1 /* @(#)rm_class.c 1.48 97/12/05 SMI */
2 /* $KAME: altq_rmclass.c,v 1.18 2003/11/06 06:32:53 kjc 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
42 #include "opt_inet6.h"
44 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
46 #include <sys/param.h>
47 #include <sys/malloc.h>
49 #include <sys/socket.h>
50 #include <sys/systm.h>
51 #include <sys/callout.h>
52 #include <sys/errno.h>
54 #include <sys/thread.h>
58 #include <net/altq/altq.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>
64 #include <sys/thread2.h>
67 static struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
68 static struct cbqtrace *cbqtrace_ptr = NULL;
69 static int cbqtrace_count;
76 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
82 static int rmc_satisfied(struct rm_class *, struct timeval *);
83 static void rmc_wrr_set_weights(struct rm_ifdat *);
84 static void rmc_depth_compute(struct rm_class *);
85 static void rmc_depth_recompute(rm_class_t *);
87 static struct mbuf *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
88 static struct mbuf *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
90 static int _rmc_addq(rm_class_t *, struct mbuf *);
91 static void _rmc_dropq(rm_class_t *);
92 static struct mbuf *_rmc_getq(rm_class_t *);
93 static struct mbuf *_rmc_pollq(rm_class_t *);
95 static int rmc_under_limit(struct rm_class *, struct timeval *);
96 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
97 static void rmc_drop_action(struct rm_class *);
98 static void rmc_restart(void *);
99 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
101 #define BORROW_OFFTIME
103 * BORROW_OFFTIME (experimental):
104 * borrow the offtime of the class borrowing from.
105 * the reason is that when its own offtime is set, the class is unable
106 * to borrow much, especially when cutoff is taking effect.
107 * but when the borrowed class is overloaded (advidle is close to minidle),
108 * use the borrowing class's offtime to avoid overload.
110 #define ADJUST_CUTOFF
112 * ADJUST_CUTOFF (experimental):
113 * if no underlimit class is found due to cutoff, increase cutoff and
114 * retry the scheduling loop.
115 * also, don't invoke delay_actions while cutoff is taking effect,
116 * since a sleeping class won't have a chance to be scheduled in the
119 * now heuristics for setting the top-level variable (cutoff_) becomes:
120 * 1. if a packet arrives for a not-overlimit class, set cutoff
121 * to the depth of the class.
122 * 2. if cutoff is i, and a packet arrives for an overlimit class
123 * with an underlimit ancestor at a lower level than i (say j),
124 * then set cutoff to j.
125 * 3. at scheduling a packet, if there is no underlimit class
126 * due to the current cutoff level, increase cutoff by 1 and
127 * then try to schedule again.
132 * rmc_newclass(...) - Create a new resource management class at priority
133 * 'pri' on the interface given by 'ifd'.
135 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
136 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
137 * than 100% of the bandwidth, this number should be the
138 * 'effective' rate for the class. Let f be the
139 * bandwidth fraction allocated to this class, and let
140 * nsPerByte be the data rate of the output link in
141 * nanoseconds/byte. Then nsecPerByte is set to
142 * nsPerByte / f. E.g., 1600 (= 800 / .5)
143 * for a class that gets 50% of an ethernet's bandwidth.
145 * action the routine to call when the class is over limit.
147 * maxq max allowable queue size for class (in packets).
149 * parent parent class pointer.
151 * borrow class to borrow from (should be either 'parent' or null).
153 * maxidle max value allowed for class 'idle' time estimate (this
154 * parameter determines how large an initial burst of packets
155 * can be before overlimit action is invoked.
157 * offtime how long 'delay' action will delay when class goes over
158 * limit (this parameter determines the steady-state burst
159 * size when a class is running over its limit).
