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>
69 #define HFSC_SUBQ_INDEX ALTQ_SUBQ_INDEX_DEFAULT
70 #define HFSC_LOCK(ifq) \
71 ALTQ_SQ_LOCK(&(ifq)->altq_subq[HFSC_SUBQ_INDEX])
72 #define HFSC_UNLOCK(ifq) \
73 ALTQ_SQ_UNLOCK(&(ifq)->altq_subq[HFSC_SUBQ_INDEX])
78 static int hfsc_clear_interface(struct hfsc_if *);
79 static int hfsc_request(struct ifaltq_subque *, int, void *);
80 static void hfsc_purge(struct hfsc_if *);
81 static struct hfsc_class *hfsc_class_create(struct hfsc_if *,
82 struct service_curve *,
83 struct service_curve *,
84 struct service_curve *,
85 struct hfsc_class *, int, int, int);
86 static int hfsc_class_destroy(struct hfsc_class *);
87 static struct hfsc_class *hfsc_nextclass(struct hfsc_class *);
88 static int hfsc_enqueue(struct ifaltq_subque *, struct mbuf *,
89 struct altq_pktattr *);
90 static struct mbuf *hfsc_dequeue(struct ifaltq_subque *, struct mbuf *, int);
92 static int hfsc_addq(struct hfsc_class *, struct mbuf *);
93 static struct mbuf *hfsc_getq(struct hfsc_class *);
94 static struct mbuf *hfsc_pollq(struct hfsc_class *);
95 static void hfsc_purgeq(struct hfsc_class *);
97 static void update_cfmin(struct hfsc_class *);
98 static void set_active(struct hfsc_class *, int);
99 static void set_passive(struct hfsc_class *);
101 static void init_ed(struct hfsc_class *, int);
102 static void update_ed(struct hfsc_class *, int);
103 static void update_d(struct hfsc_class *, int);
104 static void init_vf(struct hfsc_class *, int);
105 static void update_vf(struct hfsc_class *, int, uint64_t);
106 static ellist_t *ellist_alloc(void);
107 static void ellist_destroy(ellist_t *);
108 static void ellist_insert(struct hfsc_class *);
109 static void ellist_remove(struct hfsc_class *);
110 static void ellist_update(struct hfsc_class *);
111 struct hfsc_class *ellist_get_mindl(ellist_t *, uint64_t);
112 static actlist_t *actlist_alloc(void);
113 static void actlist_destroy(actlist_t *);
114 static void actlist_insert(struct hfsc_class *);
115 static void actlist_remove(struct hfsc_class *);
116 static void actlist_update(struct hfsc_class *);
118 static struct hfsc_class *actlist_firstfit(struct hfsc_class *, uint64_t);
120 static __inline uint64_t seg_x2y(uint64_t, uint64_t);
121 static __inline uint64_t seg_y2x(uint64_t, uint64_t);
122 static __inline uint64_t m2sm(u_int);
123 static __inline uint64_t m2ism(u_int);
124 static __inline uint64_t d2dx(u_int);
125 static u_int sm2m(uint64_t);
126 static u_int dx2d(uint64_t);
128 static void sc2isc(struct service_curve *, struct internal_sc *);
129 static void rtsc_init(struct runtime_sc *, struct internal_sc *,
131 static uint64_t rtsc_y2x(struct runtime_sc *, uint64_t);
132 static uint64_t rtsc_x2y(struct runtime_sc *, uint64_t);
133 static void rtsc_min(struct runtime_sc *, struct internal_sc *,
136 static void get_class_stats(struct hfsc_classstats *, struct hfsc_class *);
137 static struct hfsc_class *clh_to_clp(struct hfsc_if *, uint32_t);
142 #define is_a_parent_class(cl) ((cl)->cl_children != NULL)
144 #define HT_INFINITY 0xffffffffffffffffLL /* infinite time value */
147 hfsc_pfattach(struct pf_altq *a, struct ifaltq *ifq)
149 return altq_attach(ifq, ALTQT_HFSC, a->altq_disc,
150 hfsc_enqueue, hfsc_dequeue, hfsc_request, NULL, NULL);
154 hfsc_add_altq(struct pf_altq *a)
159 if ((ifp = ifunit(a->ifname)) == NULL)
161 if (!ifq_is_ready(&ifp->if_snd))
164 hif = kmalloc(sizeof(struct hfsc_if), M_ALTQ, M_WAITOK | M_ZERO);
166 hif->hif_eligible = ellist_alloc();
167 hif->hif_ifq = &ifp->if_snd;
168 ifq_purge_all(&ifp->if_snd);
170 /* keep the state in pf_altq */
177 hfsc_remove_altq(struct pf_altq *a)
181 if ((hif = a->altq_disc) == NULL)
185 hfsc_clear_interface(hif);
186 hfsc_class_destroy(hif->hif_rootclass);
