2 * Copyright (c) 1991-1997 Regents of the University of California.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. All advertising materials mentioning features or use of this software
14 * must display the following acknowledgement:
15 * This product includes software developed by the Network Research
16 * Group at Lawrence Berkeley Laboratory.
17 * 4. Neither the name of the University nor of the Laboratory may be used
18 * to endorse or promote products derived from this software without
19 * specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * LBL code modified by speer@eng.sun.com, May 1977.
34 * For questions and/or comments, please send mail to cbq@ee.lbl.gov
36 * @(#)rm_class.c 1.48 97/12/05 SMI
37 * $KAME: altq_rmclass.c,v 1.19 2005/04/13 03:44:25 suz Exp $
42 #include "opt_inet6.h"
43 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
45 #include <sys/param.h>
46 #include <sys/malloc.h>
48 #include <sys/socket.h>
49 #include <sys/systm.h>
50 #include <sys/errno.h>
54 #include <net/if_var.h>
56 #include <net/altq/if_altq.h>
57 #include <net/altq/altq.h>
58 #include <net/altq/altq_codel.h>
59 #include <net/altq/altq_rmclass.h>
60 #include <net/altq/altq_rmclass_debug.h>
61 #include <net/altq/altq_red.h>
62 #include <net/altq/altq_rio.h>
67 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
73 static int rmc_satisfied(struct rm_class *, struct timeval *);
74 static void rmc_wrr_set_weights(struct rm_ifdat *);
75 static void rmc_depth_compute(struct rm_class *);
76 static void rmc_depth_recompute(rm_class_t *);
78 static mbuf_t *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
79 static mbuf_t *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
81 static int _rmc_addq(rm_class_t *, mbuf_t *);
82 static void _rmc_dropq(rm_class_t *);
83 static mbuf_t *_rmc_getq(rm_class_t *);
84 static mbuf_t *_rmc_pollq(rm_class_t *);
86 static int rmc_under_limit(struct rm_class *, struct timeval *);
87 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
88 static void rmc_drop_action(struct rm_class *);
89 static void rmc_restart(void *);
90 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
92 #define BORROW_OFFTIME
94 * BORROW_OFFTIME (experimental):
95 * borrow the offtime of the class borrowing from.
96 * the reason is that when its own offtime is set, the class is unable
97 * to borrow much, especially when cutoff is taking effect.
98 * but when the borrowed class is overloaded (advidle is close to minidle),
99 * use the borrowing class's offtime to avoid overload.
101 #define ADJUST_CUTOFF
103 * ADJUST_CUTOFF (experimental):
104 * if no underlimit class is found due to cutoff, increase cutoff and
105 * retry the scheduling loop.
106 * also, don't invoke delay_actions while cutoff is taking effect,
107 * since a sleeping class won't have a chance to be scheduled in the
110 * now heuristics for setting the top-level variable (cutoff_) becomes:
111 * 1. if a packet arrives for a not-overlimit class, set cutoff
112 * to the depth of the class.
113 * 2. if cutoff is i, and a packet arrives for an overlimit class
114 * with an underlimit ancestor at a lower level than i (say j),
115 * then set cutoff to j.
116 * 3. at scheduling a packet, if there is no underlimit class
117 * due to the current cutoff level, increase cutoff by 1 and
118 * then try to schedule again.
123 * rmc_newclass(...) - Create a new resource management class at priority
124 * 'pri' on the interface given by 'ifd'.
126 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
127 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
128 * than 100% of the bandwidth, this number should be the
129 * 'effective' rate for the class. Let f be the
130 * bandwidth fraction allocated to this class, and let
131 * nsPerByte be the data rate of the output link in
132 * nanoseconds/byte. Then nsecPerByte is set to
133 * nsPerByte / f. E.g., 1600 (= 800 / .5)
134 * for a class that gets 50% of an ethernet's bandwidth.
136 * action the routine to call when the class is over limit.
138 * maxq max allowable queue size for class (in packets).
140 * parent parent class pointer.
142 * borrow class to borrow from (should be either 'parent' or null).
144 * maxidle max value allowed for class 'idle' time estimate (this
145 * parameter determines how large an initial burst of packets
146 * can be before overlimit action is invoked.
148 * offtime how long 'delay' action will delay when class goes over
149 * limit (this parameter determines the steady-state burst
150 * size when a class is running over its limit).
