2 /* $KAME: altq_rmclass.c,v 1.19 2005/04/13 03:44:25 suz Exp $ */
5 * Copyright (c) 1991-1997 Regents of the University of California.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the Network Research
19 * Group at Lawrence Berkeley Laboratory.
20 * 4. Neither the name of the University nor of the Laboratory may be used
21 * to endorse or promote products derived from this software without
22 * specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * LBL code modified by speer@eng.sun.com, May 1977.
37 * For questions and/or comments, please send mail to cbq@ee.lbl.gov
39 * @(#)rm_class.c 1.48 97/12/05 SMI
41 #if defined(__FreeBSD__) || defined(__NetBSD__)
45 #include "opt_inet6.h"
47 #endif /* __FreeBSD__ || __NetBSD__ */
48 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ 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>
58 #include <sys/kernel.h>
62 #include <net/if_var.h>
64 #include <netinet/in.h>
65 #include <netinet/in_systm.h>
66 #include <netinet/ip.h>
69 #include <altq/if_altq.h>
70 #include <altq/altq.h>
71 #include <altq/altq_rmclass.h>
72 #include <altq/altq_rmclass_debug.h>
73 #include <altq/altq_red.h>
74 #include <altq/altq_rio.h>
80 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
86 static int rmc_satisfied(struct rm_class *, struct timeval *);
87 static void rmc_wrr_set_weights(struct rm_ifdat *);
88 static void rmc_depth_compute(struct rm_class *);
89 static void rmc_depth_recompute(rm_class_t *);
91 static mbuf_t *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
92 static mbuf_t *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
94 static int _rmc_addq(rm_class_t *, mbuf_t *);
95 static void _rmc_dropq(rm_class_t *);
96 static mbuf_t *_rmc_getq(rm_class_t *);
97 static mbuf_t *_rmc_pollq(rm_class_t *);
99 static int rmc_under_limit(struct rm_class *, struct timeval *);
100 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
101 static void rmc_drop_action(struct rm_class *);
102 static void rmc_restart(struct rm_class *);
103 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
105 #define BORROW_OFFTIME
107 * BORROW_OFFTIME (experimental):
108 * borrow the offtime of the class borrowing from.
109 * the reason is that when its own offtime is set, the class is unable
110 * to borrow much, especially when cutoff is taking effect.
111 * but when the borrowed class is overloaded (advidle is close to minidle),
112 * use the borrowing class's offtime to avoid overload.
114 #define ADJUST_CUTOFF
116 * ADJUST_CUTOFF (experimental):
117 * if no underlimit class is found due to cutoff, increase cutoff and
118 * retry the scheduling loop.
119 * also, don't invoke delay_actions while cutoff is taking effect,
120 * since a sleeping class won't have a chance to be scheduled in the
123 * now heuristics for setting the top-level variable (cutoff_) becomes:
124 * 1. if a packet arrives for a not-overlimit class, set cutoff
125 * to the depth of the class.
126 * 2. if cutoff is i, and a packet arrives for an overlimit class
127 * with an underlimit ancestor at a lower level than i (say j),
128 * then set cutoff to j.
129 * 3. at scheduling a packet, if there is no underlimit class
130 * due to the current cutoff level, increase cutoff by 1 and
131 * then try to schedule again.
136 * rmc_newclass(...) - Create a new resource management class at priority
137 * 'pri' on the interface given by 'ifd'.
139 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
140 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
141 * than 100% of the bandwidth, this number should be the
142 * 'effective' rate for the class. Let f be the
143 * bandwidth fraction allocated to this class, and let
144 * nsPerByte be the data rate of the output link in
145 * nanoseconds/byte. Then nsecPerByte is set to
146 * nsPerByte / f. E.g., 1600 (= 800 / .5)
147 * for a class that gets 50% of an ethernet's bandwidth.
149 * action the routine to call when the class is over limit.
151 * maxq max allowable queue size for class (in packets).
153 * parent parent class pointer.
155 * borrow class to borrow from (should be either 'parent' or null).
157 * maxidle max value allowed for class 'idle' time estimate (this
158 * parameter determines how large an initial burst of packets
159 * can be before overlimit action is invoked.
161 * offtime how long 'delay' action will delay when class goes over
162 * limit (this parameter determines the steady-state burst
163 * size when a class is running over its limit).
