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>
53 #include <sys/kernel.h>
57 #include <net/if_var.h>
59 #include <netinet/in.h>
60 #include <netinet/in_systm.h>
61 #include <netinet/ip.h>
64 #include <net/altq/if_altq.h>
65 #include <net/altq/altq.h>
66 #include <net/altq/altq_rmclass.h>
67 #include <net/altq/altq_rmclass_debug.h>
68 #include <net/altq/altq_red.h>
69 #include <net/altq/altq_rio.h>
75 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
81 static int rmc_satisfied(struct rm_class *, struct timeval *);
82 static void rmc_wrr_set_weights(struct rm_ifdat *);
83 static void rmc_depth_compute(struct rm_class *);
84 static void rmc_depth_recompute(rm_class_t *);
86 static mbuf_t *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
87 static mbuf_t *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
89 static int _rmc_addq(rm_class_t *, mbuf_t *);
90 static void _rmc_dropq(rm_class_t *);
91 static mbuf_t *_rmc_getq(rm_class_t *);
92 static mbuf_t *_rmc_pollq(rm_class_t *);
94 static int rmc_under_limit(struct rm_class *, struct timeval *);
95 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
96 static void rmc_drop_action(struct rm_class *);
97 static void rmc_restart(struct rm_class *);
98 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
100 #define BORROW_OFFTIME
102 * BORROW_OFFTIME (experimental):
103 * borrow the offtime of the class borrowing from.
104 * the reason is that when its own offtime is set, the class is unable
105 * to borrow much, especially when cutoff is taking effect.
106 * but when the borrowed class is overloaded (advidle is close to minidle),
107 * use the borrowing class's offtime to avoid overload.
109 #define ADJUST_CUTOFF
111 * ADJUST_CUTOFF (experimental):
112 * if no underlimit class is found due to cutoff, increase cutoff and
113 * retry the scheduling loop.
114 * also, don't invoke delay_actions while cutoff is taking effect,
115 * since a sleeping class won't have a chance to be scheduled in the
118 * now heuristics for setting the top-level variable (cutoff_) becomes:
119 * 1. if a packet arrives for a not-overlimit class, set cutoff
120 * to the depth of the class.
121 * 2. if cutoff is i, and a packet arrives for an overlimit class
122 * with an underlimit ancestor at a lower level than i (say j),
123 * then set cutoff to j.
124 * 3. at scheduling a packet, if there is no underlimit class
125 * due to the current cutoff level, increase cutoff by 1 and
126 * then try to schedule again.
131 * rmc_newclass(...) - Create a new resource management class at priority
132 * 'pri' on the interface given by 'ifd'.
134 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
135 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
136 * than 100% of the bandwidth, this number should be the
137 * 'effective' rate for the class. Let f be the
138 * bandwidth fraction allocated to this class, and let
139 * nsPerByte be the data rate of the output link in
140 * nanoseconds/byte. Then nsecPerByte is set to
141 * nsPerByte / f. E.g., 1600 (= 800 / .5)
142 * for a class that gets 50% of an ethernet's bandwidth.
144 * action the routine to call when the class is over limit.
146 * maxq max allowable queue size for class (in packets).
148 * parent parent class pointer.
150 * borrow class to borrow from (should be either 'parent' or null).
152 * maxidle max value allowed for class 'idle' time estimate (this
153 * parameter determines how large an initial burst of packets
154 * can be before overlimit action is invoked.
156 * offtime how long 'delay' action will delay when class goes over
157 * limit (this parameter determines the steady-state burst
158 * size when a class is running over its limit).