188 ellist_destroy(hif->hif_eligible);
196 hfsc_add_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
198 struct hfsc_class *cl, *parent;
199 struct hfsc_opts *opts;
200 struct service_curve rtsc, lssc, ulsc;
202 KKASSERT(a->qid != 0);
204 opts = &a->pq_u.hfsc_opts;
206 if (a->parent_qid == HFSC_NULLCLASS_HANDLE && hif->hif_rootclass == NULL)
208 else if ((parent = clh_to_clp(hif, a->parent_qid)) == NULL)
211 if (clh_to_clp(hif, a->qid) != NULL)
214 rtsc.m1 = opts->rtsc_m1;
215 rtsc.d = opts->rtsc_d;
216 rtsc.m2 = opts->rtsc_m2;
217 lssc.m1 = opts->lssc_m1;
218 lssc.d = opts->lssc_d;
219 lssc.m2 = opts->lssc_m2;
220 ulsc.m1 = opts->ulsc_m1;
221 ulsc.d = opts->ulsc_d;
222 ulsc.m2 = opts->ulsc_m2;
224 cl = hfsc_class_create(hif, &rtsc, &lssc, &ulsc, parent, a->qlimit,
225 opts->flags, a->qid);
233 hfsc_add_queue(struct pf_altq *a)
242 /* XXX not MP safe */
243 if ((hif = a->altq_disc) == NULL)
248 error = hfsc_add_queue_locked(a, hif);
255 hfsc_remove_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
257 struct hfsc_class *cl;
259 if ((cl = clh_to_clp(hif, a->qid)) == NULL)
262 return (hfsc_class_destroy(cl));
266 hfsc_remove_queue(struct pf_altq *a)
272 /* XXX not MP safe */
273 if ((hif = a->altq_disc) == NULL)
278 error = hfsc_remove_queue_locked(a, hif);
285 hfsc_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
288 struct hfsc_class *cl;
289 struct hfsc_classstats stats;
293 if (*nbytes < sizeof(stats))
296 /* XXX not MP safe */
297 if ((hif = altq_lookup(a->ifname, ALTQT_HFSC)) == NULL)
303 if ((cl = clh_to_clp(hif, a->qid)) == NULL) {
308 get_class_stats(&stats, cl);
312 if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
314 *nbytes = sizeof(stats);
319 * bring the interface back to the initial state by discarding
320 * all the filters and classes except the root class.
323 hfsc_clear_interface(struct hfsc_if *hif)
325 struct hfsc_class *cl;
327 if (hif->hif_rootclass == NULL)
331 /* clear out the classes */
332 while ((cl = hif->hif_rootclass->cl_children) != NULL) {
334 * remove the first leaf class found in the hierarchy
337 for (; cl != NULL; cl = hfsc_nextclass(cl)) {
338 if (!is_a_parent_class(cl)) {
339 hfsc_class_destroy(cl);
349 hfsc_request(struct ifaltq_subque *ifsq, int req, void *arg)
351 struct ifaltq *ifq = ifsq->ifsq_altq;
352 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
357 if (ifsq_get_index(ifsq) == HFSC_SUBQ_INDEX) {
361 * Race happened, the unrelated subqueue was
362 * picked during the packet scheduler transition.
364 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
372 /* discard all the queued packets on the interface */
374 hfsc_purge(struct hfsc_if *hif)
376 struct hfsc_class *cl;
378 for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl)) {
379 if (!qempty(cl->cl_q))
382 if (ifq_is_enabled(hif->hif_ifq))
383 hif->hif_ifq->altq_subq[HFSC_SUBQ_INDEX].ifq_len = 0;
387 hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
388 struct service_curve *fsc, struct service_curve *usc,
389 struct hfsc_class *parent, int qlimit, int flags, int qid)
391 struct hfsc_class *cl, *p;
394 if (hif->hif_classes >= HFSC_MAX_CLASSES)
398 if (flags & HFCF_RED) {
400 kprintf("hfsc_class_create: RED not configured for HFSC!\n");
406 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
407 cl->cl_q = kmalloc(sizeof(*cl->cl_q), M_ALTQ, M_WAITOK | M_ZERO);
408 cl->cl_actc = actlist_alloc();
411 qlimit = 50; /* use default */
412 qlimit(cl->cl_q) = qlimit;
413 qtype(cl->cl_q) = Q_DROPTAIL;
415 cl->cl_flags = flags;
417 if (flags & (HFCF_RED|HFCF_RIO)) {
418 int red_flags, red_pkttime;
422 if (rsc != NULL && rsc->m2 > m2)
424 if (fsc != NULL && fsc->m2 > m2)
426 if (usc != NULL && usc->m2 > m2)
430 if (flags & HFCF_ECN)
