cl->cl_q))
594 red_destroy(cl->cl_red);
598 if (cl == hif->hif_rootclass)
599 hif->hif_rootclass = NULL;
600 if (cl == hif->hif_defaultclass)
601 hif->hif_defaultclass = NULL;
602 if (cl == hif->hif_pollcache)
603 hif->hif_pollcache = NULL;
605 if (cl->cl_usc != NULL)
606 kfree(cl->cl_usc, M_ALTQ);
607 if (cl->cl_fsc != NULL)
608 kfree(cl->cl_fsc, M_ALTQ);
609 if (cl->cl_rsc != NULL)
610 kfree(cl->cl_rsc, M_ALTQ);
611 kfree(cl->cl_q, M_ALTQ);
618 * hfsc_nextclass returns the next class in the tree.
620 * for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl))
623 static struct hfsc_class *
624 hfsc_nextclass(struct hfsc_class *cl)
626 if (cl->cl_children != NULL) {
627 cl = cl->cl_children;
628 } else if (cl->cl_siblings != NULL) {
629 cl = cl->cl_siblings;
631 while ((cl = cl->cl_parent) != NULL) {
632 if (cl->cl_siblings != NULL) {
633 cl = cl->cl_siblings;
643 * hfsc_enqueue is an enqueue function to be registered to
644 * (*altq_enqueue) in struct ifaltq.
647 hfsc_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
649 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
650 struct hfsc_class *cl;
653 /* grab class set by classifier */
654 if ((m->m_flags & M_PKTHDR) == 0) {
655 /* should not happen */
656 if_printf(ifq->altq_ifp, "altq: packet does not have pkthdr\n");
661 if (m->m_pkthdr.fw_flags & PF_MBUF_STRUCTURE)
662 cl = clh_to_clp(hif, m->m_pkthdr.pf.qid);
665 if (cl == NULL || is_a_parent_class(cl)) {
666 cl = hif->hif_defaultclass;
673 cl->cl_pktattr = NULL;
675 if (hfsc_addq(cl, m) != 0) {
676 /* drop occurred. mbuf was freed in hfsc_addq. */
677 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, len);
682 cl->cl_hif->hif_packets++;
684 /* successfully queued. */
685 if (qlen(cl->cl_q) == 1)
686 set_active(cl, m_pktlen(m));
692 * hfsc_dequeue is a dequeue function to be registered to
693 * (*altq_dequeue) in struct ifaltq.
695 * note: ALTDQ_POLL returns the next packet without removing the packet
696 * from the queue. ALTDQ_REMOVE is a normal dequeue operation.
697 * ALTDQ_REMOVE must return the same packet if called immediately
701 hfsc_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
703 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
704 struct hfsc_class *cl;
710 if (hif->hif_packets == 0) {
711 /* no packet in the tree */
716 cur_time = read_machclk();
718 if (op == ALTDQ_REMOVE && hif->hif_pollcache != NULL) {
719 cl = hif->hif_pollcache;
720 hif->hif_pollcache = NULL;
721 /* check if the class was scheduled by real-time criteria */
722 if (cl->cl_rsc != NULL)
723 realtime = (cl->cl_e <= cur_time);
726 * if there are eligible classes, use real-time criteria.
727 * find the class with the minimum deadline among
728 * the eligible classes.
730 if ((cl = ellist_get_mindl(hif->hif_eligible, cur_time)) != NULL) {
737 * use link-sharing criteria
738 * get the class with the minimum vt in the hierarchy
740 cl = hif->hif_rootclass;
741 while (is_a_parent_class(cl)) {
743 cl = actlist_firstfit(cl, cur_time);
747 kprintf("%d fit but none found\n",fits);
753 * update parent's cl_cvtmin.
754 * don't update if the new vt is smaller.