165 * Maxidle and offtime have to be computed from the following: If the
166 * average packet size is s, the bandwidth fraction allocated to this
167 * class is f, we want to allow b packet bursts, and the gain of the
168 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
170 * ptime = s * nsPerByte * (1 - f) / f
171 * maxidle = ptime * (1 - g^b) / g^b
172 * minidle = -ptime * (1 / (f - 1))
173 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
175 * Operationally, it's convenient to specify maxidle & offtime in units
176 * independent of the link bandwidth so the maxidle & offtime passed to
177 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
178 * (The constant factor is a scale factor needed to make the parameters
179 * integers. This scaling also means that the 'unscaled' values of
180 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
181 * not nanoseconds.) Also note that the 'idle' filter computation keeps
182 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
183 * maxidle also must be scaled upward by this value. Thus, the passed
184 * values for maxidle and offtime can be computed as follows:
186 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
187 * offtime = offtime * 8 / (1000 * nsecPerByte)
189 * When USE_HRTIME is employed, then maxidle and offtime become:
190 * maxidle = maxilde * (8.0 / nsecPerByte);
191 * offtime = offtime * (8.0 / nsecPerByte);
194 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
195 void (*action)(rm_class_t *, rm_class_t *), int maxq,
196 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
197 int minidle, u_int offtime, int pktsize, int flags)
200 struct rm_class *peer;
203 if (pri >= RM_MAXPRIO)
206 if (flags & RMCF_RED) {
208 printf("rmc_newclass: RED not configured for CBQ!\n");
214 if (flags & RMCF_RIO) {
216 printf("rmc_newclass: RIO not configured for CBQ!\n");
222 cl = malloc(sizeof(struct rm_class), M_DEVBUF, M_NOWAIT | M_ZERO);
225 CALLOUT_INIT(&cl->callout_);
226 cl->q_ = malloc(sizeof(class_queue_t), M_DEVBUF, M_NOWAIT | M_ZERO);
227 if (cl->q_ == NULL) {
233 * Class initialization.
235 cl->children_ = NULL;
236 cl->parent_ = parent;
237 cl->borrow_ = borrow;
241 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
244 cl->ns_per_byte_ = nsecPerByte;
246 qlimit(cl->q_) = maxq;
247 qtype(cl->q_) = Q_DROPHEAD;
251 #if 1 /* minidle is also scaled in ALTQ */
252 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
253 if (cl->minidle_ > 0)
256 cl->minidle_ = minidle;
258 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
259 if (cl->maxidle_ == 0)
261 #if 1 /* offtime is also scaled in ALTQ */
262 cl->avgidle_ = cl->maxidle_;
263 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
264 if (cl->offtime_ == 0)
268 cl->offtime_ = (offtime * nsecPerByte) / 8;
270 cl->overlimit = action;
273 if (flags & (RMCF_RED|RMCF_RIO)) {
274 int red_flags, red_pkttime;
277 if (flags & RMCF_ECN)
278 red_flags |= REDF_ECN;
279 if (flags & RMCF_FLOWVALVE)
280 red_flags |= REDF_FLOWVALVE;
282 if (flags & RMCF_CLEARDSCP)
283 red_flags |= RIOF_CLEARDSCP;
285 red_pkttime = nsecPerByte * pktsize / 1000;
287 if (flags & RMCF_RED) {
288 cl->red_ = red_alloc(0, 0,
289 qlimit(cl->q_) * 10/100,
290 qlimit(cl->q_) * 30/100,
291 red_flags, red_pkttime);
292 if (cl->red_ != NULL)
293 qtype(cl->q_) = Q_RED;
297 cl->red_ = (red_t *)rio_alloc(0, NULL,
298 red_flags, red_pkttime);
299 if (cl->red_ != NULL)
300 qtype(cl->q_) = Q_RIO;
304 #endif /* ALTQ_RED */
307 * put the class into the class tree
315 if ((peer = ifd->active_[pri]) != NULL) {
316 /* find the last class at this pri */
318 while (peer->peer_ != ifd->active_[pri])
322 ifd->active_[pri] = cl;
327 cl->next_ = parent->children_;
328 parent->children_ = cl;
333 * Compute the depth of this class and its ancestors in the class
336 rmc_depth_compute(cl);
339 * If CBQ's WRR is enabled, then initialize the class WRR state.
343 ifd->alloc_[pri] += cl->allotment_;
344 rmc_wrr_set_weights(ifd);
346 IFQ_UNLOCK(ifd->ifq_);
352 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
353 int minidle, u_int offtime, int pktsize)
355 struct rm_ifdat *ifd;
360 old_allotment = cl->allotment_;
368 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
370 cl->ns_per_byte_ = nsecPerByte;
372 qlimit(cl->q_) = maxq;
374 #if 1 /* minidle is also scaled in ALTQ */
375 cl->minidle_ = (minidle * nsecPerByte) / 8;
376 if (cl->minidle_ > 0)
379 cl->minidle_ = minidle;
381 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
382 if (cl->maxidle_ == 0)
384 #if 1 /* offtime is also scaled in ALTQ */
385 cl->avgidle_ = cl->maxidle_;
386 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
387 if (cl->offtime_ == 0)
391 cl->offtime_ = (offtime * nsecPerByte) / 8;
395 * If CBQ's WRR is enabled, then initialize the class WRR state.