152 * Maxidle and offtime have to be computed from the following: If the
153 * average packet size is s, the bandwidth fraction allocated to this
154 * class is f, we want to allow b packet bursts, and the gain of the
155 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
157 * ptime = s * nsPerByte * (1 - f) / f
158 * maxidle = ptime * (1 - g^b) / g^b
159 * minidle = -ptime * (1 / (f - 1))
160 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
162 * Operationally, it's convenient to specify maxidle & offtime in units
163 * independent of the link bandwidth so the maxidle & offtime passed to
164 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
165 * (The constant factor is a scale factor needed to make the parameters
166 * integers. This scaling also means that the 'unscaled' values of
167 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
168 * not nanoseconds.) Also note that the 'idle' filter computation keeps
169 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
170 * maxidle also must be scaled upward by this value. Thus, the passed
171 * values for maxidle and offtime can be computed as follows:
173 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
174 * offtime = offtime * 8 / (1000 * nsecPerByte)
176 * When USE_HRTIME is employed, then maxidle and offtime become:
177 * maxidle = maxilde * (8.0 / nsecPerByte);
178 * offtime = offtime * (8.0 / nsecPerByte);
181 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
182 void (*action)(rm_class_t *, rm_class_t *), int maxq,
183 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
184 int minidle, u_int offtime, int pktsize, int flags)
187 struct rm_class *peer;
190 if (pri >= RM_MAXPRIO)
193 if (flags & RMCF_RED) {
195 printf("rmc_newclass: RED not configured for CBQ!\n");
201 if (flags & RMCF_RIO) {
203 printf("rmc_newclass: RIO not configured for CBQ!\n");
209 if (flags & RMCF_CODEL) {
211 printf("rmc_newclass: CODEL not configured for CBQ!\n");
217 cl = malloc(sizeof(struct rm_class), M_DEVBUF, M_NOWAIT | M_ZERO);
220 CALLOUT_INIT(&cl->callout_);
221 cl->q_ = malloc(sizeof(class_queue_t), M_DEVBUF, M_NOWAIT | M_ZERO);
222 if (cl->q_ == NULL) {
228 * Class initialization.
230 cl->children_ = NULL;
231 cl->parent_ = parent;
232 cl->borrow_ = borrow;
236 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
239 cl->ns_per_byte_ = nsecPerByte;
241 qlimit(cl->q_) = maxq;
242 qtype(cl->q_) = Q_DROPHEAD;
246 #if 1 /* minidle is also scaled in ALTQ */
247 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
248 if (cl->minidle_ > 0)
251 cl->minidle_ = minidle;
253 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
254 if (cl->maxidle_ == 0)
256 #if 1 /* offtime is also scaled in ALTQ */
257 cl->avgidle_ = cl->maxidle_;
258 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
259 if (cl->offtime_ == 0)
263 cl->offtime_ = (offtime * nsecPerByte) / 8;
265 cl->overlimit = action;
268 if (flags & (RMCF_RED|RMCF_RIO)) {
269 int red_flags, red_pkttime;
272 if (flags & RMCF_ECN)
273 red_flags |= REDF_ECN;
274 if (flags & RMCF_FLOWVALVE)
275 red_flags |= REDF_FLOWVALVE;
277 if (flags & RMCF_CLEARDSCP)
278 red_flags |= RIOF_CLEARDSCP;
280 red_pkttime = nsecPerByte * pktsize / 1000;
282 if (flags & RMCF_RED) {
283 cl->red_ = red_alloc(0, 0,
284 qlimit(cl->q_) * 10/100,
285 qlimit(cl->q_) * 30/100,
286 red_flags, red_pkttime);
287 if (cl->red_ != NULL)
288 qtype(cl->q_) = Q_RED;
292 cl->red_ = (red_t *)rio_alloc(0, NULL,
293 red_flags, red_pkttime);
294 if (cl->red_ != NULL)
295 qtype(cl->q_) = Q_RIO;
299 #endif /* ALTQ_RED */
301 if (flags & RMCF_CODEL) {
302 cl->codel_ = codel_alloc(5, 100, 0);
303 if (cl->codel_ != NULL)
304 qtype(cl->q_) = Q_CODEL;
309 * put the class into the class tree
313 if ((peer = ifd->active_[pri]) != NULL) {
314 /* find the last class at this pri */
316 while (peer->peer_ != ifd->active_[pri])
320 ifd->active_[pri] = cl;
325 cl->next_ = parent->children_;
326 parent->children_ = cl;
331 * Compute the depth of this class and its ancestors in the class
334 rmc_depth_compute(cl);
337 * If CBQ's WRR is enabled, then initialize the class WRR state.
341 ifd->alloc_[pri] += cl->allotment_;
342 rmc_wrr_set_weights(ifd);
344 IFQ_UNLOCK(ifd->ifq_);
350 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
351 int minidle, u_int offtime, int pktsize)
353 struct rm_ifdat *ifd;
358 old_allotment = cl->allotment_;
362 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
364 cl->ns_per_byte_ = nsecPerByte;
366 qlimit(cl->q_) = maxq;
368 #if 1 /* minidle is also scaled in ALTQ */
369 cl->minidle_ = (minidle * nsecPerByte) / 8;
370 if (cl->minidle_ > 0)
373 cl->minidle_ = minidle;
375 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
376 if (cl->maxidle_ == 0)
378 #if 1 /* offtime is also scaled in ALTQ */
379 cl->avgidle_ = cl->maxidle_;
380 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
381 if (cl->offtime_ == 0)
385 cl->offtime_ = (offtime * nsecPerByte) / 8;
389 * If CBQ's WRR is enabled, then initialize the class WRR state.