431 red_flags |= REDF_ECN;
433 if (flags & HFCF_CLEARDSCP)
434 red_flags |= RIOF_CLEARDSCP;
437 red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
439 red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
440 * 1000 * 1000 * 1000 / (m2 / 8);
441 if (flags & HFCF_RED) {
442 cl->cl_red = red_alloc(0, 0,
443 qlimit(cl->cl_q) * 10/100,
444 qlimit(cl->cl_q) * 30/100,
445 red_flags, red_pkttime);
446 if (cl->cl_red != NULL)
447 qtype(cl->cl_q) = Q_RED;
451 cl->cl_red = (red_t *)rio_alloc(0, NULL,
452 red_flags, red_pkttime);
453 if (cl->cl_red != NULL)
454 qtype(cl->cl_q) = Q_RIO;
458 #endif /* ALTQ_RED */
460 if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
461 cl->cl_rsc = kmalloc(sizeof(*cl->cl_rsc), M_ALTQ, M_WAITOK);
462 sc2isc(rsc, cl->cl_rsc);
463 rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
464 rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
466 if (fsc != NULL && (fsc->m1 != 0 || fsc->m2 != 0)) {
467 cl->cl_fsc = kmalloc(sizeof(*cl->cl_fsc), M_ALTQ, M_WAITOK);
468 sc2isc(fsc, cl->cl_fsc);
469 rtsc_init(&cl->cl_virtual, cl->cl_fsc, 0, 0);
471 if (usc != NULL && (usc->m1 != 0 || usc->m2 != 0)) {
472 cl->cl_usc = kmalloc(sizeof(*cl->cl_usc), M_ALTQ, M_WAITOK);
473 sc2isc(usc, cl->cl_usc);
474 rtsc_init(&cl->cl_ulimit, cl->cl_usc, 0, 0);
477 cl->cl_id = hif->hif_classid++;
480 cl->cl_parent = parent;
486 * find a free slot in the class table. if the slot matching
487 * the lower bits of qid is free, use this slot. otherwise,
488 * use the first free slot.
490 i = qid % HFSC_MAX_CLASSES;
491 if (hif->hif_class_tbl[i] == NULL)
492 hif->hif_class_tbl[i] = cl;
494 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
495 if (hif->hif_class_tbl[i] == NULL) {
496 hif->hif_class_tbl[i] = cl;
500 if (i == HFSC_MAX_CLASSES) {
506 if (flags & HFCF_DEFAULTCLASS)
507 hif->hif_defaultclass = cl;
509 if (parent == NULL) {
510 /* this is root class */
511 hif->hif_rootclass = cl;
512 } else if (parent->cl_children == NULL) {
513 /* add this class to the children list of the parent */
514 parent->cl_children = cl;
516 p = parent->cl_children;
517 while (p->cl_siblings != NULL)
526 if (cl->cl_actc != NULL)
527 actlist_destroy(cl->cl_actc);
528 if (cl->cl_red != NULL) {
530 if (q_is_rio(cl->cl_q))
531 rio_destroy((rio_t *)cl->cl_red);
534 if (q_is_red(cl->cl_q))
535 red_destroy(cl->cl_red);
538 if (cl->cl_fsc != NULL)
539 kfree(cl->cl_fsc, M_ALTQ);
540 if (cl->cl_rsc != NULL)
541 kfree(cl->cl_rsc, M_ALTQ);
542 if (cl->cl_usc != NULL)
543 kfree(cl->cl_usc, M_ALTQ);
544 if (cl->cl_q != NULL)
545 kfree(cl->cl_q, M_ALTQ);
551 hfsc_class_destroy(struct hfsc_class *cl)
560 if (is_a_parent_class(cl))
565 if (!qempty(cl->cl_q))
568 if (cl->cl_parent == NULL) {
569 /* this is root class */
571 struct hfsc_class *p = cl->cl_parent->cl_children;
574 cl->cl_parent->cl_children = cl->cl_siblings;
577 if (p->cl_siblings == cl) {
578 p->cl_siblings = cl->cl_siblings;
581 } while ((p = p->cl_siblings) != NULL);
586 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
587 if (hif->hif_class_tbl[i] == cl) {
588 hif->hif_class_tbl[i] = NULL;
596 actlist_destroy(cl->cl_actc);
598 if (cl->cl_red != NULL) {
600 if (q_is_rio(cl->cl_q))
601 rio_destroy((rio_t *)cl->cl_red);
604 if (q_is_red(cl->cl_q))
605 red_destroy(cl->cl_red);
609 if (cl == hif->hif_rootclass)
610 hif->hif_rootclass = NULL;
611 if (cl == hif->hif_defaultclass)
612 hif->hif_defaultclass = NULL;
613 if (cl == hif->hif_pollcache)
614 hif->hif_pollcache = NULL;
616 if (cl->cl_usc != NULL)
617 kfree(cl->cl_usc, M_ALTQ);
618 if (cl->cl_fsc != NULL)
619 kfree(cl->cl_fsc, M_ALTQ);
620 if (cl->cl_rsc != NULL)
621 kfree(cl->cl_rsc, M_ALTQ);
622 kfree(cl->cl_q, M_ALTQ);
629 * hfsc_nextclass returns the next class in the tree.