398 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
399 rmc_wrr_set_weights(ifd);
401 IFQ_UNLOCK(ifd->ifq_);
408 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
409 * the appropriate run robin weights for the CBQ weighted round robin
416 rmc_wrr_set_weights(struct rm_ifdat *ifd)
419 struct rm_class *cl, *clh;
421 for (i = 0; i < RM_MAXPRIO; i++) {
423 * This is inverted from that of the simulator to
424 * maintain precision.
426 if (ifd->num_[i] == 0)
429 ifd->M_[i] = ifd->alloc_[i] /
430 (ifd->num_[i] * ifd->maxpkt_);
432 * Compute the weighted allotment for each class.
433 * This takes the expensive div instruction out
434 * of the main loop for the wrr scheduling path.
435 * These only get recomputed when a class comes or
438 if (ifd->active_[i] != NULL) {
439 clh = cl = ifd->active_[i];
441 /* safe-guard for slow link or alloc_ == 0 */
443 cl->w_allotment_ = 0;
445 cl->w_allotment_ = cl->allotment_ /
448 } while ((cl != NULL) && (cl != clh));
454 rmc_get_weight(struct rm_ifdat *ifd, int pri)
456 if ((pri >= 0) && (pri < RM_MAXPRIO))
457 return (ifd->M_[pri]);
464 * rmc_depth_compute(struct rm_class *cl) - This function computes the
465 * appropriate depth of class 'cl' and its ancestors.
471 rmc_depth_compute(struct rm_class *cl)
473 rm_class_t *t = cl, *p;
476 * Recompute the depth for the branch of the tree.
480 if (p && (t->depth_ >= p->depth_)) {
481 p->depth_ = t->depth_ + 1;
490 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
491 * the depth of the tree after a class has been deleted.
497 rmc_depth_recompute(rm_class_t *cl)
504 if ((t = p->children_) == NULL) {
510 if (t->depth_ > cdepth)
515 if (p->depth_ == cdepth + 1)
516 /* no change to this parent */
519 p->depth_ = cdepth + 1;
527 if (cl->depth_ >= 1) {
528 if (cl->children_ == NULL) {
530 } else if ((t = cl->children_) != NULL) {
532 if (t->children_ != NULL)
533 rmc_depth_recompute(t);
537 rmc_depth_compute(cl);
544 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
545 * function deletes a class from the link-sharing structure and frees
546 * all resources associated with the class.
552 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
554 struct rm_class *p, *head, *previous;
557 ASSERT(cl->children_ == NULL);
560 CALLOUT_STOP(&cl->callout_);
569 * Free packets in the packet queue.
570 * XXX - this may not be a desired behavior. Packets should be
576 * If the class has a parent, then remove the class from the
577 * class from the parent's children chain.
579 if (cl->parent_ != NULL) {
580 head = cl->parent_->children_;
582 if (head->next_ == NULL) {
584 cl->parent_->children_ = NULL;
585 cl->parent_->leaf_ = 1;
586 } else while (p != NULL) {
589 cl->parent_->children_ = cl->next_;
591 previous->next_ = cl->next_;
602 * Delete class from class priority peer list.
604 if ((p = ifd->active_[cl->pri_]) != NULL) {
606 * If there is more than one member of this priority
607 * level, then look for class(cl) in the priority level.
610 while (p->peer_ != cl)
612 p->peer_ = cl->peer_;
614 if (ifd->active_[cl->pri_] == cl)
615 ifd->active_[cl->pri_] = cl->peer_;
618 ifd->active_[cl->pri_] = NULL;
623 * Recompute the WRR weights.
626 ifd->alloc_[cl->pri_] -= cl->allotment_;
627 ifd->num_[cl->pri_]--;
628 rmc_wrr_set_weights(ifd);
632 * Re-compute the depth of the tree.
635 rmc_depth_recompute(cl->parent_);
637 rmc_depth_recompute(ifd->root_);
640 IFQ_UNLOCK(ifd->ifq_);
644 * Free the class structure.
646 if (cl->red_ != NULL) {
648 if (q_is_rio(cl->q_))
649 rio_destroy((rio_t *)cl->red_);
652 if (q_is_red(cl->q_))
653 red_destroy(cl->red_);
656 free(cl->q_, M_DEVBUF);
663 * rmc_init(...) - Initialize the resource management data structures
664 * associated with the output portion of interface 'ifp'. 'ifd' is
665 * where the structures will be built (for backwards compatibility, the
666 * structures aren't kept in the ifnet struct). 'nsecPerByte'
667 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
668 * 'restart' is the driver-specific routine that the generic 'delay
669 * until under limit' action will call to restart output. `maxq'
670 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
671 * is the maximum number of packets that the resource management
672 * code will allow to be queued 'downstream' (this is typically 1).