161 * Maxidle and offtime have to be computed from the following: If the
162 * average packet size is s, the bandwidth fraction allocated to this
163 * class is f, we want to allow b packet bursts, and the gain of the
164 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
166 * ptime = s * nsPerByte * (1 - f) / f
167 * maxidle = ptime * (1 - g^b) / g^b
168 * minidle = -ptime * (1 / (f - 1))
169 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
171 * Operationally, it's convenient to specify maxidle & offtime in units
172 * independent of the link bandwidth so the maxidle & offtime passed to
173 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
174 * (The constant factor is a scale factor needed to make the parameters
175 * integers. This scaling also means that the 'unscaled' values of
176 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
177 * not nanoseconds.) Also note that the 'idle' filter computation keeps
178 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
179 * maxidle also must be scaled upward by this value. Thus, the passed
180 * values for maxidle and offtime can be computed as follows:
182 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
183 * offtime = offtime * 8 / (1000 * nsecPerByte)
185 * When USE_HRTIME is employed, then maxidle and offtime become:
186 * maxidle = maxilde * (8.0 / nsecPerByte);
187 * offtime = offtime * (8.0 / nsecPerByte);
190 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
191 void (*action)(rm_class_t *, rm_class_t *), int maxq,
192 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
193 int minidle, u_int offtime, int pktsize, int flags)
196 struct rm_class *peer;
198 if (pri >= RM_MAXPRIO)
201 if (flags & RMCF_RED) {
203 kprintf("rmc_newclass: RED not configured for CBQ!\n");
209 if (flags & RMCF_RIO) {
211 kprintf("rmc_newclass: RIO not configured for CBQ!\n");
217 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
218 callout_init(&cl->callout_);
219 cl->q_ = kmalloc(sizeof(*cl->q_), M_ALTQ, M_WAITOK | M_ZERO);
222 * Class initialization.
224 cl->children_ = NULL;
225 cl->parent_ = parent;
226 cl->borrow_ = borrow;
230 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
233 cl->ns_per_byte_ = nsecPerByte;
235 qlimit(cl->q_) = maxq;
236 qtype(cl->q_) = Q_DROPHEAD;
240 #if 1 /* minidle is also scaled in ALTQ */
241 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
242 if (cl->minidle_ > 0)
245 cl->minidle_ = minidle;
247 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
248 if (cl->maxidle_ == 0)
250 #if 1 /* offtime is also scaled in ALTQ */
251 cl->avgidle_ = cl->maxidle_;
252 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
253 if (cl->offtime_ == 0)
257 cl->offtime_ = (offtime * nsecPerByte) / 8;
259 cl->overlimit = action;
262 if (flags & (RMCF_RED|RMCF_RIO)) {
263 int red_flags, red_pkttime;
266 if (flags & RMCF_ECN)
267 red_flags |= REDF_ECN;
269 if (flags & RMCF_CLEARDSCP)
270 red_flags |= RIOF_CLEARDSCP;
272 red_pkttime = nsecPerByte * pktsize / 1000;
274 if (flags & RMCF_RED) {
275 cl->red_ = red_alloc(0, 0,
276 qlimit(cl->q_) * 10/100,
277 qlimit(cl->q_) * 30/100,
278 red_flags, red_pkttime);
279 if (cl->red_ != NULL)
280 qtype(cl->q_) = Q_RED;
284 cl->red_ = (red_t *)rio_alloc(0, NULL,
285 red_flags, red_pkttime);
286 if (cl->red_ != NULL)
287 qtype(cl->q_) = Q_RIO;
291 #endif /* ALTQ_RED */
294 * put the class into the class tree
297 if ((peer = ifd->active_[pri]) != NULL) {
298 /* find the last class at this pri */
300 while (peer->peer_ != ifd->active_[pri])
304 ifd->active_[pri] = cl;
309 cl->next_ = parent->children_;
310 parent->children_ = cl;
315 * Compute the depth of this class and its ancestors in the class
318 rmc_depth_compute(cl);
321 * If CBQ's WRR is enabled, then initialize the class WRR state.
325 ifd->alloc_[pri] += cl->allotment_;
326 rmc_wrr_set_weights(ifd);
333 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
334 int minidle, u_int offtime, int pktsize)
336 struct rm_ifdat *ifd;
340 old_allotment = cl->allotment_;
343 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
345 cl->ns_per_byte_ = nsecPerByte;
347 qlimit(cl->q_) = maxq;
349 #if 1 /* minidle is also scaled in ALTQ */
350 cl->minidle_ = (minidle * nsecPerByte) / 8;
351 if (cl->minidle_ > 0)
354 cl->minidle_ = minidle;
356 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
357 if (cl->maxidle_ == 0)
359 #if 1 /* offtime is also scaled in ALTQ */
360 cl->avgidle_ = cl->maxidle_;
361 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
362 if (cl->offtime_ == 0)
366 cl->offtime_ = (offtime * nsecPerByte) / 8;
370 * If CBQ's WRR is enabled, then initialize the class WRR state.