392 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
393 rmc_wrr_set_weights(ifd);
395 IFQ_UNLOCK(ifd->ifq_);
402 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
403 * the appropriate run robin weights for the CBQ weighted round robin
410 rmc_wrr_set_weights(struct rm_ifdat *ifd)
413 struct rm_class *cl, *clh;
415 for (i = 0; i < RM_MAXPRIO; i++) {
417 * This is inverted from that of the simulator to
418 * maintain precision.
420 if (ifd->num_[i] == 0)
423 ifd->M_[i] = ifd->alloc_[i] /
424 (ifd->num_[i] * ifd->maxpkt_);
426 * Compute the weighted allotment for each class.
427 * This takes the expensive div instruction out
428 * of the main loop for the wrr scheduling path.
429 * These only get recomputed when a class comes or
432 if (ifd->active_[i] != NULL) {
433 clh = cl = ifd->active_[i];
435 /* safe-guard for slow link or alloc_ == 0 */
437 cl->w_allotment_ = 0;
439 cl->w_allotment_ = cl->allotment_ /
442 } while ((cl != NULL) && (cl != clh));
448 rmc_get_weight(struct rm_ifdat *ifd, int pri)
450 if ((pri >= 0) && (pri < RM_MAXPRIO))
451 return (ifd->M_[pri]);
458 * rmc_depth_compute(struct rm_class *cl) - This function computes the
459 * appropriate depth of class 'cl' and its ancestors.
465 rmc_depth_compute(struct rm_class *cl)
467 rm_class_t *t = cl, *p;
470 * Recompute the depth for the branch of the tree.
474 if (p && (t->depth_ >= p->depth_)) {
475 p->depth_ = t->depth_ + 1;
484 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
485 * the depth of the tree after a class has been deleted.
491 rmc_depth_recompute(rm_class_t *cl)
498 if ((t = p->children_) == NULL) {
504 if (t->depth_ > cdepth)
509 if (p->depth_ == cdepth + 1)
510 /* no change to this parent */
513 p->depth_ = cdepth + 1;
521 if (cl->depth_ >= 1) {
522 if (cl->children_ == NULL) {
524 } else if ((t = cl->children_) != NULL) {
526 if (t->children_ != NULL)
527 rmc_depth_recompute(t);
531 rmc_depth_compute(cl);
538 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
539 * function deletes a class from the link-sharing structure and frees
540 * all resources associated with the class.
546 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
548 struct rm_class *p, *head, *previous;
551 ASSERT(cl->children_ == NULL);
554 CALLOUT_STOP(&cl->callout_);
559 * Free packets in the packet queue.
560 * XXX - this may not be a desired behavior. Packets should be
566 * If the class has a parent, then remove the class from the
567 * class from the parent's children chain.
569 if (cl->parent_ != NULL) {
570 head = cl->parent_->children_;
572 if (head->next_ == NULL) {
574 cl->parent_->children_ = NULL;
575 cl->parent_->leaf_ = 1;
576 } else while (p != NULL) {
579 cl->parent_->children_ = cl->next_;
581 previous->next_ = cl->next_;
592 * Delete class from class priority peer list.
594 if ((p = ifd->active_[cl->pri_]) != NULL) {
596 * If there is more than one member of this priority
597 * level, then look for class(cl) in the priority level.
600 while (p->peer_ != cl)
602 p->peer_ = cl->peer_;
604 if (ifd->active_[cl->pri_] == cl)
605 ifd->active_[cl->pri_] = cl->peer_;
608 ifd->active_[cl->pri_] = NULL;
613 * Recompute the WRR weights.
616 ifd->alloc_[cl->pri_] -= cl->allotment_;
617 ifd->num_[cl->pri_]--;
618 rmc_wrr_set_weights(ifd);
622 * Re-compute the depth of the tree.
625 rmc_depth_recompute(cl->parent_);
627 rmc_depth_recompute(ifd->root_);
630 IFQ_UNLOCK(ifd->ifq_);
634 * Free the class structure.
636 if (cl->red_ != NULL) {
638 if (q_is_rio(cl->q_))
639 rio_destroy((rio_t *)cl->red_);
642 if (q_is_red(cl->q_))
643 red_destroy(cl->red_);
646 if (q_is_codel(cl->q_))
647 codel_destroy(cl->codel_);
650 free(cl->q_, M_DEVBUF);
656 * rmc_init(...) - Initialize the resource management data structures
657 * associated with the output portion of interface 'ifp'. 'ifd' is
658 * where the structures will be built (for backwards compatibility, the
659 * structures aren't kept in the ifnet struct). 'nsecPerByte'
660 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
661 * 'restart' is the driver-specific routine that the generic 'delay
662 * until under limit' action will call to restart output. `maxq'
663 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
664 * is the maximum number of packets that the resource management
665 * code will allow to be queued 'downstream' (this is typically 1).
