2 * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version 2
7 * of the License, or (at your option) any later version.
9 * 2003-10-17 - Ported from altq
12 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
14 * Permission to use, copy, modify, and distribute this software and
15 * its documentation is hereby granted (including for commercial or
16 * for-profit use), provided that both the copyright notice and this
17 * permission notice appear in all copies of the software, derivative
18 * works, or modified versions, and any portions thereof.
20 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
21 * WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON PROVIDES THIS
22 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
23 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
25 * DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
28 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
29 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
30 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
32 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
35 * Carnegie Mellon encourages (but does not require) users of this
36 * software to return any improvements or extensions that they make,
37 * and to grant Carnegie Mellon the rights to redistribute these
38 * changes without encumbrance.
41 * H-FSC is described in Proceedings of SIGCOMM'97,
42 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
43 * Real-Time and Priority Service"
44 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
46 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
47 * when a class has an upperlimit, the fit-time is computed from the
48 * upperlimit service curve. the link-sharing scheduler does not schedule
49 * a class whose fit-time exceeds the current time.
52 #include <linux/kernel.h>
53 #include <linux/module.h>
54 #include <linux/types.h>
55 #include <linux/errno.h>
56 #include <linux/compiler.h>
57 #include <linux/spinlock.h>
58 #include <linux/skbuff.h>
59 #include <linux/string.h>
60 #include <linux/slab.h>
61 #include <linux/list.h>
62 #include <linux/rbtree.h>
63 #include <linux/init.h>
64 #include <linux/rtnetlink.h>
65 #include <linux/pkt_sched.h>
66 #include <net/netlink.h>
67 #include <net/pkt_sched.h>
68 #include <net/pkt_cls.h>
69 #include <asm/div64.h>
72 * kernel internal service curve representation:
73 * coordinates are given by 64 bit unsigned integers.
74 * x-axis: unit is clock count.
75 * y-axis: unit is byte.
77 * The service curve parameters are converted to the internal
78 * representation. The slope values are scaled to avoid overflow.
79 * the inverse slope values as well as the y-projection of the 1st
80 * segment are kept in order to avoid 64-bit divide operations
81 * that are expensive on 32-bit architectures.
85 u64 sm1; /* scaled slope of the 1st segment */
86 u64 ism1; /* scaled inverse-slope of the 1st segment */
87 u64 dx; /* the x-projection of the 1st segment */
88 u64 dy; /* the y-projection of the 1st segment */
89 u64 sm2; /* scaled slope of the 2nd segment */
90 u64 ism2; /* scaled inverse-slope of the 2nd segment */
93 /* runtime service curve */
95 u64 x; /* current starting position on x-axis */
96 u64 y; /* current starting position on y-axis */
97 u64 sm1; /* scaled slope of the 1st segment */
98 u64 ism1; /* scaled inverse-slope of the 1st segment */
99 u64 dx; /* the x-projection of the 1st segment */
100 u64 dy; /* the y-projection of the 1st segment */
101 u64 sm2; /* scaled slope of the 2nd segment */
102 u64 ism2; /* scaled inverse-slope of the 2nd segment */
105 enum hfsc_class_flags {
112 struct Qdisc_class_common cl_common;
113 unsigned int refcnt; /* usage count */
115 struct gnet_stats_basic_packed bstats;
116 struct gnet_stats_queue qstats;
117 struct gnet_stats_rate_est rate_est;
118 unsigned int level; /* class level in hierarchy */
119 struct tcf_proto *filter_list; /* filter list */
120 unsigned int filter_cnt; /* filter count */
122 struct hfsc_sched *sched; /* scheduler data */
123 struct hfsc_class *cl_parent; /* parent class */
124 struct list_head siblings; /* sibling classes */
125 struct list_head children; /* child classes */
126 struct Qdisc *qdisc; /* leaf qdisc */
128 struct rb_node el_node; /* qdisc's eligible tree member */
129 struct rb_root vt_tree; /* active children sorted by cl_vt */
130 struct rb_node vt_node; /* parent's vt_tree member */
131 struct rb_root cf_tree; /* active children sorted by cl_f */
132 struct rb_node cf_node; /* parent's cf_heap member */
133 struct list_head dlist; /* drop list member */