678 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
679 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
680 int minidle, u_int offtime, int flags)
685 * Initialize the CBQ tracing/debug facility.
689 bzero((char *)ifd, sizeof (*ifd));
690 mtu = ifq->altq_ifp->if_mtu;
692 ifd->restart = restart;
693 ifd->maxqueued_ = maxqueued;
694 ifd->ns_per_byte_ = nsecPerByte;
696 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
697 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
699 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
700 if (mtu * nsecPerByte > 10 * 1000000)
701 ifd->maxiftime_ /= 4;
705 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
708 * Initialize the CBQ's WRR state.
710 for (i = 0; i < RM_MAXPRIO; i++) {
715 ifd->active_[i] = NULL;
719 * Initialize current packet state.
723 for (i = 0; i < RM_MAXQUEUED; i++) {
724 ifd->class_[i] = NULL;
726 ifd->borrowed_[i] = NULL;
730 * Create the root class of the link-sharing structure.
732 if ((ifd->root_ = rmc_newclass(0, ifd,
734 rmc_root_overlimit, maxq, 0, 0,
735 maxidle, minidle, offtime,
737 printf("rmc_init: root class not allocated\n");
740 ifd->root_->depth_ = 0;
745 * rmc_queue_packet(struct rm_class *cl, mbuf_t *m) - Add packet given by
746 * mbuf 'm' to queue for resource class 'cl'. This routine is called
747 * by a driver's if_output routine. This routine must be called with
748 * output packet completion interrupts locked out (to avoid racing with
751 * Returns: 0 on successful queueing
752 * -1 when packet drop occurs
755 rmc_queue_packet(struct rm_class *cl, mbuf_t *m)
758 struct rm_ifdat *ifd = cl->ifdat_;
760 int is_empty = qempty(cl->q_);
763 if (ifd->cutoff_ > 0) {
764 if (TV_LT(&cl->undertime_, &now)) {
765 if (ifd->cutoff_ > cl->depth_)
766 ifd->cutoff_ = cl->depth_;
767 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
772 * the class is overlimit. if the class has
773 * underlimit ancestors, set cutoff to the lowest
776 struct rm_class *borrow = cl->borrow_;
778 while (borrow != NULL &&
779 borrow->depth_ < ifd->cutoff_) {
780 if (TV_LT(&borrow->undertime_, &now)) {
781 ifd->cutoff_ = borrow->depth_;
782 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
785 borrow = borrow->borrow_;
789 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
790 if (TV_LT(&cl->borrow_->undertime_, &now)) {
791 ifd->cutoff_ = cl->borrow_->depth_;
792 CBQTRACE(rmc_queue_packet, 'ffob',
793 cl->borrow_->depth_);
799 if (_rmc_addq(cl, m) < 0)
804 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
808 if (qlen(cl->q_) > qlimit(cl->q_)) {
809 /* note: qlimit can be set to 0 or 1 */
818 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
819 * classes to see if there are satified.
823 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
828 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
829 if ((bp = ifd->active_[i]) != NULL) {
832 if (!rmc_satisfied(p, now)) {
833 ifd->cutoff_ = p->depth_;
845 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
849 rmc_satisfied(struct rm_class *cl, struct timeval *now)
855 if (TV_LT(now, &cl->undertime_))
857 if (cl->depth_ == 0) {
858 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
863 if (cl->children_ != NULL) {
866 if (!rmc_satisfied(p, now))
876 * Return 1 if class 'cl' is under limit or can borrow from a parent,
877 * 0 if overlimit. As a side-effect, this routine will invoke the
878 * class overlimit action if the class if overlimit.
882 rmc_under_limit(struct rm_class *cl, struct timeval *now)
886 struct rm_ifdat *ifd = cl->ifdat_;
888 ifd->borrowed_[ifd->qi_] = NULL;
890 * If cl is the root class, then always return that it is
891 * underlimit. Otherwise, check to see if the class is underlimit.
893 if (cl->parent_ == NULL)
897 if (TV_LT(now, &cl->undertime_))
900 CALLOUT_STOP(&cl->callout_);
902 cl->undertime_.tv_sec = 0;
907 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
908 if (((cl = cl->borrow_) == NULL) ||
909 (cl->depth_ > ifd->cutoff_)) {
912 /* cutoff is taking effect, just
913 return false without calling
917 #ifdef BORROW_OFFTIME
919 * check if the class can borrow offtime too.
920 * borrow offtime from the top of the borrow
921 * chain if the top class is not overloaded.
924 /* cutoff is taking effect, use this class as top. */
926 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
928 if (top != NULL && top->avgidle_ == top->minidle_)
931 (p->overlimit)(p, top);
934 (p->overlimit)(p, NULL);
942 ifd->borrowed_[ifd->qi_] = cl;
947 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
948 * Packet-by-packet round robin.