671 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
672 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
673 int minidle, u_int offtime, int flags)
678 * Initialize the CBQ tracing/debug facility.
682 bzero((char *)ifd, sizeof (*ifd));
683 mtu = ifq->altq_ifp->if_mtu;
685 ifd->restart = restart;
686 ifd->maxqueued_ = maxqueued;
687 ifd->ns_per_byte_ = nsecPerByte;
689 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
690 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
692 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
693 if (mtu * nsecPerByte > 10 * 1000000)
694 ifd->maxiftime_ /= 4;
698 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
701 * Initialize the CBQ's WRR state.
703 for (i = 0; i < RM_MAXPRIO; i++) {
708 ifd->active_[i] = NULL;
712 * Initialize current packet state.
716 for (i = 0; i < RM_MAXQUEUED; i++) {
717 ifd->class_[i] = NULL;
719 ifd->borrowed_[i] = NULL;
723 * Create the root class of the link-sharing structure.
725 if ((ifd->root_ = rmc_newclass(0, ifd,
727 rmc_root_overlimit, maxq, 0, 0,
728 maxidle, minidle, offtime,
730 printf("rmc_init: root class not allocated\n");
733 ifd->root_->depth_ = 0;
738 * rmc_queue_packet(struct rm_class *cl, mbuf_t *m) - Add packet given by
739 * mbuf 'm' to queue for resource class 'cl'. This routine is called
740 * by a driver's if_output routine. This routine must be called with
741 * output packet completion interrupts locked out (to avoid racing with
744 * Returns: 0 on successful queueing
745 * -1 when packet drop occurs
748 rmc_queue_packet(struct rm_class *cl, mbuf_t *m)
751 struct rm_ifdat *ifd = cl->ifdat_;
753 int is_empty = qempty(cl->q_);
756 if (ifd->cutoff_ > 0) {
757 if (TV_LT(&cl->undertime_, &now)) {
758 if (ifd->cutoff_ > cl->depth_)
759 ifd->cutoff_ = cl->depth_;
760 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
765 * the class is overlimit. if the class has
766 * underlimit ancestors, set cutoff to the lowest
769 struct rm_class *borrow = cl->borrow_;
771 while (borrow != NULL &&
772 borrow->depth_ < ifd->cutoff_) {
773 if (TV_LT(&borrow->undertime_, &now)) {
774 ifd->cutoff_ = borrow->depth_;
775 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
778 borrow = borrow->borrow_;
782 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
783 if (TV_LT(&cl->borrow_->undertime_, &now)) {
784 ifd->cutoff_ = cl->borrow_->depth_;
785 CBQTRACE(rmc_queue_packet, 'ffob',
786 cl->borrow_->depth_);
792 if (_rmc_addq(cl, m) < 0)
797 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
801 if (qlen(cl->q_) > qlimit(cl->q_)) {
802 /* note: qlimit can be set to 0 or 1 */
811 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
812 * classes to see if there are satified.
816 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
821 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
822 if ((bp = ifd->active_[i]) != NULL) {
825 if (!rmc_satisfied(p, now)) {
826 ifd->cutoff_ = p->depth_;
838 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
842 rmc_satisfied(struct rm_class *cl, struct timeval *now)
848 if (TV_LT(now, &cl->undertime_))
850 if (cl->depth_ == 0) {
851 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
856 if (cl->children_ != NULL) {
859 if (!rmc_satisfied(p, now))
869 * Return 1 if class 'cl' is under limit or can borrow from a parent,
870 * 0 if overlimit. As a side-effect, this routine will invoke the
871 * class overlimit action if the class if overlimit.
875 rmc_under_limit(struct rm_class *cl, struct timeval *now)
879 struct rm_ifdat *ifd = cl->ifdat_;
881 ifd->borrowed_[ifd->qi_] = NULL;
883 * If cl is the root class, then always return that it is
884 * underlimit. Otherwise, check to see if the class is underlimit.
886 if (cl->parent_ == NULL)
890 if (TV_LT(now, &cl->undertime_))
893 CALLOUT_STOP(&cl->callout_);
895 cl->undertime_.tv_sec = 0;
900 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
901 if (((cl = cl->borrow_) == NULL) ||
902 (cl->depth_ > ifd->cutoff_)) {
905 /* cutoff is taking effect, just
906 return false without calling
910 #ifdef BORROW_OFFTIME
912 * check if the class can borrow offtime too.
913 * borrow offtime from the top of the borrow
914 * chain if the top class is not overloaded.