135 u64 cl_total; /* total work in bytes */
136 u64 cl_cumul; /* cumulative work in bytes done by
137 real-time criteria */
139 u64 cl_d; /* deadline*/
140 u64 cl_e; /* eligible time */
141 u64 cl_vt; /* virtual time */
142 u64 cl_f; /* time when this class will fit for
143 link-sharing, max(myf, cfmin) */
144 u64 cl_myf; /* my fit-time (calculated from this
145 class's own upperlimit curve) */
146 u64 cl_myfadj; /* my fit-time adjustment (to cancel
147 history dependence) */
148 u64 cl_cfmin; /* earliest children's fit-time (used
149 with cl_myf to obtain cl_f) */
150 u64 cl_cvtmin; /* minimal virtual time among the
151 children fit for link-sharing
152 (monotonic within a period) */
153 u64 cl_vtadj; /* intra-period cumulative vt
155 u64 cl_vtoff; /* inter-period cumulative vt offset */
156 u64 cl_cvtmax; /* max child's vt in the last period */
157 u64 cl_cvtoff; /* cumulative cvtmax of all periods */
158 u64 cl_pcvtoff; /* parent's cvtoff at initialization
161 struct internal_sc cl_rsc; /* internal real-time service curve */
162 struct internal_sc cl_fsc; /* internal fair service curve */
163 struct internal_sc cl_usc; /* internal upperlimit service curve */
164 struct runtime_sc cl_deadline; /* deadline curve */
165 struct runtime_sc cl_eligible; /* eligible curve */
166 struct runtime_sc cl_virtual; /* virtual curve */
167 struct runtime_sc cl_ulimit; /* upperlimit curve */
169 unsigned long cl_flags; /* which curves are valid */
170 unsigned long cl_vtperiod; /* vt period sequence number */
171 unsigned long cl_parentperiod;/* parent's vt period sequence number*/
172 unsigned long cl_nactive; /* number of active children */
176 u16 defcls; /* default class id */
177 struct hfsc_class root; /* root class */
178 struct Qdisc_class_hash clhash; /* class hash */
179 struct rb_root eligible; /* eligible tree */
180 struct list_head droplist; /* active leaf class list (for
182 struct qdisc_watchdog watchdog; /* watchdog timer */
185 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
189 * eligible tree holds backlogged classes being sorted by their eligible times.
190 * there is one eligible tree per hfsc instance.
194 eltree_insert(struct hfsc_class *cl)
196 struct rb_node **p = &cl->sched->eligible.rb_node;
197 struct rb_node *parent = NULL;
198 struct hfsc_class *cl1;
202 cl1 = rb_entry(parent, struct hfsc_class, el_node);
203 if (cl->cl_e >= cl1->cl_e)
204 p = &parent->rb_right;
206 p = &parent->rb_left;
208 rb_link_node(&cl->el_node, parent, p);
209 rb_insert_color(&cl->el_node, &cl->sched->eligible);
213 eltree_remove(struct hfsc_class *cl)
215 rb_erase(&cl->el_node, &cl->sched->eligible);
219 eltree_update(struct hfsc_class *cl)
225 /* find the class with the minimum deadline among the eligible classes */
226 static inline struct hfsc_class *
227 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
229 struct hfsc_class *p, *cl = NULL;
232 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
233 p = rb_entry(n, struct hfsc_class, el_node);
234 if (p->cl_e > cur_time)
236 if (cl == NULL || p->cl_d < cl->cl_d)
242 /* find the class with minimum eligible time among the eligible classes */
243 static inline struct hfsc_class *
244 eltree_get_minel(struct hfsc_sched *q)
248 n = rb_first(&q->eligible);
251 return rb_entry(n, struct hfsc_class, el_node);
255 * vttree holds holds backlogged child classes being sorted by their virtual
256 * time. each intermediate class has one vttree.
259 vttree_insert(struct hfsc_class *cl)
261 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
262 struct rb_node *parent = NULL;
263 struct hfsc_class *cl1;
267 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
268 if (cl->cl_vt >= cl1->cl_vt)
269 p = &parent->rb_right;
271 p = &parent->rb_left;
273 rb_link_node(&cl->vt_node, parent, p);
274 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
278 vttree_remove(struct hfsc_class *cl)
280 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
284 vttree_update(struct hfsc_class *cl)
290 static inline struct hfsc_class *
291 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
293 struct hfsc_class *p;
296 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
297 p = rb_entry(n, struct hfsc_class, vt_node);
298 if (p->cl_f <= cur_time)
305 * get the leaf class with the minimum vt in the hierarchy
307 static struct hfsc_class *
308 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
310 /* if root-class's cfmin is bigger than cur_time nothing to do */
311 if (cl->cl_cfmin > cur_time)
314 while (cl->level > 0) {
315 cl = vttree_firstfit(cl, cur_time);
319 * update parent's cl_cvtmin.