160 * Maxidle and offtime have to be computed from the following: If the
161 * average packet size is s, the bandwidth fraction allocated to this
162 * class is f, we want to allow b packet bursts, and the gain of the
163 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
165 * ptime = s * nsPerByte * (1 - f) / f
166 * maxidle = ptime * (1 - g^b) / g^b
167 * minidle = -ptime * (1 / (f - 1))
168 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
170 * Operationally, it's convenient to specify maxidle & offtime in units
171 * independent of the link bandwidth so the maxidle & offtime passed to
172 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
173 * (The constant factor is a scale factor needed to make the parameters
174 * integers. This scaling also means that the 'unscaled' values of
175 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
176 * not nanoseconds.) Also note that the 'idle' filter computation keeps
177 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
178 * maxidle also must be scaled upward by this value. Thus, the passed
179 * values for maxidle and offtime can be computed as follows:
181 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
182 * offtime = offtime * 8 / (1000 * nsecPerByte)
184 * When USE_HRTIME is employed, then maxidle and offtime become:
185 * maxidle = maxilde * (8.0 / nsecPerByte);
186 * offtime = offtime * (8.0 / nsecPerByte);
189 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
190 void (*action)(rm_class_t *, rm_class_t *), int maxq,
191 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
192 int minidle, u_int offtime, int pktsize, int flags)
195 struct rm_class *peer;
198 if (pri >= RM_MAXPRIO)
201 if (flags & RMCF_RED) {
203 printf("rmc_newclass: RED not configured for CBQ!\n");
209 if (flags & RMCF_RIO) {
211 printf("rmc_newclass: RIO 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 */
302 * put the class into the class tree
306 if ((peer = ifd->active_[pri]) != NULL) {
307 /* find the last class at this pri */
309 while (peer->peer_ != ifd->active_[pri])
313 ifd->active_[pri] = cl;
318 cl->next_ = parent->children_;
319 parent->children_ = cl;
324 * Compute the depth of this class and its ancestors in the class
327 rmc_depth_compute(cl);
330 * If CBQ's WRR is enabled, then initialize the class WRR state.
334 ifd->alloc_[pri] += cl->allotment_;
335 rmc_wrr_set_weights(ifd);
337 IFQ_UNLOCK(ifd->ifq_);
343 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
344 int minidle, u_int offtime, int pktsize)
346 struct rm_ifdat *ifd;
351 old_allotment = cl->allotment_;
355 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
357 cl->ns_per_byte_ = nsecPerByte;
359 qlimit(cl->q_) = maxq;
361 #if 1 /* minidle is also scaled in ALTQ */
362 cl->minidle_ = (minidle * nsecPerByte) / 8;
363 if (cl->minidle_ > 0)
366 cl->minidle_ = minidle;
368 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
369 if (cl->maxidle_ == 0)
371 #if 1 /* offtime is also scaled in ALTQ */
372 cl->avgidle_ = cl->maxidle_;
373 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
374 if (cl->offtime_ == 0)
378 cl->offtime_ = (offtime * nsecPerByte) / 8;
382 * If CBQ's WRR is enabled, then initialize the class WRR state.
385 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
386 rmc_wrr_set_weights(ifd);
388 IFQ_UNLOCK(ifd->ifq_);
395 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
396 * the appropriate run robin weights for the CBQ weighted round robin
403 rmc_wrr_set_weights(struct rm_ifdat *ifd)
406 struct rm_class *cl, *clh;
408 for (i = 0; i < RM_MAXPRIO; i++) {
410 * This is inverted from that of the simulator to
411 * maintain precision.
413 if (ifd->num_[i] == 0)
416 ifd->M_[i] = ifd->alloc_[i] /
417 (ifd->num_[i] * ifd->maxpkt_);
419 * Compute the weighted allotment for each class.
420 * This takes the expensive div instruction out
421 * of the main loop for the wrr scheduling path.
422 * These only get recomputed when a class comes or
425 if (ifd->active_[i] != NULL) {
426 clh = cl = ifd->active_[i];
428 /* safe-guard for slow link or alloc_ == 0 */
430 cl->w_allotment_ = 0;
432 cl->w_allotment_ = cl->allotment_ /
435 } while ((cl != NULL) && (cl != clh));
441 rmc_get_weight(struct rm_ifdat *ifd, int pri)
443 if ((pri >= 0) && (pri < RM_MAXPRIO))
444 return (ifd->M_[pri]);
451 * rmc_depth_compute(struct rm_class *cl) - This function computes the
452 * appropriate depth of class 'cl' and its ancestors.
458 rmc_depth_compute(struct rm_class *cl)
460 rm_class_t *t = cl, *p;
463 * Recompute the depth for the branch of the tree.
467 if (p && (t->depth_ >= p->depth_)) {
468 p->depth_ = t->depth_ + 1;
477 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
478 * the depth of the tree after a class has been deleted.
484 rmc_depth_recompute(rm_class_t *cl)
491 if ((t = p->children_) == NULL) {
497 if (t->depth_ > cdepth)
502 if (p->depth_ == cdepth + 1)
503 /* no change to this parent */
506 p->depth_ = cdepth + 1;
514 if (cl->depth_ >= 1) {
515 if (cl->children_ == NULL) {
517 } else if ((t = cl->children_) != NULL) {
519 if (t->children_ != NULL)
520 rmc_depth_recompute(t);
524 rmc_depth_compute(cl);
531 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
532 * function deletes a class from the link-sharing structure and frees
533 * all resources associated with the class.