917 /* cutoff is taking effect, use this class as top. */
919 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
921 if (top != NULL && top->avgidle_ == top->minidle_)
924 (p->overlimit)(p, top);
927 (p->overlimit)(p, NULL);
935 ifd->borrowed_[ifd->qi_] = cl;
940 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
941 * Packet-by-packet round robin.
943 * The heart of the weighted round-robin scheduler, which decides which
944 * class next gets to send a packet. Highest priority first, then
945 * weighted round-robin within priorites.
947 * Each able-to-send class gets to send until its byte allocation is
948 * exhausted. Thus, the active pointer is only changed after a class has
949 * exhausted its allocation.
951 * If the scheduler finds no class that is underlimit or able to borrow,
952 * then the first class found that had a nonzero queue and is allowed to
953 * borrow gets to send.
957 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
959 struct rm_class *cl = NULL, *first = NULL;
968 * if the driver polls the top of the queue and then removes
969 * the polled packet, we must return the same packet.
971 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
972 cl = ifd->pollcache_;
974 if (ifd->efficient_) {
975 /* check if this class is overlimit */
976 if (cl->undertime_.tv_sec != 0 &&
977 rmc_under_limit(cl, &now) == 0)
980 ifd->pollcache_ = NULL;
984 /* mode == ALTDQ_POLL || pollcache == NULL */
985 ifd->pollcache_ = NULL;
986 ifd->borrowed_[ifd->qi_] = NULL;
991 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
992 if (ifd->na_[cpri] == 0)
996 * Loop through twice for a priority level, if some class
997 * was unable to send a packet the first round because
998 * of the weighted round-robin mechanism.
999 * During the second loop at this level, deficit==2.
1000 * (This second loop is not needed if for every class,
1001 * "M[cl->pri_])" times "cl->allotment" is greater than
1002 * the byte size for the largest packet in the class.)
1005 cl = ifd->active_[cpri];
1008 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
1009 cl->bytes_alloc_ += cl->w_allotment_;
1010 if (!qempty(cl->q_)) {
1011 if ((cl->undertime_.tv_sec == 0) ||
1012 rmc_under_limit(cl, &now)) {
1013 if (cl->bytes_alloc_ > 0 || deficit > 1)
1016 /* underlimit but no alloc */
1019 ifd->borrowed_[ifd->qi_] = NULL;
1022 else if (first == NULL && cl->borrow_ != NULL)
1023 first = cl; /* borrowing candidate */
1026 cl->bytes_alloc_ = 0;
1028 } while (cl != ifd->active_[cpri]);
1031 /* first loop found an underlimit class with deficit */
1032 /* Loop on same priority level, with new deficit. */
1038 #ifdef ADJUST_CUTOFF
1040 * no underlimit class found. if cutoff is taking effect,
1041 * increase cutoff and try again.
1043 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1045 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1048 #endif /* ADJUST_CUTOFF */
1050 * If LINK_EFFICIENCY is turned on, then the first overlimit
1051 * class we encounter will send a packet if all the classes
1052 * of the link-sharing structure are overlimit.
1055 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1057 if (!ifd->efficient_ || first == NULL)
1062 #if 0 /* too time-consuming for nothing */
1064 CALLOUT_STOP(&cl->callout_);
1066 cl->undertime_.tv_sec = 0;
1068 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1069 ifd->cutoff_ = cl->borrow_->depth_;
1072 * Deque the packet and do the book keeping...
1075 if (op == ALTDQ_REMOVE) {
1078 panic("_rmc_wrr_dequeue_next");
1083 * Update class statistics and link data.
1085 if (cl->bytes_alloc_ > 0)
1086 cl->bytes_alloc_ -= m_pktlen(m);
1088 if ((cl->bytes_alloc_ <= 0) || first == cl)
1089 ifd->active_[cl->pri_] = cl->peer_;
1091 ifd->active_[cl->pri_] = cl;
1093 ifd->class_[ifd->qi_] = cl;
1094 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1095 ifd->now_[ifd->qi_] = now;
1096 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1099 /* mode == ALTDQ_PPOLL */
1101 ifd->pollcache_ = cl;
1107 * Dequeue & return next packet from the highest priority class that
1108 * has a packet to send & has enough allocation to send it. This
1109 * routine is called by a driver whenever it needs a new packet to
1113 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1117 struct rm_class *cl, *first = NULL;
1123 * if the driver polls the top of the queue and then removes
1124 * the polled packet, we must return the same packet.
1126 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1127 cl = ifd->pollcache_;
1129 ifd->pollcache_ = NULL;
1132 /* mode == ALTDQ_POLL || pollcache == NULL */
1133 ifd->pollcache_ = NULL;
1134 ifd->borrowed_[ifd->qi_] = NULL;
1136 #ifdef ADJUST_CUTOFF
1139 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1140 if (ifd->na_[cpri] == 0)
1142 cl = ifd->active_[cpri];
1145 if (!qempty(cl->q_)) {
1146 if ((cl->undertime_.tv_sec == 0) ||
1147 rmc_under_limit(cl, &now))
1149 if (first == NULL && cl->borrow_ != NULL)
1153 } while (cl != ifd->active_[cpri]);
1156 #ifdef ADJUST_CUTOFF
1158 * no underlimit class found. if cutoff is taking effect, increase
1159 * cutoff and try again.