373 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
374 rmc_wrr_set_weights(ifd);
382 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
383 * the appropriate run robin weights for the CBQ weighted round robin
390 rmc_wrr_set_weights(struct rm_ifdat *ifd)
393 struct rm_class *cl, *clh;
395 for (i = 0; i < RM_MAXPRIO; i++) {
397 * This is inverted from that of the simulator to
398 * maintain precision.
400 if (ifd->num_[i] == 0)
403 ifd->M_[i] = ifd->alloc_[i] /
404 (ifd->num_[i] * ifd->maxpkt_);
406 * Compute the weighted allotment for each class.
407 * This takes the expensive div instruction out
408 * of the main loop for the wrr scheduling path.
409 * These only get recomputed when a class comes or
412 if (ifd->active_[i] != NULL) {
413 clh = cl = ifd->active_[i];
415 /* safe-guard for slow link or alloc_ == 0 */
417 cl->w_allotment_ = 0;
419 cl->w_allotment_ = cl->allotment_ /
422 } while ((cl != NULL) && (cl != clh));
428 rmc_get_weight(struct rm_ifdat *ifd, int pri)
430 if ((pri >= 0) && (pri < RM_MAXPRIO))
431 return (ifd->M_[pri]);
438 * rmc_depth_compute(struct rm_class *cl) - This function computes the
439 * appropriate depth of class 'cl' and its ancestors.
445 rmc_depth_compute(struct rm_class *cl)
447 rm_class_t *t = cl, *p;
450 * Recompute the depth for the branch of the tree.
454 if (p && (t->depth_ >= p->depth_)) {
455 p->depth_ = t->depth_ + 1;
464 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
465 * the depth of the tree after a class has been deleted.
471 rmc_depth_recompute(rm_class_t *cl)
478 if ((t = p->children_) == NULL) {
484 if (t->depth_ > cdepth)
489 if (p->depth_ == cdepth + 1)
490 /* no change to this parent */
493 p->depth_ = cdepth + 1;
501 if (cl->depth_ >= 1) {
502 if (cl->children_ == NULL) {
504 } else if ((t = cl->children_) != NULL) {
506 if (t->children_ != NULL)
507 rmc_depth_recompute(t);
511 rmc_depth_compute(cl);
518 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
519 * function deletes a class from the link-sharing structure and frees
520 * all resources associated with the class.
526 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
528 struct rm_class *p, *head, *previous;
530 KKASSERT(cl->children_ == NULL);
533 callout_stop(&cl->callout_);
537 if (ifd->pollcache_ == cl)
538 ifd->pollcache_ = NULL;
541 * Free packets in the packet queue.
542 * XXX - this may not be a desired behavior. Packets should be
548 * If the class has a parent, then remove the class from the
549 * class from the parent's children chain.
551 if (cl->parent_ != NULL) {
552 head = cl->parent_->children_;
554 if (head->next_ == NULL) {
555 KKASSERT(head == cl);
556 cl->parent_->children_ = NULL;
557 cl->parent_->leaf_ = 1;
558 } else while (p != NULL) {
561 cl->parent_->children_ = cl->next_;
563 previous->next_ = cl->next_;
574 * Delete class from class priority peer list.
576 if ((p = ifd->active_[cl->pri_]) != NULL) {
578 * If there is more than one member of this priority
579 * level, then look for class(cl) in the priority level.
582 while (p->peer_ != cl)
584 p->peer_ = cl->peer_;
586 if (ifd->active_[cl->pri_] == cl)
587 ifd->active_[cl->pri_] = cl->peer_;
590 ifd->active_[cl->pri_] = NULL;
595 * Recompute the WRR weights.
598 ifd->alloc_[cl->pri_] -= cl->allotment_;
599 ifd->num_[cl->pri_]--;
600 rmc_wrr_set_weights(ifd);
604 * Re-compute the depth of the tree.
607 rmc_depth_recompute(cl->parent_);
609 rmc_depth_recompute(ifd->root_);
615 * Free the class structure.