321 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
322 cl->cl_parent->cl_cvtmin = cl->cl_vt;
328 cftree_insert(struct hfsc_class *cl)
330 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
331 struct rb_node *parent = NULL;
332 struct hfsc_class *cl1;
336 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
337 if (cl->cl_f >= cl1->cl_f)
338 p = &parent->rb_right;
340 p = &parent->rb_left;
342 rb_link_node(&cl->cf_node, parent, p);
343 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
347 cftree_remove(struct hfsc_class *cl)
349 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
353 cftree_update(struct hfsc_class *cl)
360 * service curve support functions
362 * external service curve parameters
365 * internal service curve parameters
366 * sm: (bytes/psched_us) << SM_SHIFT
367 * ism: (psched_us/byte) << ISM_SHIFT
370 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
372 * sm and ism are scaled in order to keep effective digits.
373 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
374 * digits in decimal using the following table.
376 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
377 * ------------+-------------------------------------------------------
378 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
380 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
382 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
384 #define SM_SHIFT (30 - PSCHED_SHIFT)
385 #define ISM_SHIFT (8 + PSCHED_SHIFT)
387 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
388 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
391 seg_x2y(u64 x, u64 sm)
397 * y = x * sm >> SM_SHIFT
398 * but divide it for the upper and lower bits to avoid overflow
400 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
405 seg_y2x(u64 y, u64 ism)
411 else if (ism == HT_INFINITY)
414 x = (y >> ISM_SHIFT) * ism
415 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
420 /* Convert m (bps) into sm (bytes/psched us) */
426 sm = ((u64)m << SM_SHIFT);
427 sm += PSCHED_TICKS_PER_SEC - 1;
428 do_div(sm, PSCHED_TICKS_PER_SEC);
432 /* convert m (bps) into ism (psched us/byte) */
441 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
448 /* convert d (us) into dx (psched us) */
454 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
455 dx += USEC_PER_SEC - 1;
456 do_div(dx, USEC_PER_SEC);
460 /* convert sm (bytes/psched us) into m (bps) */
466 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
470 /* convert dx (psched us) into d (us) */
476 d = dx * USEC_PER_SEC;
477 do_div(d, PSCHED_TICKS_PER_SEC);
482 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
484 isc->sm1 = m2sm(sc->m1);
485 isc->ism1 = m2ism(sc->m1);
486 isc->dx = d2dx(sc->d);
487 isc->dy = seg_x2y(isc->dx, isc->sm1);
488 isc->sm2 = m2sm(sc->m2);
489 isc->ism2 = m2ism(sc->m2);
493 * initialize the runtime service curve with the given internal
494 * service curve starting at (x, y).
497 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
501 rtsc->sm1 = isc->sm1;
502 rtsc->ism1 = isc->ism1;
505 rtsc->sm2 = isc->sm2;
506 rtsc->ism2 = isc->ism2;
510 * calculate the y-projection of the runtime service curve by the
511 * given x-projection value
514 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
520 else if (y <= rtsc->y + rtsc->dy) {
521 /* x belongs to the 1st segment */
523 x = rtsc->x + rtsc->dx;
525 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
527 /* x belongs to the 2nd segment */
528 x = rtsc->x + rtsc->dx
529 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
535 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
541 else if (x <= rtsc->x + rtsc->dx)
542 /* y belongs to the 1st segment */
543 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
545 /* y belongs to the 2nd segment */
546 y = rtsc->y + rtsc->dy
547 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
552 * update the runtime service curve by taking the minimum of the current
553 * runtime service curve and the service curve starting at (x, y).
556 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
561 if (isc->sm1 <= isc->sm2) {
562 /* service curve is convex */
563 y1 = rtsc_x2y(rtsc, x);
565 /* the current rtsc is smaller */
573 * service curve is concave
574 * compute the two y values of the current rtsc
578 y1 = rtsc_x2y(rtsc, x);
580 /* rtsc is below isc, no change to rtsc */
584 y2 = rtsc_x2y(rtsc, x + isc->dx);
585 if (y2 >= y + isc->dy) {
586 /* rtsc is above isc, replace rtsc by isc */
595 * the two curves intersect
596 * compute the offsets (dx, dy) using the reverse
597 * function of seg_x2y()
598 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
600 dx = (y1 - y) << SM_SHIFT;
601 dsm = isc->sm1 - isc->sm2;
604 * check if (x, y1) belongs to the 1st segment of rtsc.
605 * if so, add the offset.