1161 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1165 #endif /* ADJUST_CUTOFF */
1167 * If LINK_EFFICIENCY is turned on, then the first overlimit
1168 * class we encounter will send a packet if all the classes
1169 * of the link-sharing structure are overlimit.
1172 if (!ifd->efficient_ || first == NULL)
1177 #if 0 /* too time-consuming for nothing */
1179 CALLOUT_STOP(&cl->callout_);
1181 cl->undertime_.tv_sec = 0;
1183 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1184 ifd->cutoff_ = cl->borrow_->depth_;
1187 * Deque the packet and do the book keeping...
1190 if (op == ALTDQ_REMOVE) {
1193 panic("_rmc_prr_dequeue_next");
1197 ifd->active_[cpri] = cl->peer_;
1199 ifd->class_[ifd->qi_] = cl;
1200 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1201 ifd->now_[ifd->qi_] = now;
1202 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1205 /* mode == ALTDQ_POLL */
1207 ifd->pollcache_ = cl;
1214 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1215 * is invoked by the packet driver to get the next packet to be
1216 * dequeued and output on the link. If WRR is enabled, then the
1217 * WRR dequeue next routine will determine the next packet to sent.
1218 * Otherwise, packet-by-packet round robin is invoked.
1220 * Returns: NULL, if a packet is not available or if all
1221 * classes are overlimit.
1223 * Otherwise, Pointer to the next packet.
1227 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1229 if (ifd->queued_ >= ifd->maxqueued_)
1232 return (_rmc_wrr_dequeue_next(ifd, mode));
1234 return (_rmc_prr_dequeue_next(ifd, mode));
1238 * Update the utilization estimate for the packet that just completed.
1239 * The packet's class & the parent(s) of that class all get their
1240 * estimators updated. This routine is called by the driver's output-
1241 * packet-completion interrupt service routine.
1245 * a macro to approximate "divide by 1000" that gives 0.000999,
1246 * if a value has enough effective digits.
1247 * (on pentium, mul takes 9 cycles but div takes 46!)
1249 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1251 rmc_update_class_util(struct rm_ifdat *ifd)
1253 int idle, avgidle, pktlen;
1254 int pkt_time, tidle;
1255 rm_class_t *cl, *borrowed;
1256 rm_class_t *borrows;
1257 struct timeval *nowp;
1260 * Get the most recent completed class.
1262 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1265 pktlen = ifd->curlen_[ifd->qo_];
1266 borrowed = ifd->borrowed_[ifd->qo_];
1269 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1272 * Run estimator on class and its ancestors.
1275 * rm_update_class_util is designed to be called when the
1276 * transfer is completed from a xmit complete interrupt,
1277 * but most drivers don't implement an upcall for that.
1278 * so, just use estimated completion time.
1279 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1281 nowp = &ifd->now_[ifd->qo_];
1282 /* get pkt_time (for link) in usec */
1283 #if 1 /* use approximation */
1284 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1285 pkt_time = NSEC_TO_USEC(pkt_time);
1287 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1289 #if 1 /* ALTQ4PPP */
1290 if (TV_LT(nowp, &ifd->ifnow_)) {
1294 * make sure the estimated completion time does not go
1295 * too far. it can happen when the link layer supports
1296 * data compression or the interface speed is set to
1297 * a much lower value.
1299 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1300 if (iftime+pkt_time < ifd->maxiftime_) {
1301 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1303 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1306 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1309 if (TV_LT(nowp, &ifd->ifnow_)) {
1310 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1312 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1316 while (cl != NULL) {
1317 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1318 if (idle >= 2000000)
1320 * this class is idle enough, reset avgidle.
1321 * (TV_DELTA returns 2000000 us when delta is large.)
1323 cl->avgidle_ = cl->maxidle_;
1325 /* get pkt_time (for class) in usec */
1326 #if 1 /* use approximation */
1327 pkt_time = pktlen * cl->ns_per_byte_;
1328 pkt_time = NSEC_TO_USEC(pkt_time);
1330 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1334 avgidle = cl->avgidle_;
1335 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1336 cl->avgidle_ = avgidle;
1338 /* Are we overlimit ? */
1340 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1343 * need some lower bound for avgidle, otherwise
1344 * a borrowing class gets unbounded penalty.