617 if (cl->red_ != NULL) {
619 if (q_is_rio(cl->q_))
620 rio_destroy((rio_t *)cl->red_);
623 if (q_is_red(cl->q_))
624 red_destroy(cl->red_);
627 kfree(cl->q_, M_ALTQ);
633 * rmc_init(...) - Initialize the resource management data structures
634 * associated with the output portion of interface 'ifp'. 'ifd' is
635 * where the structures will be built (for backwards compatibility, the
636 * structures aren't kept in the ifnet struct). 'nsecPerByte'
637 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
638 * 'restart' is the driver-specific routine that the generic 'delay
639 * until under limit' action will call to restart output. `maxq'
640 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
641 * is the maximum number of packets that the resource management
642 * code will allow to be queued 'downstream' (this is typically 1).
648 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
649 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
650 int minidle, u_int offtime, int flags)
655 * Initialize the CBQ tracing/debug facility.
659 bzero(ifd, sizeof (*ifd));
660 mtu = ifq->altq_ifp->if_mtu;
662 ifd->restart = restart;
663 ifd->maxqueued_ = maxqueued;
664 ifd->ns_per_byte_ = nsecPerByte;
666 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
667 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
669 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
670 if (mtu * nsecPerByte > 10 * 1000000)
671 ifd->maxiftime_ /= 4;
675 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
678 * Initialize the CBQ's WRR state.
680 for (i = 0; i < RM_MAXPRIO; i++) {
685 ifd->active_[i] = NULL;
689 * Initialize current packet state.
693 for (i = 0; i < RM_MAXQUEUED; i++) {
694 ifd->class_[i] = NULL;
696 ifd->borrowed_[i] = NULL;
700 * Create the root class of the link-sharing structure.
702 ifd->root_ = rmc_newclass(0, ifd, nsecPerByte, rmc_root_overlimit,
703 maxq, 0, 0, maxidle, minidle, offtime, 0, 0);
704 if (ifd->root_ == NULL) {
705 kprintf("rmc_init: root class not allocated\n");
708 ifd->root_->depth_ = 0;
713 * rmc_queue_packet(struct rm_class *cl, struct mbuf *m) - Add packet given by
714 * mbuf 'm' to queue for resource class 'cl'. This routine is called
715 * by a driver's if_output routine. This routine must be called with
716 * output packet completion interrupts locked out (to avoid racing with
719 * Returns: 0 on successful queueing
720 * -1 when packet drop occurs
723 rmc_queue_packet(struct rm_class *cl, struct mbuf *m)
726 struct rm_ifdat *ifd = cl->ifdat_;
728 int is_empty = qempty(cl->q_);
731 if (ifd->cutoff_ > 0) {
732 if (TV_LT(&cl->undertime_, &now)) {
733 if (ifd->cutoff_ > cl->depth_)
734 ifd->cutoff_ = cl->depth_;
735 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
740 * the class is overlimit. if the class has
741 * underlimit ancestors, set cutoff to the lowest
744 struct rm_class *borrow = cl->borrow_;
746 while (borrow != NULL &&
747 borrow->depth_ < ifd->cutoff_) {
748 if (TV_LT(&borrow->undertime_, &now)) {
749 ifd->cutoff_ = borrow->depth_;
750 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
753 borrow = borrow->borrow_;
757 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
758 if (TV_LT(&cl->borrow_->undertime_, &now)) {
759 ifd->cutoff_ = cl->borrow_->depth_;
760 CBQTRACE(rmc_queue_packet, 'ffob',
761 cl->borrow_->depth_);
767 if (_rmc_addq(cl, m) < 0)
772 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
776 if (qlen(cl->q_) > qlimit(cl->q_)) {
777 /* note: qlimit can be set to 0 or 1 */
786 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
787 * classes to see if there are satified.
791 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
796 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
797 if ((bp = ifd->active_[i]) != NULL) {
800 if (!rmc_satisfied(p, now)) {
801 ifd->cutoff_ = p->depth_;
813 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
817 rmc_satisfied(struct rm_class *cl, struct timeval *now)
823 if (TV_LT(now, &cl->undertime_))
825 if (cl->depth_ == 0) {
826 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
831 if (cl->children_ != NULL) {
834 if (!rmc_satisfied(p, now))
844 * Return 1 if class 'cl' is under limit or can borrow from a parent,
845 * 0 if overlimit. As a side-effect, this routine will invoke the
846 * class overlimit action if the class if overlimit.
850 rmc_under_limit(struct rm_class *cl, struct timeval *now)
854 struct rm_ifdat *ifd = cl->ifdat_;
856 ifd->borrowed_[ifd->qi_] = NULL;
858 * If cl is the root class, then always return that it is
859 * underlimit. Otherwise, check to see if the class is underlimit.