607 if (rtsc->x + rtsc->dx > x)
608 dx += rtsc->x + rtsc->dx - x;
609 dy = seg_x2y(dx, isc->sm1);
618 init_ed(struct hfsc_class *cl, unsigned int next_len)
620 u64 cur_time = psched_get_time();
622 /* update the deadline curve */
623 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
626 * update the eligible curve.
627 * for concave, it is equal to the deadline curve.
628 * for convex, it is a linear curve with slope m2.
630 cl->cl_eligible = cl->cl_deadline;
631 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
632 cl->cl_eligible.dx = 0;
633 cl->cl_eligible.dy = 0;
636 /* compute e and d */
637 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
638 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
644 update_ed(struct hfsc_class *cl, unsigned int next_len)
646 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
647 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
653 update_d(struct hfsc_class *cl, unsigned int next_len)
655 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
659 update_cfmin(struct hfsc_class *cl)
661 struct rb_node *n = rb_first(&cl->cf_tree);
662 struct hfsc_class *p;
668 p = rb_entry(n, struct hfsc_class, cf_node);
669 cl->cl_cfmin = p->cl_f;
673 init_vf(struct hfsc_class *cl, unsigned int len)
675 struct hfsc_class *max_cl;
682 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
683 if (go_active && cl->cl_nactive++ == 0)
689 n = rb_last(&cl->cl_parent->vt_tree);
691 max_cl = rb_entry(n, struct hfsc_class, vt_node);
693 * set vt to the average of the min and max
694 * classes. if the parent's period didn't
695 * change, don't decrease vt of the class.
698 if (cl->cl_parent->cl_cvtmin != 0)
699 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
701 if (cl->cl_parent->cl_vtperiod !=
702 cl->cl_parentperiod || vt > cl->cl_vt)
706 * first child for a new parent backlog period.
707 * add parent's cvtmax to cvtoff to make a new
708 * vt (vtoff + vt) larger than the vt in the
709 * last period for all children.
711 vt = cl->cl_parent->cl_cvtmax;
712 cl->cl_parent->cl_cvtoff += vt;
713 cl->cl_parent->cl_cvtmax = 0;
714 cl->cl_parent->cl_cvtmin = 0;
718 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
721 /* update the virtual curve */
722 vt = cl->cl_vt + cl->cl_vtoff;
723 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
725 if (cl->cl_virtual.x == vt) {
726 cl->cl_virtual.x -= cl->cl_vtoff;
731 cl->cl_vtperiod++; /* increment vt period */
732 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
733 if (cl->cl_parent->cl_nactive == 0)
734 cl->cl_parentperiod++;
740 if (cl->cl_flags & HFSC_USC) {
741 /* class has upper limit curve */
743 cur_time = psched_get_time();
745 /* update the ulimit curve */
746 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
749 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
755 f = max(cl->cl_myf, cl->cl_cfmin);
760 update_cfmin(cl->cl_parent);
765 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
767 u64 f; /* , myf_bound, delta; */
770 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
773 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
776 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
779 if (go_passive && --cl->cl_nactive == 0)
785 /* no more active child, going passive */
787 /* update cvtmax of the parent class */
788 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
789 cl->cl_parent->cl_cvtmax = cl->cl_vt;
791 /* remove this class from the vt tree */
795 update_cfmin(cl->cl_parent);
803 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
804 - cl->cl_vtoff + cl->cl_vtadj;
807 * if vt of the class is smaller than cvtmin,
808 * the class was skipped in the past due to non-fit.
809 * if so, we need to adjust vtadj.
811 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
812 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
813 cl->cl_vt = cl->cl_parent->cl_cvtmin;
816 /* update the vt tree */
819 if (cl->cl_flags & HFSC_USC) {
820 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
824 * This code causes classes to stay way under their
825 * limit when multiple classes are used at gigabit
826 * speed. needs investigation. -kaber
829 * if myf lags behind by more than one clock tick
830 * from the current time, adjust myfadj to prevent
831 * a rate-limited class from going greedy.
832 * in a steady state under rate-limiting, myf
833 * fluctuates within one clock tick.
835 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
836 if (cl->cl_myf < myf_bound) {
837 delta = cur_time - cl->cl_myf;
838 cl->cl_myfadj += delta;
844 f = max(cl->cl_myf, cl->cl_cfmin);
848 update_cfmin(cl->cl_parent);
854 set_active(struct hfsc_class *cl, unsigned int len)
856 if (cl->cl_flags & HFSC_RSC)
858 if (cl->cl_flags & HFSC_FSC)
861 list_add_tail(&cl->dlist, &cl->sched->droplist);
865 set_passive(struct hfsc_class *cl)
867 if (cl->cl_flags & HFSC_RSC)
870 list_del(&cl->dlist);
873 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
874 * needs to be called explicitly to remove a class from vttree.