950 * The heart of the weighted round-robin scheduler, which decides which
951 * class next gets to send a packet. Highest priority first, then
952 * weighted round-robin within priorites.
954 * Each able-to-send class gets to send until its byte allocation is
955 * exhausted. Thus, the active pointer is only changed after a class has
956 * exhausted its allocation.
958 * If the scheduler finds no class that is underlimit or able to borrow,
959 * then the first class found that had a nonzero queue and is allowed to
960 * borrow gets to send.
964 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
966 struct rm_class *cl = NULL, *first = NULL;
975 * if the driver polls the top of the queue and then removes
976 * the polled packet, we must return the same packet.
978 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
979 cl = ifd->pollcache_;
981 if (ifd->efficient_) {
982 /* check if this class is overlimit */
983 if (cl->undertime_.tv_sec != 0 &&
984 rmc_under_limit(cl, &now) == 0)
987 ifd->pollcache_ = NULL;
991 /* mode == ALTDQ_POLL || pollcache == NULL */
992 ifd->pollcache_ = NULL;
993 ifd->borrowed_[ifd->qi_] = NULL;
998 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
999 if (ifd->na_[cpri] == 0)
1003 * Loop through twice for a priority level, if some class
1004 * was unable to send a packet the first round because
1005 * of the weighted round-robin mechanism.
1006 * During the second loop at this level, deficit==2.
1007 * (This second loop is not needed if for every class,
1008 * "M[cl->pri_])" times "cl->allotment" is greater than
1009 * the byte size for the largest packet in the class.)
1012 cl = ifd->active_[cpri];
1015 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
1016 cl->bytes_alloc_ += cl->w_allotment_;
1017 if (!qempty(cl->q_)) {
1018 if ((cl->undertime_.tv_sec == 0) ||
1019 rmc_under_limit(cl, &now)) {
1020 if (cl->bytes_alloc_ > 0 || deficit > 1)
1023 /* underlimit but no alloc */
1026 ifd->borrowed_[ifd->qi_] = NULL;
1029 else if (first == NULL && cl->borrow_ != NULL)
1030 first = cl; /* borrowing candidate */
1033 cl->bytes_alloc_ = 0;
1035 } while (cl != ifd->active_[cpri]);
1038 /* first loop found an underlimit class with deficit */
1039 /* Loop on same priority level, with new deficit. */
1045 #ifdef ADJUST_CUTOFF
1047 * no underlimit class found. if cutoff is taking effect,
1048 * increase cutoff and try again.
1050 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1052 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1055 #endif /* ADJUST_CUTOFF */
1057 * If LINK_EFFICIENCY is turned on, then the first overlimit
1058 * class we encounter will send a packet if all the classes
1059 * of the link-sharing structure are overlimit.
1062 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1064 if (!ifd->efficient_ || first == NULL)
1069 #if 0 /* too time-consuming for nothing */
1071 CALLOUT_STOP(&cl->callout_);
1073 cl->undertime_.tv_sec = 0;
1075 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1076 ifd->cutoff_ = cl->borrow_->depth_;
1079 * Deque the packet and do the book keeping...
1082 if (op == ALTDQ_REMOVE) {
1085 panic("_rmc_wrr_dequeue_next");
1090 * Update class statistics and link data.
1092 if (cl->bytes_alloc_ > 0)
1093 cl->bytes_alloc_ -= m_pktlen(m);
1095 if ((cl->bytes_alloc_ <= 0) || first == cl)
1096 ifd->active_[cl->pri_] = cl->peer_;
1098 ifd->active_[cl->pri_] = cl;
1100 ifd->class_[ifd->qi_] = cl;
1101 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1102 ifd->now_[ifd->qi_] = now;
1103 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1106 /* mode == ALTDQ_PPOLL */
1108 ifd->pollcache_ = cl;
1114 * Dequeue & return next packet from the highest priority class that
1115 * has a packet to send & has enough allocation to send it. This
1116 * routine is called by a driver whenever it needs a new packet to
1120 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1124 struct rm_class *cl, *first = NULL;
1130 * if the driver polls the top of the queue and then removes
1131 * the polled packet, we must return the same packet.
1133 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1134 cl = ifd->pollcache_;
1136 ifd->pollcache_ = NULL;
1139 /* mode == ALTDQ_POLL || pollcache == NULL */
1140 ifd->pollcache_ = NULL;
1141 ifd->borrowed_[ifd->qi_] = NULL;
1143 #ifdef ADJUST_CUTOFF
1146 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1147 if (ifd->na_[cpri] == 0)
1149 cl = ifd->active_[cpri];
1152 if (!qempty(cl->q_)) {
1153 if ((cl->undertime_.tv_sec == 0) ||
1154 rmc_under_limit(cl, &now))
1156 if (first == NULL && cl->borrow_ != NULL)
1160 } while (cl != ifd->active_[cpri]);
1163 #ifdef ADJUST_CUTOFF
1165 * no underlimit class found. if cutoff is taking effect, increase
1166 * cutoff and try again.