631 * for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl))
634 static struct hfsc_class *
635 hfsc_nextclass(struct hfsc_class *cl)
637 if (cl->cl_children != NULL) {
638 cl = cl->cl_children;
639 } else if (cl->cl_siblings != NULL) {
640 cl = cl->cl_siblings;
642 while ((cl = cl->cl_parent) != NULL) {
643 if (cl->cl_siblings != NULL) {
644 cl = cl->cl_siblings;
654 * hfsc_enqueue is an enqueue function to be registered to
655 * (*altq_enqueue) in struct ifaltq.
658 hfsc_enqueue(struct ifaltq_subque *ifsq, struct mbuf *m,
659 struct altq_pktattr *pktattr)
661 struct ifaltq *ifq = ifsq->ifsq_altq;
662 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
663 struct hfsc_class *cl;
666 if (ifsq_get_index(ifsq) != HFSC_SUBQ_INDEX) {
668 * Race happened, the unrelated subqueue was
669 * picked during the packet scheduler transition.
671 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
676 /* grab class set by classifier */
677 if ((m->m_flags & M_PKTHDR) == 0) {
678 /* should not happen */
679 if_printf(ifq->altq_ifp, "altq: packet does not have pkthdr\n");
684 if (m->m_pkthdr.fw_flags & PF_MBUF_STRUCTURE)
685 cl = clh_to_clp(hif, m->m_pkthdr.pf.qid);
688 if (cl == NULL || is_a_parent_class(cl)) {
689 cl = hif->hif_defaultclass;
696 cl->cl_pktattr = NULL;
698 if (hfsc_addq(cl, m) != 0) {
699 /* drop occurred. mbuf was freed in hfsc_addq. */
700 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, len);
705 cl->cl_hif->hif_packets++;
707 /* successfully queued. */
708 if (qlen(cl->cl_q) == 1)
709 set_active(cl, m_pktlen(m));
715 * hfsc_dequeue is a dequeue function to be registered to
716 * (*altq_dequeue) in struct ifaltq.
718 * note: ALTDQ_POLL returns the next packet without removing the packet
719 * from the queue. ALTDQ_REMOVE is a normal dequeue operation.
720 * ALTDQ_REMOVE must return the same packet if called immediately
724 hfsc_dequeue(struct ifaltq_subque *ifsq, struct mbuf *mpolled, int op)
726 struct ifaltq *ifq = ifsq->ifsq_altq;
727 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
728 struct hfsc_class *cl;
734 if (ifsq_get_index(ifsq) != HFSC_SUBQ_INDEX) {
736 * Race happened, the unrelated subqueue was
737 * picked during the packet scheduler transition.
739 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
743 if (hif->hif_packets == 0) {
744 /* no packet in the tree */
749 cur_time = read_machclk();
751 if (op == ALTDQ_REMOVE && hif->hif_pollcache != NULL) {
752 cl = hif->hif_pollcache;
753 hif->hif_pollcache = NULL;
754 /* check if the class was scheduled by real-time criteria */
755 if (cl->cl_rsc != NULL)
756 realtime = (cl->cl_e <= cur_time);
759 * if there are eligible classes, use real-time criteria.
760 * find the class with the minimum deadline among
761 * the eligible classes.
763 if ((cl = ellist_get_mindl(hif->hif_eligible, cur_time)) != NULL) {
770 * use link-sharing criteria
771 * get the class with the minimum vt in the hierarchy
773 cl = hif->hif_rootclass;
774 while (is_a_parent_class(cl)) {
776 cl = actlist_firstfit(cl, cur_time);
780 kprintf("%d fit but none found\n",fits);
786 * update parent's cl_cvtmin.
787 * don't update if the new vt is smaller.