539 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
541 struct rm_class *p, *head, *previous;
544 ASSERT(cl->children_ == NULL);
547 CALLOUT_STOP(&cl->callout_);
552 * Free packets in the packet queue.
553 * XXX - this may not be a desired behavior. Packets should be
559 * If the class has a parent, then remove the class from the
560 * class from the parent's children chain.
562 if (cl->parent_ != NULL) {
563 head = cl->parent_->children_;
565 if (head->next_ == NULL) {
567 cl->parent_->children_ = NULL;
568 cl->parent_->leaf_ = 1;
569 } else while (p != NULL) {
572 cl->parent_->children_ = cl->next_;
574 previous->next_ = cl->next_;
585 * Delete class from class priority peer list.
587 if ((p = ifd->active_[cl->pri_]) != NULL) {
589 * If there is more than one member of this priority
590 * level, then look for class(cl) in the priority level.
593 while (p->peer_ != cl)
595 p->peer_ = cl->peer_;
597 if (ifd->active_[cl->pri_] == cl)
598 ifd->active_[cl->pri_] = cl->peer_;
601 ifd->active_[cl->pri_] = NULL;
606 * Recompute the WRR weights.
609 ifd->alloc_[cl->pri_] -= cl->allotment_;
610 ifd->num_[cl->pri_]--;
611 rmc_wrr_set_weights(ifd);
615 * Re-compute the depth of the tree.
618 rmc_depth_recompute(cl->parent_);
620 rmc_depth_recompute(ifd->root_);
623 IFQ_UNLOCK(ifd->ifq_);
627 * Free the class structure.
629 if (cl->red_ != NULL) {
631 if (q_is_rio(cl->q_))
632 rio_destroy((rio_t *)cl->red_);
635 if (q_is_red(cl->q_))
636 red_destroy(cl->red_);
639 free(cl->q_, M_DEVBUF);
646 * rmc_init(...) - Initialize the resource management data structures
647 * associated with the output portion of interface 'ifp'. 'ifd' is
648 * where the structures will be built (for backwards compatibility, the
649 * structures aren't kept in the ifnet struct). 'nsecPerByte'
650 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
651 * 'restart' is the driver-specific routine that the generic 'delay
652 * until under limit' action will call to restart output. `maxq'
653 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
654 * is the maximum number of packets that the resource management
655 * code will allow to be queued 'downstream' (this is typically 1).
661 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
662 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
663 int minidle, u_int offtime, int flags)
668 * Initialize the CBQ tracing/debug facility.
672 bzero((char *)ifd, sizeof (*ifd));
673 mtu = ifq->altq_ifp->if_mtu;
675 ifd->restart = restart;
676 ifd->maxqueued_ = maxqueued;
677 ifd->ns_per_byte_ = nsecPerByte;
679 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
680 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
682 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
683 if (mtu * nsecPerByte > 10 * 1000000)
684 ifd->maxiftime_ /= 4;
688 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
691 * Initialize the CBQ's WRR state.
693 for (i = 0; i < RM_MAXPRIO; i++) {
698 ifd->active_[i] = NULL;
702 * Initialize current packet state.
706 for (i = 0; i < RM_MAXQUEUED; i++) {
707 ifd->class_[i] = NULL;
709 ifd->borrowed_[i] = NULL;
713 * Create the root class of the link-sharing structure.