756 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
757 cl->cl_parent->cl_cvtmin = cl->cl_vt;
764 if (op == ALTDQ_POLL) {
765 hif->hif_pollcache = cl;
773 panic("hfsc_dequeue:");
775 cl->cl_hif->hif_packets--;
777 PKTCNTR_ADD(&cl->cl_stats.xmit_cnt, len);
779 update_vf(cl, len, cur_time);
783 if (!qempty(cl->cl_q)) {
784 if (cl->cl_rsc != NULL) {
786 next_len = m_pktlen(qhead(cl->cl_q));
789 update_ed(cl, next_len);
791 update_d(cl, next_len);
794 /* the class becomes passive */
799 KKASSERT(mpolled == NULL || m == mpolled);
804 hfsc_addq(struct hfsc_class *cl, struct mbuf *m)
808 if (q_is_rio(cl->cl_q))
809 return rio_addq((rio_t *)cl->cl_red, cl->cl_q,
813 if (q_is_red(cl->cl_q))
814 return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
816 if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
821 if (cl->cl_flags & HFCF_CLEARDSCP)
822 write_dsfield(m, cl->cl_pktattr, 0);
830 hfsc_getq(struct hfsc_class *cl)
833 if (q_is_rio(cl->cl_q))
834 return rio_getq((rio_t *)cl->cl_red, cl->cl_q);
837 if (q_is_red(cl->cl_q))
838 return red_getq(cl->cl_red, cl->cl_q);
840 return _getq(cl->cl_q);
844 hfsc_pollq(struct hfsc_class *cl)
846 return qhead(cl->cl_q);
850 hfsc_purgeq(struct hfsc_class *cl)
854 if (qempty(cl->cl_q))
857 while ((m = _getq(cl->cl_q)) != NULL) {
858 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, m_pktlen(m));
860 cl->cl_hif->hif_packets--;
861 cl->cl_hif->hif_ifq->ifq_len--;
863 KKASSERT(qlen(cl->cl_q) == 0);
865 update_vf(cl, 0, 0); /* remove cl from the actlist */
870 set_active(struct hfsc_class *cl, int len)
872 if (cl->cl_rsc != NULL)
874 if (cl->cl_fsc != NULL)
877 cl->cl_stats.period++;
881 set_passive(struct hfsc_class *cl)
883 if (cl->cl_rsc != NULL)
887 * actlist is now handled in update_vf() so that update_vf(cl, 0, 0)
888 * needs to be called explicitly to remove a class from actlist
893 init_ed(struct hfsc_class *cl, int next_len)
897 cur_time = read_machclk();
899 /* update the deadline curve */
900 rtsc_min(&cl->cl_deadline, cl->cl_rsc, cur_time, cl->cl_cumul);
903 * update the eligible curve.
904 * for concave, it is equal to the deadline curve.
905 * for convex, it is a linear curve with slope m2.
907 cl->cl_eligible = cl->cl_deadline;
908 if (cl->cl_rsc->sm1 <= cl->cl_rsc->sm2) {
909 cl->cl_eligible.dx = 0;
910 cl->cl_eligible.dy = 0;
913 /* compute e and d */
914 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
915 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
921 update_ed(struct hfsc_class *cl, int next_len)
923 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
924 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
930 update_d(struct hfsc_class *cl, int next_len)
932 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
936 init_vf(struct hfsc_class *cl, int len)
938 struct hfsc_class *max_cl, *p;
939 uint64_t vt, f, cur_time;
944 for ( ; cl->cl_parent != NULL; cl = cl->cl_parent) {
945 if (go_active && cl->cl_nactive++ == 0)
951 max_cl = actlist_last(cl->cl_parent->cl_actc);
952 if (max_cl != NULL) {
954 * set vt to the average of the min and max
955 * classes. if the parent's period didn't
956 * change, don't decrease vt of the class.
959 if (cl->cl_parent->cl_cvtmin != 0)
960 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
962 if (cl->cl_parent->cl_vtperiod !=
963 cl->cl_parentperiod || vt > cl->cl_vt)
967 * first child for a new parent backlog period.
968 * add parent's cvtmax to vtoff of children
969 * to make a new vt (vtoff + vt) larger than
970 * the vt in the last period for all children.