789 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
790 cl->cl_parent->cl_cvtmin = cl->cl_vt;
797 if (op == ALTDQ_POLL) {
798 hif->hif_pollcache = cl;
806 panic("hfsc_dequeue:");
808 cl->cl_hif->hif_packets--;
810 PKTCNTR_ADD(&cl->cl_stats.xmit_cnt, len);
812 update_vf(cl, len, cur_time);
816 if (!qempty(cl->cl_q)) {
817 if (cl->cl_rsc != NULL) {
819 next_len = m_pktlen(qhead(cl->cl_q));
822 update_ed(cl, next_len);
824 update_d(cl, next_len);
827 /* the class becomes passive */
832 KKASSERT(mpolled == NULL || m == mpolled);
837 hfsc_addq(struct hfsc_class *cl, struct mbuf *m)
841 if (q_is_rio(cl->cl_q))
842 return rio_addq((rio_t *)cl->cl_red, cl->cl_q,
846 if (q_is_red(cl->cl_q))
847 return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
849 if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
854 if (cl->cl_flags & HFCF_CLEARDSCP)
855 write_dsfield(m, cl->cl_pktattr, 0);
863 hfsc_getq(struct hfsc_class *cl)
866 if (q_is_rio(cl->cl_q))
867 return rio_getq((rio_t *)cl->cl_red, cl->cl_q);
870 if (q_is_red(cl->cl_q))
871 return red_getq(cl->cl_red, cl->cl_q);
873 return _getq(cl->cl_q);
877 hfsc_pollq(struct hfsc_class *cl)
879 return qhead(cl->cl_q);
883 hfsc_purgeq(struct hfsc_class *cl)
887 if (qempty(cl->cl_q))
890 while ((m = _getq(cl->cl_q)) != NULL) {
891 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, m_pktlen(m));
893 cl->cl_hif->hif_packets--;
894 cl->cl_hif->hif_ifq->altq_subq[HFSC_SUBQ_INDEX].ifq_len--;
896 KKASSERT(qlen(cl->cl_q) == 0);
898 update_vf(cl, 0, 0); /* remove cl from the actlist */
903 set_active(struct hfsc_class *cl, int len)
905 if (cl->cl_rsc != NULL)
907 if (cl->cl_fsc != NULL)
910 cl->cl_stats.period++;
914 set_passive(struct hfsc_class *cl)
916 if (cl->cl_rsc != NULL)
920 * actlist is now handled in update_vf() so that update_vf(cl, 0, 0)
921 * needs to be called explicitly to remove a class from actlist
926 init_ed(struct hfsc_class *cl, int next_len)
930 cur_time = read_machclk();
932 /* update the deadline curve */
933 rtsc_min(&cl->cl_deadline, cl->cl_rsc, cur_time, cl->cl_cumul);
936 * update the eligible curve.
937 * for concave, it is equal to the deadline curve.
938 * for convex, it is a linear curve with slope m2.
940 cl->cl_eligible = cl->cl_deadline;
941 if (cl->cl_rsc->sm1 <= cl->cl_rsc->sm2) {
942 cl->cl_eligible.dx = 0;
943 cl->cl_eligible.dy = 0;
946 /* compute e and d */
947 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
948 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
954 update_ed(struct hfsc_class *cl, int next_len)
956 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
957 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
963 update_d(struct hfsc_class *cl, int next_len)
965 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
969 init_vf(struct hfsc_class *cl, int len)
971 struct hfsc_class *max_cl, *p;
972 uint64_t vt, f, cur_time;
977 for ( ; cl->cl_parent != NULL; cl = cl->cl_parent) {
978 if (go_active && cl->cl_nactive++ == 0)
984 max_cl = actlist_last(cl->cl_parent->cl_actc);
985 if (max_cl != NULL) {
987 * set vt to the average of the min and max
988 * classes. if the parent's period didn't
989 * change, don't decrease vt of the class.
992 if (cl->cl_parent->cl_cvtmin != 0)
993 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
995 if (cl->cl_parent->cl_vtperiod !=
996 cl->cl_parentperiod || vt > cl->cl_vt)
1000 * first child for a new parent backlog period.
1001 * add parent's cvtmax to vtoff of children
1002 * to make a new vt (vtoff + vt) larger than
1003 * the vt in the last period for all children.
1005 vt = cl->cl_parent->cl_cvtmax;
1006 for (p = cl->cl_parent->cl_children; p != NULL;
1010 cl->cl_parent->cl_cvtmax = 0;
1011 cl->cl_parent->cl_cvtmin = 0;
1013 cl->cl_initvt = cl->cl_vt;
1015 /* update the virtual curve */
1016 vt = cl->cl_vt + cl->cl_vtoff;
1017 rtsc_min(&cl->cl_virtual, cl->cl_fsc, vt, cl->cl_total);
1018 if (cl->cl_virtual.x == vt) {
1019 cl->cl_virtual.x -= cl->cl_vtoff;
1024 cl->cl_vtperiod++; /* increment vt period */
1025 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
1026 if (cl->cl_parent->cl_nactive == 0)
1027 cl->cl_parentperiod++;
1032 if (cl->cl_usc != NULL) {
1033 /* class has upper limit curve */
1035 cur_time = read_machclk();
1037 /* update the ulimit curve */
1038 rtsc_min(&cl->cl_ulimit, cl->cl_usc, cur_time,
1041 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
1047 if (cl->cl_myf > cl->cl_cfmin)
1051 if (f != cl->cl_f) {
1053 update_cfmin(cl->cl_parent);
1059 update_vf(struct hfsc_class *cl, int len, uint64_t cur_time)
1061 uint64_t f, myf_bound, delta;
1064 go_passive = qempty(cl->cl_q);
1066 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
1067 cl->cl_total += len;
1069 if (cl->cl_fsc == NULL || cl->cl_nactive == 0)
1072 if (go_passive && --cl->cl_nactive == 0)
1078 /* no more active child, going passive */
1080 /* update cvtmax of the parent class */
1081 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
1082 cl->cl_parent->cl_cvtmax = cl->cl_vt;
1084 /* remove this class from the vt list */
1087 update_cfmin(cl->cl_parent);
1095 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
1096 - cl->cl_vtoff + cl->cl_vtadj;
1099 * if vt of the class is smaller than cvtmin,
1100 * the class was skipped in the past due to non-fit.