715 if ((ifd->root_ = rmc_newclass(0, ifd,
717 rmc_root_overlimit, maxq, 0, 0,
718 maxidle, minidle, offtime,
720 printf("rmc_init: root class not allocated\n");
723 ifd->root_->depth_ = 0;
728 * rmc_queue_packet(struct rm_class *cl, mbuf_t *m) - Add packet given by
729 * mbuf 'm' to queue for resource class 'cl'. This routine is called
730 * by a driver's if_output routine. This routine must be called with
731 * output packet completion interrupts locked out (to avoid racing with
734 * Returns: 0 on successful queueing
735 * -1 when packet drop occurs
738 rmc_queue_packet(struct rm_class *cl, mbuf_t *m)
741 struct rm_ifdat *ifd = cl->ifdat_;
743 int is_empty = qempty(cl->q_);
746 if (ifd->cutoff_ > 0) {
747 if (TV_LT(&cl->undertime_, &now)) {
748 if (ifd->cutoff_ > cl->depth_)
749 ifd->cutoff_ = cl->depth_;
750 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
755 * the class is overlimit. if the class has
756 * underlimit ancestors, set cutoff to the lowest
759 struct rm_class *borrow = cl->borrow_;
761 while (borrow != NULL &&
762 borrow->depth_ < ifd->cutoff_) {
763 if (TV_LT(&borrow->undertime_, &now)) {
764 ifd->cutoff_ = borrow->depth_;
765 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
768 borrow = borrow->borrow_;
772 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
773 if (TV_LT(&cl->borrow_->undertime_, &now)) {
774 ifd->cutoff_ = cl->borrow_->depth_;
775 CBQTRACE(rmc_queue_packet, 'ffob',
776 cl->borrow_->depth_);
782 if (_rmc_addq(cl, m) < 0)
787 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
791 if (qlen(cl->q_) > qlimit(cl->q_)) {
792 /* note: qlimit can be set to 0 or 1 */
801 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
802 * classes to see if there are satified.
806 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
811 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
812 if ((bp = ifd->active_[i]) != NULL) {
815 if (!rmc_satisfied(p, now)) {
816 ifd->cutoff_ = p->depth_;
828 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
832 rmc_satisfied(struct rm_class *cl, struct timeval *now)
838 if (TV_LT(now, &cl->undertime_))
840 if (cl->depth_ == 0) {
841 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
846 if (cl->children_ != NULL) {
849 if (!rmc_satisfied(p, now))
859 * Return 1 if class 'cl' is under limit or can borrow from a parent,
860 * 0 if overlimit. As a side-effect, this routine will invoke the
861 * class overlimit action if the class if overlimit.
865 rmc_under_limit(struct rm_class *cl, struct timeval *now)
869 struct rm_ifdat *ifd = cl->ifdat_;
871 ifd->borrowed_[ifd->qi_] = NULL;
873 * If cl is the root class, then always return that it is
874 * underlimit. Otherwise, check to see if the class is underlimit.
876 if (cl->parent_ == NULL)
880 if (TV_LT(now, &cl->undertime_))
883 CALLOUT_STOP(&cl->callout_);
885 cl->undertime_.tv_sec = 0;
890 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
891 if (((cl = cl->borrow_) == NULL) ||
892 (cl->depth_ > ifd->cutoff_)) {
895 /* cutoff is taking effect, just
896 return false without calling
900 #ifdef BORROW_OFFTIME
902 * check if the class can borrow offtime too.
903 * borrow offtime from the top of the borrow
904 * chain if the top class is not overloaded.
907 /* cutoff is taking effect, use this class as top. */
909 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
911 if (top != NULL && top->avgidle_ == top->minidle_)
914 (p->overlimit)(p, top);
917 (p->overlimit)(p, NULL);
925 ifd->borrowed_[ifd->qi_] = cl;
930 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
931 * Packet-by-packet round robin.
933 * The heart of the weighted round-robin scheduler, which decides which
934 * class next gets to send a packet. Highest priority first, then
935 * weighted round-robin within priorites.
937 * Each able-to-send class gets to send until its byte allocation is
938 * exhausted. Thus, the active pointer is only changed after a class has
939 * exhausted its allocation.
941 * If the scheduler finds no class that is underlimit or able to borrow,
942 * then the first class found that had a nonzero queue and is allowed to
943 * borrow gets to send.
947 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
949 struct rm_class *cl = NULL, *first = NULL;
958 * if the driver polls the top of the queue and then removes
959 * the polled packet, we must return the same packet.
961 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
962 cl = ifd->pollcache_;
964 if (ifd->efficient_) {
965 /* check if this class is overlimit */
966 if (cl->undertime_.tv_sec != 0 &&
967 rmc_under_limit(cl, &now) == 0)
970 ifd->pollcache_ = NULL;
974 /* mode == ALTDQ_POLL || pollcache == NULL */
975 ifd->pollcache_ = NULL;
976 ifd->borrowed_[ifd->qi_] = NULL;
981 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
982 if (ifd->na_[cpri] == 0)
986 * Loop through twice for a priority level, if some class
987 * was unable to send a packet the first round because
988 * of the weighted round-robin mechanism.
989 * During the second loop at this level, deficit==2.