1168 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1172 #endif /* ADJUST_CUTOFF */
1174 * If LINK_EFFICIENCY is turned on, then the first overlimit
1175 * class we encounter will send a packet if all the classes
1176 * of the link-sharing structure are overlimit.
1179 if (!ifd->efficient_ || first == NULL)
1184 #if 0 /* too time-consuming for nothing */
1186 CALLOUT_STOP(&cl->callout_);
1188 cl->undertime_.tv_sec = 0;
1190 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1191 ifd->cutoff_ = cl->borrow_->depth_;
1194 * Deque the packet and do the book keeping...
1197 if (op == ALTDQ_REMOVE) {
1200 panic("_rmc_prr_dequeue_next");
1204 ifd->active_[cpri] = cl->peer_;
1206 ifd->class_[ifd->qi_] = cl;
1207 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1208 ifd->now_[ifd->qi_] = now;
1209 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1212 /* mode == ALTDQ_POLL */
1214 ifd->pollcache_ = cl;
1221 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1222 * is invoked by the packet driver to get the next packet to be
1223 * dequeued and output on the link. If WRR is enabled, then the
1224 * WRR dequeue next routine will determine the next packet to sent.
1225 * Otherwise, packet-by-packet round robin is invoked.
1227 * Returns: NULL, if a packet is not available or if all
1228 * classes are overlimit.
1230 * Otherwise, Pointer to the next packet.
1234 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1236 if (ifd->queued_ >= ifd->maxqueued_)
1239 return (_rmc_wrr_dequeue_next(ifd, mode));
1241 return (_rmc_prr_dequeue_next(ifd, mode));
1245 * Update the utilization estimate for the packet that just completed.
1246 * The packet's class & the parent(s) of that class all get their
1247 * estimators updated. This routine is called by the driver's output-
1248 * packet-completion interrupt service routine.
1252 * a macro to approximate "divide by 1000" that gives 0.000999,
1253 * if a value has enough effective digits.
1254 * (on pentium, mul takes 9 cycles but div takes 46!)
1256 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1258 rmc_update_class_util(struct rm_ifdat *ifd)
1260 int idle, avgidle, pktlen;
1261 int pkt_time, tidle;
1262 rm_class_t *cl, *borrowed;
1263 rm_class_t *borrows;
1264 struct timeval *nowp;
1267 * Get the most recent completed class.
1269 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1272 pktlen = ifd->curlen_[ifd->qo_];
1273 borrowed = ifd->borrowed_[ifd->qo_];
1276 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1279 * Run estimator on class and its ancestors.
1282 * rm_update_class_util is designed to be called when the
1283 * transfer is completed from a xmit complete interrupt,
1284 * but most drivers don't implement an upcall for that.
1285 * so, just use estimated completion time.
1286 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1288 nowp = &ifd->now_[ifd->qo_];
1289 /* get pkt_time (for link) in usec */
1290 #if 1 /* use approximation */
1291 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1292 pkt_time = NSEC_TO_USEC(pkt_time);
1294 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1296 #if 1 /* ALTQ4PPP */
1297 if (TV_LT(nowp, &ifd->ifnow_)) {
1301 * make sure the estimated completion time does not go
1302 * too far. it can happen when the link layer supports
1303 * data compression or the interface speed is set to
1304 * a much lower value.
1306 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1307 if (iftime+pkt_time < ifd->maxiftime_) {
1308 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1310 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1313 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1316 if (TV_LT(nowp, &ifd->ifnow_)) {
1317 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1319 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1323 while (cl != NULL) {
1324 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1325 if (idle >= 2000000)
1327 * this class is idle enough, reset avgidle.
1328 * (TV_DELTA returns 2000000 us when delta is large.)
1330 cl->avgidle_ = cl->maxidle_;
1332 /* get pkt_time (for class) in usec */
1333 #if 1 /* use approximation */
1334 pkt_time = pktlen * cl->ns_per_byte_;
1335 pkt_time = NSEC_TO_USEC(pkt_time);
1337 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1341 avgidle = cl->avgidle_;
1342 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1343 cl->avgidle_ = avgidle;
1345 /* Are we overlimit ? */
1347 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1350 * need some lower bound for avgidle, otherwise
1351 * a borrowing class gets unbounded penalty.