861 if (cl->parent_ == NULL)
865 if (TV_LT(now, &cl->undertime_))
868 callout_stop(&cl->callout_);
870 cl->undertime_.tv_sec = 0;
875 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
876 if (((cl = cl->borrow_) == NULL) ||
877 (cl->depth_ > ifd->cutoff_)) {
880 /* cutoff is taking effect, just
881 return false without calling
885 #ifdef BORROW_OFFTIME
887 * check if the class can borrow offtime too.
888 * borrow offtime from the top of the borrow
889 * chain if the top class is not overloaded.
892 /* cutoff is taking effect, use this class as top. */
894 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
896 if (top != NULL && top->avgidle_ == top->minidle_)
899 (p->overlimit)(p, top);
902 (p->overlimit)(p, NULL);
910 ifd->borrowed_[ifd->qi_] = cl;
915 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
916 * Packet-by-packet round robin.
918 * The heart of the weighted round-robin scheduler, which decides which
919 * class next gets to send a packet. Highest priority first, then
920 * weighted round-robin within priorites.
922 * Each able-to-send class gets to send until its byte allocation is
923 * exhausted. Thus, the active pointer is only changed after a class has
924 * exhausted its allocation.
926 * If the scheduler finds no class that is underlimit or able to borrow,
927 * then the first class found that had a nonzero queue and is allowed to
928 * borrow gets to send.
932 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
934 struct rm_class *cl = NULL, *first = NULL;
943 * if the driver polls the top of the queue and then removes
944 * the polled packet, we must return the same packet.
946 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
947 cl = ifd->pollcache_;
949 if (ifd->efficient_) {
950 /* check if this class is overlimit */
951 if (cl->undertime_.tv_sec != 0 &&
952 rmc_under_limit(cl, &now) == 0)
955 ifd->pollcache_ = NULL;
959 /* mode == ALTDQ_POLL || pollcache == NULL */
960 ifd->pollcache_ = NULL;
961 ifd->borrowed_[ifd->qi_] = NULL;
966 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
967 if (ifd->na_[cpri] == 0)
971 * Loop through twice for a priority level, if some class
972 * was unable to send a packet the first round because
973 * of the weighted round-robin mechanism.
974 * During the second loop at this level, deficit==2.
975 * (This second loop is not needed if for every class,
976 * "M[cl->pri_])" times "cl->allotment" is greater than
977 * the byte size for the largest packet in the class.)
980 cl = ifd->active_[cpri];
981 KKASSERT(cl != NULL);
983 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
984 cl->bytes_alloc_ += cl->w_allotment_;
985 if (!qempty(cl->q_)) {
986 if ((cl->undertime_.tv_sec == 0) ||
987 rmc_under_limit(cl, &now)) {
988 if (cl->bytes_alloc_ > 0 || deficit > 1)
991 /* underlimit but no alloc */
994 ifd->borrowed_[ifd->qi_] = NULL;
997 else if (first == NULL && cl->borrow_ != NULL)
998 first = cl; /* borrowing candidate */
1001 cl->bytes_alloc_ = 0;
1003 } while (cl != ifd->active_[cpri]);
1006 /* first loop found an underlimit class with deficit */
1007 /* Loop on same priority level, with new deficit. */
1013 #ifdef ADJUST_CUTOFF
1015 * no underlimit class found. if cutoff is taking effect,
1016 * increase cutoff and try again.
1018 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1020 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1023 #endif /* ADJUST_CUTOFF */
1025 * If LINK_EFFICIENCY is turned on, then the first overlimit
1026 * class we encounter will send a packet if all the classes
1027 * of the link-sharing structure are overlimit.
1030 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1032 if (!ifd->efficient_ || first == NULL)
1037 #if 0 /* too time-consuming for nothing */
1039 callout_stop(&cl->callout_);
1041 cl->undertime_.tv_sec = 0;
1043 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1044 ifd->cutoff_ = cl->borrow_->depth_;
1047 * Deque the packet and do the book keeping...
1050 if (op == ALTDQ_REMOVE) {
1053 panic("_rmc_wrr_dequeue_next");
1058 * Update class statistics and link data.