990 * (This second loop is not needed if for every class,
991 * "M[cl->pri_])" times "cl->allotment" is greater than
992 * the byte size for the largest packet in the class.)
995 cl = ifd->active_[cpri];
998 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
999 cl->bytes_alloc_ += cl->w_allotment_;
1000 if (!qempty(cl->q_)) {
1001 if ((cl->undertime_.tv_sec == 0) ||
1002 rmc_under_limit(cl, &now)) {
1003 if (cl->bytes_alloc_ > 0 || deficit > 1)
1006 /* underlimit but no alloc */
1009 ifd->borrowed_[ifd->qi_] = NULL;
1012 else if (first == NULL && cl->borrow_ != NULL)
1013 first = cl; /* borrowing candidate */
1016 cl->bytes_alloc_ = 0;
1018 } while (cl != ifd->active_[cpri]);
1021 /* first loop found an underlimit class with deficit */
1022 /* Loop on same priority level, with new deficit. */
1028 #ifdef ADJUST_CUTOFF
1030 * no underlimit class found. if cutoff is taking effect,
1031 * increase cutoff and try again.
1033 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1035 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1038 #endif /* ADJUST_CUTOFF */
1040 * If LINK_EFFICIENCY is turned on, then the first overlimit
1041 * class we encounter will send a packet if all the classes
1042 * of the link-sharing structure are overlimit.
1045 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1047 if (!ifd->efficient_ || first == NULL)
1052 #if 0 /* too time-consuming for nothing */
1054 CALLOUT_STOP(&cl->callout_);
1056 cl->undertime_.tv_sec = 0;
1058 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1059 ifd->cutoff_ = cl->borrow_->depth_;
1062 * Deque the packet and do the book keeping...
1065 if (op == ALTDQ_REMOVE) {
1068 panic("_rmc_wrr_dequeue_next");
1073 * Update class statistics and link data.
1075 if (cl->bytes_alloc_ > 0)
1076 cl->bytes_alloc_ -= m_pktlen(m);
1078 if ((cl->bytes_alloc_ <= 0) || first == cl)
1079 ifd->active_[cl->pri_] = cl->peer_;
1081 ifd->active_[cl->pri_] = cl;
1083 ifd->class_[ifd->qi_] = cl;
1084 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1085 ifd->now_[ifd->qi_] = now;
1086 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1089 /* mode == ALTDQ_PPOLL */
1091 ifd->pollcache_ = cl;
1097 * Dequeue & return next packet from the highest priority class that
1098 * has a packet to send & has enough allocation to send it. This
1099 * routine is called by a driver whenever it needs a new packet to
1103 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1107 struct rm_class *cl, *first = NULL;
1113 * if the driver polls the top of the queue and then removes
1114 * the polled packet, we must return the same packet.
1116 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1117 cl = ifd->pollcache_;
1119 ifd->pollcache_ = NULL;
1122 /* mode == ALTDQ_POLL || pollcache == NULL */
1123 ifd->pollcache_ = NULL;
1124 ifd->borrowed_[ifd->qi_] = NULL;
1126 #ifdef ADJUST_CUTOFF
1129 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1130 if (ifd->na_[cpri] == 0)
1132 cl = ifd->active_[cpri];
1135 if (!qempty(cl->q_)) {
1136 if ((cl->undertime_.tv_sec == 0) ||
1137 rmc_under_limit(cl, &now))
1139 if (first == NULL && cl->borrow_ != NULL)
1143 } while (cl != ifd->active_[cpri]);
1146 #ifdef ADJUST_CUTOFF
1148 * no underlimit class found. if cutoff is taking effect, increase
1149 * cutoff and try again.
1151 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1155 #endif /* ADJUST_CUTOFF */
1157 * If LINK_EFFICIENCY is turned on, then the first overlimit
1158 * class we encounter will send a packet if all the classes
1159 * of the link-sharing structure are overlimit.
1162 if (!ifd->efficient_ || first == NULL)
1167 #if 0 /* too time-consuming for nothing */
1169 CALLOUT_STOP(&cl->callout_);
1171 cl->undertime_.tv_sec = 0;
1173 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1174 ifd->cutoff_ = cl->borrow_->depth_;
1177 * Deque the packet and do the book keeping...