972 vt = cl->cl_parent->cl_cvtmax;
973 for (p = cl->cl_parent->cl_children; p != NULL;
977 cl->cl_parent->cl_cvtmax = 0;
978 cl->cl_parent->cl_cvtmin = 0;
980 cl->cl_initvt = cl->cl_vt;
982 /* update the virtual curve */
983 vt = cl->cl_vt + cl->cl_vtoff;
984 rtsc_min(&cl->cl_virtual, cl->cl_fsc, vt, cl->cl_total);
985 if (cl->cl_virtual.x == vt) {
986 cl->cl_virtual.x -= cl->cl_vtoff;
991 cl->cl_vtperiod++; /* increment vt period */
992 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
993 if (cl->cl_parent->cl_nactive == 0)
994 cl->cl_parentperiod++;
999 if (cl->cl_usc != NULL) {
1000 /* class has upper limit curve */
1002 cur_time = read_machclk();
1004 /* update the ulimit curve */
1005 rtsc_min(&cl->cl_ulimit, cl->cl_usc, cur_time,
1008 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
1014 if (cl->cl_myf > cl->cl_cfmin)
1018 if (f != cl->cl_f) {
1020 update_cfmin(cl->cl_parent);
1026 update_vf(struct hfsc_class *cl, int len, uint64_t cur_time)
1028 uint64_t f, myf_bound, delta;
1031 go_passive = qempty(cl->cl_q);
1033 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
1034 cl->cl_total += len;
1036 if (cl->cl_fsc == NULL || cl->cl_nactive == 0)
1039 if (go_passive && --cl->cl_nactive == 0)
1045 /* no more active child, going passive */
1047 /* update cvtmax of the parent class */
1048 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
1049 cl->cl_parent->cl_cvtmax = cl->cl_vt;
1051 /* remove this class from the vt list */
1054 update_cfmin(cl->cl_parent);
1062 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
1063 - cl->cl_vtoff + cl->cl_vtadj;
1066 * if vt of the class is smaller than cvtmin,
1067 * the class was skipped in the past due to non-fit.
1068 * if so, we need to adjust vtadj.
1070 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
1071 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
1072 cl->cl_vt = cl->cl_parent->cl_cvtmin;
1075 /* update the vt list */
1078 if (cl->cl_usc != NULL) {
1079 cl->cl_myf = cl->cl_myfadj
1080 + rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
1083 * if myf lags behind by more than one clock tick
1084 * from the current time, adjust myfadj to prevent
1085 * a rate-limited class from going greedy.
1086 * in a steady state under rate-limiting, myf
1087 * fluctuates within one clock tick.
1089 myf_bound = cur_time - machclk_per_tick;
1090 if (cl->cl_myf < myf_bound) {
1091 delta = cur_time - cl->cl_myf;
1092 cl->cl_myfadj += delta;
1093 cl->cl_myf += delta;
1097 /* cl_f is max(cl_myf, cl_cfmin) */
1098 if (cl->cl_myf > cl->cl_cfmin)
1102 if (f != cl->cl_f) {
1104 update_cfmin(cl->cl_parent);
1110 update_cfmin(struct hfsc_class *cl)
1112 struct hfsc_class *p;
1115 if (TAILQ_EMPTY(cl->cl_actc)) {
1119 cfmin = HT_INFINITY;
1120 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1125 if (p->cl_f < cfmin)
1128 cl->cl_cfmin = cfmin;
1132 * TAILQ based ellist and actlist implementation
1133 * (ion wanted to make a calendar queue based implementation)
1136 * eligible list holds backlogged classes being sorted by their eligible times.
1137 * there is one eligible list per interface.