1060 if (cl->bytes_alloc_ > 0)
1061 cl->bytes_alloc_ -= m_pktlen(m);
1063 if ((cl->bytes_alloc_ <= 0) || first == cl)
1064 ifd->active_[cl->pri_] = cl->peer_;
1066 ifd->active_[cl->pri_] = cl;
1068 ifd->class_[ifd->qi_] = cl;
1069 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1070 ifd->now_[ifd->qi_] = now;
1071 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1074 /* mode == ALTDQ_PPOLL */
1076 ifd->pollcache_ = cl;
1082 * Dequeue & return next packet from the highest priority class that
1083 * has a packet to send & has enough allocation to send it. This
1084 * routine is called by a driver whenever it needs a new packet to
1087 static struct mbuf *
1088 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1092 struct rm_class *cl, *first = NULL;
1098 * if the driver polls the top of the queue and then removes
1099 * the polled packet, we must return the same packet.
1101 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1102 cl = ifd->pollcache_;
1104 ifd->pollcache_ = NULL;
1107 /* mode == ALTDQ_POLL || pollcache == NULL */
1108 ifd->pollcache_ = NULL;
1109 ifd->borrowed_[ifd->qi_] = NULL;
1111 #ifdef ADJUST_CUTOFF
1114 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1115 if (ifd->na_[cpri] == 0)
1117 cl = ifd->active_[cpri];
1118 KKASSERT(cl != NULL);
1120 if (!qempty(cl->q_)) {
1121 if ((cl->undertime_.tv_sec == 0) ||
1122 rmc_under_limit(cl, &now))
1124 if (first == NULL && cl->borrow_ != NULL)
1128 } while (cl != ifd->active_[cpri]);
1131 #ifdef ADJUST_CUTOFF
1133 * no underlimit class found. if cutoff is taking effect, increase
1134 * cutoff and try again.
1136 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1140 #endif /* ADJUST_CUTOFF */
1142 * If LINK_EFFICIENCY is turned on, then the first overlimit
1143 * class we encounter will send a packet if all the classes
1144 * of the link-sharing structure are overlimit.
1147 if (!ifd->efficient_ || first == NULL)
1152 #if 0 /* too time-consuming for nothing */
1154 callout_stop(&cl->callout_);
1156 cl->undertime_.tv_sec = 0;
1158 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1159 ifd->cutoff_ = cl->borrow_->depth_;
1162 * Deque the packet and do the book keeping...
1165 if (op == ALTDQ_REMOVE) {
1168 panic("_rmc_prr_dequeue_next");
1172 ifd->active_[cpri] = cl->peer_;
1174 ifd->class_[ifd->qi_] = cl;
1175 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1176 ifd->now_[ifd->qi_] = now;
1177 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1180 /* mode == ALTDQ_POLL */
1182 ifd->pollcache_ = cl;
1189 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1190 * is invoked by the packet driver to get the next packet to be
1191 * dequeued and output on the link. If WRR is enabled, then the
1192 * WRR dequeue next routine will determine the next packet to sent.
1193 * Otherwise, packet-by-packet round robin is invoked.
1195 * Returns: NULL, if a packet is not available or if all
1196 * classes are overlimit.
1198 * Otherwise, Pointer to the next packet.
1202 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1204 if (ifd->queued_ >= ifd->maxqueued_)
1207 return (_rmc_wrr_dequeue_next(ifd, mode));
1209 return (_rmc_prr_dequeue_next(ifd, mode));
1213 * Update the utilization estimate for the packet that just completed.
1214 * The packet's class & the parent(s) of that class all get their
1215 * estimators updated. This routine is called by the driver's output-
1216 * packet-completion interrupt service routine.
1220 * a macro to approximate "divide by 1000" that gives 0.000999,
1221 * if a value has enough effective digits.
1222 * (on pentium, mul takes 9 cycles but div takes 46!)
1224 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1226 rmc_update_class_util(struct rm_ifdat *ifd)
1228 int idle, avgidle, pktlen;
1229 int pkt_time, tidle;
1230 rm_class_t *cl, *borrowed;
1231 rm_class_t *borrows;
1232 struct timeval *nowp;
1235 * Get the most recent completed class.
1237 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1240 pktlen = ifd->curlen_[ifd->qo_];
1241 borrowed = ifd->borrowed_[ifd->qo_];
1244 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1247 * Run estimator on class and its ancestors.
1250 * rm_update_class_util is designed to be called when the
1251 * transfer is completed from a xmit complete interrupt,
1252 * but most drivers don't implement an upcall for that.