879 qdisc_peek_len(struct Qdisc *sch)
884 skb = sch->ops->peek(sch);
886 qdisc_warn_nonwc("qdisc_peek_len", sch);
889 len = qdisc_pkt_len(skb);
895 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
897 unsigned int len = cl->qdisc->q.qlen;
899 qdisc_reset(cl->qdisc);
900 qdisc_tree_decrease_qlen(cl->qdisc, len);
904 hfsc_adjust_levels(struct hfsc_class *cl)
906 struct hfsc_class *p;
911 list_for_each_entry(p, &cl->children, siblings) {
912 if (p->level >= level)
913 level = p->level + 1;
916 } while ((cl = cl->cl_parent) != NULL);
919 static inline struct hfsc_class *
920 hfsc_find_class(u32 classid, struct Qdisc *sch)
922 struct hfsc_sched *q = qdisc_priv(sch);
923 struct Qdisc_class_common *clc;
925 clc = qdisc_class_find(&q->clhash, classid);
928 return container_of(clc, struct hfsc_class, cl_common);
932 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
935 sc2isc(rsc, &cl->cl_rsc);
936 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
937 cl->cl_eligible = cl->cl_deadline;
938 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
939 cl->cl_eligible.dx = 0;
940 cl->cl_eligible.dy = 0;
942 cl->cl_flags |= HFSC_RSC;
946 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
948 sc2isc(fsc, &cl->cl_fsc);
949 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
950 cl->cl_flags |= HFSC_FSC;
954 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
957 sc2isc(usc, &cl->cl_usc);
958 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
959 cl->cl_flags |= HFSC_USC;
962 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
963 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
964 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
965 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
969 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
970 struct nlattr **tca, unsigned long *arg)
972 struct hfsc_sched *q = qdisc_priv(sch);
973 struct hfsc_class *cl = (struct hfsc_class *)*arg;
974 struct hfsc_class *parent = NULL;
975 struct nlattr *opt = tca[TCA_OPTIONS];
976 struct nlattr *tb[TCA_HFSC_MAX + 1];
977 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
984 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
988 if (tb[TCA_HFSC_RSC]) {
989 rsc = nla_data(tb[TCA_HFSC_RSC]);
990 if (rsc->m1 == 0 && rsc->m2 == 0)
994 if (tb[TCA_HFSC_FSC]) {
995 fsc = nla_data(tb[TCA_HFSC_FSC]);
996 if (fsc->m1 == 0 && fsc->m2 == 0)
1000 if (tb[TCA_HFSC_USC]) {
1001 usc = nla_data(tb[TCA_HFSC_USC]);
1002 if (usc->m1 == 0 && usc->m2 == 0)
1008 if (cl->cl_parent &&
1009 cl->cl_parent->cl_common.classid != parentid)
1011 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1014 cur_time = psched_get_time();
1016 if (tca[TCA_RATE]) {
1017 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1018 qdisc_root_sleeping_lock(sch),
1026 hfsc_change_rsc(cl, rsc, cur_time);
1028 hfsc_change_fsc(cl, fsc);
1030 hfsc_change_usc(cl, usc, cur_time);
1032 if (cl->qdisc->q.qlen != 0) {
1033 if (cl->cl_flags & HFSC_RSC)
1034 update_ed(cl, qdisc_peek_len(cl->qdisc));
1035 if (cl->cl_flags & HFSC_FSC)
1036 update_vf(cl, 0, cur_time);
1038 sch_tree_unlock(sch);
1043 if (parentid == TC_H_ROOT)
1048 parent = hfsc_find_class(parentid, sch);
1053 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1055 if (hfsc_find_class(classid, sch))
1058 if (rsc == NULL && fsc == NULL)
1061 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1065 if (tca[TCA_RATE]) {
1066 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1067 qdisc_root_sleeping_lock(sch),
1076 hfsc_change_rsc(cl, rsc, 0);
1078 hfsc_change_fsc(cl, fsc);
1080 hfsc_change_usc(cl, usc, 0);
1082 cl->cl_common.classid = classid;
1085 cl->cl_parent = parent;
1086 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
1087 &pfifo_qdisc_ops, classid);
1088 if (cl->qdisc == NULL)
1089 cl->qdisc = &noop_qdisc;
1090 INIT_LIST_HEAD(&cl->children);
1091 cl->vt_tree = RB_ROOT;
1092 cl->cf_tree = RB_ROOT;
1095 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1096 list_add_tail(&cl->siblings, &parent->children);
1097 if (parent->level == 0)
1098 hfsc_purge_queue(sch, parent);
1099 hfsc_adjust_levels(parent);
1100 cl->cl_pcvtoff = parent->cl_cvtoff;
1101 sch_tree_unlock(sch);
1103 qdisc_class_hash_grow(sch, &q->clhash);
1105 *arg = (unsigned long)cl;
1110 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1112 struct hfsc_sched *q = qdisc_priv(sch);
1114 tcf_destroy_chain(&cl->filter_list);
1115 qdisc_destroy(cl->qdisc);
1116 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1122 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1124 struct hfsc_sched *q = qdisc_priv(sch);
1125 struct hfsc_class *cl = (struct hfsc_class *)arg;
1127 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1132 list_del(&cl->siblings);
1133 hfsc_adjust_levels(cl->cl_parent);
1135 hfsc_purge_queue(sch, cl);
1136 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1138 BUG_ON(--cl->refcnt == 0);
1140 * This shouldn't happen: we "hold" one cops->get() when called
1141 * from tc_ctl_tclass; the destroy method is done from cops->put().