1145 head = kmalloc(sizeof(ellist_t *), M_ALTQ, M_WAITOK);
1151 ellist_destroy(ellist_t *head)
1153 kfree(head, M_ALTQ);
1157 ellist_insert(struct hfsc_class *cl)
1159 struct hfsc_if *hif = cl->cl_hif;
1160 struct hfsc_class *p;
1162 /* check the last entry first */
1163 if ((p = TAILQ_LAST(hif->hif_eligible, _eligible)) == NULL ||
1164 p->cl_e <= cl->cl_e) {
1165 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1169 TAILQ_FOREACH(p, hif->hif_eligible, cl_ellist) {
1170 if (cl->cl_e < p->cl_e) {
1171 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1175 KKASSERT(0); /* should not reach here */
1179 ellist_remove(struct hfsc_class *cl)
1181 struct hfsc_if *hif = cl->cl_hif;
1183 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1187 ellist_update(struct hfsc_class *cl)
1189 struct hfsc_if *hif = cl->cl_hif;
1190 struct hfsc_class *p, *last;
1193 * the eligible time of a class increases monotonically.
1194 * if the next entry has a larger eligible time, nothing to do.
1196 p = TAILQ_NEXT(cl, cl_ellist);
1197 if (p == NULL || cl->cl_e <= p->cl_e)
1200 /* check the last entry */
1201 last = TAILQ_LAST(hif->hif_eligible, _eligible);
1202 KKASSERT(last != NULL);
1203 if (last->cl_e <= cl->cl_e) {
1204 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1205 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1210 * the new position must be between the next entry
1211 * and the last entry
1213 while ((p = TAILQ_NEXT(p, cl_ellist)) != NULL) {
1214 if (cl->cl_e < p->cl_e) {
1215 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1216 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1220 KKASSERT(0); /* should not reach here */
1223 /* find the class with the minimum deadline among the eligible classes */
1225 ellist_get_mindl(ellist_t *head, uint64_t cur_time)
1227 struct hfsc_class *p, *cl = NULL;
1229 TAILQ_FOREACH(p, head, cl_ellist) {
1230 if (p->cl_e > cur_time)
1232 if (cl == NULL || p->cl_d < cl->cl_d)
1239 * active children list holds backlogged child classes being sorted
1240 * by their virtual time.
1241 * each intermediate class has one active children list.
1248 head = kmalloc(sizeof(*head), M_ALTQ, M_WAITOK);
1254 actlist_destroy(actlist_t *head)
1256 kfree(head, M_ALTQ);
1259 actlist_insert(struct hfsc_class *cl)
1261 struct hfsc_class *p;
1263 /* check the last entry first */
1264 if ((p = TAILQ_LAST(cl->cl_parent->cl_actc, _active)) == NULL
1265 || p->cl_vt <= cl->cl_vt) {
1266 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1270 TAILQ_FOREACH(p, cl->cl_parent->cl_actc, cl_actlist) {
1271 if (cl->cl_vt < p->cl_vt) {
1272 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1276 KKASSERT(0); /* should not reach here */
1280 actlist_remove(struct hfsc_class *cl)
1282 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1286 actlist_update(struct hfsc_class *cl)
1288 struct hfsc_class *p, *last;
1291 * the virtual time of a class increases monotonically during its
1292 * backlogged period.
1293 * if the next entry has a larger virtual time, nothing to do.
1295 p = TAILQ_NEXT(cl, cl_actlist);
1296 if (p == NULL || cl->cl_vt < p->cl_vt)
1299 /* check the last entry */
1300 last = TAILQ_LAST(cl->cl_parent->cl_actc, _active);
1301 KKASSERT(last != NULL);
1302 if (last->cl_vt <= cl->cl_vt) {
1303 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1304 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1309 * the new position must be between the next entry
1310 * and the last entry
1312 while ((p = TAILQ_NEXT(p, cl_actlist)) != NULL) {
1313 if (cl->cl_vt < p->cl_vt) {
1314 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1315 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1319 KKASSERT(0); /* should not reach here */
1322 static struct hfsc_class *
1323 actlist_firstfit(struct hfsc_class *cl, uint64_t cur_time)
1325 struct hfsc_class *p;
1327 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1328 if (p->cl_f <= cur_time)
1335 * service curve support functions
1337 * external service curve parameters
1340 * internal service curve parameters
1341 * sm: (bytes/tsc_interval) << SM_SHIFT
1342 * ism: (tsc_count/byte) << ISM_SHIFT
1345 * SM_SHIFT and ISM_SHIFT are scaled in order to keep effective digits.