1180 if (op == ALTDQ_REMOVE) {
1183 panic("_rmc_prr_dequeue_next");
1187 ifd->active_[cpri] = cl->peer_;
1189 ifd->class_[ifd->qi_] = cl;
1190 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1191 ifd->now_[ifd->qi_] = now;
1192 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1195 /* mode == ALTDQ_POLL */
1197 ifd->pollcache_ = cl;
1204 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1205 * is invoked by the packet driver to get the next packet to be
1206 * dequeued and output on the link. If WRR is enabled, then the
1207 * WRR dequeue next routine will determine the next packet to sent.
1208 * Otherwise, packet-by-packet round robin is invoked.
1210 * Returns: NULL, if a packet is not available or if all
1211 * classes are overlimit.
1213 * Otherwise, Pointer to the next packet.
1217 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1219 if (ifd->queued_ >= ifd->maxqueued_)
1222 return (_rmc_wrr_dequeue_next(ifd, mode));
1224 return (_rmc_prr_dequeue_next(ifd, mode));
1228 * Update the utilization estimate for the packet that just completed.
1229 * The packet's class & the parent(s) of that class all get their
1230 * estimators updated. This routine is called by the driver's output-
1231 * packet-completion interrupt service routine.
1235 * a macro to approximate "divide by 1000" that gives 0.000999,
1236 * if a value has enough effective digits.
1237 * (on pentium, mul takes 9 cycles but div takes 46!)
1239 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1241 rmc_update_class_util(struct rm_ifdat *ifd)
1243 int idle, avgidle, pktlen;
1244 int pkt_time, tidle;
1245 rm_class_t *cl, *borrowed;
1246 rm_class_t *borrows;
1247 struct timeval *nowp;
1250 * Get the most recent completed class.
1252 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1255 pktlen = ifd->curlen_[ifd->qo_];
1256 borrowed = ifd->borrowed_[ifd->qo_];
1259 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1262 * Run estimator on class and its ancestors.
1265 * rm_update_class_util is designed to be called when the
1266 * transfer is completed from a xmit complete interrupt,
1267 * but most drivers don't implement an upcall for that.
1268 * so, just use estimated completion time.
1269 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1271 nowp = &ifd->now_[ifd->qo_];
1272 /* get pkt_time (for link) in usec */
1273 #if 1 /* use approximation */
1274 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1275 pkt_time = NSEC_TO_USEC(pkt_time);
1277 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1279 #if 1 /* ALTQ4PPP */
1280 if (TV_LT(nowp, &ifd->ifnow_)) {
1284 * make sure the estimated completion time does not go
1285 * too far. it can happen when the link layer supports
1286 * data compression or the interface speed is set to
1287 * a much lower value.
1289 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1290 if (iftime+pkt_time < ifd->maxiftime_) {
1291 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1293 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1296 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1299 if (TV_LT(nowp, &ifd->ifnow_)) {
1300 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1302 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1306 while (cl != NULL) {
1307 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1308 if (idle >= 2000000)
1310 * this class is idle enough, reset avgidle.
1311 * (TV_DELTA returns 2000000 us when delta is large.)
1313 cl->avgidle_ = cl->maxidle_;
1315 /* get pkt_time (for class) in usec */
1316 #if 1 /* use approximation */
1317 pkt_time = pktlen * cl->ns_per_byte_;
1318 pkt_time = NSEC_TO_USEC(pkt_time);
1320 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1324 avgidle = cl->avgidle_;
1325 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1326 cl->avgidle_ = avgidle;
1328 /* Are we overlimit ? */
1330 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1333 * need some lower bound for avgidle, otherwise
1334 * a borrowing class gets unbounded penalty.
1336 if (avgidle < cl->minidle_)
1337 avgidle = cl->avgidle_ = cl->minidle_;
1339 /* set next idle to make avgidle 0 */
1341 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1342 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1346 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1347 cl->undertime_.tv_sec = 0;
1348 if (cl->sleeping_) {
1349 CALLOUT_STOP(&cl->callout_);
1354 if (borrows != NULL) {
1356 ++cl->stats_.borrows;
1360 cl->last_ = ifd->ifnow_;
1361 cl->last_pkttime_ = pkt_time;
1364 if (cl->parent_ == NULL) {
1365 /* take stats of root class */
1366 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1374 * Check to see if cutoff needs to set to a new level.