1353 if (avgidle < cl->minidle_)
1354 avgidle = cl->avgidle_ = cl->minidle_;
1356 /* set next idle to make avgidle 0 */
1358 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1359 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1363 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1364 cl->undertime_.tv_sec = 0;
1365 if (cl->sleeping_) {
1366 CALLOUT_STOP(&cl->callout_);
1371 if (borrows != NULL) {
1373 ++cl->stats_.borrows;
1377 cl->last_ = ifd->ifnow_;
1378 cl->last_pkttime_ = pkt_time;
1381 if (cl->parent_ == NULL) {
1382 /* take stats of root class */
1383 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1391 * Check to see if cutoff needs to set to a new level.
1393 cl = ifd->class_[ifd->qo_];
1394 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1396 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1397 rmc_tl_satisfied(ifd, nowp);
1398 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1400 ifd->cutoff_ = borrowed->depth_;
1401 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1404 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1407 rmc_tl_satisfied(ifd, &now);
1409 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1411 ifd->cutoff_ = borrowed->depth_;
1412 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1418 * Release class slot
1420 ifd->borrowed_[ifd->qo_] = NULL;
1421 ifd->class_[ifd->qo_] = NULL;
1422 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1428 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1429 * over-limit action routines. These get invoked by rmc_under_limit()
1430 * if a class with packets to send if over its bandwidth limit & can't
1431 * borrow from a parent class.
1437 rmc_drop_action(struct rm_class *cl)
1439 struct rm_ifdat *ifd = cl->ifdat_;
1441 ASSERT(qlen(cl->q_) > 0);
1444 ifd->na_[cl->pri_]--;
1447 void rmc_dropall(struct rm_class *cl)
1449 struct rm_ifdat *ifd = cl->ifdat_;
1451 if (!qempty(cl->q_)) {
1454 ifd->na_[cl->pri_]--;
1458 #if (__FreeBSD_version > 300000)
1459 /* hzto() is removed from FreeBSD-3.0 */
1460 static int hzto(struct timeval *);
1469 t2.tv_sec = tv->tv_sec - t2.tv_sec;
1470 t2.tv_usec = tv->tv_usec - t2.tv_usec;
1471 return (tvtohz(&t2));
1473 #endif /* __FreeBSD_version > 300000 */
1477 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1478 * delay action routine. It is invoked via rmc_under_limit when the
1479 * packet is discoverd to be overlimit.
1481 * If the delay action is result of borrow class being overlimit, then
1482 * delay for the offtime of the borrowing class that is overlimit.
1488 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1490 int delay, t, extradelay;
1492 cl->stats_.overactions++;
1493 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1494 #ifndef BORROW_OFFTIME
1495 delay += cl->offtime_;
1498 if (!cl->sleeping_) {
1499 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1500 #ifdef BORROW_OFFTIME
1502 extradelay = borrow->offtime_;
1505 extradelay = cl->offtime_;
1509 * XXX recalculate suspend time:
1510 * current undertime is (tidle + pkt_time) calculated
1511 * from the last transmission.
1512 * tidle: time required to bring avgidle back to 0
1513 * pkt_time: target waiting time for this class
1514 * we need to replace pkt_time by offtime
1516 extradelay -= cl->last_pkttime_;
1518 if (extradelay > 0) {
1519 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1520 delay += extradelay;
1524 cl->stats_.delays++;
1527 * Since packets are phased randomly with respect to the
1528 * clock, 1 tick (the next clock tick) can be an arbitrarily
1529 * short time so we have to wait for at least two ticks.
1530 * NOTE: If there's no other traffic, we need the timer as
1531 * a 'backstop' to restart this class.
1533 if (delay > tick * 2) {
1535 /* FreeBSD rounds up the tick */
1536 t = hzto(&cl->undertime_);
1538 /* other BSDs round down the tick */
1539 t = hzto(&cl->undertime_) + 1;
1543 CALLOUT_RESET(&cl->callout_, t,
1544 (timeout_t *)rmc_restart, (caddr_t)cl);
1550 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1551 * called by the system timer code & is responsible checking if the
1552 * class is still sleeping (it might have been restarted as a side
1553 * effect of the queue scan on a packet arrival) and, if so, restarting
1554 * output for the class. Inspecting the class state & restarting output
1555 * require locking the class structure. In general the driver is
1556 * responsible for locking but this is the only routine that is not
1557 * called directly or indirectly from the interface driver so it has
1558 * know about system locking conventions. Under bsd, locking is done
1559 * by raising IPL to splimp so that's what's implemented here. On a
1560 * different system this would probably need to be changed.
1566 rmc_restart(struct rm_class *cl)
1568 struct rm_ifdat *ifd = cl->ifdat_;
1576 IFQ_LOCK(ifd->ifq_);
1577 if (cl->sleeping_) {
1579 cl->undertime_.tv_sec = 0;
1581 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1582 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1583 (ifd->restart)(ifd->ifq_);
1586 IFQ_UNLOCK(ifd->ifq_);
1592 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1593 * handling routine for the root class of the link sharing structure.