1346 * we should be able to handle 100K-1Gbps linkspeed with 200Hz-1GHz CPU
1347 * speed. SM_SHIFT and ISM_SHIFT are selected to have at least 3 effective
1348 * digits in decimal using the following table.
1350 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
1351 * ----------+-------------------------------------------------------
1352 * bytes/nsec 12.5e-6 125e-6 1250e-6 12500e-6 125000e-6
1353 * sm(500MHz) 25.0e-6 250e-6 2500e-6 25000e-6 250000e-6
1354 * sm(200MHz) 62.5e-6 625e-6 6250e-6 62500e-6 625000e-6
1356 * nsec/byte 80000 8000 800 80 8
1357 * ism(500MHz) 40000 4000 400 40 4
1358 * ism(200MHz) 16000 1600 160 16 1.6
1361 #define ISM_SHIFT 10
1363 #define SM_MASK ((1LL << SM_SHIFT) - 1)
1364 #define ISM_MASK ((1LL << ISM_SHIFT) - 1)
1366 static __inline uint64_t
1367 seg_x2y(uint64_t x, uint64_t sm)
1373 * y = x * sm >> SM_SHIFT
1374 * but divide it for the upper and lower bits to avoid overflow
1376 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
1380 static __inline uint64_t
1381 seg_y2x(uint64_t y, uint64_t ism)
1387 else if (ism == HT_INFINITY)
1390 x = (y >> ISM_SHIFT) * ism + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
1395 static __inline uint64_t
1400 sm = ((uint64_t)m << SM_SHIFT) / 8 / machclk_freq;
1404 static __inline uint64_t
1412 ism = ((uint64_t)machclk_freq << ISM_SHIFT) * 8 / m;
1416 static __inline uint64_t
1421 dx = ((uint64_t)d * machclk_freq) / 1000;
1430 m = (sm * 8 * machclk_freq) >> SM_SHIFT;
1439 d = dx * 1000 / machclk_freq;
1444 sc2isc(struct service_curve *sc, struct internal_sc *isc)
1446 isc->sm1 = m2sm(sc->m1);
1447 isc->ism1 = m2ism(sc->m1);
1448 isc->dx = d2dx(sc->d);
1449 isc->dy = seg_x2y(isc->dx, isc->sm1);
1450 isc->sm2 = m2sm(sc->m2);
1451 isc->ism2 = m2ism(sc->m2);
1455 * initialize the runtime service curve with the given internal
1456 * service curve starting at (x, y).
1459 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1463 rtsc->sm1 = isc->sm1;
1464 rtsc->ism1 = isc->ism1;
1467 rtsc->sm2 = isc->sm2;
1468 rtsc->ism2 = isc->ism2;
1472 * calculate the y-projection of the runtime service curve by the
1473 * given x-projection value
1476 rtsc_y2x(struct runtime_sc *rtsc, uint64_t y)
1482 } else if (y <= rtsc->y + rtsc->dy) {
1483 /* x belongs to the 1st segment */
1485 x = rtsc->x + rtsc->dx;
1487 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
1489 /* x belongs to the 2nd segment */
1490 x = rtsc->x + rtsc->dx
1491 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
1497 rtsc_x2y(struct runtime_sc *rtsc, uint64_t x)
1503 } else if (x <= rtsc->x + rtsc->dx) {
1504 /* y belongs to the 1st segment */
1505 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
1507 /* y belongs to the 2nd segment */
1508 y = rtsc->y + rtsc->dy
1509 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
1514 * update the runtime service curve by taking the minimum of the current
1515 * runtime service curve and the service curve starting at (x, y).