1376 cl = ifd->class_[ifd->qo_];
1377 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1379 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1380 rmc_tl_satisfied(ifd, nowp);
1381 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1383 ifd->cutoff_ = borrowed->depth_;
1384 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1387 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1390 rmc_tl_satisfied(ifd, &now);
1392 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1394 ifd->cutoff_ = borrowed->depth_;
1395 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1401 * Release class slot
1403 ifd->borrowed_[ifd->qo_] = NULL;
1404 ifd->class_[ifd->qo_] = NULL;
1405 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1411 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1412 * over-limit action routines. These get invoked by rmc_under_limit()
1413 * if a class with packets to send if over its bandwidth limit & can't
1414 * borrow from a parent class.
1420 rmc_drop_action(struct rm_class *cl)
1422 struct rm_ifdat *ifd = cl->ifdat_;
1424 ASSERT(qlen(cl->q_) > 0);
1427 ifd->na_[cl->pri_]--;
1430 void rmc_dropall(struct rm_class *cl)
1432 struct rm_ifdat *ifd = cl->ifdat_;
1434 if (!qempty(cl->q_)) {
1437 ifd->na_[cl->pri_]--;
1441 #if (__FreeBSD_version > 300000)
1442 /* hzto() is removed from FreeBSD-3.0 */
1443 static int hzto(struct timeval *);
1452 t2.tv_sec = tv->tv_sec - t2.tv_sec;
1453 t2.tv_usec = tv->tv_usec - t2.tv_usec;
1454 return (tvtohz(&t2));
1456 #endif /* __FreeBSD_version > 300000 */
1460 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1461 * delay action routine. It is invoked via rmc_under_limit when the
1462 * packet is discoverd to be overlimit.
1464 * If the delay action is result of borrow class being overlimit, then
1465 * delay for the offtime of the borrowing class that is overlimit.
1471 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1473 int delay, t, extradelay;
1475 cl->stats_.overactions++;
1476 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1477 #ifndef BORROW_OFFTIME
1478 delay += cl->offtime_;
1481 if (!cl->sleeping_) {
1482 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1483 #ifdef BORROW_OFFTIME
1485 extradelay = borrow->offtime_;
1488 extradelay = cl->offtime_;
1492 * XXX recalculate suspend time:
1493 * current undertime is (tidle + pkt_time) calculated
1494 * from the last transmission.
1495 * tidle: time required to bring avgidle back to 0
1496 * pkt_time: target waiting time for this class
1497 * we need to replace pkt_time by offtime
1499 extradelay -= cl->last_pkttime_;
1501 if (extradelay > 0) {
1502 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1503 delay += extradelay;
1507 cl->stats_.delays++;
1510 * Since packets are phased randomly with respect to the
1511 * clock, 1 tick (the next clock tick) can be an arbitrarily
1512 * short time so we have to wait for at least two ticks.
1513 * NOTE: If there's no other traffic, we need the timer as
1514 * a 'backstop' to restart this class.
1516 if (delay > tick * 2) {
1517 /* FreeBSD rounds up the tick */
1518 t = hzto(&cl->undertime_);
1521 CALLOUT_RESET(&cl->callout_, t,
1522 (timeout_t *)rmc_restart, (caddr_t)cl);
1528 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1529 * called by the system timer code & is responsible checking if the
1530 * class is still sleeping (it might have been restarted as a side
1531 * effect of the queue scan on a packet arrival) and, if so, restarting
1532 * output for the class. Inspecting the class state & restarting output
1533 * require locking the class structure. In general the driver is
1534 * responsible for locking but this is the only routine that is not
1535 * called directly or indirectly from the interface driver so it has
1536 * know about system locking conventions. Under bsd, locking is done
1537 * by raising IPL to splimp so that's what's implemented here. On a
1538 * different system this would probably need to be changed.
1544 rmc_restart(struct rm_class *cl)
1546 struct rm_ifdat *ifd = cl->ifdat_;
1550 IFQ_LOCK(ifd->ifq_);
1551 if (cl->sleeping_) {
1553 cl->undertime_.tv_sec = 0;
1555 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1556 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1557 (ifd->restart)(ifd->ifq_);
1560 IFQ_UNLOCK(ifd->ifq_);
1566 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1567 * handling routine for the root class of the link sharing structure.