1346 if (avgidle < cl->minidle_)
1347 avgidle = cl->avgidle_ = cl->minidle_;
1349 /* set next idle to make avgidle 0 */
1351 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1352 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1356 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1357 cl->undertime_.tv_sec = 0;
1358 if (cl->sleeping_) {
1359 CALLOUT_STOP(&cl->callout_);
1364 if (borrows != NULL) {
1366 ++cl->stats_.borrows;
1370 cl->last_ = ifd->ifnow_;
1371 cl->last_pkttime_ = pkt_time;
1374 if (cl->parent_ == NULL) {
1375 /* take stats of root class */
1376 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1384 * Check to see if cutoff needs to set to a new level.
1386 cl = ifd->class_[ifd->qo_];
1387 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1389 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1390 rmc_tl_satisfied(ifd, nowp);
1391 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1393 ifd->cutoff_ = borrowed->depth_;
1394 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1397 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1400 rmc_tl_satisfied(ifd, &now);
1402 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1404 ifd->cutoff_ = borrowed->depth_;
1405 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1411 * Release class slot
1413 ifd->borrowed_[ifd->qo_] = NULL;
1414 ifd->class_[ifd->qo_] = NULL;
1415 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1421 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1422 * over-limit action routines. These get invoked by rmc_under_limit()
1423 * if a class with packets to send if over its bandwidth limit & can't
1424 * borrow from a parent class.
1430 rmc_drop_action(struct rm_class *cl)
1432 struct rm_ifdat *ifd = cl->ifdat_;
1434 ASSERT(qlen(cl->q_) > 0);
1437 ifd->na_[cl->pri_]--;
1440 void rmc_dropall(struct rm_class *cl)
1442 struct rm_ifdat *ifd = cl->ifdat_;
1444 if (!qempty(cl->q_)) {
1447 ifd->na_[cl->pri_]--;
1451 #if (__FreeBSD_version > 300000)
1452 /* hzto() is removed from FreeBSD-3.0 */
1453 static int hzto(struct timeval *);
1462 t2.tv_sec = tv->tv_sec - t2.tv_sec;
1463 t2.tv_usec = tv->tv_usec - t2.tv_usec;
1464 return (tvtohz(&t2));
1466 #endif /* __FreeBSD_version > 300000 */
1470 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1471 * delay action routine. It is invoked via rmc_under_limit when the
1472 * packet is discoverd to be overlimit.
1474 * If the delay action is result of borrow class being overlimit, then
1475 * delay for the offtime of the borrowing class that is overlimit.
1481 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1483 int delay, t, extradelay;
1485 cl->stats_.overactions++;
1486 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1487 #ifndef BORROW_OFFTIME
1488 delay += cl->offtime_;
1491 if (!cl->sleeping_) {
1492 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1493 #ifdef BORROW_OFFTIME
1495 extradelay = borrow->offtime_;
1498 extradelay = cl->offtime_;
1502 * XXX recalculate suspend time:
1503 * current undertime is (tidle + pkt_time) calculated
1504 * from the last transmission.
1505 * tidle: time required to bring avgidle back to 0
1506 * pkt_time: target waiting time for this class
1507 * we need to replace pkt_time by offtime
1509 extradelay -= cl->last_pkttime_;
1511 if (extradelay > 0) {
1512 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1513 delay += extradelay;
1517 cl->stats_.delays++;
1520 * Since packets are phased randomly with respect to the
1521 * clock, 1 tick (the next clock tick) can be an arbitrarily
1522 * short time so we have to wait for at least two ticks.
1523 * NOTE: If there's no other traffic, we need the timer as
1524 * a 'backstop' to restart this class.
1526 if (delay > tick * 2) {
1527 /* FreeBSD rounds up the tick */
1528 t = hzto(&cl->undertime_);
1531 CALLOUT_RESET(&cl->callout_, t, rmc_restart, cl);
1537 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1538 * called by the system timer code & is responsible checking if the
1539 * class is still sleeping (it might have been restarted as a side
1540 * effect of the queue scan on a packet arrival) and, if so, restarting
1541 * output for the class. Inspecting the class state & restarting output
1542 * require locking the class structure. In general the driver is
1543 * responsible for locking but this is the only routine that is not
1544 * called directly or indirectly from the interface driver so it has
1545 * know about system locking conventions. Under bsd, locking is done
1546 * by raising IPL to splimp so that's what's implemented here. On a
1547 * different system this would probably need to be changed.
1553 rmc_restart(void *arg)
1555 struct rm_class *cl = arg;
1556 struct rm_ifdat *ifd = cl->ifdat_;
1560 IFQ_LOCK(ifd->ifq_);
1561 if (cl->sleeping_) {
1563 cl->undertime_.tv_sec = 0;
1565 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1566 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1567 (ifd->restart)(ifd->ifq_);
1570 IFQ_UNLOCK(ifd->ifq_);
1576 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1577 * handling routine for the root class of the link sharing structure.