1253 * so, just use estimated completion time.
1254 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1256 nowp = &ifd->now_[ifd->qo_];
1257 /* get pkt_time (for link) in usec */
1258 #if 1 /* use approximation */
1259 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1260 pkt_time = NSEC_TO_USEC(pkt_time);
1262 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1264 #if 1 /* ALTQ4PPP */
1265 if (TV_LT(nowp, &ifd->ifnow_)) {
1269 * make sure the estimated completion time does not go
1270 * too far. it can happen when the link layer supports
1271 * data compression or the interface speed is set to
1272 * a much lower value.
1274 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1275 if (iftime+pkt_time < ifd->maxiftime_) {
1276 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1278 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1281 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1284 if (TV_LT(nowp, &ifd->ifnow_)) {
1285 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1287 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1291 while (cl != NULL) {
1292 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1293 if (idle >= 2000000)
1295 * this class is idle enough, reset avgidle.
1296 * (TV_DELTA returns 2000000 us when delta is large.)
1298 cl->avgidle_ = cl->maxidle_;
1300 /* get pkt_time (for class) in usec */
1301 #if 1 /* use approximation */
1302 pkt_time = pktlen * cl->ns_per_byte_;
1303 pkt_time = NSEC_TO_USEC(pkt_time);
1305 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1309 avgidle = cl->avgidle_;
1310 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1311 cl->avgidle_ = avgidle;
1313 /* Are we overlimit ? */
1315 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1318 * need some lower bound for avgidle, otherwise
1319 * a borrowing class gets unbounded penalty.
1321 if (avgidle < cl->minidle_)
1322 avgidle = cl->avgidle_ = cl->minidle_;
1324 /* set next idle to make avgidle 0 */
1326 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1327 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1331 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1332 cl->undertime_.tv_sec = 0;
1333 if (cl->sleeping_) {
1334 callout_stop(&cl->callout_);
1339 if (borrows != NULL) {
1341 ++cl->stats_.borrows;
1345 cl->last_ = ifd->ifnow_;
1346 cl->last_pkttime_ = pkt_time;
1349 if (cl->parent_ == NULL) {
1350 /* take stats of root class */
1351 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1359 * Check to see if cutoff needs to set to a new level.
1361 cl = ifd->class_[ifd->qo_];
1362 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1364 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1365 rmc_tl_satisfied(ifd, nowp);
1366 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1368 ifd->cutoff_ = borrowed->depth_;
1369 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1372 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1375 rmc_tl_satisfied(ifd, &now);
1377 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1379 ifd->cutoff_ = borrowed->depth_;
1380 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1386 * Release class slot
1388 ifd->borrowed_[ifd->qo_] = NULL;
1389 ifd->class_[ifd->qo_] = NULL;
1390 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1396 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1397 * over-limit action routines. These get invoked by rmc_under_limit()
1398 * if a class with packets to send if over its bandwidth limit & can't
1399 * borrow from a parent class.
1405 rmc_drop_action(struct rm_class *cl)
1407 struct rm_ifdat *ifd = cl->ifdat_;
1409 KKASSERT(qlen(cl->q_) > 0);
1412 ifd->na_[cl->pri_]--;
1416 rmc_dropall(struct rm_class *cl)
1418 struct rm_ifdat *ifd = cl->ifdat_;
1420 if (!qempty(cl->q_)) {
1423 ifd->na_[cl->pri_]--;
1429 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1430 * delay action routine. It is invoked via rmc_under_limit when the
1431 * packet is discoverd to be overlimit.
1433 * If the delay action is result of borrow class being overlimit, then
1434 * delay for the offtime of the borrowing class that is overlimit.
1440 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1442 int delay, t, extradelay;
1444 cl->stats_.overactions++;
1445 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1446 #ifndef BORROW_OFFTIME
1447 delay += cl->offtime_;
1450 if (!cl->sleeping_) {
1451 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1452 #ifdef BORROW_OFFTIME
1454 extradelay = borrow->offtime_;
1457 extradelay = cl->offtime_;
1461 * XXX recalculate suspend time:
1462 * current undertime is (tidle + pkt_time) calculated
1463 * from the last transmission.