1144 sch_tree_unlock(sch);
1148 static struct hfsc_class *
1149 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1151 struct hfsc_sched *q = qdisc_priv(sch);
1152 struct hfsc_class *head, *cl;
1153 struct tcf_result res;
1154 struct tcf_proto *tcf;
1157 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1158 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1162 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1164 tcf = q->root.filter_list;
1165 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1166 #ifdef CONFIG_NET_CLS_ACT
1170 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1175 cl = (struct hfsc_class *)res.class;
1177 cl = hfsc_find_class(res.classid, sch);
1179 break; /* filter selected invalid classid */
1180 if (cl->level >= head->level)
1181 break; /* filter may only point downwards */
1185 return cl; /* hit leaf class */
1187 /* apply inner filter chain */
1188 tcf = cl->filter_list;
1192 /* classification failed, try default class */
1193 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1194 if (cl == NULL || cl->level > 0)
1201 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1204 struct hfsc_class *cl = (struct hfsc_class *)arg;
1209 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1210 cl->cl_common.classid);
1216 hfsc_purge_queue(sch, cl);
1219 sch_tree_unlock(sch);
1223 static struct Qdisc *
1224 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1226 struct hfsc_class *cl = (struct hfsc_class *)arg;
1235 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1237 struct hfsc_class *cl = (struct hfsc_class *)arg;
1239 if (cl->qdisc->q.qlen == 0) {
1240 update_vf(cl, 0, 0);
1245 static unsigned long
1246 hfsc_get_class(struct Qdisc *sch, u32 classid)
1248 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1253 return (unsigned long)cl;
1257 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1259 struct hfsc_class *cl = (struct hfsc_class *)arg;
1261 if (--cl->refcnt == 0)
1262 hfsc_destroy_class(sch, cl);
1265 static unsigned long
1266 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1268 struct hfsc_class *p = (struct hfsc_class *)parent;
1269 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1272 if (p != NULL && p->level <= cl->level)
1277 return (unsigned long)cl;
1281 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1283 struct hfsc_class *cl = (struct hfsc_class *)arg;
1288 static struct tcf_proto **
1289 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1291 struct hfsc_sched *q = qdisc_priv(sch);
1292 struct hfsc_class *cl = (struct hfsc_class *)arg;
1297 return &cl->filter_list;
1301 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1303 struct tc_service_curve tsc;
1305 tsc.m1 = sm2m(sc->sm1);
1306 tsc.d = dx2d(sc->dx);
1307 tsc.m2 = sm2m(sc->sm2);
1308 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1309 goto nla_put_failure;
1318 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1320 if ((cl->cl_flags & HFSC_RSC) &&
1321 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1322 goto nla_put_failure;
1324 if ((cl->cl_flags & HFSC_FSC) &&
1325 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1326 goto nla_put_failure;
1328 if ((cl->cl_flags & HFSC_USC) &&
1329 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1330 goto nla_put_failure;
1339 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1342 struct hfsc_class *cl = (struct hfsc_class *)arg;
1343 struct nlattr *nest;
1345 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1347 tcm->tcm_handle = cl->cl_common.classid;
1349 tcm->tcm_info = cl->qdisc->handle;
1351 nest = nla_nest_start(skb, TCA_OPTIONS);
1353 goto nla_put_failure;
1354 if (hfsc_dump_curves(skb, cl) < 0)