1518 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1520 uint64_t y1, y2, dx, dy;
1522 if (isc->sm1 <= isc->sm2) {
1523 /* service curve is convex */
1524 y1 = rtsc_x2y(rtsc, x);
1526 /* the current rtsc is smaller */
1534 * service curve is concave
1535 * compute the two y values of the current rtsc
1539 y1 = rtsc_x2y(rtsc, x);
1541 /* rtsc is below isc, no change to rtsc */
1545 y2 = rtsc_x2y(rtsc, x + isc->dx);
1546 if (y2 >= y + isc->dy) {
1547 /* rtsc is above isc, replace rtsc by isc */
1556 * the two curves intersect
1557 * compute the offsets (dx, dy) using the reverse
1558 * function of seg_x2y()
1559 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
1561 dx = ((y1 - y) << SM_SHIFT) / (isc->sm1 - isc->sm2);
1563 * check if (x, y1) belongs to the 1st segment of rtsc.
1564 * if so, add the offset.
1566 if (rtsc->x + rtsc->dx > x)
1567 dx += rtsc->x + rtsc->dx - x;
1568 dy = seg_x2y(dx, isc->sm1);
1577 get_class_stats(struct hfsc_classstats *sp, struct hfsc_class *cl)
1579 sp->class_id = cl->cl_id;
1580 sp->class_handle = cl->cl_handle;
1582 if (cl->cl_rsc != NULL) {
1583 sp->rsc.m1 = sm2m(cl->cl_rsc->sm1);
1584 sp->rsc.d = dx2d(cl->cl_rsc->dx);
1585 sp->rsc.m2 = sm2m(cl->cl_rsc->sm2);
1591 if (cl->cl_fsc != NULL) {
1592 sp->fsc.m1 = sm2m(cl->cl_fsc->sm1);
1593 sp->fsc.d = dx2d(cl->cl_fsc->dx);
1594 sp->fsc.m2 = sm2m(cl->cl_fsc->sm2);
1600 if (cl->cl_usc != NULL) {
1601 sp->usc.m1 = sm2m(cl->cl_usc->sm1);
1602 sp->usc.d = dx2d(cl->cl_usc->dx);
1603 sp->usc.m2 = sm2m(cl->cl_usc->sm2);
1610 sp->total = cl->cl_total;
1611 sp->cumul = cl->cl_cumul;
1618 sp->initvt = cl->cl_initvt;
1619 sp->vtperiod = cl->cl_vtperiod;
1620 sp->parentperiod = cl->cl_parentperiod;
1621 sp->nactive = cl->cl_nactive;
1622 sp->vtoff = cl->cl_vtoff;
1623 sp->cvtmax = cl->cl_cvtmax;
1624 sp->myf = cl->cl_myf;
1625 sp->cfmin = cl->cl_cfmin;
1626 sp->cvtmin = cl->cl_cvtmin;
1627 sp->myfadj = cl->cl_myfadj;
1628 sp->vtadj = cl->cl_vtadj;
1630 sp->cur_time = read_machclk();
1631 sp->machclk_freq = machclk_freq;
1633 sp->qlength = qlen(cl->cl_q);
1634 sp->qlimit = qlimit(cl->cl_q);
1635 sp->xmit_cnt = cl->cl_stats.xmit_cnt;
1636 sp->drop_cnt = cl->cl_stats.drop_cnt;
1637 sp->period = cl->cl_stats.period;
1639 sp->qtype = qtype(cl->cl_q);
1641 if (q_is_red(cl->cl_q))
1642 red_getstats(cl->cl_red, &sp->red[0]);
1645 if (q_is_rio(cl->cl_q))
1646 rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
1650 /* convert a class handle to the corresponding class pointer */
1651 static struct hfsc_class *
1652 clh_to_clp(struct hfsc_if *hif, uint32_t chandle)
1655 struct hfsc_class *cl;
1660 * first, try optimistically the slot matching the lower bits of
1661 * the handle. if it fails, do the linear table search.
1663 i = chandle % HFSC_MAX_CLASSES;
1664 if ((cl = hif->hif_class_tbl[i]) != NULL && cl->cl_handle == chandle)
1666 for (i = 0; i < HFSC_MAX_CLASSES; i++)
1667 if ((cl = hif->hif_class_tbl[i]) != NULL &&
1668 cl->cl_handle == chandle)
1673 #endif /* ALTQ_HFSC */