1599 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1601 panic("rmc_root_overlimit");
1605 * Packet Queue handling routines. Eventually, this is to localize the
1606 * effects on the code whether queues are red queues or droptail
1611 _rmc_addq(rm_class_t *cl, mbuf_t *m)
1614 if (q_is_rio(cl->q_))
1615 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1618 if (q_is_red(cl->q_))
1619 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1620 #endif /* ALTQ_RED */
1622 if (cl->flags_ & RMCF_CLEARDSCP)
1623 write_dsfield(m, cl->pktattr_, 0);
1629 /* note: _rmc_dropq is not called for red */
1631 _rmc_dropq(rm_class_t *cl)
1635 if ((m = _getq(cl->q_)) != NULL)
1640 _rmc_getq(rm_class_t *cl)
1643 if (q_is_rio(cl->q_))
1644 return rio_getq((rio_t *)cl->red_, cl->q_);
1647 if (q_is_red(cl->q_))
1648 return red_getq(cl->red_, cl->q_);
1650 return _getq(cl->q_);
1654 _rmc_pollq(rm_class_t *cl)
1656 return qhead(cl->q_);
1661 struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
1662 struct cbqtrace *cbqtrace_ptr = NULL;
1666 * DDB hook to trace cbq events:
1667 * the last 1024 events are held in a circular buffer.
1668 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1670 void cbqtrace_dump(int);
1671 static char *rmc_funcname(void *);
1673 static struct rmc_funcs {
1678 rmc_init, "rmc_init",
1679 rmc_queue_packet, "rmc_queue_packet",
1680 rmc_under_limit, "rmc_under_limit",
1681 rmc_update_class_util, "rmc_update_class_util",
1682 rmc_delay_action, "rmc_delay_action",
1683 rmc_restart, "rmc_restart",
1684 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1688 static char *rmc_funcname(void *func)
1690 struct rmc_funcs *fp;
1692 for (fp = rmc_funcs; fp->func != NULL; fp++)
1693 if (fp->func == func)
1698 void cbqtrace_dump(int counter)
1703 counter = counter % NCBQTRACE;
1704 p = (int *)&cbqtrace_buffer[counter];
1706 for (i=0; i<20; i++) {
1707 printf("[0x%x] ", *p++);
1708 printf("%s: ", rmc_funcname((void *)*p++));
1710 printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1711 printf("%d\n",*p++);
1713 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1714 p = (int *)cbqtrace_buffer;
1717 #endif /* CBQ_TRACE */
1718 #endif /* ALTQ_CBQ */
1720 #if defined(ALTQ_CBQ) || defined(ALTQ_RED) || defined(ALTQ_RIO) || defined(ALTQ_HFSC) || defined(ALTQ_PRIQ)
1721 #if !defined(__GNUC__) || defined(ALTQ_DEBUG)
1724 _addq(class_queue_t *q, mbuf_t *m)
1728 if ((m0 = qtail(q)) != NULL)
1729 m->m_nextpkt = m0->m_nextpkt;
1738 _getq(class_queue_t *q)
1742 if ((m = qtail(q)) == NULL)
1744 if ((m0 = m->m_nextpkt) != m)
1745 m->m_nextpkt = m0->m_nextpkt;
1747 ASSERT(qlen(q) == 1);
1751 m0->m_nextpkt = NULL;
1755 /* drop a packet at the tail of the queue */
1757 _getq_tail(class_queue_t *q)
1759 mbuf_t *m, *m0, *prev;
1761 if ((m = m0 = qtail(q)) == NULL)
1767 prev->m_nextpkt = m->m_nextpkt;
1769 ASSERT(qlen(q) == 1);
1774 m->m_nextpkt = NULL;
1778 /* randomly select a packet in the queue */
1780 _getq_random(class_queue_t *q)
1785 if ((m = qtail(q)) == NULL)
1787 if (m->m_nextpkt == m) {
1788 ASSERT(qlen(q) == 1);
1791 struct mbuf *prev = NULL;
1793 n = arc4random() % qlen(q) + 1;
1794 for (i = 0; i < n; i++) {
1798 prev->m_nextpkt = m->m_nextpkt;
1803 m->m_nextpkt = NULL;
1808 _removeq(class_queue_t *q, mbuf_t *m)
1817 prev->m_nextpkt = m->m_nextpkt;
1820 else if (qtail(q) == m)
1826 _flushq(class_queue_t *q)
1830 while ((m = _getq(q)) != NULL)
1832 ASSERT(qlen(q) == 0);
1835 #endif /* !__GNUC__ || ALTQ_DEBUG */
1836 #endif /* ALTQ_CBQ || ALTQ_RED || ALTQ_RIO || ALTQ_HFSC || ALTQ_PRIQ */