1464 * tidle: time required to bring avgidle back to 0
1465 * pkt_time: target waiting time for this class
1466 * we need to replace pkt_time by offtime
1468 extradelay -= cl->last_pkttime_;
1470 if (extradelay > 0) {
1471 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1472 delay += extradelay;
1476 cl->stats_.delays++;
1479 * Since packets are phased randomly with respect to the
1480 * clock, 1 tick (the next clock tick) can be an arbitrarily
1481 * short time so we have to wait for at least two ticks.
1482 * NOTE: If there's no other traffic, we need the timer as
1483 * a 'backstop' to restart this class.
1485 if (delay > ustick * 2)
1486 t = (delay + ustick - 1) / ustick;
1489 callout_reset(&cl->callout_, t, rmc_restart, cl);
1495 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1496 * called by the system timer code & is responsible checking if the
1497 * class is still sleeping (it might have been restarted as a side
1498 * effect of the queue scan on a packet arrival) and, if so, restarting
1499 * output for the class. Inspecting the class state & restarting output
1500 * require locking the class structure. In general the driver is
1501 * responsible for locking but this is the only routine that is not
1502 * called directly or indirectly from the interface driver so it has
1503 * know about system locking conventions. Under bsd, locking is done
1504 * by raising IPL to splimp so that's what's implemented here. On a
1505 * different system this would probably need to be changed.
1507 * Since this function is called from an independant timeout, we
1508 * have to set up the lock conditions expected for the ALTQ operation.
1509 * Note that the restart will probably fall through to an if_start.
1515 rmc_restart(void *arg)
1517 struct rm_class *cl = arg;
1518 struct rm_ifdat *ifd = cl->ifdat_;
1520 ALTQ_LOCK(ifd->ifq_);
1521 if (cl->sleeping_) {
1523 cl->undertime_.tv_sec = 0;
1525 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1526 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1527 (ifd->restart)(ifd->ifq_);
1530 ALTQ_UNLOCK(ifd->ifq_);
1535 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1536 * handling routine for the root class of the link sharing structure.
1542 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1544 panic("rmc_root_overlimit");
1548 * Packet Queue handling routines. Eventually, this is to localize the
1549 * effects on the code whether queues are red queues or droptail
1554 _rmc_addq(rm_class_t *cl, struct mbuf *m)
1557 if (q_is_rio(cl->q_))
1558 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1561 if (q_is_red(cl->q_))
1562 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1563 #endif /* ALTQ_RED */
1565 if (cl->flags_ & RMCF_CLEARDSCP)
1566 write_dsfield(m, cl->pktattr_, 0);
1572 /* note: _rmc_dropq is not called for red */
1574 _rmc_dropq(rm_class_t *cl)
1578 if ((m = _getq(cl->q_)) != NULL)
1582 static struct mbuf *
1583 _rmc_getq(rm_class_t *cl)
1586 if (q_is_rio(cl->q_))
1587 return rio_getq((rio_t *)cl->red_, cl->q_);
1590 if (q_is_red(cl->q_))
1591 return red_getq(cl->red_, cl->q_);
1593 return _getq(cl->q_);
1596 static struct mbuf *
1597 _rmc_pollq(rm_class_t *cl)
1599 return qhead(cl->q_);
1604 * DDB hook to trace cbq events:
1605 * the last 1024 events are held in a circular buffer.
1606 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1608 void cbqtrace_dump(int);
1609 static char *rmc_funcname(void *);
1611 static struct rmc_funcs {
1615 rmc_init, "rmc_init",
1616 rmc_queue_packet, "rmc_queue_packet",
1617 rmc_under_limit, "rmc_under_limit",
1618 rmc_update_class_util, "rmc_update_class_util",
1619 rmc_delay_action, "rmc_delay_action",
1620 rmc_restart, "rmc_restart",
1621 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1626 rmc_funcname(void *func)
1628 struct rmc_funcs *fp;
1630 for (fp = rmc_funcs; fp->func != NULL; fp++) {
1631 if (fp->func == func)
1639 cbqtrace_dump(int counter)
1644 counter = counter % NCBQTRACE;
1645 p = (int *)&cbqtrace_buffer[counter];
1647 for (i=0; i<20; i++) {
1648 kprintf("[0x%x] ", *p++);
1649 kprintf("%s: ", rmc_funcname((void *)*p++));
1651 kprintf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1652 kprintf("%d\n",*p++);
1654 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1655 p = (int *)cbqtrace_buffer;
1658 #endif /* CBQ_TRACE */
1659 #endif /* ALTQ_CBQ */