1101 * if so, we need to adjust vtadj.
1103 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
1104 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
1105 cl->cl_vt = cl->cl_parent->cl_cvtmin;
1108 /* update the vt list */
1111 if (cl->cl_usc != NULL) {
1112 cl->cl_myf = cl->cl_myfadj
1113 + rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
1116 * if myf lags behind by more than one clock tick
1117 * from the current time, adjust myfadj to prevent
1118 * a rate-limited class from going greedy.
1119 * in a steady state under rate-limiting, myf
1120 * fluctuates within one clock tick.
1122 myf_bound = cur_time - machclk_per_tick;
1123 if (cl->cl_myf < myf_bound) {
1124 delta = cur_time - cl->cl_myf;
1125 cl->cl_myfadj += delta;
1126 cl->cl_myf += delta;
1130 /* cl_f is max(cl_myf, cl_cfmin) */
1131 if (cl->cl_myf > cl->cl_cfmin)
1135 if (f != cl->cl_f) {
1137 update_cfmin(cl->cl_parent);
1143 update_cfmin(struct hfsc_class *cl)
1145 struct hfsc_class *p;
1148 if (TAILQ_EMPTY(cl->cl_actc)) {
1152 cfmin = HT_INFINITY;
1153 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1158 if (p->cl_f < cfmin)
1161 cl->cl_cfmin = cfmin;
1165 * TAILQ based ellist and actlist implementation
1166 * (ion wanted to make a calendar queue based implementation)
1169 * eligible list holds backlogged classes being sorted by their eligible times.
1170 * there is one eligible list per interface.
1178 head = kmalloc(sizeof(ellist_t *), M_ALTQ, M_WAITOK);
1184 ellist_destroy(ellist_t *head)
1186 kfree(head, M_ALTQ);
1190 ellist_insert(struct hfsc_class *cl)
1192 struct hfsc_if *hif = cl->cl_hif;
1193 struct hfsc_class *p;
1195 /* check the last entry first */
1196 if ((p = TAILQ_LAST(hif->hif_eligible, _eligible)) == NULL ||
1197 p->cl_e <= cl->cl_e) {
1198 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1202 TAILQ_FOREACH(p, hif->hif_eligible, cl_ellist) {
1203 if (cl->cl_e < p->cl_e) {
1204 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1208 KKASSERT(0); /* should not reach here */
1212 ellist_remove(struct hfsc_class *cl)
1214 struct hfsc_if *hif = cl->cl_hif;
1216 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1220 ellist_update(struct hfsc_class *cl)
1222 struct hfsc_if *hif = cl->cl_hif;
1223 struct hfsc_class *p, *last;
1226 * the eligible time of a class increases monotonically.
1227 * if the next entry has a larger eligible time, nothing to do.
1229 p = TAILQ_NEXT(cl, cl_ellist);
1230 if (p == NULL || cl->cl_e <= p->cl_e)
1233 /* check the last entry */
1234 last = TAILQ_LAST(hif->hif_eligible, _eligible);
1235 KKASSERT(last != NULL);
1236 if (last->cl_e <= cl->cl_e) {
1237 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1238 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1243 * the new position must be between the next entry
1244 * and the last entry
1246 while ((p = TAILQ_NEXT(p, cl_ellist)) != NULL) {
1247 if (cl->cl_e < p->cl_e) {
1248 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1249 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1253 KKASSERT(0); /* should not reach here */
1256 /* find the class with the minimum deadline among the eligible classes */
1258 ellist_get_mindl(ellist_t *head, uint64_t cur_time)
1260 struct hfsc_class *p, *cl = NULL;
1262 TAILQ_FOREACH(p, head, cl_ellist) {
1263 if (p->cl_e > cur_time)
1265 if (cl == NULL || p->cl_d < cl->cl_d)
1272 * active children list holds backlogged child classes being sorted
1273 * by their virtual time.
1274 * each intermediate class has one active children list.