1355 goto nla_put_failure;
1356 nla_nest_end(skb, nest);
1360 nla_nest_cancel(skb, nest);
1365 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1366 struct gnet_dump *d)
1368 struct hfsc_class *cl = (struct hfsc_class *)arg;
1369 struct tc_hfsc_stats xstats;
1371 cl->qstats.qlen = cl->qdisc->q.qlen;
1372 cl->qstats.backlog = cl->qdisc->qstats.backlog;
1373 xstats.level = cl->level;
1374 xstats.period = cl->cl_vtperiod;
1375 xstats.work = cl->cl_total;
1376 xstats.rtwork = cl->cl_cumul;
1378 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1379 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
1380 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1383 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1389 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1391 struct hfsc_sched *q = qdisc_priv(sch);
1392 struct hfsc_class *cl;
1398 for (i = 0; i < q->clhash.hashsize; i++) {
1399 hlist_for_each_entry(cl, &q->clhash.hash[i],
1401 if (arg->count < arg->skip) {
1405 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1415 hfsc_schedule_watchdog(struct Qdisc *sch)
1417 struct hfsc_sched *q = qdisc_priv(sch);
1418 struct hfsc_class *cl;
1421 cl = eltree_get_minel(q);
1423 next_time = cl->cl_e;
1424 if (q->root.cl_cfmin != 0) {
1425 if (next_time == 0 || next_time > q->root.cl_cfmin)
1426 next_time = q->root.cl_cfmin;
1428 WARN_ON(next_time == 0);
1429 qdisc_watchdog_schedule(&q->watchdog, next_time);
1433 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1435 struct hfsc_sched *q = qdisc_priv(sch);
1436 struct tc_hfsc_qopt *qopt;
1439 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1441 qopt = nla_data(opt);
1443 q->defcls = qopt->defcls;
1444 err = qdisc_class_hash_init(&q->clhash);
1447 q->eligible = RB_ROOT;
1448 INIT_LIST_HEAD(&q->droplist);
1450 q->root.cl_common.classid = sch->handle;
1453 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1455 if (q->root.qdisc == NULL)
1456 q->root.qdisc = &noop_qdisc;
1457 INIT_LIST_HEAD(&q->root.children);
1458 q->root.vt_tree = RB_ROOT;
1459 q->root.cf_tree = RB_ROOT;
1461 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1462 qdisc_class_hash_grow(sch, &q->clhash);
1464 qdisc_watchdog_init(&q->watchdog, sch);
1470 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1472 struct hfsc_sched *q = qdisc_priv(sch);
1473 struct tc_hfsc_qopt *qopt;
1475 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1477 qopt = nla_data(opt);
1480 q->defcls = qopt->defcls;
1481 sch_tree_unlock(sch);
1487 hfsc_reset_class(struct hfsc_class *cl)
1500 cl->cl_vtperiod = 0;
1501 cl->cl_parentperiod = 0;
1508 cl->vt_tree = RB_ROOT;
1509 cl->cf_tree = RB_ROOT;
1510 qdisc_reset(cl->qdisc);
1512 if (cl->cl_flags & HFSC_RSC)
1513 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1514 if (cl->cl_flags & HFSC_FSC)
1515 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1516 if (cl->cl_flags & HFSC_USC)
1517 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1521 hfsc_reset_qdisc(struct Qdisc *sch)
1523 struct hfsc_sched *q = qdisc_priv(sch);
1524 struct hfsc_class *cl;
1527 for (i = 0; i < q->clhash.hashsize; i++) {
1528 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1529 hfsc_reset_class(cl);
1531 q->eligible = RB_ROOT;
1532 INIT_LIST_HEAD(&q->droplist);
1533 qdisc_watchdog_cancel(&q->watchdog);
1538 hfsc_destroy_qdisc(struct Qdisc *sch)
1540 struct hfsc_sched *q = qdisc_priv(sch);
1541 struct hlist_node *next;
1542 struct hfsc_class *cl;
1545 for (i = 0; i < q->clhash.hashsize; i++) {
1546 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1547 tcf_destroy_chain(&cl->filter_list);
1549 for (i = 0; i < q->clhash.hashsize; i++) {
1550 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1552 hfsc_destroy_class(sch, cl);