1573 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1575 panic("rmc_root_overlimit");
1579 * Packet Queue handling routines. Eventually, this is to localize the
1580 * effects on the code whether queues are red queues or droptail
1585 _rmc_addq(rm_class_t *cl, mbuf_t *m)
1588 if (q_is_rio(cl->q_))
1589 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1592 if (q_is_red(cl->q_))
1593 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1594 #endif /* ALTQ_RED */
1596 if (cl->flags_ & RMCF_CLEARDSCP)
1597 write_dsfield(m, cl->pktattr_, 0);
1603 /* note: _rmc_dropq is not called for red */
1605 _rmc_dropq(rm_class_t *cl)
1609 if ((m = _getq(cl->q_)) != NULL)
1614 _rmc_getq(rm_class_t *cl)
1617 if (q_is_rio(cl->q_))
1618 return rio_getq((rio_t *)cl->red_, cl->q_);
1621 if (q_is_red(cl->q_))
1622 return red_getq(cl->red_, cl->q_);
1624 return _getq(cl->q_);
1628 _rmc_pollq(rm_class_t *cl)
1630 return qhead(cl->q_);
1635 struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
1636 struct cbqtrace *cbqtrace_ptr = NULL;
1640 * DDB hook to trace cbq events:
1641 * the last 1024 events are held in a circular buffer.
1642 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1644 void cbqtrace_dump(int);
1645 static char *rmc_funcname(void *);
1647 static struct rmc_funcs {
1652 rmc_init, "rmc_init",
1653 rmc_queue_packet, "rmc_queue_packet",
1654 rmc_under_limit, "rmc_under_limit",
1655 rmc_update_class_util, "rmc_update_class_util",
1656 rmc_delay_action, "rmc_delay_action",
1657 rmc_restart, "rmc_restart",
1658 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1662 static char *rmc_funcname(void *func)
1664 struct rmc_funcs *fp;
1666 for (fp = rmc_funcs; fp->func != NULL; fp++)
1667 if (fp->func == func)
1672 void cbqtrace_dump(int counter)
1677 counter = counter % NCBQTRACE;
1678 p = (int *)&cbqtrace_buffer[counter];
1680 for (i=0; i<20; i++) {
1681 printf("[0x%x] ", *p++);
1682 printf("%s: ", rmc_funcname((void *)*p++));
1684 printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1685 printf("%d\n",*p++);
1687 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1688 p = (int *)cbqtrace_buffer;
1691 #endif /* CBQ_TRACE */
1692 #endif /* ALTQ_CBQ */
1694 #if defined(ALTQ_CBQ) || defined(ALTQ_RED) || defined(ALTQ_RIO) || defined(ALTQ_HFSC) || defined(ALTQ_PRIQ)
1695 #if !defined(__GNUC__) || defined(ALTQ_DEBUG)
1698 _addq(class_queue_t *q, mbuf_t *m)
1702 if ((m0 = qtail(q)) != NULL)
1703 m->m_nextpkt = m0->m_nextpkt;
1712 _getq(class_queue_t *q)
1716 if ((m = qtail(q)) == NULL)
1718 if ((m0 = m->m_nextpkt) != m)
1719 m->m_nextpkt = m0->m_nextpkt;
1721 ASSERT(qlen(q) == 1);
1725 m0->m_nextpkt = NULL;
1729 /* drop a packet at the tail of the queue */
1731 _getq_tail(class_queue_t *q)
1733 mbuf_t *m, *m0, *prev;
1735 if ((m = m0 = qtail(q)) == NULL)
1741 prev->m_nextpkt = m->m_nextpkt;
1743 ASSERT(qlen(q) == 1);
1748 m->m_nextpkt = NULL;
1752 /* randomly select a packet in the queue */
1754 _getq_random(class_queue_t *q)
1759 if ((m = qtail(q)) == NULL)
1761 if (m->m_nextpkt == m) {
1762 ASSERT(qlen(q) == 1);
1765 struct mbuf *prev = NULL;
1767 n = arc4random() % qlen(q) + 1;
1768 for (i = 0; i < n; i++) {
1772 prev->m_nextpkt = m->m_nextpkt;
1777 m->m_nextpkt = NULL;
1782 _removeq(class_queue_t *q, mbuf_t *m)
1791 prev->m_nextpkt = m->m_nextpkt;
1794 else if (qtail(q) == m)
1800 _flushq(class_queue_t *q)
1804 while ((m = _getq(q)) != NULL)
1806 ASSERT(qlen(q) == 0);
1809 #endif /* !__GNUC__ || ALTQ_DEBUG */
1810 #endif /* ALTQ_CBQ || ALTQ_RED || ALTQ_RIO || ALTQ_HFSC || ALTQ_PRIQ */