1281 head = kmalloc(sizeof(*head), M_ALTQ, M_WAITOK);
1287 actlist_destroy(actlist_t *head)
1289 kfree(head, M_ALTQ);
1292 actlist_insert(struct hfsc_class *cl)
1294 struct hfsc_class *p;
1296 /* check the last entry first */
1297 if ((p = TAILQ_LAST(cl->cl_parent->cl_actc, _active)) == NULL
1298 || p->cl_vt <= cl->cl_vt) {
1299 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1303 TAILQ_FOREACH(p, cl->cl_parent->cl_actc, cl_actlist) {
1304 if (cl->cl_vt < p->cl_vt) {
1305 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1309 KKASSERT(0); /* should not reach here */
1313 actlist_remove(struct hfsc_class *cl)
1315 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1319 actlist_update(struct hfsc_class *cl)
1321 struct hfsc_class *p, *last;
1324 * the virtual time of a class increases monotonically during its
1325 * backlogged period.
1326 * if the next entry has a larger virtual time, nothing to do.
1328 p = TAILQ_NEXT(cl, cl_actlist);
1329 if (p == NULL || cl->cl_vt < p->cl_vt)
1332 /* check the last entry */
1333 last = TAILQ_LAST(cl->cl_parent->cl_actc, _active);
1334 KKASSERT(last != NULL);
1335 if (last->cl_vt <= cl->cl_vt) {
1336 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1337 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1342 * the new position must be between the next entry
1343 * and the last entry
1345 while ((p = TAILQ_NEXT(p, cl_actlist)) != NULL) {
1346 if (cl->cl_vt < p->cl_vt) {
1347 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1348 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1352 KKASSERT(0); /* should not reach here */
1355 static struct hfsc_class *
1356 actlist_firstfit(struct hfsc_class *cl, uint64_t cur_time)
1358 struct hfsc_class *p;
1360 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1361 if (p->cl_f <= cur_time)
1368 * service curve support functions
1370 * external service curve parameters
1373 * internal service curve parameters
1374 * sm: (bytes/tsc_interval) << SM_SHIFT
1375 * ism: (tsc_count/byte) << ISM_SHIFT
1378 * SM_SHIFT and ISM_SHIFT are scaled in order to keep effective digits.
1379 * we should be able to handle 100K-1Gbps linkspeed with 200Hz-1GHz CPU
1380 * speed. SM_SHIFT and ISM_SHIFT are selected to have at least 3 effective
1381 * digits in decimal using the following table.
1383 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
1384 * ----------+-------------------------------------------------------
1385 * bytes/nsec 12.5e-6 125e-6 1250e-6 12500e-6 125000e-6
1386 * sm(500MHz) 25.0e-6 250e-6 2500e-6 25000e-6 250000e-6
1387 * sm(200MHz) 62.5e-6 625e-6 6250e-6 62500e-6 625000e-6
1389 * nsec/byte 80000 8000 800 80 8
1390 * ism(500MHz) 40000 4000 400 40 4
1391 * ism(200MHz) 16000 1600 160 16 1.6
1394 #define ISM_SHIFT 10
1396 #define SM_MASK ((1LL << SM_SHIFT) - 1)
1397 #define ISM_MASK ((1LL << ISM_SHIFT) - 1)
1399 static __inline uint64_t
1400 seg_x2y(uint64_t x, uint64_t sm)
1406 * y = x * sm >> SM_SHIFT
1407 * but divide it for the upper and lower bits to avoid overflow
1409 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
1413 static __inline uint64_t
1414 seg_y2x(uint64_t y, uint64_t ism)
1420 else if (ism == HT_INFINITY)
1423 x = (y >> ISM_SHIFT) * ism + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
1428 static __inline uint64_t
1433 sm = ((uint64_t)m << SM_SHIFT) / 8 / machclk_freq;
1437 static __inline uint64_t
1445 ism = ((uint64_t)machclk_freq << ISM_SHIFT) * 8 / m;
1449 static __inline uint64_t
1454 dx = ((uint64_t)d * machclk_freq) / 1000;
1463 m = (sm * 8 * machclk_freq) >> SM_SHIFT;
1472 d = dx * 1000 / machclk_freq;
1477 sc2isc(struct service_curve *sc, struct internal_sc *isc)
1479 isc->sm1 = m2sm(sc->m1);
1480 isc->ism1 = m2ism(sc->m1);
1481 isc->dx = d2dx(sc->d);
1482 isc->dy = seg_x2y(isc->dx, isc->sm1);
1483 isc->sm2 = m2sm(sc->m2);
1484 isc->ism2 = m2ism(sc->m2);
1488 * initialize the runtime service curve with the given internal
1489 * service curve starting at (x, y).
1492 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1496 rtsc->sm1 = isc->sm1;
1497 rtsc->ism1 = isc->ism1;
1500 rtsc->sm2 = isc->sm2;
1501 rtsc->ism2 = isc->ism2;
1505 * calculate the y-projection of the runtime service curve by the
1506 * given x-projection value
1509 rtsc_y2x(struct runtime_sc *rtsc, uint64_t y)
1515 } else if (y <= rtsc->y + rtsc->dy) {
1516 /* x belongs to the 1st segment */
1518 x = rtsc->x + rtsc->dx;
1520 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
1522 /* x belongs to the 2nd segment */
1523 x = rtsc->x + rtsc->dx
1524 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
1530 rtsc_x2y(struct runtime_sc *rtsc, uint64_t x)
1536 } else if (x <= rtsc->x + rtsc->dx) {
1537 /* y belongs to the 1st segment */
1538 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
1540 /* y belongs to the 2nd segment */
1541 y = rtsc->y + rtsc->dy
1542 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
1547 * update the runtime service curve by taking the minimum of the current
1548 * runtime service curve and the service curve starting at (x, y).