1583 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1585 panic("rmc_root_overlimit");
1589 * Packet Queue handling routines. Eventually, this is to localize the
1590 * effects on the code whether queues are red queues or droptail
1595 _rmc_addq(rm_class_t *cl, mbuf_t *m)
1598 if (q_is_rio(cl->q_))
1599 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1602 if (q_is_red(cl->q_))
1603 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1604 #endif /* ALTQ_RED */
1606 if (q_is_codel(cl->q_))
1607 return codel_addq(cl->codel_, cl->q_, m);
1610 if (cl->flags_ & RMCF_CLEARDSCP)
1611 write_dsfield(m, cl->pktattr_, 0);
1617 /* note: _rmc_dropq is not called for red */
1619 _rmc_dropq(rm_class_t *cl)
1623 if ((m = _getq(cl->q_)) != NULL)
1628 _rmc_getq(rm_class_t *cl)
1631 if (q_is_rio(cl->q_))
1632 return rio_getq((rio_t *)cl->red_, cl->q_);
1635 if (q_is_red(cl->q_))
1636 return red_getq(cl->red_, cl->q_);
1639 if (q_is_codel(cl->q_))
1640 return codel_getq(cl->codel_, cl->q_);
1642 return _getq(cl->q_);
1646 _rmc_pollq(rm_class_t *cl)
1648 return qhead(cl->q_);
1653 struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
1654 struct cbqtrace *cbqtrace_ptr = NULL;
1658 * DDB hook to trace cbq events:
1659 * the last 1024 events are held in a circular buffer.
1660 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1662 void cbqtrace_dump(int);
1663 static char *rmc_funcname(void *);
1665 static struct rmc_funcs {
1670 rmc_init, "rmc_init",
1671 rmc_queue_packet, "rmc_queue_packet",
1672 rmc_under_limit, "rmc_under_limit",
1673 rmc_update_class_util, "rmc_update_class_util",
1674 rmc_delay_action, "rmc_delay_action",
1675 rmc_restart, "rmc_restart",
1676 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1680 static char *rmc_funcname(void *func)
1682 struct rmc_funcs *fp;
1684 for (fp = rmc_funcs; fp->func != NULL; fp++)
1685 if (fp->func == func)
1690 void cbqtrace_dump(int counter)
1695 counter = counter % NCBQTRACE;
1696 p = (int *)&cbqtrace_buffer[counter];
1698 for (i=0; i<20; i++) {
1699 printf("[0x%x] ", *p++);
1700 printf("%s: ", rmc_funcname((void *)*p++));
1702 printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1703 printf("%d\n",*p++);
1705 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1706 p = (int *)cbqtrace_buffer;
1709 #endif /* CBQ_TRACE */
1710 #endif /* ALTQ_CBQ */
1712 #if defined(ALTQ_CBQ) || defined(ALTQ_RED) || defined(ALTQ_RIO) || \
1713 defined(ALTQ_HFSC) || defined(ALTQ_PRIQ) || defined(ALTQ_CODEL)
1714 #if !defined(__GNUC__) || defined(ALTQ_DEBUG)
1717 _addq(class_queue_t *q, mbuf_t *m)
1721 if ((m0 = qtail(q)) != NULL)
1722 m->m_nextpkt = m0->m_nextpkt;
1731 _getq(class_queue_t *q)
1735 if ((m = qtail(q)) == NULL)
1737 if ((m0 = m->m_nextpkt) != m)
1738 m->m_nextpkt = m0->m_nextpkt;
1740 ASSERT(qlen(q) == 1);
1744 m0->m_nextpkt = NULL;
1748 /* drop a packet at the tail of the queue */
1750 _getq_tail(class_queue_t *q)
1752 mbuf_t *m, *m0, *prev;
1754 if ((m = m0 = qtail(q)) == NULL)
1760 prev->m_nextpkt = m->m_nextpkt;
1762 ASSERT(qlen(q) == 1);
1767 m->m_nextpkt = NULL;
1771 /* randomly select a packet in the queue */
1773 _getq_random(class_queue_t *q)
1778 if ((m = qtail(q)) == NULL)
1780 if (m->m_nextpkt == m) {
1781 ASSERT(qlen(q) == 1);
1784 struct mbuf *prev = NULL;
1786 n = arc4random() % qlen(q) + 1;
1787 for (i = 0; i < n; i++) {
1791 prev->m_nextpkt = m->m_nextpkt;
1796 m->m_nextpkt = NULL;
1801 _removeq(class_queue_t *q, mbuf_t *m)
1810 prev->m_nextpkt = m->m_nextpkt;
1813 else if (qtail(q) == m)
1819 _flushq(class_queue_t *q)
1823 while ((m = _getq(q)) != NULL)
1825 ASSERT(qlen(q) == 0);
1828 #endif /* !__GNUC__ || ALTQ_DEBUG */
1829 #endif /* ALTQ_CBQ || ALTQ_RED || ALTQ_RIO || ALTQ_HFSC || ALTQ_PRIQ */