1554 qdisc_class_hash_destroy(&q->clhash);
1555 qdisc_watchdog_cancel(&q->watchdog);
1559 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1561 struct hfsc_sched *q = qdisc_priv(sch);
1562 unsigned char *b = skb_tail_pointer(skb);
1563 struct tc_hfsc_qopt qopt;
1564 struct hfsc_class *cl;
1567 sch->qstats.backlog = 0;
1568 for (i = 0; i < q->clhash.hashsize; i++) {
1569 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1570 sch->qstats.backlog += cl->qdisc->qstats.backlog;
1573 qopt.defcls = q->defcls;
1574 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1575 goto nla_put_failure;
1584 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1586 struct hfsc_class *cl;
1587 int uninitialized_var(err);
1589 cl = hfsc_classify(skb, sch, &err);
1591 if (err & __NET_XMIT_BYPASS)
1592 sch->qstats.drops++;
1597 err = qdisc_enqueue(skb, cl->qdisc);
1598 if (unlikely(err != NET_XMIT_SUCCESS)) {
1599 if (net_xmit_drop_count(err)) {
1601 sch->qstats.drops++;
1606 if (cl->qdisc->q.qlen == 1)
1607 set_active(cl, qdisc_pkt_len(skb));
1611 return NET_XMIT_SUCCESS;
1614 static struct sk_buff *
1615 hfsc_dequeue(struct Qdisc *sch)
1617 struct hfsc_sched *q = qdisc_priv(sch);
1618 struct hfsc_class *cl;
1619 struct sk_buff *skb;
1621 unsigned int next_len;
1624 if (sch->q.qlen == 0)
1627 cur_time = psched_get_time();
1630 * if there are eligible classes, use real-time criteria.
1631 * find the class with the minimum deadline among
1632 * the eligible classes.
1634 cl = eltree_get_mindl(q, cur_time);
1639 * use link-sharing criteria
1640 * get the class with the minimum vt in the hierarchy
1642 cl = vttree_get_minvt(&q->root, cur_time);
1644 sch->qstats.overlimits++;
1645 hfsc_schedule_watchdog(sch);
1650 skb = qdisc_dequeue_peeked(cl->qdisc);
1652 qdisc_warn_nonwc("HFSC", cl->qdisc);
1656 bstats_update(&cl->bstats, skb);
1657 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1659 cl->cl_cumul += qdisc_pkt_len(skb);
1661 if (cl->qdisc->q.qlen != 0) {
1662 if (cl->cl_flags & HFSC_RSC) {
1664 next_len = qdisc_peek_len(cl->qdisc);
1666 update_ed(cl, next_len);
1668 update_d(cl, next_len);
1671 /* the class becomes passive */
1675 qdisc_unthrottled(sch);
1676 qdisc_bstats_update(sch, skb);
1683 hfsc_drop(struct Qdisc *sch)
1685 struct hfsc_sched *q = qdisc_priv(sch);
1686 struct hfsc_class *cl;
1689 list_for_each_entry(cl, &q->droplist, dlist) {
1690 if (cl->qdisc->ops->drop != NULL &&
1691 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1692 if (cl->qdisc->q.qlen == 0) {
1693 update_vf(cl, 0, 0);
1696 list_move_tail(&cl->dlist, &q->droplist);
1699 sch->qstats.drops++;
1707 static const struct Qdisc_class_ops hfsc_class_ops = {
1708 .change = hfsc_change_class,
1709 .delete = hfsc_delete_class,
1710 .graft = hfsc_graft_class,
1711 .leaf = hfsc_class_leaf,
1712 .qlen_notify = hfsc_qlen_notify,
1713 .get = hfsc_get_class,
1714 .put = hfsc_put_class,
1715 .bind_tcf = hfsc_bind_tcf,
1716 .unbind_tcf = hfsc_unbind_tcf,
1717 .tcf_chain = hfsc_tcf_chain,
1718 .dump = hfsc_dump_class,
1719 .dump_stats = hfsc_dump_class_stats,
1723 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1725 .init = hfsc_init_qdisc,
1726 .change = hfsc_change_qdisc,
1727 .reset = hfsc_reset_qdisc,
1728 .destroy = hfsc_destroy_qdisc,
1729 .dump = hfsc_dump_qdisc,
1730 .enqueue = hfsc_enqueue,
1731 .dequeue = hfsc_dequeue,
1732 .peek = qdisc_peek_dequeued,
1734 .cl_ops = &hfsc_class_ops,
1735 .priv_size = sizeof(struct hfsc_sched),
1736 .owner = THIS_MODULE
1742 return register_qdisc(&hfsc_qdisc_ops);
1748 unregister_qdisc(&hfsc_qdisc_ops);
1751 MODULE_LICENSE("GPL");
1752 module_init(hfsc_init);
1753 module_exit(hfsc_cleanup);