1551 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1553 uint64_t y1, y2, dx, dy;
1555 if (isc->sm1 <= isc->sm2) {
1556 /* service curve is convex */
1557 y1 = rtsc_x2y(rtsc, x);
1559 /* the current rtsc is smaller */
1567 * service curve is concave
1568 * compute the two y values of the current rtsc
1572 y1 = rtsc_x2y(rtsc, x);
1574 /* rtsc is below isc, no change to rtsc */
1578 y2 = rtsc_x2y(rtsc, x + isc->dx);
1579 if (y2 >= y + isc->dy) {
1580 /* rtsc is above isc, replace rtsc by isc */
1589 * the two curves intersect
1590 * compute the offsets (dx, dy) using the reverse
1591 * function of seg_x2y()
1592 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
1594 dx = ((y1 - y) << SM_SHIFT) / (isc->sm1 - isc->sm2);
1596 * check if (x, y1) belongs to the 1st segment of rtsc.
1597 * if so, add the offset.
1599 if (rtsc->x + rtsc->dx > x)
1600 dx += rtsc->x + rtsc->dx - x;
1601 dy = seg_x2y(dx, isc->sm1);
1610 get_class_stats(struct hfsc_classstats *sp, struct hfsc_class *cl)
1612 sp->class_id = cl->cl_id;
1613 sp->class_handle = cl->cl_handle;
1615 if (cl->cl_rsc != NULL) {
1616 sp->rsc.m1 = sm2m(cl->cl_rsc->sm1);
1617 sp->rsc.d = dx2d(cl->cl_rsc->dx);
1618 sp->rsc.m2 = sm2m(cl->cl_rsc->sm2);
1624 if (cl->cl_fsc != NULL) {
1625 sp->fsc.m1 = sm2m(cl->cl_fsc->sm1);
1626 sp->fsc.d = dx2d(cl->cl_fsc->dx);
1627 sp->fsc.m2 = sm2m(cl->cl_fsc->sm2);
1633 if (cl->cl_usc != NULL) {
1634 sp->usc.m1 = sm2m(cl->cl_usc->sm1);
1635 sp->usc.d = dx2d(cl->cl_usc->dx);
1636 sp->usc.m2 = sm2m(cl->cl_usc->sm2);
1643 sp->total = cl->cl_total;
1644 sp->cumul = cl->cl_cumul;
1651 sp->initvt = cl->cl_initvt;
1652 sp->vtperiod = cl->cl_vtperiod;
1653 sp->parentperiod = cl->cl_parentperiod;
1654 sp->nactive = cl->cl_nactive;
1655 sp->vtoff = cl->cl_vtoff;
1656 sp->cvtmax = cl->cl_cvtmax;
1657 sp->myf = cl->cl_myf;
1658 sp->cfmin = cl->cl_cfmin;
1659 sp->cvtmin = cl->cl_cvtmin;
1660 sp->myfadj = cl->cl_myfadj;
1661 sp->vtadj = cl->cl_vtadj;
1663 sp->cur_time = read_machclk();
1664 sp->machclk_freq = machclk_freq;
1666 sp->qlength = qlen(cl->cl_q);
1667 sp->qlimit = qlimit(cl->cl_q);
1668 sp->xmit_cnt = cl->cl_stats.xmit_cnt;
1669 sp->drop_cnt = cl->cl_stats.drop_cnt;
1670 sp->period = cl->cl_stats.period;
1672 sp->qtype = qtype(cl->cl_q);
1674 if (q_is_red(cl->cl_q))
1675 red_getstats(cl->cl_red, &sp->red[0]);
1678 if (q_is_rio(cl->cl_q))
1679 rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
1683 /* convert a class handle to the corresponding class pointer */
1684 static struct hfsc_class *
1685 clh_to_clp(struct hfsc_if *hif, uint32_t chandle)
1688 struct hfsc_class *cl;
1693 * first, try optimistically the slot matching the lower bits of
1694 * the handle. if it fails, do the linear table search.
1696 i = chandle % HFSC_MAX_CLASSES;
1697 if ((cl = hif->hif_class_tbl[i]) != NULL && cl->cl_handle == chandle)
1699 for (i = 0; i < HFSC_MAX_CLASSES; i++)
1700 if ((cl = hif->hif_class_tbl[i]) != NULL &&
1701 cl->cl_handle == chandle)
1706 #endif /* ALTQ_HFSC */